x86_64 assembler pack update from HEAD.

3 files changed, 4216 insertions, 0 deletions

diff --git a/crypto/aes/asm/aes-x86_64.pl b/crypto/aes/asm/aes-x86_64.pl
index 7d09b35ac2..48fa857d5b 100755
--- a/crypto/aes/asm/aes-x86_64.pl
+++ b/crypto/aes/asm/aes-x86_64.pl
@@ -588,6 +588,9 @@ $code.=<<___;
 .globl	AES_encrypt
 .type	AES_encrypt,\@function,3
 .align	16
+.globl	asm_AES_encrypt
+.hidden	asm_AES_encrypt
+asm_AES_encrypt:
 AES_encrypt:
 	push	%rbx
 	push	%rbp
@@ -1184,6 +1187,9 @@ $code.=<<___;
 .globl	AES_decrypt
 .type	AES_decrypt,\@function,3
 .align	16
+.globl	asm_AES_decrypt
+.hidden	asm_AES_decrypt
+asm_AES_decrypt:
 AES_decrypt:
 	push	%rbx
 	push	%rbp
@@ -1648,6 +1654,9 @@ $code.=<<___;
 .type	AES_cbc_encrypt,\@function,6
 .align	16
 .extern	OPENSSL_ia32cap_P
+.globl	asm_AES_cbc_encrypt
+.hidden	asm_AES_cbc_encrypt
+asm_AES_cbc_encrypt:
 AES_cbc_encrypt:
 	cmp	\$0,%rdx	# check length
 	je	.Lcbc_epilogue
diff --git a/crypto/aes/asm/bsaes-x86_64.pl b/crypto/aes/asm/bsaes-x86_64.pl
new file mode 100644
index 0000000000..edc70fa1be
--- /dev/null
+++ b/crypto/aes/asm/bsaes-x86_64.pl
@@ -0,0 +1,3003 @@
+#!/usr/bin/env perl
+
+###################################################################
+### AES-128 [originally in CTR mode]				###
+### bitsliced implementation for Intel Core 2 processors	###
+### requires support of SSE extensions up to SSSE3		###
+### Author: Emilia Käsper and Peter Schwabe			###
+### Date: 2009-03-19						###
+### Public domain						###
+###								###
+### See http://homes.esat.kuleuven.be/~ekasper/#software for	###
+### further information.					###
+###################################################################
+#
+# September 2011.
+#
+# Started as transliteration to "perlasm" the original code has
+# undergone following changes:
+#
+# - code was made position-independent;
+# - rounds were folded into a loop resulting in >5x size reduction
+#   from 12.5KB to 2.2KB;
+# - above was possibile thanks to mixcolumns() modification that
+#   allowed to feed its output back to aesenc[last], this was
+#   achieved at cost of two additional inter-registers moves;
+# - some instruction reordering and interleaving;
+# - this module doesn't implement key setup subroutine, instead it
+#   relies on conversion of "conventional" key schedule as returned
+#   by AES_set_encrypt_key (see discussion below);
+# - first and last round keys are treated differently, which allowed
+#   to skip one shiftrows(), reduce bit-sliced key schedule and
+#   speed-up conversion by 22%;
+# - support for 192- and 256-bit keys was added;
+#
+# Resulting performance in CPU cycles spent to encrypt one byte out
+# of 4096-byte buffer with 128-bit key is:
+#
+#		Emilia's	this(*)		difference
+#
+# Core 2    	9.30		8.69		+7%
+# Nehalem(**) 	7.63		6.98		+9%
+# Atom	    	17.1		17.4		-2%(***)
+#
+# (*)	Comparison is not completely fair, because "this" is ECB,
+#	i.e. no extra processing such as counter values calculation
+#	and xor-ing input as in Emilia's CTR implementation is
+#	performed. However, the CTR calculations stand for not more
+#	than 1% of total time, so comparison is *rather* fair.
+#
+# (**)	Results were collected on Westmere, which is considered to
+#	be equivalent to Nehalem for this code.
+#
+# (***)	Slowdown on Atom is rather strange per se, because original
+#	implementation has a number of 9+-bytes instructions, which
+#	are bad for Atom front-end, and which I eliminated completely.
+#	In attempt to address deterioration sbox() was tested in FP
+#	SIMD "domain" (movaps instead of movdqa, xorps instead of
+#	pxor, etc.). While it resulted in nominal 4% improvement on
+#	Atom, it hurted Westmere by more than 2x factor.
+#
+# As for key schedule conversion subroutine. Interface to OpenSSL
+# relies on per-invocation on-the-fly conversion. This naturally
+# has impact on performance, especially for short inputs. Conversion
+# time in CPU cycles and its ratio to CPU cycles spent in 8x block
+# function is:
+#
+# 		conversion	conversion/8x block
+# Core 2	410		0.37
+# Nehalem	310		0.35
+# Atom		570		0.26
+#
+# The ratio values mean that 128-byte blocks will be processed
+# 21-27% slower, 256-byte blocks - 12-16%, 384-byte blocks - 8-11%,
+# etc. Then keep in mind that input sizes not divisible by 128 are
+# *effectively* slower, especially shortest ones, e.g. consecutive
+# 144-byte blocks are processed 44% slower than one would expect,
+# 272 - 29%, 400 - 22%, etc. Yet, despite all these "shortcomings"
+# it's still faster than ["hyper-threading-safe" code path in]
+# aes-x86_64.pl on all lengths above 64 bytes...
+#
+# October 2011.
+#
+# Add decryption procedure. Performance in CPU cycles spent to decrypt
+# one byte out of 4096-byte buffer with 128-bit key is:
+#
+# Core 2	11.0
+# Nehalem	9.16
+#
+# November 2011.
+#
+# Add bsaes_xts_[en|de]crypt. Less-than-80-bytes-block performance is
+# suboptimal, but XTS is meant to be used with larger blocks...
+#
+#						<appro@openssl.org>
+
+$flavour = shift;
+$output  = shift;
+if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
+
+$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $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 $flavour $output";
+
+my ($inp,$out,$len,$key,$ivp)=("%rdi","%rsi","%rdx","%rcx");
+my @XMM=map("%xmm$_",(15,0..14));	# best on Atom, +10% over (0..15)
+my $ecb=0;	# suppress unreferenced ECB subroutines, spare some space...
