X-Git-Url: https://git.openssl.org/?p=openssl.git;a=blobdiff_plain;f=crypto%2Fsha%2Fasm%2Fsha512-x86_64.pl;h=e7140defb6c6ae981c6152db91227b98f4b8c246;hp=7589e6e6f4c4b6f981cbd89816da4005f2a32a60;hb=b0d3442efc10b635863b915c2d014345f6e5a219;hpb=147cca8f530b8ddf22087e5709927082e095a062

diff --git a/crypto/sha/asm/sha512-x86_64.pl b/crypto/sha/asm/sha512-x86_64.pl
index 7589e6e6f4..e7140defb6 100755
--- a/crypto/sha/asm/sha512-x86_64.pl
+++ b/crypto/sha/asm/sha512-x86_64.pl
@@ -1,9 +1,16 @@
-#!/usr/bin/env perl
+#! /usr/bin/env perl
+# Copyright 2005-2016 The OpenSSL Project Authors. All Rights Reserved.
+#
+# Licensed under the Apache License 2.0 (the "License").  You may not use
+# this file except in compliance with the License.  You can obtain a copy
+# in the file LICENSE in the source distribution or at
+# https://www.openssl.org/source/license.html
+
 #
 # ====================================================================
 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
 # project. Rights for redistribution and usage in source and binary
-# forms are granted according to the OpenSSL license.
+# forms are granted according to the License.
 # ====================================================================
 #
 # sha256/512_block procedure for x86_64.
@@ -34,7 +41,7 @@
 # level parallelism, on a given CPU implementation in this case.
 #
 # Special note on Intel EM64T. While Opteron CPU exhibits perfect
-# perfromance ratio of 1.5 between 64- and 32-bit flavors [see above],
+# performance ratio of 1.5 between 64- and 32-bit flavors [see above],
 # [currently available] EM64T CPUs apparently are far from it. On the
 # contrary, 64-bit version, sha512_block, is ~30% *slower* than 32-bit
 # sha256_block:-( This is presumably because 64-bit shifts/rotates
@@ -67,7 +74,12 @@
 # significant 128-bit halves and data from second to most significant.
 # The data is then processed with same SIMD instruction sequence as
 # for AVX, but with %ymm as operands. Side effect is increased stack
-# frame, 448 additional bytes in SHA256 and 1152 in SHA512.
+# frame, 448 additional bytes in SHA256 and 1152 in SHA512, and 1.2KB
+# code size increase.
+#
+# March 2014.
+#
+# Add support for Intel SHA Extensions.
 
 ######################################################################
 # Current performance in cycles per processed byte (less is better):
@@ -81,20 +93,26 @@
 # Sandy Bridge	17.4	14.2(+23%)  11.6(+50%(**))  11.2    8.10(+38%(**))
 # Ivy Bridge	12.6	10.5(+20%)  10.3(+22%)	    8.17    7.22(+13%)
 # Haswell	12.2	9.28(+31%)  7.80(+56%)	    7.66    5.40(+42%)
+# Skylake	11.4	9.03(+26%)  7.70(+48%)      7.25    5.20(+40%)
 # Bulldozer	21.1	13.6(+54%)  13.6(+54%(***)) 13.5    8.58(+57%)
+# Ryzen		11.0	9.02(+22%)  2.05(+440%)     7.05    5.67(+20%)
 # VIA Nano	23.0	16.5(+39%)  -		    14.7    -
 # Atom		23.0	18.9(+22%)  -		    14.7    -
+# Silvermont	27.4	20.6(+33%)  -               17.5    -
+# Knights L	27.4	21.0(+30%)  19.6(+40%)	    17.5    12.8(+37%)
+# Goldmont	18.9	14.3(+32%)  4.16(+350%)     12.0    -
 #
-# (*)	whichever best applicable;
+# (*)	whichever best applicable, including SHAEXT;
 # (**)	switch from ror to shrd stands for fair share of improvement;
 # (***)	execution time is fully determined by remaining integer-only
 #	part, body_00_15; reducing the amount of SIMD instructions
 #	below certain limit makes no difference/sense; to conserve
 #	space SHA256 XOP code path is therefore omitted;
 
-$flavour = shift;
-$output  = shift;
-if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
+# $output is the last argument if it looks like a file (it has an extension)
+# $flavour is the first argument if it doesn't look like a file
+$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
+$flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
 
 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
 
@@ -118,7 +136,15 @@ if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
 	$avx = ($1>=10) + ($1>=11);
 }
 
