Rolls BoringSSL forward

third_party/boringssl/linux-x86/crypto/sha/sha1-586.S

Index: third_party/boringssl/linux-x86/crypto/sha/sha1-586.S
diff --git a/third_party/boringssl/linux-x86/crypto/sha/sha1-586.S b/third_party/boringssl/linux-x86/crypto/sha/sha1-586.S
index 808ccac5178f731a0373308b699d132fb7bd3f14..58d0bc127732abeaaed31c38c79fb15116e8e15c 100644
--- a/third_party/boringssl/linux-x86/crypto/sha/sha1-586.S
+++ b/third_party/boringssl/linux-x86/crypto/sha/sha1-586.S
@@ -23,8 +23,11 @@ sha1_block_data_order:
 	movl	8(%esi),%ecx
 	testl	$16777216,%eax
 	jz	.L001x86
-	testl	$536870912,%ecx
-	jnz	.Lshaext_shortcut
+	andl	$268435456,%edx
+	andl	$1073741824,%eax
+	orl	%edx,%eax
+	cmpl	$1342177280,%eax
+	je	.Lavx_shortcut
 	jmp	.Lssse3_shortcut
 .align	16
 .L001x86:
@@ -1393,177 +1396,6 @@ sha1_block_data_order:
 	popl	%ebp
 	ret
 .size	sha1_block_data_order,.-.L_sha1_block_data_order_begin
-.hidden	_sha1_block_data_order_shaext
-.type	_sha1_block_data_order_shaext,@function
-.align	16
-_sha1_block_data_order_shaext:
-	pushl	%ebp
-	pushl	%ebx
-	pushl	%esi
-	pushl	%edi
-	call	.L003pic_point
-.L003pic_point:
-	popl	%ebp
-	leal	.LK_XX_XX-.L003pic_point(%ebp),%ebp
-.Lshaext_shortcut:
-	movl	20(%esp),%edi
-	movl	%esp,%ebx
-	movl	24(%esp),%esi
-	movl	28(%esp),%ecx
-	subl	$32,%esp
-	movdqu	(%edi),%xmm0
-	movd	16(%edi),%xmm1
-	andl	$-32,%esp
-	movdqa	80(%ebp),%xmm3
-	movdqu	(%esi),%xmm4
-	pshufd	$27,%xmm0,%xmm0
-	movdqu	16(%esi),%xmm5
-	pshufd	$27,%xmm1,%xmm1
-	movdqu	32(%esi),%xmm6
-.byte	102,15,56,0,227
-	movdqu	48(%esi),%xmm7
-.byte	102,15,56,0,235
-.byte	102,15,56,0,243
-.byte	102,15,56,0,251
-	jmp	.L004loop_shaext
-.align	16
-.L004loop_shaext:
-	decl	%ecx
-	leal	64(%esi),%eax
-	movdqa	%xmm1,(%esp)
-	paddd	%xmm4,%xmm1
-	cmovnel	%eax,%esi
-	movdqa	%xmm0,16(%esp)
-.byte	15,56,201,229
-	movdqa	%xmm0,%xmm2
-.byte	15,58,204,193,0
-.byte	15,56,200,213
-	pxor	%xmm6,%xmm4
-.byte	15,56,201,238
-.byte	15,56,202,231
-	movdqa	%xmm0,%xmm1
-.byte	15,58,204,194,0
-.byte	15,56,200,206
-	pxor	%xmm7,%xmm5
-.byte	15,56,202,236
-.byte	15,56,201,247
-	movdqa	%xmm0,%xmm2
-.byte	15,58,204,193,0
-.byte	15,56,200,215
-	pxor	%xmm4,%xmm6
-.byte	15,56,201,252
-.byte	15,56,202,245
-	movdqa	%xmm0,%xmm1
-.byte	15,58,204,194,0
-.byte	15,56,200,204
-	pxor	%xmm5,%xmm7
-.byte	15,56,202,254
-.byte	15,56,201,229
-	movdqa	%xmm0,%xmm2
-.byte	15,58,204,193,0
-.byte	15,56,200,213
-	pxor	%xmm6,%xmm4
-.byte	15,56,201,238
-.byte	15,56,202,231
-	movdqa	%xmm0,%xmm1
-.byte	15,58,204,194,1
-.byte	15,56,200,206
-	pxor	%xmm7,%xmm5
-.byte	15,56,202,236
-.byte	15,56,201,247
-	movdqa	%xmm0,%xmm2
-.byte	15,58,204,193,1
-.byte	15,56,200,215
-	pxor	%xmm4,%xmm6
-.byte	15,56,201,252
-.byte	15,56,202,245
-	movdqa	%xmm0,%xmm1
-.byte	15,58,204,194,1
-.byte	15,56,200,204
-	pxor	%xmm5,%xmm7
-.byte	15,56,202,254
-.byte	15,56,201,229
-	movdqa	%xmm0,%xmm2
-.byte	15,58,204,193,1
-.byte	15,56,200,213
-	pxor	%xmm6,%xmm4
-.byte	15,56,201,238
-.byte	15,56,202,231
-	movdqa	%xmm0,%xmm1
-.byte	15,58,204,194,1
-.byte	15,56,200,206
-	pxor	%xmm7,%xmm5
-.byte	15,56,202,236
-.byte	15,56,201,247
-	movdqa	%xmm0,%xmm2
-.byte	15,58,204,193,2
-.byte	15,56,200,215
-	pxor	%xmm4,%xmm6
-.byte	15,56,201,252
-.byte	15,56,202,245
-	movdqa	%xmm0,%xmm1
-.byte	15,58,204,194,2
-.byte	15,56,200,204
-	pxor	%xmm5,%xmm7
-.byte	15,56,202,254
-.byte	15,56,201,229
-	movdqa	%xmm0,%xmm2
-.byte	15,58,204,193,2
-.byte	15,56,200,213
-	pxor	%xmm6,%xmm4
-.byte	15,56,201,238
-.byte	15,56,202,231
-	movdqa	%xmm0,%xmm1
-.byte	15,58,204,194,2
-.byte	15,56,200,206
-	pxor	%xmm7,%xmm5
-.byte	15,56,202,236
-.byte	15,56,201,247
-	movdqa	%xmm0,%xmm2
-.byte	15,58,204,193,2
-.byte	15,56,200,215
-	pxor	%xmm4,%xmm6
-.byte	15,56,201,252
-.byte	15,56,202,245
-	movdqa	%xmm0,%xmm1
-.byte	15,58,204,194,3
-.byte	15,56,200,204
-	pxor	%xmm5,%xmm7
