From c4558efbf3a44a1b5e68dce46347dd3888db4760 Mon Sep 17 00:00:00 2001
From: Andy Polyakov <appro@openssl.org>
Date: Thu, 14 Feb 2013 15:39:42 +0100
Subject: [PATCH] sha512-x86_64.pl: add AVX2 code path.

---
 crypto/sha/asm/sha512-x86_64.pl | 650 ++++++++++++++++++++++++++++----
 1 file changed, 568 insertions(+), 82 deletions(-)

diff --git a/crypto/sha/asm/sha512-x86_64.pl b/crypto/sha/asm/sha512-x86_64.pl
index 8f10f973e9..da094f0b3c 100755
--- a/crypto/sha/asm/sha512-x86_64.pl
+++ b/crypto/sha/asm/sha512-x86_64.pl
@@ -59,6 +59,15 @@
 # higher coefficients are observed on VIA Nano and Bulldozer has more
 # to do with specifics of their architecture [which is topic for
 # separate discussion].
+#
+# November 2012.
+#
+# Add AVX2 code path. Two consecutive input blocks are loaded to
+# 256-bit %ymm registers, with data from first block to least
+# 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.
 
 ######################################################################
 # Current performance in cycles per processed byte (less is better):
@@ -69,13 +78,13 @@
 # P4		17.5	-	    -		    33.4    -
 # Core 2	15.5	13.9(+11%)  -		    10.3    -
 # Westmere	15.1	12.5(+21%)  -		    9.72    -
-# Atom		23.0	21.6(+6%)   -		    14.7    -
-# VIA Nano	23.0	16.3(+41%)  -		    14.7    -
 # Sandy Bridge	17.4	14.0(+24%)  11.6(+50%(**))  11.2    8.10(+38%(**))
 # Ivy Bridge	12.6	10.3(+22%)  10.3(+22%)	    8.17    7.22(+13%)
 # Bulldozer	21.5	13.7(+57%)  13.7(+57%(***)) 13.5    8.58(+57%)
+# VIA Nano	23.0	16.3(+41%)  -		    14.7    -
+# Atom		23.0	21.6(+6%)   -		    14.7    -
 #
-# (*)	whichever applicable;
+# (*)	whichever best applicable;
 # (**)	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
@@ -93,15 +102,20 @@ $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
 die "can't locate x86_64-xlate.pl";
 
-$avx=1 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
-		=~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
-	   $1>=2.19);
-$avx=1 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
-	   `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
-	   $1>=2.09);
-$avx=1 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
-	   `ml64 2>&1` =~ /Version ([0-9]+)\./ &&
-	   $1>=10);
+if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
+		=~ /GNU assembler version ([2-9]\.[0-9]+)/) {
+	$avx = ($1>=2.19) + ($1>=2.22);
+}
+
+if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
+	   `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
+	$avx = ($1>=2.09) + ($1>=2.10);
+}
+
+if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
+	   `ml64 2>&1` =~ /Version ([0-9]+)\./) {
+	$avx = ($1>=10) + ($1>=11);
+}
 
 open OUT,"| \"$^X\" $xlate $flavour $output";
 *STDOUT=*OUT;
@@ -145,6 +159,8 @@ $framesz="16*$SZ+4*8";
 
 sub ROUND_00_15()
 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
+  my $STRIDE=$SZ;
+     $STRIDE += 16 if ($i%(16/$SZ)==(16/$SZ-1));
 
 $code.=<<___;
 	ror	\$`$Sigma1[2]-$Sigma1[1]`,$a0
@@ -186,7 +202,7 @@ $code.=<<___ if ($i>=15);
 	mov	`$SZ*(($i+2)&0xf)`(%rsp),$a0
 ___
 $code.=<<___;
-	lea	$SZ($Tbl),$Tbl		# round++
+	lea	$STRIDE($Tbl),$Tbl	# round++
 	add	$a1,$h			# h+=Sigma0(a)
 