-open OUT,"| \"$^X\" $xlate $flavour $output";
+if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([3-9]\.[0-9]+)/) {
+	$avx = ($2>=3.0) + ($2>3.0);
+}
+
+$shaext=1;	### set to zero if compiling for 1.0.1
+$avx=1		if (!$shaext && $avx);
+
+open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""
+    or die "can't call $xlate: $!";
 *STDOUT=*OUT;
 
 if ($output =~ /512/) {
@@ -154,7 +180,7 @@ $Tbl="%rbp";
 $_ctx="16*$SZ+0*8(%rsp)";
 $_inp="16*$SZ+1*8(%rsp)";
 $_end="16*$SZ+2*8(%rsp)";
-$_rsp="16*$SZ+3*8(%rsp)";
+$_rsp="`16*$SZ+3*8`(%rsp)";
 $framesz="16*$SZ+4*8";
 
 
@@ -247,6 +273,7 @@ $code=<<___;
 .type	$func,\@function,3
 .align	16
 $func:
+.cfi_startproc
 ___
 $code.=<<___ if ($SZ==4 || $avx);
 	lea	OPENSSL_ia32cap_P(%rip),%r11
@@ -254,6 +281,10 @@ $code.=<<___ if ($SZ==4 || $avx);
 	mov	4(%r11),%r10d
 	mov	8(%r11),%r11d
 ___
+$code.=<<___ if ($SZ==4 && $shaext);
+	test	\$`1<<29`,%r11d		# check for SHA
+	jnz	_shaext_shortcut
+___
 $code.=<<___ if ($avx && $SZ==8);
 	test	\$`1<<11`,%r10d		# check for XOP
 	jnz	.Lxop_shortcut
@@ -275,13 +306,20 @@ $code.=<<___ if ($SZ==4);
 	jnz	.Lssse3_shortcut
 ___
 $code.=<<___;
+	mov	%rsp,%rax		# copy %rsp
+.cfi_def_cfa_register	%rax
 	push	%rbx
+.cfi_push	%rbx
 	push	%rbp
+.cfi_push	%rbp
 	push	%r12
+.cfi_push	%r12
 	push	%r13
+.cfi_push	%r13
 	push	%r14
+.cfi_push	%r14
 	push	%r15
-	mov	%rsp,%r11		# copy %rsp
+.cfi_push	%r15
 	shl	\$4,%rdx		# num*16
 	sub	\$$framesz,%rsp
 	lea	($inp,%rdx,$SZ),%rdx	# inp+num*16*$SZ
@@ -289,7 +327,8 @@ $code.=<<___;
 	mov	$ctx,$_ctx		# save ctx, 1st arg
 	mov	$inp,$_inp		# save inp, 2nd arh
 	mov	%rdx,$_end		# save end pointer, "3rd" arg
-	mov	%r11,$_rsp		# save copy of %rsp
+	mov	%rax,$_rsp		# save copy of %rsp
+.cfi_cfa_expression	$_rsp,deref,+8
 .Lprologue:
 
 	mov	$SZ*0($ctx),$A
@@ -356,15 +395,24 @@ $code.=<<___;
 	jb	.Lloop
 
 	mov	$_rsp,%rsi
-	mov	(%rsi),%r15
-	mov	8(%rsi),%r14
-	mov	16(%rsi),%r13
-	mov	24(%rsi),%r12
-	mov	32(%rsi),%rbp
-	mov	40(%rsi),%rbx
-	lea	48(%rsi),%rsp
+.cfi_def_cfa	%rsi,8
+	mov	-48(%rsi),%r15
+.cfi_restore	%r15
+	mov	-40(%rsi),%r14
+.cfi_restore	%r14
+	mov	-32(%rsi),%r13
+.cfi_restore	%r13
+	mov	-24(%rsi),%r12
+.cfi_restore	%r12
+	mov	-16(%rsi),%rbp
+.cfi_restore	%rbp
+	mov	-8(%rsi),%rbx
+.cfi_restore	%rbx
+	lea	(%rsi),%rsp
+.cfi_def_cfa_register	%rsp
 .Lepilogue:
 	ret
+.cfi_endproc
 .size	$func,.-$func
 ___
 
