1#! /usr/bin/env perl 2# Copyright 1998-2016 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# [Re]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# "[Re]written" was achieved in two major overhauls. In 2004 BODY_* 18# functions were re-implemented to address P4 performance issue [see 19# commentary below], and in 2006 the rest was rewritten in order to 20# gain freedom to liberate licensing terms. 21 22# January, September 2004. 23# 24# It was noted that Intel IA-32 C compiler generates code which 25# performs ~30% *faster* on P4 CPU than original *hand-coded* 26# SHA1 assembler implementation. To address this problem (and 27# prove that humans are still better than machines:-), the 28# original code was overhauled, which resulted in following 29# performance changes: 30# 31# compared with original compared with Intel cc 32# assembler impl. generated code 33# Pentium -16% +48% 34# PIII/AMD +8% +16% 35# P4 +85%(!) +45% 36# 37# As you can see Pentium came out as looser:-( Yet I reckoned that 38# improvement on P4 outweighs the loss and incorporate this 39# re-tuned code to 0.9.7 and later. 40# ---------------------------------------------------------------- 41 42# August 2009. 43# 44# George Spelvin has tipped that F_40_59(b,c,d) can be rewritten as 45# '(c&d) + (b&(c^d))', which allows to accumulate partial results 46# and lighten "pressure" on scratch registers. This resulted in 47# >12% performance improvement on contemporary AMD cores (with no 48# degradation on other CPUs:-). Also, the code was revised to maximize 49# "distance" between instructions producing input to 'lea' instruction 50# and the 'lea' instruction itself, which is essential for Intel Atom 51# core and resulted in ~15% improvement. 52 53# October 2010. 54# 55# Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it 56# is to offload message schedule denoted by Wt in NIST specification, 57# or Xupdate in OpenSSL source, to SIMD unit. The idea is not novel, 58# and in SSE2 context was first explored by Dean Gaudet in 2004, see 59# http://arctic.org/~dean/crypto/sha1.html. Since then several things 60# have changed that made it interesting again: 61# 62# a) XMM units became faster and wider; 63# b) instruction set became more versatile; 64# c) an important observation was made by Max Locktykhin, which made 65# it possible to reduce amount of instructions required to perform 66# the operation in question, for further details see 67# http://software.intel.com/en-us/articles/improving-the-performance-of-the-secure-hash-algorithm-1/. 68 69# April 2011. 70# 71# Add AVX code path, probably most controversial... The thing is that 72# switch to AVX alone improves performance by as little as 4% in 73# comparison to SSSE3 code path. But below result doesn't look like 74# 4% improvement... Trouble is that Sandy Bridge decodes 'ro[rl]' as 75# pair of µ-ops, and it's the additional µ-ops, two per round, that 76# make it run slower than Core2 and Westmere. But 'sh[rl]d' is decoded 77# as single µ-op by Sandy Bridge and it's replacing 'ro[rl]' with 78# equivalent 'sh[rl]d' that is responsible for the impressive 5.1 79# cycles per processed byte. But 'sh[rl]d' is not something that used 80# to be fast, nor does it appear to be fast in upcoming Bulldozer 81# [according to its optimization manual]. Which is why AVX code path 82# is guarded by *both* AVX and synthetic bit denoting Intel CPUs. 83# One can argue that it's unfair to AMD, but without 'sh[rl]d' it 84# makes no sense to keep the AVX code path. If somebody feels that 85# strongly, it's probably more appropriate to discuss possibility of 86# using vector rotate XOP on AMD... 87 88# March 2014. 89# 90# Add support for Intel SHA Extensions. 91 92###################################################################### 93# Current performance is summarized in following table. Numbers are 94# CPU clock cycles spent to process single byte (less is better). 95# 96# x86 SSSE3 AVX 97# Pentium 15.7 - 98# PIII 11.5 - 99# P4 10.6 - 100# AMD K8 7.1 - 101# Core2 7.3 6.0/+22% - 102# Westmere 7.3 5.5/+33% - 103# Sandy Bridge 8.8 6.2/+40% 5.1(**)/+73% 104# Ivy Bridge 7.2 4.8/+51% 4.7(**)/+53% 105# Haswell 6.5 4.3/+51% 4.1(**)/+58% 106# Skylake 6.4 4.1/+55% 4.1(**)/+55% 107# Bulldozer 11.6 6.0/+92% 108# VIA Nano 10.6 7.5/+41% 109# Atom 12.5 9.3(*)/+35% 110# Silvermont 14.5 9.9(*)/+46% 111# Goldmont 8.8 6.7/+30% 1.7(***)/+415% 112# 113# (*) Loop is 1056 instructions long and expected result is ~8.25. 114# The discrepancy is because of front-end limitations, so 115# called MS-ROM penalties, and on Silvermont even rotate's 116# limited parallelism. 117# 118# (**) As per above comment, the result is for AVX *plus* sh[rl]d. 119# 120# (***) SHAEXT result 121 122$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 123push(@INC,"${dir}","${dir}../../../perlasm"); 124require "x86asm.pl"; 125 126$output=pop; 127open STDOUT,">$output"; 128 129&asm_init($ARGV[0]); 130 131$xmm = 1; 132 133# In upstream, this is controlled by shelling out to the compiler to check 134# versions, but BoringSSL is intended to be used with pre-generated perlasm 135# output, so this isn't useful anyway. 136$ymm = 1; 137 138$shaext=$xmm; ### set to zero if compiling for 1.0.1 139 140# TODO(davidben): Consider enabling the Intel SHA Extensions code once it's 141# been tested. 142$shaext = 0; 143 144 145$A="eax"; 146$B="ebx"; 147$C="ecx"; 148$D="edx"; 149$E="edi"; 150$T="esi"; 151$tmp1="ebp"; 152 153@V=($A,$B,$C,$D,$E,$T); 154 155$alt=0; # 1 denotes alternative IALU implementation, which performs 156 # 8% *worse* on P4, same on Westmere and Atom, 2% better on 157 # Sandy Bridge... 