 ___
@@ -229,28 +245,34 @@ $code=<<___;
 
 .extern	OPENSSL_ia32cap_P
 .globl	$func
-.type	$func,\@function,4
+.type	$func,\@function,3
 .align	16
 $func:
 ___
 $code.=<<___ if ($SZ==4 || $avx);
 	lea	OPENSSL_ia32cap_P(%rip),%r11
-	mov	0(%r11),%r10d
-	mov	4(%r11),%r11d
+	mov	0(%r11),%r9d
+	mov	4(%r11),%r10d
+	mov	8(%r11),%r11d
 ___
 $code.=<<___ if ($avx && $SZ==8);
-	test	\$`1<<11`,%r11d		# check for XOP
+	test	\$`1<<11`,%r10d		# check for XOP
 	jnz	.Lxop_shortcut
 ___
+$code.=<<___ if ($avx>1);
+	and	\$`1<<8|1<<5|1<<3`,%r11d	# check for BMI2+AVX2+BMI1
+	cmp	\$`1<<8|1<<5|1<<3`,%r11d
+	je	.Lavx2_shortcut
+___
 $code.=<<___ if ($avx);
-	and	\$`1<<30`,%r10d		# mask "Intel CPU" bit
-	and	\$`1<<28|1<<9`,%r11d	# mask AVX and SSSE3 bits
-	or	%r10d,%r11d
-	cmp	\$`1<<28|1<<9|1<<30`,%r11d
+	and	\$`1<<30`,%r9d		# mask "Intel CPU" bit
+	and	\$`1<<28|1<<9`,%r10d	# mask AVX and SSSE3 bits
+	or	%r9d,%r10d
+	cmp	\$`1<<28|1<<9|1<<30`,%r10d
 	je	.Lavx_shortcut
 ___
 $code.=<<___ if ($SZ==4);
-	test	\$`1<<9`,%r11d
+	test	\$`1<<9`,%r10d
 	jnz	.Lssse3_shortcut
 ___
 $code.=<<___;
@@ -352,24 +374,43 @@ $code.=<<___;
 .type	$TABLE,\@object
 $TABLE:
 	.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
 	.asciz	"SHA256 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 ___
@@ -379,48 +420,89 @@ $code.=<<___;
 .type	$TABLE,\@object
 $TABLE:
 	.quad	0x428a2f98d728ae22,0x7137449123ef65cd
+	.quad	0x428a2f98d728ae22,0x7137449123ef65cd
+	.quad	0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
 	.quad	0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
 	.quad	0x3956c25bf348b538,0x59f111f1b605d019
+	.quad	0x3956c25bf348b538,0x59f111f1b605d019
+	.quad	0x923f82a4af194f9b,0xab1c5ed5da6d8118
 	.quad	0x923f82a4af194f9b,0xab1c5ed5da6d8118
 	.quad	0xd807aa98a3030242,0x12835b0145706fbe
+	.quad	0xd807aa98a3030242,0x12835b0145706fbe
+	.quad	0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
 	.quad	0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
 	.quad	0x72be5d74f27b896f,0x80deb1fe3b1696b1
+	.quad	0x72be5d74f27b896f,0x80deb1fe3b1696b1
 	.quad	0x9bdc06a725c71235,0xc19bf174cf692694
+	.quad	0x9bdc06a725c71235,0xc19bf174cf692694
+	.quad	0xe49b69c19ef14ad2,0xefbe4786384f25e3
 	.quad	0xe49b69c19ef14ad2,0xefbe4786384f25e3
 	.quad	0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
+	.quad	0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
+	.quad	0x2de92c6f592b0275,0x4a7484aa6ea6e483
 	.quad	0x2de92c6f592b0275,0x4a7484aa6ea6e483
 	.quad	0x5cb0a9dcbd41fbd4,0x76f988da831153b5
+	.quad	0x5cb0a9dcbd41fbd4,0x76f988da831153b5
+	.quad	0x983e5152ee66dfab,0xa831c66d2db43210
 	.quad	0x983e5152ee66dfab,0xa831c66d2db43210
 	.quad	0xb00327c898fb213f,0xbf597fc7beef0ee4
+	.quad	0xb00327c898fb213f,0xbf597fc7beef0ee4
+	.quad	0xc6e00bf33da88fc2,0xd5a79147930aa725
 	.quad	0xc6e00bf33da88fc2,0xd5a79147930aa725
 	.quad	0x06ca6351e003826f,0x142929670a0e6e70
+	.quad	0x06ca6351e003826f,0x142929670a0e6e70
+	.quad	0x27b70a8546d22ffc,0x2e1b21385c26c926
 	.quad	0x27b70a8546d22ffc,0x2e1b21385c26c926
 	.quad	0x4d2c6dfc5ac42aed,0x53380d139d95b3df
+	.quad	0x4d2c6dfc5ac42aed,0x53380d139d95b3df
+	.quad	0x650a73548baf63de,0x766a0abb3c77b2a8
 	.quad	0x650a73548baf63de,0x766a0abb3c77b2a8
 	.quad	0x81c2c92e47edaee6,0x92722c851482353b
+	.quad	0x81c2c92e47edaee6,0x92722c851482353b
+	.quad	0xa2bfe8a14cf10364,0xa81a664bbc423001
 	.quad	0xa2bfe8a14cf10364,0xa81a664bbc423001
 	.quad	0xc24b8b70d0f89791,0xc76c51a30654be30
+	.quad	0xc24b8b70d0f89791,0xc76c51a30654be30
+	.quad	0xd192e819d6ef5218,0xd69906245565a910
 	.quad	0xd192e819d6ef5218,0xd69906245565a910
 	.quad	0xf40e35855771202a,0x106aa07032bbd1b8
+	.quad	0xf40e35855771202a,0x106aa07032bbd1b8
 	.quad	0x19a4c116b8d2d0c8,0x1e376c085141ab53
+	.quad	0x19a4c116b8d2d0c8,0x1e376c085141ab53
+	.quad	0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
 	.quad	0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
 	.quad	0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
+	.quad	0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
+	.quad	0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
 	.quad	0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
 	.quad	0x748f82ee5defb2fc,0x78a5636f43172f60
+	.quad	0x748f82ee5defb2fc,0x78a5636f43172f60
+	.quad	0x84c87814a1f0ab72,0x8cc702081a6439ec
 	.quad	0x84c87814a1f0ab72,0x8cc702081a6439ec
 	.quad	0x90befffa23631e28,0xa4506cebde82bde9
+	.quad	0x90befffa23631e28,0xa4506cebde82bde9
+	.quad	0xbef9a3f7b2c67915,0xc67178f2e372532b
 	.quad	0xbef9a3f7b2c67915,0xc67178f2e372532b
 	.quad	0xca273eceea26619c,0xd186b8c721c0c207
+	.quad	0xca273eceea26619c,0xd186b8c721c0c207
+	.quad	0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
 	.quad	0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
 	.quad	0x06f067aa72176fba,0x0a637dc5a2c898a6
+	.quad	0x06f067aa72176fba,0x0a637dc5a2c898a6
+	.quad	0x113f9804bef90dae,0x1b710b35131c471b
 	.quad	0x113f9804bef90dae,0x1b710b35131c471b
 	.quad	0x28db77f523047d84,0x32caab7b40c72493
+	.quad	0x28db77f523047d84,0x32caab7b40c72493
+	.quad	0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
 	.quad	0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
 	.quad	0x4cc5d4becb3e42b6,0x597f299cfc657e2a
+	.quad	0x4cc5d4becb3e42b6,0x597f299cfc657e2a
+	.quad	0x5fcb6fab3ad6faec,0x6c44198c4a475817
 	.quad	0x5fcb6fab3ad6faec,0x6c44198c4a475817
 
 	.quad	0x0001020304050607,0x08090a0b0c0d0e0f
-	.asciz	"SHA512 block transfort for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
+	.quad	0x0001020304050607,0x08090a0b0c0d0e0f
+	.asciz	"SHA512 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 ___
 }
 
@@ -489,7 +571,7 @@ my @X = map("%xmm$_",(0..3));
 my ($t0,$t1,$t2,$t3, $t4,$t5) = map("%xmm$_",(4..9));
 
 $code.=<<___;
-.type	${func}_ssse3,\@function,4
+.type	${func}_ssse3,\@function,3
 .align	64
 ${func}_ssse3:
 .Lssse3_shortcut:
@@ -529,12 +611,12 @@ $code.=<<___;
 ___
 
