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

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

1 files changed, 3062 insertions, 0 deletions

diff --git a/crypto/aes/asm/aesni-x86_64.pl b/crypto/aes/asm/aesni-x86_64.pl
new file mode 100644
index 0000000000..ae0ad7f809
--- /dev/null
+++ b/crypto/aes/asm/aesni-x86_64.pl
@@ -0,0 +1,3062 @@
+#!/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-x86_64.pl [see below for
+# details].
+#
+# Performance.
+#
+# Given aes(enc|dec) instructions' latency asymptotic performance for
+# non-parallelizable modes such as CBC encrypt is 3.75 cycles per byte
+# processed with 128-bit key. And given their throughput asymptotic
+# performance for parallelizable modes is 1.25 cycles per byte. Being
+# asymptotic limit it's not something you commonly achieve in reality,
+# but how close does one get? Below are results collected for
+# different modes and block sized. Pairs of numbers are for en-/
+# decryption.
+#
+#	16-byte     64-byte     256-byte    1-KB        8-KB
+# ECB	4.25/4.25   1.38/1.38   1.28/1.28   1.26/1.26	1.26/1.26
+# CTR	5.42/5.42   1.92/1.92   1.44/1.44   1.28/1.28   1.26/1.26
+# CBC	4.38/4.43   4.15/1.43   4.07/1.32   4.07/1.29   4.06/1.28
+# CCM	5.66/9.42   4.42/5.41   4.16/4.40   4.09/4.15   4.06/4.07   
+# OFB	5.42/5.42   4.64/4.64   4.44/4.44   4.39/4.39   4.38/4.38
+# CFB	5.73/5.85   5.56/5.62   5.48/5.56   5.47/5.55   5.47/5.55
+#
+# ECB, CTR, CBC and CCM results are free from EVP overhead. This means
+# that otherwise used 'openssl speed -evp aes-128-??? -engine aesni
+# [-decrypt]' will exhibit 10-15% worse results for smaller blocks.
+# The results were collected with specially crafted speed.c benchmark
+# in order to compare them with results reported in "Intel Advanced
+# Encryption Standard (AES) New Instruction Set" White Paper Revision
+# 3.0 dated May 2010. All above results are consistently better. This
+# module also provides better performance for block sizes smaller than
+# 128 bytes in points *not* represented in the above table.
+#
+# Looking at the results for 8-KB buffer.
+#
+# CFB and OFB results are far from the limit, because implementation
+# uses "generic" CRYPTO_[c|o]fb128_encrypt interfaces relying on
+# single-block aesni_encrypt, which is not the most optimal way to go.
+# CBC encrypt result is unexpectedly high and there is no documented
+# explanation for it. Seemingly there is a small penalty for feeding
+# the result back to AES unit the way it's done in CBC mode. There is
+# nothing one can do and the result appears optimal. CCM result is
+# identical to CBC, because CBC-MAC is essentially CBC encrypt without
+# saving output. CCM CTR "stays invisible," because it's neatly
+# interleaved wih CBC-MAC. This provides ~30% improvement over
+# "straghtforward" CCM implementation with CTR and CBC-MAC performed
+# disjointly. Parallelizable modes practically achieve the theoretical
+# limit.
+#
+# Looking at how results vary with buffer size.
+#
+# Curves are practically saturated at 1-KB buffer size. In most cases
+# "256-byte" performance is >95%, and "64-byte" is ~90% of "8-KB" one.
+# CTR curve doesn't follow this pattern and is "slowest" changing one
+# with "256-byte" result being 87% of "8-KB." This is because overhead
+# in CTR mode is most computationally intensive. Small-block CCM
+# decrypt is slower than encrypt, because first CTR and last CBC-MAC
+# iterations can't be interleaved.
+#
+# Results for 192- and 256-bit keys.
+#
+# EVP-free results were observed to scale perfectly with number of
+# rounds for larger block sizes, i.e. 192-bit result being 10/12 times
+# lower and 256-bit one - 10/14. Well, in CBC encrypt case differences
+# are a tad smaller, because the above mentioned penalty biases all
+# results by same constant value. In similar way function call
+# overhead affects small-block performance, as well as OFB and CFB
+# results. Differences are not large, most common coefficients are
+# 10/11.7 and 10/13.4 (as opposite to 10/12.0 and 10/14.0), but one
+# observe even 10/11.2 and 10/12.4 (CTR, OFB, CFB)...
+
+# January 2011
+#
+# While Westmere processor features 6 cycles latency for aes[enc|dec]
+# instructions, which can be scheduled every second cycle, Sandy
+# Bridge spends 8 cycles per instruction, but it can schedule them
+# every cycle. This means that code targeting Westmere would perform
+# suboptimally on Sandy Bridge. Therefore this update.
