diff options
| author | marha <marha@users.sourceforge.net> | 2010-03-29 17:08:02 +0000 |
|---|---|---|
| committer | marha <marha@users.sourceforge.net> | 2010-03-29 17:08:02 +0000 |
| commit | 15272ab4ed1e6250412fccd48200ed9eae59608f (patch) | |
| tree | a5996ea67966a778a16565f19dfc2e7c7f49b376 /openssl/crypto/aes/asm/aes-586.pl | |
| parent | 3827301b2ea5a45ac009c3bf9f08586ff40b8506 (diff) | |
| download | vcxsrv-15272ab4ed1e6250412fccd48200ed9eae59608f.tar.gz vcxsrv-15272ab4ed1e6250412fccd48200ed9eae59608f.tar.bz2 vcxsrv-15272ab4ed1e6250412fccd48200ed9eae59608f.zip | |
Updated to openssl 1.0.0
Diffstat (limited to 'openssl/crypto/aes/asm/aes-586.pl')
| -rw-r--r-- | openssl/crypto/aes/asm/aes-586.pl | 2401 |
1 files changed, 1924 insertions, 477 deletions
diff --git a/openssl/crypto/aes/asm/aes-586.pl b/openssl/crypto/aes/asm/aes-586.pl index 3bc46a968..aab40e6f1 100644 --- a/openssl/crypto/aes/asm/aes-586.pl +++ b/openssl/crypto/aes/asm/aes-586.pl @@ -2,11 +2,12 @@ # # ==================================================================== # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL -# project. Rights for redistribution and usage in source and binary -# forms are granted according to the OpenSSL license. +# 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/. # ==================================================================== # -# Version 3.6. +# Version 4.3. # # You might fail to appreciate this module performance from the first # try. If compared to "vanilla" linux-ia32-icc target, i.e. considered @@ -81,11 +82,117 @@ # AMD K8 20 19 # PIII 25 23 # Pentium 81 78 - -push(@INC,"perlasm","../../perlasm"); +# +# Version 3.7 reimplements outer rounds as "compact." Meaning that +# first and last rounds reference compact 256 bytes S-box. This means +# that first round consumes a lot more CPU cycles and that encrypt +# and decrypt performance becomes asymmetric. Encrypt performance +# drops by 10-12%, while decrypt - by 20-25%:-( 256 bytes S-box is +# aggressively pre-fetched. +# +# Version 4.0 effectively rolls back to 3.6 and instead implements +# additional set of functions, _[x86|sse]_AES_[en|de]crypt_compact, +# which use exclusively 256 byte S-box. These functions are to be +# called in modes not concealing plain text, such as ECB, or when +# we're asked to process smaller amount of data [or unconditionally +# on hyper-threading CPU]. Currently it's called unconditionally from +# AES_[en|de]crypt, which affects all modes, but CBC. CBC routine +# still needs to be modified to switch between slower and faster +# mode when appropriate... But in either case benchmark landscape +# changes dramatically and below numbers are CPU cycles per processed +# byte for 128-bit key. +# +# ECB encrypt ECB decrypt CBC large chunk +# P4 56[60] 84[100] 23 +# AMD K8 48[44] 70[79] 18 +# PIII 41[50] 61[91] 24 +# Core 2 32[38] 45[70] 18.5 +# Pentium 120 160 77 +# +# Version 4.1 switches to compact S-box even in key schedule setup. +# +# Version 4.2 prefetches compact S-box in every SSE round or in other +# words every cache-line is *guaranteed* to be accessed within ~50 +# cycles window. Why just SSE? Because it's needed on hyper-threading +# CPU! Which is also why it's prefetched with 64 byte stride. Best +# part is that it has no negative effect on performance:-) +# +# Version 4.3 implements switch between compact and non-compact block +# functions in AES_cbc_encrypt depending on how much data was asked +# to be processed in one stroke. +# +###################################################################### +# Timing attacks are classified in two classes: synchronous when +# attacker consciously initiates cryptographic operation and collects +# timing data of various character afterwards, and asynchronous when +# malicious code is executed on same CPU simultaneously with AES, +# instruments itself and performs statistical analysis of this data. +# +# As far as synchronous attacks go the root to the AES timing +# vulnerability is twofold. Firstly, of 256 S-box elements at most 160 +# are referred to in single 128-bit block operation. Well, in C +# implementation with 4 distinct tables it's actually as little as 40 +# references per 256 elements table, but anyway... Secondly, even +# though S-box elements are clustered into smaller amount of cache- +# lines, smaller than 160 and even 40, it turned out that for certain +# plain-text pattern[s] or simply put chosen plain-text and given key +# few cache-lines remain unaccessed during block operation. Now, if +# attacker can figure out this access pattern, he can deduct the key +# [or at least part of it]. The natural way to mitigate this kind of +# attacks is to minimize the amount of cache-lines in S-box and/or +# prefetch them to ensure that every one is accessed for more uniform +# timing. But note that *if* plain-text was concealed in such way that +# input to block function is distributed *uniformly*, then attack +# wouldn't apply. Now note that some encryption modes, most notably +# CBC, do mask the plain-text in this exact way [secure cipher output +# is distributed uniformly]. Yes, one still might find input that +# would reveal the information about given key, but if amount of +# candidate inputs to be tried is larger than amount of possible key +# combinations then attack becomes infeasible. This is why revised +# AES_cbc_encrypt "dares" to switch to larger S-box when larger chunk +# of data is to be processed in one stroke. The current size limit of +# 512 bytes is chosen to provide same [diminishigly low] probability +# for cache-line to remain untouched in large chunk operation with +# large S-box as for single block operation with compact S-box and +# surely needs more careful consideration... +# +# As for asynchronous attacks. There are two flavours: attacker code +# being interleaved with AES on hyper-threading CPU at *instruction* +# level, and two processes time sharing single core. As for latter. +# Two vectors. 1. Given that attacker process has higher priority, +# yield execution to process performing AES just before timer fires +# off the scheduler, immediately regain control of CPU and analyze the +# cache state. For this attack to be efficient attacker would have to +# effectively slow down the operation by several *orders* of magnitute, +# by ratio of time slice to duration of handful of AES rounds, which +# unlikely to remain unnoticed. Not to mention that this also means +# that he would spend correspondigly more time to collect enough +# statistical data to mount the attack. It's probably appropriate to +# say that if adeversary reckons that this attack is beneficial and +# risks to be noticed, you probably have larger problems having him +# mere opportunity. In other words suggested code design expects you +# to preclude/mitigate this attack by overall system security design. +# 2. Attacker manages to make his code interrupt driven. In order for +# this kind of attack to be feasible, interrupt rate has to be high +# enough, again comparable to duration of handful of AES rounds. But +# is there interrupt source of such rate? Hardly, not even 1Gbps NIC +# generates interrupts at such raging rate... +# +# And now back to the former, hyper-threading CPU or more specifically +# Intel P4. Recall that asynchronous attack implies that malicious +# code instruments itself. And naturally instrumentation granularity +# has be noticeably lower than duration of codepath accessing S-box. +# Given that all cache-lines are accessed during that time that is. +# Current implementation accesses *all* cache-lines within ~50 cycles +# window, which is actually *less* than RDTSC latency on Intel P4! + +$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; +push(@INC,"${dir}","${dir}../../perlasm"); require "x86asm.pl"; -&asm_init($ARGV[0],"aes-586.pl",$ARGV[$#ARGV] eq "386"); +&asm_init($ARGV[0],"aes-586.pl",$x86only = $ARGV[$#ARGV] eq "386"); +&static_label("AES_Te"); +&static_label("AES_Td"); $s0="eax"; $s1="ebx"; @@ -93,21 +200,36 @@ $s2="ecx"; $s3="edx"; $key="edi"; $acc="esi"; +$tbl="ebp"; + +# stack frame layout in _[x86|sse]_AES_* routines, frame is allocated +# by caller +$__ra=&DWP(0,"esp"); # return address +$__s0=&DWP(4,"esp"); # s0 backing store +$__s1=&DWP(8,"esp"); # s1 backing store +$__s2=&DWP(12,"esp"); # s2 backing store +$__s3=&DWP(16,"esp"); # s3 backing store +$__key=&DWP(20,"esp"); # pointer to key schedule +$__end=&DWP(24,"esp"); # pointer to end of key schedule +$__tbl=&DWP(28,"esp"); # %ebp backing store + +# stack frame layout in AES_[en|crypt] routines, which differs from +# above by 4 and overlaps by %ebp backing store +$_tbl=&DWP(24,"esp"); +$_esp=&DWP(28,"esp"); -$compromise=0; # $compromise=128 abstains from copying key - # schedule to stack when encrypting inputs - # shorter than 128 bytes at the cost of - # risksing aliasing with S-boxes. In return - # you get way better, up to +70%, small block - # performance. +sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } } + +$speed_limit=512; # chunks smaller than $speed_limit are + # processed with compact routine in CBC mode $small_footprint=1; # $small_footprint=1 code is ~5% slower [on # recent µ-archs], but ~5 times smaller! # I favor compact code to minimize cache # contention and in hope to "collect" 5% back # in real-life applications... + $vertical_spin=0; # shift "verticaly" defaults to 0, because of # its proof-of-concept status... - # Note that there is no decvert(), as well as last encryption round is # performed with "horizontal" shifts. This is because this "vertical" # implementation [one which groups shifts on a given $s[i] to form a @@ -170,17 +292,484 @@ sub encvert() &movz ($v0,&HB($v1)); &and ($v1,0xFF); &xor ($s[3],&DWP(2,$te,$v1,8)); # s1>>16 - &mov ($key,&DWP(12,"esp")); # reincarnate v1 as key + &mov ($key,$__key); # reincarnate v1 as key &xor ($s[2],&DWP(1,$te,$v0,8)); # s1>>24 } +# Another experimental routine, which features "horizontal spin," but +# eliminates one reference to stack. Strangely enough runs slower... +sub enchoriz() +{ my $v0 = $key, $v1 = $acc; + + &movz ($v0,&LB($s0)); # 3, 2, 1, 0* + &rotr ($s2,8); # 8,11,10, 9 + &mov ($v1,&DWP(0,$te,$v0,8)); # 0 + &movz ($v0,&HB($s1)); # 7, 6, 5*, 4 + &rotr ($s3,16); # 13,12,15,14 + &xor ($v1,&DWP(3,$te,$v0,8)); # 5 + &movz ($v0,&HB($s2)); # 8,11,10*, 9 + &rotr ($s0,16); # 1, 0, 3, 2 + &xor ($v1,&DWP(2,$te,$v0,8)); # 10 + &movz ($v0,&HB($s3)); # 13,12,15*,14 + &xor ($v1,&DWP(1,$te,$v0,8)); # 15, t[0] collected + &mov ($__s0,$v1); # t[0] saved + + &movz ($v0,&LB($s1)); # 7, 6, 5, 4* + &shr ($s1,16); # -, -, 7, 6 + &mov ($v1,&DWP(0,$te,$v0,8)); # 4 + &movz ($v0,&LB($s3)); # 13,12,15,14* + &xor ($v1,&DWP(2,$te,$v0,8)); # 14 + &movz ($v0,&HB($s0)); # 1, 0, 3*, 2 + &and ($s3,0xffff0000); # 13,12, -, - + &xor ($v1,&DWP(1,$te,$v0,8)); # 3 + &movz ($v0,&LB($s2)); # 8,11,10, 9* + &or ($s3,$s1); # 13,12, 7, 6 + &xor ($v1,&DWP(3,$te,$v0,8)); # 9, t[1] collected + &mov ($s1,$v1); # s[1]=t[1] + + &movz ($v0,&LB($s0)); # 1, 0, 3, 2* + &shr ($s2,16); # -, -, 8,11 + &mov ($v1,&DWP(2,$te,$v0,8)); # 2 + &movz ($v0,&HB($s3)); # 13,12, 7*, 6 + &xor ($v1,&DWP(1,$te,$v0,8)); # 7 + &movz ($v0,&HB($s2)); # -, -, 8*,11 + &xor ($v1,&DWP(0,$te,$v0,8)); # 8 + &mov ($v0,$s3); + &shr ($v0,24); # 13 + &xor ($v1,&DWP(3,$te,$v0,8)); # 13, t[2] collected + + &movz ($v0,&LB($s2)); # -, -, 8,11* + &shr ($s0,24); # 1* + &mov ($s2,&DWP(1,$te,$v0,8)); # 11 + &xor ($s2,&DWP(3,$te,$s0,8)); # 1 + &mov ($s0,$__s0); # s[0]=t[0] + &movz ($v0,&LB($s3)); # 13,12, 7, 6* + &shr ($s3,16); # , ,13,12 + &xor ($s2,&DWP(2,$te,$v0,8)); # 6 + &mov ($key,$__key); # reincarnate v0 as key + &and ($s3,0xff); # , ,13,12* + &mov ($s3,&DWP(0,$te,$s3,8)); # 12 + &xor ($s3,$s2); # s[2]=t[3] collected + &mov ($s2,$v1); # s[2]=t[2] +} + +# More experimental code... SSE one... Even though this one eliminates +# *all* references to stack, it's not faster... +sub sse_encbody() +{ + &movz ($acc,&LB("eax")); # 0 + &mov ("ecx",&DWP(0,$tbl,$acc,8)); # 0 + &pshufw ("mm2","mm0",0x0d); # 7, 6, 3, 2 + &movz ("edx",&HB("eax")); # 1 + &mov ("edx",&DWP(3,$tbl,"edx",8)); # 1 + &shr ("eax",16); # 