-.byte	15,56,202,254
-	movdqu	(%esi),%xmm4
-	movdqa	%xmm0,%xmm2
-.byte	15,58,204,193,3
-.byte	15,56,200,213
-	movdqu	16(%esi),%xmm5
-.byte	102,15,56,0,227
-	movdqa	%xmm0,%xmm1
-.byte	15,58,204,194,3
-.byte	15,56,200,206
-	movdqu	32(%esi),%xmm6
-.byte	102,15,56,0,235
-	movdqa	%xmm0,%xmm2
-.byte	15,58,204,193,3
-.byte	15,56,200,215
-	movdqu	48(%esi),%xmm7
-.byte	102,15,56,0,243
-	movdqa	%xmm0,%xmm1
-.byte	15,58,204,194,3
-	movdqa	(%esp),%xmm2
-.byte	102,15,56,0,251
-.byte	15,56,200,202
-	paddd	16(%esp),%xmm0
-	jnz	.L004loop_shaext
-	pshufd	$27,%xmm0,%xmm0
-	pshufd	$27,%xmm1,%xmm1
-	movdqu	%xmm0,(%edi)
-	movd	%xmm1,16(%edi)
-	movl	%ebx,%esp
-	popl	%edi
-	popl	%esi
-	popl	%ebx
-	popl	%ebp
-	ret
-.size	_sha1_block_data_order_shaext,.-_sha1_block_data_order_shaext
 .hidden	_sha1_block_data_order_ssse3
 .type	_sha1_block_data_order_ssse3,@function
 .align	16
@@ -1572,10 +1404,10 @@ _sha1_block_data_order_ssse3:
 	pushl	%ebx
 	pushl	%esi
 	pushl	%edi
-	call	.L005pic_point
-.L005pic_point:
+	call	.L003pic_point
+.L003pic_point:
 	popl	%ebp
-	leal	.LK_XX_XX-.L005pic_point(%ebp),%ebp
+	leal	.LK_XX_XX-.L003pic_point(%ebp),%ebp
 .Lssse3_shortcut:
 	movdqa	(%ebp),%xmm7
 	movdqa	16(%ebp),%xmm0
@@ -1628,9 +1460,9 @@ _sha1_block_data_order_ssse3:
 	xorl	%edx,%ebp
 	pshufd	$238,%xmm0,%xmm4
 	andl	%ebp,%esi
-	jmp	.L006loop
+	jmp	.L004loop
 .align	16
-.L006loop:
+.L004loop:
 	rorl	$2,%ebx
 	xorl	%edx,%esi
 	movl	%eax,%ebp
@@ -2533,7 +2365,7 @@ _sha1_block_data_order_ssse3:
 	addl	%edx,%ecx
 	movl	196(%esp),%ebp
 	cmpl	200(%esp),%ebp
-	je	.L007done
+	je	.L005done
 	movdqa	160(%esp),%xmm7
 	movdqa	176(%esp),%xmm6
 	movdqu	(%ebp),%xmm0
@@ -2668,9 +2500,9 @@ _sha1_block_data_order_ssse3:
 	pshufd	$238,%xmm0,%xmm4
 	andl	%ebx,%esi
 	movl	%ebp,%ebx
-	jmp	.L006loop
+	jmp	.L004loop
 .align	16
-.L007done:
+.L005done:
 	addl	16(%esp),%ebx
 	xorl	%edi,%esi
 	movl	%ecx,%ebp
@@ -2784,6 +2616,1177 @@ _sha1_block_data_order_ssse3:
 	popl	%ebp
 	ret
 .size	_sha1_block_data_order_ssse3,.-_sha1_block_data_order_ssse3
+.hidden	_sha1_block_data_order_avx
+.type	_sha1_block_data_order_avx,@function
+.align	16
+_sha1_block_data_order_avx:
+	pushl	%ebp
+	pushl	%ebx
+	pushl	%esi
+	pushl	%edi
+	call	.L006pic_point
+.L006pic_point:
+	popl	%ebp
+	leal	.LK_XX_XX-.L006pic_point(%ebp),%ebp
+.Lavx_shortcut:
+	vzeroall
+	vmovdqa	(%ebp),%xmm7
+	vmovdqa	16(%ebp),%xmm0
+	vmovdqa	32(%ebp),%xmm1
+	vmovdqa	48(%ebp),%xmm2
+	vmovdqa	64(%ebp),%xmm6
+	movl	20(%esp),%edi
+	movl	24(%esp),%ebp
+	movl	28(%esp),%edx
+	movl	%esp,%esi
+	subl	$208,%esp
+	andl	$-64,%esp
+	vmovdqa	%xmm0,112(%esp)
+	vmovdqa	%xmm1,128(%esp)
+	vmovdqa	%xmm2,144(%esp)
+	shll	$6,%edx
+	vmovdqa	%xmm7,160(%esp)
+	addl	%ebp,%edx
+	vmovdqa	%xmm6,176(%esp)
+	addl	$64,%ebp
+	movl	%edi,192(%esp)
+	movl	%ebp,196(%esp)
+	movl	%edx,200(%esp)
+	movl	%esi,204(%esp)
+	movl	(%edi),%eax
+	movl	4(%edi),%ebx
+	movl	8(%edi),%ecx
+	movl	12(%edi),%edx
+	movl	16(%edi),%edi
+	movl	%ebx,%esi
+	vmovdqu	-64(%ebp),%xmm0
+	vmovdqu	-48(%ebp),%xmm1
+	vmovdqu	-32(%ebp),%xmm2
+	vmovdqu	-16(%ebp),%xmm3
+	vpshufb	%xmm6,%xmm0,%xmm0
+	vpshufb	%xmm6,%xmm1,%xmm1
+	vpshufb	%xmm6,%xmm2,%xmm2
+	vmovdqa	%xmm7,96(%esp)
+	vpshufb	%xmm6,%xmm3,%xmm3
+	vpaddd	%xmm7,%xmm0,%xmm4
+	vpaddd	%xmm7,%xmm1,%xmm5
+	vpaddd	%xmm7,%xmm2,%xmm6
+	vmovdqa	%xmm4,(%esp)
+	movl	%ecx,%ebp
+	vmovdqa	%xmm5,16(%esp)
+	xorl	%edx,%ebp
+	vmovdqa	%xmm6,32(%esp)
+	andl	%ebp,%esi
+	jmp	.L007loop
+.align	16
+.L007loop:
+	shrdl	$2,%ebx,%ebx
+	xorl	%edx,%esi
+	vpalignr	$8,%xmm0,%xmm1,%xmm4
+	movl	%eax,%ebp
+	addl	(%esp),%edi
+	vpaddd	%xmm3,%xmm7,%xmm7
+	vmovdqa	%xmm0,64(%esp)
+	xorl	%ecx,%ebx
+	shldl	$5,%eax,%eax
+	vpsrldq	$4,%xmm3,%xmm6
+	addl	%esi,%edi
+	andl	%ebx,%ebp
+	vpxor	%xmm0,%xmm4,%xmm4
+	xorl	%ecx,%ebx
+	addl	%eax,%edi
+	vpxor	%xmm2,%xmm6,%xmm6
+	shrdl	$7,%eax,%eax
