Alexandre Savard | 1b09e31 | 2012-08-07 20:33:29 -0400 | [diff] [blame] | 1 | #!/usr/bin/env perl |
| 2 | # |
| 3 | # ==================================================================== |
| 4 | # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL |
| 5 | # project. The module is, however, dual licensed under OpenSSL and |
| 6 | # CRYPTOGAMS licenses depending on where you obtain it. For further |
| 7 | # details see http://www.openssl.org/~appro/cryptogams/. |
| 8 | # ==================================================================== |
| 9 | # |
| 10 | # sha1_block procedure for x86_64. |
| 11 | # |
| 12 | # It was brought to my attention that on EM64T compiler-generated code |
| 13 | # was far behind 32-bit assembler implementation. This is unlike on |
| 14 | # Opteron where compiler-generated code was only 15% behind 32-bit |
| 15 | # assembler, which originally made it hard to motivate the effort. |
| 16 | # There was suggestion to mechanically translate 32-bit code, but I |
| 17 | # dismissed it, reasoning that x86_64 offers enough register bank |
| 18 | # capacity to fully utilize SHA-1 parallelism. Therefore this fresh |
| 19 | # implementation:-) However! While 64-bit code does perform better |
| 20 | # on Opteron, I failed to beat 32-bit assembler on EM64T core. Well, |
| 21 | # x86_64 does offer larger *addressable* bank, but out-of-order core |
| 22 | # reaches for even more registers through dynamic aliasing, and EM64T |
| 23 | # core must have managed to run-time optimize even 32-bit code just as |
| 24 | # good as 64-bit one. Performance improvement is summarized in the |
| 25 | # following table: |
| 26 | # |
| 27 | # gcc 3.4 32-bit asm cycles/byte |
| 28 | # Opteron +45% +20% 6.8 |
| 29 | # Xeon P4 +65% +0% 9.9 |
| 30 | # Core2 +60% +10% 7.0 |
| 31 | |
| 32 | # August 2009. |
| 33 | # |
| 34 | # The code was revised to minimize code size and to maximize |
| 35 | # "distance" between instructions producing input to 'lea' |
| 36 | # instruction and the 'lea' instruction itself, which is essential |
| 37 | # for Intel Atom core. |
| 38 | |
| 39 | # October 2010. |
| 40 | # |
| 41 | # Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it |
| 42 | # is to offload message schedule denoted by Wt in NIST specification, |
| 43 | # or Xupdate in OpenSSL source, to SIMD unit. See sha1-586.pl module |
| 44 | # for background and implementation details. The only difference from |
| 45 | # 32-bit code is that 64-bit code doesn't have to spill @X[] elements |
| 46 | # to free temporary registers. |
| 47 | |
| 48 | # April 2011. |
| 49 | # |
| 50 | # Add AVX code path. See sha1-586.pl for further information. |
| 51 | |
| 52 | ###################################################################### |
| 53 | # Current performance is summarized in following table. Numbers are |
| 54 | # CPU clock cycles spent to process single byte (less is better). |
| 55 | # |
| 56 | # x86_64 SSSE3 AVX |
| 57 | # P4 9.8 - |
| 58 | # Opteron 6.6 - |
| 59 | # Core2 6.7 6.1/+10% - |
| 60 | # Atom 11.0 9.7/+13% - |
| 61 | # Westmere 7.1 5.6/+27% - |
| 62 | # Sandy Bridge 7.9 6.3/+25% 5.2/+51% |
| 63 | |
| 64 | $flavour = shift; |
| 65 | $output = shift; |
| 66 | if ($flavour =~ /\./) { $output = $flavour; undef $flavour; } |
| 67 | |
| 68 | $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); |
| 69 | |
| 70 | $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; |
| 71 | ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or |
| 72 | ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or |
| 73 | die "can't locate x86_64-xlate.pl"; |
| 74 | |
| 75 | $avx=1 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1` |
| 76 | =~ /GNU assembler version ([2-9]\.[0-9]+)/ && |
| 77 | $1>=2.19); |
| 78 | $avx=1 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) && |
| 79 | `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ && |
| 80 | $1>=2.09); |
| 81 | $avx=1 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) && |
| 82 | `ml64 2>&1` =~ /Version ([0-9]+)\./ && |
| 83 | $1>=10); |
| 84 | |
| 85 | open STDOUT,"| $^X $xlate $flavour $output"; |
| 86 | |
| 87 | $ctx="%rdi"; # 1st arg |
| 88 | $inp="%rsi"; # 2nd arg |
| 89 | $num="%rdx"; # 3rd arg |
| 90 | |
| 91 | # reassign arguments in order to produce more compact code |
| 92 | $ctx="%r8"; |
| 93 | $inp="%r9"; |
| 94 | $num="%r10"; |
| 95 | |
| 96 | $t0="%eax"; |
| 97 | $t1="%ebx"; |
| 98 | $t2="%ecx"; |
| 99 | @xi=("%edx","%ebp"); |
| 100 | $A="%esi"; |
| 101 | $B="%edi"; |
| 102 | $C="%r11d"; |
| 103 | $D="%r12d"; |
| 104 | $E="%r13d"; |
| 105 | |
| 106 | @V=($A,$B,$C,$D,$E); |
| 107 | |
| 108 | sub BODY_00_19 { |
| 109 | my ($i,$a,$b,$c,$d,$e)=@_; |
| 110 | my $j=$i+1; |
| 111 | $code.=<<___ if ($i==0); |
| 112 | mov `4*$i`($inp),$xi[0] |
| 113 | bswap $xi[0] |
| 114 | mov $xi[0],`4*$i`(%rsp) |
| 115 | ___ |
| 116 | $code.=<<___ if ($i<15); |
| 117 | mov $c,$t0 |
| 118 | mov `4*$j`($inp),$xi[1] |
| 119 | mov $a,$t2 |
| 120 | xor $d,$t0 |
| 121 | bswap $xi[1] |
| 122 | rol \$5,$t2 |
| 123 | lea 0x5a827999($xi[0],$e),$e |
| 124 | and $b,$t0 |
| 125 | mov $xi[1],`4*$j`(%rsp) |
| 126 | add $t2,$e |
| 127 | xor $d,$t0 |
| 128 | rol \$30,$b |
| 129 | add $t0,$e |
| 130 | ___ |
