1#! /usr/bin/env perl 2# Copyright 2007-2016 The OpenSSL Project Authors. All Rights Reserved. 3# 4# Licensed under the OpenSSL license (the "License"). You may not use 5# this file except in compliance with the License. You can obtain a copy 6# in the file LICENSE in the source distribution or at 7# https://www.openssl.org/source/license.html 8 9# 10# ==================================================================== 11# Written by Andy Polyakov <[email protected]> for the OpenSSL 12# project. The module is, however, dual licensed under OpenSSL and 13# CRYPTOGAMS licenses depending on where you obtain it. For further 14# details see http://www.openssl.org/~appro/cryptogams/. 15# ==================================================================== 16# 17# SHA256 block transform for x86. September 2007. 18# 19# Performance improvement over compiler generated code varies from 20# 10% to 40% [see below]. Not very impressive on some µ-archs, but 21# it's 5 times smaller and optimizes amount of writes. 22# 23# May 2012. 24# 25# Optimization including two of Pavel Semjanov's ideas, alternative 26# Maj and full unroll, resulted in ~20-25% improvement on most CPUs, 27# ~7% on Pentium, ~40% on Atom. As fully unrolled loop body is almost 28# 15x larger, 8KB vs. 560B, it's fired only for longer inputs. But not 29# on P4, where it kills performance, nor Sandy Bridge, where folded 30# loop is approximately as fast... 31# 32# June 2012. 33# 34# Add AMD XOP-specific code path, >30% improvement on Bulldozer over 35# May version, >60% over original. Add AVX+shrd code path, >25% 36# improvement on Sandy Bridge over May version, 60% over original. 37# 38# May 2013. 39# 40# Replace AMD XOP code path with SSSE3 to cover more processors. 41# (Biggest improvement coefficient is on upcoming Atom Silvermont, 42# not shown.) Add AVX+BMI code path. 43# 44# March 2014. 45# 46# Add support for Intel SHA Extensions. 47# 48# Performance in clock cycles per processed byte (less is better): 49# 50# gcc icc x86 asm(*) SIMD x86_64 asm(**) 51# Pentium 46 57 40/38 - - 52# PIII 36 33 27/24 - - 53# P4 41 38 28 - 17.3 54# AMD K8 27 25 19/15.5 - 14.9 55# Core2 26 23 18/15.6 14.3 13.8 56# Westmere 27 - 19/15.7 13.4 12.3 57# Sandy Bridge 25 - 15.9 12.4 11.6 58# Ivy Bridge 24 - 15.0 11.4 10.3 59# Haswell 22 - 13.9 9.46 7.80 60# Skylake 20 - 14.9 9.50 7.70 61# Bulldozer 36 - 27/22 17.0 13.6 62# VIA Nano 36 - 25/22 16.8 16.5 63# Atom 50 - 30/25 21.9 18.9 64# Silvermont 40 - 34/31 22.9 20.6 65# Goldmont 29 - 20 16.3(***) 66# 67# (*) numbers after slash are for unrolled loop, where applicable; 68# (**) x86_64 assembly performance is presented for reference 69# purposes, results are best-available; 70# (***) SHAEXT result is 4.1, strangely enough better than 64-bit one; 71 72$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 73push(@INC,"${dir}","${dir}../../../perlasm"); 74require "x86asm.pl"; 75 76$output=pop; 77open STDOUT,">$output"; 78 79&asm_init($ARGV[0]); 80 81$xmm = 1; 82 83# In upstream, this is controlled by shelling out to the compiler to check 84# versions, but BoringSSL is intended to be used with pre-generated perlasm 85# output, so this isn't useful anyway. 86# 87# TODO(davidben): Enable AVX+BMI2 code after testing by setting $avx to 2. 88$avx = 1; 89 90$avx = 0 unless ($xmm); 91 92$shaext=$xmm; ### set to zero if compiling for 1.0.1 93 94# TODO(davidben): Consider enabling the Intel SHA Extensions code once it's 95# been tested. 96$shaext = 0; 97 98$unroll_after = 64*4; # If pre-evicted from L1P cache first spin of 99 # fully unrolled loop was measured to run about 100 # 3-4x slower. If slowdown coefficient is N and 101 # unrolled loop is m times faster, then you break 102 # even at (N-1)/(m-1) blocks. Then it needs to be 103 # adjusted for probability of code being evicted, 104 # code size/cache size=1/4. Typical m is 1.15... 105 106$A="eax"; 107$E="edx"; 108$T="ebx"; 109$Aoff=&DWP(4,"esp"); 110$Boff=&DWP(8,"esp"); 111$Coff=&DWP(12,"esp"); 112$Doff=&DWP(16,"esp"); 113$Eoff=&DWP(20,"esp"); 114$Foff=&DWP(24,"esp"); 115$Goff=&DWP(28,"esp"); 116$Hoff=&DWP(32,"esp"); 117$Xoff=&DWP(36,"esp"); 118$K256="ebp"; 119 120sub BODY_16_63() { 121 &mov ($T,"ecx"); # "ecx" is preloaded 122 &mov ("esi",&DWP(4*(9+15+16-14),"esp")); 123 &ror ("ecx",18-7); 124 &mov ("edi","esi"); 125 &ror ("esi",19-17); 126 &xor ("ecx",$T); 127 &shr ($T,3); 128 &ror ("ecx",7); 129 &xor ("esi","edi"); 130 &xor ($T,"ecx"); # T = sigma0(X[-15]) 131 &ror ("esi",17); 132 &add ($T,&DWP(4*(9+15+16),"esp")); # T += X[-16] 133 &shr ("edi",10); 134 &add ($T,&DWP(4*(9+15+16-9),"esp")); # T += X[-7] 135 #&xor ("edi","esi") # sigma1(X[-2]) 136 # &add ($T,"edi"); # T += sigma1(X[-2]) 137 # &mov (&DWP(4*(9+15),"esp"),$T); # save X[0] 138 139 &BODY_00_15(1); 140} 141sub BODY_00_15() { 142 my $in_16_63=shift; 143 144 &mov ("ecx",$E); 145 &xor ("edi","esi") if ($in_16_63); # sigma1(X[-2]) 146 &mov ("esi",$Foff); 147 &ror ("ecx",25-11); 148 &add ($T,"edi") if ($in_16_63); # T += sigma1(X[-2]) 149 &mov ("edi",$Goff); 150 &xor ("ecx",$E); 151 &xor ("esi","edi"); 152 &mov ($T,&DWP(4*(9+15),"esp")) if (!