+#
+# In addition, non-parallelizable CBC encrypt (as well as CCM) is
+# optimized. Relative improvement might appear modest, 8% on Westmere,
+# but in absolute terms it's 3.77 cycles per byte encrypted with
+# 128-bit key on Westmere, and 5.07 - on Sandy Bridge. These numbers
+# should be compared to asymptotic limits of 3.75 for Westmere and
+# 5.00 for Sandy Bridge. Actually, the fact that they get this close
+# to asymptotic limits is quite amazing. Indeed, the limit is
+# calculated as latency times number of rounds, 10 for 128-bit key,
+# and divided by 16, the number of bytes in block, or in other words
+# it accounts *solely* for aesenc instructions. But there are extra
+# instructions, and numbers so close to the asymptotic limits mean
+# that it's as if it takes as little as *one* additional cycle to
+# execute all of them. How is it possible? It is possible thanks to
+# out-of-order execution logic, which manages to overlap post-
+# processing of previous block, things like saving the output, with
+# actual encryption of current block, as well as pre-processing of
+# current block, things like fetching input and xor-ing it with
+# 0-round element of the key schedule, with actual encryption of
+# previous block. Keep this in mind...
+#
+# For parallelizable modes, such as ECB, CBC decrypt, CTR, higher
+# performance is achieved by interleaving instructions working on
+# independent blocks. In which case asymptotic limit for such modes
+# can be obtained by dividing above mentioned numbers by AES
+# instructions' interleave factor. Westmere can execute at most 3 
+# instructions at a time, meaning that optimal interleave factor is 3,
+# and that's where the "magic" number of 1.25 come from. "Optimal
+# interleave factor" means that increase of interleave factor does
+# not improve performance. The formula has proven to reflect reality
+# pretty well on Westmere... Sandy Bridge on the other hand can
+# execute up to 8 AES instructions at a time, so how does varying
+# interleave factor affect the performance? Here is table for ECB
+# (numbers are cycles per byte processed with 128-bit key):
+#
+# instruction interleave factor		3x	6x	8x
+# theoretical asymptotic limit		1.67	0.83	0.625
+# measured performance for 8KB block	1.05	0.86	0.84
+#
+# "as if" interleave factor		4.7x	5.8x	6.0x
+#
+# Further data for other parallelizable modes:
+#
+# CBC decrypt				1.16	0.93	0.93
+# CTR					1.14	0.91	n/a
+#
+# Well, given 3x column it's probably inappropriate to call the limit
+# asymptotic, if it can be surpassed, isn't it? What happens there?
+# Rewind to CBC paragraph for the answer. Yes, out-of-order execution
+# magic is responsible for this. Processor overlaps not only the
+# additional instructions with AES ones, but even AES instuctions
+# processing adjacent triplets of independent blocks. In the 6x case
+# additional instructions  still claim disproportionally small amount
+# of additional cycles, but in 8x case number of instructions must be
+# a tad too high for out-of-order logic to cope with, and AES unit
+# remains underutilized... As you can see 8x interleave is hardly
+# justifiable, so there no need to feel bad that 32-bit aesni-x86.pl
+# utilizies 6x interleave because of limited register bank capacity.
+#
+# Higher interleave factors do have negative impact on Westmere
+# performance. While for ECB mode it's negligible ~1.5%, other
+# parallelizables perform ~5% worse, which is outweighed by ~25%
+# improvement on Sandy Bridge. To balance regression on Westmere
+# CTR mode was implemented with 6x aesenc interleave factor.
+
+# April 2011
+#
+# Add aesni_xts_[en|de]crypt. Westmere spends 1.33 cycles processing
+# one byte out of 8KB with 128-bit key, Sandy Bridge - 0.97. Just like
+# in CTR mode AES instruction interleave factor was chosen to be 6x.
+
+$PREFIX="aesni";	# if $PREFIX is set to "AES", the script
+			# generates drop-in replacement for
+			# crypto/aes/asm/aes-x86_64.pl:-)
+
+$flavour = shift;
+$output  = shift;
+if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
+
+$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
+( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
+die "can't locate x86_64-xlate.pl";
+
+open STDOUT,"| $^X $xlate $flavour $output";
+
+$movkey = $PREFIX eq "aesni" ? "movups" : "movups";
+@_4args=$win64?	("%rcx","%rdx","%r8", "%r9") :	# Win64 order
+		("%rdi","%rsi","%rdx","%rcx");	# Unix order
+
+$code=".text\n";
+
+$rounds="%eax";	# input to and changed by aesni_[en|de]cryptN !!!
+# this is natural Unix argument order for public $PREFIX_[ecb|cbc]_encrypt ...
+$inp="%rdi";
+$out="%rsi";
+$len="%rdx";
+$key="%rcx";	# input to and changed by aesni_[en|de]cryptN !!!
+$ivp="%r8";	# cbc, ctr, ...
+
+$rnds_="%r10d";	# backup copy for $rounds
+$key_="%r11";	# backup copy for $key
+
+# %xmm register layout
+$rndkey0="%xmm0";	$rndkey1="%xmm1";
+$inout0="%xmm2";	$inout1="%xmm3";
+$inout2="%xmm4";	$inout3="%xmm5";
+$inout4="%xmm6";	$inout5="%xmm7";
+$inout6="%xmm8";	$inout7="%xmm9";
+
+$in2="%xmm6";		$in1="%xmm7";	# used in CBC decrypt, CTR, ...
+$in0="%xmm8";		$iv="%xmm9";
+
+# Inline version of internal aesni_[en|de]crypt1.
+#
+# Why folded loop? Because aes[enc|dec] is slow enough to accommodate
+# cycles which take care of loop variables...