| 131 | $code.=<<___ if ($i>=15); |
| 132 | mov `4*($j%16)`(%rsp),$xi[1] |
| 133 | mov $c,$t0 |
| 134 | mov $a,$t2 |
| 135 | xor `4*(($j+2)%16)`(%rsp),$xi[1] |
| 136 | xor $d,$t0 |
| 137 | rol \$5,$t2 |
| 138 | xor `4*(($j+8)%16)`(%rsp),$xi[1] |
| 139 | and $b,$t0 |
| 140 | lea 0x5a827999($xi[0],$e),$e |
| 141 | xor `4*(($j+13)%16)`(%rsp),$xi[1] |
| 142 | xor $d,$t0 |
| 143 | rol \$1,$xi[1] |
| 144 | add $t2,$e |
| 145 | rol \$30,$b |
| 146 | mov $xi[1],`4*($j%16)`(%rsp) |
| 147 | add $t0,$e |
| 148 | ___ |
| 149 | unshift(@xi,pop(@xi)); |
| 150 | } |
| 151 | |
| 152 | sub BODY_20_39 { |
| 153 | my ($i,$a,$b,$c,$d,$e)=@_; |
| 154 | my $j=$i+1; |
| 155 | my $K=($i<40)?0x6ed9eba1:0xca62c1d6; |
| 156 | $code.=<<___ if ($i<79); |
| 157 | mov `4*($j%16)`(%rsp),$xi[1] |
| 158 | mov $c,$t0 |
| 159 | mov $a,$t2 |
| 160 | xor `4*(($j+2)%16)`(%rsp),$xi[1] |
| 161 | xor $b,$t0 |
| 162 | rol \$5,$t2 |
| 163 | lea $K($xi[0],$e),$e |
| 164 | xor `4*(($j+8)%16)`(%rsp),$xi[1] |
| 165 | xor $d,$t0 |
| 166 | add $t2,$e |
| 167 | xor `4*(($j+13)%16)`(%rsp),$xi[1] |
| 168 | rol \$30,$b |
| 169 | add $t0,$e |
| 170 | rol \$1,$xi[1] |
| 171 | ___ |
| 172 | $code.=<<___ if ($i<76); |
| 173 | mov $xi[1],`4*($j%16)`(%rsp) |
| 174 | ___ |
| 175 | $code.=<<___ if ($i==79); |
| 176 | mov $c,$t0 |
| 177 | mov $a,$t2 |
| 178 | xor $b,$t0 |
| 179 | lea $K($xi[0],$e),$e |
| 180 | rol \$5,$t2 |
| 181 | xor $d,$t0 |
| 182 | add $t2,$e |
| 183 | rol \$30,$b |
| 184 | add $t0,$e |
| 185 | ___ |
| 186 | unshift(@xi,pop(@xi)); |
| 187 | } |
| 188 | |
| 189 | sub BODY_40_59 { |
| 190 | my ($i,$a,$b,$c,$d,$e)=@_; |
| 191 | my $j=$i+1; |
| 192 | $code.=<<___; |
| 193 | mov `4*($j%16)`(%rsp),$xi[1] |
| 194 | mov $c,$t0 |
| 195 | mov $c,$t1 |
| 196 | xor `4*(($j+2)%16)`(%rsp),$xi[1] |
| 197 | and $d,$t0 |
| 198 | mov $a,$t2 |
| 199 | xor `4*(($j+8)%16)`(%rsp),$xi[1] |
| 200 | xor $d,$t1 |
| 201 | lea 0x8f1bbcdc($xi[0],$e),$e |
| 202 | rol \$5,$t2 |
| 203 | xor `4*(($j+13)%16)`(%rsp),$xi[1] |
| 204 | add $t0,$e |
| 205 | and $b,$t1 |
| 206 | rol \$1,$xi[1] |
| 207 | add $t1,$e |
| 208 | rol \$30,$b |
| 209 | mov $xi[1],`4*($j%16)`(%rsp) |
| 210 | add $t2,$e |
| 211 | ___ |
| 212 | unshift(@xi,pop(@xi)); |
| 213 | } |
| 214 | |
| 215 | $code.=<<___; |
| 216 | .text |
| 217 | .extern OPENSSL_ia32cap_P |
| 218 | |
| 219 | .globl sha1_block_data_order |
| 220 | .type sha1_block_data_order,\@function,3 |
| 221 | .align 16 |
| 222 | sha1_block_data_order: |
| 223 | mov OPENSSL_ia32cap_P+0(%rip),%r9d |
| 224 | mov OPENSSL_ia32cap_P+4(%rip),%r8d |
| 225 | test \$`1<<9`,%r8d # check SSSE3 bit |
| 226 | jz .Lialu |
| 227 | ___ |
| 228 | $code.=<<___ if ($avx); |
| 229 | and \$`1<<28`,%r8d # mask AVX bit |
| 230 | and \$`1<<30`,%r9d # mask "Intel CPU" bit |
| 231 | or %r9d,%r8d |
| 232 | cmp \$`1<<28|1<<30`,%r8d |
| 233 | je _avx_shortcut |
| 234 | ___ |
| 235 | $code.=<<___; |
| 236 | jmp _ssse3_shortcut |
| 237 | |
| 238 | .align 16 |
| 239 | .Lialu: |
| 240 | push %rbx |
| 241 | push %rbp |
| 242 | push %r12 |
| 243 | push %r13 |
| 244 | mov %rsp,%r11 |
| 245 | mov %rdi,$ctx # reassigned argument |
| 246 | sub \$`8+16*4`,%rsp |
| 247 | mov %rsi,$inp # reassigned argument |
| 248 | and \$-64,%rsp |
| 249 | mov %rdx,$num # reassigned argument |
| 250 | mov %r11,`16*4`(%rsp) |
| 251 | .Lprologue: |
| 252 | |
| 253 | mov 0($ctx),$A |
| 254 | mov 4($ctx),$B |
| 255 | mov 8($ctx),$C |
| 256 | mov 12($ctx),$D |
| 257 | mov 16($ctx),$E |
| 258 | jmp .Lloop |
| 259 | |
| 260 | .align 16 |
| 261 | .Lloop: |
| 262 | ___ |
| 263 | for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); } |
| 264 | for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); } |
| 265 | for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); } |
| 266 | for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); } |
| 267 | $code.=<<___; |
| 268 | add 0($ctx),$A |
| 269 | add 4($ctx),$B |
| 270 | add 8($ctx),$C |
| 271 | add 12($ctx),$D |
| 272 | add 16($ctx),$E |
| 273 | mov $A,0($ctx) |
| 274 | mov $B,4($ctx) |
| 275 | mov $C,8($ctx) |
| 276 | mov $D,12($ctx) |
| 277 | mov $E,16($ctx) |
| 278 | |
| 279 | sub \$1,$num |
| 280 | lea `16*4`($inp),$inp |
| 281 | jnz .Lloop |
| 282 | |
| 283 | mov `16*4`(%rsp),%rsi |
| 284 | mov (%rsi),%r13 |
| 285 | mov 8(%rsi),%r12 |
| 286 | mov 16(%rsi),%rbp |
| 287 | mov 24(%rsi),%rbx |
| 288 | lea 32(%rsi),%rsp |
| 289 | .Lepilogue: |
| 290 | ret |
| 291 | .size sha1_block_data_order,.-sha1_block_data_order |
| 292 | ___ |
| 293 | {{{ |
| 294 | my $Xi=4; |
| 295 | my @X=map("%xmm$_",(4..7,0..3)); |
| 296 | my @Tx=map("%xmm$_",(8..10)); |
| 297 | my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization |
| 298 | my @T=("%esi","%edi"); |
| 299 | my $j=0; |
| 300 | my $K_XX_XX="%r11"; |
| 301 | |
| 302 | my $_rol=sub { &rol(@_) }; |
| 303 | my $_ror=sub { &ror(@_) }; |
| 304 | |
| 305 | $code.=<<___; |
| 306 | .type sha1_block_data_order_ssse3,\@function,3 |
| 307 | .align 16 |
| 308 | sha1_block_data_order_ssse3: |
| 309 | _ssse3_shortcut: |
| 310 | push %rbx |
| 311 | push %rbp |
| 312 | push %r12 |
| 313 | lea `-64-($win64?5*16:0)`(%rsp),%rsp |
| 314 | ___ |
| 315 | $code.=<<___ if ($win64); |
| 316 | movaps %xmm6,64+0(%rsp) |
| 317 | movaps %xmm7,64+16(%rsp) |