+	xorl	%ecx,%ebp
+	vmovdqa	%xmm7,48(%esp)
+	movl	%edi,%esi
+	addl	4(%esp),%edx
+	vpxor	%xmm6,%xmm4,%xmm4
+	xorl	%ebx,%eax
+	shldl	$5,%edi,%edi
+	addl	%ebp,%edx
+	andl	%eax,%esi
+	vpsrld	$31,%xmm4,%xmm6
+	xorl	%ebx,%eax
+	addl	%edi,%edx
+	shrdl	$7,%edi,%edi
+	xorl	%ebx,%esi
+	vpslldq	$12,%xmm4,%xmm0
+	vpaddd	%xmm4,%xmm4,%xmm4
+	movl	%edx,%ebp
+	addl	8(%esp),%ecx
+	xorl	%eax,%edi
+	shldl	$5,%edx,%edx
+	vpsrld	$30,%xmm0,%xmm7
+	vpor	%xmm6,%xmm4,%xmm4
+	addl	%esi,%ecx
+	andl	%edi,%ebp
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	vpslld	$2,%xmm0,%xmm0
+	shrdl	$7,%edx,%edx
+	xorl	%eax,%ebp
+	vpxor	%xmm7,%xmm4,%xmm4
+	movl	%ecx,%esi
+	addl	12(%esp),%ebx
+	xorl	%edi,%edx
+	shldl	$5,%ecx,%ecx
+	vpxor	%xmm0,%xmm4,%xmm4
+	addl	%ebp,%ebx
+	andl	%edx,%esi
+	vmovdqa	96(%esp),%xmm0
+	xorl	%edi,%edx
+	addl	%ecx,%ebx
+	shrdl	$7,%ecx,%ecx
+	xorl	%edi,%esi
+	vpalignr	$8,%xmm1,%xmm2,%xmm5
+	movl	%ebx,%ebp
+	addl	16(%esp),%eax
+	vpaddd	%xmm4,%xmm0,%xmm0
+	vmovdqa	%xmm1,80(%esp)
+	xorl	%edx,%ecx
+	shldl	$5,%ebx,%ebx
+	vpsrldq	$4,%xmm4,%xmm7
+	addl	%esi,%eax
+	andl	%ecx,%ebp
+	vpxor	%xmm1,%xmm5,%xmm5
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	vpxor	%xmm3,%xmm7,%xmm7
+	shrdl	$7,%ebx,%ebx
+	xorl	%edx,%ebp
+	vmovdqa	%xmm0,(%esp)
+	movl	%eax,%esi
+	addl	20(%esp),%edi
+	vpxor	%xmm7,%xmm5,%xmm5
+	xorl	%ecx,%ebx
+	shldl	$5,%eax,%eax
+	addl	%ebp,%edi
+	andl	%ebx,%esi
+	vpsrld	$31,%xmm5,%xmm7
+	xorl	%ecx,%ebx
+	addl	%eax,%edi
+	shrdl	$7,%eax,%eax
+	xorl	%ecx,%esi
+	vpslldq	$12,%xmm5,%xmm1
+	vpaddd	%xmm5,%xmm5,%xmm5
+	movl	%edi,%ebp
+	addl	24(%esp),%edx
+	xorl	%ebx,%eax
+	shldl	$5,%edi,%edi
+	vpsrld	$30,%xmm1,%xmm0
+	vpor	%xmm7,%xmm5,%xmm5
+	addl	%esi,%edx
+	andl	%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%edi,%edx
+	vpslld	$2,%xmm1,%xmm1
+	shrdl	$7,%edi,%edi
+	xorl	%ebx,%ebp
+	vpxor	%xmm0,%xmm5,%xmm5
+	movl	%edx,%esi
+	addl	28(%esp),%ecx
+	xorl	%eax,%edi
+	shldl	$5,%edx,%edx
+	vpxor	%xmm1,%xmm5,%xmm5
+	addl	%ebp,%ecx
+	andl	%edi,%esi
+	vmovdqa	112(%esp),%xmm1
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	shrdl	$7,%edx,%edx
+	xorl	%eax,%esi
+	vpalignr	$8,%xmm2,%xmm3,%xmm6
+	movl	%ecx,%ebp
+	addl	32(%esp),%ebx
+	vpaddd	%xmm5,%xmm1,%xmm1
+	vmovdqa	%xmm2,96(%esp)
+	xorl	%edi,%edx
+	shldl	$5,%ecx,%ecx
+	vpsrldq	$4,%xmm5,%xmm0
+	addl	%esi,%ebx
+	andl	%edx,%ebp
+	vpxor	%xmm2,%xmm6,%xmm6
+	xorl	%edi,%edx
+	addl	%ecx,%ebx
+	vpxor	%xmm4,%xmm0,%xmm0
+	shrdl	$7,%ecx,%ecx
+	xorl	%edi,%ebp
+	vmovdqa	%xmm1,16(%esp)
+	movl	%ebx,%esi
+	addl	36(%esp),%eax
+	vpxor	%xmm0,%xmm6,%xmm6
+	xorl	%edx,%ecx
+	shldl	$5,%ebx,%ebx
+	addl	%ebp,%eax
+	andl	%ecx,%esi
+	vpsrld	$31,%xmm6,%xmm0
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	shrdl	$7,%ebx,%ebx
+	xorl	%edx,%esi
+	vpslldq	$12,%xmm6,%xmm2
+	vpaddd	%xmm6,%xmm6,%xmm6
+	movl	%eax,%ebp
+	addl	40(%esp),%edi
+	xorl	%ecx,%ebx
+	shldl	$5,%eax,%eax
+	vpsrld	$30,%xmm2,%xmm1
+	vpor	%xmm0,%xmm6,%xmm6
+	addl	%esi,%edi
+	andl	%ebx,%ebp
+	xorl	%ecx,%ebx
+	addl	%eax,%edi
+	vpslld	$2,%xmm2,%xmm2
+	vmovdqa	64(%esp),%xmm0
+	shrdl	$7,%eax,%eax
+	xorl	%ecx,%ebp
+	vpxor	%xmm1,%xmm6,%xmm6
+	movl	%edi,%esi
+	addl	44(%esp),%edx
+	xorl	%ebx,%eax
+	shldl	$5,%edi,%edi
+	vpxor	%xmm2,%xmm6,%xmm6
+	addl	%ebp,%edx
+	andl	%eax,%esi
+	vmovdqa	112(%esp),%xmm2
+	xorl	%ebx,%eax
+	addl	%edi,%edx
+	shrdl	$7,%edi,%edi
+	xorl	%ebx,%esi
+	vpalignr	$8,%xmm3,%xmm4,%xmm7
+	movl	%edx,%ebp
+	addl	48(%esp),%ecx
+	vpaddd	%xmm6,%xmm2,%xmm2
+	vmovdqa	%xmm3,64(%esp)
+	xorl	%eax,%edi
+	shldl	$5,%edx,%edx
+	vpsrldq	$4,%xmm6,%xmm1
+	addl	%esi,%ecx
+	andl	%edi,%ebp
+	vpxor	%xmm3,%xmm7,%xmm7
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	vpxor	%xmm5,%xmm1,%xmm1
+	shrdl	$7,%edx,%edx
+	xorl	%eax,%ebp
+	vmovdqa	%xmm2,32(%esp)
+	movl	%ecx,%esi
+	addl	52(%esp),%ebx
+	vpxor	%xmm1,%xmm7,%xmm7
+	xorl	%edi,%edx
+	shldl	$5,%ecx,%ecx
