X-Git-Url: https://git.openssl.org/gitweb/?p=openssl.git;a=blobdiff_plain;f=crypto%2Faes%2Fasm%2Fbsaes-x86_64.pl;h=c9c6312fa74a15ab9c3868dcc949dd4e9ff16133;hp=34ea2f818d2b004cee5577683cf50f2700aec437;hb=f9ef874a212f936b6503a8d149f5eb26e2b7af2b;hpb=5a326467dcec154dae3664ed8d1a742153c9248c

diff --git a/crypto/aes/asm/bsaes-x86_64.pl b/crypto/aes/asm/bsaes-x86_64.pl
index 34ea2f818d..c9c6312fa7 100644
--- a/crypto/aes/asm/bsaes-x86_64.pl
+++ b/crypto/aes/asm/bsaes-x86_64.pl
@@ -65,12 +65,12 @@
 # function is:
 #
 # 		conversion	conversion/8x block
-# Core 2	410		0.37
-# Nehalem	310		0.35
-# Atom		570		0.26
+# Core 2	240		0.22
+# Nehalem	180		0.20
+# Atom		430		0.19
 #
 # The ratio values mean that 128-byte blocks will be processed
-# 21-27% slower, 256-byte blocks - 12-16%, 384-byte blocks - 8-11%,
+# 16-18% slower, 256-byte blocks - 9-10%, 384-byte blocks - 6-7%,
 # 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,
@@ -78,6 +78,20 @@
 # 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
+# Atom		20.9
+#
+# 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;
@@ -95,11 +109,12 @@ 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 {
+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];
@@ -117,9 +132,9 @@ my @b=@_[0..7];
 $code.=<<___;
 	pxor	@b[6], @b[5]
 	pxor	@b[1], @b[2]
-	pxor 	@b[0], @b[5]
+	pxor	@b[0], @b[3]
 	pxor	@b[2], @b[6]
-	pxor 	@b[0], @b[3]
+	pxor 	@b[0], @b[5]
 
 	pxor	@b[3], @b[6]
 	pxor	@b[7], @b[3]
@@ -154,6 +169,57 @@ $code.=<<___;
 ___
 }
 
+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) *
@@ -361,7 +427,7 @@ ___
 
 # AES linear components
 
-sub shiftrows {
+sub ShiftRows {
 my @x=@_[0..7];
 my $mask=pop;
 $code.=<<___;
@@ -385,7 +451,7 @@ $code.=<<___;
 ___
 }
 
-sub mixcolumns {
+sub MixColumns {
 # modified to emit output in order suitable for feeding back to aesenc[last]
 my @x=@_[0..7];
 my @t=@_[8..15];
@@ -440,15 +506,169 @@ $code.=<<___;
 ___
 }
 
+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);
+	&ShiftRows	(@b,@t[0]);
+	&Sbox		(@b,@t);
+	&MixColumns	(@b[0,1,4,6,3,7,2,5],@t);
 }
 
 sub aesenclast {			# not used
@@ -457,8 +677,8 @@ my @t=@_[8..15];
 $code.=<<___;
 	movdqa	0x40($const),@t[0]	# .LSRM0
 ___
-	&shiftrows	(@b,@t[0]);
-	&sbox		(@b,@t);
+	&ShiftRows	(@b,@t[0]);
+	&Sbox		(@b,@t);
 $code.=<<___
 	pxor	0x00($key),@b[0]
 	pxor	0x10($key),@b[1]
@@ -525,7 +745,8 @@ ___
 $code.=<<___;
 .text
 
-.extern	AES_encrypt
+.extern	asm_AES_encrypt
+.extern	asm_AES_decrypt
 
 .type	_bsaes_encrypt8,\@abi-omnipotent
 .align	64
@@ -534,7 +755,7 @@ _bsaes_encrypt8:
 
 	movdqa	($key), @XMM[9]		# round 0 key
 	lea	0x10($key), $key
-	movdqa	0x60($const), @XMM[8]	# .LM0SR
+	movdqa	0x50($const), @XMM[8]	# .LM0SR
 	pxor	@XMM[9], @XMM[0]	# xor with round0 key
 	pxor	@XMM[9], @XMM[1]
 	 pshufb	@XMM[8], @XMM[0]
@@ -560,14 +781,14 @@ $code.=<<___;
 .align	16
 .Lenc_loop:
 ___
-	&shiftrows	(@XMM[0..7, 8]);
+	&ShiftRows	(@XMM[0..7, 8]);
 $code.=".Lenc_sbox:\n";
-	&sbox		(@XMM[0..7, 8..15]);
+	&Sbox		(@XMM[0..7, 8..15]);
 $code.=<<___;
 	dec	$rounds
 	jl	.Lenc_done
 ___
-	&mixcolumns	(@XMM[0,1,4,6,3,7,2,5, 8..15]);
+	&MixColumns	(@XMM[0,1,4,6,3,7,2,5, 8..15]);
 $code.=<<___;
 	movdqa	0x30($const), @XMM[8]	# .LSR
 	jnz	.Lenc_loop
@@ -580,16 +801,78 @@ ___
 	&bitslice	(@XMM[0,1,4,6,3,7,2,5, 8..11]);
 $code.=<<___;
 	movdqa	($key), @XMM[8]		# last round key
-	pxor	@XMM[8], @XMM[0]
-	pxor	@XMM[8], @XMM[1]
 	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
 ___
 }
 {
@@ -620,56 +903,92 @@ ___
 }
 
 $code.=<<___;
-.type	_bsaes_enc_key_convert,\@abi-omnipotent
+.type	_bsaes_key_convert,\@abi-omnipotent
 .align	16
-_bsaes_enc_key_convert:
-	lea	.LBS1(%rip), $const
+_bsaes_key_convert:
+	lea	.Lmasks(%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), $inp
+	movdqa	0x00($const), %xmm0	# 0x01...
+	movdqa	0x10($const), %xmm1	# 0x02...
+	movdqa	0x20($const), %xmm2	# 0x04...
+	movdqa	0x30($const), %xmm3	# 0x08...
+	movdqa	0x40($const), %xmm4	# .LM0
+	pcmpeqd	%xmm5, %xmm5		# .LNOT
+
+	movdqu	($inp), %xmm6		# load round 1 key
 	movdqa	%xmm7, ($out)		# save round 0 key
 	lea	0x10($out), $out
 	dec	$rounds
 	jmp	.Lkey_loop
 .align	16
 .Lkey_loop:
-	pshufb	%xmm13, %xmm6
-	movdqa	%xmm6, %xmm7
-___
-	&bitslice_key	(map("%xmm$_",(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($out)
+	pshufb	%xmm4, %xmm6		# .LM0
+
+	movdqa	%xmm0,	%xmm8
+	movdqa	%xmm1,	%xmm9
+
+	pand	%xmm6,	%xmm8
+	pand	%xmm6,	%xmm9
+	movdqa	%xmm2,	%xmm10
+	pcmpeqb	%xmm0,	%xmm8
+	psllq	\$4,	%xmm0		# 0x10...
+	movdqa	%xmm3,	%xmm11
+	pcmpeqb	%xmm1,	%xmm9
+	psllq	\$4,	%xmm1		# 0x20...
+
+	pand	%xmm6,	%xmm10
+	pand	%xmm6,	%xmm11
+	movdqa	%xmm0,	%xmm12
+	pcmpeqb	%xmm2,	%xmm10
+	psllq	\$4,	%xmm2		# 0x40...
+	movdqa	%xmm1,	%xmm13
+	pcmpeqb	%xmm3,	%xmm11
+	psllq	\$4,	%xmm3		# 0x80...
+
+	movdqa	%xmm2,	%xmm14
+	movdqa	%xmm3,	%xmm15
+	 pxor	%xmm5,	%xmm8		# "pnot"
+	 pxor	%xmm5,	%xmm9
+
+	pand	%xmm6,	%xmm12
+	pand	%xmm6,	%xmm13
+	 movdqa	%xmm8, 0x00($out)	# write bit-sliced round key
+	pcmpeqb	%xmm0,	%xmm12
+	psrlq	\$4,	%xmm0		# 0x01...
+	 movdqa	%xmm9, 0x10($out)
+	pcmpeqb	%xmm1,	%xmm13
+	psrlq	\$4,	%xmm1		# 0x02...
+	 lea	0x10($inp), $inp
+
+	pand	%xmm6,	%xmm14
+	pand	%xmm6,	%xmm15
+	 movdqa	%xmm10, 0x20($out)
+	pcmpeqb	%xmm2,	%xmm14
+	psrlq	\$4,	%xmm2		# 0x04...
+	 movdqa	%xmm11, 0x30($out)
+	pcmpeqb	%xmm3,	%xmm15
+	psrlq	\$4,	%xmm3		# 0x08...
+	 movdqu	($inp), %xmm6		# load next round key
+
+	pxor	%xmm5, %xmm13		# "pnot"
+	pxor	%xmm5, %xmm14
+	movdqa	%xmm12, 0x40($out)
+	movdqa	%xmm13, 0x50($out)
+	movdqa	%xmm14, 0x60($out)
+	movdqa	%xmm15, 0x70($out)
 	lea	0x80($out),$out
-	movdqu	($inp), %xmm6		# load next round key
 	dec	$rounds
 	jnz	.Lkey_loop
 
-	pxor	0x70($const), %xmm6	# .L63
-	movdqa	%xmm6, ($out)		# save last round key
+	movdqa	0x50($const), %xmm7	# .L63
+	#movdqa	%xmm6, ($out)		# don't save last round key
 	ret
-.size	_bsaes_enc_key_convert,.-_bsaes_enc_key_convert
+.size	_bsaes_key_convert,.-_bsaes_key_convert
 ___
 }
 
-if (1 && !$win64) {	# following two functions are unsupported interface
+if (0 && !$win64) {	# following four functions are unsupported interface
 			# used for benchmarking...
 $code.=<<___;
 .globl	bsaes_enc_key_convert
@@ -679,7 +998,9 @@ bsaes_enc_key_convert:
 	mov	240($inp),%r10d		# pass rounds
 	mov	$inp,%rcx		# pass key
 	mov	$out,%rax		# pass key schedule
-	call	_bsaes_enc_key_convert
+	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
 
