Initial commit of BoringSSL for Android.

3 files changed, 2969 insertions, 0 deletions

diff --git a/mac-x86/crypto/bn/bn-586.S b/mac-x86/crypto/bn/bn-586.S
new file mode 100644
index 0000000..34cf56f
--- /dev/null
+++ b/mac-x86/crypto/bn/bn-586.S
@@ -0,0 +1,1379 @@
+#if defined(__i386__)
+.file	"src/crypto/bn/asm/bn-586.S"
+.text
+.globl	_bn_mul_add_words
+.private_extern	_bn_mul_add_words
+.align	4
+_bn_mul_add_words:
+L_bn_mul_add_words_begin:
+	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	L000maw_finish
+.align	4,0x90
+L001maw_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	L001maw_loop
+L000maw_finish:
+	movl	32(%esp),%ecx
+	andl	$7,%ecx
+	jnz	L002maw_finish2
+	jmp	L003maw_end
+L002maw_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	L003maw_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	L003maw_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	L003maw_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	L003maw_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	L003maw_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	L003maw_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
+L003maw_end:
+	movl	%esi,%eax
+	popl	%ecx
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_bn_mul_words
+.private_extern	_bn_mul_words
+.align	4
+_bn_mul_words:
+L_bn_mul_words_begin:
+	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	L004mw_finish
+L005mw_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	L004mw_finish
+	jmp	L005mw_loop
+L004mw_finish:
+	movl	28(%esp),%ebp
+	andl	$7,%ebp
+	jnz	L006mw_finish2
+	jmp	L007mw_end
+L006mw_finish2:
+	# Tail Round 0 
+	movl	(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,(%edi)
+	movl	%edx,%esi
+	decl	%ebp
+	jz	L007mw_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	L007mw_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	L007mw_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	L007mw_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	L007mw_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	L007mw_end
+	# Tail Round 6 
+	movl	24(%ebx),%eax
+	mull	%ecx
+	addl	%esi,%eax
+	adcl	$0,%edx
+	movl	%eax,24(%edi)
+	movl	%edx,%esi
+L007mw_end:
+	movl	%esi,%eax
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_bn_sqr_words
+.private_extern	_bn_sqr_words
+.align	4
+_bn_sqr_words:
+L_bn_sqr_words_begin:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+
+	movl	20(%esp),%esi
+	movl	24(%esp),%edi
+	movl	28(%esp),%ebx
+	andl	$4294967288,%ebx
+	jz	L008sw_finish
+L009sw_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	L009sw_loop
+L008sw_finish:
+	movl	28(%esp),%ebx
+	andl	$7,%ebx
+	jz	L010sw_end
+	# Tail Round 0 
+	movl	(%edi),%eax
+	mull	%eax
+	movl	%eax,(%esi)
+	decl	%ebx
+	movl	%edx,4(%esi)
+	jz	L010sw_end
+	# Tail Round 1 
+	movl	4(%edi),%eax
+	mull	%eax
+	movl	%eax,8(%esi)
+	decl	%ebx
+	movl	%edx,12(%esi)
+	jz	L010sw_end
+	# Tail Round 2 
+	movl	8(%edi),%eax
+	mull	%eax
+	movl	%eax,16(%esi)
+	decl	%ebx
+	movl	%edx,20(%esi)
+	jz	L010sw_end
+	# Tail Round 3 
+	movl	12(%edi),%eax
+	mull	%eax
+	movl	%eax,24(%esi)
+	decl	%ebx
+	movl	%edx,28(%esi)
+	jz	L010sw_end
+	# Tail Round 4 
+	movl	16(%edi),%eax
+	mull	%eax
+	movl	%eax,32(%esi)
+	decl	%ebx
+	movl	%edx,36(%esi)
+	jz	L010sw_end
+	# Tail Round 5 
+	movl	20(%edi),%eax
+	mull	%eax
+	movl	%eax,40(%esi)
+	decl	%ebx
+	movl	%edx,44(%esi)
+	jz	L010sw_end
