diff options
Diffstat (limited to 'crypto/openssl/crypto/aes')
| -rw-r--r-- | crypto/openssl/crypto/aes/Makefile | 40 | ||||
| -rw-r--r-- | crypto/openssl/crypto/aes/aes.h | 10 | ||||
| -rw-r--r-- | crypto/openssl/crypto/aes/aes_cbc.c | 44 | ||||
| -rw-r--r-- | crypto/openssl/crypto/aes/aes_cfb.c | 67 | ||||
| -rw-r--r-- | crypto/openssl/crypto/aes/aes_core.c | 25 | ||||
| -rw-r--r-- | crypto/openssl/crypto/aes/aes_locl.h | 6 | ||||
| -rwxr-xr-x | crypto/openssl/crypto/aes/asm/aes-586.pl | 1531 | ||||
| -rw-r--r-- | crypto/openssl/crypto/aes/asm/aes-ia64.S | 1652 |
8 files changed, 3289 insertions, 86 deletions
diff --git a/crypto/openssl/crypto/aes/Makefile b/crypto/openssl/crypto/aes/Makefile index 71087bc18f1e..057b70bdd157 100644 --- a/crypto/openssl/crypto/aes/Makefile +++ b/crypto/openssl/crypto/aes/Makefile @@ -8,16 +8,14 @@ CC= cc CPP= $(CC) -E INCLUDES= CFLAG=-g -INSTALL_PREFIX= -OPENSSLDIR= /usr/local/ssl -INSTALLTOP= /usr/local/ssl -MAKEDEPPROG= makedepend -MAKEDEPEND= $(TOP)/util/domd $(TOP) -MD $(MAKEDEPPROG) MAKEFILE= Makefile AR= ar r -# CFLAGS= -mpentiumpro $(INCLUDES) $(CFLAG) -O3 -fexpensive-optimizations -funroll-loops -fforce-addr +AES_ASM_OBJ=aes_core.o aes_cbc.o + CFLAGS= $(INCLUDES) $(CFLAG) +ASFLAGS= $(INCLUDES) $(ASFLAG) +AFLAGS= $(ASFLAGS) GENERAL=Makefile #TEST=aestest.c @@ -26,7 +24,7 @@ APPS= LIB=$(TOP)/libcrypto.a LIBSRC=aes_core.c aes_misc.c aes_ecb.c aes_cbc.c aes_cfb.c aes_ofb.c aes_ctr.c -LIBOBJ=aes_core.o aes_misc.o aes_ecb.o aes_cbc.o aes_cfb.o aes_ofb.o aes_ctr.o +LIBOBJ=aes_misc.o aes_ecb.o aes_cfb.o aes_ofb.o aes_ctr.o $(AES_ASM_OBJ) SRC= $(LIBSRC) @@ -47,6 +45,16 @@ lib: $(LIBOBJ) $(LIBOBJ): $(LIBSRC) +aes-ia64.s: asm/aes-ia64.S + $(CC) $(CFLAGS) -E asm/aes-ia64.S > $@ + +ax86-elf.s: asm/aes-586.pl ../perlasm/x86asm.pl + (cd asm; $(PERL) aes-586.pl elf $(CFLAGS) $(PROCESSOR) > ../$@) +ax86-cof.s: asm/aes-586.pl ../perlasm/x86asm.pl + (cd asm; $(PERL) aes-586.pl coff $(CFLAGS) $(PROCESSOR) > ../$@) +ax86-out.s: asm/aes-586.pl ../perlasm/x86asm.pl + (cd asm; $(PERL) aes-586.pl a.out $(CFLAGS) $(PROCESSOR) > ../$@) + files: $(PERL) $(TOP)/util/files.pl Makefile >> $(TOP)/MINFO @@ -55,10 +63,9 @@ links: @$(PERL) $(TOP)/util/mklink.pl ../../test $(TEST) @$(PERL) $(TOP)/util/mklink.pl ../../apps $(APPS) -install: installs - -installs: - @for i in $(EXHEADER) ; \ +install: + @[ -n "$(INSTALLTOP)" ] # should be set by top Makefile... + @headerlist="$(EXHEADER)"; for i in $$headerlist ; \ do \ (cp $$i $(INSTALL_PREFIX)$(INSTALLTOP)/include/openssl/$$i; \ chmod 644 $(INSTALL_PREFIX)$(INSTALLTOP)/include/openssl/$$i ); \ @@ -73,6 +80,7 @@ lint: lint -DLINT $(INCLUDES) $(SRC)>fluff depend: + @[ -n "$(MAKEDEPEND)" ] # should be set by upper Makefile... $(MAKEDEPEND) -- $(CFLAG) $(INCLUDES) $(DEPFLAG) -- $(PROGS) $(LIBSRC) dclean: @@ -80,17 +88,17 @@ dclean: mv -f Makefile.new $(MAKEFILE) clean: - rm -f *.o *.obj lib tags core .pure .nfs* *.old *.bak fluff + rm -f *.s *.o *.obj lib tags core .pure .nfs* *.old *.bak fluff # DO NOT DELETE THIS LINE -- make depend depends on it. aes_cbc.o: ../../include/openssl/aes.h ../../include/openssl/e_os2.h aes_cbc.o: ../../include/openssl/opensslconf.h aes_cbc.c aes_locl.h -aes_cfb.o: ../../include/openssl/aes.h ../../include/openssl/e_os2.h -aes_cfb.o: ../../include/openssl/opensslconf.h aes_cfb.c aes_locl.h +aes_cfb.o: ../../e_os.h ../../include/openssl/aes.h +aes_cfb.o: ../../include/openssl/e_os2.h ../../include/openssl/opensslconf.h +aes_cfb.o: aes_cfb.c aes_locl.h aes_core.o: ../../include/openssl/aes.h ../../include/openssl/e_os2.h -aes_core.o: ../../include/openssl/fips.h ../../include/openssl/opensslconf.h -aes_core.o: aes_core.c aes_locl.h +aes_core.o: ../../include/openssl/opensslconf.h aes_core.c aes_locl.h aes_ctr.o: ../../include/openssl/aes.h ../../include/openssl/e_os2.h aes_ctr.o: ../../include/openssl/opensslconf.h aes_ctr.c aes_locl.h aes_ecb.o: ../../include/openssl/aes.h ../../include/openssl/e_os2.h diff --git a/crypto/openssl/crypto/aes/aes.h b/crypto/openssl/crypto/aes/aes.h index 8a3ea0b88366..9ffcc9ff2a30 100644 --- a/crypto/openssl/crypto/aes/aes.h +++ b/crypto/openssl/crypto/aes/aes.h @@ -52,7 +52,7 @@ #ifndef HEADER_AES_H #define HEADER_AES_H -#include <openssl/e_os2.h> +#include <openssl/opensslconf.h> #ifdef OPENSSL_NO_AES #error AES is disabled. @@ -66,17 +66,17 @@ #define AES_MAXNR 14 #define AES_BLOCK_SIZE 16 -#if defined(OPENSSL_FIPS) -#define FIPS_AES_SIZE_T int -#endif - #ifdef __cplusplus extern "C" { #endif /* This should be a hidden type, but EVP requires that the size be known */ struct aes_key_st { +#ifdef AES_LONG unsigned long rd_key[4 *(AES_MAXNR + 1)]; +#else + unsigned int rd_key[4 *(AES_MAXNR + 1)]; +#endif int rounds; }; typedef struct aes_key_st AES_KEY; diff --git a/crypto/openssl/crypto/aes/aes_cbc.c b/crypto/openssl/crypto/aes/aes_cbc.c index 1222a21002c1..d2ba6bcdb465 100644 --- a/crypto/openssl/crypto/aes/aes_cbc.c +++ b/crypto/openssl/crypto/aes/aes_cbc.c @@ -66,6 +66,7 @@ void AES_cbc_encrypt(const unsigned char *in, unsigned char *out, unsigned long n; unsigned long len = length; unsigned char tmp[AES_BLOCK_SIZE]; + const unsigned char *iv = ivec; assert(in && out && key && ivec); assert((AES_ENCRYPT == enc)||(AES_DECRYPT == enc)); @@ -73,22 +74,39 @@ void AES_cbc_encrypt(const unsigned char *in, unsigned char *out, if (AES_ENCRYPT == enc) { while (len >= AES_BLOCK_SIZE) { for(n=0; n < AES_BLOCK_SIZE; ++n) - tmp[n] = in[n] ^ ivec[n]; - AES_encrypt(tmp, out, key); - memcpy(ivec, out, AES_BLOCK_SIZE); + out[n] = in[n] ^ iv[n]; + AES_encrypt(out, out, key); + iv = out; len -= AES_BLOCK_SIZE; in += AES_BLOCK_SIZE; out += AES_BLOCK_SIZE; } if (len) { for(n=0; n < len; ++n) - tmp[n] = in[n] ^ ivec[n]; + out[n] = in[n] ^ iv[n]; for(n=len; n < AES_BLOCK_SIZE; ++n) - tmp[n] = ivec[n]; - AES_encrypt(tmp, tmp, key); - memcpy(out, tmp, AES_BLOCK_SIZE); - memcpy(ivec, tmp, AES_BLOCK_SIZE); - } + out[n] = iv[n]; + AES_encrypt(out, out, key); + iv = out; + } + memcpy(ivec,iv,AES_BLOCK_SIZE); + } else if (in != out) { + while (len >= AES_BLOCK_SIZE) { + AES_decrypt(in, out, key); + for(n=0; n < AES_BLOCK_SIZE; ++n) + out[n] ^= iv[n]; + iv = in; + len -= AES_BLOCK_SIZE; + in += AES_BLOCK_SIZE; + out += AES_BLOCK_SIZE; + } + if (len) { + AES_decrypt(in,tmp,key); + for(n=0; n < len; ++n) + out[n] = tmp[n] ^ iv[n]; + iv = in; + } + memcpy(ivec,iv,AES_BLOCK_SIZE); } else { while (len >= AES_BLOCK_SIZE) { memcpy(tmp, in, AES_BLOCK_SIZE); @@ -102,10 +120,12 @@ void AES_cbc_encrypt(const unsigned char *in, unsigned char *out, } if (len) { memcpy(tmp, in, AES_BLOCK_SIZE); - AES_decrypt(tmp, tmp, key); + AES_decrypt(tmp, out, key); for(n=0; n < len; ++n) - out[n] = tmp[n] ^ ivec[n]; + out[n] ^= ivec[n]; + for(n=len; n < AES_BLOCK_SIZE; ++n) + out[n] = tmp[n]; memcpy(ivec, tmp, AES_BLOCK_SIZE); - } + } } } diff --git a/crypto/openssl/crypto/aes/aes_cfb.c b/crypto/openssl/crypto/aes/aes_cfb.c index 2e0c41ec2b6b..49f0411010c3 100644 --- a/crypto/openssl/crypto/aes/aes_cfb.c +++ b/crypto/openssl/crypto/aes/aes_cfb.c @@ -114,6 +114,7 @@ #include <openssl/aes.h> #include "aes_locl.h" +#include "e_os.h" /* The input and output encrypted as though 128bit cfb mode is being * used. The extra state information to record how much of the @@ -157,61 +158,35 @@ void AES_cfb128_encrypt(const unsigned char *in, unsigned char *out, /* This expects a single block of size nbits for both in and out. Note that it corrupts any extra bits in the last byte of out */ -/* Untested, once it is working, it will be optimised */ void AES_cfbr_encrypt_block(const unsigned char *in,unsigned char *out, const int nbits,const AES_KEY *key, unsigned char *ivec,const int enc) { - int n; + int n,rem,num; unsigned char ovec[AES_BLOCK_SIZE*2]; - assert(in && out && key && ivec); - if(enc) - { - /* construct the new IV */ - AES_encrypt(ivec,ovec,key); - /* encrypt the input */ - for(n=0 ; n < (nbits+7)/8 ; ++n) - out[n]=in[n]^ovec[n]; - /* fill in the first half of the new IV with the current IV */ - memcpy(ovec,ivec,AES_BLOCK_SIZE); - /* and put the ciphertext in the second half */ - memcpy(ovec+AES_BLOCK_SIZE,out,(nbits+7)/8); - /* shift ovec left most of the bits... */ - memmove(ovec,ovec+nbits/8,AES_BLOCK_SIZE+(nbits%8 ? 1 : 0)); - /* now the remaining bits */ - if(nbits%8 != 0) - for(n=0 ; n < AES_BLOCK_SIZE ; ++n) - { - ovec[n]<<=nbits%8; - ovec[n]|=ovec[n+1]>>(8-nbits%8); - } - /* finally, move it back into place */ - memcpy(ivec,ovec,AES_BLOCK_SIZE); - } - else - { - /* construct the new IV in the first half of ovec */ - AES_encrypt(ivec,ovec,key); - /* decrypt the input */ - for(n=0 ; n < (nbits+7)/8 ; ++n) - out[n]=in[n]^ovec[n]; + if (nbits<=0 || nbits>128) return; + /* fill in the first half of the new IV with the current IV */ memcpy(ovec,ivec,AES_BLOCK_SIZE); - /* append the ciphertext */ - memcpy(ovec+AES_BLOCK_SIZE,in,(nbits+7)/8); - /* shift ovec left most of the bits... */ - memmove(ovec,ovec+nbits/8,AES_BLOCK_SIZE+(nbits%8 ? 1 : 0)); - /* now the remaining bits */ - if(nbits%8 != 0) + /* construct the new IV */ + AES_encrypt(ivec,ivec,key); + num = (nbits+7)/8; + if (enc) /* encrypt the input */ + for(n=0 ; n < num ; ++n) + out[n] = (ovec[AES_BLOCK_SIZE+n] = in[n] ^ ivec[n]); + else /* decrypt the input */ + for(n=0 ; n < num ; ++n) + out[n] = (ovec[AES_BLOCK_SIZE+n] = in[n]) ^ ivec[n]; + /* shift ovec left... */ + rem = nbits%8; + num = nbits/8; + if(rem==0) + memcpy(ivec,ovec+num,AES_BLOCK_SIZE); + else for(n=0 ; n < AES_BLOCK_SIZE ; ++n) - { - ovec[n]<<=nbits%8; - ovec[n]|=ovec[n+1]>>(8-nbits%8); - } - /* finally, move it back into place */ - memcpy(ivec,ovec,AES_BLOCK_SIZE); - } + ivec[n] = ovec[n+num]<<rem | ovec[n+num+1]>>(8-rem); + /* it is not necessary to cleanse ovec, since the IV is not secret */ } diff --git a/crypto/openssl/crypto/aes/aes_core.c b/crypto/openssl/crypto/aes/aes_core.c index ed566a81233f..410ae2e8e88a 100644 --- a/crypto/openssl/crypto/aes/aes_core.c +++ b/crypto/openssl/crypto/aes/aes_core.c @@ -37,11 +37,8 @@ #include <stdlib.h> #include <openssl/aes.h> -#include <openssl/fips.h> #include "aes_locl.h" -#ifndef OPENSSL_FIPS - /* Te0[x] = S [x].[02, 01, 01, 03]; Te1[x] = S [x].[03, 02, 01, 01]; @@ -56,6 +53,13 @@ Td3[x] = Si[x].[09, 0d, 0b, 0e]; Td4[x] = Si[x].[01, 01, 01, 01]; */ +#ifdef AES_ASM +extern const u32 AES_Te[5][256]; +#define Te0 AES_Te[0] +#define Te1 AES_Te[1] +#define Te2 AES_Te[2] +#define Te3 AES_Te[3] +#else static const u32 Te0[256] = { 0xc66363a5U, 0xf87c7c84U, 0xee777799U, 0xf67b7b8dU, 0xfff2f20dU, 0xd66b6bbdU, 0xde6f6fb1U, 0x91c5c554U, @@ -255,7 +259,6 @@ static const u32 Te2[256] = { 0xb0cb7bb0U, 0x54fca854U, 0xbbd66dbbU, 0x163a2c16U, }; static const u32 Te3[256] = { - 0x6363a5c6U, 0x7c7c84f8U, 0x777799eeU, 0x7b7b8df6U, 0xf2f20dffU, 0x6b6bbdd6U, 0x6f6fb1deU, 0xc5c55491U, 0x30305060U, 0x01010302U, 0x6767a9ceU, 0x2b2b7d56U, @@ -321,6 +324,7 @@ static const u32 Te3[256] = { 0x4141c382U, 0x9999b029U, 0x2d2d775aU, 0x0f0f111eU, 0xb0b0cb7bU, 0x5454fca8U, 0xbbbbd66dU, 0x16163a2cU, }; +#endif static const u32 Te4[256] = { 0x63636363U, 0x7c7c7c7cU, 0x77777777U, 0x7b7b7b7bU, 0xf2f2f2f2U, 0x6b6b6b6bU, 0x6f6f6f6fU, 0xc5c5c5c5U, @@ -387,6 +391,14 @@ static const u32 Te4[256] = { 0x41414141U, 0x99999999U, 0x2d2d2d2dU, 0x0f0f0f0fU, 0xb0b0b0b0U, 0x54545454U, 0xbbbbbbbbU, 0x16161616U, }; + +#ifdef AES_ASM +extern const u32 AES_Td[5][256]; +#define Td0 AES_Td[0] +#define Td1 AES_Td[1] +#define Td2 AES_Td[2] +#define Td3 AES_Td[3] +#else static const u32 Td0[256] = { 0x51f4a750U, 0x7e416553U, 0x1a17a4c3U, 0x3a275e96U, 0x3bab6bcbU, 0x1f9d45f1U, 0xacfa58abU, 0x4be30393U, @@ -540,7 +552,6 @@ static const u32 Td2[256] = { 0xf4cd65daU, 0xbed50605U, 0x621fd134U, 0xfe8ac4a6U, 0x539d342eU, 0x55a0a2f3U, 0xe132058aU, 0xeb75a4f6U, 0xec390b83U, 0xefaa4060U, 0x9f065e71U, 0x1051bd6eU, - 0x8af93e21U, 0x063d96ddU, 0x05aedd3eU, 0xbd464de6U, 0x8db59154U, 0x5d0571c4U, 0xd46f0406U, 0x15ff6050U, 0xfb241998U, 0xe997d6bdU, 0x43cc8940U, 0x9e7767d9U, @@ -652,6 +663,7 @@ static const u32 Td3[256] = { 0xa8017139U, 0x0cb3de08U, 0xb4e49cd8U, 0x56c19064U, 0xcb84617bU, 0x32b670d5U, 0x6c5c7448U, 0xb85742d0U, }; +#endif static const u32 Td4[256] = { 0x52525252U, 0x09090909U, 0x6a6a6a6aU, 0xd5d5d5d5U, 0x30303030U, 0x36363636U, 0xa5a5a5a5U, 0x38383838U, @@ -876,6 +888,7 @@ int AES_set_decrypt_key(const unsigned char *userKey, const int bits, return 0; } +#ifndef AES_ASM /* * Encrypt a single block * in and out can overlap @@ -1258,4 +1271,4 @@ void AES_decrypt(const unsigned char *in, unsigned char *out, PUTU32(out + 12, s3); } -#endif /* ndef OPENSSL_FIPS */ +#endif /* AES_ASM */ diff --git a/crypto/openssl/crypto/aes/aes_locl.h b/crypto/openssl/crypto/aes/aes_locl.h index 4184729e344d..054b442d416f 100644 --- a/crypto/openssl/crypto/aes/aes_locl.h +++ b/crypto/openssl/crypto/aes/aes_locl.h @@ -62,7 +62,7 @@ #include <stdlib.h> #include <string.h> -#if defined(_MSC_VER) && !defined(_M_IA64) && !defined(OPENSSL_SYS_WINCE) +#if defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_AMD64) || defined(_M_X64)) # define SWAP(x) (_lrotl(x, 8) & 0x00ff00ff | _lrotr(x, 8) & 0xff00ff00) # define GETU32(p) SWAP(*((u32 *)(p))) # define PUTU32(ct, st) { *((u32 *)(ct)) = SWAP((st)); } @@ -71,7 +71,11 @@ # define PUTU32(ct, st) { (ct)[0] = (u8)((st) >> 24); (ct)[1] = (u8)((st) >> 16); (ct)[2] = (u8)((st) >> 8); (ct)[3] = (u8)(st); } #endif +#ifdef AES_LONG typedef unsigned long u32; +#else +typedef unsigned int u32; +#endif typedef unsigned short u16; typedef unsigned char u8; diff --git a/crypto/openssl/crypto/aes/asm/aes-586.pl b/crypto/openssl/crypto/aes/asm/aes-586.pl new file mode 100755 index 000000000000..c1206238819a --- /dev/null +++ b/crypto/openssl/crypto/aes/asm/aes-586.pl @@ -0,0 +1,1531 @@ +#!/usr/bin/env perl +# +# ==================================================================== +# 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. +# ==================================================================== +# +# Version 3.4. +# +# You might fail to appreciate this module performance from the first +# try. If compared to "vanilla" linux-ia32-icc target, i.e. considered +# to be *the* best Intel C compiler without -KPIC, performance appears +# to be virtually identical... But try to re-configure with shared +# library support... Aha! Intel compiler "suddenly" lags behind by 30% +# [on P4, more on others]:-) And if compared to position-independent +# code generated by GNU C, this code performs *more* than *twice* as +# fast! Yes, all this buzz about PIC means that unlike other hand- +# coded implementations, this one was explicitly designed to be safe +# to use even in shared library context... This also means that this +# code isn't necessarily absolutely fastest "ever," because in order +# to achieve position independence an extra register has to be +# off-loaded to stack, which affects the benchmark result. +# +# Special note about instruction choice. Do you recall RC4_INT code +# performing poorly on P4? It might be the time to figure out why. +# RC4_INT code implies effective address calculations in base+offset*4 +# form. Trouble is that it seems that offset scaling turned to be +# critical path... At least eliminating scaling resulted in 2.8x RC4 +# performance improvement [as you might recall]. As AES code is hungry +# for scaling too, I [try to] avoid the latter by favoring off-by-2 +# shifts and masking the result with 0xFF<<2 instead of "boring" 0xFF. +# +# As was shown by Dean Gaudet <dean@arctic.org>, the above note turned +# void. Performance improvement with off-by-2 shifts was observed on +# intermediate implementation, which was spilling yet another register +# to stack... Final offset*4 code below runs just a tad faster on P4, +# but exhibits up to 10% improvement on other cores. +# +# Second version is "monolithic" replacement for aes_core.c, which in +# addition to AES_[de|en]crypt implements AES_set_[de|en]cryption_key. +# This made it possible to implement little-endian variant of the +# algorithm without modifying the base C code. Motivating factor for +# the undertaken effort was that it appeared that in tight IA-32 +# register window little-endian flavor could achieve slightly higher +# Instruction Level Parallelism, and it indeed resulted in up to 15% +# better performance on most recent µ-archs... +# +# Third version adds AES_cbc_encrypt implementation, which resulted in +# up to 40% performance imrovement of CBC benchmark results. 