+	addl	%ebp,%ebx
+	andl	%edx,%esi
+	vpsrld	$31,%xmm7,%xmm1
+	xorl	%edi,%edx
+	addl	%ecx,%ebx
+	shrdl	$7,%ecx,%ecx
+	xorl	%edi,%esi
+	vpslldq	$12,%xmm7,%xmm3
+	vpaddd	%xmm7,%xmm7,%xmm7
+	movl	%ebx,%ebp
+	addl	56(%esp),%eax
+	xorl	%edx,%ecx
+	shldl	$5,%ebx,%ebx
+	vpsrld	$30,%xmm3,%xmm2
+	vpor	%xmm1,%xmm7,%xmm7
+	addl	%esi,%eax
+	andl	%ecx,%ebp
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	vpslld	$2,%xmm3,%xmm3
+	vmovdqa	80(%esp),%xmm1
+	shrdl	$7,%ebx,%ebx
+	xorl	%edx,%ebp
+	vpxor	%xmm2,%xmm7,%xmm7
+	movl	%eax,%esi
+	addl	60(%esp),%edi
+	xorl	%ecx,%ebx
+	shldl	$5,%eax,%eax
+	vpxor	%xmm3,%xmm7,%xmm7
+	addl	%ebp,%edi
+	andl	%ebx,%esi
+	vmovdqa	112(%esp),%xmm3
+	xorl	%ecx,%ebx
+	addl	%eax,%edi
+	vpalignr	$8,%xmm6,%xmm7,%xmm2
+	vpxor	%xmm4,%xmm0,%xmm0
+	shrdl	$7,%eax,%eax
+	xorl	%ecx,%esi
+	movl	%edi,%ebp
+	addl	(%esp),%edx
+	vpxor	%xmm1,%xmm0,%xmm0
+	vmovdqa	%xmm4,80(%esp)
+	xorl	%ebx,%eax
+	shldl	$5,%edi,%edi
+	vmovdqa	%xmm3,%xmm4
+	vpaddd	%xmm7,%xmm3,%xmm3
+	addl	%esi,%edx
+	andl	%eax,%ebp
+	vpxor	%xmm2,%xmm0,%xmm0
+	xorl	%ebx,%eax
+	addl	%edi,%edx
+	shrdl	$7,%edi,%edi
+	xorl	%ebx,%ebp
+	vpsrld	$30,%xmm0,%xmm2
+	vmovdqa	%xmm3,48(%esp)
+	movl	%edx,%esi
+	addl	4(%esp),%ecx
+	xorl	%eax,%edi
+	shldl	$5,%edx,%edx
+	vpslld	$2,%xmm0,%xmm0
+	addl	%ebp,%ecx
+	andl	%edi,%esi
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	shrdl	$7,%edx,%edx
+	xorl	%eax,%esi
+	movl	%ecx,%ebp
+	addl	8(%esp),%ebx
+	vpor	%xmm2,%xmm0,%xmm0
+	xorl	%edi,%edx
+	shldl	$5,%ecx,%ecx
+	vmovdqa	96(%esp),%xmm2
+	addl	%esi,%ebx
+	andl	%edx,%ebp
+	xorl	%edi,%edx
+	addl	%ecx,%ebx
+	addl	12(%esp),%eax
+	xorl	%edi,%ebp
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%ebp,%eax
+	xorl	%edx,%esi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	vpalignr	$8,%xmm7,%xmm0,%xmm3
+	vpxor	%xmm5,%xmm1,%xmm1
+	addl	16(%esp),%edi
+	xorl	%ecx,%esi
+	movl	%eax,%ebp
+	shldl	$5,%eax,%eax
+	vpxor	%xmm2,%xmm1,%xmm1
+	vmovdqa	%xmm5,96(%esp)
+	addl	%esi,%edi
+	xorl	%ecx,%ebp
+	vmovdqa	%xmm4,%xmm5
+	vpaddd	%xmm0,%xmm4,%xmm4
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%edi
+	vpxor	%xmm3,%xmm1,%xmm1
+	addl	20(%esp),%edx
+	xorl	%ebx,%ebp
+	movl	%edi,%esi
+	shldl	$5,%edi,%edi
+	vpsrld	$30,%xmm1,%xmm3
+	vmovdqa	%xmm4,(%esp)
+	addl	%ebp,%edx
+	xorl	%ebx,%esi
+	shrdl	$7,%eax,%eax
+	addl	%edi,%edx
+	vpslld	$2,%xmm1,%xmm1
+	addl	24(%esp),%ecx
+	xorl	%eax,%esi
+	movl	%edx,%ebp
+	shldl	$5,%edx,%edx
+	addl	%esi,%ecx
+	xorl	%eax,%ebp
+	shrdl	$7,%edi,%edi
+	addl	%edx,%ecx
+	vpor	%xmm3,%xmm1,%xmm1
+	addl	28(%esp),%ebx
+	xorl	%edi,%ebp
+	vmovdqa	64(%esp),%xmm3
+	movl	%ecx,%esi
+	shldl	$5,%ecx,%ecx
+	addl	%ebp,%ebx
+	xorl	%edi,%esi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	vpalignr	$8,%xmm0,%xmm1,%xmm4
+	vpxor	%xmm6,%xmm2,%xmm2
+	addl	32(%esp),%eax
+	xorl	%edx,%esi
+	movl	%ebx,%ebp
+	shldl	$5,%ebx,%ebx
+	vpxor	%xmm3,%xmm2,%xmm2
+	vmovdqa	%xmm6,64(%esp)
+	addl	%esi,%eax
+	xorl	%edx,%ebp
+	vmovdqa	128(%esp),%xmm6
+	vpaddd	%xmm1,%xmm5,%xmm5
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	vpxor	%xmm4,%xmm2,%xmm2
+	addl	36(%esp),%edi
+	xorl	%ecx,%ebp
+	movl	%eax,%esi
+	shldl	$5,%eax,%eax
+	vpsrld	$30,%xmm2,%xmm4
+	vmovdqa	%xmm5,16(%esp)
+	addl	%ebp,%edi
+	xorl	%ecx,%esi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%edi
+	vpslld	$2,%xmm2,%xmm2
+	addl	40(%esp),%edx
+	xorl	%ebx,%esi
+	movl	%edi,%ebp
+	shldl	$5,%edi,%edi
+	addl	%esi,%edx
+	xorl	%ebx,%ebp
+	shrdl	$7,%eax,%eax
+	addl	%edi,%edx
+	vpor	%xmm4,%xmm2,%xmm2
+	addl	44(%esp),%ecx
+	xorl	%eax,%ebp
+	vmovdqa	80(%esp),%xmm4
+	movl	%edx,%esi
+	shldl	$5,%edx,%edx
+	addl	%ebp,%ecx
+	xorl	%eax,%esi
+	shrdl	$7,%edi,%edi
+	addl	%edx,%ecx
+	vpalignr	$8,%xmm1,%xmm2,%xmm5
+	vpxor	%xmm7,%xmm3,%xmm3
+	addl	48(%esp),%ebx
+	xorl	%edi,%esi
+	movl	%ecx,%ebp
+	shldl	$5,%ecx,%ecx
+	vpxor	%xmm4,%xmm3,%xmm3
+	vmovdqa	%xmm7,80(%esp)