5, 4 + + &movz ($acc,&LB("ebx")); # 10 + &xor ("ecx",&DWP(2,$tbl,$acc,8)); # 10 + &pshufw ("mm6","mm4",0x08); # 13,12, 9, 8 + &movz ($acc,&HB("ebx")); # 11 + &xor ("edx",&DWP(1,$tbl,$acc,8)); # 11 + &shr ("ebx",16); # 15,14 + + &movz ($acc,&HB("eax")); # 5 + &xor ("ecx",&DWP(3,$tbl,$acc,8)); # 5 + &movq ("mm3",QWP(16,$key)); + &movz ($acc,&HB("ebx")); # 15 + &xor ("ecx",&DWP(1,$tbl,$acc,8)); # 15 + &movd ("mm0","ecx"); # t[0] collected + + &movz ($acc,&LB("eax")); # 4 + &mov ("ecx",&DWP(0,$tbl,$acc,8)); # 4 + &movd ("eax","mm2"); # 7, 6, 3, 2 + &movz ($acc,&LB("ebx")); # 14 + &xor ("ecx",&DWP(2,$tbl,$acc,8)); # 14 + &movd ("ebx","mm6"); # 13,12, 9, 8 + + &movz ($acc,&HB("eax")); # 3 + &xor ("ecx",&DWP(1,$tbl,$acc,8)); # 3 + &movz ($acc,&HB("ebx")); # 9 + &xor ("ecx",&DWP(3,$tbl,$acc,8)); # 9 + &movd ("mm1","ecx"); # t[1] collected + + &movz ($acc,&LB("eax")); # 2 + &mov ("ecx",&DWP(2,$tbl,$acc,8)); # 2 + &shr ("eax",16); # 7, 6 + &punpckldq ("mm0","mm1"); # t[0,1] collected + &movz ($acc,&LB("ebx")); # 8 + &xor ("ecx",&DWP(0,$tbl,$acc,8)); # 8 + &shr ("ebx",16); # 13,12 + + &movz ($acc,&HB("eax")); # 7 + &xor ("ecx",&DWP(1,$tbl,$acc,8)); # 7 + &pxor ("mm0","mm3"); + &movz ("eax",&LB("eax")); # 6 + &xor ("edx",&DWP(2,$tbl,"eax",8)); # 6 + &pshufw ("mm1","mm0",0x08); # 5, 4, 1, 0 + &movz ($acc,&HB("ebx")); # 13 + &xor ("ecx",&DWP(3,$tbl,$acc,8)); # 13 + &xor ("ecx",&DWP(24,$key)); # t[2] + &movd ("mm4","ecx"); # t[2] collected + &movz ("ebx",&LB("ebx")); # 12 + &xor ("edx",&DWP(0,$tbl,"ebx",8)); # 12 + &shr ("ecx",16); + &movd ("eax","mm1"); # 5, 4, 1, 0 + &mov ("ebx",&DWP(28,$key)); # t[3] + &xor ("ebx","edx"); + &movd ("mm5","ebx"); # t[3] collected + &and ("ebx",0xffff0000); + &or ("ebx","ecx"); + + &punpckldq ("mm4","mm5"); # t[2,3] collected +} + +###################################################################### +# "Compact" block function +###################################################################### + +sub enccompact() +{ my $Fn = mov; + while ($#_>5) { pop(@_); $Fn=sub{}; } + my ($i,$te,@s)=@_; + my $tmp = $key; + my $out = $i==3?$s[0]:$acc; + + # $Fn is used in first compact round and its purpose is to + # void restoration of some values from stack, so that after + # 4xenccompact with extra argument $key value is left there... + if ($i==3) { &$Fn ($key,$__key); }##%edx + else { &mov ($out,$s[0]); } + &and ($out,0xFF); + if ($i==1) { &shr ($s[0],16); }#%ebx[1] + if ($i==2) { &shr ($s[0],24); }#%ecx[2] + &movz ($out,&BP(-128,$te,$out,1)); + + if ($i==3) { $tmp=$s[1]; }##%eax + &movz ($tmp,&HB($s[1])); + &movz ($tmp,&BP(-128,$te,$tmp,1)); + &shl ($tmp,8); + &xor ($out,$tmp); + + if ($i==3) { $tmp=$s[2]; &mov ($s[1],$__s0); }##%ebx + else { &mov ($tmp,$s[2]); + &shr ($tmp,16); } + if ($i==2) { &and ($s[1],0xFF); }#%edx[2] + &and ($tmp,0xFF); + &movz ($tmp,&BP(-128,$te,$tmp,1)); + &shl ($tmp,16); + &xor ($out,$tmp); + + if ($i==3) { $tmp=$s[3]; &mov ($s[2],$__s1); }##%ecx + elsif($i==2){ &movz ($tmp,&HB($s[3])); }#%ebx[2] + else { &mov ($tmp,$s[3]); + &shr ($tmp,24); } + &movz ($tmp,&BP(-128,$te,$tmp,1)); + &shl ($tmp,24); + &xor ($out,$tmp); + if ($i<2) { &mov (&DWP(4+4*$i,"esp"),$out); } + if ($i==3) { &mov ($s[3],$acc); } + &comment(); +} + +sub enctransform() +{ my @s = ($s0,$s1,$s2,$s3); + my $i = shift; + my $tmp = $tbl; + my $r2 = $key ; + + &mov ($acc,$s[$i]); + &and ($acc,0x80808080); + &mov ($tmp,$acc); + &shr ($tmp,7); + &lea ($r2,&DWP(0,$s[$i],$s[$i])); + &sub ($acc,$tmp); + &and ($r2,0xfefefefe); + &and ($acc,0x1b1b1b1b); + &mov ($tmp,$s[$i]); + &xor ($acc,$r2); # r2 + + &xor ($s[$i],$acc); # r0 ^ r2 + &rotl ($s[$i],24); + &xor ($s[$i],$acc) # ROTATE(r2^r0,24) ^ r2 + &rotr ($tmp,16); + &xor ($s[$i],$tmp); + &rotr ($tmp,8); + &xor ($s[$i],$tmp); +} + +&function_begin_B("_x86_AES_encrypt_compact"); + # note that caller is expected to allocate stack frame for me! + &mov ($__key,$key); # save key + + &xor ($s0,&DWP(0,$key)); # xor with key + &xor ($s1,&DWP(4,$key)); + &xor ($s2,&DWP(8,$key)); + &xor ($s3,&DWP(12,$key)); + + &mov ($acc,&DWP(240,$key)); # load key->rounds + &lea ($acc,&DWP(-2,$acc,$acc)); + &lea ($acc,&DWP(0,$key,$acc,8)); + &mov ($__end,$acc); # end of key schedule + + # prefetch Te4 + &mov ($key,&DWP(0-128,$tbl)); + &mov ($acc,&DWP(32-128,$tbl)); + &mov ($key,&DWP(64-128,$tbl)); + &mov ($acc,&DWP(96-128,$tbl)); + &mov ($key,&DWP(128-128,$tbl)); + &mov ($acc,&DWP(160-128,$tbl)); + &mov ($key,&DWP(192-128,$tbl)); + &mov ($acc,&DWP(224-128,$tbl)); + + &set_label("loop",16); + + &enccompact(0,$tbl,$s0,$s1,$s2,$s3,1); + &enccompact(1,$tbl,$s1,$s2,$s3,$s0,1); + &enccompact(2,$tbl,$s2,$s3,$s0,$s1,1); + &enccompact(3,$tbl,$s3,$s0,$s1,$s2,1); + &enctransform(2); + &enctransform(3); + &enctransform(0); + &enctransform(1); + &mov ($key,$__key); + &mov ($tbl,$__tbl); + &add ($key,16); # advance rd_key + &xor ($s0,&DWP(0,$key)); + &xor ($s1,&DWP(4,$key)); + &xor ($s2,&DWP(8,$key)); + &xor ($s3,&DWP(12,$key)); + + &cmp ($key,$__end); + &mov ($__key,$key); + &jb (&label("loop")); + + &enccompact(0,$tbl,$s0,$s1,$s2,$s3); + &enccompact(1,$tbl,$s1,$s2,$s3,$s0); + &enccompact(2,$tbl,$s2,$s3,$s0,$s1); + &enccompact(3,$tbl,$s3,$s0,$s1,$s2); + + &xor ($s0,&DWP(16,$key)); + &xor ($s1,&DWP(20,$key)); + &xor ($s2,&DWP(24,$key)); + &xor ($s3,&DWP(28,$key)); + + &ret (); +&function_end_B("_x86_AES_encrypt_compact"); + +###################################################################### +# "Compact" SSE block function. +###################################################################### +# +# Performance is not actually extraordinary in comparison to pure +# x86 code. In particular encrypt performance is virtually the same. +# Decrypt performance on the other hand is 15-20% better on newer +# µ-archs [but we're thankful for *any* improvement here], and ~50% +# better on PIII:-) And additionally on the pros side this code +# eliminates redundant references to stack and thus relieves/ +# minimizes the pressure on the memory bus. +# +# MMX register layout lsb +# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ +# | mm4 | mm0 | +# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ +# | s3 | s2 | s1 | s0 | +# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ +# |15|14|13|12|11|10| 9| 8| 7| 6| 5| 4| 3| 2| 1| 0| +# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ +# +# Indexes translate as s[N/4]>>(8*(N%4)), e.g. 5 means s1>>8. +# In this terms encryption and decryption "compact" permutation +# matrices can be depicted as following: +# +# encryption lsb # decryption lsb +# +----++----+----+----+----+ # +----++----+----+----+----+ +# | t0 || 15 | 10 | 5 | 0 | # | t0 || 7 | 10 | 13 | 0 | +# +----++----+----+----+----+ # +----++----+----+----+----+ +# | t1 || 3 | 14 | 9 | 4 | # | t1 || 11 | 14 | 1 | 4 | +# +----++----+----+----+----+ # +----++----+----+----+----+ +# | t2 || 7 | 2 | 13 | 8 | # | t2 || 15 | 2 | 5 | 8 | +# +----++----+----+----+----+ # +----++----+----+----+----+ +# | t3 || 11 | 6 | 1 | 12 | # | t3 || 3 | 6 | 9 | 12 | +# +----++----+----+----+----+ # +----++----+----+----+----+ +# +###################################################################### +# Why not xmm registers? Short answer. It was actually tested and +# was not any faster, but *contrary*, most notably on Intel CPUs. +# Longer answer. Main advantage of using mm registers is that movd +# latency is lower, especially on Intel P4. While arithmetic +# instructions are twice as many, they can be scheduled every cycle +# and not every second one when they are operating on xmm register, +# so that "arithmetic throughput" remains virtually the same. And +# finally the code can be executed even on elder SSE-only CPUs:-) + +sub sse_enccompact() +{ + &pshufw ("mm1","mm0",0x08); # 5, 4, 1, 0 + &pshufw ("mm5","mm4",0x0d); # 15,14,11,10 + &movd ("eax","mm1"); # 5, 4, 1, 0 + &movd ("ebx","mm5"); # 15,14,11,10 + + &movz ($acc,&LB("eax")); # 0 + &movz ("ecx",&BP(-128,$tbl,$acc,1)); # 0 + &pshufw ("mm2","mm0",0x0d); # 7, 6, 3, 2 + &movz ("edx",&HB("eax")); # 1 + &movz ("edx",&BP(-128,$tbl,"edx",1)); # 1 + &shl ("edx",8); # 1 + &shr ("eax",16); # 5, 4 + + &movz ($acc,&LB("ebx")); # 10 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 10 + &shl ($acc,16); # 10 + &or ("ecx",$acc); # 10 + &pshufw ("mm6","mm4",0x08); # 13,12, 9, 8 + &movz ($acc,&HB("ebx")); # 11 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 11 + &shl ($acc,24); # 11 + &or ("edx",$acc); # 11 + &shr ("ebx",16); # 15,14 + + &movz ($acc,&HB("eax")); # 5 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 5 + &shl ($acc,8); # 5 + &or ("ecx",$acc); # 5 + &movz ($acc,&HB("ebx")); # 15 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 15 + &shl ($acc,24); # 15 + &or ("ecx",$acc); # 15 + &movd ("mm0","ecx"); # t[0] collected + + &movz ($acc,&LB("eax")); # 4 + &movz ("ecx",&BP(-128,$tbl,$acc,1)); # 4 + &movd ("eax","mm2"); # 7, 6, 3, 2 + &movz ($acc,&LB("ebx")); # 14 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 14 + &shl ($acc,16); # 14 + &or ("ecx",$acc); # 14 + + &movd ("ebx","mm6"); # 13,12, 9, 8 + &movz ($acc,&HB("eax")); # 3 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 3 + &shl ($acc,24); # 3 + &or ("ecx",$acc); # 3 + &movz ($acc,&HB("ebx")); # 9 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 9 + &shl ($acc,8); # 9 + &or ("ecx",$acc); # 9 + &movd ("mm1","ecx"); # t[1] collected + + &movz ($acc,&LB("ebx")); # 8 + &movz ("ecx",&BP(-128,$tbl,$acc,1)); # 8 + &shr ("ebx",16); # 13,12 + &movz ($acc,&LB("eax")); # 2 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 2 + &shl ($acc,16); # 2 + &or ("ecx",$acc); # 2 + &shr ("eax",16); # 7, 6 + + &punpckldq ("mm0","mm1"); # t[0,1] collected + + &movz ($acc,&HB("eax")); # 7 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 7 + &shl ($acc,24); # 7 + &or ("ecx",$acc); # 7 + &and ("eax",0xff); # 6 + &movz ("eax",&BP(-128,$tbl,"eax",1)); # 6 + &shl ("eax",16); # 6 + &or ("edx","eax"); # 6 + &movz ($acc,&HB("ebx")); # 13 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 13 + &shl ($acc,8); # 13 + &or ("ecx",$acc); # 13 + &movd ("mm4","ecx"); # t[2] collected + &and ("ebx",0xff); # 12 + &movz ("ebx",&BP(-128,$tbl,"ebx",1)); # 12 + &or ("edx","ebx"); # 12 + &movd ("mm5","edx"); # t[3] collected + + &punpckldq ("mm4","mm5"); # t[2,3] collected +} + + if (!$x86only) { +&function_begin_B("_sse_AES_encrypt_compact"); + &pxor ("mm0",&QWP(0,$key)); # 7, 6, 5, 4, 3, 2, 1, 0 + &pxor ("mm4",&QWP(8,$key)); # 15,14,13,12,11,10, 9, 8 + + # note that caller is expected to allocate stack frame for me! + &mov ($acc,&DWP(240,$key)); # load key->rounds + &lea ($acc,&DWP(-2,$acc,$acc)); + &lea ($acc,&DWP(0,$key,$acc,8)); + &mov ($__end,$acc); # end of key schedule + + &mov ($s0,0x1b1b1b1b); # magic constant + &mov (&DWP(8,"esp"),$s0); + &mov (&DWP(12,"esp"),$s0); + + # prefetch Te4 + &mov ($s0,&DWP(0-128,$tbl)); + &mov ($s1,&DWP(32-128,$tbl)); + &mov ($s2,&DWP(64-128,$tbl)); + &mov ($s3,&DWP(96-128,$tbl)); + &mov ($s0,&DWP(128-128,$tbl)); + &mov ($s1,&DWP(160-128,$tbl)); + &mov ($s2,&DWP(192-128,$tbl)); + &mov ($s3,&DWP(224-128,$tbl)); + + &set_label("loop",16); + &sse_enccompact(); + &add ($key,16); + &cmp ($key,$__end); + &ja (&label("out")); + + &movq ("mm2",&QWP(8,"esp")); + &pxor ("mm3","mm3"); &pxor ("mm7","mm7"); + &movq ("mm1","mm0"); &movq ("mm5","mm4"); # r0 + &pcmpgtb("mm3","mm0"); &pcmpgtb("mm7","mm4"); + &pand ("mm3","mm2"); &pand ("mm7","mm2"); + &pshufw ("mm2","mm0",0xb1); &pshufw ("mm6","mm4",0xb1);# ROTATE(r0,16) + &paddb ("mm0","mm0"); &paddb ("mm4","mm4"); + &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # = r2 + &pshufw ("mm3","mm2",0xb1); &pshufw ("mm7","mm6",0xb1);# r0 + &pxor ("mm1","mm0"); &pxor ("mm5","mm4"); # r0^r2 + &pxor ("mm0","mm2"); &pxor ("mm4","mm6"); # ^= ROTATE(r0,16) + + &movq ("mm2","mm3"); &movq ("mm6","mm7"); + &pslld ("mm3",8); &pslld ("mm7",8); + &psrld ("mm2",24); &psrld ("mm6",24); + &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # ^= r0<<8 + &pxor ("mm0","mm2"); &pxor ("mm4","mm6"); # ^= r0>>24 + + &movq ("mm3","mm1"); &movq ("mm7","mm5"); + &movq ("mm2",&QWP(0,$key)); &movq ("mm6",&QWP(8,$key)); + &psrld ("mm1",8); &psrld ("mm5",8); + &mov ($s0,&DWP(0-128,$tbl)); + &pslld ("mm3",24); &pslld ("mm7",24); + &mov ($s1,&DWP(64-128,$tbl)); + &pxor ("mm0","mm1"); &pxor ("mm4","mm5"); # ^= (r2^r0)<<8 + &mov ($s2,&DWP(128-128,$tbl)); + &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # ^= (r2^r0)>>24 + &mov ($s3,&DWP(192-128,$tbl)); + + &pxor ("mm0","mm2"); &pxor ("mm4","mm6"); + &jmp (&label("loop")); + + &set_label("out",16); + &pxor ("mm0",&QWP(0,$key)); + &pxor ("mm4",&QWP(8,$key)); + + &ret (); +&function_end_B("_sse_AES_encrypt_compact"); + } + +###################################################################### +# Vanilla block function. +###################################################################### + sub encstep() { my ($i,$te,@s) = @_; my $tmp = $key; my $out = $i==3?