crypto/sha/asm/sha256-586.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the Apache License 2.0 (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 # SHA256 block transform for x86. September 2007.
  18 #
  19 # Performance improvement over compiler generated code varies from
  20 # 10% to 40% [see below]. Not very impressive on some µ-archs, but
  21 # it's 5 times smaller and optimizes amount of writes.
  22 #
  23 # May 2012.
  24 #
  25 # Optimization including two of Pavel Semjanov's ideas, alternative
  26 # Maj and full unroll, resulted in ~20-25% improvement on most CPUs,
  27 # ~7% on Pentium, ~40% on Atom. As fully unrolled loop body is almost
  28 # 15x larger, 8KB vs. 560B, it's fired only for longer inputs. But not
  29 # on P4, where it kills performance, nor Sandy Bridge, where folded
  30 # loop is approximately as fast...
  31 #
  32 # June 2012.
  33 #
  34 # Add AMD XOP-specific code path, >30% improvement on Bulldozer over
  35 # May version, >60% over original. Add AVX+shrd code path, >25%
  36 # improvement on Sandy Bridge over May version, 60% over original.
  37 #
  38 # May 2013.
  39 #
  40 # Replace AMD XOP code path with SSSE3 to cover more processors.
  41 # (Biggest improvement coefficient is on upcoming Atom Silvermont,
  42 # not shown.) Add AVX+BMI code path.
  43 #
  44 # March 2014.
  45 #
  46 # Add support for Intel SHA Extensions.
  47 #
  48 # Performance in clock cycles per processed byte (less is better):
  49 #
  50 #               gcc     icc     x86 asm(*)      SIMD    x86_64 asm(**)
  51 # Pentium       46      57      40/38           -       -
  52 # PIII          36      33      27/24           -       -
  53 # P4            41      38      28              -       17.3
  54 # AMD K8        27      25      19/15.5         -       14.9
  55 # Core2         26      23      18/15.6         14.3    13.8
  56 # Westmere      27      -       19/15.7         13.4    12.3
  57 # Sandy Bridge  25      -       15.9            12.4    11.6
  58 # Ivy Bridge    24      -       15.0            11.4    10.3
  59 # Haswell       22      -       13.9            9.46    7.80
  60 # Skylake       20      -       14.9            9.50    7.70
  61 # Bulldozer     36      -       27/22           17.0    13.6
  62 # VIA Nano      36      -       25/22           16.8    16.5
  63 # Atom          50      -       30/25           21.9    18.9
  64 # Silvermont    40      -       34/31           22.9    20.6
  65 # Goldmont      29      -       20              16.3(***)
  66 #
  67 # (*)   numbers after slash are for unrolled loop, where applicable;
  68 # (**)  x86_64 assembly performance is presented for reference
  69 #       purposes, results are best-available;
  70 # (***) SHAEXT result is 4.1, strangely enough better than 64-bit one;
  71
  72 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  73 push(@INC,"${dir}","${dir}../../perlasm");
  74 require "x86asm.pl";
  75
  76 $output=pop and open STDOUT,">$output";
  77
  78 &asm_init($ARGV[0],$ARGV[$#ARGV] eq "386");
  79
  80 $xmm=$avx=0;
  81 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
  82
  83 if ($xmm &&     `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
  84                         =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
  85         $avx = ($1>=2.19) + ($1>=2.22);
  86 }
  87
  88 if ($xmm && !$avx && $ARGV[0] eq "win32n" &&
  89                 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
  90         $avx = ($1>=2.03) + ($1>=2.10);
  91 }
  92
  93 if ($xmm && !$avx && $ARGV[0] eq "win32" &&
  94                 `ml 2>&1` =~ /Version ([0-9]+)\./) {
  95         $avx = ($1>=10) + ($1>=11);
  96 }
  97
  98 if ($xmm && !$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|based on LLVM) ([0-9]+\.[0-9]+)/) {
  99         $avx = ($2>=3.0) + ($2>3.0);
 100 }
 101
 102 $shaext=$xmm;   ### set to zero if compiling for 1.0.1
 103
 104 $unroll_after = 64*4;   # If pre-evicted from L1P cache first spin of
 105                         # fully unrolled loop was measured to run about
 106                         # 3-4x slower. If slowdown coefficient is N and
 107                         # unrolled loop is m times faster, then you break
 108                         # even at (N-1)/(m-1) blocks. Then it needs to be
 109                         # adjusted for probability of code being evicted,
 110                         # code size/cache size=1/4. Typical m is 1.15...
 111
 112 $A="eax";
 113 $E="edx";
 114 $T="ebx";
 115 $Aoff=&DWP(4,"esp");
 116 $Boff=&DWP(8,"esp");
 117 $Coff=&DWP(12,"esp");
 118 $Doff=&DWP(16,"esp");
 119 $Eoff=&DWP(20,"esp");
 120 $Foff=&DWP(24,"esp");
 121 $Goff=&DWP(28,"esp");
 122 $Hoff=&DWP(32,"esp");
 123 $Xoff=&DWP(36,"esp");
 124 $K256="ebp";
 125
 126 sub BODY_16_63() {
 127         &mov    ($T,"ecx");                     # "ecx" is preloaded
 128          &mov   ("esi",&DWP(4*(9+15+16-14),"esp"));
 129         &ror    ("ecx",18-7);
 130          &mov   ("edi","esi");
 131         &ror    ("esi",19-17);
 132          &xor   ("ecx",$T);
 133          &shr   ($T,3);
 134         &ror    ("ecx",7);
 135          &xor   ("esi","edi");
 136          &xor   ($T,"ecx");                     # T = sigma0(X[-15])
 137         &ror    ("esi",17);
 138          &add   ($T,&DWP(4*(9+15+16),"esp"));   # T += X[-16]
 139         &shr    ("edi",10);
 140          &add   ($T,&DWP(4*(9+15+16-9),"esp")); # T += X[-7]
 141         #&xor   ("edi","esi")                   # sigma1(X[-2])
 142         # &add  ($T,"edi");                     # T += sigma1(X[-2])
 143         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
 144
 145         &BODY_00_15(1);
 146 }
 147 sub BODY_00_15() {
 148     my $in_16_63=shift;
 149
 150         &mov    ("ecx",$E);
 151          &xor   ("edi","esi")                   if ($in_16_63); # sigma1(X[-2])
 152          &mov   ("esi",$Foff);
 153         &ror    ("ecx",25-11);
 154          &add   ($T,"edi")                      if ($in_16_63); # T += sigma1(X[-2])
 155          &mov   ("edi",$Goff);
 156         &xor    ("ecx",$E);
 157          &xor   ("esi","edi");
 158          &mov   ($T,&DWP(4*(9+15),"esp"))       if (!$in_16_63);
 159          &mov   (&DWP(4*(9+15),"esp"),$T)       if ($in_16_63); # save X[0]
 160         &ror    ("ecx",11-6);
 161          &and   ("esi",$E);
 162          &mov   ($Eoff,$E);             # modulo-scheduled
 163         &xor    ($E,"ecx");
 164          &add   ($T,$Hoff);             # T += h
 165          &xor   ("esi","edi");          # Ch(e,f,g)
 166         &ror    ($E,6);                 # Sigma1(e)
 167          &mov   ("ecx",$A);
 168          &add   ($T,"esi");             # T += Ch(e,f,g)
 169
 170         &ror    ("ecx",22-13);
 171          &add   ($T,$E);                # T += Sigma1(e)
 172          &mov   ("edi",$Boff);
 173         &xor    ("ecx",$A);
 174          &mov   ($Aoff,$A);             # modulo-scheduled
 175          &lea   ("esp",&DWP(-4,"esp"));
 176         &ror    ("ecx",13-2);
 177          &mov   ("esi",&DWP(0,$K256));
 178         &xor    ("ecx",$A);
 179          &mov   ($E,$Eoff);             # e in next iteration, d in this one
 180          &xor   ($A,"edi");             # a ^= b
 181         &ror    ("ecx",2);              # Sigma0(a)
 182
 183          &add   ($T,"esi");             # T+= K[i]
 184          &mov   (&DWP(0,"esp"),$A);     # (b^c) in next round
 185         &add    ($E,$T);                # d += T
 186          &and   ($A,&DWP(4,"esp"));     # a &= (b^c)
 187         &add    ($T,"ecx");             # T += Sigma0(a)
 188          &xor   ($A,"edi");             # h = Maj(a,b,c) = Ch(a^b,c,b)
 189          &mov   ("ecx",&DWP(4*(9+15+16-1),"esp"))       if ($in_16_63); # preload T
 190         &add    ($K256,4);
 191          &add   ($A,$T);                # h += T
 192 }
 193
 194 &external_label("OPENSSL_ia32cap_P")            if (!$i386);
 195
 196 &function_begin("sha256_block_data_order");
 197         &mov    ("esi",wparam(0));      # ctx
 198         &mov    ("edi",wparam(1));      # inp
 199         &mov    ("eax",wparam(2));      # num
 200         &mov    ("ebx","esp");          # saved sp
 201
 202         &call   (&label("pic_point"));  # make it PIC!
