From e1612ea59da067c76fa360cf976a6b38216cddf2 Mon Sep 17 00:00:00 2001
From: Andy Polyakov <appro@openssl.org>
Date: Fri, 13 Jul 2007 17:39:40 +0000
Subject: [PATCH] Add _x86_64_AES_[en|de]crypt_compact.

---
 crypto/aes/asm/aes-x86_64.pl | 1071 ++++++++++++++++++++++++++++++----
 1 file changed, 952 insertions(+), 119 deletions(-)

diff --git a/crypto/aes/asm/aes-x86_64.pl b/crypto/aes/asm/aes-x86_64.pl
index d02cd5bd2b..75e1749b64 100755
--- a/crypto/aes/asm/aes-x86_64.pl
+++ b/crypto/aes/asm/aes-x86_64.pl
@@ -2,11 +2,12 @@
 #
 # ====================================================================
 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
-# project. Rights for redistribution and usage in source and binary
-# forms are granted according to the OpenSSL license.
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
 # ====================================================================
 #
-# Version 1.2.
+# Version 2.0.
 #
 # aes-*-cbc benchmarks are improved by >70% [compared to gcc 3.3.2 on
 # Opteron 240 CPU] plus all the bells-n-whistles from 32-bit version
@@ -17,17 +18,21 @@
 #
 # Performance in number of cycles per processed byte for 128-bit key:
 #
-#		ECB		CBC encrypt
-# AMD64		13.7		13.0(*)
-# EM64T		20.2		18.6(*)
-#
-# (*)	CBC benchmarks are better than ECB thanks to custom ABI used
-#	by the private block encryption function.
+#		ECB encrypt	ECB decrypt	CBC large chunk
+# AMD64		33		41		13.0
+# EM64T		38		59		18.6
+# Core 2	30		43		14.5
 
 $verticalspin=1;	# unlike 32-bit version $verticalspin performs
 			# ~15% better on both AMD and Intel cores
 $output=shift;
-open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output";
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
+( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
+die "can't locate x86_64-xlate.pl";
+
+open STDOUT,"| $^X $xlate $output";
 
 $code=".text\n";
 
@@ -35,9 +40,9 @@ $s0="%eax";
 $s1="%ebx";
 $s2="%ecx";
 $s3="%edx";
-$acc0="%esi";
-$acc1="%edi";
-$acc2="%ebp";
+$acc0="%esi";	$mask80="%rsi";
+$acc1="%edi";	$maskfe="%rdi";
+$acc2="%ebp";	$mask1b="%rbp";
 $inp="%r8";
 $out="%r9";
 $t0="%r10d";
@@ -51,6 +56,8 @@ sub hi() { my $r=shift;	$r =~ s/%[er]([a-d])x/%\1h/;	$r; }
 sub lo() { my $r=shift;	$r =~ s/%[er]([a-d])x/%\1l/;
 			$r =~ s/%[er]([sd]i)/%\1l/;
 			$r =~ s/%(r[0-9]+)[d]?/%\1b/;	$r; }
+sub LO() { my $r=shift; $r =~ s/%r([a-z]+)/%e\1/;
+			$r =~ s/%r([0-9]+)/%r\1d/;	$r; }
 sub _data_word()
 { my $i;
     while(defined($i=shift)) { $code.=sprintf".long\t0x%08x,0x%08x\n",$i,$i; }
@@ -345,6 +352,235 @@ $code.=<<___;
 .size	_x86_64_AES_encrypt,.-_x86_64_AES_encrypt
 ___
 
