crypto/sha/asm/sha256-586.pl

   1 #!/usr/bin/env perl
   2 #
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@openssl.org> 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 # SHA256 block transform for x86. September 2007.
  11 #
  12 # Performance improvement over compiler generated code varies from
  13 # 10% to 40% [see below]. Not very impressive on some µ-archs, but
  14 # it's 5 times smaller and optimizies amount of writes.
  15 #
  16 # May 2012.
  17 #
  18 # Optimization including two of Pavel Semjanov's ideas, alternative
  19 # Maj and full unroll, resulted in ~20-25% improvement on most CPUs,
  20 # ~7% on Pentium, ~40% on Atom. As fully unrolled loop body is almost
  21 # 15x larger, 8KB vs. 560B, it's fired only for longer inputs. But not
  22 # on P4, where it kills performance, nor Sandy Bridge, where folded
  23 # loop is approximately as fast...
  24 #
  25 # June 2012.
  26 #
  27 # Add AMD XOP-specific code path, >30% improvement on Bulldozer over
  28 # May version, >60% over original. Add AVX+shrd code path, >25%
  29 # improvement on Sandy Bridge over May version, 60% over original.
  30 #
  31 # May 2013.
  32 #
  33 # Replace AMD XOP code path with SSSE3 to cover more processors.
  34 # (Biggest improvement coefficient is on upcoming Atom Silvermont,
  35 # not shown.) Add AVX+BMI code path.
  36 #
  37 # March 2014.
  38 #
  39 # Add support for Intel SHA Extensions.
  40 #
  41 # Performance in clock cycles per processed byte (less is better):
  42 #
  43 #               gcc     icc     x86 asm(*)      SIMD    x86_64 asm(**)
  44 # Pentium       46      57      40/38           -       -
  45 # PIII          36      33      27/24           -       -
  46 # P4            41      38      28              -       17.3
  47 # AMD K8        27      25      19/15.5         -       14.9
  48 # Core2         26      23      18/15.6         14.3    13.8
  49 # Westmere      27      -       19/15.7         13.4    12.3
  50 # Sandy Bridge  25      -       15.9            12.4    11.6
  51 # Ivy Bridge    24      -       15.0            11.4    10.3
  52 # Haswell       22      -       13.9            9.46    7.80
  53 # Bulldozer     36      -       27/22           17.0    13.6
  54 # VIA Nano      36      -       25/22           16.8    16.5
  55 # Atom          50      -       30/25           21.9    18.9
  56 #
  57 # (*)   numbers after slash are for unrolled loop, where applicable;
  58 # (**)  x86_64 assembly performance is presented for reference
  59 #       purposes, results are best-available;
  60
  61 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  62 push(@INC,"${dir}","${dir}../../perlasm");
  63 require "x86asm.pl";
  64
  65 &asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");
  66
  67 $xmm=$avx=0;
  68 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
  69
  70 if ($xmm &&     `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
  71                         =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
  72         $avx = ($1>=2.19) + ($1>=2.22);
  73 }
  74
  75 if ($xmm && !$avx && $ARGV[0] eq "win32n" &&
  76                 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
  77         $avx = ($1>=2.03) + ($1>=2.10);
  78 }
  79
  80 if ($xmm && !$avx && $ARGV[0] eq "win32" &&
  81                 `ml 2>&1` =~ /Version ([0-9]+)\./) {
  82         $avx = ($1>=10) + ($1>=11);
  83 }
  84
  85 if ($xmm && !$avx && `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/) {
  86         $avx = ($2>=3.0) + ($2>3.0);
  87 }
  88
  89 $shaext=$xmm;   ### set to zero if compiling for 1.0.1
  90
  91 $unroll_after = 64*4;   # If pre-evicted from L1P cache first spin of
  92                         # fully unrolled loop was measured to run about
  93                         # 3-4x slower. If slowdown coefficient is N and
  94                         # unrolled loop is m times faster, then you break
  95                         # even at (N-1)/(m-1) blocks. Then it needs to be
  96                         # adjusted for probability of code being evicted,
  97                         # code size/cache size=1/4. Typical m is 1.15...