@@ -715,6 +1036,53 @@ bsaes_encrypt_128:
 	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
 ___
 }
 {
@@ -722,15 +1090,18 @@ ___
 #
 # OpenSSL interface
 #
-my ($arg1,$arg2,$arg3,$arg4,$arg5) = $win64	? ("%rcx","%rdx","%r8","%r9","%r10")
-						: ("%rdi","%rsi","%rdx","%rcx","%r8");
+my ($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
@@ -770,7 +1141,9 @@ $code.=<<___;
 	mov	%rsp,%rax		# pass key schedule
 	mov	$key,%rcx		# pass key
 	mov	%ebx,%r10d		# pass rounds
-	call	_bsaes_enc_key_convert
+	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:
@@ -880,7 +1253,7 @@ $code.=<<___;
 	lea	($inp), $arg1
 	lea	($out), $arg2
 	lea	($key), $arg3
-	call	AES_encrypt
+	call	asm_AES_encrypt
 	lea	16($inp), $inp
 	lea	16($out), $out
 	dec	$len
@@ -917,26 +1290,28 @@ $code.=<<___;
 	mov	0x58(%rsp), %r13
 	mov	0x60(%rsp), %r12
 	mov	0x68(%rsp), %rbx
-	mov	0x70(%rsp), %rbp
+	mov	0x70(%rsp), %rax
 	lea	0x78(%rsp), %rsp
+	mov	%rax, %rbp
 .Lecb_enc_epilogue:
 	ret
 .size	bsaes_ecb_encrypt_blocks,.-bsaes_ecb_encrypt_blocks
 
-.globl	bsaes_ctr32_encrypt_blocks
-.type	bsaes_ctr32_encrypt_blocks,\@abi-omnipotent
+.globl	bsaes_ecb_decrypt_blocks
+.type	bsaes_ecb_decrypt_blocks,\@abi-omnipotent
 .align	16
-bsaes_ctr32_encrypt_blocks:
+bsaes_ecb_decrypt_blocks:
+	mov	%rsp, %rax
+.Lecb_dec_prologue:
 	push	%rbp
 	push	%rbx
 	push	%r12
 	push	%r13
 	push	%r14
 	push	%r15
-	lea	-0x48(%rsp), %rsp
+	lea	-0x48(%rsp),%rsp
 ___
 $code.=<<___ if ($win64);
-	mov	0xa0(%rsp),$arg5	# pull ivp
 	lea	-0xa0(%rsp), %rsp
 	movaps	%xmm6, 0x40(%rsp)
 	movaps	%xmm7, 0x50(%rsp)
@@ -948,179 +1323,1470 @@ $code.=<<___ if ($win64);
 	movaps	%xmm13, 0xb0(%rsp)
 	movaps	%xmm14, 0xc0(%rsp)
 	movaps	%xmm15, 0xd0(%rsp)
-.Lctr_enc_body:
+.Lecb_dec_body:
 ___
 $code.=<<___;
-	mov	%rsp, %rbp		# backup %rsp
-	movdqu	($arg5), %xmm0		# load counter
-	mov	240($arg4), %eax	# rounds
-	mov	$arg1, $inp		# backup arguments
-	mov	$arg2, $out
-	mov	$arg3, $len
-	mov	$arg4, $key
-	movdqa	%xmm0, 0x20(%rbp)	# copy counter
-	cmp	\$8, $arg3
-	jb	.Lctr_enc_short
+	mov	%rsp,%rbp		# backup %rsp
+	mov	240($arg4),%eax		# rounds
+	mov	$arg1,$inp		# backup arguments
+	mov	$arg2,$out
+	mov	$arg3,$len
+	mov	$arg4,$key
+	cmp	\$8,$arg3
+	jb	.Lecb_dec_short
 
-	mov	%eax, %ebx		# rounds
-	shl	\$7, %rax		# 128 bytes per inner round key
-	sub	\$`128-32`, %rax	# size of bit-sliced key schedule
-	sub	%rax, %rsp
+	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	(%rsp),%xmm7		# fix up 0 round key
+	movdqa	%xmm6,(%rax)		# save last round key
+	movdqa	%xmm7,(%rsp)
 
+	sub	\$8,$len
+.Lecb_dec_loop:
+	movdqu	0x00($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
-	mov	$key, %rcx		# pass key
-	mov	%ebx, %r10d		# pass rounds
-	call	_bsaes_enc_key_convert
-
-	movdqa	(%rsp), @XMM[9]		# load round0 key
-	lea	.LADD1(%rip), %r11
-	movdqa	0x20(%rbp), @XMM[0]	# counter copy
-	movdqa	-0x20(%r11), @XMM[8]	# .LSWPUP
-	pshufb	@XMM[8], @XMM[9]	# byte swap upper part
-	pshufb	@XMM[8], @XMM[0]
-	movdqa	@XMM[9], (%rsp)		# save adjusted round0 key
-	jmp	.Lctr_enc_loop
-.align	16
-.Lctr_enc_loop:
-	movdqa	@XMM[0], 0x20(%rbp)	# save counter
-	movdqa	@XMM[0], @XMM[1]	# prepare 8 counter values
-	movdqa	@XMM[0], @XMM[2]
-	paddd	0x00(%r11), @XMM[1]	# .LADD1
-	movdqa	@XMM[0], @XMM[3]
-	paddd	0x10(%r11), @XMM[2]	# .LADD2
-	movdqa	@XMM[0], @XMM[4]
-	paddd	0x20(%r11), @XMM[3]	# .LADD3
-	movdqa	@XMM[0], @XMM[5]
-	paddd	0x30(%r11), @XMM[4]	# .LADD4
-	movdqa	@XMM[0], @XMM[6]
-	paddd	0x40(%r11), @XMM[5]	# .LADD5
-	movdqa	@XMM[0], @XMM[7]
-	paddd	0x50(%r11), @XMM[6]	# .LADD6
-	paddd	0x60(%r11), @XMM[7]	# .LADD7
-
-	# Borrow prologue from _bsaes_encrypt8 to use the opportunity
-	# to flip byte order in 32-bit counter
-	movdqa	(%rsp), @XMM[9]		# round 0 key
-	lea	0x10(%rsp), %rax	# pass key schedule
-	movdqa	-0x10(%r11), @XMM[8]	# .LSWPUPM0SR
-	pxor	@XMM[9], @XMM[0]	# xor with round0 key
-	pxor	@XMM[9], @XMM[1]
-	 pshufb	@XMM[8], @XMM[0]
-	pxor	@XMM[9], @XMM[2]
-	 pshufb	@XMM[8], @XMM[1]
-	pxor	@XMM[9], @XMM[3]
-	 pshufb	@XMM[8], @XMM[2]
-	pxor	@XMM[9], @XMM[4]
-	 pshufb	@XMM[8], @XMM[3]
-	pxor	@XMM[9], @XMM[5]
-	 pshufb	@XMM[8], @XMM[4]
-	pxor	@XMM[9], @XMM[6]
-	 pshufb	@XMM[8], @XMM[5]
-	pxor	@XMM[9], @XMM[7]
-	 pshufb	@XMM[8], @XMM[6]
-	lea	.LBS0(%rip), %r11	# constants table
-	 pshufb	@XMM[8], @XMM[7]
+	movdqu	0x60($inp), @XMM[6]
 	mov	%ebx,%r10d		# pass rounds
+	movdqu	0x70($inp), @XMM[7]
+	lea	0x80($inp), $inp
 
-	call	_bsaes_encrypt8_bitslice
+	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), $out
 	sub	\$8,$len
-	jc	.Lctr_enc_loop_done
+	jnc	.Lecb_dec_loop
 
