4 files changed, 3592 insertions, 0 deletions

diff --git a/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/bn-586.s b/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/bn-586.s
new file mode 100644
index 0000000000..80c8db4d29
--- /dev/null
+++ b/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/bn-586.s
@@ -0,0 +1,1525 @@
+.text
+.globl	_bn_mul_add_words
+.type	_bn_mul_add_words,@function
+.align	4
+_bn_mul_add_words:
+L_bn_mul_add_words_begin:
+	call	L000PIC_me_up
+L000PIC_me_up:
+	popl	%eax
+	leal	__GLOBAL_OFFSET_TABLE_+[.-L000PIC_me_up](%eax),%eax
+	movl	_OPENSSL_ia32cap_P@GOT(%eax),%eax
+	btl	$26,(%eax)
+	jnc	L001maw_non_sse2
+	movl	4(%esp),%eax
+	movl	8(%esp),%edx
+	movl	12(%esp),%ecx
+	movd	16(%esp),%mm0
+	pxor	%mm1,%mm1
+	jmp	L002maw_sse2_entry
+.align	4,0x90
+L003maw_sse2_unrolled:
+	movd	(%eax),%mm3
+	paddq	%mm3,%mm1
+	movd	(%edx),%mm2
+	pmuludq	%mm0,%mm2
+	movd	4(%edx),%mm4
+	pmuludq	%mm0,%mm4
+	movd	8(%edx),%mm6
+	pmuludq	%mm0,%mm6
+	movd	12(%edx),%mm7
+	pmuludq	%mm0,%mm7
+	paddq	%mm2,%mm1
+	movd	4(%eax),%mm3
+	paddq	%mm4,%mm3
+	movd	8(%eax),%mm5
+	paddq	%mm6,%mm5
+	movd	12(%eax),%mm4
+	paddq	%mm4,%mm7
+	movd	%mm1,(%eax)
+	movd	16(%edx),%mm2
+	pmuludq	%mm0,%mm2
+	psrlq	$32,%mm1
+	movd	20(%edx),%mm4
+	pmuludq	%mm0,%mm4
+	paddq	%mm3,%mm1
+	movd	24(%edx),%mm6
+	pmuludq	%mm0,%mm6
+	movd	%mm1,4(%eax)
+	psrlq	$32,%mm1
+	movd	28(%edx),%mm3
+	addl	$32,%edx
+	pmuludq	%mm0,%mm3
+	paddq	%mm5,%mm1
+	movd	16(%eax),%mm5
+	paddq	%mm5,%mm2
+	movd	%mm1,8(%eax)
+	psrlq	$32,%mm1
+	paddq	%mm7,%mm1
+	movd	20(%eax),%mm5
+	paddq	%mm5,%mm4
+	movd	%mm1,12(%eax)
+	psrlq	$32,%mm1
+	paddq	%mm2,%mm1
+	movd	24(%eax),%mm5
+	paddq	%mm5,%mm6
+	movd	%mm1,16(%eax)
+	psrlq	$32,%mm1
+	paddq	%mm4,%mm1
+	movd	28(%eax),%mm5
+	paddq	%mm5,%mm3
+	movd	%mm1,20(%eax)
+	psrlq	$32,%mm1
+	paddq	%mm6,%mm1
+	movd	%mm1,24(%eax)
+	psrlq	$32,%mm1
+	paddq	%mm3,%mm1
+	movd	%mm1,28(%eax)
+	leal	32(%eax),%eax
+	psrlq	$32,%mm1
+	subl	$8,%ecx
+	jz	L004maw_sse2_exit
+L002maw_sse2_entry:
+	testl	$4294967288,%ecx
+	jnz	L003maw_sse2_unrolled
+.align	2,0x90
+L005maw_sse2_loop:
+	movd	(%edx),%mm2
+	movd	(%eax),%mm3
+	pmuludq	%mm0,%mm2
+	leal	4(%edx),%edx
+	paddq	%mm3,%mm1
+	paddq	%mm2,%mm1
+	movd	%mm1,(%eax)
+	subl	$1,%ecx
+	psrlq	$32,%mm1
+	leal	4(%eax),%eax
+	jnz	L005maw_sse2_loop
+L004maw_sse2_exit:
+	movd	%mm1,%eax
+	emms
+	ret
+.align	4,0x90
+L001maw_non_sse2:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+	xorl	%esi,%esi
+	movl	20(%esp),%edi
+	movl	28(%esp),%ecx
+	movl	24(%esp),%ebx
+	andl	$4294967288,%ecx
+	movl	32(%esp),%ebp
+	pushl	%ecx
+	jz	L006maw_finish
+.align	4,0x90
+L007maw_loop:
+	# Round 0
+	movl	(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	(%edi),%eax
+	adcl	$0,%edx
+	movl	%eax,(%edi)
+	movl	%edx,%esi
+	# Round 4
+	movl	4(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	4(%edi),%eax
+	adcl	$0,%edx
+	movl	%eax,4(%edi)
+	movl	%edx,%esi
+	# Round 8
+	movl	8(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	8(%edi),%eax
+	adcl	$0,%edx
+	movl	%eax,8(%edi)
+	movl	%edx,%esi
+	# Round 12
+	movl	12(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	12(%edi),%eax
+	adcl	$0,%edx
+	movl	%eax,12(%edi)
+	movl	%edx,%esi
+	# Round 16
+	movl	16(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	16(%edi),%eax
+	adcl	$0,%edx
+	movl	%eax,16(%edi)
+	movl	%edx,%esi
+	# Round 20
+	movl	20(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	20(%edi),%eax
+	adcl	$0,%edx
+	movl	%eax,20(%edi)
+	movl	%edx,%esi
+	# Round 24
+	movl	24(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	24(%edi),%eax
+	adcl	$0,%edx
+	movl	%eax,24(%edi)
+	movl	%edx,%esi
+	# Round 28
+	movl	28(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	28(%edi),%eax
+	adcl	$0,%edx
+	movl	%eax,28(%edi)
+	movl	%edx,%esi
+
+	subl	$8,%ecx
+	leal	32(%ebx),%ebx
+	leal	32(%edi),%edi
+	jnz	L007maw_loop
+L006maw_finish:
+	movl	32(%esp),%ecx
+	andl	$7,%ecx
+	jnz	L008maw_finish2
+	jmp	L009maw_end
+L008maw_finish2:
+	# Tail Round 0
+	movl	(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	(%edi),%eax
+	adcl	$0,%edx
+	decl	%ecx
+	movl	%eax,(%edi)
+	movl	%edx,%esi
+	jz	L009maw_end
+	# Tail Round 1
+	movl	4(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	4(%edi),%eax
+	adcl	$0,%edx
+	decl	%ecx
+	movl	%eax,4(%edi)
+	movl	%edx,%esi
+	jz	L009maw_end
+	# Tail Round 2
+	movl	8(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	8(%edi),%eax
+	adcl	$0,%edx
+	decl	%ecx
+	movl	%eax,8(%edi)
+	movl	%edx,%esi
+	jz	L009maw_end
+	# Tail Round 3
+	movl	12(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	12(%edi),%eax
+	adcl	$0,%edx
+	decl	%ecx
+	movl	%eax,12(%edi)
+	movl	%edx,%esi
+	jz	L009maw_end
+	# Tail Round 4
+	movl	16(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	16(%edi),%eax
+	adcl	$0,%edx
+	decl	%ecx
+	movl	%eax,16(%edi)
+	movl	%edx,%esi
+	jz	L009maw_end
+	# Tail Round 5
+	movl	20(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	20(%edi),%eax
+	adcl	$0,%edx
+	decl	%ecx
+	movl	%eax,20(%edi)
+	movl	%edx,%esi
+	jz	L009maw_end
+	# Tail Round 6
+	movl	24(%ebx),%eax
+	mull	%ebp
+	addl	%esi,%eax
+	adcl	$0,%edx
+	addl	24(%edi),%eax
+	adcl	$0,%edx
+	movl	%eax,24(%edi)
+	movl	%edx,%esi
+L009maw_end:
+	movl	%esi,%eax
+	popl	%ecx
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_bn_mul_words
+.type	_bn_mul_words,@function
+.align	4
+_bn_mul_words:
+L_bn_mul_words_begin:
+	call	L010PIC_me_up
+L010PIC_me_up:
+	popl	%eax
+	leal	__GLOBAL_OFFSET_TABLE_+[.-L010PIC_me_up](%eax),%eax
+	movl	_OPENSSL_ia32cap_P@GOT(%eax),%eax
+	btl	$26,(%eax)
+	jnc	L011mw_non_sse2
+	movl	4(%esp),%eax
+	movl	8(%esp),%edx
+	movl	12(%esp),%ecx
+	movd	16(%esp),%mm0
+	pxor	%mm1,%mm1
+.align	4,0x90
+L012mw_sse2_loop:
+	movd	(%edx),%mm2
+	pmuludq	%mm0,%mm2
+	leal	4(%edx),%edx
+	paddq	%mm2,%mm1
+	movd	%mm1,(%eax)
+	subl	$1,%ecx
+	psrlq	$32,%mm1
+	leal	4(%eax),%eax
+	jnz	L012mw_sse2_loop
+	movd	%mm1,%eax
+	emms
+	ret
+.align	4,0x90
+L011mw_non_sse2:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+	xorl	%esi,%esi
+	movl	20(%esp),%edi
+	movl	24(%esp),%ebx
+	movl	28(%esp),%ebp
+	movl	32(%esp),%ecx
+	andl	$4294967288,%ebp
+	jz	L013mw_finish
+L014mw_loop:
+	# Round 0
+	movl	(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,(%edi)
+	movl	%edx,%esi
+	# Round 4
+	movl	4(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,4(%edi)
+	movl	%edx,%esi
+	# Round 8
+	movl	8(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,8(%edi)
+	movl	%edx,%esi
+	# Round 12
+	movl	12(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,12(%edi)
+	movl	%edx,%esi
+	# Round 16
+	movl	16(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,16(%edi)
+	movl	%edx,%esi
+	# Round 20
+	movl	20(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,20(%edi)
+	movl	%edx,%esi
+	# Round 24
+	movl	24(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,24(%edi)
+	movl	%edx,%esi
+	# Round 28
+	movl	28(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,28(%edi)
+	movl	%edx,%esi
+
+	addl	$32,%ebx
+	addl	$32,%edi
+	subl	$8,%ebp
+	jz	L013mw_finish
+	jmp	L014mw_loop
+L013mw_finish:
+	movl	28(%esp),%ebp
+	andl	$7,%ebp
+	jnz	L015mw_finish2
+	jmp	L016mw_end
+L015mw_finish2:
+	# Tail Round 0
+	movl	(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,(%edi)
+	movl	%edx,%esi
+	decl	%ebp
+	jz	L016mw_end
+	# Tail Round 1
+	movl	4(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,4(%edi)
+	movl	%edx,%esi
+	decl	%ebp
+	jz	L016mw_end
+	# Tail Round 2
+	movl	8(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,8(%edi)
+	movl	%edx,%esi
+	decl	%ebp
+	jz	L016mw_end
+	# Tail Round 3
+	movl	12(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,12(%edi)
+	movl	%edx,%esi
+	decl	%ebp
+	jz	L016mw_end
+	# Tail Round 4
+	movl	16(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,16(%edi)
+	movl	%edx,%esi
+	decl	%ebp
+	jz	L016mw_end
+	# Tail Round 5
+	movl	20(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,20(%edi)
+	movl	%edx,%esi