+	# Tail Round 6 
+	movl	24(%edi),%eax
+	mull	%eax
+	movl	%eax,48(%esi)
+	movl	%edx,52(%esi)
+L010sw_end:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_bn_div_words
+.private_extern	_bn_div_words
+.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
+.private_extern	_bn_add_words
+.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	L011aw_finish
+L012aw_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	L012aw_loop
+L011aw_finish:
+	movl	32(%esp),%ebp
+	andl	$7,%ebp
+	jz	L013aw_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	L013aw_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	L013aw_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	L013aw_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	L013aw_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	L013aw_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	L013aw_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)
+L013aw_end:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_bn_sub_words
+.private_extern	_bn_sub_words
+.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	L014aw_finish
+L015aw_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	L015aw_loop
+L014aw_finish:
+	movl	32(%esp),%ebp
+	andl	$7,%ebp
+	jz	L016aw_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	L016aw_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	L016aw_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	L016aw_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	L016aw_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	L016aw_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	L016aw_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)
+L016aw_end:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.globl	_bn_sub_part_words
+.private_extern	_bn_sub_part_words
+.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	L017aw_finish
+L018aw_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	L018aw_loop
+L017aw_finish:
+	movl	32(%esp),%ebp
+	andl	$7,%ebp
+	jz	L019aw_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	L019aw_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	L019aw_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	L019aw_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	L019aw_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	L019aw_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	L019aw_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
+L019aw_end:
+	cmpl	$0,36(%esp)
+	je	L020pw_end
+	movl	36(%esp),%ebp
+	cmpl	$0,%ebp
+	je	L020pw_end
+	jge	L021pw_pos
+	# pw_neg 
+	movl	$0,%edx
+	subl	%ebp,%edx
+	movl	%edx,%ebp
+	andl	$4294967288,%ebp
+	jz	L022pw_neg_finish
+L023pw_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	L023pw_neg_loop
+L022pw_neg_finish:
+	movl	36(%esp),%edx
+	movl	$0,%ebp
+	subl	%edx,%ebp
+	andl	$7,%ebp
+	jz	L020pw_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	L020pw_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	L020pw_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	L020pw_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	L020pw_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	L020pw_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	L020pw_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	L020pw_end
+L021pw_pos:
+	andl	$4294967288,%ebp
+	jz	L024pw_pos_finish
+L025pw_pos_loop:
+	# dl>0 Round 0 
