1#! /usr/bin/env perl 2# Copyright 2006-2016 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 <appro@openssl.org> 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# sha1_block procedure for x86_64. 18# 19# It was brought to my attention that on EM64T compiler-generated code 20# was far behind 32-bit assembler implementation. This is unlike on 21# Opteron where compiler-generated code was only 15% behind 32-bit 22# assembler, which originally made it hard to motivate the effort. 23# There was suggestion to mechanically translate 32-bit code, but I 24# dismissed it, reasoning that x86_64 offers enough register bank 25# capacity to fully utilize SHA-1 parallelism. Therefore this fresh 26# implementation:-) However! While 64-bit code does perform better 27# on Opteron, I failed to beat 32-bit assembler on EM64T core. Well, 28# x86_64 does offer larger *addressable* bank, but out-of-order core 29# reaches for even more registers through dynamic aliasing, and EM64T 30# core must have managed to run-time optimize even 32-bit code just as 31# good as 64-bit one. Performance improvement is summarized in the 32# following table: 33# 34# gcc 3.4 32-bit asm cycles/byte 35# Opteron +45% +20% 6.8 36# Xeon P4 +65% +0% 9.9 37# Core2 +60% +10% 7.0 38 39# August 2009. 40# 41# The code was revised to minimize code size and to maximize 42# "distance" between instructions producing input to 'lea' 43# instruction and the 'lea' instruction itself, which is essential 44# for Intel Atom core. 45 46# October 2010. 47# 48# Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it 49# is to offload message schedule denoted by Wt in NIST specification, 50# or Xupdate in OpenSSL source, to SIMD unit. See sha1-586.pl module 51# for background and implementation details. The only difference from 52# 32-bit code is that 64-bit code doesn't have to spill @X[] elements 53# to free temporary registers. 54 55# April 2011. 56# 57# Add AVX code path. See sha1-586.pl for further information. 58 59# May 2013. 60# 61# Add AVX2+BMI code path. Initial attempt (utilizing BMI instructions 62# and loading pair of consecutive blocks to 256-bit %ymm registers) 63# did not provide impressive performance improvement till a crucial 64# hint regarding the number of Xupdate iterations to pre-compute in 65# advance was provided by Ilya Albrekht of Intel Corp. 66 67# March 2014. 68# 69# Add support for Intel SHA Extensions. 70 71###################################################################### 72# Current performance is summarized in following table. Numbers are 73# CPU clock cycles spent to process single byte (less is better). 74# 75# x86_64 SSSE3 AVX[2] 76# P4 9.05 - 77# Opteron 6.26 - 78# Core2 6.55 6.05/+8% - 79# Westmere 6.73 5.30/+27% - 80# Sandy Bridge 7.70 6.10/+26% 4.99/+54% 81# Ivy Bridge 6.06 4.67/+30% 4.60/+32% 82# Haswell 5.45 4.15/+31% 3.57/+53% 83# Skylake 5.18 4.06/+28% 3.54/+46% 84# Bulldozer 9.11 5.95/+53% 85# Ryzen 4.75 3.80/+24% 1.93/+150%(**) 86# VIA Nano 9.32 7.15/+30% 87# Atom 10.3 9.17/+12% 88# Silvermont 13.1(*) 9.37/+40% 89# Knights L 13.2(*) 9.68/+36% 8.30/+59% 90# Goldmont 8.13 6.42/+27% 1.70/+380%(**) 91# 92# (*) obviously suboptimal result, nothing was done about it, 93# because SSSE3 code is compiled unconditionally; 94# (**) SHAEXT result 95 96$flavour = shift; 97$output = shift; 98if ($flavour =~ /\./) { $output = $flavour; undef $flavour; } 99 100$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); 101 102$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 103( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or 104( $xlate="${dir}../../../perlasm/x86_64-xlate.pl" and -f $xlate) or 105die "can't locate x86_64-xlate.pl"; 106 107# In upstream, this is controlled by shelling out to the compiler to check 108# versions, but BoringSSL is intended to be used with pre-generated perlasm 109# output, so this isn't useful anyway. 110# 111# TODO(davidben): Enable AVX2 code after testing by setting $avx to 2. Is it 112# necessary to disable AVX2 code when SHA Extensions code is disabled? Upstream 113# did not tie them together until after $shaext was added. 114$avx = 1; 115 116# TODO(davidben): Consider enabling the Intel SHA Extensions code once it's 117# been tested. 118$shaext=0; ### set to zero if compiling for 1.0.1 119$avx=1 if (!$shaext && $avx); 120 121open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""; 122*STDOUT=*OUT; 123 124$ctx="%rdi"; # 1st arg 125$inp="%rsi"; # 2nd arg 126$num="%rdx"; # 3rd arg 127 128# reassign arguments in order to produce more compact code 129$ctx="%r8"; 130$inp="%r9"; 131$num="%r10"; 132 133$t0="%eax"; 134$t1="%ebx"; 135$t2="%ecx"; 136@xi=("%edx","%ebp","%r14d"); 137$A="%esi"; 138$B="%edi"; 139$C="%r11d"; 140$D="%r12d"; 141$E="%r13d"; 142 143@V=($A,$B,$C,$D,$E); 144 145sub BODY_00_19 { 146my ($i,$a,$b,$c,$d,$e)=@_; 147my $j=$i+1; 148$code.=<<___ if ($i==0); 149 mov `4*$i`($inp),$xi[0] 150 bswap $xi[0] 151___ 152$code.=<<___ if ($i<15); 153 mov `4*$j`($inp),$xi[1] 154 mov $d,$t0 155 mov $xi[0],`4*$i`(%rsp) 156 mov $a,$t2 157 bswap $xi[1] 158 xor $c,$t0 159 rol \$5,$t2 160 and $b,$t0 161 lea 0x5a827999($xi[0],$e),$e 162 add $t2,$e 163 xor $d,$t0 164 rol \$30,$b 165 add $t0,$e 166___ 167$code.=<<___ if ($i>=15); 168 xor `4*($j%16)`(%rsp),$xi[1] 169 mov $d,$t0 170 mov $xi[0],`4*($i%16)`(%rsp) 171 mov $a,$t2 172 xor `4*(($j+2)%16)`(%rsp),$xi[1] 173 xor $c,$t0 174 rol \$5,$t2 175 xor `4*(($j+8)%16)`(%rsp),$xi[1] 176 and $b,$t0 177 lea 0x5a827999($xi[0],$e),$e 178 rol \$30,$b 179 xor $d,$t0 180 add $t2,$e 181 rol \$1,$xi[1] 182 add $t0,$e 183___ 184push(@xi,shift(@xi)); 185} 186 187sub BODY_20_39 { 188my ($i,$a,$b,$c,$d,$e)=@_; 189my $j=$i+1; 190my $K=($i<40)?0x6ed9eba1:0xca62c1d6; 191$code.=<<___ if ($i<79); 192 xor `4*($j%16)`(%rsp),$xi[1] 193 mov $b,$t0 194 `"mov $xi[0],".4*($i%16)."(%rsp)" if ($i<72)` 195 mov $a,$t2 196 xor `4*(($j+2)%16)`(%rsp),$xi[1] 197 xor $d,$t0 198 rol \$5,$t2 199 xor `4*(($j+8)%16)`(%rsp),$xi[1] 200 lea $K($xi[0],$e),$e 201 xor $c,$t0 202 add $t2,$e 203 rol \$30,$b 204 add $t0,$e 205 rol \$1,$xi[1] 206___ 207$code.=<<___ if ($i==79); 208 mov $b,$t0 209 mov $a,$t2 210 xor $d,$t0 211 lea $K($xi[0],$e),$e 212 rol \$5,$t2 213 xor $c,$t0 214 add $t2,$e 215 rol \$30,$b 216 add $t0,$e 217___ 218push(@xi,shift(@xi)); 219} 220 221sub BODY_40_59 { 222my ($i,$a,$b,$c,$d,$e)=@_; 223my $j=$i+1; 224$code.=<<___; 225 xor `4*($j%16)`(%rsp),$xi[1] 226 mov $d,$t0 227 mov $xi[0],`4*($i%16)`(%rsp) 228 mov $d,$t1 229 xor `4*(($j+2)%16)`(%rsp),$xi[1] 230 and $c,$t0 231 mov $a,$t2 232 xor `4*(($j+8)%16)`(%rsp),$xi[1] 233 lea 0x8f1bbcdc($xi[0],$e),$e 234 xor $c,$t1 235 rol \$5,$t2 236 add $t0,$e 237 rol \$1,$xi[1] 238 and $b,$t1 239 add $t2,$e 240 rol \$30,$b 241 add $t1,$e 242___ 243push(@xi,shift(@xi)); 244} 245 246$code.=<<___; 247.text 248.extern OPENSSL_ia32cap_P 249 250.globl sha1_block_data_order 251.type sha1_block_data_order,\@function,3 252.align 16 253sha1_block_data_order: 254.cfi_startproc 255 leaq OPENSSL_ia32cap_P(%rip),%r10 256 mov 0(%r10),%r9d 257 mov 4(%r10),%r8d 258 mov 8(%r10),%r10d 259 test \$`1<<9`,%r8d # check SSSE3 bit 260 jz .Lialu 261___ 262$code.=<<___ if ($shaext); 263 test \$`1<<29`,%r10d # check SHA bit 264 jnz _shaext_shortcut 265___ 266$code.