+	decl	%ebp
+	jz	L016mw_end
+	# Tail Round 6
+	movl	24(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,24(%edi)
+	movl	%edx,%esi
+L016mw_end:
+	movl	%esi,%eax
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_bn_sqr_words
+.type	_bn_sqr_words,@function
+.align	4
+_bn_sqr_words:
+L_bn_sqr_words_begin:
+	call	L017PIC_me_up
+L017PIC_me_up:
+	popl	%eax
+	leal	__GLOBAL_OFFSET_TABLE_+[.-L017PIC_me_up](%eax),%eax
+	movl	_OPENSSL_ia32cap_P@GOT(%eax),%eax
+	btl	$26,(%eax)
+	jnc	L018sqr_non_sse2
+	movl	4(%esp),%eax
+	movl	8(%esp),%edx
+	movl	12(%esp),%ecx
+.align	4,0x90
+L019sqr_sse2_loop:
+	movd	(%edx),%mm0
+	pmuludq	%mm0,%mm0
+	leal	4(%edx),%edx
+	movq	%mm0,(%eax)
+	subl	$1,%ecx
+	leal	8(%eax),%eax
+	jnz	L019sqr_sse2_loop
+	emms
+	ret
+.align	4,0x90
+L018sqr_non_sse2:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+	movl	20(%esp),%esi
+	movl	24(%esp),%edi
+	movl	28(%esp),%ebx
+	andl	$4294967288,%ebx
+	jz	L020sw_finish
+L021sw_loop:
+	# Round 0
+	movl	(%edi),%eax
+	mull	%eax
+	movl	%eax,(%esi)
+	movl	%edx,4(%esi)
+	# Round 4
+	movl	4(%edi),%eax
+	mull	%eax
+	movl	%eax,8(%esi)
+	movl	%edx,12(%esi)
+	# Round 8
+	movl	8(%edi),%eax
+	mull	%eax
+	movl	%eax,16(%esi)
+	movl	%edx,20(%esi)
+	# Round 12
+	movl	12(%edi),%eax
+	mull	%eax
+	movl	%eax,24(%esi)
+	movl	%edx,28(%esi)
+	# Round 16
+	movl	16(%edi),%eax
+	mull	%eax
+	movl	%eax,32(%esi)
+	movl	%edx,36(%esi)
+	# Round 20
+	movl	20(%edi),%eax
+	mull	%eax
+	movl	%eax,40(%esi)
+	movl	%edx,44(%esi)
+	# Round 24
+	movl	24(%edi),%eax
+	mull	%eax
+	movl	%eax,48(%esi)
+	movl	%edx,52(%esi)
+	# Round 28
+	movl	28(%edi),%eax
+	mull	%eax
+	movl	%eax,56(%esi)
+	movl	%edx,60(%esi)
+
+	addl	$32,%edi
+	addl	$64,%esi
+	subl	$8,%ebx
+	jnz	L021sw_loop
+L020sw_finish:
+	movl	28(%esp),%ebx
+	andl	$7,%ebx
+	jz	L022sw_end
+	# Tail Round 0
+	movl	(%edi),%eax
+	mull	%eax
+	movl	%eax,(%esi)
+	decl	%ebx
+	movl	%edx,4(%esi)
+	jz	L022sw_end
+	# Tail Round 1
+	movl	4(%edi),%eax
+	mull	%eax
+	movl	%eax,8(%esi)
+	decl	%ebx
+	movl	%edx,12(%esi)
+	jz	L022sw_end
+	# Tail Round 2
+	movl	8(%edi),%eax
+	mull	%eax
+	movl	%eax,16(%esi)
+	decl	%ebx
+	movl	%edx,20(%esi)
+	jz	L022sw_end
+	# Tail Round 3
+	movl	12(%edi),%eax
+	mull	%eax
+	movl	%eax,24(%esi)
+	decl	%ebx
+	movl	%edx,28(%esi)
+	jz	L022sw_end
+	# Tail Round 4
+	movl	16(%edi),%eax
+	mull	%eax
+	movl	%eax,32(%esi)
+	decl	%ebx
+	movl	%edx,36(%esi)
+	jz	L022sw_end
+	# Tail Round 5
+	movl	20(%edi),%eax
+	mull	%eax
+	movl	%eax,40(%esi)
+	decl	%ebx
+	movl	%edx,44(%esi)
+	jz	L022sw_end
+	# Tail Round 6
+	movl	24(%edi),%eax
+	mull	%eax
+	movl	%eax,48(%esi)
+	movl	%edx,52(%esi)
+L022sw_end:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_bn_div_words
+.type	_bn_div_words,@function
+.align	4
+_bn_div_words:
+L_bn_div_words_begin:
+	movl	4(%esp),%edx
+	movl	8(%esp),%eax
+	movl	12(%esp),%ecx
+	divl	%ecx
+	ret
+.globl	_bn_add_words
+.type	_bn_add_words,@function
+.align	4
+_bn_add_words:
+L_bn_add_words_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+	movl	20(%esp),%ebx
+	movl	24(%esp),%esi
+	movl	28(%esp),%edi
+	movl	32(%esp),%ebp
+	xorl	%eax,%eax
+	andl	$4294967288,%ebp
+	jz	L023aw_finish
+L024aw_loop:
+	# Round 0
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	# Round 1
+	movl	4(%esi),%ecx
+	movl	4(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,4(%ebx)
+	# Round 2
+	movl	8(%esi),%ecx
+	movl	8(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,8(%ebx)
+	# Round 3
+	movl	12(%esi),%ecx
+	movl	12(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,12(%ebx)
+	# Round 4
+	movl	16(%esi),%ecx
+	movl	16(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,16(%ebx)
+	# Round 5
+	movl	20(%esi),%ecx
+	movl	20(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,20(%ebx)
+	# Round 6
+	movl	24(%esi),%ecx
+	movl	24(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,24(%ebx)
+	# Round 7
+	movl	28(%esi),%ecx
+	movl	28(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,28(%ebx)
+
+	addl	$32,%esi
+	addl	$32,%edi
+	addl	$32,%ebx
+	subl	$8,%ebp
+	jnz	L024aw_loop
+L023aw_finish:
+	movl	32(%esp),%ebp
+	andl	$7,%ebp
+	jz	L025aw_end
+	# Tail Round 0
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,(%ebx)
+	jz	L025aw_end
+	# Tail Round 1
+	movl	4(%esi),%ecx
+	movl	4(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,4(%ebx)
+	jz	L025aw_end
+	# Tail Round 2
+	movl	8(%esi),%ecx
+	movl	8(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,8(%ebx)
+	jz	L025aw_end
+	# Tail Round 3
+	movl	12(%esi),%ecx
+	movl	12(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,12(%ebx)
+	jz	L025aw_end
+	# Tail Round 4
+	movl	16(%esi),%ecx
+	movl	16(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,16(%ebx)
+	jz	L025aw_end
+	# Tail Round 5
+	movl	20(%esi),%ecx
+	movl	20(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,20(%ebx)
+	jz	L025aw_end
+	# Tail Round 6
+	movl	24(%esi),%ecx
+	movl	24(%edi),%edx
+	addl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	addl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,24(%ebx)
+L025aw_end:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_bn_sub_words
+.type	_bn_sub_words,@function
+.align	4
+_bn_sub_words:
+L_bn_sub_words_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+	movl	20(%esp),%ebx
+	movl	24(%esp),%esi
+	movl	28(%esp),%edi
+	movl	32(%esp),%ebp
+	xorl	%eax,%eax
+	andl	$4294967288,%ebp
+	jz	L026aw_finish
+L027aw_loop:
+	# Round 0
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	# Round 1
+	movl	4(%esi),%ecx
+	movl	4(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,4(%ebx)
+	# Round 2
+	movl	8(%esi),%ecx
+	movl	8(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,8(%ebx)
+	# Round 3
+	movl	12(%esi),%ecx
+	movl	12(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,12(%ebx)
+	# Round 4
+	movl	16(%esi),%ecx
+	movl	16(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,16(%ebx)
+	# Round 5
+	movl	20(%esi),%ecx
+	movl	20(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,20(%ebx)
+	# Round 6
+	movl	24(%esi),%ecx
+	movl	24(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,24(%ebx)
+	# Round 7
+	movl	28(%esi),%ecx
+	movl	28(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,28(%ebx)
+
+	addl	$32,%esi
+	addl	$32,%edi
+	addl	$32,%ebx
+	subl	$8,%ebp
+	jnz	L027aw_loop
+L026aw_finish:
+	movl	32(%esp),%ebp
+	andl	$7,%ebp
+	jz	L028aw_end
+	# Tail Round 0
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,(%ebx)
+	jz	L028aw_end
+	# Tail Round 1
+	movl	4(%esi),%ecx
+	movl	4(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,4(%ebx)
+	jz	L028aw_end
+	# Tail Round 2
+	movl	8(%esi),%ecx
+	movl	8(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,8(%ebx)
+	jz	L028aw_end
+	# Tail Round 3
+	movl	12(%esi),%ecx
+	movl	12(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,12(%ebx)
+	jz	L028aw_end
+	# Tail Round 4
+	movl	16(%esi),%ecx
+	movl	16(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,16(%ebx)
+	jz	L028aw_end
+	# Tail Round 5
+	movl	20(%esi),%ecx
+	movl	20(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,20(%ebx)
+	jz	L028aw_end
+	# Tail Round 6
+	movl	24(%esi),%ecx
+	movl	24(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,24(%ebx)
+L028aw_end:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_bn_sub_part_words
+.type	_bn_sub_part_words,@function
+.align	4
+_bn_sub_part_words:
+L_bn_sub_part_words_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+	movl	20(%esp),%ebx
+	movl	24(%esp),%esi
+	movl	28(%esp),%edi
+	movl	32(%esp),%ebp
+	xorl	%eax,%eax
+	andl	$4294967288,%ebp
+	jz	L029aw_finish
+L030aw_loop:
+	# Round 0
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	# Round 1
+	movl	4(%esi),%ecx
+	movl	4(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,4(%ebx)
+	# Round 2
+	movl	8(%esi),%ecx
+	movl	8(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,8(%ebx)
+	# Round 3
+	movl	12(%esi),%ecx
+	movl	12(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,12(%ebx)
