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 This is unlike on 21# Opteron where compiler-generated code was only 15% behind 32-bit 22# assembler, which originally made it hard to motivate the effort. 23# There was suggestion to mechanically translate 32-bit code, but I 24# dismissed it, reasoning that x86_64 offers enough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