AES-NI backport from HEAD. Note that e_aes.c doesn't implement all modes

from HEAD yet, more will be back-ported later.

3 files changed, 5253 insertions, 0 deletions

diff --git a/crypto/aes/Makefile b/crypto/aes/Makefile
index c501a43a8f..3517465bd0 100644
--- a/crypto/aes/Makefile
+++ b/crypto/aes/Makefile
@@ -50,9 +50,13 @@ aes-ia64.s: asm/aes-ia64.S
 
 aes-586.s:	asm/aes-586.pl ../perlasm/x86asm.pl
 	$(PERL) asm/aes-586.pl $(PERLASM_SCHEME) $(CFLAGS) $(PROCESSOR) > $@
+aesni-x86.s:	asm/aesni-x86.pl ../perlasm/x86asm.pl
+	$(PERL) asm/aesni-x86.pl $(PERLASM_SCHEME) $(CFLAGS) $(PROCESSOR) > $@
 
 aes-x86_64.s: asm/aes-x86_64.pl
 	$(PERL) asm/aes-x86_64.pl $(PERLASM_SCHEME) > $@
+aesni-x86_64.s: asm/aesni-x86_64.pl
+	$(PERL) asm/aesni-x86_64.pl $(PERLASM_SCHEME) > $@
 
 aes-sparcv9.s: asm/aes-sparcv9.pl
 	$(PERL) asm/aes-sparcv9.pl $(CFLAGS) > $@
diff --git a/crypto/aes/asm/aesni-x86.pl b/crypto/aes/asm/aesni-x86.pl
new file mode 100644
index 0000000000..b3c8d1f60a
--- /dev/null
+++ b/crypto/aes/asm/aesni-x86.pl
@@ -0,0 +1,2187 @@
+#!/usr/bin/env perl
+
+# ====================================================================
+# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
+# ====================================================================
+#
+# This module implements support for Intel AES-NI extension. In
+# OpenSSL context it's used with Intel engine, but can also be used as
+# drop-in replacement for crypto/aes/asm/aes-586.pl [see below for
+# details].
+#
+# Performance.
+#
+# To start with see corresponding paragraph in aesni-x86_64.pl...
+# Instead of filling table similar to one found there I've chosen to
+# summarize *comparison* results for raw ECB, CTR and CBC benchmarks.
+# The simplified table below represents 32-bit performance relative
+# to 64-bit one in every given point. Ratios vary for different
+# encryption modes, therefore interval values.
+#
+#	16-byte     64-byte     256-byte    1-KB        8-KB
+#	53-67%      67-84%      91-94%      95-98%      97-99.5%
+#
+# Lower ratios for smaller block sizes are perfectly understandable,
+# because function call overhead is higher in 32-bit mode. Largest
+# 8-KB block performance is virtually same: 32-bit code is less than
+# 1% slower for ECB, CBC and CCM, and ~3% slower otherwise.
+
+# January 2011
+#
+# See aesni-x86_64.pl for details. Unlike x86_64 version this module
+# interleaves at most 6 aes[enc|dec] instructions, because there are
+# not enough registers for 8x interleave [which should be optimal for
+# Sandy Bridge]. Actually, performance results for 6x interleave
+# factor presented in aesni-x86_64.pl (except for CTR) are for this
+# module.
+
+# April 2011
+#
+# Add aesni_xts_[en|de]crypt. Westmere spends 1.50 cycles processing
+# one byte out of 8KB with 128-bit key, Sandy Bridge - 1.09.
+
+$PREFIX="aesni";	# if $PREFIX is set to "AES", the script
+			# generates drop-in replacement for
+			# crypto/aes/asm/aes-586.pl:-)
+$inline=1;		# inline _aesni_[en|de]crypt
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+push(@INC,"${dir}","${dir}../../perlasm");
+require "x86asm.pl";
+
+&asm_init($ARGV[0],$0);
+
+if ($PREFIX eq "aesni")	{ $movekey=*movups; }
+else			{ $movekey=*movups; }
+
+$len="eax";
+$rounds="ecx";
+$key="edx";
+$inp="esi";
+$out="edi";
+$rounds_="ebx";	# backup copy for $rounds
+$key_="ebp";	# backup copy for $key
+
+$rndkey0="xmm0";
+$rndkey1="xmm1";
+$inout0="xmm2";
+$inout1="xmm3";
+$inout2="xmm4";
+$inout3="xmm5";	$in1="xmm5";
+$inout4="xmm6";	$in0="xmm6";
+$inout5="xmm7";	$ivec="xmm7";
+
+# AESNI extenstion
+sub aeskeygenassist
+{ my($dst,$src,$imm)=@_;
+    if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
+    {	&data_byte(0x66,0x0f,0x3a,0xdf,0xc0|($1<<3)|$2,$imm);	}
+}
+sub aescommon
+{ my($opcodelet,$dst,$src)=@_;
+    if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
+    {	&data_byte(0x66,0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2);}
+}
+sub aesimc	{ aescommon(0xdb,@_); }
+sub aesenc	{ aescommon(0xdc,@_); }
+sub aesenclast	{ aescommon(0xdd,@_); }