+	movl	(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,(%ebx)
+	jnc	L026pw_nc0
+	# dl>0 Round 1 
+	movl	4(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,4(%ebx)
+	jnc	L027pw_nc1
+	# dl>0 Round 2 
+	movl	8(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,8(%ebx)
+	jnc	L028pw_nc2
+	# dl>0 Round 3 
+	movl	12(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,12(%ebx)
+	jnc	L029pw_nc3
+	# dl>0 Round 4 
+	movl	16(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,16(%ebx)
+	jnc	L030pw_nc4
+	# dl>0 Round 5 
+	movl	20(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,20(%ebx)
+	jnc	L031pw_nc5
+	# dl>0 Round 6 
+	movl	24(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,24(%ebx)
+	jnc	L032pw_nc6
+	# dl>0 Round 7 
+	movl	28(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,28(%ebx)
+	jnc	L033pw_nc7
+
+	addl	$32,%esi
+	addl	$32,%ebx
+	subl	$8,%ebp
+	jnz	L025pw_pos_loop
+L024pw_pos_finish:
+	movl	36(%esp),%ebp
+	andl	$7,%ebp
+	jz	L020pw_end
+	# dl>0 Tail Round 0 
+	movl	(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,(%ebx)
+	jnc	L034pw_tail_nc0
+	decl	%ebp
+	jz	L020pw_end
+	# dl>0 Tail Round 1 
+	movl	4(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,4(%ebx)
+	jnc	L035pw_tail_nc1
+	decl	%ebp
+	jz	L020pw_end
+	# dl>0 Tail Round 2 
+	movl	8(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,8(%ebx)
+	jnc	L036pw_tail_nc2
+	decl	%ebp
+	jz	L020pw_end
+	# dl>0 Tail Round 3 
+	movl	12(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,12(%ebx)
+	jnc	L037pw_tail_nc3
+	decl	%ebp
+	jz	L020pw_end
+	# dl>0 Tail Round 4 
+	movl	16(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,16(%ebx)
+	jnc	L038pw_tail_nc4
+	decl	%ebp
+	jz	L020pw_end
+	# dl>0 Tail Round 5 
+	movl	20(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,20(%ebx)
+	jnc	L039pw_tail_nc5
+	decl	%ebp
+	jz	L020pw_end
+	# dl>0 Tail Round 6 
+	movl	24(%esi),%ecx
+	subl	%eax,%ecx
+	movl	%ecx,24(%ebx)
+	jnc	L040pw_tail_nc6
+	movl	$1,%eax
+	jmp	L020pw_end
+L041pw_nc_loop:
+	movl	(%esi),%ecx
+	movl	%ecx,(%ebx)
+L026pw_nc0:
+	movl	4(%esi),%ecx
+	movl	%ecx,4(%ebx)
+L027pw_nc1:
+	movl	8(%esi),%ecx
+	movl	%ecx,8(%ebx)
+L028pw_nc2:
+	movl	12(%esi),%ecx
+	movl	%ecx,12(%ebx)
+L029pw_nc3:
+	movl	16(%esi),%ecx
+	movl	%ecx,16(%ebx)
+L030pw_nc4:
+	movl	20(%esi),%ecx
+	movl	%ecx,20(%ebx)
+L031pw_nc5:
+	movl	24(%esi),%ecx
+	movl	%ecx,24(%ebx)
+L032pw_nc6:
+	movl	28(%esi),%ecx
+	movl	%ecx,28(%ebx)
+L033pw_nc7:
+
+	addl	$32,%esi
+	addl	$32,%ebx
+	subl	$8,%ebp
+	jnz	L041pw_nc_loop
+	movl	36(%esp),%ebp
+	andl	$7,%ebp
+	jz	L042pw_nc_end
+	movl	(%esi),%ecx
+	movl	%ecx,(%ebx)
+L034pw_tail_nc0:
+	decl	%ebp
+	jz	L042pw_nc_end
+	movl	4(%esi),%ecx
+	movl	%ecx,4(%ebx)
+L035pw_tail_nc1:
+	decl	%ebp
+	jz	L042pw_nc_end
+	movl	8(%esi),%ecx
+	movl	%ecx,8(%ebx)
+L036pw_tail_nc2:
+	decl	%ebp
+	jz	L042pw_nc_end
+	movl	12(%esi),%ecx
+	movl	%ecx,12(%ebx)
+L037pw_tail_nc3:
+	decl	%ebp
+	jz	L042pw_nc_end
+	movl	16(%esi),%ecx
+	movl	%ecx,16(%ebx)
+L038pw_tail_nc4:
+	decl	%ebp
+	jz	L042pw_nc_end
+	movl	20(%esi),%ecx
+	movl	%ecx,20(%ebx)
+L039pw_tail_nc5:
+	decl	%ebp
+	jz	L042pw_nc_end
+	movl	24(%esi),%ecx
+	movl	%ecx,24(%ebx)
+L040pw_tail_nc6:
+L042pw_nc_end:
+	movl	$0,%eax
+L020pw_end:
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+#endif
diff --git a/mac-x86/crypto/bn/co-586.S b/mac-x86/crypto/bn/co-586.S
new file mode 100644