40% was +# observed on P4 core, where "overall" imrovement coefficient, i.e. if +# compared to PIC generated by GCC and in CBC mode, was observed to be +# as large as 4x:-) CBC performance is virtually identical to ECB now +# and on some platforms even better, e.g. 17.6 "small" cycles/byte on +# Opteron, because certain function prologues and epilogues are +# effectively taken out of the loop... +# +# Version 3.2 implements compressed tables and prefetch of these tables +# in CBC[!] mode. Former means that 3/4 of table references are now +# misaligned, which unfortunately has negative impact on elder IA-32 +# implementations, Pentium suffered 30% penalty, PIII - 10%. +# +# Version 3.3 avoids L1 cache aliasing between stack frame and +# S-boxes, and 3.4 - L1 cache aliasing even between key schedule. The +# latter is achieved by copying the key schedule to controlled place in +# stack. This unfortunately has rather strong impact on small block CBC +# performance, ~2x deterioration on 16-byte block if compared to 3.3. +# +# Current ECB performance numbers for 128-bit key in CPU cycles per +# processed byte [measure commonly used by AES benchmarkers] are: +# +# small footprint fully unrolled +# P4 24 22 +# AMD K8 20 19 +# PIII 25 23 +# Pentium 81 78 + +push(@INC,"perlasm","../../perlasm"); +require "x86asm.pl"; + +&asm_init($ARGV[0],"aes-586.pl",$ARGV[$#ARGV] eq "386"); + +$s0="eax"; +$s1="ebx"; +$s2="ecx"; +$s3="edx"; +$key="edi"; +$acc="esi"; + +$compromise=0; # $compromise=128 abstains from copying key + # schedule to stack when encrypting inputs + # shorter than 128 bytes at the cost of + # risksing aliasing with S-boxes. In return + # you get way better, up to +70%, small block + # performance. +$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 +# "column," unlike "horizontal" one, which groups shifts on different +# $s[i] to form a "row"] is work in progress. It was observed to run +# few percents faster on Intel cores, but not AMD. On AMD K8 core it's +# whole 12% slower:-( So we face a trade-off... Shall it be resolved +# some day? Till then the code is considered experimental and by +# default remains dormant... + +sub encvert() +{ my ($te,@s) = @_; + my $v0 = $acc, $v1 = $key; + + &mov ($v0,$s[3]); # copy s3 + &mov (&DWP(4,"esp"),$s[2]); # save s2 + &mov ($v1,$s[0]); # copy s0 + &mov (&DWP(8,"esp"),$s[1]); # save s1 + + &movz ($s[2],&HB($s[0])); + &and ($s[0],0xFF); + &mov ($s[0],&DWP(0,$te,$s[0],8)); # s0>>0 + &shr ($v1,16); + &mov ($s[3],&DWP(3,$te,$s[2],8)); # s0>>8 + &movz ($s[1],&HB($v1)); + &and ($v1,0xFF); + &mov ($s[2],&DWP(2,$te,$v1,8)); # s0>>16 + &mov ($v1,$v0); + &mov ($s[1],&DWP(1,$te,$s[1],8)); # s0>>24 + + &and ($v0,0xFF); + &xor ($s[3],&DWP(0,$te,$v0,8)); # s3>>0 + &movz ($v0,&HB($v1)); + &shr ($v1,16); + &xor ($s[2],&DWP(3,$te,$v0,8)); # s3>>8 + &movz ($v0,&HB($v1)); + &and ($v1,0xFF); + &xor ($s[1],&DWP(2,$te,$v1,8)); # s3>>16 + &mov ($v1,&DWP(4,"esp")); # restore s2 + &xor ($s[0],&DWP(1,$te,$v0,8)); # s3>>24 + + &mov ($v0,$v1); + &and ($v1,0xFF); + &xor ($s[2],&DWP(0,$te,$v1,8)); # s2>>0 + &movz ($v1,&HB($v0)); + &shr ($v0,16); + &xor ($s[1],&DWP(3,$te,$v1,8)); # s2>>8 + &movz ($v1,&HB($v0)); + &and ($v0,0xFF); + &xor ($s[0],&DWP(2,$te,$v0,8)); # s2>>16 + &mov ($v0,&DWP(8,"esp")); # restore s1 + &xor ($s[3],&DWP(1,$te,$v1,8)); # s2>>24 + + &mov ($v1,$v0); + &and ($v0,0xFF); + &xor ($s[1],&DWP(0,$te,$v0,8)); # s1>>0 + &movz ($v0,&HB($v1)); + &shr ($v1,16); + &xor ($s[0],&DWP(3,$te,$v0,8)); # s1>>8 + &movz ($v0,&HB($v1)); + &and ($v1,0xFF); + &xor ($s[3],&DWP(2,$te,$v1,8)); # s1>>16 + &mov ($key,&DWP(12,"esp")); # reincarnate v1 as key + &xor ($s[2],&DWP(1,$te,$v0,8)); # s1>>24 +} + +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 ($key,&DWP(12,"esp")); }##%edx + else { &mov ($out,$s[0]); + &and ($out,0xFF); } + if ($i==1) { &shr ($s[0],16); }#%ebx[1] + if ($i==2) { &shr ($s[0],24); }#%ecx[2] + &mov ($out,&DWP(0,$te,$out,8)); + + if ($i==3) { $tmp=$s[1]; }##%eax + &movz ($tmp,&HB($s[1])); + &xor ($out,&DWP(3,$te,$tmp,8)); + + if ($i==3) { $tmp=$s[2]; &mov ($s[1],&DWP(4,"esp")); }##%ebx + else { &mov ($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 ($s[2],&DWP(8,"esp")); }##%ecx + elsif($i==2){ &movz ($tmp,&HB($s[3])); }#%ebx[2] + else { &mov ($tmp,$s[3]); + &shr ($tmp,24) } + &xor ($out,&DWP(1,$te,$tmp,8)); + if ($i<2) { &mov (&DWP(4+4*$i,"esp"),$out); } + if ($i==3) { &mov ($s[3],$acc); } + &comment(); +} + +sub enclast() +{ my ($i,$te,@s)=@_; + my $tmp = $key; + my $out = $i==3?$s[0]:$acc; + + if ($i==3) { &mov ($key,&DWP(12,"esp")); }##%edx + else { &mov ($out,$s[0]); } + &and ($out,0xFF); + if ($i==1) { &shr ($s[0],16); }#%ebx[1] + if ($i==2) { &shr ($s[0],24); }#%ecx[2] + &mov ($out,&DWP(2,$te,$out,8)); + &and ($out,0x000000ff); + + if ($i==3) { $tmp=$s[1]; }##%eax + &movz ($tmp,&HB($s[1])); + &mov ($tmp,&DWP(0,$te,$tmp,8)); + &and ($tmp,0x0000ff00); + &xor ($out,$tmp); + + if ($i==3) { $tmp=$s[2]; &mov ($s[1],&DWP(4,"esp")); }##%ebx + else { mov ($tmp,$s[2]); + &shr ($tmp,16); } + if ($i==2) { &and ($s[1],0xFF); }#%edx[2] + &and ($tmp,0xFF); + &mov ($tmp,&DWP(0,$te,$tmp,8)); + &and ($tmp,0x00ff0000); + &xor ($out,$tmp); + + if ($i==3) { $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }##%ecx + elsif($i==2){ &movz ($tmp,&HB($s[3])); }#%ebx[2] + else { &mov ($tmp,$s[3]); + &shr ($tmp,24); } + &mov ($tmp,&DWP(2,$te,$tmp,8)); + &and ($tmp,0xff000000); + &xor ($out,$tmp); + if ($i<2) { &mov (&DWP(4+4*$i,"esp"),$out); } + 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 ($vertical_spin) { + # I need high parts of volatile registers to be accessible... + &exch ($s1="edi",$key="ebx"); + &mov ($s2="esi",$acc="ecx"); + } + + # note that caller is expected to allocate stack frame for me! + &mov (&DWP(12,"esp"),$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 + + 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"); + if ($vertical_spin) { + &encvert("ebp",$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,$s2); + } + &add ($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); + &jb (&label("loop")); + } + else { + &cmp ($acc,10); + &jle (&label("10rounds")); + &cmp ($acc,12); + &jle (&label("12rounds")); + + &set_label("14rounds"); + for ($i=1;$i<3;$i++) { + if ($vertical_spin) { + &encvert("ebp",$s0,$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); + } + &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"); + for ($i=1;$i<3;$i++) { + if ($vertical_spin) { + &encvert("ebp",$s0,$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); + } + &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"); + for ($i=1;$i<10;$i++) { + if ($vertical_spin) { + &encvert("ebp",$s0,$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); + } + &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)); + } + } + + if ($vertical_spin) { + # "reincarnate" some registers for "horizontal" spin... + &mov ($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,$s2); + + &add ($key,$small_footprint?16:160); + &xor ($s0,&DWP(0,$key)); + &xor ($s1,&DWP(4,$key)); + &xor ($s2,&DWP(8,$key)); + &xor ($s3,&DWP(12,$key)); + + &ret (); + +&set_label("AES_Te",64); # Yes! I keep it in the code segment! + &_data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6); + &_data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591); + &_data_word(0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56); + &_data_word(0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec); + &_data_word(0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa); + &_data_word(0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb); + &_data_word(0xecadad41, 0x67d4d4b3, 0xfda2a25f, 0xeaafaf45); + &_data_word(0xbf9c9c23, 0xf7a4a453, 0x967272e4, 0x5bc0c09b); + &_data_word(0xc2b7b775, 0x1cfdfde1, 0xae93933d, 0x6a26264c); + &_data_word(0x5a36366c, 0x413f3f7e, 0x02f7f7f5, 0x4fcccc83); + &_data_word(0x5c343468, 0xf4a5a551, 0x34e5e5d1, 0x08f1f1f9); + &_data_word(0x937171e2, 0x73d8d8ab, 0x53313162, 0x3f15152a); + &_data_word(0x0c040408, 0x52c7c795, 0x65232346, 0x5ec3c39d); + &_data_word(0x28181830, 0xa1969637, 0x0f05050a, 0xb59a9a2f); + &_data_word(0x0907070e, 0x36121224, 0x9b80801b, 0x3de2e2df); + &_data_word(0x26ebebcd, 0x6927274e, 0xcdb2b27f, 0x9f7575ea); + &_data_word(0x1b090912, 0x9e83831d, 0x742c2c58, 0x2e1a1a34); + &_data_word(0x2d1b1b36, 0xb26e6edc, 0xee5a5ab4, 0xfba0a05b); + &_data_word(0xf65252a4, 0x4d3b3b76, 0x61d6d6b7, 0xceb3b37d); + &_data_word(0x7b292952, 0x3ee3e3dd, 0x712f2f5e, 0x97848413); + &_data_word(0xf55353a6, 0x68d1d1b9, 0x00000000, 0x2cededc1); + &_data_word(0x60202040, 0x1ffcfce3, 0xc8b1b179, 0xed5b5bb6); + &_data_word(0xbe6a6ad4, 0x46cbcb8d, 0xd9bebe67, 0x4b393972); + &_data_word(0xde4a4a94, 0xd44c4c98, 0xe85858b0, 0x4acfcf85); + &_data_word(0x6bd0d0bb, 0x2aefefc5, 0xe5aaaa4f, 0x16fbfbed); + &_data_word(0xc5434386, 0xd74d4d9a, 0x55333366, 0x94858511); + &_data_word(0xcf45458a, 0x10f9f9e9, 0x06020204, 0x817f7ffe); + &_data_word(0xf05050a0, 0x443c3c78, 0xba9f9f25, 0xe3a8a84b); + &_data_word(0xf35151a2, 0xfea3a35d, 0xc0404080, 0x8a8f8f05); + &_data_word(0xad92923f, 0xbc9d9d21, 0x48383870, 0x04f5f5f1); + &_data_word(0xdfbcbc63, 0xc1b6b677, 0x75dadaaf, 0x63212142); + &_data_word(0x30101020, 0x1affffe5, 0x0ef3f3fd, 0x6dd2d2bf); + &_data_word(0x4ccdcd81, 0x140c0c18, 0x35131326, 0x2fececc3); + &_data_word(0xe15f5fbe, 0xa2979735, 0xcc444488, 0x3917172e); + &_data_word(0x57c4c493, 0xf2a7a755, 0x827e7efc, 0x473d3d7a); + &_data_word(0xac6464c8, 0xe75d5dba, 0x2b191932, 0x957373e6); + &_data_word(0xa06060c0, 0x98818119, 0xd14f4f9e, 0x7fdcdca3); + &_data_word(0x66222244, 0x7e2a2a54, 0xab90903b, 0x8388880b); + &_data_word(0xca46468c, 0x29eeeec7, 0xd3b8b86b, 0x3c141428); + &_data_word(0x79dedea7, 0xe25e5ebc, 0x1d0b0b16, 0x76dbdbad); + &_data_word(0x3be0e0db, 0x56323264, 0x4e3a3a74, 0x1e0a0a14); + &_data_word(0xdb494992, 0x0a06060c, 0x6c242448, 0xe45c5cb8); + &_data_word(0x5dc2c29f, 0x6ed3d3bd, 0xefacac43, 0xa66262c4); + &_data_word(0xa8919139, 0xa4959531, 0x37e4e4d3, 0x8b7979f2); + &_data_word(0x32e7e7d5, 0x43c8c88b, 0x5937376e, 0xb76d6dda); + &_data_word(0x8c8d8d01, 0x64d5d5b1, 0xd24e4e9c, 0xe0a9a949); + &_data_word(0xb46c6cd8, 0xfa5656ac, 0x07f4f4f3, 0x25eaeacf); + &_data_word(0xaf6565ca, 0x8e7a7af4, 0xe9aeae47, 0x18080810); + &_data_word(0xd5baba6f, 0x887878f0, 0x6f25254a, 0x722e2e5c); + &_data_word(0x241c1c38, 0xf1a6a657, 0xc7b4b473, 0x51c6c697); + &_data_word(0x23e8e8cb, 0x7cdddda1, 0x9c7474e8, 0x211f1f3e); + &_data_word(0xdd4b4b96, 0xdcbdbd61, 0x868b8b0d, 0x858a8a0f); + &_data_word(0x907070e0, 0x423e3e7c, 0xc4b5b571, 0xaa6666cc); + &_data_word(0xd8484890, 0x05030306, 0x01f6f6f7, 0x120e0e1c); + &_data_word(0xa36161c2, 0x5f35356a, 0xf95757ae, 0xd0b9b969); + &_data_word(0x91868617, 0x58c1c199, 0x271d1d3a, 0xb99e9e27); + &_data_word(0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122); + &_data_word(0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433); + &_data_word(0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9); + &_data_word(0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5); + &_data_word(0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a); + &_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0); + &_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e); + &_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c); +#rcon: + &data_word(0x00000001, 0x00000002, 0x00000004, 0x00000008); + &data_word(0x00000010, 0x00000020, 0x00000040, 0x00000080); + &data_word(0x0000001b, 0x00000036, 0, 0, 0, 0, 0, 0); +&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"),$s0); + + &call (&label("pic_point")); # make it PIC! + &set_label("pic_point"); + &blindpop("ebp"); + &lea ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp")); + + &mov ($s0,&DWP(0,$acc)); # load input data + &mov ($s1,&DWP(4,$acc)); + &mov ($s2,&DWP(8,$acc)); + &mov ($s3,&DWP(12,$acc)); + + &call ("_x86_AES_encrypt"); + + &mov ("esp",&DWP(16,"esp")); + + &mov ($acc,&wparam(1)); # load out + &mov (&DWP(0,$acc),$s0); # write output data + &mov (&DWP(4,$acc),$s1); + &mov (&DWP(8,$acc),$s2); + &mov (&DWP(12,$acc),$s3); +&function_end("AES_encrypt"); + +#------------------------------------------------------------------# + +sub decstep() +{ my ($i,$td,@s) = @_; + my $tmp = $key; + my $out = $i==3?$s[0]:$acc; + + # no instructions are reordered, as performance appears + # optimal... or rather that all attempts to reorder didn't + # result in better performance [which by the way is not a + # bit lower than ecryption]. + if($i==3) { &mov ($key,&DWP(12,"esp")); } + else { &mov ($out,$s[0]); } + &and ($out,0xFF); + &mov ($out,&DWP(0,$td,$out,8)); + + if ($i==3) { $tmp=$s[1]; } + &movz ($tmp,&HB($s[1])); + &xor ($out,&DWP(3,$td,$tmp,8)); + + if ($i==3) { $tmp=$s[2]; &mov ($s[1],$acc); } + else { &mov ($tmp,$s[2]); } + &shr ($tmp,16); + &and ($tmp,0xFF); + &xor ($out,&DWP(2,$td,$tmp,8)); + + if ($i==3) { $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); } + else { &mov ($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")); } + &comment(); +} + +sub declast() +{ my ($i,$td,@s)=@_; + my $tmp = $key; + my $out = $i==3?$s[0]:$acc; + + if($i==3) { &mov ($key,&DWP(12,"esp")); } + else { &mov ($out,$s[0]); } + &and ($out,0xFF); + &mov ($out,&DWP(2048,$td,$out,4)); + &and ($out,0x000000ff); + + if ($i==3) { $tmp=$s[1]; } + &movz ($tmp,&HB($s[1])); + &mov ($tmp,&DWP(2048,$td,$tmp,4)); + &and ($tmp,0x0000ff00); + &xor ($out,$tmp); + + if ($i==3) { $tmp=$s[2]; &mov ($s[1],$acc); } + else { mov ($tmp,$s[2]); } + &shr ($tmp,16); + &and ($tmp,0xFF); + &mov ($tmp,&DWP(2048,$td,$tmp,4)); + &and ($tmp,0x00ff0000); + &xor ($out,$tmp); + + if ($i==3) { $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); } + else { &mov ($tmp,$s[3]); } + &shr ($tmp,24); + &mov ($tmp,&DWP(2048,$td,$tmp,4)); + &and ($tmp,0xff000000); + &xor ($out,$tmp); + if ($i<2) { &mov (&DWP(4+4*$i,"esp"),$out); } + if ($i==3) { &mov ($s[3],&DWP(4,"esp")); } +} + +&public_label("AES_Td"); +&function_begin_B("_x86_AES_decrypt"); + # note that caller is expected to allocate stack frame for me! + &mov (&DWP(12,"esp"),$key); # save key + + &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 + + 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); + &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"),$key); + &jb (&label("loop")); + } + else { + &cmp ($acc,10); + &jle (&label("10rounds")); + &cmp ($acc,12); + &jle (&label("12rounds")); + + &set_label("14rounds"); + for ($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); + &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"); + 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,$s0); + &xor ($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"),$key); # advance rd_key + &set_label("10rounds"); + for ($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); + &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)); + } + } + + &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); + + &add ($key,$small_footprint?16:160); + &xor ($s0,&DWP(0,$key)); + &xor ($s1,&DWP(4,$key)); + &xor ($s2,&DWP(8,$key)); + &xor ($s3,&DWP(12,$key)); + + &ret (); + +&set_label("AES_Td",64); # Yes! I keep it in the code segment! + &_data_word(0x50a7f451, 0x5365417e, 0xc3a4171a, 0x965e273a); + &_data_word(0xcb6bab3b, 0xf1459d1f, 0xab58faac, 0x9303e34b); + &_data_word(0x55fa3020, 0xf66d76ad, 0x9176cc88, 0x254c02f5); + &_data_word(0xfcd7e54f, 0xd7cb2ac5, 0x80443526, 0x8fa362b5); + &_data_word(0x495ab1de, 0x671bba25, 0x980eea45, 0xe1c0fe5d); + &_data_word(0x02752fc3, 0x12f04c81, 0xa397468d, 0xc6f9d36b); + &_data_word(0xe75f8f03, 0x959c9215, 0xeb7a6dbf, 0xda595295); + &_data_word(0x2d83bed4, 0xd3217458, 0x2969e049, 0x44c8c98e); + &_data_word(0x6a89c275, 0x78798ef4, 0x6b3e5899, 0xdd71b927); + &_data_word(0xb64fe1be, 