+sub aesdec	{ aescommon(0xde,@_); }
+sub aesdeclast	{ aescommon(0xdf,@_); }
+
+# Inline version of internal aesni_[en|de]crypt1
+{ my $sn;
+sub aesni_inline_generate1
+{ my ($p,$inout,$ivec)=@_; $inout=$inout0 if (!defined($inout));
+  $sn++;
+
+    &$movekey		($rndkey0,&QWP(0,$key));
+    &$movekey		($rndkey1,&QWP(16,$key));
+    &xorps		($ivec,$rndkey0)	if (defined($ivec));
+    &lea		($key,&DWP(32,$key));
+    &xorps		($inout,$ivec)		if (defined($ivec));
+    &xorps		($inout,$rndkey0)	if (!defined($ivec));
+    &set_label("${p}1_loop_$sn");
+	eval"&aes${p}	($inout,$rndkey1)";
+	&dec		($rounds);
+	&$movekey	($rndkey1,&QWP(0,$key));
+	&lea		($key,&DWP(16,$key));
+    &jnz		(&label("${p}1_loop_$sn"));
+    eval"&aes${p}last	($inout,$rndkey1)";
+}}
+
+sub aesni_generate1	# fully unrolled loop
+{ my ($p,$inout)=@_; $inout=$inout0 if (!defined($inout));
+
+    &function_begin_B("_aesni_${p}rypt1");
+	&movups		($rndkey0,&QWP(0,$key));
+	&$movekey	($rndkey1,&QWP(0x10,$key));
+	&xorps		($inout,$rndkey0);
+	&$movekey	($rndkey0,&QWP(0x20,$key));
+	&lea		($key,&DWP(0x30,$key));
+	&cmp		($rounds,11);
+	&jb		(&label("${p}128"));
+	&lea		($key,&DWP(0x20,$key));
+	&je		(&label("${p}192"));
+	&lea		($key,&DWP(0x20,$key));
+	eval"&aes${p}	($inout,$rndkey1)";
+	&$movekey	($rndkey1,&QWP(-0x40,$key));
+	eval"&aes${p}	($inout,$rndkey0)";
+	&$movekey	($rndkey0,&QWP(-0x30,$key));
+    &set_label("${p}192");
+	eval"&aes${p}	($inout,$rndkey1)";
+	&$movekey	($rndkey1,&QWP(-0x20,$key));
+	eval"&aes${p}	($inout,$rndkey0)";
+	&$movekey	($rndkey0,&QWP(-0x10,$key));
+    &set_label("${p}128");
+	eval"&aes${p}	($inout,$rndkey1)";
+	&$movekey	($rndkey1,&QWP(0,$key));
+	eval"&aes${p}	($inout,$rndkey0)";
+	&$movekey	($rndkey0,&QWP(0x10,$key));
+	eval"&aes${p}	($inout,$rndkey1)";
+	&$movekey	($rndkey1,&QWP(0x20,$key));
+	eval"&aes${p}	($inout,$rndkey0)";
+	&$movekey	($rndkey0,&QWP(0x30,$key));
+	eval"&aes${p}	($inout,$rndkey1)";
+	&$movekey	($rndkey1,&QWP(0x40,$key));
+	eval"&aes${p}	($inout,$rndkey0)";
+	&$movekey	($rndkey0,&QWP(0x50,$key));
+	eval"&aes${p}	($inout,$rndkey1)";
+	&$movekey	($rndkey1,&QWP(0x60,$key));
+	eval"&aes${p}	($inout,$rndkey0)";
+	&$movekey	($rndkey0,&QWP(0x70,$key));
+	eval"&aes${p}	($inout,$rndkey1)";
+    eval"&aes${p}last	($inout,$rndkey0)";
+    &ret();
+    &function_end_B("_aesni_${p}rypt1");
+}
+
+# void $PREFIX_encrypt (const void *inp,void *out,const AES_KEY *key);
+&aesni_generate1("enc") if (!$inline);
+&function_begin_B("${PREFIX}_encrypt");
+	&mov	("eax",&wparam(0));
+	&mov	($key,&wparam(2));
+	&movups	($inout0,&QWP(0,"eax"));
+	&mov	($rounds,&DWP(240,$key));
+	&mov	("eax",&wparam(1));
+	if ($inline)
+	{   &aesni_inline_generate1("enc");	}
+	else
+	{   &call	("_aesni_encrypt1");	}
+	&movups	(&QWP(0,"eax"),$inout0);
+	&ret	();
+&function_end_B("${PREFIX}_encrypt");
+
+# void $PREFIX_decrypt (const void *inp,void *out,const AES_KEY *key);
+&aesni_generate1("dec") if(!$inline);
+&function_begin_B("${PREFIX}_decrypt");
+	&mov	("eax",&wparam(0));
+	&mov	($key,&wparam(2));
+	&movups	($inout0,&QWP(0,"eax"));
+	&mov	($rounds,&DWP(240,$key));
+	&mov	("eax",&wparam(1));
+	if ($inline)
+	{   &aesni_inline_generate1("dec");	}
+	else
+	{   &call	("_aesni_decrypt1");	}
+	&movups	(&QWP(0,"eax"),$inout0);
+	&ret	();
+&function_end_B("${PREFIX}_decrypt");
+
+# _aesni_[en|de]cryptN are private interfaces, N denotes interleave
+# factor. Why 3x subroutine were originally used in loops? Even though
+# aes[enc|dec] latency was originally 6, it could be scheduled only
+# every *2nd* cycle. Thus 3x interleave was the one providing optimal
+# utilization, i.e. when subroutine's throughput is virtually same as
+# of non-interleaved subroutine [for number of input blocks up to 3].
+# This is why it makes no sense to implement 2x subroutine.
+# aes[enc|dec] latency in next processor generation is 8, but the
+# instructions can be scheduled every cycle. Optimal interleave for
+# new processor is therefore 8x, but it's unfeasible to accommodate it
+# in XMM registers addreassable in 32-bit mode and therefore 6x is
+# used instead...