 $code.=<<___;
-	movdqa	$TABLE+`$SZ*$rounds`+16(%rip),$t4
-	movdqa	$TABLE+`$SZ*$rounds`+32(%rip),$t5
+	movdqa	$TABLE+`$SZ*2*$rounds`+32(%rip),$t4
+	movdqa	$TABLE+`$SZ*2*$rounds`+64(%rip),$t5
 	jmp	.Lloop_ssse3
 .align	16
 .Lloop_ssse3:
-	movdqa	$TABLE+`$SZ*$rounds`(%rip),$t3
+	movdqa	$TABLE+`$SZ*2*$rounds`(%rip),$t3
 	movdqu	0x00($inp),@X[0]
 	movdqu	0x10($inp),@X[1]
 	movdqu	0x20($inp),@X[2]
@@ -544,11 +626,11 @@ $code.=<<___;
 	pshufb	$t3,@X[1]
 	movdqa	0x00($Tbl),$t0
 	pshufb	$t3,@X[2]
-	movdqa	0x10($Tbl),$t1
+	movdqa	0x20($Tbl),$t1
 	paddd	@X[0],$t0
-	movdqa	0x20($Tbl),$t2
+	movdqa	0x40($Tbl),$t2
 	pshufb	$t3,@X[3]
-	movdqa	0x30($Tbl),$t3
+	movdqa	0x60($Tbl),$t3
 	paddd	@X[1],$t1
 	paddd	@X[2],$t2
 	paddd	@X[3],$t3
@@ -564,7 +646,7 @@ $code.=<<___;
 
 .align	16
 .Lssse3_00_47:
-	add	\$16*$SZ,$Tbl
+	sub	\$-16*2*$SZ,$Tbl	# size optimization
 ___
 sub Xupdate_256_SSSE3 () {
 	(
@@ -601,7 +683,7 @@ sub Xupdate_256_SSSE3 () {
 	 '&pxor		($t3,$t2);',
 	 '&psrlq	($t2,$sigma1[1]-$sigma1[0])',
 	 '&pxor		($t3,$t2);',
-	'&movdqa	($t2,16*$j."($Tbl)")',
+	'&movdqa	($t2,16*2*$j."($Tbl)")',
 	 '&pshufb	($t3,$t5)',
 	'&paddd		(@X[0],$t3)'		# X[2..3] += sigma1(X[16..17])
 	);
@@ -744,7 +826,7 @@ my @insns = (&$body,&$body,&$body,&$body);	# 104 instructions
 	 &pxor		($t3,$t2);
 	  eval(shift(@insns));
 	  eval(shift(@insns));
-	&movdqa		($t2,16*$j."($Tbl)");
+	&movdqa		($t2,16*2*$j."($Tbl)");
 	  eval(shift(@insns));	#@
 	  eval(shift(@insns));
 	 &pshufb	($t3,$t5);
@@ -767,7 +849,7 @@ my @insns = (&$body,&$body,&$body,&$body);	# 104 instructions
 	&SSSE3_256_00_47($j,\&body_00_15,@X);
 	push(@X,shift(@X));			# rotate(@X)
     }
-	&cmpb	($SZ-1+16*$SZ."($Tbl)",0);
+	&cmpb	($SZ-1+16*2*$SZ."($Tbl)",0);
 	&jne	(".Lssse3_00_47");
 
     for ($i=0; $i<16; ) {
@@ -827,7 +909,7 @@ if ($avx) {{
 #
 if ($SZ==8) {	# SHA512 only
 $code.=<<___;
-.type	${func}_xop,\@function,4
+.type	${func}_xop,\@function,3
 .align	64
 ${func}_xop:
 .Lxop_shortcut:
@@ -878,7 +960,7 @@ ___
 $code.=<<___;
 .align	16
 .Lloop_xop:
-	vmovdqa	$TABLE+`$SZ*$rounds`(%rip),$t3
+	vmovdqa	$TABLE+`$SZ*2*$rounds`(%rip),$t3
 	vmovdqu	0x00($inp),@X[0]
 	vmovdqu	0x10($inp),@X[1]
 	vmovdqu	0x20($inp),@X[2]
@@ -889,9 +971,9 @@ $code.=<<___;
 	vpshufb	$t3,@X[2],@X[2]
 	vpaddd	0x00($Tbl),@X[0],$t0
 	vpshufb	$t3,@X[3],@X[3]
-	vpaddd	0x10($Tbl),@X[1],$t1
-	vpaddd	0x20($Tbl),@X[2],$t2
-	vpaddd	0x30($Tbl),@X[3],$t3
+	vpaddd	0x20($Tbl),@X[1],$t1
+	vpaddd	0x40($Tbl),@X[2],$t2
+	vpaddd	0x60($Tbl),@X[3],$t3
 	vmovdqa	$t0,0x00(%rsp)
 	mov	$A,$a1
 	vmovdqa	$t1,0x10(%rsp)
@@ -904,7 +986,7 @@ $code.=<<___;
 
 .align	16
 .Lxop_00_47:
-	add	\$16*$SZ,$Tbl
+	sub	\$-16*2*$SZ,$Tbl	# size optimization
 ___
 sub XOP_256_00_47 () {
 my $j = shift;
@@ -1001,7 +1083,7 @@ my @insns = (&$body,&$body,&$body,&$body);	# 104 instructions
 	  eval(shift(@insns));
 	  eval(shift(@insns));
 	  eval(shift(@insns));
-	&vpaddd		($t2,@X[0],16*$j."($Tbl)");
+	&vpaddd		($t2,@X[0],16*2*$j."($Tbl)");
 	  foreach (@insns) { eval; }		# remaining instructions
 	&vmovdqa	(16*$j."(%rsp)",$t2);
 }
@@ -1010,7 +1092,7 @@ my @insns = (&$body,&$body,&$body,&$body);	# 104 instructions
 	&XOP_256_00_47($j,\&body_00_15,@X);
 	push(@X,shift(@X));			# rotate(@X)
     }
-	&cmpb	($SZ-1+16*$SZ."($Tbl)",0);
+	&cmpb	($SZ-1+16*2*$SZ."($Tbl)",0);
 	&jne	(".Lxop_00_47");
 
