1#!/usr/bin/env perl 2 3# ==================================================================== 4# [Re]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# "[Re]written" was achieved in two major overhauls. In 2004 BODY_* 11# functions were re-implemented to address P4 performance issue [see 12# commentary below], and in 2006 the rest was rewritten in order to 13# gain freedom to liberate licensing terms. 14 15# January, September 2004. 16# 17# It was noted that Intel IA-32 C compiler generates code which 18# performs ~30% *faster* on P4 CPU than original *hand-coded* 19# SHA1 assembler implementation. To address this problem (and 20# prove that humans are still better than machines:-), the 21# original code was overhauled, which resulted in following 22# performance changes: 23# 24# compared with original compared with Intel cc 25# assembler impl. generated code 26# Pentium -16% +48% 27# PIII/AMD +8% +16% 28# P4 +85%(!) +45% 29# 30# As you can see Pentium came out as looser:-( Yet I reckoned that 31# improvement on P4 outweights the loss and incorporate this 32# re-tuned code to 0.9.7 and later. 33# ---------------------------------------------------------------- 34# <appro@fy.chalmers.se> 35 36# August 2009. 37# 38# George Spelvin has tipped that F_40_59(b,c,d) can be rewritten as 39# '(c&d) + (b&(c^d))', which allows to accumulate partial results 40# and lighten "pressure" on scratch registers. This resulted in 41# >12% performance improvement on contemporary AMD cores (with no 42# degradation on other CPUs:-). Also, the code was revised to maximize 43# "distance" between instructions producing input to 'lea' instruction 44# and the 'lea' instruction itself, which is essential for Intel Atom 45# core and resulted in ~15% improvement. 46 47# October 2010. 48# 49# Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it 50# is to offload message schedule denoted by Wt in NIST specification, 51# or Xupdate in OpenSSL source, to SIMD unit. The idea is not novel, 52# and in SSE2 context was first explored by Dean Gaudet in 2004, see 53# http://arctic.org/~dean/crypto/sha1.html. Since then several things 54# have changed that made it interesting again: 55# 56# a) XMM units became faster and wider; 57# b) instruction set became more versatile; 58# c) an important observation was made by Max Locktykhin, which made 59# it possible to reduce amount of instructions required to perform 60# the operation in question, for further details see 61# http://software.intel.com/en-us/articles/improving-the-performance-of-the-secure-hash-algorithm-1/. 62 63# April 2011. 64# 65# Add AVX code path, probably most controversial... The thing is that 66# switch to AVX alone improves performance by as little as 4% in 67# comparison to SSSE3 code path. But below result doesn't look like 68# 4% improvement... Trouble is that Sandy Bridge decodes 'ro[rl]' as 69# pair of �-ops, and it's the additional �-ops, two per round, that 70# make it run slower than Core2 and Westmere. But 'sh[rl]d' is decoded 71# as single �-op by Sandy Bridge and it's replacing 'ro[rl]' with 72# equivalent 'sh[rl]d' that is responsible for the impressive 5.1 73# cycles per processed byte. But 'sh[rl]d' is not something that used 74# to be fast, nor does it appear to be fast in upcoming Bulldozer 75# [according to its optimization manual]. Which is why AVX code path 76# is guarded by *both* AVX and synthetic bit denoting Intel CPUs. 77# One can argue that it's unfair to AMD, but without 'sh[rl]d' it 78# makes no sense to keep the AVX code path. If somebody feels that 79# strongly, it's probably more appropriate to discuss possibility of 80# using vector rotate XOP on AMD... 81 82###################################################################### 83# Current performance is summarized in following table. Numbers are 84# CPU clock cycles spent to process single byte (less is better). 85# 86# x86 SSSE3 AVX 87# Pentium 15.7 - 88# PIII 11.5 - 89# P4 10.6 - 90# AMD K8 7.1 - 91# Core2 7.3 6.1/+20% - 92# Atom 12.5 9.5(*)/+32% - 93# Westmere 7.3 5.6/+30% - 94# Sandy Bridge 8.8 6.2/+40% 5.1(**)/+70% 95# 96# (*) Loop is 1056 instructions long and expected result is ~8.25. 97# It remains mystery [to me] why ILP is limited to 1.7. 98# 99# (**) As per above comment, the result is for AVX *plus* sh[rl]d. 100 101$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 102push(@INC,"${dir}","${dir}../../perlasm"); 103require "x86asm.pl"; 104 105&asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386"); 106 107$xmm=$ymm=0; 108for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); } 109 110$ymm=1 if ($xmm && 111 `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1` 112 =~ /GNU assembler version ([2-9]\.[0-9]+)/ && 113 $1>=2.19); # first version supporting AVX 114 115&external_label("OPENSSL_ia32cap_P") if ($xmm); 116 117 118$A="eax"; 119$B="ebx"; 120$C="ecx"; 121$D="edx"; 122$E="edi"; 123$T="esi"; 124$tmp1="ebp"; 125 126@V=($A,$B,$C,$D,$E,$T); 127 128$alt=0; # 1 denotes alternative IALU implementation, which performs 129 # 8% *worse* on P4, same on Westmere and Atom, 2% better on 130 # Sandy Bridge... 