+
+{
+my ($key,$rounds,$const)=("%rax","%r10d","%r11");
+
+sub Sbox {
+# input in  lsb > [b0, b1, b2, b3, b4, b5, b6, b7] < msb
+# output in lsb > [b0, b1, b4, b6, b3, b7, b2, b5] < msb
+my @b=@_[0..7];
+my @t=@_[8..11];
+my @s=@_[12..15];
+	&InBasisChange	(@b);
+	&Inv_GF256	(@b[6,5,0,3,7,1,4,2],@t,@s);
+	&OutBasisChange	(@b[7,1,4,2,6,5,0,3]);
+}
+
+sub InBasisChange {
+# input in  lsb > [b0, b1, b2, b3, b4, b5, b6, b7] < msb
+# output in lsb > [b6, b5, b0, b3, b7, b1, b4, b2] < msb 
+my @b=@_[0..7];
+$code.=<<___;
+	pxor	@b[6], @b[5]
+	pxor	@b[1], @b[2]
+	pxor	@b[0], @b[3]
+	pxor	@b[2], @b[6]
+	pxor 	@b[0], @b[5]
+
+	pxor	@b[3], @b[6]
+	pxor	@b[7], @b[3]
+	pxor	@b[5], @b[7]
+	pxor	@b[4], @b[3]
+	pxor	@b[5], @b[4]
+	pxor	@b[1], @b[3]
+
+	pxor	@b[7], @b[2]
+	pxor	@b[5], @b[1]
+___
+}
+
+sub OutBasisChange {
+# input in  lsb > [b0, b1, b2, b3, b4, b5, b6, b7] < msb
+# output in lsb > [b6, b1, b2, b4, b7, b0, b3, b5] < msb
+my @b=@_[0..7];
+$code.=<<___;
+	pxor	@b[6], @b[0]
+	pxor	@b[4], @b[1]
+	pxor	@b[0], @b[2]
+	pxor	@b[6], @b[4]
+	pxor	@b[1], @b[6]
+
+	pxor	@b[5], @b[1]
+	pxor	@b[3], @b[5]
+	pxor	@b[7], @b[3]
+	pxor	@b[5], @b[7]
+	pxor	@b[5], @b[2]
+
+	pxor	@b[7], @b[4]
+___
+}
+
+sub InvSbox {
+# input in lsb 	> [b0, b1, b2, b3, b4, b5, b6, b7] < msb
+# output in lsb	> [b0, b1, b6, b4, b2, b7, b3, b5] < msb
+my @b=@_[0..7];
+my @t=@_[8..11];
+my @s=@_[12..15];
+	&InvInBasisChange	(@b);
+	&Inv_GF256		(@b[5,1,2,6,3,7,0,4],@t,@s);
+	&InvOutBasisChange	(@b[3,7,0,4,5,1,2,6]);
+}
+
+sub InvInBasisChange {		# OutBasisChange in reverse
+my @b=@_[5,1,2,6,3,7,0,4];
+$code.=<<___
+	pxor	@b[7], @b[4]
+
+	pxor	@b[5], @b[7]
+	pxor	@b[5], @b[2]
+	pxor	@b[7], @b[3]
+	pxor	@b[3], @b[5]
+	pxor	@b[5], @b[1]
+
+	pxor	@b[1], @b[6]
+	pxor	@b[0], @b[2]
+	pxor	@b[6], @b[4]
+	pxor	@b[6], @b[0]
+	pxor	@b[4], @b[1]
+___
+}
+
+sub InvOutBasisChange {		# InBasisChange in reverse
+my @b=@_[2,5,7,3,6,1,0,4];
+$code.=<<___;
+	pxor	@b[5], @b[1]
+	pxor	@b[7], @b[2]
+
+	pxor	@b[1], @b[3]
+	pxor	@b[5], @b[4]
+	pxor	@b[5], @b[7]
+	pxor	@b[4], @b[3]
+	 pxor 	@b[0], @b[5]
+	pxor	@b[7], @b[3]
+	 pxor	@b[2], @b[6]
+	 pxor	@b[1], @b[2]
+	pxor	@b[3], @b[6]
+
+	pxor	@b[0], @b[3]
+	pxor	@b[6], @b[5]
+___
+}
+
+sub Mul_GF4 {
+#;*************************************************************
+#;* Mul_GF4: Input x0-x1,y0-y1 Output x0-x1 Temp t0 (8) *
+#;*************************************************************
+my ($x0,$x1,$y0,$y1,$t0)=@_;
+$code.=<<___;
+	movdqa	$y0, $t0
+	pxor 	$y1, $t0
+	pand	$x0, $t0
+	pxor	$x1, $x0
+	pand	$y0, $x1
+	pand	$y1, $x0
+	pxor	$x1, $x0
+	pxor	$t0, $x1
+___
+}
+
+sub Mul_GF4_N {				# not used, see next subroutine
+# multiply and scale by N
+my ($x0,$x1,$y0,$y1,$t0)=@_;
+$code.=<<___;
+	movdqa	$y0, $t0
+	pxor	$y1, $t0
+	pand	$x0, $t0
+	pxor	$x1, $x0
+	pand	$y0, $x1
+	pand	$y1, $x0
+	pxor	$x0, $x1
+	pxor	$t0, $x0
+___
+}
+
+sub Mul_GF4_N_GF4 {
+# interleaved Mul_GF4_N and Mul_GF4
+my ($x0,$x1,$y0,$y1,$t0,
+    $x2,$x3,$y2,$y3,$t1)=@_;
+$code.=<<___;
+	movdqa	$y0, $t0
+	 movdqa	$y2, $t1
+	pxor	$y1, $t0
+	 pxor 	$y3, $t1
+	pand	$x0, $t0
+	 pand	$x2, $t1
+	pxor	$x1, $x0
+	 pxor	$x3, $x2
+	pand	$y0, $x1
+	 pand	$y2, $x3
+	pand	$y1, $x0
+	 pand	$y3, $x2
+	pxor	$x0, $x1
+	 pxor	$x3, $x2
+	pxor	$t0, $x0
+	 pxor	$t1, $x3
+___
+}
+sub Mul_GF16_2 {
+my @x=@_[0..7];
+my @y=@_[8..11];
+my @t=@_[12..15];