158 159sub BODY_00_15 160 { 161 local($n,$a,$b,$c,$d,$e,$f)=@_; 162 163 &comment("00_15 $n"); 164 165 &mov($f,$c); # f to hold F_00_19(b,c,d) 166 if ($n==0) { &mov($tmp1,$a); } 167 else { &mov($a,$tmp1); } 168 &rotl($tmp1,5); # tmp1=ROTATE(a,5) 169 &xor($f,$d); 170 &add($tmp1,$e); # tmp1+=e; 171 &mov($e,&swtmp($n%16)); # e becomes volatile and is loaded 172 # with xi, also note that e becomes 173 # f in next round... 174 &and($f,$b); 175 &rotr($b,2); # b=ROTATE(b,30) 176 &xor($f,$d); # f holds F_00_19(b,c,d) 177 &lea($tmp1,&DWP(0x5a827999,$tmp1,$e)); # tmp1+=K_00_19+xi 178 179 if ($n==15) { &mov($e,&swtmp(($n+1)%16));# pre-fetch f for next round 180 &add($f,$tmp1); } # f+=tmp1 181 else { &add($tmp1,$f); } # f becomes a in next round 182 &mov($tmp1,$a) if ($alt && $n==15); 183 } 184 185sub BODY_16_19 186 { 187 local($n,$a,$b,$c,$d,$e,$f)=@_; 188 189 &comment("16_19 $n"); 190 191if ($alt) { 192 &xor($c,$d); 193 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 194 &and($tmp1,$c); # tmp1 to hold F_00_19(b,c,d), b&=c^d 195 &xor($f,&swtmp(($n+8)%16)); 196 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d) 197 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 198 &rotl($f,1); # f=ROTATE(f,1) 199 &add($e,$tmp1); # e+=F_00_19(b,c,d) 200 &xor($c,$d); # restore $c 201 &mov($tmp1,$a); # b in next round 202 &rotr($b,$n==16?2:7); # b=ROTATE(b,30) 203 &mov(&swtmp($n%16),$f); # xi=f 204 &rotl($a,5); # ROTATE(a,5) 205 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e 206 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round 207 &add($f,$a); # f+=ROTATE(a,5) 208} else { 209 &mov($tmp1,$c); # tmp1 to hold F_00_19(b,c,d) 210 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 211 &xor($tmp1,$d); 212 &xor($f,&swtmp(($n+8)%16)); 213 &and($tmp1,$b); 214 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 215 &rotl($f,1); # f=ROTATE(f,1) 216 &xor($tmp1,$d); # tmp1=F_00_19(b,c,d) 217 &add($e,$tmp1); # e+=F_00_19(b,c,d) 218 &mov($tmp1,$a); 219 &rotr($b,2); # b=ROTATE(b,30) 220 &mov(&swtmp($n%16),$f); # xi=f 221 &rotl($tmp1,5); # ROTATE(a,5) 222 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e 223 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round 224 &add($f,$tmp1); # f+=ROTATE(a,5) 225} 226 } 227 228sub BODY_20_39 229 { 230 local($n,$a,$b,$c,$d,$e,$f)=@_; 231 local $K=($n<40)?0x6ed9eba1:0xca62c1d6; 232 233 &comment("20_39 $n"); 234 235if ($alt) { 236 &xor($tmp1,$c); # tmp1 to hold F_20_39(b,c,d), b^=c 237 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 238 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d) 239 &xor($f,&swtmp(($n+8)%16)); 240 &add($e,$tmp1); # e+=F_20_39(b,c,d) 241 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 242 &rotl($f,1); # f=ROTATE(f,1) 243 &mov($tmp1,$a); # b in next round 244 &rotr($b,7); # b=ROTATE(b,30) 245 &mov(&swtmp($n%16),$f) if($n<77);# xi=f 246 &rotl($a,5); # ROTATE(a,5) 247 &xor($b,$c) if($n==39);# warm up for BODY_40_59 248 &and($tmp1,$b) if($n==39); 249 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY 250 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round 251 &add($f,$a); # f+=ROTATE(a,5) 252 &rotr($a,5) if ($n==79); 253} else { 254 &mov($tmp1,$b); # tmp1 to hold F_20_39(b,c,d) 255 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 256 &xor($tmp1,$c); 257 &xor($f,&swtmp(($n+8)%16)); 258 &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d) 259 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 260 &rotl($f,1); # f=ROTATE(f,1) 261 &add($e,$tmp1); # e+=F_20_39(b,c,d) 262 &rotr($b,2); # b=ROTATE(b,30) 263 &mov($tmp1,$a); 264 &rotl($tmp1,5); # ROTATE(a,5) 265 &mov(&swtmp($n%16),$f) if($n<77);# xi=f 266 &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY 267 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round 268 &add($f,$tmp1); # f+=ROTATE(a,5) 269} 270 } 271 272sub BODY_40_59 273 { 274 local($n,$a,$b,$c,$d,$e,$f)=@_; 275 276 &comment("40_59 $n"); 277 278if ($alt) { 279 &add($e,$tmp1); # e+=b&(c^d) 280 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 281 &mov($tmp1,$d); 282 &xor($f,&swtmp(($n+8)%16)); 283 &xor($c,$d); # restore $c 284 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 285 &rotl($f,1); # f=ROTATE(f,1) 286 &and($tmp1,$c); 287 &rotr($b,7); # b=ROTATE(b,30) 288 &add($e,$tmp1); # e+=c&d 289 &mov($tmp1,$a); # b in next round 290 &mov(&swtmp($n%16),$f); # xi=f 291 &rotl($a,5); # ROTATE(a,5) 292 &xor($b,$c) if ($n<59); 293 &and($tmp1,$b) if ($n<59);# tmp1 to hold F_40_59(b,c,d) 294 &lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e+(b&(c^d)) 295 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round 296 &add($f,$a); # f+=ROTATE(a,5) 297} else { 298 &mov($tmp1,$c); # tmp1 to hold F_40_59(b,c,d) 299 &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd) 300 &xor($tmp1,$d); 301 &xor($f,&swtmp(($n+8)%16)); 302 &and($tmp1,$b); 303 &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd 304 &rotl($f,1); # f=ROTATE(f,1) 305 &add($tmp1,$e); # b&(c^d)+=e 306 &rotr($b,2); # b=ROTATE(b,30) 307 &mov($e,$a); # e becomes volatile 308 &rotl($e,5); # ROTATE(a,5) 309 &mov(&swtmp($n%16),$f); # xi=f 310 &lea($f,&DWP(0x8f1bbcdc,$f,$tmp1));# f+=K_40_59+e+(b&(c^d)) 311 &mov($tmp1,$c); 312 &add($f,$e); # f+=ROTATE(a,5) 313 &and($tmp1,$d); 314 &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round 315 &add($f,$tmp1); # f+=c&d 316} 317 } 318 319&static_label("K_XX_XX"); 320 321&function_begin("sha1_block_data_order_nohw"); 322 &mov($tmp1,&wparam(0)); # SHA_CTX *c 323 &mov($T,&wparam(1)); # const void *input 324 &mov($A,&wparam(2)); # size_t num 325 &stack_push(16+3); # allocate X[16] 326 &shl($A,6); 327 &add($A,$T); 328 &mov(&wparam(2),$A); # pointer beyond the end of input 329 &mov($E,&DWP(16,$tmp1));# pre-load E 330 &jmp(&label("loop")); 331 332&set_label("loop",16); 333 334 # copy input chunk to X, but reversing byte order! 