+{ my $sn;
+sub aesni_generate1 {
+my ($p,$key,$rounds,$inout,$ivec)=@_;	$inout=$inout0 if (!defined($inout));
+++$sn;
+$code.=<<___;
+	$movkey	($key),$rndkey0
+	$movkey	16($key),$rndkey1
+___
+$code.=<<___ if (defined($ivec));
+	xorps	$rndkey0,$ivec
+	lea	32($key),$key
+	xorps	$ivec,$inout
+___
+$code.=<<___ if (!defined($ivec));
+	lea	32($key),$key
+	xorps	$rndkey0,$inout
+___
+$code.=<<___;
+.Loop_${p}1_$sn:
+	aes${p}	$rndkey1,$inout
+	dec	$rounds
+	$movkey	($key),$rndkey1
+	lea	16($key),$key
+	jnz	.Loop_${p}1_$sn	# loop body is 16 bytes
+	aes${p}last	$rndkey1,$inout
+___
+}}
+# void $PREFIX_[en|de]crypt (const void *inp,void *out,const AES_KEY *key);
+#
+{ my ($inp,$out,$key) = @_4args;
+
+$code.=<<___;
+.globl	${PREFIX}_encrypt
+.type	${PREFIX}_encrypt,\@abi-omnipotent
+.align	16
+${PREFIX}_encrypt:
+	movups	($inp),$inout0		# load input
+	mov	240($key),$rounds	# key->rounds
+___
+	&aesni_generate1("enc",$key,$rounds);
+$code.=<<___;
+	movups	$inout0,($out)		# output
+	ret
+.size	${PREFIX}_encrypt,.-${PREFIX}_encrypt
+
+.globl	${PREFIX}_decrypt
+.type	${PREFIX}_decrypt,\@abi-omnipotent
+.align	16
+${PREFIX}_decrypt:
+	movups	($inp),$inout0		# load input
+	mov	240($key),$rounds	# key->rounds
+___
+	&aesni_generate1("dec",$key,$rounds);
+$code.=<<___;
+	movups	$inout0,($out)		# output
+	ret
+.size	${PREFIX}_decrypt, .-${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...
+sub aesni_generate3 {
+my $dir=shift;
+# As already mentioned it takes in $key and $rounds, which are *not*
+# preserved. $inout[0-2] is cipher/clear text...
+$code.=<<___;
+.type	_aesni_${dir}rypt3,\@abi-omnipotent
+.align	16
+_aesni_${dir}rypt3:
+	$movkey	($key),$rndkey0
+	shr	\$1,$rounds
+	$movkey	16($key),$rndkey1
+	lea	32($key),$key
+	xorps	$rndkey0,$inout0
+	xorps	$rndkey0,$inout1
+	xorps	$rndkey0,$inout2
+	$movkey		($key),$rndkey0
+
+.L${dir}_loop3:
+	aes${dir}	$rndkey1,$inout0
+	aes${dir}	$rndkey1,$inout1
+	dec		$rounds
+	aes${dir}	$rndkey1,$inout2
+	$movkey		16($key),$rndkey1
+	aes${dir}	$rndkey0,$inout0
+	aes${dir}	$rndkey0,$inout1
+	lea		32($key),$key
+	aes${dir}	$rndkey0,$inout2
+	$movkey		($key),$rndkey0
+	jnz		.L${dir}_loop3
+
+	aes${dir}	$rndkey1,$inout0
+	aes${dir}	$rndkey1,$inout1
+	aes${dir}	$rndkey1,$inout2
+	aes${dir}last	$rndkey0,$inout0
+	aes${dir}last	$rndkey0,$inout1
+	aes${dir}last	$rndkey0,$inout2
+	ret
+.size	_aesni_${dir}rypt3,.-_aesni_${dir}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 $dir=shift;
+# As already mentioned it takes in $key and $rounds, which are *not*
+# preserved. $inout[0-3] is cipher/clear text...
+$code.=<<___;
+.type	_aesni_${dir}rypt4,\@abi-omnipotent
+.align	16
+_aesni_${dir}rypt4:
+	$movkey	($key),$rndkey0
+	shr	\$1,$rounds
+	$movkey	16($key),$rndkey1
+	lea	32($key),$key
+	xorps	$rndkey0,$inout0
+	xorps	$rndkey0,$inout1
+	xorps	$rndkey0,$inout2
+	xorps	$rndkey0,$inout3
+	$movkey	($key),$rndkey0
+
+.L${dir}_loop4:
+	aes${dir}	$rndkey1,$inout0
+	aes${dir}	$rndkey1,$inout1
+	dec		$rounds
+	aes${dir}	$rndkey1,$inout2
+	aes${dir}	$rndkey1,$inout3
+	$movkey		16($key),$rndkey1
+	aes${dir}	$rndkey0,$inout0
+	aes${dir}	$rndkey0,$inout1
+	lea		32($key),$key
+	aes${dir}	$rndkey0,$inout2
+	aes${dir}	$rndkey0,$inout3
+	$movkey		($key),$rndkey0
+	jnz		.L${dir}_loop4
+
+	aes${dir}	$rndkey1,$inout0
+	aes${dir}	$rndkey1,$inout1
+	aes${dir}	$rndkey1,$inout2
+	aes${dir}	$rndkey1,$inout3
+	aes${dir}last	$rndkey0,$inout0
+	aes${dir}last	$rndkey0,$inout1
+	aes${dir}last	$rndkey0,$inout2
+	aes${dir}last	$rndkey0,$inout3
+	ret
+.size	_aesni_${dir}rypt4,.-_aesni_${dir}rypt4
+___
+}
+sub aesni_generate6 {
+my $dir=shift;
+# As already mentioned it takes in $key and $rounds, which are *not*
+# preserved. $inout[0-5] is cipher/clear text...
