Rolls BoringSSL forward

third_party/boringssl/mac-x86_64/crypto/sha/sha1-x86_64.S

Index: third_party/boringssl/mac-x86_64/crypto/sha/sha1-x86_64.S
diff --git a/third_party/boringssl/mac-x86_64/crypto/sha/sha1-x86_64.S b/third_party/boringssl/mac-x86_64/crypto/sha/sha1-x86_64.S
index 044dc5b7cd0f478c23c655e8785ce98215f004f9..0509d451635a1f4f4aa3341a890cba7a99146db8 100644
--- a/third_party/boringssl/mac-x86_64/crypto/sha/sha1-x86_64.S
+++ b/third_party/boringssl/mac-x86_64/crypto/sha/sha1-x86_64.S
@@ -12,6 +12,11 @@ _sha1_block_data_order:
 	movl	_OPENSSL_ia32cap_P+8(%rip),%r10d
 	testl	$512,%r8d
 	jz	L$ialu
+	andl	$268435456,%r8d
+	andl	$1073741824,%r9d
+	orl	%r9d,%r8d
+	cmpl	$1342177280,%r8d
+	je	_avx_shortcut
 	jmp	_ssse3_shortcut
 
 .p2align	4
@@ -2407,6 +2412,1122 @@ L$done_ssse3:
 L$epilogue_ssse3:
 	.byte	0xf3,0xc3
 
+
+.p2align	4
+sha1_block_data_order_avx:
+_avx_shortcut:
+	movq	%rsp,%rax
+	pushq	%rbx
+	pushq	%rbp
+	pushq	%r12
+	pushq	%r13
+	pushq	%r14
+	leaq	-64(%rsp),%rsp
+	vzeroupper
+	movq	%rax,%r14
+	andq	$-64,%rsp
+	movq	%rdi,%r8
+	movq	%rsi,%r9
+	movq	%rdx,%r10
+
+	shlq	$6,%r10
+	addq	%r9,%r10
+	leaq	K_XX_XX+64(%rip),%r11
+
+	movl	0(%r8),%eax
+	movl	4(%r8),%ebx
+	movl	8(%r8),%ecx
+	movl	12(%r8),%edx
+	movl	%ebx,%esi
+	movl	16(%r8),%ebp
+	movl	%ecx,%edi
+	xorl	%edx,%edi
+	andl	%edi,%esi
+
+	vmovdqa	64(%r11),%xmm6
+	vmovdqa	-64(%r11),%xmm11
+	vmovdqu	0(%r9),%xmm0
+	vmovdqu	16(%r9),%xmm1
+	vmovdqu	32(%r9),%xmm2
+	vmovdqu	48(%r9),%xmm3
+	vpshufb	%xmm6,%xmm0,%xmm0
+	addq	$64,%r9
+	vpshufb	%xmm6,%xmm1,%xmm1
+	vpshufb	%xmm6,%xmm2,%xmm2
+	vpshufb	%xmm6,%xmm3,%xmm3
+	vpaddd	%xmm11,%xmm0,%xmm4
+	vpaddd	%xmm11,%xmm1,%xmm5
+	vpaddd	%xmm11,%xmm2,%xmm6
+	vmovdqa	%xmm4,0(%rsp)
+	vmovdqa	%xmm5,16(%rsp)
+	vmovdqa	%xmm6,32(%rsp)
+	jmp	L$oop_avx
+.p2align	4
+L$oop_avx:
+	shrdl	$2,%ebx,%ebx
+	xorl	%edx,%esi
+	vpalignr	$8,%xmm0,%xmm1,%xmm4
+	movl	%eax,%edi
+	addl	0(%rsp),%ebp
+	vpaddd	%xmm3,%xmm11,%xmm9
+	xorl	%ecx,%ebx
+	shldl	$5,%eax,%eax
+	vpsrldq	$4,%xmm3,%xmm8
+	addl	%esi,%ebp
+	andl	%ebx,%edi
+	vpxor	%xmm0,%xmm4,%xmm4
+	xorl	%ecx,%ebx
+	addl	%eax,%ebp
+	vpxor	%xmm2,%xmm8,%xmm8
+	shrdl	$7,%eax,%eax
+	xorl	%ecx,%edi
+	movl	%ebp,%esi
+	addl	4(%rsp),%edx
+	vpxor	%xmm8,%xmm4,%xmm4
+	xorl	%ebx,%eax
+	shldl	$5,%ebp,%ebp
+	vmovdqa	%xmm9,48(%rsp)
+	addl	%edi,%edx
+	andl	%eax,%esi
+	vpsrld	$31,%xmm4,%xmm8
+	xorl	%ebx,%eax
+	addl	%ebp,%edx
+	shrdl	$7,%ebp,%ebp
+	xorl	%ebx,%esi
+	vpslldq	$12,%xmm4,%xmm10
+	vpaddd	%xmm4,%xmm4,%xmm4
+	movl	%edx,%edi
+	addl	8(%rsp),%ecx
+	xorl	%eax,%ebp
+	shldl	$5,%edx,%edx
+	vpsrld	$30,%xmm10,%xmm9
+	vpor	%xmm8,%xmm4,%xmm4
+	addl	%esi,%ecx
+	andl	%ebp,%edi
+	xorl	%eax,%ebp
+	addl	%edx,%ecx
+	vpslld	$2,%xmm10,%xmm10
+	vpxor	%xmm9,%xmm4,%xmm4
+	shrdl	$7,%edx,%edx
+	xorl	%eax,%edi
+	movl	%ecx,%esi
+	addl	12(%rsp),%ebx
+	vpxor	%xmm10,%xmm4,%xmm4
+	xorl	%ebp,%edx
+	shldl	$5,%ecx,%ecx
+	addl	%edi,%ebx
+	andl	%edx,%esi
+	xorl	%ebp,%edx
+	addl	%ecx,%ebx
+	shrdl	$7,%ecx,%ecx