131 132sub BODY_00_15 133 { 134 local($n,$a,$b,$c,$d,$e,$f)=@_; 135 136 &comment("00_15 $n"); 137 138 &mov($f,$c); # f to hold F_00_19(b,c,d) 139 if ($n==0) { &mov($tmp1,$a); } 140 else { &mov($a,$tmp1); } 141 &rotl($tmp1,5); # tmp1=ROTATE(a,5) 142 &xor($f,$d); 143 &add($tmp1,$e); # tmp1+=e; 144 &mov($e,&swtmp($n%16)); # e becomes volatile and is loaded 145 # with xi, also note that e becomes 146 # f in next round... 147 &and($f,$b); 148 &rotr($b,2); # b=ROTATE(b,30) 149 &xor($f,$d); # f holds F_00_19(b,c,d) 150 &lea($tmp1,&DWP(0x5a827999,$tmp1,$e)); # tmp1+=K_00_19+xi 151 152 if ($n==15) { &mov($e,&swtmp(($n+1)%16));# pre-fetch f for next round 153 &add($f,$tmp1); } # f+=tmp1 154 else { &add($tmp1,$f); } # f becomes a in next round 155 &mov($tmp1,$a) if ($alt && $n==15); 156 } 157 158sub BODY_16_19 159 { 160 local($n,$a,$b,$c,$d,$e,$f)=@_; 161 162 &comment("16_19 $n"); 163 164if ($alt) { 165 &xor($c,$d); 166 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 167 &and($tmp1,$c); # tmp1 to hold F_00_19(b,c,d), b&=c^d 168 &xor($f,&swtmp(($n+8)%16)); 169 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d) 170 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 171 &rotl($f,1); # f=ROTATE(f,1) 172 &add($e,$tmp1); # e+=F_00_19(b,c,d) 173 &xor($c,$d); # restore $c 174 &mov($tmp1,$a); # b in next round 175 &rotr($b,$n==16?2:7); # b=ROTATE(b,30) 176 &mov(&swtmp($n%16),$f); # xi=f 177 &rotl($a,5); # ROTATE(a,5) 178 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e 179 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round 180 &add($f,$a); # f+=ROTATE(a,5) 181} else { 182 &mov($tmp1,$c); # tmp1 to hold F_00_19(b,c,d) 183 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 184 &xor($tmp1,$d); 185 &xor($f,&swtmp(($n+8)%16)); 186 &and($tmp1,$b); 187 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 188 &rotl($f,1); # f=ROTATE(f,1) 189 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d) 190 &add($e,$tmp1); # e+=F_00_19(b,c,d) 191 &mov($tmp1,$a); 192 &rotr($b,2); # b=ROTATE(b,30) 193 &mov(&swtmp($n%16),$f); # xi=f 194 &rotl($tmp1,5); # ROTATE(a,5) 195 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e 196 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round 197 &add($f,$tmp1); # f+=ROTATE(a,5) 198} 199 } 200 201sub BODY_20_39 202 { 203 local($n,$a,$b,$c,$d,$e,$f)=@_; 204 local $K=($n<40)?0x6ed9eba1:0xca62c1d6; 205 206 &comment("20_39 $n"); 207 208if ($alt) { 209 &xor($tmp1,$c); # tmp1 to hold F_20_39(b,c,d), b^=c 210 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 211 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d) 212 &xor($f,&swtmp(($n+8)%16)); 213 &add($e,$tmp1); # e+=F_20_39(b,c,d) 214 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 215 &rotl($f,1); # f=ROTATE(f,1) 216 &mov($tmp1,$a); # b in next round 217 &rotr($b,7); # b=ROTATE(b,30) 218 &mov(&swtmp($n%16),$f) if($n<77);# xi=f 219 &rotl($a,5); # ROTATE(a,5) 220 &xor($b,$c) if($n==39);# warm up for BODY_40_59 221 &and($tmp1,$b) if($n==39); 222 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY 223 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round 224 &add($f,$a); # f+=ROTATE(a,5) 225 &rotr($a,5) if ($n==79); 226} else { 227 &mov($tmp1,$b); # tmp1 to hold F_20_39(b,c,d) 228 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 229 &xor($tmp1,$c); 230 &xor($f,&swtmp(($n+8)%16)); 231 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d) 232 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 233 &rotl($f,1); # f=ROTATE(f,1) 234 &add($e,$tmp1); # e+=F_20_39(b,c,d) 235 &rotr($b,2); # b=ROTATE(b,30) 236 &mov($tmp1,$a); 237 &rotl($tmp1,5); # ROTATE(a,5) 238 &mov(&swtmp($n%16),$f) if($n<77);# xi=f 239 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY 240 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round 241 &add($f,$tmp1); # f+=ROTATE(a,5) 242} 243 } 244 245sub BODY_40_59 246 { 247 local($n,$a,$b,$c,$d,$e,$f)=@_; 248 249 &comment("40_59 $n"); 250 251if ($alt) { 252 &add($e,$tmp1); # e+=b&(c^d) 253 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 254 &mov($tmp1,$d); 255 &xor($f,&swtmp(($n+8)%16)); 256 &xor($c,$d); # restore $c 257 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 258 &rotl($f,1); # f=ROTATE(f,1) 259 &and($tmp1,$c); 260 &rotr($b,7); # b=ROTATE(b,30) 261 &add($e,$tmp1); # e+=c&d 262 &mov($tmp1,$a); # b in next round 263 &mov(&swtmp($n%16),$f); # xi=f 264 &rotl($a,5); # ROTATE(a,5) 265 &xor($b,$c) if ($n<59); 266 &and($tmp1,$b) if ($n<59);# tmp1 to hold F_40_59(b,c,d) 267 &lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e+(b&(c^d)) 268 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round 269 &add($f,$a); # f+=ROTATE(a,5) 270} else { 271 &mov($tmp1,$c); # tmp1 to hold F_40_59(b,c,d) 272 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 273 &xor($tmp1,$d); 274 &xor($f,&swtmp(($n+8)%16)); 