+	# Round 4
+	movl	16(%esi),%ecx
+	movl	16(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,16(%ebx)
+	# Round 5
+	movl	20(%esi),%ecx
+	movl	20(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,20(%ebx)
+	# Round 6
+	movl	24(%esi),%ecx
+	movl	24(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,24(%ebx)
+	# Round 7
+	movl	28(%esi),%ecx
+	movl	28(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,28(%ebx)
+
+	addl	$32,%esi
+	addl	$32,%edi
+	addl	$32,%ebx
+	subl	$8,%ebp
+	jnz	L030aw_loop
+L029aw_finish:
+	movl	32(%esp),%ebp
+	andl	$7,%ebp
+	jz	L031aw_end
+	# Tail Round 0
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	addl	$4,%esi
+	addl	$4,%edi
+	addl	$4,%ebx
+	decl	%ebp
+	jz	L031aw_end
+	# Tail Round 1
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	addl	$4,%esi
+	addl	$4,%edi
+	addl	$4,%ebx
+	decl	%ebp
+	jz	L031aw_end
+	# Tail Round 2
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	addl	$4,%esi
+	addl	$4,%edi
+	addl	$4,%ebx
+	decl	%ebp
+	jz	L031aw_end
+	# Tail Round 3
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	addl	$4,%esi
+	addl	$4,%edi
+	addl	$4,%ebx
+	decl	%ebp
+	jz	L031aw_end
+	# Tail Round 4
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	addl	$4,%esi
+	addl	$4,%edi
+	addl	$4,%ebx
+	decl	%ebp
+	jz	L031aw_end
+	# Tail Round 5
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	addl	$4,%esi
+	addl	$4,%edi
+	addl	$4,%ebx
+	decl	%ebp
+	jz	L031aw_end
+	# Tail Round 6
+	movl	(%esi),%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	addl	$4,%esi
+	addl	$4,%edi
+	addl	$4,%ebx
+L031aw_end:
+	cmpl	$0,36(%esp)
+	je	L032pw_end
+	movl	36(%esp),%ebp
+	cmpl	$0,%ebp
+	je	L032pw_end
+	jge	L033pw_pos
+	# pw_neg
+	movl	$0,%edx
+	subl	%ebp,%edx
+	movl	%edx,%ebp
+	andl	$4294967288,%ebp
+	jz	L034pw_neg_finish
+L035pw_neg_loop:
+	# dl<0 Round 0
+	movl	$0,%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,(%ebx)
+	# dl<0 Round 1
+	movl	$0,%ecx
+	movl	4(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,4(%ebx)
+	# dl<0 Round 2
+	movl	$0,%ecx
+	movl	8(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,8(%ebx)
+	# dl<0 Round 3
+	movl	$0,%ecx
+	movl	12(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,12(%ebx)
+	# dl<0 Round 4
+	movl	$0,%ecx
+	movl	16(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,16(%ebx)
+	# dl<0 Round 5
+	movl	$0,%ecx
+	movl	20(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,20(%ebx)
+	# dl<0 Round 6
+	movl	$0,%ecx
+	movl	24(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,24(%ebx)
+	# dl<0 Round 7
+	movl	$0,%ecx
+	movl	28(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,28(%ebx)
+
+	addl	$32,%edi
+	addl	$32,%ebx
+	subl	$8,%ebp
+	jnz	L035pw_neg_loop
+L034pw_neg_finish:
+	movl	36(%esp),%edx
+	movl	$0,%ebp
+	subl	%edx,%ebp
+	andl	$7,%ebp
+	jz	L032pw_end
+	# dl<0 Tail Round 0
+	movl	$0,%ecx
+	movl	(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,(%ebx)
+	jz	L032pw_end
+	# dl<0 Tail Round 1
+	movl	$0,%ecx
+	movl	4(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,4(%ebx)
+	jz	L032pw_end
+	# dl<0 Tail Round 2
+	movl	$0,%ecx
+	movl	8(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,8(%ebx)
+	jz	L032pw_end
+	# dl<0 Tail Round 3
+	movl	$0,%ecx
+	movl	12(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,12(%ebx)
+	jz	L032pw_end
+	# dl<0 Tail Round 4
+	movl	$0,%ecx
+	movl	16(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,16(%ebx)
+	jz	L032pw_end
+	# dl<0 Tail Round 5
+	movl	$0,%ecx
+	movl	20(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	decl	%ebp
+	movl	%ecx,20(%ebx)
+	jz	L032pw_end
+	# dl<0 Tail Round 6
+	movl	$0,%ecx
+	movl	24(%edi),%edx
+	subl	%eax,%ecx
+	movl	$0,%eax
+	adcl	%eax,%eax
+	subl	%edx,%ecx
+	adcl	$0,%eax
+	movl	%ecx,24(%ebx)
+	jmp	L032pw_end
+L033pw_pos:
+	andl	$4294967288,%ebp
+	jz	L036pw_pos_finish
+L037pw_pos_loop:
+	# dl>0 Round 0
+	movl	(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,(%ebx)
+	jnc	L038pw_nc0
+	# dl>0 Round 1
+	movl	4(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,4(%ebx)
+	jnc	L039pw_nc1
+	# dl>0 Round 2
+	movl	8(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,8(%ebx)
+	jnc	L040pw_nc2
+	# dl>0 Round 3
+	movl	12(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,12(%ebx)
+	jnc	L041pw_nc3
+	# dl>0 Round 4
+	movl	16(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,16(%ebx)
+	jnc	L042pw_nc4
+	# dl>0 Round 5
+	movl	20(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,20(%ebx)
+	jnc	L043pw_nc5
+	# dl>0 Round 6
+	movl	24(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,24(%ebx)
+	jnc	L044pw_nc6
+	# dl>0 Round 7
+	movl	28(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,28(%ebx)
+	jnc	L045pw_nc7
+
+	addl	$32,%esi
+	addl	$32,%ebx
+	subl	$8,%ebp
+	jnz	L037pw_pos_loop
+L036pw_pos_finish:
+	movl	36(%esp),%ebp
+	andl	$7,%ebp
+	jz	L032pw_end
+	# dl>0 Tail Round 0
+	movl	(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,(%ebx)
+	jnc	L046pw_tail_nc0
+	decl	%ebp
+	jz	L032pw_end
+	# dl>0 Tail Round 1
+	movl	4(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,4(%ebx)
+	jnc	L047pw_tail_nc1
+	decl	%ebp
+	jz	L032pw_end
+	# dl>0 Tail Round 2
+	movl	8(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,8(%ebx)
+	jnc	L048pw_tail_nc2
+	decl	%ebp
+	jz	L032pw_end
+	# dl>0 Tail Round 3
+	movl	12(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,12(%ebx)
+	jnc	L049pw_tail_nc3
+	decl	%ebp
+	jz	L032pw_end
+	# dl>0 Tail Round 4
+	movl	16(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,16(%ebx)
+	jnc	L050pw_tail_nc4
+	decl	%ebp
+	jz	L032pw_end
+	# dl>0 Tail Round 5
+	movl	20(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,20(%ebx)
+	jnc	L051pw_tail_nc5
+	decl	%ebp
+	jz	L032pw_end
+	# dl>0 Tail Round 6
+	movl	24(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,24(%ebx)
+	jnc	L052pw_tail_nc6
+	movl	$1,%eax
+	jmp	L032pw_end
+L053pw_nc_loop:
+	movl	(%esi),%ecx
+	movl	%ecx,(%ebx)
+L038pw_nc0:
+	movl	4(%esi),%ecx
+	movl	%ecx,4(%ebx)
+L039pw_nc1:
+	movl	8(%esi),%ecx
+	movl	%ecx,8(%ebx)
+L040pw_nc2:
+	movl	12(%esi),%ecx
+	movl	%ecx,12(%ebx)
+L041pw_nc3:
+	movl	16(%esi),%ecx
+	movl	%ecx,16(%ebx)
+L042pw_nc4:
+	movl	20(%esi),%ecx
+	movl	%ecx,20(%ebx)
+L043pw_nc5:
+	movl	24(%esi),%ecx
+	movl	%ecx,24(%ebx)
+L044pw_nc6:
+	movl	28(%esi),%ecx
+	movl	%ecx,28(%ebx)
+L045pw_nc7:
+
+	addl	$32,%esi
+	addl	$32,%ebx
+	subl	$8,%ebp
+	jnz	L053pw_nc_loop
+	movl	36(%esp),%ebp
+	andl	$7,%ebp
+	jz	L054pw_nc_end
+	movl	(%esi),%ecx
+	movl	%ecx,(%ebx)
+L046pw_tail_nc0:
+	decl	%ebp
+	jz	L054pw_nc_end
+	movl	4(%esi),%ecx
+	movl	%ecx,4(%ebx)
+L047pw_tail_nc1:
+	decl	%ebp
+	jz	L054pw_nc_end
+	movl	8(%esi),%ecx
+	movl	%ecx,8(%ebx)
+L048pw_tail_nc2:
+	decl	%ebp
+	jz	L054pw_nc_end
+	movl	12(%esi),%ecx
+	movl	%ecx,12(%ebx)
+L049pw_tail_nc3:
+	decl	%ebp
+	jz	L054pw_nc_end
+	movl	16(%esi),%ecx
+	movl	%ecx,16(%ebx)
+L050pw_tail_nc4:
+	decl	%ebp
+	jz	L054pw_nc_end
+	movl	20(%esi),%ecx
+	movl	%ecx,20(%ebx)
+L051pw_tail_nc5:
+	decl	%ebp
+	jz	L054pw_nc_end
+	movl	24(%esi),%ecx
+	movl	%ecx,24(%ebx)
+L052pw_tail_nc6:
+L054pw_nc_end:
+	movl	$0,%eax
+L032pw_end:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.comm	_OPENSSL_ia32cap_P,16
diff --git a/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/co-586.s b/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/co-586.s
new file mode 100644
index 0000000000..0196f42b78
--- /dev/null
+++ b/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/co-586.s
@@ -0,0 +1,1249 @@
+.text
+.globl	_bn_mul_comba8
+.type	_bn_mul_comba8,@function
+.align	4
+_bn_mul_comba8:
+L_bn_mul_comba8_begin:
+	pushl	%esi
+	movl	12(%esp),%esi
+	pushl	%edi
+	movl	20(%esp),%edi
+	pushl	%ebp
+	pushl	%ebx
+	xorl	%ebx,%ebx
+	movl	(%esi),%eax
+	xorl	%ecx,%ecx
+	movl	(%edi),%edx
+	# ################## Calculate word 0
+	xorl	%ebp,%ebp
+	# mul a[0]*b[0]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esp),%eax
+	adcl	%edx,%ecx