@@ -509,6 +557,168 @@ ___
 ######################################################################
 # SIMD code paths
 #
+if ($SZ==4 && $shaext) {{{
+######################################################################
+# Intel SHA Extensions implementation of SHA256 update function.
+#
+my ($ctx,$inp,$num,$Tbl)=("%rdi","%rsi","%rdx","%rcx");
+
+my ($Wi,$ABEF,$CDGH,$TMP,$BSWAP,$ABEF_SAVE,$CDGH_SAVE)=map("%xmm$_",(0..2,7..10));
+my @MSG=map("%xmm$_",(3..6));
+
+$code.=<<___;
+.type	sha256_block_data_order_shaext,\@function,3
+.align	64
+sha256_block_data_order_shaext:
+_shaext_shortcut:
+.cfi_startproc
+___
+$code.=<<___ if ($win64);
+	lea	`-8-5*16`(%rsp),%rsp
+	movaps	%xmm6,-8-5*16(%rax)
+	movaps	%xmm7,-8-4*16(%rax)
+	movaps	%xmm8,-8-3*16(%rax)
+	movaps	%xmm9,-8-2*16(%rax)
+	movaps	%xmm10,-8-1*16(%rax)
+.Lprologue_shaext:
+___
+$code.=<<___;
+	lea		K256+0x80(%rip),$Tbl
+	movdqu		($ctx),$ABEF		# DCBA
+	movdqu		16($ctx),$CDGH		# HGFE
+	movdqa		0x200-0x80($Tbl),$TMP	# byte swap mask
+
+	pshufd		\$0x1b,$ABEF,$Wi	# ABCD
+	pshufd		\$0xb1,$ABEF,$ABEF	# CDAB
+	pshufd		\$0x1b,$CDGH,$CDGH	# EFGH
+	movdqa		$TMP,$BSWAP		# offload
+	palignr		\$8,$CDGH,$ABEF		# ABEF
+	punpcklqdq	$Wi,$CDGH		# CDGH
+	jmp		.Loop_shaext
+
+.align	16
+.Loop_shaext:
+	movdqu		($inp),@MSG[0]
+	movdqu		0x10($inp),@MSG[1]
+	movdqu		0x20($inp),@MSG[2]
+	pshufb		$TMP,@MSG[0]
+	movdqu		0x30($inp),@MSG[3]
+
+	movdqa		0*32-0x80($Tbl),$Wi
+	paddd		@MSG[0],$Wi
+	pshufb		$TMP,@MSG[1]
+	movdqa		$CDGH,$CDGH_SAVE	# offload
+	sha256rnds2	$ABEF,$CDGH		# 0-3
+	pshufd		\$0x0e,$Wi,$Wi
+	nop
+	movdqa		$ABEF,$ABEF_SAVE	# offload
+	sha256rnds2	$CDGH,$ABEF
+
+	movdqa		1*32-0x80($Tbl),$Wi
+	paddd		@MSG[1],$Wi
+	pshufb		$TMP,@MSG[2]
+	sha256rnds2	$ABEF,$CDGH		# 4-7
+	pshufd		\$0x0e,$Wi,$Wi
+	lea		0x40($inp),$inp
+	sha256msg1	@MSG[1],@MSG[0]
+	sha256rnds2	$CDGH,$ABEF
+
+	movdqa		2*32-0x80($Tbl),$Wi
+	paddd		@MSG[2],$Wi
+	pshufb		$TMP,@MSG[3]
+	sha256rnds2	$ABEF,$CDGH		# 8-11
+	pshufd		\$0x0e,$Wi,$Wi
+	movdqa		@MSG[3],$TMP
+	palignr		\$4,@MSG[2],$TMP
+	nop
+	paddd		$TMP,@MSG[0]
+	sha256msg1	@MSG[2],@MSG[1]
+	sha256rnds2	$CDGH,$ABEF
+
+	movdqa		3*32-0x80($Tbl),$Wi
+	paddd		@MSG[3],$Wi
+	sha256msg2	@MSG[3],@MSG[0]
+	sha256rnds2	$ABEF,$CDGH		# 12-15
+	pshufd		\$0x0e,$Wi,$Wi
+	movdqa		@MSG[0],$TMP
+	palignr		\$4,@MSG[3],$TMP
+	nop
+	paddd		$TMP,@MSG[1]
+	sha256msg1	@MSG[3],@MSG[2]
+	sha256rnds2	$CDGH,$ABEF
+___
+for($i=4;$i<16-3;$i++) {
+$code.=<<___;
+	movdqa		$i*32-0x80($Tbl),$Wi
+	paddd		@MSG[0],$Wi
+	sha256msg2	@MSG[0],@MSG[1]
+	sha256rnds2	$ABEF,$CDGH		# 16-19...
+	pshufd		\$0x0e,$Wi,$Wi
+	movdqa		@MSG[1],$TMP
+	palignr		\$4,@MSG[0],$TMP
+	nop
+	paddd		$TMP,@MSG[2]
+	sha256msg1	@MSG[0],@MSG[3]
+	sha256rnds2	$CDGH,$ABEF
+___
+	push(@MSG,shift(@MSG));
+}
+$code.=<<___;
+	movdqa		13*32-0x80($Tbl),$Wi
+	paddd		@MSG[0],$Wi
+	sha256msg2	@MSG[0],@MSG[1]
+	sha256rnds2	$ABEF,$CDGH		# 52-55
+	pshufd		\$0x0e,$Wi,$Wi
+	movdqa		@MSG[1],$TMP
+	palignr		\$4,@MSG[0],$TMP
+	sha256rnds2	$CDGH,$ABEF
+	paddd		$TMP,@MSG[2]
+
+	movdqa		14*32-0x80($Tbl),$Wi
+	paddd		@MSG[1],$Wi
+	sha256rnds2	$ABEF,$CDGH		# 56-59
+	pshufd		\$0x0e,$Wi,$Wi
+	sha256msg2	@MSG[1],@MSG[2]
+	movdqa		$BSWAP,$TMP
+	sha256rnds2	$CDGH,$ABEF
+
+	movdqa		15*32-0x80($Tbl),$Wi
+	paddd		@MSG[2],$Wi
+	nop
+	sha256rnds2	$ABEF,$CDGH		# 60-63
+	pshufd		\$0x0e,$Wi,$Wi
+	dec		$num
+	nop
+	sha256rnds2	$CDGH,$ABEF
+
+	paddd		$CDGH_SAVE,$CDGH
+	paddd		$ABEF_SAVE,$ABEF
+	jnz		.Loop_shaext
+
+	pshufd		\$0xb1,$CDGH,$CDGH	# DCHG
+	pshufd		\$0x1b,$ABEF,$TMP	# FEBA
+	pshufd		\$0xb1,$ABEF,$ABEF	# BAFE
+	punpckhqdq	$CDGH,$ABEF		# DCBA
+	palignr		\$8,$TMP,$CDGH		# HGFE
+
+	movdqu	$ABEF,($ctx)
+	movdqu	$CDGH,16($ctx)
+___
+$code.=<<___ if ($win64);
+	movaps	-8-5*16(%rax),%xmm6
+	movaps	-8-4*16(%rax),%xmm7
+	movaps	-8-3*16(%rax),%xmm8
+	movaps	-8-2*16(%rax),%xmm9
+	movaps	-8-1*16(%rax),%xmm10
+	mov	%rax,%rsp
+.Lepilogue_shaext:
+___
+$code.=<<___;
+	ret
+.cfi_endproc
+.size	sha256_block_data_order_shaext,.-sha256_block_data_order_shaext
+___
+}}}
 {{{
 