+$code.=<<___;
+	movdqa	@x[0], @t[0]
+	movdqa	@x[1], @t[1]
+___
+	&Mul_GF4  	(@x[0], @x[1], @y[0], @y[1], @t[2]);
+$code.=<<___;
+	pxor	@x[2], @t[0]
+	pxor	@x[3], @t[1]
+	pxor	@y[2], @y[0]
+	pxor	@y[3], @y[1]
+___
+	Mul_GF4_N_GF4	(@t[0], @t[1], @y[0], @y[1], @t[3],
+			 @x[2], @x[3], @y[2], @y[3], @t[2]);
+$code.=<<___;
+	pxor	@t[0], @x[0]
+	pxor	@t[0], @x[2]
+	pxor	@t[1], @x[1]
+	pxor	@t[1], @x[3]
+
+	movdqa	@x[4], @t[0]
+	movdqa	@x[5], @t[1]
+	pxor	@x[6], @t[0]
+	pxor	@x[7], @t[1]
+___
+	&Mul_GF4_N_GF4	(@t[0], @t[1], @y[0], @y[1], @t[3],
+			 @x[6], @x[7], @y[2], @y[3], @t[2]);
+$code.=<<___;
+	pxor	@y[2], @y[0]
+	pxor	@y[3], @y[1]
+___
+	&Mul_GF4  	(@x[4], @x[5], @y[0], @y[1], @t[3]);
+$code.=<<___;
+	pxor	@t[0], @x[4]
+	pxor	@t[0], @x[6]
+	pxor	@t[1], @x[5]
+	pxor	@t[1], @x[7]
+___
+}
+sub Inv_GF256 {
+#;********************************************************************
+#;* Inv_GF256: Input x0-x7 Output x0-x7 Temp t0-t3,s0-s3 (144)       *
+#;********************************************************************
+my @x=@_[0..7];
+my @t=@_[8..11];
+my @s=@_[12..15];
+# direct optimizations from hardware
+$code.=<<___;
+	movdqa	@x[4], @t[3]
+	movdqa	@x[5], @t[2]
+	movdqa	@x[1], @t[1]
+	movdqa	@x[7], @s[1]
+	movdqa	@x[0], @s[0]
+
+	pxor	@x[6], @t[3]
+	pxor	@x[7], @t[2]
+	pxor	@x[3], @t[1]
+	 movdqa	@t[3], @s[2]
+	pxor	@x[6], @s[1]
+	 movdqa	@t[2], @t[0]
+	pxor	@x[2], @s[0]
+	 movdqa	@t[3], @s[3]
+
+	por	@t[1], @t[2]
+	por	@s[0], @t[3]
+	pxor	@t[0], @s[3]
+	pand	@s[0], @s[2]
+	pxor	@t[1], @s[0]
+	pand	@t[1], @t[0]
+	pand	@s[0], @s[3]
+	movdqa	@x[3], @s[0]
+	pxor	@x[2], @s[0]
+	pand	@s[0], @s[1]
+	pxor	@s[1], @t[3]
+	pxor	@s[1], @t[2]
+	movdqa	@x[4], @s[1]
+	movdqa	@x[1], @s[0]
+	pxor	@x[5], @s[1]
+	pxor	@x[0], @s[0]
+	movdqa	@s[1], @t[1]
+	pand	@s[0], @s[1]
+	por	@s[0], @t[1]
+	pxor	@s[1], @t[0]
+	pxor	@s[3], @t[3]
+	pxor	@s[2], @t[2]
+	pxor	@s[3], @t[1]
+	movdqa	@x[7], @s[0]
+	pxor	@s[2], @t[0]
+	movdqa	@x[6], @s[1]
+	pxor	@s[2], @t[1]
+	movdqa	@x[5], @s[2]
+	pand	@x[3], @s[0]
+	movdqa	@x[4], @s[3]
+	pand	@x[2], @s[1]
+	pand	@x[1], @s[2]
+	por	@x[0], @s[3]
+	pxor	@s[0], @t[3]
+	pxor	@s[1], @t[2]
+	pxor	@s[2], @t[1]
+	pxor	@s[3], @t[0] 
+
+	#Inv_GF16 \t0, \t1, \t2, \t3, \s0, \s1, \s2, \s3
+
+	# new smaller inversion
+
+	movdqa	@t[3], @s[0]
+	pand	@t[1], @t[3]
+	pxor	@t[2], @s[0]
+
+	movdqa	@t[0], @s[2]
+	movdqa	@s[0], @s[3]
+	pxor	@t[3], @s[2]
+	pand	@s[2], @s[3]
+
+	movdqa	@t[1], @s[1]
+	pxor	@t[2], @s[3]
+	pxor	@t[0], @s[1]
+
+	pxor	@t[2], @t[3]
+
+	pand	@t[3], @s[1]
+
+	movdqa	@s[2], @t[2]
+	pxor	@t[0], @s[1]
+
+	pxor	@s[1], @t[2]
+	pxor	@s[1], @t[1]
+
+	pand	@t[0], @t[2]
+
+	pxor	@t[2], @s[2]
+	pxor	@t[2], @t[1]
+
+	pand	@s[3], @s[2]
+
+	pxor	@s[0], @s[2]
+___
+# output in s3, s2, s1, t1
+
+# Mul_GF16_2 \x0, \x1, \x2, \x3, \x4, \x5, \x6, \x7, \t2, \t3, \t0, \t1, \s0, \s1, \s2, \s3
+
+# Mul_GF16_2 \x0, \x1, \x2, \x3, \x4, \x5, \x6, \x7, \s3, \s2, \s1, \t1, \s0, \t0, \t2, \t3
+	&Mul_GF16_2(@x,@s[3,2,1],@t[1],@s[0],@t[0,2,3]);
+
+### output msb > [x3,x2,x1,x0,x7,x6,x5,x4] < lsb
+}
+
+# AES linear components
+
+sub ShiftRows {
+my @x=@_[0..7];
+my $mask=pop;
+$code.=<<___;
+	pxor	0x00($key),@x[0]
+	pxor	0x10($key),@x[1]
+	pshufb	$mask,@x[0]
+	pxor	0x20($key),@x[2]
+	pshufb	$mask,@x[1]
+	pxor	0x30($key),@x[3]
+	pshufb	$mask,@x[2]
+	pxor	0x40($key),@x[4]
+	pshufb	$mask,@x[3]
