5 files changed, 56 insertions, 8896 deletions

diff --git a/secure/lib/libcrypto/i386/chacha-x86.S b/secure/lib/libcrypto/i386/chacha-x86.S
index d6b2936a5381..566285310e06 100644
--- a/secure/lib/libcrypto/i386/chacha-x86.S
+++ b/secure/lib/libcrypto/i386/chacha-x86.S
@@ -385,8 +385,6 @@ ChaCha20_ssse3:
 	pushl	%esi
 	pushl	%edi
 .Lssse3_shortcut:
-	testl	$2048,4(%ebp)
-	jnz	.Lxop_shortcut
 	movl	20(%esp),%edi
 	movl	24(%esp),%esi
 	movl	28(%esp),%ecx
@@ -530,484 +528,6 @@ ChaCha20_ssse3:
 .byte	44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32
 .byte	60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111
 .byte	114,103,62,0
-.globl	ChaCha20_xop
-.type	ChaCha20_xop,@function
-.align	16
-ChaCha20_xop:
-.L_ChaCha20_xop_begin:
-	pushl	%ebp
-	pushl	%ebx
-	pushl	%esi
-	pushl	%edi
-.Lxop_shortcut:
-	movl	20(%esp),%edi
-	movl	24(%esp),%esi
-	movl	28(%esp),%ecx
-	movl	32(%esp),%edx
-	movl	36(%esp),%ebx
-	vzeroupper
-	movl	%esp,%ebp
-	subl	$524,%esp
-	andl	$-64,%esp
-	movl	%ebp,512(%esp)
-	leal	.Lssse3_data-.Lpic_point(%eax),%eax
-	vmovdqu	(%ebx),%xmm3
-	cmpl	$256,%ecx
-	jb	.L0141x
-	movl	%edx,516(%esp)
-	movl	%ebx,520(%esp)
-	subl	$256,%ecx
-	leal	384(%esp),%ebp
-	vmovdqu	(%edx),%xmm7
-	vpshufd	$0,%xmm3,%xmm0
-	vpshufd	$85,%xmm3,%xmm1
-	vpshufd	$170,%xmm3,%xmm2
-	vpshufd	$255,%xmm3,%xmm3
-	vpaddd	48(%eax),%xmm0,%xmm0
-	vpshufd	$0,%xmm7,%xmm4
-	vpshufd	$85,%xmm7,%xmm5
-	vpsubd	64(%eax),%xmm0,%xmm0
-	vpshufd	$170,%xmm7,%xmm6
-	vpshufd	$255,%xmm7,%xmm7
-	vmovdqa	%xmm0,64(%ebp)
-	vmovdqa	%xmm1,80(%ebp)
-	vmovdqa	%xmm2,96(%ebp)
-	vmovdqa	%xmm3,112(%ebp)
-	vmovdqu	16(%edx),%xmm3
-	vmovdqa	%xmm4,-64(%ebp)
-	vmovdqa	%xmm5,-48(%ebp)
-	vmovdqa	%xmm6,-32(%ebp)
-	vmovdqa	%xmm7,-16(%ebp)
-	vmovdqa	32(%eax),%xmm7
-	leal	128(%esp),%ebx
-	vpshufd	$0,%xmm3,%xmm0
-	vpshufd	$85,%xmm3,%xmm1
-	vpshufd	$170,%xmm3,%xmm2
-	vpshufd	$255,%xmm3,%xmm3
-	vpshufd	$0,%xmm7,%xmm4
-	vpshufd	$85,%xmm7,%xmm5
-	vpshufd	$170,%xmm7,%xmm6
-	vpshufd	$255,%xmm7,%xmm7
-	vmovdqa	%xmm0,(%ebp)
-	vmovdqa	%xmm1,16(%ebp)
-	vmovdqa	%xmm2,32(%ebp)
-	vmovdqa	%xmm3,48(%ebp)
-	vmovdqa	%xmm4,-128(%ebp)
-	vmovdqa	%xmm5,-112(%ebp)
-	vmovdqa	%xmm6,-96(%ebp)
-	vmovdqa	%xmm7,-80(%ebp)
-	leal	128(%esi),%esi
-	leal	128(%edi),%edi
-	jmp	.L015outer_loop
-.align	32
-.L015outer_loop:
-	vmovdqa	-112(%ebp),%xmm1
-	vmovdqa	-96(%ebp),%xmm2
-	vmovdqa	-80(%ebp),%xmm3
-	vmovdqa	-48(%ebp),%xmm5
-	vmovdqa	-32(%ebp),%xmm6
-	vmovdqa	-16(%ebp),%xmm7
-	vmovdqa	%xmm1,-112(%ebx)
-	vmovdqa	%xmm2,-96(%ebx)
-	vmovdqa	%xmm3,-80(%ebx)
-	vmovdqa	%xmm5,-48(%ebx)
-	vmovdqa	%xmm6,-32(%ebx)
-	vmovdqa	%xmm7,-16(%ebx)
-	vmovdqa	32(%ebp),%xmm2
-	vmovdqa	48(%ebp),%xmm3
-	vmovdqa	64(%ebp),%xmm4
-	vmovdqa	80(%ebp),%xmm5
-	vmovdqa	96(%ebp),%xmm6
-	vmovdqa	112(%ebp),%xmm7
-	vpaddd	64(%eax),%xmm4,%xmm4
-	vmovdqa	%xmm2,32(%ebx)
-	vmovdqa	%xmm3,48(%ebx)
-	vmovdqa	%xmm4,64(%ebx)
-	vmovdqa	%xmm5,80(%ebx)
-	vmovdqa	%xmm6,96(%ebx)
-	vmovdqa	%xmm7,112(%ebx)
-	vmovdqa	%xmm4,64(%ebp)
-	vmovdqa	-128(%ebp),%xmm0
-	vmovdqa	%xmm4,%xmm6
-	vmovdqa	-64(%ebp),%xmm3
-	vmovdqa	(%ebp),%xmm4
-	vmovdqa	16(%ebp),%xmm5
-	movl	$10,%edx
-	nop
-.align	32
-.L016loop:
-	vpaddd	%xmm3,%xmm0,%xmm0
-	vpxor	%xmm0,%xmm6,%xmm6
-.byte	143,232,120,194,246,16
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vpxor	%xmm4,%xmm3,%xmm2
-	vmovdqa	-112(%ebx),%xmm1
-.byte	143,232,120,194,210,12
-	vmovdqa	-48(%ebx),%xmm3
-	vpaddd	%xmm2,%xmm0,%xmm0
-	vmovdqa	80(%ebx),%xmm7
-	vpxor	%xmm0,%xmm6,%xmm6
-	vpaddd	%xmm3,%xmm1,%xmm1
-.byte	143,232,120,194,246,8
-	vmovdqa	%xmm0,-128(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	%xmm6,64(%ebx)
-	vpxor	%xmm4,%xmm2,%xmm2
-	vpxor	%xmm1,%xmm7,%xmm7
-.byte	143,232,120,194,210,7
-	vmovdqa	%xmm4,(%ebx)
-.byte	143,232,120,194,255,16
-	vmovdqa	%xmm2,-64(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vmovdqa	32(%ebx),%xmm4
-	vpxor	%xmm5,%xmm3,%xmm3
-	vmovdqa	-96(%ebx),%xmm0
-.byte	143,232,120,194,219,12
-	vmovdqa	-32(%ebx),%xmm2
-	vpaddd	%xmm3,%xmm1,%xmm1
-	vmovdqa	96(%ebx),%xmm6
-	vpxor	%xmm1,%xmm7,%xmm7
-	vpaddd	%xmm2,%xmm0,%xmm0
-.byte	143,232,120,194,255,8
-	vmovdqa	%xmm1,-112(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vmovdqa	%xmm7,80(%ebx)
-	vpxor	%xmm5,%xmm3,%xmm3
-	vpxor	%xmm0,%xmm6,%xmm6
-.byte	143,232,120,194,219,7
-	vmovdqa	%xmm5,16(%ebx)
-.byte	143,232,120,194,246,16
-	vmovdqa	%xmm3,-48(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	48(%ebx),%xmm5
-	vpxor	%xmm4,%xmm2,%xmm2
-	vmovdqa	-80(%ebx),%xmm1
-.byte	143,232,120,194,210,12
-	vmovdqa	-16(%ebx),%xmm3
-	vpaddd	%xmm2,%xmm0,%xmm0
-	vmovdqa	112(%ebx),%xmm7
-	vpxor	%xmm0,%xmm6,%xmm6
-	vpaddd	%xmm3,%xmm1,%xmm1
-.byte	143,232,120,194,246,8
-	vmovdqa	%xmm0,-96(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	%xmm6,96(%ebx)
-	vpxor	%xmm4,%xmm2,%xmm2
-	vpxor	%xmm1,%xmm7,%xmm7
-.byte	143,232,120,194,210,7
-.byte	143,232,120,194,255,16
-	vmovdqa	%xmm2,-32(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vpxor	%xmm5,%xmm3,%xmm3
-	vmovdqa	-128(%ebx),%xmm0
-.byte	143,232,120,194,219,12
-	vmovdqa	-48(%ebx),%xmm2
-	vpaddd	%xmm3,%xmm1,%xmm1
-	vpxor	%xmm1,%xmm7,%xmm7
-	vpaddd	%xmm2,%xmm0,%xmm0
-.byte	143,232,120,194,255,8
-	vmovdqa	%xmm1,-80(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vpxor	%xmm5,%xmm3,%xmm3
-	vpxor	%xmm0,%xmm7,%xmm6
-.byte	143,232,120,194,219,7
-.byte	143,232,120,194,246,16
-	vmovdqa	%xmm3,-16(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vpxor	%xmm4,%xmm2,%xmm2
-	vmovdqa	-112(%ebx),%xmm1
-.byte	143,232,120,194,210,12
-	vmovdqa	-32(%ebx),%xmm3
-	vpaddd	%xmm2,%xmm0,%xmm0
-	vmovdqa	64(%ebx),%xmm7
-	vpxor	%xmm0,%xmm6,%xmm6
-	vpaddd	%xmm3,%xmm1,%xmm1
-.byte	143,232,120,194,246,8
-	vmovdqa	%xmm0,-128(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	%xmm6,112(%ebx)
-	vpxor	%xmm4,%xmm2,%xmm2
-	vpxor	%xmm1,%xmm7,%xmm7
-.byte	143,232,120,194,210,7
-	vmovdqa	%xmm4,32(%ebx)
-.byte	143,232,120,194,255,16
-	vmovdqa	%xmm2,-48(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vmovdqa	(%ebx),%xmm4
-	vpxor	%xmm5,%xmm3,%xmm3
-	vmovdqa	-96(%ebx),%xmm0
-.byte	143,232,120,194,219,12
-	vmovdqa	-16(%ebx),%xmm2
-	vpaddd	%xmm3,%xmm1,%xmm1
-	vmovdqa	80(%ebx),%xmm6
-	vpxor	%xmm1,%xmm7,%xmm7
-	vpaddd	%xmm2,%xmm0,%xmm0
-.byte	143,232,120,194,255,8
-	vmovdqa	%xmm1,-112(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vmovdqa	%xmm7,64(%ebx)
-	vpxor	%xmm5,%xmm3,%xmm3
-	vpxor	%xmm0,%xmm6,%xmm6
-.byte	143,232,120,194,219,7
-	vmovdqa	%xmm5,48(%ebx)
-.byte	143,232,120,194,246,16
-	vmovdqa	%xmm3,-32(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	16(%ebx),%xmm5
-	vpxor	%xmm4,%xmm2,%xmm2
-	vmovdqa	-80(%ebx),%xmm1
-.byte	143,232,120,194,210,12
-	vmovdqa	-64(%ebx),%xmm3
-	vpaddd	%xmm2,%xmm0,%xmm0
-	vmovdqa	96(%ebx),%xmm7
-	vpxor	%xmm0,%xmm6,%xmm6
-	vpaddd	%xmm3,%xmm1,%xmm1
-.byte	143,232,120,194,246,8
-	vmovdqa	%xmm0,-96(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	%xmm6,80(%ebx)
-	vpxor	%xmm4,%xmm2,%xmm2
-	vpxor	%xmm1,%xmm7,%xmm7
-.byte	143,232,120,194,210,7
-.byte	143,232,120,194,255,16
-	vmovdqa	%xmm2,-16(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vpxor	%xmm5,%xmm3,%xmm3
-	vmovdqa	-128(%ebx),%xmm0
-.byte	143,232,120,194,219,12
-	vpaddd	%xmm3,%xmm1,%xmm1
-	vmovdqa	64(%ebx),%xmm6
-	vpxor	%xmm1,%xmm7,%xmm7
-.byte	143,232,120,194,255,8
-	vmovdqa	%xmm1,-80(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vmovdqa	%xmm7,96(%ebx)
-	vpxor	%xmm5,%xmm3,%xmm3
-.byte	143,232,120,194,219,7
-	decl	%edx
-	jnz	.L016loop
-	vmovdqa	%xmm3,-64(%ebx)
-	vmovdqa	%xmm4,(%ebx)
-	vmovdqa	%xmm5,16(%ebx)
-	vmovdqa	%xmm6,64(%ebx)
-	vmovdqa	%xmm7,96(%ebx)
-	vmovdqa	-112(%ebx),%xmm1
-	vmovdqa	-96(%ebx),%xmm2
-	vmovdqa	-80(%ebx),%xmm3
-	vpaddd	-128(%ebp),%xmm0,%xmm0
-	vpaddd	-112(%ebp),%xmm1,%xmm1
-	vpaddd	-96(%ebp),%xmm2,%xmm2
-	vpaddd	-80(%ebp),%xmm3,%xmm3
-	vpunpckldq	%xmm1,%xmm0,%xmm6
-	vpunpckldq	%xmm3,%xmm2,%xmm7
-	vpunpckhdq	%xmm1,%xmm0,%xmm0
-	vpunpckhdq	%xmm3,%xmm2,%xmm2
-	vpunpcklqdq	%xmm7,%xmm6,%xmm1
-	vpunpckhqdq	%xmm7,%xmm6,%xmm6
-	vpunpcklqdq	%xmm2,%xmm0,%xmm7
-	vpunpckhqdq	%xmm2,%xmm0,%xmm3
-	vpxor	-128(%esi),%xmm1,%xmm4
-	vpxor	-64(%esi),%xmm6,%xmm5
-	vpxor	(%esi),%xmm7,%xmm6
-	vpxor	64(%esi),%xmm3,%xmm7
-	leal	16(%esi),%esi
-	vmovdqa	-64(%ebx),%xmm0
-	vmovdqa	-48(%ebx),%xmm1
-	vmovdqa	-32(%ebx),%xmm2
-	vmovdqa	-16(%ebx),%xmm3
-	vmovdqu	%xmm4,-128(%edi)
-	vmovdqu	%xmm5,-64(%edi)
-	vmovdqu	%xmm6,(%edi)
-	vmovdqu	%xmm7,64(%edi)
-	leal	16(%edi),%edi
-	vpaddd	-64(%ebp),%xmm0,%xmm0
-	vpaddd	-48(%ebp),%xmm1,%xmm1
-	vpaddd	-32(%ebp),%xmm2,%xmm2
-	vpaddd	-16(%ebp),%xmm3,%xmm3
-	vpunpckldq	%xmm1,%xmm0,%xmm6
-	vpunpckldq	%xmm3,%xmm2,%xmm7
-	vpunpckhdq	%xmm1,%xmm0,%xmm0
-	vpunpckhdq	%xmm3,%xmm2,%xmm2
-	vpunpcklqdq	%xmm7,%xmm6,%xmm1
-	vpunpckhqdq	%xmm7,%xmm6,%xmm6
-	vpunpcklqdq	%xmm2,%xmm0,%xmm7
-	vpunpckhqdq	%xmm2,%xmm0,%xmm3
-	vpxor	-128(%esi),%xmm1,%xmm4
-	vpxor	-64(%esi),%xmm6,%xmm5
-	vpxor	(%esi),%xmm7,%xmm6
-	vpxor	64(%esi),%xmm3,%xmm7
-	leal	16(%esi),%esi
-	vmovdqa	(%ebx),%xmm0
-	vmovdqa	16(%ebx),%xmm1
-	vmovdqa	32(%ebx),%xmm2
-	vmovdqa	48(%ebx),%xmm3
-	vmovdqu	%xmm4,-128(%edi)
-	vmovdqu	%xmm5,-64(%edi)
-	vmovdqu	%xmm6,(%edi)
-	vmovdqu	%xmm7,64(%edi)
-	leal	16(%edi),%edi
-	vpaddd	(%ebp),%xmm0,%xmm0
-	vpaddd	16(%ebp),%xmm1,%xmm1
-	vpaddd	32(%ebp),%xmm2,%xmm2
-	vpaddd	48(%ebp),%xmm3,%xmm3
-	vpunpckldq	%xmm1,%xmm0,%xmm6
-	vpunpckldq	%xmm3,%xmm2,%xmm7
-	vpunpckhdq	%xmm1,%xmm0,%xmm0
-	vpunpckhdq	%xmm3,%xmm2,%xmm2
-	vpunpcklqdq	%xmm7,%xmm6,%xmm1
-	vpunpckhqdq	%xmm7,%xmm6,%xmm6
-	vpunpcklqdq	%xmm2,%xmm0,%xmm7
-	vpunpckhqdq	%xmm2,%xmm0,%xmm3
-	vpxor	-128(%esi),%xmm1,%xmm4
-	vpxor	-64(%esi),%xmm6,%xmm5
-	vpxor	(%esi),%xmm7,%xmm6
-	vpxor	64(%esi),%xmm3,%xmm7
-	leal	16(%esi),%esi
-	vmovdqa	64(%ebx),%xmm0
-	vmovdqa	80(%ebx),%xmm1
-	vmovdqa	96(%ebx),%xmm2
-	vmovdqa	112(%ebx),%xmm3
-	vmovdqu	%xmm4,-128(%edi)
-	vmovdqu	%xmm5,-64(%edi)
-	vmovdqu	%xmm6,(%edi)
-	vmovdqu	%xmm7,64(%edi)
-	leal	16(%edi),%edi
-	vpaddd	64(%ebp),%xmm0,%xmm0
-	vpaddd	80(%ebp),%xmm1,%xmm1
-	vpaddd	96(%ebp),%xmm2,%xmm2
-	vpaddd	112(%ebp),%xmm3,%xmm3
-	vpunpckldq	%xmm1,%xmm0,%xmm6
-	vpunpckldq	%xmm3,%xmm2,%xmm7
-	vpunpckhdq	%xmm1,%xmm0,%xmm0
-	vpunpckhdq	%xmm3,%xmm2,%xmm2
-	vpunpcklqdq	%xmm7,%xmm6,%xmm1
-	vpunpckhqdq	%xmm7,%xmm6,%xmm6
-	vpunpcklqdq	%xmm2,%xmm0,%xmm7
-	vpunpckhqdq	%xmm2,%xmm0,%xmm3
-	vpxor	-128(%esi),%xmm1,%xmm4
-	vpxor	-64(%esi),%xmm6,%xmm5
-	vpxor	(%esi),%xmm7,%xmm6
-	vpxor	64(%esi),%xmm3,%xmm7
-	leal	208(%esi),%esi
-	vmovdqu	%xmm4,-128(%edi)
-	vmovdqu	%xmm5,-64(%edi)
-	vmovdqu	%xmm6,(%edi)
-	vmovdqu	%xmm7,64(%edi)
-	leal	208(%edi),%edi
-	subl	$256,%ecx
-	jnc	.L015outer_loop
-	addl	$256,%ecx
-	jz	.L017done
-	movl	520(%esp),%ebx
-	leal	-128(%esi),%esi
-	movl	516(%esp),%edx
-	leal	-128(%edi),%edi
-	vmovd	64(%ebp),%xmm2
-	vmovdqu	(%ebx),%xmm3
-	vpaddd	96(%eax),%xmm2,%xmm2
-	vpand	112(%eax),%xmm3,%xmm3
-	vpor	%xmm2,%xmm3,%xmm3
-.L0141x:
-	vmovdqa	32(%eax),%xmm0
-	vmovdqu	(%edx),%xmm1
-	vmovdqu	16(%edx),%xmm2
-	vmovdqa	(%eax),%xmm6
-	vmovdqa	16(%eax),%xmm7
-	movl	%ebp,48(%esp)
-	vmovdqa	%xmm0,(%esp)
-	vmovdqa	%xmm1,16(%esp)
-	vmovdqa	%xmm2,32(%esp)
-	vmovdqa	%xmm3,48(%esp)
-	movl	$10,%edx
-	jmp	.L018loop1x
-.align	16
-.L019outer1x:
-	vmovdqa	80(%eax),%xmm3
-	vmovdqa	(%esp),%xmm0
-	vmovdqa	16(%esp),%xmm1
-	vmovdqa	32(%esp),%xmm2
-	vpaddd	48(%esp),%xmm3,%xmm3
-	movl	$10,%edx
-	vmovdqa	%xmm3,48(%esp)
-	jmp	.L018loop1x
-.align	16
-.L018loop1x:
-	vpaddd	%xmm1,%xmm0,%xmm0
-	vpxor	%xmm0,%xmm3,%xmm3
-.byte	143,232,120,194,219,16
-	vpaddd	%xmm3,%xmm2,%xmm2
-	vpxor	%xmm2,%xmm1,%xmm1
-.byte	143,232,120,194,201,12
-	vpaddd	%xmm1,%xmm0,%xmm0
-	vpxor	%xmm0,%xmm3,%xmm3
-.byte	143,232,120,194,219,8
-	vpaddd	%xmm3,%xmm2,%xmm2
-	vpxor	%xmm2,%xmm1,%xmm1
-.byte	143,232,120,194,201,7
-	vpshufd	$78,%xmm2,%xmm2
-	vpshufd	$57,%xmm1,%xmm1
-	vpshufd	$147,%xmm3,%xmm3
-	vpaddd	%xmm1,%xmm0,%xmm0
-	vpxor	%xmm0,%xmm3,%xmm3
-.byte	143,232,120,194,219,16
-	vpaddd	%xmm3,%xmm2,%xmm2
-	vpxor	%xmm2,%xmm1,%xmm1
-.byte	143,232,120,194,201,12
-	vpaddd	%xmm1,%xmm0,%xmm0
-	vpxor	%xmm0,%xmm3,%xmm3
-.byte	143,232,120,194,219,8
-	vpaddd	%xmm3,%xmm2,%xmm2
-	vpxor	%xmm2,%xmm1,%xmm1
-.byte	143,232,120,194,201,7
-	vpshufd	$78,%xmm2,%xmm2
-	vpshufd	$147,%xmm1,%xmm1
-	vpshufd	$57,%xmm3,%xmm3
-	decl	%edx
-	jnz	.L018loop1x
-	vpaddd	(%esp),%xmm0,%xmm0
-	vpaddd	16(%esp),%xmm1,%xmm1
-	vpaddd	32(%esp),%xmm2,%xmm2
-	vpaddd	48(%esp),%xmm3,%xmm3
-	cmpl	$64,%ecx
-	jb	.L020tail
-	vpxor	(%esi),%xmm0,%xmm0
-	vpxor	16(%esi),%xmm1,%xmm1
-	vpxor	32(%esi),%xmm2,%xmm2
-	vpxor	48(%esi),%xmm3,%xmm3
-	leal	64(%esi),%esi
-	vmovdqu	%xmm0,(%edi)
-	vmovdqu	%xmm1,16(%edi)
-	vmovdqu	%xmm2,32(%edi)
-	vmovdqu	%xmm3,48(%edi)
-	leal	64(%edi),%edi
-	subl	$64,%ecx
-	jnz	.L019outer1x
-	jmp	.L017done
-.L020tail:
-	vmovdqa	%xmm0,(%esp)
-	vmovdqa	%xmm1,16(%esp)
-	vmovdqa	%xmm2,32(%esp)
-	vmovdqa	%xmm3,48(%esp)
-	xorl	%eax,%eax
-	xorl	%edx,%edx
-	xorl	%ebp,%ebp
-.L021tail_loop:
-	movb	(%esp,%ebp,1),%al
-	movb	(%esi,%ebp,1),%dl
-	leal	1(%ebp),%ebp
-	xorb	%dl,%al
-	movb	%al,-1(%edi,%ebp,1)
-	decl	%ecx
-	jnz	.L021tail_loop
-.L017done:
-	vzeroupper
-	movl	512(%esp),%esp
-	popl	%edi
-	popl	%esi
-	popl	%ebx
-	popl	%ebp
-	ret
-.size	ChaCha20_xop,.-.L_ChaCha20_xop_begin
 .comm	OPENSSL_ia32cap_P,16,4
 #else
 .text
@@ -1394,8 +914,6 @@ ChaCha20_ssse3:
 	pushl	%esi
 	pushl	%edi
 .Lssse3_shortcut:
-	testl	$2048,4(%ebp)
-	jnz	.Lxop_shortcut
 	movl	20(%esp),%edi
 	movl	24(%esp),%esi
 	movl	28(%esp),%ecx
@@ -1539,483 +1057,5 @@ ChaCha20_ssse3:
 .byte	44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32
 .byte	60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111
 .byte	114,103,62,0
-.globl	ChaCha20_xop
-.type	ChaCha20_xop,@function
-.align	16
-ChaCha20_xop:
-.L_ChaCha20_xop_begin:
-	pushl	%ebp
-	pushl	%ebx
-	pushl	%esi
-	pushl	%edi
-.Lxop_shortcut:
-	movl	20(%esp),%edi
-	movl	24(%esp),%esi
-	movl	28(%esp),%ecx
-	movl	32(%esp),%edx
-	movl	36(%esp),%ebx
-	vzeroupper
-	movl	%esp,%ebp
-	subl	$524,%esp
-	andl	$-64,%esp
-	movl	%ebp,512(%esp)
-	leal	.Lssse3_data-.Lpic_point(%eax),%eax
-	vmovdqu	(%ebx),%xmm3
-	cmpl	$256,%ecx
-	jb	.L0141x
-	movl	%edx,516(%esp)
-	movl	%ebx,520(%esp)
-	subl	$256,%ecx
-	leal	384(%esp),%ebp
-	vmovdqu	(%edx),%xmm7
-	vpshufd	$0,%xmm3,%xmm0
-	vpshufd	$85,%xmm3,%xmm1
-	vpshufd	$170,%xmm3,%xmm2
-	vpshufd	$255,%xmm3,%xmm3
-	vpaddd	48(%eax),%xmm0,%xmm0
-	vpshufd	$0,%xmm7,%xmm4
-	vpshufd	$85,%xmm7,%xmm5
-	vpsubd	64(%eax),%xmm0,%xmm0
-	vpshufd	$170,%xmm7,%xmm6
-	vpshufd	$255,%xmm7,%xmm7
-	vmovdqa	%xmm0,64(%ebp)
-	vmovdqa	%xmm1,80(%ebp)
-	vmovdqa	%xmm2,96(%ebp)
-	vmovdqa	%xmm3,112(%ebp)
-	vmovdqu	16(%edx),%xmm3
-	vmovdqa	%xmm4,-64(%ebp)
-	vmovdqa	%xmm5,-48(%ebp)
-	vmovdqa	%xmm6,-32(%ebp)
-	vmovdqa	%xmm7,-16(%ebp)
-	vmovdqa	32(%eax),%xmm7
-	leal	128(%esp),%ebx
-	vpshufd	$0,%xmm3,%xmm0
-	vpshufd	$85,%xmm3,%xmm1
-	vpshufd	$170,%xmm3,%xmm2
-	vpshufd	$255,%xmm3,%xmm3
-	vpshufd	$0,%xmm7,%xmm4
-	vpshufd	$85,%xmm7,%xmm5
-	vpshufd	$170,%xmm7,%xmm6
-	vpshufd	$255,%xmm7,%xmm7
-	vmovdqa	%xmm0,(%ebp)
-	vmovdqa	%xmm1,16(%ebp)
-	vmovdqa	%xmm2,32(%ebp)
-	vmovdqa	%xmm3,48(%ebp)
-	vmovdqa	%xmm4,-128(%ebp)
-	vmovdqa	%xmm5,-112(%ebp)
-	vmovdqa	%xmm6,-96(%ebp)
-	vmovdqa	%xmm7,-80(%ebp)
-	leal	128(%esi),%esi
-	leal	128(%edi),%edi
-	jmp	.L015outer_loop
-.align	32
-.L015outer_loop:
-	vmovdqa	-112(%ebp),%xmm1
-	vmovdqa	-96(%ebp),%xmm2
-	vmovdqa	-80(%ebp),%xmm3
-	vmovdqa	-48(%ebp),%xmm5
-	vmovdqa	-32(%ebp),%xmm6
-	vmovdqa	-16(%ebp),%xmm7
-	vmovdqa	%xmm1,-112(%ebx)
-	vmovdqa	%xmm2,-96(%ebx)
-	vmovdqa	%xmm3,-80(%ebx)
-	vmovdqa	%xmm5,-48(%ebx)
-	vmovdqa	%xmm6,-32(%ebx)
-	vmovdqa	%xmm7,-16(%ebx)
-	vmovdqa	32(%ebp),%xmm2
-	vmovdqa	48(%ebp),%xmm3
-	vmovdqa	64(%ebp),%xmm4
-	vmovdqa	80(%ebp),%xmm5
-	vmovdqa	96(%ebp),%xmm6
-	vmovdqa	112(%ebp),%xmm7
-	vpaddd	64(%eax),%xmm4,%xmm4
-	vmovdqa	%xmm2,32(%ebx)
-	vmovdqa	%xmm3,48(%ebx)
-	vmovdqa	%xmm4,64(%ebx)
-	vmovdqa	%xmm5,80(%ebx)
-	vmovdqa	%xmm6,96(%ebx)
-	vmovdqa	%xmm7,112(%ebx)
-	vmovdqa	%xmm4,64(%ebp)
-	vmovdqa	-128(%ebp),%xmm0
-	vmovdqa	%xmm4,%xmm6
-	vmovdqa	-64(%ebp),%xmm3
-	vmovdqa	(%ebp),%xmm4
-	vmovdqa	16(%ebp),%xmm5
-	movl	$10,%edx
-	nop
-.align	32
-.L016loop:
-	vpaddd	%xmm3,%xmm0,%xmm0
-	vpxor	%xmm0,%xmm6,%xmm6
-.byte	143,232,120,194,246,16
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vpxor	%xmm4,%xmm3,%xmm2
-	vmovdqa	-112(%ebx),%xmm1
-.byte	143,232,120,194,210,12
-	vmovdqa	-48(%ebx),%xmm3
-	vpaddd	%xmm2,%xmm0,%xmm0
-	vmovdqa	80(%ebx),%xmm7
-	vpxor	%xmm0,%xmm6,%xmm6
-	vpaddd	%xmm3,%xmm1,%xmm1
-.byte	143,232,120,194,246,8
-	vmovdqa	%xmm0,-128(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	%xmm6,64(%ebx)
-	vpxor	%xmm4,%xmm2,%xmm2
-	vpxor	%xmm1,%xmm7,%xmm7
-.byte	143,232,120,194,210,7
-	vmovdqa	%xmm4,(%ebx)
-.byte	143,232,120,194,255,16
-	vmovdqa	%xmm2,-64(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vmovdqa	32(%ebx),%xmm4
-	vpxor	%xmm5,%xmm3,%xmm3
-	vmovdqa	-96(%ebx),%xmm0
-.byte	143,232,120,194,219,12
-	vmovdqa	-32(%ebx),%xmm2
-	vpaddd	%xmm3,%xmm1,%xmm1
-	vmovdqa	96(%ebx),%xmm6
-	vpxor	%xmm1,%xmm7,%xmm7
-	vpaddd	%xmm2,%xmm0,%xmm0
-.byte	143,232,120,194,255,8
-	vmovdqa	%xmm1,-112(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vmovdqa	%xmm7,80(%ebx)
-	vpxor	%xmm5,%xmm3,%xmm3
-	vpxor	%xmm0,%xmm6,%xmm6
-.byte	143,232,120,194,219,7
-	vmovdqa	%xmm5,16(%ebx)
-.byte	143,232,120,194,246,16
-	vmovdqa	%xmm3,-48(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	48(%ebx),%xmm5
-	vpxor	%xmm4,%xmm2,%xmm2
-	vmovdqa	-80(%ebx),%xmm1
-.byte	143,232,120,194,210,12
-	vmovdqa	-16(%ebx),%xmm3
-	vpaddd	%xmm2,%xmm0,%xmm0
-	vmovdqa	112(%ebx),%xmm7
-	vpxor	%xmm0,%xmm6,%xmm6