+	movl	(%edi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,(%eax)
+	movl	4(%esi),%eax
+	# saved r[0]
+	# ################## Calculate word 1
+	xorl	%ebx,%ebx
+	# mul a[1]*b[0]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	(%esi),%eax
+	adcl	%edx,%ebp
+	movl	4(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[0]*b[1]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esp),%eax
+	adcl	%edx,%ebp
+	movl	(%edi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,4(%eax)
+	movl	8(%esi),%eax
+	# saved r[1]
+	# ################## Calculate word 2
+	xorl	%ecx,%ecx
+	# mul a[2]*b[0]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	4(%esi),%eax
+	adcl	%edx,%ebx
+	movl	4(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[1]*b[1]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	(%esi),%eax
+	adcl	%edx,%ebx
+	movl	8(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[0]*b[2]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esp),%eax
+	adcl	%edx,%ebx
+	movl	(%edi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,8(%eax)
+	movl	12(%esi),%eax
+	# saved r[2]
+	# ################## Calculate word 3
+	xorl	%ebp,%ebp
+	# mul a[3]*b[0]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	8(%esi),%eax
+	adcl	%edx,%ecx
+	movl	4(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[2]*b[1]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	4(%esi),%eax
+	adcl	%edx,%ecx
+	movl	8(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[1]*b[2]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	(%esi),%eax
+	adcl	%edx,%ecx
+	movl	12(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[0]*b[3]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esp),%eax
+	adcl	%edx,%ecx
+	movl	(%edi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,12(%eax)
+	movl	16(%esi),%eax
+	# saved r[3]
+	# ################## Calculate word 4
+	xorl	%ebx,%ebx
+	# mul a[4]*b[0]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	12(%esi),%eax
+	adcl	%edx,%ebp
+	movl	4(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[3]*b[1]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	8(%esi),%eax
+	adcl	%edx,%ebp
+	movl	8(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[2]*b[2]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	4(%esi),%eax
+	adcl	%edx,%ebp
+	movl	12(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[1]*b[3]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	(%esi),%eax
+	adcl	%edx,%ebp
+	movl	16(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[0]*b[4]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esp),%eax
+	adcl	%edx,%ebp
+	movl	(%edi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,16(%eax)
+	movl	20(%esi),%eax
+	# saved r[4]
+	# ################## Calculate word 5
+	xorl	%ecx,%ecx
+	# mul a[5]*b[0]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	16(%esi),%eax
+	adcl	%edx,%ebx
+	movl	4(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[4]*b[1]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	12(%esi),%eax
+	adcl	%edx,%ebx
+	movl	8(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[3]*b[2]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	8(%esi),%eax
+	adcl	%edx,%ebx
+	movl	12(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[2]*b[3]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	4(%esi),%eax
+	adcl	%edx,%ebx
+	movl	16(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[1]*b[4]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	(%esi),%eax
+	adcl	%edx,%ebx
+	movl	20(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[0]*b[5]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esp),%eax
+	adcl	%edx,%ebx
+	movl	(%edi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,20(%eax)
+	movl	24(%esi),%eax
+	# saved r[5]
+	# ################## Calculate word 6
+	xorl	%ebp,%ebp
+	# mul a[6]*b[0]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esi),%eax
+	adcl	%edx,%ecx
+	movl	4(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[5]*b[1]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	16(%esi),%eax
+	adcl	%edx,%ecx
+	movl	8(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[4]*b[2]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	12(%esi),%eax
+	adcl	%edx,%ecx
+	movl	12(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[3]*b[3]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	8(%esi),%eax
+	adcl	%edx,%ecx
+	movl	16(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[2]*b[4]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	4(%esi),%eax
+	adcl	%edx,%ecx
+	movl	20(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[1]*b[5]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	(%esi),%eax
+	adcl	%edx,%ecx
+	movl	24(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[0]*b[6]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esp),%eax
+	adcl	%edx,%ecx
+	movl	(%edi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,24(%eax)
+	movl	28(%esi),%eax
+	# saved r[6]
+	# ################## Calculate word 7
+	xorl	%ebx,%ebx
+	# mul a[7]*b[0]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	24(%esi),%eax
+	adcl	%edx,%ebp
+	movl	4(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[6]*b[1]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esi),%eax
+	adcl	%edx,%ebp
+	movl	8(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[5]*b[2]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	16(%esi),%eax
+	adcl	%edx,%ebp
+	movl	12(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[4]*b[3]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	12(%esi),%eax
+	adcl	%edx,%ebp
+	movl	16(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[3]*b[4]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	8(%esi),%eax
+	adcl	%edx,%ebp
+	movl	20(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[2]*b[5]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	4(%esi),%eax
+	adcl	%edx,%ebp
+	movl	24(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[1]*b[6]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	(%esi),%eax
+	adcl	%edx,%ebp
+	movl	28(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[0]*b[7]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esp),%eax
+	adcl	%edx,%ebp
+	movl	4(%edi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,28(%eax)
+	movl	28(%esi),%eax
+	# saved r[7]
+	# ################## Calculate word 8
+	xorl	%ecx,%ecx
+	# mul a[7]*b[1]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	24(%esi),%eax
+	adcl	%edx,%ebx
+	movl	8(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[6]*b[2]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esi),%eax
+	adcl	%edx,%ebx
+	movl	12(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[5]*b[3]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	16(%esi),%eax
+	adcl	%edx,%ebx
+	movl	16(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[4]*b[4]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	12(%esi),%eax
+	adcl	%edx,%ebx
+	movl	20(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[3]*b[5]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	8(%esi),%eax
+	adcl	%edx,%ebx
+	movl	24(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[2]*b[6]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	4(%esi),%eax
+	adcl	%edx,%ebx
+	movl	28(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[1]*b[7]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esp),%eax
+	adcl	%edx,%ebx
+	movl	8(%edi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,32(%eax)
+	movl	28(%esi),%eax
+	# saved r[8]
+	# ################## Calculate word 9
+	xorl	%ebp,%ebp
+	# mul a[7]*b[2]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	24(%esi),%eax
+	adcl	%edx,%ecx
+	movl	12(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[6]*b[3]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esi),%eax
+	adcl	%edx,%ecx
+	movl	16(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[5]*b[4]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	16(%esi),%eax
