visaljs 0.0.2 → 0.0.3

package/runtime/third_party/boringssl/gen/bcm/sha256-x86_64-apple.S

@@ -0,0 +1,4173 @@
+// This file is generated from a similarly-named Perl script in the BoringSSL
+// source tree. Do not edit by hand.
+
+#include <openssl/asm_base.h>
+
+#if !defined(OPENSSL_NO_ASM) && defined(OPENSSL_X86_64) && defined(__APPLE__)
+.text	
+
+.globl	_sha256_block_data_order_nohw
+.private_extern _sha256_block_data_order_nohw
+
+.p2align	4
+_sha256_block_data_order_nohw:
+
+_CET_ENDBR
+	movq	%rsp,%rax  // copy %rsp
+
+	pushq	%rbx
+
+	pushq	%rbp
+
+	pushq	%r12
+
+	pushq	%r13
+
+	pushq	%r14
+
+	pushq	%r15
+
+	shlq	$4,%rdx  // num*16
+	subq	$64+32,%rsp
+	leaq	(%rsi,%rdx,4),%rdx  // inp+num*16*4
+	andq	$-64,%rsp  // align stack frame
+	movq	%rdi,64+0(%rsp)  // save ctx, 1st arg
+	movq	%rsi,64+8(%rsp)  // save inp, 2nd arh
+	movq	%rdx,64+16(%rsp)  // save end pointer, "3rd" arg
+	movq	%rax,88(%rsp)  // save copy of %rsp
+
+L$prologue:
+
+	movl	0(%rdi),%eax
+	movl	4(%rdi),%ebx
+	movl	8(%rdi),%ecx
+	movl	12(%rdi),%edx
+	movl	16(%rdi),%r8d
+	movl	20(%rdi),%r9d
+	movl	24(%rdi),%r10d
+	movl	28(%rdi),%r11d
+	jmp	L$loop
+
+.p2align	4
+L$loop:
+	movl	%ebx,%edi
+	leaq	K256(%rip),%rbp
+	xorl	%ecx,%edi  // magic
+	movl	0(%rsi),%r12d
+	movl	%r8d,%r13d
+	movl	%eax,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%r9d,%r15d
+
+	xorl	%r8d,%r13d
+	rorl	$9,%r14d
+	xorl	%r10d,%r15d  // f^g
+
+	movl	%r12d,0(%rsp)
+	xorl	%eax,%r14d
+	andl	%r8d,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r11d,%r12d  // T1+=h
+	xorl	%r10d,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r8d,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%eax,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%eax,%r14d
+
+	xorl	%ebx,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%ebx,%r11d
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%r11d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%edx  // d+=T1
+	addl	%r12d,%r11d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%r11d  // h+=Sigma0(a)
+	movl	4(%rsi),%r12d
+	movl	%edx,%r13d
+	movl	%r11d,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%r8d,%edi
+
+	xorl	%edx,%r13d
+	rorl	$9,%r14d
+	xorl	%r9d,%edi  // f^g
+
+	movl	%r12d,4(%rsp)
+	xorl	%r11d,%r14d
+	andl	%edx,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r10d,%r12d  // T1+=h
+	xorl	%r9d,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%edx,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r11d,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r11d,%r14d
+
+	xorl	%eax,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%eax,%r10d
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%r10d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%ecx  // d+=T1
+	addl	%r12d,%r10d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%r10d  // h+=Sigma0(a)
+	movl	8(%rsi),%r12d
+	movl	%ecx,%r13d
+	movl	%r10d,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%edx,%r15d
+
+	xorl	%ecx,%r13d
+	rorl	$9,%r14d
+	xorl	%r8d,%r15d  // f^g
+
+	movl	%r12d,8(%rsp)
+	xorl	%r10d,%r14d
+	andl	%ecx,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r9d,%r12d  // T1+=h
+	xorl	%r8d,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%ecx,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r10d,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r10d,%r14d
+
+	xorl	%r11d,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r11d,%r9d
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%r9d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%ebx  // d+=T1
+	addl	%r12d,%r9d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%r9d  // h+=Sigma0(a)
+	movl	12(%rsi),%r12d
+	movl	%ebx,%r13d
+	movl	%r9d,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%ecx,%edi
+
+	xorl	%ebx,%r13d
+	rorl	$9,%r14d
+	xorl	%edx,%edi  // f^g
+
+	movl	%r12d,12(%rsp)
+	xorl	%r9d,%r14d
+	andl	%ebx,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r8d,%r12d  // T1+=h
+	xorl	%edx,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%ebx,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r9d,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r9d,%r14d
+
+	xorl	%r10d,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r10d,%r8d
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%r8d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%eax  // d+=T1
+	addl	%r12d,%r8d  // h+=T1
+
+	leaq	20(%rbp),%rbp  // round++
+	addl	%r14d,%r8d  // h+=Sigma0(a)
+	movl	16(%rsi),%r12d
+	movl	%eax,%r13d
+	movl	%r8d,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%ebx,%r15d
+
+	xorl	%eax,%r13d
+	rorl	$9,%r14d
+	xorl	%ecx,%r15d  // f^g
+
+	movl	%r12d,16(%rsp)
+	xorl	%r8d,%r14d
+	andl	%eax,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%edx,%r12d  // T1+=h
+	xorl	%ecx,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%eax,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r8d,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r8d,%r14d
+
+	xorl	%r9d,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r9d,%edx
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%edx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r11d  // d+=T1
+	addl	%r12d,%edx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%edx  // h+=Sigma0(a)
+	movl	20(%rsi),%r12d
+	movl	%r11d,%r13d
+	movl	%edx,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%eax,%edi
+
+	xorl	%r11d,%r13d
+	rorl	$9,%r14d
+	xorl	%ebx,%edi  // f^g
+
+	movl	%r12d,20(%rsp)
+	xorl	%edx,%r14d
+	andl	%r11d,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%ecx,%r12d  // T1+=h
+	xorl	%ebx,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r11d,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%edx,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%edx,%r14d
+
+	xorl	%r8d,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r8d,%ecx
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%ecx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r10d  // d+=T1
+	addl	%r12d,%ecx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%ecx  // h+=Sigma0(a)
+	movl	24(%rsi),%r12d
+	movl	%r10d,%r13d
+	movl	%ecx,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%r11d,%r15d
+
+	xorl	%r10d,%r13d
+	rorl	$9,%r14d
+	xorl	%eax,%r15d  // f^g
+
+	movl	%r12d,24(%rsp)
+	xorl	%ecx,%r14d
+	andl	%r10d,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%ebx,%r12d  // T1+=h
+	xorl	%eax,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r10d,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%ecx,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%ecx,%r14d
+
+	xorl	%edx,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%edx,%ebx
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%ebx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r9d  // d+=T1
+	addl	%r12d,%ebx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%ebx  // h+=Sigma0(a)
+	movl	28(%rsi),%r12d
+	movl	%r9d,%r13d
+	movl	%ebx,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%r10d,%edi
+
+	xorl	%r9d,%r13d
+	rorl	$9,%r14d
+	xorl	%r11d,%edi  // f^g
+
+	movl	%r12d,28(%rsp)
+	xorl	%ebx,%r14d
+	andl	%r9d,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%eax,%r12d  // T1+=h
+	xorl	%r11d,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r9d,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%ebx,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%ebx,%r14d
+
+	xorl	%ecx,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%ecx,%eax
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%eax  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r8d  // d+=T1
+	addl	%r12d,%eax  // h+=T1
+
+	leaq	20(%rbp),%rbp  // round++
+	addl	%r14d,%eax  // h+=Sigma0(a)
+	movl	32(%rsi),%r12d
+	movl	%r8d,%r13d
+	movl	%eax,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%r9d,%r15d
+
+	xorl	%r8d,%r13d
+	rorl	$9,%r14d
+	xorl	%r10d,%r15d  // f^g
+
+	movl	%r12d,32(%rsp)
+	xorl	%eax,%r14d
+	andl	%r8d,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r11d,%r12d  // T1+=h
+	xorl	%r10d,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r8d,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%eax,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%eax,%r14d
+
+	xorl	%ebx,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%ebx,%r11d
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%r11d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%edx  // d+=T1
+	addl	%r12d,%r11d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%r11d  // h+=Sigma0(a)
+	movl	36(%rsi),%r12d
+	movl	%edx,%r13d
+	movl	%r11d,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%r8d,%edi
+
+	xorl	%edx,%r13d
+	rorl	$9,%r14d
+	xorl	%r9d,%edi  // f^g
+
+	movl	%r12d,36(%rsp)
+	xorl	%r11d,%r14d
+	andl	%edx,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r10d,%r12d  // T1+=h
+	xorl	%r9d,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%edx,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r11d,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r11d,%r14d
+
+	xorl	%eax,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%eax,%r10d
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%r10d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%ecx  // d+=T1
+	addl	%r12d,%r10d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%r10d  // h+=Sigma0(a)
+	movl	40(%rsi),%r12d
+	movl	%ecx,%r13d
+	movl	%r10d,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%edx,%r15d
+
+	xorl	%ecx,%r13d
+	rorl	$9,%r14d
+	xorl	%r8d,%r15d  // f^g
+
+	movl	%r12d,40(%rsp)
+	xorl	%r10d,%r14d
+	andl	%ecx,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r9d,%r12d  // T1+=h
+	xorl	%r8d,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%ecx,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r10d,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r10d,%r14d
+
+	xorl	%r11d,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r11d,%r9d
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%r9d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%ebx  // d+=T1
+	addl	%r12d,%r9d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%r9d  // h+=Sigma0(a)
+	movl	44(%rsi),%r12d
+	movl	%ebx,%r13d
+	movl	%r9d,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%ecx,%edi
+
+	xorl	%ebx,%r13d
+	rorl	$9,%r14d
+	xorl	%edx,%edi  // f^g
+
+	movl	%r12d,44(%rsp)
+	xorl	%r9d,%r14d
+	andl	%ebx,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r8d,%r12d  // T1+=h
+	xorl	%edx,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%ebx,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r9d,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r9d,%r14d
+
+	xorl	%r10d,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r10d,%r8d
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%r8d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%eax  // d+=T1
+	addl	%r12d,%r8d  // h+=T1
+
+	leaq	20(%rbp),%rbp  // round++
+	addl	%r14d,%r8d  // h+=Sigma0(a)
+	movl	48(%rsi),%r12d
+	movl	%eax,%r13d
+	movl	%r8d,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%ebx,%r15d
+
+	xorl	%eax,%r13d
+	rorl	$9,%r14d
+	xorl	%ecx,%r15d  // f^g
+
+	movl	%r12d,48(%rsp)
+	xorl	%r8d,%r14d
+	andl	%eax,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%edx,%r12d  // T1+=h
+	xorl	%ecx,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%eax,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r8d,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r8d,%r14d
+
+	xorl	%r9d,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r9d,%edx
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%edx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r11d  // d+=T1
+	addl	%r12d,%edx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%edx  // h+=Sigma0(a)
+	movl	52(%rsi),%r12d
+	movl	%r11d,%r13d
+	movl	%edx,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%eax,%edi
+
+	xorl	%r11d,%r13d
+	rorl	$9,%r14d
+	xorl	%ebx,%edi  // f^g
+
+	movl	%r12d,52(%rsp)
+	xorl	%edx,%r14d
+	andl	%r11d,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%ecx,%r12d  // T1+=h
+	xorl	%ebx,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r11d,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%edx,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%edx,%r14d
+
+	xorl	%r8d,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r8d,%ecx
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%ecx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r10d  // d+=T1
+	addl	%r12d,%ecx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%ecx  // h+=Sigma0(a)
+	movl	56(%rsi),%r12d
+	movl	%r10d,%r13d
+	movl	%ecx,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%r11d,%r15d
+
+	xorl	%r10d,%r13d
+	rorl	$9,%r14d
+	xorl	%eax,%r15d  // f^g
+
+	movl	%r12d,56(%rsp)
+	xorl	%ecx,%r14d
+	andl	%r10d,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%ebx,%r12d  // T1+=h
+	xorl	%eax,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r10d,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%ecx,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%ecx,%r14d
+
+	xorl	%edx,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%edx,%ebx
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%ebx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r9d  // d+=T1
+	addl	%r12d,%ebx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	addl	%r14d,%ebx  // h+=Sigma0(a)
