Initial commit.

5 files changed, 1505 insertions, 0 deletions

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<multiplier, this being true of a big calculation
+C done piece by piece.  But let's worry about that if/when mul_1c is
+C actually used.
+
+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)
+
+	movl	PARAM_SIZE, %ecx
+	movl	PARAM_SRC, %edx
+
+	cmpl	$1, %ecx
+	jne	L(two_or_more)
+
+	C one limb only
+
+	movl	PARAM_MULTIPLIER, %eax
+	movl	PARAM_DST, %ecx
+
+	mull	(%edx)
+
+	movl	%eax, (%ecx)
+	movl	%edx, %eax
+
+	ret
+
+
+L(two_or_more):
+	C eax	size
+	C ebx
+	C ecx	carry
+	C edx
+	C esi	src
+	C edi
+	C ebp
+
+	pushl	%esi		FRAME_pushl()
+	pushl	%edi		FRAME_pushl()
+
+	movl	%edx, %esi		C src
+	movl	PARAM_DST, %edi
+
+	movl	PARAM_MULTIPLIER, %eax
+	pushl	%ebx		FRAME_pushl()
+
+	leal	(%esi,%ecx,4), %esi	C src end
+	leal	(%edi,%ecx,4), %edi	C dst end
+
+	negl	%ecx			C -size
+
+	pushl	%ebp		FRAME_pushl()
+	cmpl	$65536, %eax
+
+	jb	L(small)
+
+
+L(big):
+	xorl	%ebx, %ebx		C carry limb
+	sarl	%ecx			C -size/2
+
+	jnc	L(top)			C with carry flag clear
+
+
+	C size was odd, process one limb separately
+
+	mull	4(%esi,%ecx,8)		C m * src[0]
+
+	movl	%eax, 4(%edi,%ecx,8)
+	incl	%ecx
+
+	orl	%edx, %ebx		C carry limb, and clear carry flag
+
+
+L(top):
+	C eax
+	C ebx	carry
+	C ecx	counter, negative
+	C edx
+	C esi	src end
+	C edi	dst end
+	C ebp	(scratch carry)
+
+	adcl	$0, %ebx
+	movl	(%esi,%ecx,8), %eax
+
+	mull	PARAM_MULTIPLIER
+
+	movl	%edx, %ebp
+	addl	%eax, %ebx
+
+	adcl	$0, %ebp
+	movl	4(%esi,%ecx,8), %eax
+
+	mull	PARAM_MULTIPLIER
+
+	movl	%ebx, (%edi,%ecx,8)
+	addl	%ebp, %eax
+
+	movl	%eax, 4(%edi,%ecx,8)
+	incl	%ecx
+
+	movl	%edx, %ebx
+	jnz	L(top)
+
+
+	adcl	$0, %ebx
+	popl	%ebp
+
+	movl	%ebx, %eax
+	popl	%ebx
+
+	popl	%edi
+	popl	%esi
+
+	ret
+
+
+L(small):
+	C Special case for 16-bit multiplier.
+	C
+	C eax	multiplier
+	C ebx
+	C ecx	-size
+	C edx	src
+	C esi	src end
+	C edi	dst end
+	C ebp	multiplier
+
+	C size<3 not supported here.  At size==3 we're already a couple of
+	C cycles faster, so there's no threshold as such, just use the MMX
+	C as soon as possible.
+
+	cmpl	$-3, %ecx
+	ja	L(big)
+
+	movd	%eax, %mm7		C m
+	pxor	%mm6, %mm6		C initial carry word
+
+	punpcklwd %mm7, %mm7		C m replicated 2 times
+	addl	$2, %ecx		C -size+2
+
+	punpckldq %mm7, %mm7		C m replicated 4 times
+	andl	$4, %edx		C test alignment, clear carry flag
+
+	movq	%mm7, %mm0		C m
+	jz	L(small_entry)
+
+
+	C Source is unaligned, process one limb separately.