$in_16_63); 153 &mov (&DWP(4*(9+15),"esp"),$T) if ($in_16_63); # save X[0] 154 &ror ("ecx",11-6); 155 &and ("esi",$E); 156 &mov ($Eoff,$E); # modulo-scheduled 157 &xor ($E,"ecx"); 158 &add ($T,$Hoff); # T += h 159 &xor ("esi","edi"); # Ch(e,f,g) 160 &ror ($E,6); # Sigma1(e) 161 &mov ("ecx",$A); 162 &add ($T,"esi"); # T += Ch(e,f,g) 163 164 &ror ("ecx",22-13); 165 &add ($T,$E); # T += Sigma1(e) 166 &mov ("edi",$Boff); 167 &xor ("ecx",$A); 168 &mov ($Aoff,$A); # modulo-scheduled 169 &lea ("esp",&DWP(-4,"esp")); 170 &ror ("ecx",13-2); 171 &mov ("esi",&DWP(0,$K256)); 172 &xor ("ecx",$A); 173 &mov ($E,$Eoff); # e in next iteration, d in this one 174 &xor ($A,"edi"); # a ^= b 175 &ror ("ecx",2); # Sigma0(a) 176 177 &add ($T,"esi"); # T+= K[i] 178 &mov (&DWP(0,"esp"),$A); # (b^c) in next round 179 &add ($E,$T); # d += T 180 &and ($A,&DWP(4,"esp")); # a &= (b^c) 181 &add ($T,"ecx"); # T += Sigma0(a) 182 &xor ($A,"edi"); # h = Maj(a,b,c) = Ch(a^b,c,b) 183 &mov ("ecx",&DWP(4*(9+15+16-1),"esp")) if ($in_16_63); # preload T 184 &add ($K256,4); 185 &add ($A,$T); # h += T 186} 187 188&static_label("K256"); 189 190&function_begin("sha256_block_data_order_nohw"); 191 &mov ("esi",wparam(0)); # ctx 192 &mov ("edi",wparam(1)); # inp 193 &mov ("eax",wparam(2)); # num 194 &mov ("ebx","esp"); # saved sp 195 196 &call (&label("pic_point")); # make it PIC! 197&set_label("pic_point"); 198 &blindpop($K256); 199 &lea ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256)); 200 201 &sub ("esp",16); 202 &and ("esp",-64); 203 204 &shl ("eax",6); 205 &add ("eax","edi"); 206 &mov (&DWP(0,"esp"),"esi"); # ctx 207 &mov (&DWP(4,"esp"),"edi"); # inp 208 &mov (&DWP(8,"esp"),"eax"); # inp+num*128 209 &mov (&DWP(12,"esp"),"ebx"); # saved sp 210 if (!$i386 && $xmm) { 211 if ($unroll_after) { 212&set_label("no_xmm"); 213 &sub ("eax","edi"); 214 &cmp ("eax",$unroll_after); 215 &jae (&label("unrolled")); 216 } } 217 &jmp (&label("loop")); 218 219sub COMPACT_LOOP() { 220my $suffix=shift; 221 222&set_label("loop$suffix",$suffix?32:16); 223 # copy input block to stack reversing byte and dword order 224 for($i=0;$i<4;$i++) { 225 &mov ("eax",&DWP($i*16+0,"edi")); 226 &mov ("ebx",&DWP($i*16+4,"edi")); 227 &mov ("ecx",&DWP($i*16+8,"edi")); 228 &bswap ("eax"); 229 &mov ("edx",&DWP($i*16+12,"edi")); 230 &bswap ("ebx"); 231 &push ("eax"); 232 &bswap ("ecx"); 233 &push ("ebx"); 234 &bswap ("edx"); 235 &push ("ecx"); 236 &push ("edx"); 237 } 238 &add ("edi",64); 239 &lea ("esp",&DWP(-4*9,"esp"));# place for A,B,C,D,E,F,G,H 240 &mov (&DWP(4*(9+16)+4,"esp"),"edi"); 241 242 # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack 243 &mov ($A,&DWP(0,"esi")); 244 &mov ("ebx",&DWP(4,"esi")); 245 &mov ("ecx",&DWP(8,"esi")); 246 &mov ("edi",&DWP(12,"esi")); 247 # &mov ($Aoff,$A); 248 &mov ($Boff,"ebx"); 249 &xor ("ebx","ecx"); 250 &mov ($Coff,"ecx"); 251 &mov ($Doff,"edi"); 252 &mov (&DWP(0,"esp"),"ebx"); # magic 253 &mov ($E,&DWP(16,"esi")); 254 &mov ("ebx",&DWP(20,"esi")); 255 &mov ("ecx",&DWP(24,"esi")); 256 &mov ("edi",&DWP(28,"esi")); 257 # &mov ($Eoff,$E); 258 &mov ($Foff,"ebx"); 259 &mov ($Goff,"ecx"); 260 &mov ($Hoff,"edi"); 261 262&set_label("00_15$suffix",16); 263 264 &BODY_00_15(); 265 266 &cmp ("esi",0xc19bf174); 267 &jne (&label("00_15$suffix")); 268 269 &mov ("ecx",&DWP(4*(9+15+16-1),"esp")); # preloaded in BODY_00_15(1) 270 &jmp (&label("16_63$suffix")); 271 272&set_label("16_63$suffix",16); 273 274 &BODY_16_63(); 275 276 &cmp ("esi",0xc67178f2); 277 &jne (&label("16_63$suffix")); 278 279 &mov ("esi",&DWP(4*(9+16+64)+0,"esp"));#ctx 280 # &mov ($A,$Aoff); 281 &mov ("ebx",$Boff); 282 # &mov ("edi",$Coff); 283 &mov ("ecx",$Doff); 284 &add ($A,&DWP(0,"esi")); 285 &add ("ebx",&DWP(4,"esi")); 286 &add ("edi",&DWP(8,"esi")); 287 &add ("ecx",&DWP(12,"esi")); 288 &mov (&DWP(0,"esi"),$A); 289 &mov (&DWP(4,"esi"),"ebx"); 290 &mov (&DWP(8,"esi"),"edi"); 291 &mov (&DWP(12,"esi"),"ecx"); 292 # &mov ($E,$Eoff); 293 &mov ("eax",$Foff); 294 &mov ("ebx",$Goff); 295 &mov ("ecx",$Hoff); 296 &mov ("edi",&DWP(4*(9+16+64)+4,"esp"));#inp 297 &add ($E,&DWP(16,"esi")); 298 &add ("eax",&DWP(20,"esi")); 299 &add ("ebx",&DWP(24,"esi")); 300 &add ("ecx",&DWP(28,"esi")); 301 &mov (&DWP(16,"esi"),$E); 302 &mov (&DWP(20,"esi"),"eax"); 303 &mov (&DWP(24,"esi"),"ebx"); 304 &mov (&DWP(28,"esi"),"ecx"); 305 306 &lea ("esp",&DWP(4*(9+16+64),"esp"));# destroy frame 307 &sub ($K256,4*64); # rewind K 308 309 &cmp ("edi",&DWP(8,"esp")); # are we done yet? 310 &jb (&label("loop$suffix")); 311} 312 &COMPACT_LOOP(); 313 &mov ("esp",&DWP(12,"esp")); # restore sp 314&function_end_A(); 315 if (!$i386 && !$xmm) { 316 # ~20% improvement on Sandy Bridge 317 local *ror = sub { &shrd(@_[0],@_) }; 318 &COMPACT_LOOP("_shrd"); 319 &mov ("esp",&DWP(12,"esp")); # restore sp 320&function_end_A(); 321 } 322 323&set_label("K256",64); # Yes! I keep it in the code segment! 324@K256=( 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5, 325 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5, 326 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3, 327 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174, 328 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc, 329 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da, 330 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7, 331 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967, 332 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13, 333 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85, 334 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3, 335 0xd192e819,0xd6990624,0xf40e3585,0x106aa070, 336 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5, 337 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3, 338 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208, 339 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2 ); 340&data_word(@K256); 341&data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f); # byte swap mask 342&asciz("SHA256 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>"); 343 344($a,$b,$c,$d,$e,$f,$g,$h)=(0..7); # offsets 345sub off { &DWP(4*(((shift)-$i)&7),"esp"); } 346 347if (!