  98
  99 $A="eax";
 100 $E="edx";
 101 $T="ebx";
 102 $Aoff=&DWP(4,"esp");
 103 $Boff=&DWP(8,"esp");
 104 $Coff=&DWP(12,"esp");
 105 $Doff=&DWP(16,"esp");
 106 $Eoff=&DWP(20,"esp");
 107 $Foff=&DWP(24,"esp");
 108 $Goff=&DWP(28,"esp");
 109 $Hoff=&DWP(32,"esp");
 110 $Xoff=&DWP(36,"esp");
 111 $K256="ebp";
 112
 113 sub BODY_16_63() {
 114         &mov    ($T,"ecx");                     # "ecx" is preloaded
 115          &mov   ("esi",&DWP(4*(9+15+16-14),"esp"));
 116         &ror    ("ecx",18-7);
 117          &mov   ("edi","esi");
 118         &ror    ("esi",19-17);
 119          &xor   ("ecx",$T);
 120          &shr   ($T,3);
 121         &ror    ("ecx",7);
 122          &xor   ("esi","edi");
 123          &xor   ($T,"ecx");                     # T = sigma0(X[-15])
 124         &ror    ("esi",17);
 125          &add   ($T,&DWP(4*(9+15+16),"esp"));   # T += X[-16]
 126         &shr    ("edi",10);
 127          &add   ($T,&DWP(4*(9+15+16-9),"esp")); # T += X[-7]
 128         #&xor   ("edi","esi")                   # sigma1(X[-2])
 129         # &add  ($T,"edi");                     # T += sigma1(X[-2])
 130         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
 131
 132         &BODY_00_15(1);
 133 }
 134 sub BODY_00_15() {
 135     my $in_16_63=shift;
 136
 137         &mov    ("ecx",$E);
 138          &xor   ("edi","esi")                   if ($in_16_63); # sigma1(X[-2])
 139          &mov   ("esi",$Foff);
 140         &ror    ("ecx",25-11);
 141          &add   ($T,"edi")                      if ($in_16_63); # T += sigma1(X[-2])
 142          &mov   ("edi",$Goff);
 143         &xor    ("ecx",$E);
 144          &xor   ("esi","edi");
 145          &mov   ($T,&DWP(4*(9+15),"esp"))       if (!$in_16_63);
 146          &mov   (&DWP(4*(9+15),"esp"),$T)       if ($in_16_63); # save X[0]
 147         &ror    ("ecx",11-6);
 148          &and   ("esi",$E);
 149          &mov   ($Eoff,$E);             # modulo-scheduled
 150         &xor    ($E,"ecx");
 151          &add   ($T,$Hoff);             # T += h
 152          &xor   ("esi","edi");          # Ch(e,f,g)
 153         &ror    ($E,6);                 # Sigma1(e)
 154          &mov   ("ecx",$A);
 155          &add   ($T,"esi");             # T += Ch(e,f,g)
 156
 157         &ror    ("ecx",22-13);
 158          &add   ($T,$E);                # T += Sigma1(e)
 159          &mov   ("edi",$Boff);
 160         &xor    ("ecx",$A);
 161          &mov   ($Aoff,$A);             # modulo-scheduled
 162          &lea   ("esp",&DWP(-4,"esp"));
 163         &ror    ("ecx",13-2);
 164          &mov   ("esi",&DWP(0,$K256));
 165         &xor    ("ecx",$A);
 166          &mov   ($E,$Eoff);             # e in next iteration, d in this one
 167          &xor   ($A,"edi");             # a ^= b
 168         &ror    ("ecx",2);              # Sigma0(a)
 169
 170          &add   ($T,"esi");             # T+= K[i]
 171          &mov   (&DWP(0,"esp"),$A);     # (b^c) in next round
 172         &add    ($E,$T);                # d += T
 173          &and   ($A,&DWP(4,"esp"));     # a &= (b^c)
 174         &add    ($T,"ecx");             # T += Sigma0(a)
 175          &xor   ($A,"edi");             # h = Maj(a,b,c) = Ch(a^b,c,b)
 176          &mov   ("ecx",&DWP(4*(9+15+16-1),"esp"))       if ($in_16_63); # preload T
 177         &add    ($K256,4);
 178          &add   ($A,$T);                # h += T
 179 }
 180
 181 &external_label("OPENSSL_ia32cap_P")            if (!$i386);
 182
 183 &function_begin("sha256_block_data_order");
 184         &mov    ("esi",wparam(0));      # ctx
 185         &mov    ("edi",wparam(1));      # inp
 186         &mov    ("eax",wparam(2));      # num
 187         &mov    ("ebx","esp");          # saved sp
 188
 189         &call   (&label("pic_point"));  # make it PIC!