-	movdqu	0x00($inp), @XMM[8]	# load input
-	movdqu	0x10($inp), @XMM[9]
-	movdqu	0x20($inp), @XMM[10]
-	movdqu	0x30($inp), @XMM[11]
-	movdqu	0x40($inp), @XMM[12]
-	movdqu	0x50($inp), @XMM[13]
-	movdqu	0x60($inp), @XMM[14]
-	movdqu	0x70($inp), @XMM[15]
-	lea	0x80($inp),$inp
-	pxor	@XMM[0], @XMM[8]
-	movdqa	0x20(%rbp), @XMM[0]	# load counter
-	pxor	@XMM[9], @XMM[1]
-	movdqu	@XMM[8], 0x00($out)	# write output
-	pxor	@XMM[10], @XMM[4]
+	add	\$8,$len
+	jz	.Lecb_dec_done
+
+	movdqu	0x00($inp), @XMM[0]	# load input
+	mov	%rsp, %rax		# pass key schedule
+	mov	%ebx,%r10d		# pass rounds
+	cmp	\$2,$len
+	jb	.Lecb_dec_one
+	movdqu	0x10($inp), @XMM[1]
+	je	.Lecb_dec_two
+	movdqu	0x20($inp), @XMM[2]
+	cmp	\$4,$len
+	jb	.Lecb_dec_three
+	movdqu	0x30($inp), @XMM[3]
+	je	.Lecb_dec_four
+	movdqu	0x40($inp), @XMM[4]
+	cmp	\$6,$len
+	jb	.Lecb_dec_five
+	movdqu	0x50($inp), @XMM[5]
+	je	.Lecb_dec_six
+	movdqu	0x60($inp), @XMM[6]
+	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($out)
+	jmp	.Lecb_dec_done
+.align	16
+.Lecb_dec_six:
+	call	_bsaes_decrypt8
+	movdqu	@XMM[0], 0x00($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($out)
+	jmp	.Lecb_dec_done
+.align	16
+.Lecb_dec_five:
+	call	_bsaes_decrypt8
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	movdqu	@XMM[6], 0x20($out)
+	movdqu	@XMM[4], 0x30($out)
+	movdqu	@XMM[2], 0x40($out)
+	jmp	.Lecb_dec_done
+.align	16
+.Lecb_dec_four:
+	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)
+	jmp	.Lecb_dec_done
+.align	16
+.Lecb_dec_three:
+	call	_bsaes_decrypt8
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	movdqu	@XMM[6], 0x20($out)
+	jmp	.Lecb_dec_done
+.align	16
+.Lecb_dec_two:
+	call	_bsaes_decrypt8
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	jmp	.Lecb_dec_done
+.align	16
+.Lecb_dec_one:
+	call	_bsaes_decrypt8
+	movdqu	@XMM[0], 0x00($out)	# write output
+	jmp	.Lecb_dec_done
+.align	16
+.Lecb_dec_short:
+	lea	($inp), $arg1
+	lea	($out), $arg2
+	lea	($key), $arg3
+	call	asm_AES_decrypt
+	lea	16($inp), $inp
+	lea	16($out), $out
+	dec	$len
+	jnz	.Lecb_dec_short
+
+.Lecb_dec_done:
+	lea	(%rsp),%rax
+	pxor	%xmm0, %xmm0
+.Lecb_dec_bzero:			# wipe key schedule [if any]
+	movdqa	%xmm0, 0x00(%rax)
+	movdqa	%xmm0, 0x10(%rax)
+	lea	0x20(%rax), %rax
+	cmp	%rax, %rbp
+	jb	.Lecb_dec_bzero
+
+	lea	(%rbp),%rsp		# restore %rsp
+___
+$code.=<<___ if ($win64);
+	movaps	0x40(%rbp), %xmm6
+	movaps	0x50(%rbp), %xmm7
+	movaps	0x60(%rbp), %xmm8
+	movaps	0x70(%rbp), %xmm9
+	movaps	0x80(%rbp), %xmm10
+	movaps	0x90(%rbp), %xmm11
+	movaps	0xa0(%rbp), %xmm12
+	movaps	0xb0(%rbp), %xmm13
+	movaps	0xc0(%rbp), %xmm14
+	movaps	0xd0(%rbp), %xmm15
+	lea	0xa0(%rbp), %rsp
+___
+$code.=<<___;
+	mov	0x48(%rsp), %r15
+	mov	0x50(%rsp), %r14
+	mov	0x58(%rsp), %r13
+	mov	0x60(%rsp), %r12
+	mov	0x68(%rsp), %rbx
+	mov	0x70(%rsp), %rax
+	lea	0x78(%rsp), %rsp
+	mov	%rax, %rbp
+.Lecb_dec_epilogue:
+	ret
+.size	bsaes_ecb_decrypt_blocks,.-bsaes_ecb_decrypt_blocks
+___
+}
+$code.=<<___;
+.extern	asm_AES_cbc_encrypt
+.globl	bsaes_cbc_encrypt
+.type	bsaes_cbc_encrypt,\@abi-omnipotent
+.align	16
+bsaes_cbc_encrypt:
+___
+$code.=<<___ if ($win64);
+	mov	48(%rsp),$arg6		# pull direction flag
+___
+$code.=<<___;
+	cmp	\$0,$arg6
+	jne	asm_AES_cbc_encrypt
+	cmp	\$128,$arg3
+	jb	asm_AES_cbc_encrypt
+
+	mov	%rsp, %rax
+.Lcbc_dec_prologue:
+	push	%rbp
+	push	%rbx
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	lea	-0x48(%rsp), %rsp
+___
+$code.=<<___ if ($win64);
+	mov	0xa0(%rsp),$arg5	# pull ivp
+	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)
+.Lcbc_dec_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
+	mov	$arg5, %rbx
+	shr	\$4, $len		# bytes to blocks
+
+	mov	%eax, %edx		# 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	$key, %rcx		# pass key
+	mov	%edx, %r10d		# pass rounds
+	call	_bsaes_key_convert
+	pxor	(%rsp),%xmm7		# fix up 0 round key
+	movdqa	%xmm6,(%rax)		# save last round key
+	movdqa	%xmm7,(%rsp)
+
+	movdqu	(%rbx), @XMM[15]	# load IV
+	sub	\$8,$len
+.Lcbc_dec_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($inp), @XMM[6]
+	mov	%edx,%r10d		# pass rounds
+	movdqu	0x70($inp), @XMM[7]
+	movdqa	@XMM[15], 0x20(%rbp)	# put aside IV
+
+	call	_bsaes_decrypt8
+
+	pxor	0x20(%rbp), @XMM[0]	# ^= IV
+	movdqu	0x00($inp), @XMM[8]	# re-load input
+	movdqu	0x10($inp), @XMM[9]
+	pxor	@XMM[8], @XMM[1]
+	movdqu	0x20($inp), @XMM[10]
+	pxor	@XMM[9], @XMM[6]
+	movdqu	0x30($inp), @XMM[11]
+	pxor	@XMM[10], @XMM[4]
+	movdqu	0x40($inp), @XMM[12]
+	pxor	@XMM[11], @XMM[2]
+	movdqu	0x50($inp), @XMM[13]
+	pxor	@XMM[12], @XMM[7]
+	movdqu	0x60($inp), @XMM[14]
+	pxor	@XMM[13], @XMM[3]
+	movdqu	0x70($inp), @XMM[15]	# IV
+	pxor	@XMM[14], @XMM[5]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	lea	0x80($inp), $inp
+	movdqu	@XMM[1], 0x10($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), $out
+	sub	\$8,$len
+	jnc	.Lcbc_dec_loop
+
+	add	\$8,$len
+	jz	.Lcbc_dec_done
+
+	movdqu	0x00($inp), @XMM[0]	# load input
+	mov	%rsp, %rax		# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+	cmp	\$2,$len
+	jb	.Lcbc_dec_one
+	movdqu	0x10($inp), @XMM[1]
+	je	.Lcbc_dec_two
+	movdqu	0x20($inp), @XMM[2]
+	cmp	\$4,$len
+	jb	.Lcbc_dec_three
+	movdqu	0x30($inp), @XMM[3]
+	je	.Lcbc_dec_four
+	movdqu	0x40($inp), @XMM[4]
+	cmp	\$6,$len
+	jb	.Lcbc_dec_five
+	movdqu	0x50($inp), @XMM[5]
+	je	.Lcbc_dec_six
+	movdqu	0x60($inp), @XMM[6]
+	movdqa	@XMM[15], 0x20(%rbp)	# put aside IV
+	call	_bsaes_decrypt8
+	pxor	0x20(%rbp), @XMM[0]	# ^= IV
+	movdqu	0x00($inp), @XMM[8]	# re-load input
+	movdqu	0x10($inp), @XMM[9]
+	pxor	@XMM[8], @XMM[1]
+	movdqu	0x20($inp), @XMM[10]
+	pxor	@XMM[9], @XMM[6]
+	movdqu	0x30($inp), @XMM[11]
+	pxor	@XMM[10], @XMM[4]
+	movdqu	0x40($inp), @XMM[12]
+	pxor	@XMM[11], @XMM[2]
+	movdqu	0x50($inp), @XMM[13]
+	pxor	@XMM[12], @XMM[7]
+	movdqu	0x60($inp), @XMM[15]	# IV
+	pxor	@XMM[13], @XMM[3]
+	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($out)
+	jmp	.Lcbc_dec_done
+.align	16
+.Lcbc_dec_six:
+	movdqa	@XMM[15], 0x20(%rbp)	# put aside IV
+	call	_bsaes_decrypt8
+	pxor	0x20(%rbp), @XMM[0]	# ^= IV
+	movdqu	0x00($inp), @XMM[8]	# re-load input
+	movdqu	0x10($inp), @XMM[9]
+	pxor	@XMM[8], @XMM[1]
+	movdqu	0x20($inp), @XMM[10]
+	pxor	@XMM[9], @XMM[6]
+	movdqu	0x30($inp), @XMM[11]
+	pxor	@XMM[10], @XMM[4]
+	movdqu	0x40($inp), @XMM[12]
+	pxor	@XMM[11], @XMM[2]
+	movdqu	0x50($inp), @XMM[15]	# IV
+	pxor	@XMM[12], @XMM[7]
+	movdqu	@XMM[0], 0x00($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($out)
+	jmp	.Lcbc_dec_done
+.align	16
+.Lcbc_dec_five:
+	movdqa	@XMM[15], 0x20(%rbp)	# put aside IV
+	call	_bsaes_decrypt8
+	pxor	0x20(%rbp), @XMM[0]	# ^= IV
+	movdqu	0x00($inp), @XMM[8]	# re-load input
+	movdqu	0x10($inp), @XMM[9]
+	pxor	@XMM[8], @XMM[1]
+	movdqu	0x20($inp), @XMM[10]
+	pxor	@XMM[9], @XMM[6]
+	movdqu	0x30($inp), @XMM[11]
+	pxor	@XMM[10], @XMM[4]
+	movdqu	0x40($inp), @XMM[15]	# IV
+	pxor	@XMM[11], @XMM[2]
+	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($out)
+	jmp	.Lcbc_dec_done
+.align	16
+.Lcbc_dec_four:
+	movdqa	@XMM[15], 0x20(%rbp)	# put aside IV
+	call	_bsaes_decrypt8
+	pxor	0x20(%rbp), @XMM[0]	# ^= IV
+	movdqu	0x00($inp), @XMM[8]	# re-load input
+	movdqu	0x10($inp), @XMM[9]
+	pxor	@XMM[8], @XMM[1]
+	movdqu	0x20($inp), @XMM[10]
+	pxor	@XMM[9], @XMM[6]
+	movdqu	0x30($inp), @XMM[15]	# IV
+	pxor	@XMM[10], @XMM[4]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	movdqu	@XMM[6], 0x20($out)
+	movdqu	@XMM[4], 0x30($out)
+	jmp	.Lcbc_dec_done
+.align	16
+.Lcbc_dec_three:
+	movdqa	@XMM[15], 0x20(%rbp)	# put aside IV
+	call	_bsaes_decrypt8
+	pxor	0x20(%rbp), @XMM[0]	# ^= IV
+	movdqu	0x00($inp), @XMM[8]	# re-load input
+	movdqu	0x10($inp), @XMM[9]
+	pxor	@XMM[8], @XMM[1]
+	movdqu	0x20($inp), @XMM[15]	# IV
+	pxor	@XMM[9], @XMM[6]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	movdqu	@XMM[6], 0x20($out)
+	jmp	.Lcbc_dec_done
+.align	16
+.Lcbc_dec_two:
+	movdqa	@XMM[15], 0x20(%rbp)	# put aside IV
+	call	_bsaes_decrypt8
+	pxor	0x20(%rbp), @XMM[0]	# ^= IV
+	movdqu	0x00($inp), @XMM[8]	# re-load input
+	movdqu	0x10($inp), @XMM[15]	# IV
+	pxor	@XMM[8], @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	jmp	.Lcbc_dec_done
+.align	16
+.Lcbc_dec_one:
+	lea	($inp), $arg1
+	lea	0x20(%rbp), $arg2	# buffer output
+	lea	($key), $arg3
+	call	asm_AES_decrypt		# doesn't touch %xmm
+	pxor	0x20(%rbp), @XMM[15]	# ^= IV
+	movdqu	@XMM[15], ($out)	# write output
+	movdqa	@XMM[0], @XMM[15]	# IV
+
+.Lcbc_dec_done:
+	movdqu	@XMM[15], (%rbx)	# return IV
+	lea	(%rsp), %rax
+	pxor	%xmm0, %xmm0
+.Lcbc_dec_bzero:			# wipe key schedule [if any]
+	movdqa	%xmm0, 0x00(%rax)
+	movdqa	%xmm0, 0x10(%rax)
+	lea	0x20(%rax), %rax
+	cmp	%rax, %rbp
+	ja	.Lcbc_dec_bzero
+
+	lea	(%rbp),%rsp		# restore %rsp
+___
+$code.=<<___ if ($win64);
+	movaps	0x40(%rbp), %xmm6
+	movaps	0x50(%rbp), %xmm7
+	movaps	0x60(%rbp), %xmm8
+	movaps	0x70(%rbp), %xmm9
+	movaps	0x80(%rbp), %xmm10
+	movaps	0x90(%rbp), %xmm11
+	movaps	0xa0(%rbp), %xmm12
+	movaps	0xb0(%rbp), %xmm13
+	movaps	0xc0(%rbp), %xmm14
+	movaps	0xd0(%rbp), %xmm15
+	lea	0xa0(%rbp), %rsp
+___
+$code.=<<___;
+	mov	0x48(%rsp), %r15
+	mov	0x50(%rsp), %r14
+	mov	0x58(%rsp), %r13
+	mov	0x60(%rsp), %r12
+	mov	0x68(%rsp), %rbx
+	mov	0x70(%rsp), %rax
+	lea	0x78(%rsp), %rsp
+	mov	%rax, %rbp
+.Lcbc_dec_epilogue:
+	ret
+.size	bsaes_cbc_encrypt,.-bsaes_cbc_encrypt
+
+.globl	bsaes_ctr32_encrypt_blocks
+.type	bsaes_ctr32_encrypt_blocks,\@abi-omnipotent
+.align	16
+bsaes_ctr32_encrypt_blocks:
+	mov	%rsp, %rax
+.Lctr_enc_prologue:
+	push	%rbp
+	push	%rbx
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	lea	-0x48(%rsp), %rsp
+___
+$code.=<<___ if ($win64);
+	mov	0xa0(%rsp),$arg5	# pull ivp
+	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)
+.Lctr_enc_body:
+___
+$code.=<<___;
+	mov	%rsp, %rbp		# backup %rsp
+	movdqu	($arg5), %xmm0		# load counter
+	mov	240($arg4), %eax	# rounds
+	mov	$arg1, $inp		# backup arguments
+	mov	$arg2, $out
+	mov	$arg3, $len
+	mov	$arg4, $key
+	movdqa	%xmm0, 0x20(%rbp)	# copy counter
+	cmp	\$8, $arg3
+	jb	.Lctr_enc_short
+
+	mov	%eax, %ebx		# 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	$key, %rcx		# pass key
+	mov	%ebx, %r10d		# pass rounds
+	call	_bsaes_key_convert
+	pxor	%xmm6,%xmm7		# fix up last round key
+	movdqa	%xmm7,(%rax)		# save last round key
+
+	movdqa	(%rsp), @XMM[9]		# load round0 key
+	lea	.LADD1(%rip), %r11
+	movdqa	0x20(%rbp), @XMM[0]	# counter copy
+	movdqa	-0x20(%r11), @XMM[8]	# .LSWPUP
+	pshufb	@XMM[8], @XMM[9]	# byte swap upper part
+	pshufb	@XMM[8], @XMM[0]
+	movdqa	@XMM[9], (%rsp)		# save adjusted round0 key
+	jmp	.Lctr_enc_loop
+.align	16
+.Lctr_enc_loop:
+	movdqa	@XMM[0], 0x20(%rbp)	# save counter
+	movdqa	@XMM[0], @XMM[1]	# prepare 8 counter values
+	movdqa	@XMM[0], @XMM[2]
+	paddd	0x00(%r11), @XMM[1]	# .LADD1
+	movdqa	@XMM[0], @XMM[3]
+	paddd	0x10(%r11), @XMM[2]	# .LADD2
+	movdqa	@XMM[0], @XMM[4]
+	paddd	0x20(%r11), @XMM[3]	# .LADD3
+	movdqa	@XMM[0], @XMM[5]
+	paddd	0x30(%r11), @XMM[4]	# .LADD4
+	movdqa	@XMM[0], @XMM[6]
+	paddd	0x40(%r11), @XMM[5]	# .LADD5
+	movdqa	@XMM[0], @XMM[7]
+	paddd	0x50(%r11), @XMM[6]	# .LADD6
+	paddd	0x60(%r11), @XMM[7]	# .LADD7
+
+	# Borrow prologue from _bsaes_encrypt8 to use the opportunity
+	# to flip byte order in 32-bit counter
+	movdqa	(%rsp), @XMM[9]		# round 0 key
+	lea	0x10(%rsp), %rax	# pass key schedule
+	movdqa	-0x10(%r11), @XMM[8]	# .LSWPUPM0SR
+	pxor	@XMM[9], @XMM[0]	# xor with round0 key
+	pxor	@XMM[9], @XMM[1]
+	 pshufb	@XMM[8], @XMM[0]
+	pxor	@XMM[9], @XMM[2]
+	 pshufb	@XMM[8], @XMM[1]
+	pxor	@XMM[9], @XMM[3]
+	 pshufb	@XMM[8], @XMM[2]
+	pxor	@XMM[9], @XMM[4]
+	 pshufb	@XMM[8], @XMM[3]
+	pxor	@XMM[9], @XMM[5]
+	 pshufb	@XMM[8], @XMM[4]
+	pxor	@XMM[9], @XMM[6]
+	 pshufb	@XMM[8], @XMM[5]
+	pxor	@XMM[9], @XMM[7]
+	 pshufb	@XMM[8], @XMM[6]
+	lea	.LBS0(%rip), %r11	# constants table
+	 pshufb	@XMM[8], @XMM[7]
+	mov	%ebx,%r10d		# pass rounds
+
+	call	_bsaes_encrypt8_bitslice
+
+	sub	\$8,$len
+	jc	.Lctr_enc_loop_done
+
+	movdqu	0x00($inp), @XMM[8]	# load input
+	movdqu	0x10($inp), @XMM[9]
+	movdqu	0x20($inp), @XMM[10]
+	movdqu	0x30($inp), @XMM[11]
+	movdqu	0x40($inp), @XMM[12]
+	movdqu	0x50($inp), @XMM[13]
+	movdqu	0x60($inp), @XMM[14]
+	movdqu	0x70($inp), @XMM[15]
+	lea	0x80($inp),$inp
+	pxor	@XMM[0], @XMM[8]
+	movdqa	0x20(%rbp), @XMM[0]	# load counter
+	pxor	@XMM[9], @XMM[1]
+	movdqu	@XMM[8], 0x00($out)	# write output
+	pxor	@XMM[10], @XMM[4]
+	movdqu	@XMM[1], 0x10($out)
+	pxor	@XMM[11], @XMM[6]
+	movdqu	@XMM[4], 0x20($out)
+	pxor	@XMM[12], @XMM[3]
+	movdqu	@XMM[6], 0x30($out)
+	pxor	@XMM[13], @XMM[7]
+	movdqu	@XMM[3], 0x40($out)
+	pxor	@XMM[14], @XMM[2]
+	movdqu	@XMM[7], 0x50($out)
+	pxor	@XMM[15], @XMM[5]
+	movdqu	@XMM[2], 0x60($out)
+	lea	.LADD1(%rip), %r11
+	movdqu	@XMM[5], 0x70($out)
+	lea	0x80($out), $out
+	paddd	0x70(%r11), @XMM[0]	# .LADD8
+	jnz	.Lctr_enc_loop
+
+	jmp	.Lctr_enc_done
+.align	16
+.Lctr_enc_loop_done:
+	add	\$8, $len
+	movdqu	0x00($inp), @XMM[8]	# load input
+	pxor	@XMM[8], @XMM[0]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	cmp	\$2,$len
+	jb	.Lctr_enc_done
+	movdqu	0x10($inp), @XMM[9]
+	pxor	@XMM[9], @XMM[1]
+	movdqu	@XMM[1], 0x10($out)
+	je	.Lctr_enc_done
+	movdqu	0x20($inp), @XMM[10]
+	pxor	@XMM[10], @XMM[4]
+	movdqu	@XMM[4], 0x20($out)
+	cmp	\$4,$len
+	jb	.Lctr_enc_done
+	movdqu	0x30($inp), @XMM[11]
+	pxor	@XMM[11], @XMM[6]
+	movdqu	@XMM[6], 0x30($out)
+	je	.Lctr_enc_done
+	movdqu	0x40($inp), @XMM[12]