275 &and($tmp1,$b); 276 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 277 &rotl($f,1); # f=ROTATE(f,1) 278 &add($tmp1,$e); # b&(c^d)+=e 279 &rotr($b,2); # b=ROTATE(b,30) 280 &mov($e,$a); # e becomes volatile 281 &rotl($e,5); # ROTATE(a,5) 282 &mov(&swtmp($n%16),$f); # xi=f 283 &lea($f,&DWP(0x8f1bbcdc,$f,$tmp1));# f+=K_40_59+e+(b&(c^d)) 284 &mov($tmp1,$c); 285 &add($f,$e); # f+=ROTATE(a,5) 286 &and($tmp1,$d); 287 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round 288 &add($f,$tmp1); # f+=c&d 289} 290 } 291 292&function_begin("sha1_block_data_order"); 293if ($xmm) { 294 &static_label("ssse3_shortcut"); 295 &static_label("avx_shortcut") if ($ymm); 296 &static_label("K_XX_XX"); 297 298 &call (&label("pic_point")); # make it PIC! 299 &set_label("pic_point"); 300 &blindpop($tmp1); 301 &picmeup($T,"OPENSSL_ia32cap_P",$tmp1,&label("pic_point")); 302 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1)); 303 304 &mov ($A,&DWP(0,$T)); 305 &mov ($D,&DWP(4,$T)); 306 &test ($D,"\$IA32CAP_MASK1_SSSE3"); # check SSSE3 bit 307 &jz (&label("x86")); 308 &test ($A,"\$IA32CAP_MASK0_FXSR"); # check FXSR bit 309 &jz (&label("x86")); 310 if ($ymm) { 311 &and ($D,"\$IA32CAP_MASK1_AVX"); # mask AVX bit 312 &and ($A,"\$IA32CAP_MASK0_INTEL"); # mask "Intel CPU" bit 313 &or ($A,$D); 314 &cmp ($A,"\$(IA32CAP_MASK1_AVX | IA32CAP_MASK0_INTEL)"); 315 &je (&label("avx_shortcut")); 316 } 317 &jmp (&label("ssse3_shortcut")); 318 &set_label("x86",16); 319} 320 &mov($tmp1,&wparam(0)); # SHA_CTX *c 321 &mov($T,&wparam(1)); # const void *input 322 &mov($A,&wparam(2)); # size_t num 323 &stack_push(16+3); # allocate X[16] 324 &shl($A,6); 325 &add($A,$T); 326 &mov(&wparam(2),$A); # pointer beyond the end of input 327 &mov($E,&DWP(16,$tmp1));# pre-load E 328 &jmp(&label("loop")); 329 330&set_label("loop",16); 331 332 # copy input chunk to X, but reversing byte order! 333 for ($i=0; $i<16; $i+=4) 334 { 335 &mov($A,&DWP(4*($i+0),$T)); 336 &mov($B,&DWP(4*($i+1),$T)); 337 &mov($C,&DWP(4*($i+2),$T)); 338 &mov($D,&DWP(4*($i+3),$T)); 339 &bswap($A); 340 &bswap($B); 341 &bswap($C); 342 &bswap($D); 343 &mov(&swtmp($i+0),$A); 344 &mov(&swtmp($i+1),$B); 345 &mov(&swtmp($i+2),$C); 346 &mov(&swtmp($i+3),$D); 347 } 348 &mov(&wparam(1),$T); # redundant in 1st spin 349 350 &mov($A,&DWP(0,$tmp1)); # load SHA_CTX 351 &mov($B,&DWP(4,$tmp1)); 352 &mov($C,&DWP(8,$tmp1)); 353 &mov($D,&DWP(12,$tmp1)); 354 # E is pre-loaded 355 356 for($i=0;$i<16;$i++) { &BODY_00_15($i,@V); unshift(@V,pop(@V)); } 357 for(;$i<20;$i++) { &BODY_16_19($i,@V); unshift(@V,pop(@V)); } 358 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); } 359 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); } 360 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); } 361 362 (($V[5] eq $D) and ($V[0] eq $E)) or die; # double-check 363 364 &mov($tmp1,&wparam(0)); # re-load SHA_CTX* 365 &mov($D,&wparam(1)); # D is last "T" and is discarded 366 367 &add($E,&DWP(0,$tmp1)); # E is last "A"... 368 &add($T,&DWP(4,$tmp1)); 369 &add($A,&DWP(8,$tmp1)); 370 &add($B,&DWP(12,$tmp1)); 371 &add($C,&DWP(16,$tmp1)); 372 373 &mov(&DWP(0,$tmp1),$E); # update SHA_CTX 374 &add($D,64); # advance input pointer 375 &mov(&DWP(4,$tmp1),$T); 376 &cmp($D,&wparam(2)); # have we reached the end yet? 377 &mov(&DWP(8,$tmp1),$A); 378 &mov($E,$C); # C is last "E" which needs to be "pre-loaded" 379 &mov(&DWP(12,$tmp1),$B); 380 &mov($T,$D); # input pointer 381 &mov(&DWP(16,$tmp1),$C); 382 &jb(&label("loop")); 383 384 &stack_pop(16+3); 385&function_end("sha1_block_data_order"); 386 387if ($xmm) { 388###################################################################### 389# The SSSE3 implementation. 390# 391# %xmm[0-7] are used as ring @X[] buffer containing quadruples of last 392# 32 elements of the message schedule or Xupdate outputs. First 4 393# quadruples are simply byte-swapped input, next 4 are calculated 394# according to method originally suggested by Dean Gaudet (modulo 395# being implemented in SSSE3). Once 8 quadruples or 32 elements are 396# collected, it switches to routine proposed by Max Locktyukhin. 397# 398# Calculations inevitably require temporary reqisters, and there are 399# no %xmm registers left to spare. For this reason part of the ring 400# buffer, X[2..4] to be specific, is offloaded to 3 quadriples ring 401# buffer on the stack. Keep in mind that X[2] is alias X[-6], X[3] - 402# X[-5], and X[4] - X[-4]... 403# 404# Another notable optimization is aggressive stack frame compression 405# aiming to minimize amount of 9-byte instructions... 406# 407# Yet another notable optimization is "jumping" $B variable. It means 408# that there is no register permanently allocated for $B value. This 409# allowed to eliminate one instruction from body_20_39... 