0x17ad88f0, 0x66ac20c9, 0xb43ace7d); + &_data_word(0x184adf63, 0x82311ae5, 0x60335197, 0x457f5362); + &_data_word(0xe07764b1, 0x84ae6bbb, 0x1ca081fe, 0x942b08f9); + &_data_word(0x58684870, 0x19fd458f, 0x876cde94, 0xb7f87b52); + &_data_word(0x23d373ab, 0xe2024b72, 0x578f1fe3, 0x2aab5566); + &_data_word(0x0728ebb2, 0x03c2b52f, 0x9a7bc586, 0xa50837d3); + &_data_word(0xf2872830, 0xb2a5bf23, 0xba6a0302, 0x5c8216ed); + &_data_word(0x2b1ccf8a, 0x92b479a7, 0xf0f207f3, 0xa1e2694e); + &_data_word(0xcdf4da65, 0xd5be0506, 0x1f6234d1, 0x8afea6c4); + &_data_word(0x9d532e34, 0xa055f3a2, 0x32e18a05, 0x75ebf6a4); + &_data_word(0x39ec830b, 0xaaef6040, 0x069f715e, 0x51106ebd); + &_data_word(0xf98a213e, 0x3d06dd96, 0xae053edd, 0x46bde64d); + &_data_word(0xb58d5491, 0x055dc471, 0x6fd40604, 0xff155060); + &_data_word(0x24fb9819, 0x97e9bdd6, 0xcc434089, 0x779ed967); + &_data_word(0xbd42e8b0, 0x888b8907, 0x385b19e7, 0xdbeec879); + &_data_word(0x470a7ca1, 0xe90f427c, 0xc91e84f8, 0x00000000); + &_data_word(0x83868009, 0x48ed2b32, 0xac70111e, 0x4e725a6c); + &_data_word(0xfbff0efd, 0x5638850f, 0x1ed5ae3d, 0x27392d36); + &_data_word(0x64d90f0a, 0x21a65c68, 0xd1545b9b, 0x3a2e3624); + &_data_word(0xb1670a0c, 0x0fe75793, 0xd296eeb4, 0x9e919b1b); + &_data_word(0x4fc5c080, 0xa220dc61, 0x694b775a, 0x161a121c); + &_data_word(0x0aba93e2, 0xe52aa0c0, 0x43e0223c, 0x1d171b12); + &_data_word(0x0b0d090e, 0xadc78bf2, 0xb9a8b62d, 0xc8a91e14); + &_data_word(0x8519f157, 0x4c0775af, 0xbbdd99ee, 0xfd607fa3); + &_data_word(0x9f2601f7, 0xbcf5725c, 0xc53b6644, 0x347efb5b); + &_data_word(0x7629438b, 0xdcc623cb, 0x68fcedb6, 0x63f1e4b8); + &_data_word(0xcadc31d7, 0x10856342, 0x40229713, 0x2011c684); + &_data_word(0x7d244a85, 0xf83dbbd2, 0x1132f9ae, 0x6da129c7); + &_data_word(0x4b2f9e1d, 0xf330b2dc, 0xec52860d, 0xd0e3c177); + &_data_word(0x6c16b32b, 0x99b970a9, 0xfa489411, 0x2264e947); + &_data_word(0xc48cfca8, 0x1a3ff0a0, 0xd82c7d56, 0xef903322); + &_data_word(0xc74e4987, 0xc1d138d9, 0xfea2ca8c, 0x360bd498); + &_data_word(0xcf81f5a6, 0x28de7aa5, 0x268eb7da, 0xa4bfad3f); + &_data_word(0xe49d3a2c, 0x0d927850, 0x9bcc5f6a, 0x62467e54); + &_data_word(0xc2138df6, 0xe8b8d890, 0x5ef7392e, 0xf5afc382); + &_data_word(0xbe805d9f, 0x7c93d069, 0xa92dd56f, 0xb31225cf); + &_data_word(0x3b99acc8, 0xa77d1810, 0x6e639ce8, 0x7bbb3bdb); + &_data_word(0x097826cd, 0xf418596e, 0x01b79aec, 0xa89a4f83); + &_data_word(0x656e95e6, 0x7ee6ffaa, 0x08cfbc21, 0xe6e815ef); + &_data_word(0xd99be7ba, 0xce366f4a, 0xd4099fea, 0xd67cb029); + &_data_word(0xafb2a431, 0x31233f2a, 0x3094a5c6, 0xc066a235); + &_data_word(0x37bc4e74, 0xa6ca82fc, 0xb0d090e0, 0x15d8a733); + &_data_word(0x4a9804f1, 0xf7daec41, 0x0e50cd7f, 0x2ff69117); + &_data_word(0x8dd64d76, 0x4db0ef43, 0x544daacc, 0xdf0496e4); + &_data_word(0xe3b5d19e, 0x1b886a4c, 0xb81f2cc1, 0x7f516546); + &_data_word(0x04ea5e9d, 0x5d358c01, 0x737487fa, 0x2e410bfb); + &_data_word(0x5a1d67b3, 0x52d2db92, 0x335610e9, 0x1347d66d); + &_data_word(0x8c61d79a, 0x7a0ca137, 0x8e14f859, 0x893c13eb); + &_data_word(0xee27a9ce, 0x35c961b7, 0xede51ce1, 0x3cb1477a); + &_data_word(0x59dfd29c, 0x3f73f255, 0x79ce1418, 0xbf37c773); + &_data_word(0xeacdf753, 0x5baafd5f, 0x146f3ddf, 0x86db4478); + &_data_word(0x81f3afca, 0x3ec468b9, 0x2c342438, 0x5f40a3c2); + &_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff); + &_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664); + &_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0); +#Td4: + &data_word(0x52525252, 0x09090909, 0x6a6a6a6a, 0xd5d5d5d5); + &data_word(0x30303030, 0x36363636, 0xa5a5a5a5, 0x38383838); + &data_word(0xbfbfbfbf, 0x40404040, 0xa3a3a3a3, 0x9e9e9e9e); + &data_word(0x81818181, 0xf3f3f3f3, 0xd7d7d7d7, 0xfbfbfbfb); + &data_word(0x7c7c7c7c, 0xe3e3e3e3, 0x39393939, 0x82828282); + &data_word(0x9b9b9b9b, 0x2f2f2f2f, 0xffffffff, 0x87878787); + &data_word(0x34343434, 0x8e8e8e8e, 0x43434343, 0x44444444); + &data_word(0xc4c4c4c4, 0xdededede, 0xe9e9e9e9, 0xcbcbcbcb); + &data_word(0x54545454, 0x7b7b7b7b, 0x94949494, 0x32323232); + &data_word(0xa6a6a6a6, 0xc2c2c2c2, 0x23232323, 0x3d3d3d3d); + &data_word(0xeeeeeeee, 0x4c4c4c4c, 0x95959595, 0x0b0b0b0b); + &data_word(0x42424242, 0xfafafafa, 0xc3c3c3c3, 0x4e4e4e4e); + &data_word(0x08080808, 0x2e2e2e2e, 0xa1a1a1a1, 0x66666666); + &data_word(0x28282828, 0xd9d9d9d9, 0x24242424, 0xb2b2b2b2); + &data_word(0x76767676, 0x5b5b5b5b, 0xa2a2a2a2, 0x49494949); + &data_word(0x6d6d6d6d, 0x8b8b8b8b, 0xd1d1d1d1, 0x25252525); + &data_word(0x72727272, 0xf8f8f8f8, 0xf6f6f6f6, 0x64646464); + &data_word(0x86868686, 0x68686868, 0x98989898, 0x16161616); + &data_word(0xd4d4d4d4, 0xa4a4a4a4, 0x5c5c5c5c, 0xcccccccc); + &data_word(0x5d5d5d5d, 0x65656565, 0xb6b6b6b6, 0x92929292); + &data_word(0x6c6c6c6c, 0x70707070, 0x48484848, 0x50505050); + &data_word(0xfdfdfdfd, 0xedededed, 0xb9b9b9b9, 0xdadadada); + &data_word(0x5e5e5e5e, 0x15151515, 0x46464646, 0x57575757); + &data_word(0xa7a7a7a7, 0x8d8d8d8d, 0x9d9d9d9d, 0x84848484); + &data_word(0x90909090, 0xd8d8d8d8, 0xabababab, 0x00000000); + &data_word(0x8c8c8c8c, 0xbcbcbcbc, 0xd3d3d3d3, 0x0a0a0a0a); + &data_word(0xf7f7f7f7, 0xe4e4e4e4, 0x58585858, 0x05050505); + &data_word(0xb8b8b8b8, 0xb3b3b3b3, 0x45454545, 0x06060606); + &data_word(0xd0d0d0d0, 0x2c2c2c2c, 0x1e1e1e1e, 0x8f8f8f8f); + &data_word(0xcacacaca, 0x3f3f3f3f, 0x0f0f0f0f, 0x02020202); + &data_word(0xc1c1c1c1, 0xafafafaf, 0xbdbdbdbd, 0x03030303); + &data_word(0x01010101, 0x13131313, 0x8a8a8a8a, 0x6b6b6b6b); + &data_word(0x3a3a3a3a, 0x91919191, 0x11111111, 0x41414141); + &data_word(0x4f4f4f4f, 0x67676767, 0xdcdcdcdc, 0xeaeaeaea); + &data_word(0x97979797, 0xf2f2f2f2, 0xcfcfcfcf, 0xcececece); + &data_word(0xf0f0f0f0, 0xb4b4b4b4, 0xe6e6e6e6, 0x73737373); + &data_word(0x96969696, 0xacacacac, 0x74747474, 0x22222222); + &data_word(0xe7e7e7e7, 0xadadadad, 0x35353535, 0x85858585); + &data_word(0xe2e2e2e2, 0xf9f9f9f9, 0x37373737, 0xe8e8e8e8); + &data_word(0x1c1c1c1c, 0x75757575, 0xdfdfdfdf, 0x6e6e6e6e); + &data_word(0x47474747, 0xf1f1f1f1, 0x1a1a1a1a, 0x71717171); + &data_word(0x1d1d1d1d, 0x29292929, 0xc5c5c5c5, 0x89898989); + &data_word(0x6f6f6f6f, 0xb7b7b7b7, 0x62626262, 0x0e0e0e0e); + &data_word(0xaaaaaaaa, 0x18181818, 0xbebebebe, 0x1b1b1b1b); + &data_word(0xfcfcfcfc, 0x56565656, 0x3e3e3e3e, 0x4b4b4b4b); + &data_word(0xc6c6c6c6, 0xd2d2d2d2, 0x79797979, 0x20202020); + &data_word(0x9a9a9a9a, 0xdbdbdbdb, 0xc0c0c0c0, 0xfefefefe); + &data_word(0x78787878, 0xcdcdcdcd, 0x5a5a5a5a, 0xf4f4f4f4); + &data_word(0x1f1f1f1f, 0xdddddddd, 0xa8a8a8a8, 0x33333333); + &data_word(0x88888888, 0x07070707, 0xc7c7c7c7, 0x31313131); + &data_word(0xb1b1b1b1, 0x12121212, 0x10101010, 0x59595959); + &data_word(0x27272727, 0x80808080, 0xecececec, 0x5f5f5f5f); + &data_word(0x60606060, 0x51515151, 0x7f7f7f7f, 0xa9a9a9a9); + &data_word(0x19191919, 0xb5b5b5b5, 0x4a4a4a4a, 0x0d0d0d0d); + &data_word(0x2d2d2d2d, 0xe5e5e5e5, 0x7a7a7a7a, 0x9f9f9f9f); + &data_word(0x93939393, 0xc9c9c9c9, 0x9c9c9c9c, 0xefefefef); + &data_word(0xa0a0a0a0, 0xe0e0e0e0, 0x3b3b3b3b, 0x4d4d4d4d); + &data_word(0xaeaeaeae, 0x2a2a2a2a, 0xf5f5f5f5, 0xb0b0b0b0); + &data_word(0xc8c8c8c8, 0xebebebeb, 0xbbbbbbbb, 0x3c3c3c3c); + &data_word(0x83838383, 0x53535353, 0x99999999, 0x61616161); + &data_word(0x17171717, 0x2b2b2b2b, 0x04040404, 0x7e7e7e7e); + &data_word(0xbabababa, 0x77777777, 0xd6d6d6d6, 0x26262626); + &data_word(0xe1e1e1e1, 0x69696969, 0x14141414, 0x63636363); + &data_word(0x55555555, 0x21212121, 0x0c0c0c0c, 0x7d7d7d7d); +&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); + + &call (&label("pic_point")); # make it PIC! + &set_label("pic_point"); + &blindpop("ebp"); + &lea ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp")); + + &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")); + + &mov ($acc,&wparam(1)); # load out + &mov (&DWP(0,$acc),$s0); # write output data + &mov (&DWP(4,$acc),$s1); + &mov (&DWP(8,$acc),$s2); + &mov (&DWP(12,$acc),$s3); +&function_end("AES_decrypt"); + +# void AES_cbc_encrypt (const void char *inp, unsigned char *out, +# size_t length, const AES_KEY *key, +# 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 $aes_key=&DWP(60,"esp"); #copy of aes_key + +&public_label("AES_Te"); +&public_label("AES_Td"); +&function_begin("AES_cbc_encrypt"); + &mov ($s2 eq "ecx"? $s2 : "",&wparam(2)); # load len + &cmp ($s2,0); + &je (&label("enc_out")); + + &call (&label("pic_point")); # make it PIC! + &set_label("pic_point"); + &blindpop("ebp"); + + &pushf (); + &cld (); + + &cmp (&wparam(5),0); + &je (&label("DECRYPT")); + + &lea ("ebp",&DWP(&label("AES_Te")."-".&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_Te modulo 4096 + &mov ($s0,"ebp"); + &lea ($s1,&DWP(2048,"ebp")); + &mov ($s3,$key); + &and ($s0,0xfff); # s = %ebp&0xfff + &and ($s1,0xfff); # e = (%ebp+2048)&0xfff + &and ($s3,0xfff); # p = %esp&0xfff + + &cmp ($s3,$s1); # if (p>=e) %esp =- (p-e); + &jb (&label("te_break_out")); + &sub ($s3,$s1); + &sub ($key,$s3); + &jmp (&label("te_ok")); + &set_label("te_break_out"); # else %esp -= (p-s)&0xfff + framesz; + &sub ($s3,$s0); + &and ($s3,0xfff); + &add ($s3,64+256); + &sub ($key,$s3); + &align (4); + &set_label("te_ok"); + + &mov ($s0,&wparam(0)); # load inp + &mov ($s1,&wparam(1)); # load out + &mov ($s3,&wparam(3)); # load key + &mov ($acc,&wparam(4)); # load ivp + + &exch ("esp",$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 + + if ($compromise) { + &cmp ($s2,$compromise); + &jb (&label("skip_ecopy")); + } + # copy key schedule to stack + &mov ("ecx",244/4); + &mov ("esi",$s3); + &lea ("edi",$aes_key); + &mov ($_key,"edi"); + &align (4); + &data_word(0xF689A5F3); # rep movsd + &set_label("skip_ecopy") if ($compromise); + + &mov ($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,$_len); + &mov ($key,$_ivp); + &test ($s2,0xFFFFFFF0); + &jz (&label("enc_tail")); # short input... + + &mov ($s0,&DWP(0,$key)); # load iv + &mov ($s1,&DWP(4,$key)); + + &align (4); + &set_label("enc_loop"); + &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"); + + &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 + + &lea ($acc,&DWP(16,$acc)); + &mov ($_inp,$acc); # save inp + + &lea ($s3,&DWP(16,$key)); + &mov ($_out,$s3); # save out + + &sub ($s2,16); + &test ($s2,0xFFFFFFF0); + &mov ($_len,$s2); # save len + &jnz (&label("enc_loop")); + &test ($s2,15); + &jnz (&label("enc_tail")); + &mov ($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),$s3); + + &mov ("edi",$_key); + &mov ("esp",$_esp); + if ($compromise) { + &cmp (&wparam(2),$compromise); + &jb (&label("skip_ezero")); + } + # zero copy of key schedule + &mov ("ecx",240/4); + &xor ("eax","eax"); + &align (4); + &data_word(0xF689ABF3); # rep stosd + &set_label("skip_ezero") if ($compromise); + &popf (); + &set_label("enc_out"); + &function_end_A(); + &pushf (); # kludge, never executed + + &align (4); + &set_label("enc_tail"); + &push ($key eq "edi" ? $key : ""); # push ivp + &mov ($key,$_out); # load out + &mov ($s1,16); + &sub ($s1,$s2); + &cmp ($key,$acc); # compare with inp + &je (&label("enc_in_place")); + &align (4); + &data_word(0xF689A4F3); # 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,$s0); + &align (4); + &data_word(0xF689AAF3); # rep stosb # zero tail + &pop ($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 ($s1,&DWP(3072,"ebp")); + &mov ($s3,$key); + &and ($s0,0xfff); # s = %ebp&0xfff + &and ($s1,0xfff); # e = (%ebp+3072)&0xfff + &and ($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,$acc); # save copy of ivp + + if ($compromise) { + &cmp ($s2,$compromise); + &jb (&label("skip_dcopy")); + } + # copy key schedule to stack + &mov ("ecx",244/4); + &mov ("esi",$s3); + &lea ("edi",$aes_key); + &mov ($_key,"edi"); + &align (4); + &data_word(0xF689A5F3); # rep movsd + &set_label("skip_dcopy") if ($compromise); + + &mov ($acc,$s0); + &mov ($key,24); + &align (4); + &set_label("prefetch_td"); + &mov ($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",3072); + + &cmp ($acc,$_out); + &je (&label("dec_in_place")); # in-place processing... + + &mov ($key,$_ivp); # load ivp + &mov ($_tmp,$key); + + &align (4); + &set_label("dec_loop"); + &mov ($s0,&DWP(0,$acc)); # read input + &mov ($s1,&DWP(4,$acc)); + &mov ($s2,&DWP(8,$acc)); + &mov ($s3,&DWP(12,$acc)); + + &mov ($key,$_key); # load key + &call ("_x86_AES_decrypt"); + + &mov ($key,$_tmp); # 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 ($acc,16); + &jc (&label("dec_partial")); + &mov ($_len,$acc); # save len + &mov ($acc,$_inp); # load inp + &mov ($key,$_out); # load out + + &mov (&DWP(0,$key),$s0); # write output + &mov (&DWP(4,$key),$s1); + &mov (&DWP(8,$key),$s2); + &mov (&DWP(12,$key),$s3); + + &mov ($_tmp,$acc); # save ivp + &lea ($acc,&DWP(16,$acc)); + &mov ($_inp,$acc); # save inp + + &lea ($key,&DWP(16,$key)); + &mov ($_out,$key); # save out + + &jnz (&label("dec_loop")); + &mov ($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,$key)); + &mov ($s2,&DWP(8,$key)); + &mov ($s3,&DWP(12,$key)); + &mov (&DWP(0,$acc),$s0); # copy back to user + &mov (&DWP(4,$acc),$s1); + &mov (&DWP(8,$acc),$s2); + &mov (&DWP(12,$acc),$s3); + &jmp (&label("dec_out")); + + &align (4); + &set_label("dec_partial"); + &lea ($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(0xF689A4F3); # 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); + &mov ($s0,&DWP(0,$acc)); # read input + &mov ($s1,&DWP(4,$acc)); + &mov ($s2,&DWP(8,$acc)); + &mov ($s3,&DWP(12,$acc)); + + &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"); + + &mov ($key,$_ivp); # load ivp + &mov ($acc,$_out); # load out + &xor ($s0,&DWP(0,$key)); # xor iv + &xor ($s1,&DWP(4,$key)); + &xor ($s2,&DWP(8,$key)); + &xor ($s3,&DWP(12,$key)); + + &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)); + &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 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)); + &mov ($_inp,$acc); # save inp + + &mov ($s2,$_len); # load len + &sub ($s2,16); + &jc (&label("dec_in_place_partial")); + &mov ($_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); + &lea ($key,&DWP(0,$key,$s2)); + &lea ($acc,&DWP(16,$acc,$s2)); + &neg ($s2 eq "ecx" ? $s2 : ""); + &data_word(0xF689A4F3); # rep movsb # restore tail + + &align (4); + &set_label("dec_out"); + &mov ("edi",$_key); + &mov ("esp",$_esp); + if ($compromise) { + &cmp (&wparam(2),$compromise); + &jb (&label("skip_dzero")); + } + # zero copy of key schedule + &mov ("ecx",240/4); + &xor ("eax","eax"); + &align (4); + &data_word(0xF689ABF3); # rep stosd + &set_label("skip_dzero") if ($compromise); + &popf (); +&function_end("AES_cbc_encrypt"); +} + +#------------------------------------------------------------------# + +sub enckey() +{ + &movz ("esi",&LB("edx")); # rk[i]>>0 + &mov ("ebx",&DWP(2,"ebp","esi",8)); + &movz ("esi",&HB("edx")); # rk[i]>>8 + &and ("ebx",0xFF000000); + &xor ("eax","ebx"); + + &mov ("ebx",&DWP(2,"ebp","esi",8)); + &shr ("edx",16); + &and ("ebx",0x000000FF); + &movz ("esi",&LB("edx")); # rk[i]>>16 + &xor ("eax","ebx"); + + &mov ("ebx",&DWP(0,"ebp","esi",8)); + &movz ("esi",&HB("edx")); # rk[i]>>24 + &and ("ebx",0x0000FF00); + &xor ("eax","ebx"); + + &mov ("ebx",&DWP(0,"ebp","esi",8)); + &and ("ebx",0x00FF0000); + &xor ("eax","ebx"); + + &xor ("eax",&DWP(2048,"ebp","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 + + &test ("esi",-1); + &jz (&label("badpointer")); + &test ("edi",-1); + &jz (&label("badpointer")); + + &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 + &cmp ("ecx",128); + &je (&label("10rounds")); + &cmp ("ecx",192); + &je (&label("12rounds")); + &cmp ("ecx",256); + &je (&label("14rounds")); + &mov ("eax",-2); # invalid number of bits + &jmp (&label("exit")); + + &set_label("10rounds"); + &mov ("eax",&DWP(0,"esi")); # copy first 4 dwords + &mov ("ebx",&DWP(4,"esi")); + &mov ("ecx",&DWP(8,"esi")); + &mov ("edx",&DWP(12,"esi")); + &mov (&DWP(0,"edi"),"eax"); + &mov (&DWP(4,"edi"),"ebx"); + &mov (&DWP(8,"edi"),"ecx"); + &mov (&DWP(12,"edi"),"edx"); + + &xor ("ecx","ecx"); + &jmp (&label("10shortcut")); + + &align (4); + &set_label("10loop"); + &mov ("eax",&DWP(0,"edi")); # rk[0] + &mov ("edx",&DWP(12,"edi")); # rk[3] + &set_label("10shortcut"); + &enckey (); + + &mov (&DWP(16,"edi"),"eax"); # rk[4] + &xor ("eax",&DWP(4,"edi")); + &mov (&DWP(20,"edi"),"eax"); # rk[5] + &xor ("eax",&DWP(8,"edi")); + &mov (&DWP(24,"edi"),"eax"); # rk[6] + &xor ("eax",&DWP(12,"edi")); + &mov (&DWP(28,"edi"),"eax"); # rk[7] + &inc ("ecx"); + &add ("edi",16); + &cmp ("ecx",10); + &jl (&label("10loop")); + + &mov (&DWP(80,"edi"),10); # setup number of rounds + &xor ("eax","eax"); + &jmp (&label("exit")); + + &set_label("12rounds"); + &mov ("eax",&DWP(0,"esi")); # copy first 6 dwords + &mov ("ebx",&DWP(4,"esi")); + &mov ("ecx",&DWP(8,"esi")); + &mov ("edx",&DWP(12,"esi")); + &mov (&DWP(0,"edi"),"eax"); + &mov (&DWP(4,"edi"),"ebx"); + &mov (&DWP(8,"edi"),"ecx"); + &mov (&DWP(12,"edi"),"edx"); + &mov ("ecx",&DWP(16,"esi")); + &mov ("edx",&DWP(20,"esi")); + &mov (&DWP(16,"edi"),"ecx"); + &mov (&DWP(20,"edi"),"edx"); + + &xor ("ecx","ecx"); + &jmp (&label("12shortcut")); + + &align (4); + &set_label("12loop"); + &mov ("eax",&DWP(0,"edi")); # rk[0] + &mov ("edx",&DWP(20,"edi")); # rk[5] + &set_label("12shortcut"); + &enckey (); + + &mov (&DWP(24,"edi"),"eax"); # rk[6] + &xor ("eax",&DWP(4,"edi")); + &mov (&DWP(28,"edi"),"eax"); # rk[7] + &xor ("eax",&DWP(8,"edi")); + &mov (&DWP(32,"edi"),"eax"); # rk[8] + &xor ("eax",&DWP(12,"edi")); + &mov (&DWP(36,"edi"),"eax"); # rk[9] + + &cmp ("ecx",7); + &je (&label("12break")); + &inc ("ecx"); + + &xor ("eax",&DWP(16,"edi")); + &mov (&DWP(40,"edi"),"eax"); # rk[10] + &xor ("eax",&DWP(20,"edi")); + &mov (&DWP(44,"edi"),"eax"); # rk[11] + + &add ("edi",24); + &jmp (&label("12loop")); + + &set_label("12break"); + &mov (&DWP(72,"edi"),12); # setup number of rounds + &xor ("eax","eax"); + &jmp (&label("exit")); + + &set_label("14rounds"); + &mov ("eax",&DWP(0,"esi")); # copy first 8 dwords + &mov ("ebx",&DWP(4,"esi")); + &mov ("ecx",&DWP(8,"esi")); + &mov ("edx",&DWP(12,"esi")); + &mov (&DWP(0,"edi"),"eax"); + &mov (&DWP(4,"edi"),"ebx"); + &mov (&DWP(8,"edi"),"ecx"); + &mov (&DWP(12,"edi"),"edx"); + &mov ("eax",&DWP(16,"esi")); + &mov ("ebx",&DWP(20,"esi")); + &mov ("ecx",&DWP(24,"esi")); + &mov ("edx",&DWP(28,"esi")); + &mov (&DWP(16,"edi"),"eax"); + &mov (&DWP(20,"edi"),"ebx"); + &mov (&DWP(24,"edi"),"ecx"); + &mov (&DWP(28,"edi"),"edx"); + + &xor ("ecx","ecx"); + &jmp (&label("14shortcut")); + + &align (4); + &set_label("14loop"); + &mov ("edx",&DWP(28,"edi")); # rk[7] + &set_label("14shortcut"); + &mov ("eax",&DWP(0,"edi")); # rk[0] + + &enckey (); + + &mov (&DWP(32,"edi"),"eax"); # rk[8] + &xor ("eax",&DWP(4,"edi")); + &mov (&DWP(36,"edi"),"eax"); # rk[9] + &xor ("eax",&DWP(8,"edi")); + &mov (&DWP(40,"edi"),"eax"); # rk[10] + &xor ("eax",&DWP(12,"edi")); + &mov (&DWP(44,"edi"),"eax"); # rk[11] + + &cmp ("ecx",6); + &je (&label("14break")); + &inc ("ecx"); + + &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 ("esi",&HB("edx")); # rk[11]>>8 + &and ("ebx",0x000000FF); + &xor ("eax","ebx"); + + &mov ("ebx",&DWP(0,"ebp","esi",8)); + &shr ("edx",16); + &and ("ebx",0x0000FF00); + &movz ("esi",&LB("edx")); # rk[11]>>16 + &xor ("eax","ebx"); + + &mov ("ebx",&DWP(0,"ebp","esi",8)); + &movz ("esi",&HB("edx")); # rk[11]>>24 + &and ("ebx",0x00FF0000); + &xor ("eax","ebx"); + + &mov ("ebx",&DWP(2,"ebp","esi",8)); + &and ("ebx",0xFF000000); + &xor ("eax","ebx"); + + &mov (&DWP(48,"edi"),"eax"); # rk[12] + &xor ("eax",&DWP(20,"edi")); + &mov (&DWP(52,"edi"),"eax"); # rk[13] + &xor ("eax",&DWP(24,"edi")); + &mov (&DWP(56,"edi"),"eax"); # rk[14] + &xor ("eax",&DWP(28,"edi")); + &mov (&DWP(60,"edi"),"eax"); # rk[15] + + &add ("edi",32); + &jmp (&label("14loop")); + + &set_label("14break"); + &mov (&DWP(48,"edi"),14); # setup number of rounds + &xor ("eax","eax"); + &jmp (&label("exit")); + + &set_label("badpointer"); + &mov ("eax",-1); + &set_label("exit"); +&function_end("AES_set_encrypt_key"); + +sub deckey() +{ my ($i,$ptr,$te,$td) = @_; + + &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"); +} + +# 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); + &cmp ("eax",0); + &je (&label("proceed")); + &ret (); + + &set_label("proceed"); + &push ("ebp"); + &push ("ebx"); + &push ("esi"); + &push ("edi"); + + &mov ("esi",&wparam(2)); + &mov ("ecx",&DWP(240,"esi")); # pull number of rounds + &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 + &mov ("eax",&DWP(0,"esi")); + &mov ("ebx",&DWP(4,"esi")); + &mov ("ecx",&DWP(0,"edi")); + &mov ("edx",&DWP(4,"edi")); + &mov (&DWP(0,"edi"),"eax"); + &mov (&DWP(4,"edi"),"ebx"); + &mov (&DWP(0,"esi"),"ecx"); + &mov (&DWP(4,"esi"),"edx"); + &mov ("eax",&DWP(8,"esi")); + &mov ("ebx",&DWP(12,"esi")); + &mov ("ecx",&DWP(8,"edi")); + &mov ("edx",&DWP(12,"edi")); + &mov (&DWP(8,"edi"),"eax"); + &mov (&DWP(12,"edi"),"ebx"); + &mov (&DWP(8,"esi"),"ecx"); + &mov (&DWP(12,"esi"),"edx"); + &add ("esi",16); + &sub ("edi",16); + &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")); + + &xor ("eax","eax"); # return success +&function_end("AES_set_decrypt_key"); + +&asm_finish(); diff --git a/crypto/openssl/crypto/aes/asm/aes-ia64.S b/crypto/openssl/crypto/aes/asm/aes-ia64.S new file mode 100644 index 000000000000..542cf335e995 --- /dev/null +++ b/crypto/openssl/crypto/aes/asm/aes-ia64.S @@ -0,0 +1,1652 @@ +// ==================================================================== +// 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. +// ==================================================================== +// +// What's wrong with compiler generated code? Compiler never uses +// variable 'shr' which is pairable with 'extr'/'dep' instructions. +// Then it uses 'zxt' which is an I-type, but can be replaced with +// 'and' which in turn can be assigned to M-port [there're double as +// much M-ports as there're I-ports on Itanium 2]. By sacrificing few +// registers for small constants (255, 24 and 16) to be used with +// 'shr' and 'and' instructions I can achieve better ILP, Intruction +// Level Parallelism, and performance. This code outperforms GCC 3.3 +// generated code by over factor of 2 (two), GCC 3.4 - by 70% and +// HP C - by 40%. Measured best-case scenario, i.e. aligned +// big-endian input, ECB timing on Itanium 2 is (18 + 13*rounds) +// ticks per block, or 9.25 CPU cycles per byte for 128 bit key. + +.ident "aes-ia64.S, version 1.1" +.ident "IA-64 ISA artwork by Andy Polyakov <appro@fy.chalmers.se>" +.explicit +.text + +rk0=r8; rk1=r9; + +prsave=r10; +maskff=r11; +twenty4=r14; +sixteen=r15; + +te00=r16; te11=r17; te22=r18; te33=r19; +te01=r20; te12=r21; te23=r22; te30=r23; +te02=r24; te13=r25; te20=r26; te31=r27; +te03=r28; te10=r29; te21=r30; te32=r31; + +// these are rotating... +t0=r32; s0=r33; +t1=r34; s1=r35; +t2=r36; s2=r37; +t3=r38; s3=r39; + +te0=r40; te1=r41; te2=r42; te3=r43; + +#if defined(_HPUX_SOURCE) && !defined(_LP64) +# define ADDP addp4 +# define KSZ 4 +# define LDKEY ld4 +#else +# define ADDP add +#endif + +// This implies that AES_KEY comprises 32-bit key schedule elements +// even on LP64 platforms. +#ifndef KSZ +# define KSZ 4 +# define LDKEY ld4 +#endif + +.proc _ia64_AES_encrypt# +// Input: rk0-rk1 +// te0 +// te3 as AES_KEY->rounds!!! +// s0-s3 +// maskff,twenty4,sixteen +// Output: r16,r20,r24,r28 as s0-s3 +// Clobber: r16-r31,rk0-rk1,r32-r43 +.align 32 +_ia64_AES_encrypt: +{ .mmi; alloc r16=ar.pfs,12,0,0,8 + LDKEY t0=[rk0],2*KSZ + mov pr.rot=1<<16 } +{ .mmi; LDKEY t1=[rk1],2*KSZ + add te1=1024,te0 + add te3=-3,te3 };; +{ .mib; LDKEY t2=[rk0],2*KSZ + mov ar.ec=3 } +{ .mib; LDKEY t3=[rk1],2*KSZ + add te2=2048,te0 + brp.loop.imp .Le_top,.Le_end-16 };; + +{ .mmi; xor s0=s0,t0 + xor s1=s1,t1 + mov ar.lc=te3 } +{ .mmi; xor s2=s2,t2 + xor s3=s3,t3 + add te3=3072,te0 };; + +.align 32 +.Le_top: +{ .mmi; (p0) LDKEY t0=[rk0],2*KSZ // 0/0:rk[0] + (p0) and te33=s3,maskff // 0/0:s3&0xff + (p0) extr.u te22=s2,8,8 } // 0/0:s2>>8&0xff +{ .mmi; (p0) LDKEY t1=[rk1],2*KSZ // 0/1:rk[1] + (p0) and te30=s0,maskff // 0/1:s0&0xff + (p0) shr.u te00=s0,twenty4 };; // 0/0:s0>>24 +{ .mmi; (p0) LDKEY t2=[rk0],2*KSZ // 1/2:rk[2] + (p0) shladd te33=te33,2,te3 // 1/0:te0+s0>>24 + (p0) extr.u te23=s3,8,8 } // 1/1:s3>>8&0xff +{ .mmi; (p0) LDKEY t3=[rk1],2*KSZ // 1/3:rk[3] + (p0) shladd te30=te30,2,te3 // 1/1:te3+s0 + (p0) shr.u te01=s1,twenty4 };; // 1/1:s1>>24 +{ .mmi; (p0) ld4 te33=[te33] // 2/0:te3[s3&0xff] + (p0) shladd te22=te22,2,te2 // 2/0:te2+s2>>8&0xff + (p0) extr.u te20=s0,8,8 } // 2/2:s0>>8&0xff +{ .mmi; (p0) ld4 te30=[te30] // 2/1:te3[s0] + (p0) shladd te23=te23,2,te2 // 2/1:te2+s3>>8 + (p0) shr.u te02=s2,twenty4 };; // 2/2:s2>>24 +{ .mmi; (p0) ld4 te22=[te22] // 3/0:te2[s2>>8] + (p0) shladd te20=te20,2,te2 // 3/2:te2+s0>>8 + (p0) extr.u te21=s1,8,8 } // 3/3:s1>>8&0xff +{ .mmi; (p0) ld4 te23=[te23] // 3/1:te2[s3>>8] + (p0) shladd te00=te00,2,te0 // 3/0:te0+s0>>24 + (p0) shr.u te03=s3,twenty4 };; // 3/3:s3>>24 +{ .mmi; (p0) ld4 te20=[te20] // 4/2:te2[s0>>8] + (p0) shladd te21=te21,2,te2 // 4/3:te3+s2 + (p0) extr.u te11=s1,16,8 } // 4/0:s1>>16&0xff +{ .mmi; (p0) ld4 te00=[te00] // 4/0:te0[s0>>24] + (p0) shladd te01=te01,2,te0 // 4/1:te0+s1>>24 + (p0) shr.u te13=s3,sixteen };; // 4/2:s3>>16 +{ .mmi; (p0) ld4 te21=[te21] // 5/3:te2[s1>>8] + (p0) shladd te11=te11,2,te1 // 5/0:te1+s1>>16 + (p0) extr.u te12=s2,16,8 } // 5/1:s2>>16&0xff +{ .mmi; (p0) ld4 te01=[te01] // 5/1:te0[s1>>24] + (p0) shladd te02=te02,2,te0 // 5/2:te0+s2>>24 + (p0) and te31=s1,maskff };; // 5/2:s1&0xff + +{ .mmi; (p0) ld4 te11=[te11] // 6/0:te1[s1>>16] + (p0) shladd te12=te12,2,te1 // 6/1:te1+s2>>16 + (p0) extr.u te10=s0,16,8 } // 6/3:s0>>16&0xff +{ .mmi; (p0) ld4 te02=[te02] // 6/2:te0[s2>>24] + (p0) shladd te03=te03,2,te0 // 6/3:te1+s0>>16 + (p0) and te32=s2,maskff };; // 6/3:s2&0xff +{ .mmi; (p0) ld4 te12=[te12] // 7/1:te1[s2>>16] + (p0) shladd te31=te31,2,te3 // 7/2:te3+s1&0xff + (p0) and te13=te13,maskff} // 7/2:s3>>16&0xff +{ .mmi; (p0) ld4 te03=[te03] // 7/3:te0[s3>>24] + (p0) shladd te32=te32,2,te3 // 7/3:te3+s2 + (p0) xor t0=t0,te33 };; // 7/0: +{ .mmi; (p0) ld4 te31=[te31] // 8/2:te3[s1] + (p0) shladd te13=te13,2,te1 // 8/2:te1+s3>>16 + (p0) xor t0=t0,te22 } // 8/0: +{ .mmi; (p0) ld4 te32=[te32] // 8/3:te3[s2] + (p0) shladd te10=te10,2,te1 // 8/3:te1+s0>>16 + (p0) xor t1=t1,te30 };; // 8/1: +{ .mmi; (p0) ld4 te13=[te13] // 9/2:te1[s3>>16] + (p0) xor t0=t0,te00 // 9/0: + (p0) xor t1=t1,te23 } // 9/1: +{ .mmi; (p0) ld4 te10=[te10] // 9/3:te1[s0>>16] + (p0) xor t2=t2,te20 // 9/2: + (p0) xor t3=t3,te21 };; // 9/3: +{ .mmi; (p0) xor t0=t0,te11 // 10/0:done! + (p0) xor t1=t1,te01 // 10/1: + (p0) xor t2=t2,te02 } // 10/2: +{ .mmi; (p0) xor t3=t3,te03 // 10/3: + (p16) cmp.eq p0,p17=r0,r0 };; // 10/clear (p17) +{ .mmi; (p0) xor t1=t1,te12 // 11/1:done! + (p0) xor t2=t2,te31 // 11/2: + (p0) xor t3=t3,te32 } // 11/3: +{ .mmi; (p17) add te0=4096,te0 // 11/ + (p17) add te1=4096,te1 };; // 11/ +{ .mib; (p0) xor t2=t2,te13 // 12/2:done! + (p0) xor t3=t3,te10 } // 12/3:done! +{ .mib; (p17) add te2=4096,te2 // 12/ + (p17) add te3=4096,te3 // 12/ + br.ctop.sptk .Le_top };; +.Le_end: +{ .mib; mov r16=s0 + mov r20=s1 } +{ .mib; mov r24=s2 + mov r28=s3 + br.ret.sptk b6 };; +.endp _ia64_AES_encrypt# + +// void AES_encrypt (const void *in,void *out,const AES_KEY *key); +.global AES_encrypt# +.proc AES_encrypt# +.align 32 +.skip 16 +AES_encrypt: + .prologue + .fframe 0 + .save ar.pfs,r2 + .save ar.lc,r3 +{ .mmi; alloc r2=ar.pfs,3,0,12,0 + addl out8=@ltoff(AES_Te#),gp + mov r3=ar.lc } +{ .mmi; and out0=3,in0 + ADDP in0=0,in0 + ADDP out11=KSZ*60,in2 };; // &AES_KEY->rounds + + .body +{ .mmi; ld8 out8=[out8] // Te0 + ld4 out11=[out11] // AES_KEY->rounds + mov prsave=pr } + +#if defined(_HPUX_SOURCE) // HPUX is big-endian, cut 15+15 cycles... +{ .mib; cmp.ne p6,p0=out0,r0 + add out0=4,in0 +(p6) br.dpnt.many .Le_i_unaligned };; + +{ .mmi; ld4 out1=[in0],8 // s0 + and out9=3,in1 + mov twenty4=24 } +{ .mmi; ld4 out3=[out0],8 // s1 + ADDP rk0=0,in2 + mov sixteen=16 };; +{ .mmi; ld4 out5=[in0] // s2 + cmp.ne p6,p0=out9,r0 + mov maskff=0xff } +{ .mmb; ld4 out7=[out0] // s3 + ADDP rk1=KSZ,in2 + br.call.sptk.many b6=_ia64_AES_encrypt };; + +{ .mib; ADDP in0=4,in1 + ADDP in1=0,in1 +(p6) br.spnt .Le_o_unaligned };; + +{ .mii; mov ar.pfs=r2 + mov ar.lc=r3 } +{ .mmi; st4 [in1]=r16,8 // s0 + st4 [in0]=r20,8 // s1 + mov pr=prsave,0x1ffff };; +{ .mmb; st4 [in1]=r24 // s2 + st4 [in0]=r28 // s3 + br.ret.sptk.many b0 };; +#endif + +.align 32 +.Le_i_unaligned: +{ .mmi; add out0=1,in0 + add out2=2,in0 + add out4=3,in0 };; +{ .mmi; ld1 r16=[in0],4 + ld1 r17=[out0],4 }//;; +{ .mmi; ld1 r18=[out2],4 + ld1 out1=[out4],4 };; // s0 +{ .mmi; ld1 r20=[in0],4 + ld1 r21=[out0],4 }//;; +{ .mmi; ld1 r22=[out2],4 + ld1 out3=[out4],4 };; // s1 +{ .mmi; ld1 r24=[in0],4 + ld1 r25=[out0],4 }//;; +{ .mmi; ld1 r26=[out2],4 + ld1 out5=[out4],4 };; // s2 +{ .mmi; ld1 r28=[in0] + ld1 r29=[out0] }//;; +{ .mmi; ld1 r30=[out2] + ld1 out7=[out4] };; // s3 + +{ .mii; + dep out1=r16,out1,24,8 //;; + dep out3=r20,out3,24,8 }//;; +{ .mii; ADDP rk0=0,in2 + dep out5=r24,out5,24,8 //;; + dep out7=r28,out7,24,8 };; +{ .mii; ADDP rk1=KSZ,in2 + dep out1=r17,out1,16,8 //;; + dep out3=r21,out3,16,8 }//;; +{ .mii; mov twenty4=24 + dep out5=r25,out5,16,8 //;; + dep out7=r29,out7,16,8 };; +{ .mii; mov sixteen=16 + dep out1=r18,out1,8,8 //;; + dep out3=r22,out3,8,8 }//;; +{ .mii; mov maskff=0xff + dep out5=r26,out5,8,8 //;; + dep out7=r30,out7,8,8 };; + +{ .mib; br.call.sptk.many b6=_ia64_AES_encrypt };; + +.Le_o_unaligned: +{ .mii; ADDP out0=0,in1 + extr.u r17=r16,8,8 // s0 + shr.u r19=r16,twenty4 }//;; +{ .mii; ADDP out1=1,in1 + extr.u r18=r16,16,8 + shr.u r23=r20,twenty4 }//;; // s1 +{ .mii; ADDP out2=2,in1 + extr.u r21=r20,8,8 + shr.u r22=r20,sixteen }//;; +{ .mii; ADDP out3=3,in1 + extr.u r25=r24,8,8 // s2 + shr.u r27=r24,twenty4 };; +{ .mii; st1 [out3]=r16,4 + extr.u r26=r24,16,8 + shr.u r31=r28,twenty4 }//;; // s3 +{ .mii; st1 [out2]=r17,4 + extr.u r29=r28,8,8 + shr.u r30=r28,sixteen }//;; + +{ .mmi; st1 [out1]=r18,4 + st1 [out0]=r19,4 };; +{ .mmi; st1 [out3]=r20,4 + st1 [out2]=r21,4 }//;; +{ .mmi; st1 [out1]=r22,4 + st1 [out0]=r23,4 };; +{ .mmi; st1 [out3]=r24,4 + st1 [out2]=r25,4 + mov pr=prsave,0x1ffff }//;; +{ .mmi; st1 [out1]=r26,4 + st1 [out0]=r27,4 + mov ar.pfs=r2 };; +{ .mmi; st1 [out3]=r28 + st1 [out2]=r29 + mov ar.lc=r3 }//;; +{ .mmb; st1 [out1]=r30 + st1 [out0]=r31 + br.ret.sptk.many b0 };; +.endp AES_encrypt# + +// *AES_decrypt are autogenerated by the following script: +#if 0 +#!/usr/bin/env perl +print "// *AES_decrypt are autogenerated by the following script:\n#if 0\n"; +open(PROG,'<'.$0); while(<PROG>) { print; } close(PROG); +print "#endif\n"; +while(<>) { + $process=1 if (/\.proc\s+_ia64_AES_encrypt/); + next if (!