From 11da511c784eca003deb90c23570f0873954e0de Mon Sep 17 00:00:00 2001 From: Duncan Wilkie Date: Sat, 18 Nov 2023 06:11:09 -0600 Subject: Initial commit. --- gmp-6.3.0/mpn/x86/README | 525 +++++++++++ gmp-6.3.0/mpn/x86/aors_n.asm | 202 +++++ gmp-6.3.0/mpn/x86/aorsmul_1.asm | 214 +++++ gmp-6.3.0/mpn/x86/atom/aorrlsh1_n.asm | 53 ++ gmp-6.3.0/mpn/x86/atom/aorrlsh2_n.asm | 53 ++ gmp-6.3.0/mpn/x86/atom/aorrlshC_n.asm | 156 ++++ gmp-6.3.0/mpn/x86/atom/aors_n.asm | 159 ++++ gmp-6.3.0/mpn/x86/atom/aorslshC_n.asm | 247 ++++++ gmp-6.3.0/mpn/x86/atom/bdiv_q_1.asm | 35 + gmp-6.3.0/mpn/x86/atom/cnd_add_n.asm | 113 +++ gmp-6.3.0/mpn/x86/atom/cnd_sub_n.asm | 124 +++ gmp-6.3.0/mpn/x86/atom/dive_1.asm | 34 + gmp-6.3.0/mpn/x86/atom/gmp-mparam.h | 214 +++++ gmp-6.3.0/mpn/x86/atom/logops_n.asm | 151 ++++ gmp-6.3.0/mpn/x86/atom/lshift.asm | 218 +++++ gmp-6.3.0/mpn/x86/atom/lshiftc.asm | 159 ++++ gmp-6.3.0/mpn/x86/atom/mmx/copyd.asm | 34 + gmp-6.3.0/mpn/x86/atom/mmx/copyi.asm | 34 + gmp-6.3.0/mpn/x86/atom/mmx/hamdist.asm | 34 + gmp-6.3.0/mpn/x86/atom/mod_34lsub1.asm | 34 + gmp-6.3.0/mpn/x86/atom/mode1o.asm | 34 + gmp-6.3.0/mpn/x86/atom/rshift.asm | 152 ++++ gmp-6.3.0/mpn/x86/atom/sse2/aorsmul_1.asm | 174 ++++ gmp-6.3.0/mpn/x86/atom/sse2/bdiv_dbm1c.asm | 34 + gmp-6.3.0/mpn/x86/atom/sse2/divrem_1.asm | 34 + gmp-6.3.0/mpn/x86/atom/sse2/mod_1_1.asm | 34 + gmp-6.3.0/mpn/x86/atom/sse2/mod_1_4.asm | 34 + gmp-6.3.0/mpn/x86/atom/sse2/mul_1.asm | 124 +++ gmp-6.3.0/mpn/x86/atom/sse2/mul_basecase.asm | 501 +++++++++++ gmp-6.3.0/mpn/x86/atom/sse2/popcount.asm | 35 + gmp-6.3.0/mpn/x86/atom/sse2/sqr_basecase.asm | 634 ++++++++++++++ gmp-6.3.0/mpn/x86/atom/sublsh1_n.asm | 34 + gmp-6.3.0/mpn/x86/atom/sublsh2_n.asm | 57 ++ gmp-6.3.0/mpn/x86/bd1/gmp-mparam.h | 211 +++++ gmp-6.3.0/mpn/x86/bd2/gmp-mparam.h | 214 +++++ gmp-6.3.0/mpn/x86/bd4/gmp-mparam.h | 225 +++++ gmp-6.3.0/mpn/x86/bdiv_dbm1c.asm | 129 +++ gmp-6.3.0/mpn/x86/bdiv_q_1.asm | 208 +++++ gmp-6.3.0/mpn/x86/bt1/gmp-mparam.h | 218 +++++ gmp-6.3.0/mpn/x86/bt2/gmp-mparam.h | 214 +++++ gmp-6.3.0/mpn/x86/cnd_aors_n.asm | 124 +++ gmp-6.3.0/mpn/x86/copyd.asm | 91 ++ gmp-6.3.0/mpn/x86/copyi.asm | 99 +++ gmp-6.3.0/mpn/x86/core2/gmp-mparam.h | 210 +++++ gmp-6.3.0/mpn/x86/coreibwl/gmp-mparam.h | 216 +++++ gmp-6.3.0/mpn/x86/coreihwl/gmp-mparam.h | 216 +++++ gmp-6.3.0/mpn/x86/coreinhm/gmp-mparam.h | 223 +++++ gmp-6.3.0/mpn/x86/coreisbr/gmp-mparam.h | 215 +++++ gmp-6.3.0/mpn/x86/darwin.m4 | 102 +++ gmp-6.3.0/mpn/x86/dive_1.asm | 190 ++++ gmp-6.3.0/mpn/x86/divrem_1.asm | 233 +++++ gmp-6.3.0/mpn/x86/divrem_2.asm | 199 +++++ gmp-6.3.0/mpn/x86/fat/com.c | 32 + gmp-6.3.0/mpn/x86/fat/fat.c | 530 +++++++++++ gmp-6.3.0/mpn/x86/fat/fat_entry.asm | 243 +++++ gmp-6.3.0/mpn/x86/fat/gmp-mparam.h | 71 ++ gmp-6.3.0/mpn/x86/fat/lshiftc.c | 32 + gmp-6.3.0/mpn/x86/fat/mod_1.c | 32 + gmp-6.3.0/mpn/x86/fat/mod_1_1.c | 36 + gmp-6.3.0/mpn/x86/fat/mod_1_2.c | 36 + gmp-6.3.0/mpn/x86/fat/mod_1_4.c | 36 + gmp-6.3.0/mpn/x86/fat/mode1o.c | 32 + gmp-6.3.0/mpn/x86/fat/mullo_basecase.c | 32 + gmp-6.3.0/mpn/x86/fat/redc_1.c | 32 + gmp-6.3.0/mpn/x86/fat/redc_2.c | 32 + gmp-6.3.0/mpn/x86/gcd_11.asm | 126 +++ gmp-6.3.0/mpn/x86/geode/gmp-mparam.h | 141 +++ gmp-6.3.0/mpn/x86/gmp-mparam.h | 38 + gmp-6.3.0/mpn/x86/goldmont/gmp-mparam.h | 219 +++++ gmp-6.3.0/mpn/x86/i486/gmp-mparam.h | 69 ++ gmp-6.3.0/mpn/x86/k10/gmp-mparam.h | 217 +++++ gmp-6.3.0/mpn/x86/k6/README | 251 ++++++ gmp-6.3.0/mpn/x86/k6/aors_n.asm | 337 +++++++ gmp-6.3.0/mpn/x86/k6/aorsmul_1.asm | 391 +++++++++ gmp-6.3.0/mpn/x86/k6/cross.pl | 182 ++++ gmp-6.3.0/mpn/x86/k6/divrem_1.asm | 203 +++++ gmp-6.3.0/mpn/x86/k6/gmp-mparam.h | 166 ++++ gmp-6.3.0/mpn/x86/k6/k62mmx/copyd.asm | 118 +++ gmp-6.3.0/mpn/x86/k6/k62mmx/lshift.asm | 294 +++++++ gmp-6.3.0/mpn/x86/k6/k62mmx/rshift.asm | 293 +++++++ gmp-6.3.0/mpn/x86/k6/mmx/com.asm | 103 +++ gmp-6.3.0/mpn/x86/k6/mmx/dive_1.asm | 282 ++++++ gmp-6.3.0/mpn/x86/k6/mmx/logops_n.asm | 226 +++++ gmp-6.3.0/mpn/x86/k6/mmx/lshift.asm | 130 +++ gmp-6.3.0/mpn/x86/k6/mmx/popham.asm | 236 +++++ gmp-6.3.0/mpn/x86/k6/mmx/rshift.asm | 130 +++ gmp-6.3.0/mpn/x86/k6/mod_34lsub1.asm | 190 ++++ gmp-6.3.0/mpn/x86/k6/mode1o.asm | 176 ++++ gmp-6.3.0/mpn/x86/k6/mul_1.asm | 292 ++++++ gmp-6.3.0/mpn/x86/k6/mul_basecase.asm | 612 +++++++++++++ gmp-6.3.0/mpn/x86/k6/pre_mod_1.asm | 146 +++ gmp-6.3.0/mpn/x86/k6/sqr_basecase.asm | 680 ++++++++++++++ gmp-6.3.0/mpn/x86/k7/README | 174 ++++ gmp-6.3.0/mpn/x86/k7/addlsh1_n.asm | 196 +++++ gmp-6.3.0/mpn/x86/k7/aors_n.asm | 258 ++++++ gmp-6.3.0/mpn/x86/k7/aorsmul_1.asm | 167 ++++ gmp-6.3.0/mpn/x86/k7/bdiv_q_1.asm | 245 ++++++ gmp-6.3.0/mpn/x86/k7/dive_1.asm | 208 +++++ gmp-6.3.0/mpn/x86/k7/gcd_11.asm | 107 +++ gmp-6.3.0/mpn/x86/k7/gmp-mparam.h | 263 ++++++ gmp-6.3.0/mpn/x86/k7/invert_limb.asm | 194 ++++ gmp-6.3.0/mpn/x86/k7/mmx/com.asm | 125 +++ gmp-6.3.0/mpn/x86/k7/mmx/copyd.asm | 144 +++ gmp-6.3.0/mpn/x86/k7/mmx/copyi.asm | 157 ++++ gmp-6.3.0/mpn/x86/k7/mmx/divrem_1.asm | 832 ++++++++++++++++++ gmp-6.3.0/mpn/x86/k7/mmx/lshift.asm | 481 ++++++++++ gmp-6.3.0/mpn/x86/k7/mmx/popham.asm | 213 +++++ gmp-6.3.0/mpn/x86/k7/mmx/rshift.asm | 480 ++++++++++ gmp-6.3.0/mpn/x86/k7/mod_1_1.asm | 221 +++++ gmp-6.3.0/mpn/x86/k7/mod_1_4.asm | 260 ++++++ gmp-6.3.0/mpn/x86/k7/mod_34lsub1.asm | 188 ++++ gmp-6.3.0/mpn/x86/k7/mode1o.asm | 181 ++++ gmp-6.3.0/mpn/x86/k7/mul_1.asm | 237 +++++ gmp-6.3.0/mpn/x86/k7/mul_basecase.asm | 602 +++++++++++++ gmp-6.3.0/mpn/x86/k7/sqr_basecase.asm | 635 ++++++++++++++ gmp-6.3.0/mpn/x86/k7/sublsh1_n.asm | 173 ++++ gmp-6.3.0/mpn/x86/k8/gmp-mparam.h | 215 +++++ gmp-6.3.0/mpn/x86/lshift.asm | 106 +++ gmp-6.3.0/mpn/x86/mmx/sec_tabselect.asm | 163 ++++ gmp-6.3.0/mpn/x86/mod_34lsub1.asm | 183 ++++ gmp-6.3.0/mpn/x86/mul_1.asm | 140 +++ gmp-6.3.0/mpn/x86/mul_basecase.asm | 223 +++++ gmp-6.3.0/mpn/x86/nano/gmp-mparam.h | 162 ++++ gmp-6.3.0/mpn/x86/p6/README | 125 +++ gmp-6.3.0/mpn/x86/p6/aors_n.asm | 156 ++++ gmp-6.3.0/mpn/x86/p6/aorsmul_1.asm | 320 +++++++ gmp-6.3.0/mpn/x86/p6/bdiv_q_1.asm | 287 ++++++ gmp-6.3.0/mpn/x86/p6/copyd.asm | 178 ++++ gmp-6.3.0/mpn/x86/p6/dive_1.asm | 267 ++++++ gmp-6.3.0/mpn/x86/p6/gcd_11.asm | 83 ++ gmp-6.3.0/mpn/x86/p6/gmp-mparam.h | 194 ++++ gmp-6.3.0/mpn/x86/p6/lshsub_n.asm | 169 ++++ gmp-6.3.0/mpn/x86/p6/mmx/divrem_1.asm | 767 ++++++++++++++++ gmp-6.3.0/mpn/x86/p6/mmx/gmp-mparam.h | 218 +++++ gmp-6.3.0/mpn/x86/p6/mmx/lshift.asm | 38 + gmp-6.3.0/mpn/x86/p6/mmx/popham.asm | 39 + gmp-6.3.0/mpn/x86/p6/mmx/rshift.asm | 38 + gmp-6.3.0/mpn/x86/p6/mod_34lsub1.asm | 190 ++++ gmp-6.3.0/mpn/x86/p6/mode1o.asm | 170 ++++ gmp-6.3.0/mpn/x86/p6/mul_basecase.asm | 607 +++++++++++++ gmp-6.3.0/mpn/x86/p6/p3mmx/popham.asm | 42 + gmp-6.3.0/mpn/x86/p6/sqr_basecase.asm | 649 ++++++++++++++ gmp-6.3.0/mpn/x86/p6/sse2/addmul_1.asm | 37 + gmp-6.3.0/mpn/x86/p6/sse2/gmp-mparam.h | 200 +++++ gmp-6.3.0/mpn/x86/p6/sse2/mod_1_1.asm | 34 + gmp-6.3.0/mpn/x86/p6/sse2/mod_1_4.asm | 34 + gmp-6.3.0/mpn/x86/p6/sse2/mul_1.asm | 38 + gmp-6.3.0/mpn/x86/p6/sse2/mul_basecase.asm | 35 + gmp-6.3.0/mpn/x86/p6/sse2/popcount.asm | 35 + gmp-6.3.0/mpn/x86/p6/sse2/sqr_basecase.asm | 35 + gmp-6.3.0/mpn/x86/p6/sse2/submul_1.asm | 35 + gmp-6.3.0/mpn/x86/pentium/README | 181 ++++ gmp-6.3.0/mpn/x86/pentium/aors_n.asm | 203 +++++ gmp-6.3.0/mpn/x86/pentium/aorsmul_1.asm | 144 +++ gmp-6.3.0/mpn/x86/pentium/bdiv_q_1.asm | 266 ++++++ gmp-6.3.0/mpn/x86/pentium/com.asm | 181 ++++ gmp-6.3.0/mpn/x86/pentium/copyd.asm | 146 +++ gmp-6.3.0/mpn/x86/pentium/copyi.asm | 164 ++++ gmp-6.3.0/mpn/x86/pentium/dive_1.asm | 264 ++++++ gmp-6.3.0/mpn/x86/pentium/gmp-mparam.h | 76 ++ gmp-6.3.0/mpn/x86/pentium/hamdist.asm | 154 ++++ gmp-6.3.0/mpn/x86/pentium/logops_n.asm | 176 ++++ gmp-6.3.0/mpn/x86/pentium/lshift.asm | 243 +++++ gmp-6.3.0/mpn/x86/pentium/mmx/gmp-mparam.h | 163 ++++ gmp-6.3.0/mpn/x86/pentium/mmx/hamdist.asm | 40 + gmp-6.3.0/mpn/x86/pentium/mmx/lshift.asm | 463 ++++++++++ gmp-6.3.0/mpn/x86/pentium/mmx/mul_1.asm | 371 ++++++++ gmp-6.3.0/mpn/x86/pentium/mmx/rshift.asm | 468 ++++++++++ gmp-6.3.0/mpn/x86/pentium/mod_34lsub1.asm | 192 ++++ gmp-6.3.0/mpn/x86/pentium/mode1o.asm | 279 ++++++ gmp-6.3.0/mpn/x86/pentium/mul_1.asm | 177 ++++ gmp-6.3.0/mpn/x86/pentium/mul_2.asm | 150 ++++ gmp-6.3.0/mpn/x86/pentium/mul_basecase.asm | 142 +++ gmp-6.3.0/mpn/x86/pentium/popcount.asm | 146 +++ gmp-6.3.0/mpn/x86/pentium/rshift.asm | 243 +++++ gmp-6.3.0/mpn/x86/pentium/sqr_basecase.asm | 528 +++++++++++ gmp-6.3.0/mpn/x86/pentium4/README | 124 +++ gmp-6.3.0/mpn/x86/pentium4/copyd.asm | 71 ++ gmp-6.3.0/mpn/x86/pentium4/copyi.asm | 93 ++ gmp-6.3.0/mpn/x86/pentium4/mmx/lshift.asm | 39 + gmp-6.3.0/mpn/x86/pentium4/mmx/popham.asm | 203 +++++ gmp-6.3.0/mpn/x86/pentium4/mmx/rshift.asm | 39 + gmp-6.3.0/mpn/x86/pentium4/sse2/add_n.asm | 101 +++ gmp-6.3.0/mpn/x86/pentium4/sse2/addlsh1_n.asm | 108 +++ gmp-6.3.0/mpn/x86/pentium4/sse2/addmul_1.asm | 189 ++++ gmp-6.3.0/mpn/x86/pentium4/sse2/bdiv_dbm1c.asm | 141 +++ gmp-6.3.0/mpn/x86/pentium4/sse2/bdiv_q_1.asm | 234 +++++ gmp-6.3.0/mpn/x86/pentium4/sse2/cnd_add_n.asm | 95 ++ gmp-6.3.0/mpn/x86/pentium4/sse2/cnd_sub_n.asm | 114 +++ gmp-6.3.0/mpn/x86/pentium4/sse2/dive_1.asm | 216 +++++ gmp-6.3.0/mpn/x86/pentium4/sse2/divrem_1.asm | 645 ++++++++++++++ gmp-6.3.0/mpn/x86/pentium4/sse2/gmp-mparam.h | 213 +++++ gmp-6.3.0/mpn/x86/pentium4/sse2/mod_1_1.asm | 166 ++++ gmp-6.3.0/mpn/x86/pentium4/sse2/mod_1_4.asm | 269 ++++++ gmp-6.3.0/mpn/x86/pentium4/sse2/mod_34lsub1.asm | 175 ++++ gmp-6.3.0/mpn/x86/pentium4/sse2/mode1o.asm | 175 ++++ gmp-6.3.0/mpn/x86/pentium4/sse2/mul_1.asm | 164 ++++ gmp-6.3.0/mpn/x86/pentium4/sse2/mul_basecase.asm | 662 ++++++++++++++ gmp-6.3.0/mpn/x86/pentium4/sse2/popcount.asm | 281 ++++++ gmp-6.3.0/mpn/x86/pentium4/sse2/rsh1add_n.asm | 126 +++ gmp-6.3.0/mpn/x86/pentium4/sse2/sqr_basecase.asm | 705 +++++++++++++++ gmp-6.3.0/mpn/x86/pentium4/sse2/sub_n.asm | 119 +++ gmp-6.3.0/mpn/x86/pentium4/sse2/submul_1.asm | 182 ++++ gmp-6.3.0/mpn/x86/rshift.asm | 108 +++ gmp-6.3.0/mpn/x86/sec_tabselect.asm | 106 +++ gmp-6.3.0/mpn/x86/silvermont/gmp-mparam.h | 222 +++++ gmp-6.3.0/mpn/x86/skylake/gmp-mparam.h | 211 +++++ gmp-6.3.0/mpn/x86/sqr_basecase.asm | 359 ++++++++ gmp-6.3.0/mpn/x86/t-zdisp.sh | 71 ++ gmp-6.3.0/mpn/x86/t-zdisp2.pl | 147 ++++ gmp-6.3.0/mpn/x86/udiv.asm | 52 ++ gmp-6.3.0/mpn/x86/umul.asm | 51 ++ gmp-6.3.0/mpn/x86/x86-defs.m4 | 1024 ++++++++++++++++++++++ gmp-6.3.0/mpn/x86/zn1/gmp-mparam.h | 220 +++++ gmp-6.3.0/mpn/x86/zn2/gmp-mparam.h | 226 +++++ 215 files changed, 42594 insertions(+) create mode 100644 gmp-6.3.0/mpn/x86/README create mode 100644 gmp-6.3.0/mpn/x86/aors_n.asm create mode 100644 gmp-6.3.0/mpn/x86/aorsmul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/aorrlsh1_n.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/aorrlsh2_n.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/aorrlshC_n.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/aors_n.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/aorslshC_n.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/bdiv_q_1.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/cnd_add_n.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/cnd_sub_n.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/dive_1.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/atom/logops_n.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/lshift.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/lshiftc.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/mmx/copyd.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/mmx/copyi.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/mmx/hamdist.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/mod_34lsub1.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/mode1o.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/rshift.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sse2/aorsmul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sse2/bdiv_dbm1c.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sse2/divrem_1.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sse2/mod_1_1.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sse2/mod_1_4.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sse2/mul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sse2/mul_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sse2/popcount.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sse2/sqr_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sublsh1_n.asm create mode 100644 gmp-6.3.0/mpn/x86/atom/sublsh2_n.asm create mode 100644 gmp-6.3.0/mpn/x86/bd1/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/bd2/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/bd4/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/bdiv_dbm1c.asm create mode 100644 gmp-6.3.0/mpn/x86/bdiv_q_1.asm create mode 100644 gmp-6.3.0/mpn/x86/bt1/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/bt2/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/cnd_aors_n.asm create mode 100644 gmp-6.3.0/mpn/x86/copyd.asm create mode 100644 gmp-6.3.0/mpn/x86/copyi.asm create mode 100644 gmp-6.3.0/mpn/x86/core2/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/coreibwl/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/coreihwl/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/coreinhm/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/coreisbr/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/darwin.m4 create mode 100644 gmp-6.3.0/mpn/x86/dive_1.asm create mode 100644 gmp-6.3.0/mpn/x86/divrem_1.asm create mode 100644 gmp-6.3.0/mpn/x86/divrem_2.asm create mode 100644 gmp-6.3.0/mpn/x86/fat/com.c create mode 100644 gmp-6.3.0/mpn/x86/fat/fat.c create mode 100644 gmp-6.3.0/mpn/x86/fat/fat_entry.asm create mode 100644 gmp-6.3.0/mpn/x86/fat/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/fat/lshiftc.c create mode 100644 gmp-6.3.0/mpn/x86/fat/mod_1.c create mode 100644 gmp-6.3.0/mpn/x86/fat/mod_1_1.c create mode 100644 gmp-6.3.0/mpn/x86/fat/mod_1_2.c create mode 100644 gmp-6.3.0/mpn/x86/fat/mod_1_4.c create mode 100644 gmp-6.3.0/mpn/x86/fat/mode1o.c create mode 100644 gmp-6.3.0/mpn/x86/fat/mullo_basecase.c create mode 100644 gmp-6.3.0/mpn/x86/fat/redc_1.c create mode 100644 gmp-6.3.0/mpn/x86/fat/redc_2.c create mode 100644 gmp-6.3.0/mpn/x86/gcd_11.asm create mode 100644 gmp-6.3.0/mpn/x86/geode/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/goldmont/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/i486/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/k10/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/k6/README create mode 100644 gmp-6.3.0/mpn/x86/k6/aors_n.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/aorsmul_1.asm create mode 100755 gmp-6.3.0/mpn/x86/k6/cross.pl create mode 100644 gmp-6.3.0/mpn/x86/k6/divrem_1.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/k6/k62mmx/copyd.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/k62mmx/lshift.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/k62mmx/rshift.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/mmx/com.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/mmx/dive_1.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/mmx/logops_n.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/mmx/lshift.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/mmx/popham.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/mmx/rshift.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/mod_34lsub1.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/mode1o.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/mul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/mul_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/pre_mod_1.asm create mode 100644 gmp-6.3.0/mpn/x86/k6/sqr_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/README create mode 100644 gmp-6.3.0/mpn/x86/k7/addlsh1_n.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/aors_n.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/aorsmul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/bdiv_q_1.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/dive_1.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/gcd_11.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/k7/invert_limb.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mmx/com.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mmx/copyd.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mmx/copyi.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mmx/divrem_1.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mmx/lshift.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mmx/popham.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mmx/rshift.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mod_1_1.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mod_1_4.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mod_34lsub1.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mode1o.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/mul_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/sqr_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/k7/sublsh1_n.asm create mode 100644 gmp-6.3.0/mpn/x86/k8/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/lshift.asm create mode 100644 gmp-6.3.0/mpn/x86/mmx/sec_tabselect.asm create mode 100644 gmp-6.3.0/mpn/x86/mod_34lsub1.asm create mode 100644 gmp-6.3.0/mpn/x86/mul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/mul_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/nano/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/p6/README create mode 100644 gmp-6.3.0/mpn/x86/p6/aors_n.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/aorsmul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/bdiv_q_1.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/copyd.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/dive_1.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/gcd_11.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/p6/lshsub_n.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/mmx/divrem_1.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/mmx/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/p6/mmx/lshift.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/mmx/popham.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/mmx/rshift.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/mod_34lsub1.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/mode1o.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/mul_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/p3mmx/popham.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/sqr_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/sse2/addmul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/sse2/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/p6/sse2/mod_1_1.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/sse2/mod_1_4.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/sse2/mul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/sse2/mul_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/sse2/popcount.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/sse2/sqr_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/p6/sse2/submul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/README create mode 100644 gmp-6.3.0/mpn/x86/pentium/aors_n.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/aorsmul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/bdiv_q_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/com.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/copyd.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/copyi.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/dive_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/pentium/hamdist.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/logops_n.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/lshift.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/mmx/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/pentium/mmx/hamdist.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/mmx/lshift.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/mmx/mul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/mmx/rshift.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/mod_34lsub1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/mode1o.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/mul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/mul_2.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/mul_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/popcount.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/rshift.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium/sqr_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/README create mode 100644 gmp-6.3.0/mpn/x86/pentium4/copyd.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/copyi.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/mmx/lshift.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/mmx/popham.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/mmx/rshift.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/add_n.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/addlsh1_n.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/addmul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/bdiv_dbm1c.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/bdiv_q_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/cnd_add_n.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/cnd_sub_n.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/dive_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/divrem_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/mod_1_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/mod_1_4.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/mod_34lsub1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/mode1o.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/mul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/mul_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/popcount.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/rsh1add_n.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/sqr_basecase.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/sub_n.asm create mode 100644 gmp-6.3.0/mpn/x86/pentium4/sse2/submul_1.asm create mode 100644 gmp-6.3.0/mpn/x86/rshift.asm create mode 100644 gmp-6.3.0/mpn/x86/sec_tabselect.asm create mode 100644 gmp-6.3.0/mpn/x86/silvermont/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/skylake/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/sqr_basecase.asm create mode 100755 gmp-6.3.0/mpn/x86/t-zdisp.sh create mode 100755 gmp-6.3.0/mpn/x86/t-zdisp2.pl create mode 100644 gmp-6.3.0/mpn/x86/udiv.asm create mode 100644 gmp-6.3.0/mpn/x86/umul.asm create mode 100644 gmp-6.3.0/mpn/x86/x86-defs.m4 create mode 100644 gmp-6.3.0/mpn/x86/zn1/gmp-mparam.h create mode 100644 gmp-6.3.0/mpn/x86/zn2/gmp-mparam.h (limited to 'gmp-6.3.0/mpn/x86') diff --git a/gmp-6.3.0/mpn/x86/README b/gmp-6.3.0/mpn/x86/README new file mode 100644 index 0000000..8d7ac90 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/README @@ -0,0 +1,525 @@ +Copyright 1999-2002 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. + + + + + + X86 MPN SUBROUTINES + + +This directory contains mpn functions for various 80x86 chips. + + +CODE ORGANIZATION + + x86 i386, generic + x86/i486 i486 + x86/pentium Intel Pentium (P5, P54) + x86/pentium/mmx Intel Pentium with MMX (P55) + x86/p6 Intel Pentium Pro + x86/p6/mmx Intel Pentium II, III + x86/p6/p3mmx Intel Pentium III + x86/k6 \ AMD K6 + x86/k6/mmx / + x86/k6/k62mmx AMD K6-2 + x86/k7 \ AMD Athlon + x86/k7/mmx / + x86/pentium4 \ + x86/pentium4/mmx | Intel Pentium 4 + x86/pentium4/sse2 / + + +The top-level x86 directory contains blended style code, meant to be +reasonable on all x86s. + + + +STATUS + +The code is well-optimized for AMD and Intel chips, but there's nothing +specific for Cyrix chips, nor for actual 80386 and 80486 chips. + + + +ASM FILES + +The x86 .asm files are BSD style assembler code, first put through m4 for +macro processing. The generic mpn/asm-defs.m4 is used, together with +mpn/x86/x86-defs.m4. See comments in those files. + +The code is meant for use with GNU "gas" or a system "as". There's no +support for assemblers that demand Intel style code. + + + +STACK FRAME + +m4 macros are used to define the parameters passed on the stack, and these +act like comments on what the stack frame looks like too. For example, +mpn_mul_1() has the following. + + defframe(PARAM_MULTIPLIER, 16) + defframe(PARAM_SIZE, 12) + defframe(PARAM_SRC, 8) + defframe(PARAM_DST, 4) + +PARAM_MULTIPLIER becomes `FRAME+16(%esp)', and the others similarly. The +return address is at offset 0, but there's not normally any need to access +that. + +FRAME is redefined as necessary through the code so it's the number of bytes +pushed on the stack, and hence the offsets in the parameter macros stay +correct. At the start of a routine FRAME should be zero. + + deflit(`FRAME',0) + ... + deflit(`FRAME',4) + ... + deflit(`FRAME',8) + ... + +Helper macros FRAME_pushl(), FRAME_popl(), FRAME_addl_esp() and +FRAME_subl_esp() exist to adjust FRAME for the effect of those instructions, +and can be used instead of explicit definitions if preferred. +defframe_pushl() is a combination FRAME_pushl() and defframe(). + +There's generally some slackness in redefining FRAME. If new values aren't +going to get used then the redefinitions are omitted to keep from cluttering +up the code. This happens for instance at the end of a routine, where there +might be just four pops and then a ret, so FRAME isn't getting used. + +Local variables and saved registers can be similarly defined, with negative +offsets representing stack space below the initial stack pointer. For +example, + + defframe(SAVE_ESI, -4) + defframe(SAVE_EDI, -8) + defframe(VAR_COUNTER,-12) + + deflit(STACK_SPACE, 12) + +Here STACK_SPACE gets used in a "subl $STACK_SPACE, %esp" to allocate the +space, and that instruction must be followed by a redefinition of FRAME +(setting it equal to STACK_SPACE) to reflect the change in %esp. + +Definitions for pushed registers are only put in when they're going to be +used. If registers are just saved and restored with pushes and pops then +definitions aren't made. + + + +ASSEMBLER EXPRESSIONS + +Only addition and subtraction seem to be universally available, certainly +that's all the Solaris 8 "as" seems to accept. If expressions are wanted +then m4 eval() should be used. + +In particular note that a "/" anywhere in a line starts a comment in Solaris +"as", and in some configurations of gas too. + + addl $32/2, %eax <-- wrong + + addl $eval(32/2), %eax <-- right + +Binutils gas/config/tc-i386.c has a choice between "/" being a comment +anywhere in a line, or only at the start. FreeBSD patches 2.9.1 to select +the latter, and from 2.9.5 it's the default for GNU/Linux too. + + + +ASSEMBLER COMMENTS + +Solaris "as" doesn't support "#" commenting, using /* */ instead. For that +reason "C" commenting is used (see asm-defs.m4) and the intermediate ".s" +files have no comments. + +Any comments before include(`../config.m4') must use m4 "dnl", since it's +only after the include that "C" is available. By convention "dnl" is also +used for comments about m4 macros. + + + +TEMPORARY LABELS + +Temporary numbered labels like "1:" used as "1f" or "1b" are available in +"gas" and Solaris "as", but not in SCO "as". Normal L() labels should be +used instead, possibly with a counter to make them unique, see jadcl0() in +x86-defs.m4 for instance. A separate counter for each macro makes it +possible to nest them, for instance movl_text_address() can be used within +an ASSERT(). + +"1:" etc must be avoided in gcc __asm__ blocks too. "%=" for generating a +unique number looks like a good alternative, but is that actually a +documented feature? In any case this problem doesn't currently arise. + + + +ZERO DISPLACEMENTS + +In a couple of places addressing modes like 0(%ebx) with a byte-sized zero +displacement are wanted, rather than (%ebx) with no displacement. These are +either for computed jumps or to get desirable code alignment. Explicit +.byte sequences are used to ensure the assembler doesn't turn 0(%ebx) into +(%ebx). The Zdisp() macro in x86-defs.m4 is used for this. + +Current gas 2.9.5 or recent 2.9.1 leave 0(%ebx) as written, but old gas +1.92.3 changes it. In general changing would be the sort of "optimization" +an assembler might perform, hence explicit ".byte"s are used where +necessary. + + + +SHLD/SHRD INSTRUCTIONS + +The %cl count forms of double shift instructions like "shldl %cl,%eax,%ebx" +must be written "shldl %eax,%ebx" for some assemblers. gas takes either, +Solaris "as" doesn't allow %cl, gcc generates %cl for gas and NeXT (which is +gas), and omits %cl elsewhere. + +For GMP an autoconf test GMP_ASM_X86_SHLDL_CL is used to determine whether +%cl should be used, and the macros shldl, shrdl, shldw and shrdw in +mpn/x86/x86-defs.m4 pass through or omit %cl as necessary. See the comments +with those macros for usage. + + + +IMUL INSTRUCTION + +GCC config/i386/i386.md (cvs rev 1.187, 21 Oct 00) under *mulsi3_1 notes +that the following two forms produce identical object code + + imul $12, %eax + imul $12, %eax, %eax + +but that the former isn't accepted by some assemblers, in particular the SCO +OSR5 COFF assembler. GMP follows GCC and uses only the latter form. + +(This applies only to immediate operands, the three operand form is only +valid with an immediate.) + + + +DIRECTION FLAG + +The x86 calling conventions say that the direction flag should be clear at +function entry and exit. (See iBCS2 and SVR4 ABI books, references below.) +Although this has been so since the year dot, it's not absolutely clear +whether it's universally respected. Since it's better to be safe than +sorry, GMP follows glibc and does a "cld" if it depends on the direction +flag being clear. This happens only in a few places. + + + +POSITION INDEPENDENT CODE + + Coding Style + + Defining the symbol PIC in m4 processing selects SVR4 / ELF style + position independent code. This is necessary for shared libraries + because they can be mapped into different processes at different virtual + addresses. Actually, relocations are allowed but text pages with + relocations aren't shared, defeating the purpose of a shared library. + + The GOT is used to access global data, and the PLT is used for + functions. The use of the PLT adds a fixed cost to every function call, + and the GOT adds a cost to any function accessing global variables. + These are small but might be noticeable when working with small + operands. + + Scope + + It's intended, as a matter of policy, that references within libgmp are + resolved within libgmp. Certainly there's no need for an application to + replace any internals, and we take the view that there's no value in an + application subverting anything documented either. + + Resolving references within libgmp in theory means calls can be made with a + plain PC-relative call instruction, which is faster and smaller than going + through the PLT, and data references can be similarly PC-relative, saving a + GOT entry and fetch from there. Unfortunately the normal linker behaviour + doesn't allow us to do this. + + By default an R_386_PC32 PC-relative reference, either for a call or for + data, is left in libgmp.so by the linker so that it can be resolved at + runtime to a location in the application or another shared library. This + means a text segment relocation which we don't want. + + -Bsymbolic + + Under the "-Bsymbolic" option, the linker resolves references to symbols + within libgmp.so. This gives us the desired effect for R_386_PC32, + ie. it's resolved at link time. It also resolves R_386_PLT32 calls + directly to their target without creating a PLT entry (though if this is + done to normal compiler-generated code it still leaves a setup of %ebx + to _GLOBAL_OFFSET_TABLE_ which may then be unnecessary). + + Unfortunately -Bsymbolic does bad things to global variables defined in + a shared library but accessed by non-PIC code from the mainline (or a + static library). + + The problem is that the mainline needs a fixed data address to avoid + text segment relocations, so space is allocated in its data segment and + the value from the variable is copied from the shared library's data + segment when the library is loaded. Under -Bsymbolic, however, + references in the shared library are then resolved still to the shared + library data area. Not surprisingly it bombs badly to have mainline + code and library code accessing different locations for what should be + one variable. + + Note that this -Bsymbolic effect for the shared library is not just for + R_386_PC32 offsets which might have been cooked up in assembler, but is + done also for the contents of GOT entries. -Bsymbolic simply applies a + general rule that symbols are resolved first from the local module. + + Visibility Attributes + + GCC __attribute__ ((visibility ("protected"))), which is available in + recent versions, eg. 3.3, is probably what we'd like to use. It makes + gcc generate plain PC-relative calls to indicated functions, and directs + the linker to resolve references to the given function within the link + module. + + Unfortunately, as of debian binutils 2.13.90.0.16 at least, the + resulting libgmp.so comes out with text segment relocations, references + are not resolved at link time. If the gcc description is to be believed + this is this not how it should work. If a symbol cannot be overridden + by another module then surely references within that module can be + resolved immediately (ie. at link time). + + Present + + In any case, all this means that we have no optimizations we can + usefully make to function or variable usages, neither for assembler nor + C code. Perhaps in the future the visibility attribute will work as + we'd like. + + + + +GLOBAL OFFSET TABLE + +The magic _GLOBAL_OFFSET_TABLE_ used by code establishing the address of the +GOT sometimes requires an extra underscore prefix. SVR4 systems and NetBSD +don't need a prefix, OpenBSD does need one. Note that NetBSD and OpenBSD +are both a.out underscore systems, so the prefix for _GLOBAL_OFFSET_TABLE_ +is not simply the same as the prefix for ordinary globals. + +In any case in the asm code we write _GLOBAL_OFFSET_TABLE_ and let a macro +in x86-defs.m4 add an extra underscore if required (according to a configure +test). + +Old gas 1.92.3 which comes with FreeBSD 2.2.8 gets a segmentation fault when +asked to assemble the following, + + L1: + addl $_GLOBAL_OFFSET_TABLE_+[.-L1], %ebx + +It seems that using the label in the same instruction it refers to is the +problem, since a nop in between works. But the simplest workaround is to +follow gcc and omit the +[.-L1] since it does nothing, + + addl $_GLOBAL_OFFSET_TABLE_, %ebx + +Current gas 2.10 generates incorrect object code when %eax is used in such a +construction (with or without +[.-L1]), + + addl $_GLOBAL_OFFSET_TABLE_, %eax + +The R_386_GOTPC gets a displacement of 2 rather than the 1 appropriate for +the 1 byte opcode of "addl $n,%eax". The best workaround is just to use any +other register, since then it's a two byte opcode+mod/rm. GCC for example +always uses %ebx (which is needed for calls through the PLT). + +A similar problem occurs in an leal (again with or without a +[.-L1]), + + leal _GLOBAL_OFFSET_TABLE_(%edi), %ebx + +This time the R_386_GOTPC gets a displacement of 0 rather than the 2 +appropriate for the opcode and mod/rm, making this form unusable. + + + + +SIMPLE LOOPS + +The overheads in setting up for an unrolled loop can mean that at small +sizes a simple loop is faster. Making small sizes go fast is important, +even if it adds a cycle or two to bigger sizes. To this end various +routines choose between a simple loop and an unrolled loop according to +operand size. The path to the simple loop, or to special case code for +small sizes, is always as fast as possible. + +Adding a simple loop requires a conditional jump to choose between the +simple and unrolled code. The size of a branch misprediction penalty +affects whether a simple loop is worthwhile. + +The convention is for an m4 definition UNROLL_THRESHOLD to set the crossover +point, with sizes < UNROLL_THRESHOLD using the simple loop, sizes >= +UNROLL_THRESHOLD using the unrolled loop. If position independent code adds +a couple of cycles to an unrolled loop setup, the threshold will vary with +PIC or non-PIC. Something like the following is typical. + + deflit(UNROLL_THRESHOLD, ifdef(`PIC',10,8)) + +There's no automated way to determine the threshold. Setting it to a small +value and then to a big value makes it possible to measure the simple and +unrolled loops each over a range of sizes, from which the crossover point +can be determined. Alternately, just adjust the threshold up or down until +there's no more speedups. + + + +UNROLLED LOOP CODING + +The x86 addressing modes allow a byte displacement of -128 to +127, making +it possible to access 256 bytes, which is 64 limbs, without adjusting +pointer registers within the loop. Dword sized displacements can be used +too, but they increase code size, and unrolling to 64 ought to be enough. + +When unrolling to the full 64 limbs/loop, the limb at the top of the loop +will have a displacement of -128, so pointers have to have a corresponding ++128 added before entering the loop. When unrolling to 32 limbs/loop +displacements 0 to 127 can be used with 0 at the top of the loop and no +adjustment needed to the pointers. + +Where 64 limbs/loop is supported, the +128 adjustment is done only when 64 +limbs/loop is selected. Usually the gain in speed using 64 instead of 32 or +16 is small, so support for 64 limbs/loop is generally only for comparison. + + + +COMPUTED JUMPS + +When working from least significant limb to most significant limb (most +routines) the computed jump and pointer calculations in preparation for an +unrolled loop are as follows. + + S = operand size in limbs + N = number of limbs per loop (UNROLL_COUNT) + L = log2 of unrolling (UNROLL_LOG2) + M = mask for unrolling (UNROLL_MASK) + C = code bytes per limb in the loop + B = bytes per limb (4 for x86) + + computed jump (-S & M) * C + entrypoint + subtract from pointers (-S & M) * B + initial loop counter (S-1) >> L + displacements 0 to B*(N-1) + +The loop counter is decremented at the end of each loop, and the looping +stops when the decrement takes the counter to -1. The displacements are for +the addressing accessing each limb, eg. a load with "movl disp(%ebx), %eax". + +Usually the multiply by "C" can be handled without an imul, using instead an +leal, or a shift and subtract. + +When working from most significant to least significant limb (eg. mpn_lshift +and mpn_copyd), the calculations change as follows. + + add to pointers (-S & M) * B + displacements 0 to -B*(N-1) + + + +OLD GAS 1.92.3 + +This version comes with FreeBSD 2.2.8 and has a couple of gremlins that +affect GMP code. + +Firstly, an expression involving two forward references to labels comes out +as zero. For example, + + addl $bar-foo, %eax + foo: + nop + bar: + +This should lead to "addl $1, %eax", but it comes out as "addl $0, %eax". +When only one forward reference is involved, it works correctly, as for +example, + + foo: + addl $bar-foo, %eax + nop + bar: + +Secondly, an expression involving two labels can't be used as the +displacement for an leal. For example, + + foo: + nop + bar: + leal bar-foo(%eax,%ebx,8), %ecx + +A slightly cryptic error is given, "Unimplemented segment type 0 in +parse_operand". When only one label is used it's ok, and the label can be a +forward reference too, as for example, + + leal foo(%eax,%ebx,8), %ecx + nop + foo: + +These problems only affect PIC computed jump calculations. The workarounds +are just to do an leal without a displacement and then an addl, and to make +sure the code is placed so that there's at most one forward reference in the +addl. + + + +REFERENCES + +"Intel Architecture Software Developer's Manual", volumes 1, 2a, 2b, 3a, 3b, +2006, order numbers 253665 through 253669. Available on-line, + + ftp://download.intel.com/design/Pentium4/manuals/25366518.pdf + ftp://download.intel.com/design/Pentium4/manuals/25366618.pdf + ftp://download.intel.com/design/Pentium4/manuals/25366718.pdf + ftp://download.intel.com/design/Pentium4/manuals/25366818.pdf + ftp://download.intel.com/design/Pentium4/manuals/25366918.pdf + + +"System V Application Binary Interface", Unix System Laboratories Inc, 1992, +published by Prentice Hall, ISBN 0-13-880410-9. And the "Intel386 Processor +Supplement", AT&T, 1991, ISBN 0-13-877689-X. These have details of calling +conventions and ELF shared library PIC coding. Versions of both available +on-line, + + http://www.sco.com/developer/devspecs + +"Intel386 Family Binary Compatibility Specification 2", Intel Corporation, +published by McGraw-Hill, 1991, ISBN 0-07-031219-2. (Same as the above 386 +ABI supplement.) + + + +---------------- +Local variables: +mode: text +fill-column: 76 +End: diff --git a/gmp-6.3.0/mpn/x86/aors_n.asm b/gmp-6.3.0/mpn/x86/aors_n.asm new file mode 100644 index 0000000..5d359f5 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/aors_n.asm @@ -0,0 +1,202 @@ +dnl x86 mpn_add_n/mpn_sub_n -- mpn addition and subtraction. + +dnl Copyright 1992, 1994-1996, 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P5 3.375 +C P6 3.125 +C K6 3.5 +C K7 2.25 +C P4 8.75 + + +ifdef(`OPERATION_add_n',` + define(M4_inst, adcl) + define(M4_function_n, mpn_add_n) + define(M4_function_nc, mpn_add_nc) + +',`ifdef(`OPERATION_sub_n',` + define(M4_inst, sbbl) + define(M4_function_n, mpn_sub_n) + define(M4_function_nc, mpn_sub_nc) + +',`m4_error(`Need OPERATION_add_n or OPERATION_sub_n +')')') + +MULFUNC_PROLOGUE(mpn_add_n mpn_add_nc mpn_sub_n mpn_sub_nc) + + +C mp_limb_t M4_function_n (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size); +C mp_limb_t M4_function_nc (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size, mp_limb_t carry); + +defframe(PARAM_CARRY,20) +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC2, 12) +defframe(PARAM_SRC1, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) + +PROLOGUE(M4_function_nc) +deflit(`FRAME',0) + + pushl %edi FRAME_pushl() + pushl %esi FRAME_pushl() + + movl PARAM_DST,%edi + movl PARAM_SRC1,%esi + movl PARAM_SRC2,%edx + movl PARAM_SIZE,%ecx + + movl %ecx,%eax + shrl $3,%ecx C compute count for unrolled loop + negl %eax + andl $7,%eax C get index where to start loop + jz L(oopgo) C necessary special case for 0 + incl %ecx C adjust loop count + shll $2,%eax C adjustment for pointers... + subl %eax,%edi C ... since they are offset ... + subl %eax,%esi C ... by a constant when we ... + subl %eax,%edx C ... enter the loop + shrl $2,%eax C restore previous value + +ifdef(`PIC',` + C Calculate start address in loop for PIC. Due to limitations in + C old gas, LF(M4_function_n,oop)-L(0a)-3 cannot be put into the leal + call L(0a) +L(0a): leal (%eax,%eax,8),%eax + addl (%esp),%eax + addl $L(oop)-L(0a)-3,%eax + addl $4,%esp +',` + C Calculate start address in loop for non-PIC. + leal L(oop)-3(%eax,%eax,8),%eax +') + + C These lines initialize carry from the 5th parameter. Should be + C possible to simplify. + pushl %ebp FRAME_pushl() + movl PARAM_CARRY,%ebp + shrl %ebp C shift bit 0 into carry + popl %ebp FRAME_popl() + + jmp *%eax C jump into loop + +EPILOGUE() + + + ALIGN(16) +PROLOGUE(M4_function_n) +deflit(`FRAME',0) + + pushl %edi FRAME_pushl() + pushl %esi FRAME_pushl() + + movl PARAM_DST,%edi + movl PARAM_SRC1,%esi + movl PARAM_SRC2,%edx + movl PARAM_SIZE,%ecx + + movl %ecx,%eax + shrl $3,%ecx C compute count for unrolled loop + negl %eax + andl $7,%eax C get index where to start loop + jz L(oop) C necessary special case for 0 + incl %ecx C adjust loop count + shll $2,%eax C adjustment for pointers... + subl %eax,%edi C ... since they are offset ... + subl %eax,%esi C ... by a constant when we ... + subl %eax,%edx C ... enter the loop + shrl $2,%eax C restore previous value + +ifdef(`PIC',` + C Calculate start address in loop for PIC. Due to limitations in + C some assemblers, L(oop)-L(0b)-3 cannot be put into the leal + call L(0b) +L(0b): leal (%eax,%eax,8),%eax + addl (%esp),%eax + addl $L(oop)-L(0b)-3,%eax + addl $4,%esp +',` + C Calculate start address in loop for non-PIC. + leal L(oop)-3(%eax,%eax,8),%eax +') + jmp *%eax C jump into loop + +L(oopgo): + pushl %ebp FRAME_pushl() + movl PARAM_CARRY,%ebp + shrl %ebp C shift bit 0 into carry + popl %ebp FRAME_popl() + + ALIGN(16) +L(oop): movl (%esi),%eax + M4_inst (%edx),%eax + movl %eax,(%edi) + movl 4(%esi),%eax + M4_inst 4(%edx),%eax + movl %eax,4(%edi) + movl 8(%esi),%eax + M4_inst 8(%edx),%eax + movl %eax,8(%edi) + movl 12(%esi),%eax + M4_inst 12(%edx),%eax + movl %eax,12(%edi) + movl 16(%esi),%eax + M4_inst 16(%edx),%eax + movl %eax,16(%edi) + movl 20(%esi),%eax + M4_inst 20(%edx),%eax + movl %eax,20(%edi) + movl 24(%esi),%eax + M4_inst 24(%edx),%eax + movl %eax,24(%edi) + movl 28(%esi),%eax + M4_inst 28(%edx),%eax + movl %eax,28(%edi) + leal 32(%edi),%edi + leal 32(%esi),%esi + leal 32(%edx),%edx + decl %ecx + jnz L(oop) + + sbbl %eax,%eax + negl %eax + + popl %esi + popl %edi + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/aorsmul_1.asm b/gmp-6.3.0/mpn/x86/aorsmul_1.asm new file mode 100644 index 0000000..8c13997 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/aorsmul_1.asm @@ -0,0 +1,214 @@ +dnl x86 __gmpn_addmul_1 (for 386 and 486) -- Multiply a limb vector with a +dnl limb and add the result to a second limb vector. + +dnl Copyright 1992, 1994, 1997, 1999-2002, 2005 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom +C AMD K6 +C AMD K7 +C AMD K8 3.875 +C AMD K10 + + +ifdef(`OPERATION_addmul_1',` + define(ADDSUB, addl) + define(M4_function_1, mpn_addmul_1) + define(M4_function_1c, mpn_addmul_1c) + +',`ifdef(`OPERATION_submul_1',` + define(ADDSUB, subl) + define(M4_function_1, mpn_submul_1) + +',`m4_error(`Need OPERATION_addmul_1 or OPERATION_submul_1 +')')') + +MULFUNC_PROLOGUE(mpn_addmul_1 mpn_submul_1 mpn_addmul_1c) + + +C mp_limb_t M4_function_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult); + +define(PARAM_CARRY, `FRAME+20(%esp)') +define(PARAM_MULTIPLIER, `FRAME+16(%esp)') +define(PARAM_SIZE, `FRAME+12(%esp)') +define(PARAM_SRC, `FRAME+8(%esp)') +define(PARAM_DST, `FRAME+4(%esp)') + + TEXT + ALIGN(32) +PROLOGUE(M4_function_1) +deflit(`FRAME',0) + + pushl %edi + pushl %esi + pushl %ebx + pushl %ebp +deflit(`FRAME',16) + + movl PARAM_SRC, %esi + movl PARAM_SIZE, %ecx + movl PARAM_DST, %edi + + movl (%esi), %eax + mull PARAM_MULTIPLIER + + testb $1, %cl + jnz L(bx1) + +L(bx0): movl %eax, %ebx + movl %edx, %ebp + shrl $2, %ecx + jnc L(lo0) + +L(b10): leal -8(%esi), %esi + leal -8(%edi), %edi + incl %ecx + jmp L(lo2) + +L(bx1): movl %eax, %ebp + movl %edx, %ebx + shrl $2, %ecx + jc L(b11) + +L(b01): leal 4(%edi), %edi + jz L(end) + leal 4(%esi), %esi + jmp L(top) + +L(b11): leal -4(%esi), %esi + leal -4(%edi), %edi + incl %ecx + jmp L(lo3) + + ALIGN(16) +L(top): movl (%esi), %eax + mull PARAM_MULTIPLIER + ADDSUB %ebp, -4(%edi) + adcl %eax, %ebx + movl $0, %ebp + adcl %edx, %ebp +L(lo0): movl 4(%esi), %eax + mull PARAM_MULTIPLIER + ADDSUB %ebx, (%edi) + adcl %eax, %ebp + movl $0, %ebx + adcl %edx, %ebx +L(lo3): movl 8(%esi), %eax + mull PARAM_MULTIPLIER + ADDSUB %ebp, 4(%edi) + adcl %eax, %ebx + movl $0, %ebp + adcl %edx, %ebp +L(lo2): movl 12(%esi), %eax + mull PARAM_MULTIPLIER + ADDSUB %ebx, 8(%edi) + adcl %eax, %ebp + movl $0, %ebx + adcl %edx, %ebx + + leal 16(%esi), %esi + leal 16(%edi), %edi + decl %ecx + jnz L(top) + +L(end): xor %eax, %eax + ADDSUB %ebp, -4(%edi) + adcl %ebx, %eax + popl %ebp + popl %ebx + popl %esi + popl %edi + ret + +EPILOGUE() +ifdef(`OPERATION_addmul_1',` + ALIGN(32) +PROLOGUE(M4_function_1c) +deflit(`FRAME',0) + + pushl %edi + pushl %esi + pushl %ebx + pushl %ebp +deflit(`FRAME',16) + + movl PARAM_SRC, %esi + movl PARAM_SIZE, %ecx + movl PARAM_DST, %edi + + movl (%esi), %eax + mull PARAM_MULTIPLIER + + testb $1, %cl + jnz L(cx1) + + movl PARAM_CARRY, %ebx + xorl %ebp, %ebp + +L(cx0): addl %eax, %ebx + adcl %edx, %ebp + shrl $2, %ecx + jnc L(lo0) + +L(c10): leal -8(%esi), %esi + leal -8(%edi), %edi + incl %ecx + jmp L(lo2) + +L(cx1): movl PARAM_CARRY, %ebp + xorl %ebx, %ebx + + addl %eax, %ebp + adcl %edx, %ebx + shrl $2, %ecx + jc L(c11) + +L(c01): leal 4(%edi), %edi + jz L(end) + leal 4(%esi), %esi + jmp L(top) + +L(c11): leal -4(%esi), %esi + leal -4(%edi), %edi + incl %ecx + jmp L(lo3) +EPILOGUE() +') diff --git a/gmp-6.3.0/mpn/x86/atom/aorrlsh1_n.asm b/gmp-6.3.0/mpn/x86/atom/aorrlsh1_n.asm new file mode 100644 index 0000000..cd1a650 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/aorrlsh1_n.asm @@ -0,0 +1,53 @@ +dnl Intel Atom mpn_rsblsh1_n -- rp[] = (vp[] << 1) - up[] + +dnl Contributed to the GNU project by Marco Bodrato. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +define(LSH, 1) +define(RSH, 31) + +ifdef(`OPERATION_addlsh1_n', ` + define(M4_inst, adc) + define(M4_opp, sub) + define(M4_function, mpn_addlsh1_n) + define(M4_function_c, mpn_addlsh1_nc) +',`ifdef(`OPERATION_rsblsh1_n', ` + define(M4_inst, sbb) + define(M4_opp, add) + define(M4_function, mpn_rsblsh1_n) + define(M4_function_c, mpn_rsblsh1_nc) +',`m4_error(`Need OPERATION_addlsh1_n or OPERATION_rsblsh1_n +')')') + +MULFUNC_PROLOGUE(mpn_addlsh1_n mpn_addlsh1_nc mpn_rsblsh1_n mpn_rsblsh1_nc) + +include_mpn(`x86/atom/aorrlshC_n.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/aorrlsh2_n.asm b/gmp-6.3.0/mpn/x86/atom/aorrlsh2_n.asm new file mode 100644 index 0000000..10f4419 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/aorrlsh2_n.asm @@ -0,0 +1,53 @@ +dnl Intel Atom mpn_addlsh2_n/mpn_rsblsh2_n -- rp[] = (vp[] << 2) +- up[] + +dnl Contributed to the GNU project by Marco Bodrato. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +define(LSH, 2) +define(RSH, 30) + +ifdef(`OPERATION_addlsh2_n', ` + define(M4_inst, adcl) + define(M4_opp, subl) + define(M4_function, mpn_addlsh2_n) + define(M4_function_c, mpn_addlsh2_nc) +',`ifdef(`OPERATION_rsblsh2_n', ` + define(M4_inst, sbbl) + define(M4_opp, addl) + define(M4_function, mpn_rsblsh2_n) + define(M4_function_c, mpn_rsblsh2_nc) +',`m4_error(`Need OPERATION_addlsh2_n or OPERATION_rsblsh2_n +')')') + +MULFUNC_PROLOGUE(mpn_addlsh2_n mpn_addlsh2_nc mpn_rsblsh2_n mpn_rsblsh2_nc) + +include_mpn(`x86/atom/aorrlshC_n.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/aorrlshC_n.asm b/gmp-6.3.0/mpn/x86/atom/aorrlshC_n.asm new file mode 100644 index 0000000..71cfe49 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/aorrlshC_n.asm @@ -0,0 +1,156 @@ +dnl Intel Atom mpn_addlshC_n/mpn_rsblshC_n -- rp[] = (vp[] << C) +- up[] + +dnl Contributed to the GNU project by Marco Bodrato. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C mp_limb_t mpn_addlshC_n (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size); +C mp_limb_t mpn_addlshC_nc (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size, mp_limb_t carry); +C mp_limb_t mpn_rsblshC_n (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size); +C mp_limb_t mpn_rsblshC_nc (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size, mp_signed_limb_t carry); + +C cycles/limb +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 6 +C AMD K6 +C AMD K7 +C AMD K8 +C AMD K10 + +defframe(PARAM_CORB, 20) +defframe(PARAM_SIZE, 16) +defframe(PARAM_DBLD, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(VAR_COUNT,`PARAM_SIZE') +define(SAVE_EBP,`PARAM_DBLD') +define(SAVE_VP,`PARAM_SRC') +define(SAVE_UP,`PARAM_DST') + +define(M, eval(m4_lshift(1,LSH))) +define(`rp', `%edi') +define(`up', `%esi') +define(`vp', `%ebx') + +ASM_START() + TEXT + ALIGN(8) + +PROLOGUE(M4_function_c) +deflit(`FRAME',0) + movl PARAM_CORB, %eax + movl %eax, %edx + shr $LSH, %edx + andl $1, %edx + M4_opp %edx, %eax + jmp L(start_nc) +EPILOGUE() + +PROLOGUE(M4_function) +deflit(`FRAME',0) + + xor %eax, %eax + xor %edx, %edx +L(start_nc): + push rp FRAME_pushl() + + mov PARAM_SIZE, %ecx C size + mov PARAM_DST, rp + mov up, SAVE_UP + incl %ecx C size + 1 + mov PARAM_SRC, up + mov vp, SAVE_VP + shr %ecx C (size+1)\2 + mov PARAM_DBLD, vp + mov %ebp, SAVE_EBP + mov %ecx, VAR_COUNT + jnc L(entry) C size odd + + shr %edx C size even + mov (vp), %ecx + lea 4(vp), vp + lea (%eax,%ecx,M), %edx + mov %ecx, %eax + lea -4(up), up + lea -4(rp), rp + jmp L(enteven) + + ALIGN(16) +L(oop): + lea (%eax,%ecx,M), %ebp + shr $RSH, %ecx + mov 4(vp), %eax + shr %edx + lea 8(vp), vp + M4_inst (up), %ebp + lea (%ecx,%eax,M), %edx + mov %ebp, (rp) +L(enteven): + M4_inst 4(up), %edx + lea 8(up), up + mov %edx, 4(rp) + adc %edx, %edx + shr $RSH, %eax + lea 8(rp), rp +L(entry): + mov (vp), %ecx + decl VAR_COUNT + jnz L(oop) + + lea (%eax,%ecx,M), %ebp + shr $RSH, %ecx + shr %edx + mov SAVE_VP, vp + M4_inst (up), %ebp + mov %ecx, %eax + mov SAVE_UP, up + M4_inst $0, %eax + mov %ebp, (rp) + mov SAVE_EBP, %ebp + pop rp FRAME_popl() + ret +EPILOGUE() + +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/atom/aors_n.asm b/gmp-6.3.0/mpn/x86/atom/aors_n.asm new file mode 100644 index 0000000..45ec287 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/aors_n.asm @@ -0,0 +1,159 @@ +dnl Intel Atom mpn_add_n/mpn_sub_n -- rp[] = up[] +- vp[]. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl Contributed to the GNU project by Marco Bodrato. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 3 +C AMD K6 +C AMD K7 +C AMD K8 +C AMD K10 + +ifdef(`OPERATION_add_n', ` + define(M4_inst, adcl) + define(M4_function_n, mpn_add_n) + define(M4_function_nc, mpn_add_nc) + define(M4_description, add) +',`ifdef(`OPERATION_sub_n', ` + define(M4_inst, sbbl) + define(M4_function_n, mpn_sub_n) + define(M4_function_nc, mpn_sub_nc) + define(M4_description, subtract) +',`m4_error(`Need OPERATION_add_n or OPERATION_sub_n +')')') + +MULFUNC_PROLOGUE(mpn_add_n mpn_add_nc mpn_sub_n mpn_sub_nc) + +C mp_limb_t M4_function_n (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size); +C mp_limb_t M4_function_nc (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size, mp_limb_t carry); +C +C Calculate src1,size M4_description src2,size, and store the result in +C dst,size. The return value is the carry bit from the top of the result (1 +C or 0). +C +C The _nc version accepts 1 or 0 for an initial carry into the low limb of +C the calculation. Note values other than 1 or 0 here will lead to garbage +C results. + +defframe(PARAM_CARRY,20) +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC2, 12) +defframe(PARAM_SRC1, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(SAVE_RP,`PARAM_SIZE') +define(SAVE_VP,`PARAM_SRC1') +define(SAVE_UP,`PARAM_DST') + +define(`rp', `%edi') +define(`up', `%esi') +define(`vp', `%ebx') +define(`cy', `%ecx') +define(`r1', `%ecx') +define(`r2', `%edx') + +ASM_START() + TEXT + ALIGN(16) +deflit(`FRAME',0) + +PROLOGUE(M4_function_n) + xor cy, cy C carry +L(start): + mov PARAM_SIZE, %eax C size + mov rp, SAVE_RP + mov PARAM_DST, rp + mov up, SAVE_UP + mov PARAM_SRC1, up + shr %eax C size >> 1 + mov vp, SAVE_VP + mov PARAM_SRC2, vp + jz L(one) C size == 1 + jc L(three) C size % 2 == 1 + + shr cy + mov (up), r2 + lea 4(up), up + lea 4(vp), vp + lea -4(rp), rp + jmp L(entry) +L(one): + shr cy + mov (up), r1 + jmp L(end) +L(three): + shr cy + mov (up), r1 + + ALIGN(16) +L(oop): + M4_inst (vp), r1 + lea 8(up), up + mov -4(up), r2 + lea 8(vp), vp + mov r1, (rp) +L(entry): + M4_inst -4(vp), r2 + lea 8(rp), rp + dec %eax + mov (up), r1 + mov r2, -4(rp) + jnz L(oop) + +L(end): C %eax is zero here + mov SAVE_UP, up + M4_inst (vp), r1 + mov SAVE_VP, vp + mov r1, (rp) + adc %eax, %eax + mov SAVE_RP, rp + ret +EPILOGUE() + +PROLOGUE(M4_function_nc) + mov PARAM_CARRY, cy C carry + jmp L(start) +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/atom/aorslshC_n.asm b/gmp-6.3.0/mpn/x86/atom/aorslshC_n.asm new file mode 100644 index 0000000..75ace65 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/aorslshC_n.asm @@ -0,0 +1,247 @@ +dnl Intel Atom mpn_addlshC_n/mpn_sublshC_n -- rp[] = up[] +- (vp[] << C) + +dnl Contributed to the GNU project by Marco Bodrato. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C mp_limb_t mpn_addlshC_n_ip1 (mp_ptr dst, mp_srcptr src, mp_size_t size); +C mp_limb_t mpn_addlshC_nc_ip1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t carry); +C mp_limb_t mpn_sublshC_n_ip1 (mp_ptr dst, mp_srcptr src, mp_size_t size,); +C mp_limb_t mpn_sublshC_nc_ip1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_signed_limb_t borrow); + +defframe(PARAM_CORB, 16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +C mp_limb_t mpn_addlshC_n (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size,); +C mp_limb_t mpn_addlshC_nc (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size, mp_limb_t carry); +C mp_limb_t mpn_sublshC_n (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size,); +C mp_limb_t mpn_sublshC_nc (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size, mp_limb_t borrow); + +C if src1 == dst, _ip1 is used + +C cycles/limb +C dst!=src1,src2 dst==src1 +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 7 6 +C AMD K6 +C AMD K7 +C AMD K8 +C AMD K10 + +defframe(GPARAM_CORB, 20) +defframe(GPARAM_SIZE, 16) +defframe(GPARAM_SRC2, 12) + +dnl re-use parameter space +define(SAVE_EBP,`PARAM_SIZE') +define(SAVE_EBX,`PARAM_SRC') +define(SAVE_UP,`PARAM_DST') + +define(M, eval(m4_lshift(1,LSH))) +define(`rp', `%edi') +define(`up', `%esi') + +ASM_START() + TEXT + ALIGN(8) + +PROLOGUE(M4_ip_function_c) +deflit(`FRAME',0) + movl PARAM_CORB, %ecx + movl %ecx, %edx + shr $LSH, %edx + andl $1, %edx + M4_opp %edx, %ecx + jmp L(start_nc) +EPILOGUE() + +PROLOGUE(M4_ip_function) +deflit(`FRAME',0) + + xor %ecx, %ecx + xor %edx, %edx +L(start_nc): + push rp FRAME_pushl() + mov PARAM_DST, rp + mov up, SAVE_UP + mov PARAM_SRC, up + mov %ebx, SAVE_EBX + mov PARAM_SIZE, %ebx C size +L(inplace): + incl %ebx C size + 1 + shr %ebx C (size+1)\2 + mov %ebp, SAVE_EBP + jnc L(entry) C size odd + + add %edx, %edx C size even + mov %ecx, %ebp + mov (up), %ecx + lea -4(rp), rp + lea (%ebp,%ecx,M), %eax + lea 4(up), up + jmp L(enteven) + + ALIGN(16) +L(oop): + lea (%ecx,%eax,M), %ebp + shr $RSH, %eax + mov 4(up), %ecx + add %edx, %edx + lea 8(up), up + M4_inst %ebp, (rp) + lea (%eax,%ecx,M), %eax + +L(enteven): + M4_inst %eax, 4(rp) + lea 8(rp), rp + + sbb %edx, %edx + shr $RSH, %ecx + +L(entry): + mov (up), %eax + decl %ebx + jnz L(oop) + + lea (%ecx,%eax,M), %ebp + shr $RSH, %eax + shr %edx + M4_inst %ebp, (rp) + mov SAVE_UP, up + adc $0, %eax + mov SAVE_EBP, %ebp + mov SAVE_EBX, %ebx + pop rp FRAME_popl() + ret +EPILOGUE() + +PROLOGUE(M4_function_c) +deflit(`FRAME',0) + movl GPARAM_CORB, %ecx + movl %ecx, %edx + shr $LSH, %edx + andl $1, %edx + M4_opp %edx, %ecx + jmp L(generic_nc) +EPILOGUE() + +PROLOGUE(M4_function) +deflit(`FRAME',0) + + xor %ecx, %ecx + xor %edx, %edx +L(generic_nc): + push rp FRAME_pushl() + mov PARAM_DST, rp + mov up, SAVE_UP + mov PARAM_SRC, up + cmp rp, up + mov %ebx, SAVE_EBX + jne L(general) + mov GPARAM_SIZE, %ebx C size + mov GPARAM_SRC2, up + jmp L(inplace) + +L(general): + mov GPARAM_SIZE, %eax C size + mov %ebx, SAVE_EBX + incl %eax C size + 1 + mov up, %ebx C vp + mov GPARAM_SRC2, up C up + shr %eax C (size+1)\2 + mov %ebp, SAVE_EBP + mov %eax, GPARAM_SIZE + jnc L(entry2) C size odd + + add %edx, %edx C size even + mov %ecx, %ebp + mov (up), %ecx + lea -4(rp), rp + lea -4(%ebx), %ebx + lea (%ebp,%ecx,M), %eax + lea 4(up), up + jmp L(enteven2) + + ALIGN(16) +L(oop2): + lea (%ecx,%eax,M), %ebp + shr $RSH, %eax + mov 4(up), %ecx + add %edx, %edx + lea 8(up), up + mov (%ebx), %edx + M4_inst %ebp, %edx + lea (%eax,%ecx,M), %eax + mov %edx, (rp) +L(enteven2): + mov 4(%ebx), %edx + lea 8(%ebx), %ebx + M4_inst %eax, %edx + mov %edx, 4(rp) + sbb %edx, %edx + shr $RSH, %ecx + lea 8(rp), rp +L(entry2): + mov (up), %eax + decl GPARAM_SIZE + jnz L(oop2) + + lea (%ecx,%eax,M), %ebp + shr $RSH, %eax + shr %edx + mov (%ebx), %edx + M4_inst %ebp, %edx + mov %edx, (rp) + mov SAVE_UP, up + adc $0, %eax + mov SAVE_EBP, %ebp + mov SAVE_EBX, %ebx + pop rp FRAME_popl() + ret +EPILOGUE() + +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/atom/bdiv_q_1.asm b/gmp-6.3.0/mpn/x86/atom/bdiv_q_1.asm new file mode 100644 index 0000000..31e908e --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/bdiv_q_1.asm @@ -0,0 +1,35 @@ +dnl Intel Atom mpn_bdiv_q_1, mpn_pi1_bdiv_q_1 -- schoolbook Hensel +dnl division by 1-limb divisor, returning quotient only. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_bdiv_q_1 mpn_pi1_bdiv_q_1) +include_mpn(`x86/pentium/bdiv_q_1.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/cnd_add_n.asm b/gmp-6.3.0/mpn/x86/atom/cnd_add_n.asm new file mode 100644 index 0000000..50bf2ad --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/cnd_add_n.asm @@ -0,0 +1,113 @@ +dnl X86 mpn_cnd_add_n optimised for Intel Atom. + +dnl Copyright 2013 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C P5 ? +C P6 model 0-8,10-12 ? +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) ? +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) ? +C P4 model 3-4 (Prescott) ? +C Intel atom 4.67 +C AMD K6 ? +C AMD K7 ? +C AMD K8 ? + + +define(`rp', `%edi') +define(`up', `%esi') +define(`vp', `%ebp') +define(`n', `%ecx') +define(`cnd', `20(%esp)') + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(mpn_cnd_add_n) + push %edi + push %esi + push %ebx + push %ebp + + mov cnd, %eax C make cnd into a mask (1) + mov 24(%esp), rp + neg %eax C make cnd into a mask (1) + mov 28(%esp), up + sbb %eax, %eax C make cnd into a mask (1) + mov 32(%esp), vp + mov %eax, cnd C make cnd into a mask (1) + mov 36(%esp), n + + xor %edx, %edx + + shr $1, n + jnc L(top) + + mov 0(vp), %eax + and cnd, %eax + lea 4(vp), vp + add 0(up), %eax + lea 4(rp), rp + lea 4(up), up + sbb %edx, %edx + mov %eax, -4(rp) + inc n + dec n + je L(end) + +L(top): sbb %edx, %edx + mov 0(vp), %eax + and cnd, %eax + lea 8(vp), vp + lea 8(rp), rp + mov -4(vp), %ebx + and cnd, %ebx + add %edx, %edx + adc 0(up), %eax + lea 8(up), up + mov %eax, -8(rp) + adc -4(up), %ebx + dec n + mov %ebx, -4(rp) + jne L(top) + +L(end): mov $0, %eax + adc %eax, %eax + + pop %ebp + pop %ebx + pop %esi + pop %edi + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/atom/cnd_sub_n.asm b/gmp-6.3.0/mpn/x86/atom/cnd_sub_n.asm new file mode 100644 index 0000000..221bedc --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/cnd_sub_n.asm @@ -0,0 +1,124 @@ +dnl X86 mpn_cnd_sub_n optimised for Intel Atom. + +dnl Copyright 2013 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C P5 ? +C P6 model 0-8,10-12 ? +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) ? +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) ? +C P4 model 3-4 (Prescott) ? +C Intel atom 5.67 +C AMD K6 ? +C AMD K7 ? +C AMD K8 ? + + +define(`rp', `%edi') +define(`up', `%esi') +define(`vp', `%ebp') +define(`n', `%ecx') +define(`cnd', `20(%esp)') + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(mpn_cnd_sub_n) + push %edi + push %esi + push %ebx + push %ebp + + mov cnd, %eax C make cnd into a mask (1) + mov 24(%esp), rp + neg %eax C make cnd into a mask (1) + mov 28(%esp), up + sbb %eax, %eax C make cnd into a mask (1) + mov 32(%esp), vp + mov %eax, cnd C make cnd into a mask (1) + mov 36(%esp), n + + xor %edx, %edx + + inc n + shr n + jnc L(ent) + + mov 0(vp), %eax + and cnd, %eax + lea 4(vp), vp + mov 0(up), %edx + sub %eax, %edx + lea 4(rp), rp + lea 4(up), up + mov %edx, -4(rp) + sbb %edx, %edx C save cy + +L(ent): mov 0(vp), %ebx + and cnd, %ebx + add %edx, %edx C restore cy + mov 0(up), %edx + dec n + je L(end) + +L(top): sbb %ebx, %edx + mov 4(vp), %eax + mov %edx, 0(rp) + sbb %edx, %edx C save cy + mov 8(vp), %ebx + lea 8(up), up + and cnd, %ebx + and cnd, %eax + add %edx, %edx C restore cy + mov -4(up), %edx + lea 8(rp), rp + sbb %eax, %edx + mov %edx, -4(rp) + dec n + mov 0(up), %edx + lea 8(vp), vp + jne L(top) + +L(end): sbb %ebx, %edx + mov %edx, 0(rp) + + mov $0, %eax + adc %eax, %eax + + pop %ebp + pop %ebx + pop %esi + pop %edi + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/atom/dive_1.asm b/gmp-6.3.0/mpn/x86/atom/dive_1.asm new file mode 100644 index 0000000..71036a1 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/dive_1.asm @@ -0,0 +1,34 @@ +dnl Intel Atom mpn_divexact_1 -- mpn by limb exact division. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_divexact_1) +include_mpn(`x86/pentium/dive_1.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/gmp-mparam.h b/gmp-6.3.0/mpn/x86/atom/gmp-mparam.h new file mode 100644 index 0000000..e025bb7 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/gmp-mparam.h @@ -0,0 +1,214 @@ +/* Intel Atom/32 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 1600 MHz Diamondville (Atom 330) */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-18, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 3 +#define MOD_1_UNNORM_THRESHOLD 5 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 11 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 5 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 10 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 17 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 72.60% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 35 + +#define DIV_1_VS_MUL_1_PERCENT 236 + +#define MUL_TOOM22_THRESHOLD 22 +#define MUL_TOOM33_THRESHOLD 81 +#define MUL_TOOM44_THRESHOLD 178 +#define MUL_TOOM6H_THRESHOLD 270 +#define MUL_TOOM8H_THRESHOLD 399 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 97 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 126 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 115 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 129 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 115 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 32 +#define SQR_TOOM3_THRESHOLD 117 +#define SQR_TOOM4_THRESHOLD 178 +#define SQR_TOOM6_THRESHOLD 366 +#define SQR_TOOM8_THRESHOLD 527 + +#define MULMID_TOOM42_THRESHOLD 50 + +#define MULMOD_BNM1_THRESHOLD 13 +#define SQRMOD_BNM1_THRESHOLD 17 + +#define MUL_FFT_MODF_THRESHOLD 404 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 404, 5}, { 21, 6}, { 11, 5}, { 23, 6}, \ + { 12, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 21, 7}, { 11, 6}, { 25, 7}, { 13, 6}, \ + { 27, 7}, { 15, 6}, { 31, 7}, { 21, 8}, \ + { 11, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 39, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 39, 9}, { 23, 8}, \ + { 51,10}, { 15, 9}, { 31, 8}, { 67, 9}, \ + { 39, 8}, { 79, 9}, { 47, 8}, { 95,10}, \ + { 31, 9}, { 79,10}, { 47, 9}, { 95,11}, \ + { 31,10}, { 63, 9}, { 135,10}, { 79, 9}, \ + { 159,10}, { 95, 9}, { 191,10}, { 111,11}, \ + { 63,10}, { 127, 9}, { 255, 8}, { 511,10}, \ + { 143, 9}, { 287, 8}, { 575, 9}, { 303,10}, \ + { 159,11}, { 95,10}, { 191, 9}, { 383,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 511,10}, \ + { 271, 9}, { 543,10}, { 287, 9}, { 575,10}, \ + { 303,11}, { 159,10}, { 351, 9}, { 703,10}, \ + { 367, 9}, { 735,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 415,11}, { 223,10}, { 447,12}, \ + { 127,11}, { 255,10}, { 543,11}, { 287,10}, \ + { 607,11}, { 319,10}, { 671,11}, { 351,10}, \ + { 735,12}, { 191,11}, { 383,10}, { 767,11}, \ + { 415,10}, { 831,11}, { 447,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,12}, \ + { 319,11}, { 735,12}, { 383,11}, { 831,12}, \ + { 447,11}, { 959,13}, { 255,12}, { 511,11}, \ + { 1087,12}, { 575,11}, { 1215,12}, { 639,11}, \ + { 1279,12}, { 703,11}, { 1407,13}, { 383,12}, \ + { 831,11}, { 1663,12}, { 959,14}, { 255,13}, \ + { 511,12}, { 1215,13}, { 639,12}, { 1471,13}, \ + { 767,12}, { 1599,13}, { 895,12}, { 1791,14}, \ + { 511,13}, { 1023,12}, { 2111,13}, { 1151,12}, \ + { 2431,13}, { 1407,14}, { 767,13}, { 1663,12}, \ + { 3455,13}, { 1791,15}, { 511,14}, { 1023,13}, \ + { 2431,14}, { 1279,13}, { 2943,12}, { 5887,14}, \ + { 1535,13}, { 3455,14}, { 1791,13}, { 3839,15}, \ + { 1023,14}, { 2047,13}, { 4223,14}, { 2303,13}, \ + { 4991,12}, { 9983,14}, { 2815,13}, { 5887,15}, \ + { 1535,14}, { 3839,16} } +#define MUL_FFT_TABLE3_SIZE 158 +#define MUL_FFT_THRESHOLD 4544 + +#define SQR_FFT_MODF_THRESHOLD 368 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 368, 5}, { 23, 6}, { 12, 5}, { 25, 6}, \ + { 13, 5}, { 27, 6}, { 25, 7}, { 13, 6}, \ + { 28, 7}, { 15, 6}, { 31, 7}, { 17, 6}, \ + { 35, 7}, { 21, 8}, { 11, 7}, { 27, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 47, 8}, { 27, 9}, { 15, 8}, \ + { 39, 9}, { 23, 8}, { 51,10}, { 15, 9}, \ + { 31, 8}, { 63, 9}, { 39, 8}, { 79, 9}, \ + { 47,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 127, 8}, \ + { 255, 9}, { 135,10}, { 79, 9}, { 159, 8}, \ + { 319,10}, { 95, 9}, { 191,11}, { 63,10}, \ + { 127, 9}, { 255, 8}, { 511, 9}, { 271,10}, \ + { 143, 9}, { 287, 8}, { 575, 9}, { 303,10}, \ + { 159, 9}, { 319,11}, { 95,10}, { 191, 9}, \ + { 383,12}, { 63,11}, { 127,10}, { 255, 9}, \ + { 511,10}, { 271, 9}, { 543,10}, { 287, 9}, \ + { 575,10}, { 303, 9}, { 607,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335, 9}, { 671,10}, \ + { 351, 9}, { 703,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 415,11}, { 223,10}, { 447,12}, \ + { 127,11}, { 255,10}, { 543,11}, { 287,10}, \ + { 607,11}, { 319,10}, { 671,11}, { 351,10}, \ + { 703,12}, { 191,11}, { 383,10}, { 767,11}, \ + { 415,10}, { 831,11}, { 447,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 735,12}, \ + { 383,11}, { 831,12}, { 447,11}, { 959,13}, \ + { 255,12}, { 511,11}, { 1087,12}, { 575,11}, \ + { 1215,12}, { 639,11}, { 1343,12}, { 703,13}, \ + { 383,12}, { 959,14}, { 255,13}, { 511,12}, \ + { 1215,13}, { 639,12}, { 1471,13}, { 767,12}, \ + { 1599,13}, { 895,14}, { 511,13}, { 1023,12}, \ + { 2111,13}, { 1151,12}, { 2431,13}, { 1407,14}, \ + { 767,13}, { 1663,12}, { 3455,15}, { 511,14}, \ + { 1023,13}, { 2175,12}, { 4351,13}, { 2431,14}, \ + { 1279,13}, { 2943,12}, { 5887,14}, { 1535,13}, \ + { 3455,14}, { 1791,13}, { 3839,15}, { 1023,14}, \ + { 2047,13}, { 4351,14}, { 2303,13}, { 4991,12}, \ + { 9983,14}, { 2815,13}, { 5887,15}, { 1535,14}, \ + { 3839,16} } +#define SQR_FFT_TABLE3_SIZE 161 +#define SQR_FFT_THRESHOLD 3712 + +#define MULLO_BASECASE_THRESHOLD 0 /* always */ +#define MULLO_DC_THRESHOLD 56 +#define MULLO_MUL_N_THRESHOLD 8907 +#define SQRLO_BASECASE_THRESHOLD 6 +#define SQRLO_DC_THRESHOLD 111 +#define SQRLO_SQR_THRESHOLD 6654 + +#define DC_DIV_QR_THRESHOLD 67 +#define DC_DIVAPPR_Q_THRESHOLD 252 +#define DC_BDIV_QR_THRESHOLD 63 +#define DC_BDIV_Q_THRESHOLD 172 + +#define INV_MULMOD_BNM1_THRESHOLD 42 +#define INV_NEWTON_THRESHOLD 250 +#define INV_APPR_THRESHOLD 250 + +#define BINV_NEWTON_THRESHOLD 276 +#define REDC_1_TO_REDC_N_THRESHOLD 68 + +#define MU_DIV_QR_THRESHOLD 1334 +#define MU_DIVAPPR_Q_THRESHOLD 1442 +#define MUPI_DIV_QR_THRESHOLD 116 +#define MU_BDIV_QR_THRESHOLD 1142 +#define MU_BDIV_Q_THRESHOLD 1341 + +#define POWM_SEC_TABLE 1,16,98,376,1259 + +#define GET_STR_DC_THRESHOLD 12 +#define GET_STR_PRECOMPUTE_THRESHOLD 23 +#define SET_STR_DC_THRESHOLD 298 +#define SET_STR_PRECOMPUTE_THRESHOLD 1037 + +#define FAC_DSC_THRESHOLD 171 +#define FAC_ODD_THRESHOLD 34 + +#define MATRIX22_STRASSEN_THRESHOLD 17 +#define HGCD2_DIV1_METHOD 3 /* 3.71% faster than 1 */ +#define HGCD_THRESHOLD 128 +#define HGCD_APPR_THRESHOLD 186 +#define HGCD_REDUCE_THRESHOLD 2479 +#define GCD_DC_THRESHOLD 465 +#define GCDEXT_DC_THRESHOLD 339 +#define JACOBI_BASE_METHOD 3 /* 2.58% faster than 2 */ + +/* Tuneup completed successfully, took 214190 seconds */ diff --git a/gmp-6.3.0/mpn/x86/atom/logops_n.asm b/gmp-6.3.0/mpn/x86/atom/logops_n.asm new file mode 100644 index 0000000..3cb6d73 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/logops_n.asm @@ -0,0 +1,151 @@ +dnl Intel Atom mpn_and_n,...,mpn_xnor_n -- bitwise logical operations. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl Contributed to the GNU project by Marco Bodrato. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C op nop opn +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 3 3.5 3.5 +C AMD K6 +C AMD K7 +C AMD K8 +C AMD K10 + +define(M4_choose_op, +`ifdef(`OPERATION_$1',` +define(`M4_function', `mpn_$1') +define(`M4_want_pre', `$4') +define(`M4_inst', `$3') +define(`M4_want_post',`$2') +')') +define(M4pre, `ifelse(M4_want_pre, yes,`$1')') +define(M4post,`ifelse(M4_want_post,yes,`$1')') + +M4_choose_op( and_n, , andl, ) +M4_choose_op( andn_n, , andl, yes) +M4_choose_op( nand_n, yes, andl, ) +M4_choose_op( ior_n, , orl, ) +M4_choose_op( iorn_n, , orl, yes) +M4_choose_op( nior_n, yes, orl, ) +M4_choose_op( xor_n, , xorl, ) +M4_choose_op( xnor_n, yes, xorl, ) + +ifdef(`M4_function',, +`m4_error(`Unrecognised or undefined OPERATION symbol +')') + +MULFUNC_PROLOGUE(mpn_and_n mpn_andn_n mpn_nand_n mpn_ior_n mpn_iorn_n mpn_nior_n mpn_xor_n mpn_xnor_n) + +C void M4_function (mp_ptr dst, mp_srcptr src2, mp_srcptr src1, mp_size_t size); +C + +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC1, 12) +defframe(PARAM_SRC2, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(SAVE_RP,`PARAM_SIZE') +define(SAVE_VP,`PARAM_SRC1') +define(SAVE_UP,`PARAM_DST') + +define(`rp', `%edi') +define(`up', `%esi') +define(`vp', `%ebx') +define(`cnt', `%eax') +define(`r1', `%ecx') +define(`r2', `%edx') + +ASM_START() + TEXT + ALIGN(16) +deflit(`FRAME',0) + +PROLOGUE(M4_function) + mov PARAM_SIZE, cnt C size + mov rp, SAVE_RP + mov PARAM_DST, rp + mov up, SAVE_UP + mov PARAM_SRC1, up + shr cnt C size >> 1 + mov vp, SAVE_VP + mov PARAM_SRC2, vp + mov (up), r1 + jz L(end) C size == 1 + jnc L(even) C size % 2 == 0 + + ALIGN(16) +L(oop): +M4pre(` notl_or_xorl_GMP_NUMB_MASK(r1)') + M4_inst (vp), r1 + lea 8(up), up + mov -4(up), r2 +M4post(` notl_or_xorl_GMP_NUMB_MASK(r1)') + lea 8(vp), vp + mov r1, (rp) +L(entry): +M4pre(` notl_or_xorl_GMP_NUMB_MASK(r2)') + M4_inst -4(vp), r2 + lea 8(rp), rp +M4post(` notl_or_xorl_GMP_NUMB_MASK(r2)') + dec cnt + mov (up), r1 + mov r2, -4(rp) + jnz L(oop) + +L(end): +M4pre(` notl_or_xorl_GMP_NUMB_MASK(r1)') + mov SAVE_UP, up + M4_inst (vp), r1 +M4post(`notl_or_xorl_GMP_NUMB_MASK(r1)') + mov SAVE_VP, vp + mov r1, (rp) + mov SAVE_RP, rp + ret + +L(even): + mov r1, r2 + lea 4(up), up + lea 4(vp), vp + lea -4(rp), rp + jmp L(entry) +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/atom/lshift.asm b/gmp-6.3.0/mpn/x86/atom/lshift.asm new file mode 100644 index 0000000..f2c70dd --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/lshift.asm @@ -0,0 +1,218 @@ +dnl Intel Atom mpn_lshift -- mpn left shift. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl Contributed to the GNU project by Torbjorn Granlund and Marco Bodrato. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C mp_limb_t mpn_lshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned cnt); + +C cycles/limb +C cnt!=1 cnt==1 +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 5 2.5 +C AMD K6 +C AMD K7 +C AMD K8 +C AMD K10 + +defframe(PARAM_CNT, 16) +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(SAVE_UP,`PARAM_CNT') +define(VAR_COUNT,`PARAM_SIZE') +define(SAVE_EBX,`PARAM_SRC') +define(SAVE_EBP,`PARAM_DST') + +define(`rp', `%edi') +define(`up', `%esi') +define(`cnt', `%ecx') + +ASM_START() + TEXT + ALIGN(8) +deflit(`FRAME',0) +PROLOGUE(mpn_lshift) + mov PARAM_CNT, cnt + mov PARAM_SIZE, %edx + mov up, SAVE_UP + mov PARAM_SRC, up + push rp FRAME_pushl() + mov PARAM_DST, rp + +C We can use faster code for shift-by-1 under certain conditions. + cmp $1,cnt + jne L(normal) + cmpl rp, up + jnc L(special) C jump if s_ptr + 1 >= res_ptr + leal (up,%edx,4),%eax + cmpl %eax,rp + jnc L(special) C jump if res_ptr >= s_ptr + size + +L(normal): + lea -4(up,%edx,4), up + mov %ebx, SAVE_EBX + lea -4(rp,%edx,4), rp + + shr %edx + mov (up), %eax + mov %edx, VAR_COUNT + jnc L(evn) + + mov %eax, %ebx + shl %cl, %ebx + neg cnt + shr %cl, %eax + test %edx, %edx + jnz L(gt1) + mov %ebx, (rp) + jmp L(quit) + +L(gt1): mov %ebp, SAVE_EBP + push %eax + mov -4(up), %eax + mov %eax, %ebp + shr %cl, %eax + jmp L(lo1) + +L(evn): mov %ebp, SAVE_EBP + neg cnt + mov %eax, %ebp + mov -4(up), %edx + shr %cl, %eax + mov %edx, %ebx + shr %cl, %edx + neg cnt + decl VAR_COUNT + lea 4(rp), rp + lea -4(up), up + jz L(end) + push %eax FRAME_pushl() + + ALIGN(8) +L(top): shl %cl, %ebp + or %ebp, %edx + shl %cl, %ebx + neg cnt + mov -4(up), %eax + mov %eax, %ebp + mov %edx, -4(rp) + shr %cl, %eax + lea -8(rp), rp +L(lo1): mov -8(up), %edx + or %ebx, %eax + mov %edx, %ebx + shr %cl, %edx + lea -8(up), up + neg cnt + mov %eax, (rp) + decl VAR_COUNT + jg L(top) + + pop %eax FRAME_popl() +L(end): + shl %cl, %ebp + shl %cl, %ebx + or %ebp, %edx + mov SAVE_EBP, %ebp + mov %edx, -4(rp) + mov %ebx, -8(rp) + +L(quit): + mov SAVE_UP, up + mov SAVE_EBX, %ebx + pop rp FRAME_popl() + ret + +L(special): +deflit(`FRAME',4) + lea 3(%edx), %eax C size + 3 + dec %edx C size - 1 + mov (up), %ecx + shr $2, %eax C (size + 3) / 4 + and $3, %edx C (size - 1) % 4 + jz L(goloop) C jmp if size == 1 (mod 4) + shr %edx + jnc L(odd) C jum if size == 3 (mod 4) + + add %ecx, %ecx + lea 4(up), up + mov %ecx, (rp) + mov (up), %ecx + lea 4(rp), rp + + dec %edx + jnz L(goloop) C jump if size == 0 (mod 4) +L(odd): lea -8(up), up + lea -8(rp), rp + jmp L(sentry) C reached if size == 2 or 3 (mod 4) + +L(sloop): + adc %ecx, %ecx + mov 4(up), %edx + mov %ecx, (rp) + adc %edx, %edx + mov 8(up), %ecx + mov %edx, 4(rp) +L(sentry): + adc %ecx, %ecx + mov 12(up), %edx + mov %ecx, 8(rp) + adc %edx, %edx + lea 16(up), up + mov %edx, 12(rp) + lea 16(rp), rp + mov (up), %ecx +L(goloop): + decl %eax + jnz L(sloop) + +L(squit): + adc %ecx, %ecx + mov %ecx, (rp) + adc %eax, %eax + + mov SAVE_UP, up + pop rp FRAME_popl() + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/atom/lshiftc.asm b/gmp-6.3.0/mpn/x86/atom/lshiftc.asm new file mode 100644 index 0000000..5be53ed --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/lshiftc.asm @@ -0,0 +1,159 @@ +dnl Intel Atom mpn_lshiftc -- mpn left shift with complement. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl Contributed to the GNU project by Torbjorn Granlund and Marco Bodrato. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C mp_limb_t mpn_lshiftc (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned cnt); + +C cycles/limb +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 5.5 +C AMD K6 +C AMD K7 +C AMD K8 +C AMD K10 + +defframe(PARAM_CNT, 16) +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(SAVE_UP,`PARAM_CNT') +define(VAR_COUNT,`PARAM_SIZE') +define(SAVE_EBX,`PARAM_SRC') +define(SAVE_EBP,`PARAM_DST') + +define(`rp', `%edi') +define(`up', `%esi') +define(`cnt', `%ecx') + +ASM_START() + TEXT + +PROLOGUE(mpn_lshiftc) +deflit(`FRAME',0) + mov PARAM_CNT, cnt + mov PARAM_SIZE, %edx + mov up, SAVE_UP + mov PARAM_SRC, up + push rp FRAME_pushl() + mov PARAM_DST, rp + + lea -4(up,%edx,4), up + mov %ebx, SAVE_EBX + lea -4(rp,%edx,4), rp + + shr %edx + mov (up), %eax + mov %edx, VAR_COUNT + jnc L(evn) + + mov %eax, %ebx + shl %cl, %ebx + neg cnt + shr %cl, %eax + test %edx, %edx + jnz L(gt1) + not %ebx + mov %ebx, (rp) + jmp L(quit) + +L(gt1): mov %ebp, SAVE_EBP + push %eax + mov -4(up), %eax + mov %eax, %ebp + shr %cl, %eax + jmp L(lo1) + +L(evn): mov %ebp, SAVE_EBP + neg cnt + mov %eax, %ebp + mov -4(up), %edx + shr %cl, %eax + mov %edx, %ebx + shr %cl, %edx + neg cnt + decl VAR_COUNT + lea 4(rp), rp + lea -4(up), up + jz L(end) + push %eax FRAME_pushl() + +L(top): shl %cl, %ebp + or %ebp, %edx + shl %cl, %ebx + neg cnt + not %edx + mov -4(up), %eax + mov %eax, %ebp + mov %edx, -4(rp) + shr %cl, %eax + lea -8(rp), rp +L(lo1): mov -8(up), %edx + or %ebx, %eax + mov %edx, %ebx + shr %cl, %edx + not %eax + lea -8(up), up + neg cnt + mov %eax, (rp) + decl VAR_COUNT + jg L(top) + + pop %eax FRAME_popl() +L(end): + shl %cl, %ebp + shl %cl, %ebx + or %ebp, %edx + mov SAVE_EBP, %ebp + not %edx + not %ebx + mov %edx, -4(rp) + mov %ebx, -8(rp) + +L(quit): + mov SAVE_UP, up + mov SAVE_EBX, %ebx + pop rp FRAME_popl() + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/atom/mmx/copyd.asm b/gmp-6.3.0/mpn/x86/atom/mmx/copyd.asm new file mode 100644 index 0000000..b80fb03 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/mmx/copyd.asm @@ -0,0 +1,34 @@ +dnl Intel Atom mpn_copyd -- copy limb vector, decrementing. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_copyd) +include_mpn(`x86/k7/mmx/copyd.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/mmx/copyi.asm b/gmp-6.3.0/mpn/x86/atom/mmx/copyi.asm new file mode 100644 index 0000000..49b6b8d --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/mmx/copyi.asm @@ -0,0 +1,34 @@ +dnl Intel Atom mpn_copyi -- copy limb vector, incrementing. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_copyi) +include_mpn(`x86/k7/mmx/copyi.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/mmx/hamdist.asm b/gmp-6.3.0/mpn/x86/atom/mmx/hamdist.asm new file mode 100644 index 0000000..3fe8253 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/mmx/hamdist.asm @@ -0,0 +1,34 @@ +dnl Intel Atom mpn_hamdist -- hamming distance. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_hamdist) +include_mpn(`x86/k7/mmx/popham.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/mod_34lsub1.asm b/gmp-6.3.0/mpn/x86/atom/mod_34lsub1.asm new file mode 100644 index 0000000..6d57ba3 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/mod_34lsub1.asm @@ -0,0 +1,34 @@ +dnl Intel Atom mpn_mod_34lsub1 -- remainder modulo 2^24-1. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_mod_34lsub1) +include_mpn(`x86/p6/mod_34lsub1.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/mode1o.asm b/gmp-6.3.0/mpn/x86/atom/mode1o.asm new file mode 100644 index 0000000..c9ee6bd --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/mode1o.asm @@ -0,0 +1,34 @@ +dnl Intel Atom mpn_modexact_1_odd -- exact division style remainder. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_modexact_1_odd mpn_modexact_1c_odd) +include_mpn(`x86/pentium/mode1o.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/rshift.asm b/gmp-6.3.0/mpn/x86/atom/rshift.asm new file mode 100644 index 0000000..1cb5dbe --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/rshift.asm @@ -0,0 +1,152 @@ +dnl Intel Atom mpn_rshift -- mpn right shift. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl Converted from AMD64 by Marco Bodrato. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C mp_limb_t mpn_rshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned cnt); + +C cycles/limb +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 5 +C AMD K6 +C AMD K7 +C AMD K8 +C AMD K10 + +defframe(PARAM_CNT, 16) +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(SAVE_UP,`PARAM_CNT') +define(VAR_COUNT,`PARAM_SIZE') +define(SAVE_EBX,`PARAM_SRC') +define(SAVE_EBP,`PARAM_DST') + +define(`rp', `%edi') +define(`up', `%esi') +define(`cnt', `%ecx') + +ASM_START() + TEXT + ALIGN(8) +deflit(`FRAME',0) +PROLOGUE(mpn_rshift) + mov PARAM_CNT, cnt + mov PARAM_SIZE, %edx + mov up, SAVE_UP + mov PARAM_SRC, up + push rp FRAME_pushl() + mov PARAM_DST, rp + mov %ebx, SAVE_EBX + + shr %edx + mov (up), %eax + mov %edx, VAR_COUNT + jnc L(evn) + + mov %eax, %ebx + shr %cl, %ebx + neg cnt + shl %cl, %eax + test %edx, %edx + jnz L(gt1) + mov %ebx, (rp) + jmp L(quit) + +L(gt1): mov %ebp, SAVE_EBP + push %eax + mov 4(up), %eax + mov %eax, %ebp + shl %cl, %eax + jmp L(lo1) + +L(evn): mov %ebp, SAVE_EBP + neg cnt + mov %eax, %ebp + mov 4(up), %edx + shl %cl, %eax + mov %edx, %ebx + shl %cl, %edx + neg cnt + decl VAR_COUNT + lea -4(rp), rp + lea 4(up), up + jz L(end) + push %eax FRAME_pushl() + + ALIGN(8) +L(top): shr %cl, %ebp + or %ebp, %edx + shr %cl, %ebx + neg cnt + mov 4(up), %eax + mov %eax, %ebp + mov %edx, 4(rp) + shl %cl, %eax + lea 8(rp), rp +L(lo1): mov 8(up), %edx + or %ebx, %eax + mov %edx, %ebx + shl %cl, %edx + lea 8(up), up + neg cnt + mov %eax, (rp) + decl VAR_COUNT + jg L(top) + + pop %eax FRAME_popl() +L(end): + shr %cl, %ebp + shr %cl, %ebx + or %ebp, %edx + mov SAVE_EBP, %ebp + mov %edx, 4(rp) + mov %ebx, 8(rp) + +L(quit): + mov SAVE_UP, up + mov SAVE_EBX, %ebx + pop rp FRAME_popl() + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/atom/sse2/aorsmul_1.asm b/gmp-6.3.0/mpn/x86/atom/sse2/aorsmul_1.asm new file mode 100644 index 0000000..969a14a --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sse2/aorsmul_1.asm @@ -0,0 +1,174 @@ +dnl x86-32 mpn_addmul_1 and mpn_submul_1 optimised for Intel Atom. + +dnl Contributed to the GNU project by Torbjorn Granlund and Marco Bodrato. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C cycles/limb +C P5 - +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 8 +C AMD K6 +C AMD K7 - +C AMD K8 +C AMD K10 + +define(`rp', `%edi') +define(`up', `%esi') +define(`n', `%ecx') + +ifdef(`OPERATION_addmul_1',` + define(ADDSUB, add) + define(func_1, mpn_addmul_1) + define(func_1c, mpn_addmul_1c)') +ifdef(`OPERATION_submul_1',` + define(ADDSUB, sub) + define(func_1, mpn_submul_1) + define(func_1c, mpn_submul_1c)') + +MULFUNC_PROLOGUE(mpn_addmul_1 mpn_addmul_1c mpn_submul_1 mpn_submul_1c) + + TEXT + ALIGN(16) +PROLOGUE(func_1) + xor %edx, %edx +L(ent): push %edi + push %esi + push %ebx + mov 16(%esp), rp + mov 20(%esp), up + mov 24(%esp), n + movd 28(%esp), %mm7 + test $1, n + jz L(fi0or2) + movd (up), %mm0 + pmuludq %mm7, %mm0 + shr $2, n + jnc L(fi1) + +L(fi3): lea -8(up), up + lea -8(rp), rp + movd 12(up), %mm1 + movd %mm0, %ebx + pmuludq %mm7, %mm1 + add $1, n C increment and clear carry + jmp L(lo3) + +L(fi1): movd %mm0, %ebx + jz L(wd1) + movd 4(up), %mm1 + pmuludq %mm7, %mm1 + jmp L(lo1) + +L(fi0or2): + movd (up), %mm1 + pmuludq %mm7, %mm1 + shr $2, n + movd 4(up), %mm0 + jc L(fi2) + lea -4(up), up + lea -4(rp), rp + movd %mm1, %eax + pmuludq %mm7, %mm0 + jmp L(lo0) + +L(fi2): lea 4(up), up + add $1, n C increment and clear carry + movd %mm1, %eax + lea -12(rp), rp + jmp L(lo2) + +C ALIGN(16) C alignment seems irrelevant +L(top): movd 4(up), %mm1 + adc $0, %edx + ADDSUB %eax, 12(rp) + movd %mm0, %ebx + pmuludq %mm7, %mm1 + lea 16(rp), rp +L(lo1): psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %edx + movd %mm1, %eax + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + adc $0, %edx + ADDSUB %ebx, (rp) +L(lo0): psrlq $32, %mm1 + adc %edx, %eax + movd %mm1, %edx + movd %mm0, %ebx + movd 12(up), %mm1 + pmuludq %mm7, %mm1 + adc $0, %edx + ADDSUB %eax, 4(rp) +L(lo3): psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %edx + movd %mm1, %eax + lea 16(up), up + movd (up), %mm0 + adc $0, %edx + ADDSUB %ebx, 8(rp) +L(lo2): psrlq $32, %mm1 + adc %edx, %eax + movd %mm1, %edx + pmuludq %mm7, %mm0 + dec n + jnz L(top) + +L(end): adc n, %edx C n is zero here + ADDSUB %eax, 12(rp) + movd %mm0, %ebx + lea 16(rp), rp +L(wd1): psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %eax + adc n, %eax + ADDSUB %ebx, (rp) + emms + adc n, %eax + pop %ebx + pop %esi + pop %edi + ret +EPILOGUE() +PROLOGUE(func_1c) + mov 20(%esp), %edx C carry + jmp L(ent) +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/atom/sse2/bdiv_dbm1c.asm b/gmp-6.3.0/mpn/x86/atom/sse2/bdiv_dbm1c.asm new file mode 100644 index 0000000..782e914 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sse2/bdiv_dbm1c.asm @@ -0,0 +1,34 @@ +dnl Intel Atom mpn_bdiv_dbm1. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_bdiv_dbm1c) +include_mpn(`x86/pentium4/sse2/bdiv_dbm1c.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/sse2/divrem_1.asm b/gmp-6.3.0/mpn/x86/atom/sse2/divrem_1.asm new file mode 100644 index 0000000..f84709a --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sse2/divrem_1.asm @@ -0,0 +1,34 @@ +dnl Intel Atom mpn_divrem_1 -- mpn by limb division. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_preinv_divrem_1 mpn_divrem_1c mpn_divrem_1) +include_mpn(`x86/pentium4/sse2/divrem_1.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/sse2/mod_1_1.asm b/gmp-6.3.0/mpn/x86/atom/sse2/mod_1_1.asm new file mode 100644 index 0000000..ae6581d --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sse2/mod_1_1.asm @@ -0,0 +1,34 @@ +dnl Intel Atom/SSE2 mpn_mod_1_1. + +dnl Copyright 2009, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_mod_1_1p) +include_mpn(`x86/pentium4/sse2/mod_1_1.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/sse2/mod_1_4.asm b/gmp-6.3.0/mpn/x86/atom/sse2/mod_1_4.asm new file mode 100644 index 0000000..31faa3f --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sse2/mod_1_4.asm @@ -0,0 +1,34 @@ +dnl Intel Atom/SSE2 mpn_mod_1_4. + +dnl Copyright 2009, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_mod_1s_4p) +include_mpn(`x86/pentium4/sse2/mod_1_4.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/sse2/mul_1.asm b/gmp-6.3.0/mpn/x86/atom/sse2/mul_1.asm new file mode 100644 index 0000000..aa3bb97 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sse2/mul_1.asm @@ -0,0 +1,124 @@ +dnl Intel Atom mpn_mul_1. + +dnl Contributed to the GNU project by Torbjorn Granlund and Marco Bodrato. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C cycles/limb +C P5 - +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 7.5 +C AMD K6 - +C AMD K7 - +C AMD K8 +C AMD K10 + +defframe(PARAM_CARRY,20) +defframe(PARAM_MUL, 16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +define(`rp', `%edx') +define(`up', `%esi') +define(`n', `%ecx') + +ASM_START() + TEXT + ALIGN(16) +deflit(`FRAME',0) + +PROLOGUE(mpn_mul_1c) + movd PARAM_CARRY, %mm6 C carry + jmp L(ent) +EPILOGUE() + + ALIGN(8) C for compact code +PROLOGUE(mpn_mul_1) + pxor %mm6, %mm6 +L(ent): push %esi FRAME_pushl() + mov PARAM_SRC, up + mov PARAM_SIZE, %eax C size + movd PARAM_MUL, %mm7 + movd (up), %mm0 + mov %eax, n + and $3, %eax + pmuludq %mm7, %mm0 + mov PARAM_DST, rp + jz L(lo0) + cmp $2, %eax + lea -16(up,%eax,4),up + lea -16(rp,%eax,4),rp + jc L(lo1) + jz L(lo2) + jmp L(lo3) + + ALIGN(16) +L(top): movd (up), %mm0 + pmuludq %mm7, %mm0 + psrlq $32, %mm6 + lea 16(rp), rp +L(lo0): paddq %mm0, %mm6 + movd 4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, (rp) + psrlq $32, %mm6 +L(lo3): paddq %mm0, %mm6 + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, 4(rp) + psrlq $32, %mm6 +L(lo2): paddq %mm0, %mm6 + movd 12(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, 8(rp) + psrlq $32, %mm6 +L(lo1): paddq %mm0, %mm6 + sub $4, n + movd %mm6, 12(rp) + lea 16(up), up + ja L(top) + + psrlq $32, %mm6 + movd %mm6, %eax + emms + pop %esi FRAME_popl() + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/atom/sse2/mul_basecase.asm b/gmp-6.3.0/mpn/x86/atom/sse2/mul_basecase.asm new file mode 100644 index 0000000..97d3aeb --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sse2/mul_basecase.asm @@ -0,0 +1,501 @@ +dnl x86 mpn_mul_basecase -- Multiply two limb vectors and store the result in +dnl a third limb vector. + +dnl Contributed to the GNU project by Torbjorn Granlund and Marco Bodrato. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C TODO +C * Check if 'jmp N(%esp)' is well-predicted enough to allow us to combine the +C 4 large loops into one; we could use it for the outer loop branch. +C * Optimise code outside of inner loops. +C * Write combined addmul_1 feed-in a wind-down code, and use when iterating +C outer each loop. ("Overlapping software pipelining") +C * Postpone push of ebx until we know vn > 1. Perhaps use caller-saves regs +C for inlined mul_1, allowing us to postpone all pushes. +C * Perhaps write special code for vn <= un < M, for some small M. + +C void mpn_mul_basecase (mp_ptr wp, +C mp_srcptr xp, mp_size_t xn, +C mp_srcptr yp, mp_size_t yn); +C + +define(`rp', `%edi') +define(`up', `%esi') +define(`un', `%ecx') +define(`vp', `%ebp') +define(`vn', `36(%esp)') + + TEXT + ALIGN(16) +PROLOGUE(mpn_mul_basecase) + push %edi + push %esi + push %ebx + push %ebp + mov 20(%esp), rp + mov 24(%esp), up + mov 28(%esp), un + mov 32(%esp), vp + + movd (up), %mm0 + movd (vp), %mm7 + pmuludq %mm7, %mm0 + pxor %mm6, %mm6 + + mov un, %eax + and $3, %eax + jz L(of0) + cmp $2, %eax + jc L(of1) + jz L(of2) + +C ================================================================ + jmp L(m3) + ALIGN(16) +L(lm3): movd -4(up), %mm0 + pmuludq %mm7, %mm0 + psrlq $32, %mm6 + lea 16(rp), rp + paddq %mm0, %mm6 + movd (up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -4(rp) + psrlq $32, %mm6 +L(m3): paddq %mm0, %mm6 + movd 4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, (rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, 4(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + sub $4, un + movd %mm6, 8(rp) + lea 16(up), up + ja L(lm3) + + psrlq $32, %mm6 + movd %mm6, 12(rp) + + decl vn + jz L(done) + lea -8(rp), rp + +L(ol3): mov 28(%esp), un + neg un + lea 4(vp), vp + movd (vp), %mm7 C read next V limb + mov 24(%esp), up + lea 16(rp,un,4), rp + + movd (up), %mm0 + pmuludq %mm7, %mm0 + sar $2, un + movd 4(up), %mm1 + movd %mm0, %ebx + pmuludq %mm7, %mm1 + lea -8(up), up + xor %edx, %edx C zero edx and CF + jmp L(a3) + +L(la3): movd 4(up), %mm1 + adc $0, %edx + add %eax, 12(rp) + movd %mm0, %ebx + pmuludq %mm7, %mm1 + lea 16(rp), rp + psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %edx + movd %mm1, %eax + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + adc $0, %edx + add %ebx, (rp) + psrlq $32, %mm1 + adc %edx, %eax + movd %mm1, %edx + movd %mm0, %ebx + movd 12(up), %mm1 + pmuludq %mm7, %mm1 + adc $0, %edx + add %eax, 4(rp) +L(a3): psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %edx + movd %mm1, %eax + lea 16(up), up + movd (up), %mm0 + adc $0, %edx + add %ebx, 8(rp) + psrlq $32, %mm1 + adc %edx, %eax + movd %mm1, %edx + pmuludq %mm7, %mm0 + inc un + jnz L(la3) + + adc un, %edx C un is zero here + add %eax, 12(rp) + movd %mm0, %ebx + psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %eax + adc un, %eax + add %ebx, 16(rp) + adc un, %eax + mov %eax, 20(rp) + + decl vn + jnz L(ol3) + jmp L(done) + +C ================================================================ + ALIGN(16) +L(lm0): movd (up), %mm0 + pmuludq %mm7, %mm0 + psrlq $32, %mm6 + lea 16(rp), rp +L(of0): paddq %mm0, %mm6 + movd 4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, (rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, 4(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd 12(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, 8(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + sub $4, un + movd %mm6, 12(rp) + lea 16(up), up + ja L(lm0) + + psrlq $32, %mm6 + movd %mm6, 16(rp) + + decl vn + jz L(done) + lea -4(rp), rp + +L(ol0): mov 28(%esp), un + neg un + lea 4(vp), vp + movd (vp), %mm7 C read next V limb + mov 24(%esp), up + lea 20(rp,un,4), rp + + movd (up), %mm1 + pmuludq %mm7, %mm1 + sar $2, un + movd 4(up), %mm0 + lea -4(up), up + movd %mm1, %eax + pmuludq %mm7, %mm0 + xor %edx, %edx C zero edx and CF + jmp L(a0) + +L(la0): movd 4(up), %mm1 + adc $0, %edx + add %eax, 12(rp) + movd %mm0, %ebx + pmuludq %mm7, %mm1 + lea 16(rp), rp + psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %edx + movd %mm1, %eax + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + adc $0, %edx + add %ebx, (rp) +L(a0): psrlq $32, %mm1 + adc %edx, %eax + movd %mm1, %edx + movd %mm0, %ebx + movd 12(up), %mm1 + pmuludq %mm7, %mm1 + adc $0, %edx + add %eax, 4(rp) + psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %edx + movd %mm1, %eax + lea 16(up), up + movd (up), %mm0 + adc $0, %edx + add %ebx, 8(rp) + psrlq $32, %mm1 + adc %edx, %eax + movd %mm1, %edx + pmuludq %mm7, %mm0 + inc un + jnz L(la0) + + adc un, %edx C un is zero here + add %eax, 12(rp) + movd %mm0, %ebx + psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %eax + adc un, %eax + add %ebx, 16(rp) + adc un, %eax + mov %eax, 20(rp) + + decl vn + jnz L(ol0) + jmp L(done) + +C ================================================================ + ALIGN(16) +L(lm1): movd -12(up), %mm0 + pmuludq %mm7, %mm0 + psrlq $32, %mm6 + lea 16(rp), rp + paddq %mm0, %mm6 + movd -8(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -12(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd -4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -8(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd (up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -4(rp) + psrlq $32, %mm6 +L(of1): paddq %mm0, %mm6 + sub $4, un + movd %mm6, (rp) + lea 16(up), up + ja L(lm1) + + psrlq $32, %mm6 + movd %mm6, 4(rp) + + decl vn + jz L(done) + lea -16(rp), rp + +L(ol1): mov 28(%esp), un + neg un + lea 4(vp), vp + movd (vp), %mm7 C read next V limb + mov 24(%esp), up + lea 24(rp,un,4), rp + + movd (up), %mm0 + pmuludq %mm7, %mm0 + sar $2, un + movd %mm0, %ebx + movd 4(up), %mm1 + pmuludq %mm7, %mm1 + xor %edx, %edx C zero edx and CF + inc un + jmp L(a1) + +L(la1): movd 4(up), %mm1 + adc $0, %edx + add %eax, 12(rp) + movd %mm0, %ebx + pmuludq %mm7, %mm1 + lea 16(rp), rp +L(a1): psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %edx + movd %mm1, %eax + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + adc $0, %edx + add %ebx, (rp) + psrlq $32, %mm1 + adc %edx, %eax + movd %mm1, %edx + movd %mm0, %ebx + movd 12(up), %mm1 + pmuludq %mm7, %mm1 + adc $0, %edx + add %eax, 4(rp) + psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %edx + movd %mm1, %eax + lea 16(up), up + movd (up), %mm0 + adc $0, %edx + add %ebx, 8(rp) + psrlq $32, %mm1 + adc %edx, %eax + movd %mm1, %edx + pmuludq %mm7, %mm0 + inc un + jnz L(la1) + + adc un, %edx C un is zero here + add %eax, 12(rp) + movd %mm0, %ebx + psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %eax + adc un, %eax + add %ebx, 16(rp) + adc un, %eax + mov %eax, 20(rp) + + decl vn + jnz L(ol1) + jmp L(done) + +C ================================================================ + ALIGN(16) +L(lm2): movd -8(up), %mm0 + pmuludq %mm7, %mm0 + psrlq $32, %mm6 + lea 16(rp), rp + paddq %mm0, %mm6 + movd -4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -8(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd (up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -4(rp) + psrlq $32, %mm6 +L(of2): paddq %mm0, %mm6 + movd 4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, (rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + sub $4, un + movd %mm6, 4(rp) + lea 16(up), up + ja L(lm2) + + psrlq $32, %mm6 + movd %mm6, 8(rp) + + decl vn + jz L(done) + lea -12(rp), rp + +L(ol2): mov 28(%esp), un + neg un + lea 4(vp), vp + movd (vp), %mm7 C read next V limb + mov 24(%esp), up + lea 12(rp,un,4), rp + + movd (up), %mm1 + pmuludq %mm7, %mm1 + sar $2, un + movd 4(up), %mm0 + lea 4(up), up + movd %mm1, %eax + xor %edx, %edx C zero edx and CF + jmp L(lo2) + +L(la2): movd 4(up), %mm1 + adc $0, %edx + add %eax, 12(rp) + movd %mm0, %ebx + pmuludq %mm7, %mm1 + lea 16(rp), rp + psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %edx + movd %mm1, %eax + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + adc $0, %edx + add %ebx, (rp) + psrlq $32, %mm1 + adc %edx, %eax + movd %mm1, %edx + movd %mm0, %ebx + movd 12(up), %mm1 + pmuludq %mm7, %mm1 + adc $0, %edx + add %eax, 4(rp) + psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %edx + movd %mm1, %eax + lea 16(up), up + movd (up), %mm0 + adc $0, %edx + add %ebx, 8(rp) +L(lo2): psrlq $32, %mm1 + adc %edx, %eax + movd %mm1, %edx + pmuludq %mm7, %mm0 + inc un + jnz L(la2) + + adc un, %edx C un is zero here + add %eax, 12(rp) + movd %mm0, %ebx + psrlq $32, %mm0 + adc %edx, %ebx + movd %mm0, %eax + adc un, %eax + add %ebx, 16(rp) + adc un, %eax + mov %eax, 20(rp) + + decl vn + jnz L(ol2) +C jmp L(done) + +C ================================================================ +L(done): + emms + pop %ebp + pop %ebx + pop %esi + pop %edi + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/atom/sse2/popcount.asm b/gmp-6.3.0/mpn/x86/atom/sse2/popcount.asm new file mode 100644 index 0000000..7847aec --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sse2/popcount.asm @@ -0,0 +1,35 @@ +dnl Intel Atom mpn_popcount -- population count. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +MULFUNC_PROLOGUE(mpn_popcount) +include_mpn(`x86/pentium4/sse2/popcount.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/sse2/sqr_basecase.asm b/gmp-6.3.0/mpn/x86/atom/sse2/sqr_basecase.asm new file mode 100644 index 0000000..af19ed8 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sse2/sqr_basecase.asm @@ -0,0 +1,634 @@ +dnl x86 mpn_sqr_basecase -- square an mpn number, optimised for atom. + +dnl Contributed to the GNU project by Torbjorn Granlund and Marco Bodrato. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C TODO +C * Check if 'jmp N(%esp)' is well-predicted enough to allow us to combine the +C 4 large loops into one; we could use it for the outer loop branch. +C * Optimise code outside of inner loops. +C * Write combined addmul_1 feed-in a wind-down code, and use when iterating +C outer each loop. ("Overlapping software pipelining") +C * Perhaps use caller-saves regs for inlined mul_1, allowing us to postpone +C all pushes. +C * Perhaps write special code for n < M, for some small M. +C * Replace inlined addmul_1 with smaller code from aorsmul_1.asm, or perhaps +C with even less pipelined code. +C * We run the outer loop until we have a 2-limb by 1-limb addmul_1 left. +C Consider breaking out earlier, saving high the cost of short loops. + +C void mpn_sqr_basecase (mp_ptr wp, +C mp_srcptr xp, mp_size_t xn); + +define(`rp', `%edi') +define(`up', `%esi') +define(`n', `%ecx') + +define(`un', `%ebp') + + TEXT + ALIGN(16) +PROLOGUE(mpn_sqr_basecase) + push %edi + push %esi + mov 12(%esp), rp + mov 16(%esp), up + mov 20(%esp), n + + lea 4(rp), rp C write triangular product starting at rp[1] + dec n + movd (up), %mm7 + + jz L(one) + lea 4(up), up + push %ebx + push %ebp + mov n, %eax + + movd (up), %mm0 + neg n + pmuludq %mm7, %mm0 + pxor %mm6, %mm6 + mov n, un + + and $3, %eax + jz L(of0) + cmp $2, %eax + jc L(of1) + jz L(of2) + +C ================================================================ + jmp L(m3) + ALIGN(16) +L(lm3): movd -4(up), %mm0 + pmuludq %mm7, %mm0 + psrlq $32, %mm6 + lea 16(rp), rp + paddq %mm0, %mm6 + movd (up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -4(rp) + psrlq $32, %mm6 +L(m3): paddq %mm0, %mm6 + movd 4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, (rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, 4(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + add $4, un + movd %mm6, 8(rp) + lea 16(up), up + js L(lm3) + + psrlq $32, %mm6 + movd %mm6, 12(rp) + + inc n +C jz L(done) + lea -12(up), up + lea 4(rp), rp + jmp L(ol2) + +C ================================================================ + ALIGN(16) +L(lm0): movd (up), %mm0 + pmuludq %mm7, %mm0 + psrlq $32, %mm6 + lea 16(rp), rp +L(of0): paddq %mm0, %mm6 + movd 4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, (rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, 4(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd 12(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, 8(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + add $4, un + movd %mm6, 12(rp) + lea 16(up), up + js L(lm0) + + psrlq $32, %mm6 + movd %mm6, 16(rp) + + inc n +C jz L(done) + lea -8(up), up + lea 8(rp), rp + jmp L(ol3) + +C ================================================================ + ALIGN(16) +L(lm1): movd -12(up), %mm0 + pmuludq %mm7, %mm0 + psrlq $32, %mm6 + lea 16(rp), rp + paddq %mm0, %mm6 + movd -8(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -12(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd -4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -8(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd (up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -4(rp) + psrlq $32, %mm6 +L(of1): paddq %mm0, %mm6 + add $4, un + movd %mm6, (rp) + lea 16(up), up + js L(lm1) + + psrlq $32, %mm6 + movd %mm6, 4(rp) + + inc n + jz L(done) C goes away when we add special n=2 code + lea -20(up), up + lea -4(rp), rp + jmp L(ol0) + +C ================================================================ + ALIGN(16) +L(lm2): movd -8(up), %mm0 + pmuludq %mm7, %mm0 + psrlq $32, %mm6 + lea 16(rp), rp + paddq %mm0, %mm6 + movd -4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -8(rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + movd (up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, -4(rp) + psrlq $32, %mm6 +L(of2): paddq %mm0, %mm6 + movd 4(up), %mm0 + pmuludq %mm7, %mm0 + movd %mm6, (rp) + psrlq $32, %mm6 + paddq %mm0, %mm6 + add $4, un + movd %mm6, 4(rp) + lea 16(up), up + js L(lm2) + + psrlq $32, %mm6 + movd %mm6, 8(rp) + + inc n +C jz L(done) + lea -16(up), up +C lea (rp), rp +C jmp L(ol1) + +C ================================================================ + +L(ol1): lea 4(up,n,4), up + movd (up), %mm7 C read next U invariant limb + lea 8(rp,n,4), rp + mov n, un + + movd 4(up), %mm1 + pmuludq %mm7, %mm1 + sar $2, un + movd %mm1, %ebx + inc un + jz L(re1) + + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + xor %edx, %edx C zero edx and CF + jmp L(a1) + +L(la1): adc $0, %edx + add %ebx, 12(rp) + movd %mm0, %eax + pmuludq %mm7, %mm1 + lea 16(rp), rp + psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + adc $0, %edx + add %eax, (rp) +L(a1): psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %edx + movd %mm0, %eax + movd 12(up), %mm1 + pmuludq %mm7, %mm1 + adc $0, %edx + add %ebx, 4(rp) + psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + lea 16(up), up + movd (up), %mm0 + adc $0, %edx + add %eax, 8(rp) + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %edx + pmuludq %mm7, %mm0 + inc un + movd 4(up), %mm1 + jnz L(la1) + + adc un, %edx C un is zero here + add %ebx, 12(rp) + movd %mm0, %eax + pmuludq %mm7, %mm1 + lea 16(rp), rp + psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + adc un, %edx + add %eax, (rp) + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %eax + adc un, %eax + add %ebx, 4(rp) + adc un, %eax + mov %eax, 8(rp) + + inc n + +C ================================================================ + +L(ol0): lea (up,n,4), up + movd 4(up), %mm7 C read next U invariant limb + lea 4(rp,n,4), rp + mov n, un + + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + sar $2, un + movd 12(up), %mm1 + movd %mm0, %eax + pmuludq %mm7, %mm1 + xor %edx, %edx C zero edx and CF + jmp L(a0) + +L(la0): adc $0, %edx + add %ebx, 12(rp) + movd %mm0, %eax + pmuludq %mm7, %mm1 + lea 16(rp), rp + psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + adc $0, %edx + add %eax, (rp) + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %edx + movd %mm0, %eax + movd 12(up), %mm1 + pmuludq %mm7, %mm1 + adc $0, %edx + add %ebx, 4(rp) +L(a0): psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + lea 16(up), up + movd (up), %mm0 + adc $0, %edx + add %eax, 8(rp) + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %edx + pmuludq %mm7, %mm0 + inc un + movd 4(up), %mm1 + jnz L(la0) + + adc un, %edx C un is zero here + add %ebx, 12(rp) + movd %mm0, %eax + pmuludq %mm7, %mm1 + lea 16(rp), rp + psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + adc un, %edx + add %eax, (rp) + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %eax + adc un, %eax + add %ebx, 4(rp) + adc un, %eax + mov %eax, 8(rp) + + inc n + +C ================================================================ + +L(ol3): lea 12(up,n,4), up + movd -8(up), %mm7 C read next U invariant limb + lea (rp,n,4), rp C put rp back + mov n, un + + movd -4(up), %mm1 + pmuludq %mm7, %mm1 + sar $2, un + movd %mm1, %ebx + movd (up), %mm0 + xor %edx, %edx C zero edx and CF + jmp L(a3) + +L(la3): adc $0, %edx + add %ebx, 12(rp) + movd %mm0, %eax + pmuludq %mm7, %mm1 + lea 16(rp), rp + psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + adc $0, %edx + add %eax, (rp) + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %edx + movd %mm0, %eax + movd 12(up), %mm1 + pmuludq %mm7, %mm1 + adc $0, %edx + add %ebx, 4(rp) + psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + lea 16(up), up + movd (up), %mm0 + adc $0, %edx + add %eax, 8(rp) +L(a3): psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %edx + pmuludq %mm7, %mm0 + inc un + movd 4(up), %mm1 + jnz L(la3) + + adc un, %edx C un is zero here + add %ebx, 12(rp) + movd %mm0, %eax + pmuludq %mm7, %mm1 + lea 16(rp), rp + psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + adc un, %edx + add %eax, (rp) + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %eax + adc un, %eax + add %ebx, 4(rp) + adc un, %eax + mov %eax, 8(rp) + + inc n + +C ================================================================ + +L(ol2): lea 8(up,n,4), up + movd -4(up), %mm7 C read next U invariant limb + lea 12(rp,n,4), rp + mov n, un + + movd (up), %mm0 + pmuludq %mm7, %mm0 + xor %edx, %edx + sar $2, un + movd 4(up), %mm1 + test un, un C clear carry + movd %mm0, %eax + pmuludq %mm7, %mm1 + inc un + jnz L(a2) + jmp L(re2) + +L(la2): adc $0, %edx + add %ebx, 12(rp) + movd %mm0, %eax + pmuludq %mm7, %mm1 + lea 16(rp), rp +L(a2): psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + movd 8(up), %mm0 + pmuludq %mm7, %mm0 + adc $0, %edx + add %eax, (rp) + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %edx + movd %mm0, %eax + movd 12(up), %mm1 + pmuludq %mm7, %mm1 + adc $0, %edx + add %ebx, 4(rp) + psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + lea 16(up), up + movd (up), %mm0 + adc $0, %edx + add %eax, 8(rp) + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %edx + pmuludq %mm7, %mm0 + inc un + movd 4(up), %mm1 + jnz L(la2) + + adc un, %edx C un is zero here + add %ebx, 12(rp) + movd %mm0, %eax + pmuludq %mm7, %mm1 + lea 16(rp), rp + psrlq $32, %mm0 + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + adc un, %edx + add %eax, (rp) + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %eax + adc un, %eax + add %ebx, 4(rp) + adc un, %eax + mov %eax, 8(rp) + + inc n + jmp L(ol1) + +C ================================================================ +L(re2): psrlq $32, %mm0 + movd (up), %mm7 C read next U invariant limb + adc %edx, %eax + movd %mm0, %edx + movd %mm1, %ebx + adc un, %edx + add %eax, (rp) + lea 4(rp), rp + psrlq $32, %mm1 + adc %edx, %ebx + movd %mm1, %eax + movd 4(up), %mm1 + adc un, %eax + add %ebx, (rp) + pmuludq %mm7, %mm1 + adc un, %eax + mov %eax, 4(rp) + movd %mm1, %ebx + +L(re1): psrlq $32, %mm1 + add %ebx, 4(rp) + movd %mm1, %eax + adc un, %eax + xor n, n C make n zeroness assumption below true + mov %eax, 8(rp) + +L(done): C n is zero here + mov 24(%esp), up + mov 28(%esp), %eax + + movd (up), %mm0 + inc %eax + pmuludq %mm0, %mm0 + lea 4(up), up + mov 20(%esp), rp + shr %eax + movd %mm0, (rp) + psrlq $32, %mm0 + lea -12(rp), rp + mov %eax, 28(%esp) + jnc L(odd) + + movd %mm0, %ebp + movd (up), %mm0 + lea 8(rp), rp + pmuludq %mm0, %mm0 + lea -4(up), up + add 8(rp), %ebp + movd %mm0, %edx + adc 12(rp), %edx + rcr n + jmp L(ent) + +C ALIGN(16) C alignment seems irrelevant +L(top): movd (up), %mm1 + adc n, n + movd %mm0, %eax + pmuludq %mm1, %mm1 + movd 4(up), %mm0 + adc (rp), %eax + movd %mm1, %ebx + pmuludq %mm0, %mm0 + psrlq $32, %mm1 + adc 4(rp), %ebx + movd %mm1, %ebp + movd %mm0, %edx + adc 8(rp), %ebp + adc 12(rp), %edx + rcr n C FIXME: isn't this awfully slow on atom??? + adc %eax, (rp) + adc %ebx, 4(rp) +L(ent): lea 8(up), up + adc %ebp, 8(rp) + psrlq $32, %mm0 + adc %edx, 12(rp) +L(odd): decl 28(%esp) + lea 16(rp), rp + jnz L(top) + +L(end): adc n, n + movd %mm0, %eax + adc n, %eax + mov %eax, (rp) + +L(rtn): emms + pop %ebp + pop %ebx + pop %esi + pop %edi + ret + +L(one): pmuludq %mm7, %mm7 + movq %mm7, -4(rp) + emms + pop %esi + pop %edi + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/atom/sublsh1_n.asm b/gmp-6.3.0/mpn/x86/atom/sublsh1_n.asm new file mode 100644 index 0000000..d3e7e5b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sublsh1_n.asm @@ -0,0 +1,34 @@ +dnl Intel Atom mpn_sublsh1_n -- rp[] = up[] - (vp[] << 1) + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_sublsh1_n_ip1) +include_mpn(`x86/k7/sublsh1_n.asm') diff --git a/gmp-6.3.0/mpn/x86/atom/sublsh2_n.asm b/gmp-6.3.0/mpn/x86/atom/sublsh2_n.asm new file mode 100644 index 0000000..79405cf --- /dev/null +++ b/gmp-6.3.0/mpn/x86/atom/sublsh2_n.asm @@ -0,0 +1,57 @@ +dnl Intel Atom mpn_addlsh2_n/mpn_sublsh2_n -- rp[] = up[] +- (vp[] << 2). + +dnl Contributed to the GNU project by Marco Bodrato. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +define(LSH, 2) +define(RSH, 30) + +ifdef(`OPERATION_addlsh2_n', ` + define(M4_inst, adcl) + define(M4_opp, subl) + define(M4_function, mpn_addlsh2_n) + define(M4_function_c, mpn_addlsh2_nc) + define(M4_ip_function_c, mpn_addlsh2_nc_ip1) + define(M4_ip_function, mpn_addlsh2_n_ip1) +',`ifdef(`OPERATION_sublsh2_n', ` + define(M4_inst, sbbl) + define(M4_opp, addl) + define(M4_function, mpn_sublsh2_n) + define(M4_function_c, mpn_sublsh2_nc) + define(M4_ip_function_c, mpn_sublsh2_nc_ip1) + define(M4_ip_function, mpn_sublsh2_n_ip1) +',`m4_error(`Need OPERATION_addlsh2_n or OPERATION_sublsh2_n +')')') + +MULFUNC_PROLOGUE(mpn_sublsh2_n mpn_sublsh2_nc mpn_sublsh2_n_ip1 mpn_sublsh2_nc_ip1) + +include_mpn(`x86/atom/aorslshC_n.asm') diff --git a/gmp-6.3.0/mpn/x86/bd1/gmp-mparam.h b/gmp-6.3.0/mpn/x86/bd1/gmp-mparam.h new file mode 100644 index 0000000..254cfea --- /dev/null +++ b/gmp-6.3.0/mpn/x86/bd1/gmp-mparam.h @@ -0,0 +1,211 @@ +/* AMD bd1 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 3600-3800 MHz Bulldozer Zambezi */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-27, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 0 /* always */ +#define MOD_1_UNNORM_THRESHOLD 0 /* always */ +#define MOD_1N_TO_MOD_1_1_THRESHOLD 7 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 4 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 15 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 13 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 59.59% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 5 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 27 + +#define DIV_1_VS_MUL_1_PERCENT 245 + +#define MUL_TOOM22_THRESHOLD 32 +#define MUL_TOOM33_THRESHOLD 89 +#define MUL_TOOM44_THRESHOLD 154 +#define MUL_TOOM6H_THRESHOLD 230 +#define MUL_TOOM8H_THRESHOLD 351 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 89 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 110 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 101 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 111 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 130 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 46 +#define SQR_TOOM3_THRESHOLD 87 +#define SQR_TOOM4_THRESHOLD 216 +#define SQR_TOOM6_THRESHOLD 294 +#define SQR_TOOM8_THRESHOLD 442 + +#define MULMID_TOOM42_THRESHOLD 50 + +#define MULMOD_BNM1_THRESHOLD 22 +#define SQRMOD_BNM1_THRESHOLD 26 + +#define MUL_FFT_MODF_THRESHOLD 636 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 636, 5}, { 28, 6}, { 15, 5}, { 31, 6}, \ + { 28, 7}, { 15, 6}, { 33, 7}, { 17, 6}, \ + { 35, 7}, { 19, 6}, { 39, 7}, { 23, 6}, \ + { 47, 7}, { 29, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 7}, { 55, 8}, { 31, 7}, { 63, 8}, \ + { 43, 9}, { 23, 8}, { 55, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 79, 9}, { 47, 8}, \ + { 95, 9}, { 55,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95, 9}, \ + { 191,11}, { 63, 7}, { 1023, 8}, { 543,11}, \ + { 95,10}, { 191,12}, { 63,11}, { 127,10}, \ + { 255, 9}, { 511,10}, { 271,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335,11}, { 191,10}, \ + { 399,11}, { 223,12}, { 127,11}, { 255,10}, \ + { 543,11}, { 287,10}, { 607,11}, { 319,10}, \ + { 639,12}, { 191,11}, { 383,10}, { 799,11}, \ + { 415,13}, { 127,12}, { 255,11}, { 543,10}, \ + { 1087,11}, { 607,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,12}, { 383,11}, { 799,10}, \ + { 1599,11}, { 863,12}, { 447,11}, { 895,13}, \ + { 255,12}, { 511,11}, { 1087,12}, { 575,11}, \ + { 1215,12}, { 639,11}, { 1343,12}, { 703,11}, \ + { 1471,13}, { 383,12}, { 767,11}, { 1599,12}, \ + { 831,11}, { 1727,12}, { 895,14}, { 255,13}, \ + { 511,12}, { 1087,11}, { 2239,10}, { 4479,12}, \ + { 1215,13}, { 639,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1727,11}, { 3455,13}, { 895,12}, \ + { 1919,14}, { 511,13}, { 1023,12}, { 2239,11}, \ + { 4479,13}, { 1151,12}, { 2495,11}, { 4991,13}, \ + { 1279,12}, { 2623,13}, { 1407,12}, { 2943,14}, \ + { 767,13}, { 1663,12}, { 3455,13}, { 1919,15}, \ + { 511,14}, { 1023,13}, { 2175,12}, { 4479,13}, \ + { 2431,12}, { 4991,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3967,12}, { 7935,15}, { 1023,14}, { 2047,13}, \ + { 4479,14}, { 2303,13}, { 4991,12}, { 9983,14}, \ + { 2815,13}, { 5887,15}, { 1535,14}, { 3327,13}, \ + { 6911,14}, { 3839,13}, { 7935,16} } +#define MUL_FFT_TABLE3_SIZE 159 +#define MUL_FFT_THRESHOLD 6784 + +#define SQR_FFT_MODF_THRESHOLD 565 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 565, 5}, { 29, 6}, { 15, 5}, { 32, 6}, \ + { 17, 5}, { 35, 6}, { 29, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 23, 6}, { 47, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 49, 8}, \ + { 27, 7}, { 55, 8}, { 43, 9}, { 23, 8}, \ + { 55, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 79, 9}, { 47, 8}, { 95, 9}, { 55,10}, \ + { 31, 9}, { 79,10}, { 47, 9}, { 95,11}, \ + { 31,10}, { 63, 9}, { 135,10}, { 79, 9}, \ + { 159,10}, { 95,11}, { 63,10}, { 159,11}, \ + { 127,10}, { 255, 9}, { 511,10}, { 271, 9}, \ + { 543,11}, { 159,10}, { 319, 9}, { 639,10}, \ + { 335, 9}, { 671,11}, { 191,10}, { 415,11}, \ + { 223,12}, { 127,11}, { 255,10}, { 543,11}, \ + { 287,10}, { 607,11}, { 319,10}, { 671,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,10}, \ + { 831,13}, { 127,12}, { 255,11}, { 543,10}, \ + { 1087,11}, { 607,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,12}, { 383,11}, { 863,12}, \ + { 447,11}, { 959,13}, { 255,12}, { 511,11}, \ + { 1087,12}, { 575,11}, { 1215,12}, { 639,11}, \ + { 1343,12}, { 703,13}, { 383,12}, { 767,11}, \ + { 1535,12}, { 831,11}, { 1727,12}, { 895,11}, \ + { 1791,12}, { 959,14}, { 255,13}, { 511,12}, \ + { 1087,11}, { 2239,10}, { 4479,12}, { 1215,13}, \ + { 639,12}, { 1471,11}, { 2943,13}, { 767,12}, \ + { 1727,13}, { 895,12}, { 1919,14}, { 511,13}, \ + { 1023,12}, { 2239,11}, { 4479,13}, { 1151,12}, \ + { 2495,11}, { 4991,13}, { 1279,12}, { 2623,13}, \ + { 1407,12}, { 2943,14}, { 767,13}, { 1663,12}, \ + { 3455,13}, { 1919,15}, { 511,14}, { 1023,13}, \ + { 2175,12}, { 4479,13}, { 2431,12}, { 4991,14}, \ + { 1279,13}, { 2943,12}, { 5887,14}, { 1535,13}, \ + { 3455,14}, { 1791,13}, { 3967,15}, { 1023,14}, \ + { 2047,13}, { 4479,14}, { 2303,13}, { 4991,12}, \ + { 9983,14}, { 2815,13}, { 5887,15}, { 1535,14}, \ + { 3327,13}, { 6783,14}, { 3839,13}, { 7679,16} } +#define SQR_FFT_TABLE3_SIZE 152 +#define SQR_FFT_THRESHOLD 5760 + +#define MULLO_BASECASE_THRESHOLD 3 +#define MULLO_DC_THRESHOLD 31 +#define MULLO_MUL_N_THRESHOLD 13463 +#define SQRLO_BASECASE_THRESHOLD 0 /* always */ +#define SQRLO_DC_THRESHOLD 33 +#define SQRLO_SQR_THRESHOLD 11278 + +#define DC_DIV_QR_THRESHOLD 52 +#define DC_DIVAPPR_Q_THRESHOLD 198 +#define DC_BDIV_QR_THRESHOLD 48 +#define DC_BDIV_Q_THRESHOLD 126 + +#define INV_MULMOD_BNM1_THRESHOLD 82 +#define INV_NEWTON_THRESHOLD 212 +#define INV_APPR_THRESHOLD 202 + +#define BINV_NEWTON_THRESHOLD 238 +#define REDC_1_TO_REDC_N_THRESHOLD 55 + +#define MU_DIV_QR_THRESHOLD 1652 +#define MU_DIVAPPR_Q_THRESHOLD 1528 +#define MUPI_DIV_QR_THRESHOLD 110 +#define MU_BDIV_QR_THRESHOLD 1442 +#define MU_BDIV_Q_THRESHOLD 1528 + +#define POWM_SEC_TABLE 1,20,96,386,1221,2698 + +#define GET_STR_DC_THRESHOLD 11 +#define GET_STR_PRECOMPUTE_THRESHOLD 21 +#define SET_STR_DC_THRESHOLD 100 +#define SET_STR_PRECOMPUTE_THRESHOLD 762 + +#define FAC_DSC_THRESHOLD 118 +#define FAC_ODD_THRESHOLD 34 + +#define MATRIX22_STRASSEN_THRESHOLD 16 +#define HGCD2_DIV1_METHOD 4 /* 1.22% faster than 3 */ +#define HGCD_THRESHOLD 67 +#define HGCD_APPR_THRESHOLD 150 +#define HGCD_REDUCE_THRESHOLD 3389 +#define GCD_DC_THRESHOLD 483 +#define GCDEXT_DC_THRESHOLD 345 +#define JACOBI_BASE_METHOD 4 /* 5.07% faster than 1 */ + +/* Tuneup completed successfully, took 65358 seconds */ diff --git a/gmp-6.3.0/mpn/x86/bd2/gmp-mparam.h b/gmp-6.3.0/mpn/x86/bd2/gmp-mparam.h new file mode 100644 index 0000000..6893da7 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/bd2/gmp-mparam.h @@ -0,0 +1,214 @@ +/* AMD bd2 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 4000-4200 MHz Piledriver Vishera */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-23, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 3 +#define MOD_1_UNNORM_THRESHOLD 4 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 6 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 4 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 18 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 12 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 40.87% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 5 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 24 + +#define DIV_1_VS_MUL_1_PERCENT 254 + +#define MUL_TOOM22_THRESHOLD 32 +#define MUL_TOOM33_THRESHOLD 73 +#define MUL_TOOM44_THRESHOLD 151 +#define MUL_TOOM6H_THRESHOLD 222 +#define MUL_TOOM8H_THRESHOLD 351 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 85 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 110 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 100 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 110 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 130 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 44 +#define SQR_TOOM3_THRESHOLD 93 +#define SQR_TOOM4_THRESHOLD 212 +#define SQR_TOOM6_THRESHOLD 318 +#define SQR_TOOM8_THRESHOLD 466 + +#define MULMID_TOOM42_THRESHOLD 66 + +#define MULMOD_BNM1_THRESHOLD 20 +#define SQRMOD_BNM1_THRESHOLD 23 + +#define MUL_FFT_MODF_THRESHOLD 595 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 595, 5}, { 27, 6}, { 29, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 23, 6}, { 47, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 49, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 55, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 83, 9}, { 47, 8}, \ + { 95, 9}, { 55,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95, 9}, \ + { 191,11}, { 63,10}, { 143, 7}, { 1215, 9}, \ + { 319, 8}, { 639, 9}, { 335, 8}, { 671, 9}, \ + { 351,10}, { 191,12}, { 63,11}, { 127,10}, \ + { 271,11}, { 159,10}, { 319, 9}, { 639,10}, \ + { 335,11}, { 191,10}, { 399,11}, { 223,12}, \ + { 127,11}, { 255,10}, { 543,11}, { 287,10}, \ + { 607,11}, { 319,10}, { 671,12}, { 191,11}, \ + { 383,10}, { 799,11}, { 415,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 735,10}, \ + { 1471,12}, { 383,11}, { 799,10}, { 1599,11}, \ + { 863,12}, { 447,11}, { 895,13}, { 255,12}, \ + { 511,11}, { 1087,12}, { 575,11}, { 1215,12}, \ + { 639,11}, { 1343,12}, { 703,11}, { 1471,13}, \ + { 383,12}, { 767,11}, { 1599,12}, { 831,11}, \ + { 1727,12}, { 895,14}, { 255,13}, { 511,12}, \ + { 1087,11}, { 2239,12}, { 1215,13}, { 639,12}, \ + { 1471,11}, { 2943,13}, { 767,12}, { 1727,13}, \ + { 895,12}, { 1919,14}, { 511,13}, { 1023,12}, \ + { 2239,13}, { 1151,12}, { 2431,13}, { 1279,12}, \ + { 2623,13}, { 1407,12}, { 2943,14}, { 767,13}, \ + { 1535,12}, { 3135,13}, { 1663,12}, { 3455,13}, \ + { 1919,15}, { 511,14}, { 1023,13}, { 2175,12}, \ + { 4479,13}, { 2431,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3967,12}, { 7935,11}, { 15871,15}, { 1023,14}, \ + { 2047,13}, { 4479,14}, { 2303,13}, { 4991,12}, \ + { 9983,14}, { 2815,13}, { 5887,15}, { 1535,14}, \ + { 3839,13}, { 7935,12}, { 15871,16} } +#define MUL_FFT_TABLE3_SIZE 155 +#define MUL_FFT_THRESHOLD 6784 + +#define SQR_FFT_MODF_THRESHOLD 555 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 555, 5}, { 28, 6}, { 15, 5}, { 31, 6}, \ + { 16, 5}, { 33, 6}, { 29, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 36, 7}, { 19, 6}, \ + { 39, 7}, { 23, 6}, { 47, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 43, 8}, \ + { 23, 7}, { 49, 8}, { 31, 7}, { 63, 8}, \ + { 43, 9}, { 23, 8}, { 55, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 79, 9}, { 47, 8}, \ + { 95, 9}, { 55,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95,11}, \ + { 63,10}, { 143, 9}, { 287,10}, { 159,11}, \ + { 95,10}, { 191, 6}, { 3071, 5}, { 6399, 6}, \ + { 3455, 7}, { 1791, 8}, { 959,10}, { 255, 9}, \ + { 511,10}, { 271,11}, { 159,10}, { 319, 9}, \ + { 639,10}, { 335, 9}, { 671,10}, { 351,11}, \ + { 191,10}, { 399, 9}, { 799,10}, { 415,11}, \ + { 223,12}, { 127,11}, { 255,10}, { 543,11}, \ + { 287,10}, { 607,11}, { 319,10}, { 671,11}, \ + { 351,12}, { 191,11}, { 383,10}, { 799,11}, \ + { 415,13}, { 127,12}, { 255,11}, { 543,10}, \ + { 1087,11}, { 607,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,12}, { 383,11}, { 799,10}, \ + { 1599,11}, { 863,12}, { 447,11}, { 927,13}, \ + { 255,12}, { 511,11}, { 1055,10}, { 2111,11}, \ + { 1087,12}, { 575,11}, { 1215,12}, { 639,11}, \ + { 1343,12}, { 703,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1727,10}, { 3455,12}, \ + { 959,14}, { 255,13}, { 511,12}, { 1023,11}, \ + { 2111,12}, { 1087,11}, { 2239,10}, { 4479,12}, \ + { 1215,13}, { 639,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1727,11}, { 3455,13}, { 895,12}, \ + { 1855,14}, { 511,13}, { 1023,12}, { 2239,13}, \ + { 1151,12}, { 2495,13}, { 1279,12}, { 2623,13}, \ + { 1407,12}, { 2943,14}, { 767,13}, { 1663,12}, \ + { 3455,13}, { 1791,15}, { 511,14}, { 1023,13}, \ + { 2175,12}, { 4479,13}, { 2431,14}, { 1279,13}, \ + { 2943,12}, { 5887,14}, { 1535,13}, { 3455,14}, \ + { 1791,13}, { 3967,12}, { 7935,15}, { 1023,14}, \ + { 2047,13}, { 4479,14}, { 2303,13}, { 4991,12}, \ + { 9983,14}, { 2815,13}, { 5887,15}, { 1535,14}, \ + { 3839,13}, { 7935,16} } +#define SQR_FFT_TABLE3_SIZE 166 +#define SQR_FFT_THRESHOLD 5760 + +#define MULLO_BASECASE_THRESHOLD 3 +#define MULLO_DC_THRESHOLD 34 +#define MULLO_MUL_N_THRESHOLD 13463 +#define SQRLO_BASECASE_THRESHOLD 8 +#define SQRLO_DC_THRESHOLD 43 +#define SQRLO_SQR_THRESHOLD 11278 + +#define DC_DIV_QR_THRESHOLD 75 +#define DC_DIVAPPR_Q_THRESHOLD 200 +#define DC_BDIV_QR_THRESHOLD 71 +#define DC_BDIV_Q_THRESHOLD 119 + +#define INV_MULMOD_BNM1_THRESHOLD 74 +#define INV_NEWTON_THRESHOLD 266 +#define INV_APPR_THRESHOLD 214 + +#define BINV_NEWTON_THRESHOLD 278 +#define REDC_1_TO_REDC_N_THRESHOLD 71 + +#define MU_DIV_QR_THRESHOLD 1652 +#define MU_DIVAPPR_Q_THRESHOLD 1589 +#define MUPI_DIV_QR_THRESHOLD 122 +#define MU_BDIV_QR_THRESHOLD 1442 +#define MU_BDIV_Q_THRESHOLD 1597 + +#define POWM_SEC_TABLE 1,22,96,289,1259 + +#define GET_STR_DC_THRESHOLD 11 +#define GET_STR_PRECOMPUTE_THRESHOLD 20 +#define SET_STR_DC_THRESHOLD 173 +#define SET_STR_PRECOMPUTE_THRESHOLD 454 + +#define FAC_DSC_THRESHOLD 90 +#define FAC_ODD_THRESHOLD 34 + +#define MATRIX22_STRASSEN_THRESHOLD 19 +#define HGCD2_DIV1_METHOD 1 /* 5.80% faster than 3 */ +#define HGCD_THRESHOLD 74 +#define HGCD_APPR_THRESHOLD 50 +#define HGCD_REDUCE_THRESHOLD 3389 +#define GCD_DC_THRESHOLD 456 +#define GCDEXT_DC_THRESHOLD 345 +#define JACOBI_BASE_METHOD 4 /* 17.07% faster than 1 */ + +/* Tuneup completed successfully, took 53914 seconds */ diff --git a/gmp-6.3.0/mpn/x86/bd4/gmp-mparam.h b/gmp-6.3.0/mpn/x86/bd4/gmp-mparam.h new file mode 100644 index 0000000..6c20d0f --- /dev/null +++ b/gmp-6.3.0/mpn/x86/bd4/gmp-mparam.h @@ -0,0 +1,225 @@ +/* AMD bd4 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 3800-4200 MHz Excavator/Bristol Ridge */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-23, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 0 /* always */ +#define MOD_1_UNNORM_THRESHOLD 0 /* always */ +#define MOD_1N_TO_MOD_1_1_THRESHOLD 8 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 6 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 27 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 50 +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 13 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 28.45% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 4 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD 13 +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 28 + +#define DIV_1_VS_MUL_1_PERCENT 314 + +#define MUL_TOOM22_THRESHOLD 32 +#define MUL_TOOM33_THRESHOLD 73 +#define MUL_TOOM44_THRESHOLD 166 +#define MUL_TOOM6H_THRESHOLD 270 +#define MUL_TOOM8H_THRESHOLD 357 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 69 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 114 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 103 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 121 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 154 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 42 +#define SQR_TOOM3_THRESHOLD 89 +#define SQR_TOOM4_THRESHOLD 208 +#define SQR_TOOM6_THRESHOLD 306 +#define SQR_TOOM8_THRESHOLD 454 + +#define MULMID_TOOM42_THRESHOLD 68 + +#define MULMOD_BNM1_THRESHOLD 19 +#define SQRMOD_BNM1_THRESHOLD 18 + +#define MUL_FFT_MODF_THRESHOLD 570 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 570, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 25, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \ + { 32, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 51,10}, { 15, 9}, \ + { 31, 8}, { 67, 9}, { 39, 8}, { 79, 9}, \ + { 47, 8}, { 95,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95, 9}, \ + { 191,11}, { 63,10}, { 143, 6}, { 2303, 5}, \ + { 4735, 4}, { 9471, 5}, { 4863, 7}, { 1279, 9}, \ + { 335, 8}, { 671, 9}, { 351, 8}, { 703,10}, \ + { 191,12}, { 63,11}, { 127,10}, { 255, 9}, \ + { 511,10}, { 271, 9}, { 543,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335, 9}, { 671, 8}, \ + { 1343,10}, { 351, 9}, { 703,10}, { 367, 9}, \ + { 735,11}, { 191,10}, { 383, 9}, { 767,10}, \ + { 399, 9}, { 799, 8}, { 1599,10}, { 415,11}, \ + { 223,12}, { 127,11}, { 255,10}, { 543, 9}, \ + { 1087,11}, { 287,10}, { 607, 9}, { 1215,11}, \ + { 319,10}, { 671, 9}, { 1343,11}, { 351,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,10}, \ + { 863,13}, { 127,12}, { 255,11}, { 543,10}, \ + { 1087,11}, { 607,10}, { 1215, 9}, { 2431,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 735,10}, \ + { 1471, 9}, { 2943,12}, { 383,11}, { 799,10}, \ + { 1599,11}, { 863,10}, { 1727,12}, { 447,11}, \ + { 959,10}, { 1919,13}, { 255,12}, { 511,11}, \ + { 1087,12}, { 575,11}, { 1215,10}, { 2431,12}, \ + { 639,11}, { 1343,12}, { 703,11}, { 1471,10}, \ + { 2943,13}, { 383,12}, { 767,11}, { 1599,12}, \ + { 831,11}, { 1727,10}, { 3455,12}, { 959,11}, \ + { 1919,10}, { 3839,13}, { 511,12}, { 1087,11}, \ + { 2239,12}, { 1215,11}, { 2431,13}, { 639,12}, \ + { 1471,11}, { 2943,10}, { 5887,13}, { 767,12}, \ + { 1727,11}, { 3455,13}, { 895,12}, { 1919,11}, \ + { 3839,14}, { 511,13}, { 1023,12}, { 2239,13}, \ + { 1151,12}, { 2431,13}, { 1279,12}, { 2559,13}, \ + { 1407,12}, { 2943,11}, { 5887,14}, { 767,13}, \ + { 1663,12}, { 3455,13}, { 1919,12}, { 3839,15}, \ + { 511,14}, { 1023,13}, { 2175,12}, { 4479,13}, \ + { 2431,14}, { 1279,13}, { 2943,12}, { 5887,14}, \ + { 1535,13}, { 3455,14}, { 1791,13}, { 3967,12}, \ + { 7935,15}, { 1023,14}, { 2047,13}, { 4479,14}, \ + { 2303,13}, { 4991,12}, { 9983,14}, { 2815,13}, \ + { 5887,15}, { 1535,14}, { 3839,13}, { 7935,16} } +#define MUL_FFT_TABLE3_SIZE 192 +#define MUL_FFT_THRESHOLD 5760 + +#define SQR_FFT_MODF_THRESHOLD 476 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 476, 5}, { 28, 6}, { 15, 5}, { 31, 6}, \ + { 16, 5}, { 33, 6}, { 29, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 36, 7}, { 19, 6}, \ + { 39, 7}, { 29, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 39, 9}, { 23, 8}, \ + { 51, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 79, 9}, { 47, 8}, { 95,10}, { 31, 9}, \ + { 79,10}, { 47, 9}, { 95,11}, { 31,10}, \ + { 63, 9}, { 135,10}, { 95, 9}, { 191,10}, \ + { 111,11}, { 63,10}, { 127, 9}, { 255,10}, \ + { 143, 9}, { 287, 8}, { 575,10}, { 159,11}, \ + { 95,10}, { 191,12}, { 63,10}, { 255, 9}, \ + { 511,10}, { 271, 9}, { 543,10}, { 287, 9}, \ + { 575,11}, { 159,10}, { 319, 9}, { 639,10}, \ + { 335, 9}, { 671,10}, { 351, 9}, { 735,11}, \ + { 191,10}, { 383, 9}, { 767,10}, { 399, 9}, \ + { 799,10}, { 415, 9}, { 863,12}, { 127,11}, \ + { 255,10}, { 511, 9}, { 1023,10}, { 543,11}, \ + { 287,10}, { 607, 9}, { 1215,11}, { 319,10}, \ + { 671, 9}, { 1343,11}, { 351,10}, { 735,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,10}, \ + { 863,13}, { 127,12}, { 255,11}, { 511,10}, \ + { 1055,11}, { 543,10}, { 1087,11}, { 607,10}, \ + { 1215,12}, { 319,11}, { 671,10}, { 1343,11}, \ + { 735,10}, { 1471,12}, { 383,11}, { 799,10}, \ + { 1599,11}, { 863,10}, { 1727,12}, { 447,11}, \ + { 959,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1471,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1727,12}, { 959,11}, \ + { 1919,14}, { 255,13}, { 511,12}, { 1023,11}, \ + { 2047,12}, { 1087,11}, { 2239,12}, { 1215,11}, \ + { 2431,13}, { 639,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1727,13}, { 895,12}, { 1983,14}, \ + { 511,13}, { 1023,12}, { 2239,13}, { 1151,12}, \ + { 2431,13}, { 1279,12}, { 2559,13}, { 1407,12}, \ + { 2943,14}, { 767,13}, { 1663,12}, { 3455,13}, \ + { 1919,12}, { 3839,15}, { 511,14}, { 1023,13}, \ + { 2175,12}, { 4479,13}, { 2431,14}, { 1279,13}, \ + { 2943,12}, { 5887,14}, { 1535,13}, { 3455,14}, \ + { 1791,13}, { 3967,15}, { 1023,14}, { 2047,13}, \ + { 4479,14}, { 2303,13}, { 4991,12}, { 9983,14}, \ + { 2815,13}, { 5887,15}, { 1535,14}, { 3839,16} } +#define SQR_FFT_TABLE3_SIZE 176 +#define SQR_FFT_THRESHOLD 4736 + +#define MULLO_BASECASE_THRESHOLD 3 +#define MULLO_DC_THRESHOLD 54 +#define MULLO_MUL_N_THRESHOLD 10950 +#define SQRLO_BASECASE_THRESHOLD 10 +#define SQRLO_DC_THRESHOLD 77 +#define SQRLO_SQR_THRESHOLD 9449 + +#define DC_DIV_QR_THRESHOLD 84 +#define DC_DIVAPPR_Q_THRESHOLD 252 +#define DC_BDIV_QR_THRESHOLD 79 +#define DC_BDIV_Q_THRESHOLD 80 + +#define INV_MULMOD_BNM1_THRESHOLD 71 +#define INV_NEWTON_THRESHOLD 254 +#define INV_APPR_THRESHOLD 266 + +#define BINV_NEWTON_THRESHOLD 294 +#define REDC_1_TO_REDC_N_THRESHOLD 79 + +#define MU_DIV_QR_THRESHOLD 1652 +#define MU_DIVAPPR_Q_THRESHOLD 1528 +#define MUPI_DIV_QR_THRESHOLD 122 +#define MU_BDIV_QR_THRESHOLD 1387 +#define MU_BDIV_Q_THRESHOLD 1528 + +#define POWM_SEC_TABLE 1,16,96,480,960 + +#define GET_STR_DC_THRESHOLD 12 +#define GET_STR_PRECOMPUTE_THRESHOLD 19 +#define SET_STR_DC_THRESHOLD 264 +#define SET_STR_PRECOMPUTE_THRESHOLD 542 + +#define FAC_DSC_THRESHOLD 91 +#define FAC_ODD_THRESHOLD 29 + +#define MATRIX22_STRASSEN_THRESHOLD 19 +#define HGCD2_DIV1_METHOD 1 /* 9.73% faster than 3 */ +#define HGCD_THRESHOLD 55 +#define HGCD_APPR_THRESHOLD 50 +#define HGCD_REDUCE_THRESHOLD 3389 +#define GCD_DC_THRESHOLD 562 +#define GCDEXT_DC_THRESHOLD 416 +#define JACOBI_BASE_METHOD 4 /* 16.50% faster than 1 */ + +/* Tuneup completed successfully, took 49179 seconds */ diff --git a/gmp-6.3.0/mpn/x86/bdiv_dbm1c.asm b/gmp-6.3.0/mpn/x86/bdiv_dbm1c.asm new file mode 100644 index 0000000..0288c47 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/bdiv_dbm1c.asm @@ -0,0 +1,129 @@ +dnl x86 mpn_bdiv_dbm1. + +dnl Copyright 2008, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C P5 +C P6 model 0-8,10-12) +C P6 model 9 (Banias) +C P6 model 13 (Dothan) 5.1 +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) 13.67 +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom +C AMD K6 +C AMD K7 3.5 +C AMD K8 +C AMD K10 + + +C TODO +C * Optimize for more x86 processors + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(mpn_bdiv_dbm1c) + mov 16(%esp), %ecx C d + push %esi + mov 12(%esp), %esi C ap + push %edi + mov 12(%esp), %edi C qp + push %ebp + mov 24(%esp), %ebp C n + push %ebx + + mov (%esi), %eax + mul %ecx + mov 36(%esp), %ebx + sub %eax, %ebx + mov %ebx, (%edi) + sbb %edx, %ebx + + mov %ebp, %eax + and $3, %eax + jz L(b0) + cmp $2, %eax + jc L(b1) + jz L(b2) + +L(b3): lea -8(%esi), %esi + lea 8(%edi), %edi + add $-3, %ebp + jmp L(3) + +L(b0): mov 4(%esi), %eax + lea -4(%esi), %esi + lea 12(%edi), %edi + add $-4, %ebp + jmp L(0) + +L(b2): mov 4(%esi), %eax + lea 4(%esi), %esi + lea 4(%edi), %edi + add $-2, %ebp + jmp L(2) + + ALIGN(8) +L(top): mov 4(%esi), %eax + mul %ecx + lea 16(%edi), %edi + sub %eax, %ebx + mov 8(%esi), %eax + mov %ebx, -12(%edi) + sbb %edx, %ebx +L(0): mul %ecx + sub %eax, %ebx + mov %ebx, -8(%edi) + sbb %edx, %ebx +L(3): mov 12(%esi), %eax + mul %ecx + sub %eax, %ebx + mov %ebx, -4(%edi) + mov 16(%esi), %eax + lea 16(%esi), %esi + sbb %edx, %ebx +L(2): mul %ecx + sub %eax, %ebx + mov %ebx, 0(%edi) + sbb %edx, %ebx +L(b1): add $-4, %ebp + jns L(top) + + mov %ebx, %eax + pop %ebx + pop %ebp + pop %edi + pop %esi + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/bdiv_q_1.asm b/gmp-6.3.0/mpn/x86/bdiv_q_1.asm new file mode 100644 index 0000000..132de06 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/bdiv_q_1.asm @@ -0,0 +1,208 @@ +dnl x86 mpn_bdiv_q_1 -- mpn by limb exact division. + +dnl Rearranged from mpn/x86/dive_1.asm by Marco Bodrato. + +dnl Copyright 2001, 2002, 2007, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P54 30.0 +C P55 29.0 +C P6 13.0 odd divisor, 12.0 even (strangely) +C K6 14.0 +C K7 12.0 +C P4 42.0 + +MULFUNC_PROLOGUE(mpn_bdiv_q_1 mpn_pi1_bdiv_q_1) + +defframe(PARAM_SHIFT, 24) +defframe(PARAM_INVERSE,20) +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(VAR_INVERSE,`PARAM_SRC') + + TEXT + +C mp_limb_t +C mpn_pi1_bdiv_q_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, mp_limb_t divisor, +C mp_limb_t inverse, int shift) + + ALIGN(16) +PROLOGUE(mpn_pi1_bdiv_q_1) +deflit(`FRAME',0) + + movl PARAM_SHIFT, %ecx + pushl %ebp FRAME_pushl() + + movl PARAM_INVERSE, %eax + movl PARAM_SIZE, %ebp + pushl %ebx FRAME_pushl() +L(common): + pushl %edi FRAME_pushl() + pushl %esi FRAME_pushl() + + movl PARAM_SRC, %esi + movl PARAM_DST, %edi + + leal (%esi,%ebp,4), %esi C src end + leal (%edi,%ebp,4), %edi C dst end + negl %ebp C -size + + movl %eax, VAR_INVERSE + movl (%esi,%ebp,4), %eax C src[0] + + xorl %ebx, %ebx + xorl %edx, %edx + + incl %ebp + jz L(one) + + movl (%esi,%ebp,4), %edx C src[1] + + shrdl( %cl, %edx, %eax) + + movl VAR_INVERSE, %edx + jmp L(entry) + + + ALIGN(8) + nop C k6 code alignment + nop +L(top): + C eax q + C ebx carry bit, 0 or -1 + C ecx shift + C edx carry limb + C esi src end + C edi dst end + C ebp counter, limbs, negative + + movl -4(%esi,%ebp,4), %eax + subl %ebx, %edx C accumulate carry bit + + movl (%esi,%ebp,4), %ebx + + shrdl( %cl, %ebx, %eax) + + subl %edx, %eax C apply carry limb + movl VAR_INVERSE, %edx + + sbbl %ebx, %ebx + +L(entry): + imull %edx, %eax + + movl %eax, -4(%edi,%ebp,4) + movl PARAM_DIVISOR, %edx + + mull %edx + + incl %ebp + jnz L(top) + + + movl -4(%esi), %eax C src high limb +L(one): + shrl %cl, %eax + popl %esi FRAME_popl() + + addl %ebx, %eax C apply carry bit + + subl %edx, %eax C apply carry limb + + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi) + + popl %edi + popl %ebx + popl %ebp + + ret + +EPILOGUE() + +C mp_limb_t mpn_bdiv_q_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C + + ALIGN(16) +PROLOGUE(mpn_bdiv_q_1) +deflit(`FRAME',0) + + movl PARAM_DIVISOR, %eax + pushl %ebp FRAME_pushl() + + movl $-1, %ecx C shift count + movl PARAM_SIZE, %ebp + + pushl %ebx FRAME_pushl() + +L(strip_twos): + incl %ecx + + shrl %eax + jnc L(strip_twos) + + leal 1(%eax,%eax), %ebx C d without twos + andl $127, %eax C d/2, 7 bits + +ifdef(`PIC',` + LEA( binvert_limb_table, %edx) + movzbl (%eax,%edx), %eax C inv 8 bits +',` + movzbl binvert_limb_table(%eax), %eax C inv 8 bits +') + + leal (%eax,%eax), %edx C 2*inv + movl %ebx, PARAM_DIVISOR C d without twos + imull %eax, %eax C inv*inv + imull %ebx, %eax C inv*inv*d + subl %eax, %edx C inv = 2*inv - inv*inv*d + + leal (%edx,%edx), %eax C 2*inv + imull %edx, %edx C inv*inv + imull %ebx, %edx C inv*inv*d + subl %edx, %eax C inv = 2*inv - inv*inv*d + + ASSERT(e,` C expect d*inv == 1 mod 2^GMP_LIMB_BITS + pushl %eax FRAME_pushl() + imull PARAM_DIVISOR, %eax + cmpl $1, %eax + popl %eax FRAME_popl()') + + jmp L(common) +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/bt1/gmp-mparam.h b/gmp-6.3.0/mpn/x86/bt1/gmp-mparam.h new file mode 100644 index 0000000..302dbc6 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/bt1/gmp-mparam.h @@ -0,0 +1,218 @@ +/* x86/bobcat gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* NOTE: In a fat binary build SQR_TOOM2_THRESHOLD here cannot be greater than + the value in mpn/x86/k7/gmp-mparam.h. The latter is used as a hard limit in + k7/sqr_basecase.asm. */ + +/* 1600 MHz AMD Bobcat Zacate E-350 */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-17, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 0 /* always */ +#define MOD_1_UNNORM_THRESHOLD 0 /* always */ +#define MOD_1N_TO_MOD_1_1_THRESHOLD 10 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 6 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 16 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 21 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 57.16% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 3 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 36 + +#define DIV_1_VS_MUL_1_PERCENT 199 + +#define MUL_TOOM22_THRESHOLD 28 +#define MUL_TOOM33_THRESHOLD 93 +#define MUL_TOOM44_THRESHOLD 166 +#define MUL_TOOM6H_THRESHOLD 270 +#define MUL_TOOM8H_THRESHOLD 478 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 102 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 177 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 169 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 113 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 143 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 50 +#define SQR_TOOM3_THRESHOLD 89 +#define SQR_TOOM4_THRESHOLD 248 +#define SQR_TOOM6_THRESHOLD 342 +#define SQR_TOOM8_THRESHOLD 470 + +#define MULMID_TOOM42_THRESHOLD 72 + +#define MULMOD_BNM1_THRESHOLD 20 +#define SQRMOD_BNM1_THRESHOLD 21 + +#define MUL_FFT_MODF_THRESHOLD 630 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 630, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 15, 5}, { 31, 6}, { 27, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 23, 6}, { 47, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 49, 8}, { 27, 7}, { 55, 9}, \ + { 15, 8}, { 31, 7}, { 63, 8}, { 43, 9}, \ + { 23, 8}, { 55, 9}, { 31, 8}, { 67, 9}, \ + { 39, 8}, { 79, 9}, { 47, 8}, { 95, 9}, \ + { 55,10}, { 31, 9}, { 79,10}, { 47, 6}, \ + { 767, 7}, { 399, 6}, { 799, 7}, { 415, 8}, \ + { 235, 7}, { 479, 9}, { 135,10}, { 79, 9}, \ + { 159,10}, { 95, 9}, { 191,11}, { 63,10}, \ + { 159,11}, { 95,10}, { 191,12}, { 63,11}, \ + { 127,10}, { 255, 9}, { 511,10}, { 271, 9}, \ + { 543,11}, { 159,10}, { 319, 9}, { 639,10}, \ + { 335, 9}, { 671,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 399, 9}, { 799,11}, { 223,12}, \ + { 127,11}, { 255,10}, { 543, 9}, { 1087,11}, \ + { 287,10}, { 607, 9}, { 1215,11}, { 319,10}, \ + { 671,11}, { 351,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,13}, { 127,12}, { 255,11}, \ + { 543,10}, { 1087,11}, { 607,10}, { 1215,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 735,10}, \ + { 1471,12}, { 383,11}, { 799,10}, { 1599,11}, \ + { 863,12}, { 447,11}, { 991,13}, { 255,12}, \ + { 511,11}, { 1087,12}, { 575,11}, { 1215,12}, \ + { 639,11}, { 1343,12}, { 703,11}, { 1471,13}, \ + { 383,12}, { 767,11}, { 1599,12}, { 831,11}, \ + { 1727,12}, { 959,14}, { 255,13}, { 511,12}, \ + { 1215,13}, { 639,12}, { 1471,13}, { 767,12}, \ + { 1727,13}, { 895,12}, { 1919,14}, { 511,13}, \ + { 1023,12}, { 2111,13}, { 1151,12}, { 2431,13}, \ + { 1407,14}, { 767,13}, { 1663,12}, { 3455,13}, \ + { 1919,15}, { 511,14}, { 1023,13}, { 2175,12}, \ + { 4479,13}, { 2431,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3967,15}, { 1023,14}, { 2047,13}, { 4479,14}, \ + { 2303,13}, { 4991,12}, { 9983,14}, { 2815,13}, \ + { 5887,15}, { 1535,14}, { 3839,16} } +#define MUL_FFT_TABLE3_SIZE 159 +#define MUL_FFT_THRESHOLD 7424 + +#define SQR_FFT_MODF_THRESHOLD 500 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 500, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 28, 7}, { 15, 6}, { 32, 7}, { 17, 6}, \ + { 35, 7}, { 19, 6}, { 39, 7}, { 23, 6}, \ + { 47, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 51, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 79, 9}, { 47, 8}, \ + { 95, 9}, { 55,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 127, 6}, { 1087, 7}, { 575, 8}, { 303, 9}, \ + { 159,10}, { 95,11}, { 63,10}, { 127, 9}, \ + { 255,10}, { 143, 9}, { 287,10}, { 159,11}, \ + { 95,10}, { 191,12}, { 63,11}, { 127,10}, \ + { 255, 9}, { 511,10}, { 271, 9}, { 543,10}, \ + { 287, 9}, { 575,11}, { 159,10}, { 319, 9}, \ + { 639,10}, { 335, 9}, { 671,10}, { 351,11}, \ + { 191,10}, { 383, 9}, { 767,10}, { 399, 9}, \ + { 799,10}, { 415, 9}, { 831,12}, { 127,11}, \ + { 255,10}, { 543,11}, { 287,10}, { 607,11}, \ + { 319,10}, { 671,11}, { 351,10}, { 703,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,10}, \ + { 831,13}, { 127,12}, { 255,11}, { 543,10}, \ + { 1087,11}, { 607,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,10}, { 1471,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,12}, { 447,11}, \ + { 959,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1471,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1727,12}, { 959,14}, \ + { 255,13}, { 511,12}, { 1215,13}, { 639,12}, \ + { 1471,13}, { 767,12}, { 1727,13}, { 895,12}, \ + { 1919,14}, { 511,13}, { 1023,12}, { 2111,13}, \ + { 1151,12}, { 2431,13}, { 1407,14}, { 767,13}, \ + { 1663,12}, { 3455,13}, { 1919,15}, { 511,14}, \ + { 1023,13}, { 2175,12}, { 4479,13}, { 2431,14}, \ + { 1279,13}, { 2943,12}, { 5887,14}, { 1535,13}, \ + { 3455,14}, { 1791,13}, { 3839,15}, { 1023,14}, \ + { 2047,13}, { 4479,14}, { 2303,13}, { 4991,12}, \ + { 9983,14}, { 2815,13}, { 5887,15}, { 1535,14}, \ + { 3839,16} } +#define SQR_FFT_TABLE3_SIZE 161 +#define SQR_FFT_THRESHOLD 5760 + +#define MULLO_BASECASE_THRESHOLD 9 +#define MULLO_DC_THRESHOLD 48 +#define MULLO_MUL_N_THRESHOLD 14281 +#define SQRLO_BASECASE_THRESHOLD 7 +#define SQRLO_DC_THRESHOLD 146 +#define SQRLO_SQR_THRESHOLD 11278 + +#define DC_DIV_QR_THRESHOLD 77 +#define DC_DIVAPPR_Q_THRESHOLD 240 +#define DC_BDIV_QR_THRESHOLD 83 +#define DC_BDIV_Q_THRESHOLD 182 + +#define INV_MULMOD_BNM1_THRESHOLD 74 +#define INV_NEWTON_THRESHOLD 252 +#define INV_APPR_THRESHOLD 252 + +#define BINV_NEWTON_THRESHOLD 252 +#define REDC_1_TO_REDC_N_THRESHOLD 79 + +#define MU_DIV_QR_THRESHOLD 1787 +#define MU_DIVAPPR_Q_THRESHOLD 1718 +#define MUPI_DIV_QR_THRESHOLD 122 +#define MU_BDIV_QR_THRESHOLD 1470 +#define MU_BDIV_Q_THRESHOLD 1713 + +#define POWM_SEC_TABLE 1,16,96,563,1317,1867 + +#define GET_STR_DC_THRESHOLD 19 +#define GET_STR_PRECOMPUTE_THRESHOLD 32 +#define SET_STR_DC_THRESHOLD 254 +#define SET_STR_PRECOMPUTE_THRESHOLD 907 + +#define FAC_DSC_THRESHOLD 224 +#define FAC_ODD_THRESHOLD 55 + +#define MATRIX22_STRASSEN_THRESHOLD 23 +#define HGCD2_DIV1_METHOD 3 /* 3.59% faster than 5 */ +#define HGCD_THRESHOLD 85 +#define HGCD_APPR_THRESHOLD 152 +#define HGCD_REDUCE_THRESHOLD 3389 +#define GCD_DC_THRESHOLD 531 +#define GCDEXT_DC_THRESHOLD 386 +#define JACOBI_BASE_METHOD 3 /* 0.92% faster than 1 */ + +/* Tuneup completed successfully, took 159946 seconds */ diff --git a/gmp-6.3.0/mpn/x86/bt2/gmp-mparam.h b/gmp-6.3.0/mpn/x86/bt2/gmp-mparam.h new file mode 100644 index 0000000..f936cb7 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/bt2/gmp-mparam.h @@ -0,0 +1,214 @@ +/* x86/bobcat gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* NOTE: In a fat binary build SQR_TOOM2_THRESHOLD here cannot be greater than + the value in mpn/x86/k7/gmp-mparam.h. The latter is used as a hard limit in + k7/sqr_basecase.asm. */ + +/* 2050 MHz AMD Jaguar/Kabini */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-24, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 4 +#define MOD_1_UNNORM_THRESHOLD 6 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 4 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 4 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 18 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 9 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 47.53% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 27 + +#define DIV_1_VS_MUL_1_PERCENT 243 + +#define MUL_TOOM22_THRESHOLD 32 +#define MUL_TOOM33_THRESHOLD 90 +#define MUL_TOOM44_THRESHOLD 154 +#define MUL_TOOM6H_THRESHOLD 286 +#define MUL_TOOM8H_THRESHOLD 478 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 97 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 152 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 103 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 113 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 154 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 38 +#define SQR_TOOM3_THRESHOLD 126 +#define SQR_TOOM4_THRESHOLD 220 +#define SQR_TOOM6_THRESHOLD 318 +#define SQR_TOOM8_THRESHOLD 502 + +#define MULMID_TOOM42_THRESHOLD 68 + +#define MULMOD_BNM1_THRESHOLD 19 +#define SQRMOD_BNM1_THRESHOLD 25 + +#define MUL_FFT_MODF_THRESHOLD 570 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 570, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 15, 5}, { 31, 6}, { 28, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 23, 6}, { 47, 7}, { 27, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 49, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 55, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 79, 9}, { 47, 8}, \ + { 95, 9}, { 55,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95,11}, \ + { 63,10}, { 159,11}, { 95,10}, { 191,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 511,10}, \ + { 271, 9}, { 543,10}, { 287,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335, 9}, { 671,11}, \ + { 191,10}, { 383, 9}, { 767,10}, { 399, 9}, \ + { 799,10}, { 415,11}, { 223,12}, { 127,11}, \ + { 255,10}, { 543,11}, { 287,10}, { 607, 9}, \ + { 1215,11}, { 319,10}, { 671,11}, { 351,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,13}, \ + { 127,12}, { 255,11}, { 543,10}, { 1087,11}, \ + { 607,10}, { 1215,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,10}, { 1471,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,12}, { 447,11}, \ + { 991,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1471,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1727,12}, { 959,14}, \ + { 255,13}, { 511,12}, { 1215,13}, { 639,12}, \ + { 1471,13}, { 767,12}, { 1727,13}, { 895,12}, \ + { 1919,14}, { 511,13}, { 1023,12}, { 2111,13}, \ + { 1151,12}, { 2431,13}, { 1407,14}, { 767,13}, \ + { 1663,12}, { 3455,13}, { 1919,15}, { 511,14}, \ + { 1023,13}, { 2175,12}, { 4479,13}, { 2431,14}, \ + { 1279,13}, { 2943,12}, { 5887,14}, { 1535,13}, \ + { 3455,14}, { 1791,13}, { 3967,15}, { 1023,14}, \ + { 2047,13}, { 4479,14}, { 2303,13}, { 4991,12}, \ + { 9983,14}, { 2815,13}, { 5887,15}, { 1535,14}, \ + { 3839,16} } +#define MUL_FFT_TABLE3_SIZE 153 +#define MUL_FFT_THRESHOLD 5760 + +#define SQR_FFT_MODF_THRESHOLD 530 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 530, 5}, { 27, 6}, { 15, 5}, { 31, 6}, \ + { 28, 7}, { 15, 6}, { 33, 7}, { 17, 6}, \ + { 35, 7}, { 19, 6}, { 39, 7}, { 23, 6}, \ + { 47, 7}, { 29, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 49, 8}, \ + { 31, 7}, { 63, 8}, { 39, 9}, { 23, 8}, \ + { 55, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 79, 9}, { 47, 8}, { 95, 9}, { 55,10}, \ + { 31, 9}, { 79,10}, { 47, 9}, { 95,11}, \ + { 31,10}, { 63, 9}, { 135,10}, { 95,11}, \ + { 63,10}, { 143, 9}, { 287,10}, { 159,11}, \ + { 95,12}, { 63,11}, { 127,10}, { 255, 9}, \ + { 511,10}, { 271, 9}, { 543,10}, { 287,11}, \ + { 159,10}, { 319, 9}, { 639,10}, { 335, 9}, \ + { 671,10}, { 351,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 399, 9}, { 799,12}, { 127,11}, \ + { 255,10}, { 543,11}, { 287,10}, { 607, 9}, \ + { 1215,11}, { 319,10}, { 671,11}, { 351,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,10}, \ + { 831,13}, { 127,12}, { 255,11}, { 543,10}, \ + { 1087,11}, { 607,10}, { 1215,12}, { 319,11}, \ + { 671,10}, { 1343,11}, { 735,10}, { 1471,12}, \ + { 383,11}, { 799,10}, { 1599,11}, { 863,12}, \ + { 447,11}, { 991,13}, { 255,12}, { 511,11}, \ + { 1087,12}, { 575,11}, { 1215,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,12}, \ + { 959,11}, { 1919,14}, { 255,13}, { 511,12}, \ + { 1215,13}, { 639,12}, { 1471,13}, { 767,12}, \ + { 1727,13}, { 895,12}, { 1919,14}, { 511,13}, \ + { 1023,12}, { 2111,13}, { 1151,12}, { 2495,13}, \ + { 1407,14}, { 767,13}, { 1663,12}, { 3455,13}, \ + { 1919,15}, { 511,14}, { 1023,13}, { 2175,12}, \ + { 4479,13}, { 2431,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3967,15}, { 1023,14}, { 2047,13}, { 4479,14}, \ + { 2303,13}, { 4991,12}, { 9983,14}, { 2815,13}, \ + { 5887,15}, { 1535,14}, { 3839,16} } +#define SQR_FFT_TABLE3_SIZE 151 +#define SQR_FFT_THRESHOLD 4736 + +#define MULLO_BASECASE_THRESHOLD 8 +#define MULLO_DC_THRESHOLD 44 +#define MULLO_MUL_N_THRESHOLD 11278 +#define SQRLO_BASECASE_THRESHOLD 13 +#define SQRLO_DC_THRESHOLD 62 +#define SQRLO_SQR_THRESHOLD 8907 + +#define DC_DIV_QR_THRESHOLD 79 +#define DC_DIVAPPR_Q_THRESHOLD 228 +#define DC_BDIV_QR_THRESHOLD 75 +#define DC_BDIV_Q_THRESHOLD 136 + +#define INV_MULMOD_BNM1_THRESHOLD 90 +#define INV_NEWTON_THRESHOLD 260 +#define INV_APPR_THRESHOLD 236 + +#define BINV_NEWTON_THRESHOLD 294 +#define REDC_1_TO_REDC_N_THRESHOLD 80 + +#define MU_DIV_QR_THRESHOLD 1787 +#define MU_DIVAPPR_Q_THRESHOLD 1718 +#define MUPI_DIV_QR_THRESHOLD 118 +#define MU_BDIV_QR_THRESHOLD 1442 +#define MU_BDIV_Q_THRESHOLD 1652 + +#define POWM_SEC_TABLE 1,16,96,615,865,1442 + +#define GET_STR_DC_THRESHOLD 16 +#define GET_STR_PRECOMPUTE_THRESHOLD 27 +#define SET_STR_DC_THRESHOLD 252 +#define SET_STR_PRECOMPUTE_THRESHOLD 638 + +#define FAC_DSC_THRESHOLD 141 +#define FAC_ODD_THRESHOLD 39 + +#define MATRIX22_STRASSEN_THRESHOLD 19 +#define HGCD2_DIV1_METHOD 1 /* 13.65% faster than 3 */ +#define HGCD_THRESHOLD 81 +#define HGCD_APPR_THRESHOLD 66 +#define HGCD_REDUCE_THRESHOLD 3389 +#define GCD_DC_THRESHOLD 531 +#define GCDEXT_DC_THRESHOLD 345 +#define JACOBI_BASE_METHOD 1 /* 0.84% faster than 4 */ + +/* Tuneup completed successfully, took 103818 seconds */ diff --git a/gmp-6.3.0/mpn/x86/cnd_aors_n.asm b/gmp-6.3.0/mpn/x86/cnd_aors_n.asm new file mode 100644 index 0000000..74f4917 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/cnd_aors_n.asm @@ -0,0 +1,124 @@ +dnl X86 mpn_cnd_add_n, mpn_cnd_sub_n + +dnl Copyright 2013 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C P5 ? +C P6 model 0-8,10-12 ? +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) 5.4 +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) 14.5 +C P4 model 3-4 (Prescott) 21 +C Intel atom 11 +C AMD K6 ? +C AMD K7 3.4 +C AMD K8 ? + + +define(`rp', `%edi') +define(`up', `%esi') +define(`vp', `%ebp') +define(`n', `%ecx') +define(`cnd', `20(%esp)') +define(`cy', `%edx') + +ifdef(`OPERATION_cnd_add_n', ` + define(ADDSUB, add) + define(ADCSBB, adc) + define(func, mpn_cnd_add_n)') +ifdef(`OPERATION_cnd_sub_n', ` + define(ADDSUB, sub) + define(ADCSBB, sbb) + define(func, mpn_cnd_sub_n)') + +MULFUNC_PROLOGUE(mpn_cnd_add_n mpn_cnd_sub_n) + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(func) + add $-16, %esp + mov %ebp, (%esp) + mov %ebx, 4(%esp) + mov %esi, 8(%esp) + mov %edi, 12(%esp) + + C make cnd into a full mask + mov cnd, %eax + neg %eax + sbb %eax, %eax + mov %eax, cnd + + C load parameters into registers + mov 24(%esp), rp + mov 28(%esp), up + mov 32(%esp), vp + mov 36(%esp), n + + mov (vp), %eax + mov (up), %ebx + + C put operand pointers just beyond their last limb + lea (vp,n,4), vp + lea (up,n,4), up + lea -4(rp,n,4), rp + neg n + + and cnd, %eax + ADDSUB %eax, %ebx + sbb cy, cy + inc n + je L(end) + + ALIGN(16) +L(top): mov (vp,n,4), %eax + and cnd, %eax + mov %ebx, (rp,n,4) + mov (up,n,4), %ebx + add cy, cy + ADCSBB %eax, %ebx + sbb cy, cy + inc n + jne L(top) + +L(end): mov %ebx, (rp) + xor %eax, %eax + sub cy, %eax + + mov (%esp), %ebp + mov 4(%esp), %ebx + mov 8(%esp), %esi + mov 12(%esp), %edi + add $16, %esp + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/copyd.asm b/gmp-6.3.0/mpn/x86/copyd.asm new file mode 100644 index 0000000..51fa195 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/copyd.asm @@ -0,0 +1,91 @@ +dnl x86 mpn_copyd -- copy limb vector, decrementing. + +dnl Copyright 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb startup (approx) +C P5 1.0 40 +C P6 2.4 70 +C K6 1.0 55 +C K7 1.3 75 +C P4 2.6 175 +C +C (Startup time includes some function call overheads.) + + +C void mpn_copyd (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C Copy src,size to dst,size, working from high to low addresses. +C +C The code here is very generic and can be expected to be reasonable on all +C the x86 family. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) +deflit(`FRAME',0) + + TEXT + ALIGN(32) + +PROLOGUE(mpn_copyd) + C eax saved esi + C ebx + C ecx counter + C edx saved edi + C esi src + C edi dst + C ebp + + movl PARAM_SIZE, %ecx + movl %esi, %eax + + movl PARAM_SRC, %esi + movl %edi, %edx + + movl PARAM_DST, %edi + leal -4(%esi,%ecx,4), %esi + + leal -4(%edi,%ecx,4), %edi + + std + + rep + movsl + + cld + + movl %eax, %esi + movl %edx, %edi + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/copyi.asm b/gmp-6.3.0/mpn/x86/copyi.asm new file mode 100644 index 0000000..f6b0354 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/copyi.asm @@ -0,0 +1,99 @@ +dnl x86 mpn_copyi -- copy limb vector, incrementing. + +dnl Copyright 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb startup (approx) +C P5 1.0 35 +C P6 0.75 45 +C K6 1.0 30 +C K7 1.3 65 +C P4 1.0 120 +C +C (Startup time includes some function call overheads.) + + +C void mpn_copyi (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C Copy src,size to dst,size, working from low to high addresses. +C +C The code here is very generic and can be expected to be reasonable on all +C the x86 family. +C +C P6 - An MMX based copy was tried, but was found to be slower than a rep +C movs in all cases. The fastest MMX found was 0.8 cycles/limb (when +C fully aligned). A rep movs seems to have a startup time of about 15 +C cycles, but doing something special for small sizes could lead to a +C branch misprediction that would destroy any saving. For now a plain +C rep movs seems ok. +C +C K62 - We used to have a big chunk of code doing an MMX copy at 0.56 c/l if +C aligned or a 1.0 rep movs if not. But that seemed excessive since +C it only got an advantage half the time, and even then only showed it +C above 50 limbs or so. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) +deflit(`FRAME',0) + + TEXT + ALIGN(32) + + C eax saved esi + C ebx + C ecx counter + C edx saved edi + C esi src + C edi dst + C ebp + +PROLOGUE(mpn_copyi) + + movl PARAM_SIZE, %ecx + movl %esi, %eax + + movl PARAM_SRC, %esi + movl %edi, %edx + + movl PARAM_DST, %edi + + cld C better safe than sorry, see mpn/x86/README + + rep + movsl + + movl %eax, %esi + movl %edx, %edi + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/core2/gmp-mparam.h b/gmp-6.3.0/mpn/x86/core2/gmp-mparam.h new file mode 100644 index 0000000..8a44ad1 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/core2/gmp-mparam.h @@ -0,0 +1,210 @@ +/* x86/core2 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 3000 MHz Penryn */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-20, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define MOD_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define MOD_1N_TO_MOD_1_1_THRESHOLD 9 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 8 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 9 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 3 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 2 /* 22.20% faster than 1 */ +#define DIV_QR_1_NORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD 9 +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 18 + +#define DIV_1_VS_MUL_1_PERCENT 277 + +#define MUL_TOOM22_THRESHOLD 24 +#define MUL_TOOM33_THRESHOLD 93 +#define MUL_TOOM44_THRESHOLD 136 +#define MUL_TOOM6H_THRESHOLD 300 +#define MUL_TOOM8H_THRESHOLD 478 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 91 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 153 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 93 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 94 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 130 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 34 +#define SQR_TOOM3_THRESHOLD 117 +#define SQR_TOOM4_THRESHOLD 184 +#define SQR_TOOM6_THRESHOLD 262 +#define SQR_TOOM8_THRESHOLD 597 + +#define MULMID_TOOM42_THRESHOLD 70 + +#define MULMOD_BNM1_THRESHOLD 17 +#define SQRMOD_BNM1_THRESHOLD 25 + +#define MUL_FFT_MODF_THRESHOLD 505 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 505, 5}, { 28, 6}, { 15, 5}, { 31, 6}, \ + { 29, 7}, { 15, 6}, { 32, 7}, { 17, 6}, \ + { 35, 7}, { 19, 6}, { 39, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 43, 8}, \ + { 23, 7}, { 47, 8}, { 27, 9}, { 15, 8}, \ + { 31, 7}, { 63, 8}, { 43, 9}, { 23, 8}, \ + { 55, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 79, 9}, { 55,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95,11}, \ + { 63, 9}, { 255,10}, { 159,11}, { 95,10}, \ + { 191,12}, { 63,11}, { 127,10}, { 271, 9}, \ + { 543,10}, { 287,11}, { 159,10}, { 335, 9}, \ + { 671,10}, { 351,11}, { 191,10}, { 399, 9}, \ + { 799,11}, { 223,12}, { 127,11}, { 255,10}, \ + { 543,11}, { 287,10}, { 607,11}, { 319,10}, \ + { 671,11}, { 351,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,13}, { 127,12}, { 255,11}, \ + { 543,10}, { 1087,11}, { 607,12}, { 319,11}, \ + { 671,10}, { 1343,11}, { 735,10}, { 1471,12}, \ + { 383,11}, { 799,10}, { 1599,11}, { 863,12}, \ + { 447,11}, { 959,13}, { 255,12}, { 511,11}, \ + { 1087,12}, { 575,11}, { 1215,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,12}, \ + { 959,14}, { 255,13}, { 511,12}, { 1087,11}, \ + { 2239,12}, { 1215,13}, { 639,12}, { 1471,11}, \ + { 2943,13}, { 767,12}, { 1727,13}, { 895,12}, \ + { 1919,14}, { 511,13}, { 1023,12}, { 2239,13}, \ + { 1151,12}, { 2431,13}, { 1407,12}, { 2943,14}, \ + { 767,13}, { 1663,12}, { 3455,13}, { 1919,15}, \ + { 511,14}, { 1023,13}, { 2175,12}, { 4479,13}, \ + { 2431,14}, { 1279,13}, { 2943,12}, { 5887,14}, \ + { 1535,13}, { 3455,14}, { 1791,13}, { 3967,12}, \ + { 7935,15}, { 1023,14}, { 2047,13}, { 4479,14}, \ + { 2303,13}, { 4991,12}, { 9983,14}, { 2815,13}, \ + { 5887,15}, { 1535,14}, { 3839,16} } +#define MUL_FFT_TABLE3_SIZE 147 +#define MUL_FFT_THRESHOLD 6784 + +#define SQR_FFT_MODF_THRESHOLD 464 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 464, 5}, { 28, 6}, { 15, 5}, { 31, 6}, \ + { 29, 7}, { 15, 6}, { 32, 7}, { 17, 6}, \ + { 35, 7}, { 19, 6}, { 39, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 49, 8}, { 27, 9}, { 15, 8}, \ + { 39, 9}, { 23, 8}, { 51,10}, { 15, 9}, \ + { 31, 8}, { 67, 9}, { 39, 8}, { 79, 9}, \ + { 47, 8}, { 95, 9}, { 55,10}, { 31, 9}, \ + { 79,10}, { 47, 9}, { 95,11}, { 31,10}, \ + { 63, 9}, { 127,10}, { 79, 9}, { 159,10}, \ + { 95,11}, { 63,10}, { 127, 9}, { 255,10}, \ + { 143, 9}, { 287, 5}, { 4863, 6}, { 2495, 7}, \ + { 1343, 8}, { 703, 9}, { 367,12}, { 63,11}, \ + { 127,10}, { 303,11}, { 159,10}, { 319, 9}, \ + { 639,10}, { 335, 9}, { 671,10}, { 351, 9}, \ + { 703,10}, { 367,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 399, 9}, { 799,10}, { 415, 9}, \ + { 831,12}, { 127,11}, { 255,10}, { 543,11}, \ + { 287,10}, { 607,11}, { 319,10}, { 671,11}, \ + { 351,10}, { 703,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,10}, { 863,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 735,10}, \ + { 1471,12}, { 383,11}, { 799,10}, { 1599,11}, \ + { 863,12}, { 447,11}, { 959,13}, { 255,12}, \ + { 511,11}, { 1087,12}, { 575,11}, { 1215,12}, \ + { 639,11}, { 1343,12}, { 703,11}, { 1407,13}, \ + { 383,12}, { 767,11}, { 1599,12}, { 831,11}, \ + { 1727,12}, { 959,14}, { 255,13}, { 511,12}, \ + { 1215,13}, { 639,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1727,13}, { 895,12}, { 1919,14}, \ + { 511,13}, { 1023,12}, { 2111,13}, { 1151,12}, \ + { 2431,13}, { 1407,12}, { 2943,14}, { 767,13}, \ + { 1663,12}, { 3455,13}, { 1919,15}, { 511,14}, \ + { 1023,13}, { 2175,12}, { 4479,13}, { 2431,14}, \ + { 1279,13}, { 2943,12}, { 5887,14}, { 1535,13}, \ + { 3455,14}, { 1791,13}, { 3967,15}, { 1023,14}, \ + { 2047,13}, { 4479,14}, { 2303,13}, { 4991,12}, \ + { 9983,14}, { 2815,13}, { 5887,15}, { 1535,14}, \ + { 3839,16} } +#define SQR_FFT_TABLE3_SIZE 157 +#define SQR_FFT_THRESHOLD 5312 + +#define MULLO_BASECASE_THRESHOLD 0 /* always */ +#define MULLO_DC_THRESHOLD 36 +#define MULLO_MUL_N_THRESHOLD 13463 +#define SQRLO_BASECASE_THRESHOLD 0 /* always */ +#define SQRLO_DC_THRESHOLD 140 +#define SQRLO_SQR_THRESHOLD 10393 + +#define DC_DIV_QR_THRESHOLD 32 +#define DC_DIVAPPR_Q_THRESHOLD 116 +#define DC_BDIV_QR_THRESHOLD 76 +#define DC_BDIV_Q_THRESHOLD 180 + +#define INV_MULMOD_BNM1_THRESHOLD 46 +#define INV_NEWTON_THRESHOLD 138 +#define INV_APPR_THRESHOLD 123 + +#define BINV_NEWTON_THRESHOLD 306 +#define REDC_1_TO_REDC_N_THRESHOLD 82 + +#define MU_DIV_QR_THRESHOLD 1499 +#define MU_DIVAPPR_Q_THRESHOLD 1442 +#define MUPI_DIV_QR_THRESHOLD 63 +#define MU_BDIV_QR_THRESHOLD 1442 +#define MU_BDIV_Q_THRESHOLD 1589 + +#define POWM_SEC_TABLE 1,22,66,428,1035 + +#define GET_STR_DC_THRESHOLD 13 +#define GET_STR_PRECOMPUTE_THRESHOLD 18 +#define SET_STR_DC_THRESHOLD 732 +#define SET_STR_PRECOMPUTE_THRESHOLD 1118 + +#define FAC_DSC_THRESHOLD 115 +#define FAC_ODD_THRESHOLD 50 + +#define MATRIX22_STRASSEN_THRESHOLD 25 +#define HGCD2_DIV1_METHOD 1 /* 5.78% faster than 3 */ +#define HGCD_THRESHOLD 121 +#define HGCD_APPR_THRESHOLD 151 +#define HGCD_REDUCE_THRESHOLD 3259 +#define GCD_DC_THRESHOLD 368 +#define GCDEXT_DC_THRESHOLD 306 +#define JACOBI_BASE_METHOD 4 /* 14.19% faster than 1 */ + +/* Tuneup completed successfully, took 67142 seconds */ diff --git a/gmp-6.3.0/mpn/x86/coreibwl/gmp-mparam.h b/gmp-6.3.0/mpn/x86/coreibwl/gmp-mparam.h new file mode 100644 index 0000000..7b58cad --- /dev/null +++ b/gmp-6.3.0/mpn/x86/coreibwl/gmp-mparam.h @@ -0,0 +1,216 @@ +/* x86/coreibwl gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 3400-3800 MHz Intel Xeon E3-1285Lv4 Broadwell */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-20, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 15 +#define MOD_1_UNNORM_THRESHOLD 16 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 10 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 8 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 0 /* never mpn_mod_1_1p */ +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 10 +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 11 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 21.34% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 14 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD 29 +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 19 + +#define DIV_1_VS_MUL_1_PERCENT 295 + +#define MUL_TOOM22_THRESHOLD 26 +#define MUL_TOOM33_THRESHOLD 97 +#define MUL_TOOM44_THRESHOLD 220 +#define MUL_TOOM6H_THRESHOLD 306 +#define MUL_TOOM8H_THRESHOLD 454 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 93 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 153 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 154 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 169 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 136 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 44 +#define SQR_TOOM3_THRESHOLD 134 +#define SQR_TOOM4_THRESHOLD 242 +#define SQR_TOOM6_THRESHOLD 342 +#define SQR_TOOM8_THRESHOLD 502 + +#define MULMID_TOOM42_THRESHOLD 98 + +#define MULMOD_BNM1_THRESHOLD 20 +#define SQRMOD_BNM1_THRESHOLD 23 + +#define MUL_FFT_MODF_THRESHOLD 540 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 540, 5}, { 29, 6}, { 15, 5}, { 31, 6}, \ + { 16, 5}, { 33, 6}, { 17, 5}, { 36, 6}, \ + { 25, 7}, { 13, 6}, { 29, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 36, 7}, { 19, 6}, \ + { 39, 7}, { 21, 6}, { 43, 7}, { 23, 6}, \ + { 47, 7}, { 29, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 43, 8}, { 23, 7}, { 49, 8}, \ + { 27, 7}, { 55, 9}, { 15, 8}, { 31, 7}, \ + { 63, 8}, { 43, 9}, { 23, 8}, { 55,10}, \ + { 15, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 83, 9}, { 47, 8}, { 95, 9}, { 55,10}, \ + { 31, 9}, { 79,10}, { 47, 9}, { 95,11}, \ + { 31,10}, { 63, 9}, { 135,10}, { 79, 9}, \ + { 159,10}, { 95, 9}, { 191,10}, { 111,11}, \ + { 63,10}, { 143, 9}, { 287,10}, { 159,11}, \ + { 95, 7}, { 1599, 8}, { 831, 9}, { 431, 8}, \ + { 863, 9}, { 447,10}, { 239, 9}, { 479,10}, \ + { 255, 9}, { 511,10}, { 287,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335, 9}, { 671,11}, \ + { 191,10}, { 383, 9}, { 767,10}, { 399,11}, \ + { 223,12}, { 127,11}, { 255,10}, { 511, 9}, \ + { 1023,11}, { 287,10}, { 607,11}, { 319,10}, \ + { 671,11}, { 351,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,13}, { 127,12}, { 255,11}, \ + { 543,10}, { 1119,11}, { 607,12}, { 319,11}, \ + { 671,10}, { 1343,11}, { 735,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,12}, { 447,11}, \ + { 959,13}, { 255,12}, { 511,11}, { 1119,12}, \ + { 575,11}, { 1215,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1407,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1727,12}, { 959,14}, \ + { 255,13}, { 511,12}, { 1215,13}, { 639,12}, \ + { 1471,13}, { 767,12}, { 1727,13}, { 895,12}, \ + { 1919,14}, { 511,13}, { 1023,12}, { 2239,13}, \ + { 1151,12}, { 2431,13}, { 1279,12}, { 2623,13}, \ + { 1407,12}, { 2815,14}, { 767,13}, { 1535,12}, \ + { 3135,13}, { 1663,12}, { 3455,13}, { 1919,15}, \ + { 511,14}, { 1023,13}, { 2175,12}, { 4479,13}, \ + { 2431,14}, { 1279,13}, { 2943,12}, { 5887,14}, \ + { 1535,13}, { 3455,14}, { 1791,13}, { 3839,15}, \ + { 1023,14}, { 2047,13}, { 4479,14}, { 2303,13}, \ + { 4991,12}, { 9983,14}, { 2559,13}, { 5247,14}, \ + { 2815,13}, { 5887,15}, { 1535,14}, { 3839,16} } +#define MUL_FFT_TABLE3_SIZE 172 +#define MUL_FFT_THRESHOLD 7424 + +#define SQR_FFT_MODF_THRESHOLD 472 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 472, 5}, { 29, 6}, { 15, 5}, { 33, 6}, \ + { 37, 7}, { 19, 6}, { 40, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 43, 8}, \ + { 23, 7}, { 49, 8}, { 27, 9}, { 15, 8}, \ + { 31, 7}, { 63, 8}, { 43, 9}, { 23, 8}, \ + { 55,10}, { 15, 9}, { 31, 8}, { 67, 9}, \ + { 39, 8}, { 83, 9}, { 47, 8}, { 95, 9}, \ + { 55,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 135,10}, \ + { 79, 9}, { 159,10}, { 95,11}, { 63,10}, \ + { 127, 9}, { 255,10}, { 143, 9}, { 287,10}, \ + { 159,11}, { 95,12}, { 63,11}, { 127,10}, \ + { 271, 9}, { 543, 6}, { 4479, 7}, { 2431, 8}, \ + { 1247, 7}, { 2495, 8}, { 1279,10}, { 351,11}, \ + { 191,10}, { 399, 9}, { 799,10}, { 415,12}, \ + { 127,11}, { 255,10}, { 543,11}, { 287,10}, \ + { 607,11}, { 319,10}, { 639,11}, { 351,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,10}, \ + { 831,13}, { 127,12}, { 255,11}, { 511,10}, \ + { 1023,11}, { 543,10}, { 1087,11}, { 607,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 735,12}, \ + { 383,11}, { 799,10}, { 1599,11}, { 863,12}, \ + { 447,11}, { 927,13}, { 255,12}, { 511,11}, \ + { 1087,12}, { 575,11}, { 1215,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1663,12}, \ + { 895,11}, { 1855,14}, { 255,13}, { 511,12}, \ + { 1023,11}, { 2047,12}, { 1087,11}, { 2239,12}, \ + { 1215,13}, { 639,12}, { 1471,13}, { 767,12}, \ + { 1663,13}, { 895,12}, { 1983,14}, { 511,13}, \ + { 1023,12}, { 2239,13}, { 1151,12}, { 2495,13}, \ + { 1279,12}, { 2623,13}, { 1407,14}, { 767,13}, \ + { 1535,12}, { 3135,13}, { 1663,12}, { 3455,13}, \ + { 1919,15}, { 511,14}, { 1023,13}, { 2175,12}, \ + { 4479,13}, { 2431,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3839,15}, { 1023,14}, { 2047,13}, { 4479,14}, \ + { 2303,13}, { 4991,12}, { 9983,14}, { 2815,13}, \ + { 5887,15}, { 1535,14}, { 3327,13}, { 6783,14}, \ + { 3839,16} } +#define SQR_FFT_TABLE3_SIZE 157 +#define SQR_FFT_THRESHOLD 5568 + +#define MULLO_BASECASE_THRESHOLD 16 +#define MULLO_DC_THRESHOLD 37 +#define MULLO_MUL_N_THRESHOLD 14281 +#define SQRLO_BASECASE_THRESHOLD 0 /* always */ +#define SQRLO_DC_THRESHOLD 137 +#define SQRLO_SQR_THRESHOLD 10821 + +#define DC_DIV_QR_THRESHOLD 54 +#define DC_DIVAPPR_Q_THRESHOLD 146 +#define DC_BDIV_QR_THRESHOLD 98 +#define DC_BDIV_Q_THRESHOLD 218 + +#define INV_MULMOD_BNM1_THRESHOLD 50 +#define INV_NEWTON_THRESHOLD 173 +#define INV_APPR_THRESHOLD 165 + +#define BINV_NEWTON_THRESHOLD 278 +#define REDC_1_TO_REDC_N_THRESHOLD 79 + +#define MU_DIV_QR_THRESHOLD 1787 +#define MU_DIVAPPR_Q_THRESHOLD 1787 +#define MUPI_DIV_QR_THRESHOLD 78 +#define MU_BDIV_QR_THRESHOLD 1589 +#define MU_BDIV_Q_THRESHOLD 1830 + +#define POWM_SEC_TABLE 1,16,126,416,932 + +#define GET_STR_DC_THRESHOLD 11 +#define GET_STR_PRECOMPUTE_THRESHOLD 17 +#define SET_STR_DC_THRESHOLD 306 +#define SET_STR_PRECOMPUTE_THRESHOLD 894 + +#define FAC_DSC_THRESHOLD 141 +#define FAC_ODD_THRESHOLD 34 + +#define MATRIX22_STRASSEN_THRESHOLD 20 +#define HGCD2_DIV1_METHOD 3 /* 5.97% faster than 1 */ +#define HGCD_THRESHOLD 73 +#define HGCD_APPR_THRESHOLD 123 +#define HGCD_REDUCE_THRESHOLD 3664 +#define GCD_DC_THRESHOLD 562 +#define GCDEXT_DC_THRESHOLD 465 +#define JACOBI_BASE_METHOD 1 /* 31.16% faster than 3 */ + +/* Tuneup completed successfully, took 35114 seconds */ diff --git a/gmp-6.3.0/mpn/x86/coreihwl/gmp-mparam.h b/gmp-6.3.0/mpn/x86/coreihwl/gmp-mparam.h new file mode 100644 index 0000000..4c1b388 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/coreihwl/gmp-mparam.h @@ -0,0 +1,216 @@ +/* x86/coreihwl gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 3600-4000 MHz Intel Xeon E3-1271v3 Haswell */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-21, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 17 +#define MOD_1_UNNORM_THRESHOLD 17 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 10 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 9 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 0 /* never mpn_mod_1_1p */ +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 10 +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 5 +#define USE_PREINV_DIVREM_1 1 /* native */ +/* From sky.gmplib.org, 2023-07-20 */ +#define DIV_QR_1N_PI1_METHOD 3 /* 5.86% faster than 1 */ +#define DIV_QR_1_NORM_THRESHOLD 13 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 21 + +#define DIV_1_VS_MUL_1_PERCENT 296 + +#define MUL_TOOM22_THRESHOLD 28 +#define MUL_TOOM33_THRESHOLD 108 +#define MUL_TOOM44_THRESHOLD 232 +#define MUL_TOOM6H_THRESHOLD 306 +#define MUL_TOOM8H_THRESHOLD 478 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 109 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 183 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 113 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 113 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 136 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 44 +#define SQR_TOOM3_THRESHOLD 141 +#define SQR_TOOM4_THRESHOLD 384 +#define SQR_TOOM6_THRESHOLD 517 +#define SQR_TOOM8_THRESHOLD 698 + +#define MULMID_TOOM42_THRESHOLD 98 + +#define MULMOD_BNM1_THRESHOLD 20 +#define SQRMOD_BNM1_THRESHOLD 23 + +#define MUL_FFT_MODF_THRESHOLD 565 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 565, 5}, { 28, 6}, { 15, 5}, { 31, 6}, \ + { 16, 5}, { 33, 6}, { 29, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 36, 7}, { 19, 6}, \ + { 39, 7}, { 21, 6}, { 43, 7}, { 23, 6}, \ + { 47, 7}, { 29, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 43, 8}, { 23, 7}, { 47, 8}, \ + { 27, 7}, { 55, 9}, { 15, 8}, { 31, 7}, \ + { 63, 8}, { 43, 9}, { 23, 8}, { 55, 9}, \ + { 31, 8}, { 71, 9}, { 39, 8}, { 83, 9}, \ + { 47, 8}, { 95, 9}, { 55,10}, { 31, 9}, \ + { 79,10}, { 47, 9}, { 95,11}, { 31,10}, \ + { 63, 9}, { 135,10}, { 79, 9}, { 159,10}, \ + { 95, 9}, { 191,10}, { 111,11}, { 63,10}, \ + { 143, 9}, { 287,10}, { 159,11}, { 95,10}, \ + { 191, 6}, { 3199, 7}, { 1727, 9}, { 447,10}, \ + { 239, 9}, { 479,10}, { 287,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335, 9}, { 671,11}, \ + { 191,10}, { 399, 9}, { 799,10}, { 415,11}, \ + { 223,12}, { 127,11}, { 255,10}, { 511, 9}, \ + { 1023,10}, { 527,11}, { 287,10}, { 607,11}, \ + { 319,10}, { 671,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,13}, { 127,12}, { 255,11}, \ + { 511,10}, { 1023,11}, { 543,10}, { 1087,11}, \ + { 607,12}, { 319,11}, { 671,10}, { 1343,11}, \ + { 735,12}, { 383,11}, { 799,10}, { 1599,11}, \ + { 863,12}, { 447,11}, { 991,13}, { 255,12}, \ + { 511,11}, { 1087,12}, { 575,11}, { 1215,12}, \ + { 639,11}, { 1343,12}, { 703,11}, { 1407,13}, \ + { 383,12}, { 767,11}, { 1599,12}, { 831,11}, \ + { 1727,12}, { 959,11}, { 1919,14}, { 255,13}, \ + { 511,12}, { 1087,11}, { 2239,12}, { 1215,13}, \ + { 639,12}, { 1471,13}, { 767,12}, { 1727,13}, \ + { 895,12}, { 1919,14}, { 511,13}, { 1023,12}, \ + { 2239,13}, { 1151,12}, { 2431,13}, { 1279,12}, \ + { 2623,13}, { 1407,14}, { 767,13}, { 1663,12}, \ + { 3455,13}, { 1919,15}, { 511,14}, { 1023,13}, \ + { 2175,12}, { 4479,13}, { 2431,14}, { 1279,13}, \ + { 2943,12}, { 5887,14}, { 1535,13}, { 3455,14}, \ + { 1791,13}, { 3967,15}, { 1023,14}, { 2047,13}, \ + { 4479,14}, { 2303,13}, { 4991,14}, { 2559,13}, \ + { 5375,14}, { 2815,13}, { 5887,15}, { 1535,14}, \ + { 3839,16} } +#define MUL_FFT_TABLE3_SIZE 165 +#define MUL_FFT_THRESHOLD 7808 + +#define SQR_FFT_MODF_THRESHOLD 560 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 560, 5}, { 29, 6}, { 15, 5}, { 31, 6}, \ + { 16, 5}, { 33, 6}, { 17, 5}, { 36, 6}, \ + { 29, 7}, { 15, 6}, { 33, 7}, { 17, 6}, \ + { 36, 7}, { 19, 6}, { 40, 7}, { 21, 6}, \ + { 43, 7}, { 23, 6}, { 47, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 43, 8}, \ + { 23, 7}, { 49, 8}, { 27, 7}, { 55, 9}, \ + { 15, 8}, { 31, 7}, { 63, 8}, { 43, 9}, \ + { 23, 8}, { 55,10}, { 15, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 79, 9}, { 47, 8}, \ + { 95, 9}, { 55,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95,11}, \ + { 63,10}, { 143, 9}, { 287,10}, { 159,11}, \ + { 95,12}, { 63,11}, { 127, 9}, { 511, 5}, \ + { 8959, 7}, { 2431, 8}, { 1247, 7}, { 2495, 8}, \ + { 1279, 9}, { 671,10}, { 367,11}, { 191,10}, \ + { 399, 9}, { 799,10}, { 415,12}, { 127,11}, \ + { 255,10}, { 527,11}, { 287,10}, { 607,11}, \ + { 319,10}, { 671,11}, { 351,10}, { 703,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,10}, \ + { 831,13}, { 127,11}, { 543,10}, { 1119,11}, \ + { 607,12}, { 319,11}, { 671,10}, { 1343,11}, \ + { 735,12}, { 383,11}, { 863,12}, { 447,11}, \ + { 991,12}, { 511,11}, { 1119,12}, { 575,11}, \ + { 1215,12}, { 639,11}, { 1343,12}, { 703,13}, \ + { 383,12}, { 767,11}, { 1599,12}, { 831,11}, \ + { 1727,12}, { 959,11}, { 1983,13}, { 511,12}, \ + { 1087,11}, { 2239,12}, { 1215,13}, { 639,12}, \ + { 1471,13}, { 767,12}, { 1727,13}, { 895,12}, \ + { 1983,14}, { 511,13}, { 1023,12}, { 2239,13}, \ + { 1151,12}, { 2495,13}, { 1279,12}, { 2623,13}, \ + { 1407,14}, { 767,13}, { 1663,12}, { 3455,13}, \ + { 1919,15}, { 511,14}, { 1023,13}, { 2175,12}, \ + { 4479,13}, { 2431,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3967,15}, { 1023,14}, { 2047,13}, { 4479,14}, \ + { 2303,13}, { 4991,12}, { 9983,14}, { 2559,13}, \ + { 5119,14}, { 2815,13}, { 5887,15}, { 1535,14}, \ + { 3327,13}, { 6911,14}, { 3839,16} } +#define SQR_FFT_TABLE3_SIZE 159 +#define SQR_FFT_THRESHOLD 5568 + +#define MULLO_BASECASE_THRESHOLD 17 +#define MULLO_DC_THRESHOLD 40 +#define MULLO_MUL_N_THRESHOLD 14281 +#define SQRLO_BASECASE_THRESHOLD 0 /* always */ +#define SQRLO_DC_THRESHOLD 141 +#define SQRLO_SQR_THRESHOLD 10821 + +#define DC_DIV_QR_THRESHOLD 30 +#define DC_DIVAPPR_Q_THRESHOLD 190 +#define DC_BDIV_QR_THRESHOLD 67 +#define DC_BDIV_Q_THRESHOLD 254 + +#define INV_MULMOD_BNM1_THRESHOLD 54 +#define INV_NEWTON_THRESHOLD 157 +#define INV_APPR_THRESHOLD 163 + +#define BINV_NEWTON_THRESHOLD 236 +#define REDC_1_TO_REDC_N_THRESHOLD 79 + +#define MU_DIV_QR_THRESHOLD 1895 +#define MU_DIVAPPR_Q_THRESHOLD 1718 +#define MUPI_DIV_QR_THRESHOLD 54 +#define MU_BDIV_QR_THRESHOLD 1589 +#define MU_BDIV_Q_THRESHOLD 1898 + +#define POWM_SEC_TABLE 1,16,95,480,1442 + +#define GET_STR_DC_THRESHOLD 10 +#define GET_STR_PRECOMPUTE_THRESHOLD 16 +#define SET_STR_DC_THRESHOLD 372 +#define SET_STR_PRECOMPUTE_THRESHOLD 1037 + +#define FAC_DSC_THRESHOLD 141 +#define FAC_ODD_THRESHOLD 34 + +#define MATRIX22_STRASSEN_THRESHOLD 21 +#define HGCD2_DIV1_METHOD 3 /* 6.26% faster than 1 */ +#define HGCD_THRESHOLD 70 +#define HGCD_APPR_THRESHOLD 129 +#define HGCD_REDUCE_THRESHOLD 3664 +#define GCD_DC_THRESHOLD 573 +#define GCDEXT_DC_THRESHOLD 483 +#define JACOBI_BASE_METHOD 1 /* 27.01% faster than 3 */ + +/* Tuneup completed successfully, took 35232 seconds */ diff --git a/gmp-6.3.0/mpn/x86/coreinhm/gmp-mparam.h b/gmp-6.3.0/mpn/x86/coreinhm/gmp-mparam.h new file mode 100644 index 0000000..4428b4b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/coreinhm/gmp-mparam.h @@ -0,0 +1,223 @@ +/* x86/coreinhm gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 2933-3200 MHz Intel Xeon X3470 Nehalem */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-23, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 36 +#define MOD_1_UNNORM_THRESHOLD 40 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 8 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 7 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 12 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 3 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 42.59% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD 9 +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 17 + +#define DIV_1_VS_MUL_1_PERCENT 288 + +#define MUL_TOOM22_THRESHOLD 24 +#define MUL_TOOM33_THRESHOLD 93 +#define MUL_TOOM44_THRESHOLD 214 +#define MUL_TOOM6H_THRESHOLD 306 +#define MUL_TOOM8H_THRESHOLD 430 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 93 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 134 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 145 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 94 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 118 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 38 +#define SQR_TOOM3_THRESHOLD 133 +#define SQR_TOOM4_THRESHOLD 212 +#define SQR_TOOM6_THRESHOLD 318 +#define SQR_TOOM8_THRESHOLD 620 + +#define MULMID_TOOM42_THRESHOLD 68 + +#define MULMOD_BNM1_THRESHOLD 17 +#define SQRMOD_BNM1_THRESHOLD 23 + +#define MUL_FFT_MODF_THRESHOLD 595 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 595, 5}, { 28, 6}, { 15, 5}, { 31, 6}, \ + { 17, 5}, { 35, 6}, { 28, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 36, 7}, { 19, 6}, \ + { 39, 7}, { 23, 6}, { 47, 7}, { 27, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 43, 8}, \ + { 23, 7}, { 49, 8}, { 27, 9}, { 15, 8}, \ + { 31, 7}, { 63, 8}, { 43, 9}, { 23, 8}, \ + { 51, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 79, 9}, { 47, 8}, { 99, 9}, { 55,10}, \ + { 31, 9}, { 63, 8}, { 127, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95, 9}, \ + { 191,11}, { 63, 9}, { 255,10}, { 159,11}, \ + { 95,10}, { 191,12}, { 63,11}, { 127,10}, \ + { 255, 9}, { 511,10}, { 271, 9}, { 543,11}, \ + { 159,10}, { 335,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 399,12}, { 127,11}, { 255,10}, \ + { 511, 9}, { 1023,10}, { 543, 9}, { 1087,11}, \ + { 287,10}, { 607,11}, { 319,10}, { 671,12}, \ + { 191,11}, { 383,10}, { 767,13}, { 127,12}, \ + { 255,11}, { 511,10}, { 1023,11}, { 543,10}, \ + { 1119,11}, { 607,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,10}, { 1471,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,10}, { 1727,12}, \ + { 447,11}, { 959,13}, { 255,12}, { 511,11}, \ + { 1119,12}, { 575,11}, { 1215,10}, { 2431,12}, \ + { 639,11}, { 1343,12}, { 703,11}, { 1471,10}, \ + { 2943,13}, { 383,12}, { 767,11}, { 1599,12}, \ + { 831,11}, { 1727,10}, { 3455,12}, { 959,14}, \ + { 255,13}, { 511,12}, { 1087,11}, { 2239,10}, \ + { 4479,12}, { 1215,11}, { 2431,13}, { 639,12}, \ + { 1471,11}, { 2943,13}, { 767,12}, { 1727,11}, \ + { 3455,13}, { 895,12}, { 1983,14}, { 511,13}, \ + { 1023,12}, { 2239,11}, { 4479,13}, { 1151,12}, \ + { 2431,13}, { 1279,12}, { 2559,13}, { 1407,12}, \ + { 2943,11}, { 5887,14}, { 767,13}, { 1663,12}, \ + { 3455,13}, { 1919,12}, { 3839,15}, { 511,14}, \ + { 1023,13}, { 2175,12}, { 4479,13}, { 2431,14}, \ + { 1279,13}, { 2943,12}, { 5887,14}, { 1535,13}, \ + { 3455,14}, { 1791,13}, { 3967,15}, { 1023,14}, \ + { 2047,13}, { 4479,14}, { 2303,13}, { 4991,12}, \ + { 9983,14}, { 2815,13}, { 6015,15}, { 1535,14}, \ + { 3839,13}, { 7679,16} } +#define MUL_FFT_TABLE3_SIZE 170 +#define MUL_FFT_THRESHOLD 6784 + +#define SQR_FFT_MODF_THRESHOLD 525 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 525, 5}, { 29, 6}, { 15, 5}, { 33, 6}, \ + { 17, 5}, { 35, 6}, { 29, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 23, 6}, { 47, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 49, 8}, { 27, 7}, { 55, 9}, \ + { 15, 8}, { 31, 7}, { 63, 8}, { 39, 9}, \ + { 23, 8}, { 55, 9}, { 31, 8}, { 67, 9}, \ + { 39, 8}, { 79, 9}, { 47, 8}, { 95, 9}, \ + { 55,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 135,10}, \ + { 79, 9}, { 159,10}, { 95,11}, { 63,10}, \ + { 143, 9}, { 287,10}, { 159, 6}, { 2687, 7}, \ + { 1407, 9}, { 367, 8}, { 735, 9}, { 383,10}, \ + { 207, 9}, { 415,11}, { 127,10}, { 271, 9}, \ + { 543,10}, { 287,11}, { 159,10}, { 319, 9}, \ + { 639,10}, { 335, 9}, { 671,10}, { 351,11}, \ + { 191,10}, { 383, 9}, { 767,10}, { 399, 9}, \ + { 799,10}, { 415,12}, { 127,11}, { 255,10}, \ + { 511, 9}, { 1023,10}, { 543,11}, { 287,10}, \ + { 607,11}, { 319,10}, { 671,11}, { 351,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,13}, \ + { 127,12}, { 255,11}, { 511,10}, { 1023,11}, \ + { 543,10}, { 1087,11}, { 607,10}, { 1215,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 735,10}, \ + { 1471,12}, { 383,11}, { 799,10}, { 1599,11}, \ + { 863,10}, { 1727,12}, { 447,11}, { 991,10}, \ + { 1983,13}, { 255,12}, { 511,11}, { 1119,12}, \ + { 575,11}, { 1215,10}, { 2431,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,10}, \ + { 3455,12}, { 895,11}, { 1791,12}, { 959,11}, \ + { 1983,14}, { 255,13}, { 511,12}, { 1023,11}, \ + { 2047,12}, { 1087,11}, { 2239,12}, { 1215,11}, \ + { 2431,13}, { 639,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1727,11}, { 3455,13}, { 895,12}, \ + { 1983,11}, { 3967,14}, { 511,13}, { 1023,12}, \ + { 2239,13}, { 1151,12}, { 2495,13}, { 1279,12}, \ + { 2623,13}, { 1407,12}, { 2943,14}, { 767,13}, \ + { 1663,12}, { 3455,13}, { 1919,12}, { 3967,15}, \ + { 511,14}, { 1023,13}, { 2175,12}, { 4479,13}, \ + { 2431,12}, { 4863,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3967,12}, { 7935,15}, { 1023,14}, { 2047,13}, \ + { 4479,14}, { 2303,13}, { 4991,12}, { 9983,14}, \ + { 2815,13}, { 5887,15}, { 1535,14}, { 3327,13}, \ + { 6655,14}, { 3839,13}, { 7935,16} } +#define SQR_FFT_TABLE3_SIZE 187 +#define SQR_FFT_THRESHOLD 5312 + +#define MULLO_BASECASE_THRESHOLD 0 /* always */ +#define MULLO_DC_THRESHOLD 43 +#define MULLO_MUL_N_THRESHOLD 13463 +#define SQRLO_BASECASE_THRESHOLD 9 +#define SQRLO_DC_THRESHOLD 42 +#define SQRLO_SQR_THRESHOLD 10323 + +#define DC_DIV_QR_THRESHOLD 43 +#define DC_DIVAPPR_Q_THRESHOLD 132 +#define DC_BDIV_QR_THRESHOLD 83 +#define DC_BDIV_Q_THRESHOLD 130 + +#define INV_MULMOD_BNM1_THRESHOLD 46 +#define INV_NEWTON_THRESHOLD 189 +#define INV_APPR_THRESHOLD 167 + +#define BINV_NEWTON_THRESHOLD 372 +#define REDC_1_TO_REDC_N_THRESHOLD 83 + +#define MU_DIV_QR_THRESHOLD 1589 +#define MU_DIVAPPR_Q_THRESHOLD 1589 +#define MUPI_DIV_QR_THRESHOLD 97 +#define MU_BDIV_QR_THRESHOLD 1589 +#define MU_BDIV_Q_THRESHOLD 1718 + +#define POWM_SEC_TABLE 1,28,96,473,803 + +#define GET_STR_DC_THRESHOLD 12 +#define GET_STR_PRECOMPUTE_THRESHOLD 16 +#define SET_STR_DC_THRESHOLD 145 +#define SET_STR_PRECOMPUTE_THRESHOLD 419 + +#define FAC_DSC_THRESHOLD 114 +#define FAC_ODD_THRESHOLD 57 + +#define MATRIX22_STRASSEN_THRESHOLD 20 +#define HGCD2_DIV1_METHOD 1 /* 1.03% faster than 3 */ +#define HGCD_THRESHOLD 117 +#define HGCD_APPR_THRESHOLD 137 +#define HGCD_REDUCE_THRESHOLD 3524 +#define GCD_DC_THRESHOLD 389 +#define GCDEXT_DC_THRESHOLD 318 +#define JACOBI_BASE_METHOD 4 /* 6.10% faster than 1 */ + +/* Tuneup completed successfully, took 67994 seconds */ diff --git a/gmp-6.3.0/mpn/x86/coreisbr/gmp-mparam.h b/gmp-6.3.0/mpn/x86/coreisbr/gmp-mparam.h new file mode 100644 index 0000000..23d708a --- /dev/null +++ b/gmp-6.3.0/mpn/x86/coreisbr/gmp-mparam.h @@ -0,0 +1,215 @@ +/* x86/coreisbr gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 3400-3800 MHz Intel Xeon E3-1270 Sandy Bridge */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-24, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 28 +#define MOD_1_UNNORM_THRESHOLD 26 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 9 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 7 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 11 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 4 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 2 /* 88.29% faster than 1 */ +#define DIV_QR_1_NORM_THRESHOLD 21 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD 14 +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 20 + +#define DIV_1_VS_MUL_1_PERCENT 297 + +#define MUL_TOOM22_THRESHOLD 32 +#define MUL_TOOM33_THRESHOLD 105 +#define MUL_TOOM44_THRESHOLD 190 +#define MUL_TOOM6H_THRESHOLD 294 +#define MUL_TOOM8H_THRESHOLD 478 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 109 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 144 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 116 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 129 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 160 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 48 +#define SQR_TOOM3_THRESHOLD 163 +#define SQR_TOOM4_THRESHOLD 250 +#define SQR_TOOM6_THRESHOLD 354 +#define SQR_TOOM8_THRESHOLD 502 + +#define MULMID_TOOM42_THRESHOLD 98 + +#define MULMOD_BNM1_THRESHOLD 19 +#define SQRMOD_BNM1_THRESHOLD 23 + +#define MUL_FFT_MODF_THRESHOLD 666 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 666, 5}, { 28, 6}, { 15, 5}, { 31, 6}, \ + { 28, 7}, { 15, 6}, { 33, 7}, { 17, 6}, \ + { 36, 7}, { 19, 6}, { 39, 7}, { 23, 6}, \ + { 47, 7}, { 29, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 49, 8}, \ + { 27, 7}, { 55, 8}, { 31, 7}, { 63, 8}, \ + { 43, 9}, { 23, 8}, { 55, 9}, { 31, 8}, \ + { 71, 9}, { 39, 8}, { 79, 9}, { 47, 8}, \ + { 99, 9}, { 55,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95, 9}, \ + { 191,11}, { 63,10}, { 159, 7}, { 1343, 8}, \ + { 703, 9}, { 367, 8}, { 735, 9}, { 383,10}, \ + { 207,11}, { 127,10}, { 255, 9}, { 511,10}, \ + { 271, 9}, { 543,10}, { 287,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335,11}, { 191,10}, \ + { 383, 9}, { 767,11}, { 223,12}, { 127,11}, \ + { 255,10}, { 543,11}, { 287,10}, { 607, 9}, \ + { 1215,11}, { 319,10}, { 671,12}, { 191,11}, \ + { 383,10}, { 799,13}, { 127,12}, { 255,11}, \ + { 511,10}, { 1023,11}, { 543,10}, { 1087,11}, \ + { 607,10}, { 1215,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,10}, { 1471,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,12}, { 447,11}, \ + { 959,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1471,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1727,12}, { 959,14}, \ + { 255,13}, { 511,12}, { 1087,11}, { 2239,12}, \ + { 1215,13}, { 639,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1727,11}, { 3455,13}, { 895,12}, \ + { 1983,14}, { 511,13}, { 1023,12}, { 2239,13}, \ + { 1151,12}, { 2495,13}, { 1279,12}, { 2623,13}, \ + { 1407,12}, { 2943,14}, { 767,13}, { 1535,12}, \ + { 3071,13}, { 1663,12}, { 3455,13}, { 1919,15}, \ + { 511,14}, { 1023,13}, { 2175,12}, { 4479,13}, \ + { 2431,14}, { 1279,13}, { 2943,12}, { 5887,14}, \ + { 1535,13}, { 3455,14}, { 1791,13}, { 3967,15}, \ + { 1023,14}, { 2047,13}, { 4479,14}, { 2303,13}, \ + { 4991,12}, { 9983,14}, { 2815,13}, { 5887,15}, \ + { 1535,14}, { 3839,13}, { 7679,16} } +#define MUL_FFT_TABLE3_SIZE 163 +#define MUL_FFT_THRESHOLD 7552 + +#define SQR_FFT_MODF_THRESHOLD 570 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 570, 5}, { 28, 6}, { 15, 5}, { 32, 6}, \ + { 17, 5}, { 35, 6}, { 29, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 36, 7}, { 19, 6}, \ + { 40, 7}, { 23, 6}, { 47, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 43, 8}, \ + { 23, 7}, { 49, 8}, { 27, 7}, { 55, 8}, \ + { 31, 7}, { 63, 8}, { 43, 9}, { 23, 8}, \ + { 55, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 79, 9}, { 47, 8}, { 95, 9}, { 55,10}, \ + { 31, 9}, { 79,10}, { 47, 9}, { 95,11}, \ + { 31,10}, { 63, 9}, { 135,10}, { 79, 9}, \ + { 159,10}, { 95,11}, { 63,10}, { 159,11}, \ + { 95,10}, { 191,12}, { 63, 8}, { 1023, 9}, \ + { 543,11}, { 159,10}, { 319, 9}, { 639,10}, \ + { 335,11}, { 191,10}, { 383, 9}, { 767,10}, \ + { 399, 9}, { 799,12}, { 127,11}, { 255,10}, \ + { 511, 9}, { 1023,10}, { 543,11}, { 287,10}, \ + { 607,11}, { 319,10}, { 671,11}, { 351,12}, \ + { 191,11}, { 383,10}, { 799,13}, { 127,12}, \ + { 255,11}, { 511,10}, { 1023,11}, { 543,10}, \ + { 1087,11}, { 607,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,10}, { 1471,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,12}, { 447,11}, \ + { 991,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1471,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1727,12}, { 959,11}, \ + { 1919,14}, { 255,13}, { 511,12}, { 1023,11}, \ + { 2047,12}, { 1087,11}, { 2239,12}, { 1215,11}, \ + { 2431,13}, { 639,12}, { 1471,13}, { 767,12}, \ + { 1727,13}, { 895,12}, { 1983,14}, { 511,13}, \ + { 1023,12}, { 2239,13}, { 1151,12}, { 2495,13}, \ + { 1279,12}, { 2623,13}, { 1407,12}, { 2943,14}, \ + { 767,13}, { 1663,12}, { 3455,13}, { 1919,12}, \ + { 3967,15}, { 511,14}, { 1023,13}, { 2175,12}, \ + { 4479,13}, { 2431,12}, { 4863,14}, { 1279,13}, \ + { 2943,12}, { 5887,14}, { 1535,13}, { 3455,14}, \ + { 1791,13}, { 3967,15}, { 1023,14}, { 2047,13}, \ + { 4479,14}, { 2303,13}, { 4991,12}, { 9983,14}, \ + { 2559,13}, { 5119,14}, { 2815,13}, { 5887,15}, \ + { 1535,14}, { 3839,13}, { 7679,16} } +#define SQR_FFT_TABLE3_SIZE 163 +#define SQR_FFT_THRESHOLD 5760 + +#define MULLO_BASECASE_THRESHOLD 16 +#define MULLO_DC_THRESHOLD 46 +#define MULLO_MUL_N_THRESHOLD 14281 +#define SQRLO_BASECASE_THRESHOLD 0 /* always */ +#define SQRLO_DC_THRESHOLD 159 +#define SQRLO_SQR_THRESHOLD 11317 + +#define DC_DIV_QR_THRESHOLD 47 +#define DC_DIVAPPR_Q_THRESHOLD 191 +#define DC_BDIV_QR_THRESHOLD 107 +#define DC_BDIV_Q_THRESHOLD 232 + +#define INV_MULMOD_BNM1_THRESHOLD 62 +#define INV_NEWTON_THRESHOLD 181 +#define INV_APPR_THRESHOLD 182 + +#define BINV_NEWTON_THRESHOLD 378 +#define REDC_1_TO_REDC_N_THRESHOLD 91 + +#define MU_DIV_QR_THRESHOLD 1858 +#define MU_DIVAPPR_Q_THRESHOLD 1858 +#define MUPI_DIV_QR_THRESHOLD 77 +#define MU_BDIV_QR_THRESHOLD 1830 +#define MU_BDIV_Q_THRESHOLD 2166 + +#define POWM_SEC_TABLE 1,16,126,428,1442 + +#define GET_STR_DC_THRESHOLD 10 +#define GET_STR_PRECOMPUTE_THRESHOLD 16 +#define SET_STR_DC_THRESHOLD 418 +#define SET_STR_PRECOMPUTE_THRESHOLD 1104 + +#define FAC_DSC_THRESHOLD 149 +#define FAC_ODD_THRESHOLD 34 + +#define MATRIX22_STRASSEN_THRESHOLD 21 +#define HGCD2_DIV1_METHOD 1 /* 5.54% faster than 4 */ +#define HGCD_THRESHOLD 66 +#define HGCD_APPR_THRESHOLD 135 +#define HGCD_REDUCE_THRESHOLD 4284 +#define GCD_DC_THRESHOLD 642 +#define GCDEXT_DC_THRESHOLD 465 +#define JACOBI_BASE_METHOD 3 /* 14.76% faster than 4 */ + +/* Tuneup completed successfully, took 44241 seconds */ diff --git a/gmp-6.3.0/mpn/x86/darwin.m4 b/gmp-6.3.0/mpn/x86/darwin.m4 new file mode 100644 index 0000000..c449216 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/darwin.m4 @@ -0,0 +1,102 @@ +divert(-1) +dnl Copyright 2007, 2011, 2012, 2014 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +define(`DARWIN') + + +dnl Usage LEA(symbol,reg) +dnl Usage LEAL(symbol_local_to_file,reg) +dnl +dnl We maintain lists of stuff to append in load_eip and darwin_bd. The +dnl `index' stuff is needed to suppress repeated definitions. To avoid +dnl getting fooled by "var" and "var1", we add 'bol ' (the end of +dnl 'indirect_symbol') at the beginning and and a newline at the end. This +dnl might be a bit fragile. + +define(`LEA', +m4_assert_numargs(2) +`ifdef(`PIC',` +ifelse(index(defn(`load_eip'), `$2'),-1, +`m4append(`load_eip', +` TEXT + ALIGN(16) +L(movl_eip_`'substr($2,1)): + movl (%esp), $2 + ret_internal +')') +ifelse(index(defn(`darwin_bd'), `bol $1 +'),-1, +`m4append(`darwin_bd', +` .section __IMPORT,__pointers,non_lazy_symbol_pointers +L($1`'$non_lazy_ptr): + .indirect_symbol $1 + .long 0 +')') + call L(movl_eip_`'substr($2,1)) + movl L($1`'$non_lazy_ptr)-.