blob: a9b31515b1cad8613e7ef89669416ac6d2d3e792 [file] [log] [blame]
#! /usr/bin/env perl
# Copyright 2009-2016 The OpenSSL Project Authors. All Rights Reserved.
#
# Licensed under the OpenSSL license (the "License"). You may not use
# this file except in compliance with the License. You can obtain a copy
# in the file LICENSE in the source distribution or at
# https://www.openssl.org/source/license.html
#
# ====================================================================
# Written by Andy Polyakov <appro@openssl.org> 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 0.74
# CTR 1.14 0.91 0.74
#
# 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 1.25 cycles processing
# one byte out of 8KB with 128-bit key, Sandy Bridge - 0.90. Just like
# in CTR mode AES instruction interleave factor was chosen to be 6x.
# November 2015
#
# Add aesni_ocb_[en|de]crypt. AES instruction interleave factor was
# chosen to be 6x.
######################################################################
# Current large-block performance in cycles per byte processed with
# 128-bit key (less is better).
#
# CBC en-/decrypt CTR XTS ECB OCB
# Westmere 3.77/1.25 1.25 1.25 1.26
# * Bridge 5.07/0.74 0.75 0.90 0.85 0.98
# Haswell 4.44/0.63 0.63 0.73 0.63 0.70
# Skylake 2.62/0.63 0.63 0.63 0.63
# Silvermont 5.75/3.54 3.56 4.12 3.87(*) 4.11
# Knights L 2.54/0.77 0.78 0.85 - 1.50
# Goldmont 3.82/1.26 1.26 1.29 1.29 1.50
# Bulldozer 5.77/0.70 0.72 0.90 0.70 0.95
# Ryzen 2.71/0.35 0.35 0.44 0.38 0.49
#
# (*) Atom Silvermont ECB result is suboptimal because of penalties
# incurred by operations on %xmm8-15. As ECB is not considered
# critical, nothing was done to mitigate the problem.
$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 $xlate) or
die "can't locate x86_64-xlate.pl";
open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"";
*STDOUT=*OUT;
$movkey = $PREFIX eq "aesni" ? "movups" : "movups";
@_4args=$win64? ("%rcx","%rdx","%r8", "%r9") : # Win64 order
("%rdi","%rsi","%rdx","%rcx"); # Unix order
$code=".text\n";
$code.=".extern OPENSSL_ia32cap_P\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.=<<___;
pxor $rndkey0,$rndkey0 # clear register bank
pxor $rndkey1,$rndkey1
movups $inout0,($out) # output
pxor $inout0,$inout0
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.=<<___;
pxor $rndkey0,$rndkey0 # clear register bank
pxor $rndkey1,$rndkey1
movups $inout0,($out) # output
pxor $inout0,$inout0
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 originally made no sense to implement 2x subroutine.
# But times change and it became appropriate to spend extra 192 bytes
# on 2x subroutine on Atom Silvermont account. For processors that
# can schedule aes[enc|dec] every cycle optimal interleave factor
# equals to corresponding instructions latency. 8x is optimal for
# * Bridge and "super-optimal" for other Intel CPUs...
sub aesni_generate2 {
my $dir=shift;
# As already mentioned it takes in $key and $rounds, which are *not*
# preserved. $inout[0-1] is cipher/clear text...
$code.=<<___;
.type _aesni_${dir}rypt2,\@abi-omnipotent
.align 16
_aesni_${dir}rypt2:
$movkey ($key),$rndkey0
shl \$4,$rounds
$movkey 16($key),$rndkey1
xorps $rndkey0,$inout0
xorps $rndkey0,$inout1
$movkey 32($key),$rndkey0
lea 32($key,$rounds),$key
neg %rax # $rounds
add \$16,%rax
.L${dir}_loop2:
aes${dir} $rndkey1,$inout0
aes${dir} $rndkey1,$inout1
$movkey ($key,%rax),$rndkey1
add \$32,%rax
aes${dir} $rndkey0,$inout0
aes${dir} $rndkey0,$inout1
$movkey -16($key,%rax),$rndkey0
jnz .L${dir}_loop2
aes${dir} $rndkey1,$inout0
aes${dir} $rndkey1,$inout1
aes${dir}last $rndkey0,$inout0
aes${dir}last $rndkey0,$inout1
ret
.size _aesni_${dir}rypt2,.-_aesni_${dir}rypt2
___
}
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),$rndkey0
shl \$4,$rounds
$movkey 16($key),$rndkey1
xorps $rndkey0,$inout0
xorps $rndkey0,$inout1
xorps $rndkey0,$inout2
$movkey 32($key),$rndkey0
lea 32($key,$rounds),$key
neg %rax # $rounds
add \$16,%rax
.L${dir}_loop3:
aes${dir} $rndkey1,$inout0
aes${dir} $rndkey1,$inout1
aes${dir} $rndkey1,$inout2
$movkey ($key,%rax),$rndkey1
add \$32,%rax
aes${dir} $rndkey0,$inout0
aes${dir} $rndkey0,$inout1
aes${dir} $rndkey0,$inout2
$movkey -16($key,%rax),$rndkey0
jnz .L${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),$rndkey0
shl \$4,$rounds
$movkey 16($key),$rndkey1
xorps $rndkey0,$inout0
xorps $rndkey0,$inout1
xorps $rndkey0,$inout2
xorps $rndkey0,$inout3
$movkey 32($key),$rndkey0
lea 32($key,$rounds),$key
neg %rax # $rounds
.byte 0x0f,0x1f,0x00
add \$16,%rax
.L${dir}_loop4:
aes${dir} $rndkey1,$inout0
aes${dir} $rndkey1,$inout1
aes${dir} $rndkey1,$inout2
aes${dir} $rndkey1,$inout3
$movkey ($key,%rax),$rndkey1
add \$32,%rax
aes${dir} $rndkey0,$inout0
aes${dir} $rndkey0,$inout1
aes${dir} $rndkey0,$inout2
aes${dir} $rndkey0,$inout3
$movkey -16($key,%rax),$rndkey0
jnz .L${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),$rndkey0
shl \$4,$rounds
$movkey 16($key),$rndkey1
xorps $rndkey0,$inout0
pxor $rndkey0,$inout1
pxor $rndkey0,$inout2
aes${dir} $rndkey1,$inout0
lea 32($key,$rounds),$key
neg %rax # $rounds
aes${dir} $rndkey1,$inout1
pxor $rndkey0,$inout3
pxor $rndkey0,$inout4
aes${dir} $rndkey1,$inout2
pxor $rndkey0,$inout5
$movkey ($key,%rax),$rndkey0
add \$16,%rax
jmp .L${dir}_loop6_enter
.align 16
.L${dir}_loop6:
aes${dir} $rndkey1,$inout0
aes${dir} $rndkey1,$inout1
aes${dir} $rndkey1,$inout2
.L${dir}_loop6_enter:
aes${dir} $rndkey1,$inout3
aes${dir} $rndkey1,$inout4
aes${dir} $rndkey1,$inout5
$movkey ($key,%rax),$rndkey1
add \$32,%rax
aes${dir} $rndkey0,$inout0
aes${dir} $rndkey0,$inout1
aes${dir} $rndkey0,$inout2
aes${dir} $rndkey0,$inout3
aes${dir} $rndkey0,$inout4
aes${dir} $rndkey0,$inout5
$movkey -16($key,%rax),$rndkey0
jnz .L${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),$rndkey0
shl \$4,$rounds
$movkey 16($key),$rndkey1
xorps $rndkey0,$inout0
xorps $rndkey0,$inout1
pxor $rndkey0,$inout2
pxor $rndkey0,$inout3
pxor $rndkey0,$inout4
lea 32($key,$rounds),$key
neg %rax # $rounds
aes${dir} $rndkey1,$inout0
pxor $rndkey0,$inout5
pxor $rndkey0,$inout6
aes${dir} $rndkey1,$inout1
pxor $rndkey0,$inout7
$movkey ($key,%rax),$rndkey0
add \$16,%rax
jmp .L${dir}_loop8_inner
.align 16
.L${dir}_loop8:
aes${dir} $rndkey1,$inout0
aes${dir} $rndkey1,$inout1
.L${dir}_loop8_inner:
aes${dir} $rndkey1,$inout2
aes${dir} $rndkey1,$inout3
aes${dir} $rndkey1,$inout4
aes${dir} $rndkey1,$inout5
aes${dir} $rndkey1,$inout6
aes${dir} $rndkey1,$inout7
.L${dir}_loop8_enter:
$movkey ($key,%rax),$rndkey1
add \$32,%rax
aes${dir} $rndkey0,$inout0
aes${dir} $rndkey0,$inout1
aes${dir} $rndkey0,$inout2
aes${dir} $rndkey0,$inout3
aes${dir} $rndkey0,$inout4
aes${dir} $rndkey0,$inout5
aes${dir} $rndkey0,$inout6
aes${dir} $rndkey0,$inout7
$movkey -16($key,%rax),$rndkey0
jnz .L${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_generate2("enc") if ($PREFIX eq "aesni");
&aesni_generate2("dec");
&aesni_generate3("enc") if ($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:
___
$code.=<<___ if ($win64);
lea -0x58(%rsp),%rsp
movaps %xmm6,(%rsp) # offload $inout4..7
movaps %xmm7,0x10(%rsp)
movaps %xmm8,0x20(%rsp)
movaps %xmm9,0x30(%rsp)
.Lecb_enc_body:
___
$code.=<<___;
and \$-16,$len # if ($len<16)
jz .Lecb_ret # return
mov 240($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 # if ($len<8*16)
jb .Lecb_enc_tail # short input
movdqu ($inp),$inout0 # load 8 input blocks
movdqu 0x10($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 # $inp+=8*16
sub \$0x80,$len # $len-=8*16 (can be zero)
jmp .Lecb_enc_loop8_enter
.align 16
.Lecb_enc_loop8:
movups $inout0,($out) # store 8 output blocks
mov $key_,$key # restore $key
movdqu ($inp),$inout0 # load 8 input blocks
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 # $out+=8*16
movdqu 0x70($inp),$inout7
lea 0x80($inp),$inp # $inp+=8*16
.Lecb_enc_loop8_enter:
call _aesni_encrypt8
sub \$0x80,$len
jnc .Lecb_enc_loop8 # loop if $len-=8*16 didn't borrow
movups $inout0,($out) # store 8 output blocks
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 # $out+=8*16
add \$0x80,$len # restore real remaining $len
jz .Lecb_ret # done if ($len==0)
.Lecb_enc_tail: # $len is less than 8*16
movups ($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
xorps $inout7,$inout7
call _aesni_encrypt8
movups $inout0,($out) # store 7 output blocks
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) # store one output block
jmp .Lecb_ret
.align 16
.Lecb_enc_two:
call _aesni_encrypt2
movups $inout0,($out) # store 2 output blocks
movups $inout1,0x10($out)
jmp .Lecb_ret
.align 16
.Lecb_enc_three:
call _aesni_encrypt3
movups $inout0,($out) # store 3 output blocks
movups $inout1,0x10($out)
movups $inout2,0x20($out)
jmp .Lecb_ret
.align 16
.Lecb_enc_four:
call _aesni_encrypt4
movups $inout0,($out) # store 4 output blocks
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) # store 5 output blocks
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) # store 6 output blocks
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,$len # if ($len<8*16)
jb .Lecb_dec_tail # short input
movdqu ($inp),$inout0 # load 8 input blocks
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 # $inp+=8*16
sub \$0x80,$len # $len-=8*16 (can be zero)
jmp .Lecb_dec_loop8_enter
.align 16
.Lecb_dec_loop8:
movups $inout0,($out) # store 8 output blocks
mov $key_,$key # restore $key
movdqu ($inp),$inout0 # load 8 input blocks
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 # $out+=8*16
movdqu 0x70($inp),$inout7
lea 0x80($inp),$inp # $inp+=8*16
.Lecb_dec_loop8_enter:
call _aesni_decrypt8
$movkey ($key_),$rndkey0
sub \$0x80,$len
jnc .Lecb_dec_loop8 # loop if $len-=8*16 didn't borrow
movups $inout0,($out) # store 8 output blocks
pxor $inout0,$inout0 # clear register bank
mov $key_,$key # restore $key
movups $inout1,0x10($out)
pxor $inout1,$inout1
mov $rnds_,$rounds # restore $rounds
movups $inout2,0x20($out)
pxor $inout2,$inout2
movups $inout3,0x30($out)
pxor $inout3,$inout3
movups $inout4,0x40($out)
pxor $inout4,$inout4
movups $inout5,0x50($out)
pxor $inout5,$inout5
movups $inout6,0x60($out)
pxor $inout6,$inout6
movups $inout7,0x70($out)
pxor $inout7,$inout7
lea 0x80($out),$out # $out+=8*16
add \$0x80,$len # restore real remaining $len
jz .Lecb_ret # done if ($len==0)
.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
xorps $inout7,$inout7
call _aesni_decrypt8
movups $inout0,($out) # store 7 output blocks
pxor $inout0,$inout0 # clear register bank
movups $inout1,0x10($out)
pxor $inout1,$inout1
movups $inout2,0x20($out)
pxor $inout2,$inout2
movups $inout3,0x30($out)
pxor $inout3,$inout3
movups $inout4,0x40($out)
pxor $inout4,$inout4
movups $inout5,0x50($out)
pxor $inout5,$inout5
movups $inout6,0x60($out)
pxor $inout6,$inout6
pxor $inout7,$inout7
jmp .Lecb_ret
.align 16
.Lecb_dec_one:
___
&aesni_generate1("dec",$key,$rounds);
$code.=<<___;
movups $inout0,($out) # store one output block
pxor $inout0,$inout0 # clear register bank
jmp .Lecb_ret
.align 16
.Lecb_dec_two:
call _aesni_decrypt2
movups $inout0,($out) # store 2 output blocks
pxor $inout0,$inout0 # clear register bank
movups $inout1,0x10($out)
pxor $inout1,$inout1
jmp .Lecb_ret
.align 16
.Lecb_dec_three:
call _aesni_decrypt3
movups $inout0,($out) # store 3 output blocks
pxor $inout0,$inout0 # clear register bank
movups $inout1,0x10($out)
pxor $inout1,$inout1
movups $inout2,0x20($out)
pxor $inout2,$inout2
jmp .Lecb_ret
.align 16
.Lecb_dec_four:
call _aesni_decrypt4
movups $inout0,($out) # store 4 output blocks
pxor $inout0,$inout0 # clear register bank
movups $inout1,0x10($out)
pxor $inout1,$inout1
movups $inout2,0x20($out)
pxor $inout2,$inout2
movups $inout3,0x30($out)
pxor $inout3,$inout3
jmp .Lecb_ret
.align 16
.Lecb_dec_five:
xorps $inout5,$inout5
call _aesni_decrypt6
movups $inout0,($out) # store 5 output blocks
pxor $inout0,$inout0 # clear register bank
movups $inout1,0x10($out)
pxor $inout1,$inout1
movups $inout2,0x20($out)
pxor $inout2,$inout2
movups $inout3,0x30($out)
pxor $inout3,$inout3
movups $inout4,0x40($out)
pxor $inout4,$inout4
pxor $inout5,$inout5
jmp .Lecb_ret
.align 16
.Lecb_dec_six:
call _aesni_decrypt6
movups $inout0,($out) # store 6 output blocks
pxor $inout0,$inout0 # clear register bank
movups $inout1,0x10($out)
pxor $inout1,$inout1
movups $inout2,0x20($out)
pxor $inout2,$inout2
movups $inout3,0x30($out)
pxor $inout3,$inout3
movups $inout4,0x40($out)
pxor $inout4,$inout4
movups $inout5,0x50($out)
pxor $inout5,$inout5
.Lecb_ret:
xorps $rndkey0,$rndkey0 # %xmm0
pxor $rndkey1,$rndkey1
___
$code.=<<___ if ($win64);
movaps (%rsp),%xmm6
movaps %xmm0,(%rsp) # clear stack
movaps 0x10(%rsp),%xmm7
movaps %xmm0,0x10(%rsp)
movaps 0x20(%rsp),%xmm8
movaps %xmm0,0x20(%rsp)
movaps 0x30(%rsp),%xmm9
movaps %xmm0,0x30(%rsp)
lea 0x58(%rsp),%rsp
.Lecb_enc_ret:
___
$code.=<<___;
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="%xmm9";
my $iv="%xmm6";
my $bswap_mask="%xmm7";
$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) # $iv
movaps %xmm7,0x10(%rsp) # $bswap_mask
movaps %xmm8,0x20(%rsp) # $in0
movaps %xmm9,0x30(%rsp) # $increment
.Lccm64_enc_body:
___
$code.=<<___;
mov 240($key),$rounds # key->rounds
movdqu ($ivp),$iv
movdqa .Lincrement64(%rip),$increment
movdqa .Lbswap_mask(%rip),$bswap_mask
shl \$4,$rounds
mov \$16,$rnds_
lea 0($key),$key_
movdqu ($cmac),$inout1
movdqa $iv,$inout0
lea 32($key,$rounds),$key # end of key schedule
pshufb $bswap_mask,$iv
sub %rax,%r10 # twisted $rounds
jmp .Lccm64_enc_outer
.align 16
.Lccm64_enc_outer:
$movkey ($key_),$rndkey0
mov %r10,%rax
movups ($inp),$in0 # load inp
xorps $rndkey0,$inout0 # counter
$movkey 16($key_),$rndkey1
xorps $in0,$rndkey0
xorps $rndkey0,$inout1 # cmac^=inp
$movkey 32($key_),$rndkey0
.Lccm64_enc2_loop:
aesenc $rndkey1,$inout0
aesenc $rndkey1,$inout1
$movkey ($key,%rax),$rndkey1
add \$32,%rax
aesenc $rndkey0,$inout0
aesenc $rndkey0,$inout1
$movkey -16($key,%rax),$rndkey0
jnz .Lccm64_enc2_loop
aesenc $rndkey1,$inout0
aesenc $rndkey1,$inout1
paddq $increment,$iv
dec $len # $len-- ($len is in blocks)
aesenclast $rndkey0,$inout0
aesenclast $rndkey0,$inout1
lea 16($inp),$inp
xorps $inout0,$in0 # inp ^= E(iv)
movdqa $iv,$inout0
movups $in0,($out) # save output
pshufb $bswap_mask,$inout0
lea 16($out),$out # $out+=16
jnz .Lccm64_enc_outer # loop if ($len!=0)
pxor $rndkey0,$rndkey0 # clear register bank
pxor $rndkey1,$rndkey1
pxor $inout0,$inout0
movups $inout1,($cmac) # store resulting mac
pxor $inout1,$inout1
pxor $in0,$in0
pxor $iv,$iv
___
$code.=<<___ if ($win64);
movaps (%rsp),%xmm6
movaps %xmm0,(%rsp) # clear stack
movaps 0x10(%rsp),%xmm7
movaps %xmm0,0x10(%rsp)
movaps 0x20(%rsp),%xmm8
movaps %xmm0,0x20(%rsp)
movaps 0x30(%rsp),%xmm9
movaps %xmm0,0x30(%rsp)
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) # $iv
movaps %xmm7,0x10(%rsp) # $bswap_mask
movaps %xmm8,0x20(%rsp) # $in8
movaps %xmm9,0x30(%rsp) # $increment
.Lccm64_dec_body:
___
$code.=<<___;
mov 240($key),$rounds # key->rounds
movups ($ivp),$iv
movdqu ($cmac),$inout1
movdqa .Lincrement64(%rip),$increment
movdqa .Lbswap_mask(%rip),$bswap_mask
movaps $iv,$inout0
mov $rounds,$rnds_
mov $key,$key_
pshufb $bswap_mask,$iv
___
&aesni_generate1("enc",$key,$rounds);
$code.=<<___;
shl \$4,$rnds_
mov \$16,$rounds
movups ($inp),$in0 # load inp
paddq $increment,$iv
lea 16($inp),$inp # $inp+=16
sub %r10,%rax # twisted $rounds
lea 32($key_,$rnds_),$key # end of key schedule
mov %rax,%r10
jmp .Lccm64_dec_outer
.align 16
.Lccm64_dec_outer:
xorps $inout0,$in0 # inp ^= E(iv)
movdqa $iv,$inout0
movups $in0,($out) # save output
lea 16($out),$out # $out+=16
pshufb $bswap_mask,$inout0
sub \$1,$len # $len-- ($len is in blocks)
jz .Lccm64_dec_break # if ($len==0) break
$movkey ($key_),$rndkey0
mov %r10,%rax
$movkey 16($key_),$rndkey1
xorps $rndkey0,$in0
xorps $rndkey0,$inout0
xorps $in0,$inout1 # cmac^=out
$movkey 32($key_),$rndkey0
jmp .Lccm64_dec2_loop
.align 16
.Lccm64_dec2_loop:
aesenc $rndkey1,$inout0
aesenc $rndkey1,$inout1
$movkey ($key,%rax),$rndkey1
add \$32,%rax
aesenc $rndkey0,$inout0
aesenc $rndkey0,$inout1
$movkey -16($key,%rax),$rndkey0
jnz .Lccm64_dec2_loop
movups ($inp),$in0 # load input
paddq $increment,$iv
aesenc $rndkey1,$inout0
aesenc $rndkey1,$inout1
aesenclast $rndkey0,$inout0
aesenclast $rndkey0,$inout1
lea 16($inp),$inp # $inp+=16
jmp .Lccm64_dec_outer
.align 16
.Lccm64_dec_break:
#xorps $in0,$inout1 # cmac^=out
mov 240($key_),$rounds
___
&aesni_generate1("enc",$key_,$rounds,$inout1,$in0);
$code.=<<___;
pxor $rndkey0,$rndkey0 # clear register bank
pxor $rndkey1,$rndkey1
pxor $inout0,$inout0
movups $inout1,($cmac) # store resulting mac
pxor $inout1,$inout1
pxor $in0,$in0
pxor $iv,$iv
___
$code.=<<___ if ($win64);
movaps (%rsp),%xmm6
movaps %xmm0,(%rsp) # clear stack
movaps 0x10(%rsp),%xmm7
movaps %xmm0,0x10(%rsp)
movaps 0x20(%rsp),%xmm8
movaps %xmm0,0x20(%rsp)
movaps 0x30(%rsp),%xmm9
movaps %xmm0,0x30(%rsp)
lea 0x58(%rsp),%rsp
.Lccm64_dec_ret:
___
$code.=<<___;