+	movl	60(%rsi),%r12d
+	movl	%r9d,%r13d
+	movl	%ebx,%r14d
+	bswapl	%r12d
+	rorl	$14,%r13d
+	movl	%r10d,%edi
+
+	xorl	%r9d,%r13d
+	rorl	$9,%r14d
+	xorl	%r11d,%edi  // f^g
+
+	movl	%r12d,60(%rsp)
+	xorl	%ebx,%r14d
+	andl	%r9d,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%eax,%r12d  // T1+=h
+	xorl	%r11d,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r9d,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%ebx,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%ebx,%r14d
+
+	xorl	%ecx,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%ecx,%eax
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%eax  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r8d  // d+=T1
+	addl	%r12d,%eax  // h+=T1
+
+	leaq	20(%rbp),%rbp  // round++
+	jmp	L$rounds_16_xx
+.p2align	4
+L$rounds_16_xx:
+	movl	4(%rsp),%r13d
+	movl	56(%rsp),%r15d
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%eax  // modulo-scheduled h+=Sigma0(a)
+	movl	%r15d,%r14d
+	rorl	$2,%r15d
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%r15d
+	shrl	$10,%r14d
+
+	rorl	$17,%r15d
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%r15d  // sigma1(X[(i+14)&0xf])
+	addl	36(%rsp),%r12d
+
+	addl	0(%rsp),%r12d
+	movl	%r8d,%r13d
+	addl	%r15d,%r12d
+	movl	%eax,%r14d
+	rorl	$14,%r13d
+	movl	%r9d,%r15d
+
+	xorl	%r8d,%r13d
+	rorl	$9,%r14d
+	xorl	%r10d,%r15d  // f^g
+
+	movl	%r12d,0(%rsp)
+	xorl	%eax,%r14d
+	andl	%r8d,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r11d,%r12d  // T1+=h
+	xorl	%r10d,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r8d,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%eax,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%eax,%r14d
+
+	xorl	%ebx,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%ebx,%r11d
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%r11d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%edx  // d+=T1
+	addl	%r12d,%r11d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	8(%rsp),%r13d
+	movl	60(%rsp),%edi
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%r11d  // modulo-scheduled h+=Sigma0(a)
+	movl	%edi,%r14d
+	rorl	$2,%edi
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%edi
+	shrl	$10,%r14d
+
+	rorl	$17,%edi
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%edi  // sigma1(X[(i+14)&0xf])
+	addl	40(%rsp),%r12d
+
+	addl	4(%rsp),%r12d
+	movl	%edx,%r13d
+	addl	%edi,%r12d
+	movl	%r11d,%r14d
+	rorl	$14,%r13d
+	movl	%r8d,%edi
+
+	xorl	%edx,%r13d
+	rorl	$9,%r14d
+	xorl	%r9d,%edi  // f^g
+
+	movl	%r12d,4(%rsp)
+	xorl	%r11d,%r14d
+	andl	%edx,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r10d,%r12d  // T1+=h
+	xorl	%r9d,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%edx,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r11d,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r11d,%r14d
+
+	xorl	%eax,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%eax,%r10d
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%r10d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%ecx  // d+=T1
+	addl	%r12d,%r10d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	12(%rsp),%r13d
+	movl	0(%rsp),%r15d
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%r10d  // modulo-scheduled h+=Sigma0(a)
+	movl	%r15d,%r14d
+	rorl	$2,%r15d
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%r15d
+	shrl	$10,%r14d
+
+	rorl	$17,%r15d
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%r15d  // sigma1(X[(i+14)&0xf])
+	addl	44(%rsp),%r12d
+
+	addl	8(%rsp),%r12d
+	movl	%ecx,%r13d
+	addl	%r15d,%r12d
+	movl	%r10d,%r14d
+	rorl	$14,%r13d
+	movl	%edx,%r15d
+
+	xorl	%ecx,%r13d
+	rorl	$9,%r14d
+	xorl	%r8d,%r15d  // f^g
+
+	movl	%r12d,8(%rsp)
+	xorl	%r10d,%r14d
+	andl	%ecx,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r9d,%r12d  // T1+=h
+	xorl	%r8d,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%ecx,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r10d,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r10d,%r14d
+
+	xorl	%r11d,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r11d,%r9d
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%r9d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%ebx  // d+=T1
+	addl	%r12d,%r9d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	16(%rsp),%r13d
+	movl	4(%rsp),%edi
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%r9d  // modulo-scheduled h+=Sigma0(a)
+	movl	%edi,%r14d
+	rorl	$2,%edi
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%edi
+	shrl	$10,%r14d
+
+	rorl	$17,%edi
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%edi  // sigma1(X[(i+14)&0xf])
+	addl	48(%rsp),%r12d
+
+	addl	12(%rsp),%r12d
+	movl	%ebx,%r13d
+	addl	%edi,%r12d
+	movl	%r9d,%r14d
+	rorl	$14,%r13d
+	movl	%ecx,%edi
+
+	xorl	%ebx,%r13d
+	rorl	$9,%r14d
+	xorl	%edx,%edi  // f^g
+
+	movl	%r12d,12(%rsp)
+	xorl	%r9d,%r14d
+	andl	%ebx,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r8d,%r12d  // T1+=h
+	xorl	%edx,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%ebx,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r9d,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r9d,%r14d
+
+	xorl	%r10d,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r10d,%r8d
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%r8d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%eax  // d+=T1
+	addl	%r12d,%r8d  // h+=T1
+
+	leaq	20(%rbp),%rbp  // round++
+	movl	20(%rsp),%r13d
+	movl	8(%rsp),%r15d
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%r8d  // modulo-scheduled h+=Sigma0(a)
+	movl	%r15d,%r14d
+	rorl	$2,%r15d
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%r15d
+	shrl	$10,%r14d
+
+	rorl	$17,%r15d
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%r15d  // sigma1(X[(i+14)&0xf])
+	addl	52(%rsp),%r12d
+
+	addl	16(%rsp),%r12d
+	movl	%eax,%r13d
+	addl	%r15d,%r12d
+	movl	%r8d,%r14d
+	rorl	$14,%r13d
+	movl	%ebx,%r15d
+
+	xorl	%eax,%r13d
+	rorl	$9,%r14d
+	xorl	%ecx,%r15d  // f^g
+
+	movl	%r12d,16(%rsp)
+	xorl	%r8d,%r14d
+	andl	%eax,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%edx,%r12d  // T1+=h
+	xorl	%ecx,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%eax,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r8d,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r8d,%r14d
+
+	xorl	%r9d,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r9d,%edx
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%edx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r11d  // d+=T1
+	addl	%r12d,%edx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	24(%rsp),%r13d
+	movl	12(%rsp),%edi
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%edx  // modulo-scheduled h+=Sigma0(a)
+	movl	%edi,%r14d
+	rorl	$2,%edi
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%edi
+	shrl	$10,%r14d
+
+	rorl	$17,%edi
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%edi  // sigma1(X[(i+14)&0xf])
+	addl	56(%rsp),%r12d
+
+	addl	20(%rsp),%r12d
+	movl	%r11d,%r13d
+	addl	%edi,%r12d
+	movl	%edx,%r14d
+	rorl	$14,%r13d
+	movl	%eax,%edi
+
+	xorl	%r11d,%r13d
+	rorl	$9,%r14d
+	xorl	%ebx,%edi  // f^g
+
+	movl	%r12d,20(%rsp)
+	xorl	%edx,%r14d
+	andl	%r11d,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%ecx,%r12d  // T1+=h
+	xorl	%ebx,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r11d,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%edx,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%edx,%r14d
+
+	xorl	%r8d,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r8d,%ecx
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%ecx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r10d  // d+=T1
+	addl	%r12d,%ecx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	28(%rsp),%r13d
+	movl	16(%rsp),%r15d
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%ecx  // modulo-scheduled h+=Sigma0(a)
+	movl	%r15d,%r14d
+	rorl	$2,%r15d
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%r15d
+	shrl	$10,%r14d
+
+	rorl	$17,%r15d
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%r15d  // sigma1(X[(i+14)&0xf])
+	addl	60(%rsp),%r12d
+
+	addl	24(%rsp),%r12d
+	movl	%r10d,%r13d
+	addl	%r15d,%r12d
+	movl	%ecx,%r14d
+	rorl	$14,%r13d
+	movl	%r11d,%r15d
+
+	xorl	%r10d,%r13d
+	rorl	$9,%r14d
+	xorl	%eax,%r15d  // f^g
+
+	movl	%r12d,24(%rsp)
+	xorl	%ecx,%r14d
+	andl	%r10d,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%ebx,%r12d  // T1+=h
+	xorl	%eax,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r10d,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%ecx,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%ecx,%r14d
+
+	xorl	%edx,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%edx,%ebx
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%ebx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r9d  // d+=T1
+	addl	%r12d,%ebx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	32(%rsp),%r13d
+	movl	20(%rsp),%edi
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%ebx  // modulo-scheduled h+=Sigma0(a)
+	movl	%edi,%r14d
+	rorl	$2,%edi
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%edi
+	shrl	$10,%r14d
+
+	rorl	$17,%edi
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%edi  // sigma1(X[(i+14)&0xf])
+	addl	0(%rsp),%r12d
+
+	addl	28(%rsp),%r12d
+	movl	%r9d,%r13d
+	addl	%edi,%r12d
+	movl	%ebx,%r14d
+	rorl	$14,%r13d
+	movl	%r10d,%edi
+
+	xorl	%r9d,%r13d
+	rorl	$9,%r14d
+	xorl	%r11d,%edi  // f^g
+
+	movl	%r12d,28(%rsp)
+	xorl	%ebx,%r14d
+	andl	%r9d,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%eax,%r12d  // T1+=h
+	xorl	%r11d,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r9d,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%ebx,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%ebx,%r14d
+
+	xorl	%ecx,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%ecx,%eax
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%eax  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r8d  // d+=T1
+	addl	%r12d,%eax  // h+=T1
+
+	leaq	20(%rbp),%rbp  // round++
+	movl	36(%rsp),%r13d
+	movl	24(%rsp),%r15d
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%eax  // modulo-scheduled h+=Sigma0(a)
+	movl	%r15d,%r14d
+	rorl	$2,%r15d
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%r15d
+	shrl	$10,%r14d
+
+	rorl	$17,%r15d
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%r15d  // sigma1(X[(i+14)&0xf])
+	addl	4(%rsp),%r12d
+
+	addl	32(%rsp),%r12d
+	movl	%r8d,%r13d
+	addl	%r15d,%r12d
+	movl	%eax,%r14d
+	rorl	$14,%r13d
+	movl	%r9d,%r15d
+
+	xorl	%r8d,%r13d
+	rorl	$9,%r14d
+	xorl	%r10d,%r15d  // f^g
+
+	movl	%r12d,32(%rsp)
+	xorl	%eax,%r14d
+	andl	%r8d,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r11d,%r12d  // T1+=h
+	xorl	%r10d,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r8d,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%eax,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%eax,%r14d
+
+	xorl	%ebx,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%ebx,%r11d
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%r11d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%edx  // d+=T1
+	addl	%r12d,%r11d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	40(%rsp),%r13d
+	movl	28(%rsp),%edi
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%r11d  // modulo-scheduled h+=Sigma0(a)
+	movl	%edi,%r14d
+	rorl	$2,%edi
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%edi
+	shrl	$10,%r14d
+
+	rorl	$17,%edi
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%edi  // sigma1(X[(i+14)&0xf])
+	addl	8(%rsp),%r12d
+
+	addl	36(%rsp),%r12d
+	movl	%edx,%r13d
+	addl	%edi,%r12d
+	movl	%r11d,%r14d
+	rorl	$14,%r13d
+	movl	%r8d,%edi
+
+	xorl	%edx,%r13d
+	rorl	$9,%r14d
+	xorl	%r9d,%edi  // f^g
+
+	movl	%r12d,36(%rsp)
+	xorl	%r11d,%r14d
+	andl	%edx,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r10d,%r12d  // T1+=h
+	xorl	%r9d,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%edx,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r11d,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r11d,%r14d
+
+	xorl	%eax,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%eax,%r10d
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%r10d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%ecx  // d+=T1
+	addl	%r12d,%r10d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	44(%rsp),%r13d
+	movl	32(%rsp),%r15d
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%r10d  // modulo-scheduled h+=Sigma0(a)
+	movl	%r15d,%r14d
+	rorl	$2,%r15d
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%r15d
+	shrl	$10,%r14d
+
+	rorl	$17,%r15d
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%r15d  // sigma1(X[(i+14)&0xf])
+	addl	12(%rsp),%r12d
+
+	addl	40(%rsp),%r12d
+	movl	%ecx,%r13d
+	addl	%r15d,%r12d
+	movl	%r10d,%r14d
+	rorl	$14,%r13d
+	movl	%edx,%r15d
+
+	xorl	%ecx,%r13d
+	rorl	$9,%r14d
+	xorl	%r8d,%r15d  // f^g
+
+	movl	%r12d,40(%rsp)
+	xorl	%r10d,%r14d
+	andl	%ecx,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r9d,%r12d  // T1+=h
+	xorl	%r8d,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%ecx,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r10d,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r10d,%r14d
+
+	xorl	%r11d,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r11d,%r9d
