external/boringssl: Sync to a0b49d63fdc33e54eac93674c86891d15d181d87.

This includes the following changes:

https://boringssl.googlesource.com/boringssl/+log/2fb729d4f36beaf263ad85e24a790b571652679c..a0b49d63fdc33e54eac93674c86891d15d181d87

Test: atest CtsLibcoreTestCases
Change-Id: Ida4794d56d237422351b9ddcc7d0bc9295e816e1

3 files changed, 1868 insertions, 3 deletions

diff --git a/linux-x86_64/crypto/cipher_extra/chacha20_poly1305_x86_64.S b/linux-x86_64/crypto/cipher_extra/chacha20_poly1305_x86_64.S
index e3133488..aefa5432 100644
--- a/linux-x86_64/crypto/cipher_extra/chacha20_poly1305_x86_64.S
+++ b/linux-x86_64/crypto/cipher_extra/chacha20_poly1305_x86_64.S
@@ -3935,7 +3935,7 @@ do_length_block:
 	popq	%rbp
 .cfi_adjust_cfa_offset	-8
 	.byte	0xf3,0xc3
-.cfi_adjust_cfa_offset	(8 * 6) + 288 + 32
+.cfi_adjust_cfa_offset	(8 * 7) + 288 + 32
 
 seal_sse_128:
 	movdqu	.chacha20_consts(%rip),%xmm0
diff --git a/linux-x86_64/crypto/fipsmodule/ghash-x86_64.S b/linux-x86_64/crypto/fipsmodule/ghash-x86_64.S
index 8fff7eb0..3eb1af43 100644
--- a/linux-x86_64/crypto/fipsmodule/ghash-x86_64.S
+++ b/linux-x86_64/crypto/fipsmodule/ghash-x86_64.S
@@ -1119,8 +1119,6 @@ gcm_ghash_avx:
 .byte	1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
 .L7_mask:
 .long	7,0,7,0
-.L7_mask_poly:
-.long	7,0,450,0
 .align	64
 
 .byte	71,72,65,83,72,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
diff --git a/linux-x86_64/crypto/fipsmodule/sha1-x86_64.S b/linux-x86_64/crypto/fipsmodule/sha1-x86_64.S
index a4ce81ff..964687dc 100644
--- a/linux-x86_64/crypto/fipsmodule/sha1-x86_64.S
+++ b/linux-x86_64/crypto/fipsmodule/sha1-x86_64.S
@@ -27,6 +27,11 @@ sha1_block_data_order:
 	movl	8(%r10),%r10d
 	testl	$512,%r8d
 	jz	.Lialu
+	testl	$536870912,%r10d
+	jnz	_shaext_shortcut
+	andl	$296,%r10d
+	cmpl	$296,%r10d
+	je	_avx2_shortcut
 	andl	$268435456,%r8d
 	andl	$1073741824,%r9d
 	orl	%r9d,%r8d
@@ -1266,6 +1271,175 @@ sha1_block_data_order:
 	.byte	0xf3,0xc3
 .cfi_endproc	
 .size	sha1_block_data_order,.-sha1_block_data_order
+.type	sha1_block_data_order_shaext,@function
+.align	32
+sha1_block_data_order_shaext:
+_shaext_shortcut:
+.cfi_startproc	
+	movdqu	(%rdi),%xmm0
+	movd	16(%rdi),%xmm1
+	movdqa	K_XX_XX+160(%rip),%xmm3
+
+	movdqu	(%rsi),%xmm4
+	pshufd	$27,%xmm0,%xmm0
+	movdqu	16(%rsi),%xmm5
+	pshufd	$27,%xmm1,%xmm1
+	movdqu	32(%rsi),%xmm6
+.byte	102,15,56,0,227
+	movdqu	48(%rsi),%xmm7
+.byte	102,15,56,0,235
+.byte	102,15,56,0,243
+	movdqa	%xmm1,%xmm9
+.byte	102,15,56,0,251
+	jmp	.Loop_shaext
+
+.align	16
+.Loop_shaext:
+	decq	%rdx
+	leaq	64(%rsi),%r8
+	paddd	%xmm4,%xmm1
+	cmovneq	%r8,%rsi
+	movdqa	%xmm0,%xmm8
+.byte	15,56,201,229
+	movdqa	%xmm0,%xmm2
+.byte	15,58,204,193,0
+.byte	15,56,200,213
+	pxor	%xmm6,%xmm4
+.byte	15,56,201,238
+.byte	15,56,202,231
+
+	movdqa	%xmm0,%xmm1
+.byte	15,58,204,194,0
+.byte	15,56,200,206
+	pxor	%xmm7,%xmm5
+.byte	15,56,202,236
+.byte	15,56,201,247
+	movdqa	%xmm0,%xmm2
+.byte	15,58,204,193,0
+.byte	15,56,200,215
+	pxor	%xmm4,%xmm6
+.byte	15,56,201,252
+.byte	15,56,202,245
+
+	movdqa	%xmm0,%xmm1
+.byte	15,58,204,194,0
+.byte	15,56,200,204
+	pxor	%xmm5,%xmm7
+.byte	15,56,202,254
+.byte	15,56,201,229
+	movdqa	%xmm0,%xmm2
+.byte	15,58,204,193,0
+.byte	15,56,200,213
+	pxor	%xmm6,%xmm4
+.byte	15,56,201,238
+.byte	15,56,202,231
+
+	movdqa	%xmm0,%xmm1
+.byte	15,58,204,194,1
+.byte	15,56,200,206
+	pxor	%xmm7,%xmm5
+.byte	15,56,202,236
+.byte	15,56,201,247
+	movdqa	%xmm0,%xmm2
+.byte	15,58,204,193,1
+.byte	15,56,200,215
+	pxor	%xmm4,%xmm6
+.byte	15,56,201,252
+.byte	15,56,202,245
+
+	movdqa	%xmm0,%xmm1
+.byte	15,58,204,194,1
+.byte	15,56,200,204
+	pxor	%xmm5,%xmm7
+.byte	15,56,202,254
+.byte	15,56,201,229
+	movdqa	%xmm0,%xmm2
+.byte	15,58,204,193,1
+.byte	15,56,200,213
+	pxor	%xmm6,%xmm4
+.byte	15,56,201,238
+.byte	15,56,202,231
+
+	movdqa	%xmm0,%xmm1
+.byte	15,58,204,194,1
+.byte	15,56,200,206
+	pxor	%xmm7,%xmm5
+.byte	15,56,202,236
+.byte	15,56,201,247
+	movdqa	%xmm0,%xmm2
+.byte	15,58,204,193,2
+.byte	15,56,200,215
+	pxor	%xmm4,%xmm6
+.byte	15,56,201,252
+.byte	15,56,202,245
+
+	movdqa	%xmm0,%xmm1
+.byte	15,58,204,194,2
+.byte	15,56,200,204
+	pxor	%xmm5,%xmm7
+.byte	15,56,202,254
+.byte	15,56,201,229
+	movdqa	%xmm0,%xmm2
+.byte	15,58,204,193,2
+.byte	15,56,200,213
+	pxor	%xmm6,%xmm4
+.byte	15,56,201,238
+.byte	15,56,202,231
+
+	movdqa	%xmm0,%xmm1
+.byte	15,58,204,194,2
+.byte	15,56,200,206
+	pxor	%xmm7,%xmm5
+.byte	15,56,202,236
+.byte	15,56,201,247
