321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# ====================================================================
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# project. The module is, however, dual licensed under OpenSSL and
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# CRYPTOGAMS licenses depending on where you obtain it. For further
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# ====================================================================
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# sha1_block procedure for x86_64.
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# It was brought to my attention that on EM64T compiler-generated code
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# was far behind 32-bit assembler implementation. This is unlike on
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# Opteron where compiler-generated code was only 15% behind 32-bit
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# assembler, which originally made it hard to motivate the effort.
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# There was suggestion to mechanically translate 32-bit code, but I
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# dismissed it, reasoning that x86_64 offers enough register bank
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# capacity to fully utilize SHA-1 parallelism. Therefore this fresh
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# implementation:-) However! While 64-bit code does performs better
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# on Opteron, I failed to beat 32-bit assembler on EM64T core. Well,
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# x86_64 does offer larger *addressable* bank, but out-of-order core
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# reaches for even more registers through dynamic aliasing, and EM64T
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# core must have managed to run-time optimize even 32-bit code just as
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# good as 64-bit one. Performance improvement is summarized in the
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# following table:
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# gcc 3.4 32-bit asm cycles/byte
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# Opteron +45% +20% 6.8
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# Xeon P4 +65% +0% 9.9
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# Core2 +60% +10% 7.0
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# OpenSolaris OS modifications
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# Sun elects to use this software under the BSD license.
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# This source originates from OpenSSL file sha1-x86_64.pl at
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# ftp://ftp.openssl.org/snapshot/openssl-0.9.8-stable-SNAP-20080131.tar.gz
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# (presumably for future OpenSSL release 0.9.8h), with these changes:
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# 1. Added perl "use strict" and declared variables.
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# 2. Added OpenSolaris ENTRY_NP/SET_SIZE macros from
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# /usr/include/sys/asm_linkage.h, .ident keywords, and lint(1B) guards.
8de5c4f463386063e184a851437d58080c6c626cDan OpenSolaris Anderson# 3. Removed x86_64-xlate.pl script (not needed for as(1) or gas(1) assemblers).
321502cd0930b1eb6d4805e17f16234f3e3ff4b2dause strict;
321502cd0930b1eb6d4805e17f16234f3e3ff4b2damy ($code, $ctx, $inp, $num, $xi, $t0, $t1, $i, @V, $A, $B, $C, $D, $E, $T);
321502cd0930b1eb6d4805e17f16234f3e3ff4b2damy $output = shift;
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# void sha1_block_data_order(SHA1_CTX *ctx, const void *inpp, size_t blocks);
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# Arguments:
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# reassign arguments in order to produce more compact code
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# Temporaries:
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# State information from SHA-1 context:
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da$E="%r11d";
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da# Temporary:
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da$T="%r12d";
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da@V=($A,$B,$C,$D,$E,$T);
321502cd0930b1eb6d4805e17f16234f3e3ff4b2damy ($i,$a,$b,$c,$d,$e,$f,$host)=@_;
321502cd0930b1eb6d4805e17f16234f3e3ff4b2damy $j=$i+1;
321502cd0930b1eb6d4805e17f16234f3e3ff4b2damy ($i,$a,$b,$c,$d,$e,$f)=@_;
321502cd0930b1eb6d4805e17f16234f3e3ff4b2damy $j=$i+1;
321502cd0930b1eb6d4805e17f16234f3e3ff4b2damy ($i,$a,$b,$c,$d,$e,$f)=@_;
321502cd0930b1eb6d4805e17f16234f3e3ff4b2damy $j=$i+1;
8de5c4f463386063e184a851437d58080c6c626cDan OpenSolaris Anderson# Execution begins here
8de5c4f463386063e184a851437d58080c6c626cDan OpenSolaris Anderson#if defined(lint) || defined(__lint)
8de5c4f463386063e184a851437d58080c6c626cDan OpenSolaris Andersonsha1_block_data_order(SHA1_CTX *ctx, const void *inpp, size_t blocks)
321502cd0930b1eb6d4805e17f16234f3e3ff4b2dafor($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); }
321502cd0930b1eb6d4805e17f16234f3e3ff4b2dafor(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
321502cd0930b1eb6d4805e17f16234f3e3ff4b2dafor(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
321502cd0930b1eb6d4805e17f16234f3e3ff4b2dafor(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da # mov $C,$E
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da.asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
8de5c4f463386063e184a851437d58080c6c626cDan OpenSolaris Anderson#endif /* lint || __lint */
321502cd0930b1eb6d4805e17f16234f3e3ff4b2da####################################################################