 my $a4=$T1;
@@ -574,14 +784,22 @@ $code.=<<___;
 .type	${func}_ssse3,\@function,3
 .align	64
 ${func}_ssse3:
+.cfi_startproc
 .Lssse3_shortcut:
+	mov	%rsp,%rax		# copy %rsp
+.cfi_def_cfa_register	%rax
 	push	%rbx
+.cfi_push	%rbx
 	push	%rbp
+.cfi_push	%rbp
 	push	%r12
+.cfi_push	%r12
 	push	%r13
+.cfi_push	%r13
 	push	%r14
+.cfi_push	%r14
 	push	%r15
-	mov	%rsp,%r11		# copy %rsp
+.cfi_push	%r15
 	shl	\$4,%rdx		# num*16
 	sub	\$`$framesz+$win64*16*4`,%rsp
 	lea	($inp,%rdx,$SZ),%rdx	# inp+num*16*$SZ
@@ -589,7 +807,8 @@ ${func}_ssse3:
 	mov	$ctx,$_ctx		# save ctx, 1st arg
 	mov	$inp,$_inp		# save inp, 2nd arh
 	mov	%rdx,$_end		# save end pointer, "3rd" arg
-	mov	%r11,$_rsp		# save copy of %rsp
+	mov	%rax,$_rsp		# save copy of %rsp
+.cfi_cfa_expression	$_rsp,deref,+8
 ___
 $code.=<<___ if ($win64);
 	movaps	%xmm6,16*$SZ+32(%rsp)
@@ -620,13 +839,13 @@ $code.=<<___;
 	movdqu	0x00($inp),@X[0]
 	movdqu	0x10($inp),@X[1]
 	movdqu	0x20($inp),@X[2]
-	movdqu	0x30($inp),@X[3]
 	pshufb	$t3,@X[0]
+	movdqu	0x30($inp),@X[3]
 	lea	$TABLE(%rip),$Tbl
 	pshufb	$t3,@X[1]
 	movdqa	0x00($Tbl),$t0
-	pshufb	$t3,@X[2]
 	movdqa	0x20($Tbl),$t1
+	pshufb	$t3,@X[2]
 	paddd	@X[0],$t0
 	movdqa	0x40($Tbl),$t2
 	pshufb	$t3,@X[3]
@@ -888,6 +1107,7 @@ $code.=<<___;
 	jb	.Lloop_ssse3
 