+
+sub aesni_generate3
+{ my $p=shift;
+
+    &function_begin_B("_aesni_${p}rypt3");
+	&$movekey	($rndkey0,&QWP(0,$key));
+	&shr		($rounds,1);
+	&$movekey	($rndkey1,&QWP(16,$key));
+	&lea		($key,&DWP(32,$key));
+	&xorps		($inout0,$rndkey0);
+	&pxor		($inout1,$rndkey0);
+	&pxor		($inout2,$rndkey0);
+	&$movekey	($rndkey0,&QWP(0,$key));
+
+    &set_label("${p}3_loop");
+	eval"&aes${p}	($inout0,$rndkey1)";
+	eval"&aes${p}	($inout1,$rndkey1)";
+	&dec		($rounds);
+	eval"&aes${p}	($inout2,$rndkey1)";
+	&$movekey	($rndkey1,&QWP(16,$key));
+	eval"&aes${p}	($inout0,$rndkey0)";
+	eval"&aes${p}	($inout1,$rndkey0)";
+	&lea		($key,&DWP(32,$key));
+	eval"&aes${p}	($inout2,$rndkey0)";
+	&$movekey	($rndkey0,&QWP(0,$key));
+	&jnz		(&label("${p}3_loop"));
+    eval"&aes${p}	($inout0,$rndkey1)";
+    eval"&aes${p}	($inout1,$rndkey1)";
+    eval"&aes${p}	($inout2,$rndkey1)";
+    eval"&aes${p}last	($inout0,$rndkey0)";
+    eval"&aes${p}last	($inout1,$rndkey0)";
+    eval"&aes${p}last	($inout2,$rndkey0)";
+    &ret();
+    &function_end_B("_aesni_${p}rypt3");
+}
+
+# 4x interleave is implemented to improve small block performance,
+# most notably [and naturally] 4 block by ~30%. One can argue that one
+# should have implemented 5x as well, but improvement  would be <20%,
+# so it's not worth it...
+sub aesni_generate4
+{ my $p=shift;
+
+    &function_begin_B("_aesni_${p}rypt4");
+	&$movekey	($rndkey0,&QWP(0,$key));
+	&$movekey	($rndkey1,&QWP(16,$key));
+	&shr		($rounds,1);
+	&lea		($key,&DWP(32,$key));
+	&xorps		($inout0,$rndkey0);
+	&pxor		($inout1,$rndkey0);
+	&pxor		($inout2,$rndkey0);
+	&pxor		($inout3,$rndkey0);
+	&$movekey	($rndkey0,&QWP(0,$key));
+
+    &set_label("${p}4_loop");
+	eval"&aes${p}	($inout0,$rndkey1)";
+	eval"&aes${p}	($inout1,$rndkey1)";
+	&dec		($rounds);
+	eval"&aes${p}	($inout2,$rndkey1)";
+	eval"&aes${p}	($inout3,$rndkey1)";
+	&$movekey	($rndkey1,&QWP(16,$key));
+	eval"&aes${p}	($inout0,$rndkey0)";
+	eval"&aes${p}	($inout1,$rndkey0)";
+	&lea		($key,&DWP(32,$key));
+	eval"&aes${p}	($inout2,$rndkey0)";
+	eval"&aes${p}	($inout3,$rndkey0)";
+	&$movekey	($rndkey0,&QWP(0,$key));
+    &jnz		(&label("${p}4_loop"));
+
+    eval"&aes${p}	($inout0,$rndkey1)";
+    eval"&aes${p}	($inout1,$rndkey1)";
+    eval"&aes${p}	($inout2,$rndkey1)";
+    eval"&aes${p}	($inout3,$rndkey1)";
+    eval"&aes${p}last	($inout0,$rndkey0)";
+    eval"&aes${p}last	($inout1,$rndkey0)";
+    eval"&aes${p}last	($inout2,$rndkey0)";
+    eval"&aes${p}last	($inout3,$rndkey0)";
+    &ret();
+    &function_end_B("_aesni_${p}rypt4");
+}
+
+sub aesni_generate6
+{ my $p=shift;
+
+    &function_begin_B("_aesni_${p}rypt6");
+    &static_label("_aesni_${p}rypt6_enter");
+	&$movekey	($rndkey0,&QWP(0,$key));
+	&shr		($rounds,1);
+	&$movekey	($rndkey1,&QWP(16,$key));
+	&lea		($key,&DWP(32,$key));
+	&xorps		($inout0,$rndkey0);
+	&pxor		($inout1,$rndkey0);	# pxor does better here
+	eval"&aes${p}	($inout0,$rndkey1)";
+	&pxor		($inout2,$rndkey0);
+	eval"&aes${p}	($inout1,$rndkey1)";
+	&pxor		($inout3,$rndkey0);
+	&dec		($rounds);
+	eval"&aes${p}	($inout2,$rndkey1)";
+	&pxor		($inout4,$rndkey0);
+	eval"&aes${p}	($inout3,$rndkey1)";
+	&pxor		($inout5,$rndkey0);
+	eval"&aes${p}	($inout4,$rndkey1)";
+	&$movekey	($rndkey0,&QWP(0,$key));
+	eval"&aes${p}	($inout5,$rndkey1)";
+	&jmp		(&label("_aesni_${p}rypt6_enter"));
+
+    &set_label("${p}6_loop",16);
+	eval"&aes${p}	($inout0,$rndkey1)";
+	eval"&aes${p}	($inout1,$rndkey1)";
+	&dec		($rounds);
+	eval"&aes${p}	($inout2,$rndkey1)";
+	eval"&aes${p}	($inout3,$rndkey1)";
+	eval"&aes${p}	($inout4,$rndkey1)";
+	eval"&aes${p}	($inout5,$rndkey1)";
+    &set_label("_aesni_${p}rypt6_enter",16);
+	&$movekey	($rndkey1,&QWP(16,$key));
+	eval"&aes${p}	($inout0,$rndkey0)";
+	eval"&aes${p}	($inout1,$rndkey0)";
+	&lea		($key,&DWP(32,$key));
+	eval"&aes${p}	($inout2,$rndkey0)";
+	eval"&aes${p}	($inout3,$rndkey0)";
+	eval"&aes${p}	($inout4,$rndkey0)";
+	eval"&aes${p}	($inout5,$rndkey0)";
+	&$movekey	($rndkey0,&QWP(0,$key));