+	adcl	%edx,%ecx
+	movl	20(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[4]*b[5]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	12(%esi),%eax
+	adcl	%edx,%ecx
+	movl	24(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[3]*b[6]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	8(%esi),%eax
+	adcl	%edx,%ecx
+	movl	28(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[2]*b[7]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esp),%eax
+	adcl	%edx,%ecx
+	movl	12(%edi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,36(%eax)
+	movl	28(%esi),%eax
+	# saved r[9]
+	# ################## Calculate word 10
+	xorl	%ebx,%ebx
+	# mul a[7]*b[3]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	24(%esi),%eax
+	adcl	%edx,%ebp
+	movl	16(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[6]*b[4]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esi),%eax
+	adcl	%edx,%ebp
+	movl	20(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[5]*b[5]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	16(%esi),%eax
+	adcl	%edx,%ebp
+	movl	24(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[4]*b[6]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	12(%esi),%eax
+	adcl	%edx,%ebp
+	movl	28(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[3]*b[7]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esp),%eax
+	adcl	%edx,%ebp
+	movl	16(%edi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,40(%eax)
+	movl	28(%esi),%eax
+	# saved r[10]
+	# ################## Calculate word 11
+	xorl	%ecx,%ecx
+	# mul a[7]*b[4]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	24(%esi),%eax
+	adcl	%edx,%ebx
+	movl	20(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[6]*b[5]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esi),%eax
+	adcl	%edx,%ebx
+	movl	24(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[5]*b[6]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	16(%esi),%eax
+	adcl	%edx,%ebx
+	movl	28(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[4]*b[7]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esp),%eax
+	adcl	%edx,%ebx
+	movl	20(%edi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,44(%eax)
+	movl	28(%esi),%eax
+	# saved r[11]
+	# ################## Calculate word 12
+	xorl	%ebp,%ebp
+	# mul a[7]*b[5]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	24(%esi),%eax
+	adcl	%edx,%ecx
+	movl	24(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[6]*b[6]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esi),%eax
+	adcl	%edx,%ecx
+	movl	28(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[5]*b[7]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esp),%eax
+	adcl	%edx,%ecx
+	movl	24(%edi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,48(%eax)
+	movl	28(%esi),%eax
+	# saved r[12]
+	# ################## Calculate word 13
+	xorl	%ebx,%ebx
+	# mul a[7]*b[6]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	24(%esi),%eax
+	adcl	%edx,%ebp
+	movl	28(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[6]*b[7]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esp),%eax
+	adcl	%edx,%ebp
+	movl	28(%edi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,52(%eax)
+	movl	28(%esi),%eax
+	# saved r[13]
+	# ################## Calculate word 14
+	xorl	%ecx,%ecx
+	# mul a[7]*b[7]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esp),%eax
+	adcl	%edx,%ebx
+	adcl	$0,%ecx
+	movl	%ebp,56(%eax)
+	# saved r[14]
+	# save r[15]
+	movl	%ebx,60(%eax)
+	popl	%ebx
+	popl	%ebp
+	popl	%edi
+	popl	%esi
+	ret
+.globl	_bn_mul_comba4
+.type	_bn_mul_comba4,@function
+.align	4
+_bn_mul_comba4:
+L_bn_mul_comba4_begin:
+	pushl	%esi
+	movl	12(%esp),%esi
+	pushl	%edi
+	movl	20(%esp),%edi
+	pushl	%ebp
+	pushl	%ebx
+	xorl	%ebx,%ebx
+	movl	(%esi),%eax
+	xorl	%ecx,%ecx
+	movl	(%edi),%edx
+	# ################## Calculate word 0
+	xorl	%ebp,%ebp
+	# mul a[0]*b[0]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esp),%eax
+	adcl	%edx,%ecx
+	movl	(%edi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,(%eax)
+	movl	4(%esi),%eax
+	# saved r[0]
+	# ################## Calculate word 1
+	xorl	%ebx,%ebx
+	# mul a[1]*b[0]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	(%esi),%eax
+	adcl	%edx,%ebp
+	movl	4(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[0]*b[1]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esp),%eax
+	adcl	%edx,%ebp
+	movl	(%edi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,4(%eax)
+	movl	8(%esi),%eax
+	# saved r[1]
+	# ################## Calculate word 2
+	xorl	%ecx,%ecx
+	# mul a[2]*b[0]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	4(%esi),%eax
+	adcl	%edx,%ebx
+	movl	4(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[1]*b[1]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	(%esi),%eax
+	adcl	%edx,%ebx
+	movl	8(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[0]*b[2]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esp),%eax
+	adcl	%edx,%ebx
+	movl	(%edi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,8(%eax)
+	movl	12(%esi),%eax
+	# saved r[2]
+	# ################## Calculate word 3
+	xorl	%ebp,%ebp
+	# mul a[3]*b[0]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	8(%esi),%eax
+	adcl	%edx,%ecx
+	movl	4(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[2]*b[1]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	4(%esi),%eax
+	adcl	%edx,%ecx
+	movl	8(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[1]*b[2]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	(%esi),%eax
+	adcl	%edx,%ecx
+	movl	12(%edi),%edx
+	adcl	$0,%ebp
+	# mul a[0]*b[3]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esp),%eax
+	adcl	%edx,%ecx
+	movl	4(%edi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,12(%eax)
+	movl	12(%esi),%eax
+	# saved r[3]
+	# ################## Calculate word 4
+	xorl	%ebx,%ebx
+	# mul a[3]*b[1]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	8(%esi),%eax
+	adcl	%edx,%ebp
+	movl	8(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[2]*b[2]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	4(%esi),%eax
+	adcl	%edx,%ebp
+	movl	12(%edi),%edx
+	adcl	$0,%ebx
+	# mul a[1]*b[3]
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esp),%eax
+	adcl	%edx,%ebp
+	movl	8(%edi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,16(%eax)
+	movl	12(%esi),%eax
+	# saved r[4]
+	# ################## Calculate word 5
+	xorl	%ecx,%ecx
+	# mul a[3]*b[2]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	8(%esi),%eax
+	adcl	%edx,%ebx
+	movl	12(%edi),%edx
+	adcl	$0,%ecx
+	# mul a[2]*b[3]
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esp),%eax
+	adcl	%edx,%ebx
+	movl	12(%edi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,20(%eax)
+	movl	12(%esi),%eax
+	# saved r[5]
+	# ################## Calculate word 6
+	xorl	%ebp,%ebp
+	# mul a[3]*b[3]
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esp),%eax
+	adcl	%edx,%ecx
+	adcl	$0,%ebp
+	movl	%ebx,24(%eax)
+	# saved r[6]
+	# save r[7]
+	movl	%ecx,28(%eax)
+	popl	%ebx
+	popl	%ebp
+	popl	%edi
+	popl	%esi
+	ret
+.globl	_bn_sqr_comba8
+.type	_bn_sqr_comba8,@function
+.align	4
+_bn_sqr_comba8:
+L_bn_sqr_comba8_begin:
+	pushl	%esi
+	pushl	%edi
+	pushl	%ebp
+	pushl	%ebx
+	movl	20(%esp),%edi
+	movl	24(%esp),%esi
+	xorl	%ebx,%ebx
+	xorl	%ecx,%ecx
+	movl	(%esi),%eax
+	# ############### Calculate word 0
+	xorl	%ebp,%ebp
+	# sqr a[0]*a[0]
+	mull	%eax
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	(%esi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,(%edi)
+	movl	4(%esi),%eax
+	# saved r[0]
+	# ############### Calculate word 1
+	xorl	%ebx,%ebx
+	# sqr a[1]*a[0]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	8(%esi),%eax
+	adcl	$0,%ebx
+	movl	%ecx,4(%edi)
+	movl	(%esi),%edx
+	# saved r[1]
+	# ############### Calculate word 2
+	xorl	%ecx,%ecx
+	# sqr a[2]*a[0]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	4(%esi),%eax
