Breaking Change: merge BoringSSL branch into master

third_party/boringssl/linux-x86/crypto/bn/co-586.S

Index: third_party/boringssl/linux-x86/crypto/bn/co-586.S
diff --git a/third_party/boringssl/linux-x86/crypto/bn/co-586.S b/third_party/boringssl/linux-x86/crypto/bn/co-586.S
new file mode 100644
index 0000000000000000000000000000000000000000..e41c3a1dffdacaea1798eded18743e6344c6a9f1
--- /dev/null
+++ b/third_party/boringssl/linux-x86/crypto/bn/co-586.S
@@ -0,0 +1,1260 @@
+#if defined(__i386__)
+.file	"src/crypto/bn/asm/co-586.S"
+.text
+.globl	bn_mul_comba8
+.hidden	bn_mul_comba8
+.type	bn_mul_comba8,@function
+.align	16
+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
+
+	xorl	%ebp,%ebp
+
+	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
+
+
+	xorl	%ebx,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	(%esi),%eax
+	adcl	%edx,%ebp
+	movl	4(%edi),%edx
+	adcl	$0,%ebx
+
+	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
+
+
+	xorl	%ecx,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	4(%esi),%eax
+	adcl	%edx,%ebx
+	movl	4(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	(%esi),%eax
+	adcl	%edx,%ebx
+	movl	8(%edi),%edx
+	adcl	$0,%ecx
+
+	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
+
+
+	xorl	%ebp,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	8(%esi),%eax
+	adcl	%edx,%ecx
+	movl	4(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	4(%esi),%eax
+	adcl	%edx,%ecx
+	movl	8(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	(%esi),%eax
+	adcl	%edx,%ecx
+	movl	12(%edi),%edx
+	adcl	$0,%ebp
+
+	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
+
+
+	xorl	%ebx,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	12(%esi),%eax
+	adcl	%edx,%ebp
+	movl	4(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	8(%esi),%eax
+	adcl	%edx,%ebp
+	movl	8(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	4(%esi),%eax
+	adcl	%edx,%ebp
+	movl	12(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	(%esi),%eax
+	adcl	%edx,%ebp
+	movl	16(%edi),%edx
+	adcl	$0,%ebx
+
+	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
+
+
+	xorl	%ecx,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	16(%esi),%eax
+	adcl	%edx,%ebx
+	movl	4(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	12(%esi),%eax
+	adcl	%edx,%ebx
+	movl	8(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	8(%esi),%eax
+	adcl	%edx,%ebx
+	movl	12(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	4(%esi),%eax
+	adcl	%edx,%ebx
+	movl	16(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	(%esi),%eax
+	adcl	%edx,%ebx
+	movl	20(%edi),%edx
+	adcl	$0,%ecx
+
+	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
+
+
+	xorl	%ebp,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esi),%eax
+	adcl	%edx,%ecx
+	movl	4(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	16(%esi),%eax
+	adcl	%edx,%ecx
+	movl	8(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	12(%esi),%eax
+	adcl	%edx,%ecx
+	movl	12(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	8(%esi),%eax
+	adcl	%edx,%ecx
+	movl	16(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	4(%esi),%eax
+	adcl	%edx,%ecx
+	movl	20(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	(%esi),%eax
+	adcl	%edx,%ecx
+	movl	24(%edi),%edx
+	adcl	$0,%ebp
+
+	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
+
+
+	xorl	%ebx,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	24(%esi),%eax
+	adcl	%edx,%ebp
+	movl	4(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esi),%eax
+	adcl	%edx,%ebp
+	movl	8(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	16(%esi),%eax
+	adcl	%edx,%ebp
+	movl	12(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	12(%esi),%eax
+	adcl	%edx,%ebp
+	movl	16(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	8(%esi),%eax
+	adcl	%edx,%ebp
+	movl	20(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	4(%esi),%eax
+	adcl	%edx,%ebp
+	movl	24(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	(%esi),%eax
+	adcl	%edx,%ebp
+	movl	28(%edi),%edx
+	adcl	$0,%ebx
+
+	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
+
+
+	xorl	%ecx,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	24(%esi),%eax
+	adcl	%edx,%ebx
+	movl	8(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esi),%eax
+	adcl	%edx,%ebx
+	movl	12(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	16(%esi),%eax
+	adcl	%edx,%ebx
+	movl	16(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	12(%esi),%eax
+	adcl	%edx,%ebx
+	movl	20(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	8(%esi),%eax
+	adcl	%edx,%ebx
+	movl	24(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	4(%esi),%eax