+    &jnz		(&label("${p}6_loop"));
+
+    eval"&aes${p}	($inout0,$rndkey1)";
+    eval"&aes${p}	($inout1,$rndkey1)";
+    eval"&aes${p}	($inout2,$rndkey1)";
+    eval"&aes${p}	($inout3,$rndkey1)";
+    eval"&aes${p}	($inout4,$rndkey1)";
+    eval"&aes${p}	($inout5,$rndkey1)";
+    eval"&aes${p}last	($inout0,$rndkey0)";
+    eval"&aes${p}last	($inout1,$rndkey0)";
+    eval"&aes${p}last	($inout2,$rndkey0)";
+    eval"&aes${p}last	($inout3,$rndkey0)";
+    eval"&aes${p}last	($inout4,$rndkey0)";
+    eval"&aes${p}last	($inout5,$rndkey0)";
+    &ret();
+    &function_end_B("_aesni_${p}rypt6");
+}
+&aesni_generate3("enc") if ($PREFIX eq "aesni");
+&aesni_generate3("dec");
+&aesni_generate4("enc") if ($PREFIX eq "aesni");
+&aesni_generate4("dec");
+&aesni_generate6("enc") if ($PREFIX eq "aesni");
+&aesni_generate6("dec");
+
+if ($PREFIX eq "aesni") {
+######################################################################
+# void aesni_ecb_encrypt (const void *in, void *out,
+#                         size_t length, const AES_KEY *key,
+#                         int enc);
+&function_begin("aesni_ecb_encrypt");
+	&mov	($inp,&wparam(0));
+	&mov	($out,&wparam(1));
+	&mov	($len,&wparam(2));
+	&mov	($key,&wparam(3));
+	&mov	($rounds_,&wparam(4));
+	&and	($len,-16);
+	&jz	(&label("ecb_ret"));
+	&mov	($rounds,&DWP(240,$key));
+	&test	($rounds_,$rounds_);
+	&jz	(&label("ecb_decrypt"));
+
+	&mov	($key_,$key);		# backup $key
+	&mov	($rounds_,$rounds);	# backup $rounds
+	&cmp	($len,0x60);
+	&jb	(&label("ecb_enc_tail"));
+
+	&movdqu	($inout0,&QWP(0,$inp));
+	&movdqu	($inout1,&QWP(0x10,$inp));
+	&movdqu	($inout2,&QWP(0x20,$inp));
+	&movdqu	($inout3,&QWP(0x30,$inp));
+	&movdqu	($inout4,&QWP(0x40,$inp));
+	&movdqu	($inout5,&QWP(0x50,$inp));
+	&lea	($inp,&DWP(0x60,$inp));
+	&sub	($len,0x60);
+	&jmp	(&label("ecb_enc_loop6_enter"));
+
+&set_label("ecb_enc_loop6",16);
+	&movups	(&QWP(0,$out),$inout0);
+	&movdqu	($inout0,&QWP(0,$inp));
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movdqu	($inout1,&QWP(0x10,$inp));
+	&movups	(&QWP(0x20,$out),$inout2);
+	&movdqu	($inout2,&QWP(0x20,$inp));
+	&movups	(&QWP(0x30,$out),$inout3);
+	&movdqu	($inout3,&QWP(0x30,$inp));
+	&movups	(&QWP(0x40,$out),$inout4);
+	&movdqu	($inout4,&QWP(0x40,$inp));
+	&movups	(&QWP(0x50,$out),$inout5);
+	&lea	($out,&DWP(0x60,$out));
+	&movdqu	($inout5,&QWP(0x50,$inp));
+	&lea	($inp,&DWP(0x60,$inp));
+&set_label("ecb_enc_loop6_enter");
+
+	&call	("_aesni_encrypt6");
+
+	&mov	($key,$key_);		# restore $key
+	&mov	($rounds,$rounds_);	# restore $rounds
+	&sub	($len,0x60);
+	&jnc	(&label("ecb_enc_loop6"));
+
+	&movups	(&QWP(0,$out),$inout0);
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movups	(&QWP(0x20,$out),$inout2);
+	&movups	(&QWP(0x30,$out),$inout3);
+	&movups	(&QWP(0x40,$out),$inout4);
+	&movups	(&QWP(0x50,$out),$inout5);
+	&lea	($out,&DWP(0x60,$out));
+	&add	($len,0x60);
+	&jz	(&label("ecb_ret"));
+
+&set_label("ecb_enc_tail");
+	&movups	($inout0,&QWP(0,$inp));
+	&cmp	($len,0x20);
+	&jb	(&label("ecb_enc_one"));
+	&movups	($inout1,&QWP(0x10,$inp));
+	&je	(&label("ecb_enc_two"));
+	&movups	($inout2,&QWP(0x20,$inp));
+	&cmp	($len,0x40);
+	&jb	(&label("ecb_enc_three"));
+	&movups	($inout3,&QWP(0x30,$inp));
+	&je	(&label("ecb_enc_four"));
+	&movups	($inout4,&QWP(0x40,$inp));
+	&xorps	($inout5,$inout5);
+	&call	("_aesni_encrypt6");
+	&movups	(&QWP(0,$out),$inout0);
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movups	(&QWP(0x20,$out),$inout2);
+	&movups	(&QWP(0x30,$out),$inout3);
+	&movups	(&QWP(0x40,$out),$inout4);
+	jmp	(&label("ecb_ret"));
+
+&set_label("ecb_enc_one",16);
+	if ($inline)
+	{   &aesni_inline_generate1("enc");	}
+	else
+	{   &call	("_aesni_encrypt1");	}
+	&movups	(&QWP(0,$out),$inout0);
+	&jmp	(&label("ecb_ret"));
+
+&set_label("ecb_enc_two",16);
+	&xorps	($inout2,$inout2);
+	&call	("_aesni_encrypt3");
+	&movups	(&QWP(0,$out),$inout0);
+	&movups	(&QWP(0x10,$out),$inout1);
+	&jmp	(&label("ecb_ret"));
+
+&set_label("ecb_enc_three",16);