$process); + + #s/te00=s0/td00=s0/; s/te00/td00/g; + s/te11=s1/td13=s3/; s/te11/td13/g; + #s/te22=s2/td22=s2/; s/te22/td22/g; + s/te33=s3/td31=s1/; s/te33/td31/g; + + #s/te01=s1/td01=s1/; s/te01/td01/g; + s/te12=s2/td10=s0/; s/te12/td10/g; + #s/te23=s3/td23=s3/; s/te23/td23/g; + s/te30=s0/td32=s2/; s/te30/td32/g; + + #s/te02=s2/td02=s2/; s/te02/td02/g; + s/te13=s3/td11=s1/; s/te13/td11/g; + #s/te20=s0/td20=s0/; s/te20/td20/g; + s/te31=s1/td33=s3/; s/te31/td33/g; + + #s/te03=s3/td03=s3/; s/te03/td03/g; + s/te10=s0/td12=s2/; s/te10/td12/g; + #s/te21=s1/td21=s1/; s/te21/td21/g; + s/te32=s2/td30=s0/; s/te32/td30/g; + + s/td/te/g; + + s/AES_encrypt/AES_decrypt/g; + s/\.Le_/.Ld_/g; + s/AES_Te#/AES_Td#/g; + + print; + + exit if (/\.endp\s+AES_decrypt/); +} +#endif +.proc _ia64_AES_decrypt# +// Input: rk0-rk1 +// te0 +// te3 as AES_KEY->rounds!!! +// s0-s3 +// maskff,twenty4,sixteen +// Output: r16,r20,r24,r28 as s0-s3 +// Clobber: r16-r31,rk0-rk1,r32-r43 +.align 32 +_ia64_AES_decrypt: +{ .mmi; alloc r16=ar.pfs,12,0,0,8 + LDKEY t0=[rk0],2*KSZ + mov pr.rot=1<<16 } +{ .mmi; LDKEY t1=[rk1],2*KSZ + add te1=1024,te0 + add te3=-3,te3 };; +{ .mib; LDKEY t2=[rk0],2*KSZ + mov ar.ec=3 } +{ .mib; LDKEY t3=[rk1],2*KSZ + add te2=2048,te0 + brp.loop.imp .Ld_top,.Ld_end-16 };; + +{ .mmi; xor s0=s0,t0 + xor s1=s1,t1 + mov ar.lc=te3 } +{ .mmi; xor s2=s2,t2 + xor s3=s3,t3 + add te3=3072,te0 };; + +.align 32 +.Ld_top: +{ .mmi; (p0) LDKEY t0=[rk0],2*KSZ // 0/0:rk[0] + (p0) and te31=s1,maskff // 0/0:s3&0xff + (p0) extr.u te22=s2,8,8 } // 0/0:s2>>8&0xff +{ .mmi; (p0) LDKEY t1=[rk1],2*KSZ // 0/1:rk[1] + (p0) and te32=s2,maskff // 0/1:s0&0xff + (p0) shr.u te00=s0,twenty4 };; // 0/0:s0>>24 +{ .mmi; (p0) LDKEY t2=[rk0],2*KSZ // 1/2:rk[2] + (p0) shladd te31=te31,2,te3 // 1/0:te0+s0>>24 + (p0) extr.u te23=s3,8,8 } // 1/1:s3>>8&0xff +{ .mmi; (p0) LDKEY t3=[rk1],2*KSZ // 1/3:rk[3] + (p0) shladd te32=te32,2,te3 // 1/1:te3+s0 + (p0) shr.u te01=s1,twenty4 };; // 1/1:s1>>24 +{ .mmi; (p0) ld4 te31=[te31] // 2/0:te3[s3&0xff] + (p0) shladd te22=te22,2,te2 // 2/0:te2+s2>>8&0xff + (p0) extr.u te20=s0,8,8 } // 2/2:s0>>8&0xff +{ .mmi; (p0) ld4 te32=[te32] // 2/1:te3[s0] + (p0) shladd te23=te23,2,te2 // 2/1:te2+s3>>8 + (p0) shr.u te02=s2,twenty4 };; // 2/2:s2>>24 +{ .mmi; (p0) ld4 te22=[te22] // 3/0:te2[s2>>8] + (p0) shladd te20=te20,2,te2 // 3/2:te2+s0>>8 + (p0) extr.u te21=s1,8,8 } // 3/3:s1>>8&0xff +{ .mmi; (p0) ld4 te23=[te23] // 3/1:te2[s3>>8] + (p0) shladd te00=te00,2,te0 // 3/0:te0+s0>>24 + (p0) shr.u te03=s3,twenty4 };; // 3/3:s3>>24 +{ .mmi; (p0) ld4 te20=[te20] // 4/2:te2[s0>>8] + (p0) shladd te21=te21,2,te2 // 4/3:te3+s2 + (p0) extr.u te13=s3,16,8 } // 4/0:s1>>16&0xff +{ .mmi; (p0) ld4 te00=[te00] // 4/0:te0[s0>>24] + (p0) shladd te01=te01,2,te0 // 4/1:te0+s1>>24 + (p0) shr.u te11=s1,sixteen };; // 4/2:s3>>16 +{ .mmi; (p0) ld4 te21=[te21] // 5/3:te2[s1>>8] + (p0) shladd te13=te13,2,te1 // 5/0:te1+s1>>16 + (p0) extr.u te10=s0,16,8 } // 5/1:s2>>16&0xff +{ .mmi; (p0) ld4 te01=[te01] // 5/1:te0[s1>>24] + (p0) shladd te02=te02,2,te0 // 5/2:te0+s2>>24 + (p0) and te33=s3,maskff };; // 5/2:s1&0xff + +{ .mmi; (p0) ld4 te13=[te13] // 6/0:te1[s1>>16] + (p0) shladd te10=te10,2,te1 // 6/1:te1+s2>>16 + (p0) extr.u te12=s2,16,8 } // 6/3:s0>>16&0xff +{ .mmi; (p0) ld4 te02=[te02] // 6/2:te0[s2>>24] + (p0) shladd te03=te03,2,te0 // 6/3:te1+s0>>16 + (p0) and te30=s0,maskff };; // 6/3:s2&0xff +{ .mmi; (p0) ld4 te10=[te10] // 7/1:te1[s2>>16] + (p0) shladd te33=te33,2,te3 // 7/2:te3+s1&0xff + (p0) and te11=te11,maskff} // 7/2:s3>>16&0xff +{ .mmi; (p0) ld4 te03=[te03] // 7/3:te0[s3>>24] + (p0) shladd te30=te30,2,te3 // 7/3:te3+s2 + (p0) xor t0=t0,te31 };; // 7/0: +{ .mmi; (p0) ld4 te33=[te33] // 8/2:te3[s1] + (p0) shladd te11=te11,2,te1 // 8/2:te1+s3>>16 + (p0) xor t0=t0,te22 } // 8/0: +{ .mmi; (p0) ld4 te30=[te30] // 8/3:te3[s2] + (p0) shladd te12=te12,2,te1 // 8/3:te1+s0>>16 + (p0) xor t1=t1,te32 };; // 8/1: +{ .mmi; (p0) ld4 te11=[te11] // 9/2:te1[s3>>16] + (p0) xor t0=t0,te00 // 9/0: + (p0) xor t1=t1,te23 } // 9/1: +{ .mmi; (p0) ld4 te12=[te12] // 9/3:te1[s0>>16] + (p0) xor t2=t2,te20 // 9/2: + (p0) xor t3=t3,te21 };; // 9/3: +{ .mmi; (p0) xor t0=t0,te13 // 10/0:done! + (p0) xor t1=t1,te01 // 10/1: + (p0) xor t2=t2,te02 } // 10/2: +{ .mmi; (p0) xor t3=t3,te03 // 10/3: + (p16) cmp.eq p0,p17=r0,r0 };; // 10/clear (p17) +{ .mmi; (p0) xor t1=t1,te10 // 11/1:done! + (p0) xor t2=t2,te33 // 11/2: + (p0) xor t3=t3,te30 } // 11/3: +{ .mmi; (p17) add te0=4096,te0 // 11/ + (p17) add te1=4096,te1 };; // 11/ +{ .mib; (p0) xor t2=t2,te11 // 12/2:done! + (p0) xor t3=t3,te12 } // 12/3:done! +{ .mib; (p17) add te2=4096,te2 // 12/ + (p17) add te3=4096,te3 // 12/ + br.ctop.sptk .Ld_top };; +.Ld_end: +{ .mib; mov r16=s0 + mov r20=s1 } +{ .mib; mov r24=s2 + mov r28=s3 + br.ret.sptk b6 };; +.endp _ia64_AES_decrypt# + +// void AES_decrypt (const void *in,void *out,const AES_KEY *key); +.global AES_decrypt# +.proc AES_decrypt# +.align 32 +.skip 16 +AES_decrypt: + .prologue + .fframe 0 + .save ar.pfs,r2 + .save ar.lc,r3 +{ .mmi; alloc r2=ar.pfs,3,0,12,0 + addl out8=@ltoff(AES_Td#),gp + mov r3=ar.lc } +{ .mmi; and out0=3,in0 + ADDP in0=0,in0 + ADDP out11=KSZ*60,in2 };; // &AES_KEY->rounds + + .body +{ .mmi; ld8 out8=[out8] // Te0 + ld4 out11=[out11] // AES_KEY->rounds + mov prsave=pr } + +#if defined(_HPUX_SOURCE) // HPUX is big-endian, cut 15+15 cycles... +{ .mib; cmp.ne p6,p0=out0,r0 + add out0=4,in0 +(p6) br.dpnt.many .Ld_i_unaligned };; + +{ .mmi; ld4 out1=[in0],8 // s0 + and out9=3,in1 + mov twenty4=24 } +{ .mmi; ld4 out3=[out0],8 // s1 + ADDP rk0=0,in2 + mov sixteen=16 };; +{ .mmi; ld4 out5=[in0] // s2 + cmp.ne p6,p0=out9,r0 + mov maskff=0xff } +{ .mmb; ld4 out7=[out0] // s3 + ADDP rk1=KSZ,in2 + br.call.sptk.many b6=_ia64_AES_decrypt };; + +{ .mib; ADDP in0=4,in1 + ADDP in1=0,in1 +(p6) br.spnt .Ld_o_unaligned };; + +{ .mii; mov ar.pfs=r2 + mov ar.lc=r3 } +{ .mmi; st4 [in1]=r16,8 // s0 + st4 [in0]=r20,8 // s1 + mov pr=prsave,0x1ffff };; +{ .mmb; st4 [in1]=r24 // s2 + st4 [in0]=r28 // s3 + br.ret.sptk.many b0 };; +#endif + +.align 32 +.Ld_i_unaligned: +{ .mmi; add out0=1,in0 + add out2=2,in0 + add out4=3,in0 };; +{ .mmi; ld1 r16=[in0],4 + ld1 r17=[out0],4 }//;; +{ .mmi; ld1 r18=[out2],4 + ld1 out1=[out4],4 };; // s0 +{ .mmi; ld1 r20=[in0],4 + ld1 r21=[out0],4 }//;; +{ .mmi; ld1 r22=[out2],4 + ld1 out3=[out4],4 };; // s1 +{ .mmi; ld1 r24=[in0],4 + ld1 r25=[out0],4 }//;; +{ .mmi; ld1 r26=[out2],4 + ld1 out5=[out4],4 };; // s2 +{ .mmi; ld1 r28=[in0] + ld1 r29=[out0] }//;; +{ .mmi; ld1 r30=[out2] + ld1 out7=[out4] };; // s3 + +{ .mii; + dep out1=r16,out1,24,8 //;; + dep out3=r20,out3,24,8 }//;; +{ .mii; ADDP rk0=0,in2 + dep out5=r24,out5,24,8 //;; + dep out7=r28,out7,24,8 };; +{ .mii; ADDP rk1=KSZ,in2 + dep out1=r17,out1,16,8 //;; + dep out3=r21,out3,16,8 }//;; +{ .mii; mov twenty4=24 + dep out5=r25,out5,16,8 //;; + dep out7=r29,out7,16,8 };; +{ .mii; mov sixteen=16 + dep out1=r18,out1,8,8 //;; + dep out3=r22,out3,8,8 }//;; +{ .mii; mov maskff=0xff + dep out5=r26,out5,8,8 //;; + dep out7=r30,out7,8,8 };; + +{ .mib; br.call.sptk.many b6=_ia64_AES_decrypt };; + +.Ld_o_unaligned: +{ .mii; ADDP out0=0,in1 + extr.u r17=r16,8,8 // s0 + shr.u r19=r16,twenty4 }//;; +{ .mii; ADDP out1=1,in1 + extr.u r18=r16,16,8 + shr.u r23=r20,twenty4 }//;; // s1 +{ .mii; ADDP out2=2,in1 + extr.u r21=r20,8,8 + shr.u r22=r20,sixteen }//;; +{ .mii; ADDP out3=3,in1 + extr.u r25=r24,8,8 // s2 + shr.u r27=r24,twenty4 };; +{ .mii; st1 [out3]=r16,4 + extr.u r26=r24,16,8 + shr.u r31=r28,twenty4 }//;; // s3 +{ .mii; st1 [out2]=r17,4 + extr.u r29=r28,8,8 + shr.u r30=r28,sixteen }//;; + +{ .mmi; st1 [out1]=r18,4 + st1 [out0]=r19,4 };; +{ .mmi; st1 [out3]=r20,4 + st1 [out2]=r21,4 }//;; +{ .mmi; st1 [out1]=r22,4 + st1 [out0]=r23,4 };; +{ .mmi; st1 [out3]=r24,4 + st1 [out2]=r25,4 + mov pr=prsave,0x1ffff }//;; +{ .mmi; st1 [out1]=r26,4 + st1 [out0]=r27,4 + mov ar.pfs=r2 };; +{ .mmi; st1 [out3]=r28 + st1 [out2]=r29 + mov ar.lc=r3 }//;; +{ .mmb; st1 [out1]=r30 + st1 [out0]=r31 + br.ret.sptk.many b0 };; +.endp AES_decrypt# + +// leave it in .text segment... +.align 64 +.global AES_Te# +.type AES_Te#,@object +AES_Te: data4 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d + data4 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554 + data4 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d + data4 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a + data4 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87 + data4 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b + data4 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea + data4 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b + data4 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a + data4 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f + data4 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108 + data4 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f + data4 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e + data4 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5 + data4 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d + data4 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f + data4 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e + data4 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb + data4 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce + data4 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497 + data4 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c + data4 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed + data4 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b + data4 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a + data4 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16 + data4 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594 + data4 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81 + data4 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3 + data4 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a + data4 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504 + data4 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163 + data4 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d + data4 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f + data4 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739 + data4 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47 + data4 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395 + data4 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f + data4 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883 + data4 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c + data4 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76 + data4 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e + data4 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4 + data4 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6 + data4 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b + data4 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7 + data4 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0 + data4 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25 + data4 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818 + data4 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72 + data4 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651 + data4 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21 + data4 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85 + data4 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa + data4 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12 + data4 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0 + data4 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9 + data4 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133 + data4 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7 + data4 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920 + data4 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a + data4 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17 + data4 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8 + data4 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11 + data4 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a +// Te1: + data4 0xa5c66363, 0x84f87c7c, 0x99ee7777, 0x8df67b7b + data4 0x0dfff2f2, 0xbdd66b6b, 0xb1de6f6f, 0x5491c5c5 + data4 0x50603030, 0x03020101, 0xa9ce6767, 0x7d562b2b + data4 0x19e7fefe, 0x62b5d7d7, 0xe64dabab, 0x9aec7676 + data4 0x458fcaca, 0x9d1f8282, 0x4089c9c9, 0x87fa7d7d + data4 0x15effafa, 0xebb25959, 0xc98e4747, 0x0bfbf0f0 + data4 0xec41adad, 0x67b3d4d4, 0xfd5fa2a2, 0xea45afaf + data4 0xbf239c9c, 0xf753a4a4, 0x96e47272, 0x5b9bc0c0 + data4 0xc275b7b7, 0x1ce1fdfd, 0xae3d9393, 0x6a4c2626 + data4 0x5a6c3636, 0x417e3f3f, 0x02f5f7f7, 0x4f83cccc + data4 0x5c683434, 0xf451a5a5, 0x34d1e5e5, 0x08f9f1f1 + data4 0x93e27171, 0x73abd8d8, 0x53623131, 0x3f2a1515 + data4 0x0c080404, 0x5295c7c7, 0x65462323, 0x5e9dc3c3 + data4 0x28301818, 0xa1379696, 0x0f0a0505, 0xb52f9a9a + data4 0x090e0707, 0x36241212, 0x9b1b8080, 0x3ddfe2e2 + data4 0x26cdebeb, 0x694e2727, 0xcd7fb2b2, 0x9fea7575 + data4 0x1b120909, 0x9e1d8383, 0x74582c2c, 0x2e341a1a + data4 0x2d361b1b, 0xb2dc6e6e, 0xeeb45a5a, 0xfb5ba0a0 + data4 0xf6a45252, 0x4d763b3b, 0x61b7d6d6, 0xce7db3b3 + data4 0x7b522929, 0x3edde3e3, 0x715e2f2f, 0x97138484 + data4 0xf5a65353, 0x68b9d1d1, 0x00000000, 0x2cc1eded + data4 0x60402020, 0x1fe3fcfc, 0xc879b1b1, 0xedb65b5b + data4 0xbed46a6a, 0x468dcbcb, 0xd967bebe, 0x4b723939 + data4 0xde944a4a, 0xd4984c4c, 0xe8b05858, 0x4a85cfcf + data4 0x6bbbd0d0, 0x2ac5efef, 0xe54faaaa, 0x16edfbfb + data4 0xc5864343, 0xd79a4d4d, 0x55663333, 0x94118585 + data4 0xcf8a4545, 0x10e9f9f9, 0x06040202, 0x81fe7f7f + data4 0xf0a05050, 0x44783c3c, 0xba259f9f, 0xe34ba8a8 + data4 0xf3a25151, 0xfe5da3a3, 0xc0804040, 0x8a058f8f + data4 0xad3f9292, 0xbc219d9d, 0x48703838, 0x04f1f5f5 + data4 0xdf63bcbc, 0xc177b6b6, 0x75afdada, 0x63422121 + data4 0x30201010, 0x1ae5ffff, 0x0efdf3f3, 0x6dbfd2d2 + data4 0x4c81cdcd, 0x14180c0c, 0x35261313, 0x2fc3ecec + data4 0xe1be5f5f, 0xa2359797, 0xcc884444, 0x392e1717 + data4 0x5793c4c4, 0xf255a7a7, 0x82fc7e7e, 0x477a3d3d + data4 0xacc86464, 0xe7ba5d5d, 0x2b321919, 0x95e67373 + data4 0xa0c06060, 0x98198181, 0xd19e4f4f, 0x7fa3dcdc + data4 0x66442222, 0x7e542a2a, 0xab3b9090, 0x830b8888 + data4 0xca8c4646, 0x29c7eeee, 0xd36bb8b8, 0x3c281414 + data4 0x79a7dede, 0xe2bc5e5e, 0x1d160b0b, 0x76addbdb + data4 0x3bdbe0e0, 0x56643232, 0x4e743a3a, 0x1e140a0a + data4 0xdb924949, 0x0a0c0606, 0x6c482424, 0xe4b85c5c + data4 0x5d9fc2c2, 0x6ebdd3d3, 0xef43acac, 0xa6c46262 + data4 0xa8399191, 0xa4319595, 0x37d3e4e4, 0x8bf27979 + data4 0x32d5e7e7, 0x438bc8c8, 0x596e3737, 0xb7da6d6d + data4 0x8c018d8d, 0x64b1d5d5, 0xd29c4e4e, 0xe049a9a9 + data4 0xb4d86c6c, 0xfaac5656, 0x07f3f4f4, 0x25cfeaea + data4 0xafca6565, 0x8ef47a7a, 0xe947aeae, 0x18100808 + data4 0xd56fbaba, 0x88f07878, 0x6f4a2525, 0x725c2e2e + data4 0x24381c1c, 0xf157a6a6, 0xc773b4b4, 0x5197c6c6 + data4 0x23cbe8e8, 0x7ca1dddd, 0x9ce87474, 0x213e1f1f + data4 0xdd964b4b, 0xdc61bdbd, 0x860d8b8b, 0x850f8a8a + data4 0x90e07070, 0x427c3e3e, 0xc471b5b5, 0xaacc6666 + data4 0xd8904848, 0x05060303, 0x01f7f6f6, 0x121c0e0e + data4 0xa3c26161, 0x5f6a3535, 0xf9ae5757, 0xd069b9b9 + data4 0x91178686, 0x5899c1c1, 0x273a1d1d, 0xb9279e9e + data4 0x38d9e1e1, 0x13ebf8f8, 0xb32b9898, 0x33221111 + data4 0xbbd26969, 0x70a9d9d9, 0x89078e8e, 0xa7339494 + data4 0xb62d9b9b, 0x223c1e1e, 0x92158787, 0x20c9e9e9 + data4 0x4987cece, 0xffaa5555, 0x78502828, 0x7aa5dfdf + data4 0x8f038c8c, 0xf859a1a1, 0x80098989, 0x171a0d0d + data4 0xda65bfbf, 0x31d7e6e6, 0xc6844242, 0xb8d06868 + data4 0xc3824141, 0xb0299999, 0x775a2d2d, 0x111e0f0f + data4 0xcb7bb0b0, 0xfca85454, 0xd66dbbbb, 0x3a2c1616 +// Te2: + data4 0x63a5c663, 0x7c84f87c, 0x7799ee77, 0x7b8df67b + data4 0xf20dfff2, 0x6bbdd66b, 0x6fb1de6f, 0xc55491c5 + data4 0x30506030, 0x01030201, 0x67a9ce67, 0x2b7d562b + data4 0xfe19e7fe, 0xd762b5d7, 0xabe64dab, 0x769aec76 + data4 0xca458fca, 0x829d1f82, 0xc94089c9, 0x7d87fa7d + data4 0xfa15effa, 0x59ebb259, 0x47c98e47, 0xf00bfbf0 + data4 0xadec41ad, 0xd467b3d4, 0xa2fd5fa2, 0xafea45af + data4 0x9cbf239c, 0xa4f753a4, 0x7296e472, 0xc05b9bc0 + data4 0xb7c275b7, 0xfd1ce1fd, 0x93ae3d93, 0x266a4c26 + data4 0x365a6c36, 0x3f417e3f, 0xf702f5f7, 0xcc4f83cc + data4 0x345c6834, 0xa5f451a5, 0xe534d1e5, 0xf108f9f1 + data4 0x7193e271, 0xd873abd8, 0x31536231, 0x153f2a15 + data4 0x040c0804, 0xc75295c7, 0x23654623, 0xc35e9dc3 + data4 0x18283018, 0x96a13796, 0x050f0a05, 0x9ab52f9a + data4 0x07090e07, 0x12362412, 0x809b1b80, 0xe23ddfe2 + data4 0xeb26cdeb, 0x27694e27, 0xb2cd7fb2, 0x759fea75 + data4 0x091b1209, 0x839e1d83, 0x2c74582c, 0x1a2e341a + data4 0x1b2d361b, 0x6eb2dc6e, 0x5aeeb45a, 0xa0fb5ba0 + data4 0x52f6a452, 0x3b4d763b, 0xd661b7d6, 0xb3ce7db3 + data4 0x297b5229, 0xe33edde3, 0x2f715e2f, 0x84971384 + data4 0x53f5a653, 0xd168b9d1, 0x00000000, 0xed2cc1ed + data4 0x20604020, 0xfc1fe3fc, 0xb1c879b1, 0x5bedb65b + data4 0x6abed46a, 0xcb468dcb, 0xbed967be, 0x394b7239 + data4 0x4ade944a, 0x4cd4984c, 0x58e8b058, 0xcf4a85cf + data4 0xd06bbbd0, 0xef2ac5ef, 0xaae54faa, 0xfb16edfb + data4 0x43c58643, 0x4dd79a4d, 0x33556633, 0x85941185 + data4 0x45cf8a45, 0xf910e9f9, 0x02060402, 0x7f81fe7f + data4 0x50f0a050, 0x3c44783c, 0x9fba259f, 0xa8e34ba8 + data4 0x51f3a251, 0xa3fe5da3, 0x40c08040, 0x8f8a058f + data4 0x92ad3f92, 0x9dbc219d, 0x38487038, 0xf504f1f5 + data4 0xbcdf63bc, 0xb6c177b6, 0xda75afda, 0x21634221 + data4 0x10302010, 0xff1ae5ff, 0xf30efdf3, 0xd26dbfd2 + data4 0xcd4c81cd, 0x0c14180c, 0x13352613, 0xec2fc3ec + data4 0x5fe1be5f, 0x97a23597, 