+$code.=<<___;
+.type	_aesni_${dir}rypt6,\@abi-omnipotent
+.align	16
+_aesni_${dir}rypt6:
+	$movkey		($key),$rndkey0
+	shr		\$1,$rounds
+	$movkey		16($key),$rndkey1
+	lea		32($key),$key
+	xorps		$rndkey0,$inout0
+	pxor		$rndkey0,$inout1
+	aes${dir}	$rndkey1,$inout0
+	pxor		$rndkey0,$inout2
+	aes${dir}	$rndkey1,$inout1
+	pxor		$rndkey0,$inout3
+	aes${dir}	$rndkey1,$inout2
+	pxor		$rndkey0,$inout4
+	aes${dir}	$rndkey1,$inout3
+	pxor		$rndkey0,$inout5
+	dec		$rounds
+	aes${dir}	$rndkey1,$inout4
+	$movkey		($key),$rndkey0
+	aes${dir}	$rndkey1,$inout5
+	jmp		.L${dir}_loop6_enter
+.align	16
+.L${dir}_loop6:
+	aes${dir}	$rndkey1,$inout0
+	aes${dir}	$rndkey1,$inout1
+	dec		$rounds
+	aes${dir}	$rndkey1,$inout2
+	aes${dir}	$rndkey1,$inout3
+	aes${dir}	$rndkey1,$inout4
+	aes${dir}	$rndkey1,$inout5
+.L${dir}_loop6_enter:				# happens to be 16-byte aligned
+	$movkey		16($key),$rndkey1
+	aes${dir}	$rndkey0,$inout0
+	aes${dir}	$rndkey0,$inout1
+	lea		32($key),$key
+	aes${dir}	$rndkey0,$inout2
+	aes${dir}	$rndkey0,$inout3
+	aes${dir}	$rndkey0,$inout4
+	aes${dir}	$rndkey0,$inout5
+	$movkey		($key),$rndkey0
+	jnz		.L${dir}_loop6
+
+	aes${dir}	$rndkey1,$inout0
+	aes${dir}	$rndkey1,$inout1
+	aes${dir}	$rndkey1,$inout2
+	aes${dir}	$rndkey1,$inout3
+	aes${dir}	$rndkey1,$inout4
+	aes${dir}	$rndkey1,$inout5
+	aes${dir}last	$rndkey0,$inout0
+	aes${dir}last	$rndkey0,$inout1
+	aes${dir}last	$rndkey0,$inout2
+	aes${dir}last	$rndkey0,$inout3
+	aes${dir}last	$rndkey0,$inout4
+	aes${dir}last	$rndkey0,$inout5
+	ret
+.size	_aesni_${dir}rypt6,.-_aesni_${dir}rypt6
+___
+}
+sub aesni_generate8 {
+my $dir=shift;
+# As already mentioned it takes in $key and $rounds, which are *not*
+# preserved. $inout[0-7] is cipher/clear text...
+$code.=<<___;
+.type	_aesni_${dir}rypt8,\@abi-omnipotent
+.align	16
+_aesni_${dir}rypt8:
+	$movkey		($key),$rndkey0
+	shr		\$1,$rounds
+	$movkey		16($key),$rndkey1
+	lea		32($key),$key
+	xorps		$rndkey0,$inout0
+	xorps		$rndkey0,$inout1
+	aes${dir}	$rndkey1,$inout0
+	pxor		$rndkey0,$inout2
+	aes${dir}	$rndkey1,$inout1
+	pxor		$rndkey0,$inout3
+	aes${dir}	$rndkey1,$inout2
+	pxor		$rndkey0,$inout4
+	aes${dir}	$rndkey1,$inout3
+	pxor		$rndkey0,$inout5
+	dec		$rounds
+	aes${dir}	$rndkey1,$inout4
+	pxor		$rndkey0,$inout6
+	aes${dir}	$rndkey1,$inout5
+	pxor		$rndkey0,$inout7
+	$movkey		($key),$rndkey0
+	aes${dir}	$rndkey1,$inout6
+	aes${dir}	$rndkey1,$inout7
+	$movkey		16($key),$rndkey1
+	jmp		.L${dir}_loop8_enter
+.align	16
+.L${dir}_loop8:
+	aes${dir}	$rndkey1,$inout0
+	aes${dir}	$rndkey1,$inout1
+	dec		$rounds
+	aes${dir}	$rndkey1,$inout2
+	aes${dir}	$rndkey1,$inout3
+	aes${dir}	$rndkey1,$inout4
+	aes${dir}	$rndkey1,$inout5
+	aes${dir}	$rndkey1,$inout6
+	aes${dir}	$rndkey1,$inout7
+	$movkey		16($key),$rndkey1
+.L${dir}_loop8_enter:				# happens to be 16-byte aligned
+	aes${dir}	$rndkey0,$inout0
+	aes${dir}	$rndkey0,$inout1
+	lea		32($key),$key
+	aes${dir}	$rndkey0,$inout2
+	aes${dir}	$rndkey0,$inout3
+	aes${dir}	$rndkey0,$inout4
+	aes${dir}	$rndkey0,$inout5
+	aes${dir}	$rndkey0,$inout6
+	aes${dir}	$rndkey0,$inout7
+	$movkey		($key),$rndkey0
+	jnz		.L${dir}_loop8
+
+	aes${dir}	$rndkey1,$inout0
+	aes${dir}	$rndkey1,$inout1
+	aes${dir}	$rndkey1,$inout2
+	aes${dir}	$rndkey1,$inout3
+	aes${dir}	$rndkey1,$inout4
+	aes${dir}	$rndkey1,$inout5
+	aes${dir}	$rndkey1,$inout6
+	aes${dir}	$rndkey1,$inout7