+	pxor	0x50($key),@x[5]
+	pshufb	$mask,@x[4]
+	pxor	0x60($key),@x[6]
+	pshufb	$mask,@x[5]
+	pxor	0x70($key),@x[7]
+	pshufb	$mask,@x[6]
+	lea	0x80($key),$key
+	pshufb	$mask,@x[7]
+___
+}
+
+sub MixColumns {
+# modified to emit output in order suitable for feeding back to aesenc[last]
+my @x=@_[0..7];
+my @t=@_[8..15];
+$code.=<<___;
+	pshufd	\$0x93, @x[0], @t[0]	# x0 <<< 32
+	pshufd	\$0x93, @x[1], @t[1]
+	 pxor	@t[0], @x[0]		# x0 ^ (x0 <<< 32)
+	pshufd	\$0x93, @x[2], @t[2]
+	 pxor	@t[1], @x[1]
+	pshufd	\$0x93, @x[3], @t[3]
+	 pxor	@t[2], @x[2]
+	pshufd	\$0x93, @x[4], @t[4]
+	 pxor	@t[3], @x[3]
+	pshufd	\$0x93, @x[5], @t[5]
+	 pxor	@t[4], @x[4]
+	pshufd	\$0x93, @x[6], @t[6]
+	 pxor	@t[5], @x[5]
+	pshufd	\$0x93, @x[7], @t[7]
+	 pxor	@t[6], @x[6]
+	 pxor	@t[7], @x[7]
+
+	pxor	@x[0], @t[1]
+	pxor	@x[7], @t[0]
+	pxor	@x[7], @t[1]
+	 pshufd	\$0x4E, @x[0], @x[0] 	# (x0 ^ (x0 <<< 32)) <<< 64)
+	pxor	@x[1], @t[2]
+	 pshufd	\$0x4E, @x[1], @x[1]
+	pxor	@x[4], @t[5]
+	 pxor	@t[0], @x[0]
+	pxor	@x[5], @t[6]
+	 pxor	@t[1], @x[1]
+	pxor	@x[3], @t[4]
+	 pshufd	\$0x4E, @x[4], @t[0]
+	pxor	@x[6], @t[7]
+	 pshufd	\$0x4E, @x[5], @t[1]
+	pxor	@x[2], @t[3]
+	 pshufd	\$0x4E, @x[3], @x[4]
+	pxor	@x[7], @t[3]
+	 pshufd	\$0x4E, @x[7], @x[5]
+	pxor	@x[7], @t[4]
+	 pshufd	\$0x4E, @x[6], @x[3]
+	pxor	@t[4], @t[0]
+	 pshufd	\$0x4E, @x[2], @x[6]
+	pxor	@t[5], @t[1]
+
+	pxor	@t[3], @x[4]
+	pxor	@t[7], @x[5]
+	pxor	@t[6], @x[3]
+	 movdqa	@t[0], @x[2]
+	pxor	@t[2], @x[6]
+	 movdqa	@t[1], @x[7]
+___
+}
+
+sub InvMixColumns {
+my @x=@_[0..7];
+my @t=@_[8..15];
+
+$code.=<<___;
+	# multiplication by 0x0e
+	pshufd	\$0x93, @x[7], @t[7]
+	movdqa	@x[2], @t[2]
+	pxor	@x[5], @x[7]		# 7 5
+	pxor	@x[5], @x[2]		# 2 5
+	pshufd	\$0x93, @x[0], @t[0]
+	movdqa	@x[5], @t[5]
+	pxor	@x[0], @x[5]		# 5 0		[1]
+	pxor	@x[1], @x[0]		# 0 1
+	pshufd	\$0x93, @x[1], @t[1]
+	pxor	@x[2], @x[1]		# 1 25
+	pxor	@x[6], @x[0]		# 01 6		[2]
+	pxor	@x[3], @x[1]		# 125 3		[4]
+	pshufd	\$0x93, @x[3], @t[3]
+	pxor	@x[0], @x[2]		# 25 016	[3]
+	pxor	@x[7], @x[3]		# 3 75
+	pxor	@x[6], @x[7]		# 75 6		[0]
+	pshufd	\$0x93, @x[6], @t[6]
+	movdqa	@x[4], @t[4]
+	pxor	@x[4], @x[6]		# 6 4
+	pxor	@x[3], @x[4]		# 4 375		[6]
+	pxor	@x[7], @x[3]		# 375 756=36
+	pxor	@t[5], @x[6]		# 64 5		[7]
+	pxor	@t[2], @x[3]		# 36 2
+	pxor	@t[4], @x[3]		# 362 4		[5]
+	pshufd	\$0x93, @t[5], @t[5]
+___
+					my @y = @x[7,5,0,2,1,3,4,6];
+$code.=<<___;
+	# multiplication by 0x0b
+	pxor	@y[0], @y[1]
+	pxor	@t[0], @y[0]
+	pxor	@t[1], @y[1]
+	pshufd	\$0x93, @t[2], @t[2]
+	pxor	@t[5], @y[0]
+	pxor	@t[6], @y[1]
+	pxor	@t[7], @y[0]
+	pshufd	\$0x93, @t[4], @t[4]
+	pxor	@t[6], @t[7]		# clobber t[7]
+	pxor	@y[0], @y[1]
+
+	pxor	@t[0], @y[3]
+	pshufd	\$0x93, @t[0], @t[0]
+	pxor	@t[1], @y[2]
+	pxor	@t[1], @y[4]
+	pxor	@t[2], @y[2]
+	pshufd	\$0x93, @t[1], @t[1]
+	pxor	@t[2], @y[3]
+	pxor	@t[2], @y[5]
+	pxor	@t[7], @y[2]
+	pshufd	\$0x93, @t[2], @t[2]
+	pxor	@t[3], @y[3]
+	pxor	@t[3], @y[6]
+	pxor	@t[3], @y[4]
+	pshufd	\$0x93, @t[3], @t[3]
+	pxor	@t[4], @y[7]
+	pxor	@t[4], @y[5]
+	pxor	@t[7], @y[7]
+	pxor	@t[5], @y[3]
+	pxor	@t[4], @y[4]
+	pxor	@t[5], @t[7]		# clobber t[7] even more
+
+	pxor	@t[7], @y[5]
+	pshufd	\$0x93, @t[4], @t[4]
+	pxor	@t[7], @y[6]
+	pxor	@t[7], @y[4]
+
+	pxor	@t[5], @t[7]
+	pshufd	\$0x93, @t[5], @t[5]
+	pxor	@t[6], @t[7]		# restore t[7]
+
+	# multiplication by 0x0d
