1#!/usr/bin/env perl 2# 3# ==================================================================== 4# Written by Andy Polyakov <appro@openssl.org> 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# SHA512 block transform for x86. September 2007. 11# 12# May 2013. 13# 14# Add SSSE3 code path, 20-25% improvement [over original SSE2 code]. 15# 16# Performance in clock cycles per processed byte (less is better): 17# 18# gcc icc x86 asm SIMD(*) x86_64(**) 19# Pentium 100 97 61 - - 20# PIII 75 77 56 - - 21# P4 116 95 82 34.6 30.8 22# AMD K8 54 55 36 20.7 9.57 23# Core2 66 57 40 15.9 9.97 24# Westmere 70 - 38 12.2 9.58 25# Sandy Bridge 58 - 35 11.9 11.2 26# Ivy Bridge 50 - 33 11.5 8.17 27# Haswell 46 - 29 11.3 7.66 28# Skylake 40 - 26 13.3 7.25 29# Bulldozer 121 - 50 14.0 13.5 30# VIA Nano 91 - 52 33 14.7 31# Atom 126 - 68 48(***) 14.7 32# Silvermont 97 - 58 42(***) 17.5 33# Goldmont 80 - 48 19.5 12.0 34# 35# (*) whichever best applicable. 36# (**) x86_64 assembler performance is presented for reference 37# purposes, the results are for integer-only code. 38# (***) paddq is increadibly slow on Atom. 39# 40# IALU code-path is optimized for elder Pentiums. On vanilla Pentium 41# performance improvement over compiler generated code reaches ~60%, 42# while on PIII - ~35%. On newer µ-archs improvement varies from 15% 43# to 50%, but it's less important as they are expected to execute SSE2 44# code-path, which is commonly ~2-3x faster [than compiler generated 45# code]. SSE2 code-path is as fast as original sha512-sse2.pl, even 46# though it does not use 128-bit operations. The latter means that 47# SSE2-aware kernel is no longer required to execute the code. Another 48# difference is that new code optimizes amount of writes, but at the 49# cost of increased data cache "footprint" by 1/2KB. 50 51$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 52push(@INC,"${dir}","${dir}../../../perlasm"); 53require "x86asm.pl"; 54 55$output=pop; 56open STDOUT,">$output"; 57 58&asm_init($ARGV[0],$ARGV[$#ARGV] eq "386"); 59 60$sse2=0; 61for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); } 62 63&external_label("OPENSSL_ia32cap_P") if ($sse2); 64 65$Tlo=&DWP(0,"esp"); $Thi=&DWP(4,"esp"); 66$Alo=&DWP(8,"esp"); $Ahi=&DWP(8+4,"esp"); 67$Blo=&DWP(16,"esp"); $Bhi=&DWP(16+4,"esp"); 68$Clo=&DWP(24,"esp"); $Chi=&DWP(24+4,"esp"); 69$Dlo=&DWP(32,"esp"); $Dhi=&DWP(32+4,"esp"); 70$Elo=&DWP(40,"esp"); $Ehi=&DWP(40+4,"esp"); 71$Flo=&DWP(48,"esp"); $Fhi=&DWP(48+4,"esp"); 72$Glo=&DWP(56,"esp"); $Ghi=&DWP(56+4,"esp"); 73$Hlo=&DWP(64,"esp"); $Hhi=&DWP(64+4,"esp"); 74$K512="ebp"; 75 76$Asse2=&QWP(0,"esp"); 77$Bsse2=&QWP(8,"esp"); 78$Csse2=&QWP(16,"esp"); 79$Dsse2=&QWP(24,"esp"); 80$Esse2=&QWP(32,"esp"); 81$Fsse2=&QWP(40,"esp"); 82$Gsse2=&QWP(48,"esp"); 83$Hsse2=&QWP(56,"esp"); 84 85$A="mm0"; # B-D and 86$E="mm4"; # F-H are commonly loaded to respectively mm1-mm3 and 87 # mm5-mm7, but it's done on on-demand basis... 88$BxC="mm2"; # ... except for B^C 89 90sub BODY_00_15_sse2 { 91 my $phase=shift; 92 93 #&movq ("mm5",$Fsse2); # load f 94 #&movq ("mm6",$Gsse2); # load g 95 96 &movq ("mm1",$E); # %mm1 is sliding right 97 &pxor ("mm5","mm6"); # f^=g 98 &psrlq ("mm1",14); 99 &movq ($Esse2,$E); # modulo-scheduled save e 100 &pand ("mm5",$E); # f&=e 101 &psllq ($E,23); # $E is sliding left 102 &movq ($A,"mm3") if ($phase<2); 103 &movq (&QWP(8*9,"esp"),"mm7") # save X[i] 104 &movq ("mm3","mm1"); # %mm3 is T1 105 &psrlq ("mm1",4); 106 &pxor ("mm5","mm6"); # Ch(e,f,g) 107 &pxor ("mm3",$E); 108 &psllq ($E,23); 109 &pxor ("mm3","mm1"); 110 &movq ($Asse2,$A); # modulo-scheduled save a 111 &paddq ("mm7","mm5"); # X[i]+=Ch(e,f,g) 112 &pxor ("mm3",$E); 113 &psrlq ("mm1",23); 114 &paddq ("mm7",$Hsse2); # X[i]+=h 115 &pxor ("mm3","mm1"); 116 &psllq ($E,4); 117 &paddq ("mm7",QWP(0,$K512)); # X[i]+=K512[i] 118 &pxor ("mm3",$E); # T1=Sigma1_512(e) 119 120 &movq ($E,$Dsse2); # e = load d, e in next round 121 &paddq ("mm3","mm7"); # T1+=X[i] 122 &movq ("mm5",$A); # %mm5 is sliding right 123 &psrlq ("mm5",28); 124 &paddq ($E,"mm3"); # d += T1 125 &movq ("mm6",$A); # %mm6 is sliding left 126 &movq ("mm7","mm5"); 127 &psllq ("mm6",25); 128 &movq ("mm1",$Bsse2); # load b 129 &psrlq ("mm5",6); 130 &pxor ("mm7","mm6"); 131 &sub ("esp",8); 132 &psllq ("mm6",5); 133 &pxor ("mm7","mm5"); 134 &pxor ($A,"mm1"); # a^b, b^c in next round 135 &psrlq ("mm5",5); 136 &pxor ("mm7","mm6"); 137 &pand ($BxC,$A); # (b^c)&(a^b) 138 &psllq ("mm6",6); 139 &pxor ("mm7","mm5"); 140 &pxor ($BxC,"mm1"); # [h=]Maj(a,b,c) 141 &pxor ("mm6","mm7"); # Sigma0_512(a) 142 &movq ("mm7",&QWP(8*(9+16-1),"esp")) if ($phase!=0); # pre-fetch 143 &movq ("mm5",$Fsse2) if ($phase==0); # load f 144 145 if ($phase>1) { 146 &paddq ($BxC,"mm6"); # h+=Sigma0(a) 147 &add ($K512,8); 148 #&paddq ($BxC,"mm3"); # h+=T1 149 150 ($A,$BxC) = ($BxC,$A); # rotate registers 151 } else { 152 &paddq ("mm3",$BxC); # T1+=Maj(a,b,c) 153 &movq ($BxC,$A); 154 &add ($K512,8); 155 &paddq ("mm3","mm6"); # T1+=Sigma0(a) 156 &movq ("mm6",$Gsse2) if ($phase==0); # load g 157 #&movq ($A,"mm3"); # h=T1 158 } 159} 160 161sub BODY_00_15_x86 { 162 #define Sigma1(x) (ROTR((x),14) ^ ROTR((x),18) ^ ROTR((x),41)) 163 # LO lo>>14^hi<<18 ^ lo>>18^hi<<14 ^ hi>>9^lo<<23 164 # HI hi>>14^lo<<18 ^ hi>>18^lo<<14 ^ lo>>9^hi<<23 165 &mov ("ecx",$Elo); 166 &mov ("edx",$Ehi); 167 &mov ("esi","ecx"); 168 169 &shr ("ecx",9); # lo>>9 170 &mov ("edi","edx"); 171 &shr ("edx",9); # hi>>9 172 &mov ("ebx","ecx"); 173 &shl ("esi",14); # lo<<14 174 &mov ("eax","edx"); 175 &shl ("edi",14); # hi<<14 176 &xor ("ebx","esi"); 177 178 &shr ("ecx",14-9); # lo>>14 179 &xor ("eax","edi"); 180 &shr ("edx",14-9); # hi>>14 181 &xor ("eax","ecx"); 182 &shl ("esi",18-14); # lo<<18 183 &xor ("ebx","edx"); 184 &shl ("edi",18-14); # hi<<18 185 &xor ("ebx","esi"); 186 187 &shr ("ecx",18-14); # lo>>18 188 &xor ("eax","edi"); 189 &shr ("edx",18-14); # hi>>18 190 &xor ("eax","ecx"); 191 &shl ("esi",23-18); # lo<<23 192 &xor ("ebx","edx"); 193 &shl ("edi",23-18); # hi<<23 194 &xor ("eax","esi"); 195 &xor ("ebx","edi"); # T1 = Sigma1(e) 196 197 &mov ("ecx",$Flo); 198 &mov ("edx",$Fhi); 199 &mov ("esi",$Glo); 200 &mov ("edi",$Ghi); 201 &add ("eax",$Hlo); 202 &adc ("ebx",$Hhi); # T1 += h 203 &xor ("ecx","esi"); 204 &xor ("edx","edi"); 205 &and ("ecx",$Elo); 206 &and ("edx",$Ehi); 207 &add ("eax",&DWP(8*(9+15)+0,"esp")); 208 &adc ("ebx",&DWP(8*(9+15)+4,"esp")); # T1 += X[0] 209 &xor ("ecx","esi"); 210 &xor ("edx","edi"); # Ch(e,f,g) = (f^g)&e)^g 211 212 &mov ("esi",&DWP(0,$K512)); 213 &mov ("edi",&DWP(4,$K512)); # K[i] 214 &add ("eax","ecx"); 215 &adc ("ebx","edx"); # T1 += Ch(e,f,g) 216 &mov ("ecx",$Dlo); 217 &mov ("edx",$Dhi); 218 &add ("eax","esi"); 219 &adc ("ebx","edi"); # T1 += K[i] 220 &mov ($Tlo,"eax"); 221 &mov ($Thi,"ebx"); # put T1 away 222 &add ("eax","ecx"); 223 &adc ("ebx","edx"); # d += T1 224 225 #define Sigma0(x) (ROTR((x),28) ^ ROTR((x),34) ^ ROTR((x),39)) 226 # LO lo>>28^hi<<4 ^ hi>>2^lo<<30 ^ hi>>7^lo<<25 227 # HI hi>>28^lo<<4 ^ lo>>2^hi<<30 ^ lo>>7^hi<<25 228 &mov ("ecx",$Alo); 229 &mov ("edx",$Ahi); 