$i386 && $unroll_after) { 348my @AH=($A,$K256); 349 350&set_label("unrolled",16); 351 &lea ("esp",&DWP(-96,"esp")); 352 # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack 353 &mov ($AH[0],&DWP(0,"esi")); 354 &mov ($AH[1],&DWP(4,"esi")); 355 &mov ("ecx",&DWP(8,"esi")); 356 &mov ("ebx",&DWP(12,"esi")); 357 #&mov (&DWP(0,"esp"),$AH[0]); 358 &mov (&DWP(4,"esp"),$AH[1]); 359 &xor ($AH[1],"ecx"); # magic 360 &mov (&DWP(8,"esp"),"ecx"); 361 &mov (&DWP(12,"esp"),"ebx"); 362 &mov ($E,&DWP(16,"esi")); 363 &mov ("ebx",&DWP(20,"esi")); 364 &mov ("ecx",&DWP(24,"esi")); 365 &mov ("esi",&DWP(28,"esi")); 366 #&mov (&DWP(16,"esp"),$E); 367 &mov (&DWP(20,"esp"),"ebx"); 368 &mov (&DWP(24,"esp"),"ecx"); 369 &mov (&DWP(28,"esp"),"esi"); 370 &jmp (&label("grand_loop")); 371 372&set_label("grand_loop",16); 373 # copy input block to stack reversing byte order 374 for($i=0;$i<5;$i++) { 375 &mov ("ebx",&DWP(12*$i+0,"edi")); 376 &mov ("ecx",&DWP(12*$i+4,"edi")); 377 &bswap ("ebx"); 378 &mov ("esi",&DWP(12*$i+8,"edi")); 379 &bswap ("ecx"); 380 &mov (&DWP(32+12*$i+0,"esp"),"ebx"); 381 &bswap ("esi"); 382 &mov (&DWP(32+12*$i+4,"esp"),"ecx"); 383 &mov (&DWP(32+12*$i+8,"esp"),"esi"); 384 } 385 &mov ("ebx",&DWP($i*12,"edi")); 386 &add ("edi",64); 387 &bswap ("ebx"); 388 &mov (&DWP(96+4,"esp"),"edi"); 389 &mov (&DWP(32+12*$i,"esp"),"ebx"); 390 391 my ($t1,$t2) = ("ecx","esi"); 392 393 for ($i=0;$i<64;$i++) { 394 395 if ($i>=16) { 396 &mov ($T,$t1); # $t1 is preloaded 397 # &mov ($t2,&DWP(32+4*(($i+14)&15),"esp")); 398 &ror ($t1,18-7); 399 &mov ("edi",$t2); 400 &ror ($t2,19-17); 401 &xor ($t1,$T); 402 &shr ($T,3); 403 &ror ($t1,7); 404 &xor ($t2,"edi"); 405 &xor ($T,$t1); # T = sigma0(X[-15]) 406 &ror ($t2,17); 407 &add ($T,&DWP(32+4*($i&15),"esp")); # T += X[-16] 408 &shr ("edi",10); 409 &add ($T,&DWP(32+4*(($i+9)&15),"esp")); # T += X[-7] 410 #&xor ("edi",$t2) # sigma1(X[-2]) 411 # &add ($T,"edi"); # T += sigma1(X[-2]) 412 # &mov (&DWP(4*(9+15),"esp"),$T); # save X[0] 413 } 414 &mov ($t1,$E); 415 &xor ("edi",$t2) if ($i>=16); # sigma1(X[-2]) 416 &mov ($t2,&off($f)); 417 &ror ($E,25-11); 418 &add ($T,"edi") if ($i>=16); # T += sigma1(X[-2]) 419 &mov ("edi",&off($g)); 420 &xor ($E,$t1); 421 &mov ($T,&DWP(32+4*($i&15),"esp")) if ($i<16); # X[i] 422 &mov (&DWP(32+4*($i&15),"esp"),$T) if ($i>=16 && $i<62); # save X[0] 423 &xor ($t2,"edi"); 424 &ror ($E,11-6); 425 &and ($t2,$t1); 426 &mov (&off($e),$t1); # save $E, modulo-scheduled 427 &xor ($E,$t1); 428 &add ($T,&off($h)); # T += h 429 &xor ("edi",$t2); # Ch(e,f,g) 430 &ror ($E,6); # Sigma1(e) 431 &mov ($t1,$AH[0]); 432 &add ($T,"edi"); # T += Ch(e,f,g) 433 434 &ror ($t1,22-13); 435 &mov ($t2,$AH[0]); 436 &mov ("edi",&off($b)); 437 &xor ($t1,$AH[0]); 438 &mov (&off($a),$AH[0]); # save $A, modulo-scheduled 439 &xor ($AH[0],"edi"); # a ^= b, (b^c) in next round 440 &ror ($t1,13-2); 441 &and ($AH[1],$AH[0]); # (b^c) &= (a^b) 442 &lea ($E,&DWP(@K256[$i],$T,$E)); # T += Sigma1(1)+K[i] 443 &xor ($t1,$t2); 444 &xor ($AH[1],"edi"); # h = Maj(a,b,c) = Ch(a^b,c,b) 445 &mov ($t2,&DWP(32+4*(($i+2)&15),"esp")) if ($i>=15 && $i<63); 446 &ror ($t1,2); # Sigma0(a) 447 448 &add ($AH[1],$E); # h += T 449 &add ($E,&off($d)); # d += T 450 &add ($AH[1],$t1); # h += Sigma0(a) 451 &mov ($t1,&DWP(32+4*(($i+15)&15),"esp")) if ($i>=15 && $i<63); 452 453 @AH = reverse(@AH); # rotate(a,h) 454 ($t1,$t2) = ($t2,$t1); # rotate(t1,t2) 455 } 456 &mov ("esi",&DWP(96,"esp")); #ctx 457 #&mov ($AH[0],&DWP(0,"esp")); 458 &xor ($AH[1],"edi"); #&mov ($AH[1],&DWP(4,"esp")); 459 #&mov ("edi", &DWP(8,"esp")); 460 &mov ("ecx",&DWP(12,"esp")); 461 &add ($AH[0],&DWP(0,"esi")); 462 &add ($AH[1],&DWP(4,"esi")); 463 &add ("edi",&DWP(8,"esi")); 464 &add ("ecx",&DWP(12,"esi")); 465 &mov (&DWP(0,"esi"),$AH[0]); 466 &mov (&DWP(4,"esi"),$AH[1]); 467 &mov (&DWP(8,"esi"),"edi"); 468 &mov (&DWP(12,"esi"),"ecx"); 469 #&mov (&DWP(0,"esp"),$AH[0]); 470 &mov (&DWP(4,"esp"),$AH[1]); 471 &xor ($AH[1],"edi"); # magic 472 &mov (&DWP(8,"esp"),"edi"); 473 &mov (&DWP(12,"esp"),"ecx"); 474 #&mov ($E,&DWP(16,"esp")); 475 &mov ("edi",&DWP(20,"esp")); 476 &mov ("ebx",&DWP(24,"esp")); 477 &mov ("ecx",&DWP(28,"esp")); 478 &add ($E,&DWP(16,"esi")); 479 &add ("edi",&DWP(20,"esi")); 480 &add ("ebx",&DWP(24,"esi")); 481 &add ("ecx",&DWP(28,"esi")); 482 &mov (&DWP(16,"esi"),$E); 483 &mov (&DWP(20,"esi"),"edi"); 484 &mov (&DWP(24,"esi"),"ebx"); 485 &mov (&DWP(28,"esi"),"ecx"); 486 #&mov (&DWP(16,"esp"),$E); 487 &mov (&DWP(20,"esp"),"edi"); 488 &mov ("edi",&DWP(96+4,"esp")); # inp 489 &mov (&DWP(24,"esp"),"ebx"); 490 &mov (&DWP(28,"esp"),"ecx"); 491 492 &cmp ("edi",&DWP(96+8,"esp")); # are we done yet? 493 &jb (&label("grand_loop")); 494 495 &mov ("esp",&DWP(96+12,"esp")); # restore sp 496&function_end_A(); 497} 498&function_end_B("sha256_block_data_order_nohw"); 499 500 if (!$i386 && $xmm) {{{ 501if ($shaext) { 502###################################################################### 503# Intel SHA Extensions implementation of SHA256 update function. 