$s[0]:$acc; # lines marked with #%e?x[i] denote "reordered" instructions... - if ($i==3) { &mov ($key,&DWP(12,"esp")); }##%edx + if ($i==3) { &mov ($key,$__key); }##%edx else { &mov ($out,$s[0]); &and ($out,0xFF); } if ($i==1) { &shr ($s[0],16); }#%ebx[1] @@ -191,14 +780,14 @@ sub encstep() &movz ($tmp,&HB($s[1])); &xor ($out,&DWP(3,$te,$tmp,8)); - if ($i==3) { $tmp=$s[2]; &mov ($s[1],&DWP(4,"esp")); }##%ebx + if ($i==3) { $tmp=$s[2]; &mov ($s[1],$__s0); }##%ebx else { &mov ($tmp,$s[2]); &shr ($tmp,16); } if ($i==2) { &and ($s[1],0xFF); }#%edx[2] &and ($tmp,0xFF); &xor ($out,&DWP(2,$te,$tmp,8)); - if ($i==3) { $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }##%ecx + if ($i==3) { $tmp=$s[3]; &mov ($s[2],$__s1); }##%ecx elsif($i==2){ &movz ($tmp,&HB($s[3])); }#%ebx[2] else { &mov ($tmp,$s[3]); &shr ($tmp,24) } @@ -213,7 +802,7 @@ sub enclast() my $tmp = $key; my $out = $i==3?$s[0]:$acc; - if ($i==3) { &mov ($key,&DWP(12,"esp")); }##%edx + if ($i==3) { &mov ($key,$__key); }##%edx else { &mov ($out,$s[0]); } &and ($out,0xFF); if ($i==1) { &shr ($s[0],16); }#%ebx[1] @@ -227,8 +816,8 @@ sub enclast() &and ($tmp,0x0000ff00); &xor ($out,$tmp); - if ($i==3) { $tmp=$s[2]; &mov ($s[1],&DWP(4,"esp")); }##%ebx - else { mov ($tmp,$s[2]); + if ($i==3) { $tmp=$s[2]; &mov ($s[1],$__s0); }##%ebx + else { &mov ($tmp,$s[2]); &shr ($tmp,16); } if ($i==2) { &and ($s[1],0xFF); }#%edx[2] &and ($tmp,0xFF); @@ -236,7 +825,7 @@ sub enclast() &and ($tmp,0x00ff0000); &xor ($out,$tmp); - if ($i==3) { $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }##%ecx + if ($i==3) { $tmp=$s[3]; &mov ($s[2],$__s1); }##%ecx elsif($i==2){ &movz ($tmp,&HB($s[3])); }#%ebx[2] else { &mov ($tmp,$s[3]); &shr ($tmp,24); } @@ -247,9 +836,6 @@ sub enclast() if ($i==3) { &mov ($s[3],$acc); } } -sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } } - -&public_label("AES_Te"); &function_begin_B("_x86_AES_encrypt"); if ($vertical_spin) { # I need high parts of volatile registers to be accessible... @@ -258,7 +844,7 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } } } # note that caller is expected to allocate stack frame for me! - &mov (&DWP(12,"esp"),$key); # save key + &mov ($__key,$key); # save key &xor ($s0,&DWP(0,$key)); # xor with key &xor ($s1,&DWP(4,$key)); @@ -270,24 +856,24 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } } if ($small_footprint) { &lea ($acc,&DWP(-2,$acc,$acc)); &lea ($acc,&DWP(0,$key,$acc,8)); - &mov (&DWP(16,"esp"),$acc); # end of key schedule - &align (4); - &set_label("loop"); + &mov ($__end,$acc); # end of key schedule + + &set_label("loop",16); if ($vertical_spin) { - &encvert("ebp",$s0,$s1,$s2,$s3); + &encvert($tbl,$s0,$s1,$s2,$s3); } else { - &encstep(0,"ebp",$s0,$s1,$s2,$s3); - &encstep(1,"ebp",$s1,$s2,$s3,$s0); - &encstep(2,"ebp",$s2,$s3,$s0,$s1); - &encstep(3,"ebp",$s3,$s0,$s1,$s2); + &encstep(0,$tbl,$s0,$s1,$s2,$s3); + &encstep(1,$tbl,$s1,$s2,$s3,$s0); + &encstep(2,$tbl,$s2,$s3,$s0,$s1); + &encstep(3,$tbl,$s3,$s0,$s1,$s2); } &add ($key,16); # advance rd_key &xor ($s0,&DWP(0,$key)); &xor ($s1,&DWP(4,$key)); &xor ($s2,&DWP(8,$key)); &xor ($s3,&DWP(12,$key)); - &cmp ($key,&DWP(16,"esp")); - &mov (&DWP(12,"esp"),$key); + &cmp ($key,$__end); + &mov ($__key,$key); &jb (&label("loop")); } else { @@ -296,15 +882,15 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } } &cmp ($acc,12); &jle (&label("12rounds")); - &set_label("14rounds"); + &set_label("14rounds",4); for ($i=1;$i<3;$i++) { if ($vertical_spin) { - &encvert("ebp",$s0,$s1,$s2,$s3); + &encvert($tbl,$s0,$s1,$s2,$s3); } else { - &encstep(0,"ebp",$s0,$s1,$s2,$s3); - &encstep(1,"ebp",$s1,$s2,$s3,$s0); - &encstep(2,"ebp",$s2,$s3,$s0,$s1); - &encstep(3,"ebp",$s3,$s0,$s1,$s2); + &encstep(0,$tbl,$s0,$s1,$s2,$s3); + &encstep(1,$tbl,$s1,$s2,$s3,$s0); + &encstep(2,$tbl,$s2,$s3,$s0,$s1); + &encstep(3,$tbl,$s3,$s0,$s1,$s2); } &xor ($s0,&DWP(16*$i+0,$key)); &xor ($s1,&DWP(16*$i+4,$key)); @@ -312,16 +898,16 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } } &xor ($s3,&DWP(16*$i+12,$key)); } &add ($key,32); - &mov (&DWP(12,"esp"),$key); # advance rd_key - &set_label("12rounds"); + &mov ($__key,$key); # advance rd_key + &set_label("12rounds",4); for ($i=1;$i<3;$i++) { if ($vertical_spin) { - &encvert("ebp",$s0,$s1,$s2,$s3); + &encvert($tbl,$s0,$s1,$s2,$s3); } else { - &encstep(0,"ebp",$s0,$s1,$s2,$s3); - &encstep(1,"ebp",$s1,$s2,$s3,$s0); - &encstep(2,"ebp",$s2,$s3,$s0,$s1); - &encstep(3,"ebp",$s3,$s0,$s1,$s2); + &encstep(0,$tbl,$s0,$s1,$s2,$s3); + &encstep(1,$tbl,$s1,$s2,$s3,$s0); + &encstep(2,$tbl,$s2,$s3,$s0,$s1); + &encstep(3,$tbl,$s3,$s0,$s1,$s2); } &xor ($s0,&DWP(16*$i+0,$key)); &xor ($s1,&DWP(16*$i+4,$key)); @@ -329,16 +915,16 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } } &xor ($s3,&DWP(16*$i+12,$key)); } &add ($key,32); - &mov (&DWP(12,"esp"),$key); # advance rd_key - &set_label("10rounds"); + &mov ($__key,$key); # advance rd_key + &set_label("10rounds",4); for ($i=1;$i<10;$i++) { if ($vertical_spin) { - &encvert("ebp",$s0,$s1,$s2,$s3); + &encvert($tbl,$s0,$s1,$s2,$s3); } else { - &encstep(0,"ebp",$s0,$s1,$s2,$s3); - &encstep(1,"ebp",$s1,$s2,$s3,$s0); - &encstep(2,"ebp",$s2,$s3,$s0,$s1); - &encstep(3,"ebp",$s3,$s0,$s1,$s2); + &encstep(0,$tbl,$s0,$s1,$s2,$s3); + &encstep(1,$tbl,$s1,$s2,$s3,$s0); + &encstep(2,$tbl,$s2,$s3,$s0,$s1); + &encstep(3,$tbl,$s3,$s0,$s1,$s2); } &xor ($s0,&DWP(16*$i+0,$key)); &xor ($s1,&DWP(16*$i+4,$key)); @@ -352,10 +938,10 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } } &mov ($s1="ebx",$key="edi"); &mov ($s2="ecx",$acc="esi"); } - &enclast(0,"ebp",$s0,$s1,$s2,$s3); - &enclast(1,"ebp",$s1,$s2,$s3,$s0); - &enclast(2,"ebp",$s2,$s3,$s0,$s1); - &enclast(3,"ebp",$s3,$s0,$s1,$s2); + &enclast(0,$tbl,$s0,$s1,$s2,$s3); + &enclast(1,$tbl,$s1,$s2,$s3,$s0); + &enclast(2,$tbl,$s2,$s3,$s0,$s1); + &enclast(3,$tbl,$s3,$s0,$s1,$s2); &add ($key,$small_footprint?16:160); &xor ($s0,&DWP(0,$key)); @@ -430,38 +1016,198 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } } &_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0); &_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e); &_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c); + +#Te4 # four copies of Te4 to choose from to avoid L1 aliasing + &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5); + &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76); + &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0); + &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0); + &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc); + &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15); + &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a); + &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75); + &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0); + &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84); + &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b); + &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf); + &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85); + &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8); + &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5); + &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2); + &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17); + &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73); + &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88); + &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb); + &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c); + &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79); + &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9); + &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08); + &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6); + &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a); + &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e); + &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e); + &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94); + &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf); + &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68); + &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16); + + &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5); + &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76); + &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0); + &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0); + &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc); + &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15); + &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a); + &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75); + &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0); + &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84); + &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b); + &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf); + &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85); + &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8); + &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5); + &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2); + &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17); + &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73); + &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88); + &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb); + &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c); + &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79); + &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9); + &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08); + &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6); + &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a); + &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e); + &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e); + &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94); + &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf); + &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68); + &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16); + + &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5); + &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76); + &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0); + &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0); + &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc); + &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15); + &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a); + &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75); + &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0); + &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84); + &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b); + &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf); + &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85); + &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8); + &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5); + &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2); + &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17); + &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73); + &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88); + &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb); + &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c); + &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79); + &