+	movdqa	%xmm0,%xmm2
+.byte	15,58,204,193,2
+.byte	15,56,200,215
+	pxor	%xmm4,%xmm6
+.byte	15,56,201,252
+.byte	15,56,202,245
+
+	movdqa	%xmm0,%xmm1
+.byte	15,58,204,194,3
+.byte	15,56,200,204
+	pxor	%xmm5,%xmm7
+.byte	15,56,202,254
+	movdqu	(%rsi),%xmm4
+	movdqa	%xmm0,%xmm2
+.byte	15,58,204,193,3
+.byte	15,56,200,213
+	movdqu	16(%rsi),%xmm5
+.byte	102,15,56,0,227
+
+	movdqa	%xmm0,%xmm1
+.byte	15,58,204,194,3
+.byte	15,56,200,206
+	movdqu	32(%rsi),%xmm6
+.byte	102,15,56,0,235
+
+	movdqa	%xmm0,%xmm2
+.byte	15,58,204,193,3
+.byte	15,56,200,215
+	movdqu	48(%rsi),%xmm7
+.byte	102,15,56,0,243
+
+	movdqa	%xmm0,%xmm1
+.byte	15,58,204,194,3
+.byte	65,15,56,200,201
+.byte	102,15,56,0,251
+
+	paddd	%xmm8,%xmm0
+	movdqa	%xmm1,%xmm9
+
+	jnz	.Loop_shaext
+
+	pshufd	$27,%xmm0,%xmm0
+	pshufd	$27,%xmm1,%xmm1
+	movdqu	%xmm0,(%rdi)
+	movd	%xmm1,16(%rdi)
+	.byte	0xf3,0xc3
+.cfi_endproc	
+.size	sha1_block_data_order_shaext,.-sha1_block_data_order_shaext
 .type	sha1_block_data_order_ssse3,@function
 .align	16
 sha1_block_data_order_ssse3:
@@ -3582,6 +3756,1699 @@ _avx_shortcut:
 	.byte	0xf3,0xc3
 .cfi_endproc	
 .size	sha1_block_data_order_avx,.-sha1_block_data_order_avx
+.type	sha1_block_data_order_avx2,@function
+.align	16
+sha1_block_data_order_avx2:
+_avx2_shortcut:
+.cfi_startproc	
+	movq	%rsp,%r11
+.cfi_def_cfa_register	%r11
+	pushq	%rbx
+.cfi_offset	%rbx,-16
+	pushq	%rbp
+.cfi_offset	%rbp,-24
+	pushq	%r12
+.cfi_offset	%r12,-32
+	pushq	%r13
+.cfi_offset	%r13,-40
+	pushq	%r14
+.cfi_offset	%r14,-48
+	vzeroupper
+	movq	%rdi,%r8
+	movq	%rsi,%r9
+	movq	%rdx,%r10
+
+	leaq	-640(%rsp),%rsp
+	shlq	$6,%r10
+	leaq	64(%r9),%r13
+	andq	$-128,%rsp
+	addq	%r9,%r10
+	leaq	K_XX_XX+64(%rip),%r14
+
+	movl	0(%r8),%eax
+	cmpq	%r10,%r13
+	cmovaeq	%r9,%r13
+	movl	4(%r8),%ebp
+	movl	8(%r8),%ecx
+	movl	12(%r8),%edx
+	movl	16(%r8),%esi
+	vmovdqu	64(%r14),%ymm6
+
+	vmovdqu	(%r9),%xmm0
+	vmovdqu	16(%r9),%xmm1
+	vmovdqu	32(%r9),%xmm2
+	vmovdqu	48(%r9),%xmm3
+	leaq	64(%r9),%r9
+	vinserti128	$1,(%r13),%ymm0,%ymm0
+	vinserti128	$1,16(%r13),%ymm1,%ymm1
+	vpshufb	%ymm6,%ymm0,%ymm0
+	vinserti128	$1,32(%r13),%ymm2,%ymm2
+	vpshufb	%ymm6,%ymm1,%ymm1
+	vinserti128	$1,48(%r13),%ymm3,%ymm3
+	vpshufb	%ymm6,%ymm2,%ymm2
+	vmovdqu	-64(%r14),%ymm11
+	vpshufb	%ymm6,%ymm3,%ymm3
+
+	vpaddd	%ymm11,%ymm0,%ymm4
+	vpaddd	%ymm11,%ymm1,%ymm5
+	vmovdqu	%ymm4,0(%rsp)
+	vpaddd	%ymm11,%ymm2,%ymm6
+	vmovdqu	%ymm5,32(%rsp)
+	vpaddd	%ymm11,%ymm3,%ymm7
+	vmovdqu	%ymm6,64(%rsp)
+	vmovdqu	%ymm7,96(%rsp)
+	vpalignr	$8,%ymm0,%ymm1,%ymm4
+	vpsrldq	$4,%ymm3,%ymm8
+	vpxor	%ymm0,%ymm4,%ymm4
+	vpxor	%ymm2,%ymm8,%ymm8
+	vpxor	%ymm8,%ymm4,%ymm4
+	vpsrld	$31,%ymm4,%ymm8
+	vpslldq	$12,%ymm4,%ymm10
+	vpaddd	%ymm4,%ymm4,%ymm4
+	vpsrld	$30,%ymm10,%ymm9
+	vpor	%ymm8,%ymm4,%ymm4
+	vpslld	$2,%ymm10,%ymm10
+	vpxor	%ymm9,%ymm4,%ymm4
+	vpxor	%ymm10,%ymm4,%ymm4
+	vpaddd	%ymm11,%ymm4,%ymm9
+	vmovdqu	%ymm9,128(%rsp)
+	vpalignr	$8,%ymm1,%ymm2,%ymm5
+	vpsrldq	$4,%ymm4,%ymm8
+	vpxor	%ymm1,%ymm5,%ymm5
+	vpxor	%ymm3,%ymm8,%ymm8
+	vpxor	%ymm8,%ymm5,%ymm5
+	vpsrld	$31,%ymm5,%ymm8
+	vmovdqu	-32(%r14),%ymm11
+	vpslldq	$12,%ymm5,%ymm10
+	vpaddd	%ymm5,%ymm5,%ymm5
+	vpsrld	$30,%ymm10,%ymm9
+	vpor	%ymm8,%ymm5,%ymm5
+	vpslld	$2,%ymm10,%ymm10
+	vpxor	%ymm9,%ymm5,%ymm5
+	vpxor	%ymm10,%ymm5,%ymm5
+	vpaddd	%ymm11,%ymm5,%ymm9
+	vmovdqu	%ymm9,160(%rsp)
+	vpalignr	$8,%ymm2,%ymm3,%ymm6
+	vpsrldq	$4,%ymm5,%ymm8
+	vpxor	%ymm2,%ymm6,%ymm6
+	vpxor	%ymm4,%ymm8,%ymm8
+	vpxor	%ymm8,%ymm6,%ymm6
+	vpsrld	$31,%ymm6,%ymm8
+	vpslldq	$12,%ymm6,%ymm10
+	vpaddd	%ymm6,%ymm6,%ymm6
+	vpsrld	$30,%ymm10,%ymm9
+	vpor	%ymm8,%ymm6,%ymm6
+	vpslld	$2,%ymm10,%ymm10
+	vpxor	%ymm9,%ymm6,%ymm6
+	vpxor	%ymm10,%ymm6,%ymm6
+	vpaddd	%ymm11,%ymm6,%ymm9
+	vmovdqu	%ymm9,192(%rsp)
+	vpalignr	$8,%ymm3,%ymm4,%ymm7
+	vpsrldq	$4,%ymm6,%ymm8
+	vpxor	%ymm3,%ymm7,%ymm7
+	vpxor	%ymm5,%ymm8,%ymm8
+	vpxor	%ymm8,%ymm7,%ymm7
+	vpsrld	$31,%ymm7,%ymm8
+	vpslldq	$12,%ymm7,%ymm10
+	vpaddd	%ymm7,%ymm7,%ymm7
+	vpsrld	$30,%ymm10,%ymm9
+	vpor	%ymm8,%ymm7,%ymm7
+	vpslld	$2,%ymm10,%ymm10
+	vpxor	%ymm9,%ymm7,%ymm7
+	vpxor	%ymm10,%ymm7,%ymm7
+	vpaddd	%ymm11,%ymm7,%ymm9
+	vmovdqu	%ymm9,224(%rsp)
+	leaq	128(%rsp),%r13
+	jmp	.Loop_avx2
+.align	32
+.Loop_avx2:
+	rorxl	$2,%ebp,%ebx
+	andnl	%edx,%ebp,%edi
+	andl	%ecx,%ebp
+	xorl	%edi,%ebp
+	jmp	.Lalign32_1
+.align	32
+.Lalign32_1:
+	vpalignr	$8,%ymm6,%ymm7,%ymm8
+	vpxor	%ymm4,%ymm0,%ymm0
+	addl	-128(%r13),%esi
+	andnl	%ecx,%eax,%edi
+	vpxor	%ymm1,%ymm0,%ymm0
+	addl	%ebp,%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	vpxor	%ymm8,%ymm0,%ymm0