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%r9d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%ebx  // d+=T1
+	addl	%r12d,%r9d  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	48(%rsp),%r13d
+	movl	36(%rsp),%edi
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%r9d  // modulo-scheduled h+=Sigma0(a)
+	movl	%edi,%r14d
+	rorl	$2,%edi
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%edi
+	shrl	$10,%r14d
+
+	rorl	$17,%edi
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%edi  // sigma1(X[(i+14)&0xf])
+	addl	16(%rsp),%r12d
+
+	addl	44(%rsp),%r12d
+	movl	%ebx,%r13d
+	addl	%edi,%r12d
+	movl	%r9d,%r14d
+	rorl	$14,%r13d
+	movl	%ecx,%edi
+
+	xorl	%ebx,%r13d
+	rorl	$9,%r14d
+	xorl	%edx,%edi  // f^g
+
+	movl	%r12d,44(%rsp)
+	xorl	%r9d,%r14d
+	andl	%ebx,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%r8d,%r12d  // T1+=h
+	xorl	%edx,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%ebx,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r9d,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r9d,%r14d
+
+	xorl	%r10d,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r10d,%r8d
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%r8d  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%eax  // d+=T1
+	addl	%r12d,%r8d  // h+=T1
+
+	leaq	20(%rbp),%rbp  // round++
+	movl	52(%rsp),%r13d
+	movl	40(%rsp),%r15d
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%r8d  // modulo-scheduled h+=Sigma0(a)
+	movl	%r15d,%r14d
+	rorl	$2,%r15d
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%r15d
+	shrl	$10,%r14d
+
+	rorl	$17,%r15d
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%r15d  // sigma1(X[(i+14)&0xf])
+	addl	20(%rsp),%r12d
+
+	addl	48(%rsp),%r12d
+	movl	%eax,%r13d
+	addl	%r15d,%r12d
+	movl	%r8d,%r14d
+	rorl	$14,%r13d
+	movl	%ebx,%r15d
+
+	xorl	%eax,%r13d
+	rorl	$9,%r14d
+	xorl	%ecx,%r15d  // f^g
+
+	movl	%r12d,48(%rsp)
+	xorl	%r8d,%r14d
+	andl	%eax,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%edx,%r12d  // T1+=h
+	xorl	%ecx,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%eax,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%r8d,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%r8d,%r14d
+
+	xorl	%r9d,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r9d,%edx
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%edx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r11d  // d+=T1
+	addl	%r12d,%edx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	56(%rsp),%r13d
+	movl	44(%rsp),%edi
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%edx  // modulo-scheduled h+=Sigma0(a)
+	movl	%edi,%r14d
+	rorl	$2,%edi
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%edi
+	shrl	$10,%r14d
+
+	rorl	$17,%edi
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%edi  // sigma1(X[(i+14)&0xf])
+	addl	24(%rsp),%r12d
+
+	addl	52(%rsp),%r12d
+	movl	%r11d,%r13d
+	addl	%edi,%r12d
+	movl	%edx,%r14d
+	rorl	$14,%r13d
+	movl	%eax,%edi
+
+	xorl	%r11d,%r13d
+	rorl	$9,%r14d
+	xorl	%ebx,%edi  // f^g
+
+	movl	%r12d,52(%rsp)
+	xorl	%edx,%r14d
+	andl	%r11d,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%ecx,%r12d  // T1+=h
+	xorl	%ebx,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r11d,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%edx,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%edx,%r14d
+
+	xorl	%r8d,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%r8d,%ecx
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%ecx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r10d  // d+=T1
+	addl	%r12d,%ecx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	60(%rsp),%r13d
+	movl	48(%rsp),%r15d
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%ecx  // modulo-scheduled h+=Sigma0(a)
+	movl	%r15d,%r14d
+	rorl	$2,%r15d
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%r15d
+	shrl	$10,%r14d
+
+	rorl	$17,%r15d
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%r15d  // sigma1(X[(i+14)&0xf])
+	addl	28(%rsp),%r12d
+
+	addl	56(%rsp),%r12d
+	movl	%r10d,%r13d
+	addl	%r15d,%r12d
+	movl	%ecx,%r14d
+	rorl	$14,%r13d
+	movl	%r11d,%r15d
+
+	xorl	%r10d,%r13d
+	rorl	$9,%r14d
+	xorl	%eax,%r15d  // f^g
+
+	movl	%r12d,56(%rsp)
+	xorl	%ecx,%r14d
+	andl	%r10d,%r15d  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%ebx,%r12d  // T1+=h
+	xorl	%eax,%r15d  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r10d,%r13d
+	addl	%r15d,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%ecx,%r15d
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%ecx,%r14d
+
+	xorl	%edx,%r15d  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%edx,%ebx
+
+	andl	%r15d,%edi
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%edi,%ebx  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r9d  // d+=T1
+	addl	%r12d,%ebx  // h+=T1
+
+	leaq	4(%rbp),%rbp  // round++
+	movl	0(%rsp),%r13d
+	movl	52(%rsp),%edi
+
+	movl	%r13d,%r12d
+	rorl	$11,%r13d
+	addl	%r14d,%ebx  // modulo-scheduled h+=Sigma0(a)
+	movl	%edi,%r14d
+	rorl	$2,%edi
+
+	xorl	%r12d,%r13d
+	shrl	$3,%r12d
+	rorl	$7,%r13d
+	xorl	%r14d,%edi
+	shrl	$10,%r14d
+
+	rorl	$17,%edi
+	xorl	%r13d,%r12d  // sigma0(X[(i+1)&0xf])
+	xorl	%r14d,%edi  // sigma1(X[(i+14)&0xf])
+	addl	32(%rsp),%r12d
+
+	addl	60(%rsp),%r12d
+	movl	%r9d,%r13d
+	addl	%edi,%r12d
+	movl	%ebx,%r14d
+	rorl	$14,%r13d
+	movl	%r10d,%edi
+
+	xorl	%r9d,%r13d
+	rorl	$9,%r14d
+	xorl	%r11d,%edi  // f^g
+
+	movl	%r12d,60(%rsp)
+	xorl	%ebx,%r14d
+	andl	%r9d,%edi  // (f^g)&e
+
+	rorl	$5,%r13d
+	addl	%eax,%r12d  // T1+=h
+	xorl	%r11d,%edi  // Ch(e,f,g)=((f^g)&e)^g
+
+	rorl	$11,%r14d
+	xorl	%r9d,%r13d
+	addl	%edi,%r12d  // T1+=Ch(e,f,g)
+
+	movl	%ebx,%edi
+	addl	(%rbp),%r12d  // T1+=K[round]
+	xorl	%ebx,%r14d
+
+	xorl	%ecx,%edi  // a^b, b^c in next round
+	rorl	$6,%r13d  // Sigma1(e)
+	movl	%ecx,%eax
+
+	andl	%edi,%r15d
+	rorl	$2,%r14d  // Sigma0(a)
+	addl	%r13d,%r12d  // T1+=Sigma1(e)
+
+	xorl	%r15d,%eax  // h=Maj(a,b,c)=Ch(a^b,c,b)
+	addl	%r12d,%r8d  // d+=T1
+	addl	%r12d,%eax  // h+=T1
+
+	leaq	20(%rbp),%rbp  // round++
+	cmpb	$0,3(%rbp)
+	jnz	L$rounds_16_xx
+
+	movq	64+0(%rsp),%rdi
+	addl	%r14d,%eax  // modulo-scheduled h+=Sigma0(a)
+	leaq	64(%rsi),%rsi
+
+	addl	0(%rdi),%eax
+	addl	4(%rdi),%ebx
+	addl	8(%rdi),%ecx
+	addl	12(%rdi),%edx
+	addl	16(%rdi),%r8d
+	addl	20(%rdi),%r9d
+	addl	24(%rdi),%r10d
+	addl	28(%rdi),%r11d
+
+	cmpq	64+16(%rsp),%rsi
+
+	movl	%eax,0(%rdi)
+	movl	%ebx,4(%rdi)
+	movl	%ecx,8(%rdi)
+	movl	%edx,12(%rdi)
+	movl	%r8d,16(%rdi)
+	movl	%r9d,20(%rdi)
+	movl	%r10d,24(%rdi)
+	movl	%r11d,28(%rdi)
+	jb	L$loop
+
+	movq	88(%rsp),%rsi
+
+	movq	-48(%rsi),%r15
+
+	movq	-40(%rsi),%r14
+
+	movq	-32(%rsi),%r13
+
+	movq	-24(%rsi),%r12
+
+	movq	-16(%rsi),%rbp
+
+	movq	-8(%rsi),%rbx
+
+	leaq	(%rsi),%rsp
+
+L$epilogue:
+	ret
+
+
+.section	__DATA,__const
+.p2align	6
+
+K256:
+.long	0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
+.long	0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
+.long	0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
+.long	0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
+.long	0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
+.long	0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
+.long	0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
+.long	0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
+.long	0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
+.long	0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
+.long	0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
+.long	0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
+.long	0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
+.long	0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
+.long	0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
+.long	0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
+.long	0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
+.long	0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
+.long	0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
+.long	0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
+.long	0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
+.long	0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
+.long	0xd192e819,0xd6990624,0xf40e3585,0x106aa070
+.long	0xd192e819,0xd6990624,0xf40e3585,0x106aa070
+.long	0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
+.long	0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
+.long	0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
+.long	0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
+.long	0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
+.long	0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
+.long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
+.long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
+
+.long	0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
+.long	0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
+.long	0x03020100,0x0b0a0908,0xffffffff,0xffffffff
+.long	0x03020100,0x0b0a0908,0xffffffff,0xffffffff
+.long	0xffffffff,0xffffffff,0x03020100,0x0b0a0908
+.long	0xffffffff,0xffffffff,0x03020100,0x0b0a0908
+.byte	83,72,65,50,53,54,32,98,108,111,99,107,32,116,114,97,110,115,102,111,114,109,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
+.text	
+.globl	_sha256_block_data_order_hw
+.private_extern _sha256_block_data_order_hw
+
+.p2align	6
+.alt_entry _sha256_block_data_order_hw
+_sha256_block_data_order_hw:
+
+_CET_ENDBR
+	leaq	K256+128(%rip),%rcx
+	movdqu	(%rdi),%xmm1  // DCBA
+	movdqu	16(%rdi),%xmm2  // HGFE
+	movdqa	512-128(%rcx),%xmm7  // byte swap mask
+
+	pshufd	$0x1b,%xmm1,%xmm0  // ABCD
+	pshufd	$0xb1,%xmm1,%xmm1  // CDAB
+	pshufd	$0x1b,%xmm2,%xmm2  // EFGH
+	movdqa	%xmm7,%xmm8  // offload
+	palignr	$8,%xmm2,%xmm1  // ABEF
+	punpcklqdq	%xmm0,%xmm2  // CDGH
+	jmp	L$oop_shaext
+
+.p2align	4
+L$oop_shaext:
+	movdqu	(%rsi),%xmm3
+	movdqu	16(%rsi),%xmm4
+	movdqu	32(%rsi),%xmm5
+	pshufb	%xmm7,%xmm3
+	movdqu	48(%rsi),%xmm6
+
+	movdqa	0-128(%rcx),%xmm0
+	paddd	%xmm3,%xmm0
+	pshufb	%xmm7,%xmm4
+	movdqa	%xmm2,%xmm10  // offload
+	sha256rnds2	%xmm1,%xmm2  // 0-3
+	pshufd	$0x0e,%xmm0,%xmm0
+	nop
+	movdqa	%xmm1,%xmm9  // offload
+	sha256rnds2	%xmm2,%xmm1
+
+	movdqa	32-128(%rcx),%xmm0
+	paddd	%xmm4,%xmm0
+	pshufb	%xmm7,%xmm5
+	sha256rnds2	%xmm1,%xmm2  // 4-7
+	pshufd	$0x0e,%xmm0,%xmm0
+	leaq	64(%rsi),%rsi
+	sha256msg1	%xmm4,%xmm3
+	sha256rnds2	%xmm2,%xmm1
+
+	movdqa	64-128(%rcx),%xmm0
+	paddd	%xmm5,%xmm0
+	pshufb	%xmm7,%xmm6
+	sha256rnds2	%xmm1,%xmm2  // 8-11
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm6,%xmm7
+	palignr	$4,%xmm5,%xmm7
+	nop
+	paddd	%xmm7,%xmm3
+	sha256msg1	%xmm5,%xmm4
+	sha256rnds2	%xmm2,%xmm1
+
+	movdqa	96-128(%rcx),%xmm0
+	paddd	%xmm6,%xmm0
+	sha256msg2	%xmm6,%xmm3
+	sha256rnds2	%xmm1,%xmm2  // 12-15
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm3,%xmm7
+	palignr	$4,%xmm6,%xmm7
+	nop
+	paddd	%xmm7,%xmm4
+	sha256msg1	%xmm6,%xmm5
+	sha256rnds2	%xmm2,%xmm1
+	movdqa	128-128(%rcx),%xmm0
+	paddd	%xmm3,%xmm0
+	sha256msg2	%xmm3,%xmm4
+	sha256rnds2	%xmm1,%xmm2  // 16-19...
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm4,%xmm7
+	palignr	$4,%xmm3,%xmm7
+	nop
+	paddd	%xmm7,%xmm5
+	sha256msg1	%xmm3,%xmm6
+	sha256rnds2	%xmm2,%xmm1
+	movdqa	160-128(%rcx),%xmm0
+	paddd	%xmm4,%xmm0
+	sha256msg2	%xmm4,%xmm5
+	sha256rnds2	%xmm1,%xmm2  // 16-19...
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm5,%xmm7
+	palignr	$4,%xmm4,%xmm7
+	nop
+	paddd	%xmm7,%xmm6
+	sha256msg1	%xmm4,%xmm3
+	sha256rnds2	%xmm2,%xmm1
+	movdqa	192-128(%rcx),%xmm0
+	paddd	%xmm5,%xmm0
+	sha256msg2	%xmm5,%xmm6
+	sha256rnds2	%xmm1,%xmm2  // 16-19...
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm6,%xmm7
+	palignr	$4,%xmm5,%xmm7
+	nop
+	paddd	%xmm7,%xmm3
+	sha256msg1	%xmm5,%xmm4
+	sha256rnds2	%xmm2,%xmm1
+	movdqa	224-128(%rcx),%xmm0
+	paddd	%xmm6,%xmm0
+	sha256msg2	%xmm6,%xmm3
+	sha256rnds2	%xmm1,%xmm2  // 16-19...
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm3,%xmm7
+	palignr	$4,%xmm6,%xmm7
+	nop
+	paddd	%xmm7,%xmm4
+	sha256msg1	%xmm6,%xmm5
+	sha256rnds2	%xmm2,%xmm1
+	movdqa	256-128(%rcx),%xmm0
+	paddd	%xmm3,%xmm0
+	sha256msg2	%xmm3,%xmm4
+	sha256rnds2	%xmm1,%xmm2  // 16-19...
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm4,%xmm7
+	palignr	$4,%xmm3,%xmm7
+	nop
+	paddd	%xmm7,%xmm5
+	sha256msg1	%xmm3,%xmm6
+	sha256rnds2	%xmm2,%xmm1
+	movdqa	288-128(%rcx),%xmm0
+	paddd	%xmm4,%xmm0
+	sha256msg2	%xmm4,%xmm5
+	sha256rnds2	%xmm1,%xmm2  // 16-19...
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm5,%xmm7
+	palignr	$4,%xmm4,%xmm7
+	nop
+	paddd	%xmm7,%xmm6
+	sha256msg1	%xmm4,%xmm3
+	sha256rnds2	%xmm2,%xmm1
+	movdqa	320-128(%rcx),%xmm0
+	paddd	%xmm5,%xmm0
+	sha256msg2	%xmm5,%xmm6
+	sha256rnds2	%xmm1,%xmm2  // 16-19...
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm6,%xmm7
+	palignr	$4,%xmm5,%xmm7
+	nop
+	paddd	%xmm7,%xmm3
+	sha256msg1	%xmm5,%xmm4
+	sha256rnds2	%xmm2,%xmm1
+	movdqa	352-128(%rcx),%xmm0
+	paddd	%xmm6,%xmm0
+	sha256msg2	%xmm6,%xmm3
+	sha256rnds2	%xmm1,%xmm2  // 16-19...
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm3,%xmm7
+	palignr	$4,%xmm6,%xmm7
+	nop
+	paddd	%xmm7,%xmm4
+	sha256msg1	%xmm6,%xmm5
+	sha256rnds2	%xmm2,%xmm1
+	movdqa	384-128(%rcx),%xmm0
+	paddd	%xmm3,%xmm0
+	sha256msg2	%xmm3,%xmm4
+	sha256rnds2	%xmm1,%xmm2  // 16-19...