 190 &set_label("pic_point");
 191         &blindpop($K256);
 192         &lea    ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256));
 193
 194         &sub    ("esp",16);
 195         &and    ("esp",-64);
 196
 197         &shl    ("eax",6);
 198         &add    ("eax","edi");
 199         &mov    (&DWP(0,"esp"),"esi");  # ctx
 200         &mov    (&DWP(4,"esp"),"edi");  # inp
 201         &mov    (&DWP(8,"esp"),"eax");  # inp+num*128
 202         &mov    (&DWP(12,"esp"),"ebx"); # saved sp
 203                                                 if (!$i386 && $xmm) {
 204         &picmeup("edx","OPENSSL_ia32cap_P",$K256,&label("K256"));
 205         &mov    ("ecx",&DWP(0,"edx"));
 206         &mov    ("ebx",&DWP(4,"edx"));
 207         &test   ("ecx",1<<20);          # check for P4
 208         &jnz    (&label("loop"));
 209         &mov    ("edx",&DWP(8,"edx"))   if ($xmm);
 210         &test   ("ecx",1<<24);          # check for FXSR
 211         &jz     ($unroll_after?&label("no_xmm"):&label("loop"));
 212         &and    ("ecx",1<<30);          # mask "Intel CPU" bit
 213         &and    ("ebx",1<<28|1<<9);     # mask AVX and SSSE3 bits
 214         &test   ("edx",1<<29)           if ($shaext);   # check for SHA
 215         &jnz    (&label("shaext"))      if ($shaext);
 216         &or     ("ecx","ebx");
 217         &and    ("ecx",1<<28|1<<30);
 218         &cmp    ("ecx",1<<28|1<<30);
 219                                         if ($xmm) {
 220         &je     (&label("AVX"))         if ($avx);
 221         &test   ("ebx",1<<9);           # check for SSSE3
 222         &jnz    (&label("SSSE3"));
 223                                         } else {
 224         &je     (&label("loop_shrd"));
 225                                         }
 226                                                 if ($unroll_after) {
 227 &set_label("no_xmm");
 228         &sub    ("eax","edi");
 229         &cmp    ("eax",$unroll_after);
 230         &jae    (&label("unrolled"));
 231                                                 } }
 232         &jmp    (&label("loop"));
 233
 234 sub COMPACT_LOOP() {
 235 my $suffix=shift;
 236
 237 &set_label("loop$suffix",$suffix?32:16);
 238     # copy input block to stack reversing byte and dword order
 239     for($i=0;$i<4;$i++) {
 240         &mov    ("eax",&DWP($i*16+0,"edi"));
 241         &mov    ("ebx",&DWP($i*16+4,"edi"));
 242         &mov    ("ecx",&DWP($i*16+8,"edi"));
 243         &bswap  ("eax");
 244         &mov    ("edx",&DWP($i*16+12,"edi"));
 245         &bswap  ("ebx");
 246         &push   ("eax");
 247         &bswap  ("ecx");
 248         &push   ("ebx");
 249         &bswap  ("edx");
 250         &push   ("ecx");
 251         &push   ("edx");
 252     }