+# it's possible to implement this by shifting tN by 8, filling least
+# significant byte with byte load and finally bswap-ing at the end,
+# but partial byte load kills Core 2...
+sub enccompactvert()
+{ my ($t3,$t4,$t5)=("%r8d","%r9d","%r13d");
+
+$code.=<<___;
+	movzb	`&lo("$s0")`,$t0
+	movzb	`&lo("$s1")`,$t1
+	movzb	`&lo("$s2")`,$t2
+	movzb	($sbox,$t0,1),$t0
+	movzb	($sbox,$t1,1),$t1
+	movzb	($sbox,$t2,1),$t2
+
+	movzb	`&lo("$s3")`,$t3
+	movzb	`&hi("$s1")`,$acc0
+	movzb	`&hi("$s2")`,$acc1
+	movzb	($sbox,$t3,1),$t3
+	movzb	($sbox,$acc0,1),$t4	#$t0
+	movzb	($sbox,$acc1,1),$t5	#$t1
+
+	movzb	`&hi("$s3")`,$acc2
+	movzb	`&hi("$s0")`,$acc0
+	shr	\$16,$s2
+	movzb	($sbox,$acc2,1),$acc2	#$t2
+	movzb	($sbox,$acc0,1),$acc0	#$t3
+	shr	\$16,$s3
+
+	movzb	`&lo("$s2")`,$acc1
+	shl	\$8,$t4
+	shl	\$8,$t5
+	movzb	($sbox,$acc1,1),$acc1	#$t0
+	xor	$t4,$t0
+	xor	$t5,$t1
+
+	movzb	`&lo("$s3")`,$t4
+	shr	\$16,$s0
+	shr	\$16,$s1
+	movzb	`&lo("$s0")`,$t5
+	shl	\$8,$acc2
+	shl	\$8,$acc0
+	movzb	($sbox,$t4,1),$t4	#$t1
+	movzb	($sbox,$t5,1),$t5	#$t2
+	xor	$acc2,$t2
+	xor	$acc0,$t3
+
+	movzb	`&lo("$s1")`,$acc2
+	movzb	`&hi("$s3")`,$acc0
+	shl	\$16,$acc1
+	movzb	($sbox,$acc2,1),$acc2	#$t3
+	movzb	($sbox,$acc0,1),$acc0	#$t0
+	xor	$acc1,$t0
+
+	movzb	`&hi("$s0")`,$acc1
+	shr	\$8,$s2
+	shr	\$8,$s1
+	movzb	($sbox,$acc1,1),$acc1	#$t1
+	movzb	($sbox,$s2,1),$s3	#$t3
+	movzb	($sbox,$s1,1),$s2	#$t2
+	shl	\$16,$t4
+	shl	\$16,$t5
+	shl	\$16,$acc2
+	xor	$t4,$t1
+	xor	$t5,$t2
+	xor	$acc2,$t3
+
+	shl	\$24,$acc0
+	shl	\$24,$acc1
+	shl	\$24,$s3
+	xor	$acc0,$t0
+	shl	\$24,$s2
+	xor	$acc1,$t1
+	mov	$t0,$s0
+	mov	$t1,$s1
+	xor	$t2,$s2
+	xor	$t3,$s3
+___
+}
+
+sub enctransform_ref()
+{ my $sn = shift;
+  my ($acc,$r2,$tmp)=("%r8d","%r9d","%r13d");
+
+$code.=<<___;
+	mov	$sn,$acc
+	and	\$0x80808080,$acc
+	mov	$acc,$tmp
+	shr	\$7,$tmp
+	lea	($sn,$sn),$r2
+	sub	$tmp,$acc
+	and	\$0xfefefefe,$r2
+	and	\$0x1b1b1b1b,$acc
+	mov	$sn,$tmp
+	xor	$acc,$r2
+
+	xor	$r2,$sn
+	rol	\$24,$sn
+	xor	$r2,$sn
+	ror	\$16,$tmp
+	xor	$tmp,$sn
+	ror	\$8,$tmp
+	xor	$tmp,$sn
+___
+}
+
+# unlike decrypt case it does not pay off to parallelize enctransform
+sub enctransform()
+{ my ($t3,$r20,$r21)=($acc2,"%r8d","%r9d");
+
+$code.=<<___;
+	mov	$s0,$acc0
+	mov	$s1,$acc1
+	and	\$0x80808080,$acc0
+	and	\$0x80808080,$acc1
+	mov	$acc0,$t0
+	mov	$acc1,$t1
+	shr	\$7,$t0
+	lea	($s0,$s0),$r20
+	shr	\$7,$t1
+	lea	($s1,$s1),$r21
+	sub	$t0,$acc0
+	sub	$t1,$acc1
+	and	\$0xfefefefe,$r20
+	and	\$0xfefefefe,$r21
+	and	\$0x1b1b1b1b,$acc0
+	and	\$0x1b1b1b1b,$acc1
+	mov	$s0,$t0
+	mov	$s1,$t1
+	xor	$acc0,$r20
+	xor	$acc1,$r21
+
+	xor	$r20,$s0
+	xor	$r21,$s1
+	 mov	$s2,$acc0
+	 mov	$s3,$acc1
+	rol	\$24,$s0
+	rol	\$24,$s1
+	 and	\$0x80808080,$acc0
+	 and	\$0x80808080,$acc1
+	xor	$r20,$s0
+	xor	$r21,$s1
+	 mov	$acc0,$t2
+	 mov	$acc1,$t3
+	ror	\$16,$t0
+	ror	\$16,$t1
+	 shr	\$7,$t2
+	 lea	($s2,$s2),$r20
+	xor	$t0,$s0
+	xor	$t1,$s1
+	 shr	\$7,$t3
+	 lea	($s3,$s3),$r21
+	ror	\$8,$t0
+	ror	\$8,$t1
+	 sub	$t2,$acc0
+	 sub	$t3,$acc1
+	xor	$t0,$s0
+	xor	$t1,$s1
+
+	and	\$0xfefefefe,$r20
+	and	\$0xfefefefe,$r21
+	and	\$0x1b1b1b1b,$acc0
+	and	\$0x1b1b1b1b,$acc1
+	mov	$s2,$t2
+	mov	$s3,$t3
+	xor	$acc0,$r20
+	xor	$acc1,$r21
+
+	xor	$r20,$s2
+	xor	$r21,$s3
+	rol	\$24,$s2
+	rol	\$24,$s3
+	xor	$r20,$s2
+	xor	$r21,$s3
+	ror	\$16,$t2
+	ror	\$16,$t3
+	xor	$t2,$s2
+	xor	$t3,$s3
+	ror	\$8,$t2
+	ror	\$8,$t3
+	xor	$t2,$s2
+	xor	$t3,$s3
+
+___
+}
+
+$code.=<<___;
+.type	_x86_64_AES_encrypt_compact,\@abi-omnipotent
+.align	16
+_x86_64_AES_encrypt_compact:
+	lea	128($sbox),$acc0		# size optimization
+	mov	0-128($acc0),$acc1		# prefetch Te4
+	mov	32-128($acc0),$acc2
+	mov	64-128($acc0),$t0
+	mov	96-128($acc0),$t1
+	mov	128-128($acc0),$acc1
+	mov	160-128($acc0),$acc2
+	mov	192-128($acc0),$t0
+	mov	224-128($acc0),$t1
+
+	xor	0($key),$s0			# xor with key
+	xor	4($key),$s1
+	xor	8($key),$s2
+	xor	12($key),$s3
+	lea	16($key),$key
+	jmp	.Lenc_loop_compact
+.align	16
+.Lenc_loop_compact:
+___
+		&enccompactvert();
+		&enctransform();
+$code.=<<___;
+		xor	0($key),$s0
+		xor	4($key),$s1
+		xor	8($key),$s2
+		xor	12($key),$s3
+		lea	16($key),$key
+		cmp	16(%rsp),$key
+	jne	.Lenc_loop_compact
+___
+	&enccompactvert();
+$code.=<<___;
+	xor	0($key),$s0
+	xor	4($key),$s1
+	xor	8($key),$s2
+	xor	12($key),$s3
+	.byte	0xf3,0xc3			# rep ret
+.size	_x86_64_AES_encrypt_compact,.-_x86_64_AES_encrypt_compact
+___
+
 # void AES_encrypt (const void *inp,void *out,const AES_KEY *key);
 $code.=<<___;
 .globl	AES_encrypt
@@ -358,25 +594,49 @@ AES_encrypt:
 	push	%r14
 	push	%r15
 