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm4,%xmm7
+	palignr	$4,%xmm3,%xmm7
+	nop
+	paddd	%xmm7,%xmm5
+	sha256msg1	%xmm3,%xmm6
+	sha256rnds2	%xmm2,%xmm1
+	movdqa	416-128(%rcx),%xmm0
+	paddd	%xmm4,%xmm0
+	sha256msg2	%xmm4,%xmm5
+	sha256rnds2	%xmm1,%xmm2  // 52-55
+	pshufd	$0x0e,%xmm0,%xmm0
+	movdqa	%xmm5,%xmm7
+	palignr	$4,%xmm4,%xmm7
+	sha256rnds2	%xmm2,%xmm1
+	paddd	%xmm7,%xmm6
+
+	movdqa	448-128(%rcx),%xmm0
+	paddd	%xmm5,%xmm0
+	sha256rnds2	%xmm1,%xmm2  // 56-59
+	pshufd	$0x0e,%xmm0,%xmm0
+	sha256msg2	%xmm5,%xmm6
+	movdqa	%xmm8,%xmm7
+	sha256rnds2	%xmm2,%xmm1
+
+	movdqa	480-128(%rcx),%xmm0
+	paddd	%xmm6,%xmm0
+	nop
+	sha256rnds2	%xmm1,%xmm2  // 60-63
+	pshufd	$0x0e,%xmm0,%xmm0
+	decq	%rdx
+	nop
+	sha256rnds2	%xmm2,%xmm1
+
+	paddd	%xmm10,%xmm2
+	paddd	%xmm9,%xmm1
+	jnz	L$oop_shaext
+
+	pshufd	$0xb1,%xmm2,%xmm2  // DCHG
+	pshufd	$0x1b,%xmm1,%xmm7  // FEBA
+	pshufd	$0xb1,%xmm1,%xmm1  // BAFE
+	punpckhqdq	%xmm2,%xmm1  // DCBA
+	palignr	$8,%xmm7,%xmm2  // HGFE
+
+	movdqu	%xmm1,(%rdi)
+	movdqu	%xmm2,16(%rdi)
+	ret
+
+
+.globl	_sha256_block_data_order_ssse3
+.private_extern _sha256_block_data_order_ssse3
+
+.p2align	6
+.alt_entry _sha256_block_data_order_ssse3
+_sha256_block_data_order_ssse3:
+
+_CET_ENDBR
+	movq	%rsp,%rax  // copy %rsp
+
+	pushq	%rbx
+
+	pushq	%rbp
+
+	pushq	%r12
+
+	pushq	%r13
+
+	pushq	%r14
+
+	pushq	%r15
+
+	shlq	$4,%rdx  // num*16
+	subq	$96,%rsp
+	leaq	(%rsi,%rdx,4),%rdx  // inp+num*16*4
+	andq	$-64,%rsp  // align stack frame
+	movq	%rdi,64+0(%rsp)  // save ctx, 1st arg
+	movq	%rsi,64+8(%rsp)  // save inp, 2nd arh
+	movq	%rdx,64+16(%rsp)  // save end pointer, "3rd" arg
+	movq	%rax,88(%rsp)  // save copy of %rsp
+
+L$prologue_ssse3:
+
+	movl	0(%rdi),%eax
+	movl	4(%rdi),%ebx
+	movl	8(%rdi),%ecx
+	movl	12(%rdi),%edx
+	movl	16(%rdi),%r8d
+	movl	20(%rdi),%r9d
+	movl	24(%rdi),%r10d
+	movl	28(%rdi),%r11d
+// movdqa	K256+512+32(%rip),%xmm8
+// movdqa	K256+512+64(%rip),%xmm9
+	jmp	L$loop_ssse3
+.p2align	4
+L$loop_ssse3:
+	movdqa	K256+512(%rip),%xmm7
+	movdqu	0(%rsi),%xmm0
+	movdqu	16(%rsi),%xmm1
+	movdqu	32(%rsi),%xmm2
+	pshufb	%xmm7,%xmm0
+	movdqu	48(%rsi),%xmm3
+	leaq	K256(%rip),%rbp
+	pshufb	%xmm7,%xmm1
+	movdqa	0(%rbp),%xmm4
+	movdqa	32(%rbp),%xmm5
+	pshufb	%xmm7,%xmm2
+	paddd	%xmm0,%xmm4
+	movdqa	64(%rbp),%xmm6
+	pshufb	%xmm7,%xmm3
+	movdqa	96(%rbp),%xmm7
+	paddd	%xmm1,%xmm5
+	paddd	%xmm2,%xmm6
+	paddd	%xmm3,%xmm7
+	movdqa	%xmm4,0(%rsp)
+	movl	%eax,%r14d
+	movdqa	%xmm5,16(%rsp)
+	movl	%ebx,%edi
+	movdqa	%xmm6,32(%rsp)
+	xorl	%ecx,%edi  // magic
+	movdqa	%xmm7,48(%rsp)
+	movl	%r8d,%r13d
+	jmp	L$ssse3_00_47
+
+.p2align	4
+L$ssse3_00_47:
+	subq	$-128,%rbp  // size optimization
+	rorl	$14,%r13d
+	movdqa	%xmm1,%xmm4
+	movl	%r14d,%eax
+	movl	%r9d,%r12d
+	movdqa	%xmm3,%xmm7
+	rorl	$9,%r14d
+	xorl	%r8d,%r13d
+	xorl	%r10d,%r12d
+	rorl	$5,%r13d
+	xorl	%eax,%r14d
+	palignr	$4,%xmm0,%xmm4
+	andl	%r8d,%r12d
+	xorl	%r8d,%r13d
+	palignr	$4,%xmm2,%xmm7
+	addl	0(%rsp),%r11d
+	movl	%eax,%r15d
+	xorl	%r10d,%r12d
+	rorl	$11,%r14d
+	movdqa	%xmm4,%xmm5
+	xorl	%ebx,%r15d
+	addl	%r12d,%r11d
+	movdqa	%xmm4,%xmm6
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	psrld	$3,%xmm4
+	xorl	%eax,%r14d
+	addl	%r13d,%r11d
+	xorl	%ebx,%edi
+	paddd	%xmm7,%xmm0
+	rorl	$2,%r14d
+	addl	%r11d,%edx
+	psrld	$7,%xmm6
+	addl	%edi,%r11d
+	movl	%edx,%r13d
+	pshufd	$250,%xmm3,%xmm7
+	addl	%r11d,%r14d
+	rorl	$14,%r13d
+	pslld	$14,%xmm5
+	movl	%r14d,%r11d
+	movl	%r8d,%r12d
+	pxor	%xmm6,%xmm4
+	rorl	$9,%r14d
+	xorl	%edx,%r13d
+	xorl	%r9d,%r12d
+	rorl	$5,%r13d
+	psrld	$11,%xmm6
+	xorl	%r11d,%r14d
+	pxor	%xmm5,%xmm4
+	andl	%edx,%r12d
+	xorl	%edx,%r13d
+	pslld	$11,%xmm5
+	addl	4(%rsp),%r10d
+	movl	%r11d,%edi
+	pxor	%xmm6,%xmm4
+	xorl	%r9d,%r12d
+	rorl	$11,%r14d
+	movdqa	%xmm7,%xmm6
+	xorl	%eax,%edi
+	addl	%r12d,%r10d
+	pxor	%xmm5,%xmm4
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%r11d,%r14d
+	psrld	$10,%xmm7
+	addl	%r13d,%r10d
+	xorl	%eax,%r15d
+	paddd	%xmm4,%xmm0
+	rorl	$2,%r14d
+	addl	%r10d,%ecx
+	psrlq	$17,%xmm6
+	addl	%r15d,%r10d
+	movl	%ecx,%r13d
+	addl	%r10d,%r14d
+	pxor	%xmm6,%xmm7
+	rorl	$14,%r13d
+	movl	%r14d,%r10d
+	movl	%edx,%r12d
+	rorl	$9,%r14d
+	psrlq	$2,%xmm6
+	xorl	%ecx,%r13d
+	xorl	%r8d,%r12d
+	pxor	%xmm6,%xmm7
+	rorl	$5,%r13d
+	xorl	%r10d,%r14d
+	andl	%ecx,%r12d
+	pshufd	$128,%xmm7,%xmm7
+	xorl	%ecx,%r13d
+	addl	8(%rsp),%r9d
+	movl	%r10d,%r15d
+	psrldq	$8,%xmm7
+	xorl	%r8d,%r12d
+	rorl	$11,%r14d
+	xorl	%r11d,%r15d
+	addl	%r12d,%r9d
+	rorl	$6,%r13d
+	paddd	%xmm7,%xmm0
+	andl	%r15d,%edi
+	xorl	%r10d,%r14d
+	addl	%r13d,%r9d
+	pshufd	$80,%xmm0,%xmm7
+	xorl	%r11d,%edi
+	rorl	$2,%r14d
+	addl	%r9d,%ebx
+	movdqa	%xmm7,%xmm6
+	addl	%edi,%r9d
+	movl	%ebx,%r13d
+	psrld	$10,%xmm7
+	addl	%r9d,%r14d
+	rorl	$14,%r13d
+	psrlq	$17,%xmm6
+	movl	%r14d,%r9d
+	movl	%ecx,%r12d
+	pxor	%xmm6,%xmm7
+	rorl	$9,%r14d
+	xorl	%ebx,%r13d
+	xorl	%edx,%r12d
+	rorl	$5,%r13d
+	xorl	%r9d,%r14d
+	psrlq	$2,%xmm6
+	andl	%ebx,%r12d
+	xorl	%ebx,%r13d
+	addl	12(%rsp),%r8d
+	pxor	%xmm6,%xmm7
+	movl	%r9d,%edi
+	xorl	%edx,%r12d
+	rorl	$11,%r14d
+	pshufd	$8,%xmm7,%xmm7
+	xorl	%r10d,%edi
+	addl	%r12d,%r8d
+	movdqa	0(%rbp),%xmm6
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	pslldq	$8,%xmm7
+	xorl	%r9d,%r14d
+	addl	%r13d,%r8d
+	xorl	%r10d,%r15d
+	paddd	%xmm7,%xmm0
+	rorl	$2,%r14d
+	addl	%r8d,%eax
+	addl	%r15d,%r8d
+	paddd	%xmm0,%xmm6
+	movl	%eax,%r13d
+	addl	%r8d,%r14d
+	movdqa	%xmm6,0(%rsp)
+	rorl	$14,%r13d
+	movdqa	%xmm2,%xmm4
+	movl	%r14d,%r8d
+	movl	%ebx,%r12d
+	movdqa	%xmm0,%xmm7
+	rorl	$9,%r14d
+	xorl	%eax,%r13d
+	xorl	%ecx,%r12d
+	rorl	$5,%r13d
+	xorl	%r8d,%r14d
+	palignr	$4,%xmm1,%xmm4
+	andl	%eax,%r12d
+	xorl	%eax,%r13d
+	palignr	$4,%xmm3,%xmm7
+	addl	16(%rsp),%edx
+	movl	%r8d,%r15d
+	xorl	%ecx,%r12d
+	rorl	$11,%r14d
+	movdqa	%xmm4,%xmm5
+	xorl	%r9d,%r15d
+	addl	%r12d,%edx
+	movdqa	%xmm4,%xmm6
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	psrld	$3,%xmm4
+	xorl	%r8d,%r14d
+	addl	%r13d,%edx
+	xorl	%r9d,%edi
+	paddd	%xmm7,%xmm1
+	rorl	$2,%r14d
+	addl	%edx,%r11d
+	psrld	$7,%xmm6
+	addl	%edi,%edx
+	movl	%r11d,%r13d
+	pshufd	$250,%xmm0,%xmm7
+	addl	%edx,%r14d
+	rorl	$14,%r13d
+	pslld	$14,%xmm5
+	movl	%r14d,%edx
+	movl	%eax,%r12d
+	pxor	%xmm6,%xmm4
+	rorl	$9,%r14d
+	xorl	%r11d,%r13d
+	xorl	%ebx,%r12d
+	rorl	$5,%r13d
+	psrld	$11,%xmm6
+	xorl	%edx,%r14d
+	pxor	%xmm5,%xmm4
+	andl	%r11d,%r12d
+	xorl	%r11d,%r13d
+	pslld	$11,%xmm5
+	addl	20(%rsp),%ecx
+	movl	%edx,%edi
+	pxor	%xmm6,%xmm4
+	xorl	%ebx,%r12d
+	rorl	$11,%r14d
+	movdqa	%xmm7,%xmm6
+	xorl	%r8d,%edi
+	addl	%r12d,%ecx
+	pxor	%xmm5,%xmm4
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%edx,%r14d
+	psrld	$10,%xmm7
+	addl	%r13d,%ecx
+	xorl	%r8d,%r15d
+	paddd	%xmm4,%xmm1
+	rorl	$2,%r14d
+	addl	%ecx,%r10d
+	psrlq	$17,%xmm6
+	addl	%r15d,%ecx
+	movl	%r10d,%r13d
+	addl	%ecx,%r14d
+	pxor	%xmm6,%xmm7
+	rorl	$14,%r13d
+	movl	%r14d,%ecx
+	movl	%r11d,%r12d
+	rorl	$9,%r14d
+	psrlq	$2,%xmm6
+	xorl	%r10d,%r13d
+	xorl	%eax,%r12d
+	pxor	%xmm6,%xmm7
+	rorl	$5,%r13d
+	xorl	%ecx,%r14d
+	andl	%r10d,%r12d
+	pshufd	$128,%xmm7,%xmm7
+	xorl	%r10d,%r13d
+	addl	24(%rsp),%ebx
+	movl	%ecx,%r15d
+	psrldq	$8,%xmm7
+	xorl	%eax,%r12d
+	rorl	$11,%r14d
+	xorl	%edx,%r15d
+	addl	%r12d,%ebx
+	rorl	$6,%r13d
+	paddd	%xmm7,%xmm1
+	andl	%r15d,%edi
+	xorl	%ecx,%r14d
+	addl	%r13d,%ebx
+	pshufd	$80,%xmm1,%xmm7
+	xorl	%edx,%edi
+	rorl	$2,%r14d
+	addl	%ebx,%r9d
+	movdqa	%xmm7,%xmm6
+	addl	%edi,%ebx
+	movl	%r9d,%r13d
+	psrld	$10,%xmm7
+	addl	%ebx,%r14d
+	rorl	$14,%r13d
+	psrlq	$17,%xmm6
+	movl	%r14d,%ebx
+	movl	%r10d,%r12d
+	pxor	%xmm6,%xmm7
+	rorl	$9,%r14d
+	xorl	%r9d,%r13d
+	xorl	%r11d,%r12d
+	rorl	$5,%r13d
+	xorl	%ebx,%r14d
+	psrlq	$2,%xmm6
+	andl	%r9d,%r12d
+	xorl	%r9d,%r13d
+	addl	28(%rsp),%eax
+	pxor	%xmm6,%xmm7
+	movl	%ebx,%edi
+	xorl	%r11d,%r12d
+	rorl	$11,%r14d
+	pshufd	$8,%xmm7,%xmm7
+	xorl	%ecx,%edi
+	addl	%r12d,%eax
+	movdqa	32(%rbp),%xmm6
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	pslldq	$8,%xmm7
+	xorl	%ebx,%r14d
+	addl	%r13d,%eax
+	xorl	%ecx,%r15d
+	paddd	%xmm7,%xmm1
+	rorl	$2,%r14d
+	addl	%eax,%r8d
+	addl	%r15d,%eax
+	paddd	%xmm1,%xmm6
+	movl	%r8d,%r13d
+	addl	%eax,%r14d
+	movdqa	%xmm6,16(%rsp)
+	rorl	$14,%r13d
+	movdqa	%xmm3,%xmm4
+	movl	%r14d,%eax
+	movl	%r9d,%r12d
+	movdqa	%xmm1,%xmm7
+	rorl	$9,%r14d
+	xorl	%r8d,%r13d
+	xorl	%r10d,%r12d
+	rorl	$5,%r13d
+	xorl	%eax,%r14d
+	palignr	$4,%xmm2,%xmm4
+	andl	%r8d,%r12d
+	xorl	%r8d,%r13d
+	palignr	$4,%xmm0,%xmm7
+	addl	32(%rsp),%r11d
+	movl	%eax,%r15d
+	xorl	%r10d,%r12d
+	rorl	$11,%r14d
+	movdqa	%xmm4,%xmm5
+	xorl	%ebx,%r15d
+	addl	%r12d,%r11d
+	movdqa	%xmm4,%xmm6
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	psrld	$3,%xmm4
+	xorl	%eax,%r14d
+	addl	%r13d,%r11d
+	xorl	%ebx,%edi
+	paddd	%xmm7,%xmm2
+	rorl	$2,%r14d
+	addl	%r11d,%edx
+	psrld	$7,%xmm6
+	addl	%edi,%r11d
+	movl	%edx,%r13d
+	pshufd	$250,%xmm1,%xmm7
+	addl	%r11d,%r14d
+	rorl	$14,%r13d
+	pslld	$14,%xmm5
+	movl	%r14d,%r11d
+	movl	%r8d,%r12d
+	pxor	%xmm6,%xmm4
+	rorl	$9,%r14d
+	xorl	%edx,%r13d
+	xorl	%r9d,%r12d
+	rorl	$5,%r13d
+	psrld	$11,%xmm6
+	xorl	%r11d,%r14d
+	pxor	%xmm5,%xmm4
+	andl	%edx,%r12d
+	xorl	%edx,%r13d
+	pslld	$11,%xmm5
+	addl	36(%rsp),%r10d
+	movl	%r11d,%edi
+	pxor	%xmm6,%xmm4