+	andl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%edi,%eax
+	vpsrld	$30,%ymm0,%ymm8
+	vpslld	$2,%ymm0,%ymm0
+	addl	-124(%r13),%edx
+	andnl	%ebx,%esi,%edi
+	addl	%eax,%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	andl	%ebp,%esi
+	vpor	%ymm8,%ymm0,%ymm0
+	addl	%r12d,%edx
+	xorl	%edi,%esi
+	addl	-120(%r13),%ecx
+	andnl	%ebp,%edx,%edi
+	vpaddd	%ymm11,%ymm0,%ymm9
+	addl	%esi,%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	andl	%eax,%edx
+	vmovdqu	%ymm9,256(%rsp)
+	addl	%r12d,%ecx
+	xorl	%edi,%edx
+	addl	-116(%r13),%ebx
+	andnl	%eax,%ecx,%edi
+	addl	%edx,%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	andl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%edi,%ecx
+	addl	-96(%r13),%ebp
+	andnl	%esi,%ebx,%edi
+	addl	%ecx,%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	andl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%edi,%ebx
+	vpalignr	$8,%ymm7,%ymm0,%ymm8
+	vpxor	%ymm5,%ymm1,%ymm1
+	addl	-92(%r13),%eax
+	andnl	%edx,%ebp,%edi
+	vpxor	%ymm2,%ymm1,%ymm1
+	addl	%ebx,%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	vpxor	%ymm8,%ymm1,%ymm1
+	andl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edi,%ebp
+	vpsrld	$30,%ymm1,%ymm8
+	vpslld	$2,%ymm1,%ymm1
+	addl	-88(%r13),%esi
+	andnl	%ecx,%eax,%edi
+	addl	%ebp,%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	andl	%ebx,%eax
+	vpor	%ymm8,%ymm1,%ymm1
+	addl	%r12d,%esi
+	xorl	%edi,%eax
+	addl	-84(%r13),%edx
+	andnl	%ebx,%esi,%edi
+	vpaddd	%ymm11,%ymm1,%ymm9
+	addl	%eax,%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	andl	%ebp,%esi
+	vmovdqu	%ymm9,288(%rsp)
+	addl	%r12d,%edx
+	xorl	%edi,%esi
+	addl	-64(%r13),%ecx
+	andnl	%ebp,%edx,%edi
+	addl	%esi,%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	andl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%edi,%edx
+	addl	-60(%r13),%ebx
+	andnl	%eax,%ecx,%edi
+	addl	%edx,%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	andl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%edi,%ecx
+	vpalignr	$8,%ymm0,%ymm1,%ymm8
+	vpxor	%ymm6,%ymm2,%ymm2
+	addl	-56(%r13),%ebp
+	andnl	%esi,%ebx,%edi
+	vpxor	%ymm3,%ymm2,%ymm2
+	vmovdqu	0(%r14),%ymm11
+	addl	%ecx,%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	vpxor	%ymm8,%ymm2,%ymm2
+	andl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%edi,%ebx
+	vpsrld	$30,%ymm2,%ymm8
+	vpslld	$2,%ymm2,%ymm2
+	addl	-52(%r13),%eax
+	andnl	%edx,%ebp,%edi
+	addl	%ebx,%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	andl	%ecx,%ebp
+	vpor	%ymm8,%ymm2,%ymm2
+	addl	%r12d,%eax
+	xorl	%edi,%ebp
+	addl	-32(%r13),%esi
+	andnl	%ecx,%eax,%edi
+	vpaddd	%ymm11,%ymm2,%ymm9
+	addl	%ebp,%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	andl	%ebx,%eax
+	vmovdqu	%ymm9,320(%rsp)
+	addl	%r12d,%esi
+	xorl	%edi,%eax
+	addl	-28(%r13),%edx
+	andnl	%ebx,%esi,%edi
+	addl	%eax,%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	andl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%edi,%esi
+	addl	-24(%r13),%ecx
+	andnl	%ebp,%edx,%edi
+	addl	%esi,%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	andl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%edi,%edx
+	vpalignr	$8,%ymm1,%ymm2,%ymm8
+	vpxor	%ymm7,%ymm3,%ymm3
+	addl	-20(%r13),%ebx
+	andnl	%eax,%ecx,%edi
+	vpxor	%ymm4,%ymm3,%ymm3
+	addl	%edx,%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	vpxor	%ymm8,%ymm3,%ymm3
+	andl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%edi,%ecx
+	vpsrld	$30,%ymm3,%ymm8
+	vpslld	$2,%ymm3,%ymm3
+	addl	0(%r13),%ebp
+	andnl	%esi,%ebx,%edi
+	addl	%ecx,%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	andl	%edx,%ebx
+	vpor	%ymm8,%ymm3,%ymm3
+	addl	%r12d,%ebp
+	xorl	%edi,%ebx
+	addl	4(%r13),%eax
+	andnl	%edx,%ebp,%edi
+	vpaddd	%ymm11,%ymm3,%ymm9
+	addl	%ebx,%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	andl	%ecx,%ebp
+	vmovdqu	%ymm9,352(%rsp)
+	addl	%r12d,%eax
+	xorl	%edi,%ebp
+	addl	8(%r13),%esi
+	andnl	%ecx,%eax,%edi
+	addl	%ebp,%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	andl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%edi,%eax
+	addl	12(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	vpalignr	$8,%ymm2,%ymm3,%ymm8
+	vpxor	%ymm0,%ymm4,%ymm4
+	addl	32(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	vpxor	%ymm5,%ymm4,%ymm4
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	vpxor	%ymm8,%ymm4,%ymm4
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	36(%r13),%ebx
+	vpsrld	$30,%ymm4,%ymm8