| 318 | movaps %xmm8,64+32(%rsp) |
| 319 | movaps %xmm9,64+48(%rsp) |
| 320 | movaps %xmm10,64+64(%rsp) |
| 321 | .Lprologue_ssse3: |
| 322 | ___ |
| 323 | $code.=<<___; |
| 324 | mov %rdi,$ctx # reassigned argument |
| 325 | mov %rsi,$inp # reassigned argument |
| 326 | mov %rdx,$num # reassigned argument |
| 327 | |
| 328 | shl \$6,$num |
| 329 | add $inp,$num |
| 330 | lea K_XX_XX(%rip),$K_XX_XX |
| 331 | |
| 332 | mov 0($ctx),$A # load context |
| 333 | mov 4($ctx),$B |
| 334 | mov 8($ctx),$C |
| 335 | mov 12($ctx),$D |
| 336 | mov $B,@T[0] # magic seed |
| 337 | mov 16($ctx),$E |
| 338 | |
| 339 | movdqa 64($K_XX_XX),@X[2] # pbswap mask |
| 340 | movdqa 0($K_XX_XX),@Tx[1] # K_00_19 |
| 341 | movdqu 0($inp),@X[-4&7] # load input to %xmm[0-3] |
| 342 | movdqu 16($inp),@X[-3&7] |
| 343 | movdqu 32($inp),@X[-2&7] |
| 344 | movdqu 48($inp),@X[-1&7] |
| 345 | pshufb @X[2],@X[-4&7] # byte swap |
| 346 | add \$64,$inp |
| 347 | pshufb @X[2],@X[-3&7] |
| 348 | pshufb @X[2],@X[-2&7] |
| 349 | pshufb @X[2],@X[-1&7] |
| 350 | paddd @Tx[1],@X[-4&7] # add K_00_19 |
| 351 | paddd @Tx[1],@X[-3&7] |
| 352 | paddd @Tx[1],@X[-2&7] |
| 353 | movdqa @X[-4&7],0(%rsp) # X[]+K xfer to IALU |
| 354 | psubd @Tx[1],@X[-4&7] # restore X[] |
| 355 | movdqa @X[-3&7],16(%rsp) |
| 356 | psubd @Tx[1],@X[-3&7] |
| 357 | movdqa @X[-2&7],32(%rsp) |
| 358 | psubd @Tx[1],@X[-2&7] |
| 359 | jmp .Loop_ssse3 |
| 360 | ___ |
| 361 | |
| 362 | sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm |
| 363 | { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; |
| 364 | my $arg = pop; |
| 365 | $arg = "\$$arg" if ($arg*1 eq $arg); |
| 366 | $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n"; |
| 367 | } |
| 368 | |
| 369 | sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4 |
| 370 | { use integer; |
| 371 | my $body = shift; |
| 372 | my @insns = (&$body,&$body,&$body,&$body); # 40 instructions |
| 373 | my ($a,$b,$c,$d,$e); |
| 374 | |
| 375 | &movdqa (@X[0],@X[-3&7]); |
| 376 | eval(shift(@insns)); |
| 377 | eval(shift(@insns)); |
| 378 | &movdqa (@Tx[0],@X[-1&7]); |
| 379 | &palignr(@X[0],@X[-4&7],8); # compose "X[-14]" in "X[0]" |
| 380 | eval(shift(@insns)); |
| 381 | eval(shift(@insns)); |
| 382 | |
| 383 | &paddd (@Tx[1],@X[-1&7]); |
| 384 | eval(shift(@insns)); |
| 385 | eval(shift(@insns)); |
| 386 | &psrldq (@Tx[0],4); # "X[-3]", 3 dwords |
| 387 | eval(shift(@insns)); |
| 388 | eval(shift(@insns)); |
| 389 | &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]" |
| 390 | eval(shift(@insns)); |
| 391 | eval(shift(@insns)); |
| 392 | |
| 393 | &pxor (@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]" |
| 394 | eval(shift(@insns)); |
| 395 | eval(shift(@insns)); |
| 396 | eval(shift(@insns)); |
| 397 | eval(shift(@insns)); |
| 398 | |
| 399 | &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]" |
| 400 | eval(shift(@insns)); |
| 401 | eval(shift(@insns)); |
| 402 | &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU |
| 403 | eval(shift(@insns)); |
| 404 | eval(shift(@insns)); |
| 405 | |
| 406 | &movdqa (@Tx[2],@X[0]); |
| 407 | &movdqa (@Tx[0],@X[0]); |
| 408 | eval(shift(@insns)); |
| 409 | eval(shift(@insns)); |
| 410 | eval(shift(@insns)); |
| 411 | eval(shift(@insns)); |
| 412 | |
| 413 | &pslldq (@Tx[2],12); # "X[0]"<<96, extract one dword |
| 414 | &paddd (@X[0],@X[0]); |
| 415 | eval(shift(@insns)); |
| 416 | eval(shift(@insns)); |
| 417 | eval(shift(@insns)); |
| 418 | eval(shift(@insns)); |
| 419 | |
| 420 | &psrld (@Tx[0],31); |
| 421 | eval(shift(@insns)); |
| 422 | eval(shift(@insns)); |
| 423 | &movdqa (@Tx[1],@Tx[2]); |
| 424 | eval(shift(@insns)); |
| 425 | eval(shift(@insns)); |
| 426 | |
| 427 | &psrld (@Tx[2],30); |
| 428 | &por (@X[0],@Tx[0]); # "X[0]"<<<=1 |
| 429 | eval(shift(@insns)); |
| 430 | eval(shift(@insns)); |
| 431 | eval(shift(@insns)); |
| 432 | eval(shift(@insns)); |
| 433 | |
| 434 | &pslld (@Tx[1],2); |
| 435 | &pxor (@X[0],@Tx[2]); |
| 436 | eval(shift(@insns)); |
| 437 | eval(shift(@insns)); |
| 438 | &movdqa (@Tx[2],eval(16*(($Xi)/5))."($K_XX_XX)"); # K_XX_XX |
| 439 | eval(shift(@insns)); |
| 440 | eval(shift(@insns)); |
| 441 | |
| 442 | &pxor (@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2 |
| 443 | |
| 444 | foreach (@insns) { eval; } # remaining instructions [if any] |
| 445 | |
| 446 | $Xi++; push(@X,shift(@X)); # "rotate" X[] |
| 447 | push(@Tx,shift(@Tx)); |
| 448 | } |
| 449 | |
| 450 | sub Xupdate_ssse3_32_79() |
| 451 | { use integer; |
| 452 | my $body = shift; |
| 453 | my @insns = (&$body,&$body,&$body,&$body); # 32 to 48 instructions |
| 454 | my ($a,$b,$c,$d,$e); |
| 455 | |
| 456 | &movdqa (@Tx[0],@X[-1&7]) if ($Xi==8); |
| 457 | eval(shift(@insns)); # body_20_39 |
| 458 | &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" |
| 459 | &palignr(@Tx[0],@X[-2&7],8); # compose "X[-6]" |
| 460 | eval(shift(@insns)); |
| 461 | eval(shift(@insns)); |
| 462 | eval(shift(@insns)); # rol |
| 463 | |
| 464 | &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]" |
| 465 | eval(shift(@insns)); |
| 466 | eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/); |
| 467 | if ($Xi%5) { |
| 468 | &movdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX... |