 253         &add    ("edi",64);
 254         &lea    ("esp",&DWP(-4*9,"esp"));# place for A,B,C,D,E,F,G,H
 255         &mov    (&DWP(4*(9+16)+4,"esp"),"edi");
 256
 257         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 258         &mov    ($A,&DWP(0,"esi"));
 259         &mov    ("ebx",&DWP(4,"esi"));
 260         &mov    ("ecx",&DWP(8,"esi"));
 261         &mov    ("edi",&DWP(12,"esi"));
 262         # &mov  ($Aoff,$A);
 263         &mov    ($Boff,"ebx");
 264         &xor    ("ebx","ecx");
 265         &mov    ($Coff,"ecx");
 266         &mov    ($Doff,"edi");
 267         &mov    (&DWP(0,"esp"),"ebx");  # magic
 268         &mov    ($E,&DWP(16,"esi"));
 269         &mov    ("ebx",&DWP(20,"esi"));
 270         &mov    ("ecx",&DWP(24,"esi"));
 271         &mov    ("edi",&DWP(28,"esi"));
 272         # &mov  ($Eoff,$E);
 273         &mov    ($Foff,"ebx");
 274         &mov    ($Goff,"ecx");
 275         &mov    ($Hoff,"edi");
 276
 277 &set_label("00_15$suffix",16);
 278
 279         &BODY_00_15();
 280
 281         &cmp    ("esi",0xc19bf174);
 282         &jne    (&label("00_15$suffix"));
 283
 284         &mov    ("ecx",&DWP(4*(9+15+16-1),"esp"));      # preloaded in BODY_00_15(1)
 285         &jmp    (&label("16_63$suffix"));
 286
 287 &set_label("16_63$suffix",16);
 288
 289         &BODY_16_63();
 290
 291         &cmp    ("esi",0xc67178f2);
 292         &jne    (&label("16_63$suffix"));
 293
 294         &mov    ("esi",&DWP(4*(9+16+64)+0,"esp"));#ctx
 295         # &mov  ($A,$Aoff);
 296         &mov    ("ebx",$Boff);
 297         # &mov  ("edi",$Coff);
 298         &mov    ("ecx",$Doff);
 299         &add    ($A,&DWP(0,"esi"));
 300         &add    ("ebx",&DWP(4,"esi"));
 301         &add    ("edi",&DWP(8,"esi"));
 302         &add    ("ecx",&DWP(12,"esi"));
 303         &mov    (&DWP(0,"esi"),$A);
 304         &mov    (&DWP(4,"esi"),"ebx");
 305         &mov    (&DWP(8,"esi"),"edi");
 306         &mov    (&DWP(12,"esi"),"ecx");
 307         # &mov  ($E,$Eoff);
 308         &mov    ("eax",$Foff);
 309         &mov    ("ebx",$Goff);
 310         &mov    ("ecx",$Hoff);
 311         &mov    ("edi",&DWP(4*(9+16+64)+4,"esp"));#inp
 312         &add    ($E,&DWP(16,"esi"));
 313         &add    ("eax",&DWP(20,"esi"));
 314         &add    ("ebx",&DWP(24,"esi"));
 315         &add    ("ecx",&DWP(28,"esi"));
 316         &mov    (&DWP(16,"esi"),$E);
 317         &mov    (&DWP(20,"esi"),"eax");
 318         &mov    (&DWP(24,"esi"),"ebx");
 319         &mov    (&DWP(28,"esi"),"ecx");
 320
 321         &lea    ("esp",&DWP(4*(9+16+64),"esp"));# destroy frame
 322         &sub    ($K256,4*64);                   # rewind K
 323
 324         &cmp    ("edi",&DWP(8,"esp"));          # are we done yet?