+	aes${dir}last	$rndkey0,$inout0
+	aes${dir}last	$rndkey0,$inout1
+	aes${dir}last	$rndkey0,$inout2
+	aes${dir}last	$rndkey0,$inout3
+	aes${dir}last	$rndkey0,$inout4
+	aes${dir}last	$rndkey0,$inout5
+	aes${dir}last	$rndkey0,$inout6
+	aes${dir}last	$rndkey0,$inout7
+	ret
+.size	_aesni_${dir}rypt8,.-_aesni_${dir}rypt8
+___
+}
+&aesni_generate3("enc") if ($PREFIX eq "aesni");
+&aesni_generate3("dec");
+&aesni_generate4("enc") if ($PREFIX eq "aesni");
+&aesni_generate4("dec");
+&aesni_generate6("enc") if ($PREFIX eq "aesni");
+&aesni_generate6("dec");
+&aesni_generate8("enc") if ($PREFIX eq "aesni");
+&aesni_generate8("dec");
+
+if ($PREFIX eq "aesni") {
+########################################################################
+# void aesni_ecb_encrypt (const void *in, void *out,
+#			  size_t length, const AES_KEY *key,
+#			  int enc);
+$code.=<<___;
+.globl	aesni_ecb_encrypt
+.type	aesni_ecb_encrypt,\@function,5
+.align	16
+aesni_ecb_encrypt:
+	and	\$-16,$len
+	jz	.Lecb_ret
+
+	mov	240($key),$rounds	# key->rounds
+	$movkey	($key),$rndkey0
+	mov	$key,$key_		# backup $key
+	mov	$rounds,$rnds_		# backup $rounds
+	test	%r8d,%r8d		# 5th argument
+	jz	.Lecb_decrypt
+#--------------------------- ECB ENCRYPT ------------------------------#
+	cmp	\$0x80,$len
+	jb	.Lecb_enc_tail
+
+	movdqu	($inp),$inout0
+	movdqu	0x10($inp),$inout1
+	movdqu	0x20($inp),$inout2
+	movdqu	0x30($inp),$inout3
+	movdqu	0x40($inp),$inout4
+	movdqu	0x50($inp),$inout5
+	movdqu	0x60($inp),$inout6
+	movdqu	0x70($inp),$inout7
+	lea	0x80($inp),$inp
+	sub	\$0x80,$len
+	jmp	.Lecb_enc_loop8_enter
+.align 16
+.Lecb_enc_loop8:
+	movups	$inout0,($out)
+	mov	$key_,$key		# restore $key
+	movdqu	($inp),$inout0
+	mov	$rnds_,$rounds		# restore $rounds
+	movups	$inout1,0x10($out)
+	movdqu	0x10($inp),$inout1
+	movups	$inout2,0x20($out)
+	movdqu	0x20($inp),$inout2
+	movups	$inout3,0x30($out)
+	movdqu	0x30($inp),$inout3
+	movups	$inout4,0x40($out)
+	movdqu	0x40($inp),$inout4
+	movups	$inout5,0x50($out)
+	movdqu	0x50($inp),$inout5
+	movups	$inout6,0x60($out)
+	movdqu	0x60($inp),$inout6
+	movups	$inout7,0x70($out)
+	lea	0x80($out),$out
+	movdqu	0x70($inp),$inout7
+	lea	0x80($inp),$inp
+.Lecb_enc_loop8_enter:
+
+	call	_aesni_encrypt8
+
+	sub	\$0x80,$len
+	jnc	.Lecb_enc_loop8
+
+	movups	$inout0,($out)
+	mov	$key_,$key		# restore $key
+	movups	$inout1,0x10($out)
+	mov	$rnds_,$rounds		# restore $rounds
+	movups	$inout2,0x20($out)
+	movups	$inout3,0x30($out)
+	movups	$inout4,0x40($out)
+	movups	$inout5,0x50($out)
+	movups	$inout6,0x60($out)
+	movups	$inout7,0x70($out)
+	lea	0x80($out),$out
+	add	\$0x80,$len
+	jz	.Lecb_ret
+
+.Lecb_enc_tail:
+	movups	($inp),$inout0
+	cmp	\$0x20,$len
+	jb	.Lecb_enc_one
+	movups	0x10($inp),$inout1
+	je	.Lecb_enc_two
+	movups	0x20($inp),$inout2
+	cmp	\$0x40,$len
+	jb	.Lecb_enc_three
+	movups	0x30($inp),$inout3
+	je	.Lecb_enc_four
+	movups	0x40($inp),$inout4
+	cmp	\$0x60,$len
+	jb	.Lecb_enc_five
+	movups	0x50($inp),$inout5
+	je	.Lecb_enc_six
+	movdqu	0x60($inp),$inout6
+	call	_aesni_encrypt8
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	movups	$inout2,0x20($out)
+	movups	$inout3,0x30($out)
+	movups	$inout4,0x40($out)
+	movups	$inout5,0x50($out)
+	movups	$inout6,0x60($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_enc_one:
+___
+	&aesni_generate1("enc",$key,$rounds);
+$code.=<<___;