ret
.size aesni_ccm64_decrypt_blocks,.-aesni_ccm64_decrypt_blocks
___
}
######################################################################
# void aesni_ctr32_encrypt_blocks (const void *in, void *out,
# size_t blocks, const AES_KEY *key,
# const char *ivec);
#
# Handles only complete blocks, operates on 32-bit counter and
# does not update *ivec! (see crypto/modes/ctr128.c for details)
#
# Overhaul based on suggestions from Shay Gueron and Vlad Krasnov,
# http://rt.openssl.org/Ticket/Display.html?id=3021&user=guest&pass=guest.
# Keywords are full unroll and modulo-schedule counter calculations
# with zero-round key xor.
{
my ($in0,$in1,$in2,$in3,$in4,$in5)=map("%xmm$_",(10..15));
my ($key0,$ctr)=("%ebp","${ivp}d");
my $frame_size = 0x80 + ($win64?160:0);
$code.=<<___;
.globl aesni_ctr32_encrypt_blocks
.type aesni_ctr32_encrypt_blocks,\@function,5
.align 16
aesni_ctr32_encrypt_blocks:
cmp \$1,$len
jne .Lctr32_bulk
# handle single block without allocating stack frame,
# useful when handling edges
movups ($ivp),$inout0
movups ($inp),$inout1
mov 240($key),%edx # key->rounds
___
&aesni_generate1("enc",$key,"%edx");
$code.=<<___;
pxor $rndkey0,$rndkey0 # clear register bank
pxor $rndkey1,$rndkey1
xorps $inout1,$inout0
pxor $inout1,$inout1
movups $inout0,($out)
xorps $inout0,$inout0
jmp .Lctr32_epilogue
.align 16
.Lctr32_bulk:
lea (%rsp),$key_ # use $key_ as frame pointer
push %rbp
sub \$$frame_size,%rsp
and \$-16,%rsp # Linux kernel stack can be incorrectly seeded
___
$code.=<<___ if ($win64);
movaps %xmm6,-0xa8($key_) # offload everything
movaps %xmm7,-0x98($key_)
movaps %xmm8,-0x88($key_)
movaps %xmm9,-0x78($key_)
movaps %xmm10,-0x68($key_)
movaps %xmm11,-0x58($key_)
movaps %xmm12,-0x48($key_)
movaps %xmm13,-0x38($key_)
movaps %xmm14,-0x28($key_)
movaps %xmm15,-0x18($key_)
.Lctr32_body:
___
$code.=<<___;
# 8 16-byte words on top of stack are counter values
# xor-ed with zero-round key
movdqu ($ivp),$inout0
movdqu ($key),$rndkey0
mov 12($ivp),$ctr # counter LSB
pxor $rndkey0,$inout0
mov 12($key),$key0 # 0-round key LSB
movdqa $inout0,0x00(%rsp) # populate counter block
bswap $ctr
movdqa $inout0,$inout1
movdqa $inout0,$inout2
movdqa $inout0,$inout3
movdqa $inout0,0x40(%rsp)
movdqa $inout0,0x50(%rsp)
movdqa $inout0,0x60(%rsp)
mov %rdx,%r10 # about to borrow %rdx
movdqa $inout0,0x70(%rsp)
lea 1($ctr),%rax
lea 2($ctr),%rdx
bswap %eax
bswap %edx
xor $key0,%eax
xor $key0,%edx
pinsrd \$3,%eax,$inout1
lea 3($ctr),%rax
movdqa $inout1,0x10(%rsp)
pinsrd \$3,%edx,$inout2
bswap %eax
mov %r10,%rdx # restore %rdx
lea 4($ctr),%r10
movdqa $inout2,0x20(%rsp)
xor $key0,%eax
bswap %r10d
pinsrd \$3,%eax,$inout3
xor $key0,%r10d
movdqa $inout3,0x30(%rsp)
lea 5($ctr),%r9
mov %r10d,0x40+12(%rsp)
bswap %r9d
lea 6($ctr),%r10
mov 240($key),$rounds # key->rounds
xor $key0,%r9d
bswap %r10d
mov %r9d,0x50+12(%rsp)
xor $key0,%r10d
lea 7($ctr),%r9
mov %r10d,0x60+12(%rsp)
bswap %r9d
leaq OPENSSL_ia32cap_P(%rip),%r10
mov 4(%r10),%r10d
xor $key0,%r9d
and \$`1<<26|1<<22`,%r10d # isolate XSAVE+MOVBE
mov %r9d,0x70+12(%rsp)
$movkey 0x10($key),$rndkey1
movdqa 0x40(%rsp),$inout4
movdqa 0x50(%rsp),$inout5
cmp \$8,$len # $len is in blocks
jb .Lctr32_tail # short input if ($len<8)
sub \$6,$len # $len is biased by -6
cmp \$`1<<22`,%r10d # check for MOVBE without XSAVE
je .Lctr32_6x # [which denotes Atom Silvermont]
lea 0x80($key),$key # size optimization
sub \$2,$len # $len is biased by -8
jmp .Lctr32_loop8
.align 16
.Lctr32_6x:
shl \$4,$rounds
mov \$48,$rnds_
bswap $key0
lea 32($key,$rounds),$key # end of key schedule
sub %rax,%r10 # twisted $rounds
jmp .Lctr32_loop6
.align 16
.Lctr32_loop6:
add \$6,$ctr # next counter value
$movkey -48($key,$rnds_),$rndkey0
aesenc $rndkey1,$inout0
mov $ctr,%eax
xor $key0,%eax
aesenc $rndkey1,$inout1
movbe %eax,`0x00+12`(%rsp) # store next counter value
lea 1($ctr),%eax
aesenc $rndkey1,$inout2
xor $key0,%eax
movbe %eax,`0x10+12`(%rsp)
aesenc $rndkey1,$inout3
lea 2($ctr),%eax
xor $key0,%eax
aesenc $rndkey1,$inout4
movbe %eax,`0x20+12`(%rsp)
lea 3($ctr),%eax
aesenc $rndkey1,$inout5
$movkey -32($key,$rnds_),$rndkey1
xor $key0,%eax
aesenc $rndkey0,$inout0
movbe %eax,`0x30+12`(%rsp)
lea 4($ctr),%eax
aesenc $rndkey0,$inout1
xor $key0,%eax
movbe %eax,`0x40+12`(%rsp)
aesenc $rndkey0,$inout2
lea 5($ctr),%eax
xor $key0,%eax
aesenc $rndkey0,$inout3
movbe %eax,`0x50+12`(%rsp)
mov %r10,%rax # mov $rnds_,$rounds
aesenc $rndkey0,$inout4
aesenc $rndkey0,$inout5
$movkey -16($key,$rnds_),$rndkey0
call .Lenc_loop6
movdqu ($inp),$inout6 # load 6 input blocks
movdqu 0x10($inp),$inout7
movdqu 0x20($inp),$in0
movdqu 0x30($inp),$in1
movdqu 0x40($inp),$in2
movdqu 0x50($inp),$in3
lea 0x60($inp),$inp # $inp+=6*16
$movkey -64($key,$rnds_),$rndkey1
pxor $inout0,$inout6 # inp^=E(ctr)
movaps 0x00(%rsp),$inout0 # load next counter [xor-ed with 0 round]
pxor $inout1,$inout7
movaps 0x10(%rsp),$inout1
pxor $inout2,$in0
movaps 0x20(%rsp),$inout2
pxor $inout3,$in1
movaps 0x30(%rsp),$inout3
pxor $inout4,$in2
movaps 0x40(%rsp),$inout4
pxor $inout5,$in3
movaps 0x50(%rsp),$inout5
movdqu $inout6,($out) # store 6 output blocks
movdqu $inout7,0x10($out)
movdqu $in0,0x20($out)
movdqu $in1,0x30($out)
movdqu $in2,0x40($out)
movdqu $in3,0x50($out)
lea 0x60($out),$out # $out+=6*16
sub \$6,$len
jnc .Lctr32_loop6 # loop if $len-=6 didn't borrow
add \$6,$len # restore real remaining $len
jz .Lctr32_done # done if ($len==0)
lea -48($rnds_),$rounds
lea -80($key,$rnds_),$key # restore $key
neg $rounds
shr \$4,$rounds # restore $rounds
jmp .Lctr32_tail
.align 32
.Lctr32_loop8:
add \$8,$ctr # next counter value
movdqa 0x60(%rsp),$inout6
aesenc $rndkey1,$inout0
mov $ctr,%r9d
movdqa 0x70(%rsp),$inout7
aesenc $rndkey1,$inout1
bswap %r9d
$movkey 0x20-0x80($key),$rndkey0
aesenc $rndkey1,$inout2
xor $key0,%r9d
nop
aesenc $rndkey1,$inout3
mov %r9d,0x00+12(%rsp) # store next counter value
lea 1($ctr),%r9
aesenc $rndkey1,$inout4
aesenc $rndkey1,$inout5
aesenc $rndkey1,$inout6
aesenc $rndkey1,$inout7
$movkey 0x30-0x80($key),$rndkey1
___
for($i=2;$i<8;$i++) {
my $rndkeyx = ($i&1)?$rndkey1:$rndkey0;
$code.=<<___;
bswap %r9d
aesenc $rndkeyx,$inout0
aesenc $rndkeyx,$inout1
xor $key0,%r9d
.byte 0x66,0x90
aesenc $rndkeyx,$inout2
aesenc $rndkeyx,$inout3
mov %r9d,`0x10*($i-1)`+12(%rsp)
lea $i($ctr),%r9
aesenc $rndkeyx,$inout4
aesenc $rndkeyx,$inout5
aesenc $rndkeyx,$inout6
aesenc $rndkeyx,$inout7
$movkey `0x20+0x10*$i`-0x80($key),$rndkeyx
___
}
$code.=<<___;
bswap %r9d
aesenc $rndkey0,$inout0
aesenc $rndkey0,$inout1
aesenc $rndkey0,$inout2
xor $key0,%r9d
movdqu 0x00($inp),$in0 # start loading input