335 for ($i=0; $i<16; $i+=4) 336 { 337 &mov($A,&DWP(4*($i+0),$T)); 338 &mov($B,&DWP(4*($i+1),$T)); 339 &mov($C,&DWP(4*($i+2),$T)); 340 &mov($D,&DWP(4*($i+3),$T)); 341 &bswap($A); 342 &bswap($B); 343 &bswap($C); 344 &bswap($D); 345 &mov(&swtmp($i+0),$A); 346 &mov(&swtmp($i+1),$B); 347 &mov(&swtmp($i+2),$C); 348 &mov(&swtmp($i+3),$D); 349 } 350 &mov(&wparam(1),$T); # redundant in 1st spin 351 352 &mov($A,&DWP(0,$tmp1)); # load SHA_CTX 353 &mov($B,&DWP(4,$tmp1)); 354 &mov($C,&DWP(8,$tmp1)); 355 &mov($D,&DWP(12,$tmp1)); 356 # E is pre-loaded 357 358 for($i=0;$i<16;$i++) { &BODY_00_15($i,@V); unshift(@V,pop(@V)); } 359 for(;$i<20;$i++) { &BODY_16_19($i,@V); unshift(@V,pop(@V)); } 360 for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); } 361 for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); } 362 for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); } 363 364 (($V[5] eq $D) and ($V[0] eq $E)) or die; # double-check 365 366 &mov($tmp1,&wparam(0)); # re-load SHA_CTX* 367 &mov($D,&wparam(1)); # D is last "T" and is discarded 368 369 &add($E,&DWP(0,$tmp1)); # E is last "A"... 370 &add($T,&DWP(4,$tmp1)); 371 &add($A,&DWP(8,$tmp1)); 372 &add($B,&DWP(12,$tmp1)); 373 &add($C,&DWP(16,$tmp1)); 374 375 &mov(&DWP(0,$tmp1),$E); # update SHA_CTX 376 &add($D,64); # advance input pointer 377 &mov(&DWP(4,$tmp1),$T); 378 &cmp($D,&wparam(2)); # have we reached the end yet? 379 &mov(&DWP(8,$tmp1),$A); 380 &mov($E,$C); # C is last "E" which needs to be "pre-loaded" 381 &mov(&DWP(12,$tmp1),$B); 382 &mov($T,$D); # input pointer 383 &mov(&DWP(16,$tmp1),$C); 384 &jb(&label("loop")); 385 386 &stack_pop(16+3); 387&function_end("sha1_block_data_order_nohw"); 388 389if ($xmm) { 390if ($shaext) { 391###################################################################### 392# Intel SHA Extensions implementation of SHA1 update function. 393# 394my ($ctx,$inp,$num)=("edi","esi","ecx"); 395my ($ABCD,$E,$E_,$BSWAP)=map("xmm$_",(0..3)); 396my @MSG=map("xmm$_",(4..7)); 397 398sub sha1rnds4 { 399 my ($dst,$src,$imm)=@_; 400 if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/) 401 { &data_byte(0x0f,0x3a,0xcc,0xc0|($1<<3)|$2,$imm); } 402} 403sub sha1op38 { 404 my ($opcodelet,$dst,$src)=@_; 405 if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/) 406 { &data_byte(0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2); } 407} 408sub sha1nexte { sha1op38(0xc8,@_); } 409sub sha1msg1 { sha1op38(0xc9,@_); } 410sub sha1msg2 { sha1op38(0xca,@_); } 411 412&function_begin("sha1_block_data_order_shaext"); 413 &call (&label("pic_point")); # make it PIC! 414 &set_label("pic_point"); 415 &blindpop($tmp1); 416 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1)); 417 &mov ($ctx,&wparam(0)); 418 &mov ("ebx","esp"); 419 &mov ($inp,&wparam(1)); 420 &mov ($num,&wparam(2)); 421 &sub ("esp",32); 422 423 &movdqu ($ABCD,&QWP(0,$ctx)); 424 &movd ($E,&DWP(16,$ctx)); 425 &and ("esp",-32); 426 &movdqa ($BSWAP,&QWP(0x50,$tmp1)); # byte-n-word swap 427 428 &movdqu (@MSG[0],&QWP(0,$inp)); 429 &pshufd ($ABCD,$ABCD,0b00011011); # flip word order 430 &movdqu (@MSG[1],&QWP(0x10,$inp)); 431 &pshufd ($E,$E,0b00011011); # flip word order 432 &movdqu (@MSG[2],&QWP(0x20,$inp)); 433 &pshufb (@MSG[0],$BSWAP); 434 &movdqu (@MSG[3],&QWP(0x30,$inp)); 435 &pshufb (@MSG[1],$BSWAP); 436 &pshufb (@MSG[2],$BSWAP); 437 &pshufb (@MSG[3],$BSWAP); 438 &jmp (&label("loop_shaext")); 439 440&set_label("loop_shaext",16); 441 &dec ($num); 442 &lea ("eax",&DWP(0x40,$inp)); 443 &movdqa (&QWP(0,"esp"),$E); # offload $E 444 &paddd ($E,@MSG[0]); 445 &cmovne ($inp,"eax"); 446 &movdqa (&QWP(16,"esp"),$ABCD); # offload $ABCD 447 448for($i=0;$i<20-4;$i+=2) { 449 &sha1msg1 (@MSG[0],@MSG[1]); 450 &movdqa ($E_,$ABCD); 451 &sha1rnds4 ($ABCD,$E,int($i/5)); # 0-3... 452 &sha1nexte ($E_,@MSG[1]); 453 &pxor (@MSG[0],@MSG[2]); 454 &sha1msg1 (@MSG[1],@MSG[2]); 455 &sha1msg2 (@MSG[0],@MSG[3]); 456 457 &movdqa ($E,$ABCD); 458 &sha1rnds4 ($ABCD,$E_,int(($i+1)/5)); 459 &sha1nexte ($E,@MSG[2]); 460 &pxor (@MSG[1],@MSG[3]); 461 &sha1msg2 (@MSG[1],@MSG[0]); 462 463 push(@MSG,shift(@MSG)); push(@MSG,shift(@MSG)); 464} 465 &movdqu (@MSG[0],&QWP(0,$inp)); 466 &movdqa ($E_,$ABCD); 467 &sha1rnds4 ($ABCD,$E,3); # 64-67 468 &sha1nexte ($E_,@MSG[1]); 469 &movdqu (@MSG[1],&QWP(0x10,$inp)); 470 &pshufb (@MSG[0],$BSWAP); 471 472 &movdqa ($E,$ABCD); 473 &sha1rnds4 ($ABCD,$E_,3); # 68-71 474 &sha1nexte ($E,@MSG[2]); 475 &movdqu (@MSG[2],&QWP(0x20,$inp)); 476 &pshufb (@MSG[1],$BSWAP); 477 478 &movdqa ($E_,$ABCD); 479 &sha1rnds4 ($ABCD,$E,3); # 72-75 480 &sha1nexte ($E_,@MSG[3]); 481 &movdqu (@MSG[3],&QWP(0x30,$inp)); 482 &pshufb (@MSG[2],$BSWAP); 483 484 &movdqa ($E,$ABCD); 485 &sha1rnds4 ($ABCD,$E_,3); # 76-79 486 &movdqa ($E_,&QWP(0,"esp")); 487 &pshufb (@MSG[3],$BSWAP); 488 &sha1nexte ($E,$E_); 489 &paddd ($ABCD,&QWP(16,"esp")); 490 491 &jnz (&label("loop_shaext")); 492 493 &pshufd ($ABCD,$ABCD,0b00011011); 494 &pshufd ($E,$E,0b00011011); 495 &movdqu (&QWP(0,$ctx),$ABCD) 496 &movd (&DWP(16,$ctx),$E); 497 &mov ("esp","ebx"); 498&function_end("sha1_block_data_order_shaext"); 499} 500###################################################################### 501# The SSSE3 implementation. 502# 503# %xmm[0-7] are used as ring @X[] buffer containing quadruples of last 504# 32 elements of the message schedule or Xupdate outputs. First 4 505# quadruples are simply byte-swapped input, next 4 are calculated 506# according to method originally suggested by Dean Gaudet (modulo 507# being implemented in SSSE3). Once 8 quadruples or 32 elements are 508# collected, it switches to routine proposed by Max Locktyukhin. 509# 510# Calculations inevitably require temporary registers, and there are 511# no %xmm registers left to spare. For this reason part of the ring 512# buffer, X[2..4] to be specific, is offloaded to 3 quadriples ring 513# buffer on the stack. Keep in mind that X[2] is alias X[-6], X[3] - 514# X[-5], and X[4] - X[-4]... 