+	xorl	%ebp,%esi
+	vpalignr	$8,%xmm1,%xmm2,%xmm5
+	movl	%ebx,%edi
+	addl	16(%rsp),%eax
+	vpaddd	%xmm4,%xmm11,%xmm9
+	xorl	%edx,%ecx
+	shldl	$5,%ebx,%ebx
+	vpsrldq	$4,%xmm4,%xmm8
+	addl	%esi,%eax
+	andl	%ecx,%edi
+	vpxor	%xmm1,%xmm5,%xmm5
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	vpxor	%xmm3,%xmm8,%xmm8
+	shrdl	$7,%ebx,%ebx
+	xorl	%edx,%edi
+	movl	%eax,%esi
+	addl	20(%rsp),%ebp
+	vpxor	%xmm8,%xmm5,%xmm5
+	xorl	%ecx,%ebx
+	shldl	$5,%eax,%eax
+	vmovdqa	%xmm9,0(%rsp)
+	addl	%edi,%ebp
+	andl	%ebx,%esi
+	vpsrld	$31,%xmm5,%xmm8
+	xorl	%ecx,%ebx
+	addl	%eax,%ebp
+	shrdl	$7,%eax,%eax
+	xorl	%ecx,%esi
+	vpslldq	$12,%xmm5,%xmm10
+	vpaddd	%xmm5,%xmm5,%xmm5
+	movl	%ebp,%edi
+	addl	24(%rsp),%edx
+	xorl	%ebx,%eax
+	shldl	$5,%ebp,%ebp
+	vpsrld	$30,%xmm10,%xmm9
+	vpor	%xmm8,%xmm5,%xmm5
+	addl	%esi,%edx
+	andl	%eax,%edi
+	xorl	%ebx,%eax
+	addl	%ebp,%edx
+	vpslld	$2,%xmm10,%xmm10
+	vpxor	%xmm9,%xmm5,%xmm5
+	shrdl	$7,%ebp,%ebp
+	xorl	%ebx,%edi
+	movl	%edx,%esi
+	addl	28(%rsp),%ecx
+	vpxor	%xmm10,%xmm5,%xmm5
+	xorl	%eax,%ebp
+	shldl	$5,%edx,%edx
+	vmovdqa	-32(%r11),%xmm11
+	addl	%edi,%ecx
+	andl	%ebp,%esi
+	xorl	%eax,%ebp
+	addl	%edx,%ecx
+	shrdl	$7,%edx,%edx
+	xorl	%eax,%esi
+	vpalignr	$8,%xmm2,%xmm3,%xmm6
+	movl	%ecx,%edi
+	addl	32(%rsp),%ebx
+	vpaddd	%xmm5,%xmm11,%xmm9
+	xorl	%ebp,%edx
+	shldl	$5,%ecx,%ecx
+	vpsrldq	$4,%xmm5,%xmm8
+	addl	%esi,%ebx
+	andl	%edx,%edi
+	vpxor	%xmm2,%xmm6,%xmm6
+	xorl	%ebp,%edx
+	addl	%ecx,%ebx
+	vpxor	%xmm4,%xmm8,%xmm8
+	shrdl	$7,%ecx,%ecx
+	xorl	%ebp,%edi
+	movl	%ebx,%esi
+	addl	36(%rsp),%eax
+	vpxor	%xmm8,%xmm6,%xmm6
+	xorl	%edx,%ecx
+	shldl	$5,%ebx,%ebx
+	vmovdqa	%xmm9,16(%rsp)
+	addl	%edi,%eax
+	andl	%ecx,%esi
+	vpsrld	$31,%xmm6,%xmm8
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	shrdl	$7,%ebx,%ebx
+	xorl	%edx,%esi
+	vpslldq	$12,%xmm6,%xmm10
+	vpaddd	%xmm6,%xmm6,%xmm6
+	movl	%eax,%edi
+	addl	40(%rsp),%ebp
+	xorl	%ecx,%ebx
+	shldl	$5,%eax,%eax
+	vpsrld	$30,%xmm10,%xmm9
+	vpor	%xmm8,%xmm6,%xmm6
+	addl	%esi,%ebp
+	andl	%ebx,%edi
+	xorl	%ecx,%ebx
+	addl	%eax,%ebp
+	vpslld	$2,%xmm10,%xmm10
+	vpxor	%xmm9,%xmm6,%xmm6
+	shrdl	$7,%eax,%eax
+	xorl	%ecx,%edi
+	movl	%ebp,%esi
+	addl	44(%rsp),%edx
+	vpxor	%xmm10,%xmm6,%xmm6
+	xorl	%ebx,%eax
+	shldl	$5,%ebp,%ebp
+	addl	%edi,%edx
+	andl	%eax,%esi
+	xorl	%ebx,%eax
+	addl	%ebp,%edx
+	shrdl	$7,%ebp,%ebp
+	xorl	%ebx,%esi
+	vpalignr	$8,%xmm3,%xmm4,%xmm7
+	movl	%edx,%edi
+	addl	48(%rsp),%ecx
+	vpaddd	%xmm6,%xmm11,%xmm9
+	xorl	%eax,%ebp
+	shldl	$5,%edx,%edx
+	vpsrldq	$4,%xmm6,%xmm8
+	addl	%esi,%ecx
+	andl	%ebp,%edi
+	vpxor	%xmm3,%xmm7,%xmm7
+	xorl	%eax,%ebp
+	addl	%edx,%ecx
+	vpxor	%xmm5,%xmm8,%xmm8
+	shrdl	$7,%edx,%edx
+	xorl	%eax,%edi
+	movl	%ecx,%esi
+	addl	52(%rsp),%ebx
+	vpxor	%xmm8,%xmm7,%xmm7
+	xorl	%ebp,%edx
+	shldl	$5,%ecx,%ecx
+	vmovdqa	%xmm9,32(%rsp)
+	addl	%edi,%ebx
+	andl	%edx,%esi
+	vpsrld	$31,%xmm7,%xmm8
+	xorl	%ebp,%edx
+	addl	%ecx,%ebx
+	shrdl	$7,%ecx,%ecx
+	xorl	%ebp,%esi
+	vpslldq	$12,%xmm7,%xmm10