aesenc $rndkey0,$inout3
mov %r9d,0x70+12(%rsp)
cmp \$11,$rounds
aesenc $rndkey0,$inout4
aesenc $rndkey0,$inout5
aesenc $rndkey0,$inout6
aesenc $rndkey0,$inout7
$movkey 0xa0-0x80($key),$rndkey0
jb .Lctr32_enc_done
aesenc $rndkey1,$inout0
aesenc $rndkey1,$inout1
aesenc $rndkey1,$inout2
aesenc $rndkey1,$inout3
aesenc $rndkey1,$inout4
aesenc $rndkey1,$inout5
aesenc $rndkey1,$inout6
aesenc $rndkey1,$inout7
$movkey 0xb0-0x80($key),$rndkey1
aesenc $rndkey0,$inout0
aesenc $rndkey0,$inout1
aesenc $rndkey0,$inout2
aesenc $rndkey0,$inout3
aesenc $rndkey0,$inout4
aesenc $rndkey0,$inout5
aesenc $rndkey0,$inout6
aesenc $rndkey0,$inout7
$movkey 0xc0-0x80($key),$rndkey0
je .Lctr32_enc_done
aesenc $rndkey1,$inout0
aesenc $rndkey1,$inout1
aesenc $rndkey1,$inout2
aesenc $rndkey1,$inout3
aesenc $rndkey1,$inout4
aesenc $rndkey1,$inout5
aesenc $rndkey1,$inout6
aesenc $rndkey1,$inout7
$movkey 0xd0-0x80($key),$rndkey1
aesenc $rndkey0,$inout0
aesenc $rndkey0,$inout1
aesenc $rndkey0,$inout2
aesenc $rndkey0,$inout3
aesenc $rndkey0,$inout4
aesenc $rndkey0,$inout5
aesenc $rndkey0,$inout6
aesenc $rndkey0,$inout7
$movkey 0xe0-0x80($key),$rndkey0
jmp .Lctr32_enc_done
.align 16
.Lctr32_enc_done:
movdqu 0x10($inp),$in1
pxor $rndkey0,$in0 # input^=round[last]
movdqu 0x20($inp),$in2
pxor $rndkey0,$in1
movdqu 0x30($inp),$in3
pxor $rndkey0,$in2
movdqu 0x40($inp),$in4
pxor $rndkey0,$in3
movdqu 0x50($inp),$in5
pxor $rndkey0,$in4
pxor $rndkey0,$in5
aesenc $rndkey1,$inout0
aesenc $rndkey1,$inout1
aesenc $rndkey1,$inout2
aesenc $rndkey1,$inout3
aesenc $rndkey1,$inout4
aesenc $rndkey1,$inout5
aesenc $rndkey1,$inout6
aesenc $rndkey1,$inout7
movdqu 0x60($inp),$rndkey1 # borrow $rndkey1 for inp[6]
lea 0x80($inp),$inp # $inp+=8*16
aesenclast $in0,$inout0 # $inN is inp[N]^round[last]
pxor $rndkey0,$rndkey1 # borrowed $rndkey
movdqu 0x70-0x80($inp),$in0
aesenclast $in1,$inout1
pxor $rndkey0,$in0
movdqa 0x00(%rsp),$in1 # load next counter block
aesenclast $in2,$inout2
aesenclast $in3,$inout3
movdqa 0x10(%rsp),$in2
movdqa 0x20(%rsp),$in3
aesenclast $in4,$inout4
aesenclast $in5,$inout5
movdqa 0x30(%rsp),$in4
movdqa 0x40(%rsp),$in5
aesenclast $rndkey1,$inout6
movdqa 0x50(%rsp),$rndkey0
$movkey 0x10-0x80($key),$rndkey1#real 1st-round key
aesenclast $in0,$inout7
movups $inout0,($out) # store 8 output blocks
movdqa $in1,$inout0
movups $inout1,0x10($out)
movdqa $in2,$inout1
movups $inout2,0x20($out)
movdqa $in3,$inout2
movups $inout3,0x30($out)
movdqa $in4,$inout3
movups $inout4,0x40($out)
movdqa $in5,$inout4
movups $inout5,0x50($out)
movdqa $rndkey0,$inout5
movups $inout6,0x60($out)
movups $inout7,0x70($out)
lea 0x80($out),$out # $out+=8*16
sub \$8,$len
jnc .Lctr32_loop8 # loop if $len-=8 didn't borrow
add \$8,$len # restore real remainig $len
jz .Lctr32_done # done if ($len==0)
lea -0x80($key),$key
.Lctr32_tail:
# note that at this point $inout0..5 are populated with
# counter values xor-ed with 0-round key
lea 16($key),$key
cmp \$4,$len
jb .Lctr32_loop3
je .Lctr32_loop4
# if ($len>4) compute 7 E(counter)
shl \$4,$rounds
movdqa 0x60(%rsp),$inout6
pxor $inout7,$inout7
$movkey 16($key),$rndkey0
aesenc $rndkey1,$inout0
aesenc $rndkey1,$inout1
lea 32-16($key,$rounds),$key# prepare for .Lenc_loop8_enter
neg %rax
aesenc $rndkey1,$inout2
add \$16,%rax # prepare for .Lenc_loop8_enter
movups ($inp),$in0
aesenc $rndkey1,$inout3
aesenc $rndkey1,$inout4
movups 0x10($inp),$in1 # pre-load input
movups 0x20($inp),$in2
aesenc $rndkey1,$inout5
aesenc $rndkey1,$inout6
call .Lenc_loop8_enter
movdqu 0x30($inp),$in3
pxor $in0,$inout0
movdqu 0x40($inp),$in0
pxor $in1,$inout1
movdqu $inout0,($out) # store output
pxor $in2,$inout2
movdqu $inout1,0x10($out)
pxor $in3,$inout3
movdqu $inout2,0x20($out)
pxor $in0,$inout4
movdqu $inout3,0x30($out)
movdqu $inout4,0x40($out)
cmp \$6,$len
jb .Lctr32_done # $len was 5, stop store
movups 0x50($inp),$in1
xorps $in1,$inout5
movups $inout5,0x50($out)
je .Lctr32_done # $len was 6, stop store
movups 0x60($inp),$in2
xorps $in2,$inout6
movups $inout6,0x60($out)
jmp .Lctr32_done # $len was 7, stop store
.align 32
.Lctr32_loop4:
aesenc $rndkey1,$inout0
lea 16($key),$key
dec $rounds
aesenc $rndkey1,$inout1
aesenc $rndkey1,$inout2
aesenc $rndkey1,$inout3
$movkey ($key),$rndkey1
jnz .Lctr32_loop4
aesenclast $rndkey1,$inout0
aesenclast $rndkey1,$inout1
movups ($inp),$in0 # load input
movups 0x10($inp),$in1
aesenclast $rndkey1,$inout2
aesenclast $rndkey1,$inout3
movups 0x20($inp),$in2
movups 0x30($inp),$in3
xorps $in0,$inout0
movups $inout0,($out) # store output
xorps $in1,$inout1
movups $inout1,0x10($out)
pxor $in2,$inout2
movdqu $inout2,0x20($out)
pxor $in3,$inout3
movdqu $inout3,0x30($out)
jmp .Lctr32_done # $len was 4, stop store
.align 32
.Lctr32_loop3:
aesenc $rndkey1,$inout0
lea 16($key),$key
dec $rounds
aesenc $rndkey1,$inout1
aesenc $rndkey1,$inout2
$movkey ($key),$rndkey1
jnz .Lctr32_loop3
aesenclast $rndkey1,$inout0
aesenclast $rndkey1,$inout1
aesenclast $rndkey1,$inout2
movups ($inp),$in0 # load input
xorps $in0,$inout0
movups $inout0,($out) # store output
cmp \$2,$len
jb .Lctr32_done # $len was 1, stop store
movups 0x10($inp),$in1
xorps $in1,$inout1
movups $inout1,0x10($out)
je .Lctr32_done # $len was 2, stop store
movups 0x20($inp),$in2
xorps $in2,$inout2
movups $inout2,0x20($out) # $len was 3, stop store
.Lctr32_done:
xorps %xmm0,%xmm0 # clear regiser bank
xor $key0,$key0
pxor %xmm1,%xmm1
pxor %xmm2,%xmm2
pxor %xmm3,%xmm3
pxor %xmm4,%xmm4
pxor %xmm5,%xmm5
___
$code.=<<___ if (!$win64);
pxor %xmm6,%xmm6
pxor %xmm7,%xmm7
movaps %xmm0,0x00(%rsp) # clear stack
pxor %xmm8,%xmm8
movaps %xmm0,0x10(%rsp)
pxor %xmm9,%xmm9
movaps %xmm0,0x20(%rsp)
pxor %xmm10,%xmm10
movaps %xmm0,0x30(%rsp)
pxor %xmm11,%xmm11
movaps %xmm0,0x40(%rsp)
pxor %xmm12,%xmm12
movaps %xmm0,0x50(%rsp)
pxor %xmm13,%xmm13
movaps %xmm0,0x60(%rsp)
pxor %xmm14,%xmm14
movaps %xmm0,0x70(%rsp)
pxor %xmm15,%xmm15
___
$code.=<<___ if ($win64);
movaps -0xa8($key_),%xmm6
movaps %xmm0,-0xa8($key_) # clear stack
movaps -0x98($key_),%xmm7
movaps %xmm0,-0x98($key_)
movaps -0x88($key_),%xmm8
movaps %xmm0,-0x88($key_)
movaps -0x78($key_),%xmm9
movaps %xmm0,-0x78($key_)
movaps -0x68($key_),%xmm10
movaps %xmm0,-0x68($key_)
movaps -0x58($key_),%xmm11
movaps %xmm0,-0x58($key_)
movaps -0x48($key_),%xmm12
movaps %xmm0,-0x48($key_)
movaps -0x38($key_),%xmm13
movaps %xmm0,-0x38($key_)
movaps -0x28($key_),%xmm14
movaps %xmm0,-0x28($key_)
movaps -0x18($key_),%xmm15
movaps %xmm0,-0x18($key_)
movaps %xmm0,0x00(%rsp)
movaps %xmm0,0x10(%rsp)
movaps %xmm0,0x20(%rsp)
movaps %xmm0,0x30(%rsp)
movaps %xmm0,0x40(%rsp)
movaps %xmm0,0x50(%rsp)