| 469 | } else { # ... or load next one |
| 470 | &movdqa (@Tx[2],eval(16*($Xi/5))."($K_XX_XX)"); |
| 471 | } |
| 472 | &paddd (@Tx[1],@X[-1&7]); |
| 473 | eval(shift(@insns)); # ror |
| 474 | eval(shift(@insns)); |
| 475 | |
| 476 | &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-6]" |
| 477 | eval(shift(@insns)); # body_20_39 |
| 478 | eval(shift(@insns)); |
| 479 | eval(shift(@insns)); |
| 480 | eval(shift(@insns)); # rol |
| 481 | |
| 482 | &movdqa (@Tx[0],@X[0]); |
| 483 | &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU |
| 484 | eval(shift(@insns)); |
| 485 | eval(shift(@insns)); |
| 486 | eval(shift(@insns)); # ror |
| 487 | eval(shift(@insns)); |
| 488 | |
| 489 | &pslld (@X[0],2); |
| 490 | eval(shift(@insns)); # body_20_39 |
| 491 | eval(shift(@insns)); |
| 492 | &psrld (@Tx[0],30); |
| 493 | eval(shift(@insns)); |
| 494 | eval(shift(@insns)); # rol |
| 495 | eval(shift(@insns)); |
| 496 | eval(shift(@insns)); |
| 497 | eval(shift(@insns)); # ror |
| 498 | eval(shift(@insns)); |
| 499 | |
| 500 | &por (@X[0],@Tx[0]); # "X[0]"<<<=2 |
| 501 | eval(shift(@insns)); # body_20_39 |
| 502 | eval(shift(@insns)); |
| 503 | &movdqa (@Tx[1],@X[0]) if ($Xi<19); |
| 504 | eval(shift(@insns)); |
| 505 | eval(shift(@insns)); # rol |
| 506 | eval(shift(@insns)); |
| 507 | eval(shift(@insns)); |
| 508 | eval(shift(@insns)); # rol |
| 509 | eval(shift(@insns)); |
| 510 | |
| 511 | foreach (@insns) { eval; } # remaining instructions |
| 512 | |
| 513 | $Xi++; push(@X,shift(@X)); # "rotate" X[] |
| 514 | push(@Tx,shift(@Tx)); |
| 515 | } |
| 516 | |
| 517 | sub Xuplast_ssse3_80() |
| 518 | { use integer; |
| 519 | my $body = shift; |
| 520 | my @insns = (&$body,&$body,&$body,&$body); # 32 instructions |
| 521 | my ($a,$b,$c,$d,$e); |
| 522 | |
| 523 | eval(shift(@insns)); |
| 524 | &paddd (@Tx[1],@X[-1&7]); |
| 525 | eval(shift(@insns)); |
| 526 | eval(shift(@insns)); |
| 527 | eval(shift(@insns)); |
| 528 | eval(shift(@insns)); |
| 529 | |
| 530 | &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU |
| 531 | |
| 532 | foreach (@insns) { eval; } # remaining instructions |
| 533 | |
| 534 | &cmp ($inp,$num); |
| 535 | &je (".Ldone_ssse3"); |
| 536 | |
| 537 | unshift(@Tx,pop(@Tx)); |
| 538 | |
| 539 | &movdqa (@X[2],"64($K_XX_XX)"); # pbswap mask |
| 540 | &movdqa (@Tx[1],"0($K_XX_XX)"); # K_00_19 |
| 541 | &movdqu (@X[-4&7],"0($inp)"); # load input |
| 542 | &movdqu (@X[-3&7],"16($inp)"); |
| 543 | &movdqu (@X[-2&7],"32($inp)"); |
| 544 | &movdqu (@X[-1&7],"48($inp)"); |
| 545 | &pshufb (@X[-4&7],@X[2]); # byte swap |
| 546 | &add ($inp,64); |
| 547 | |
| 548 | $Xi=0; |
| 549 | } |
| 550 | |
| 551 | sub Xloop_ssse3() |
| 552 | { use integer; |
| 553 | my $body = shift; |
| 554 | my @insns = (&$body,&$body,&$body,&$body); # 32 instructions |
| 555 | my ($a,$b,$c,$d,$e); |
| 556 | |
| 557 | eval(shift(@insns)); |
| 558 | eval(shift(@insns)); |
| 559 | &pshufb (@X[($Xi-3)&7],@X[2]); |
| 560 | eval(shift(@insns)); |
| 561 | eval(shift(@insns)); |
| 562 | &paddd (@X[($Xi-4)&7],@Tx[1]); |
| 563 | eval(shift(@insns)); |
| 564 | eval(shift(@insns)); |
| 565 | eval(shift(@insns)); |
| 566 | eval(shift(@insns)); |
| 567 | &movdqa (eval(16*$Xi)."(%rsp)",@X[($Xi-4)&7]); # X[]+K xfer to IALU |
| 568 | eval(shift(@insns)); |
| 569 | eval(shift(@insns)); |
| 570 | &psubd (@X[($Xi-4)&7],@Tx[1]); |
| 571 | |
| 572 | foreach (@insns) { eval; } |
| 573 | $Xi++; |
| 574 | } |
| 575 | |
| 576 | sub Xtail_ssse3() |
| 577 | { use integer; |
| 578 | my $body = shift; |
| 579 | my @insns = (&$body,&$body,&$body,&$body); # 32 instructions |
| 580 | my ($a,$b,$c,$d,$e); |
| 581 | |
| 582 | foreach (@insns) { eval; } |
| 583 | } |
| 584 | |
| 585 | sub body_00_19 () { |
| 586 | ( |
| 587 | '($a,$b,$c,$d,$e)=@V;'. |
| 588 | '&add ($e,eval(4*($j&15))."(%rsp)");', # X[]+K xfer |
| 589 | '&xor ($c,$d);', |
| 590 | '&mov (@T[1],$a);', # $b in next round |
| 591 | '&$_rol ($a,5);', |
| 592 | '&and (@T[0],$c);', # ($b&($c^$d)) |
| 593 | '&xor ($c,$d);', # restore $c |
| 594 | '&xor (@T[0],$d);', |
| 595 | '&add ($e,$a);', |
| 596 | '&$_ror ($b,$j?7:2);', # $b>>>2 |
| 597 | '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' |
| 598 | ); |
| 599 | } |
| 600 | |
| 601 | sub body_20_39 () { |
| 602 | ( |
| 603 | '($a,$b,$c,$d,$e)=@V;'. |
| 604 | '&add ($e,eval(4*($j++&15))."(%rsp)");', # X[]+K xfer |
| 605 | '&xor (@T[0],$d);', # ($b^$d) |
| 606 | '&mov (@T[1],$a);', # $b in next round |
| 607 | '&$_rol ($a,5);', |
| 608 | '&xor (@T[0],$c);', # ($b^$d^$c) |
| 609 | '&add ($e,$a);', |
| 610 | '&$_ror ($b,7);', # $b>>>2 |
| 611 | '&add ($e,@T[0]);' .'unshift(@V,pop(@V)); unshift(@T,pop(@T));' |
| 612 | ); |
| 613 | } |
| 614 | |
| 615 | sub body_40_59 () { |
| 616 | ( |
| 617 | '($a,$b,$c,$d,$e)=@V;'. |
| 618 | '&mov (@T[1],$c);', |
| 619 | '&xor ($c,$d);', |
| 620 | '&add ($e,eval(4*($j++&15))."(%rsp)");', # X[]+K xfer |
| 621 | '&and (@T[1],$d);', |
| 622 | '&and (@T[0],$c);', # ($b&($c^$d)) |
| 623 | '&$_ror ($b,7);', # $b>>>2 |
| 624 | '&add ($e,@T[1]);', |
| 625 | '&mov (@T[1],$a);', # $b in next round |
| 626 | '&$_rol ($a,5);', |
| 627 | '&add ($e,@T[0]);', |
| 628 | '&xor ($c,$d);', # restore $c |
| 629 | '&add ($e,$a);' .'unshift(@V,pop(@V)); unshift(@T,pop(@T));' |