515# 516# Another notable optimization is aggressive stack frame compression 517# aiming to minimize amount of 9-byte instructions... 518# 519# Yet another notable optimization is "jumping" $B variable. It means 520# that there is no register permanently allocated for $B value. This 521# allowed to eliminate one instruction from body_20_39... 522# 523my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded 524my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4 525my @V=($A,$B,$C,$D,$E); 526my $j=0; # hash round 527my $rx=0; 528my @T=($T,$tmp1); 529my $inp; 530 531my $_rol=sub { &rol(@_) }; 532my $_ror=sub { &ror(@_) }; 533 534&function_begin("sha1_block_data_order_ssse3"); 535 &call (&label("pic_point")); # make it PIC! 536 &set_label("pic_point"); 537 &blindpop($tmp1); 538 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1)); 539 540 &movdqa (@X[3],&QWP(0,$tmp1)); # K_00_19 541 &movdqa (@X[4],&QWP(16,$tmp1)); # K_20_39 542 &movdqa (@X[5],&QWP(32,$tmp1)); # K_40_59 543 &movdqa (@X[6],&QWP(48,$tmp1)); # K_60_79 544 &movdqa (@X[2],&QWP(64,$tmp1)); # pbswap mask 545 546 &mov ($E,&wparam(0)); # load argument block 547 &mov ($inp=@T[1],&wparam(1)); 548 &mov ($D,&wparam(2)); 549 &mov (@T[0],"esp"); 550 551 # stack frame layout 552 # 553 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area 554 # X[4]+K X[5]+K X[6]+K X[7]+K 555 # X[8]+K X[9]+K X[10]+K X[11]+K 556 # X[12]+K X[13]+K X[14]+K X[15]+K 557 # 558 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area 559 # X[4] X[5] X[6] X[7] 560 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19 561 # 562 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants 563 # K_40_59 K_40_59 K_40_59 K_40_59 564 # K_60_79 K_60_79 K_60_79 K_60_79 565 # K_00_19 K_00_19 K_00_19 K_00_19 566 # pbswap mask 567 # 568 # +192 ctx # argument block 569 # +196 inp 570 # +200 end 571 # +204 esp 572 &sub ("esp",208); 573 &and ("esp",-64); 574 575 &movdqa (&QWP(112+0,"esp"),@X[4]); # copy constants 576 &movdqa (&QWP(112+16,"esp"),@X[5]); 577 &movdqa (&QWP(112+32,"esp"),@X[6]); 578 &shl ($D,6); # len*64 579 &movdqa (&QWP(112+48,"esp"),@X[3]); 580 &add ($D,$inp); # end of input 581 &movdqa (&QWP(112+64,"esp"),@X[2]); 582 &add ($inp,64); 583 &mov (&DWP(192+0,"esp"),$E); # save argument block 584 &mov (&DWP(192+4,"esp"),$inp); 585 &mov (&DWP(192+8,"esp"),$D); 586 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp 587 588 &mov ($A,&DWP(0,$E)); # load context 589 &mov ($B,&DWP(4,$E)); 590 &mov ($C,&DWP(8,$E)); 591 &mov ($D,&DWP(12,$E)); 592 &mov ($E,&DWP(16,$E)); 593 &mov (@T[0],$B); # magic seed 594 595 &movdqu (@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3] 596 &movdqu (@X[-3&7],&QWP(-48,$inp)); 597 &movdqu (@X[-2&7],&QWP(-32,$inp)); 598 &movdqu (@X[-1&7],&QWP(-16,$inp)); 599 &pshufb (@X[-4&7],@X[2]); # byte swap 600 &pshufb (@X[-3&7],@X[2]); 601 &pshufb (@X[-2&7],@X[2]); 602 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot 603 &pshufb (@X[-1&7],@X[2]); 604 &paddd (@X[-4&7],@X[3]); # add K_00_19 605 &paddd (@X[-3&7],@X[3]); 606 &paddd (@X[-2&7],@X[3]); 607 &movdqa (&QWP(0,"esp"),@X[-4&7]); # X[]+K xfer to IALU 608 &psubd (@X[-4&7],@X[3]); # restore X[] 609 &movdqa (&QWP(0+16,"esp"),@X[-3&7]); 610 &psubd (@X[-3&7],@X[3]); 611 &movdqa (&QWP(0+32,"esp"),@X[-2&7]); 612 &mov (@T[1],$C); 613 &psubd (@X[-2&7],@X[3]); 614 &xor (@T[1],$D); 615 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]); 616 &and (@T[0],@T[1]); 617 &jmp (&label("loop")); 618 619###################################################################### 620# SSE instruction sequence is first broken to groups of independent 621# instructions, independent in respect to their inputs and shifter 622# (not all architectures have more than one). Then IALU instructions 623# are "knitted in" between the SSE groups. Distance is maintained for 624# SSE latency of 2 in hope that it fits better upcoming AMD Bulldozer 625# [which allegedly also implements SSSE3]... 626# 627# Temporary registers usage. X[2] is volatile at the entry and at the 628# end is restored from backtrace ring buffer. X[3] is expected to 629# contain current K_XX_XX constant and is used to calculate X[-1]+K 630# from previous round, it becomes volatile the moment the value is 631# saved to stack for transfer to IALU. X[4] becomes volatile whenever 632# X[-4] is accumulated and offloaded to backtrace ring buffer, at the 633# end it is loaded with next K_XX_XX [which becomes X[3] in next 634# round]... 635# 636sub Xupdate_ssse3_16_31() # recall that $Xi starts with 4 637{ use integer; 638 my $body = shift; 639 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions 640 my ($a,$b,$c,$d,$e); 641 642 eval(shift(@insns)); # ror 643 eval(shift(@insns)); 644 eval(shift(@insns)); 645 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8); 646 &movdqa (@X[2],@X[-1&7]); 647 eval(shift(@insns)); 648 eval(shift(@insns)); 649 650 &paddd (@X[3],@X[-1&7]); 651 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer 652 eval(shift(@insns)); # rol 653 eval(shift(@insns)); 654 &psrldq (@X[2],4); # "X[-3]", 3 dwords 655 eval(shift(@insns)); 656 eval(shift(@insns)); 657 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]" 658 eval(shift(@insns)); 659 eval(shift(@insns)); # ror 660 661 &pxor (@X[2],@X[-2&7]); # "X[-3]"^"X[-8]" 662 eval(shift(@insns)); 663 eval(shift(@insns)); 664 eval(shift(@insns)); 665 666 &pxor (@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]" 667 eval(shift(@insns)); 668 eval(shift(@insns)); # rol 669 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU 670 eval(shift(@insns)); 671 eval(shift(@insns)); 672 673 &movdqa (@X[4],@X[0]); 674 eval(shift(@insns)); 675 eval(shift(@insns)); 