=<<___ if ($avx>1); 267 and \$`1<<3|1<<5|1<<8`,%r10d # check AVX2+BMI1+BMI2 268 cmp \$`1<<3|1<<5|1<<8`,%r10d 269 je _avx2_shortcut 270___ 271$code.=<<___ if ($avx); 272 and \$`1<<28`,%r8d # mask AVX bit 273 and \$`1<<30`,%r9d # mask "Intel CPU" bit 274 or %r9d,%r8d 275 cmp \$`1<<28|1<<30`,%r8d 276 je _avx_shortcut 277___ 278$code.=<<___; 279 jmp _ssse3_shortcut 280 281.align 16 282.Lialu: 283 mov %rsp,%rax 284.cfi_def_cfa_register %rax 285 push %rbx 286.cfi_push %rbx 287 push %rbp 288.cfi_push %rbp 289 push %r12 290.cfi_push %r12 291 push %r13 292.cfi_push %r13 293 push %r14 294.cfi_push %r14 295 mov %rdi,$ctx # reassigned argument 296 sub \$`8+16*4`,%rsp 297 mov %rsi,$inp # reassigned argument 298 and \$-64,%rsp 299 mov %rdx,$num # reassigned argument 300 mov %rax,`16*4`(%rsp) 301.cfi_cfa_expression %rsp+64,deref,+8 302.Lprologue: 303 304 mov 0($ctx),$A 305 mov 4($ctx),$B 306 mov 8($ctx),$C 307 mov 12($ctx),$D 308 mov 16($ctx),$E 309 jmp .Lloop 310 311.align 16 312.Lloop: 313___ 314for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); } 315for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); } 316for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); } 317for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); } 318$code.=<<___; 319 add 0($ctx),$A 320 add 4($ctx),$B 321 add 8($ctx),$C 322 add 12($ctx),$D 323 add 16($ctx),$E 324 mov $A,0($ctx) 325 mov $B,4($ctx) 326 mov $C,8($ctx) 327 mov $D,12($ctx) 328 mov $E,16($ctx) 329 330 sub \$1,$num 331 lea `16*4`($inp),$inp 332 jnz .Lloop 333 334 mov `16*4`(%rsp),%rsi 335.cfi_def_cfa %rsi,8 336 mov -40(%rsi),%r14 337.cfi_restore %r14 338 mov -32(%rsi),%r13 339.cfi_restore %r13 340 mov -24(%rsi),%r12 341.cfi_restore %r12 342 mov -16(%rsi),%rbp 343.cfi_restore %rbp 344 mov -8(%rsi),%rbx 345.cfi_restore %rbx 346 lea (%rsi),%rsp 347.cfi_def_cfa_register %rsp 348.Lepilogue: 349 ret 350.cfi_endproc 351.size sha1_block_data_order,.-sha1_block_data_order 352___ 353if ($shaext) {{{ 354###################################################################### 355# Intel SHA Extensions implementation of SHA1 update function. 356# 357my ($ctx,$inp,$num)=("%rdi","%rsi","%rdx"); 358my ($ABCD,$E,$E_,$BSWAP,$ABCD_SAVE,$E_SAVE)=map("%xmm$_",(0..3,8,9)); 359my @MSG=map("%xmm$_",(4..7)); 360 361$code.=<<___; 362.type sha1_block_data_order_shaext,\@function,3 363.align 32 364sha1_block_data_order_shaext: 365_shaext_shortcut: 366.cfi_startproc 367___ 368$code.=<<___ if ($win64); 369 lea `-8-4*16`(%rsp),%rsp 370 movaps %xmm6,-8-4*16(%rax) 371 movaps %xmm7,-8-3*16(%rax) 372 movaps %xmm8,-8-2*16(%rax) 373 movaps %xmm9,-8-1*16(%rax) 374.Lprologue_shaext: 375___ 376$code.=<<___; 377 movdqu ($ctx),$ABCD 378 movd 16($ctx),$E 379 movdqa K_XX_XX+0xa0(%rip),$BSWAP # byte-n-word swap 380 381 movdqu ($inp),@MSG[0] 382 pshufd \$0b00011011,$ABCD,$ABCD # flip word order 383 movdqu 0x10($inp),@MSG[1] 384 pshufd \$0b00011011,$E,$E # flip word order 385 movdqu 0x20($inp),@MSG[2] 386 pshufb $BSWAP,@MSG[0] 387 movdqu 0x30($inp),@MSG[3] 388 pshufb $BSWAP,@MSG[1] 389 pshufb $BSWAP,@MSG[2] 390 movdqa $E,$E_SAVE # offload $E 391 pshufb $BSWAP,@MSG[3] 392 jmp .Loop_shaext 393 394.align 16 395.Loop_shaext: 396 dec $num 397 lea 0x40($inp),%r8 # next input block 398 paddd @MSG[0],$E 399 cmovne %r8,$inp 400 movdqa $ABCD,$ABCD_SAVE # offload $ABCD 401___ 402for($i=0;$i<20-4;$i+=2) { 403$code.=<<___; 404 sha1msg1 @MSG[1],@MSG[0] 405 movdqa $ABCD,$E_ 406 sha1rnds4 \$`int($i/5)`,$E,$ABCD # 0-3... 407 sha1nexte @MSG[1],$E_ 408 pxor @MSG[2],@MSG[0] 409 sha1msg1 @MSG[2],@MSG[1] 410 sha1msg2 @MSG[3],@MSG[0] 411 412 movdqa $ABCD,$E 413 sha1rnds4 \$`int(($i+1)/5)`,$E_,$ABCD 414 sha1nexte @MSG[2],$E 415 pxor @MSG[3],@MSG[1] 416 sha1msg2 @MSG[0],@MSG[1] 417___ 418 push(@MSG,shift(@MSG)); push(@MSG,shift(@MSG)); 419} 420$code.=<<___; 421 movdqu ($inp),@MSG[0] 422 movdqa $ABCD,$E_ 423 sha1rnds4 \$3,$E,$ABCD # 64-67 424 sha1nexte @MSG[1],$E_ 425 movdqu 0x10($inp),@MSG[1] 426 pshufb $BSWAP,@MSG[0] 427 428 movdqa $ABCD,$E 429 sha1rnds4 \$3,$E_,$ABCD # 68-71 430 sha1nexte @MSG[2],$E 431 movdqu 0x20($inp),@MSG[2] 432 pshufb $BSWAP,@MSG[1] 433 434 movdqa $ABCD,$E_ 435 sha1rnds4 \$3,$E,$ABCD # 72-75 436 sha1nexte @MSG[3],$E_ 437 movdqu 0x30($inp),@MSG[3] 438 pshufb $BSWAP,@MSG[2] 439 440 movdqa $ABCD,$E 441 sha1rnds4 \$3,$E_,$ABCD # 76-79 442 sha1nexte $E_SAVE,$E 443 pshufb $BSWAP,@MSG[3] 444 445 paddd $ABCD_SAVE,$ABCD 446 movdqa $E,$E_SAVE # offload $E 447 448 jnz .Loop_shaext 449 450 pshufd \$0b00011011,$ABCD,$ABCD 451 pshufd \$0b00011011,$E,$E 452 movdqu $ABCD,($ctx) 453 movd $E,16($ctx) 454___ 455$code.=<<___ if ($win64); 456 movaps -8-4*16(%rax),%xmm6 457 movaps -8-3*16(%rax),%xmm7 458 movaps -8-2*16(%rax),%xmm8 459 movaps -8-1*16(%rax),%xmm9 460 mov %rax,%rsp 461.Lepilogue_shaext: 462___ 463$code.=<<___; 464.cfi_endproc 465 ret 466.size sha1_block_data_order_shaext,.-sha1_block_data_order_shaext 467___ 468}}} 469{{{ 470my $Xi=4; 471my @X=map("%xmm$_",(4..7,0..3)); 472my @Tx=map("%xmm$_",(8..10)); 473my $Kx="%xmm11"; 474my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization 475my @T=("%esi","%edi"); 476my $j=0; 477my $rx=0; 478my $K_XX_XX="%r14"; 479my $fp="%r11"; 480 481my $_rol=sub { &rol(@_) }; 482my $_ror=sub { &ror(@_) }; 483 484{ my $sn; 485sub align32() { 486 ++$sn; 487$code.=<<___; 488 jmp .Lalign32_$sn # see "Decoded ICache" in manual 489.align 32 490.Lalign32_$sn: 491___ 492} 493} 494 495$code.=<<___; 496.type sha1_block_data_order_ssse3,\@function,3 497.align 16 498sha1_block_data_order_ssse3: 499_ssse3_shortcut: 500.cfi_startproc 501 mov %rsp,$fp # frame pointer 502.cfi_def_cfa_register $fp 503 push %rbx 504.cfi_push %rbx 505 push %rbp 506.cfi_push %rbp 507 push %r12 508.cfi_push %r12 509 push %r13 # redundant, done to share Win64 SE handler 510.cfi_push %r13 511 push %r14 512.cfi_push %r14 513 lea `-64-($win64?6*16:0)`(%rsp),%rsp 514___ 515$code.=<<___ if ($win64); 516 movaps %xmm6,-40-6*16($fp) 517 movaps %xmm7,-40-5*16($fp) 518 movaps %xmm8,-40-4*16($fp) 519 movaps %xmm9,-40-3*16($fp) 520 movaps %xmm10,-40-2*16($fp) 521 movaps %xmm11,-40-1*16($fp) 522.Lprologue_ssse3: 523___ 524$code.=<<___; 525 and \$-64,%rsp 526 mov %rdi,$ctx # reassigned argument 527 mov %rsi,$inp # reassigned argument 528 mov %rdx,$num # reassigned argument 529 530 shl \$6,$num 531 add $inp,$num 532 lea K_XX_XX+64(%rip),$K_XX_XX 533 534 mov 0($ctx),$A # load context 535 mov 4($ctx),$B 536 mov 8($ctx),$C 537 mov 12($ctx),$D 538 mov $B,@T[0] # magic seed 539 mov 16($ctx),$E 540 mov $C,@T[1] 541 xor $D,@T[1] 542 and @T[1],@T[0] 543 544 movdqa 64($K_XX_XX),@X[2] # pbswap mask 545 movdqa -64($K_XX_XX),@Tx[1] # K_00_19 546 movdqu 0($inp),@X[-4&7] # load input to %xmm[0-3] 547 movdqu 16($inp),@X[-3&7] 548 movdqu 32($inp),@X[-2&7] 549 movdqu 48($inp),@X[-1&7] 550 pshufb @X[2],@X[-4&7] # byte swap 551 pshufb @X[2],@X[-3&7] 552 pshufb @X[2],@X[-2&7] 553 add \$64,$inp 554 paddd @Tx[1],@X[-4&7] # add K_00_19 555 pshufb @X[2],@X[-1&7] 556 paddd @Tx[1],@X[-3&7] 557 paddd @Tx[1],@X[-2&7] 558 movdqa @X[-4&7],0(%rsp) # X[]+K xfer to IALU 559 psubd @Tx[1],@X[-4&7] # restore X[] 560 movdqa @X[-3&7],16(%rsp) 561 psubd @Tx[1],@X[-3&7] 562 movdqa @X[-2&7],32(%rsp) 563 psubd @Tx[1],@X[-2&7] 564 jmp .Loop_ssse3 565___ 566 567sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm 568{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; 569 my $arg = pop; 570 $arg = "\$$arg" if ($arg*1 eq $arg); 571 $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n"; 572} 573 574sub Xupdate_ssse3_16_31() # recall that $Xi starts with 4 575{ use integer; 576 my $body = shift; 577 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions 578 my ($a,$b,$c,$d,$e); 579 580 eval(shift(@insns)); # ror 581 &pshufd (@X[0],@X[-4&7],0xee); # was &movdqa (@X[0],@X[-3&7]); 582 eval(shift(@insns)); 583 &movdqa (@Tx[0],@X[-1&7]); 584 &paddd (@Tx[1],@X[-1&7]); 585 eval(shift(@insns)); 586 eval(shift(@insns)); 587 588 &punpcklqdq(@X[0],@X[-3&7]); # compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8); 589 eval(shift(@insns)); 590 eval(shift(@insns)); # rol 591 eval(shift(@insns)); 592 &psrldq (@Tx[0],4); # "X[-3]", 3 dwords 593 eval(shift(@insns)); 594 eval(shift(@insns)); 595 596 &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]" 597 eval(shift(@insns)); 598 eval(shift(@insns)); # ror 599 &pxor (@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]" 600 eval(shift(@insns)); 601 eval(shift(@insns)); 602 eval(shift(@insns)); 603 604 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]" 605 eval(shift(@insns)); 606 eval(shift(@insns)); # rol 607 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU 608 eval(shift(@insns)); 609 eval(shift(@insns)); 610 611 &movdqa (@Tx[2],@X[0]); 612 eval(shift(@insns)); 613 eval(shift(@insns)); 614 eval(shift(@insns)); # ror 615 &movdqa (@Tx[0],@X[0]); 616 eval(shift(@insns)); 617 618 &pslldq (@Tx[2],12); # "X[0]"<<96, extract one dword 619 &paddd (@X[0],@X[0]); 620 eval(shift(@insns)); 621 eval(shift(@insns)); 622 623 &psrld (@Tx[0],31); 624 eval(shift(@insns)); 625 eval(shift(@insns)); # rol 626 eval(shift(@insns)); 627 &movdqa (@Tx[1],@Tx[2]); 628 eval(shift(@insns)); 629 eval(shift(@insns)); 630 631 &psrld (@Tx[2],30); 632 eval(shift(@insns)); 633 eval(shift(@insns)); # ror 634 &por (@X[0],@Tx[0]); # "X[0]"<<<=1 635 eval(shift(@insns)); 636 eval(shift(@insns)); 637 eval(shift(@insns)); 638 639 &pslld (@Tx[1],2); 640 &pxor (@X[0],@Tx[2]); 641 eval(shift(@insns)); 642 &movdqa (@Tx[2],eval(2*16*(($Xi)/5)-64)."($K_XX_XX)"); # K_XX_XX 643 eval(shift(@insns)); # rol 644 eval(shift(@insns)); 645 eval(shift(@insns)); 646 647 &pxor (@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2 648 &pshufd (@Tx[1],@X[-1&7],0xee) if ($Xi==7); # was &movdqa (@Tx[0],@X[-1&7]) in Xupdate_ssse3_32_79 649 650 foreach (@insns) { eval; } # remaining instructions [if any] 651 652 $Xi++; push(@X,shift(@X)); # "rotate" X[] 653 push(@Tx,shift(@Tx)); 654} 655 656sub Xupdate_ssse3_32_79() 657{ use integer; 658 my $body = shift; 659 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions 660 my ($a,$b,$c,$d,$e); 661 662 eval(shift(@insns)) if ($Xi==8); 663 &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" 664 eval(shift(@insns)) if ($Xi==8); 665 eval(shift(@insns)); # body_20_39 666 eval(shift(@insns)); 667 eval(shift(@insns)) if (@insns[1] =~ /_ror/); 668 eval(shift(@insns)) if (@insns[0] =~ /_ror/); 669 &punpcklqdq(@Tx[0],@X[-1&7]); # compose "X[-6]", was &palignr(@Tx[0],@X[-2&7],8); 670 eval(shift(@insns)); 671 eval(shift(@insns)); # rol 672 673 &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]" 674 eval(shift(@insns)); 675 eval(shift(@insns)); 676 if ($Xi%5) { 677 &movdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX... 678 } else { # ... or load next one 679 &movdqa (@Tx[2],eval(2*16*($Xi/5)-64)."($K_XX_XX)"); 680 } 681 eval(shift(@insns)); # ror 682 &paddd (@Tx[1],@X[-1&7]); 683 eval(shift(@insns)); 684 685 &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-6]" 686 eval(shift(@insns)); # body_20_39 687 eval(shift(@insns)); 688 eval(shift(@insns)); 689 eval(shift(@insns)); # rol 690 eval(shift(@insns)) if (@insns[0] =~ /_ror/); 691 692 &movdqa (@Tx[0],@X[0]); 693 eval(shift(@insns)); 694 eval(shift(@insns)); 695 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU 696 eval(shift(@insns)); # ror 697 eval(shift(@insns)); 698 eval(shift(@insns)); # body_20_39 699 700 &pslld (@X[0],2); 701 eval(shift(@insns)); 702 eval(shift(@insns)); 703 &psrld (@Tx[0],30); 704 eval(shift(@insns)) if (@insns[0] =~ /_rol/);# rol 705 eval(shift(@insns)); 706 eval(shift(@insns)); 707 eval(shift(@insns)); # ror 708 709 &por (@X[0],@Tx[0]); # "X[0]"<<<=2 710 eval(shift(@insns)); 711 eval(shift(@insns)); # body_20_39 712 eval(shift(@insns)) if (@insns[1] =~ /_rol/); 713 eval(shift(@insns)) if (@insns[0] =~ /_rol/); 714 &pshufd(@Tx[1],@X[-1&7],0xee) if ($Xi<19); # was &movdqa (@Tx[1],@X[0]) 715 eval(shift(@insns)); 716 eval(shift(@insns)); # rol 717 eval(shift(@insns)); 718 eval(shift(@insns)); 719 eval(shift(@insns)); # rol 720 eval(shift(@insns)); 721 722 foreach (@insns) { eval; } # remaining instructions 723 724 $Xi++; push(@X,shift(@X)); # "rotate" X[] 725 push(@Tx,shift(@Tx)); 726} 727 728sub Xuplast_ssse3_80() 729{ use integer; 730 my $body = shift; 731 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 732 my ($a,$b,$c,$d,$e); 733 734 eval(shift(@insns)); 735 eval(shift(@insns)); 736 eval(shift(@insns)); 737 eval(shift(@insns)); 738 &paddd (@Tx[1],@X[-1&7]); 739 eval(shift(@insns)); 740 eval(shift(@insns)); 741 742 &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU 743 744 foreach (@insns) { eval; } # remaining instructions 745 746 &cmp ($inp,$num); 747 &je (".Ldone_ssse3"); 748 749 unshift(@Tx,pop(@Tx)); 750 751 &movdqa (@X[2],"64($K_XX_XX)"); # pbswap mask 752 &movdqa (@Tx[1],"-64($K_XX_XX)"); # K_00_19 753 &movdqu (@X[-4&7],"0($inp)"); # load input 754 &movdqu (@X[-3&7],"16($inp)"); 755 &movdqu (@X[-2&7],"32($inp)"); 756 &movdqu (@X[-1&7],"48($inp)"); 757 &pshufb (@X[-4&7],@X[2]); # byte swap 758 &add ($inp,64); 759 760 $Xi=0; 761} 762 763sub Xloop_ssse3() 764{ use integer; 765 my $body = shift; 766 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 767 my ($a,$b,$c,$d,$e); 768 769 eval(shift(@insns)); 770 eval(shift(@insns)); 771 eval(shift(@insns)); 772 &pshufb (@X[($Xi-3)&7],@X[2]); 773 eval(shift(@insns)); 774 eval(shift(@insns)); 775 eval(shift(@insns)); 776 eval(shift(@insns)); 777 &paddd (@X[($Xi-4)&7],@Tx[1]); 778 eval(shift(@insns)); 779 eval(shift(@insns)); 780 eval(shift(@insns)); 781 eval(shift(@insns)); 782 &movdqa (eval(16*$Xi)."(%rsp)",@X[($Xi-4)&7]); # X[]+K xfer to IALU 783 eval(shift(@insns)); 784 eval(shift(@insns)); 785 eval(shift(@insns)); 786 eval(shift(@insns)); 787 &psubd (@X[($Xi-4)&7],@Tx[1]); 788 789 foreach (@insns) { eval; } 790 $Xi++; 791} 792 793sub Xtail_ssse3() 794{ use integer; 795 my $body = shift; 796 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 797 my ($a,$b,$c,$d,$e); 798 799 foreach (@insns) { eval; } 800} 801 802sub body_00_19 () { # ((c^d)&b)^d 803 # on start @T[0]=(c^d)&b 804 return &body_20_39() if ($rx==19); $rx++; 805 ( 806 '($a,$b,$c,$d,$e)=@V;'. 807 '&$_ror ($b,$j?7:2)', # $b>>>2 808 '&xor (@T[0],$d)', 809 '&mov (@T[1],$a)', # $b for next round 810 811 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer 812 '&xor ($b,$c)', # $c^$d for next round 813 814 '&$_rol ($a,5)', 815 '&add ($e,@T[0])', 816 '&and (@T[1],$b)', # ($b&($c^$d)) for next round 817 818 '&xor ($b,$c)', # restore $b 819 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' 820 ); 821} 822 823sub body_20_39 () { # b^d^c 824 # on entry @T[0]=b^d 825 return &body_40_59() if ($rx==39); $rx++; 826 ( 827 '($a,$b,$c,$d,$e)=@V;'. 