 203 &set_label("pic_point");
 204         &blindpop($K256);
 205         &lea    ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256));
 206
 207         &sub    ("esp",16);
 208         &and    ("esp",-64);
 209
 210         &shl    ("eax",6);
 211         &add    ("eax","edi");
 212         &mov    (&DWP(0,"esp"),"esi");  # ctx
 213         &mov    (&DWP(4,"esp"),"edi");  # inp
 214         &mov    (&DWP(8,"esp"),"eax");  # inp+num*128
 215         &mov    (&DWP(12,"esp"),"ebx"); # saved sp
 216                                                 if (!$i386 && $xmm) {
 217         &picmeup("edx","OPENSSL_ia32cap_P",$K256,&label("K256"));
 218         &mov    ("ecx",&DWP(0,"edx"));
 219         &mov    ("ebx",&DWP(4,"edx"));
 220         &test   ("ecx",1<<20);          # check for P4
 221         &jnz    (&label("loop"));
 222         &mov    ("edx",&DWP(8,"edx"))   if ($xmm);
 223         &test   ("ecx",1<<24);          # check for FXSR
 224         &jz     ($unroll_after?&label("no_xmm"):&label("loop"));
 225         &and    ("ecx",1<<30);          # mask "Intel CPU" bit
 226         &and    ("ebx",1<<28|1<<9);     # mask AVX and SSSE3 bits
 227         &test   ("edx",1<<29)           if ($shaext);   # check for SHA
 228         &jnz    (&label("shaext"))      if ($shaext);
 229         &or     ("ecx","ebx");
 230         &and    ("ecx",1<<28|1<<30);
 231         &cmp    ("ecx",1<<28|1<<30);
 232                                         if ($xmm) {
 233         &je     (&label("AVX"))         if ($avx);
 234         &test   ("ebx",1<<9);           # check for SSSE3
 235         &jnz    (&label("SSSE3"));
 236                                         } else {
 237         &je     (&label("loop_shrd"));
 238                                         }
 239                                                 if ($unroll_after) {
 240 &set_label("no_xmm");
 241         &sub    ("eax","edi");
 242         &cmp    ("eax",$unroll_after);
 243         &jae    (&label("unrolled"));
 244                                                 } }
 245         &jmp    (&label("loop"));
 246
 247 sub COMPACT_LOOP() {
 248 my $suffix=shift;
 249
 250 &set_label("loop$suffix",$suffix?32:16);
 251     # copy input block to stack reversing byte and dword order
 252     for($i=0;$i<4;$i++) {
 253         &mov    ("eax",&DWP($i*16+0,"edi"));
 254         &mov    ("ebx",&DWP($i*16+4,"edi"));
 255         &mov    ("ecx",&DWP($i*16+8,"edi"));
 256         &bswap  ("eax");
 257         &mov    ("edx",&DWP($i*16+12,"edi"));
 258         &bswap  ("ebx");
 259         &push   ("eax");
 260         &bswap  ("ecx");
 261         &push   ("ebx");
 262         &bswap  ("edx");
 263         &push   ("ecx");
 264         &push   ("edx");
 265     }
 266         &add    ("edi",64);
 267         &lea    ("esp",&DWP(-4*9,"esp"));# place for A,B,C,D,E,F,G,H
 268         &mov    (&DWP(4*(9+16)+4,"esp"),"edi");
 269
 270         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 271         &mov    ($A,&DWP(0,"esi"));
 272         &mov    ("ebx",&DWP(4,"esi"));
 273         &mov    ("ecx",&DWP(8,"esi"));
 274         &mov    ("edi",&DWP(12,"esi"));
 275         # &mov  ($Aoff,$A);
 276         &mov    ($Boff,"ebx");
 277         &xor    ("ebx","ecx");
 278         &mov    ($Coff,"ecx");
 279         &mov    ($Doff,"edi");
 280         &mov    (&DWP(0,"esp"),"ebx");  # magic
 281         &mov    ($E,&DWP(16,"esi"));
 282         &mov    ("ebx",&DWP(20,"esi"));
 283         &mov    ("ecx",&DWP(24,"esi"));
 284         &mov    ("edi",&DWP(28,"esi"));
 285         # &mov  ($Eoff,$E);
 286         &mov    ($Foff,"ebx");
 287         &mov    ($Goff,"ecx");
 288         &mov    ($Hoff,"edi");
 289
 290 &set_label("00_15$suffix",16);
 291
 292         &BODY_00_15();
 293
 294         &cmp    ("esi",0xc19bf174);
 295         &jne    (&label("00_15$suffix"));
 296
 297         &mov    ("ecx",&DWP(4*(9+15+16-1),"esp"));      # preloaded in BODY_00_15(1)
 298         &jmp    (&label("16_63$suffix"));
 299
 300 &set_label("16_63$suffix",16);
 301
 302         &BODY_16_63();
 303
 304         &cmp    ("esi",0xc67178f2);
 305         &jne    (&label("16_63$suffix"));
 306
 307         &mov    ("esi",&DWP(4*(9+16+64)+0,"esp"));#ctx
 308         # &mov  ($A,$Aoff);
 309         &mov    ("ebx",$Boff);
 310         # &mov  ("edi",$Coff);
 311         &mov    ("ecx",$Doff);
 312         &add    ($A,&DWP(0,"esi"));
 313         &add    ("ebx",&DWP(4,"esi"));
 314         &add    ("edi",&DWP(8,"esi"));
 315         &add    ("ecx",&DWP(12,"esi"));
 316         &mov    (&DWP(0,"esi"),$A);
 317         &mov    (&DWP(4,"esi"),"ebx");
 318         &mov    (&DWP(8,"esi"),"edi");
 319         &mov    (&DWP(12,"esi"),"ecx");
 320         # &mov  ($E,$Eoff);
 321         &mov    ("eax",$Foff);
 322         &mov    ("ebx",$Goff);
 323         &mov    ("ecx",$Hoff);
 324         &mov    ("edi",&DWP(4*(9+16+64)+4,"esp"));#inp
 325         &add    ($E,&DWP(16,"esi"));
 326         &add    ("eax",&DWP(20,"esi"));
 327         &add    ("ebx",&DWP(24,"esi"));
 328         &add    ("ecx",&DWP(28,"esi"));
 329         &mov    (&DWP(16,"esi"),$E);
 330         &mov    (&DWP(20,"esi"),"eax");
 331         &mov    (&DWP(24,"esi"),"ebx");
 332         &mov    (&DWP(28,"esi"),"ecx");
 333
 334         &lea    ("esp",&DWP(4*(9+16+64),"esp"));# destroy frame
 335         &sub    ($K256,4*64);                   # rewind K
 336
 337         &cmp    ("edi",&DWP(8,"esp"));          # are we done yet?
 338         &jb     (&label("loop$suffix"));
 339 }
 340         &COMPACT_LOOP();
 341         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 342 &function_end_A();
 343                                                 if (!$i386 && !$xmm) {
 344         # ~20% improvement on Sandy Bridge
 345         local *ror = sub { &shrd(@_[0],@_) };
 346         &COMPACT_LOOP("_shrd");
 347         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 348 &function_end_A();
 349                                                 }
 350
 351 &set_label("K256",64);  # Yes! I keep it in the code segment!