     for ($i=0; $i<16; ) {
@@ -1024,9 +1106,9 @@ my @insns = (&$body,&$body,&$body,&$body);	# 104 instructions
 $code.=<<___;
 .align	16
 .Lloop_xop:
-	vmovdqa	$TABLE+`$SZ*$rounds`(%rip),$t3
+	vmovdqa	$TABLE+`$SZ*2*$rounds`(%rip),$t3
 	vmovdqu	0x00($inp),@X[0]
-	lea	$TABLE(%rip),$Tbl
+	lea	$TABLE+0x80(%rip),$Tbl	# size optimization
 	vmovdqu	0x10($inp),@X[1]
 	vmovdqu	0x20($inp),@X[2]
 	vpshufb	$t3,@X[0],@X[0]
@@ -1040,20 +1122,20 @@ $code.=<<___;
 	vpshufb	$t3,@X[4],@X[4]
 	vmovdqu	0x70($inp),@X[7]
 	vpshufb	$t3,@X[5],@X[5]
-	vpaddq	0x00($Tbl),@X[0],$t0
+	vpaddq	-0x80($Tbl),@X[0],$t0
 	vpshufb	$t3,@X[6],@X[6]
-	vpaddq	0x10($Tbl),@X[1],$t1
+	vpaddq	-0x60($Tbl),@X[1],$t1
 	vpshufb	$t3,@X[7],@X[7]
-	vpaddq	0x20($Tbl),@X[2],$t2
-	vpaddq	0x30($Tbl),@X[3],$t3
+	vpaddq	-0x40($Tbl),@X[2],$t2
+	vpaddq	-0x20($Tbl),@X[3],$t3
 	vmovdqa	$t0,0x00(%rsp)
-	vpaddq	0x40($Tbl),@X[4],$t0
+	vpaddq	0x00($Tbl),@X[4],$t0
 	vmovdqa	$t1,0x10(%rsp)
-	vpaddq	0x50($Tbl),@X[5],$t1
+	vpaddq	0x20($Tbl),@X[5],$t1
 	vmovdqa	$t2,0x20(%rsp)
-	vpaddq	0x60($Tbl),@X[6],$t2
+	vpaddq	0x40($Tbl),@X[6],$t2
 	vmovdqa	$t3,0x30(%rsp)
-	vpaddq	0x70($Tbl),@X[7],$t3
+	vpaddq	0x60($Tbl),@X[7],$t3
 	vmovdqa	$t0,0x40(%rsp)
 	mov	$A,$a1
 	vmovdqa	$t1,0x50(%rsp)
@@ -1066,7 +1148,7 @@ $code.=<<___;
 
 .align	16
 .Lxop_00_47:
-	add	\$16*$SZ,$Tbl
+	add	\$16*2*$SZ,$Tbl
 ___
 sub XOP_512_00_47 () {
 my $j = shift;
@@ -1129,7 +1211,7 @@ my @insns = (&$body,&$body);			# 52 instructions
 	  eval(shift(@insns));
 	  eval(shift(@insns));
 	  eval(shift(@insns));
-	&vpaddq		($t2,@X[0],16*$j."($Tbl)");
+	&vpaddq		($t2,@X[0],16*2*$j-0x80."($Tbl)");
 	  foreach (@insns) { eval; }		# remaining instructions
 	&vmovdqa	(16*$j."(%rsp)",$t2);
 }
@@ -1138,7 +1220,7 @@ my @insns = (&$body,&$body);			# 52 instructions
 	&XOP_512_00_47($j,\&body_00_15,@X);
 	push(@X,shift(@X));			# rotate(@X)
     }
-	&cmpb	($SZ-1+16*$SZ."($Tbl)",0);
+	&cmpb	($SZ-1+16*2*$SZ-0x80."($Tbl)",0);
 	&jne	(".Lxop_00_47");
 
     for ($i=0; $i<16; ) {
@@ -1203,7 +1285,7 @@ ___
 local *ror = sub { &shrd(@_[0],@_) };
 
 $code.=<<___;
-.type	${func}_avx,\@function,4
+.type	${func}_avx,\@function,3
 .align	64
 ${func}_avx:
 .Lavx_shortcut:
@@ -1251,12 +1333,12 @@ ___
     my ($t0,$t1,$t2,$t3, $t4,$t5) = map("%xmm$_",(4..9));
 
 $code.=<<___;
-	vmovdqa	$TABLE+`$SZ*$rounds`+16(%rip),$t4
-	vmovdqa	$TABLE+`$SZ*$rounds`+32(%rip),$t5
+	vmovdqa	$TABLE+`$SZ*2*$rounds`+32(%rip),$t4
+	vmovdqa	$TABLE+`$SZ*2*$rounds`+64(%rip),$t5
 	jmp	.Lloop_avx
 .align	16
 .Lloop_avx:
-	vmovdqa	$TABLE+`$SZ*$rounds`(%rip),$t3
+	vmovdqa	$TABLE+`$SZ*2*$rounds`(%rip),$t3
 	vmovdqu	0x00($inp),@X[0]
 	vmovdqu	0x10($inp),@X[1]
 	vmovdqu	0x20($inp),@X[2]
@@ -1267,9 +1349,9 @@ $code.=<<___;
 	vpshufb	$t3,@X[2],@X[2]
 	vpaddd	0x00($Tbl),@X[0],$t0
 	vpshufb	$t3,@X[3],@X[3]
-	vpaddd	0x10($Tbl),@X[1],$t1
-	vpaddd	0x20($Tbl),@X[2],$t2
-	vpaddd	0x30($Tbl),@X[3],$t3
+	vpaddd	0x20($Tbl),@X[1],$t1
+	vpaddd	0x40($Tbl),@X[2],$t2
+	vpaddd	0x60($Tbl),@X[3],$t3
 	vmovdqa	$t0,0x00(%rsp)
 	mov	$A,$a1
 	vmovdqa	$t1,0x10(%rsp)
@@ -1282,7 +1364,7 @@ $code.=<<___;
 