+	pxor	@XMM[12], @XMM[3]
+	movdqu	@XMM[3], 0x40($out)
+	cmp	\$6,$len
+	jb	.Lctr_enc_done
+	movdqu	0x50($inp), @XMM[13]
+	pxor	@XMM[13], @XMM[7]
+	movdqu	@XMM[7], 0x50($out)
+	je	.Lctr_enc_done
+	movdqu	0x60($inp), @XMM[14]
+	pxor	@XMM[14], @XMM[2]
+	movdqu	@XMM[2], 0x60($out)
+	jmp	.Lctr_enc_done
+
+.align	16
+.Lctr_enc_short:
+	lea	0x20(%rbp), $arg1
+	lea	0x30(%rbp), $arg2
+	lea	($key), $arg3
+	call	asm_AES_encrypt
+	movdqu	($inp), @XMM[1]
+	lea	16($inp), $inp
+	mov	0x2c(%rbp), %eax	# load 32-bit counter
+	bswap	%eax
+	pxor	0x30(%rbp), @XMM[1]
+	inc	%eax			# increment
+	movdqu	@XMM[1], ($out)
+	bswap	%eax
+	lea	16($out), $out
+	mov	%eax, 0x2c(%rsp)	# save 32-bit counter
+	dec	$len
+	jnz	.Lctr_enc_short
+
+.Lctr_enc_done:
+	lea	(%rsp), %rax
+	pxor	%xmm0, %xmm0
+.Lctr_enc_bzero:			# wipe key schedule [if any]
+	movdqa	%xmm0, 0x00(%rax)
+	movdqa	%xmm0, 0x10(%rax)
+	lea	0x20(%rax), %rax
+	cmp	%rax, %rbp
+	ja	.Lctr_enc_bzero
+
+	lea	(%rbp),%rsp		# restore %rsp
+___
+$code.=<<___ if ($win64);
+	movaps	0x40(%rbp), %xmm6
+	movaps	0x50(%rbp), %xmm7
+	movaps	0x60(%rbp), %xmm8
+	movaps	0x70(%rbp), %xmm9
+	movaps	0x80(%rbp), %xmm10
+	movaps	0x90(%rbp), %xmm11
+	movaps	0xa0(%rbp), %xmm12
+	movaps	0xb0(%rbp), %xmm13
+	movaps	0xc0(%rbp), %xmm14
+	movaps	0xd0(%rbp), %xmm15
+	lea	0xa0(%rbp), %rsp
+___
+$code.=<<___;
+	mov	0x48(%rsp), %r15
+	mov	0x50(%rsp), %r14
+	mov	0x58(%rsp), %r13
+	mov	0x60(%rsp), %r12
+	mov	0x68(%rsp), %rbx
+	mov	0x70(%rsp), %rax
+	lea	0x78(%rsp), %rsp
+	mov	%rax, %rbp
+.Lctr_enc_epilogue:
+	ret
+.size	bsaes_ctr32_encrypt_blocks,.-bsaes_ctr32_encrypt_blocks
+___
+######################################################################
+# void bsaes_xts_[en|de]crypt(const char *inp,char *out,size_t len,
+#	const AES_KEY *key1, const AES_KEY *key2,
+#	const unsigned char iv[16]);
+#
+my ($twmask,$twres,$twtmp)=@XMM[13..15];
+$code.=<<___;
+.globl	bsaes_xts_encrypt
+.type	bsaes_xts_encrypt,\@abi-omnipotent
+.align	16
+bsaes_xts_encrypt:
+	mov	%rsp, %rax
+.Lxts_enc_prologue:
+	push	%rbp
+	push	%rbx
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	lea	-0x48(%rsp), %rsp
+___
+$code.=<<___ if ($win64);
+	mov	0xa0(%rsp),$arg5	# pull key2
+	mov	0xa8(%rsp),$arg6	# pull ivp
+	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)
+.Lxts_enc_body:
+___
+$code.=<<___;
+	mov	%rsp, %rbp		# backup %rsp
+	mov	$arg1, $inp		# backup arguments
+	mov	$arg2, $out
+	mov	$arg3, $len
+	mov	$arg4, $key
+
+	lea	($arg6), $arg1
+	lea	0x20(%rbp), $arg2
+	lea	($arg5), $arg3
+	call	asm_AES_encrypt		# generate initial tweak
+
+	mov	240($key), %eax		# rounds
+	mov	$len, %rbx		# backup $len
+
+	mov	%eax, %edx		# 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	%edx, %r10d		# pass rounds
+	call	_bsaes_key_convert
+	pxor	%xmm6, %xmm7		# fix up last round key
+	movdqa	%xmm7, (%rax)		# save last round key
+
+	and	\$-16, $len
+	sub	\$0x80, %rsp		# place for tweak[8]
+	movdqa	0x20(%rbp), @XMM[7]	# initial tweak
+
+	pxor	$twtmp, $twtmp
+	movdqa	.Lxts_magic(%rip), $twmask
+	pcmpgtd	@XMM[7], $twtmp		# broadcast upper bits
+
+	sub	\$0x80, $len
+	jc	.Lxts_enc_short
+	jmp	.Lxts_enc_loop
+
+.align	16
+.Lxts_enc_loop:
+___
+    for ($i=0;$i<7;$i++) {
+    $code.=<<___;
+	pshufd	\$0x13, $twtmp, $twres
+	pxor	$twtmp, $twtmp
+	movdqa	@XMM[7], @XMM[$i]
+	movdqa	@XMM[7], `0x10*$i`(%rsp)# save tweak[$i]
+	paddq	@XMM[7], @XMM[7]	# psllq	1,$tweak
+	pand	$twmask, $twres		# isolate carry and residue
+	pcmpgtd	@XMM[7], $twtmp		# broadcast upper bits
+	pxor	$twres, @XMM[7]
+___
+    $code.=<<___ if ($i>=1);
+	movdqu	`0x10*($i-1)`($inp), @XMM[8+$i-1]
+___
+    $code.=<<___ if ($i>=2);
+	pxor	@XMM[8+$i-2], @XMM[$i-2]# input[] ^ tweak[]
+___
+    }
+$code.=<<___;
+	movdqu	0x60($inp), @XMM[8+6]
+	pxor	@XMM[8+5], @XMM[5]
+	movdqu	0x70($inp), @XMM[8+7]
+	lea	0x80($inp), $inp
+	movdqa	@XMM[7], 0x70(%rsp)
+	pxor	@XMM[8+6], @XMM[6]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	pxor	@XMM[8+7], @XMM[7]
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_encrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[4]
+	movdqu	@XMM[1], 0x10($out)
+	pxor	0x30(%rsp), @XMM[6]
+	movdqu	@XMM[4], 0x20($out)
+	pxor	0x40(%rsp), @XMM[3]
+	movdqu	@XMM[6], 0x30($out)
+	pxor	0x50(%rsp), @XMM[7]
+	movdqu	@XMM[3], 0x40($out)
+	pxor	0x60(%rsp), @XMM[2]
+	movdqu	@XMM[7], 0x50($out)
+	pxor	0x70(%rsp), @XMM[5]
+	movdqu	@XMM[2], 0x60($out)
+	movdqu	@XMM[5], 0x70($out)
+	lea	0x80($out), $out
+
+	movdqa	0x70(%rsp), @XMM[7]	# prepare next iteration tweak
+	pxor	$twtmp, $twtmp
+	movdqa	.Lxts_magic(%rip), $twmask
+	pcmpgtd	@XMM[7], $twtmp
+	pshufd	\$0x13, $twtmp, $twres
+	pxor	$twtmp, $twtmp
+	paddq	@XMM[7], @XMM[7]	# psllq	1,$tweak
+	pand	$twmask, $twres		# isolate carry and residue
+	pcmpgtd	@XMM[7], $twtmp		# broadcast upper bits
+	pxor	$twres, @XMM[7]
+
+	sub	\$0x80,$len
+	jnc	.Lxts_enc_loop
+
+.Lxts_enc_short:
+	add	\$0x80, $len
+	jz	.Lxts_enc_done
+___
+    for ($i=0;$i<7;$i++) {
+    $code.=<<___;
+	pshufd	\$0x13, $twtmp, $twres
+	pxor	$twtmp, $twtmp
+	movdqa	@XMM[7], @XMM[$i]
+	movdqa	@XMM[7], `0x10*$i`(%rsp)# save tweak[$i]
+	paddq	@XMM[7], @XMM[7]	# psllq	1,$tweak
+	pand	$twmask, $twres		# isolate carry and residue
+	pcmpgtd	@XMM[7], $twtmp		# broadcast upper bits
+	pxor	$twres, @XMM[7]
+___
+    $code.=<<___ if ($i>=1);
+	movdqu	`0x10*($i-1)`($inp), @XMM[8+$i-1]
+	cmp	\$`0x10*$i`,$len
+	je	.Lxts_enc_$i
+___
+    $code.=<<___ if ($i>=2);
+	pxor	@XMM[8+$i-2], @XMM[$i-2]# input[] ^ tweak[]
+___
+    }
+$code.=<<___;
+	movdqu	0x60($inp), @XMM[8+6]
+	pxor	@XMM[8+5], @XMM[5]
+	movdqa	@XMM[7], 0x70(%rsp)
+	lea	0x70($inp), $inp
+	pxor	@XMM[8+6], @XMM[6]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_encrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[4]
+	movdqu	@XMM[1], 0x10($out)
+	pxor	0x30(%rsp), @XMM[6]
+	movdqu	@XMM[4], 0x20($out)
+	pxor	0x40(%rsp), @XMM[3]
+	movdqu	@XMM[6], 0x30($out)
+	pxor	0x50(%rsp), @XMM[7]
+	movdqu	@XMM[3], 0x40($out)
+	pxor	0x60(%rsp), @XMM[2]
+	movdqu	@XMM[7], 0x50($out)
+	movdqu	@XMM[2], 0x60($out)
+	lea	0x70($out), $out
+
+	movdqa	0x70(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_enc_done
+.align	16
+.Lxts_enc_6:
+	pxor	@XMM[8+4], @XMM[4]
+	lea	0x60($inp), $inp
+	pxor	@XMM[8+5], @XMM[5]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_encrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[4]
+	movdqu	@XMM[1], 0x10($out)
+	pxor	0x30(%rsp), @XMM[6]
+	movdqu	@XMM[4], 0x20($out)
+	pxor	0x40(%rsp), @XMM[3]
+	movdqu	@XMM[6], 0x30($out)
+	pxor	0x50(%rsp), @XMM[7]
+	movdqu	@XMM[3], 0x40($out)
+	movdqu	@XMM[7], 0x50($out)
+	lea	0x60($out), $out
+
+	movdqa	0x60(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_enc_done
+.align	16
+.Lxts_enc_5:
+	pxor	@XMM[8+3], @XMM[3]
+	lea	0x50($inp), $inp
+	pxor	@XMM[8+4], @XMM[4]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_encrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[4]
+	movdqu	@XMM[1], 0x10($out)
+	pxor	0x30(%rsp), @XMM[6]
+	movdqu	@XMM[4], 0x20($out)
+	pxor	0x40(%rsp), @XMM[3]
+	movdqu	@XMM[6], 0x30($out)
+	movdqu	@XMM[3], 0x40($out)
+	lea	0x50($out), $out
+
+	movdqa	0x50(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_enc_done
+.align	16
+.Lxts_enc_4:
+	pxor	@XMM[8+2], @XMM[2]
+	lea	0x40($inp), $inp
+	pxor	@XMM[8+3], @XMM[3]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_encrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[4]
 	movdqu	@XMM[1], 0x10($out)
-	pxor	@XMM[11], @XMM[6]
+	pxor	0x30(%rsp), @XMM[6]
 	movdqu	@XMM[4], 0x20($out)
-	pxor	@XMM[12], @XMM[3]
 	movdqu	@XMM[6], 0x30($out)
-	pxor	@XMM[13], @XMM[7]
-	movdqu	@XMM[3], 0x40($out)
-	pxor	@XMM[14], @XMM[2]
+	lea	0x40($out), $out
+
+	movdqa	0x40(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_enc_done
+.align	16
+.Lxts_enc_3:
+	pxor	@XMM[8+1], @XMM[1]
+	lea	0x30($inp), $inp
+	pxor	@XMM[8+2], @XMM[2]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_encrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[4]
+	movdqu	@XMM[1], 0x10($out)
+	movdqu	@XMM[4], 0x20($out)
+	lea	0x30($out), $out
+
+	movdqa	0x30(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_enc_done
+.align	16
+.Lxts_enc_2:
+	pxor	@XMM[8+0], @XMM[0]
+	lea	0x20($inp), $inp
+	pxor	@XMM[8+1], @XMM[1]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_encrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	lea	0x20($out), $out
+
+	movdqa	0x20(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_enc_done
+.align	16
+.Lxts_enc_1:
+	pxor	@XMM[0], @XMM[8]
+	lea	0x10($inp), $inp
+	movdqa	@XMM[8], 0x20(%rbp)
+	lea	0x20(%rbp), $arg1
+	lea	0x20(%rbp), $arg2
+	lea	($key), $arg3
+	call	asm_AES_encrypt		# doesn't touch %xmm
+	pxor	0x20(%rbp), @XMM[0]	# ^= tweak[]
+	#pxor	@XMM[8], @XMM[0]
+	#lea	0x80(%rsp), %rax	# pass key schedule
+	#mov	%edx, %r10d		# pass rounds
+	#call	_bsaes_encrypt8
+	#pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	lea	0x10($out), $out
+
+	movdqa	0x10(%rsp), @XMM[7]	# next iteration tweak
+
+.Lxts_enc_done:
+	and	\$15, %ebx
+	jz	.Lxts_enc_ret
+	mov	$out, %rdx
+
+.Lxts_enc_steal:
+	movzb	($inp), %eax
+	movzb	-16(%rdx), %ecx
+	lea	1($inp), $inp
+	mov	%al, -16(%rdx)
+	mov	%cl, 0(%rdx)
+	lea	1(%rdx), %rdx
+	sub	\$1,%ebx
+	jnz	.Lxts_enc_steal
+
+	movdqu	-16($out), @XMM[0]
+	lea	0x20(%rbp), $arg1
+	pxor	@XMM[7], @XMM[0]
+	lea	0x20(%rbp), $arg2
+	movdqa	@XMM[0], 0x20(%rbp)
+	lea	($key), $arg3
+	call	asm_AES_encrypt		# doesn't touch %xmm
+	pxor	0x20(%rbp), @XMM[7]
+	movdqu	@XMM[7], -16($out)
+
+.Lxts_enc_ret:
+	lea	(%rsp), %rax
+	pxor	%xmm0, %xmm0
+.Lxts_enc_bzero:			# wipe key schedule [if any]
+	movdqa	%xmm0, 0x00(%rax)
+	movdqa	%xmm0, 0x10(%rax)
+	lea	0x20(%rax), %rax
+	cmp	%rax, %rbp
+	ja	.Lxts_enc_bzero
+
+	lea	(%rbp),%rsp		# restore %rsp
+___
+$code.=<<___ if ($win64);
+	movaps	0x40(%rbp), %xmm6
+	movaps	0x50(%rbp), %xmm7
+	movaps	0x60(%rbp), %xmm8
+	movaps	0x70(%rbp), %xmm9
+	movaps	0x80(%rbp), %xmm10
+	movaps	0x90(%rbp), %xmm11
+	movaps	0xa0(%rbp), %xmm12
+	movaps	0xb0(%rbp), %xmm13
+	movaps	0xc0(%rbp), %xmm14
+	movaps	0xd0(%rbp), %xmm15
+	lea	0xa0(%rbp), %rsp
+___
+$code.=<<___;
+	mov	0x48(%rsp), %r15
+	mov	0x50(%rsp), %r14
+	mov	0x58(%rsp), %r13
+	mov	0x60(%rsp), %r12
+	mov	0x68(%rsp), %rbx
+	mov	0x70(%rsp), %rax
+	lea	0x78(%rsp), %rsp
+	mov	%rax, %rbp
+.Lxts_enc_epilogue:
+	ret
+.size	bsaes_xts_encrypt,.-bsaes_xts_encrypt
+
+.globl	bsaes_xts_decrypt
+.type	bsaes_xts_decrypt,\@abi-omnipotent
+.align	16
+bsaes_xts_decrypt:
+	mov	%rsp, %rax
+.Lxts_dec_prologue:
+	push	%rbp
+	push	%rbx
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	lea	-0x48(%rsp), %rsp
+___
+$code.=<<___ if ($win64);
+	mov	0xa0(%rsp),$arg5	# pull key2
+	mov	0xa8(%rsp),$arg6	# pull ivp
+	lea	-0xa0(%rsp), %rsp
+	movaps	%xmm6, 0x40(%rsp)
+	movaps	%xmm7, 0x50(%rsp)
+	movaps	%xmm8, 0x60(%rsp)
+	movaps	%xmm9, 0x70(%rsp)
+	movaps	%xmm10, 0x80(%rsp)
+	movaps	%xmm11, 0x90(%rsp)
+	movaps	%xmm12, 0xa0(%rsp)
+	movaps	%xmm13, 0xb0(%rsp)
+	movaps	%xmm14, 0xc0(%rsp)
+	movaps	%xmm15, 0xd0(%rsp)
+.Lxts_dec_body:
+___
+$code.=<<___;
+	mov	%rsp, %rbp		# backup %rsp
+	mov	$arg1, $inp		# backup arguments
+	mov	$arg2, $out
+	mov	$arg3, $len
+	mov	$arg4, $key
+
+	lea	($arg6), $arg1
+	lea	0x20(%rbp), $arg2
+	lea	($arg5), $arg3
+	call	asm_AES_encrypt		# generate initial tweak
+
+	mov	240($key), %eax		# rounds
+	mov	$len, %rbx		# backup $len
+
+	mov	%eax, %edx		# 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	$key, %rcx		# pass key
+	mov	%edx, %r10d		# pass rounds
+	call	_bsaes_key_convert
+	pxor	(%rsp), %xmm7		# fix up round 0 key
+	movdqa	%xmm6, (%rax)		# save last round key
+	movdqa	%xmm7, (%rsp)
+
+	xor	%eax, %eax		# if ($len%16) len-=16;
+	and	\$-16, $len
+	test	\$15, %ebx
+	setnz	%al
+	shl	\$4, %rax
+	sub	%rax, $len
+
+	sub	\$0x80, %rsp		# place for tweak[8]
+	movdqa	0x20(%rbp), @XMM[7]	# initial tweak
+
+	pxor	$twtmp, $twtmp
+	movdqa	.Lxts_magic(%rip), $twmask
+	pcmpgtd	@XMM[7], $twtmp		# broadcast upper bits
+
+	sub	\$0x80, $len
+	jc	.Lxts_dec_short
+	jmp	.Lxts_dec_loop
+
+.align	16
+.Lxts_dec_loop:
+___
+    for ($i=0;$i<7;$i++) {
+    $code.=<<___;
+	pshufd	\$0x13, $twtmp, $twres
+	pxor	$twtmp, $twtmp
+	movdqa	@XMM[7], @XMM[$i]
+	movdqa	@XMM[7], `0x10*$i`(%rsp)# save tweak[$i]
+	paddq	@XMM[7], @XMM[7]	# psllq	1,$tweak
+	pand	$twmask, $twres		# isolate carry and residue
+	pcmpgtd	@XMM[7], $twtmp		# broadcast upper bits
+	pxor	$twres, @XMM[7]
+___
+    $code.=<<___ if ($i>=1);
+	movdqu	`0x10*($i-1)`($inp), @XMM[8+$i-1]
+___
+    $code.=<<___ if ($i>=2);
+	pxor	@XMM[8+$i-2], @XMM[$i-2]# input[] ^ tweak[]
+___
+    }
+$code.=<<___;
+	movdqu	0x60($inp), @XMM[8+6]
+	pxor	@XMM[8+5], @XMM[5]
+	movdqu	0x70($inp), @XMM[8+7]
+	lea	0x80($inp), $inp
+	movdqa	@XMM[7], 0x70(%rsp)
+	pxor	@XMM[8+6], @XMM[6]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	pxor	@XMM[8+7], @XMM[7]
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_decrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[6]
+	movdqu	@XMM[1], 0x10($out)
+	pxor	0x30(%rsp), @XMM[4]
+	movdqu	@XMM[6], 0x20($out)
+	pxor	0x40(%rsp), @XMM[2]
+	movdqu	@XMM[4], 0x30($out)
+	pxor	0x50(%rsp), @XMM[7]
+	movdqu	@XMM[2], 0x40($out)
+	pxor	0x60(%rsp), @XMM[3]
 	movdqu	@XMM[7], 0x50($out)
-	pxor	@XMM[15], @XMM[5]
-	movdqu	@XMM[2], 0x60($out)
-	lea	.LADD1(%rip), %r11
+	pxor	0x70(%rsp), @XMM[5]
+	movdqu	@XMM[3], 0x60($out)
 	movdqu	@XMM[5], 0x70($out)
 	lea	0x80($out), $out
-	paddd	0x70(%r11), @XMM[0]	# .LADD8
-	jnz	.Lctr_enc_loop
 