676 eval(shift(@insns)); # ror 677 &movdqa (@X[2],@X[0]); 678 eval(shift(@insns)); 679 680 &pslldq (@X[4],12); # "X[0]"<<96, extract one dword 681 &paddd (@X[0],@X[0]); 682 eval(shift(@insns)); 683 eval(shift(@insns)); 684 685 &psrld (@X[2],31); 686 eval(shift(@insns)); 687 eval(shift(@insns)); # rol 688 &movdqa (@X[3],@X[4]); 689 eval(shift(@insns)); 690 eval(shift(@insns)); 691 eval(shift(@insns)); 692 693 &psrld (@X[4],30); 694 eval(shift(@insns)); 695 eval(shift(@insns)); # ror 696 &por (@X[0],@X[2]); # "X[0]"<<<=1 697 eval(shift(@insns)); 698 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer 699 eval(shift(@insns)); 700 eval(shift(@insns)); 701 702 &pslld (@X[3],2); 703 eval(shift(@insns)); 704 eval(shift(@insns)); # rol 705 &pxor (@X[0],@X[4]); 706 &movdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX 707 eval(shift(@insns)); 708 eval(shift(@insns)); 709 710 &pxor (@X[0],@X[3]); # "X[0]"^=("X[0]"<<96)<<<2 711 &pshufd (@X[1],@X[-3&7],0xee) if ($Xi<7); # was &movdqa (@X[1],@X[-2&7]) 712 &pshufd (@X[3],@X[-1&7],0xee) if ($Xi==7); 713 eval(shift(@insns)); 714 eval(shift(@insns)); 715 716 foreach (@insns) { eval; } # remaining instructions [if any] 717 718 $Xi++; push(@X,shift(@X)); # "rotate" X[] 719} 720 721sub Xupdate_ssse3_32_79() 722{ use integer; 723 my $body = shift; 724 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions 725 my ($a,$b,$c,$d,$e); 726 727 eval(shift(@insns)); # body_20_39 728 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" 729 &punpcklqdq(@X[2],@X[-1&7]); # compose "X[-6]", was &palignr(@X[2],@X[-2&7],8) 730 eval(shift(@insns)); 731 eval(shift(@insns)); 732 eval(shift(@insns)); # rol 733 734 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]" 735 &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer 736 eval(shift(@insns)); 737 eval(shift(@insns)); 738 eval(shift(@insns)) if (@insns[0] =~ /_rol/); 739 if ($Xi%5) { 740 &movdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX... 741 } else { # ... or load next one 742 &movdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp")); 743 } 744 eval(shift(@insns)); # ror 745 &paddd (@X[3],@X[-1&7]); 746 eval(shift(@insns)); 747 748 &pxor (@X[0],@X[2]); # "X[0]"^="X[-6]" 749 eval(shift(@insns)); # body_20_39 750 eval(shift(@insns)); 751 eval(shift(@insns)); 752 eval(shift(@insns)); # rol 753 754 &movdqa (@X[2],@X[0]); 755 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU 756 eval(shift(@insns)); 757 eval(shift(@insns)); 758 eval(shift(@insns)); # ror 759 eval(shift(@insns)); 760 eval(shift(@insns)) if (@insns[0] =~ /_rol/); 761 762 &pslld (@X[0],2); 763 eval(shift(@insns)); # body_20_39 764 eval(shift(@insns)); 765 &psrld (@X[2],30); 766 eval(shift(@insns)); 767 eval(shift(@insns)); # rol 768 eval(shift(@insns)); 769 eval(shift(@insns)); 770 eval(shift(@insns)); # ror 771 eval(shift(@insns)); 772 eval(shift(@insns)) if (@insns[1] =~ /_rol/); 773 eval(shift(@insns)) if (@insns[0] =~ /_rol/); 774 775 &por (@X[0],@X[2]); # "X[0]"<<<=2 776 eval(shift(@insns)); # body_20_39 777 eval(shift(@insns)); 778 &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer 779 eval(shift(@insns)); 780 eval(shift(@insns)); # rol 781 eval(shift(@insns)); 782 eval(shift(@insns)); 783 eval(shift(@insns)); # ror 784 &pshufd (@X[3],@X[-1],0xee) if ($Xi<19); # was &movdqa (@X[3],@X[0]) 785 eval(shift(@insns)); 786 787 foreach (@insns) { eval; } # remaining instructions 788 789 $Xi++; push(@X,shift(@X)); # "rotate" X[] 790} 791 792sub Xuplast_ssse3_80() 793{ use integer; 794 my $body = shift; 795 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 796 my ($a,$b,$c,$d,$e); 797 798 eval(shift(@insns)); 799 eval(shift(@insns)); 800 eval(shift(@insns)); 801 eval(shift(@insns)); 802 eval(shift(@insns)); 803 eval(shift(@insns)); 804 eval(shift(@insns)); 805 &paddd (@X[3],@X[-1&7]); 806 eval(shift(@insns)); 807 eval(shift(@insns)); 808 eval(shift(@insns)); 809 eval(shift(@insns)); 810 811 &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU 812 813 foreach (@insns) { eval; } # remaining instructions 814 815 &mov ($inp=@T[1],&DWP(192+4,"esp")); 816 &cmp ($inp,&DWP(192+8,"esp")); 817 &je (&label("done")); 818 819 &movdqa (@X[3],&QWP(112+48,"esp")); # K_00_19 820 &movdqa (@X[2],&QWP(112+64,"esp")); # pbswap mask 821 &movdqu (@X[-4&7],&QWP(0,$inp)); # load input 822 &movdqu (@X[-3&7],&QWP(16,$inp)); 823 &movdqu (@X[-2&7],&QWP(32,$inp)); 824 &movdqu (@X[-1&7],&QWP(48,$inp)); 825 &add ($inp,64); 826 &pshufb (@X[-4&7],@X[2]); # byte swap 827 &mov (&DWP(192+4,"esp"),$inp); 828 &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot 829 830 $Xi=0; 831} 832 833sub Xloop_ssse3() 834{ use integer; 835 my $body = shift; 836 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 837 my ($a,$b,$c,$d,$e); 838 839 eval(shift(@insns)); 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 &pshufb (@X[($Xi-3)&7],@X[2]); 847 eval(shift(@insns)); 848 eval(shift(@insns)); 849 eval(shift(@insns)); 850 eval(shift(@insns)); 851 &paddd (@X[($Xi-4)&7],@X[3]); 852 eval(shift(@insns)); 853 eval(shift(@insns)); 854 eval(shift(@insns)); 855 eval(shift(@insns)); 856 &movdqa (&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]); # X[]+K xfer to IALU 857 eval(shift(@insns)); 858 eval(shift(@insns)); 859 eval(shift(@insns)); 860 eval(shift(@insns)); 861 &psubd (@X[($Xi-4)&7],@X[3]); 862 863 foreach (@insns) { eval; } 864 $Xi++; 865} 866 867sub Xtail_ssse3() 868{ use integer; 869 my $body = shift; 870 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 871 my ($a,$b,$c,$d,$e); 872 873 foreach (@insns) { eval; } 874} 875 876sub body_00_19 () { # ((c^d)&b)^d 877 # on start @T[0]=(c^d)&b 878 return &body_20_39() if ($rx==19); $rx++; 879 ( 880 '($a,$b,$c,$d,$e)=@V;'. 