504# 505my ($ctx,$inp,$end)=("esi","edi","eax"); 506my ($Wi,$ABEF,$CDGH,$TMP)=map("xmm$_",(0..2,7)); 507my @MSG=map("xmm$_",(3..6)); 508 509sub sha256op38 { 510 my ($opcodelet,$dst,$src)=@_; 511 if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/) 512 { &data_byte(0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2); } 513} 514sub sha256rnds2 { sha256op38(0xcb,@_); } 515sub sha256msg1 { sha256op38(0xcc,@_); } 516sub sha256msg2 { sha256op38(0xcd,@_); } 517 518&function_begin("sha256_block_data_order_hw"); 519 &mov ("esi",wparam(0)); # ctx 520 &mov ("edi",wparam(1)); # inp 521 &mov ("eax",wparam(2)); # num 522 &mov ("ebx","esp"); # saved sp 523 524 &call (&label("pic_point")); # make it PIC! 525&set_label("pic_point"); 526 &blindpop($K256); 527 &lea ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256)); 528 529 &sub ("esp",16); 530 &and ("esp",-64); 531 532 &shl ("eax",6); 533 &add ("eax","edi"); 534 &mov (&DWP(0,"esp"),"esi"); # ctx 535 &mov (&DWP(4,"esp"),"edi"); # inp 536 &mov (&DWP(8,"esp"),"eax"); # inp+num*128 537 &mov (&DWP(12,"esp"),"ebx"); # saved sp 538 539 # TODO(davidben): The preamble above this point comes from the original 540 # merged sha256_block_data_order function, which performed some common 541 # setup and then jumped to the particular SHA-256 implementation. The 542 # parts of the preamble that do not apply to this function can be 543 # removed. 544 545 &sub ("esp",32); 546 547 &movdqu ($ABEF,&QWP(0,$ctx)); # DCBA 548 &lea ($K256,&DWP(0x80,$K256)); 549 &movdqu ($CDGH,&QWP(16,$ctx)); # HGFE 550 &movdqa ($TMP,&QWP(0x100-0x80,$K256)); # byte swap mask 551 552 &pshufd ($Wi,$ABEF,0x1b); # ABCD 553 &pshufd ($ABEF,$ABEF,0xb1); # CDAB 554 &pshufd ($CDGH,$CDGH,0x1b); # EFGH 555 &palignr ($ABEF,$CDGH,8); # ABEF 556 &punpcklqdq ($CDGH,$Wi); # CDGH 557 &jmp (&label("loop_shaext")); 558 559&set_label("loop_shaext",16); 560 &movdqu (@MSG[0],&QWP(0,$inp)); 561 &movdqu (@MSG[1],&QWP(0x10,$inp)); 562 &movdqu (@MSG[2],&QWP(0x20,$inp)); 563 &pshufb (@MSG[0],$TMP); 564 &movdqu (@MSG[3],&QWP(0x30,$inp)); 565 &movdqa (&QWP(16,"esp"),$CDGH); # offload 566 567 &movdqa ($Wi,&QWP(0*16-0x80,$K256)); 568 &paddd ($Wi,@MSG[0]); 569 &pshufb (@MSG[1],$TMP); 570 &sha256rnds2 ($CDGH,$ABEF); # 0-3 571 &pshufd ($Wi,$Wi,0x0e); 572 &nop (); 573 &movdqa (&QWP(0,"esp"),$ABEF); # offload 574 &sha256rnds2 ($ABEF,$CDGH); 575 576 &movdqa ($Wi,&QWP(1*16-0x80,$K256)); 577 &paddd ($Wi,@MSG[1]); 578 &pshufb (@MSG[2],$TMP); 579 &sha256rnds2 ($CDGH,$ABEF); # 4-7 580 &pshufd ($Wi,$Wi,0x0e); 581 &lea ($inp,&DWP(0x40,$inp)); 582 &sha256msg1 (@MSG[0],@MSG[1]); 583 &sha256rnds2 ($ABEF,$CDGH); 584 585 &movdqa ($Wi,&QWP(2*16-0x80,$K256)); 586 &paddd ($Wi,@MSG[2]); 587 &pshufb (@MSG[3],$TMP); 588 &sha256rnds2 ($CDGH,$ABEF); # 8-11 589 &pshufd ($Wi,$Wi,0x0e); 590 &movdqa ($TMP,@MSG[3]); 591 &palignr ($TMP,@MSG[2],4); 592 &nop (); 593 &paddd (@MSG[0],$TMP); 594 &sha256msg1 (@MSG[1],@MSG[2]); 595 &sha256rnds2 ($ABEF,$CDGH); 596 597 &movdqa ($Wi,&QWP(3*16-0x80,$K256)); 598 &paddd ($Wi,@MSG[3]); 599 &sha256msg2 (@MSG[0],@MSG[3]); 600 &sha256rnds2 ($CDGH,$ABEF); # 12-15 601 &pshufd ($Wi,$Wi,0x0e); 602 &movdqa ($TMP,@MSG[0]); 603 &palignr ($TMP,@MSG[3],4); 604 &nop (); 605 &paddd (@MSG[1],$TMP); 606 &sha256msg1 (@MSG[2],@MSG[3]); 607 &sha256rnds2 ($ABEF,$CDGH); 608 609for($i=4;$i<16-3;$i++) { 610 &movdqa ($Wi,&QWP($i*16-0x80,$K256)); 611 &paddd ($Wi,@MSG[0]); 612 &sha256msg2 (@MSG[1],@MSG[0]); 613 &sha256rnds2 ($CDGH,$ABEF); # 16-19... 614 &pshufd ($Wi,$Wi,0x0e); 615 &movdqa ($TMP,@MSG[1]); 616 &palignr ($TMP,@MSG[0],4); 617 &nop (); 618 &paddd (@MSG[2],$TMP); 619 &sha256msg1 (@MSG[3],@MSG[0]); 620 &sha256rnds2 ($ABEF,$CDGH); 621 622 push(@MSG,shift(@MSG)); 623} 624 &movdqa ($Wi,&QWP(13*16-0x80,$K256)); 625 &paddd ($Wi,@MSG[0]); 626 &sha256msg2 (@MSG[1],@MSG[0]); 627 &sha256rnds2 ($CDGH,$ABEF); # 52-55 628 &pshufd ($Wi,$Wi,0x0e); 629 &movdqa ($TMP,@MSG[1]) 630 &palignr ($TMP,@MSG[0],4); 631 &sha256rnds2 ($ABEF,$CDGH); 632 &paddd (@MSG[2],$TMP); 633 634 &movdqa ($Wi,&QWP(14*16-0x80,$K256)); 635 &paddd ($Wi,@MSG[1]); 636 &sha256rnds2 ($CDGH,$ABEF); # 56-59 637 &pshufd ($Wi,$Wi,0x0e); 638 &sha256msg2 (@MSG[2],@MSG[1]); 639 &movdqa ($TMP,&QWP(0x100-0x80,$K256)); # byte swap mask 640 &sha256rnds2 ($ABEF,$CDGH); 641 642 &movdqa ($Wi,&QWP(15*16-0x80,$K256)); 643 &paddd ($Wi,@MSG[2]); 644 &nop (); 645 &sha256rnds2 ($CDGH,$ABEF); # 60-63 646 &pshufd ($Wi,$Wi,0x0e); 647 &cmp ($end,$inp); 648 &nop (); 649 &sha256rnds2 ($ABEF,$CDGH); 650 651 &paddd ($CDGH,&QWP(16,"esp")); 652 &paddd ($ABEF,&QWP(0,"esp")); 653 &jnz (&label("loop_shaext")); 654 655 &pshufd ($CDGH,$CDGH,0xb1); # DCHG 656 &pshufd ($TMP,$ABEF,0x1b); # FEBA 657 &pshufd ($ABEF,$ABEF,0xb1); # BAFE 658 &punpckhqdq ($ABEF,$CDGH); # DCBA 659 &palignr ($CDGH,$TMP,8); # HGFE 660 661 &mov ("esp",&DWP(32+12,"esp")); 662 &movdqu (&QWP(0,$ctx),$ABEF); 663 &movdqu (&QWP(16,$ctx),$CDGH); 664&function_end("sha256_block_data_order_shaext"); 665} 666 667my @X = map("xmm$_",(0..3)); 668my ($t0,$t1,$t2,$t3) = map("xmm$_",(4..7)); 669my @AH = ($A,$T); 670 671&function_begin("sha256_block_data_order_ssse3"); 672 &mov ("esi",wparam(0)); # ctx 673 &mov ("edi",wparam(1)); # inp 674 &mov ("eax",wparam(2)); # num 675 &mov ("ebx","esp"); # saved sp 676 677 &call (&label("pic_point")); # make it PIC! 678&set_label("pic_point"); 679 &blindpop($K256); 680 &lea ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256)); 681 682 &sub ("esp",16); 683 &and ("esp",-64); 684 685 &shl ("eax",6); 686 &add ("eax","edi"); 687 &mov (&DWP(0,"esp"),"esi"); # ctx 688 &mov (&DWP(4,"esp"),"edi"); # inp 689 &mov (&DWP(8,"esp"),"eax"); # inp+num*128 690 &mov (&DWP(12,"esp"),"ebx"); # saved sp 691 692 # TODO(davidben): The preamble above this point comes from the original 693 # merged sha256_block_data_order function, which performed some common 694 # setup and then jumped to the particular SHA-256 implementation. The 695 # parts of the preamble that do not apply to this function can be 696 # removed. 