-	jmp	.Lctr_enc_done
+	movdqa	0x70(%rsp), @XMM[7]	# prepare next iteration tweak
+	pxor	$twtmp, $twtmp
+	movdqa	.Lxts_magic(%rip), $twmask
+	pcmpgtd	@XMM[7], $twtmp
+	pshufd	\$0x13, $twtmp, $twres
+	pxor	$twtmp, $twtmp
+	paddq	@XMM[7], @XMM[7]	# psllq	1,$tweak
+	pand	$twmask, $twres		# isolate carry and residue
+	pcmpgtd	@XMM[7], $twtmp		# broadcast upper bits
+	pxor	$twres, @XMM[7]
+
+	sub	\$0x80,$len
+	jnc	.Lxts_dec_loop
+
+.Lxts_dec_short:
+	add	\$0x80, $len
+	jz	.Lxts_dec_done
+___
+    for ($i=0;$i<7;$i++) {
+    $code.=<<___;
+	pshufd	\$0x13, $twtmp, $twres
+	pxor	$twtmp, $twtmp
+	movdqa	@XMM[7], @XMM[$i]
+	movdqa	@XMM[7], `0x10*$i`(%rsp)# save tweak[$i]
+	paddq	@XMM[7], @XMM[7]	# psllq	1,$tweak
+	pand	$twmask, $twres		# isolate carry and residue
+	pcmpgtd	@XMM[7], $twtmp		# broadcast upper bits
+	pxor	$twres, @XMM[7]
+___
+    $code.=<<___ if ($i>=1);
+	movdqu	`0x10*($i-1)`($inp), @XMM[8+$i-1]
+	cmp	\$`0x10*$i`,$len
+	je	.Lxts_dec_$i
+___
+    $code.=<<___ if ($i>=2);
+	pxor	@XMM[8+$i-2], @XMM[$i-2]# input[] ^ tweak[]
+___
+    }
+$code.=<<___;
+	movdqu	0x60($inp), @XMM[8+6]
+	pxor	@XMM[8+5], @XMM[5]
+	movdqa	@XMM[7], 0x70(%rsp)
+	lea	0x70($inp), $inp
+	pxor	@XMM[8+6], @XMM[6]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_decrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[6]
+	movdqu	@XMM[1], 0x10($out)
+	pxor	0x30(%rsp), @XMM[4]
+	movdqu	@XMM[6], 0x20($out)
+	pxor	0x40(%rsp), @XMM[2]
+	movdqu	@XMM[4], 0x30($out)
+	pxor	0x50(%rsp), @XMM[7]
+	movdqu	@XMM[2], 0x40($out)
+	pxor	0x60(%rsp), @XMM[3]
+	movdqu	@XMM[7], 0x50($out)
+	movdqu	@XMM[3], 0x60($out)
+	lea	0x70($out), $out
+
+	movdqa	0x70(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_dec_done
 .align	16
-.Lctr_enc_loop_done:
-	movdqu	0x00($inp), @XMM[8]	# load input
-	pxor	@XMM[8], @XMM[0]
+.Lxts_dec_6:
+	pxor	@XMM[8+4], @XMM[4]
+	lea	0x60($inp), $inp
+	pxor	@XMM[8+5], @XMM[5]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_decrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
 	movdqu	@XMM[0], 0x00($out)	# write output
-	cmp	\$2,$len
-	jb	.Lctr_enc_done
-	movdqu	0x10($inp), @XMM[9]
-	pxor	@XMM[9], @XMM[1]
+	pxor	0x20(%rsp), @XMM[6]
 	movdqu	@XMM[1], 0x10($out)
-	je	.Lctr_enc_done
-	movdqu	0x20($inp), @XMM[10]
-	pxor	@XMM[10], @XMM[4]
-	movdqu	@XMM[4], 0x20($out)
-	cmp	\$4,$len
-	jb	.Lctr_enc_done
-	movdqu	0x30($inp), @XMM[11]
-	pxor	@XMM[11], @XMM[6]
-	movdqu	@XMM[6], 0x30($out)
-	je	.Lctr_enc_done
-	movdqu	0x40($inp), @XMM[12]
-	pxor	@XMM[12], @XMM[3]
-	movdqu	@XMM[3], 0x40($out)
-	cmp	\$6,$len
-	jb	.Lctr_enc_done
-	movdqu	0x50($inp), @XMM[13]
-	pxor	@XMM[13], @XMM[7]
+	pxor	0x30(%rsp), @XMM[4]
+	movdqu	@XMM[6], 0x20($out)
+	pxor	0x40(%rsp), @XMM[2]
+	movdqu	@XMM[4], 0x30($out)
+	pxor	0x50(%rsp), @XMM[7]
+	movdqu	@XMM[2], 0x40($out)
 	movdqu	@XMM[7], 0x50($out)
-	je	.Lctr_enc_done
-	movdqu	0x60($inp), @XMM[14]
-	pxor	@XMM[14], @XMM[2]
-	movdqu	@XMM[2], 0x60($out)
-	jmp	.Lctr_enc_done
+	lea	0x60($out), $out
 