+	C
+	C Plain integer code is used here, since it's smaller and is about
+	C the same 13 cycles as an mmx block would be.
+	C
+	C An "addl $1,%ecx" doesn't clear the carry flag when size==3, hence
+	C the use of separate incl and orl.
+
+	mull	-8(%esi,%ecx,4)		C m * src[0]
+
+	movl	%eax, -8(%edi,%ecx,4)	C dst[0]
+	incl	%ecx			C one limb processed
+
+	movd	%edx, %mm6		C initial carry
+
+	orl	%eax, %eax		C clear carry flag
+	jmp	L(small_entry)
+
+
+C The scheduling here is quite tricky, since so many instructions have
+C pairing restrictions.  In particular the js won't pair with a movd, and
+C can't be paired with an adc since it wants flags from the inc, so
+C instructions are rotated to the top of the loop to find somewhere useful
+C for it.
+C
+C Trouble has been taken to avoid overlapping successive loop iterations,
+C since that would greatly increase the size of the startup and finishup
+C code.  Actually there's probably not much advantage to be had from
+C overlapping anyway, since the difficulties are mostly with pairing, not
+C with latencies as such.
+C
+C In the comments x represents the src data and m the multiplier (16
+C bits, but replicated 4 times).
+C
+C The m signs calculated in %mm3 are a loop invariant and could be held in
+C say %mm5, but that would save only one instruction and hence be no faster.
+
+L(small_top):
+	C eax	l.low, then l.high
+	C ebx	(h.low)
+	C ecx	counter, -size+2 to 0 or 1
+	C edx	(h.high)
+	C esi	&src[size]
+	C edi	&dst[size]
+	C ebp
+	C
+	C %mm0	(high products)
+	C %mm1	(low products)
+	C %mm2	(adjust for m using x signs)
+	C %mm3	(adjust for x using m signs)
+	C %mm4
+	C %mm5
+	C %mm6	h.low, then carry
+	C %mm7	m replicated 4 times
+
+	movd	%mm6, %ebx		C h.low
+	psrlq	$32, %mm1		C l.high
+
+	movd	%mm0, %edx		C h.high
+	movq	%mm0, %mm6		C new c
+
+	adcl	%eax, %ebx
+	incl	%ecx
+
+	movd	%mm1, %eax		C l.high
+	movq	%mm7, %mm0
+
+	adcl	%eax, %edx
+	movl	%ebx, -16(%edi,%ecx,4)
+
+	movl	%edx, -12(%edi,%ecx,4)
+	psrlq	$32, %mm6		C c
+
+L(small_entry):
+	pmulhw	-8(%esi,%ecx,4), %mm0	C h = (x*m).high
+	movq	%mm7, %mm1
+
+	pmullw	-8(%esi,%ecx,4), %mm1	C l = (x*m).low
+	movq	%mm7, %mm3
+
+	movq	-8(%esi,%ecx,4), %mm2	C x
+	psraw	$15, %mm3		C m signs
+
+	pand	-8(%esi,%ecx,4), %mm3	C x selected by m signs
+	psraw	$15, %mm2		C x signs
+
+	paddw	%mm3, %mm0		C add x to h if m neg
+	pand	%mm7, %mm2		C m selected by x signs
+
+	paddw	%mm2, %mm0		C add m to h if x neg
+	incl	%ecx
+
+	movd	%mm1, %eax		C l.low
+	punpcklwd %mm0, %mm6		C c + h.low << 16
+
+	psrlq	$16, %mm0		C h.high
+	js	L(small_top)
+
+
+
+
+	movd	%mm6, %ebx		C h.low
+	psrlq	$32, %mm1		C l.high
+
+	adcl	%eax, %ebx
+	popl	%ebp		FRAME_popl()
+
+	movd	%mm0, %edx		C h.high
+	psrlq	$32, %mm0		C l.high
+
+	movd	%mm1, %eax		C l.high
+
+	adcl	%eax, %edx
+	movl	%ebx, -12(%edi,%ecx,4)
+
+	movd	%mm0, %eax		C c
+
+	adcl	$0, %eax
+	movl	%edx, -8(%edi,%ecx,4)
+
+	orl	%ecx, %ecx
+	jnz	L(small_done)		C final %ecx==1 means even, ==0 odd
+
+
+	C Size odd, one extra limb to process.