 325         &jb     (&label("loop$suffix"));
 326 }
 327         &COMPACT_LOOP();
 328         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 329 &function_end_A();
 330                                                 if (!$i386 && !$xmm) {
 331         # ~20% improvement on Sandy Bridge
 332         local *ror = sub { &shrd(@_[0],@_) };
 333         &COMPACT_LOOP("_shrd");
 334         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 335 &function_end_A();
 336                                                 }
 337
 338 &set_label("K256",64);  # Yes! I keep it in the code segment!
 339 @K256=( 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5,
 340         0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5,
 341         0xd807aa98,0x12835b01,0x243185be,0x550c7dc3,
 342         0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174,
 343         0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc,
 344         0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da,
 345         0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7,
 346         0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967,
 347         0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13,
 348         0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85,
 349         0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3,
 350         0xd192e819,0xd6990624,0xf40e3585,0x106aa070,
 351         0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5,
 352         0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3,
 353         0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208,
 354         0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2     );
 355 &data_word(@K256);
 356 &data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f);        # byte swap mask
 357 &asciz("SHA256 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
 358
 359 ($a,$b,$c,$d,$e,$f,$g,$h)=(0..7);       # offsets
 360 sub off { &DWP(4*(((shift)-$i)&7),"esp"); }
 361
 362 if (!$i386 && $unroll_after) {
 363 my @AH=($A,$K256);
 364
 365 &set_label("unrolled",16);
 366         &lea    ("esp",&DWP(-96,"esp"));
 367         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 368         &mov    ($AH[0],&DWP(0,"esi"));
 369         &mov    ($AH[1],&DWP(4,"esi"));
 370         &mov    ("ecx",&DWP(8,"esi"));
 371         &mov    ("ebx",&DWP(12,"esi"));
 372         #&mov   (&DWP(0,"esp"),$AH[0]);
 373         &mov    (&DWP(4,"esp"),$AH[1]);
 374         &xor    ($AH[1],"ecx");         # magic
 375         &mov    (&DWP(8,"esp"),"ecx");
 376         &mov    (&DWP(12,"esp"),"ebx");
 377         &mov    ($E,&DWP(16,"esi"));
 378         &mov    ("ebx",&DWP(20,"esi"));
 379         &mov    ("ecx",&DWP(24,"esi"));
 380         &mov    ("esi",&DWP(28,"esi"));
 381         #&mov   (&DWP(16,"esp"),$E);
 382         &mov    (&DWP(20,"esp"),"ebx");
 383         &mov    (&DWP(24,"esp"),"ecx");
 384         &mov    (&DWP(28,"esp"),"esi");
 385         &jmp    (&label("grand_loop"));
 386
 387 &set_label("grand_loop",16);
 388     # copy input block to stack reversing byte order
 389     for($i=0;$i<5;$i++) {
 390         &mov    ("ebx",&DWP(12*$i+0,"edi"));
 391         &mov    ("ecx",&DWP(12*$i+4,"edi"));
 392         &bswap  ("ebx");
 393         &mov    ("esi",&DWP(12*$i+8,"edi"));
 394         &bswap  ("ecx");
 395         &mov    (&DWP(32+12*$i+0,"esp"),"ebx");
 396         &bswap  ("esi");
 397         &mov    (&DWP(32+12*$i+4,"esp"),"ecx");