828 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer 829 '&xor (@T[0],$d) if($j==19);'. 830 '&xor (@T[0],$c) if($j> 19)', # ($b^$d^$c) 831 '&mov (@T[1],$a)', # $b for next round 832 833 '&$_rol ($a,5)', 834 '&add ($e,@T[0])', 835 '&xor (@T[1],$c) if ($j< 79)', # $b^$d for next round 836 837 '&$_ror ($b,7)', # $b>>>2 838 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' 839 ); 840} 841 842sub body_40_59 () { # ((b^c)&(c^d))^c 843 # on entry @T[0]=(b^c), (c^=d) 844 $rx++; 845 ( 846 '($a,$b,$c,$d,$e)=@V;'. 847 '&add ($e,eval(4*($j&15))."(%rsp)")', # X[]+K xfer 848 '&and (@T[0],$c) if ($j>=40)', # (b^c)&(c^d) 849 '&xor ($c,$d) if ($j>=40)', # restore $c 850 851 '&$_ror ($b,7)', # $b>>>2 852 '&mov (@T[1],$a)', # $b for next round 853 '&xor (@T[0],$c)', 854 855 '&$_rol ($a,5)', 856 '&add ($e,@T[0])', 857 '&xor (@T[1],$c) if ($j==59);'. 858 '&xor (@T[1],$b) if ($j< 59)', # b^c for next round 859 860 '&xor ($b,$c) if ($j< 59)', # c^d for next round 861 '&add ($e,$a);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' 862 ); 863} 864$code.=<<___; 865.align 16 866.Loop_ssse3: 867___ 868 &Xupdate_ssse3_16_31(\&body_00_19); 869 &Xupdate_ssse3_16_31(\&body_00_19); 870 &Xupdate_ssse3_16_31(\&body_00_19); 871 &Xupdate_ssse3_16_31(\&body_00_19); 872 &Xupdate_ssse3_32_79(\&body_00_19); 873 &Xupdate_ssse3_32_79(\&body_20_39); 874 &Xupdate_ssse3_32_79(\&body_20_39); 875 &Xupdate_ssse3_32_79(\&body_20_39); 876 &Xupdate_ssse3_32_79(\&body_20_39); 877 &Xupdate_ssse3_32_79(\&body_20_39); 878 &Xupdate_ssse3_32_79(\&body_40_59); 879 &Xupdate_ssse3_32_79(\&body_40_59); 880 &Xupdate_ssse3_32_79(\&body_40_59); 881 &Xupdate_ssse3_32_79(\&body_40_59); 882 &Xupdate_ssse3_32_79(\&body_40_59); 883 &Xupdate_ssse3_32_79(\&body_20_39); 884 &Xuplast_ssse3_80(\&body_20_39); # can jump to "done" 885 886 $saved_j=$j; @saved_V=@V; 887 888 &Xloop_ssse3(\&body_20_39); 889 &Xloop_ssse3(\&body_20_39); 890 &Xloop_ssse3(\&body_20_39); 891 892$code.=<<___; 893 add 0($ctx),$A # update context 894 add 4($ctx),@T[0] 895 add 8($ctx),$C 896 add 12($ctx),$D 897 mov $A,0($ctx) 898 add 16($ctx),$E 899 mov @T[0],4($ctx) 900 mov @T[0],$B # magic seed 901 mov $C,8($ctx) 902 mov $C,@T[1] 903 mov $D,12($ctx) 904 xor $D,@T[1] 905 mov $E,16($ctx) 906 and @T[1],@T[0] 907 jmp .Loop_ssse3 908 909.align 16 910.Ldone_ssse3: 911___ 912 $j=$saved_j; @V=@saved_V; 913 914 &Xtail_ssse3(\&body_20_39); 915 &Xtail_ssse3(\&body_20_39); 916 &Xtail_ssse3(\&body_20_39); 917 918$code.=<<___; 919 add 0($ctx),$A # update context 920 add 4($ctx),@T[0] 921 add 8($ctx),$C 922 mov $A,0($ctx) 923 add 12($ctx),$D 924 mov @T[0],4($ctx) 925 add 16($ctx),$E 926 mov $C,8($ctx) 927 mov $D,12($ctx) 928 mov $E,16($ctx) 929___ 930$code.=<<___ if ($win64); 931 movaps -40-6*16($fp),%xmm6 932 movaps -40-5*16($fp),%xmm7 933 movaps -40-4*16($fp),%xmm8 934 movaps -40-3*16($fp),%xmm9 935 movaps -40-2*16($fp),%xmm10 936 movaps -40-1*16($fp),%xmm11 937___ 938$code.=<<___; 939 mov -40($fp),%r14 940.cfi_restore %r14 941 mov -32($fp),%r13 942.cfi_restore %r13 943 mov -24($fp),%r12 944.cfi_restore %r12 945 mov -16($fp),%rbp 946.cfi_restore %rbp 947 mov -8($fp),%rbx 948.cfi_restore %rbx 949 lea ($fp),%rsp 950.cfi_def_cfa_register %rsp 951.Lepilogue_ssse3: 952 ret 953.cfi_endproc 954.size sha1_block_data_order_ssse3,.-sha1_block_data_order_ssse3 955___ 956 957if ($avx) { 958$Xi=4; # reset variables 959@X=map("%xmm$_",(4..7,0..3)); 960@Tx=map("%xmm$_",(8..10)); 961$j=0; 962$rx=0; 963 964my $done_avx_label=".Ldone_avx"; 965 966my $_rol=sub { &shld(@_[0],@_) }; 967my $_ror=sub { &shrd(@_[0],@_) }; 968 969$code.=<<___; 970.type sha1_block_data_order_avx,\@function,3 971.align 16 972sha1_block_data_order_avx: 973_avx_shortcut: 974.cfi_startproc 975 mov %rsp,$fp 976.cfi_def_cfa_register $fp 977 push %rbx 978.cfi_push %rbx 979 push %rbp 980.cfi_push %rbp 981 push %r12 982.cfi_push %r12 983 push %r13 # redundant, done to share Win64 SE handler 984.cfi_push %r13 985 push %r14 986.cfi_push %r14 987 lea `-64-($win64?6*16:0)`(%rsp),%rsp 988 vzeroupper 989___ 990$code.=<<___ if ($win64); 991 vmovaps %xmm6,-40-6*16($fp) 992 vmovaps %xmm7,-40-5*16($fp) 993 vmovaps %xmm8,-40-4*16($fp) 994 vmovaps %xmm9,-40-3*16($fp) 995 vmovaps %xmm10,-40-2*16($fp) 996 vmovaps %xmm11,-40-1*16($fp) 997.Lprologue_avx: 998___ 999$code.=<<___; 1000 and \$-64,%rsp 1001 mov %rdi,$ctx # reassigned argument 1002 mov %rsi,$inp # reassigned argument 1003 mov %rdx,$num # reassigned argument 1004 1005 shl \$6,$num 1006 add $inp,$num 1007 lea K_XX_XX+64(%rip),$K_XX_XX 1008 1009 mov 0($ctx),$A # load context 1010 mov 4($ctx),$B 1011 mov 8($ctx),$C 1012 mov 12($ctx),$D 1013 mov $B,@T[0] # magic seed 1014 mov 16($ctx),$E 1015 mov $C,@T[1] 1016 xor $D,@T[1] 1017 and @T[1],@T[0] 1018 1019 vmovdqa 64($K_XX_XX),@X[2] # pbswap mask 1020 vmovdqa -64($K_XX_XX),$Kx # K_00_19 1021 vmovdqu 0($inp),@X[-4&7] # load input to %xmm[0-3] 1022 vmovdqu 16($inp),@X[-3&7] 1023 vmovdqu 32($inp),@X[-2&7] 1024 vmovdqu 48($inp),@X[-1&7] 1025 vpshufb @X[2],@X[-4&7],@X[-4&7] # byte swap 1026 add \$64,$inp 1027 vpshufb @X[2],@X[-3&7],@X[-3&7] 1028 vpshufb @X[2],@X[-2&7],@X[-2&7] 1029 vpshufb @X[2],@X[-1&7],@X[-1&7] 1030 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19 1031 vpaddd $Kx,@X[-3&7],@X[1] 1032 vpaddd $Kx,@X[-2&7],@X[2] 1033 vmovdqa @X[0],0(%rsp) # X[]+K xfer to IALU 1034 vmovdqa @X[1],16(%rsp) 1035 vmovdqa @X[2],32(%rsp) 1036 jmp .Loop_avx 1037___ 1038 1039sub Xupdate_avx_16_31() # recall that $Xi starts with 4 1040{ use integer; 1041 my $body = shift; 1042 my @insns = (&$body,&$body,&$body,&$body); # 40 instructions 1043 my ($a,$b,$c,$d,$e); 1044 1045 eval(shift(@insns)); 1046 eval(shift(@insns)); 1047 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]" 1048 eval(shift(@insns)); 1049 eval(shift(@insns)); 1050 1051 &vpaddd (@Tx[1],$Kx,@X[-1&7]); 1052 eval(shift(@insns)); 1053 eval(shift(@insns)); 1054 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords 1055 eval(shift(@insns)); 1056 eval(shift(@insns)); 1057 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]" 1058 eval(shift(@insns)); 1059 eval(shift(@insns)); 1060 1061 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]" 1062 eval(shift(@insns)); 1063 eval(shift(@insns)); 1064 eval(shift(@insns)); 1065 eval(shift(@insns)); 1066 1067 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]" 1068 eval(shift(@insns)); 1069 eval(shift(@insns)); 1070 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU 1071 eval(shift(@insns)); 1072 eval(shift(@insns)); 1073 1074 &vpsrld (@Tx[0],@X[0],31); 1075 eval(shift(@insns)); 1076 eval(shift(@insns)); 1077 eval(shift(@insns)); 1078 eval(shift(@insns)); 1079 1080 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword 1081 &vpaddd (@X[0],@X[0],@X[0]); 1082 eval(shift(@insns)); 1083 eval(shift(@insns)); 1084 eval(shift(@insns)); 1085 eval(shift(@insns)); 1086 1087 &vpsrld (@Tx[1],@Tx[2],30); 1088 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1 1089 eval(shift(@insns)); 1090 eval(shift(@insns)); 1091 eval(shift(@insns)); 1092 eval(shift(@insns)); 1093 1094 &vpslld (@Tx[2],@Tx[2],2); 