410# 411my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded 412my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4 413my @V=($A,$B,$C,$D,$E); 414my $j=0; # hash round 415my @T=($T,$tmp1); 416my $inp; 417 418my $_rol=sub { &rol(@_) }; 419my $_ror=sub { &ror(@_) }; 420 421&function_begin("_sha1_block_data_order_ssse3"); 422 &call (&label("pic_point")); # make it PIC! 423 &set_label("pic_point"); 424 &blindpop($tmp1); 425 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1)); 426&set_label("ssse3_shortcut"); 427 428 &movdqa (@X[3],&QWP(0,$tmp1)); # K_00_19 429 &movdqa (@X[4],&QWP(16,$tmp1)); # K_20_39 430 &movdqa (@X[5],&QWP(32,$tmp1)); # K_40_59 431 &movdqa (@X[6],&QWP(48,$tmp1)); # K_60_79 432 &movdqa (@X[2],&QWP(64,$tmp1)); # pbswap mask 433 434 &mov ($E,&wparam(0)); # load argument block 435 &mov ($inp=@T[1],&wparam(1)); 436 &mov ($D,&wparam(2)); 437 &mov (@T[0],"esp"); 438 439 # stack frame layout 440 # 441 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area 442 # X[4]+K X[5]+K X[6]+K X[7]+K 443 # X[8]+K X[9]+K X[10]+K X[11]+K 444 # X[12]+K X[13]+K X[14]+K X[15]+K 445 # 446 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area 447 # X[4] X[5] X[6] X[7] 448 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19 449 # 450 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants 451 # K_40_59 K_40_59 K_40_59 K_40_59 452 # K_60_79 K_60_79 K_60_79 K_60_79 453 # K_00_19 K_00_19 K_00_19 K_00_19 454 # pbswap mask 455 # 456 # +192 ctx # argument block 457 # +196 inp 458 # +200 end 459 # +204 esp 460 &sub ("esp",208); 461 &and ("esp",-64); 462 463 &movdqa (&QWP(112+0,"esp"),@X[4]); # copy constants 464 &movdqa (&QWP(112+16,"esp"),@X[5]); 465 &movdqa (&QWP(112+32,"esp"),@X[6]); 466 &shl ($D,6); # len*64 467 &movdqa (&QWP(112+48,"esp"),@X[3]); 468 &add ($D,$inp); # end of input 469 &movdqa (&QWP(112+64,"esp"),@X[2]); 470 &add ($inp,64); 471 &mov (&DWP(192+0,"esp"),$E); # save argument block 472 &mov (&DWP(192+4,"esp"),$inp); 473 &mov (&DWP(192+8,"esp"),$D); 474 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp 475 476 &mov ($A,&DWP(0,$E)); # load context 477 &mov ($B,&DWP(4,$E)); 478 &mov ($C,&DWP(8,$E)); 479 &mov ($D,&DWP(12,$E)); 480 &mov ($E,&DWP(16,$E)); 481 &mov (@T[0],$B); # magic seed 482 483 &movdqu (@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3] 484 &movdqu (@X[-3&7],&QWP(-48,$inp)); 485 &movdqu (@X[-2&7],&QWP(-32,$inp)); 486 &movdqu (@X[-1&7],&QWP(-16,$inp)); 487 &pshufb (@X[-4&7],@X[2]); # byte swap 488 &pshufb (@X[-3&7],@X[2]); 489 &pshufb (@X[-2&7],@X[2]); 490 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot 491 &pshufb (@X[-1&7],@X[2]); 492 &paddd (@X[-4&7],@X[3]); # add K_00_19 493 &paddd (@X[-3&7],@X[3]); 494 &paddd (@X[-2&7],@X[3]); 495 &movdqa (&QWP(0,"esp"),@X[-4&7]); # X[]+K xfer to IALU 496 &psubd (@X[-4&7],@X[3]); # restore X[] 497 &movdqa (&QWP(0+16,"esp"),@X[-3&7]); 498 &psubd (@X[-3&7],@X[3]); 499 &movdqa (&QWP(0+32,"esp"),@X[-2&7]); 500 &psubd (@X[-2&7],@X[3]); 501 &movdqa (@X[0],@X[-3&7]); 502 &jmp (&label("loop")); 503 504###################################################################### 505# SSE instruction sequence is first broken to groups of independent 506# instructions, independent in respect to their inputs and shifter 507# (not all architectures have more than one). Then IALU instructions 508# are "knitted in" between the SSE groups. Distance is maintained for 509# SSE latency of 2 in hope that it fits better upcoming AMD Bulldozer 510# [which allegedly also implements SSSE3]... 511# 512# Temporary registers usage. X[2] is volatile at the entry and at the 513# end is restored from backtrace ring buffer. X[3] is expected to 514# contain current K_XX_XX constant and is used to caclulate X[-1]+K 515# from previous round, it becomes volatile the moment the value is 516# saved to stack for transfer to IALU. X[4] becomes volatile whenever 517# X[-4] is accumulated and offloaded to backtrace ring buffer, at the 518# end it is loaded with next K_XX_XX [which becomes X[3] in next 519# round]... 520# 521sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4 522{ use integer; 523 my $body = shift; 524 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions 525 my ($a,$b,$c,$d,$e); 526 527 eval(shift(@insns)); 528 eval(shift(@insns)); 529 &palignr(@X[0],@X[-4&7],8); # compose "X[-14]" in "X[0]" 530 &movdqa (@X[2],@X[-1&7]); 531 eval(shift(@insns)); 532 eval(shift(@insns)); 533 534 &paddd (@X[3],@X[-1&7]); 535 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer 536 eval(shift(@insns)); 537 eval(shift(@insns)); 538 &psrldq (@X[2],4); # "X[-3]", 3 dwords 539 eval(shift(@insns)); 540 eval(shift(@insns)); 541 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]" 542 eval(shift(@insns)); 543 eval(shift(@insns)); 544 545 &pxor (@X[2],@X[-2&7]); # "X[-3]"^"X[-8]" 546 eval(shift(@insns)); 547 eval(shift(@insns)); 548 eval(shift(@insns)); 549 eval(shift(@insns)); 550 551 &pxor (@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]" 552 eval(shift(@insns)); 553 eval(shift(@insns)); 554 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU 555 eval(shift(@insns)); 556 eval(shift(@insns)); 557 558 &movdqa (@X[4],@X[0]); 559 &movdqa (@X[2],@X[0]); 560 eval(shift(@insns)); 561 eval(shift(@insns)); 562 eval(shift(@insns)); 563 eval(shift(@insns)); 564 565 &pslldq (@X[4],12); # "X[0]"<<96, extract one dword 566 &paddd (@X[0],@X[0]); 567 eval(shift(@insns)); 568 eval(shift(@insns)); 569 eval(shift(@insns)); 570 eval(shift(@insns)); 571 572 &psrld (@X[2],31); 573 eval(shift(@insns)); 574 eval(shift(@insns)); 575 &movdqa (@X[3],@X[4]); 576 eval(shift(@insns)); 577 eval(shift(@insns)); 578 579 &psrld (@X[4],30); 580 &por (@X[0],@X[2]); # "X[0]"<<<=1 581 eval(shift(@insns)); 582 eval(shift(@insns)); 583 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer 584 eval(shift(@insns)); 585 eval(shift(@insns)); 586 587 &pslld (@X[3],2); 588 &pxor (@X[0],@X[4]); 589 eval(shift(@insns)); 590 eval(shift(@insns)); 591 &movdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX 592 eval(shift(@insns)); 593 eval(shift(@insns)); 594 595 &pxor (@X[0],@X[3]); # "X[0]"^=("X[0]"<<96)<<<2 596 &movdqa (@X[1],@X[-2&7]) if ($Xi<7); 597 eval(shift(@insns)); 598 eval(shift(@insns)); 599 600 foreach (@insns) { eval; } # remaining instructions [if any] 601 602 $Xi++; push(@X,shift(@X)); # "rotate" X[] 603} 604 605sub Xupdate_ssse3_32_79() 606{ use integer; 607 my $body = shift; 608 my @insns = (&$body,&$body,&$body,&$body); # 32 to 48 instructions 609 my ($a,$b,$c,$d,$e); 610 611 &movdqa (@X[2],@X[-1&7]) if ($Xi==8); 612 eval(shift(@insns)); # body_20_39 613 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" 614 &palignr(@X[2],@X[-2&7],8); # compose "X[-6]" 615 eval(shift(@insns)); 616 eval(shift(@insns)); 617 eval(shift(@insns)); # rol 618 619 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]" 620 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer 621 eval(shift(@insns)); 622 eval(shift(@insns)); 623 if ($Xi%5) { 624 &movdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX... 625 } else { # ... or load next one 626 &movdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp")); 627 } 628 &paddd (@X[3],@X[-1&7]); 629 eval(shift(@insns)); # ror 630 eval(shift(@insns)); 631 632 &pxor (@X[0],@X[2]); # "X[0]"^="X[-6]" 633 eval(shift(@insns)); # body_20_39 634 eval(shift(@insns)); 635 eval(shift(@insns)); 636 eval(shift(@insns)); # rol 637 638 &movdqa (@X[2],@X[0]); 639 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU 640 eval(shift(@insns)); 641 eval(shift(@insns)); 642 eval(shift(@insns)); # ror 643 eval(shift(@insns)); 644 645 &pslld (@X[0],2); 646 eval(shift(@insns)); # body_20_39 647 eval(shift(@insns)); 648 &psrld (@X[2],30); 649 eval(shift(@insns)); 650 eval(shift(@insns)); # rol 651 eval(shift(@insns)); 652 eval(shift(@insns)); 653 eval(shift(@insns)); # ror 654 eval(shift(@insns)); 655 656 &por (@X[0],@X[2]); # "X[0]"<<<=2 657 eval(shift(@insns)); # body_20_39 658 eval(shift(@insns)); 659 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer 660 eval(shift(@insns)); 661 eval(shift(@insns)); # rol 662 eval(shift(@insns)); 663 eval(shift(@insns)); 664 eval(shift(@insns)); # ror 665 &movdqa (@X[3],@X[0]) if ($Xi<19); 666 eval(shift(@insns)); 667 668 foreach (@insns) { eval; } # remaining instructions 669 670 $Xi++; push(@X,shift(@X)); # "rotate" X[] 671} 672 673sub Xuplast_ssse3_80() 674{ use integer; 675 my $body = shift; 676 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 677 my ($a,$b,$c,$d,$e); 678 679 eval(shift(@insns)); 680 &paddd (@X[3],@X[-1&7]); 681 eval(shift(@insns)); 682 eval(shift(@insns)); 683 eval(shift(@insns)); 684 eval(shift(@insns)); 685 686 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU 687 688 foreach (@insns) { eval; } # remaining instructions 689 690 &mov ($inp=@T[1],&DWP(192+4,"esp")); 691 &cmp ($inp,&DWP(192+8,"esp")); 692 &je (&label("done")); 693 694 &movdqa (@X[3],&QWP(112+48,"esp")); # K_00_19 695 &movdqa (@X[2],&QWP(112+64,"esp")); # pbswap mask 696 &movdqu (@X[-4&7],&QWP(0,$inp)); # load input 697 &movdqu (@X[-3&7],&QWP(16,$inp)); 698 &movdqu (@X[-2&7],&QWP(32,$inp)); 699 &movdqu (@X[-1&7],&QWP(48,$inp)); 700 &add ($inp,64); 701 &pshufb (@X[-4&7],@X[2]); # byte swap 702 &mov (&DWP(192+4,"esp"),$inp); 703 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot 704 705 $Xi=0; 706} 707 708sub Xloop_ssse3() 709{ use integer; 710 my $body = shift; 711 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 712 my ($a,$b,$c,$d,$e); 713 714 eval(shift(@insns)); 715 eval(shift(@insns)); 716 &pshufb (@X[($Xi-3)&7],@X[2]); 717 eval(shift(@insns)); 718 eval(shift(@insns)); 719 &paddd (@X[($Xi-4)&7],@X[3]); 720 eval(shift(@insns)); 721 eval(shift(@insns)); 722 eval(shift(@insns)); 723 eval(shift(@insns)); 724 &movdqa (&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]); # X[]+K xfer to IALU 725 eval(shift(@insns)); 726 eval(shift(@insns)); 727 &psubd (@X[($Xi-4)&7],@X[3]); 728 729 foreach (@insns) { eval; } 730 $Xi++; 731} 732 733sub Xtail_ssse3() 734{ use integer; 735 my $body = shift; 736 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 737 my ($a,$b,$c,$d,$e); 738 739 foreach (@insns) { eval; } 740} 741 742sub body_00_19 () { 743 ( 744 '($a,$b,$c,$d,$e)=@V;'. 