+	vpaddd	%xmm7,%xmm7,%xmm7
+	movl	%ebx,%edi
+	addl	56(%rsp),%eax
+	xorl	%edx,%ecx
+	shldl	$5,%ebx,%ebx
+	vpsrld	$30,%xmm10,%xmm9
+	vpor	%xmm8,%xmm7,%xmm7
+	addl	%esi,%eax
+	andl	%ecx,%edi
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	vpslld	$2,%xmm10,%xmm10
+	vpxor	%xmm9,%xmm7,%xmm7
+	shrdl	$7,%ebx,%ebx
+	xorl	%edx,%edi
+	movl	%eax,%esi
+	addl	60(%rsp),%ebp
+	vpxor	%xmm10,%xmm7,%xmm7
+	xorl	%ecx,%ebx
+	shldl	$5,%eax,%eax
+	addl	%edi,%ebp
+	andl	%ebx,%esi
+	xorl	%ecx,%ebx
+	addl	%eax,%ebp
+	vpalignr	$8,%xmm6,%xmm7,%xmm8
+	vpxor	%xmm4,%xmm0,%xmm0
+	shrdl	$7,%eax,%eax
+	xorl	%ecx,%esi
+	movl	%ebp,%edi
+	addl	0(%rsp),%edx
+	vpxor	%xmm1,%xmm0,%xmm0
+	xorl	%ebx,%eax
+	shldl	$5,%ebp,%ebp
+	vpaddd	%xmm7,%xmm11,%xmm9
+	addl	%esi,%edx
+	andl	%eax,%edi
+	vpxor	%xmm8,%xmm0,%xmm0
+	xorl	%ebx,%eax
+	addl	%ebp,%edx
+	shrdl	$7,%ebp,%ebp
+	xorl	%ebx,%edi
+	vpsrld	$30,%xmm0,%xmm8
+	vmovdqa	%xmm9,48(%rsp)
+	movl	%edx,%esi
+	addl	4(%rsp),%ecx
+	xorl	%eax,%ebp
+	shldl	$5,%edx,%edx
+	vpslld	$2,%xmm0,%xmm0
+	addl	%edi,%ecx
+	andl	%ebp,%esi
+	xorl	%eax,%ebp
+	addl	%edx,%ecx
+	shrdl	$7,%edx,%edx
+	xorl	%eax,%esi
+	movl	%ecx,%edi
+	addl	8(%rsp),%ebx
+	vpor	%xmm8,%xmm0,%xmm0
+	xorl	%ebp,%edx
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	andl	%edx,%edi
+	xorl	%ebp,%edx
+	addl	%ecx,%ebx
+	addl	12(%rsp),%eax
+	xorl	%ebp,%edi
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%edi,%eax
+	xorl	%edx,%esi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	vpalignr	$8,%xmm7,%xmm0,%xmm8
+	vpxor	%xmm5,%xmm1,%xmm1
+	addl	16(%rsp),%ebp
+	xorl	%ecx,%esi
+	movl	%eax,%edi
+	shldl	$5,%eax,%eax
+	vpxor	%xmm2,%xmm1,%xmm1
+	addl	%esi,%ebp
+	xorl	%ecx,%edi
+	vpaddd	%xmm0,%xmm11,%xmm9
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%ebp
+	vpxor	%xmm8,%xmm1,%xmm1
+	addl	20(%rsp),%edx
+	xorl	%ebx,%edi
+	movl	%ebp,%esi
+	shldl	$5,%ebp,%ebp
+	vpsrld	$30,%xmm1,%xmm8
+	vmovdqa	%xmm9,0(%rsp)
+	addl	%edi,%edx
+	xorl	%ebx,%esi
+	shrdl	$7,%eax,%eax
+	addl	%ebp,%edx
+	vpslld	$2,%xmm1,%xmm1
+	addl	24(%rsp),%ecx
+	xorl	%eax,%esi
+	movl	%edx,%edi
+	shldl	$5,%edx,%edx
+	addl	%esi,%ecx
+	xorl	%eax,%edi
+	shrdl	$7,%ebp,%ebp
+	addl	%edx,%ecx
+	vpor	%xmm8,%xmm1,%xmm1
+	addl	28(%rsp),%ebx
+	xorl	%ebp,%edi
+	movl	%ecx,%esi
+	shldl	$5,%ecx,%ecx
+	addl	%edi,%ebx
+	xorl	%ebp,%esi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	vpalignr	$8,%xmm0,%xmm1,%xmm8
+	vpxor	%xmm6,%xmm2,%xmm2
+	addl	32(%rsp),%eax
+	xorl	%edx,%esi
+	movl	%ebx,%edi
+	shldl	$5,%ebx,%ebx
+	vpxor	%xmm3,%xmm2,%xmm2
+	addl	%esi,%eax
+	xorl	%edx,%edi
+	vpaddd	%xmm1,%xmm11,%xmm9
+	vmovdqa	0(%r11),%xmm11
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	vpxor	%xmm8,%xmm2,%xmm2
+	addl	36(%rsp),%ebp
+	xorl	%ecx,%edi
+	movl	%eax,%esi
+	shldl	$5,%eax,%eax
+	vpsrld	$30,%xmm2,%xmm8
+	vmovdqa	%xmm9,16(%rsp)
+	addl	%edi,%ebp
+	xorl	%ecx,%esi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%ebp
+	vpslld	$2,%xmm2,%xmm2
+	addl	40(%rsp),%edx
+	xorl	%ebx,%esi
+	movl	%ebp,%edi