1095 &vpxor (@X[0],@X[0],@Tx[1]); 1096 eval(shift(@insns)); 1097 eval(shift(@insns)); 1098 eval(shift(@insns)); 1099 eval(shift(@insns)); 1100 1101 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2 1102 eval(shift(@insns)); 1103 eval(shift(@insns)); 1104 &vmovdqa ($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX 1105 eval(shift(@insns)); 1106 eval(shift(@insns)); 1107 1108 1109 foreach (@insns) { eval; } # remaining instructions [if any] 1110 1111 $Xi++; push(@X,shift(@X)); # "rotate" X[] 1112} 1113 1114sub Xupdate_avx_32_79() 1115{ use integer; 1116 my $body = shift; 1117 my @insns = (&$body,&$body,&$body,&$body); # 32 to 44 instructions 1118 my ($a,$b,$c,$d,$e); 1119 1120 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]" 1121 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" 1122 eval(shift(@insns)); # body_20_39 1123 eval(shift(@insns)); 1124 eval(shift(@insns)); 1125 eval(shift(@insns)); # rol 1126 1127 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]" 1128 eval(shift(@insns)); 1129 eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/); 1130 &vpaddd (@Tx[1],$Kx,@X[-1&7]); 1131 &vmovdqa ($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0); 1132 eval(shift(@insns)); # ror 1133 eval(shift(@insns)); 1134 1135 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]" 1136 eval(shift(@insns)); # body_20_39 1137 eval(shift(@insns)); 1138 eval(shift(@insns)); 1139 eval(shift(@insns)); # rol 1140 1141 &vpsrld (@Tx[0],@X[0],30); 1142 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU 1143 eval(shift(@insns)); 1144 eval(shift(@insns)); 1145 eval(shift(@insns)); # ror 1146 eval(shift(@insns)); 1147 1148 &vpslld (@X[0],@X[0],2); 1149 eval(shift(@insns)); # body_20_39 1150 eval(shift(@insns)); 1151 eval(shift(@insns)); 1152 eval(shift(@insns)); # rol 1153 eval(shift(@insns)); 1154 eval(shift(@insns)); 1155 eval(shift(@insns)); # ror 1156 eval(shift(@insns)); 1157 1158 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2 1159 eval(shift(@insns)); # body_20_39 1160 eval(shift(@insns)); 1161 eval(shift(@insns)); 1162 eval(shift(@insns)); # rol 1163 eval(shift(@insns)); 1164 eval(shift(@insns)); 1165 eval(shift(@insns)); # rol 1166 eval(shift(@insns)); 1167 1168 foreach (@insns) { eval; } # remaining instructions 1169 1170 $Xi++; push(@X,shift(@X)); # "rotate" X[] 1171} 1172 1173sub Xuplast_avx_80() 1174{ use integer; 1175 my $body = shift; 1176 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 1177 my ($a,$b,$c,$d,$e); 1178 1179 eval(shift(@insns)); 1180 &vpaddd (@Tx[1],$Kx,@X[-1&7]); 1181 eval(shift(@insns)); 1182 eval(shift(@insns)); 1183 eval(shift(@insns)); 1184 eval(shift(@insns)); 1185 1186 &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU 1187 1188 foreach (@insns) { eval; } # remaining instructions 1189 1190 &cmp ($inp,$num); 1191 &je ($done_avx_label); 1192 1193 &vmovdqa(@X[2],"64($K_XX_XX)"); # pbswap mask 1194 &vmovdqa($Kx,"-64($K_XX_XX)"); # K_00_19 1195 &vmovdqu(@X[-4&7],"0($inp)"); # load input 1196 &vmovdqu(@X[-3&7],"16($inp)"); 1197 &vmovdqu(@X[-2&7],"32($inp)"); 1198 &vmovdqu(@X[-1&7],"48($inp)"); 1199 &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap 1200 &add ($inp,64); 1201 1202 $Xi=0; 1203} 1204 1205sub Xloop_avx() 1206{ use integer; 1207 my $body = shift; 1208 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 1209 my ($a,$b,$c,$d,$e); 1210 1211 eval(shift(@insns)); 1212 eval(shift(@insns)); 1213 &vpshufb(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]); 1214 eval(shift(@insns)); 1215 eval(shift(@insns)); 1216 &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],$Kx); 1217 eval(shift(@insns)); 1218 eval(shift(@insns)); 1219 eval(shift(@insns)); 1220 eval(shift(@insns)); 1221 &vmovdqa(eval(16*$Xi)."(%rsp)",@X[$Xi&7]); # X[]+K xfer to IALU 1222 eval(shift(@insns)); 1223 eval(shift(@insns)); 1224 1225 foreach (@insns) { eval; } 1226 $Xi++; 1227} 1228 1229sub Xtail_avx() 1230{ use integer; 1231 my $body = shift; 1232 my @insns = (&$body,&$body,&$body,&$body); # 32 instructions 1233 my ($a,$b,$c,$d,$e); 1234 1235 foreach (@insns) { eval; } 1236} 1237 1238$code.=<<___; 1239.align 16 1240.Loop_avx: 1241___ 1242 &Xupdate_avx_16_31(\&body_00_19); 1243 &Xupdate_avx_16_31(\&body_00_19); 1244 &Xupdate_avx_16_31(\&body_00_19); 1245 &Xupdate_avx_16_31(\&body_00_19); 1246 &Xupdate_avx_32_79(\&body_00_19); 1247 &Xupdate_avx_32_79(\&body_20_39); 1248 &Xupdate_avx_32_79(\&body_20_39); 1249 &Xupdate_avx_32_79(\&body_20_39); 1250 &Xupdate_avx_32_79(\&body_20_39); 1251 &Xupdate_avx_32_79(\&body_20_39); 1252 &Xupdate_avx_32_79(\&body_40_59); 1253 &Xupdate_avx_32_79(\&body_40_59); 1254 &Xupdate_avx_32_79(\&body_40_59); 1255 &Xupdate_avx_32_79(\&body_40_59); 1256 &Xupdate_avx_32_79(\&body_40_59); 1257 &Xupdate_avx_32_79(\&body_20_39); 1258 &Xuplast_avx_80(\&body_20_39); # can jump to "done" 1259 1260 $saved_j=$j; @saved_V=@V; 1261 1262 &Xloop_avx(\&body_20_39); 1263 &Xloop_avx(\&body_20_39); 1264 &Xloop_avx(\&body_20_39); 1265 1266$code.=<<___; 1267 add 0($ctx),$A # update context 1268 add 4($ctx),@T[0] 1269 add 8($ctx),$C 1270 add 12($ctx),$D 1271 mov $A,0($ctx) 1272 add 16($ctx),$E 1273 mov @T[0],4($ctx) 1274 mov @T[0],$B # magic seed 1275 mov $C,8($ctx) 1276 mov $C,@T[1] 1277 mov $D,12($ctx) 1278 xor $D,@T[1] 1279 mov $E,16($ctx) 1280 and @T[1],@T[0] 1281 jmp .Loop_avx 1282 1283.align 16 1284$done_avx_label: 1285___ 1286 $j=$saved_j; @V=@saved_V; 1287 1288 &Xtail_avx(\&body_20_39); 1289 &Xtail_avx(\&body_20_39); 1290 &Xtail_avx(\&body_20_39); 1291 1292$code.=<<___; 1293 vzeroupper 1294 1295 add 0($ctx),$A # update context 1296 add 4($ctx),@T[0] 1297 add 8($ctx),$C 1298 mov $A,0($ctx) 1299 add 12($ctx),$D 1300 mov @T[0],4($ctx) 1301 add 16($ctx),$E 1302 mov $C,8($ctx) 1303 mov $D,12($ctx) 1304 mov $E,16($ctx) 1305___ 1306$code.=<<___ if ($win64); 1307 movaps -40-6*16($fp),%xmm6 1308 movaps -40-5*16($fp),%xmm7 1309 movaps -40-4*16($fp),%xmm8 1310 movaps -40-3*16($fp),%xmm9 1311 movaps -40-2*16($fp),%xmm10 1312 movaps -40-1*16($fp),%xmm11 1313___ 1314$code.=<<___; 1315 mov -40($fp),%r14 1316.cfi_restore %r14 1317 mov -32($fp),%r13 1318.cfi_restore %r13 1319 mov -24($fp),%r12 1320.cfi_restore %r12 1321 mov -16($fp),%rbp 1322.cfi_restore %rbp 1323 mov -8($fp),%rbx 1324.cfi_restore %rbx 1325 lea ($fp),%rsp 1326.cfi_def_cfa_register %rsp 1327.Lepilogue_avx: 1328 ret 1329.cfi_endproc 1330.size sha1_block_data_order_avx,.-sha1_block_data_order_avx 1331___ 1332 1333if ($avx>1) { 1334use integer; 1335$Xi=4; # reset variables 1336@X=map("%ymm$_",(4..7,0..3)); 1337@Tx=map("%ymm$_",(8..10)); 1338$Kx="%ymm11"; 1339$j=0; 1340 1341my @ROTX=("%eax","%ebp","%ebx","%ecx","%edx","%esi"); 1342my ($a5,$t0)=("%r12d","%edi"); 1343 1344my ($A,$F,$B,$C,$D,$E)=@ROTX; 1345my $rx=0; 1346my $frame="%r13"; 1347 1348$code.=<<___; 1349.type sha1_block_data_order_avx2,\@function,3 1350.align 16 1351sha1_block_data_order_avx2: 1352_avx2_shortcut: 1353.cfi_startproc 1354 mov %rsp,$fp 1355.cfi_def_cfa_register $fp 1356 push %rbx 1357.cfi_push %rbx 1358 push %rbp 1359.cfi_push %rbp 1360 push %r12 1361.cfi_push %r12 1362 push %r13 1363.cfi_push %r13 1364 push %r14 1365.cfi_push %r14 1366 vzeroupper 1367___ 1368$code.