1#!/usr/bin/env perl 2# 3# ==================================================================== 4# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL 5# project. The module is, however, dual licensed under OpenSSL and 6# CRYPTOGAMS licenses depending on where you obtain it. For further 7# details see http://www.openssl.org/~appro/cryptogams/. 8# ==================================================================== 9# 10# sha1_block procedure for x86_64. 11# 12# It was brought to my attention that on EM64T compiler-generated code 13# was far behind 32-bit assembler implementation. This is unlike on 14# Opteron where compiler-generated code was only 15% behind 32-bit 15# assembler, which originally made it hard to motivate the effort. 16# There was suggestion to mechanically translate 32-bit code, but I 17# dismissed it, reasoning that x86_64 offers enough register bank 18# capacity to fully utilize SHA-1 parallelism. Therefore this fresh 19# implementation:-) However! While 64-bit code does performs better 20# on Opteron, I failed to beat 32-bit assembler on EM64T core. Well, 21# x86_64 does offer larger *addressable* bank, but out-of-order core 22# reaches for even more registers through dynamic aliasing, and EM64T 23# core must have managed to run-time optimize even 32-bit code just as 24# good as 64-bit one. Performance improvement is summarized in the 25# following table: 26# 27# gcc 3.4 32-bit asm cycles/byte 28# Opteron +45% +20% 6.8 29# Xeon P4 +65% +0% 9.9 30# Core2 +60% +10% 7.0 31 32# 33# OpenSolaris OS modifications 34# 35# Sun elects to use this software under the BSD license. 36# 37# This source originates from OpenSSL file sha1-x86_64.pl at 38# ftp://ftp.openssl.org/snapshot/openssl-0.9.8-stable-SNAP-20080131.tar.gz 39# (presumably for future OpenSSL release 0.9.8h), with these changes: 40# 41# 1. Added perl "use strict" and declared variables. 42# 43# 2. Added OpenSolaris ENTRY_NP/SET_SIZE macros from 44# /usr/include/sys/asm_linkage.h, .ident keywords, and lint(1B) guards. 45# 46# 3. Added perl function &lea_offset_eax_register_register() to handle 47# Solaris as(1) bug. 48# 49# 4. Removed x86_64-xlate.pl script (not needed for as(1) or gas(1) assemblers). 50# 51 52use strict; 53my ($code, $ctx, $inp, $num, $xi, $t0, $t1, $i, @V, $A, $B, $C, $D, $E, $T); 54my $output = shift; 55open STDOUT,">$output"; 56 57 58sub lea_offset_eax_register_register 59# Workaround for a Solaris "gas" assembler bug where compiling the source 60# errors out and does not generate a valid "lea" instruction. Specifically, 61# &lea OFFSET(%eax, SOURCE_REGISTER),DESTINATION_REGISTER 62# 63# For Solaris as, "as -a32" must be used to compile this. 64# For Solaris gas 2.15, this errors out with this message: 65# Error: `0x5a827999(%eax,%r11d)' is not a valid 64 bit base/index expression 66# 67# This should be fixed in Solaris gas 2.16. 68# It assembles with the Linux "as --64" gas 2.17 assembler and runs OK. 69# 70# For the ONBLD NV tools, the aw wrapper script fails when -a32 is used: 71# /ws/onnv-tools/onbld/bin/i386/aw -xarch=amd64 -P -a32 -o lea.o lea.s 72# aw: as->gas mapping failed at or near arg '-a32' 73# 74# For more information, see CRs 6644870 and 6628627. 75{ 76 use Switch; 77 my ($offset, $reg_src, $reg_dest) = @_; 78 79 # Failed "lea" instruction. 