 	mov	$_rsp,%rsi
+.cfi_def_cfa	%rsi,8
 ___
 $code.=<<___ if ($win64);
 	movaps	16*$SZ+32(%rsp),%xmm6
@@ -896,15 +1116,23 @@ $code.=<<___ if ($win64);
 	movaps	16*$SZ+80(%rsp),%xmm9
 ___
 $code.=<<___;
-	mov	(%rsi),%r15
-	mov	8(%rsi),%r14
-	mov	16(%rsi),%r13
-	mov	24(%rsi),%r12
-	mov	32(%rsi),%rbp
-	mov	40(%rsi),%rbx
-	lea	48(%rsi),%rsp
+	mov	-48(%rsi),%r15
+.cfi_restore	%r15
+	mov	-40(%rsi),%r14
+.cfi_restore	%r14
+	mov	-32(%rsi),%r13
+.cfi_restore	%r13
+	mov	-24(%rsi),%r12
+.cfi_restore	%r12
+	mov	-16(%rsi),%rbp
+.cfi_restore	%rbp
+	mov	-8(%rsi),%rbx
+.cfi_restore	%rbx
+	lea	(%rsi),%rsp
+.cfi_def_cfa_register	%rsp
 .Lepilogue_ssse3:
 	ret
+.cfi_endproc
 .size	${func}_ssse3,.-${func}_ssse3
 ___
 }
@@ -918,14 +1146,22 @@ $code.=<<___;
 .type	${func}_xop,\@function,3
 .align	64
 ${func}_xop:
+.cfi_startproc
 .Lxop_shortcut:
+	mov	%rsp,%rax		# copy %rsp
+.cfi_def_cfa_register	%rax
 	push	%rbx
+.cfi_push	%rbx
 	push	%rbp
+.cfi_push	%rbp
 	push	%r12
+.cfi_push	%r12
 	push	%r13
+.cfi_push	%r13
 	push	%r14
+.cfi_push	%r14
 	push	%r15
-	mov	%rsp,%r11		# copy %rsp
+.cfi_push	%r15
 	shl	\$4,%rdx		# num*16
 	sub	\$`$framesz+$win64*16*($SZ==4?4:6)`,%rsp
 	lea	($inp,%rdx,$SZ),%rdx	# inp+num*16*$SZ
@@ -933,7 +1169,8 @@ ${func}_xop:
 	mov	$ctx,$_ctx		# save ctx, 1st arg
 	mov	$inp,$_inp		# save inp, 2nd arh
 	mov	%rdx,$_end		# save end pointer, "3rd" arg
-	mov	%r11,$_rsp		# save copy of %rsp
+	mov	%rax,$_rsp		# save copy of %rsp
+.cfi_cfa_expression	$_rsp,deref,+8
 ___
 $code.=<<___ if ($win64);
 	movaps	%xmm6,16*$SZ+32(%rsp)
@@ -1260,6 +1497,7 @@ $code.=<<___;
 	jb	.Lloop_xop
 
 	mov	$_rsp,%rsi
+.cfi_def_cfa	%rsi,8
 	vzeroupper
 ___
 $code.=<<___ if ($win64);
@@ -1273,15 +1511,23 @@ $code.=<<___ if ($win64 && $SZ>4);
 	movaps	16*$SZ+112(%rsp),%xmm11
 ___
 $code.=<<___;
-	mov	(%rsi),%r15
-	mov	8(%rsi),%r14
-	mov	16(%rsi),%r13
-	mov	24(%rsi),%r12
-	mov	32(%rsi),%rbp
-	mov	40(%rsi),%rbx
-	lea	48(%rsi),%rsp
+	mov	-48(%rsi),%r15
+.cfi_restore	%r15
+	mov	-40(%rsi),%r14
+.cfi_restore	%r14
+	mov	-32(%rsi),%r13
+.cfi_restore	%r13
+	mov	-24(%rsi),%r12
+.cfi_restore	%r12
+	mov	-16(%rsi),%rbp
+.cfi_restore	%rbp
+	mov	-8(%rsi),%rbx
+.cfi_restore	%rbx
+	lea	(%rsi),%rsp
+.cfi_def_cfa_register	%rsp
 .Lepilogue_xop:
 	ret
+.cfi_endproc
 .size	${func}_xop,.-${func}_xop
 ___
 }
@@ -1294,14 +1540,22 @@ $code.=<<___;
 .type	${func}_avx,\@function,3
 .align	64
 ${func}_avx:
+.cfi_startproc
 .Lavx_shortcut:
+	mov	%rsp,%rax		# copy %rsp
+.cfi_def_cfa_register	%rax
 	push	%rbx
+.cfi_push	%rbx
 	push	%rbp
+.cfi_push	%rbp
 	push	%r12
+.cfi_push	%r12
 	push	%r13
+.cfi_push	%r13
 	push	%r14
+.cfi_push	%r14
 	push	%r15
-	mov	%rsp,%r11		# copy %rsp
+.cfi_push	%r15
 	shl	\$4,%rdx		# num*16
 	sub	\$`$framesz+$win64*16*($SZ==4?4:6)`,%rsp
 	lea	($inp,%rdx,$SZ),%rdx	# inp+num*16*$SZ
@@ -1309,7 +1563,8 @@ ${func}_avx:
 	mov	$ctx,$_ctx		# save ctx, 1st arg
 	mov	$inp,$_inp		# save inp, 2nd arh
 	mov	%rdx,$_end		# save end pointer, "3rd" arg
-	mov	%r11,$_rsp		# save copy of %rsp
+	mov	%rax,$_rsp		# save copy of %rsp
+.cfi_cfa_expression	$_rsp,deref,+8
 ___
 $code.=<<___ if ($win64);
 	movaps	%xmm6,16*$SZ+32(%rsp)
@@ -1568,6 +1823,7 @@ $code.=<<___;
 	jb	.Lloop_avx
 