-	vpaddd	%xmm3,%xmm1,%xmm1
-.byte	143,232,120,194,246,8
-	vmovdqa	%xmm0,-96(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	%xmm6,96(%ebx)
-	vpxor	%xmm4,%xmm2,%xmm2
-	vpxor	%xmm1,%xmm7,%xmm7
-.byte	143,232,120,194,210,7
-.byte	143,232,120,194,255,16
-	vmovdqa	%xmm2,-32(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vpxor	%xmm5,%xmm3,%xmm3
-	vmovdqa	-128(%ebx),%xmm0
-.byte	143,232,120,194,219,12
-	vmovdqa	-48(%ebx),%xmm2
-	vpaddd	%xmm3,%xmm1,%xmm1
-	vpxor	%xmm1,%xmm7,%xmm7
-	vpaddd	%xmm2,%xmm0,%xmm0
-.byte	143,232,120,194,255,8
-	vmovdqa	%xmm1,-80(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vpxor	%xmm5,%xmm3,%xmm3
-	vpxor	%xmm0,%xmm7,%xmm6
-.byte	143,232,120,194,219,7
-.byte	143,232,120,194,246,16
-	vmovdqa	%xmm3,-16(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vpxor	%xmm4,%xmm2,%xmm2
-	vmovdqa	-112(%ebx),%xmm1
-.byte	143,232,120,194,210,12
-	vmovdqa	-32(%ebx),%xmm3
-	vpaddd	%xmm2,%xmm0,%xmm0
-	vmovdqa	64(%ebx),%xmm7
-	vpxor	%xmm0,%xmm6,%xmm6
-	vpaddd	%xmm3,%xmm1,%xmm1
-.byte	143,232,120,194,246,8
-	vmovdqa	%xmm0,-128(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	%xmm6,112(%ebx)
-	vpxor	%xmm4,%xmm2,%xmm2
-	vpxor	%xmm1,%xmm7,%xmm7
-.byte	143,232,120,194,210,7
-	vmovdqa	%xmm4,32(%ebx)
-.byte	143,232,120,194,255,16
-	vmovdqa	%xmm2,-48(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vmovdqa	(%ebx),%xmm4
-	vpxor	%xmm5,%xmm3,%xmm3
-	vmovdqa	-96(%ebx),%xmm0
-.byte	143,232,120,194,219,12
-	vmovdqa	-16(%ebx),%xmm2
-	vpaddd	%xmm3,%xmm1,%xmm1
-	vmovdqa	80(%ebx),%xmm6
-	vpxor	%xmm1,%xmm7,%xmm7
-	vpaddd	%xmm2,%xmm0,%xmm0
-.byte	143,232,120,194,255,8
-	vmovdqa	%xmm1,-112(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vmovdqa	%xmm7,64(%ebx)
-	vpxor	%xmm5,%xmm3,%xmm3
-	vpxor	%xmm0,%xmm6,%xmm6
-.byte	143,232,120,194,219,7
-	vmovdqa	%xmm5,48(%ebx)
-.byte	143,232,120,194,246,16
-	vmovdqa	%xmm3,-32(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	16(%ebx),%xmm5
-	vpxor	%xmm4,%xmm2,%xmm2
-	vmovdqa	-80(%ebx),%xmm1
-.byte	143,232,120,194,210,12
-	vmovdqa	-64(%ebx),%xmm3
-	vpaddd	%xmm2,%xmm0,%xmm0
-	vmovdqa	96(%ebx),%xmm7
-	vpxor	%xmm0,%xmm6,%xmm6
-	vpaddd	%xmm3,%xmm1,%xmm1
-.byte	143,232,120,194,246,8
-	vmovdqa	%xmm0,-96(%ebx)
-	vpaddd	%xmm6,%xmm4,%xmm4
-	vmovdqa	%xmm6,80(%ebx)
-	vpxor	%xmm4,%xmm2,%xmm2
-	vpxor	%xmm1,%xmm7,%xmm7
-.byte	143,232,120,194,210,7
-.byte	143,232,120,194,255,16
-	vmovdqa	%xmm2,-16(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vpxor	%xmm5,%xmm3,%xmm3
-	vmovdqa	-128(%ebx),%xmm0
-.byte	143,232,120,194,219,12
-	vpaddd	%xmm3,%xmm1,%xmm1
-	vmovdqa	64(%ebx),%xmm6
-	vpxor	%xmm1,%xmm7,%xmm7
-.byte	143,232,120,194,255,8
-	vmovdqa	%xmm1,-80(%ebx)
-	vpaddd	%xmm7,%xmm5,%xmm5
-	vmovdqa	%xmm7,96(%ebx)
-	vpxor	%xmm5,%xmm3,%xmm3
-.byte	143,232,120,194,219,7
-	decl	%edx
-	jnz	.L016loop
-	vmovdqa	%xmm3,-64(%ebx)
-	vmovdqa	%xmm4,(%ebx)
-	vmovdqa	%xmm5,16(%ebx)
-	vmovdqa	%xmm6,64(%ebx)
-	vmovdqa	%xmm7,96(%ebx)
-	vmovdqa	-112(%ebx),%xmm1
-	vmovdqa	-96(%ebx),%xmm2
-	vmovdqa	-80(%ebx),%xmm3
-	vpaddd	-128(%ebp),%xmm0,%xmm0
-	vpaddd	-112(%ebp),%xmm1,%xmm1
-	vpaddd	-96(%ebp),%xmm2,%xmm2
-	vpaddd	-80(%ebp),%xmm3,%xmm3
-	vpunpckldq	%xmm1,%xmm0,%xmm6
-	vpunpckldq	%xmm3,%xmm2,%xmm7
-	vpunpckhdq	%xmm1,%xmm0,%xmm0
-	vpunpckhdq	%xmm3,%xmm2,%xmm2
-	vpunpcklqdq	%xmm7,%xmm6,%xmm1
-	vpunpckhqdq	%xmm7,%xmm6,%xmm6
-	vpunpcklqdq	%xmm2,%xmm0,%xmm7
-	vpunpckhqdq	%xmm2,%xmm0,%xmm3
-	vpxor	-128(%esi),%xmm1,%xmm4
-	vpxor	-64(%esi),%xmm6,%xmm5
-	vpxor	(%esi),%xmm7,%xmm6
-	vpxor	64(%esi),%xmm3,%xmm7
-	leal	16(%esi),%esi
-	vmovdqa	-64(%ebx),%xmm0
-	vmovdqa	-48(%ebx),%xmm1
-	vmovdqa	-32(%ebx),%xmm2
-	vmovdqa	-16(%ebx),%xmm3
-	vmovdqu	%xmm4,-128(%edi)
-	vmovdqu	%xmm5,-64(%edi)
-	vmovdqu	%xmm6,(%edi)
-	vmovdqu	%xmm7,64(%edi)
-	leal	16(%edi),%edi
-	vpaddd	-64(%ebp),%xmm0,%xmm0
-	vpaddd	-48(%ebp),%xmm1,%xmm1
-	vpaddd	-32(%ebp),%xmm2,%xmm2
-	vpaddd	-16(%ebp),%xmm3,%xmm3
-	vpunpckldq	%xmm1,%xmm0,%xmm6
-	vpunpckldq	%xmm3,%xmm2,%xmm7
-	vpunpckhdq	%xmm1,%xmm0,%xmm0
-	vpunpckhdq	%xmm3,%xmm2,%xmm2
-	vpunpcklqdq	%xmm7,%xmm6,%xmm1
-	vpunpckhqdq	%xmm7,%xmm6,%xmm6
-	vpunpcklqdq	%xmm2,%xmm0,%xmm7
-	vpunpckhqdq	%xmm2,%xmm0,%xmm3
-	vpxor	-128(%esi),%xmm1,%xmm4
-	vpxor	-64(%esi),%xmm6,%xmm5
-	vpxor	(%esi),%xmm7,%xmm6
-	vpxor	64(%esi),%xmm3,%xmm7
-	leal	16(%esi),%esi
-	vmovdqa	(%ebx),%xmm0
-	vmovdqa	16(%ebx),%xmm1
-	vmovdqa	32(%ebx),%xmm2
-	vmovdqa	48(%ebx),%xmm3
-	vmovdqu	%xmm4,-128(%edi)
-	vmovdqu	%xmm5,-64(%edi)
-	vmovdqu	%xmm6,(%edi)
-	vmovdqu	%xmm7,64(%edi)
-	leal	16(%edi),%edi
-	vpaddd	(%ebp),%xmm0,%xmm0
-	vpaddd	16(%ebp),%xmm1,%xmm1
-	vpaddd	32(%ebp),%xmm2,%xmm2
-	vpaddd	48(%ebp),%xmm3,%xmm3
-	vpunpckldq	%xmm1,%xmm0,%xmm6
-	vpunpckldq	%xmm3,%xmm2,%xmm7
-	vpunpckhdq	%xmm1,%xmm0,%xmm0
-	vpunpckhdq	%xmm3,%xmm2,%xmm2
-	vpunpcklqdq	%xmm7,%xmm6,%xmm1
-	vpunpckhqdq	%xmm7,%xmm6,%xmm6
-	vpunpcklqdq	%xmm2,%xmm0,%xmm7
-	vpunpckhqdq	%xmm2,%xmm0,%xmm3
-	vpxor	-128(%esi),%xmm1,%xmm4
-	vpxor	-64(%esi),%xmm6,%xmm5
-	vpxor	(%esi),%xmm7,%xmm6
-	vpxor	64(%esi),%xmm3,%xmm7
-	leal	16(%esi),%esi
-	vmovdqa	64(%ebx),%xmm0
-	vmovdqa	80(%ebx),%xmm1
-	vmovdqa	96(%ebx),%xmm2
-	vmovdqa	112(%ebx),%xmm3
-	vmovdqu	%xmm4,-128(%edi)
-	vmovdqu	%xmm5,-64(%edi)
-	vmovdqu	%xmm6,(%edi)
-	vmovdqu	%xmm7,64(%edi)
-	leal	16(%edi),%edi
-	vpaddd	64(%ebp),%xmm0,%xmm0
-	vpaddd	80(%ebp),%xmm1,%xmm1
-	vpaddd	96(%ebp),%xmm2,%xmm2
-	vpaddd	112(%ebp),%xmm3,%xmm3
-	vpunpckldq	%xmm1,%xmm0,%xmm6
-	vpunpckldq	%xmm3,%xmm2,%xmm7
-	vpunpckhdq	%xmm1,%xmm0,%xmm0
-	vpunpckhdq	%xmm3,%xmm2,%xmm2
-	vpunpcklqdq	%xmm7,%xmm6,%xmm1
-	vpunpckhqdq	%xmm7,%xmm6,%xmm6
-	vpunpcklqdq	%xmm2,%xmm0,%xmm7
-	vpunpckhqdq	%xmm2,%xmm0,%xmm3
-	vpxor	-128(%esi),%xmm1,%xmm4
-	vpxor	-64(%esi),%xmm6,%xmm5
-	vpxor	(%esi),%xmm7,%xmm6
-	vpxor	64(%esi),%xmm3,%xmm7
-	leal	208(%esi),%esi
-	vmovdqu	%xmm4,-128(%edi)
-	vmovdqu	%xmm5,-64(%edi)
-	vmovdqu	%xmm6,(%edi)
-	vmovdqu	%xmm7,64(%edi)
-	leal	208(%edi),%edi
-	subl	$256,%ecx
-	jnc	.L015outer_loop
-	addl	$256,%ecx
-	jz	.L017done
-	movl	520(%esp),%ebx
-	leal	-128(%esi),%esi
-	movl	516(%esp),%edx
-	leal	-128(%edi),%edi
-	vmovd	64(%ebp),%xmm2
-	vmovdqu	(%ebx),%xmm3
-	vpaddd	96(%eax),%xmm2,%xmm2
-	vpand	112(%eax),%xmm3,%xmm3
-	vpor	%xmm2,%xmm3,%xmm3
-.L0141x:
-	vmovdqa	32(%eax),%xmm0
-	vmovdqu	(%edx),%xmm1
-	vmovdqu	16(%edx),%xmm2
-	vmovdqa	(%eax),%xmm6
-	vmovdqa	16(%eax),%xmm7
-	movl	%ebp,48(%esp)
-	vmovdqa	%xmm0,(%esp)
-	vmovdqa	%xmm1,16(%esp)
-	vmovdqa	%xmm2,32(%esp)
-	vmovdqa	%xmm3,48(%esp)
-	movl	$10,%edx
-	jmp	.L018loop1x
-.align	16
-.L019outer1x:
-	vmovdqa	80(%eax),%xmm3
-	vmovdqa	(%esp),%xmm0
-	vmovdqa	16(%esp),%xmm1
-	vmovdqa	32(%esp),%xmm2
-	vpaddd	48(%esp),%xmm3,%xmm3
-	movl	$10,%edx
-	vmovdqa	%xmm3,48(%esp)
-	jmp	.L018loop1x
-.align	16
-.L018loop1x:
-	vpaddd	%xmm1,%xmm0,%xmm0
-	vpxor	%xmm0,%xmm3,%xmm3
-.byte	143,232,120,194,219,16
-	vpaddd	%xmm3,%xmm2,%xmm2
-	vpxor	%xmm2,%xmm1,%xmm1
-.byte	143,232,120,194,201,12
-	vpaddd	%xmm1,%xmm0,%xmm0
-	vpxor	%xmm0,%xmm3,%xmm3
-.byte	143,232,120,194,219,8
-	vpaddd	%xmm3,%xmm2,%xmm2
-	vpxor	%xmm2,%xmm1,%xmm1
-.byte	143,232,120,194,201,7
-	vpshufd	$78,%xmm2,%xmm2
-	vpshufd	$57,%xmm1,%xmm1
-	vpshufd	$147,%xmm3,%xmm3
-	vpaddd	%xmm1,%xmm0,%xmm0
-	vpxor	%xmm0,%xmm3,%xmm3
-.byte	143,232,120,194,219,16
-	vpaddd	%xmm3,%xmm2,%xmm2
-	vpxor	%xmm2,%xmm1,%xmm1
-.byte	143,232,120,194,201,12
-	vpaddd	%xmm1,%xmm0,%xmm0
-	vpxor	%xmm0,%xmm3,%xmm3
-.byte	143,232,120,194,219,8
-	vpaddd	%xmm3,%xmm2,%xmm2
-	vpxor	%xmm2,%xmm1,%xmm1
-.byte	143,232,120,194,201,7
-	vpshufd	$78,%xmm2,%xmm2
-	vpshufd	$147,%xmm1,%xmm1
-	vpshufd	$57,%xmm3,%xmm3
-	decl	%edx
-	jnz	.L018loop1x
-	vpaddd	(%esp),%xmm0,%xmm0
-	vpaddd	16(%esp),%xmm1,%xmm1
-	vpaddd	32(%esp),%xmm2,%xmm2
-	vpaddd	48(%esp),%xmm3,%xmm3
-	cmpl	$64,%ecx
-	jb	.L020tail
-	vpxor	(%esi),%xmm0,%xmm0
-	vpxor	16(%esi),%xmm1,%xmm1
-	vpxor	32(%esi),%xmm2,%xmm2
-	vpxor	48(%esi),%xmm3,%xmm3
-	leal	64(%esi),%esi
-	vmovdqu	%xmm0,(%edi)
-	vmovdqu	%xmm1,16(%edi)
-	vmovdqu	%xmm2,32(%edi)
-	vmovdqu	%xmm3,48(%edi)
-	leal	64(%edi),%edi
-	subl	$64,%ecx
-	jnz	.L019outer1x
-	jmp	.L017done
-.L020tail:
-	vmovdqa	%xmm0,(%esp)
-	vmovdqa	%xmm1,16(%esp)
-	vmovdqa	%xmm2,32(%esp)
-	vmovdqa	%xmm3,48(%esp)
-	xorl	%eax,%eax
-	xorl	%edx,%edx
-	xorl	%ebp,%ebp
-.L021tail_loop:
-	movb	(%esp,%ebp,1),%al
-	movb	(%esi,%ebp,1),%dl
-	leal	1(%ebp),%ebp
-	xorb	%dl,%al
-	movb	%al,-1(%edi,%ebp,1)
-	decl	%ecx
-	jnz	.L021tail_loop
-.L017done:
-	vzeroupper
-	movl	512(%esp),%esp
-	popl	%edi
-	popl	%esi
-	popl	%ebx
-	popl	%ebp
-	ret
-.size	ChaCha20_xop,.-.L_ChaCha20_xop_begin
 .comm	OPENSSL_ia32cap_P,16,4
 #endif
diff --git a/secure/lib/libcrypto/i386/ecp_nistz256-x86.S b/secure/lib/libcrypto/i386/ecp_nistz256-x86.S
index 7d0c1b9eb9df..eb413d9f1a73 100644
--- a/secure/lib/libcrypto/i386/ecp_nistz256-x86.S
+++ b/secure/lib/libcrypto/i386/ecp_nistz256-x86.S
@@ -4422,19 +4422,15 @@ ecp_nistz256_point_add:
 	orl	4(%edi),%eax
 	orl	8(%edi),%eax
 	orl	12(%edi),%eax
+	movl	576(%esp),%ebx
+	notl	%ebx
+	orl	%ebx,%eax
+	movl	580(%esp),%ebx
+	notl	%ebx
+	orl	%ebx,%eax
+	orl	584(%esp),%eax
 .byte	62
 	jnz	.L010add_proceed
-	movl	576(%esp),%eax
-	andl	580(%esp),%eax
-	movl	584(%esp),%ebx
-	jz	.L010add_proceed
-	testl	%ebx,%ebx
-	jz	.L011add_double
-	movl	616(%esp),%edi
-	xorl	%eax,%eax
-	movl	$24,%ecx
-.byte	252,243,171
-	jmp	.L012add_done
 .align	16
 .L011add_double:
 	movl	620(%esp),%esi
@@ -9590,19 +9586,15 @@ ecp_nistz256_point_add:
 	orl	4(%edi),%eax
 	orl	8(%edi),%eax
 	orl	12(%edi),%eax
+	movl	576(%esp),%ebx
+	notl	%ebx
+	orl	%ebx,%eax
+	movl	580(%esp),%ebx
+	notl	%ebx
+	orl	%ebx,%eax
+	orl	584(%esp),%eax
 .byte	62
 	jnz	.L010add_proceed
-	movl	576(%esp),%eax
-	andl	580(%esp),%eax
-	movl	584(%esp),%ebx
-	jz	.L010add_proceed
-	testl	%ebx,%ebx
-	jz	.L011add_double
-	movl	616(%esp),%edi
-	xorl	%eax,%eax
-	movl	$24,%ecx
-.byte	252,243,171
-	jmp	.L012add_done
 .align	16
 .L011add_double:
 	movl	620(%esp),%esi
diff --git a/secure/lib/libcrypto/i386/poly1305-x86.S b/secure/lib/libcrypto/i386/poly1305-x86.S
index 100deee40bf2..b394500278d5 100644
--- a/secure/lib/libcrypto/i386/poly1305-x86.S
+++ b/secure/lib/libcrypto/i386/poly1305-x86.S
@@ -36,10 +36,6 @@ poly1305_init:
 	jne	.L002no_sse2
 	leal	_poly1305_blocks_sse2-.L001pic_point(%ebx),%eax
 	leal	_poly1305_emit_sse2-.L001pic_point(%ebx),%edx
-	movl	8(%edi),%ecx
-	testl	$32,%ecx
-	jz	.L002no_sse2
-	leal	_poly1305_blocks_avx2-.L001pic_point(%ebx),%eax
 .L002no_sse2:
 	movl	20(%esp),%edi
 	movl	%eax,(%ebp)
@@ -1348,557 +1344,6 @@ _poly1305_emit_sse2:
 	popl	%ebp
 	ret
 .size	_poly1305_emit_sse2,.-_poly1305_emit_sse2
-.align	32
-.type	_poly1305_init_avx2,@function
-.align	16
-_poly1305_init_avx2:
-	vmovdqu	24(%edi),%xmm4
-	leal	48(%edi),%edi
-	movl	%esp,%ebp
-	subl	$224,%esp
-	andl	$-16,%esp
-	vmovdqa	64(%ebx),%xmm7
-	vpand	%xmm7,%xmm4,%xmm0
-	vpsrlq	$26,%xmm4,%xmm1
-	vpsrldq	$6,%xmm4,%xmm3
-	vpand	%xmm7,%xmm1,%xmm1
-	vpsrlq	$4,%xmm3,%xmm2
-	vpsrlq	$30,%xmm3,%xmm3
-	vpand	%xmm7,%xmm2,%xmm2
-	vpand	%xmm7,%xmm3,%xmm3
-	vpsrldq	$13,%xmm4,%xmm4
-	leal	144(%esp),%edx
-	movl	$2,%ecx
-.L018square:
-	vmovdqa	%xmm0,(%esp)
-	vmovdqa	%xmm1,16(%esp)
-	vmovdqa	%xmm2,32(%esp)
-	vmovdqa	%xmm3,48(%esp)
-	vmovdqa	%xmm4,64(%esp)
-	vpslld	$2,%xmm1,%xmm6
-	vpslld	$2,%xmm2,%xmm5
-	vpaddd	%xmm1,%xmm6,%xmm6
-	vpaddd	%xmm2,%xmm5,%xmm5
-	vmovdqa	%xmm6,80(%esp)
-	vmovdqa	%xmm5,96(%esp)
-	vpslld	$2,%xmm3,%xmm6
-	vpslld	$2,%xmm4,%xmm5
-	vpaddd	%xmm3,%xmm6,%xmm6
-	vpaddd	%xmm4,%xmm5,%xmm5
-	vmovdqa	%xmm6,112(%esp)
-	vmovdqa	%xmm5,128(%esp)
-	vpshufd	$68,%xmm0,%xmm5
-	vmovdqa	%xmm1,%xmm6
-	vpshufd	$68,%xmm1,%xmm1
-	vpshufd	$68,%xmm2,%xmm2
-	vpshufd	$68,%xmm3,%xmm3
-	vpshufd	$68,%xmm4,%xmm4
-	vmovdqa	%xmm5,(%edx)
-	vmovdqa	%xmm1,16(%edx)
-	vmovdqa	%xmm2,32(%edx)
-	vmovdqa	%xmm3,48(%edx)
-	vmovdqa	%xmm4,64(%edx)
-	vpmuludq	%xmm0,%xmm4,%xmm4
-	vpmuludq	%xmm0,%xmm3,%xmm3
-	vpmuludq	%xmm0,%xmm2,%xmm2
-	vpmuludq	%xmm0,%xmm1,%xmm1
-	vpmuludq	%xmm0,%xmm5,%xmm0
-	vpmuludq	48(%edx),%xmm6,%xmm5
-	vpaddq	%xmm5,%xmm4,%xmm4
-	vpmuludq	32(%edx),%xmm6,%xmm7
-	vpaddq	%xmm7,%xmm3,%xmm3
-	vpmuludq	16(%edx),%xmm6,%xmm5
-	vpaddq	%xmm5,%xmm2,%xmm2
-	vmovdqa	80(%esp),%xmm7
-	vpmuludq	(%edx),%xmm6,%xmm6
-	vpaddq	%xmm6,%xmm1,%xmm1
-	vmovdqa	32(%esp),%xmm5
-	vpmuludq	64(%edx),%xmm7,%xmm7
-	vpaddq	%xmm7,%xmm0,%xmm0
-	vpmuludq	32(%edx),%xmm5,%xmm6
-	vpaddq	%xmm6,%xmm4,%xmm4
-	vpmuludq	16(%edx),%xmm5,%xmm7
-	vpaddq	%xmm7,%xmm3,%xmm3
-	vmovdqa	96(%esp),%xmm6
-	vpmuludq	(%edx),%xmm5,%xmm5
-	vpaddq	%xmm5,%xmm2,%xmm2
-	vpmuludq	64(%edx),%xmm6,%xmm7
-	vpaddq	%xmm7,%xmm1,%xmm1
-	vmovdqa	48(%esp),%xmm5
-	vpmuludq	48(%edx),%xmm6,%xmm6
-	vpaddq	%xmm6,%xmm0,%xmm0
-	vpmuludq	16(%edx),%xmm5,%xmm7
-	vpaddq	%xmm7,%xmm4,%xmm4
-	vmovdqa	112(%esp),%xmm6
-	vpmuludq	(%edx),%xmm5,%xmm5
-	vpaddq	%xmm5,%xmm3,%xmm3
-	vpmuludq	64(%edx),%xmm6,%xmm7
-	vpaddq	%xmm7,%xmm2,%xmm2
-	vpmuludq	48(%edx),%xmm6,%xmm5
-	vpaddq	%xmm5,%xmm1,%xmm1
-	vmovdqa	64(%esp),%xmm7
-	vpmuludq	32(%edx),%xmm6,%xmm6
-	vpaddq	%xmm6,%xmm0,%xmm0
-	vmovdqa	128(%esp),%xmm5
-	vpmuludq	(%edx),%xmm7,%xmm7
-	vpaddq	%xmm7,%xmm4,%xmm4
-	vpmuludq	64(%edx),%xmm5,%xmm6
-	vpaddq	%xmm6,%xmm3,%xmm3
-	vpmuludq	16(%edx),%xmm5,%xmm7
-	vpaddq	%xmm7,%xmm0,%xmm0
-	vpmuludq	32(%edx),%xmm5,%xmm6
-	vpaddq	%xmm6,%xmm1,%xmm1
-	vmovdqa	64(%ebx),%xmm7
-	vpmuludq	48(%edx),%xmm5,%xmm5
-	vpaddq	%xmm5,%xmm2,%xmm2
-	vpsrlq	$26,%xmm3,%xmm5
-	vpand	%xmm7,%xmm3,%xmm3
-	vpsrlq	$26,%xmm0,%xmm6
-	vpand	%xmm7,%xmm0,%xmm0
-	vpaddq	%xmm5,%xmm4,%xmm4
-	vpaddq	%xmm6,%xmm1,%xmm1
-	vpsrlq	$26,%xmm4,%xmm5
-	vpand	%xmm7,%xmm4,%xmm4
-	vpsrlq	$26,%xmm1,%xmm6
-	vpand	%xmm7,%xmm1,%xmm1
-	vpaddq	%xmm6,%xmm2,%xmm2
-	vpaddd	%xmm5,%xmm0,%xmm0
-	vpsllq	$2,%xmm5,%xmm5
-	vpsrlq	$26,%xmm2,%xmm6
-	vpand	%xmm7,%xmm2,%xmm2
-	vpaddd	%xmm5,%xmm0,%xmm0
-	vpaddd	%xmm6,%xmm3,%xmm3
-	vpsrlq	$26,%xmm3,%xmm6
-	vpsrlq	$26,%xmm0,%xmm5
-	vpand	%xmm7,%xmm0,%xmm0
-	vpand	%xmm7,%xmm3,%xmm3
-	vpaddd	%xmm5,%xmm1,%xmm1
-	vpaddd	%xmm6,%xmm4,%xmm4
-	decl	%ecx
-	jz	.L019square_break
-	vpunpcklqdq	(%esp),%xmm0,%xmm0
-	vpunpcklqdq	16(%esp),%xmm1,%xmm1
-	vpunpcklqdq	32(%esp),%xmm2,%xmm2
-	vpunpcklqdq	48(%esp),%xmm3,%xmm3
-	vpunpcklqdq	64(%esp),%xmm4,%xmm4
-	jmp	.L018square
-.L019square_break:
-	vpsllq	$32,%xmm0,%xmm0
-	vpsllq	$32,%xmm1,%xmm1
-	vpsllq	$32,%xmm2,%xmm2
-	vpsllq	$32,%xmm3,%xmm3
-	vpsllq	$32,%xmm4,%xmm4
-	vpor	(%esp),%xmm0,%xmm0
-	vpor	16(%esp),%xmm1,%xmm1
-	vpor	32(%esp),%xmm2,%xmm2
-	vpor	48(%esp),%xmm3,%xmm3
-	vpor	64(%esp),%xmm4,%xmm4
-	vpshufd	$141,%xmm0,%xmm0
-	vpshufd	$141,%xmm1,%xmm1
-	vpshufd	$141,%xmm2,%xmm2
-	vpshufd	$141,%xmm3,%xmm3
-	vpshufd	$141,%xmm4,%xmm4
-	vmovdqu	%xmm0,(%edi)
-	vmovdqu	%xmm1,16(%edi)
-	vmovdqu	%xmm2,32(%edi)
-	vmovdqu	%xmm3,48(%edi)
-	vmovdqu	%xmm4,64(%edi)
-	vpslld	$2,%xmm1,%xmm6
-	vpslld	$2,%xmm2,%xmm5
-	vpaddd	%xmm1,%xmm6,%xmm6
-	vpaddd	%xmm2,%xmm5,%xmm5
-	vmovdqu	%xmm6,80(%edi)
-	vmovdqu	%xmm5,96(%edi)
-	vpslld	$2,%xmm3,%xmm6
-	vpslld	$2,%xmm4,%xmm5
-	vpaddd	%xmm3,%xmm6,%xmm6
-	vpaddd	%xmm4,%xmm5,%xmm5
-	vmovdqu	%xmm6,112(%edi)
-	vmovdqu	%xmm5,128(%edi)
-	movl	%ebp,%esp
-	leal	-48(%edi),%edi
-	ret
-.size	_poly1305_init_avx2,.-_poly1305_init_avx2
-.align	32
-.type	_poly1305_blocks_avx2,@function
-.align	16
-_poly1305_blocks_avx2:
-	pushl	%ebp
-	pushl	%ebx
-	pushl	%esi
-	pushl	%edi
-	movl	20(%esp),%edi
-	movl	24(%esp),%esi
-	movl	28(%esp),%ecx
-	movl	20(%edi),%eax
-	andl	$-16,%ecx
-	jz	.L020nodata
-	cmpl	$64,%ecx
-	jae	.L021enter_avx2
-	testl	%eax,%eax
-	jz	.Lenter_blocks
-.L021enter_avx2:
-	vzeroupper
-	call	.L022pic_point
-.L022pic_point:
-	popl	%ebx
-	leal	.Lconst_sse2-.L022pic_point(%ebx),%ebx
-	testl	%eax,%eax
-	jnz	.L023base2_26
-	call	_poly1305_init_avx2
-	movl	(%edi),%eax
-	movl	3(%edi),%ecx
-	movl	6(%edi),%edx
-	movl	9(%edi),%esi
-	movl	13(%edi),%ebp
-	shrl	$2,%ecx
-	andl	$67108863,%eax
-	shrl	$4,%edx
-	andl	$67108863,%ecx
-	shrl	$6,%esi
-	andl	$67108863,%edx
-	movl	%eax,(%edi)
-	movl	%ecx,4(%edi)
-	movl	%edx,8(%edi)
-	movl	%esi,12(%edi)
-	movl	%ebp,16(%edi)
-	movl	$1,20(%edi)
-	movl	24(%esp),%esi
-	movl	28(%esp),%ecx
-.L023base2_26:
-	movl	32(%esp),%eax
-	movl	%esp,%ebp
-	subl	$448,%esp
-	andl	$-512,%esp
-	vmovdqu	48(%edi),%xmm0
-	leal	288(%esp),%edx
-	vmovdqu	64(%edi),%xmm1
-	vmovdqu	80(%edi),%xmm2
-	vmovdqu	96(%edi),%xmm3
-	vmovdqu	112(%edi),%xmm4
-	leal	48(%edi),%edi
-	vpermq	$64,%ymm0,%ymm0
-	vpermq	$64,%ymm1,%ymm1
-	vpermq	$64,%ymm2,%ymm2
-	vpermq	$64,%ymm3,%ymm3
-	vpermq	$64,%ymm4,%ymm4
-	vpshufd	$200,%ymm0,%ymm0
-	vpshufd	$200,%ymm1,%ymm1
-	vpshufd	$200,%ymm2,%ymm2
-	vpshufd	$200,%ymm3,%ymm3
-	vpshufd	$200,%ymm4,%ymm4
-	vmovdqa	%ymm0,-128(%edx)
-	vmovdqu	80(%edi),%xmm0
-	vmovdqa	%ymm1,-96(%edx)
-	vmovdqu	96(%edi),%xmm1
-	vmovdqa	%ymm2,-64(%edx)
-	vmovdqu	112(%edi),%xmm2
-	vmovdqa	%ymm3,-32(%edx)
-	vmovdqu	128(%edi),%xmm3
-	vmovdqa	%ymm4,(%edx)
-	vpermq	$64,%ymm0,%ymm0
-	vpermq	$64,%ymm1,%ymm1
-	vpermq	$64,%ymm2,%ymm2
-	vpermq	$64,%ymm3,%ymm3
-	vpshufd	$200,%ymm0,%ymm0
-	vpshufd	$200,%ymm1,%ymm1
-	vpshufd	$200,%ymm2,%ymm2
-	vpshufd	$200,%ymm3,%ymm3
-	vmovdqa	%ymm0,32(%edx)
-	vmovd	-48(%edi),%xmm0
-	vmovdqa	%ymm1,64(%edx)
-	vmovd	-44(%edi),%xmm1
-	vmovdqa	%ymm2,96(%edx)
-	vmovd	-40(%edi),%xmm2
-	vmovdqa	%ymm3,128(%edx)
-	vmovd	-36(%edi),%xmm3
-	vmovd	-32(%edi),%xmm4
-	vmovdqa	64(%ebx),%ymm7
-	negl	%eax
-	testl	$63,%ecx
-	jz	.L024even
-	movl	%ecx,%edx
-	andl	$-64,%ecx
-	andl	$63,%edx
-	vmovdqu	(%esi),%xmm5
-	cmpl	$32,%edx
-	jb	.L025one
-	vmovdqu	16(%esi),%xmm6
-	je	.L026two
-	vinserti128	$1,32(%esi),%ymm5,%ymm5
-	leal	48(%esi),%esi
-	leal	8(%ebx),%ebx
-	leal	296(%esp),%edx
-	jmp	.L027tail
-.L026two:
-	leal	32(%esi),%esi
-	leal	16(%ebx),%ebx
-	leal	304(%esp),%edx
-	jmp	.L027tail
-.L025one:
-	leal	16(%esi),%esi
-	vpxor	%ymm6,%ymm6,%ymm6
-	leal	32(%ebx,%eax,8),%ebx
-	leal	312(%esp),%edx
-	jmp	.L027tail
-.align	32
-.L024even:
-	vmovdqu	(%esi),%xmm5
-	vmovdqu	16(%esi),%xmm6
-	vinserti128	$1,32(%esi),%ymm5,%ymm5
-	vinserti128	$1,48(%esi),%ymm6,%ymm6
-	leal	64(%esi),%esi
-	subl	$64,%ecx
-	jz	.L027tail
-.L028loop:
-	vmovdqa	%ymm2,64(%esp)
-	vpsrldq	$6,%ymm5,%ymm2
-	vmovdqa	%ymm0,(%esp)
-	vpsrldq	$6,%ymm6,%ymm0
-	vmovdqa	%ymm1,32(%esp)
-	vpunpckhqdq	%ymm6,%ymm5,%ymm1
-	vpunpcklqdq	%ymm6,%ymm5,%ymm5
-	vpunpcklqdq	%ymm0,%ymm2,%ymm2
-	vpsrlq	$30,%ymm2,%ymm0
-	vpsrlq	$4,%ymm2,%ymm2
-	vpsrlq	$26,%ymm5,%ymm6
-	vpsrlq	$40,%ymm1,%ymm1
-	vpand	%ymm7,%ymm2,%ymm2
-	vpand	%ymm7,%ymm5,%ymm5
-	vpand	%ymm7,%ymm6,%ymm6
-	vpand	%ymm7,%ymm0,%ymm0
-	vpor	(%ebx),%ymm1,%ymm1
-	vpaddq	64(%esp),%ymm2,%ymm2
-	vpaddq	(%esp),%ymm5,%ymm5
-	vpaddq	32(%esp),%ymm6,%ymm6
-	vpaddq	%ymm3,%ymm0,%ymm0
-	vpaddq	%ymm4,%ymm1,%ymm1
-	vpmuludq	-96(%edx),%ymm2,%ymm3
-	vmovdqa	%ymm6,32(%esp)
-	vpmuludq	-64(%edx),%ymm2,%ymm4
-	vmovdqa	%ymm0,96(%esp)
-	vpmuludq	96(%edx),%ymm2,%ymm0
-	vmovdqa	%ymm1,128(%esp)
-	vpmuludq	128(%edx),%ymm2,%ymm1
-	vpmuludq	-128(%edx),%ymm2,%ymm2
-	vpmuludq	-32(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm3,%ymm3
-	vpmuludq	(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm4,%ymm4
-	vpmuludq	-128(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm0,%ymm0
-	vmovdqa	32(%esp),%ymm7
-	vpmuludq	-96(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpmuludq	-64(%edx),%ymm5,%ymm5
-	vpaddq	%ymm5,%ymm2,%ymm2
-	vpmuludq	-64(%edx),%ymm7,%ymm6
-	vpaddq	%ymm6,%ymm3,%ymm3
-	vpmuludq	-32(%edx),%ymm7,%ymm5
-	vpaddq	%ymm5,%ymm4,%ymm4
-	vpmuludq	128(%edx),%ymm7,%ymm6
-	vpaddq	%ymm6,%ymm0,%ymm0
-	vmovdqa	96(%esp),%ymm6
-	vpmuludq	-128(%edx),%ymm7,%ymm5
-	vpaddq	%ymm5,%ymm1,%ymm1
-	vpmuludq	-96(%edx),%ymm7,%ymm7
-	vpaddq	%ymm7,%ymm2,%ymm2
-	vpmuludq	-128(%edx),%ymm6,%ymm5
-	vpaddq	%ymm5,%ymm3,%ymm3
-	vpmuludq	-96(%edx),%ymm6,%ymm7
-	vpaddq	%ymm7,%ymm4,%ymm4
-	vpmuludq	64(%edx),%ymm6,%ymm5
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vmovdqa	128(%esp),%ymm5
-	vpmuludq	96(%edx),%ymm6,%ymm7
-	vpaddq	%ymm7,%ymm1,%ymm1
-	vpmuludq	128(%edx),%ymm6,%ymm6
-	vpaddq	%ymm6,%ymm2,%ymm2
-	vpmuludq	128(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm3,%ymm3
-	vpmuludq	32(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm0,%ymm0
-	vpmuludq	-128(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm4,%ymm4
-	vmovdqa	64(%ebx),%ymm7
-	vpmuludq	64(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpmuludq	96(%edx),%ymm5,%ymm5
-	vpaddq	%ymm5,%ymm2,%ymm2
-	vpsrlq	$26,%ymm3,%ymm5
-	vpand	%ymm7,%ymm3,%ymm3
-	vpsrlq	$26,%ymm0,%ymm6
-	vpand	%ymm7,%ymm0,%ymm0
-	vpaddq	%ymm5,%ymm4,%ymm4
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpsrlq	$26,%ymm4,%ymm5
-	vpand	%ymm7,%ymm4,%ymm4
-	vpsrlq	$26,%ymm1,%ymm6
-	vpand	%ymm7,%ymm1,%ymm1
-	vpaddq	%ymm6,%ymm2,%ymm2
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vpsllq	$2,%ymm5,%ymm5
-	vpsrlq	$26,%ymm2,%ymm6
-	vpand	%ymm7,%ymm2,%ymm2
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vpaddq	%ymm6,%ymm3,%ymm3
-	vpsrlq	$26,%ymm3,%ymm6
-	vpsrlq	$26,%ymm0,%ymm5
-	vpand	%ymm7,%ymm0,%ymm0
-	vpand	%ymm7,%ymm3,%ymm3
-	vpaddq	%ymm5,%ymm1,%ymm1
-	vpaddq	%ymm6,%ymm4,%ymm4
-	vmovdqu	(%esi),%xmm5
-	vmovdqu	16(%esi),%xmm6
-	vinserti128	$1,32(%esi),%ymm5,%ymm5
-	vinserti128	$1,48(%esi),%ymm6,%ymm6
-	leal	64(%esi),%esi
-	subl	$64,%ecx
-	jnz	.L028loop
-.L027tail:
-	vmovdqa	%ymm2,64(%esp)
-	vpsrldq	$6,%ymm5,%ymm2
-	vmovdqa	%ymm0,(%esp)
-	vpsrldq	$6,%ymm6,%ymm0
-	vmovdqa	%ymm1,32(%esp)
-	vpunpckhqdq	%ymm6,%ymm5,%ymm1
-	vpunpcklqdq	%ymm6,%ymm5,%ymm5
-	vpunpcklqdq	%ymm0,%ymm2,%ymm2
-	vpsrlq	$30,%ymm2,%ymm0
-	vpsrlq	$4,%ymm2,%ymm2
-	vpsrlq	$26,%ymm5,%ymm6
-	vpsrlq	$40,%ymm1,%ymm1
-	vpand	%ymm7,%ymm2,%ymm2
-	vpand	%ymm7,%ymm5,%ymm5
-	vpand	%ymm7,%ymm6,%ymm6
-	vpand	%ymm7,%ymm0,%ymm0
-	vpor	(%ebx),%ymm1,%ymm1
-	andl	$-64,%ebx
-	vpaddq	64(%esp),%ymm2,%ymm2
-	vpaddq	(%esp),%ymm5,%ymm5
-	vpaddq	32(%esp),%ymm6,%ymm6
-	vpaddq	%ymm3,%ymm0,%ymm0
-	vpaddq	%ymm4,%ymm1,%ymm1
-	vpmuludq	-92(%edx),%ymm2,%ymm3
-	vmovdqa	%ymm6,32(%esp)
-	vpmuludq	-60(%edx),%ymm2,%ymm4
-	vmovdqa	%ymm0,96(%esp)
-	vpmuludq	100(%edx),%ymm2,%ymm0
-	vmovdqa	%ymm1,128(%esp)
-	vpmuludq	132(%edx),%ymm2,%ymm1
-	vpmuludq	-124(%edx),%ymm2,%ymm2
-	vpmuludq	-28(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm3,%ymm3
-	vpmuludq	4(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm4,%ymm4
-	vpmuludq	-124(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm0,%ymm0
-	vmovdqa	32(%esp),%ymm7
-	vpmuludq	-92(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpmuludq	-60(%edx),%ymm5,%ymm5
-	vpaddq	%ymm5,%ymm2,%ymm2
-	vpmuludq	-60(%edx),%ymm7,%ymm6
-	vpaddq	%ymm6,%ymm3,%ymm3
-	vpmuludq	-28(%edx),%ymm7,%ymm5
-	vpaddq	%ymm5,%ymm4,%ymm4
-	vpmuludq	132(%edx),%ymm7,%ymm6
-	vpaddq	%ymm6,%ymm0,%ymm0
-	vmovdqa	96(%esp),%ymm6
-	vpmuludq	-124(%edx),%ymm7,%ymm5
-	vpaddq	%ymm5,%ymm1,%ymm1
-	vpmuludq	-92(%edx),%ymm7,%ymm7
-	vpaddq	%ymm7,%ymm2,%ymm2
-	vpmuludq	-124(%edx),%ymm6,%ymm5
-	vpaddq	%ymm5,%ymm3,%ymm3
-	vpmuludq	-92(%edx),%ymm6,%ymm7
-	vpaddq	%ymm7,%ymm4,%ymm4
-	vpmuludq	68(%edx),%ymm6,%ymm5
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vmovdqa	128(%esp),%ymm5
-	vpmuludq	100(%edx),%ymm6,%ymm7
-	vpaddq	%ymm7,%ymm1,%ymm1
-	vpmuludq	132(%edx),%ymm6,%ymm6
-	vpaddq	%ymm6,%ymm2,%ymm2
-	vpmuludq	132(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm3,%ymm3
-	vpmuludq	36(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm0,%ymm0
-	vpmuludq	-124(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm4,%ymm4
-	vmovdqa	64(%ebx),%ymm7
-	vpmuludq	68(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpmuludq	100(%edx),%ymm5,%ymm5
-	vpaddq	%ymm5,%ymm2,%ymm2
-	vpsrldq	$8,%ymm4,%ymm5
-	vpsrldq	$8,%ymm3,%ymm6
-	vpaddq	%ymm5,%ymm4,%ymm4
-	vpsrldq	$8,%ymm0,%ymm5
-	vpaddq	%ymm6,%ymm3,%ymm3
-	vpsrldq	$8,%ymm1,%ymm6
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vpsrldq	$8,%ymm2,%ymm5
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpermq	$2,%ymm4,%ymm6
-	vpaddq	%ymm5,%ymm2,%ymm2
-	vpermq	$2,%ymm3,%ymm5
-	vpaddq	%ymm6,%ymm4,%ymm4
-	vpermq	$2,%ymm0,%ymm6
-	vpaddq	%ymm5,%ymm3,%ymm3
-	vpermq	$2,%ymm1,%ymm5
-	vpaddq	%ymm6,%ymm0,%ymm0
-	vpermq	$2,%ymm2,%ymm6
-	vpaddq	%ymm5,%ymm1,%ymm1
-	vpaddq	%ymm6,%ymm2,%ymm2
-	vpsrlq	$26,%ymm3,%ymm5
-	vpand	%ymm7,%ymm3,%ymm3
-	vpsrlq	$26,%ymm0,%ymm6
-	vpand	%ymm7,%ymm0,%ymm0
-	vpaddq	%ymm5,%ymm4,%ymm4
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpsrlq	$26,%ymm4,%ymm5
-	vpand	%ymm7,%ymm4,%ymm4
-	vpsrlq	$26,%ymm1,%ymm6
-	vpand	%ymm7,%ymm1,%ymm1
-	vpaddq	%ymm6,%ymm2,%ymm2
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vpsllq	$2,%ymm5,%ymm5
-	vpsrlq	$26,%ymm2,%ymm6
-	vpand	%ymm7,%ymm2,%ymm2
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vpaddq	%ymm6,%ymm3,%ymm3
-	vpsrlq	$26,%ymm3,%ymm6
-	vpsrlq	$26,%ymm0,%ymm5
-	vpand	%ymm7,%ymm0,%ymm0
-	vpand	%ymm7,%ymm3,%ymm3
-	vpaddq	%ymm5,%ymm1,%ymm1
-	vpaddq	%ymm6,%ymm4,%ymm4
-	cmpl	$0,%ecx
-	je	.L029done
-	vpshufd	$252,%xmm0,%xmm0
-	leal	288(%esp),%edx
-	vpshufd	$252,%xmm1,%xmm1
-	vpshufd	$252,%xmm2,%xmm2
-	vpshufd	$252,%xmm3,%xmm3
-	vpshufd	$252,%xmm4,%xmm4
-	jmp	.L024even
-.align	16
-.L029done:
-	vmovd	%xmm0,-48(%edi)
-	vmovd	%xmm1,-44(%edi)
-	vmovd	%xmm2,-40(%edi)
-	vmovd	%xmm3,-36(%edi)
-	vmovd	%xmm4,-32(%edi)
-	vzeroupper
-	movl	%ebp,%esp
-.L020nodata:
-	popl	%edi
-	popl	%esi
-	popl	%ebx
-	popl	%ebp
-	ret
-.size	_poly1305_blocks_avx2,.-_poly1305_blocks_avx2
 .align	64
 .Lconst_sse2:
 .long	16777216,0,16777216,0,16777216,0,16777216,0
@@ -1947,10 +1392,6 @@ poly1305_init:
 	jne	.L002no_sse2
 	leal	_poly1305_blocks_sse2-.L001pic_point(%ebx),%eax
 	leal	_poly1305_emit_sse2-.L001pic_point(%ebx),%edx
-	movl	8(%edi),%ecx
-	testl	$32,%ecx
-	jz	.L002no_sse2
-	leal	_poly1305_blocks_avx2-.L001pic_point(%ebx),%eax
 .L002no_sse2:
 	movl	20(%esp),%edi
 	movl	%eax,(%ebp)
@@ -3259,557 +2700,6 @@ _poly1305_emit_sse2:
 	popl	%ebp
 	ret
 .size	_poly1305_emit_sse2,.-_poly1305_emit_sse2
-.align	32
-.type	_poly1305_init_avx2,@function
-.align	16
-_poly1305_init_avx2:
-	vmovdqu	24(%edi),%xmm4
-	leal	48(%edi),%edi
-	movl	%esp,%ebp
-	subl	$224,%esp
-	andl	$-16,%esp
-	vmovdqa	64(%ebx),%xmm7
-	vpand	%xmm7,%xmm4,%xmm0
-	vpsrlq	$26,%xmm4,%xmm1
-	vpsrldq	$6,%xmm4,%xmm3
-	vpand	%xmm7,%xmm1,%xmm1
-	vpsrlq	$4,%xmm3,%xmm2
-	vpsrlq	$30,%xmm3,%xmm3
-	vpand	%xmm7,%xmm2,%xmm2
-	vpand	%xmm7,%xmm3,%xmm3
-	vpsrldq	$13,%xmm4,%xmm4
-	leal	144(%esp),%edx
-	movl	$2,%ecx
-.L018square:
-	vmovdqa	%xmm0,(%esp)
-	vmovdqa	%xmm1,16(%esp)
-	vmovdqa	%xmm2,32(%esp)
-	vmovdqa	%xmm3,48(%esp)
-	vmovdqa	%xmm4,64(%esp)
-	vpslld	$2,%xmm1,%xmm6
-	vpslld	$2,%xmm2,%xmm5
-	vpaddd	%xmm1,%xmm6,%xmm6
-	vpaddd	%xmm2,%xmm5,%xmm5
-	vmovdqa	%xmm6,80(%esp)
-	vmovdqa	%xmm5,96(%esp)
-	vpslld	$2,%xmm3,%xmm6
-	vpslld	$2,%xmm4,%xmm5
-	vpaddd	%xmm3,%xmm6,%xmm6
-	vpaddd	%xmm4,%xmm5,%xmm5
-	vmovdqa	%xmm6,112(%esp)
-	vmovdqa	%xmm5,128(%esp)
-	vpshufd	$68,%xmm0,%xmm5
-	vmovdqa	%xmm1,%xmm6
-	vpshufd	$68,%xmm1,%xmm1
-	vpshufd	$68,%xmm2,%xmm2
-	vpshufd	$68,%xmm3,%xmm3
-	vpshufd	$68,%xmm4,%xmm4
-	vmovdqa	%xmm5,(%edx)
-	vmovdqa	%xmm1,16(%edx)
-	vmovdqa	%xmm2,32(%edx)
-	vmovdqa	%xmm3,48(%edx)
-	vmovdqa	%xmm4,64(%edx)
-	vpmuludq	%xmm0,%xmm4,%xmm4
-	vpmuludq	%xmm0,%xmm3,%xmm3
-	vpmuludq	%xmm0,%xmm2,%xmm2
-	vpmuludq	%xmm0,%xmm1,%xmm1
-	vpmuludq	%xmm0,%xmm5,%xmm0
-	vpmuludq	48(%edx),%xmm6,%xmm5
-	vpaddq	%xmm5,%xmm4,%xmm4
-	vpmuludq	32(%edx),%xmm6,%xmm7
-	vpaddq	%xmm7,%xmm3,%xmm3
-	vpmuludq	16(%edx),%xmm6,%xmm5
-	vpaddq	%xmm5,%xmm2,%xmm2
-	vmovdqa	80(%esp),%xmm7
-	vpmuludq	(%edx),%xmm6,%xmm6
-	vpaddq	%xmm6,%xmm1,%xmm1
-	vmovdqa	32(%esp),%xmm5
-	vpmuludq	64(%edx),%xmm7,%xmm7
-	vpaddq	%xmm7,%xmm0,%xmm0
-	vpmuludq	32(%edx),%xmm5,%xmm6
-	vpaddq	%xmm6,%xmm4,%xmm4
-	vpmuludq	16(%edx),%xmm5,%xmm7
-	vpaddq	%xmm7,%xmm3,%xmm3
-	vmovdqa	96(%esp),%xmm6
-	vpmuludq	(%edx),%xmm5,%xmm5
-	vpaddq	%xmm5,%xmm2,%xmm2
-	vpmuludq	64(%edx),%xmm6,%xmm7
-	vpaddq	%xmm7,%xmm1,%xmm1
-	vmovdqa	48(%esp),%xmm5
-	vpmuludq	48(%edx),%xmm6,%xmm6
-	vpaddq	%xmm6,%xmm0,%xmm0
-	vpmuludq	16(%edx),%xmm5,%xmm7
-	vpaddq	%xmm7,%xmm4,%xmm4
-	vmovdqa	112(%esp),%xmm6
-	vpmuludq	(%edx),%xmm5,%xmm5
-	vpaddq	%xmm5,%xmm3,%xmm3
-	vpmuludq	64(%edx),%xmm6,%xmm7
-	vpaddq	%xmm7,%xmm2,%xmm2
-	vpmuludq	48(%edx),%xmm6,%xmm5
-	vpaddq	%xmm5,%xmm1,%xmm1
-	vmovdqa	64(%esp),%xmm7
-	vpmuludq	32(%edx),%xmm6,%xmm6
-	vpaddq	%xmm6,%xmm0,%xmm0
-	vmovdqa	128(%esp),%xmm5
-	vpmuludq	(%edx),%xmm7,%xmm7
-	vpaddq	%xmm7,%xmm4,%xmm4
-	vpmuludq	64(%edx),%xmm5,%xmm6
-	vpaddq	%xmm6,%xmm3,%xmm3
-	vpmuludq	16(%edx),%xmm5,%xmm7
-	vpaddq	%xmm7,%xmm0,%xmm0
-	vpmuludq	32(%edx),%xmm5,%xmm6
-	vpaddq	%xmm6,%xmm1,%xmm1
-	vmovdqa	64(%ebx),%xmm7
-	vpmuludq	48(%edx),%xmm5,%xmm5
-	vpaddq	%xmm5,%xmm2,%xmm2
-	vpsrlq	$26,%xmm3,%xmm5
-	vpand	%xmm7,%xmm3,%xmm3
-	vpsrlq	$26,%xmm0,%xmm6
-	vpand	%xmm7,%xmm0,%xmm0
-	vpaddq	%xmm5,%xmm4,%xmm4
-	vpaddq	%xmm6,%xmm1,%xmm1
-	vpsrlq	$26,%xmm4,%xmm5
-	vpand	%xmm7,%xmm4,%xmm4
-	vpsrlq	$26,%xmm1,%xmm6
-	vpand	%xmm7,%xmm1,%xmm1
-	vpaddq	%xmm6,%xmm2,%xmm2
-	vpaddd	%xmm5,%xmm0,%xmm0
-	vpsllq	$2,%xmm5,%xmm5
-	vpsrlq	$26,%xmm2,%xmm6
-	vpand	%xmm7,%xmm2,%xmm2
-	vpaddd	%xmm5,%xmm0,%xmm0
-	vpaddd	%xmm6,%xmm3,%xmm3
-	vpsrlq	$26,%xmm3,%xmm6
-	vpsrlq	$26,%xmm0,%xmm5
-	vpand	%xmm7,%xmm0,%xmm0
-	vpand	%xmm7,%xmm3,%xmm3
-	vpaddd	%xmm5,%xmm1,%xmm1
-	vpaddd	%xmm6,%xmm4,%xmm4
-	decl	%ecx
-	jz	.L019square_break
-	vpunpcklqdq	(%esp),%xmm0,%xmm0
-	vpunpcklqdq	16(%esp),%xmm1,%xmm1
-	vpunpcklqdq	32(%esp),%xmm2,%xmm2
-	vpunpcklqdq	48(%esp),%xmm3,%xmm3
-	vpunpcklqdq	64(%esp),%xmm4,%xmm4
-	jmp	.L018square
-.L019square_break:
-	vpsllq	$32,%xmm0,%xmm0
-	vpsllq	$32,%xmm1,%xmm1
-	vpsllq	$32,%xmm2,%xmm2
-	vpsllq	$32,%xmm3,%xmm3
-	vpsllq	$32,%xmm4,%xmm4
-	vpor	(%esp),%xmm0,%xmm0
-	vpor	16(%esp),%xmm1,%xmm1
-	vpor	32(%esp),%xmm2,%xmm2
-	vpor	48(%esp),%xmm3,%xmm3
-	vpor	64(%esp),%xmm4,%xmm4
-	vpshufd	$141,%xmm0,%xmm0
-	vpshufd	$141,%xmm1,%xmm1
-	vpshufd	$141,%xmm2,%xmm2
-	vpshufd	$141,%xmm3,%xmm3
-	vpshufd	$141,%xmm4,%xmm4
-	vmovdqu	%xmm0,(%edi)
-	vmovdqu	%xmm1,16(%edi)
-	vmovdqu	%xmm2,32(%edi)
-	vmovdqu	%xmm3,48(%edi)
-	vmovdqu	%xmm4,64(%edi)
-	vpslld	$2,%xmm1,%xmm6
-	vpslld	$2,%xmm2,%xmm5
-	vpaddd	%xmm1,%xmm6,%xmm6
-	vpaddd	%xmm2,%xmm5,%xmm5
-	vmovdqu	%xmm6,80(%edi)
-	vmovdqu	%xmm5,96(%edi)
-	vpslld	$2,%xmm3,%xmm6
-	vpslld	$2,%xmm4,%xmm5
-	vpaddd	%xmm3,%xmm6,%xmm6
-	vpaddd	%xmm4,%xmm5,%xmm5
-	vmovdqu	%xmm6,112(%edi)
-	vmovdqu	%xmm5,128(%edi)
-	movl	%ebp,%esp
-	leal	-48(%edi),%edi
-	ret
-.size	_poly1305_init_avx2,.-_poly1305_init_avx2
-.align	32
-.type	_poly1305_blocks_avx2,@function
-.align	16
-_poly1305_blocks_avx2:
-	pushl	%ebp
-	pushl	%ebx
-	pushl	%esi
-	pushl	%edi
-	movl	20(%esp),%edi
-	movl	24(%esp),%esi
-	movl	28(%esp),%ecx
-	movl	20(%edi),%eax
-	andl	$-16,%ecx
-	jz	.L020nodata
-	cmpl	$64,%ecx
-	jae	.L021enter_avx2
-	testl	%eax,%eax
-	jz	.Lenter_blocks
-.L021enter_avx2:
-	vzeroupper
-	call	.L022pic_point
-.L022pic_point:
-	popl	%ebx
-	leal	.Lconst_sse2-.L022pic_point(%ebx),%ebx
-	testl	%eax,%eax
-	jnz	.L023base2_26
-	call	_poly1305_init_avx2
-	movl	(%edi),%eax
-	movl	3(%edi),%ecx
-	movl	6(%edi),%edx
-	movl	9(%edi),%esi
-	movl	13(%edi),%ebp
-	shrl	$2,%ecx
-	andl	$67108863,%eax
-	shrl	$4,%edx
-	andl	$67108863,%ecx
-	shrl	$6,%esi
-	andl	$67108863,%edx
-	movl	%eax,(%edi)
-	movl	%ecx,4(%edi)
-	movl	%edx,8(%edi)
-	movl	%esi,12(%edi)
-	movl	%ebp,16(%edi)
-	movl	$1,20(%edi)
-	movl	24(%esp),%esi
-	movl	28(%esp),%ecx
-.L023base2_26:
-	movl	32(%esp),%eax
-	movl	%esp,%ebp
-	subl	$448,%esp
-	andl	$-512,%esp
-	vmovdqu	48(%edi),%xmm0
-	leal	288(%esp),%edx
-	vmovdqu	64(%edi),%xmm1
-	vmovdqu	80(%edi),%xmm2
-	vmovdqu	96(%edi),%xmm3
-	vmovdqu	112(%edi),%xmm4
-	leal	48(%edi),%edi
-	vpermq	$64,%ymm0,%ymm0
-	vpermq	$64,%ymm1,%ymm1
-	vpermq	$64,%ymm2,%ymm2
-	vpermq	$64,%ymm3,%ymm3
-	vpermq	$64,%ymm4,%ymm4
-	vpshufd	$200,%ymm0,%ymm0
-	vpshufd	$200,%ymm1,%ymm1
-	vpshufd	$200,%ymm2,%ymm2
-	vpshufd	$200,%ymm3,%ymm3
-	vpshufd	$200,%ymm4,%ymm4
-	vmovdqa	%ymm0,-128(%edx)
-	vmovdqu	80(%edi),%xmm0
-	vmovdqa	%ymm1,-96(%edx)
-	vmovdqu	96(%edi),%xmm1
-	vmovdqa	%ymm2,-64(%edx)
-	vmovdqu	112(%edi),%xmm2
-	vmovdqa	%ymm3,-32(%edx)
-	vmovdqu	128(%edi),%xmm3
-	vmovdqa	%ymm4,(%edx)
-	vpermq	$64,%ymm0,%ymm0
-	vpermq	$64,%ymm1,%ymm1
-	vpermq	$64,%ymm2,%ymm2
-	vpermq	$64,%ymm3,%ymm3
-	vpshufd	$200,%ymm0,%ymm0
-	vpshufd	$200,%ymm1,%ymm1
-	vpshufd	$200,%ymm2,%ymm2
-	vpshufd	$200,%ymm3,%ymm3
-	vmovdqa	%ymm0,32(%edx)
-	vmovd	-48(%edi),%xmm0
-	vmovdqa	%ymm1,64(%edx)
-	vmovd	-44(%edi),%xmm1
-	vmovdqa	%ymm2,96(%edx)
-	vmovd	-40(%edi),%xmm2
-	vmovdqa	%ymm3,128(%edx)
-	vmovd	-36(%edi),%xmm3
-	vmovd	-32(%edi),%xmm4
-	vmovdqa	64(%ebx),%ymm7
-	negl	%eax
-	testl	$63,%ecx
-	jz	.L024even
-	movl	%ecx,%edx
-	andl	$-64,%ecx
-	andl	$63,%edx
-	vmovdqu	(%esi),%xmm5
-	cmpl	$32,%edx
-	jb	.L025one
-	vmovdqu	16(%esi),%xmm6
-	je	.L026two
-	vinserti128	$1,32(%esi),%ymm5,%ymm5
-	leal	48(%esi),%esi
-	leal	8(%ebx),%ebx
-	leal	296(%esp),%edx
-	jmp	.L027tail
-.L026two:
-	leal	32(%esi),%esi
-	leal	16(%ebx),%ebx
-	leal	304(%esp),%edx
-	jmp	.L027tail
-.L025one:
-	leal	16(%esi),%esi
-	vpxor	%ymm6,%ymm6,%ymm6
-	leal	32(%ebx,%eax,8),%ebx
-	leal	312(%esp),%edx
-	jmp	.L027tail
-.align	32
-.L024even:
-	vmovdqu	(%esi),%xmm5
-	vmovdqu	16(%esi),%xmm6
-	vinserti128	$1,32(%esi),%ymm5,%ymm5
-	vinserti128	$1,48(%esi),%ymm6,%ymm6
-	leal	64(%esi),%esi
-	subl	$64,%ecx
-	jz	.L027tail
-.L028loop:
-	vmovdqa	%ymm2,64(%esp)
-	vpsrldq	$6,%ymm5,%ymm2
-	vmovdqa	%ymm0,(%esp)
-	vpsrldq	$6,%ymm6,%ymm0
-	vmovdqa	%ymm1,32(%esp)
-	vpunpckhqdq	%ymm6,%ymm5,%ymm1
-	vpunpcklqdq	%ymm6,%ymm5,%ymm5
-	vpunpcklqdq	%ymm0,%ymm2,%ymm2
-	vpsrlq	$30,%ymm2,%ymm0
-	vpsrlq	$4,%ymm2,%ymm2
-	vpsrlq	$26,%ymm5,%ymm6
-	vpsrlq	$40,%ymm1,%ymm1
-	vpand	%ymm7,%ymm2,%ymm2
-	vpand	%ymm7,%ymm5,%ymm5
-	vpand	%ymm7,%ymm6,%ymm6
-	vpand	%ymm7,%ymm0,%ymm0
-	vpor	(%ebx),%ymm1,%ymm1
-	vpaddq	64(%esp),%ymm2,%ymm2
-	vpaddq	(%esp),%ymm5,%ymm5
-	vpaddq	32(%esp),%ymm6,%ymm6
-	vpaddq	%ymm3,%ymm0,%ymm0
-	vpaddq	%ymm4,%ymm1,%ymm1
-	vpmuludq	-96(%edx),%ymm2,%ymm3
-	vmovdqa	%ymm6,32(%esp)
-	vpmuludq	-64(%edx),%ymm2,%ymm4
-	vmovdqa	%ymm0,96(%esp)
-	vpmuludq	96(%edx),%ymm2,%ymm0
-	vmovdqa	%ymm1,128(%esp)
-	vpmuludq	128(%edx),%ymm2,%ymm1
-	vpmuludq	-128(%edx),%ymm2,%ymm2
-	vpmuludq	-32(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm3,%ymm3
-	vpmuludq	(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm4,%ymm4
-	vpmuludq	-128(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm0,%ymm0
-	vmovdqa	32(%esp),%ymm7
-	vpmuludq	-96(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpmuludq	-64(%edx),%ymm5,%ymm5
-	vpaddq	%ymm5,%ymm2,%ymm2
-	vpmuludq	-64(%edx),%ymm7,%ymm6
-	vpaddq	%ymm6,%ymm3,%ymm3
-	vpmuludq	-32(%edx),%ymm7,%ymm5
-	vpaddq	%ymm5,%ymm4,%ymm4
-	vpmuludq	128(%edx),%ymm7,%ymm6
-	vpaddq	%ymm6,%ymm0,%ymm0
-	vmovdqa	96(%esp),%ymm6
-	vpmuludq	-128(%edx),%ymm7,%ymm5
-	vpaddq	%ymm5,%ymm1,%ymm1
-	vpmuludq	-96(%edx),%ymm7,%ymm7
-	vpaddq	%ymm7,%ymm2,%ymm2
-	vpmuludq	-128(%edx),%ymm6,%ymm5
-	vpaddq	%ymm5,%ymm3,%ymm3
-	vpmuludq	-96(%edx),%ymm6,%ymm7
-	vpaddq	%ymm7,%ymm4,%ymm4
-	vpmuludq	64(%edx),%ymm6,%ymm5
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vmovdqa	128(%esp),%ymm5
-	vpmuludq	96(%edx),%ymm6,%ymm7
-	vpaddq	%ymm7,%ymm1,%ymm1
-	vpmuludq	128(%edx),%ymm6,%ymm6
-	vpaddq	%ymm6,%ymm2,%ymm2
-	vpmuludq	128(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm3,%ymm3