881 '&$_ror ($b,$j?7:2);', # $b>>>2 882 '&xor (@T[0],$d);', 883 '&mov (@T[1],$a);', # $b in next round 884 885 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer 886 '&xor ($b,$c);', # $c^$d for next round 887 888 '&$_rol ($a,5);', 889 '&add ($e,@T[0]);', 890 '&and (@T[1],$b);', # ($b&($c^$d)) for next round 891 892 '&xor ($b,$c);', # restore $b 893 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' 894 ); 895} 896 897sub body_20_39 () { # b^d^c 898 # on entry @T[0]=b^d 899 return &body_40_59() if ($rx==39); $rx++; 900 ( 901 '($a,$b,$c,$d,$e)=@V;'. 902 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer 903 '&xor (@T[0],$d) if($j==19);'. 904 '&xor (@T[0],$c) if($j> 19);', # ($b^$d^$c) 905 '&mov (@T[1],$a);', # $b in next round 906 907 '&$_rol ($a,5);', 908 '&add ($e,@T[0]);', 909 '&xor (@T[1],$c) if ($j< 79);', # $b^$d for next round 910 911 '&$_ror ($b,7);', # $b>>>2 912 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' 913 ); 914} 915 916sub body_40_59 () { # ((b^c)&(c^d))^c 917 # on entry @T[0]=(b^c), (c^=d) 918 $rx++; 919 ( 920 '($a,$b,$c,$d,$e)=@V;'. 921 '&add ($e,&DWP(4*($j&15),"esp"));', # X[]+K xfer 922 '&and (@T[0],$c) if ($j>=40);', # (b^c)&(c^d) 923 '&xor ($c,$d) if ($j>=40);', # restore $c 924 925 '&$_ror ($b,7);', # $b>>>2 926 '&mov (@T[1],$a);', # $b for next round 927 '&xor (@T[0],$c);', 928 929 '&$_rol ($a,5);', 930 '&add ($e,@T[0]);', 931 '&xor (@T[1],$c) if ($j==59);'. 932 '&xor (@T[1],$b) if ($j< 59);', # b^c for next round 933 934 '&xor ($b,$c) if ($j< 59);', # c^d for next round 935 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' 936 ); 937} 938###### 939sub bodyx_00_19 () { # ((c^d)&b)^d 940 # on start @T[0]=(b&c)^(~b&d), $e+=X[]+K 941 return &bodyx_20_39() if ($rx==19); $rx++; 942 ( 943 '($a,$b,$c,$d,$e)=@V;'. 944 945 '&rorx ($b,$b,2) if ($j==0);'. # $b>>>2 946 '&rorx ($b,@T[1],7) if ($j!=0);', # $b>>>2 947 '&lea ($e,&DWP(0,$e,@T[0]));', 948 '&rorx (@T[0],$a,5);', 949 950 '&andn (@T[1],$a,$c);', 951 '&and ($a,$b)', 952 '&add ($d,&DWP(4*(($j+1)&15),"esp"));', # X[]+K xfer 953 954 '&xor (@T[1],$a)', 955 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' 956 ); 957} 958 959sub bodyx_20_39 () { # b^d^c 960 # on start $b=b^c^d 961 return &bodyx_40_59() if ($rx==39); $rx++; 962 ( 963 '($a,$b,$c,$d,$e)=@V;'. 964 965 '&add ($e,($j==19?@T[0]:$b))', 966 '&rorx ($b,@T[1],7);', # $b>>>2 967 '&rorx (@T[0],$a,5);', 968 969 '&xor ($a,$b) if ($j<79);', 970 '&add ($d,&DWP(4*(($j+1)&15),"esp")) if ($j<79);', # X[]+K xfer 971 '&xor ($a,$c) if ($j<79);', 972 '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' 973 ); 974} 975 976sub bodyx_40_59 () { # ((b^c)&(c^d))^c 977 # on start $b=((b^c)&(c^d))^c 978 return &bodyx_20_39() if ($rx==59); $rx++; 979 ( 980 '($a,$b,$c,$d,$e)=@V;'. 981 982 '&rorx (@T[0],$a,5)', 983 '&lea ($e,&DWP(0,$e,$b))', 984 '&rorx ($b,@T[1],7)', # $b>>>2 985 '&add ($d,&DWP(4*(($j+1)&15),"esp"))', # X[]+K xfer 986 987 '&mov (@T[1],$c)', 988 '&xor ($a,$b)', # b^c for next round 989 '&xor (@T[1],$b)', # c^d for next round 990 991 '&and ($a,@T[1])', 992 '&add ($e,@T[0])', 993 '&xor ($a,$b)' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' 994 ); 995} 996 997&set_label("loop",16); 998 &Xupdate_ssse3_16_31(\&body_00_19); 999 &Xupdate_ssse3_16_31(\&body_00_19); 1000 &Xupdate_ssse3_16_31(\&body_00_19); 1001 &Xupdate_ssse3_16_31(\&body_00_19); 1002 &Xupdate_ssse3_32_79(\&body_00_19); 1003 &Xupdate_ssse3_32_79(\&body_20_39); 1004 &Xupdate_ssse3_32_79(\&body_20_39); 1005 &Xupdate_ssse3_32_79(\&body_20_39); 1006 &Xupdate_ssse3_32_79(\&body_20_39); 1007 &Xupdate_ssse3_32_79(\&body_20_39); 1008 &Xupdate_ssse3_32_79(\&body_40_59); 1009 &Xupdate_ssse3_32_79(\&body_40_59); 1010 &Xupdate_ssse3_32_79(\&body_40_59); 1011 &Xupdate_ssse3_32_79(\&body_40_59); 1012 &Xupdate_ssse3_32_79(\&body_40_59); 1013 &Xupdate_ssse3_32_79(\&body_20_39); 1014 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done" 1015 1016 $saved_j=$j; @saved_V=@V; 1017 1018 &Xloop_ssse3(\&body_20_39); 1019 &Xloop_ssse3(\&body_20_39); 1020 &Xloop_ssse3(\&body_20_39); 1021 1022 &mov (@T[1],&DWP(192,"esp")); # update context 1023 &add ($A,&DWP(0,@T[1])); 1024 &add (@T[0],&DWP(4,@T[1])); # $b 1025 &add ($C,&DWP(8,@T[1])); 1026 &mov (&DWP(0,@T[1]),$A); 1027 &add ($D,&DWP(12,@T[1])); 1028 &mov (&DWP(4,@T[1]),@T[0]); 1029 &add ($E,&DWP(16,@T[1])); 1030 &mov (&DWP(8,@T[1]),$C); 1031 &mov ($B,$C); 1032 &mov (&DWP(12,@T[1]),$D); 1033 &xor ($B,$D); 1034 &mov (&DWP(16,@T[1]),$E); 1035 &mov (@T[1],@T[0]); 1036 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]); 1037 &and (@T[0],$B); 1038 &mov ($B,$T[1]); 1039 1040 &jmp (&label("loop")); 1041 1042&set_label("done",16); $j=$saved_j; @V=@saved_V; 1043 1044 &Xtail_ssse3(\&body_20_39); 1045 &Xtail_ssse3(\&body_20_39); 1046 &Xtail_ssse3(\&body_20_39); 1047 1048 &mov (@T[1],&DWP(192,"esp")); # update context 1049 &add ($A,&DWP(0,@T[1])); 1050 &mov ("esp",&DWP(192+12,"esp")); # restore %esp 1051 &add (@T[0],&DWP(4,@T[1])); # $b 1052 &add ($C,&DWP(8,@T[1])); 1053 &mov (&DWP(0,@T[1]),$A); 1054 &add ($D,&DWP(12,@T[1])); 1055 &mov (&DWP(4,@T[1]),@T[0]); 1056 &add ($E,&DWP(16,@T[1])); 1057 &mov (&DWP(8,@T[1]),$C); 1058 &mov (&DWP(12,@T[1]),$D); 1059 &mov (&DWP(16,@T[1]),$E); 1060 1061&function_end("sha1_block_data_order_ssse3"); 1062 1063$rx=0; # reset 1064 1065if ($ymm) { 1066my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded 1067my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4 1068my @V=($A,$B,$C,$D,$E); 1069my $j=0; # hash round 1070my @T=($T,$tmp1); 1071my $inp; 1072 1073my $_rol=sub { &shld(@_[0],@_) }; 1074my $_ror=sub { &shrd(@_[0],@_) }; 1075 1076&function_begin("sha1_block_data_order_avx"); 1077 &call (&label("pic_point")); # make it PIC! 1078 &set_label("pic_point"); 1079 &blindpop($tmp1); 1080 &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1)); 