697 698 &lea ("esp",&DWP(-96,"esp")); 699 # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack 700 &mov ($AH[0],&DWP(0,"esi")); 701 &mov ($AH[1],&DWP(4,"esi")); 702 &mov ("ecx",&DWP(8,"esi")); 703 &mov ("edi",&DWP(12,"esi")); 704 #&mov (&DWP(0,"esp"),$AH[0]); 705 &mov (&DWP(4,"esp"),$AH[1]); 706 &xor ($AH[1],"ecx"); # magic 707 &mov (&DWP(8,"esp"),"ecx"); 708 &mov (&DWP(12,"esp"),"edi"); 709 &mov ($E,&DWP(16,"esi")); 710 &mov ("edi",&DWP(20,"esi")); 711 &mov ("ecx",&DWP(24,"esi")); 712 &mov ("esi",&DWP(28,"esi")); 713 #&mov (&DWP(16,"esp"),$E); 714 &mov (&DWP(20,"esp"),"edi"); 715 &mov ("edi",&DWP(96+4,"esp")); # inp 716 &mov (&DWP(24,"esp"),"ecx"); 717 &mov (&DWP(28,"esp"),"esi"); 718 &movdqa ($t3,&QWP(256,$K256)); 719 &jmp (&label("grand_ssse3")); 720 721&set_label("grand_ssse3",16); 722 # load input, reverse byte order, add K256[0..15], save to stack 723 &movdqu (@X[0],&QWP(0,"edi")); 724 &movdqu (@X[1],&QWP(16,"edi")); 725 &movdqu (@X[2],&QWP(32,"edi")); 726 &movdqu (@X[3],&QWP(48,"edi")); 727 &add ("edi",64); 728 &pshufb (@X[0],$t3); 729 &mov (&DWP(96+4,"esp"),"edi"); 730 &pshufb (@X[1],$t3); 731 &movdqa ($t0,&QWP(0,$K256)); 732 &pshufb (@X[2],$t3); 733 &movdqa ($t1,&QWP(16,$K256)); 734 &paddd ($t0,@X[0]); 735 &pshufb (@X[3],$t3); 736 &movdqa ($t2,&QWP(32,$K256)); 737 &paddd ($t1,@X[1]); 738 &movdqa ($t3,&QWP(48,$K256)); 739 &movdqa (&QWP(32+0,"esp"),$t0); 740 &paddd ($t2,@X[2]); 741 &movdqa (&QWP(32+16,"esp"),$t1); 742 &paddd ($t3,@X[3]); 743 &movdqa (&QWP(32+32,"esp"),$t2); 744 &movdqa (&QWP(32+48,"esp"),$t3); 745 &jmp (&label("ssse3_00_47")); 746 747&set_label("ssse3_00_47",16); 748 &add ($K256,64); 749 750sub SSSE3_00_47 () { 751my $j = shift; 752my $body = shift; 753my @X = @_; 754my @insns = (&$body,&$body,&$body,&$body); # 120 instructions 755 756 eval(shift(@insns)); 757 &movdqa ($t0,@X[1]); 758 eval(shift(@insns)); # @ 759 eval(shift(@insns)); 760 &movdqa ($t3,@X[3]); 761 eval(shift(@insns)); 762 eval(shift(@insns)); 763 &palignr ($t0,@X[0],4); # X[1..4] 764 eval(shift(@insns)); 765 eval(shift(@insns)); # @ 766 eval(shift(@insns)); 767 &palignr ($t3,@X[2],4); # X[9..12] 768 eval(shift(@insns)); 769 eval(shift(@insns)); 770 eval(shift(@insns)); 771 &movdqa ($t1,$t0); 772 eval(shift(@insns)); # @ 773 eval(shift(@insns)); 774 &movdqa ($t2,$t0); 775 eval(shift(@insns)); 776 eval(shift(@insns)); 777 &psrld ($t0,3); 778 eval(shift(@insns)); 779 eval(shift(@insns)); # @ 780 &paddd (@X[0],$t3); # X[0..3] += X[9..12] 781 eval(shift(@insns)); 782 eval(shift(@insns)); 783 &psrld ($t2,7); 784 eval(shift(@insns)); 785 eval(shift(@insns)); 786 eval(shift(@insns)); # @ 787 eval(shift(@insns)); 788 &pshufd ($t3,@X[3],0b11111010); # X[14..15] 789 eval(shift(@insns)); 790 eval(shift(@insns)); 791 &pslld ($t1,32-18); 792 eval(shift(@insns)); 793 eval(shift(@insns)); # @ 794 &pxor ($t0,$t2); 795 eval(shift(@insns)); 796 eval(shift(@insns)); 797 &psrld ($t2,18-7); 798 eval(shift(@insns)); 799 eval(shift(@insns)); 800 eval(shift(@insns)); # @ 801 &pxor ($t0,$t1); 802 eval(shift(@insns)); 803 eval(shift(@insns)); 804 &pslld ($t1,18-7); 805 eval(shift(@insns)); 806 eval(shift(@insns)); 807 eval(shift(@insns)); # @ 808 &pxor ($t0,$t2); 809 eval(shift(@insns)); 810 eval(shift(@insns)); 811 &movdqa ($t2,$t3); 812 eval(shift(@insns)); 813 eval(shift(@insns)); 814 eval(shift(@insns)); # @ 815 &pxor ($t0,$t1); # sigma0(X[1..4]) 816 eval(shift(@insns)); 817 eval(shift(@insns)); 818 &psrld ($t3,10); 819 eval(shift(@insns)); 820 eval(shift(@insns)); 821 eval(shift(@insns)); # @ 822 &paddd (@X[0],$t0); # X[0..3] += sigma0(X[1..4]) 823 eval(shift(@insns)); 824 eval(shift(@insns)); 825 &psrlq ($t2,17); 826 eval(shift(@insns)); 827 eval(shift(@insns)); 828 eval(shift(@insns)); # @ 829 &pxor ($t3,$t2); 830 eval(shift(@insns)); 831 eval(shift(@insns)); 832 &psrlq ($t2,19-17); 833 eval(shift(@insns)); 834 eval(shift(@insns)); 835 eval(shift(@insns)); # @ 836 &pxor ($t3,$t2); 837 eval(shift(@insns)); 838 eval(shift(@insns)); 839 &pshufd ($t3,$t3,0b10000000); 840 eval(shift(@insns)); 841 eval(shift(@insns)); 842 eval(shift(@insns)); # @ 843 eval(shift(@insns)); 844 eval(shift(@insns)); 845 eval(shift(@insns)); 846 eval(shift(@insns)); 847 eval(shift(@insns)); # @ 848 eval(shift(@insns)); 849 &psrldq ($t3,8); 850 eval(shift(@insns)); 851 eval(shift(@insns)); 852 eval(shift(@insns)); 853 &paddd (@X[0],$t3); # X[0..1] += sigma1(X[14..15]) 854 eval(shift(@insns)); # @ 855 eval(shift(@insns)); 856 eval(shift(@insns)); 857 eval(shift(@insns)); 858 eval(shift(@insns)); 859 eval(shift(@insns)); # @ 860 eval(shift(@insns)); 861 &pshufd ($t3,@X[0],0b01010000); # X[16..17] 862 eval(shift(@insns)); 863 eval(shift(@insns)); 864 eval(shift(@insns)); 865 &movdqa ($t2,$t3); 866 eval(shift(@insns)); # @ 867 &psrld ($t3,10); 868 eval(shift(@insns)); 869 &psrlq ($t2,17); 870 eval(shift(@insns)); 871 eval(shift(@insns)); 872 eval(shift(@insns)); 873 eval(shift(@insns)); # @ 874 &pxor ($t3,$t2); 875 eval(shift(@insns)); 876 eval(shift(@insns)); 877 &psrlq ($t2,19-17); 878 eval(shift(@insns)); 879 eval(shift(@insns)); 880 eval(shift(@insns)); # @ 881 &pxor ($t3,$t2); 882 eval(shift(@insns)); 883 eval(shift(@insns)); 884 eval(shift(@insns)); 885 &pshufd ($t3,$t3,0b00001000); 886 eval(shift(@insns)); 887 eval(shift(@insns)); # @ 888 &movdqa ($t2,&QWP(16*$j,$K256)); 889 eval(shift(@insns)); 890 eval(shift(@insns)); 891 &pslldq ($t3,8); 892 eval(shift(@insns)); 893 eval(shift(@insns)); 894 eval(shift(@insns)); # @ 895 eval(shift(@insns)); 896 eval(shift(@insns)); 897 eval(shift(@insns)); 898 eval(shift(@insns)); 899 eval(shift(@insns)); # @ 900 &paddd (@X[0],$t3); # X[2..3] += sigma1(X[16..17]) 901 eval(shift(@insns)); 902 eval(shift(@insns)); 903 eval(shift(@insns)); 904 eval(shift(@insns)); 905 &paddd ($t2,@X[0]); 906 eval(shift(@insns)); # @ 907 908 foreach (@insns) { eval; } # remaining instructions 909 910 &movdqa (&QWP(32+16*$j,"esp"),$t2); 911} 912 913sub body_00_15 () { 914 ( 915 '&mov ("ecx",$E);', 916 '&ror ($E,25-11);', 917 '&mov ("esi",&off($f));', 918 '&xor ($E,"ecx");', 919 '&mov ("edi",&off($g));', 920 '&xor ("esi","edi");', 921 '&ror ($E,11-6);', 922 '&and ("esi","ecx");', 923 '&mov (&off($e),"ecx");', # save $E, modulo-scheduled 924 '&xor ($E,"ecx");', 925 '&xor ("edi","esi");', # Ch(e,f,g) 926 '&ror ($E,6);', # T = Sigma1(e) 927 '&mov ("ecx",$AH[0]);', 928 '&add ($E,"edi");', # T += Ch(e,f,g) 929 '&mov ("edi",&off($b));', 930 '&mov ("esi",$AH[0]);', 931 932 '&ror ("ecx",22-13);', 933 '&mov (&off($a),$AH[0]);', # save $A, modulo-scheduled 934 '&xor ("ecx",$AH[0]);', 935 '&xor ($AH[0],"edi");', # a ^= b, (b^c) in next round 936 '&add ($E,&off($h));', # T += h 937 '&ror ("ecx",13-2);', 938 '&and ($AH[1],$AH[0]);', # (b^c) &= (a^b) 939 '&xor ("ecx","esi");', 940 '&add ($E,&DWP(32+4*($i&15),"esp"));', # T += K[i]+X[i] 941 '&xor ($AH[1],"edi");', # h = Maj(a,b,c) = Ch(a^b,c,b) 942 '&ror ("ecx",2);', # Sigma0(a) 943 944 '&add ($AH[1],$E);', # h += T 945 '&add ($E,&off($d));', # d += T 946 '&add ($AH[1],"ecx");'. # h += Sigma0(a) 947 948 '@AH = reverse(@AH); $i++;' # rotate(a,h) 949 ); 950} 951 952 for ($i=0,$j=0; $j<4; $j++) { 953 &SSSE3_00_47($j,\&body_00_15,@X); 954 push(@X,shift(@X)); # rotate(@X) 955 } 956 &cmp (&DWP(16*$j,$K256),0x00010203); 957 &jne (&label("ssse3_00_47")); 958 959 for ($i=0; $i<16; ) { 960 foreach(body_00_15()) { eval; } 961 } 962 963 &mov ("esi",&DWP(96,"esp")); #ctx 964 #&mov ($AH[0],&DWP(0,"esp")); 965 &xor ($AH[1],"edi"); #&mov ($AH[1],&DWP(4,"esp")); 966 #&mov ("edi", &DWP(8,"esp")); 967 &mov ("ecx",&DWP(12,"esp")); 968 &add ($AH[0],&DWP(0,"esi")); 969 &add ($AH[1],&DWP(4,"esi")); 970 &add ("edi",&DWP(8,"esi")); 971 &add ("ecx",&DWP(12,"esi")); 972 &mov (&DWP(0,"esi"),$AH[0]); 973 &mov (&DWP(4,"esi"),$AH[1]); 974 &mov (&DWP(8,"esi"),"edi"); 975 &mov (&DWP(12,"esi"),"ecx"); 976 #&mov (&DWP(0,"esp"),$AH[0]); 977 &mov (&DWP(4,"esp"),$AH[1]); 978 &xor ($AH[1],"edi"); # magic 979 &mov (&DWP(8,"esp"),"edi"); 980 &mov (&DWP(12,"esp"),"ecx"); 981 #&mov ($E,&DWP(16,"esp")); 982 &mov ("edi",&DWP(20,"esp")); 983 &mov ("ecx",&DWP(24,"esp")); 984 &add ($E,&DWP(16,"esi")); 985 &add ("edi",&DWP(20,"esi")); 986 &add ("ecx",&DWP(24,"esi")); 987 &mov (&DWP(16,"esi"),$E); 988 &mov (&DWP(20,"esi"),"edi"); 989 &mov (&DWP(20,"esp"),"edi"); 990 &mov ("edi",&DWP(28,"esp")); 991 &mov (&DWP(24,"esi"),"ecx"); 992 #&mov (&DWP(16,"esp"),$E); 993 &add ("edi",&DWP(28,"esi")); 994 &mov (&DWP(24,"esp"),"ecx"); 995 &mov (&DWP(28,"esi"),"edi"); 996 &mov (&DWP(28,"esp"),"edi"); 997 &mov ("edi",&DWP(96+4,"esp")); # inp 998 999 &movdqa ($t3,&QWP(64,$K256)); 1000 &sub ($K256,3*64); # rewind K 1001 &cmp ("edi",&DWP(96+8,"esp")); # are we done yet? 1002 &jb (&label("grand_ssse3")); 1003 1004 &mov ("esp",&DWP(96+12,"esp")); # restore sp 1005&function_end("sha256_block_data_order_ssse3"); 1006 1007 if ($avx) { 1008&function_begin("sha256_block_data_order_avx"); 1009 &mov ("esi",wparam(0)); # ctx 1010 &mov ("edi",wparam(1)); # inp 1011 &mov ("eax",wparam(2)); # num 1012 &mov ("ebx","esp"); # saved sp 1013 1014 &call (&label("pic_point")); # make it PIC! 1015&set_label("pic_point"); 1016 &blindpop($K256); 1017 &lea ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256)); 1018 1019 &sub ("esp",16); 1020 &and ("esp",-64); 1021 1022 &shl ("eax",6); 1023 &add ("eax","edi"); 1024 &mov (&DWP(0,"esp"),"esi"); # ctx 1025 &mov (&DWP(4,"esp"),"edi"); # inp 1026 &mov (&DWP(8,"esp"),"eax"); # inp+num*128 1027 &mov (&DWP(12,"esp"),"ebx"); # saved sp 1028 1029 # TODO(davidben): The preamble above this point comes from the original 1030 # merged sha256_block_data_order function, which performed some common 1031 # setup and then jumped to the particular SHA-256 implementation. The 1032 # parts of the preamble that do not apply to this function can be 1033 # removed. 