+	xorl	%r9d,%r12d
+	rorl	$11,%r14d
+	movdqa	%xmm7,%xmm6
+	xorl	%eax,%edi
+	addl	%r12d,%r10d
+	pxor	%xmm5,%xmm4
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%r11d,%r14d
+	psrld	$10,%xmm7
+	addl	%r13d,%r10d
+	xorl	%eax,%r15d
+	paddd	%xmm4,%xmm2
+	rorl	$2,%r14d
+	addl	%r10d,%ecx
+	psrlq	$17,%xmm6
+	addl	%r15d,%r10d
+	movl	%ecx,%r13d
+	addl	%r10d,%r14d
+	pxor	%xmm6,%xmm7
+	rorl	$14,%r13d
+	movl	%r14d,%r10d
+	movl	%edx,%r12d
+	rorl	$9,%r14d
+	psrlq	$2,%xmm6
+	xorl	%ecx,%r13d
+	xorl	%r8d,%r12d
+	pxor	%xmm6,%xmm7
+	rorl	$5,%r13d
+	xorl	%r10d,%r14d
+	andl	%ecx,%r12d
+	pshufd	$128,%xmm7,%xmm7
+	xorl	%ecx,%r13d
+	addl	40(%rsp),%r9d
+	movl	%r10d,%r15d
+	psrldq	$8,%xmm7
+	xorl	%r8d,%r12d
+	rorl	$11,%r14d
+	xorl	%r11d,%r15d
+	addl	%r12d,%r9d
+	rorl	$6,%r13d
+	paddd	%xmm7,%xmm2
+	andl	%r15d,%edi
+	xorl	%r10d,%r14d
+	addl	%r13d,%r9d
+	pshufd	$80,%xmm2,%xmm7
+	xorl	%r11d,%edi
+	rorl	$2,%r14d
+	addl	%r9d,%ebx
+	movdqa	%xmm7,%xmm6
+	addl	%edi,%r9d
+	movl	%ebx,%r13d
+	psrld	$10,%xmm7
+	addl	%r9d,%r14d
+	rorl	$14,%r13d
+	psrlq	$17,%xmm6
+	movl	%r14d,%r9d
+	movl	%ecx,%r12d
+	pxor	%xmm6,%xmm7
+	rorl	$9,%r14d
+	xorl	%ebx,%r13d
+	xorl	%edx,%r12d
+	rorl	$5,%r13d
+	xorl	%r9d,%r14d
+	psrlq	$2,%xmm6
+	andl	%ebx,%r12d
+	xorl	%ebx,%r13d
+	addl	44(%rsp),%r8d
+	pxor	%xmm6,%xmm7
+	movl	%r9d,%edi
+	xorl	%edx,%r12d
+	rorl	$11,%r14d
+	pshufd	$8,%xmm7,%xmm7
+	xorl	%r10d,%edi
+	addl	%r12d,%r8d
+	movdqa	64(%rbp),%xmm6
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	pslldq	$8,%xmm7
+	xorl	%r9d,%r14d
+	addl	%r13d,%r8d
+	xorl	%r10d,%r15d
+	paddd	%xmm7,%xmm2
+	rorl	$2,%r14d
+	addl	%r8d,%eax
+	addl	%r15d,%r8d
+	paddd	%xmm2,%xmm6
+	movl	%eax,%r13d
+	addl	%r8d,%r14d
+	movdqa	%xmm6,32(%rsp)
+	rorl	$14,%r13d
+	movdqa	%xmm0,%xmm4
+	movl	%r14d,%r8d
+	movl	%ebx,%r12d
+	movdqa	%xmm2,%xmm7
+	rorl	$9,%r14d
+	xorl	%eax,%r13d
+	xorl	%ecx,%r12d
+	rorl	$5,%r13d
+	xorl	%r8d,%r14d
+	palignr	$4,%xmm3,%xmm4
+	andl	%eax,%r12d
+	xorl	%eax,%r13d
+	palignr	$4,%xmm1,%xmm7
+	addl	48(%rsp),%edx
+	movl	%r8d,%r15d
+	xorl	%ecx,%r12d
+	rorl	$11,%r14d
+	movdqa	%xmm4,%xmm5
+	xorl	%r9d,%r15d
+	addl	%r12d,%edx
+	movdqa	%xmm4,%xmm6
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	psrld	$3,%xmm4
+	xorl	%r8d,%r14d
+	addl	%r13d,%edx
+	xorl	%r9d,%edi
+	paddd	%xmm7,%xmm3
+	rorl	$2,%r14d
+	addl	%edx,%r11d
+	psrld	$7,%xmm6
+	addl	%edi,%edx
+	movl	%r11d,%r13d
+	pshufd	$250,%xmm2,%xmm7
+	addl	%edx,%r14d
+	rorl	$14,%r13d
+	pslld	$14,%xmm5
+	movl	%r14d,%edx
+	movl	%eax,%r12d
+	pxor	%xmm6,%xmm4
+	rorl	$9,%r14d
+	xorl	%r11d,%r13d
+	xorl	%ebx,%r12d
+	rorl	$5,%r13d
+	psrld	$11,%xmm6
+	xorl	%edx,%r14d
+	pxor	%xmm5,%xmm4
+	andl	%r11d,%r12d
+	xorl	%r11d,%r13d
+	pslld	$11,%xmm5
+	addl	52(%rsp),%ecx
+	movl	%edx,%edi
+	pxor	%xmm6,%xmm4
+	xorl	%ebx,%r12d
+	rorl	$11,%r14d
+	movdqa	%xmm7,%xmm6
+	xorl	%r8d,%edi
+	addl	%r12d,%ecx
+	pxor	%xmm5,%xmm4
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%edx,%r14d
+	psrld	$10,%xmm7
+	addl	%r13d,%ecx
+	xorl	%r8d,%r15d
+	paddd	%xmm4,%xmm3
+	rorl	$2,%r14d
+	addl	%ecx,%r10d
+	psrlq	$17,%xmm6
+	addl	%r15d,%ecx
+	movl	%r10d,%r13d
+	addl	%ecx,%r14d
+	pxor	%xmm6,%xmm7
+	rorl	$14,%r13d
+	movl	%r14d,%ecx
+	movl	%r11d,%r12d
+	rorl	$9,%r14d
+	psrlq	$2,%xmm6
+	xorl	%r10d,%r13d
+	xorl	%eax,%r12d
+	pxor	%xmm6,%xmm7
+	rorl	$5,%r13d
+	xorl	%ecx,%r14d
+	andl	%r10d,%r12d
+	pshufd	$128,%xmm7,%xmm7
+	xorl	%r10d,%r13d
+	addl	56(%rsp),%ebx
+	movl	%ecx,%r15d
+	psrldq	$8,%xmm7
+	xorl	%eax,%r12d
+	rorl	$11,%r14d
+	xorl	%edx,%r15d
+	addl	%r12d,%ebx
+	rorl	$6,%r13d
+	paddd	%xmm7,%xmm3
+	andl	%r15d,%edi
+	xorl	%ecx,%r14d
+	addl	%r13d,%ebx
+	pshufd	$80,%xmm3,%xmm7
+	xorl	%edx,%edi
+	rorl	$2,%r14d
+	addl	%ebx,%r9d
+	movdqa	%xmm7,%xmm6
+	addl	%edi,%ebx
+	movl	%r9d,%r13d
+	psrld	$10,%xmm7
+	addl	%ebx,%r14d
+	rorl	$14,%r13d
+	psrlq	$17,%xmm6
+	movl	%r14d,%ebx
+	movl	%r10d,%r12d
+	pxor	%xmm6,%xmm7
+	rorl	$9,%r14d
+	xorl	%r9d,%r13d
+	xorl	%r11d,%r12d
+	rorl	$5,%r13d
+	xorl	%ebx,%r14d
+	psrlq	$2,%xmm6
+	andl	%r9d,%r12d
+	xorl	%r9d,%r13d
+	addl	60(%rsp),%eax
+	pxor	%xmm6,%xmm7
+	movl	%ebx,%edi
+	xorl	%r11d,%r12d
+	rorl	$11,%r14d
+	pshufd	$8,%xmm7,%xmm7
+	xorl	%ecx,%edi
+	addl	%r12d,%eax
+	movdqa	96(%rbp),%xmm6
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	pslldq	$8,%xmm7
+	xorl	%ebx,%r14d
+	addl	%r13d,%eax
+	xorl	%ecx,%r15d
+	paddd	%xmm7,%xmm3
+	rorl	$2,%r14d
+	addl	%eax,%r8d
+	addl	%r15d,%eax
+	paddd	%xmm3,%xmm6
+	movl	%r8d,%r13d
+	addl	%eax,%r14d
+	movdqa	%xmm6,48(%rsp)
+	cmpb	$0,131(%rbp)
+	jne	L$ssse3_00_47
+	rorl	$14,%r13d
+	movl	%r14d,%eax
+	movl	%r9d,%r12d
+	rorl	$9,%r14d
+	xorl	%r8d,%r13d
+	xorl	%r10d,%r12d
+	rorl	$5,%r13d
+	xorl	%eax,%r14d
+	andl	%r8d,%r12d
+	xorl	%r8d,%r13d
+	addl	0(%rsp),%r11d
+	movl	%eax,%r15d
+	xorl	%r10d,%r12d
+	rorl	$11,%r14d
+	xorl	%ebx,%r15d
+	addl	%r12d,%r11d
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	xorl	%eax,%r14d
+	addl	%r13d,%r11d
+	xorl	%ebx,%edi
+	rorl	$2,%r14d
+	addl	%r11d,%edx
+	addl	%edi,%r11d
+	movl	%edx,%r13d
+	addl	%r11d,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%r11d
+	movl	%r8d,%r12d
+	rorl	$9,%r14d
+	xorl	%edx,%r13d
+	xorl	%r9d,%r12d
+	rorl	$5,%r13d
+	xorl	%r11d,%r14d
+	andl	%edx,%r12d
+	xorl	%edx,%r13d
+	addl	4(%rsp),%r10d
+	movl	%r11d,%edi
+	xorl	%r9d,%r12d
+	rorl	$11,%r14d
+	xorl	%eax,%edi
+	addl	%r12d,%r10d
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%r11d,%r14d
+	addl	%r13d,%r10d
+	xorl	%eax,%r15d
+	rorl	$2,%r14d
+	addl	%r10d,%ecx
+	addl	%r15d,%r10d
+	movl	%ecx,%r13d
+	addl	%r10d,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%r10d
+	movl	%edx,%r12d
+	rorl	$9,%r14d
+	xorl	%ecx,%r13d
+	xorl	%r8d,%r12d
+	rorl	$5,%r13d
+	xorl	%r10d,%r14d
+	andl	%ecx,%r12d
+	xorl	%ecx,%r13d
+	addl	8(%rsp),%r9d
+	movl	%r10d,%r15d
+	xorl	%r8d,%r12d
+	rorl	$11,%r14d
+	xorl	%r11d,%r15d
+	addl	%r12d,%r9d
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	xorl	%r10d,%r14d
+	addl	%r13d,%r9d
+	xorl	%r11d,%edi
+	rorl	$2,%r14d
+	addl	%r9d,%ebx
+	addl	%edi,%r9d
+	movl	%ebx,%r13d
+	addl	%r9d,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%r9d
+	movl	%ecx,%r12d
+	rorl	$9,%r14d
+	xorl	%ebx,%r13d
+	xorl	%edx,%r12d
+	rorl	$5,%r13d
+	xorl	%r9d,%r14d
+	andl	%ebx,%r12d
+	xorl	%ebx,%r13d
+	addl	12(%rsp),%r8d
+	movl	%r9d,%edi
+	xorl	%edx,%r12d
+	rorl	$11,%r14d
+	xorl	%r10d,%edi
+	addl	%r12d,%r8d
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%r9d,%r14d
+	addl	%r13d,%r8d
+	xorl	%r10d,%r15d
+	rorl	$2,%r14d
+	addl	%r8d,%eax
+	addl	%r15d,%r8d
+	movl	%eax,%r13d
+	addl	%r8d,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%r8d
+	movl	%ebx,%r12d
+	rorl	$9,%r14d
+	xorl	%eax,%r13d
+	xorl	%ecx,%r12d
+	rorl	$5,%r13d
+	xorl	%r8d,%r14d
+	andl	%eax,%r12d
+	xorl	%eax,%r13d
+	addl	16(%rsp),%edx
+	movl	%r8d,%r15d
+	xorl	%ecx,%r12d
+	rorl	$11,%r14d
+	xorl	%r9d,%r15d
+	addl	%r12d,%edx
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	xorl	%r8d,%r14d
+	addl	%r13d,%edx
+	xorl	%r9d,%edi
+	rorl	$2,%r14d
+	addl	%edx,%r11d
+	addl	%edi,%edx
+	movl	%r11d,%r13d
+	addl	%edx,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%edx
+	movl	%eax,%r12d
+	rorl	$9,%r14d
+	xorl	%r11d,%r13d
+	xorl	%ebx,%r12d
+	rorl	$5,%r13d
+	xorl	%edx,%r14d
+	andl	%r11d,%r12d
+	xorl	%r11d,%r13d
+	addl	20(%rsp),%ecx
+	movl	%edx,%edi
+	xorl	%ebx,%r12d
+	rorl	$11,%r14d
+	xorl	%r8d,%edi
+	addl	%r12d,%ecx
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%edx,%r14d
+	addl	%r13d,%ecx
+	xorl	%r8d,%r15d
+	rorl	$2,%r14d
+	addl	%ecx,%r10d
+	addl	%r15d,%ecx
+	movl	%r10d,%r13d
+	addl	%ecx,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%ecx
+	movl	%r11d,%r12d
+	rorl	$9,%r14d
+	xorl	%r10d,%r13d
+	xorl	%eax,%r12d
+	rorl	$5,%r13d
+	xorl	%ecx,%r14d
+	andl	%r10d,%r12d
+	xorl	%r10d,%r13d
+	addl	24(%rsp),%ebx
+	movl	%ecx,%r15d
+	xorl	%eax,%r12d
+	rorl	$11,%r14d
+	xorl	%edx,%r15d
+	addl	%r12d,%ebx
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	xorl	%ecx,%r14d
+	addl	%r13d,%ebx
+	xorl	%edx,%edi
+	rorl	$2,%r14d
+	addl	%ebx,%r9d
+	addl	%edi,%ebx
+	movl	%r9d,%r13d
+	addl	%ebx,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%ebx
+	movl	%r10d,%r12d
+	rorl	$9,%r14d
+	xorl	%r9d,%r13d
+	xorl	%r11d,%r12d
+	rorl	$5,%r13d
+	xorl	%ebx,%r14d
+	andl	%r9d,%r12d
+	xorl	%r9d,%r13d
+	addl	28(%rsp),%eax
+	movl	%ebx,%edi
+	xorl	%r11d,%r12d
+	rorl	$11,%r14d
+	xorl	%ecx,%edi
+	addl	%r12d,%eax
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%ebx,%r14d
+	addl	%r13d,%eax
+	xorl	%ecx,%r15d
+	rorl	$2,%r14d
+	addl	%eax,%r8d
+	addl	%r15d,%eax
+	movl	%r8d,%r13d
+	addl	%eax,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%eax
+	movl	%r9d,%r12d
+	rorl	$9,%r14d
+	xorl	%r8d,%r13d
+	xorl	%r10d,%r12d
+	rorl	$5,%r13d
+	xorl	%eax,%r14d
+	andl	%r8d,%r12d
+	xorl	%r8d,%r13d
+	addl	32(%rsp),%r11d
+	movl	%eax,%r15d
+	xorl	%r10d,%r12d
+	rorl	$11,%r14d
+	xorl	%ebx,%r15d
+	addl	%r12d,%r11d
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	xorl	%eax,%r14d
+	addl	%r13d,%r11d
+	xorl	%ebx,%edi
+	rorl	$2,%r14d
+	addl	%r11d,%edx
+	addl	%edi,%r11d
+	movl	%edx,%r13d
+	addl	%r11d,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%r11d
+	movl	%r8d,%r12d
+	rorl	$9,%r14d
+	xorl	%edx,%r13d
+	xorl	%r9d,%r12d
+	rorl	$5,%r13d
+	xorl	%r11d,%r14d
+	andl	%edx,%r12d
+	xorl	%edx,%r13d
+	addl	36(%rsp),%r10d