230 &mov ($Dlo,"eax"); 231 &mov ($Dhi,"ebx"); 232 &mov ("esi","ecx"); 233 234 &shr ("ecx",2); # lo>>2 235 &mov ("edi","edx"); 236 &shr ("edx",2); # hi>>2 237 &mov ("ebx","ecx"); 238 &shl ("esi",4); # lo<<4 239 &mov ("eax","edx"); 240 &shl ("edi",4); # hi<<4 241 &xor ("ebx","esi"); 242 243 &shr ("ecx",7-2); # lo>>7 244 &xor ("eax","edi"); 245 &shr ("edx",7-2); # hi>>7 246 &xor ("ebx","ecx"); 247 &shl ("esi",25-4); # lo<<25 248 &xor ("eax","edx"); 249 &shl ("edi",25-4); # hi<<25 250 &xor ("eax","esi"); 251 252 &shr ("ecx",28-7); # lo>>28 253 &xor ("ebx","edi"); 254 &shr ("edx",28-7); # hi>>28 255 &xor ("eax","ecx"); 256 &shl ("esi",30-25); # lo<<30 257 &xor ("ebx","edx"); 258 &shl ("edi",30-25); # hi<<30 259 &xor ("eax","esi"); 260 &xor ("ebx","edi"); # Sigma0(a) 261 262 &mov ("ecx",$Alo); 263 &mov ("edx",$Ahi); 264 &mov ("esi",$Blo); 265 &mov ("edi",$Bhi); 266 &add ("eax",$Tlo); 267 &adc ("ebx",$Thi); # T1 = Sigma0(a)+T1 268 &or ("ecx","esi"); 269 &or ("edx","edi"); 270 &and ("ecx",$Clo); 271 &and ("edx",$Chi); 272 &and ("esi",$Alo); 273 &and ("edi",$Ahi); 274 &or ("ecx","esi"); 275 &or ("edx","edi"); # Maj(a,b,c) = ((a|b)&c)|(a&b) 276 277 &add ("eax","ecx"); 278 &adc ("ebx","edx"); # T1 += Maj(a,b,c) 279 &mov ($Tlo,"eax"); 280 &mov ($Thi,"ebx"); 281 282 &mov (&LB("edx"),&BP(0,$K512)); # pre-fetch LSB of *K 283 &sub ("esp",8); 284 &lea ($K512,&DWP(8,$K512)); # K++ 285} 286 287 288&function_begin("sha512_block_data_order"); 289 &mov ("esi",wparam(0)); # ctx 290 &mov ("edi",wparam(1)); # inp 291 &mov ("eax",wparam(2)); # num 292 &mov ("ebx","esp"); # saved sp 293 294 &call (&label("pic_point")); # make it PIC! 295&set_label("pic_point"); 296 &blindpop($K512); 297 &lea ($K512,&DWP(&label("K512")."-".&label("pic_point"),$K512)); 298 299 &sub ("esp",16); 300 &and ("esp",-64); 301 302 &shl ("eax",7); 303 &add ("eax","edi"); 304 &mov (&DWP(0,"esp"),"esi"); # ctx 305 &mov (&DWP(4,"esp"),"edi"); # inp 306 &mov (&DWP(8,"esp"),"eax"); # inp+num*128 307 &mov (&DWP(12,"esp"),"ebx"); # saved sp 308 309if ($sse2) { 310 &picmeup("edx","OPENSSL_ia32cap_P",$K512,&label("K512")); 311 &mov ("ecx",&DWP(0,"edx")); 312 &test ("ecx",1<<26); 313 &jz (&label("loop_x86")); 314 315 &mov ("edx",&DWP(4,"edx")); 316 317 # load ctx->h[0-7] 318 &movq ($A,&QWP(0,"esi")); 319 &and ("ecx",1<<24); # XMM registers availability 320 &movq ("mm1",&QWP(8,"esi")); 321 &and ("edx",1<<9); # SSSE3 bit 322 &movq ($BxC,&QWP(16,"esi")); 323 &or ("ecx","edx"); 324 &movq ("mm3",&QWP(24,"esi")); 325 &movq ($E,&QWP(32,"esi")); 326 &movq ("mm5",&QWP(40,"esi")); 327 &movq ("mm6",&QWP(48,"esi")); 328 &movq ("mm7",&QWP(56,"esi")); 329 &cmp ("ecx",1<<24|1<<9); 330 &je (&label("SSSE3")); 331 &sub ("esp",8*10); 332 &jmp (&label("loop_sse2")); 333 334&set_label("loop_sse2",16); 335 #&movq ($Asse2,$A); 336 &movq ($Bsse2,"mm1"); 337 &movq ($Csse2,$BxC); 338 &movq ($Dsse2,"mm3"); 339 #&movq ($Esse2,$E); 340 &movq ($Fsse2,"mm5"); 341 &movq ($Gsse2,"mm6"); 342 &pxor ($BxC,"mm1"); # magic 343 &movq ($Hsse2,"mm7"); 344 &movq ("mm3",$A); # magic 345 346 &mov ("eax",&DWP(0,"edi")); 347 &mov ("ebx",&DWP(4,"edi")); 348 &add ("edi",8); 349 &mov ("edx",15); # counter 350 &bswap ("eax"); 351 &bswap ("ebx"); 352 &jmp (&label("00_14_sse2")); 353 354&set_label("00_14_sse2",16); 355 &movd ("mm1","eax"); 356 &mov ("eax",&DWP(0,"edi")); 357 &movd ("mm7","ebx"); 358 &mov ("ebx",&DWP(4,"edi")); 359 &add ("edi",8); 360 &bswap ("eax"); 361 &bswap ("ebx"); 362 &punpckldq("mm7","mm1"); 363 364 &BODY_00_15_sse2(); 365 366 &dec ("edx"); 367 &jnz (&label("00_14_sse2")); 368 369 &movd ("mm1","eax"); 370 &movd ("mm7","ebx"); 371 &punpckldq("mm7","mm1"); 372 373 &BODY_00_15_sse2(1); 374 375 &pxor ($A,$A); # A is in %mm3 376 &mov ("edx",32); # counter 377 &jmp (&label("16_79_sse2")); 378 379&set_label("16_79_sse2",16); 380 for ($j=0;$j<2;$j++) { # 2x unroll 381 #&movq ("mm7",&QWP(8*(9+16-1),"esp")); # prefetched in BODY_00_15 382 &movq ("mm5",&QWP(8*(9+16-14),"esp")); 383 &movq ("mm1","mm7"); 384 &psrlq ("mm7",1); 385 &movq ("mm6","mm5"); 386 &psrlq ("mm5",6); 387 &psllq ("mm1",56); 388 &paddq ($A,"mm3"); # from BODY_00_15 389 &movq ("mm3","mm7"); 390 &psrlq ("mm7",7-1); 391 &pxor ("mm3","mm1"); 392 &psllq ("mm1",63-56); 393 &pxor ("mm3","mm7"); 394 &psrlq ("mm7",8-7); 395 &pxor ("mm3","mm1"); 396 &movq ("mm1","mm5"); 397 &psrlq ("mm5",19-6); 398 &pxor ("mm7","mm3"); # sigma0 399 400 &psllq ("mm6",3); 401 &pxor ("mm1","mm5"); 402 &paddq ("mm7",&QWP(8*(9+16),"esp")); 403 &pxor ("mm1","mm6"); 404 &psrlq ("mm5",61-19); 405 &paddq ("mm7",&QWP(8*(9+16-9),"esp")); 406 &pxor ("mm1","mm5"); 407 &psllq ("mm6",45-3); 408 &movq ("mm5",$Fsse2); # load f 409 &pxor ("mm1","mm6"); # sigma1 410 &movq ("mm6",$Gsse2); # load g 411 412 &paddq ("mm7","mm1"); # X[i] 413 #&movq (&QWP(8*9,"esp"),"mm7"); # moved to BODY_00_15 414 415 &BODY_00_15_sse2(2); 416 } 417 &dec ("edx"); 418 &jnz (&label("16_79_sse2")); 419 420 #&movq ($A,$Asse2); 421 &paddq ($A,"mm3"); # from BODY_00_15 422 &movq ("mm1",$Bsse2); 423 #&movq ($BxC,$Csse2); 424 &movq ("mm3",$Dsse2); 425 #&movq ($E,$Esse2); 426 &movq ("mm5",$Fsse2); 427 &movq ("mm6",$Gsse2); 428 &movq ("mm7",$Hsse2); 429 430 &pxor ($BxC,"mm1"); # de-magic 431 &paddq ($A,&QWP(0,"esi")); 432 &paddq ("mm1",&QWP(8,"esi")); 433 &paddq ($BxC,&QWP(16,"esi")); 434 &paddq ("mm3",&QWP(24,"esi")); 435 &paddq ($E,&QWP(32,"esi")); 436 &paddq ("mm5",&QWP(40,"esi")); 437 &paddq ("mm6",&QWP(48,"esi")); 438 &paddq ("mm7",&QWP(56,"esi")); 439 440 &mov ("eax",8*80); 441 &movq (&QWP(0,"esi"),$A); 442 &movq (&QWP(8,"esi"),"mm1"); 443 &movq (&QWP(16,"esi"),$BxC); 444 &movq (&QWP(24,"esi"),"mm3"); 445 &movq (&QWP(32,"esi"),$E); 446 &movq (&QWP(40,"esi"),"mm5"); 447 &movq (&QWP(48,"esi"),"mm6"); 448 &movq (&QWP(56,"esi"),"mm7"); 449 450 &lea ("esp",&DWP(0,"esp","eax")); # destroy frame 451 &sub ($K512,"eax"); # rewind K 452 453 &cmp ("edi",&DWP(8*10+8,"esp")); # are we done yet? 454 &jb (&label("loop_sse2")); 455 456 &mov ("esp",&DWP(8*10+12,"esp")); # restore sp 457 &emms (); 458&function_end_A(); 459 460&set_label("SSSE3",32); 461{ my ($cnt,$frame)=("ecx","edx"); 462 my @X=map("xmm$_",(0..7)); 463 my $j; 464 my $i=0; 465 466 &lea ($frame,&DWP(-64,"esp")); 467 &sub ("esp",256); 468 469 # fixed stack frame layout 470 # 471 # +0 A B C D E F G H # backing store 472 # +64 X[0]+K[i] .. X[15]+K[i] # XMM->MM xfer area 473 # +192 # XMM off-load ring buffer 474 # +256 # saved parameters 475 476 &movdqa (@X[1],&QWP(80*8,$K512)); # byte swap mask 477 &movdqu (@X[0],&QWP(0,"edi")); 478 &pshufb (@X[0],@X[1]); 479 for ($j=0;$j<8;$j++) { 480 &movdqa (&QWP(16*(($j-1)%4),$frame),@X[3]) if ($j>4); # off-load 481 &movdqa (@X[3],&QWP(16*($j%8),$K512)); 482 &movdqa (@X[2],@X[1]) if ($j<7); # perpetuate byte swap mask 483 &movdqu (@X[1],&QWP(16*($j+1),"edi")) if ($j<7); # next input 484 &movdqa (@X[1],&QWP(16*(($j+1)%4),$frame)) if ($j==7);# restore @X[0] 485 &paddq (@X[3],@X[0]); 486 &pshufb (@X[1],@X[2]) if ($j<7); 487 &movdqa (&QWP(16*($j%8)-128,$frame),@X[3]); # xfer X[i]+K[i] 488 489 push(@X,shift(@X)); # rotate(@X) 490 } 491 #&jmp (&label("loop_ssse3")); 