+	C Plain integer code is used here, since it's smaller and is about
+	C the same speed as another mmx block would be.
+
+	movl	%eax, %ecx
+	movl	PARAM_MULTIPLIER, %eax
+
+	mull	-4(%esi)
+
+	addl	%ecx, %eax
+
+	adcl	$0, %edx
+	movl	%eax, -4(%edi)
+
+	movl	%edx, %eax
+L(small_done):
+	popl	%ebx
+
+	popl	%edi
+	popl	%esi
+
+	emms
+
+	ret
+
+EPILOGUE()
diff --git a/gmp-6.3.0/mpn/x86/pentium/mmx/rshift.asm b/gmp-6.3.0/mpn/x86/pentium/mmx/rshift.asm
new file mode 100644
index 0000000..e3b274b
--- /dev/null
+++ b/gmp-6.3.0/mpn/x86/pentium/mmx/rshift.asm
@@ -0,0 +1,468 @@
+dnl  Intel P5 mpn_rshift -- mpn right 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_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.  Return the bits shifted out at the
+C right.
+C
+C It takes 6 mmx instructions to process 2 limbs, making 1.5 cycles/limb,
+C and with a 4 limb loop and 1 cycle of loop overhead the total is 1.75 c/l.
+C
+C Full speed depends on source and destination being aligned.  Unaligned mmx
+C loads and stores on P5 don't pair and have a 2 cycle penalty.  Some hairy
+C setups and finish-ups are done to ensure alignment for the loop.
+C
+C MMX shifts work out a bit faster even for the simple loop.
+
+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_rshift)
+
+	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)
+
+	decl	%eax
+	movl	(%ebx), %edi		C src low limb
+
+	jnz	L(simple)
+
+	shrdl(	%cl, %edi, %eax)	C eax was decremented to zero
+
+	shrl	%cl, %edi
+
+	movl	%edi, (%edx)		C dst low limb
+	popl	%edi			C risk of data cache bank clash
+
+	popl	%ebx
+
+	ret
+
+
+C -----------------------------------------------------------------------------
+	ALIGN(8)
+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), %mm5		C src[0]
+	leal	(%ebx,%eax,4), %ebx	C &src[size-1]
+
+	movd	%ecx, %mm6		C rshift
+	leal	-4(%edx,%eax,4), %edx	C &dst[size-2]
+
+	psllq	$32, %mm5
+	negl	%eax
+
+
+C This loop is 5 or 8 cycles, with every second load unaligned and a wasted
+C cycle waiting for the mm0 result to be ready.  For comparison a shrdl is 4
+C cycles and would be 8 in a simple loop.  Using mmx helps the return value
+C and last limb calculations too.