+	movdqa	0x60(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_dec_done
 .align	16
-.Lctr_enc_short:
+.Lxts_dec_5:
+	pxor	@XMM[8+3], @XMM[3]
+	lea	0x50($inp), $inp
+	pxor	@XMM[8+4], @XMM[4]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_decrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[6]
+	movdqu	@XMM[1], 0x10($out)
+	pxor	0x30(%rsp), @XMM[4]
+	movdqu	@XMM[6], 0x20($out)
+	pxor	0x40(%rsp), @XMM[2]
+	movdqu	@XMM[4], 0x30($out)
+	movdqu	@XMM[2], 0x40($out)
+	lea	0x50($out), $out
+
+	movdqa	0x50(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_dec_done
+.align	16
+.Lxts_dec_4:
+	pxor	@XMM[8+2], @XMM[2]
+	lea	0x40($inp), $inp
+	pxor	@XMM[8+3], @XMM[3]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_decrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[6]
+	movdqu	@XMM[1], 0x10($out)
+	pxor	0x30(%rsp), @XMM[4]
+	movdqu	@XMM[6], 0x20($out)
+	movdqu	@XMM[4], 0x30($out)
+	lea	0x40($out), $out
+
+	movdqa	0x40(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_dec_done
+.align	16
+.Lxts_dec_3:
+	pxor	@XMM[8+1], @XMM[1]
+	lea	0x30($inp), $inp
+	pxor	@XMM[8+2], @XMM[2]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_decrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	pxor	0x20(%rsp), @XMM[6]
+	movdqu	@XMM[1], 0x10($out)
+	movdqu	@XMM[6], 0x20($out)
+	lea	0x30($out), $out
+
+	movdqa	0x30(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_dec_done
+.align	16
+.Lxts_dec_2:
+	pxor	@XMM[8+0], @XMM[0]
+	lea	0x20($inp), $inp
+	pxor	@XMM[8+1], @XMM[1]
+	lea	0x80(%rsp), %rax	# pass key schedule
+	mov	%edx, %r10d		# pass rounds
+
+	call	_bsaes_decrypt8
+
+	pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	pxor	0x10(%rsp), @XMM[1]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	movdqu	@XMM[1], 0x10($out)
+	lea	0x20($out), $out
+
+	movdqa	0x20(%rsp), @XMM[7]	# next iteration tweak
+	jmp	.Lxts_dec_done
+.align	16
+.Lxts_dec_1:
+	pxor	@XMM[0], @XMM[8]
+	lea	0x10($inp), $inp
+	movdqa	@XMM[8], 0x20(%rbp)
 	lea	0x20(%rbp), $arg1
-	lea	0x30(%rbp), $arg2
+	lea	0x20(%rbp), $arg2
 	lea	($key), $arg3
-	call	AES_encrypt
-	movdqu	($inp), @XMM[1]
-	lea	16($inp), $inp
-	mov	0x2c(%rbp), %eax	# load 32-bit counter
-	bswap	%eax
-	pxor	0x30(%rbp), @XMM[1]
-	inc	%eax			# increment
-	movdqu	@XMM[1], ($out)
-	bswap	%eax
-	lea	16($out), $out
-	mov	%eax, 0x2c(%rsp)	# save 32-bit counter
-	dec	$len
-	jnz	.Lctr_enc_short
+	call	asm_AES_decrypt		# doesn't touch %xmm
+	pxor	0x20(%rbp), @XMM[0]	# ^= tweak[]
+	#pxor	@XMM[8], @XMM[0]
+	#lea	0x80(%rsp), %rax	# pass key schedule
+	#mov	%edx, %r10d		# pass rounds
+	#call	_bsaes_decrypt8
+	#pxor	0x00(%rsp), @XMM[0]	# ^= tweak[]
+	movdqu	@XMM[0], 0x00($out)	# write output
+	lea	0x10($out), $out
 