-	vpmuludq	32(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm0,%ymm0
-	vpmuludq	-128(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm4,%ymm4
-	vmovdqa	64(%ebx),%ymm7
-	vpmuludq	64(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpmuludq	96(%edx),%ymm5,%ymm5
-	vpaddq	%ymm5,%ymm2,%ymm2
-	vpsrlq	$26,%ymm3,%ymm5
-	vpand	%ymm7,%ymm3,%ymm3
-	vpsrlq	$26,%ymm0,%ymm6
-	vpand	%ymm7,%ymm0,%ymm0
-	vpaddq	%ymm5,%ymm4,%ymm4
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpsrlq	$26,%ymm4,%ymm5
-	vpand	%ymm7,%ymm4,%ymm4
-	vpsrlq	$26,%ymm1,%ymm6
-	vpand	%ymm7,%ymm1,%ymm1
-	vpaddq	%ymm6,%ymm2,%ymm2
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vpsllq	$2,%ymm5,%ymm5
-	vpsrlq	$26,%ymm2,%ymm6
-	vpand	%ymm7,%ymm2,%ymm2
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vpaddq	%ymm6,%ymm3,%ymm3
-	vpsrlq	$26,%ymm3,%ymm6
-	vpsrlq	$26,%ymm0,%ymm5
-	vpand	%ymm7,%ymm0,%ymm0
-	vpand	%ymm7,%ymm3,%ymm3
-	vpaddq	%ymm5,%ymm1,%ymm1
-	vpaddq	%ymm6,%ymm4,%ymm4
-	vmovdqu	(%esi),%xmm5
-	vmovdqu	16(%esi),%xmm6
-	vinserti128	$1,32(%esi),%ymm5,%ymm5
-	vinserti128	$1,48(%esi),%ymm6,%ymm6
-	leal	64(%esi),%esi
-	subl	$64,%ecx
-	jnz	.L028loop
-.L027tail:
-	vmovdqa	%ymm2,64(%esp)
-	vpsrldq	$6,%ymm5,%ymm2
-	vmovdqa	%ymm0,(%esp)
-	vpsrldq	$6,%ymm6,%ymm0
-	vmovdqa	%ymm1,32(%esp)
-	vpunpckhqdq	%ymm6,%ymm5,%ymm1
-	vpunpcklqdq	%ymm6,%ymm5,%ymm5
-	vpunpcklqdq	%ymm0,%ymm2,%ymm2
-	vpsrlq	$30,%ymm2,%ymm0
-	vpsrlq	$4,%ymm2,%ymm2
-	vpsrlq	$26,%ymm5,%ymm6
-	vpsrlq	$40,%ymm1,%ymm1
-	vpand	%ymm7,%ymm2,%ymm2
-	vpand	%ymm7,%ymm5,%ymm5
-	vpand	%ymm7,%ymm6,%ymm6
-	vpand	%ymm7,%ymm0,%ymm0
-	vpor	(%ebx),%ymm1,%ymm1
-	andl	$-64,%ebx
-	vpaddq	64(%esp),%ymm2,%ymm2
-	vpaddq	(%esp),%ymm5,%ymm5
-	vpaddq	32(%esp),%ymm6,%ymm6
-	vpaddq	%ymm3,%ymm0,%ymm0
-	vpaddq	%ymm4,%ymm1,%ymm1
-	vpmuludq	-92(%edx),%ymm2,%ymm3
-	vmovdqa	%ymm6,32(%esp)
-	vpmuludq	-60(%edx),%ymm2,%ymm4
-	vmovdqa	%ymm0,96(%esp)
-	vpmuludq	100(%edx),%ymm2,%ymm0
-	vmovdqa	%ymm1,128(%esp)
-	vpmuludq	132(%edx),%ymm2,%ymm1
-	vpmuludq	-124(%edx),%ymm2,%ymm2
-	vpmuludq	-28(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm3,%ymm3
-	vpmuludq	4(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm4,%ymm4
-	vpmuludq	-124(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm0,%ymm0
-	vmovdqa	32(%esp),%ymm7
-	vpmuludq	-92(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpmuludq	-60(%edx),%ymm5,%ymm5
-	vpaddq	%ymm5,%ymm2,%ymm2
-	vpmuludq	-60(%edx),%ymm7,%ymm6
-	vpaddq	%ymm6,%ymm3,%ymm3
-	vpmuludq	-28(%edx),%ymm7,%ymm5
-	vpaddq	%ymm5,%ymm4,%ymm4
-	vpmuludq	132(%edx),%ymm7,%ymm6
-	vpaddq	%ymm6,%ymm0,%ymm0
-	vmovdqa	96(%esp),%ymm6
-	vpmuludq	-124(%edx),%ymm7,%ymm5
-	vpaddq	%ymm5,%ymm1,%ymm1
-	vpmuludq	-92(%edx),%ymm7,%ymm7
-	vpaddq	%ymm7,%ymm2,%ymm2
-	vpmuludq	-124(%edx),%ymm6,%ymm5
-	vpaddq	%ymm5,%ymm3,%ymm3
-	vpmuludq	-92(%edx),%ymm6,%ymm7
-	vpaddq	%ymm7,%ymm4,%ymm4
-	vpmuludq	68(%edx),%ymm6,%ymm5
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vmovdqa	128(%esp),%ymm5
-	vpmuludq	100(%edx),%ymm6,%ymm7
-	vpaddq	%ymm7,%ymm1,%ymm1
-	vpmuludq	132(%edx),%ymm6,%ymm6
-	vpaddq	%ymm6,%ymm2,%ymm2
-	vpmuludq	132(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm3,%ymm3
-	vpmuludq	36(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm0,%ymm0
-	vpmuludq	-124(%edx),%ymm5,%ymm7
-	vpaddq	%ymm7,%ymm4,%ymm4
-	vmovdqa	64(%ebx),%ymm7
-	vpmuludq	68(%edx),%ymm5,%ymm6
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpmuludq	100(%edx),%ymm5,%ymm5
-	vpaddq	%ymm5,%ymm2,%ymm2
-	vpsrldq	$8,%ymm4,%ymm5
-	vpsrldq	$8,%ymm3,%ymm6
-	vpaddq	%ymm5,%ymm4,%ymm4
-	vpsrldq	$8,%ymm0,%ymm5
-	vpaddq	%ymm6,%ymm3,%ymm3
-	vpsrldq	$8,%ymm1,%ymm6
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vpsrldq	$8,%ymm2,%ymm5
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpermq	$2,%ymm4,%ymm6
-	vpaddq	%ymm5,%ymm2,%ymm2
-	vpermq	$2,%ymm3,%ymm5
-	vpaddq	%ymm6,%ymm4,%ymm4
-	vpermq	$2,%ymm0,%ymm6
-	vpaddq	%ymm5,%ymm3,%ymm3
-	vpermq	$2,%ymm1,%ymm5
-	vpaddq	%ymm6,%ymm0,%ymm0
-	vpermq	$2,%ymm2,%ymm6
-	vpaddq	%ymm5,%ymm1,%ymm1
-	vpaddq	%ymm6,%ymm2,%ymm2
-	vpsrlq	$26,%ymm3,%ymm5
-	vpand	%ymm7,%ymm3,%ymm3
-	vpsrlq	$26,%ymm0,%ymm6
-	vpand	%ymm7,%ymm0,%ymm0
-	vpaddq	%ymm5,%ymm4,%ymm4
-	vpaddq	%ymm6,%ymm1,%ymm1
-	vpsrlq	$26,%ymm4,%ymm5
-	vpand	%ymm7,%ymm4,%ymm4
-	vpsrlq	$26,%ymm1,%ymm6
-	vpand	%ymm7,%ymm1,%ymm1
-	vpaddq	%ymm6,%ymm2,%ymm2
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vpsllq	$2,%ymm5,%ymm5
-	vpsrlq	$26,%ymm2,%ymm6
-	vpand	%ymm7,%ymm2,%ymm2
-	vpaddq	%ymm5,%ymm0,%ymm0
-	vpaddq	%ymm6,%ymm3,%ymm3
-	vpsrlq	$26,%ymm3,%ymm6
-	vpsrlq	$26,%ymm0,%ymm5
-	vpand	%ymm7,%ymm0,%ymm0
-	vpand	%ymm7,%ymm3,%ymm3
-	vpaddq	%ymm5,%ymm1,%ymm1
-	vpaddq	%ymm6,%ymm4,%ymm4
-	cmpl	$0,%ecx
-	je	.L029done
-	vpshufd	$252,%xmm0,%xmm0
-	leal	288(%esp),%edx
-	vpshufd	$252,%xmm1,%xmm1
-	vpshufd	$252,%xmm2,%xmm2
-	vpshufd	$252,%xmm3,%xmm3
-	vpshufd	$252,%xmm4,%xmm4
-	jmp	.L024even
-.align	16
-.L029done:
-	vmovd	%xmm0,-48(%edi)
-	vmovd	%xmm1,-44(%edi)
-	vmovd	%xmm2,-40(%edi)
-	vmovd	%xmm3,-36(%edi)
-	vmovd	%xmm4,-32(%edi)
-	vzeroupper
-	movl	%ebp,%esp
-.L020nodata:
-	popl	%edi
-	popl	%esi
-	popl	%ebx
-	popl	%ebp
-	ret
-.size	_poly1305_blocks_avx2,.-_poly1305_blocks_avx2
 .align	64
 .Lconst_sse2:
 .long	16777216,0,16777216,0,16777216,0,16777216,0
diff --git a/secure/lib/libcrypto/i386/sha1-586.S b/secure/lib/libcrypto/i386/sha1-586.S
index 7e90e2d9b1d2..49e7482b8161 100644
--- a/secure/lib/libcrypto/i386/sha1-586.S
+++ b/secure/lib/libcrypto/i386/sha1-586.S
@@ -25,11 +25,6 @@ sha1_block_data_order:
 	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:
@@ -2787,1176 +2782,6 @@ _sha1_block_data_order_ssse3:
 	popl	%ebp
 	ret
 .size	_sha1_block_data_order_ssse3,.-_sha1_block_data_order_ssse3
-.type	_sha1_block_data_order_avx,@function
-.align	16
-_sha1_block_data_order_avx:
-	pushl	%ebp
-	pushl	%ebx
-	pushl	%esi
-	pushl	%edi
-	call	.L008pic_point
-.L008pic_point:
-	popl	%ebp
-	leal	.LK_XX_XX-.L008pic_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	.L009loop
-.align	16
-.L009loop:
-	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	.L010done
-	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	.L009loop
-.align	16
-.L010done:
-	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
@@ -3995,11 +2820,6 @@ sha1_block_data_order:
 	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:
@@ -6757,1176 +5577,6 @@ _sha1_block_data_order_ssse3:
 	popl	%ebp
 	ret
 .size	_sha1_block_data_order_ssse3,.-_sha1_block_data_order_ssse3
-.type	_sha1_block_data_order_avx,@function
-.align	16
-_sha1_block_data_order_avx:
-	pushl	%ebp
-	pushl	%ebx
-	pushl	%esi
-	pushl	%edi
-	call	.L008pic_point
-.L008pic_point:
-	popl	%ebp
-	leal	.LK_XX_XX-.L008pic_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	.L009loop
-.align	16
-.L009loop:
-	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	.L010done
-	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	.L009loop
-.align	16
-.L010done:
-	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
diff --git a/secure/lib/libcrypto/i386/sha256-586.S b/secure/lib/libcrypto/i386/sha256-586.S
index 7b4205352bdf..5d8476c1e1bb 100644
--- a/secure/lib/libcrypto/i386/sha256-586.S
+++ b/secure/lib/libcrypto/i386/sha256-586.S
@@ -42,13 +42,12 @@ sha256_block_data_order:
 	orl	%ebx,%ecx
 	andl	$1342177280,%ecx
 	cmpl	$1342177280,%ecx
-	je	.L005AVX
 	testl	$512,%ebx
-	jnz	.L006SSSE3
+	jnz	.L005SSSE3
 .L003no_xmm:
 	subl	%edi,%eax
 	cmpl	$256,%eax
-	jae	.L007unrolled
+	jae	.L006unrolled
 	jmp	.L002loop
 .align	16
 .L002loop:
@@ -120,7 +119,7 @@ sha256_block_data_order:
 	movl	%ecx,28(%esp)
 	movl	%edi,32(%esp)
 .align	16
-.L00800_15:
+.L00700_15:
 	movl	%edx,%ecx
 	movl	24(%esp),%esi
 	rorl	$14,%ecx
@@ -158,11 +157,11 @@ sha256_block_data_order:
 	addl	$4,%ebp
 	addl	%ebx,%eax
 	cmpl	$3248222580,%esi
-	jne	.L00800_15
+	jne	.L00700_15
 	movl	156(%esp),%ecx
-	jmp	.L00916_63
+	jmp	.L00816_63
 .align	16
-.L00916_63:
+.L00816_63:
 	movl	%ecx,%ebx
 	movl	104(%esp),%esi
 	rorl	$11,%ecx
@@ -217,7 +216,7 @@ sha256_block_data_order:
 	addl	$4,%ebp
 	addl	%ebx,%eax
 	cmpl	$3329325298,%esi
-	jne	.L00916_63
+	jne	.L00816_63
 	movl	356(%esp),%esi
 	movl	8(%esp),%ebx
 	movl	16(%esp),%ecx
@@ -261,7 +260,7 @@ sha256_block_data_order:
 .byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
 .byte	62,0
 .align	16
-.L007unrolled:
+.L006unrolled:
 	leal	-96(%esp),%esp
 	movl	(%esi),%eax
 	movl	4(%esi),%ebp
@@ -278,9 +277,9 @@ sha256_block_data_order:
 	movl	%ebx,20(%esp)
 	movl	%ecx,24(%esp)
 	movl	%esi,28(%esp)
-	jmp	.L010grand_loop
+	jmp	.L009grand_loop
 .align	16
-.L010grand_loop:
+.L009grand_loop:
 	movl	(%edi),%ebx
 	movl	4(%edi),%ecx
 	bswap	%ebx
@@ -3160,7 +3159,7 @@ sha256_block_data_order:
 	movl	%ebx,24(%esp)
 	movl	%ecx,28(%esp)
 	cmpl	104(%esp),%edi
-	jb	.L010grand_loop
+	jb	.L009grand_loop
 	movl	108(%esp),%esp
 	popl	%edi
 	popl	%esi
@@ -3179,9 +3178,9 @@ sha256_block_data_order:
 	pshufd	$27,%xmm2,%xmm2
 .byte	102,15,58,15,202,8
 	punpcklqdq	%xmm0,%xmm2
-	jmp	.L011loop_shaext
+	jmp	.L010loop_shaext
 .align	16
-.L011loop_shaext:
+.L010loop_shaext:
 	movdqu	(%edi),%xmm3
 	movdqu	16(%edi),%xmm4
 	movdqu	32(%edi),%xmm5
@@ -3351,7 +3350,7 @@ sha256_block_data_order:
 .byte	15,56,203,202
 	paddd	16(%esp),%xmm2
 	paddd	(%esp),%xmm1
-	jnz	.L011loop_shaext
+	jnz	.L010loop_shaext
 	pshufd	$177,%xmm2,%xmm2
 	pshufd	$27,%xmm1,%xmm7
 	pshufd	$177,%xmm1,%xmm1
@@ -3366,7 +3365,7 @@ sha256_block_data_order:
 	popl	%ebp
 	ret
 .align	32
-.L006SSSE3:
+.L005SSSE3:
 	leal	-96(%esp),%esp
 	movl	(%esi),%eax
 	movl	4(%esi),%ebx
@@ -3385,9 +3384,9 @@ sha256_block_data_order:
 	movl	%ecx,24(%esp)
 	movl	%esi,28(%esp)
 	movdqa	256(%ebp),%xmm7
-	jmp	.L012grand_ssse3
+	jmp	.L011grand_ssse3
 .align	16
-.L012grand_ssse3:
+.L011grand_ssse3:
 	movdqu	(%edi),%xmm0
 	movdqu	16(%edi),%xmm1
 	movdqu	32(%edi),%xmm2
@@ -3410,9 +3409,9 @@ sha256_block_data_order:
 	paddd	%xmm3,%xmm7
 	movdqa	%xmm6,64(%esp)
 	movdqa	%xmm7,80(%esp)
-	jmp	.L013ssse3_00_47
+	jmp	.L012ssse3_00_47
 .align	16
-.L013ssse3_00_47:
+.L012ssse3_00_47:
 	addl	$64,%ebp
 	movl	%edx,%ecx
 	movdqa	%xmm1,%xmm4
@@ -4055,7 +4054,7 @@ sha256_block_data_order:
 	addl	%ecx,%eax
 	movdqa	%xmm6,80(%esp)
 	cmpl	$66051,64(%ebp)
-	jne	.L013ssse3_00_47
+	jne	.L012ssse3_00_47
 	movl	%edx,%ecx
 	rorl	$14,%edx
 	movl	20(%esp),%esi
@@ -4569,2218 +4568,13 @@ sha256_block_data_order:
 	movdqa	64(%ebp),%xmm7
 	subl	$192,%ebp
 	cmpl	104(%esp),%edi
-	jb	.L012grand_ssse3
+	jb	.L011grand_ssse3
 	movl	108(%esp),%esp
 	popl	%edi
 	popl	%esi
 	popl	%ebx
 	popl	%ebp
 	ret
-.align	32
-.L005AVX:
-	andl	$264,%edx
-	cmpl	$264,%edx
-	je	.L014AVX_BMI
-	leal	-96(%esp),%esp
-	vzeroall
-	movl	(%esi),%eax
-	movl	4(%esi),%ebx
-	movl	8(%esi),%ecx
-	movl	12(%esi),%edi
-	movl	%ebx,4(%esp)
-	xorl	%ecx,%ebx
-	movl	%ecx,8(%esp)
-	movl	%edi,12(%esp)
-	movl	16(%esi),%edx
-	movl	20(%esi),%edi
-	movl	24(%esi),%ecx
-	movl	28(%esi),%esi
-	movl	%edi,20(%esp)
-	movl	100(%esp),%edi
-	movl	%ecx,24(%esp)
-	movl	%esi,28(%esp)
-	vmovdqa	256(%ebp),%xmm7
-	jmp	.L015grand_avx
-.align	32
-.L015grand_avx:
-	vmovdqu	(%edi),%xmm0
-	vmovdqu	16(%edi),%xmm1
-	vmovdqu	32(%edi),%xmm2
-	vmovdqu	48(%edi),%xmm3
-	addl	$64,%edi
-	vpshufb	%xmm7,%xmm0,%xmm0
-	movl	%edi,100(%esp)
-	vpshufb	%xmm7,%xmm1,%xmm1
-	vpshufb	%xmm7,%xmm2,%xmm2
-	vpaddd	(%ebp),%xmm0,%xmm4
-	vpshufb	%xmm7,%xmm3,%xmm3
-	vpaddd	16(%ebp),%xmm1,%xmm5
-	vpaddd	32(%ebp),%xmm2,%xmm6
-	vpaddd	48(%ebp),%xmm3,%xmm7
-	vmovdqa	%xmm4,32(%esp)
-	vmovdqa	%xmm5,48(%esp)
-	vmovdqa	%xmm6,64(%esp)
-	vmovdqa	%xmm7,80(%esp)
-	jmp	.L016avx_00_47
-.align	16
-.L016avx_00_47:
-	addl	$64,%ebp
-	vpalignr	$4,%xmm0,%xmm1,%xmm4
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	20(%esp),%esi
-	vpalignr	$4,%xmm2,%xmm3,%xmm7
-	xorl	%ecx,%edx
-	movl	24(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,16(%esp)
-	vpaddd	%xmm7,%xmm0,%xmm0
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrld	$3,%xmm4,%xmm7
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	4(%esp),%edi
-	vpslld	$14,%xmm4,%xmm5
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,(%esp)
-	vpxor	%xmm6,%xmm7,%xmm4
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	vpshufd	$250,%xmm3,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpsrld	$11,%xmm6,%xmm6
-	addl	32(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpxor	%xmm5,%xmm4,%xmm4
-	addl	%edx,%ebx
-	addl	12(%esp),%edx
-	addl	%ecx,%ebx
-	vpslld	$11,%xmm5,%xmm5
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	16(%esp),%esi
-	vpxor	%xmm6,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	movl	20(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,12(%esp)
-	vpxor	%xmm5,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	(%esp),%edi
-	vpaddd	%xmm4,%xmm0,%xmm0
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,28(%esp)
-	vpxor	%xmm5,%xmm6,%xmm6
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	vpsrlq	$19,%xmm7,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	addl	36(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	vpshufd	$132,%xmm6,%xmm7
-	addl	%edx,%eax
-	addl	8(%esp),%edx
-	addl	%ecx,%eax
-	vpsrldq	$8,%xmm7,%xmm7
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	12(%esp),%esi
-	vpaddd	%xmm7,%xmm0,%xmm0
-	xorl	%ecx,%edx
-	movl	16(%esp),%edi
-	xorl	%edi,%esi
-	vpshufd	$80,%xmm0,%xmm7
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,8(%esp)
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	28(%esp),%edi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,24(%esp)
-	vpsrlq	$19,%xmm7,%xmm7
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpshufd	$232,%xmm6,%xmm7
-	addl	40(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpslldq	$8,%xmm7,%xmm7
-	addl	%edx,%ebx
-	addl	4(%esp),%edx
-	addl	%ecx,%ebx
-	vpaddd	%xmm7,%xmm0,%xmm0
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	8(%esp),%esi
-	vpaddd	(%ebp),%xmm0,%xmm6
-	xorl	%ecx,%edx
-	movl	12(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,4(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	24(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,20(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	44(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	(%esp),%edx
-	addl	%ecx,%eax
-	vmovdqa	%xmm6,32(%esp)
-	vpalignr	$4,%xmm1,%xmm2,%xmm4
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	4(%esp),%esi
-	vpalignr	$4,%xmm3,%xmm0,%xmm7
-	xorl	%ecx,%edx
-	movl	8(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,(%esp)
-	vpaddd	%xmm7,%xmm1,%xmm1
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrld	$3,%xmm4,%xmm7
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	20(%esp),%edi
-	vpslld	$14,%xmm4,%xmm5
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,16(%esp)
-	vpxor	%xmm6,%xmm7,%xmm4
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	vpshufd	$250,%xmm0,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpsrld	$11,%xmm6,%xmm6
-	addl	48(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpxor	%xmm5,%xmm4,%xmm4
-	addl	%edx,%ebx
-	addl	28(%esp),%edx
-	addl	%ecx,%ebx
-	vpslld	$11,%xmm5,%xmm5
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	(%esp),%esi
-	vpxor	%xmm6,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	movl	4(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,28(%esp)
-	vpxor	%xmm5,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	16(%esp),%edi
-	vpaddd	%xmm4,%xmm1,%xmm1
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,12(%esp)
-	vpxor	%xmm5,%xmm6,%xmm6
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	vpsrlq	$19,%xmm7,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	addl	52(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	vpshufd	$132,%xmm6,%xmm7
-	addl	%edx,%eax
-	addl	24(%esp),%edx
-	addl	%ecx,%eax
-	vpsrldq	$8,%xmm7,%xmm7
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	28(%esp),%esi
-	vpaddd	%xmm7,%xmm1,%xmm1
-	xorl	%ecx,%edx
-	movl	(%esp),%edi
-	xorl	%edi,%esi
-	vpshufd	$80,%xmm1,%xmm7
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,24(%esp)
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	12(%esp),%edi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,8(%esp)
-	vpsrlq	$19,%xmm7,%xmm7
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpshufd	$232,%xmm6,%xmm7
-	addl	56(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpslldq	$8,%xmm7,%xmm7
-	addl	%edx,%ebx
-	addl	20(%esp),%edx
-	addl	%ecx,%ebx
-	vpaddd	%xmm7,%xmm1,%xmm1
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	24(%esp),%esi
-	vpaddd	16(%ebp),%xmm1,%xmm6
-	xorl	%ecx,%edx
-	movl	28(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,20(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	8(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,4(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	60(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	16(%esp),%edx
-	addl	%ecx,%eax
-	vmovdqa	%xmm6,48(%esp)
-	vpalignr	$4,%xmm2,%xmm3,%xmm4
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	20(%esp),%esi
-	vpalignr	$4,%xmm0,%xmm1,%xmm7
-	xorl	%ecx,%edx
-	movl	24(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,16(%esp)
-	vpaddd	%xmm7,%xmm2,%xmm2
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrld	$3,%xmm4,%xmm7
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	4(%esp),%edi
-	vpslld	$14,%xmm4,%xmm5
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,(%esp)
-	vpxor	%xmm6,%xmm7,%xmm4
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	vpshufd	$250,%xmm1,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpsrld	$11,%xmm6,%xmm6