movaps %xmm0,0x60(%rsp)
movaps %xmm0,0x70(%rsp)
___
$code.=<<___;
mov -8($key_),%rbp
lea ($key_),%rsp
.Lctr32_epilogue:
ret
.size aesni_ctr32_encrypt_blocks,.-aesni_ctr32_encrypt_blocks
___
}
######################################################################
# void aesni_xts_[en|de]crypt(const char *inp,char *out,size_t len,
# const AES_KEY *key1, const AES_KEY *key2
# const unsigned char iv[16]);
#
{
my @tweak=map("%xmm$_",(10..15));
my ($twmask,$twres,$twtmp)=("%xmm8","%xmm9",@tweak[4]);
my ($key2,$ivp,$len_)=("%r8","%r9","%r9");
my $frame_size = 0x70 + ($win64?160:0);
my $key_ = "%rbp"; # override so that we can use %r11 as FP
$code.=<<___;
.globl aesni_xts_encrypt
.type aesni_xts_encrypt,\@function,6
.align 16
aesni_xts_encrypt:
lea (%rsp),%r11 # frame pointer
push %rbp
sub \$$frame_size,%rsp
and \$-16,%rsp # Linux kernel stack can be incorrectly seeded
___
$code.=<<___ if ($win64);
movaps %xmm6,-0xa8(%r11) # offload everything
movaps %xmm7,-0x98(%r11)
movaps %xmm8,-0x88(%r11)
movaps %xmm9,-0x78(%r11)
movaps %xmm10,-0x68(%r11)
movaps %xmm11,-0x58(%r11)
movaps %xmm12,-0x48(%r11)
movaps %xmm13,-0x38(%r11)
movaps %xmm14,-0x28(%r11)
movaps %xmm15,-0x18(%r11)
.Lxts_enc_body:
___
$code.=<<___;
movups ($ivp),$inout0 # load clear-text tweak
mov 240(%r8),$rounds # key2->rounds
mov 240($key),$rnds_ # key1->rounds
___
# generate the tweak
&aesni_generate1("enc",$key2,$rounds,$inout0);
$code.=<<___;
$movkey ($key),$rndkey0 # zero round key
mov $key,$key_ # backup $key
mov $rnds_,$rounds # backup $rounds
shl \$4,$rnds_
mov $len,$len_ # backup $len
and \$-16,$len
$movkey 16($key,$rnds_),$rndkey1 # last round key
movdqa .Lxts_magic(%rip),$twmask
movdqa $inout0,@tweak[5]
pshufd \$0x5f,$inout0,$twres
pxor $rndkey0,$rndkey1
___
# alternative tweak calculation algorithm is based on suggestions
# by Shay Gueron. psrad doesn't conflict with AES-NI instructions
# and should help in the future...
for ($i=0;$i<4;$i++) {
$code.=<<___;
movdqa $twres,$twtmp
paddd $twres,$twres
movdqa @tweak[5],@tweak[$i]
psrad \$31,$twtmp # broadcast upper bits
paddq @tweak[5],@tweak[5]
pand $twmask,$twtmp
pxor $rndkey0,@tweak[$i]
pxor $twtmp,@tweak[5]
___
}
$code.=<<___;
movdqa @tweak[5],@tweak[4]
psrad \$31,$twres
paddq @tweak[5],@tweak[5]
pand $twmask,$twres
pxor $rndkey0,@tweak[4]
pxor $twres,@tweak[5]
movaps $rndkey1,0x60(%rsp) # save round[0]^round[last]
sub \$16*6,$len
jc .Lxts_enc_short # if $len-=6*16 borrowed
mov \$16+96,$rounds
lea 32($key_,$rnds_),$key # end of key schedule
sub %r10,%rax # twisted $rounds
$movkey 16($key_),$rndkey1
mov %rax,%r10 # backup twisted $rounds
lea .Lxts_magic(%rip),%r8
jmp .Lxts_enc_grandloop
.align 32
.Lxts_enc_grandloop:
movdqu `16*0`($inp),$inout0 # load input
movdqa $rndkey0,$twmask
movdqu `16*1`($inp),$inout1
pxor @tweak[0],$inout0 # input^=tweak^round[0]
movdqu `16*2`($inp),$inout2
pxor @tweak[1],$inout1
aesenc $rndkey1,$inout0
movdqu `16*3`($inp),$inout3
pxor @tweak[2],$inout2
aesenc $rndkey1,$inout1
movdqu `16*4`($inp),$inout4
pxor @tweak[3],$inout3
aesenc $rndkey1,$inout2
movdqu `16*5`($inp),$inout5
pxor @tweak[5],$twmask # round[0]^=tweak[5]
movdqa 0x60(%rsp),$twres # load round[0]^round[last]
pxor @tweak[4],$inout4
aesenc $rndkey1,$inout3
$movkey 32($key_),$rndkey0
lea `16*6`($inp),$inp
pxor $twmask,$inout5
pxor $twres,@tweak[0] # calclulate tweaks^round[last]
aesenc $rndkey1,$inout4
pxor $twres,@tweak[1]
movdqa @tweak[0],`16*0`(%rsp) # put aside tweaks^round[last]
aesenc $rndkey1,$inout5
$movkey 48($key_),$rndkey1
pxor $twres,@tweak[2]
aesenc $rndkey0,$inout0
pxor $twres,@tweak[3]
movdqa @tweak[1],`16*1`(%rsp)
aesenc $rndkey0,$inout1
pxor $twres,@tweak[4]
movdqa @tweak[2],`16*2`(%rsp)
aesenc $rndkey0,$inout2
aesenc $rndkey0,$inout3
pxor $twres,$twmask
movdqa @tweak[4],`16*4`(%rsp)
aesenc $rndkey0,$inout4
aesenc $rndkey0,$inout5
$movkey 64($key_),$rndkey0
movdqa $twmask,`16*5`(%rsp)
pshufd \$0x5f,@tweak[5],$twres
jmp .Lxts_enc_loop6
.align 32
.Lxts_enc_loop6:
aesenc $rndkey1,$inout0
aesenc $rndkey1,$inout1
aesenc $rndkey1,$inout2
aesenc $rndkey1,$inout3
aesenc $rndkey1,$inout4
aesenc $rndkey1,$inout5
$movkey -64($key,%rax),$rndkey1
add \$32,%rax
aesenc $rndkey0,$inout0
aesenc $rndkey0,$inout1
aesenc $rndkey0,$inout2
aesenc $rndkey0,$inout3
aesenc $rndkey0,$inout4
aesenc $rndkey0,$inout5
$movkey -80($key,%rax),$rndkey0
jnz .Lxts_enc_loop6
movdqa (%r8),$twmask # start calculating next tweak
movdqa $twres,$twtmp
paddd $twres,$twres
aesenc $rndkey1,$inout0
paddq @tweak[5],@tweak[5]
psrad \$31,$twtmp
aesenc $rndkey1,$inout1
pand $twmask,$twtmp
$movkey ($key_),@tweak[0] # load round[0]
aesenc $rndkey1,$inout2
aesenc $rndkey1,$inout3
aesenc $rndkey1,$inout4
pxor $twtmp,@tweak[5]
movaps @tweak[0],@tweak[1] # copy round[0]
aesenc $rndkey1,$inout5
$movkey -64($key),$rndkey1
movdqa $twres,$twtmp
aesenc $rndkey0,$inout0
paddd $twres,$twres
pxor @tweak[5],@tweak[0]
aesenc $rndkey0,$inout1
psrad \$31,$twtmp
paddq @tweak[5],@tweak[5]
aesenc $rndkey0,$inout2
aesenc $rndkey0,$inout3
pand $twmask,$twtmp
movaps @tweak[1],@tweak[2]
aesenc $rndkey0,$inout4
pxor $twtmp,@tweak[5]
movdqa $twres,$twtmp
aesenc $rndkey0,$inout5
$movkey -48($key),$rndkey0
paddd $twres,$twres
aesenc $rndkey1,$inout0
pxor @tweak[5],@tweak[1]
psrad \$31,$twtmp
aesenc $rndkey1,$inout1
paddq @tweak[5],@tweak[5]
pand $twmask,$twtmp
aesenc $rndkey1,$inout2
aesenc $rndkey1,$inout3
movdqa @tweak[3],`16*3`(%rsp)
pxor $twtmp,@tweak[5]
aesenc $rndkey1,$inout4
movaps @tweak[2],@tweak[3]
movdqa $twres,$twtmp
aesenc $rndkey1,$inout5
$movkey -32($key),$rndkey1
paddd $twres,$twres
aesenc $rndkey0,$inout0
pxor @tweak[5],@tweak[2]
psrad \$31,$twtmp
aesenc $rndkey0,$inout1
paddq @tweak[5],@tweak[5]
pand $twmask,$twtmp
aesenc $rndkey0,$inout2
aesenc $rndkey0,$inout3
aesenc $rndkey0,$inout4
pxor $twtmp,@tweak[5]
movaps @tweak[3],@tweak[4]
aesenc $rndkey0,$inout5
movdqa $twres,$rndkey0
paddd $twres,$twres
aesenc $rndkey1,$inout0
pxor @tweak[5],@tweak[3]
psrad \$31,$rndkey0
aesenc $rndkey1,$inout1
paddq @tweak[5],@tweak[5]
pand $twmask,$rndkey0
aesenc $rndkey1,$inout2
aesenc $rndkey1,$inout3
pxor $rndkey0,@tweak[5]
$movkey ($key_),$rndkey0
aesenc $rndkey1,$inout4
aesenc $rndkey1,$inout5
$movkey 16($key_),$rndkey1
pxor @tweak[5],@tweak[4]
aesenclast `16*0`(%rsp),$inout0
psrad \$31,$twres
paddq @tweak[5],@tweak[5]
aesenclast `16*1`(%rsp),$inout1
aesenclast `16*2`(%rsp),$inout2
pand $twmask,$twres
mov %r10,%rax # restore $rounds
aesenclast `16*3`(%rsp),$inout3
aesenclast `16*4`(%rsp),$inout4
aesenclast `16*5`(%rsp),$inout5
pxor $twres,@tweak[5]
lea `16*6`($out),$out # $out+=6*16