+	movl	%r11d,%edi
+	xorl	%r9d,%r12d
+	rorl	$11,%r14d
+	xorl	%eax,%edi
+	addl	%r12d,%r10d
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%r11d,%r14d
+	addl	%r13d,%r10d
+	xorl	%eax,%r15d
+	rorl	$2,%r14d
+	addl	%r10d,%ecx
+	addl	%r15d,%r10d
+	movl	%ecx,%r13d
+	addl	%r10d,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%r10d
+	movl	%edx,%r12d
+	rorl	$9,%r14d
+	xorl	%ecx,%r13d
+	xorl	%r8d,%r12d
+	rorl	$5,%r13d
+	xorl	%r10d,%r14d
+	andl	%ecx,%r12d
+	xorl	%ecx,%r13d
+	addl	40(%rsp),%r9d
+	movl	%r10d,%r15d
+	xorl	%r8d,%r12d
+	rorl	$11,%r14d
+	xorl	%r11d,%r15d
+	addl	%r12d,%r9d
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	xorl	%r10d,%r14d
+	addl	%r13d,%r9d
+	xorl	%r11d,%edi
+	rorl	$2,%r14d
+	addl	%r9d,%ebx
+	addl	%edi,%r9d
+	movl	%ebx,%r13d
+	addl	%r9d,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%r9d
+	movl	%ecx,%r12d
+	rorl	$9,%r14d
+	xorl	%ebx,%r13d
+	xorl	%edx,%r12d
+	rorl	$5,%r13d
+	xorl	%r9d,%r14d
+	andl	%ebx,%r12d
+	xorl	%ebx,%r13d
+	addl	44(%rsp),%r8d
+	movl	%r9d,%edi
+	xorl	%edx,%r12d
+	rorl	$11,%r14d
+	xorl	%r10d,%edi
+	addl	%r12d,%r8d
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%r9d,%r14d
+	addl	%r13d,%r8d
+	xorl	%r10d,%r15d
+	rorl	$2,%r14d
+	addl	%r8d,%eax
+	addl	%r15d,%r8d
+	movl	%eax,%r13d
+	addl	%r8d,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%r8d
+	movl	%ebx,%r12d
+	rorl	$9,%r14d
+	xorl	%eax,%r13d
+	xorl	%ecx,%r12d
+	rorl	$5,%r13d
+	xorl	%r8d,%r14d
+	andl	%eax,%r12d
+	xorl	%eax,%r13d
+	addl	48(%rsp),%edx
+	movl	%r8d,%r15d
+	xorl	%ecx,%r12d
+	rorl	$11,%r14d
+	xorl	%r9d,%r15d
+	addl	%r12d,%edx
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	xorl	%r8d,%r14d
+	addl	%r13d,%edx
+	xorl	%r9d,%edi
+	rorl	$2,%r14d
+	addl	%edx,%r11d
+	addl	%edi,%edx
+	movl	%r11d,%r13d
+	addl	%edx,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%edx
+	movl	%eax,%r12d
+	rorl	$9,%r14d
+	xorl	%r11d,%r13d
+	xorl	%ebx,%r12d
+	rorl	$5,%r13d
+	xorl	%edx,%r14d
+	andl	%r11d,%r12d
+	xorl	%r11d,%r13d
+	addl	52(%rsp),%ecx
+	movl	%edx,%edi
+	xorl	%ebx,%r12d
+	rorl	$11,%r14d
+	xorl	%r8d,%edi
+	addl	%r12d,%ecx
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%edx,%r14d
+	addl	%r13d,%ecx
+	xorl	%r8d,%r15d
+	rorl	$2,%r14d
+	addl	%ecx,%r10d
+	addl	%r15d,%ecx
+	movl	%r10d,%r13d
+	addl	%ecx,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%ecx
+	movl	%r11d,%r12d
+	rorl	$9,%r14d
+	xorl	%r10d,%r13d
+	xorl	%eax,%r12d
+	rorl	$5,%r13d
+	xorl	%ecx,%r14d
+	andl	%r10d,%r12d
+	xorl	%r10d,%r13d
+	addl	56(%rsp),%ebx
+	movl	%ecx,%r15d
+	xorl	%eax,%r12d
+	rorl	$11,%r14d
+	xorl	%edx,%r15d
+	addl	%r12d,%ebx
+	rorl	$6,%r13d
+	andl	%r15d,%edi
+	xorl	%ecx,%r14d
+	addl	%r13d,%ebx
+	xorl	%edx,%edi
+	rorl	$2,%r14d
+	addl	%ebx,%r9d
+	addl	%edi,%ebx
+	movl	%r9d,%r13d
+	addl	%ebx,%r14d
+	rorl	$14,%r13d
+	movl	%r14d,%ebx
+	movl	%r10d,%r12d
+	rorl	$9,%r14d
+	xorl	%r9d,%r13d
+	xorl	%r11d,%r12d
+	rorl	$5,%r13d
+	xorl	%ebx,%r14d
+	andl	%r9d,%r12d
+	xorl	%r9d,%r13d
+	addl	60(%rsp),%eax
+	movl	%ebx,%edi
+	xorl	%r11d,%r12d
+	rorl	$11,%r14d
+	xorl	%ecx,%edi
+	addl	%r12d,%eax
+	rorl	$6,%r13d
+	andl	%edi,%r15d
+	xorl	%ebx,%r14d
+	addl	%r13d,%eax
+	xorl	%ecx,%r15d
+	rorl	$2,%r14d
+	addl	%eax,%r8d
+	addl	%r15d,%eax
+	movl	%r8d,%r13d
+	addl	%eax,%r14d
+	movq	64+0(%rsp),%rdi
+	movl	%r14d,%eax
+
+	addl	0(%rdi),%eax
+	leaq	64(%rsi),%rsi
+	addl	4(%rdi),%ebx
+	addl	8(%rdi),%ecx
+	addl	12(%rdi),%edx
+	addl	16(%rdi),%r8d
+	addl	20(%rdi),%r9d
+	addl	24(%rdi),%r10d
+	addl	28(%rdi),%r11d
+
+	cmpq	64+16(%rsp),%rsi
+
+	movl	%eax,0(%rdi)
+	movl	%ebx,4(%rdi)
+	movl	%ecx,8(%rdi)
+	movl	%edx,12(%rdi)
+	movl	%r8d,16(%rdi)
+	movl	%r9d,20(%rdi)
+	movl	%r10d,24(%rdi)
+	movl	%r11d,28(%rdi)
+	jb	L$loop_ssse3
+
+	movq	88(%rsp),%rsi
+
+	movq	-48(%rsi),%r15
+
+	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_ssse3:
+	ret
+
+
+.globl	_sha256_block_data_order_avx
+.private_extern _sha256_block_data_order_avx
+
+.p2align	6
+.alt_entry _sha256_block_data_order_avx
+_sha256_block_data_order_avx:
+
+_CET_ENDBR
+	movq	%rsp,%rax  // copy %rsp
+
+	pushq	%rbx
+
+	pushq	%rbp
+
+	pushq	%r12
+
+	pushq	%r13
+
+	pushq	%r14
+
+	pushq	%r15
+
+	shlq	$4,%rdx  // num*16
+	subq	$96,%rsp
+	leaq	(%rsi,%rdx,4),%rdx  // inp+num*16*4
+	andq	$-64,%rsp  // align stack frame
+	movq	%rdi,64+0(%rsp)  // save ctx, 1st arg
+	movq	%rsi,64+8(%rsp)  // save inp, 2nd arh
+	movq	%rdx,64+16(%rsp)  // save end pointer, "3rd" arg
+	movq	%rax,88(%rsp)  // save copy of %rsp
+
+L$prologue_avx:
+
+	vzeroupper
+	movl	0(%rdi),%eax
+	movl	4(%rdi),%ebx
+	movl	8(%rdi),%ecx
+	movl	12(%rdi),%edx
+	movl	16(%rdi),%r8d
+	movl	20(%rdi),%r9d
+	movl	24(%rdi),%r10d
+	movl	28(%rdi),%r11d
+	vmovdqa	K256+512+32(%rip),%xmm8
+	vmovdqa	K256+512+64(%rip),%xmm9
+	jmp	L$loop_avx
+.p2align	4
+L$loop_avx:
+	vmovdqa	K256+512(%rip),%xmm7
+	vmovdqu	0(%rsi),%xmm0
+	vmovdqu	16(%rsi),%xmm1
+	vmovdqu	32(%rsi),%xmm2
+	vmovdqu	48(%rsi),%xmm3
+	vpshufb	%xmm7,%xmm0,%xmm0
+	leaq	K256(%rip),%rbp
+	vpshufb	%xmm7,%xmm1,%xmm1
+	vpshufb	%xmm7,%xmm2,%xmm2
+	vpaddd	0(%rbp),%xmm0,%xmm4
+	vpshufb	%xmm7,%xmm3,%xmm3
+	vpaddd	32(%rbp),%xmm1,%xmm5
+	vpaddd	64(%rbp),%xmm2,%xmm6
+	vpaddd	96(%rbp),%xmm3,%xmm7
+	vmovdqa	%xmm4,0(%rsp)
+	movl	%eax,%r14d
+	vmovdqa	%xmm5,16(%rsp)
+	movl	%ebx,%edi
+	vmovdqa	%xmm6,32(%rsp)
+	xorl	%ecx,%edi  // magic
+	vmovdqa	%xmm7,48(%rsp)
+	movl	%r8d,%r13d
+	jmp	L$avx_00_47
+
+.p2align	4
+L$avx_00_47:
+	subq	$-128,%rbp  // size optimization
+	vpalignr	$4,%xmm0,%xmm1,%xmm4
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%eax
+	movl	%r9d,%r12d
+	vpalignr	$4,%xmm2,%xmm3,%xmm7
+	shrdl	$9,%r14d,%r14d
+	xorl	%r8d,%r13d
+	xorl	%r10d,%r12d
+	vpsrld	$7,%xmm4,%xmm6
+	shrdl	$5,%r13d,%r13d
+	xorl	%eax,%r14d
+	andl	%r8d,%r12d
+	vpaddd	%xmm7,%xmm0,%xmm0
+	xorl	%r8d,%r13d
+	addl	0(%rsp),%r11d
+	movl	%eax,%r15d
+	vpsrld	$3,%xmm4,%xmm7
+	xorl	%r10d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%ebx,%r15d
+	vpslld	$14,%xmm4,%xmm5
+	addl	%r12d,%r11d
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	vpxor	%xmm6,%xmm7,%xmm4
+	xorl	%eax,%r14d
+	addl	%r13d,%r11d
+	xorl	%ebx,%edi
+	vpshufd	$250,%xmm3,%xmm7
+	shrdl	$2,%r14d,%r14d
+	addl	%r11d,%edx
+	addl	%edi,%r11d
+	vpsrld	$11,%xmm6,%xmm6
+	movl	%edx,%r13d
+	addl	%r11d,%r14d
+	shrdl	$14,%r13d,%r13d
+	vpxor	%xmm5,%xmm4,%xmm4
+	movl	%r14d,%r11d
+	movl	%r8d,%r12d
+	shrdl	$9,%r14d,%r14d
+	vpslld	$11,%xmm5,%xmm5
+	xorl	%edx,%r13d
+	xorl	%r9d,%r12d
+	shrdl	$5,%r13d,%r13d
+	vpxor	%xmm6,%xmm4,%xmm4
+	xorl	%r11d,%r14d
+	andl	%edx,%r12d
+	xorl	%edx,%r13d
+	vpsrld	$10,%xmm7,%xmm6
+	addl	4(%rsp),%r10d
+	movl	%r11d,%edi
+	xorl	%r9d,%r12d
+	vpxor	%xmm5,%xmm4,%xmm4
+	shrdl	$11,%r14d,%r14d
+	xorl	%eax,%edi
+	addl	%r12d,%r10d
+	vpsrlq	$17,%xmm7,%xmm7
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%r11d,%r14d
+	vpaddd	%xmm4,%xmm0,%xmm0
+	addl	%r13d,%r10d
+	xorl	%eax,%r15d
+	shrdl	$2,%r14d,%r14d
+	vpxor	%xmm7,%xmm6,%xmm6
+	addl	%r10d,%ecx
+	addl	%r15d,%r10d
+	movl	%ecx,%r13d
+	vpsrlq	$2,%xmm7,%xmm7
+	addl	%r10d,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r10d
+	vpxor	%xmm7,%xmm6,%xmm6
+	movl	%edx,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%ecx,%r13d
+	vpshufb	%xmm8,%xmm6,%xmm6
+	xorl	%r8d,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%r10d,%r14d
+	vpaddd	%xmm6,%xmm0,%xmm0
+	andl	%ecx,%r12d
+	xorl	%ecx,%r13d
+	addl	8(%rsp),%r9d
+	vpshufd	$80,%xmm0,%xmm7
+	movl	%r10d,%r15d
+	xorl	%r8d,%r12d
+	shrdl	$11,%r14d,%r14d
+	vpsrld	$10,%xmm7,%xmm6
+	xorl	%r11d,%r15d
+	addl	%r12d,%r9d
+	shrdl	$6,%r13d,%r13d
+	vpsrlq	$17,%xmm7,%xmm7
+	andl	%r15d,%edi
+	xorl	%r10d,%r14d
+	addl	%r13d,%r9d
+	vpxor	%xmm7,%xmm6,%xmm6
+	xorl	%r11d,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%r9d,%ebx
+	vpsrlq	$2,%xmm7,%xmm7
+	addl	%edi,%r9d
+	movl	%ebx,%r13d
+	addl	%r9d,%r14d
+	vpxor	%xmm7,%xmm6,%xmm6
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r9d
+	movl	%ecx,%r12d
+	vpshufb	%xmm9,%xmm6,%xmm6
+	shrdl	$9,%r14d,%r14d
+	xorl	%ebx,%r13d
+	xorl	%edx,%r12d
+	vpaddd	%xmm6,%xmm0,%xmm0
+	shrdl	$5,%r13d,%r13d
+	xorl	%r9d,%r14d
+	andl	%ebx,%r12d
+	vpaddd	0(%rbp),%xmm0,%xmm6
+	xorl	%ebx,%r13d
+	addl	12(%rsp),%r8d
+	movl	%r9d,%edi
+	xorl	%edx,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r10d,%edi
+	addl	%r12d,%r8d
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%r9d,%r14d
+	addl	%r13d,%r8d
+	xorl	%r10d,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%r8d,%eax
+	addl	%r15d,%r8d
+	movl	%eax,%r13d
+	addl	%r8d,%r14d
+	vmovdqa	%xmm6,0(%rsp)
+	vpalignr	$4,%xmm1,%xmm2,%xmm4
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r8d
+	movl	%ebx,%r12d
+	vpalignr	$4,%xmm3,%xmm0,%xmm7
+	shrdl	$9,%r14d,%r14d
+	xorl	%eax,%r13d
+	xorl	%ecx,%r12d
+	vpsrld	$7,%xmm4,%xmm6
+	shrdl	$5,%r13d,%r13d
+	xorl	%r8d,%r14d
+	andl	%eax,%r12d
+	vpaddd	%xmm7,%xmm1,%xmm1
+	xorl	%eax,%r13d
+	addl	16(%rsp),%edx
+	movl	%r8d,%r15d
+	vpsrld	$3,%xmm4,%xmm7
+	xorl	%ecx,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r9d,%r15d
+	vpslld	$14,%xmm4,%xmm5
+	addl	%r12d,%edx
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	vpxor	%xmm6,%xmm7,%xmm4
+	xorl	%r8d,%r14d
+	addl	%r13d,%edx
+	xorl	%r9d,%edi
+	vpshufd	$250,%xmm0,%xmm7
+	shrdl	$2,%r14d,%r14d
+	addl	%edx,%r11d
+	addl	%edi,%edx
+	vpsrld	$11,%xmm6,%xmm6