+	vpslld	$2,%ymm4,%ymm4
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	vpor	%ymm8,%ymm4,%ymm4
+	addl	40(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	vpaddd	%ymm11,%ymm4,%ymm9
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	addl	44(%r13),%eax
+	vmovdqu	%ymm9,384(%rsp)
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	addl	64(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	vpalignr	$8,%ymm3,%ymm4,%ymm8
+	vpxor	%ymm1,%ymm5,%ymm5
+	addl	68(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	vpxor	%ymm6,%ymm5,%ymm5
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	vpxor	%ymm8,%ymm5,%ymm5
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	addl	72(%r13),%ecx
+	vpsrld	$30,%ymm5,%ymm8
+	vpslld	$2,%ymm5,%ymm5
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	vpor	%ymm8,%ymm5,%ymm5
+	addl	76(%r13),%ebx
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	vpaddd	%ymm11,%ymm5,%ymm9
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	addl	96(%r13),%ebp
+	vmovdqu	%ymm9,416(%rsp)
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	addl	100(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	vpalignr	$8,%ymm4,%ymm5,%ymm8
+	vpxor	%ymm2,%ymm6,%ymm6
+	addl	104(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	vpxor	%ymm7,%ymm6,%ymm6
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	vpxor	%ymm8,%ymm6,%ymm6
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	addl	108(%r13),%edx
+	leaq	256(%r13),%r13
+	vpsrld	$30,%ymm6,%ymm8
+	vpslld	$2,%ymm6,%ymm6
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	vpor	%ymm8,%ymm6,%ymm6
+	addl	-128(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	vpaddd	%ymm11,%ymm6,%ymm9
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	-124(%r13),%ebx
+	vmovdqu	%ymm9,448(%rsp)
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	addl	-120(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	vpalignr	$8,%ymm5,%ymm6,%ymm8
+	vpxor	%ymm3,%ymm7,%ymm7
+	addl	-116(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	vpxor	%ymm0,%ymm7,%ymm7
+	vmovdqu	32(%r14),%ymm11
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	vpxor	%ymm8,%ymm7,%ymm7
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	addl	-96(%r13),%esi
+	vpsrld	$30,%ymm7,%ymm8
+	vpslld	$2,%ymm7,%ymm7
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	vpor	%ymm8,%ymm7,%ymm7
+	addl	-92(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	vpaddd	%ymm11,%ymm7,%ymm9
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	addl	-88(%r13),%ecx
+	vmovdqu	%ymm9,480(%rsp)
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	-84(%r13),%ebx
+	movl	%esi,%edi
+	xorl	%eax,%edi
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	andl	%edi,%ecx
+	jmp	.Lalign32_2
+.align	32
+.Lalign32_2:
+	vpalignr	$8,%ymm6,%ymm7,%ymm8
+	vpxor	%ymm4,%ymm0,%ymm0
+	addl	-64(%r13),%ebp
+	xorl	%esi,%ecx
+	vpxor	%ymm1,%ymm0,%ymm0
+	movl	%edx,%edi
+	xorl	%esi,%edi
+	leal	(%rcx,%rbp,1),%ebp
+	vpxor	%ymm8,%ymm0,%ymm0
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	vpsrld	$30,%ymm0,%ymm8
+	vpslld	$2,%ymm0,%ymm0
+	addl	%r12d,%ebp
+	andl	%edi,%ebx
+	addl	-60(%r13),%eax
+	xorl	%edx,%ebx
+	movl	%ecx,%edi
+	xorl	%edx,%edi
+	vpor	%ymm8,%ymm0,%ymm0
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	vpaddd	%ymm11,%ymm0,%ymm9
+	addl	%r12d,%eax
+	andl	%edi,%ebp
+	addl	-56(%r13),%esi
+	xorl	%ecx,%ebp
+	vmovdqu	%ymm9,512(%rsp)
+	movl	%ebx,%edi
+	xorl	%ecx,%edi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	andl	%edi,%eax
+	addl	-52(%r13),%edx
+	xorl	%ebx,%eax
+	movl	%ebp,%edi
+	xorl	%ebx,%edi
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	andl	%edi,%esi
+	addl	-32(%r13),%ecx
+	xorl	%ebp,%esi
+	movl	%eax,%edi
+	xorl	%ebp,%edi
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	andl	%edi,%edx
+	vpalignr	$8,%ymm7,%ymm0,%ymm8
+	vpxor	%ymm5,%ymm1,%ymm1
+	addl	-28(%r13),%ebx
+	xorl	%eax,%edx
+	vpxor	%ymm2,%ymm1,%ymm1
+	movl	%esi,%edi
+	xorl	%eax,%edi
+	leal	(%rbx,%rdx,1),%ebx
+	vpxor	%ymm8,%ymm1,%ymm1
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	vpsrld	$30,%ymm1,%ymm8
+	vpslld	$2,%ymm1,%ymm1
+	addl	%r12d,%ebx
+	andl	%edi,%ecx
+	addl	-24(%r13),%ebp
+	xorl	%esi,%ecx
+	movl	%edx,%edi
+	xorl	%esi,%edi
+	vpor	%ymm8,%ymm1,%ymm1
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	vpaddd	%ymm11,%ymm1,%ymm9
+	addl	%r12d,%ebp
+	andl	%edi,%ebx
+	addl	-20(%r13),%eax
+	xorl	%edx,%ebx
+	vmovdqu	%ymm9,544(%rsp)
+	movl	%ecx,%edi
+	xorl	%edx,%edi