+	&call	("_aesni_encrypt3");
+	&movups	(&QWP(0,$out),$inout0);
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movups	(&QWP(0x20,$out),$inout2);
+	&jmp	(&label("ecb_ret"));
+
+&set_label("ecb_enc_four",16);
+	&call	("_aesni_encrypt4");
+	&movups	(&QWP(0,$out),$inout0);
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movups	(&QWP(0x20,$out),$inout2);
+	&movups	(&QWP(0x30,$out),$inout3);
+	&jmp	(&label("ecb_ret"));
+######################################################################
+&set_label("ecb_decrypt",16);
+	&mov	($key_,$key);		# backup $key
+	&mov	($rounds_,$rounds);	# backup $rounds
+	&cmp	($len,0x60);
+	&jb	(&label("ecb_dec_tail"));
+
+	&movdqu	($inout0,&QWP(0,$inp));
+	&movdqu	($inout1,&QWP(0x10,$inp));
+	&movdqu	($inout2,&QWP(0x20,$inp));
+	&movdqu	($inout3,&QWP(0x30,$inp));
+	&movdqu	($inout4,&QWP(0x40,$inp));
+	&movdqu	($inout5,&QWP(0x50,$inp));
+	&lea	($inp,&DWP(0x60,$inp));
+	&sub	($len,0x60);
+	&jmp	(&label("ecb_dec_loop6_enter"));
+
+&set_label("ecb_dec_loop6",16);
+	&movups	(&QWP(0,$out),$inout0);
+	&movdqu	($inout0,&QWP(0,$inp));
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movdqu	($inout1,&QWP(0x10,$inp));
+	&movups	(&QWP(0x20,$out),$inout2);
+	&movdqu	($inout2,&QWP(0x20,$inp));
+	&movups	(&QWP(0x30,$out),$inout3);
+	&movdqu	($inout3,&QWP(0x30,$inp));
+	&movups	(&QWP(0x40,$out),$inout4);
+	&movdqu	($inout4,&QWP(0x40,$inp));
+	&movups	(&QWP(0x50,$out),$inout5);
+	&lea	($out,&DWP(0x60,$out));
+	&movdqu	($inout5,&QWP(0x50,$inp));
+	&lea	($inp,&DWP(0x60,$inp));
+&set_label("ecb_dec_loop6_enter");
+
+	&call	("_aesni_decrypt6");
+
+	&mov	($key,$key_);		# restore $key
+	&mov	($rounds,$rounds_);	# restore $rounds
+	&sub	($len,0x60);
+	&jnc	(&label("ecb_dec_loop6"));
+
+	&movups	(&QWP(0,$out),$inout0);
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movups	(&QWP(0x20,$out),$inout2);
+	&movups	(&QWP(0x30,$out),$inout3);
+	&movups	(&QWP(0x40,$out),$inout4);
+	&movups	(&QWP(0x50,$out),$inout5);
+	&lea	($out,&DWP(0x60,$out));
+	&add	($len,0x60);
+	&jz	(&label("ecb_ret"));
+
+&set_label("ecb_dec_tail");
+	&movups	($inout0,&QWP(0,$inp));
+	&cmp	($len,0x20);
+	&jb	(&label("ecb_dec_one"));
+	&movups	($inout1,&QWP(0x10,$inp));
+	&je	(&label("ecb_dec_two"));
+	&movups	($inout2,&QWP(0x20,$inp));
+	&cmp	($len,0x40);
+	&jb	(&label("ecb_dec_three"));
+	&movups	($inout3,&QWP(0x30,$inp));
+	&je	(&label("ecb_dec_four"));
+	&movups	($inout4,&QWP(0x40,$inp));
+	&xorps	($inout5,$inout5);
+	&call	("_aesni_decrypt6");
+	&movups	(&QWP(0,$out),$inout0);
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movups	(&QWP(0x20,$out),$inout2);
+	&movups	(&QWP(0x30,$out),$inout3);
+	&movups	(&QWP(0x40,$out),$inout4);
+	&jmp	(&label("ecb_ret"));
+
+&set_label("ecb_dec_one",16);
+	if ($inline)
+	{   &aesni_inline_generate1("dec");	}
+	else
+	{   &call	("_aesni_decrypt1");	}
+	&movups	(&QWP(0,$out),$inout0);
+	&jmp	(&label("ecb_ret"));
+
+&set_label("ecb_dec_two",16);
+	&xorps	($inout2,$inout2);
+	&call	("_aesni_decrypt3");
+	&movups	(&QWP(0,$out),$inout0);
+	&movups	(&QWP(0x10,$out),$inout1);
+	&jmp	(&label("ecb_ret"));
+
+&set_label("ecb_dec_three",16);
+	&call	("_aesni_decrypt3");
+	&movups	(&QWP(0,$out),$inout0);
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movups	(&QWP(0x20,$out),$inout2);
+	&jmp	(&label("ecb_ret"));
+
+&set_label("ecb_dec_four",16);
+	&call	("_aesni_decrypt4");
+	&movups	(&QWP(0,$out),$inout0);
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movups	(&QWP(0x20,$out),$inout2);
+	&movups	(&QWP(0x30,$out),$inout3);
+
+&set_label("ecb_ret");
+&function_end("aesni_ecb_encrypt");
+
+######################################################################
+# void aesni_ccm64_[en|de]crypt_blocks (const void *in, void *out,
+#                         size_t blocks, const AES_KEY *key,
+#                         const char *ivec,char *cmac);
+#
+# Handles only complete blocks, operates on 64-bit counter and
+# does not update *ivec! Nor does it finalize CMAC value