+
+L(simple_top):
+	C eax	counter, limbs, negative
+	C ebx	&src[size-1]
+	C ecx	return value
+	C edx	&dst[size-2]
+	C
+	C mm0	scratch
+	C mm5	return value
+	C mm6	shift
+
+	movq	(%ebx,%eax,4), %mm0
+	incl	%eax
+
+	psrlq	%mm6, %mm0
+
+	movd	%mm0, (%edx,%eax,4)
+	jnz	L(simple_top)
+
+
+	movd	(%ebx), %mm0
+	psrlq	%mm6, %mm5		C return value
+
+	psrlq	%mm6, %mm0
+	popl	%edi
+
+	movd	%mm5, %eax
+	popl	%ebx
+
+	movd	%mm0, 4(%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	(%ebx), %mm5		C src[0]
+	movl	$4, %edi
+
+	movd	%ecx, %mm6		C rshift
+	testl	%edi, %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                  ebx
+	C --+-------+-------+-------+
+	C           |          xxx  |
+	C --+-------+-------+-------+
+	C         4mod8   0mod8   4mod8
+	C
+	C         dest            edx
+	C         --+-------+-------+
+	C           |       |  xxx  |
+	C         --+-------+-------+
+
+	movq	(%ebx), %mm0		C unaligned load
+
+	psrlq	%mm6, %mm0
+	addl	$4, %ebx
+
+	decl	%eax
+
+	movd	%mm0, (%edx)
+	addl	$4, %edx
+L(start_src_aligned):
+
+
+	movq	(%ebx), %mm1
+	testl	%edi, %edx
+
+	psrlq	%mm6, %mm5		C retval
+	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
+	C separately.
+	C
+	C          source          ebx
+	C          --+-------+-------+
+	C            |      mm1      |
+	C          --+-------+-------+
+	C                  4mod8   0mod8
+	C
+	C  dest                    edx
+	C  --+-------+-------+-------+
+	C                    |  xxx  |
+	C  --+-------+-------+-------+
+	C          4mod8   0mod8   4mod8
+
+	movq	%mm1, %mm0
+	addl	$32, %ecx		C new shift
+
+	psrlq	%mm6, %mm0
+
+	movd	%ecx, %mm6
+
+	movd	%mm0, (%edx)
+	addl	$4, %edx
+L(start_dst_aligned):
+
+
+	movq	8(%ebx), %mm3
+	negl	%ecx
+
+	movq	%mm3, %mm2		C mm2 src qword
+	addl	$64, %ecx
+
+	movd	%ecx, %mm7
+	psrlq	%mm6, %mm1
+
+	leal	-12(%ebx,%eax,4), %ebx
+	leal	-20(%edx,%eax,4), %edx
+
+	psllq	%mm7, %mm3
+	subl	$7, %eax		C size-7
+
+	por	%mm1, %mm3		C mm3 ready to store
+	negl	%eax			C -(size-7)
+
+	jns	L(finish)
+
+
+	C This loop is the important bit, the rest is just support.  Careful
+	C instruction scheduling achieves the claimed 1.75 c/l.  The
+	C relevant parts of the pairing rules are:
+	C
+	C - mmx loads and stores execute only in the U pipe
+	C - only one mmx shift in a pair
+	C - wait one cycle before storing an mmx register result
+	C - the usual address generation interlock
+	C
+	C Two qword calculations are slightly interleaved.  The instructions
+	C marked "C" belong to the second qword, and the "C prev" one is for
+	C the second qword from the previous iteration.