($2), $2 +',` + movl `$'$1, $2 +')') + +define(`LEAL', +m4_assert_numargs(2) +`ifdef(`PIC',` +ifelse(index(defn(`load_eip'), `$2'),-1, +`m4append(`load_eip', +` TEXT + ALIGN(16) +L(movl_eip_`'substr($2,1)): + movl (%esp), $2 + ret_internal +')') + call L(movl_eip_`'substr($2,1)) + leal $1-.($2), $2 +',` + movl `$'$1, $2 +')') + + +dnl ASM_END + +define(`ASM_END',`load_eip`'darwin_bd') + +define(`load_eip', `') dnl updated in LEA +define(`darwin_bd', `') dnl updated in LEA + + +dnl Usage: CALL(funcname) +dnl + +define(`CALL', +m4_assert_numargs(1) +`call GSYM_PREFIX`'$1') + +undefine(`PIC_WITH_EBX') + +divert`'dnl diff --git a/gmp-6.3.0/mpn/x86/dive_1.asm b/gmp-6.3.0/mpn/x86/dive_1.asm new file mode 100644 index 0000000..5bb0f45 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/dive_1.asm @@ -0,0 +1,190 @@ +dnl x86 mpn_divexact_1 -- mpn by limb exact division. + +dnl Copyright 2001, 2002, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P54 30.0 +C P55 29.0 +C P6 13.0 odd divisor, 12.0 even (strangely) +C K6 14.0 +C K7 12.0 +C P4 42.0 + + +C mp_limb_t mpn_divexact_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C + +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(VAR_INVERSE,`PARAM_SRC') + + TEXT + + ALIGN(16) +PROLOGUE(mpn_divexact_1) +deflit(`FRAME',0) + + movl PARAM_DIVISOR, %eax + pushl %ebp FRAME_pushl() + + movl PARAM_SIZE, %ebp + pushl %edi FRAME_pushl() + + pushl %ebx FRAME_pushl() + movl $-1, %ecx C shift count + + pushl %esi FRAME_pushl() + +L(strip_twos): + incl %ecx + + shrl %eax + jnc L(strip_twos) + + leal 1(%eax,%eax), %ebx C d without twos + andl $127, %eax C d/2, 7 bits + +ifdef(`PIC',` + LEA( binvert_limb_table, %edx) + movzbl (%eax,%edx), %eax C inv 8 bits +',` + movzbl binvert_limb_table(%eax), %eax C inv 8 bits +') + + leal (%eax,%eax), %edx C 2*inv + movl %ebx, PARAM_DIVISOR C d without twos + + imull %eax, %eax C inv*inv + + movl PARAM_SRC, %esi + movl PARAM_DST, %edi + + imull %ebx, %eax C inv*inv*d + + subl %eax, %edx C inv = 2*inv - inv*inv*d + leal (%edx,%edx), %eax C 2*inv + + imull %edx, %edx C inv*inv + + leal (%esi,%ebp,4), %esi C src end + leal (%edi,%ebp,4), %edi C dst end + negl %ebp C -size + + imull %ebx, %edx C inv*inv*d + + subl %edx, %eax C inv = 2*inv - inv*inv*d + + ASSERT(e,` C expect d*inv == 1 mod 2^GMP_LIMB_BITS + pushl %eax FRAME_pushl() + imull PARAM_DIVISOR, %eax + cmpl $1, %eax + popl %eax FRAME_popl()') + + movl %eax, VAR_INVERSE + movl (%esi,%ebp,4), %eax C src[0] + + xorl %ebx, %ebx + xorl %edx, %edx + + incl %ebp + jz L(one) + + movl (%esi,%ebp,4), %edx C src[1] + + shrdl( %cl, %edx, %eax) + + movl VAR_INVERSE, %edx + jmp L(entry) + + + ALIGN(8) + nop C k6 code alignment + nop +L(top): + C eax q + C ebx carry bit, 0 or -1 + C ecx shift + C edx carry limb + C esi src end + C edi dst end + C ebp counter, limbs, negative + + movl -4(%esi,%ebp,4), %eax + subl %ebx, %edx C accumulate carry bit + + movl (%esi,%ebp,4), %ebx + + shrdl( %cl, %ebx, %eax) + + subl %edx, %eax C apply carry limb + movl VAR_INVERSE, %edx + + sbbl %ebx, %ebx + +L(entry): + imull %edx, %eax + + movl %eax, -4(%edi,%ebp,4) + movl PARAM_DIVISOR, %edx + + mull %edx + + incl %ebp + jnz L(top) + + + movl -4(%esi), %eax C src high limb +L(one): + shrl %cl, %eax + popl %esi FRAME_popl() + + addl %ebx, %eax C apply carry bit + popl %ebx FRAME_popl() + + subl %edx, %eax C apply carry limb + + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi) + + popl %edi + popl %ebp + + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/divrem_1.asm b/gmp-6.3.0/mpn/x86/divrem_1.asm new file mode 100644 index 0000000..255d493 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/divrem_1.asm @@ -0,0 +1,233 @@ +dnl x86 mpn_divrem_1 -- mpn by limb division extending to fractional quotient. + +dnl Copyright 1999-2003, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C 486 approx 43 maybe +C P5 44 +C P6 39 +C P6MMX 39 +C K6 22 +C K7 42 +C P4 58 + + +C mp_limb_t mpn_divrem_1 (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, mp_limb_t divisor); +C mp_limb_t mpn_divrem_1c (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, mp_limb_t divisor, +C mp_limb_t carry); +C +C Divide src,size by divisor and store the quotient in dst+xsize,size. +C Extend the division to fractional quotient limbs in dst,xsize. Return the +C remainder. Either or both xsize and size can be 0. +C +C mpn_divrem_1c takes a carry parameter which is an initial high limb, +C effectively one extra limb at the top of src,size. Must have +C carry /* for printf */ +#include /* for getenv */ +#include + +#include "gmp-impl.h" + +/* Change this to "#define TRACE(x) x" for some traces. */ +#define TRACE(x) + + +/* fat_entry.asm */ +long __gmpn_cpuid (char [12], int); +int __gmpn_cpuid_available (void); + + +#if WANT_FAKE_CPUID +/* The "name"s in the table are values for the GMP_CPU_TYPE environment + variable. Anything can be used, but for now it's the canonical cpu types + as per config.guess/config.sub. */ + +#define __gmpn_cpuid fake_cpuid +#define __gmpn_cpuid_available fake_cpuid_available + +#define MAKE_FMS(family, model) \ + ((((family) & 0xf) << 8) + (((family) & 0xff0) << 20) \ + + (((model) & 0xf) << 4) + (((model) & 0xf0) << 12)) + +static struct { + const char *name; + const char *vendor; + unsigned fms; +} fake_cpuid_table[] = { + { "i386", "" }, + { "i486", "GenuineIntel", MAKE_FMS (4, 0) }, + { "pentium", "GenuineIntel", MAKE_FMS (5, 0) }, + { "pentiummmx", "GenuineIntel", MAKE_FMS (5, 4) }, + { "pentiumpro", "GenuineIntel", MAKE_FMS (6, 0) }, + { "pentium2", "GenuineIntel", MAKE_FMS (6, 2) }, + { "pentium3", "GenuineIntel", MAKE_FMS (6, 7) }, + { "pentium4", "GenuineIntel", MAKE_FMS (15, 2) }, + { "prescott", "GenuineIntel", MAKE_FMS (15, 3) }, + { "nocona", "GenuineIntel", MAKE_FMS (15, 4) }, + { "core2", "GenuineIntel", MAKE_FMS (6, 0xf) }, + { "nehalem", "GenuineIntel", MAKE_FMS (6, 0x1a) }, + { "nhm", "GenuineIntel", MAKE_FMS (6, 0x1a) }, + { "atom", "GenuineIntel", MAKE_FMS (6, 0x1c) }, + { "westmere", "GenuineIntel", MAKE_FMS (6, 0x25) }, + { "wsm", "GenuineIntel", MAKE_FMS (6, 0x25) }, + { "sandybridge","GenuineIntel", MAKE_FMS (6, 0x2a) }, + { "sbr", "GenuineIntel", MAKE_FMS (6, 0x2a) }, + { "silvermont", "GenuineIntel", MAKE_FMS (6, 0x37) }, + { "slm", "GenuineIntel", MAKE_FMS (6, 0x37) }, + { "haswell", "GenuineIntel", MAKE_FMS (6, 0x3c) }, + { "hwl", "GenuineIntel", MAKE_FMS (6, 0x3c) }, + { "broadwell", "GenuineIntel", MAKE_FMS (6, 0x3d) }, + { "bwl", "GenuineIntel", MAKE_FMS (6, 0x3d) }, + { "skylake", "GenuineIntel", MAKE_FMS (6, 0x5e) }, + { "sky", "GenuineIntel", MAKE_FMS (6, 0x5e) }, + + { "k5", "AuthenticAMD", MAKE_FMS (5, 0) }, + { "k6", "AuthenticAMD", MAKE_FMS (5, 3) }, + { "k62", "AuthenticAMD", MAKE_FMS (5, 8) }, + { "k63", "AuthenticAMD", MAKE_FMS (5, 9) }, + { "athlon", "AuthenticAMD", MAKE_FMS (6, 0) }, + { "k8", "AuthenticAMD", MAKE_FMS (15, 0) }, + { "k10", "AuthenticAMD", MAKE_FMS (16, 0) }, + { "bobcat", "AuthenticAMD", MAKE_FMS (20, 1) }, + { "bulldozer", "AuthenticAMD", MAKE_FMS (21, 1) }, + { "piledriver", "AuthenticAMD", MAKE_FMS (21, 2) }, + { "steamroller","AuthenticAMD", MAKE_FMS (21, 0x30) }, + { "excavator", "AuthenticAMD", MAKE_FMS (21, 0x60) }, + { "jaguar", "AuthenticAMD", MAKE_FMS (22, 1) }, + { "zen", "AuthenticAMD", MAKE_FMS (23, 1) }, + + { "viac3", "CentaurHauls", MAKE_FMS (6, 0) }, + { "viac32", "CentaurHauls", MAKE_FMS (6, 9) }, + { "nano", "CentaurHauls", MAKE_FMS (6, 15) }, +}; + +static int +fake_cpuid_lookup (void) +{ + char *s; + int i; + + s = getenv ("GMP_CPU_TYPE"); + if (s == NULL) + { + printf ("Need GMP_CPU_TYPE environment variable for fake cpuid\n"); + abort (); + } + + for (i = 0; i < numberof (fake_cpuid_table); i++) + if (strcmp (s, fake_cpuid_table[i].name) == 0) + return i; + + printf ("GMP_CPU_TYPE=%s unknown\n", s); + abort (); +} + +static int +fake_cpuid_available (void) +{ + return fake_cpuid_table[fake_cpuid_lookup()].vendor[0] != '\0'; +} + +static long +fake_cpuid (char dst[12], int id) +{ + int i = fake_cpuid_lookup(); + + switch (id) { + case 0: + memcpy (dst, fake_cpuid_table[i].vendor, 12); + return 0; + case 1: + return fake_cpuid_table[i].fms; + default: + printf ("fake_cpuid(): oops, unknown id %d\n", id); + abort (); + } +} +#endif + + +typedef DECL_preinv_divrem_1 ((*preinv_divrem_1_t)); +typedef DECL_preinv_mod_1 ((*preinv_mod_1_t)); + +struct cpuvec_t __gmpn_cpuvec = { + __MPN(add_n_init), + 0, + 0, + __MPN(addmul_1_init), + 0, + __MPN(bdiv_dbm1c_init), + __MPN(cnd_add_n_init), + __MPN(cnd_sub_n_init), + __MPN(com_init), + __MPN(copyd_init), + __MPN(copyi_init), + __MPN(divexact_1_init), + __MPN(divrem_1_init), + __MPN(gcd_11_init), + __MPN(lshift_init), + __MPN(lshiftc_init), + __MPN(mod_1_init), + __MPN(mod_1_1p_init), + __MPN(mod_1_1p_cps_init), + __MPN(mod_1s_2p_init), + __MPN(mod_1s_2p_cps_init), + __MPN(mod_1s_4p_init), + __MPN(mod_1s_4p_cps_init), + __MPN(mod_34lsub1_init), + __MPN(modexact_1c_odd_init), + __MPN(mul_1_init), + __MPN(mul_basecase_init), + __MPN(mullo_basecase_init), + __MPN(preinv_divrem_1_init), + __MPN(preinv_mod_1_init), + __MPN(redc_1_init), + __MPN(redc_2_init), + __MPN(rshift_init), + __MPN(sqr_basecase_init), + __MPN(sub_n_init), + 0, + __MPN(submul_1_init), + 0 +}; + +int __gmpn_cpuvec_initialized = 0; + +/* The following setups start with generic x86, then overwrite with + specifics for a chip, and higher versions of that chip. + + The arrangement of the setups here will normally be the same as the $path + selections in configure.in for the respective chips. + + This code is reentrant and thread safe. We always calculate the same + decided_cpuvec, so if two copies of the code are running it doesn't + matter which completes first, both write the same to __gmpn_cpuvec. + + We need to go via decided_cpuvec because if one thread has completed + __gmpn_cpuvec then it may be making use of the threshold values in that + vector. If another thread is still running __gmpn_cpuvec_init then we + don't want it to write different values to those fields since some of the + asm routines only operate correctly up to their own defined threshold, + not an arbitrary value. */ + +void +__gmpn_cpuvec_init (void) +{ + struct cpuvec_t decided_cpuvec; + + TRACE (printf ("__gmpn_cpuvec_init:\n")); + + memset (&decided_cpuvec, '\0', sizeof (decided_cpuvec)); + + CPUVEC_SETUP_x86; + CPUVEC_SETUP_fat; + + if (! __gmpn_cpuid_available ()) + { + TRACE (printf (" 80386, or early 80486 without cpuid\n")); + } + else + { + char vendor_string[13]; + char dummy_string[12]; + long fms; + int family, model; + + __gmpn_cpuid (vendor_string, 0); + vendor_string[12] = 0; + + fms = __gmpn_cpuid (dummy_string, 1); + family = ((fms >> 8) & 0xf) + ((fms >> 20) & 0xff); + model = ((fms >> 4) & 0xf) + ((fms >> 12) & 0xf0); + + if (strcmp (vendor_string, "GenuineIntel") == 0) + { + switch (family) + { + case 4: + TRACE (printf (" 80486 with cpuid\n")); + break; + + case 5: + TRACE (printf (" pentium\n")); + CPUVEC_SETUP_pentium; + if (model == 4 || model == 8) + { + TRACE (printf (" pentiummmx\n")); + CPUVEC_SETUP_pentium_mmx; + } + break; + + case 6: + TRACE (printf (" p6\n")); + CPUVEC_SETUP_p6; + switch (model) + { + case 0x00: + case 0x01: + TRACE (printf (" pentiumpro\n")); + break; + + case 0x02: + case 0x03: + case 0x04: + case 0x05: + case 0x06: + TRACE (printf (" pentium2\n")); + CPUVEC_SETUP_p6_mmx; + break; + + case 0x07: + case 0x08: + case 0x0a: + case 0x0b: + case 0x0c: + TRACE (printf (" pentium3\n")); + CPUVEC_SETUP_p6_mmx; + CPUVEC_SETUP_p6_p3mmx; + break; + + case 0x09: /* Banias */ + case 0x0d: /* Dothan */ + case 0x0e: /* Yonah */ + TRACE (printf (" Banias/Dothan/Yonah\n")); + CPUVEC_SETUP_p6_mmx; + CPUVEC_SETUP_p6_p3mmx; + CPUVEC_SETUP_p6_sse2; + break; + + case 0x0f: /* Conroe Merom Kentsfield Allendale */ + case 0x10: + case 0x11: + case 0x12: + case 0x13: + case 0x14: + case 0x15: + case 0x16: + case 0x17: /* PNR Wolfdale Yorkfield */ + case 0x18: + case 0x19: + case 0x1d: /* PNR Dunnington */ + TRACE (printf (" Conroe\n")); + CPUVEC_SETUP_p6_mmx; + CPUVEC_SETUP_p6_p3mmx; + CPUVEC_SETUP_p6_sse2; + CPUVEC_SETUP_core2; + break; + + case 0x1c: /* Atom Silverthorne */ + case 0x26: /* Atom Lincroft */ + case 0x27: /* Atom Saltwell */ + case 0x36: /* Atom Cedarview/Saltwell */ + TRACE (printf (" atom\n")); + CPUVEC_SETUP_atom; + CPUVEC_SETUP_atom_mmx; + CPUVEC_SETUP_atom_sse2; + break; + + case 0x37: /* Silvermont */ + case 0x4a: /* Silvermont */ + case 0x4c: /* Airmont */ + case 0x4d: /* Silvermont/Avoton */ + case 0x5a: /* Silvermont */ + TRACE (printf (" silvermont\n")); + CPUVEC_SETUP_atom; + CPUVEC_SETUP_atom_mmx; + CPUVEC_SETUP_atom_sse2; + CPUVEC_SETUP_silvermont; + break; + + case 0x5c: /* Goldmont */ + case 0x5f: /* Goldmont */ + case 0x7a: /* Goldmont Plus */ + TRACE (printf (" goldmont\n")); + CPUVEC_SETUP_atom; + CPUVEC_SETUP_atom_mmx; + CPUVEC_SETUP_atom_sse2; + CPUVEC_SETUP_goldmont; + break; + + case 0x1a: /* NHM Gainestown */ + case 0x1b: + case 0x1e: /* NHM Lynnfield/Jasper */ + case 0x1f: + case 0x20: + case 0x21: + case 0x22: + case 0x23: + case 0x24: + case 0x25: /* WSM Clarkdale/Arrandale */ + case 0x28: + case 0x29: + case 0x2b: + case 0x2c: /* WSM Gulftown */ + case 0x2e: /* NHM Beckton */ + case 0x2f: /* WSM Eagleton */ + TRACE (printf (" nehalem/westmere\n")); + CPUVEC_SETUP_p6_mmx; + CPUVEC_SETUP_p6_p3mmx; + CPUVEC_SETUP_p6_sse2; + CPUVEC_SETUP_core2; + CPUVEC_SETUP_coreinhm; + break; + + case 0x2a: /* SBR */ + case 0x2d: /* SBR-EP */ + case 0x3a: /* IBR */ + case 0x3e: /* IBR Ivytown */ + case 0x3c: /* Haswell client */ + case 0x3f: /* Haswell server */ + case 0x45: /* Haswell ULT */ + case 0x46: /* Crystal Well */ + case 0x3d: /* Broadwell */ + case 0x47: /* Broadwell */ + case 0x4f: /* Broadwell server */ + case 0x56: /* Broadwell microserver */ + case 0x4e: /* Skylake client */ + case 0x55: /* Skylake server */ + case 0x5e: /* Skylake */ + case 0x8e: /* Kabylake */ + case 0x9e: /* Kabylake */ + TRACE (printf (" sandybridge\n")); + CPUVEC_SETUP_p6_mmx; + CPUVEC_SETUP_p6_p3mmx; + CPUVEC_SETUP_p6_sse2; + CPUVEC_SETUP_core2; + CPUVEC_SETUP_coreinhm; + CPUVEC_SETUP_coreisbr; + break; + } + break; + + case 15: + TRACE (printf (" pentium4\n")); + CPUVEC_SETUP_pentium4; + CPUVEC_SETUP_pentium4_mmx; + CPUVEC_SETUP_pentium4_sse2; + break; + } + } + else if (strcmp (vendor_string, "AuthenticAMD") == 0) + { + switch (family) + { + case 5: + if (model <= 3) + { + TRACE (printf (" k5\n")); + } + else + { + TRACE (printf (" k6\n")); + CPUVEC_SETUP_k6; + CPUVEC_SETUP_k6_mmx; + if (model >= 8) + { + TRACE (printf (" k62\n")); + CPUVEC_SETUP_k6_k62mmx; + } + if (model >= 9) + { + TRACE (printf (" k63\n")); + } + } + break; + case 6: + TRACE (printf (" athlon\n")); + CPUVEC_SETUP_k7; + CPUVEC_SETUP_k7_mmx; + break; + + case 0x0f: /* k8 */ + case 0x11: /* "fam 11h", mix of k8 and k10 */ + case 0x13: /* unknown, conservatively assume k8 */ + TRACE (printf (" k8\n")); + CPUVEC_SETUP_k7; + CPUVEC_SETUP_k7_mmx; + CPUVEC_SETUP_k8; + break; + + case 0x10: /* k10 */ + case 0x12: /* k10 (llano) */ + TRACE (printf (" k10\n")); + CPUVEC_SETUP_k7; + CPUVEC_SETUP_k7_mmx; + break; + + case 0x14: /* bobcat */ + case 0x16: /* jaguar */ + TRACE (printf (" bobcat\n")); + CPUVEC_SETUP_k7; + CPUVEC_SETUP_k7_mmx; + CPUVEC_SETUP_bt1; + break; + + case 0x15: /* bulldozer */ + TRACE (printf (" bulldozer\n")); + CPUVEC_SETUP_k7; + CPUVEC_SETUP_k7_mmx; + CPUVEC_SETUP_bd1; + break; + + case 0x17: /* zen */ + case 0x19: /* zen3 */ + TRACE (printf (" zen\n")); + CPUVEC_SETUP_k7; + CPUVEC_SETUP_k7_mmx; + break; + } + } + else if (strcmp (vendor_string, "CentaurHauls") == 0) + { + switch (family) + { + case 6: + TRACE (printf (" viac3\n")); + if (model >= 9) + { + TRACE (printf (" viac32\n")); + } + if (model >= 15) + { + TRACE (printf (" nano\n")); + CPUVEC_SETUP_nano; + } + break; + } + } + else if (strcmp (vendor_string, "CyrixInstead") == 0) + { + /* Should recognize Cyrix' processors too. */ + TRACE (printf (" cyrix something\n")); + } + } + + /* There's no x86 generic mpn_preinv_divrem_1 or mpn_preinv_mod_1. + Instead default to the plain versions from whichever CPU we detected. + The function arguments are compatible, no need for any glue code. */ + if (decided_cpuvec.preinv_divrem_1 == NULL) + decided_cpuvec.preinv_divrem_1 =(preinv_divrem_1_t)decided_cpuvec.divrem_1; + if (decided_cpuvec.preinv_mod_1 == NULL) + decided_cpuvec.preinv_mod_1 =(preinv_mod_1_t) decided_cpuvec.mod_1; + + ASSERT_CPUVEC (decided_cpuvec); + CPUVEC_INSTALL (decided_cpuvec); + + /* Set this once the threshold fields are ready. + Use volatile to prevent it getting moved. */ + *((volatile int *) &__gmpn_cpuvec_initialized) = 1; +} diff --git a/gmp-6.3.0/mpn/x86/fat/fat_entry.asm b/gmp-6.3.0/mpn/x86/fat/fat_entry.asm new file mode 100644 index 0000000..25655cf --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/fat_entry.asm @@ -0,0 +1,243 @@ +dnl x86 fat binary entrypoints. + +dnl Copyright 2003, 2012, 2014 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +dnl Forcibly disable profiling. +dnl +dnl The entrypoints and inits are small enough not to worry about, the real +dnl routines arrived at will have any profiling. Also, the way the code +dnl here ends with a jump means we won't work properly with the +dnl "instrument" profiling scheme anyway. + +define(`WANT_PROFILING',no) + + + TEXT + + +dnl Usage: FAT_ENTRY(name, offset) +dnl +dnl Emit a fat binary entrypoint function of the given name. This is the +dnl normal entry for applications, eg. __gmpn_add_n. +dnl +dnl The code simply jumps through the function pointer in __gmpn_cpuvec at +dnl the given "offset" (in bytes). +dnl +dnl For non-PIC, the jumps are 5 bytes each, aligning them to 8 should be +dnl fine for all x86s. +dnl +dnl For PIC, the jumps are 20 bytes each, and are best aligned to 16 to +dnl ensure at least the first two instructions don't cross a cache line +dnl boundary. +dnl +dnl Note the extra `' ahead of PROLOGUE obscures it from the HAVE_NATIVE +dnl grepping in configure, stopping that code trying to eval something with +dnl $1 in it. + +define(FAT_ENTRY, +m4_assert_numargs(2) +` ALIGN(ifdef(`PIC',16,8)) +`'PROLOGUE($1)dnl +ifdef(`PIC',`dnl +ifdef(`DARWIN',` + call L(movl_eip_edx) + movl L(___gmpn_cpuvec)$non_lazy_ptr-.(%edx), %edx + jmp *m4_empty_if_zero($2)(%edx) +',`dnl + call L(movl_eip_edx) +L(entry_here$2): + addl $_GLOBAL_OFFSET_TABLE_+[.-L(entry_here$2)], %edx + movl GSYM_PREFIX`'__gmpn_cpuvec@GOT(%edx), %edx + jmp *m4_empty_if_zero($2)(%edx) +') +',`dnl non-PIC + jmp *GSYM_PREFIX`'__gmpn_cpuvec+$2 +') +EPILOGUE() +') + + +dnl FAT_ENTRY for each CPUVEC_FUNCS_LIST +dnl + +define(`CPUVEC_offset',0) +foreach(i, +`FAT_ENTRY(MPN(i),CPUVEC_offset) +define(`CPUVEC_offset',eval(CPUVEC_offset + 4))', +CPUVEC_FUNCS_LIST) + +ifdef(`PIC',` + ALIGN(8) +L(movl_eip_edx): + movl (%esp), %edx + ret_internal +ifdef(`DARWIN',` + .section __IMPORT,__pointers,non_lazy_symbol_pointers +L(___gmpn_cpuvec)$non_lazy_ptr: + .indirect_symbol ___gmpn_cpuvec + .long 0 + TEXT +') +') + + +dnl Usage: FAT_INIT(name, offset) +dnl +dnl Emit a fat binary initializer function of the given name. These +dnl functions are the initial values for the pointers in __gmpn_cpuvec. +dnl +dnl The code simply calls __gmpn_cpuvec_init, and then jumps back through +dnl the __gmpn_cpuvec pointer, at the given "offset" (in bytes). +dnl __gmpn_cpuvec_init will have stored the address of the selected +dnl implementation there. +dnl +dnl Only one of these routines will be executed, and only once, since after +dnl that all the __gmpn_cpuvec pointers go to real routines. So there's no +dnl need for anything special here, just something small and simple. To +dnl keep code size down, "fat_init" is a shared bit of code, arrived at +dnl with the offset in %al. %al is used since the movb instruction is 2 +dnl bytes where %eax would be 4. +dnl +dnl Note having `PROLOGUE in FAT_INIT obscures that PROLOGUE from the +dnl HAVE_NATIVE grepping in configure, preventing that code trying to eval +dnl something with $1 in it. + +define(FAT_INIT, +m4_assert_numargs(2) +`PROLOGUE($1)dnl + movb $`'$2, %al + jmp L(fat_init) +EPILOGUE() +') + +L(fat_init): + C al __gmpn_cpuvec byte offset + + movzbl %al, %eax + pushl %eax + +ifdef(`PIC',`dnl +ifdef(`DARWIN',` + sub $8, %esp + CALL( __gmpn_cpuvec_init) + add $8, %esp + call L(movl_eip_edx) + movl L(___gmpn_cpuvec)$non_lazy_ptr-.(%edx), %edx +',`dnl + pushl %ebx + call L(movl_eip_ebx) +L(init_here): + addl $_GLOBAL_OFFSET_TABLE_+[.-L(init_here)], %ebx + CALL( __gmpn_cpuvec_init) + movl GSYM_PREFIX`'__gmpn_cpuvec@GOT(%ebx), %edx + popl %ebx +') + popl %eax + jmp *(%edx,%eax) + +L(movl_eip_ebx): + movl (%esp), %ebx + ret_internal +',`dnl non-PIC + sub $8, %esp C needed on Darwin, harmless elsewhere + CALL( __gmpn_cpuvec_init) + add $8, %esp C needed on Darwin, harmless elsewhere + popl %eax + jmp *GSYM_PREFIX`'__gmpn_cpuvec(%eax) +') + +dnl FAT_INIT for each CPUVEC_FUNCS_LIST +dnl + +define(`CPUVEC_offset',0) +foreach(i, +`FAT_INIT(MPN(i`'_init),CPUVEC_offset) +define(`CPUVEC_offset',eval(CPUVEC_offset + 4))', +CPUVEC_FUNCS_LIST) + + + +C long __gmpn_cpuid (char dst[12], int id); +C +C This is called only once, so just something simple and compact is fine. + +defframe(PARAM_ID, 8) +defframe(PARAM_DST, 4) +deflit(`FRAME',0) + +PROLOGUE(__gmpn_cpuid) + pushl %esi FRAME_pushl() + pushl %ebx FRAME_pushl() + movl PARAM_ID, %eax + cpuid + movl PARAM_DST, %esi + movl %ebx, (%esi) + movl %edx, 4(%esi) + movl %ecx, 8(%esi) + popl %ebx + popl %esi + ret +EPILOGUE() + + +C int __gmpn_cpuid_available (void); +C +C Return non-zero if the cpuid instruction is available, which means late +C model 80486 and higher. 80386 and early 80486 don't have cpuid. +C +C The test follows Intel AP-485 application note, namely that if bit 21 is +C modifiable then cpuid is supported. This test is reentrant and thread +C safe, since of course any interrupt or context switch will preserve the +C flags while we're tinkering with them. +C +C This is called only once, so just something simple and compact is fine. + +PROLOGUE(__gmpn_cpuid_available) + pushf + popl %ecx C old flags + + movl %ecx, %edx + xorl $0x200000, %edx + pushl %edx + popf + pushf + popl %edx C tweaked flags + + movl $1, %eax + cmpl %ecx, %edx + jne L(available) + xorl %eax, %eax C not changed, so cpuid not available + +L(available): + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/fat/gmp-mparam.h b/gmp-6.3.0/mpn/x86/fat/gmp-mparam.h new file mode 100644 index 0000000..3641a6b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/gmp-mparam.h @@ -0,0 +1,71 @@ +/* Fat binary x86 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991, 1993, 1994, 2000-2003, 2011 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + + +/* mpn_divexact_1 is faster than mpn_divrem_1 at all sizes. The only time + this might not be true currently is for actual 80386 and 80486 chips, + where mpn/x86/dive_1.asm might be slower than mpn/x86/divrem_1.asm, but + that's not worth worrying about. */ +#define DIVEXACT_1_THRESHOLD 0 + +/* Only some of the x86s have an mpn_preinv_divrem_1, but we set + USE_PREINV_DIVREM_1 so that all callers use it, and then let the + __gmpn_cpuvec pointer go to plain mpn_divrem_1 if there's not an actual + preinv. */ +#define USE_PREINV_DIVREM_1 1 + +#define BMOD_1_TO_MOD_1_THRESHOLD 20 + +/* mpn_sqr_basecase is faster than mpn_mul_basecase at all sizes, no need + for mpn_sqr to call the latter. */ +#define SQR_BASECASE_THRESHOLD 0 + +/* Sensible fallbacks for these, when not taken from a cpu-specific + gmp-mparam.h. */ +#define MUL_TOOM22_THRESHOLD 20 +#define MUL_TOOM33_THRESHOLD 130 +#define SQR_TOOM2_THRESHOLD 30 +#define SQR_TOOM3_THRESHOLD 200 + +/* These are values more or less in the middle of what the typical x86 chips + come out as. For a fat binary it's necessary to have values for these, + since the defaults for MUL_FFT_TABLE and SQR_FFT_TABLE otherwise come out + as non-constant array initializers. FIXME: Perhaps these should be done + in the cpuvec structure like other thresholds. */ +#define MUL_FFT_TABLE { 464, 928, 1920, 3584, 10240, 40960, 0 } +#define MUL_FFT_MODF_THRESHOLD 400 +#define MUL_FFT_THRESHOLD 2000 + +#define SQR_FFT_TABLE { 528, 1184, 1920, 4608, 14336, 40960, 0 } +#define SQR_FFT_MODF_THRESHOLD 500 +#define SQR_FFT_THRESHOLD 3000 diff --git a/gmp-6.3.0/mpn/x86/fat/lshiftc.c b/gmp-6.3.0/mpn/x86/fat/lshiftc.c new file mode 100644 index 0000000..9ecf489 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/lshiftc.c @@ -0,0 +1,32 @@ +/* Fat binary fallback mpn_lshiftc. + +Copyright 2003, 2009, 2011 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#include "mpn/generic/lshiftc.c" diff --git a/gmp-6.3.0/mpn/x86/fat/mod_1.c b/gmp-6.3.0/mpn/x86/fat/mod_1.c new file mode 100644 index 0000000..4f149cc --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/mod_1.c @@ -0,0 +1,32 @@ +/* Fat binary fallback mpn_mod_1. + +Copyright 2003, 2009 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#include "mpn/generic/mod_1.c" diff --git a/gmp-6.3.0/mpn/x86/fat/mod_1_1.c b/gmp-6.3.0/mpn/x86/fat/mod_1_1.c new file mode 100644 index 0000000..92eaa7a --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/mod_1_1.c @@ -0,0 +1,36 @@ +/* Fat binary fallback mpn_mod_1_1p. + +Copyright 2003, 2009, 2011 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +/* +PROLOGUE(mpn_mod_1_1p_cps) +*/ + +#define OPERATION_mod_1_1_cps 1 +#include "mpn/generic/mod_1_1.c" diff --git a/gmp-6.3.0/mpn/x86/fat/mod_1_2.c b/gmp-6.3.0/mpn/x86/fat/mod_1_2.c new file mode 100644 index 0000000..9095a61 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/mod_1_2.c @@ -0,0 +1,36 @@ +/* Fat binary fallback mpn_mod_1s_2p. + +Copyright 2003, 2009, 2011 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +/* +PROLOGUE(mpn_mod_1s_2p_cps) +*/ + +#define OPERATION_mod_1_2_cps 1 +#include "mpn/generic/mod_1_2.c" diff --git a/gmp-6.3.0/mpn/x86/fat/mod_1_4.c b/gmp-6.3.0/mpn/x86/fat/mod_1_4.c new file mode 100644 index 0000000..51c0def --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/mod_1_4.c @@ -0,0 +1,36 @@ +/* Fat binary fallback mpn_mod_1s_4p. + +Copyright 2003, 2009, 2011 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +/* +PROLOGUE(mpn_mod_1s_4p_cps) +*/ + +#define OPERATION_mod_1_4_cps 1 +#include "mpn/generic/mod_1_4.c" diff --git a/gmp-6.3.0/mpn/x86/fat/mode1o.c b/gmp-6.3.0/mpn/x86/fat/mode1o.c new file mode 100644 index 0000000..870ddb8 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/mode1o.c @@ -0,0 +1,32 @@ +/* Fat binary fallback mpn_modexact_1c_odd. + +Copyright 2003 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#include "mpn/generic/mode1o.c" diff --git a/gmp-6.3.0/mpn/x86/fat/mullo_basecase.c b/gmp-6.3.0/mpn/x86/fat/mullo_basecase.c new file mode 100644 index 0000000..7f86be6 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/mullo_basecase.c @@ -0,0 +1,32 @@ +/* Fat binary fallback mpn_mullo_basecase. + +Copyright 2012 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#include "mpn/generic/mullo_basecase.c" diff --git a/gmp-6.3.0/mpn/x86/fat/redc_1.c b/gmp-6.3.0/mpn/x86/fat/redc_1.c new file mode 100644 index 0000000..0025403 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/redc_1.c @@ -0,0 +1,32 @@ +/* Fat binary fallback mpn_redc_1. + +Copyright 2012 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#include "mpn/generic/redc_1.c" diff --git a/gmp-6.3.0/mpn/x86/fat/redc_2.c b/gmp-6.3.0/mpn/x86/fat/redc_2.c new file mode 100644 index 0000000..1932d58 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/fat/redc_2.c @@ -0,0 +1,32 @@ +/* Fat binary fallback mpn_redc_2. + +Copyright 2012 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#include "mpn/generic/redc_2.c" diff --git a/gmp-6.3.0/mpn/x86/gcd_11.asm b/gmp-6.3.0/mpn/x86/gcd_11.asm new file mode 100644 index 0000000..af69135 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/gcd_11.asm @@ -0,0 +1,126 @@ +dnl x86 mpn_gcd_11 optimised for processors with slow BSF. + +dnl Based on C version. + +dnl Copyright 2019 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +dnl Rudimentary code for x86-32, i.e. for CPUs without cmov. Also, the bsf +dnl instruction is assumed to be so slow it is useless. Instead a teble is +dnl used. +dnl +dnl The loop benefits from OoO, in-order CPUs might want a different loop. +dnl The ebx and ecx registers could be combined if the assigment of ecx were +dnl postponed until ebx died, but that would at least hurt in-order CPUs. + +C cycles/bit (approx) +C AMD K7 ? +C AMD K8,K9 ? +C AMD K10 ? +C AMD bd1 ? +C AMD bd2 ? +C AMD bd3 ? +C AMD bd4 ? +C AMD bt1 ? +C AMD bt2 ? +C AMD zn1 ? +C AMD zn2 ? +C Intel P4-2 ? +C Intel P4-3/4 ? +C Intel P6/13 ? +C Intel CNR ? +C Intel NHM ? +C Intel SBR ? +C Intel IBR ? +C Intel HWL ? +C Intel BWL ? +C Intel SKL ? +C Intel atom ? +C Intel SLM ? +C Intel GLM ? +C Intel GLM+ ? +C VIA nano ? +C Numbers measured with: speed -CD -s8-32 -t24 mpn_gcd_1 + +deflit(MAXSHIFT, 6) +deflit(MASK, eval((m4_lshift(1,MAXSHIFT))-1)) + +DEF_OBJECT(ctz_table,64) + .byte MAXSHIFT +forloop(i,1,MASK, +` .byte m4_count_trailing_zeros(i) +') +END_OBJECT(ctz_table) + +define(`u0', `%eax') +define(`v0', `%edx') + + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(mpn_gcd_11) + push %edi + push %esi + push %ebx + + mov 16(%esp), u0 + mov 20(%esp), v0 + LEAL( ctz_table, %esi) + sub v0, u0 C u = u - v 0 + jz L(end) + + ALIGN(16) +L(top): sbb %ebx, %ebx C mask 1 + mov u0, %edi C 1 + mov u0, %ecx C 1 + and %ebx, %edi C 2 + xor %ebx, u0 C 2 + add %edi, v0 C v = min(u.v) 3 + sub %ebx, u0 C u = |u - v| 3 +L(mid): and $MASK, %ecx C 2 + movzbl (%esi,%ecx), %ecx C 3 + jz L(shift_alot) + shr %cl, u0 C 4 + sub v0, u0 C u = u - v 0,5 + jnz L(top) + +L(end): mov v0, %eax + pop %ebx + pop %esi + pop %edi + ret + +L(shift_alot): + shr $MAXSHIFT, u0 + mov u0, %ecx + jmp L(mid) +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/geode/gmp-mparam.h b/gmp-6.3.0/mpn/x86/geode/gmp-mparam.h new file mode 100644 index 0000000..cc9c9f1 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/geode/gmp-mparam.h @@ -0,0 +1,141 @@ +/* Generic x86 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991, 1993, 1994, 2000-2002, 2011 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* Generated by tuneup.c, 2011-01-30, gcc 3.4 */ + +#define MOD_1_NORM_THRESHOLD 6 +#define MOD_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define MOD_1N_TO_MOD_1_1_THRESHOLD 17 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 9 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 0 /* never mpn_mod_1_1p */ +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 14 +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD MP_SIZE_T_MAX /* never */ +#define USE_PREINV_DIVREM_1 0 +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 42 + +#define MUL_TOOM22_THRESHOLD 18 +#define MUL_TOOM33_THRESHOLD 66 +#define MUL_TOOM44_THRESHOLD 105 +#define MUL_TOOM6H_THRESHOLD 141 +#define MUL_TOOM8H_THRESHOLD 212 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 62 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 69 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 65 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 67 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 33 +#define SQR_TOOM3_THRESHOLD 60 +#define SQR_TOOM4_THRESHOLD 136 +#define SQR_TOOM6_THRESHOLD 196 +#define SQR_TOOM8_THRESHOLD 292 + +#define MULMOD_BNM1_THRESHOLD 14 +#define SQRMOD_BNM1_THRESHOLD 16 + +#define MUL_FFT_MODF_THRESHOLD 468 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 468, 5}, { 17, 6}, { 9, 5}, { 19, 6}, \ + { 11, 5}, { 23, 6}, { 21, 7}, { 11, 6}, \ + { 25, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \ + { 31, 7}, { 21, 8}, { 11, 7}, { 27, 8}, \ + { 15, 7}, { 33, 8}, { 19, 7}, { 39, 8}, \ + { 23, 7}, { 47, 8}, { 27, 9}, { 15, 8}, \ + { 39, 9}, { 23, 8}, { 47,10}, { 15, 9}, \ + { 31, 8}, { 67, 9}, { 39, 8}, { 79, 9}, \ + { 47, 8}, { 95, 9}, { 55,10}, { 31, 9}, \ + { 63, 8}, { 127, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 135,10}, \ + { 79, 9}, { 159,10}, { 95, 9}, { 191,11}, \ + { 63,10}, { 127, 9}, { 255,10}, { 143, 9}, \ + { 287,10}, { 159,11}, { 95,10}, { 191, 9}, \ + { 383,12}, { 4096,13}, { 8192,14}, { 16384,15}, \ + { 32768,16} } +#define MUL_FFT_TABLE3_SIZE 61 +#define MUL_FFT_THRESHOLD 5504 + +#define SQR_FFT_MODF_THRESHOLD 396 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 396, 5}, { 21, 6}, { 11, 5}, { 23, 6}, \ + { 21, 7}, { 11, 6}, { 24, 7}, { 13, 6}, \ + { 27, 7}, { 15, 6}, { 31, 7}, { 21, 8}, \ + { 11, 7}, { 27, 8}, { 15, 7}, { 33, 8}, \ + { 19, 7}, { 39, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 39, 9}, { 23, 8}, \ + { 51,10}, { 15, 9}, { 31, 8}, { 67, 9}, \ + { 39, 8}, { 79, 9}, { 47, 8}, { 95, 9}, \ + { 55,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 127, 8}, \ + { 255, 9}, { 135,10}, { 79, 9}, { 159, 8}, \ + { 319,10}, { 95, 9}, { 191,11}, { 63,10}, \ + { 127, 9}, { 255, 8}, { 511,10}, { 143, 9}, \ + { 287, 8}, { 575,10}, { 159,11}, { 95,10}, \ + { 191,12}, { 4096,13}, { 8192,14}, { 16384,15}, \ + { 32768,16} } +#define SQR_FFT_TABLE3_SIZE 61 +#define SQR_FFT_THRESHOLD 3712 + +#define MULLO_BASECASE_THRESHOLD 3 +#define MULLO_DC_THRESHOLD 37 +#define MULLO_MUL_N_THRESHOLD 10950 + +#define DC_DIV_QR_THRESHOLD 59 +#define DC_DIVAPPR_Q_THRESHOLD 189 +#define DC_BDIV_QR_THRESHOLD 55 +#define DC_BDIV_Q_THRESHOLD 136 + +#define INV_MULMOD_BNM1_THRESHOLD 50 +#define INV_NEWTON_THRESHOLD 183 +#define INV_APPR_THRESHOLD 181 + +#define BINV_NEWTON_THRESHOLD 204 +#define REDC_1_TO_REDC_N_THRESHOLD 54 + +#define MU_DIV_QR_THRESHOLD 1142 +#define MU_DIVAPPR_Q_THRESHOLD 1142 +#define MUPI_DIV_QR_THRESHOLD 81 +#define MU_BDIV_QR_THRESHOLD 889 +#define MU_BDIV_Q_THRESHOLD 998 + +#define MATRIX22_STRASSEN_THRESHOLD 13 +#define HGCD_THRESHOLD 133 +#define GCD_DC_THRESHOLD 451 +#define GCDEXT_DC_THRESHOLD 318 +#define JACOBI_BASE_METHOD 1 + +#define GET_STR_DC_THRESHOLD 15 +#define GET_STR_PRECOMPUTE_THRESHOLD 30 +#define SET_STR_DC_THRESHOLD 547 +#define SET_STR_PRECOMPUTE_THRESHOLD 1049 diff --git a/gmp-6.3.0/mpn/x86/gmp-mparam.h b/gmp-6.3.0/mpn/x86/gmp-mparam.h new file mode 100644 index 0000000..2cb1984 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/gmp-mparam.h @@ -0,0 +1,38 @@ +/* Generic x86 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991, 1993, 1994, 2000-2002 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + + +/* Generic x86 mpn_divexact_1 is faster than generic x86 mpn_divrem_1 on all + of p5, p6, k6 and k7, so use it always. It's probably slower on 386 and + 486, but that's too bad. */ +#define DIVEXACT_1_THRESHOLD 0 diff --git a/gmp-6.3.0/mpn/x86/goldmont/gmp-mparam.h b/gmp-6.3.0/mpn/x86/goldmont/gmp-mparam.h new file mode 100644 index 0000000..3d37fa3 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/goldmont/gmp-mparam.h @@ -0,0 +1,219 @@ +/* Intel Goldmont/32 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 2200 MHz Intel Atom C3758 Goldmont/Denverton */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-22, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 7 +#define MOD_1_UNNORM_THRESHOLD 12 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 9 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 7 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 10 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 12 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 32.79% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 32 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 23 + +#define DIV_1_VS_MUL_1_PERCENT 228 + +#define MUL_TOOM22_THRESHOLD 18 +#define MUL_TOOM33_THRESHOLD 81 +#define MUL_TOOM44_THRESHOLD 193 +#define MUL_TOOM6H_THRESHOLD 286 +#define MUL_TOOM8H_THRESHOLD 399 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 81 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 138 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 125 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 137 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 185 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 32 +#define SQR_TOOM3_THRESHOLD 113 +#define SQR_TOOM4_THRESHOLD 280 +#define SQR_TOOM6_THRESHOLD 399 +#define SQR_TOOM8_THRESHOLD 547 + +#define MULMID_TOOM42_THRESHOLD 60 + +#define MULMOD_BNM1_THRESHOLD 13 +#define SQRMOD_BNM1_THRESHOLD 15 + +#define MUL_FFT_MODF_THRESHOLD 368 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 368, 5}, { 21, 6}, { 11, 5}, { 23, 6}, \ + { 21, 7}, { 11, 6}, { 25, 7}, { 13, 6}, \ + { 27, 7}, { 15, 6}, { 31, 7}, { 21, 8}, \ + { 11, 7}, { 27, 8}, { 15, 7}, { 33, 8}, \ + { 19, 7}, { 39, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 39, 9}, { 23, 8}, \ + { 47,10}, { 15, 9}, { 31, 8}, { 63, 9}, \ + { 39, 8}, { 79, 9}, { 47,10}, { 31, 9}, \ + { 79,10}, { 47, 9}, { 95,11}, { 31,10}, \ + { 63, 9}, { 127, 8}, { 255, 9}, { 135,10}, \ + { 79, 9}, { 159,10}, { 95, 9}, { 191,11}, \ + { 63,10}, { 127, 9}, { 255, 8}, { 511,10}, \ + { 143, 9}, { 287, 8}, { 575, 9}, { 303,10}, \ + { 159,11}, { 95,10}, { 191,12}, { 63,11}, \ + { 127,10}, { 255, 9}, { 511,10}, { 271, 9}, \ + { 543,10}, { 287, 9}, { 575,10}, { 303, 9}, \ + { 607,11}, { 159,10}, { 319, 9}, { 639,10}, \ + { 351, 9}, { 703,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 415, 9}, { 831,11}, { 223,10}, \ + { 447,12}, { 127,11}, { 255,10}, { 543, 9}, \ + { 1087,11}, { 287,10}, { 607, 9}, { 1215,11}, \ + { 319,10}, { 671,11}, { 351,10}, { 703,12}, \ + { 191,11}, { 383,10}, { 767,11}, { 415,10}, \ + { 831,11}, { 447,13}, { 127,12}, { 255,11}, \ + { 543,10}, { 1087,11}, { 607,10}, { 1215,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 703,10}, \ + { 1407,11}, { 735,12}, { 383,11}, { 831,12}, \ + { 447,11}, { 959,13}, { 255,12}, { 511,11}, \ + { 1087,12}, { 575,11}, { 1215,10}, { 2431,12}, \ + { 639,11}, { 1343,12}, { 703,11}, { 1407,13}, \ + { 383,12}, { 831,11}, { 1663,12}, { 959,11}, \ + { 1919,14}, { 255,13}, { 511,12}, { 1215,11}, \ + { 2431,13}, { 639,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1727,13}, { 895,12}, { 1919,11}, \ + { 3839,14}, { 511,13}, { 1023,12}, { 2111,13}, \ + { 1151,12}, { 2431,13}, { 1407,12}, { 2943,14}, \ + { 767,13}, { 1663,12}, { 3455,13}, { 1919,12}, \ + { 3839,15}, { 511,14}, { 1023,13}, { 2431,14}, \ + { 1279,13}, { 2943,12}, { 5887,14}, { 1535,13}, \ + { 3455,14}, { 1791,13}, { 3839,12}, { 7679,15}, \ + { 1023,14}, { 2303,13}, { 4991,12}, { 9983,14}, \ + { 2559,13}, { 5119,14}, { 2815,13}, { 5887,15}, \ + { 1535,14}, { 3839,13}, { 7679,16} } +#define MUL_FFT_TABLE3_SIZE 171 +#define MUL_FFT_THRESHOLD 3712 + +#define SQR_FFT_MODF_THRESHOLD 340 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 340, 5}, { 21, 6}, { 11, 5}, { 23, 6}, \ + { 12, 5}, { 25, 6}, { 21, 7}, { 11, 6}, \ + { 25, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \ + { 31, 7}, { 21, 8}, { 11, 7}, { 27, 8}, \ + { 15, 7}, { 33, 8}, { 19, 7}, { 39, 8}, \ + { 23, 7}, { 47, 8}, { 27, 9}, { 15, 8}, \ + { 39, 9}, { 23, 8}, { 47,10}, { 15, 9}, \ + { 31, 8}, { 67, 9}, { 39, 8}, { 79, 9}, \ + { 47,10}, { 31, 9}, { 79,10}, { 47,11}, \ + { 31,10}, { 63, 9}, { 127, 8}, { 255,10}, \ + { 79, 9}, { 159, 8}, { 319,10}, { 95, 9}, \ + { 191,11}, { 63,10}, { 127, 9}, { 255, 8}, \ + { 511, 9}, { 271,10}, { 143, 9}, { 287, 8}, \ + { 575, 9}, { 303, 8}, { 607, 9}, { 319,11}, \ + { 95,10}, { 191,12}, { 63,11}, { 127,10}, \ + { 255, 9}, { 511,10}, { 271, 9}, { 543,10}, \ + { 287, 9}, { 575,10}, { 303, 9}, { 607,11}, \ + { 159,10}, { 319, 9}, { 639,10}, { 335, 9}, \ + { 671,10}, { 351, 9}, { 703,11}, { 191,10}, \ + { 383, 9}, { 767,10}, { 415, 9}, { 831,11}, \ + { 223,10}, { 479,12}, { 127,11}, { 255,10}, \ + { 543, 9}, { 1087,11}, { 287,10}, { 607, 9}, \ + { 1215,11}, { 319,10}, { 671,11}, { 351,10}, \ + { 703,12}, { 191,11}, { 383,10}, { 767,11}, \ + { 415,10}, { 831,11}, { 479,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,10}, \ + { 1215,12}, { 319,11}, { 671,10}, { 1343,11}, \ + { 735,12}, { 383,11}, { 831,12}, { 447,11}, \ + { 959,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1471,13}, { 383,12}, { 831,11}, \ + { 1663,12}, { 959,11}, { 1919,14}, { 255,13}, \ + { 511,12}, { 1215,13}, { 639,12}, { 1471,11}, \ + { 2943,13}, { 767,12}, { 1727,13}, { 895,12}, \ + { 1919,14}, { 511,13}, { 1023,12}, { 2111,13}, \ + { 1151,12}, { 2431,13}, { 1407,12}, { 2943,14}, \ + { 767,13}, { 1663,12}, { 3455,13}, { 1919,15}, \ + { 511,14}, { 1023,13}, { 2431,14}, { 1279,13}, \ + { 2943,12}, { 5887,14}, { 1535,13}, { 3455,14}, \ + { 1791,13}, { 3839,12}, { 7679,15}, { 1023,14}, \ + { 2047,13}, { 4095,14}, { 2303,13}, { 4991,12}, \ + { 9983,14}, { 2815,13}, { 5887,15}, { 1535,14}, \ + { 3839,13}, { 7679,16} } +#define SQR_FFT_TABLE3_SIZE 170 +#define SQR_FFT_THRESHOLD 3520 + +#define MULLO_BASECASE_THRESHOLD 5 +#define MULLO_DC_THRESHOLD 50 +#define MULLO_MUL_N_THRESHOLD 6633 +#define SQRLO_BASECASE_THRESHOLD 0 /* always */ +#define SQRLO_DC_THRESHOLD 95 +#define SQRLO_SQR_THRESHOLD 6633 + +#define DC_DIV_QR_THRESHOLD 68 +#define DC_DIVAPPR_Q_THRESHOLD 204 +#define DC_BDIV_QR_THRESHOLD 64 +#define DC_BDIV_Q_THRESHOLD 108 + +#define INV_MULMOD_BNM1_THRESHOLD 34 +#define INV_NEWTON_THRESHOLD 276 +#define INV_APPR_THRESHOLD 226 + +#define BINV_NEWTON_THRESHOLD 298 +#define REDC_1_TO_REDC_N_THRESHOLD 65 + +#define MU_DIV_QR_THRESHOLD 1528 +#define MU_DIVAPPR_Q_THRESHOLD 1589 +#define MUPI_DIV_QR_THRESHOLD 140 +#define MU_BDIV_QR_THRESHOLD 1334 +#define MU_BDIV_Q_THRESHOLD 1499 + +#define POWM_SEC_TABLE 3,16,96,428,1317 + +#define GET_STR_DC_THRESHOLD 13 +#define GET_STR_PRECOMPUTE_THRESHOLD 18 +#define SET_STR_DC_THRESHOLD 704 +#define SET_STR_PRECOMPUTE_THRESHOLD 1358 + +#define FAC_DSC_THRESHOLD 95 +#define FAC_ODD_THRESHOLD 29 + +#define MATRIX22_STRASSEN_THRESHOLD 15 +#define HGCD2_DIV1_METHOD 1 /* 5.53% faster than 3 */ +#define HGCD_THRESHOLD 172 +#define HGCD_APPR_THRESHOLD 204 +#define HGCD_REDUCE_THRESHOLD 2479 +#define GCD_DC_THRESHOLD 610 +#define GCDEXT_DC_THRESHOLD 443 +#define JACOBI_BASE_METHOD 4 /* 6.53% faster than 3 */ + +/* Tuneup completed successfully, took 101563 seconds */ diff --git a/gmp-6.3.0/mpn/x86/i486/gmp-mparam.h b/gmp-6.3.0/mpn/x86/i486/gmp-mparam.h new file mode 100644 index 0000000..aa7dbad --- /dev/null +++ b/gmp-6.3.0/mpn/x86/i486/gmp-mparam.h @@ -0,0 +1,69 @@ +/* 80486 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2001-2003 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + + +/* 100MHz DX4 */ + +/* Generated by tuneup.c, 2003-02-13, gcc 2.95 */ + +#define MUL_TOOM22_THRESHOLD 18 +#define MUL_TOOM33_THRESHOLD 228 + +#define SQR_BASECASE_THRESHOLD 13 +#define SQR_TOOM2_THRESHOLD 49 +#define SQR_TOOM3_THRESHOLD 238 + +#define DIV_SB_PREINV_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_DC_THRESHOLD 72 +#define POWM_THRESHOLD 38 + +#define GCD_ACCEL_THRESHOLD 3 +#define JACOBI_BASE_METHOD 2 + +#define USE_PREINV_DIVREM_1 0 +#define USE_PREINV_MOD_1 0 +#define DIVREM_2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define MODEXACT_1_ODD_THRESHOLD 17 + +#define GET_STR_DC_THRESHOLD 32 +#define GET_STR_PRECOMPUTE_THRESHOLD 82 +#define SET_STR_THRESHOLD 3524 + +#define MUL_FFT_TABLE { 464, 928, 1920, 4608, 10240, 40960, 0 } +#define MUL_FFT_MODF_THRESHOLD 392 +#define MUL_FFT_THRESHOLD 2816 + +#define SQR_FFT_TABLE { 432, 928, 1920, 4608, 14336, 40960, 0 } +#define SQR_FFT_MODF_THRESHOLD 392 +#define SQR_FFT_THRESHOLD 2816 diff --git a/gmp-6.3.0/mpn/x86/k10/gmp-mparam.h b/gmp-6.3.0/mpn/x86/k10/gmp-mparam.h new file mode 100644 index 0000000..eceaaae --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k10/gmp-mparam.h @@ -0,0 +1,217 @@ +/* x86/k10 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991, 1993, 1994, 2000-2011, 2014-2015 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 3200-3600 MHz K10 Thuban */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-19, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 0 /* always */ +#define MOD_1_UNNORM_THRESHOLD 0 /* always */ +#define MOD_1N_TO_MOD_1_1_THRESHOLD 14 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 7 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 0 /* never mpn_mod_1_1p */ +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 18 +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 22 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 29.33% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 2 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 35 + +#define DIV_1_VS_MUL_1_PERCENT 258 + +#define MUL_TOOM22_THRESHOLD 22 +#define MUL_TOOM33_THRESHOLD 73 +#define MUL_TOOM44_THRESHOLD 124 +#define MUL_TOOM6H_THRESHOLD 274 +#define MUL_TOOM8H_THRESHOLD 430 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 81 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 99 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 85 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 88 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 113 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 26 +#define SQR_TOOM3_THRESHOLD 105 +#define SQR_TOOM4_THRESHOLD 154 +#define SQR_TOOM6_THRESHOLD 238 +#define SQR_TOOM8_THRESHOLD 309 + +#define MULMID_TOOM42_THRESHOLD 50 + +#define MULMOD_BNM1_THRESHOLD 15 +#define SQRMOD_BNM1_THRESHOLD 18 + +#define MUL_FFT_MODF_THRESHOLD 570 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 570, 5}, { 21, 6}, { 11, 5}, { 25, 6}, \ + { 13, 5}, { 27, 6}, { 15, 5}, { 31, 6}, \ + { 25, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \ + { 32, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 55, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 79, 9}, { 47, 8}, \ + { 95, 9}, { 55,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95, 9}, \ + { 191,11}, { 63,10}, { 143, 9}, { 287,10}, \ + { 159,11}, { 95,10}, { 191,12}, { 63,11}, \ + { 127,10}, { 255, 9}, { 511,10}, { 271, 9}, \ + { 543,10}, { 287,11}, { 159,10}, { 319, 9}, \ + { 639,10}, { 335, 9}, { 671,11}, { 191,10}, \ + { 383, 9}, { 767,10}, { 399, 9}, { 799,11}, \ + { 223,12}, { 127,11}, { 255,10}, { 543,11}, \ + { 287,10}, { 607, 9}, { 1215,11}, { 319,10}, \ + { 671,12}, { 191,11}, { 383,10}, { 799,11}, \ + { 415,13}, { 127,12}, { 255,11}, { 543,10}, \ + { 1087,11}, { 607,10}, { 1215,12}, { 319,11}, \ + { 671,10}, { 1343,11}, { 735,10}, { 1471, 9}, \ + { 2943,12}, { 383,11}, { 799,10}, { 1599,11}, \ + { 863,12}, { 447,11}, { 959,13}, { 255,12}, \ + { 511,11}, { 1087,12}, { 575,11}, { 1215,10}, \ + { 2431,12}, { 639,11}, { 1343,12}, { 703,11}, \ + { 1471,10}, { 2943,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1727,10}, { 3455,12}, \ + { 959,11}, { 1919,14}, { 255,13}, { 511,12}, \ + { 1087,11}, { 2239,12}, { 1215,11}, { 2431,13}, \ + { 639,12}, { 1471,11}, { 2943,13}, { 767,12}, \ + { 1727,11}, { 3455,13}, { 895,12}, { 1983,14}, \ + { 511,13}, { 1023,12}, { 2239,13}, { 1151,12}, \ + { 2431,13}, { 1407,12}, { 2943,14}, { 767,13}, \ + { 1663,12}, { 3455,13}, { 1919,15}, { 511,14}, \ + { 1023,13}, { 2175,12}, { 4479,13}, { 2431,14}, \ + { 1279,13}, { 2943,12}, { 5887,14}, { 1535,13}, \ + { 3455,14}, { 1791,13}, { 3967,15}, { 1023,14}, \ + { 2047,13}, { 4479,14}, { 2303,13}, { 4991,14}, \ + { 2815,13}, { 5887,15}, { 1535,14}, { 3839,16} } +#define MUL_FFT_TABLE3_SIZE 168 +#define MUL_FFT_THRESHOLD 7424 + +#define SQR_FFT_MODF_THRESHOLD 525 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 525, 5}, { 25, 6}, { 13, 5}, { 28, 6}, \ + { 25, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \ + { 31, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 51, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 79, 9}, { 55,10}, \ + { 31, 9}, { 79,10}, { 47, 9}, { 95,11}, \ + { 31,10}, { 63, 9}, { 127,10}, { 79, 9}, \ + { 159,10}, { 95,11}, { 63,10}, { 143, 9}, \ + { 287,10}, { 159,11}, { 95,10}, { 191,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 511,10}, \ + { 271, 9}, { 543,10}, { 287,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335, 9}, { 671,10}, \ + { 351,11}, { 191,10}, { 383, 9}, { 767,10}, \ + { 399, 9}, { 799,10}, { 415,12}, { 127,11}, \ + { 255,10}, { 543,11}, { 287,10}, { 607,11}, \ + { 319,10}, { 671, 9}, { 1343,11}, { 351,10}, \ + { 703,12}, { 191,11}, { 383,10}, { 799, 9}, \ + { 1599,11}, { 415,10}, { 831,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 735,10}, \ + { 1471,12}, { 383,11}, { 799,10}, { 1599,11}, \ + { 863,10}, { 1727,12}, { 447,11}, { 959,10}, \ + { 1919,11}, { 991,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,10}, { 2431,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,10}, \ + { 3455,12}, { 959,11}, { 1919,13}, { 511,12}, \ + { 1087,11}, { 2239,12}, { 1215,11}, { 2431,13}, \ + { 639,12}, { 1471,11}, { 2943,13}, { 767,12}, \ + { 1727,11}, { 3455,13}, { 895,12}, { 1919,14}, \ + { 511,13}, { 1023,12}, { 2239,13}, { 1151,12}, \ + { 2495,13}, { 1279,12}, { 2623,13}, { 1407,12}, \ + { 2943,14}, { 767,13}, { 1663,12}, { 3455,13}, \ + { 1919,15}, { 511,14}, { 1023,13}, { 2175,12}, \ + { 4351,13}, { 2431,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3967,15}, { 1023,14}, { 2047,13}, { 4351,14}, \ + { 2303,13}, { 4991,14}, { 2815,13}, { 5887,15}, \ + { 1535,14}, { 3839,16} } +#define SQR_FFT_TABLE3_SIZE 166 +#define SQR_FFT_THRESHOLD 5312 + +#define MULLO_BASECASE_THRESHOLD 6 +#define MULLO_DC_THRESHOLD 40 +#define MULLO_MUL_N_THRESHOLD 14281 +#define SQRLO_BASECASE_THRESHOLD 8 +#define SQRLO_DC_THRESHOLD 113 +#define SQRLO_SQR_THRESHOLD 10323 + +#define DC_DIV_QR_THRESHOLD 56 +#define DC_DIVAPPR_Q_THRESHOLD 248 +#define DC_BDIV_QR_THRESHOLD 55 +#define DC_BDIV_Q_THRESHOLD 158 + +#define INV_MULMOD_BNM1_THRESHOLD 42 +#define INV_NEWTON_THRESHOLD 254 +#define INV_APPR_THRESHOLD 252 + +#define BINV_NEWTON_THRESHOLD 292 +#define REDC_1_TO_REDC_N_THRESHOLD 67 + +#define MU_DIV_QR_THRESHOLD 1589 +#define MU_DIVAPPR_Q_THRESHOLD 1558 +#define MUPI_DIV_QR_THRESHOLD 114 +#define MU_BDIV_QR_THRESHOLD 1442 +#define MU_BDIV_Q_THRESHOLD 1524 + +#define POWM_SEC_TABLE 1,16,102,416,1378 + +#define GET_STR_DC_THRESHOLD 13 +#define GET_STR_PRECOMPUTE_THRESHOLD 21 +#define SET_STR_DC_THRESHOLD 270 +#define SET_STR_PRECOMPUTE_THRESHOLD 1105 + +#define FAC_DSC_THRESHOLD 159 +#define FAC_ODD_THRESHOLD 34 + +#define MATRIX22_STRASSEN_THRESHOLD 17 +#define HGCD2_DIV1_METHOD 3 /* 0.70% faster than 4 */ +#define HGCD_THRESHOLD 130 +#define HGCD_APPR_THRESHOLD 163 +#define HGCD_REDUCE_THRESHOLD 3389 +#define GCD_DC_THRESHOLD 573 +#define GCDEXT_DC_THRESHOLD 393 +#define JACOBI_BASE_METHOD 4 /* 9.13% faster than 1 */ + +/* Tuneup completed successfully, took 52901 seconds */ diff --git a/gmp-6.3.0/mpn/x86/k6/README b/gmp-6.3.0/mpn/x86/k6/README new file mode 100644 index 0000000..1d65af3 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/README @@ -0,0 +1,251 @@ +Copyright 2000, 2001 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. + + + + + AMD K6 MPN SUBROUTINES + + + +This directory contains code optimized for AMD K6 CPUs, meaning K6, K6-2 and +K6-3. + +The mmx subdirectory has MMX code suiting plain K6, the k62mmx subdirectory +has MMX code suiting K6-2 and K6-3. All chips in the K6 family have MMX, +the separate directories are just so that ./configure can omit them if the +assembler doesn't support MMX. + + + + +STATUS + +Times for the loops, with all code and data in L1 cache, are as follows. + + cycles/limb + + mpn_add_n/sub_n 3.25 normal, 2.75 in-place + + mpn_mul_1 6.25 + mpn_add/submul_1 7.65-8.4 (varying with data values) + + mpn_mul_basecase 9.25 cycles/crossproduct (approx) + mpn_sqr_basecase 4.7 cycles/crossproduct (approx) + or 9.2 cycles/triangleproduct (approx) + + mpn_l/rshift 3.0 + + mpn_divrem_1 20.0 + mpn_mod_1 20.0 + mpn_divexact_by3 11.0 + + mpn_copyi 1.0 + mpn_copyd 1.0 + + +K6-2 and K6-3 have dual-issue MMX and get the following improvements. + + mpn_l/rshift 1.75 + + +Prefetching of sources hasn't yet given any joy. With the 3DNow "prefetch" +instruction, code seems to run slower, and with just "mov" loads it doesn't +seem faster. Results so far are inconsistent. The K6 does a hardware +prefetch of the second cache line in a sector, so the penalty for not +prefetching in software is reduced. + + + + +NOTES + +All K6 family chips have MMX, but only K6-2 and K6-3 have 3DNow. + +Plain K6 executes MMX instructions only in the X pipe, but K6-2 and K6-3 can +execute them in both X and Y (and in both together). + +Branch misprediction penalty is 1 to 4 cycles (Optimization Manual +chapter 6 table 12). + +Write-allocate L1 data cache means prefetching of destinations is unnecessary. +Store queue is 7 entries of 64 bits each. + +Floating point multiplications can be done in parallel with integer +multiplications, but there doesn't seem to be any way to make use of this. + + + +OPTIMIZATIONS + +Unrolled loops are used to reduce looping overhead. The unrolling is +configurable up to 32 limbs/loop for most routines, up to 64 for some. + +Sometimes computed jumps into the unrolling are used to handle sizes not a +multiple of the unrolling. An attractive feature of this is that times +smoothly increase with operand size, but an indirect jump is about 6 cycles +and the setups about another 6, so it depends on how much the unrolled code +is faster than a simple loop as to whether a computed jump ought to be used. + +Position independent code is implemented using a call to get eip for +computed jumps and a ret is always done, rather than an addl $4,%esp or a +popl, so the CPU return address branch prediction stack stays synchronised +with the actual stack in memory. Such a call however still costs 4 to 7 +cycles. + +Branch prediction, in absence of any history, will guess forward jumps are +not taken and backward jumps are taken. Where possible it's arranged that +the less likely or less important case is under a taken forward jump. + + + +MMX + +Putting emms or femms as late as possible in a routine seems to be fastest. +Perhaps an emms or femms stalls until all outstanding MMX instructions have +completed, so putting it later gives them a chance to complete on their own, +in parallel with other operations (like register popping). + +The Optimization Manual chapter 5 recommends using a femms on K6-2 and K6-3 +at the start of a routine, in case it's been preceded by x87 floating point +operations. This isn't done because in gmp programs it's expected that x87 +floating point won't be much used and that chances are an mpn routine won't +have been preceded by any x87 code. + + + +CODING + +Instructions in general code are shown paired if they can decode and execute +together, meaning two short decode instructions with the second not +depending on the first, only the first using the shifter, no more than one +load, and no more than one store. + +K6 does some out of order execution so the pairings aren't essential, they +just show what slots might be available. When decoding is the limiting +factor things can be scheduled that might not execute until later. + + + +NOTES + +Code alignment + +- if an opcode/modrm or 0Fh/opcode/modrm crosses a cache line boundary, + short decode is inhibited. The cross.pl script detects this. + +- loops and branch targets should be aligned to 16 bytes, or ensure at least + 2 instructions before a 32 byte boundary. This makes use of the 16 byte + cache in the BTB. + +Addressing modes + +- (%esi) degrades decoding from short to vector. 0(%esi) doesn't have this + problem, and can be used as an equivalent, or easier is just to use a + different register, like %ebx. + +- K6 and pre-CXT core K6-2 have the following problem. (K6-2 CXT and K6-3 + have it fixed, these being cpuid function 1 signatures 0x588 to 0x58F). + + If more than 3 bytes are needed to determine instruction length then + decoding degrades from direct to long, or from long to vector. This + happens with forms like "0F opcode mod/rm" with mod/rm=00-xxx-100 since + with mod=00 the sib determines whether there's a displacement. + + This affects all MMX and 3DNow instructions, and others with an 0F prefix, + like movzbl. The modes affected are anything with an index and no + displacement, or an index but no base, and this includes (%esp) which is + really (,%esp,1). + + The cross.pl script detects problem cases. The workaround is to always + use a displacement, and to do this with Zdisp if it's zero so the + assembler doesn't discard it. + + See Optimization Manual rev D page 67 and 3DNow Porting Guide rev B pages + 13-14 and 36-37. + +Calls + +- indirect jumps and calls are not branch predicted, they measure about 6 + cycles. + +Various + +- adcl 2 cycles of decode, maybe 2 cycles executing in the X pipe +- bsf 12-27 cycles +- emms 5 cycles +- femms 3 cycles +- jecxz 2 cycles taken, 13 not taken (optimization manual says 7 not taken) +- divl 20 cycles back-to-back +- imull 2 decode, 3 execute +- mull 2 decode, 3 execute (optimization manual decoding sample) +- prefetch 2 cycles +- rcll/rcrl implicit by one bit: 2 cycles + immediate or %cl count: 11 + 2 per bit for dword + 13 + 4 per bit for byte +- setCC 2 cycles +- xchgl %eax,reg 1.5 cycles, back-to-back (strange) + reg,reg 2 cycles, back-to-back + + + + +REFERENCES + +"AMD-K6 Processor Code Optimization Application Note", AMD publication +number 21924, revision D amendment 0, January 2000. This describes K6-2 and +K6-3. Available on-line, + +http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/21924.pdf + +"AMD-K6 MMX Enhanced Processor x86 Code Optimization Application Note", AMD +publication number 21828, revision A amendment 0, August 1997. This is an +older edition of the above document, describing plain K6. Available +on-line, + +http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/21828.pdf + +"3DNow Technology Manual", AMD publication number 21928G/0-March 2000. +This describes the femms and prefetch instructions, but nothing else from +3DNow has been used. Available on-line, + +http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/21928.pdf + +"3DNow Instruction Porting Guide", AMD publication number 22621, revision B, +August 1999. This has some notes on general K6 optimizations as well as +3DNow. Available on-line, + +http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22621.pdf + + + +---------------- +Local variables: +mode: text +fill-column: 76 +End: diff --git a/gmp-6.3.0/mpn/x86/k6/aors_n.asm b/gmp-6.3.0/mpn/x86/k6/aors_n.asm new file mode 100644 index 0000000..168f9b4 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/aors_n.asm @@ -0,0 +1,337 @@ +dnl AMD K6 mpn_add/sub_n -- mpn addition or subtraction. + +dnl Copyright 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K6: normal 3.25 cycles/limb, in-place 2.75 cycles/limb. + + +ifdef(`OPERATION_add_n', ` + define(M4_inst, adcl) + define(M4_function_n, mpn_add_n) + define(M4_function_nc, mpn_add_nc) + define(M4_description, add) +',`ifdef(`OPERATION_sub_n', ` + define(M4_inst, sbbl) + define(M4_function_n, mpn_sub_n) + define(M4_function_nc, mpn_sub_nc) + define(M4_description, subtract) +',`m4_error(`Need OPERATION_add_n or OPERATION_sub_n +')')') + +MULFUNC_PROLOGUE(mpn_add_n mpn_add_nc mpn_sub_n mpn_sub_nc) + + +C mp_limb_t M4_function_n (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size); +C mp_limb_t M4_function_nc (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size, mp_limb_t carry); +C +C Calculate src1,size M4_description src2,size, and store the result in +C dst,size. The return value is the carry bit from the top of the result +C (1 or 0). +C +C The _nc version accepts 1 or 0 for an initial carry into the low limb of +C the calculation. Note values other than 1 or 0 here will lead to garbage +C results. +C +C Instruction decoding limits a normal dst=src1+src2 operation to 3 c/l, and +C an in-place dst+=src to 2.5 c/l. The unrolled loops have 1 cycle/loop of +C loop control, which with 4 limbs/loop means an extra 0.25 c/l. + +define(PARAM_CARRY, `FRAME+20(%esp)') +define(PARAM_SIZE, `FRAME+16(%esp)') +define(PARAM_SRC2, `FRAME+12(%esp)') +define(PARAM_SRC1, `FRAME+8(%esp)') +define(PARAM_DST, `FRAME+4(%esp)') +deflit(`FRAME',0) + +dnl minimum 5 because the unrolled code can't handle less +deflit(UNROLL_THRESHOLD, 5) + + TEXT + ALIGN(32) + +PROLOGUE(M4_function_nc) + movl PARAM_CARRY, %eax + jmp L(start) +EPILOGUE() + + +PROLOGUE(M4_function_n) + xorl %eax, %eax +L(start): + movl PARAM_SIZE, %ecx + pushl %ebx +FRAME_pushl() + + movl PARAM_SRC1, %ebx + pushl %edi +FRAME_pushl() + + movl PARAM_SRC2, %edx + cmpl $UNROLL_THRESHOLD, %ecx + + movl PARAM_DST, %edi + jae L(unroll) + + + shrl %eax C initial carry flag + + C offset 0x21 here, close enough to aligned +L(simple): + C eax scratch + C ebx src1 + C ecx counter + C edx src2 + C esi + C edi dst + C ebp + C + C The store to (%edi) could be done with a stosl; it'd be smaller + C code, but there's no speed gain and a cld would have to be added + C (per mpn/x86/README). + + movl (%ebx), %eax + leal 4(%ebx), %ebx + + M4_inst (%edx), %eax + + movl %eax, (%edi) + leal 4(%edi), %edi + + leal 4(%edx), %edx + loop L(simple) + + + movl $0, %eax + popl %edi + + setc %al + + popl %ebx + ret + + +C ----------------------------------------------------------------------------- +L(unroll): + C eax carry + C ebx src1 + C ecx counter + C edx src2 + C esi + C edi dst + C ebp + + cmpl %edi, %ebx + pushl %esi + + je L(inplace) + +ifdef(`OPERATION_add_n',` + cmpl %edi, %edx + + je L(inplace_reverse) +') + + movl %ecx, %esi + + andl $-4, %ecx + andl $3, %esi + + leal (%ebx,%ecx,4), %ebx + leal (%edx,%ecx,4), %edx + leal (%edi,%ecx,4), %edi + + negl %ecx + shrl %eax + + ALIGN(32) +L(normal_top): + C eax counter, qwords, negative + C ebx src1 + C ecx scratch + C edx src2 + C esi + C edi dst + C ebp + + movl (%ebx,%ecx,4), %eax + leal 5(%ecx), %ecx + M4_inst -20(%edx,%ecx,4), %eax + movl %eax, -20(%edi,%ecx,4) + + movl 4-20(%ebx,%ecx,4), %eax + M4_inst 4-20(%edx,%ecx,4), %eax + movl %eax, 4-20(%edi,%ecx,4) + + movl 8-20(%ebx,%ecx,4), %eax + M4_inst 8-20(%edx,%ecx,4), %eax + movl %eax, 8-20(%edi,%ecx,4) + + movl 12-20(%ebx,%ecx,4), %eax + M4_inst 12-20(%edx,%ecx,4), %eax + movl %eax, 12-20(%edi,%ecx,4) + + loop L(normal_top) + + + decl %esi + jz L(normal_finish_one) + js L(normal_done) + + C two or three more limbs + + movl (%ebx), %eax + M4_inst (%edx), %eax + movl %eax, (%edi) + + movl 4(%ebx), %eax + M4_inst 4(%edx), %eax + decl %esi + movl %eax, 4(%edi) + + jz L(normal_done) + movl $2, %ecx + +L(normal_finish_one): + movl (%ebx,%ecx,4), %eax + M4_inst (%edx,%ecx,4), %eax + movl %eax, (%edi,%ecx,4) + +L(normal_done): + popl %esi + popl %edi + + movl $0, %eax + popl %ebx + + setc %al + + ret + + +C ----------------------------------------------------------------------------- + +ifdef(`OPERATION_add_n',` +L(inplace_reverse): + C dst==src2 + + movl %ebx, %edx +') + +L(inplace): + C eax initial carry + C ebx + C ecx size + C edx src + C esi + C edi dst + C ebp + + leal -1(%ecx), %esi + decl %ecx + + andl $-4, %ecx + andl $3, %esi + + movl (%edx), %ebx C src low limb + leal (%edx,%ecx,4), %edx + + leal (%edi,%ecx,4), %edi + negl %ecx + + shrl %eax + + + ALIGN(32) +L(inplace_top): + C eax + C ebx next src limb + C ecx size + C edx src + C esi + C edi dst + C ebp + + M4_inst %ebx, (%edi,%ecx,4) + + movl 4(%edx,%ecx,4), %eax + leal 5(%ecx), %ecx + + M4_inst %eax, 4-20(%edi,%ecx,4) + + movl 8-20(%edx,%ecx,4), %eax + movl 12-20(%edx,%ecx,4), %ebx + + M4_inst %eax, 8-20(%edi,%ecx,4) + M4_inst %ebx, 12-20(%edi,%ecx,4) + + movl 16-20(%edx,%ecx,4), %ebx + loop L(inplace_top) + + + C now %esi is 0 to 3 representing respectively 1 to 4 limbs more + + M4_inst %ebx, (%edi) + + decl %esi + jz L(inplace_finish_one) + js L(inplace_done) + + C two or three more limbs + + movl 4(%edx), %eax + movl 8(%edx), %ebx + M4_inst %eax, 4(%edi) + M4_inst %ebx, 8(%edi) + + decl %esi + movl $2, %ecx + + jz L(normal_done) + +L(inplace_finish_one): + movl 4(%edx,%ecx,4), %eax + M4_inst %eax, 4(%edi,%ecx,4) + +L(inplace_done): + popl %esi + popl %edi + + movl $0, %eax + popl %ebx + + setc %al + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k6/aorsmul_1.asm b/gmp-6.3.0/mpn/x86/k6/aorsmul_1.asm new file mode 100644 index 0000000..eaa92eb --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/aorsmul_1.asm @@ -0,0 +1,391 @@ +dnl AMD K6 mpn_addmul_1/mpn_submul_1 -- add or subtract mpn multiple. + +dnl Copyright 1999-2003, 2005 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P5 +C P6 model 0-8,10-12 5.94 +C P6 model 9 (Banias) 5.51 +C P6 model 13 (Dothan) 5.57 +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C AMD K6 7.65-8.5 (data dependent) +C AMD K7 +C AMD K8 + + +dnl K6: large multipliers small multipliers +dnl UNROLL_COUNT cycles/limb cycles/limb +dnl 4 9.5 7.78 +dnl 8 9.0 7.78 +dnl 16 8.4 7.65 +dnl 32 8.4 8.2 +dnl +dnl Maximum possible unrolling with the current code is 32. +dnl +dnl Unrolling to 16 limbs/loop makes the unrolled loop fit exactly in a 256 +dnl byte block, which might explain the good speed at that unrolling. + +deflit(UNROLL_COUNT, 16) + + +ifdef(`OPERATION_addmul_1', ` + define(M4_inst, addl) + define(M4_function_1, mpn_addmul_1) + define(M4_function_1c, mpn_addmul_1c) +',`ifdef(`OPERATION_submul_1', ` + define(M4_inst, subl) + define(M4_function_1, mpn_submul_1) + define(M4_function_1c, mpn_submul_1c) +',`m4_error(`Need OPERATION_addmul_1 or OPERATION_submul_1 +')')') + +MULFUNC_PROLOGUE(mpn_addmul_1 mpn_addmul_1c mpn_submul_1 mpn_submul_1c) + + +C mp_limb_t mpn_addmul_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult); +C mp_limb_t mpn_addmul_1c (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult, mp_limb_t carry); +C mp_limb_t mpn_submul_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult); +C mp_limb_t mpn_submul_1c (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult, mp_limb_t carry); +C +C The jadcl0()s in the unrolled loop makes the speed data dependent. Small +C multipliers (most significant few bits clear) result in few carry bits and +C speeds up to 7.65 cycles/limb are attained. Large multipliers (most +C significant few bits set) make the carry bits 50/50 and lead to something +C more like 8.4 c/l. With adcl's both of these would be 9.3 c/l. +C +C It's important that the gains for jadcl0 on small multipliers don't come +C at the cost of slowing down other data. Tests on uniformly distributed +C random data, designed to confound branch prediction, show about a 7% +C speed-up using jadcl0 over adcl (8.93 versus 9.57 cycles/limb, with all +C overheads included). +C +C In the simple loop, jadcl0() measures slower than adcl (11.9-14.7 versus +C 11.0 cycles/limb), and hence isn't used. +C +C In the simple loop, note that running ecx from negative to zero and using +C it as an index in the two movs wouldn't help. It would save one +C instruction (2*addl+loop becoming incl+jnz), but there's nothing unpaired +C that would be collapsed by this. +C +C Attempts at a simpler main loop, with less unrolling, haven't yielded much +C success, generally running over 9 c/l. +C +C +C jadcl0 +C ------ +C +C jadcl0() being faster than adcl $0 seems to be an artifact of two things, +C firstly the instruction decoding and secondly the fact that there's a +C carry bit for the jadcl0 only on average about 1/4 of the time. +C +C The code in the unrolled loop decodes something like the following. +C +C decode cycles +C mull %ebp 2 +C M4_inst %esi, disp(%edi) 1 +C adcl %eax, %ecx 2 +C movl %edx, %esi \ 1 +C jnc 1f / +C incl %esi \ 1 +C 1: movl disp(%ebx), %eax / +C --- +C 7 +C +C In a back-to-back style test this measures 7 with the jnc not taken, or 8 +C with it taken (both when correctly predicted). This is opposite to the +C measurements showing small multipliers running faster than large ones. +C Don't really know why. +C +C It's not clear how much branch misprediction might be costing. The K6 +C doco says it will be 1 to 4 cycles, but presumably it's near the low end +C of that range to get the measured results. +C +C +C In the code the two carries are more or less the preceding mul product and +C the calculation is roughly +C +C x*y + u*b+v +C +C where b=2^32 is the size of a limb, x*y is the two carry limbs, and u and +C v are the two limbs it's added to (being the low of the next mul, and a +C limb from the destination). +C +C To get a carry requires x*y+u*b+v >= b^2, which is u*b+v >= b^2-x*y, and +C there are b^2-(b^2-x*y) = x*y many such values, giving a probability of +C x*y/b^2. If x, y, u and v are random and uniformly distributed between 0 +C and b-1, then the total probability can be summed over x and y, +C +C 1 b-1 b-1 x*y 1 b*(b-1) b*(b-1) +C --- * sum sum --- = --- * ------- * ------- = 1/4 +C b^2 x=0 y=1 b^2 b^4 2 2 +C +C Actually it's a very tiny bit less than 1/4 of course. If y is fixed, +C then the probability is 1/2*y/b thus varying linearly between 0 and 1/2. + + +ifdef(`PIC',` +deflit(UNROLL_THRESHOLD, 9) +',` +deflit(UNROLL_THRESHOLD, 6) +') + +defframe(PARAM_CARRY, 20) +defframe(PARAM_MULTIPLIER,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(32) + +PROLOGUE(M4_function_1c) + pushl %esi +deflit(`FRAME',4) + movl PARAM_CARRY, %esi + jmp L(start_nc) +EPILOGUE() + +PROLOGUE(M4_function_1) + push %esi +deflit(`FRAME',4) + xorl %esi, %esi C initial carry + +L(start_nc): + movl PARAM_SIZE, %ecx + pushl %ebx +deflit(`FRAME',8) + + movl PARAM_SRC, %ebx + pushl %edi +deflit(`FRAME',12) + + cmpl $UNROLL_THRESHOLD, %ecx + movl PARAM_DST, %edi + + pushl %ebp +deflit(`FRAME',16) + jae L(unroll) + + + C simple loop + + movl PARAM_MULTIPLIER, %ebp + +L(simple): + C eax scratch + C ebx src + C ecx counter + C edx scratch + C esi carry + C edi dst + C ebp multiplier + + movl (%ebx), %eax + addl $4, %ebx + + mull %ebp + + addl $4, %edi + addl %esi, %eax + + adcl $0, %edx + + M4_inst %eax, -4(%edi) + + adcl $0, %edx + + movl %edx, %esi + loop L(simple) + + + popl %ebp + popl %edi + + popl %ebx + movl %esi, %eax + + popl %esi + ret + + + +C ----------------------------------------------------------------------------- +C The unrolled loop uses a "two carry limbs" scheme. At the top of the loop +C the carries are ecx=lo, esi=hi, then they swap for each limb processed. +C For the computed jump an odd size means they start one way around, an even +C size the other. +C +C VAR_JUMP holds the computed jump temporarily because there's not enough +C registers at the point of doing the mul for the initial two carry limbs. +C +C The add/adc for the initial carry in %esi is necessary only for the +C mpn_addmul/submul_1c entry points. Duplicating the startup code to +C eliminate this for the plain mpn_add/submul_1 doesn't seem like a good +C idea. + +dnl overlapping with parameters already fetched +define(VAR_COUNTER, `PARAM_SIZE') +define(VAR_JUMP, `PARAM_DST') + +L(unroll): + C eax + C ebx src + C ecx size + C edx + C esi initial carry + C edi dst + C ebp + + movl %ecx, %edx + decl %ecx + + subl $2, %edx + negl %ecx + + shrl $UNROLL_LOG2, %edx + andl $UNROLL_MASK, %ecx + + movl %edx, VAR_COUNTER + movl %ecx, %edx + + shll $4, %edx + negl %ecx + + C 15 code bytes per limb +ifdef(`PIC',` + call L(pic_calc) +L(here): +',` + leal L(entry) (%edx,%ecx,1), %edx +') + movl (%ebx), %eax C src low limb + + movl PARAM_MULTIPLIER, %ebp + movl %edx, VAR_JUMP + + mull %ebp + + addl %esi, %eax C initial carry (from _1c) + jadcl0( %edx) + + + leal 4(%ebx,%ecx,4), %ebx + movl %edx, %esi C high carry + + movl VAR_JUMP, %edx + leal (%edi,%ecx,4), %edi + + testl $1, %ecx + movl %eax, %ecx C low carry + + jz L(noswap) + movl %esi, %ecx C high,low carry other way around + + movl %eax, %esi +L(noswap): + + jmp *%edx + + +ifdef(`PIC',` +L(pic_calc): + C See mpn/x86/README about old gas bugs + leal (%edx,%ecx,1), %edx + addl $L(entry)-L(here), %edx + addl (%esp), %edx + ret_internal +') + + +C ----------------------------------------------------------- + ALIGN(32) +L(top): +deflit(`FRAME',16) + C eax scratch + C ebx src + C ecx carry lo + C edx scratch + C esi carry hi + C edi dst + C ebp multiplier + C + C 15 code bytes per limb + + leal UNROLL_BYTES(%edi), %edi + +L(entry): +forloop(`i', 0, UNROLL_COUNT/2-1, ` + deflit(`disp0', eval(2*i*4)) + deflit(`disp1', eval(disp0 + 4)) + +Zdisp( movl, disp0,(%ebx), %eax) + mull %ebp +Zdisp( M4_inst,%ecx, disp0,(%edi)) + adcl %eax, %esi + movl %edx, %ecx + jadcl0( %ecx) + + movl disp1(%ebx), %eax + mull %ebp + M4_inst %esi, disp1(%edi) + adcl %eax, %ecx + movl %edx, %esi + jadcl0( %esi) +') + + decl VAR_COUNTER + + leal UNROLL_BYTES(%ebx), %ebx + jns L(top) + + + popl %ebp + M4_inst %ecx, UNROLL_BYTES(%edi) + + popl %edi + movl %esi, %eax + + popl %ebx + jadcl0( %eax) + + popl %esi + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k6/cross.pl b/gmp-6.3.0/mpn/x86/k6/cross.pl new file mode 100755 index 0000000..fc921a5 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/cross.pl @@ -0,0 +1,182 @@ +#! /usr/bin/perl + +# Copyright 2000, 2001 Free Software Foundation, Inc. +# +# This file is part of the GNU MP Library. +# +# The GNU MP Library is free software; you can redistribute it and/or modify +# it under the terms of either: +# +# * the GNU Lesser General Public License as published by the Free +# Software Foundation; either version 3 of the License, or (at your +# option) any later version. +# +# or +# +# * the GNU General Public License as published by the Free Software +# Foundation; either version 2 of the License, or (at your option) any +# later version. +# +# or both in parallel, as here. +# +# The GNU MP Library is distributed in the hope that it will be useful, but +# WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +# or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +# for more details. +# +# You should have received copies of the GNU General Public License and the +# GNU Lesser General Public License along with the GNU MP Library. If not, +# see https://www.gnu.org/licenses/. + + +# Usage: cross.pl [filename.o]... +# +# Produce an annotated disassembly of the given object files, indicating +# certain code alignment and addressing mode problems afflicting K6 chips. +# "ZZ" is used on all annotations, so this can be searched for. +# +# With no arguments, all .o files corresponding to .asm files are processed. +# This is good in the mpn object directory of a k6*-*-* build. +# +# Code alignments of 8 bytes or more are handled. When 32 is used, cache +# line boundaries will fall in at offsets 0x20,0x40,etc and problems are +# flagged at those locations. When 16 is used, the line boundaries can also +# fall at offsets 0x10,0x30,0x50,etc, depending where the file is loaded, so +# problems are identified there too. Likewise when 8 byte alignment is used +# problems are flagged additionally at 0x08,0x18,0x28,etc. +# +# Usually 32 byte alignment is used for k6 routines, but less is certainly +# possible if through good luck, or a little tweaking, cache line crossing +# problems can be avoided at the extra locations. +# +# Bugs: +# +# Instructions without mod/rm bytes or which are already vector decoded are +# unaffected by cache line boundary crossing, but not all of these have yet +# been put in as exceptions. All that occur in practice in GMP are present +# though. +# +# There's no messages for using the vector decoded addressing mode (%esi), +# but that's easy to avoid when coding. +# +# Future: +# +# Warn about jump targets that are poorly aligned (less than 2 instructions +# before a cache line boundary). + +use strict; + +sub disassemble { + my ($file) = @_; + my ($addr,$b1,$b2,$b3, $prefix,$opcode,$modrm); + my $align; + + open (IN, "objdump -Srfh $file |") + || die "Cannot open pipe from objdump\n"; + while () { + print; + + if (/^[ \t]*[0-9]+[ \t]+\.text[ \t]/ && /2\*\*([0-9]+)$/) { + $align = 1 << $1; + if ($align < 8) { + print "ZZ cross.pl cannot handle alignment < 2**3\n"; + $align = 8 + } + } + + if (/^[ \t]*([0-9a-f]*):[ \t]*([0-9a-f]+)[ \t]+([0-9a-f]+)[ \t]+([0-9a-f]+)/) { + ($addr,$b1,$b2,$b3) = ($1,$2,$3,$4); + + } elsif (/^[ \t]*([0-9a-f]*):[ \t]*([0-9a-f]+)[ \t]+([0-9a-f]+)/) { + ($addr,$b1,$b2,$b3) = ($1,$2,$3,''); + + } elsif (/^[ \t]*([0-9a-f]*):[ \t]*([0-9a-f]+)/) { + ($addr,$b1,$b2,$b3) = ($1,$2,'',''); + + } else { + next; + } + + if ($b1 =~ /0f/) { + $prefix = $b1; + $opcode = $b2; + $modrm = $b3; + } else { + $prefix = ''; + $opcode = $b1; + $modrm = $b2; + } + + # modrm of the form 00-xxx-100 with an 0F prefix is the problem case + # for K6 and pre-CXT K6-2 + if ($prefix =~ /0f/ + && $opcode !~ /^8/ # jcond disp32 + && $modrm =~ /^[0-3][4c]/) { + print "ZZ ($file) >3 bytes to determine instruction length [K6]\n"; + } + + # with just an opcode, starting 1f mod 20h + if (($align==32 && $addr =~ /[13579bdf]f$/ + || $align==16 && $addr =~ /f$/ + || $align==8 && $addr =~ /[7f]$/) + && $prefix !~ /0f/ + && $opcode !~ /1[012345]/ # adc + && $opcode !~ /1[89abcd]/ # sbb + && $opcode !~ /^4/ # inc/dec reg + && $opcode !~ /^5/ # push/pop reg + && $opcode !~ /68/ # push $imm32 + && $opcode !~ /^7/ # jcond disp8 + && $opcode !~ /a[89]/ # test+imm + && $opcode !~ /a[a-f]/ # stos/lods/scas + && $opcode !~ /b8/ # movl $imm32,%eax + && $opcode !~ /d[0123]/ # rcl + && $opcode !~ /e[0123]/ # loop/loopz/loopnz/jcxz + && $opcode !~ /e8/ # call disp32 + && $opcode !~ /e[9b]/ # jmp disp32/disp8 + && $opcode !~ /f[89abcd]/ # clc,stc,cli,sti,cld,std + && !($opcode =~ /f[67]/ # grp 1 + && $modrm =~ /^[2367abef]/) # mul, imul, div, idiv + && $modrm !~ /^$/) { + print "ZZ ($file) opcode/modrm cross 32-byte boundary\n"; + } + + # with an 0F prefix, anything starting at 1f mod 20h + if (($align==32 && $addr =~ /[13579bdf][f]$/ + || $align==16 && $addr =~ /f$/ + || $align==8 && $addr =~ /[7f]$/) + && $prefix =~ /0f/ + && $opcode !~ /af/ # imul + && $opcode !~ /a[45]/ # shldl + && $opcode !~ /a[cd]/ # shrdl + ) { + print "ZZ ($file) prefix/opcode cross 32-byte boundary\n"; + } + + # with an 0F prefix, anything with mod/rm starting at 1e mod 20h + if (($align==32 && $addr =~ /[13579bdf][e]$/ + || $align==16 && $addr =~ /[e]$/ + || $align==8 && $addr =~ /[6e]$/) + && $prefix =~ /0f/ + && $opcode !~ /^8/ # jcond disp32 + && $opcode !~ /af/ # imull reg,reg + && $opcode !~ /a[45]/ # shldl + && $opcode !~ /a[cd]/ # shrdl + && $modrm !~ /^$/) { + print "ZZ ($file) prefix/opcode/modrm cross 32-byte boundary\n"; + } + } + close IN || die "Error from objdump (or objdump not available)\n"; +} + + +my @files; +if ($#ARGV >= 0) { + @files = @ARGV; +} else { + @files = glob "*.asm"; + map {s/.asm/.o/} @files; +} + +foreach (@files) { + disassemble($_); +} diff --git a/gmp-6.3.0/mpn/x86/k6/divrem_1.asm b/gmp-6.3.0/mpn/x86/k6/divrem_1.asm new file mode 100644 index 0000000..b4cea4f --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/divrem_1.asm @@ -0,0 +1,203 @@ +dnl AMD K6 mpn_divrem_1 -- mpn by limb division. + +dnl Copyright 1999-2003, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K6: 20 cycles/limb + + +C mp_limb_t mpn_divrem_1 (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, mp_limb_t divisor); +C mp_limb_t mpn_divrem_1c (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, mp_limb_t divisor, +C mp_limb_t carry); +C +C The code here is basically the same as mpn/x86/divrem_1.asm, but uses loop +C instead of decl+jnz, since it comes out 2 cycles/limb faster. +C +C A test is done to see if the high limb is less than the divisor, and if so +C one less div is done. A div is 20 cycles, so assuming high= 0 and + C and leaves 0 to 3 which can be tested with test $1 and $2. + +L(top): + C eax counter, -(size-7) step by +4 until >=0 + C ebx src end - 32 + C ecx dst end - 32 + C edx retval + C + C mm0 src next qword + C mm1 scratch + C mm2 src prev qword + C mm6 shift + C mm7 64-shift + + psrlq %mm6, %mm2 + addl $4, %eax + + movq %mm0, %mm1 + psllq %mm7, %mm0 + + por %mm0, %mm2 + movq 4(%ebx,%eax,4), %mm0 + + psrlq %mm6, %mm1 + movq %mm2, -12(%ecx,%eax,4) + + movq %mm0, %mm2 + psllq %mm7, %mm0 + + por %mm0, %mm1 + movq 12(%ebx,%eax,4), %mm0 + + movq %mm1, -4(%ecx,%eax,4) + ja L(top) C jump if no carry and not zero + + + + C Now have the four limbs in mm2 (low) and mm0 (high), and %eax is 0 + C to 3 representing respectively 3 to 0 further limbs. + + testl $2, %eax C testl to avoid bad cache line crossings + jnz L(finish_nottwo) + + C Two or three extra limbs: rshift mm2, OR it with lshifted mm0, mm0 + C becomes new mm2 and a new mm0 is loaded. + + psrlq %mm6, %mm2 + movq %mm0, %mm1 + + psllq %mm7, %mm0 + addl $2, %eax + + por %mm0, %mm2 + movq 12(%ebx,%eax,4), %mm0 + + movq %mm2, -4(%ecx,%eax,4) + movq %mm1, %mm2 +L(finish_nottwo): + + + testb $1, %al + psrlq %mm6, %mm2 + + movq %mm0, %mm1 + psllq %mm7, %mm0 + + por %mm0, %mm2 + psrlq %mm6, %mm1 + + movq %mm2, 4(%ecx,%eax,4) + jnz L(finish_even) + + + C one further extra limb to process + + movd 32-4(%ebx), %mm0 C src[size-1], most significant limb + popl %ebx + + movq %mm0, %mm2 + psllq %mm7, %mm0 + + por %mm0, %mm1 + psrlq %mm6, %mm2 + + movq %mm1, 32-12(%ecx) C dst[size-3,size-2] + movd %mm2, 32-4(%ecx) C dst[size-1] + + movl %edx, %eax C retval + + femms + ret + + + nop C avoid bad cache line crossing +L(finish_even): + C no further extra limbs + + movq %mm1, 32-8(%ecx) C dst[size-2,size-1] + movl %edx, %eax C retval + + popl %ebx + + femms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k6/mmx/com.asm b/gmp-6.3.0/mpn/x86/k6/mmx/com.asm new file mode 100644 index 0000000..b747454 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/mmx/com.asm @@ -0,0 +1,103 @@ +dnl AMD K6-2 mpn_com -- mpn bitwise one's complement. + +dnl Copyright 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +NAILS_SUPPORT(0-31) + + +C alignment dst/src, A=0mod8 N=4mod8 +C A/A A/N N/A N/N +C K6-2 1.0 1.18 1.18 1.18 cycles/limb +C K6 1.5 1.85 1.75 1.85 + + +C void mpn_com (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C Take the bitwise ones-complement of src,size and write it to dst,size. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(16) +PROLOGUE(mpn_com) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + movl PARAM_SRC, %eax + movl PARAM_DST, %edx + shrl %ecx + jnz L(two_or_more) + + movl (%eax), %eax + notl_or_xorl_GMP_NUMB_MASK( %eax) + movl %eax, (%edx) + ret + + +L(two_or_more): + pushl %ebx FRAME_pushl() + pcmpeqd %mm7, %mm7 C all ones + + movl %ecx, %ebx +ifelse(GMP_NAIL_BITS,0,, +` psrld $GMP_NAIL_BITS, %mm7') C clear nails + + + + ALIGN(8) +L(top): + C eax src + C ebx floor(size/2) + C ecx counter + C edx dst + C + C mm0 scratch + C mm7 mask + + movq -8(%eax,%ecx,8), %mm0 + pxor %mm7, %mm0 + movq %mm0, -8(%edx,%ecx,8) + loop L(top) + + + jnc L(no_extra) + movl (%eax,%ebx,8), %eax + notl_or_xorl_GMP_NUMB_MASK( %eax) + movl %eax, (%edx,%ebx,8) +L(no_extra): + + popl %ebx + emms_or_femms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k6/mmx/dive_1.asm b/gmp-6.3.0/mpn/x86/k6/mmx/dive_1.asm new file mode 100644 index 0000000..1bbad3a --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/mmx/dive_1.asm @@ -0,0 +1,282 @@ +dnl AMD K6 mpn_divexact_1 -- mpn by limb exact division. + +dnl Copyright 2000-2002, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C divisor +C odd even +C K6: 10.0 12.0 cycles/limb +C K6-2: 10.0 11.5 + + +C void mpn_divexact_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C +C A simple divl is used for size==1. This is about 10 cycles faster for an +C odd divisor or 20 cycles for an even divisor. +C +C The loops are quite sensitive to code alignment, speeds should be +C rechecked (odd and even divisor, pic and non-pic) if contemplating +C changing anything. + +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(VAR_INVERSE,`PARAM_DST') + + TEXT + + ALIGN(32) +PROLOGUE(mpn_divexact_1) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + + movl PARAM_SRC, %eax + xorl %edx, %edx + + cmpl $1, %ecx + jnz L(two_or_more) + + movl (%eax), %eax + + divl PARAM_DIVISOR + + movl PARAM_DST, %ecx + movl %eax, (%ecx) + + ret + + +L(two_or_more): + movl PARAM_DIVISOR, %eax + pushl %ebx FRAME_pushl() + + movl PARAM_SRC, %ebx + pushl %ebp FRAME_pushl() + +L(strip_twos): + shrl %eax + incl %edx C will get shift+1 + + jnc L(strip_twos) + pushl %esi FRAME_pushl() + + leal 1(%eax,%eax), %esi C d without twos + andl $127, %eax C d/2, 7 bits + +ifdef(`PIC',` + LEA( binvert_limb_table, %ebp) +Zdisp( movzbl, 0,(%eax,%ebp), %eax) +',` + movzbl binvert_limb_table(%eax), %eax C inv 8 bits +') + pushl %edi FRAME_pushl() + + leal (%eax,%eax), %ebp C 2*inv + + imull %eax, %eax C inv*inv + + movl PARAM_DST, %edi + + imull %esi, %eax C inv*inv*d + + subl %eax, %ebp C inv = 2*inv - inv*inv*d + leal (%ebp,%ebp), %eax C 2*inv + + imull %ebp, %ebp C inv*inv + + movl %esi, PARAM_DIVISOR C d without twos + leal (%ebx,%ecx,4), %ebx C src end + + imull %esi, %ebp C inv*inv*d + + leal (%edi,%ecx,4), %edi C dst end + negl %ecx C -size + + subl %ebp, %eax C inv = 2*inv - inv*inv*d + subl $1, %edx C shift amount, and clear carry + + ASSERT(e,` C expect d*inv == 1 mod 2^GMP_LIMB_BITS + pushl %eax FRAME_pushl() + imull PARAM_DIVISOR, %eax + cmpl $1, %eax + popl %eax FRAME_popl()') + + movl %eax, VAR_INVERSE + jnz L(even) + + movl (%ebx,%ecx,4), %esi C src low limb + jmp L(odd_entry) + + + ALIGN(16) + nop C code alignment +L(odd_top): + C eax scratch + C ebx src end + C ecx counter, limbs, negative + C edx inverse + C esi next limb, adjusted for carry + C edi dst end + C ebp carry bit, 0 or -1 + + imull %edx, %esi + + movl PARAM_DIVISOR, %eax + movl %esi, -4(%edi,%ecx,4) + + mull %esi C carry limb in edx + + subl %ebp, %edx C apply carry bit + movl (%ebx,%ecx,4), %esi + +L(odd_entry): + subl %edx, %esi C apply carry limb + movl VAR_INVERSE, %edx + + sbbl %ebp, %ebp C 0 or -1 + + incl %ecx + jnz L(odd_top) + + + imull %edx, %esi + + movl %esi, -4(%edi,%ecx,4) + + popl %edi + popl %esi + + popl %ebp + popl %ebx + + ret + + +L(even): + C eax + C ebx src end + C ecx -size + C edx twos + C esi + C edi dst end + C ebp + + xorl %ebp, %ebp +Zdisp( movq, 0,(%ebx,%ecx,4), %mm0) C src[0,1] + + movd %edx, %mm7 + movl VAR_INVERSE, %edx + + addl $2, %ecx + psrlq %mm7, %mm0 + + movd %mm0, %esi + jz L(even_two) C if only two limbs + + +C Out-of-order execution is good enough to hide the load/rshift/movd +C latency. Having imul at the top of the loop gives 11.5 c/l instead of 12, +C on K6-2. In fact there's only 11 of decode, but nothing running at 11 has +C been found. Maybe the fact every second movq is unaligned costs the extra +C 0.5. + +L(even_top): + C eax scratch + C ebx src end + C ecx counter, limbs, negative + C edx inverse + C esi next limb, adjusted for carry + C edi dst end + C ebp carry bit, 0 or -1 + C + C mm0 scratch, source limbs + C mm7 twos + + imull %edx, %esi + + movl %esi, -8(%edi,%ecx,4) + movl PARAM_DIVISOR, %eax + + mull %esi C carry limb in edx + + movq -4(%ebx,%ecx,4), %mm0 + psrlq %mm7, %mm0 + + movd %mm0, %esi + subl %ebp, %edx C apply carry bit + + subl %edx, %esi C apply carry limb + movl VAR_INVERSE, %edx + + sbbl %ebp, %ebp C 0 or -1 + + incl %ecx + jnz L(even_top) + + +L(even_two): + movd -4(%ebx), %mm0 C src high limb + psrlq %mm7, %mm0 + + imull %edx, %esi + + movl %esi, -8(%edi) + movl PARAM_DIVISOR, %eax + + mull %esi C carry limb in edx + + movd %mm0, %esi + subl %ebp, %edx C apply carry bit + + movl VAR_INVERSE, %eax + subl %edx, %esi C apply carry limb + + imull %eax, %esi + + movl %esi, -4(%edi) + + popl %edi + popl %esi + + popl %ebp + popl %ebx + + emms_or_femms + + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/k6/mmx/logops_n.asm b/gmp-6.3.0/mpn/x86/k6/mmx/logops_n.asm new file mode 100644 index 0000000..e17930b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/mmx/logops_n.asm @@ -0,0 +1,226 @@ +dnl AMD K6-2 mpn_and_n, mpn_andn_n, mpn_nand_n, mpn_ior_n, mpn_iorn_n, +dnl mpn_nior_n, mpn_xor_n, mpn_xnor_n -- mpn bitwise logical operations. + +dnl Copyright 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +NAILS_SUPPORT(0-31) + + +C alignment dst/src1/src2, A=0mod8, N=4mod8 +C A/A/A A/A/N A/N/A A/N/N N/A/A N/A/N N/N/A N/N/N +C +C K6-2 1.2 1.5 1.5 1.2 1.2 1.5 1.5 1.2 and,andn,ior,xor +C K6-2 1.5 1.75 2.0 1.75 1.75 2.0 1.75 1.5 iorn,xnor +C K6-2 1.75 2.0 2.0 2.0 2.0 2.0 2.0 1.75 nand,nior +C +C K6 1.5 1.68 1.75 1.2 1.75 1.75 1.68 1.5 and,andn,ior,xor +C K6 2.0 2.0 2.25 2.25 2.25 2.25 2.0 2.0 iorn,xnor +C K6 2.0 2.25 2.25 2.25 2.25 2.25 2.25 2.0 nand,nior + + +dnl M4_p and M4_i are the MMX and integer instructions +dnl M4_*_neg_dst means whether to negate the final result before writing +dnl M4_*_neg_src2 means whether to negate the src2 values before using them + +define(M4_choose_op, +m4_assert_numargs(7) +`ifdef(`OPERATION_$1',` +define(`M4_function', `mpn_$1') +define(`M4_operation', `$1') +define(`M4_p', `$2') +define(`M4_p_neg_dst', `$3') +define(`M4_p_neg_src2',`$4') +define(`M4_i', `$5') +define(`M4_i_neg_dst', `$6') +define(`M4_i_neg_src2',`$7') +')') + +dnl xnor is done in "iorn" style because it's a touch faster than "nior" +dnl style (the two are equivalent for xor). +dnl +dnl pandn can't be used with nails. + +M4_choose_op( and_n, pand,0,0, andl,0,0) +ifelse(GMP_NAIL_BITS,0, +`M4_choose_op(andn_n, pandn,0,0, andl,0,1)', +`M4_choose_op(andn_n, pand,0,1, andl,0,1)') +M4_choose_op( nand_n, pand,1,0, andl,1,0) +M4_choose_op( ior_n, por,0,0, orl,0,0) +M4_choose_op( iorn_n, por,0,1, orl,0,1) +M4_choose_op( nior_n, por,1,0, orl,1,0) +M4_choose_op( xor_n, pxor,0,0, xorl,0,0) +M4_choose_op( xnor_n, pxor,0,1, xorl,0,1) + +ifdef(`M4_function',, +`m4_error(`Unrecognised or undefined OPERATION symbol +')') + +MULFUNC_PROLOGUE(mpn_and_n mpn_andn_n mpn_nand_n mpn_ior_n mpn_iorn_n mpn_nior_n mpn_xor_n mpn_xnor_n) + + +C void M4_function (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size); +C +C Do src1,size M4_operation src2,size, storing the result in dst,size. +C +C Unaligned movq loads and stores are a bit slower than aligned ones. The +C test at the start of the routine checks the alignment of src1 and if +C necessary processes one limb separately at the low end to make it aligned. +C +C The raw speeds without this alignment switch are as follows. +C +C alignment dst/src1/src2, A=0mod8, N=4mod8 +C A/A/A A/A/N A/N/A A/N/N N/A/A N/A/N N/N/A N/N/N +C +C K6 1.5 2.0 1.5 2.0 and,andn,ior,xor +C K6 1.75 2.2 2.0 2.28 iorn,xnor +C K6 2.0 2.25 2.35 2.28 nand,nior +C +C +C Future: +C +C K6 can do one 64-bit load per cycle so each of these routines should be +C able to approach 1.0 c/l, if aligned. The basic and/andn/ior/xor might be +C able to get 1.0 with just a 4 limb loop, being 3 instructions per 2 limbs. +C The others are 4 instructions per 2 limbs, and so can only approach 1.0 +C because there's nowhere to hide some loop control. + +defframe(PARAM_SIZE,16) +defframe(PARAM_SRC2,12) +defframe(PARAM_SRC1,8) +defframe(PARAM_DST, 4) +deflit(`FRAME',0) + + TEXT + ALIGN(32) +PROLOGUE(M4_function) + movl PARAM_SIZE, %ecx + pushl %ebx FRAME_pushl() + + movl PARAM_SRC1, %eax + + movl PARAM_SRC2, %ebx + cmpl $1, %ecx + + movl PARAM_DST, %edx + ja L(two_or_more) + + + movl (%ebx), %ecx + popl %ebx +ifelse(M4_i_neg_src2,1,`notl_or_xorl_GMP_NUMB_MASK( %ecx)') + M4_i (%eax), %ecx +ifelse(M4_i_neg_dst,1,` notl_or_xorl_GMP_NUMB_MASK( %ecx)') + movl %ecx, (%edx) + + ret + + +L(two_or_more): + C eax src1 + C ebx src2 + C ecx size + C edx dst + C esi + C edi + C ebp + + pushl %esi FRAME_pushl() + testl $4, %eax + jz L(alignment_ok) + + movl (%ebx), %esi + addl $4, %ebx +ifelse(M4_i_neg_src2,1,`notl_or_xorl_GMP_NUMB_MASK( %esi)') + M4_i (%eax), %esi + addl $4, %eax +ifelse(M4_i_neg_dst,1,` notl_or_xorl_GMP_NUMB_MASK( %esi)') + movl %esi, (%edx) + addl $4, %edx + decl %ecx + +L(alignment_ok): + movl %ecx, %esi + shrl %ecx + jnz L(still_two_or_more) + + movl (%ebx), %ecx + popl %esi +ifelse(M4_i_neg_src2,1,`notl_or_xorl_GMP_NUMB_MASK( %ecx)') + M4_i (%eax), %ecx +ifelse(M4_i_neg_dst,1,` notl_or_xorl_GMP_NUMB_MASK( %ecx)') + popl %ebx + movl %ecx, (%edx) + ret + + +L(still_two_or_more): +ifelse(eval(M4_p_neg_src2 || M4_p_neg_dst),1,` + pcmpeqd %mm7, %mm7 C all ones +ifelse(GMP_NAIL_BITS,0,,`psrld $GMP_NAIL_BITS, %mm7') C clear nails +') + + ALIGN(16) +L(top): + C eax src1 + C ebx src2 + C ecx counter + C edx dst + C esi + C edi + C ebp + C + C carry bit is low of size + + movq -8(%ebx,%ecx,8), %mm0 +ifelse(M4_p_neg_src2,1,`pxor %mm7, %mm0') + M4_p -8(%eax,%ecx,8), %mm0 +ifelse(M4_p_neg_dst,1,` pxor %mm7, %mm0') + movq %mm0, -8(%edx,%ecx,8) + + loop L(top) + + + jnc L(no_extra) + + movl -4(%ebx,%esi,4), %ebx +ifelse(M4_i_neg_src2,1,`notl_or_xorl_GMP_NUMB_MASK( %ebx)') + M4_i -4(%eax,%esi,4), %ebx +ifelse(M4_i_neg_dst,1,` notl_or_xorl_GMP_NUMB_MASK( %ebx)') + movl %ebx, -4(%edx,%esi,4) +L(no_extra): + + popl %esi + popl %ebx + emms_or_femms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k6/mmx/lshift.asm b/gmp-6.3.0/mpn/x86/k6/mmx/lshift.asm new file mode 100644 index 0000000..45be582 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/mmx/lshift.asm @@ -0,0 +1,130 @@ +dnl AMD K6 mpn_lshift -- mpn left shift. + +dnl Copyright 1999, 2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K6: 3.0 cycles/limb + + +C mp_limb_t mpn_lshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned shift); +C +C The loop runs at 3 cycles/limb, limited by decoding and by having 3 mmx +C instructions. This is despite every second fetch being unaligned. + + +defframe(PARAM_SHIFT,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(32) + +PROLOGUE(mpn_lshift) +deflit(`FRAME',0) + + C The 1 limb case can be done without the push %ebx, but it's then + C still the same speed. The push is left as a free helping hand for + C the two_or_more code. + + movl PARAM_SIZE, %eax + pushl %ebx FRAME_pushl() + + movl PARAM_SRC, %ebx + decl %eax + + movl PARAM_SHIFT, %ecx + jnz L(two_or_more) + + movl (%ebx), %edx C src limb + movl PARAM_DST, %ebx + + shldl( %cl, %edx, %eax) C return value + + shll %cl, %edx + + movl %edx, (%ebx) C dst limb + popl %ebx + + ret + + + ALIGN(16) C avoid offset 0x1f + nop C avoid bad cache line crossing +L(two_or_more): + C eax size-1 + C ebx src + C ecx shift + C edx + + movl (%ebx,%eax,4), %edx C src high limb + negl %ecx + + movd PARAM_SHIFT, %mm6 + addl $32, %ecx C 32-shift + + shrl %cl, %edx + + movd %ecx, %mm7 + movl PARAM_DST, %ecx + +L(top): + C eax counter, size-1 to 1 + C ebx src + C ecx dst + C edx retval + C + C mm0 scratch + C mm6 shift + C mm7 32-shift + + movq -4(%ebx,%eax,4), %mm0 + decl %eax + + psrlq %mm7, %mm0 + + movd %mm0, 4(%ecx,%eax,4) + jnz L(top) + + + movd (%ebx), %mm0 + popl %ebx + + psllq %mm6, %mm0 + movl %edx, %eax + + movd %mm0, (%ecx) + + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k6/mmx/popham.asm b/gmp-6.3.0/mpn/x86/k6/mmx/popham.asm new file mode 100644 index 0000000..2b19d0b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/mmx/popham.asm @@ -0,0 +1,236 @@ +dnl AMD K6-2 mpn_popcount, mpn_hamdist -- mpn bit population count and +dnl hamming distance. + +dnl Copyright 2000-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C popcount hamdist +C K6-2: 9.0 11.5 cycles/limb +C K6: 12.5 13.0 + + +C unsigned long mpn_popcount (mp_srcptr src, mp_size_t size); +C unsigned long mpn_hamdist (mp_srcptr src, mp_srcptr src2, mp_size_t size); +C +C The code here isn't optimal, but it's already a 2x speedup over the plain +C integer mpn/generic/popcount.c,hamdist.c. + + +ifdef(`OPERATION_popcount',, +`ifdef(`OPERATION_hamdist',, +`m4_error(`Need OPERATION_popcount or OPERATION_hamdist +')m4exit(1)')') + +define(HAM, +m4_assert_numargs(1) +`ifdef(`OPERATION_hamdist',`$1')') + +define(POP, +m4_assert_numargs(1) +`ifdef(`OPERATION_popcount',`$1')') + +HAM(` +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC2, 8) +defframe(PARAM_SRC, 4) +define(M4_function,mpn_hamdist) +') +POP(` +defframe(PARAM_SIZE, 8) +defframe(PARAM_SRC, 4) +define(M4_function,mpn_popcount) +') + +MULFUNC_PROLOGUE(mpn_popcount mpn_hamdist) + + +ifdef(`PIC',,` + dnl non-PIC + + RODATA + ALIGN(8) + +L(rodata_AAAAAAAAAAAAAAAA): + .long 0xAAAAAAAA + .long 0xAAAAAAAA + +L(rodata_3333333333333333): + .long 0x33333333 + .long 0x33333333 + +L(rodata_0F0F0F0F0F0F0F0F): + .long 0x0F0F0F0F + .long 0x0F0F0F0F + +L(rodata_000000FF000000FF): + .long 0x000000FF + .long 0x000000FF +') + + TEXT + ALIGN(32) + +POP(`ifdef(`PIC', ` + C avoid shrl crossing a 32-byte boundary + nop')') + +PROLOGUE(M4_function) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + +ifdef(`PIC',` + movl $0xAAAAAAAA, %eax + movl $0x33333333, %edx + + movd %eax, %mm7 + movd %edx, %mm6 + + movl $0x0F0F0F0F, %eax + movl $0x000000FF, %edx + + punpckldq %mm7, %mm7 + punpckldq %mm6, %mm6 + + movd %eax, %mm5 + movd %edx, %mm4 + + punpckldq %mm5, %mm5 + punpckldq %mm4, %mm4 +',` + + movq L(rodata_AAAAAAAAAAAAAAAA), %mm7 + movq L(rodata_3333333333333333), %mm6 + movq L(rodata_0F0F0F0F0F0F0F0F), %mm5 + movq L(rodata_000000FF000000FF), %mm4 +') + +define(REG_AAAAAAAAAAAAAAAA, %mm7) +define(REG_3333333333333333, %mm6) +define(REG_0F0F0F0F0F0F0F0F, %mm5) +define(REG_000000FF000000FF, %mm4) + + + movl PARAM_SRC, %eax +HAM(` movl PARAM_SRC2, %edx') + + pxor %mm2, %mm2 C total + + shrl %ecx + jnc L(top) + +Zdisp( movd, 0,(%eax,%ecx,8), %mm1) + +HAM(` +Zdisp( movd, 0,(%edx,%ecx,8), %mm0) + pxor %mm0, %mm1 +') + + incl %ecx + jmp L(loaded) + + + ALIGN(16) +POP(` nop C alignment to avoid crossing 32-byte boundaries') + +L(top): + C eax src + C ebx + C ecx counter, qwords, decrementing + C edx [hamdist] src2 + C + C mm0 (scratch) + C mm1 (scratch) + C mm2 total (low dword) + C mm3 + C mm4 \ + C mm5 | special constants + C mm6 | + C mm7 / + + movq -8(%eax,%ecx,8), %mm1 +HAM(` pxor -8(%edx,%ecx,8), %mm1') + +L(loaded): + movq %mm1, %mm0 + pand REG_AAAAAAAAAAAAAAAA, %mm1 + + psrlq $1, %mm1 +HAM(` nop C code alignment') + + psubd %mm1, %mm0 C bit pairs +HAM(` nop C code alignment') + + + movq %mm0, %mm1 + psrlq $2, %mm0 + + pand REG_3333333333333333, %mm0 + pand REG_3333333333333333, %mm1 + + paddd %mm1, %mm0 C nibbles + + + movq %mm0, %mm1 + psrlq $4, %mm0 + + pand REG_0F0F0F0F0F0F0F0F, %mm0 + pand REG_0F0F0F0F0F0F0F0F, %mm1 + + paddd %mm1, %mm0 C bytes + + movq %mm0, %mm1 + psrlq $8, %mm0 + + + paddb %mm1, %mm0 C words + + + movq %mm0, %mm1 + psrlq $16, %mm0 + + paddd %mm1, %mm0 C dwords + + pand REG_000000FF000000FF, %mm0 + + paddd %mm0, %mm2 C low to total + psrlq $32, %mm0 + + paddd %mm0, %mm2 C high to total + loop L(top) + + + + movd %mm2, %eax + emms_or_femms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k6/mmx/rshift.asm b/gmp-6.3.0/mpn/x86/k6/mmx/rshift.asm new file mode 100644 index 0000000..cd0382f --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/mmx/rshift.asm @@ -0,0 +1,130 @@ +dnl AMD K6 mpn_rshift -- mpn right shift. + +dnl Copyright 1999, 2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K6: 3.0 cycles/limb + + +C mp_limb_t mpn_rshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned shift); +C +C The loop runs at 3 cycles/limb, limited by decoding and by having 3 mmx +C instructions. This is despite every second fetch being unaligned. + + +defframe(PARAM_SHIFT,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) +deflit(`FRAME',0) + + TEXT + ALIGN(32) + +PROLOGUE(mpn_rshift) +deflit(`FRAME',0) + + C The 1 limb case can be done without the push %ebx, but it's then + C still the same speed. The push is left as a free helping hand for + C the two_or_more code. + + movl PARAM_SIZE, %eax + pushl %ebx FRAME_pushl() + + movl PARAM_SRC, %ebx + decl %eax + + movl PARAM_SHIFT, %ecx + jnz L(two_or_more) + + movl (%ebx), %edx C src limb + movl PARAM_DST, %ebx + + shrdl( %cl, %edx, %eax) C return value + + shrl %cl, %edx + + movl %edx, (%ebx) C dst limb + popl %ebx + + ret + + + ALIGN(16) C avoid offset 0x1f +L(two_or_more): + C eax size-1 + C ebx src + C ecx shift + C edx + + movl (%ebx), %edx C src low limb + negl %ecx + + addl $32, %ecx C 32-shift + movd PARAM_SHIFT, %mm6 + + shll %cl, %edx C retval + movl PARAM_DST, %ecx + + leal (%ebx,%eax,4), %ebx + + leal -4(%ecx,%eax,4), %ecx + negl %eax + + +L(simple): + C eax counter (negative) + C ebx &src[size-1] + C ecx &dst[size-1] + C edx retval + C + C mm0 scratch + C mm6 shift + +Zdisp( movq, 0,(%ebx,%eax,4), %mm0) + incl %eax + + psrlq %mm6, %mm0 + +Zdisp( movd, %mm0, 0,(%ecx,%eax,4)) + jnz L(simple) + + + movq %mm0, (%ecx) + movl %edx, %eax + + popl %ebx + + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k6/mod_34lsub1.asm b/gmp-6.3.0/mpn/x86/k6/mod_34lsub1.asm new file mode 100644 index 0000000..7e30503 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/mod_34lsub1.asm @@ -0,0 +1,190 @@ +dnl AMD K6 mpn_mod_34lsub1 -- mpn remainder modulo 2**24-1. + +dnl Copyright 2000-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K6: 2.66 cycles/limb + + +C mp_limb_t mpn_mod_34lsub1 (mp_srcptr src, mp_size_t size) +C +C An attempt was made to use a loop like +C +C L(top): +C adcl (%edx), %eax +C adcl 4(%edx), %ebx +C adcl 8(%edx), %esi +C leal 12(%edx), %edx +C loop L(top) +C +C with %ecx starting from floor(size/3), but it still measured 2.66 c/l. +C The form used instead can save about 6 cycles by not dividing by 3. +C +C In the code used, putting the "leal"s at the top of the loop is necessary +C for the claimed speed, anywhere else costs an extra cycle per loop. +C Perhaps a tight loop like this needs short decode instructions at the +C branch target, which would explain the leal/loop form above taking 8 +C cycles instead of 7 too. + +defframe(PARAM_SIZE, 8) +defframe(PARAM_SRC, 4) + +dnl re-use parameter space +define(SAVE_EBX, `PARAM_SIZE') +define(SAVE_ESI, `PARAM_SRC') + + TEXT + ALIGN(16) +PROLOGUE(mpn_mod_34lsub1) +deflit(`FRAME',0) + + movl PARAM_SIZE, %eax + movl PARAM_SRC, %edx + + subl $2, %eax + ja L(three_or_more) + +Zdisp( movl, 0,(%edx), %eax) C avoid code cache line boundary + jne L(one) + + movl %eax, %ecx + movl 4(%edx), %edx + + shrl $24, %eax C src[0] high + andl $0x00FFFFFF, %ecx C src[0] low + + addl %ecx, %eax + movl %edx, %ecx + + shll $8, %edx + andl $0x00FFFF00, %edx C src[1] high + + shrl $16, %ecx C src[1] low + addl %ecx, %eax + + addl %edx, %eax + +L(one): + ret + + +L(three_or_more): + C eax size-2 + C ebx + C ecx + C edx src + + movl %ebx, SAVE_EBX + xorl %ebx, %ebx + + movl %esi, SAVE_ESI + pushl %edi FRAME_pushl() + + xorl %esi, %esi + xorl %edi, %edi C and clear carry flag + +L(top): + C eax counter, limbs + C ebx acc 0mod3 + C ecx + C edx src, incrementing + C esi acc 1mod3 + C edi acc 2mod3 + C ebp + + leal -2(%eax), %eax + leal 12(%edx), %edx + + adcl -12(%edx), %ebx + adcl -8(%edx), %esi + adcl -4(%edx), %edi + + decl %eax + jg L(top) + + + C ecx is -3, -2 or -1 representing 0, 1 or 2 more limbs, respectively + + movb $0, %cl + incl %eax + + js L(combine) C 0 more + +Zdisp( adcl, 0,(%edx), %ebx) C avoid code cache line crossings + + movb $8, %cl + decl %eax + + js L(combine) C 1 more + + adcl 4(%edx), %esi + + movb $16, %cl + + +L(combine): + sbbl %edx, %edx + + shll %cl, %edx C carry + movl %ebx, %eax C 0mod3 + + shrl $24, %eax C 0mod3 high + andl $0x00FFFFFF, %ebx C 0mod3 low + + subl %edx, %eax C apply carry + movl %esi, %ecx C 1mod3 + + shrl $16, %esi C 1mod3 high + addl %ebx, %eax C apply 0mod3 low + + andl $0x0000FFFF, %ecx + addl %esi, %eax C apply 1mod3 high + + shll $8, %ecx C 1mod3 low + movl %edi, %edx C 2mod3 + + shrl $8, %edx C 2mod3 high + addl %ecx, %eax C apply 1mod3 low + + addl %edx, %eax C apply 2mod3 high + andl $0x000000FF, %edi + + shll $16, %edi C 2mod3 low + movl SAVE_EBX, %ebx + + addl %edi, %eax C apply 2mod3 low + movl SAVE_ESI, %esi + + popl %edi + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k6/mode1o.asm b/gmp-6.3.0/mpn/x86/k6/mode1o.asm new file mode 100644 index 0000000..4a338bd --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k6/mode1o.asm @@ -0,0 +1,176 @@ +dnl AMD K6 mpn_modexact_1_odd -- exact division style remainder. + +dnl Copyright 2000-2003, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K6: 10.0 cycles/limb + + +C mp_limb_t mpn_modexact_1_odd (mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C mp_limb_t mpn_modexact_1c_odd (mp_srcptr src, mp_size_t size, +C mp_limb_t divisor, mp_limb_t carry); +C +C A special case for high31),1, +eval((UNROLL_COUNT-31)*4), +0)) + + C eax + C ebx &src[size] + C ecx + C edx + C esi carry + C edi &dst[size] + C ebp + + movl PARAM_SIZE, %ecx + movl %esi, (%edi) + + subl $4, %ecx + jz L(corner) + + movl %ecx, %edx +ifelse(OFFSET,0,, +` subl $OFFSET, %ebx') + + shll $4, %ecx +ifelse(OFFSET,0,, +` subl $OFFSET, %edi') + + negl %ecx + +ifdef(`PIC',` + call L(pic_calc) +L(here): +',` + leal L(unroll_inner_end)-eval(2*CODE_BYTES_PER_LIMB)(%ecx,%edx), %ecx +') + negl %edx + + + C The calculated jump mustn't be before the start of the available + C code. This is the limitation UNROLL_COUNT puts on the src operand + C size, but checked here using the jump address directly. + C + ASSERT(ae,` + movl_text_address( L(unroll_inner_start), %eax) + cmpl %eax, %ecx + ') + + +C ----------------------------------------------------------------------------- + ALIGN(16) +L(unroll_outer_top): + C eax + C ebx &src[size], constant + C ecx VAR_JMP + C edx VAR_COUNTER, limbs, negative + C esi high limb to store + C edi dst ptr, high of last addmul + C ebp + + movl -12+OFFSET(%ebx,%edx,4), %ebp C multiplier + movl %edx, VAR_COUNTER + + movl -8+OFFSET(%ebx,%edx,4), %eax C first limb of multiplicand + + mull %ebp + + testb $1, %cl + + movl %edx, %esi C high carry + movl %ecx, %edx C jump + + movl %eax, %ecx C low carry + leal CODE_BYTES_PER_LIMB(%edx), %edx + + movl %edx, VAR_JMP + leal 4(%edi), %edi + + C A branch-free version of this using some xors was found to be a + C touch slower than just a conditional jump, despite the jump + C switching between taken and not taken on every loop. + +ifelse(eval(UNROLL_COUNT%2),0, + jz,jnz) L(unroll_noswap) + movl %esi, %eax C high,low carry other way around + + movl %ecx, %esi + movl %eax, %ecx +L(unroll_noswap): + + jmp *%edx + + + C Must be on an even address here so the low bit of the jump address + C will indicate which way around ecx/esi should start. + C + C An attempt was made at padding here to get the end of the unrolled + C code to come out on a good alignment, to save padding before + C L(corner). This worked, but turned out to run slower than just an + C ALIGN(2). The reason for this is not clear, it might be related + C to the different speeds on different UNROLL_COUNTs noted above. + + ALIGN(2) + +L(unroll_inner_start): + C eax scratch + C ebx src + C ecx carry low + C edx scratch + C esi carry high + C edi dst + C ebp multiplier + C + C 15 code bytes each limb + C ecx/esi swapped on each chunk + +forloop(`i', UNROLL_COUNT, 1, ` + deflit(`disp_src', eval(-i*4 + OFFSET)) + deflit(`disp_dst', eval(disp_src - 4)) + + m4_assert(`disp_src>=-128 && disp_src<128') + m4_assert(`disp_dst>=-128 && disp_dst<128') + +ifelse(eval(i%2),0,` +Zdisp( movl, disp_src,(%ebx), %eax) + mull %ebp +Zdisp( addl, %esi, disp_dst,(%edi)) + adcl %eax, %ecx + movl %edx, %esi + jadcl0( %esi) +',` + dnl this one comes out last +Zdisp( movl, disp_src,(%ebx), %eax) + mull %ebp +Zdisp( addl, %ecx, disp_dst,(%edi)) + adcl %eax, %esi + movl %edx, %ecx + jadcl0( %ecx) +') +') +L(unroll_inner_end): + + addl %esi, -4+OFFSET(%edi) + + movl VAR_COUNTER, %edx + jadcl0( %ecx) + + movl %ecx, m4_empty_if_zero(OFFSET)(%edi) + movl VAR_JMP, %ecx + + incl %edx + jnz L(unroll_outer_top) + + +ifelse(OFFSET,0,,` + addl $OFFSET, %ebx + addl $OFFSET, %edi +') + + +C ----------------------------------------------------------------------------- + ALIGN(16) +L(corner): + C ebx &src[size] + C edi &dst[2*size-5] + + movl -12(%ebx), %ebp + + movl -8(%ebx), %eax + movl %eax, %ecx + + mull %ebp + + addl %eax, -4(%edi) + adcl $0, %edx + + movl -4(%ebx), %eax + movl %edx, %esi + movl %eax, %ebx + + mull %ebp + + addl %esi, %eax + adcl $0, %edx + + addl %eax, (%edi) + adcl $0, %edx + + movl %edx, %esi + movl %ebx, %eax + + mull %ecx + + addl %esi, %eax + movl %eax, 4(%edi) + + adcl $0, %edx + + movl %edx, 8(%edi) + + +C ----------------------------------------------------------------------------- +C Left shift of dst[1..2*size-2], the bit shifted out becomes dst[2*size-1]. +C The loop measures about 6 cycles/iteration, though it looks like it should +C decode in 5. + +L(lshift_start): + movl PARAM_SIZE, %ecx + + movl PARAM_DST, %edi + subl $1, %ecx C size-1 and clear carry + + movl PARAM_SRC, %ebx + movl %ecx, %edx + + xorl %eax, %eax C ready for adcl + + + ALIGN(16) +L(lshift): + C eax + C ebx src (for later use) + C ecx counter, decrementing + C edx size-1 (for later use) + C esi + C edi dst, incrementing + C ebp + + rcll 4(%edi) + rcll 8(%edi) + leal 8(%edi), %edi + loop L(lshift) + + + adcl %eax, %eax + + movl %eax, 4(%edi) C dst most significant limb + movl (%ebx), %eax C src[0] + + leal 4(%ebx,%edx,4), %ebx C &src[size] + subl %edx, %ecx C -(size-1) + + +C ----------------------------------------------------------------------------- +C Now add in the squares on the diagonal, src[0]^2, src[1]^2, ..., +C src[size-1]^2. dst[0] hasn't yet been set at all yet, and just gets the +C low limb of src[0]^2. + + + mull %eax + + movl %eax, (%edi,%ecx,8) C dst[0] + + + ALIGN(16) +L(diag): + C eax scratch + C ebx &src[size] + C ecx counter, negative + C edx carry + C esi scratch + C edi dst[2*size-2] + C ebp + + movl (%ebx,%ecx,4), %eax + movl %edx, %esi + + mull %eax + + addl %esi, 4(%edi,%ecx,8) + adcl %eax, 8(%edi,%ecx,8) + adcl $0, %edx + + incl %ecx + jnz L(diag) + + + movl SAVE_EBX, %ebx + movl SAVE_ESI, %esi + + addl %edx, 4(%edi) C dst most significant limb + + movl SAVE_EDI, %edi + movl SAVE_EBP, %ebp + addl $FRAME, %esp + ret + + + +C ----------------------------------------------------------------------------- +ifdef(`PIC',` +L(pic_calc): + C See mpn/x86/README about old gas bugs + addl (%esp), %ecx + addl $L(unroll_inner_end)-L(here)-eval(2*CODE_BYTES_PER_LIMB), %ecx + addl %edx, %ecx + ret_internal +') + + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/README b/gmp-6.3.0/mpn/x86/k7/README new file mode 100644 index 0000000..5711b61 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/README @@ -0,0 +1,174 @@ +Copyright 2000, 2001 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. + + + + + AMD K7 MPN SUBROUTINES + + +This directory contains code optimized for the AMD Athlon CPU. + +The mmx subdirectory has routines using MMX instructions. All Athlons have +MMX, the separate directory is just so that configure can omit it if the +assembler doesn't support MMX. + + + +STATUS + +Times for the loops, with all code and data in L1 cache. + + cycles/limb + mpn_add/sub_n 1.6 + + mpn_copyi 0.75 or 1.0 \ varying with data alignment + mpn_copyd 0.75 or 1.0 / + + mpn_divrem_1 17.0 integer part, 15.0 fractional part + mpn_mod_1 17.0 + mpn_divexact_by3 8.0 + + mpn_l/rshift 1.2 + + mpn_mul_1 3.4 + mpn_addmul/submul_1 3.9 + + mpn_mul_basecase 4.42 cycles/crossproduct (approx) + mpn_sqr_basecase 2.3 cycles/crossproduct (approx) + or 4.55 cycles/triangleproduct (approx) + +Prefetching of sources hasn't yet been tried. + + + +NOTES + +cmov, MMX, 3DNow and some extensions to MMX and 3DNow are available. + +Write-allocate L1 data cache means prefetching of destinations is unnecessary. + +Floating point multiplications can be done in parallel with integer +multiplications, but there doesn't seem to be any way to make use of this. + +Unsigned "mul"s can be issued every 3 cycles. This suggests 3 is a limit on +the speed of the multiplication routines. The documentation shows mul +executing in IEU0 (or maybe in IEU0 and IEU1 together), so it might be that, +to get near 3 cycles code has to be arranged so that nothing else is issued +to IEU0. A busy IEU0 could explain why some code takes 4 cycles and other +apparently equivalent code takes 5. + + + +OPTIMIZATIONS + +Unrolled loops are used to reduce looping overhead. The unrolling is +configurable up to 32 limbs/loop for most routines and up to 64 for some. +The K7 has 64k L1 code cache so quite big unrolling is allowable. + +Computed jumps into the unrolling are used to handle sizes not a multiple of +the unrolling. An attractive feature of this is that times increase +smoothly with operand size, but it may be that some routines should just +have simple loops to finish up, especially when PIC adds between 2 and 16 +cycles to get %eip. + +Position independent code is implemented using a call to get %eip for the +computed jumps and a ret is always done, rather than an addl $4,%esp or a +popl, so the CPU return address branch prediction stack stays synchronised +with the actual stack in memory. + +Branch prediction, in absence of any history, will guess forward jumps are +not taken and backward jumps are taken. Where possible it's arranged that +the less likely or less important case is under a taken forward jump. + + + +CODING + +Instructions in general code have been shown grouped if they can execute +together, which means up to three direct-path instructions which have no +successive dependencies. K7 always decodes three and has out-of-order +execution, but the groupings show what slots might be available and what +dependency chains exist. + +When there's vector-path instructions an effort is made to get triplets of +direct-path instructions in between them, even if there's dependencies, +since this maximizes decoding throughput and might save a cycle or two if +decoding is the limiting factor. + + + +INSTRUCTIONS + +adcl direct +divl 39 cycles back-to-back +lodsl,etc vector +loop 1 cycle vector (decl/jnz opens up one decode slot) +movd reg vector +movd mem direct +mull issue every 3 cycles, latency 4 cycles low word, 6 cycles high word +popl vector (use movl for more than one pop) +pushl direct, will pair with a load +shrdl %cl vector, 3 cycles, seems to be 3 decode too +xorl r,r false read dependency recognised + + + +REFERENCES + +"AMD Athlon Processor X86 Code Optimization Guide", AMD publication number +22007, revision K, February 2002. Available on-line, + +http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22007.pdf + +"3DNow Technology Manual", AMD publication number 21928G/0-March 2000. +This describes the femms and prefetch instructions. Available on-line, + +http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/21928.pdf + +"AMD Extensions to the 3DNow and MMX Instruction Sets Manual", AMD +publication number 22466, revision D, March 2000. This describes +instructions added in the Athlon processor, such as pswapd and the extra +prefetch forms. Available on-line, + +http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22466.pdf + +"3DNow Instruction Porting Guide", AMD publication number 22621, revision B, +August 1999. This has some notes on general Athlon optimizations as well as +3DNow. Available on-line, + +http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22621.pdf + + + + +---------------- +Local variables: +mode: text +fill-column: 76 +End: diff --git a/gmp-6.3.0/mpn/x86/k7/addlsh1_n.asm b/gmp-6.3.0/mpn/x86/k7/addlsh1_n.asm new file mode 100644 index 0000000..2cba1eb --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/addlsh1_n.asm @@ -0,0 +1,196 @@ +dnl AMD K7 mpn_addlsh1_n -- rp[] = up[] + (vp[] << 1) + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl Contributed to the GNU project by Torbjorn Granlund and Marco Bodrato. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C This is an attempt at an addlsh1_n for x86-32, not relying on sse2 insns. +C The innerloop is 2*3-way unrolled, which is best we can do with the available +C registers. It seems tricky to use the same structure for rsblsh1_n, since we +C cannot feed carry between operations there. + +C cycles/limb +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) 5.4 (worse than add_n + lshift) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 6 +C AMD K6 ? +C AMD K7 2.5 +C AMD K8 + +C This is a basic addlsh1_n for k7, atom, and perhaps some other x86-32 +C processors. It uses 2*3-way unrolling, for good reasons. Unfortunately, +C that means we need an initial magic multiply. +C +C It is not clear how to do sublsh1_n or rsblsh1_n using the same pattern. We +C cannot do rsblsh1_n since we feed carry from the shift blocks to the +C add/subtract blocks, which is right for addition but reversed for +C subtraction. We could perhaps do sublsh1_n, with some extra move insns, +C without losing any time, since we're not issue limited but carry recurrency +C latency. +C +C Breaking carry recurrency might be a good idea. We would then need separate +C registers for the shift carry and add/subtract carry, which in turn would +C force us to 2*2-way unrolling. + +defframe(PARAM_SIZE, 16) +defframe(PARAM_DBLD, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(VAR_COUNT,`PARAM_DST') +define(VAR_TMP,`PARAM_DBLD') + +ASM_START() + TEXT + ALIGN(8) +PROLOGUE(mpn_addlsh1_n) +deflit(`FRAME',0) + +define(`rp', `%edi') +define(`up', `%esi') +define(`vp', `%ebp') + + mov $0x2aaaaaab, %eax + + push %ebx FRAME_pushl() + mov PARAM_SIZE, %ebx C size + + push rp FRAME_pushl() + mov PARAM_DST, rp + + mul %ebx + + push up FRAME_pushl() + mov PARAM_SRC, up + + not %edx C count = -(size\8)-1 + mov %edx, VAR_COUNT + + push vp FRAME_pushl() + mov PARAM_DBLD, vp + + lea 3(%edx,%edx,2), %ecx C count*3+3 = -(size\6)*3 + xor %edx, %edx + lea (%ebx,%ecx,2), %ebx C size + (count*3+3)*2 = size % 6 + or %ebx, %ebx + jz L(exact) + +L(oop): +ifdef(`CPU_P6',` + shr %edx ') C restore 2nd saved carry bit + mov (vp), %eax + adc %eax, %eax + rcr %edx C restore 1st saved carry bit + lea 4(vp), vp + adc (up), %eax + lea 4(up), up + adc %edx, %edx C save a carry bit in edx +ifdef(`CPU_P6',` + adc %edx, %edx ') C save another carry bit in edx + dec %ebx + mov %eax, (rp) + lea 4(rp), rp + jnz L(oop) + mov vp, VAR_TMP +L(exact): + incl VAR_COUNT + jz L(end) + + ALIGN(16) +L(top): +ifdef(`CPU_P6',` + shr %edx ') C restore 2nd saved carry bit + mov (vp), %eax + adc %eax, %eax + mov 4(vp), %ebx + adc %ebx, %ebx + mov 8(vp), %ecx + adc %ecx, %ecx + + rcr %edx C restore 1st saved carry bit + + adc (up), %eax + mov %eax, (rp) + adc 4(up), %ebx + mov %ebx, 4(rp) + adc 8(up), %ecx + mov %ecx, 8(rp) + + mov 12(vp), %eax + adc %eax, %eax + mov 16(vp), %ebx + adc %ebx, %ebx + mov 20(vp), %ecx + adc %ecx, %ecx + + lea 24(vp), vp + adc %edx, %edx C save a carry bit in edx + + adc 12(up), %eax + mov %eax, 12(rp) + adc 16(up), %ebx + mov %ebx, 16(rp) + adc 20(up), %ecx + + lea 24(up), up + +ifdef(`CPU_P6',` + adc %edx, %edx ') C save another carry bit in edx + mov %ecx, 20(rp) + incl VAR_COUNT + lea 24(rp), rp + jne L(top) + +L(end): + pop vp FRAME_popl() + pop up FRAME_popl() + +ifdef(`CPU_P6',` + xor %eax, %eax + shr $1, %edx + adc %edx, %eax +',` + adc $0, %edx + mov %edx, %eax +') + pop rp FRAME_popl() + pop %ebx FRAME_popl() + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/k7/aors_n.asm b/gmp-6.3.0/mpn/x86/k7/aors_n.asm new file mode 100644 index 0000000..1a08072 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/aors_n.asm @@ -0,0 +1,258 @@ +dnl AMD K7 mpn_add_n/mpn_sub_n -- mpn add or subtract. + +dnl Copyright 1999-2003 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K7: 1.64 cycles/limb (at 16 limbs/loop). + + + +dnl K7: UNROLL_COUNT cycles/limb +dnl 8 1.9 +dnl 16 1.64 +dnl 32 1.7 +dnl 64 2.0 +dnl Maximum possible with the current code is 64. + +deflit(UNROLL_COUNT, 16) + + +ifdef(`OPERATION_add_n', ` + define(M4_inst, adcl) + define(M4_function_n, mpn_add_n) + define(M4_function_nc, mpn_add_nc) + define(M4_description, add) +',`ifdef(`OPERATION_sub_n', ` + define(M4_inst, sbbl) + define(M4_function_n, mpn_sub_n) + define(M4_function_nc, mpn_sub_nc) + define(M4_description, subtract) +',`m4_error(`Need OPERATION_add_n or OPERATION_sub_n +')')') + +MULFUNC_PROLOGUE(mpn_add_n mpn_add_nc mpn_sub_n mpn_sub_nc) + + +C mp_limb_t M4_function_n (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size); +C mp_limb_t M4_function_nc (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size, mp_limb_t carry); +C +C Calculate src1,size M4_description src2,size, and store the result in +C dst,size. The return value is the carry bit from the top of the result (1 +C or 0). +C +C The _nc version accepts 1 or 0 for an initial carry into the low limb of +C the calculation. Note values other than 1 or 0 here will lead to garbage +C results. +C +C This code runs at 1.64 cycles/limb, which might be the best possible with +C plain integer operations. Each limb is 2 loads and 1 store, any 2 of +C which can be done each cycle, leading to 1.5 c/l. + +dnl Must have UNROLL_THRESHOLD >= 2, since the unrolled loop can't handle 1. +ifdef(`PIC',` +deflit(UNROLL_THRESHOLD, 8) +',` +deflit(UNROLL_THRESHOLD, 8) +') + +defframe(PARAM_CARRY,20) +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC2, 12) +defframe(PARAM_SRC1, 8) +defframe(PARAM_DST, 4) + +defframe(SAVE_EBP, -4) +defframe(SAVE_ESI, -8) +defframe(SAVE_EBX, -12) +defframe(SAVE_EDI, -16) +deflit(STACK_SPACE, 16) + + TEXT + ALIGN(32) +deflit(`FRAME',0) + +PROLOGUE(M4_function_nc) + movl PARAM_CARRY, %eax + jmp L(start) +EPILOGUE() + +PROLOGUE(M4_function_n) + + xorl %eax, %eax C carry +L(start): + movl PARAM_SIZE, %ecx + subl $STACK_SPACE, %esp +deflit(`FRAME',STACK_SPACE) + + movl %edi, SAVE_EDI + movl %ebx, SAVE_EBX + cmpl $UNROLL_THRESHOLD, %ecx + + movl PARAM_SRC2, %edx + movl PARAM_SRC1, %ebx + jae L(unroll) + + movl PARAM_DST, %edi + leal (%ebx,%ecx,4), %ebx + leal (%edx,%ecx,4), %edx + + leal (%edi,%ecx,4), %edi + negl %ecx + shrl %eax + + C This loop in in a single 16 byte code block already, so no + C alignment necessary. +L(simple): + C eax scratch + C ebx src1 + C ecx counter + C edx src2 + C esi + C edi dst + C ebp + + movl (%ebx,%ecx,4), %eax + M4_inst (%edx,%ecx,4), %eax + movl %eax, (%edi,%ecx,4) + incl %ecx + jnz L(simple) + + movl $0, %eax + movl SAVE_EDI, %edi + + movl SAVE_EBX, %ebx + setc %al + addl $STACK_SPACE, %esp + + ret + + +C ----------------------------------------------------------------------------- + C This is at 0x55, close enough to aligned. +L(unroll): +deflit(`FRAME',STACK_SPACE) + movl %ebp, SAVE_EBP + andl $-2, %ecx C size low bit masked out + andl $1, PARAM_SIZE C size low bit kept + + movl %ecx, %edi + decl %ecx + movl PARAM_DST, %ebp + + shrl $UNROLL_LOG2, %ecx + negl %edi + movl %esi, SAVE_ESI + + andl $UNROLL_MASK, %edi + +ifdef(`PIC',` + call L(pic_calc) +L(here): +',` + leal L(entry) (%edi,%edi,8), %esi C 9 bytes per +') + negl %edi + shrl %eax + + leal ifelse(UNROLL_BYTES,256,128) (%ebx,%edi,4), %ebx + leal ifelse(UNROLL_BYTES,256,128) (%edx,%edi,4), %edx + leal ifelse(UNROLL_BYTES,256,128) (%ebp,%edi,4), %edi + + jmp *%esi + + +ifdef(`PIC',` +L(pic_calc): + C See mpn/x86/README about old gas bugs + leal (%edi,%edi,8), %esi + addl $L(entry)-L(here), %esi + addl (%esp), %esi + ret_internal +') + + +C ----------------------------------------------------------------------------- + ALIGN(32) +L(top): + C eax zero + C ebx src1 + C ecx counter + C edx src2 + C esi scratch (was computed jump) + C edi dst + C ebp scratch + + leal UNROLL_BYTES(%edx), %edx + +L(entry): +deflit(CHUNK_COUNT, 2) +forloop(i, 0, UNROLL_COUNT/CHUNK_COUNT-1, ` + deflit(`disp0', eval(i*CHUNK_COUNT*4 ifelse(UNROLL_BYTES,256,-128))) + deflit(`disp1', eval(disp0 + 4)) + +Zdisp( movl, disp0,(%ebx), %esi) + movl disp1(%ebx), %ebp +Zdisp( M4_inst,disp0,(%edx), %esi) +Zdisp( movl, %esi, disp0,(%edi)) + M4_inst disp1(%edx), %ebp + movl %ebp, disp1(%edi) +') + + decl %ecx + leal UNROLL_BYTES(%ebx), %ebx + leal UNROLL_BYTES(%edi), %edi + jns L(top) + + + mov PARAM_SIZE, %esi + movl SAVE_EBP, %ebp + movl $0, %eax + + decl %esi + js L(even) + + movl (%ebx), %ecx + M4_inst UNROLL_BYTES(%edx), %ecx + movl %ecx, (%edi) +L(even): + + movl SAVE_EDI, %edi + movl SAVE_EBX, %ebx + setc %al + + movl SAVE_ESI, %esi + addl $STACK_SPACE, %esp + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/aorsmul_1.asm b/gmp-6.3.0/mpn/x86/k7/aorsmul_1.asm new file mode 100644 index 0000000..eec8df6 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/aorsmul_1.asm @@ -0,0 +1,167 @@ +dnl AMD K7 mpn_addmul_1/mpn_submul_1 -- add or subtract mpn multiple. + +dnl Copyright 1999-2002, 2005, 2008 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) 6.5 +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C AMD K6 +C AMD K7 3.75 +C AMD K8 + +C TODO +C * Improve feed-in and wind-down code. We beat the old code for all n != 1, +C but lose by 2x for n == 1. + +ifdef(`OPERATION_addmul_1',` + define(`ADDSUB', `add') + define(`func', `mpn_addmul_1') +') +ifdef(`OPERATION_submul_1',` + define(`ADDSUB', `sub') + define(`func', `mpn_submul_1') +') + +MULFUNC_PROLOGUE(mpn_addmul_1 mpn_submul_1) + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(func) + add $-16, %esp + mov %ebp, (%esp) + mov %ebx, 4(%esp) + mov %esi, 8(%esp) + mov %edi, 12(%esp) + + mov 20(%esp), %edi + mov 24(%esp), %esi + mov 28(%esp), %eax + mov 32(%esp), %ecx + mov %eax, %ebx + shr $2, %eax + mov %eax, 28(%esp) + mov (%esi), %eax + and $3, %ebx + jz L(b0) + cmp $2, %ebx + jz L(b2) + jg L(b3) + +L(b1): lea -4(%esi), %esi + lea -4(%edi), %edi + mul %ecx + mov %eax, %ebx + mov %edx, %ebp + cmpl $0, 28(%esp) + jz L(cj1) + mov 8(%esi), %eax + jmp L(1) + +L(b2): mul %ecx + mov %eax, %ebp + mov 4(%esi), %eax + mov %edx, %ebx + cmpl $0, 28(%esp) + jne L(2) + jmp L(cj2) + +L(b3): lea -12(%esi), %esi + lea -12(%edi), %edi + mul %ecx + mov %eax, %ebx + mov %edx, %ebp + mov 16(%esi), %eax + incl 28(%esp) + jmp L(3) + +L(b0): lea -8(%esi), %esi + lea -8(%edi), %edi + mul %ecx + mov %eax, %ebp + mov 12(%esi), %eax + mov %edx, %ebx + jmp L(0) + + ALIGN(16) +L(top): lea 16(%edi), %edi +L(2): mul %ecx + ADDSUB %ebp, 0(%edi) + mov $0, %ebp + adc %eax, %ebx + mov 8(%esi), %eax + adc %edx, %ebp +L(1): mul %ecx + ADDSUB %ebx, 4(%edi) + mov $0, %ebx + adc %eax, %ebp + mov 12(%esi), %eax + adc %edx, %ebx +L(0): mul %ecx + ADDSUB %ebp, 8(%edi) + mov $0, %ebp + adc %eax, %ebx + adc %edx, %ebp + mov 16(%esi), %eax +L(3): mul %ecx + ADDSUB %ebx, 12(%edi) + adc %eax, %ebp + mov 20(%esi), %eax + lea 16(%esi), %esi + mov $0, %ebx + adc %edx, %ebx + decl 28(%esp) + jnz L(top) + +L(end): lea 16(%edi), %edi +L(cj2): mul %ecx + ADDSUB %ebp, (%edi) + adc %eax, %ebx + adc $0, %edx +L(cj1): ADDSUB %ebx, 4(%edi) + adc $0, %edx + mov %edx, %eax + mov (%esp), %ebp + mov 4(%esp), %ebx + mov 8(%esp), %esi + mov 12(%esp), %edi + add $16, %esp + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/k7/bdiv_q_1.asm b/gmp-6.3.0/mpn/x86/k7/bdiv_q_1.asm new file mode 100644 index 0000000..2af7bb9 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/bdiv_q_1.asm @@ -0,0 +1,245 @@ +dnl AMD K7 mpn_bdiv_q_1 -- mpn by limb exact division. + +dnl Rearranged from mpn/x86/k7/dive_1.asm by Marco Bodrato. + +dnl Copyright 2001, 2002, 2004, 2007, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C Athlon: 11.0 +C Hammer: 9.0 + + +C void mpn_divexact_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C +C The dependent chain is mul+imul+sub for 11 cycles and that speed is +C achieved with no special effort. The load and shrld latencies are hidden +C by out of order execution. +C +C It's a touch faster on size==1 to use the mul-by-inverse than divl. + +defframe(PARAM_SHIFT, 24) +defframe(PARAM_INVERSE,20) +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +defframe(SAVE_EBX, -4) +defframe(SAVE_ESI, -8) +defframe(SAVE_EDI, -12) +defframe(SAVE_EBP, -16) +defframe(VAR_INVERSE, -20) +defframe(VAR_DST_END, -24) + +deflit(STACK_SPACE, 24) + + TEXT + +C mp_limb_t +C mpn_pi1_bdiv_q_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, mp_limb_t divisor, +C mp_limb_t inverse, int shift) + ALIGN(16) +PROLOGUE(mpn_pi1_bdiv_q_1) +deflit(`FRAME',0) + + subl $STACK_SPACE, %esp deflit(`FRAME',STACK_SPACE) + movl PARAM_SHIFT, %ecx C shift count + + movl %ebp, SAVE_EBP + movl PARAM_SIZE, %ebp + + movl %esi, SAVE_ESI + movl PARAM_SRC, %esi + + movl %edi, SAVE_EDI + movl PARAM_DST, %edi + + movl %ebx, SAVE_EBX + + leal (%esi,%ebp,4), %esi C src end + leal (%edi,%ebp,4), %edi C dst end + negl %ebp C -size + + movl PARAM_INVERSE, %eax C inv + +L(common): + movl %eax, VAR_INVERSE + movl (%esi,%ebp,4), %eax C src[0] + + incl %ebp + jz L(one) + + movl (%esi,%ebp,4), %edx C src[1] + + shrdl( %cl, %edx, %eax) + + movl %edi, VAR_DST_END + xorl %ebx, %ebx + jmp L(entry) + + ALIGN(8) +L(top): + C eax q + C ebx carry bit, 0 or 1 + C ecx shift + C edx + C esi src end + C edi dst end + C ebp counter, limbs, negative + + mull PARAM_DIVISOR C carry limb in edx + + movl -4(%esi,%ebp,4), %eax + movl (%esi,%ebp,4), %edi + + shrdl( %cl, %edi, %eax) + + subl %ebx, %eax C apply carry bit + setc %bl + movl VAR_DST_END, %edi + + subl %edx, %eax C apply carry limb + adcl $0, %ebx + +L(entry): + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi,%ebp,4) + incl %ebp + jnz L(top) + + + mull PARAM_DIVISOR C carry limb in edx + + movl -4(%esi), %eax C src high limb + shrl %cl, %eax + movl SAVE_ESI, %esi + + subl %ebx, %eax C apply carry bit + movl SAVE_EBX, %ebx + movl SAVE_EBP, %ebp + + subl %edx, %eax C apply carry limb + + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi) + movl SAVE_EDI, %edi + addl $STACK_SPACE, %esp + + ret + +L(one): + shrl %cl, %eax + movl SAVE_ESI, %esi + movl SAVE_EBX, %ebx + + imull VAR_INVERSE, %eax + + movl SAVE_EBP, %ebp + + movl %eax, -4(%edi) + movl SAVE_EDI, %edi + addl $STACK_SPACE, %esp + + ret +EPILOGUE() + +C mp_limb_t mpn_bdiv_q_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C + + ALIGN(16) +PROLOGUE(mpn_bdiv_q_1) +deflit(`FRAME',0) + + movl PARAM_DIVISOR, %eax + subl $STACK_SPACE, %esp deflit(`FRAME',STACK_SPACE) + movl $-1, %ecx C shift count + + movl %ebp, SAVE_EBP + movl PARAM_SIZE, %ebp + + movl %esi, SAVE_ESI + movl %edi, SAVE_EDI + + C If there's usually only one or two trailing zero bits then this + C should be faster than bsfl. +L(strip_twos): + incl %ecx + shrl %eax + jnc L(strip_twos) + + movl %ebx, SAVE_EBX + leal 1(%eax,%eax), %ebx C d without twos + andl $127, %eax C d/2, 7 bits + +ifdef(`PIC',` + LEA( binvert_limb_table, %edx) + movzbl (%eax,%edx), %eax C inv 8 bits +',` + movzbl binvert_limb_table(%eax), %eax C inv 8 bits +') + + leal (%eax,%eax), %edx C 2*inv + movl %ebx, PARAM_DIVISOR C d without twos + + imull %eax, %eax C inv*inv + + movl PARAM_SRC, %esi + movl PARAM_DST, %edi + + imull %ebx, %eax C inv*inv*d + + subl %eax, %edx C inv = 2*inv - inv*inv*d + leal (%edx,%edx), %eax C 2*inv + + imull %edx, %edx C inv*inv + + leal (%esi,%ebp,4), %esi C src end + leal (%edi,%ebp,4), %edi C dst end + negl %ebp C -size + + imull %ebx, %edx C inv*inv*d + + subl %edx, %eax C inv = 2*inv - inv*inv*d + + ASSERT(e,` C expect d*inv == 1 mod 2^GMP_LIMB_BITS + pushl %eax FRAME_pushl() + imull PARAM_DIVISOR, %eax + cmpl $1, %eax + popl %eax FRAME_popl()') + + jmp L(common) +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/k7/dive_1.asm b/gmp-6.3.0/mpn/x86/k7/dive_1.asm new file mode 100644 index 0000000..458bd02 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/dive_1.asm @@ -0,0 +1,208 @@ +dnl AMD K7 mpn_divexact_1 -- mpn by limb exact division. + +dnl Copyright 2001, 2002, 2004, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C Athlon: 11.0 +C Hammer: 9.0 + + +C void mpn_divexact_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C +C The dependent chain is mul+imul+sub for 11 cycles and that speed is +C achieved with no special effort. The load and shrld latencies are hidden +C by out of order execution. +C +C It's a touch faster on size==1 to use the mul-by-inverse than divl. + +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +defframe(SAVE_EBX, -4) +defframe(SAVE_ESI, -8) +defframe(SAVE_EDI, -12) +defframe(SAVE_EBP, -16) +defframe(VAR_INVERSE, -20) +defframe(VAR_DST_END, -24) + +deflit(STACK_SPACE, 24) + + TEXT + + ALIGN(16) +PROLOGUE(mpn_divexact_1) +deflit(`FRAME',0) + + movl PARAM_DIVISOR, %eax + subl $STACK_SPACE, %esp deflit(`FRAME',STACK_SPACE) + movl $-1, %ecx C shift count + + movl %ebp, SAVE_EBP + movl PARAM_SIZE, %ebp + + movl %esi, SAVE_ESI + movl %edi, SAVE_EDI + + C If there's usually only one or two trailing zero bits then this + C should be faster than bsfl. +L(strip_twos): + incl %ecx + shrl %eax + jnc L(strip_twos) + + movl %ebx, SAVE_EBX + leal 1(%eax,%eax), %ebx C d without twos + andl $127, %eax C d/2, 7 bits + +ifdef(`PIC',` + LEA( binvert_limb_table, %edx) + movzbl (%eax,%edx), %eax C inv 8 bits +',` + movzbl binvert_limb_table(%eax), %eax C inv 8 bits +') + + leal (%eax,%eax), %edx C 2*inv + movl %ebx, PARAM_DIVISOR C d without twos + + imull %eax, %eax C inv*inv + + movl PARAM_SRC, %esi + movl PARAM_DST, %edi + + imull %ebx, %eax C inv*inv*d + + subl %eax, %edx C inv = 2*inv - inv*inv*d + leal (%edx,%edx), %eax C 2*inv + + imull %edx, %edx C inv*inv + + leal (%esi,%ebp,4), %esi C src end + leal (%edi,%ebp,4), %edi C dst end + negl %ebp C -size + + imull %ebx, %edx C inv*inv*d + + subl %edx, %eax C inv = 2*inv - inv*inv*d + + ASSERT(e,` C expect d*inv == 1 mod 2^GMP_LIMB_BITS + pushl %eax FRAME_pushl() + imull PARAM_DIVISOR, %eax + cmpl $1, %eax + popl %eax FRAME_popl()') + + movl %eax, VAR_INVERSE + movl (%esi,%ebp,4), %eax C src[0] + + incl %ebp + jz L(one) + + movl (%esi,%ebp,4), %edx C src[1] + + shrdl( %cl, %edx, %eax) + + movl %edi, VAR_DST_END + xorl %ebx, %ebx + jmp L(entry) + + ALIGN(8) +L(top): + C eax q + C ebx carry bit, 0 or 1 + C ecx shift + C edx + C esi src end + C edi dst end + C ebp counter, limbs, negative + + mull PARAM_DIVISOR C carry limb in edx + + movl -4(%esi,%ebp,4), %eax + movl (%esi,%ebp,4), %edi + + shrdl( %cl, %edi, %eax) + + subl %ebx, %eax C apply carry bit + setc %bl + movl VAR_DST_END, %edi + + subl %edx, %eax C apply carry limb + adcl $0, %ebx + +L(entry): + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi,%ebp,4) + incl %ebp + jnz L(top) + + + mull PARAM_DIVISOR C carry limb in edx + + movl -4(%esi), %eax C src high limb + shrl %cl, %eax + movl SAVE_ESI, %esi + + subl %ebx, %eax C apply carry bit + movl SAVE_EBX, %ebx + movl SAVE_EBP, %ebp + + subl %edx, %eax C apply carry limb + + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi) + movl SAVE_EDI, %edi + addl $STACK_SPACE, %esp + + ret + + +L(one): + shrl %cl, %eax + movl SAVE_ESI, %esi + movl SAVE_EBX, %ebx + + imull VAR_INVERSE, %eax + + movl SAVE_EBP, %ebp + movl %eax, -4(%edi) + + movl SAVE_EDI, %edi + addl $STACK_SPACE, %esp + + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/k7/gcd_11.asm b/gmp-6.3.0/mpn/x86/k7/gcd_11.asm new file mode 100644 index 0000000..2648dfd --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/gcd_11.asm @@ -0,0 +1,107 @@ +dnl x86 mpn_gcd_11 optimised for AMD K7. + +dnl Contributed to the GNU project by by Kevin Ryde. Rehacked by Torbjorn +dnl Granlund. + +dnl Copyright 2000-2002, 2005, 2009, 2011, 2012, 2014, 2015 Free Software +dnl Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/bit (approx) +C AMD K7 5.31 +C AMD K8,K9 5.33 +C AMD K10 5.30 +C AMD bd1 ? +C AMD bobcat 7.02 +C Intel P4-2 10.1 +C Intel P4-3/4 10.0 +C Intel P6/13 5.88 +C Intel core2 6.26 +C Intel NHM 6.83 +C Intel SBR 8.50 +C Intel atom 8.90 +C VIA nano ? +C Numbers measured with: speed -CD -s16-32 -t16 mpn_gcd_1 + + +C ctz_table[n] is the number of trailing zeros on n, or MAXSHIFT if n==0. + +deflit(MAXSHIFT, 6) +deflit(MASK, eval((m4_lshift(1,MAXSHIFT))-1)) + +DEF_OBJECT(ctz_table,64) + .byte MAXSHIFT +forloop(i,1,MASK, +` .byte m4_count_trailing_zeros(i) +') +END_OBJECT(ctz_table) + + +define(`u0', `%eax') +define(`v0', `%edx') + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(mpn_gcd_11) + push %edi + push %esi + + mov 12(%esp), %eax + mov 16(%esp), %edx + + LEAL( ctz_table, %esi) + jmp L(odd) + + ALIGN(16) C +L(top): cmovc( %ecx, %eax) C u = |v - u| + cmovc( %edi, %edx) C v = min(u,v) +L(mid): and $MASK, %ecx C + movzbl (%esi,%ecx), %ecx C + jz L(shift_alot) C + shr %cl, %eax C +L(odd): mov %eax, %edi C + mov %edx, %ecx C + sub %eax, %ecx C + sub %edx, %eax C + jnz L(top) C + +L(end): mov %edx, %eax + pop %esi + pop %edi + ret + +L(shift_alot): + shr $MAXSHIFT, %eax + mov %eax, %ecx + jmp L(mid) +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/k7/gmp-mparam.h b/gmp-6.3.0/mpn/x86/k7/gmp-mparam.h new file mode 100644 index 0000000..a09507d --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/gmp-mparam.h @@ -0,0 +1,263 @@ +/* AMD K7 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 2083 MHz K7 Barton */ +/* FFT tuning limit = 49,770,069 */ +/* Generated by tuneup.c, 2019-11-09, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 0 /* always */ +#define MOD_1_UNNORM_THRESHOLD 3 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 8 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 4 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 24 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 13 +#define USE_PREINV_DIVREM_1 1 /* native */ +/* From mati.gmplib.org, 2023-07-21 */ +#define DIV_QR_1N_PI1_METHOD 3 /* 9.52% faster than 1 */ +#define DIV_QR_1_NORM_THRESHOLD 4 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 26 + +#define DIV_1_VS_MUL_1_PERCENT 182 + +#define MUL_TOOM22_THRESHOLD 28 +#define MUL_TOOM33_THRESHOLD 85 +#define MUL_TOOM44_THRESHOLD 154 +#define MUL_TOOM6H_THRESHOLD 208 +#define MUL_TOOM8H_THRESHOLD 309 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 97 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 99 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 97 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 102 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 121 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 50 +#define SQR_TOOM3_THRESHOLD 86 +#define SQR_TOOM4_THRESHOLD 220 +#define SQR_TOOM6_THRESHOLD 270 +#define SQR_TOOM8_THRESHOLD 446 + +#define MULMID_TOOM42_THRESHOLD 50 + +#define MULMOD_BNM1_THRESHOLD 18 +#define SQRMOD_BNM1_THRESHOLD 19 + +#define MUL_FFT_MODF_THRESHOLD 606 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 606, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 15, 5}, { 31, 6}, { 28, 7}, { 15, 6}, \ + { 32, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 23, 6}, { 47, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 49, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 55, 9}, { 31, 8}, \ + { 63, 7}, { 127, 8}, { 71, 9}, { 39, 6}, \ + { 319, 9}, { 47, 8}, { 99, 6}, { 399, 9}, \ + { 55,10}, { 31, 9}, { 63, 8}, { 127, 9}, \ + { 79,10}, { 47, 9}, { 95, 8}, { 191, 4}, \ + { 3135, 5}, { 1599, 4}, { 3455, 6}, { 959, 8}, \ + { 247,10}, { 79, 9}, { 167,10}, { 95, 9}, \ + { 199,10}, { 111,11}, { 63,10}, { 127, 9}, \ + { 255,10}, { 143, 9}, { 287, 8}, { 575,10}, \ + { 159, 9}, { 319, 8}, { 639, 7}, { 1279,11}, \ + { 95,10}, { 191, 9}, { 383, 8}, { 799,10}, \ + { 207,12}, { 63,11}, { 127,10}, { 255, 9}, \ + { 511, 8}, { 1023,10}, { 271, 9}, { 543, 8}, \ + { 1087, 9}, { 575,11}, { 159, 9}, { 639,10}, \ + { 335, 9}, { 671, 8}, { 1343,10}, { 351, 9}, \ + { 703,11}, { 191,10}, { 383, 9}, { 799, 8}, \ + { 1599,11}, { 223,10}, { 447,12}, { 127,11}, \ + { 255,10}, { 511, 9}, { 1023,10}, { 543, 9}, \ + { 1087,10}, { 575, 9}, { 1151,10}, { 607, 9}, \ + { 1215,11}, { 319,10}, { 639, 9}, { 1343,10}, \ + { 703, 9}, { 1407,12}, { 191,11}, { 383,10}, \ + { 767, 9}, { 1535,10}, { 799, 9}, { 1599,10}, \ + { 831, 9}, { 1727, 8}, { 3455,11}, { 447,13}, \ + { 127,12}, { 255,11}, { 511,10}, { 1023, 9}, \ + { 2047,11}, { 543,10}, { 1087,11}, { 575,10}, \ + { 1151, 9}, { 2303,11}, { 607,10}, { 1215,12}, \ + { 319,11}, { 639,10}, { 1279,11}, { 671,10}, \ + { 1343,11}, { 703,10}, { 1407,11}, { 735,10}, \ + { 1471, 9}, { 2943,12}, { 383,11}, { 767,10}, \ + { 1535,11}, { 799,10}, { 1599,11}, { 831,10}, \ + { 1663,11}, { 863,10}, { 1727,12}, { 447,11}, \ + { 895,10}, { 1791,11}, { 959,10}, { 1919,13}, \ + { 255,12}, { 511,11}, { 1023,10}, { 2111,11}, \ + { 1087,10}, { 2175,12}, { 575,11}, { 1151,10}, \ + { 2303,11}, { 1215,10}, { 2431,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1407,10}, { 2815,11}, \ + { 1471,10}, { 2943,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1663,10}, { 3327,11}, \ + { 1727,10}, { 3455,12}, { 895,11}, { 1855,12}, \ + { 959,11}, { 1919,10}, { 3839,14}, { 255,13}, \ + { 511,12}, { 1023,11}, { 2111,12}, { 1087,11}, \ + { 2239,12}, { 1151,11}, { 2303,12}, { 1215,11}, \ + { 2431,13}, { 639,12}, { 1343,11}, { 2687,12}, \ + { 1407,11}, { 2815,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1663,11}, { 3327,12}, { 1727,11}, \ + { 3455,13}, { 895,12}, { 1919,11}, { 3839,12}, \ + { 1983,14}, { 511,13}, { 1023,12}, { 2239,13}, \ + { 1151,12}, { 2495,13}, { 1279,12}, { 2687,13}, \ + { 1407,12}, { 2943,14}, { 767,13}, { 1535,12}, \ + { 3135,13}, { 1663,12}, { 3455,13}, { 1791,12}, \ + { 3583,13}, { 1919,12}, { 3967,15}, { 511,14}, \ + { 1023,13}, { 2047,12}, { 4095,13}, { 2175,12}, \ + { 4479,13}, { 2431,12}, { 4863,14}, { 1279,13}, \ + { 2559,12}, { 5119,13}, { 2943,12}, { 5887,14}, \ + { 1535,13}, { 3455,14}, { 1791,13}, { 3967,15}, \ + { 1023,14}, { 2047,13}, { 4479,14}, { 2303,13}, \ + { 4991,14}, { 2559,13}, { 5119,14}, { 2815,13}, \ + { 5887,15}, { 32768,16} } +#define MUL_FFT_TABLE3_SIZE 254 +#define MUL_FFT_THRESHOLD 7552 + +#define SQR_FFT_MODF_THRESHOLD 492 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 492, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 28, 7}, { 15, 6}, { 32, 7}, { 17, 6}, \ + { 35, 7}, { 19, 6}, { 39, 7}, { 27, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 47, 8}, { 27, 9}, { 15, 8}, \ + { 31, 7}, { 63, 8}, { 39, 9}, { 23, 8}, \ + { 51, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 79, 9}, { 47, 8}, { 95, 9}, { 55,10}, \ + { 31, 9}, { 79,10}, { 47, 9}, { 103,11}, \ + { 31,10}, { 63, 9}, { 135, 8}, { 271, 9}, \ + { 143,10}, { 79, 9}, { 167,10}, { 95, 9}, \ + { 191, 8}, { 383,10}, { 111,11}, { 63,10}, \ + { 127, 9}, { 255, 8}, { 511,10}, { 143, 9}, \ + { 303,10}, { 159, 9}, { 319, 8}, { 639,11}, \ + { 95,10}, { 191, 9}, { 383, 8}, { 767, 9}, \ + { 399,10}, { 207,12}, { 63,11}, { 127,10}, \ + { 255, 9}, { 511,10}, { 271, 9}, { 543, 8}, \ + { 1087,10}, { 287, 9}, { 575,10}, { 303,11}, \ + { 159,10}, { 319, 9}, { 639,10}, { 335, 9}, \ + { 671, 8}, { 1343, 9}, { 703,11}, { 191,10}, \ + { 383, 9}, { 767, 8}, { 1535,10}, { 399, 9}, \ + { 799, 8}, { 1599, 9}, { 863,11}, { 223,10}, \ + { 447,12}, { 127,11}, { 255,10}, { 511, 9}, \ + { 1087,10}, { 575, 9}, { 1215,10}, { 639, 9}, \ + { 1279,10}, { 671, 9}, { 1343,11}, { 351,10}, \ + { 703, 9}, { 1407,10}, { 735, 9}, { 1471,12}, \ + { 191,11}, { 383,10}, { 767, 9}, { 1535,10}, \ + { 799, 9}, { 1599,11}, { 415,10}, { 831, 9}, \ + { 1663,10}, { 863, 9}, { 1727, 8}, { 3455,11}, \ + { 447,10}, { 895,13}, { 127,12}, { 255,11}, \ + { 511,10}, { 1023, 9}, { 2047,11}, { 543,10}, \ + { 1087, 9}, { 2175,11}, { 575,10}, { 1151, 9}, \ + { 2303,11}, { 607,10}, { 1215, 9}, { 2431,12}, \ + { 319,11}, { 639,10}, { 1279,11}, { 671,10}, \ + { 1343,11}, { 703,10}, { 1407, 9}, { 2815,11}, \ + { 735,10}, { 1471, 9}, { 2943,12}, { 383,11}, \ + { 767,10}, { 1599,11}, { 831,10}, { 1663, 9}, \ + { 3327,10}, { 1727,12}, { 447,11}, { 895,10}, \ + { 1791,11}, { 959,10}, { 1919,13}, { 255,12}, \ + { 511,11}, { 1023,10}, { 2111,11}, { 1087,10}, \ + { 2175,12}, { 575,11}, { 1151,10}, { 2303,11}, \ + { 1215,10}, { 2431,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1407,10}, { 2815,11}, { 1471,10}, \ + { 2943,13}, { 383,12}, { 767,11}, { 1599,12}, \ + { 831,11}, { 1663,10}, { 3327,11}, { 1727,10}, \ + { 3455,12}, { 895,11}, { 1791,12}, { 959,11}, \ + { 1919,10}, { 3839,14}, { 255,13}, { 511,12}, \ + { 1023,11}, { 2111,12}, { 1087,11}, { 2239,12}, \ + { 1151,11}, { 2303,12}, { 1215,11}, { 2431,13}, \ + { 639,12}, { 1343,11}, { 2687,12}, { 1407,11}, \ + { 2815,12}, { 1471,11}, { 2943,13}, { 767,12}, \ + { 1599,11}, { 3199,12}, { 1663,11}, { 3327,12}, \ + { 1727,11}, { 3455,13}, { 895,12}, { 1791,11}, \ + { 3583,12}, { 1919,11}, { 3839,12}, { 1983,14}, \ + { 511,13}, { 1023,12}, { 2239,13}, { 1151,12}, \ + { 2431,13}, { 1279,12}, { 2687,13}, { 1407,12}, \ + { 2943,14}, { 767,13}, { 1535,12}, { 3199,13}, \ + { 1663,12}, { 3455,13}, { 1791,12}, { 3583,13}, \ + { 1919,12}, { 3967,15}, { 511,14}, { 1023,13}, \ + { 2047,12}, { 4095,13}, { 2175,12}, { 4351,13}, \ + { 2431,14}, { 1279,13}, { 2943,12}, { 5887,14}, \ + { 1535,13}, { 3455,14}, { 1791,13}, { 3967,15}, \ + { 1023,14}, { 2047,13}, { 4351,14}, { 2303,13}, \ + { 4991,14}, { 2559,13}, { 5119,14}, { 2815,13}, \ + { 5887,15}, { 32768,16} } +#define SQR_FFT_TABLE3_SIZE 258 +#define SQR_FFT_THRESHOLD 5504 + +#define MULLO_BASECASE_THRESHOLD 3 +#define MULLO_DC_THRESHOLD 34 +#define MULLO_MUL_N_THRESHOLD 14281 +#define SQRLO_BASECASE_THRESHOLD 6 +#define SQRLO_DC_THRESHOLD 137 +#define SQRLO_SQR_THRESHOLD 10821 + +#define DC_DIV_QR_THRESHOLD 45 +#define DC_DIVAPPR_Q_THRESHOLD 206 +#define DC_BDIV_QR_THRESHOLD 39 +#define DC_BDIV_Q_THRESHOLD 144 + +#define INV_MULMOD_BNM1_THRESHOLD 54 +#define INV_NEWTON_THRESHOLD 202 +#define INV_APPR_THRESHOLD 206 + +#define BINV_NEWTON_THRESHOLD 224 +#define REDC_1_TO_REDC_N_THRESHOLD 63 + +#define MU_DIV_QR_THRESHOLD 1442 +#define MU_DIVAPPR_Q_THRESHOLD 1387 +#define MUPI_DIV_QR_THRESHOLD 82 +#define MU_BDIV_QR_THRESHOLD 1308 +#define MU_BDIV_Q_THRESHOLD 1387 + +#define POWM_SEC_TABLE 1,16,102,428,1221 + +#define GET_STR_DC_THRESHOLD 14 +#define GET_STR_PRECOMPUTE_THRESHOLD 28 +#define SET_STR_DC_THRESHOLD 254 +#define SET_STR_PRECOMPUTE_THRESHOLD 890 + +#define FAC_DSC_THRESHOLD 206 +#define FAC_ODD_THRESHOLD 29 + +#define MATRIX22_STRASSEN_THRESHOLD 17 +#define HGCD2_DIV1_METHOD 3 /* 3.84% faster than 4 */ +#define HGCD_THRESHOLD 123 +#define HGCD_APPR_THRESHOLD 151 +#define HGCD_REDUCE_THRESHOLD 3389 +#define GCD_DC_THRESHOLD 435 +#define GCDEXT_DC_THRESHOLD 318 +#define JACOBI_BASE_METHOD 4 /* 8.04% faster than 3 */ + +/* Tuneup completed successfully, took 175382 seconds */ diff --git a/gmp-6.3.0/mpn/x86/k7/invert_limb.asm b/gmp-6.3.0/mpn/x86/k7/invert_limb.asm new file mode 100644 index 0000000..31a867e --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/invert_limb.asm @@ -0,0 +1,194 @@ +dnl x86 mpn_invert_limb + +dnl Contributed to the GNU project by Niels Möller + +dnl Copyright 2009, 2011, 2015 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles (approx) div +C P5 ? +C P6 model 0-8,10-12 ? +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) ? +C P4 model 0 (Willamette) ? +C P4 model 1 (?) ? +C P4 model 2 (Northwood) ? +C P4 model 3 (Prescott) ? +C P4 model 4 (Nocona) ? +C AMD K6 ? +C AMD K7 41 53 +C AMD K8 ? + +C TODO +C * These c/l numbers are for a non-PIC build. Consider falling back to using +C the 'div' instruction for PIC builds. +C * Perhaps use this file--or at least the algorithm--for more machines than k7. + +C Register usage: +C Input D in %edi +C Current approximation is in %eax and/or %ecx +C %ebx and %edx are temporaries +C %esi and %ebp are unused + +defframe(PARAM_DIVISOR,4) + +ASM_START() + +C Make approx_tab global to work around Apple relocation bug. +ifdef(`DARWIN',` + deflit(`approx_tab', MPN(invert_limb_tab)) + GLOBL approx_tab') + + TEXT + ALIGN(16) +PROLOGUE(mpn_invert_limb) +deflit(`FRAME', 0) + mov PARAM_DIVISOR, %eax + C Avoid push/pop on k7. + sub $8, %esp FRAME_subl_esp(8) + mov %ebx, (%esp) + mov %edi, 4(%esp) + + mov %eax, %edi + shr $22, %eax +ifdef(`PIC',` + LEAL( approx_tab, %ebx) + movzwl -1024(%ebx, %eax, 2), %eax +',` + movzwl -1024+approx_tab(%eax, %eax), %eax C %eax = v0 +') + + C v1 = (v0 << 4) - ((v0*v0*d_21) >> 32) - 1 + mov %eax, %ecx + imul %eax, %eax + mov %edi, %ebx + shr $11, %ebx + inc %ebx + mul %ebx + mov %edi, %ebx C Prepare + shr %ebx + sbb %eax, %eax + sub %eax, %ebx C %ebx = d_31, %eax = mask + shl $4, %ecx + dec %ecx + sub %edx, %ecx C %ecx = v1 + + C v_2 = (v1 << 15) + ((v1 *(2^48 - v1 * d31 + (v1 >> 1) & mask)) >> 33) + imul %ecx, %ebx + and %ecx, %eax + shr %eax + sub %ebx, %eax + mul %ecx + mov %edi, %eax C Prepare for next mul + shl $15, %ecx + shr %edx + add %edx, %ecx C %ecx = v2 + + mul %ecx + add %edi, %eax + mov %ecx, %eax + adc %edi, %edx + sub %edx, %eax C %eax = v3 + + mov (%esp), %ebx + mov 4(%esp), %edi + add $8, %esp + + ret + +EPILOGUE() + +DEF_OBJECT(approx_tab,2) + .value 0x7fe1,0x7fa1,0x7f61,0x7f22,0x7ee3,0x7ea4,0x7e65,0x7e27 + .value 0x7de9,0x7dab,0x7d6d,0x7d30,0x7cf3,0x7cb6,0x7c79,0x7c3d + .value 0x7c00,0x7bc4,0x7b89,0x7b4d,0x7b12,0x7ad7,0x7a9c,0x7a61 + .value 0x7a27,0x79ec,0x79b2,0x7979,0x793f,0x7906,0x78cc,0x7894 + .value 0x785b,0x7822,0x77ea,0x77b2,0x777a,0x7742,0x770b,0x76d3 + .value 0x769c,0x7665,0x762f,0x75f8,0x75c2,0x758c,0x7556,0x7520 + .value 0x74ea,0x74b5,0x7480,0x744b,0x7416,0x73e2,0x73ad,0x7379 + .value 0x7345,0x7311,0x72dd,0x72aa,0x7277,0x7243,0x7210,0x71de + .value 0x71ab,0x7179,0x7146,0x7114,0x70e2,0x70b1,0x707f,0x704e + .value 0x701c,0x6feb,0x6fba,0x6f8a,0x6f59,0x6f29,0x6ef9,0x6ec8 + .value 0x6e99,0x6e69,0x6e39,0x6e0a,0x6ddb,0x6dab,0x6d7d,0x6d4e + .value 0x6d1f,0x6cf1,0x6cc2,0x6c94,0x6c66,0x6c38,0x6c0a,0x6bdd + .value 0x6bb0,0x6b82,0x6b55,0x6b28,0x6afb,0x6acf,0x6aa2,0x6a76 + .value 0x6a49,0x6a1d,0x69f1,0x69c6,0x699a,0x696e,0x6943,0x6918 + .value 0x68ed,0x68c2,0x6897,0x686c,0x6842,0x6817,0x67ed,0x67c3 + .value 0x6799,0x676f,0x6745,0x671b,0x66f2,0x66c8,0x669f,0x6676 + .value 0x664d,0x6624,0x65fc,0x65d3,0x65aa,0x6582,0x655a,0x6532 + .value 0x650a,0x64e2,0x64ba,0x6493,0x646b,0x6444,0x641c,0x63f5 + .value 0x63ce,0x63a7,0x6381,0x635a,0x6333,0x630d,0x62e7,0x62c1 + .value 0x629a,0x6275,0x624f,0x6229,0x6203,0x61de,0x61b8,0x6193 + .value 0x616e,0x6149,0x6124,0x60ff,0x60da,0x60b6,0x6091,0x606d + .value 0x6049,0x6024,0x6000,0x5fdc,0x5fb8,0x5f95,0x5f71,0x5f4d + .value 0x5f2a,0x5f07,0x5ee3,0x5ec0,0x5e9d,0x5e7a,0x5e57,0x5e35 + .value 0x5e12,0x5def,0x5dcd,0x5dab,0x5d88,0x5d66,0x5d44,0x5d22 + .value 0x5d00,0x5cde,0x5cbd,0x5c9b,0x5c7a,0x5c58,0x5c37,0x5c16 + .value 0x5bf5,0x5bd4,0x5bb3,0x5b92,0x5b71,0x5b51,0x5b30,0x5b10 + .value 0x5aef,0x5acf,0x5aaf,0x5a8f,0x5a6f,0x5a4f,0x5a2f,0x5a0f + .value 0x59ef,0x59d0,0x59b0,0x5991,0x5972,0x5952,0x5933,0x5914 + .value 0x58f5,0x58d6,0x58b7,0x5899,0x587a,0x585b,0x583d,0x581f + .value 0x5800,0x57e2,0x57c4,0x57a6,0x5788,0x576a,0x574c,0x572e + .value 0x5711,0x56f3,0x56d5,0x56b8,0x569b,0x567d,0x5660,0x5643 + .value 0x5626,0x5609,0x55ec,0x55cf,0x55b2,0x5596,0x5579,0x555d + .value 0x5540,0x5524,0x5507,0x54eb,0x54cf,0x54b3,0x5497,0x547b + .value 0x545f,0x5443,0x5428,0x540c,0x53f0,0x53d5,0x53b9,0x539e + .value 0x5383,0x5368,0x534c,0x5331,0x5316,0x52fb,0x52e0,0x52c6 + .value 0x52ab,0x5290,0x5276,0x525b,0x5240,0x5226,0x520c,0x51f1 + .value 0x51d7,0x51bd,0x51a3,0x5189,0x516f,0x5155,0x513b,0x5121 + .value 0x5108,0x50ee,0x50d5,0x50bb,0x50a2,0x5088,0x506f,0x5056 + .value 0x503c,0x5023,0x500a,0x4ff1,0x4fd8,0x4fbf,0x4fa6,0x4f8e + .value 0x4f75,0x4f5c,0x4f44,0x4f2b,0x4f13,0x4efa,0x4ee2,0x4eca + .value 0x4eb1,0x4e99,0x4e81,0x4e69,0x4e51,0x4e39,0x4e21,0x4e09 + .value 0x4df1,0x4dda,0x4dc2,0x4daa,0x4d93,0x4d7b,0x4d64,0x4d4d + .value 0x4d35,0x4d1e,0x4d07,0x4cf0,0x4cd8,0x4cc1,0x4caa,0x4c93 + .value 0x4c7d,0x4c66,0x4c4f,0x4c38,0x4c21,0x4c0b,0x4bf4,0x4bde + .value 0x4bc7,0x4bb1,0x4b9a,0x4b84,0x4b6e,0x4b58,0x4b41,0x4b2b + .value 0x4b15,0x4aff,0x4ae9,0x4ad3,0x4abd,0x4aa8,0x4a92,0x4a7c + .value 0x4a66,0x4a51,0x4a3b,0x4a26,0x4a10,0x49fb,0x49e5,0x49d0 + .value 0x49bb,0x49a6,0x4990,0x497b,0x4966,0x4951,0x493c,0x4927 + .value 0x4912,0x48fe,0x48e9,0x48d4,0x48bf,0x48ab,0x4896,0x4881 + .value 0x486d,0x4858,0x4844,0x482f,0x481b,0x4807,0x47f3,0x47de + .value 0x47ca,0x47b6,0x47a2,0x478e,0x477a,0x4766,0x4752,0x473e + .value 0x472a,0x4717,0x4703,0x46ef,0x46db,0x46c8,0x46b4,0x46a1 + .value 0x468d,0x467a,0x4666,0x4653,0x4640,0x462c,0x4619,0x4606 + .value 0x45f3,0x45e0,0x45cd,0x45ba,0x45a7,0x4594,0x4581,0x456e + .value 0x455b,0x4548,0x4536,0x4523,0x4510,0x44fe,0x44eb,0x44d8 + .value 0x44c6,0x44b3,0x44a1,0x448f,0x447c,0x446a,0x4458,0x4445 + .value 0x4433,0x4421,0x440f,0x43fd,0x43eb,0x43d9,0x43c7,0x43b5 + .value 0x43a3,0x4391,0x437f,0x436d,0x435c,0x434a,0x4338,0x4327 + .value 0x4315,0x4303,0x42f2,0x42e0,0x42cf,0x42bd,0x42ac,0x429b + .value 0x4289,0x4278,0x4267,0x4256,0x4244,0x4233,0x4222,0x4211 + .value 0x4200,0x41ef,0x41de,0x41cd,0x41bc,0x41ab,0x419a,0x418a + .value 0x4179,0x4168,0x4157,0x4147,0x4136,0x4125,0x4115,0x4104 + .value 0x40f4,0x40e3,0x40d3,0x40c2,0x40b2,0x40a2,0x4091,0x4081 + .value 0x4071,0x4061,0x4050,0x4040,0x4030,0x4020,0x4010,0x4000 +END_OBJECT(approx_tab) +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/k7/mmx/com.asm b/gmp-6.3.0/mpn/x86/k7/mmx/com.asm new file mode 100644 index 0000000..a258c22 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mmx/com.asm @@ -0,0 +1,125 @@ +dnl AMD Athlon mpn_com -- mpn bitwise one's complement. + +dnl Copyright 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K7: 1.0 cycles/limb + + +C void mpn_com (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C The loop form below is necessary for the claimed speed. It needs to be +C aligned to a 16 byte boundary and only 16 bytes long. Maybe that's so it +C fits in a BTB entry. The adjustments to %eax and %edx avoid offsets on +C the movq's and achieve the necessary size. +C +C If both src and dst are 4mod8, the loop runs at 1.5 c/l. So long as one +C of the two is 0mod8, it runs at 1.0 c/l. On that basis dst is checked +C (offset by the size, as per the loop addressing) and one high limb +C processed separately to get alignment. +C +C The padding for the nails case is unattractive, but shouldn't cost any +C cycles. Explicit .byte's guarantee the desired instructions, at a point +C where we're probably stalled waiting for loads anyway. +C +C Enhancements: +C +C The combination load/pxor/store might be able to be unrolled to approach +C 0.5 c/l if desired. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(16) + +PROLOGUE(mpn_com) +deflit(`FRAME',0) + + movl PARAM_DST, %edx + movl PARAM_SIZE, %ecx + pcmpeqd %mm7, %mm7 + + leal (%edx,%ecx,4), %eax + andl $4, %eax +ifelse(GMP_NAIL_BITS,0,, +` psrld $GMP_NAIL_BITS, %mm7') C GMP_NUMB_MASK + + movl PARAM_SRC, %eax + movd -4(%eax,%ecx,4), %mm0 C src high limb + +ifelse(GMP_NAIL_BITS,0,, +` C padding for alignment below + .byte 0x8d, 0xb6, 0x00, 0x00, 0x00, 0x00 C lea 0(%esi),%esi + .byte 0x8d, 0xbf, 0x00, 0x00, 0x00, 0x00 C lea 0(%edi),%edi +') + + jz L(aligned) + + pxor %mm7, %mm0 + movd %mm0, -4(%edx,%ecx,4) C dst high limb + decl %ecx + jz L(done) +L(aligned): + + addl $4, %eax + addl $4, %edx + decl %ecx + jz L(one) + + C offset 0x30 for no nails, or 0x40 for nails + ALIGN(16) +L(top): + C eax src + C ebx + C ecx counter + C edx dst + + subl $2, %ecx + movq (%eax,%ecx,4), %mm0 + pxor %mm7, %mm0 + movq %mm0, (%edx,%ecx,4) + jg L(top) + + jnz L(done) C if size even + +L(one): + movd -4(%eax), %mm0 C src low limb + pxor %mm7, %mm0 + movd %mm0, -4(%edx) C dst low limb + +L(done): + emms + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/mmx/copyd.asm b/gmp-6.3.0/mpn/x86/k7/mmx/copyd.asm new file mode 100644 index 0000000..59ece40 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mmx/copyd.asm @@ -0,0 +1,144 @@ +dnl AMD K7 mpn_copyd -- copy limb vector, decrementing. + +dnl Copyright 1999, 2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C alignment dst/src, A=0mod8 N=4mod8 +C A/A A/N N/A N/N +C K7 0.75 1.0 1.0 0.75 + + +C void mpn_copyd (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C The various comments in mpn/x86/k7/copyi.asm apply here too. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) +deflit(`FRAME',0) + +dnl parameter space reused +define(SAVE_EBX,`PARAM_SIZE') +define(SAVE_ESI,`PARAM_SRC') + +dnl minimum 5 since the unrolled code can't handle less than 5 +deflit(UNROLL_THRESHOLD, 5) + + TEXT + ALIGN(32) +PROLOGUE(mpn_copyd) + + movl PARAM_SIZE, %ecx + movl %ebx, SAVE_EBX + + movl PARAM_SRC, %eax + movl PARAM_DST, %edx + + cmpl $UNROLL_THRESHOLD, %ecx + jae L(unroll) + + orl %ecx, %ecx + jz L(simple_done) + +L(simple): + C eax src + C ebx scratch + C ecx counter + C edx dst + C + C this loop is 2 cycles/limb + + movl -4(%eax,%ecx,4), %ebx + movl %ebx, -4(%edx,%ecx,4) + decl %ecx + jnz L(simple) + +L(simple_done): + movl SAVE_EBX, %ebx + ret + + +L(unroll): + movl %esi, SAVE_ESI + leal (%eax,%ecx,4), %ebx + leal (%edx,%ecx,4), %esi + + andl %esi, %ebx + movl SAVE_ESI, %esi + subl $4, %ecx C size-4 + + testl $4, %ebx C testl to pad code closer to 16 bytes for L(top) + jz L(aligned) + + C both src and dst unaligned, process one limb to align them + movl 12(%eax,%ecx,4), %ebx + movl %ebx, 12(%edx,%ecx,4) + decl %ecx +L(aligned): + + + ALIGN(16) +L(top): + C eax src + C ebx + C ecx counter, limbs + C edx dst + + movq 8(%eax,%ecx,4), %mm0 + movq (%eax,%ecx,4), %mm1 + subl $4, %ecx + movq %mm0, 16+8(%edx,%ecx,4) + movq %mm1, 16(%edx,%ecx,4) + jns L(top) + + + C now %ecx is -4 to -1 representing respectively 0 to 3 limbs remaining + + testb $2, %cl + jz L(finish_not_two) + + movq 8(%eax,%ecx,4), %mm0 + movq %mm0, 8(%edx,%ecx,4) +L(finish_not_two): + + testb $1, %cl + jz L(done) + + movl (%eax), %ebx + movl %ebx, (%edx) + +L(done): + movl SAVE_EBX, %ebx + emms + ret + + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/mmx/copyi.asm b/gmp-6.3.0/mpn/x86/k7/mmx/copyi.asm new file mode 100644 index 0000000..9a28f92 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mmx/copyi.asm @@ -0,0 +1,157 @@ +dnl AMD K7 mpn_copyi -- copy limb vector, incrementing. + +dnl Copyright 1999, 2000, 2002, 2003 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C alignment dst/src, A=0mod8 N=4mod8 +C A/A A/N N/A N/N +C K7 0.75 1.0 1.0 0.75 + + +C void mpn_copyi (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C Copy src,size to dst,size. +C +C This code at 0.75 or 1.0 c/l is always faster than a plain rep movsl at +C 1.33 c/l. +C +C The K7 can do a 64-bit load and 64-bit store in one cycle (optimization +C guile 22007 appendix B), so 0.5 c/l should be possible, however nothing +C under 0.7 c/l is known. Apparently only two 32-bit stores can be done in +C one cycle, so perhaps some scheduling is needed to ensure it's a +C load+store in each cycle, not store+store. +C +C If both source and destination are unaligned then one limb is processed at +C the start to make them aligned and so get 0.75 c/l, whereas if they'd been +C used unaligned it would be 1.5 c/l. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl parameter space reused +define(SAVE_EBX,`PARAM_SIZE') + +dnl minimum 5 since the unrolled code can't handle less than 5 +deflit(UNROLL_THRESHOLD, 5) + + TEXT + ALIGN(32) +PROLOGUE(mpn_copyi) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + movl %ebx, SAVE_EBX + + movl PARAM_SRC, %eax + movl PARAM_DST, %edx + + cmpl $UNROLL_THRESHOLD, %ecx + jae L(unroll) + + orl %ecx, %ecx + jz L(simple_done) + +L(simple): + C eax src, incrementing + C ebx scratch + C ecx counter + C edx dst, incrementing + C + C this loop is 2 cycles/limb + + movl (%eax), %ebx + movl %ebx, (%edx) + decl %ecx + leal 4(%eax), %eax + leal 4(%edx), %edx + jnz L(simple) + +L(simple_done): + movl SAVE_EBX, %ebx + ret + + +L(unroll): + movl %eax, %ebx + leal -12(%eax,%ecx,4), %eax C src end - 12 + subl $3, %ecx C size-3 + + andl %edx, %ebx + leal (%edx,%ecx,4), %edx C dst end - 12 + negl %ecx + + testl $4, %ebx C testl to pad code closer to 16 bytes for L(top) + jz L(aligned) + + C both src and dst unaligned, process one limb to align them + movl (%eax,%ecx,4), %ebx + movl %ebx, (%edx,%ecx,4) + incl %ecx +L(aligned): + + + ALIGN(16) +L(top): + C eax src end - 12 + C ebx + C ecx counter, negative, limbs + C edx dst end - 12 + + movq (%eax,%ecx,4), %mm0 + movq 8(%eax,%ecx,4), %mm1 + addl $4, %ecx + movq %mm0, -16(%edx,%ecx,4) + movq %mm1, -16+8(%edx,%ecx,4) + ja L(top) C jump no carry and not zero + + + C now %ecx is 0 to 3 representing respectively 3 to 0 limbs remaining + + testb $2, %cl + jnz L(finish_not_two) + + movq (%eax,%ecx,4), %mm0 + movq %mm0, (%edx,%ecx,4) +L(finish_not_two): + + testb $1, %cl + jnz L(done) + + movl 8(%eax), %ebx + movl %ebx, 8(%edx) + +L(done): + movl SAVE_EBX, %ebx + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/mmx/divrem_1.asm b/gmp-6.3.0/mpn/x86/k7/mmx/divrem_1.asm new file mode 100644 index 0000000..cf34328 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mmx/divrem_1.asm @@ -0,0 +1,832 @@ +dnl AMD K7 mpn_divrem_1, mpn_divrem_1c, mpn_preinv_divrem_1 -- mpn by limb +dnl division. + +dnl Copyright 1999-2002, 2004 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K7: 17.0 cycles/limb integer part, 15.0 cycles/limb fraction part. + + +C mp_limb_t mpn_divrem_1 (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C mp_limb_t mpn_divrem_1c (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, +C mp_limb_t divisor, mp_limb_t carry); +C mp_limb_t mpn_preinv_divrem_1 (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, +C mp_limb_t divisor, mp_limb_t inverse, +C unsigned shift); +C +C Algorithm: +C +C The method and nomenclature follow part 8 of "Division by Invariant +C Integers using Multiplication" by Granlund and Montgomery, reference in +C gmp.texi. +C +C The "and"s shown in the paper are done here with "cmov"s. "m" is written +C for m', and "d" for d_norm, which won't cause any confusion since it's +C only the normalized divisor that's of any use in the code. "b" is written +C for 2^N, the size of a limb, N being 32 here. +C +C The step "sdword dr = n - 2^N*d + (2^N-1-q1) * d" is instead done as +C "n-(q1+1)*d"; this rearrangement gives the same two-limb answer. If +C q1==0xFFFFFFFF, then q1+1 would overflow. We branch to a special case +C "q1_ff" if this occurs. Since the true quotient is either q1 or q1+1 then +C if q1==0xFFFFFFFF that must be the right value. +C +C For the last and second last steps q1==0xFFFFFFFF is instead handled by an +C sbbl to go back to 0xFFFFFFFF if an overflow occurs when adding 1. This +C then goes through as normal, and finding no addback required. sbbl costs +C an extra cycle over what the main loop code does, but it keeps code size +C and complexity down. +C +C Notes: +C +C mpn_divrem_1 and mpn_preinv_divrem_1 avoid one division if the src high +C limb is less than the divisor. mpn_divrem_1c doesn't check for a zero +C carry, since in normal circumstances that will be a very rare event. +C +C The test for skipping a division is branch free (once size>=1 is tested). +C The store to the destination high limb is 0 when a divide is skipped, or +C if it's not skipped then a copy of the src high limb is used. The latter +C is in case src==dst. +C +C There's a small bias towards expecting xsize==0, by having code for +C xsize==0 in a straight line and xsize!=0 under forward jumps. +C +C Alternatives: +C +C If the divisor is normalized (high bit set) then a division step can +C always be skipped, since the high destination limb is always 0 or 1 in +C that case. It doesn't seem worth checking for this though, since it +C probably occurs infrequently, in particular note that big_base for a +C decimal mpn_get_str is not normalized in a 32-bit limb. + + +dnl MUL_THRESHOLD is the value of xsize+size at which the multiply by +dnl inverse method is used, rather than plain "divl"s. Minimum value 1. +dnl +dnl The inverse takes about 50 cycles to calculate, but after that the +dnl multiply is 17 c/l versus division at 42 c/l. +dnl +dnl At 3 limbs the mul is a touch faster than div on the integer part, and +dnl even more so on the fractional part. + +deflit(MUL_THRESHOLD, 3) + + +defframe(PARAM_PREINV_SHIFT, 28) dnl mpn_preinv_divrem_1 +defframe(PARAM_PREINV_INVERSE, 24) dnl mpn_preinv_divrem_1 +defframe(PARAM_CARRY, 24) dnl mpn_divrem_1c +defframe(PARAM_DIVISOR,20) +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC, 12) +defframe(PARAM_XSIZE, 8) +defframe(PARAM_DST, 4) + +defframe(SAVE_EBX, -4) +defframe(SAVE_ESI, -8) +defframe(SAVE_EDI, -12) +defframe(SAVE_EBP, -16) + +defframe(VAR_NORM, -20) +defframe(VAR_INVERSE, -24) +defframe(VAR_SRC, -28) +defframe(VAR_DST, -32) +defframe(VAR_DST_STOP,-36) + +deflit(STACK_SPACE, 36) + + TEXT + ALIGN(32) + +PROLOGUE(mpn_preinv_divrem_1) +deflit(`FRAME',0) + movl PARAM_XSIZE, %ecx + movl PARAM_DST, %edx + subl $STACK_SPACE, %esp FRAME_subl_esp(STACK_SPACE) + + movl %esi, SAVE_ESI + movl PARAM_SRC, %esi + + movl %ebx, SAVE_EBX + movl PARAM_SIZE, %ebx + + leal 8(%edx,%ecx,4), %edx C &dst[xsize+2] + movl %ebp, SAVE_EBP + movl PARAM_DIVISOR, %ebp + + movl %edx, VAR_DST_STOP C &dst[xsize+2] + movl %edi, SAVE_EDI + xorl %edi, %edi C carry + + movl -4(%esi,%ebx,4), %eax C src high limb + xor %ecx, %ecx + + C + + C + + cmpl %ebp, %eax C high cmp divisor + + cmovc( %eax, %edi) C high is carry if high n2 + leal (%ebp,%esi), %edx + + movd %mm0, %esi + + cmovc( %edx, %edi) C n - q1*d if underflow from using q1+1 + sbbl $0, %ebx C q + cmpl %eax, %ecx + + movl %ebx, (%ecx) + movl %ecx, VAR_DST + jne L(integer_top) + + +L(integer_loop_done): + + +C ----------------------------------------------------------------------------- +C +C Here, and in integer_one_left below, an sbbl $0 is used rather than a jz +C q1_ff special case. This make the code a bit smaller and simpler, and +C costs only 1 cycle (each). + +L(integer_two_left): + C eax scratch + C ebx scratch (nadj, q1) + C ecx scratch (src, dst) + C edx scratch + C esi n10 + C edi n2 + C ebp divisor + C + C mm7 rshift + + cmpl $0x80000000, %esi C n1 as 0=c, 1=nc + movl %edi, %eax C n2 + movl PARAM_SRC, %ecx + + leal (%ebp,%esi), %ebx + cmovc( %esi, %ebx) C nadj = n10 + (-n1 & d), ignoring overflow + sbbl $-1, %eax C n2+n1 + + mull VAR_INVERSE C m*(n2+n1) + + movd (%ecx), %mm0 C src low limb + + movl VAR_DST_STOP, %ecx + + C + + addl %ebx, %eax C m*(n2+n1) + nadj, low giving carry flag + leal 1(%edi), %ebx C n2+1 + movl %ebp, %eax C d + + adcl %edx, %ebx C 1 + high(n2<<32 + m*(n2+n1) + nadj) = q1+1 + + sbbl $0, %ebx + + mull %ebx C (q1+1)*d + + psllq $32, %mm0 + + psrlq %mm7, %mm0 + + C + + subl %eax, %esi + + C + + sbbl %edx, %edi C n - (q1+1)*d + movl %esi, %edi C remainder -> n2 + leal (%ebp,%esi), %edx + + movd %mm0, %esi + + cmovc( %edx, %edi) C n - q1*d if underflow from using q1+1 + sbbl $0, %ebx C q + + movl %ebx, -4(%ecx) + + +C ----------------------------------------------------------------------------- +L(integer_one_left): + C eax scratch + C ebx scratch (nadj, q1) + C ecx dst + C edx scratch + C esi n10 + C edi n2 + C ebp divisor + C + C mm7 rshift + + movl VAR_DST_STOP, %ecx + cmpl $0x80000000, %esi C n1 as 0=c, 1=nc + movl %edi, %eax C n2 + + leal (%ebp,%esi), %ebx + cmovc( %esi, %ebx) C nadj = n10 + (-n1 & d), ignoring overflow + sbbl $-1, %eax C n2+n1 + + mull VAR_INVERSE C m*(n2+n1) + + C + + C + + C + + addl %ebx, %eax C m*(n2+n1) + nadj, low giving carry flag + leal 1(%edi), %ebx C n2+1 + movl %ebp, %eax C d + + C + + adcl %edx, %ebx C 1 + high(n2<<32 + m*(n2+n1) + nadj) = q1+1 + + sbbl $0, %ebx C q1 if q1+1 overflowed + + mull %ebx + + C + + C + + C + + subl %eax, %esi + + C + + sbbl %edx, %edi C n - (q1+1)*d + movl %esi, %edi C remainder -> n2 + leal (%ebp,%esi), %edx + + cmovc( %edx, %edi) C n - q1*d if underflow from using q1+1 + sbbl $0, %ebx C q + + movl %ebx, -8(%ecx) + subl $8, %ecx + + + +L(integer_none): + cmpl $0, PARAM_XSIZE + jne L(fraction_some) + + movl %edi, %eax +L(fraction_done): + movl VAR_NORM, %ecx +L(zero_done): + movl SAVE_EBP, %ebp + + movl SAVE_EDI, %edi + movl SAVE_ESI, %esi + + movl SAVE_EBX, %ebx + addl $STACK_SPACE, %esp + + shrl %cl, %eax + emms + + ret + + +C ----------------------------------------------------------------------------- +C +C Special case for q1=0xFFFFFFFF, giving q=0xFFFFFFFF meaning the low dword +C of q*d is simply -d and the remainder n-q*d = n10+d + +L(q1_ff): + C eax (divisor) + C ebx (q1+1 == 0) + C ecx + C edx + C esi n10 + C edi n2 + C ebp divisor + + movl VAR_DST, %ecx + movl VAR_DST_STOP, %edx + subl $4, %ecx + + psrlq %mm7, %mm0 + leal (%ebp,%esi), %edi C n-q*d remainder -> next n2 + movl %ecx, VAR_DST + + movd %mm0, %esi C next n10 + + movl $-1, (%ecx) + cmpl %ecx, %edx + jne L(integer_top) + + jmp L(integer_loop_done) + + + +C ----------------------------------------------------------------------------- +C +C Being the fractional part, the "source" limbs are all zero, meaning +C n10=0, n1=0, and hence nadj=0, leading to many instructions eliminated. +C +C The loop runs at 15 cycles. The dependent chain is the same as the +C general case above, but without the n2+n1 stage (due to n1==0), so 15 +C would seem to be the lower bound. +C +C A not entirely obvious simplification is that q1+1 never overflows a limb, +C and so there's no need for the sbbl $0 or jz q1_ff from the general case. +C q1 is the high word of m*n2+b*n2 and the following shows q1<=b-2 always. +C rnd() means rounding down to a multiple of d. +C +C m*n2 + b*n2 <= m*(d-1) + b*(d-1) +C = m*d + b*d - m - b +C = floor((b(b-d)-1)/d)*d + b*d - m - b +C = rnd(b(b-d)-1) + b*d - m - b +C = rnd(b(b-d)-1 + b*d) - m - b +C = rnd(b*b-1) - m - b +C <= (b-2)*b +C +C Unchanged from the general case is that the final quotient limb q can be +C either q1 or q1+1, and the q1+1 case occurs often. This can be seen from +C equation 8.4 of the paper which simplifies as follows when n1==0 and +C n0==0. +C +C n-q1*d = (n2*k+q0*d)/b <= d + (d*d-2d)/b +C +C As before, the instruction groupings and empty comments show a naive +C in-order view of the code, which is made a nonsense by out of order +C execution. There's 17 cycles shown, but it executes at 15. +C +C Rotating the store q and remainder->n2 instructions up to the top of the +C loop gets the run time down from 16 to 15. + + ALIGN(16) +L(fraction_some): + C eax + C ebx + C ecx + C edx + C esi + C edi carry + C ebp divisor + + movl PARAM_DST, %esi + movl VAR_DST_STOP, %ecx C &dst[xsize+2] + movl %edi, %eax + + subl $8, %ecx C &dst[xsize] + jmp L(fraction_entry) + + + ALIGN(16) +L(fraction_top): + C eax n2 carry, then scratch + C ebx scratch (nadj, q1) + C ecx dst, decrementing + C edx scratch + C esi dst stop point + C edi (will be n2) + C ebp divisor + + movl %ebx, (%ecx) C previous q + movl %eax, %edi C remainder->n2 + +L(fraction_entry): + mull VAR_INVERSE C m*n2 + + movl %ebp, %eax C d + subl $4, %ecx C dst + leal 1(%edi), %ebx + + C + + C + + C + + C + + addl %edx, %ebx C 1 + high(n2<<32 + m*n2) = q1+1 + + mull %ebx C (q1+1)*d + + C + + C + + C + + negl %eax C low of n - (q1+1)*d + + C + + sbbl %edx, %edi C high of n - (q1+1)*d, caring only about carry + leal (%ebp,%eax), %edx + + cmovc( %edx, %eax) C n - q1*d if underflow from using q1+1 + sbbl $0, %ebx C q + cmpl %esi, %ecx + + jne L(fraction_top) + + + movl %ebx, (%ecx) + jmp L(fraction_done) + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/mmx/lshift.asm b/gmp-6.3.0/mpn/x86/k7/mmx/lshift.asm new file mode 100644 index 0000000..b3383cf --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mmx/lshift.asm @@ -0,0 +1,481 @@ +dnl AMD K7 mpn_lshift -- mpn left shift. + +dnl Copyright 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K7: 1.21 cycles/limb (at 16 limbs/loop). + + + +dnl K7: UNROLL_COUNT cycles/limb +dnl 4 1.51 +dnl 8 1.26 +dnl 16 1.21 +dnl 32 1.2 +dnl Maximum possible with the current code is 64. + +deflit(UNROLL_COUNT, 16) + + +C mp_limb_t mpn_lshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned shift); +C +C Shift src,size left by shift many bits and store the result in dst,size. +C Zeros are shifted in at the right. The bits shifted out at the left are +C the return value. +C +C The comments in mpn_rshift apply here too. + +ifdef(`PIC',` +deflit(UNROLL_THRESHOLD, 10) +',` +deflit(UNROLL_THRESHOLD, 10) +') + +defframe(PARAM_SHIFT,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +defframe(SAVE_EDI, -4) +defframe(SAVE_ESI, -8) +defframe(SAVE_EBX, -12) +deflit(SAVE_SIZE, 12) + + TEXT + ALIGN(32) + +PROLOGUE(mpn_lshift) +deflit(`FRAME',0) + + movl PARAM_SIZE, %eax + movl PARAM_SRC, %edx + subl $SAVE_SIZE, %esp +deflit(`FRAME',SAVE_SIZE) + + movl PARAM_SHIFT, %ecx + movl %edi, SAVE_EDI + + movl PARAM_DST, %edi + decl %eax + jnz L(more_than_one_limb) + + movl (%edx), %edx + + shldl( %cl, %edx, %eax) C eax was decremented to zero + + shll %cl, %edx + + movl %edx, (%edi) + movl SAVE_EDI, %edi + addl $SAVE_SIZE, %esp + + ret + + +C ----------------------------------------------------------------------------- +L(more_than_one_limb): + C eax size-1 + C ebx + C ecx shift + C edx src + C esi + C edi dst + C ebp + + movd PARAM_SHIFT, %mm6 + movd (%edx,%eax,4), %mm5 C src high limb + cmp $UNROLL_THRESHOLD-1, %eax + + jae L(unroll) + negl %ecx + movd (%edx), %mm4 C src low limb + + addl $32, %ecx + + movd %ecx, %mm7 + +L(simple_top): + C eax loop counter, limbs + C ebx + C ecx + C edx src + C esi + C edi dst + C ebp + C + C mm0 scratch + C mm4 src low limb + C mm5 src high limb + C mm6 shift + C mm7 32-shift + + movq -4(%edx,%eax,4), %mm0 + decl %eax + + psrlq %mm7, %mm0 + + movd %mm0, 4(%edi,%eax,4) + jnz L(simple_top) + + + psllq %mm6, %mm5 + psllq %mm6, %mm4 + + psrlq $32, %mm5 + movd %mm4, (%edi) C dst low limb + + movd %mm5, %eax C return value + + movl SAVE_EDI, %edi + addl $SAVE_SIZE, %esp + emms + + ret + + +C ----------------------------------------------------------------------------- + ALIGN(16) +L(unroll): + C eax size-1 + C ebx (saved) + C ecx shift + C edx src + C esi + C edi dst + C ebp + C + C mm5 src high limb, for return value + C mm6 lshift + + movl %esi, SAVE_ESI + movl %ebx, SAVE_EBX + leal -4(%edx,%eax,4), %edx C &src[size-2] + + testb $4, %dl + movq (%edx), %mm1 C src high qword + + jz L(start_src_aligned) + + + C src isn't aligned, process high limb (marked xxx) separately to + C make it so + C + C source -4(edx,%eax,4) + C | + C +-------+-------+-------+-- + C | xxx | + C +-------+-------+-------+-- + C 0mod8 4mod8 0mod8 + C + C dest -4(edi,%eax,4) + C | + C +-------+-------+-- + C | xxx | | + C +-------+-------+-- + + psllq %mm6, %mm1 + subl $4, %edx + movl %eax, PARAM_SIZE C size-1 + + psrlq $32, %mm1 + decl %eax C size-2 is new size-1 + + movd %mm1, 4(%edi,%eax,4) + movq (%edx), %mm1 C new src high qword +L(start_src_aligned): + + + leal -4(%edi,%eax,4), %edi C &dst[size-2] + psllq %mm6, %mm5 + + testl $4, %edi + psrlq $32, %mm5 C return value + + jz L(start_dst_aligned) + + + C dst isn't aligned, subtract 4 bytes to make it so, and pretend the + C shift is 32 bits extra. High limb of dst (marked xxx) handled + C here separately. + C + C source %edx + C +-------+-------+-- + C | mm1 | + C +-------+-------+-- + C 0mod8 4mod8 + C + C dest %edi + C +-------+-------+-------+-- + C | xxx | + C +-------+-------+-------+-- + C 0mod8 4mod8 0mod8 + + movq %mm1, %mm0 + psllq %mm6, %mm1 + addl $32, %ecx C shift+32 + + psrlq $32, %mm1 + + movd %mm1, 4(%edi) + movq %mm0, %mm1 + subl $4, %edi + + movd %ecx, %mm6 C new lshift +L(start_dst_aligned): + + decl %eax C size-2, two last limbs handled at end + movq %mm1, %mm2 C copy of src high qword + negl %ecx + + andl $-2, %eax C round size down to even + addl $64, %ecx + + movl %eax, %ebx + negl %eax + + andl $UNROLL_MASK, %eax + decl %ebx + + shll %eax + + movd %ecx, %mm7 C rshift = 64-lshift + +ifdef(`PIC',` + call L(pic_calc) +L(here): +',` + leal L(entry) (%eax,%eax,4), %esi +') + shrl $UNROLL_LOG2, %ebx C loop counter + + leal ifelse(UNROLL_BYTES,256,128) -8(%edx,%eax,2), %edx + leal ifelse(UNROLL_BYTES,256,128) (%edi,%eax,2), %edi + movl PARAM_SIZE, %eax C for use at end + jmp *%esi + + +ifdef(`PIC',` +L(pic_calc): + C See mpn/x86/README about old gas bugs + leal (%eax,%eax,4), %esi + addl $L(entry)-L(here), %esi + addl (%esp), %esi + + ret_internal +') + + +C ----------------------------------------------------------------------------- + ALIGN(32) +L(top): + C eax size (for use at end) + C ebx loop counter + C ecx rshift + C edx src + C esi computed jump + C edi dst + C ebp + C + C mm0 scratch + C mm1 \ carry (alternating, mm2 first) + C mm2 / + C mm6 lshift + C mm7 rshift + C + C 10 code bytes/limb + C + C The two chunks differ in whether mm1 or mm2 hold the carry. + C The computed jump puts the initial carry in both mm1 and mm2. + +L(entry): +deflit(CHUNK_COUNT, 4) +forloop(i, 0, UNROLL_COUNT/CHUNK_COUNT-1, ` + deflit(`disp0', eval(-i*CHUNK_COUNT*4 ifelse(UNROLL_BYTES,256,-128))) + deflit(`disp1', eval(disp0 - 8)) + +Zdisp( movq, disp0,(%edx), %mm0) + psllq %mm6, %mm2 + + movq %mm0, %mm1 + psrlq %mm7, %mm0 + + por %mm2, %mm0 +Zdisp( movq, %mm0, disp0,(%edi)) + + +Zdisp( movq, disp1,(%edx), %mm0) + psllq %mm6, %mm1 + + movq %mm0, %mm2 + psrlq %mm7, %mm0 + + por %mm1, %mm0 +Zdisp( movq, %mm0, disp1,(%edi)) +') + + subl $UNROLL_BYTES, %edx + subl $UNROLL_BYTES, %edi + decl %ebx + + jns L(top) + + + +define(`disp', `m4_empty_if_zero(eval($1 ifelse(UNROLL_BYTES,256,-128)))') + +L(end): + testb $1, %al + movl SAVE_EBX, %ebx + psllq %mm6, %mm2 C wanted left shifted in all cases below + + movd %mm5, %eax + + movl SAVE_ESI, %esi + jz L(end_even) + + +L(end_odd): + + C Size odd, destination was aligned. + C + C source edx+8 edx+4 + C --+---------------+-------+ + C | mm2 | | + C --+---------------+-------+ + C + C dest edi + C --+---------------+---------------+-------+ + C | written | | | + C --+---------------+---------------+-------+ + C + C mm6 = shift + C mm7 = ecx = 64-shift + + + C Size odd, destination was unaligned. + C + C source edx+8 edx+4 + C --+---------------+-------+ + C | mm2 | | + C --+---------------+-------+ + C + C dest edi + C --+---------------+---------------+ + C | written | | + C --+---------------+---------------+ + C + C mm6 = shift+32 + C mm7 = ecx = 64-(shift+32) + + + C In both cases there's one extra limb of src to fetch and combine + C with mm2 to make a qword at (%edi), and in the aligned case + C there's an extra limb of dst to be formed from that extra src limb + C left shifted. + + movd disp(4) (%edx), %mm0 + testb $32, %cl + + movq %mm0, %mm1 + psllq $32, %mm0 + + psrlq %mm7, %mm0 + psllq %mm6, %mm1 + + por %mm2, %mm0 + + movq %mm0, disp(0) (%edi) + jz L(end_odd_unaligned) + movd %mm1, disp(-4) (%edi) +L(end_odd_unaligned): + + movl SAVE_EDI, %edi + addl $SAVE_SIZE, %esp + emms + + ret + + +L(end_even): + + C Size even, destination was aligned. + C + C source edx+8 + C --+---------------+ + C | mm2 | + C --+---------------+ + C + C dest edi + C --+---------------+---------------+ + C | written | | + C --+---------------+---------------+ + C + C mm6 = shift + C mm7 = ecx = 64-shift + + + C Size even, destination was unaligned. + C + C source edx+8 + C --+---------------+ + C | mm2 | + C --+---------------+ + C + C dest edi+4 + C --+---------------+-------+ + C | written | | + C --+---------------+-------+ + C + C mm6 = shift+32 + C mm7 = ecx = 64-(shift+32) + + + C The movq for the aligned case overwrites the movd for the + C unaligned case. + + movq %mm2, %mm0 + psrlq $32, %mm2 + + testb $32, %cl + movd %mm2, disp(4) (%edi) + + jz L(end_even_unaligned) + movq %mm0, disp(0) (%edi) +L(end_even_unaligned): + + movl SAVE_EDI, %edi + addl $SAVE_SIZE, %esp + emms + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/mmx/popham.asm b/gmp-6.3.0/mpn/x86/k7/mmx/popham.asm new file mode 100644 index 0000000..95965b7 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mmx/popham.asm @@ -0,0 +1,213 @@ +dnl AMD K7 mpn_popcount, mpn_hamdist -- population count and hamming +dnl distance. + +dnl Copyright 2000-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C popcount hamdist +C P3 generic 6.5 7 +C P3 model 9 (Banias) 5.7 6.1 +C P3 model 13 (Dothan) 5.75 6 +C K7 5 6 + +C unsigned long mpn_popcount (mp_srcptr src, mp_size_t size); +C unsigned long mpn_hamdist (mp_srcptr src, mp_srcptr src2, mp_size_t size); +C +C The code here is almost certainly not optimal, but is already a 3x speedup +C over the generic C code. The main improvement would be to interleave +C processing of two qwords in the loop so as to fully exploit the available +C execution units, possibly leading to 3.25 c/l (13 cycles for 4 limbs). +C +C The loop is based on the example "Efficient 64-bit population count using +C MMX instructions" in the Athlon Optimization Guide, AMD document 22007, +C page 158 of rev E (reference in mpn/x86/k7/README). + +ifdef(`OPERATION_popcount',, +`ifdef(`OPERATION_hamdist',, +`m4_error(`Need OPERATION_popcount or OPERATION_hamdist defined +')')') + +define(HAM, +m4_assert_numargs(1) +`ifdef(`OPERATION_hamdist',`$1')') + +define(POP, +m4_assert_numargs(1) +`ifdef(`OPERATION_popcount',`$1')') + +HAM(` +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC2, 8) +defframe(PARAM_SRC, 4) +define(M4_function,mpn_hamdist) +') +POP(` +defframe(PARAM_SIZE, 8) +defframe(PARAM_SRC, 4) +define(M4_function,mpn_popcount) +') + +MULFUNC_PROLOGUE(mpn_popcount mpn_hamdist) + + +ifdef(`PIC',,` + dnl non-PIC + + RODATA + ALIGN(8) + +L(rodata_AAAAAAAAAAAAAAAA): + .long 0xAAAAAAAA + .long 0xAAAAAAAA + +L(rodata_3333333333333333): + .long 0x33333333 + .long 0x33333333 + +L(rodata_0F0F0F0F0F0F0F0F): + .long 0x0F0F0F0F + .long 0x0F0F0F0F +') + + TEXT + ALIGN(32) + +PROLOGUE(M4_function) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + +ifdef(`PIC',` + movl $0xAAAAAAAA, %eax + movl $0x33333333, %edx + + movd %eax, %mm7 + movd %edx, %mm6 + + movl $0x0F0F0F0F, %eax + + punpckldq %mm7, %mm7 + punpckldq %mm6, %mm6 + + movd %eax, %mm5 + movd %edx, %mm4 + + punpckldq %mm5, %mm5 + +',` + movq L(rodata_AAAAAAAAAAAAAAAA), %mm7 + movq L(rodata_3333333333333333), %mm6 + movq L(rodata_0F0F0F0F0F0F0F0F), %mm5 +') + pxor %mm4, %mm4 + +define(REG_AAAAAAAAAAAAAAAA,%mm7) +define(REG_3333333333333333,%mm6) +define(REG_0F0F0F0F0F0F0F0F,%mm5) +define(REG_0000000000000000,%mm4) + + + movl PARAM_SRC, %eax +HAM(` movl PARAM_SRC2, %edx') + + pxor %mm2, %mm2 C total + + shrl %ecx + jnc L(top) + + movd (%eax,%ecx,8), %mm1 + +HAM(` movd (%edx,%ecx,8), %mm0 + pxor %mm0, %mm1 +') + orl %ecx, %ecx + jmp L(loaded) + + + ALIGN(16) +L(top): + C eax src + C ebx + C ecx counter, qwords, decrementing + C edx [hamdist] src2 + C + C mm0 (scratch) + C mm1 (scratch) + C mm2 total (low dword) + C mm3 + C mm4 \ + C mm5 | special constants + C mm6 | + C mm7 / + + movq -8(%eax,%ecx,8), %mm1 + +HAM(` pxor -8(%edx,%ecx,8), %mm1') + decl %ecx + +L(loaded): + movq %mm1, %mm0 + pand REG_AAAAAAAAAAAAAAAA, %mm1 + + psrlq $1, %mm1 + + psubd %mm1, %mm0 C bit pairs + + + movq %mm0, %mm1 + psrlq $2, %mm0 + + pand REG_3333333333333333, %mm0 + pand REG_3333333333333333, %mm1 + + paddd %mm1, %mm0 C nibbles + + + movq %mm0, %mm1 + psrlq $4, %mm0 + + pand REG_0F0F0F0F0F0F0F0F, %mm0 + pand REG_0F0F0F0F0F0F0F0F, %mm1 + + paddd %mm1, %mm0 C bytes + + + psadbw( %mm4, %mm0) + + paddd %mm0, %mm2 C add to total + jnz L(top) + + + movd %mm2, %eax + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/mmx/rshift.asm b/gmp-6.3.0/mpn/x86/k7/mmx/rshift.asm new file mode 100644 index 0000000..345d23a --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mmx/rshift.asm @@ -0,0 +1,480 @@ +dnl AMD K7 mpn_rshift -- mpn right shift. + +dnl Copyright 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C K7: 1.21 cycles/limb (at 16 limbs/loop). + + + +dnl K7: UNROLL_COUNT cycles/limb +dnl 4 1.51 +dnl 8 1.26 +dnl 16 1.21 +dnl 32 1.2 +dnl Maximum possible with the current code is 64. + +deflit(UNROLL_COUNT, 16) + + +C mp_limb_t mpn_rshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned shift); +C +C Shift src,size right by shift many bits and store the result in dst,size. +C Zeros are shifted in at the left. The bits shifted out at the right are +C the return value. +C +C This code uses 64-bit MMX operations, which makes it possible to handle +C two limbs at a time, for a theoretical 1.0 cycles/limb. Plain integer +C code, on the other hand, suffers from shrd being a vector path decode and +C running at 3 cycles back-to-back. +C +C Full speed depends on source and destination being aligned, and some hairy +C setups and finish-ups are done to arrange this for the loop. + +ifdef(`PIC',` +deflit(UNROLL_THRESHOLD, 10) +',` +deflit(UNROLL_THRESHOLD, 10) +') + +defframe(PARAM_SHIFT,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +defframe(SAVE_EDI, -4) +defframe(SAVE_ESI, -8) +defframe(SAVE_EBX, -12) +deflit(SAVE_SIZE, 12) + + TEXT + ALIGN(32) + +PROLOGUE(mpn_rshift) +deflit(`FRAME',0) + + movl PARAM_SIZE, %eax + movl PARAM_SRC, %edx + subl $SAVE_SIZE, %esp +deflit(`FRAME',SAVE_SIZE) + + movl PARAM_SHIFT, %ecx + movl %edi, SAVE_EDI + + movl PARAM_DST, %edi + decl %eax + jnz L(more_than_one_limb) + + movl (%edx), %edx C src limb + + shrdl( %cl, %edx, %eax) C eax was decremented to zero + + shrl %cl, %edx + + movl %edx, (%edi) C dst limb + movl SAVE_EDI, %edi + addl $SAVE_SIZE, %esp + + ret + + +C ----------------------------------------------------------------------------- +L(more_than_one_limb): + C eax size-1 + C ebx + C ecx shift + C edx src + C esi + C edi dst + C ebp + + movd PARAM_SHIFT, %mm6 C rshift + movd (%edx), %mm5 C src low limb + cmp $UNROLL_THRESHOLD-1, %eax + + jae L(unroll) + leal (%edx,%eax,4), %edx C &src[size-1] + leal -4(%edi,%eax,4), %edi C &dst[size-2] + + movd (%edx), %mm4 C src high limb + negl %eax + + +L(simple_top): + C eax loop counter, limbs, negative + C ebx + C ecx shift + C edx carry + C edx &src[size-1] + C edi &dst[size-2] + C ebp + C + C mm0 scratch + C mm4 src high limb + C mm5 src low limb + C mm6 shift + + movq (%edx,%eax,4), %mm0 + incl %eax + + psrlq %mm6, %mm0 + + movd %mm0, (%edi,%eax,4) + jnz L(simple_top) + + + psllq $32, %mm5 + psrlq %mm6, %mm4 + + psrlq %mm6, %mm5 + movd %mm4, 4(%edi) C dst high limb + + movd %mm5, %eax C return value + + movl SAVE_EDI, %edi + addl $SAVE_SIZE, %esp + emms + + ret + + +C ----------------------------------------------------------------------------- + ALIGN(16) +L(unroll): + C eax size-1 + C ebx + C ecx shift + C edx src + C esi + C edi dst + C ebp + C + C mm5 src low limb + C mm6 rshift + + testb $4, %dl + movl %esi, SAVE_ESI + movl %ebx, SAVE_EBX + + psllq $32, %mm5 + jz L(start_src_aligned) + + + C src isn't aligned, process low limb separately (marked xxx) and + C step src and dst by one limb, making src aligned. + C + C source edx + C --+-------+-------+-------+ + C | xxx | + C --+-------+-------+-------+ + C 4mod8 0mod8 4mod8 + C + C dest edi + C --+-------+-------+ + C | | xxx | + C --+-------+-------+ + + movq (%edx), %mm0 C src low two limbs + addl $4, %edx + movl %eax, PARAM_SIZE C size-1 + + addl $4, %edi + decl %eax C size-2 is new size-1 + + psrlq %mm6, %mm0 + movl %edi, PARAM_DST C new dst + + movd %mm0, -4(%edi) +L(start_src_aligned): + + + movq (%edx), %mm1 C src low two limbs + decl %eax C size-2, two last limbs handled at end + testl $4, %edi + + psrlq %mm6, %mm5 + jz L(start_dst_aligned) + + + C dst isn't aligned, add 4 to make it so, and pretend the shift is + C 32 bits extra. Low limb of dst (marked xxx) handled here separately. + C + C source edx + C --+-------+-------+ + C | mm1 | + C --+-------+-------+ + C 4mod8 0mod8 + C + C dest edi + C --+-------+-------+-------+ + C | xxx | + C --+-------+-------+-------+ + C 4mod8 0mod8 4mod8 + + movq %mm1, %mm0 + psrlq %mm6, %mm1 + addl $32, %ecx C shift+32 + + movd %mm1, (%edi) + movq %mm0, %mm1 + addl $4, %edi C new dst + + movd %ecx, %mm6 +L(start_dst_aligned): + + + movq %mm1, %mm2 C copy of src low two limbs + negl %ecx + andl $-2, %eax C round size down to even + + movl %eax, %ebx + negl %eax + addl $64, %ecx + + andl $UNROLL_MASK, %eax + decl %ebx + + shll %eax + + movd %ecx, %mm7 C lshift = 64-rshift + +ifdef(`PIC',` + call L(pic_calc) +L(here): +',` + leal L(entry) (%eax,%eax,4), %esi + negl %eax +') + shrl $UNROLL_LOG2, %ebx C loop counter + + leal ifelse(UNROLL_BYTES,256,128+) 8(%edx,%eax,2), %edx + leal ifelse(UNROLL_BYTES,256,128) (%edi,%eax,2), %edi + movl PARAM_SIZE, %eax C for use at end + + jmp *%esi + + +ifdef(`PIC',` +L(pic_calc): + C See mpn/x86/README about old gas bugs + leal (%eax,%eax,4), %esi + addl $L(entry)-L(here), %esi + addl (%esp), %esi + negl %eax + + ret_internal +') + + +C ----------------------------------------------------------------------------- + ALIGN(64) +L(top): + C eax size, for use at end + C ebx loop counter + C ecx lshift + C edx src + C esi was computed jump + C edi dst + C ebp + C + C mm0 scratch + C mm1 \ carry (alternating) + C mm2 / + C mm6 rshift + C mm7 lshift + C + C 10 code bytes/limb + C + C The two chunks differ in whether mm1 or mm2 hold the carry. + C The computed jump puts the initial carry in both mm1 and mm2. + +L(entry): +deflit(CHUNK_COUNT, 4) +forloop(i, 0, UNROLL_COUNT/CHUNK_COUNT-1, ` + deflit(`disp0', eval(i*CHUNK_COUNT*4 ifelse(UNROLL_BYTES,256,-128))) + deflit(`disp1', eval(disp0 + 8)) + +Zdisp( movq, disp0,(%edx), %mm0) + psrlq %mm6, %mm2 + + movq %mm0, %mm1 + psllq %mm7, %mm0 + + por %mm2, %mm0 +Zdisp( movq, %mm0, disp0,(%edi)) + + +Zdisp( movq, disp1,(%edx), %mm0) + psrlq %mm6, %mm1 + + movq %mm0, %mm2 + psllq %mm7, %mm0 + + por %mm1, %mm0 +Zdisp( movq, %mm0, disp1,(%edi)) +') + + addl $UNROLL_BYTES, %edx + addl $UNROLL_BYTES, %edi + decl %ebx + + jns L(top) + + +deflit(`disp0', ifelse(UNROLL_BYTES,256,-128)) +deflit(`disp1', eval(disp0-0 + 8)) + + testb $1, %al + psrlq %mm6, %mm2 C wanted rshifted in all cases below + movl SAVE_ESI, %esi + + movd %mm5, %eax C return value + + movl SAVE_EBX, %ebx + jz L(end_even) + + + C Size odd, destination was aligned. + C + C source + C edx + C +-------+---------------+-- + C | | mm2 | + C +-------+---------------+-- + C + C dest edi + C +-------+---------------+---------------+-- + C | | | written | + C +-------+---------------+---------------+-- + C + C mm6 = shift + C mm7 = ecx = 64-shift + + + C Size odd, destination was unaligned. + C + C source + C edx + C +-------+---------------+-- + C | | mm2 | + C +-------+---------------+-- + C + C dest edi + C +---------------+---------------+-- + C | | written | + C +---------------+---------------+-- + C + C mm6 = shift+32 + C mm7 = ecx = 64-(shift+32) + + + C In both cases there's one extra limb of src to fetch and combine + C with mm2 to make a qword to store, and in the aligned case there's + C a further extra limb of dst to be formed. + + + movd disp0(%edx), %mm0 + movq %mm0, %mm1 + + psllq %mm7, %mm0 + testb $32, %cl + + por %mm2, %mm0 + psrlq %mm6, %mm1 + + movq %mm0, disp0(%edi) + jz L(finish_odd_unaligned) + + movd %mm1, disp1(%edi) +L(finish_odd_unaligned): + + movl SAVE_EDI, %edi + addl $SAVE_SIZE, %esp + emms + + ret + + +L(end_even): + + C Size even, destination was aligned. + C + C source + C +---------------+-- + C | mm2 | + C +---------------+-- + C + C dest edi + C +---------------+---------------+-- + C | | mm3 | + C +---------------+---------------+-- + C + C mm6 = shift + C mm7 = ecx = 64-shift + + + C Size even, destination was unaligned. + C + C source + C +---------------+-- + C | mm2 | + C +---------------+-- + C + C dest edi + C +-------+---------------+-- + C | | mm3 | + C +-------+---------------+-- + C + C mm6 = shift+32 + C mm7 = 64-(shift+32) + + + C The movd for the unaligned case is the same data as the movq for + C the aligned case, it's just a choice between whether one or two + C limbs should be written. + + + testb $32, %cl + movd %mm2, disp0(%edi) + + jz L(end_even_unaligned) + + movq %mm2, disp0(%edi) +L(end_even_unaligned): + + movl SAVE_EDI, %edi + addl $SAVE_SIZE, %esp + emms + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/mod_1_1.asm b/gmp-6.3.0/mpn/x86/k7/mod_1_1.asm new file mode 100644 index 0000000..1bbe6f9 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mod_1_1.asm @@ -0,0 +1,221 @@ +dnl x86-32 mpn_mod_1_1p, requiring cmov. + +dnl Contributed to the GNU project by Niels Möller and Torbjorn Granlund. + +dnl Copyright 2010, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C P5 ? +C P6 model 0-8,10-12 ? +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) ? +C P4 model 0 (Willamette) ? +C P4 model 1 (?) ? +C P4 model 2 (Northwood) ? +C P4 model 3 (Prescott) ? +C P4 model 4 (Nocona) ? +C AMD K6 ? +C AMD K7 7 +C AMD K8 ? + +define(`B2mb', `%ebx') +define(`r0', `%esi') +define(`r2', `%ebp') +define(`t0', `%edi') +define(`ap', `%ecx') C Also shift count + +C Stack frame +C pre 36(%esp) +C b 32(%esp) +C n 28(%esp) +C ap 24(%esp) +C return 20(%esp) +C %ebp 16(%esp) +C %edi 12(%esp) +C %esi 8(%esp) +C %ebx 4(%esp) +C B2mod (%esp) + +define(`B2modb', `(%esp)') +define(`n', `28(%esp)') +define(`b', `32(%esp)') +define(`pre', `36(%esp)') + +C mp_limb_t +C mpn_mod_1_1p (mp_srcptr ap, mp_size_t n, mp_limb_t b, mp_limb_t pre[4]) +C +C The pre array contains bi, cnt, B1modb, B2modb +C Note: This implementation needs B1modb only when cnt > 0 + +ASM_START() + TEXT + ALIGN(8) +PROLOGUE(mpn_mod_1_1p) + push %ebp + push %edi + push %esi + push %ebx + mov 32(%esp), %ebp C pre[] + + mov 12(%ebp), %eax C B2modb + push %eax C Put it on stack + + mov n, %edx + mov 24(%esp), ap + + lea (ap, %edx, 4), ap + mov -4(ap), %eax + cmp $3, %edx + jnc L(first) + mov -8(ap), r0 + jmp L(reduce_two) + +L(first): + C First iteration, no r2 + mull B2modb + mov -12(ap), r0 + add %eax, r0 + mov -8(ap), %eax + adc %edx, %eax + sbb r2, r2 + subl $3, n + lea -16(ap), ap + jz L(reduce_three) + + mov B2modb, B2mb + sub b, B2mb + lea (B2mb, r0), t0 + jmp L(mid) + + ALIGN(16) +L(top): C Loopmixed to 7 c/l on k7 + add %eax, r0 + lea (B2mb, r0), t0 + mov r2, %eax + adc %edx, %eax + sbb r2, r2 +L(mid): mull B2modb + and B2modb, r2 + add r0, r2 + decl n + mov (ap), r0 + cmovc( t0, r2) + lea -4(ap), ap + jnz L(top) + + add %eax, r0 + mov r2, %eax + adc %edx, %eax + sbb r2, r2 + +L(reduce_three): + C Eliminate r2 + and b, r2 + sub r2, %eax + +L(reduce_two): + mov pre, %ebp + movb 4(%ebp), %cl + test %cl, %cl + jz L(normalized) + + C Unnormalized, use B1modb to reduce to size < B b + mull 8(%ebp) + xor t0, t0 + add %eax, r0 + adc %edx, t0 + mov t0, %eax + + C Left-shift to normalize + shld %cl, r0, %eax C Always use shld? + + shl %cl, r0 + jmp L(udiv) + +L(normalized): + mov %eax, t0 + sub b, t0 + cmovnc( t0, %eax) + +L(udiv): + lea 1(%eax), t0 + mull (%ebp) + mov b, %ebx C Needed in register for lea + add r0, %eax + adc t0, %edx + imul %ebx, %edx + sub %edx, r0 + cmp r0, %eax + lea (%ebx, r0), %eax + cmovnc( r0, %eax) + cmp %ebx, %eax + jnc L(fix) +L(ok): shr %cl, %eax + + add $4, %esp + pop %ebx + pop %esi + pop %edi + pop %ebp + + ret +L(fix): sub %ebx, %eax + jmp L(ok) +EPILOGUE() + +PROLOGUE(mpn_mod_1_1p_cps) + push %ebp + mov 12(%esp), %ebp + push %esi + bsr %ebp, %ecx + push %ebx + xor $31, %ecx + mov 16(%esp), %esi + sal %cl, %ebp + mov %ebp, %edx + not %edx + mov $-1, %eax + div %ebp C On K7, invert_limb would be a few cycles faster. + mov %eax, (%esi) C store bi + mov %ecx, 4(%esi) C store cnt + neg %ebp + mov $1, %edx + shld %cl, %eax, %edx + imul %ebp, %edx + shr %cl, %edx + imul %ebp, %eax + mov %edx, 8(%esi) C store B1modb + mov %eax, 12(%esi) C store B2modb + pop %ebx + pop %esi + pop %ebp + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/mod_1_4.asm b/gmp-6.3.0/mpn/x86/k7/mod_1_4.asm new file mode 100644 index 0000000..bb7597e --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mod_1_4.asm @@ -0,0 +1,260 @@ +dnl x86-32 mpn_mod_1s_4p, requiring cmov. + +dnl Contributed to the GNU project by Torbjorn Granlund. + +dnl Copyright 2009, 2010 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C P5 ? +C P6 model 0-8,10-12 ? +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) 6 +C P4 model 0 (Willamette) ? +C P4 model 1 (?) ? +C P4 model 2 (Northwood) 15.5 +C P4 model 3 (Prescott) ? +C P4 model 4 (Nocona) ? +C AMD K6 ? +C AMD K7 4.75 +C AMD K8 ? + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(mpn_mod_1s_4p) + push %ebp + push %edi + push %esi + push %ebx + sub $28, %esp + mov 60(%esp), %edi C cps[] + mov 8(%edi), %eax + mov 12(%edi), %edx + mov 16(%edi), %ecx + mov 20(%edi), %esi + mov 24(%edi), %edi + mov %eax, 4(%esp) + mov %edx, 8(%esp) + mov %ecx, 12(%esp) + mov %esi, 16(%esp) + mov %edi, 20(%esp) + mov 52(%esp), %eax C n + xor %edi, %edi + mov 48(%esp), %esi C up + lea -12(%esi,%eax,4), %esi + and $3, %eax + je L(b0) + cmp $2, %eax + jc L(b1) + je L(b2) + +L(b3): mov 4(%esi), %eax + mull 4(%esp) + mov (%esi), %ebp + add %eax, %ebp + adc %edx, %edi + mov 8(%esi), %eax + mull 8(%esp) + lea -12(%esi), %esi + jmp L(m0) + +L(b0): mov (%esi), %eax + mull 4(%esp) + mov -4(%esi), %ebp + add %eax, %ebp + adc %edx, %edi + mov 4(%esi), %eax + mull 8(%esp) + add %eax, %ebp + adc %edx, %edi + mov 8(%esi), %eax + mull 12(%esp) + lea -16(%esi), %esi + jmp L(m0) + +L(b1): mov 8(%esi), %ebp + lea -4(%esi), %esi + jmp L(m1) + +L(b2): mov 8(%esi), %edi + mov 4(%esi), %ebp + lea -8(%esi), %esi + jmp L(m1) + + ALIGN(16) +L(top): mov (%esi), %eax + mull 4(%esp) + mov -4(%esi), %ebx + xor %ecx, %ecx + add %eax, %ebx + adc %edx, %ecx + mov 4(%esi), %eax + mull 8(%esp) + add %eax, %ebx + adc %edx, %ecx + mov 8(%esi), %eax + mull 12(%esp) + add %eax, %ebx + adc %edx, %ecx + lea -16(%esi), %esi + mov 16(%esp), %eax + mul %ebp + add %eax, %ebx + adc %edx, %ecx + mov 20(%esp), %eax + mul %edi + mov %ebx, %ebp + mov %ecx, %edi +L(m0): add %eax, %ebp + adc %edx, %edi +L(m1): subl $4, 52(%esp) + ja L(top) + +L(end): mov 4(%esp), %eax + mul %edi + mov 60(%esp), %edi + add %eax, %ebp + adc $0, %edx + mov 4(%edi), %ecx + mov %edx, %esi + mov %ebp, %eax + sal %cl, %esi + mov %ecx, %ebx + neg %ecx + shr %cl, %eax + or %esi, %eax + lea 1(%eax), %esi + mull (%edi) + mov %ebx, %ecx + mov %eax, %ebx + mov %ebp, %eax + mov 56(%esp), %ebp + sal %cl, %eax + add %eax, %ebx + adc %esi, %edx + imul %ebp, %edx + sub %edx, %eax + lea (%eax,%ebp), %edx + cmp %eax, %ebx + cmovc( %edx, %eax) + mov %eax, %edx + sub %ebp, %eax + cmovc( %edx, %eax) + add $28, %esp + pop %ebx + pop %esi + pop %edi + pop %ebp + shr %cl, %eax + ret +EPILOGUE() + + ALIGN(16) +PROLOGUE(mpn_mod_1s_4p_cps) +C CAUTION: This is the same code as in pentium4/sse2/mod_1_4.asm + push %ebp + push %edi + push %esi + push %ebx + mov 20(%esp), %ebp C FIXME: avoid bp for 0-idx + mov 24(%esp), %ebx + bsr %ebx, %ecx + xor $31, %ecx + sal %cl, %ebx C b << cnt + mov %ebx, %edx + not %edx + mov $-1, %eax + div %ebx + xor %edi, %edi + sub %ebx, %edi + mov $1, %esi + mov %eax, (%ebp) C store bi + mov %ecx, 4(%ebp) C store cnt + shld %cl, %eax, %esi + imul %edi, %esi + mov %eax, %edi + mul %esi + + add %esi, %edx + shr %cl, %esi + mov %esi, 8(%ebp) C store B1modb + + not %edx + imul %ebx, %edx + lea (%edx,%ebx), %esi + cmp %edx, %eax + cmovnc( %edx, %esi) + mov %edi, %eax + mul %esi + + add %esi, %edx + shr %cl, %esi + mov %esi, 12(%ebp) C store B2modb + + not %edx + imul %ebx, %edx + lea (%edx,%ebx), %esi + cmp %edx, %eax + cmovnc( %edx, %esi) + mov %edi, %eax + mul %esi + + add %esi, %edx + shr %cl, %esi + mov %esi, 16(%ebp) C store B3modb + + not %edx + imul %ebx, %edx + lea (%edx,%ebx), %esi + cmp %edx, %eax + cmovnc( %edx, %esi) + mov %edi, %eax + mul %esi + + add %esi, %edx + shr %cl, %esi + mov %esi, 20(%ebp) C store B4modb + + not %edx + imul %ebx, %edx + add %edx, %ebx + cmp %edx, %eax + cmovnc( %edx, %ebx) + + shr %cl, %ebx + mov %ebx, 24(%ebp) C store B5modb + + pop %ebx + pop %esi + pop %edi + pop %ebp + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/mod_34lsub1.asm b/gmp-6.3.0/mpn/x86/k7/mod_34lsub1.asm new file mode 100644 index 0000000..ee3ad04 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mod_34lsub1.asm @@ -0,0 +1,188 @@ +dnl AMD K7 mpn_mod_34lsub1 -- remainder modulo 2^24-1. + +dnl Copyright 2000-2002, 2004, 2005, 2008 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C Athlon: 1 +C Hammer: 1 + + +C mp_limb_t mpn_mod_34lsub1 (mp_srcptr src, mp_size_t size) +C +C The loop form below and the 64 byte code alignment seem necessary for the +C claimed speed. This is a bit strange, since normally k7 isn't very +C sensitive to such things. Perhaps there has to be 6 instructions in the +C first 16 bytes for the BTB entry or something. + +defframe(PARAM_SIZE, 8) +defframe(PARAM_SRC, 4) + +dnl re-use parameter space +define(SAVE_EDI, `PARAM_SIZE') + + TEXT + ALIGN(64) +PROLOGUE(mpn_mod_34lsub1) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + movl PARAM_SRC, %edx + + subl $2, %ecx + ja L(three_or_more) + + movl (%edx), %eax + jb L(one) + + movl 4(%edx), %ecx + movl %eax, %edx + shrl $24, %eax C src[0] low + + andl $0xFFFFFF, %edx C src[0] high + addl %edx, %eax + movl %ecx, %edx + + andl $0xFFFF, %ecx + shrl $16, %edx C src[1] high + addl %edx, %eax + + shll $8, %ecx C src[1] low + addl %ecx, %eax + +L(one): + ret + + +L(three_or_more): + C eax + C ebx + C ecx size-2 + C edx src + C esi + C edi + + pushl %ebx FRAME_pushl() + xorl %eax, %eax + xorl %ebx, %ebx + + movl %edi, SAVE_EDI + pushl %esi FRAME_pushl() + xorl %esi, %esi C and clear carry flag + + + C code offset 0x40 at this point +L(top): + C eax acc 0mod3 + C ebx acc 1mod3 + C ecx counter, limbs + C edx src + C esi acc 2mod3 + C edi + + leal 24(%edx), %edx + leal -2(%ecx), %ecx + adcl -24(%edx), %eax + adcl -20(%edx), %ebx + adcl -16(%edx), %esi + + decl %ecx + jng L(done_loop) + + leal -2(%ecx), %ecx + adcl -12(%edx), %eax + adcl -8(%edx), %ebx + adcl -4(%edx), %esi + + decl %ecx + jg L(top) + + + leal 12(%edx), %edx + + +L(done_loop): + C ecx is -2, -1 or 0 representing 0, 1 or 2 more limbs, respectively + + incl %ecx + movl $0xFFFFFFFF, %edi + js L(combine) + + adcl -12(%edx), %eax + decl %ecx + movl $0xFFFFFF00, %edi + js L(combine) + + adcl -8(%edx), %ebx + movl $0xFFFF0000, %edi + + +L(combine): + C eax acc 0mod3 + C ebx acc 1mod3 + C ecx + C edx + C esi acc 2mod3 + C edi mask + + sbbl %ecx, %ecx C carry + movl %eax, %edx C 0mod3 + shrl $24, %eax C 0mod3 high + + andl %edi, %ecx C carry masked + andl $0x00FFFFFF, %edx C 0mod3 low + movl %ebx, %edi C 1mod3 + + subl %ecx, %eax C apply carry + shrl $16, %ebx C 1mod3 high + andl $0xFFFF, %edi + + addl %edx, %eax C apply 0mod3 low + movl %esi, %edx C 2mod3 + shll $8, %edi C 1mod3 low + + addl %ebx, %eax C apply 1mod3 high + shrl $8, %esi C 2mod3 high + movzbl %dl, %edx C 2mod3 low + + addl %edi, %eax C apply 1mod3 low + shll $16, %edx C 2mod3 low + + addl %esi, %eax C apply 2mod3 high + popl %esi FRAME_popl() + + movl SAVE_EDI, %edi + addl %edx, %eax C apply 2mod3 low + popl %ebx FRAME_popl() + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/mode1o.asm b/gmp-6.3.0/mpn/x86/k7/mode1o.asm new file mode 100644 index 0000000..2394033 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/mode1o.asm @@ -0,0 +1,181 @@ +dnl AMD K7 mpn_modexact_1_odd -- exact division style remainder. + +dnl Copyright 2000-2002, 2004, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C Athlon: 11.0 +C Hammer: 7.0 + + +C mp_limb_t mpn_modexact_1_odd (mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C mp_limb_t mpn_modexact_1c_odd (mp_srcptr src, mp_size_t size, +C mp_limb_t divisor, mp_limb_t carry); +C +C With the loop running at just 11 cycles it doesn't seem worth bothering to +C check for high31),1, +eval((UNROLL_COUNT-31)*4), +0)) + +dnl Because the last chunk of code is generated differently, a label placed +dnl at the end doesn't work. Instead calculate the implied end using the +dnl start and how many chunks of code there are. + +deflit(UNROLL_INNER_END, +`L(unroll_inner_start)+eval(UNROLL_COUNT*CODE_BYTES_PER_LIMB)') + + C eax + C ebx carry + C ecx + C edx + C esi &src[size] + C edi &dst[size] + C ebp + + movl PARAM_SIZE, %ecx + movl %ebx, (%edi) + + subl $4, %ecx + jz L(corner) + + negl %ecx +ifelse(OFFSET,0,,`subl $OFFSET, %edi') +ifelse(OFFSET,0,,`subl $OFFSET, %esi') + + movl %ecx, %edx + shll $4, %ecx + +ifdef(`PIC',` + call L(pic_calc) +L(here): +',` + leal UNROLL_INNER_END-eval(2*CODE_BYTES_PER_LIMB)(%ecx,%edx), %ecx +') + + + C The calculated jump mustn't come out to before the start of the + C code available. This is the limit UNROLL_COUNT puts on the src + C operand size, but checked here directly using the jump address. + ASSERT(ae, + `movl_text_address(L(unroll_inner_start), %eax) + cmpl %eax, %ecx') + + +C------------------------------------------------------------------------------ + ALIGN(16) +L(unroll_outer_top): + C eax + C ebx high limb to store + C ecx VAR_JMP + C edx VAR_COUNTER, limbs, negative + C esi &src[size], constant + C edi dst ptr, high of last addmul + C ebp + + movl -12+OFFSET(%esi,%edx,4), %ebp C next multiplier + movl -8+OFFSET(%esi,%edx,4), %eax C first of multiplicand + + movl %edx, VAR_COUNTER + + mull %ebp + +define(cmovX,`ifelse(eval(UNROLL_COUNT%2),0,`cmovz($@)',`cmovnz($@)')') + + testb $1, %cl + movl %edx, %ebx C high carry + movl %ecx, %edx C jump + + movl %eax, %ecx C low carry + cmovX( %ebx, %ecx) C high carry reverse + cmovX( %eax, %ebx) C low carry reverse + + leal CODE_BYTES_PER_LIMB(%edx), %eax + xorl %edx, %edx + leal 4(%edi), %edi + + movl %eax, VAR_JMP + + jmp *%eax + + +ifdef(`PIC',` +L(pic_calc): + addl (%esp), %ecx + addl $UNROLL_INNER_END-eval(2*CODE_BYTES_PER_LIMB)-L(here), %ecx + addl %edx, %ecx + ret_internal +') + + + C Must be an even address to preserve the significance of the low + C bit of the jump address indicating which way around ecx/ebx should + C start. + ALIGN(2) + +L(unroll_inner_start): + C eax next limb + C ebx carry high + C ecx carry low + C edx scratch + C esi src + C edi dst + C ebp multiplier + +forloop(`i', UNROLL_COUNT, 1, ` + deflit(`disp_src', eval(-i*4 + OFFSET)) + deflit(`disp_dst', eval(disp_src - 4)) + + m4_assert(`disp_src>=-128 && disp_src<128') + m4_assert(`disp_dst>=-128 && disp_dst<128') + +ifelse(eval(i%2),0,` +Zdisp( movl, disp_src,(%esi), %eax) + adcl %edx, %ebx + + mull %ebp + +Zdisp( addl, %ecx, disp_dst,(%edi)) + movl $0, %ecx + + adcl %eax, %ebx + +',` + dnl this bit comes out last +Zdisp( movl, disp_src,(%esi), %eax) + adcl %edx, %ecx + + mull %ebp + +Zdisp( addl, %ebx, disp_dst,(%edi)) + +ifelse(forloop_last,0, +` movl $0, %ebx') + + adcl %eax, %ecx +') +') + + C eax next limb + C ebx carry high + C ecx carry low + C edx scratch + C esi src + C edi dst + C ebp multiplier + + adcl $0, %edx + addl %ecx, -4+OFFSET(%edi) + movl VAR_JMP, %ecx + + adcl $0, %edx + + movl %edx, m4_empty_if_zero(OFFSET) (%edi) + movl VAR_COUNTER, %edx + + incl %edx + jnz L(unroll_outer_top) + + +ifelse(OFFSET,0,,` + addl $OFFSET, %esi + addl $OFFSET, %edi +') + + +C------------------------------------------------------------------------------ +L(corner): + C esi &src[size] + C edi &dst[2*size-5] + + movl -12(%esi), %ebp + movl -8(%esi), %eax + movl %eax, %ecx + + mull %ebp + + addl %eax, -4(%edi) + movl -4(%esi), %eax + + adcl $0, %edx + movl %edx, %ebx + movl %eax, %esi + + mull %ebp + + addl %ebx, %eax + + adcl $0, %edx + addl %eax, (%edi) + movl %esi, %eax + + adcl $0, %edx + movl %edx, %ebx + + mull %ecx + + addl %ebx, %eax + movl %eax, 4(%edi) + + adcl $0, %edx + movl %edx, 8(%edi) + + + +C Left shift of dst[1..2*size-2], high bit shifted out becomes dst[2*size-1]. + +L(lshift_start): + movl PARAM_SIZE, %eax + movl PARAM_DST, %edi + xorl %ecx, %ecx C clear carry + + leal (%edi,%eax,8), %edi + notl %eax C -size-1, preserve carry + + leal 2(%eax), %eax C -(size-1) + +L(lshift): + C eax counter, negative + C ebx + C ecx + C edx + C esi + C edi dst, pointing just after last limb + C ebp + + rcll -4(%edi,%eax,8) + rcll (%edi,%eax,8) + incl %eax + jnz L(lshift) + + setc %al + + movl PARAM_SRC, %esi + movl %eax, -4(%edi) C dst most significant limb + + movl PARAM_SIZE, %ecx + + +C Now add in the squares on the diagonal, src[0]^2, src[1]^2, ..., +C src[size-1]^2. dst[0] hasn't yet been set at all yet, and just gets the +C low limb of src[0]^2. + + movl (%esi), %eax C src[0] + + mull %eax + + leal (%esi,%ecx,4), %esi C src point just after last limb + negl %ecx + + movl %eax, (%edi,%ecx,8) C dst[0] + incl %ecx + +L(diag): + C eax scratch + C ebx scratch + C ecx counter, negative + C edx carry + C esi src just after last limb + C edi dst just after last limb + C ebp + + movl (%esi,%ecx,4), %eax + movl %edx, %ebx + + mull %eax + + addl %ebx, -4(%edi,%ecx,8) + adcl %eax, (%edi,%ecx,8) + adcl $0, %edx + + incl %ecx + jnz L(diag) + + + movl SAVE_ESI, %esi + movl SAVE_EBX, %ebx + + addl %edx, -4(%edi) C dst most significant limb + movl SAVE_EDI, %edi + + movl SAVE_EBP, %ebp + addl $FRAME, %esp + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/k7/sublsh1_n.asm b/gmp-6.3.0/mpn/x86/k7/sublsh1_n.asm new file mode 100644 index 0000000..8851683 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k7/sublsh1_n.asm @@ -0,0 +1,173 @@ +dnl AMD K7 mpn_sublsh1_n_ip1 -- rp[] = rp[] - (up[] << 1) + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl Contributed to the GNU project by Torbjorn Granlund and Marco Bodrato. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C This is an attempt at a sublsh1_n for x86-32, not relying on sse2 insns. The +C innerloop is 2*3-way unrolled, which is best we can do with the available +C registers. It seems tricky to use the same structure for rsblsh1_n, since we +C cannot feed carry between operations there. + +C cycles/limb +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 6.75 +C AMD K6 +C AMD K7 +C AMD K8 + +C This is a basic sublsh1_n for k7, atom, and perhaps some other x86-32 +C processors. It uses 2*4-way unrolling, for good reasons. +C +C Breaking carry recurrency might be a good idea. We would then need separate +C registers for the shift carry and add/subtract carry, which in turn would +C force us to 2*2-way unrolling. + +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(VAR_COUNT,`PARAM_SIZE') +define(SAVE_EBX,`PARAM_SRC') +define(SAVE_EBP,`PARAM_DST') + +ASM_START() + TEXT + ALIGN(8) +PROLOGUE(mpn_sublsh1_n_ip1) +deflit(`FRAME',0) + +define(`rp', `%edi') +define(`up', `%esi') + + mov PARAM_SIZE, %eax C size + push up FRAME_pushl() + push rp FRAME_pushl() + xor %edx, %edx + mov PARAM_SRC, up + mov PARAM_DST, rp + mov %ebx, SAVE_EBX + mov %eax, %ebx + shr $3, %eax + + not %eax C count = -(size\8)-i + and $7, %ebx C size % 8 + jz L(exact) + +L(oop): +ifdef(`CPU_P6',` + shr %edx ') C restore 2nd saved carry bit + mov (up), %ecx + adc %ecx, %ecx + rcr %edx C restore 1st saved carry bit + lea 4(up), up + sbb %ecx, (rp) + lea 4(rp), rp + adc %edx, %edx C save a carry bit in edx +ifdef(`CPU_P6',` + adc %edx, %edx ') C save another carry bit in edx + dec %ebx + jnz L(oop) +L(exact): + inc %eax + jz L(end) + mov %eax, VAR_COUNT + mov %ebp, SAVE_EBP + + ALIGN(16) +L(top): +ifdef(`CPU_P6',` + shr %edx ') C restore 2nd saved carry bit + mov (up), %eax + adc %eax, %eax + mov 4(up), %ebx + adc %ebx, %ebx + mov 8(up), %ecx + adc %ecx, %ecx + mov 12(up), %ebp + adc %ebp, %ebp + + rcr %edx C restore 1st saved carry bit + + sbb %eax, (rp) + sbb %ebx, 4(rp) + sbb %ecx, 8(rp) + sbb %ebp, 12(rp) + + mov 16(up), %eax + adc %eax, %eax + mov 20(up), %ebx + adc %ebx, %ebx + mov 24(up), %ecx + adc %ecx, %ecx + mov 28(up), %ebp + adc %ebp, %ebp + + lea 32(up), up + adc %edx, %edx C save a carry bit in edx + + sbb %eax, 16(rp) + sbb %ebx, 20(rp) + sbb %ecx, 24(rp) + sbb %ebp, 28(rp) + +ifdef(`CPU_P6',` + adc %edx, %edx ') C save another carry bit in edx + incl VAR_COUNT + lea 32(rp), rp + jne L(top) + + mov SAVE_EBP, %ebp +L(end): + mov SAVE_EBX, %ebx + +ifdef(`CPU_P6',` + xor %eax, %eax + shr $1, %edx + adc %edx, %eax +',` + adc $0, %edx + mov %edx, %eax +') + pop rp FRAME_popl() + pop up FRAME_popl() + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/k8/gmp-mparam.h b/gmp-6.3.0/mpn/x86/k8/gmp-mparam.h new file mode 100644 index 0000000..fa71292 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/k8/gmp-mparam.h @@ -0,0 +1,215 @@ +/* x86/k8 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991, 1993, 1994, 2000-2011, 2014 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 2500 MHz K8 Brisbane */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-20, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 0 /* always */ +#define MOD_1_UNNORM_THRESHOLD 0 /* always */ +#define MOD_1N_TO_MOD_1_1_THRESHOLD 11 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 4 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 12 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 21 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 36.85% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 3 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 44 + +#define DIV_1_VS_MUL_1_PERCENT 251 + +#define MUL_TOOM22_THRESHOLD 26 +#define MUL_TOOM33_THRESHOLD 78 +#define MUL_TOOM44_THRESHOLD 136 +#define MUL_TOOM6H_THRESHOLD 270 +#define MUL_TOOM8H_THRESHOLD 430 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 85 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 91 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 89 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 96 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 121 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 46 +#define SQR_TOOM3_THRESHOLD 81 +#define SQR_TOOM4_THRESHOLD 202 +#define SQR_TOOM6_THRESHOLD 300 +#define SQR_TOOM8_THRESHOLD 430 + +#define MULMID_TOOM42_THRESHOLD 50 + +#define MULMOD_BNM1_THRESHOLD 18 +#define SQRMOD_BNM1_THRESHOLD 22 + +#define MUL_FFT_MODF_THRESHOLD 606 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 606, 5}, { 27, 6}, { 15, 5}, { 31, 6}, \ + { 25, 7}, { 13, 6}, { 29, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 23, 6}, { 47, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 31, 7}, { 63, 8}, { 39, 9}, { 23, 8}, \ + { 55, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 79, 9}, { 47, 8}, { 95, 9}, { 55,10}, \ + { 31, 9}, { 79,10}, { 47, 9}, { 103,11}, \ + { 31,10}, { 63, 9}, { 135,10}, { 79, 9}, \ + { 167,10}, { 95, 9}, { 191,10}, { 111,11}, \ + { 63,10}, { 159,11}, { 95,10}, { 191,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 511,10}, \ + { 271, 9}, { 543,10}, { 287,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335, 9}, { 671,11}, \ + { 191,10}, { 383, 9}, { 767,10}, { 399, 9}, \ + { 799,11}, { 223,12}, { 127,11}, { 255,10}, \ + { 511, 9}, { 1023,10}, { 543,11}, { 287,10}, \ + { 607,11}, { 319,10}, { 671,11}, { 351,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,13}, \ + { 127,12}, { 255,11}, { 511,10}, { 1023,11}, \ + { 543,10}, { 1087,11}, { 607,12}, { 319,11}, \ + { 671,10}, { 1343,11}, { 735,10}, { 1471,12}, \ + { 383,11}, { 799,10}, { 1599,11}, { 863,12}, \ + { 447,11}, { 927,13}, { 255,12}, { 511,11}, \ + { 1087,12}, { 575,11}, { 1215,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,12}, \ + { 895,11}, { 1791,12}, { 959,14}, { 255,13}, \ + { 511,12}, { 1087,11}, { 2239,12}, { 1215,13}, \ + { 639,12}, { 1471,13}, { 767,12}, { 1727,13}, \ + { 895,12}, { 1919,14}, { 511,13}, { 1023,12}, \ + { 2239,13}, { 1151,12}, { 2431,13}, { 1279,12}, \ + { 2623,13}, { 1407,12}, { 2943,14}, { 767,13}, \ + { 1663,12}, { 3455,13}, { 1919,15}, { 511,14}, \ + { 1023,13}, { 2175,12}, { 4351,13}, { 2431,14}, \ + { 1279,13}, { 2943,14}, { 1535,13}, { 3455,14}, \ + { 1791,13}, { 3967,15}, { 1023,14}, { 2047,13}, \ + { 4351,14}, { 2303,13}, { 4991,14}, { 2815,15}, \ + { 1535,14}, { 3839,16} } +#define MUL_FFT_TABLE3_SIZE 158 +#define MUL_FFT_THRESHOLD 7296 + +#define SQR_FFT_MODF_THRESHOLD 500 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 500, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 25, 7}, { 13, 6}, { 29, 7}, { 15, 6}, \ + { 32, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 29, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 51, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 79, 9}, { 47, 8}, \ + { 95,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 135,10}, \ + { 79, 9}, { 159,10}, { 95, 9}, { 191,10}, \ + { 111,11}, { 63,10}, { 127, 9}, { 255,10}, \ + { 143, 9}, { 287, 8}, { 575,10}, { 159,11}, \ + { 95,10}, { 191,12}, { 63,11}, { 127,10}, \ + { 255, 9}, { 511,10}, { 271, 9}, { 543,10}, \ + { 287,11}, { 159,10}, { 319, 9}, { 639,10}, \ + { 335, 9}, { 671,10}, { 351,11}, { 191,10}, \ + { 383, 9}, { 767,10}, { 399, 9}, { 799,10}, \ + { 415,11}, { 223,12}, { 127,11}, { 255,10}, \ + { 511, 9}, { 1023,10}, { 543,11}, { 287,10}, \ + { 607, 9}, { 1215,11}, { 319,10}, { 671,11}, \ + { 351,10}, { 703,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,10}, { 831,13}, { 127,12}, \ + { 255,11}, { 511,10}, { 1023,11}, { 543,10}, \ + { 1087,11}, { 607,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,10}, { 1471,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,12}, { 447,11}, \ + { 959,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,10}, { 2431,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,12}, \ + { 959,14}, { 255,13}, { 511,12}, { 1087,11}, \ + { 2239,12}, { 1215,11}, { 2431,13}, { 639,12}, \ + { 1471,13}, { 767,12}, { 1727,13}, { 895,12}, \ + { 1919,14}, { 511,13}, { 1023,12}, { 2239,13}, \ + { 1151,12}, { 2431,13}, { 1279,12}, { 2623,13}, \ + { 1407,12}, { 2943,14}, { 767,13}, { 1663,12}, \ + { 3455,13}, { 1919,15}, { 511,14}, { 1023,13}, \ + { 2431,14}, { 1279,13}, { 2943,14}, { 1535,13}, \ + { 3455,14}, { 1791,13}, { 3839,15}, { 1023,14}, \ + { 2047,13}, { 4223,14}, { 2303,13}, { 4863,14}, \ + { 2815,15}, { 1535,14}, { 3839,16} } +#define SQR_FFT_TABLE3_SIZE 167 +#define SQR_FFT_THRESHOLD 5504 + +#define MULLO_BASECASE_THRESHOLD 4 +#define MULLO_DC_THRESHOLD 29 +#define MULLO_MUL_N_THRESHOLD 14281 +#define SQRLO_BASECASE_THRESHOLD 6 +#define SQRLO_DC_THRESHOLD 193 +#define SQRLO_SQR_THRESHOLD 10704 + +#define DC_DIV_QR_THRESHOLD 84 +#define DC_DIVAPPR_Q_THRESHOLD 278 +#define DC_BDIV_QR_THRESHOLD 87 +#define DC_BDIV_Q_THRESHOLD 216 + +#define INV_MULMOD_BNM1_THRESHOLD 50 +#define INV_NEWTON_THRESHOLD 268 +#define INV_APPR_THRESHOLD 268 + +#define BINV_NEWTON_THRESHOLD 276 +#define REDC_1_TO_REDC_N_THRESHOLD 78 + +#define MU_DIV_QR_THRESHOLD 1652 +#define MU_DIVAPPR_Q_THRESHOLD 1528 +#define MUPI_DIV_QR_THRESHOLD 114 +#define MU_BDIV_QR_THRESHOLD 1442 +#define MU_BDIV_Q_THRESHOLD 1466 + +#define POWM_SEC_TABLE 1,22,102,452,1357 + +#define GET_STR_DC_THRESHOLD 14 +#define GET_STR_PRECOMPUTE_THRESHOLD 24 +#define SET_STR_DC_THRESHOLD 270 +#define SET_STR_PRECOMPUTE_THRESHOLD 1149 + +#define FAC_DSC_THRESHOLD 208 +#define FAC_ODD_THRESHOLD 48 + +#define MATRIX22_STRASSEN_THRESHOLD 16 +#define HGCD2_DIV1_METHOD 3 /* 4.69% faster than 1 */ +#define HGCD_THRESHOLD 139 +#define HGCD_APPR_THRESHOLD 174 +#define HGCD_REDUCE_THRESHOLD 3389 +#define GCD_DC_THRESHOLD 599 +#define GCDEXT_DC_THRESHOLD 419 +#define JACOBI_BASE_METHOD 1 /* 1.57% faster than 4 */ + +/* Tuneup completed successfully, took 83851 seconds */ diff --git a/gmp-6.3.0/mpn/x86/lshift.asm b/gmp-6.3.0/mpn/x86/lshift.asm new file mode 100644 index 0000000..6ee6153 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/lshift.asm @@ -0,0 +1,106 @@ +dnl x86 mpn_lshift -- mpn left shift. + +dnl Copyright 1992, 1994, 1996, 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P54 7.5 +C P55 7.0 +C P6 2.5 +C K6 4.5 +C K7 5.0 +C P4 14.5 + + +C mp_limb_t mpn_lshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned shift); + +defframe(PARAM_SHIFT,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_lshift) + + pushl %edi + pushl %esi + pushl %ebx +deflit(`FRAME',12) + + movl PARAM_DST,%edi + movl PARAM_SRC,%esi + movl PARAM_SIZE,%edx + movl PARAM_SHIFT,%ecx + + subl $4,%esi C adjust src + + movl (%esi,%edx,4),%ebx C read most significant limb + xorl %eax,%eax + shldl( %cl, %ebx, %eax) C compute carry limb + decl %edx + jz L(end) + pushl %eax C push carry limb onto stack + testb $1,%dl + jnz L(1) C enter loop in the middle + movl %ebx,%eax + + ALIGN(8) +L(oop): movl (%esi,%edx,4),%ebx C load next lower limb + shldl( %cl, %ebx, %eax) C compute result limb + movl %eax,(%edi,%edx,4) C store it + decl %edx +L(1): movl (%esi,%edx,4),%eax + shldl( %cl, %eax, %ebx) + movl %ebx,(%edi,%edx,4) + decl %edx + jnz L(oop) + + shll %cl,%eax C compute least significant limb + movl %eax,(%edi) C store it + + popl %eax C pop carry limb + + popl %ebx + popl %esi + popl %edi + ret + +L(end): shll %cl,%ebx C compute least significant limb + movl %ebx,(%edi) C store it + + popl %ebx + popl %esi + popl %edi + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/mmx/sec_tabselect.asm b/gmp-6.3.0/mpn/x86/mmx/sec_tabselect.asm new file mode 100644 index 0000000..aae158a --- /dev/null +++ b/gmp-6.3.0/mpn/x86/mmx/sec_tabselect.asm @@ -0,0 +1,163 @@ +dnl X86 MMX mpn_sec_tabselect. + +dnl Contributed to the GNU project by Torbjörn Granlund. + +dnl Copyright 2011-2013 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb cycles/limb +C ali,evn n unal,evn n +C P5 +C P6 model 0-8,10-12 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) 1.33 1.87 +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) 2.1 2.63 +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) 1.7 2.57 +C Intel Atom 1.85 2.7 +C AMD K6 +C AMD K7 1.33 1.33 +C AMD K8 +C AMD K10 + +define(`rp', `%edi') +define(`tp', `%esi') +define(`n', `%edx') +define(`nents', `%ecx') +define(`which', `') + +define(`i', `%ebp') +define(`j', `%ebx') + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(mpn_sec_tabselect) + push %ebx + push %esi + push %edi + push %ebp + + mov 20(%esp), rp + mov 24(%esp), tp + mov 28(%esp), n + mov 32(%esp), nents + + movd 36(%esp), %mm6 + punpckldq %mm6, %mm6 C 2 copies of `which' + + mov $1, %ebx + movd %ebx, %mm7 + punpckldq %mm7, %mm7 C 2 copies of 1 + + mov n, j + add $-4, j + js L(outer_end) + +L(outer_top): + mov nents, i + mov tp, %eax + pxor %mm1, %mm1 + pxor %mm4, %mm4 + pxor %mm5, %mm5 + ALIGN(16) +L(top): movq %mm6, %mm0 + pcmpeqd %mm1, %mm0 + paddd %mm7, %mm1 + movq (tp), %mm2 + movq 8(tp), %mm3 + pand %mm0, %mm2 + pand %mm0, %mm3 + por %mm2, %mm4 + por %mm3, %mm5 + lea (tp,n,4), tp + add $-1, i + jne L(top) + + movq %mm4, (rp) + movq %mm5, 8(rp) + + lea 16(%eax), tp + lea 16(rp), rp + add $-4, j + jns L(outer_top) +L(outer_end): + + test $2, %dl + jz L(b0x) + +L(b1x): mov nents, i + mov tp, %eax + pxor %mm1, %mm1 + pxor %mm4, %mm4 + ALIGN(16) +L(tp2): movq %mm6, %mm0 + pcmpeqd %mm1, %mm0 + paddd %mm7, %mm1 + movq (tp), %mm2 + pand %mm0, %mm2 + por %mm2, %mm4 + lea (tp,n,4), tp + add $-1, i + jne L(tp2) + + movq %mm4, (rp) + + lea 8(%eax), tp + lea 8(rp), rp + +L(b0x): test $1, %dl + jz L(b00) + +L(b01): mov nents, i + pxor %mm1, %mm1 + pxor %mm4, %mm4 + ALIGN(16) +L(tp1): movq %mm6, %mm0 + pcmpeqd %mm1, %mm0 + paddd %mm7, %mm1 + movd (tp), %mm2 + pand %mm0, %mm2 + por %mm2, %mm4 + lea (tp,n,4), tp + add $-1, i + jne L(tp1) + + movd %mm4, (rp) + +L(b00): pop %ebp + pop %edi + pop %esi + pop %ebx + emms + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/mod_34lsub1.asm b/gmp-6.3.0/mpn/x86/mod_34lsub1.asm new file mode 100644 index 0000000..e09e702 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/mod_34lsub1.asm @@ -0,0 +1,183 @@ +dnl Generic x86 mpn_mod_34lsub1 -- mpn remainder modulo 2^24-1. + +dnl Copyright 2000-2002, 2004 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P5 3.0 +C P6 3.66 +C K6 3.0 +C K7 1.3 +C P4 9 + + +C mp_limb_t mpn_mod_34lsub1 (mp_srcptr src, mp_size_t size) +C + +defframe(PARAM_SIZE, 8) +defframe(PARAM_SRC, 4) + +dnl re-use parameter space +define(SAVE_EBX, `PARAM_SRC') + + TEXT + ALIGN(16) +PROLOGUE(mpn_mod_34lsub1) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + movl PARAM_SRC, %edx + + subl $2, %ecx + ja L(three_or_more) + + movl (%edx), %eax + jb L(one) + + movl 4(%edx), %ecx + movl %eax, %edx + shrl $24, %eax C src[0] low + + andl $0xFFFFFF, %edx C src[0] high + addl %edx, %eax + movl %ecx, %edx + + andl $0xFFFF, %ecx + shrl $16, %edx C src[1] high + addl %edx, %eax + + shll $8, %ecx C src[1] low + addl %ecx, %eax + +L(one): + ret + + +L(three_or_more): + C eax + C ebx + C ecx size-2 + C edx src + C esi + C edi + C ebp + + movl %ebx, SAVE_EBX C and arrange 16-byte loop alignment + xorl %ebx, %ebx + + pushl %esi FRAME_pushl() + xorl %esi, %esi + + pushl %edi FRAME_pushl() + xorl %eax, %eax C and clear carry flag + + + C offset 0x40 here +L(top): + C eax acc 0mod3 + C ebx acc 1mod3 + C ecx counter, limbs + C edx src + C esi acc 2mod3 + C edi + C ebp + + leal 12(%edx), %edx + leal -2(%ecx), %ecx + + adcl -12(%edx), %eax + adcl -8(%edx), %ebx + adcl -4(%edx), %esi + + decl %ecx + jg L(top) + + + C ecx is -2, -1 or 0 representing 0, 1 or 2 more limbs, respectively + + movl $0xFFFFFFFF, %edi + incl %ecx + js L(combine) + + adcl (%edx), %eax + movl $0xFFFFFF00, %edi + decl %ecx + js L(combine) + + adcl 4(%edx), %ebx + movl $0xFFFF0000, %edi + + +L(combine): + C eax acc 0mod3 + C ebx acc 1mod3 + C ecx + C edx + C esi acc 2mod3 + C edi mask + C ebp + + sbbl %ecx, %ecx C carry + movl %eax, %edx C 0mod3 + + shrl $24, %eax C 0mod3 high + andl %edi, %ecx C carry masked + + subl %ecx, %eax C apply carry + movl %ebx, %edi C 1mod3 + + shrl $16, %ebx C 1mod3 high + andl $0x00FFFFFF, %edx C 0mod3 low + + addl %edx, %eax C apply 0mod3 low + andl $0xFFFF, %edi + + shll $8, %edi C 1mod3 low + addl %ebx, %eax C apply 1mod3 high + + addl %edi, %eax C apply 1mod3 low + movl %esi, %edx C 2mod3 + + shrl $8, %esi C 2mod3 high + andl $0xFF, %edx C 2mod3 low + + shll $16, %edx C 2mod3 low + addl %esi, %eax C apply 2mod3 high + + addl %edx, %eax C apply 2mod3 low + popl %edi FRAME_popl() + + movl SAVE_EBX, %ebx + popl %esi FRAME_popl() + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/mul_1.asm b/gmp-6.3.0/mpn/x86/mul_1.asm new file mode 100644 index 0000000..421de62 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/mul_1.asm @@ -0,0 +1,140 @@ +dnl x86 mpn_mul_1 (for 386, 486, and Pentium Pro) -- Multiply a limb vector +dnl with a limb and store the result in a second limb vector. + +dnl Copyright 1992, 1994, 1997-2002, 2005 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P5 12.5 +C P6 model 0-8,10-12 5.5 +C P6 model 9 (Banias) +C P6 model 13 (Dothan) 5.25 +C P4 model 0 (Willamette) 19.0 +C P4 model 1 (?) 19.0 +C P4 model 2 (Northwood) 19.0 +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C AMD K6 10.5 +C AMD K7 4.5 +C AMD K8 + + +C mp_limb_t mpn_mul_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t multiplier); + +defframe(PARAM_MULTIPLIER,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_mul_1) +deflit(`FRAME',0) + + pushl %edi + pushl %esi + pushl %ebx + pushl %ebp +deflit(`FRAME',16) + + movl PARAM_DST,%edi + movl PARAM_SRC,%esi + movl PARAM_SIZE,%ecx + + xorl %ebx,%ebx + andl $3,%ecx + jz L(end0) + +L(oop0): + movl (%esi),%eax + mull PARAM_MULTIPLIER + leal 4(%esi),%esi + addl %ebx,%eax + movl $0,%ebx + adcl %ebx,%edx + movl %eax,(%edi) + movl %edx,%ebx C propagate carry into cylimb + + leal 4(%edi),%edi + decl %ecx + jnz L(oop0) + +L(end0): + movl PARAM_SIZE,%ecx + shrl $2,%ecx + jz L(end) + + + ALIGN(8) +L(oop): movl (%esi),%eax + mull PARAM_MULTIPLIER + addl %eax,%ebx + movl $0,%ebp + adcl %edx,%ebp + + movl 4(%esi),%eax + mull PARAM_MULTIPLIER + movl %ebx,(%edi) + addl %eax,%ebp C new lo + cylimb + movl $0,%ebx + adcl %edx,%ebx + + movl 8(%esi),%eax + mull PARAM_MULTIPLIER + movl %ebp,4(%edi) + addl %eax,%ebx C new lo + cylimb + movl $0,%ebp + adcl %edx,%ebp + + movl 12(%esi),%eax + mull PARAM_MULTIPLIER + movl %ebx,8(%edi) + addl %eax,%ebp C new lo + cylimb + movl $0,%ebx + adcl %edx,%ebx + + movl %ebp,12(%edi) + + leal 16(%esi),%esi + leal 16(%edi),%edi + decl %ecx + jnz L(oop) + +L(end): movl %ebx,%eax + + popl %ebp + popl %ebx + popl %esi + popl %edi + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/mul_basecase.asm b/gmp-6.3.0/mpn/x86/mul_basecase.asm new file mode 100644 index 0000000..8339732 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/mul_basecase.asm @@ -0,0 +1,223 @@ +dnl x86 mpn_mul_basecase -- Multiply two limb vectors and store the result +dnl in a third limb vector. + +dnl Copyright 1996-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/crossproduct +C P5 15 +C P6 7.5 +C K6 12.5 +C K7 5.5 +C P4 24 + + +C void mpn_mul_basecase (mp_ptr wp, +C mp_srcptr xp, mp_size_t xsize, +C mp_srcptr yp, mp_size_t ysize); +C +C This was written in a haste since the Pentium optimized code that was used +C for all x86 machines was slow for the Pentium II. This code would benefit +C from some cleanup. +C +C To shave off some percentage of the run-time, one should make 4 variants +C of the Louter loop, for the four different outcomes of un mod 4. That +C would avoid Loop0 altogether. Code expansion would be > 4-fold for that +C part of the function, but since it is not very large, that would be +C acceptable. +C +C The mul loop (at L(oopM)) might need some tweaking. It's current speed is +C unknown. + +defframe(PARAM_YSIZE,20) +defframe(PARAM_YP, 16) +defframe(PARAM_XSIZE,12) +defframe(PARAM_XP, 8) +defframe(PARAM_WP, 4) + +defframe(VAR_MULTIPLIER, -4) +defframe(VAR_COUNTER, -8) +deflit(VAR_STACK_SPACE, 8) + + TEXT + ALIGN(8) + +PROLOGUE(mpn_mul_basecase) +deflit(`FRAME',0) + + subl $VAR_STACK_SPACE,%esp + pushl %esi + pushl %ebp + pushl %edi +deflit(`FRAME',eval(VAR_STACK_SPACE+12)) + + movl PARAM_XP,%esi + movl PARAM_WP,%edi + movl PARAM_YP,%ebp + + movl (%esi),%eax C load xp[0] + mull (%ebp) C multiply by yp[0] + movl %eax,(%edi) C store to wp[0] + movl PARAM_XSIZE,%ecx C xsize + decl %ecx C If xsize = 1, ysize = 1 too + jz L(done) + + pushl %ebx +FRAME_pushl() + movl %edx,%ebx + + leal 4(%esi),%esi + leal 4(%edi),%edi + +L(oopM): + movl (%esi),%eax C load next limb at xp[j] + leal 4(%esi),%esi + mull (%ebp) + addl %ebx,%eax + movl %edx,%ebx + adcl $0,%ebx + movl %eax,(%edi) + leal 4(%edi),%edi + decl %ecx + jnz L(oopM) + + movl %ebx,(%edi) C most significant limb of product + addl $4,%edi C increment wp + movl PARAM_XSIZE,%eax + shll $2,%eax + subl %eax,%edi + subl %eax,%esi + + movl PARAM_YSIZE,%eax C ysize + decl %eax + jz L(skip) + movl %eax,VAR_COUNTER C set index i to ysize + +L(outer): + movl PARAM_YP,%ebp C yp + addl $4,%ebp C make ebp point to next v limb + movl %ebp,PARAM_YP + movl (%ebp),%eax C copy y limb ... + movl %eax,VAR_MULTIPLIER C ... to stack slot + movl PARAM_XSIZE,%ecx + + xorl %ebx,%ebx + andl $3,%ecx + jz L(end0) + +L(oop0): + movl (%esi),%eax + mull VAR_MULTIPLIER + leal 4(%esi),%esi + addl %ebx,%eax + movl $0,%ebx + adcl %ebx,%edx + addl %eax,(%edi) + adcl %edx,%ebx C propagate carry into cylimb + + leal 4(%edi),%edi + decl %ecx + jnz L(oop0) + +L(end0): + movl PARAM_XSIZE,%ecx + shrl $2,%ecx + jz L(endX) + + ALIGN(8) +L(oopX): + movl (%esi),%eax + mull VAR_MULTIPLIER + addl %eax,%ebx + movl $0,%ebp + adcl %edx,%ebp + + movl 4(%esi),%eax + mull VAR_MULTIPLIER + addl %ebx,(%edi) + adcl %eax,%ebp C new lo + cylimb + movl $0,%ebx + adcl %edx,%ebx + + movl 8(%esi),%eax + mull VAR_MULTIPLIER + addl %ebp,4(%edi) + adcl %eax,%ebx C new lo + cylimb + movl $0,%ebp + adcl %edx,%ebp + + movl 12(%esi),%eax + mull VAR_MULTIPLIER + addl %ebx,8(%edi) + adcl %eax,%ebp C new lo + cylimb + movl $0,%ebx + adcl %edx,%ebx + + addl %ebp,12(%edi) + adcl $0,%ebx C propagate carry into cylimb + + leal 16(%esi),%esi + leal 16(%edi),%edi + decl %ecx + jnz L(oopX) + +L(endX): + movl %ebx,(%edi) + addl $4,%edi + + C we incremented wp and xp in the loop above; compensate + movl PARAM_XSIZE,%eax + shll $2,%eax + subl %eax,%edi + subl %eax,%esi + + movl VAR_COUNTER,%eax + decl %eax + movl %eax,VAR_COUNTER + jnz L(outer) + +L(skip): + popl %ebx + popl %edi + popl %ebp + popl %esi + addl $8,%esp + ret + +L(done): + movl %edx,4(%edi) C store to wp[1] + popl %edi + popl %ebp + popl %esi + addl $8,%esp + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/nano/gmp-mparam.h b/gmp-6.3.0/mpn/x86/nano/gmp-mparam.h new file mode 100644 index 0000000..cd8ac4e --- /dev/null +++ b/gmp-6.3.0/mpn/x86/nano/gmp-mparam.h @@ -0,0 +1,162 @@ +/* x86/nano gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991, 1993, 1994, 2000-2011 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* Generated by tuneup.c, 2011-11-25, gcc 4.2 */ + +#define MOD_1_1P_METHOD 1 +#define MOD_1_NORM_THRESHOLD 3 +#define MOD_1_UNNORM_THRESHOLD 3 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 10 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 9 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 0 /* never mpn_mod_1_1p */ +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 53 +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 12 +#define USE_PREINV_DIVREM_1 1 +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 32 + +#define MUL_TOOM22_THRESHOLD 16 +#define MUL_TOOM33_THRESHOLD 132 +#define MUL_TOOM44_THRESHOLD 195 +#define MUL_TOOM6H_THRESHOLD 270 +#define MUL_TOOM8H_THRESHOLD 478 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 129 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 138 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 130 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 135 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 28 +#define SQR_TOOM3_THRESHOLD 194 +#define SQR_TOOM4_THRESHOLD 502 +#define SQR_TOOM6_THRESHOLD 746 +#define SQR_TOOM8_THRESHOLD 1005 + +#define MULMID_TOOM42_THRESHOLD 40 + +#define MULMOD_BNM1_THRESHOLD 14 +#define SQRMOD_BNM1_THRESHOLD 19 + +#define POWM_SEC_TABLE 4,23,258,828,2246 + +#define MUL_FFT_MODF_THRESHOLD 308 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 308, 5}, { 13, 6}, { 7, 5}, { 17, 6}, \ + { 9, 5}, { 19, 6}, { 11, 5}, { 23, 6}, \ + { 13, 7}, { 7, 6}, { 17, 7}, { 9, 6}, \ + { 19, 7}, { 11, 6}, { 24, 7}, { 15, 6}, \ + { 31, 7}, { 19, 8}, { 11, 7}, { 25, 8}, \ + { 15, 7}, { 33, 8}, { 19, 7}, { 39, 8}, \ + { 23, 7}, { 47, 9}, { 15, 8}, { 31, 7}, \ + { 63, 8}, { 39, 9}, { 23, 8}, { 47,10}, \ + { 15, 9}, { 31, 8}, { 63, 9}, { 47,10}, \ + { 31, 9}, { 71,10}, { 47, 9}, { 95,11}, \ + { 31,10}, { 63, 9}, { 127, 8}, { 255,10}, \ + { 79, 9}, { 159,10}, { 95, 9}, { 191,11}, \ + { 63,10}, { 127, 9}, { 255, 8}, { 543, 9}, \ + { 287, 8}, { 575, 7}, { 1215,10}, { 159,11}, \ + { 95,10}, { 191,12}, { 63,11}, { 127,10}, \ + { 255, 9}, { 543, 8}, { 1087,10}, { 287, 9}, \ + { 607, 8}, { 1215,11}, { 159,10}, { 319, 9}, \ + { 639,10}, { 351, 9}, { 703, 8}, { 1407, 9}, \ + { 735, 8}, { 1471,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 415, 9}, { 831,11}, { 223,10}, \ + { 447, 9}, { 895,10}, { 479, 9}, { 959, 8}, \ + { 1919,12}, { 4096,13}, { 8192,14}, { 16384,15}, \ + { 32768,16} } +#define MUL_FFT_TABLE3_SIZE 89 +#define MUL_FFT_THRESHOLD 1856 + +#define SQR_FFT_MODF_THRESHOLD 396 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 396, 5}, { 13, 6}, { 7, 5}, { 21, 6}, \ + { 11, 5}, { 23, 6}, { 21, 7}, { 11, 6}, \ + { 25, 7}, { 15, 6}, { 31, 7}, { 19, 6}, \ + { 39, 7}, { 21, 8}, { 11, 7}, { 23, 6}, \ + { 47, 7}, { 27, 8}, { 15, 7}, { 33, 8}, \ + { 19, 7}, { 39, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 47,10}, { 15, 9}, \ + { 31, 8}, { 63, 9}, { 39, 8}, { 79, 9}, \ + { 47,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 127,10}, \ + { 79, 9}, { 159,10}, { 95,11}, { 63,10}, \ + { 127, 9}, { 255, 8}, { 543,10}, { 143, 9}, \ + { 287, 8}, { 607, 7}, { 1215, 6}, { 2431,10}, \ + { 159, 8}, { 639,11}, { 95,10}, { 191,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 543, 8}, \ + { 1087,10}, { 287, 9}, { 607, 8}, { 1215,11}, \ + { 159,10}, { 319, 9}, { 671,10}, { 351, 9}, \ + { 703, 8}, { 1407, 9}, { 735, 8}, { 1471, 7}, \ + { 2943,11}, { 191,10}, { 383, 9}, { 799,10}, \ + { 415, 9}, { 895,10}, { 479,12}, { 4096,13}, \ + { 8192,14}, { 16384,15}, { 32768,16} } +#define SQR_FFT_TABLE3_SIZE 87 +#define SQR_FFT_THRESHOLD 2368 + +#define MULLO_BASECASE_THRESHOLD 0 /* always */ +#define MULLO_DC_THRESHOLD 51 +#define MULLO_MUL_N_THRESHOLD 3369 + +#define DC_DIV_QR_THRESHOLD 56 +#define DC_DIVAPPR_Q_THRESHOLD 183 +#define DC_BDIV_QR_THRESHOLD 55 +#define DC_BDIV_Q_THRESHOLD 118 + +#define INV_MULMOD_BNM1_THRESHOLD 30 +#define INV_NEWTON_THRESHOLD 266 +#define INV_APPR_THRESHOLD 218 + +#define BINV_NEWTON_THRESHOLD 268 +#define REDC_1_TO_REDC_N_THRESHOLD 56 + +#define MU_DIV_QR_THRESHOLD 1308 +#define MU_DIVAPPR_Q_THRESHOLD 1528 +#define MUPI_DIV_QR_THRESHOLD 124 +#define MU_BDIV_QR_THRESHOLD 855 +#define MU_BDIV_Q_THRESHOLD 1334 + +#define MATRIX22_STRASSEN_THRESHOLD 14 +#define HGCD_THRESHOLD 104 +#define HGCD_APPR_THRESHOLD 139 +#define HGCD_REDUCE_THRESHOLD 2121 +#define GCD_DC_THRESHOLD 456 +#define GCDEXT_DC_THRESHOLD 321 +#define JACOBI_BASE_METHOD 4 + +#define GET_STR_DC_THRESHOLD 11 +#define GET_STR_PRECOMPUTE_THRESHOLD 25 +#define SET_STR_DC_THRESHOLD 542 +#define SET_STR_PRECOMPUTE_THRESHOLD 840 diff --git a/gmp-6.3.0/mpn/x86/p6/README b/gmp-6.3.0/mpn/x86/p6/README new file mode 100644 index 0000000..f19d47b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/README @@ -0,0 +1,125 @@ +Copyright 2000, 2001 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. + + + + + + INTEL P6 MPN SUBROUTINES + + + +This directory contains code optimized for Intel P6 class CPUs, meaning +PentiumPro, Pentium II and Pentium III. The mmx and p3mmx subdirectories +have routines using MMX instructions. + + + +STATUS + +Times for the loops, with all code and data in L1 cache, are as follows. +Some of these might be able to be improved. + + cycles/limb + + mpn_add_n/sub_n 3.7 + + mpn_copyi 0.75 + mpn_copyd 1.75 (or 0.75 if no overlap) + + mpn_divrem_1 39.0 + mpn_mod_1 21.5 + mpn_divexact_by3 8.5 + + mpn_mul_1 5.5 + mpn_addmul/submul_1 6.35 + + mpn_l/rshift 2.5 + + mpn_mul_basecase 8.2 cycles/crossproduct (approx) + mpn_sqr_basecase 4.0 cycles/crossproduct (approx) + or 7.75 cycles/triangleproduct (approx) + +Pentium II and III have MMX and get the following improvements. + + mpn_divrem_1 25.0 integer part, 17.5 fractional part + + mpn_l/rshift 1.75 + + + + +NOTES + +Write-allocate L1 data cache means prefetching of destinations is unnecessary. + +Mispredicted branches have a penalty of between 9 and 15 cycles, and even up +to 26 cycles depending how far speculative execution has gone. The 9 cycle +minimum penalty comes from the issue pipeline being 9 stages. + +A copy with rep movs seems to copy 16 bytes at a time, since speeds for 4, +5, 6 or 7 limb operations are all the same. The 0.75 cycles/limb would be 3 +cycles per 16 byte block. + + + + +CODING + +Instructions in general code have been shown grouped if they can execute +together, which means up to three instructions with no successive +dependencies, and with only the first being a multiple micro-op. + +P6 has out-of-order execution, so the groupings are really only showing +dependent paths where some shuffling might allow some latencies to be +hidden. + + + + +REFERENCES + +"Intel Architecture Optimization Reference Manual", 1999, revision 001 dated +02/99, order number 245127 (order number 730795-001 is in the document too). +Available on-line: + + http://download.intel.com/design/PentiumII/manuals/245127.htm + +"Intel Architecture Optimization Manual", 1997, order number 242816. This +is an older document mostly about P5 and not as good as the above. +Available on-line: + + http://download.intel.com/design/PentiumII/manuals/242816.htm + + + +---------------- +Local variables: +mode: text +fill-column: 76 +End: diff --git a/gmp-6.3.0/mpn/x86/p6/aors_n.asm b/gmp-6.3.0/mpn/x86/p6/aors_n.asm new file mode 100644 index 0000000..df51c2e --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/aors_n.asm @@ -0,0 +1,156 @@ +dnl Intel P6 mpn_add_n/mpn_sub_n -- mpn add or subtract. + +dnl Copyright 2006 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C TODO: +C * Avoid indexed addressing, it makes us stall on the two-ported register +C file. + +C cycles/limb +C P6 model 0-8,10-12 3.17 +C P6 model 9 (Banias) 2.15 +C P6 model 13 (Dothan) 2.25 + + +define(`rp', `%edi') +define(`up', `%esi') +define(`vp', `%ebx') +define(`n', `%ecx') + +ifdef(`OPERATION_add_n', ` + define(ADCSBB, adc) + define(func, mpn_add_n) + define(func_nc, mpn_add_nc)') +ifdef(`OPERATION_sub_n', ` + define(ADCSBB, sbb) + define(func, mpn_sub_n) + define(func_nc, mpn_sub_nc)') + +MULFUNC_PROLOGUE(mpn_add_n mpn_add_nc mpn_sub_n mpn_sub_nc) + +ASM_START() + + TEXT + ALIGN(16) + +PROLOGUE(func) + xor %edx, %edx +L(start): + push %edi + push %esi + push %ebx + + mov 16(%esp), rp + mov 20(%esp), up + mov 24(%esp), vp + mov 28(%esp), n + + lea (up,n,4), up + lea (vp,n,4), vp + lea (rp,n,4), rp + + neg n + mov n, %eax + and $-8, n + and $7, %eax + shl $2, %eax C 4x +ifdef(`PIC',` + call L(pic_calc) +L(here): +',` + lea L(ent) (%eax,%eax,2), %eax C 12x +') + + shr %edx C set cy flag + jmp *%eax + +ifdef(`PIC',` +L(pic_calc): + C See mpn/x86/README about old gas bugs + lea (%eax,%eax,2), %eax + add $L(ent)-L(here), %eax + add (%esp), %eax + ret_internal +') + +L(end): + sbb %eax, %eax + neg %eax + pop %ebx + pop %esi + pop %edi + ret + + ALIGN(16) +L(top): + jecxz L(end) +L(ent): +Zdisp( mov, 0,(up,n,4), %eax) +Zdisp( ADCSBB, 0,(vp,n,4), %eax) +Zdisp( mov, %eax, 0,(rp,n,4)) + + mov 4(up,n,4), %edx + ADCSBB 4(vp,n,4), %edx + mov %edx, 4(rp,n,4) + + mov 8(up,n,4), %eax + ADCSBB 8(vp,n,4), %eax + mov %eax, 8(rp,n,4) + + mov 12(up,n,4), %edx + ADCSBB 12(vp,n,4), %edx + mov %edx, 12(rp,n,4) + + mov 16(up,n,4), %eax + ADCSBB 16(vp,n,4), %eax + mov %eax, 16(rp,n,4) + + mov 20(up,n,4), %edx + ADCSBB 20(vp,n,4), %edx + mov %edx, 20(rp,n,4) + + mov 24(up,n,4), %eax + ADCSBB 24(vp,n,4), %eax + mov %eax, 24(rp,n,4) + + mov 28(up,n,4), %edx + ADCSBB 28(vp,n,4), %edx + mov %edx, 28(rp,n,4) + + lea 8(n), n + jmp L(top) + +EPILOGUE() + +PROLOGUE(func_nc) + movl 20(%esp), %edx + jmp L(start) +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/p6/aorsmul_1.asm b/gmp-6.3.0/mpn/x86/p6/aorsmul_1.asm new file mode 100644 index 0000000..bc8c49c --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/aorsmul_1.asm @@ -0,0 +1,320 @@ +dnl Intel P6 mpn_addmul_1/mpn_submul_1 -- add or subtract mpn multiple. + +dnl Copyright 1999-2002, 2005 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P5 +C P6 model 0-8,10-12 6.44 +C P6 model 9 (Banias) 6.15 +C P6 model 13 (Dothan) 6.11 +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C AMD K6 +C AMD K7 +C AMD K8 + + +dnl P6 UNROLL_COUNT cycles/limb +dnl 8 6.7 +dnl 16 6.35 +dnl 32 6.3 +dnl 64 6.3 +dnl Maximum possible with the current code is 64. + +deflit(UNROLL_COUNT, 16) + + +ifdef(`OPERATION_addmul_1', ` + define(M4_inst, addl) + define(M4_function_1, mpn_addmul_1) + define(M4_function_1c, mpn_addmul_1c) + define(M4_description, add it to) + define(M4_desc_retval, carry) +',`ifdef(`OPERATION_submul_1', ` + define(M4_inst, subl) + define(M4_function_1, mpn_submul_1) + define(M4_function_1c, mpn_submul_1c) + define(M4_description, subtract it from) + define(M4_desc_retval, borrow) +',`m4_error(`Need OPERATION_addmul_1 or OPERATION_submul_1 +')')') + +MULFUNC_PROLOGUE(mpn_addmul_1 mpn_addmul_1c mpn_submul_1 mpn_submul_1c) + + +C mp_limb_t M4_function_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult); +C mp_limb_t M4_function_1c (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult, mp_limb_t carry); +C +C Calculate src,size multiplied by mult and M4_description dst,size. +C Return the M4_desc_retval limb from the top of the result. +C +C This code is pretty much the same as the K6 code. The unrolled loop is +C the same, but there's just a few scheduling tweaks in the setups and the +C simple loop. +C +C A number of variations have been tried for the unrolled loop, with one or +C two carries, and with loads scheduled earlier, but nothing faster than 6 +C cycles/limb has been found. + +ifdef(`PIC',` +deflit(UNROLL_THRESHOLD, 5) +',` +deflit(UNROLL_THRESHOLD, 5) +') + +defframe(PARAM_CARRY, 20) +defframe(PARAM_MULTIPLIER,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(32) + +PROLOGUE(M4_function_1c) + pushl %ebx +deflit(`FRAME',4) + movl PARAM_CARRY, %ebx + jmp L(start_nc) +EPILOGUE() + +PROLOGUE(M4_function_1) + push %ebx +deflit(`FRAME',4) + xorl %ebx, %ebx C initial carry + +L(start_nc): + movl PARAM_SIZE, %ecx + pushl %esi +deflit(`FRAME',8) + + movl PARAM_SRC, %esi + pushl %edi +deflit(`FRAME',12) + + movl PARAM_DST, %edi + pushl %ebp +deflit(`FRAME',16) + cmpl $UNROLL_THRESHOLD, %ecx + + movl PARAM_MULTIPLIER, %ebp + jae L(unroll) + + + C simple loop + C this is offset 0x22, so close enough to aligned +L(simple): + C eax scratch + C ebx carry + C ecx counter + C edx scratch + C esi src + C edi dst + C ebp multiplier + + movl (%esi), %eax + addl $4, %edi + + mull %ebp + + addl %ebx, %eax + adcl $0, %edx + + M4_inst %eax, -4(%edi) + movl %edx, %ebx + + adcl $0, %ebx + decl %ecx + + leal 4(%esi), %esi + jnz L(simple) + + + popl %ebp + popl %edi + + popl %esi + movl %ebx, %eax + + popl %ebx + ret + + + +C------------------------------------------------------------------------------ +C VAR_JUMP holds the computed jump temporarily because there's not enough +C registers when doing the mul for the initial two carry limbs. +C +C The add/adc for the initial carry in %ebx is necessary only for the +C mpn_add/submul_1c entry points. Duplicating the startup code to +C eliminate this for the plain mpn_add/submul_1 doesn't seem like a good +C idea. + +dnl overlapping with parameters already fetched +define(VAR_COUNTER,`PARAM_SIZE') +define(VAR_JUMP, `PARAM_DST') + + C this is offset 0x43, so close enough to aligned +L(unroll): + C eax + C ebx initial carry + C ecx size + C edx + C esi src + C edi dst + C ebp + + movl %ecx, %edx + decl %ecx + + subl $2, %edx + negl %ecx + + shrl $UNROLL_LOG2, %edx + andl $UNROLL_MASK, %ecx + + movl %edx, VAR_COUNTER + movl %ecx, %edx + + C 15 code bytes per limb +ifdef(`PIC',` + call L(pic_calc) +L(here): +',` + shll $4, %edx + negl %ecx + + leal L(entry) (%edx,%ecx,1), %edx +') + movl (%esi), %eax C src low limb + + movl %edx, VAR_JUMP + leal ifelse(UNROLL_BYTES,256,128+) 4(%esi,%ecx,4), %esi + + mull %ebp + + addl %ebx, %eax C initial carry (from _1c) + adcl $0, %edx + + movl %edx, %ebx C high carry + leal ifelse(UNROLL_BYTES,256,128) (%edi,%ecx,4), %edi + + movl VAR_JUMP, %edx + testl $1, %ecx + movl %eax, %ecx C low carry + + cmovnz( %ebx, %ecx) C high,low carry other way around + cmovnz( %eax, %ebx) + + jmp *%edx + + +ifdef(`PIC',` +L(pic_calc): + shll $4, %edx + negl %ecx + + C See mpn/x86/README about old gas bugs + leal (%edx,%ecx,1), %edx + addl $L(entry)-L(here), %edx + + addl (%esp), %edx + + ret_internal +') + + +C ----------------------------------------------------------- + ALIGN(32) +L(top): +deflit(`FRAME',16) + C eax scratch + C ebx carry hi + C ecx carry lo + C edx scratch + C esi src + C edi dst + C ebp multiplier + C + C VAR_COUNTER loop counter + C + C 15 code bytes per limb + + addl $UNROLL_BYTES, %edi + +L(entry): +deflit(CHUNK_COUNT,2) +forloop(`i', 0, UNROLL_COUNT/CHUNK_COUNT-1, ` + deflit(`disp0', eval(i*4*CHUNK_COUNT ifelse(UNROLL_BYTES,256,-128))) + deflit(`disp1', eval(disp0 + 4)) + +Zdisp( movl, disp0,(%esi), %eax) + mull %ebp +Zdisp( M4_inst,%ecx, disp0,(%edi)) + adcl %eax, %ebx + movl %edx, %ecx + adcl $0, %ecx + + movl disp1(%esi), %eax + mull %ebp + M4_inst %ebx, disp1(%edi) + adcl %eax, %ecx + movl %edx, %ebx + adcl $0, %ebx +') + + decl VAR_COUNTER + leal UNROLL_BYTES(%esi), %esi + + jns L(top) + + +deflit(`disp0', eval(UNROLL_BYTES ifelse(UNROLL_BYTES,256,-128))) + + M4_inst %ecx, disp0(%edi) + movl %ebx, %eax + + popl %ebp + popl %edi + + popl %esi + popl %ebx + adcl $0, %eax + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/p6/bdiv_q_1.asm b/gmp-6.3.0/mpn/x86/p6/bdiv_q_1.asm new file mode 100644 index 0000000..a0a9d90 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/bdiv_q_1.asm @@ -0,0 +1,287 @@ +dnl Intel P6 mpn_modexact_1_odd -- exact division style remainder. + +dnl Rearranged from mpn/x86/p6/dive_1.asm by Marco Bodrato. + +dnl Copyright 2001, 2002, 2007, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C odd even divisor +C P6: 10.0 12.0 cycles/limb + +C MULFUNC_PROLOGUE(mpn_bdiv_q_1 mpn_pi1_bdiv_q_1) + +C The odd case is basically the same as mpn_modexact_1_odd, just with an +C extra store, and it runs at the same 10 cycles which is the dependent +C chain. +C +C The shifts for the even case aren't on the dependent chain so in principle +C it could run the same too, but nothing running at 10 has been found. +C Perhaps there's too many uops (an extra 4 over the odd case). + +defframe(PARAM_SHIFT, 24) +defframe(PARAM_INVERSE,20) +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +defframe(SAVE_EBX, -4) +defframe(SAVE_ESI, -8) +defframe(SAVE_EDI, -12) +defframe(SAVE_EBP, -16) +deflit(STACK_SPACE, 16) + +dnl re-use parameter space +define(VAR_INVERSE,`PARAM_SRC') + + TEXT + +C mp_limb_t +C mpn_pi1_bdiv_q_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, mp_limb_t divisor, +C mp_limb_t inverse, int shift) + + ALIGN(16) +PROLOGUE(mpn_pi1_bdiv_q_1) +deflit(`FRAME',0) + + subl $STACK_SPACE, %esp FRAME_subl_esp(STACK_SPACE) + + movl %esi, SAVE_ESI + movl PARAM_SRC, %esi + + movl %ebx, SAVE_EBX + movl PARAM_SIZE, %ebx + + movl %ebp, SAVE_EBP + movl PARAM_INVERSE, %ebp + + movl PARAM_SHIFT, %ecx C trailing twos + +L(common): + movl %edi, SAVE_EDI + movl PARAM_DST, %edi + + leal (%esi,%ebx,4), %esi C src end + + leal (%edi,%ebx,4), %edi C dst end + negl %ebx C -size + + movl (%esi,%ebx,4), %eax C src[0] + + orl %ecx, %ecx + jz L(odd_entry) + + movl %edi, PARAM_DST + movl %ebp, VAR_INVERSE + +L(even): + C eax src[0] + C ebx counter, limbs, negative + C ecx shift + C edx + C esi + C edi + C ebp + + xorl %ebp, %ebp C initial carry bit + xorl %edx, %edx C initial carry limb (for size==1) + + incl %ebx + jz L(even_one) + + movl (%esi,%ebx,4), %edi C src[1] + + shrdl( %cl, %edi, %eax) + + jmp L(even_entry) + + +L(even_top): + C eax scratch + C ebx counter, limbs, negative + C ecx shift + C edx scratch + C esi &src[size] + C edi &dst[size] and scratch + C ebp carry bit + + movl (%esi,%ebx,4), %edi + + mull PARAM_DIVISOR + + movl -4(%esi,%ebx,4), %eax + shrdl( %cl, %edi, %eax) + + subl %ebp, %eax + + sbbl %ebp, %ebp + subl %edx, %eax + + sbbl $0, %ebp + +L(even_entry): + imull VAR_INVERSE, %eax + + movl PARAM_DST, %edi + negl %ebp + + movl %eax, -4(%edi,%ebx,4) + incl %ebx + jnz L(even_top) + + mull PARAM_DIVISOR + + movl -4(%esi), %eax + +L(even_one): + shrl %cl, %eax + movl SAVE_ESI, %esi + + subl %ebp, %eax + movl SAVE_EBP, %ebp + + subl %edx, %eax + movl SAVE_EBX, %ebx + + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi) + movl SAVE_EDI, %edi + addl $STACK_SPACE, %esp + + ret + +C The dependent chain here is +C +C subl %edx, %eax 1 +C imull %ebp, %eax 4 +C mull PARAM_DIVISOR 5 +C ---- +C total 10 +C +C and this is the measured speed. No special scheduling is necessary, out +C of order execution hides the load latency. + +L(odd_top): + C eax scratch (src limb) + C ebx counter, limbs, negative + C ecx carry bit + C edx carry limb, high of last product + C esi &src[size] + C edi &dst[size] + C ebp inverse + + mull PARAM_DIVISOR + + movl (%esi,%ebx,4), %eax + subl %ecx, %eax + + sbbl %ecx, %ecx + subl %edx, %eax + + sbbl $0, %ecx + +L(odd_entry): + imull %ebp, %eax + + movl %eax, (%edi,%ebx,4) + negl %ecx + + incl %ebx + jnz L(odd_top) + + + movl SAVE_ESI, %esi + + movl SAVE_EDI, %edi + + movl SAVE_EBP, %ebp + + movl SAVE_EBX, %ebx + addl $STACK_SPACE, %esp + + ret + +EPILOGUE() + +C mp_limb_t mpn_bdiv_q_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C + + ALIGN(16) +PROLOGUE(mpn_bdiv_q_1) +deflit(`FRAME',0) + + movl PARAM_DIVISOR, %eax + subl $STACK_SPACE, %esp FRAME_subl_esp(STACK_SPACE) + + movl %esi, SAVE_ESI + movl PARAM_SRC, %esi + + movl %ebx, SAVE_EBX + movl PARAM_SIZE, %ebx + + bsfl %eax, %ecx C trailing twos + + movl %ebp, SAVE_EBP + + shrl %cl, %eax C d without twos + + movl %eax, %edx + shrl %eax C d/2 without twos + + movl %edx, PARAM_DIVISOR + andl $127, %eax + +ifdef(`PIC',` + LEA( binvert_limb_table, %ebp) + movzbl (%eax,%ebp), %ebp C inv 8 bits +',` + movzbl binvert_limb_table(%eax), %ebp C inv 8 bits +') + + leal (%ebp,%ebp), %eax C 2*inv + + imull %ebp, %ebp C inv*inv + imull %edx, %ebp C inv*inv*d + + subl %ebp, %eax C inv = 2*inv - inv*inv*d + leal (%eax,%eax), %ebp C 2*inv + + imull %eax, %eax C inv*inv + imull %edx, %eax C inv*inv*d + + subl %eax, %ebp C inv = 2*inv - inv*inv*d + + jmp L(common) + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/p6/copyd.asm b/gmp-6.3.0/mpn/x86/p6/copyd.asm new file mode 100644 index 0000000..1be7636 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/copyd.asm @@ -0,0 +1,178 @@ +dnl Intel P6 mpn_copyd -- copy limb vector backwards. + +dnl Copyright 2001, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P6: 1.75 cycles/limb, or 0.75 if no overlap + + +C void mpn_copyd (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C An explicit loop is used because a decrementing rep movsl is a bit slow at +C 2.4 c/l. That rep movsl also has about a 40 cycle startup time, and the +C code here stands a chance of being faster if the branches predict well. +C +C The slightly strange loop form seems necessary for the claimed speed. +C Maybe load/store ordering affects it. +C +C The source and destination are checked to see if they're actually +C overlapping, since it might be possible to use an incrementing rep movsl +C at 0.75 c/l. (It doesn't suffer the bad startup time of the decrementing +C version.) +C +C Enhancements: +C +C Top speed for an all-integer copy is probably 1.0 c/l, being one load and +C one store each cycle. Unrolling the loop below would approach 1.0, but +C it'd be good to know why something like store/load/subl + store/load/jnz +C doesn't already run at 1.0 c/l. It looks like it should decode in 2 +C cycles, but doesn't run that way. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-using parameter space +define(SAVE_ESI,`PARAM_SIZE') +define(SAVE_EDI,`PARAM_SRC') + + TEXT + ALIGN(16) + +PROLOGUE(mpn_copyd) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + + movl %esi, SAVE_ESI + movl PARAM_SRC, %esi + + movl %edi, SAVE_EDI + movl PARAM_DST, %edi + + subl $1, %ecx + jb L(zero) + + movl (%esi,%ecx,4), %eax C src[size-1] + jz L(one) + + movl -4(%esi,%ecx,4), %edx C src[size-2] + subl $2, %ecx + jbe L(done_loop) C 2 or 3 limbs only + + + C The usual overlap is + C + C high low + C +------------------+ + C | dst| + C +------------------+ + C +------------------+ + C | src| + C +------------------+ + C + C We can use an incrementing copy in the following circumstances. + C + C src+4*size<=dst, since then the regions are disjoint + C + C src==dst, clearly (though this shouldn't occur normally) + C + C src>dst, since in that case it's a requirement of the + C parameters that src>=dst+size*4, and hence the + C regions are disjoint + C + + leal (%edi,%ecx,4), %edx + cmpl %edi, %esi + jae L(use_movsl) C src >= dst + + cmpl %edi, %edx + movl 4(%esi,%ecx,4), %edx C src[size-2] again + jbe L(use_movsl) C src+4*size <= dst + + +L(top): + C eax prev high limb + C ebx + C ecx counter, size-3 down to 0 or -1, inclusive, by 2s + C edx prev low limb + C esi src + C edi dst + C ebp + + movl %eax, 8(%edi,%ecx,4) + movl (%esi,%ecx,4), %eax + + movl %edx, 4(%edi,%ecx,4) + movl -4(%esi,%ecx,4), %edx + + subl $2, %ecx + jnbe L(top) + + +L(done_loop): + movl %eax, 8(%edi,%ecx,4) + movl %edx, 4(%edi,%ecx,4) + + C copy low limb (needed if size was odd, but will already have been + C done in the loop if size was even) + movl (%esi), %eax +L(one): + movl %eax, (%edi) + movl SAVE_EDI, %edi + movl SAVE_ESI, %esi + + ret + + +L(use_movsl): + C eax + C ebx + C ecx size-3 + C edx + C esi src + C edi dst + C ebp + + addl $3, %ecx + + cld C better safe than sorry, see mpn/x86/README + + rep + movsl + +L(zero): + movl SAVE_ESI, %esi + movl SAVE_EDI, %edi + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/p6/dive_1.asm b/gmp-6.3.0/mpn/x86/p6/dive_1.asm new file mode 100644 index 0000000..7d61a18 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/dive_1.asm @@ -0,0 +1,267 @@ +dnl Intel P6 mpn_modexact_1_odd -- exact division style remainder. + +dnl Copyright 2001, 2002, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C odd even divisor +C P6: 10.0 12.0 cycles/limb + + +C void mpn_divexact_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C +C The odd case is basically the same as mpn_modexact_1_odd, just with an +C extra store, and it runs at the same 10 cycles which is the dependent +C chain. +C +C The shifts for the even case aren't on the dependent chain so in principle +C it could run the same too, but nothing running at 10 has been found. +C Perhaps there's too many uops (an extra 4 over the odd case). + +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +defframe(SAVE_EBX, -4) +defframe(SAVE_ESI, -8) +defframe(SAVE_EDI, -12) +defframe(SAVE_EBP, -16) +defframe(VAR_INVERSE, -20) +deflit(STACK_SPACE, 20) + + TEXT + + ALIGN(16) +PROLOGUE(mpn_divexact_1) +deflit(`FRAME',0) + + movl PARAM_DIVISOR, %eax + subl $STACK_SPACE, %esp FRAME_subl_esp(STACK_SPACE) + + movl %esi, SAVE_ESI + movl PARAM_SRC, %esi + + movl %ebx, SAVE_EBX + movl PARAM_SIZE, %ebx + + bsfl %eax, %ecx C trailing twos + + movl %ebp, SAVE_EBP + + shrl %cl, %eax C d without twos + + movl %eax, %edx + shrl %eax C d/2 without twos + + movl %edx, PARAM_DIVISOR + andl $127, %eax + +ifdef(`PIC',` + LEA( binvert_limb_table, %ebp) + movzbl (%eax,%ebp), %ebp C inv 8 bits +',` + movzbl binvert_limb_table(%eax), %ebp C inv 8 bits +') + + leal (%ebp,%ebp), %eax C 2*inv + + imull %ebp, %ebp C inv*inv + + movl %edi, SAVE_EDI + movl PARAM_DST, %edi + + leal (%esi,%ebx,4), %esi C src end + + imull PARAM_DIVISOR, %ebp C inv*inv*d + + subl %ebp, %eax C inv = 2*inv - inv*inv*d + leal (%eax,%eax), %ebp C 2*inv + + imull %eax, %eax C inv*inv + + leal (%edi,%ebx,4), %edi C dst end + negl %ebx C -size + + movl %edi, PARAM_DST + + imull PARAM_DIVISOR, %eax C inv*inv*d + + subl %eax, %ebp C inv = 2*inv - inv*inv*d + + ASSERT(e,` C d*inv == 1 mod 2^GMP_LIMB_BITS + movl PARAM_DIVISOR, %eax + imull %ebp, %eax + cmpl $1, %eax') + + movl %ebp, VAR_INVERSE + movl (%esi,%ebx,4), %eax C src[0] + + orl %ecx, %ecx + jnz L(even) + + C ecx initial carry is zero + jmp L(odd_entry) + + +C The dependent chain here is +C +C subl %edx, %eax 1 +C imull %ebp, %eax 4 +C mull PARAM_DIVISOR 5 +C ---- +C total 10 +C +C and this is the measured speed. No special scheduling is necessary, out +C of order execution hides the load latency. + +L(odd_top): + C eax scratch (src limb) + C ebx counter, limbs, negative + C ecx carry bit + C edx carry limb, high of last product + C esi &src[size] + C edi &dst[size] + C ebp + + mull PARAM_DIVISOR + + movl (%esi,%ebx,4), %eax + subl %ecx, %eax + + sbbl %ecx, %ecx + subl %edx, %eax + + sbbl $0, %ecx + +L(odd_entry): + imull VAR_INVERSE, %eax + + movl %eax, (%edi,%ebx,4) + negl %ecx + + incl %ebx + jnz L(odd_top) + + + movl SAVE_ESI, %esi + + movl SAVE_EDI, %edi + + movl SAVE_EBP, %ebp + + movl SAVE_EBX, %ebx + addl $STACK_SPACE, %esp + + ret + + +L(even): + C eax src[0] + C ebx counter, limbs, negative + C ecx shift + C edx + C esi + C edi + C ebp + + xorl %ebp, %ebp C initial carry bit + xorl %edx, %edx C initial carry limb (for size==1) + + incl %ebx + jz L(even_one) + + movl (%esi,%ebx,4), %edi C src[1] + + shrdl( %cl, %edi, %eax) + + jmp L(even_entry) + + +L(even_top): + C eax scratch + C ebx counter, limbs, negative + C ecx shift + C edx scratch + C esi &src[size] + C edi &dst[size] and scratch + C ebp carry bit + + movl (%esi,%ebx,4), %edi + + mull PARAM_DIVISOR + + movl -4(%esi,%ebx,4), %eax + shrdl( %cl, %edi, %eax) + + subl %ebp, %eax + + sbbl %ebp, %ebp + subl %edx, %eax + + sbbl $0, %ebp + +L(even_entry): + imull VAR_INVERSE, %eax + + movl PARAM_DST, %edi + negl %ebp + + movl %eax, -4(%edi,%ebx,4) + incl %ebx + jnz L(even_top) + + + + mull PARAM_DIVISOR + + movl -4(%esi), %eax + +L(even_one): + shrl %cl, %eax + movl SAVE_ESI, %esi + + subl %ebp, %eax + movl SAVE_EBP, %ebp + + subl %edx, %eax + movl SAVE_EBX, %ebx + + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi) + movl SAVE_EDI, %edi + addl $STACK_SPACE, %esp + + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/p6/gcd_11.asm b/gmp-6.3.0/mpn/x86/p6/gcd_11.asm new file mode 100644 index 0000000..80e055e --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/gcd_11.asm @@ -0,0 +1,83 @@ +dnl x86 mpn_gcd_11 optimised for processors with fast BSF. + +dnl Based on the K7 gcd_1.asm, by Kevin Ryde. Rehacked by Torbjorn Granlund. + +dnl Copyright 2000-2002, 2005, 2009, 2011, 2012, 2015 Free Software +dnl Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/bit (approx) +C AMD K7 7.80 +C AMD K8,K9 7.79 +C AMD K10 4.08 +C AMD bd1 ? +C AMD bobcat 7.82 +C Intel P4-2 14.9 +C Intel P4-3/4 14.0 +C Intel P6/13 5.09 +C Intel core2 4.22 +C Intel NHM 5.00 +C Intel SBR 5.00 +C Intel atom 17.1 +C VIA nano ? +C Numbers measured with: speed -CD -s16-32 -t16 mpn_gcd_1 + + +define(`u0', `%eax') +define(`v0', `%edx') + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(mpn_gcd_11) + push %edi + push %esi + + mov 12(%esp), %eax + mov 16(%esp), %edx + jmp L(odd) + + ALIGN(16) C K10 BD C2 NHM SBR +L(top): cmovc( %esi, %eax) C u = |v - u| 0,3 0,3 0,6 0,5 0,5 + cmovc( %edi, %edx) C v = min(u,v) 0,3 0,3 2,8 1,7 1,7 + shr %cl, %eax C 1,7 1,6 2,8 2,8 2,8 +L(odd): mov %edx, %esi C 1 1 4 3 3 + sub %eax, %esi C 2 2 5 4 4 + bsf %esi, %ecx C 3 3 6 5 5 + mov %eax, %edi C 2 2 3 3 4 + sub %edx, %eax C 2 2 4 3 4 + jnz L(top) C + +L(end): mov %edx, %eax + pop %esi + pop %edi + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/p6/gmp-mparam.h b/gmp-6.3.0/mpn/x86/p6/gmp-mparam.h new file mode 100644 index 0000000..96c96fd --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/gmp-mparam.h @@ -0,0 +1,194 @@ +/* Intel P6 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991, 1993, 1994, 1999-2003, 2008-2010, 2012 Free Software +Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + + +/* NOTE: In a fat binary build SQR_TOOM2_THRESHOLD here cannot be more than the + value in mpn/x86/p6/gmp-mparam.h. The latter is used as a hard limit in + mpn/x86/p6/sqr_basecase.asm. */ + + +/* 1867 MHz P6 model 13 */ + +#define MOD_1_NORM_THRESHOLD 4 +#define MOD_1_UNNORM_THRESHOLD 4 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 5 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 4 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 11 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 8 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 21 + +#define MUL_TOOM22_THRESHOLD 20 +#define MUL_TOOM33_THRESHOLD 74 +#define MUL_TOOM44_THRESHOLD 181 +#define MUL_TOOM6H_THRESHOLD 252 +#define MUL_TOOM8H_THRESHOLD 363 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 73 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 114 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 115 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 80 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 30 +#define SQR_TOOM3_THRESHOLD 101 +#define SQR_TOOM4_THRESHOLD 154 +#define SQR_TOOM6_THRESHOLD 222 +#define SQR_TOOM8_THRESHOLD 527 + +#define MULMID_TOOM42_THRESHOLD 58 + +#define MULMOD_BNM1_THRESHOLD 13 +#define SQRMOD_BNM1_THRESHOLD 17 + +#define POWM_SEC_TABLE 4,23,258,768,2388 + +#define MUL_FFT_MODF_THRESHOLD 565 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 565, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 25, 7}, { 13, 6}, { 28, 7}, { 15, 6}, \ + { 31, 7}, { 17, 6}, { 35, 7}, { 27, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 47, 8}, { 27, 9}, { 15, 8}, \ + { 31, 7}, { 63, 8}, { 39, 9}, { 23, 5}, \ + { 383, 4}, { 991, 5}, { 511, 6}, { 267, 7}, \ + { 157, 8}, { 91, 9}, { 47, 8}, { 111, 9}, \ + { 63, 8}, { 127, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 135,10}, \ + { 79, 9}, { 159,10}, { 95,11}, { 63,10}, \ + { 143, 9}, { 287,10}, { 159,11}, { 95,10}, \ + { 191,12}, { 63,11}, { 127,10}, { 255, 9}, \ + { 511,10}, { 271, 9}, { 543,10}, { 287,11}, \ + { 159,10}, { 335, 9}, { 671,11}, { 191,10}, \ + { 383, 9}, { 767,10}, { 399, 9}, { 799,10}, \ + { 415,11}, { 223,12}, { 127,11}, { 255,10}, \ + { 543, 9}, { 1087,11}, { 287,10}, { 607,11}, \ + { 319,10}, { 671,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,10}, { 831,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,10}, \ + { 1215,12}, { 319,11}, { 671,10}, { 1343,11}, \ + { 735,10}, { 1471,12}, { 383,11}, { 799,10}, \ + { 1599,11}, { 863,12}, { 447,11}, { 959,13}, \ + { 255,12}, { 511,11}, { 1087,12}, { 575,11}, \ + { 1215,12}, { 639,11}, { 1343,12}, { 703,11}, \ + { 1471,13}, { 383,12}, { 831,11}, { 1727,12}, \ + { 959,14}, { 255,13}, { 511,12}, { 1215,13}, \ + { 639,12}, { 1471,11}, { 2943,13}, { 767,12}, \ + { 1727,13}, { 895,12}, { 1919,14}, { 511,13}, \ + { 1023,12}, { 2111,13}, { 1151,12}, { 2431,13}, \ + { 1407,12}, { 2815,14}, { 767,13}, { 1663,12}, \ + { 3455,13}, { 8192,14}, { 16384,15}, { 32768,16} } +#define MUL_FFT_TABLE3_SIZE 132 +#define MUL_FFT_THRESHOLD 6784 + +#define SQR_FFT_MODF_THRESHOLD 472 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 472, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 25, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \ + { 31, 7}, { 17, 6}, { 35, 7}, { 27, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 49, 8}, { 27, 9}, { 15, 8}, \ + { 39, 9}, { 23, 8}, { 51,10}, { 15, 9}, \ + { 31, 8}, { 63, 4}, { 1023, 8}, { 67, 9}, \ + { 39, 5}, { 639, 4}, { 1471, 6}, { 383, 7}, \ + { 209, 8}, { 119, 9}, { 63, 7}, { 255, 8}, \ + { 139, 9}, { 71, 8}, { 143, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159, 8}, { 319, 9}, \ + { 167,10}, { 95,11}, { 63,10}, { 143, 9}, \ + { 287,10}, { 159,11}, { 95,10}, { 191,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 543, 8}, \ + { 1087,10}, { 287, 9}, { 575,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335, 9}, { 671,10}, \ + { 351, 9}, { 703,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 399, 9}, { 799,10}, { 415, 9}, \ + { 831,11}, { 223,12}, { 127,11}, { 255,10}, \ + { 543, 9}, { 1087,11}, { 287,10}, { 607, 9}, \ + { 1215,11}, { 319,10}, { 671, 9}, { 1343,11}, \ + { 351,10}, { 703,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,10}, { 831,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 735,12}, \ + { 383,11}, { 799,10}, { 1599,11}, { 863,12}, \ + { 447,11}, { 959,13}, { 255,12}, { 511,11}, \ + { 1087,12}, { 575,11}, { 1215,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,12}, \ + { 959,14}, { 255,13}, { 511,12}, { 1215,13}, \ + { 639,12}, { 1471,13}, { 767,12}, { 1727,13}, \ + { 895,12}, { 1919,14}, { 511,13}, { 1023,12}, \ + { 2111,13}, { 1151,12}, { 2431,13}, { 1407,14}, \ + { 767,13}, { 1663,12}, { 3455,13}, { 8192,14}, \ + { 16384,15}, { 32768,16} } +#define SQR_FFT_TABLE3_SIZE 146 +#define SQR_FFT_THRESHOLD 5760 + +#define MULLO_BASECASE_THRESHOLD 0 /* always */ +#define MULLO_DC_THRESHOLD 33 +#define MULLO_MUL_N_THRESHOLD 13463 + +#define DC_DIV_QR_THRESHOLD 20 +#define DC_DIVAPPR_Q_THRESHOLD 56 +#define DC_BDIV_QR_THRESHOLD 60 +#define DC_BDIV_Q_THRESHOLD 134 + +#define INV_MULMOD_BNM1_THRESHOLD 38 +#define INV_NEWTON_THRESHOLD 66 +#define INV_APPR_THRESHOLD 63 + +#define BINV_NEWTON_THRESHOLD 250 +#define REDC_1_TO_REDC_N_THRESHOLD 63 + +#define MU_DIV_QR_THRESHOLD 1164 +#define MU_DIVAPPR_Q_THRESHOLD 979 +#define MUPI_DIV_QR_THRESHOLD 38 +#define MU_BDIV_QR_THRESHOLD 1442 +#define MU_BDIV_Q_THRESHOLD 1470 + +#define MATRIX22_STRASSEN_THRESHOLD 17 +#define HGCD_THRESHOLD 64 +#define HGCD_APPR_THRESHOLD 105 +#define HGCD_REDUCE_THRESHOLD 3524 +#define GCD_DC_THRESHOLD 386 +#define GCDEXT_DC_THRESHOLD 309 +#define JACOBI_BASE_METHOD 1 + +#define GET_STR_DC_THRESHOLD 13 +#define GET_STR_PRECOMPUTE_THRESHOLD 26 +#define SET_STR_DC_THRESHOLD 587 +#define SET_STR_PRECOMPUTE_THRESHOLD 1104 diff --git a/gmp-6.3.0/mpn/x86/p6/lshsub_n.asm b/gmp-6.3.0/mpn/x86/p6/lshsub_n.asm new file mode 100644 index 0000000..7ada213 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/lshsub_n.asm @@ -0,0 +1,169 @@ +dnl Intel P6 mpn_lshsub_n -- mpn papillion support. + +dnl Copyright 2006 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C P6/13: 3.35 cycles/limb (separate mpn_sub_n + mpn_lshift needs 4.12) + +C (1) The loop is not scheduled in any way, and scheduling attempts have not +C improved speed on P6/13. Presumably, the K7 will want scheduling, if it +C at all wants to use MMX. +C (2) We could save a register by not alternatingly using eax and edx in the +C loop. + +define(`rp', `%edi') +define(`up', `%esi') +define(`vp', `%ebx') +define(`n', `%ecx') +define(`cnt', `%mm7') + +ASM_START() + + TEXT + ALIGN(16) + +PROLOGUE(mpn_lshsub_n) + push %edi + push %esi + push %ebx + + mov 16(%esp), rp + mov 20(%esp), up + mov 24(%esp), vp + mov 28(%esp), n + mov $32, %eax + sub 32(%esp), %eax + movd %eax, cnt + + lea (up,n,4), up + lea (vp,n,4), vp + lea (rp,n,4), rp + + neg n + mov n, %eax + and $-8, n + and $7, %eax + shl %eax C eax = 2x + lea (%eax,%eax,4), %edx C edx = 10x +ifdef(`PIC',` + call L(pic_calc) +L(here): +',` + lea L(ent)(%eax,%edx,2), %eax C eax = 22x +') + + pxor %mm1, %mm1 + pxor %mm0, %mm0 + + jmp *%eax + +ifdef(`PIC',` +L(pic_calc): + C See mpn/x86/README about old gas bugs + lea (%eax,%edx,2), %eax + add $L(ent)-L(here), %eax + add (%esp), %eax + ret_internal +') + +L(end): C compute (cy<>(32-cnt)) + sbb %eax, %eax + neg %eax + mov 32(%esp), %ecx + shld %cl, %edx, %eax + + emms + + pop %ebx + pop %esi + pop %edi + ret + ALIGN(16) +L(top): jecxz L(end) +L(ent): mov 0(up,n,4), %eax + sbb 0(vp,n,4), %eax + movd %eax, %mm0 + punpckldq %mm0, %mm1 + psrlq %mm7, %mm1 + movd %mm1, 0(rp,n,4) + + mov 4(up,n,4), %edx + sbb 4(vp,n,4), %edx + movd %edx, %mm1 + punpckldq %mm1, %mm0 + psrlq %mm7, %mm0 + movd %mm0, 4(rp,n,4) + + mov 8(up,n,4), %eax + sbb 8(vp,n,4), %eax + movd %eax, %mm0 + punpckldq %mm0, %mm1 + psrlq %mm7, %mm1 + movd %mm1, 8(rp,n,4) + + mov 12(up,n,4), %edx + sbb 12(vp,n,4), %edx + movd %edx, %mm1 + punpckldq %mm1, %mm0 + psrlq %mm7, %mm0 + movd %mm0, 12(rp,n,4) + + mov 16(up,n,4), %eax + sbb 16(vp,n,4), %eax + movd %eax, %mm0 + punpckldq %mm0, %mm1 + psrlq %mm7, %mm1 + movd %mm1, 16(rp,n,4) + + mov 20(up,n,4), %edx + sbb 20(vp,n,4), %edx + movd %edx, %mm1 + punpckldq %mm1, %mm0 + psrlq %mm7, %mm0 + movd %mm0, 20(rp,n,4) + + mov 24(up,n,4), %eax + sbb 24(vp,n,4), %eax + movd %eax, %mm0 + punpckldq %mm0, %mm1 + psrlq %mm7, %mm1 + movd %mm1, 24(rp,n,4) + + mov 28(up,n,4), %edx + sbb 28(vp,n,4), %edx + movd %edx, %mm1 + punpckldq %mm1, %mm0 + psrlq %mm7, %mm0 + movd %mm0, 28(rp,n,4) + + lea 8(n), n + jmp L(top) + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/p6/mmx/divrem_1.asm b/gmp-6.3.0/mpn/x86/p6/mmx/divrem_1.asm new file mode 100644 index 0000000..5300616 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/mmx/divrem_1.asm @@ -0,0 +1,767 @@ +dnl Intel Pentium-II mpn_divrem_1 -- mpn by limb division. + +dnl Copyright 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P6MMX: 25.0 cycles/limb integer part, 17.5 cycles/limb fraction part. + + +C mp_limb_t mpn_divrem_1 (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C mp_limb_t mpn_divrem_1c (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, +C mp_limb_t divisor, mp_limb_t carry); +C mp_limb_t mpn_preinv_divrem_1 (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, +C mp_limb_t divisor, mp_limb_t inverse, +C unsigned shift); +C +C This code is a lightly reworked version of mpn/x86/k7/mmx/divrem_1.asm, +C see that file for some comments. It's possible what's here can be improved. + + +dnl MUL_THRESHOLD is the value of xsize+size at which the multiply by +dnl inverse method is used, rather than plain "divl"s. Minimum value 1. +dnl +dnl The different speeds of the integer and fraction parts means that using +dnl xsize+size isn't quite right. The threshold wants to be a bit higher +dnl for the integer part and a bit lower for the fraction part. (Or what's +dnl really wanted is to speed up the integer part!) +dnl +dnl The threshold is set to make the integer part right. At 4 limbs the +dnl div and mul are about the same there, but on the fractional part the +dnl mul is much faster. + +deflit(MUL_THRESHOLD, 4) + + +defframe(PARAM_PREINV_SHIFT, 28) dnl mpn_preinv_divrem_1 +defframe(PARAM_PREINV_INVERSE, 24) dnl mpn_preinv_divrem_1 +defframe(PARAM_CARRY, 24) dnl mpn_divrem_1c +defframe(PARAM_DIVISOR,20) +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC, 12) +defframe(PARAM_XSIZE, 8) +defframe(PARAM_DST, 4) + +defframe(SAVE_EBX, -4) +defframe(SAVE_ESI, -8) +defframe(SAVE_EDI, -12) +defframe(SAVE_EBP, -16) + +defframe(VAR_NORM, -20) +defframe(VAR_INVERSE, -24) +defframe(VAR_SRC, -28) +defframe(VAR_DST, -32) +defframe(VAR_DST_STOP,-36) + +deflit(STACK_SPACE, 36) + + TEXT + ALIGN(16) + +PROLOGUE(mpn_preinv_divrem_1) +deflit(`FRAME',0) + movl PARAM_XSIZE, %ecx + subl $STACK_SPACE, %esp FRAME_subl_esp(STACK_SPACE) + + movl %esi, SAVE_ESI + movl PARAM_SRC, %esi + + movl %ebx, SAVE_EBX + movl PARAM_SIZE, %ebx + + movl %ebp, SAVE_EBP + movl PARAM_DIVISOR, %ebp + + movl %edi, SAVE_EDI + movl PARAM_DST, %edx + + movl -4(%esi,%ebx,4), %eax C src high limb + xorl %edi, %edi C initial carry (if can't skip a div) + + C + + leal 8(%edx,%ecx,4), %edx C &dst[xsize+2] + xor %ecx, %ecx + + movl %edx, VAR_DST_STOP C &dst[xsize+2] + cmpl %ebp, %eax C high cmp divisor + + cmovc( %eax, %edi) C high is carry if high n2 + leal (%ebp,%esi), %edx + + cmovc( %edx, %edi) C n - q1*d if underflow from using q1+1 + movd %mm0, %esi + + sbbl $0, %ebx C q + subl $4, %ecx + + movl %ebx, (%ecx) + cmpl %eax, %ecx + + movl %ecx, VAR_DST + jne L(integer_top) + + +L(integer_loop_done): + + +C ----------------------------------------------------------------------------- +C +C Here, and in integer_one_left below, an sbbl $0 is used rather than a jz +C q1_ff special case. This make the code a bit smaller and simpler, and +C costs only 2 cycles (each). + +L(integer_two_left): + C eax scratch + C ebx scratch (nadj, q1) + C ecx scratch (src, dst) + C edx scratch + C esi n10 + C edi n2 + C ebp divisor + C + C mm7 rshift + + + movl %esi, %eax + movl %ebp, %ebx + + sarl $31, %eax C -n1 + movl PARAM_SRC, %ecx + + andl %eax, %ebx C -n1 & d + negl %eax C n1 + + addl %esi, %ebx C nadj = n10 + (-n1 & d), ignoring overflow + addl %edi, %eax C n2+n1 + + mull VAR_INVERSE C m*(n2+n1) + + movd (%ecx), %mm0 C src low limb + + movl VAR_DST_STOP, %ecx + + C + + C + + addl %ebx, %eax C m*(n2+n1) + nadj, low giving carry flag + leal 1(%edi), %ebx C n2+1 + movl %ebp, %eax C d + + adcl %edx, %ebx C 1 + high(n2<<32 + m*(n2+n1) + nadj) = q1+1 + + sbbl $0, %ebx + + mull %ebx C (q1+1)*d + + psllq $32, %mm0 + + psrlq %mm7, %mm0 + + C + + C + + subl %eax, %esi + + sbbl %edx, %edi C n - (q1+1)*d + movl %esi, %edi C remainder -> n2 + leal (%ebp,%esi), %edx + + cmovc( %edx, %edi) C n - q1*d if underflow from using q1+1 + movd %mm0, %esi + + sbbl $0, %ebx C q + + movl %ebx, -4(%ecx) + + +C ----------------------------------------------------------------------------- +L(integer_one_left): + C eax scratch + C ebx scratch (nadj, q1) + C ecx scratch (dst) + C edx scratch + C esi n10 + C edi n2 + C ebp divisor + C + C mm7 rshift + + + movl %esi, %eax + movl %ebp, %ebx + + sarl $31, %eax C -n1 + movl VAR_DST_STOP, %ecx + + andl %eax, %ebx C -n1 & d + negl %eax C n1 + + addl %esi, %ebx C nadj = n10 + (-n1 & d), ignoring overflow + addl %edi, %eax C n2+n1 + + mull VAR_INVERSE C m*(n2+n1) + + C + + C + + C + + addl %ebx, %eax C m*(n2+n1) + nadj, low giving carry flag + leal 1(%edi), %ebx C n2+1 + movl %ebp, %eax C d + + C + + adcl %edx, %ebx C 1 + high(n2<<32 + m*(n2+n1) + nadj) = q1+1 + + sbbl $0, %ebx C q1 if q1+1 overflowed + + mull %ebx + + C + + C + + C + + C + + subl %eax, %esi + movl PARAM_XSIZE, %eax + + sbbl %edx, %edi C n - (q1+1)*d + movl %esi, %edi C remainder -> n2 + leal (%ebp,%esi), %edx + + cmovc( %edx, %edi) C n - q1*d if underflow from using q1+1 + + sbbl $0, %ebx C q + + movl %ebx, -8(%ecx) + subl $8, %ecx + + + + orl %eax, %eax C xsize + jnz L(fraction_some) + + movl %edi, %eax +L(fraction_done): + movl VAR_NORM, %ecx +L(zero_done): + movl SAVE_EBP, %ebp + + movl SAVE_EDI, %edi + + movl SAVE_ESI, %esi + + movl SAVE_EBX, %ebx + addl $STACK_SPACE, %esp + + shrl %cl, %eax + emms + + ret + + +C ----------------------------------------------------------------------------- +C +C Special case for q1=0xFFFFFFFF, giving q=0xFFFFFFFF meaning the low dword +C of q*d is simply -d and the remainder n-q*d = n10+d + +L(q1_ff): + C eax (divisor) + C ebx (q1+1 == 0) + C ecx + C edx + C esi n10 + C edi n2 + C ebp divisor + + movl VAR_DST, %ecx + movl VAR_DST_STOP, %edx + subl $4, %ecx + + movl %ecx, VAR_DST + psrlq %mm7, %mm0 + leal (%ebp,%esi), %edi C n-q*d remainder -> next n2 + + movl $-1, (%ecx) + movd %mm0, %esi C next n10 + + cmpl %ecx, %edx + jne L(integer_top) + + jmp L(integer_loop_done) + + + +C ----------------------------------------------------------------------------- +C +C In the current implementation, the following successively dependent +C micro-ops seem to exist. +C +C uops +C mul 5 +C q1+1 1 (addl) +C mul 5 +C sub 3 (negl/sbbl) +C addback 2 (cmov) +C --- +C 16 +C +C The loop in fact runs at about 17.5 cycles. Using a sarl/andl/addl for +C the addback was found to be a touch slower. + + + ALIGN(16) +L(fraction_some): + C eax + C ebx + C ecx + C edx + C esi + C edi carry + C ebp divisor + + movl PARAM_DST, %esi + movl VAR_DST_STOP, %ecx C &dst[xsize+2] + movl %edi, %eax + + subl $8, %ecx C &dst[xsize] + + + ALIGN(16) +L(fraction_top): + C eax n2, then scratch + C ebx scratch (nadj, q1) + C ecx dst, decrementing + C edx scratch + C esi dst stop point + C edi n2 + C ebp divisor + + mull VAR_INVERSE C m*n2 + + movl %ebp, %eax C d + subl $4, %ecx C dst + leal 1(%edi), %ebx + + C + + C + + C + + addl %edx, %ebx C 1 + high(n2<<32 + m*n2) = q1+1 + + mull %ebx C (q1+1)*d + + C + + C + + C + + C + + negl %eax C low of n - (q1+1)*d + + sbbl %edx, %edi C high of n - (q1+1)*d, caring only about carry + leal (%ebp,%eax), %edx + + cmovc( %edx, %eax) C n - q1*d if underflow from using q1+1 + + sbbl $0, %ebx C q + movl %eax, %edi C remainder->n2 + cmpl %esi, %ecx + + movl %ebx, (%ecx) C previous q + jne L(fraction_top) + + + jmp L(fraction_done) + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/p6/mmx/gmp-mparam.h b/gmp-6.3.0/mpn/x86/p6/mmx/gmp-mparam.h new file mode 100644 index 0000000..ef29061 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/mmx/gmp-mparam.h @@ -0,0 +1,218 @@ +/* Intel P6/mmx gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991-2017 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + + +/* NOTE: In a fat binary build SQR_TOOM2_THRESHOLD here cannot be more than the + value in mpn/x86/p6/gmp-mparam.h. The latter is used as a hard limit in + mpn/x86/p6/sqr_basecase.asm. */ + + +/* 800 MHz P6 model 8 */ +/* Generated by tuneup.c, 2017-02-03, gcc 4.8 */ + +#define MOD_1_1P_METHOD 2 +#define MOD_1_NORM_THRESHOLD 3 +#define MOD_1_UNNORM_THRESHOLD 4 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 7 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 7 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 8 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 30 +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 14 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 +#define DIV_QR_1_NORM_THRESHOLD 4 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 62 + +#define DIV_1_VS_MUL_1_PERCENT 168 + +#define MUL_TOOM22_THRESHOLD 22 +#define MUL_TOOM33_THRESHOLD 73 +#define MUL_TOOM44_THRESHOLD 195 +#define MUL_TOOM6H_THRESHOLD 254 +#define MUL_TOOM8H_THRESHOLD 381 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 81 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 122 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 73 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 80 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 100 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 30 /* WRONG value, see comment above */ +#define SQR_TOOM3_THRESHOLD 83 +#define SQR_TOOM4_THRESHOLD 196 +#define SQR_TOOM6_THRESHOLD 214 +#define SQR_TOOM8_THRESHOLD 381 + +#define MULMID_TOOM42_THRESHOLD 56 + +#define MULMOD_BNM1_THRESHOLD 16 +#define SQRMOD_BNM1_THRESHOLD 17 + +#define MUL_FFT_MODF_THRESHOLD 476 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 476, 5}, { 21, 6}, { 11, 5}, { 23, 6}, \ + { 21, 7}, { 11, 6}, { 25, 7}, { 13, 6}, \ + { 27, 7}, { 15, 6}, { 31, 7}, { 21, 8}, \ + { 11, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 51,10}, { 15, 9}, \ + { 31, 8}, { 67, 9}, { 39, 8}, { 79, 9}, \ + { 47, 8}, { 95, 9}, { 55,10}, { 31, 9}, \ + { 63, 8}, { 127, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 135,10}, \ + { 79, 9}, { 167,10}, { 95, 9}, { 199,10}, \ + { 111,11}, { 63,10}, { 127, 9}, { 255, 8}, \ + { 511,10}, { 143, 9}, { 287, 8}, { 575,10}, \ + { 159,11}, { 95,10}, { 191, 9}, { 383,10}, \ + { 207,12}, { 63,11}, { 127,10}, { 255, 9}, \ + { 511,10}, { 271, 9}, { 543, 8}, { 1087,10}, \ + { 287, 9}, { 575,11}, { 159,10}, { 319, 9}, \ + { 639,10}, { 351, 9}, { 703,11}, { 191,10}, \ + { 383, 9}, { 767,10}, { 415, 9}, { 831,11}, \ + { 223,10}, { 447,12}, { 127,11}, { 255,10}, \ + { 543, 9}, { 1087,11}, { 287,10}, { 607, 9}, \ + { 1215,11}, { 319,10}, { 671,11}, { 351,10}, \ + { 703,12}, { 191,11}, { 383,10}, { 767,11}, \ + { 415,10}, { 831,11}, { 447,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,10}, \ + { 1215,12}, { 319,11}, { 671,10}, { 1343,11}, \ + { 703,10}, { 1407,11}, { 735,12}, { 383,11}, \ + { 831,12}, { 447,11}, { 959,10}, { 1919,13}, \ + { 255,12}, { 511,11}, { 1087,12}, { 575,11}, \ + { 1215,10}, { 2431,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1471,13}, { 383,12}, { 767,11}, \ + { 1535,12}, { 831,11}, { 1727,12}, { 959,11}, \ + { 1919,14}, { 255,13}, { 511,12}, { 1215,11}, \ + { 2431,13}, { 639,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1727,13}, { 895,12}, { 1919,11}, \ + { 3839,14}, { 511,13}, { 1023,12}, { 2111,13}, \ + { 1151,12}, { 2431,13}, { 1279,12}, { 2559,13}, \ + { 1407,12}, { 2943,14}, { 767,13}, { 1663,12}, \ + { 3327,13}, { 1919,12}, { 3839,15}, { 32768,16} } +#define MUL_FFT_TABLE3_SIZE 160 +#define MUL_FFT_THRESHOLD 7040 + +#define SQR_FFT_MODF_THRESHOLD 376 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 376, 5}, { 21, 6}, { 11, 5}, { 23, 6}, \ + { 21, 7}, { 11, 6}, { 24, 7}, { 13, 6}, \ + { 27, 7}, { 15, 6}, { 31, 7}, { 21, 8}, \ + { 11, 7}, { 27, 8}, { 15, 7}, { 33, 8}, \ + { 19, 7}, { 39, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 39, 9}, { 23, 8}, \ + { 51,10}, { 15, 9}, { 31, 8}, { 67, 9}, \ + { 39, 8}, { 79, 9}, { 47, 8}, { 95, 9}, \ + { 55,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 127, 8}, \ + { 255, 9}, { 135,10}, { 79, 9}, { 167,10}, \ + { 95, 9}, { 191, 8}, { 383,10}, { 111,11}, \ + { 63,10}, { 127, 9}, { 255, 8}, { 511, 9}, \ + { 271,10}, { 143, 9}, { 287, 8}, { 575, 9}, \ + { 303, 8}, { 607,10}, { 159, 9}, { 319,11}, \ + { 95,10}, { 191, 9}, { 383,10}, { 207,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 511,10}, \ + { 271, 9}, { 543,10}, { 287, 9}, { 575,10}, \ + { 303,11}, { 159,10}, { 319, 9}, { 639,10}, \ + { 351, 9}, { 703,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 415, 9}, { 831,11}, { 223,10}, \ + { 479,12}, { 127,11}, { 255,10}, { 543, 9}, \ + { 1087,11}, { 287,10}, { 607, 9}, { 1215,11}, \ + { 319,10}, { 671,11}, { 351,10}, { 703,12}, \ + { 191,11}, { 383,10}, { 767,11}, { 415,10}, \ + { 831,11}, { 479,13}, { 127,12}, { 255,11}, \ + { 543,10}, { 1087,11}, { 607,10}, { 1215,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 703,10}, \ + { 1407,11}, { 735,12}, { 383,11}, { 831,12}, \ + { 447,11}, { 959,10}, { 1919,13}, { 255,12}, \ + { 511,11}, { 1087,12}, { 575,11}, { 1215,10}, \ + { 2431,12}, { 639,11}, { 1343,12}, { 703,11}, \ + { 1407,13}, { 383,12}, { 831,11}, { 1727,12}, \ + { 959,11}, { 1919,14}, { 255,13}, { 511,12}, \ + { 1215,11}, { 2431,13}, { 639,12}, { 1471,11}, \ + { 2943,13}, { 767,12}, { 1727,13}, { 895,12}, \ + { 1919,11}, { 3839,14}, { 511,13}, { 1023,12}, \ + { 2111,13}, { 1151,12}, { 2431,13}, { 1407,12}, \ + { 2943,14}, { 767,13}, { 1535,12}, { 3071,13}, \ + { 1663,12}, { 3455,13}, { 1919,12}, { 3839,15}, \ + { 32768,16} } +#define SQR_FFT_TABLE3_SIZE 161 +#define SQR_FFT_THRESHOLD 3712 + +#define MULLO_BASECASE_THRESHOLD 0 /* always */ +#define MULLO_DC_THRESHOLD 62 +#define MULLO_MUL_N_THRESHOLD 13463 +#define SQRLO_BASECASE_THRESHOLD 0 /* always */ +#define SQRLO_DC_THRESHOLD 177 +#define SQRLO_SQR_THRESHOLD 8937 + +#define DC_DIV_QR_THRESHOLD 80 +#define DC_DIVAPPR_Q_THRESHOLD 240 +#define DC_BDIV_QR_THRESHOLD 76 +#define DC_BDIV_Q_THRESHOLD 166 + +#define INV_MULMOD_BNM1_THRESHOLD 42 +#define INV_NEWTON_THRESHOLD 262 +#define INV_APPR_THRESHOLD 250 + +#define BINV_NEWTON_THRESHOLD 272 +#define REDC_1_TO_REDC_N_THRESHOLD 72 + +#define MU_DIV_QR_THRESHOLD 1499 +#define MU_DIVAPPR_Q_THRESHOLD 1470 +#define MUPI_DIV_QR_THRESHOLD 124 +#define MU_BDIV_QR_THRESHOLD 1142 +#define MU_BDIV_Q_THRESHOLD 1341 + +#define POWM_SEC_TABLE 1,16,96,416,1259 + +#define GET_STR_DC_THRESHOLD 14 +#define GET_STR_PRECOMPUTE_THRESHOLD 27 +#define SET_STR_DC_THRESHOLD 270 +#define SET_STR_PRECOMPUTE_THRESHOLD 1084 + +#define FAC_DSC_THRESHOLD 194 +#define FAC_ODD_THRESHOLD 25 + +#define MATRIX22_STRASSEN_THRESHOLD 16 +#define HGCD_THRESHOLD 124 +#define HGCD_APPR_THRESHOLD 152 +#define HGCD_REDUCE_THRESHOLD 3014 +#define GCD_DC_THRESHOLD 474 +#define GCDEXT_DC_THRESHOLD 321 +#define JACOBI_BASE_METHOD 1 diff --git a/gmp-6.3.0/mpn/x86/p6/mmx/lshift.asm b/gmp-6.3.0/mpn/x86/p6/mmx/lshift.asm new file mode 100644 index 0000000..febd1c0 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/mmx/lshift.asm @@ -0,0 +1,38 @@ +dnl Intel Pentium-II mpn_lshift -- mpn left shift. + +dnl Copyright 2001 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + + +dnl The P55 code runs well on P-II/III, but could stand some minor tweaks +dnl at some stage probably. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_lshift) +include_mpn(`x86/pentium/mmx/lshift.asm') diff --git a/gmp-6.3.0/mpn/x86/p6/mmx/popham.asm b/gmp-6.3.0/mpn/x86/p6/mmx/popham.asm new file mode 100644 index 0000000..fd340e4 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/mmx/popham.asm @@ -0,0 +1,39 @@ +dnl Intel Pentium-II mpn_popcount, mpn_hamdist -- population count and +dnl hamming distance. + +dnl Copyright 2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P6MMX: popcount 11 cycles/limb (approx), hamdist 11.5 cycles/limb (approx) + + +MULFUNC_PROLOGUE(mpn_popcount mpn_hamdist) +include_mpn(`x86/k6/mmx/popham.asm') diff --git a/gmp-6.3.0/mpn/x86/p6/mmx/rshift.asm b/gmp-6.3.0/mpn/x86/p6/mmx/rshift.asm new file mode 100644 index 0000000..77aa190 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/mmx/rshift.asm @@ -0,0 +1,38 @@ +dnl Intel Pentium-II mpn_rshift -- mpn left shift. + +dnl Copyright 2001 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + + +dnl The P55 code runs well on P-II/III, but could stand some minor tweaks +dnl at some stage probably. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_rshift) +include_mpn(`x86/pentium/mmx/rshift.asm') diff --git a/gmp-6.3.0/mpn/x86/p6/mod_34lsub1.asm b/gmp-6.3.0/mpn/x86/p6/mod_34lsub1.asm new file mode 100644 index 0000000..b88ab5d --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/mod_34lsub1.asm @@ -0,0 +1,190 @@ +dnl Intel P6 mpn_mod_34lsub1 -- remainder modulo 2^24-1. + +dnl Copyright 2000-2002, 2004 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P6: 2.0 cycles/limb + +C TODO +C Experiments with more unrolling indicate that 1.5 c/l is possible on P6-13 +C with the current carry handling scheme. + +C mp_limb_t mpn_mod_34lsub1 (mp_srcptr src, mp_size_t size) +C +C Groups of three limbs are handled, with carry bits from 0mod3 into 1mod3 +C into 2mod3, but at that point going into a separate carries total so we +C don't keep the carry flag live across the loop control. Avoiding decl +C lets us get to 2.0 c/l, as compared to the generic x86 code at 3.66. +C + +defframe(PARAM_SIZE, 8) +defframe(PARAM_SRC, 4) + +dnl re-use parameter space +define(SAVE_EBX, `PARAM_SIZE') +define(SAVE_ESI, `PARAM_SRC') + + TEXT + ALIGN(16) +PROLOGUE(mpn_mod_34lsub1) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + movl PARAM_SRC, %edx + + subl $2, %ecx C size-2 + movl (%edx), %eax C src[0] + ja L(three_or_more) + jb L(one) + + C size==2 + + movl 4(%edx), %ecx C src[1] + + movl %eax, %edx C src[0] + shrl $24, %eax C src[0] high + + andl $0xFFFFFF, %edx C src[0] low + + addl %edx, %eax + movl %ecx, %edx C src[1] + shrl $16, %ecx C src[1] high + + andl $0xFFFF, %edx + addl %ecx, %eax + + shll $8, %edx C src[1] low + + addl %edx, %eax +L(one): + ret + + +L(three_or_more): + C eax src[0], initial acc 0mod3 + C ebx + C ecx size-2 + C edx src + C esi + C edi + C ebp + + movl %ebx, SAVE_EBX + movl 4(%edx), %ebx C src[1], initial 1mod3 + subl $3, %ecx C size-5 + + movl %esi, SAVE_ESI + movl 8(%edx), %esi C src[2], initial 2mod3 + + pushl %edi FRAME_pushl() + movl $0, %edi C initial carries 0mod3 + jng L(done) C if size < 6 + + +L(top): + C eax acc 0mod3 + C ebx acc 1mod3 + C ecx counter, limbs + C edx src + C esi acc 2mod3 + C edi carrys into 0mod3 + C ebp + + addl 12(%edx), %eax + adcl 16(%edx), %ebx + adcl 20(%edx), %esi + leal 12(%edx), %edx + adcl $0, %edi + + subl $3, %ecx + jg L(top) C at least 3 more to process + + +L(done): + C ecx is -2, -1 or 0 representing 0, 1 or 2 more limbs respectively + cmpl $-1, %ecx + jl L(done_0) C if -2, meaning 0 more limbs + + C 1 or 2 more limbs + movl $0, %ecx + je L(done_1) C if -1, meaning 1 more limb only + movl 16(%edx), %ecx +L(done_1): + addl 12(%edx), %eax C 0mod3 + adcl %ecx, %ebx C 1mod3 + adcl $0, %esi C 2mod3 + adcl $0, %edi C carries 0mod3 + +L(done_0): + C eax acc 0mod3 + C ebx acc 1mod3 + C ecx + C edx + C esi acc 2mod3 + C edi carries 0mod3 + C ebp + + movl %eax, %ecx C 0mod3 + shrl $24, %eax C 0mod3 high initial total + + andl $0xFFFFFF, %ecx C 0mod3 low + movl %edi, %edx C carries + shrl $24, %edi C carries high + + addl %ecx, %eax C add 0mod3 low + andl $0xFFFFFF, %edx C carries 0mod3 low + movl %ebx, %ecx C 1mod3 + + shrl $16, %ebx C 1mod3 high + addl %edi, %eax C add carries high + addl %edx, %eax C add carries 0mod3 low + + andl $0xFFFF, %ecx C 1mod3 low mask + addl %ebx, %eax C add 1mod3 high + movl SAVE_EBX, %ebx + + shll $8, %ecx C 1mod3 low + movl %esi, %edx C 2mod3 + popl %edi FRAME_popl() + + shrl $8, %esi C 2mod3 high + andl $0xFF, %edx C 2mod3 low mask + addl %ecx, %eax C add 1mod3 low + + shll $16, %edx C 2mod3 low + addl %esi, %eax C add 2mod3 high + movl SAVE_ESI, %esi + + addl %edx, %eax C add 2mod3 low + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/p6/mode1o.asm b/gmp-6.3.0/mpn/x86/p6/mode1o.asm new file mode 100644 index 0000000..7083195 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/mode1o.asm @@ -0,0 +1,170 @@ +dnl Intel P6 mpn_modexact_1_odd -- exact division style remainder. + +dnl Copyright 2000-2002, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P6: 10.0 cycles/limb + + +C mp_limb_t mpn_modexact_1_odd (mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C mp_limb_t mpn_modexact_1c_odd (mp_srcptr src, mp_size_t size, +C mp_limb_t divisor, mp_limb_t carry); +C +C It's not worth skipping a step at the end when high32),1, +eval((UNROLL_COUNT-32)*4), +0)) + + C eax + C ebx carry + C ecx + C edx + C esi &src[size] + C edi &dst[size-1] + C ebp + + movl PARAM_SIZE, %ecx + + subl $4, %ecx + jz L(corner) + + movl %ecx, %edx + negl %ecx + + shll $4, %ecx +ifelse(OFFSET,0,,`subl $OFFSET, %esi') + +ifdef(`PIC',` + call L(pic_calc) +L(here): +',` + leal L(unroll_inner_end)-eval(2*CODE_BYTES_PER_LIMB)(%ecx,%edx), %ecx +') + negl %edx + +ifelse(OFFSET,0,,`subl $OFFSET, %edi') + + C The calculated jump mustn't be before the start of the available + C code. This is the limit that UNROLL_COUNT puts on the src operand + C size, but checked here using the jump address directly. + + ASSERT(ae, + `movl_text_address( L(unroll_inner_start), %eax) + cmpl %eax, %ecx') + + +C ----------------------------------------------------------------------------- + ALIGN(16) +L(unroll_outer_top): + C eax + C ebx high limb to store + C ecx VAR_JMP + C edx VAR_COUNTER, limbs, negative + C esi &src[size], constant + C edi dst ptr, second highest limb of last addmul + C ebp + + movl -12+OFFSET(%esi,%edx,4), %ebp C multiplier + movl %edx, VAR_COUNTER + + movl -8+OFFSET(%esi,%edx,4), %eax C first limb of multiplicand + + mull %ebp + +define(cmovX,`ifelse(eval(UNROLL_COUNT%2),1,`cmovz($@)',`cmovnz($@)')') + + testb $1, %cl + + movl %edx, %ebx C high carry + leal 4(%edi), %edi + + movl %ecx, %edx C jump + + movl %eax, %ecx C low carry + leal CODE_BYTES_PER_LIMB(%edx), %edx + + cmovX( %ebx, %ecx) C high carry reverse + cmovX( %eax, %ebx) C low carry reverse + movl %edx, VAR_JMP + jmp *%edx + + + C Must be on an even address here so the low bit of the jump address + C will indicate which way around ecx/ebx should start. + + ALIGN(2) + +L(unroll_inner_start): + C eax scratch + C ebx carry high + C ecx carry low + C edx scratch + C esi src pointer + C edi dst pointer + C ebp multiplier + C + C 15 code bytes each limb + C ecx/ebx reversed on each chunk + +forloop(`i', UNROLL_COUNT, 1, ` + deflit(`disp_src', eval(-i*4 + OFFSET)) + deflit(`disp_dst', eval(disp_src)) + + m4_assert(`disp_src>=-128 && disp_src<128') + m4_assert(`disp_dst>=-128 && disp_dst<128') + +ifelse(eval(i%2),0,` +Zdisp( movl, disp_src,(%esi), %eax) + mull %ebp +Zdisp( addl, %ebx, disp_dst,(%edi)) + adcl %eax, %ecx + movl %edx, %ebx + adcl $0, %ebx +',` + dnl this one comes out last +Zdisp( movl, disp_src,(%esi), %eax) + mull %ebp +Zdisp( addl, %ecx, disp_dst,(%edi)) + adcl %eax, %ebx + movl %edx, %ecx + adcl $0, %ecx +') +') +L(unroll_inner_end): + + addl %ebx, m4_empty_if_zero(OFFSET)(%edi) + + movl VAR_COUNTER, %edx + adcl $0, %ecx + + movl %ecx, m4_empty_if_zero(OFFSET+4)(%edi) + movl VAR_JMP, %ecx + + incl %edx + jnz L(unroll_outer_top) + + +ifelse(OFFSET,0,,` + addl $OFFSET, %esi + addl $OFFSET, %edi +') + + +C ----------------------------------------------------------------------------- + ALIGN(16) +L(corner): + C eax + C ebx + C ecx + C edx + C esi &src[size] + C edi &dst[2*size-5] + C ebp + + movl -12(%esi), %eax + + mull -8(%esi) + + addl %eax, (%edi) + movl -12(%esi), %eax + movl $0, %ebx + + adcl %edx, %ebx + + mull -4(%esi) + + addl %eax, %ebx + movl -8(%esi), %eax + + adcl $0, %edx + + addl %ebx, 4(%edi) + movl $0, %ebx + + adcl %edx, %ebx + + mull -4(%esi) + + movl PARAM_SIZE, %ecx + addl %ebx, %eax + + adcl $0, %edx + + movl %eax, 8(%edi) + + movl %edx, 12(%edi) + movl PARAM_DST, %edi + + +C Left shift of dst[1..2*size-2], the bit shifted out becomes dst[2*size-1]. + + subl $1, %ecx C size-1 + xorl %eax, %eax C ready for final adcl, and clear carry + + movl %ecx, %edx + movl PARAM_SRC, %esi + + +L(lshift): + C eax + C ebx + C ecx counter, size-1 to 1 + C edx size-1 (for later use) + C esi src (for later use) + C edi dst, incrementing + C ebp + + rcll 4(%edi) + rcll 8(%edi) + + leal 8(%edi), %edi + decl %ecx + jnz L(lshift) + + + adcl %eax, %eax + + movl %eax, 4(%edi) C dst most significant limb + movl (%esi), %eax C src[0] + + leal 4(%esi,%edx,4), %esi C &src[size] + subl %edx, %ecx C -(size-1) + + +C Now add in the squares on the diagonal, src[0]^2, src[1]^2, ..., +C src[size-1]^2. dst[0] hasn't yet been set at all yet, and just gets the +C low limb of src[0]^2. + + + mull %eax + + movl %eax, (%edi,%ecx,8) C dst[0] + + +L(diag): + C eax scratch + C ebx scratch + C ecx counter, negative + C edx carry + C esi &src[size] + C edi dst[2*size-2] + C ebp + + movl (%esi,%ecx,4), %eax + movl %edx, %ebx + + mull %eax + + addl %ebx, 4(%edi,%ecx,8) + adcl %eax, 8(%edi,%ecx,8) + adcl $0, %edx + + incl %ecx + jnz L(diag) + + + movl SAVE_ESI, %esi + movl SAVE_EBX, %ebx + + addl %edx, 4(%edi) C dst most significant limb + + movl SAVE_EDI, %edi + movl SAVE_EBP, %ebp + addl $FRAME, %esp + ret + + + +C ----------------------------------------------------------------------------- +ifdef(`PIC',` +L(pic_calc): + addl (%esp), %ecx + addl $L(unroll_inner_end)-L(here)-eval(2*CODE_BYTES_PER_LIMB), %ecx + addl %edx, %ecx + ret_internal +') + + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/p6/sse2/addmul_1.asm b/gmp-6.3.0/mpn/x86/p6/sse2/addmul_1.asm new file mode 100644 index 0000000..144b627 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/sse2/addmul_1.asm @@ -0,0 +1,37 @@ +dnl Intel P6/SSE2 mpn_addmul_1. + +dnl Copyright 2008 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C TODO +C * Write P6 specific SSE2 code. + +MULFUNC_PROLOGUE(mpn_addmul_1) +include_mpn(`x86/pentium4/sse2/addmul_1.asm') diff --git a/gmp-6.3.0/mpn/x86/p6/sse2/gmp-mparam.h b/gmp-6.3.0/mpn/x86/p6/sse2/gmp-mparam.h new file mode 100644 index 0000000..a1e261b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/sse2/gmp-mparam.h @@ -0,0 +1,200 @@ +/* Intel P6/sse2 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991, 1993, 1994, 1999-2003, 2008-2010 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + + +/* NOTE: In a fat binary build SQR_TOOM2_THRESHOLD here cannot be more than the + value in mpn/x86/p6/gmp-mparam.h. The latter is used as a hard limit in + mpn/x86/p6/sqr_basecase.asm. */ + + +/* 1867 MHz P6 model 13 */ + +#define MOD_1_NORM_THRESHOLD 4 +#define MOD_1_UNNORM_THRESHOLD 4 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 5 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 4 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 11 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 8 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 21 + +#define MUL_TOOM22_THRESHOLD 20 +#define MUL_TOOM33_THRESHOLD 77 +#define MUL_TOOM44_THRESHOLD 169 +#define MUL_TOOM6H_THRESHOLD 246 +#define MUL_TOOM8H_THRESHOLD 381 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 73 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 114 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 97 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 80 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 106 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 30 +#define SQR_TOOM3_THRESHOLD 101 +#define SQR_TOOM4_THRESHOLD 154 +#define SQR_TOOM6_THRESHOLD 222 +#define SQR_TOOM8_THRESHOLD 527 + +#define MULMID_TOOM42_THRESHOLD 58 + +#define MULMOD_BNM1_THRESHOLD 13 +#define SQRMOD_BNM1_THRESHOLD 17 + +#define MUL_FFT_MODF_THRESHOLD 690 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 565, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 25, 7}, { 13, 6}, { 28, 7}, { 15, 6}, \ + { 31, 7}, { 17, 6}, { 35, 7}, { 27, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 47, 8}, { 27, 9}, { 15, 8}, \ + { 31, 7}, { 63, 8}, { 39, 9}, { 23, 5}, \ + { 383, 4}, { 991, 5}, { 511, 6}, { 267, 7}, \ + { 157, 8}, { 91, 9}, { 47, 8}, { 111, 9}, \ + { 63, 8}, { 127, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 135,10}, \ + { 79, 9}, { 159,10}, { 95,11}, { 63,10}, \ + { 143, 9}, { 287,10}, { 159,11}, { 95,10}, \ + { 191,12}, { 63,11}, { 127,10}, { 255, 9}, \ + { 511,10}, { 271, 9}, { 543,10}, { 287,11}, \ + { 159,10}, { 335, 9}, { 671,11}, { 191,10}, \ + { 383, 9}, { 767,10}, { 399, 9}, { 799,10}, \ + { 415,11}, { 223,12}, { 127,11}, { 255,10}, \ + { 543, 9}, { 1087,11}, { 287,10}, { 607,11}, \ + { 319,10}, { 671,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,10}, { 831,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,10}, \ + { 1215,12}, { 319,11}, { 671,10}, { 1343,11}, \ + { 735,10}, { 1471,12}, { 383,11}, { 799,10}, \ + { 1599,11}, { 863,12}, { 447,11}, { 959,13}, \ + { 255,12}, { 511,11}, { 1087,12}, { 575,11}, \ + { 1215,12}, { 639,11}, { 1343,12}, { 703,11}, \ + { 1471,13}, { 383,12}, { 831,11}, { 1727,12}, \ + { 959,14}, { 255,13}, { 511,12}, { 1215,13}, \ + { 639,12}, { 1471,11}, { 2943,13}, { 767,12}, \ + { 1727,13}, { 895,12}, { 1919,14}, { 511,13}, \ + { 1023,12}, { 2111,13}, { 1151,12}, { 2431,13}, \ + { 1407,12}, { 2815,14}, { 767,13}, { 1663,12}, \ + { 3455,13}, { 8192,14}, { 16384,15}, { 32768,16} } +#define MUL_FFT_TABLE3_SIZE 132 +#define MUL_FFT_THRESHOLD 7424 + +#define SQR_FFT_MODF_THRESHOLD 565 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 472, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 25, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \ + { 31, 7}, { 17, 6}, { 35, 7}, { 27, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 49, 8}, { 27, 9}, { 15, 8}, \ + { 39, 9}, { 23, 8}, { 51,10}, { 15, 9}, \ + { 31, 8}, { 63, 4}, { 1023, 8}, { 67, 9}, \ + { 39, 5}, { 639, 4}, { 1471, 6}, { 383, 7}, \ + { 209, 8}, { 119, 9}, { 63, 7}, { 255, 8}, \ + { 139, 9}, { 71, 8}, { 143, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159, 8}, { 319, 9}, \ + { 167,10}, { 95,11}, { 63,10}, { 143, 9}, \ + { 287,10}, { 159,11}, { 95,10}, { 191,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 543, 8}, \ + { 1087,10}, { 287, 9}, { 575,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335, 9}, { 671,10}, \ + { 351, 9}, { 703,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 399, 9}, { 799,10}, { 415, 9}, \ + { 831,11}, { 223,12}, { 127,11}, { 255,10}, \ + { 543, 9}, { 1087,11}, { 287,10}, { 607, 9}, \ + { 1215,11}, { 319,10}, { 671, 9}, { 1343,11}, \ + { 351,10}, { 703,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,10}, { 831,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,12}, \ + { 319,11}, { 671,10}, { 1343,11}, { 735,12}, \ + { 383,11}, { 799,10}, { 1599,11}, { 863,12}, \ + { 447,11}, { 959,13}, { 255,12}, { 511,11}, \ + { 1087,12}, { 575,11}, { 1215,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,12}, \ + { 959,14}, { 255,13}, { 511,12}, { 1215,13}, \ + { 639,12}, { 1471,13}, { 767,12}, { 1727,13}, \ + { 895,12}, { 1919,14}, { 511,13}, { 1023,12}, \ + { 2111,13}, { 1151,12}, { 2431,13}, { 1407,14}, \ + { 767,13}, { 1663,12}, { 3455,13}, { 8192,14}, \ + { 16384,15}, { 32768,16} } +#define SQR_FFT_TABLE3_SIZE 146 +#define SQR_FFT_THRESHOLD 5760 + +#define MULLO_BASECASE_THRESHOLD 0 /* always */ +#define MULLO_DC_THRESHOLD 31 +#define MULLO_MUL_N_THRESHOLD 13463 +#define SQRLO_BASECASE_THRESHOLD 0 /* always */ +#define SQRLO_DC_THRESHOLD 100 +#define SQRLO_SQR_THRESHOLD 9236 + +#define DC_DIV_QR_THRESHOLD 25 +#define DC_DIVAPPR_Q_THRESHOLD 55 +#define DC_BDIV_QR_THRESHOLD 60 +#define DC_BDIV_Q_THRESHOLD 132 + +#define INV_MULMOD_BNM1_THRESHOLD 38 +#define INV_NEWTON_THRESHOLD 65 +#define INV_APPR_THRESHOLD 65 + +#define BINV_NEWTON_THRESHOLD 252 +#define REDC_1_TO_REDC_N_THRESHOLD 62 + +#define MU_DIV_QR_THRESHOLD 1164 +#define MU_DIVAPPR_Q_THRESHOLD 748 +#define MUPI_DIV_QR_THRESHOLD 38 +#define MU_BDIV_QR_THRESHOLD 1360 +#define MU_BDIV_Q_THRESHOLD 1470 + +#define POWM_SEC_TABLE 2,23,258,879,2246 + +#define GET_STR_DC_THRESHOLD 13 +#define GET_STR_PRECOMPUTE_THRESHOLD 25 +#define SET_STR_DC_THRESHOLD 582 +#define SET_STR_PRECOMPUTE_THRESHOLD 1118 + +#define FAC_DSC_THRESHOLD 178 +#define FAC_ODD_THRESHOLD 34 + +#define MATRIX22_STRASSEN_THRESHOLD 17 +#define HGCD_THRESHOLD 69 +#define HGCD_APPR_THRESHOLD 112 +#define HGCD_REDUCE_THRESHOLD 3389 +#define GCD_DC_THRESHOLD 386 +#define GCDEXT_DC_THRESHOLD 303 +#define JACOBI_BASE_METHOD 1 diff --git a/gmp-6.3.0/mpn/x86/p6/sse2/mod_1_1.asm b/gmp-6.3.0/mpn/x86/p6/sse2/mod_1_1.asm new file mode 100644 index 0000000..8b7b7ad --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/sse2/mod_1_1.asm @@ -0,0 +1,34 @@ +dnl Intel P6/SSE2 mpn_mod_1_1. + +dnl Copyright 2009, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_mod_1_1p) +include_mpn(`x86/pentium4/sse2/mod_1_1.asm') diff --git a/gmp-6.3.0/mpn/x86/p6/sse2/mod_1_4.asm b/gmp-6.3.0/mpn/x86/p6/sse2/mod_1_4.asm new file mode 100644 index 0000000..49c96c6 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/sse2/mod_1_4.asm @@ -0,0 +1,34 @@ +dnl Intel P6/SSE2 mpn_mod_1_4. + +dnl Copyright 2009, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +MULFUNC_PROLOGUE(mpn_mod_1s_4p) +include_mpn(`x86/pentium4/sse2/mod_1_4.asm') diff --git a/gmp-6.3.0/mpn/x86/p6/sse2/mul_1.asm b/gmp-6.3.0/mpn/x86/p6/sse2/mul_1.asm new file mode 100644 index 0000000..50e5b69 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/sse2/mul_1.asm @@ -0,0 +1,38 @@ +dnl Intel P6/SSE2 mpn_mul_1. + +dnl Copyright 2008 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C TODO +C * Write P6 specific SSE2 code. It should reach 3 c/l. +C The Pentium4 code runs at 4.2 c/l. + +MULFUNC_PROLOGUE(mpn_mul_1) +include_mpn(`x86/pentium4/sse2/mul_1.asm') diff --git a/gmp-6.3.0/mpn/x86/p6/sse2/mul_basecase.asm b/gmp-6.3.0/mpn/x86/p6/sse2/mul_basecase.asm new file mode 100644 index 0000000..4687625 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/sse2/mul_basecase.asm @@ -0,0 +1,35 @@ +dnl Intel P6/SSE2 mpn_mul_basecase. + +dnl Copyright 2008 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +MULFUNC_PROLOGUE(mpn_mul_basecase) +include_mpn(`x86/pentium4/sse2/mul_basecase.asm') diff --git a/gmp-6.3.0/mpn/x86/p6/sse2/popcount.asm b/gmp-6.3.0/mpn/x86/p6/sse2/popcount.asm new file mode 100644 index 0000000..4c02b93 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/sse2/popcount.asm @@ -0,0 +1,35 @@ +dnl Intel P6/SSE2 mpn_popcount -- population count. + +dnl Copyright 2008 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +MULFUNC_PROLOGUE(mpn_popcount) +include_mpn(`x86/pentium4/sse2/popcount.asm') diff --git a/gmp-6.3.0/mpn/x86/p6/sse2/sqr_basecase.asm b/gmp-6.3.0/mpn/x86/p6/sse2/sqr_basecase.asm new file mode 100644 index 0000000..76b574b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/sse2/sqr_basecase.asm @@ -0,0 +1,35 @@ +dnl Intel P6/SSE2 mpn_sqr_basecase. + +dnl Copyright 2008 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +MULFUNC_PROLOGUE(mpn_sqr_basecase) +include_mpn(`x86/pentium4/sse2/sqr_basecase.asm') diff --git a/gmp-6.3.0/mpn/x86/p6/sse2/submul_1.asm b/gmp-6.3.0/mpn/x86/p6/sse2/submul_1.asm new file mode 100644 index 0000000..69d940d --- /dev/null +++ b/gmp-6.3.0/mpn/x86/p6/sse2/submul_1.asm @@ -0,0 +1,35 @@ +dnl Intel P6/SSE2 mpn_submul_1. + +dnl Copyright 2008 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +MULFUNC_PROLOGUE(mpn_submul_1) +include_mpn(`x86/pentium4/sse2/submul_1.asm') diff --git a/gmp-6.3.0/mpn/x86/pentium/README b/gmp-6.3.0/mpn/x86/pentium/README new file mode 100644 index 0000000..305936b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/README @@ -0,0 +1,181 @@ +Copyright 1996, 1999-2001, 2003 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. + + + + + + INTEL PENTIUM P5 MPN SUBROUTINES + + +This directory contains mpn functions optimized for Intel Pentium (P5,P54) +processors. The mmx subdirectory has additional code for Pentium with MMX +(P55). + + +STATUS + + cycles/limb + + mpn_add_n/sub_n 2.375 + + mpn_mul_1 12.0 + mpn_add/submul_1 14.0 + + mpn_mul_basecase 14.2 cycles/crossproduct (approx) + + mpn_sqr_basecase 8 cycles/crossproduct (approx) + or 15.5 cycles/triangleproduct (approx) + + mpn_l/rshift 5.375 normal (6.0 on P54) + 1.875 special shift by 1 bit + + mpn_divrem_1 44.0 + mpn_mod_1 28.0 + mpn_divexact_by3 15.0 + + mpn_copyi/copyd 1.0 + +Pentium MMX gets the following improvements + + mpn_l/rshift 1.75 + + mpn_mul_1 12.0 normal, 7.0 for 16-bit multiplier + + +mpn_add_n and mpn_sub_n run at asymptotically 2 cycles/limb. Due to loop +overhead and other delays (cache refill?), they run at or near 2.5 +cycles/limb. + +mpn_mul_1, mpn_addmul_1, mpn_submul_1 all run 1 cycle faster than they +should. Intel documentation says a mul instruction is 10 cycles, but it +measures 9 and the routines using it run as 9. + + + +P55 MMX AND X87 + +The cost of switching between MMX and x87 floating point on P55 is about 100 +cycles (fld1/por/emms for instance). In order to avoid that the two aren't +mixed and currently that means using MMX and not x87. + +MMX offers a big speedup for lshift and rshift, and a nice speedup for +16-bit multipliers in mpn_mul_1. If fast code using x87 is found then +perhaps the preference for MMX will be reversed. + + + + +P54 SHLDL + +mpn_lshift and mpn_rshift run at about 6 cycles/limb on P5 and P54, but the +documentation indicates that they should take only 43/8 = 5.375 cycles/limb, +or 5 cycles/limb asymptotically. The P55 runs them at the expected speed. + +It seems that on P54 a shldl or shrdl allows pairing in one following cycle, +but not two. For example, back to back repetitions of the following + + shldl( %cl, %eax, %ebx) + xorl %edx, %edx + xorl %esi, %esi + +run at 5 cycles, as expected, but repetitions of the following run at 7 +cycles, whereas 6 would be expected (and is achieved on P55), + + shldl( %cl, %eax, %ebx) + xorl %edx, %edx + xorl %esi, %esi + xorl %edi, %edi + xorl %ebp, %ebp + +Three xorls run at 7 cycles too, so it doesn't seem to be just that pairing +inhibited is only in the second following cycle (or something like that). + +Avoiding this problem would bring P54 shifts down from 6.0 c/l to 5.5 with a +pattern of shift, 2 loads, shift, 2 stores, shift, etc. A start has been +made on something like that, but it's not yet complete. + + + + +OTHER NOTES + +Prefetching Destinations + + Pentium doesn't allocate cache lines on writes, unlike most other modern + processors. Since the functions in the mpn class do array writes, we + have to handle allocating the destination cache lines by reading a word + from it in the loops, to achieve the best performance. + +Prefetching Sources + + Prefetching of sources is pointless since there's no out-of-order loads. + Any load instruction blocks until the line is brought to L1, so it may + as well be the load that wants the data which blocks. + +Data Cache Bank Clashes + + Pairing of memory operations requires that the two issued operations + refer to different cache banks (ie. different addresses modulo 32 + bytes). The simplest way to ensure this is to read/write two words from + the same object. If we make operations on different objects, they might + or might not be to the same cache bank. + +PIC %eip Fetching + + A simple call $+5 and popl can be used to get %eip, there's no need to + balance calls and returns since P5 doesn't have any return stack branch + prediction. + +Float Multiplies + + fmul is pairable and can be issued every 2 cycles (with a 4 cycle + latency for data ready to use). This is a lot better than integer mull + or imull at 9 cycles non-pairing. Unfortunately the advantage is + quickly eaten away by needing to throw data through memory back to the + integer registers to adjust for fild and fist being signed, and to do + things like propagating carry bits. + + + + + +REFERENCES + +"Intel Architecture Optimization Manual", 1997, order number 242816. This +is mostly about P5, the parts about P6 aren't relevant. Available on-line: + + http://download.intel.com/design/PentiumII/manuals/242816.htm + + + +---------------- +Local variables: +mode: text +fill-column: 76 +End: diff --git a/gmp-6.3.0/mpn/x86/pentium/aors_n.asm b/gmp-6.3.0/mpn/x86/pentium/aors_n.asm new file mode 100644 index 0000000..01ebfb9 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/aors_n.asm @@ -0,0 +1,203 @@ +dnl Intel Pentium mpn_add_n/mpn_sub_n -- mpn addition and subtraction. + +dnl Copyright 1992, 1994-1996, 1999, 2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 2.375 cycles/limb + + +ifdef(`OPERATION_add_n',` + define(M4_inst, adcl) + define(M4_function_n, mpn_add_n) + define(M4_function_nc, mpn_add_nc) + +',`ifdef(`OPERATION_sub_n',` + define(M4_inst, sbbl) + define(M4_function_n, mpn_sub_n) + define(M4_function_nc, mpn_sub_nc) + +',`m4_error(`Need OPERATION_add_n or OPERATION_sub_n +')')') + +MULFUNC_PROLOGUE(mpn_add_n mpn_add_nc mpn_sub_n mpn_sub_nc) + + +C mp_limb_t M4_function_n (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size); +C mp_limb_t M4_function_nc (mp_ptr dst, mp_srcptr src1, mp_srcptr src2, +C mp_size_t size, mp_limb_t carry); + +defframe(PARAM_CARRY,20) +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC2, 12) +defframe(PARAM_SRC1, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(M4_function_nc) + + pushl %edi + pushl %esi + pushl %ebx + pushl %ebp +deflit(`FRAME',16) + + movl PARAM_DST,%edi + movl PARAM_SRC1,%esi + movl PARAM_SRC2,%ebp + movl PARAM_SIZE,%ecx + + movl (%ebp),%ebx + + decl %ecx + movl %ecx,%edx + shrl $3,%ecx + andl $7,%edx + testl %ecx,%ecx C zero carry flag + jz L(endgo) + + pushl %edx +FRAME_pushl() + movl PARAM_CARRY,%eax + shrl %eax C shift bit 0 into carry + jmp L(oop) + +L(endgo): +deflit(`FRAME',16) + movl PARAM_CARRY,%eax + shrl %eax C shift bit 0 into carry + jmp L(end) + +EPILOGUE() + + + ALIGN(8) +PROLOGUE(M4_function_n) + + pushl %edi + pushl %esi + pushl %ebx + pushl %ebp +deflit(`FRAME',16) + + movl PARAM_DST,%edi + movl PARAM_SRC1,%esi + movl PARAM_SRC2,%ebp + movl PARAM_SIZE,%ecx + + movl (%ebp),%ebx + + decl %ecx + movl %ecx,%edx + shrl $3,%ecx + andl $7,%edx + testl %ecx,%ecx C zero carry flag + jz L(end) + pushl %edx +FRAME_pushl() + + ALIGN(8) +L(oop): movl 28(%edi),%eax C fetch destination cache line + leal 32(%edi),%edi + +L(1): movl (%esi),%eax + movl 4(%esi),%edx + M4_inst %ebx,%eax + movl 4(%ebp),%ebx + M4_inst %ebx,%edx + movl 8(%ebp),%ebx + movl %eax,-32(%edi) + movl %edx,-28(%edi) + +L(2): movl 8(%esi),%eax + movl 12(%esi),%edx + M4_inst %ebx,%eax + movl 12(%ebp),%ebx + M4_inst %ebx,%edx + movl 16(%ebp),%ebx + movl %eax,-24(%edi) + movl %edx,-20(%edi) + +L(3): movl 16(%esi),%eax + movl 20(%esi),%edx + M4_inst %ebx,%eax + movl 20(%ebp),%ebx + M4_inst %ebx,%edx + movl 24(%ebp),%ebx + movl %eax,-16(%edi) + movl %edx,-12(%edi) + +L(4): movl 24(%esi),%eax + movl 28(%esi),%edx + M4_inst %ebx,%eax + movl 28(%ebp),%ebx + M4_inst %ebx,%edx + movl 32(%ebp),%ebx + movl %eax,-8(%edi) + movl %edx,-4(%edi) + + leal 32(%esi),%esi + leal 32(%ebp),%ebp + decl %ecx + jnz L(oop) + + popl %edx +FRAME_popl() +L(end): + decl %edx C test %edx w/o clobbering carry + js L(end2) + incl %edx +L(oop2): + leal 4(%edi),%edi + movl (%esi),%eax + M4_inst %ebx,%eax + movl 4(%ebp),%ebx + movl %eax,-4(%edi) + leal 4(%esi),%esi + leal 4(%ebp),%ebp + decl %edx + jnz L(oop2) +L(end2): + movl (%esi),%eax + M4_inst %ebx,%eax + movl %eax,(%edi) + + sbbl %eax,%eax + negl %eax + + popl %ebp + popl %ebx + popl %esi + popl %edi + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/aorsmul_1.asm b/gmp-6.3.0/mpn/x86/pentium/aorsmul_1.asm new file mode 100644 index 0000000..d83cc45 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/aorsmul_1.asm @@ -0,0 +1,144 @@ +dnl Intel Pentium mpn_addmul_1 -- mpn by limb multiplication. + +dnl Copyright 1992, 1994, 1996, 1999, 2000, 2002 Free Software Foundation, +dnl Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 14.0 cycles/limb + + +ifdef(`OPERATION_addmul_1', ` + define(M4_inst, addl) + define(M4_function_1, mpn_addmul_1) + define(M4_function_1c, mpn_addmul_1c) + +',`ifdef(`OPERATION_submul_1', ` + define(M4_inst, subl) + define(M4_function_1, mpn_submul_1) + define(M4_function_1c, mpn_submul_1c) + +',`m4_error(`Need OPERATION_addmul_1 or OPERATION_submul_1 +')')') + +MULFUNC_PROLOGUE(mpn_addmul_1 mpn_addmul_1c mpn_submul_1 mpn_submul_1c) + + +C mp_limb_t mpn_addmul_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult); +C mp_limb_t mpn_addmul_1c (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult, mp_limb_t carry); +C +C mp_limb_t mpn_submul_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult); +C mp_limb_t mpn_submul_1c (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t mult, mp_limb_t carry); +C + +defframe(PARAM_CARRY, 20) +defframe(PARAM_MULTIPLIER,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + + ALIGN(8) +PROLOGUE(M4_function_1c) +deflit(`FRAME',0) + + movl PARAM_CARRY, %ecx + pushl %esi FRAME_pushl() + + jmp L(start_1c) + +EPILOGUE() + + + ALIGN(8) +PROLOGUE(M4_function_1) +deflit(`FRAME',0) + + xorl %ecx, %ecx + pushl %esi FRAME_pushl() + +L(start_1c): + movl PARAM_SRC, %esi + movl PARAM_SIZE, %eax + + pushl %edi FRAME_pushl() + pushl %ebx FRAME_pushl() + + movl PARAM_DST, %edi + leal -1(%eax), %ebx C size-1 + + leal (%esi,%eax,4), %esi + xorl $-1, %ebx C -size, and clear carry + + leal (%edi,%eax,4), %edi + +L(top): + C eax + C ebx counter, negative + C ecx carry + C edx + C esi src end + C edi dst end + C ebp + + adcl $0, %ecx + movl (%esi,%ebx,4), %eax + + mull PARAM_MULTIPLIER + + addl %ecx, %eax + movl (%edi,%ebx,4), %ecx + + adcl $0, %edx + M4_inst %eax, %ecx + + movl %ecx, (%edi,%ebx,4) + incl %ebx + + movl %edx, %ecx + jnz L(top) + + + adcl $0, %ecx + popl %ebx + + movl %ecx, %eax + popl %edi + + popl %esi + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/bdiv_q_1.asm b/gmp-6.3.0/mpn/x86/pentium/bdiv_q_1.asm new file mode 100644 index 0000000..c2c4f58 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/bdiv_q_1.asm @@ -0,0 +1,266 @@ +dnl Intel Pentium mpn_divexact_1 -- mpn by limb exact division. + +dnl Rearranged from mpn/x86/pentium/dive_1.asm by Marco Bodrato. + +dnl Copyright 2001, 2002, 2011, 2014 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C divisor +C odd even +C P54: 24.5 30.5 cycles/limb +C P55: 23.0 28.0 + +MULFUNC_PROLOGUE(mpn_bdiv_q_1 mpn_pi1_bdiv_q_1) + +C The P55 speeds noted above, 23 cycles odd or 28 cycles even, are as +C expected. On P54 in the even case the shrdl pairing nonsense (see +C mpn/x86/pentium/README) costs 1 cycle, but it's not clear why there's a +C further 1.5 slowdown for both odd and even. + +defframe(PARAM_SHIFT, 24) +defframe(PARAM_INVERSE,20) +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(VAR_INVERSE,`PARAM_DST') + + TEXT + + ALIGN(32) +C mp_limb_t mpn_bdiv_q_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C +PROLOGUE(mpn_bdiv_q_1) +deflit(`FRAME',0) + + movl $-1, %ecx + movl PARAM_DIVISOR, %eax + +L(strip_twos): + ASSERT(nz, `orl %eax, %eax') + shrl %eax + incl %ecx C shift count + + jnc L(strip_twos) + + leal 1(%eax,%eax), %edx C d + andl $127, %eax C d/2, 7 bits + + pushl %ebx FRAME_pushl() + pushl %ebp FRAME_pushl() + +ifdef(`PIC',` +ifdef(`DARWIN',` + LEA( binvert_limb_table, %ebp) + movzbl (%eax,%ebp), %eax +',` + call L(here) +L(here): + popl %ebp C eip + + addl $_GLOBAL_OFFSET_TABLE_+[.-L(here)], %ebp + C AGI + movl binvert_limb_table@GOT(%ebp), %ebp + C AGI + movzbl (%eax,%ebp), %eax +') +',` + +dnl non-PIC + movzbl binvert_limb_table(%eax), %eax C inv 8 bits +') + + movl %eax, %ebp C inv + addl %eax, %eax C 2*inv + + imull %ebp, %ebp C inv*inv + + imull %edx, %ebp C inv*inv*d + + subl %ebp, %eax C inv = 2*inv - inv*inv*d + movl PARAM_SIZE, %ebx + + movl %eax, %ebp + addl %eax, %eax C 2*inv + + imull %ebp, %ebp C inv*inv + + imull %edx, %ebp C inv*inv*d + + subl %ebp, %eax C inv = 2*inv - inv*inv*d + movl %edx, PARAM_DIVISOR C d without twos + + ASSERT(e,` C expect d*inv == 1 mod 2^GMP_LIMB_BITS + pushl %eax FRAME_pushl() + imull PARAM_DIVISOR, %eax + cmpl $1, %eax + popl %eax FRAME_popl()') + + jmp L(common) +EPILOGUE() + +C mp_limb_t +C mpn_pi1_bdiv_q_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, mp_limb_t divisor, +C mp_limb_t inverse, int shift) + ALIGN(32) +PROLOGUE(mpn_pi1_bdiv_q_1) +deflit(`FRAME',0) + + movl PARAM_SHIFT, %ecx + + pushl %ebx FRAME_pushl() + pushl %ebp FRAME_pushl() + + movl PARAM_SIZE, %ebx + movl PARAM_INVERSE, %eax + +L(common): + pushl %esi FRAME_pushl() + push %edi FRAME_pushl() + + movl PARAM_SRC, %esi + movl PARAM_DST, %edi + movl %eax, VAR_INVERSE + + leal (%esi,%ebx,4), %esi C src end + leal (%edi,%ebx,4), %edi C dst end + + negl %ebx C -size + + xorl %ebp, %ebp C initial carry bit + + orl %ecx, %ecx C shift + movl (%esi,%ebx,4), %eax C src low limb + jz L(odd_entry) + + xorl %edx, %edx C initial carry limb (for even, if one) + incl %ebx + jz L(one) + + movl (%esi,%ebx,4), %edx C src second limb (for even) + shrdl( %cl, %edx, %eax) + + jmp L(even_entry) + + + ALIGN(8) +L(odd_top): + C eax scratch + C ebx counter, limbs, negative + C ecx + C edx + C esi src end + C edi dst end + C ebp carry bit, 0 or -1 + + mull PARAM_DIVISOR + + movl (%esi,%ebx,4), %eax + subl %ebp, %edx + + subl %edx, %eax + + sbbl %ebp, %ebp + +L(odd_entry): + imull VAR_INVERSE, %eax + + movl %eax, (%edi,%ebx,4) + + incl %ebx + jnz L(odd_top) + + popl %edi + popl %esi + + popl %ebp + popl %ebx + + ret + +L(even_top): + C eax scratch + C ebx counter, limbs, negative + C ecx twos + C edx + C esi src end + C edi dst end + C ebp carry bit, 0 or -1 + + mull PARAM_DIVISOR + + subl %ebp, %edx C carry bit + movl -4(%esi,%ebx,4), %eax C src limb + + movl (%esi,%ebx,4), %ebp C and one above it + + shrdl( %cl, %ebp, %eax) + + subl %edx, %eax C carry limb + + sbbl %ebp, %ebp + +L(even_entry): + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi,%ebx,4) + incl %ebx + + jnz L(even_top) + + mull PARAM_DIVISOR + + movl -4(%esi), %eax C src high limb + subl %ebp, %edx + +L(one): + shrl %cl, %eax + + subl %edx, %eax C no carry if division is exact + + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi) C dst high limb + nop C protect against cache bank clash + + popl %edi + popl %esi + + popl %ebp + popl %ebx + + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/pentium/com.asm b/gmp-6.3.0/mpn/x86/pentium/com.asm new file mode 100644 index 0000000..b080545 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/com.asm @@ -0,0 +1,181 @@ +dnl Intel Pentium mpn_com -- mpn ones complement. + +dnl Copyright 1996, 2001, 2002, 2006 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 1.75 cycles/limb + + +NAILS_SUPPORT(0-31) + + +C void mpn_com (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C This code is similar to mpn_copyi, basically there's just some "xorl +C $GMP_NUMB_MASK"s inserted. +C +C Alternatives: +C +C On P55 some MMX code could be 1.25 c/l (8 limb unrolled) if src and dst +C are the same alignment mod 8, but it doesn't seem worth the trouble for +C just that case (there'd need to be some plain integer available too for +C the unaligned case). + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_com) +deflit(`FRAME',0) + + movl PARAM_SRC, %eax + movl PARAM_SIZE, %ecx + + pushl %esi FRAME_pushl() + pushl %edi FRAME_pushl() + + leal (%eax,%ecx,4), %eax + xorl $-1, %ecx C -size-1 + + movl PARAM_DST, %edx + addl $8, %ecx C -size+7 + + jns L(end) + + movl (%edx), %esi C fetch destination cache line + nop + +L(top): + C eax &src[size] + C ebx + C ecx counter, limbs, negative + C edx dst, incrementing + C esi scratch + C edi scratch + C ebp + + movl 28(%edx), %esi C destination prefetch + addl $32, %edx + + movl -28(%eax,%ecx,4), %esi + movl -24(%eax,%ecx,4), %edi + xorl $GMP_NUMB_MASK, %esi + xorl $GMP_NUMB_MASK, %edi + movl %esi, -32(%edx) + movl %edi, -28(%edx) + + movl -20(%eax,%ecx,4), %esi + movl -16(%eax,%ecx,4), %edi + xorl $GMP_NUMB_MASK, %esi + xorl $GMP_NUMB_MASK, %edi + movl %esi, -24(%edx) + movl %edi, -20(%edx) + + movl -12(%eax,%ecx,4), %esi + movl -8(%eax,%ecx,4), %edi + xorl $GMP_NUMB_MASK, %esi + xorl $GMP_NUMB_MASK, %edi + movl %esi, -16(%edx) + movl %edi, -12(%edx) + + movl -4(%eax,%ecx,4), %esi + movl (%eax,%ecx,4), %edi + xorl $GMP_NUMB_MASK, %esi + xorl $GMP_NUMB_MASK, %edi + movl %esi, -8(%edx) + movl %edi, -4(%edx) + + addl $8, %ecx + js L(top) + + +L(end): + C eax &src[size] + C ecx 0 to 7, representing respectively 7 to 0 limbs remaining + C edx dst, next location to store + + subl $4, %ecx + nop + + jns L(no4) + + movl -12(%eax,%ecx,4), %esi + movl -8(%eax,%ecx,4), %edi + xorl $GMP_NUMB_MASK, %esi + xorl $GMP_NUMB_MASK, %edi + movl %esi, (%edx) + movl %edi, 4(%edx) + + movl -4(%eax,%ecx,4), %esi + movl (%eax,%ecx,4), %edi + xorl $GMP_NUMB_MASK, %esi + xorl $GMP_NUMB_MASK, %edi + movl %esi, 8(%edx) + movl %edi, 12(%edx) + + addl $16, %edx + addl $4, %ecx +L(no4): + + subl $2, %ecx + nop + + jns L(no2) + + movl -4(%eax,%ecx,4), %esi + movl (%eax,%ecx,4), %edi + xorl $GMP_NUMB_MASK, %esi + xorl $GMP_NUMB_MASK, %edi + movl %esi, (%edx) + movl %edi, 4(%edx) + + addl $8, %edx + addl $2, %ecx +L(no2): + + popl %edi + jnz L(done) + + movl -4(%eax), %ecx + + xorl $GMP_NUMB_MASK, %ecx + popl %esi + + movl %ecx, (%edx) + ret + +L(done): + popl %esi + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/copyd.asm b/gmp-6.3.0/mpn/x86/pentium/copyd.asm new file mode 100644 index 0000000..72a543b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/copyd.asm @@ -0,0 +1,146 @@ +dnl Intel Pentium mpn_copyd -- copy limb vector, decrementing. + +dnl Copyright 1996, 2001, 2002, 2006 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 1.25 cycles/limb + + +C void mpn_copyd (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C See comments in copyi.asm. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_copyd) +deflit(`FRAME',0) + + movl PARAM_SRC, %eax + movl PARAM_SIZE, %ecx + + pushl %esi FRAME_pushl() + pushl %edi FRAME_pushl() + + leal -4(%eax,%ecx,4), %eax C &src[size-1] + movl PARAM_DST, %edx + + subl $7, %ecx C size-7 + jle L(end) + + movl 28-4(%edx,%ecx,4), %esi C prefetch cache, dst[size-1] + nop + +L(top): + C eax src, decrementing + C ebx + C ecx counter, limbs + C edx dst + C esi scratch + C edi scratch + C ebp + + movl 28-32(%edx,%ecx,4), %esi C prefetch dst cache line + subl $8, %ecx + + movl (%eax), %esi C read words pairwise + movl -4(%eax), %edi + movl %esi, 56(%edx,%ecx,4) C store words pairwise + movl %edi, 52(%edx,%ecx,4) + + movl -8(%eax), %esi + movl -12(%eax), %edi + movl %esi, 48(%edx,%ecx,4) + movl %edi, 44(%edx,%ecx,4) + + movl -16(%eax), %esi + movl -20(%eax), %edi + movl %esi, 40(%edx,%ecx,4) + movl %edi, 36(%edx,%ecx,4) + + movl -24(%eax), %esi + movl -28(%eax), %edi + movl %esi, 32(%edx,%ecx,4) + movl %edi, 28(%edx,%ecx,4) + + leal -32(%eax), %eax + jg L(top) + + +L(end): + C ecx -7 to 0, representing respectively 0 to 7 limbs remaining + C eax src end + C edx dst, next location to store + + addl $4, %ecx + jle L(no4) + + movl (%eax), %esi + movl -4(%eax), %edi + movl %esi, 8(%edx,%ecx,4) + movl %edi, 4(%edx,%ecx,4) + + movl -8(%eax), %esi + movl -12(%eax), %edi + movl %esi, (%edx,%ecx,4) + movl %edi, -4(%edx,%ecx,4) + + subl $16, %eax + subl $4, %ecx +L(no4): + + addl $2, %ecx + jle L(no2) + + movl (%eax), %esi + movl -4(%eax), %edi + movl %esi, (%edx,%ecx,4) + movl %edi, -4(%edx,%ecx,4) + + subl $8, %eax + subl $2, %ecx +L(no2): + + jnz L(done) + + movl (%eax), %ecx + movl %ecx, (%edx) C risk of cache bank clash here + +L(done): + popl %edi + popl %esi + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/copyi.asm b/gmp-6.3.0/mpn/x86/pentium/copyi.asm new file mode 100644 index 0000000..d983d6b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/copyi.asm @@ -0,0 +1,164 @@ +dnl Intel Pentium mpn_copyi -- copy limb vector, incrementing. + +dnl Copyright 1996, 2001, 2002, 2006 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 1.25 cycles/limb + + +C void mpn_copyi (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C Destination prefetching is done to avoid repeated write-throughs on lines +C not already in L1. +C +C At least one of the src or dst pointer needs to be incremented rather than +C using indexing, so that there's somewhere to put the loop control without +C an AGI. Incrementing one and not two lets us keep loop overhead to 2 +C cycles. Making it the src pointer incremented avoids an AGI on the %ecx +C subtracts in the finishup code. +C +C The block of finishup code is almost as big as the main loop itself, which +C is unfortunate, but it's faster that way than with say rep movsl, by about +C 10 cycles for instance on P55. +C +C There's nothing to be gained from MMX on P55, since it can do only one +C movq load (or store) per cycle, so the throughput would be the same as the +C code here (and even then only if src and dst have the same alignment mod +C 8). + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_copyi) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + movl PARAM_DST, %edx + + pushl %ebx FRAME_pushl() + pushl %esi FRAME_pushl() + + leal (%edx,%ecx,4), %edx C &dst[size-1] + xorl $-1, %ecx C -size-1 + + movl PARAM_SRC, %esi + addl $8, %ecx C -size+7 + + jns L(end) + + movl -28(%edx,%ecx,4), %eax C fetch destination cache line, dst[0] + nop + +L(top): + C eax scratch + C ebx scratch + C ecx counter, limbs, negative + C edx &dst[size-1] + C esi src, incrementing + C edi + C ebp + + movl (%edx,%ecx,4), %eax C fetch destination cache line + addl $8, %ecx + + movl (%esi), %eax C read words pairwise + movl 4(%esi), %ebx + movl %eax, -60(%edx,%ecx,4) C store words pairwise + movl %ebx, -56(%edx,%ecx,4) + + movl 8(%esi), %eax + movl 12(%esi), %ebx + movl %eax, -52(%edx,%ecx,4) + movl %ebx, -48(%edx,%ecx,4) + + movl 16(%esi), %eax + movl 20(%esi), %ebx + movl %eax, -44(%edx,%ecx,4) + movl %ebx, -40(%edx,%ecx,4) + + movl 24(%esi), %eax + movl 28(%esi), %ebx + movl %eax, -36(%edx,%ecx,4) + movl %ebx, -32(%edx,%ecx,4) + + leal 32(%esi), %esi + js L(top) + + +L(end): + C ecx 0 to 7, representing respectively 7 to 0 limbs remaining + C esi src end + C edx dst, next location to store + + subl $4, %ecx + jns L(no4) + + movl (%esi), %eax + movl 4(%esi), %ebx + movl %eax, -12(%edx,%ecx,4) + movl %ebx, -8(%edx,%ecx,4) + + movl 8(%esi), %eax + movl 12(%esi), %ebx + movl %eax, -4(%edx,%ecx,4) + movl %ebx, (%edx,%ecx,4) + + addl $16, %esi + addl $4, %ecx +L(no4): + + subl $2, %ecx + jns L(no2) + + movl (%esi), %eax + movl 4(%esi), %ebx + movl %eax, -4(%edx,%ecx,4) + movl %ebx, (%edx,%ecx,4) + + addl $8, %esi + addl $2, %ecx +L(no2): + + jnz L(done) + + movl (%esi), %eax + movl %eax, -4(%edx,%ecx,4) C risk of cache bank clash here + +L(done): + popl %esi + popl %ebx + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/dive_1.asm b/gmp-6.3.0/mpn/x86/pentium/dive_1.asm new file mode 100644 index 0000000..21b5287 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/dive_1.asm @@ -0,0 +1,264 @@ +dnl Intel Pentium mpn_divexact_1 -- mpn by limb exact division. + +dnl Copyright 2001, 2002, 2014 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C divisor +C odd even +C P54: 24.5 30.5 cycles/limb +C P55: 23.0 28.0 + + +C void mpn_divexact_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C +C Plain divl is used for small sizes, since the inverse takes a while to +C setup. Multiplying works out faster for size>=3 when the divisor is odd, +C or size>=4 when the divisor is even. Actually on P55 size==2 for odd or +C size==3 for even are about the same speed for both divl or mul, but the +C former is used since it will use up less code cache. +C +C The P55 speeds noted above, 23 cycles odd or 28 cycles even, are as +C expected. On P54 in the even case the shrdl pairing nonsense (see +C mpn/x86/pentium/README) costs 1 cycle, but it's not clear why there's a +C further 1.5 slowdown for both odd and even. + +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(VAR_INVERSE,`PARAM_DST') + + TEXT + + ALIGN(32) +PROLOGUE(mpn_divexact_1) +deflit(`FRAME',0) + + movl PARAM_DIVISOR, %eax + movl PARAM_SIZE, %ecx + + pushl %esi FRAME_pushl() + push %edi FRAME_pushl() + + movl PARAM_SRC, %esi + andl $1, %eax + + movl PARAM_DST, %edi + addl %ecx, %eax C size if even, size+1 if odd + + cmpl $4, %eax + jae L(mul_by_inverse) + + + xorl %edx, %edx +L(div_top): + movl -4(%esi,%ecx,4), %eax + + divl PARAM_DIVISOR + + movl %eax, -4(%edi,%ecx,4) + decl %ecx + + jnz L(div_top) + + popl %edi + popl %esi + + ret + + + +L(mul_by_inverse): + movl PARAM_DIVISOR, %eax + movl $-1, %ecx + +L(strip_twos): + ASSERT(nz, `orl %eax, %eax') + shrl %eax + incl %ecx C shift count + + jnc L(strip_twos) + + leal 1(%eax,%eax), %edx C d + andl $127, %eax C d/2, 7 bits + + pushl %ebx FRAME_pushl() + pushl %ebp FRAME_pushl() + +ifdef(`PIC',`dnl + LEA( binvert_limb_table, %ebp) + movzbl (%eax,%ebp), %eax C inv 8 bits +',` + movzbl binvert_limb_table(%eax), %eax C inv 8 bits +') + + movl %eax, %ebp C inv + addl %eax, %eax C 2*inv + + imull %ebp, %ebp C inv*inv + + imull %edx, %ebp C inv*inv*d + + subl %ebp, %eax C inv = 2*inv - inv*inv*d + movl PARAM_SIZE, %ebx + + movl %eax, %ebp + addl %eax, %eax C 2*inv + + imull %ebp, %ebp C inv*inv + + imull %edx, %ebp C inv*inv*d + + subl %ebp, %eax C inv = 2*inv - inv*inv*d + movl %edx, PARAM_DIVISOR C d without twos + + leal (%esi,%ebx,4), %esi C src end + leal (%edi,%ebx,4), %edi C dst end + + negl %ebx C -size + + ASSERT(e,` C expect d*inv == 1 mod 2^GMP_LIMB_BITS + pushl %eax FRAME_pushl() + imull PARAM_DIVISOR, %eax + cmpl $1, %eax + popl %eax FRAME_popl()') + + movl %eax, VAR_INVERSE + xorl %ebp, %ebp C initial carry bit + + movl (%esi,%ebx,4), %eax C src low limb + orl %ecx, %ecx C shift + + movl 4(%esi,%ebx,4), %edx C src second limb (for even) + jz L(odd_entry) + + shrdl( %cl, %edx, %eax) + + incl %ebx + jmp L(even_entry) + + + ALIGN(8) +L(odd_top): + C eax scratch + C ebx counter, limbs, negative + C ecx + C edx + C esi src end + C edi dst end + C ebp carry bit, 0 or -1 + + mull PARAM_DIVISOR + + movl (%esi,%ebx,4), %eax + subl %ebp, %edx + + subl %edx, %eax + + sbbl %ebp, %ebp + +L(odd_entry): + imull VAR_INVERSE, %eax + + movl %eax, (%edi,%ebx,4) + + incl %ebx + jnz L(odd_top) + + + popl %ebp + popl %ebx + + popl %edi + popl %esi + + ret + + +L(even_top): + C eax scratch + C ebx counter, limbs, negative + C ecx twos + C edx + C esi src end + C edi dst end + C ebp carry bit, 0 or -1 + + mull PARAM_DIVISOR + + subl %ebp, %edx C carry bit + movl -4(%esi,%ebx,4), %eax C src limb + + movl (%esi,%ebx,4), %ebp C and one above it + + shrdl( %cl, %ebp, %eax) + + subl %edx, %eax C carry limb + + sbbl %ebp, %ebp + +L(even_entry): + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi,%ebx,4) + incl %ebx + + jnz L(even_top) + + + + mull PARAM_DIVISOR + + movl -4(%esi), %eax C src high limb + subl %ebp, %edx + + shrl %cl, %eax + + subl %edx, %eax C no carry if division is exact + + imull VAR_INVERSE, %eax + + movl %eax, -4(%edi) C dst high limb + nop C protect against cache bank clash + + popl %ebp + popl %ebx + + popl %edi + popl %esi + + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/pentium/gmp-mparam.h b/gmp-6.3.0/mpn/x86/pentium/gmp-mparam.h new file mode 100644 index 0000000..befa6e2 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/gmp-mparam.h @@ -0,0 +1,76 @@ +/* Intel P54 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991, 1993, 1994, 1999-2002, 2004 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + + +/* For mpn/x86/pentium/mod_1.asm */ +#define COUNT_LEADING_ZEROS_NEED_CLZ_TAB + + +/* 166MHz P54 */ + +/* Generated by tuneup.c, 2004-02-10, gcc 2.95 */ + +#define MUL_TOOM22_THRESHOLD 16 +#define MUL_TOOM33_THRESHOLD 90 + +#define SQR_BASECASE_THRESHOLD 0 /* always */ +#define SQR_TOOM2_THRESHOLD 22 +#define SQR_TOOM3_THRESHOLD 122 + +#define DIV_SB_PREINV_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_DC_THRESHOLD 52 +#define POWM_THRESHOLD 77 + +#define HGCD_THRESHOLD 121 +#define GCD_ACCEL_THRESHOLD 3 +#define GCD_DC_THRESHOLD 615 +#define JACOBI_BASE_METHOD 2 + +#define USE_PREINV_DIVREM_1 0 +#define USE_PREINV_MOD_1 1 /* native */ +#define DIVREM_2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define MODEXACT_1_ODD_THRESHOLD 0 /* always (native) */ + +#define GET_STR_DC_THRESHOLD 23 +#define GET_STR_PRECOMPUTE_THRESHOLD 33 +#define SET_STR_THRESHOLD 2788 + +#define MUL_FFT_TABLE { 432, 928, 1664, 3584, 10240, 40960, 0 } +#define MUL_FFT_MODF_THRESHOLD 448 +#define MUL_FFT_THRESHOLD 3328 + +#define SQR_FFT_TABLE { 496, 928, 1920, 4608, 10240, 40960, 0 } +#define SQR_FFT_MODF_THRESHOLD 512 +#define SQR_FFT_THRESHOLD 3328 diff --git a/gmp-6.3.0/mpn/x86/pentium/hamdist.asm b/gmp-6.3.0/mpn/x86/pentium/hamdist.asm new file mode 100644 index 0000000..6c6c1a1 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/hamdist.asm @@ -0,0 +1,154 @@ +dnl Intel P5 mpn_hamdist -- mpn hamming distance. + +dnl Copyright 2001, 2002, 2014, 2015 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 14.0 cycles/limb + + +C unsigned long mpn_hamdist (mp_srcptr src1, mp_srcptr src2, mp_size_t size); +C +C It might be possible to shave 1 cycle from the loop, and hence 2 +C cycles/limb. The xorb is taking 2 cycles, but a separate load and xor +C would be 1, if the right schedule could be found (not found so far). +C Wanting to avoid potential cache bank clashes makes it tricky. + +C The slightly strange quoting here helps the renaming done by tune/many.pl. +deflit(TABLE_NAME, +m4_assert_defined(`GSYM_PREFIX') +GSYM_PREFIX`'mpn_popcount``'_table') + +C FIXME: referencing popcount.asm's table is incorrect as it hurt incremental +C linking. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC2, 8) +defframe(PARAM_SRC1, 4) + + TEXT + ALIGN(8) + +PROLOGUE(mpn_hamdist) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + pushl %esi FRAME_pushl() + + shll %ecx C size in byte pairs + pushl %edi FRAME_pushl() + +ifdef(`PIC',` + pushl %ebx FRAME_pushl() + pushl %ebp FRAME_pushl() +ifdef(`DARWIN',` + movl PARAM_SRC1, %esi + movl PARAM_SRC2, %edi + LEA( TABLE_NAME, %ebp) + xorl %ebx, %ebx C byte + xorl %edx, %edx C byte + xorl %eax, %eax C total +',` + call L(here) FRAME_pushl() +L(here): + movl PARAM_SRC1, %esi + popl %ebp FRAME_popl() + + movl PARAM_SRC2, %edi + addl $_GLOBAL_OFFSET_TABLE_+[.-L(here)], %ebp + + xorl %ebx, %ebx C byte + xorl %edx, %edx C byte + + movl TABLE_NAME@GOT(%ebp), %ebp + xorl %eax, %eax C total +') +define(TABLE,`(%ebp,$1)') +',` +dnl non-PIC + movl PARAM_SRC1, %esi + movl PARAM_SRC2, %edi + + xorl %eax, %eax C total + pushl %ebx FRAME_pushl() + + xorl %edx, %edx C byte + xorl %ebx, %ebx C byte + +define(TABLE,`TABLE_NAME($1)') +') + + + C The nop after the xorb seems necessary. Although a movb might be + C expected to go down the V pipe in the second cycle of the xorb, it + C doesn't and costs an extra 2 cycles. +L(top): + C eax total + C ebx byte + C ecx counter, 2*size to 2 + C edx byte + C esi src1 + C edi src2 + C ebp [PIC] table + + addl %ebx, %eax + movb -1(%esi,%ecx,2), %bl + + addl %edx, %eax + movb -1(%edi,%ecx,2), %dl + + xorb %dl, %bl + movb -2(%esi,%ecx,2), %dl + + xorb -2(%edi,%ecx,2), %dl + nop + + movb TABLE(%ebx), %bl + decl %ecx + + movb TABLE(%edx), %dl + jnz L(top) + + +ifdef(`PIC',` + popl %ebp +') + addl %ebx, %eax + popl %ebx + + addl %edx, %eax + popl %edi + + popl %esi + + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/pentium/logops_n.asm b/gmp-6.3.0/mpn/x86/pentium/logops_n.asm new file mode 100644 index 0000000..1877317 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/logops_n.asm @@ -0,0 +1,176 @@ +dnl Intel Pentium mpn_and_n,...,mpn_xnor_n -- bitwise logical operations. + +dnl Copyright 2001, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 3.0 c/l and, ior, xor +C 3.5 c/l andn, iorn, nand, nior, xnor + + +define(M4_choose_op, +`ifdef(`OPERATION_$1',` +define(`M4_function', `mpn_$1') +define(`M4_want_pre', `$4') +define(`M4op', `$3') +define(`M4_want_post',`$2') +')') +define(M4pre, `ifelse(M4_want_pre, yes,`$1')') +define(M4post,`ifelse(M4_want_post,yes,`$1')') + +M4_choose_op( and_n, , andl, ) +M4_choose_op( andn_n, , andl, yes) +M4_choose_op( nand_n, yes, andl, ) +M4_choose_op( ior_n, , orl, ) +M4_choose_op( iorn_n, , orl, yes) +M4_choose_op( nior_n, yes, orl, ) +M4_choose_op( xor_n, , xorl, ) +M4_choose_op( xnor_n, yes, xorl, ) + +ifdef(`M4_function',, +`m4_error(`Unrecognised or undefined OPERATION symbol +')') + +MULFUNC_PROLOGUE(mpn_and_n mpn_andn_n mpn_nand_n mpn_ior_n mpn_iorn_n mpn_nior_n mpn_xor_n mpn_xnor_n) + +NAILS_SUPPORT(0-31) + + +C void M4_function (mp_ptr wp, mp_srcptr xp, mp_srcptr yp, mp_size_t size); +C +C Nothing complicated here, just some care to avoid data cache bank clashes +C and AGIs. +C +C We're one register short of being able to do a simple 4 loads, 2 ops, 2 +C stores. Instead %ebp is juggled a bit and nops are introduced to keep the +C pairings as intended. An in-place operation would free up a register, for +C an 0.5 c/l speedup, if that's worth bothering with. +C +C This code seems best for P55 too. Data alignment is a big problem for MMX +C and the pairing restrictions on movq and integer instructions make life +C difficult. + +defframe(PARAM_SIZE,16) +defframe(PARAM_YP, 12) +defframe(PARAM_XP, 8) +defframe(PARAM_WP, 4) + + TEXT + ALIGN(8) + +PROLOGUE(M4_function) +deflit(`FRAME',0) + + pushl %ebx FRAME_pushl() + pushl %esi FRAME_pushl() + + pushl %edi FRAME_pushl() + pushl %ebp FRAME_pushl() + + movl PARAM_SIZE, %ecx + movl PARAM_XP, %ebx + + movl PARAM_YP, %esi + movl PARAM_WP, %edi + + shrl %ecx + jnc L(entry) + + movl (%ebx,%ecx,8), %eax C risk of data cache bank clash here + movl (%esi,%ecx,8), %edx + +M4pre(` notl_or_xorl_GMP_NUMB_MASK(%edx)') + + M4op %edx, %eax + +M4post(`xorl $GMP_NUMB_MASK, %eax') + orl %ecx, %ecx + + movl %eax, (%edi,%ecx,8) + jz L(done) + + jmp L(entry) + + +L(top): + C eax + C ebx xp + C ecx counter, limb pairs, decrementing + C edx + C esi yp + C edi wp + C ebp + + M4op %ebp, %edx + nop + +M4post(`xorl $GMP_NUMB_MASK, %eax') +M4post(`xorl $GMP_NUMB_MASK, %edx') + + movl %eax, 4(%edi,%ecx,8) + movl %edx, (%edi,%ecx,8) + +L(entry): + movl -4(%ebx,%ecx,8), %ebp + nop + + movl -4(%esi,%ecx,8), %eax + movl -8(%esi,%ecx,8), %edx + +M4pre(` xorl $GMP_NUMB_MASK, %eax') +M4pre(` xorl $GMP_NUMB_MASK, %edx') + + M4op %ebp, %eax + movl -8(%ebx,%ecx,8), %ebp + + decl %ecx + jnz L(top) + + + M4op %ebp, %edx + nop + +M4post(`xorl $GMP_NUMB_MASK, %eax') +M4post(`xorl $GMP_NUMB_MASK, %edx') + + movl %eax, 4(%edi,%ecx,8) + movl %edx, (%edi,%ecx,8) + + +L(done): + popl %ebp + popl %edi + + popl %esi + popl %ebx + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/lshift.asm b/gmp-6.3.0/mpn/x86/pentium/lshift.asm new file mode 100644 index 0000000..2a31f36 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/lshift.asm @@ -0,0 +1,243 @@ +dnl Intel Pentium mpn_lshift -- mpn left shift. + +dnl Copyright 1992, 1994-1996, 1999, 2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P5,P54: 6.0 +C P55: 5.375 + + +C mp_limb_t mpn_lshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned shift); +C +C The main shift-by-N loop should run at 5.375 c/l and that's what P55 does, +C but P5 and P54 run only at 6.0 c/l, which is 4 cycles lost somewhere. + +defframe(PARAM_SHIFT,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_lshift) + + pushl %edi + pushl %esi + pushl %ebx + pushl %ebp +deflit(`FRAME',16) + + movl PARAM_DST,%edi + movl PARAM_SRC,%esi + movl PARAM_SIZE,%ebp + movl PARAM_SHIFT,%ecx + +C We can use faster code for shift-by-1 under certain conditions. + cmp $1,%ecx + jne L(normal) + leal 4(%esi),%eax + cmpl %edi,%eax + jnc L(special) C jump if s_ptr + 1 >= res_ptr + leal (%esi,%ebp,4),%eax + cmpl %eax,%edi + jnc L(special) C jump if res_ptr >= s_ptr + size + +L(normal): + leal -4(%edi,%ebp,4),%edi + leal -4(%esi,%ebp,4),%esi + + movl (%esi),%edx + subl $4,%esi + xorl %eax,%eax + shldl( %cl, %edx, %eax) C compute carry limb + pushl %eax C push carry limb onto stack + + decl %ebp + pushl %ebp + shrl $3,%ebp + jz L(end) + + movl (%edi),%eax C fetch destination cache line + + ALIGN(4) +L(oop): movl -28(%edi),%eax C fetch destination cache line + movl %edx,%ebx + + movl (%esi),%eax + movl -4(%esi),%edx + shldl( %cl, %eax, %ebx) + shldl( %cl, %edx, %eax) + movl %ebx,(%edi) + movl %eax,-4(%edi) + + movl -8(%esi),%ebx + movl -12(%esi),%eax + shldl( %cl, %ebx, %edx) + shldl( %cl, %eax, %ebx) + movl %edx,-8(%edi) + movl %ebx,-12(%edi) + + movl -16(%esi),%edx + movl -20(%esi),%ebx + shldl( %cl, %edx, %eax) + shldl( %cl, %ebx, %edx) + movl %eax,-16(%edi) + movl %edx,-20(%edi) + + movl -24(%esi),%eax + movl -28(%esi),%edx + shldl( %cl, %eax, %ebx) + shldl( %cl, %edx, %eax) + movl %ebx,-24(%edi) + movl %eax,-28(%edi) + + subl $32,%esi + subl $32,%edi + decl %ebp + jnz L(oop) + +L(end): popl %ebp + andl $7,%ebp + jz L(end2) +L(oop2): + movl (%esi),%eax + shldl( %cl,%eax,%edx) + movl %edx,(%edi) + movl %eax,%edx + subl $4,%esi + subl $4,%edi + decl %ebp + jnz L(oop2) + +L(end2): + shll %cl,%edx C compute least significant limb + movl %edx,(%edi) C store it + + popl %eax C pop carry limb + + popl %ebp + popl %ebx + popl %esi + popl %edi + ret + + +C We loop from least significant end of the arrays, which is only +C permissable if the source and destination don't overlap, since the +C function is documented to work for overlapping source and destination. + +L(special): + movl (%esi),%edx + addl $4,%esi + + decl %ebp + pushl %ebp + shrl $3,%ebp + + addl %edx,%edx + incl %ebp + decl %ebp + jz L(Lend) + + movl (%edi),%eax C fetch destination cache line + + ALIGN(4) +L(Loop): + movl 28(%edi),%eax C fetch destination cache line + movl %edx,%ebx + + movl (%esi),%eax + movl 4(%esi),%edx + adcl %eax,%eax + movl %ebx,(%edi) + adcl %edx,%edx + movl %eax,4(%edi) + + movl 8(%esi),%ebx + movl 12(%esi),%eax + adcl %ebx,%ebx + movl %edx,8(%edi) + adcl %eax,%eax + movl %ebx,12(%edi) + + movl 16(%esi),%edx + movl 20(%esi),%ebx + adcl %edx,%edx + movl %eax,16(%edi) + adcl %ebx,%ebx + movl %edx,20(%edi) + + movl 24(%esi),%eax + movl 28(%esi),%edx + adcl %eax,%eax + movl %ebx,24(%edi) + adcl %edx,%edx + movl %eax,28(%edi) + + leal 32(%esi),%esi C use leal not to clobber carry + leal 32(%edi),%edi + decl %ebp + jnz L(Loop) + +L(Lend): + popl %ebp + sbbl %eax,%eax C save carry in %eax + andl $7,%ebp + jz L(Lend2) + addl %eax,%eax C restore carry from eax +L(Loop2): + movl %edx,%ebx + movl (%esi),%edx + adcl %edx,%edx + movl %ebx,(%edi) + + leal 4(%esi),%esi C use leal not to clobber carry + leal 4(%edi),%edi + decl %ebp + jnz L(Loop2) + + jmp L(L1) +L(Lend2): + addl %eax,%eax C restore carry from eax +L(L1): movl %edx,(%edi) C store last limb + + sbbl %eax,%eax + negl %eax + + popl %ebp + popl %ebx + popl %esi + popl %edi + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/mmx/gmp-mparam.h b/gmp-6.3.0/mpn/x86/pentium/mmx/gmp-mparam.h new file mode 100644 index 0000000..02a0def --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/mmx/gmp-mparam.h @@ -0,0 +1,163 @@ +/* Intel P55 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 1991, 1993, 1994, 1999-2002, 2004, 2009, 2010 Free Software +Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + + +/* For mpn/x86/pentium/mod_1.asm */ +#define COUNT_LEADING_ZEROS_NEED_CLZ_TAB + + +/* 233MHz P55 */ + +#define MOD_1_NORM_THRESHOLD 5 +#define MOD_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define MOD_1N_TO_MOD_1_1_THRESHOLD MP_SIZE_T_MAX /* never */ +#define MOD_1U_TO_MOD_1_1_THRESHOLD 12 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 0 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 11 +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 63 +#define USE_PREINV_DIVREM_1 0 +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 51 + +#define MUL_TOOM22_THRESHOLD 16 +#define MUL_TOOM33_THRESHOLD 53 +#define MUL_TOOM44_THRESHOLD 128 +#define MUL_TOOM6H_THRESHOLD 189 +#define MUL_TOOM8H_THRESHOLD 260 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 89 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 91 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 90 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 88 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 20 +#define SQR_TOOM3_THRESHOLD 73 +#define SQR_TOOM4_THRESHOLD 178 +#define SQR_TOOM6_THRESHOLD 210 +#define SQR_TOOM8_THRESHOLD 375 + +#define MULMOD_BNM1_THRESHOLD 11 +#define SQRMOD_BNM1_THRESHOLD 12 + +#define MUL_FFT_MODF_THRESHOLD 364 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 364, 5}, { 15, 6}, { 8, 5}, { 17, 6}, \ + { 9, 5}, { 19, 6}, { 17, 7}, { 9, 6}, \ + { 21, 7}, { 11, 6}, { 23, 7}, { 15, 6}, \ + { 31, 7}, { 21, 8}, { 11, 7}, { 27, 8}, \ + { 15, 7}, { 33, 8}, { 19, 7}, { 39, 8}, \ + { 23, 7}, { 47, 8}, { 27, 9}, { 15, 8}, \ + { 31, 7}, { 63, 8}, { 39, 9}, { 23, 8}, \ + { 47,10}, { 15, 9}, { 31, 8}, { 67, 9}, \ + { 39, 8}, { 79, 9}, { 47, 8}, { 95, 9}, \ + { 55,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 135,10}, \ + { 79, 9}, { 159, 8}, { 319, 9}, { 167,10}, \ + { 95, 9}, { 191, 8}, { 383,11}, { 63,10}, \ + { 127, 9}, { 255,10}, { 143, 9}, { 287,10}, \ + { 159, 9}, { 319,11}, { 95,10}, { 191, 9}, \ + { 383,12}, { 63,11}, { 127,10}, { 271, 9}, \ + { 543,10}, { 287,11}, { 159,10}, { 351,11}, \ + { 191,10}, { 415,11}, { 223,12}, { 127,11}, \ + { 255,10}, { 511,11}, { 287,10}, { 575,11}, \ + { 351,12}, { 191,11}, { 415,13}, { 127,12}, \ + { 255,11}, { 575,12}, { 319,11}, { 703,12}, \ + { 383,11}, { 831,12}, { 447,13}, { 8192,14}, \ + { 16384,15}, { 32768,16} } +#define MUL_FFT_TABLE3_SIZE 90 +#define MUL_FFT_THRESHOLD 3520 + +#define SQR_FFT_MODF_THRESHOLD 340 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 340, 5}, { 17, 6}, { 9, 5}, { 19, 6}, \ + { 17, 7}, { 9, 6}, { 21, 7}, { 11, 6}, \ + { 23, 7}, { 15, 6}, { 31, 7}, { 21, 8}, \ + { 11, 7}, { 29, 8}, { 15, 7}, { 33, 8}, \ + { 19, 7}, { 39, 8}, { 27, 7}, { 55, 9}, \ + { 15, 8}, { 31, 7}, { 65, 8}, { 43, 9}, \ + { 23, 8}, { 47,10}, { 15, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 83, 9}, { 47, 8}, \ + { 95,10}, { 31, 9}, { 63, 8}, { 127, 9}, \ + { 79,10}, { 47, 9}, { 95,11}, { 31,10}, \ + { 63, 9}, { 127, 8}, { 255, 9}, { 135,10}, \ + { 79, 9}, { 159, 8}, { 319,10}, { 95, 9}, \ + { 191,11}, { 63,10}, { 127, 9}, { 255, 8}, \ + { 511, 9}, { 271,10}, { 143, 9}, { 287, 8}, \ + { 575, 9}, { 303,10}, { 159, 9}, { 319,11}, \ + { 95,10}, { 191, 9}, { 383,10}, { 207,12}, \ + { 63,11}, { 127,10}, { 271, 9}, { 543,10}, \ + { 287, 9}, { 575,10}, { 303,11}, { 159,10}, \ + { 351,11}, { 191,10}, { 415,11}, { 223,10}, \ + { 447,12}, { 127,11}, { 255,10}, { 543,11}, \ + { 287,10}, { 607,11}, { 351,12}, { 191,11}, \ + { 479,13}, { 127,12}, { 255,11}, { 575,12}, \ + { 319,11}, { 703,12}, { 383,11}, { 767,12}, \ + { 447,13}, { 8192,14}, { 16384,15}, { 32768,16} } +#define SQR_FFT_TABLE3_SIZE 96 +#define SQR_FFT_THRESHOLD 5504 + +#define MULLO_BASECASE_THRESHOLD 0 /* always */ +#define MULLO_DC_THRESHOLD 48 +#define MULLO_MUL_N_THRESHOLD 6633 + +#define DC_DIV_QR_THRESHOLD 43 +#define DC_DIVAPPR_Q_THRESHOLD 170 +#define DC_BDIV_QR_THRESHOLD 43 +#define DC_BDIV_Q_THRESHOLD 110 + +#define INV_MULMOD_BNM1_THRESHOLD 30 +#define INV_NEWTON_THRESHOLD 177 +#define INV_APPR_THRESHOLD 171 + +#define BINV_NEWTON_THRESHOLD 194 +#define REDC_1_TO_REDC_N_THRESHOLD 50 + +#define MU_DIV_QR_THRESHOLD 1142 +#define MU_DIVAPPR_Q_THRESHOLD 1142 +#define MUPI_DIV_QR_THRESHOLD 90 +#define MU_BDIV_QR_THRESHOLD 942 +#define MU_BDIV_Q_THRESHOLD 1017 + +#define MATRIX22_STRASSEN_THRESHOLD 13 +#define HGCD_THRESHOLD 92 +#define GCD_DC_THRESHOLD 283 +#define GCDEXT_DC_THRESHOLD 221 +#define JACOBI_BASE_METHOD 2 + +#define GET_STR_DC_THRESHOLD 18 +#define GET_STR_PRECOMPUTE_THRESHOLD 31 +#define SET_STR_DC_THRESHOLD 490 +#define SET_STR_PRECOMPUTE_THRESHOLD 994 diff --git a/gmp-6.3.0/mpn/x86/pentium/mmx/hamdist.asm b/gmp-6.3.0/mpn/x86/pentium/mmx/hamdist.asm new file mode 100644 index 0000000..72e3196 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/mmx/hamdist.asm @@ -0,0 +1,40 @@ +dnl Intel P55 mpn_hamdist -- mpn hamming distance. + +dnl Copyright 2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P55: hamdist 12.0 cycles/limb + +C For reference, this code runs at 11.5 cycles/limb for popcount, which is +C slower than the plain integer mpn/x86/pentium/popcount.asm. + +MULFUNC_PROLOGUE(mpn_hamdist) +include_mpn(`x86/k6/mmx/popham.asm') diff --git a/gmp-6.3.0/mpn/x86/pentium/mmx/lshift.asm b/gmp-6.3.0/mpn/x86/pentium/mmx/lshift.asm new file mode 100644 index 0000000..04b0ddc --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/mmx/lshift.asm @@ -0,0 +1,463 @@ +dnl Intel P5 mpn_lshift -- mpn left shift. + +dnl Copyright 2000-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 1.75 cycles/limb. + + +C mp_limb_t mpn_lshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned shift); +C +C Shift src,size left by shift many bits and store the result in dst,size. +C Zeros are shifted in at the right. Return the bits shifted out at the +C left. +C +C The comments in mpn_rshift apply here too. + +defframe(PARAM_SHIFT,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) +deflit(`FRAME',0) + +dnl minimum 5, because the unrolled loop can't handle less +deflit(UNROLL_THRESHOLD, 5) + + TEXT + ALIGN(8) + +PROLOGUE(mpn_lshift) + + pushl %ebx + pushl %edi +deflit(`FRAME',8) + + movl PARAM_SIZE, %eax + movl PARAM_DST, %edx + + movl PARAM_SRC, %ebx + movl PARAM_SHIFT, %ecx + + cmp $UNROLL_THRESHOLD, %eax + jae L(unroll) + + movl -4(%ebx,%eax,4), %edi C src high limb + decl %eax + + jnz L(simple) + + shldl( %cl, %edi, %eax) C eax was decremented to zero + + shll %cl, %edi + + movl %edi, (%edx) C dst low limb + popl %edi C risk of data cache bank clash + + popl %ebx + + ret + + +C ----------------------------------------------------------------------------- +L(simple): + C eax size-1 + C ebx src + C ecx shift + C edx dst + C esi + C edi + C ebp +deflit(`FRAME',8) + + movd (%ebx,%eax,4), %mm5 C src high limb + + movd %ecx, %mm6 C lshift + negl %ecx + + psllq %mm6, %mm5 + addl $32, %ecx + + movd %ecx, %mm7 + psrlq $32, %mm5 C retval + + +L(simple_top): + C eax counter, limbs, negative + C ebx src + C ecx + C edx dst + C esi + C edi + C + C mm0 scratch + C mm5 return value + C mm6 shift + C mm7 32-shift + + movq -4(%ebx,%eax,4), %mm0 + decl %eax + + psrlq %mm7, %mm0 + + C + + movd %mm0, 4(%edx,%eax,4) + jnz L(simple_top) + + + movd (%ebx), %mm0 + + movd %mm5, %eax + psllq %mm6, %mm0 + + popl %edi + popl %ebx + + movd %mm0, (%edx) + + emms + + ret + + +C ----------------------------------------------------------------------------- + ALIGN(8) +L(unroll): + C eax size + C ebx src + C ecx shift + C edx dst + C esi + C edi + C ebp +deflit(`FRAME',8) + + movd -4(%ebx,%eax,4), %mm5 C src high limb + leal (%ebx,%eax,4), %edi + + movd %ecx, %mm6 C lshift + andl $4, %edi + + psllq %mm6, %mm5 + jz L(start_src_aligned) + + + C src isn't aligned, process high limb separately (marked xxx) to + C make it so. + C + C source -8(ebx,%eax,4) + C | + C +-------+-------+-------+-- + C | | + C +-------+-------+-------+-- + C 0mod8 4mod8 0mod8 + C + C dest + C -4(edx,%eax,4) + C | + C +-------+-------+-- + C | xxx | | + C +-------+-------+-- + + movq -8(%ebx,%eax,4), %mm0 C unaligned load + + psllq %mm6, %mm0 + decl %eax + + psrlq $32, %mm0 + + C + + movd %mm0, (%edx,%eax,4) +L(start_src_aligned): + + movq -8(%ebx,%eax,4), %mm1 C src high qword + leal (%edx,%eax,4), %edi + + andl $4, %edi + psrlq $32, %mm5 C return value + + movq -16(%ebx,%eax,4), %mm3 C src second highest qword + jz L(start_dst_aligned) + + C dst isn't aligned, subtract 4 to make it so, and pretend the shift + C is 32 bits extra. High limb of dst (marked xxx) handled here + C separately. + C + C source -8(ebx,%eax,4) + C | + C +-------+-------+-- + C | mm1 | + C +-------+-------+-- + C 0mod8 4mod8 + C + C dest + C -4(edx,%eax,4) + C | + C +-------+-------+-------+-- + C | xxx | | + C +-------+-------+-------+-- + C 0mod8 4mod8 0mod8 + + movq %mm1, %mm0 + addl $32, %ecx C new shift + + psllq %mm6, %mm0 + + movd %ecx, %mm6 + psrlq $32, %mm0 + + C wasted cycle here waiting for %mm0 + + movd %mm0, -4(%edx,%eax,4) + subl $4, %edx +L(start_dst_aligned): + + + psllq %mm6, %mm1 + negl %ecx C -shift + + addl $64, %ecx C 64-shift + movq %mm3, %mm2 + + movd %ecx, %mm7 + subl $8, %eax C size-8 + + psrlq %mm7, %mm3 + + por %mm1, %mm3 C mm3 ready to store + jc L(finish) + + + C The comments in mpn_rshift apply here too. + + ALIGN(8) +L(unroll_loop): + C eax counter, limbs + C ebx src + C ecx + C edx dst + C esi + C edi + C + C mm0 + C mm1 + C mm2 src qword from 16(%ebx,%eax,4) + C mm3 dst qword ready to store to 24(%edx,%eax,4) + C + C mm5 return value + C mm6 lshift + C mm7 rshift + + movq 8(%ebx,%eax,4), %mm0 + psllq %mm6, %mm2 + + movq %mm0, %mm1 + psrlq %mm7, %mm0 + + movq %mm3, 24(%edx,%eax,4) C prev + por %mm2, %mm0 + + movq (%ebx,%eax,4), %mm3 C + psllq %mm6, %mm1 C + + movq %mm0, 16(%edx,%eax,4) + movq %mm3, %mm2 C + + psrlq %mm7, %mm3 C + subl $4, %eax + + por %mm1, %mm3 C + jnc L(unroll_loop) + + + +L(finish): + C eax -4 to -1 representing respectively 0 to 3 limbs remaining + + testb $2, %al + + jz L(finish_no_two) + + movq 8(%ebx,%eax,4), %mm0 + psllq %mm6, %mm2 + + movq %mm0, %mm1 + psrlq %mm7, %mm0 + + movq %mm3, 24(%edx,%eax,4) C prev + por %mm2, %mm0 + + movq %mm1, %mm2 + movq %mm0, %mm3 + + subl $2, %eax +L(finish_no_two): + + + C eax -4 or -3 representing respectively 0 or 1 limbs remaining + C + C mm2 src prev qword, from 16(%ebx,%eax,4) + C mm3 dst qword, for 24(%edx,%eax,4) + + testb $1, %al + movd %mm5, %eax C retval + + popl %edi + jz L(finish_zero) + + + C One extra src limb, destination was aligned. + C + C source ebx + C --+---------------+-------+ + C | mm2 | | + C --+---------------+-------+ + C + C dest edx+12 edx+4 edx + C --+---------------+---------------+-------+ + C | mm3 | | | + C --+---------------+---------------+-------+ + C + C mm6 = shift + C mm7 = ecx = 64-shift + + + C One extra src limb, destination was unaligned. + C + C source ebx + C --+---------------+-------+ + C | mm2 | | + C --+---------------+-------+ + C + C dest edx+12 edx+4 + C --+---------------+---------------+ + C | mm3 | | + C --+---------------+---------------+ + C + C mm6 = shift+32 + C mm7 = ecx = 64-(shift+32) + + + C In both cases there's one extra limb of src to fetch and combine + C with mm2 to make a qword at 4(%edx), and in the aligned case + C there's an extra limb of dst to be formed from that extra src limb + C left shifted. + + + movd (%ebx), %mm0 + psllq %mm6, %mm2 + + movq %mm3, 12(%edx) + psllq $32, %mm0 + + movq %mm0, %mm1 + psrlq %mm7, %mm0 + + por %mm2, %mm0 + psllq %mm6, %mm1 + + movq %mm0, 4(%edx) + psrlq $32, %mm1 + + andl $32, %ecx + popl %ebx + + jz L(finish_one_unaligned) + + movd %mm1, (%edx) +L(finish_one_unaligned): + + emms + + ret + + +L(finish_zero): + + C No extra src limbs, destination was aligned. + C + C source ebx + C --+---------------+ + C | mm2 | + C --+---------------+ + C + C dest edx+8 edx + C --+---------------+---------------+ + C | mm3 | | + C --+---------------+---------------+ + C + C mm6 = shift + C mm7 = ecx = 64-shift + + + C No extra src limbs, destination was unaligned. + C + C source ebx + C --+---------------+ + C | mm2 | + C --+---------------+ + C + C dest edx+8 edx+4 + C --+---------------+-------+ + C | mm3 | | + C --+---------------+-------+ + C + C mm6 = shift+32 + C mm7 = ecx = 64-(shift+32) + + + C The movd for the unaligned case writes the same data to 4(%edx) + C that the movq does for the aligned case. + + + movq %mm3, 8(%edx) + andl $32, %ecx + + psllq %mm6, %mm2 + jz L(finish_zero_unaligned) + + movq %mm2, (%edx) +L(finish_zero_unaligned): + + psrlq $32, %mm2 + popl %ebx + + movd %mm5, %eax C retval + + movd %mm2, 4(%edx) + + emms + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/mmx/mul_1.asm b/gmp-6.3.0/mpn/x86/pentium/mmx/mul_1.asm new file mode 100644 index 0000000..4ced577 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/mmx/mul_1.asm @@ -0,0 +1,371 @@ +dnl Intel Pentium MMX mpn_mul_1 -- mpn by limb multiplication. + +dnl Copyright 2000-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P5: 12.0 for 32-bit multiplier +C 7.0 for 16-bit multiplier + + +C mp_limb_t mpn_mul_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t multiplier); +C +C When the multiplier is 16 bits some special case MMX code is used. Small +C multipliers might arise reasonably often from mpz_mul_ui etc. If the size +C is odd there's roughly a 5 cycle penalty, so times for say size==7 and +C size==8 end up being quite close. If src isn't aligned to an 8 byte +C boundary then one limb is processed separately with roughly a 5 cycle +C penalty, so in that case it's say size==8 and size==9 which are close. +C +C Alternatives: +C +C MMX is not believed to be of any use for 32-bit multipliers, since for +C instance the current method would just have to be more or less duplicated +C for the high and low halves of the multiplier, and would probably +C therefore run at about 14 cycles, which is slower than the plain integer +C at 12. +C +C Adding the high and low MMX products using integer code seems best. An +C attempt at using paddd and carry bit propagation with pcmpgtd didn't give +C any joy. Perhaps something could be done keeping the values signed and +C thereby avoiding adjustments to make pcmpgtd into an unsigned compare, or +C perhaps not. +C +C Future: +C +C An mpn_mul_1c entrypoint would need a double carry out of the low result +C limb in the 16-bit code, unless it could be assumed the carry fits in 16 +C bits, possibly as carry=a + + divl PARAM_DIVISOR + + movl %edx, %eax + + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/pentium/mul_1.asm b/gmp-6.3.0/mpn/x86/pentium/mul_1.asm new file mode 100644 index 0000000..a0858af --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/mul_1.asm @@ -0,0 +1,177 @@ +dnl Intel Pentium mpn_mul_1 -- mpn by limb multiplication. + +dnl Copyright 1992, 1994, 1996, 1999, 2000, 2002 Free Software Foundation, +dnl Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 12.0 cycles/limb + + +C mp_limb_t mpn_mul_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t multiplier); +C mp_limb_t mpn_mul_1c (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t multiplier, mp_limb_t carry); +C + +defframe(PARAM_CARRY, 20) +defframe(PARAM_MULTIPLIER,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_mul_1c) +deflit(`FRAME',0) + + movl PARAM_CARRY, %ecx + pushl %esi FRAME_pushl() + + jmp L(start_1c) + +EPILOGUE() + + + ALIGN(8) +PROLOGUE(mpn_mul_1) +deflit(`FRAME',0) + + xorl %ecx, %ecx + pushl %esi FRAME_pushl() + +L(start_1c): + movl PARAM_SRC, %esi + movl PARAM_SIZE, %eax + + shrl %eax + jnz L(two_or_more) + + + C one limb only + + movl (%esi), %eax + + mull PARAM_MULTIPLIER + + addl %eax, %ecx + movl PARAM_DST, %eax + + adcl $0, %edx + popl %esi + + movl %ecx, (%eax) + movl %edx, %eax + + ret + + +L(two_or_more): + C eax size/2 + C ebx + C ecx carry + C edx + C esi src + C edi + C ebp + + pushl %edi FRAME_pushl() + pushl %ebx FRAME_pushl() + + movl PARAM_DST, %edi + leal -1(%eax), %ebx C size/2-1 + + notl %ebx C -size, preserve carry + + leal (%esi,%eax,8), %esi C src end + leal (%edi,%eax,8), %edi C dst end + + pushl %ebp FRAME_pushl() + jnc L(top) + + + C size was odd, process one limb separately + + movl (%esi,%ebx,8), %eax + addl $4, %esi + + mull PARAM_MULTIPLIER + + addl %ecx, %eax + movl %edx, %ecx + + movl %eax, (%edi,%ebx,8) + leal 4(%edi), %edi + + +L(top): + C eax + C ebx counter, negative + C ecx carry + C edx + C esi src end + C edi dst end + C ebp + + adcl $0, %ecx + movl (%esi,%ebx,8), %eax + + mull PARAM_MULTIPLIER + + movl %edx, %ebp + addl %eax, %ecx + + adcl $0, %ebp + movl 4(%esi,%ebx,8), %eax + + mull PARAM_MULTIPLIER + + movl %ecx, (%edi,%ebx,8) + addl %ebp, %eax + + movl %eax, 4(%edi,%ebx,8) + incl %ebx + + movl %edx, %ecx + jnz L(top) + + + adcl $0, %ecx + popl %ebp + + movl %ecx, %eax + popl %ebx + + popl %edi + popl %esi + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/mul_2.asm b/gmp-6.3.0/mpn/x86/pentium/mul_2.asm new file mode 100644 index 0000000..4c7beb5 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/mul_2.asm @@ -0,0 +1,150 @@ +dnl Intel Pentium mpn_mul_2 -- mpn by 2-limb multiplication. + +dnl Copyright 2001, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 24.0 cycles/limb + + +C mp_limb_t mpn_mul_2 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_srcptr mult); +C +C At 24 c/l this is only 2 cycles faster than a separate mul_1 and addmul_1, +C but has the advantage of making just one pass over the operands. +C +C There's not enough registers to use PARAM_MULT directly, so the multiplier +C limbs are transferred to local variables on the stack. + +defframe(PARAM_MULT, 16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(VAR_MULT_LOW, `PARAM_SRC') +define(VAR_MULT_HIGH,`PARAM_DST') + + TEXT + ALIGN(8) +PROLOGUE(mpn_mul_2) +deflit(`FRAME',0) + + pushl %esi FRAME_pushl() + pushl %edi FRAME_pushl() + + movl PARAM_SRC, %esi + movl PARAM_DST, %edi + + movl PARAM_MULT, %eax + movl PARAM_SIZE, %ecx + + movl 4(%eax), %edx C mult high + movl (%eax), %eax C mult low + + movl %eax, VAR_MULT_LOW + movl %edx, VAR_MULT_HIGH + + pushl %ebx FRAME_pushl() + pushl %ebp FRAME_pushl() + + mull (%esi) C src[0] * mult[0] + + movl %eax, %ebp C in case src==dst + movl (%esi), %eax C src[0] + + movl %ebp, (%edi) C dst[0] + movl %edx, %ebx C initial low carry + + xorl %ebp, %ebp C initial high carry + leal (%edi,%ecx,4), %edi C dst end + + mull VAR_MULT_HIGH C src[0] * mult[1] + + subl $2, %ecx C size-2 + js L(done) + + leal 8(%esi,%ecx,4), %esi C &src[size] + xorl $-1, %ecx C -(size-1) + + + +L(top): + C eax low prod + C ebx low carry + C ecx counter, negative + C edx high prod + C esi src end + C edi dst end + C ebp high carry (0 or -1) + + andl $1, %ebp C 1 or 0 + addl %eax, %ebx + + adcl %edx, %ebp + ASSERT(nc) + movl (%esi,%ecx,4), %eax + + mull VAR_MULT_LOW + + addl %eax, %ebx C low carry + movl (%esi,%ecx,4), %eax + + adcl %ebp, %edx C high carry + movl %ebx, (%edi,%ecx,4) + + sbbl %ebp, %ebp C new high carry, -1 or 0 + movl %edx, %ebx C new low carry + + mull VAR_MULT_HIGH + + incl %ecx + jnz L(top) + + +L(done): + andl $1, %ebp C 1 or 0 + addl %ebx, %eax + + adcl %ebp, %edx + ASSERT(nc) + movl %eax, (%edi) C store carry low + + movl %edx, %eax C return carry high + + popl %ebp + popl %ebx + + popl %edi + popl %esi + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/mul_basecase.asm b/gmp-6.3.0/mpn/x86/pentium/mul_basecase.asm new file mode 100644 index 0000000..e1d0f05 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/mul_basecase.asm @@ -0,0 +1,142 @@ +dnl Intel Pentium mpn_mul_basecase -- mpn by mpn multiplication. + +dnl Copyright 1996, 1998-2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 14.2 cycles/crossproduct (approx) + + +C void mpn_mul_basecase (mp_ptr wp, +C mp_srcptr xp, mp_size_t xsize, +C mp_srcptr yp, mp_size_t ysize); + +defframe(PARAM_YSIZE, 20) +defframe(PARAM_YP, 16) +defframe(PARAM_XSIZE, 12) +defframe(PARAM_XP, 8) +defframe(PARAM_WP, 4) + +defframe(VAR_COUNTER, -4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_mul_basecase) + + pushl %eax C dummy push for allocating stack slot + pushl %esi + pushl %ebp + pushl %edi +deflit(`FRAME',16) + + movl PARAM_XP,%esi + movl PARAM_WP,%edi + movl PARAM_YP,%ebp + + movl (%esi),%eax C load xp[0] + mull (%ebp) C multiply by yp[0] + movl %eax,(%edi) C store to wp[0] + movl PARAM_XSIZE,%ecx C xsize + decl %ecx C If xsize = 1, ysize = 1 too + jz L(done) + + movl PARAM_XSIZE,%eax + pushl %ebx +FRAME_pushl() + movl %edx,%ebx + leal (%esi,%eax,4),%esi C make xp point at end + leal (%edi,%eax,4),%edi C offset wp by xsize + negl %ecx C negate j size/index for inner loop + xorl %eax,%eax C clear carry + + ALIGN(8) +L(oop1): adcl $0,%ebx + movl (%esi,%ecx,4),%eax C load next limb at xp[j] + mull (%ebp) + addl %ebx,%eax + movl %eax,(%edi,%ecx,4) + incl %ecx + movl %edx,%ebx + jnz L(oop1) + + adcl $0,%ebx + movl PARAM_YSIZE,%eax + movl %ebx,(%edi) C most significant limb of product + addl $4,%edi C increment wp + decl %eax + jz L(skip) + movl %eax,VAR_COUNTER C set index i to ysize + +L(outer): + addl $4,%ebp C make ebp point to next y limb + movl PARAM_XSIZE,%ecx + negl %ecx + xorl %ebx,%ebx + + C code at 0x61 here, close enough to aligned +L(oop2): + adcl $0,%ebx + movl (%esi,%ecx,4),%eax + mull (%ebp) + addl %ebx,%eax + movl (%edi,%ecx,4),%ebx + adcl $0,%edx + addl %eax,%ebx + movl %ebx,(%edi,%ecx,4) + incl %ecx + movl %edx,%ebx + jnz L(oop2) + + adcl $0,%ebx + + movl %ebx,(%edi) + addl $4,%edi + movl VAR_COUNTER,%eax + decl %eax + movl %eax,VAR_COUNTER + jnz L(outer) + +L(skip): + popl %ebx + popl %edi + popl %ebp + popl %esi + addl $4,%esp + ret + +L(done): + movl %edx,4(%edi) C store to wp[1] + popl %edi + popl %ebp + popl %esi + popl %eax C dummy pop for deallocating stack slot + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/popcount.asm b/gmp-6.3.0/mpn/x86/pentium/popcount.asm new file mode 100644 index 0000000..0e82144 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/popcount.asm @@ -0,0 +1,146 @@ +dnl Intel P5 mpn_popcount -- mpn bit population count. + +dnl Copyright 2001, 2002, 2014, 2015 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: 8.0 cycles/limb + + +C unsigned long mpn_popcount (mp_srcptr src, mp_size_t size); +C +C An arithmetic approach has been found to be slower than the table lookup, +C due to needing too many instructions. + +C The slightly strange quoting here helps the renaming done by tune/many.pl. +deflit(TABLE_NAME, +m4_assert_defined(`GSYM_PREFIX') +GSYM_PREFIX`'mpn_popcount``'_table') + +C FIXME: exporting the table to hamdist is incorrect as it hurt incremental +C linking. + + RODATA + ALIGN(8) + GLOBL TABLE_NAME +TABLE_NAME: +forloop(i,0,255, +` .byte m4_popcount(i) +') + +defframe(PARAM_SIZE,8) +defframe(PARAM_SRC, 4) + + TEXT + ALIGN(8) + +PROLOGUE(mpn_popcount) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + pushl %esi FRAME_pushl() + +ifdef(`PIC',` + pushl %ebx FRAME_pushl() + pushl %ebp FRAME_pushl() +ifdef(`DARWIN',` + shll %ecx C size in byte pairs + LEA( TABLE_NAME, %ebp) + movl PARAM_SRC, %esi + xorl %eax, %eax C total + xorl %ebx, %ebx C byte + xorl %edx, %edx C byte +',` + call L(here) +L(here): + popl %ebp + shll %ecx C size in byte pairs + + addl $_GLOBAL_OFFSET_TABLE_+[.-L(here)], %ebp + movl PARAM_SRC, %esi + + xorl %eax, %eax C total + xorl %ebx, %ebx C byte + + movl TABLE_NAME@GOT(%ebp), %ebp + xorl %edx, %edx C byte +') +define(TABLE,`(%ebp,$1)') +',` +dnl non-PIC + shll %ecx C size in byte pairs + movl PARAM_SRC, %esi + + pushl %ebx FRAME_pushl() + xorl %eax, %eax C total + + xorl %ebx, %ebx C byte + xorl %edx, %edx C byte + +define(TABLE,`TABLE_NAME`'($1)') +') + + + ALIGN(8) C necessary on P55 for claimed speed +L(top): + C eax total + C ebx byte + C ecx counter, 2*size to 2 + C edx byte + C esi src + C edi + C ebp [PIC] table + + addl %ebx, %eax + movb -1(%esi,%ecx,2), %bl + + addl %edx, %eax + movb -2(%esi,%ecx,2), %dl + + movb TABLE(%ebx), %bl + decl %ecx + + movb TABLE(%edx), %dl + jnz L(top) + + +ifdef(`PIC',` + popl %ebp +') + addl %ebx, %eax + popl %ebx + + addl %edx, %eax + popl %esi + + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/pentium/rshift.asm b/gmp-6.3.0/mpn/x86/pentium/rshift.asm new file mode 100644 index 0000000..2105c4c --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/rshift.asm @@ -0,0 +1,243 @@ +dnl Intel Pentium mpn_rshift -- mpn right shift. + +dnl Copyright 1992, 1994-1996, 1999, 2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P5,P54: 6.0 +C P55: 5.375 + + +C mp_limb_t mpn_rshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned shift); +C +C The main shift-by-N loop should run at 5.375 c/l and that's what P55 does, +C but P5 and P54 run only at 6.0 c/l, which is 4 cycles lost somewhere. + +defframe(PARAM_SHIFT,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_rshift) + + pushl %edi + pushl %esi + pushl %ebx + pushl %ebp +deflit(`FRAME',16) + + movl PARAM_DST,%edi + movl PARAM_SRC,%esi + movl PARAM_SIZE,%ebp + movl PARAM_SHIFT,%ecx + +C We can use faster code for shift-by-1 under certain conditions. + cmp $1,%ecx + jne L(normal) + leal 4(%edi),%eax + cmpl %esi,%eax + jnc L(special) C jump if res_ptr + 1 >= s_ptr + leal (%edi,%ebp,4),%eax + cmpl %eax,%esi + jnc L(special) C jump if s_ptr >= res_ptr + size + +L(normal): + movl (%esi),%edx + addl $4,%esi + xorl %eax,%eax + shrdl( %cl, %edx, %eax) C compute carry limb + pushl %eax C push carry limb onto stack + + decl %ebp + pushl %ebp + shrl $3,%ebp + jz L(end) + + movl (%edi),%eax C fetch destination cache line + + ALIGN(4) +L(oop): movl 28(%edi),%eax C fetch destination cache line + movl %edx,%ebx + + movl (%esi),%eax + movl 4(%esi),%edx + shrdl( %cl, %eax, %ebx) + shrdl( %cl, %edx, %eax) + movl %ebx,(%edi) + movl %eax,4(%edi) + + movl 8(%esi),%ebx + movl 12(%esi),%eax + shrdl( %cl, %ebx, %edx) + shrdl( %cl, %eax, %ebx) + movl %edx,8(%edi) + movl %ebx,12(%edi) + + movl 16(%esi),%edx + movl 20(%esi),%ebx + shrdl( %cl, %edx, %eax) + shrdl( %cl, %ebx, %edx) + movl %eax,16(%edi) + movl %edx,20(%edi) + + movl 24(%esi),%eax + movl 28(%esi),%edx + shrdl( %cl, %eax, %ebx) + shrdl( %cl, %edx, %eax) + movl %ebx,24(%edi) + movl %eax,28(%edi) + + addl $32,%esi + addl $32,%edi + decl %ebp + jnz L(oop) + +L(end): popl %ebp + andl $7,%ebp + jz L(end2) +L(oop2): + movl (%esi),%eax + shrdl( %cl,%eax,%edx) C compute result limb + movl %edx,(%edi) + movl %eax,%edx + addl $4,%esi + addl $4,%edi + decl %ebp + jnz L(oop2) + +L(end2): + shrl %cl,%edx C compute most significant limb + movl %edx,(%edi) C store it + + popl %eax C pop carry limb + + popl %ebp + popl %ebx + popl %esi + popl %edi + ret + + +C We loop from least significant end of the arrays, which is only +C permissable if the source and destination don't overlap, since the +C function is documented to work for overlapping source and destination. + +L(special): + leal -4(%edi,%ebp,4),%edi + leal -4(%esi,%ebp,4),%esi + + movl (%esi),%edx + subl $4,%esi + + decl %ebp + pushl %ebp + shrl $3,%ebp + + shrl %edx + incl %ebp + decl %ebp + jz L(Lend) + + movl (%edi),%eax C fetch destination cache line + + ALIGN(4) +L(Loop): + movl -28(%edi),%eax C fetch destination cache line + movl %edx,%ebx + + movl (%esi),%eax + movl -4(%esi),%edx + rcrl %eax + movl %ebx,(%edi) + rcrl %edx + movl %eax,-4(%edi) + + movl -8(%esi),%ebx + movl -12(%esi),%eax + rcrl %ebx + movl %edx,-8(%edi) + rcrl %eax + movl %ebx,-12(%edi) + + movl -16(%esi),%edx + movl -20(%esi),%ebx + rcrl %edx + movl %eax,-16(%edi) + rcrl %ebx + movl %edx,-20(%edi) + + movl -24(%esi),%eax + movl -28(%esi),%edx + rcrl %eax + movl %ebx,-24(%edi) + rcrl %edx + movl %eax,-28(%edi) + + leal -32(%esi),%esi C use leal not to clobber carry + leal -32(%edi),%edi + decl %ebp + jnz L(Loop) + +L(Lend): + popl %ebp + sbbl %eax,%eax C save carry in %eax + andl $7,%ebp + jz L(Lend2) + addl %eax,%eax C restore carry from eax +L(Loop2): + movl %edx,%ebx + movl (%esi),%edx + rcrl %edx + movl %ebx,(%edi) + + leal -4(%esi),%esi C use leal not to clobber carry + leal -4(%edi),%edi + decl %ebp + jnz L(Loop2) + + jmp L(L1) +L(Lend2): + addl %eax,%eax C restore carry from eax +L(L1): movl %edx,(%edi) C store last limb + + movl $0,%eax + rcrl %eax + + popl %ebp + popl %ebx + popl %esi + popl %edi + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium/sqr_basecase.asm b/gmp-6.3.0/mpn/x86/pentium/sqr_basecase.asm new file mode 100644 index 0000000..b11d767 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium/sqr_basecase.asm @@ -0,0 +1,528 @@ +dnl Intel P5 mpn_sqr_basecase -- square an mpn number. + +dnl Copyright 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P5: approx 8 cycles per crossproduct, or 15.5 cycles per triangular +C product at around 20x20 limbs. + + +C void mpn_sqr_basecase (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C Calculate src,size squared, storing the result in dst,2*size. +C +C The algorithm is basically the same as mpn/generic/sqr_basecase.c, but a +C lot of function call overheads are avoided, especially when the size is +C small. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_sqr_basecase) +deflit(`FRAME',0) + + movl PARAM_SIZE, %edx + movl PARAM_SRC, %eax + + cmpl $2, %edx + movl PARAM_DST, %ecx + + je L(two_limbs) + + movl (%eax), %eax + ja L(three_or_more) + +C ----------------------------------------------------------------------------- +C one limb only + C eax src + C ebx + C ecx dst + C edx + + mull %eax + + movl %eax, (%ecx) + movl %edx, 4(%ecx) + + ret + +C ----------------------------------------------------------------------------- + ALIGN(8) +L(two_limbs): + C eax src + C ebx + C ecx dst + C edx size + + pushl %ebp + pushl %edi + + pushl %esi + pushl %ebx + + movl %eax, %ebx + movl (%eax), %eax + + mull %eax C src[0]^2 + + movl %eax, (%ecx) C dst[0] + movl %edx, %esi C dst[1] + + movl 4(%ebx), %eax + + mull %eax C src[1]^2 + + movl %eax, %edi C dst[2] + movl %edx, %ebp C dst[3] + + movl (%ebx), %eax + + mull 4(%ebx) C src[0]*src[1] + + addl %eax, %esi + popl %ebx + + adcl %edx, %edi + + adcl $0, %ebp + addl %esi, %eax + + adcl %edi, %edx + movl %eax, 4(%ecx) + + adcl $0, %ebp + popl %esi + + movl %edx, 8(%ecx) + movl %ebp, 12(%ecx) + + popl %edi + popl %ebp + + ret + + +C ----------------------------------------------------------------------------- + ALIGN(8) +L(three_or_more): + C eax src low limb + C ebx + C ecx dst + C edx size + + cmpl $4, %edx + pushl %ebx +deflit(`FRAME',4) + + movl PARAM_SRC, %ebx + jae L(four_or_more) + + +C ----------------------------------------------------------------------------- +C three limbs + C eax src low limb + C ebx src + C ecx dst + C edx size + + pushl %ebp + pushl %edi + + mull %eax C src[0] ^ 2 + + movl %eax, (%ecx) + movl %edx, 4(%ecx) + + movl 4(%ebx), %eax + xorl %ebp, %ebp + + mull %eax C src[1] ^ 2 + + movl %eax, 8(%ecx) + movl %edx, 12(%ecx) + + movl 8(%ebx), %eax + pushl %esi C risk of cache bank clash + + mull %eax C src[2] ^ 2 + + movl %eax, 16(%ecx) + movl %edx, 20(%ecx) + + movl (%ebx), %eax + + mull 4(%ebx) C src[0] * src[1] + + movl %eax, %esi + movl %edx, %edi + + movl (%ebx), %eax + + mull 8(%ebx) C src[0] * src[2] + + addl %eax, %edi + movl %edx, %ebp + + adcl $0, %ebp + movl 4(%ebx), %eax + + mull 8(%ebx) C src[1] * src[2] + + xorl %ebx, %ebx + addl %eax, %ebp + + C eax + C ebx zero, will be dst[5] + C ecx dst + C edx dst[4] + C esi dst[1] + C edi dst[2] + C ebp dst[3] + + adcl $0, %edx + addl %esi, %esi + + adcl %edi, %edi + + adcl %ebp, %ebp + + adcl %edx, %edx + movl 4(%ecx), %eax + + adcl $0, %ebx + addl %esi, %eax + + movl %eax, 4(%ecx) + movl 8(%ecx), %eax + + adcl %edi, %eax + movl 12(%ecx), %esi + + adcl %ebp, %esi + movl 16(%ecx), %edi + + movl %eax, 8(%ecx) + movl %esi, 12(%ecx) + + adcl %edx, %edi + popl %esi + + movl 20(%ecx), %eax + movl %edi, 16(%ecx) + + popl %edi + popl %ebp + + adcl %ebx, %eax C no carry out of this + popl %ebx + + movl %eax, 20(%ecx) + + ret + + +C ----------------------------------------------------------------------------- + ALIGN(8) +L(four_or_more): + C eax src low limb + C ebx src + C ecx dst + C edx size + C esi + C edi + C ebp + C + C First multiply src[0]*src[1..size-1] and store at dst[1..size]. + +deflit(`FRAME',4) + + pushl %edi +FRAME_pushl() + pushl %esi +FRAME_pushl() + + pushl %ebp +FRAME_pushl() + leal (%ecx,%edx,4), %edi C dst end of this mul1 + + leal (%ebx,%edx,4), %esi C src end + movl %ebx, %ebp C src + + negl %edx C -size + xorl %ebx, %ebx C clear carry limb and carry flag + + leal 1(%edx), %ecx C -(size-1) + +L(mul1): + C eax scratch + C ebx carry + C ecx counter, negative + C edx scratch + C esi &src[size] + C edi &dst[size] + C ebp src + + adcl $0, %ebx + movl (%esi,%ecx,4), %eax + + mull (%ebp) + + addl %eax, %ebx + + movl %ebx, (%edi,%ecx,4) + incl %ecx + + movl %edx, %ebx + jnz L(mul1) + + + C Add products src[n]*src[n+1..size-1] at dst[2*n-1...], for + C n=1..size-2. + C + C The last two products, which are the end corner of the product + C triangle, are handled separately to save looping overhead. These + C are src[size-3]*src[size-2,size-1] and src[size-2]*src[size-1]. + C If size is 4 then it's only these that need to be done. + C + C In the outer loop %esi is a constant, and %edi just advances by 1 + C limb each time. The size of the operation decreases by 1 limb + C each time. + + C eax + C ebx carry (needing carry flag added) + C ecx + C edx + C esi &src[size] + C edi &dst[size] + C ebp + + adcl $0, %ebx + movl PARAM_SIZE, %edx + + movl %ebx, (%edi) + subl $4, %edx + + negl %edx + jz L(corner) + + +L(outer): + C ebx previous carry limb to store + C edx outer loop counter (negative) + C esi &src[size] + C edi dst, pointing at stored carry limb of previous loop + + pushl %edx C new outer loop counter + leal -2(%edx), %ecx + + movl %ebx, (%edi) + addl $4, %edi + + addl $4, %ebp + xorl %ebx, %ebx C initial carry limb, clear carry flag + +L(inner): + C eax scratch + C ebx carry (needing carry flag added) + C ecx counter, negative + C edx scratch + C esi &src[size] + C edi dst end of this addmul + C ebp &src[j] + + adcl $0, %ebx + movl (%esi,%ecx,4), %eax + + mull (%ebp) + + addl %ebx, %eax + movl (%edi,%ecx,4), %ebx + + adcl $0, %edx + addl %eax, %ebx + + movl %ebx, (%edi,%ecx,4) + incl %ecx + + movl %edx, %ebx + jnz L(inner) + + + adcl $0, %ebx + popl %edx C outer loop counter + + incl %edx + jnz L(outer) + + + movl %ebx, (%edi) + +L(corner): + C esi &src[size] + C edi &dst[2*size-4] + + movl -8(%esi), %eax + movl -4(%edi), %ebx C risk of data cache bank clash here + + mull -12(%esi) C src[size-2]*src[size-3] + + addl %eax, %ebx + movl %edx, %ecx + + adcl $0, %ecx + movl -4(%esi), %eax + + mull -12(%esi) C src[size-1]*src[size-3] + + addl %ecx, %eax + movl (%edi), %ecx + + adcl $0, %edx + movl %ebx, -4(%edi) + + addl %eax, %ecx + movl %edx, %ebx + + adcl $0, %ebx + movl -4(%esi), %eax + + mull -8(%esi) C src[size-1]*src[size-2] + + movl %ecx, (%edi) + addl %eax, %ebx + + adcl $0, %edx + movl PARAM_SIZE, %eax + + negl %eax + movl %ebx, 4(%edi) + + addl $1, %eax C -(size-1) and clear carry + movl %edx, 8(%edi) + + +C ----------------------------------------------------------------------------- +C Left shift of dst[1..2*size-2], high bit shifted out becomes dst[2*size-1]. + +L(lshift): + C eax counter, negative + C ebx next limb + C ecx + C edx + C esi + C edi &dst[2*size-4] + C ebp + + movl 12(%edi,%eax,8), %ebx + + rcll %ebx + movl 16(%edi,%eax,8), %ecx + + rcll %ecx + movl %ebx, 12(%edi,%eax,8) + + movl %ecx, 16(%edi,%eax,8) + incl %eax + + jnz L(lshift) + + + adcl %eax, %eax C high bit out + movl PARAM_SRC, %esi + + movl PARAM_SIZE, %ecx C risk of cache bank clash + movl %eax, 12(%edi) C dst most significant limb + + +C ----------------------------------------------------------------------------- +C Now add in the squares on the diagonal, namely src[0]^2, src[1]^2, ..., +C src[size-1]^2. dst[0] hasn't yet been set at all yet, and just gets the +C low limb of src[0]^2. + + movl (%esi), %eax C src[0] + leal (%esi,%ecx,4), %esi C src end + + negl %ecx + + mull %eax + + movl %eax, 16(%edi,%ecx,8) C dst[0] + movl %edx, %ebx + + addl $1, %ecx C size-1 and clear carry + +L(diag): + C eax scratch (low product) + C ebx carry limb + C ecx counter, negative + C edx scratch (high product) + C esi &src[size] + C edi &dst[2*size-4] + C ebp scratch (fetched dst limbs) + + movl (%esi,%ecx,4), %eax + adcl $0, %ebx + + mull %eax + + movl 16-4(%edi,%ecx,8), %ebp + + addl %ebp, %ebx + movl 16(%edi,%ecx,8), %ebp + + adcl %eax, %ebp + movl %ebx, 16-4(%edi,%ecx,8) + + movl %ebp, 16(%edi,%ecx,8) + incl %ecx + + movl %edx, %ebx + jnz L(diag) + + + adcl $0, %edx + movl 16-4(%edi), %eax C dst most significant limb + + addl %eax, %edx + popl %ebp + + movl %edx, 16-4(%edi) + popl %esi C risk of cache bank clash + + popl %edi + popl %ebx + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/README b/gmp-6.3.0/mpn/x86/pentium4/README new file mode 100644 index 0000000..90f752e --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/README @@ -0,0 +1,124 @@ +Copyright 2001 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. + + + + + INTEL PENTIUM-4 MPN SUBROUTINES + + +This directory contains mpn functions optimized for Intel Pentium-4. + +The mmx subdirectory has routines using MMX instructions, the sse2 +subdirectory has routines using SSE2 instructions. All P4s have these, the +separate directories are just so configure can omit that code if the +assembler doesn't support it. + + +STATUS + + cycles/limb + + mpn_add_n/sub_n 4 normal, 6 in-place + + mpn_mul_1 4 normal, 6 in-place + mpn_addmul_1 6 + mpn_submul_1 7 + + mpn_mul_basecase 6 cycles/crossproduct (approx) + + mpn_sqr_basecase 3.5 cycles/crossproduct (approx) + or 7.0 cycles/triangleproduct (approx) + + mpn_l/rshift 1.75 + + + +The shifts ought to be able to go at 1.5 c/l, but not much effort has been +applied to them yet. + +In-place operations, and all addmul, submul, mul_basecase and sqr_basecase +calls, suffer from pipeline anomalies associated with write combining and +movd reads and writes to the same or nearby locations. The movq +instructions do not trigger the same hardware problems. Unfortunately, +using movq and splitting/combining seems to require too many extra +instructions to help. Perhaps future chip steppings will be better. + + + +NOTES + +The Pentium-4 pipeline "Netburst", provides for quite a number of surprises. +Many traditional x86 instructions run very slowly, requiring use of +alterative instructions for acceptable performance. + +adcl and sbbl are quite slow at 8 cycles for reg->reg. paddq of 32-bits +within a 64-bit mmx register seems better, though the combination +paddq/psrlq when propagating a carry is still a 4 cycle latency. + +incl and decl should be avoided, instead use add $1 and sub $1. Apparently +the carry flag is not separately renamed, so incl and decl depend on all +previous flags-setting instructions. + +shll and shrl have a 4 cycle latency, or 8 times the latency of the fastest +integer instructions (addl, subl, orl, andl, and some more). shldl and +shrdl seem to have 13 and 15 cycles latency, respectively. Bizarre. + +movq mmx -> mmx does have 6 cycle latency, as noted in the documentation. +pxor/por or similar combination at 2 cycles latency can be used instead. +The movq however executes in the float unit, thereby saving MMX execution +resources. With the right juggling, data moves shouldn't be on a dependent +chain. + +L1 is write-through, but the write-combining sounds like it does enough to +not require explicit destination prefetching. + +xmm registers so far haven't found a use, but not much effort has been +expended. A configure test for whether the operating system knows +fxsave/fxrestor will be needed if they're used. + + + +REFERENCES + +Intel Pentium-4 processor manuals, + + http://developer.intel.com/design/pentium4/manuals + +"Intel Pentium 4 Processor Optimization Reference Manual", Intel, 2001, +order number 248966. Available on-line: + + http://developer.intel.com/design/pentium4/manuals/248966.htm + + + +---------------- +Local variables: +mode: text +fill-column: 76 +End: diff --git a/gmp-6.3.0/mpn/x86/pentium4/copyd.asm b/gmp-6.3.0/mpn/x86/pentium4/copyd.asm new file mode 100644 index 0000000..82af81c --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/copyd.asm @@ -0,0 +1,71 @@ +dnl Pentium-4 mpn_copyd -- copy limb vector, decrementing. + +dnl Copyright 1999-2001 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + + +dnl The std/rep/movsl/cld is very slow for small blocks on pentium4. Its +dnl startup time seems to be about 165 cycles. It then needs 2.6 c/l. +dnl We therefore use an open-coded 2 c/l copying loop. + +dnl Ultimately, we may want to use 64-bit movq or 128-bit movdqu in some +dnl nifty unrolled arrangement. Clearly, that could reach much higher +dnl speeds, at least for large blocks. + +include(`../config.m4') + + +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) + +PROLOGUE(mpn_copyd) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + + movl PARAM_SRC, %eax + movl PARAM_DST, %edx + movl %ebx, PARAM_SIZE + addl $-1, %ecx + js L(end) + +L(loop): + movl (%eax,%ecx,4), %ebx + movl %ebx, (%edx,%ecx,4) + addl $-1, %ecx + + jns L(loop) +L(end): + movl PARAM_SIZE, %ebx + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/copyi.asm b/gmp-6.3.0/mpn/x86/pentium4/copyi.asm new file mode 100644 index 0000000..b614887 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/copyi.asm @@ -0,0 +1,93 @@ +dnl Pentium-4 mpn_copyi -- copy limb vector, incrementing. + +dnl Copyright 1999-2001 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + + +dnl The rep/movsl is very slow for small blocks on pentium4. Its startup +dnl time seems to be about 110 cycles. It then copies at a rate of one +dnl limb per cycle. We therefore fall back to an open-coded 2 c/l copying +dnl loop for smaller sizes. + +dnl Ultimately, we may want to use 64-bit movd or 128-bit movdqu in some +dnl nifty unrolled arrangement. Clearly, that could reach much higher +dnl speeds, at least for large blocks. + +include(`../config.m4') + + +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) + +PROLOGUE(mpn_copyi) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + cmpl $150, %ecx + jg L(replmovs) + + movl PARAM_SRC, %eax + movl PARAM_DST, %edx + movl %ebx, PARAM_SIZE + testl %ecx, %ecx + jz L(end) + +L(loop): + movl (%eax), %ebx + leal 4(%eax), %eax + addl $-1, %ecx + movl %ebx, (%edx) + leal 4(%edx), %edx + + jnz L(loop) + +L(end): + movl PARAM_SIZE, %ebx + ret + +L(replmovs): + cld C better safe than sorry, see mpn/x86/README + + movl %esi, %eax + movl PARAM_SRC, %esi + movl %edi, %edx + movl PARAM_DST, %edi + + rep + movsl + + movl %eax, %esi + movl %edx, %edi + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/mmx/lshift.asm b/gmp-6.3.0/mpn/x86/pentium4/mmx/lshift.asm new file mode 100644 index 0000000..b5eca66 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/mmx/lshift.asm @@ -0,0 +1,39 @@ +dnl Intel Pentium-4 mpn_lshift -- left shift. + +dnl Copyright 2001, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P4 Willamette, Northwood: 1.75 cycles/limb +C P4 Prescott: 2.0 cycles/limb + + +MULFUNC_PROLOGUE(mpn_lshift) +include_mpn(`x86/pentium/mmx/lshift.asm') diff --git a/gmp-6.3.0/mpn/x86/pentium4/mmx/popham.asm b/gmp-6.3.0/mpn/x86/pentium4/mmx/popham.asm new file mode 100644 index 0000000..9563cb5 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/mmx/popham.asm @@ -0,0 +1,203 @@ +dnl Intel Pentium 4 mpn_popcount, mpn_hamdist -- population count and +dnl hamming distance. + +dnl Copyright 2000-2002, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C popcount hamdist +C P3 model 9 (Banias) ? ? +C P3 model 13 (Dothan) 6 6 +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) 8 9 +C P4 model 3 (Prescott) 8 9 +C P4 model 4 (Nocona) + +C unsigned long mpn_popcount (mp_srcptr src, mp_size_t size); +C unsigned long mpn_hamdist (mp_srcptr src, mp_srcptr src2, mp_size_t size); +C +C Loading with unaligned movq's costs an extra 1 c/l and hence is avoided. +C Two movd's and a punpckldq seems to be the same speed as an aligned movq, +C and using them saves fiddling about with alignment testing on entry. +C +C For popcount there's 13 mmx instructions in the loop, so perhaps 6.5 c/l +C might be possible, but 8 c/l relying on out-of-order execution is already +C quite reasonable. + +ifdef(`OPERATION_popcount',, +`ifdef(`OPERATION_hamdist',, +`m4_error(`Need OPERATION_popcount or OPERATION_hamdist defined +')')') + +define(HAM, +m4_assert_numargs(1) +`ifdef(`OPERATION_hamdist',`$1')') + +define(POP, +m4_assert_numargs(1) +`ifdef(`OPERATION_popcount',`$1')') + +HAM(` +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC2, 8) +defframe(PARAM_SRC, 4) +define(M4_function,mpn_hamdist) +') +POP(` +defframe(PARAM_SIZE, 8) +defframe(PARAM_SRC, 4) +define(M4_function,mpn_popcount) +') + +MULFUNC_PROLOGUE(mpn_popcount mpn_hamdist) + + +ifdef(`PIC',,` + dnl non-PIC + RODATA + ALIGN(8) +L(rodata_AAAAAAAAAAAAAAAA): + .long 0xAAAAAAAA + .long 0xAAAAAAAA +L(rodata_3333333333333333): + .long 0x33333333 + .long 0x33333333 +L(rodata_0F0F0F0F0F0F0F0F): + .long 0x0F0F0F0F + .long 0x0F0F0F0F +') + + TEXT + ALIGN(16) + +PROLOGUE(M4_function) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + movl PARAM_SRC, %eax + +ifdef(`PIC',` + movl $0xAAAAAAAA, %edx + movd %edx, %mm7 + punpckldq %mm7, %mm7 + + movl $0x33333333, %edx + movd %edx, %mm6 + punpckldq %mm6, %mm6 + + movl $0x0F0F0F0F, %edx + movd %edx, %mm5 + punpckldq %mm5, %mm5 + +HAM(` movl PARAM_SRC2, %edx') + +',` + dnl non-PIC +HAM(` movl PARAM_SRC2, %edx') + movq L(rodata_AAAAAAAAAAAAAAAA), %mm7 + movq L(rodata_3333333333333333), %mm6 + movq L(rodata_0F0F0F0F0F0F0F0F), %mm5 +') + + pxor %mm4, %mm4 C zero + pxor %mm0, %mm0 C total + + subl $1, %ecx + ja L(top) + +L(last): + movd (%eax,%ecx,4), %mm1 C src high limb +HAM(` movd (%edx,%ecx,4), %mm2 + pxor %mm2, %mm1 +') + jmp L(loaded) + + +L(top): + C eax src + C ebx + C ecx counter, size-1 to 2 or 1, inclusive + C edx [hamdist] src2 + C + C mm0 total (low dword) + C mm1 (scratch) + C mm2 (scratch) + C mm3 + C mm4 0x0000000000000000 + C mm5 0x0F0F0F0F0F0F0F0F + C mm6 0x3333333333333333 + C mm7 0xAAAAAAAAAAAAAAAA + + movd (%eax), %mm1 + movd 4(%eax), %mm2 + punpckldq %mm2, %mm1 + addl $8, %eax + +HAM(` movd (%edx), %mm2 + movd 4(%edx), %mm3 + punpckldq %mm3, %mm2 + pxor %mm2, %mm1 + addl $8, %edx +') + +L(loaded): + movq %mm7, %mm2 + pand %mm1, %mm2 + psrlq $1, %mm2 + psubd %mm2, %mm1 C bit pairs + + movq %mm6, %mm2 + pand %mm1, %mm2 + psrlq $2, %mm1 + pand %mm6, %mm1 + paddd %mm2, %mm1 C nibbles + + movq %mm5, %mm2 + pand %mm1, %mm2 + psrlq $4, %mm1 + pand %mm5, %mm1 + paddd %mm2, %mm1 C bytes + + psadbw( %mm4, %mm1) + paddd %mm1, %mm0 C to total + + subl $2, %ecx + jg L(top) + + C ecx is 0 or -1 representing respectively 1 or 0 further limbs + jz L(last) + + + movd %mm0, %eax + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/mmx/rshift.asm b/gmp-6.3.0/mpn/x86/pentium4/mmx/rshift.asm new file mode 100644 index 0000000..3ac0094 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/mmx/rshift.asm @@ -0,0 +1,39 @@ +dnl Intel Pentium-4 mpn_rshift -- right shift. + +dnl Copyright 2001, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P4 Willamette, Northwood: 1.75 cycles/limb +C P4 Prescott: 2.0 cycles/limb + + +MULFUNC_PROLOGUE(mpn_rshift) +include_mpn(`x86/pentium/mmx/rshift.asm') diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/add_n.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/add_n.asm new file mode 100644 index 0000000..8e2380e --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/add_n.asm @@ -0,0 +1,101 @@ +dnl Intel Pentium-4 mpn_add_n -- mpn addition. + +dnl Copyright 2001, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C dst!=src1,2 dst==src1 dst==src2 +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) ? +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) 4 6 6 +C P4 model 3-4 (Prescott) 4.25 7.5 7.5 + +defframe(PARAM_CARRY,20) +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC2, 12) +defframe(PARAM_SRC1, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(SAVE_EBX,`PARAM_SRC1') + + TEXT + ALIGN(8) + +PROLOGUE(mpn_add_nc) +deflit(`FRAME',0) + movd PARAM_CARRY, %mm0 + jmp L(start_nc) +EPILOGUE() + + ALIGN(8) +PROLOGUE(mpn_add_n) +deflit(`FRAME',0) + pxor %mm0, %mm0 +L(start_nc): + mov PARAM_SRC1, %eax + mov %ebx, SAVE_EBX + mov PARAM_SRC2, %ebx + mov PARAM_DST, %edx + mov PARAM_SIZE, %ecx + + lea (%eax,%ecx,4), %eax C src1 end + lea (%ebx,%ecx,4), %ebx C src2 end + lea (%edx,%ecx,4), %edx C dst end + neg %ecx C -size + +L(top): + C eax src1 end + C ebx src2 end + C ecx counter, limbs, negative + C edx dst end + C mm0 carry bit + + movd (%eax,%ecx,4), %mm1 + movd (%ebx,%ecx,4), %mm2 + paddq %mm2, %mm1 + + paddq %mm1, %mm0 + movd %mm0, (%edx,%ecx,4) + + psrlq $32, %mm0 + + add $1, %ecx + jnz L(top) + + movd %mm0, %eax + mov SAVE_EBX, %ebx + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/addlsh1_n.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/addlsh1_n.asm new file mode 100644 index 0000000..93b63b2 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/addlsh1_n.asm @@ -0,0 +1,108 @@ +dnl Intel Pentium-4 mpn_addlsh1_n -- mpn x+2*y. + +dnl Copyright 2001-2004, 2006 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C dst!=src1,2 dst==src1 dst==src2 +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) ? +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) 4.25 6 6 +C P4 model 3-4 (Prescott) 5 8.5 8.5 + +C The slightly strange combination of indexing and pointer incrementing +C that's used seems to work best. Not sure why, but %ecx,4 with src1 and/or +C src2 is a slowdown. +C +C The dependent chain is simply the paddq of x+2*y to the previous carry, +C then psrlq to get the new carry. That makes 4 c/l the target speed, which +C is almost achieved for separate src/dst but when src==dst the write +C combining anomalies slow it down. + +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC2, 12) +defframe(PARAM_SRC1, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(SAVE_EBX,`PARAM_SRC1') + + TEXT + ALIGN(8) + +PROLOGUE(mpn_addlsh1_n) +deflit(`FRAME',0) + + mov PARAM_SRC1, %eax + mov %ebx, SAVE_EBX + + mov PARAM_SRC2, %ebx + pxor %mm0, %mm0 C initial carry + + mov PARAM_DST, %edx + + mov PARAM_SIZE, %ecx + + lea (%edx,%ecx,4), %edx C dst end + neg %ecx C -size + +L(top): + C eax src1 end + C ebx src2 end + C ecx counter, limbs, negative + C edx dst end + C mm0 carry + + movd (%ebx), %mm2 + movd (%eax), %mm1 + psrlq $32, %mm0 + lea 4(%eax), %eax + lea 4(%ebx), %ebx + + psllq $1, %mm2 + paddq %mm2, %mm1 + + paddq %mm1, %mm0 + + movd %mm0, (%edx,%ecx,4) + add $1, %ecx + jnz L(top) + + + psrlq $32, %mm0 + mov SAVE_EBX, %ebx + movd %mm0, %eax + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/addmul_1.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/addmul_1.asm new file mode 100644 index 0000000..7810207 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/addmul_1.asm @@ -0,0 +1,189 @@ +dnl mpn_addmul_1 for Pentium 4 and P6 models with SSE2 (i.e., 9,D,E,F). + +dnl Copyright 2005, 2007, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + + +include(`../config.m4') + +C cycles/limb +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) 5.24 +C P6 model 13 (Dothan) 5.24 +C P4 model 0-1 (Willamette) 5 +C P4 model 2 (Northwood) 5 +C P4 model 3-4 (Prescott) 5 + +C TODO: +C * Tweak eax/edx offsets in loop as to save some lea's +C * Perhaps software pipeline small-case code + +C INPUT PARAMETERS +C rp sp + 4 +C up sp + 8 +C n sp + 12 +C v0 sp + 16 + + TEXT + ALIGN(16) +PROLOGUE(mpn_addmul_1) + pxor %mm6, %mm6 +L(ent): mov 4(%esp), %edx + mov 8(%esp), %eax + mov 12(%esp), %ecx + movd 16(%esp), %mm7 + cmp $4, %ecx + jnc L(big) + +L(lp0): movd (%eax), %mm0 + lea 4(%eax), %eax + movd (%edx), %mm4 + lea 4(%edx), %edx + pmuludq %mm7, %mm0 + paddq %mm0, %mm4 + paddq %mm4, %mm6 + movd %mm6, -4(%edx) + psrlq $32, %mm6 + dec %ecx + jnz L(lp0) + movd %mm6, %eax + emms + ret + +L(big): and $3, %ecx + je L(0) + cmp $2, %ecx + jc L(1) + je L(2) + jmp L(3) C FIXME: one case should fall through + +L(0): movd (%eax), %mm3 + sub 12(%esp), %ecx C loop count + lea -16(%eax), %eax + lea -12(%edx), %edx + pmuludq %mm7, %mm3 + movd 20(%eax), %mm0 + movd 12(%edx), %mm5 + pmuludq %mm7, %mm0 + movd 24(%eax), %mm1 + paddq %mm3, %mm5 + movd 16(%edx), %mm4 + jmp L(00) + +L(1): movd (%eax), %mm2 + sub 12(%esp), %ecx + lea -12(%eax), %eax + lea -8(%edx), %edx + movd 8(%edx), %mm4 + pmuludq %mm7, %mm2 + movd 16(%eax), %mm3 + pmuludq %mm7, %mm3 + movd 20(%eax), %mm0 + paddq %mm2, %mm4 + movd 12(%edx), %mm5 + jmp L(01) + +L(2): movd (%eax), %mm1 + sub 12(%esp), %ecx + lea -8(%eax), %eax + lea -4(%edx), %edx + pmuludq %mm7, %mm1 + movd 12(%eax), %mm2 + movd 4(%edx), %mm5 + pmuludq %mm7, %mm2 + movd 16(%eax), %mm3 + paddq %mm1, %mm5 + movd 8(%edx), %mm4 + jmp L(10) + +L(3): movd (%eax), %mm0 + sub 12(%esp), %ecx + lea -4(%eax), %eax + pmuludq %mm7, %mm0 + movd 8(%eax), %mm1 + movd (%edx), %mm4 + pmuludq %mm7, %mm1 + movd 12(%eax), %mm2 + paddq %mm0, %mm4 + movd 4(%edx), %mm5 + + ALIGN(16) +L(top): pmuludq %mm7, %mm2 + paddq %mm4, %mm6 + movd 16(%eax), %mm3 + paddq %mm1, %mm5 + movd 8(%edx), %mm4 + movd %mm6, 0(%edx) + psrlq $32, %mm6 +L(10): pmuludq %mm7, %mm3 + paddq %mm5, %mm6 + movd 20(%eax), %mm0 + paddq %mm2, %mm4 + movd 12(%edx), %mm5 + movd %mm6, 4(%edx) + psrlq $32, %mm6 +L(01): pmuludq %mm7, %mm0 + paddq %mm4, %mm6 + movd 24(%eax), %mm1 + paddq %mm3, %mm5 + movd 16(%edx), %mm4 + movd %mm6, 8(%edx) + psrlq $32, %mm6 +L(00): pmuludq %mm7, %mm1 + paddq %mm5, %mm6 + movd 28(%eax), %mm2 + paddq %mm0, %mm4 + movd 20(%edx), %mm5 + movd %mm6, 12(%edx) + psrlq $32, %mm6 + lea 16(%eax), %eax + lea 16(%edx), %edx + add $4, %ecx + jnz L(top) + +L(end): pmuludq %mm7, %mm2 + paddq %mm4, %mm6 + paddq %mm1, %mm5 + movd 8(%edx), %mm4 + movd %mm6, 0(%edx) + psrlq $32, %mm6 + paddq %mm5, %mm6 + paddq %mm2, %mm4 + movd %mm6, 4(%edx) + psrlq $32, %mm6 + paddq %mm4, %mm6 + movd %mm6, 8(%edx) + psrlq $32, %mm6 + movd %mm6, %eax + emms + ret +EPILOGUE() +PROLOGUE(mpn_addmul_1c) + movd 20(%esp), %mm6 + jmp L(ent) +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/bdiv_dbm1c.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/bdiv_dbm1c.asm new file mode 100644 index 0000000..354300e --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/bdiv_dbm1c.asm @@ -0,0 +1,141 @@ +dnl Intel Atom mpn_bdiv_dbm1. + +dnl Contributed to the GNU project by Torbjorn Granlund and Marco Bodrato. + +dnl Copyright 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C cycles/limb +C cycles/limb +C P5 - +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) 9.75 +C P6 model 13 (Dothan) +C P4 model 0 (Willamette) +C P4 model 1 (?) +C P4 model 2 (Northwood) 8.25 +C P4 model 3 (Prescott) +C P4 model 4 (Nocona) +C Intel Atom 8 +C AMD K6 - +C AMD K7 - +C AMD K8 +C AMD K10 + +C TODO: This code was optimised for atom-32, consider moving it back to atom +C dir(atom currently grabs this code), and write a 4-way version(7c/l). + +defframe(PARAM_CARRY,20) +defframe(PARAM_MUL, 16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(SAVE_RP,`PARAM_MUL') +define(SAVE_UP,`PARAM_SIZE') + +define(`rp', `%edi') +define(`up', `%esi') +define(`n', `%ecx') +define(`reg', `%edx') +define(`cy', `%eax') C contains the return value + +ASM_START() + TEXT + ALIGN(16) +deflit(`FRAME',0) + +PROLOGUE(mpn_bdiv_dbm1c) + mov PARAM_SIZE, n C size + mov up, SAVE_UP + mov PARAM_SRC, up + movd PARAM_MUL, %mm7 + mov rp, SAVE_RP + mov PARAM_DST, rp + + movd (up), %mm0 + pmuludq %mm7, %mm0 + shr n + mov PARAM_CARRY, cy + jz L(eq1) + + movd 4(up), %mm1 + jc L(odd) + + lea 4(up), up + pmuludq %mm7, %mm1 + movd %mm0, reg + psrlq $32, %mm0 + sub reg, cy + movd %mm0, reg + movq %mm1, %mm0 + dec n + mov cy, (rp) + lea 4(rp), rp + jz L(end) + +C ALIGN(16) +L(top): movd 4(up), %mm1 + sbb reg, cy +L(odd): movd %mm0, reg + psrlq $32, %mm0 + pmuludq %mm7, %mm1 + sub reg, cy + lea 8(up), up + movd %mm0, reg + movd (up), %mm0 + mov cy, (rp) + sbb reg, cy + movd %mm1, reg + psrlq $32, %mm1 + sub reg, cy + movd %mm1, reg + pmuludq %mm7, %mm0 + dec n + mov cy, 4(rp) + lea 8(rp), rp + jnz L(top) + +L(end): sbb reg, cy + +L(eq1): movd %mm0, reg + psrlq $32, %mm0 + mov SAVE_UP, up + sub reg, cy + movd %mm0, reg + emms + mov cy, (rp) + sbb reg, cy + + mov SAVE_RP, rp + ret +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/bdiv_q_1.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/bdiv_q_1.asm new file mode 100644 index 0000000..d5008f4 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/bdiv_q_1.asm @@ -0,0 +1,234 @@ +dnl Intel Pentium-4 mpn_divexact_1 -- mpn by limb exact division. + +dnl Rearranged from mpn/x86/pentium4/sse2/dive_1.asm by Marco Bodrato. + +dnl Copyright 2001, 2002, 2007, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P4: 19.0 cycles/limb + +C Pairs of movd's are used to avoid unaligned loads. Despite the loads not +C being on the dependent chain and there being plenty of cycles available, +C using an unaligned movq on every second iteration measured about 23 c/l. +C + +defframe(PARAM_SHIFT, 24) +defframe(PARAM_INVERSE,20) +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + +C mp_limb_t +C mpn_pi1_bdiv_q_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, mp_limb_t divisor, +C mp_limb_t inverse, int shift) + ALIGN(32) +PROLOGUE(mpn_pi1_bdiv_q_1) +deflit(`FRAME',0) + + movl PARAM_SIZE, %edx + + movl PARAM_SRC, %eax + + movl PARAM_DIVISOR, %ecx + + movd %ecx, %mm6 + movl PARAM_SHIFT, %ecx + + movd %ecx, %mm7 C shift + + C + + movl PARAM_INVERSE, %ecx + movd %ecx, %mm5 C inv + + movl PARAM_DST, %ecx + pxor %mm1, %mm1 C initial carry limb + pxor %mm0, %mm0 C initial carry bit + + subl $1, %edx + jz L(done) + + pcmpeqd %mm4, %mm4 + psrlq $32, %mm4 C 0x00000000FFFFFFFF + +C The dependent chain here is as follows. +C +C latency +C psubq s = (src-cbit) - climb 2 +C pmuludq q = s*inverse 8 +C pmuludq prod = q*divisor 8 +C psrlq climb = high(prod) 2 +C -- +C 20 +C +C Yet the loop measures 19.0 c/l, so obviously there's something gained +C there over a straight reading of the chip documentation. + +L(top): + C eax src, incrementing + C ebx + C ecx dst, incrementing + C edx counter, size-1 iterations + C + C mm0 carry bit + C mm1 carry limb + C mm4 0x00000000FFFFFFFF + C mm5 inverse + C mm6 divisor + C mm7 shift + + movd (%eax), %mm2 + movd 4(%eax), %mm3 + addl $4, %eax + punpckldq %mm3, %mm2 + + psrlq %mm7, %mm2 + pand %mm4, %mm2 C src + psubq %mm0, %mm2 C src - cbit + + psubq %mm1, %mm2 C src - cbit - climb + movq %mm2, %mm0 + psrlq $63, %mm0 C new cbit + + pmuludq %mm5, %mm2 C s*inverse + movd %mm2, (%ecx) C q + addl $4, %ecx + + movq %mm6, %mm1 + pmuludq %mm2, %mm1 C q*divisor + psrlq $32, %mm1 C new climb + +L(entry): + subl $1, %edx + jnz L(top) + +L(done): + movd (%eax), %mm2 + psrlq %mm7, %mm2 C src + psubq %mm0, %mm2 C src - cbit + + psubq %mm1, %mm2 C src - cbit - climb + + pmuludq %mm5, %mm2 C s*inverse + movd %mm2, (%ecx) C q + + emms + ret + +EPILOGUE() + + ALIGN(16) +C mp_limb_t mpn_bdiv_q_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C +PROLOGUE(mpn_bdiv_q_1) +deflit(`FRAME',0) + + movl PARAM_SIZE, %edx + + movl PARAM_DIVISOR, %ecx + + C eax src + C ebx + C ecx divisor + C edx size-1 + + movl %ecx, %eax + bsfl %ecx, %ecx C trailing twos + + shrl %cl, %eax C d = divisor without twos + movd %eax, %mm6 + movd %ecx, %mm7 C shift + + shrl %eax C d/2 + + andl $127, %eax C d/2, 7 bits + +ifdef(`PIC',` + LEA( binvert_limb_table, %ecx) + movzbl (%eax,%ecx), %eax C inv 8 bits +',` + movzbl binvert_limb_table(%eax), %eax C inv 8 bits +') + + C + + movd %eax, %mm5 C inv + + movd %eax, %mm0 C inv + + pmuludq %mm5, %mm5 C inv*inv + + C + + pmuludq %mm6, %mm5 C inv*inv*d + paddd %mm0, %mm0 C 2*inv + + C + + psubd %mm5, %mm0 C inv = 2*inv - inv*inv*d + pxor %mm5, %mm5 + + paddd %mm0, %mm5 + pmuludq %mm0, %mm0 C inv*inv + + pcmpeqd %mm4, %mm4 + psrlq $32, %mm4 C 0x00000000FFFFFFFF + + C + + pmuludq %mm6, %mm0 C inv*inv*d + paddd %mm5, %mm5 C 2*inv + + movl PARAM_SRC, %eax + movl PARAM_DST, %ecx + pxor %mm1, %mm1 C initial carry limb + + C + + psubd %mm0, %mm5 C inv = 2*inv - inv*inv*d + + ASSERT(e,` C expect d*inv == 1 mod 2^GMP_LIMB_BITS + pushl %eax FRAME_pushl() + movq %mm6, %mm0 + pmuludq %mm5, %mm0 + movd %mm0, %eax + cmpl $1, %eax + popl %eax FRAME_popl()') + + pxor %mm0, %mm0 C initial carry bit + jmp L(entry) + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/cnd_add_n.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/cnd_add_n.asm new file mode 100644 index 0000000..b3f3474 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/cnd_add_n.asm @@ -0,0 +1,95 @@ +dnl Intel Pentium-4 mpn_cnd_add_n -- mpn addition. + +dnl Copyright 2001, 2002, 2013 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) 4.67 +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) 5 +C P4 model 3-4 (Prescott) 5.25 + +defframe(PARAM_SIZE, 20) +defframe(PARAM_SRC2, 16) +defframe(PARAM_SRC1, 12) +defframe(PARAM_DST, 8) +defframe(PARAM_CND, 4) + +dnl re-use parameter space +define(SAVE_EBX,`PARAM_SRC1') + +define(`cnd', `%mm3') + + TEXT + ALIGN(8) + + ALIGN(8) +PROLOGUE(mpn_cnd_add_n) +deflit(`FRAME',0) + pxor %mm0, %mm0 + + mov PARAM_CND, %eax + neg %eax + sbb %eax, %eax + movd %eax, cnd + + mov PARAM_SRC1, %eax + mov %ebx, SAVE_EBX + mov PARAM_SRC2, %ebx + mov PARAM_DST, %edx + mov PARAM_SIZE, %ecx + + lea (%eax,%ecx,4), %eax C src1 end + lea (%ebx,%ecx,4), %ebx C src2 end + lea (%edx,%ecx,4), %edx C dst end + neg %ecx C -size + +L(top): movd (%ebx,%ecx,4), %mm2 + movd (%eax,%ecx,4), %mm1 + pand cnd, %mm2 + paddq %mm2, %mm1 + + paddq %mm1, %mm0 + movd %mm0, (%edx,%ecx,4) + + psrlq $32, %mm0 + + add $1, %ecx + jnz L(top) + + movd %mm0, %eax + mov SAVE_EBX, %ebx + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/cnd_sub_n.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/cnd_sub_n.asm new file mode 100644 index 0000000..339a23e --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/cnd_sub_n.asm @@ -0,0 +1,114 @@ +dnl Intel Pentium-4 mpn_cnd_sub_n -- mpn subtraction. + +dnl Copyright 2001, 2002, 2013 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) 4.67 +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) 5 +C P4 model 3-4 (Prescott) 5.25 + +defframe(PARAM_SIZE, 20) +defframe(PARAM_SRC2, 16) +defframe(PARAM_SRC1, 12) +defframe(PARAM_DST, 8) +defframe(PARAM_CND, 4) + +dnl re-use parameter space +define(SAVE_EBX,`PARAM_SRC1') + +define(`cnd', `%mm3') + + TEXT + ALIGN(8) + + ALIGN(8) +PROLOGUE(mpn_cnd_sub_n) +deflit(`FRAME',0) + pxor %mm0, %mm0 + + mov PARAM_CND, %eax + neg %eax + sbb %eax, %eax + movd %eax, cnd + + mov PARAM_SRC1, %eax + mov %ebx, SAVE_EBX + mov PARAM_SRC2, %ebx + mov PARAM_DST, %edx + mov PARAM_SIZE, %ecx + + lea (%eax,%ecx,4), %eax C src1 end + lea (%ebx,%ecx,4), %ebx C src2 end + lea (%edx,%ecx,4), %edx C dst end + neg %ecx C -size + +L(top): movd (%ebx,%ecx,4), %mm2 + movd (%eax,%ecx,4), %mm1 + pand cnd, %mm2 + psubq %mm2, %mm1 + + psubq %mm0, %mm1 + movd %mm1, (%edx,%ecx,4) + + psrlq $63, %mm1 + + add $1, %ecx + jz L(done_mm1) + + movd (%ebx,%ecx,4), %mm2 + movd (%eax,%ecx,4), %mm0 + pand cnd, %mm2 + psubq %mm2, %mm0 + + psubq %mm1, %mm0 + movd %mm0, (%edx,%ecx,4) + + psrlq $63, %mm0 + + add $1, %ecx + jnz L(top) + + movd %mm0, %eax + mov SAVE_EBX, %ebx + emms + ret + +L(done_mm1): + movd %mm1, %eax + mov SAVE_EBX, %ebx + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/dive_1.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/dive_1.asm new file mode 100644 index 0000000..0ceef5b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/dive_1.asm @@ -0,0 +1,216 @@ +dnl Intel Pentium-4 mpn_divexact_1 -- mpn by limb exact division. + +dnl Copyright 2001, 2002, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P4: 19.0 cycles/limb + + +C void mpn_divexact_1 (mp_ptr dst, mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C +C Pairs of movd's are used to avoid unaligned loads. Despite the loads not +C being on the dependent chain and there being plenty of cycles available, +C using an unaligned movq on every second iteration measured about 23 c/l. +C +C Using divl for size==1 seems a touch quicker than mul-by-inverse. The mul +C will be about 9+2*4+2*2+10*4+19+12 = 92 cycles latency, though some of +C that might be hidden by out-of-order execution, whereas divl is around 60. +C At size==2 an extra 19 for the mul versus 60 for the divl will see the mul +C faster. + +defframe(PARAM_DIVISOR,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + + ALIGN(16) +PROLOGUE(mpn_divexact_1) +deflit(`FRAME',0) + + movl PARAM_SIZE, %edx + + movl PARAM_SRC, %eax + + movl PARAM_DIVISOR, %ecx + subl $1, %edx + jnz L(two_or_more) + + movl (%eax), %eax + xorl %edx, %edx + + divl %ecx + movl PARAM_DST, %ecx + + movl %eax, (%ecx) + ret + + +L(two_or_more): + C eax src + C ebx + C ecx divisor + C edx size-1 + + movl %ecx, %eax + bsfl %ecx, %ecx C trailing twos + + shrl %cl, %eax C d = divisor without twos + movd %eax, %mm6 + movd %ecx, %mm7 C shift + + shrl %eax C d/2 + + andl $127, %eax C d/2, 7 bits + +ifdef(`PIC',` + LEA( binvert_limb_table, %ecx) + movzbl (%eax,%ecx), %eax C inv 8 bits +',` + movzbl binvert_limb_table(%eax), %eax C inv 8 bits +') + + C + + movd %eax, %mm5 C inv + + movd %eax, %mm0 C inv + + pmuludq %mm5, %mm5 C inv*inv + + C + + pmuludq %mm6, %mm5 C inv*inv*d + paddd %mm0, %mm0 C 2*inv + + C + + psubd %mm5, %mm0 C inv = 2*inv - inv*inv*d + pxor %mm5, %mm5 + + paddd %mm0, %mm5 + pmuludq %mm0, %mm0 C inv*inv + + pcmpeqd %mm4, %mm4 + psrlq $32, %mm4 C 0x00000000FFFFFFFF + + C + + pmuludq %mm6, %mm0 C inv*inv*d + paddd %mm5, %mm5 C 2*inv + + movl PARAM_SRC, %eax + movl PARAM_DST, %ecx + pxor %mm1, %mm1 C initial carry limb + + C + + psubd %mm0, %mm5 C inv = 2*inv - inv*inv*d + + ASSERT(e,` C expect d*inv == 1 mod 2^GMP_LIMB_BITS + pushl %eax FRAME_pushl() + movq %mm6, %mm0 + pmuludq %mm5, %mm0 + movd %mm0, %eax + cmpl $1, %eax + popl %eax FRAME_popl()') + + pxor %mm0, %mm0 C initial carry bit + + +C The dependent chain here is as follows. +C +C latency +C psubq s = (src-cbit) - climb 2 +C pmuludq q = s*inverse 8 +C pmuludq prod = q*divisor 8 +C psrlq climb = high(prod) 2 +C -- +C 20 +C +C Yet the loop measures 19.0 c/l, so obviously there's something gained +C there over a straight reading of the chip documentation. + +L(top): + C eax src, incrementing + C ebx + C ecx dst, incrementing + C edx counter, size-1 iterations + C + C mm0 carry bit + C mm1 carry limb + C mm4 0x00000000FFFFFFFF + C mm5 inverse + C mm6 divisor + C mm7 shift + + movd (%eax), %mm2 + movd 4(%eax), %mm3 + addl $4, %eax + punpckldq %mm3, %mm2 + + psrlq %mm7, %mm2 + pand %mm4, %mm2 C src + psubq %mm0, %mm2 C src - cbit + + psubq %mm1, %mm2 C src - cbit - climb + movq %mm2, %mm0 + psrlq $63, %mm0 C new cbit + + pmuludq %mm5, %mm2 C s*inverse + movd %mm2, (%ecx) C q + addl $4, %ecx + + movq %mm6, %mm1 + pmuludq %mm2, %mm1 C q*divisor + psrlq $32, %mm1 C new climb + + subl $1, %edx + jnz L(top) + + +L(done): + movd (%eax), %mm2 + psrlq %mm7, %mm2 C src + psubq %mm0, %mm2 C src - cbit + + psubq %mm1, %mm2 C src - cbit - climb + + pmuludq %mm5, %mm2 C s*inverse + movd %mm2, (%ecx) C q + + emms + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/divrem_1.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/divrem_1.asm new file mode 100644 index 0000000..0146fab --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/divrem_1.asm @@ -0,0 +1,645 @@ +dnl Intel Pentium-4 mpn_divrem_1 -- mpn by limb division. + +dnl Copyright 1999-2004 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P4: 32 cycles/limb integer part, 30 cycles/limb fraction part. + + +C mp_limb_t mpn_divrem_1 (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C mp_limb_t mpn_divrem_1c (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, +C mp_limb_t divisor, mp_limb_t carry); +C mp_limb_t mpn_preinv_divrem_1 (mp_ptr dst, mp_size_t xsize, +C mp_srcptr src, mp_size_t size, +C mp_limb_t divisor, mp_limb_t inverse, +C unsigned shift); +C +C Algorithm: +C +C The method and nomenclature follow part 8 of "Division by Invariant +C Integers using Multiplication" by Granlund and Montgomery, reference in +C gmp.texi. +C +C "m" is written for what is m' in the paper, and "d" for d_norm, which +C won't cause any confusion since it's only the normalized divisor that's of +C any use in the code. "b" is written for 2^N, the size of a limb, N being +C 32 here. +C +C The step "sdword dr = n - 2^N*d + (2^N-1-q1) * d" is instead done as +C "n-d - q1*d". This rearrangement gives the same two-limb answer but lets +C us have just a psubq on the dependent chain. +C +C For reference, the way the k7 code uses "n-(q1+1)*d" would not suit here, +C detecting an overflow of q1+1 when q1=0xFFFFFFFF would cost too much. +C +C Notes: +C +C mpn_divrem_1 and mpn_preinv_divrem_1 avoid one division if the src high +C limb is less than the divisor. mpn_divrem_1c doesn't check for a zero +C carry, since in normal circumstances that will be a very rare event. +C +C The test for skipping a division is branch free (once size>=1 is tested). +C The store to the destination high limb is 0 when a divide is skipped, or +C if it's not skipped then a copy of the src high limb is stored. The +C latter is in case src==dst. +C +C There's a small bias towards expecting xsize==0, by having code for +C xsize==0 in a straight line and xsize!=0 under forward jumps. +C +C Enhancements: +C +C The loop measures 32 cycles, but the dependent chain would suggest it +C could be done with 30. Not sure where to start looking for the extras. +C +C Alternatives: +C +C If the divisor is normalized (high bit set) then a division step can +C always be skipped, since the high destination limb is always 0 or 1 in +C that case. It doesn't seem worth checking for this though, since it +C probably occurs infrequently. + + +dnl MUL_THRESHOLD is the value of xsize+size at which the multiply by +dnl inverse method is used, rather than plain "divl"s. Minimum value 1. +dnl +dnl The inverse takes about 80-90 cycles to calculate, but after that the +dnl multiply is 32 c/l versus division at about 58 c/l. +dnl +dnl At 4 limbs the div is a touch faster than the mul (and of course +dnl simpler), so start the mul from 5 limbs. + +deflit(MUL_THRESHOLD, 5) + + +defframe(PARAM_PREINV_SHIFT, 28) dnl mpn_preinv_divrem_1 +defframe(PARAM_PREINV_INVERSE, 24) dnl mpn_preinv_divrem_1 +defframe(PARAM_CARRY, 24) dnl mpn_divrem_1c +defframe(PARAM_DIVISOR,20) +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC, 12) +defframe(PARAM_XSIZE, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(SAVE_ESI,`PARAM_SIZE') +define(SAVE_EBP,`PARAM_SRC') +define(SAVE_EDI,`PARAM_DIVISOR') +define(SAVE_EBX,`PARAM_DST') + + TEXT + + ALIGN(16) +PROLOGUE(mpn_preinv_divrem_1) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + xorl %edx, %edx C carry if can't skip a div + + movl %esi, SAVE_ESI + movl PARAM_SRC, %esi + + movl %ebp, SAVE_EBP + movl PARAM_DIVISOR, %ebp + + movl %edi, SAVE_EDI + movl PARAM_DST, %edi + + movl -4(%esi,%ecx,4), %eax C src high limb + + movl %ebx, SAVE_EBX + movl PARAM_XSIZE, %ebx + + movd PARAM_PREINV_INVERSE, %mm4 + + movd PARAM_PREINV_SHIFT, %mm7 C l + cmpl %ebp, %eax C high cmp divisor + + cmovc( %eax, %edx) C high is carry if high new n2 + psrlq $32, %mm3 C high n - (q1+1)*d, 0 or -1 + + ASSERT(be,`C 0 or -1 + movd %mm3, %eax + addl $1, %eax + cmpl $1, %eax') + + paddd %mm3, %mm2 C q + pand %mm5, %mm3 C mask & d + + paddd %mm3, %mm0 C addback if necessary + movd %mm2, (%edi) + leal -4(%edi), %edi + + subl $1, %ecx + ja L(integer_top) + + +L(integer_last): + C eax + C ebx xsize + C ecx + C edx + C esi &src[0] + C edi &dst[xsize] + C + C mm0 n2 + C mm4 m + C mm5 d + C mm6 + C mm7 l + + ASSERT(b,`C n2 n2 + psrlq $32, %mm3 C high n - (q1+1)*d, 0 or -1 + + ASSERT(be,`C 0 or -1 + movd %mm3, %eax + addl $1, %eax + cmpl $1, %eax') + + paddd %mm3, %mm2 C q + pand %mm5, %mm3 C mask & d + + paddd %mm3, %mm0 C addback if necessary + movd %mm2, (%edi) + leal -4(%edi), %edi + + +L(integer_none): + C eax + C ebx xsize + + orl %ebx, %ebx + jnz L(fraction_some) C if xsize!=0 + + +L(fraction_done): + movl SAVE_EBP, %ebp + psrld %mm7, %mm0 C remainder + + movl SAVE_EDI, %edi + movd %mm0, %eax + + movl SAVE_ESI, %esi + movl SAVE_EBX, %ebx + emms + ret + + + +C ----------------------------------------------------------------------------- +C + +L(fraction_some): + C eax + C ebx xsize + C ecx + C edx + C esi + C edi &dst[xsize-1] + C ebp + + +L(fraction_top): + C eax + C ebx counter, xsize iterations + C ecx + C edx + C esi src, decrementing + C edi dst, decrementing + C + C mm0 n2 + C mm4 m + C mm5 d + C mm6 32-l + C mm7 l + + ASSERT(b,`C n2 n2 + psrlq $32, %mm1 C high n - (q1+1)*d, 0 or -1 + + ASSERT(be,`C 0 or -1 + movd %mm1, %eax + addl $1, %eax + cmpl $1, %eax') + + paddd %mm1, %mm2 C q + pand %mm5, %mm1 C mask & d + + paddd %mm1, %mm0 C addback if necessary + movd %mm2, (%edi) + leal -4(%edi), %edi + + subl $1, %ebx + jne L(fraction_top) + + + jmp L(fraction_done) + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/gmp-mparam.h b/gmp-6.3.0/mpn/x86/pentium4/sse2/gmp-mparam.h new file mode 100644 index 0000000..a047a51 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/gmp-mparam.h @@ -0,0 +1,213 @@ +/* Intel Pentium-4 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 2600 MHz P4 Northwood */ +/* FFT tuning limit = 23,700,309 */ +/* Generated by tuneup.c, 2019-11-09, gcc 8.2 */ + +#define MOD_1_NORM_THRESHOLD 5 +#define MOD_1_UNNORM_THRESHOLD 14 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 7 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 5 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 13 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 7 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 2 /* 4.36% faster than 1 */ +#define DIV_QR_1_NORM_THRESHOLD 16 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 21 + +#define DIV_1_VS_MUL_1_PERCENT 358 + +#define MUL_TOOM22_THRESHOLD 26 +#define MUL_TOOM33_THRESHOLD 101 +#define MUL_TOOM44_THRESHOLD 284 +#define MUL_TOOM6H_THRESHOLD 406 +#define MUL_TOOM8H_THRESHOLD 592 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 101 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 191 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 189 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 195 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 151 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 51 +#define SQR_TOOM3_THRESHOLD 163 +#define SQR_TOOM4_THRESHOLD 254 +#define SQR_TOOM6_THRESHOLD 614 +#define SQR_TOOM8_THRESHOLD 842 + +#define MULMID_TOOM42_THRESHOLD 58 + +#define MULMOD_BNM1_THRESHOLD 19 +#define SQRMOD_BNM1_THRESHOLD 23 + +#define MUL_FFT_MODF_THRESHOLD 824 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 824, 5}, { 29, 6}, { 15, 5}, { 33, 6}, \ + { 17, 5}, { 36, 6}, { 19, 5}, { 39, 6}, \ + { 29, 7}, { 15, 6}, { 32, 7}, { 17, 6}, \ + { 36, 7}, { 19, 6}, { 39, 7}, { 21, 6}, \ + { 43, 7}, { 23, 6}, { 48, 7}, { 29, 8}, \ + { 15, 7}, { 37, 8}, { 19, 7}, { 43, 8}, \ + { 23, 7}, { 49, 8}, { 27, 7}, { 55, 8}, \ + { 31, 7}, { 63, 8}, { 43, 9}, { 23, 8}, \ + { 55, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 79, 9}, { 47, 8}, { 99, 9}, { 55,10}, \ + { 31, 9}, { 79,10}, { 47, 9}, { 103,11}, \ + { 31,10}, { 63, 9}, { 143,10}, { 79, 9}, \ + { 167,10}, { 95, 9}, { 191,10}, { 111,11}, \ + { 63,10}, { 127, 9}, { 255,10}, { 159,11}, \ + { 95,10}, { 191,12}, { 63,11}, { 127,10}, \ + { 271,11}, { 159,10}, { 319, 9}, { 639,10}, \ + { 335,11}, { 191,10}, { 383, 9}, { 799,10}, \ + { 415,11}, { 223,12}, { 127,11}, { 255,10}, \ + { 527,11}, { 287,10}, { 607, 9}, { 1215,11}, \ + { 319,10}, { 671,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,10}, { 863,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1119, 9}, { 2239,11}, \ + { 607,10}, { 1215,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,10}, { 1471, 9}, { 2943,12}, \ + { 383,11}, { 799,10}, { 1599,11}, { 863,12}, \ + { 447,11}, { 927,10}, { 1855,11}, { 959,13}, \ + { 255,12}, { 511,11}, { 1119,12}, { 575,11}, \ + { 1215,10}, { 2431,11}, { 1247,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,10}, { 2943,13}, \ + { 383,12}, { 767,11}, { 1599,12}, { 831,11}, \ + { 1727,10}, { 3455,12}, { 895,14}, { 255,13}, \ + { 511,12}, { 1087,11}, { 2239,10}, { 4479,12}, \ + { 1215,11}, { 2431,13}, { 639,12}, { 1343,11}, \ + { 2687,12}, { 1471,11}, { 2943,13}, { 767,12}, \ + { 1727,11}, { 3455,13}, { 895,12}, { 1983,14}, \ + { 511,13}, { 1023,12}, { 2239,13}, { 1151,12}, \ + { 2495,11}, { 4991,13}, { 1407,12}, { 2943,14}, \ + { 767,13}, { 1535,12}, { 3135,13}, { 1663,12}, \ + { 3455,13}, { 1919,12}, { 3967,15}, { 511,14}, \ + { 1023,13}, { 2175,12}, { 4479,13}, { 2431,12}, \ + { 4991,14}, { 1279,13}, { 2687,12}, { 5503,13}, \ + { 8192,14}, { 16384,15}, { 32768,16} } +#define MUL_FFT_TABLE3_SIZE 167 +#define MUL_FFT_THRESHOLD 7808 + +#define SQR_FFT_MODF_THRESHOLD 560 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 560, 5}, { 33, 6}, { 17, 5}, { 35, 6}, \ + { 33, 7}, { 17, 6}, { 36, 7}, { 19, 6}, \ + { 39, 7}, { 35, 8}, { 19, 7}, { 43, 8}, \ + { 23, 7}, { 47, 8}, { 27, 7}, { 55, 8}, \ + { 31, 7}, { 63, 8}, { 43, 9}, { 23, 8}, \ + { 55, 9}, { 31, 8}, { 67, 9}, { 39, 8}, \ + { 79, 9}, { 47, 8}, { 95, 9}, { 55,10}, \ + { 31, 9}, { 79,10}, { 47, 9}, { 95,11}, \ + { 31,10}, { 63, 9}, { 135,10}, { 79, 9}, \ + { 159,10}, { 111,11}, { 63,10}, { 143, 9}, \ + { 287,10}, { 159,11}, { 95,10}, { 191,12}, \ + { 63,11}, { 127, 9}, { 511, 8}, { 1023, 9}, \ + { 527,11}, { 159,10}, { 319, 9}, { 639,10}, \ + { 351,11}, { 191,10}, { 431,11}, { 223,12}, \ + { 127,11}, { 255,10}, { 543,11}, { 287,10}, \ + { 607, 9}, { 1215,11}, { 319,10}, { 639,11}, \ + { 351,12}, { 191,11}, { 383,10}, { 767,11}, \ + { 415,10}, { 831,13}, { 127,12}, { 255,11}, \ + { 543,10}, { 1119,11}, { 607,12}, { 319,11}, \ + { 671,10}, { 1343,11}, { 735,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,12}, { 447,11}, \ + { 927,10}, { 1855,11}, { 991,13}, { 255,12}, \ + { 511,11}, { 1055,10}, { 2111,11}, { 1087,12}, \ + { 575,11}, { 1215,10}, { 2431,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1407,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,10}, \ + { 3455,12}, { 895,11}, { 1855,12}, { 959,14}, \ + { 255,13}, { 511,12}, { 1087,11}, { 2239,12}, \ + { 1215,11}, { 2431,13}, { 639,12}, { 1471,11}, \ + { 2943,13}, { 767,12}, { 1727,11}, { 3455,13}, \ + { 895,12}, { 1983,14}, { 511,13}, { 1023,12}, \ + { 2239,13}, { 1151,12}, { 2495,11}, { 4991,13}, \ + { 1279,12}, { 2623,13}, { 1407,12}, { 2943,14}, \ + { 767,13}, { 1663,12}, { 3455,13}, { 1919,12}, \ + { 3839,15}, { 511,14}, { 1023,13}, { 2175,12}, \ + { 4479,13}, { 2431,12}, { 4991,14}, { 1279,13}, \ + { 2687,12}, { 5503,13}, { 8192,14}, { 16384,15}, \ + { 32768,16} } +#define SQR_FFT_TABLE3_SIZE 149 +#define SQR_FFT_THRESHOLD 4800 + +#define MULLO_BASECASE_THRESHOLD 12 +#define MULLO_DC_THRESHOLD 44 +#define MULLO_MUL_N_THRESHOLD 14281 +#define SQRLO_BASECASE_THRESHOLD 13 +#define SQRLO_DC_THRESHOLD 42 +#define SQRLO_SQR_THRESHOLD 9449 + +#define DC_DIV_QR_THRESHOLD 38 +#define DC_DIVAPPR_Q_THRESHOLD 105 +#define DC_BDIV_QR_THRESHOLD 52 +#define DC_BDIV_Q_THRESHOLD 83 + +#define INV_MULMOD_BNM1_THRESHOLD 50 +#define INV_NEWTON_THRESHOLD 158 +#define INV_APPR_THRESHOLD 118 + +#define BINV_NEWTON_THRESHOLD 342 +#define REDC_1_TO_REDC_N_THRESHOLD 67 + +#define MU_DIV_QR_THRESHOLD 2130 +#define MU_DIVAPPR_Q_THRESHOLD 1895 +#define MUPI_DIV_QR_THRESHOLD 60 +#define MU_BDIV_QR_THRESHOLD 1652 +#define MU_BDIV_Q_THRESHOLD 2089 + +#define POWM_SEC_TABLE 1,22,96,446,723,1378 + +#define GET_STR_DC_THRESHOLD 13 +#define GET_STR_PRECOMPUTE_THRESHOLD 20 +#define SET_STR_DC_THRESHOLD 298 +#define SET_STR_PRECOMPUTE_THRESHOLD 960 + +#define FAC_DSC_THRESHOLD 212 +#define FAC_ODD_THRESHOLD 71 + +#define MATRIX22_STRASSEN_THRESHOLD 26 +#define HGCD2_DIV1_METHOD 3 /* 0.68% faster than 1 */ +#define HGCD_THRESHOLD 80 +#define HGCD_APPR_THRESHOLD 138 +#define HGCD_REDUCE_THRESHOLD 4455 +#define GCD_DC_THRESHOLD 365 +#define GCDEXT_DC_THRESHOLD 245 +#define JACOBI_BASE_METHOD 4 /* 23.41% faster than 1 */ + +/* Tuneup completed successfully, took 63807 seconds */ diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/mod_1_1.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/mod_1_1.asm new file mode 100644 index 0000000..ee88bab --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/mod_1_1.asm @@ -0,0 +1,166 @@ +dnl x86-32 mpn_mod_1_1p for Pentium 4 and P6 models with SSE2 (i.e., 9,D,E,F). + +dnl Contributed to the GNU project by Torbjorn Granlund. + +dnl Copyright 2009, 2010 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C TODO: +C * Optimize. The present code was written quite straightforwardly. +C * Optimize post-loop reduction code; it is from mod_1s_4p, thus overkill. +C * Write a cps function that uses sse2 insns. + +C cycles/limb +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) ? +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) 16 +C P4 model 3-4 (Prescott) 18 + +C INPUT PARAMETERS +C ap sp + 4 +C n sp + 8 +C b sp + 12 +C cps sp + 16 + +define(`B1modb', `%mm1') +define(`B2modb', `%mm2') +define(`ap', `%edx') +define(`n', `%eax') + + TEXT + ALIGN(16) +PROLOGUE(mpn_mod_1_1p) + push %ebx + mov 8(%esp), ap + mov 12(%esp), n + mov 20(%esp), %ecx + movd 8(%ecx), B1modb + movd 12(%ecx), B2modb + + lea -4(ap,n,4), ap + +C FIXME: See comment in generic/mod_1_1.c. + movd (ap), %mm7 + movd -4(ap), %mm4 + pmuludq B1modb, %mm7 + paddq %mm4, %mm7 + add $-2, n + jz L(end) + + ALIGN(8) +L(top): movq %mm7, %mm6 + psrlq $32, %mm7 C rh + movd -8(ap), %mm0 + add $-4, ap + pmuludq B2modb, %mm7 + pmuludq B1modb, %mm6 + add $-1, n + paddq %mm0, %mm7 + paddq %mm6, %mm7 + jnz L(top) + +L(end): pcmpeqd %mm4, %mm4 + psrlq $32, %mm4 C 0x00000000FFFFFFFF + pand %mm7, %mm4 C rl + psrlq $32, %mm7 C rh + pmuludq B1modb, %mm7 C rh,cl + paddq %mm4, %mm7 C rh,rl + movd 4(%ecx), %mm4 C cnt + psllq %mm4, %mm7 C rh,rl normalized + movq %mm7, %mm2 C rl in low half + psrlq $32, %mm7 C rh + movd (%ecx), %mm1 C bi + pmuludq %mm7, %mm1 C qh,ql + paddq %mm2, %mm1 C qh-1,ql + movd %mm1, %ecx C ql + psrlq $32, %mm1 C qh-1 + movd 16(%esp), %mm3 C b + pmuludq %mm1, %mm3 C (qh-1) * b + psubq %mm3, %mm2 C r in low half (could use psubd) + movd %mm2, %eax C r + mov 16(%esp), %ebx + sub %ebx, %eax C r + cmp %eax, %ecx + lea (%eax,%ebx), %edx + cmovc( %edx, %eax) + movd %mm4, %ecx C cnt + cmp %ebx, %eax + jae L(fix) + emms + pop %ebx + shr %cl, %eax + ret + +L(fix): sub %ebx, %eax + emms + pop %ebx + shr %cl, %eax + ret +EPILOGUE() + +PROLOGUE(mpn_mod_1_1p_cps) +C CAUTION: This is the same code as in k7/mod_1_1.asm + push %ebp + mov 12(%esp), %ebp + push %esi + bsr %ebp, %ecx + push %ebx + xor $31, %ecx + mov 16(%esp), %esi + sal %cl, %ebp + mov %ebp, %edx + not %edx + mov $-1, %eax + div %ebp + mov %eax, (%esi) C store bi + mov %ecx, 4(%esi) C store cnt + xor %ebx, %ebx + sub %ebp, %ebx + mov $1, %edx + shld %cl, %eax, %edx + imul %edx, %ebx + mul %ebx + add %ebx, %edx + not %edx + imul %ebp, %edx + add %edx, %ebp + cmp %edx, %eax + cmovc( %ebp, %edx) + shr %cl, %ebx + mov %ebx, 8(%esi) C store B1modb + shr %cl, %edx + mov %edx, 12(%esi) C store B2modb + pop %ebx + pop %esi + pop %ebp + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/mod_1_4.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/mod_1_4.asm new file mode 100644 index 0000000..eb2edb6 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/mod_1_4.asm @@ -0,0 +1,269 @@ +dnl x86-32 mpn_mod_1s_4p for Pentium 4 and P6 models with SSE2 (i.e. 9,D,E,F). + +dnl Contributed to the GNU project by Torbjorn Granlund. + +dnl Copyright 2009, 2010 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C TODO: +C * Optimize. The present code was written quite straightforwardly. +C * Optimize post-loop reduction code. +C * Write a cps function that uses sse2 insns. + +C cycles/limb +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) 3.4 +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) 4 +C P4 model 3-4 (Prescott) 4.5 + +C INPUT PARAMETERS +C ap sp + 4 +C n sp + 8 +C b sp + 12 +C cps sp + 16 + +define(`B1modb', `%mm1') +define(`B2modb', `%mm2') +define(`B3modb', `%mm3') +define(`B4modb', `%mm4') +define(`B5modb', `%mm5') +define(`ap', `%edx') +define(`n', `%eax') + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(mpn_mod_1s_4p) + push %ebx + mov 8(%esp), ap + mov 12(%esp), n + mov 20(%esp), %ecx + + movd 8(%ecx), B1modb + movd 12(%ecx), B2modb + movd 16(%ecx), B3modb + movd 20(%ecx), B4modb + movd 24(%ecx), B5modb + + mov n, %ebx + lea -4(ap,n,4), ap + and $3, %ebx + je L(b0) + cmp $2, %ebx + jc L(b1) + je L(b2) + +L(b3): movd -4(ap), %mm7 + pmuludq B1modb, %mm7 + movd -8(ap), %mm6 + paddq %mm6, %mm7 + movd (ap), %mm6 + pmuludq B2modb, %mm6 + paddq %mm6, %mm7 + lea -24(ap), ap + add $-3, n + jz L(end) + jmp L(top) + +L(b0): movd -8(ap), %mm7 + pmuludq B1modb, %mm7 + movd -12(ap), %mm6 + paddq %mm6, %mm7 + movd -4(ap), %mm6 + pmuludq B2modb, %mm6 + paddq %mm6, %mm7 + movd (ap), %mm6 + pmuludq B3modb, %mm6 + paddq %mm6, %mm7 + lea -28(ap), ap + add $-4, n + jz L(end) + jmp L(top) + +L(b1): movd (ap), %mm7 + lea -16(ap), ap + dec n + jz L(x) + jmp L(top) + +L(b2): movd -4(ap), %mm7 C rl + punpckldq (ap), %mm7 C rh + lea -20(ap), ap + add $-2, n + jz L(end) + + ALIGN(8) +L(top): movd 4(ap), %mm0 + pmuludq B1modb, %mm0 + movd 0(ap), %mm6 + paddq %mm6, %mm0 + + movd 8(ap), %mm6 + pmuludq B2modb, %mm6 + paddq %mm6, %mm0 + + movd 12(ap), %mm6 + pmuludq B3modb, %mm6 + paddq %mm6, %mm0 + + movq %mm7, %mm6 + psrlq $32, %mm7 C rh + pmuludq B5modb, %mm7 + pmuludq B4modb, %mm6 + + paddq %mm0, %mm7 + paddq %mm6, %mm7 + + add $-16, ap + add $-4, n + jnz L(top) + +L(end): pcmpeqd %mm4, %mm4 + psrlq $32, %mm4 C 0x00000000FFFFFFFF + pand %mm7, %mm4 C rl + psrlq $32, %mm7 C rh + pmuludq B1modb, %mm7 C rh,cl + paddq %mm4, %mm7 C rh,rl +L(x): movd 4(%ecx), %mm4 C cnt + psllq %mm4, %mm7 C rh,rl normalized + movq %mm7, %mm2 C rl in low half + psrlq $32, %mm7 C rh + movd (%ecx), %mm1 C bi + pmuludq %mm7, %mm1 C qh,ql + paddq %mm2, %mm1 C qh-1,ql + movd %mm1, %ecx C ql + psrlq $32, %mm1 C qh-1 + movd 16(%esp), %mm3 C b + pmuludq %mm1, %mm3 C (qh-1) * b + psubq %mm3, %mm2 C r in low half (could use psubd) + movd %mm2, %eax C r + mov 16(%esp), %ebx + sub %ebx, %eax C r + cmp %eax, %ecx + lea (%eax,%ebx), %edx + cmovc( %edx, %eax) + movd %mm4, %ecx C cnt + cmp %ebx, %eax + jae L(fix) + emms + pop %ebx + shr %cl, %eax + ret + +L(fix): sub %ebx, %eax + emms + pop %ebx + shr %cl, %eax + ret +EPILOGUE() + + ALIGN(16) +PROLOGUE(mpn_mod_1s_4p_cps) +C CAUTION: This is the same code as in k7/mod_1_4.asm + push %ebp + push %edi + push %esi + push %ebx + mov 20(%esp), %ebp C FIXME: avoid bp for 0-idx + mov 24(%esp), %ebx + bsr %ebx, %ecx + xor $31, %ecx + sal %cl, %ebx C b << cnt + mov %ebx, %edx + not %edx + mov $-1, %eax + div %ebx + xor %edi, %edi + sub %ebx, %edi + mov $1, %esi + mov %eax, (%ebp) C store bi + mov %ecx, 4(%ebp) C store cnt + shld %cl, %eax, %esi + imul %edi, %esi + mov %eax, %edi + mul %esi + + add %esi, %edx + shr %cl, %esi + mov %esi, 8(%ebp) C store B1modb + + not %edx + imul %ebx, %edx + lea (%edx,%ebx), %esi + cmp %edx, %eax + cmovnc( %edx, %esi) + mov %edi, %eax + mul %esi + + add %esi, %edx + shr %cl, %esi + mov %esi, 12(%ebp) C store B2modb + + not %edx + imul %ebx, %edx + lea (%edx,%ebx), %esi + cmp %edx, %eax + cmovnc( %edx, %esi) + mov %edi, %eax + mul %esi + + add %esi, %edx + shr %cl, %esi + mov %esi, 16(%ebp) C store B3modb + + not %edx + imul %ebx, %edx + lea (%edx,%ebx), %esi + cmp %edx, %eax + cmovnc( %edx, %esi) + mov %edi, %eax + mul %esi + + add %esi, %edx + shr %cl, %esi + mov %esi, 20(%ebp) C store B4modb + + not %edx + imul %ebx, %edx + add %edx, %ebx + cmp %edx, %eax + cmovnc( %edx, %ebx) + + shr %cl, %ebx + mov %ebx, 24(%ebp) C store B5modb + + pop %ebx + pop %esi + pop %edi + pop %ebp + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/mod_34lsub1.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/mod_34lsub1.asm new file mode 100644 index 0000000..31e25b7 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/mod_34lsub1.asm @@ -0,0 +1,175 @@ +dnl Intel Pentium 4 mpn_mod_34lsub1 -- remainder modulo 2^24-1. + +dnl Copyright 2000-2003 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C Pentium4: 1.0 cycles/limb + + +C mp_limb_t mpn_mod_34lsub1 (mp_srcptr src, mp_size_t size) +C +C Enhancements: +C +C There might a couple of cycles to save by using plain integer code for +C more small sizes. 2 limbs measures about 20 cycles, but 3 limbs jumps to +C about 46 (inclusive of some function call overheads). + +defframe(PARAM_SIZE, 8) +defframe(PARAM_SRC, 4) + +dnl re-use parameter space +define(SAVE_EBX, `PARAM_SRC') +define(SAVE_ESI, `PARAM_SIZE') + + TEXT + ALIGN(16) +PROLOGUE(mpn_mod_34lsub1) +deflit(`FRAME',0) + + movl PARAM_SIZE, %ecx + movl PARAM_SRC, %edx + movl (%edx), %eax + + subl $2, %ecx + ja L(three_or_more) + jne L(one) + + movl 4(%edx), %edx + movl %eax, %ecx + shrl $24, %eax C src[0] high + + andl $0x00FFFFFF, %ecx C src[0] low + addl %ecx, %eax + + movl %edx, %ecx + shll $8, %edx + + shrl $16, %ecx C src[1] low + addl %ecx, %eax + + andl $0x00FFFF00, %edx C src[1] high + addl %edx, %eax + +L(one): + ret + + +L(three_or_more): + pxor %mm0, %mm0 + pxor %mm1, %mm1 + pxor %mm2, %mm2 + + pcmpeqd %mm7, %mm7 + psrlq $32, %mm7 C 0x00000000FFFFFFFF, low 32 bits + + pcmpeqd %mm6, %mm6 + psrlq $40, %mm6 C 0x0000000000FFFFFF, low 24 bits + +L(top): + C eax + C ebx + C ecx counter, size-2 to 0, -1 or -2 + C edx src, incrementing + C + C mm0 sum 0mod3 + C mm1 sum 1mod3 + C mm2 sum 2mod3 + C mm3 + C mm4 + C mm5 + C mm6 0x0000000000FFFFFF + C mm7 0x00000000FFFFFFFF + + movd (%edx), %mm3 + paddq %mm3, %mm0 + + movd 4(%edx), %mm3 + paddq %mm3, %mm1 + + movd 8(%edx), %mm3 + paddq %mm3, %mm2 + + addl $12, %edx + subl $3, %ecx + ja L(top) + + + C ecx is -2, -1 or 0 representing 0, 1 or 2 more limbs, respectively + + addl $1, %ecx + js L(combine) C 0 more + + movd (%edx), %mm3 + paddq %mm3, %mm0 + + jz L(combine) C 1 more + + movd 4(%edx), %mm3 + paddq %mm3, %mm1 + +L(combine): + movq %mm7, %mm3 C low halves + pand %mm0, %mm3 + + movq %mm7, %mm4 + pand %mm1, %mm4 + + movq %mm7, %mm5 + pand %mm2, %mm5 + + psrlq $32, %mm0 C high halves + psrlq $32, %mm1 + psrlq $32, %mm2 + + paddq %mm0, %mm4 C fold high halves to give 33 bits each + paddq %mm1, %mm5 + paddq %mm2, %mm3 + + psllq $8, %mm4 C combine at respective offsets + psllq $16, %mm5 + paddq %mm4, %mm3 + paddq %mm5, %mm3 C 0x000cxxxxxxxxxxxx, 50 bits + + pand %mm3, %mm6 C fold at 24 bits + psrlq $24, %mm3 + + paddq %mm6, %mm3 + movd %mm3, %eax + + ASSERT(z, C nothing left in high dword + `psrlq $32, %mm3 + movd %mm3, %ecx + orl %ecx, %ecx') + + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/mode1o.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/mode1o.asm new file mode 100644 index 0000000..aa9ef31 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/mode1o.asm @@ -0,0 +1,175 @@ +dnl Intel Pentium-4 mpn_modexact_1_odd -- mpn by limb exact remainder. + +dnl Copyright 2001, 2002, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C P4: 19.0 cycles/limb + + +C mp_limb_t mpn_modexact_1_odd (mp_srcptr src, mp_size_t size, +C mp_limb_t divisor); +C mp_limb_t mpn_modexact_1c_odd (mp_srcptr src, mp_size_t size, +C mp_limb_t divisor, mp_limb_t carry); +C + +defframe(PARAM_CARRY, 16) +defframe(PARAM_DIVISOR,12) +defframe(PARAM_SIZE, 8) +defframe(PARAM_SRC, 4) + + TEXT + + ALIGN(16) +PROLOGUE(mpn_modexact_1c_odd) +deflit(`FRAME',0) + + movd PARAM_CARRY, %mm1 + jmp L(start_1c) + +EPILOGUE() + + + ALIGN(16) +PROLOGUE(mpn_modexact_1_odd) +deflit(`FRAME',0) + + pxor %mm1, %mm1 C carry limb +L(start_1c): + movl PARAM_DIVISOR, %eax + + movd PARAM_DIVISOR, %mm7 + + shrl %eax + + andl $127, %eax C d/2, 7 bits + +ifdef(`PIC',` + LEA( binvert_limb_table, %edx) + movzbl (%eax,%edx), %eax C inv 8 bits +',` + movzbl binvert_limb_table(%eax), %eax C inv 8 bits +') + + C + + movd %eax, %mm6 C inv + + movd %eax, %mm0 C inv + + pmuludq %mm6, %mm6 C inv*inv + + C + + pmuludq %mm7, %mm6 C inv*inv*d + paddd %mm0, %mm0 C 2*inv + + C + + psubd %mm6, %mm0 C inv = 2*inv - inv*inv*d + pxor %mm6, %mm6 + + paddd %mm0, %mm6 + pmuludq %mm0, %mm0 C inv*inv + + C + + pmuludq %mm7, %mm0 C inv*inv*d + paddd %mm6, %mm6 C 2*inv + + + movl PARAM_SRC, %eax + movl PARAM_SIZE, %ecx + + C + + psubd %mm0, %mm6 C inv = 2*inv - inv*inv*d + + ASSERT(e,` C expect d*inv == 1 mod 2^GMP_LIMB_BITS + pushl %eax FRAME_pushl() + movd %mm6, %eax + imul PARAM_DIVISOR, %eax + cmpl $1, %eax + popl %eax FRAME_popl()') + + pxor %mm0, %mm0 C carry bit + + +C The dependent chain here is as follows. +C +C latency +C psubq s = (src-cbit) - climb 2 +C pmuludq q = s*inverse 8 +C pmuludq prod = q*divisor 8 +C psrlq climb = high(prod) 2 +C -- +C 20 +C +C Yet the loop measures 19.0 c/l, so obviously there's something gained +C there over a straight reading of the chip documentation. + +L(top): + C eax src, incrementing + C ebx + C ecx counter, limbs + C edx + C + C mm0 carry bit + C mm1 carry limb + C mm6 inverse + C mm7 divisor + + movd (%eax), %mm2 + addl $4, %eax + + psubq %mm0, %mm2 C src - cbit + + psubq %mm1, %mm2 C src - cbit - climb + movq %mm2, %mm0 + psrlq $63, %mm0 C new cbit + + pmuludq %mm6, %mm2 C s*inverse + + movq %mm7, %mm1 + pmuludq %mm2, %mm1 C q*divisor + psrlq $32, %mm1 C new climb + + subl $1, %ecx + jnz L(top) + + +L(done): + paddq %mm1, %mm0 + movd %mm0, %eax + emms + ret + +EPILOGUE() +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/mul_1.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/mul_1.asm new file mode 100644 index 0000000..6347b8b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/mul_1.asm @@ -0,0 +1,164 @@ +dnl mpn_mul_1 for Pentium 4 and P6 models with SSE2 (i.e., 9,D,E,F). + +dnl Copyright 2005, 2007, 2011 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + + +include(`../config.m4') + +C cycles/limb +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) 4.17 +C P6 model 13 (Dothan) 4.17 +C P4 model 0-1 (Willamette) 4 +C P4 model 2 (Northwood) 4 +C P4 model 3-4 (Prescott) 4.55 + +C TODO: +C * Tweak eax/edx offsets in loop as to save some lea's +C * Perhaps software pipeline small-case code + +C INPUT PARAMETERS +C rp sp + 4 +C up sp + 8 +C n sp + 12 +C v0 sp + 16 + + TEXT + ALIGN(16) +PROLOGUE(mpn_mul_1) + pxor %mm6, %mm6 +L(ent): mov 4(%esp), %edx + mov 8(%esp), %eax + mov 12(%esp), %ecx + movd 16(%esp), %mm7 + cmp $4, %ecx + jnc L(big) + +L(lp0): movd (%eax), %mm0 + lea 4(%eax), %eax + lea 4(%edx), %edx + pmuludq %mm7, %mm0 + paddq %mm0, %mm6 + movd %mm6, -4(%edx) + psrlq $32, %mm6 + dec %ecx + jnz L(lp0) + movd %mm6, %eax + emms + ret + +L(big): and $3, %ecx + je L(0) + cmp $2, %ecx + jc L(1) + je L(2) + jmp L(3) C FIXME: one case should fall through + +L(0): movd (%eax), %mm3 + sub 12(%esp), %ecx C loop count + lea -16(%eax), %eax + lea -12(%edx), %edx + pmuludq %mm7, %mm3 + movd 20(%eax), %mm0 + pmuludq %mm7, %mm0 + movd 24(%eax), %mm1 + jmp L(00) + +L(1): movd (%eax), %mm2 + sub 12(%esp), %ecx + lea -12(%eax), %eax + lea -8(%edx), %edx + pmuludq %mm7, %mm2 + movd 16(%eax), %mm3 + pmuludq %mm7, %mm3 + movd 20(%eax), %mm0 + jmp L(01) + +L(2): movd (%eax), %mm1 + sub 12(%esp), %ecx + lea -8(%eax), %eax + lea -4(%edx), %edx + pmuludq %mm7, %mm1 + movd 12(%eax), %mm2 + pmuludq %mm7, %mm2 + movd 16(%eax), %mm3 + jmp L(10) + +L(3): movd (%eax), %mm0 + sub 12(%esp), %ecx + lea -4(%eax), %eax + pmuludq %mm7, %mm0 + movd 8(%eax), %mm1 + pmuludq %mm7, %mm1 + movd 12(%eax), %mm2 + + ALIGN(16) +L(top): pmuludq %mm7, %mm2 + paddq %mm0, %mm6 + movd 16(%eax), %mm3 + movd %mm6, 0(%edx) + psrlq $32, %mm6 +L(10): pmuludq %mm7, %mm3 + paddq %mm1, %mm6 + movd 20(%eax), %mm0 + movd %mm6, 4(%edx) + psrlq $32, %mm6 +L(01): pmuludq %mm7, %mm0 + paddq %mm2, %mm6 + movd 24(%eax), %mm1 + movd %mm6, 8(%edx) + psrlq $32, %mm6 +L(00): pmuludq %mm7, %mm1 + paddq %mm3, %mm6 + movd 28(%eax), %mm2 + movd %mm6, 12(%edx) + psrlq $32, %mm6 + lea 16(%eax), %eax + lea 16(%edx), %edx + add $4, %ecx + ja L(top) + +L(end): pmuludq %mm7, %mm2 + paddq %mm0, %mm6 + movd %mm6, 0(%edx) + psrlq $32, %mm6 + paddq %mm1, %mm6 + movd %mm6, 4(%edx) + psrlq $32, %mm6 + paddq %mm2, %mm6 + movd %mm6, 8(%edx) + psrlq $32, %mm6 + movd %mm6, %eax + emms + ret +EPILOGUE() +PROLOGUE(mpn_mul_1c) + movd 20(%esp), %mm6 + jmp L(ent) +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/mul_basecase.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/mul_basecase.asm new file mode 100644 index 0000000..6e3775a --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/mul_basecase.asm @@ -0,0 +1,662 @@ +dnl mpn_mul_basecase for Pentium 4 and P6 models with SSE2 (i.e., 9,D,E,F). + +dnl Copyright 2001, 2002, 2005, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C TODO: +C * Improve ad-hoc outer loop code and register handling. Some feed-in +C scheduling could improve things by several cycles per outer iteration. +C * In code for un <= 3, try keeping accumulation operands in registers, +C without storing intermediates to rp. +C * We might want to keep 32 in a free mm register, since the register form is +C 3 bytes and the immediate form is 4 bytes. About 70 bytes to save. +C * Look into different loop alignment, we now expand the code about 50 bytes +C with possibly needless alignment. +C * Perhaps rewrap loops 00,01,02 (6 loops) to allow fall-through entry. +C * Use OSP, should solve feed-in latency problems. +C * Save a few tens of bytes by doing cross-jumping for Loel0, etc. +C * Save around 120 bytes by remapping "m 0", "m 1", "m 2" and "m 3" registers +C so that they can share feed-in code, and changing the branch targets from +C L to Lm. + +C cycles/limb +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) 5.24 +C P6 model 14 (Yonah) ? +C P4 model 0-1 (Willamette): 5 +C P4 model 2 (Northwood): 4.60 at 32 limbs +C P4 model 3-4 (Prescott): 4.94 at 32 limbs + +C INPUT PARAMETERS +C rp sp + 4 +C up sp + 8 +C un sp + 12 +C vp sp + 16 +C vn sp + 20 + + TEXT + ALIGN(16) +PROLOGUE(mpn_mul_basecase) + push %esi + push %ebx + mov 12(%esp), %edx C rp + mov 16(%esp), %eax C up + mov 20(%esp), %ecx C un + mov 24(%esp), %esi C vp + mov 28(%esp), %ebx C vn + movd (%esi), %mm7 C +L(ent): cmp $3, %ecx + ja L(big) + movd (%eax), %mm6 + pmuludq %mm7, %mm6 + jz L(un3) + cmp $2, %ecx + jz L(un2) + +L(un1): movd %mm6, (%edx) C un=1 + psrlq $32, %mm6 C un=1 + movd %mm6, 4(%edx) C un=1 + jmp L(rtr) C un=1 + +L(un2): movd 4(%eax), %mm1 C un=2 + pmuludq %mm7, %mm1 C un=2 + movd %mm6, (%edx) C un=2 + psrlq $32, %mm6 C un=2 + paddq %mm1, %mm6 C un=2 + movd %mm6, 4(%edx) C un=2 + psrlq $32, %mm6 C un=2 + movd %mm6, 8(%edx) C un=2 + dec %ebx C un=2 + jz L(rtr) C un=2 + movd 4(%esi), %mm7 C un=2 + movd (%eax), %mm6 C un=2 + pmuludq %mm7, %mm6 C un=2 + movd 4(%eax), %mm1 C un=2 + movd 4(%edx), %mm4 C un=2 + pmuludq %mm7, %mm1 C un=2 + movd 8(%edx), %mm5 C un=2 + paddq %mm4, %mm6 C un=2 + paddq %mm1, %mm5 C un=2 + movd %mm6, 4(%edx) C un=2 + psrlq $32, %mm6 C un=2 + paddq %mm5, %mm6 C un=2 + movd %mm6, 8(%edx) C un=2 + psrlq $32, %mm6 C un=2 + movd %mm6, 12(%edx) C un=2 +L(rtr): emms + pop %ebx + pop %esi + ret + +L(un3): movd 4(%eax), %mm1 C un=3 + pmuludq %mm7, %mm1 C un=3 + movd 8(%eax), %mm2 C un=3 + pmuludq %mm7, %mm2 C un=3 + movd %mm6, (%edx) C un=3 + psrlq $32, %mm6 C un=3 + paddq %mm1, %mm6 C un=3 + movd %mm6, 4(%edx) C un=3 + psrlq $32, %mm6 C un=3 + paddq %mm2, %mm6 C un=3 + movd %mm6, 8(%edx) C un=3 + psrlq $32, %mm6 C un=3 + movd %mm6, 12(%edx) C un=3 + dec %ebx C un=3 + jz L(rtr) C un=3 + movd 4(%esi), %mm7 C un=3 + movd (%eax), %mm6 C un=3 + pmuludq %mm7, %mm6 C un=3 + movd 4(%eax), %mm1 C un=3 + movd 4(%edx), %mm4 C un=3 + pmuludq %mm7, %mm1 C un=3 + movd 8(%eax), %mm2 C un=3 + movd 8(%edx), %mm5 C un=3 + pmuludq %mm7, %mm2 C un=3 + paddq %mm4, %mm6 C un=3 + paddq %mm1, %mm5 C un=3 + movd 12(%edx), %mm4 C un=3 + movd %mm6, 4(%edx) C un=3 + psrlq $32, %mm6 C un=3 + paddq %mm5, %mm6 C un=3 + paddq %mm2, %mm4 C un=3 + movd %mm6, 8(%edx) C un=3 + psrlq $32, %mm6 C un=3 + paddq %mm4, %mm6 C un=3 + movd %mm6, 12(%edx) C un=3 + psrlq $32, %mm6 C un=3 + movd %mm6, 16(%edx) C un=3 + dec %ebx C un=3 + jz L(rtr) C un=3 + movd 8(%esi), %mm7 C un=3 + movd (%eax), %mm6 C un=3 + pmuludq %mm7, %mm6 C un=3 + movd 4(%eax), %mm1 C un=3 + movd 8(%edx), %mm4 C un=3 + pmuludq %mm7, %mm1 C un=3 + movd 8(%eax), %mm2 C un=3 + movd 12(%edx), %mm5 C un=3 + pmuludq %mm7, %mm2 C un=3 + paddq %mm4, %mm6 C un=3 + paddq %mm1, %mm5 C un=3 + movd 16(%edx), %mm4 C un=3 + movd %mm6, 8(%edx) C un=3 + psrlq $32, %mm6 C un=3 + paddq %mm5, %mm6 C un=3 + paddq %mm2, %mm4 C un=3 + movd %mm6, 12(%edx) C un=3 + psrlq $32, %mm6 C un=3 + paddq %mm4, %mm6 C un=3 + movd %mm6, 16(%edx) C un=3 + psrlq $32, %mm6 C un=3 + movd %mm6, 20(%edx) C un=3 + jmp L(rtr) + + +L(big): push %edi + pxor %mm6, %mm6 + lea 4(%esi), %esi + and $3, %ecx + jz L(0) + cmp $2, %ecx + jc L(1) + jz L(2) + jmp L(3) C FIXME: one case should fall through + + +L(0): movd (%eax), %mm3 C m 0 + sub 24(%esp), %ecx C inner loop count m 0 + mov %ecx, 24(%esp) C update loop count for later m 0 + pmuludq %mm7, %mm3 C m 0 + movd 4(%eax), %mm0 C m 0 + pmuludq %mm7, %mm0 C m 0 + movd 8(%eax), %mm1 C m 0 + jmp L(m00) C m 0 + ALIGN(16) C m 0 +L(lpm0): + pmuludq %mm7, %mm4 C m 0 + paddq %mm0, %mm6 C m 0 + movd (%eax), %mm3 C m 0 + movd %mm6, -12(%edx) C m 0 + psrlq $32, %mm6 C m 0 + pmuludq %mm7, %mm3 C m 0 + paddq %mm1, %mm6 C m 0 + movd 4(%eax), %mm0 C m 0 + movd %mm6, -8(%edx) C m 0 + psrlq $32, %mm6 C m 0 + pmuludq %mm7, %mm0 C m 0 + paddq %mm4, %mm6 C m 0 + movd 8(%eax), %mm1 C m 0 + movd %mm6, -4(%edx) C m 0 + psrlq $32, %mm6 C m 0 +L(m00): pmuludq %mm7, %mm1 C m 0 + paddq %mm3, %mm6 C m 0 + movd 12(%eax), %mm4 C m 0 + movd %mm6, (%edx) C m 0 + psrlq $32, %mm6 C m 0 + lea 16(%eax), %eax C m 0 + lea 16(%edx), %edx C m 0 + add $4, %ecx C m 0 + ja L(lpm0) C m 0 + pmuludq %mm7, %mm4 C m 0 + paddq %mm0, %mm6 C m 0 + movd %mm6, -12(%edx) C m 0 + psrlq $32, %mm6 C m 0 + paddq %mm1, %mm6 C m 0 + mov 16(%esp), %edi C rp 0 + jmp L(x0) + +L(olp0): + lea 4(%edi), %edi C am 0 + movd (%esi), %mm7 C am 0 + lea 4(%esi), %esi C am 0 + mov %edi, %edx C rp am 0 + mov 20(%esp), %eax C up am 0 + movd (%eax), %mm3 C am 0 + mov 24(%esp), %ecx C inner loop count am 0 + pxor %mm6, %mm6 C am 0 + pmuludq %mm7, %mm3 C am 0 + movd 4(%eax), %mm0 C am 0 + movd (%edx), %mm5 C am 0 + pmuludq %mm7, %mm0 C am 0 + movd 8(%eax), %mm1 C am 0 + paddq %mm3, %mm5 C am 0 + movd 4(%edx), %mm4 C am 0 + jmp L(am00) C am 0 + ALIGN(16) C mm 0 +L(lam0): + pmuludq %mm7, %mm2 C am 0 + paddq %mm4, %mm6 C am 0 + movd (%eax), %mm3 C am 0 + paddq %mm1, %mm5 C am 0 + movd -4(%edx), %mm4 C am 0 + movd %mm6, -12(%edx) C am 0 + psrlq $32, %mm6 C am 0 + pmuludq %mm7, %mm3 C am 0 + paddq %mm5, %mm6 C am 0 + movd 4(%eax), %mm0 C am 0 + paddq %mm2, %mm4 C am 0 + movd (%edx), %mm5 C am 0 + movd %mm6, -8(%edx) C am 0 + psrlq $32, %mm6 C am 0 + pmuludq %mm7, %mm0 C am 0 + paddq %mm4, %mm6 C am 0 + movd 8(%eax), %mm1 C am 0 + paddq %mm3, %mm5 C am 0 + movd 4(%edx), %mm4 C am 0 + movd %mm6, -4(%edx) C am 0 + psrlq $32, %mm6 C am 0 +L(am00): + pmuludq %mm7, %mm1 C am 0 + paddq %mm5, %mm6 C am 0 + movd 12(%eax), %mm2 C am 0 + paddq %mm0, %mm4 C am 0 + movd 8(%edx), %mm5 C am 0 + movd %mm6, (%edx) C am 0 + psrlq $32, %mm6 C am 0 + lea 16(%eax), %eax C am 0 + lea 16(%edx), %edx C am 0 + add $4, %ecx C am 0 + jnz L(lam0) C am 0 + pmuludq %mm7, %mm2 C am 0 + paddq %mm4, %mm6 C am 0 + paddq %mm1, %mm5 C am 0 + movd -4(%edx), %mm4 C am 0 + movd %mm6, -12(%edx) C am 0 + psrlq $32, %mm6 C am 0 + paddq %mm5, %mm6 C am 0 + paddq %mm2, %mm4 C am 0 +L(x0): movd %mm6, -8(%edx) C am 0 + psrlq $32, %mm6 C am 0 + paddq %mm4, %mm6 C am 0 + movd %mm6, -4(%edx) C am 0 + psrlq $32, %mm6 C am 0 + movd %mm6, (%edx) C am 0 + dec %ebx C am 0 + jnz L(olp0) C am 0 +L(oel0): + emms C 0 + pop %edi C 0 + pop %ebx C 0 + pop %esi C 0 + ret C 0 + + +L(1): movd (%eax), %mm4 C m 1 + sub 24(%esp), %ecx C m 1 + mov %ecx, 24(%esp) C update loop count for later m 1 + pmuludq %mm7, %mm4 C m 1 + movd 4(%eax), %mm3 C m 1 + pmuludq %mm7, %mm3 C m 1 + movd 8(%eax), %mm0 C m 1 + jmp L(m01) C m 1 + ALIGN(16) C m 1 +L(lpm1): + pmuludq %mm7, %mm4 C m 1 + paddq %mm0, %mm6 C m 1 + movd 4(%eax), %mm3 C m 1 + movd %mm6, -8(%edx) C m 1 + psrlq $32, %mm6 C m 1 + pmuludq %mm7, %mm3 C m 1 + paddq %mm1, %mm6 C m 1 + movd 8(%eax), %mm0 C m 1 + movd %mm6, -4(%edx) C m 1 + psrlq $32, %mm6 C m 1 +L(m01): pmuludq %mm7, %mm0 C m 1 + paddq %mm4, %mm6 C m 1 + movd 12(%eax), %mm1 C m 1 + movd %mm6, (%edx) C m 1 + psrlq $32, %mm6 C m 1 + pmuludq %mm7, %mm1 C m 1 + paddq %mm3, %mm6 C m 1 + movd 16(%eax), %mm4 C m 1 + movd %mm6, 4(%edx) C m 1 + psrlq $32, %mm6 C m 1 + lea 16(%eax), %eax C m 1 + lea 16(%edx), %edx C m 1 + add $4, %ecx C m 1 + ja L(lpm1) C m 1 + pmuludq %mm7, %mm4 C m 1 + paddq %mm0, %mm6 C m 1 + movd %mm6, -8(%edx) C m 1 + psrlq $32, %mm6 C m 1 + paddq %mm1, %mm6 C m 1 + mov 16(%esp), %edi C rp 1 + jmp L(x1) + +L(olp1): + lea 4(%edi), %edi C am 1 + movd (%esi), %mm7 C am 1 + lea 4(%esi), %esi C am 1 + mov %edi, %edx C rp am 1 + mov 20(%esp), %eax C up am 1 + movd (%eax), %mm2 C am 1 + mov 24(%esp), %ecx C inner loop count am 1 + pxor %mm6, %mm6 C am 1 + pmuludq %mm7, %mm2 C am 1 + movd 4(%eax), %mm3 C am 1 + movd (%edx), %mm4 C am 1 + pmuludq %mm7, %mm3 C am 1 + movd 8(%eax), %mm0 C am 1 + paddq %mm2, %mm4 C am 1 + movd 4(%edx), %mm5 C am 1 + jmp L(am01) C am 1 + ALIGN(16) C am 1 +L(lam1): + pmuludq %mm7, %mm2 C am 1 + paddq %mm4, %mm6 C am 1 + movd 4(%eax), %mm3 C am 1 + paddq %mm1, %mm5 C am 1 + movd (%edx), %mm4 C am 1 + movd %mm6, -8(%edx) C am 1 + psrlq $32, %mm6 C am 1 + pmuludq %mm7, %mm3 C am 1 + paddq %mm5, %mm6 C am 1 + movd 8(%eax), %mm0 C am 1 + paddq %mm2, %mm4 C am 1 + movd 4(%edx), %mm5 C am 1 + movd %mm6, -4(%edx) C am 1 + psrlq $32, %mm6 C am 1 +L(am01): + pmuludq %mm7, %mm0 C am 1 + paddq %mm4, %mm6 C am 1 + movd 12(%eax), %mm1 C am 1 + paddq %mm3, %mm5 C am 1 + movd 8(%edx), %mm4 C am 1 + movd %mm6, (%edx) C am 1 + psrlq $32, %mm6 C am 1 + pmuludq %mm7, %mm1 C am 1 + paddq %mm5, %mm6 C am 1 + movd 16(%eax), %mm2 C am 1 + paddq %mm0, %mm4 C am 1 + movd 12(%edx), %mm5 C am 1 + movd %mm6, 4(%edx) C am 1 + psrlq $32, %mm6 C am 1 + lea 16(%eax), %eax C am 1 + lea 16(%edx), %edx C am 1 + add $4, %ecx C am 1 + jnz L(lam1) C am 1 + pmuludq %mm7, %mm2 C am 1 + paddq %mm4, %mm6 C am 1 + paddq %mm1, %mm5 C am 1 + movd (%edx), %mm4 C am 1 + movd %mm6, -8(%edx) C am 1 + psrlq $32, %mm6 C am 1 + paddq %mm5, %mm6 C am 1 + paddq %mm2, %mm4 C am 1 +L(x1): movd %mm6, -4(%edx) C am 1 + psrlq $32, %mm6 C am 1 + paddq %mm4, %mm6 C am 1 + movd %mm6, (%edx) C am 1 + psrlq $32, %mm6 C am 1 + movd %mm6, 4(%edx) C am 1 + dec %ebx C am 1 + jnz L(olp1) C am 1 +L(oel1): + emms C 1 + pop %edi C 1 + pop %ebx C 1 + pop %esi C 1 + ret C 1 + + +L(2): movd (%eax), %mm1 C m 2 + sub 24(%esp), %ecx C m 2 + mov %ecx, 24(%esp) C update loop count for later m 2 + pmuludq %mm7, %mm1 C m 2 + movd 4(%eax), %mm4 C m 2 + pmuludq %mm7, %mm4 C m 2 + movd 8(%eax), %mm3 C m 2 + jmp L(m10) C m 2 + ALIGN(16) C m 2 +L(lpm2): + pmuludq %mm7, %mm4 C m 2 + paddq %mm0, %mm6 C m 2 + movd 8(%eax), %mm3 C m 2 + movd %mm6, -4(%edx) C m 2 + psrlq $32, %mm6 C m 2 +L(m10): pmuludq %mm7, %mm3 C m 2 + paddq %mm1, %mm6 C m 2 + movd 12(%eax), %mm0 C m 2 + movd %mm6, (%edx) C m 2 + psrlq $32, %mm6 C m 2 + pmuludq %mm7, %mm0 C m 2 + paddq %mm4, %mm6 C m 2 + movd 16(%eax), %mm1 C m 2 + movd %mm6, 4(%edx) C m 2 + psrlq $32, %mm6 C m 2 + pmuludq %mm7, %mm1 C m 2 + paddq %mm3, %mm6 C m 2 + movd 20(%eax), %mm4 C m 2 + movd %mm6, 8(%edx) C m 2 + psrlq $32, %mm6 C m 2 + lea 16(%eax), %eax C m 2 + lea 16(%edx), %edx C m 2 + add $4, %ecx C m 2 + ja L(lpm2) C m 2 + pmuludq %mm7, %mm4 C m 2 + paddq %mm0, %mm6 C m 2 + movd %mm6, -4(%edx) C m 2 + psrlq $32, %mm6 C m 2 + paddq %mm1, %mm6 C m 2 + mov 16(%esp), %edi C rp 2 + jmp L(x2) + +L(olp2): + lea 4(%edi), %edi C am 2 + movd (%esi), %mm7 C am 2 + lea 4(%esi), %esi C am 2 + mov %edi, %edx C rp am 2 + mov 20(%esp), %eax C up am 2 + movd (%eax), %mm1 C am 2 + mov 24(%esp), %ecx C inner loop count am 2 + pxor %mm6, %mm6 C am 2 + pmuludq %mm7, %mm1 C am 2 + movd 4(%eax), %mm2 C am 2 + movd (%edx), %mm5 C am 2 + pmuludq %mm7, %mm2 C am 2 + movd 8(%eax), %mm3 C am 2 + paddq %mm1, %mm5 C am 2 + movd 4(%edx), %mm4 C am 2 + jmp L(am10) C am 2 + ALIGN(16) C am 2 +L(lam2): + pmuludq %mm7, %mm2 C am 2 + paddq %mm4, %mm6 C am 2 + movd 8(%eax), %mm3 C am 2 + paddq %mm1, %mm5 C am 2 + movd 4(%edx), %mm4 C am 2 + movd %mm6, -4(%edx) C am 2 + psrlq $32, %mm6 C am 2 +L(am10): + pmuludq %mm7, %mm3 C am 2 + paddq %mm5, %mm6 C am 2 + movd 12(%eax), %mm0 C am 2 + paddq %mm2, %mm4 C am 2 + movd 8(%edx), %mm5 C am 2 + movd %mm6, (%edx) C am 2 + psrlq $32, %mm6 C am 2 + pmuludq %mm7, %mm0 C am 2 + paddq %mm4, %mm6 C am 2 + movd 16(%eax), %mm1 C am 2 + paddq %mm3, %mm5 C am 2 + movd 12(%edx), %mm4 C am 2 + movd %mm6, 4(%edx) C am 2 + psrlq $32, %mm6 C am 2 + pmuludq %mm7, %mm1 C am 2 + paddq %mm5, %mm6 C am 2 + movd 20(%eax), %mm2 C am 2 + paddq %mm0, %mm4 C am 2 + movd 16(%edx), %mm5 C am 2 + movd %mm6, 8(%edx) C am 2 + psrlq $32, %mm6 C am 2 + lea 16(%eax), %eax C am 2 + lea 16(%edx), %edx C am 2 + add $4, %ecx C am 2 + jnz L(lam2) C am 2 + pmuludq %mm7, %mm2 C am 2 + paddq %mm4, %mm6 C am 2 + paddq %mm1, %mm5 C am 2 + movd 4(%edx), %mm4 C am 2 + movd %mm6, -4(%edx) C am 2 + psrlq $32, %mm6 C am 2 + paddq %mm5, %mm6 C am 2 + paddq %mm2, %mm4 C am 2 +L(x2): movd %mm6, (%edx) C am 2 + psrlq $32, %mm6 C am 2 + paddq %mm4, %mm6 C am 2 + movd %mm6, 4(%edx) C am 2 + psrlq $32, %mm6 C am 2 + movd %mm6, 8(%edx) C am 2 + dec %ebx C am 2 + jnz L(olp2) C am 2 +L(oel2): + emms C 2 + pop %edi C 2 + pop %ebx C 2 + pop %esi C 2 + ret C 2 + + +L(3): movd (%eax), %mm0 C m 3 + sub 24(%esp), %ecx C m 3 + mov %ecx, 24(%esp) C update loop count for later m 3 + pmuludq %mm7, %mm0 C m 3 + movd 4(%eax), %mm1 C m 3 + pmuludq %mm7, %mm1 C m 3 + movd 8(%eax), %mm4 C m 3 + jmp L(lpm3) C m 3 + ALIGN(16) C m 3 +L(lpm3): + pmuludq %mm7, %mm4 C m 3 + paddq %mm0, %mm6 C m 3 + movd 12(%eax), %mm3 C m 3 + movd %mm6, (%edx) C m 3 + psrlq $32, %mm6 C m 3 + pmuludq %mm7, %mm3 C m 3 + paddq %mm1, %mm6 C m 3 + movd 16(%eax), %mm0 C m 3 + movd %mm6, 4(%edx) C m 3 + psrlq $32, %mm6 C m 3 + pmuludq %mm7, %mm0 C m 3 + paddq %mm4, %mm6 C m 3 + movd 20(%eax), %mm1 C m 3 + movd %mm6, 8(%edx) C m 3 + psrlq $32, %mm6 C m 3 + pmuludq %mm7, %mm1 C m 3 + paddq %mm3, %mm6 C m 3 + movd 24(%eax), %mm4 C m 3 + movd %mm6, 12(%edx) C m 3 + psrlq $32, %mm6 C m 3 + lea 16(%eax), %eax C m 3 + lea 16(%edx), %edx C m 3 + add $4, %ecx C m 3 + ja L(lpm3) C m 3 + pmuludq %mm7, %mm4 C m 3 + paddq %mm0, %mm6 C m 3 + movd %mm6, (%edx) C m 3 + psrlq $32, %mm6 C m 3 + paddq %mm1, %mm6 C m 3 + mov 16(%esp), %edi C rp 3 + jmp L(x3) + +L(olp3): + lea 4(%edi), %edi C am 3 + movd (%esi), %mm7 C am 3 + lea 4(%esi), %esi C am 3 + mov %edi, %edx C rp am 3 + mov 20(%esp), %eax C up am 3 + movd (%eax), %mm0 C am 3 + mov 24(%esp), %ecx C inner loop count am 3 + pxor %mm6, %mm6 C am 3 + pmuludq %mm7, %mm0 C am 3 + movd 4(%eax), %mm1 C am 3 + movd (%edx), %mm4 C am 3 + pmuludq %mm7, %mm1 C am 3 + movd 8(%eax), %mm2 C am 3 + paddq %mm0, %mm4 C am 3 + movd 4(%edx), %mm5 C am 3 + jmp L(lam3) C am 3 + ALIGN(16) C am 3 +L(lam3): + pmuludq %mm7, %mm2 C am 3 + paddq %mm4, %mm6 C am 3 + movd 12(%eax), %mm3 C am 3 + paddq %mm1, %mm5 C am 3 + movd 8(%edx), %mm4 C am 3 + movd %mm6, (%edx) C am 3 + psrlq $32, %mm6 C am 3 + pmuludq %mm7, %mm3 C am 3 + paddq %mm5, %mm6 C am 3 + movd 16(%eax), %mm0 C am 3 + paddq %mm2, %mm4 C am 3 + movd 12(%edx), %mm5 C am 3 + movd %mm6, 4(%edx) C am 3 + psrlq $32, %mm6 C am 3 + pmuludq %mm7, %mm0 C am 3 + paddq %mm4, %mm6 C am 3 + movd 20(%eax), %mm1 C am 3 + paddq %mm3, %mm5 C am 3 + movd 16(%edx), %mm4 C am 3 + movd %mm6, 8(%edx) C am 3 + psrlq $32, %mm6 C am 3 + pmuludq %mm7, %mm1 C am 3 + paddq %mm5, %mm6 C am 3 + movd 24(%eax), %mm2 C am 3 + paddq %mm0, %mm4 C am 3 + movd 20(%edx), %mm5 C am 3 + movd %mm6, 12(%edx) C am 3 + psrlq $32, %mm6 C am 3 + lea 16(%eax), %eax C am 3 + lea 16(%edx), %edx C am 3 + add $4, %ecx C am 3 + jnz L(lam3) C am 3 + pmuludq %mm7, %mm2 C am 3 + paddq %mm4, %mm6 C am 3 + paddq %mm1, %mm5 C am 3 + movd 8(%edx), %mm4 C am 3 + movd %mm6, (%edx) C am 3 + psrlq $32, %mm6 C am 3 + paddq %mm5, %mm6 C am 3 + paddq %mm2, %mm4 C am 3 +L(x3): movd %mm6, 4(%edx) C am 3 + psrlq $32, %mm6 C am 3 + paddq %mm4, %mm6 C am 3 + movd %mm6, 8(%edx) C am 3 + psrlq $32, %mm6 C am 3 + movd %mm6, 12(%edx) C am 3 + dec %ebx C am 3 + jnz L(olp3) C am 3 +L(oel3): + emms C 3 + pop %edi C 3 + pop %ebx C 3 + pop %esi C 3 + ret C 3 +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/popcount.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/popcount.asm new file mode 100644 index 0000000..c7f4426 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/popcount.asm @@ -0,0 +1,281 @@ +dnl X86-32 and X86-64 mpn_popcount using SSE2. + +dnl Copyright 2006, 2007, 2011, 2015, 2020 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + + +include(`../config.m4') + + +C 32-bit popcount hamdist +C cycles/limb cycles/limb +C P5 - +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) 4 +C P4 model 0 (Willamette) ? +C P4 model 1 (?) ? +C P4 model 2 (Northwood) 3.9 +C P4 model 3 (Prescott) ? +C P4 model 4 (Nocona) ? +C AMD K6 - +C AMD K7 - +C AMD K8 ? + +C 64-bit popcount hamdist +C cycles/limb cycles/limb +C P4 model 4 (Nocona): 8 +C AMD K8,K9 7.5 +C AMD K10 3.5 +C Intel core2 3.68 +C Intel corei 3.15 +C Intel atom 10.8 +C VIA nano 6.5 + +C TODO +C * Make an mpn_hamdist based on this. Alignment could either be handled by +C using movdqu for one operand and movdqa for the other, or by painfully +C shifting as we go. Unfortunately, there seem to be no usable shift +C instruction, except for one that takes an immediate count. +C * It would probably be possible to cut a few cycles/limb using software +C pipelining. +C * There are 35 decode slots unused by the SSE2 instructions. Loop control +C needs just 2 or 3 slots, leaving around 32 slots. This allows a parallel +C integer based popcount. Such a combined loop would handle 6 limbs in +C about 30 cycles on K8. +C * We could save a byte or two by using 32-bit operations on areg. +C * Check if using movdqa to a temp of and then register-based pand is faster. + +ifelse(GMP_LIMB_BITS,`32', +` define(`up', `%edx') + define(`n', `%ecx') + define(`areg',`%eax') + define(`breg',`%ebx') + define(`zero',`%xmm4') + define(`LIMB32',` $1') + define(`LIMB64',`dnl') +',` + define(`up', `%rdi') + define(`n', `%rsi') + define(`areg',`%rax') + define(`breg',`%rdx') + define(`zero',`%xmm8') + define(`LIMB32',`dnl') + define(`LIMB64',` $1') +') + +define(`mm01010101',`%xmm6') +define(`mm00110011',`%xmm7') +define(`mm00001111',`%xmm2') + +define(`GMP_LIMB_BYTES', eval(GMP_LIMB_BITS/8)) +define(`LIMBS_PER_XMM', eval(16/GMP_LIMB_BYTES)) +define(`LIMBS_PER_2XMM', eval(32/GMP_LIMB_BYTES)) + +undefine(`psadbw') C override inherited m4 version + +C This file is shared between 32-bit and 64-bit builds. Only the former has +C LEAL. Default LEAL as an alias of LEA. +ifdef(`LEAL',,`define(`LEAL', `LEA($1,$2)')') + +ASM_START() + +C Make cnsts global to work around Apple relocation bug. +ifdef(`DARWIN',` + define(`cnsts', MPN(popccnsts)) + GLOBL cnsts') + + TEXT + ALIGN(32) +PROLOGUE(mpn_popcount) + +LIMB32(`mov 4(%esp), up ') +LIMB32(`mov 8(%esp), n ') +LIMB32(`push %ebx ') + + pxor %xmm3, %xmm3 C zero grand total count +LIMB64(`pxor zero, zero ') + + LEAL( cnsts, breg) + + movdqa -48(breg), mm01010101 + movdqa -32(breg), mm00110011 + movdqa -16(breg), mm00001111 + + mov up, areg + and $-16, up C round `up' down to 128-bit boundary + and $12, areg C 32:areg = 0, 4, 8, 12 + C 64:areg = 0, 8 + movdqa (up), %xmm0 + pand 64(breg,areg,4), %xmm0 + shr $m4_log2(GMP_LIMB_BYTES), %eax + add areg, n C compensate n for rounded down `up' + + pxor %xmm4, %xmm4 + sub $LIMBS_PER_XMM, n + jbe L(sum) + + sub $LIMBS_PER_XMM, n + ja L(ent) + jmp L(lsum) + + ALIGN(16) +L(top): movdqa (up), %xmm0 +L(ent): movdqa 16(up), %xmm4 + + movdqa %xmm0, %xmm1 + movdqa %xmm4, %xmm5 + psrld $1, %xmm0 + psrld $1, %xmm4 + pand mm01010101, %xmm0 + pand mm01010101, %xmm4 + psubd %xmm0, %xmm1 + psubd %xmm4, %xmm5 + + movdqa %xmm1, %xmm0 + movdqa %xmm5, %xmm4 + psrlq $2, %xmm1 + psrlq $2, %xmm5 + pand mm00110011, %xmm0 + pand mm00110011, %xmm4 + pand mm00110011, %xmm1 + pand mm00110011, %xmm5 + paddq %xmm0, %xmm1 + paddq %xmm4, %xmm5 + +LIMB32(`pxor zero, zero ') + + add $32, up + sub $LIMBS_PER_2XMM, n + + paddq %xmm5, %xmm1 + movdqa %xmm1, %xmm0 + psrlq $4, %xmm1 + pand mm00001111, %xmm0 + pand mm00001111, %xmm1 + paddq %xmm0, %xmm1 + + psadbw zero, %xmm1 + paddq %xmm1, %xmm3 C add to grand total + + jnc L(top) +L(end): + add $LIMBS_PER_2XMM, n + jz L(rt) + movdqa (up), %xmm0 + pxor %xmm4, %xmm4 + sub $LIMBS_PER_XMM, n + jbe L(sum) +L(lsum): + movdqa %xmm0, %xmm4 + movdqa 16(up), %xmm0 +L(sum): + shl $m4_log2(GMP_LIMB_BYTES), n + and $12, n + pand (breg,n,4), %xmm0 + + movdqa %xmm0, %xmm1 + movdqa %xmm4, %xmm5 + psrld $1, %xmm0 + psrld $1, %xmm4 + pand mm01010101, %xmm0 + pand mm01010101, %xmm4 + psubd %xmm0, %xmm1 + psubd %xmm4, %xmm5 + + movdqa %xmm1, %xmm0 + movdqa %xmm5, %xmm4 + psrlq $2, %xmm1 + psrlq $2, %xmm5 + pand mm00110011, %xmm0 + pand mm00110011, %xmm4 + pand mm00110011, %xmm1 + pand mm00110011, %xmm5 + paddq %xmm0, %xmm1 + paddq %xmm4, %xmm5 + +LIMB32(`pxor zero, zero ') + + paddq %xmm5, %xmm1 + movdqa %xmm1, %xmm0 + psrlq $4, %xmm1 + pand mm00001111, %xmm0 + pand mm00001111, %xmm1 + paddq %xmm0, %xmm1 + + psadbw zero, %xmm1 + paddq %xmm1, %xmm3 C add to grand total + + +C Add the two 64-bit halves of the grand total counter +L(rt): movdqa %xmm3, %xmm0 + psrldq $8, %xmm3 + paddq %xmm3, %xmm0 + movd %xmm0, areg C movq avoided due to gas bug + +LIMB32(`pop %ebx ') + ret + +EPILOGUE() +DEF_OBJECT(dummy,16) +C Three magic constants used for masking out bits + .byte 0x55,0x55,0x55,0x55,0x55,0x55,0x55,0x55 + .byte 0x55,0x55,0x55,0x55,0x55,0x55,0x55,0x55 + + .byte 0x33,0x33,0x33,0x33,0x33,0x33,0x33,0x33 + .byte 0x33,0x33,0x33,0x33,0x33,0x33,0x33,0x33 + + .byte 0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f + .byte 0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f,0x0f +cnsts: +C Masks for high end of number + .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff + .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff + + .byte 0xff,0xff,0xff,0xff,0x00,0x00,0x00,0x00 + .byte 0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00 + + .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff + .byte 0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00 + + .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff + .byte 0xff,0xff,0xff,0xff,0x00,0x00,0x00,0x00 +C Masks for low end of number + .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff + .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff + + .byte 0x00,0x00,0x00,0x00,0xff,0xff,0xff,0xff + .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff + + .byte 0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00 + .byte 0xff,0xff,0xff,0xff,0xff,0xff,0xff,0xff + + .byte 0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00 + .byte 0x00,0x00,0x00,0x00,0xff,0xff,0xff,0xff +END_OBJECT(dummy) +ASM_END() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/rsh1add_n.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/rsh1add_n.asm new file mode 100644 index 0000000..f421d13 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/rsh1add_n.asm @@ -0,0 +1,126 @@ +dnl Intel Pentium-4 mpn_rsh1add_n -- mpn (x+y)/2 + +dnl Copyright 2001-2004 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb (approx) +C dst!=src1,2 dst==src1 dst==src2 +C P4: 4.5 6.5 6.5 + + +C mp_limb_t mpn_rsh1add_n (mp_ptr wp, mp_srcptr xp, mp_srcptr yp, +C mp_size_t size); +C +C The slightly strange combination of indexing and pointer incrementing +C that's used seems to work best. Not sure why, but for instance leal +C incrementing on %esi is a 1 or 2 cycle slowdown. +C +C The dependent chain is paddq combining the carry and next (shifted) part, +C plus psrlq to move the new carry down. That, and just 4 mmx instructions +C in total, makes 4 c/l the target speed, which is almost achieved for +C separate src/dst but when src==dst the write combining anomalies slow it +C down. + +defframe(PARAM_SIZE, 16) +defframe(PARAM_YP, 12) +defframe(PARAM_XP, 8) +defframe(PARAM_WP, 4) + +dnl re-use parameter space +define(SAVE_EBX,`PARAM_XP') +define(SAVE_ESI,`PARAM_YP') + + TEXT + ALIGN(8) + +PROLOGUE(mpn_rsh1add_n) +deflit(`FRAME',0) + + movl PARAM_XP, %edx + movl %ebx, SAVE_EBX + + movl PARAM_YP, %ebx + movl %esi, SAVE_ESI + + movl PARAM_WP, %esi + + movd (%edx), %mm0 C xp[0] + + movd (%ebx), %mm1 C yp[0] + movl PARAM_SIZE, %ecx + + movl (%edx), %eax C xp[0] + + addl (%ebx), %eax C xp[0]+yp[0] + + paddq %mm1, %mm0 C xp[0]+yp[0] + leal (%esi,%ecx,4), %esi C wp end + negl %ecx C -size + + psrlq $1, %mm0 C (xp[0]+yp[0])/2 + and $1, %eax C return value, rsh1 bit of xp[0]+yp[0] + addl $1, %ecx C -(size-1) + jz L(done) + + +L(top): + C eax return value + C ebx yp end + C ecx counter, limbs, -(size-1) to -1 inclusive + C edx xp end + C esi wp end + C mm0 carry (32 bits) + + movd 4(%edx), %mm1 C xp[i+1] + movd 4(%ebx), %mm2 C yp[i+1] + leal 4(%edx), %edx + leal 4(%ebx), %ebx + paddq %mm2, %mm1 C xp[i+1]+yp[i+1] + psllq $31, %mm1 C low bit at 31, further 32 above + + paddq %mm1, %mm0 C 31 and carry from prev add + movd %mm0, -4(%esi,%ecx,4) C low ready to store dst[i] + + psrlq $32, %mm0 C high becomes new carry + + addl $1, %ecx + jnz L(top) + + +L(done): + movd %mm0, -4(%esi) C dst[size-1] + movl SAVE_EBX, %ebx + + movl SAVE_ESI, %esi + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/sqr_basecase.