492 &nop (); 493 494&set_label("loop_ssse3",32); 495 &movdqa (@X[2],&QWP(16*(($j+1)%4),$frame)); # pre-restore @X[1] 496 &movdqa (&QWP(16*(($j-1)%4),$frame),@X[3]); # off-load @X[3] 497 &lea ($K512,&DWP(16*8,$K512)); 498 499 #&movq ($Asse2,$A); # off-load A-H 500 &movq ($Bsse2,"mm1"); 501 &mov ("ebx","edi"); 502 &movq ($Csse2,$BxC); 503 &lea ("edi",&DWP(128,"edi")); # advance input 504 &movq ($Dsse2,"mm3"); 505 &cmp ("edi","eax"); 506 #&movq ($Esse2,$E); 507 &movq ($Fsse2,"mm5"); 508 &cmovb ("ebx","edi"); 509 &movq ($Gsse2,"mm6"); 510 &mov ("ecx",4); # loop counter 511 &pxor ($BxC,"mm1"); # magic 512 &movq ($Hsse2,"mm7"); 513 &pxor ("mm3","mm3"); # magic 514 515 &jmp (&label("00_47_ssse3")); 516 517sub BODY_00_15_ssse3 { # "phase-less" copy of BODY_00_15_sse2 518 ( 519 '&movq ("mm1",$E)', # %mm1 is sliding right 520 '&movq ("mm7",&QWP(((-8*$i)%128)-128,$frame))',# X[i]+K[i] 521 '&pxor ("mm5","mm6")', # f^=g 522 '&psrlq ("mm1",14)', 523 '&movq (&QWP(8*($i+4)%64,"esp"),$E)', # modulo-scheduled save e 524 '&pand ("mm5",$E)', # f&=e 525 '&psllq ($E,23)', # $E is sliding left 526 '&paddq ($A,"mm3")', # [h+=Maj(a,b,c)] 527 '&movq ("mm3","mm1")', # %mm3 is T1 528 '&psrlq("mm1",4)', 529 '&pxor ("mm5","mm6")', # Ch(e,f,g) 530 '&pxor ("mm3",$E)', 531 '&psllq($E,23)', 532 '&pxor ("mm3","mm1")', 533 '&movq (&QWP(8*$i%64,"esp"),$A)', # modulo-scheduled save a 534 '&paddq("mm7","mm5")', # X[i]+=Ch(e,f,g) 535 '&pxor ("mm3",$E)', 536 '&psrlq("mm1",23)', 537 '&paddq("mm7",&QWP(8*($i+7)%64,"esp"))', # X[i]+=h 538 '&pxor ("mm3","mm1")', 539 '&psllq($E,4)', 540 '&pxor ("mm3",$E)', # T1=Sigma1_512(e) 541 542 '&movq ($E,&QWP(8*($i+3)%64,"esp"))', # e = load d, e in next round 543 '&paddq ("mm3","mm7")', # T1+=X[i] 544 '&movq ("mm5",$A)', # %mm5 is sliding right 545 '&psrlq("mm5",28)', 546 '&paddq ($E,"mm3")', # d += T1 547 '&movq ("mm6",$A)', # %mm6 is sliding left 548 '&movq ("mm7","mm5")', 549 '&psllq("mm6",25)', 550 '&movq ("mm1",&QWP(8*($i+1)%64,"esp"))', # load b 551 '&psrlq("mm5",6)', 552 '&pxor ("mm7","mm6")', 553 '&psllq("mm6",5)', 554 '&pxor ("mm7","mm5")', 555 '&pxor ($A,"mm1")', # a^b, b^c in next round 556 '&psrlq("mm5",5)', 557 '&pxor ("mm7","mm6")', 558 '&pand ($BxC,$A)', # (b^c)&(a^b) 559 '&psllq("mm6",6)', 560 '&pxor ("mm7","mm5")', 561 '&pxor ($BxC,"mm1")', # [h=]Maj(a,b,c) 562 '&pxor ("mm6","mm7")', # Sigma0_512(a) 563 '&movq ("mm5",&QWP(8*($i+5-1)%64,"esp"))', # pre-load f 564 '&paddq ($BxC,"mm6")', # h+=Sigma0(a) 565 '&movq ("mm6",&QWP(8*($i+6-1)%64,"esp"))', # pre-load g 566 567 '($A,$BxC) = ($BxC,$A); $i--;' 568 ); 569} 570 571&set_label("00_47_ssse3",32); 572 573 for(;$j<16;$j++) { 574 my ($t0,$t2,$t1)=@X[2..4]; 575 my @insns = (&BODY_00_15_ssse3(),&BODY_00_15_ssse3()); 576 577 &movdqa ($t2,@X[5]); 578 &movdqa (@X[1],$t0); # restore @X[1] 579 &palignr ($t0,@X[0],8); # X[1..2] 580 &movdqa (&QWP(16*($j%4),$frame),@X[4]); # off-load @X[4] 581 &palignr ($t2,@X[4],8); # X[9..10] 582 583 &movdqa ($t1,$t0); 584 &psrlq ($t0,7); 585 &paddq (@X[0],$t2); # X[0..1] += X[9..10] 586 &movdqa ($t2,$t1); 587 &psrlq ($t1,1); 588 &psllq ($t2,64-8); 589 &pxor ($t0,$t1); 590 &psrlq ($t1,8-1); 591 &pxor ($t0,$t2); 592 &psllq ($t2,8-1); 593 &pxor ($t0,$t1); 594 &movdqa ($t1,@X[7]); 595 &pxor ($t0,$t2); # sigma0(X[1..2]) 596 &movdqa ($t2,@X[7]); 597 &psrlq ($t1,6); 598 &paddq (@X[0],$t0); # X[0..1] += sigma0(X[1..2]) 599 600 &movdqa ($t0,@X[7]); 601 &psrlq ($t2,19); 602 &psllq ($t0,64-61); 603 &pxor ($t1,$t2); 604 &psrlq ($t2,61-19); 605 &pxor ($t1,$t0); 606 &psllq ($t0,61-19); 607 &pxor ($t1,$t2); 608 &movdqa ($t2,&QWP(16*(($j+2)%4),$frame));# pre-restore @X[1] 609 &pxor ($t1,$t0); # sigma0(X[1..2]) 610 &movdqa ($t0,&QWP(16*($j%8),$K512)); 611 eval(shift(@insns)); 612 &paddq (@X[0],$t1); # X[0..1] += sigma0(X[14..15]) 613 eval(shift(@insns)); 614 eval(shift(@insns)); 615 eval(shift(@insns)); 616 eval(shift(@insns)); 617 &paddq ($t0,@X[0]); 618 foreach(@insns) { eval; } 619 &movdqa (&QWP(16*($j%8)-128,$frame),$t0);# xfer X[i]+K[i] 620 621 push(@X,shift(@X)); # rotate(@X) 622 } 623 &lea ($K512,&DWP(16*8,$K512)); 624 &dec ("ecx"); 625 &jnz (&label("00_47_ssse3")); 626 627 &movdqa (@X[1],&QWP(0,$K512)); # byte swap mask 628 &lea ($K512,&DWP(-80*8,$K512)); # rewind 629 &movdqu (@X[0],&QWP(0,"ebx")); 630 &pshufb (@X[0],@X[1]); 631 632 for ($j=0;$j<8;$j++) { # load next or same block 633 my @insns = (&BODY_00_15_ssse3(),&BODY_00_15_ssse3()); 634 635 &movdqa (&QWP(16*(($j-1)%4),$frame),@X[3]) if ($j>4); # off-load 636 &movdqa (@X[3],&QWP(16*($j%8),$K512)); 637 &movdqa (@X[2],@X[1]) if ($j<7); # perpetuate byte swap mask 638 &movdqu (@X[1],&QWP(16*($j+1),"ebx")) if ($j<7); # next input 639 &movdqa (@X[1],&QWP(16*(($j+1)%4),$frame)) if ($j==7);# restore @X[0] 640 &paddq (@X[3],@X[0]); 641 &pshufb (@X[1],@X[2]) if ($j<7); 642 foreach(@insns) { eval; } 643 &movdqa (&QWP(16*($j%8)-128,$frame),@X[3]);# xfer X[i]+K[i] 644 645 push(@X,shift(@X)); # rotate(@X) 646 } 647 648 #&movq ($A,$Asse2); # load A-H 649 &movq ("mm1",$Bsse2); 650 &paddq ($A,"mm3"); # from BODY_00_15 651 #&movq ($BxC,$Csse2); 652 &movq ("mm3",$Dsse2); 653 #&movq ($E,$Esse2); 654 #&movq ("mm5",$Fsse2); 655 #&movq ("mm6",$Gsse2); 656 &movq ("mm7",$Hsse2); 657 658 &pxor ($BxC,"mm1"); # de-magic 659 &paddq ($A,&QWP(0,"esi")); 660 &paddq ("mm1",&QWP(8,"esi")); 661 &paddq ($BxC,&QWP(16,"esi")); 662 &paddq ("mm3",&QWP(24,"esi")); 663 &paddq ($E,&QWP(32,"esi")); 664 &paddq ("mm5",&QWP(40,"esi")); 665 &paddq ("mm6",&QWP(48,"esi")); 666 &paddq ("mm7",&QWP(56,"esi")); 667 668 &movq (&QWP(0,"esi"),$A); 669 &movq (&QWP(8,"esi"),"mm1"); 670 &movq (&QWP(16,"esi"),$BxC); 671 &movq (&QWP(24,"esi"),"mm3"); 672 &movq (&QWP(32,"esi"),$E); 673 &movq (&QWP(40,"esi"),"mm5"); 674 &movq (&QWP(48,"esi"),"mm6"); 675 &movq (&QWP(56,"esi"),"mm7"); 676 677 &cmp ("edi","eax") # are we done yet? 678 &jb (&label("loop_ssse3")); 679 680 &mov ("esp",&DWP(64+12,$frame)); # restore sp 681 &emms (); 682} 683&function_end_A(); 684} 685&set_label("loop_x86",16); 686 # copy input block to stack reversing byte and qword order 687 for ($i=0;$i<8;$i++) { 688 &mov ("eax",&DWP($i*16+0,"edi")); 689 &mov ("ebx",&DWP($i*16+4,"edi")); 690 &mov ("ecx",&DWP($i*16+8,"edi")); 691 &mov ("edx",&DWP($i*16+12,"edi")); 692 &bswap ("eax"); 693 &bswap ("ebx"); 694 &bswap ("ecx"); 695 &bswap ("edx"); 696 &push ("eax"); 697 &push ("ebx"); 698 &push ("ecx"); 699 &push ("edx"); 700 } 701 &add ("edi",128); 702 &sub ("esp",9*8); # place for T,A,B,C,D,E,F,G,H 703 &mov (&DWP(8*(9+16)+4,"esp"),"edi"); 704 705 # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack 706 &lea ("edi",&DWP(8,"esp")); 707 &mov ("ecx",16); 708 &data_word(0xA5F3F689); # rep movsd 709 710&set_label("00_15_x86",16); 711 &BODY_00_15_x86(); 712 713 &cmp (&LB("edx"),0x94); 714 &jne (&label("00_15_x86")); 715 716&set_label("16_79_x86",16); 717 #define sigma0(x) (ROTR((x),1) ^ ROTR((x),8) ^ ((x)>>7)) 718 # LO lo>>1^hi<<31 ^ lo>>8^hi<<24 ^ lo>>7^hi<<25 719 # HI hi>>1^lo<<31 ^ hi>>8^lo<<24 ^ hi>>7 720 &mov ("ecx",&DWP(8*(9+15+16-1)+0,"esp")); 721 &mov ("edx",&DWP(8*(9+15+16-1)+4,"esp")); 722 &mov ("esi","ecx"); 723 724 &shr ("ecx",1); # lo>>1 725 &mov ("edi","edx"); 726 &shr ("edx",1); # hi>>1 727 &mov ("eax","ecx"); 728 &shl ("esi",24); # lo<<24 729 &mov ("ebx","edx"); 730 &shl ("edi",24); # hi<<24 731 &xor ("ebx","esi"); 732 733 &shr ("ecx",7-1); # lo>>7 734 &xor ("eax","edi"); 735 &shr ("edx",7-1); # hi>>7 736 &xor ("eax","ecx"); 737 &shl ("esi",31-24); # lo<<31 738 &xor ("ebx","edx"); 739 &shl ("edi",25-24); # hi<<25 740 &xor ("ebx","esi"); 741 742 &shr ("ecx",8-7); # lo>>8 743 &xor ("eax","edi"); 744 &shr ("edx",8-7); # hi>>8 745 &xor ("eax","ecx"); 746 &shl ("edi",31-25); # hi<<31 747 &xor ("ebx","edx"); 748 &xor ("eax","edi"); # T1 = sigma0(X[-15]) 749 750 &mov (&DWP(0,"esp"),"eax"); 751 &mov (&DWP(4,"esp"),"ebx"); # put T1 away 752 753 #define sigma1(x) (ROTR((x),19) ^ ROTR((x),61) ^ ((x)>>6)) 754 # LO lo>>19^hi<<13 ^ hi>>29^lo<<3 ^ lo>>6^hi<<26 755 # HI hi>>19^lo<<13 ^ lo>>29^hi<<3 ^ hi>>6 756 &mov ("ecx",&DWP(8*(9+15+16-14)+0,"esp")); 757 &mov ("edx",&DWP(8*(9+15+16-14)+4,"esp")); 758 &mov ("esi","ecx"); 759 760 &shr ("ecx",6); # lo>>6 761 &mov ("edi","edx"); 762 &shr ("edx",6); # hi>>6 763 &mov ("eax","ecx"); 764 &shl ("esi",3); # lo<<3 765 &mov ("ebx","edx"); 766 &shl ("edi",3); # hi<<3 767 &xor ("eax","esi"); 768 769 &shr ("ecx",19-6); # lo>>19 770 &xor ("ebx","edi"); 771 &shr ("edx",19-6); # hi>>19 772 &xor ("eax","ecx"); 773 &shl ("esi",13-3); # lo<<13 774 &xor ("ebx","edx"); 775 &shl ("edi",13-3); # hi<<13 776 &xor ("ebx","esi"); 777 778 &shr ("ecx",29-19); # lo>>29 779 &xor ("eax","edi"); 780 &shr ("edx",29-19); # hi>>29 781 &xor ("ebx","ecx"); 782 &shl ("edi",26-13); # hi<<26 783 &xor ("eax","edx"); 784 &xor ("eax","edi"); # sigma1(X[-2]) 785 786 &mov ("ecx",&DWP(8*(9+15+16)+0,"esp")); 787 &mov ("edx",&DWP(8*(9+15+16)+4,"esp")); 788 &add ("eax",&DWP(0,"esp")); 789 &adc ("ebx",&DWP(4,"esp")); # T1 = sigma1(X[-2])+T1 790 &mov ("esi",&DWP(8*(9+15+16-9)+0,"esp")); 791 &mov ("edi",&DWP(8*(9+15+16-9)+4,"esp")); 792 &add ("eax","ecx"); 793 &adc ("ebx","edx"); # T1 += X[-16] 794 &add ("eax","esi"); 795 &adc ("ebx","edi"); # T1 += X[-7] 796 &mov (&DWP(8*(9+15)+0,"esp"),"eax"); 797 &mov (&DWP(8*(9+15)+4,"esp"),"ebx"); # save X[0] 798 799 &BODY_00_15_x86(); 800 801 &cmp (&LB("edx"),0x17); 802 &jne (&label("16_79_x86")); 803 804 &mov ("esi",&DWP(8*(9+16+80)+0,"esp"));# ctx 805 &mov ("edi",&DWP(8*(9+16+80)+4,"esp"));# inp 806 for($i=0;$i<4;$i++) { 807 &mov ("eax",&DWP($i*16+0,"esi")); 808 &mov ("ebx",&DWP($i*16+4,"esi")); 809 &mov ("ecx",&DWP($i*16+8,"esi")); 810 &mov ("edx",&DWP($i*16+12,"esi")); 811 &add ("eax",&DWP(8+($i*16)+0,"esp")); 812 &adc ("ebx",&DWP(8+($i*16)+4,"esp")); 813 &mov (&DWP($i*16+0,"esi"),"eax"); 814 &mov (&DWP($i*16+4,"esi"),"ebx"); 815 &add ("ecx",&DWP(8+($i*16)+8,"esp")); 816 &adc ("edx",&DWP(8+($i*16)+12,"esp")); 817 &mov (&DWP($i*16+8,"esi"),"ecx"); 818 &mov (&DWP($i*16+12,"esi"),"edx"); 819 } 820 &add ("esp",8*(9+16+80)); # destroy frame 821 &sub ($K512,8*80); # rewind K 822 823 &cmp ("edi",&DWP(8,"esp")); # are we done yet? 824 &jb (&label("loop_x86")); 825 826 &mov ("esp",&DWP(12,"esp")); # restore sp 827&function_end_A(); 828 829&set_label("K512",64); # Yes! I keep it in the code segment! 