+	# allocate frame "above" key schedule
+	mov	%rsp,%rax
 	mov	%rdx,$key
-	mov	%rdi,$inp
-	mov	%rsi,$out
+	lea	-63(%rdx),%rcx
+	and	\$-64,%rsp
+	sub	%rsp,%rcx
+	neg	%rcx
+	and	\$0x3c0,%rcx
+	sub	%rcx,%rsp
 
-	.picmeup	$sbox
-	lea	AES_Te-.($sbox),$sbox
+	push	%rax		# save real stack pointer
+	push	%rsi		# save out
 
-	mov	0($inp),$s0
-	mov	4($inp),$s1
-	mov	8($inp),$s2
-	mov	12($inp),$s3
+	mov	240($key),$rnds	# load rounds
 
-	call	_x86_64_AES_encrypt
+	mov	0(%rdi),$s0	# load input vector
+	mov	4(%rdi),$s1
+	mov	8(%rdi),$s2
+	mov	12(%rdi),$s3
 
-	mov	$s0,0($out)
+	shl	\$4,$rnds
+	lea	($key,$rnds),%rbp
+	push	%rbp
+	push	$key
+
+	# pick Te4 copy which can't "overlap" with stack frame or key schedule
+	.picmeup	$sbox
+	lea	AES_Te+2048-.($sbox),$sbox
+	lea	768-32(%rsp),%rbp
+	sub	$sbox,%rbp
+	and	\$0x300,%rbp
+	lea	($sbox,%rbp),$sbox
+
+	call	_x86_64_AES_encrypt_compact
+
+	lea	16(%rsp),%rsp
+	pop	$out		# restore out
+	mov	$s0,0($out)	# write output vector
 	mov	$s1,4($out)
 	mov	$s2,8($out)
 	mov	$s3,12($out)
 
+	mov	(%rsp),%rsp
 	pop	%r15
 	pop	%r14
 	pop	%r13
@@ -453,19 +713,20 @@ sub declastvert()
 { my $t3="%r8d";	# zaps $inp!
 
 $code.=<<___;
+	lea	2048($sbox),$sbox	# size optimization
 	movzb	`&lo("$s0")`,$acc0
 	movzb	`&lo("$s1")`,$acc1
 	movzb	`&lo("$s2")`,$acc2
-	movzb	2048($sbox,$acc0,1),$t0
-	movzb	2048($sbox,$acc1,1),$t1
-	movzb	2048($sbox,$acc2,1),$t2
+	movzb	($sbox,$acc0,1),$t0
+	movzb	($sbox,$acc1,1),$t1
+	movzb	($sbox,$acc2,1),$t2
 
 	movzb	`&lo("$s3")`,$acc0
 	movzb	`&hi("$s3")`,$acc1
 	movzb	`&hi("$s0")`,$acc2
-	movzb	2048($sbox,$acc0,1),$t3
-	movzb	2048($sbox,$acc1,1),$acc1	#$t0
-	movzb	2048($sbox,$acc2,1),$acc2	#$t1
+	movzb	($sbox,$acc0,1),$t3
+	movzb	($sbox,$acc1,1),$acc1	#$t0
+	movzb	($sbox,$acc2,1),$acc2	#$t1
 
 	shl	\$8,$acc1
 	shl	\$8,$acc2
@@ -477,8 +738,8 @@ $code.=<<___;
 	movzb	`&hi("$s1")`,$acc0
 	movzb	`&hi("$s2")`,$acc1
 	shr	\$16,$s0
-	movzb	2048($sbox,$acc0,1),$acc0	#$t2
-	movzb	2048($sbox,$acc1,1),$acc1	#$t3
+	movzb	($sbox,$acc0,1),$acc0	#$t2
+	movzb	($sbox,$acc1,1),$acc1	#$t3
 
 	shl	\$8,$acc0
 	shl	\$8,$acc1
@@ -490,9 +751,9 @@ $code.=<<___;
 	movzb	`&lo("$s2")`,$acc0
 	movzb	`&lo("$s3")`,$acc1
 	movzb	`&lo("$s0")`,$acc2
-	movzb	2048($sbox,$acc0,1),$acc0	#$t0
-	movzb	2048($sbox,$acc1,1),$acc1	#$t1
-	movzb	2048($sbox,$acc2,1),$acc2	#$t2
+	movzb	($sbox,$acc0,1),$acc0	#$t0
+	movzb	($sbox,$acc1,1),$acc1	#$t1
+	movzb	($sbox,$acc2,1),$acc2	#$t2
 