+	shldl	$5,%ebp,%ebp
+	addl	%esi,%edx
+	xorl	%ebx,%edi
+	shrdl	$7,%eax,%eax
+	addl	%ebp,%edx
+	vpor	%xmm8,%xmm2,%xmm2
+	addl	44(%rsp),%ecx
+	xorl	%eax,%edi
+	movl	%edx,%esi
+	shldl	$5,%edx,%edx
+	addl	%edi,%ecx
+	xorl	%eax,%esi
+	shrdl	$7,%ebp,%ebp
+	addl	%edx,%ecx
+	vpalignr	$8,%xmm1,%xmm2,%xmm8
+	vpxor	%xmm7,%xmm3,%xmm3
+	addl	48(%rsp),%ebx
+	xorl	%ebp,%esi
+	movl	%ecx,%edi
+	shldl	$5,%ecx,%ecx
+	vpxor	%xmm4,%xmm3,%xmm3
+	addl	%esi,%ebx
+	xorl	%ebp,%edi
+	vpaddd	%xmm2,%xmm11,%xmm9
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	vpxor	%xmm8,%xmm3,%xmm3
+	addl	52(%rsp),%eax
+	xorl	%edx,%edi
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	vpsrld	$30,%xmm3,%xmm8
+	vmovdqa	%xmm9,32(%rsp)
+	addl	%edi,%eax
+	xorl	%edx,%esi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	vpslld	$2,%xmm3,%xmm3
+	addl	56(%rsp),%ebp
+	xorl	%ecx,%esi
+	movl	%eax,%edi
+	shldl	$5,%eax,%eax
+	addl	%esi,%ebp
+	xorl	%ecx,%edi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%ebp
+	vpor	%xmm8,%xmm3,%xmm3
+	addl	60(%rsp),%edx
+	xorl	%ebx,%edi
+	movl	%ebp,%esi
+	shldl	$5,%ebp,%ebp
+	addl	%edi,%edx
+	xorl	%ebx,%esi
+	shrdl	$7,%eax,%eax
+	addl	%ebp,%edx
+	vpalignr	$8,%xmm2,%xmm3,%xmm8
+	vpxor	%xmm0,%xmm4,%xmm4
+	addl	0(%rsp),%ecx
+	xorl	%eax,%esi
+	movl	%edx,%edi
+	shldl	$5,%edx,%edx
+	vpxor	%xmm5,%xmm4,%xmm4
+	addl	%esi,%ecx
+	xorl	%eax,%edi
+	vpaddd	%xmm3,%xmm11,%xmm9
+	shrdl	$7,%ebp,%ebp
+	addl	%edx,%ecx
+	vpxor	%xmm8,%xmm4,%xmm4
+	addl	4(%rsp),%ebx
+	xorl	%ebp,%edi
+	movl	%ecx,%esi
+	shldl	$5,%ecx,%ecx
+	vpsrld	$30,%xmm4,%xmm8
+	vmovdqa	%xmm9,48(%rsp)
+	addl	%edi,%ebx
+	xorl	%ebp,%esi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	vpslld	$2,%xmm4,%xmm4
+	addl	8(%rsp),%eax
+	xorl	%edx,%esi
+	movl	%ebx,%edi
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	xorl	%edx,%edi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	vpor	%xmm8,%xmm4,%xmm4
+	addl	12(%rsp),%ebp
+	xorl	%ecx,%edi
+	movl	%eax,%esi
+	shldl	$5,%eax,%eax
+	addl	%edi,%ebp
+	xorl	%ecx,%esi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%ebp
+	vpalignr	$8,%xmm3,%xmm4,%xmm8
+	vpxor	%xmm1,%xmm5,%xmm5
+	addl	16(%rsp),%edx
+	xorl	%ebx,%esi
+	movl	%ebp,%edi
+	shldl	$5,%ebp,%ebp
+	vpxor	%xmm6,%xmm5,%xmm5
+	addl	%esi,%edx
+	xorl	%ebx,%edi
+	vpaddd	%xmm4,%xmm11,%xmm9
+	shrdl	$7,%eax,%eax
+	addl	%ebp,%edx
+	vpxor	%xmm8,%xmm5,%xmm5
+	addl	20(%rsp),%ecx
+	xorl	%eax,%edi
+	movl	%edx,%esi
+	shldl	$5,%edx,%edx
+	vpsrld	$30,%xmm5,%xmm8
+	vmovdqa	%xmm9,0(%rsp)
+	addl	%edi,%ecx
+	xorl	%eax,%esi
+	shrdl	$7,%ebp,%ebp
+	addl	%edx,%ecx
+	vpslld	$2,%xmm5,%xmm5
+	addl	24(%rsp),%ebx
+	xorl	%ebp,%esi
+	movl	%ecx,%edi
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	xorl	%ebp,%edi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	vpor	%xmm8,%xmm5,%xmm5
+	addl	28(%rsp),%eax
+	shrdl	$7,%ecx,%ecx
+	movl	%ebx,%esi
+	xorl	%edx,%edi
+	shldl	$5,%ebx,%ebx
+	addl	%edi,%eax
+	xorl	%ecx,%esi
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	vpalignr	$8,%xmm4,%xmm5,%xmm8