+	pxor	@y[7], @y[4]
+	pxor	@t[4], @y[7]
+	pshufd	\$0x93, @t[6], @t[6]
+	pxor	@t[0], @y[2]
+	pxor	@t[5], @y[7]
+	pxor	@t[2], @y[2]
+	pshufd	\$0x93, @t[7], @t[7]
+
+	pxor	@y[1], @y[3]
+	pxor	@t[1], @y[1]
+	pxor	@t[0], @y[0]
+	pxor	@t[0], @y[3]
+	pxor	@t[5], @y[1]
+	pxor	@t[5], @y[0]
+	pxor	@t[7], @y[1]
+	pshufd	\$0x93, @t[0], @t[0]
+	pxor	@t[6], @y[0]
+	pxor	@y[1], @y[3]
+	pxor	@t[1], @y[4]
+	pshufd	\$0x93, @t[1], @t[1]
+
+	pxor	@t[7], @y[7]
+	pxor	@t[2], @y[4]
+	pxor	@t[2], @y[5]
+	pshufd	\$0x93, @t[2], @t[2]
+	pxor	@t[6], @y[2]
+	pxor	@t[3], @t[6]		# clobber t[6]
+	pxor	@y[7], @y[4]
+	pxor	@t[6], @y[3]
+
+	pxor	@t[6], @y[6]
+	pxor	@t[5], @y[5]
+	pxor	@t[4], @y[6]
+	pshufd	\$0x93, @t[4], @t[4]
+	pxor	@t[6], @y[5]
+	pxor	@t[7], @y[6]
+	pxor	@t[3], @t[6]		# restore t[6]
+
+	pshufd	\$0x93, @t[5], @t[5]
+	pshufd	\$0x93, @t[6], @t[6]
+	pshufd	\$0x93, @t[7], @t[7]
+	pshufd	\$0x93, @t[3], @t[3]
+
+	# multiplication by 0x09
+	pxor	@y[1], @y[4]
+	pxor	@y[1], @t[1]		# t[1]=y[1]
+	pxor	@t[5], @t[0]		# clobber t[0]
+	pxor	@t[5], @t[1]
+	pxor	@t[0], @y[3]
+	pxor	@y[0], @t[0]		# t[0]=y[0]
+	pxor	@t[6], @t[1]
+	pxor	@t[7], @t[6]		# clobber t[6]
+	pxor	@t[1], @y[4]
+	pxor	@t[4], @y[7]
+	pxor	@y[4], @t[4]		# t[4]=y[4]
+	pxor	@t[3], @y[6]
+	pxor	@y[3], @t[3]		# t[3]=y[3]
+	pxor	@t[2], @y[5]
+	pxor	@y[2], @t[2]		# t[2]=y[2]
+	pxor	@t[7], @t[3]
+	pxor	@y[5], @t[5]		# t[5]=y[5]
+	pxor	@t[6], @t[2]
+	pxor	@t[6], @t[5]
+	pxor	@y[6], @t[6]		# t[6]=y[6]
+	pxor	@y[7], @t[7]		# t[7]=y[7]
+
+	movdqa	@t[0],@XMM[0]
+	movdqa	@t[1],@XMM[1]
+	movdqa	@t[2],@XMM[2]
+	movdqa	@t[3],@XMM[3]
+	movdqa	@t[4],@XMM[4]
+	movdqa	@t[5],@XMM[5]
+	movdqa	@t[6],@XMM[6]
+	movdqa	@t[7],@XMM[7]
+___
+}
+
+sub aesenc {				# not used
+my @b=@_[0..7];
+my @t=@_[8..15];
+$code.=<<___;
+	movdqa	0x30($const),@t[0]	# .LSR
+___
+	&ShiftRows	(@b,@t[0]);
+	&Sbox		(@b,@t);
+	&MixColumns	(@b[0,1,4,6,3,7,2,5],@t);
+}
+
+sub aesenclast {			# not used
+my @b=@_[0..7];
+my @t=@_[8..15];
+$code.=<<___;
+	movdqa	0x40($const),@t[0]	# .LSRM0
+___
+	&ShiftRows	(@b,@t[0]);
+	&Sbox		(@b,@t);
+$code.=<<___
+	pxor	0x00($key),@b[0]
+	pxor	0x10($key),@b[1]
+	pxor	0x20($key),@b[4]
+	pxor	0x30($key),@b[6]
+	pxor	0x40($key),@b[3]
+	pxor	0x50($key),@b[7]
+	pxor	0x60($key),@b[2]
+	pxor	0x70($key),@b[5]
+___
+}
+
+sub swapmove {
+my ($a,$b,$n,$mask,$t)=@_;
+$code.=<<___;
+	movdqa	$b,$t
+	psrlq	\$$n,$b
+	pxor  	$a,$b
+	pand	$mask,$b
+	pxor	$b,$a
+	psllq	\$$n,$b
+	pxor	$t,$b
+___
+}
+sub swapmove2x {
+my ($a0,$b0,$a1,$b1,$n,$mask,$t0,$t1)=@_;
+$code.=<<___;
+	movdqa	$b0,$t0
+	psrlq	\$$n,$b0
+	 movdqa	$b1,$t1
+	 psrlq	\$$n,$b1
+	pxor  	$a0,$b0
+	 pxor  	$a1,$b1
+	pand	$mask,$b0
+	 pand	$mask,$b1
+	pxor	$b0,$a0
+	psllq	\$$n,$b0
+	 pxor	$b1,$a1
+	 psllq	\$$n,$b1
+	pxor	$t0,$b0
+	 pxor	$t1,$b1
+___
+}
+
+sub bitslice {
+my @x=reverse(@_[0..7]);
+my ($t0,$t1,$t2,$t3)=@_[8..11];
+$code.=<<___;
+	movdqa	0x00($const),$t0	# .LBS0
+	movdqa	0x10($const),$t1	# .LBS1
+___
+	&swapmove2x(@x[0,1,2,3],1,$t0,$t2,$t3);
+	&swapmove2x(@x[4,5,6,7],1,$t0,$t2,$t3);
+$code.=<<___;
+	movdqa	0x20($const),$t0	# .LBS2
+___
+	&swapmove2x(@x[0,2,1,3],2,$t1,$t2,$t3);
+	&swapmove2x(@x[4,6,5,7],2,$t1,$t2,$t3);
+
+	&swapmove2x(@x[0,4,1,5],4,$t0,$t2,$t3);
+	&swapmove2x(@x[2,6,3,7],4,$t0,$t2,$t3);
+}
+
+$code.=<<___;
+.text
+
+.extern	asm_AES_encrypt
+.extern	asm_AES_decrypt
+
+.type	_bsaes_encrypt8,\@abi-omnipotent