+	addl	%esi,%ebx
+	xorl	%edi,%ebp
+	vmovdqa	%xmm6,%xmm7
+	vpaddd	%xmm2,%xmm6,%xmm6
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	vpxor	%xmm5,%xmm3,%xmm3
+	addl	52(%esp),%eax
+	xorl	%edx,%ebp
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	vpsrld	$30,%xmm3,%xmm5
+	vmovdqa	%xmm6,32(%esp)
+	addl	%ebp,%eax
+	xorl	%edx,%esi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	vpslld	$2,%xmm3,%xmm3
+	addl	56(%esp),%edi
+	xorl	%ecx,%esi
+	movl	%eax,%ebp
+	shldl	$5,%eax,%eax
+	addl	%esi,%edi
+	xorl	%ecx,%ebp
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%edi
+	vpor	%xmm5,%xmm3,%xmm3
+	addl	60(%esp),%edx
+	xorl	%ebx,%ebp
+	vmovdqa	96(%esp),%xmm5
+	movl	%edi,%esi
+	shldl	$5,%edi,%edi
+	addl	%ebp,%edx
+	xorl	%ebx,%esi
+	shrdl	$7,%eax,%eax
+	addl	%edi,%edx
+	vpalignr	$8,%xmm2,%xmm3,%xmm6
+	vpxor	%xmm0,%xmm4,%xmm4
+	addl	(%esp),%ecx
+	xorl	%eax,%esi
+	movl	%edx,%ebp
+	shldl	$5,%edx,%edx
+	vpxor	%xmm5,%xmm4,%xmm4
+	vmovdqa	%xmm0,96(%esp)
+	addl	%esi,%ecx
+	xorl	%eax,%ebp
+	vmovdqa	%xmm7,%xmm0
+	vpaddd	%xmm3,%xmm7,%xmm7
+	shrdl	$7,%edi,%edi
+	addl	%edx,%ecx
+	vpxor	%xmm6,%xmm4,%xmm4
+	addl	4(%esp),%ebx
+	xorl	%edi,%ebp
+	movl	%ecx,%esi
+	shldl	$5,%ecx,%ecx
+	vpsrld	$30,%xmm4,%xmm6
+	vmovdqa	%xmm7,48(%esp)
+	addl	%ebp,%ebx
+	xorl	%edi,%esi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	vpslld	$2,%xmm4,%xmm4
+	addl	8(%esp),%eax
+	xorl	%edx,%esi
+	movl	%ebx,%ebp
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	xorl	%edx,%ebp
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	vpor	%xmm6,%xmm4,%xmm4
+	addl	12(%esp),%edi
+	xorl	%ecx,%ebp
+	vmovdqa	64(%esp),%xmm6
+	movl	%eax,%esi
+	shldl	$5,%eax,%eax
+	addl	%ebp,%edi
+	xorl	%ecx,%esi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%edi
+	vpalignr	$8,%xmm3,%xmm4,%xmm7
+	vpxor	%xmm1,%xmm5,%xmm5
+	addl	16(%esp),%edx
+	xorl	%ebx,%esi
+	movl	%edi,%ebp
+	shldl	$5,%edi,%edi
+	vpxor	%xmm6,%xmm5,%xmm5
+	vmovdqa	%xmm1,64(%esp)
+	addl	%esi,%edx
+	xorl	%ebx,%ebp
+	vmovdqa	%xmm0,%xmm1
+	vpaddd	%xmm4,%xmm0,%xmm0
+	shrdl	$7,%eax,%eax
+	addl	%edi,%edx
+	vpxor	%xmm7,%xmm5,%xmm5
+	addl	20(%esp),%ecx
+	xorl	%eax,%ebp
+	movl	%edx,%esi
+	shldl	$5,%edx,%edx
+	vpsrld	$30,%xmm5,%xmm7
+	vmovdqa	%xmm0,(%esp)
+	addl	%ebp,%ecx
+	xorl	%eax,%esi
+	shrdl	$7,%edi,%edi
+	addl	%edx,%ecx
+	vpslld	$2,%xmm5,%xmm5
+	addl	24(%esp),%ebx
+	xorl	%edi,%esi
+	movl	%ecx,%ebp
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	xorl	%edi,%ebp
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	vpor	%xmm7,%xmm5,%xmm5
+	addl	28(%esp),%eax
+	vmovdqa	80(%esp),%xmm7
+	shrdl	$7,%ecx,%ecx
+	movl	%ebx,%esi
+	xorl	%edx,%ebp
+	shldl	$5,%ebx,%ebx
+	addl	%ebp,%eax
+	xorl	%ecx,%esi
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	vpalignr	$8,%xmm4,%xmm5,%xmm0
+	vpxor	%xmm2,%xmm6,%xmm6
+	addl	32(%esp),%edi
+	andl	%ecx,%esi
+	xorl	%edx,%ecx
+	shrdl	$7,%ebx,%ebx
+	vpxor	%xmm7,%xmm6,%xmm6
+	vmovdqa	%xmm2,80(%esp)
+	movl	%eax,%ebp
+	xorl	%ecx,%esi
+	vmovdqa	%xmm1,%xmm2
+	vpaddd	%xmm5,%xmm1,%xmm1
+	shldl	$5,%eax,%eax
+	addl	%esi,%edi
+	vpxor	%xmm0,%xmm6,%xmm6
+	xorl	%ebx,%ebp
+	xorl	%ecx,%ebx
+	addl	%eax,%edi
+	addl	36(%esp),%edx
+	vpsrld	$30,%xmm6,%xmm0
+	vmovdqa	%xmm1,16(%esp)
+	andl	%ebx,%ebp
+	xorl	%ecx,%ebx
+	shrdl	$7,%eax,%eax
+	movl	%edi,%esi
+	vpslld	$2,%xmm6,%xmm6
+	xorl	%ebx,%ebp
+	shldl	$5,%edi,%edi
+	addl	%ebp,%edx
+	xorl	%eax,%esi
+	xorl	%ebx,%eax
+	addl	%edi,%edx
+	addl	40(%esp),%ecx
+	andl	%eax,%esi
+	vpor	%xmm0,%xmm6,%xmm6
+	xorl	%ebx,%eax
+	shrdl	$7,%edi,%edi
+	vmovdqa	96(%esp),%xmm0
+	movl	%edx,%ebp
+	xorl	%eax,%esi
+	shldl	$5,%edx,%edx
+	addl	%esi,%ecx
+	xorl	%edi,%ebp
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	addl	44(%esp),%ebx
+	andl	%edi,%ebp
+	xorl	%eax,%edi
+	shrdl	$7,%edx,%edx
+	movl	%ecx,%esi
+	xorl	%edi,%ebp
+	shldl	$5,%ecx,%ecx
+	addl	%ebp,%ebx
+	xorl	%edx,%esi