 	mov	$_rsp,%rsi
+.cfi_def_cfa	%rsi,8
 	vzeroupper
 ___
 $code.=<<___ if ($win64);
@@ -1581,15 +1837,23 @@ $code.=<<___ if ($win64 && $SZ>4);
 	movaps	16*$SZ+112(%rsp),%xmm11
 ___
 $code.=<<___;
-	mov	(%rsi),%r15
-	mov	8(%rsi),%r14
-	mov	16(%rsi),%r13
-	mov	24(%rsi),%r12
-	mov	32(%rsi),%rbp
-	mov	40(%rsi),%rbx
-	lea	48(%rsi),%rsp
+	mov	-48(%rsi),%r15
+.cfi_restore	%r15
+	mov	-40(%rsi),%r14
+.cfi_restore	%r14
+	mov	-32(%rsi),%r13
+.cfi_restore	%r13
+	mov	-24(%rsi),%r12
+.cfi_restore	%r12
+	mov	-16(%rsi),%rbp
+.cfi_restore	%rbp
+	mov	-8(%rsi),%rbx
+.cfi_restore	%rbx
+	lea	(%rsi),%rsp
+.cfi_def_cfa_register	%rsp
 .Lepilogue_avx:
 	ret
+.cfi_endproc
 .size	${func}_avx,.-${func}_avx
 ___
 
@@ -1597,7 +1861,7 @@ if ($avx>1) {{
 ######################################################################
 # AVX2+BMI code path
 #
-my $a5=$SZ==4?"%esi":"%rsi";	# zap $inp 
+my $a5=$SZ==4?"%esi":"%rsi";	# zap $inp
 my $PUSH8=8*2*$SZ;
 use integer;
 
@@ -1645,14 +1909,22 @@ $code.=<<___;
 .type	${func}_avx2,\@function,3
 .align	64
 ${func}_avx2:
+.cfi_startproc
 .Lavx2_shortcut:
+	mov	%rsp,%rax		# copy %rsp
+.cfi_def_cfa_register	%rax
 	push	%rbx
+.cfi_push	%rbx
 	push	%rbp
+.cfi_push	%rbp
 	push	%r12
+.cfi_push	%r12
 	push	%r13
+.cfi_push	%r13
 	push	%r14
+.cfi_push	%r14
 	push	%r15
-	mov	%rsp,%r11		# copy %rsp
+.cfi_push	%r15
 	sub	\$`2*$SZ*$rounds+4*8+$win64*16*($SZ==4?4:6)`,%rsp
 	shl	\$4,%rdx		# num*16
 	and	\$-256*$SZ,%rsp		# align stack frame
@@ -1661,7 +1933,8 @@ ${func}_avx2:
 	mov	$ctx,$_ctx		# save ctx, 1st arg
 	mov	$inp,$_inp		# save inp, 2nd arh
 	mov	%rdx,$_end		# save end pointer, "3rd" arg
-	mov	%r11,$_rsp		# save copy of %rsp
+	mov	%rax,$_rsp		# save copy of %rsp
+.cfi_cfa_expression	$_rsp,deref,+8
 ___
 $code.=<<___ if ($win64);
 	movaps	%xmm6,16*$SZ+32(%rsp)
@@ -1723,7 +1996,23 @@ $code.=<<___;
 	vmovdqa	$t0,0x00(%rsp)
 	xor	$a1,$a1
 	vmovdqa	$t1,0x20(%rsp)
+___
+$code.=<<___ if (!$win64);
+# temporarily use %rdi as frame pointer
+	mov	$_rsp,%rdi
+.cfi_def_cfa	%rdi,8
+___
+$code.=<<___;
 	lea	-$PUSH8(%rsp),%rsp
+___
+$code.=<<___ if (!$win64);
+# the frame info is at $_rsp, but the stack is moving...
+# so a second frame pointer is saved at -8(%rsp)
+# that is in the red zone
+	mov	%rdi,-8(%rsp)
+.cfi_cfa_expression	%rsp-8,deref,+8
+___
+$code.=<<___;
 	mov	$B,$a3
 	vmovdqa	$t2,0x00(%rsp)
 	xor	$C,$a3			# magic
@@ -1743,7 +2032,18 @@ my @X = @_;
 my @insns = (&$body,&$body,&$body,&$body);	# 96 instructions
 my $base = "+2*$PUSH8(%rsp)";
 