0x44cc8844, 0x17392e17 + data4 0xc45793c4, 0xa7f255a7, 0x7e82fc7e, 0x3d477a3d + data4 0x64acc864, 0x5de7ba5d, 0x192b3219, 0x7395e673 + data4 0x60a0c060, 0x81981981, 0x4fd19e4f, 0xdc7fa3dc + data4 0x22664422, 0x2a7e542a, 0x90ab3b90, 0x88830b88 + data4 0x46ca8c46, 0xee29c7ee, 0xb8d36bb8, 0x143c2814 + data4 0xde79a7de, 0x5ee2bc5e, 0x0b1d160b, 0xdb76addb + data4 0xe03bdbe0, 0x32566432, 0x3a4e743a, 0x0a1e140a + data4 0x49db9249, 0x060a0c06, 0x246c4824, 0x5ce4b85c + data4 0xc25d9fc2, 0xd36ebdd3, 0xacef43ac, 0x62a6c462 + data4 0x91a83991, 0x95a43195, 0xe437d3e4, 0x798bf279 + data4 0xe732d5e7, 0xc8438bc8, 0x37596e37, 0x6db7da6d + data4 0x8d8c018d, 0xd564b1d5, 0x4ed29c4e, 0xa9e049a9 + data4 0x6cb4d86c, 0x56faac56, 0xf407f3f4, 0xea25cfea + data4 0x65afca65, 0x7a8ef47a, 0xaee947ae, 0x08181008 + data4 0xbad56fba, 0x7888f078, 0x256f4a25, 0x2e725c2e + data4 0x1c24381c, 0xa6f157a6, 0xb4c773b4, 0xc65197c6 + data4 0xe823cbe8, 0xdd7ca1dd, 0x749ce874, 0x1f213e1f + data4 0x4bdd964b, 0xbddc61bd, 0x8b860d8b, 0x8a850f8a + data4 0x7090e070, 0x3e427c3e, 0xb5c471b5, 0x66aacc66 + data4 0x48d89048, 0x03050603, 0xf601f7f6, 0x0e121c0e + data4 0x61a3c261, 0x355f6a35, 0x57f9ae57, 0xb9d069b9 + data4 0x86911786, 0xc15899c1, 0x1d273a1d, 0x9eb9279e + data4 0xe138d9e1, 0xf813ebf8, 0x98b32b98, 0x11332211 + data4 0x69bbd269, 0xd970a9d9, 0x8e89078e, 0x94a73394 + data4 0x9bb62d9b, 0x1e223c1e, 0x87921587, 0xe920c9e9 + data4 0xce4987ce, 0x55ffaa55, 0x28785028, 0xdf7aa5df + data4 0x8c8f038c, 0xa1f859a1, 0x89800989, 0x0d171a0d + data4 0xbfda65bf, 0xe631d7e6, 0x42c68442, 0x68b8d068 + data4 0x41c38241, 0x99b02999, 0x2d775a2d, 0x0f111e0f + data4 0xb0cb7bb0, 0x54fca854, 0xbbd66dbb, 0x163a2c16 +// Te3: + data4 0x6363a5c6, 0x7c7c84f8, 0x777799ee, 0x7b7b8df6 + data4 0xf2f20dff, 0x6b6bbdd6, 0x6f6fb1de, 0xc5c55491 + data4 0x30305060, 0x01010302, 0x6767a9ce, 0x2b2b7d56 + data4 0xfefe19e7, 0xd7d762b5, 0xababe64d, 0x76769aec + data4 0xcaca458f, 0x82829d1f, 0xc9c94089, 0x7d7d87fa + data4 0xfafa15ef, 0x5959ebb2, 0x4747c98e, 0xf0f00bfb + data4 0xadadec41, 0xd4d467b3, 0xa2a2fd5f, 0xafafea45 + data4 0x9c9cbf23, 0xa4a4f753, 0x727296e4, 0xc0c05b9b + data4 0xb7b7c275, 0xfdfd1ce1, 0x9393ae3d, 0x26266a4c + data4 0x36365a6c, 0x3f3f417e, 0xf7f702f5, 0xcccc4f83 + data4 0x34345c68, 0xa5a5f451, 0xe5e534d1, 0xf1f108f9 + data4 0x717193e2, 0xd8d873ab, 0x31315362, 0x15153f2a + data4 0x04040c08, 0xc7c75295, 0x23236546, 0xc3c35e9d + data4 0x18182830, 0x9696a137, 0x05050f0a, 0x9a9ab52f + data4 0x0707090e, 0x12123624, 0x80809b1b, 0xe2e23ddf + data4 0xebeb26cd, 0x2727694e, 0xb2b2cd7f, 0x75759fea + data4 0x09091b12, 0x83839e1d, 0x2c2c7458, 0x1a1a2e34 + data4 0x1b1b2d36, 0x6e6eb2dc, 0x5a5aeeb4, 0xa0a0fb5b + data4 0x5252f6a4, 0x3b3b4d76, 0xd6d661b7, 0xb3b3ce7d + data4 0x29297b52, 0xe3e33edd, 0x2f2f715e, 0x84849713 + data4 0x5353f5a6, 0xd1d168b9, 0x00000000, 0xeded2cc1 + data4 0x20206040, 0xfcfc1fe3, 0xb1b1c879, 0x5b5bedb6 + data4 0x6a6abed4, 0xcbcb468d, 0xbebed967, 0x39394b72 + data4 0x4a4ade94, 0x4c4cd498, 0x5858e8b0, 0xcfcf4a85 + data4 0xd0d06bbb, 0xefef2ac5, 0xaaaae54f, 0xfbfb16ed + data4 0x4343c586, 0x4d4dd79a, 0x33335566, 0x85859411 + data4 0x4545cf8a, 0xf9f910e9, 0x02020604, 0x7f7f81fe + data4 0x5050f0a0, 0x3c3c4478, 0x9f9fba25, 0xa8a8e34b + data4 0x5151f3a2, 0xa3a3fe5d, 0x4040c080, 0x8f8f8a05 + data4 0x9292ad3f, 0x9d9dbc21, 0x38384870, 0xf5f504f1 + data4 0xbcbcdf63, 0xb6b6c177, 0xdada75af, 0x21216342 + data4 0x10103020, 0xffff1ae5, 0xf3f30efd, 0xd2d26dbf + data4 0xcdcd4c81, 0x0c0c1418, 0x13133526, 0xecec2fc3 + data4 0x5f5fe1be, 0x9797a235, 0x4444cc88, 0x1717392e + data4 0xc4c45793, 0xa7a7f255, 0x7e7e82fc, 0x3d3d477a + data4 0x6464acc8, 0x5d5de7ba, 0x19192b32, 0x737395e6 + data4 0x6060a0c0, 0x81819819, 0x4f4fd19e, 0xdcdc7fa3 + data4 0x22226644, 0x2a2a7e54, 0x9090ab3b, 0x8888830b + data4 0x4646ca8c, 0xeeee29c7, 0xb8b8d36b, 0x14143c28 + data4 0xdede79a7, 0x5e5ee2bc, 0x0b0b1d16, 0xdbdb76ad + data4 0xe0e03bdb, 0x32325664, 0x3a3a4e74, 0x0a0a1e14 + data4 0x4949db92, 0x06060a0c, 0x24246c48, 0x5c5ce4b8 + data4 0xc2c25d9f, 0xd3d36ebd, 0xacacef43, 0x6262a6c4 + data4 0x9191a839, 0x9595a431, 0xe4e437d3, 0x79798bf2 + data4 0xe7e732d5, 0xc8c8438b, 0x3737596e, 0x6d6db7da + data4 0x8d8d8c01, 0xd5d564b1, 0x4e4ed29c, 0xa9a9e049 + data4 0x6c6cb4d8, 0x5656faac, 0xf4f407f3, 0xeaea25cf + data4 0x6565afca, 0x7a7a8ef4, 0xaeaee947, 0x08081810 + data4 0xbabad56f, 0x787888f0, 0x25256f4a, 0x2e2e725c + data4 0x1c1c2438, 0xa6a6f157, 0xb4b4c773, 0xc6c65197 + data4 0xe8e823cb, 0xdddd7ca1, 0x74749ce8, 0x1f1f213e + data4 0x4b4bdd96, 0xbdbddc61, 0x8b8b860d, 0x8a8a850f + data4 0x707090e0, 0x3e3e427c, 0xb5b5c471, 0x6666aacc + data4 0x4848d890, 0x03030506, 0xf6f601f7, 0x0e0e121c + data4 0x6161a3c2, 0x35355f6a, 0x5757f9ae, 0xb9b9d069 + data4 0x86869117, 0xc1c15899, 0x1d1d273a, 0x9e9eb927 + data4 0xe1e138d9, 0xf8f813eb, 0x9898b32b, 0x11113322 + data4 0x6969bbd2, 0xd9d970a9, 0x8e8e8907, 0x9494a733 + data4 0x9b9bb62d, 0x1e1e223c, 0x87879215, 0xe9e920c9 + data4 0xcece4987, 0x5555ffaa, 0x28287850, 0xdfdf7aa5 + data4 0x8c8c8f03, 0xa1a1f859, 0x89898009, 0x0d0d171a + data4 0xbfbfda65, 0xe6e631d7, 0x4242c684, 0x6868b8d0 + data4 0x4141c382, 0x9999b029, 0x2d2d775a, 0x0f0f111e + data4 0xb0b0cb7b, 0x5454fca8, 0xbbbbd66d, 0x16163a2c +// Te4: + data4 0x63000000, 0x7c000000, 0x77000000, 0x7b000000 + data4 0xf2000000, 0x6b000000, 0x6f000000, 0xc5000000 + data4 0x30000000, 0x01000000, 0x67000000, 0x2b000000 + data4 0xfe000000, 0xd7000000, 0xab000000, 0x76000000 + data4 0xca000000, 0x82000000, 0xc9000000, 0x7d000000 + data4 0xfa000000, 0x59000000, 0x47000000, 0xf0000000 + data4 0xad000000, 0xd4000000, 0xa2000000, 0xaf000000 + data4 0x9c000000, 0xa4000000, 0x72000000, 0xc0000000 + data4 0xb7000000, 0xfd000000, 0x93000000, 0x26000000 + data4 0x36000000, 0x3f000000, 0xf7000000, 0xcc000000 + data4 0x34000000, 0xa5000000, 0xe5000000, 0xf1000000 + data4 0x71000000, 0xd8000000, 0x31000000, 0x15000000 + data4 0x04000000, 0xc7000000, 0x23000000, 0xc3000000 + data4 0x18000000, 0x96000000, 0x05000000, 0x9a000000 + data4 0x07000000, 0x12000000, 0x80000000, 0xe2000000 + data4 0xeb000000, 0x27000000, 0xb2000000, 0x75000000 + data4 0x09000000, 0x83000000, 0x2c000000, 0x1a000000 + data4 0x1b000000, 0x6e000000, 0x5a000000, 0xa0000000 + data4 0x52000000, 0x3b000000, 0xd6000000, 0xb3000000 + data4 0x29000000, 0xe3000000, 0x2f000000, 0x84000000 + data4 0x53000000, 0xd1000000, 0x00000000, 0xed000000 + data4 0x20000000, 0xfc000000, 0xb1000000, 0x5b000000 + data4 0x6a000000, 0xcb000000, 0xbe000000, 0x39000000 + data4 0x4a000000, 0x4c000000, 0x58000000, 0xcf000000 + data4 0xd0000000, 0xef000000, 0xaa000000, 0xfb000000 + data4 0x43000000, 0x4d000000, 0x33000000, 0x85000000 + data4 0x45000000, 0xf9000000, 0x02000000, 0x7f000000 + data4 0x50000000, 0x3c000000, 0x9f000000, 0xa8000000 + data4 0x51000000, 0xa3000000, 0x40000000, 0x8f000000 + data4 0x92000000, 0x9d000000, 0x38000000, 0xf5000000 + data4 0xbc000000, 0xb6000000, 0xda000000, 0x21000000 + data4 0x10000000, 0xff000000, 0xf3000000, 0xd2000000 + data4 0xcd000000, 0x0c000000, 0x13000000, 0xec000000 + data4 0x5f000000, 0x97000000, 0x44000000, 0x17000000 + data4 0xc4000000, 0xa7000000, 0x7e000000, 0x3d000000 + data4 0x64000000, 0x5d000000, 0x19000000, 0x73000000 + data4 0x60000000, 0x81000000, 0x4f000000, 0xdc000000 + data4 0x22000000, 0x2a000000, 0x90000000, 0x88000000 + data4 0x46000000, 0xee000000, 0xb8000000, 0x14000000 + data4 0xde000000, 0x5e000000, 0x0b000000, 0xdb000000 + data4 0xe0000000, 0x32000000, 0x3a000000, 0x0a000000 + data4 0x49000000, 0x06000000, 0x24000000, 0x5c000000 + data4 0xc2000000, 0xd3000000, 0xac000000, 0x62000000 + data4 0x91000000, 0x95000000, 0xe4000000, 0x79000000 + data4 0xe7000000, 0xc8000000, 0x37000000, 0x6d000000 + data4 0x8d000000, 0xd5000000, 0x4e000000, 0xa9000000 + data4 0x6c000000, 0x56000000, 0xf4000000, 0xea000000 + data4 0x65000000, 0x7a000000, 0xae000000, 0x08000000 + data4 0xba000000, 0x78000000, 0x25000000, 0x2e000000 + data4 0x1c000000, 0xa6000000, 0xb4000000, 0xc6000000 + data4 0xe8000000, 0xdd000000, 0x74000000, 0x1f000000 + data4 0x4b000000, 0xbd000000, 0x8b000000, 0x8a000000 + data4 0x70000000, 0x3e000000, 0xb5000000, 0x66000000 + data4 0x48000000, 0x03000000, 0xf6000000, 0x0e000000 + data4 0x61000000, 0x35000000, 0x57000000, 0xb9000000 + data4 0x86000000, 0xc1000000, 0x1d000000, 0x9e000000 + data4 0xe1000000, 0xf8000000, 0x98000000, 0x11000000 + data4 0x69000000, 0xd9000000, 0x8e000000, 0x94000000 + data4 0x9b000000, 0x1e000000, 0x87000000, 0xe9000000 + data4 0xce000000, 0x55000000, 0x28000000, 0xdf000000 + data4 0x8c000000, 0xa1000000, 0x89000000, 0x0d000000 + data4 0xbf000000, 0xe6000000, 0x42000000, 0x68000000 + data4 0x41000000, 0x99000000, 0x2d000000, 0x0f000000 + data4 0xb0000000, 0x54000000, 0xbb000000, 0x16000000 +// Te5: + data4 0x00630000, 0x007c0000, 0x00770000, 0x007b0000 + data4 0x00f20000, 0x006b0000, 0x006f0000, 0x00c50000 + data4 0x00300000, 0x00010000, 0x00670000, 0x002b0000 + data4 0x00fe0000, 0x00d70000, 0x00ab0000, 0x00760000 + data4 0x00ca0000, 0x00820000, 0x00c90000, 0x007d0000 + data4 0x00fa0000, 0x00590000, 0x00470000, 0x00f00000 + data4 0x00ad0000, 0x00d40000, 0x00a20000, 0x00af0000 + data4 0x009c0000, 0x00a40000, 0x00720000, 0x00c00000 + data4 0x00b70000, 0x00fd0000, 0x00930000, 0x00260000 + data4 0x00360000, 0x003f0000, 0x00f70000, 0x00cc0000 + data4 0x00340000, 0x00a50000, 0x00e50000, 0x00f10000 + data4 0x00710000, 0x00d80000, 0x00310000, 0x00150000 + data4 0x00040000, 0x00c70000, 0x00230000, 0x00c30000 + data4 0x00180000, 0x00960000, 0x00050000, 0x009a0000 + data4 0x00070000, 0x00120000, 0x00800000, 0x00e20000 + data4 0x00eb0000, 0x00270000, 0x00b20000, 0x00750000 + data4 0x00090000, 0x00830000, 0x002c0000, 0x001a0000 + data4 0x001b0000, 0x006e0000, 0x005a0000, 0x00a00000 + data4 0x00520000, 0x003b0000, 0x00d60000, 0x00b30000 + data4 0x00290000, 0x00e30000, 0x002f0000, 0x00840000 + data4 0x00530000, 0x00d10000, 0x00000000, 0x00ed0000 + data4 0x00200000, 0x00fc0000, 0x00b10000, 0x005b0000 + data4 0x006a0000, 0x00cb0000, 0x00be0000, 0x00390000 + data4 0x004a0000, 0x004c0000, 0x00580000, 0x00cf0000 + data4 0x00d00000, 0x00ef0000, 0x00aa0000, 0x00fb0000 + data4 0x00430000, 0x004d0000, 0x00330000, 0x00850000 + data4 0x00450000, 0x00f90000, 0x00020000, 0x007f0000 + data4 0x00500000, 0x003c0000, 0x009f0000, 0x00a80000 + data4 0x00510000, 0x00a30000, 0x00400000, 0x008f0000 + data4 0x00920000, 0x009d0000, 0x00380000, 0x00f50000 + data4 0x00bc0000, 0x00b60000, 0x00da0000, 0x00210000 + data4 0x00100000, 0x00ff0000, 0x00f30000, 0x00d20000 + data4 0x00cd0000, 0x000c0000, 0x00130000, 0x00ec0000 + data4 0x005f0000, 0x00970000, 0x00440000, 0x00170000 + data4 0x00c40000, 0x00a70000, 0x007e0000, 0x003d0000 + data4 0x00640000, 0x005d0000, 0x00190000, 0x00730000 + data4 0x00600000, 0x00810000, 0x004f0000, 0x00dc0000 + data4 0x00220000, 0x002a0000, 0x00900000, 0x00880000 + data4 0x00460000, 0x00ee0000, 0x00b80000, 0x00140000 + data4 0x00de0000, 0x005e0000, 0x000b0000, 0x00db0000 + data4 0x00e00000, 0x00320000, 0x003a0000, 0x000a0000 + data4 0x00490000, 0x00060000, 0x00240000, 0x005c0000 + data4 0x00c20000, 0x00d30000, 0x00ac0000, 0x00620000 + data4 0x00910000, 0x00950000, 0x00e40000, 0x00790000 + data4 0x00e70000, 0x00c80000, 0x00370000, 0x006d0000 + data4 0x008d0000, 0x00d50000, 0x004e0000, 0x00a90000 + data4 0x006c0000, 0x00560000, 0x00f40000, 0x00ea0000 + data4 0x00650000, 0x007a0000, 0x00ae0000, 0x00080000 + data4 0x00ba0000, 0x00780000, 0x00250000, 0x002e0000 + data4 0x001c0000, 0x00a60000, 0x00b40000, 0x00c60000 + data4 0x00e80000, 0x00dd0000, 0x00740000, 0x001f0000 + data4 0x004b0000, 0x00bd0000, 0x008b0000, 0x008a0000 + data4 0x00700000, 0x003e0000, 0x00b50000, 0x00660000 + data4 0x00480000, 0x00030000, 0x00f60000, 0x000e0000 + data4 0x00610000, 0x00350000, 0x00570000, 0x00b90000 + data4 0x00860000, 0x00c10000, 0x001d0000, 0x009e0000 + data4 0x00e10000, 0x00f80000, 0x00980000, 0x00110000 + data4 0x00690000, 0x00d90000, 0x008e0000, 0x00940000 + data4 0x009b0000, 0x001e0000, 0x00870000, 0x00e90000 + data4 0x00ce0000, 0x00550000, 0x00280000, 0x00df0000 + data4 0x008c0000, 0x00a10000, 0x00890000, 0x000d0000 + data4 0x00bf0000, 0x00e60000, 0x00420000, 0x00680000 + data4 0x00410000, 0x00990000, 0x002d0000, 0x000f0000 + data4 0x00b00000, 0x00540000, 0x00bb0000, 0x00160000 +// Te6: + data4 0x00006300, 0x00007c00, 0x00007700, 0x00007b00 + data4 0x0000f200, 0x00006b00, 0x00006f00, 0x0000c500 + data4 0x00003000, 0x00000100, 0x00006700, 0x00002b00 + data4 0x0000fe00, 0x0000d700, 0x0000ab00, 0x00007600 + data4 0x0000ca00, 0x00008200, 0x0000c900, 0x00007d00 + data4 0x0000fa00, 0x00005900, 0x00004700, 0x0000f000 + data4 0x0000ad00, 0x0000d400, 0x0000a200, 0x0000af00 + data4 0x00009c00, 0x0000a400, 0x00007200, 0x0000c000 + data4 0x0000b700, 0x0000fd00, 0x00009300, 0x00002600 + data4 0x00003600, 0x00003f00, 0x0000f700, 0x0000cc00 + data4 0x00003400, 0x0000a500, 0x0000e500, 0x0000f100 + data4 0x00007100, 0x0000d800, 0x00003100, 0x00001500 + data4 0x00000400, 0x0000c700, 0x00002300, 0x0000c300 + data4 0x00001800, 0x00009600, 0x00000500, 0x00009a00 + data4 0x00000700, 0x00001200, 0x00008000, 0x0000e200 + data4 0x0000eb00, 0x00002700, 0x0000b200, 0x00007500 + data4 0x00000900, 0x00008300, 0x00002c00, 0x00001a00 + data4 0x00001b00, 0x00006e00, 0x00005a00, 0x0000a000 + data4 0x00005200, 0x00003b00, 0x0000d600, 0x0000b300 + data4 0x00002900, 0x0000e300, 0x00002f00, 0x00008400 + data4 0x00005300, 0x0000d100, 0x00000000, 0x0000ed00 + data4 0x00002000, 0x0000fc00, 0x0000b100, 0x00005b00 + data4 0x00006a00, 0x0000cb00, 0x0000be00, 0x00003900 + data4 0x00004a00, 0x00004c00, 0x00005800, 0x0000cf00 + data4 0x0000d000, 0x0000ef00, 0x0000aa00, 0x0000fb00 + data4 0x00004300, 0x00004d00, 0x00003300, 0x00008500 + data4 0x00004500, 0x0000f900, 0x00000200, 0x00007f00 + data4 0x00005000, 0x00003c00, 0x00009f00, 0x0000a800 + data4 0x00005100, 0x0000a300, 0x00004000, 0x00008f00 + data4 0x00009200, 0x00009d00, 0x00003800, 0x0000f500 + data4 0x0000bc00, 0x0000b600, 0x0000da00, 0x00002100 + data4 0x00001000, 0x0000ff00, 0x0000f300, 0x0000d200 + data4 0x0000cd00, 0x00000c00, 0x00001300, 0x0000ec00 + data4 0x00005f00, 0x00009700, 0x00004400, 0x00001700 + data4 0x0000c400, 0x0000a700, 0x00007e00, 0x00003d00 + data4 0x00006400, 0x00005d00, 0x00001900, 0x00007300 + data4 0x00006000, 0x00008100, 0x00004f00, 0x0000dc00 + data4 0x00002200, 0x00002a00, 0x00009000, 0x00008800 + data4 0x00004600, 0x0000ee00, 0x0000b800, 0x00001400 + data4 0x0000de00, 0x00005e00, 0x00000b00, 0x0000db00 + data4 0x0000e000, 0x00003200, 0x00003a00, 0x00000a00 + data4 0x00004900, 0x00000600, 0x00002400, 0x00005c00 + data4 0x0000c200, 0x0000d300, 0x0000ac00, 0x00006200 + data4 0x00009100, 0x00009500, 0x0000e400, 0x00007900 + data4 0x0000e700, 0x0000c800, 0x00003700, 0x00006d00 + data4 0x00008d00, 0x0000d500, 0x00004e00, 0x0000a900 + data4 0x00006c00, 0x00005600, 0x0000f400, 0x0000ea00 + data4 0x00006500, 0x00007a00, 0x0000ae00, 0x00000800 + data4 0x0000ba00, 0x00007800, 0x00002500, 0x00002e00 + data4 0x00001c00, 0x0000a600, 0x0000b400, 0x0000c600 + data4 0x0000e800, 0x0000dd00, 0x00007400, 0x00001f00 + data4 0x00004b00, 0x0000bd00, 0x00008b00, 0x00008a00 + data4 0x00007000, 0x00003e00, 0x0000b500, 0x00006600 + data4 0x00004800, 0x00000300, 0x0000f600, 0x00000e00 + data4 0x00006100, 0x00003500, 0x00005700, 0x0000b900 + data4 0x00008600, 0x0000c100, 0x00001d00, 0x00009e00 + data4 0x0000e100, 0x0000f800, 0x00009800, 0x00001100 + data4 0x00006900, 0x0000d900, 0x00008e00, 0x00009400 + data4 0x00009b00, 0x00001e00, 0x00008700, 0x0000e900 + data4 0x0000ce00, 0x00005500, 0x00002800, 0x0000df00 + data4 0x00008c00, 0x0000a100, 0x00008900, 0x00000d00 + data4 0x0000bf00, 0x0000e600, 0x00004200, 0x00006800 + data4 0x00004100, 0x00009900, 0x00002d00, 0x00000f00 + data4 0x0000b000, 0x00005400, 0x0000bb00, 0x00001600 +// Te7: + data4 0x00000063, 0x0000007c, 0x00000077, 0x0000007b + data4 0x000000f2, 0x0000006b, 0x0000006f, 0x000000c5 + data4 0x00000030, 0x00000001, 0x00000067, 0x0000002b + data4 0x000000fe, 0x000000d7, 0x000000ab, 0x00000076 + data4 0x000000ca, 0x00000082, 0x000000c9, 0x0000007d + data4 0x000000fa, 