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	andl	%edi,%ebp
+	addl	0(%r13),%esi
+	xorl	%ecx,%ebp
+	movl	%ebx,%edi
+	xorl	%ecx,%edi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	andl	%edi,%eax
+	addl	4(%r13),%edx
+	xorl	%ebx,%eax
+	movl	%ebp,%edi
+	xorl	%ebx,%edi
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	andl	%edi,%esi
+	vpalignr	$8,%ymm0,%ymm1,%ymm8
+	vpxor	%ymm6,%ymm2,%ymm2
+	addl	8(%r13),%ecx
+	xorl	%ebp,%esi
+	vpxor	%ymm3,%ymm2,%ymm2
+	movl	%eax,%edi
+	xorl	%ebp,%edi
+	leal	(%rcx,%rsi,1),%ecx
+	vpxor	%ymm8,%ymm2,%ymm2
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	vpsrld	$30,%ymm2,%ymm8
+	vpslld	$2,%ymm2,%ymm2
+	addl	%r12d,%ecx
+	andl	%edi,%edx
+	addl	12(%r13),%ebx
+	xorl	%eax,%edx
+	movl	%esi,%edi
+	xorl	%eax,%edi
+	vpor	%ymm8,%ymm2,%ymm2
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	vpaddd	%ymm11,%ymm2,%ymm9
+	addl	%r12d,%ebx
+	andl	%edi,%ecx
+	addl	32(%r13),%ebp
+	xorl	%esi,%ecx
+	vmovdqu	%ymm9,576(%rsp)
+	movl	%edx,%edi
+	xorl	%esi,%edi
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	andl	%edi,%ebx
+	addl	36(%r13),%eax
+	xorl	%edx,%ebx
+	movl	%ecx,%edi
+	xorl	%edx,%edi
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	andl	%edi,%ebp
+	addl	40(%r13),%esi
+	xorl	%ecx,%ebp
+	movl	%ebx,%edi
+	xorl	%ecx,%edi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	andl	%edi,%eax
+	vpalignr	$8,%ymm1,%ymm2,%ymm8
+	vpxor	%ymm7,%ymm3,%ymm3
+	addl	44(%r13),%edx
+	xorl	%ebx,%eax
+	vpxor	%ymm4,%ymm3,%ymm3
+	movl	%ebp,%edi
+	xorl	%ebx,%edi
+	leal	(%rdx,%rax,1),%edx
+	vpxor	%ymm8,%ymm3,%ymm3
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	vpsrld	$30,%ymm3,%ymm8
+	vpslld	$2,%ymm3,%ymm3
+	addl	%r12d,%edx
+	andl	%edi,%esi
+	addl	64(%r13),%ecx
+	xorl	%ebp,%esi
+	movl	%eax,%edi
+	xorl	%ebp,%edi
+	vpor	%ymm8,%ymm3,%ymm3
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	vpaddd	%ymm11,%ymm3,%ymm9
+	addl	%r12d,%ecx
+	andl	%edi,%edx
+	addl	68(%r13),%ebx
+	xorl	%eax,%edx
+	vmovdqu	%ymm9,608(%rsp)
+	movl	%esi,%edi
+	xorl	%eax,%edi
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	andl	%edi,%ecx
+	addl	72(%r13),%ebp
+	xorl	%esi,%ecx
+	movl	%edx,%edi
+	xorl	%esi,%edi
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	andl	%edi,%ebx
+	addl	76(%r13),%eax
+	xorl	%edx,%ebx
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	addl	96(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	addl	100(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	addl	104(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	108(%r13),%ebx
+	leaq	256(%r13),%r13
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	addl	-128(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	addl	-124(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	addl	-120(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	addl	-116(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	addl	-96(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	-92(%r13),%ebx
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	addl	-88(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	addl	-84(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	addl	-64(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	addl	-60(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	addl	-56(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	-52(%r13),%ebx
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	addl	-32(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	addl	-28(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	addl	-24(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	addl	-20(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	addl	%r12d,%edx
+	leaq	128(%r9),%r13
+	leaq	128(%r9),%rdi
+	cmpq	%r10,%r13
+	cmovaeq	%r9,%r13
+
+
+	addl	0(%r8),%edx
+	addl	4(%r8),%esi
+	addl	8(%r8),%ebp
+	movl	%edx,0(%r8)
+	addl	12(%r8),%ebx
+	movl	%esi,4(%r8)
+	movl	%edx,%eax
+	addl	16(%r8),%ecx
+	movl	%ebp,%r12d
+	movl	%ebp,8(%r8)
+	movl	%ebx,%edx
+
+	movl	%ebx,12(%r8)
+	movl	%esi,%ebp
+	movl	%ecx,16(%r8)
+
+	movl	%ecx,%esi
+	movl	%r12d,%ecx
+
+
+	cmpq	%r10,%r9
+	je	.Ldone_avx2
+	vmovdqu	64(%r14),%ymm6
+	cmpq	%r10,%rdi
+	ja	.Last_avx2
+
+	vmovdqu	-64(%rdi),%xmm0
+	vmovdqu	-48(%rdi),%xmm1
+	vmovdqu	-32(%rdi),%xmm2
+	vmovdqu	-16(%rdi),%xmm3
+	vinserti128	$1,0(%r13),%ymm0,%ymm0
+	vinserti128	$1,16(%r13),%ymm1,%ymm1
+	vinserti128	$1,32(%r13),%ymm2,%ymm2
+	vinserti128	$1,48(%r13),%ymm3,%ymm3
+	jmp	.Last_avx2
+
+.align	32
+.Last_avx2:
+	leaq	128+16(%rsp),%r13
+	rorxl	$2,%ebp,%ebx
+	andnl	%edx,%ebp,%edi