1081 &vzeroall(); 1082 1083 &vmovdqa(@X[3],&QWP(0,$tmp1)); # K_00_19 1084 &vmovdqa(@X[4],&QWP(16,$tmp1)); # K_20_39 1085 &vmovdqa(@X[5],&QWP(32,$tmp1)); # K_40_59 1086 &vmovdqa(@X[6],&QWP(48,$tmp1)); # K_60_79 1087 &vmovdqa(@X[2],&QWP(64,$tmp1)); # pbswap mask 1088 1089 &mov ($E,&wparam(0)); # load argument block 1090 &mov ($inp=@T[1],&wparam(1)); 1091 &mov ($D,&wparam(2)); 1092 &mov (@T[0],"esp"); 1093 1094 # stack frame layout 1095 # 1096 # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area 1097 # X[4]+K X[5]+K X[6]+K X[7]+K 1098 # X[8]+K X[9]+K X[10]+K X[11]+K 1099 # X[12]+K X[13]+K X[14]+K X[15]+K 1100 # 1101 # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area 1102 # X[4] X[5] X[6] X[7] 1103 # X[8] X[9] X[10] X[11] # even borrowed for K_00_19 1104 # 1105 # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants 1106 # K_40_59 K_40_59 K_40_59 K_40_59 1107 # K_60_79 K_60_79 K_60_79 K_60_79 1108 # K_00_19 K_00_19 K_00_19 K_00_19 1109 # pbswap mask 1110 # 1111 # +192 ctx # argument block 1112 # +196 inp 1113 # +200 end 1114 # +204 esp 1115 &sub ("esp",208); 1116 &and ("esp",-64); 1117 1118 &vmovdqa(&QWP(112+0,"esp"),@X[4]); # copy constants 1119 &vmovdqa(&QWP(112+16,"esp"),@X[5]); 1120 &vmovdqa(&QWP(112+32,"esp"),@X[6]); 1121 &shl ($D,6); # len*64 1122 &vmovdqa(&QWP(112+48,"esp"),@X[3]); 1123 &add ($D,$inp); # end of input 1124 &vmovdqa(&QWP(112+64,"esp"),@X[2]); 1125 &add ($inp,64); 1126 &mov (&DWP(192+0,"esp"),$E); # save argument block 1127 &mov (&DWP(192+4,"esp"),$inp); 1128 &mov (&DWP(192+8,"esp"),$D); 1129 &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp 1130 1131 &mov ($A,&DWP(0,$E)); # load context 1132 &mov ($B,&DWP(4,$E)); 1133 &mov ($C,&DWP(8,$E)); 1134 &mov ($D,&DWP(12,$E)); 1135 &mov ($E,&DWP(16,$E)); 1136 &mov (@T[0],$B); # magic seed 1137 1138 &vmovdqu(@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3] 1139 &vmovdqu(@X[-3&7],&QWP(-48,$inp)); 1140 &vmovdqu(@X[-2&7],&QWP(-32,$inp)); 1141 &vmovdqu(@X[-1&7],&QWP(-16,$inp)); 1142 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap 1143 &vpshufb(@X[-3&7],@X[-3&7],@X[2]); 1144 &vpshufb(@X[-2&7],@X[-2&7],@X[2]); 1145 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot 1146 &vpshufb(@X[-1&7],@X[-1&7],@X[2]); 1147 &vpaddd (@X[0],@X[-4&7],@X[3]); # add K_00_19 1148 &vpaddd (@X[1],@X[-3&7],@X[3]); 1149 &vpaddd (@X[2],@X[-2&7],@X[3]); 1150 &vmovdqa(&QWP(0,"esp"),@X[0]); # X[]+K xfer to IALU 1151 &mov (@T[1],$C); 1152 &vmovdqa(&QWP(0+16,"esp"),@X[1]); 1153 &xor (@T[1],$D); 1154 &vmovdqa(&QWP(0+32,"esp"),@X[2]); 1155 &and (@T[0],@T[1]); 1156 &jmp (&label("loop")); 1157 1158sub Xupdate_avx_16_31() # recall that $Xi starts with 4 1159{ use integer; 1160 my $body = shift; 1161 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions 1162 my ($a,$b,$c,$d,$e); 1163 1164 eval(shift(@insns)); 1165 eval(shift(@insns)); 1166 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]" 1167 eval(shift(@insns)); 1168 eval(shift(@insns)); 1169 1170 &vpaddd (@X[3],@X[3],@X[-1&7]); 1171 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer 1172 eval(shift(@insns)); 1173 eval(shift(@insns)); 1174 &vpsrldq(@X[2],@X[-1&7],4); # "X[-3]", 3 dwords 1175 eval(shift(@insns)); 1176 eval(shift(@insns)); 1177 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]" 1178 eval(shift(@insns)); 1179 eval(shift(@insns)); 1180 1181 &vpxor (@X[2],@X[2],@X[-2&7]); # "X[-3]"^"X[-8]" 1182 eval(shift(@insns)); 1183 eval(shift(@insns)); 1184 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU 1185 eval(shift(@insns)); 1186 eval(shift(@insns)); 1187 1188 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]" 1189 eval(shift(@insns)); 1190 eval(shift(@insns)); 1191 eval(shift(@insns)); 1192 eval(shift(@insns)); 1193 1194 &vpsrld (@X[2],@X[0],31); 1195 eval(shift(@insns)); 1196 eval(shift(@insns)); 1197 eval(shift(@insns)); 1198 eval(shift(@insns)); 1199 1200 &vpslldq(@X[4],@X[0],12); # "X[0]"<<96, extract one dword 1201 &vpaddd (@X[0],@X[0],@X[0]); 1202 eval(shift(@insns)); 1203 eval(shift(@insns)); 1204 eval(shift(@insns)); 1205 eval(shift(@insns)); 1206 1207 &vpsrld (@X[3],@X[4],30); 1208 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=1 1209 eval(shift(@insns)); 1210 eval(shift(@insns)); 1211 eval(shift(@insns)); 1212 eval(shift(@insns)); 1213 1214 &vpslld (@X[4],@X[4],2); 1215 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer 1216 eval(shift(@insns)); 1217 eval(shift(@insns)); 1218 &vpxor (@X[0],@X[0],@X[3]); 1219 eval(shift(@insns)); 1220 eval(shift(@insns)); 1221 eval(shift(@insns)); 1222 eval(shift(@insns)); 1223 1224 &vpxor (@X[0],@X[0],@X[4]); # "X[0]"^=("X[0]"<<96)<<<2 1225 eval(shift(@insns)); 1226 eval(shift(@insns)); 1227 &vmovdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX 1228 eval(shift(@insns)); 1229 eval(shift(@insns)); 1230 1231 foreach (@insns) { eval; } # remaining instructions [if any] 1232 1233 $Xi++; push(@X,shift(@X)); # "rotate" X[] 1234} 1235 1236sub Xupdate_avx_32_79() 1237{ use integer; 1238 my $body = shift; 1239 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions 1240 my ($a,$b,$c,$d,$e); 1241 1242 &vpalignr(@X[2],@X[-1&7],@X[-2&7],8); # compose "X[-6]" 1243 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" 1244 eval(shift(@insns)); # body_20_39 1245 eval(shift(@insns)); 1246 eval(shift(@insns)); 1247 eval(shift(@insns)); # rol 1248 1249 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]" 1250 &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer 1251 eval(shift(@insns)); 1252 eval(shift(@insns)); 1253 if ($Xi%5) { 1254 &vmovdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX... 1255 } else { # ... or load next one 1256 &vmovdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp")); 1257 } 1258 &vpaddd (@X[3],@X[3],@X[-1&7]); 1259 eval(shift(@insns)); # ror 