80 # This instruction errors out from the Solaris as assembler. 81 # It assembles with the Linux "as --64" assembler and runs OK. 82 $code .= " /lea $offset(%eax,$reg_src),$reg_dest\n"; 83 84 # Workaround 85 # This workaround hand-generates hex machine code for lea. 86 $code .= " / Solaris as assembly bug CR 6628627 errors out for\n"; 87 $code .= " / the above, so we specify the machine code in hex:\n"; 88 $code .= " .byte 0x67 / lea\n"; 89 90 switch ($reg_src) { 91 case "%ebp" { 92 switch ($reg_dest) { 93 case "%r11d" { $code .= 94 " .byte 0x44,0x8d,0x9c,0x28 " 95 . "/ (%eax,$reg_src),$reg_dest\n"; } 96 else { $code .= "Unknown register $reg_dest\n"; } 97 } 98 } 99 case "%edi" { 100 switch ($reg_dest) { 101 case "%ebp" { $code .= 102 " .byte 0x8d,0xac,0x38 " 103 . "/ (%eax,$reg_src),$reg_dest\n"; } 104 else { $code .= "Unknown register $reg_dest\n"; } 105 } 106 } 107 case "%edx" { 108 switch ($reg_dest) { 109 case "%esi" { $code .= 110 " .byte 0x8d,0xb4,0x10 " 111 . "/ (%eax,$reg_src),$reg_dest\n"; } 112 else { $code .= "Unknown register $reg_dest\n"; } 113 } 114 } 115 case "%esi" { 116 switch ($reg_dest) { 117 case "%edi" { $code .= 118 " .byte 0x8d,0xbc,0x30 " 119 . "/ (%eax,$reg_src),$reg_dest\n"; } 120 else { $code .= "Unknown register $reg_dest\n"; } 121 } 122 } 123 case "%r11d" { 124 switch ($reg_dest) { 125 case "%r12d" { $code .= 126 " .byte 0x46,0x8d,0xa4,0x18 " 127 . "/ (%eax,$reg_src),$reg_dest\n"; } 128 else { $code .= "Unknown register $reg_dest\n"; } 129 } 130 } 131 case "%r12d" { 132 switch ($reg_dest) { 133 case "%edx" { $code .= 134 " .byte 0x42,0x8d,0x94,0x20 " 135 . "/ (%eax,$reg_src),$reg_dest\n"; } 136 else { $code .= "Unknown register $reg_dest\n"; } 137 } 138 } 139 else { $code .= "Unknown register $reg_src\n"; } 140 } 141 142 $code .= " .long $offset / offset\n"; 143} 144 145 146# 147# void sha1_block_data_order(SHA1_CTX *ctx, const void *inpp, size_t blocks); 148# 149 150# Arguments: 151$ctx="%rdi"; # 1st arg 152$inp="%rsi"; # 2nd arg 153$num="%rdx"; # 3rd arg 154 155# reassign arguments in order to produce more compact code 156$ctx="%r8"; 157$inp="%r9"; 158$num="%r10"; 159 160# Temporaries: 161$xi="%eax"; 162$t0="%ebx"; 163$t1="%ecx"; 164# State information from SHA-1 context: 165$A="%edx"; 166$B="%esi"; 167$C="%edi"; 168$D="%ebp"; 169$E="%r11d"; 170# Temporary: 171$T="%r12d"; 172 173@V=($A,$B,$C,$D,$E,$T); 174 175sub PROLOGUE { 176my $func=shift; 177$code.=<<___; 178ENTRY_NP($func) 179 push %rbx 180 push %rbp 181 push %r12 182 mov %rsp,%rax 183 mov %rdi,$ctx # reassigned argument 184 sub \$`8+16*4`,%rsp 185 mov %rsi,$inp # reassigned argument 186 and \$-64,%rsp 187 mov %rdx,$num # reassigned argument 188 mov %rax,`16*4`(%rsp) 189 190 mov 0($ctx),$A 191 mov 4($ctx),$B 192 mov 8($ctx),$C 193 mov 12($ctx),$D 194 mov 16($ctx),$E 195___ 196} 197 198sub EPILOGUE { 199my $func=shift; 200$code.=<<___; 201 mov `16*4`(%rsp),%rsp 202 pop %r12 203 pop %rbp 204 pop %rbx 205 ret 206SET_SIZE($func) 207___ 208} 209 210sub BODY_00_19 { 211my ($i,$a,$b,$c,$d,$e,$f,$host)=@_; 212my $j=$i+1; 213$code.=<<___ if ($i==0); 214 mov `4*$i`($inp),$xi 215 `"bswap $xi" if(!defined($host))` 216 mov $xi,`4*$i`(%rsp) 217___ 218 &lea_offset_eax_register_register("0x5a827999", $e, $f) if ($i < 15); 219$code.