+	andl	%ecx,%ebp
+	xorl	%edi,%ebp
+	subq	$-128,%r9
+	addl	-128(%r13),%esi
+	andnl	%ecx,%eax,%edi
+	addl	%ebp,%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	andl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%edi,%eax
+	addl	-124(%r13),%edx
+	andnl	%ebx,%esi,%edi
+	addl	%eax,%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	andl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%edi,%esi
+	addl	-120(%r13),%ecx
+	andnl	%ebp,%edx,%edi
+	addl	%esi,%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	andl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%edi,%edx
+	addl	-116(%r13),%ebx
+	andnl	%eax,%ecx,%edi
+	addl	%edx,%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	andl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%edi,%ecx
+	addl	-96(%r13),%ebp
+	andnl	%esi,%ebx,%edi
+	addl	%ecx,%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	andl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%edi,%ebx
+	addl	-92(%r13),%eax
+	andnl	%edx,%ebp,%edi
+	addl	%ebx,%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	andl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edi,%ebp
+	addl	-88(%r13),%esi
+	andnl	%ecx,%eax,%edi
+	addl	%ebp,%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	andl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%edi,%eax
+	addl	-84(%r13),%edx
+	andnl	%ebx,%esi,%edi
+	addl	%eax,%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	andl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%edi,%esi
+	addl	-64(%r13),%ecx
+	andnl	%ebp,%edx,%edi
+	addl	%esi,%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	andl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%edi,%edx
+	addl	-60(%r13),%ebx
+	andnl	%eax,%ecx,%edi
+	addl	%edx,%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	andl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%edi,%ecx
+	addl	-56(%r13),%ebp
+	andnl	%esi,%ebx,%edi
+	addl	%ecx,%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	andl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%edi,%ebx
+	addl	-52(%r13),%eax
+	andnl	%edx,%ebp,%edi
+	addl	%ebx,%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	andl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edi,%ebp
+	addl	-32(%r13),%esi
+	andnl	%ecx,%eax,%edi
+	addl	%ebp,%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	andl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%edi,%eax
+	addl	-28(%r13),%edx
+	andnl	%ebx,%esi,%edi
+	addl	%eax,%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	andl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%edi,%esi
+	addl	-24(%r13),%ecx
+	andnl	%ebp,%edx,%edi
+	addl	%esi,%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	andl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%edi,%edx
+	addl	-20(%r13),%ebx
+	andnl	%eax,%ecx,%edi
+	addl	%edx,%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	andl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%edi,%ecx
+	addl	0(%r13),%ebp
+	andnl	%esi,%ebx,%edi
+	addl	%ecx,%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	andl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%edi,%ebx
+	addl	4(%r13),%eax
+	andnl	%edx,%ebp,%edi
+	addl	%ebx,%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	andl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edi,%ebp
+	addl	8(%r13),%esi
+	andnl	%ecx,%eax,%edi
+	addl	%ebp,%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	andl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%edi,%eax
+	addl	12(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	addl	32(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	36(%r13),%ebx
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	addl	40(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	addl	44(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	addl	64(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	vmovdqu	-64(%r14),%ymm11
+	vpshufb	%ymm6,%ymm0,%ymm0
+	addl	68(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	addl	72(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	76(%r13),%ebx
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	addl	96(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	addl	100(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	vpshufb	%ymm6,%ymm1,%ymm1
+	vpaddd	%ymm11,%ymm0,%ymm8
+	addl	104(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	addl	108(%r13),%edx
+	leaq	256(%r13),%r13
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	addl	-128(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	-124(%r13),%ebx