 352 @K256=( 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5,
 353         0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5,
 354         0xd807aa98,0x12835b01,0x243185be,0x550c7dc3,
 355         0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174,
 356         0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc,
 357         0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da,
 358         0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7,
 359         0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967,
 360         0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13,
 361         0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85,
 362         0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3,
 363         0xd192e819,0xd6990624,0xf40e3585,0x106aa070,
 364         0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5,
 365         0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3,
 366         0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208,
 367         0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2     );
 368 &data_word(@K256);
 369 &data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f);        # byte swap mask
 370 &asciz("SHA256 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
 371
 372 ($a,$b,$c,$d,$e,$f,$g,$h)=(0..7);       # offsets
 373 sub off { &DWP(4*(((shift)-$i)&7),"esp"); }
 374
 375 if (!$i386 && $unroll_after) {
 376 my @AH=($A,$K256);
 377
 378 &set_label("unrolled",16);
 379         &lea    ("esp",&DWP(-96,"esp"));
 380         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 381         &mov    ($AH[0],&DWP(0,"esi"));
 382         &mov    ($AH[1],&DWP(4,"esi"));
 383         &mov    ("ecx",&DWP(8,"esi"));
 384         &mov    ("ebx",&DWP(12,"esi"));
 385         #&mov   (&DWP(0,"esp"),$AH[0]);
 386         &mov    (&DWP(4,"esp"),$AH[1]);
 387         &xor    ($AH[1],"ecx");         # magic
 388         &mov    (&DWP(8,"esp"),"ecx");
 389         &mov    (&DWP(12,"esp"),"ebx");
 390         &mov    ($E,&DWP(16,"esi"));
 391         &mov    ("ebx",&DWP(20,"esi"));
 392         &mov    ("ecx",&DWP(24,"esi"));
 393         &mov    ("esi",&DWP(28,"esi"));
 394         #&mov   (&DWP(16,"esp"),$E);
 395         &mov    (&DWP(20,"esp"),"ebx");
 396         &mov    (&DWP(24,"esp"),"ecx");
 397         &mov    (&DWP(28,"esp"),"esi");
 398         &jmp    (&label("grand_loop"));
 399
 400 &set_label("grand_loop",16);
 401     # copy input block to stack reversing byte order
 402     for($i=0;$i<5;$i++) {
 403         &mov    ("ebx",&DWP(12*$i+0,"edi"));
 404         &mov    ("ecx",&DWP(12*$i+4,"edi"));
 405         &bswap  ("ebx");
 406         &mov    ("esi",&DWP(12*$i+8,"edi"));
 407         &bswap  ("ecx");
 408         &mov    (&DWP(32+12*$i+0,"esp"),"ebx");
 409         &bswap  ("esi");
 410         &mov    (&DWP(32+12*$i+4,"esp"),"ecx");
 411         &mov    (&DWP(32+12*$i+8,"esp"),"esi");
 412     }
 413         &mov    ("ebx",&DWP($i*12,"edi"));
 414         &add    ("edi",64);
 415         &bswap  ("ebx");
 416         &mov    (&DWP(96+4,"esp"),"edi");
 417         &mov    (&DWP(32+12*$i,"esp"),"ebx");
 418
 419     my ($t1,$t2) = ("ecx","esi");
 420
 421     for ($i=0;$i<64;$i++) {
 422
 423       if ($i>=16) {
 424         &mov    ($T,$t1);                       # $t1 is preloaded
 425         # &mov  ($t2,&DWP(32+4*(($i+14)&15),"esp"));
 426         &ror    ($t1,18-7);
 427          &mov   ("edi",$t2);
 428         &ror    ($t2,19-17);
 429          &xor   ($t1,$T);
 430          &shr   ($T,3);
 431         &ror    ($t1,7);
 432          &xor   ($t2,"edi");
 433          &xor   ($T,$t1);                       # T = sigma0(X[-15])
 434         &ror    ($t2,17);
 435          &add   ($T,&DWP(32+4*($i&15),"esp"));  # T += X[-16]
 436         &shr    ("edi",10);
 437          &add   ($T,&DWP(32+4*(($i+9)&15),"esp"));      # T += X[-7]
 438         #&xor   ("edi",$t2)                     # sigma1(X[-2])
 439         # &add  ($T,"edi");                     # T += sigma1(X[-2])
 440         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
 441       }
 442         &mov    ($t1,$E);
 443          &xor   ("edi",$t2)                     if ($i>=16);    # sigma1(X[-2])
 444          &mov   ($t2,&off($f));
 445         &ror    ($E,25-11);
 446          &add   ($T,"edi")                      if ($i>=16);    # T += sigma1(X[-2])
 447          &mov   ("edi",&off($g));
 448         &xor    ($E,$t1);
 449          &mov   ($T,&DWP(32+4*($i&15),"esp"))   if ($i<16);     # X[i]
 450          &mov   (&DWP(32+4*($i&15),"esp"),$T)   if ($i>=16 && $i<62);   # save X[0]
 451          &xor   ($t2,"edi");
 452         &ror    ($E,11-6);
 453          &and   ($t2,$t1);
 454          &mov   (&off($e),$t1);         # save $E, modulo-scheduled
 455         &xor    ($E,$t1);
 456          &add   ($T,&off($h));          # T += h
 457          &xor   ("edi",$t2);            # Ch(e,f,g)
 458         &ror    ($E,6);                 # Sigma1(e)
 459          &mov   ($t1,$AH[0]);
 460          &add   ($T,"edi");             # T += Ch(e,f,g)
 461
 462         &ror    ($t1,22-13);
 463          &mov   ($t2,$AH[0]);
 464          &mov   ("edi",&off($b));
 465         &xor    ($t1,$AH[0]);
 466          &mov   (&off($a),$AH[0]);      # save $A, modulo-scheduled
 467          &xor   ($AH[0],"edi");         # a ^= b, (b^c) in next round
 468         &ror    ($t1,13-2);
 469          &and   ($AH[1],$AH[0]);        # (b^c) &= (a^b)
 470          &lea   ($E,&DWP(@K256[$i],$T,$E));     # T += Sigma1(1)+K[i]
 471         &xor    ($t1,$t2);
 472          &xor   ($AH[1],"edi");         # h = Maj(a,b,c) = Ch(a^b,c,b)
 473          &mov   ($t2,&DWP(32+4*(($i+2)&15),"esp"))      if ($i>=15 && $i<63);
 474         &ror    ($t1,2);                # Sigma0(a)
 475
 476          &add   ($AH[1],$E);            # h += T
 477          &add   ($E,&off($d));          # d += T
 478         &add    ($AH[1],$t1);           # h += Sigma0(a)
 479          &mov   ($t1,&DWP(32+4*(($i+15)&15),"esp"))     if ($i>=15 && $i<63);
 480
 481         @AH = reverse(@AH);             # rotate(a,h)
 482         ($t1,$t2) = ($t2,$t1);          # rotate(t1,t2)
 483     }
 484         &mov    ("esi",&DWP(96,"esp")); #ctx
 485                                         #&mov   ($AH[0],&DWP(0,"esp"));
 486         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
 487                                         #&mov   ("edi", &DWP(8,"esp"));
 488         &mov    ("ecx",&DWP(12,"esp"));
 489         &add    ($AH[0],&DWP(0,"esi"));
 490         &add    ($AH[1],&DWP(4,"esi"));
 491         &add    ("edi",&DWP(8,"esi"));
 492         &add    ("ecx",&DWP(12,"esi"));
 493         &mov    (&DWP(0,"esi"),$AH[0]);
 494         &mov    (&DWP(4,"esi"),$AH[1]);
 495         &mov    (&DWP(8,"esi"),"edi");
 496         &mov    (&DWP(12,"esi"),"ecx");
 497          #&mov  (&DWP(0,"esp"),$AH[0]);
 498          &mov   (&DWP(4,"esp"),$AH[1]);
 499          &xor   ($AH[1],"edi");         # magic
 500          &mov   (&DWP(8,"esp"),"edi");
 501          &mov   (&DWP(12,"esp"),"ecx");
 502         #&mov   ($E,&DWP(16,"esp"));
 503         &mov    ("edi",&DWP(20,"esp"));
 504         &mov    ("ebx",&DWP(24,"esp"));
 505         &mov    ("ecx",&DWP(28,"esp"));
 506         &add    ($E,&DWP(16,"esi"));
 507         &add    ("edi",&DWP(20,"esi"));
 508         &add    ("ebx",&DWP(24,"esi"));
 509         &add    ("ecx",&DWP(28,"esi"));
 510         &mov    (&DWP(16,"esi"),$E);
 511         &mov    (&DWP(20,"esi"),"edi");
 512         &mov    (&DWP(24,"esi"),"ebx");
 513         &mov    (&DWP(28,"esi"),"ecx");
 514          #&mov  (&DWP(16,"esp"),$E);
 515          &mov   (&DWP(20,"esp"),"edi");
 516         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 517          &mov   (&DWP(24,"esp"),"ebx");
 518          &mov   (&DWP(28,"esp"),"ecx");
 519
 520         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
 521         &jb     (&label("grand_loop"));
 522
 523         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
 524 &function_end_A();
 525 }
 526                                                 if (!$i386 && $xmm) {{{
 527 if ($shaext) {
 528 ######################################################################
 529 # Intel SHA Extensions implementation of SHA256 update function.