 	shl	\$16,$acc0
 	shl	\$16,$acc1
@@ -505,9 +766,9 @@ $code.=<<___;
 	movzb	`&lo("$s1")`,$acc0
 	movzb	`&hi("$s1")`,$acc1
 	movzb	`&hi("$s2")`,$acc2
-	movzb	2048($sbox,$acc0,1),$acc0	#$t3
-	movzb	2048($sbox,$acc1,1),$acc1	#$t0
-	movzb	2048($sbox,$acc2,1),$acc2	#$t1
+	movzb	($sbox,$acc0,1),$acc0	#$t3
+	movzb	($sbox,$acc1,1),$acc1	#$t0
+	movzb	($sbox,$acc2,1),$acc2	#$t1
 
 	shl	\$16,$acc0
 	shl	\$24,$acc1
@@ -520,8 +781,8 @@ $code.=<<___;
 	movzb	`&hi("$s3")`,$acc0
 	movzb	`&hi("$s0")`,$acc1
 	mov	16+12($key),$s3
-	movzb	2048($sbox,$acc0,1),$acc0	#$t2
-	movzb	2048($sbox,$acc1,1),$acc1	#$t3
+	movzb	($sbox,$acc0,1),$acc0	#$t2
+	movzb	($sbox,$acc1,1),$acc1	#$t3
 	mov	16+0($key),$s0
 
 	shl	\$24,$acc0
@@ -532,6 +793,7 @@ $code.=<<___;
 
 	mov	16+4($key),$s1
 	mov	16+8($key),$s2
+	lea	-2048($sbox),$sbox
 	xor	$t0,$s0
 	xor	$t1,$s1
 	xor	$t2,$s2
@@ -659,6 +921,260 @@ $code.=<<___;
 .size	_x86_64_AES_decrypt,.-_x86_64_AES_decrypt
 ___
 