 .align	16
 .Lavx_00_47:
-	add	\$16*$SZ,$Tbl
+	sub	\$-16*2*$SZ,$Tbl	# size optimization
 ___
 sub Xupdate_256_AVX () {
 	(
@@ -1330,7 +1412,7 @@ my @insns = (&$body,&$body,&$body,&$body);	# 104 instructions
 	    eval(shift(@insns));
 	    eval(shift(@insns));
 	}
-	&vpaddd		($t2,@X[0],16*$j."($Tbl)");
+	&vpaddd		($t2,@X[0],16*2*$j."($Tbl)");
 	  foreach (@insns) { eval; }		# remaining instructions
 	&vmovdqa	(16*$j."(%rsp)",$t2);
 }
@@ -1339,7 +1421,7 @@ my @insns = (&$body,&$body,&$body,&$body);	# 104 instructions
 	&AVX_256_00_47($j,\&body_00_15,@X);
 	push(@X,shift(@X));			# rotate(@X)
     }
-	&cmpb	($SZ-1+16*$SZ."($Tbl)",0);
+	&cmpb	($SZ-1+16*2*$SZ."($Tbl)",0);
 	&jne	(".Lavx_00_47");
 
     for ($i=0; $i<16; ) {
@@ -1354,9 +1436,9 @@ $code.=<<___;
 	jmp	.Lloop_avx
 .align	16
 .Lloop_avx:
-	vmovdqa	$TABLE+`$SZ*$rounds`(%rip),$t3
+	vmovdqa	$TABLE+`$SZ*2*$rounds`(%rip),$t3
 	vmovdqu	0x00($inp),@X[0]
-	lea	$TABLE(%rip),$Tbl
+	lea	$TABLE+0x80(%rip),$Tbl	# size optimization
 	vmovdqu	0x10($inp),@X[1]
 	vmovdqu	0x20($inp),@X[2]
 	vpshufb	$t3,@X[0],@X[0]
@@ -1370,20 +1452,20 @@ $code.=<<___;
 	vpshufb	$t3,@X[4],@X[4]
 	vmovdqu	0x70($inp),@X[7]
 	vpshufb	$t3,@X[5],@X[5]
-	vpaddq	0x00($Tbl),@X[0],$t0
+	vpaddq	-0x80($Tbl),@X[0],$t0
 	vpshufb	$t3,@X[6],@X[6]
-	vpaddq	0x10($Tbl),@X[1],$t1
+	vpaddq	-0x60($Tbl),@X[1],$t1
 	vpshufb	$t3,@X[7],@X[7]
-	vpaddq	0x20($Tbl),@X[2],$t2
-	vpaddq	0x30($Tbl),@X[3],$t3
+	vpaddq	-0x40($Tbl),@X[2],$t2
+	vpaddq	-0x20($Tbl),@X[3],$t3
 	vmovdqa	$t0,0x00(%rsp)
-	vpaddq	0x40($Tbl),@X[4],$t0
+	vpaddq	0x00($Tbl),@X[4],$t0
 	vmovdqa	$t1,0x10(%rsp)
-	vpaddq	0x50($Tbl),@X[5],$t1
+	vpaddq	0x20($Tbl),@X[5],$t1
 	vmovdqa	$t2,0x20(%rsp)
-	vpaddq	0x60($Tbl),@X[6],$t2
+	vpaddq	0x40($Tbl),@X[6],$t2
 	vmovdqa	$t3,0x30(%rsp)
-	vpaddq	0x70($Tbl),@X[7],$t3
+	vpaddq	0x60($Tbl),@X[7],$t3
 	vmovdqa	$t0,0x40(%rsp)
 	mov	$A,$a1
 	vmovdqa	$t1,0x50(%rsp)
@@ -1396,14 +1478,14 @@ $code.=<<___;
 
 .align	16
 .Lavx_00_47:
-	add	\$16*$SZ,$Tbl
+	add	\$16*2*$SZ,$Tbl
 ___
 sub Xupdate_512_AVX () {
 	(
 	'&vpalignr	($t0,@X[1],@X[0],$SZ)',	# X[1..2]
 	 '&vpalignr	($t3,@X[5],@X[4],$SZ)',	# X[9..10]
-	'&vpsrlq	($t2,$t0,$sigma0[0]);',
-	 '&vpaddq	(@X[0],@X[0],$t3)',	# X[0..1] += X[9..10]
+	'&vpsrlq	($t2,$t0,$sigma0[0])',
+	 '&vpaddq	(@X[0],@X[0],$t3);',	# X[0..1] += X[9..10]
 	'&vpsrlq	($t3,$t0,$sigma0[2])',
 	'&vpsllq	($t1,$t0,8*$SZ-$sigma0[1]);',
 	 '&vpxor	($t0,$t3,$t2)',
@@ -1413,7 +1495,7 @@ sub Xupdate_512_AVX () {
 	 '&vpxor	($t0,$t0,$t2)',
 	 '&vpsrlq	($t3,@X[7],$sigma1[2]);',
 	'&vpxor		($t0,$t0,$t1)',		# sigma0(X[1..2])
-	 '&vpsllq	($t2,@X[7],8*$SZ-$sigma1[1])',
+	 '&vpsllq	($t2,@X[7],8*$SZ-$sigma1[1]);',
 	'&vpaddq	(@X[0],@X[0],$t0)',	# X[0..1] += sigma0(X[1..2])
 	 '&vpsrlq	($t1,@X[7],$sigma1[0]);',
 	 '&vpxor	($t3,$t3,$t2)',
@@ -1437,7 +1519,7 @@ my @insns = (&$body,&$body);			# 52 instructions
 	    eval(shift(@insns));
 	    eval(shift(@insns));
 	}
-	&vpaddq		($t2,@X[0],16*$j."($Tbl)");
+	&vpaddq		($t2,@X[0],16*2*$j-0x80."($Tbl)");
 	  foreach (@insns) { eval; }		# remaining instructions
 	&vmovdqa	(16*$j."(%rsp)",$t2);
 }
@@ -1446,7 +1528,7 @@ my @insns = (&$body,&$body);			# 52 instructions
 	&AVX_512_00_47($j,\&body_00_15,@X);
 	push(@X,shift(@X));			# rotate(@X)
     }
-	&cmpb	($SZ-1+16*$SZ."($Tbl)",0);
+	&cmpb	($SZ-1+16*2*$SZ-0x80."($Tbl)",0);
 	&jne	(".Lavx_00_47");
 