+	adcl	%edx,%ebx
+	movl	28(%edi),%edx
+	adcl	$0,%ecx
+
+	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
+
+
+	xorl	%ebp,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	24(%esi),%eax
+	adcl	%edx,%ecx
+	movl	12(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esi),%eax
+	adcl	%edx,%ecx
+	movl	16(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	16(%esi),%eax
+	adcl	%edx,%ecx
+	movl	20(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	12(%esi),%eax
+	adcl	%edx,%ecx
+	movl	24(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	8(%esi),%eax
+	adcl	%edx,%ecx
+	movl	28(%edi),%edx
+	adcl	$0,%ebp
+
+	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
+
+
+	xorl	%ebx,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	24(%esi),%eax
+	adcl	%edx,%ebp
+	movl	16(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	20(%esi),%eax
+	adcl	%edx,%ebp
+	movl	20(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	16(%esi),%eax
+	adcl	%edx,%ebp
+	movl	24(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	12(%esi),%eax
+	adcl	%edx,%ebp
+	movl	28(%edi),%edx
+	adcl	$0,%ebx
+
+	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
+
+
+	xorl	%ecx,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	24(%esi),%eax
+	adcl	%edx,%ebx
+	movl	20(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esi),%eax
+	adcl	%edx,%ebx
+	movl	24(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	16(%esi),%eax
+	adcl	%edx,%ebx
+	movl	28(%edi),%edx
+	adcl	$0,%ecx
+
+	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
+
+
+	xorl	%ebp,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	24(%esi),%eax
+	adcl	%edx,%ecx
+	movl	24(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esi),%eax
+	adcl	%edx,%ecx
+	movl	28(%edi),%edx
+	adcl	$0,%ebp
+
+	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
+
+
+	xorl	%ebx,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	24(%esi),%eax
+	adcl	%edx,%ebp
+	movl	28(%edi),%edx
+	adcl	$0,%ebx
+
+	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
+
+
+	xorl	%ecx,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	20(%esp),%eax
+	adcl	%edx,%ebx
+	adcl	$0,%ecx
+	movl	%ebp,56(%eax)
+
+
+	movl	%ebx,60(%eax)
+	popl	%ebx
+	popl	%ebp
+	popl	%edi
+	popl	%esi
+	ret
+.size	bn_mul_comba8,.-.L_bn_mul_comba8_begin
+.globl	bn_mul_comba4
+.hidden	bn_mul_comba4
+.type	bn_mul_comba4,@function
+.align	16
+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
+
+	xorl	%ebp,%ebp
+
+	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
+
+
+	xorl	%ebx,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	(%esi),%eax
+	adcl	%edx,%ebp
+	movl	4(%edi),%edx
+	adcl	$0,%ebx
+
+	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
+
+
+	xorl	%ecx,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	4(%esi),%eax
+	adcl	%edx,%ebx
+	movl	4(%edi),%edx
+	adcl	$0,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	(%esi),%eax
+	adcl	%edx,%ebx
+	movl	8(%edi),%edx
+	adcl	$0,%ecx
+
+	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
+
+
+	xorl	%ebp,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	8(%esi),%eax
+	adcl	%edx,%ecx
+	movl	4(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	4(%esi),%eax
+	adcl	%edx,%ecx
+	movl	8(%edi),%edx
+	adcl	$0,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	(%esi),%eax
+	adcl	%edx,%ecx
+	movl	12(%edi),%edx
+	adcl	$0,%ebp
+
+	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
+
+
+	xorl	%ebx,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	8(%esi),%eax
+	adcl	%edx,%ebp
+	movl	8(%edi),%edx
+	adcl	$0,%ebx
+
+	mull	%edx
+	addl	%eax,%ecx
+	movl	4(%esi),%eax
+	adcl	%edx,%ebp
+	movl	12(%edi),%edx
+	adcl	$0,%ebx
+
+	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
+
+
+	xorl	%ecx,%ecx
+
+	mull	%edx
+	addl	%eax,%ebp
+	movl	8(%esi),%eax
+	adcl	%edx,%ebx
+	movl	12(%edi),%edx
+	adcl	$0,%ecx
+
+	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
+
+
+	xorl	%ebp,%ebp
+
+	mull	%edx
+	addl	%eax,%ebx
+	movl	20(%esp),%eax
+	adcl	%edx,%ecx
+	adcl	$0,%ebp
+	movl	%ebx,24(%eax)
+
+
+	movl	%ecx,28(%eax)
+	popl	%ebx
+	popl	%ebp
+	popl	%edi
+	popl	%esi
+	ret
+.size	bn_mul_comba4,.-.L_bn_mul_comba4_begin
+.globl	bn_sqr_comba8
+.hidden	bn_sqr_comba8
+.type	bn_sqr_comba8,@function
+.align	16
+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
+
+	xorl	%ebp,%ebp
+
+	mull	%eax
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	(%esi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,(%edi)
+	movl	4(%esi),%eax
+
+
+	xorl	%ebx,%ebx
+
+	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
+
+
+	xorl	%ecx,%ecx
+