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9); + &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08); + &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6); + &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a); + &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e); + &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e); + &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94); + &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf); + &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68); + &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16); + + &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5); + &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76); + &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0); + &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0); + &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc); + &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15); + &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a); + &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75); + &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0); + &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84); + &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b); + &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf); + &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85); + &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8); + &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5); + &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2); + &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17); + &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73); + &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88); + &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb); + &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c); + &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79); + &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9); + &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08); + &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6); + &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a); + &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e); + &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e); + &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94); + &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf); + &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68); + &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16); #rcon: &data_word(0x00000001, 0x00000002, 0x00000004, 0x00000008); &data_word(0x00000010, 0x00000020, 0x00000040, 0x00000080); - &data_word(0x0000001b, 0x00000036, 0, 0, 0, 0, 0, 0); + &data_word(0x0000001b, 0x00000036, 0x00000000, 0x00000000); + &data_word(0x00000000, 0x00000000, 0x00000000, 0x00000000); &function_end_B("_x86_AES_encrypt"); # void AES_encrypt (const void *inp,void *out,const AES_KEY *key); -&public_label("AES_Te"); &function_begin("AES_encrypt"); &mov ($acc,&wparam(0)); # load inp &mov ($key,&wparam(2)); # load key &mov ($s0,"esp"); - &sub ("esp",24); - &and ("esp",-64); - &add ("esp",4); - &mov (&DWP(16,"esp"),$s0); + &sub ("esp",36); + &and ("esp",-64); # align to cache-line + + # place stack frame just "above" the key schedule + &lea ($s1,&DWP(-64-63,$key)); + &sub ($s1,"esp"); + &neg ($s1); + &and ($s1,0x3C0); # modulo 1024, but aligned to cache-line + &sub ("esp",$s1); + &add ("esp",4); # 4 is reserved for caller's return address + &mov ($_esp,$s0); # save stack pointer &call (&label("pic_point")); # make it PIC! &set_label("pic_point"); - &blindpop("ebp"); - &lea ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp")); - + &blindpop($tbl); + &picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if (!$x86only); + &lea ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl)); + + # pick Te4 copy which can't "overlap" with stack frame or key schedule + &lea ($s1,&DWP(768-4,"esp")); + &sub ($s1,$tbl); + &and ($s1,0x300); + &lea ($tbl,&DWP(2048+128,$tbl,$s1)); + + if (!$x86only) { + &bt (&DWP(0,$s0),25); # check for SSE bit + &jnc (&label("x86")); + + &movq ("mm0",&QWP(0,$acc)); + &movq ("mm4",&QWP(8,$acc)); + &call ("_sse_AES_encrypt_compact"); + &mov ("esp",$_esp); # restore stack pointer + &mov ($acc,&wparam(1)); # load out + &movq (&QWP(0,$acc),"mm0"); # write output data + &movq (&QWP(8,$acc),"mm4"); + &emms (); + &function_end_A(); + } + &set_label("x86",16); + &mov ($_tbl,$tbl); &mov ($s0,&DWP(0,$acc)); # load input data &mov ($s1,&DWP(4,$acc)); &mov ($s2,&DWP(8,$acc)); &mov ($s3,&DWP(12,$acc)); - - &call ("_x86_AES_encrypt"); - - &mov ("esp",&DWP(16,"esp")); - + &call ("_x86_AES_encrypt_compact"); + &mov ("esp",$_esp); # restore stack pointer &mov ($acc,&wparam(1)); # load out &mov (&DWP(0,$acc),$s0); # write output data &mov (&DWP(4,$acc),$s1); @@ -469,7 +1215,370 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } } &mov (&DWP(12,$acc),$s3); &function_end("AES_encrypt"); -#------------------------------------------------------------------# +#--------------------------------------------------------------------# + +###################################################################### +# "Compact" block function +###################################################################### + +sub deccompact() +{ my $Fn = mov; + while ($#_>5) { pop(@_); $Fn=sub{}; } + my ($i,$td,@s)=@_; + my $tmp = $key; + my $out = $i==3?$s[0]:$acc; + + # $Fn is used in first compact round and its purpose is to + # void restoration of some values from stack, so that after + # 4xdeccompact with extra argument $key, $s0 and $s1 values + # are left there... + if($i==3) { &$Fn ($key,$__key); } + else { &mov ($out,$s[0]); } + &and ($out,0xFF); + &movz ($out,&BP(-128,$td,$out,1)); + + if ($i==3) { $tmp=$s[1]; } + &movz ($tmp,&HB($s[1])); + &movz ($tmp,&BP(-128,$td,$tmp,1)); + &shl ($tmp,8); + &xor ($out,$tmp); + + if ($i==3) { $tmp=$s[2]; &mov ($s[1],$acc); } + else { mov ($tmp,$s[2]); } + &shr ($tmp,16); + &and ($tmp,0xFF); + &movz ($tmp,&BP(-128,$td,$tmp,1)); + &shl ($tmp,16); + &xor ($out,$tmp); + + if ($i==3) { $tmp=$s[3]; &$Fn ($s[2],$__s1); } + else { &mov ($tmp,$s[3]); } + &shr ($tmp,24); + &movz ($tmp,&BP(-128,$td,$tmp,1)); + &shl ($tmp,24); + &xor ($out,$tmp); + if ($i<2) { &mov (&DWP(4+4*$i,"esp"),$out); } + if ($i==3) { &$Fn ($s[3],$__s0); } +} + +# must be called with 2,3,0,1 as argument sequence!!! +sub dectransform() +{ my @s = ($s0,$s1,$s2,$s3); + my $i = shift; + my $tmp = $key; + my $tp2 = @s[($i+2)%4]; $tp2 = @s[2] if ($i==1); + my $tp4 = @s[($i+3)%4]; $tp4 = @s[3] if ($i==1); + my $tp8 = $tbl; + + &mov ($acc,$s[$i]); + &and ($acc,0x80808080); + &mov ($tmp,$acc); + &shr ($tmp,7); + &lea ($tp2,&DWP(0,$s[$i],$s[$i])); + &sub ($acc,$tmp); + &and ($tp2,0xfefefefe); + &and ($acc,0x1b1b1b1b); + &xor ($acc,$tp2); + &mov ($tp2,$acc); + + &and ($acc,0x80808080); + &mov ($tmp,$acc); + &shr ($tmp,7); + &lea ($tp4,&DWP(0,$tp2,$tp2)); + &sub ($acc,$tmp); + &and ($tp4,0xfefefefe); + &and ($acc,0x1b1b1b1b); + &xor ($tp2,$s[$i]); # tp2^tp1 + &xor ($acc,$tp4); + &mov ($tp4,$acc); + + &and ($acc,0x80808080); + &mov ($tmp,$acc); + &shr ($tmp,7); + &lea ($tp8,&DWP(0,$tp4,$tp4)); + &sub ($acc,$tmp); + &and ($tp8,0xfefefefe); + &and ($acc,0x1b1b1b1b); + &xor ($tp4,$s[$i]); # tp4^tp1 + &rotl ($s[$i],8); # = ROTATE(tp1,8) + &xor ($tp8,$acc); + + &xor ($s[$i],$tp2); + &xor ($tp2,$tp8); + &rotl ($tp2,24); + &xor ($s[$i],$tp4); + &xor ($tp4,$tp8); + &rotl ($tp4,16); + &xor ($s[$i],$tp8); # ^= tp8^(tp4^tp1)^(tp2^tp1) + &rotl ($tp8,8); + &xor ($s[$i],$tp2); # ^= ROTATE(tp8^tp2^tp1,24) + &xor ($s[$i],$tp4); # ^= ROTATE(tp8^tp4^tp1,16) + &mov ($s[0],$__s0) if($i==2); #prefetch $s0 + &mov ($s[1],$__s1) if($i==3); #prefetch $s1 + &mov ($s[2],$__s2) if($i==1); + &xor ($s[$i],$tp8); # ^= ROTATE(tp8,8) + + &mov ($s[3],$__s3) if($i==1); + &mov (&DWP(4+4*$i,"esp"),$s[$i]) if($i>=2); +} + +&function_begin_B("_x86_AES_decrypt_compact"); + # note that caller is expected to allocate stack frame for me! + &mov ($__key,$key); # save key + + &xor ($s0,&DWP(0,$key)); # xor with key + &xor ($s1,&DWP(4,$key)); + &xor ($s2,&DWP(8,$key)); + &xor ($s3,&DWP(12,$key)); + + &mov ($acc,&DWP(240,$key)); # load key->rounds + + &lea ($acc,&DWP(-2,$acc,$acc)); + &lea ($acc,&DWP(0,$key,$acc,8)); + &mov ($__end,$acc); # end of key schedule + + # prefetch Td4 + &mov ($key,&DWP(0-128,$tbl)); + &mov ($acc,&DWP(32-128,$tbl)); + &mov ($key,&DWP(64-128,$tbl)); + &mov ($acc,&DWP(96-128,$tbl)); + &mov ($key,&DWP(128-128,$tbl)); + &mov ($acc,&DWP(160-128,$tbl)); + &mov ($key,&DWP(192-128,$tbl)); + &mov ($acc,&DWP(224-128,$tbl)); + + &set_label("loop",16); + + &deccompact(0,$tbl,$s0,$s3,$s2,$s1,1); + &deccompact(1,$tbl,$s1,$s0,$s3,$s2,1); + &deccompact(2,$tbl,$s2,$s1,$s0,$s3,1); + &deccompact(3,$tbl,$s3,$s2,$s1,$s0,1); + &dectransform(2); + &dectransform(3); + &dectransform(0); + &dectransform(1); + &mov ($key,$__key); + &mov ($tbl,$__tbl); + &add ($key,16); # advance rd_key + &xor ($s0,&DWP(0,$key)); + &xor ($s1,&DWP(4,$key)); + &xor ($s2,&DWP(8,$key)); + &xor ($s3,&DWP(12,$key)); + + &cmp ($key,$__end); + &mov ($__key,$key); + &jb (&label("loop")); + + &deccompact(0,$tbl,$s0,$s3,$s2,$s1); + &deccompact(1,$tbl,$s1,$s0,$s3,$s2); + &deccompact(2,$tbl,$s2,$s1,$s0,$s3); + &deccompact(3,$tbl,$s3,$s2,$s1,$s0); + + &xor ($s0,&DWP(16,$key)); + &xor ($s1,&DWP(20,$key)); + &xor ($s2,&DWP(24,$key)); + &xor ($s3,&DWP(28,$key)); + + &ret (); +&function_end_B("_x86_AES_decrypt_compact"); + +###################################################################### +# "Compact" SSE block function. +###################################################################### + +sub sse_deccompact() +{ + &pshufw ("mm1","mm0",0x0c); # 7, 6, 1, 0 + &movd ("eax","mm1"); # 7, 6, 1, 0 + + &pshufw ("mm5","mm4",0x09); # 13,12,11,10 + &movz ($acc,&LB("eax")); # 0 + &movz ("ecx",&BP(-128,$tbl,$acc,1)); # 0 + &movd ("ebx","mm5"); # 13,12,11,10 + &movz ("edx",&HB("eax")); # 1 + &movz ("edx",&BP(-128,$tbl,"edx",1)); # 1 + &shl ("edx",8); # 1 + + &pshufw ("mm2","mm0",0x06); # 3, 2, 5, 4 + &movz ($acc,&LB("ebx")); # 10 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 10 + &shl ($acc,16); # 10 + &or ("ecx",$acc); # 10 + &shr ("eax",16); # 7, 6 + &movz ($acc,&HB("ebx")); # 11 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 11 + &shl ($acc,24); # 11 + &or ("edx",$acc); # 11 + &shr ("ebx",16); # 13,12 + + &pshufw ("mm6","mm4",0x03); # 9, 8,15,14 + &movz ($acc,&HB("eax")); # 7 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 7 + &shl ($acc,24); # 7 + &or ("ecx",$acc); # 7 + &movz ($acc,&HB("ebx")); # 13 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 13 + &shl ($acc,8); # 13 + &or ("ecx",$acc); # 13 + &movd ("mm0","ecx"); # t[0] collected + + &movz ($acc,&LB("eax")); # 6 + &movd ("eax","mm2"); # 3, 2, 5, 4 + &movz ("ecx",&BP(-128,$tbl,$acc,1)); # 6 + &shl ("ecx",16); # 6 + &movz ($acc,&LB("ebx")); # 12 + &movd ("ebx","mm6"); # 9, 8,15,14 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 12 + &or ("ecx",$acc); # 12 + + &movz ($acc,&LB("eax")); # 4 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 4 + &or ("edx",$acc); # 4 + &movz ($acc,&LB("ebx")); # 14 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 14 + &shl ($acc,16); # 14 + &or ("edx",$acc); # 14 + &movd ("mm1","edx"); # t[1] collected + + &movz ($acc,&HB("eax")); # 5 + &movz ("edx",&BP(-128,$tbl,$acc,1)); # 5 + &shl ("edx",8); # 5 + &movz ($acc,&HB("ebx")); # 15 + &shr ("eax",16); # 3, 2 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 15 + &shl ($acc,24); # 15 + &or ("edx",$acc); # 15 + &shr ("ebx",16); # 9, 8 + + &punpckldq ("mm0","mm1"); # t[0,1] collected + + &movz ($acc,&HB("ebx")); # 9 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 9 + &shl ($acc,8); # 9 + &or ("ecx",$acc); # 9 + &and ("ebx",0xff); # 8 + &movz ("ebx",&BP(-128,$tbl,"ebx",1)); # 8 + &or ("edx","ebx"); # 8 + &movz ($acc,&LB("eax")); # 2 + &movz ($acc,&BP(-128,$tbl,$acc,1)); # 2 + &shl ($acc,16); # 2 + &or ("edx",$acc); # 2 + &movd ("mm4","edx"); # t[2] collected + &movz ("eax",&HB("eax")); # 3 + &movz ("eax",&BP(-128,$tbl,"eax",1)); # 3 + &shl ("eax",24); # 3 + &or ("ecx","eax"); # 3 + &movd ("mm5","ecx"); # t[3] collected + + &punpckldq ("mm4","mm5"); # t[2,3] collected +} + + if (!