-.Lctr_enc_done:
+	movdqa	0x10(%rsp), @XMM[7]	# next iteration tweak
+
+.Lxts_dec_done:
+	and	\$15, %ebx
+	jz	.Lxts_dec_ret
+
+	pxor	$twtmp, $twtmp
+	movdqa	.Lxts_magic(%rip), $twmask
+	pcmpgtd	@XMM[7], $twtmp
+	pshufd	\$0x13, $twtmp, $twres
+	movdqa	@XMM[7], @XMM[6]
+	paddq	@XMM[7], @XMM[7]	# psllq 1,$tweak
+	pand	$twmask, $twres		# isolate carry and residue
+	movdqu	($inp), @XMM[0]
+	pxor	$twres, @XMM[7]
+
+	lea	0x20(%rbp), $arg1
+	pxor	@XMM[7], @XMM[0]
+	lea	0x20(%rbp), $arg2
+	movdqa	@XMM[0], 0x20(%rbp)
+	lea	($key), $arg3
+	call	asm_AES_decrypt		# doesn't touch %xmm
+	pxor	0x20(%rbp), @XMM[7]
+	mov	$out, %rdx
+	movdqu	@XMM[7], ($out)
+
+.Lxts_dec_steal:
+	movzb	16($inp), %eax
+	movzb	(%rdx), %ecx
+	lea	1($inp), $inp
+	mov	%al, (%rdx)
+	mov	%cl, 16(%rdx)
+	lea	1(%rdx), %rdx
+	sub	\$1,%ebx
+	jnz	.Lxts_dec_steal
+
+	movdqu	($out), @XMM[0]
+	lea	0x20(%rbp), $arg1
+	pxor	@XMM[6], @XMM[0]
+	lea	0x20(%rbp), $arg2
+	movdqa	@XMM[0], 0x20(%rbp)
+	lea	($key), $arg3
+	call	asm_AES_decrypt		# doesn't touch %xmm
+	pxor	0x20(%rbp), @XMM[6]
+	movdqu	@XMM[6], ($out)
+
+.Lxts_dec_ret:
 	lea	(%rsp), %rax
 	pxor	%xmm0, %xmm0
-.Lctr_enc_bzero:			# wipe key schedule [if any]
+.Lxts_dec_bzero:			# wipe key schedule [if any]
 	movdqa	%xmm0, 0x00(%rax)
 	movdqa	%xmm0, 0x10(%rax)
 	lea	0x20(%rax), %rax
 	cmp	%rax, %rbp
-	ja	.Lctr_enc_bzero
+	ja	.Lxts_dec_bzero
 