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/sqr_basecase.asm new file mode 100644 index 0000000..0d548e0 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/sqr_basecase.asm @@ -0,0 +1,705 @@ +dnl mpn_sqr_basecase for Pentium 4 and P6 models with SSE2 (i.e., 9,D,E,F). + +dnl Copyright 2001, 2002, 2007 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + +C TODO: +C * Improve ad-hoc outer loop code and register handling. Some feed-in +C scheduling could improve things by several cycles per outer iteration. +C * In Lam3...Lam1 code for, keep accumulation operands in registers, without +C storing intermediates to rp. +C * We might want to keep 32 in a free mm register, since the register form is +C 3 bytes and the immediate form is 4 bytes. About 80 bytes to save. +C * Look into different loop alignment, we now expand the code about 50 bytes +C with possibly needless alignment. +C * Use OSP, should solve feed-in latency problems. +C * Address relative slowness for un<=3 for Pentium M. The old code is there +C considerably faster. (1:20/14, 2:34/32, 3:66/57) + +C INPUT PARAMETERS +C rp sp + 4 +C up sp + 8 +C un sp + 12 + + TEXT + ALIGN(16) +PROLOGUE(mpn_sqr_basecase) + mov 4(%esp), %edx C rp + mov 8(%esp), %eax C up + mov 12(%esp), %ecx C un + + cmp $2, %ecx + jc L(un1) + jz L(un2) + cmp $4, %ecx + jc L(un3) + jz L(un4) + jmp L(big) + +L(un1): mov (%eax), %eax + mov %edx, %ecx + mul %eax + mov %eax, (%ecx) + mov %edx, 4(%ecx) + ret +L(un2): movd (%eax), %mm0 C un=2 + movd (%eax), %mm2 C un=2 + movd 4(%eax), %mm1 C un=2 + pmuludq %mm0, %mm0 C 64b weight 0 un=2 + pmuludq %mm1, %mm2 C 64b weight 32 un=2 + pmuludq %mm1, %mm1 C 64b weight 64 un=2 + movd %mm0, (%edx) C un=2 + psrlq $32, %mm0 C 32b weight 32 un=2 + pcmpeqd %mm7, %mm7 C un=2 + psrlq $33, %mm7 C 0x000000007FFFFFFF un=2 + pand %mm2, %mm7 C 31b weight 32 un=2 + psrlq $31, %mm2 C 33b weight 65 un=2 + psllq $1, %mm7 C 31b weight 33 un=2 + paddq %mm7, %mm0 C un=2 + movd %mm0, 4(%edx) C un=2 + psrlq $32, %mm0 C un=2 + paddq %mm2, %mm1 C un=2 + paddq %mm0, %mm1 C un=2 + movd %mm1, 8(%edx) C un=2 + psrlq $32, %mm1 C un=2 + movd %mm1, 12(%edx) C un=2 + emms + ret +L(un3): movd (%eax), %mm7 C un=3 + movd 4(%eax), %mm6 C un=3 + pmuludq %mm7, %mm6 C un=3 + movd 8(%eax), %mm2 C un=3 + pmuludq %mm7, %mm2 C un=3 + movd %mm6, 4(%edx) C un=3 + psrlq $32, %mm6 C un=3 + paddq %mm2, %mm6 C un=3 + movd %mm6, 8(%edx) C un=3 + psrlq $32, %mm6 C un=3 + movd %mm6, 12(%edx) C un=3 + lea 4(%edx), %edx C un=3 + lea 4(%eax), %eax C un=3 + jmp L(am1) +L(un4): movd (%eax), %mm7 C un=4 + movd 4(%eax), %mm6 C un=4 + pmuludq %mm7, %mm6 C un=4 + movd 8(%eax), %mm0 C un=4 + pmuludq %mm7, %mm0 C un=4 + movd 12(%eax), %mm1 C un=4 + pmuludq %mm7, %mm1 C un=4 + movd %mm6, 4(%edx) C un=4 + psrlq $32, %mm6 C un=4 + paddq %mm0, %mm6 C un=4 + movd %mm6, 8(%edx) C un=4 + psrlq $32, %mm6 C un=4 + paddq %mm1, %mm6 C un=4 + movd %mm6, 12(%edx) C un=4 + psrlq $32, %mm6 C un=4 + movd %mm6, 16(%edx) C un=4 + lea 4(%edx), %edx C un=4 + lea 4(%eax), %eax C un=4 + jmp L(am2) + +L(big): push %esi + push %ebx + push %edi + pxor %mm6, %mm6 + movd (%eax), %mm7 C + lea 4(%eax), %esi C init up, up++ + lea 4(%eax), %eax C up2++ FIXME: should fix offsets + lea 4(%edx), %edi C init rp, rp++ + lea 4(%edx), %edx C rp2++ + lea -4(%ecx), %ebx C loop count + and $3, %ecx + jz L(3m) + cmp $2, %ecx + ja L(2m) + jb L(0m) + +L(1m): + movd (%eax), %mm4 C m 1 + lea (%ebx), %ecx C inner loop count m 1 + pmuludq %mm7, %mm4 C m 1 + movd 4(%eax), %mm3 C m 1 + pmuludq %mm7, %mm3 C m 1 + movd 8(%eax), %mm0 C m 1 + jmp L(m01) C m 1 + ALIGN(16) C m 1 +L(lpm1): + pmuludq %mm7, %mm4 C m 1 + paddq %mm0, %mm6 C m 1 + movd 4(%eax), %mm3 C m 1 + movd %mm6, -8(%edx) C m 1 + psrlq $32, %mm6 C m 1 + pmuludq %mm7, %mm3 C m 1 + paddq %mm1, %mm6 C m 1 + movd 8(%eax), %mm0 C m 1 + movd %mm6, -4(%edx) C m 1 + psrlq $32, %mm6 C m 1 +L(m01): pmuludq %mm7, %mm0 C m 1 + paddq %mm4, %mm6 C m 1 + movd 12(%eax), %mm1 C m 1 + movd %mm6, (%edx) C m 1 + psrlq $32, %mm6 C m 1 + pmuludq %mm7, %mm1 C m 1 + paddq %mm3, %mm6 C m 1 + movd 16(%eax), %mm4 C m 1 + movd %mm6, 4(%edx) C m 1 + psrlq $32, %mm6 C m 1 + lea 16(%eax), %eax C m 1 + lea 16(%edx), %edx C m 1 + sub $4, %ecx C m 1 + ja L(lpm1) C m 1 + pmuludq %mm7, %mm4 C m 1 + paddq %mm0, %mm6 C m 1 + movd %mm6, -8(%edx) C m 1 + psrlq $32, %mm6 C m 1 + paddq %mm1, %mm6 C m 1 + jmp L(0) + +L(2m): + movd (%eax), %mm1 C m 2 + lea (%ebx), %ecx C inner loop count m 2 + pmuludq %mm7, %mm1 C m 2 + movd 4(%eax), %mm4 C m 2 + pmuludq %mm7, %mm4 C m 2 + movd 8(%eax), %mm3 C m 2 + jmp L(m10) C m 2 + ALIGN(16) C m 2 +L(lpm2): + pmuludq %mm7, %mm4 C m 2 + paddq %mm0, %mm6 C m 2 + movd 8(%eax), %mm3 C m 2 + movd %mm6, -4(%edx) C m 2 + psrlq $32, %mm6 C m 2 +L(m10): pmuludq %mm7, %mm3 C m 2 + paddq %mm1, %mm6 C m 2 + movd 12(%eax), %mm0 C m 2 + movd %mm6, (%edx) C m 2 + psrlq $32, %mm6 C m 2 + pmuludq %mm7, %mm0 C m 2 + paddq %mm4, %mm6 C m 2 + movd 16(%eax), %mm1 C m 2 + movd %mm6, 4(%edx) C m 2 + psrlq $32, %mm6 C m 2 + pmuludq %mm7, %mm1 C m 2 + paddq %mm3, %mm6 C m 2 + movd 20(%eax), %mm4 C m 2 + movd %mm6, 8(%edx) C m 2 + psrlq $32, %mm6 C m 2 + lea 16(%eax), %eax C m 2 + lea 16(%edx), %edx C m 2 + sub $4, %ecx C m 2 + ja L(lpm2) C m 2 + pmuludq %mm7, %mm4 C m 2 + paddq %mm0, %mm6 C m 2 + movd %mm6, -4(%edx) C m 2 + psrlq $32, %mm6 C m 2 + paddq %mm1, %mm6 C m 2 + jmp L(1) + +L(3m): + movd (%eax), %mm0 C m 3 + lea (%ebx), %ecx C inner loop count m 3 + pmuludq %mm7, %mm0 C m 3 + movd 4(%eax), %mm1 C m 3 + pmuludq %mm7, %mm1 C m 3 + movd 8(%eax), %mm4 C m 3 + jmp L(lpm3) C m 3 + ALIGN(16) C m 3 +L(lpm3): + pmuludq %mm7, %mm4 C m 3 + paddq %mm0, %mm6 C m 3 + movd 12(%eax), %mm3 C m 3 + movd %mm6, (%edx) C m 3 + psrlq $32, %mm6 C m 3 + pmuludq %mm7, %mm3 C m 3 + paddq %mm1, %mm6 C m 3 + movd 16(%eax), %mm0 C m 3 + movd %mm6, 4(%edx) C m 3 + psrlq $32, %mm6 C m 3 + pmuludq %mm7, %mm0 C m 3 + paddq %mm4, %mm6 C m 3 + movd 20(%eax), %mm1 C m 3 + movd %mm6, 8(%edx) C m 3 + psrlq $32, %mm6 C m 3 + pmuludq %mm7, %mm1 C m 3 + paddq %mm3, %mm6 C m 3 + movd 24(%eax), %mm4 C m 3 + movd %mm6, 12(%edx) C m 3 + psrlq $32, %mm6 C m 3 + lea 16(%eax), %eax C m 3 + lea 16(%edx), %edx C m 3 + sub $4, %ecx C m 3 + ja L(lpm3) C m 3 + pmuludq %mm7, %mm4 C m 3 + paddq %mm0, %mm6 C m 3 + movd %mm6, (%edx) C m 3 + psrlq $32, %mm6 C m 3 + paddq %mm1, %mm6 C m 3 + jmp L(2) + +L(0m): + movd (%eax), %mm3 C m 0 + lea (%ebx), %ecx C inner loop count m 0 + pmuludq %mm7, %mm3 C m 0 + movd 4(%eax), %mm0 C m 0 + pmuludq %mm7, %mm0 C m 0 + movd 8(%eax), %mm1 C m 0 + jmp L(m00) C m 0 + ALIGN(16) C m 0 +L(lpm0): + pmuludq %mm7, %mm4 C m 0 + paddq %mm0, %mm6 C m 0 + movd (%eax), %mm3 C m 0 + movd %mm6, -12(%edx) C m 0 + psrlq $32, %mm6 C m 0 + pmuludq %mm7, %mm3 C m 0 + paddq %mm1, %mm6 C m 0 + movd 4(%eax), %mm0 C m 0 + movd %mm6, -8(%edx) C m 0 + psrlq $32, %mm6 C m 0 + pmuludq %mm7, %mm0 C m 0 + paddq %mm4, %mm6 C m 0 + movd 8(%eax), %mm1 C m 0 + movd %mm6, -4(%edx) C m 0 + psrlq $32, %mm6 C m 0 +L(m00): pmuludq %mm7, %mm1 C m 0 + paddq %mm3, %mm6 C m 0 + movd 12(%eax), %mm4 C m 0 + movd %mm6, (%edx) C m 0 + psrlq $32, %mm6 C m 0 + lea 16(%eax), %eax C m 0 + lea 16(%edx), %edx C m 0 + sub $4, %ecx C m 0 + ja L(lpm0) C m 0 + pmuludq %mm7, %mm4 C m 0 + paddq %mm0, %mm6 C m 0 + movd %mm6, -12(%edx) C m 0 + psrlq $32, %mm6 C m 0 + paddq %mm1, %mm6 C m 0 + jmp L(3) + +L(outer): + lea 8(%edi), %edi C rp += 2 + movd (%esi), %mm7 C am 3 + mov %edi, %edx C rp2 = rp am 3 + lea 4(%esi), %esi C up++ am 3 + lea (%esi), %eax C up2 = up am 3 + movd (%eax), %mm0 C am 3 + lea (%ebx), %ecx C inner loop count am 3 + pxor %mm6, %mm6 C am 3 + pmuludq %mm7, %mm0 C am 3 + movd 4(%eax), %mm1 C am 3 + movd (%edx), %mm4 C am 3 + pmuludq %mm7, %mm1 C am 3 + movd 8(%eax), %mm2 C am 3 + paddq %mm0, %mm4 C am 3 + movd 4(%edx), %mm5 C am 3 + jmp L(lam3) C am 3 + ALIGN(16) C am 3 +L(lam3): + pmuludq %mm7, %mm2 C am 3 + paddq %mm4, %mm6 C am 3 + movd 12(%eax), %mm3 C am 3 + paddq %mm1, %mm5 C am 3 + movd 8(%edx), %mm4 C am 3 + movd %mm6, (%edx) C am 3 + psrlq $32, %mm6 C am 3 + pmuludq %mm7, %mm3 C am 3 + paddq %mm5, %mm6 C am 3 + movd 16(%eax), %mm0 C am 3 + paddq %mm2, %mm4 C am 3 + movd 12(%edx), %mm5 C am 3 + movd %mm6, 4(%edx) C am 3 + psrlq $32, %mm6 C am 3 + pmuludq %mm7, %mm0 C am 3 + paddq %mm4, %mm6 C am 3 + movd 20(%eax), %mm1 C am 3 + paddq %mm3, %mm5 C am 3 + movd 16(%edx), %mm4 C am 3 + movd %mm6, 8(%edx) C am 3 + psrlq $32, %mm6 C am 3 + pmuludq %mm7, %mm1 C am 3 + paddq %mm5, %mm6 C am 3 + movd 24(%eax), %mm2 C am 3 + paddq %mm0, %mm4 C am 3 + movd 20(%edx), %mm5 C am 3 + movd %mm6, 12(%edx) C am 3 + psrlq $32, %mm6 C am 3 + lea 16(%eax), %eax C am 3 + lea 16(%edx), %edx C am 3 + sub $4, %ecx C am 3 + ja L(lam3) C am 3 + pmuludq %mm7, %mm2 C am 3 + paddq %mm4, %mm6 C am 3 + paddq %mm1, %mm5 C am 3 + movd 8(%edx), %mm4 C am 3 + movd %mm6, (%edx) C am 3 + psrlq $32, %mm6 C am 3 + paddq %mm5, %mm6 C am 3 + paddq %mm2, %mm4 C am 3 +L(2): movd %mm6, 4(%edx) C am 3 + psrlq $32, %mm6 C am 3 + paddq %mm4, %mm6 C am 3 + movd %mm6, 8(%edx) C am 3 + psrlq $32, %mm6 C am 3 + movd %mm6, 12(%edx) C am 3 + + lea 8(%edi), %edi C rp += 2 + movd (%esi), %mm7 C am 2 + mov %edi, %edx C rp2 = rp am 2 + lea 4(%esi), %esi C up++ am 2 + lea (%esi), %eax C up2 = up am 2 + movd (%eax), %mm1 C am 2 + lea (%ebx), %ecx C inner loop count am 2 + pxor %mm6, %mm6 C am 2 + pmuludq %mm7, %mm1 C am 2 + movd 4(%eax), %mm2 C am 2 + movd (%edx), %mm5 C am 2 + pmuludq %mm7, %mm2 C am 2 + movd 8(%eax), %mm3 C am 2 + paddq %mm1, %mm5 C am 2 + movd 4(%edx), %mm4 C am 2 + jmp L(am10) C am 2 + ALIGN(16) C am 2 +L(lam2): + pmuludq %mm7, %mm2 C am 2 + paddq %mm4, %mm6 C am 2 + movd 8(%eax), %mm3 C am 2 + paddq %mm1, %mm5 C am 2 + movd 4(%edx), %mm4 C am 2 + movd %mm6, -4(%edx) C am 2 + psrlq $32, %mm6 C am 2 +L(am10): + pmuludq %mm7, %mm3 C am 2 + paddq %mm5, %mm6 C am 2 + movd 12(%eax), %mm0 C am 2 + paddq %mm2, %mm4 C am 2 + movd 8(%edx), %mm5 C am 2 + movd %mm6, (%edx) C am 2 + psrlq $32, %mm6 C am 2 + pmuludq %mm7, %mm0 C am 2 + paddq %mm4, %mm6 C am 2 + movd 16(%eax), %mm1 C am 2 + paddq %mm3, %mm5 C am 2 + movd 12(%edx), %mm4 C am 2 + movd %mm6, 4(%edx) C am 2 + psrlq $32, %mm6 C am 2 + pmuludq %mm7, %mm1 C am 2 + paddq %mm5, %mm6 C am 2 + movd 20(%eax), %mm2 C am 2 + paddq %mm0, %mm4 C am 2 + movd 16(%edx), %mm5 C am 2 + movd %mm6, 8(%edx) C am 2 + psrlq $32, %mm6 C am 2 + lea 16(%eax), %eax C am 2 + lea 16(%edx), %edx C am 2 + sub $4, %ecx C am 2 + ja L(lam2) C am 2 + pmuludq %mm7, %mm2 C am 2 + paddq %mm4, %mm6 C am 2 + paddq %mm1, %mm5 C am 2 + movd 4(%edx), %mm4 C am 2 + movd %mm6, -4(%edx) C am 2 + psrlq $32, %mm6 C am 2 + paddq %mm5, %mm6 C am 2 + paddq %mm2, %mm4 C am 2 +L(1): movd %mm6, (%edx) C am 2 + psrlq $32, %mm6 C am 2 + paddq %mm4, %mm6 C am 2 + movd %mm6, 4(%edx) C am 2 + psrlq $32, %mm6 C am 2 + movd %mm6, 8(%edx) C am 2 + + lea 8(%edi), %edi C rp += 2 + movd (%esi), %mm7 C am 1 + mov %edi, %edx C rp2 = rp am 1 + lea 4(%esi), %esi C up++ am 1 + lea (%esi), %eax C up2 = up am 1 + movd (%eax), %mm2 C am 1 + lea (%ebx), %ecx C inner loop count am 1 + pxor %mm6, %mm6 C am 1 + pmuludq %mm7, %mm2 C am 1 + movd 4(%eax), %mm3 C am 1 + movd (%edx), %mm4 C am 1 + pmuludq %mm7, %mm3 C am 1 + movd 8(%eax), %mm0 C am 1 + paddq %mm2, %mm4 C am 1 + movd 4(%edx), %mm5 C am 1 + jmp L(am01) C am 1 + ALIGN(16) C am 1 +L(lam1): + pmuludq %mm7, %mm2 C am 1 + paddq %mm4, %mm6 C am 1 + movd 4(%eax), %mm3 C am 1 + paddq %mm1, %mm5 C am 1 + movd (%edx), %mm4 C am 1 + movd %mm6, -8(%edx) C am 1 + psrlq $32, %mm6 C am 1 + pmuludq %mm7, %mm3 C am 1 + paddq %mm5, %mm6 C am 1 + movd 8(%eax), %mm0 C am 1 + paddq %mm2, %mm4 C am 1 + movd 4(%edx), %mm5 C am 1 + movd %mm6, -4(%edx) C am 1 + psrlq $32, %mm6 C am 1 +L(am01): + pmuludq %mm7, %mm0 C am 1 + paddq %mm4, %mm6 C am 1 + movd 12(%eax), %mm1 C am 1 + paddq %mm3, %mm5 C am 1 + movd 8(%edx), %mm4 C am 1 + movd %mm6, (%edx) C am 1 + psrlq $32, %mm6 C am 1 + pmuludq %mm7, %mm1 C am 1 + paddq %mm5, %mm6 C am 1 + movd 16(%eax), %mm2 C am 1 + paddq %mm0, %mm4 C am 1 + movd 12(%edx), %mm5 C am 1 + movd %mm6, 4(%edx) C am 1 + psrlq $32, %mm6 C am 1 + lea 16(%eax), %eax C am 1 + lea 16(%edx), %edx C am 1 + sub $4, %ecx C am 1 + ja L(lam1) C am 1 + pmuludq %mm7, %mm2 C am 1 + paddq %mm4, %mm6 C am 1 + paddq %mm1, %mm5 C am 1 + movd (%edx), %mm4 C am 1 + movd %mm6, -8(%edx) C am 1 + psrlq $32, %mm6 C am 1 + paddq %mm5, %mm6 C am 1 + paddq %mm2, %mm4 C am 1 +L(0): movd %mm6, -4(%edx) C am 1 + psrlq $32, %mm6 C am 1 + paddq %mm4, %mm6 C am 1 + movd %mm6, (%edx) C am 1 + psrlq $32, %mm6 C am 1 + movd %mm6, 4(%edx) C am 1 + + lea 8(%edi), %edi C rp += 2 + movd (%esi), %mm7 C am 0 + mov %edi, %edx C rp2 = rp am 0 + lea 4(%esi), %esi C up++ am 0 + lea (%esi), %eax C up2 = up am 0 + movd (%eax), %mm3 C am 0 + lea (%ebx), %ecx C inner loop count am 0 + pxor %mm6, %mm6 C am 0 + pmuludq %mm7, %mm3 C am 0 + movd 4(%eax), %mm0 C am 0 + movd (%edx), %mm5 C am 0 + pmuludq %mm7, %mm0 C am 0 + movd 8(%eax), %mm1 C am 0 + paddq %mm3, %mm5 C am 0 + movd 4(%edx), %mm4 C am 0 + jmp L(am00) C am 0 + ALIGN(16) C am 0 +L(lam0): + pmuludq %mm7, %mm2 C am 0 + paddq %mm4, %mm6 C am 0 + movd (%eax), %mm3 C am 0 + paddq %mm1, %mm5 C am 0 + movd -4(%edx), %mm4 C am 0 + movd %mm6, -12(%edx) C am 0 + psrlq $32, %mm6 C am 0 + pmuludq %mm7, %mm3 C am 0 + paddq %mm5, %mm6 C am 0 + movd 4(%eax), %mm0 C am 0 + paddq %mm2, %mm4 C am 0 + movd (%edx), %mm5 C am 0 + movd %mm6, -8(%edx) C am 0 + psrlq $32, %mm6 C am 0 + pmuludq %mm7, %mm0 C am 0 + paddq %mm4, %mm6 C am 0 + movd 8(%eax), %mm1 C am 0 + paddq %mm3, %mm5 C am 0 + movd 4(%edx), %mm4 C am 0 + movd %mm6, -4(%edx) C am 0 + psrlq $32, %mm6 C am 0 +L(am00): + pmuludq %mm7, %mm1 C am 0 + paddq %mm5, %mm6 C am 0 + movd 12(%eax), %mm2 C am 0 + paddq %mm0, %mm4 C am 0 + movd 8(%edx), %mm5 C am 0 + movd %mm6, (%edx) C am 0 + psrlq $32, %mm6 C am 0 + lea 16(%eax), %eax C am 0 + lea 16(%edx), %edx C am 0 + sub $4, %ecx C am 0 + ja L(lam0) C am 0 + pmuludq %mm7, %mm2 C am 0 + paddq %mm4, %mm6 C am 0 + paddq %mm1, %mm5 C am 0 + movd -4(%edx), %mm4 C am 0 + movd %mm6, -12(%edx) C am 0 + psrlq $32, %mm6 C am 0 + paddq %mm5, %mm6 C am 0 + paddq %mm2, %mm4 C am 0 +L(3): movd %mm6, -8(%edx) C am 0 + psrlq $32, %mm6 C am 0 + paddq %mm4, %mm6 C am 0 + movd %mm6, -4(%edx) C am 0 + psrlq $32, %mm6 C am 0 + movd %mm6, (%edx) C am 0 + sub $4, %ebx C am 0 + ja L(outer) C am 0 + + mov %edi, %edx + mov %esi, %eax + pop %edi + pop %ebx + pop %esi + +L(am3): C up[un-1..un-3] x up[un-4] + lea 8(%edx), %edx C rp2 += 2 + movd (%eax), %mm7 + movd 4(%eax), %mm1 + movd 8(%eax), %mm2 + movd 12(%eax), %mm3 + movd (%edx), %mm4 + pmuludq %mm7, %mm1 + movd 4(%edx), %mm5 + pmuludq %mm7, %mm2 + movd 8(%edx), %mm6 + pmuludq %mm7, %mm3 + paddq %mm1, %mm4 + paddq %mm2, %mm5 + paddq %mm3, %mm6 + movd %mm4, (%edx) + psrlq $32, %mm4 + paddq %mm5, %mm4 + movd %mm4, 4(%edx) + psrlq $32, %mm4 + paddq %mm6, %mm4 + movd %mm4, 8(%edx) + psrlq $32, %mm4 + movd %mm4, 12(%edx) C FIXME feed through! + lea 4(%eax), %eax + +L(am2): C up[un-1..un-2] x up[un-3] + lea 8(%edx), %edx C rp2 += 2 + movd (%eax), %mm7 + movd 4(%eax), %mm1 + movd 8(%eax), %mm2 + movd (%edx), %mm4 + movd 4(%edx), %mm5 + pmuludq %mm7, %mm1 + pmuludq %mm7, %mm2 + paddq %mm1, %mm4 + paddq %mm2, %mm5 + movd %mm4, (%edx) + psrlq $32, %mm4 + paddq %mm5, %mm4 + movd %mm4, 4(%edx) + psrlq $32, %mm4 + movd %mm4, 8(%edx) C FIXME feed through! + lea 4(%eax), %eax + +L(am1): C up[un-1] x up[un-2] + lea 8(%edx), %edx C rp2 += 2 + movd (%eax), %mm7 + movd 4(%eax), %mm2 + movd (%edx), %mm4 + pmuludq %mm7, %mm2 + paddq %mm2, %mm4 + movd %mm4, (%edx) + psrlq $32, %mm4 + movd %mm4, 4(%edx) + +C *** diag stuff, use elementary code for now + + mov 4(%esp), %edx C rp + mov 8(%esp), %eax C up + mov 12(%esp), %ecx C un + + movd (%eax), %mm2 + pmuludq %mm2, %mm2 C src[0]^2 + + pcmpeqd %mm7, %mm7 + psrlq $32, %mm7 + + movd 4(%edx), %mm3 C dst[1] + + movd %mm2, (%edx) + psrlq $32, %mm2 + + psllq $1, %mm3 C 2*dst[1] + paddq %mm3, %mm2 + movd %mm2, 4(%edx) + psrlq $32, %mm2 + + sub $2, %ecx + +L(diag): + movd 4(%eax), %mm0 C src limb + add $4, %eax + pmuludq %mm0, %mm0 + movq %mm7, %mm1 + pand %mm0, %mm1 C diagonal low + psrlq $32, %mm0 C diagonal high + + movd 8(%edx), %mm3 + psllq $1, %mm3 C 2*dst[i] + paddq %mm3, %mm1 + paddq %mm1, %mm2 + movd %mm2, 8(%edx) + psrlq $32, %mm2 + + movd 12(%edx), %mm3 + psllq $1, %mm3 C 2*dst[i+1] + paddq %mm3, %mm0 + paddq %mm0, %mm2 + movd %mm2, 12(%edx) + add $8, %edx + psrlq $32, %mm2 + + sub $1, %ecx + jnz L(diag) + + movd 4(%eax), %mm0 C src[size-1] + pmuludq %mm0, %mm0 + pand %mm0, %mm7 C diagonal low + psrlq $32, %mm0 C diagonal high + + movd 8(%edx), %mm3 C dst[2*size-2] + psllq $1, %mm3 + paddq %mm3, %mm7 + paddq %mm7, %mm2 + movd %mm2, 8(%edx) + psrlq $32, %mm2 + + paddq %mm0, %mm2 + movd %mm2, 12(%edx) C dst[2*size-1] + + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/sub_n.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/sub_n.asm new file mode 100644 index 0000000..5ba1c01 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/sub_n.asm @@ -0,0 +1,119 @@ +dnl Intel Pentium-4 mpn_sub_n -- mpn subtraction. + +dnl Copyright 2001, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C dst!=src1,2 dst==src1 dst==src2 +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) ? +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) 4 6 6 +C P4 model 3-4 (Prescott) 4.25 7.5 7.5 + +defframe(PARAM_CARRY,20) +defframe(PARAM_SIZE, 16) +defframe(PARAM_SRC2, 12) +defframe(PARAM_SRC1, 8) +defframe(PARAM_DST, 4) + +dnl re-use parameter space +define(SAVE_EBX,`PARAM_SRC1') + + TEXT + ALIGN(8) + +PROLOGUE(mpn_sub_nc) +deflit(`FRAME',0) + movd PARAM_CARRY, %mm0 + jmp L(start_nc) +EPILOGUE() + + ALIGN(8) +PROLOGUE(mpn_sub_n) +deflit(`FRAME',0) + pxor %mm0, %mm0 +L(start_nc): + mov PARAM_SRC1, %eax + mov %ebx, SAVE_EBX + mov PARAM_SRC2, %ebx + mov PARAM_DST, %edx + mov PARAM_SIZE, %ecx + + lea (%eax,%ecx,4), %eax C src1 end + lea (%ebx,%ecx,4), %ebx C src2 end + lea (%edx,%ecx,4), %edx C dst end + neg %ecx C -size + +L(top): + C eax src1 end + C ebx src2 end + C ecx counter, limbs, negative + C edx dst end + C mm0 carry bit + + movd (%eax,%ecx,4), %mm1 + movd (%ebx,%ecx,4), %mm2 + psubq %mm2, %mm1 + + psubq %mm0, %mm1 + movd %mm1, (%edx,%ecx,4) + + psrlq $63, %mm1 + + add $1, %ecx + jz L(done_mm1) + + movd (%eax,%ecx,4), %mm0 + movd (%ebx,%ecx,4), %mm2 + psubq %mm2, %mm0 + + psubq %mm1, %mm0 + movd %mm0, (%edx,%ecx,4) + + psrlq $63, %mm0 + + add $1, %ecx + jnz L(top) + + movd %mm0, %eax + mov SAVE_EBX, %ebx + emms + ret + +L(done_mm1): + movd %mm1, %eax + mov SAVE_EBX, %ebx + emms + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/pentium4/sse2/submul_1.asm b/gmp-6.3.0/mpn/x86/pentium4/sse2/submul_1.asm new file mode 100644 index 0000000..020675b --- /dev/null +++ b/gmp-6.3.0/mpn/x86/pentium4/sse2/submul_1.asm @@ -0,0 +1,182 @@ +dnl Intel Pentium-4 mpn_submul_1 -- Multiply a limb vector with a limb and +dnl subtract the result from a second limb vector. + +dnl Copyright 2001, 2002, 2008, 2010 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P6 model 0-8,10-12 - +C P6 model 9 (Banias) 6.8 +C P6 model 13 (Dothan) 6.9 +C P4 model 0-1 (Willamette) ? +C P4 model 2 (Northwood) 5.87 +C P4 model 3-4 (Prescott) 6.5 + +C This code represents a step forwards compared to the code available before +C GMP 5.1, but it is not carefully tuned for either P6 or P4. In fact, it is +C not good for P6. For P4 it saved a bit over 1 c/l for both Northwood and +C Prescott compared to the old code. +C +C The arrangements made here to get a two instruction dependent chain are +C slightly subtle. In the loop the carry (or borrow rather) is a negative so +C that a paddq can be used to give a low limb ready to store, and a high limb +C ready to become the new carry after a psrlq. +C +C If the carry was a simple twos complement negative then the psrlq shift would +C need to bring in 0 bits or 1 bits according to whether the high was zero or +C non-zero, since a non-zero value would represent a negative needing sign +C extension. That wouldn't be particularly easy to arrange and certainly would +C add an instruction to the dependent chain, so instead an offset is applied so +C that the high limb will be 0xFFFFFFFF+c. With c in the range -0xFFFFFFFF to +C 0, the value 0xFFFFFFFF+c is in the range 0 to 0xFFFFFFFF and is therefore +C always positive and can always have 0 bits shifted in, which is what psrlq +C does. +C +C The extra 0xFFFFFFFF must be subtracted before c is used, but that can be +C done off the dependent chain. The total adjustment then is to add +C 0xFFFFFFFF00000000 to offset the new carry, and subtract 0x00000000FFFFFFFF +C to remove the offset from the current carry, for a net add of +C 0xFFFFFFFE00000001. In the code this is applied to the destination limb when +C fetched. +C +C It's also possible to view the 0xFFFFFFFF adjustment as a ones-complement +C negative, which is how it's undone for the return value, but that doesn't +C seem as clear. + +defframe(PARAM_CARRY, 20) +defframe(PARAM_MULTIPLIER,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(16) + +PROLOGUE(mpn_submul_1c) +deflit(`FRAME',0) + movd PARAM_CARRY, %mm1 + jmp L(start_1c) +EPILOGUE() + +PROLOGUE(mpn_submul_1) +deflit(`FRAME',0) + pxor %mm1, %mm1 C initial borrow + +L(start_1c): + mov PARAM_SRC, %eax + pcmpeqd %mm0, %mm0 + + movd PARAM_MULTIPLIER, %mm7 + pcmpeqd %mm6, %mm6 + + mov PARAM_DST, %edx + psrlq $32, %mm0 C 0x00000000FFFFFFFF + + mov PARAM_SIZE, %ecx + psllq $32, %mm6 C 0xFFFFFFFF00000000 + + psubq %mm0, %mm6 C 0xFFFFFFFE00000001 + + psubq %mm1, %mm0 C 0xFFFFFFFF - borrow + + + movd (%eax), %mm3 C up + movd (%edx), %mm4 C rp + + add $-1, %ecx + paddq %mm6, %mm4 C add 0xFFFFFFFE00000001 + pmuludq %mm7, %mm3 + jnz L(gt1) + psubq %mm3, %mm4 C prod + paddq %mm4, %mm0 C borrow + movd %mm0, (%edx) C result + jmp L(rt) + +L(gt1): movd 4(%eax), %mm1 C up + movd 4(%edx), %mm2 C rp + + add $-1, %ecx + jz L(eev) + + ALIGN(16) +L(top): paddq %mm6, %mm2 C add 0xFFFFFFFE00000001 + pmuludq %mm7, %mm1 + psubq %mm3, %mm4 C prod + movd 8(%eax), %mm3 C up + paddq %mm4, %mm0 C borrow + movd 8(%edx), %mm4 C rp + movd %mm0, (%edx) C result + psrlq $32, %mm0 + + add $-1, %ecx + jz L(eod) + + paddq %mm6, %mm4 C add 0xFFFFFFFE00000001 + pmuludq %mm7, %mm3 + psubq %mm1, %mm2 C prod + movd 12(%eax), %mm1 C up + paddq %mm2, %mm0 C borrow + movd 12(%edx), %mm2 C rp + movd %mm0, 4(%edx) C result + psrlq $32, %mm0 + + lea 8(%eax), %eax + lea 8(%edx), %edx + add $-1, %ecx + jnz L(top) + + +L(eev): paddq %mm6, %mm2 C add 0xFFFFFFFE00000001 + pmuludq %mm7, %mm1 + psubq %mm3, %mm4 C prod + paddq %mm4, %mm0 C borrow + movd %mm0, (%edx) C result + psrlq $32, %mm0 + psubq %mm1, %mm2 C prod + paddq %mm2, %mm0 C borrow + movd %mm0, 4(%edx) C result +L(rt): psrlq $32, %mm0 + movd %mm0, %eax + not %eax + emms + ret + +L(eod): paddq %mm6, %mm4 C add 0xFFFFFFFE00000001 + pmuludq %mm7, %mm3 + psubq %mm1, %mm2 C prod + paddq %mm2, %mm0 C borrow + movd %mm0, 4(%edx) C result + psrlq $32, %mm0 + psubq %mm3, %mm4 C prod + paddq %mm4, %mm0 C borrow + movd %mm0, 8(%edx) C result + jmp L(rt) +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/rshift.asm b/gmp-6.3.0/mpn/x86/rshift.asm new file mode 100644 index 0000000..a60dcaa --- /dev/null +++ b/gmp-6.3.0/mpn/x86/rshift.asm @@ -0,0 +1,108 @@ +dnl x86 mpn_rshift -- mpn right shift. + +dnl Copyright 1992, 1994, 1996, 1999-2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P54 7.5 +C P55 7.0 +C P6 2.5 +C K6 4.5 +C K7 5.0 +C P4 16.5 + + +C mp_limb_t mpn_rshift (mp_ptr dst, mp_srcptr src, mp_size_t size, +C unsigned shift); + +defframe(PARAM_SHIFT,16) +defframe(PARAM_SIZE, 12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_rshift) + + pushl %edi + pushl %esi + pushl %ebx +deflit(`FRAME',12) + + movl PARAM_DST,%edi + movl PARAM_SRC,%esi + movl PARAM_SIZE,%edx + movl PARAM_SHIFT,%ecx + + leal -4(%edi,%edx,4),%edi + leal (%esi,%edx,4),%esi + negl %edx + + movl (%esi,%edx,4),%ebx C read least significant limb + xorl %eax,%eax + shrdl( %cl, %ebx, %eax) C compute carry limb + incl %edx + jz L(end) + pushl %eax C push carry limb onto stack + testb $1,%dl + jnz L(1) C enter loop in the middle + movl %ebx,%eax + + ALIGN(8) +L(oop): movl (%esi,%edx,4),%ebx C load next higher limb + shrdl( %cl, %ebx, %eax) C compute result limb + movl %eax,(%edi,%edx,4) C store it + incl %edx +L(1): movl (%esi,%edx,4),%eax + shrdl( %cl, %eax, %ebx) + movl %ebx,(%edi,%edx,4) + incl %edx + jnz L(oop) + + shrl %cl,%eax C compute most significant limb + movl %eax,(%edi) C store it + + popl %eax C pop carry limb + + popl %ebx + popl %esi + popl %edi + ret + +L(end): shrl %cl,%ebx C compute most significant limb + movl %ebx,(%edi) C store it + + popl %ebx + popl %esi + popl %edi + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/sec_tabselect.asm b/gmp-6.3.0/mpn/x86/sec_tabselect.asm new file mode 100644 index 0000000..d9d9952 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/sec_tabselect.asm @@ -0,0 +1,106 @@ +dnl x86 mpn_sec_tabselect. + +dnl Copyright 2011, 2021 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C cycles/limb +C P5 ? +C P6 model 0-8,10-12 ? +C P6 model 9 (Banias) ? +C P6 model 13 (Dothan) ? +C P4 model 0 (Willamette) ? +C P4 model 1 (?) ? +C P4 model 2 (Northwood) 4.5 +C P4 model 3 (Prescott) ? +C P4 model 4 (Nocona) ? +C Intel Atom ? +C AMD K6 ? +C AMD K7 3.4 +C AMD K8 ? +C AMD K10 ? + +C NOTES +C * This has not been tuned for any specific processor. Its speed should not +C be too bad, though. +C * Using SSE2 could result in many-fold speedup. + +C mpn_sec_tabselect (mp_limb_t *rp, mp_limb_t *tp, mp_size_t n, mp_size_t nents, mp_size_t which) +define(`rp', `%edi') +define(`tp', `%esi') +define(`n', `%ebx') +define(`nents', `32(%esp)') +define(`which', `36(%esp)') + +define(`i', `%ebp') +define(`mask', `%ecx') + +ASM_START() + TEXT + ALIGN(16) +PROLOGUE(mpn_sec_tabselect) + push %edi + push %esi + push %ebx + push %ebp + mov 20(%esp), rp + mov 24(%esp), tp + mov 28(%esp), n + + lea (rp,n,4), rp + lea (tp,n,4), tp +L(outer): + subl $1, which + sbb mask, mask + + mov n, i + neg i + + ALIGN(16) +L(top): mov (tp,i,4), %eax + mov (rp,i,4), %edx + xor %edx, %eax + and mask, %eax + xor %edx, %eax + mov %eax, (rp,i,4) + inc i + js L(top) + +L(end): lea (tp,n,4), tp + decl nents + jne L(outer) + +L(outer_end): + pop %ebp + pop %ebx + pop %esi + pop %edi + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/silvermont/gmp-mparam.h b/gmp-6.3.0/mpn/x86/silvermont/gmp-mparam.h new file mode 100644 index 0000000..e9f1d8f --- /dev/null +++ b/gmp-6.3.0/mpn/x86/silvermont/gmp-mparam.h @@ -0,0 +1,222 @@ +/* Intel Silvermont/32 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 2400 MHz Intel Atom C2758 Silvermont/Rangeley */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-30, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 3 +#define MOD_1_UNNORM_THRESHOLD 5 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 9 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 5 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 11 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 16 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 64.62% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 32 + +#define DIV_1_VS_MUL_1_PERCENT 204 + +#define MUL_TOOM22_THRESHOLD 26 +#define MUL_TOOM33_THRESHOLD 105 +#define MUL_TOOM44_THRESHOLD 236 +#define MUL_TOOM6H_THRESHOLD 351 +#define MUL_TOOM8H_THRESHOLD 502 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 105 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 163 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 137 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 174 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 215 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 36 +#define SQR_TOOM3_THRESHOLD 138 +#define SQR_TOOM4_THRESHOLD 360 +#define SQR_TOOM6_THRESHOLD 494 +#define SQR_TOOM8_THRESHOLD 620 + +#define MULMID_TOOM42_THRESHOLD 58 + +#define MULMOD_BNM1_THRESHOLD 15 +#define SQRMOD_BNM1_THRESHOLD 19 + +#define MUL_FFT_MODF_THRESHOLD 460 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 460, 5}, { 21, 6}, { 11, 5}, { 23, 6}, \ + { 12, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 21, 7}, { 11, 6}, { 25, 7}, { 13, 6}, \ + { 27, 7}, { 15, 6}, { 31, 7}, { 17, 6}, \ + { 35, 7}, { 19, 6}, { 39, 7}, { 21, 8}, \ + { 11, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 51,10}, { 15, 9}, \ + { 31, 8}, { 67, 9}, { 39, 8}, { 79, 9}, \ + { 47, 8}, { 95,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95,11}, \ + { 63,10}, { 127, 9}, { 255,10}, { 143, 9}, \ + { 287,10}, { 159,11}, { 95,10}, { 191,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 511,10}, \ + { 271, 9}, { 543,10}, { 287, 9}, { 575,11}, \ + { 159,10}, { 319, 9}, { 639,10}, { 335, 9}, \ + { 671,10}, { 351, 9}, { 703,11}, { 191,10}, \ + { 383, 9}, { 767,10}, { 415, 9}, { 831,12}, \ + { 127,11}, { 255,10}, { 543, 9}, { 1087,11}, \ + { 287,10}, { 607, 9}, { 1215,11}, { 319,10}, \ + { 671,11}, { 351,10}, { 735,12}, { 191,11}, \ + { 383,10}, { 767,11}, { 415,10}, { 831,13}, \ + { 127,12}, { 255,11}, { 543,10}, { 1087,11}, \ + { 607,10}, { 1215,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,10}, { 1471,12}, { 383,11}, \ + { 863,10}, { 1727,12}, { 447,11}, { 959,13}, \ + { 255,12}, { 511,11}, { 1087,12}, { 575,11}, \ + { 1215,10}, { 2431,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1471,13}, { 383,12}, { 767,11}, \ + { 1535,12}, { 831,11}, { 1727,10}, { 3455,12}, \ + { 959,11}, { 1919,14}, { 255,13}, { 511,12}, \ + { 1215,11}, { 2431,13}, { 639,12}, { 1471,11}, \ + { 2943,13}, { 767,12}, { 1727,11}, { 3455,13}, \ + { 895,12}, { 1919,14}, { 511,13}, { 1023,12}, \ + { 2111,13}, { 1151,12}, { 2431,13}, { 1407,12}, \ + { 2943,14}, { 767,13}, { 1663,12}, { 3455,13}, \ + { 1919,15}, { 511,14}, { 1023,13}, { 2175,12}, \ + { 4479,13}, { 2431,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3967,12}, { 7935,15}, { 1023,14}, { 2047,13}, \ + { 4479,14}, { 2303,13}, { 4991,12}, { 9983,14}, \ + { 2815,13}, { 5887,15}, { 1535,14}, { 3839,13}, \ + { 7935,16} } +#define MUL_FFT_TABLE3_SIZE 177 +#define MUL_FFT_THRESHOLD 4544 + +#define SQR_FFT_MODF_THRESHOLD 400 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 400, 5}, { 21, 6}, { 11, 5}, { 23, 6}, \ + { 12, 5}, { 25, 6}, { 13, 5}, { 28, 6}, \ + { 21, 7}, { 11, 6}, { 25, 7}, { 13, 6}, \ + { 28, 7}, { 15, 6}, { 32, 7}, { 17, 6}, \ + { 35, 7}, { 19, 6}, { 39, 7}, { 21, 8}, \ + { 11, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 51,10}, { 15, 9}, \ + { 31, 8}, { 67, 9}, { 39, 8}, { 79, 9}, \ + { 55,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 127,10}, \ + { 79, 9}, { 159,10}, { 95,11}, { 63,10}, \ + { 127, 9}, { 255,10}, { 143, 9}, { 287, 8}, \ + { 575,10}, { 159,11}, { 95,12}, { 63,11}, \ + { 127,10}, { 255, 9}, { 511,10}, { 271, 9}, \ + { 543,10}, { 287, 9}, { 575,11}, { 159,10}, \ + { 319, 9}, { 639,10}, { 335, 9}, { 671,10}, \ + { 351, 9}, { 735,11}, { 191,10}, { 383, 9}, \ + { 799,10}, { 415, 9}, { 831,11}, { 223,12}, \ + { 127,11}, { 255,10}, { 543, 9}, { 1087,11}, \ + { 287,10}, { 607, 9}, { 1215,11}, { 319,10}, \ + { 671,11}, { 351,10}, { 735, 9}, { 1471,12}, \ + { 191,11}, { 383,10}, { 799,11}, { 415,10}, \ + { 863,13}, { 127,12}, { 255,11}, { 543,10}, \ + { 1087,11}, { 607,10}, { 1215,12}, { 319,11}, \ + { 671,10}, { 1343,11}, { 735,10}, { 1471,12}, \ + { 383,11}, { 863,10}, { 1727,12}, { 447,11}, \ + { 959,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,10}, { 2431,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1535,12}, { 831,11}, { 1727,12}, \ + { 959,14}, { 255,13}, { 511,12}, { 1215,11}, \ + { 2431,13}, { 639,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1727,11}, { 3455,13}, { 895,12}, \ + { 1919,14}, { 511,13}, { 1023,12}, { 2111,13}, \ + { 1151,12}, { 2431,13}, { 1407,12}, { 2943,14}, \ + { 767,13}, { 1663,12}, { 3455,13}, { 1919,15}, \ + { 511,14}, { 1023,13}, { 2175,12}, { 4479,13}, \ + { 2431,14}, { 1279,13}, { 2943,12}, { 5887,14}, \ + { 1535,13}, { 3455,14}, { 1791,13}, { 3967,15}, \ + { 1023,14}, { 2047,13}, { 4479,14}, { 2303,13}, \ + { 4991,12}, { 9983,14}, { 2815,13}, { 5887,15}, \ + { 1535,14}, { 3839,13}, { 7679,16} } +#define SQR_FFT_TABLE3_SIZE 175 +#define SQR_FFT_THRESHOLD 3712 + +#define MULLO_BASECASE_THRESHOLD 0 /* always */ +#define MULLO_DC_THRESHOLD 56 +#define MULLO_MUL_N_THRESHOLD 8907 +#define SQRLO_BASECASE_THRESHOLD 0 /* always */ +#define SQRLO_DC_THRESHOLD 137 +#define SQRLO_SQR_THRESHOLD 7373 + +#define DC_DIV_QR_THRESHOLD 76 +#define DC_DIVAPPR_Q_THRESHOLD 336 +#define DC_BDIV_QR_THRESHOLD 66 +#define DC_BDIV_Q_THRESHOLD 218 + +#define INV_MULMOD_BNM1_THRESHOLD 50 +#define INV_NEWTON_THRESHOLD 345 +#define INV_APPR_THRESHOLD 342 + +#define BINV_NEWTON_THRESHOLD 366 +#define REDC_1_TO_REDC_N_THRESHOLD 91 + +#define MU_DIV_QR_THRESHOLD 1652 +#define MU_DIVAPPR_Q_THRESHOLD 1858 +#define MUPI_DIV_QR_THRESHOLD 171 +#define MU_BDIV_QR_THRESHOLD 1442 +#define MU_BDIV_Q_THRESHOLD 1830 + +#define POWM_SEC_TABLE 3,17,102,404,1185 + +#define GET_STR_DC_THRESHOLD 14 +#define GET_STR_PRECOMPUTE_THRESHOLD 21 +#define SET_STR_DC_THRESHOLD 272 +#define SET_STR_PRECOMPUTE_THRESHOLD 788 + +#define FAC_DSC_THRESHOLD 132 +#define FAC_ODD_THRESHOLD 34 + +#define MATRIX22_STRASSEN_THRESHOLD 19 +#define HGCD2_DIV1_METHOD 1 /* 0.59% faster than 3 */ +#define HGCD_THRESHOLD 142 +#define HGCD_APPR_THRESHOLD 181 +#define HGCD_REDUCE_THRESHOLD 2681 +#define GCD_DC_THRESHOLD 492 +#define GCDEXT_DC_THRESHOLD 365 +#define JACOBI_BASE_METHOD 1 /* 0.41% faster than 2 */ + +/* Tuneup completed successfully, took 147027 seconds */ diff --git a/gmp-6.3.0/mpn/x86/skylake/gmp-mparam.h b/gmp-6.3.0/mpn/x86/skylake/gmp-mparam.h new file mode 100644 index 0000000..fb87957 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/skylake/gmp-mparam.h @@ -0,0 +1,211 @@ +/* x86/skylake gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 3600-4000 MHz Intel Xeon E3-1270v5 Skylake */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-21, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 15 +#define MOD_1_UNNORM_THRESHOLD 16 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 10 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 8 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 0 /* never mpn_mod_1_1p */ +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 10 +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 9 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 5.63% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 12 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD 17 +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 18 + +#define DIV_1_VS_MUL_1_PERCENT 348 + +#define MUL_TOOM22_THRESHOLD 24 +#define MUL_TOOM33_THRESHOLD 81 +#define MUL_TOOM44_THRESHOLD 208 +#define MUL_TOOM6H_THRESHOLD 303 +#define MUL_TOOM8H_THRESHOLD 454 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 81 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 149 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 137 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 145 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 196 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 40 +#define SQR_TOOM3_THRESHOLD 129 +#define SQR_TOOM4_THRESHOLD 220 +#define SQR_TOOM6_THRESHOLD 354 +#define SQR_TOOM8_THRESHOLD 608 + +#define MULMID_TOOM42_THRESHOLD 72 + +#define MULMOD_BNM1_THRESHOLD 17 +#define SQRMOD_BNM1_THRESHOLD 21 + +#define MUL_FFT_MODF_THRESHOLD 530 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 530, 5}, { 29, 6}, { 15, 5}, { 31, 6}, \ + { 29, 7}, { 15, 6}, { 33, 7}, { 17, 6}, \ + { 36, 7}, { 19, 6}, { 39, 7}, { 21, 6}, \ + { 43, 7}, { 23, 6}, { 47, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 43, 8}, \ + { 23, 7}, { 49, 8}, { 27, 7}, { 55, 9}, \ + { 15, 8}, { 31, 7}, { 63, 8}, { 43, 9}, \ + { 23, 8}, { 51, 9}, { 31, 8}, { 67, 9}, \ + { 39, 8}, { 83, 9}, { 47, 8}, { 95, 9}, \ + { 55,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 135,10}, \ + { 79, 9}, { 159,10}, { 95, 9}, { 191,10}, \ + { 111,11}, { 63,10}, { 143, 9}, { 287,10}, \ + { 159,11}, { 95,10}, { 191,12}, { 63,11}, \ + { 127,10}, { 255, 9}, { 511,10}, { 271, 9}, \ + { 543,10}, { 287,11}, { 159,10}, { 351,11}, \ + { 191,10}, { 415,12}, { 127,11}, { 255,10}, \ + { 543,11}, { 287,10}, { 607,11}, { 319,10}, \ + { 671,11}, { 351,12}, { 191,11}, { 383,10}, \ + { 799,11}, { 415,13}, { 127,12}, { 255,11}, \ + { 543,10}, { 1087,11}, { 607,12}, { 319,11}, \ + { 671,10}, { 1343,11}, { 735,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,12}, { 447,11}, \ + { 959,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,12}, { 639,11}, { 1343,12}, \ + { 703,13}, { 383,12}, { 767,11}, { 1599,12}, \ + { 831,11}, { 1727,12}, { 959,14}, { 255,13}, \ + { 511,12}, { 1087,11}, { 2239,12}, { 1215,11}, \ + { 2431,13}, { 639,12}, { 1471,13}, { 767,12}, \ + { 1727,13}, { 895,12}, { 1919,14}, { 511,13}, \ + { 1023,12}, { 2239,13}, { 1151,12}, { 2431,13}, \ + { 1279,12}, { 2623,13}, { 1407,12}, { 2815,14}, \ + { 767,13}, { 1663,12}, { 3455,13}, { 1919,15}, \ + { 511,14}, { 1023,13}, { 2175,12}, { 4479,13}, \ + { 2431,14}, { 1279,13}, { 2943,12}, { 5887,14}, \ + { 1535,13}, { 3455,14}, { 1791,13}, { 3967,15}, \ + { 1023,14}, { 2047,13}, { 4479,14}, { 2303,13}, \ + { 4991,12}, { 9983,14}, { 2815,13}, { 5887,15}, \ + { 1535,14}, { 3839,16} } +#define MUL_FFT_TABLE3_SIZE 154 +#define MUL_FFT_THRESHOLD 6784 + +#define SQR_FFT_MODF_THRESHOLD 460 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 460, 5}, { 29, 6}, { 15, 5}, { 31, 6}, \ + { 29, 7}, { 15, 6}, { 33, 7}, { 17, 6}, \ + { 36, 7}, { 19, 6}, { 39, 7}, { 29, 8}, \ + { 15, 7}, { 35, 8}, { 19, 7}, { 41, 8}, \ + { 23, 7}, { 49, 8}, { 27, 7}, { 55, 9}, \ + { 15, 8}, { 31, 7}, { 63, 8}, { 43, 9}, \ + { 23, 8}, { 55,10}, { 15, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 79, 9}, { 47, 8}, \ + { 95, 9}, { 55,10}, { 31, 9}, { 79,10}, \ + { 47, 9}, { 95,11}, { 31,10}, { 63, 9}, \ + { 135,10}, { 79, 9}, { 159,10}, { 95,11}, \ + { 63,10}, { 127, 9}, { 255,10}, { 143, 9}, \ + { 287,10}, { 159,11}, { 95,12}, { 63,11}, \ + { 127,10}, { 271, 9}, { 543,10}, { 287,11}, \ + { 159,10}, { 319, 9}, { 639,10}, { 351,11}, \ + { 191,10}, { 415,12}, { 127,11}, { 255,10}, \ + { 543,11}, { 287,10}, { 575,11}, { 319,10}, \ + { 671,11}, { 351,10}, { 703,12}, { 191,11}, \ + { 383,10}, { 799,11}, { 415,10}, { 831,13}, \ + { 127,12}, { 255,11}, { 511,10}, { 1023,11}, \ + { 543,10}, { 1087,11}, { 607,12}, { 319,11}, \ + { 671,10}, { 1343,11}, { 735,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,12}, { 447,11}, \ + { 927,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1407,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1727,12}, { 895,11}, \ + { 1791,14}, { 255,13}, { 511,12}, { 1087,11}, \ + { 2239,12}, { 1215,13}, { 639,12}, { 1471,13}, \ + { 767,12}, { 1727,13}, { 895,12}, { 1919,14}, \ + { 511,13}, { 1023,12}, { 2239,13}, { 1151,12}, \ + { 2431,13}, { 1279,12}, { 2623,13}, { 1407,12}, \ + { 2815,14}, { 767,13}, { 1663,12}, { 3455,13}, \ + { 1919,15}, { 511,14}, { 1023,13}, { 2175,12}, \ + { 4479,13}, { 2431,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3967,15}, { 1023,14}, { 2047,13}, { 4479,14}, \ + { 2303,13}, { 4991,12}, { 9983,14}, { 2815,13}, \ + { 5887,15}, { 1535,14}, { 3839,16} } +#define SQR_FFT_TABLE3_SIZE 155 +#define SQR_FFT_THRESHOLD 5568 + +#define MULLO_BASECASE_THRESHOLD 0 /* always */ +#define MULLO_DC_THRESHOLD 68 +#define MULLO_MUL_N_THRESHOLD 13555 +#define SQRLO_BASECASE_THRESHOLD 0 /* always */ +#define SQRLO_DC_THRESHOLD 117 +#define SQRLO_SQR_THRESHOLD 10988 + +#define DC_DIV_QR_THRESHOLD 42 +#define DC_DIVAPPR_Q_THRESHOLD 163 +#define DC_BDIV_QR_THRESHOLD 66 +#define DC_BDIV_Q_THRESHOLD 160 + +#define INV_MULMOD_BNM1_THRESHOLD 46 +#define INV_NEWTON_THRESHOLD 165 +#define INV_APPR_THRESHOLD 157 + +#define BINV_NEWTON_THRESHOLD 300 +#define REDC_1_TO_REDC_N_THRESHOLD 68 + +#define MU_DIV_QR_THRESHOLD 1718 +#define MU_DIVAPPR_Q_THRESHOLD 1685 +#define MUPI_DIV_QR_THRESHOLD 62 +#define MU_BDIV_QR_THRESHOLD 1589 +#define MU_BDIV_Q_THRESHOLD 1830 + +#define POWM_SEC_TABLE 1,17,129,547,1317 + +#define GET_STR_DC_THRESHOLD 10 +#define GET_STR_PRECOMPUTE_THRESHOLD 16 +#define SET_STR_DC_THRESHOLD 354 +#define SET_STR_PRECOMPUTE_THRESHOLD 860 + +#define FAC_DSC_THRESHOLD 141 +#define FAC_ODD_THRESHOLD 34 + +#define MATRIX22_STRASSEN_THRESHOLD 20 +#define HGCD2_DIV1_METHOD 5 /* 1.04% faster than 3 */ +#define HGCD_THRESHOLD 114 +#define HGCD_APPR_THRESHOLD 132 +#define HGCD_REDUCE_THRESHOLD 3524 +#define GCD_DC_THRESHOLD 474 +#define GCDEXT_DC_THRESHOLD 379 +#define JACOBI_BASE_METHOD 1 /* 27.39% faster than 4 */ + +/* Tuneup completed successfully, took 31721 seconds */ diff --git a/gmp-6.3.0/mpn/x86/sqr_basecase.asm b/gmp-6.3.0/mpn/x86/sqr_basecase.asm new file mode 100644 index 0000000..39f8a89 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/sqr_basecase.asm @@ -0,0 +1,359 @@ +dnl x86 generic mpn_sqr_basecase -- square an mpn number. + +dnl Copyright 1999, 2000, 2002, 2003 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + + +include(`../config.m4') + + +C cycles/crossproduct cycles/triangleproduct +C P5 +C P6 +C K6 +C K7 +C P4 + + +C void mpn_sqr_basecase (mp_ptr dst, mp_srcptr src, mp_size_t size); +C +C The algorithm is basically the same as mpn/generic/sqr_basecase.c, but a +C lot of function call overheads are avoided, especially when the size is +C small. +C +C The mul1 loop is not unrolled like mul_1.asm, it doesn't seem worth the +C code size to do so here. +C +C Enhancements: +C +C The addmul loop here is also not unrolled like aorsmul_1.asm and +C mul_basecase.asm are. Perhaps it should be done. It'd add to the +C complexity, but if it's worth doing in the other places then it should be +C worthwhile here. +C +C A fully-unrolled style like other sqr_basecase.asm versions (k6, k7, p6) +C might be worth considering. That'd add quite a bit to the code size, but +C only as much as is used would be dragged into L1 cache. + +defframe(PARAM_SIZE,12) +defframe(PARAM_SRC, 8) +defframe(PARAM_DST, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_sqr_basecase) +deflit(`FRAME',0) + + movl PARAM_SIZE, %edx + + movl PARAM_SRC, %eax + + cmpl $2, %edx + movl PARAM_DST, %ecx + + je L(two_limbs) + ja L(three_or_more) + + +C ----------------------------------------------------------------------------- +C one limb only + C eax src + C ebx + C ecx dst + C edx + + movl (%eax), %eax + mull %eax + movl %eax, (%ecx) + movl %edx, 4(%ecx) + ret + + +C ----------------------------------------------------------------------------- + ALIGN(8) +L(two_limbs): + C eax src + C ebx + C ecx dst + C edx + + pushl %ebx + pushl %ebp + + movl %eax, %ebx + movl (%eax), %eax + + mull %eax C src[0]^2 + + pushl %esi + pushl %edi + + movl %edx, %esi C dst[1] + movl %eax, (%ecx) C dst[0] + + movl 4(%ebx), %eax + mull %eax C src[1]^2 + + movl %eax, %edi C dst[2] + movl %edx, %ebp C dst[3] + + movl (%ebx), %eax + mull 4(%ebx) C src[0]*src[1] + + addl %eax, %esi + + adcl %edx, %edi + + adcl $0, %ebp + addl %esi, %eax + + adcl %edi, %edx + movl %eax, 4(%ecx) + + adcl $0, %ebp + + movl %edx, 8(%ecx) + movl %ebp, 12(%ecx) + + popl %edi + popl %esi + + popl %ebp + popl %ebx + + ret + + +C ----------------------------------------------------------------------------- + ALIGN(8) +L(three_or_more): +deflit(`FRAME',0) + C eax src + C ebx + C ecx dst + C edx size + + pushl %ebx FRAME_pushl() + pushl %edi FRAME_pushl() + + pushl %esi FRAME_pushl() + pushl %ebp FRAME_pushl() + + leal (%ecx,%edx,4), %edi C &dst[size], end of this mul1 + leal (%eax,%edx,4), %esi C &src[size] + +C First multiply src[0]*src[1..size-1] and store at dst[1..size]. + + movl (%eax), %ebp C src[0], multiplier + movl %edx, %ecx + + negl %ecx C -size + xorl %ebx, %ebx C clear carry limb + + incl %ecx C -(size-1) + +L(mul1): + C eax scratch + C ebx carry + C ecx counter, limbs, negative + C edx scratch + C esi &src[size] + C edi &dst[size] + C ebp multiplier + + movl (%esi,%ecx,4), %eax + mull %ebp + addl %eax, %ebx + adcl $0, %edx + movl %ebx, (%edi,%ecx,4) + movl %edx, %ebx + incl %ecx + jnz L(mul1) + + movl %ebx, (%edi) + + + C Add products src[n]*src[n+1..size-1] at dst[2*n-1...], for + C n=1..size-2. + C + C The last products src[size-2]*src[size-1], which is the end corner + C of the product triangle, is handled separately at the end to save + C looping overhead. If size is 3 then it's only this that needs to + C be done. + C + C In the outer loop %esi is a constant, and %edi just advances by 1 + C limb each time. The size of the operation decreases by 1 limb + C each time. + + C eax + C ebx carry (needing carry flag added) + C ecx + C edx + C esi &src[size] + C edi &dst[size] + C ebp + + movl PARAM_SIZE, %ecx + subl $3, %ecx + jz L(corner) + + negl %ecx + +dnl re-use parameter space +define(VAR_OUTER,`PARAM_DST') + +L(outer): + C eax + C ebx + C ecx + C edx outer loop counter, -(size-3) to -1 + C esi &src[size] + C edi dst, pointing at stored carry limb of previous loop + C ebp + + movl %ecx, VAR_OUTER + addl $4, %edi C advance dst end + + movl -8(%esi,%ecx,4), %ebp C next multiplier + subl $1, %ecx + + xorl %ebx, %ebx C initial carry limb + +L(inner): + C eax scratch + C ebx carry (needing carry flag added) + C ecx counter, -n-1 to -1 + C edx scratch + C esi &src[size] + C edi dst end of this addmul + C ebp multiplier + + movl (%esi,%ecx,4), %eax + mull %ebp + addl %ebx, %eax + adcl $0, %edx + addl %eax, (%edi,%ecx,4) + adcl $0, %edx + movl %edx, %ebx + addl $1, %ecx + jl L(inner) + + + movl %ebx, (%edi) + movl VAR_OUTER, %ecx + incl %ecx + jnz L(outer) + + +L(corner): + C esi &src[size] + C edi &dst[2*size-3] + + movl -4(%esi), %eax + mull -8(%esi) C src[size-1]*src[size-2] + addl %eax, 0(%edi) + adcl $0, %edx + movl %edx, 4(%edi) C dst high limb + + +C ----------------------------------------------------------------------------- +C Left shift of dst[1..2*size-2], high bit shifted out becomes dst[2*size-1]. + + movl PARAM_SIZE, %eax + negl %eax + addl $1, %eax C -(size-1) and clear carry + +L(lshift): + C eax counter, negative + C ebx next limb + C ecx + C edx + C esi + C edi &dst[2*size-4] + C ebp + + rcll 8(%edi,%eax,8) + rcll 12(%edi,%eax,8) + incl %eax + jnz L(lshift) + + + adcl %eax, %eax C high bit out + movl %eax, 8(%edi) C dst most significant limb + + +C Now add in the squares on the diagonal, namely src[0]^2, src[1]^2, ..., +C src[size-1]^2. dst[0] hasn't yet been set at all yet, and just gets the +C low limb of src[0]^2. + + movl PARAM_SRC, %esi + movl (%esi), %eax C src[0] + mull %eax C src[0]^2 + + movl PARAM_SIZE, %ecx + leal (%esi,%ecx,4), %esi C src end + + negl %ecx C -size + movl %edx, %ebx C initial carry + + movl %eax, 12(%edi,%ecx,8) C dst[0] + incl %ecx C -(size-1) + +L(diag): + C eax scratch (low product) + C ebx carry limb + C ecx counter, -(size-1) to -1 + C edx scratch (high product) + C esi &src[size] + C edi &dst[2*size-3] + C ebp scratch (fetched dst limbs) + + movl (%esi,%ecx,4), %eax + mull %eax + + addl %ebx, 8(%edi,%ecx,8) + movl %edx, %ebx + + adcl %eax, 12(%edi,%ecx,8) + adcl $0, %ebx + + incl %ecx + jnz L(diag) + + + addl %ebx, 8(%edi) C dst most significant limb + + popl %ebp + popl %esi + + popl %edi + popl %ebx + + ret + +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/t-zdisp.sh b/gmp-6.3.0/mpn/x86/t-zdisp.sh new file mode 100755 index 0000000..61efdd6 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/t-zdisp.sh @@ -0,0 +1,71 @@ +#! /bin/sh +# +# Copyright 2000 Free Software Foundation, Inc. +# +# This file is part of the GNU MP Library. +# +# The GNU MP Library is free software; you can redistribute it and/or modify +# it under the terms of either: +# +# * the GNU Lesser General Public License as published by the Free +# Software Foundation; either version 3 of the License, or (at your +# option) any later version. +# +# or +# +# * the GNU General Public License as published by the Free Software +# Foundation; either version 2 of the License, or (at your option) any +# later version. +# +# or both in parallel, as here. +# +# The GNU MP Library is distributed in the hope that it will be useful, but +# WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +# or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +# for more details. +# +# You should have received copies of the GNU General Public License and the +# GNU Lesser General Public License along with the GNU MP Library. If not, +# see https://www.gnu.org/licenses/. + + +# Usage: cd $(builddir)/mpn +# $(srcdir)/x86/t-zdisp.sh +# +# Run the Zdisp() macro instructions through the assembler to check +# the encodings used. Mismatches are printed, no output means all ok. +# +# This program is only meant for use during development. It can be +# run in the mpn build directory of any x86 configuration. +# +# For this test the assembler needs to generate byte sized 0 +# displacements when given something like 0(%eax). Recent versions of +# gas are suitable (eg. 2.9.x or 2.10.x). + +set -e + +cat >tmp-zdisptest.asm <<\EOF + +include(`../config.m4') + +dnl Redefine Zdisp_match to output its pattern and encoding. +define(`Zdisp_match', +`define(`Zdisp_found',1)dnl +ifelse(`$2',0,` $1 $2$3, $4')`'dnl +ifelse(`$3',0,` $1 $2, $3$4')`'dnl + + .byte $5 +') + .text + Zdisp() +EOF + +m4 tmp-zdisptest.asm >tmp-zdisptest.s +as -o tmp-zdisptest.o tmp-zdisptest.s + +# Demand duplicates from the instruction patterns and byte encodings. +objdump -d tmp-zdisptest.o | awk ' +/^ *[a-z0-9]+:/ { + sub(/^ *[a-z0-9]+:/,"") + print +}' | sort | uniq -u diff --git a/gmp-6.3.0/mpn/x86/t-zdisp2.pl b/gmp-6.3.0/mpn/x86/t-zdisp2.pl new file mode 100755 index 0000000..b441b65 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/t-zdisp2.pl @@ -0,0 +1,147 @@ +#!/usr/bin/perl -w +# +# Copyright 2001, 2002 Free Software Foundation, Inc. +# +# This file is part of the GNU MP Library. +# +# The GNU MP Library is free software; you can redistribute it and/or modify +# it under the terms of either: +# +# * the GNU Lesser General Public License as published by the Free +# Software Foundation; either version 3 of the License, or (at your +# option) any later version. +# +# or +# +# * the GNU General Public License as published by the Free Software +# Foundation; either version 2 of the License, or (at your option) any +# later version. +# +# or both in parallel, as here. +# +# The GNU MP Library is distributed in the hope that it will be useful, but +# WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +# or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +# for more details. +# +# You should have received copies of the GNU General Public License and the +# GNU Lesser General Public License along with the GNU MP Library. If not, +# see https://www.gnu.org/licenses/. + + +# Usage: cd $(builddir)/mpn +# $(srcdir)/x86/t-zdisp2.pl +# +# Grep for any "0(reg...)" addressing modes coming out of the x86 .asm +# files. Additive expressions like "12+4-16" are recognised too. +# +# Old gas doesn't preserve the "0" displacement, so if it's wanted then +# Zdisp ought to be used to give explicit .byte sequences. See +# mpn/x86/README. +# +# No output means everything is ok. All the asm files are put through m4 in +# PIC and non-PIC modes, and in each multi-function form, all of which can +# take a while to run. +# +# This program is only meant for use during development. + +use strict; +use File::Find; +use File::Basename; +use Getopt::Std; + +my %opt; +getopts('t', \%opt); + + +my $srcdir; +open IN, ') { + if (/^srcdir[ \t]*=[ \t]*(.*)/) { + $srcdir = $1; + last; + } +} +close IN or die; +defined $srcdir or die "Cannot find \$srcdir in Makefile\n"; + +my $filecount = 0; + +my $tempfile = 't-zdisp2.tmp'; +open KARA, ">$tempfile" or die; +close KARA or die; + +find({ wanted => \&process, preprocess => \&process_mparam, no_chdir => 1 }, + "$srcdir/x86"); + +sub process { + if (/gmp-mparam.h$/) { + process_mparam($_); + } elsif (/\.asm$/) { + process_asm($_); + } +} + +# Ensure we're using the right SQR_TOOM2_THRESHOLD for the part of the +# tree being processed. +sub process_mparam { + my $file = "$File::Find::dir/gmp-mparam.h"; + if (-f $file) { + print "$file\n" if $opt{'t'}; + open MPARAM, "<$file" or die; + while () { + if (/^#define SQR_TOOM2_THRESHOLD[ \t]*([0-9][0-9]*)/) { + open KARA, ">$tempfile" or die; + print KARA "define(\`SQR_TOOM2_THRESHOLD',$1)\n\n"; + print "define(\`SQR_TOOM2_THRESHOLD',$1)\n" if $opt{'t'}; + close KARA or die; + last; + } + } + close MPARAM or die; + } + return @_; +} + +sub process_asm { + my ($file) = @_; + my $base = basename ($file, '.asm'); + + my @funs; + if ($base eq 'aors_n') { @funs = qw(add_n sub_n); } + elsif ($base eq 'aorsmul_1') { @funs = qw(addmul_1 submul_1); } + elsif ($base eq 'popham') { @funs = qw(popcount hamdist); } + elsif ($base eq 'logops_n') { @funs = qw(and_n andn_n nand_n ior_n iorn_n nior_n xor_n xnor_n); } + elsif ($base eq 'lorrshift') { @funs = qw(lshift rshift); } + else { @funs = ($base); } + + foreach my $fun (@funs) { + foreach my $pic ('', ' -DPIC') { + my $header = "$file: 0: $pic\n"; + $filecount++; + + my $m4 = "m4 -DHAVE_HOST_CPU_athlon -DOPERATION_$fun $pic ../config.m4 $tempfile $file"; + print "$m4\n" if $opt{'t'}; + + open IN, "$m4 |" or die; + while () { + next unless /([0-9+-][0-9 \t+-]*)\(%/; + my $pat=$1; + $pat = eval($pat); + next if ($pat != 0); + print "$header$_"; + $header=''; + } + close IN or die; + } + } +} + +unlink($tempfile); +print "total $filecount processed\n"; +exit 0; + + +# Local variables: +# perl-indent-level: 2 +# End: diff --git a/gmp-6.3.0/mpn/x86/udiv.asm b/gmp-6.3.0/mpn/x86/udiv.asm new file mode 100644 index 0000000..a3ee088 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/udiv.asm @@ -0,0 +1,52 @@ +dnl x86 mpn_udiv_qrnnd -- 2 by 1 limb division + +dnl Copyright 1999, 2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C mp_limb_t mpn_udiv_qrnnd (mp_limb_t *remptr, mp_limb_t high, mp_limb_t low, +C mp_limb_t divisor); + +defframe(PARAM_DIVISOR, 16) +defframe(PARAM_LOW, 12) +defframe(PARAM_HIGH, 8) +defframe(PARAM_REMPTR, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_udiv_qrnnd) +deflit(`FRAME',0) + movl PARAM_LOW, %eax + movl PARAM_HIGH, %edx + divl PARAM_DIVISOR + movl PARAM_REMPTR, %ecx + movl %edx, (%ecx) + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/umul.asm b/gmp-6.3.0/mpn/x86/umul.asm new file mode 100644 index 0000000..34fe434 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/umul.asm @@ -0,0 +1,51 @@ +dnl mpn_umul_ppmm -- 1x1->2 limb multiplication + +dnl Copyright 1999, 2000, 2002 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + +include(`../config.m4') + + +C mp_limb_t mpn_umul_ppmm (mp_limb_t *lowptr, mp_limb_t m1, mp_limb_t m2); +C + +defframe(PARAM_M2, 12) +defframe(PARAM_M1, 8) +defframe(PARAM_LOWPTR, 4) + + TEXT + ALIGN(8) +PROLOGUE(mpn_umul_ppmm) +deflit(`FRAME',0) + movl PARAM_LOWPTR, %ecx + movl PARAM_M1, %eax + mull PARAM_M2 + movl %eax, (%ecx) + movl %edx, %eax + ret +EPILOGUE() diff --git a/gmp-6.3.0/mpn/x86/x86-defs.m4 b/gmp-6.3.0/mpn/x86/x86-defs.m4 new file mode 100644 index 0000000..81309b2 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/x86-defs.m4 @@ -0,0 +1,1024 @@ +divert(-1) + +dnl m4 macros for x86 assembler. + +dnl Copyright 1999-2003, 2007, 2010, 2012, 2014 Free Software Foundation, Inc. + +dnl This file is part of the GNU MP Library. +dnl +dnl The GNU MP Library is free software; you can redistribute it and/or modify +dnl it under the terms of either: +dnl +dnl * the GNU Lesser General Public License as published by the Free +dnl Software Foundation; either version 3 of the License, or (at your +dnl option) any later version. +dnl +dnl or +dnl +dnl * the GNU General Public License as published by the Free Software +dnl Foundation; either version 2 of the License, or (at your option) any +dnl later version. +dnl +dnl or both in parallel, as here. +dnl +dnl The GNU MP Library is distributed in the hope that it will be useful, but +dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +dnl for more details. +dnl +dnl You should have received copies of the GNU General Public License and the +dnl GNU Lesser General Public License along with the GNU MP Library. If not, +dnl see https://www.gnu.org/licenses/. + + +dnl Notes: +dnl +dnl m4 isn't perfect for processing BSD style x86 assembler code, the main +dnl problems are, +dnl +dnl 1. Doing define(foo,123) and then using foo in an addressing mode like +dnl foo(%ebx) expands as a macro rather than a constant. This is worked +dnl around by using deflit() from asm-defs.m4, instead of define(). +dnl +dnl 2. Immediates in macro definitions need a space or `' to stop the $ +dnl looking like a macro parameter. For example, +dnl +dnl define(foo, `mov $ 123, %eax') +dnl +dnl This is only a problem in macro definitions, not in ordinary text, +dnl and not in macro parameters like text passed to forloop() or ifdef(). + + +deflit(GMP_LIMB_BYTES, 4) + + +dnl Libtool gives -DPIC -DDLL_EXPORT to indicate a cygwin or mingw DLL. We +dnl undefine PIC since we don't need to be position independent in this +dnl case and definitely don't want the ELF style _GLOBAL_OFFSET_TABLE_ etc. + +ifdef(`DLL_EXPORT',`undefine(`PIC')') + + +dnl Usage: CPUVEC_FUNCS_LIST +dnl +dnl A list of the functions from gmp-impl.h x86 struct cpuvec_t, in the +dnl order they appear in that structure. + +define(CPUVEC_FUNCS_LIST, +``add_n', +`addlsh1_n', +`addlsh2_n', +`addmul_1', +`addmul_2', +`bdiv_dbm1c', +`cnd_add_n', +`cnd_sub_n', +`com', +`copyd', +`copyi', +`divexact_1', +`divrem_1', +`gcd_11', +`lshift', +`lshiftc', +`mod_1', +`mod_1_1p', +`mod_1_1p_cps', +`mod_1s_2p', +`mod_1s_2p_cps', +`mod_1s_4p', +`mod_1s_4p_cps', +`mod_34lsub1', +`modexact_1c_odd', +`mul_1', +`mul_basecase', +`mullo_basecase', +`preinv_divrem_1', +`preinv_mod_1', +`redc_1', +`redc_2', +`rshift', +`sqr_basecase', +`sub_n', +`sublsh1_n', +`submul_1'') + + +dnl Called: PROLOGUE_cpu(GSYM_PREFIX`'foo) +dnl +dnl In the x86 code we use explicit TEXT and ALIGN() calls in the code, +dnl since different alignments are wanted in various circumstances. So for +dnl instance, +dnl +dnl TEXT +dnl ALIGN(16) +dnl PROLOGUE(mpn_add_n) +dnl ... +dnl EPILOGUE() + +define(`PROLOGUE_cpu', +m4_assert_numargs(1) +m4_assert_defined(`WANT_PROFILING') + `GLOBL $1 + TYPE($1,`function') + COFF_TYPE($1) +$1: +ifelse(WANT_PROFILING,`prof', ` call_mcount') +ifelse(WANT_PROFILING,`gprof', ` call_mcount') +ifelse(WANT_PROFILING,`instrument',` call_instrument(enter)') +') + + +dnl Usage: COFF_TYPE(GSYM_PREFIX`'foo) +dnl +dnl Emit COFF style ".def ... .endef" type information for a function, when +dnl supported. The argument should include any GSYM_PREFIX. +dnl +dnl See autoconf macro GMP_ASM_COFF_TYPE for HAVE_COFF_TYPE. + +define(COFF_TYPE, +m4_assert_numargs(1) +m4_assert_defined(`HAVE_COFF_TYPE') +`ifelse(HAVE_COFF_TYPE,yes, + `.def $1 + .scl 2 + .type 32 + .endef')') + + +dnl Usage: call_mcount +dnl +dnl For `gprof' style profiling, %ebp is setup as a frame pointer. None of +dnl the assembler routines use %ebp this way, so it's done only for the +dnl benefit of mcount. glibc sysdeps/i386/i386-mcount.S shows how mcount +dnl gets the current function from (%esp) and the parent from 4(%ebp). +dnl +dnl For `prof' style profiling gcc generates mcount calls without setting +dnl up %ebp, and the same is done here. + +define(`call_mcount', +m4_assert_numargs(-1) +m4_assert_defined(`WANT_PROFILING') +m4_assert_defined(`MCOUNT_PIC_REG') +m4_assert_defined(`MCOUNT_NONPIC_REG') +m4_assert_defined(`MCOUNT_PIC_CALL') +m4_assert_defined(`MCOUNT_NONPIC_CALL') +`ifelse(ifdef(`PIC',`MCOUNT_PIC_REG',`MCOUNT_NONPIC_REG'),,, +` DATA + ALIGN(4) +L(mcount_data_`'mcount_counter): + W32 0 + TEXT +')dnl +ifelse(WANT_PROFILING,`gprof', +` pushl %ebp + movl %esp, %ebp +')dnl +ifdef(`PIC', +` pushl %ebx + call_movl_eip_to_ebx +L(mcount_here_`'mcount_counter): + addl $_GLOBAL_OFFSET_TABLE_+[.-L(mcount_here_`'mcount_counter)], %ebx +ifelse(MCOUNT_PIC_REG,,, +` leal L(mcount_data_`'mcount_counter)@GOTOFF(%ebx), MCOUNT_PIC_REG') +MCOUNT_PIC_CALL + popl %ebx +',`dnl non-PIC +ifelse(MCOUNT_NONPIC_REG,,, +` movl `$'L(mcount_data_`'mcount_counter), MCOUNT_NONPIC_REG +')dnl +MCOUNT_NONPIC_CALL +')dnl +ifelse(WANT_PROFILING,`gprof', +` popl %ebp +') +define(`mcount_counter',incr(mcount_counter)) +') + +define(mcount_counter,1) + + +dnl Usage: call_instrument(enter|exit) +dnl +dnl Call __cyg_profile_func_enter or __cyg_profile_func_exit. +dnl +dnl For PIC, most routines don't require _GLOBAL_OFFSET_TABLE_ themselves +dnl so %ebx is just setup for these calls. It's a bit wasteful to repeat +dnl the setup for the exit call having done it earlier for the enter, but +dnl there's nowhere very convenient to hold %ebx through the length of a +dnl routine, in general. +dnl +dnl For PIC, because instrument_current_function will be within the current +dnl object file we can get it just as an offset from %eip, there's no need +dnl to use the GOT. +dnl +dnl No attempt is made to maintain the stack alignment gcc generates with +dnl -mpreferred-stack-boundary. This wouldn't be hard, but it seems highly +dnl unlikely the instrumenting functions would be doing anything that'd +dnl benefit from alignment, in particular they're unlikely to be using +dnl doubles or long doubles on the stack. +dnl +dnl The FRAME scheme is used to conveniently account for the register saves +dnl before accessing the return address. Any previous value is saved and +dnl restored, since plenty of code keeps a value across a "ret" in the +dnl middle of a routine. + +define(call_instrument, +m4_assert_numargs(1) +` pushdef(`FRAME',0) +ifelse($1,exit, +` pushl %eax FRAME_pushl() C return value +') +ifdef(`PIC', +` pushl %ebx FRAME_pushl() + call_movl_eip_to_ebx +L(instrument_here_`'instrument_count): + movl %ebx, %ecx + addl $_GLOBAL_OFFSET_TABLE_+[.-L(instrument_here_`'instrument_count)], %ebx + C use addl rather than leal to avoid old gas bugs, see mpn/x86/README + addl $instrument_current_function-L(instrument_here_`'instrument_count), %ecx + pushl m4_empty_if_zero(FRAME)(%esp) FRAME_pushl() C return addr + pushl %ecx FRAME_pushl() C this function + call GSYM_PREFIX`'__cyg_profile_func_$1@PLT + addl $`'8, %esp + popl %ebx +', +` C non-PIC + pushl m4_empty_if_zero(FRAME)(%esp) FRAME_pushl() C return addr + pushl $instrument_current_function FRAME_pushl() C this function + call GSYM_PREFIX`'__cyg_profile_func_$1 + addl $`'8, %esp +') +ifelse($1,exit, +` popl %eax C return value +') + popdef(`FRAME') +define(`instrument_count',incr(instrument_count)) +') +define(instrument_count,1) + + +dnl Usage: instrument_current_function +dnl +dnl Return the current function name for instrumenting purposes. This is +dnl PROLOGUE_current_function, but it sticks at the first such name seen. +dnl +dnl Sticking to the first name seen ensures that multiple-entrypoint +dnl functions like mpn_add_nc and mpn_add_n will make enter and exit calls +dnl giving the same function address. + +define(instrument_current_function, +m4_assert_numargs(-1) +`ifdef(`instrument_current_function_seen', +`instrument_current_function_seen', +`define(`instrument_current_function_seen',PROLOGUE_current_function)dnl +PROLOGUE_current_function')') + + +dnl Usage: call_movl_eip_to_ebx +dnl +dnl Generate a call to L(movl_eip_to_ebx), and record the need for that +dnl routine. + +define(call_movl_eip_to_ebx, +m4_assert_numargs(-1) +`call L(movl_eip_to_ebx) +define(`movl_eip_to_ebx_needed',1)') + +dnl Usage: generate_movl_eip_to_ebx +dnl +dnl Emit a L(movl_eip_to_ebx) routine, if needed and not already generated. + +define(generate_movl_eip_to_ebx, +m4_assert_numargs(-1) +`ifelse(movl_eip_to_ebx_needed,1, +`ifelse(movl_eip_to_ebx_done,1,, +`L(movl_eip_to_ebx): + movl (%esp), %ebx + ret_internal +define(`movl_eip_to_ebx_done',1) +')')') + + +dnl Usage: ret +dnl +dnl Generate a "ret", but if doing instrumented profiling then call +dnl __cyg_profile_func_exit first. + +define(ret, +m4_assert_numargs(-1) +m4_assert_defined(`WANT_PROFILING') +`ifelse(WANT_PROFILING,instrument, +`ret_instrument', +`ret_internal') +generate_movl_eip_to_ebx +') + + +dnl Usage: ret_internal +dnl +dnl A plain "ret", without any __cyg_profile_func_exit call. This can be +dnl used for a return which is internal to some function, such as when +dnl getting %eip for PIC. + +define(ret_internal, +m4_assert_numargs(-1) +``ret'') + + +dnl Usage: ret_instrument +dnl +dnl Generate call to __cyg_profile_func_exit and then a ret. If a ret has +dnl already been seen from this function then jump to that chunk of code, +dnl rather than emitting it again. + +define(ret_instrument, +m4_assert_numargs(-1) +`ifelse(m4_unquote(ret_instrument_seen_`'instrument_current_function),1, +`jmp L(instrument_exit_`'instrument_current_function)', +`define(ret_instrument_seen_`'instrument_current_function,1) +L(instrument_exit_`'instrument_current_function): +call_instrument(exit) + ret_internal')') + + +dnl Usage: _GLOBAL_OFFSET_TABLE_ +dnl +dnl Expand to _GLOBAL_OFFSET_TABLE_ plus any necessary underscore prefix. +dnl This lets us write plain _GLOBAL_OFFSET_TABLE_ in SVR4 style, but still +dnl work with systems requiring an extra underscore such as OpenBSD. +dnl +dnl deflit is used so "leal _GLOBAL_OFFSET_TABLE_(%eax), %ebx" will come +dnl out right, though that form doesn't work properly in gas (see +dnl mpn/x86/README). + +deflit(_GLOBAL_OFFSET_TABLE_, +m4_assert_defined(`GOT_GSYM_PREFIX') +`GOT_GSYM_PREFIX`_GLOBAL_OFFSET_TABLE_'') + + +dnl -------------------------------------------------------------------------- +dnl Various x86 macros. +dnl + + +dnl Usage: ALIGN_OFFSET(bytes,offset) +dnl +dnl Align to `offset' away from a multiple of `bytes'. +dnl +dnl This is useful for testing, for example align to something very strict +dnl and see what effect offsets from it have, "ALIGN_OFFSET(256,32)". +dnl +dnl Generally you wouldn't execute across the padding, but it's done with +dnl nop's so it'll work. + +define(ALIGN_OFFSET, +m4_assert_numargs(2) +`ALIGN($1) +forloop(`i',1,$2,` nop +')') + + +dnl Usage: defframe(name,offset) +dnl +dnl Make a definition like the following with which to access a parameter +dnl or variable on the stack. +dnl +dnl define(name,`FRAME+offset(%esp)') +dnl +dnl Actually m4_empty_if_zero(FRAME+offset) is used, which will save one +dnl byte if FRAME+offset is zero, by putting (%esp) rather than 0(%esp). +dnl Use define(`defframe_empty_if_zero_disabled',1) if for some reason the +dnl zero offset is wanted. +dnl +dnl The new macro also gets a check that when it's used FRAME is actually +dnl defined, and that the final %esp offset isn't negative, which would +dnl mean an attempt to access something below the current %esp. +dnl +dnl deflit() is used rather than a plain define(), so the new macro won't +dnl delete any following parenthesized expression. name(%edi) will come +dnl out say as 16(%esp)(%edi). This isn't valid assembler and should +dnl provoke an error, which is better than silently giving just 16(%esp). +dnl +dnl See README for more on the suggested way to access the stack frame. + +define(defframe, +m4_assert_numargs(2) +`deflit(`$1', +m4_assert_defined(`FRAME') +`defframe_check_notbelow(`$1',$2,FRAME)dnl +defframe_empty_if_zero(FRAME+($2))(%esp)')') + +dnl Called: defframe_empty_if_zero(expression) +define(defframe_empty_if_zero, +m4_assert_numargs(1) +`ifelse(defframe_empty_if_zero_disabled,1, +`eval($1)', +`m4_empty_if_zero($1)')') + +dnl Called: defframe_check_notbelow(`name',offset,FRAME) +define(defframe_check_notbelow, +m4_assert_numargs(3) +`ifelse(eval(($3)+($2)<0),1, +`m4_error(`$1 at frame offset $2 used when FRAME is only $3 bytes +')')') + + +dnl Usage: FRAME_pushl() +dnl FRAME_popl() +dnl FRAME_addl_esp(n) +dnl FRAME_subl_esp(n) +dnl +dnl Adjust FRAME appropriately for a pushl or popl, or for an addl or subl +dnl %esp of n bytes. +dnl +dnl Using these macros is completely optional. Sometimes it makes more +dnl sense to put explicit deflit(`FRAME',N) forms, especially when there's +dnl jumps and different sequences of FRAME values need to be used in +dnl different places. + +define(FRAME_pushl, +m4_assert_numargs(0) +m4_assert_defined(`FRAME') +`deflit(`FRAME',eval(FRAME+4))') + +define(FRAME_popl, +m4_assert_numargs(0) +m4_assert_defined(`FRAME') +`deflit(`FRAME',eval(FRAME-4))') + +define(FRAME_addl_esp, +m4_assert_numargs(1) +m4_assert_defined(`FRAME') +`deflit(`FRAME',eval(FRAME-($1)))') + +define(FRAME_subl_esp, +m4_assert_numargs(1) +m4_assert_defined(`FRAME') +`deflit(`FRAME',eval(FRAME+($1)))') + + +dnl Usage: defframe_pushl(name) +dnl +dnl Do a combination FRAME_pushl() and a defframe() to name the stack +dnl location just pushed. This should come after a pushl instruction. +dnl Putting it on the same line works and avoids lengthening the code. For +dnl example, +dnl +dnl pushl %eax defframe_pushl(VAR_COUNTER) +dnl +dnl Notice the defframe() is done with an unquoted -FRAME thus giving its +dnl current value without tracking future changes. + +define(defframe_pushl, +m4_assert_numargs(1) +`FRAME_pushl()defframe(`$1',-FRAME)') + + +dnl -------------------------------------------------------------------------- +dnl Assembler instruction macros. +dnl + + +dnl Usage: emms_or_femms +dnl femms_available_p +dnl +dnl femms_available_p expands to 1 or 0 according to whether the AMD 3DNow +dnl femms instruction is available. emms_or_femms expands to femms if +dnl available, or emms if not. +dnl +dnl emms_or_femms is meant for use in the K6 directory where plain K6 +dnl (without femms) and K6-2 and K6-3 (with a slightly faster femms) are +dnl supported together. +dnl +dnl On K7 femms is no longer faster and is just an alias for emms, so plain +dnl emms may as well be used. + +define(femms_available_p, +m4_assert_numargs(-1) +`m4_ifdef_anyof_p( + `HAVE_HOST_CPU_k62', + `HAVE_HOST_CPU_k63', + `HAVE_HOST_CPU_athlon')') + +define(emms_or_femms, +m4_assert_numargs(-1) +`ifelse(femms_available_p,1,`femms',`emms')') + + +dnl Usage: femms +dnl +dnl Gas 2.9.1 which comes with FreeBSD 3.4 doesn't support femms, so the +dnl following is a replacement using .byte. + +define(femms, +m4_assert_numargs(-1) +`.byte 15,14 C AMD 3DNow femms') + + +dnl Usage: jadcl0(op) +dnl +dnl Generate a jnc/incl as a substitute for adcl $0,op. Note this isn't an +dnl exact replacement, since it doesn't set the flags like adcl does. +dnl +dnl This finds a use in K6 mpn_addmul_1, mpn_submul_1, mpn_mul_basecase and +dnl mpn_sqr_basecase because on K6 an adcl is slow, the branch +dnl misprediction penalty is small, and the multiply algorithm used leads +dnl to a carry bit on average only 1/4 of the time. +dnl +dnl jadcl0_disabled can be set to 1 to instead generate an ordinary adcl +dnl for comparison. For example, +dnl +dnl define(`jadcl0_disabled',1) +dnl +dnl When using a register operand, eg. "jadcl0(%edx)", the jnc/incl code is +dnl the same size as an adcl. This makes it possible to use the exact same +dnl computed jump code when testing the relative speed of the two. + +define(jadcl0, +m4_assert_numargs(1) +`ifelse(jadcl0_disabled,1, + `adcl $`'0, $1', + `jnc L(jadcl0_`'jadcl0_counter) + incl $1 +L(jadcl0_`'jadcl0_counter): +define(`jadcl0_counter',incr(jadcl0_counter))')') + +define(jadcl0_counter,1) + + +dnl Usage: x86_lookup(target, key,value, key,value, ...) +dnl x86_lookup_p(target, key,value, key,value, ...) +dnl +dnl Look for `target' among the `key' parameters. +dnl +dnl x86_lookup expands to the corresponding `value', or generates an error +dnl if `target' isn't found. +dnl +dnl x86_lookup_p expands to 1 if `target' is found, or 0 if not. + +define(x86_lookup, +m4_assert_numargs_range(1,999) +`ifelse(eval($#<3),1, +`m4_error(`unrecognised part of x86 instruction: $1 +')', +`ifelse(`$1',`$2', `$3', +`x86_lookup(`$1',shift(shift(shift($@))))')')') + +define(x86_lookup_p, +m4_assert_numargs_range(1,999) +`ifelse(eval($#<3),1, `0', +`ifelse(`$1',`$2', `1', +`x86_lookup_p(`$1',shift(shift(shift($@))))')')') + + +dnl Usage: x86_opcode_reg32(reg) +dnl x86_opcode_reg32_p(reg) +dnl +dnl x86_opcode_reg32 expands to the standard 3 bit encoding for the given +dnl 32-bit register, eg. `%ebp' turns into 5. +dnl +dnl x86_opcode_reg32_p expands to 1 if reg is a valid 32-bit register, or 0 +dnl if not. + +define(x86_opcode_reg32, +m4_assert_numargs(1) +`x86_lookup(`$1',x86_opcode_reg32_list)') + +define(x86_opcode_reg32_p, +m4_assert_onearg() +`x86_lookup_p(`$1',x86_opcode_reg32_list)') + +define(x86_opcode_reg32_list, +``%eax',0, +`%ecx',1, +`%edx',2, +`%ebx',3, +`%esp',4, +`%ebp',5, +`%esi',6, +`%edi',7') + + +dnl Usage: x86_opcode_tttn(cond) +dnl +dnl Expand to the 4-bit "tttn" field value for the given x86 branch +dnl condition (like `c', `ae', etc). + +define(x86_opcode_tttn, +m4_assert_numargs(1) +`x86_lookup(`$1',x86_opcode_ttn_list)') + +define(x86_opcode_tttn_list, +``o', 0, +`no', 1, +`b', 2, `c', 2, `nae',2, +`nb', 3, `nc', 3, `ae', 3, +`e', 4, `z', 4, +`ne', 5, `nz', 5, +`be', 6, `na', 6, +`nbe', 7, `a', 7, +`s', 8, +`ns', 9, +`p', 10, `pe', 10, `npo',10, +`np', 11, `npe',11, `po', 11, +`l', 12, `nge',12, +`nl', 13, `ge', 13, +`le', 14, `ng', 14, +`nle',15, `g', 15') + + +dnl Usage: cmovCC(%srcreg,%dstreg) +dnl +dnl Emit a cmov instruction, using a .byte sequence, since various past +dnl versions of gas don't know cmov. For example, +dnl +dnl cmovz( %eax, %ebx) +dnl +dnl The source operand can only be a plain register. (m4 code implementing +dnl full memory addressing modes exists, believe it or not, but isn't +dnl currently needed and isn't included.) +dnl +dnl All the standard conditions are defined. Attempting to use one without +dnl the macro parentheses, such as just "cmovbe %eax, %ebx", will provoke +dnl an error. This protects against writing something old gas wouldn't +dnl understand. + +dnl Called: define_cmov_many(cond,tttn,cond,tttn,...) +define(define_cmov_many, +`ifelse(m4_length(`$1'),0,, +`define_cmov(`$1',`$2')define_cmov_many(shift(shift($@)))')') + +dnl Called: define_cmov(cond,tttn) +dnl Emit basically define(cmov,`cmov_internal(,,`$1',`$2')') +define(define_cmov, +m4_assert_numargs(2) +`define(`cmov$1', +m4_instruction_wrapper() +m4_assert_numargs(2) +`cmov_internal'(m4_doublequote($`'0),``$2'',dnl +m4_doublequote($`'1),m4_doublequote($`'2)))') + +define_cmov_many(x86_opcode_tttn_list) + +dnl Called: cmov_internal(name,tttn,src,dst) +define(cmov_internal, +m4_assert_numargs(4) +`.byte dnl +15, dnl +eval(64+$2), dnl +eval(192+8*x86_opcode_reg32(`$4')+x86_opcode_reg32(`$3')) dnl + C `$1 $3, $4'') + + +dnl Usage: x86_opcode_regmmx(reg) +dnl +dnl Validate the given mmx register, and return its number, 0 to 7. + +define(x86_opcode_regmmx, +m4_assert_numargs(1) +`x86_lookup(`$1',x86_opcode_regmmx_list)') + +define(x86_opcode_regmmx_list, +``%mm0',0, +`%mm1',1, +`%mm2',2, +`%mm3',3, +`%mm4',4, +`%mm5',5, +`%mm6',6, +`%mm7',7') + + +dnl Usage: psadbw(%srcreg,%dstreg) +dnl +dnl Oldish versions of gas don't know psadbw, in particular gas 2.9.1 on +dnl FreeBSD 3.3 and 3.4 doesn't, so instead emit .byte sequences. For +dnl example, +dnl +dnl psadbw( %mm1, %mm2) +dnl +dnl Only register->register forms are supported here, which suffices for +dnl the current code. + +define(psadbw, +m4_instruction_wrapper() +m4_assert_numargs(2) +`.byte 0x0f,0xf6,dnl +eval(192+x86_opcode_regmmx(`$2')*8+x86_opcode_regmmx(`$1')) dnl + C `psadbw $1, $2'') + + +dnl Usage: Zdisp(inst,op,op,op) +dnl +dnl Generate explicit .byte sequences if necessary to force a byte-sized +dnl zero displacement on an instruction. For example, +dnl +dnl Zdisp( movl, 0,(%esi), %eax) +dnl +dnl expands to +dnl +dnl .byte 139,70,0 C movl 0(%esi), %eax +dnl +dnl If the displacement given isn't 0, then normal assembler code is +dnl generated. For example, +dnl +dnl Zdisp( movl, 4,(%esi), %eax) +dnl +dnl expands to +dnl +dnl movl 4(%esi), %eax +dnl +dnl This means a single Zdisp() form can be used with an expression for the +dnl displacement, and .byte will be used only if necessary. The +dnl displacement argument is eval()ed. +dnl +dnl Because there aren't many places a 0(reg) form is wanted, Zdisp is +dnl implemented with a table of instructions and encodings. A new entry is +dnl needed for any different operation or registers. The table is split +dnl into separate macros to avoid overflowing BSD m4 macro expansion space. + +define(Zdisp, +m4_assert_numargs(4) +`define(`Zdisp_found',0)dnl +Zdisp_1($@)dnl +Zdisp_2($@)dnl +Zdisp_3($@)dnl +Zdisp_4($@)dnl +ifelse(Zdisp_found,0, +`m4_error(`unrecognised instruction in Zdisp: $1 $2 $3 $4 +')')') + +define(Zdisp_1,`dnl +Zdisp_match( adcl, 0,(%edx), %eax, `0x13,0x42,0x00', $@)`'dnl +Zdisp_match( adcl, 0,(%edx), %ebx, `0x13,0x5a,0x00', $@)`'dnl +Zdisp_match( adcl, 0,(%edx), %esi, `0x13,0x72,0x00', $@)`'dnl +Zdisp_match( addl, %ebx, 0,(%edi), `0x01,0x5f,0x00', $@)`'dnl +Zdisp_match( addl, %ecx, 0,(%edi), `0x01,0x4f,0x00', $@)`'dnl +Zdisp_match( addl, %esi, 0,(%edi), `0x01,0x77,0x00', $@)`'dnl +Zdisp_match( sbbl, 0,(%edx), %eax, `0x1b,0x42,0x00', $@)`'dnl +Zdisp_match( sbbl, 0,(%edx), %esi, `0x1b,0x72,0x00', $@)`'dnl +Zdisp_match( subl, %ecx, 0,(%edi), `0x29,0x4f,0x00', $@)`'dnl +Zdisp_match( movzbl, 0,(%eax,%ebp), %eax, `0x0f,0xb6,0x44,0x28,0x00', $@)`'dnl +Zdisp_match( movzbl, 0,(%ecx,%edi), %edi, `0x0f,0xb6,0x7c,0x39,0x00', $@)`'dnl +Zdisp_match( adc, 0,(%ebx,%ecx,4), %eax, `0x13,0x44,0x8b,0x00', $@)`'dnl +Zdisp_match( sbb, 0,(%ebx,%ecx,4), %eax, `0x1b,0x44,0x8b,0x00', $@)`'dnl +') +define(Zdisp_2,`dnl +Zdisp_match( movl, %eax, 0,(%edi), `0x89,0x47,0x00', $@)`'dnl +Zdisp_match( movl, %ebx, 0,(%edi), `0x89,0x5f,0x00', $@)`'dnl +Zdisp_match( movl, %esi, 0,(%edi), `0x89,0x77,0x00', $@)`'dnl +Zdisp_match( movl, 0,(%ebx), %eax, `0x8b,0x43,0x00', $@)`'dnl +Zdisp_match( movl, 0,(%ebx), %esi, `0x8b,0x73,0x00', $@)`'dnl +Zdisp_match( movl, 0,(%edx), %eax, `0x8b,0x42,0x00', $@)`'dnl +Zdisp_match( movl, 0,(%esi), %eax, `0x8b,0x46,0x00', $@)`'dnl +Zdisp_match( movl, 0,(%esi,%ecx,4), %eax, `0x8b,0x44,0x8e,0x00', $@)`'dnl +Zdisp_match( mov, 0,(%esi,%ecx,4), %eax, `0x8b,0x44,0x8e,0x00', $@)`'dnl +Zdisp_match( mov, %eax, 0,(%edi,%ecx,4), `0x89,0x44,0x8f,0x00', $@)`'dnl +') +define(Zdisp_3,`dnl +Zdisp_match( movq, 0,(%eax,%ecx,8), %mm0, `0x0f,0x6f,0x44,0xc8,0x00', $@)`'dnl +Zdisp_match( movq, 0,(%ebx,%eax,4), %mm0, `0x0f,0x6f,0x44,0x83,0x00', $@)`'dnl +Zdisp_match( movq, 0,(%ebx,%eax,4), %mm2, `0x0f,0x6f,0x54,0x83,0x00', $@)`'dnl +Zdisp_match( movq, 0,(%ebx,%ecx,4), %mm0, `0x0f,0x6f,0x44,0x8b,0x00', $@)`'dnl +Zdisp_match( movq, 0,(%edx), %mm0, `0x0f,0x6f,0x42,0x00', $@)`'dnl +Zdisp_match( movq, 0,(%esi), %mm0, `0x0f,0x6f,0x46,0x00', $@)`'dnl +Zdisp_match( movq, %mm0, 0,(%edi), `0x0f,0x7f,0x47,0x00', $@)`'dnl +Zdisp_match( movq, %mm2, 0,(%ecx,%eax,4), `0x0f,0x7f,0x54,0x81,0x00', $@)`'dnl +Zdisp_match( movq, %mm2, 0,(%edx,%eax,4), `0x0f,0x7f,0x54,0x82,0x00', $@)`'dnl +Zdisp_match( movq, %mm0, 0,(%edx,%ecx,8), `0x0f,0x7f,0x44,0xca,0x00', $@)`'dnl +') +define(Zdisp_4,`dnl +Zdisp_match( movd, 0,(%eax,%ecx,4), %mm0, `0x0f,0x6e,0x44,0x88,0x00', $@)`'dnl +Zdisp_match( movd, 0,(%eax,%ecx,8), %mm1, `0x0f,0x6e,0x4c,0xc8,0x00', $@)`'dnl +Zdisp_match( movd, 0,(%edx,%ecx,8), %mm0, `0x0f,0x6e,0x44,0xca,0x00', $@)`'dnl +Zdisp_match( movd, %mm0, 0,(%eax,%ecx,4), `0x0f,0x7e,0x44,0x88,0x00', $@)`'dnl +Zdisp_match( movd, %mm0, 0,(%ecx,%eax,4), `0x0f,0x7e,0x44,0x81,0x00', $@)`'dnl +Zdisp_match( movd, %mm2, 0,(%ecx,%eax,4), `0x0f,0x7e,0x54,0x81,0x00', $@)`'dnl +Zdisp_match( movd, %mm0, 0,(%edx,%ecx,4), `0x0f,0x7e,0x44,0x8a,0x00', $@)`'dnl +') + +define(Zdisp_match, +m4_assert_numargs(9) +`ifelse(eval(m4_stringequal_p(`$1',`$6') + && m4_stringequal_p(`$2',0) + && m4_stringequal_p(`$3',`$8') + && m4_stringequal_p(`$4',`$9')),1, +`define(`Zdisp_found',1)dnl +ifelse(eval(`$7'),0, +` .byte $5 C `$1 0$3, $4'', +` $6 $7$8, $9')', + +`ifelse(eval(m4_stringequal_p(`$1',`$6') + && m4_stringequal_p(`$2',`$7') + && m4_stringequal_p(`$3',0) + && m4_stringequal_p(`$4',`$9')),1, +`define(`Zdisp_found',1)dnl +ifelse(eval(`$8'),0, +` .byte $5 C `$1 $2, 0$4'', +` $6 $7, $8$9')')')') + + +dnl Usage: shldl(count,src,dst) +dnl shrdl(count,src,dst) +dnl shldw(count,src,dst) +dnl shrdw(count,src,dst) +dnl +dnl Generate a double-shift instruction, possibly omitting a %cl count +dnl parameter if that's what the assembler requires, as indicated by +dnl WANT_SHLDL_CL in config.m4. For example, +dnl +dnl shldl( %cl, %eax, %ebx) +dnl +dnl turns into either +dnl +dnl shldl %cl, %eax, %ebx +dnl or +dnl shldl %eax, %ebx +dnl +dnl Immediate counts are always passed through unchanged. For example, +dnl +dnl shrdl( $2, %esi, %edi) +dnl becomes +dnl shrdl $2, %esi, %edi +dnl +dnl +dnl If you forget to use the macro form "shldl( ...)" and instead write +dnl just a plain "shldl ...", an error results. This ensures the necessary +dnl variant treatment of %cl isn't accidentally bypassed. + +define(define_shd_instruction, +m4_assert_numargs(1) +`define($1, +m4_instruction_wrapper() +m4_assert_numargs(3) +`shd_instruction'(m4_doublequote($`'0),m4_doublequote($`'1),dnl +m4_doublequote($`'2),m4_doublequote($`'3)))') + +dnl Effectively: define(shldl,`shd_instruction(`$0',`$1',`$2',`$3')') etc +define_shd_instruction(shldl) +define_shd_instruction(shrdl) +define_shd_instruction(shldw) +define_shd_instruction(shrdw) + +dnl Called: shd_instruction(op,count,src,dst) +define(shd_instruction, +m4_assert_numargs(4) +m4_assert_defined(`WANT_SHLDL_CL') +`ifelse(eval(m4_stringequal_p(`$2',`%cl') && !WANT_SHLDL_CL),1, +``$1' `$3', `$4'', +``$1' `$2', `$3', `$4'')') + + +dnl Usage: ASSERT([cond][,instructions]) +dnl +dnl If WANT_ASSERT is 1, output the given instructions and expect the given +dnl flags condition to then be satisfied. For example, +dnl +dnl ASSERT(ne, `cmpl %eax, %ebx') +dnl +dnl The instructions can be omitted to just assert a flags condition with +dnl no extra calculation. For example, +dnl +dnl ASSERT(nc) +dnl +dnl When `instructions' is not empty, a pushf/popf is added to preserve the +dnl flags, but the instructions themselves must preserve any registers that +dnl matter. FRAME is adjusted for the push and pop, so the instructions +dnl given can use defframe() stack variables. +dnl +dnl The condition can be omitted to just output the given instructions when +dnl assertion checking is wanted. In this case the pushf/popf is omitted. +dnl For example, +dnl +dnl ASSERT(, `movl %eax, VAR_KEEPVAL') + +define(ASSERT, +m4_assert_numargs_range(1,2) +m4_assert_defined(`WANT_ASSERT') +`ifelse(WANT_ASSERT,1, +`ifelse(`$1',, + `$2', + `C ASSERT +ifelse(`$2',,,` pushf ifdef(`FRAME',`FRAME_pushl()')') + $2 + j`$1' L(ASSERT_ok`'ASSERT_counter) + ud2 C assertion failed +L(ASSERT_ok`'ASSERT_counter): +ifelse(`$2',,,` popf ifdef(`FRAME',`FRAME_popl()')') +define(`ASSERT_counter',incr(ASSERT_counter))')')') + +define(ASSERT_counter,1) + + +dnl Usage: movl_text_address(label,register) +dnl +dnl Get the address of a text segment label, using either a plain movl or a +dnl position-independent calculation, as necessary. For example, +dnl +dnl movl_code_address(L(foo),%eax) +dnl +dnl This macro is only meant for use in ASSERT()s or when testing, since +dnl the PIC sequence it generates will want to be done with a ret balancing +dnl the call on CPUs with return address branch prediction. +dnl +dnl The addl generated here has a backward reference to the label, and so +dnl won't suffer from the two forwards references bug in old gas (described +dnl in mpn/x86/README). + +define(movl_text_address, +m4_assert_numargs(2) +`ifdef(`PIC', + `call L(movl_text_address_`'movl_text_address_counter) +L(movl_text_address_`'movl_text_address_counter): + popl $2 C %eip + addl `$'$1-L(movl_text_address_`'movl_text_address_counter), $2 +define(`movl_text_address_counter',incr(movl_text_address_counter))', + `movl `$'$1, $2')') + +define(movl_text_address_counter,1) + + +dnl Usage: notl_or_xorl_GMP_NUMB_MASK(reg) +dnl +dnl Expand to either "notl `reg'" or "xorl $GMP_NUMB_BITS,`reg'" as +dnl appropriate for nails in use or not. + +define(notl_or_xorl_GMP_NUMB_MASK, +m4_assert_numargs(1) +`ifelse(GMP_NAIL_BITS,0, +`notl `$1'', +`xorl $GMP_NUMB_MASK, `$1'')') + + +dnl Usage LEA(symbol,reg) +dnl Usage LEAL(symbol_local_to_file,reg) + +define(`LEA', +m4_assert_numargs(2) +`ifdef(`PIC',`dnl +ifelse(index(defn(`load_eip'), `$2'),-1, +`m4append(`load_eip', +` TEXT + ALIGN(16) +L(movl_eip_`'substr($2,1)): + movl (%esp), $2 + ret_internal +')')dnl + call L(movl_eip_`'substr($2,1)) + addl $_GLOBAL_OFFSET_TABLE_, $2 + movl $1@GOT($2), $2 +',` + movl `$'$1, $2 +')') + +define(`LEAL', +m4_assert_numargs(2) +`ifdef(`PIC',`dnl +ifelse(index(defn(`load_eip'), `$2'),-1, +`m4append(`load_eip', +` TEXT + ALIGN(16) +L(movl_eip_`'substr($2,1)): + movl (%esp), $2 + ret_internal +')')dnl + call L(movl_eip_`'substr($2,1)) + addl $_GLOBAL_OFFSET_TABLE_, $2 + leal $1@GOTOFF($2), $2 +',` + movl `$'$1, $2 +')') + +dnl ASM_END + +define(`ASM_END',`load_eip') + +define(`load_eip', `') dnl updated in LEA/LEAL + + +define(`DEF_OBJECT', +m4_assert_numargs_range(1,2) + `RODATA + ALIGN(ifelse($#,1,2,$2)) +$1: +') + +define(`END_OBJECT', +m4_assert_numargs(1) +` SIZE(`$1',.-`$1')') + +dnl Usage: CALL(funcname) +dnl + +define(`CALL', +m4_assert_numargs(1) +`ifdef(`PIC', + `call GSYM_PREFIX`'$1@PLT', + `call GSYM_PREFIX`'$1')') + +ifdef(`PIC', +`define(`PIC_WITH_EBX')', +`undefine(`PIC_WITH_EBX')') + +divert`'dnl diff --git a/gmp-6.3.0/mpn/x86/zn1/gmp-mparam.h b/gmp-6.3.0/mpn/x86/zn1/gmp-mparam.h new file mode 100644 index 0000000..8e6c052 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/zn1/gmp-mparam.h @@ -0,0 +1,220 @@ +/* AMD zn1/32 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 3700-4300 MHz Pinnacle Ridge */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-21, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 0 /* always */ +#define MOD_1_UNNORM_THRESHOLD 3 +#define MOD_1N_TO_MOD_1_1_THRESHOLD 4 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 3 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 10 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 9 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 14.00% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 4 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 22 + +#define DIV_1_VS_MUL_1_PERCENT 248 + +#define MUL_TOOM22_THRESHOLD 28 +#define MUL_TOOM33_THRESHOLD 91 +#define MUL_TOOM44_THRESHOLD 137 +#define MUL_TOOM6H_THRESHOLD 222 +#define MUL_TOOM8H_THRESHOLD 454 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 85 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 103 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 88 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 105 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 130 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 63 +#define SQR_TOOM3_THRESHOLD 98 +#define SQR_TOOM4_THRESHOLD 172 +#define SQR_TOOM6_THRESHOLD 286 +#define SQR_TOOM8_THRESHOLD 478 + +#define MULMID_TOOM42_THRESHOLD 64 + +#define MULMOD_BNM1_THRESHOLD 21 +#define SQRMOD_BNM1_THRESHOLD 17 + +#define MUL_FFT_MODF_THRESHOLD 606 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 606, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 15, 5}, { 31, 6}, { 27, 7}, { 15, 6}, \ + { 33, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 29, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 43, 9}, { 23, 8}, { 51, 9}, { 31, 8}, \ + { 67, 9}, { 39, 8}, { 79, 9}, { 47, 8}, \ + { 95,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 127,10}, \ + { 79, 9}, { 159,10}, { 95,11}, { 63,10}, \ + { 159,11}, { 95,10}, { 191,12}, { 63,11}, \ + { 127,10}, { 255, 9}, { 511,10}, { 271, 9}, \ + { 543, 8}, { 1087,11}, { 159,10}, { 319, 9}, \ + { 639,10}, { 335, 9}, { 671,11}, { 191,10}, \ + { 383, 9}, { 767,10}, { 399,11}, { 223,12}, \ + { 127,11}, { 255,10}, { 543, 9}, { 1087,11}, \ + { 287,10}, { 607, 9}, { 1215,11}, { 319,10}, \ + { 671, 9}, { 1343,11}, { 351,12}, { 191,11}, \ + { 383,10}, { 799,11}, { 415,13}, { 127,12}, \ + { 255,11}, { 543,10}, { 1087,11}, { 607,10}, \ + { 1215,12}, { 319,11}, { 671,10}, { 1343,11}, \ + { 735,10}, { 1471,12}, { 383,11}, { 799,10}, \ + { 1599,11}, { 863,10}, { 1727,12}, { 447,11}, \ + { 959,10}, { 1919,11}, { 991,13}, { 255,12}, \ + { 511,11}, { 1087,12}, { 575,11}, { 1215,10}, \ + { 2431,12}, { 639,11}, { 1343,12}, { 703,11}, \ + { 1471,10}, { 2943,13}, { 383,12}, { 767,11}, \ + { 1599,12}, { 831,11}, { 1727,10}, { 3455,12}, \ + { 959,11}, { 1919,14}, { 255,13}, { 511,12}, \ + { 1087,11}, { 2239,12}, { 1215,11}, { 2431,13}, \ + { 639,12}, { 1471,11}, { 2943,13}, { 767,12}, \ + { 1727,11}, { 3455,13}, { 895,12}, { 1983,14}, \ + { 511,13}, { 1023,12}, { 2239,13}, { 1151,12}, \ + { 2495,13}, { 1279,12}, { 2623,13}, { 1407,12}, \ + { 2943,14}, { 767,13}, { 1663,12}, { 3455,13}, \ + { 1919,12}, { 3839,15}, { 511,14}, { 1023,13}, \ + { 2175,12}, { 4479,13}, { 2431,14}, { 1279,13}, \ + { 2943,12}, { 5887,14}, { 1535,13}, { 3455,14}, \ + { 1791,13}, { 3967,12}, { 7935,11}, { 15871,15}, \ + { 1023,14}, { 2047,13}, { 4479,14}, { 2303,13}, \ + { 4991,12}, { 9983,14}, { 2815,13}, { 5887,15}, \ + { 1535,14}, { 3839,13}, { 7935,12}, { 15871,16} } +#define MUL_FFT_TABLE3_SIZE 172 +#define MUL_FFT_THRESHOLD 5760 + +#define SQR_FFT_MODF_THRESHOLD 464 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 464, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 25, 7}, { 13, 6}, { 28, 7}, { 15, 6}, \ + { 31, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 39, 9}, { 23, 8}, \ + { 51,10}, { 15, 9}, { 31, 8}, { 67, 9}, \ + { 39, 8}, { 79, 9}, { 47, 8}, { 95, 9}, \ + { 55,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 135,10}, \ + { 79, 9}, { 159,10}, { 95,11}, { 63,10}, \ + { 127, 9}, { 255,10}, { 143, 9}, { 287,10}, \ + { 159,11}, { 95,12}, { 63,11}, { 127,10}, \ + { 255, 9}, { 511,10}, { 271, 9}, { 543,10}, \ + { 287, 9}, { 575,11}, { 159, 9}, { 639,10}, \ + { 335, 9}, { 671,10}, { 351, 9}, { 703,11}, \ + { 191,10}, { 383, 9}, { 767,10}, { 399, 9}, \ + { 799,10}, { 415,12}, { 127,11}, { 255,10}, \ + { 543,11}, { 287,10}, { 607,11}, { 319,10}, \ + { 671,11}, { 351,10}, { 703,12}, { 191,11}, \ + { 383,10}, { 799,11}, { 415,10}, { 831,13}, \ + { 127,12}, { 255,11}, { 543,10}, { 1087,11}, \ + { 607,10}, { 1215,12}, { 319,11}, { 671,10}, \ + { 1343,11}, { 735,10}, { 1471,12}, { 383,11}, \ + { 799,10}, { 1599,11}, { 863,12}, { 447,11}, \ + { 959,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,10}, { 2431,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,10}, \ + { 3455,12}, { 959,11}, { 1919,14}, { 255,13}, \ + { 511,12}, { 1087,11}, { 2239,12}, { 1215,11}, \ + { 2431,13}, { 639,12}, { 1471,11}, { 2943,13}, \ + { 767,12}, { 1727,11}, { 3455,13}, { 895,12}, \ + { 1919,14}, { 511,13}, { 1023,12}, { 2239,13}, \ + { 1151,12}, { 2431,13}, { 1279,12}, { 2623,13}, \ + { 1407,12}, { 2943,14}, { 767,13}, { 1663,12}, \ + { 3455,13}, { 1919,12}, { 3839,15}, { 511,14}, \ + { 1023,13}, { 2175,12}, { 4479,13}, { 2431,14}, \ + { 1279,13}, { 2943,12}, { 5887,14}, { 1535,13}, \ + { 3455,14}, { 1791,13}, { 3839,12}, { 7679,13}, \ + { 3967,12}, { 7935,15}, { 1023,14}, { 2047,13}, \ + { 4479,14}, { 2303,13}, { 4991,12}, { 9983,14}, \ + { 2815,13}, { 5887,15}, { 1535,14}, { 3839,13}, \ + { 7935,16} } +#define SQR_FFT_TABLE3_SIZE 173 +#define SQR_FFT_THRESHOLD 4736 + +#define MULLO_BASECASE_THRESHOLD 3 +#define MULLO_DC_THRESHOLD 60 +#define MULLO_MUL_N_THRESHOLD 11278 +#define SQRLO_BASECASE_THRESHOLD 8 +#define SQRLO_DC_THRESHOLD 161 +#define SQRLO_SQR_THRESHOLD 9335 + +#define DC_DIV_QR_THRESHOLD 71 +#define DC_DIVAPPR_Q_THRESHOLD 206 +#define DC_BDIV_QR_THRESHOLD 63 +#define DC_BDIV_Q_THRESHOLD 126 + +#define INV_MULMOD_BNM1_THRESHOLD 78 +#define INV_NEWTON_THRESHOLD 274 +#define INV_APPR_THRESHOLD 228 + +#define BINV_NEWTON_THRESHOLD 274 +#define REDC_1_TO_REDC_N_THRESHOLD 71 + +#define MU_DIV_QR_THRESHOLD 1652 +#define MU_DIVAPPR_Q_THRESHOLD 1718 +#define MUPI_DIV_QR_THRESHOLD 122 +#define MU_BDIV_QR_THRESHOLD 1470 +#define MU_BDIV_Q_THRESHOLD 1589 + +#define POWM_SEC_TABLE 3,28,54,386,1337 + +#define GET_STR_DC_THRESHOLD 13 +#define GET_STR_PRECOMPUTE_THRESHOLD 19 +#define SET_STR_DC_THRESHOLD 262 +#define SET_STR_PRECOMPUTE_THRESHOLD 558 + +#define FAC_DSC_THRESHOLD 109 +#define FAC_ODD_THRESHOLD 39 + +#define MATRIX22_STRASSEN_THRESHOLD 21 +#define HGCD2_DIV1_METHOD 1 /* 7.49% faster than 3 */ +#define HGCD_THRESHOLD 74 +#define HGCD_APPR_THRESHOLD 70 +#define HGCD_REDUCE_THRESHOLD 3389 +#define GCD_DC_THRESHOLD 440 +#define GCDEXT_DC_THRESHOLD 327 +#define JACOBI_BASE_METHOD 1 /* 11.98% faster than 3 */ + +/* Tuneup completed successfully, took 36916 seconds */ diff --git a/gmp-6.3.0/mpn/x86/zn2/gmp-mparam.h b/gmp-6.3.0/mpn/x86/zn2/gmp-mparam.h new file mode 100644 index 0000000..152e6b7 --- /dev/null +++ b/gmp-6.3.0/mpn/x86/zn2/gmp-mparam.h @@ -0,0 +1,226 @@ +/* AMD zn2/32 gmp-mparam.h -- Compiler/machine parameter header file. + +Copyright 2019 Free Software Foundation, Inc. + +This file is part of the GNU MP Library. + +The GNU MP Library is free software; you can redistribute it and/or modify +it under the terms of either: + + * the GNU Lesser General Public License as published by the Free + Software Foundation; either version 3 of the License, or (at your + option) any later version. + +or + + * the GNU General Public License as published by the Free Software + Foundation; either version 2 of the License, or (at your option) any + later version. + +or both in parallel, as here. + +The GNU MP Library is distributed in the hope that it will be useful, but +WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License +for more details. + +You should have received copies of the GNU General Public License and the +GNU Lesser General Public License along with the GNU MP Library. If not, +see https://www.gnu.org/licenses/. */ + +#define GMP_LIMB_BITS 32 +#define GMP_LIMB_BYTES 4 + +/* 3600-4400 MHz Matisse */ +/* FFT tuning limit = 67,000,000 */ +/* Generated by tuneup.c, 2019-10-23, gcc 8.3 */ + +#define MOD_1_NORM_THRESHOLD 0 /* always */ +#define MOD_1_UNNORM_THRESHOLD 0 /* always */ +#define MOD_1N_TO_MOD_1_1_THRESHOLD 3 +#define MOD_1U_TO_MOD_1_1_THRESHOLD 3 +#define MOD_1_1_TO_MOD_1_2_THRESHOLD 15 +#define MOD_1_2_TO_MOD_1_4_THRESHOLD 0 /* never mpn_mod_1s_2p */ +#define PREINV_MOD_1_TO_MOD_1_THRESHOLD 9 +#define USE_PREINV_DIVREM_1 1 /* native */ +#define DIV_QR_1N_PI1_METHOD 1 /* 4.78% faster than 2 */ +#define DIV_QR_1_NORM_THRESHOLD 3 +#define DIV_QR_1_UNNORM_THRESHOLD MP_SIZE_T_MAX /* never */ +#define DIV_QR_2_PI2_THRESHOLD 7 +#define DIVEXACT_1_THRESHOLD 0 /* always (native) */ +#define BMOD_1_TO_MOD_1_THRESHOLD 23 + +#define DIV_1_VS_MUL_1_PERCENT 274 + +#define MUL_TOOM22_THRESHOLD 24 +#define MUL_TOOM33_THRESHOLD 85 +#define MUL_TOOM44_THRESHOLD 166 +#define MUL_TOOM6H_THRESHOLD 290 +#define MUL_TOOM8H_THRESHOLD 430 + +#define MUL_TOOM32_TO_TOOM43_THRESHOLD 97 +#define MUL_TOOM32_TO_TOOM53_THRESHOLD 114 +#define MUL_TOOM42_TO_TOOM53_THRESHOLD 97 +#define MUL_TOOM42_TO_TOOM63_THRESHOLD 113 +#define MUL_TOOM43_TO_TOOM54_THRESHOLD 130 + +#define SQR_BASECASE_THRESHOLD 0 /* always (native) */ +#define SQR_TOOM2_THRESHOLD 26 +#define SQR_TOOM3_THRESHOLD 153 +#define SQR_TOOM4_THRESHOLD 214 +#define SQR_TOOM6_THRESHOLD 318 +#define SQR_TOOM8_THRESHOLD 478 + +#define MULMID_TOOM42_THRESHOLD 48 + +#define MULMOD_BNM1_THRESHOLD 18 +#define SQRMOD_BNM1_THRESHOLD 24 + +#define MUL_FFT_MODF_THRESHOLD 444 /* k = 5 */ +#define MUL_FFT_TABLE3 \ + { { 444, 5}, { 21, 6}, { 11, 5}, { 23, 6}, \ + { 12, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 25, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \ + { 31, 7}, { 17, 6}, { 35, 7}, { 19, 6}, \ + { 39, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 31, 7}, { 63, 8}, \ + { 39, 9}, { 23, 8}, { 51,10}, { 15, 9}, \ + { 31, 8}, { 67, 9}, { 39, 8}, { 79, 9}, \ + { 47,10}, { 31, 9}, { 79,10}, { 47, 9}, \ + { 95,11}, { 31,10}, { 63, 9}, { 127,10}, \ + { 79, 9}, { 159,10}, { 95,11}, { 63,10}, \ + { 127, 9}, { 255, 8}, { 511,10}, { 143, 9}, \ + { 287, 8}, { 575,10}, { 159,11}, { 95,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 511,10}, \ + { 271, 9}, { 543,10}, { 287, 9}, { 575,11}, \ + { 159,10}, { 319, 9}, { 639,10}, { 335, 9}, \ + { 671, 8}, { 1343,10}, { 351, 9}, { 703,10}, \ + { 367, 9}, { 735,11}, { 191,10}, { 383, 9}, \ + { 767,10}, { 415,11}, { 223,10}, { 447,12}, \ + { 127,11}, { 255,10}, { 543, 9}, { 1087,11}, \ + { 287,10}, { 607,11}, { 319,10}, { 671, 9}, \ + { 1343,11}, { 351,10}, { 735,12}, { 191,11}, \ + { 383,10}, { 767,11}, { 415,10}, { 831,11}, \ + { 447,13}, { 127,12}, { 255,11}, { 543,10}, \ + { 1087,11}, { 607,10}, { 1215,12}, { 319,11}, \ + { 671,10}, { 1343,11}, { 735,10}, { 1471, 9}, \ + { 2943,12}, { 383,11}, { 799,10}, { 1599,11}, \ + { 863,12}, { 447,11}, { 959,10}, { 1919,13}, \ + { 255,12}, { 511,11}, { 1087,12}, { 575,11}, \ + { 1215,10}, { 2431,12}, { 639,11}, { 1343,12}, \ + { 703,11}, { 1471,10}, { 2943,13}, { 383,12}, \ + { 767,11}, { 1599,12}, { 831,11}, { 1727,10}, \ + { 3455,12}, { 959,11}, { 1919,10}, { 3839,14}, \ + { 255,13}, { 511,12}, { 1215,11}, { 2431,13}, \ + { 639,12}, { 1471,11}, { 2943,10}, { 5887,13}, \ + { 767,12}, { 1727,11}, { 3455,13}, { 895,12}, \ + { 1919,11}, { 3839,14}, { 511,13}, { 1023,12}, \ + { 2111,13}, { 1151,12}, { 2431,13}, { 1407,12}, \ + { 2943,11}, { 5887,14}, { 767,13}, { 1663,12}, \ + { 3455,13}, { 1919,12}, { 3839,15}, { 511,14}, \ + { 1023,13}, { 2431,14}, { 1279,13}, { 2943,12}, \ + { 5887,14}, { 1535,13}, { 3455,14}, { 1791,13}, \ + { 3839,12}, { 7679,13}, { 3967,12}, { 7935,11}, \ + { 15871,15}, { 1023,14}, { 2047,13}, { 4351,14}, \ + { 2303,13}, { 4991,12}, { 9983,14}, { 2815,13}, \ + { 5887,15}, { 1535,14}, { 3839,13}, { 7935,12}, \ + { 15871,16} } +#define MUL_FFT_TABLE3_SIZE 189 +#define MUL_FFT_THRESHOLD 4736 + +#define SQR_FFT_MODF_THRESHOLD 404 /* k = 5 */ +#define SQR_FFT_TABLE3 \ + { { 404, 5}, { 21, 6}, { 11, 5}, { 23, 6}, \ + { 12, 5}, { 25, 6}, { 13, 5}, { 27, 6}, \ + { 25, 7}, { 13, 6}, { 27, 7}, { 15, 6}, \ + { 31, 7}, { 19, 6}, { 39, 7}, { 23, 6}, \ + { 47, 7}, { 27, 8}, { 15, 7}, { 35, 8}, \ + { 19, 7}, { 41, 8}, { 23, 7}, { 47, 8}, \ + { 27, 9}, { 15, 8}, { 39, 9}, { 23, 8}, \ + { 47,10}, { 15, 9}, { 31, 8}, { 63, 9}, \ + { 39, 8}, { 79, 9}, { 47,10}, { 31, 9}, \ + { 79,10}, { 47,11}, { 31,10}, { 63, 9}, \ + { 127,10}, { 95,11}, { 63,10}, { 127, 9}, \ + { 255, 8}, { 511, 9}, { 271,10}, { 143, 9}, \ + { 287, 8}, { 607, 7}, { 1215,11}, { 95,12}, \ + { 63,11}, { 127,10}, { 255, 9}, { 511,10}, \ + { 271, 9}, { 543, 8}, { 1087, 9}, { 607, 8}, \ + { 1215,11}, { 159, 9}, { 671, 8}, { 1343,10}, \ + { 351, 9}, { 735, 8}, { 1471,11}, { 191,10}, \ + { 383, 9}, { 767,10}, { 415,11}, { 223,12}, \ + { 127,11}, { 255,10}, { 543, 9}, { 1087,10}, \ + { 607, 9}, { 1215, 8}, { 2431,10}, { 671, 9}, \ + { 1343,10}, { 735, 9}, { 1471,12}, { 191,11}, \ + { 383,10}, { 767,11}, { 415,10}, { 831,13}, \ + { 127,12}, { 255,11}, { 543,10}, { 1087,11}, \ + { 607,10}, { 1215, 9}, { 2431,11}, { 671,10}, \ + { 1343,11}, { 735,10}, { 1471, 9}, { 2943,12}, \ + { 383,11}, { 863,12}, { 447,11}, { 959,10}, \ + { 1919,13}, { 255,12}, { 511,11}, { 1087,12}, \ + { 575,11}, { 1215,10}, { 2431,12}, { 639,11}, \ + { 1343,12}, { 703,11}, { 1471,10}, { 2943, 9}, \ + { 5887,12}, { 767,11}, { 1599,12}, { 831,11}, \ + { 1727,12}, { 959,11}, { 1919,10}, { 3839,14}, \ + { 255,13}, { 511,12}, { 1215,11}, { 2431,13}, \ + { 639,12}, { 1471,11}, { 2943,10}, { 5887,13}, \ + { 767,12}, { 1727,13}, { 895,12}, { 1919,11}, \ + { 3839,14}, { 511,13}, { 1023,12}, { 2111,13}, \ + { 1151,12}, { 2431,13}, { 1279,12}, { 2623,13}, \ + { 1407,12}, { 2943,11}, { 5887,14}, { 767,13}, \ + { 1663,12}, { 3455,13}, { 1919,12}, { 3839,15}, \ + { 511,14}, { 1023,13}, { 2431,14}, { 1279,13}, \ + { 2943,12}, { 5887,14}, { 1535,13}, { 3455,14}, \ + { 1791,13}, { 3839,12}, { 7679,13}, { 3967,12}, \ + { 7935,11}, { 15871,15}, { 1023,14}, { 2047,13}, \ + { 4223,14}, { 2303,13}, { 4991,12}, { 9983,14}, \ + { 2815,13}, { 5887,15}, { 1535,14}, { 3839,13}, \ + { 7935,12}, { 15871,16} } +#define SQR_FFT_TABLE3_SIZE 178 +#define SQR_FFT_THRESHOLD 3712 + +#define MULLO_BASECASE_THRESHOLD 4 +#define MULLO_DC_THRESHOLD 62 +#define MULLO_MUL_N_THRESHOLD 8907 +#define SQRLO_BASECASE_THRESHOLD 8 +#define SQRLO_DC_THRESHOLD 107 +#define SQRLO_SQR_THRESHOLD 6633 + +#define DC_DIV_QR_THRESHOLD 54 +#define DC_DIVAPPR_Q_THRESHOLD 206 +#define DC_BDIV_QR_THRESHOLD 55 +#define DC_BDIV_Q_THRESHOLD 136 + +#define INV_MULMOD_BNM1_THRESHOLD 74 +#define INV_NEWTON_THRESHOLD 212 +#define INV_APPR_THRESHOLD 204 + +#define BINV_NEWTON_THRESHOLD 292 +#define REDC_1_TO_REDC_N_THRESHOLD 67 + +#define MU_DIV_QR_THRESHOLD 1442 +#define MU_DIVAPPR_Q_THRESHOLD 1528 +#define MUPI_DIV_QR_THRESHOLD 97 +#define MU_BDIV_QR_THRESHOLD 1142 +#define MU_BDIV_Q_THRESHOLD 1470 + +#define POWM_SEC_TABLE 1,16,96,386,1555 + +#define GET_STR_DC_THRESHOLD 10 +#define GET_STR_PRECOMPUTE_THRESHOLD 16 +#define SET_STR_DC_THRESHOLD 303 +#define SET_STR_PRECOMPUTE_THRESHOLD 748 + +#define FAC_DSC_THRESHOLD 141 +#define FAC_ODD_THRESHOLD 55 + +#define MATRIX22_STRASSEN_THRESHOLD 20 +#define HGCD2_DIV1_METHOD 1 /* 14.03% faster than 3 */ +#define HGCD_THRESHOLD 103 +#define HGCD_APPR_THRESHOLD 127 +#define HGCD_REDUCE_THRESHOLD 3014 +#define GCD_DC_THRESHOLD 396 +#define GCDEXT_DC_THRESHOLD 265 +#define JACOBI_BASE_METHOD 1 /* 47.88% faster than 4 */ + +/* Tuneup completed successfully, took 29014 seconds */ -- cgit v1.2.3