745 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer 746 '&xor ($c,$d);', 747 '&mov (@T[1],$a);', # $b in next round 748 '&$_rol ($a,5);', 749 '&and (@T[0],$c);', # ($b&($c^$d)) 750 '&xor ($c,$d);', # restore $c 751 '&xor (@T[0],$d);', 752 '&add ($e,$a);', 753 '&$_ror ($b,$j?7:2);', # $b>>>2 754 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' 755 ); 756} 757 758sub body_20_39 () { 759 ( 760 '($a,$b,$c,$d,$e)=@V;'. 761 '&add ($e,&DWP(4*($j++&15),"esp"));', # X[]+K xfer 762 '&xor (@T[0],$d);', # ($b^$d) 763 '&mov (@T[1],$a);', # $b in next round 764 '&$_rol ($a,5);', 765 '&xor (@T[0],$c);', # ($b^$d^$c) 766 '&add ($e,$a);', 767 '&$_ror ($b,7);', # $b>>>2 768 '&add ($e,@T[0]);' .'unshift(@V,pop(@V)); unshift(@T,pop(@T));' 769 ); 770} 771 772sub body_40_59 () { 773 ( 774 '($a,$b,$c,$d,$e)=@V;'. 775 '&mov (@T[1],$c);', 776 '&xor ($c,$d);', 777 '&add ($e,&DWP(4*($j++&15),"esp"));', # X[]+K xfer 778 '&and (@T[1],$d);', 779 '&and (@T[0],$c);', # ($b&($c^$d)) 780 '&$_ror ($b,7);', # $b>>>2 781 '&add ($e,@T[1]);', 782 '&mov (@T[1],$a);', # $b in next round 783 '&$_rol ($a,5);', 784 '&add ($e,@T[0]);', 785 '&xor ($c,$d);', # restore $c 786 '&add ($e,$a);' .'unshift(@V,pop(@V)); unshift(@T,pop(@T));' 787 ); 788} 789 790&set_label("loop",16); 791 &Xupdate_ssse3_16_31(\&body_00_19); 792 &Xupdate_ssse3_16_31(\&body_00_19); 793 &Xupdate_ssse3_16_31(\&body_00_19); 794 &Xupdate_ssse3_16_31(\&body_00_19); 795 &Xupdate_ssse3_32_79(\&body_00_19); 796 &Xupdate_ssse3_32_79(\&body_20_39); 797 &Xupdate_ssse3_32_79(\&body_20_39); 798 &Xupdate_ssse3_32_79(\&body_20_39); 799 &Xupdate_ssse3_32_79(\&body_20_39); 800 &Xupdate_ssse3_32_79(\&body_20_39); 801 &Xupdate_ssse3_32_79(\&body_40_59); 802 &Xupdate_ssse3_32_79(\&body_40_59); 803 &Xupdate_ssse3_32_79(\&body_40_59); 804 &Xupdate_ssse3_32_79(\&body_40_59); 805 &Xupdate_ssse3_32_79(\&body_40_59); 806 &Xupdate_ssse3_32_79(\&body_20_39); 807 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done" 808 809 $saved_j=$j; @saved_V=@V; 810 811 &Xloop_ssse3(\&body_20_39); 812 &Xloop_ssse3(\&body_20_39); 813 &Xloop_ssse3(\&body_20_39); 814 815 &mov (@T[1],&DWP(192,"esp")); # update context 816 &add ($A,&DWP(0,@T[1])); 817 &add (@T[0],&DWP(4,@T[1])); # $b 818 &add ($C,&DWP(8,@T[1])); 819 &mov (&DWP(0,@T[1]),$A); 820 &add ($D,&DWP(12,@T[1])); 821 &mov (&DWP(4,@T[1]),@T[0]); 822 &add ($E,&DWP(16,@T[1])); 823 &mov (&DWP(8,@T[1]),$C); 824 &mov ($B,@T[0]); 825 &mov (&DWP(12,@T[1]),$D); 826 &mov (&DWP(16,@T[1]),$E); 827 &movdqa (@X[0],@X[-3&7]); 828 829 &jmp (&label("loop")); 830 831&set_label("done",16); $j=$saved_j; @V=@saved_V; 832 833 &Xtail_ssse3(\&body_20_39); 834 &Xtail_ssse3(\&body_20_39); 835 &Xtail_ssse3(\&body_20_39); 836 837 &mov (@T[1],&DWP(192,"esp")); # update context 838 &add ($A,&DWP(0,@T[1])); 839 &mov ("esp",&DWP(192+12,"esp")); # restore %esp 840 &add (@T[0],&DWP(4,@T[1])); # $b 841 &add ($C,&DWP(8,@T[1])); 842 &mov (&DWP(0,@T[1]),$A); 843 &add ($D,&DWP(12,@T[1])); 844 &mov (&DWP(4,@T[1]),@T[0]); 845 &add ($E,&DWP(16,@T[1])); 846 &mov (&DWP(8,@T[1]),$C); 847 &mov (&DWP(12,@T[1]),$D); 848 &mov (&DWP(16,@T[1]),$E); 849 850&function_end("_sha1_block_data_order_ssse3"); 851 852if ($ymm) { 853my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded 854my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4 855my @V=($A,$B,$C,$D,$E); 856my $j=0; # hash round 857my @T=($T,$tmp1); 858my $inp; 859 860my $_rol=sub { &shld(@_[0],@_) }; 861my $_ror=sub { &shrd(@_[0],@_) }; 862 863&function_begin("_sha1_block_data_order_avx"); 864 &call (&label("pic_point")); # make it PIC! 865 &set_label("pic_point"); 866 &blindpop($tmp1); 867 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1)); 868&set_label("avx_shortcut"); 869 &vzeroall(); 870 871 &vmovdqa(@X[3],&QWP(0,$tmp1)); # K_00_19 872 &vmovdqa(@X[4],&QWP(16,$tmp1)); # K_20_39 873 &vmovdqa(@X[5],&QWP(32,$tmp1)); # K_40_59 874 &vmovdqa(@X[6],&QWP(48,$tmp1)); # K_60_79 875 &vmovdqa(@X[2],&QWP(64,$tmp1)); # pbswap mask 876 877 &mov ($E,&wparam(0)); # load argument block 878 &mov ($inp=@T[1],&wparam(1)); 879 &mov ($D,&wparam(2)); 880 &mov (@T[0],"esp"); 881 882 # stack frame layout 883 # 884 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area 885 # X[4]+K X[5]+K X[6]+K X[7]+K 886 # X[8]+K X[9]+K X[10]+K X[11]+K 887 # X[12]+K X[13]+K X[14]+K X[15]+K 888 # 889 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area 890 # X[4] X[5] X[6] X[7] 891 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19 892 # 893 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants 894 # K_40_59 K_40_59 K_40_59 K_40_59 895 # K_60_79 K_60_79 K_60_79 K_60_79 896 # K_00_19 K_00_19 K_00_19 K_00_19 897 # pbswap mask 898 # 899 # +192 ctx # argument block 900 # +196 inp 901 # +200 end 902 # +204 esp 903 &sub ("esp",208); 904 &and ("esp",-64); 905 906 &vmovdqa(&QWP(112+0,"esp"),@X[4]); # copy constants 907 &vmovdqa(&QWP(112+16,"esp"),@X[5]); 908 &vmovdqa(&QWP(112+32,"esp"),@X[6]); 909 &shl ($D,6); # len*64 910 &vmovdqa(&QWP(112+48,"esp"),@X[3]); 911 &add ($D,$inp); # end of input 912 &vmovdqa(&QWP(112+64,"esp"),@X[2]); 913 &add ($inp,64); 914 &mov (&DWP(192+0,"esp"),$E); # save argument block 915 &mov (&DWP(192+4,"esp"),$inp); 916 &mov (&DWP(192+8,"esp"),$D); 917 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp 918 919 &mov ($A,&DWP(0,$E)); # load context 920 &mov ($B,&DWP(4,$E)); 921 &mov ($C,&DWP(8,$E)); 922 &mov ($D,&DWP(12,$E)); 923 &mov ($E,&DWP(16,$E)); 924 &mov (@T[0],$B); # magic seed 925 926 &vmovdqu(@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3] 927 &vmovdqu(@X[-3&7],&QWP(-48,$inp)); 928 &vmovdqu(@X[-2&7],&QWP(-32,$inp)); 929 &vmovdqu(@X[-1&7],&QWP(-16,$inp)); 930 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap 931 &vpshufb(@X[-3&7],@X[-3&7],@X[2]); 932 &vpshufb(@X[-2&7],@X[-2&7],@X[2]); 933 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot 934 &vpshufb(@X[-1&7],@X[-1&7],@X[2]); 935 &vpaddd (@X[0],@X[-4&7],@X[3]); # add K_00_19 936 &vpaddd (@X[1],@X[-3&7],@X[3]); 937 &vpaddd (@X[2],@X[-2&7],@X[3]); 938 &vmovdqa(&QWP(0,"esp"),@X[0]); # X[]+K xfer to IALU 939 &vmovdqa(&QWP(0+16,"esp"),@X[1]); 940 &vmovdqa(&QWP(0+32,"esp"),@X[2]); 941 &jmp (&label("loop")); 942 943sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4 944{ use integer; 945 my $body = shift; 946 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions 947 my ($a,$b,$c,$d,$e); 948 949 eval(shift(@insns)); 950 eval(shift(@insns)); 951 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]" 952 eval(shift(@insns)); 953 eval(shift(@insns)); 954 955 &vpaddd (@X[3],@X[3],@X[-1&7]); 956 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer 957 eval(shift(@insns)); 958 eval(shift(@insns)); 959 &vpsrldq(@X[2],@X[-1&7],4); # "X[-3]", 3 dwords 960 eval(shift(@insns)); 961 eval(shift(@insns)); 962 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]" 963 eval(shift(@insns)); 964 eval(shift(@insns)); 965 966 &vpxor (@X[2],@X[2],@X[-2&7]); # "X[-3]"^"X[-8]" 967 eval(shift(@insns)); 968 eval(shift(@insns)); 969 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU 970 eval(shift(@insns)); 971 eval(shift(@insns)); 972 973 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]" 974 eval(shift(@insns)); 975 eval(shift(@insns)); 976 eval(shift(@insns)); 977 eval(shift(@insns)); 978 979 &vpsrld (@X[2],@X[0],31); 980 eval(shift(@insns)); 981 eval(shift(@insns)); 982 eval(shift(@insns)); 983 eval(shift(@insns)); 984 985 &vpslldq(@X[4],@X[0],12); # "X[0]"<<96, extract one dword 986 &vpaddd (@X[0],@X[0],@X[0]); 987 eval(shift(@insns)); 988 eval(shift(@insns)); 989 eval(shift(@insns)); 990 eval(shift(@insns)); 991 992 &vpsrld (@X[3],@X[4],30); 993 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=1 994 eval(shift(@insns)); 995 eval(shift(@insns)); 996 eval(shift(@insns)); 997 eval(shift(@insns)); 998 999 &vpslld (@X[4],@X[4],2); 1000 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer 1001 eval(shift(@insns)); 1002 eval(shift(@insns)); 1003 &vpxor (@X[0],@X[0],@X[3]); 1004 eval(shift(@insns)); 1005 eval(shift(@insns)); 1006 eval(shift(@insns)); 1007 eval(shift(@insns)); 1008 1009 &vpxor (@X[0],@X[0],@X[4]); # "X[0]"^=("X[0]"<<96)<<<2 1010 eval(shift(@insns)); 1011 eval(shift(@insns)); 1012 &vmovdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX 1013 eval(shift(@insns)); 1014 eval(shift(@insns)); 1015 1016 foreach (@insns) { eval; } # remaining instructions [if any] 1017 1018 $Xi++; push(@X,shift(@X)); # "rotate" X[] 1019} 1020 1021sub Xupdate_avx_32_79() 1022{ use integer; 1023 my $body = shift; 1024 my @insns = (&$body,&$body,&$body,&$body); # 32 to 48 instructions 1025 my ($a,$b,$c,$d,$e); 1026 1027 &vpalignr(@X[2],@X[-1&7],@X[-2&7],8); # compose "X[-6]" 1028 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" 1029 eval(shift(@insns)); # body_20_39 1030 eval(shift(@insns)); 1031 eval(shift(@insns)); 1032 eval(shift(@insns)); # rol 1033 1034 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]" 1035 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer 1036 eval(shift(@insns)); 1037 eval(shift(@insns)); 1038 if ($Xi%5) { 1039 &vmovdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX... 1040 } else { # ... or load next one 1041 &vmovdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp")); 1042 } 1043 &vpaddd (@X[3],@X[3],@X[-1&7]); 1044 eval(shift(@insns)); # ror 1045 eval(shift(@insns)); 1046 1047 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-6]" 1048 eval(shift(@insns)); # body_20_39 1049 eval(shift(@insns)); 1050 eval(shift(@insns)); 1051 eval(shift(@insns)); # rol 1052 1053 &vpsrld (@X[2],@X[0],30); 1054 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU 1055 eval(shift(@insns)); 1056 eval(shift(@insns)); 1057 eval(shift(@insns)); # ror 1058 eval(shift(@insns)); 1059 1060 &vpslld (@X[0],@X[0],2); 1061 eval(shift(@insns)); # body_20_39 1062 eval(shift(@insns)); 1063 eval(shift(@insns)); 1064 