 	lea	(%rbp),%rsp		# restore %rsp
 ___
@@ -1143,15 +2809,24 @@ $code.=<<___;
 	mov	0x58(%rsp), %r13
 	mov	0x60(%rsp), %r12
 	mov	0x68(%rsp), %rbx
-	mov	0x70(%rsp), %rbp
+	mov	0x70(%rsp), %rax
 	lea	0x78(%rsp), %rsp
-.Lctr_enc_epilogue:
+	mov	%rax, %rbp
+.Lxts_dec_epilogue:
 	ret
-.size	bsaes_ctr32_encrypt_blocks,.-bsaes_ctr32_encrypt_blocks
+.size	bsaes_xts_decrypt,.-bsaes_xts_decrypt
 ___
 }
 $code.=<<___;
+.type	_bsaes_const,\@object
 .align	64
+_bsaes_const:
+.LM0ISR:	# InvShiftRows constants
+	.quad	0x0a0e0206070b0f03, 0x0004080c0d010509
+.LISRM0:
+	.quad	0x01040b0e0205080f, 0x0306090c00070a0d
+.LISR:
+	.quad	0x0504070602010003, 0x0f0e0d0c080b0a09
 .LBS0:		# bit-slice constants
 	.quad	0x5555555555555555, 0x5555555555555555
 .LBS1:
@@ -1162,14 +2837,8 @@ $code.=<<___;
 	.quad	0x0504070600030201, 0x0f0e0d0c0a09080b
 .LSRM0:
 	.quad	0x0304090e00050a0f, 0x01060b0c0207080d
-.LM0:
-	.quad	0x02060a0e03070b0f, 0x0004080c0105090d
 .LM0SR:
 	.quad	0x0a0e02060f03070b, 0x0004080c05090d01
-.LNOT:		# magic constants
-	.quad	0xffffffffffffffff, 0xffffffffffffffff
-.L63:
-	.quad	0x6363636363636363, 0x6363636363636363
 .LSWPUP:	# byte-swap upper dword
 	.quad	0x0706050403020100, 0x0c0d0e0f0b0a0908
 .LSWPUPM0SR:
@@ -1190,9 +2859,183 @@ $code.=<<___;
 	.quad	0x0000000000000000, 0x0000000700000000
 .LADD8:
 	.quad	0x0000000000000000, 0x0000000800000000
-.asciz	"Bit-sliced AES for x86_64/SSSE3, Emilia Käsper and Peter Schwabe"
+.Lxts_magic:
+	.long	0x87,0,1,0
+.Lmasks:
+	.quad	0x0101010101010101, 0x0101010101010101
+	.quad	0x0202020202020202, 0x0202020202020202
+	.quad	0x0404040404040404, 0x0404040404040404
+	.quad	0x0808080808080808, 0x0808080808080808
+.LM0:
+	.quad	0x02060a0e03070b0f, 0x0004080c0105090d
+.L63:
+	.quad	0x6363636363636363, 0x6363636363636363
+.asciz	"Bit-sliced AES for x86_64/SSSE3, Emilia Käsper, Peter Schwabe, Andy Polyakov"
 .align	64
+.size	_bsaes_const,.-_bsaes_const
+___
+
+# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
+#		CONTEXT *context,DISPATCHER_CONTEXT *disp)
+if ($win64) {
+$rec="%rcx";
+$frame="%rdx";
+$context="%r8";
+$disp="%r9";
+
+$code.=<<___;
+.extern	__imp_RtlVirtualUnwind
+.type	se_handler,\@abi-omnipotent
+.align	16
+se_handler:
+	push	%rsi
+	push	%rdi
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	pushfq
+	sub	\$64,%rsp
+
+	mov	120($context),%rax	# pull context->Rax
+	mov	248($context),%rbx	# pull context->Rip
+
+	mov	8($disp),%rsi		# disp->ImageBase
+	mov	56($disp),%r11		# disp->HandlerData
+
+	mov	0(%r11),%r10d		# HandlerData[0]
+	lea	(%rsi,%r10),%r10	# prologue label
+	cmp	%r10,%rbx		# context->Rip<prologue label
+	jb	.Lin_prologue
+
+	mov	152($context),%rax	# pull context->Rsp
+
+	mov	4(%r11),%r10d		# HandlerData[1]
+	lea	(%rsi,%r10),%r10	# epilogue label
+	cmp	%r10,%rbx		# context->Rip>=epilogue label
+	jae	.Lin_prologue
+
+	mov	160($context),%rax	# pull context->Rbp
+
+	lea	0x40(%rax),%rsi		# %xmm save area
+	lea	512($context),%rdi	# &context.Xmm6
+	mov	\$20,%ecx		# 10*sizeof(%xmm0)/sizeof(%rax)
+	.long	0xa548f3fc		# cld; rep movsq
+	lea	0xa0(%rax),%rax		# adjust stack pointer
+
+	mov	0x70(%rax),%rbp
+	mov	0x68(%rax),%rbx
+	mov	0x60(%rax),%r12
+	mov	0x58(%rax),%r13
+	mov	0x50(%rax),%r14
+	mov	0x48(%rax),%r15
+	lea	0x78(%rax),%rax		# adjust stack pointer
+	mov	%rbx,144($context)	# restore context->Rbx
+	mov	%rbp,160($context)	# restore context->Rbp
+	mov	%r12,216($context)	# restore context->R12
+	mov	%r13,224($context)	# restore context->R13
+	mov	%r14,232($context)	# restore context->R14
+	mov	%r15,240($context)	# restore context->R15
+
+.Lin_prologue:
+	mov	%rax,152($context)	# restore context->Rsp
+
+	mov	40($disp),%rdi		# disp->ContextRecord
+	mov	$context,%rsi		# context
+	mov	\$`1232/8`,%ecx		# sizeof(CONTEXT)
+	.long	0xa548f3fc		# cld; rep movsq
+
+	mov	$disp,%rsi
+	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
+	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
+	mov	0(%rsi),%r8		# arg3, disp->ControlPc
+	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
+	mov	40(%rsi),%r10		# disp->ContextRecord
+	lea	56(%rsi),%r11		# &disp->HandlerData
+	lea	24(%rsi),%r12		# &disp->EstablisherFrame
+	mov	%r10,32(%rsp)		# arg5
+	mov	%r11,40(%rsp)		# arg6
+	mov	%r12,48(%rsp)		# arg7
+	mov	%rcx,56(%rsp)		# arg8, (NULL)
+	call	*__imp_RtlVirtualUnwind(%rip)
+
+	mov	\$1,%eax		# ExceptionContinueSearch
+	add	\$64,%rsp
+	popfq
+	pop	%r15
+	pop	%r14
+	pop	%r13
+	pop	%r12
+	pop	%rbp
+	pop	%rbx
+	pop	%rdi
+	pop	%rsi
+	ret
+.size	se_handler,.-se_handler
+
+.section	.pdata
+.align	4
+___
+$code.=<<___ if ($ecb);
+	.rva	.Lecb_enc_prologue
+	.rva	.Lecb_enc_epilogue
+	.rva	.Lecb_enc_info
+
+	.rva	.Lecb_dec_prologue
+	.rva	.Lecb_dec_epilogue
+	.rva	.Lecb_dec_info
+___
+$code.=<<___;
+	.rva	.Lcbc_dec_prologue
+	.rva	.Lcbc_dec_epilogue
+	.rva	.Lcbc_dec_info
+
+	.rva	.Lctr_enc_prologue
+	.rva	.Lctr_enc_epilogue
+	.rva	.Lctr_enc_info
+
+	.rva	.Lxts_enc_prologue
+	.rva	.Lxts_enc_epilogue
+	.rva	.Lxts_enc_info
+
+	.rva	.Lxts_dec_prologue
+	.rva	.Lxts_dec_epilogue
+	.rva	.Lxts_dec_info
+
+.section	.xdata
+.align	8
 ___
+$code.=<<___ if ($ecb);
+.Lecb_enc_info:
+	.byte	9,0,0,0
+	.rva	se_handler
+	.rva	.Lecb_enc_body,.Lecb_enc_epilogue	# HandlerData[]
+.Lecb_dec_info:
+	.byte	9,0,0,0
+	.rva	se_handler
+	.rva	.Lecb_dec_body,.Lecb_dec_epilogue	# HandlerData[]
+___
+$code.=<<___;
+.Lcbc_dec_info:
+	.byte	9,0,0,0
+	.rva	se_handler
+	.rva	.Lcbc_dec_body,.Lcbc_dec_epilogue	# HandlerData[]
+.Lctr_enc_info:
+	.byte	9,0,0,0
+	.rva	se_handler
+	.rva	.Lctr_enc_body,.Lctr_enc_epilogue	# HandlerData[]
+.Lxts_enc_info:
+	.byte	9,0,0,0
+	.rva	se_handler
+	.rva	.Lxts_enc_body,.Lxts_enc_epilogue	# HandlerData[]
+.Lxts_dec_info:
+	.byte	9,0,0,0
+	.rva	se_handler
+	.rva	.Lxts_dec_body,.Lxts_dec_epilogue	# HandlerData[]
+___
+}
 
 $code =~ s/\`([^\`]*)\`/eval($1)/gem;