+	xorl	%edi,%edx
+	addl	%ecx,%ebx
+	vpalignr	$8,%xmm5,%xmm6,%xmm1
+	vpxor	%xmm3,%xmm7,%xmm7
+	addl	48(%esp),%eax
+	andl	%edx,%esi
+	xorl	%edi,%edx
+	shrdl	$7,%ecx,%ecx
+	vpxor	%xmm0,%xmm7,%xmm7
+	vmovdqa	%xmm3,96(%esp)
+	movl	%ebx,%ebp
+	xorl	%edx,%esi
+	vmovdqa	144(%esp),%xmm3
+	vpaddd	%xmm6,%xmm2,%xmm2
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	vpxor	%xmm1,%xmm7,%xmm7
+	xorl	%ecx,%ebp
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	addl	52(%esp),%edi
+	vpsrld	$30,%xmm7,%xmm1
+	vmovdqa	%xmm2,32(%esp)
+	andl	%ecx,%ebp
+	xorl	%edx,%ecx
+	shrdl	$7,%ebx,%ebx
+	movl	%eax,%esi
+	vpslld	$2,%xmm7,%xmm7
+	xorl	%ecx,%ebp
+	shldl	$5,%eax,%eax
+	addl	%ebp,%edi
+	xorl	%ebx,%esi
+	xorl	%ecx,%ebx
+	addl	%eax,%edi
+	addl	56(%esp),%edx
+	andl	%ebx,%esi
+	vpor	%xmm1,%xmm7,%xmm7
+	xorl	%ecx,%ebx
+	shrdl	$7,%eax,%eax
+	vmovdqa	64(%esp),%xmm1
+	movl	%edi,%ebp
+	xorl	%ebx,%esi
+	shldl	$5,%edi,%edi
+	addl	%esi,%edx
+	xorl	%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%edi,%edx
+	addl	60(%esp),%ecx
+	andl	%eax,%ebp
+	xorl	%ebx,%eax
+	shrdl	$7,%edi,%edi
+	movl	%edx,%esi
+	xorl	%eax,%ebp
+	shldl	$5,%edx,%edx
+	addl	%ebp,%ecx
+	xorl	%edi,%esi
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	vpalignr	$8,%xmm6,%xmm7,%xmm2
+	vpxor	%xmm4,%xmm0,%xmm0
+	addl	(%esp),%ebx
+	andl	%edi,%esi
+	xorl	%eax,%edi
+	shrdl	$7,%edx,%edx
+	vpxor	%xmm1,%xmm0,%xmm0
+	vmovdqa	%xmm4,64(%esp)
+	movl	%ecx,%ebp
+	xorl	%edi,%esi
+	vmovdqa	%xmm3,%xmm4
+	vpaddd	%xmm7,%xmm3,%xmm3
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	vpxor	%xmm2,%xmm0,%xmm0
+	xorl	%edx,%ebp
+	xorl	%edi,%edx
+	addl	%ecx,%ebx
+	addl	4(%esp),%eax
+	vpsrld	$30,%xmm0,%xmm2
+	vmovdqa	%xmm3,48(%esp)
+	andl	%edx,%ebp
+	xorl	%edi,%edx
+	shrdl	$7,%ecx,%ecx
+	movl	%ebx,%esi
+	vpslld	$2,%xmm0,%xmm0
+	xorl	%edx,%ebp
+	shldl	$5,%ebx,%ebx
+	addl	%ebp,%eax
+	xorl	%ecx,%esi
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	addl	8(%esp),%edi
+	andl	%ecx,%esi
+	vpor	%xmm2,%xmm0,%xmm0
+	xorl	%edx,%ecx
+	shrdl	$7,%ebx,%ebx
+	vmovdqa	80(%esp),%xmm2
+	movl	%eax,%ebp
+	xorl	%ecx,%esi
+	shldl	$5,%eax,%eax
+	addl	%esi,%edi
+	xorl	%ebx,%ebp
+	xorl	%ecx,%ebx
+	addl	%eax,%edi
+	addl	12(%esp),%edx
+	andl	%ebx,%ebp
+	xorl	%ecx,%ebx
+	shrdl	$7,%eax,%eax
+	movl	%edi,%esi
+	xorl	%ebx,%ebp
+	shldl	$5,%edi,%edi
+	addl	%ebp,%edx
+	xorl	%eax,%esi
+	xorl	%ebx,%eax
+	addl	%edi,%edx
+	vpalignr	$8,%xmm7,%xmm0,%xmm3
+	vpxor	%xmm5,%xmm1,%xmm1
+	addl	16(%esp),%ecx
+	andl	%eax,%esi
+	xorl	%ebx,%eax
+	shrdl	$7,%edi,%edi
+	vpxor	%xmm2,%xmm1,%xmm1
+	vmovdqa	%xmm5,80(%esp)
+	movl	%edx,%ebp
+	xorl	%eax,%esi
+	vmovdqa	%xmm4,%xmm5
+	vpaddd	%xmm0,%xmm4,%xmm4
+	shldl	$5,%edx,%edx
+	addl	%esi,%ecx
+	vpxor	%xmm3,%xmm1,%xmm1
+	xorl	%edi,%ebp
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	addl	20(%esp),%ebx
+	vpsrld	$30,%xmm1,%xmm3
+	vmovdqa	%xmm4,(%esp)
+	andl	%edi,%ebp
+	xorl	%eax,%edi
+	shrdl	$7,%edx,%edx
+	movl	%ecx,%esi
+	vpslld	$2,%xmm1,%xmm1
+	xorl	%edi,%ebp
+	shldl	$5,%ecx,%ecx
+	addl	%ebp,%ebx
+	xorl	%edx,%esi
+	xorl	%edi,%edx
+	addl	%ecx,%ebx
+	addl	24(%esp),%eax
+	andl	%edx,%esi
+	vpor	%xmm3,%xmm1,%xmm1
+	xorl	%edi,%edx
+	shrdl	$7,%ecx,%ecx
+	vmovdqa	96(%esp),%xmm3
+	movl	%ebx,%ebp
+	xorl	%edx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	xorl	%ecx,%ebp
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	addl	28(%esp),%edi
+	andl	%ecx,%ebp
+	xorl	%edx,%ecx
+	shrdl	$7,%ebx,%ebx
+	movl	%eax,%esi
+	xorl	%ecx,%ebp
+	shldl	$5,%eax,%eax
+	addl	%ebp,%edi
+	xorl	%ebx,%esi
+	xorl	%ecx,%ebx
+	addl	%eax,%edi
+	vpalignr	$8,%xmm0,%xmm1,%xmm4
+	vpxor	%xmm6,%xmm2,%xmm2
+	addl	32(%esp),%edx
+	andl	%ebx,%esi
+	xorl	%ecx,%ebx
+	shrdl	$7,%eax,%eax
+	vpxor	%xmm3,%xmm2,%xmm2
+	vmovdqa	%xmm6,96(%esp)
+	movl	%edi,%ebp
+	xorl	%ebx,%esi
+	vmovdqa	%xmm5,%xmm6
+	vpaddd	%xmm1,%xmm5,%xmm5
+	shldl	$5,%edi,%edi