movups $inout0,`-16*6`($out) # store 6 output blocks
movups $inout1,`-16*5`($out)
movups $inout2,`-16*4`($out)
movups $inout3,`-16*3`($out)
movups $inout4,`-16*2`($out)
movups $inout5,`-16*1`($out)
sub \$16*6,$len
jnc .Lxts_enc_grandloop # loop if $len-=6*16 didn't borrow
mov \$16+96,$rounds
sub $rnds_,$rounds
mov $key_,$key # restore $key
shr \$4,$rounds # restore original value
.Lxts_enc_short:
# at the point @tweak[0..5] are populated with tweak values
mov $rounds,$rnds_ # backup $rounds
pxor $rndkey0,@tweak[0]
add \$16*6,$len # restore real remaining $len
jz .Lxts_enc_done # done if ($len==0)
pxor $rndkey0,@tweak[1]
cmp \$0x20,$len
jb .Lxts_enc_one # $len is 1*16
pxor $rndkey0,@tweak[2]
je .Lxts_enc_two # $len is 2*16
pxor $rndkey0,@tweak[3]
cmp \$0x40,$len
jb .Lxts_enc_three # $len is 3*16
pxor $rndkey0,@tweak[4]
je .Lxts_enc_four # $len is 4*16
movdqu ($inp),$inout0 # $len is 5*16
movdqu 16*1($inp),$inout1
movdqu 16*2($inp),$inout2
pxor @tweak[0],$inout0
movdqu 16*3($inp),$inout3
pxor @tweak[1],$inout1
movdqu 16*4($inp),$inout4
lea 16*5($inp),$inp # $inp+=5*16
pxor @tweak[2],$inout2
pxor @tweak[3],$inout3
pxor @tweak[4],$inout4
pxor $inout5,$inout5
call _aesni_encrypt6
xorps @tweak[0],$inout0
movdqa @tweak[5],@tweak[0]
xorps @tweak[1],$inout1
xorps @tweak[2],$inout2
movdqu $inout0,($out) # store 5 output blocks
xorps @tweak[3],$inout3
movdqu $inout1,16*1($out)
xorps @tweak[4],$inout4
movdqu $inout2,16*2($out)
movdqu $inout3,16*3($out)
movdqu $inout4,16*4($out)
lea 16*5($out),$out # $out+=5*16
jmp .Lxts_enc_done
.align 16
.Lxts_enc_one:
movups ($inp),$inout0
lea 16*1($inp),$inp # inp+=1*16
xorps @tweak[0],$inout0
___
&aesni_generate1("enc",$key,$rounds);
$code.=<<___;
xorps @tweak[0],$inout0
movdqa @tweak[1],@tweak[0]
movups $inout0,($out) # store one output block
lea 16*1($out),$out # $out+=1*16
jmp .Lxts_enc_done
.align 16
.Lxts_enc_two:
movups ($inp),$inout0
movups 16($inp),$inout1
lea 32($inp),$inp # $inp+=2*16
xorps @tweak[0],$inout0
xorps @tweak[1],$inout1
call _aesni_encrypt2
xorps @tweak[0],$inout0
movdqa @tweak[2],@tweak[0]
xorps @tweak[1],$inout1
movups $inout0,($out) # store 2 output blocks
movups $inout1,16*1($out)
lea 16*2($out),$out # $out+=2*16
jmp .Lxts_enc_done
.align 16
.Lxts_enc_three:
movups ($inp),$inout0
movups 16*1($inp),$inout1
movups 16*2($inp),$inout2
lea 16*3($inp),$inp # $inp+=3*16
xorps @tweak[0],$inout0
xorps @tweak[1],$inout1
xorps @tweak[2],$inout2
call _aesni_encrypt3
xorps @tweak[0],$inout0
movdqa @tweak[3],@tweak[0]
xorps @tweak[1],$inout1
xorps @tweak[2],$inout2
movups $inout0,($out) # store 3 output blocks
movups $inout1,16*1($out)
movups $inout2,16*2($out)
lea 16*3($out),$out # $out+=3*16
jmp .Lxts_enc_done
.align 16
.Lxts_enc_four:
movups ($inp),$inout0
movups 16*1($inp),$inout1
movups 16*2($inp),$inout2
xorps @tweak[0],$inout0
movups 16*3($inp),$inout3
lea 16*4($inp),$inp # $inp+=4*16
xorps @tweak[1],$inout1
xorps @tweak[2],$inout2
xorps @tweak[3],$inout3
call _aesni_encrypt4
pxor @tweak[0],$inout0
movdqa @tweak[4],@tweak[0]
pxor @tweak[1],$inout1
pxor @tweak[2],$inout2
movdqu $inout0,($out) # store 4 output blocks
pxor @tweak[3],$inout3
movdqu $inout1,16*1($out)
movdqu $inout2,16*2($out)
movdqu $inout3,16*3($out)
lea 16*4($out),$out # $out+=4*16
jmp .Lxts_enc_done
.align 16
.Lxts_enc_done:
and \$15,$len_ # see if $len%16 is 0
jz .Lxts_enc_ret
mov $len_,$len
.Lxts_enc_steal:
movzb ($inp),%eax # borrow $rounds ...
movzb -16($out),%ecx # ... and $key
lea 1($inp),$inp
mov %al,-16($out)
mov %cl,0($out)
lea 1($out),$out
sub \$1,$len
jnz .Lxts_enc_steal
sub $len_,$out # rewind $out
mov $key_,$key # restore $key
mov $rnds_,$rounds # restore $rounds
movups -16($out),$inout0
xorps @tweak[0],$inout0
___
&aesni_generate1("enc",$key,$rounds);
$code.=<<___;
xorps @tweak[0],$inout0
movups $inout0,-16($out)
.Lxts_enc_ret:
xorps %xmm0,%xmm0 # clear register bank
pxor %xmm1,%xmm1
pxor %xmm2,%xmm2
pxor %xmm3,%xmm3
pxor %xmm4,%xmm4
pxor %xmm5,%xmm5
___
$code.=<<___ if (!$win64);
pxor %xmm6,%xmm6
pxor %xmm7,%xmm7
movaps %xmm0,0x00(%rsp) # clear stack
pxor %xmm8,%xmm8
movaps %xmm0,0x10(%rsp)
pxor %xmm9,%xmm9
movaps %xmm0,0x20(%rsp)
pxor %xmm10,%xmm10
movaps %xmm0,0x30(%rsp)
pxor %xmm11,%xmm11
movaps %xmm0,0x40(%rsp)
pxor %xmm12,%xmm12
movaps %xmm0,0x50(%rsp)
pxor %xmm13,%xmm13
movaps %xmm0,0x60(%rsp)
pxor %xmm14,%xmm14
pxor %xmm15,%xmm15
___
$code.=<<___ if ($win64);
movaps -0xa8(%r11),%xmm6
movaps %xmm0,-0xa8(%r11) # clear stack
movaps -0x98(%r11),%xmm7
movaps %xmm0,-0x98(%r11)
movaps -0x88(%r11),%xmm8
movaps %xmm0,-0x88(%r11)
movaps -0x78(%r11),%xmm9
movaps %xmm0,-0x78(%r11)
movaps -0x68(%r11),%xmm10
movaps %xmm0,-0x68(%r11)
movaps -0x58(%r11),%xmm11
movaps %xmm0,-0x58(%r11)
movaps -0x48(%r11),%xmm12
movaps %xmm0,-0x48(%r11)
movaps -0x38(%r11),%xmm13
movaps %xmm0,-0x38(%r11)
movaps -0x28(%r11),%xmm14
movaps %xmm0,-0x28(%r11)
movaps -0x18(%r11),%xmm15
movaps %xmm0,-0x18(%r11)
movaps %xmm0,0x00(%rsp)
movaps %xmm0,0x10(%rsp)
movaps %xmm0,0x20(%rsp)
movaps %xmm0,0x30(%rsp)
movaps %xmm0,0x40(%rsp)
movaps %xmm0,0x50(%rsp)
movaps %xmm0,0x60(%rsp)
___
$code.=<<___;
mov -8(%r11),%rbp
lea (%r11),%rsp
.Lxts_enc_epilogue:
ret
.size aesni_xts_encrypt,.-aesni_xts_encrypt
___
$code.=<<___;
.globl aesni_xts_decrypt
.type aesni_xts_decrypt,\@function,6
.align 16
aesni_xts_decrypt:
lea (%rsp),%r11 # frame pointer
push %rbp
sub \$$frame_size,%rsp
and \$-16,%rsp # Linux kernel stack can be incorrectly seeded
___
$code.=<<___ if ($win64);
movaps %xmm6,-0xa8(%r11) # offload everything
movaps %xmm7,-0x98(%r11)
movaps %xmm8,-0x88(%r11)
movaps %xmm9,-0x78(%r11)
movaps %xmm10,-0x68(%r11)
movaps %xmm11,-0x58(%r11)
movaps %xmm12,-0x48(%r11)
movaps %xmm13,-0x38(%r11)
movaps %xmm14,-0x28(%r11)
movaps %xmm15,-0x18(%r11)
.Lxts_dec_body:
___
$code.=<<___;
movups ($ivp),$inout0 # load clear-text tweak
mov 240($key2),$rounds # key2->rounds
mov 240($key),$rnds_ # key1->rounds
___
# generate the tweak
&aesni_generate1("enc",$key2,$rounds,$inout0);
$code.=<<___;
xor %eax,%eax # if ($len%16) len-=16;
test \$15,$len
setnz %al
shl \$4,%rax
sub %rax,$len
$movkey ($key),$rndkey0 # zero round key
mov $key,$key_ # backup $key
mov $rnds_,$rounds # backup $rounds
shl \$4,$rnds_
mov $len,$len_ # backup $len
and \$-16,$len
$movkey 16($key,$rnds_),$rndkey1 # last round key
movdqa .Lxts_magic(%rip),$twmask
movdqa $inout0,@tweak[5]
pshufd \$0x5f,$inout0,$twres
pxor $rndkey0,$rndkey1
___
for ($i=0;$i<4;$i++) {
$code.=<<___;
movdqa $twres,$twtmp
paddd $twres,$twres
movdqa @tweak[5],@tweak[$i]
psrad \$31,$twtmp # broadcast upper bits
paddq @tweak[5],@tweak[5]