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	addl	-120(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	vmovdqu	%ymm8,0(%rsp)
+	vpshufb	%ymm6,%ymm2,%ymm2
+	vpaddd	%ymm11,%ymm1,%ymm9
+	addl	-116(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	addl	-96(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	addl	-92(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	addl	-88(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	-84(%r13),%ebx
+	movl	%esi,%edi
+	xorl	%eax,%edi
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	andl	%edi,%ecx
+	vmovdqu	%ymm9,32(%rsp)
+	vpshufb	%ymm6,%ymm3,%ymm3
+	vpaddd	%ymm11,%ymm2,%ymm6
+	addl	-64(%r13),%ebp
+	xorl	%esi,%ecx
+	movl	%edx,%edi
+	xorl	%esi,%edi
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	andl	%edi,%ebx
+	addl	-60(%r13),%eax
+	xorl	%edx,%ebx
+	movl	%ecx,%edi
+	xorl	%edx,%edi
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	andl	%edi,%ebp
+	addl	-56(%r13),%esi
+	xorl	%ecx,%ebp
+	movl	%ebx,%edi
+	xorl	%ecx,%edi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	andl	%edi,%eax
+	addl	-52(%r13),%edx
+	xorl	%ebx,%eax
+	movl	%ebp,%edi
+	xorl	%ebx,%edi
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	andl	%edi,%esi
+	addl	-32(%r13),%ecx
+	xorl	%ebp,%esi
+	movl	%eax,%edi
+	xorl	%ebp,%edi
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	andl	%edi,%edx
+	jmp	.Lalign32_3
+.align	32
+.Lalign32_3:
+	vmovdqu	%ymm6,64(%rsp)
+	vpaddd	%ymm11,%ymm3,%ymm7
+	addl	-28(%r13),%ebx
+	xorl	%eax,%edx
+	movl	%esi,%edi
+	xorl	%eax,%edi
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	andl	%edi,%ecx
+	addl	-24(%r13),%ebp
+	xorl	%esi,%ecx
+	movl	%edx,%edi
+	xorl	%esi,%edi
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	andl	%edi,%ebx
+	addl	-20(%r13),%eax
+	xorl	%edx,%ebx
+	movl	%ecx,%edi
+	xorl	%edx,%edi
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	andl	%edi,%ebp
+	addl	0(%r13),%esi
+	xorl	%ecx,%ebp
+	movl	%ebx,%edi
+	xorl	%ecx,%edi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	andl	%edi,%eax
+	addl	4(%r13),%edx
+	xorl	%ebx,%eax
+	movl	%ebp,%edi
+	xorl	%ebx,%edi
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	andl	%edi,%esi
+	vmovdqu	%ymm7,96(%rsp)
+	addl	8(%r13),%ecx
+	xorl	%ebp,%esi
+	movl	%eax,%edi
+	xorl	%ebp,%edi
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	andl	%edi,%edx
+	addl	12(%r13),%ebx
+	xorl	%eax,%edx
+	movl	%esi,%edi
+	xorl	%eax,%edi
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	andl	%edi,%ecx
+	addl	32(%r13),%ebp
+	xorl	%esi,%ecx
+	movl	%edx,%edi
+	xorl	%esi,%edi
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	andl	%edi,%ebx
+	addl	36(%r13),%eax
+	xorl	%edx,%ebx
+	movl	%ecx,%edi
+	xorl	%edx,%edi
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	andl	%edi,%ebp
+	addl	40(%r13),%esi
+	xorl	%ecx,%ebp
+	movl	%ebx,%edi
+	xorl	%ecx,%edi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	andl	%edi,%eax
+	vpalignr	$8,%ymm0,%ymm1,%ymm4
+	addl	44(%r13),%edx
+	xorl	%ebx,%eax
+	movl	%ebp,%edi
+	xorl	%ebx,%edi
+	vpsrldq	$4,%ymm3,%ymm8
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	vpxor	%ymm0,%ymm4,%ymm4
+	vpxor	%ymm2,%ymm8,%ymm8
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	vpxor	%ymm8,%ymm4,%ymm4
+	andl	%edi,%esi
+	addl	64(%r13),%ecx
+	xorl	%ebp,%esi
+	movl	%eax,%edi
+	vpsrld	$31,%ymm4,%ymm8
+	xorl	%ebp,%edi
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	vpslldq	$12,%ymm4,%ymm10
+	vpaddd	%ymm4,%ymm4,%ymm4
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	vpsrld	$30,%ymm10,%ymm9
+	vpor	%ymm8,%ymm4,%ymm4
+	addl	%r12d,%ecx
+	andl	%edi,%edx
+	vpslld	$2,%ymm10,%ymm10
+	vpxor	%ymm9,%ymm4,%ymm4
+	addl	68(%r13),%ebx
+	xorl	%eax,%edx
+	vpxor	%ymm10,%ymm4,%ymm4
+	movl	%esi,%edi
+	xorl	%eax,%edi
+	leal	(%rbx,%rdx,1),%ebx
+	vpaddd	%ymm11,%ymm4,%ymm9
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	vmovdqu	%ymm9,128(%rsp)
+	addl	%r12d,%ebx
+	andl	%edi,%ecx
+	addl	72(%r13),%ebp
+	xorl	%esi,%ecx
+	movl	%edx,%edi
+	xorl	%esi,%edi
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	andl	%edi,%ebx
+	addl	76(%r13),%eax
+	xorl	%edx,%ebx
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	vpalignr	$8,%ymm1,%ymm2,%ymm5
+	addl	96(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	vpsrldq	$4,%ymm4,%ymm8
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	vpxor	%ymm1,%ymm5,%ymm5
+	vpxor	%ymm3,%ymm8,%ymm8
+	addl	100(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	vpxor	%ymm8,%ymm5,%ymm5