| 630 | ); |
| 631 | } |
| 632 | $code.=<<___; |
| 633 | .align 16 |
| 634 | .Loop_ssse3: |
| 635 | ___ |
| 636 | &Xupdate_ssse3_16_31(\&body_00_19); |
| 637 | &Xupdate_ssse3_16_31(\&body_00_19); |
| 638 | &Xupdate_ssse3_16_31(\&body_00_19); |
| 639 | &Xupdate_ssse3_16_31(\&body_00_19); |
| 640 | &Xupdate_ssse3_32_79(\&body_00_19); |
| 641 | &Xupdate_ssse3_32_79(\&body_20_39); |
| 642 | &Xupdate_ssse3_32_79(\&body_20_39); |
| 643 | &Xupdate_ssse3_32_79(\&body_20_39); |
| 644 | &Xupdate_ssse3_32_79(\&body_20_39); |
| 645 | &Xupdate_ssse3_32_79(\&body_20_39); |
| 646 | &Xupdate_ssse3_32_79(\&body_40_59); |
| 647 | &Xupdate_ssse3_32_79(\&body_40_59); |
| 648 | &Xupdate_ssse3_32_79(\&body_40_59); |
| 649 | &Xupdate_ssse3_32_79(\&body_40_59); |
| 650 | &Xupdate_ssse3_32_79(\&body_40_59); |
| 651 | &Xupdate_ssse3_32_79(\&body_20_39); |
| 652 | &Xuplast_ssse3_80(\&body_20_39); # can jump to "done" |
| 653 | |
| 654 | $saved_j=$j; @saved_V=@V; |
| 655 | |
| 656 | &Xloop_ssse3(\&body_20_39); |
| 657 | &Xloop_ssse3(\&body_20_39); |
| 658 | &Xloop_ssse3(\&body_20_39); |
| 659 | |
| 660 | $code.=<<___; |
| 661 | add 0($ctx),$A # update context |
| 662 | add 4($ctx),@T[0] |
| 663 | add 8($ctx),$C |
| 664 | add 12($ctx),$D |
| 665 | mov $A,0($ctx) |
| 666 | add 16($ctx),$E |
| 667 | mov @T[0],4($ctx) |
| 668 | mov @T[0],$B # magic seed |
| 669 | mov $C,8($ctx) |
| 670 | mov $D,12($ctx) |
| 671 | mov $E,16($ctx) |
| 672 | jmp .Loop_ssse3 |
| 673 | |
| 674 | .align 16 |
| 675 | .Ldone_ssse3: |
| 676 | ___ |
| 677 | $j=$saved_j; @V=@saved_V; |
| 678 | |
| 679 | &Xtail_ssse3(\&body_20_39); |
| 680 | &Xtail_ssse3(\&body_20_39); |
| 681 | &Xtail_ssse3(\&body_20_39); |
| 682 | |
| 683 | $code.=<<___; |
| 684 | add 0($ctx),$A # update context |
| 685 | add 4($ctx),@T[0] |
| 686 | add 8($ctx),$C |
| 687 | mov $A,0($ctx) |
| 688 | add 12($ctx),$D |
| 689 | mov @T[0],4($ctx) |
| 690 | add 16($ctx),$E |
| 691 | mov $C,8($ctx) |
| 692 | mov $D,12($ctx) |
| 693 | mov $E,16($ctx) |
| 694 | ___ |
| 695 | $code.=<<___ if ($win64); |
| 696 | movaps 64+0(%rsp),%xmm6 |
| 697 | movaps 64+16(%rsp),%xmm7 |
| 698 | movaps 64+32(%rsp),%xmm8 |
| 699 | movaps 64+48(%rsp),%xmm9 |
| 700 | movaps 64+64(%rsp),%xmm10 |
| 701 | ___ |
| 702 | $code.=<<___; |
| 703 | lea `64+($win64?5*16:0)`(%rsp),%rsi |
| 704 | mov 0(%rsi),%r12 |
| 705 | mov 8(%rsi),%rbp |
| 706 | mov 16(%rsi),%rbx |
| 707 | lea 24(%rsi),%rsp |
| 708 | .Lepilogue_ssse3: |
| 709 | ret |
| 710 | .size sha1_block_data_order_ssse3,.-sha1_block_data_order_ssse3 |
| 711 | ___ |
| 712 | |
| 713 | if ($avx) { |
| 714 | my $Xi=4; |
| 715 | my @X=map("%xmm$_",(4..7,0..3)); |
| 716 | my @Tx=map("%xmm$_",(8..10)); |
| 717 | my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization |
| 718 | my @T=("%esi","%edi"); |
| 719 | my $j=0; |
| 720 | my $K_XX_XX="%r11"; |
| 721 | |
| 722 | my $_rol=sub { &shld(@_[0],@_) }; |
| 723 | my $_ror=sub { &shrd(@_[0],@_) }; |
| 724 | |
| 725 | $code.=<<___; |
| 726 | .type sha1_block_data_order_avx,\@function,3 |
| 727 | .align 16 |
| 728 | sha1_block_data_order_avx: |
| 729 | _avx_shortcut: |
| 730 | push %rbx |
| 731 | push %rbp |
| 732 | push %r12 |
| 733 | lea `-64-($win64?5*16:0)`(%rsp),%rsp |
| 734 | ___ |
| 735 | $code.=<<___ if ($win64); |
| 736 | movaps %xmm6,64+0(%rsp) |
| 737 | movaps %xmm7,64+16(%rsp) |
| 738 | movaps %xmm8,64+32(%rsp) |
| 739 | movaps %xmm9,64+48(%rsp) |
| 740 | movaps %xmm10,64+64(%rsp) |
| 741 | .Lprologue_avx: |
| 742 | ___ |
| 743 | $code.=<<___; |
| 744 | mov %rdi,$ctx # reassigned argument |
| 745 | mov %rsi,$inp # reassigned argument |
| 746 | mov %rdx,$num # reassigned argument |
| 747 | vzeroall |
| 748 | |
| 749 | shl \$6,$num |
| 750 | add $inp,$num |
| 751 | lea K_XX_XX(%rip),$K_XX_XX |
| 752 | |
| 753 | mov 0($ctx),$A # load context |
| 754 | mov 4($ctx),$B |
| 755 | mov 8($ctx),$C |
| 756 | mov 12($ctx),$D |
| 757 | mov $B,@T[0] # magic seed |
| 758 | mov 16($ctx),$E |
| 759 | |
| 760 | vmovdqa 64($K_XX_XX),@X[2] # pbswap mask |
| 761 | vmovdqa 0($K_XX_XX),@Tx[1] # K_00_19 |
| 762 | vmovdqu 0($inp),@X[-4&7] # load input to %xmm[0-3] |
| 763 | vmovdqu 16($inp),@X[-3&7] |
| 764 | vmovdqu 32($inp),@X[-2&7] |
| 765 | vmovdqu 48($inp),@X[-1&7] |
| 766 | vpshufb @X[2],@X[-4&7],@X[-4&7] # byte swap |
| 767 | add \$64,$inp |
| 768 | vpshufb @X[2],@X[-3&7],@X[-3&7] |
| 769 | vpshufb @X[2],@X[-2&7],@X[-2&7] |
| 770 | vpshufb @X[2],@X[-1&7],@X[-1&7] |
| 771 | vpaddd @Tx[1],@X[-4&7],@X[0] # add K_00_19 |
| 772 | vpaddd @Tx[1],@X[-3&7],@X[1] |
| 773 | vpaddd @Tx[1],@X[-2&7],@X[2] |
| 774 | vmovdqa @X[0],0(%rsp) # X[]+K xfer to IALU |
| 775 | vmovdqa @X[1],16(%rsp) |
| 776 | vmovdqa @X[2],32(%rsp) |
| 777 | jmp .Loop_avx |
| 778 | ___ |
| 779 | |
| 780 | sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4 |
| 781 | { use integer; |
| 782 | my $body = shift; |
| 783 | my @insns = (&$body,&$body,&$body,&$body); # 40 instructions |
| 784 | my ($a,$b,$c,$d,$e); |
| 785 | |
| 786 | eval(shift(@insns)); |
| 787 | eval(shift(@insns)); |
| 788 | &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]" |
| 789 | eval(shift(@insns)); |
| 790 | eval(shift(@insns)); |
| 791 | |
| 792 | &vpaddd (@Tx[1],@Tx[1],@X[-1&7]); |
| 793 | eval(shift(@insns)); |