 398         &mov    (&DWP(32+12*$i+8,"esp"),"esi");
 399     }
 400         &mov    ("ebx",&DWP($i*12,"edi"));
 401         &add    ("edi",64);
 402         &bswap  ("ebx");
 403         &mov    (&DWP(96+4,"esp"),"edi");
 404         &mov    (&DWP(32+12*$i,"esp"),"ebx");
 405
 406     my ($t1,$t2) = ("ecx","esi");
 407
 408     for ($i=0;$i<64;$i++) {
 409
 410       if ($i>=16) {
 411         &mov    ($T,$t1);                       # $t1 is preloaded
 412         # &mov  ($t2,&DWP(32+4*(($i+14)&15),"esp"));
 413         &ror    ($t1,18-7);
 414          &mov   ("edi",$t2);
 415         &ror    ($t2,19-17);
 416          &xor   ($t1,$T);
 417          &shr   ($T,3);
 418         &ror    ($t1,7);
 419          &xor   ($t2,"edi");
 420          &xor   ($T,$t1);                       # T = sigma0(X[-15])
 421         &ror    ($t2,17);
 422          &add   ($T,&DWP(32+4*($i&15),"esp"));  # T += X[-16]
 423         &shr    ("edi",10);
 424          &add   ($T,&DWP(32+4*(($i+9)&15),"esp"));      # T += X[-7]
 425         #&xor   ("edi",$t2)                     # sigma1(X[-2])
 426         # &add  ($T,"edi");                     # T += sigma1(X[-2])
 427         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
 428       }
 429         &mov    ($t1,$E);
 430          &xor   ("edi",$t2)                     if ($i>=16);    # sigma1(X[-2])
 431          &mov   ($t2,&off($f));
 432         &ror    ($E,25-11);
 433          &add   ($T,"edi")                      if ($i>=16);    # T += sigma1(X[-2])
 434          &mov   ("edi",&off($g));
 435         &xor    ($E,$t1);
 436          &mov   ($T,&DWP(32+4*($i&15),"esp"))   if ($i<16);     # X[i]
 437          &mov   (&DWP(32+4*($i&15),"esp"),$T)   if ($i>=16 && $i<62);   # save X[0]
 438          &xor   ($t2,"edi");
 439         &ror    ($E,11-6);
 440          &and   ($t2,$t1);
 441          &mov   (&off($e),$t1);         # save $E, modulo-scheduled
 442         &xor    ($E,$t1);
 443          &add   ($T,&off($h));          # T += h
 444          &xor   ("edi",$t2);            # Ch(e,f,g)
 445         &ror    ($E,6);                 # Sigma1(e)
 446          &mov   ($t1,$AH[0]);
 447          &add   ($T,"edi");             # T += Ch(e,f,g)
 448
 449         &ror    ($t1,22-13);
 450          &mov   ($t2,$AH[0]);
 451          &mov   ("edi",&off($b));
 452         &xor    ($t1,$AH[0]);
 453          &mov   (&off($a),$AH[0]);      # save $A, modulo-scheduled
 454          &xor   ($AH[0],"edi");         # a ^= b, (b^c) in next round
 455         &ror    ($t1,13-2);
 456          &and   ($AH[1],$AH[0]);        # (b^c) &= (a^b)
 457          &lea   ($E,&DWP(@K256[$i],$T,$E));     # T += Sigma1(1)+K[i]
 458         &xor    ($t1,$t2);
 459          &xor   ($AH[1],"edi");         # h = Maj(a,b,c) = Ch(a^b,c,b)
 460          &mov   ($t2,&DWP(32+4*(($i+2)&15),"esp"))      if ($i>=15 && $i<63);
 461         &ror    ($t1,2);                # Sigma0(a)
 462
 463          &add   ($AH[1],$E);            # h += T
 464          &add   ($E,&off($d));          # d += T
 465         &add    ($AH[1],$t1);           # h += Sigma0(a)
 466          &mov   ($t1,&DWP(32+4*(($i+15)&15),"esp"))     if ($i>=15 && $i<63);
 467
 468         @AH = reverse(@AH);             # rotate(a,h)
 469         ($t1,$t2) = ($t2,$t1);          # rotate(t1,t2)
 470     }
 471         &mov    ("esi",&DWP(96,"esp")); #ctx
 472                                         #&mov   ($AH[0],&DWP(0,"esp"));