+	movl	%r11d,%r13d
+	addl	%edx,%r14d
+	shrdl	$14,%r13d,%r13d
+	vpxor	%xmm5,%xmm4,%xmm4
+	movl	%r14d,%edx
+	movl	%eax,%r12d
+	shrdl	$9,%r14d,%r14d
+	vpslld	$11,%xmm5,%xmm5
+	xorl	%r11d,%r13d
+	xorl	%ebx,%r12d
+	shrdl	$5,%r13d,%r13d
+	vpxor	%xmm6,%xmm4,%xmm4
+	xorl	%edx,%r14d
+	andl	%r11d,%r12d
+	xorl	%r11d,%r13d
+	vpsrld	$10,%xmm7,%xmm6
+	addl	20(%rsp),%ecx
+	movl	%edx,%edi
+	xorl	%ebx,%r12d
+	vpxor	%xmm5,%xmm4,%xmm4
+	shrdl	$11,%r14d,%r14d
+	xorl	%r8d,%edi
+	addl	%r12d,%ecx
+	vpsrlq	$17,%xmm7,%xmm7
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%edx,%r14d
+	vpaddd	%xmm4,%xmm1,%xmm1
+	addl	%r13d,%ecx
+	xorl	%r8d,%r15d
+	shrdl	$2,%r14d,%r14d
+	vpxor	%xmm7,%xmm6,%xmm6
+	addl	%ecx,%r10d
+	addl	%r15d,%ecx
+	movl	%r10d,%r13d
+	vpsrlq	$2,%xmm7,%xmm7
+	addl	%ecx,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%ecx
+	vpxor	%xmm7,%xmm6,%xmm6
+	movl	%r11d,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%r10d,%r13d
+	vpshufb	%xmm8,%xmm6,%xmm6
+	xorl	%eax,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%ecx,%r14d
+	vpaddd	%xmm6,%xmm1,%xmm1
+	andl	%r10d,%r12d
+	xorl	%r10d,%r13d
+	addl	24(%rsp),%ebx
+	vpshufd	$80,%xmm1,%xmm7
+	movl	%ecx,%r15d
+	xorl	%eax,%r12d
+	shrdl	$11,%r14d,%r14d
+	vpsrld	$10,%xmm7,%xmm6
+	xorl	%edx,%r15d
+	addl	%r12d,%ebx
+	shrdl	$6,%r13d,%r13d
+	vpsrlq	$17,%xmm7,%xmm7
+	andl	%r15d,%edi
+	xorl	%ecx,%r14d
+	addl	%r13d,%ebx
+	vpxor	%xmm7,%xmm6,%xmm6
+	xorl	%edx,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%ebx,%r9d
+	vpsrlq	$2,%xmm7,%xmm7
+	addl	%edi,%ebx
+	movl	%r9d,%r13d
+	addl	%ebx,%r14d
+	vpxor	%xmm7,%xmm6,%xmm6
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%ebx
+	movl	%r10d,%r12d
+	vpshufb	%xmm9,%xmm6,%xmm6
+	shrdl	$9,%r14d,%r14d
+	xorl	%r9d,%r13d
+	xorl	%r11d,%r12d
+	vpaddd	%xmm6,%xmm1,%xmm1
+	shrdl	$5,%r13d,%r13d
+	xorl	%ebx,%r14d
+	andl	%r9d,%r12d
+	vpaddd	32(%rbp),%xmm1,%xmm6
+	xorl	%r9d,%r13d
+	addl	28(%rsp),%eax
+	movl	%ebx,%edi
+	xorl	%r11d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%ecx,%edi
+	addl	%r12d,%eax
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%ebx,%r14d
+	addl	%r13d,%eax
+	xorl	%ecx,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%eax,%r8d
+	addl	%r15d,%eax
+	movl	%r8d,%r13d
+	addl	%eax,%r14d
+	vmovdqa	%xmm6,16(%rsp)
+	vpalignr	$4,%xmm2,%xmm3,%xmm4
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%eax
+	movl	%r9d,%r12d
+	vpalignr	$4,%xmm0,%xmm1,%xmm7
+	shrdl	$9,%r14d,%r14d
+	xorl	%r8d,%r13d
+	xorl	%r10d,%r12d
+	vpsrld	$7,%xmm4,%xmm6
+	shrdl	$5,%r13d,%r13d
+	xorl	%eax,%r14d
+	andl	%r8d,%r12d
+	vpaddd	%xmm7,%xmm2,%xmm2
+	xorl	%r8d,%r13d
+	addl	32(%rsp),%r11d
+	movl	%eax,%r15d
+	vpsrld	$3,%xmm4,%xmm7
+	xorl	%r10d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%ebx,%r15d
+	vpslld	$14,%xmm4,%xmm5
+	addl	%r12d,%r11d
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	vpxor	%xmm6,%xmm7,%xmm4
+	xorl	%eax,%r14d
+	addl	%r13d,%r11d
+	xorl	%ebx,%edi
+	vpshufd	$250,%xmm1,%xmm7
+	shrdl	$2,%r14d,%r14d
+	addl	%r11d,%edx
+	addl	%edi,%r11d
+	vpsrld	$11,%xmm6,%xmm6
+	movl	%edx,%r13d
+	addl	%r11d,%r14d
+	shrdl	$14,%r13d,%r13d
+	vpxor	%xmm5,%xmm4,%xmm4
+	movl	%r14d,%r11d
+	movl	%r8d,%r12d
+	shrdl	$9,%r14d,%r14d
+	vpslld	$11,%xmm5,%xmm5
+	xorl	%edx,%r13d
+	xorl	%r9d,%r12d
+	shrdl	$5,%r13d,%r13d
+	vpxor	%xmm6,%xmm4,%xmm4
+	xorl	%r11d,%r14d
+	andl	%edx,%r12d
+	xorl	%edx,%r13d
+	vpsrld	$10,%xmm7,%xmm6
+	addl	36(%rsp),%r10d
+	movl	%r11d,%edi
+	xorl	%r9d,%r12d
+	vpxor	%xmm5,%xmm4,%xmm4
+	shrdl	$11,%r14d,%r14d
+	xorl	%eax,%edi
+	addl	%r12d,%r10d
+	vpsrlq	$17,%xmm7,%xmm7
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%r11d,%r14d
+	vpaddd	%xmm4,%xmm2,%xmm2
+	addl	%r13d,%r10d
+	xorl	%eax,%r15d
+	shrdl	$2,%r14d,%r14d
+	vpxor	%xmm7,%xmm6,%xmm6
+	addl	%r10d,%ecx
+	addl	%r15d,%r10d
+	movl	%ecx,%r13d
+	vpsrlq	$2,%xmm7,%xmm7
+	addl	%r10d,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r10d
+	vpxor	%xmm7,%xmm6,%xmm6
+	movl	%edx,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%ecx,%r13d
+	vpshufb	%xmm8,%xmm6,%xmm6
+	xorl	%r8d,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%r10d,%r14d
+	vpaddd	%xmm6,%xmm2,%xmm2
+	andl	%ecx,%r12d
+	xorl	%ecx,%r13d
+	addl	40(%rsp),%r9d
+	vpshufd	$80,%xmm2,%xmm7
+	movl	%r10d,%r15d
+	xorl	%r8d,%r12d
+	shrdl	$11,%r14d,%r14d
+	vpsrld	$10,%xmm7,%xmm6
+	xorl	%r11d,%r15d
+	addl	%r12d,%r9d
+	shrdl	$6,%r13d,%r13d
+	vpsrlq	$17,%xmm7,%xmm7
+	andl	%r15d,%edi
+	xorl	%r10d,%r14d
+	addl	%r13d,%r9d
+	vpxor	%xmm7,%xmm6,%xmm6
+	xorl	%r11d,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%r9d,%ebx
+	vpsrlq	$2,%xmm7,%xmm7
+	addl	%edi,%r9d
+	movl	%ebx,%r13d
+	addl	%r9d,%r14d
+	vpxor	%xmm7,%xmm6,%xmm6
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r9d
+	movl	%ecx,%r12d
+	vpshufb	%xmm9,%xmm6,%xmm6
+	shrdl	$9,%r14d,%r14d
+	xorl	%ebx,%r13d
+	xorl	%edx,%r12d
+	vpaddd	%xmm6,%xmm2,%xmm2
+	shrdl	$5,%r13d,%r13d
+	xorl	%r9d,%r14d
+	andl	%ebx,%r12d
+	vpaddd	64(%rbp),%xmm2,%xmm6
+	xorl	%ebx,%r13d
+	addl	44(%rsp),%r8d
+	movl	%r9d,%edi
+	xorl	%edx,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r10d,%edi
+	addl	%r12d,%r8d
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%r9d,%r14d
+	addl	%r13d,%r8d
+	xorl	%r10d,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%r8d,%eax
+	addl	%r15d,%r8d
+	movl	%eax,%r13d
+	addl	%r8d,%r14d
+	vmovdqa	%xmm6,32(%rsp)
+	vpalignr	$4,%xmm3,%xmm0,%xmm4
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r8d
+	movl	%ebx,%r12d
+	vpalignr	$4,%xmm1,%xmm2,%xmm7
+	shrdl	$9,%r14d,%r14d
+	xorl	%eax,%r13d
+	xorl	%ecx,%r12d
+	vpsrld	$7,%xmm4,%xmm6
+	shrdl	$5,%r13d,%r13d
+	xorl	%r8d,%r14d
+	andl	%eax,%r12d
+	vpaddd	%xmm7,%xmm3,%xmm3
+	xorl	%eax,%r13d
+	addl	48(%rsp),%edx
+	movl	%r8d,%r15d
+	vpsrld	$3,%xmm4,%xmm7
+	xorl	%ecx,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r9d,%r15d
+	vpslld	$14,%xmm4,%xmm5
+	addl	%r12d,%edx
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	vpxor	%xmm6,%xmm7,%xmm4
+	xorl	%r8d,%r14d
+	addl	%r13d,%edx
+	xorl	%r9d,%edi
+	vpshufd	$250,%xmm2,%xmm7
+	shrdl	$2,%r14d,%r14d
+	addl	%edx,%r11d
+	addl	%edi,%edx
+	vpsrld	$11,%xmm6,%xmm6
+	movl	%r11d,%r13d
+	addl	%edx,%r14d
+	shrdl	$14,%r13d,%r13d
+	vpxor	%xmm5,%xmm4,%xmm4
+	movl	%r14d,%edx
+	movl	%eax,%r12d
+	shrdl	$9,%r14d,%r14d
+	vpslld	$11,%xmm5,%xmm5
+	xorl	%r11d,%r13d
+	xorl	%ebx,%r12d
+	shrdl	$5,%r13d,%r13d
+	vpxor	%xmm6,%xmm4,%xmm4
+	xorl	%edx,%r14d
+	andl	%r11d,%r12d
+	xorl	%r11d,%r13d
+	vpsrld	$10,%xmm7,%xmm6
+	addl	52(%rsp),%ecx
+	movl	%edx,%edi
+	xorl	%ebx,%r12d
+	vpxor	%xmm5,%xmm4,%xmm4
+	shrdl	$11,%r14d,%r14d
+	xorl	%r8d,%edi
+	addl	%r12d,%ecx
+	vpsrlq	$17,%xmm7,%xmm7
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%edx,%r14d
+	vpaddd	%xmm4,%xmm3,%xmm3
+	addl	%r13d,%ecx
+	xorl	%r8d,%r15d
+	shrdl	$2,%r14d,%r14d
+	vpxor	%xmm7,%xmm6,%xmm6
+	addl	%ecx,%r10d
+	addl	%r15d,%ecx
+	movl	%r10d,%r13d
+	vpsrlq	$2,%xmm7,%xmm7
+	addl	%ecx,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%ecx
+	vpxor	%xmm7,%xmm6,%xmm6
+	movl	%r11d,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%r10d,%r13d
+	vpshufb	%xmm8,%xmm6,%xmm6
+	xorl	%eax,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%ecx,%r14d
+	vpaddd	%xmm6,%xmm3,%xmm3
+	andl	%r10d,%r12d
+	xorl	%r10d,%r13d
+	addl	56(%rsp),%ebx
+	vpshufd	$80,%xmm3,%xmm7
+	movl	%ecx,%r15d
+	xorl	%eax,%r12d
+	shrdl	$11,%r14d,%r14d
+	vpsrld	$10,%xmm7,%xmm6
+	xorl	%edx,%r15d
+	addl	%r12d,%ebx
+	shrdl	$6,%r13d,%r13d
+	vpsrlq	$17,%xmm7,%xmm7
+	andl	%r15d,%edi
+	xorl	%ecx,%r14d
+	addl	%r13d,%ebx
+	vpxor	%xmm7,%xmm6,%xmm6
+	xorl	%edx,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%ebx,%r9d
+	vpsrlq	$2,%xmm7,%xmm7
+	addl	%edi,%ebx
+	movl	%r9d,%r13d
+	addl	%ebx,%r14d
+	vpxor	%xmm7,%xmm6,%xmm6
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%ebx
+	movl	%r10d,%r12d
+	vpshufb	%xmm9,%xmm6,%xmm6
+	shrdl	$9,%r14d,%r14d
+	xorl	%r9d,%r13d
+	xorl	%r11d,%r12d
+	vpaddd	%xmm6,%xmm3,%xmm3
+	shrdl	$5,%r13d,%r13d
+	xorl	%ebx,%r14d
+	andl	%r9d,%r12d
+	vpaddd	96(%rbp),%xmm3,%xmm6
+	xorl	%r9d,%r13d
+	addl	60(%rsp),%eax
+	movl	%ebx,%edi
+	xorl	%r11d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%ecx,%edi
+	addl	%r12d,%eax
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%ebx,%r14d
+	addl	%r13d,%eax
+	xorl	%ecx,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%eax,%r8d
+	addl	%r15d,%eax
+	movl	%r8d,%r13d
+	addl	%eax,%r14d
+	vmovdqa	%xmm6,48(%rsp)
+	cmpb	$0,131(%rbp)
+	jne	L$avx_00_47
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%eax
+	movl	%r9d,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%r8d,%r13d
+	xorl	%r10d,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%eax,%r14d
+	andl	%r8d,%r12d
+	xorl	%r8d,%r13d
+	addl	0(%rsp),%r11d
+	movl	%eax,%r15d
+	xorl	%r10d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%ebx,%r15d
+	addl	%r12d,%r11d
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	xorl	%eax,%r14d
+	addl	%r13d,%r11d
+	xorl	%ebx,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%r11d,%edx
+	addl	%edi,%r11d
+	movl	%edx,%r13d
+	addl	%r11d,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r11d
+	movl	%r8d,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%edx,%r13d
+	xorl	%r9d,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%r11d,%r14d