pand $twmask,$twtmp
pxor $rndkey0,@tweak[$i]
pxor $twtmp,@tweak[5]
___
}
$code.=<<___;
movdqa @tweak[5],@tweak[4]
psrad \$31,$twres
paddq @tweak[5],@tweak[5]
pand $twmask,$twres
pxor $rndkey0,@tweak[4]
pxor $twres,@tweak[5]
movaps $rndkey1,0x60(%rsp) # save round[0]^round[last]
sub \$16*6,$len
jc .Lxts_dec_short # if $len-=6*16 borrowed
mov \$16+96,$rounds
lea 32($key_,$rnds_),$key # end of key schedule
sub %r10,%rax # twisted $rounds
$movkey 16($key_),$rndkey1
mov %rax,%r10 # backup twisted $rounds
lea .Lxts_magic(%rip),%r8
jmp .Lxts_dec_grandloop
.align 32
.Lxts_dec_grandloop:
movdqu `16*0`($inp),$inout0 # load input
movdqa $rndkey0,$twmask
movdqu `16*1`($inp),$inout1
pxor @tweak[0],$inout0 # intput^=tweak^round[0]
movdqu `16*2`($inp),$inout2
pxor @tweak[1],$inout1
aesdec $rndkey1,$inout0
movdqu `16*3`($inp),$inout3
pxor @tweak[2],$inout2
aesdec $rndkey1,$inout1
movdqu `16*4`($inp),$inout4
pxor @tweak[3],$inout3
aesdec $rndkey1,$inout2
movdqu `16*5`($inp),$inout5
pxor @tweak[5],$twmask # round[0]^=tweak[5]
movdqa 0x60(%rsp),$twres # load round[0]^round[last]
pxor @tweak[4],$inout4
aesdec $rndkey1,$inout3
$movkey 32($key_),$rndkey0
lea `16*6`($inp),$inp
pxor $twmask,$inout5
pxor $twres,@tweak[0] # calclulate tweaks^round[last]
aesdec $rndkey1,$inout4
pxor $twres,@tweak[1]
movdqa @tweak[0],`16*0`(%rsp) # put aside tweaks^last round key
aesdec $rndkey1,$inout5
$movkey 48($key_),$rndkey1
pxor $twres,@tweak[2]
aesdec $rndkey0,$inout0
pxor $twres,@tweak[3]
movdqa @tweak[1],`16*1`(%rsp)
aesdec $rndkey0,$inout1
pxor $twres,@tweak[4]
movdqa @tweak[2],`16*2`(%rsp)
aesdec $rndkey0,$inout2
aesdec $rndkey0,$inout3
pxor $twres,$twmask
movdqa @tweak[4],`16*4`(%rsp)
aesdec $rndkey0,$inout4
aesdec $rndkey0,$inout5
$movkey 64($key_),$rndkey0
movdqa $twmask,`16*5`(%rsp)
pshufd \$0x5f,@tweak[5],$twres
jmp .Lxts_dec_loop6
.align 32
.Lxts_dec_loop6:
aesdec $rndkey1,$inout0
aesdec $rndkey1,$inout1
aesdec $rndkey1,$inout2
aesdec $rndkey1,$inout3
aesdec $rndkey1,$inout4
aesdec $rndkey1,$inout5
$movkey -64($key,%rax),$rndkey1
add \$32,%rax
aesdec $rndkey0,$inout0
aesdec $rndkey0,$inout1
aesdec $rndkey0,$inout2
aesdec $rndkey0,$inout3
aesdec $rndkey0,$inout4
aesdec $rndkey0,$inout5
$movkey -80($key,%rax),$rndkey0
jnz .Lxts_dec_loop6
movdqa (%r8),$twmask # start calculating next tweak
movdqa $twres,$twtmp
paddd $twres,$twres
aesdec $rndkey1,$inout0
paddq @tweak[5],@tweak[5]
psrad \$31,$twtmp
aesdec $rndkey1,$inout1
pand $twmask,$twtmp
$movkey ($key_),@tweak[0] # load round[0]
aesdec $rndkey1,$inout2
aesdec $rndkey1,$inout3
aesdec $rndkey1,$inout4
pxor $twtmp,@tweak[5]
movaps @tweak[0],@tweak[1] # copy round[0]
aesdec $rndkey1,$inout5
$movkey -64($key),$rndkey1
movdqa $twres,$twtmp
aesdec $rndkey0,$inout0
paddd $twres,$twres
pxor @tweak[5],@tweak[0]
aesdec $rndkey0,$inout1
psrad \$31,$twtmp
paddq @tweak[5],@tweak[5]
aesdec $rndkey0,$inout2
aesdec $rndkey0,$inout3
pand $twmask,$twtmp
movaps @tweak[1],@tweak[2]
aesdec $rndkey0,$inout4
pxor $twtmp,@tweak[5]
movdqa $twres,$twtmp
aesdec $rndkey0,$inout5
$movkey -48($key),$rndkey0
paddd $twres,$twres
aesdec $rndkey1,$inout0
pxor @tweak[5],@tweak[1]
psrad \$31,$twtmp
aesdec $rndkey1,$inout1
paddq @tweak[5],@tweak[5]
pand $twmask,$twtmp
aesdec $rndkey1,$inout2
aesdec $rndkey1,$inout3
movdqa @tweak[3],`16*3`(%rsp)
pxor $twtmp,@tweak[5]
aesdec $rndkey1,$inout4
movaps @tweak[2],@tweak[3]
movdqa $twres,$twtmp
aesdec $rndkey1,$inout5
$movkey -32($key),$rndkey1
paddd $twres,$twres
aesdec $rndkey0,$inout0
pxor @tweak[5],@tweak[2]
psrad \$31,$twtmp
aesdec $rndkey0,$inout1
paddq @tweak[5],@tweak[5]
pand $twmask,$twtmp
aesdec $rndkey0,$inout2
aesdec $rndkey0,$inout3
aesdec $rndkey0,$inout4
pxor $twtmp,@tweak[5]
movaps @tweak[3],@tweak[4]
aesdec $rndkey0,$inout5
movdqa $twres,$rndkey0
paddd $twres,$twres
aesdec $rndkey1,$inout0
pxor @tweak[5],@tweak[3]
psrad \$31,$rndkey0
aesdec $rndkey1,$inout1
paddq @tweak[5],@tweak[5]
pand $twmask,$rndkey0
aesdec $rndkey1,$inout2
aesdec $rndkey1,$inout3
pxor $rndkey0,@tweak[5]
$movkey ($key_),$rndkey0
aesdec $rndkey1,$inout4
aesdec $rndkey1,$inout5
$movkey 16($key_),$rndkey1
pxor @tweak[5],@tweak[4]
aesdeclast `16*0`(%rsp),$inout0
psrad \$31,$twres
paddq @tweak[5],@tweak[5]
aesdeclast `16*1`(%rsp),$inout1
aesdeclast `16*2`(%rsp),$inout2
pand $twmask,$twres
mov %r10,%rax # restore $rounds
aesdeclast `16*3`(%rsp),$inout3
aesdeclast `16*4`(%rsp),$inout4
aesdeclast `16*5`(%rsp),$inout5
pxor $twres,@tweak[5]
lea `16*6`($out),$out # $out+=6*16
movups $inout0,`-16*6`($out) # store 6 output blocks
movups $inout1,`-16*5`($out)
movups $inout2,`-16*4`($out)
movups $inout3,`-16*3`($out)
movups $inout4,`-16*2`($out)
movups $inout5,`-16*1`($out)
sub \$16*6,$len
jnc .Lxts_dec_grandloop # loop if $len-=6*16 didn't borrow
mov \$16+96,$rounds
sub $rnds_,$rounds
mov $key_,$key # restore $key
shr \$4,$rounds # restore original value
.Lxts_dec_short:
# at the point @tweak[0..5] are populated with tweak values
mov $rounds,$rnds_ # backup $rounds
pxor $rndkey0,@tweak[0]
pxor $rndkey0,@tweak[1]
add \$16*6,$len # restore real remaining $len
jz .Lxts_dec_done # done if ($len==0)
pxor $rndkey0,@tweak[2]
cmp \$0x20,$len
jb .Lxts_dec_one # $len is 1*16
pxor $rndkey0,@tweak[3]
je .Lxts_dec_two # $len is 2*16
pxor $rndkey0,@tweak[4]
cmp \$0x40,$len
jb .Lxts_dec_three # $len is 3*16
je .Lxts_dec_four # $len is 4*16
movdqu ($inp),$inout0 # $len is 5*16
movdqu 16*1($inp),$inout1
movdqu 16*2($inp),$inout2
pxor @tweak[0],$inout0
movdqu 16*3($inp),$inout3
pxor @tweak[1],$inout1
movdqu 16*4($inp),$inout4
lea 16*5($inp),$inp # $inp+=5*16
pxor @tweak[2],$inout2
pxor @tweak[3],$inout3
pxor @tweak[4],$inout4
call _aesni_decrypt6
xorps @tweak[0],$inout0
xorps @tweak[1],$inout1
xorps @tweak[2],$inout2
movdqu $inout0,($out) # store 5 output blocks
xorps @tweak[3],$inout3
movdqu $inout1,16*1($out)
xorps @tweak[4],$inout4
movdqu $inout2,16*2($out)
pxor $twtmp,$twtmp
movdqu $inout3,16*3($out)
pcmpgtd @tweak[5],$twtmp
movdqu $inout4,16*4($out)
lea 16*5($out),$out # $out+=5*16
pshufd \$0x13,$twtmp,@tweak[1] # $twres
and \$15,$len_
jz .Lxts_dec_ret
movdqa @tweak[5],@tweak[0]
paddq @tweak[5],@tweak[5] # psllq 1,$tweak
pand $twmask,@tweak[1] # isolate carry and residue
pxor @tweak[5],@tweak[1]
jmp .Lxts_dec_done2
.align 16
.Lxts_dec_one:
movups ($inp),$inout0
lea 16*1($inp),$inp # $inp+=1*16
xorps @tweak[0],$inout0
___
&aesni_generate1("dec",$key,$rounds);
$code.=<<___;
xorps @tweak[0],$inout0
movdqa @tweak[1],@tweak[0]
movups $inout0,($out) # store one output block
movdqa @tweak[2],@tweak[1]
lea 16*1($out),$out # $out+=1*16