$x86only) { +&function_begin_B("_sse_AES_decrypt_compact"); + &pxor ("mm0",&QWP(0,$key)); # 7, 6, 5, 4, 3, 2, 1, 0 + &pxor ("mm4",&QWP(8,$key)); # 15,14,13,12,11,10, 9, 8 + + # note that caller is expected to allocate stack frame for me! + &mov ($acc,&DWP(240,$key)); # load key->rounds + &lea ($acc,&DWP(-2,$acc,$acc)); + &lea ($acc,&DWP(0,$key,$acc,8)); + &mov ($__end,$acc); # end of key schedule + + &mov ($s0,0x1b1b1b1b); # magic constant + &mov (&DWP(8,"esp"),$s0); + &mov (&DWP(12,"esp"),$s0); + + # prefetch Td4 + &mov ($s0,&DWP(0-128,$tbl)); + &mov ($s1,&DWP(32-128,$tbl)); + &mov ($s2,&DWP(64-128,$tbl)); + &mov ($s3,&DWP(96-128,$tbl)); + &mov ($s0,&DWP(128-128,$tbl)); + &mov ($s1,&DWP(160-128,$tbl)); + &mov ($s2,&DWP(192-128,$tbl)); + &mov ($s3,&DWP(224-128,$tbl)); + + &set_label("loop",16); + &sse_deccompact(); + &add ($key,16); + &cmp ($key,$__end); + &ja (&label("out")); + + # ROTATE(x^y,N) == ROTATE(x,N)^ROTATE(y,N) + &movq ("mm3","mm0"); &movq ("mm7","mm4"); + &movq ("mm2","mm0",1); &movq ("mm6","mm4",1); + &movq ("mm1","mm0"); &movq ("mm5","mm4"); + &pshufw ("mm0","mm0",0xb1); &pshufw ("mm4","mm4",0xb1);# = ROTATE(tp0,16) + &pslld ("mm2",8); &pslld ("mm6",8); + &psrld ("mm3",8); &psrld ("mm7",8); + &pxor ("mm0","mm2"); &pxor ("mm4","mm6"); # ^= tp0<<8 + &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # ^= tp0>>8 + &pslld ("mm2",16); &pslld ("mm6",16); + &psrld ("mm3",16); &psrld ("mm7",16); + &pxor ("mm0","mm2"); &pxor ("mm4","mm6"); # ^= tp0<<24 + &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # ^= tp0>>24 + + &movq ("mm3",&QWP(8,"esp")); + &pxor ("mm2","mm2"); &pxor ("mm6","mm6"); + &pcmpgtb("mm2","mm1"); &pcmpgtb("mm6","mm5"); + &pand ("mm2","mm3"); &pand ("mm6","mm3"); + &paddb ("mm1","mm1"); &paddb ("mm5","mm5"); + &pxor ("mm1","mm2"); &pxor ("mm5","mm6"); # tp2 + &movq ("mm3","mm1"); &movq ("mm7","mm5"); + &movq ("mm2","mm1"); &movq ("mm6","mm5"); + &pxor ("mm0","mm1"); &pxor ("mm4","mm5"); # ^= tp2 + &pslld ("mm3",24); &pslld ("mm7",24); + &psrld ("mm2",8); &psrld ("mm6",8); + &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # ^= tp2<<24 + &pxor ("mm0","mm2"); &pxor ("mm4","mm6"); # ^= tp2>>8 + + &movq ("mm2",&QWP(8,"esp")); + &pxor ("mm3","mm3"); &pxor ("mm7","mm7"); + &pcmpgtb("mm3","mm1"); &pcmpgtb("mm7","mm5"); + &pand ("mm3","mm2"); &pand ("mm7","mm2"); + &paddb ("mm1","mm1"); &paddb ("mm5","mm5"); + &pxor ("mm1","mm3"); &pxor ("mm5","mm7"); # tp4 + &pshufw ("mm3","mm1",0xb1); &pshufw ("mm7","mm5",0xb1); + &pxor ("mm0","mm1"); &pxor ("mm4","mm5"); # ^= tp4 + &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # ^= ROTATE(tp4,16) + + &pxor ("mm3","mm3"); &pxor ("mm7","mm7"); + &pcmpgtb("mm3","mm1"); &pcmpgtb("mm7","mm5"); + &pand ("mm3","mm2"); &pand ("mm7","mm2"); + &paddb ("mm1","mm1"); &paddb ("mm5","mm5"); + &pxor ("mm1","mm3"); &pxor ("mm5","mm7"); # tp8 + &pxor ("mm0","mm1"); &pxor ("mm4","mm5"); # ^= tp8 + &movq ("mm3","mm1"); &movq ("mm7","mm5"); + &pshufw ("mm2","mm1",0xb1); &pshufw ("mm6","mm5",0xb1); + &pxor ("mm0","mm2"); &pxor ("mm4","mm6"); # ^= ROTATE(tp8,16) + &pslld ("mm1",8); &pslld ("mm5",8); + &psrld ("mm3",8); &psrld ("mm7",8); + &movq ("mm2",&QWP(0,$key)); &movq ("mm6",&QWP(8,$key)); + &pxor ("mm0","mm1"); &pxor ("mm4","mm5"); # ^= tp8<<8 + &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # ^= tp8>>8 + &mov ($s0,&DWP(0-128,$tbl)); + &pslld ("mm1",16); &pslld ("mm5",16); + &mov ($s1,&DWP(64-128,$tbl)); + &psrld ("mm3",16); &psrld ("mm7",16); + &mov ($s2,&DWP(128-128,$tbl)); + &pxor ("mm0","mm1"); &pxor ("mm4","mm5"); # ^= tp8<<24 + &mov ($s3,&DWP(192-128,$tbl)); + &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # ^= tp8>>24 + + &pxor ("mm0","mm2"); &pxor ("mm4","mm6"); + &jmp (&label("loop")); + + &set_label("out",16); + &pxor ("mm0",&QWP(0,$key)); + &pxor ("mm4",&QWP(8,$key)); + + &ret (); +&function_end_B("_sse_AES_decrypt_compact"); + } + +###################################################################### +# Vanilla block function. +###################################################################### sub decstep() { my ($i,$td,@s) = @_; @@ -480,7 +1589,7 @@ sub decstep() # optimal... or rather that all attempts to reorder didn't # result in better performance [which by the way is not a # bit lower than ecryption]. - if($i==3) { &mov ($key,&DWP(12,"esp")); } + if($i==3) { &mov ($key,$__key); } else { &mov ($out,$s[0]); } &and ($out,0xFF); &mov ($out,&DWP(0,$td,$out,8)); @@ -495,12 +1604,12 @@ sub decstep() &and ($tmp,0xFF); &xor ($out,&DWP(2,$td,$tmp,8)); - if ($i==3) { $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); } + if ($i==3) { $tmp=$s[3]; &mov ($s[2],$__s1); } else { &mov ($tmp,$s[3]); } &shr ($tmp,24); &xor ($out,&DWP(1,$td,$tmp,8)); if ($i<2) { &mov (&DWP(4+4*$i,"esp"),$out); } - if ($i==3) { &mov ($s[3],&DWP(4,"esp")); } + if ($i==3) { &mov ($s[3],$__s0); } &comment(); } @@ -509,14 +1618,24 @@ sub declast() my $tmp = $key; my $out = $i==3?$s[0]:$acc; - if($i==3) { &mov ($key,&DWP(12,"esp")); } + if($i==0) { &lea ($td,&DWP(2048+128,$td)); + &mov ($tmp,&DWP(0-128,$td)); + &mov ($acc,&DWP(32-128,$td)); + &mov ($tmp,&DWP(64-128,$td)); + &mov ($acc,&DWP(96-128,$td)); + &mov ($tmp,&DWP(128-128,$td)); + &mov ($acc,&DWP(160-128,$td)); + &mov ($tmp,&DWP(192-128,$td)); + &mov ($acc,&DWP(224-128,$td)); + &lea ($td,&DWP(-128,$td)); } + if($i==3) { &mov ($key,$__key); } else { &mov ($out,$s[0]); } &and ($out,0xFF); - &movz ($out,&BP(2048,$td,$out,1)); + &movz ($out,&BP(0,$td,$out,1)); if ($i==3) { $tmp=$s[1]; } &movz ($tmp,&HB($s[1])); - &movz ($tmp,&BP(2048,$td,$tmp,1)); + &movz ($tmp,&BP(0,$td,$tmp,1)); &shl ($tmp,8); &xor ($out,$tmp); @@ -524,24 +1643,24 @@ sub declast() else { mov ($tmp,$s[2]); } &shr ($tmp,16); &and ($tmp,0xFF); - &movz ($tmp,&BP(2048,$td,$tmp,1)); + &movz ($tmp,&BP(0,$td,$tmp,1)); &shl ($tmp,16); &xor ($out,$tmp); - if ($i==3) { $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); } + if ($i==3) { $tmp=$s[3]; &mov ($s[2],$__s1); } else { &mov ($tmp,$s[3]); } &shr ($tmp,24); - &movz ($tmp,&BP(2048,$td,$tmp,1)); + &movz ($tmp,&BP(0,$td,$tmp,1)); &shl ($tmp,24); &xor ($out,$tmp); if ($i<2) { &mov (&DWP(4+4*$i,"esp"),$out); } - if ($i==3) { &mov ($s[3],&DWP(4,"esp")); } + if ($i==3) { &mov ($s[3],$__s0); + &lea ($td,&DWP(-2048,$td)); } } -&public_label("AES_Td"); &function_begin_B("_x86_AES_decrypt"); # note that caller is expected to allocate stack frame for me! - &mov (&DWP(12,"esp"),$key); # save key + &mov ($__key,$key); # save key &xor ($s0,&DWP(0,$key)); # xor with key &xor ($s1,&DWP(4,$key)); @@ -553,20 +1672,19 @@ sub declast() if ($small_footprint) { &lea ($acc,&DWP(-2,$acc,$acc)); &lea ($acc,&DWP(0,$key,$acc,8)); - &mov (&DWP(16,"esp"),$acc); # end of key schedule - &align (4); - &set_label("loop"); - &decstep(0,"ebp",$s0,$s3,$s2,$s1); - &decstep(1,"ebp",$s1,$s0,$s3,$s2); - &decstep(2,"ebp",$s2,$s1,$s0,$s3); - &decstep(3,"ebp",$s3,$s2,$s1,$s0); + &mov ($__end,$acc); # end of key schedule + &set_label("loop",16); + &decstep(0,$tbl,$s0,$s3,$s2,$s1); + &decstep(1,$tbl,$s1,$s0,$s3,$s2); + &decstep(2,$tbl,$s2,$s1,$s0,$s3); + &decstep(3,$tbl,$s3,$s2,$s1,$s0); &add ($key,16); # advance rd_key &xor ($s0,&DWP(0,$key)); &xor ($s1,&DWP(4,$key)); &xor ($s2,&DWP(8,$key)); &xor ($s3,&DWP(12,$key)); - &cmp ($key,&DWP(16,"esp")); - &mov (&DWP(12,"esp"),$key); + &cmp ($key,$__end); + &mov ($__key,$key); &jb (&label("loop")); } else { @@ -575,38 +1693,38 @@ sub declast() &cmp ($acc,12); &jle (&label("12rounds")); - &set_label("14rounds"); + &set_label("14rounds",4); for ($i=1;$i<3;$i++) { - &decstep(0,"ebp",$s0,$s3,$s2,$s1); - &decstep(1,"ebp",$s1,$s0,$s3,$s2); - &decstep(2,"ebp",$s2,$s1,$s0,$s3); - &decstep(3,"ebp",$s3,$s2,$s1,$s0); + &decstep(0,$tbl,$s0,$s3,$s2,$s1); + &decstep(1,$tbl,$s1,$s0,$s3,$s2); + &decstep(2,$tbl,$s2,$s1,$s0,$s3); + &decstep(3,$tbl,$s3,$s2,$s1,$s0); &xor ($s0,&DWP(16*$i+0,$key)); &xor ($s1,&DWP(16*$i+4,$key)); &xor ($s2,&DWP(16*$i+8,$key)); &xor ($s3,&DWP(16*$i+12,$key)); } &add ($key,32); - &mov (&DWP(12,"esp"),$key); # advance rd_key - &set_label("12rounds"); + &mov ($__key,$key); # advance rd_key + &set_label("12rounds",4); for ($i=1;$i<3;$i++) { - &decstep(0,"ebp",$s0,$s3,$s2,$s1); - &decstep(1,"ebp",$s1,$s0,$s3,$s2); - &decstep(2,"ebp",$s2,$s1,$s0,$s3); - &decstep(3,"ebp",$s3,$s2,$s1,$s0); + &decstep(0,$tbl,$s0,$s3,$s2,$s1); + &decstep(1,$tbl,$s1,$s0,$s3,$s2); + &decstep(2,$tbl,$s2,$s1,$s0,$s3); + &decstep(3,$tbl,$s3,$s2,$s1,$s0); &xor ($s0,&DWP(16*$i+0,$key)); &xor ($s1,&DWP(16*$i+4,$key)); &xor ($s2,&DWP(16*$i+8,$key)); &xor ($s3,&DWP(16*$i+12,$key)); } &add ($key,32); - &mov (&DWP(12,"esp"),$key); # advance rd_key - &set_label("10rounds"); + &mov ($__key,$key); # advance rd_key + &set_label("10rounds",4); for ($i=1;$i<10;$i++) { - &decstep(0,"ebp",$s0,$s3,$s2,$s1); - &decstep(1,"ebp",$s1,$s0,$s3,$s2); - &decstep(2,"ebp",$s2,$s1,$s0,$s3); - &decstep(3,"ebp",$s3,$s2,$s1,$s0); + &decstep(0,$tbl,$s0,$s3,$s2,$s1); + &decstep(1,$tbl,$s1,$s0,$s3,$s2); + &decstep(2,$tbl,$s2,$s1,$s0,$s3); + &decstep(3,$tbl,$s3,$s2,$s1,$s0); &xor ($s0,&DWP(16*$i+0,$key)); &xor ($s1,&DWP(16*$i+4,$key)); &xor ($s2,&DWP(16*$i+8,$key)); @@ -614,10 +1732,10 @@ sub declast() } } - &declast(0,"ebp",$s0,$s3,$s2,$s1); - &declast(1,"ebp",$s1,$s0,$s3,$s2); - &declast(2,"ebp",$s2,$s1,$s0,$s3); - &declast(3,"ebp",$s3,$s2,$s1,$s0); + &declast(0,$tbl,$s0,$s3,$s2,$s1); + &declast(1,$tbl,$s1,$s0,$s3,$s2); + &declast(2,$tbl,$s2,$s1,$s0,$s3); + &declast(3,$tbl,$s3,$s2,$s1,$s0); &add ($key,$small_footprint?16:160); &xor ($s0,&DWP(0,$key)); @@ -692,7 +1810,107 @@ sub declast() &_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff); &_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664); &_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0); -#Td4: + +#Td4: # four copies of Td4 to choose from to avoid L1 aliasing + &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38); + &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb); + &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87); + &data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb); + &data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d); + &data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e); + &data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2); + &data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25); + &data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16); + &data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92); + &data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda); + &data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84); + &data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a); + &data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06); + &data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02); + &data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b); + &data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea); + &data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73); + &data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85); + &data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e); + &data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89); + &data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b); + &data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20); + &data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4); + &data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31); + &data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f); + &data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d); + &data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef); + &data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0); + &data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61); + &data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26); + &data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d); + + &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38); + &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb); + &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87); + &data