+	movups	$inout0,($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_enc_two:
+	xorps	$inout2,$inout2
+	call	_aesni_encrypt3
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_enc_three:
+	call	_aesni_encrypt3
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	movups	$inout2,0x20($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_enc_four:
+	call	_aesni_encrypt4
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	movups	$inout2,0x20($out)
+	movups	$inout3,0x30($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_enc_five:
+	xorps	$inout5,$inout5
+	call	_aesni_encrypt6
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	movups	$inout2,0x20($out)
+	movups	$inout3,0x30($out)
+	movups	$inout4,0x40($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_enc_six:
+	call	_aesni_encrypt6
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	movups	$inout2,0x20($out)
+	movups	$inout3,0x30($out)
+	movups	$inout4,0x40($out)
+	movups	$inout5,0x50($out)
+	jmp	.Lecb_ret
+#--------------------------- ECB DECRYPT ------------------------------#
+.align	16
+.Lecb_decrypt:
+	cmp	\$0x80,$len
+	jb	.Lecb_dec_tail
+
+	movdqu	($inp),$inout0
+	movdqu	0x10($inp),$inout1
+	movdqu	0x20($inp),$inout2
+	movdqu	0x30($inp),$inout3
+	movdqu	0x40($inp),$inout4
+	movdqu	0x50($inp),$inout5
+	movdqu	0x60($inp),$inout6
+	movdqu	0x70($inp),$inout7
+	lea	0x80($inp),$inp
+	sub	\$0x80,$len
+	jmp	.Lecb_dec_loop8_enter
+.align 16
+.Lecb_dec_loop8:
+	movups	$inout0,($out)
+	mov	$key_,$key		# restore $key
+	movdqu	($inp),$inout0
+	mov	$rnds_,$rounds		# restore $rounds
+	movups	$inout1,0x10($out)
+	movdqu	0x10($inp),$inout1
+	movups	$inout2,0x20($out)
+	movdqu	0x20($inp),$inout2
+	movups	$inout3,0x30($out)
+	movdqu	0x30($inp),$inout3
+	movups	$inout4,0x40($out)
+	movdqu	0x40($inp),$inout4
+	movups	$inout5,0x50($out)
+	movdqu	0x50($inp),$inout5
+	movups	$inout6,0x60($out)
+	movdqu	0x60($inp),$inout6
+	movups	$inout7,0x70($out)
+	lea	0x80($out),$out
+	movdqu	0x70($inp),$inout7
+	lea	0x80($inp),$inp
+.Lecb_dec_loop8_enter:
+
+	call	_aesni_decrypt8
+
+	$movkey	($key_),$rndkey0
+	sub	\$0x80,$len
+	jnc	.Lecb_dec_loop8
+
+	movups	$inout0,($out)
+	mov	$key_,$key		# restore $key
+	movups	$inout1,0x10($out)
+	mov	$rnds_,$rounds		# restore $rounds
+	movups	$inout2,0x20($out)
+	movups	$inout3,0x30($out)
+	movups	$inout4,0x40($out)
+	movups	$inout5,0x50($out)
+	movups	$inout6,0x60($out)
+	movups	$inout7,0x70($out)
+	lea	0x80($out),$out
+	add	\$0x80,$len
+	jz	.Lecb_ret
+
+.Lecb_dec_tail:
+	movups	($inp),$inout0
+	cmp	\$0x20,$len
+	jb	.Lecb_dec_one
+	movups	0x10($inp),$inout1
+	je	.Lecb_dec_two
+	movups	0x20($inp),$inout2
+	cmp	\$0x40,$len
+	jb	.Lecb_dec_three
+	movups	0x30($inp),$inout3
+	je	.Lecb_dec_four
+	movups	0x40($inp),$inout4
+	cmp	\$0x60,$len
+	jb	.Lecb_dec_five
+	movups	0x50($inp),$inout5
+	je	.Lecb_dec_six
+	movups	0x60($inp),$inout6
+	$movkey	($key),$rndkey0
+	call	_aesni_decrypt8
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	movups	$inout2,0x20($out)
+	movups	$inout3,0x30($out)
+	movups	$inout4,0x40($out)
+	movups	$inout5,0x50($out)
+	movups	$inout6,0x60($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_dec_one:
+___
+	&aesni_generate1("dec",$key,$rounds);
+$code.=<<___;
+	movups	$inout0,($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_dec_two:
+	xorps	$inout2,$inout2
+	call	_aesni_decrypt3