1260 eval(shift(@insns)); 1261 1262 &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-6]" 1263 eval(shift(@insns)); # body_20_39 1264 eval(shift(@insns)); 1265 eval(shift(@insns)); 1266 eval(shift(@insns)); # rol 1267 1268 &vpsrld (@X[2],@X[0],30); 1269 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU 1270 eval(shift(@insns)); 1271 eval(shift(@insns)); 1272 eval(shift(@insns)); # ror 1273 eval(shift(@insns)); 1274 1275 &vpslld (@X[0],@X[0],2); 1276 eval(shift(@insns)); # body_20_39 1277 eval(shift(@insns)); 1278 eval(shift(@insns)); 1279 eval(shift(@insns)); # rol 1280 eval(shift(@insns)); 1281 eval(shift(@insns)); 1282 eval(shift(@insns)); # ror 1283 eval(shift(@insns)); 1284 1285 &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=2 1286 eval(shift(@insns)); # body_20_39 1287 eval(shift(@insns)); 1288 &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer 1289 eval(shift(@insns)); 1290 eval(shift(@insns)); # rol 1291 eval(shift(@insns)); 1292 eval(shift(@insns)); 1293 eval(shift(@insns)); # ror 1294 eval(shift(@insns)); 1295 1296 foreach (@insns) { eval; } # remaining instructions 1297 1298 $Xi++; push(@X,shift(@X)); # "rotate" X[] 1299} 1300 1301sub Xuplast_avx_80() 1302{ use integer; 1303 my $body = shift; 1304 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 1305 my ($a,$b,$c,$d,$e); 1306 1307 eval(shift(@insns)); 1308 &vpaddd (@X[3],@X[3],@X[-1&7]); 1309 eval(shift(@insns)); 1310 eval(shift(@insns)); 1311 eval(shift(@insns)); 1312 eval(shift(@insns)); 1313 1314 &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU 1315 1316 foreach (@insns) { eval; } # remaining instructions 1317 1318 &mov ($inp=@T[1],&DWP(192+4,"esp")); 1319 &cmp ($inp,&DWP(192+8,"esp")); 1320 &je (&label("done")); 1321 1322 &vmovdqa(@X[3],&QWP(112+48,"esp")); # K_00_19 1323 &vmovdqa(@X[2],&QWP(112+64,"esp")); # pbswap mask 1324 &vmovdqu(@X[-4&7],&QWP(0,$inp)); # load input 1325 &vmovdqu(@X[-3&7],&QWP(16,$inp)); 1326 &vmovdqu(@X[-2&7],&QWP(32,$inp)); 1327 &vmovdqu(@X[-1&7],&QWP(48,$inp)); 1328 &add ($inp,64); 1329 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap 1330 &mov (&DWP(192+4,"esp"),$inp); 1331 &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot 1332 1333 $Xi=0; 1334} 1335 1336sub Xloop_avx() 1337{ use integer; 1338 my $body = shift; 1339 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 1340 my ($a,$b,$c,$d,$e); 1341 1342 eval(shift(@insns)); 1343 eval(shift(@insns)); 1344 &vpshufb (@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]); 1345 eval(shift(@insns)); 1346 eval(shift(@insns)); 1347 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],@X[3]); 1348 eval(shift(@insns)); 1349 eval(shift(@insns)); 1350 eval(shift(@insns)); 1351 eval(shift(@insns)); 1352 &vmovdqa (&QWP(0+16*$Xi,"esp"),@X[$Xi&7]); # X[]+K xfer to IALU 1353 eval(shift(@insns)); 1354 eval(shift(@insns)); 1355 1356 foreach (@insns) { eval; } 1357 $Xi++; 1358} 1359 1360sub Xtail_avx() 1361{ use integer; 1362 my $body = shift; 1363 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 1364 my ($a,$b,$c,$d,$e); 1365 1366 foreach (@insns) { eval; } 1367} 1368 1369&set_label("loop",16); 1370 &Xupdate_avx_16_31(\&body_00_19); 1371 &Xupdate_avx_16_31(\&body_00_19); 1372 &Xupdate_avx_16_31(\&body_00_19); 1373 &Xupdate_avx_16_31(\&body_00_19); 1374 &Xupdate_avx_32_79(\&body_00_19); 1375 &Xupdate_avx_32_79(\&body_20_39); 1376 &Xupdate_avx_32_79(\&body_20_39); 1377 &Xupdate_avx_32_79(\&body_20_39); 1378 &Xupdate_avx_32_79(\&body_20_39); 1379 &Xupdate_avx_32_79(\&body_20_39); 1380 &Xupdate_avx_32_79(\&body_40_59); 1381 &Xupdate_avx_32_79(\&body_40_59); 1382 &Xupdate_avx_32_79(\&body_40_59); 1383 &Xupdate_avx_32_79(\&body_40_59); 1384 &Xupdate_avx_32_79(\&body_40_59); 1385 &Xupdate_avx_32_79(\&body_20_39); 1386 &Xuplast_avx_80(\&body_20_39); # can jump to "done" 1387 1388 $saved_j=$j; @saved_V=@V; 1389 1390 &Xloop_avx(\&body_20_39); 1391 &Xloop_avx(\&body_20_39); 1392 &Xloop_avx(\&body_20_39); 1393 1394 &mov (@T[1],&DWP(192,"esp")); # update context 1395 &add ($A,&DWP(0,@T[1])); 1396 &add (@T[0],&DWP(4,@T[1])); # $b 1397 &add ($C,&DWP(8,@T[1])); 1398 &mov (&DWP(0,@T[1]),$A); 1399 &add ($D,&DWP(12,@T[1])); 1400 &mov (&DWP(4,@T[1]),@T[0]); 1401 &add ($E,&DWP(16,@T[1])); 1402 &mov ($B,$C); 1403 &mov (&DWP(8,@T[1]),$C); 1404 &xor ($B,$D); 1405 &mov (&DWP(12,@T[1]),$D); 1406 &mov (&DWP(16,@T[1]),$E); 1407 &mov (@T[1],@T[0]); 1408 &and (@T[0],$B); 1409 &mov ($B,@T[1]); 1410 1411 &jmp (&label("loop")); 1412 1413&set_label("done",16); $j=$saved_j; @V=@saved_V; 1414 1415 &Xtail_avx(\&body_20_39); 1416 &Xtail_avx(\&body_20_39); 1417 &Xtail_avx(\&body_20_39); 1418 1419 &vzeroall(); 1420 1421 &mov (@T[1],&DWP(192,"esp")); # update context 1422 &add ($A,&DWP(0,@T[1])); 1423 &mov ("esp",&DWP(192+12,"esp")); # restore %esp 1424 &add (@T[0],&DWP(4,@T[1])); # $b 1425 &add ($C,&DWP(8,@T[1])); 1426 &mov (&DWP(0,@T[1]),$A); 1427 &add ($D,&DWP(12,@T[1])); 1428 &mov (&DWP(4,@T[1]),@T[0]); 1429 &add ($E,&DWP(16,@T[1])); 1430 &mov (&DWP(8,@T[1]),$C); 1431 &mov (&DWP(12,@T[1]),$D); 1432 &mov (&DWP(16,@T[1]),$E); 1433&function_end("sha1_block_data_order_avx"); 1434} 1435&set_label("K_XX_XX",64); 1436&data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999); # K_00_19 1437&data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1); # K_20_39 1438&data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc); # K_40_59 1439&data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6); # K_60_79 1440&data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f); # pbswap mask 1441&data_byte(0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0); 1442} 1443&asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>"); 1444 1445&asm_finish(); 1446 1447close STDOUT or die "error closing STDOUT: $!"; 1448