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb); + &data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d); + &data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e); + &data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2); + &data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25); + &data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16); + &data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92); + &data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda); + &data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84); + &data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a); + &data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06); + &data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02); + &data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b); + &data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea); + &data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73); + &data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85); + &data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e); + &data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89); + &data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b); + &data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20); + &data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4); + &data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31); + &data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f); + &data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d); + &data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef); + &data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0); + &data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61); + &data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26); + &data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d); + + &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38); + &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb); + &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87); + &data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb); + &data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d); + &data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e); + &data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2); + &data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25); + &data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16); + &data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92); + &data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda); + &data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84); + &data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a); + &data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06); + &data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02); + &data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b); + &data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea); + &data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73); + &data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85); + &data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e); + &data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89); + &data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b); + &data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20); + &data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4); + &data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31); + &data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f); + &data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d); + &data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef); + &data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0); + &data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61); + &data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26); + &data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d); + &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38); &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb); &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87); @@ -728,43 +1946,57 @@ sub declast() &function_end_B("_x86_AES_decrypt"); # void AES_decrypt (const void *inp,void *out,const AES_KEY *key); -&public_label("AES_Td"); &function_begin("AES_decrypt"); &mov ($acc,&wparam(0)); # load inp &mov ($key,&wparam(2)); # load key &mov ($s0,"esp"); - &sub ("esp",24); - &and ("esp",-64); - &add ("esp",4); - &mov (&DWP(16,"esp"),$s0); + &sub ("esp",36); + &and ("esp",-64); # align to cache-line + + # place stack frame just "above" the key schedule + &lea ($s1,&DWP(-64-63,$key)); + &sub ($s1,"esp"); + &neg ($s1); + &and ($s1,0x3C0); # modulo 1024, but aligned to cache-line + &sub ("esp",$s1); + &add ("esp",4); # 4 is reserved for caller's return address + &mov ($_esp,$s0); # save stack pointer &call (&label("pic_point")); # make it PIC! &set_label("pic_point"); - &blindpop("ebp"); - &lea ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp")); - - # prefetch Td4 - &lea ("ebp",&DWP(2048+128,"ebp")); - &mov ($s0,&DWP(0-128,"ebp")); - &mov ($s1,&DWP(32-128,"ebp")); - &mov ($s2,&DWP(64-128,"ebp")); - &mov ($s3,&DWP(96-128,"ebp")); - &mov ($s0,&DWP(128-128,"ebp")); - &mov ($s1,&DWP(160-128,"ebp")); - &mov ($s2,&DWP(192-128,"ebp")); - &mov ($s3,&DWP(224-128,"ebp")); - &lea ("ebp",&DWP(-2048-128,"ebp")); - + &blindpop($tbl); + &picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if(!$x86only); + &lea ($tbl,&DWP(&label("AES_Td")."-".&label("pic_point"),$tbl)); + + # pick Td4 copy which can't "overlap" with stack frame or key schedule + &lea ($s1,&DWP(768-4,"esp")); + &sub ($s1,$tbl); + &and ($s1,0x300); + &lea ($tbl,&DWP(2048+128,$tbl,$s1)); + + if (!$x86only) { + &bt (&DWP(0,$s0),25); # check for SSE bit + &jnc (&label("x86")); + + &movq ("mm0",&QWP(0,$acc)); + &movq ("mm4",&QWP(8,$acc)); + &call ("_sse_AES_decrypt_compact"); + &mov ("esp",$_esp); # restore stack pointer + &mov ($acc,&wparam(1)); # load out + &movq (&QWP(0,$acc),"mm0"); # write output data + &movq (&QWP(8,$acc),"mm4"); + &emms (); + &function_end_A(); + } + &set_label("x86",16); + &mov ($_tbl,$tbl); &mov ($s0,&DWP(0,$acc)); # load input data &mov ($s1,&DWP(4,$acc)); &mov ($s2,&DWP(8,$acc)); &mov ($s3,&DWP(12,$acc)); - - &call ("_x86_AES_decrypt"); - - &mov ("esp",&DWP(16,"esp")); - + &call ("_x86_AES_decrypt_compact"); + &mov ("esp",$_esp); # restore stack pointer &mov ($acc,&wparam(1)); # load out &mov (&DWP(0,$acc),$s0); # write output data &mov (&DWP(4,$acc),$s1); @@ -777,126 +2009,136 @@ sub declast() # unsigned char *ivp,const int enc); { # stack frame layout -# -4(%esp) 0(%esp) return address -# 0(%esp) 4(%esp) tmp1 -# 4(%esp) 8(%esp) tmp2 -# 8(%esp) 12(%esp) key -# 12(%esp) 16(%esp) end of key schedule -my $_esp=&DWP(16,"esp"); #saved %esp -my $_inp=&DWP(20,"esp"); #copy of wparam(0) -my $_out=&DWP(24,"esp"); #copy of wparam(1) -my $_len=&DWP(28,"esp"); #copy of wparam(2) -my $_key=&DWP(32,"esp"); #copy of wparam(3) -my $_ivp=&DWP(36,"esp"); #copy of wparam(4) -my $_tmp=&DWP(40,"esp"); #volatile variable -my $ivec=&DWP(44,"esp"); #ivec[16] -my $aes_key=&DWP(60,"esp"); #copy of aes_key -my $mark=&DWP(60+240,"esp"); #copy of aes_key->rounds - -&public_label("AES_Te"); -&public_label("AES_Td"); +# -4(%esp) # return address 0(%esp) +# 0(%esp) # s0 backing store 4(%esp) +# 4(%esp) # s1 backing store 8(%esp) +# 8(%esp) # s2 backing store 12(%esp) +# 12(%esp) # s3 backing store 16(%esp) +# 16(%esp) # key backup 20(%esp) +# 20(%esp) # end of key schedule 24(%esp) +# 24(%esp) # %ebp backup 28(%esp) +# 28(%esp) # %esp backup +my $_inp=&DWP(32,"esp"); # copy of wparam(0) +my $_out=&DWP(36,"esp"); # copy of wparam(1) +my $_len=&DWP(40,"esp"); # copy of wparam(2) +my $_key=&DWP(44,"esp"); # copy of wparam(3) +my $_ivp=&DWP(48,"esp"); # copy of wparam(4) +my $_tmp=&DWP(52,"esp"); # volatile variable +# +my $ivec=&DWP(60,"esp"); # ivec[16] +my $aes_key=&DWP(76,"esp"); # copy of aes_key +my $mark=&DWP(76+240,"esp"); # copy of aes_key->rounds + &function_begin("AES_cbc_encrypt"); &mov ($s2 eq "ecx"? $s2 : "",&wparam(2)); # load len &cmp ($s2,0); - &je (&label("enc_out")); + &je (&label("drop_out")); &call (&label("pic_point")); # make it PIC! &set_label("pic_point"); - &blindpop("ebp"); - - &pushf (); - &cld (); + &blindpop($tbl); + &picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if(!$x86only); &cmp (&wparam(5),0); - &je (&label("DECRYPT")); - - &lea ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp")); + &lea ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl)); + &jne (&label("picked_te")); + &lea ($tbl,&DWP(&label("AES_Td")."-".&label("AES_Te"),$tbl)); + &set_label("picked_te"); - # allocate aligned stack frame... - &lea ($key,&DWP(-64-244,"esp")); - &and ($key,-64); + # one can argue if this is required + &pushf (); + &cld (); - # ... and make sure it doesn't alias with AES_Te modulo 4096 - &mov ($s0,"ebp"); - &lea ($s1,&DWP(2048,"ebp")); - &mov ($s3,$key); + &cmp ($s2,$speed_limit); + &jb (&label("slow_way")); + &test ($s2,15); + &jnz (&label("slow_way")); + if (!$x86only) { + &bt (&DWP(0,$s0),28); # check for hyper-threading bit + &jc (&label("slow_way")); + } + # pre-allocate aligned stack frame... + &lea ($acc,&DWP(-80-244,"esp")); + &and ($acc,-64); + + # ... and make sure it doesn't alias with $tbl modulo 4096 + &mov ($s0,$tbl); + &lea ($s1,&DWP(2048+256,$tbl)); + &mov ($s3,$acc); &and ($s0,0xfff); # s = %ebp&0xfff - &and ($s1,0xfff); # e = (%ebp+2048)&0xfff + &and ($s1,0xfff); # e = (%ebp+2048+256)&0xfff &and ($s3,0xfff); # p = %esp&0xfff &cmp ($s3,$s1); # if (p>=e) %esp =- (p-e); - &jb (&label("te_break_out")); + &jb (&label("tbl_break_out")); &sub ($s3,$s1); - &sub ($key,$s3); - &jmp (&label("te_ok")); - &set_label("te_break_out"); # else %esp -= (p-s)&0xfff + framesz; + &sub ($acc,$s3); + &jmp (&label("tbl_ok")); + &set_label("tbl_break_out",4); # else %esp -= (p-s)&0xfff + framesz; &sub ($s3,$s0); &and ($s3,0xfff); - &add ($s3,64+256); - &sub ($key,$s3); - &align (4); - &set_label("te_ok"); - - &mov ($s0,&wparam(0)); # load inp - &mov ($s1,&wparam(1)); # load out - &mov ($s3,&wparam(3)); # load key - &mov ($acc,&wparam(4)); # load ivp + &add ($s3,384); + &sub ($acc,$s3); + &set_label("tbl_ok",4); - &exch ("esp",$key); + &lea ($s3,&wparam(0)); # obtain pointer to parameter block + &exch ("esp",$acc); # allocate stack frame &add ("esp",4); # reserve for return address! - &mov ($_esp,$key); # save %esp + &mov ($_tbl,$tbl); # save %ebp + &mov ($_esp,$acc); # save %esp + + &mov ($s0,&DWP(0,$s3)); # load inp + &mov ($s1,&DWP(4,$s3)); # load out + #&mov ($s2,&DWP(8,$s3)); # load len + &mov ($key,&DWP(12,$s3)); # load key + &mov ($acc,&DWP(16,$s3)); # load ivp + &mov ($s3,&DWP(20,$s3)); # load enc flag &mov ($_inp,$s0); # save copy of inp &mov ($_out,$s1); # save copy of out &mov ($_len,$s2); # save copy of len - &mov ($_key,$s3); # save copy of key + &mov ($_key,$key); # save copy of key &mov ($_ivp,$acc); # save copy of ivp &mov ($mark,0); # copy of aes_key->rounds = 0; - if ($compromise) { - &cmp ($s2,$compromise); - &jb (&label("skip_ecopy")); - } # do we copy key schedule to stack? - &mov ($s1 eq "ebx" ? $s1 : "",$s3); + &mov ($s1 eq "ebx" ? $s1 : "",$key); &mov ($s2 eq "ecx" ? $s2 : "",244/4); - &sub ($s1,"ebp"); - &mov ("esi",$s3); + &sub ($s1,$tbl); + &mov ("esi",$key); &and ($s1,0xfff); &lea ("edi",$aes_key); - &cmp ($s1,2048); - &jb (&label("do_ecopy")); + &cmp ($s1,2048+256); + &jb (&label("do_copy")); &cmp ($s1,4096-244); - &jb (&label("skip_ecopy")); - &align (4); - &set_label("do_ecopy"); + &jb (&label("skip_copy")); + &set_label("do_copy",4); &mov ($_key,"edi"); &data_word(0xA5F3F689); # rep movsd - &set_label("skip_ecopy"); + &set_label("skip_copy"); - &mov ($acc,$s0); &mov ($key,16); - &align (4); - &set_label("prefetch_te"); - &mov ($s0,&DWP(0,"ebp")); - &mov ($s1,&DWP(32,"ebp")); - &mov ($s2,&DWP(64,"ebp")); - &mov ($s3,&DWP(96,"ebp")); - &lea ("ebp",&DWP(128,"ebp")); - &dec ($key); - &jnz (&label("prefetch_te")); - &sub ("ebp",2048); - - &mov ($s2,$_len); + &set_label("prefetch_tbl",4); + &mov ($s0,&DWP(0,$tbl)); + &mov ($s1,&DWP(32,$tbl)); + &mov ($s2,&DWP(64,$tbl)); + &mov ($acc,&DWP(96,$tbl)); + &lea ($tbl,&DWP(128,$tbl)); + &sub ($key,1); + &jnz (&label("prefetch_tbl")); + &sub ($tbl,2048); + + &mov ($acc,$_inp); &mov ($key,$_ivp); - &test ($s2,0xFFFFFFF0); - &jz (&label("enc_tail")); # short input... + &cmp ($s3,0); + &je (&label("fast_decrypt")); + +#----------------------------- ENCRYPT -----------------------------# &mov ($s0,&DWP(0,$key)); # load iv &mov ($s1,&DWP(4,$key)); - &align (4); - &set_label("enc_loop"); + &set_label("fast_enc_loop",16); &mov ($s2,&DWP(8,$key)); &mov ($s3,&DWP(12,$key)); @@ -916,22 +2158,16 @@ my $mark=&DWP(60+240,"esp"); #copy of aes_key->rounds &mov (&DWP(8,$key),$s2); &mov (&DWP(12,$key),$s3); + &lea ($acc,&DWP(16,$acc)); # advance inp &mov ($s2,$_len); # load len - - &lea ($acc,&DWP(16,$acc)); &mov ($_inp,$acc); # save inp - - &lea ($s3,&DWP(16,$key)); + &lea ($s3,&DWP(16,$key)); # advance out &mov ($_out,$s3); # save out - - &sub ($s2,16); - &test ($s2,0xFFFFFFF0); + &sub ($s2,16); # decrease len &mov ($_len,$s2); # save len - &jnz (&label("enc_loop")); - &test ($s2,15); - &jnz (&label("enc_tail")); + &jnz (&label("fast_enc_loop")); &mov ($acc,$_ivp); # load ivp - &mov ($s2,&DWP(8,$key)); # restore last dwords + &mov ($s2,&DWP(8,$key)); # restore last 2 dwords &mov ($s3,&DWP(12,$key)); &mov (&DWP(0,$acc),$s0); # save ivec &mov (&DWP(4,$acc),$s1); @@ -949,125 +2185,20 @@ my $mark=&DWP(60+240,"esp"); #copy of aes_key->rounds &set_label("skip_ezero") &mov ("esp",$_esp); &popf (); - &set_label("enc_out"); + &set_label("drop_out"); &function_end_A(); &pushf (); # kludge, never executed - &align (4); - &set_label("enc_tail"); - &mov ($s0,$key eq "edi" ? $key : ""); - &mov ($key,$_out); # load out - &push ($s0); # push ivp - &mov ($s1,16); - &sub ($s1,$s2); - &cmp ($key,$acc); # compare with inp - &je (&label("enc_in_place")); - &align (4); - &data_word(0xA4F3F689); # rep movsb # copy input - &jmp (&label("enc_skip_in_place")); - &set_label("enc_in_place"); - &lea ($key,&DWP(0,$key,$s2)); - &set_label("enc_skip_in_place"); - &mov ($s2,$s1); - &xor ($s0,$s0); - &align (4); - &data_word(0xAAF3F689); # rep stosb # zero tail - &pop ($key); # pop ivp - - &mov ($acc,$_out); # output as input - &mov ($s0,&DWP(0,$key)); - &mov ($s1,&DWP(4,$key)); - &mov ($_len,16); # len=16 - &jmp (&label("enc_loop")); # one more spin... - #----------------------------- DECRYPT -----------------------------# -&align (4); -&set_label("DECRYPT"); - &lea ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp")); - - # allocate aligned stack frame... - &lea ($key,&DWP(-64-244,"esp")); - &and ($key,-64); - - # ... and make sure it doesn't alias with AES_Td modulo 4096 - &mov ($s0,"ebp"); - &lea ($s1,&DWP(2048+256,"ebp")); - &mov ($s3,$key); - &and ($s0,0xfff); # s = %ebp&0xfff - &and ($s1,0xfff); # e = (%ebp+2048+256)&0xfff - &and ($s3,0xfff); # p = %esp&0xfff - - &cmp ($s3,$s1); # if (p>=e) %esp =- (p-e); - &jb (&label("td_break_out")); - &sub ($s3,$s1); - &sub ($key,$s3); - &jmp (&label("td_ok")); - &set_label("td_break_out"); # else %esp -= (p-s)&0xfff + framesz; - &sub ($s3,$s0); - &and ($s3,0xfff); - &add ($s3,64+256); - &sub ($key,$s3); - &align (4); - &set_label("td_ok"); - - &mov ($s0,&wparam(0)); # load inp - &mov ($s1,&wparam(1)); # load out - &mov ($s3,&wparam(3)); # load key - &mov ($acc,&wparam(4)); # load ivp - - &exch ("esp",$key); - &add ("esp",4); # reserve for return address! - &mov ($_esp,$key); # save %esp - - &mov ($_inp,$s0); # save copy of inp - &mov ($_out,$s1); # save copy of out - &mov ($_len,$s2); # save copy of len - &mov ($_key,$s3); # save copy of key - &mov ($_ivp,$acc); # save copy of ivp - - &mov ($mark,0); # copy of aes_key->rounds = 0; - if ($compromise) { - &cmp ($s2,$compromise); - &jb (&label("skip_dcopy")); - } - # do we copy key schedule to stack? - &mov ($s1 eq "ebx" ? $s1 : "",$s3); - &mov ($s2 eq "ecx" ? $s2 : "",244/4); - &sub ($s1,"ebp"); - &mov ("esi",$s3); - &and ($s1,0xfff); - &lea ("edi",$aes_key); - &cmp ($s1,2048+256); - &jb (&label("do_dcopy")); - &cmp ($s1,4096-244); - &jb (&label("skip_dcopy")); - &align (4); - &set_label("do_dcopy"); - &mov ($_key,"edi"); - &data_word(0xA5F3F689); # rep movsd - &set_label("skip_dcopy"); - - &mov ($acc,$s0); - &mov ($key,18); - &align (4); - &set_label("prefetch_td"); - &mov ($s0,&DWP(0,"ebp")); - &mov ($s1,&DWP(32,"ebp")); - &mov ($s2,&DWP(64,"ebp")); - &mov ($s3,&DWP(96,"ebp")); - &lea ("ebp",&DWP(128,"ebp")); - &dec ($key); - &jnz (&label("prefetch_td")); - &sub ("ebp",2048+256); +&set_label("fast_decrypt",16); &cmp ($acc,$_out); - &je (&label("dec_in_place")); # in-place processing... + &je (&label("fast_dec_in_place")); # in-place processing... - &mov ($key,$_ivp); # load ivp &mov ($_tmp,$key); &align (4); - &set_label("dec_loop"); + &set_label("fast_dec_loop",16); &mov ($s0,&DWP(0,$acc)); # read input &mov ($s1,&DWP(4,$acc)); &mov ($s2,&DWP(8,$acc)); @@ -1083,27 +2214,24 @@ my $mark=&DWP(60+240,"esp"); #copy of aes_key->rounds &xor ($s2,&DWP(8,$key)); &xor ($s3,&DWP(12,$key)); - &sub ($acc,16); - &jc (&label("dec_partial")); - &mov ($_len,$acc); # save len - &mov ($acc,$_inp); # load inp &mov ($key,$_out); # load out + &mov ($acc,$_inp); # load inp &mov (&DWP(0,$key),$s0); # write output &mov (&DWP(4,$key),$s1); &mov (&DWP(8,$key),$s2); &mov (&DWP(12,$key),$s3); + &mov ($s2,$_len); # load len &mov ($_tmp,$acc); # save ivp - &lea ($acc,&DWP(16,$acc)); + &lea ($acc,&DWP(16,$acc)); # advance inp &mov ($_inp,$acc); # save inp - - &lea ($key,&DWP(16,$key)); + &lea ($key,&DWP(16,$key)); # advance out &mov ($_out,$key); # save out - - &jnz (&label("dec_loop")); + &sub ($s2,16); # decrease len + &mov ($_len,$s2); # save len + &jnz (&label("fast_dec_loop")); &mov ($key,$_tmp); # load temp ivp - &set_label("dec_end"); &mov ($acc,$_ivp); # load user ivp &mov ($s0,&DWP(0,$key)); # load iv &mov ($s1,&DWP(4,$key)); @@ -1113,31 +2241,16 @@ my $mark=&DWP(60+240,"esp"); #copy of aes_key->rounds &mov (&DWP(4,$acc),$s1); &mov (&DWP(8,$acc),$s2); &mov (&DWP(12,$acc),$s3); - &jmp (&label("dec_out")); + &jmp (&label("fast_dec_out")); - &align (4); - &set_label("dec_partial"); - &lea ($key,$ivec); - &mov (&DWP(0,$key),$s0); # dump output to stack - &mov (&DWP(4,$key),$s1); - &mov (&DWP(8,$key),$s2); - &mov (&DWP(12,$key),$s3); - &lea ($s2 eq "ecx" ? $s2 : "",&DWP(16,$acc)); - &mov ($acc eq "esi" ? $acc : "",$key); - &mov ($key eq "edi" ? $key : "",$_out); # load out - &data_word(0xA4F3F689); # rep movsb # copy output - &mov ($key,$_inp); # use inp as temp ivp - &jmp (&label("dec_end")); - - &align (4); - &set_label("dec_in_place"); - &set_label("dec_in_place_loop"); - &lea ($key,$ivec); + &set_label("fast_dec_in_place",16); + &set_label("fast_dec_in_place_loop"); &mov ($s0,&DWP(0,$acc)); # read input &mov ($s1,&DWP(4,$acc)); &mov ($s2,&DWP(8,$acc)); &mov ($s3,&DWP(12,$acc)); + &lea ($key,$ivec); &mov (&DWP(0,$key),$s0); # copy to temp &mov (&DWP(4,$key),$s1); &mov (&DWP(8,$key),$s2); @@ -1158,7 +2271,7 @@ my $mark=&DWP(60+240,"esp"); #copy of aes_key->rounds &mov (&DWP(8,$acc),$s2); &mov (&DWP(12,$acc),$s3); - &lea ($acc,&DWP(16,$acc)); + &lea ($acc,&DWP(16,$acc)); # advance out &mov ($_out,$acc); # save out &lea ($acc,$ivec); @@ -1173,40 +2286,340 @@ my $mark=&DWP(60+240,"esp"); #copy of aes_key->rounds &mov (&DWP(12,$key),$s3); &mov ($acc,$_inp); # load inp + &mov ($s2,$_len); # load len + &lea ($acc,&DWP(16,$acc)); # advance inp + &mov ($_inp,$acc); # save inp + &sub ($s2,16); # decrease len + &mov ($_len,$s2); # save len + &jnz (&label("fast_dec_in_place_loop")); + + &set_label("fast_dec_out",4); + &cmp ($mark,0); # was the key schedule copied? + &mov ("edi",$_key); + &je (&label("skip_dzero")); + # zero copy of key schedule + &mov ("ecx",240/4); + &xor ("eax","eax"); + &align (4); + &data_word(0xABF3F689); # rep stosd + &set_label("skip_dzero") + &mov ("esp",$_esp); + &popf (); + &function_end_A(); + &pushf (); # kludge, never executed + +#--------------------------- SLOW ROUTINE ---------------------------# +&set_label("slow_way",16); + + &mov ($s0,&DWP(0,$s0)) if (!$x86only);# load OPENSSL_ia32cap + &mov ($key,&wparam(3)); # load key + + # pre-allocate aligned stack frame... + &lea ($acc,&DWP(-80,"esp")); + &and ($acc,-64); + + # ... and make sure it doesn't alias with $key modulo 1024 + &lea ($s1,&DWP(-80-63,$key)); + &sub ($s1,$acc); + &neg ($s1); + &and ($s1,0x3C0); # modulo 1024, but aligned to cache-line + &sub ($acc,$s1); + + # pick S-box copy which can't overlap with stack frame or $key + &lea ($s1,&DWP(768,$acc)); + &sub ($s1,$tbl); + &and ($s1,0x300); + &lea ($tbl,&DWP(2048+128,$tbl,$s1)); + + &lea ($s3,&wparam(0)); # pointer to parameter block + + &exch ("esp",$acc); + &add ("esp",4); # reserve for return address! + &mov ($_tbl,$tbl); # save %ebp + &mov ($_esp,$acc); # save %esp + &mov ($_tmp,$s0); # save OPENSSL_ia32cap + + &mov ($s0,&DWP(0,$s3)); # load inp + &mov ($s1,&DWP(4,$s3)); # load out + #&mov ($s2,&DWP(8,$s3)); # load len + #&mov ($key,&DWP(12,$s3)); # load key + &mov ($acc,&DWP(16,$s3)); # load ivp + &mov ($s3,&DWP(20,$s3)); # load enc flag + + &mov ($_inp,$s0); # save copy of inp + &mov ($_out,$s1); # save copy of out + &mov ($_len,$s2); # save copy of len + &mov ($_key,$key); # save copy of key + &mov ($_ivp,$acc); # save copy of ivp + + &mov ($key,$acc); + &mov ($acc,$s0); + + &cmp ($s3,0); + &je (&label("slow_decrypt")); + +#--------------------------- SLOW ENCRYPT ---------------------------# + &cmp ($s2,16); + &mov ($s3,$s1); + &jb (&label("slow_enc_tail")); + + if (!$x86only) { + &bt ($_tmp,25); # check for SSE bit + &jnc (&label("slow_enc_x86")); - &lea ($acc,&DWP(16,$acc)); + &movq ("mm0",&QWP(0,$key)); # load iv + &movq ("mm4",&QWP(8,$key)); + + &set_label("slow_enc_loop_sse",16); + &pxor ("mm0",&QWP(0,$acc)); # xor input data + &pxor ("mm4",&QWP(8,$acc)); + + &mov ($key,$_key); + &call ("_sse_AES_encrypt_compact"); + + &mov ($acc,$_inp); # load inp + &mov ($key,$_out); # load out + &mov ($s2,$_len); # load len + + &movq (&QWP(0,$key),"mm0"); # save output data + &movq (&QWP(8,$key),"mm4"); + + &lea ($acc,&DWP(16,$acc)); # advance inp &mov ($_inp,$acc); # save inp + &lea ($s3,&DWP(16,$key)); # advance out + &mov ($_out,$s3); # save out + &sub ($s2,16); # decrease len + &cmp ($s2,16); + &mov ($_len,$s2); # save len + &jae (&label("slow_enc_loop_sse")); + &test ($s2,15); + &jnz (&label("slow_enc_tail")); + &mov ($acc,$_ivp); # load ivp + &movq (&QWP(0,$acc),"mm0"); # save ivec + &movq (&QWP(8,$acc),"mm4"); + &emms (); + &mov ("esp",$_esp); + &popf (); + &function_end_A(); + &pushf (); # kludge, never executed + } + &set_label("slow_enc_x86",16); + &mov ($s0,&DWP(0,$key)); # load iv + &mov ($s1,&DWP(4,$key)); + + &set_label("slow_enc_loop_x86",4); + &mov ($s2,&DWP(8,$key)); + &mov ($s3,&DWP(12,$key)); + + &xor ($s0,&DWP(0,$acc)); # xor input data + &xor ($s1,&DWP(4,$acc)); + &xor ($s2,&DWP(8,$acc)); + &xor ($s3,&DWP(12,$acc)); + + &mov ($key,$_key); # load key + &call ("_x86_AES_encrypt_compact"); + + &mov ($acc,$_inp); # load inp + &mov ($key,$_out); # load out + + &mov (&DWP(0,$key),$s0); # save output data + &mov (&DWP(4,$key),$s1); + &mov (&DWP(8,$key),$s2); + &mov (&DWP(12,$key),$s3); &mov ($s2,$_len); # load len - &sub ($s2,16); - &jc (&label("dec_in_place_partial")); + &lea ($acc,&DWP(16,$acc)); # advance inp + &mov ($_inp,$acc); # save inp + &lea ($s3,&DWP(16,$key)); # advance out + &mov ($_out,$s3); # save out + &sub ($s2,16); # decrease len + &cmp ($s2,16); &mov ($_len,$s2); # save len - &jnz (&label("dec_in_place_loop")); - &jmp (&label("dec_out")); - - &align (4); - &set_label("dec_in_place_partial"); - # one can argue if this is actually required... - &mov ($key eq "edi" ? $key : "",$_out); - &lea ($acc eq "esi" ? $acc : "",$ivec); + &jae (&label("slow_enc_loop_x86")); + &test ($s2,15); + &jnz (&label("slow_enc_tail")); + &mov ($acc,$_ivp); # load ivp + &mov ($s2,&DWP(8,$key)); # restore last dwords + &mov ($s3,&DWP(12,$key)); + &mov (&DWP(0,$acc),$s0); # save ivec + &mov (&DWP(4,$acc),$s1); + &mov (&DWP(8,$acc),$s2); + &mov (&DWP(12,$acc),$s3); + + &mov ("esp",$_esp); + &popf (); + &function_end_A(); + &pushf (); # kludge, never executed + + &set_label("slow_enc_tail",16); + &emms () if (!