=<<___ if ($win64); 1369 lea -6*16(%rsp),%rsp 1370 vmovaps %xmm6,-40-6*16($fp) 1371 vmovaps %xmm7,-40-5*16($fp) 1372 vmovaps %xmm8,-40-4*16($fp) 1373 vmovaps %xmm9,-40-3*16($fp) 1374 vmovaps %xmm10,-40-2*16($fp) 1375 vmovaps %xmm11,-40-1*16($fp) 1376.Lprologue_avx2: 1377___ 1378$code.=<<___; 1379 mov %rdi,$ctx # reassigned argument 1380 mov %rsi,$inp # reassigned argument 1381 mov %rdx,$num # reassigned argument 1382 1383 lea -640(%rsp),%rsp 1384 shl \$6,$num 1385 lea 64($inp),$frame 1386 and \$-128,%rsp 1387 add $inp,$num 1388 lea K_XX_XX+64(%rip),$K_XX_XX 1389 1390 mov 0($ctx),$A # load context 1391 cmp $num,$frame 1392 cmovae $inp,$frame # next or same block 1393 mov 4($ctx),$F 1394 mov 8($ctx),$C 1395 mov 12($ctx),$D 1396 mov 16($ctx),$E 1397 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask 1398 1399 vmovdqu ($inp),%xmm0 1400 vmovdqu 16($inp),%xmm1 1401 vmovdqu 32($inp),%xmm2 1402 vmovdqu 48($inp),%xmm3 1403 lea 64($inp),$inp 1404 vinserti128 \$1,($frame),@X[-4&7],@X[-4&7] 1405 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7] 1406 vpshufb @X[2],@X[-4&7],@X[-4&7] 1407 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7] 1408 vpshufb @X[2],@X[-3&7],@X[-3&7] 1409 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7] 1410 vpshufb @X[2],@X[-2&7],@X[-2&7] 1411 vmovdqu -64($K_XX_XX),$Kx # K_00_19 1412 vpshufb @X[2],@X[-1&7],@X[-1&7] 1413 1414 vpaddd $Kx,@X[-4&7],@X[0] # add K_00_19 1415 vpaddd $Kx,@X[-3&7],@X[1] 1416 vmovdqu @X[0],0(%rsp) # X[]+K xfer to IALU 1417 vpaddd $Kx,@X[-2&7],@X[2] 1418 vmovdqu @X[1],32(%rsp) 1419 vpaddd $Kx,@X[-1&7],@X[3] 1420 vmovdqu @X[2],64(%rsp) 1421 vmovdqu @X[3],96(%rsp) 1422___ 1423for (;$Xi<8;$Xi++) { # Xupdate_avx2_16_31 1424 use integer; 1425 1426 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]" 1427 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords 1428 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]" 1429 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]" 1430 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]" 1431 &vpsrld (@Tx[0],@X[0],31); 1432 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX 1433 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword 1434 &vpaddd (@X[0],@X[0],@X[0]); 1435 &vpsrld (@Tx[1],@Tx[2],30); 1436 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1 1437 &vpslld (@Tx[2],@Tx[2],2); 1438 &vpxor (@X[0],@X[0],@Tx[1]); 1439 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2 1440 &vpaddd (@Tx[1],@X[0],$Kx); 1441 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU 1442 1443 push(@X,shift(@X)); # "rotate" X[] 1444} 1445$code.=<<___; 1446 lea 128(%rsp),$frame 1447 jmp .Loop_avx2 1448.align 32 1449.Loop_avx2: 1450 rorx \$2,$F,$B 1451 andn $D,$F,$t0 1452 and $C,$F 1453 xor $t0,$F 1454___ 1455sub bodyx_00_19 () { # 8 instructions, 3 cycles critical path 1456 # at start $f=(b&c)^(~b&d), $b>>>=2 1457 return &bodyx_20_39() if ($rx==19); $rx++; 1458 ( 1459 '($a,$f,$b,$c,$d,$e)=@ROTX;'. 1460 1461 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K 1462 '&lea ($frame,"256($frame)") if ($j%32==31);', 1463 '&andn ($t0,$a,$c)', # ~b&d for next round 1464 1465 '&add ($e,$f)', # e+=(b&c)^(~b&d) 1466 '&rorx ($a5,$a,27)', # a<<<5 1467 '&rorx ($f,$a,2)', # b>>>2 for next round 1468 '&and ($a,$b)', # b&c for next round 1469 1470 '&add ($e,$a5)', # e+=a<<<5 1471 '&xor ($a,$t0);'. # f=(b&c)^(~b&d) for next round 1472 1473 'unshift(@ROTX,pop(@ROTX)); $j++;' 1474 ) 1475} 1476 1477sub bodyx_20_39 () { # 7 instructions, 2 cycles critical path 1478 # on entry $f=b^c^d, $b>>>=2 1479 return &bodyx_40_59() if ($rx==39); $rx++; 1480 ( 1481 '($a,$f,$b,$c,$d,$e)=@ROTX;'. 1482 1483 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K 1484 '&lea ($frame,"256($frame)") if ($j%32==31);', 1485 1486 '&lea ($e,"($e,$f)")', # e+=b^c^d 1487 '&rorx ($a5,$a,27)', # a<<<5 1488 '&rorx ($f,$a,2) if ($j<79)', # b>>>2 in next round 1489 '&xor ($a,$b) if ($j<79)', # b^c for next round 1490 1491 '&add ($e,$a5)', # e+=a<<<5 1492 '&xor ($a,$c) if ($j<79);'. # f=b^c^d for next round 1493 1494 'unshift(@ROTX,pop(@ROTX)); $j++;' 1495 ) 1496} 1497 1498sub bodyx_40_59 () { # 10 instructions, 3 cycles critical path 1499 # on entry $f=((b^c)&(c^d)), $b>>>=2 1500 $rx++; 1501 ( 1502 '($a,$f,$b,$c,$d,$e)=@ROTX;'. 1503 1504 '&add ($e,((32*($j/4)+4*($j%4))%256-128)."($frame)");'. # e+=X[i]+K 1505 '&lea ($frame,"256($frame)") if ($j%32==31);', 1506 '&xor ($f,$c) if ($j>39)', # (b^c)&(c^d)^c 1507 '&mov ($t0,$b) if ($j<59)', # count on zero latency 1508 '&xor ($t0,$c) if ($j<59)', # c^d for next round 1509 1510 '&lea ($e,"($e,$f)")', # e+=(b^c)&(c^d)^c 1511 '&rorx ($a5,$a,27)', # a<<<5 1512 '&rorx ($f,$a,2)', # b>>>2 in next round 1513 '&xor ($a,$b)', # b^c for next round 1514 1515 '&add ($e,$a5)', # e+=a<<<5 1516 '&and ($a,$t0) if ($j< 59);'. # f=(b^c)&(c^d) for next round 1517 '&xor ($a,$c) if ($j==59);'. # f=b^c^d for next round 1518 1519 'unshift(@ROTX,pop(@ROTX)); $j++;' 1520 ) 1521} 1522 1523sub Xupdate_avx2_16_31() # recall that $Xi starts with 4 1524{ use integer; 1525 my $body = shift; 1526 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 instructions 1527 my ($a,$b,$c,$d,$e); 1528 1529 &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]" 1530 eval(shift(@insns)); 1531 eval(shift(@insns)); 1532 eval(shift(@insns)); 1533 eval(shift(@insns)); 1534 1535 &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords 1536 eval(shift(@insns)); 1537 eval(shift(@insns)); 1538 eval(shift(@insns)); 1539 1540 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]" 1541 &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]" 1542 eval(shift(@insns)); 1543 eval(shift(@insns)); 1544 1545 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]" 1546 eval(shift(@insns)); 1547 eval(shift(@insns)); 1548 eval(shift(@insns)); 1549 eval(shift(@insns)); 1550 1551 &vpsrld (@Tx[0],@X[0],31); 1552 &vmovdqu($Kx,eval(2*16*(($Xi)/5)-64)."($K_XX_XX)") if ($Xi%5==0); # K_XX_XX 1553 eval(shift(@insns)); 1554 eval(shift(@insns)); 1555 eval(shift(@insns)); 1556 1557 &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword 1558 &vpaddd (@X[0],@X[0],@X[0]); 1559 eval(shift(@insns)); 1560 eval(shift(@insns)); 1561 1562 &vpsrld (@Tx[1],@Tx[2],30); 1563 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1 1564 eval(shift(@insns)); 1565 eval(shift(@insns)); 1566 1567 &vpslld (@Tx[2],@Tx[2],2); 1568 &vpxor (@X[0],@X[0],@Tx[1]); 1569 eval(shift(@insns)); 1570 eval(shift(@insns)); 1571 1572 &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2 1573 eval(shift(@insns)); 1574 eval(shift(@insns)); 1575 eval(shift(@insns)); 1576 1577 &vpaddd (@Tx[1],@X[0],$Kx); 1578 eval(shift(@insns)); 1579 eval(shift(@insns)); 1580 eval(shift(@insns)); 1581 &vmovdqu(eval(32*($Xi))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU 1582 1583 foreach (@insns) { eval; } # remaining instructions [if any] 1584 1585 $Xi++; 1586 push(@X,shift(@X)); # "rotate" X[] 1587} 1588 1589sub Xupdate_avx2_32_79() 1590{ use integer; 1591 my $body = shift; 1592 my @insns = (&$body,&$body,&$body,&$body,&$body); # 35 to 50 instructions 1593 my ($a,$b,$c,$d,$e); 1594 1595 &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]" 1596 &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" 1597 eval(shift(@insns)); 1598 eval(shift(@insns)); 1599 1600 &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]" 1601 &vmovdqu($Kx,eval(2*16*($Xi/5)-64)."