| 794 | eval(shift(@insns)); |
| 795 | &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords |
| 796 | eval(shift(@insns)); |
| 797 | eval(shift(@insns)); |
| 798 | &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]" |
| 799 | eval(shift(@insns)); |
| 800 | eval(shift(@insns)); |
| 801 | |
| 802 | &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]" |
| 803 | eval(shift(@insns)); |
| 804 | eval(shift(@insns)); |
| 805 | eval(shift(@insns)); |
| 806 | eval(shift(@insns)); |
| 807 | |
| 808 | &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]" |
| 809 | eval(shift(@insns)); |
| 810 | eval(shift(@insns)); |
| 811 | &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU |
| 812 | eval(shift(@insns)); |
| 813 | eval(shift(@insns)); |
| 814 | |
| 815 | &vpsrld (@Tx[0],@X[0],31); |
| 816 | eval(shift(@insns)); |
| 817 | eval(shift(@insns)); |
| 818 | eval(shift(@insns)); |
| 819 | eval(shift(@insns)); |
| 820 | |
| 821 | &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword |
| 822 | &vpaddd (@X[0],@X[0],@X[0]); |
| 823 | eval(shift(@insns)); |
| 824 | eval(shift(@insns)); |
| 825 | eval(shift(@insns)); |
| 826 | eval(shift(@insns)); |
| 827 | |
| 828 | &vpsrld (@Tx[1],@Tx[2],30); |
| 829 | &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1 |
| 830 | eval(shift(@insns)); |
| 831 | eval(shift(@insns)); |
| 832 | eval(shift(@insns)); |
| 833 | eval(shift(@insns)); |
| 834 | |
| 835 | &vpslld (@Tx[2],@Tx[2],2); |
| 836 | &vpxor (@X[0],@X[0],@Tx[1]); |
| 837 | eval(shift(@insns)); |
| 838 | eval(shift(@insns)); |
| 839 | eval(shift(@insns)); |
| 840 | eval(shift(@insns)); |
| 841 | |
| 842 | &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2 |
| 843 | eval(shift(@insns)); |
| 844 | eval(shift(@insns)); |
| 845 | &vmovdqa (@Tx[2],eval(16*(($Xi)/5))."($K_XX_XX)"); # K_XX_XX |
| 846 | eval(shift(@insns)); |
| 847 | eval(shift(@insns)); |
| 848 | |
| 849 | |
| 850 | foreach (@insns) { eval; } # remaining instructions [if any] |
| 851 | |
| 852 | $Xi++; push(@X,shift(@X)); # "rotate" X[] |
| 853 | push(@Tx,shift(@Tx)); |
| 854 | } |
| 855 | |
| 856 | sub Xupdate_avx_32_79() |
| 857 | { use integer; |
| 858 | my $body = shift; |
| 859 | my @insns = (&$body,&$body,&$body,&$body); # 32 to 48 instructions |
| 860 | my ($a,$b,$c,$d,$e); |
| 861 | |
| 862 | &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]" |
| 863 | &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" |
| 864 | eval(shift(@insns)); # body_20_39 |
| 865 | eval(shift(@insns)); |
| 866 | eval(shift(@insns)); |
| 867 | eval(shift(@insns)); # rol |
| 868 | |
| 869 | &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]" |
| 870 | eval(shift(@insns)); |
| 871 | eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/); |
| 872 | if ($Xi%5) { |
| 873 | &vmovdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX... |
| 874 | } else { # ... or load next one |
| 875 | &vmovdqa (@Tx[2],eval(16*($Xi/5))."($K_XX_XX)"); |
| 876 | } |
| 877 | &vpaddd (@Tx[1],@Tx[1],@X[-1&7]); |
| 878 | eval(shift(@insns)); # ror |
| 879 | eval(shift(@insns)); |
| 880 | |
| 881 | &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]" |
| 882 | eval(shift(@insns)); # body_20_39 |
| 883 | eval(shift(@insns)); |
| 884 | eval(shift(@insns)); |
| 885 | eval(shift(@insns)); # rol |
| 886 | |
| 887 | &vpsrld (@Tx[0],@X[0],30); |
| 888 | &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU |
| 889 | eval(shift(@insns)); |
| 890 | eval(shift(@insns)); |
| 891 | eval(shift(@insns)); # ror |
| 892 | eval(shift(@insns)); |
| 893 | |
| 894 | &vpslld (@X[0],@X[0],2); |
| 895 | eval(shift(@insns)); # body_20_39 |
| 896 | eval(shift(@insns)); |
| 897 | eval(shift(@insns)); |
| 898 | eval(shift(@insns)); # rol |
| 899 | eval(shift(@insns)); |
| 900 | eval(shift(@insns)); |
| 901 | eval(shift(@insns)); # ror |
| 902 | eval(shift(@insns)); |
| 903 | |
| 904 | &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2 |
| 905 | eval(shift(@insns)); # body_20_39 |
| 906 | eval(shift(@insns)); |
| 907 | &vmovdqa (@Tx[1],@X[0]) if ($Xi<19); |
| 908 | eval(shift(@insns)); |
| 909 | eval(shift(@insns)); # rol |
| 910 | eval(shift(@insns)); |
| 911 | eval(shift(@insns)); |
| 912 | eval(shift(@insns)); # rol |
| 913 | eval(shift(@insns)); |
| 914 | |
| 915 | foreach (@insns) { eval; } # remaining instructions |
| 916 | |
| 917 | $Xi++; push(@X,shift(@X)); # "rotate" X[] |
| 918 | push(@Tx,shift(@Tx)); |
| 919 | } |
| 920 | |
| 921 | sub Xuplast_avx_80() |
| 922 | { use integer; |
| 923 | my $body = shift; |
| 924 | my @insns = (&$body,&$body,&$body,&$body); # 32 instructions |
| 925 | my ($a,$b,$c,$d,$e); |
| 926 | |
| 927 | eval(shift(@insns)); |
| 928 | &vpaddd (@Tx[1],@Tx[1],@X[-1&7]); |
| 929 | eval(shift(@insns)); |
| 930 | eval(shift(@insns)); |
| 931 | eval(shift(@insns)); |
| 932 | eval(shift(@insns)); |
| 933 | |
| 934 | &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU |
| 935 | |
| 936 | foreach (@insns) { eval; } # remaining instructions |
| 937 | |
| 938 | &cmp ($inp,$num); |
| 939 | &je (".Ldone_avx"); |
| 940 | |
| 941 | unshift(@Tx,pop(@Tx)); |
| 942 | |
| 943 | &vmovdqa(@X[2],"64($K_XX_XX)"); # pbswap mask |
| 944 | &vmovdqa(@Tx[1],"0($K_XX_XX)"); # K_00_19 |