-	addl	64(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpxor	%xmm5,%xmm4,%xmm4
-	addl	%edx,%ebx
-	addl	12(%esp),%edx
-	addl	%ecx,%ebx
-	vpslld	$11,%xmm5,%xmm5
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	16(%esp),%esi
-	vpxor	%xmm6,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	movl	20(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,12(%esp)
-	vpxor	%xmm5,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	(%esp),%edi
-	vpaddd	%xmm4,%xmm2,%xmm2
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,28(%esp)
-	vpxor	%xmm5,%xmm6,%xmm6
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	vpsrlq	$19,%xmm7,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	addl	68(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	vpshufd	$132,%xmm6,%xmm7
-	addl	%edx,%eax
-	addl	8(%esp),%edx
-	addl	%ecx,%eax
-	vpsrldq	$8,%xmm7,%xmm7
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	12(%esp),%esi
-	vpaddd	%xmm7,%xmm2,%xmm2
-	xorl	%ecx,%edx
-	movl	16(%esp),%edi
-	xorl	%edi,%esi
-	vpshufd	$80,%xmm2,%xmm7
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,8(%esp)
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	28(%esp),%edi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,24(%esp)
-	vpsrlq	$19,%xmm7,%xmm7
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpshufd	$232,%xmm6,%xmm7
-	addl	72(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpslldq	$8,%xmm7,%xmm7
-	addl	%edx,%ebx
-	addl	4(%esp),%edx
-	addl	%ecx,%ebx
-	vpaddd	%xmm7,%xmm2,%xmm2
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	8(%esp),%esi
-	vpaddd	32(%ebp),%xmm2,%xmm6
-	xorl	%ecx,%edx
-	movl	12(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,4(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	24(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,20(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	76(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	(%esp),%edx
-	addl	%ecx,%eax
-	vmovdqa	%xmm6,64(%esp)
-	vpalignr	$4,%xmm3,%xmm0,%xmm4
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	4(%esp),%esi
-	vpalignr	$4,%xmm1,%xmm2,%xmm7
-	xorl	%ecx,%edx
-	movl	8(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,(%esp)
-	vpaddd	%xmm7,%xmm3,%xmm3
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrld	$3,%xmm4,%xmm7
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	20(%esp),%edi
-	vpslld	$14,%xmm4,%xmm5
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,16(%esp)
-	vpxor	%xmm6,%xmm7,%xmm4
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	vpshufd	$250,%xmm2,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpsrld	$11,%xmm6,%xmm6
-	addl	80(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpxor	%xmm5,%xmm4,%xmm4
-	addl	%edx,%ebx
-	addl	28(%esp),%edx
-	addl	%ecx,%ebx
-	vpslld	$11,%xmm5,%xmm5
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	(%esp),%esi
-	vpxor	%xmm6,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	movl	4(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,28(%esp)
-	vpxor	%xmm5,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	16(%esp),%edi
-	vpaddd	%xmm4,%xmm3,%xmm3
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,12(%esp)
-	vpxor	%xmm5,%xmm6,%xmm6
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	vpsrlq	$19,%xmm7,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	addl	84(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	vpshufd	$132,%xmm6,%xmm7
-	addl	%edx,%eax
-	addl	24(%esp),%edx
-	addl	%ecx,%eax
-	vpsrldq	$8,%xmm7,%xmm7
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	28(%esp),%esi
-	vpaddd	%xmm7,%xmm3,%xmm3
-	xorl	%ecx,%edx
-	movl	(%esp),%edi
-	xorl	%edi,%esi
-	vpshufd	$80,%xmm3,%xmm7
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,24(%esp)
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	12(%esp),%edi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,8(%esp)
-	vpsrlq	$19,%xmm7,%xmm7
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpshufd	$232,%xmm6,%xmm7
-	addl	88(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpslldq	$8,%xmm7,%xmm7
-	addl	%edx,%ebx
-	addl	20(%esp),%edx
-	addl	%ecx,%ebx
-	vpaddd	%xmm7,%xmm3,%xmm3
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	24(%esp),%esi
-	vpaddd	48(%ebp),%xmm3,%xmm6
-	xorl	%ecx,%edx
-	movl	28(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,20(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	8(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,4(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	92(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	16(%esp),%edx
-	addl	%ecx,%eax
-	vmovdqa	%xmm6,80(%esp)
-	cmpl	$66051,64(%ebp)
-	jne	.L016avx_00_47
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	20(%esp),%esi
-	xorl	%ecx,%edx
-	movl	24(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,16(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	4(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	32(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	12(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	16(%esp),%esi
-	xorl	%ecx,%edx
-	movl	20(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,12(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,28(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	36(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	8(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	12(%esp),%esi
-	xorl	%ecx,%edx
-	movl	16(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,8(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	28(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,24(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	40(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	4(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	8(%esp),%esi
-	xorl	%ecx,%edx
-	movl	12(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,4(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	24(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,20(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	44(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	4(%esp),%esi
-	xorl	%ecx,%edx
-	movl	8(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	20(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,16(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	48(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	28(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	(%esp),%esi
-	xorl	%ecx,%edx
-	movl	4(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,28(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	16(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,12(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	52(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	24(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	28(%esp),%esi
-	xorl	%ecx,%edx
-	movl	(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,24(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	12(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,8(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	56(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	20(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	24(%esp),%esi
-	xorl	%ecx,%edx
-	movl	28(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,20(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	8(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,4(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	60(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	16(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	20(%esp),%esi
-	xorl	%ecx,%edx
-	movl	24(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,16(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	4(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	64(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	12(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	16(%esp),%esi
-	xorl	%ecx,%edx
-	movl	20(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,12(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,28(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	68(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	8(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	12(%esp),%esi
-	xorl	%ecx,%edx
-	movl	16(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,8(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	28(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,24(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	72(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	4(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	8(%esp),%esi
-	xorl	%ecx,%edx
-	movl	12(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,4(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	24(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,20(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	76(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	4(%esp),%esi
-	xorl	%ecx,%edx
-	movl	8(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	20(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,16(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	80(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	28(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	(%esp),%esi
-	xorl	%ecx,%edx
-	movl	4(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,28(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	16(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,12(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	84(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	24(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	28(%esp),%esi
-	xorl	%ecx,%edx
-	movl	(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,24(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	12(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,8(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	88(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	20(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	24(%esp),%esi
-	xorl	%ecx,%edx
-	movl	28(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,20(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	8(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,4(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	92(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	16(%esp),%edx
-	addl	%ecx,%eax
-	movl	96(%esp),%esi
-	xorl	%edi,%ebx
-	movl	12(%esp),%ecx
-	addl	(%esi),%eax
-	addl	4(%esi),%ebx
-	addl	8(%esi),%edi
-	addl	12(%esi),%ecx
-	movl	%eax,(%esi)
-	movl	%ebx,4(%esi)
-	movl	%edi,8(%esi)
-	movl	%ecx,12(%esi)
-	movl	%ebx,4(%esp)
-	xorl	%edi,%ebx
-	movl	%edi,8(%esp)
-	movl	%ecx,12(%esp)
-	movl	20(%esp),%edi
-	movl	24(%esp),%ecx
-	addl	16(%esi),%edx
-	addl	20(%esi),%edi
-	addl	24(%esi),%ecx
-	movl	%edx,16(%esi)
-	movl	%edi,20(%esi)
-	movl	%edi,20(%esp)
-	movl	28(%esp),%edi
-	movl	%ecx,24(%esi)
-	addl	28(%esi),%edi
-	movl	%ecx,24(%esp)
-	movl	%edi,28(%esi)
-	movl	%edi,28(%esp)
-	movl	100(%esp),%edi
-	vmovdqa	64(%ebp),%xmm7
-	subl	$192,%ebp
-	cmpl	104(%esp),%edi
-	jb	.L015grand_avx
-	movl	108(%esp),%esp
-	vzeroall
-	popl	%edi
-	popl	%esi
-	popl	%ebx
-	popl	%ebp
-	ret
-.align	32
-.L014AVX_BMI:
-	leal	-96(%esp),%esp
-	vzeroall
-	movl	(%esi),%eax
-	movl	4(%esi),%ebx
-	movl	8(%esi),%ecx
-	movl	12(%esi),%edi
-	movl	%ebx,4(%esp)
-	xorl	%ecx,%ebx
-	movl	%ecx,8(%esp)
-	movl	%edi,12(%esp)
-	movl	16(%esi),%edx
-	movl	20(%esi),%edi
-	movl	24(%esi),%ecx
-	movl	28(%esi),%esi
-	movl	%edi,20(%esp)
-	movl	100(%esp),%edi
-	movl	%ecx,24(%esp)
-	movl	%esi,28(%esp)
-	vmovdqa	256(%ebp),%xmm7
-	jmp	.L017grand_avx_bmi
-.align	32
-.L017grand_avx_bmi:
-	vmovdqu	(%edi),%xmm0
-	vmovdqu	16(%edi),%xmm1
-	vmovdqu	32(%edi),%xmm2
-	vmovdqu	48(%edi),%xmm3
-	addl	$64,%edi
-	vpshufb	%xmm7,%xmm0,%xmm0
-	movl	%edi,100(%esp)
-	vpshufb	%xmm7,%xmm1,%xmm1
-	vpshufb	%xmm7,%xmm2,%xmm2
-	vpaddd	(%ebp),%xmm0,%xmm4
-	vpshufb	%xmm7,%xmm3,%xmm3
-	vpaddd	16(%ebp),%xmm1,%xmm5
-	vpaddd	32(%ebp),%xmm2,%xmm6
-	vpaddd	48(%ebp),%xmm3,%xmm7
-	vmovdqa	%xmm4,32(%esp)
-	vmovdqa	%xmm5,48(%esp)
-	vmovdqa	%xmm6,64(%esp)
-	vmovdqa	%xmm7,80(%esp)
-	jmp	.L018avx_bmi_00_47
-.align	16
-.L018avx_bmi_00_47:
-	addl	$64,%ebp
-	vpalignr	$4,%xmm0,%xmm1,%xmm4
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,16(%esp)
-	vpalignr	$4,%xmm2,%xmm3,%xmm7
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	24(%esp),%edx,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	xorl	%edi,%ecx
-	andl	20(%esp),%edx
-	movl	%eax,(%esp)
-	vpaddd	%xmm7,%xmm0,%xmm0
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrld	$3,%xmm4,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpslld	$14,%xmm4,%xmm5
-	movl	4(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpxor	%xmm6,%xmm7,%xmm4
-	addl	28(%esp),%edx
-	andl	%eax,%ebx
-	addl	32(%esp),%edx
-	vpshufd	$250,%xmm3,%xmm7
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	12(%esp),%edx
-	vpsrld	$11,%xmm6,%xmm6
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	%edx,12(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpslld	$11,%xmm5,%xmm5
-	andnl	20(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	16(%esp),%edx
-	vpxor	%xmm6,%xmm4,%xmm4
-	movl	%ebx,28(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpsrlq	$17,%xmm7,%xmm5
-	addl	24(%esp),%edx
-	andl	%ebx,%eax
-	addl	36(%esp),%edx
-	vpaddd	%xmm4,%xmm0,%xmm0
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	8(%esp),%edx
-	vpxor	%xmm5,%xmm6,%xmm6
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpsrlq	$19,%xmm7,%xmm7
-	movl	%edx,8(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	andnl	16(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	12(%esp),%edx
-	vpshufd	$132,%xmm6,%xmm7
-	movl	%eax,24(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrldq	$8,%xmm7,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpaddd	%xmm7,%xmm0,%xmm0
-	movl	28(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpshufd	$80,%xmm0,%xmm7
-	addl	20(%esp),%edx
-	andl	%eax,%ebx
-	addl	40(%esp),%edx
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	4(%esp),%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%edx,4(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpsrlq	$19,%xmm7,%xmm7
-	andnl	12(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	8(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	movl	%ebx,20(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpshufd	$232,%xmm6,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpslldq	$8,%xmm7,%xmm7
-	movl	24(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpaddd	%xmm7,%xmm0,%xmm0
-	addl	16(%esp),%edx
-	andl	%ebx,%eax
-	addl	44(%esp),%edx
-	vpaddd	(%ebp),%xmm0,%xmm6
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	vmovdqa	%xmm6,32(%esp)
-	vpalignr	$4,%xmm1,%xmm2,%xmm4
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,(%esp)
-	vpalignr	$4,%xmm3,%xmm0,%xmm7
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	8(%esp),%edx,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	xorl	%edi,%ecx
-	andl	4(%esp),%edx
-	movl	%eax,16(%esp)
-	vpaddd	%xmm7,%xmm1,%xmm1
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrld	$3,%xmm4,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpslld	$14,%xmm4,%xmm5
-	movl	20(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpxor	%xmm6,%xmm7,%xmm4
-	addl	12(%esp),%edx
-	andl	%eax,%ebx
-	addl	48(%esp),%edx
-	vpshufd	$250,%xmm0,%xmm7
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	28(%esp),%edx
-	vpsrld	$11,%xmm6,%xmm6
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	%edx,28(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpslld	$11,%xmm5,%xmm5
-	andnl	4(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	(%esp),%edx
-	vpxor	%xmm6,%xmm4,%xmm4
-	movl	%ebx,12(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	16(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpsrlq	$17,%xmm7,%xmm5
-	addl	8(%esp),%edx
-	andl	%ebx,%eax
-	addl	52(%esp),%edx
-	vpaddd	%xmm4,%xmm1,%xmm1
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	24(%esp),%edx
-	vpxor	%xmm5,%xmm6,%xmm6
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpsrlq	$19,%xmm7,%xmm7
-	movl	%edx,24(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	andnl	(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	28(%esp),%edx
-	vpshufd	$132,%xmm6,%xmm7
-	movl	%eax,8(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrldq	$8,%xmm7,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpaddd	%xmm7,%xmm1,%xmm1
-	movl	12(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpshufd	$80,%xmm1,%xmm7
-	addl	4(%esp),%edx
-	andl	%eax,%ebx
-	addl	56(%esp),%edx
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	20(%esp),%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%edx,20(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpsrlq	$19,%xmm7,%xmm7
-	andnl	28(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	24(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	movl	%ebx,4(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpshufd	$232,%xmm6,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpslldq	$8,%xmm7,%xmm7
-	movl	8(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpaddd	%xmm7,%xmm1,%xmm1
-	addl	(%esp),%edx
-	andl	%ebx,%eax
-	addl	60(%esp),%edx
-	vpaddd	16(%ebp),%xmm1,%xmm6
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	16(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	vmovdqa	%xmm6,48(%esp)
-	vpalignr	$4,%xmm2,%xmm3,%xmm4
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,16(%esp)