($K_XX_XX)") if ($Xi%5==0); 1602 eval(shift(@insns)); 1603 eval(shift(@insns)); 1604 eval(shift(@insns)); 1605 1606 &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]" 1607 eval(shift(@insns)); 1608 eval(shift(@insns)); 1609 eval(shift(@insns)); 1610 1611 &vpsrld (@Tx[0],@X[0],30); 1612 &vpslld (@X[0],@X[0],2); 1613 eval(shift(@insns)); 1614 eval(shift(@insns)); 1615 eval(shift(@insns)); 1616 1617 #&vpslld (@X[0],@X[0],2); 1618 eval(shift(@insns)); 1619 eval(shift(@insns)); 1620 eval(shift(@insns)); 1621 1622 &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2 1623 eval(shift(@insns)); 1624 eval(shift(@insns)); 1625 eval(shift(@insns)); 1626 eval(shift(@insns)); 1627 1628 &vpaddd (@Tx[1],@X[0],$Kx); 1629 eval(shift(@insns)); 1630 eval(shift(@insns)); 1631 eval(shift(@insns)); 1632 eval(shift(@insns)); 1633 1634 &vmovdqu("32*$Xi(%rsp)",@Tx[1]); # X[]+K xfer to IALU 1635 1636 foreach (@insns) { eval; } # remaining instructions 1637 1638 $Xi++; 1639 push(@X,shift(@X)); # "rotate" X[] 1640} 1641 1642sub Xloop_avx2() 1643{ use integer; 1644 my $body = shift; 1645 my @insns = (&$body,&$body,&$body,&$body,&$body); # 32 instructions 1646 my ($a,$b,$c,$d,$e); 1647 1648 foreach (@insns) { eval; } 1649} 1650 1651 &align32(); 1652 &Xupdate_avx2_32_79(\&bodyx_00_19); 1653 &Xupdate_avx2_32_79(\&bodyx_00_19); 1654 &Xupdate_avx2_32_79(\&bodyx_00_19); 1655 &Xupdate_avx2_32_79(\&bodyx_00_19); 1656 1657 &Xupdate_avx2_32_79(\&bodyx_20_39); 1658 &Xupdate_avx2_32_79(\&bodyx_20_39); 1659 &Xupdate_avx2_32_79(\&bodyx_20_39); 1660 &Xupdate_avx2_32_79(\&bodyx_20_39); 1661 1662 &align32(); 1663 &Xupdate_avx2_32_79(\&bodyx_40_59); 1664 &Xupdate_avx2_32_79(\&bodyx_40_59); 1665 &Xupdate_avx2_32_79(\&bodyx_40_59); 1666 &Xupdate_avx2_32_79(\&bodyx_40_59); 1667 1668 &Xloop_avx2(\&bodyx_20_39); 1669 &Xloop_avx2(\&bodyx_20_39); 1670 &Xloop_avx2(\&bodyx_20_39); 1671 &Xloop_avx2(\&bodyx_20_39); 1672 1673$code.=<<___; 1674 lea 128($inp),$frame 1675 lea 128($inp),%rdi # borrow $t0 1676 cmp $num,$frame 1677 cmovae $inp,$frame # next or previous block 1678 1679 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c 1680 add 0($ctx),@ROTX[0] # update context 1681 add 4($ctx),@ROTX[1] 1682 add 8($ctx),@ROTX[3] 1683 mov @ROTX[0],0($ctx) 1684 add 12($ctx),@ROTX[4] 1685 mov @ROTX[1],4($ctx) 1686 mov @ROTX[0],$A # A=d 1687 add 16($ctx),@ROTX[5] 1688 mov @ROTX[3],$a5 1689 mov @ROTX[3],8($ctx) 1690 mov @ROTX[4],$D # D=b 1691 #xchg @ROTX[5],$F # F=c, C=f 1692 mov @ROTX[4],12($ctx) 1693 mov @ROTX[1],$F # F=e 1694 mov @ROTX[5],16($ctx) 1695 #mov $F,16($ctx) 1696 mov @ROTX[5],$E # E=c 1697 mov $a5,$C # C=f 1698 #xchg $F,$E # E=c, F=e 1699 1700 cmp $num,$inp 1701 je .Ldone_avx2 1702___ 1703 1704$Xi=4; # reset variables 1705@X=map("%ymm$_",(4..7,0..3)); 1706 1707$code.=<<___; 1708 vmovdqu 64($K_XX_XX),@X[2] # pbswap mask 1709 cmp $num,%rdi # borrowed $t0 1710 ja .Last_avx2 1711 1712 vmovdqu -64(%rdi),%xmm0 # low part of @X[-4&7] 1713 vmovdqu -48(%rdi),%xmm1 1714 vmovdqu -32(%rdi),%xmm2 1715 vmovdqu -16(%rdi),%xmm3 1716 vinserti128 \$1,0($frame),@X[-4&7],@X[-4&7] 1717 vinserti128 \$1,16($frame),@X[-3&7],@X[-3&7] 1718 vinserti128 \$1,32($frame),@X[-2&7],@X[-2&7] 1719 vinserti128 \$1,48($frame),@X[-1&7],@X[-1&7] 1720 jmp .Last_avx2 1721 1722.align 32 1723.Last_avx2: 1724 lea 128+16(%rsp),$frame 1725 rorx \$2,$F,$B 1726 andn $D,$F,$t0 1727 and $C,$F 1728 xor $t0,$F 1729 sub \$-128,$inp 1730___ 1731 $rx=$j=0; @ROTX=($A,$F,$B,$C,$D,$E); 1732 1733 &Xloop_avx2 (\&bodyx_00_19); 1734 &Xloop_avx2 (\&bodyx_00_19); 1735 &Xloop_avx2 (\&bodyx_00_19); 1736 &Xloop_avx2 (\&bodyx_00_19); 1737 1738 &Xloop_avx2 (\&bodyx_20_39); 1739 &vmovdqu ($Kx,"-64($K_XX_XX)"); # K_00_19 1740 &vpshufb (@X[-4&7],@X[-4&7],@X[2]); # byte swap 1741 &Xloop_avx2 (\&bodyx_20_39); 1742 &vpshufb (@X[-3&7],@X[-3&7],@X[2]); 1743 &vpaddd (@Tx[0],@X[-4&7],$Kx); # add K_00_19 1744 &Xloop_avx2 (\&bodyx_20_39); 1745 &vmovdqu ("0(%rsp)",@Tx[0]); 1746 &vpshufb (@X[-2&7],@X[-2&7],@X[2]); 1747 &vpaddd (@Tx[1],@X[-3&7],$Kx); 1748 &Xloop_avx2 (\&bodyx_20_39); 1749 &vmovdqu ("32(%rsp)",@Tx[1]); 1750 &vpshufb (@X[-1&7],@X[-1&7],@X[2]); 1751 &vpaddd (@X[2],@X[-2&7],$Kx); 1752 1753 &Xloop_avx2 (\&bodyx_40_59); 1754 &align32 (); 1755 &vmovdqu ("64(%rsp)",@X[2]); 1756 &vpaddd (@X[3],@X[-1&7],$Kx); 1757 &Xloop_avx2 (\&bodyx_40_59); 1758 &vmovdqu ("96(%rsp)",@X[3]); 1759 &Xloop_avx2 (\&bodyx_40_59); 1760 &Xupdate_avx2_16_31(\&bodyx_40_59); 1761 1762 &Xupdate_avx2_16_31(\&bodyx_20_39); 1763 &Xupdate_avx2_16_31(\&bodyx_20_39); 1764 &Xupdate_avx2_16_31(\&bodyx_20_39); 1765 &Xloop_avx2 (\&bodyx_20_39); 1766 1767$code.=<<___; 1768 lea 128(%rsp),$frame 1769 1770 # output is d-e-[a]-f-b-c => A=d,F=e,C=f,D=b,E=c 1771 add 0($ctx),@ROTX[0] # update context 1772 add 4($ctx),@ROTX[1] 1773 add 8($ctx),@ROTX[3] 1774 mov @ROTX[0],0($ctx) 1775 add 12($ctx),@ROTX[4] 1776 mov @ROTX[1],4($ctx) 1777 mov @ROTX[0],$A # A=d 1778 add 16($ctx),@ROTX[5] 1779 mov @ROTX[3],$a5 1780 mov @ROTX[3],8($ctx) 1781 mov @ROTX[4],$D # D=b 1782 #xchg @ROTX[5],$F # F=c, C=f 1783 mov @ROTX[4],12($ctx) 1784 mov @ROTX[1],$F # F=e 1785 mov @ROTX[5],16($ctx) 1786 #mov $F,16($ctx) 1787 mov @ROTX[5],$E # E=c 1788 mov $a5,$C # C=f 1789 #xchg $F,$E # E=c, F=e 1790 1791 cmp $num,$inp 1792 jbe .Loop_avx2 1793 1794.Ldone_avx2: 1795 vzeroupper 1796___ 1797$code.=<<___ if ($win64); 1798 movaps -40-6*16($fp),%xmm6 1799 movaps -40-5*16($fp),%xmm7 1800 movaps -40-4*16($fp),%xmm8 1801 movaps -40-3*16($fp),%xmm9 1802 movaps -40-2*16($fp),%xmm10 1803 movaps -40-1*16($fp),%xmm11 1804___ 1805$code.=<<___; 1806 mov -40($fp),%r14 1807.cfi_restore %r14 1808 mov -32($fp),%r13 1809.cfi_restore %r13 1810 mov -24($fp),%r12 1811.cfi_restore %r12 1812 mov -16($fp),%rbp 1813.cfi_restore %rbp 1814 mov -8($fp),%rbx 1815.cfi_restore %rbx 1816 lea ($fp),%rsp 1817.cfi_def_cfa_register %rsp 1818.Lepilogue_avx2: 1819 ret 1820.cfi_endproc 1821.size sha1_block_data_order_avx2,.-sha1_block_data_order_avx2 1822___ 1823} 1824} 1825$code.=<<___; 1826.align 64 1827K_XX_XX: 1828.long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19 1829.long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19 1830.long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39 1831.long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39 1832.long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59 1833.long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59 1834.long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79 1835.long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79 1836.long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask 1837.long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask 1838.byte 0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0 1839___ 1840}}} 1841$code.=<<___; 1842.asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>" 1843.align 64 1844___ 1845 1846# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame, 1847# CONTEXT *context,DISPATCHER_CONTEXT *disp) 1848if ($win64) { 1849$rec="%rcx"; 1850$frame="%rdx"; 1851$context="%r8"; 1852$disp="%r9"; 1853 1854$code.