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	vpsrld	$31,%ymm5,%ymm8
+	vmovdqu	-32(%r14),%ymm11
+	xorl	%ebx,%esi
+	addl	104(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	vpslldq	$12,%ymm5,%ymm10
+	vpaddd	%ymm5,%ymm5,%ymm5
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	vpsrld	$30,%ymm10,%ymm9
+	vpor	%ymm8,%ymm5,%ymm5
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	vpslld	$2,%ymm10,%ymm10
+	vpxor	%ymm9,%ymm5,%ymm5
+	xorl	%ebp,%edx
+	addl	108(%r13),%ebx
+	leaq	256(%r13),%r13
+	vpxor	%ymm10,%ymm5,%ymm5
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	vpaddd	%ymm11,%ymm5,%ymm9
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	vmovdqu	%ymm9,160(%rsp)
+	addl	-128(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	vpalignr	$8,%ymm2,%ymm3,%ymm6
+	addl	-124(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	vpsrldq	$4,%ymm5,%ymm8
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	vpxor	%ymm2,%ymm6,%ymm6
+	vpxor	%ymm4,%ymm8,%ymm8
+	addl	-120(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	vpxor	%ymm8,%ymm6,%ymm6
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	vpsrld	$31,%ymm6,%ymm8
+	xorl	%ecx,%eax
+	addl	-116(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	vpslldq	$12,%ymm6,%ymm10
+	vpaddd	%ymm6,%ymm6,%ymm6
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	vpsrld	$30,%ymm10,%ymm9
+	vpor	%ymm8,%ymm6,%ymm6
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	vpslld	$2,%ymm10,%ymm10
+	vpxor	%ymm9,%ymm6,%ymm6
+	xorl	%ebx,%esi
+	addl	-96(%r13),%ecx
+	vpxor	%ymm10,%ymm6,%ymm6
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	vpaddd	%ymm11,%ymm6,%ymm9
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	vmovdqu	%ymm9,192(%rsp)
+	addl	-92(%r13),%ebx
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	vpalignr	$8,%ymm3,%ymm4,%ymm7
+	addl	-88(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	vpsrldq	$4,%ymm6,%ymm8
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	vpxor	%ymm3,%ymm7,%ymm7
+	vpxor	%ymm5,%ymm8,%ymm8
+	addl	-84(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	vpxor	%ymm8,%ymm7,%ymm7
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	vpsrld	$31,%ymm7,%ymm8
+	xorl	%edx,%ebp
+	addl	-64(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	vpslldq	$12,%ymm7,%ymm10
+	vpaddd	%ymm7,%ymm7,%ymm7
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	vpsrld	$30,%ymm10,%ymm9
+	vpor	%ymm8,%ymm7,%ymm7
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	vpslld	$2,%ymm10,%ymm10
+	vpxor	%ymm9,%ymm7,%ymm7
+	xorl	%ecx,%eax
+	addl	-60(%r13),%edx
+	vpxor	%ymm10,%ymm7,%ymm7
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	rorxl	$2,%esi,%eax
+	vpaddd	%ymm11,%ymm7,%ymm9
+	xorl	%ebp,%esi
+	addl	%r12d,%edx
+	xorl	%ebx,%esi
+	vmovdqu	%ymm9,224(%rsp)
+	addl	-56(%r13),%ecx
+	leal	(%rcx,%rsi,1),%ecx
+	rorxl	$27,%edx,%r12d
+	rorxl	$2,%edx,%esi
+	xorl	%eax,%edx
+	addl	%r12d,%ecx
+	xorl	%ebp,%edx
+	addl	-52(%r13),%ebx
+	leal	(%rbx,%rdx,1),%ebx
+	rorxl	$27,%ecx,%r12d
+	rorxl	$2,%ecx,%edx
+	xorl	%esi,%ecx
+	addl	%r12d,%ebx
+	xorl	%eax,%ecx
+	addl	-32(%r13),%ebp
+	leal	(%rcx,%rbp,1),%ebp
+	rorxl	$27,%ebx,%r12d
+	rorxl	$2,%ebx,%ecx
+	xorl	%edx,%ebx
+	addl	%r12d,%ebp
+	xorl	%esi,%ebx
+	addl	-28(%r13),%eax
+	leal	(%rax,%rbx,1),%eax
+	rorxl	$27,%ebp,%r12d
+	rorxl	$2,%ebp,%ebx
+	xorl	%ecx,%ebp
+	addl	%r12d,%eax
+	xorl	%edx,%ebp
+	addl	-24(%r13),%esi
+	leal	(%rsi,%rbp,1),%esi
+	rorxl	$27,%eax,%r12d
+	rorxl	$2,%eax,%ebp
+	xorl	%ebx,%eax
+	addl	%r12d,%esi
+	xorl	%ecx,%eax
+	addl	-20(%r13),%edx
+	leal	(%rdx,%rax,1),%edx
+	rorxl	$27,%esi,%r12d
+	addl	%r12d,%edx
+	leaq	128(%rsp),%r13
+
+
+	addl	0(%r8),%edx
+	addl	4(%r8),%esi
+	addl	8(%r8),%ebp
+	movl	%edx,0(%r8)
+	addl	12(%r8),%ebx
+	movl	%esi,4(%r8)
+	movl	%edx,%eax
+	addl	16(%r8),%ecx
+	movl	%ebp,%r12d
+	movl	%ebp,8(%r8)
+	movl	%ebx,%edx
+
+	movl	%ebx,12(%r8)
+	movl	%esi,%ebp
+	movl	%ecx,16(%r8)
+
+	movl	%ecx,%esi
+	movl	%r12d,%ecx
+
+
+	cmpq	%r10,%r9
+	jbe	.Loop_avx2
+
+.Ldone_avx2:
+	vzeroupper
+	movq	-40(%r11),%r14
+.cfi_restore	%r14
+	movq	-32(%r11),%r13
+.cfi_restore	%r13
+	movq	-24(%r11),%r12
+.cfi_restore	%r12
+	movq	-16(%r11),%rbp
+.cfi_restore	%rbp
+	movq	-8(%r11),%rbx
+.cfi_restore	%rbx
+	leaq	(%r11),%rsp
+.cfi_def_cfa_register	%rsp
+.Lepilogue_avx2:
+	.byte	0xf3,0xc3
+.cfi_endproc	
+.size	sha1_block_data_order_avx2,.-sha1_block_data_order_avx2
 .align	64
 K_XX_XX:
 .long	0x5a827999,0x5a827999,0x5a827999,0x5a827999