| 945 | &vmovdqu(@X[-4&7],"0($inp)"); # load input |
| 946 | &vmovdqu(@X[-3&7],"16($inp)"); |
| 947 | &vmovdqu(@X[-2&7],"32($inp)"); |
| 948 | &vmovdqu(@X[-1&7],"48($inp)"); |
| 949 | &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap |
| 950 | &add ($inp,64); |
| 951 | |
| 952 | $Xi=0; |
| 953 | } |
| 954 | |
| 955 | sub Xloop_avx() |
| 956 | { use integer; |
| 957 | my $body = shift; |
| 958 | my @insns = (&$body,&$body,&$body,&$body); # 32 instructions |
| 959 | my ($a,$b,$c,$d,$e); |
| 960 | |
| 961 | eval(shift(@insns)); |
| 962 | eval(shift(@insns)); |
| 963 | &vpshufb(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]); |
| 964 | eval(shift(@insns)); |
| 965 | eval(shift(@insns)); |
| 966 | &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],@Tx[1]); |
| 967 | eval(shift(@insns)); |
| 968 | eval(shift(@insns)); |
| 969 | eval(shift(@insns)); |
| 970 | eval(shift(@insns)); |
| 971 | &vmovdqa(eval(16*$Xi)."(%rsp)",@X[$Xi&7]); # X[]+K xfer to IALU |
| 972 | eval(shift(@insns)); |
| 973 | eval(shift(@insns)); |
| 974 | |
| 975 | foreach (@insns) { eval; } |
| 976 | $Xi++; |
| 977 | } |
| 978 | |
| 979 | sub Xtail_avx() |
| 980 | { use integer; |
| 981 | my $body = shift; |
| 982 | my @insns = (&$body,&$body,&$body,&$body); # 32 instructions |
| 983 | my ($a,$b,$c,$d,$e); |
| 984 | |
| 985 | foreach (@insns) { eval; } |
| 986 | } |
| 987 | |
| 988 | $code.=<<___; |
| 989 | .align 16 |
| 990 | .Loop_avx: |
| 991 | ___ |
| 992 | &Xupdate_avx_16_31(\&body_00_19); |
| 993 | &Xupdate_avx_16_31(\&body_00_19); |
| 994 | &Xupdate_avx_16_31(\&body_00_19); |
| 995 | &Xupdate_avx_16_31(\&body_00_19); |
| 996 | &Xupdate_avx_32_79(\&body_00_19); |
| 997 | &Xupdate_avx_32_79(\&body_20_39); |
| 998 | &Xupdate_avx_32_79(\&body_20_39); |
| 999 | &Xupdate_avx_32_79(\&body_20_39); |
| 1000 | &Xupdate_avx_32_79(\&body_20_39); |
| 1001 | &Xupdate_avx_32_79(\&body_20_39); |
| 1002 | &Xupdate_avx_32_79(\&body_40_59); |
| 1003 | &Xupdate_avx_32_79(\&body_40_59); |
| 1004 | &Xupdate_avx_32_79(\&body_40_59); |
| 1005 | &Xupdate_avx_32_79(\&body_40_59); |
| 1006 | &Xupdate_avx_32_79(\&body_40_59); |
| 1007 | &Xupdate_avx_32_79(\&body_20_39); |
| 1008 | &Xuplast_avx_80(\&body_20_39); # can jump to "done" |
| 1009 | |
| 1010 | $saved_j=$j; @saved_V=@V; |
| 1011 | |
| 1012 | &Xloop_avx(\&body_20_39); |
| 1013 | &Xloop_avx(\&body_20_39); |
| 1014 | &Xloop_avx(\&body_20_39); |
| 1015 | |
| 1016 | $code.=<<___; |
| 1017 | add 0($ctx),$A # update context |
| 1018 | add 4($ctx),@T[0] |
| 1019 | add 8($ctx),$C |
| 1020 | add 12($ctx),$D |
| 1021 | mov $A,0($ctx) |
| 1022 | add 16($ctx),$E |
| 1023 | mov @T[0],4($ctx) |
| 1024 | mov @T[0],$B # magic seed |
| 1025 | mov $C,8($ctx) |
| 1026 | mov $D,12($ctx) |
| 1027 | mov $E,16($ctx) |
| 1028 | jmp .Loop_avx |
| 1029 | |
| 1030 | .align 16 |
| 1031 | .Ldone_avx: |
| 1032 | ___ |
| 1033 | $j=$saved_j; @V=@saved_V; |
| 1034 | |
| 1035 | &Xtail_avx(\&body_20_39); |
| 1036 | &Xtail_avx(\&body_20_39); |
| 1037 | &Xtail_avx(\&body_20_39); |
| 1038 | |
| 1039 | $code.=<<___; |
| 1040 | vzeroall |
| 1041 | |
| 1042 | add 0($ctx),$A # update context |
| 1043 | add 4($ctx),@T[0] |
| 1044 | add 8($ctx),$C |
| 1045 | mov $A,0($ctx) |
| 1046 | add 12($ctx),$D |
| 1047 | mov @T[0],4($ctx) |
| 1048 | add 16($ctx),$E |
| 1049 | mov $C,8($ctx) |
| 1050 | mov $D,12($ctx) |
| 1051 | mov $E,16($ctx) |
| 1052 | ___ |
| 1053 | $code.=<<___ if ($win64); |
| 1054 | movaps 64+0(%rsp),%xmm6 |
| 1055 | movaps 64+16(%rsp),%xmm7 |
| 1056 | movaps 64+32(%rsp),%xmm8 |
| 1057 | movaps 64+48(%rsp),%xmm9 |
| 1058 | movaps 64+64(%rsp),%xmm10 |
| 1059 | ___ |
| 1060 | $code.=<<___; |
| 1061 | lea `64+($win64?5*16:0)`(%rsp),%rsi |
| 1062 | mov 0(%rsi),%r12 |
| 1063 | mov 8(%rsi),%rbp |
| 1064 | mov 16(%rsi),%rbx |
| 1065 | lea 24(%rsi),%rsp |
| 1066 | .Lepilogue_avx: |
| 1067 | ret |
| 1068 | .size sha1_block_data_order_avx,.-sha1_block_data_order_avx |
| 1069 | ___ |
| 1070 | } |
| 1071 | $code.=<<___; |
| 1072 | .align 64 |
| 1073 | K_XX_XX: |
| 1074 | .long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19 |
| 1075 | .long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39 |
| 1076 | .long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59 |
| 1077 | .long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79 |
| 1078 | .long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask |
| 1079 | ___ |
| 1080 | }}} |
| 1081 | $code.=<<___; |
| 1082 | .asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>" |
| 1083 | .align 64 |
| 1084 | ___ |
| 1085 | |
| 1086 | # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame, |
| 1087 | # CONTEXT *context,DISPATCHER_CONTEXT *disp) |
| 1088 | if ($win64) { |
| 1089 | $rec="%rcx"; |
| 1090 | $frame="%rdx"; |
| 1091 | $context="%r8"; |
| 1092 | $disp="%r9"; |
| 1093 | |
| 1094 | $code.=<<___; |
| 1095 | .extern __imp_RtlVirtualUnwind |
| 1096 | .type se_handler,\@abi-omnipotent |
| 1097 | .align 16 |
| 1098 | se_handler: |
| 1099 | push %rsi |
| 1100 | push %rdi |
| 1101 | push %rbx |
| 1102 | push %rbp |
| 1103 | push %r12 |
| 1104 | push %r13 |
| 1105 | push %r14 |
| 1106 | push %r15 |
| 1107 | pushfq |
| 1108 | sub \$64,%rsp |
| 1109 | |
| 1110 | mov 120($context),%rax # pull context->Rax |