-	&lea	("%rsp","-$PUSH8(%rsp)")	if (($j%2)==0);
+	if (($j%2)==0) {
+	&lea	("%rsp","-$PUSH8(%rsp)");
+$code.=<<___ if (!$win64);
+.cfi_cfa_expression	%rsp+`$PUSH8-8`,deref,+8
+# copy secondary frame pointer to new location again at -8(%rsp)
+	pushq	$PUSH8-8(%rsp)
+.cfi_cfa_expression	%rsp,deref,+8
+	lea	8(%rsp),%rsp
+.cfi_cfa_expression	%rsp-8,deref,+8
+___
+	}
+
 	foreach (Xupdate_256_AVX()) {		# 29 instructions
 	    eval;
 	    eval(shift(@insns));
@@ -1814,7 +2114,23 @@ $code.=<<___;
 	vmovdqa	$t2,0x40(%rsp)
 	vpaddq	0x40($Tbl),@X[6],$t2
 	vmovdqa	$t3,0x60(%rsp)
+___
+$code.=<<___ if (!$win64);
+# temporarily use %rdi as frame pointer
+	mov	$_rsp,%rdi
+.cfi_def_cfa	%rdi,8
+___
+$code.=<<___;
 	lea	-$PUSH8(%rsp),%rsp
+___
+$code.=<<___ if (!$win64);
+# the frame info is at $_rsp, but the stack is moving...
+# so a second frame pointer is saved at -8(%rsp)
+# that is in the red zone
+	mov	%rdi,-8(%rsp)
+.cfi_cfa_expression	%rsp-8,deref,+8
+___
+$code.=<<___;
 	vpaddq	0x60($Tbl),@X[7],$t3
 	vmovdqa	$t0,0x00(%rsp)
 	xor	$a1,$a1
@@ -1838,7 +2154,18 @@ my @X = @_;
 my @insns = (&$body,&$body);			# 48 instructions
 my $base = "+2*$PUSH8(%rsp)";
 
-	&lea	("%rsp","-$PUSH8(%rsp)")	if (($j%4)==0);
+	if (($j%4)==0) {
+	&lea	("%rsp","-$PUSH8(%rsp)");
+$code.=<<___ if (!$win64);
+.cfi_cfa_expression	%rsp+`$PUSH8-8`,deref,+8
+# copy secondary frame pointer to new location again at -8(%rsp)
+	pushq	$PUSH8-8(%rsp)
+.cfi_cfa_expression	%rsp,deref,+8
+	lea	8(%rsp),%rsp
+.cfi_cfa_expression	%rsp-8,deref,+8
+___
+	}
+
 	foreach (Xupdate_512_AVX()) {		# 23 instructions
 	    eval;
 	    if ($_ !~ /\;$/) {
@@ -1913,6 +2240,8 @@ $code.=<<___;
 	add	$a1,$A
 	#mov	`2*$SZ*$rounds+8`(%rsp),$inp	# $_inp
 	lea	`2*$SZ*($rounds-8)`(%rsp),%rsp
+# restore frame pointer to original location at $_rsp
+.cfi_cfa_expression	$_rsp,deref,+8
 
 	add	$SZ*0($ctx),$A
 	add	$SZ*1($ctx),$B
@@ -1938,32 +2267,43 @@ $code.=<<___;
 
 	jbe	.Loop_avx2
 	lea	(%rsp),$Tbl
+# temporarily use $Tbl as index to $_rsp
+# this avoids the need to save a secondary frame pointer at -8(%rsp)
+.cfi_cfa_expression	$Tbl+`16*$SZ+3*8`,deref,+8
 