+	adcl	$0,%ecx
+	# sqr a[1]*a[1]
+	mull	%eax
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	(%esi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,8(%edi)
+	movl	12(%esi),%eax
+	# saved r[2]
+	# ############### Calculate word 3
+	xorl	%ebp,%ebp
+	# sqr a[3]*a[0]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	8(%esi),%eax
+	adcl	$0,%ebp
+	movl	4(%esi),%edx
+	# sqr a[2]*a[1]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	16(%esi),%eax
+	adcl	$0,%ebp
+	movl	%ebx,12(%edi)
+	movl	(%esi),%edx
+	# saved r[3]
+	# ############### Calculate word 4
+	xorl	%ebx,%ebx
+	# sqr a[4]*a[0]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	12(%esi),%eax
+	adcl	$0,%ebx
+	movl	4(%esi),%edx
+	# sqr a[3]*a[1]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	8(%esi),%eax
+	adcl	$0,%ebx
+	# sqr a[2]*a[2]
+	mull	%eax
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	(%esi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,16(%edi)
+	movl	20(%esi),%eax
+	# saved r[4]
+	# ############### Calculate word 5
+	xorl	%ecx,%ecx
+	# sqr a[5]*a[0]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	16(%esi),%eax
+	adcl	$0,%ecx
+	movl	4(%esi),%edx
+	# sqr a[4]*a[1]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	12(%esi),%eax
+	adcl	$0,%ecx
+	movl	8(%esi),%edx
+	# sqr a[3]*a[2]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	24(%esi),%eax
+	adcl	$0,%ecx
+	movl	%ebp,20(%edi)
+	movl	(%esi),%edx
+	# saved r[5]
+	# ############### Calculate word 6
+	xorl	%ebp,%ebp
+	# sqr a[6]*a[0]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	20(%esi),%eax
+	adcl	$0,%ebp
+	movl	4(%esi),%edx
+	# sqr a[5]*a[1]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	16(%esi),%eax
+	adcl	$0,%ebp
+	movl	8(%esi),%edx
+	# sqr a[4]*a[2]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	12(%esi),%eax
+	adcl	$0,%ebp
+	# sqr a[3]*a[3]
+	mull	%eax
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	(%esi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,24(%edi)
+	movl	28(%esi),%eax
+	# saved r[6]
+	# ############### Calculate word 7
+	xorl	%ebx,%ebx
+	# sqr a[7]*a[0]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	24(%esi),%eax
+	adcl	$0,%ebx
+	movl	4(%esi),%edx
+	# sqr a[6]*a[1]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	20(%esi),%eax
+	adcl	$0,%ebx
+	movl	8(%esi),%edx
+	# sqr a[5]*a[2]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	16(%esi),%eax
+	adcl	$0,%ebx
+	movl	12(%esi),%edx
+	# sqr a[4]*a[3]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	28(%esi),%eax
+	adcl	$0,%ebx
+	movl	%ecx,28(%edi)
+	movl	4(%esi),%edx
+	# saved r[7]
+	# ############### Calculate word 8
+	xorl	%ecx,%ecx
+	# sqr a[7]*a[1]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	24(%esi),%eax
+	adcl	$0,%ecx
+	movl	8(%esi),%edx
+	# sqr a[6]*a[2]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	20(%esi),%eax
+	adcl	$0,%ecx
+	movl	12(%esi),%edx
+	# sqr a[5]*a[3]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	16(%esi),%eax
+	adcl	$0,%ecx
+	# sqr a[4]*a[4]
+	mull	%eax
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	8(%esi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,32(%edi)
+	movl	28(%esi),%eax
+	# saved r[8]
+	# ############### Calculate word 9
+	xorl	%ebp,%ebp
+	# sqr a[7]*a[2]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	24(%esi),%eax
+	adcl	$0,%ebp
+	movl	12(%esi),%edx
+	# sqr a[6]*a[3]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	20(%esi),%eax
+	adcl	$0,%ebp
+	movl	16(%esi),%edx
+	# sqr a[5]*a[4]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	28(%esi),%eax
+	adcl	$0,%ebp
+	movl	%ebx,36(%edi)
+	movl	12(%esi),%edx
+	# saved r[9]
+	# ############### Calculate word 10
+	xorl	%ebx,%ebx
+	# sqr a[7]*a[3]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	24(%esi),%eax
+	adcl	$0,%ebx
+	movl	16(%esi),%edx
+	# sqr a[6]*a[4]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	20(%esi),%eax
+	adcl	$0,%ebx
+	# sqr a[5]*a[5]
+	mull	%eax
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	16(%esi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,40(%edi)
+	movl	28(%esi),%eax
+	# saved r[10]
+	# ############### Calculate word 11
+	xorl	%ecx,%ecx
+	# sqr a[7]*a[4]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	24(%esi),%eax
+	adcl	$0,%ecx
+	movl	20(%esi),%edx
+	# sqr a[6]*a[5]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	28(%esi),%eax
+	adcl	$0,%ecx
+	movl	%ebp,44(%edi)
+	movl	20(%esi),%edx
+	# saved r[11]
+	# ############### Calculate word 12
+	xorl	%ebp,%ebp
+	# sqr a[7]*a[5]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	24(%esi),%eax
+	adcl	$0,%ebp
+	# sqr a[6]*a[6]
+	mull	%eax
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	24(%esi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,48(%edi)
+	movl	28(%esi),%eax
+	# saved r[12]
+	# ############### Calculate word 13
+	xorl	%ebx,%ebx
+	# sqr a[7]*a[6]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	28(%esi),%eax
+	adcl	$0,%ebx
+	movl	%ecx,52(%edi)
+	# saved r[13]
+	# ############### Calculate word 14
+	xorl	%ecx,%ecx
+	# sqr a[7]*a[7]
+	mull	%eax
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	adcl	$0,%ecx
+	movl	%ebp,56(%edi)
+	# saved r[14]
+	movl	%ebx,60(%edi)
+	popl	%ebx
+	popl	%ebp
+	popl	%edi
+	popl	%esi
+	ret
+.globl	_bn_sqr_comba4
+.type	_bn_sqr_comba4,@function
+.align	4
+_bn_sqr_comba4:
+L_bn_sqr_comba4_begin:
+	pushl	%esi
+	pushl	%edi
+	pushl	%ebp
+	pushl	%ebx
+	movl	20(%esp),%edi
+	movl	24(%esp),%esi
+	xorl	%ebx,%ebx
+	xorl	%ecx,%ecx
+	movl	(%esi),%eax
+	# ############### Calculate word 0
+	xorl	%ebp,%ebp
+	# sqr a[0]*a[0]
+	mull	%eax
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	(%esi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,(%edi)
+	movl	4(%esi),%eax
+	# saved r[0]
+	# ############### Calculate word 1
+	xorl	%ebx,%ebx
+	# sqr a[1]*a[0]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	8(%esi),%eax
+	adcl	$0,%ebx
+	movl	%ecx,4(%edi)
+	movl	(%esi),%edx
+	# saved r[1]
+	# ############### Calculate word 2
+	xorl	%ecx,%ecx
+	# sqr a[2]*a[0]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	4(%esi),%eax
+	adcl	$0,%ecx
+	# sqr a[1]*a[1]
+	mull	%eax
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	(%esi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,8(%edi)
+	movl	12(%esi),%eax
+	# saved r[2]
+	# ############### Calculate word 3
+	xorl	%ebp,%ebp
+	# sqr a[3]*a[0]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	8(%esi),%eax
+	adcl	$0,%ebp
+	movl	4(%esi),%edx
+	# sqr a[2]*a[1]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	12(%esi),%eax
+	adcl	$0,%ebp
+	movl	%ebx,12(%edi)
+	movl	4(%esi),%edx
+	# saved r[3]
+	# ############### Calculate word 4
+	xorl	%ebx,%ebx
+	# sqr a[3]*a[1]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	8(%esi),%eax
+	adcl	$0,%ebx
+	# sqr a[2]*a[2]
+	mull	%eax
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	8(%esi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,16(%edi)
+	movl	12(%esi),%eax
+	# saved r[4]
+	# ############### Calculate word 5
+	xorl	%ecx,%ecx
+	# sqr a[3]*a[2]
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	12(%esi),%eax
+	adcl	$0,%ecx
+	movl	%ebp,20(%edi)
+	# saved r[5]
+	# ############### Calculate word 6
+	xorl	%ebp,%ebp
+	# sqr a[3]*a[3]
+	mull	%eax
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	adcl	$0,%ebp
+	movl	%ebx,24(%edi)
+	# saved r[6]
+	movl	%ecx,28(%edi)
+	popl	%ebx
+	popl	%ebp
+	popl	%edi
+	popl	%esi
+	ret
diff --git a/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/x86-gf2m.s b/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/x86-gf2m.s
new file mode 100644
index 0000000000..0999445f43
--- /dev/null
+++ b/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/x86-gf2m.s
@@ -0,0 +1,343 @@
+.text
+.type	__mul_1x1_mmx,@function
+.align	4
+__mul_1x1_mmx:
+	subl	$36,%esp
+	movl	%eax,%ecx
+	leal	(%eax,%eax,1),%edx
+	andl	$1073741823,%ecx