=<<___; 1855.extern __imp_RtlVirtualUnwind 1856.type se_handler,\@abi-omnipotent 1857.align 16 1858se_handler: 1859 push %rsi 1860 push %rdi 1861 push %rbx 1862 push %rbp 1863 push %r12 1864 push %r13 1865 push %r14 1866 push %r15 1867 pushfq 1868 sub \$64,%rsp 1869 1870 mov 120($context),%rax # pull context->Rax 1871 mov 248($context),%rbx # pull context->Rip 1872 1873 lea .Lprologue(%rip),%r10 1874 cmp %r10,%rbx # context->Rip<.Lprologue 1875 jb .Lcommon_seh_tail 1876 1877 mov 152($context),%rax # pull context->Rsp 1878 1879 lea .Lepilogue(%rip),%r10 1880 cmp %r10,%rbx # context->Rip>=.Lepilogue 1881 jae .Lcommon_seh_tail 1882 1883 mov `16*4`(%rax),%rax # pull saved stack pointer 1884 1885 mov -8(%rax),%rbx 1886 mov -16(%rax),%rbp 1887 mov -24(%rax),%r12 1888 mov -32(%rax),%r13 1889 mov -40(%rax),%r14 1890 mov %rbx,144($context) # restore context->Rbx 1891 mov %rbp,160($context) # restore context->Rbp 1892 mov %r12,216($context) # restore context->R12 1893 mov %r13,224($context) # restore context->R13 1894 mov %r14,232($context) # restore context->R14 1895 1896 jmp .Lcommon_seh_tail 1897.size se_handler,.-se_handler 1898___ 1899 1900$code.=<<___ if ($shaext); 1901.type shaext_handler,\@abi-omnipotent 1902.align 16 1903shaext_handler: 1904 push %rsi 1905 push %rdi 1906 push %rbx 1907 push %rbp 1908 push %r12 1909 push %r13 1910 push %r14 1911 push %r15 1912 pushfq 1913 sub \$64,%rsp 1914 1915 mov 120($context),%rax # pull context->Rax 1916 mov 248($context),%rbx # pull context->Rip 1917 1918 lea .Lprologue_shaext(%rip),%r10 1919 cmp %r10,%rbx # context->Rip<.Lprologue 1920 jb .Lcommon_seh_tail 1921 1922 lea .Lepilogue_shaext(%rip),%r10 1923 cmp %r10,%rbx # context->Rip>=.Lepilogue 1924 jae .Lcommon_seh_tail 1925 1926 lea -8-4*16(%rax),%rsi 1927 lea 512($context),%rdi # &context.Xmm6 1928 mov \$8,%ecx 1929 .long 0xa548f3fc # cld; rep movsq 1930 1931 jmp .Lcommon_seh_tail 1932.size shaext_handler,.-shaext_handler 1933___ 1934 1935$code.=<<___; 1936.type ssse3_handler,\@abi-omnipotent 1937.align 16 1938ssse3_handler: 1939 push %rsi 1940 push %rdi 1941 push %rbx 1942 push %rbp 1943 push %r12 1944 push %r13 1945 push %r14 1946 push %r15 1947 pushfq 1948 sub \$64,%rsp 1949 1950 mov 120($context),%rax # pull context->Rax 1951 mov 248($context),%rbx # pull context->Rip 1952 1953 mov 8($disp),%rsi # disp->ImageBase 1954 mov 56($disp),%r11 # disp->HandlerData 1955 1956 mov 0(%r11),%r10d # HandlerData[0] 1957 lea (%rsi,%r10),%r10 # prologue label 1958 cmp %r10,%rbx # context->Rip<prologue label 1959 jb .Lcommon_seh_tail 1960 1961 mov 208($context),%rax # pull context->R11 1962 1963 mov 4(%r11),%r10d # HandlerData[1] 1964 lea (%rsi,%r10),%r10 # epilogue label 1965 cmp %r10,%rbx # context->Rip>=epilogue label 1966 jae .Lcommon_seh_tail 1967 1968 lea -40-6*16(%rax),%rsi 1969 lea 512($context),%rdi # &context.Xmm6 1970 mov \$12,%ecx 1971 .long 0xa548f3fc # cld; rep movsq 1972 1973 mov -8(%rax),%rbx 1974 mov -16(%rax),%rbp 1975 mov -24(%rax),%r12 1976 mov -32(%rax),%r13 1977 mov -40(%rax),%r14 1978 mov %rbx,144($context) # restore context->Rbx 1979 mov %rbp,160($context) # restore context->Rbp 1980 mov %r12,216($context) # restore context->R12 1981 mov %r13,224($context) # restore context->R13 1982 mov %r14,232($context) # restore context->R14 1983 1984.Lcommon_seh_tail: 1985 mov 8(%rax),%rdi 1986 mov 16(%rax),%rsi 1987 mov %rax,152($context) # restore context->Rsp 1988 mov %rsi,168($context) # restore context->Rsi 1989 mov %rdi,176($context) # restore context->Rdi 1990 1991 mov 40($disp),%rdi # disp->ContextRecord 1992 mov $context,%rsi # context 1993 mov \$154,%ecx # sizeof(CONTEXT) 1994 .long 0xa548f3fc # cld; rep movsq 1995 1996 mov $disp,%rsi 1997 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER 1998 mov 8(%rsi),%rdx # arg2, disp->ImageBase 1999 mov 0(%rsi),%r8 # arg3, disp->ControlPc 2000 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry 2001 mov 40(%rsi),%r10 # disp->ContextRecord 2002 lea 56(%rsi),%r11 # &disp->HandlerData 2003 lea 24(%rsi),%r12 # &disp->EstablisherFrame 2004 mov %r10,32(%rsp) # arg5 2005 mov %r11,40(%rsp) # arg6 2006 mov %r12,48(%rsp) # arg7 2007 mov %rcx,56(%rsp) # arg8, (NULL) 2008 call *__imp_RtlVirtualUnwind(%rip) 2009 2010 mov \$1,%eax # ExceptionContinueSearch 2011 add \$64,%rsp 2012 popfq 2013 pop %r15 2014 pop %r14 2015 pop %r13 2016 pop %r12 2017 pop %rbp 2018 pop %rbx 2019 pop %rdi 2020 pop %rsi 2021 ret 2022.size ssse3_handler,.-ssse3_handler 2023 2024.section .pdata 2025.align 4 2026 .rva .LSEH_begin_sha1_block_data_order 2027 .rva .LSEH_end_sha1_block_data_order 2028 .rva .LSEH_info_sha1_block_data_order 2029___ 2030$code.=<<___ if ($shaext); 2031 .rva .LSEH_begin_sha1_block_data_order_shaext 2032 .rva .LSEH_end_sha1_block_data_order_shaext 2033 .rva .LSEH_info_sha1_block_data_order_shaext 2034___ 2035$code.=<<___; 2036 .rva .LSEH_begin_sha1_block_data_order_ssse3 2037 .rva .LSEH_end_sha1_block_data_order_ssse3 2038 .rva .LSEH_info_sha1_block_data_order_ssse3 2039___ 2040$code.=<<___ if ($avx); 2041 .rva .LSEH_begin_sha1_block_data_order_avx 2042 .rva .LSEH_end_sha1_block_data_order_avx 2043 .rva .LSEH_info_sha1_block_data_order_avx 2044___ 2045$code.=<<___ if ($avx>1); 2046 .rva .LSEH_begin_sha1_block_data_order_avx2 2047 .rva .LSEH_end_sha1_block_data_order_avx2 2048 .rva .LSEH_info_sha1_block_data_order_avx2 2049___ 2050$code.=<<___; 2051.section .xdata 2052.align 8 2053.LSEH_info_sha1_block_data_order: 2054 .byte 9,0,0,0 2055 .rva se_handler 2056___ 2057$code.=<<___ if ($shaext); 2058.LSEH_info_sha1_block_data_order_shaext: 2059 .byte 9,0,0,0 2060 .rva shaext_handler 2061___ 2062$code.=<<___; 2063.LSEH_info_sha1_block_data_order_ssse3: 2064 .byte 9,0,0,0 2065 .rva ssse3_handler 2066 .rva .Lprologue_ssse3,.Lepilogue_ssse3 # HandlerData[] 2067___ 2068$code.=<<___ if ($avx); 2069.LSEH_info_sha1_block_data_order_avx: 2070 .byte 9,0,0,0 2071 .rva ssse3_handler 2072 .rva .Lprologue_avx,.Lepilogue_avx # HandlerData[] 2073___ 2074$code.=<<___ if ($avx>1); 2075.LSEH_info_sha1_block_data_order_avx2: 2076 .byte 9,0,0,0 2077 .rva ssse3_handler 2078 .rva .Lprologue_avx2,.Lepilogue_avx2 # HandlerData[] 2079___ 2080} 2081 2082#################################################################### 2083 2084sub sha1rnds4 { 2085 if (@_[0] =~ /\$([x0-9a-f]+),\s*%xmm([0-7]),\s*%xmm([0-7])/) { 2086 my @opcode=(0x0f,0x3a,0xcc); 2087 push @opcode,0xc0|($2&7)|(($3&7)<<3); # ModR/M 2088 my $c=$1; 2089 push @opcode,$c=~/^0/?oct($c):$c; 2090 return ".byte\t".join(',',@opcode); 2091 } else { 2092 return "sha1rnds4\t".@_[0]; 2093 } 2094} 2095 2096sub sha1op38 { 2097 my $instr = shift; 2098 my %opcodelet = ( 2099 "sha1nexte" => 0xc8, 2100 "sha1msg1" => 0xc9, 2101 "sha1msg2" => 0xca ); 2102 2103 if (defined($opcodelet{$instr}) && @_[0] =~ /%xmm([0-9]+),\s*%xmm([0-9]+)/) { 2104 my @opcode=(0x0f,0x38); 2105 my $rex=0; 2106 $rex|=0x04 if ($2>=8); 2107 $rex|=0x01 if ($1>=8); 2108 unshift @opcode,0x40|$rex if ($rex); 2109 push @opcode,$opcodelet{$instr}; 2110 push @opcode,0xc0|($1&7)|(($2&7)<<3); # ModR/M 2111 return ".byte\t".join(',',@opcode); 2112 } else { 2113 return $instr."\t".@_[0]; 2114 } 2115} 2116 2117foreach (split("\n",$code)) { 2118 s/\`([^\`]*)\`/eval $1/geo; 2119 2120 s/\b(sha1rnds4)\s+(.*)/sha1rnds4($2)/geo or 2121 s/\b(sha1[^\s]*)\s+(.*)/sha1op38($1,$2)/geo; 2122 2123 print $_,"\n"; 2124} 2125close STDOUT or die "error closing STDOUT"; 2126