+	vpxor	%xmm2,%xmm6,%xmm6
+	addl	32(%rsp),%ebp
+	andl	%ecx,%esi
+	xorl	%edx,%ecx
+	shrdl	$7,%ebx,%ebx
+	vpxor	%xmm7,%xmm6,%xmm6
+	movl	%eax,%edi
+	xorl	%ecx,%esi
+	vpaddd	%xmm5,%xmm11,%xmm9
+	shldl	$5,%eax,%eax
+	addl	%esi,%ebp
+	vpxor	%xmm8,%xmm6,%xmm6
+	xorl	%ebx,%edi
+	xorl	%ecx,%ebx
+	addl	%eax,%ebp
+	addl	36(%rsp),%edx
+	vpsrld	$30,%xmm6,%xmm8
+	vmovdqa	%xmm9,16(%rsp)
+	andl	%ebx,%edi
+	xorl	%ecx,%ebx
+	shrdl	$7,%eax,%eax
+	movl	%ebp,%esi
+	vpslld	$2,%xmm6,%xmm6
+	xorl	%ebx,%edi
+	shldl	$5,%ebp,%ebp
+	addl	%edi,%edx
+	xorl	%eax,%esi
+	xorl	%ebx,%eax
+	addl	%ebp,%edx
+	addl	40(%rsp),%ecx
+	andl	%eax,%esi
+	vpor	%xmm8,%xmm6,%xmm6
+	xorl	%ebx,%eax
+	shrdl	$7,%ebp,%ebp
+	movl	%edx,%edi
+	xorl	%eax,%esi
+	shldl	$5,%edx,%edx
+	addl	%esi,%ecx
+	xorl	%ebp,%edi
+	xorl	%eax,%ebp
+	addl	%edx,%ecx
+	addl	44(%rsp),%ebx
+	andl	%ebp,%edi
+	xorl	%eax,%ebp
+	shrdl	$7,%edx,%edx
+	movl	%ecx,%esi
+	xorl	%ebp,%edi
+	shldl	$5,%ecx,%ecx
+	addl	%edi,%ebx
+	xorl	%edx,%esi
+	xorl	%ebp,%edx
+	addl	%ecx,%ebx
+	vpalignr	$8,%xmm5,%xmm6,%xmm8
+	vpxor	%xmm3,%xmm7,%xmm7
+	addl	48(%rsp),%eax
+	andl	%edx,%esi
+	xorl	%ebp,%edx
+	shrdl	$7,%ecx,%ecx
+	vpxor	%xmm0,%xmm7,%xmm7
+	movl	%ebx,%edi
+	xorl	%edx,%esi
+	vpaddd	%xmm6,%xmm11,%xmm9
+	vmovdqa	32(%r11),%xmm11
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	vpxor	%xmm8,%xmm7,%xmm7
+	xorl	%ecx,%edi
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	addl	52(%rsp),%ebp
+	vpsrld	$30,%xmm7,%xmm8
+	vmovdqa	%xmm9,32(%rsp)
+	andl	%ecx,%edi
+	xorl	%edx,%ecx
+	shrdl	$7,%ebx,%ebx
+	movl	%eax,%esi
+	vpslld	$2,%xmm7,%xmm7
+	xorl	%ecx,%edi
+	shldl	$5,%eax,%eax
+	addl	%edi,%ebp
+	xorl	%ebx,%esi
+	xorl	%ecx,%ebx
+	addl	%eax,%ebp
+	addl	56(%rsp),%edx
+	andl	%ebx,%esi
+	vpor	%xmm8,%xmm7,%xmm7
+	xorl	%ecx,%ebx
+	shrdl	$7,%eax,%eax
+	movl	%ebp,%edi
+	xorl	%ebx,%esi
+	shldl	$5,%ebp,%ebp
+	addl	%esi,%edx
+	xorl	%eax,%edi
+	xorl	%ebx,%eax
+	addl	%ebp,%edx
+	addl	60(%rsp),%ecx
+	andl	%eax,%edi
+	xorl	%ebx,%eax
+	shrdl	$7,%ebp,%ebp
+	movl	%edx,%esi
+	xorl	%eax,%edi
+	shldl	$5,%edx,%edx
+	addl	%edi,%ecx
+	xorl	%ebp,%esi
+	xorl	%eax,%ebp
+	addl	%edx,%ecx
+	vpalignr	$8,%xmm6,%xmm7,%xmm8
+	vpxor	%xmm4,%xmm0,%xmm0
+	addl	0(%rsp),%ebx
+	andl	%ebp,%esi
+	xorl	%eax,%ebp
+	shrdl	$7,%edx,%edx
+	vpxor	%xmm1,%xmm0,%xmm0
+	movl	%ecx,%edi
+	xorl	%ebp,%esi
+	vpaddd	%xmm7,%xmm11,%xmm9
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	vpxor	%xmm8,%xmm0,%xmm0
+	xorl	%edx,%edi
+	xorl	%ebp,%edx
+	addl	%ecx,%ebx
+	addl	4(%rsp),%eax
+	vpsrld	$30,%xmm0,%xmm8
+	vmovdqa	%xmm9,48(%rsp)
+	andl	%edx,%edi
+	xorl	%ebp,%edx
+	shrdl	$7,%ecx,%ecx
+	movl	%ebx,%esi
+	vpslld	$2,%xmm0,%xmm0
+	xorl	%edx,%edi
+	shldl	$5,%ebx,%ebx
+	addl	%edi,%eax
+	xorl	%ecx,%esi
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	addl	8(%rsp),%ebp
+	andl	%ecx,%esi
+	vpor	%xmm8,%xmm0,%xmm0
+	xorl	%edx,%ecx
+	shrdl	$7,%ebx,%ebx
+	movl	%eax,%edi
+	xorl	%ecx,%esi
+	shldl	$5,%eax,%eax
+	addl	%esi,%ebp
+	xorl	%ebx,%edi