830 &data_word(0xd728ae22,0x428a2f98); # u64 831 &data_word(0x23ef65cd,0x71374491); # u64 832 &data_word(0xec4d3b2f,0xb5c0fbcf); # u64 833 &data_word(0x8189dbbc,0xe9b5dba5); # u64 834 &data_word(0xf348b538,0x3956c25b); # u64 835 &data_word(0xb605d019,0x59f111f1); # u64 836 &data_word(0xaf194f9b,0x923f82a4); # u64 837 &data_word(0xda6d8118,0xab1c5ed5); # u64 838 &data_word(0xa3030242,0xd807aa98); # u64 839 &data_word(0x45706fbe,0x12835b01); # u64 840 &data_word(0x4ee4b28c,0x243185be); # u64 841 &data_word(0xd5ffb4e2,0x550c7dc3); # u64 842 &data_word(0xf27b896f,0x72be5d74); # u64 843 &data_word(0x3b1696b1,0x80deb1fe); # u64 844 &data_word(0x25c71235,0x9bdc06a7); # u64 845 &data_word(0xcf692694,0xc19bf174); # u64 846 &data_word(0x9ef14ad2,0xe49b69c1); # u64 847 &data_word(0x384f25e3,0xefbe4786); # u64 848 &data_word(0x8b8cd5b5,0x0fc19dc6); # u64 849 &data_word(0x77ac9c65,0x240ca1cc); # u64 850 &data_word(0x592b0275,0x2de92c6f); # u64 851 &data_word(0x6ea6e483,0x4a7484aa); # u64 852 &data_word(0xbd41fbd4,0x5cb0a9dc); # u64 853 &data_word(0x831153b5,0x76f988da); # u64 854 &data_word(0xee66dfab,0x983e5152); # u64 855 &data_word(0x2db43210,0xa831c66d); # u64 856 &data_word(0x98fb213f,0xb00327c8); # u64 857 &data_word(0xbeef0ee4,0xbf597fc7); # u64 858 &data_word(0x3da88fc2,0xc6e00bf3); # u64 859 &data_word(0x930aa725,0xd5a79147); # u64 860 &data_word(0xe003826f,0x06ca6351); # u64 861 &data_word(0x0a0e6e70,0x14292967); # u64 862 &data_word(0x46d22ffc,0x27b70a85); # u64 863 &data_word(0x5c26c926,0x2e1b2138); # u64 864 &data_word(0x5ac42aed,0x4d2c6dfc); # u64 865 &data_word(0x9d95b3df,0x53380d13); # u64 866 &data_word(0x8baf63de,0x650a7354); # u64 867 &data_word(0x3c77b2a8,0x766a0abb); # u64 868 &data_word(0x47edaee6,0x81c2c92e); # u64 869 &data_word(0x1482353b,0x92722c85); # u64 870 &data_word(0x4cf10364,0xa2bfe8a1); # u64 871 &data_word(0xbc423001,0xa81a664b); # u64 872 &data_word(0xd0f89791,0xc24b8b70); # u64 873 &data_word(0x0654be30,0xc76c51a3); # u64 874 &data_word(0xd6ef5218,0xd192e819); # u64 875 &data_word(0x5565a910,0xd6990624); # u64 876 &data_word(0x5771202a,0xf40e3585); # u64 877 &data_word(0x32bbd1b8,0x106aa070); # u64 878 &data_word(0xb8d2d0c8,0x19a4c116); # u64 879 &data_word(0x5141ab53,0x1e376c08); # u64 880 &data_word(0xdf8eeb99,0x2748774c); # u64 881 &data_word(0xe19b48a8,0x34b0bcb5); # u64 882 &data_word(0xc5c95a63,0x391c0cb3); # u64 883 &data_word(0xe3418acb,0x4ed8aa4a); # u64 884 &data_word(0x7763e373,0x5b9cca4f); # u64 885 &data_word(0xd6b2b8a3,0x682e6ff3); # u64 886 &data_word(0x5defb2fc,0x748f82ee); # u64 887 &data_word(0x43172f60,0x78a5636f); # u64 888 &data_word(0xa1f0ab72,0x84c87814); # u64 889 &data_word(0x1a6439ec,0x8cc70208); # u64 890 &data_word(0x23631e28,0x90befffa); # u64 891 &data_word(0xde82bde9,0xa4506ceb); # u64 892 &data_word(0xb2c67915,0xbef9a3f7); # u64 893 &data_word(0xe372532b,0xc67178f2); # u64 894 &data_word(0xea26619c,0xca273ece); # u64 895 &data_word(0x21c0c207,0xd186b8c7); # u64 896 &data_word(0xcde0eb1e,0xeada7dd6); # u64 897 &data_word(0xee6ed178,0xf57d4f7f); # u64 898 &data_word(0x72176fba,0x06f067aa); # u64 899 &data_word(0xa2c898a6,0x0a637dc5); # u64 900 &data_word(0xbef90dae,0x113f9804); # u64 901 &data_word(0x131c471b,0x1b710b35); # u64 902 &data_word(0x23047d84,0x28db77f5); # u64 903 &data_word(0x40c72493,0x32caab7b); # u64 904 &data_word(0x15c9bebc,0x3c9ebe0a); # u64 905 &data_word(0x9c100d4c,0x431d67c4); # u64 906 &data_word(0xcb3e42b6,0x4cc5d4be); # u64 907 &data_word(0xfc657e2a,0x597f299c); # u64 908 &data_word(0x3ad6faec,0x5fcb6fab); # u64 909 &data_word(0x4a475817,0x6c44198c); # u64 910 911 &data_word(0x04050607,0x00010203); # byte swap 912 &data_word(0x0c0d0e0f,0x08090a0b); # mask 913&function_end_B("sha512_block_data_order"); 914&asciz("SHA512 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>"); 915 916&asm_finish(); 917 918close STDOUT; 919