+# (see engine/eng_aesni.c for details)
+#
+{ my $cmac=$inout1;
+&function_begin("aesni_ccm64_encrypt_blocks");
+	&mov	($inp,&wparam(0));
+	&mov	($out,&wparam(1));
+	&mov	($len,&wparam(2));
+	&mov	($key,&wparam(3));
+	&mov	($rounds_,&wparam(4));
+	&mov	($rounds,&wparam(5));
+	&mov	($key_,"esp");
+	&sub	("esp",60);
+	&and	("esp",-16);			# align stack
+	&mov	(&DWP(48,"esp"),$key_);
+
+	&movdqu	($ivec,&QWP(0,$rounds_));	# load ivec
+	&movdqu	($cmac,&QWP(0,$rounds));	# load cmac
+
+	# compose byte-swap control mask for pshufb on stack
+	&mov	(&DWP(0,"esp"),0x0c0d0e0f);
+	&mov	(&DWP(4,"esp"),0x08090a0b);
+	&mov	(&DWP(8,"esp"),0x04050607);
+	&mov	(&DWP(12,"esp"),0x00010203);
+
+	# compose counter increment vector on stack
+	&mov	($rounds,1);
+	&xor	($key_,$key_);
+	&mov	(&DWP(16,"esp"),$rounds);
+	&mov	(&DWP(20,"esp"),$key_);
+	&mov	(&DWP(24,"esp"),$key_);
+	&mov	(&DWP(28,"esp"),$key_);
+
+	&movdqa	($inout3,&QWP(0,"esp"));
+	&pshufb	($ivec,$inout3);		# keep iv in reverse order
+
+	&mov	($rounds,&DWP(240,$key));
+	&mov	($key_,$key);
+	&mov	($rounds_,$rounds);
+	&movdqa	($inout0,$ivec);
+
+&set_label("ccm64_enc_outer");
+	&movups		($in0,&QWP(0,$inp));
+	&pshufb		($inout0,$inout3);
+	&mov		($key,$key_);
+	&mov		($rounds,$rounds_);
+
+	&$movekey	($rndkey0,&QWP(0,$key));
+	&shr		($rounds,1);
+	&$movekey	($rndkey1,&QWP(16,$key));
+	&xorps		($in0,$rndkey0);
+	&lea		($key,&DWP(32,$key));
+	&xorps		($inout0,$rndkey0);
+	&xorps		($cmac,$in0);		# cmac^=inp
+	&$movekey	($rndkey0,&QWP(0,$key));
+
+&set_label("ccm64_enc2_loop");
+	&aesenc		($inout0,$rndkey1);
+	&dec		($rounds);
+	&aesenc		($cmac,$rndkey1);
+	&$movekey	($rndkey1,&QWP(16,$key));
+	&aesenc		($inout0,$rndkey0);
+	&lea		($key,&DWP(32,$key));
+	&aesenc		($cmac,$rndkey0);
+	&$movekey	($rndkey0,&QWP(0,$key));
+	&jnz		(&label("ccm64_enc2_loop"));
+	&aesenc		($inout0,$rndkey1);
+	&aesenc		($cmac,$rndkey1);
+	&aesenclast	($inout0,$rndkey0);
+	&aesenclast	($cmac,$rndkey0);
+
+	&paddq	($ivec,&QWP(16,"esp"));
+	&dec	($len);
+	&lea	($inp,&DWP(16,$inp));
+	&xorps	($in0,$inout0);			# inp^=E(ivec)
+	&movdqa	($inout0,$ivec);
+	&movups	(&QWP(0,$out),$in0);
+	&lea	($out,&DWP(16,$out));
+	&jnz	(&label("ccm64_enc_outer"));
+
+	&mov	("esp",&DWP(48,"esp"));
+	&mov	($out,&wparam(5));
+	&movups	(&QWP(0,$out),$cmac);
+&function_end("aesni_ccm64_encrypt_blocks");
+
+&function_begin("aesni_ccm64_decrypt_blocks");
+	&mov	($inp,&wparam(0));
+	&mov	($out,&wparam(1));
+	&mov	($len,&wparam(2));
+	&mov	($key,&wparam(3));
+	&mov	($rounds_,&wparam(4));
+	&mov	($rounds,&wparam(5));
+	&mov	($key_,"esp");
+	&sub	("esp",60);
+	&and	("esp",-16);			# align stack
+	&mov	(&DWP(48,"esp"),$key_);
+
+	&movdqu	($ivec,&QWP(0,$rounds_));	# load ivec
+	&movdqu	($cmac,&QWP(0,$rounds));	# load cmac
+
+	# compose byte-swap control mask for pshufb on stack
+	&mov	(&DWP(0,"esp"),0x0c0d0e0f);
+	&mov	(&DWP(4,"esp"),0x08090a0b);
+	&mov	(&DWP(8,"esp"),0x04050607);
+	&mov	(&DWP(12,"esp"),0x00010203);
+
+	# compose counter increment vector on stack
+	&mov	($rounds,1);
+	&xor	($key_,$key_);
+	&mov	(&DWP(16,"esp"),$rounds);
+	&mov	(&DWP(20,"esp"),$key_);
+	&mov	(&DWP(24,"esp"),$key_);
+	&mov	(&DWP(28,"esp"),$key_);
+
+	&movdqa	($inout3,&QWP(0,"esp"));	# bswap mask
+	&movdqa	($inout0,$ivec);
+	&pshufb	($ivec,$inout3);		# keep iv in reverse order
+
+	&mov	($rounds,&DWP(240,$key));
+	&mov	($key_,$key);
+	&mov	($rounds_,$rounds);
+
+	if ($inline)
+	{   &aesni_inline_generate1("enc");	}
+	else
+	{   &call	("_aesni_encrypt1");	}
+
+&set_label("ccm64_dec_outer");
+	&paddq	($ivec,&QWP(16,"esp"));
+	&movups	($in0,&QWP(0,$inp));		# load inp
+	&xorps	($in0,$inout0);
+	&movdqa	($inout0,$ivec);
+	&lea	($inp,&QWP(16,$inp));
+	&pshufb	($inout0,$inout3);
+	&mov	($key,$key_);
+	&mov	($rounds,$rounds_);
+	&movups	(&QWP(0,$out),$in0);
+	&lea	($out,&DWP(16,$out));
+
+	&sub	($len,1);
+	&jz	(&label("ccm64_dec_break"));
+
+	&$movekey	($rndkey0,&QWP(0,$key));
+	&shr		($rounds,1);
+	&$movekey	($rndkey1,&QWP(16,$key));
+	&xorps		($in0,$rndkey0);