     for ($i=0; $i<16; ) {
@@ -1504,6 +1586,389 @@ $code.=<<___;
 	ret
 .size	${func}_avx,.-${func}_avx
 ___
+
+if ($avx>1) {{
+######################################################################
+# AVX2+BMI code path
+#
+my $a5=$SZ==4?"%esi":"%rsi";	# zap $inp 
+my $PUSH8=8*2*$SZ;
+use integer;
+
+sub bodyx_00_15 () {
+	# at start $a1 should be zero, $a3 - $b^$c and $a4 copy of $f
+	(
+	'($a,$b,$c,$d,$e,$f,$g,$h)=@ROT;'.
+
+	'&add	($h,(32*($i/(16/$SZ))+$SZ*($i%(16/$SZ)))%$PUSH8.$base)',    # h+=X[i]+K[i]
+	'&and	($a4,$e)',		# f&e
+	'&rorx	($a0,$e,$Sigma1[2])',
+	'&rorx	($a2,$e,$Sigma1[1])',
+
+	'&lea	($a,"($a,$a1)")',	# h+=Sigma0(a) from the past
+	'&lea	($h,"($h,$a4)")',
+	'&andn	($a4,$e,$g)',		# ~e&g
+	'&xor	($a0,$a2)',
+
+	'&rorx	($a1,$e,$Sigma1[0])',
+	'&lea	($h,"($h,$a4)")',	# h+=Ch(e,f,g)=(e&f)+(~e&g)
+	'&xor	($a0,$a1)',		# Sigma1(e)
+	'&mov	($a2,$a)',
+
+	'&rorx	($a4,$a,$Sigma0[2])',
+	'&lea	($h,"($h,$a0)")',	# h+=Sigma1(e)
+	'&xor	($a2,$b)',		# a^b, b^c in next round
+	'&rorx	($a1,$a,$Sigma0[1])',
+
+	'&rorx	($a0,$a,$Sigma0[0])',
+	'&lea	($d,"($d,$h)")',	# d+=h
+	'&and	($a3,$a2)',		# (b^c)&(a^b)
+	'&xor	($a1,$a4)',
+
+	'&xor	($a3,$b)',		# Maj(a,b,c)=Ch(a^b,c,b)
+	'&xor	($a1,$a0)',		# Sigma0(a)
+	'&lea	($h,"($h,$a3)");'.	# h+=Maj(a,b,c)
+	'&mov	($a4,$e)',		# copy of f in future
+
+	'($a2,$a3) = ($a3,$a2); unshift(@ROT,pop(@ROT)); $i++;'
+	);
+	# and at the finish one has to $a+=$a1
+}
+
+$code.=<<___;
+.type	${func}_avx2,\@function,3
+.align	64
+${func}_avx2:
+.Lavx2_shortcut:
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	mov	%rsp,%r11		# copy %rsp
+	sub	\$`2*$SZ*$rounds+4*8+$win64*16*($SZ==4?4:6)`,%rsp
+	shl	\$4,%rdx		# num*16
+	and	\$-256*$SZ,%rsp		# align stack frame
+	lea	($inp,%rdx,$SZ),%rdx	# inp+num*16*$SZ
+	add	\$`2*$SZ*($rounds-8)`,%rsp
+	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
+___
+$code.=<<___ if ($win64);
+	movaps	%xmm6,16*$SZ+32(%rsp)
+	movaps	%xmm7,16*$SZ+48(%rsp)
+	movaps	%xmm8,16*$SZ+64(%rsp)
+	movaps	%xmm9,16*$SZ+80(%rsp)
+___
+$code.=<<___ if ($win64 && $SZ>4);
+	movaps	%xmm10,16*$SZ+96(%rsp)
+	movaps	%xmm11,16*$SZ+112(%rsp)
+___
+$code.=<<___;
+.Lprologue_avx2:
+
+	vzeroall
+	sub	\$-16*$SZ,$inp		# inp++, size optimization
+	mov	$SZ*0($ctx),$A
+	xor	%r12,%r12		# borrow $T1
+	mov	$SZ*1($ctx),$B
+	cmp	%rdx,$inp		# $_end
+	mov	$SZ*2($ctx),$C
+	sete	%r12b
+	mov	$SZ*3($ctx),$D
+	mov	$SZ*4($ctx),$E
+	mov	$SZ*5($ctx),$F
+	mov	$SZ*6($ctx),$G
+	mov	$SZ*7($ctx),$H
+___
+					if ($SZ==4) {	# SHA256
+    my @X = map("%ymm$_",(0..3));
+    my ($t0,$t1,$t2,$t3, $t4,$t5) = map("%ymm$_",(4..9));
+
+$code.=<<___;
+	vmovdqa	$TABLE+`$SZ*2*$rounds`+32(%rip),$t4
+	vmovdqa	$TABLE+`$SZ*2*$rounds`+64(%rip),$t5
+	jmp	.Loop_avx2
+.align	16
+.Loop_avx2:
+	shl	\$`log(16*$SZ)/log(2)`,%r12
+	vmovdqa	$TABLE+`$SZ*2*$rounds`(%rip),$t3
+	neg	%r12	
+	vmovdqu	-16*$SZ+0($inp),$t0
+	add	$inp,%r12
+	vmovdqu	-16*$SZ+32($inp),$t1
+	vmovdqu	(%r12),@X[2]		# next or same input block
+	vmovdqu	32(%r12),@X[3]
+
+	vperm2i128	\$0x20,@X[2],$t0,@X[0]
+	#mov		$inp,$_inp	# offload $inp
+	vperm2i128	\$0x31,@X[2],$t0,@X[1]
+	vperm2i128	\$0x20,@X[3],$t1,@X[2]
+	vperm2i128	\$0x31,@X[3],$t1,@X[3]
+
+	lea	$TABLE(%rip),$Tbl
+	vpshufb	$t3,@X[0],@X[0]
+	vpshufb	$t3,@X[1],@X[1]
+	vpshufb	$t3,@X[2],@X[2]
+	vpaddd	0x00($Tbl),@X[0],$t0
+	vpshufb	$t3,@X[3],@X[3]
+	vpaddd	0x20($Tbl),@X[1],$t1
+	vpaddd	0x40($Tbl),@X[2],$t2
+	vpaddd	0x60($Tbl),@X[3],$t3
+	vmovdqa	$t0,0x00(%rsp)
+	xor	$a1,$a1
+	vmovdqa	$t1,0x20(%rsp)
+	lea	-$PUSH8(%rsp),%rsp
+	mov	$B,$a3
+	vmovdqa	$t2,0x00(%rsp)
+	xor	$C,$a3			# magic
+	vmovdqa	$t3,0x20(%rsp)
+	mov	$F,$a4
+	sub	\$-16*2*$SZ,$Tbl	# size optimization
+	jmp	.Lavx2_00_47
+
+.align	16
+.Lavx2_00_47:
+___
+
+sub AVX2_256_00_47 () {
+my $j = shift;
+my $body = shift;
+my @X = @_;
+my @insns = (&$body,&$body,&$body,&$body);	# 96 instructions
+my $base = "+2*$PUSH8(%rsp)";
+
+	&lea	("%rsp","-$PUSH8(%rsp)")	if (($j%2)==0);
+	foreach (Xupdate_256_AVX()) {		# 29 instructions
+	    eval;
+	    eval(shift(@insns));
+	    eval(shift(@insns));
+	    eval(shift(@insns));
+	}
+	&vpaddd		($t2,@X[0],16*2*$j."($Tbl)");
+	  foreach (@insns) { eval; }		# remaining instructions
+	&vmovdqa	((32*$j)%$PUSH8."(%rsp)",$t2);
+}
+
+    for ($i=0,$j=0; $j<4; $j++) {
+	&AVX2_256_00_47($j,\&bodyx_00_15,@X);
+	push(@X,shift(@X));			# rotate(@X)
+    }
+	&lea	($Tbl,16*2*$SZ."($Tbl)");
+	&cmpb	(($SZ-1)."($Tbl)",0);
+	&jne	(".Lavx2_00_47");
+
+    for ($i=0; $i<16; ) {
+	my $base=$i<8?"+$PUSH8(%rsp)":"(%rsp)";