+	leal	(%edx,%edx,1),%ebp
+	movl	$0,(%esp)
+	andl	$2147483647,%edx
+	movd	%eax,%mm2
+	movd	%ebx,%mm3
+	movl	%ecx,4(%esp)
+	xorl	%edx,%ecx
+	pxor	%mm5,%mm5
+	pxor	%mm4,%mm4
+	movl	%edx,8(%esp)
+	xorl	%ebp,%edx
+	movl	%ecx,12(%esp)
+	pcmpgtd	%mm2,%mm5
+	paddd	%mm2,%mm2
+	xorl	%edx,%ecx
+	movl	%ebp,16(%esp)
+	xorl	%edx,%ebp
+	pand	%mm3,%mm5
+	pcmpgtd	%mm2,%mm4
+	movl	%ecx,20(%esp)
+	xorl	%ecx,%ebp
+	psllq	$31,%mm5
+	pand	%mm3,%mm4
+	movl	%edx,24(%esp)
+	movl	$7,%esi
+	movl	%ebp,28(%esp)
+	movl	%esi,%ebp
+	andl	%ebx,%esi
+	shrl	$3,%ebx
+	movl	%ebp,%edi
+	psllq	$30,%mm4
+	andl	%ebx,%edi
+	shrl	$3,%ebx
+	movd	(%esp,%esi,4),%mm0
+	movl	%ebp,%esi
+	andl	%ebx,%esi
+	shrl	$3,%ebx
+	movd	(%esp,%edi,4),%mm2
+	movl	%ebp,%edi
+	psllq	$3,%mm2
+	andl	%ebx,%edi
+	shrl	$3,%ebx
+	pxor	%mm2,%mm0
+	movd	(%esp,%esi,4),%mm1
+	movl	%ebp,%esi
+	psllq	$6,%mm1
+	andl	%ebx,%esi
+	shrl	$3,%ebx
+	pxor	%mm1,%mm0
+	movd	(%esp,%edi,4),%mm2
+	movl	%ebp,%edi
+	psllq	$9,%mm2
+	andl	%ebx,%edi
+	shrl	$3,%ebx
+	pxor	%mm2,%mm0
+	movd	(%esp,%esi,4),%mm1
+	movl	%ebp,%esi
+	psllq	$12,%mm1
+	andl	%ebx,%esi
+	shrl	$3,%ebx
+	pxor	%mm1,%mm0
+	movd	(%esp,%edi,4),%mm2
+	movl	%ebp,%edi
+	psllq	$15,%mm2
+	andl	%ebx,%edi
+	shrl	$3,%ebx
+	pxor	%mm2,%mm0
+	movd	(%esp,%esi,4),%mm1
+	movl	%ebp,%esi
+	psllq	$18,%mm1
+	andl	%ebx,%esi
+	shrl	$3,%ebx
+	pxor	%mm1,%mm0
+	movd	(%esp,%edi,4),%mm2
+	movl	%ebp,%edi
+	psllq	$21,%mm2
+	andl	%ebx,%edi
+	shrl	$3,%ebx
+	pxor	%mm2,%mm0
+	movd	(%esp,%esi,4),%mm1
+	movl	%ebp,%esi
+	psllq	$24,%mm1
+	andl	%ebx,%esi
+	shrl	$3,%ebx
+	pxor	%mm1,%mm0
+	movd	(%esp,%edi,4),%mm2
+	pxor	%mm4,%mm0
+	psllq	$27,%mm2
+	pxor	%mm2,%mm0
+	movd	(%esp,%esi,4),%mm1
+	pxor	%mm5,%mm0
+	psllq	$30,%mm1
+	addl	$36,%esp
+	pxor	%mm1,%mm0
+	ret
+.type	__mul_1x1_ialu,@function
+.align	4
+__mul_1x1_ialu:
+	subl	$36,%esp
+	movl	%eax,%ecx
+	leal	(%eax,%eax,1),%edx
+	leal	(,%eax,4),%ebp
+	andl	$1073741823,%ecx
+	leal	(%eax,%eax,1),%edi
+	sarl	$31,%eax
+	movl	$0,(%esp)
+	andl	$2147483647,%edx
+	movl	%ecx,4(%esp)
+	xorl	%edx,%ecx
+	movl	%edx,8(%esp)
+	xorl	%ebp,%edx
+	movl	%ecx,12(%esp)
+	xorl	%edx,%ecx
+	movl	%ebp,16(%esp)
+	xorl	%edx,%ebp
+	movl	%ecx,20(%esp)
+	xorl	%ecx,%ebp
+	sarl	$31,%edi
+	andl	%ebx,%eax
+	movl	%edx,24(%esp)
+	andl	%ebx,%edi
+	movl	%ebp,28(%esp)
+	movl	%eax,%edx
+	shll	$31,%eax
+	movl	%edi,%ecx
+	shrl	$1,%edx
+	movl	$7,%esi
+	shll	$30,%edi
+	andl	%ebx,%esi
+	shrl	$2,%ecx
+	xorl	%edi,%eax
+	shrl	$3,%ebx
+	movl	$7,%edi
+	andl	%ebx,%edi
+	shrl	$3,%ebx
+	xorl	%ecx,%edx
+	xorl	(%esp,%esi,4),%eax
+	movl	$7,%esi
+	andl	%ebx,%esi
+	shrl	$3,%ebx
+	movl	(%esp,%edi,4),%ebp
+	movl	$7,%edi
+	movl	%ebp,%ecx
+	shll	$3,%ebp
+	andl	%ebx,%edi
+	shrl	$29,%ecx
+	xorl	%ebp,%eax
+	shrl	$3,%ebx
+	xorl	%ecx,%edx
+	movl	(%esp,%esi,4),%ecx
+	movl	$7,%esi
+	movl	%ecx,%ebp
+	shll	$6,%ecx
+	andl	%ebx,%esi
+	shrl	$26,%ebp
+	xorl	%ecx,%eax
+	shrl	$3,%ebx
+	xorl	%ebp,%edx
+	movl	(%esp,%edi,4),%ebp
+	movl	$7,%edi
+	movl	%ebp,%ecx
+	shll	$9,%ebp
+	andl	%ebx,%edi
+	shrl	$23,%ecx
+	xorl	%ebp,%eax
+	shrl	$3,%ebx
+	xorl	%ecx,%edx
+	movl	(%esp,%esi,4),%ecx
+	movl	$7,%esi
+	movl	%ecx,%ebp
+	shll	$12,%ecx
+	andl	%ebx,%esi
+	shrl	$20,%ebp
+	xorl	%ecx,%eax
+	shrl	$3,%ebx
+	xorl	%ebp,%edx
+	movl	(%esp,%edi,4),%ebp
+	movl	$7,%edi
+	movl	%ebp,%ecx
+	shll	$15,%ebp
+	andl	%ebx,%edi
+	shrl	$17,%ecx
+	xorl	%ebp,%eax
+	shrl	$3,%ebx
+	xorl	%ecx,%edx
+	movl	(%esp,%esi,4),%ecx
+	movl	$7,%esi
+	movl	%ecx,%ebp
+	shll	$18,%ecx
+	andl	%ebx,%esi
+	shrl	$14,%ebp
+	xorl	%ecx,%eax
+	shrl	$3,%ebx
+	xorl	%ebp,%edx
+	movl	(%esp,%edi,4),%ebp
+	movl	$7,%edi
+	movl	%ebp,%ecx
+	shll	$21,%ebp
+	andl	%ebx,%edi
+	shrl	$11,%ecx
+	xorl	%ebp,%eax
+	shrl	$3,%ebx
+	xorl	%ecx,%edx
+	movl	(%esp,%esi,4),%ecx
+	movl	$7,%esi
+	movl	%ecx,%ebp
+	shll	$24,%ecx
+	andl	%ebx,%esi
+	shrl	$8,%ebp
+	xorl	%ecx,%eax
+	shrl	$3,%ebx
+	xorl	%ebp,%edx
+	movl	(%esp,%edi,4),%ebp
+	movl	%ebp,%ecx
+	shll	$27,%ebp
+	movl	(%esp,%esi,4),%edi
+	shrl	$5,%ecx
+	movl	%edi,%esi
+	xorl	%ebp,%eax
+	shll	$30,%edi
+	xorl	%ecx,%edx
+	shrl	$2,%esi
+	xorl	%edi,%eax
+	xorl	%esi,%edx
+	addl	$36,%esp
+	ret
+.globl	_bn_GF2m_mul_2x2
+.type	_bn_GF2m_mul_2x2,@function
+.align	4
+_bn_GF2m_mul_2x2:
+L_bn_GF2m_mul_2x2_begin:
+	call	L000PIC_me_up
+L000PIC_me_up:
+	popl	%edx
+	leal	__GLOBAL_OFFSET_TABLE_+[.-L000PIC_me_up](%edx),%edx
+	movl	_OPENSSL_ia32cap_P@GOT(%edx),%edx
+	movl	(%edx),%eax
+	movl	4(%edx),%edx
+	testl	$8388608,%eax
+	jz	L001ialu
+	testl	$16777216,%eax
+	jz	L002mmx
+	testl	$2,%edx
+	jz	L002mmx
+	movups	8(%esp),%xmm0
+	shufps	$177,%xmm0,%xmm0
+.byte	102,15,58,68,192,1
+	movl	4(%esp),%eax
+	movups	%xmm0,(%eax)
+	ret
+.align	4,0x90
+L002mmx:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	movl	24(%esp),%eax
+	movl	32(%esp),%ebx
+	call	__mul_1x1_mmx
+	movq	%mm0,%mm7
+	movl	28(%esp),%eax
+	movl	36(%esp),%ebx
+	call	__mul_1x1_mmx
+	movq	%mm0,%mm6
+	movl	24(%esp),%eax
+	movl	32(%esp),%ebx
+	xorl	28(%esp),%eax
+	xorl	36(%esp),%ebx
+	call	__mul_1x1_mmx
+	pxor	%mm7,%mm0
+	movl	20(%esp),%eax
+	pxor	%mm6,%mm0
+	movq	%mm0,%mm2
+	psllq	$32,%mm0
+	popl	%edi
+	psrlq	$32,%mm2
+	popl	%esi
+	pxor	%mm6,%mm0
+	popl	%ebx
+	pxor	%mm7,%mm2
+	movq	%mm0,(%eax)
+	popl	%ebp
+	movq	%mm2,8(%eax)
+	emms
+	ret
+.align	4,0x90
+L001ialu:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	subl	$20,%esp
+	movl	44(%esp),%eax
+	movl	52(%esp),%ebx
+	call	__mul_1x1_ialu
+	movl	%eax,8(%esp)
+	movl	%edx,12(%esp)
+	movl	48(%esp),%eax
+	movl	56(%esp),%ebx
+	call	__mul_1x1_ialu
+	movl	%eax,(%esp)
+	movl	%edx,4(%esp)
+	movl	44(%esp),%eax
+	movl	52(%esp),%ebx
+	xorl	48(%esp),%eax
+	xorl	56(%esp),%ebx
+	call	__mul_1x1_ialu
+	movl	40(%esp),%ebp
+	movl	(%esp),%ebx
+	movl	4(%esp),%ecx
+	movl	8(%esp),%edi
+	movl	12(%esp),%esi
+	xorl	%edx,%eax
+	xorl	%ecx,%edx
+	xorl	%ebx,%eax
+	movl	%ebx,(%ebp)
+	xorl	%edi,%edx
+	movl	%esi,12(%ebp)
+	xorl	%esi,%eax
+	addl	$20,%esp
+	xorl	%esi,%edx
+	popl	%edi
+	xorl	%edx,%eax
+	popl	%esi
+	movl	%edx,8(%ebp)
+	popl	%ebx
+	movl	%eax,4(%ebp)
+	popl	%ebp
+	ret
+.byte	71,70,40,50,94,109,41,32,77,117,108,116,105,112,108,105
+.byte	99,97,116,105,111,110,32,102,111,114,32,120,56,54,44,32
+.byte	67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97
+.byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
+.byte	62,0
+.comm	_OPENSSL_ia32cap_P,16
diff --git a/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/x86-mont.s b/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/x86-mont.s
new file mode 100644
index 0000000000..2612cf541b
--- /dev/null
+++ b/deps/openssl/config/archs/BSD-x86/asm/crypto/bn/x86-mont.s
@@ -0,0 +1,475 @@
+.text
+.globl	_bn_mul_mont
+.type	_bn_mul_mont,@function
+.align	4
+_bn_mul_mont:
+L_bn_mul_mont_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	xorl	%eax,%eax
+	movl	40(%esp),%edi
+	cmpl	$4,%edi
+	jl	L000just_leave
+	leal	20(%esp),%esi
+	leal	24(%esp),%edx
+	addl	$2,%edi
+	negl	%edi
+	leal	-32(%esp,%edi,4),%ebp
+	negl	%edi
+	movl	%ebp,%eax
+	subl	%edx,%eax
+	andl	$2047,%eax
+	subl	%eax,%ebp
+	xorl	%ebp,%edx
+	andl	$2048,%edx
+	xorl	$2048,%edx
+	subl	%edx,%ebp
+	andl	$-64,%ebp
+	movl	%esp,%eax
+	subl	%ebp,%eax
+	andl	$-4096,%eax
+	movl	%esp,%edx
+	leal	(%ebp,%eax,1),%esp
+	movl	(%esp),%eax
+	cmpl	%ebp,%esp
+	ja	L001page_walk
+	jmp	L002page_walk_done
+.align	4,0x90
+L001page_walk:
+	leal	-4096(%esp),%esp
+	movl	(%esp),%eax
+	cmpl	%ebp,%esp
+	ja	L001page_walk
+L002page_walk_done:
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%ebp
+	movl	16(%esi),%esi
+	movl	(%esi),%esi
+	movl	%eax,4(%esp)
+	movl	%ebx,8(%esp)
+	movl	%ecx,12(%esp)
+	movl	%ebp,16(%esp)
+	movl	%esi,20(%esp)
+	leal	-3(%edi),%ebx
+	movl	%edx,24(%esp)
+	call	L003PIC_me_up
+L003PIC_me_up:
+	popl	%eax
+	leal	__GLOBAL_OFFSET_TABLE_+[.-L003PIC_me_up](%eax),%eax
+	movl	_OPENSSL_ia32cap_P@GOT(%eax),%eax
+	btl	$26,(%eax)
+	jnc	L004non_sse2
+	movl	$-1,%eax
+	movd	%eax,%mm7
+	movl	8(%esp),%esi
+	movl	12(%esp),%edi
+	movl	16(%esp),%ebp
+	xorl	%edx,%edx
+	xorl	%ecx,%ecx
+	movd	(%edi),%mm4