 530 #
 531 my ($ctx,$inp,$end)=("esi","edi","eax");
 532 my ($Wi,$ABEF,$CDGH,$TMP)=map("xmm$_",(0..2,7));
 533 my @MSG=map("xmm$_",(3..6));
 534
 535 sub sha256op38 {
 536  my ($opcodelet,$dst,$src)=@_;
 537     if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
 538     {   &data_byte(0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2);       }
 539 }
 540 sub sha256rnds2 { sha256op38(0xcb,@_); }
 541 sub sha256msg1  { sha256op38(0xcc,@_); }
 542 sub sha256msg2  { sha256op38(0xcd,@_); }
 543
 544 &set_label("shaext",32);
 545         &sub            ("esp",32);
 546
 547         &movdqu         ($ABEF,&QWP(0,$ctx));           # DCBA
 548         &lea            ($K256,&DWP(0x80,$K256));
 549         &movdqu         ($CDGH,&QWP(16,$ctx));          # HGFE
 550         &movdqa         ($TMP,&QWP(0x100-0x80,$K256));  # byte swap mask
 551
 552         &pshufd         ($Wi,$ABEF,0x1b);               # ABCD
 553         &pshufd         ($ABEF,$ABEF,0xb1);             # CDAB
 554         &pshufd         ($CDGH,$CDGH,0x1b);             # EFGH
 555         &palignr        ($ABEF,$CDGH,8);                # ABEF
 556         &punpcklqdq     ($CDGH,$Wi);                    # CDGH
 557         &jmp            (&label("loop_shaext"));
 558
 559 &set_label("loop_shaext",16);
 560         &movdqu         (@MSG[0],&QWP(0,$inp));
 561         &movdqu         (@MSG[1],&QWP(0x10,$inp));
 562         &movdqu         (@MSG[2],&QWP(0x20,$inp));
 563         &pshufb         (@MSG[0],$TMP);
 564         &movdqu         (@MSG[3],&QWP(0x30,$inp));
 565         &movdqa         (&QWP(16,"esp"),$CDGH);         # offload
 566
 567         &movdqa         ($Wi,&QWP(0*16-0x80,$K256));
 568         &paddd          ($Wi,@MSG[0]);
 569         &pshufb         (@MSG[1],$TMP);
 570         &sha256rnds2    ($CDGH,$ABEF);                  # 0-3
 571         &pshufd         ($Wi,$Wi,0x0e);
 572         &nop            ();
 573         &movdqa         (&QWP(0,"esp"),$ABEF);          # offload
 574         &sha256rnds2    ($ABEF,$CDGH);
 575
 576         &movdqa         ($Wi,&QWP(1*16-0x80,$K256));
 577         &paddd          ($Wi,@MSG[1]);
 578         &pshufb         (@MSG[2],$TMP);
 579         &sha256rnds2    ($CDGH,$ABEF);                  # 4-7
 580         &pshufd         ($Wi,$Wi,0x0e);
 581         &lea            ($inp,&DWP(0x40,$inp));
 582         &sha256msg1     (@MSG[0],@MSG[1]);
 583         &sha256rnds2    ($ABEF,$CDGH);
 584
 585         &movdqa         ($Wi,&QWP(2*16-0x80,$K256));
 586         &paddd          ($Wi,@MSG[2]);
 587         &pshufb         (@MSG[3],$TMP);
 588         &sha256rnds2    ($CDGH,$ABEF);                  # 8-11
 589         &pshufd         ($Wi,$Wi,0x0e);
 590         &movdqa         ($TMP,@MSG[3]);
 591         &palignr        ($TMP,@MSG[2],4);
 592         &nop            ();
 593         &paddd          (@MSG[0],$TMP);
 594         &sha256msg1     (@MSG[1],@MSG[2]);
 595         &sha256rnds2    ($ABEF,$CDGH);
 596
 597         &movdqa         ($Wi,&QWP(3*16-0x80,$K256));
 598         &paddd          ($Wi,@MSG[3]);
 599         &sha256msg2     (@MSG[0],@MSG[3]);
 600         &sha256rnds2    ($CDGH,$ABEF);                  # 12-15
 601         &pshufd         ($Wi,$Wi,0x0e);
 602         &movdqa         ($TMP,@MSG[0]);
 603         &palignr        ($TMP,@MSG[3],4);
 604         &nop            ();
 605         &paddd          (@MSG[1],$TMP);
 606         &sha256msg1     (@MSG[2],@MSG[3]);
 607         &sha256rnds2    ($ABEF,$CDGH);
 608
 609 for($i=4;$i<16-3;$i++) {
 610         &movdqa         ($Wi,&QWP($i*16-0x80,$K256));
 611         &paddd          ($Wi,@MSG[0]);
 612         &sha256msg2     (@MSG[1],@MSG[0]);
 613         &sha256rnds2    ($CDGH,$ABEF);                  # 16-19...