-	vpalignr	$4,%xmm0,%xmm1,%xmm7
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	24(%esp),%edx,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	xorl	%edi,%ecx
-	andl	20(%esp),%edx
-	movl	%eax,(%esp)
-	vpaddd	%xmm7,%xmm2,%xmm2
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrld	$3,%xmm4,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpslld	$14,%xmm4,%xmm5
-	movl	4(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpxor	%xmm6,%xmm7,%xmm4
-	addl	28(%esp),%edx
-	andl	%eax,%ebx
-	addl	64(%esp),%edx
-	vpshufd	$250,%xmm1,%xmm7
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	12(%esp),%edx
-	vpsrld	$11,%xmm6,%xmm6
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	%edx,12(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpslld	$11,%xmm5,%xmm5
-	andnl	20(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	16(%esp),%edx
-	vpxor	%xmm6,%xmm4,%xmm4
-	movl	%ebx,28(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpsrlq	$17,%xmm7,%xmm5
-	addl	24(%esp),%edx
-	andl	%ebx,%eax
-	addl	68(%esp),%edx
-	vpaddd	%xmm4,%xmm2,%xmm2
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	8(%esp),%edx
-	vpxor	%xmm5,%xmm6,%xmm6
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpsrlq	$19,%xmm7,%xmm7
-	movl	%edx,8(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	andnl	16(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	12(%esp),%edx
-	vpshufd	$132,%xmm6,%xmm7
-	movl	%eax,24(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrldq	$8,%xmm7,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpaddd	%xmm7,%xmm2,%xmm2
-	movl	28(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpshufd	$80,%xmm2,%xmm7
-	addl	20(%esp),%edx
-	andl	%eax,%ebx
-	addl	72(%esp),%edx
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	4(%esp),%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%edx,4(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpsrlq	$19,%xmm7,%xmm7
-	andnl	12(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	8(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	movl	%ebx,20(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpshufd	$232,%xmm6,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpslldq	$8,%xmm7,%xmm7
-	movl	24(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpaddd	%xmm7,%xmm2,%xmm2
-	addl	16(%esp),%edx
-	andl	%ebx,%eax
-	addl	76(%esp),%edx
-	vpaddd	32(%ebp),%xmm2,%xmm6
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	vmovdqa	%xmm6,64(%esp)
-	vpalignr	$4,%xmm3,%xmm0,%xmm4
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,(%esp)
-	vpalignr	$4,%xmm1,%xmm2,%xmm7
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	8(%esp),%edx,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	xorl	%edi,%ecx
-	andl	4(%esp),%edx
-	movl	%eax,16(%esp)
-	vpaddd	%xmm7,%xmm3,%xmm3
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrld	$3,%xmm4,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpslld	$14,%xmm4,%xmm5
-	movl	20(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpxor	%xmm6,%xmm7,%xmm4
-	addl	12(%esp),%edx
-	andl	%eax,%ebx
-	addl	80(%esp),%edx
-	vpshufd	$250,%xmm2,%xmm7
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	28(%esp),%edx
-	vpsrld	$11,%xmm6,%xmm6
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	%edx,28(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpslld	$11,%xmm5,%xmm5
-	andnl	4(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	(%esp),%edx
-	vpxor	%xmm6,%xmm4,%xmm4
-	movl	%ebx,12(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	16(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpsrlq	$17,%xmm7,%xmm5
-	addl	8(%esp),%edx
-	andl	%ebx,%eax
-	addl	84(%esp),%edx
-	vpaddd	%xmm4,%xmm3,%xmm3
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	24(%esp),%edx
-	vpxor	%xmm5,%xmm6,%xmm6
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpsrlq	$19,%xmm7,%xmm7
-	movl	%edx,24(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	andnl	(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	28(%esp),%edx
-	vpshufd	$132,%xmm6,%xmm7
-	movl	%eax,8(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrldq	$8,%xmm7,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpaddd	%xmm7,%xmm3,%xmm3
-	movl	12(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpshufd	$80,%xmm3,%xmm7
-	addl	4(%esp),%edx
-	andl	%eax,%ebx
-	addl	88(%esp),%edx
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	20(%esp),%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%edx,20(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpsrlq	$19,%xmm7,%xmm7
-	andnl	28(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	24(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	movl	%ebx,4(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpshufd	$232,%xmm6,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpslldq	$8,%xmm7,%xmm7
-	movl	8(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpaddd	%xmm7,%xmm3,%xmm3
-	addl	(%esp),%edx
-	andl	%ebx,%eax
-	addl	92(%esp),%edx
-	vpaddd	48(%ebp),%xmm3,%xmm6
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	16(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	vmovdqa	%xmm6,80(%esp)
-	cmpl	$66051,64(%ebp)
-	jne	.L018avx_bmi_00_47
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,16(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	24(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	20(%esp),%edx
-	movl	%eax,(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	4(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	andl	%eax,%ebx
-	addl	32(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	12(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,12(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	20(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	16(%esp),%edx
-	movl	%ebx,28(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	andl	%ebx,%eax
-	addl	36(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	8(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,8(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	16(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	12(%esp),%edx
-	movl	%eax,24(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	28(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	andl	%eax,%ebx
-	addl	40(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	4(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,4(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	12(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	8(%esp),%edx
-	movl	%ebx,20(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	24(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	andl	%ebx,%eax
-	addl	44(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	8(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	4(%esp),%edx
-	movl	%eax,16(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	20(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	andl	%eax,%ebx
-	addl	48(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	28(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,28(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	4(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	(%esp),%edx
-	movl	%ebx,12(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	16(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	andl	%ebx,%eax
-	addl	52(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	24(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,24(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	28(%esp),%edx
-	movl	%eax,8(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	12(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	andl	%eax,%ebx
-	addl	56(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	20(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,20(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	28(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	24(%esp),%edx
-	movl	%ebx,4(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	8(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	andl	%ebx,%eax
-	addl	60(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	16(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,16(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	24(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	20(%esp),%edx
-	movl	%eax,(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	4(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	andl	%eax,%ebx
-	addl	64(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	12(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,12(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	20(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	16(%esp),%edx
-	movl	%ebx,28(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	andl	%ebx,%eax
-	addl	68(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	8(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,8(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	16(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	12(%esp),%edx
-	movl	%eax,24(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	28(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	andl	%eax,%ebx
-	addl	72(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	4(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,4(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	12(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	8(%esp),%edx
-	movl	%ebx,20(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	24(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	andl	%ebx,%eax
-	addl	76(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	8(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	4(%esp),%edx
-	movl	%eax,16(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	20(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	andl	%eax,%ebx
-	addl	80(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	28(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,28(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	4(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	(%esp),%edx
-	movl	%ebx,12(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	16(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	andl	%ebx,%eax
-	addl	84(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	24(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,24(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	28(%esp),%edx
-	movl	%eax,8(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	12(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	andl	%eax,%ebx
-	addl	88(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	20(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,20(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	28(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	24(%esp),%edx
-	movl	%ebx,4(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	8(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	andl	%ebx,%eax
-	addl	92(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	16(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	movl	96(%esp),%esi
-	xorl	%edi,%ebx
-	movl	12(%esp),%ecx
-	addl	(%esi),%eax
-	addl	4(%esi),%ebx
-	addl	8(%esi),%edi
-	addl	12(%esi),%ecx
-	movl	%eax,(%esi)
-	movl	%ebx,4(%esi)
-	movl	%edi,8(%esi)
-	movl	%ecx,12(%esi)
-	movl	%ebx,4(%esp)
-	xorl	%edi,%ebx
-	movl	%edi,8(%esp)
-	movl	%ecx,12(%esp)
-	movl	20(%esp),%edi
-	movl	24(%esp),%ecx
-	addl	16(%esi),%edx
-	addl	20(%esi),%edi
-	addl	24(%esi),%ecx
-	movl	%edx,16(%esi)
-	movl	%edi,20(%esi)
-	movl	%edi,20(%esp)
-	movl	28(%esp),%edi
-	movl	%ecx,24(%esi)
-	addl	28(%esi),%edi
-	movl	%ecx,24(%esp)
-	movl	%edi,28(%esi)
-	movl	%edi,28(%esp)
-	movl	100(%esp),%edi
-	vmovdqa	64(%ebp),%xmm7
-	subl	$192,%ebp
-	cmpl	104(%esp),%edi
-	jb	.L017grand_avx_bmi
-	movl	108(%esp),%esp
-	vzeroall
-	popl	%edi
-	popl	%esi
-	popl	%ebx
-	popl	%ebp
-	ret
 .size	sha256_block_data_order,.-.L_sha256_block_data_order_begin
 .comm	OPENSSL_ia32cap_P,16,4
 #else
@@ -6825,13 +4619,12 @@ sha256_block_data_order:
 	orl	%ebx,%ecx
 	andl	$1342177280,%ecx
 	cmpl	$1342177280,%ecx
-	je	.L005AVX
 	testl	$512,%ebx
-	jnz	.L006SSSE3
+	jnz	.L005SSSE3
 .L003no_xmm:
 	subl	%edi,%eax
 	cmpl	$256,%eax
-	jae	.L007unrolled
+	jae	.L006unrolled
 	jmp	.L002loop
 .align	16
 .L002loop:
@@ -6903,7 +4696,7 @@ sha256_block_data_order:
 	movl	%ecx,28(%esp)
 	movl	%edi,32(%esp)
 .align	16
-.L00800_15:
+.L00700_15:
 	movl	%edx,%ecx
 	movl	24(%esp),%esi
 	rorl	$14,%ecx
@@ -6941,11 +4734,11 @@ sha256_block_data_order:
 	addl	$4,%ebp
 	addl	%ebx,%eax
 	cmpl	$3248222580,%esi
-	jne	.L00800_15
+	jne	.L00700_15
 	movl	156(%esp),%ecx
-	jmp	.L00916_63
+	jmp	.L00816_63
 .align	16
-.L00916_63:
+.L00816_63:
 	movl	%ecx,%ebx
 	movl	104(%esp),%esi
 	rorl	$11,%ecx
@@ -7000,7 +4793,7 @@ sha256_block_data_order:
 	addl	$4,%ebp
 	addl	%ebx,%eax
 	cmpl	$3329325298,%esi
-	jne	.L00916_63
+	jne	.L00816_63
 	movl	356(%esp),%esi
 	movl	8(%esp),%ebx
 	movl	16(%esp),%ecx
@@ -7044,7 +4837,7 @@ sha256_block_data_order:
 .byte	112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103
 .byte	62,0
 .align	16
-.L007unrolled:
+.L006unrolled:
 	leal	-96(%esp),%esp
 	movl	(%esi),%eax
 	movl	4(%esi),%ebp
@@ -7061,9 +4854,9 @@ sha256_block_data_order:
 	movl	%ebx,20(%esp)
 	movl	%ecx,24(%esp)
 	movl	%esi,28(%esp)
-	jmp	.L010grand_loop
+	jmp	.L009grand_loop
 .align	16
-.L010grand_loop:
+.L009grand_loop:
 	movl	(%edi),%ebx
 	movl	4(%edi),%ecx
 	bswap	%ebx
@@ -9943,7 +7736,7 @@ sha256_block_data_order:
 	movl	%ebx,24(%esp)
 	movl	%ecx,28(%esp)
 	cmpl	104(%esp),%edi
-	jb	.L010grand_loop
+	jb	.L009grand_loop
 	movl	108(%esp),%esp
 	popl	%edi
 	popl	%esi
@@ -9962,9 +7755,9 @@ sha256_block_data_order:
 	pshufd	$27,%xmm2,%xmm2
 .byte	102,15,58,15,202,8
 	punpcklqdq	%xmm0,%xmm2
-	jmp	.L011loop_shaext
+	jmp	.L010loop_shaext
 .align	16
-.L011loop_shaext:
+.L010loop_shaext:
 	movdqu	(%edi),%xmm3
 	movdqu	16(%edi),%xmm4
 	movdqu	32(%edi),%xmm5
@@ -10134,7 +7927,7 @@ sha256_block_data_order:
 .byte	15,56,203,202
 	paddd	16(%esp),%xmm2
 	paddd	(%esp),%xmm1
-	jnz	.L011loop_shaext
+	jnz	.L010loop_shaext
 	pshufd	$177,%xmm2,%xmm2
 	pshufd	$27,%xmm1,%xmm7
 	pshufd	$177,%xmm1,%xmm1
@@ -10149,7 +7942,7 @@ sha256_block_data_order:
 	popl	%ebp
 	ret
 .align	32
-.L006SSSE3:
+.L005SSSE3:
 	leal	-96(%esp),%esp
 	movl	(%esi),%eax
 	movl	4(%esi),%ebx
@@ -10168,9 +7961,9 @@ sha256_block_data_order:
 	movl	%ecx,24(%esp)
 	movl	%esi,28(%esp)
 	movdqa	256(%ebp),%xmm7
-	jmp	.L012grand_ssse3
+	jmp	.L011grand_ssse3
 .align	16
-.L012grand_ssse3:
+.L011grand_ssse3:
 	movdqu	(%edi),%xmm0
 	movdqu	16(%edi),%xmm1
 	movdqu	32(%edi),%xmm2
@@ -10193,9 +7986,9 @@ sha256_block_data_order:
 	paddd	%xmm3,%xmm7
 	movdqa	%xmm6,64(%esp)
 	movdqa	%xmm7,80(%esp)
-	jmp	.L013ssse3_00_47
+	jmp	.L012ssse3_00_47
 .align	16
-.L013ssse3_00_47:
+.L012ssse3_00_47:
 	addl	$64,%ebp
 	movl	%edx,%ecx
 	movdqa	%xmm1,%xmm4
@@ -10838,7 +8631,7 @@ sha256_block_data_order:
 	addl	%ecx,%eax
 	movdqa	%xmm6,80(%esp)
 	cmpl	$66051,64(%ebp)
-	jne	.L013ssse3_00_47
+	jne	.L012ssse3_00_47
 	movl	%edx,%ecx
 	rorl	$14,%edx
 	movl	20(%esp),%esi
@@ -11352,2213 +9145,8 @@ sha256_block_data_order:
 	movdqa	64(%ebp),%xmm7
 	subl	$192,%ebp
 	cmpl	104(%esp),%edi
-	jb	.L012grand_ssse3
-	movl	108(%esp),%esp
-	popl	%edi
-	popl	%esi
-	popl	%ebx
-	popl	%ebp
-	ret
-.align	32
-.L005AVX:
-	andl	$264,%edx
-	cmpl	$264,%edx
-	je	.L014AVX_BMI
-	leal	-96(%esp),%esp
-	vzeroall
-	movl	(%esi),%eax
-	movl	4(%esi),%ebx
-	movl	8(%esi),%ecx
-	movl	12(%esi),%edi
-	movl	%ebx,4(%esp)
-	xorl	%ecx,%ebx
-	movl	%ecx,8(%esp)
-	movl	%edi,12(%esp)
-	movl	16(%esi),%edx
-	movl	20(%esi),%edi
-	movl	24(%esi),%ecx
-	movl	28(%esi),%esi
-	movl	%edi,20(%esp)
-	movl	100(%esp),%edi
-	movl	%ecx,24(%esp)
-	movl	%esi,28(%esp)
-	vmovdqa	256(%ebp),%xmm7
-	jmp	.L015grand_avx
-.align	32
-.L015grand_avx:
-	vmovdqu	(%edi),%xmm0
-	vmovdqu	16(%edi),%xmm1
-	vmovdqu	32(%edi),%xmm2
-	vmovdqu	48(%edi),%xmm3
-	addl	$64,%edi
-	vpshufb	%xmm7,%xmm0,%xmm0
-	movl	%edi,100(%esp)
-	vpshufb	%xmm7,%xmm1,%xmm1
-	vpshufb	%xmm7,%xmm2,%xmm2
-	vpaddd	(%ebp),%xmm0,%xmm4
-	vpshufb	%xmm7,%xmm3,%xmm3
-	vpaddd	16(%ebp),%xmm1,%xmm5
-	vpaddd	32(%ebp),%xmm2,%xmm6
-	vpaddd	48(%ebp),%xmm3,%xmm7
-	vmovdqa	%xmm4,32(%esp)
-	vmovdqa	%xmm5,48(%esp)
-	vmovdqa	%xmm6,64(%esp)
-	vmovdqa	%xmm7,80(%esp)
-	jmp	.L016avx_00_47
-.align	16
-.L016avx_00_47:
-	addl	$64,%ebp
-	vpalignr	$4,%xmm0,%xmm1,%xmm4
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	20(%esp),%esi
-	vpalignr	$4,%xmm2,%xmm3,%xmm7
-	xorl	%ecx,%edx
-	movl	24(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,16(%esp)
-	vpaddd	%xmm7,%xmm0,%xmm0
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrld	$3,%xmm4,%xmm7
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	4(%esp),%edi
-	vpslld	$14,%xmm4,%xmm5
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,(%esp)
-	vpxor	%xmm6,%xmm7,%xmm4
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	vpshufd	$250,%xmm3,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpsrld	$11,%xmm6,%xmm6
-	addl	32(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpxor	%xmm5,%xmm4,%xmm4
-	addl	%edx,%ebx
-	addl	12(%esp),%edx
-	addl	%ecx,%ebx
-	vpslld	$11,%xmm5,%xmm5
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	16(%esp),%esi
-	vpxor	%xmm6,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	movl	20(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,12(%esp)
-	vpxor	%xmm5,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	(%esp),%edi
-	vpaddd	%xmm4,%xmm0,%xmm0