1034 1035 &lea ("esp",&DWP(-96,"esp")); 1036 &vzeroall (); 1037 # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack 1038 &mov ($AH[0],&DWP(0,"esi")); 1039 &mov ($AH[1],&DWP(4,"esi")); 1040 &mov ("ecx",&DWP(8,"esi")); 1041 &mov ("edi",&DWP(12,"esi")); 1042 #&mov (&DWP(0,"esp"),$AH[0]); 1043 &mov (&DWP(4,"esp"),$AH[1]); 1044 &xor ($AH[1],"ecx"); # magic 1045 &mov (&DWP(8,"esp"),"ecx"); 1046 &mov (&DWP(12,"esp"),"edi"); 1047 &mov ($E,&DWP(16,"esi")); 1048 &mov ("edi",&DWP(20,"esi")); 1049 &mov ("ecx",&DWP(24,"esi")); 1050 &mov ("esi",&DWP(28,"esi")); 1051 #&mov (&DWP(16,"esp"),$E); 1052 &mov (&DWP(20,"esp"),"edi"); 1053 &mov ("edi",&DWP(96+4,"esp")); # inp 1054 &mov (&DWP(24,"esp"),"ecx"); 1055 &mov (&DWP(28,"esp"),"esi"); 1056 &vmovdqa ($t3,&QWP(256,$K256)); 1057 &jmp (&label("grand_avx")); 1058 1059&set_label("grand_avx",32); 1060 # load input, reverse byte order, add K256[0..15], save to stack 1061 &vmovdqu (@X[0],&QWP(0,"edi")); 1062 &vmovdqu (@X[1],&QWP(16,"edi")); 1063 &vmovdqu (@X[2],&QWP(32,"edi")); 1064 &vmovdqu (@X[3],&QWP(48,"edi")); 1065 &add ("edi",64); 1066 &vpshufb (@X[0],@X[0],$t3); 1067 &mov (&DWP(96+4,"esp"),"edi"); 1068 &vpshufb (@X[1],@X[1],$t3); 1069 &vpshufb (@X[2],@X[2],$t3); 1070 &vpaddd ($t0,@X[0],&QWP(0,$K256)); 1071 &vpshufb (@X[3],@X[3],$t3); 1072 &vpaddd ($t1,@X[1],&QWP(16,$K256)); 1073 &vpaddd ($t2,@X[2],&QWP(32,$K256)); 1074 &vpaddd ($t3,@X[3],&QWP(48,$K256)); 1075 &vmovdqa (&QWP(32+0,"esp"),$t0); 1076 &vmovdqa (&QWP(32+16,"esp"),$t1); 1077 &vmovdqa (&QWP(32+32,"esp"),$t2); 1078 &vmovdqa (&QWP(32+48,"esp"),$t3); 1079 &jmp (&label("avx_00_47")); 1080 1081&set_label("avx_00_47",16); 1082 &add ($K256,64); 1083 1084sub Xupdate_AVX () { 1085 ( 1086 '&vpalignr ($t0,@X[1],@X[0],4);', # X[1..4] 1087 '&vpalignr ($t3,@X[3],@X[2],4);', # X[9..12] 1088 '&vpsrld ($t2,$t0,7);', 1089 '&vpaddd (@X[0],@X[0],$t3);', # X[0..3] += X[9..16] 1090 '&vpsrld ($t3,$t0,3);', 1091 '&vpslld ($t1,$t0,14);', 1092 '&vpxor ($t0,$t3,$t2);', 1093 '&vpshufd ($t3,@X[3],0b11111010)',# X[14..15] 1094 '&vpsrld ($t2,$t2,18-7);', 1095 '&vpxor ($t0,$t0,$t1);', 1096 '&vpslld ($t1,$t1,25-14);', 1097 '&vpxor ($t0,$t0,$t2);', 1098 '&vpsrld ($t2,$t3,10);', 1099 '&vpxor ($t0,$t0,$t1);', # sigma0(X[1..4]) 1100 '&vpsrlq ($t1,$t3,17);', 1101 '&vpaddd (@X[0],@X[0],$t0);', # X[0..3] += sigma0(X[1..4]) 1102 '&vpxor ($t2,$t2,$t1);', 1103 '&vpsrlq ($t3,$t3,19);', 1104 '&vpxor ($t2,$t2,$t3);', # sigma1(X[14..15] 1105 '&vpshufd ($t3,$t2,0b10000100);', 1106 '&vpsrldq ($t3,$t3,8);', 1107 '&vpaddd (@X[0],@X[0],$t3);', # X[0..1] += sigma1(X[14..15]) 1108 '&vpshufd ($t3,@X[0],0b01010000)',# X[16..17] 1109 '&vpsrld ($t2,$t3,10);', 1110 '&vpsrlq ($t1,$t3,17);', 1111 '&vpxor ($t2,$t2,$t1);', 1112 '&vpsrlq ($t3,$t3,19);', 1113 '&vpxor ($t2,$t2,$t3);', # sigma1(X[16..17] 1114 '&vpshufd ($t3,$t2,0b11101000);', 1115 '&vpslldq ($t3,$t3,8);', 1116 '&vpaddd (@X[0],@X[0],$t3);' # X[2..3] += sigma1(X[16..17]) 1117 ); 1118} 1119 1120local *ror = sub { &shrd(@_[0],@_) }; 1121sub AVX_00_47 () { 1122my $j = shift; 1123my $body = shift; 1124my @X = @_; 1125my @insns = (&$body,&$body,&$body,&$body); # 120 instructions 1126my $insn; 1127 1128 foreach (Xupdate_AVX()) { # 31 instructions 1129 eval; 1130 eval(shift(@insns)); 1131 eval(shift(@insns)); 1132 eval($insn = shift(@insns)); 1133 eval(shift(@insns)) if ($insn =~ /rorx/ && @insns[0] =~ /rorx/); 1134 } 1135 &vpaddd ($t2,@X[0],&QWP(16*$j,$K256)); 1136 foreach (@insns) { eval; } # remaining instructions 1137 &vmovdqa (&QWP(32+16*$j,"esp"),$t2); 1138} 1139 1140 for ($i=0,$j=0; $j<4; $j++) { 1141 &AVX_00_47($j,\&body_00_15,@X); 1142 push(@X,shift(@X)); # rotate(@X) 1143 } 1144 &cmp (&DWP(16*$j,$K256),0x00010203); 1145 &jne (&label("avx_00_47")); 1146 1147 for ($i=0; $i<16; ) { 1148 foreach(body_00_15()) { eval; } 1149 } 1150 1151 &mov ("esi",&DWP(96,"esp")); #ctx 1152 #&mov ($AH[0],&DWP(0,"esp")); 1153 &xor ($AH[1],"edi"); #&mov ($AH[1],&DWP(4,"esp")); 1154 #&mov ("edi", &DWP(8,"esp")); 1155 &mov ("ecx",&DWP(12,"esp")); 1156 &add ($AH[0],&DWP(0,"esi")); 1157 &add ($AH[1],&DWP(4,"esi")); 1158 &add ("edi",&DWP(8,"esi")); 1159 &add ("ecx",&DWP(12,"esi")); 1160 &mov (&DWP(0,"esi"),$AH[0]); 1161 &mov (&DWP(4,"esi"),$AH[1]); 1162 &mov (&DWP(8,"esi"),"edi"); 1163 &mov (&DWP(12,"esi"),"ecx"); 1164 #&mov (&DWP(0,"esp"),$AH[0]); 1165 &mov (&DWP(4,"esp"),$AH[1]); 1166 &xor ($AH[1],"edi"); # magic 1167 &mov (&DWP(8,"esp"),"edi"); 1168 &mov (&DWP(12,"esp"),"ecx"); 1169 #&mov ($E,&DWP(16,"esp")); 1170 &mov ("edi",&DWP(20,"esp")); 1171 &mov ("ecx",&DWP(24,"esp")); 1172 &add ($E,&DWP(16,"esi")); 1173 &add ("edi",&DWP(20,"esi")); 1174 &add ("ecx",&DWP(24,"esi")); 1175 &mov (&DWP(16,"esi"),$E); 1176 &mov (&DWP(20,"esi"),"edi"); 1177 &mov (&DWP(20,"esp"),"edi"); 1178 &mov ("edi",&DWP(28,"esp")); 1179 &mov (&DWP(24,"esi"),"ecx"); 1180 #&mov (&DWP(16,"esp"),$E); 1181 &add ("edi",&DWP(28,"esi")); 1182 &mov (&DWP(24,"esp"),"ecx"); 1183 &mov (&DWP(28,"esi"),"edi"); 1184 &mov (&DWP(28,"esp"),"edi"); 1185 &mov ("edi",&DWP(96+4,"esp")); # inp 1186 1187 &vmovdqa ($t3,&QWP(64,$K256)); 1188 &sub ($K256,3*64); # rewind K 1189 &cmp ("edi",&DWP(96+8,"esp")); # are we done yet? 1190 &jb (&label("grand_avx")); 1191 1192 &mov ("esp",&DWP(96+12,"esp")); # restore sp 1193 &vzeroall (); 1194&function_end("sha256_block_data_order_avx"); 1195 1196 if ($avx>1) { 1197sub bodyx_00_15 () { # +10% 1198 ( 1199 '&rorx ("ecx",$E,6)', 1200 '&rorx ("esi",$E,11)', 1201 '&mov (&off($e),$E)', # save $E, modulo-scheduled 1202 '&rorx ("edi",$E,25)', 1203 '&xor ("ecx","esi")', 1204 '&andn ("esi",$E,&off($g))', 1205 '&xor ("ecx","edi")', # Sigma1(e) 1206 '&and ($E,&off($f))', 1207 '&mov (&off($a),$AH[0]);', # save $A, modulo-scheduled 