index 0000000..7ce8e79
--- /dev/null
+++ b/mac-x86/crypto/bn/co-586.S
@@ -0,0 +1,1252 @@
+#if defined(__i386__)
+.file	"src/crypto/bn/asm/co-586.S"
+.text
+.globl	_bn_mul_comba8
+.private_extern	_bn_mul_comba8
+.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
+.private_extern	_bn_mul_comba4
+.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
+.private_extern	_bn_sqr_comba8
+.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
+.private_extern	_bn_sqr_comba4
+.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
+#endif
diff --git a/mac-x86/crypto/bn/x86-mont.S b/mac-x86/crypto/bn/x86-mont.S
new file mode 100644
index 0000000..1b79c5f
--- /dev/null
+++ b/mac-x86/crypto/bn/x86-mont.S
@@ -0,0 +1,338 @@
+#if defined(__i386__)
+.file	"src/crypto/bn/asm/x86-mont.S"
+.text
+.globl	_bn_mul_mont
+.private_extern	_bn_mul_mont
+.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
+	movl	%esp,%ebp
+	addl	$2,%edi
+	negl	%edi
+	leal	-32(%esp,%edi,4),%esp
+	negl	%edi
+	movl	%esp,%eax
+	subl	%edx,%eax
+	andl	$2047,%eax
+	subl	%eax,%esp
+	xorl	%esp,%edx
+	andl	$2048,%edx
+	xorl	$2048,%edx
+	subl	%edx,%esp
+	andl	$-64,%esp
+	movl	(%esi),%eax
+	movl	4(%esi),%ebx
+	movl	8(%esi),%ecx
+	movl	12(%esi),%edx
+	movl	16(%esi),%esi
+	movl	(%esi),%esi
+	movl	%eax,4(%esp)
+	movl	%ebx,8(%esp)
+	movl	%ecx,12(%esp)
+	movl	%edx,16(%esp)
+	movl	%esi,20(%esp)
+	leal	-3(%edi),%ebx
+	movl	%ebp,24(%esp)
+	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	L001bn_sqr_mont
+	movl	%eax,28(%esp)
+	movl	(%esi),%eax
+	xorl	%edx,%edx
+.align	4,0x90
+L002mull:
+	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	L002mull
+	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	L0032ndmadd
+.align	4,0x90
+L0041stmadd:
+	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	L0041stmadd
+	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
+L0032ndmadd:
+	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	L0032ndmadd
+	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	L005common_tail
+	movl	(%ecx),%edi
+	movl	8(%esp),%esi
+	movl	%ecx,12(%esp)
+	xorl	%ecx,%ecx
+	xorl	%edx,%edx
+	movl	(%esi),%eax
+	jmp	L0041stmadd
+.align	4,0x90
+L001bn_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
+L006sqr:
+	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	L006sqr
+	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
+L0073rdmadd:
+	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	L0073rdmadd
+	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	L005common_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	L008sqrlast
+	movl	%edx,%ebx
+	shrl	$1,%edx
+	andl	$1,%ebx
+.align	4,0x90
+L009sqradd:
+	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	L009sqradd
+	movl	%edx,%ebp
+	addl	%edx,%edx
+	shrl	$31,%ebp
+	addl	%ebx,%edx
+	adcl	$0,%ebp
+L008sqrlast:
+	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	L0073rdmadd
+.align	4,0x90
+L005common_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
+L010sub:
+	sbbl	(%ebp,%edx,4),%eax
+	movl	%eax,(%edi,%edx,4)
+	decl	%ecx
+	movl	4(%esi,%edx,4),%eax
+	leal	1(%edx),%edx
+	jge	L010sub
+	sbbl	$0,%eax
+.align	4,0x90
+L011copy:
+	movl	(%esi,%ebx,4),%edx
+	movl	(%edi,%ebx,4),%ebp
+	xorl	%ebp,%edx
+	andl	%eax,%edx
+	xorl	%ebp,%edx
+	movl	%ecx,(%esi,%ebx,4)
+	movl	%edx,(%edi,%ebx,4)
+	decl	%ebx
+	jge	L011copy
+	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
+#endif