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_dec_three:
+	call	_aesni_decrypt3
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	movups	$inout2,0x20($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_dec_four:
+	call	_aesni_decrypt4
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	movups	$inout2,0x20($out)
+	movups	$inout3,0x30($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_dec_five:
+	xorps	$inout5,$inout5
+	call	_aesni_decrypt6
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	movups	$inout2,0x20($out)
+	movups	$inout3,0x30($out)
+	movups	$inout4,0x40($out)
+	jmp	.Lecb_ret
+.align	16
+.Lecb_dec_six:
+	call	_aesni_decrypt6
+	movups	$inout0,($out)
+	movups	$inout1,0x10($out)
+	movups	$inout2,0x20($out)
+	movups	$inout3,0x30($out)
+	movups	$inout4,0x40($out)
+	movups	$inout5,0x50($out)
+
+.Lecb_ret:
+	ret
+.size	aesni_ecb_encrypt,.-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="%r9";	# 6th argument
+
+my $increment="%xmm8";
+my $bswap_mask="%xmm9";
+
+$code.=<<___;
+.globl	aesni_ccm64_encrypt_blocks
+.type	aesni_ccm64_encrypt_blocks,\@function,6
+.align	16
+aesni_ccm64_encrypt_blocks:
+___
+$code.=<<___ if ($win64);
+	lea	-0x58(%rsp),%rsp
+	movaps	%xmm6,(%rsp)
+	movaps	%xmm7,0x10(%rsp)
+	movaps	%xmm8,0x20(%rsp)
+	movaps	%xmm9,0x30(%rsp)
+.Lccm64_enc_body:
+___
+$code.=<<___;
+	movdqu	($ivp),$iv
+	movdqu	($cmac),$inout1
+	movdqa	.Lincrement64(%rip),$increment
+	movdqa	.Lbswap_mask(%rip),$bswap_mask
+	pshufb	$bswap_mask,$iv			# keep iv in reverse order
+
+	mov	240($key),$rounds		# key->rounds
+	mov	$key,$key_
+	mov	$rounds,$rnds_
+	movdqa	$iv,$inout0
+
+.Lccm64_enc_outer:
+	movups	($inp),$in0			# load inp
+	pshufb	$bswap_mask,$inout0
+	mov	$key_,$key
+	mov	$rnds_,$rounds
+
+	$movkey	($key),$rndkey0
+	shr	\$1,$rounds
+	$movkey	16($key),$rndkey1
+	xorps	$rndkey0,$in0
+	lea	32($key),$key
+	xorps	$rndkey0,$inout0
+	xorps	$inout1,$in0			# cmac^=inp
+	$movkey	($key),$rndkey0
+
+.Lccm64_enc2_loop:
+	aesenc	$rndkey1,$inout0
+	dec	$rounds
+	aesenc	$rndkey1,$inout1
+	$movkey	16($key),$rndkey1
+	aesenc	$rndkey0,$inout0
+	lea	32($key),$key
+	aesenc	$rndkey0,$inout1
+	$movkey	0($key),$rndkey0
+	jnz	.Lccm64_enc2_loop
+	aesenc	$rndkey1,$inout0
+	aesenc	$rndkey1,$inout1
+	aesenclast	$rndkey0,$inout0
+	aesenclast	$rndkey0,$inout1
+
+	paddq	$increment,$iv
+	dec	$len
+	lea	16($inp),$inp
+	xorps	$inout0,$in0			# inp ^= E(iv)
+	movdqa	$iv,$inout0
+	movups	$in0,($out)			# save output
+	lea	16($out),$out
+	jnz	.Lccm64_enc_outer
+
+	movups	$inout1,($cmac)
+___
+$code.=<<___ if ($win64);
+	movaps	(%rsp),%xmm6
+	movaps	0x10(%rsp),%xmm7
+	movaps	0x20(%rsp),%xmm8
+	movaps	0x30(%rsp),%xmm9
+	lea	0x58(%rsp),%rsp
+.Lccm64_enc_ret:
+___
+$code.=<<___;
+	ret
+.size	aesni_ccm64_encrypt_blocks,.-aesni_ccm64_encrypt_blocks
+___
+######################################################################
+$code.=<<___;
+.globl	aesni_ccm64_decrypt_blocks
+.type	aesni_ccm64_decrypt_blocks,\@function,6
+.align	16
+aesni_ccm64_decrypt_blocks:
+___
+$code.=<<___ if ($win64);
+	lea	-0x58(%rsp),%rsp
+	movaps	%xmm6,(%rsp)
+	movaps	%xmm7,0x10(%rsp)
+	movaps	%xmm8,0x20(%rsp)
+	movaps	%xmm9,0x30(%rsp)
+.Lccm64_dec_body:
+___
+$code.=<<___;
+	movdqu	($ivp),$iv
+	movdqu	($cmac),$inout1
+	movdqa	.Lincrement64(%rip),$increment
+	movdqa	.Lbswap_mask(%rip),$bswap_mask
+
+	mov	240($key),$rounds		# key->rounds