1208 '&or ($E,"esi")', # T = Ch(e,f,g) 1209 1210 '&rorx ("edi",$AH[0],2)', 1211 '&rorx ("esi",$AH[0],13)', 1212 '&lea ($E,&DWP(0,$E,"ecx"))', # T += Sigma1(e) 1213 '&rorx ("ecx",$AH[0],22)', 1214 '&xor ("esi","edi")', 1215 '&mov ("edi",&off($b))', 1216 '&xor ("ecx","esi")', # Sigma0(a) 1217 1218 '&xor ($AH[0],"edi")', # a ^= b, (b^c) in next round 1219 '&add ($E,&off($h))', # T += h 1220 '&and ($AH[1],$AH[0])', # (b^c) &= (a^b) 1221 '&add ($E,&DWP(32+4*($i&15),"esp"))', # T += K[i]+X[i] 1222 '&xor ($AH[1],"edi")', # h = Maj(a,b,c) = Ch(a^b,c,b) 1223 1224 '&add ("ecx",$E)', # h += T 1225 '&add ($E,&off($d))', # d += T 1226 '&lea ($AH[1],&DWP(0,$AH[1],"ecx"));'. # h += Sigma0(a) 1227 1228 '@AH = reverse(@AH); $i++;' # rotate(a,h) 1229 ); 1230} 1231 1232# If enabled, this function should be gated on AVX, BMI1, and BMI2. 1233&function_begin("sha256_block_data_order_avx_bmi"); 1234 &mov ("esi",wparam(0)); # ctx 1235 &mov ("edi",wparam(1)); # inp 1236 &mov ("eax",wparam(2)); # num 1237 &mov ("ebx","esp"); # saved sp 1238 1239 &call (&label("pic_point")); # make it PIC! 1240&set_label("pic_point"); 1241 &blindpop($K256); 1242 &lea ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256)); 1243 1244 &sub ("esp",16); 1245 &and ("esp",-64); 1246 1247 &shl ("eax",6); 1248 &add ("eax","edi"); 1249 &mov (&DWP(0,"esp"),"esi"); # ctx 1250 &mov (&DWP(4,"esp"),"edi"); # inp 1251 &mov (&DWP(8,"esp"),"eax"); # inp+num*128 1252 &mov (&DWP(12,"esp"),"ebx"); # saved sp 1253 1254 # TODO(davidben): The preamble above this point comes from the original 1255 # merged sha256_block_data_order function, which performed some common 1256 # setup and then jumped to the particular SHA-256 implementation. The 1257 # parts of the preamble that do not apply to this function can be 1258 # removed. 1259 1260 &lea ("esp",&DWP(-96,"esp")); 1261 &vzeroall (); 1262 # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack 1263 &mov ($AH[0],&DWP(0,"esi")); 1264 &mov ($AH[1],&DWP(4,"esi")); 1265 &mov ("ecx",&DWP(8,"esi")); 1266 &mov ("edi",&DWP(12,"esi")); 1267 #&mov (&DWP(0,"esp"),$AH[0]); 1268 &mov (&DWP(4,"esp"),$AH[1]); 1269 &xor ($AH[1],"ecx"); # magic 1270 &mov (&DWP(8,"esp"),"ecx"); 1271 &mov (&DWP(12,"esp"),"edi"); 1272 &mov ($E,&DWP(16,"esi")); 1273 &mov ("edi",&DWP(20,"esi")); 1274 &mov ("ecx",&DWP(24,"esi")); 1275 &mov ("esi",&DWP(28,"esi")); 1276 #&mov (&DWP(16,"esp"),$E); 1277 &mov (&DWP(20,"esp"),"edi"); 1278 &mov ("edi",&DWP(96+4,"esp")); # inp 1279 &mov (&DWP(24,"esp"),"ecx"); 1280 &mov (&DWP(28,"esp"),"esi"); 1281 &vmovdqa ($t3,&QWP(256,$K256)); 1282 &jmp (&label("grand_avx_bmi")); 1283 1284&set_label("grand_avx_bmi",32); 1285 # load input, reverse byte order, add K256[0..15], save to stack 1286 &vmovdqu (@X[0],&QWP(0,"edi")); 1287 &vmovdqu (@X[1],&QWP(16,"edi")); 1288 &vmovdqu (@X[2],&QWP(32,"edi")); 1289 &vmovdqu (@X[3],&QWP(48,"edi")); 1290 &add ("edi",64); 1291 &vpshufb (@X[0],@X[0],$t3); 1292 &mov (&DWP(96+4,"esp"),"edi"); 1293 &vpshufb (@X[1],@X[1],$t3); 1294 &vpshufb (@X[2],@X[2],$t3); 1295 &vpaddd ($t0,@X[0],&QWP(0,$K256)); 1296 &vpshufb (@X[3],@X[3],$t3); 1297 &vpaddd ($t1,@X[1],&QWP(16,$K256)); 1298 &vpaddd ($t2,@X[2],&QWP(32,$K256)); 1299 &vpaddd ($t3,@X[3],&QWP(48,$K256)); 1300 &vmovdqa (&QWP(32+0,"esp"),$t0); 1301 &vmovdqa (&QWP(32+16,"esp"),$t1); 1302 &vmovdqa (&QWP(32+32,"esp"),$t2); 1303 &vmovdqa (&QWP(32+48,"esp"),$t3); 1304 &jmp (&label("avx_bmi_00_47")); 1305 1306&set_label("avx_bmi_00_47",16); 1307 &add ($K256,64); 1308 1309 for ($i=0,$j=0; $j<4; $j++) { 1310 &AVX_00_47($j,\&bodyx_00_15,@X); 1311 push(@X,shift(@X)); # rotate(@X) 1312 } 1313 &cmp (&DWP(16*$j,$K256),0x00010203); 1314 &jne (&label("avx_bmi_00_47")); 1315 1316 for ($i=0; $i<16; ) { 1317 foreach(bodyx_00_15()) { eval; } 1318 } 1319 1320 &mov ("esi",&DWP(96,"esp")); #ctx 1321 #&mov ($AH[0],&DWP(0,"esp")); 1322 &xor ($AH[1],"edi"); #&mov ($AH[1],&DWP(4,"esp")); 1323 #&mov ("edi", &DWP(8,"esp")); 1324 &mov ("ecx",&DWP(12,"esp")); 1325 &add ($AH[0],&DWP(0,"esi")); 1326 &add ($AH[1],&DWP(4,"esi")); 1327 &add ("edi",&DWP(8,"esi")); 1328 &add ("ecx",&DWP(12,"esi")); 1329 &mov (&DWP(0,"esi"),$AH[0]); 1330 &mov (&DWP(4,"esi"),$AH[1]); 1331 &mov (&DWP(8,"esi"),"edi"); 1332 &mov (&DWP(12,"esi"),"ecx"); 1333 #&mov (&DWP(0,"esp"),$AH[0]); 1334 &mov (&DWP(4,"esp"),$AH[1]); 1335 &xor ($AH[1],"edi"); # magic 1336 &mov (&DWP(8,"esp"),"edi"); 1337 &mov (&DWP(12,"esp"),"ecx"); 1338 #&mov ($E,&DWP(16,"esp")); 1339 &mov ("edi",&DWP(20,"esp")); 1340 &mov ("ecx",&DWP(24,"esp")); 1341 &add ($E,&DWP(16,"esi")); 1342 &add ("edi",&DWP(20,"esi")); 1343 &add ("ecx",&DWP(24,"esi")); 1344 &mov (&DWP(16,"esi"),$E); 1345 &mov (&DWP(20,"esi"),"edi"); 1346 &mov (&DWP(20,"esp"),"edi"); 1347 &mov ("edi",&DWP(28,"esp")); 1348 &mov (&DWP(24,"esi"),"ecx"); 1349 #&mov (&DWP(16,"esp"),$E); 1350 &add ("edi",&DWP(28,"esi")); 1351 &mov (&DWP(24,"esp"),"ecx"); 1352 &mov (&DWP(28,"esi"),"edi"); 1353 &mov (&DWP(28,"esp"),"edi"); 1354 &mov ("edi",&DWP(96+4,"esp")); # inp 1355 1356 &vmovdqa ($t3,&QWP(64,$K256)); 1357 &sub ($K256,3*64); # rewind K 1358 &cmp ("edi",&DWP(96+8,"esp")); # are we done yet? 1359 &jb (&label("grand_avx_bmi")); 1360 1361 &mov ("esp",&DWP(96+12,"esp")); # restore sp 1362 &vzeroall (); 1363&function_end("sha256_block_data_order_avx_bmi"); 1364 } 1365 } 1366 }}} 1367 1368&asm_finish(); 1369 1370close STDOUT or die "error closing STDOUT: $!"; 1371