+	foreach(bodyx_00_15()) { eval; }
+    }
+					} else {	# SHA512
+    my @X = map("%ymm$_",(0..7));
+    my ($t0,$t1,$t2,$t3) = map("%ymm$_",(8..11));
+
+$code.=<<___;
+	jmp	.Loop_avx2
+.align	16
+.Loop_avx2:
+	shl	\$`log(16*$SZ)/log(2)`,%r12
+	vmovdqu	-16*$SZ($inp),$t0
+	neg	%r12
+	vmovdqu	-16*$SZ+32($inp),$t1
+	add	$inp,%r12
+	vmovdqu	-16*$SZ+64($inp),$t2
+	vmovdqu	-16*$SZ+96($inp),$t3
+	vmovdqu	(%r12),@X[4]		# next or same block
+	vmovdqu	32(%r12),@X[5]
+	vmovdqu	64(%r12),@X[6]
+	vmovdqu	96(%r12),@X[7]
+
+	vperm2i128	\$0x20,@X[4],$t0,@X[0]
+	#mov		$inp,$_inp	# offload $inp
+	vperm2i128	\$0x31,@X[4],$t0,@X[1]
+	vperm2i128	\$0x20,@X[5],$t1,@X[2]
+	vperm2i128	\$0x31,@X[5],$t1,@X[3]
+	vperm2i128	\$0x20,@X[6],$t2,@X[4]
+	vperm2i128	\$0x31,@X[6],$t2,@X[5]
+	vmovdqa		$TABLE+`$SZ*2*$rounds`(%rip),$t2
+	vperm2i128	\$0x20,@X[7],$t3,@X[6]
+	vperm2i128	\$0x31,@X[7],$t3,@X[7]
+
+	vpshufb	$t2,@X[0],@X[0]
+	lea	$TABLE+0x80(%rip),$Tbl	# size optimization
+	vpshufb	$t2,@X[1],@X[1]
+	vpshufb	$t2,@X[2],@X[2]
+	vpshufb	$t2,@X[3],@X[3]
+	vpshufb	$t2,@X[4],@X[4]
+	vpshufb	$t2,@X[5],@X[5]
+	vpaddq	-0x80($Tbl),@X[0],$t0
+	vpshufb	$t2,@X[6],@X[6]
+	vpaddq	-0x60($Tbl),@X[1],$t1
+	vpshufb	$t2,@X[7],@X[7]
+	vpaddq	-0x40($Tbl),@X[2],$t2
+	vpaddq	-0x20($Tbl),@X[3],$t3
+	vmovdqa	$t0,0x00(%rsp)
+	vpaddq	0x00($Tbl),@X[4],$t0
+	vmovdqa	$t1,0x20(%rsp)
+	vpaddq	0x20($Tbl),@X[5],$t1
+	vmovdqa	$t2,0x40(%rsp)
+	vpaddq	0x40($Tbl),@X[6],$t2
+	vmovdqa	$t3,0x60(%rsp)
+	lea	-$PUSH8(%rsp),%rsp
+	vpaddq	0x60($Tbl),@X[7],$t3
+	vmovdqa	$t0,0x00(%rsp)
+	xor	$a1,$a1
+	vmovdqa	$t1,0x20(%rsp)
+	mov	$B,$a3
+	vmovdqa	$t2,0x40(%rsp)
+	xor	$C,$a3			# magic
+	vmovdqa	$t3,0x60(%rsp)
+	mov	$F,$a4
+	add	\$16*2*$SZ,$Tbl
+	jmp	.Lavx2_00_47
+
+.align	16
+.Lavx2_00_47:
+___
+
+sub AVX2_512_00_47 () {
+my $j = shift;
+my $body = shift;
+my @X = @_;
+my @insns = (&$body,&$body);			# 48 instructions
+my $base = "+2*$PUSH8(%rsp)";
+
+	&lea	("%rsp","-$PUSH8(%rsp)")	if (($j%4)==0);
+	foreach (Xupdate_512_AVX()) {		# 23 instructions
+	    eval;
+	    if ($_ !~ /\;$/) {
+		eval(shift(@insns));
+		eval(shift(@insns));
+		eval(shift(@insns));
+	    }
+	}
+	&vpaddq		($t2,@X[0],16*2*$j-0x80."($Tbl)");
+	  foreach (@insns) { eval; }		# remaining instructions
+	&vmovdqa	((32*$j)%$PUSH8."(%rsp)",$t2);
+}
+
+    for ($i=0,$j=0; $j<8; $j++) {
+	&AVX2_512_00_47($j,\&bodyx_00_15,@X);
+	push(@X,shift(@X));			# rotate(@X)
+    }
+	&lea	($Tbl,16*2*$SZ."($Tbl)");
+	&cmpb	(($SZ-1-0x80)."($Tbl)",0);
+	&jne	(".Lavx2_00_47");
+
+    for ($i=0; $i<16; ) {
+	my $base=$i<8?"+$PUSH8(%rsp)":"(%rsp)";
+	foreach(bodyx_00_15()) { eval; }
+    }
+}
+$code.=<<___;
+	mov	`2*$SZ*$rounds`(%rsp),$ctx	# $_ctx
+	add	$a1,$A
+	#mov	`2*$SZ*$rounds+8`(%rsp),$inp	# $_inp
+	lea	`2*$SZ*($rounds-8)`(%rsp),$Tbl
+
+	add	$SZ*0($ctx),$A
+	add	$SZ*1($ctx),$B
+	add	$SZ*2($ctx),$C
+	add	$SZ*3($ctx),$D
+	add	$SZ*4($ctx),$E
+	add	$SZ*5($ctx),$F
+	add	$SZ*6($ctx),$G
+	add	$SZ*7($ctx),$H
+
+	mov	$A,$SZ*0($ctx)
+	mov	$B,$SZ*1($ctx)
+	mov	$C,$SZ*2($ctx)
+	mov	$D,$SZ*3($ctx)
+	mov	$E,$SZ*4($ctx)
+	mov	$F,$SZ*5($ctx)
+	mov	$G,$SZ*6($ctx)
+	mov	$H,$SZ*7($ctx)
+
+	cmp	`$PUSH8+2*8`($Tbl),$inp	# $_end
+	je	.Ldone_avx2
+
+	xor	$a1,$a1
+	mov	$B,$a3
+	xor	$C,$a3			# magic
+	mov	$F,$a4
+	jmp	.Lower_avx2
+.align	16
+.Lower_avx2:
+___
+    for ($i=0; $i<8; ) {
+	my $base="+16($Tbl)";
+	foreach(bodyx_00_15()) { eval; }
+    }
+$code.=<<___;
+	lea	-$PUSH8($Tbl),$Tbl
+	cmp	%rsp,$Tbl
+	jae	.Lower_avx2
+
+	mov	`2*$SZ*$rounds`(%rsp),$ctx	# $_ctx
+	add	$a1,$A
+	#mov	`2*$SZ*$rounds+8`(%rsp),$inp	# $_inp
+	lea	`2*$SZ*($rounds-8)`(%rsp),%rsp
+
+	add	$SZ*0($ctx),$A
+	add	$SZ*1($ctx),$B
+	add	$SZ*2($ctx),$C
+	add	$SZ*3($ctx),$D
+	add	$SZ*4($ctx),$E
+	add	$SZ*5($ctx),$F
+	lea	`2*16*$SZ`($inp),$inp	# inp+=2
+	add	$SZ*6($ctx),$G
+	xor	%r12,%r12
+	add	$SZ*7($ctx),$H
+	cmp	$_end,$inp
+
+	mov	$A,$SZ*0($ctx)
+	mov	$B,$SZ*1($ctx)
+	mov	$C,$SZ*2($ctx)
+	mov	$D,$SZ*3($ctx)
+	mov	$E,$SZ*4($ctx)
+	mov	$F,$SZ*5($ctx)
+	mov	$G,$SZ*6($ctx)
+	mov	$H,$SZ*7($ctx)
+
+	sete	%r12b
+	jbe	.Loop_avx2
+	lea	(%rsp),$Tbl
+
+.Ldone_avx2:
+	lea	($Tbl),%rsp
+	mov	$_rsp,%rsi
+	vzeroall
+___
+$code.=<<___ if ($win64);
+	movaps	16*$SZ+32(%rsp),%xmm6
+	movaps	16*$SZ+48(%rsp),%xmm7
+	movaps	16*$SZ+64(%rsp),%xmm8
+	movaps	16*$SZ+80(%rsp),%xmm9
+___
+$code.=<<___ if ($win64 && $SZ>4);
+	movaps	16*$SZ+96(%rsp),%xmm10
+	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
+.Lepilogue_avx2:
+	ret
+.size	${func}_avx2,.-${func}_avx2
+___
+}}
 }}}}}
 