+	movd	(%esi),%mm5
+	movd	(%ebp),%mm3
+	pmuludq	%mm4,%mm5
+	movq	%mm5,%mm2
+	movq	%mm5,%mm0
+	pand	%mm7,%mm0
+	pmuludq	20(%esp),%mm5
+	pmuludq	%mm5,%mm3
+	paddq	%mm0,%mm3
+	movd	4(%ebp),%mm1
+	movd	4(%esi),%mm0
+	psrlq	$32,%mm2
+	psrlq	$32,%mm3
+	incl	%ecx
+.align	4,0x90
+L0051st:
+	pmuludq	%mm4,%mm0
+	pmuludq	%mm5,%mm1
+	paddq	%mm0,%mm2
+	paddq	%mm1,%mm3
+	movq	%mm2,%mm0
+	pand	%mm7,%mm0
+	movd	4(%ebp,%ecx,4),%mm1
+	paddq	%mm0,%mm3
+	movd	4(%esi,%ecx,4),%mm0
+	psrlq	$32,%mm2
+	movd	%mm3,28(%esp,%ecx,4)
+	psrlq	$32,%mm3
+	leal	1(%ecx),%ecx
+	cmpl	%ebx,%ecx
+	jl	L0051st
+	pmuludq	%mm4,%mm0
+	pmuludq	%mm5,%mm1
+	paddq	%mm0,%mm2
+	paddq	%mm1,%mm3
+	movq	%mm2,%mm0
+	pand	%mm7,%mm0
+	paddq	%mm0,%mm3
+	movd	%mm3,28(%esp,%ecx,4)
+	psrlq	$32,%mm2
+	psrlq	$32,%mm3
+	paddq	%mm2,%mm3
+	movq	%mm3,32(%esp,%ebx,4)
+	incl	%edx
+L006outer:
+	xorl	%ecx,%ecx
+	movd	(%edi,%edx,4),%mm4
+	movd	(%esi),%mm5
+	movd	32(%esp),%mm6
+	movd	(%ebp),%mm3
+	pmuludq	%mm4,%mm5
+	paddq	%mm6,%mm5
+	movq	%mm5,%mm0
+	movq	%mm5,%mm2
+	pand	%mm7,%mm0
+	pmuludq	20(%esp),%mm5
+	pmuludq	%mm5,%mm3
+	paddq	%mm0,%mm3
+	movd	36(%esp),%mm6
+	movd	4(%ebp),%mm1
+	movd	4(%esi),%mm0
+	psrlq	$32,%mm2
+	psrlq	$32,%mm3
+	paddq	%mm6,%mm2
+	incl	%ecx
+	decl	%ebx
+L007inner:
+	pmuludq	%mm4,%mm0
+	pmuludq	%mm5,%mm1
+	paddq	%mm0,%mm2
+	paddq	%mm1,%mm3
+	movq	%mm2,%mm0
+	movd	36(%esp,%ecx,4),%mm6
+	pand	%mm7,%mm0
+	movd	4(%ebp,%ecx,4),%mm1
+	paddq	%mm0,%mm3
+	movd	4(%esi,%ecx,4),%mm0
+	psrlq	$32,%mm2
+	movd	%mm3,28(%esp,%ecx,4)
+	psrlq	$32,%mm3
+	paddq	%mm6,%mm2
+	decl	%ebx
+	leal	1(%ecx),%ecx
+	jnz	L007inner
+	movl	%ecx,%ebx
+	pmuludq	%mm4,%mm0
+	pmuludq	%mm5,%mm1
+	paddq	%mm0,%mm2
+	paddq	%mm1,%mm3
+	movq	%mm2,%mm0
+	pand	%mm7,%mm0
+	paddq	%mm0,%mm3
+	movd	%mm3,28(%esp,%ecx,4)
+	psrlq	$32,%mm2
+	psrlq	$32,%mm3
+	movd	36(%esp,%ebx,4),%mm6
+	paddq	%mm2,%mm3
+	paddq	%mm6,%mm3
+	movq	%mm3,32(%esp,%ebx,4)
+	leal	1(%edx),%edx
+	cmpl	%ebx,%edx
+	jle	L006outer
+	emms
+	jmp	L008common_tail
+.align	4,0x90
+L004non_sse2:
+	movl	8(%esp),%esi
+	leal	1(%ebx),%ebp
+	movl	12(%esp),%edi
+	xorl	%ecx,%ecx
+	movl	%esi,%edx
+	andl	$1,%ebp
+	subl	%edi,%edx
+	leal	4(%edi,%ebx,4),%eax
+	orl	%edx,%ebp
+	movl	(%edi),%edi
+	jz	L009bn_sqr_mont
+	movl	%eax,28(%esp)
+	movl	(%esi),%eax
+	xorl	%edx,%edx
+.align	4,0x90
+L010mull:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%eax,%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	movl	(%esi,%ecx,4),%eax
+	cmpl	%ebx,%ecx
+	movl	%ebp,28(%esp,%ecx,4)
+	jl	L010mull
+	movl	%edx,%ebp
+	mull	%edi
+	movl	20(%esp),%edi
+	addl	%ebp,%eax
+	movl	16(%esp),%esi
+	adcl	$0,%edx
+	imull	32(%esp),%edi
+	movl	%eax,32(%esp,%ebx,4)
+	xorl	%ecx,%ecx
+	movl	%edx,36(%esp,%ebx,4)
+	movl	%ecx,40(%esp,%ebx,4)
+	movl	(%esi),%eax
+	mull	%edi
+	addl	32(%esp),%eax
+	movl	4(%esi),%eax
+	adcl	$0,%edx
+	incl	%ecx
+	jmp	L0112ndmadd
+.align	4,0x90
+L0121stmadd:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ecx,4),%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	cmpl	%ebx,%ecx
+	movl	%ebp,28(%esp,%ecx,4)
+	jl	L0121stmadd
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ebx,4),%eax
+	movl	20(%esp),%edi
+	adcl	$0,%edx
+	movl	16(%esp),%esi
+	addl	%eax,%ebp
+	adcl	$0,%edx
+	imull	32(%esp),%edi
+	xorl	%ecx,%ecx
+	addl	36(%esp,%ebx,4),%edx
+	movl	%ebp,32(%esp,%ebx,4)
+	adcl	$0,%ecx
+	movl	(%esi),%eax
+	movl	%edx,36(%esp,%ebx,4)
+	movl	%ecx,40(%esp,%ebx,4)
+	mull	%edi
+	addl	32(%esp),%eax
+	movl	4(%esi),%eax
+	adcl	$0,%edx
+	movl	$1,%ecx
+.align	4,0x90
+L0112ndmadd:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ecx,4),%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	cmpl	%ebx,%ecx
+	movl	%ebp,24(%esp,%ecx,4)
+	jl	L0112ndmadd
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ebx,4),%ebp
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	adcl	$0,%edx
+	movl	%ebp,28(%esp,%ebx,4)
+	xorl	%eax,%eax
+	movl	12(%esp),%ecx
+	addl	36(%esp,%ebx,4),%edx
+	adcl	40(%esp,%ebx,4),%eax
+	leal	4(%ecx),%ecx
+	movl	%edx,32(%esp,%ebx,4)
+	cmpl	28(%esp),%ecx
+	movl	%eax,36(%esp,%ebx,4)
+	je	L008common_tail
+	movl	(%ecx),%edi
+	movl	8(%esp),%esi
+	movl	%ecx,12(%esp)
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	movl	(%esi),%eax
+	jmp	L0121stmadd
+.align	4,0x90
+L009bn_sqr_mont:
+	movl	%ebx,(%esp)
+	movl	%ecx,12(%esp)
+	movl	%edi,%eax
+	mull	%edi
+	movl	%eax,32(%esp)
+	movl	%edx,%ebx
+	shrl	$1,%edx
+	andl	$1,%ebx
+	incl	%ecx
+.align	4,0x90
+L013sqr:
+	movl	(%esi,%ecx,4),%eax
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%ebp,%eax
+	leal	1(%ecx),%ecx
+	adcl	$0,%edx
+	leal	(%ebx,%eax,2),%ebp
+	shrl	$31,%eax
+	cmpl	(%esp),%ecx
+	movl	%eax,%ebx
+	movl	%ebp,28(%esp,%ecx,4)
+	jl	L013sqr
+	movl	(%esi,%ecx,4),%eax
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%ebp,%eax
+	movl	20(%esp),%edi
+	adcl	$0,%edx
+	movl	16(%esp),%esi
+	leal	(%ebx,%eax,2),%ebp
+	imull	32(%esp),%edi
+	shrl	$31,%eax
+	movl	%ebp,32(%esp,%ecx,4)
+	leal	(%eax,%edx,2),%ebp
+	movl	(%esi),%eax
+	shrl	$31,%edx
+	movl	%ebp,36(%esp,%ecx,4)
+	movl	%edx,40(%esp,%ecx,4)
+	mull	%edi
+	addl	32(%esp),%eax
+	movl	%ecx,%ebx
+	adcl	$0,%edx
+	movl	4(%esi),%eax
+	movl	$1,%ecx
+.align	4,0x90
+L0143rdmadd:
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ecx,4),%ebp
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	4(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	movl	%ebp,28(%esp,%ecx,4)
+	movl	%edx,%ebp
+	mull	%edi
+	addl	36(%esp,%ecx,4),%ebp
+	leal	2(%ecx),%ecx
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	movl	(%esi,%ecx,4),%eax
+	adcl	$0,%edx
+	cmpl	%ebx,%ecx
+	movl	%ebp,24(%esp,%ecx,4)
+	jl	L0143rdmadd
+	movl	%edx,%ebp
+	mull	%edi
+	addl	32(%esp,%ebx,4),%ebp
+	adcl	$0,%edx
+	addl	%eax,%ebp
+	adcl	$0,%edx
+	movl	%ebp,28(%esp,%ebx,4)
+	movl	12(%esp),%ecx
+	xorl	%eax,%eax
+	movl	8(%esp),%esi
+	addl	36(%esp,%ebx,4),%edx
+	adcl	40(%esp,%ebx,4),%eax
+	movl	%edx,32(%esp,%ebx,4)
+	cmpl	%ebx,%ecx
+	movl	%eax,36(%esp,%ebx,4)
+	je	L008common_tail
+	movl	4(%esi,%ecx,4),%edi
+	leal	1(%ecx),%ecx
+	movl	%edi,%eax
+	movl	%ecx,12(%esp)
+	mull	%edi
+	addl	32(%esp,%ecx,4),%eax
+	adcl	$0,%edx
+	movl	%eax,32(%esp,%ecx,4)
+	xorl	%ebp,%ebp
+	cmpl	%ebx,%ecx
+	leal	1(%ecx),%ecx
+	je	L015sqrlast
+	movl	%edx,%ebx
+	shrl	$1,%edx
+	andl	$1,%ebx
+.align	4,0x90
+L016sqradd:
+	movl	(%esi,%ecx,4),%eax
+	movl	%edx,%ebp
+	mull	%edi
+	addl	%ebp,%eax
+	leal	(%eax,%eax,1),%ebp
+	adcl	$0,%edx
+	shrl	$31,%eax
+	addl	32(%esp,%ecx,4),%ebp
+	leal	1(%ecx),%ecx
+	adcl	$0,%eax
+	addl	%ebx,%ebp
+	adcl	$0,%eax
+	cmpl	(%esp),%ecx
+	movl	%ebp,28(%esp,%ecx,4)
+	movl	%eax,%ebx
+	jle	L016sqradd
+	movl	%edx,%ebp
+	addl	%edx,%edx
+	shrl	$31,%ebp
+	addl	%ebx,%edx
+	adcl	$0,%ebp
+L015sqrlast:
+	movl	20(%esp),%edi
+	movl	16(%esp),%esi
+	imull	32(%esp),%edi
+	addl	32(%esp,%ecx,4),%edx
+	movl	(%esi),%eax
+	adcl	$0,%ebp
+	movl	%edx,32(%esp,%ecx,4)
+	movl	%ebp,36(%esp,%ecx,4)
+	mull	%edi
+	addl	32(%esp),%eax
+	leal	-1(%ecx),%ebx
+	adcl	$0,%edx
+	movl	$1,%ecx
+	movl	4(%esi),%eax
+	jmp	L0143rdmadd
+.align	4,0x90
+L008common_tail:
+	movl	16(%esp),%ebp
+	movl	4(%esp),%edi
+	leal	32(%esp),%esi
+	movl	(%esi),%eax
+	movl	%ebx,%ecx
+	xorl	%edx,%edx
+.align	4,0x90
+L017sub:
+	sbbl	(%ebp,%edx,4),%eax
+	movl	%eax,(%edi,%edx,4)
+	decl	%ecx
+	movl	4(%esi,%edx,4),%eax
+	leal	1(%edx),%edx
+	jge	L017sub
+	sbbl	$0,%eax
+	movl	$-1,%edx
+	xorl	%eax,%edx
+	jmp	L018copy
+.align	4,0x90
+L018copy:
+	movl	32(%esp,%ebx,4),%esi
+	movl	(%edi,%ebx,4),%ebp
+	movl	%ecx,32(%esp,%ebx,4)
+	andl	%eax,%esi
+	andl	%edx,%ebp
+	orl	%esi,%ebp
+	movl	%ebp,(%edi,%ebx,4)
+	decl	%ebx
+	jge	L018copy
+	movl	24(%esp),%esp
+	movl	$1,%eax
+L000just_leave:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.byte	77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
+.byte	112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
+.byte	54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
+.byte	32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
+.byte	111,114,103,62,0
+.comm	_OPENSSL_ia32cap_P,16