+.align	64
+_bsaes_encrypt8:
+	lea	.LBS0(%rip), $const	# constants table
+
+	movdqa	($key), @XMM[9]		# round 0 key
+	lea	0x10($key), $key
+	movdqa	0x60($const), @XMM[8]	# .LM0SR
+	pxor	@XMM[9], @XMM[0]	# xor with round0 key
+	pxor	@XMM[9], @XMM[1]
+	 pshufb	@XMM[8], @XMM[0]
+	pxor	@XMM[9], @XMM[2]
+	 pshufb	@XMM[8], @XMM[1]
+	pxor	@XMM[9], @XMM[3]
+	 pshufb	@XMM[8], @XMM[2]
+	pxor	@XMM[9], @XMM[4]
+	 pshufb	@XMM[8], @XMM[3]
+	pxor	@XMM[9], @XMM[5]
+	 pshufb	@XMM[8], @XMM[4]
+	pxor	@XMM[9], @XMM[6]
+	 pshufb	@XMM[8], @XMM[5]
+	pxor	@XMM[9], @XMM[7]
+	 pshufb	@XMM[8], @XMM[6]
+	 pshufb	@XMM[8], @XMM[7]
+_bsaes_encrypt8_bitslice:
+___
+	&bitslice	(@XMM[0..7, 8..11]);
+$code.=<<___;
+	dec	$rounds
+	jmp	.Lenc_sbox
+.align	16
+.Lenc_loop:
+___
+	&ShiftRows	(@XMM[0..7, 8]);
+$code.=".Lenc_sbox:\n";
+	&Sbox		(@XMM[0..7, 8..15]);
+$code.=<<___;
+	dec	$rounds
+	jl	.Lenc_done
+___
+	&MixColumns	(@XMM[0,1,4,6,3,7,2,5, 8..15]);
+$code.=<<___;
+	movdqa	0x30($const), @XMM[8]	# .LSR
+	jnz	.Lenc_loop
+	movdqa	0x40($const), @XMM[8]	# .LSRM0
+	jmp	.Lenc_loop
+.align	16
+.Lenc_done:
+___
+	# output in lsb > [t0, t1, t4, t6, t3, t7, t2, t5] < msb
+	&bitslice	(@XMM[0,1,4,6,3,7,2,5, 8..11]);
+$code.=<<___;
+	movdqa	($key), @XMM[8]		# last round key
+	pxor	@XMM[8], @XMM[4]
+	pxor	@XMM[8], @XMM[6]
+	pxor	@XMM[8], @XMM[3]
+	pxor	@XMM[8], @XMM[7]
+	pxor	@XMM[8], @XMM[2]
+	pxor	@XMM[8], @XMM[5]
+	pxor	@XMM[8], @XMM[0]
+	pxor	@XMM[8], @XMM[1]
+	ret
+.size	_bsaes_encrypt8,.-_bsaes_encrypt8
+
+.type	_bsaes_decrypt8,\@abi-omnipotent
+.align	64
+_bsaes_decrypt8:
+	lea	.LBS0(%rip), $const	# constants table
+
+	movdqa	($key), @XMM[9]		# round 0 key
+	lea	0x10($key), $key
+	movdqa	-0x30($const), @XMM[8]	# .LM0ISR
+	pxor	@XMM[9], @XMM[0]	# xor with round0 key
+	pxor	@XMM[9], @XMM[1]
+	 pshufb	@XMM[8], @XMM[0]
+	pxor	@XMM[9], @XMM[2]
+	 pshufb	@XMM[8], @XMM[1]
+	pxor	@XMM[9], @XMM[3]
+	 pshufb	@XMM[8], @XMM[2]
+	pxor	@XMM[9], @XMM[4]
+	 pshufb	@XMM[8], @XMM[3]
+	pxor	@XMM[9], @XMM[5]
+	 pshufb	@XMM[8], @XMM[4]
+	pxor	@XMM[9], @XMM[6]
+	 pshufb	@XMM[8], @XMM[5]
+	pxor	@XMM[9], @XMM[7]
+	 pshufb	@XMM[8], @XMM[6]
+	 pshufb	@XMM[8], @XMM[7]
+___
+	&bitslice	(@XMM[0..7, 8..11]);
+$code.=<<___;
+	dec	$rounds
+	jmp	.Ldec_sbox
+.align	16
+.Ldec_loop:
+___
+	&ShiftRows	(@XMM[0..7, 8]);
+$code.=".Ldec_sbox:\n";
+	&InvSbox	(@XMM[0..7, 8..15]);
+$code.=<<___;
+	dec	$rounds
+	jl	.Ldec_done
+___
+	&InvMixColumns	(@XMM[0,1,6,4,2,7,3,5, 8..15]);
+$code.=<<___;
+	movdqa	-0x10($const), @XMM[8]	# .LISR
+	jnz	.Ldec_loop
+	movdqa	-0x20($const), @XMM[8]	# .LISRM0
+	jmp	.Ldec_loop
+.align	16
+.Ldec_done:
+___
+	&bitslice	(@XMM[0,1,6,4,2,7,3,5, 8..11]);
+$code.=<<___;
+	movdqa	($key), @XMM[8]		# last round key
+	pxor	@XMM[8], @XMM[6]
+	pxor	@XMM[8], @XMM[4]
+	pxor	@XMM[8], @XMM[2]
+	pxor	@XMM[8], @XMM[7]
+	pxor	@XMM[8], @XMM[3]
+	pxor	@XMM[8], @XMM[5]
+	pxor	@XMM[8], @XMM[0]
+	pxor	@XMM[8], @XMM[1]
+	ret
+.size	_bsaes_decrypt8,.-_bsaes_decrypt8
+___
+}
+{
+my ($out,$inp,$rounds,$const)=("%rax","%rcx","%r10d","%r11");
+
+sub bitslice_key {
+my @x=reverse(@_[0..7]);
+my ($bs0,$bs1,$bs2,$t2,$t3)=@_[8..12];
+
+	&swapmove	(@x[0,1],1,$bs0,$t2,$t3);
+$code.=<<___;
+	#&swapmove(@x[2,3],1,$t0,$t2,$t3);
+	movdqa	@x[0], @x[2]
+	movdqa	@x[1], @x[3]
+___
+	#&swapmove2x(@x[4,5,6,7],1,$t0,$t2,$t3);
+
+	&swapmove2x	(@x[0,2,1,3],2,$bs1,$t2,$t3);