| 1111 | mov 248($context),%rbx # pull context->Rip |
| 1112 | |
| 1113 | lea .Lprologue(%rip),%r10 |
| 1114 | cmp %r10,%rbx # context->Rip<.Lprologue |
| 1115 | jb .Lcommon_seh_tail |
| 1116 | |
| 1117 | mov 152($context),%rax # pull context->Rsp |
| 1118 | |
| 1119 | lea .Lepilogue(%rip),%r10 |
| 1120 | cmp %r10,%rbx # context->Rip>=.Lepilogue |
| 1121 | jae .Lcommon_seh_tail |
| 1122 | |
| 1123 | mov `16*4`(%rax),%rax # pull saved stack pointer |
| 1124 | lea 32(%rax),%rax |
| 1125 | |
| 1126 | mov -8(%rax),%rbx |
| 1127 | mov -16(%rax),%rbp |
| 1128 | mov -24(%rax),%r12 |
| 1129 | mov -32(%rax),%r13 |
| 1130 | mov %rbx,144($context) # restore context->Rbx |
| 1131 | mov %rbp,160($context) # restore context->Rbp |
| 1132 | mov %r12,216($context) # restore context->R12 |
| 1133 | mov %r13,224($context) # restore context->R13 |
| 1134 | |
| 1135 | jmp .Lcommon_seh_tail |
| 1136 | .size se_handler,.-se_handler |
| 1137 | |
| 1138 | .type ssse3_handler,\@abi-omnipotent |
| 1139 | .align 16 |
| 1140 | ssse3_handler: |
| 1141 | push %rsi |
| 1142 | push %rdi |
| 1143 | push %rbx |
| 1144 | push %rbp |
| 1145 | push %r12 |
| 1146 | push %r13 |
| 1147 | push %r14 |
| 1148 | push %r15 |
| 1149 | pushfq |
| 1150 | sub \$64,%rsp |
| 1151 | |
| 1152 | mov 120($context),%rax # pull context->Rax |
| 1153 | mov 248($context),%rbx # pull context->Rip |
| 1154 | |
| 1155 | mov 8($disp),%rsi # disp->ImageBase |
| 1156 | mov 56($disp),%r11 # disp->HandlerData |
| 1157 | |
| 1158 | mov 0(%r11),%r10d # HandlerData[0] |
| 1159 | lea (%rsi,%r10),%r10 # prologue label |
| 1160 | cmp %r10,%rbx # context->Rip<prologue label |
| 1161 | jb .Lcommon_seh_tail |
| 1162 | |
| 1163 | mov 152($context),%rax # pull context->Rsp |
| 1164 | |
| 1165 | mov 4(%r11),%r10d # HandlerData[1] |
| 1166 | lea (%rsi,%r10),%r10 # epilogue label |
| 1167 | cmp %r10,%rbx # context->Rip>=epilogue label |
| 1168 | jae .Lcommon_seh_tail |
| 1169 | |
| 1170 | lea 64(%rax),%rsi |
| 1171 | lea 512($context),%rdi # &context.Xmm6 |
| 1172 | mov \$10,%ecx |
| 1173 | .long 0xa548f3fc # cld; rep movsq |
| 1174 | lea `24+64+5*16`(%rax),%rax # adjust stack pointer |
| 1175 | |
| 1176 | mov -8(%rax),%rbx |
| 1177 | mov -16(%rax),%rbp |
| 1178 | mov -24(%rax),%r12 |
| 1179 | mov %rbx,144($context) # restore context->Rbx |
| 1180 | mov %rbp,160($context) # restore context->Rbp |
| 1181 | mov %r12,216($context) # restore cotnext->R12 |
| 1182 | |
| 1183 | .Lcommon_seh_tail: |
| 1184 | mov 8(%rax),%rdi |
| 1185 | mov 16(%rax),%rsi |
| 1186 | mov %rax,152($context) # restore context->Rsp |
| 1187 | mov %rsi,168($context) # restore context->Rsi |
| 1188 | mov %rdi,176($context) # restore context->Rdi |
| 1189 | |
| 1190 | mov 40($disp),%rdi # disp->ContextRecord |
| 1191 | mov $context,%rsi # context |
| 1192 | mov \$154,%ecx # sizeof(CONTEXT) |
| 1193 | .long 0xa548f3fc # cld; rep movsq |
| 1194 | |
| 1195 | mov $disp,%rsi |
| 1196 | xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER |
| 1197 | mov 8(%rsi),%rdx # arg2, disp->ImageBase |
| 1198 | mov 0(%rsi),%r8 # arg3, disp->ControlPc |
| 1199 | mov 16(%rsi),%r9 # arg4, disp->FunctionEntry |
| 1200 | mov 40(%rsi),%r10 # disp->ContextRecord |
| 1201 | lea 56(%rsi),%r11 # &disp->HandlerData |
| 1202 | lea 24(%rsi),%r12 # &disp->EstablisherFrame |
| 1203 | mov %r10,32(%rsp) # arg5 |
| 1204 | mov %r11,40(%rsp) # arg6 |
| 1205 | mov %r12,48(%rsp) # arg7 |
| 1206 | mov %rcx,56(%rsp) # arg8, (NULL) |
| 1207 | call *__imp_RtlVirtualUnwind(%rip) |
| 1208 | |
| 1209 | mov \$1,%eax # ExceptionContinueSearch |
| 1210 | add \$64,%rsp |
| 1211 | popfq |
| 1212 | pop %r15 |
| 1213 | pop %r14 |
| 1214 | pop %r13 |
| 1215 | pop %r12 |
| 1216 | pop %rbp |
| 1217 | pop %rbx |
| 1218 | pop %rdi |
| 1219 | pop %rsi |
| 1220 | ret |
| 1221 | .size ssse3_handler,.-ssse3_handler |
| 1222 | |
| 1223 | .section .pdata |
| 1224 | .align 4 |
| 1225 | .rva .LSEH_begin_sha1_block_data_order |
| 1226 | .rva .LSEH_end_sha1_block_data_order |
| 1227 | .rva .LSEH_info_sha1_block_data_order |
| 1228 | .rva .LSEH_begin_sha1_block_data_order_ssse3 |
| 1229 | .rva .LSEH_end_sha1_block_data_order_ssse3 |
| 1230 | .rva .LSEH_info_sha1_block_data_order_ssse3 |
| 1231 | ___ |
| 1232 | $code.=<<___ if ($avx); |
| 1233 | .rva .LSEH_begin_sha1_block_data_order_avx |
| 1234 | .rva .LSEH_end_sha1_block_data_order_avx |
| 1235 | .rva .LSEH_info_sha1_block_data_order_avx |
| 1236 | ___ |
| 1237 | $code.=<<___; |
| 1238 | .section .xdata |
| 1239 | .align 8 |
| 1240 | .LSEH_info_sha1_block_data_order: |
| 1241 | .byte 9,0,0,0 |
| 1242 | .rva se_handler |
| 1243 | .LSEH_info_sha1_block_data_order_ssse3: |
| 1244 | .byte 9,0,0,0 |
| 1245 | .rva ssse3_handler |
| 1246 | .rva .Lprologue_ssse3,.Lepilogue_ssse3 # HandlerData[] |
| 1247 | ___ |
| 1248 | $code.=<<___ if ($avx); |
| 1249 | .LSEH_info_sha1_block_data_order_avx: |
| 1250 | .byte 9,0,0,0 |
| 1251 | .rva ssse3_handler |
| 1252 | .rva .Lprologue_avx,.Lepilogue_avx # HandlerData[] |
| 1253 | ___ |
| 1254 | } |
| 1255 | |
| 1256 | #################################################################### |
| 1257 | |
| 1258 | $code =~ s/\`([^\`]*)\`/eval $1/gem; |
| 1259 | print $code; |
| 1260 | close STDOUT; |