+sub deccompactvert()
+{ my ($t3,$t4,$t5)=("%r8d","%r9d","%r13d");
+
+$code.=<<___;
+	movzb	`&lo("$s0")`,$t0
+	movzb	`&lo("$s1")`,$t1
+	movzb	`&lo("$s2")`,$t2
+	movzb	($sbox,$t0,1),$t0
+	movzb	($sbox,$t1,1),$t1
+	movzb	($sbox,$t2,1),$t2
+
+	movzb	`&lo("$s3")`,$t3
+	movzb	`&hi("$s3")`,$acc0
+	movzb	`&hi("$s0")`,$acc1
+	movzb	($sbox,$t3,1),$t3
+	movzb	($sbox,$acc0,1),$t4	#$t0
+	movzb	($sbox,$acc1,1),$t5	#$t1
+
+	movzb	`&hi("$s1")`,$acc2
+	movzb	`&hi("$s2")`,$acc0
+	shr	\$16,$s2
+	movzb	($sbox,$acc2,1),$acc2	#$t2
+	movzb	($sbox,$acc0,1),$acc0	#$t3
+	shr	\$16,$s3
+
+	movzb	`&lo("$s2")`,$acc1
+	shl	\$8,$t4
+	shl	\$8,$t5
+	movzb	($sbox,$acc1,1),$acc1	#$t0
+	xor	$t4,$t0
+	xor	$t5,$t1
+
+	movzb	`&lo("$s3")`,$t4
+	shr	\$16,$s0
+	shr	\$16,$s1
+	movzb	`&lo("$s0")`,$t5
+	shl	\$8,$acc2
+	shl	\$8,$acc0
+	movzb	($sbox,$t4,1),$t4	#$t1
+	movzb	($sbox,$t5,1),$t5	#$t2
+	xor	$acc2,$t2
+	xor	$acc0,$t3
+
+	movzb	`&lo("$s1")`,$acc2
+	movzb	`&hi("$s1")`,$acc0
+	shl	\$16,$acc1
+	movzb	($sbox,$acc2,1),$acc2	#$t3
+	movzb	($sbox,$acc0,1),$acc0	#$t0
+	xor	$acc1,$t0
+
+	movzb	`&hi("$s2")`,$acc1
+	shl	\$16,$t4
+	shl	\$16,$t5
+	movzb	($sbox,$acc1,1),$s1	#$t1
+	xor	$t4,$t1
+	xor	$t5,$t2
+
+	movzb	`&hi("$s3")`,$acc1
+	shr	\$8,$s0
+	shl	\$16,$acc2
+	movzb	($sbox,$acc1,1),$s2	#$t2
+	movzb	($sbox,$s0,1),$s3	#$t3
+	xor	$acc2,$t3
+
+	shl	\$24,$acc0
+	shl	\$24,$s1
+	shl	\$24,$s2
+	xor	$acc0,$t0
+	shl	\$24,$s3
+	xor	$t1,$s1
+	mov	$t0,$s0
+	xor	$t2,$s2
+	xor	$t3,$s3
+___
+}
+
+# parallelized version! input is pair of 64-bit values: %rax=s1.s0
+# and %rcx=s3.s2, output is four 32-bit values in %eax=s0, %ebx=s1,
+# %ecx=s2 and %edx=s3.
+sub dectransform()
+{ my ($tp10,$tp20,$tp40,$tp80,$acc0)=("%rax","%r8", "%r9", "%r10","%rbx");
+  my ($tp18,$tp28,$tp48,$tp88,$acc8)=("%rcx","%r11","%r12","%r13","%rdx");
+
+$code.=<<___;
+	mov	$tp10,$acc0
+	mov	$tp18,$acc8
+	and	$mask80,$acc0
+	and	$mask80,$acc8
+	mov	$acc0,$tp40
+	mov	$acc8,$tp48
+	shr	\$7,$tp40
+	lea	($tp10,$tp10),$tp20
+	shr	\$7,$tp48
+	lea	($tp18,$tp18),$tp28
+	sub	$tp40,$acc0
+	sub	$tp48,$acc8
+	and	$maskfe,$tp20
+	and	$maskfe,$tp28
+	and	$mask1b,$acc0
+	and	$mask1b,$acc8
+	xor	$tp20,$acc0
+	xor	$tp28,$acc8
+	mov	$acc0,$tp20
+	mov	$acc8,$tp28
+
+	and	$mask80,$acc0
+	and	$mask80,$acc8
+	mov	$acc0,$tp80
+	mov	$acc8,$tp88
+	shr	\$7,$tp80
+	lea	($tp20,$tp20),$tp40
+	shr	\$7,$tp88
+	lea	($tp28,$tp28),$tp48
+	sub	$tp80,$acc0
+	sub	$tp88,$acc8
+	and	$maskfe,$tp40
+	and	$maskfe,$tp48
+	and	$mask1b,$acc0
+	and	$mask1b,$acc8
+	xor	$tp40,$acc0
+	xor	$tp48,$acc8
+	mov	$acc0,$tp40
+	mov	$acc8,$tp48
+
+	and	$mask80,$acc0
+	and	$mask80,$acc8
+	mov	$acc0,$tp80
+	mov	$acc8,$tp88
+	shr	\$7,$tp80
+	 xor	$tp10,$tp20		# tp2^=tp1
+	shr	\$7,$tp88
+	 xor	$tp18,$tp28		# tp2^=tp1
+	sub	$tp80,$acc0
+	sub	$tp88,$acc8
+	lea	($tp40,$tp40),$tp80
+	lea	($tp48,$tp48),$tp88
+	 xor	$tp10,$tp40		# tp4^=tp1
+	 xor	$tp18,$tp48		# tp4^=tp1
+	and	$maskfe,$tp80
+	and	$maskfe,$tp88
+	and	$mask1b,$acc0
+	and	$mask1b,$acc8
+	xor	$acc0,$tp80
+	xor	$acc8,$tp88
+
+	xor	$tp80,$tp10		# tp1^=tp8
+	xor	$tp88,$tp18		# tp1^=tp8
+	xor	$tp80,$tp20		# tp2^tp1^=tp8
+	xor	$tp88,$tp28		# tp2^tp1^=tp8
+	mov	$tp10,$acc0
+	mov	$tp18,$acc8
+	xor	$tp80,$tp40		# tp4^tp1^=tp8
+	xor	$tp88,$tp48		# tp4^tp1^=tp8
+	shr	\$32,$acc0
+	shr	\$32,$acc8
+	xor	$tp20,$tp80		# tp8^=tp8^tp2^tp1=tp2^tp1
+	xor	$tp28,$tp88		# tp8^=tp8^tp2^tp1=tp2^tp1
+	rol	\$8,`&LO("$tp10")`	# ROTATE(tp1^tp8,8)
+	rol	\$8,`&LO("$tp18")`	# ROTATE(tp1^tp8,8)
+	xor	$tp40,$tp80		# tp2^tp1^=tp8^tp4^tp1=tp8^tp4^tp2
+	xor	$tp48,$tp88		# tp2^tp1^=tp8^tp4^tp1=tp8^tp4^tp2
+
+	rol	\$8,`&LO("$acc0")`	# ROTATE(tp1^tp8,8)
+	rol	\$8,`&LO("$acc8")`	# ROTATE(tp1^tp8,8)
+	xor	`&LO("$tp80")`,`&LO("$tp10")`
+	xor	`&LO("$tp88")`,`&LO("$tp18")`
+	shr	\$32,$tp80
+	shr	\$32,$tp88
+	xor	`&LO("$tp80")`,`&LO("$acc0")`
+	xor	`&LO("$tp88")`,`&LO("$acc8")`
+
+	mov	$tp20,$tp80
+	mov	$tp28,$tp88
+	shr	\$32,$tp80
+	shr	\$32,$tp88
+	rol	\$24,`&LO("$tp20")`	# ROTATE(tp2^tp1^tp8,24)
+	rol	\$24,`&LO("$tp28")`	# ROTATE(tp2^tp1^tp8,24)
+	rol	\$24,`&LO("$tp80")`	# ROTATE(tp2^tp1^tp8,24)
+	rol	\$24,`&LO("$tp88")`	# ROTATE(tp2^tp1^tp8,24)
+	xor	`&LO("$tp20")`,`&LO("$tp10")`
+	xor	`&LO("$tp28")`,`&LO("$tp18")`
+	mov	$tp40,$tp20
+	mov	$tp48,$tp28
+	xor	`&LO("$tp80")`,`&LO("$acc0")`
+	xor	`&LO("$tp88")`,`&LO("$acc8")`
+
+	shr	\$32,$tp20
+	shr	\$32,$tp28
+	rol	\$16,`&LO("$tp40")`	# ROTATE(tp4^tp1^tp8,16)
+	rol	\$16,`&LO("$tp48")`	# ROTATE(tp4^tp1^tp8,16)
+	rol	\$16,`&LO("$tp20")`	# ROTATE(tp4^tp1^tp8,16)
+	rol	\$16,`&LO("$tp28")`	# ROTATE(tp4^tp1^tp8,16)
+	xor	`&LO("$tp40")`,`&LO("$tp10")`
+	xor	`&LO("$tp48")`,`&LO("$tp18")`
+	xor	`&LO("$tp20")`,`&LO("$acc0")`
+	xor	`&LO("$tp28")`,`&LO("$acc8")`
+
+___
+}
+
+$code.=<<___;
+.type	_x86_64_AES_decrypt_compact,\@abi-omnipotent
+.align	16
+_x86_64_AES_decrypt_compact:
+	lea	128($sbox),$acc0		# size optimization
+	mov	0-128($acc0),$acc1		# prefetch Td4
+	mov	32-128($acc0),$acc2
+	mov	64-128($acc0),$t0
+	mov	96-128($acc0),$t1
+	mov	128-128($acc0),$acc1
+	mov	160-128($acc0),$acc2
+	mov	192-128($acc0),$t0
+	mov	224-128($acc0),$t1
+
+	xor	0($key),$s0			# xor with key
+	xor	4($key),$s1
+	xor	8($key),$s2
+	xor	12($key),$s3
+	lea	16($key),$key
+
+	jmp	.Ldec_compact_loop
+.align	16
+.Ldec_compact_loop:
+___
+		&deccompactvert();
+$code.=<<___;
+		mov	256+0($sbox),$mask80
+		shl	\$32,%rbx
+		shl	\$32,%rdx
+		mov	256+8($sbox),$maskfe
+		or	%rbx,%rax
+		or	%rdx,%rcx
+		mov	256+16($sbox),$mask1b
+___
+		&dectransform();
+$code.=<<___;
+		xor	0($key),$s0
+		xor	4($key),$s1
+		xor	8($key),$s2
+		xor	12($key),$s3
+		lea	16($key),$key
+		cmp	16(%rsp),$key
+	jne	.Ldec_compact_loop
+___
+	&deccompactvert();
+$code.=<<___;
+	xor	0($key),$s0
+	xor	4($key),$s1
+	xor	8($key),$s2
+	xor	12($key),$s3
+	.byte	0xf3,0xc3			# rep ret
+.size	_x86_64_AES_decrypt_compact,.-_x86_64_AES_decrypt_compact
+___
+
 # void AES_decrypt (const void *inp,void *out,const AES_KEY *key);
 $code.=<<___;
 .globl	AES_decrypt
@@ -672,37 +1188,51 @@ AES_decrypt:
 	push	%r14
 	push	%r15
 