$x86only); + &mov ($key eq "edi"? $key:"",$s3); # load out to edi + &mov ($s1,16); + &sub ($s1,$s2); + &cmp ($key,$acc eq "esi"? $acc:""); # compare with inp + &je (&label("enc_in_place")); + &align (4); + &data_word(0xA4F3F689); # rep movsb # copy input + &jmp (&label("enc_skip_in_place")); + &set_label("enc_in_place"); &lea ($key,&DWP(0,$key,$s2)); - &lea ($acc,&DWP(16,$acc,$s2)); - &neg ($s2 eq "ecx" ? $s2 : ""); - &data_word(0xA4F3F689); # rep movsb # restore tail - - &align (4); - &set_label("dec_out"); - &cmp ($mark,0); # was the key schedule copied? - &mov ("edi",$_key); - &je (&label("skip_dzero")); - # zero copy of key schedule - &mov ("ecx",240/4); - &xor ("eax","eax"); - &align (4); - &data_word(0xABF3F689); # rep stosd - &set_label("skip_dzero") - &mov ("esp",$_esp); - &popf (); + &set_label("enc_skip_in_place"); + &mov ($s2,$s1); + &xor ($s0,$s0); + &align (4); + &data_word(0xAAF3F689); # rep stosb # zero tail + + &mov ($key,$_ivp); # restore ivp + &mov ($acc,$s3); # output as input + &mov ($s0,&DWP(0,$key)); + &mov ($s1,&DWP(4,$key)); + &mov ($_len,16); # len=16 + &jmp (&label("slow_enc_loop_x86")); # one more spin... + +#--------------------------- SLOW DECRYPT ---------------------------# +&set_label("slow_decrypt",16); + if (!$x86only) { + &bt ($_tmp,25); # check for SSE bit + &jnc (&label("slow_dec_loop_x86")); + + &set_label("slow_dec_loop_sse",4); + &movq ("mm0",&QWP(0,$acc)); # read input + &movq ("mm4",&QWP(8,$acc)); + + &mov ($key,$_key); + &call ("_sse_AES_decrypt_compact"); + + &mov ($acc,$_inp); # load inp + &lea ($s0,$ivec); + &mov ($s1,$_out); # load out + &mov ($s2,$_len); # load len + &mov ($key,$_ivp); # load ivp + + &movq ("mm1",&QWP(0,$acc)); # re-read input + &movq ("mm5",&QWP(8,$acc)); + + &pxor ("mm0",&QWP(0,$key)); # xor iv + &pxor ("mm4",&QWP(8,$key)); + + &movq (&QWP(0,$key),"mm1"); # copy input to iv + &movq (&QWP(8,$key),"mm5"); + + &sub ($s2,16); # decrease len + &jc (&label("slow_dec_partial_sse")); + + &movq (&QWP(0,$s1),"mm0"); # write output + &movq (&QWP(8,$s1),"mm4"); + + &lea ($s1,&DWP(16,$s1)); # advance out + &mov ($_out,$s1); # save out + &lea ($acc,&DWP(16,$acc)); # advance inp + &mov ($_inp,$acc); # save inp + &mov ($_len,$s2); # save len + &jnz (&label("slow_dec_loop_sse")); + &emms (); + &mov ("esp",$_esp); + &popf (); + &function_end_A(); + &pushf (); # kludge, never executed + + &set_label("slow_dec_partial_sse",16); + &movq (&QWP(0,$s0),"mm0"); # save output to temp + &movq (&QWP(8,$s0),"mm4"); + &emms (); + + &add ($s2 eq "ecx" ? "ecx":"",16); + &mov ("edi",$s1); # out + &mov ("esi",$s0); # temp + &align (4); + &data_word(0xA4F3F689); # rep movsb # copy partial output + + &mov ("esp",$_esp); + &popf (); + &function_end_A(); + &pushf (); # kludge, never executed + } + &set_label("slow_dec_loop_x86",16); + &mov ($s0,&DWP(0,$acc)); # read input + &mov ($s1,&DWP(4,$acc)); + &mov ($s2,&DWP(8,$acc)); + &mov ($s3,&DWP(12,$acc)); + + &lea ($key,$ivec); + &mov (&DWP(0,$key),$s0); # copy to temp + &mov (&DWP(4,$key),$s1); + &mov (&DWP(8,$key),$s2); + &mov (&DWP(12,$key),$s3); + + &mov ($key,$_key); # load key + &call ("_x86_AES_decrypt_compact"); + + &mov ($key,$_ivp); # load ivp + &mov ($acc,$_len); # load len + &xor ($s0,&DWP(0,$key)); # xor iv + &xor ($s1,&DWP(4,$key)); + &xor ($s2,&DWP(8,$key)); + &xor ($s3,&DWP(12,$key)); + + &sub ($acc,16); + &jc (&label("slow_dec_partial_x86")); + + &mov ($_len,$acc); # save len + &mov ($acc,$_out); # load out + + &mov (&DWP(0,$acc),$s0); # write output + &mov (&DWP(4,$acc),$s1); + &mov (&DWP(8,$acc),$s2); + &mov (&DWP(12,$acc),$s3); + + &lea ($acc,&DWP(16,$acc)); # advance out + &mov ($_out,$acc); # save out + + &lea ($acc,$ivec); + &mov ($s0,&DWP(0,$acc)); # read temp + &mov ($s1,&DWP(4,$acc)); + &mov ($s2,&DWP(8,$acc)); + &mov ($s3,&DWP(12,$acc)); + + &mov (&DWP(0,$key),$s0); # copy it to iv + &mov (&DWP(4,$key),$s1); + &mov (&DWP(8,$key),$s2); + &mov (&DWP(12,$key),$s3); + + &mov ($acc,$_inp); # load inp + &lea ($acc,&DWP(16,$acc)); # advance inp + &mov ($_inp,$acc); # save inp + &jnz (&label("slow_dec_loop_x86")); + &mov ("esp",$_esp); + &popf (); + &function_end_A(); + &pushf (); # kludge, never executed + + &set_label("slow_dec_partial_x86",16); + &lea ($acc,$ivec); + &mov (&DWP(0,$acc),$s0); # save output to temp + &mov (&DWP(4,$acc),$s1); + &mov (&DWP(8,$acc),$s2); + &mov (&DWP(12,$acc),$s3); + + &mov ($acc,$_inp); + &mov ($s0,&DWP(0,$acc)); # re-read input + &mov ($s1,&DWP(4,$acc)); + &mov ($s2,&DWP(8,$acc)); + &mov ($s3,&DWP(12,$acc)); + + &mov (&DWP(0,$key),$s0); # copy it to iv + &mov (&DWP(4,$key),$s1); + &mov (&DWP(8,$key),$s2); + &mov (&DWP(12,$key),$s3); + + &mov ("ecx",$_len); + &mov ("edi",$_out); + &lea ("esi",$ivec); + &align (4); + &data_word(0xA4F3F689); # rep movsb # copy partial output + + &mov ("esp",$_esp); + &popf (); &function_end("AES_cbc_encrypt"); } @@ -1215,35 +2628,31 @@ my $mark=&DWP(60+240,"esp"); #copy of aes_key->rounds sub enckey() { &movz ("esi",&LB("edx")); # rk[i]>>0 - &mov ("ebx",&DWP(2,"ebp","esi",8)); + &movz ("ebx",&BP(-128,$tbl,"esi",1)); &movz ("esi",&HB("edx")); # rk[i]>>8 - &and ("ebx",0xFF000000); + &shl ("ebx",24); &xor ("eax","ebx"); - &mov ("ebx",&DWP(2,"ebp","esi",8)); + &movz ("ebx",&BP(-128,$tbl,"esi",1)); &shr ("edx",16); - &and ("ebx",0x000000FF); &movz ("esi",&LB("edx")); # rk[i]>>16 &xor ("eax","ebx"); - &mov ("ebx",&DWP(0,"ebp","esi",8)); + &movz ("ebx",&BP(-128,$tbl,"esi",1)); &movz ("esi",&HB("edx")); # rk[i]>>24 - &and ("ebx",0x0000FF00); + &shl ("ebx",8); &xor ("eax","ebx"); - &mov ("ebx",&DWP(0,"ebp","esi",8)); - &and ("ebx",0x00FF0000); + &movz ("ebx",&BP(-128,$tbl,"esi",1)); + &shl ("ebx",16); &xor ("eax","ebx"); - &xor ("eax",&DWP(2048,"ebp","ecx",4)); # rcon + &xor ("eax",&DWP(1024-128,$tbl,"ecx",4)); # rcon } -# int AES_set_encrypt_key(const unsigned char *userKey, const int bits, -# AES_KEY *key) -&public_label("AES_Te"); -&function_begin("AES_set_encrypt_key"); - &mov ("esi",&wparam(0)); # user supplied key - &mov ("edi",&wparam(2)); # private key schedule +&function_begin("_x86_AES_set_encrypt_key"); + &mov ("esi",&wparam(1)); # user supplied key + &mov ("edi",&wparam(3)); # private key schedule &test ("esi",-1); &jz (&label("badpointer")); @@ -1252,10 +2661,21 @@ sub enckey() &call (&label("pic_point")); &set_label("pic_point"); - &blindpop("ebp"); - &lea ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp")); - - &mov ("ecx",&wparam(1)); # number of bits in key + &blindpop($tbl); + &lea ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl)); + &lea ($tbl,&DWP(2048+128,$tbl)); + + # prefetch Te4 + &mov ("eax",&DWP(0-128,$tbl)); + &mov ("ebx",&DWP(32-128,$tbl)); + &mov ("ecx",&DWP(64-128,$tbl)); + &mov ("edx",&DWP(96-128,$tbl)); + &mov ("eax",&DWP(128-128,$tbl)); + &mov ("ebx",&DWP(160-128,$tbl)); + &mov ("ecx",&DWP(192-128,$tbl)); + &mov ("edx",&DWP(224-128,$tbl)); + + &mov ("ecx",&wparam(2)); # number of bits in key &cmp ("ecx",128); &je (&label("10rounds")); &cmp ("ecx",192); @@ -1394,24 +2814,23 @@ sub enckey() &mov ("edx","eax"); &mov ("eax",&DWP(16,"edi")); # rk[4] &movz ("esi",&LB("edx")); # rk[11]>>0 - &mov ("ebx",&DWP(2,"ebp","esi",8)); + &movz ("ebx",&BP(-128,$tbl,"esi",1)); &movz ("esi",&HB("edx")); # rk[11]>>8 - &and ("ebx",0x000000FF); &xor ("eax","ebx"); - &mov ("ebx",&DWP(0,"ebp","esi",8)); + &movz ("ebx",&BP(-128,$tbl,"esi",1)); &shr ("edx",16); - &and ("ebx",0x0000FF00); + &shl ("ebx",8); &movz ("esi",&LB("edx")); # rk[11]>>16 &xor ("eax","ebx"); - &mov ("ebx",&DWP(0,"ebp","esi",8)); + &movz ("ebx",&BP(-128,$tbl,"esi",1)); &movz ("esi",&HB("edx")); # rk[11]>>24 - &and ("ebx",0x00FF0000); + &shl ("ebx",16); &xor ("eax","ebx"); - &mov ("ebx",&DWP(2,"ebp","esi",8)); - &and ("ebx",0xFF000000); + &movz ("ebx",&BP(-128,$tbl,"esi",1)); + &shl ("ebx",24); &xor ("eax","ebx"); &mov (&DWP(48,"edi"),"eax"); # rk[12] @@ -1433,43 +2852,74 @@ sub enckey() &set_label("badpointer"); &mov ("eax",-1); &set_label("exit"); -&function_end("AES_set_encrypt_key"); +&function_end("_x86_AES_set_encrypt_key"); -sub deckey() -{ my ($i,$ptr,$te,$td) = @_; +# int AES_set_encrypt_key(const unsigned char *userKey, const int bits, +# AES_KEY *key) +&function_begin_B("AES_set_encrypt_key"); + &call ("_x86_AES_set_encrypt_key"); + &ret (); +&function_end_B("AES_set_encrypt_key"); - &mov ("eax",&DWP($i,$ptr)); - &mov ("edx","eax"); - &movz ("ebx",&HB("eax")); - &shr ("edx",16); - &and ("eax",0xFF); - &movz ("eax",&BP(2,$te,"eax",8)); - &movz ("ebx",&BP(2,$te,"ebx",8)); - &mov ("eax",&DWP(0,$td,"eax",8)); - &xor ("eax",&DWP(3,$td,"ebx",8)); - &movz ("ebx",&HB("edx")); - &and ("edx",0xFF); - &movz ("edx",&BP(2,$te,"edx",8)); - &movz ("ebx",&BP(2,$te,"ebx",8)); - &xor ("eax",&DWP(2,$td,"edx",8)); - &xor ("eax",&DWP(1,$td,"ebx",8)); - &mov (&DWP($i,$ptr),"eax"); +sub deckey() +{ my ($i,$key,$tp1,$tp2,$tp4,$tp8) = @_; + my $tmp = $tbl; + + &mov ($acc,$tp1); + &and ($acc,0x80808080); + &mov ($tmp,$acc); + &shr ($tmp,7); + &lea ($tp2,&DWP(0,$tp1,$tp1)); + &sub ($acc,$tmp); + &and ($tp2,0xfefefefe); + &and ($acc,0x1b1b1b1b); + &xor ($acc,$tp2); + &mov ($tp2,$acc); + + &and ($acc,0x80808080); + &mov ($tmp,$acc); + &shr ($tmp,7); + &lea ($tp4,&DWP(0,$tp2,$tp2)); + &sub ($acc,$tmp); + &and ($tp4,0xfefefefe); + &and ($acc,0x1b1b1b1b); + &xor ($tp2,$tp1); # tp2^tp1 + &xor ($acc,$tp4); + &mov ($tp4,$acc); + + &and ($acc,0x80808080); + &mov ($tmp,$acc); + &shr ($tmp,7); + &lea ($tp8,&DWP(0,$tp4,$tp4)); + &xor ($tp4,$tp1); # tp4^tp1 + &sub ($acc,$tmp); + &and ($tp8,0xfefefefe); + &and ($acc,0x1b1b1b1b); + &rotl ($tp1,8); # = ROTATE(tp1,8) + &xor ($tp8,$acc); + + &mov ($tmp,&DWP(4*($i+1),$key)); # modulo-scheduled load + + &xor ($tp1,$tp2); + &xor ($tp2,$tp8); + &xor ($tp1,$tp4); + &rotl ($tp2,24); + &xor ($tp4,$tp8); + &xor ($tp1,$tp8); # ^= tp8^(tp4^tp1)^(tp2^tp1) + &rotl ($tp4,16); + &xor ($tp1,$tp2); # ^= ROTATE(tp8^tp2^tp1,24) + &rotl ($tp8,8); + &xor ($tp1,$tp4); # ^= ROTATE(tp8^tp4^tp1,16) + &mov ($tp2,$tmp); + &xor ($tp1,$tp8); # ^= ROTATE(tp8,8) + + &mov (&DWP(4*$i,$key),$tp1); } # int AES_set_decrypt_key(const unsigned char *userKey, const int bits, # AES_KEY *key) -&public_label("AES_Td"); -&public_label("AES_Te"); &function_begin_B("AES_set_decrypt_key"); - &mov ("eax",&wparam(0)); - &mov ("ecx",&wparam(1)); - &mov ("edx",&wparam(2)); - &sub ("esp",12); - &mov (&DWP(0,"esp"),"eax"); - &mov (&DWP(4,"esp"),"ecx"); - &mov (&DWP(8,"esp"),"edx"); - &call ("AES_set_encrypt_key"); - &add ("esp",12); + &call ("_x86_AES_set_encrypt_key"); &cmp ("eax",0); &je (&label("proceed")); &ret (); @@ -1485,8 +2935,7 @@ sub deckey() &lea ("ecx",&DWP(0,"","ecx",4)); &lea ("edi",&DWP(0,"esi","ecx",4)); # pointer to last chunk - &align (4); - &set_label("invert"); # invert order of chunks + &set_label("invert",4); # invert order of chunks &mov ("eax",&DWP(0,"esi")); &mov ("ebx",&DWP(4,"esi")); &mov ("ecx",&DWP(0,"edi")); @@ -1508,26 +2957,24 @@ sub deckey() &cmp ("esi","edi"); &jne (&label("invert")); - &call (&label("pic_point")); - &set_label("pic_point"); - blindpop("ebp"); - &lea ("edi",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp")); - &lea ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp")); - - &mov ("esi",&wparam(2)); - &mov ("ecx",&DWP(240,"esi")); # pull number of rounds - &dec ("ecx"); - &align (4); - &set_label("permute"); # permute the key schedule - &add ("esi",16); - &deckey (0,"esi","ebp","edi"); - &deckey (4,"esi","ebp","edi"); - &deckey (8,"esi","ebp","edi"); - &deckey (12,"esi","ebp","edi"); - &dec ("ecx"); - &jnz (&label("permute")); + &mov ($key,&wparam(2)); + &mov ($acc,&DWP(240,$key)); # pull number of rounds + &lea ($acc,&DWP(-2,$acc,$acc)); + &lea ($acc,&DWP(0,$key,$acc,8)); + &mov (&wparam(2),$acc); + + &mov ($s0,&DWP(16,$key)); # modulo-scheduled load + &set_label("permute",4); # permute the key schedule + &add ($key,16); + &deckey (0,$key,$s0,$s1,$s2,$s3); + &deckey (1,$key,$s1,$s2,$s3,$s0); + &deckey (2,$key,$s2,$s3,$s0,$s1); + &deckey (3,$key,$s3,$s0,$s1,$s2); + &cmp ($key,&wparam(2)); + &jb (&label("permute")); &xor ("eax","eax"); # return success &function_end("AES_set_decrypt_key"); +&asciz("AES for x86, CRYPTOGAMS by <appro\@openssl.org>"); &asm_finish(); |