+	&lea		($key,&DWP(32,$key));
+	&xorps		($inout0,$rndkey0);
+	&xorps		($cmac,$in0);		# cmac^=out
+	&$movekey	($rndkey0,&QWP(0,$key));
+
+&set_label("ccm64_dec2_loop");
+	&aesenc		($inout0,$rndkey1);
+	&dec		($rounds);
+	&aesenc		($cmac,$rndkey1);
+	&$movekey	($rndkey1,&QWP(16,$key));
+	&aesenc		($inout0,$rndkey0);
+	&lea		($key,&DWP(32,$key));
+	&aesenc		($cmac,$rndkey0);
+	&$movekey	($rndkey0,&QWP(0,$key));
+	&jnz		(&label("ccm64_dec2_loop"));
+	&aesenc		($inout0,$rndkey1);
+	&aesenc		($cmac,$rndkey1);
+	&aesenclast	($inout0,$rndkey0);
+	&aesenclast	($cmac,$rndkey0);
+	&jmp	(&label("ccm64_dec_outer"));
+
+&set_label("ccm64_dec_break",16);
+	if ($inline)
+	{   &aesni_inline_generate1("enc",$cmac,$in0);	}
+	else
+	{   &call	("_aesni_encrypt1",$cmac);	}
+
+	&mov	("esp",&DWP(48,"esp"));
+	&mov	($out,&wparam(5));
+	&movups	(&QWP(0,$out),$cmac);
+&function_end("aesni_ccm64_decrypt_blocks");
+}
+
+######################################################################
+# void aesni_ctr32_encrypt_blocks (const void *in, void *out,
+#                         size_t blocks, const AES_KEY *key,
+#                         const char *ivec);
+#
+# Handles only complete blocks, operates on 32-bit counter and
+# does not update *ivec! (see engine/eng_aesni.c for details)
+#
+# stack layout:
+#	0	pshufb mask
+#	16	vector addend: 0,6,6,6
+# 	32	counter-less ivec
+#	48	1st triplet of counter vector
+#	64	2nd triplet of counter vector
+#	80	saved %esp
+
+&function_begin("aesni_ctr32_encrypt_blocks");
+	&mov	($inp,&wparam(0));
+	&mov	($out,&wparam(1));
+	&mov	($len,&wparam(2));
+	&mov	($key,&wparam(3));
+	&mov	($rounds_,&wparam(4));
+	&mov	($key_,"esp");
+	&sub	("esp",88);
+	&and	("esp",-16);			# align stack
+	&mov	(&DWP(80,"esp"),$key_);
+
+	&cmp	($len,1);
+	&je	(&label("ctr32_one_shortcut"));
+
+	&movdqu	($inout5,&QWP(0,$rounds_));	# load ivec
+
+	# compose byte-swap control mask for pshufb on stack
+	&mov	(&DWP(0,"esp"),0x0c0d0e0f);
+	&mov	(&DWP(4,"esp"),0x08090a0b);
+	&mov	(&DWP(8,"esp"),0x04050607);
+	&mov	(&DWP(12,"esp"),0x00010203);
+
+	# compose counter increment vector on stack
+	&mov	($rounds,6);
+	&xor	($key_,$key_);
+	&mov	(&DWP(16,"esp"),$rounds);
+	&mov	(&DWP(20,"esp"),$rounds);
+	&mov	(&DWP(24,"esp"),$rounds);
+	&mov	(&DWP(28,"esp"),$key_);
+
+	&pextrd	($rounds_,$inout5,3);		# pull 32-bit counter
+	&pinsrd	($inout5,$key_,3);		# wipe 32-bit counter
+
+	&mov	($rounds,&DWP(240,$key));	# key->rounds
+
+	# compose 2 vectors of 3x32-bit counters
+	&bswap	($rounds_);
+	&pxor	($rndkey1,$rndkey1);
+	&pxor	($rndkey0,$rndkey0);
+	&movdqa	($inout0,&QWP(0,"esp"));	# load byte-swap mask
+	&pinsrd	($rndkey1,$rounds_,0);
+	&lea	($key_,&DWP(3,$rounds_));
+	&pinsrd	($rndkey0,$key_,0);
+	&inc	($rounds_);
+	&pinsrd	($rndkey1,$rounds_,1);
+	&inc	($key_);
+	&pinsrd	($rndkey0,$key_,1);
+	&inc	($rounds_);
+	&pinsrd	($rndkey1,$rounds_,2);
+	&inc	($key_);
+	&pinsrd	($rndkey0,$key_,2);
+	&movdqa	(&QWP(48,"esp"),$rndkey1);	# save 1st triplet
+	&pshufb	($rndkey1,$inout0);		# byte swap
+	&movdqa	(&QWP(64,"esp"),$rndkey0);	# save 2nd triplet
+	&pshufb	($rndkey0,$inout0);		# byte swap
+
+	&pshufd	($inout0,$rndkey1,3<<6);	# place counter to upper dword
+	&pshufd	($inout1,$rndkey1,2<<6);
+	&cmp	($len,6);
+	&jb	(&label("ctr32_tail"));
+	&movdqa	(&QWP(32,"esp"),$inout5);	# save counter-less ivec
+	&shr	($rounds,1);
+	&mov	($key_,$key);			# backup $key
+	&mov	($rounds_,$rounds);		# backup $rounds
+	&sub	($len,6);
+	&jmp	(&label("ctr32_loop6"));
+
+&set_label("ctr32_loop6",16);
+	&pshufd	($inout2,$rndkey1,1<<6);
+	&movdqa	($rndkey1,&QWP(32,"esp"));	# pull counter-less ivec
+	&pshufd	($inout3,$rndkey0,3<<6);
+	&por	($inout0,$rndkey1);		# merge counter-less ivec
+	&pshufd	($inout4,$rndkey0,2<<6);
+	&por	($inout1,$rndkey1);
+	&pshufd	($inout5,$rndkey0,1<<6);
+	&por	($inout2,$rndkey1);
+	&por	($inout3,$rndkey1);
+	&por	($inout4,$rndkey1);
+	&por	($inout5,$rndkey1);
+
+	# inlining _aesni_encrypt6's prologue gives ~4% improvement...