+	andl	%edx,%r12d
+	xorl	%edx,%r13d
+	addl	4(%rsp),%r10d
+	movl	%r11d,%edi
+	xorl	%r9d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%eax,%edi
+	addl	%r12d,%r10d
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%r11d,%r14d
+	addl	%r13d,%r10d
+	xorl	%eax,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%r10d,%ecx
+	addl	%r15d,%r10d
+	movl	%ecx,%r13d
+	addl	%r10d,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r10d
+	movl	%edx,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%ecx,%r13d
+	xorl	%r8d,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%r10d,%r14d
+	andl	%ecx,%r12d
+	xorl	%ecx,%r13d
+	addl	8(%rsp),%r9d
+	movl	%r10d,%r15d
+	xorl	%r8d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r11d,%r15d
+	addl	%r12d,%r9d
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	xorl	%r10d,%r14d
+	addl	%r13d,%r9d
+	xorl	%r11d,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%r9d,%ebx
+	addl	%edi,%r9d
+	movl	%ebx,%r13d
+	addl	%r9d,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r9d
+	movl	%ecx,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%ebx,%r13d
+	xorl	%edx,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%r9d,%r14d
+	andl	%ebx,%r12d
+	xorl	%ebx,%r13d
+	addl	12(%rsp),%r8d
+	movl	%r9d,%edi
+	xorl	%edx,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r10d,%edi
+	addl	%r12d,%r8d
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%r9d,%r14d
+	addl	%r13d,%r8d
+	xorl	%r10d,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%r8d,%eax
+	addl	%r15d,%r8d
+	movl	%eax,%r13d
+	addl	%r8d,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r8d
+	movl	%ebx,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%eax,%r13d
+	xorl	%ecx,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%r8d,%r14d
+	andl	%eax,%r12d
+	xorl	%eax,%r13d
+	addl	16(%rsp),%edx
+	movl	%r8d,%r15d
+	xorl	%ecx,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r9d,%r15d
+	addl	%r12d,%edx
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	xorl	%r8d,%r14d
+	addl	%r13d,%edx
+	xorl	%r9d,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%edx,%r11d
+	addl	%edi,%edx
+	movl	%r11d,%r13d
+	addl	%edx,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%edx
+	movl	%eax,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%r11d,%r13d
+	xorl	%ebx,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%edx,%r14d
+	andl	%r11d,%r12d
+	xorl	%r11d,%r13d
+	addl	20(%rsp),%ecx
+	movl	%edx,%edi
+	xorl	%ebx,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r8d,%edi
+	addl	%r12d,%ecx
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%edx,%r14d
+	addl	%r13d,%ecx
+	xorl	%r8d,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%ecx,%r10d
+	addl	%r15d,%ecx
+	movl	%r10d,%r13d
+	addl	%ecx,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%ecx
+	movl	%r11d,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%r10d,%r13d
+	xorl	%eax,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%ecx,%r14d
+	andl	%r10d,%r12d
+	xorl	%r10d,%r13d
+	addl	24(%rsp),%ebx
+	movl	%ecx,%r15d
+	xorl	%eax,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%edx,%r15d
+	addl	%r12d,%ebx
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	xorl	%ecx,%r14d
+	addl	%r13d,%ebx
+	xorl	%edx,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%ebx,%r9d
+	addl	%edi,%ebx
+	movl	%r9d,%r13d
+	addl	%ebx,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%ebx
+	movl	%r10d,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%r9d,%r13d
+	xorl	%r11d,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%ebx,%r14d
+	andl	%r9d,%r12d
+	xorl	%r9d,%r13d
+	addl	28(%rsp),%eax
+	movl	%ebx,%edi
+	xorl	%r11d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%ecx,%edi
+	addl	%r12d,%eax
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%ebx,%r14d
+	addl	%r13d,%eax
+	xorl	%ecx,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%eax,%r8d
+	addl	%r15d,%eax
+	movl	%r8d,%r13d
+	addl	%eax,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%eax
+	movl	%r9d,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%r8d,%r13d
+	xorl	%r10d,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%eax,%r14d
+	andl	%r8d,%r12d
+	xorl	%r8d,%r13d
+	addl	32(%rsp),%r11d
+	movl	%eax,%r15d
+	xorl	%r10d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%ebx,%r15d
+	addl	%r12d,%r11d
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	xorl	%eax,%r14d
+	addl	%r13d,%r11d
+	xorl	%ebx,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%r11d,%edx
+	addl	%edi,%r11d
+	movl	%edx,%r13d
+	addl	%r11d,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r11d
+	movl	%r8d,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%edx,%r13d
+	xorl	%r9d,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%r11d,%r14d
+	andl	%edx,%r12d
+	xorl	%edx,%r13d
+	addl	36(%rsp),%r10d
+	movl	%r11d,%edi
+	xorl	%r9d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%eax,%edi
+	addl	%r12d,%r10d
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%r11d,%r14d
+	addl	%r13d,%r10d
+	xorl	%eax,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%r10d,%ecx
+	addl	%r15d,%r10d
+	movl	%ecx,%r13d
+	addl	%r10d,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r10d
+	movl	%edx,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%ecx,%r13d
+	xorl	%r8d,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%r10d,%r14d
+	andl	%ecx,%r12d
+	xorl	%ecx,%r13d
+	addl	40(%rsp),%r9d
+	movl	%r10d,%r15d
+	xorl	%r8d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r11d,%r15d
+	addl	%r12d,%r9d
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	xorl	%r10d,%r14d
+	addl	%r13d,%r9d
+	xorl	%r11d,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%r9d,%ebx
+	addl	%edi,%r9d
+	movl	%ebx,%r13d
+	addl	%r9d,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r9d
+	movl	%ecx,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%ebx,%r13d
+	xorl	%edx,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%r9d,%r14d
+	andl	%ebx,%r12d
+	xorl	%ebx,%r13d
+	addl	44(%rsp),%r8d
+	movl	%r9d,%edi
+	xorl	%edx,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r10d,%edi
+	addl	%r12d,%r8d
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%r9d,%r14d
+	addl	%r13d,%r8d
+	xorl	%r10d,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%r8d,%eax
+	addl	%r15d,%r8d
+	movl	%eax,%r13d
+	addl	%r8d,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%r8d
+	movl	%ebx,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%eax,%r13d
+	xorl	%ecx,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%r8d,%r14d
+	andl	%eax,%r12d
+	xorl	%eax,%r13d
+	addl	48(%rsp),%edx
+	movl	%r8d,%r15d
+	xorl	%ecx,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r9d,%r15d
+	addl	%r12d,%edx
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	xorl	%r8d,%r14d
+	addl	%r13d,%edx
+	xorl	%r9d,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%edx,%r11d
+	addl	%edi,%edx
+	movl	%r11d,%r13d
+	addl	%edx,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%edx
+	movl	%eax,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%r11d,%r13d
+	xorl	%ebx,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%edx,%r14d
+	andl	%r11d,%r12d
+	xorl	%r11d,%r13d
+	addl	52(%rsp),%ecx
+	movl	%edx,%edi
+	xorl	%ebx,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%r8d,%edi
+	addl	%r12d,%ecx
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%edx,%r14d
+	addl	%r13d,%ecx
+	xorl	%r8d,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%ecx,%r10d
+	addl	%r15d,%ecx
+	movl	%r10d,%r13d
+	addl	%ecx,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%ecx
+	movl	%r11d,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%r10d,%r13d
+	xorl	%eax,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%ecx,%r14d
+	andl	%r10d,%r12d
+	xorl	%r10d,%r13d
+	addl	56(%rsp),%ebx
+	movl	%ecx,%r15d
+	xorl	%eax,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%edx,%r15d
+	addl	%r12d,%ebx
+	shrdl	$6,%r13d,%r13d
+	andl	%r15d,%edi
+	xorl	%ecx,%r14d
+	addl	%r13d,%ebx
+	xorl	%edx,%edi
+	shrdl	$2,%r14d,%r14d
+	addl	%ebx,%r9d
+	addl	%edi,%ebx
+	movl	%r9d,%r13d
+	addl	%ebx,%r14d
+	shrdl	$14,%r13d,%r13d
+	movl	%r14d,%ebx
+	movl	%r10d,%r12d
+	shrdl	$9,%r14d,%r14d
+	xorl	%r9d,%r13d
+	xorl	%r11d,%r12d
+	shrdl	$5,%r13d,%r13d
+	xorl	%ebx,%r14d
+	andl	%r9d,%r12d
+	xorl	%r9d,%r13d
+	addl	60(%rsp),%eax
+	movl	%ebx,%edi
+	xorl	%r11d,%r12d
+	shrdl	$11,%r14d,%r14d
+	xorl	%ecx,%edi
+	addl	%r12d,%eax
+	shrdl	$6,%r13d,%r13d
+	andl	%edi,%r15d
+	xorl	%ebx,%r14d
+	addl	%r13d,%eax
+	xorl	%ecx,%r15d
+	shrdl	$2,%r14d,%r14d
+	addl	%eax,%r8d
+	addl	%r15d,%eax
+	movl	%r8d,%r13d
+	addl	%eax,%r14d
+	movq	64+0(%rsp),%rdi
+	movl	%r14d,%eax
+
+	addl	0(%rdi),%eax
+	leaq	64(%rsi),%rsi
+	addl	4(%rdi),%ebx
+	addl	8(%rdi),%ecx
+	addl	12(%rdi),%edx
+	addl	16(%rdi),%r8d
+	addl	20(%rdi),%r9d
+	addl	24(%rdi),%r10d
+	addl	28(%rdi),%r11d
+
+	cmpq	64+16(%rsp),%rsi
+
+	movl	%eax,0(%rdi)
+	movl	%ebx,4(%rdi)
+	movl	%ecx,8(%rdi)
+	movl	%edx,12(%rdi)
+	movl	%r8d,16(%rdi)
+	movl	%r9d,20(%rdi)
+	movl	%r10d,24(%rdi)
+	movl	%r11d,28(%rdi)
+	jb	L$loop_avx
+
+	movq	88(%rsp),%rsi
+
+	vzeroupper
+	movq	-48(%rsi),%r15
+
+	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:
+	ret
+
+
+#endif