+	xorl	%ecx,%ebx
+	addl	%eax,%ebp
+	addl	12(%rsp),%edx
+	andl	%ebx,%edi
+	xorl	%ecx,%ebx
+	shrdl	$7,%eax,%eax
+	movl	%ebp,%esi
+	xorl	%ebx,%edi
+	shldl	$5,%ebp,%ebp
+	addl	%edi,%edx
+	xorl	%eax,%esi
+	xorl	%ebx,%eax
+	addl	%ebp,%edx
+	vpalignr	$8,%xmm7,%xmm0,%xmm8
+	vpxor	%xmm5,%xmm1,%xmm1
+	addl	16(%rsp),%ecx
+	andl	%eax,%esi
+	xorl	%ebx,%eax
+	shrdl	$7,%ebp,%ebp
+	vpxor	%xmm2,%xmm1,%xmm1
+	movl	%edx,%edi
+	xorl	%eax,%esi
+	vpaddd	%xmm0,%xmm11,%xmm9
+	shldl	$5,%edx,%edx
+	addl	%esi,%ecx
+	vpxor	%xmm8,%xmm1,%xmm1
+	xorl	%ebp,%edi
+	xorl	%eax,%ebp
+	addl	%edx,%ecx
+	addl	20(%rsp),%ebx
+	vpsrld	$30,%xmm1,%xmm8
+	vmovdqa	%xmm9,0(%rsp)
+	andl	%ebp,%edi
+	xorl	%eax,%ebp
+	shrdl	$7,%edx,%edx
+	movl	%ecx,%esi
+	vpslld	$2,%xmm1,%xmm1
+	xorl	%ebp,%edi
+	shldl	$5,%ecx,%ecx
+	addl	%edi,%ebx
+	xorl	%edx,%esi
+	xorl	%ebp,%edx
+	addl	%ecx,%ebx
+	addl	24(%rsp),%eax
+	andl	%edx,%esi
+	vpor	%xmm8,%xmm1,%xmm1
+	xorl	%ebp,%edx
+	shrdl	$7,%ecx,%ecx
+	movl	%ebx,%edi
+	xorl	%edx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	xorl	%ecx,%edi
+	xorl	%edx,%ecx
+	addl	%ebx,%eax
+	addl	28(%rsp),%ebp
+	andl	%ecx,%edi
+	xorl	%edx,%ecx
+	shrdl	$7,%ebx,%ebx
+	movl	%eax,%esi
+	xorl	%ecx,%edi
+	shldl	$5,%eax,%eax
+	addl	%edi,%ebp
+	xorl	%ebx,%esi
+	xorl	%ecx,%ebx
+	addl	%eax,%ebp
+	vpalignr	$8,%xmm0,%xmm1,%xmm8
+	vpxor	%xmm6,%xmm2,%xmm2
+	addl	32(%rsp),%edx
+	andl	%ebx,%esi
+	xorl	%ecx,%ebx
+	shrdl	$7,%eax,%eax
+	vpxor	%xmm3,%xmm2,%xmm2
+	movl	%ebp,%edi
+	xorl	%ebx,%esi
+	vpaddd	%xmm1,%xmm11,%xmm9
+	shldl	$5,%ebp,%ebp
+	addl	%esi,%edx
+	vpxor	%xmm8,%xmm2,%xmm2
+	xorl	%eax,%edi
+	xorl	%ebx,%eax
+	addl	%ebp,%edx
+	addl	36(%rsp),%ecx
+	vpsrld	$30,%xmm2,%xmm8
+	vmovdqa	%xmm9,16(%rsp)
+	andl	%eax,%edi
+	xorl	%ebx,%eax
+	shrdl	$7,%ebp,%ebp
+	movl	%edx,%esi
+	vpslld	$2,%xmm2,%xmm2
+	xorl	%eax,%edi
+	shldl	$5,%edx,%edx
+	addl	%edi,%ecx
+	xorl	%ebp,%esi
+	xorl	%eax,%ebp
+	addl	%edx,%ecx
+	addl	40(%rsp),%ebx
+	andl	%ebp,%esi
+	vpor	%xmm8,%xmm2,%xmm2
+	xorl	%eax,%ebp
+	shrdl	$7,%edx,%edx
+	movl	%ecx,%edi
+	xorl	%ebp,%esi
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	xorl	%edx,%edi
+	xorl	%ebp,%edx
+	addl	%ecx,%ebx
+	addl	44(%rsp),%eax
+	andl	%edx,%edi
+	xorl	%ebp,%edx
+	shrdl	$7,%ecx,%ecx
+	movl	%ebx,%esi
+	xorl	%edx,%edi
+	shldl	$5,%ebx,%ebx
+	addl	%edi,%eax
+	xorl	%edx,%esi
+	addl	%ebx,%eax
+	vpalignr	$8,%xmm1,%xmm2,%xmm8
+	vpxor	%xmm7,%xmm3,%xmm3
+	addl	48(%rsp),%ebp
+	xorl	%ecx,%esi
+	movl	%eax,%edi
+	shldl	$5,%eax,%eax
+	vpxor	%xmm4,%xmm3,%xmm3
+	addl	%esi,%ebp
+	xorl	%ecx,%edi
+	vpaddd	%xmm2,%xmm11,%xmm9
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%ebp
+	vpxor	%xmm8,%xmm3,%xmm3
+	addl	52(%rsp),%edx
+	xorl	%ebx,%edi
+	movl	%ebp,%esi
+	shldl	$5,%ebp,%ebp
+	vpsrld	$30,%xmm3,%xmm8
+	vmovdqa	%xmm9,32(%rsp)
+	addl	%edi,%edx
+	xorl	%ebx,%esi
+	shrdl	$7,%eax,%eax
+	addl	%ebp,%edx
+	vpslld	$2,%xmm3,%xmm3
+	addl	56(%rsp),%ecx
+	xorl	%eax,%esi
+	movl	%edx,%edi
+	shldl	$5,%edx,%edx
+	addl	%esi,%ecx