+$code.=<<___;
+	#&swapmove2x(@x[4,6,5,7],2,$t1,$t2,$t3);
+	movdqa	@x[0], @x[4]
+	movdqa	@x[2], @x[6]
+	movdqa	@x[1], @x[5]
+	movdqa	@x[3], @x[7]
+___
+	&swapmove2x	(@x[0,4,1,5],4,$bs2,$t2,$t3);
+	&swapmove2x	(@x[2,6,3,7],4,$bs2,$t2,$t3);
+}
+
+$code.=<<___;
+.type	_bsaes_key_convert,\@abi-omnipotent
+.align	16
+_bsaes_key_convert:
+	lea	.LBS1(%rip), $const
+	movdqu	($inp), %xmm7		# load round 0 key
+	movdqa	-0x10($const), %xmm8	# .LBS0
+	movdqa	0x00($const), %xmm9	# .LBS1
+	movdqa	0x10($const), %xmm10	# .LBS2
+	movdqa	0x40($const), %xmm13	# .LM0
+	movdqa	0x60($const), %xmm14	# .LNOT
+
+	movdqu	0x10($inp), %xmm6	# load round 1 key
+	lea	0x10($inp), $inp
+	movdqa	%xmm7, ($out)		# save round 0 key
+	lea	0x10($out), $out
+	dec	$rounds
+	jmp	.Lkey_loop
+.align	16
+.Lkey_loop:
+	pshufb	%xmm13, %xmm6		# .LM0
+	movdqa	%xmm6, %xmm7
+___
+	&bitslice_key	(map("%xmm$_",(0..7, 8..12)));
+$code.=<<___;
+	pxor	%xmm14, %xmm5		# "pnot"
+	pxor	%xmm14, %xmm6
+	pxor	%xmm14, %xmm0
+	pxor	%xmm14, %xmm1
+	lea	0x10($inp), $inp
+	movdqa	%xmm0, 0x00($out)	# write bit-sliced round key
+	movdqa	%xmm1, 0x10($out)
+	movdqa	%xmm2, 0x20($out)
+	movdqa	%xmm3, 0x30($out)
+	movdqa	%xmm4, 0x40($out)
+	movdqa	%xmm5, 0x50($out)
+	movdqa	%xmm6, 0x60($out)
+	movdqa	%xmm7, 0x70($out)
+	lea	0x80($out),$out
+	movdqu	($inp), %xmm6		# load next round key
+	dec	$rounds
+	jnz	.Lkey_loop
+
+	movdqa	0x70($const), %xmm7	# .L63
+	#movdqa	%xmm6, ($out)		# don't save last round key
+	ret
+.size	_bsaes_key_convert,.-_bsaes_key_convert
+___
+}
+
+if (0 && !$win64) {	# following four functions are unsupported interface
+			# used for benchmarking...
+$code.=<<___;
+.globl	bsaes_enc_key_convert
+.type	bsaes_enc_key_convert,\@function,2
+.align	16
+bsaes_enc_key_convert:
+	mov	240($inp),%r10d		# pass rounds
+	mov	$inp,%rcx		# pass key
+	mov	$out,%rax		# pass key schedule
+	call	_bsaes_key_convert
+	pxor	%xmm6,%xmm7		# fix up last round key
+	movdqa	%xmm7,(%rax)		# save last round key
+	ret
+.size	bsaes_enc_key_convert,.-bsaes_enc_key_convert
+
+.globl	bsaes_encrypt_128
+.type	bsaes_encrypt_128,\@function,4
+.align	16
+bsaes_encrypt_128:
+.Lenc128_loop:
+	movdqu	0x00($inp), @XMM[0]	# load input
+	movdqu	0x10($inp), @XMM[1]
+	movdqu	0x20($inp), @XMM[2]
+	movdqu	0x30($inp), @XMM[3]
+	movdqu	0x40($inp), @XMM[4]
+	movdqu	0x50($inp), @XMM[5]
+	movdqu	0x60($inp), @XMM[6]
+	movdqu	0x70($inp), @XMM[7]
+	mov	$key, %rax		# pass the $key
+	lea	0x80($inp), $inp
+	mov	\$10,%r10d
+
+	call	_bsaes_encrypt8
+
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	movdqu	@XMM[4], 0x20($out)
+	movdqu	@XMM[6], 0x30($out)
+	movdqu	@XMM[3], 0x40($out)
+	movdqu	@XMM[7], 0x50($out)
+	movdqu	@XMM[2], 0x60($out)
+	movdqu	@XMM[5], 0x70($out)
+	lea	0x80($out), $out
+	sub	\$0x80,$len
+	ja	.Lenc128_loop
+	ret
+.size	bsaes_encrypt_128,.-bsaes_encrypt_128
+
+.globl	bsaes_dec_key_convert
+.type	bsaes_dec_key_convert,\@function,2
+.align	16
+bsaes_dec_key_convert:
+	mov	240($inp),%r10d		# pass rounds
+	mov	$inp,%rcx		# pass key
+	mov	$out,%rax		# pass key schedule
+	call	_bsaes_key_convert
+	pxor	($out),%xmm7		# fix up round 0 key
+	movdqa	%xmm6,(%rax)		# save last round key
+	movdqa	%xmm7,($out)
+	ret
+.size	bsaes_dec_key_convert,.-bsaes_dec_key_convert
+
+.globl	bsaes_decrypt_128
+.type	bsaes_decrypt_128,\@function,4
+.align	16
+bsaes_decrypt_128:
+.Ldec128_loop:
+	movdqu	0x00($inp), @XMM[0]	# load input
+	movdqu	0x10($inp), @XMM[1]
+	movdqu	0x20($inp), @XMM[2]
+	movdqu	0x30($inp), @XMM[3]
+	movdqu	0x40($inp), @XMM[4]
+	movdqu	0x50($inp), @XMM[5]
+	movdqu	0x60($inp), @XMM[6]
+	movdqu	0x70($inp), @XMM[7]
+	mov	$key, %rax		# pass the $key
+	lea	0x80($inp), $inp
+	mov	\$10,%r10d
+
+	call	_bsaes_decrypt8
+
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	movdqu	@XMM[6], 0x20($out)
+	movdqu	@XMM[4], 0x30($out)
+	movdqu	@XMM[2], 0x40($out)
+	movdqu	@XMM[7], 0x50($out)
+	movdqu	@XMM[3], 0x60($out)
+	movdqu	@XMM[5], 0x70($out)
+	lea	0x80($out), $out
+	sub	\$0x80,$len
+	ja	.Ldec128_loop
+	ret
+.size	bsaes_decrypt_128,.-bsaes_decrypt_128
+___
+}
+{
+######################################################################
+#
+# OpenSSL interface
+#
+my ($arg1,$arg2,$arg3,$arg4,$arg5,$arg6)=$win64	? ("%rcx","%rdx","%r8","%r9","%r10","%r11d")
+						: ("%rdi","%rsi","%rdx","%rcx","%r8","%r9d");
+my ($inp,$out,$len,$key)=("%r12","%r13","%r14","%r15");
+
+if ($ecb) {
+$code.=<<___;
+.globl	bsaes_ecb_encrypt_blocks
+.type	bsaes_ecb_encrypt_blocks,\@abi-omnipotent
+.align	16
+bsaes_ecb_encrypt_blocks:
+	mov	%rsp, %rax
+.Lecb_enc_prologue:
+	push	%rbp
+	push	%rbx
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	lea	-0x48(%rsp),%rsp
+___
+$code.=<<___ if ($win64);
+	lea	-0xa0(%rsp), %rsp
+	movaps	%xmm6, 0x40(%rsp)
+	movaps	%xmm7, 0x50(%rsp)
+	movaps	%xmm8, 0x60(%rsp)
+	movaps	%xmm9, 0x70(%rsp)
+	movaps	%xmm10, 0x80(%rsp)
+	movaps	%xmm11, 0x90(%rsp)
+	movaps	%xmm12, 0xa0(%rsp)
+	movaps	%xmm13, 0xb0(%rsp)
+	movaps	%xmm14, 0xc0(%rsp)
+	movaps	%xmm15, 0xd0(%rsp)
+.Lecb_enc_body:
+___
+$code.=<<___;
+	mov	%rsp,%rbp		# backup %rsp
+	mov	240($arg4),%eax		# rounds
+	mov	$arg1,$inp		# backup arguments
+	mov	$arg2,$out
+	mov	$arg3,$len
+	mov	$arg4,$key
+	cmp	\$8,$arg3
+	jb	.Lecb_enc_short
+
+	mov	%eax,%ebx		# backup rounds
+	shl	\$7,%rax		# 128 bytes per inner round key
+	sub	\$`128-32`,%rax		# size of bit-sliced key schedule
+	sub	%rax,%rsp
+	mov	%rsp,%rax		# pass key schedule
+	mov	$key,%rcx		# pass key
+	mov	%ebx,%r10d		# pass rounds
+	call	_bsaes_key_convert
+	pxor	%xmm6,%xmm7		# fix up last round key
+	movdqa	%xmm7,(%rax)		# save last round key
+
+	sub	\$8,$len
+.Lecb_enc_loop:
+	movdqu	0x00($inp), @XMM[0]	# load input
+	movdqu	0x10($inp), @XMM[1]
+	movdqu	0x20($inp), @XMM[2]
+	movdqu	0x30($inp), @XMM[3]
+	movdqu	0x40($inp), @XMM[4]
+	movdqu	0x50($inp), @XMM[5]
+	mov	%rsp, %rax		# pass key schedule
+	movdqu	0x60($inp), @XMM[6]
+	mov	%ebx,%r10d		# pass rounds
+	movdqu	0x70($inp), @XMM[7]
+	lea	0x80($inp), $inp
+
+	call	_bsaes_encrypt8
+
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	movdqu	@XMM[4], 0x20($out)
+	movdqu	@XMM[6], 0x30($out)
+	movdqu	@XMM[3], 0x40($out)
+	movdqu	@XMM[7], 0x50($out)
+	movdqu	@XMM[2], 0x60($out)
+	movdqu	@XMM[5], 0x70($out)
+	lea	0x80($out), $out
+	sub	\$8,$len
+	jnc	.Lecb_enc_loop
+
+	add	\$8,$len
+	jz	.Lecb_enc_done
+
+	movdqu	0x00($inp), @XMM[0]	# load input
+	mov	%rsp, %rax		# pass key schedule
+	mov	%ebx,%r10d		# pass rounds
+	cmp	\$2,$len
+	jb	.Lecb_enc_one
+	movdqu	0x10($inp), @XMM[1]
+	je	.Lecb_enc_two
+	movdqu	0x20($inp), @XMM[2]
+	cmp	\$4,$len
+	jb	.Lecb_enc_three
+	movdqu	0x30($inp), @XMM[3]
+	je	.Lecb_enc_four
+	movdqu	0x40($inp), @XMM[4]
+	cmp	\$6,$len
+	jb	.Lecb_enc_five
+	movdqu	0x50($i