+	# allocate frame "above" key schedule
+	mov	%rsp,%rax
 	mov	%rdx,$key
-	mov	%rdi,$inp
-	mov	%rsi,$out
+	lea	-63(%rdx),%rcx
+	and	\$-64,%rsp
+	sub	%rsp,%rcx
+	neg	%rcx
+	and	\$0x3c0,%rcx
+	sub	%rcx,%rsp
 
-	.picmeup	$sbox
-	lea	AES_Td-.($sbox),$sbox
-
-	# prefetch Td4
-	lea	2048+128($sbox),$sbox;
-	mov	0-128($sbox),$s0
-	mov	32-128($sbox),$s1
-	mov	64-128($sbox),$s2
-	mov	96-128($sbox),$s3
-	mov	128-128($sbox),$s0
-	mov	160-128($sbox),$s1
-	mov	192-128($sbox),$s2
-	mov	224-128($sbox),$s3
-	lea	-2048-128($sbox),$sbox;
-
-	mov	0($inp),$s0
-	mov	4($inp),$s1
-	mov	8($inp),$s2
-	mov	12($inp),$s3
-
-	call	_x86_64_AES_decrypt
+	push	%rax		# save real stack pointer
+	push	%rsi		# save out
+
+	mov	240($key),$rnds	# load rounds
+
+	mov	0(%rdi),$s0	# load input vector
+	mov	4(%rdi),$s1
+	mov	8(%rdi),$s2
+	mov	12(%rdi),$s3
+
+	shl	\$4,$rnds
+	lea	($key,$rnds),%rbp
+	push	%rbp
+	push	$key
 
+	# pick Td4 copy which can't "overlap" with stack frame or key schedule
+	.picmeup	$sbox
+	lea	AES_Td+2048-.($sbox),$sbox
+	lea	768-32(%rsp),%rbp
+	sub	$sbox,%rbp
+	and	\$0x300,%rbp
+	lea	($sbox,%rbp),$sbox
+	shr	\$3,%rbp		# recall "magic" constants!
+	add	%rbp,$sbox
+
+	call	_x86_64_AES_decrypt_compact
+
+	lea	16(%rsp),%rsp
+	pop	$out	# restore out
 	mov	$s0,0($out)
 	mov	$s1,4($out)
 	mov	$s2,8($out)
 	mov	$s3,12($out)
 
+	mov	(%rsp),%rsp
 	pop	%r15
 	pop	%r14
 	pop	%r13
@@ -718,27 +1248,26 @@ sub enckey()
 {
 $code.=<<___;
 	movz	%dl,%esi		# rk[i]>>0
-	mov	2(%rbp,%rsi,8),%ebx
+	movzb	-128(%rbp,%rsi),%ebx
 	movz	%dh,%esi		# rk[i]>>8
-	and	\$0xFF000000,%ebx
+	shl	\$24,%ebx
 	xor	%ebx,%eax
 
-	mov	2(%rbp,%rsi,8),%ebx
+	movzb	-128(%rbp,%rsi),%ebx
 	shr	\$16,%edx
-	and	\$0x000000FF,%ebx
 	movz	%dl,%esi		# rk[i]>>16
 	xor	%ebx,%eax
 