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,28(%esp)
-	vpxor	%xmm5,%xmm6,%xmm6
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	vpsrlq	$19,%xmm7,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	addl	36(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	vpshufd	$132,%xmm6,%xmm7
-	addl	%edx,%eax
-	addl	8(%esp),%edx
-	addl	%ecx,%eax
-	vpsrldq	$8,%xmm7,%xmm7
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	12(%esp),%esi
-	vpaddd	%xmm7,%xmm0,%xmm0
-	xorl	%ecx,%edx
-	movl	16(%esp),%edi
-	xorl	%edi,%esi
-	vpshufd	$80,%xmm0,%xmm7
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,8(%esp)
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	28(%esp),%edi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,24(%esp)
-	vpsrlq	$19,%xmm7,%xmm7
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpshufd	$232,%xmm6,%xmm7
-	addl	40(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpslldq	$8,%xmm7,%xmm7
-	addl	%edx,%ebx
-	addl	4(%esp),%edx
-	addl	%ecx,%ebx
-	vpaddd	%xmm7,%xmm0,%xmm0
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	8(%esp),%esi
-	vpaddd	(%ebp),%xmm0,%xmm6
-	xorl	%ecx,%edx
-	movl	12(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,4(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	24(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,20(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	44(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	(%esp),%edx
-	addl	%ecx,%eax
-	vmovdqa	%xmm6,32(%esp)
-	vpalignr	$4,%xmm1,%xmm2,%xmm4
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	4(%esp),%esi
-	vpalignr	$4,%xmm3,%xmm0,%xmm7
-	xorl	%ecx,%edx
-	movl	8(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,(%esp)
-	vpaddd	%xmm7,%xmm1,%xmm1
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrld	$3,%xmm4,%xmm7
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	20(%esp),%edi
-	vpslld	$14,%xmm4,%xmm5
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,16(%esp)
-	vpxor	%xmm6,%xmm7,%xmm4
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	vpshufd	$250,%xmm0,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpsrld	$11,%xmm6,%xmm6
-	addl	48(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpxor	%xmm5,%xmm4,%xmm4
-	addl	%edx,%ebx
-	addl	28(%esp),%edx
-	addl	%ecx,%ebx
-	vpslld	$11,%xmm5,%xmm5
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	(%esp),%esi
-	vpxor	%xmm6,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	movl	4(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,28(%esp)
-	vpxor	%xmm5,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	16(%esp),%edi
-	vpaddd	%xmm4,%xmm1,%xmm1
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,12(%esp)
-	vpxor	%xmm5,%xmm6,%xmm6
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	vpsrlq	$19,%xmm7,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	addl	52(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	vpshufd	$132,%xmm6,%xmm7
-	addl	%edx,%eax
-	addl	24(%esp),%edx
-	addl	%ecx,%eax
-	vpsrldq	$8,%xmm7,%xmm7
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	28(%esp),%esi
-	vpaddd	%xmm7,%xmm1,%xmm1
-	xorl	%ecx,%edx
-	movl	(%esp),%edi
-	xorl	%edi,%esi
-	vpshufd	$80,%xmm1,%xmm7
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,24(%esp)
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	12(%esp),%edi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,8(%esp)
-	vpsrlq	$19,%xmm7,%xmm7
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpshufd	$232,%xmm6,%xmm7
-	addl	56(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpslldq	$8,%xmm7,%xmm7
-	addl	%edx,%ebx
-	addl	20(%esp),%edx
-	addl	%ecx,%ebx
-	vpaddd	%xmm7,%xmm1,%xmm1
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	24(%esp),%esi
-	vpaddd	16(%ebp),%xmm1,%xmm6
-	xorl	%ecx,%edx
-	movl	28(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,20(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	8(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,4(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	60(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	16(%esp),%edx
-	addl	%ecx,%eax
-	vmovdqa	%xmm6,48(%esp)
-	vpalignr	$4,%xmm2,%xmm3,%xmm4
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	20(%esp),%esi
-	vpalignr	$4,%xmm0,%xmm1,%xmm7
-	xorl	%ecx,%edx
-	movl	24(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,16(%esp)
-	vpaddd	%xmm7,%xmm2,%xmm2
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrld	$3,%xmm4,%xmm7
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	4(%esp),%edi
-	vpslld	$14,%xmm4,%xmm5
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,(%esp)
-	vpxor	%xmm6,%xmm7,%xmm4
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	vpshufd	$250,%xmm1,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpsrld	$11,%xmm6,%xmm6
-	addl	64(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpxor	%xmm5,%xmm4,%xmm4
-	addl	%edx,%ebx
-	addl	12(%esp),%edx
-	addl	%ecx,%ebx
-	vpslld	$11,%xmm5,%xmm5
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	16(%esp),%esi
-	vpxor	%xmm6,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	movl	20(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,12(%esp)
-	vpxor	%xmm5,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	(%esp),%edi
-	vpaddd	%xmm4,%xmm2,%xmm2
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,28(%esp)
-	vpxor	%xmm5,%xmm6,%xmm6
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	vpsrlq	$19,%xmm7,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	addl	68(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	vpshufd	$132,%xmm6,%xmm7
-	addl	%edx,%eax
-	addl	8(%esp),%edx
-	addl	%ecx,%eax
-	vpsrldq	$8,%xmm7,%xmm7
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	12(%esp),%esi
-	vpaddd	%xmm7,%xmm2,%xmm2
-	xorl	%ecx,%edx
-	movl	16(%esp),%edi
-	xorl	%edi,%esi
-	vpshufd	$80,%xmm2,%xmm7
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,8(%esp)
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	28(%esp),%edi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,24(%esp)
-	vpsrlq	$19,%xmm7,%xmm7
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpshufd	$232,%xmm6,%xmm7
-	addl	72(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpslldq	$8,%xmm7,%xmm7
-	addl	%edx,%ebx
-	addl	4(%esp),%edx
-	addl	%ecx,%ebx
-	vpaddd	%xmm7,%xmm2,%xmm2
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	8(%esp),%esi
-	vpaddd	32(%ebp),%xmm2,%xmm6
-	xorl	%ecx,%edx
-	movl	12(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,4(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	24(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,20(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	76(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	(%esp),%edx
-	addl	%ecx,%eax
-	vmovdqa	%xmm6,64(%esp)
-	vpalignr	$4,%xmm3,%xmm0,%xmm4
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	4(%esp),%esi
-	vpalignr	$4,%xmm1,%xmm2,%xmm7
-	xorl	%ecx,%edx
-	movl	8(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,(%esp)
-	vpaddd	%xmm7,%xmm3,%xmm3
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrld	$3,%xmm4,%xmm7
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	20(%esp),%edi
-	vpslld	$14,%xmm4,%xmm5
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,16(%esp)
-	vpxor	%xmm6,%xmm7,%xmm4
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	vpshufd	$250,%xmm2,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpsrld	$11,%xmm6,%xmm6
-	addl	80(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpxor	%xmm5,%xmm4,%xmm4
-	addl	%edx,%ebx
-	addl	28(%esp),%edx
-	addl	%ecx,%ebx
-	vpslld	$11,%xmm5,%xmm5
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	(%esp),%esi
-	vpxor	%xmm6,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	movl	4(%esp),%edi
-	xorl	%edi,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,28(%esp)
-	vpxor	%xmm5,%xmm4,%xmm4
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	16(%esp),%edi
-	vpaddd	%xmm4,%xmm3,%xmm3
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,12(%esp)
-	vpxor	%xmm5,%xmm6,%xmm6
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	vpsrlq	$19,%xmm7,%xmm7
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	addl	84(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	vpshufd	$132,%xmm6,%xmm7
-	addl	%edx,%eax
-	addl	24(%esp),%edx
-	addl	%ecx,%eax
-	vpsrldq	$8,%xmm7,%xmm7
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	28(%esp),%esi
-	vpaddd	%xmm7,%xmm3,%xmm3
-	xorl	%ecx,%edx
-	movl	(%esp),%edi
-	xorl	%edi,%esi
-	vpshufd	$80,%xmm3,%xmm7
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,24(%esp)
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	12(%esp),%edi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,8(%esp)
-	vpsrlq	$19,%xmm7,%xmm7
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	vpshufd	$232,%xmm6,%xmm7
-	addl	88(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	vpslldq	$8,%xmm7,%xmm7
-	addl	%edx,%ebx
-	addl	20(%esp),%edx
-	addl	%ecx,%ebx
-	vpaddd	%xmm7,%xmm3,%xmm3
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	24(%esp),%esi
-	vpaddd	48(%ebp),%xmm3,%xmm6
-	xorl	%ecx,%edx
-	movl	28(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,20(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	8(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,4(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	92(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	16(%esp),%edx
-	addl	%ecx,%eax
-	vmovdqa	%xmm6,80(%esp)
-	cmpl	$66051,64(%ebp)
-	jne	.L016avx_00_47
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	20(%esp),%esi
-	xorl	%ecx,%edx
-	movl	24(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,16(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	4(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	32(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	12(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	16(%esp),%esi
-	xorl	%ecx,%edx
-	movl	20(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,12(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,28(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	36(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	8(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	12(%esp),%esi
-	xorl	%ecx,%edx
-	movl	16(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,8(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	28(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,24(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	40(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	4(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	8(%esp),%esi
-	xorl	%ecx,%edx
-	movl	12(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,4(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	24(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,20(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	44(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	4(%esp),%esi
-	xorl	%ecx,%edx
-	movl	8(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	20(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,16(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	48(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	28(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	(%esp),%esi
-	xorl	%ecx,%edx
-	movl	4(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,28(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	16(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,12(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	52(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	24(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	28(%esp),%esi
-	xorl	%ecx,%edx
-	movl	(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,24(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	12(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,8(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	56(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	20(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	24(%esp),%esi
-	xorl	%ecx,%edx
-	movl	28(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,20(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	8(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,4(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	60(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	16(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	20(%esp),%esi
-	xorl	%ecx,%edx
-	movl	24(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,16(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	4(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	64(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	12(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	16(%esp),%esi
-	xorl	%ecx,%edx
-	movl	20(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,12(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,28(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	68(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	8(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	12(%esp),%esi
-	xorl	%ecx,%edx
-	movl	16(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,8(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	28(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,24(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	72(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	4(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	8(%esp),%esi
-	xorl	%ecx,%edx
-	movl	12(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,4(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	24(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,20(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	76(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	4(%esp),%esi
-	xorl	%ecx,%edx
-	movl	8(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	20(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,16(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	80(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	28(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	(%esp),%esi
-	xorl	%ecx,%edx
-	movl	4(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,28(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	16(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,12(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	84(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	24(%esp),%edx
-	addl	%ecx,%eax
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	28(%esp),%esi
-	xorl	%ecx,%edx
-	movl	(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,24(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%eax,%ecx
-	addl	%edi,%edx
-	movl	12(%esp),%edi
-	movl	%eax,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%eax,8(%esp)
-	xorl	%eax,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%eax,%ebx
-	xorl	%esi,%ecx
-	addl	88(%esp),%edx
-	xorl	%edi,%ebx
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%ebx
-	addl	20(%esp),%edx
-	addl	%ecx,%ebx
-	movl	%edx,%ecx
-	shrdl	$14,%edx,%edx
-	movl	24(%esp),%esi
-	xorl	%ecx,%edx
-	movl	28(%esp),%edi
-	xorl	%edi,%esi
-	shrdl	$5,%edx,%edx
-	andl	%ecx,%esi
-	movl	%ecx,20(%esp)
-	xorl	%ecx,%edx
-	xorl	%esi,%edi
-	shrdl	$6,%edx,%edx
-	movl	%ebx,%ecx
-	addl	%edi,%edx
-	movl	8(%esp),%edi
-	movl	%ebx,%esi
-	shrdl	$9,%ecx,%ecx
-	movl	%ebx,4(%esp)
-	xorl	%ebx,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	shrdl	$11,%ecx,%ecx
-	andl	%ebx,%eax
-	xorl	%esi,%ecx
-	addl	92(%esp),%edx
-	xorl	%edi,%eax
-	shrdl	$2,%ecx,%ecx
-	addl	%edx,%eax
-	addl	16(%esp),%edx
-	addl	%ecx,%eax
-	movl	96(%esp),%esi
-	xorl	%edi,%ebx
-	movl	12(%esp),%ecx
-	addl	(%esi),%eax
-	addl	4(%esi),%ebx
-	addl	8(%esi),%edi
-	addl	12(%esi),%ecx
-	movl	%eax,(%esi)
-	movl	%ebx,4(%esi)
-	movl	%edi,8(%esi)
-	movl	%ecx,12(%esi)
-	movl	%ebx,4(%esp)
-	xorl	%edi,%ebx
-	movl	%edi,8(%esp)
-	movl	%ecx,12(%esp)
-	movl	20(%esp),%edi
-	movl	24(%esp),%ecx
-	addl	16(%esi),%edx
-	addl	20(%esi),%edi
-	addl	24(%esi),%ecx
-	movl	%edx,16(%esi)
-	movl	%edi,20(%esi)
-	movl	%edi,20(%esp)
-	movl	28(%esp),%edi
-	movl	%ecx,24(%esi)
-	addl	28(%esi),%edi
-	movl	%ecx,24(%esp)
-	movl	%edi,28(%esi)
-	movl	%edi,28(%esp)
-	movl	100(%esp),%edi
-	vmovdqa	64(%ebp),%xmm7
-	subl	$192,%ebp
-	cmpl	104(%esp),%edi
-	jb	.L015grand_avx
-	movl	108(%esp),%esp
-	vzeroall
-	popl	%edi
-	popl	%esi
-	popl	%ebx
-	popl	%ebp
-	ret
-.align	32
-.L014AVX_BMI:
-	leal	-96(%esp),%esp
-	vzeroall
-	movl	(%esi),%eax
-	movl	4(%esi),%ebx
-	movl	8(%esi),%ecx
-	movl	12(%esi),%edi
-	movl	%ebx,4(%esp)
-	xorl	%ecx,%ebx
-	movl	%ecx,8(%esp)
-	movl	%edi,12(%esp)
-	movl	16(%esi),%edx
-	movl	20(%esi),%edi
-	movl	24(%esi),%ecx
-	movl	28(%esi),%esi
-	movl	%edi,20(%esp)
-	movl	100(%esp),%edi
-	movl	%ecx,24(%esp)
-	movl	%esi,28(%esp)
-	vmovdqa	256(%ebp),%xmm7
-	jmp	.L017grand_avx_bmi
-.align	32
-.L017grand_avx_bmi:
-	vmovdqu	(%edi),%xmm0
-	vmovdqu	16(%edi),%xmm1
-	vmovdqu	32(%edi),%xmm2
-	vmovdqu	48(%edi),%xmm3
-	addl	$64,%edi
-	vpshufb	%xmm7,%xmm0,%xmm0
-	movl	%edi,100(%esp)
-	vpshufb	%xmm7,%xmm1,%xmm1
-	vpshufb	%xmm7,%xmm2,%xmm2
-	vpaddd	(%ebp),%xmm0,%xmm4
-	vpshufb	%xmm7,%xmm3,%xmm3
-	vpaddd	16(%ebp),%xmm1,%xmm5
-	vpaddd	32(%ebp),%xmm2,%xmm6
-	vpaddd	48(%ebp),%xmm3,%xmm7
-	vmovdqa	%xmm4,32(%esp)
-	vmovdqa	%xmm5,48(%esp)
-	vmovdqa	%xmm6,64(%esp)
-	vmovdqa	%xmm7,80(%esp)
-	jmp	.L018avx_bmi_00_47
-.align	16
-.L018avx_bmi_00_47:
-	addl	$64,%ebp
-	vpalignr	$4,%xmm0,%xmm1,%xmm4
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,16(%esp)
-	vpalignr	$4,%xmm2,%xmm3,%xmm7