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	4(%esi),%eax
+	adcl	$0,%ecx
+
+	mull	%eax
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	(%esi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,8(%edi)
+	movl	12(%esi),%eax
+
+
+	xorl	%ebp,%ebp
+
+	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
+
+	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
+
+
+	xorl	%ebx,%ebx
+
+	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
+
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	8(%esi),%eax
+	adcl	$0,%ebx
+
+	mull	%eax
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	(%esi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,16(%edi)
+	movl	20(%esi),%eax
+
+
+	xorl	%ecx,%ecx
+
+	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
+
+	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
+
+	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
+
+
+	xorl	%ebp,%ebp
+
+	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
+
+	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
+
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	12(%esi),%eax
+	adcl	$0,%ebp
+
+	mull	%eax
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	(%esi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,24(%edi)
+	movl	28(%esi),%eax
+
+
+	xorl	%ebx,%ebx
+
+	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
+
+	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
+
+	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
+
+	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
+
+
+	xorl	%ecx,%ecx
+
+	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
+
+	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
+
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	16(%esi),%eax
+	adcl	$0,%ecx
+
+	mull	%eax
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	8(%esi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,32(%edi)
+	movl	28(%esi),%eax
+
+
+	xorl	%ebp,%ebp
+
+	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
+
+	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
+
+	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
+
+
+	xorl	%ebx,%ebx
+
+	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
+
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	20(%esi),%eax
+	adcl	$0,%ebx
+
+	mull	%eax
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	16(%esi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,40(%edi)
+	movl	28(%esi),%eax
+
+
+	xorl	%ecx,%ecx
+
+	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
+
+	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
+
+
+	xorl	%ebp,%ebp
+
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebp
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	24(%esi),%eax
+	adcl	$0,%ebp
+
+	mull	%eax
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	24(%esi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,48(%edi)
+	movl	28(%esi),%eax
+
+
+	xorl	%ebx,%ebx
+
+	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)
+
+
+	xorl	%ecx,%ecx
+
+	mull	%eax
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	adcl	$0,%ecx
+	movl	%ebp,56(%edi)
+
+	movl	%ebx,60(%edi)
+	popl	%ebx
+	popl	%ebp
+	popl	%edi
+	popl	%esi
+	ret
+.size	bn_sqr_comba8,.-.L_bn_sqr_comba8_begin
+.globl	bn_sqr_comba4
+.hidden	bn_sqr_comba4
+.type	bn_sqr_comba4,@function
+.align	16
+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
+
+	xorl	%ebp,%ebp
+
+	mull	%eax
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	movl	(%esi),%edx
+	adcl	$0,%ebp
+	movl	%ebx,(%edi)
+	movl	4(%esi),%eax
+
+
+	xorl	%ebx,%ebx
+
+	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
+
+
+	xorl	%ecx,%ecx
+
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ecx
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	4(%esi),%eax
+	adcl	$0,%ecx
+
+	mull	%eax
+	addl	%eax,%ebp
+	adcl	%edx,%ebx
+	movl	(%esi),%edx
+	adcl	$0,%ecx
+	movl	%ebp,8(%edi)
+	movl	12(%esi),%eax
+
+
+	xorl	%ebp,%ebp
+
+	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
+
+	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
+
+
+	xorl	%ebx,%ebx
+
+	mull	%edx
+	addl	%eax,%eax
+	adcl	%edx,%edx
+	adcl	$0,%ebx
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	8(%esi),%eax
+	adcl	$0,%ebx
+
+	mull	%eax
+	addl	%eax,%ecx
+	adcl	%edx,%ebp
+	movl	8(%esi),%edx
+	adcl	$0,%ebx
+	movl	%ecx,16(%edi)
+	movl	12(%esi),%eax
+
+
+	xorl	%ecx,%ecx
+
+	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)
+
+
+	xorl	%ebp,%ebp
+
+	mull	%eax
+	addl	%eax,%ebx
+	adcl	%edx,%ecx
+	adcl	$0,%ebp
+	movl	%ebx,24(%edi)
+
+	movl	%ecx,28(%edi)
+	popl	%ebx
+	popl	%ebp
+	popl	%edi
+	popl	%esi
+	ret
+.size	bn_sqr_comba4,.-.L_bn_sqr_comba4_begin
+#endif