 614         &pshufd         ($Wi,$Wi,0x0e);
 615         &movdqa         ($TMP,@MSG[1]);
 616         &palignr        ($TMP,@MSG[0],4);
 617         &nop            ();
 618         &paddd          (@MSG[2],$TMP);
 619         &sha256msg1     (@MSG[3],@MSG[0]);
 620         &sha256rnds2    ($ABEF,$CDGH);
 621
 622         push(@MSG,shift(@MSG));
 623 }
 624         &movdqa         ($Wi,&QWP(13*16-0x80,$K256));
 625         &paddd          ($Wi,@MSG[0]);
 626         &sha256msg2     (@MSG[1],@MSG[0]);
 627         &sha256rnds2    ($CDGH,$ABEF);                  # 52-55
 628         &pshufd         ($Wi,$Wi,0x0e);
 629         &movdqa         ($TMP,@MSG[1])
 630         &palignr        ($TMP,@MSG[0],4);
 631         &sha256rnds2    ($ABEF,$CDGH);
 632         &paddd          (@MSG[2],$TMP);
 633
 634         &movdqa         ($Wi,&QWP(14*16-0x80,$K256));
 635         &paddd          ($Wi,@MSG[1]);
 636         &sha256rnds2    ($CDGH,$ABEF);                  # 56-59
 637         &pshufd         ($Wi,$Wi,0x0e);
 638         &sha256msg2     (@MSG[2],@MSG[1]);
 639         &movdqa         ($TMP,&QWP(0x100-0x80,$K256));  # byte swap mask
 640         &sha256rnds2    ($ABEF,$CDGH);
 641
 642         &movdqa         ($Wi,&QWP(15*16-0x80,$K256));
 643         &paddd          ($Wi,@MSG[2]);
 644         &nop            ();
 645         &sha256rnds2    ($CDGH,$ABEF);                  # 60-63
 646         &pshufd         ($Wi,$Wi,0x0e);
 647         &cmp            ($end,$inp);
 648         &nop            ();
 649         &sha256rnds2    ($ABEF,$CDGH);
 650
 651         &paddd          ($CDGH,&QWP(16,"esp"));
 652         &paddd          ($ABEF,&QWP(0,"esp"));
 653         &jnz            (&label("loop_shaext"));
 654
 655         &pshufd         ($CDGH,$CDGH,0xb1);             # DCHG
 656         &pshufd         ($TMP,$ABEF,0x1b);              # FEBA
 657         &pshufd         ($ABEF,$ABEF,0xb1);             # BAFE
 658         &punpckhqdq     ($ABEF,$CDGH);                  # DCBA
 659         &palignr        ($CDGH,$TMP,8);                 # HGFE
 660
 661         &mov            ("esp",&DWP(32+12,"esp"));
 662         &movdqu         (&QWP(0,$ctx),$ABEF);
 663         &movdqu         (&QWP(16,$ctx),$CDGH);
 664 &function_end_A();
 665 }
 666
 667 my @X = map("xmm$_",(0..3));
 668 my ($t0,$t1,$t2,$t3) = map("xmm$_",(4..7));
 669 my @AH = ($A,$T);
 670
 671 &set_label("SSSE3",32);
 672         &lea    ("esp",&DWP(-96,"esp"));
 673         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 674         &mov    ($AH[0],&DWP(0,"esi"));
 675         &mov    ($AH[1],&DWP(4,"esi"));
 676         &mov    ("ecx",&DWP(8,"esi"));
 677         &mov    ("edi",&DWP(12,"esi"));
 678         #&mov   (&DWP(0,"esp"),$AH[0]);
 679         &mov    (&DWP(4,"esp"),$AH[1]);
 680         &xor    ($AH[1],"ecx");                 # magic
 681         &mov    (&DWP(8,"esp"),"ecx");
 682         &mov    (&DWP(12,"esp"),"edi");
 683         &mov    ($E,&DWP(16,"esi"));
 684         &mov    ("edi",&DWP(20,"esi"));
 685         &mov    ("ecx",&DWP(24,"esi"));
 686         &mov    ("esi",&DWP(28,"esi"));
 687         #&mov   (&DWP(16,"esp"),$E);
 688         &mov    (&DWP(20,"esp"),"edi");
 689         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 690         &mov    (&DWP(24,"esp"),"ecx");
 691         &mov    (&DWP(28,"esp"),"esi");
 692         &movdqa ($t3,&QWP(256,$K256));
 693         &jmp    (&label("grand_ssse3"));
 694
 695 &set_label("grand_ssse3",16);
 696         # load input, reverse byte order, add K256[0..15], save to stack
 697         &movdqu (@X[0],&QWP(0,"edi"));
 698         &movdqu (@X[1],&QWP(16,"edi"));
 699         &movdqu (@X[2],&QWP(32,"edi"));
 700         &movdqu (@X[3],&QWP(48,"edi"));
 701         &add    ("edi",64);
 702         &pshufb (@X[0],$t3);
 703         &mov    (&DWP(96+4,"esp"),"edi");
 704         &pshufb (@X[1],$t3);
 705         &movdqa ($t0,&QWP(0,$K256));
 706         &pshufb (@X[2],$t3);
 707         &movdqa ($t1,&QWP(16,$K256));
 708         &paddd  ($t0,@X[0]);
 709         &pshufb (@X[3],$t3);
 710         &movdqa ($t2,&QWP(32,$K256));
 711         &paddd  ($t1,@X[1]);
 712         &movdqa ($t3,&QWP(48,$K256));
 713         &movdqa (&QWP(32+0,"esp"),$t0);
 714         &paddd  ($t2,@X[2]);
 715         &movdqa (&QWP(32+16,"esp"),$t1);
 716         &paddd  ($t3,@X[3]);
 717         &movdqa (&QWP(32+32,"esp"),$t2);
 718         &movdqa (&QWP(32+48,"esp"),$t3);
 719         &jmp    (&label("ssse3_00_47"));
 720
 721 &set_label("ssse3_00_47",16);
 722         &add            ($K256,64);
 723
 724 sub SSSE3_00_47 () {
 725 my $j = shift;
 726 my $body = shift;
 727 my @X = @_;
 728 my @insns = (&$body,&$body,&$body,&$body);      # 120 instructions
 729
 730           eval(shift(@insns));
 731         &movdqa         ($t0,@X[1]);
 732           eval(shift(@insns));                  # @
 733           eval(shift(@insns));
 734         &movdqa         ($t3,@X[3]);
 735           eval(shift(@insns));
 736           eval(shift(@insns));
 737         &palignr        ($t0,@X[0],4);          # X[1..4]
 738           eval(shift(@insns));
 739           eval(shift(@insns));                  # @
 740           eval(shift(@insns));
 741          &palignr       ($t3,@X[2],4);          # X[9..12]
 742           eval(shift(@insns));
 743           eval(shift(@insns));
 744           eval(shift(@insns));
 745         &movdqa         ($t1,$t0);
 746           eval(shift(@insns));                  # @
 747           eval(shift(@insns));
 748         &movdqa         ($t2,$t0);
 749           eval(shift(@insns));
 750           eval(shift(@insns));
 751         &psrld          ($t0,3);
 752           eval(shift(@insns));
 753           eval(shift(@insns));                  # @
 754          &paddd         (@X[0],$t3);            # X[0..3] += X[9..12]
 755           eval(shift(@insns));
 756           eval(shift(@insns));
 757         &psrld          ($t2,7);
 758           eval(shift(@insns));
 759           eval(shift(@insns));
 760           eval(shift(@insns));                  # @
 761           eval(shift(@insns));
 762          &pshufd        ($t3,@X[3],0b11111010); # X[14..15]
 763           eval(shift(@insns));
 764           eval(shift(@insns));
 765         &pslld          ($t1,32-18);
 766           eval(shift(@insns));
 767           eval(shift(@insns));                  # @
 768         &pxor           ($t0,$t2);
 769           eval(shift(@insns));
 770           eval(shift(@insns));
 771         &psrld          ($t2,18-7);
 772           eval(shift(@insns));
 773           eval(shift(@insns));
 774           eval(shift(@insns));                  # @
 775         &pxor           ($t0,$t1);
 776           eval(shift(@insns));
 777           eval(shift(@insns));
 778         &pslld          ($t1,18-7);
 779           eval(shift(@insns));
 780           eval(shift(@insns));
 781           eval(shift(@insns));                  # @
 782         &pxor           ($t0,$t2);
 783           eval(shift(@insns));
 784           eval(shift(@insns));
 785          &movdqa        ($t2,$t3);
 786           eval(shift(@insns));
 787           eval(shift(@insns));
 788           eval(shift(@insns));                  # @
 789         &pxor           ($t0,$t1);              # sigma0(X[1..4])
 790           eval(shift(@insns));
