Basic constant folding implementation

9 files changed, 1604 insertions, 0 deletions

diff --git a/vendor/gmp-6.3.0/mpn/x86/atom/sse2/aorsmul_1.asm b/vendor/gmp-6.3.0/mpn/x86/atom/sse2/aorsmul_1.asm
new file mode 100644
index 0000000..969a14a
--- /dev/null
+++ b/vendor/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/vendor/gmp-6.3.0/mpn/x86/atom/sse2/bdiv_dbm1c.asm b/vendor/gmp-6.3.0/mpn/x86/atom/sse2/bdiv_dbm1c.asm
new file mode 100644
index 0000000..782e914
--- /dev/null
+++ b/vendor/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/vendor/gmp-6.3.0/mpn/x86/atom/sse2/divrem_1.asm b/vendor/gmp-6.3.0/mpn/x86/atom/sse2/divrem_1.asm
new file mode 100644
index 0000000..f84709a
--- /dev/null
+++ b/vendor/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/vendor/gmp-6.3.0/mpn/x86/atom/sse2/mod_1_1.asm b/vendor/gmp-6.3.0/mpn/x86/atom/sse2/mod_1_1.asm
new file mode 100644
index 0000000..ae6581d
--- /dev/null
+++ b/vendor/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/vendor/gmp-6.3.0/mpn/x86/atom/sse2/mod_1_4.asm b/vendor/gmp-6.3.0/mpn/x86/atom/sse2/mod_1_4.asm
new file mode 100644
index 0000000..31faa3f
--- /dev/null
+++ b/vendor/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/vendor/gmp-6.3.0/mpn/x86/atom/sse2/mul_1.asm b/vendor/gmp-6.3.0/mpn/x86/atom/sse2/mul_1.asm
new file mode 100644
index 0000000..aa3bb97
--- /dev/null
+++ b/vendor/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/vendor/gmp-6.3.0/mpn/x86/atom/sse2/mul_basecase.asm b/vendor/gmp-6.3.0/mpn/x86/atom/sse2/mul_basecase.asm
new file mode 100644
index 0000000..97d3aeb
--- /dev/null
+++ b/vendor/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/vendor/gmp-6.3.0/mpn/x86/atom/sse2/popcount.asm b/vendor/gmp-6.3.0/mpn/x86/atom/sse2/popcount.asm
new file mode 100644
index 0000000..7847aec
--- /dev/null
+++ b/vendor/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/vendor/gmp-6.3.0/mpn/x86/atom/sse2/sqr_basecase.asm b/vendor/gmp-6.3.0/mpn/x86/atom/sse2/sqr_basecase.asm
new file mode 100644
index 0000000..af19ed8
--- /dev/null
+++ b/vendor/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()