+	movdqa	$iv,$inout0
+	pshufb	$bswap_mask,$iv			# keep iv in reverse order
+	mov	$rounds,$rnds_
+	mov	$key,$key_
+___
+	&aesni_generate1("enc",$key,$rounds);
+$code.=<<___;
+.Lccm64_dec_outer:
+	paddq	$increment,$iv
+	movups	($inp),$in0			# load inp
+	xorps	$inout0,$in0
+	movdqa	$iv,$inout0
+	lea	16($inp),$inp
+	pshufb	$bswap_mask,$inout0
+	mov	$key_,$key
+	mov	$rnds_,$rounds
+	movups	$in0,($out)
+	lea	16($out),$out
+
+	sub	\$1,$len
+	jz	.Lccm64_dec_break
+
+	$movkey	($key),$rndkey0
+	shr	\$1,$rounds
+	$movkey	16($key),$rndkey1
+	xorps	$rndkey0,$in0
+	lea	32($key),$key
+	xorps	$rndkey0,$inout0
+	xorps	$in0,$inout1			# cmac^=out
+	$movkey	($key),$rndkey0
+
+.Lccm64_dec2_loop:
+	aesenc	$rndkey1,$inout0
+	dec	$rounds
+	aesenc	$rndkey1,$inout1
+	$movkey	16($key),$rndkey1
+	aesenc	$rndkey0,$inout0
+	lea	32($key),$key
+	aesenc	$rndkey0,$inout1
+	$movkey	0($key),$rndkey0
+	jnz	.Lccm64_dec2_loop
+	aesenc	$rndkey1,$inout0
+	aesenc	$rndkey1,$inout1
+	aesenclast	$rndkey0,$inout0
+	jmp	.Lccm64_dec_outer
+
+.align	16
+.Lccm64_dec_break:
+___
+	&aesni_generate1("enc",$key,$rounds,$inout1);
+$code.=<<___;
+	movups	$inout1,($cmac)
+___
+$code.=<<___ if ($win64);
+	movaps	(%rsp),%xmm6
+	movaps	0x10(%rsp),%xmm7
+	movaps	0x20(%rsp),%xmm8
+	movaps	0x30(%rsp),%xmm9
+	lea	0x58(%rsp),%rsp
+.Lccm64_dec_ret:
+___
+$code.=<<___;
+	ret
+.size	aesni_ccm64_decrypt_blocks,.-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)
+#
+{
+my $reserved = $win64?0:-0x28;
+my ($in0,$in1,$in2,$in3)=map("%xmm$_",(8..11));
+my ($iv0,$iv1,$ivec)=("%xmm12","%xmm13","%xmm14");
+my $bswap_mask="%xmm15";
+
+$code.=<<___;
+.globl	aesni_ctr32_encrypt_blocks
+.type	aesni_ctr32_encrypt_blocks,\@function,5
+.align	16
+aesni_ctr32_encrypt_blocks:
+___
+$code.=<<___ if ($win64);
+	lea	-0xc8(%rsp),%rsp
+	movaps	%xmm6,0x20(%rsp)
+	movaps	%xmm7,0x30(%rsp)
+	movaps	%xmm8,0x40(%rsp)
+	movaps	%xmm9,0x50(%rsp)
+	movaps	%xmm10,0x60(%rsp)
+	movaps	%xmm11,0x70(%rsp)
+	movaps	%xmm12,0x80(%rsp)
+	movaps	%xmm13,0x90(%rsp)
+	movaps	%xmm14,0xa0(%rsp)
+	movaps	%xmm15,0xb0(%rsp)
+.Lctr32_body:
+___
+$code.=<<___;
+	cmp	\$1,$len
+	je	.Lctr32_one_shortcut
+
+	movdqu	($ivp),$ivec
+	movdqa	.Lbswap_mask(%rip),$bswap_mask
+	xor	$rounds,$rounds
+	pextrd	\$3,$ivec,$rnds_		# pull 32-bit counter
+	pinsrd	\$3,$rounds,$ivec		# wipe 32-bit counter
+
+	mov	240($key),$rounds		# key->rounds
+	bswap	$rnds_
+	pxor	$iv0,$iv0			# vector of 3 32-bit counters
+	pxor	$iv1,$iv1			# vector of 3 32-bit counters
+	pinsrd	\$0,$rnds_,$iv0
+	lea	3($rnds_),$key_
+	pinsrd	\$0,$key_,$iv1
+	inc	$rnds_
+	pinsrd	\$1,$rnds_,$iv0
+	inc	$key_
+	pinsrd	\$1,$key_,$iv1
+	inc	$rnds_
+	pinsrd	\$2,$rnds_,$iv0
+	inc	$key_
+	pinsrd	\$2,$key_,$iv1
+	movdqa	$iv0,$reserved(%rsp)
+	pshufb	$bswap_mask,$iv0
+	movdqa	$iv1,`$reserved+0x10`(%rsp)
+	pshufb	$bswap_mask,$iv1
+
+	pshufd	\$`3<<6`,$iv0,$inout0		# place counter to upper dword
+	pshufd	\$`2<<6`,$iv0,$inout1
+	pshufd	\$`1<<6`,$iv0,$inout2
+	cmp	\$6,$len
+	jb	.Lctr32_tail
+	shr	\$1,$rounds
+	mov	$key,$key_			# backup $key
+	mov	$rounds,$rnds_			# backup $rounds
+	sub	\$6,$len
+	jmp	.Lctr32_loop6
+
+.align	16
+.Lctr32_loop6:
+	pshufd	\$`3<<6`,$iv1,$inout3
+	por	$ivec,$inout0			# merge counter-less ivec
+	 $movkey	($key_),$rndkey0
+	pshufd	\$`2<<6`,$iv1,$inout4
+	por	$ivec,$inout1
+	 $movkey	16($key_),$rndkey1
+	pshufd	\$`1<<6`,$iv1,$inout5