-	mov	0(%rbp,%rsi,8),%ebx
+	movzb	-128(%rbp,%rsi),%ebx
 	movz	%dh,%esi		# rk[i]>>24
-	and	\$0x0000FF00,%ebx
+	shl	\$8,%ebx
 	xor	%ebx,%eax
 
-	mov	0(%rbp,%rsi,8),%ebx
-	and	\$0x00FF0000,%ebx
+	movzb	-128(%rbp,%rsi),%ebx
+	shl	\$16,%ebx
 	xor	%ebx,%eax
 
-	xor	2048(%rbp,%rcx,4),%eax		# rcon
+	xor	1024-128(%rbp,%rcx,4),%eax		# rcon
 ___
 }
 
@@ -763,6 +1292,17 @@ AES_set_encrypt_key:
 
 	.picmeup %rbp
 	lea	AES_Te-.(%rbp),%rbp
+	lea	2048+128(%rbp),%rbp
+
+	# prefetch Te4
+	mov	0-128(%rbp),%eax
+	mov	32-128(%rbp),%ebx
+	mov	64-128(%rbp),%r8d
+	mov	96-128(%rbp),%edx
+	mov	128-128(%rbp),%eax
+	mov	160-128(%rbp),%ebx
+	mov	192-128(%rbp),%r8d
+	mov	224-128(%rbp),%edx
 
 	cmp	\$128,%ecx
 	je	.L10rounds
@@ -900,24 +1440,23 @@ $code.=<<___;
 		mov	%eax,%edx
 		mov	16(%rdi),%eax			# rk[4]
 		movz	%dl,%esi			# rk[11]>>0
-		mov	2(%rbp,%rsi,8),%ebx
+		movzb	-128(%rbp,%rsi),%ebx
 		movz	%dh,%esi			# rk[11]>>8
-		and	\$0x000000FF,%ebx
 		xor	%ebx,%eax
 
-		mov	0(%rbp,%rsi,8),%ebx
+		movzb	-128(%rbp,%rsi),%ebx
 		shr	\$16,%edx
-		and	\$0x0000FF00,%ebx
+		shl	\$8,%ebx
 		movz	%dl,%esi			# rk[11]>>16
 		xor	%ebx,%eax
 
-		mov	0(%rbp,%rsi,8),%ebx
+		movzb	-128(%rbp,%rsi),%ebx
 		movz	%dh,%esi			# rk[11]>>24
-		and	\$0x00FF0000,%ebx
+		shl	\$16,%ebx
 		xor	%ebx,%eax
 
-		mov	2(%rbp,%rsi,8),%ebx
-		and	\$0xFF000000,%ebx
+		movzb	-128(%rbp,%rsi),%ebx
+		shl	\$24,%ebx
 		xor	%ebx,%eax
 
 		mov	%eax,48(%rdi)			# rk[12]
@@ -944,25 +1483,57 @@ $code.=<<___;
 .size	AES_set_encrypt_key,.-AES_set_encrypt_key
 ___
 
-sub deckey()
+sub deckey_ref()
 { my ($i,$ptr,$te,$td) = @_;
+  my ($tp1,$tp2,$tp4,$tp8,$acc)=("%eax","%ebx","%edi","%edx","%r8d");
 $code.=<<___;
-	mov	$i($ptr),%eax
-	mov	%eax,%edx
-	movz	%ah,%ebx
-	shr	\$16,%edx
-	and	\$0xFF,%eax
-	movzb	2($te,%rax,8),%rax
-	movzb	2($te,%rbx,8),%rbx
-	mov	0($td,%rax,8),%eax
-	xor	3($td,%rbx,8),%eax
-	movzb	%dh,%ebx
-	and	\$0xFF,%edx
-	movzb	2($te,%rdx,8),%rdx
-	movzb	2($te,%rbx,8),%rbx
-	xor	2($td,%rdx,8),%eax
-	xor	1($td,%rbx,8),%eax
-	mov	%eax,$i($ptr)
+	mov	$i($ptr),$tp1
+	mov	$tp1,$acc
+	and	\$0x80808080,$acc
+	mov	$acc,$tp4
+	shr	\$7,$tp4
+	lea	0($tp1,$tp1),$tp2
+	sub	$tp4,$acc
+	and	\$0xfefefefe,$tp2
+	and	\$0x1b1b1b1b,$acc
+	xor	$tp2,$acc
+	mov	$acc,$tp2
+
+	and	\$0x80808080,$acc
+	mov	$acc,$tp8
+	shr	\$7,$tp8
+	lea	0($tp2,$tp2),$tp4
+	sub	$tp8,$acc
+	and	\$0xfefefefe,$tp4
+	and	\$0x1b1b1b1b,$acc
+	 xor	$tp1,$tp2		# tp2^tp1
+	xor	$tp4,$acc
+	mov	$acc,$tp4
+
+	and	\$0x80808080,$acc
+	mov	$acc,$tp8
+	shr	\$7,$tp8
+	sub	$tp8,$acc
+	lea	0($tp4,$tp4),$tp8
+	 xor	$tp1,$tp4		# tp4^tp1
+	and	\$0xfefefefe,$tp8
+	and	\$0x1b1b1b1b,$acc
+	xor	$acc,$tp8
+
+	xor	$tp8,$tp1		# tp1^tp8
+	rol	\$8,$tp1		# ROTATE(tp1^tp8,8)
+	xor	$tp8,$tp2		# tp2^tp1^tp8
+	xor	$tp8,$tp4		# tp4^tp1^tp8
+	xor	$tp2,$tp8
+	xor	$tp4,$tp8		# tp8^(tp8^tp4^tp1)^(tp8^tp2^tp1)=tp8^tp4^tp2
+
+	xor	$tp8,$tp1
+	rol	\$24,$tp2		# ROTATE(tp2^tp1^tp8,24)
+	xor	$tp2,$tp1
+	rol	\$16,$tp4		# ROTATE(tp4^tp1^tp8,16)
+	xor	$tp4,$tp1
+
+	mov	$tp1,$i($ptr)
 ___
 }
 