+	addl	%esi,%edx
+	vpxor	%xmm4,%xmm2,%xmm2
+	xorl	%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%edi,%edx
+	addl	36(%esp),%ecx
+	vpsrld	$30,%xmm2,%xmm4
+	vmovdqa	%xmm5,16(%esp)
+	andl	%eax,%ebp
+	xorl	%ebx,%eax
+	shrdl	$7,%edi,%edi
+	movl	%edx,%esi
+	vpslld	$2,%xmm2,%xmm2
+	xorl	%eax,%ebp
+	shldl	$5,%edx,%edx
+	addl	%ebp,%ecx
+	xorl	%edi,%esi
+	xorl	%eax,%edi
+	addl	%edx,%ecx
+	addl	40(%esp),%ebx
+	andl	%edi,%esi
+	vpor	%xmm4,%xmm2,%xmm2
+	xorl	%eax,%edi
+	shrdl	$7,%edx,%edx
+	vmovdqa	64(%esp),%xmm4
+	movl	%ecx,%ebp
+	xorl	%edi,%esi
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	xorl	%edx,%ebp
+	xorl	%edi,%edx
+	addl	%ecx,%ebx
+	addl	44(%esp),%eax
+	andl	%edx,%ebp
+	xorl	%edi,%edx
+	shrdl	$7,%ecx,%ecx
+	movl	%ebx,%esi
+	xorl	%edx,%ebp
+	shldl	$5,%ebx,%ebx
+	addl	%ebp,%eax
+	xorl	%edx,%esi
+	addl	%ebx,%eax
+	vpalignr	$8,%xmm1,%xmm2,%xmm5
+	vpxor	%xmm7,%xmm3,%xmm3
+	addl	48(%esp),%edi
+	xorl	%ecx,%esi
+	movl	%eax,%ebp
+	shldl	$5,%eax,%eax
+	vpxor	%xmm4,%xmm3,%xmm3
+	vmovdqa	%xmm7,64(%esp)
+	addl	%esi,%edi
+	xorl	%ecx,%ebp
+	vmovdqa	%xmm6,%xmm7
+	vpaddd	%xmm2,%xmm6,%xmm6
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%edi
+	vpxor	%xmm5,%xmm3,%xmm3
+	addl	52(%esp),%edx
+	xorl	%ebx,%ebp
+	movl	%edi,%esi
+	shldl	$5,%edi,%edi
+	vpsrld	$30,%xmm3,%xmm5
+	vmovdqa	%xmm6,32(%esp)
+	addl	%ebp,%edx
+	xorl	%ebx,%esi
+	shrdl	$7,%eax,%eax
+	addl	%edi,%edx
+	vpslld	$2,%xmm3,%xmm3
+	addl	56(%esp),%ecx
+	xorl	%eax,%esi
+	movl	%edx,%ebp
+	shldl	$5,%edx,%edx
+	addl	%esi,%ecx
+	xorl	%eax,%ebp
+	shrdl	$7,%edi,%edi
+	addl	%edx,%ecx
+	vpor	%xmm5,%xmm3,%xmm3
+	addl	60(%esp),%ebx
+	xorl	%edi,%ebp
+	movl	%ecx,%esi
+	shldl	$5,%ecx,%ecx
+	addl	%ebp,%ebx
+	xorl	%edi,%esi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	(%esp),%eax
+	vpaddd	%xmm3,%xmm7,%xmm7
+	xorl	%edx,%esi
+	movl	%ebx,%ebp
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	vmovdqa	%xmm7,48(%esp)
+	xorl	%edx,%ebp
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	4(%esp),%edi
+	xorl	%ecx,%ebp
+	movl	%eax,%esi
+	shldl	$5,%eax,%eax
+	addl	%ebp,%edi
+	xorl	%ecx,%esi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%edi
+	addl	8(%esp),%edx
+	xorl	%ebx,%esi
+	movl	%edi,%ebp
+	shldl	$5,%edi,%edi
+	addl	%esi,%edx
+	xorl	%ebx,%ebp
+	shrdl	$7,%eax,%eax
+	addl	%edi,%edx
+	addl	12(%esp),%ecx
+	xorl	%eax,%ebp
+	movl	%edx,%esi
+	shldl	$5,%edx,%edx
+	addl	%ebp,%ecx
+	xorl	%eax,%esi
+	shrdl	$7,%edi,%edi
+	addl	%edx,%ecx
+	movl	196(%esp),%ebp
+	cmpl	200(%esp),%ebp
+	je	.L008done
+	vmovdqa	160(%esp),%xmm7
+	vmovdqa	176(%esp),%xmm6
+	vmovdqu	(%ebp),%xmm0
+	vmovdqu	16(%ebp),%xmm1
+	vmovdqu	32(%ebp),%xmm2
+	vmovdqu	48(%ebp),%xmm3
+	addl	$64,%ebp
+	vpshufb	%xmm6,%xmm0,%xmm0
+	movl	%ebp,196(%esp)
+	vmovdqa	%xmm7,96(%esp)
+	addl	16(%esp),%ebx
+	xorl	%edi,%esi
+	vpshufb	%xmm6,%xmm1,%xmm1
+	movl	%ecx,%ebp
+	shldl	$5,%ecx,%ecx
+	vpaddd	%xmm7,%xmm0,%xmm4
+	addl	%esi,%ebx
+	xorl	%edi,%ebp
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	vmovdqa	%xmm4,(%esp)
+	addl	20(%esp),%eax
+	xorl	%edx,%ebp
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%ebp,%eax
+	xorl	%edx,%esi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	24(%esp),%edi
+	xorl	%ecx,%esi
+	movl	%eax,%ebp
+	shldl	$5,%eax,%eax
+	addl	%esi,%edi
+	xorl	%ecx,%ebp
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%edi
+	addl	28(%esp),%edx
+	xorl	%ebx,%ebp
+	movl	%edi,%esi
+	shldl	$5,%edi,%edi
+	addl	%ebp,%edx
+	xorl	%ebx,%esi
+	shrdl	$7,%eax,%eax
+	addl	%edi,%edx
+	addl	32(%esp),%ecx
+	xorl	%eax,%esi
+	vpshufb	%xmm6,%xmm2,%xmm2
+	movl	%edx,%ebp
+	shldl	$5,%edx,%edx
+	vpaddd	%xmm7,%xmm1,%xmm5
+	addl	%esi,%ecx
+	xorl	%eax,%ebp
+	shrdl	$7,%edi,%edi
+	addl	%edx,%ecx
+	vmovdqa	%xmm5,16(%esp)
+	addl	36(%esp),%ebx
+	xorl	%edi,%ebp
+	movl	%ecx,%esi
+	shldl	$5,%ecx,%ecx
+	addl	%ebp,%ebx