0x00000059, 0x00000047, 0x000000f0 + data4 0x000000ad, 0x000000d4, 0x000000a2, 0x000000af + data4 0x0000009c, 0x000000a4, 0x00000072, 0x000000c0 + data4 0x000000b7, 0x000000fd, 0x00000093, 0x00000026 + data4 0x00000036, 0x0000003f, 0x000000f7, 0x000000cc + data4 0x00000034, 0x000000a5, 0x000000e5, 0x000000f1 + data4 0x00000071, 0x000000d8, 0x00000031, 0x00000015 + data4 0x00000004, 0x000000c7, 0x00000023, 0x000000c3 + data4 0x00000018, 0x00000096, 0x00000005, 0x0000009a + data4 0x00000007, 0x00000012, 0x00000080, 0x000000e2 + data4 0x000000eb, 0x00000027, 0x000000b2, 0x00000075 + data4 0x00000009, 0x00000083, 0x0000002c, 0x0000001a + data4 0x0000001b, 0x0000006e, 0x0000005a, 0x000000a0 + data4 0x00000052, 0x0000003b, 0x000000d6, 0x000000b3 + data4 0x00000029, 0x000000e3, 0x0000002f, 0x00000084 + data4 0x00000053, 0x000000d1, 0x00000000, 0x000000ed + data4 0x00000020, 0x000000fc, 0x000000b1, 0x0000005b + data4 0x0000006a, 0x000000cb, 0x000000be, 0x00000039 + data4 0x0000004a, 0x0000004c, 0x00000058, 0x000000cf + data4 0x000000d0, 0x000000ef, 0x000000aa, 0x000000fb + data4 0x00000043, 0x0000004d, 0x00000033, 0x00000085 + data4 0x00000045, 0x000000f9, 0x00000002, 0x0000007f + data4 0x00000050, 0x0000003c, 0x0000009f, 0x000000a8 + data4 0x00000051, 0x000000a3, 0x00000040, 0x0000008f + data4 0x00000092, 0x0000009d, 0x00000038, 0x000000f5 + data4 0x000000bc, 0x000000b6, 0x000000da, 0x00000021 + data4 0x00000010, 0x000000ff, 0x000000f3, 0x000000d2 + data4 0x000000cd, 0x0000000c, 0x00000013, 0x000000ec + data4 0x0000005f, 0x00000097, 0x00000044, 0x00000017 + data4 0x000000c4, 0x000000a7, 0x0000007e, 0x0000003d + data4 0x00000064, 0x0000005d, 0x00000019, 0x00000073 + data4 0x00000060, 0x00000081, 0x0000004f, 0x000000dc + data4 0x00000022, 0x0000002a, 0x00000090, 0x00000088 + data4 0x00000046, 0x000000ee, 0x000000b8, 0x00000014 + data4 0x000000de, 0x0000005e, 0x0000000b, 0x000000db + data4 0x000000e0, 0x00000032, 0x0000003a, 0x0000000a + data4 0x00000049, 0x00000006, 0x00000024, 0x0000005c + data4 0x000000c2, 0x000000d3, 0x000000ac, 0x00000062 + data4 0x00000091, 0x00000095, 0x000000e4, 0x00000079 + data4 0x000000e7, 0x000000c8, 0x00000037, 0x0000006d + data4 0x0000008d, 0x000000d5, 0x0000004e, 0x000000a9 + data4 0x0000006c, 0x00000056, 0x000000f4, 0x000000ea + data4 0x00000065, 0x0000007a, 0x000000ae, 0x00000008 + data4 0x000000ba, 0x00000078, 0x00000025, 0x0000002e + data4 0x0000001c, 0x000000a6, 0x000000b4, 0x000000c6 + data4 0x000000e8, 0x000000dd, 0x00000074, 0x0000001f + data4 0x0000004b, 0x000000bd, 0x0000008b, 0x0000008a + data4 0x00000070, 0x0000003e, 0x000000b5, 0x00000066 + data4 0x00000048, 0x00000003, 0x000000f6, 0x0000000e + data4 0x00000061, 0x00000035, 0x00000057, 0x000000b9 + data4 0x00000086, 0x000000c1, 0x0000001d, 0x0000009e + data4 0x000000e1, 0x000000f8, 0x00000098, 0x00000011 + data4 0x00000069, 0x000000d9, 0x0000008e, 0x00000094 + data4 0x0000009b, 0x0000001e, 0x00000087, 0x000000e9 + data4 0x000000ce, 0x00000055, 0x00000028, 0x000000df + data4 0x0000008c, 0x000000a1, 0x00000089, 0x0000000d + data4 0x000000bf, 0x000000e6, 0x00000042, 0x00000068 + data4 0x00000041, 0x00000099, 0x0000002d, 0x0000000f + data4 0x000000b0, 0x00000054, 0x000000bb, 0x00000016 +.size AES_Te#,8*256*4 // HP-UX assembler fails to ".-AES_Te#" + +.align 64 +.global AES_Td# +.type AES_Td#,@object +AES_Td: data4 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96 + data4 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393 + data4 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25 + data4 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f + data4 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1 + data4 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6 + data4 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da + data4 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844 + data4 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd + data4 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4 + data4 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45 + data4 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94 + data4 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7 + data4 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a + data4 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5 + data4 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c + data4 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1 + data4 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a + data4 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75 + data4 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051 + data4 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46 + data4 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff + data4 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77 + data4 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb + data4 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000 + data4 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e + data4 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927 + data4 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a + data4 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e + data4 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16 + data4 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d + data4 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8 + data4 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd + data4 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34 + data4 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163 + data4 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120 + data4 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d + data4 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0 + data4 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422 + data4 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef + data4 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36 + data4 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4 + data4 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662 + data4 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5 + data4 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3 + data4 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b + data4 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8 + data4 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6 + data4 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6 + data4 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0 + data4 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815 + data4 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f + data4 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df + data4 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f + data4 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e + data4 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713 + data4 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89 + data4 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c + data4 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf + data4 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86 + data4 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f + data4 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541 + data4 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190 + data4 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742 +// Td1: + data4 0x5051f4a7, 0x537e4165, 0xc31a17a4, 0x963a275e + data4 0xcb3bab6b, 0xf11f9d45, 0xabacfa58, 0x934be303 + data4 0x552030fa, 0xf6ad766d, 0x9188cc76, 0x25f5024c + data4 0xfc4fe5d7, 0xd7c52acb, 0x80263544, 0x8fb562a3 + data4 0x49deb15a, 0x6725ba1b, 0x9845ea0e, 0xe15dfec0 + data4 0x02c32f75, 0x12814cf0, 0xa38d4697, 0xc66bd3f9 + data4 0xe7038f5f, 0x9515929c, 0xebbf6d7a, 0xda955259 + data4 0x2dd4be83, 0xd3587421, 0x2949e069, 0x448ec9c8 + data4 0x6a75c289, 0x78f48e79, 0x6b99583e, 0xdd27b971 + data4 0xb6bee14f, 0x17f088ad, 0x66c920ac, 0xb47dce3a + data4 0x1863df4a, 0x82e51a31, 0x60975133, 0x4562537f + data4 0xe0b16477, 0x84bb6bae, 0x1cfe81a0, 0x94f9082b + data4 0x58704868, 0x198f45fd, 0x8794de6c, 0xb7527bf8 + data4 0x23ab73d3, 0xe2724b02, 0x57e31f8f, 0x2a6655ab + data4 0x07b2eb28, 0x032fb5c2, 0x9a86c57b, 0xa5d33708 + data4 0xf2302887, 0xb223bfa5, 0xba02036a, 0x5ced1682 + data4 0x2b8acf1c, 0x92a779b4, 0xf0f307f2, 0xa14e69e2 + data4 0xcd65daf4, 0xd50605be, 0x1fd13462, 0x8ac4a6fe + data4 0x9d342e53, 0xa0a2f355, 0x32058ae1, 0x75a4f6eb + data4 0x390b83ec, 0xaa4060ef, 0x065e719f, 0x51bd6e10 + data4 0xf93e218a, 0x3d96dd06, 0xaedd3e05, 0x464de6bd + data4 0xb591548d, 0x0571c45d, 0x6f0406d4, 0xff605015 + data4 0x241998fb, 0x97d6bde9, 0xcc894043, 0x7767d99e + data4 0xbdb0e842, 0x8807898b, 0x38e7195b, 0xdb79c8ee + data4 0x47a17c0a, 0xe97c420f, 0xc9f8841e, 0x00000000 + data4 0x83098086, 0x48322bed, 0xac1e1170, 0x4e6c5a72 + data4 0xfbfd0eff, 0x560f8538, 0x1e3daed5, 0x27362d39 + data4 0x640a0fd9, 0x21685ca6, 0xd19b5b54, 0x3a24362e + data4 0xb10c0a67, 0x0f9357e7, 0xd2b4ee96, 0x9e1b9b91 + data4 0x4f80c0c5, 0xa261dc20, 0x695a774b, 0x161c121a + data4 0x0ae293ba, 0xe5c0a02a, 0x433c22e0, 0x1d121b17 + data4 0x0b0e090d, 0xadf28bc7, 0xb92db6a8, 0xc8141ea9 + data4 0x8557f119, 0x4caf7507, 0xbbee99dd, 0xfda37f60 + data4 0x9ff70126, 0xbc5c72f5, 0xc544663b, 0x345bfb7e + data4 0x768b4329, 0xdccb23c6, 0x68b6edfc, 0x63b8e4f1 + data4 0xcad731dc, 0x10426385, 0x40139722, 0x2084c611 + data4 0x7d854a24, 0xf8d2bb3d, 0x11aef932, 0x6dc729a1 + data4 0x4b1d9e2f, 0xf3dcb230, 0xec0d8652, 0xd077c1e3 + data4 0x6c2bb316, 0x99a970b9, 0xfa119448, 0x2247e964 + data4 0xc4a8fc8c, 0x1aa0f03f, 0xd8567d2c, 0xef223390 + data4 0xc787494e, 0xc1d938d1, 0xfe8ccaa2, 0x3698d40b + data4 0xcfa6f581, 0x28a57ade, 0x26dab78e, 0xa43fadbf + data4 0xe42c3a9d, 0x0d507892, 0x9b6a5fcc, 0x62547e46 + data4 0xc2f68d13, 0xe890d8b8, 0x5e2e39f7, 0xf582c3af + data4 0xbe9f5d80, 0x7c69d093, 0xa96fd52d, 0xb3cf2512 + data4 0x3bc8ac99, 0xa710187d, 0x6ee89c63, 0x7bdb3bbb + data4 0x09cd2678, 0xf46e5918, 0x01ec9ab7, 0xa8834f9a + data4 0x65e6956e, 0x7eaaffe6, 0x0821bccf, 0xe6ef15e8 + data4 0xd9bae79b, 0xce4a6f36, 0xd4ea9f09, 0xd629b07c + data4 0xaf31a4b2, 0x312a3f23, 0x30c6a594, 0xc035a266 + data4 0x37744ebc, 0xa6fc82ca, 0xb0e090d0, 0x1533a7d8 + data4 0x4af10498, 0xf741ecda, 0x0e7fcd50, 0x2f1791f6 + data4 0x8d764dd6, 0x4d43efb0, 0x54ccaa4d, 0xdfe49604 + data4 0xe39ed1b5, 0x1b4c6a88, 0xb8c12c1f, 0x7f466551 + data4 0x049d5eea, 0x5d018c35, 0x73fa8774, 0x2efb0b41 + data4 0x5ab3671d, 0x5292dbd2, 0x33e91056, 0x136dd647 + data4 0x8c9ad761, 0x7a37a10c, 0x8e59f814, 0x89eb133c + data4 0xeecea927, 0x35b761c9, 0xede11ce5, 0x3c7a47b1 + data4 0x599cd2df, 0x3f55f273, 0x791814ce, 0xbf73c737 + data4 0xea53f7cd, 0x5b5ffdaa, 0x14df3d6f, 0x867844db + data4 0x81caaff3, 0x3eb968c4, 0x2c382434, 0x5fc2a340 + data4 0x72161dc3, 0x0cbce225, 0x8b283c49, 0x41ff0d95 + data4 0x7139a801, 0xde080cb3, 0x9cd8b4e4, 0x906456c1 + data4 0x617bcb84, 0x70d532b6, 0x74486c5c, 0x42d0b857 +// Td2: + data4 0xa75051f4, 0x65537e41, 0xa4c31a17, 0x5e963a27 + data4 0x6bcb3bab, 0x45f11f9d, 0x58abacfa, 0x03934be3 + data4 0xfa552030, 0x6df6ad76, 0x769188cc, 0x4c25f502 + data4 0xd7fc4fe5, 0xcbd7c52a, 0x44802635, 0xa38fb562 + data4 0x5a49deb1, 0x1b6725ba, 0x0e9845ea, 0xc0e15dfe + data4 0x7502c32f, 0xf012814c, 0x97a38d46, 0xf9c66bd3 + data4 0x5fe7038f, 0x9c951592, 0x7aebbf6d, 0x59da9552 + data4 0x832dd4be, 0x21d35874, 0x692949e0, 0xc8448ec9 + data4 0x896a75c2, 0x7978f48e, 0x3e6b9958, 0x71dd27b9 + data4 0x4fb6bee1, 0xad17f088, 0xac66c920, 0x3ab47dce + data4 0x4a1863df, 0x3182e51a, 0x33609751, 0x7f456253 + data4 0x77e0b164, 0xae84bb6b, 0xa01cfe81, 0x2b94f908 + data4 0x68587048, 0xfd198f45, 0x6c8794de, 0xf8b7527b + data4 0xd323ab73, 0x02e2724b, 0x8f57e31f, 0xab2a6655 + data4 0x2807b2eb, 0xc2032fb5, 0x7b9a86c5, 0x08a5d337 + data4 0x87f23028, 0xa5b223bf, 0x6aba0203, 0x825ced16 + data4 0x1c2b8acf, 0xb492a779, 0xf2f0f307, 0xe2a14e69 + data4 0xf4cd65da, 0xbed50605, 0x621fd134, 0xfe8ac4a6 + data4 0x539d342e, 0x55a0a2f3, 0xe132058a, 0xeb75a4f6 + data4 0xec390b83, 0xefaa4060, 0x9f065e71, 0x1051bd6e + data4 0x8af93e21, 0x063d96dd, 0x05aedd3e, 0xbd464de6 + data4 0x8db59154, 0x5d0571c4, 0xd46f0406, 0x15ff6050 + data4 0xfb241998, 0xe997d6bd, 0x43cc8940, 0x9e7767d9 + data4 0x42bdb0e8, 0x8b880789, 0x5b38e719, 0xeedb79c8 + data4 0x0a47a17c, 0x0fe97c42, 0x1ec9f884, 0x00000000 + data4 0x86830980, 0xed48322b, 0x70ac1e11, 0x724e6c5a + data4 0xfffbfd0e, 0x38560f85, 0xd51e3dae, 0x3927362d + data4 0xd9640a0f, 0xa621685c, 0x54d19b5b, 0x2e3a2436 + data4 0x67b10c0a, 0xe70f9357, 0x96d2b4ee, 0x919e1b9b + data4 0xc54f80c0, 0x20a261dc, 0x4b695a77, 0x1a161c12 + data4 0xba0ae293, 0x2ae5c0a0, 0xe0433c22, 0x171d121b + data4 0x0d0b0e09, 0xc7adf28b, 0xa8b92db6, 0xa9c8141e + data4 0x198557f1, 0x074caf75, 0xddbbee99, 0x60fda37f + data4 0x269ff701, 0xf5bc5c72, 0x3bc54466, 0x7e345bfb + data4 0x29768b43, 0xc6dccb23, 0xfc68b6ed, 0xf163b8e4 + data4 0xdccad731, 0x85104263, 0x22401397, 0x112084c6 + data4 0x247d854a, 0x3df8d2bb, 0x3211aef9, 0xa16dc729 + data4 0x2f4b1d9e, 0x30f3dcb2, 0x52ec0d86, 0xe3d077c1 + data4 0x166c2bb3, 0xb999a970, 0x48fa1194, 0x642247e9 + data4 0x8cc4a8fc, 0x3f1aa0f0, 0x2cd8567d, 0x90ef2233 + data4 0x4ec78749, 0xd1c1d938, 0xa2fe8cca, 0x0b3698d4 + data4 0x81cfa6f5, 0xde28a57a, 0x8e26dab7, 0xbfa43fad + data4 0x9de42c3a, 0x920d5078, 0xcc9b6a5f, 0x4662547e + data4 0x13c2f68d, 0xb8e890d8, 0xf75e2e39, 0xaff582c3 + data4 0x80be9f5d, 0x937c69d0, 0x2da96fd5, 0x12b3cf25 + data4 0x993bc8ac, 0x7da71018, 0x636ee89c, 0xbb7bdb3b + data4 0x7809cd26, 0x18f46e59, 0xb701ec9a, 0x9aa8834f + data4 0x6e65e695, 0xe67eaaff, 0xcf0821bc, 0xe8e6ef15 + data4 0x9bd9bae7, 0x36ce4a6f, 0x09d4ea9f, 0x7cd629b0 + data4 0xb2af31a4, 0x23312a3f, 0x9430c6a5, 0x66c035a2 + data4 0xbc37744e, 0xcaa6fc82, 0xd0b0e090, 0xd81533a7 + data4 0x984af104, 0xdaf741ec, 0x500e7fcd, 0xf62f1791 + data4 0xd68d764d, 0xb04d43ef, 0x4d54ccaa, 0x04dfe496 + data4 0xb5e39ed1, 0x881b4c6a, 0x1fb8c12c, 0x517f4665 + data4 0xea049d5e, 0x355d018c, 0x7473fa87, 0x412efb0b + data4 0x1d5ab367, 0xd25292db, 0x5633e910, 0x47136dd6 + data4 0x618c9ad7, 0x0c7a37a1, 0x148e59f8, 0x3c89eb13 + data4 0x27eecea9, 0xc935b761, 0xe5ede11c, 0xb13c7a47 + data4 0xdf599cd2, 0x733f55f2, 0xce791814, 0x37bf73c7 + data4 0xcdea53f7, 0xaa5b5ffd, 0x6f14df3d, 0xdb867844 + data4 0xf381caaf, 0xc43eb968, 0x342c3824, 0x405fc2a3 + data4 0xc372161d, 0x250cbce2, 0x498b283c, 0x9541ff0d + data4 0x017139a8, 0xb3de080c, 0xe49cd8b4, 0xc1906456 + data4 0x84617bcb, 0xb670d532, 0x5c74486c, 0x5742d0b8 +// Td3: + data4 0xf4a75051, 0x4165537e, 0x17a4c31a, 0x275e963a + data4 0xab6bcb3b, 0x9d45f11f, 0xfa58abac, 0xe303934b + data4 0x30fa5520, 0x766df6ad, 0xcc769188, 0x024c25f5 + data4 0xe5d7fc4f, 0x2acbd7c5, 0x35448026, 0x62a38fb5 + data4 0xb15a49de, 0xba1b6725, 0xea0e9845, 0xfec0e15d + data4 0x2f7502c3, 0x4cf01281, 0x4697a38d, 0xd3f9c66b + data4 0x8f5fe703, 0x929c9515, 0x6d7aebbf, 0x5259da95 + data4 0xbe832dd4, 0x7421d358, 0xe0692949, 0xc9c8448e + data4 0xc2896a75, 0x8e7978f4, 0x583e6b99, 0xb971dd27 + data4 0xe14fb6be, 0x88ad17f0, 0x20ac66c9, 0xce3ab47d + data4 0xdf4a1863, 0x1a3182e5, 0x51336097, 0x537f4562 + data4 0x6477e0b1, 0x6bae84bb, 0x81a01cfe, 0x082b94f9 + data4 0x48685870, 0x45fd198f, 0xde6c8794, 0x7bf8b752 + data4 0x73d323ab, 0x4b02e272, 0x1f8f57e3, 0x55ab2a66 + data4 0xeb2807b2, 0xb5c2032f, 0xc57b9a86, 0x3708a5d3 + data4 0x2887f230, 0xbfa5b223, 0x036aba02, 0x16825ced + data4 0xcf1c2b8a, 0x79b492a7, 0x07f2f0f3, 0x69e2a14e + data4 0xdaf4cd65, 0x05bed506, 0x34621fd1, 0xa6fe8ac4 + data4 0x2e539d34, 0xf355a0a2, 0x8ae13205, 0xf6eb75a4 + data4 0x83ec390b, 0x60efaa40, 0x719f065e, 0x6e1051bd + data4 0x218af93e, 0xdd063d96, 0x3e05aedd, 0xe6bd464d + data4 0x548db591, 0xc45d0571, 0x06d46f04, 0x5015ff60 + data4 0x98fb2419, 0xbde997d6, 0x4043cc89, 0xd99e7767 + data4 0xe842bdb0, 0x898b8807, 0x195b38e7, 0xc8eedb79 + data4 0x7c0a47a1, 0x420fe97c, 0x841ec9f8, 0x00000000 + data4 0x80868309, 0x2bed4832, 0x1170ac1e, 0x5a724e6c + data4 0x0efffbfd, 0x8538560f, 0xaed51e3d, 0x2d392736 + data4 0x0fd9640a, 0x5ca62168, 0x5b54d19b, 0x362e3a24 + data4 0x0a67b10c, 0x57e70f93, 0xee96d2b4, 0x9b919e1b + data4 