 791           eval(shift(@insns));
 792          &psrld         ($t3,10);
 793           eval(shift(@insns));
 794           eval(shift(@insns));
 795           eval(shift(@insns));                  # @
 796         &paddd          (@X[0],$t0);            # X[0..3] += sigma0(X[1..4])
 797           eval(shift(@insns));
 798           eval(shift(@insns));
 799          &psrlq         ($t2,17);
 800           eval(shift(@insns));
 801           eval(shift(@insns));
 802           eval(shift(@insns));                  # @
 803          &pxor          ($t3,$t2);
 804           eval(shift(@insns));
 805           eval(shift(@insns));
 806          &psrlq         ($t2,19-17);
 807           eval(shift(@insns));
 808           eval(shift(@insns));
 809           eval(shift(@insns));                  # @
 810          &pxor          ($t3,$t2);
 811           eval(shift(@insns));
 812           eval(shift(@insns));
 813          &pshufd        ($t3,$t3,0b10000000);
 814           eval(shift(@insns));
 815           eval(shift(@insns));
 816           eval(shift(@insns));                  # @
 817           eval(shift(@insns));
 818           eval(shift(@insns));
 819           eval(shift(@insns));
 820           eval(shift(@insns));
 821           eval(shift(@insns));                  # @
 822           eval(shift(@insns));
 823          &psrldq        ($t3,8);
 824           eval(shift(@insns));
 825           eval(shift(@insns));
 826           eval(shift(@insns));
 827         &paddd          (@X[0],$t3);            # X[0..1] += sigma1(X[14..15])
 828           eval(shift(@insns));                  # @
 829           eval(shift(@insns));
 830           eval(shift(@insns));
 831           eval(shift(@insns));
 832           eval(shift(@insns));
 833           eval(shift(@insns));                  # @
 834           eval(shift(@insns));
 835          &pshufd        ($t3,@X[0],0b01010000); # X[16..17]
 836           eval(shift(@insns));
 837           eval(shift(@insns));
 838           eval(shift(@insns));
 839          &movdqa        ($t2,$t3);
 840           eval(shift(@insns));                  # @
 841          &psrld         ($t3,10);
 842           eval(shift(@insns));
 843          &psrlq         ($t2,17);
 844           eval(shift(@insns));
 845           eval(shift(@insns));
 846           eval(shift(@insns));
 847           eval(shift(@insns));                  # @
 848          &pxor          ($t3,$t2);
 849           eval(shift(@insns));
 850           eval(shift(@insns));
 851          &psrlq         ($t2,19-17);
 852           eval(shift(@insns));
 853           eval(shift(@insns));
 854           eval(shift(@insns));                  # @
 855          &pxor          ($t3,$t2);
 856           eval(shift(@insns));
 857           eval(shift(@insns));
 858           eval(shift(@insns));
 859          &pshufd        ($t3,$t3,0b00001000);
 860           eval(shift(@insns));
 861           eval(shift(@insns));                  # @
 862         &movdqa         ($t2,&QWP(16*$j,$K256));
 863           eval(shift(@insns));
 864           eval(shift(@insns));
 865          &pslldq        ($t3,8);
 866           eval(shift(@insns));
 867           eval(shift(@insns));
 868           eval(shift(@insns));                  # @
 869           eval(shift(@insns));
 870           eval(shift(@insns));
 871           eval(shift(@insns));
 872           eval(shift(@insns));
 873           eval(shift(@insns));                  # @
 874         &paddd          (@X[0],$t3);            # X[2..3] += sigma1(X[16..17])
 875           eval(shift(@insns));
 876           eval(shift(@insns));
 877           eval(shift(@insns));
 878           eval(shift(@insns));
 879         &paddd          ($t2,@X[0]);
 880           eval(shift(@insns));                  # @
 881
 882         foreach (@insns) { eval; }              # remaining instructions
 883
 884         &movdqa         (&QWP(32+16*$j,"esp"),$t2);
 885 }
 886
 887 sub body_00_15 () {
 888         (
 889         '&mov   ("ecx",$E);',
 890         '&ror   ($E,25-11);',
 891          '&mov  ("esi",&off($f));',
 892         '&xor   ($E,"ecx");',
 893          '&mov  ("edi",&off($g));',
 894          '&xor  ("esi","edi");',
 895         '&ror   ($E,11-6);',
 896          '&and  ("esi","ecx");',
 897          '&mov  (&off($e),"ecx");',     # save $E, modulo-scheduled
 898         '&xor   ($E,"ecx");',
 899          '&xor  ("edi","esi");',        # Ch(e,f,g)
 900         '&ror   ($E,6);',               # T = Sigma1(e)
 901          '&mov  ("ecx",$AH[0]);',
 902          '&add  ($E,"edi");',           # T += Ch(e,f,g)
 903          '&mov  ("edi",&off($b));',
 904         '&mov   ("esi",$AH[0]);',
 905
 906         '&ror   ("ecx",22-13);',
 907          '&mov  (&off($a),$AH[0]);',    # save $A, modulo-scheduled
 908         '&xor   ("ecx",$AH[0]);',
 909          '&xor  ($AH[0],"edi");',       # a ^= b, (b^c) in next round
 910          '&add  ($E,&off($h));',        # T += h
 911         '&ror   ("ecx",13-2);',
 912          '&and  ($AH[1],$AH[0]);',      # (b^c) &= (a^b)
 913         '&xor   ("ecx","esi");',
 914          '&add  ($E,&DWP(32+4*($i&15),"esp"));',        # T += K[i]+X[i]
 915          '&xor  ($AH[1],"edi");',       # h = Maj(a,b,c) = Ch(a^b,c,b)
 916         '&ror   ("ecx",2);',            # Sigma0(a)
 917
 918          '&add  ($AH[1],$E);',          # h += T
 919          '&add  ($E,&off($d));',        # d += T
 920         '&add   ($AH[1],"ecx");'.       # h += Sigma0(a)
 921
 922         '@AH = reverse(@AH); $i++;'     # rotate(a,h)
 923         );
 924 }
 925
 926     for ($i=0,$j=0; $j<4; $j++) {
 927         &SSSE3_00_47($j,\&body_00_15,@X);
 928         push(@X,shift(@X));             # rotate(@X)
 929     }
 930         &cmp    (&DWP(16*$j,$K256),0x00010203);
 931         &jne    (&label("ssse3_00_47"));
 932
 933     for ($i=0; $i<16; ) {
 934         foreach(body_00_15()) { eval; }
 935     }
 936
 937         &mov    ("esi",&DWP(96,"esp")); #ctx
 938                                         #&mov   ($AH[0],&DWP(0,"esp"));
 939         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
 940                                         #&mov   ("edi", &DWP(8,"esp"));
 941         &mov    ("ecx",&DWP(12,"esp"));
 942         &add    ($AH[0],&DWP(0,"esi"));
 943         &add    ($AH[1],&DWP(4,"esi"));
 944         &add    ("edi",&DWP(8,"esi"));
 945         &add    ("ecx",&DWP(12,"esi"));
 946         &mov    (&DWP(0,"esi"),$AH[0]);
 947         &mov    (&DWP(4,"esi"),$AH[1]);
 948         &mov    (&DWP(8,"esi"),"edi");
 949         &mov    (&DWP(12,"esi"),"ecx");
 950          #&mov  (&DWP(0,"esp"),$AH[0]);
 951          &mov   (&DWP(4,"esp"),$AH[1]);
 952          &xor   ($AH[1],"edi");                 # magic
 953          &mov   (&DWP(8,"esp"),"edi");
 954          &mov   (&DWP(12,"esp"),"ecx");
 955         #&mov   ($E,&DWP(16,"esp"));
 956         &mov    ("edi",&DWP(20,"esp"));
 957         &mov    ("ecx",&DWP(24,"esp"));
 958         &add    ($E,&DWP(16,"esi"));
 959         &add    ("edi",&DWP(20,"esi"));
 960         &add    ("ecx",&DWP(24,"esi"));
 961         &mov    (&DWP(16,"esi"),$E);
 962         &mov    (&DWP(20,"esi"),"edi");
 963          &mov   (&DWP(20,"esp"),"edi");
 964         &mov    ("edi",&DWP(28,"esp"));
 965         &mov    (&DWP(24,"esi"),"ecx");