 .Ldone_avx2:
-	lea	($Tbl),%rsp
-	mov	$_rsp,%rsi
+	mov	`16*$SZ+3*8`($Tbl),%rsi
+.cfi_def_cfa	%rsi,8
 	vzeroupper
 ___
 $code.=<<___ if ($win64);
-	movaps	16*$SZ+32(%rsp),%xmm6
-	movaps	16*$SZ+48(%rsp),%xmm7
-	movaps	16*$SZ+64(%rsp),%xmm8
-	movaps	16*$SZ+80(%rsp),%xmm9
+	movaps	16*$SZ+32($Tbl),%xmm6
+	movaps	16*$SZ+48($Tbl),%xmm7
+	movaps	16*$SZ+64($Tbl),%xmm8
+	movaps	16*$SZ+80($Tbl),%xmm9
 ___
 $code.=<<___ if ($win64 && $SZ>4);
-	movaps	16*$SZ+96(%rsp),%xmm10
-	movaps	16*$SZ+112(%rsp),%xmm11
+	movaps	16*$SZ+96($Tbl),%xmm10
+	movaps	16*$SZ+112($Tbl),%xmm11
 ___
 $code.=<<___;
-	mov	(%rsi),%r15
-	mov	8(%rsi),%r14
-	mov	16(%rsi),%r13
-	mov	24(%rsi),%r12
-	mov	32(%rsi),%rbp
-	mov	40(%rsi),%rbx
-	lea	48(%rsi),%rsp
+	mov	-48(%rsi),%r15
+.cfi_restore	%r15
+	mov	-40(%rsi),%r14
+.cfi_restore	%r14
+	mov	-32(%rsi),%r13
+.cfi_restore	%r13
+	mov	-24(%rsi),%r12
+.cfi_restore	%r12
+	mov	-16(%rsi),%rbp
+.cfi_restore	%rbp
+	mov	-8(%rsi),%rbx
+.cfi_restore	%rbx
+	lea	(%rsi),%rsp
+.cfi_def_cfa_register	%rsp
 .Lepilogue_avx2:
 	ret
+.cfi_endproc
 .size	${func}_avx2,.-${func}_avx2
 ___
 }}
@@ -2023,7 +2363,6 @@ ___
 $code.=<<___;
 	mov	%rax,%rsi		# put aside Rsp
 	mov	16*$SZ+3*8(%rax),%rax	# pull $_rsp
-	lea	48(%rax),%rax
 
 	mov	-8(%rax),%rbx
 	mov	-16(%rax),%rbp
@@ -2086,13 +2425,55 @@ $code.=<<___;
 	pop	%rsi
 	ret
 .size	se_handler,.-se_handler
+___
+
+$code.=<<___ if ($SZ==4 && $shaext);
+.type	shaext_handler,\@abi-omnipotent
+.align	16
+shaext_handler:
+	push	%rsi
+	push	%rdi
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	pushfq
+	sub	\$64,%rsp
+
+	mov	120($context),%rax	# pull context->Rax
+	mov	248($context),%rbx	# pull context->Rip
+
+	lea	.Lprologue_shaext(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip<.Lprologue
+	jb	.Lin_prologue
 
+	lea	.Lepilogue_shaext(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip>=.Lepilogue
+	jae	.Lin_prologue
+
+	lea	-8-5*16(%rax),%rsi
+	lea	512($context),%rdi	# &context.Xmm6
+	mov	\$10,%ecx
+	.long	0xa548f3fc		# cld; rep movsq
+
+	jmp	.Lin_prologue
+.size	shaext_handler,.-shaext_handler
+___
+
+$code.=<<___;
 .section	.pdata
 .align	4
 	.rva	.LSEH_begin_$func
 	.rva	.LSEH_end_$func
 	.rva	.LSEH_info_$func
 ___
+$code.=<<___ if ($SZ==4 && $shaext);
+	.rva	.LSEH_begin_${func}_shaext
+	.rva	.LSEH_end_${func}_shaext
+	.rva	.LSEH_info_${func}_shaext
+___
 $code.=<<___ if ($SZ==4);
 	.rva	.LSEH_begin_${func}_ssse3
 	.rva	.LSEH_end_${func}_ssse3
@@ -2121,6 +2502,11 @@ $code.=<<___;
 	.rva	se_handler
 	.rva	.Lprologue,.Lepilogue			# HandlerData[]
 ___
+$code.=<<___ if ($SZ==4 && $shaext);
+.LSEH_info_${func}_shaext:
+	.byte	9,0,0,0
+	.rva	shaext_handler
+___
 $code.=<<___ if ($SZ==4);
 .LSEH_info_${func}_ssse3:
 	.byte	9,0,0,0
@@ -2147,6 +2533,28 @@ $code.=<<___ if ($avx>1);
 ___
 }
 
-$code =~ s/\`([^\`]*)\`/eval $1/gem;
-print $code;
+sub sha256op38 {
+    my $instr = shift;
+    my %opcodelet = (
+		"sha256rnds2" => 0xcb,
+  		"sha256msg1"  => 0xcc,
+		"sha256msg2"  => 0xcd	);
+
+    if (defined($opcodelet{$instr}) && @_[0] =~ /%xmm([0-7]),\s*%xmm([0-7])/) {
+      my @opcode=(0x0f,0x38);
+	push @opcode,$opcodelet{$instr};
+	push @opcode,0xc0|($1&7)|(($2&7)<<3);		# ModR/M
+	return ".byte\t".join(',',@opcode);
+    } else {
+	return $instr."\t".@_[0];
+    }
+}
+
+foreach (split("\n",$code)) {
+	s/\`([^\`]*)\`/eval $1/geo;
+
+	s/\b(sha256[^\s]*)\s+(.*)/sha256op38($1,$2)/geo;
+
+	print $_,"\n";
+}
 close STDOUT;