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	24(%esp),%edx,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	xorl	%edi,%ecx
-	andl	20(%esp),%edx
-	movl	%eax,(%esp)
-	vpaddd	%xmm7,%xmm0,%xmm0
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrld	$3,%xmm4,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpslld	$14,%xmm4,%xmm5
-	movl	4(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpxor	%xmm6,%xmm7,%xmm4
-	addl	28(%esp),%edx
-	andl	%eax,%ebx
-	addl	32(%esp),%edx
-	vpshufd	$250,%xmm3,%xmm7
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	12(%esp),%edx
-	vpsrld	$11,%xmm6,%xmm6
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	%edx,12(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpslld	$11,%xmm5,%xmm5
-	andnl	20(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	16(%esp),%edx
-	vpxor	%xmm6,%xmm4,%xmm4
-	movl	%ebx,28(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpsrlq	$17,%xmm7,%xmm5
-	addl	24(%esp),%edx
-	andl	%ebx,%eax
-	addl	36(%esp),%edx
-	vpaddd	%xmm4,%xmm0,%xmm0
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	8(%esp),%edx
-	vpxor	%xmm5,%xmm6,%xmm6
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpsrlq	$19,%xmm7,%xmm7
-	movl	%edx,8(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	andnl	16(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	12(%esp),%edx
-	vpshufd	$132,%xmm6,%xmm7
-	movl	%eax,24(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrldq	$8,%xmm7,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpaddd	%xmm7,%xmm0,%xmm0
-	movl	28(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpshufd	$80,%xmm0,%xmm7
-	addl	20(%esp),%edx
-	andl	%eax,%ebx
-	addl	40(%esp),%edx
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	4(%esp),%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%edx,4(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpsrlq	$19,%xmm7,%xmm7
-	andnl	12(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	8(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	movl	%ebx,20(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpshufd	$232,%xmm6,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpslldq	$8,%xmm7,%xmm7
-	movl	24(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpaddd	%xmm7,%xmm0,%xmm0
-	addl	16(%esp),%edx
-	andl	%ebx,%eax
-	addl	44(%esp),%edx
-	vpaddd	(%ebp),%xmm0,%xmm6
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	vmovdqa	%xmm6,32(%esp)
-	vpalignr	$4,%xmm1,%xmm2,%xmm4
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,(%esp)
-	vpalignr	$4,%xmm3,%xmm0,%xmm7
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	8(%esp),%edx,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	xorl	%edi,%ecx
-	andl	4(%esp),%edx
-	movl	%eax,16(%esp)
-	vpaddd	%xmm7,%xmm1,%xmm1
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrld	$3,%xmm4,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpslld	$14,%xmm4,%xmm5
-	movl	20(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpxor	%xmm6,%xmm7,%xmm4
-	addl	12(%esp),%edx
-	andl	%eax,%ebx
-	addl	48(%esp),%edx
-	vpshufd	$250,%xmm0,%xmm7
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	28(%esp),%edx
-	vpsrld	$11,%xmm6,%xmm6
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	%edx,28(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpslld	$11,%xmm5,%xmm5
-	andnl	4(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	(%esp),%edx
-	vpxor	%xmm6,%xmm4,%xmm4
-	movl	%ebx,12(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	16(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpsrlq	$17,%xmm7,%xmm5
-	addl	8(%esp),%edx
-	andl	%ebx,%eax
-	addl	52(%esp),%edx
-	vpaddd	%xmm4,%xmm1,%xmm1
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	24(%esp),%edx
-	vpxor	%xmm5,%xmm6,%xmm6
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpsrlq	$19,%xmm7,%xmm7
-	movl	%edx,24(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	andnl	(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	28(%esp),%edx
-	vpshufd	$132,%xmm6,%xmm7
-	movl	%eax,8(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrldq	$8,%xmm7,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpaddd	%xmm7,%xmm1,%xmm1
-	movl	12(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpshufd	$80,%xmm1,%xmm7
-	addl	4(%esp),%edx
-	andl	%eax,%ebx
-	addl	56(%esp),%edx
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	20(%esp),%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%edx,20(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpsrlq	$19,%xmm7,%xmm7
-	andnl	28(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	24(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	movl	%ebx,4(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpshufd	$232,%xmm6,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpslldq	$8,%xmm7,%xmm7
-	movl	8(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpaddd	%xmm7,%xmm1,%xmm1
-	addl	(%esp),%edx
-	andl	%ebx,%eax
-	addl	60(%esp),%edx
-	vpaddd	16(%ebp),%xmm1,%xmm6
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	16(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	vmovdqa	%xmm6,48(%esp)
-	vpalignr	$4,%xmm2,%xmm3,%xmm4
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,16(%esp)
-	vpalignr	$4,%xmm0,%xmm1,%xmm7
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	24(%esp),%edx,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	xorl	%edi,%ecx
-	andl	20(%esp),%edx
-	movl	%eax,(%esp)
-	vpaddd	%xmm7,%xmm2,%xmm2
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrld	$3,%xmm4,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpslld	$14,%xmm4,%xmm5
-	movl	4(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpxor	%xmm6,%xmm7,%xmm4
-	addl	28(%esp),%edx
-	andl	%eax,%ebx
-	addl	64(%esp),%edx
-	vpshufd	$250,%xmm1,%xmm7
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	12(%esp),%edx
-	vpsrld	$11,%xmm6,%xmm6
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	%edx,12(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpslld	$11,%xmm5,%xmm5
-	andnl	20(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	16(%esp),%edx
-	vpxor	%xmm6,%xmm4,%xmm4
-	movl	%ebx,28(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpsrlq	$17,%xmm7,%xmm5
-	addl	24(%esp),%edx
-	andl	%ebx,%eax
-	addl	68(%esp),%edx
-	vpaddd	%xmm4,%xmm2,%xmm2
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	8(%esp),%edx
-	vpxor	%xmm5,%xmm6,%xmm6
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpsrlq	$19,%xmm7,%xmm7
-	movl	%edx,8(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	andnl	16(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	12(%esp),%edx
-	vpshufd	$132,%xmm6,%xmm7
-	movl	%eax,24(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrldq	$8,%xmm7,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpaddd	%xmm7,%xmm2,%xmm2
-	movl	28(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpshufd	$80,%xmm2,%xmm7
-	addl	20(%esp),%edx
-	andl	%eax,%ebx
-	addl	72(%esp),%edx
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	4(%esp),%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%edx,4(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpsrlq	$19,%xmm7,%xmm7
-	andnl	12(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	8(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	movl	%ebx,20(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpshufd	$232,%xmm6,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpslldq	$8,%xmm7,%xmm7
-	movl	24(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpaddd	%xmm7,%xmm2,%xmm2
-	addl	16(%esp),%edx
-	andl	%ebx,%eax
-	addl	76(%esp),%edx
-	vpaddd	32(%ebp),%xmm2,%xmm6
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	vmovdqa	%xmm6,64(%esp)
-	vpalignr	$4,%xmm3,%xmm0,%xmm4
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,(%esp)
-	vpalignr	$4,%xmm1,%xmm2,%xmm7
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	8(%esp),%edx,%esi
-	vpsrld	$7,%xmm4,%xmm6
-	xorl	%edi,%ecx
-	andl	4(%esp),%edx
-	movl	%eax,16(%esp)
-	vpaddd	%xmm7,%xmm3,%xmm3
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrld	$3,%xmm4,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpslld	$14,%xmm4,%xmm5
-	movl	20(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpxor	%xmm6,%xmm7,%xmm4
-	addl	12(%esp),%edx
-	andl	%eax,%ebx
-	addl	80(%esp),%edx
-	vpshufd	$250,%xmm2,%xmm7
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	28(%esp),%edx
-	vpsrld	$11,%xmm6,%xmm6
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	%edx,28(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpslld	$11,%xmm5,%xmm5
-	andnl	4(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	(%esp),%edx
-	vpxor	%xmm6,%xmm4,%xmm4
-	movl	%ebx,12(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpsrld	$10,%xmm7,%xmm6
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpxor	%xmm5,%xmm4,%xmm4
-	movl	16(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpsrlq	$17,%xmm7,%xmm5
-	addl	8(%esp),%edx
-	andl	%ebx,%eax
-	addl	84(%esp),%edx
-	vpaddd	%xmm4,%xmm3,%xmm3
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	24(%esp),%edx
-	vpxor	%xmm5,%xmm6,%xmm6
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpsrlq	$19,%xmm7,%xmm7
-	movl	%edx,24(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpxor	%xmm7,%xmm6,%xmm6
-	andnl	(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	28(%esp),%edx
-	vpshufd	$132,%xmm6,%xmm7
-	movl	%eax,8(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	vpsrldq	$8,%xmm7,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	vpaddd	%xmm7,%xmm3,%xmm3
-	movl	12(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	vpshufd	$80,%xmm3,%xmm7
-	addl	4(%esp),%edx
-	andl	%eax,%ebx
-	addl	88(%esp),%edx
-	vpsrld	$10,%xmm7,%xmm6
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	20(%esp),%edx
-	vpsrlq	$17,%xmm7,%xmm5
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	vpxor	%xmm5,%xmm6,%xmm6
-	movl	%edx,20(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	vpsrlq	$19,%xmm7,%xmm7
-	andnl	28(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	24(%esp),%edx
-	vpxor	%xmm7,%xmm6,%xmm6
-	movl	%ebx,4(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	vpshufd	$232,%xmm6,%xmm7
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	vpslldq	$8,%xmm7,%xmm7
-	movl	8(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	vpaddd	%xmm7,%xmm3,%xmm3
-	addl	(%esp),%edx
-	andl	%ebx,%eax
-	addl	92(%esp),%edx
-	vpaddd	48(%ebp),%xmm3,%xmm6
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	16(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	vmovdqa	%xmm6,80(%esp)
-	cmpl	$66051,64(%ebp)
-	jne	.L018avx_bmi_00_47
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,16(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	24(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	20(%esp),%edx
-	movl	%eax,(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	4(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	andl	%eax,%ebx
-	addl	32(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	12(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,12(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	20(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	16(%esp),%edx
-	movl	%ebx,28(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	andl	%ebx,%eax
-	addl	36(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	8(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,8(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	16(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	12(%esp),%edx
-	movl	%eax,24(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	28(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	andl	%eax,%ebx
-	addl	40(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	4(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,4(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	12(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	8(%esp),%edx
-	movl	%ebx,20(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	24(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	andl	%ebx,%eax
-	addl	44(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	8(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	4(%esp),%edx
-	movl	%eax,16(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	20(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	andl	%eax,%ebx
-	addl	48(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	28(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,28(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	4(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	(%esp),%edx
-	movl	%ebx,12(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	16(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	andl	%ebx,%eax
-	addl	52(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	24(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,24(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	28(%esp),%edx
-	movl	%eax,8(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	12(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	andl	%eax,%ebx
-	addl	56(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	20(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,20(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	28(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	24(%esp),%edx
-	movl	%ebx,4(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	8(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	andl	%ebx,%eax
-	addl	60(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	16(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,16(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	24(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	20(%esp),%edx
-	movl	%eax,(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	4(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	28(%esp),%edx
-	andl	%eax,%ebx
-	addl	64(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	12(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,12(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	20(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	16(%esp),%edx
-	movl	%ebx,28(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	24(%esp),%edx
-	andl	%ebx,%eax
-	addl	68(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	8(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,8(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	16(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	12(%esp),%edx
-	movl	%eax,24(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	28(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	20(%esp),%edx
-	andl	%eax,%ebx
-	addl	72(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	4(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,4(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	12(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	8(%esp),%edx
-	movl	%ebx,20(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	24(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	16(%esp),%edx
-	andl	%ebx,%eax
-	addl	76(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	8(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	4(%esp),%edx
-	movl	%eax,16(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	20(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	12(%esp),%edx
-	andl	%eax,%ebx
-	addl	80(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	28(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,28(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	4(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	(%esp),%edx
-	movl	%ebx,12(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	16(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	8(%esp),%edx
-	andl	%ebx,%eax
-	addl	84(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	24(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,24(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	28(%esp),%edx
-	movl	%eax,8(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%eax,%edi
-	rorxl	$13,%eax,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%eax,%ecx
-	xorl	%edi,%esi
-	movl	12(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%eax
-	addl	4(%esp),%edx
-	andl	%eax,%ebx
-	addl	88(%esp),%edx
-	xorl	%edi,%ebx
-	addl	%edx,%ecx
-	addl	20(%esp),%edx
-	leal	(%ebx,%ecx,1),%ebx
-	rorxl	$6,%edx,%ecx
-	rorxl	$11,%edx,%esi
-	movl	%edx,20(%esp)
-	rorxl	$25,%edx,%edi
-	xorl	%esi,%ecx
-	andnl	28(%esp),%edx,%esi
-	xorl	%edi,%ecx
-	andl	24(%esp),%edx
-	movl	%ebx,4(%esp)
-	orl	%esi,%edx
-	rorxl	$2,%ebx,%edi
-	rorxl	$13,%ebx,%esi
-	leal	(%edx,%ecx,1),%edx
-	rorxl	$22,%ebx,%ecx
-	xorl	%edi,%esi
-	movl	8(%esp),%edi
-	xorl	%esi,%ecx
-	xorl	%edi,%ebx
-	addl	(%esp),%edx
-	andl	%ebx,%eax
-	addl	92(%esp),%edx
-	xorl	%edi,%eax
-	addl	%edx,%ecx
-	addl	16(%esp),%edx
-	leal	(%eax,%ecx,1),%eax
-	movl	96(%esp),%esi
-	xorl	%edi,%ebx
-	movl	12(%esp),%ecx
-	addl	(%esi),%eax
-	addl	4(%esi),%ebx
-	addl	8(%esi),%edi
-	addl	12(%esi),%ecx
-	movl	%eax,(%esi)
-	movl	%ebx,4(%esi)
-	movl	%edi,8(%esi)
-	movl	%ecx,12(%esi)
-	movl	%ebx,4(%esp)
-	xorl	%edi,%ebx
-	movl	%edi,8(%esp)
-	movl	%ecx,12(%esp)
-	movl	20(%esp),%edi
-	movl	24(%esp),%ecx
-	addl	16(%esi),%edx
-	addl	20(%esi),%edi
-	addl	24(%esi),%ecx
-	movl	%edx,16(%esi)
-	movl	%edi,20(%esi)
-	movl	%edi,20(%esp)
-	movl	28(%esp),%edi
-	movl	%ecx,24(%esi)
-	addl	28(%esi),%edi
-	movl	%ecx,24(%esp)
-	movl	%edi,28(%esi)
-	movl	%edi,28(%esp)
-	movl	100(%esp),%edi
-	vmovdqa	64(%ebp),%xmm7
-	subl	$192,%ebp
-	cmpl	104(%esp),%edi
-	jb	.L017grand_avx_bmi
+	jb	.L011grand_ssse3
 	movl	108(%esp),%esp
-	vzeroall
 	popl	%edi
 	popl	%esi
 	popl	%ebx