+	xorl	%edi,%esi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	40(%esp),%eax
+	xorl	%edx,%esi
+	movl	%ebx,%ebp
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	xorl	%edx,%ebp
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	44(%esp),%edi
+	xorl	%ecx,%ebp
+	movl	%eax,%esi
+	shldl	$5,%eax,%eax
+	addl	%ebp,%edi
+	xorl	%ecx,%esi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%edi
+	addl	48(%esp),%edx
+	xorl	%ebx,%esi
+	vpshufb	%xmm6,%xmm3,%xmm3
+	movl	%edi,%ebp
+	shldl	$5,%edi,%edi
+	vpaddd	%xmm7,%xmm2,%xmm6
+	addl	%esi,%edx
+	xorl	%ebx,%ebp
+	shrdl	$7,%eax,%eax
+	addl	%edi,%edx
+	vmovdqa	%xmm6,32(%esp)
+	addl	52(%esp),%ecx
+	xorl	%eax,%ebp
+	movl	%edx,%esi
+	shldl	$5,%edx,%edx
+	addl	%ebp,%ecx
+	xorl	%eax,%esi
+	shrdl	$7,%edi,%edi
+	addl	%edx,%ecx
+	addl	56(%esp),%ebx
+	xorl	%edi,%esi
+	movl	%ecx,%ebp
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	xorl	%edi,%ebp
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	60(%esp),%eax
+	xorl	%edx,%ebp
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%ebp,%eax
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	movl	192(%esp),%ebp
+	addl	(%ebp),%eax
+	addl	4(%ebp),%esi
+	addl	8(%ebp),%ecx
+	movl	%eax,(%ebp)
+	addl	12(%ebp),%edx
+	movl	%esi,4(%ebp)
+	addl	16(%ebp),%edi
+	movl	%ecx,%ebx
+	movl	%ecx,8(%ebp)
+	xorl	%edx,%ebx
+	movl	%edx,12(%ebp)
+	movl	%edi,16(%ebp)
+	movl	%esi,%ebp
+	andl	%ebx,%esi
+	movl	%ebp,%ebx
+	jmp	.L007loop
+.align	16
+.L008done:
+	addl	16(%esp),%ebx
+	xorl	%edi,%esi
+	movl	%ecx,%ebp
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	xorl	%edi,%ebp
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	20(%esp),%eax
+	xorl	%edx,%ebp
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%ebp,%eax
+	xorl	%edx,%esi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	24(%esp),%edi
+	xorl	%ecx,%esi
+	movl	%eax,%ebp
+	shldl	$5,%eax,%eax
+	addl	%esi,%edi
+	xorl	%ecx,%ebp
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%edi
+	addl	28(%esp),%edx
+	xorl	%ebx,%ebp
+	movl	%edi,%esi
+	shldl	$5,%edi,%edi
+	addl	%ebp,%edx
+	xorl	%ebx,%esi
+	shrdl	$7,%eax,%eax
+	addl	%edi,%edx
+	addl	32(%esp),%ecx
+	xorl	%eax,%esi
+	movl	%edx,%ebp
+	shldl	$5,%edx,%edx
+	addl	%esi,%ecx
+	xorl	%eax,%ebp
+	shrdl	$7,%edi,%edi
+	addl	%edx,%ecx
+	addl	36(%esp),%ebx
+	xorl	%edi,%ebp
+	movl	%ecx,%esi
+	shldl	$5,%ecx,%ecx
+	addl	%ebp,%ebx
+	xorl	%edi,%esi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	40(%esp),%eax
+	xorl	%edx,%esi
+	movl	%ebx,%ebp
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	xorl	%edx,%ebp
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	44(%esp),%edi
+	xorl	%ecx,%ebp
+	movl	%eax,%esi
+	shldl	$5,%eax,%eax
+	addl	%ebp,%edi
+	xorl	%ecx,%esi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%edi
+	addl	48(%esp),%edx
+	xorl	%ebx,%esi
+	movl	%edi,%ebp
+	shldl	$5,%edi,%edi
+	addl	%esi,%edx
+	xorl	%ebx,%ebp
+	shrdl	$7,%eax,%eax
+	addl	%edi,%edx
+	addl	52(%esp),%ecx
+	xorl	%eax,%ebp
+	movl	%edx,%esi
+	shldl	$5,%edx,%edx
+	addl	%ebp,%ecx
+	xorl	%eax,%esi
+	shrdl	$7,%edi,%edi
+	addl	%edx,%ecx
+	addl	56(%esp),%ebx
+	xorl	%edi,%esi
+	movl	%ecx,%ebp
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	xorl	%edi,%ebp
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	60(%esp),%eax
+	xorl	%edx,%ebp
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%ebp,%eax
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	vzeroall
+	movl	192(%esp),%ebp
+	addl	(%ebp),%eax
+	movl	204(%esp),%esp
+	addl	4(%ebp),%esi
+	addl	8(%ebp),%ecx
+	movl	%eax,(%ebp)
+	addl	12(%ebp),%edx
+	movl	%esi,4(%ebp)
+	addl	16(%ebp),%edi
+	movl	%ecx,8(%ebp)
+	movl	%edx,12(%ebp)
+	movl	%edi,16(%ebp)
+	popl	%edi
+	popl	%esi
+	popl	%ebx
+	popl	%ebp
+	ret
+.size	_sha1_block_data_order_avx,.-_sha1_block_data_order_avx
 .align	64
 .LK_XX_XX:
 .long	1518500249,1518500249,1518500249,1518500249