 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
@@ -1547,7 +2012,17 @@ se_handler:
 	lea	(%rsi,%r10),%r10	# epilogue label
 	cmp	%r10,%rbx		# context->Rip>=epilogue label
 	jae	.Lin_prologue
-
+___
+$code.=<<___ if ($avx>1);
+	lea	.Lavx2_shortcut(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip<avx2_shortcut
+	jb	.Lnot_in_avx2
+
+	and	\$-256*$SZ,%rax
+	add	\$`2*$SZ*($rounds-8)`,%rax
+.Lnot_in_avx2:
+___
+$code.=<<___;
 	mov	%rax,%rsi		# put aside Rsp
 	mov	16*$SZ+3*8(%rax),%rax	# pull $_rsp
 	lea	48(%rax),%rax
@@ -1635,6 +2110,11 @@ $code.=<<___ if ($avx);
 	.rva	.LSEH_end_${func}_avx
 	.rva	.LSEH_info_${func}_avx
 ___
+$code.=<<___ if ($avx>1);
+	.rva	.LSEH_begin_${func}_avx2
+	.rva	.LSEH_end_${func}_avx2
+	.rva	.LSEH_info_${func}_avx2
+___
 $code.=<<___;
 .section	.xdata
 .align	8
@@ -1661,6 +2141,12 @@ $code.=<<___ if ($avx);
 	.rva	se_handler
 	.rva	.Lprologue_avx,.Lepilogue_avx		# HandlerData[]
 ___
+$code.=<<___ if ($avx>1);
+.LSEH_info_${func}_avx2:
+	.byte	9,0,0,0
+	.rva	se_handler
+	.rva	.Lprologue_avx2,.Lepilogue_avx2		# HandlerData[]
+___
 }
 
 $code =~ s/\`([^\`]*)\`/eval $1/gem;
-- 
2.25.1