 966          #&mov  (&DWP(16,"esp"),$E);
 967         &add    ("edi",&DWP(28,"esi"));
 968          &mov   (&DWP(24,"esp"),"ecx");
 969         &mov    (&DWP(28,"esi"),"edi");
 970          &mov   (&DWP(28,"esp"),"edi");
 971         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 972
 973         &movdqa ($t3,&QWP(64,$K256));
 974         &sub    ($K256,3*64);                   # rewind K
 975         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
 976         &jb     (&label("grand_ssse3"));
 977
 978         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
 979 &function_end_A();
 980                                                 if ($avx) {
 981 &set_label("AVX",32);
 982                                                 if ($avx>1) {
 983         &and    ("edx",1<<8|1<<3);              # check for BMI2+BMI1
 984         &cmp    ("edx",1<<8|1<<3);
 985         &je     (&label("AVX_BMI"));
 986                                                 }
 987         &lea    ("esp",&DWP(-96,"esp"));
 988         &vzeroall       ();
 989         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 990         &mov    ($AH[0],&DWP(0,"esi"));
 991         &mov    ($AH[1],&DWP(4,"esi"));
 992         &mov    ("ecx",&DWP(8,"esi"));
 993         &mov    ("edi",&DWP(12,"esi"));
 994         #&mov   (&DWP(0,"esp"),$AH[0]);
 995         &mov    (&DWP(4,"esp"),$AH[1]);
 996         &xor    ($AH[1],"ecx");                 # magic
 997         &mov    (&DWP(8,"esp"),"ecx");
 998         &mov    (&DWP(12,"esp"),"edi");
 999         &mov    ($E,&DWP(16,"esi"));
1000         &mov    ("edi",&DWP(20,"esi"));
1001         &mov    ("ecx",&DWP(24,"esi"));
1002         &mov    ("esi",&DWP(28,"esi"));
1003         #&mov   (&DWP(16,"esp"),$E);
1004         &mov    (&DWP(20,"esp"),"edi");
1005         &mov    ("edi",&DWP(96+4,"esp"));       # inp
1006         &mov    (&DWP(24,"esp"),"ecx");
1007         &mov    (&DWP(28,"esp"),"esi");
1008         &vmovdqa        ($t3,&QWP(256,$K256));
1009         &jmp    (&label("grand_avx"));
1010
1011 &set_label("grand_avx",32);
1012         # load input, reverse byte order, add K256[0..15], save to stack
1013         &vmovdqu        (@X[0],&QWP(0,"edi"));
1014         &vmovdqu        (@X[1],&QWP(16,"edi"));
1015         &vmovdqu        (@X[2],&QWP(32,"edi"));
1016         &vmovdqu        (@X[3],&QWP(48,"edi"));
1017         &add            ("edi",64);
1018         &vpshufb        (@X[0],@X[0],$t3);
1019         &mov            (&DWP(96+4,"esp"),"edi");
1020         &vpshufb        (@X[1],@X[1],$t3);
1021         &vpshufb        (@X[2],@X[2],$t3);
1022         &vpaddd         ($t0,@X[0],&QWP(0,$K256));
1023         &vpshufb        (@X[3],@X[3],$t3);
1024         &vpaddd         ($t1,@X[1],&QWP(16,$K256));
1025         &vpaddd         ($t2,@X[2],&QWP(32,$K256));
1026         &vpaddd         ($t3,@X[3],&QWP(48,$K256));
1027         &vmovdqa        (&QWP(32+0,"esp"),$t0);
1028         &vmovdqa        (&QWP(32+16,"esp"),$t1);
1029         &vmovdqa        (&QWP(32+32,"esp"),$t2);
1030         &vmovdqa        (&QWP(32+48,"esp"),$t3);
1031         &jmp            (&label("avx_00_47"));
1032
1033 &set_label("avx_00_47",16);
1034         &add            ($K256,64);
1035
1036 sub Xupdate_AVX () {
1037         (
1038         '&vpalignr      ($t0,@X[1],@X[0],4);',  # X[1..4]
1039          '&vpalignr     ($t3,@X[3],@X[2],4);',  # X[9..12]
1040         '&vpsrld        ($t2,$t0,7);',
1041          '&vpaddd       (@X[0],@X[0],$t3);',    # X[0..3] += X[9..16]
1042         '&vpsrld        ($t3,$t0,3);',
1043         '&vpslld        ($t1,$t0,14);',
1044         '&vpxor         ($t0,$t3,$t2);',
1045          '&vpshufd      ($t3,@X[3],0b11111010)',# X[14..15]
1046         '&vpsrld        ($t2,$t2,18-7);',
1047         '&vpxor         ($t0,$t0,$t1);',
1048         '&vpslld        ($t1,$t1,25-14);',
1049         '&vpxor         ($t0,$t0,$t2);',
1050          '&vpsrld       ($t2,$t3,10);',
1051         '&vpxor         ($t0,$t0,$t1);',        # sigma0(X[1..4])
1052          '&vpsrlq       ($t1,$t3,17);',
1053         '&vpaddd        (@X[0],@X[0],$t0);',    # X[0..3] += sigma0(X[1..4])
1054          '&vpxor        ($t2,$t2,$t1);',
1055          '&vpsrlq       ($t3,$t3,19);',
1056          '&vpxor        ($t2,$t2,$t3);',        # sigma1(X[14..15]
1057          '&vpshufd      ($t3,$t2,0b10000100);',
1058         '&vpsrldq       ($t3,$t3,8);',
1059         '&vpaddd        (@X[0],@X[0],$t3);',    # X[0..1] += sigma1(X[14..15])
1060          '&vpshufd      ($t3,@X[0],0b01010000)',# X[16..17]
1061          '&vpsrld       ($t2,$t3,10);',
1062          '&vpsrlq       ($t1,$t3,17);',
1063          '&vpxor        ($t2,$t2,$t1);',
1064          '&vpsrlq       ($t3,$t3,19);',
1065          '&vpxor        ($t2,$t2,$t3);',        # sigma1(X[16..17]
1066          '&vpshufd      ($t3,$t2,0b11101000);',
1067         '&vpslldq       ($t3,$t3,8);',
1068         '&vpaddd        (@X[0],@X[0],$t3);'     # X[2..3] += sigma1(X[16..17])
1069         );
1070 }
1071
1072 local *ror = sub { &shrd(@_[0],@_) };
1073 sub AVX_00_47 () {
1074 my $j = shift;
1075 my $body = shift;
1076 my @X = @_;
1077 my @insns = (&$body,&$body,&$body,&$body);      # 120 instructions
1078 my $insn;
1079
1080         foreach (Xupdate_AVX()) {               # 31 instructions
1081             eval;
1082             eval(shift(@insns));
1083             eval(shift(@insns));
1084             eval($insn = shift(@insns));
1085             eval(shift(@insns)) if ($insn =~ /rorx/ && @insns[0] =~ /rorx/);
1086         }
1087         &vpaddd         ($t2,@X[0],&QWP(16*$j,$K256));
1088         foreach (@insns) { eval; }              # remaining instructions
1089         &vmovdqa        (&QWP(32+16*$j,"esp"),$t2);
1090 }
1091
1092     for ($i=0,$j=0; $j<4; $j++) {
1093         &AVX_00_47($j,\&body_00_15,@X);
1094         push(@X,shift(@X));             # rotate(@X)
1095     }
1096         &cmp    (&DWP(16*$j,$K256),0x00010203);
1097         &jne    (&label("avx_00_47"));
1098
1099     for ($i=0; $i<16; ) {
1100         foreach(body_00_15()) { eval; }
1101     }
1102
1103         &mov    ("esi",&DWP(96,"esp")); #ctx
1104                                         #&mov   ($AH[0],&DWP(0,"esp"));
1105         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
1106                                         #&mov   ("edi", &DWP(8,"esp"));
1107         &mov    ("ecx",&DWP(12,"esp"));
1108         &add    ($AH[0],&DWP(0,"esi"));
1109         &add    ($AH[1],&DWP(4,"esi"));
1110         &add    ("edi",&DWP(8,"esi"));
1111         &add    ("ecx",&DWP(12,"esi"));
1112         &mov    (&DWP(0,"esi"),$AH[0]);
1113         &mov    (&DWP(4,"esi"),$AH[1]);
1114         &mov    (&DWP(8,"esi"),"edi");
1115         &mov    (&DWP(12,"esi"),"ecx");
1116          #&mov  (&DWP(0,"esp"),$AH[0]);
1117          &mov   (&DWP(4,"esp"),$AH[1]);
1118          &xor   ($AH[1],"edi");                 # magic
1119          &mov   (&DWP(8,"esp"),"edi");
1120          &mov   (&DWP(12,"esp"),"ecx");
1121         #&mov   ($E,&DWP(16,"esp"));
1122         &mov    ("edi",&DWP(20,"esp"));
1123         &mov    ("ecx",&DWP(24,"esp"));
1124         &add    ($E,&DWP(16,"esi"));
1125         &add    ("edi",&DWP(20,"esi"));
1126         &add    ("ecx",&DWP(24,"esi"));
1127         &mov    (&DWP(16,"esi"),$E);
1128         &mov    (&DWP(20,"esi"),"edi");