+
+	ALIGN(8)
+L(unroll_loop):
+	C eax	counter, limbs, negative
+	C ebx	&src[size-12]
+	C ecx
+	C edx	&dst[size-12]
+	C esi
+	C edi
+	C
+	C mm0
+	C mm1
+	C mm2	src qword from -8(%ebx,%eax,4)
+	C mm3	dst qword ready to store to -8(%edx,%eax,4)
+	C
+	C mm5	return value
+	C mm6	rshift
+	C mm7	lshift
+
+	movq	(%ebx,%eax,4), %mm0
+	psrlq	%mm6, %mm2
+
+	movq	%mm0, %mm1
+	psllq	%mm7, %mm0
+
+	movq	%mm3, -8(%edx,%eax,4)	C prev
+	por	%mm2, %mm0
+
+	movq	8(%ebx,%eax,4), %mm3	C
+	psrlq	%mm6, %mm1		C
+
+	movq	%mm0, (%edx,%eax,4)
+	movq	%mm3, %mm2		C
+
+	psllq	%mm7, %mm3		C
+	addl	$4, %eax
+
+	por	%mm1, %mm3		C
+	js	L(unroll_loop)
+
+
+L(finish):
+	C eax	0 to 3 representing respectively 3 to 0 limbs remaining
+
+	testb	$2, %al
+
+	jnz	L(finish_no_two)
+
+	movq	(%ebx,%eax,4), %mm0
+	psrlq	%mm6, %mm2
+
+	movq	%mm0, %mm1
+	psllq	%mm7, %mm0
+
+	movq	%mm3, -8(%edx,%eax,4)	C prev
+	por	%mm2, %mm0
+
+	movq	%mm1, %mm2
+	movq	%mm0, %mm3
+
+	addl	$2, %eax
+L(finish_no_two):
+
+
+	C eax	2 or 3 representing respectively 1 or 0 limbs remaining
+	C
+	C mm2	src prev qword, from -8(%ebx,%eax,4)
+	C mm3	dst qword, for -8(%edx,%eax,4)
+
+	testb	$1, %al
+	popl	%edi
+
+	movd	%mm5, %eax	C retval
+	jnz	L(finish_zero)
+
+
+	C One extra limb, destination was aligned.
+	C
+	C source                ebx
+	C +-------+---------------+--
+	C |       |      mm2      |
+	C +-------+---------------+--
+	C
+	C dest                                  edx
+	C +-------+---------------+---------------+--
+	C |       |               |      mm3      |
+	C +-------+---------------+---------------+--
+	C
+	C mm6 = shift
+	C mm7 = ecx = 64-shift
+
+
+	C One extra limb, destination was unaligned.
+	C
+	C source                ebx
+	C +-------+---------------+--
+	C |       |      mm2      |
+	C +-------+---------------+--
+	C
+	C dest                          edx
+	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 8(%edx), and in the aligned case
+	C there's a further extra limb of dst to be formed.
+
+
+	movd	8(%ebx), %mm0
+	psrlq	%mm6, %mm2
+
+	movq	%mm0, %mm1
+	psllq	%mm7, %mm0
+
+	movq	%mm3, (%edx)
+	por	%mm2, %mm0
+
+	psrlq	%mm6, %mm1
+	andl	$32, %ecx
+
+	popl	%ebx
+	jz	L(finish_one_unaligned)
+
+	C dst was aligned, must store one extra limb
+	movd	%mm1, 16(%edx)
+L(finish_one_unaligned):
+
+	movq	%mm0, 8(%edx)
+
+	emms
+
+	ret
+
+
+L(finish_zero):
+
+	C No extra limbs, destination was aligned.
+	C
+	C source        ebx
+	C +---------------+--
+	C |      mm2      |
+	C +---------------+--
+	C
+	C dest                        edx+4
+	C +---------------+---------------+--
+	C |               |      mm3      |
+	C +---------------+---------------+--
+	C
+	C mm6 = shift
+	C mm7 = ecx = 64-shift
+
+
+	C No extra limbs, destination was unaligned.
+	C
+	C source        ebx
+	C +---------------+--
+	C |      mm2      |
+	C +---------------+--
+	C
+	C dest                edx+4
+	C +-------+---------------+--
+	C |       |      mm3      |
+	C +-------+---------------+--
+	C
+	C mm6 = shift+32
+	C mm7 = 64-(shift+32)
+
+
+	C The movd for the unaligned case is clearly the same data as the
+	C movq for the aligned case, it's just a choice between whether one
+	C or two limbs should be written.
+
+
+	movq	%mm3, 4(%edx)
+	psrlq	%mm6, %mm2
+
+	movd	%mm2, 12(%edx)
+	andl	$32, %ecx
+
+	popl	%ebx
+	jz	L(finish_zero_unaligned)
+
+	movq	%mm2, 12(%edx)
+L(finish_zero_unaligned):
+
+	emms
+
+	ret
+
+EPILOGUE()