0xc0c54f80, 0xdc20a261, 0x774b695a, 0x121a161c + data4 0x93ba0ae2, 0xa02ae5c0, 0x22e0433c, 0x1b171d12 + data4 0x090d0b0e, 0x8bc7adf2, 0xb6a8b92d, 0x1ea9c814 + data4 0xf1198557, 0x75074caf, 0x99ddbbee, 0x7f60fda3 + data4 0x01269ff7, 0x72f5bc5c, 0x663bc544, 0xfb7e345b + data4 0x4329768b, 0x23c6dccb, 0xedfc68b6, 0xe4f163b8 + data4 0x31dccad7, 0x63851042, 0x97224013, 0xc6112084 + data4 0x4a247d85, 0xbb3df8d2, 0xf93211ae, 0x29a16dc7 + data4 0x9e2f4b1d, 0xb230f3dc, 0x8652ec0d, 0xc1e3d077 + data4 0xb3166c2b, 0x70b999a9, 0x9448fa11, 0xe9642247 + data4 0xfc8cc4a8, 0xf03f1aa0, 0x7d2cd856, 0x3390ef22 + data4 0x494ec787, 0x38d1c1d9, 0xcaa2fe8c, 0xd40b3698 + data4 0xf581cfa6, 0x7ade28a5, 0xb78e26da, 0xadbfa43f + data4 0x3a9de42c, 0x78920d50, 0x5fcc9b6a, 0x7e466254 + data4 0x8d13c2f6, 0xd8b8e890, 0x39f75e2e, 0xc3aff582 + data4 0x5d80be9f, 0xd0937c69, 0xd52da96f, 0x2512b3cf + data4 0xac993bc8, 0x187da710, 0x9c636ee8, 0x3bbb7bdb + data4 0x267809cd, 0x5918f46e, 0x9ab701ec, 0x4f9aa883 + data4 0x956e65e6, 0xffe67eaa, 0xbccf0821, 0x15e8e6ef + data4 0xe79bd9ba, 0x6f36ce4a, 0x9f09d4ea, 0xb07cd629 + data4 0xa4b2af31, 0x3f23312a, 0xa59430c6, 0xa266c035 + data4 0x4ebc3774, 0x82caa6fc, 0x90d0b0e0, 0xa7d81533 + data4 0x04984af1, 0xecdaf741, 0xcd500e7f, 0x91f62f17 + data4 0x4dd68d76, 0xefb04d43, 0xaa4d54cc, 0x9604dfe4 + data4 0xd1b5e39e, 0x6a881b4c, 0x2c1fb8c1, 0x65517f46 + data4 0x5eea049d, 0x8c355d01, 0x877473fa, 0x0b412efb + data4 0x671d5ab3, 0xdbd25292, 0x105633e9, 0xd647136d + data4 0xd7618c9a, 0xa10c7a37, 0xf8148e59, 0x133c89eb + data4 0xa927eece, 0x61c935b7, 0x1ce5ede1, 0x47b13c7a + data4 0xd2df599c, 0xf2733f55, 0x14ce7918, 0xc737bf73 + data4 0xf7cdea53, 0xfdaa5b5f, 0x3d6f14df, 0x44db8678 + data4 0xaff381ca, 0x68c43eb9, 0x24342c38, 0xa3405fc2 + data4 0x1dc37216, 0xe2250cbc, 0x3c498b28, 0x0d9541ff + data4 0xa8017139, 0x0cb3de08, 0xb4e49cd8, 0x56c19064 + data4 0xcb84617b, 0x32b670d5, 0x6c5c7448, 0xb85742d0 +// Td4: + data4 0x52000000, 0x09000000, 0x6a000000, 0xd5000000 + data4 0x30000000, 0x36000000, 0xa5000000, 0x38000000 + data4 0xbf000000, 0x40000000, 0xa3000000, 0x9e000000 + data4 0x81000000, 0xf3000000, 0xd7000000, 0xfb000000 + data4 0x7c000000, 0xe3000000, 0x39000000, 0x82000000 + data4 0x9b000000, 0x2f000000, 0xff000000, 0x87000000 + data4 0x34000000, 0x8e000000, 0x43000000, 0x44000000 + data4 0xc4000000, 0xde000000, 0xe9000000, 0xcb000000 + data4 0x54000000, 0x7b000000, 0x94000000, 0x32000000 + data4 0xa6000000, 0xc2000000, 0x23000000, 0x3d000000 + data4 0xee000000, 0x4c000000, 0x95000000, 0x0b000000 + data4 0x42000000, 0xfa000000, 0xc3000000, 0x4e000000 + data4 0x08000000, 0x2e000000, 0xa1000000, 0x66000000 + data4 0x28000000, 0xd9000000, 0x24000000, 0xb2000000 + data4 0x76000000, 0x5b000000, 0xa2000000, 0x49000000 + data4 0x6d000000, 0x8b000000, 0xd1000000, 0x25000000 + data4 0x72000000, 0xf8000000, 0xf6000000, 0x64000000 + data4 0x86000000, 0x68000000, 0x98000000, 0x16000000 + data4 0xd4000000, 0xa4000000, 0x5c000000, 0xcc000000 + data4 0x5d000000, 0x65000000, 0xb6000000, 0x92000000 + data4 0x6c000000, 0x70000000, 0x48000000, 0x50000000 + data4 0xfd000000, 0xed000000, 0xb9000000, 0xda000000 + data4 0x5e000000, 0x15000000, 0x46000000, 0x57000000 + data4 0xa7000000, 0x8d000000, 0x9d000000, 0x84000000 + data4 0x90000000, 0xd8000000, 0xab000000, 0x00000000 + data4 0x8c000000, 0xbc000000, 0xd3000000, 0x0a000000 + data4 0xf7000000, 0xe4000000, 0x58000000, 0x05000000 + data4 0xb8000000, 0xb3000000, 0x45000000, 0x06000000 + data4 0xd0000000, 0x2c000000, 0x1e000000, 0x8f000000 + data4 0xca000000, 0x3f000000, 0x0f000000, 0x02000000 + data4 0xc1000000, 0xaf000000, 0xbd000000, 0x03000000 + data4 0x01000000, 0x13000000, 0x8a000000, 0x6b000000 + data4 0x3a000000, 0x91000000, 0x11000000, 0x41000000 + data4 0x4f000000, 0x67000000, 0xdc000000, 0xea000000 + data4 0x97000000, 0xf2000000, 0xcf000000, 0xce000000 + data4 0xf0000000, 0xb4000000, 0xe6000000, 0x73000000 + data4 0x96000000, 0xac000000, 0x74000000, 0x22000000 + data4 0xe7000000, 0xad000000, 0x35000000, 0x85000000 + data4 0xe2000000, 0xf9000000, 0x37000000, 0xe8000000 + data4 0x1c000000, 0x75000000, 0xdf000000, 0x6e000000 + data4 0x47000000, 0xf1000000, 0x1a000000, 0x71000000 + data4 0x1d000000, 0x29000000, 0xc5000000, 0x89000000 + data4 0x6f000000, 0xb7000000, 0x62000000, 0x0e000000 + data4 0xaa000000, 0x18000000, 0xbe000000, 0x1b000000 + data4 0xfc000000, 0x56000000, 0x3e000000, 0x4b000000 + data4 0xc6000000, 0xd2000000, 0x79000000, 0x20000000 + data4 0x9a000000, 0xdb000000, 0xc0000000, 0xfe000000 + data4 0x78000000, 0xcd000000, 0x5a000000, 0xf4000000 + data4 0x1f000000, 0xdd000000, 0xa8000000, 0x33000000 + data4 0x88000000, 0x07000000, 0xc7000000, 0x31000000 + data4 0xb1000000, 0x12000000, 0x10000000, 0x59000000 + data4 0x27000000, 0x80000000, 0xec000000, 0x5f000000 + data4 0x60000000, 0x51000000, 0x7f000000, 0xa9000000 + data4 0x19000000, 0xb5000000, 0x4a000000, 0x0d000000 + data4 0x2d000000, 0xe5000000, 0x7a000000, 0x9f000000 + data4 0x93000000, 0xc9000000, 0x9c000000, 0xef000000 + data4 0xa0000000, 0xe0000000, 0x3b000000, 0x4d000000 + data4 0xae000000, 0x2a000000, 0xf5000000, 0xb0000000 + data4 0xc8000000, 0xeb000000, 0xbb000000, 0x3c000000 + data4 0x83000000, 0x53000000, 0x99000000, 0x61000000 + data4 0x17000000, 0x2b000000, 0x04000000, 0x7e000000 + data4 0xba000000, 0x77000000, 0xd6000000, 0x26000000 + data4 0xe1000000, 0x69000000, 0x14000000, 0x63000000 + data4 0x55000000, 0x21000000, 0x0c000000, 0x7d000000 +// Td5: + data4 0x00520000, 0x00090000, 0x006a0000, 0x00d50000 + data4 0x00300000, 0x00360000, 0x00a50000, 0x00380000 + data4 0x00bf0000, 0x00400000, 0x00a30000, 0x009e0000 + data4 0x00810000, 0x00f30000, 0x00d70000, 0x00fb0000 + data4 0x007c0000, 0x00e30000, 0x00390000, 0x00820000 + data4 0x009b0000, 0x002f0000, 0x00ff0000, 0x00870000 + data4 0x00340000, 0x008e0000, 0x00430000, 0x00440000 + data4 0x00c40000, 0x00de0000, 0x00e90000, 0x00cb0000 + data4 0x00540000, 0x007b0000, 0x00940000, 0x00320000 + data4 0x00a60000, 0x00c20000, 0x00230000, 0x003d0000 + data4 0x00ee0000, 0x004c0000, 0x00950000, 0x000b0000 + data4 0x00420000, 0x00fa0000, 0x00c30000, 0x004e0000 + data4 0x00080000, 0x002e0000, 0x00a10000, 0x00660000 + data4 0x00280000, 0x00d90000, 0x00240000, 0x00b20000 + data4 0x00760000, 0x005b0000, 0x00a20000, 0x00490000 + data4 0x006d0000, 0x008b0000, 0x00d10000, 0x00250000 + data4 0x00720000, 0x00f80000, 0x00f60000, 0x00640000 + data4 0x00860000, 0x00680000, 0x00980000, 0x00160000 + data4 0x00d40000, 0x00a40000, 0x005c0000, 0x00cc0000 + data4 0x005d0000, 0x00650000, 0x00b60000, 0x00920000 + data4 0x006c0000, 0x00700000, 0x00480000, 0x00500000 + data4 0x00fd0000, 0x00ed0000, 0x00b90000, 0x00da0000 + data4 0x005e0000, 0x00150000, 0x00460000, 0x00570000 + data4 0x00a70000, 0x008d0000, 0x009d0000, 0x00840000 + data4 0x00900000, 0x00d80000, 0x00ab0000, 0x00000000 + data4 0x008c0000, 0x00bc0000, 0x00d30000, 0x000a0000 + data4 0x00f70000, 0x00e40000, 0x00580000, 0x00050000 + data4 0x00b80000, 0x00b30000, 0x00450000, 0x00060000 + data4 0x00d00000, 0x002c0000, 0x001e0000, 0x008f0000 + data4 0x00ca0000, 0x003f0000, 0x000f0000, 0x00020000 + data4 0x00c10000, 0x00af0000, 0x00bd0000, 0x00030000 + data4 0x00010000, 0x00130000, 0x008a0000, 0x006b0000 + data4 0x003a0000, 0x00910000, 0x00110000, 0x00410000 + data4 0x004f0000, 0x00670000, 0x00dc0000, 0x00ea0000 + data4 0x00970000, 0x00f20000, 0x00cf0000, 0x00ce0000 + data4 0x00f00000, 0x00b40000, 0x00e60000, 0x00730000 + data4 0x00960000, 0x00ac0000, 0x00740000, 0x00220000 + data4 0x00e70000, 0x00ad0000, 0x00350000, 0x00850000 + data4 0x00e20000, 0x00f90000, 0x00370000, 0x00e80000 + data4 0x001c0000, 0x00750000, 0x00df0000, 0x006e0000 + data4 0x00470000, 0x00f10000, 0x001a0000, 0x00710000 + data4 0x001d0000, 0x00290000, 0x00c50000, 0x00890000 + data4 0x006f0000, 0x00b70000, 0x00620000, 0x000e0000 + data4 0x00aa0000, 0x00180000, 0x00be0000, 0x001b0000 + data4 0x00fc0000, 0x00560000, 0x003e0000, 0x004b0000 + data4 0x00c60000, 0x00d20000, 0x00790000, 0x00200000 + data4 0x009a0000, 0x00db0000, 0x00c00000, 0x00fe0000 + data4 0x00780000, 0x00cd0000, 0x005a0000, 0x00f40000 + data4 0x001f0000, 0x00dd0000, 0x00a80000, 0x00330000 + data4 0x00880000, 0x00070000, 0x00c70000, 0x00310000 + data4 0x00b10000, 0x00120000, 0x00100000, 0x00590000 + data4 0x00270000, 0x00800000, 0x00ec0000, 0x005f0000 + data4 0x00600000, 0x00510000, 0x007f0000, 0x00a90000 + data4 0x00190000, 0x00b50000, 0x004a0000, 0x000d0000 + data4 0x002d0000, 0x00e50000, 0x007a0000, 0x009f0000 + data4 0x00930000, 0x00c90000, 0x009c0000, 0x00ef0000 + data4 0x00a00000, 0x00e00000, 0x003b0000, 0x004d0000 + data4 0x00ae0000, 0x002a0000, 0x00f50000, 0x00b00000 + data4 0x00c80000, 0x00eb0000, 0x00bb0000, 0x003c0000 + data4 0x00830000, 0x00530000, 0x00990000, 0x00610000 + data4 0x00170000, 0x002b0000, 0x00040000, 0x007e0000 + data4 0x00ba0000, 0x00770000, 0x00d60000, 0x00260000 + data4 0x00e10000, 0x00690000, 0x00140000, 0x00630000 + data4 0x00550000, 0x00210000, 0x000c0000, 0x007d0000 +// Td6: + data4 0x00005200, 0x00000900, 0x00006a00, 0x0000d500 + data4 0x00003000, 0x00003600, 0x0000a500, 0x00003800 + data4 0x0000bf00, 0x00004000, 0x0000a300, 0x00009e00 + data4 0x00008100, 0x0000f300, 0x0000d700, 0x0000fb00 + data4 0x00007c00, 0x0000e300, 0x00003900, 0x00008200 + data4 0x00009b00, 0x00002f00, 0x0000ff00, 0x00008700 + data4 0x00003400, 0x00008e00, 0x00004300, 0x00004400 + data4 0x0000c400, 0x0000de00, 0x0000e900, 0x0000cb00 + data4 0x00005400, 0x00007b00, 0x00009400, 0x00003200 + data4 0x0000a600, 0x0000c200, 0x00002300, 0x00003d00 + data4 0x0000ee00, 0x00004c00, 0x00009500, 0x00000b00 + data4 0x00004200, 0x0000fa00, 0x0000c300, 0x00004e00 + data4 0x00000800, 0x00002e00, 0x0000a100, 0x00006600 + data4 0x00002800, 0x0000d900, 0x00002400, 0x0000b200 + data4 0x00007600, 0x00005b00, 0x0000a200, 0x00004900 + data4 0x00006d00, 0x00008b00, 0x0000d100, 0x00002500 + data4 0x00007200, 0x0000f800, 0x0000f600, 0x00006400 + data4 0x00008600, 0x00006800, 0x00009800, 0x00001600 + data4 0x0000d400, 0x0000a400, 0x00005c00, 0x0000cc00 + data4 0x00005d00, 0x00006500, 0x0000b600, 0x00009200 + data4 0x00006c00, 0x00007000, 0x00004800, 0x00005000 + data4 0x0000fd00, 0x0000ed00, 0x0000b900, 0x0000da00 + data4 0x00005e00, 0x00001500, 0x00004600, 0x00005700 + data4 0x0000a700, 0x00008d00, 0x00009d00, 0x00008400 + data4 0x00009000, 0x0000d800, 0x0000ab00, 0x00000000 + data4 0x00008c00, 0x0000bc00, 0x0000d300, 0x00000a00 + data4 0x0000f700, 0x0000e400, 0x00005800, 0x00000500 + data4 0x0000b800, 0x0000b300, 0x00004500, 0x00000600 + data4 0x0000d000, 0x00002c00, 0x00001e00, 0x00008f00 + data4 0x0000ca00, 0x00003f00, 0x00000f00, 0x00000200 + data4 0x0000c100, 0x0000af00, 0x0000bd00, 0x00000300 + data4 0x00000100, 0x00001300, 0x00008a00, 0x00006b00 + data4 0x00003a00, 0x00009100, 0x00001100, 0x00004100 + data4 0x00004f00, 0x00006700, 0x0000dc00, 0x0000ea00 + data4 0x00009700, 0x0000f200, 0x0000cf00, 0x0000ce00 + data4 0x0000f000, 0x0000b400, 0x0000e600, 0x00007300 + data4 0x00009600, 0x0000ac00, 0x00007400, 0x00002200 + data4 0x0000e700, 0x0000ad00, 0x00003500, 0x00008500 + data4 0x0000e200, 0x0000f900, 0x00003700, 0x0000e800 + data4 0x00001c00, 0x00007500, 0x0000df00, 0x00006e00 + data4 0x00004700, 0x0000f100, 0x00001a00, 0x00007100 + data4 0x00001d00, 0x00002900, 0x0000c500, 0x00008900 + data4 0x00006f00, 0x0000b700, 0x00006200, 0x00000e00 + data4 0x0000aa00, 0x00001800, 0x0000be00, 0x00001b00 + data4 0x0000fc00, 0x00005600, 0x00003e00, 0x00004b00 + data4 0x0000c600, 0x0000d200, 0x00007900, 0x00002000 + data4 0x00009a00, 0x0000db00, 0x0000c000, 0x0000fe00 + data4 0x00007800, 0x0000cd00, 0x00005a00, 0x0000f400 + data4 0x00001f00, 0x0000dd00, 0x0000a800, 0x00003300 + data4 0x00008800, 0x00000700, 0x0000c700, 0x00003100 + data4 0x0000b100, 0x00001200, 0x00001000, 0x00005900 + data4 0x00002700, 0x00008000, 0x0000ec00, 0x00005f00 + data4 0x00006000, 0x00005100, 0x00007f00, 0x0000a900 + data4 0x00001900, 0x0000b500, 0x00004a00, 0x00000d00 + data4 0x00002d00, 0x0000e500, 0x00007a00, 0x00009f00 + data4 0x00009300, 0x0000c900, 0x00009c00, 0x0000ef00 + data4 0x0000a000, 0x0000e000, 0x00003b00, 0x00004d00 + data4 0x0000ae00, 0x00002a00, 0x0000f500, 0x0000b000 + data4 0x0000c800, 0x0000eb00, 0x0000bb00, 0x00003c00 + data4 0x00008300, 0x00005300, 0x00009900, 0x00006100 + data4 0x00001700, 0x00002b00, 0x00000400, 0x00007e00 + data4 0x0000ba00, 0x00007700, 0x0000d600, 0x00002600 + data4 0x0000e100, 0x00006900, 0x00001400, 0x00006300 + data4 0x00005500, 0x00002100, 0x00000c00, 0x00007d00 +// Td7: + data4 0x00000052, 0x00000009, 0x0000006a, 0x000000d5 + data4 0x00000030, 0x00000036, 0x000000a5, 0x00000038 + data4 0x000000bf, 0x00000040, 0x000000a3, 0x0000009e + data4 0x00000081, 0x000000f3, 0x000000d7, 0x000000fb + data4 0x0000007c, 0x000000e3, 0x00000039, 0x00000082 + data4 0x0000009b, 0x0000002f, 0x000000ff, 0x00000087 + data4 0x00000034, 0x0000008e, 0x00000043, 0x00000044 + data4 0x000000c4, 0x000000de, 0x000000e9, 0x000000cb + data4 0x00000054, 0x0000007b, 0x00000094, 0x00000032 + data4 0x000000a6, 0x000000c2, 0x00000023, 0x0000003d + data4 0x000000ee, 0x0000004c, 0x00000095, 0x0000000b + data4 0x00000042, 0x000000fa, 0x000000c3, 0x0000004e + data4 0x00000008, 0x0000002e, 0x000000a1, 0x00000066 + data4 0x00000028, 0x000000d9, 0x00000024, 0x000000b2 + data4 0x00000076, 0x0000005b, 0x000000a2, 0x00000049 + data4 0x0000006d, 0x0000008b, 0x000000d1, 0x00000025 + data4 0x00000072, 0x000000f8, 0x000000f6, 0x00000064 + data4 0x00000086, 0x00000068, 0x00000098, 0x00000016 + data4 0x000000d4, 0x000000a4, 0x0000005c, 0x000000cc + data4 0x0000005d, 0x00000065, 0x000000b6, 0x00000092 + data4 0x0000006c, 0x00000070, 0x00000048, 0x00000050 + data4 0x000000fd, 0x000000ed, 0x000000b9, 0x000000da + data4 0x0000005e, 0x00000015, 0x00000046, 0x00000057 + data4 0x000000a7, 0x0000008d, 0x0000009d, 0x00000084 + data4 0x00000090, 0x000000d8, 0x000000ab, 0x00000000 + data4 0x0000008c, 0x000000bc, 0x000000d3, 0x0000000a + data4 0x000000f7, 0x000000e4, 0x00000058, 0x00000005 + data4 0x000000b8, 0x000000b3, 0x00000045, 0x00000006 + data4 0x000000d0, 0x0000002c, 0x0000001e, 0x0000008f + data4 0x000000ca, 0x0000003f, 0x0000000f, 0x00000002 + data4 0x000000c1, 0x000000af, 0x000000bd, 0x00000003 + data4 0x00000001, 0x00000013, 0x0000008a, 0x0000006b + data4 0x0000003a, 0x00000091, 0x00000011, 0x00000041 + data4 0x0000004f, 0x00000067, 0x000000dc, 0x000000ea + data4 0x00000097, 0x000000f2, 0x000000cf, 0x000000ce + data4 0x000000f0, 0x000000b4, 0x000000e6, 0x00000073 + data4 0x00000096, 0x000000ac, 0x00000074, 0x00000022 + data4 0x000000e7, 0x000000ad, 0x00000035, 0x00000085 + data4 0x000000e2, 0x000000f9, 0x00000037, 0x000000e8 + data4 0x0000001c, 0x00000075, 0x000000df, 0x0000006e + data4 0x00000047, 0x000000f1, 0x0000001a, 0x00000071 + data4 0x0000001d, 0x00000029, 0x000000c5, 0x00000089 + data4 0x0000006f, 0x000000b7, 0x00000062, 0x0000000e + data4 0x000000aa, 0x00000018, 0x000000be, 0x0000001b + data4 0x000000fc, 0x00000056, 0x0000003e, 0x0000004b + data4 0x000000c6, 0x000000d2, 0x00000079, 0x00000020 + data4 0x0000009a, 0x000000db, 0x000000c0, 0x000000fe + data4 0x00000078, 0x000000cd, 0x0000005a, 0x000000f4 + data4 0x0000001f, 0x000000dd, 0x000000a8, 0x00000033 + data4 0x00000088, 0x00000007, 0x000000c7, 0x00000031 + data4 0x000000b1, 0x00000012, 0x00000010, 0x00000059 + data4 0x00000027, 0x00000080, 0x000000ec, 0x0000005f + data4 0x00000060, 0x00000051, 0x0000007f, 0x000000a9 + data4 0x00000019, 0x000000b5, 0x0000004a, 0x0000000d + data4 0x0000002d, 0x000000e5, 0x0000007a, 0x0000009f + data4 0x00000093, 0x000000c9, 0x0000009c, 0x000000ef + data4 0x000000a0, 0x000000e0, 0x0000003b, 0x0000004d + data4 0x000000ae, 0x0000002a, 0x000000f5, 0x000000b0 + data4 0x000000c8, 0x000000eb, 0x000000bb, 0x0000003c + data4 0x00000083, 0x00000053, 0x00000099, 0x00000061 + data4 0x00000017, 0x0000002b, 0x00000004, 0x0000007e + data4 0x000000ba, 0x00000077, 0x000000d6, 0x00000026 + data4 0x000000e1, 0x00000069, 0x00000014, 0x00000063 + data4 0x00000055, 0x00000021, 0x0000000c, 0x0000007d +.size AES_Td#,8*256*4 // HP-UX assembler fails to ".-AES_Td#" |