+	&$movekey	($rndkey0,&QWP(0,$key_));
+	&$movekey	($rndkey1,&QWP(16,$key_));
+	&lea		($key,&DWP(32,$key_));
+	&dec		($rounds);
+	&pxor		($inout0,$rndkey0);
+	&pxor		($inout1,$rndkey0);
+	&aesenc		($inout0,$rndkey1);
+	&pxor		($inout2,$rndkey0);
+	&aesenc		($inout1,$rndkey1);
+	&pxor		($inout3,$rndkey0);
+	&aesenc		($inout2,$rndkey1);
+	&pxor		($inout4,$rndkey0);
+	&aesenc		($inout3,$rndkey1);
+	&pxor		($inout5,$rndkey0);
+	&aesenc		($inout4,$rndkey1);
+	&$movekey	($rndkey0,&QWP(0,$key));
+	&aesenc		($inout5,$rndkey1);
+
+	&call		(&label("_aesni_encrypt6_enter"));
+
+	&movups	($rndkey1,&QWP(0,$inp));
+	&movups	($rndkey0,&QWP(0x10,$inp));
+	&xorps	($inout0,$rndkey1);
+	&movups	($rndkey1,&QWP(0x20,$inp));
+	&xorps	($inout1,$rndkey0);
+	&movups	(&QWP(0,$out),$inout0);
+	&movdqa	($rndkey0,&QWP(16,"esp"));	# load increment
+	&xorps	($inout2,$rndkey1);
+	&movdqa	($rndkey1,&QWP(48,"esp"));	# load 1st triplet
+	&movups	(&QWP(0x10,$out),$inout1);
+	&movups	(&QWP(0x20,$out),$inout2);
+
+	&paddd	($rndkey1,$rndkey0);		# 1st triplet increment
+	&paddd	($rndkey0,&QWP(64,"esp"));	# 2nd triplet increment
+	&movdqa	($inout0,&QWP(0,"esp"));	# load byte swap mask
+
+	&movups	($inout1,&QWP(0x30,$inp));
+	&movups	($inout2,&QWP(0x40,$inp));
+	&xorps	($inout3,$inout1);
+	&movups	($inout1,&QWP(0x50,$inp));
+	&lea	($inp,&DWP(0x60,$inp));
+	&movdqa	(&QWP(48,"esp"),$rndkey1);	# save 1st triplet
+	&pshufb	($rndkey1,$inout0);		# byte swap
+	&xorps	($inout4,$inout2);
+	&movups	(&QWP(0x30,$out),$inout3);
+	&xorps	($inout5,$inout1);
+	&movdqa	(&QWP(64,"esp"),$rndkey0);	# save 2nd triplet
+	&pshufb	($rndkey0,$inout0);		# byte swap
+	&movups	(&QWP(0x40,$out),$inout4);
+	&pshufd	($inout0,$rndkey1,3<<6);
+	&movups	(&QWP(0x50,$out),$inout5);
+	&lea	($out,&DWP(0x60,$out));
+
+	&mov	($rounds,$rounds_);
+	&pshufd	($inout1,$rndkey1,2<<6);
+	&sub	($len,6);
+	&jnc	(&label("ctr32_loop6"));
+
+	&add	($len,6);
+	&jz	(&label("ctr32_ret"));
+	&mov	($key,$key_);
+	&lea	($rounds,&DWP(1,"",$rounds,2));	# restore $rounds
+	&movdqa	($inout5,&QWP(32,"esp"));	# pull count-less ivec
+
+&set_label("ctr32_tail");
+	&por	($inout0,$inout5);
+	&cmp	($len,2);
+	&jb	(&label("ctr32_one"));
+
+	&pshufd	($inout2,$rndkey1,1<<6);
+	&por	($inout1,$inout5);
+	&je	(&label("ctr32_two"));
+
+	&pshufd	($inout3,$rndkey0,3<<6);
+	&por	($inout2,$inout5);
+	&cmp	($len,4);
+	&jb	(&label("ctr32_three"));
+
+	&pshufd	($inout4,$rndkey0,2<<6);
+	&por	($inout3,$inout5);
+	&je	(&label("ctr32_four"));
+
+	&por	($inout4,$inout5);
+	&call	("_aesni_encrypt6");