=<<___ if ($i<15); 220 /lea 0x5a827999($xi,$e),$f 221 mov $c,$t0 222 mov `4*$j`($inp),$xi 223 mov $a,$e 224 xor $d,$t0 225 `"bswap $xi" if(!defined($host))` 226 rol \$5,$e 227 and $b,$t0 228 mov $xi,`4*$j`(%rsp) 229 add $e,$f 230 xor $d,$t0 231 rol \$30,$b 232 add $t0,$f 233___ 234 &lea_offset_eax_register_register("0x5a827999", $e, $f) if ($i >= 15); 235$code.=<<___ if ($i>=15); 236 /lea 0x5a827999($xi,$e),$f 237 mov `4*($j%16)`(%rsp),$xi 238 mov $c,$t0 239 mov $a,$e 240 xor `4*(($j+2)%16)`(%rsp),$xi 241 xor $d,$t0 242 rol \$5,$e 243 xor `4*(($j+8)%16)`(%rsp),$xi 244 and $b,$t0 245 add $e,$f 246 xor `4*(($j+13)%16)`(%rsp),$xi 247 xor $d,$t0 248 rol \$30,$b 249 add $t0,$f 250 rol \$1,$xi 251 mov $xi,`4*($j%16)`(%rsp) 252___ 253} 254 255sub BODY_20_39 { 256my ($i,$a,$b,$c,$d,$e,$f)=@_; 257my $j=$i+1; 258my $K=($i<40)?0x6ed9eba1:0xca62c1d6; 259 &lea_offset_eax_register_register($K, $e, $f) if ($i < 79); 260$code.=<<___ if ($i<79); 261 /lea $K($xi,$e),$f 262 mov `4*($j%16)`(%rsp),$xi 263 mov $c,$t0 264 mov $a,$e 265 xor `4*(($j+2)%16)`(%rsp),$xi 266 xor $b,$t0 267 rol \$5,$e 268 xor `4*(($j+8)%16)`(%rsp),$xi 269 xor $d,$t0 270 add $e,$f 271 xor `4*(($j+13)%16)`(%rsp),$xi 272 rol \$30,$b 273 add $t0,$f 274 rol \$1,$xi 275___ 276$code.=<<___ if ($i<76); 277 mov $xi,`4*($j%16)`(%rsp) 278___ 279 &lea_offset_eax_register_register($K, $e, $f) if ($i == 79); 280$code.=<<___ if ($i==79); 281 /lea $K($xi,$e),$f 282 mov $c,$t0 283 mov $a,$e 284 xor $b,$t0 285 rol \$5,$e 286 xor $d,$t0 287 add $e,$f 288 rol \$30,$b 289 add $t0,$f 290___ 291} 292 293sub BODY_40_59 { 294my ($i,$a,$b,$c,$d,$e,$f)=@_; 295my $j=$i+1; 296 &lea_offset_eax_register_register("0x8f1bbcdc", $e, $f); 297$code.=<<___; 298 /lea 0x8f1bbcdc($xi,$e),$f 299 mov `4*($j%16)`(%rsp),$xi 300 mov $b,$t0 301 mov $b,$t1 302 xor `4*(($j+2)%16)`(%rsp),$xi 303 mov $a,$e 304 and $c,$t0 305 xor `4*(($j+8)%16)`(%rsp),$xi 306 or $c,$t1 307 rol \$5,$e 308 xor `4*(($j+13)%16)`(%rsp),$xi 309 and $d,$t1 310 add $e,$f 311 rol \$1,$xi 312 or $t1,$t0 313 rol \$30,$b 314 mov $xi,`4*($j%16)`(%rsp) 315 add $t0,$f 316___ 317} 318 319$code=<<___; 320#if !defined(lint) && !defined(__lint) 321 .ident "%Z%%M% %I% %E% SMI" 322#include <sys/asm_linkage.h> 323___ 324 325 326&PROLOGUE("sha1_block_data_order"); 327$code.=".align 4\n.Lloop:\n"; 328for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); } 329for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); } 330for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); } 331for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); } 332$code.=<<___; 333 / Update and save state information in SHA-1 context 334 add 0($ctx),$E 335 add 4($ctx),$T 336 add 8($ctx),$A 337 add 12($ctx),$B 338 add 16($ctx),$C 339 mov $E,0($ctx) 340 mov $T,4($ctx) 341 mov $A,8($ctx) 342 mov $B,12($ctx) 343 mov $C,16($ctx) 344 345 xchg $E,$A # mov $E,$A 346 xchg $T,$B # mov $T,$B 347 xchg $E,$C # mov $A,$C 348 xchg $T,$D # mov $B,$D 349 # mov $C,$E 350 lea `16*4`($inp),$inp 351 sub \$1,$num 352 jnz .Lloop 353___ 354&EPILOGUE("sha1_block_data_order"); 355$code.=<<___; 356.asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>" 357 358#else 359 /* LINTED */ 360 /* Nothing to be linted in this file--it's pure assembly source. */ 361#endif /* !lint && !__lint */ 362___ 363 364#################################################################### 365 366$code =~ s/\`([^\`]*)\`/eval $1/gem; 367print $code; 368close STDOUT; 369