@@ -973,19 +1544,22 @@ $code.=<<___;
 .type	AES_set_decrypt_key,\@function,3
 .align	16
 AES_set_decrypt_key:
-	push	%rdx
+	push	%rdx			# save key schedule
 	call	AES_set_encrypt_key
 	cmp	\$0,%eax
-	je	.Lproceed
-	lea	24(%rsp),%rsp
-	ret
-.Lproceed:
-	mov	(%rsp),%r8		# restore key schedule
-	mov	%rbx,(%rsp)
+	pop	%r8			# restore key schedule
+	jne	.Labort
+
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
 
-	mov	240(%r8),%ecx		# pull number of rounds
+	mov	240(%r8),%r14d		# pull number of rounds
 	xor	%rdi,%rdi
-	lea	(%rdi,%rcx,4),%rcx
+	lea	(%rdi,%r14d,4),%rcx
 	mov	%r8,%rsi
 	lea	(%r8,%rcx,4),%rdi	# pointer to last chunk
 .align	4
@@ -1003,27 +1577,38 @@ AES_set_decrypt_key:
 		cmp	%rsi,%rdi
 	jne	.Linvert
 
-	.picmeup %r9
-	lea	AES_Td-.(%r9),%rdi
-	lea	AES_Te-AES_Td(%rdi),%r9
+	.picmeup %rax
+	lea	AES_Te+2048+1024-.(%rax),%rax	# rcon
 
-	mov	%r8,%rsi
-	mov	240(%r8),%ecx		# pull number of rounds
-	sub	\$1,%ecx
+	mov	40(%rax),$mask80
+	mov	48(%rax),$maskfe
+	mov	56(%rax),$mask1b
+
+	mov	%r8,$key
+	sub	\$1,%r14d
 .align	4
 .Lpermute:
-		lea	16(%rsi),%rsi
+		lea	16($key),$key
+		mov	0($key),%rax
+		mov	8($key),%rcx
 ___
-		&deckey	(0,"%rsi","%r9","%rdi");
-		&deckey	(4,"%rsi","%r9","%rdi");
-		&deckey	(8,"%rsi","%r9","%rdi");
-		&deckey	(12,"%rsi","%r9","%rdi");
+		&dectransform ();
 $code.=<<___;
-		sub	\$1,%ecx
+		mov	%eax,0($key)
+		mov	%ebx,4($key)
+		mov	%ecx,8($key)
+		mov	%edx,12($key)
+		sub	\$1,%r14d
 	jnz	.Lpermute
 
 	xor	%rax,%rax
+	pop	%r15
+	pop	%r14
+	pop	%r13
+	pop	%r12
+	pop	%rbp
 	pop	%rbx
+.Labort:
 	ret
 .size	AES_set_decrypt_key,.-AES_set_decrypt_key
 ___
@@ -1461,11 +2046,145 @@ ___
 	&_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
 	&_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
 	&_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
+
+#Te4	# four copies of Te4 to choose from to avoid L1 aliasing
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
 #rcon:
 $code.=<<___;
 	.long	0x00000001, 0x00000002, 0x00000004, 0x00000008
 	.long	0x00000010, 0x00000020, 0x00000040, 0x00000080
-	.long	0x0000001b, 0x00000036, 0, 0, 0, 0, 0, 0
+	.long	0x0000001b, 0x00000036, 0x80808080, 0x80808080
+	.long	0xfefefefe, 0xfefefefe, 0x1b1b1b1b, 0x1b1b1b1b
 ___
 $code.=<<___;
 .globl	AES_Td
@@ -1536,7 +2255,80 @@ ___
 	&_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
 	&_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
 	&_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
-#Td4:
+
+#Td4:	# four copies of Td4 to choose from to avoid L1 aliasing
+	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+$code.=<<___;
+	.long	0x80808080, 0x80808080, 0xfefefefe, 0xfefefefe
+	.long	0x1b1b1b1b, 0x1b1b1b1b, 0, 0
+___
+	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+$code.=<<___;
+	.long	0x80808080, 0x80808080, 0xfefefefe, 0xfefefefe
+	.long	0x1b1b1b1b, 0x1b1b1b1b, 0, 0
+___
 	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
 	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
 	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
@@ -1569,6 +2361,47 @@ ___
 	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
 	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
 	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+$code.=<<___;
+	.long	0x80808080, 0x80808080, 0xfefefefe, 0xfefefefe
+	.long	0x1b1b1b1b, 0x1b1b1b1b, 0, 0
+___
+	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+$code.=<<___;
+	.long	0x80808080, 0x80808080, 0xfefefefe, 0xfefefefe
+	.long	0x1b1b1b1b, 0x1b1b1b1b, 0, 0
+.asciz  "AES for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
+___
 
 $code =~ s/\`([^\`]*)\`/eval($1)/gem;
 
-- 
2.40.0