+	xorl	%eax,%edi
+	shrdl	$7,%ebp,%ebp
+	addl	%edx,%ecx
+	vpor	%xmm8,%xmm3,%xmm3
+	addl	60(%rsp),%ebx
+	xorl	%ebp,%edi
+	movl	%ecx,%esi
+	shldl	$5,%ecx,%ecx
+	addl	%edi,%ebx
+	xorl	%ebp,%esi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	0(%rsp),%eax
+	vpaddd	%xmm3,%xmm11,%xmm9
+	xorl	%edx,%esi
+	movl	%ebx,%edi
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	vmovdqa	%xmm9,48(%rsp)
+	xorl	%edx,%edi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	4(%rsp),%ebp
+	xorl	%ecx,%edi
+	movl	%eax,%esi
+	shldl	$5,%eax,%eax
+	addl	%edi,%ebp
+	xorl	%ecx,%esi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%ebp
+	addl	8(%rsp),%edx
+	xorl	%ebx,%esi
+	movl	%ebp,%edi
+	shldl	$5,%ebp,%ebp
+	addl	%esi,%edx
+	xorl	%ebx,%edi
+	shrdl	$7,%eax,%eax
+	addl	%ebp,%edx
+	addl	12(%rsp),%ecx
+	xorl	%eax,%edi
+	movl	%edx,%esi
+	shldl	$5,%edx,%edx
+	addl	%edi,%ecx
+	xorl	%eax,%esi
+	shrdl	$7,%ebp,%ebp
+	addl	%edx,%ecx
+	cmpq	%r10,%r9
+	je	L$done_avx
+	vmovdqa	64(%r11),%xmm6
+	vmovdqa	-64(%r11),%xmm11
+	vmovdqu	0(%r9),%xmm0
+	vmovdqu	16(%r9),%xmm1
+	vmovdqu	32(%r9),%xmm2
+	vmovdqu	48(%r9),%xmm3
+	vpshufb	%xmm6,%xmm0,%xmm0
+	addq	$64,%r9
+	addl	16(%rsp),%ebx
+	xorl	%ebp,%esi
+	vpshufb	%xmm6,%xmm1,%xmm1
+	movl	%ecx,%edi
+	shldl	$5,%ecx,%ecx
+	vpaddd	%xmm11,%xmm0,%xmm4
+	addl	%esi,%ebx
+	xorl	%ebp,%edi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	vmovdqa	%xmm4,0(%rsp)
+	addl	20(%rsp),%eax
+	xorl	%edx,%edi
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%edi,%eax
+	xorl	%edx,%esi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	24(%rsp),%ebp
+	xorl	%ecx,%esi
+	movl	%eax,%edi
+	shldl	$5,%eax,%eax
+	addl	%esi,%ebp
+	xorl	%ecx,%edi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%ebp
+	addl	28(%rsp),%edx
+	xorl	%ebx,%edi
+	movl	%ebp,%esi
+	shldl	$5,%ebp,%ebp
+	addl	%edi,%edx
+	xorl	%ebx,%esi
+	shrdl	$7,%eax,%eax
+	addl	%ebp,%edx
+	addl	32(%rsp),%ecx
+	xorl	%eax,%esi
+	vpshufb	%xmm6,%xmm2,%xmm2
+	movl	%edx,%edi
+	shldl	$5,%edx,%edx
+	vpaddd	%xmm11,%xmm1,%xmm5
+	addl	%esi,%ecx
+	xorl	%eax,%edi
+	shrdl	$7,%ebp,%ebp
+	addl	%edx,%ecx
+	vmovdqa	%xmm5,16(%rsp)
+	addl	36(%rsp),%ebx
+	xorl	%ebp,%edi
+	movl	%ecx,%esi
+	shldl	$5,%ecx,%ecx
+	addl	%edi,%ebx
+	xorl	%ebp,%esi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	40(%rsp),%eax
+	xorl	%edx,%esi
+	movl	%ebx,%edi
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	xorl	%edx,%edi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	44(%rsp),%ebp
+	xorl	%ecx,%edi
+	movl	%eax,%esi
+	shldl	$5,%eax,%eax
+	addl	%edi,%ebp
+	xorl	%ecx,%esi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%ebp
+	addl	48(%rsp),%edx
+	xorl	%ebx,%esi
+	vpshufb	%xmm6,%xmm3,%xmm3
+	movl	%ebp,%edi
+	shldl	$5,%ebp,%ebp
+	vpaddd	%xmm11,%xmm2,%xmm6
+	addl	%esi,%edx
+	xorl	%ebx,%edi
+	shrdl	$7,%eax,%eax
+	addl	%ebp,%edx
+	vmovdqa	%xmm6,32(%rsp)
+	addl	52(%rsp),%ecx
+	xorl	%eax,%edi
+	movl	%edx,%esi
+	shldl	$5,%edx,%edx
+	addl	%edi,%ecx
+	xorl	%eax,%esi
+	shrdl	$7,%ebp,%ebp