1129          &mov   (&DWP(20,"esp"),"edi");
1130         &mov    ("edi",&DWP(28,"esp"));
1131         &mov    (&DWP(24,"esi"),"ecx");
1132          #&mov  (&DWP(16,"esp"),$E);
1133         &add    ("edi",&DWP(28,"esi"));
1134          &mov   (&DWP(24,"esp"),"ecx");
1135         &mov    (&DWP(28,"esi"),"edi");
1136          &mov   (&DWP(28,"esp"),"edi");
1137         &mov    ("edi",&DWP(96+4,"esp"));       # inp
1138
1139         &vmovdqa        ($t3,&QWP(64,$K256));
1140         &sub    ($K256,3*64);                   # rewind K
1141         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
1142         &jb     (&label("grand_avx"));
1143
1144         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
1145         &vzeroall       ();
1146 &function_end_A();
1147                                                 if ($avx>1) {
1148 sub bodyx_00_15 () {                    # +10%
1149         (
1150         '&rorx  ("ecx",$E,6)',
1151         '&rorx  ("esi",$E,11)',
1152          '&mov  (&off($e),$E)',         # save $E, modulo-scheduled
1153         '&rorx  ("edi",$E,25)',
1154         '&xor   ("ecx","esi")',
1155          '&andn ("esi",$E,&off($g))',
1156         '&xor   ("ecx","edi")',         # Sigma1(e)
1157          '&and  ($E,&off($f))',
1158          '&mov  (&off($a),$AH[0]);',    # save $A, modulo-scheduled
1159          '&or   ($E,"esi")',            # T = Ch(e,f,g)
1160
1161         '&rorx  ("edi",$AH[0],2)',
1162         '&rorx  ("esi",$AH[0],13)',
1163          '&lea  ($E,&DWP(0,$E,"ecx"))', # T += Sigma1(e)
1164         '&rorx  ("ecx",$AH[0],22)',
1165         '&xor   ("esi","edi")',
1166          '&mov  ("edi",&off($b))',
1167         '&xor   ("ecx","esi")',         # Sigma0(a)
1168
1169          '&xor  ($AH[0],"edi")',        # a ^= b, (b^c) in next round
1170          '&add  ($E,&off($h))',         # T += h
1171          '&and  ($AH[1],$AH[0])',       # (b^c) &= (a^b)
1172          '&add  ($E,&DWP(32+4*($i&15),"esp"))', # T += K[i]+X[i]
1173          '&xor  ($AH[1],"edi")',        # h = Maj(a,b,c) = Ch(a^b,c,b)
1174
1175          '&add  ("ecx",$E)',            # h += T
1176          '&add  ($E,&off($d))',         # d += T
1177         '&lea   ($AH[1],&DWP(0,$AH[1],"ecx"));'.        # h += Sigma0(a)
1178
1179         '@AH = reverse(@AH); $i++;'     # rotate(a,h)
1180         );
1181 }
1182
1183 &set_label("AVX_BMI",32);
1184         &lea    ("esp",&DWP(-96,"esp"));
1185         &vzeroall       ();
1186         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
1187         &mov    ($AH[0],&DWP(0,"esi"));
1188         &mov    ($AH[1],&DWP(4,"esi"));
1189         &mov    ("ecx",&DWP(8,"esi"));
1190         &mov    ("edi",&DWP(12,"esi"));
1191         #&mov   (&DWP(0,"esp"),$AH[0]);
1192         &mov    (&DWP(4,"esp"),$AH[1]);
1193         &xor    ($AH[1],"ecx");                 # magic
1194         &mov    (&DWP(8,"esp"),"ecx");
1195         &mov    (&DWP(12,"esp"),"edi");
1196         &mov    ($E,&DWP(16,"esi"));
1197         &mov    ("edi",&DWP(20,"esi"));
1198         &mov    ("ecx",&DWP(24,"esi"));
1199         &mov    ("esi",&DWP(28,"esi"));
1200         #&mov   (&DWP(16,"esp"),$E);
1201         &mov    (&DWP(20,"esp"),"edi");
1202         &mov    ("edi",&DWP(96+4,"esp"));       # inp
1203         &mov    (&DWP(24,"esp"),"ecx");
1204         &mov    (&DWP(28,"esp"),"esi");
1205         &vmovdqa        ($t3,&QWP(256,$K256));
1206         &jmp    (&label("grand_avx_bmi"));
1207
1208 &set_label("grand_avx_bmi",32);
1209         # load input, reverse byte order, add K256[0..15], save to stack
1210         &vmovdqu        (@X[0],&QWP(0,"edi"));
1211         &vmovdqu        (@X[1],&QWP(16,"edi"));
1212         &vmovdqu        (@X[2],&QWP(32,"edi"));
1213         &vmovdqu        (@X[3],&QWP(48,"edi"));
1214         &add            ("edi",64);
1215         &vpshufb        (@X[0],@X[0],$t3);
1216         &mov            (&DWP(96+4,"esp"),"edi");
1217         &vpshufb        (@X[1],@X[1],$t3);
1218         &vpshufb        (@X[2],@X[2],$t3);
1219         &vpaddd         ($t0,@X[0],&QWP(0,$K256));
1220         &vpshufb        (@X[3],@X[3],$t3);
1221         &vpaddd         ($t1,@X[1],&QWP(16,$K256));
1222         &vpaddd         ($t2,@X[2],&QWP(32,$K256));
1223         &vpaddd         ($t3,@X[3],&QWP(48,$K256));
1224         &vmovdqa        (&QWP(32+0,"esp"),$t0);
1225         &vmovdqa        (&QWP(32+16,"esp"),$t1);
1226         &vmovdqa        (&QWP(32+32,"esp"),$t2);
1227         &vmovdqa        (&QWP(32+48,"esp"),$t3);
1228         &jmp            (&label("avx_bmi_00_47"));
1229
1230 &set_label("avx_bmi_00_47",16);
1231         &add            ($K256,64);
1232
1233     for ($i=0,$j=0; $j<4; $j++) {
1234         &AVX_00_47($j,\&bodyx_00_15,@X);
1235         push(@X,shift(@X));             # rotate(@X)
1236     }
1237         &cmp    (&DWP(16*$j,$K256),0x00010203);
1238         &jne    (&label("avx_bmi_00_47"));
1239
1240     for ($i=0; $i<16; ) {
1241         foreach(bodyx_00_15()) { eval; }
1242     }
1243
1244         &mov    ("esi",&DWP(96,"esp")); #ctx
1245                                         #&mov   ($AH[0],&DWP(0,"esp"));
1246         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
1247                                         #&mov   ("edi", &DWP(8,"esp"));
1248         &mov    ("ecx",&DWP(12,"esp"));
1249         &add    ($AH[0],&DWP(0,"esi"));
1250         &add    ($AH[1],&DWP(4,"esi"));
1251         &add    ("edi",&DWP(8,"esi"));
1252         &add    ("ecx",&DWP(12,"esi"));
1253         &mov    (&DWP(0,"esi"),$AH[0]);
1254         &mov    (&DWP(4,"esi"),$AH[1]);
1255         &mov    (&DWP(8,"esi"),"edi");
1256         &mov    (&DWP(12,"esi"),"ecx");
1257          #&mov  (&DWP(0,"esp"),$AH[0]);
1258          &mov   (&DWP(4,"esp"),$AH[1]);
1259          &xor   ($AH[1],"edi");                 # magic
1260          &mov   (&DWP(8,"esp"),"edi");
1261          &mov   (&DWP(12,"esp"),"ecx");
1262         #&mov   ($E,&DWP(16,"esp"));
1263         &mov    ("edi",&DWP(20,"esp"));
1264         &mov    ("ecx",&DWP(24,"esp"));
1265         &add    ($E,&DWP(16,"esi"));
1266         &add    ("edi",&DWP(20,"esi"));
1267         &add    ("ecx",&DWP(24,"esi"));
1268         &mov    (&DWP(16,"esi"),$E);
1269         &mov    (&DWP(20,"esi"),"edi");
1270          &mov   (&DWP(20,"esp"),"edi");
1271         &mov    ("edi",&DWP(28,"esp"));
1272         &mov    (&DWP(24,"esi"),"ecx");
1273          #&mov  (&DWP(16,"esp"),$E);
1274         &add    ("edi",&DWP(28,"esi"));
1275          &mov   (&DWP(24,"esp"),"ecx");
1276         &mov    (&DWP(28,"esi"),"edi");
1277          &mov   (&DWP(28,"esp"),"edi");
1278         &mov    ("edi",&DWP(96+4,"esp"));       # inp
1279
1280         &vmovdqa        ($t3,&QWP(64,$K256));
1281         &sub    ($K256,3*64);                   # rewind K
1282         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
1283         &jb     (&label("grand_avx_bmi"));
1284
1285         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
1286         &vzeroall       ();
1287 &function_end_A();
1288                                                 }
1289                                                 }
1290                                                 }}}
1291 &function_end_B("sha256_block_data_order");
1292
1293 &asm_finish();
1294
1295 close STDOUT or die "error closing STDOUT: $!";