eval(shift(@insns)); # rol 1065 eval(shift(@insns)); 1066 eval(shift(@insns)); 1067 eval(shift(@insns)); # ror 1068 eval(shift(@insns)); 1069 1070 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=2 1071 eval(shift(@insns)); # body_20_39 1072 eval(shift(@insns)); 1073 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer 1074 eval(shift(@insns)); 1075 eval(shift(@insns)); # rol 1076 eval(shift(@insns)); 1077 eval(shift(@insns)); 1078 eval(shift(@insns)); # ror 1079 eval(shift(@insns)); 1080 1081 foreach (@insns) { eval; } # remaining instructions 1082 1083 $Xi++; push(@X,shift(@X)); # "rotate" X[] 1084} 1085 1086sub Xuplast_avx_80() 1087{ use integer; 1088 my $body = shift; 1089 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 1090 my ($a,$b,$c,$d,$e); 1091 1092 eval(shift(@insns)); 1093 &vpaddd (@X[3],@X[3],@X[-1&7]); 1094 eval(shift(@insns)); 1095 eval(shift(@insns)); 1096 eval(shift(@insns)); 1097 eval(shift(@insns)); 1098 1099 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU 1100 1101 foreach (@insns) { eval; } # remaining instructions 1102 1103 &mov ($inp=@T[1],&DWP(192+4,"esp")); 1104 &cmp ($inp,&DWP(192+8,"esp")); 1105 &je (&label("done")); 1106 1107 &vmovdqa(@X[3],&QWP(112+48,"esp")); # K_00_19 1108 &vmovdqa(@X[2],&QWP(112+64,"esp")); # pbswap mask 1109 &vmovdqu(@X[-4&7],&QWP(0,$inp)); # load input 1110 &vmovdqu(@X[-3&7],&QWP(16,$inp)); 1111 &vmovdqu(@X[-2&7],&QWP(32,$inp)); 1112 &vmovdqu(@X[-1&7],&QWP(48,$inp)); 1113 &add ($inp,64); 1114 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap 1115 &mov (&DWP(192+4,"esp"),$inp); 1116 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot 1117 1118 $Xi=0; 1119} 1120 1121sub Xloop_avx() 1122{ use integer; 1123 my $body = shift; 1124 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 1125 my ($a,$b,$c,$d,$e); 1126 1127 eval(shift(@insns)); 1128 eval(shift(@insns)); 1129 &vpshufb (@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]); 1130 eval(shift(@insns)); 1131 eval(shift(@insns)); 1132 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],@X[3]); 1133 eval(shift(@insns)); 1134 eval(shift(@insns)); 1135 eval(shift(@insns)); 1136 eval(shift(@insns)); 1137 &vmovdqa (&QWP(0+16*$Xi,"esp"),@X[$Xi&7]); # X[]+K xfer to IALU 1138 eval(shift(@insns)); 1139 eval(shift(@insns)); 1140 1141 foreach (@insns) { eval; } 1142 $Xi++; 1143} 1144 1145sub Xtail_avx() 1146{ use integer; 1147 my $body = shift; 1148 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 1149 my ($a,$b,$c,$d,$e); 1150 1151 foreach (@insns) { eval; } 1152} 1153 1154&set_label("loop",16); 1155 &Xupdate_avx_16_31(\&body_00_19); 1156 &Xupdate_avx_16_31(\&body_00_19); 1157 &Xupdate_avx_16_31(\&body_00_19); 1158 &Xupdate_avx_16_31(\&body_00_19); 1159 &Xupdate_avx_32_79(\&body_00_19); 1160 &Xupdate_avx_32_79(\&body_20_39); 1161 &Xupdate_avx_32_79(\&body_20_39); 1162 &Xupdate_avx_32_79(\&body_20_39); 1163 &Xupdate_avx_32_79(\&body_20_39); 1164 &Xupdate_avx_32_79(\&body_20_39); 1165 &Xupdate_avx_32_79(\&body_40_59); 1166 &Xupdate_avx_32_79(\&body_40_59); 1167 &Xupdate_avx_32_79(\&body_40_59); 1168 &Xupdate_avx_32_79(\&body_40_59); 1169 &Xupdate_avx_32_79(\&body_40_59); 1170 &Xupdate_avx_32_79(\&body_20_39); 1171 &Xuplast_avx_80(\&body_20_39); # can jump to "done" 1172 1173 $saved_j=$j; @saved_V=@V; 1174 1175 &Xloop_avx(\&body_20_39); 1176 &Xloop_avx(\&body_20_39); 1177 &Xloop_avx(\&body_20_39); 1178 1179 &mov (@T[1],&DWP(192,"esp")); # update context 1180 &add ($A,&DWP(0,@T[1])); 1181 &add (@T[0],&DWP(4,@T[1])); # $b 1182 &add ($C,&DWP(8,@T[1])); 1183 &mov (&DWP(0,@T[1]),$A); 1184 &add ($D,&DWP(12,@T[1])); 1185 &mov (&DWP(4,@T[1]),@T[0]); 1186 &add ($E,&DWP(16,@T[1])); 1187 &mov (&DWP(8,@T[1]),$C); 1188 &mov ($B,@T[0]); 1189 &mov (&DWP(12,@T[1]),$D); 1190 &mov (&DWP(16,@T[1]),$E); 1191 1192 &jmp (&label("loop")); 1193 1194&set_label("done",16); $j=$saved_j; @V=@saved_V; 1195 1196 &Xtail_avx(\&body_20_39); 1197 &Xtail_avx(\&body_20_39); 1198 &Xtail_avx(\&body_20_39); 1199 1200 &vzeroall(); 1201 1202 &mov (@T[1],&DWP(192,"esp")); # update context 1203 &add ($A,&DWP(0,@T[1])); 1204 &mov ("esp",&DWP(192+12,"esp")); # restore %esp 1205 &add (@T[0],&DWP(4,@T[1])); # $b 1206 &add ($C,&DWP(8,@T[1])); 1207 &mov (&DWP(0,@T[1]),$A); 1208 &add ($D,&DWP(12,@T[1])); 1209 &mov (&DWP(4,@T[1]),@T[0]); 1210 &add ($E,&DWP(16,@T[1])); 1211 &mov (&DWP(8,@T[1]),$C); 1212 &mov (&DWP(12,@T[1]),$D); 1213 &mov (&DWP(16,@T[1]),$E); 1214&function_end("_sha1_block_data_order_avx"); 1215} 1216&set_label("K_XX_XX",64); 1217&data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999); # K_00_19 1218&data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1); # K_20_39 1219&data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc); # K_40_59 1220&data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6); # K_60_79 1221&data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f); # pbswap mask 1222} 1223&asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>"); 1224 1225&asm_finish(); 1226