jmp .Lxts_dec_done
.align 16
.Lxts_dec_two:
movups ($inp),$inout0
movups 16($inp),$inout1
lea 32($inp),$inp # $inp+=2*16
xorps @tweak[0],$inout0
xorps @tweak[1],$inout1
call _aesni_decrypt2
xorps @tweak[0],$inout0
movdqa @tweak[2],@tweak[0]
xorps @tweak[1],$inout1
movdqa @tweak[3],@tweak[1]
movups $inout0,($out) # store 2 output blocks
movups $inout1,16*1($out)
lea 16*2($out),$out # $out+=2*16
jmp .Lxts_dec_done
.align 16
.Lxts_dec_three:
movups ($inp),$inout0
movups 16*1($inp),$inout1
movups 16*2($inp),$inout2
lea 16*3($inp),$inp # $inp+=3*16
xorps @tweak[0],$inout0
xorps @tweak[1],$inout1
xorps @tweak[2],$inout2
call _aesni_decrypt3
xorps @tweak[0],$inout0
movdqa @tweak[3],@tweak[0]
xorps @tweak[1],$inout1
movdqa @tweak[4],@tweak[1]
xorps @tweak[2],$inout2
movups $inout0,($out) # store 3 output blocks
movups $inout1,16*1($out)
movups $inout2,16*2($out)
lea 16*3($out),$out # $out+=3*16
jmp .Lxts_dec_done
.align 16
.Lxts_dec_four:
movups ($inp),$inout0
movups 16*1($inp),$inout1
movups 16*2($inp),$inout2
xorps @tweak[0],$inout0
movups 16*3($inp),$inout3
lea 16*4($inp),$inp # $inp+=4*16
xorps @tweak[1],$inout1
xorps @tweak[2],$inout2
xorps @tweak[3],$inout3
call _aesni_decrypt4
pxor @tweak[0],$inout0
movdqa @tweak[4],@tweak[0]
pxor @tweak[1],$inout1
movdqa @tweak[5],@tweak[1]
pxor @tweak[2],$inout2
movdqu $inout0,($out) # store 4 output blocks
pxor @tweak[3],$inout3
movdqu $inout1,16*1($out)
movdqu $inout2,16*2($out)
movdqu $inout3,16*3($out)
lea 16*4($out),$out # $out+=4*16
jmp .Lxts_dec_done
.align 16
.Lxts_dec_done:
and \$15,$len_ # see if $len%16 is 0
jz .Lxts_dec_ret
.Lxts_dec_done2:
mov $len_,$len
mov $key_,$key # restore $key
mov $rnds_,$rounds # restore $rounds
movups ($inp),$inout0
xorps @tweak[1],$inout0
___
&aesni_generate1("dec",$key,$rounds);
$code.=<<___;
xorps @tweak[1],$inout0
movups $inout0,($out)
.Lxts_dec_steal:
movzb 16($inp),%eax # borrow $rounds ...
movzb ($out),%ecx # ... and $key
lea 1($inp),$inp
mov %al,($out)
mov %cl,16($out)
lea 1($out),$out
sub \$1,$len
jnz .Lxts_dec_steal
sub $len_,$out # rewind $out
mov $key_,$key # restore $key
mov $rnds_,$rounds # restore $rounds
movups ($out),$inout0
xorps @tweak[0],$inout0
___
&aesni_generate1("dec",$key,$rounds);
$code.=<<___;
xorps @tweak[0],$inout0
movups $inout0,($out)
.Lxts_dec_ret:
xorps %xmm0,%xmm0 # clear register bank
pxor %xmm1,%xmm1
pxor %xmm2,%xmm2
pxor %xmm3,%xmm3
pxor %xmm4,%xmm4
pxor %xmm5,%xmm5
___
$code.=<<___ if (!$win64);
pxor %xmm6,%xmm6
pxor %xmm7,%xmm7
movaps %xmm0,0x00(%rsp) # clear stack
pxor %xmm8,%xmm8
movaps %xmm0,0x10(%rsp)
pxor %xmm9,%xmm9
movaps %xmm0,0x20(%rsp)
pxor %xmm10,%xmm10
movaps %xmm0,0x30(%rsp)
pxor %xmm11,%xmm11
movaps %xmm0,0x40(%rsp)
pxor %xmm12,%xmm12
movaps %xmm0,0x50(%rsp)
pxor %xmm13,%xmm13
movaps %xmm0,0x60(%rsp)
pxor %xmm14,%xmm14
pxor %xmm15,%xmm15
___
$code.=<<___ if ($win64);
movaps -0xa8(%r11),%xmm6
movaps %xmm0,-0xa8(%r11) # clear stack
movaps -0x98(%r11),%xmm7
movaps %xmm0,-0x98(%r11)
movaps -0x88(%r11),%xmm8
movaps %xmm0,-0x88(%r11)
movaps -0x78(%r11),%xmm9
movaps %xmm0,-0x78(%r11)
movaps -0x68(%r11),%xmm10
movaps %xmm0,-0x68(%r11)
movaps -0x58(%r11),%xmm11
movaps %xmm0,-0x58(%r11)
movaps -0x48(%r11),%xmm12
movaps %xmm0,-0x48(%r11)
movaps -0x38(%r11),%xmm13
movaps %xmm0,-0x38(%r11)
movaps -0x28(%r11),%xmm14
movaps %xmm0,-0x28(%r11)
movaps -0x18(%r11),%xmm15
movaps %xmm0,-0x18(%r11)
movaps %xmm0,0x00(%rsp)
movaps %xmm0,0x10(%rsp)
movaps %xmm0,0x20(%rsp)
movaps %xmm0,0x30(%rsp)
movaps %xmm0,0x40(%rsp)
movaps %xmm0,0x50(%rsp)
movaps %xmm0,0x60(%rsp)
___
$code.=<<___;
mov -8(%r11),%rbp
lea (%r11),%rsp
.Lxts_dec_epilogue:
ret
.size aesni_xts_decrypt,.-aesni_xts_decrypt
___
}
######################################################################
# void aesni_ocb_[en|de]crypt(const char *inp, char *out, size_t blocks,
# const AES_KEY *key, unsigned int start_block_num,
# unsigned char offset_i[16], const unsigned char L_[][16],
# unsigned char checksum[16]);
#
{
my @offset=map("%xmm$_",(10..15));
my ($checksum,$rndkey0l)=("%xmm8","%xmm9");
my ($block_num,$offset_p)=("%r8","%r9"); # 5th and 6th arguments
my ($L_p,$checksum_p) = ("%rbx","%rbp");
my ($i1,$i3,$i5) = ("%r12","%r13","%r14");
my $seventh_arg = $win64 ? 56 : 8;
my $blocks = $len;
$code.=<<___;
.globl aesni_ocb_encrypt
.type aesni_ocb_encrypt,\@function,6
.align 32
aesni_ocb_encrypt:
lea (%rsp),%rax
push %rbx
push %rbp
push %r12
push %r13
push %r14
___
$code.=<<___ if ($win64);
lea -0xa0(%rsp),%rsp
movaps %xmm6,0x00(%rsp) # offload everything
movaps %xmm7,0x10(%rsp)
movaps %xmm8,0x20(%rsp)
movaps %xmm9,0x30(%rsp)
movaps %xmm10,0x40(%rsp)
movaps %xmm11,0x50(%rsp)
movaps %xmm12,0x60(%rsp)
movaps %xmm13,0x70(%rsp)
movaps %xmm14,0x80(%rsp)
movaps %xmm15,0x90(%rsp)
.Locb_enc_body:
___
$code.=<<___;
mov $seventh_arg(%rax),$L_p # 7th argument
mov $seventh_arg+8(%rax),$checksum_p# 8th argument
mov 240($key),$rnds_
mov $key,$key_
shl \$4,$rnds_
$movkey ($key),$rndkey0l # round[0]
$movkey 16($key,$rnds_),$rndkey1 # round[last]
movdqu ($offset_p),@offset[5] # load last offset_i
pxor $rndkey1,$rndkey0l # round[0] ^ round[last]
pxor $rndkey1,@offset[5] # offset_i ^ round[last]
mov \$16+32,$rounds
lea 32($key_,$rnds_),$key
$movkey 16($key_),$rndkey1 # round[1]
sub %r10,%rax # twisted $rounds
mov %rax,%r10 # backup twisted $rounds
movdqu ($L_p),@offset[0] # L_0 for all odd-numbered blocks
movdqu ($checksum_p),$checksum # load checksum
test \$1,$block_num # is first block number odd?
jnz .Locb_enc_odd
bsf $block_num,$i1
add \$1,$block_num
shl \$4,$i1
movdqu ($L_p,$i1),$inout5 # borrow
movdqu ($inp),$inout0
lea 16($inp),$inp
call __ocb_encrypt1
movdqa $inout5,@offset[5]
movups $inout0,($out)
lea 16($out),$out
sub \$1,$blocks
jz .Locb_enc_done
.Locb_enc_odd:
lea 1($block_num),$i1 # even-numbered blocks
lea 3($block_num),$i3
lea 5($block_num),$i5
lea 6($block_num),$block_num
bsf $i1,$i1 # ntz(block)
bsf $i3,$i3
bsf $i5,$i5
shl \$4,$i1 # ntz(block) -> table offset
shl \$4,$i3
shl \$4,$i5
sub \$6,$blocks
jc .Locb_enc_short
jmp .Locb_enc_grandloop
.align 32
.Locb_enc_grandloop:
movdqu `16*0`($inp),$inout0 # load input
movdqu `16*1`($inp),$inout1
movdqu `16*2`($inp),$inout2
movdqu `16*3`($inp),$inout3
movdqu `16*4`($inp),$inout4
movdqu `16*5`($inp),$inout5
lea `16*6`($inp),$inp
call __ocb_encrypt6
movups $inout0,`16*0`($out) # store output
movups $inout1,`16*1`($out)
movups $inout2,`16*2`($out)
movups $inout3,`16*3`($out)
movups $inout4,`16*4`($out)
movups $inout5,`16*5`($out)
lea `16*6`($out),$out
sub \$6,$blocks
jnc .Locb_enc_grandloop
.Locb_enc_short:
add \$6,$blocks
jz .<