+	addl	%edx,%ecx
+	addl	56(%rsp),%ebx
+	xorl	%ebp,%esi
+	movl	%ecx,%edi
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	xorl	%ebp,%edi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	60(%rsp),%eax
+	xorl	%edx,%edi
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%edi,%eax
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	0(%r8),%eax
+	addl	4(%r8),%esi
+	addl	8(%r8),%ecx
+	addl	12(%r8),%edx
+	movl	%eax,0(%r8)
+	addl	16(%r8),%ebp
+	movl	%esi,4(%r8)
+	movl	%esi,%ebx
+	movl	%ecx,8(%r8)
+	movl	%ecx,%edi
+	movl	%edx,12(%r8)
+	xorl	%edx,%edi
+	movl	%ebp,16(%r8)
+	andl	%edi,%esi
+	jmp	L$oop_avx
+
+.p2align	4
+L$done_avx:
+	addl	16(%rsp),%ebx
+	xorl	%ebp,%esi
+	movl	%ecx,%edi
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	xorl	%ebp,%edi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	20(%rsp),%eax
+	xorl	%edx,%edi
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%edi,%eax
+	xorl	%edx,%esi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	24(%rsp),%ebp
+	xorl	%ecx,%esi
+	movl	%eax,%edi
+	shldl	$5,%eax,%eax
+	addl	%esi,%ebp
+	xorl	%ecx,%edi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%ebp
+	addl	28(%rsp),%edx
+	xorl	%ebx,%edi
+	movl	%ebp,%esi
+	shldl	$5,%ebp,%ebp
+	addl	%edi,%edx
+	xorl	%ebx,%esi
+	shrdl	$7,%eax,%eax
+	addl	%ebp,%edx
+	addl	32(%rsp),%ecx
+	xorl	%eax,%esi
+	movl	%edx,%edi
+	shldl	$5,%edx,%edx
+	addl	%esi,%ecx
+	xorl	%eax,%edi
+	shrdl	$7,%ebp,%ebp
+	addl	%edx,%ecx
+	addl	36(%rsp),%ebx
+	xorl	%ebp,%edi
+	movl	%ecx,%esi
+	shldl	$5,%ecx,%ecx
+	addl	%edi,%ebx
+	xorl	%ebp,%esi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	40(%rsp),%eax
+	xorl	%edx,%esi
+	movl	%ebx,%edi
+	shldl	$5,%ebx,%ebx
+	addl	%esi,%eax
+	xorl	%edx,%edi
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	addl	44(%rsp),%ebp
+	xorl	%ecx,%edi
+	movl	%eax,%esi
+	shldl	$5,%eax,%eax
+	addl	%edi,%ebp
+	xorl	%ecx,%esi
+	shrdl	$7,%ebx,%ebx
+	addl	%eax,%ebp
+	addl	48(%rsp),%edx
+	xorl	%ebx,%esi
+	movl	%ebp,%edi
+	shldl	$5,%ebp,%ebp
+	addl	%esi,%edx
+	xorl	%ebx,%edi
+	shrdl	$7,%eax,%eax
+	addl	%ebp,%edx
+	addl	52(%rsp),%ecx
+	xorl	%eax,%edi
+	movl	%edx,%esi
+	shldl	$5,%edx,%edx
+	addl	%edi,%ecx
+	xorl	%eax,%esi
+	shrdl	$7,%ebp,%ebp
+	addl	%edx,%ecx
+	addl	56(%rsp),%ebx
+	xorl	%ebp,%esi
+	movl	%ecx,%edi
+	shldl	$5,%ecx,%ecx
+	addl	%esi,%ebx
+	xorl	%ebp,%edi
+	shrdl	$7,%edx,%edx
+	addl	%ecx,%ebx
+	addl	60(%rsp),%eax
+	xorl	%edx,%edi
+	movl	%ebx,%esi
+	shldl	$5,%ebx,%ebx
+	addl	%edi,%eax
+	shrdl	$7,%ecx,%ecx
+	addl	%ebx,%eax
+	vzeroupper
+
+	addl	0(%r8),%eax
+	addl	4(%r8),%esi
+	addl	8(%r8),%ecx
+	movl	%eax,0(%r8)
+	addl	12(%r8),%edx
+	movl	%esi,4(%r8)
+	addl	16(%r8),%ebp
+	movl	%ecx,8(%r8)
+	movl	%edx,12(%r8)
+	movl	%ebp,16(%r8)
+	leaq	(%r14),%rsi
+	movq	-40(%rsi),%r14
+	movq	-32(%rsi),%r13
+	movq	-24(%rsi),%r12
+	movq	-16(%rsi),%rbp
+	movq	-8(%rsi),%rbx
+	leaq	(%rsi),%rsp
+L$epilogue_avx:
+	.byte	0xf3,0xc3
+
 .p2align	6
 K_XX_XX:
 .long	0x5a827999,0x5a827999,0x5a827999,0x5a827999