 473         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
 474                                         #&mov   ("edi", &DWP(8,"esp"));
 475         &mov    ("ecx",&DWP(12,"esp"));
 476         &add    ($AH[0],&DWP(0,"esi"));
 477         &add    ($AH[1],&DWP(4,"esi"));
 478         &add    ("edi",&DWP(8,"esi"));
 479         &add    ("ecx",&DWP(12,"esi"));
 480         &mov    (&DWP(0,"esi"),$AH[0]);
 481         &mov    (&DWP(4,"esi"),$AH[1]);
 482         &mov    (&DWP(8,"esi"),"edi");
 483         &mov    (&DWP(12,"esi"),"ecx");
 484          #&mov  (&DWP(0,"esp"),$AH[0]);
 485          &mov   (&DWP(4,"esp"),$AH[1]);
 486          &xor   ($AH[1],"edi");         # magic
 487          &mov   (&DWP(8,"esp"),"edi");
 488          &mov   (&DWP(12,"esp"),"ecx");
 489         #&mov   ($E,&DWP(16,"esp"));
 490         &mov    ("edi",&DWP(20,"esp"));
 491         &mov    ("ebx",&DWP(24,"esp"));
 492         &mov    ("ecx",&DWP(28,"esp"));
 493         &add    ($E,&DWP(16,"esi"));
 494         &add    ("edi",&DWP(20,"esi"));
 495         &add    ("ebx",&DWP(24,"esi"));
 496         &add    ("ecx",&DWP(28,"esi"));
 497         &mov    (&DWP(16,"esi"),$E);
 498         &mov    (&DWP(20,"esi"),"edi");
 499         &mov    (&DWP(24,"esi"),"ebx");
 500         &mov    (&DWP(28,"esi"),"ecx");
 501          #&mov  (&DWP(16,"esp"),$E);
 502          &mov   (&DWP(20,"esp"),"edi");
 503         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 504          &mov   (&DWP(24,"esp"),"ebx");
 505          &mov   (&DWP(28,"esp"),"ecx");
 506
 507         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
 508         &jb     (&label("grand_loop"));
 509
 510         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
 511 &function_end_A();
 512 }
 513                                                 if (!$i386 && $xmm) {{{
 514 if ($shaext) {
 515 ######################################################################
 516 # Intel SHA Extensions implementation of SHA256 update function.
 517 #
 518 my ($ctx,$inp,$end)=("esi","edi","eax");
 519 my ($Wi,$ABEF,$CDGH,$TMP)=map("xmm$_",(0..2,7));
 520 my @MSG=map("xmm$_",(3..6));
 521
 522 sub sha256op38 {
 523  my ($opcodelet,$dst,$src)=@_;
 524     if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
 525     {   &data_byte(0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2);       }
 526 }
 527 sub sha256rnds2 { sha256op38(0xcb,@_); }
 528 sub sha256msg1  { sha256op38(0xcc,@_); }
 529 sub sha256msg2  { sha256op38(0xcd,@_); }
 530
 531 &set_label("shaext",32);
 532         &sub            ("esp",32);
 533
 534         &movdqu         ($ABEF,&QWP(0,$ctx));           # DCBA
 535         &lea            ($K256,&DWP(0x80,$K256));
 536         &movdqu         ($CDGH,&QWP(16,$ctx));          # HGFE
 537         &movdqa         ($TMP,&QWP(0x100-0x80,$K256));  # byte swap mask
 538
 539         &pshufd         ($Wi,$ABEF,0x1b);               # ABCD
 540         &pshufd         ($ABEF,$ABEF,0xb1);             # CDAB
 541         &pshufd         ($CDGH,$CDGH,0x1b);             # EFGH
 542         &palignr        ($ABEF,$CDGH,8);                # ABEF
 543         &punpcklqdq     ($CDGH,$Wi);                    # CDGH
 544         &jmp            (&label("loop_shaext"));
 545
 546 &set_label("loop_shaext",16);
 547         &movdqu         (@MSG[0],&QWP(0,$inp));
 548         &movdqu         (@MSG[1],&QWP(0x10,$inp));
 549         &movdqu         (@MSG[2],&QWP(0x20,$inp));
 550         &pshufb         (@MSG[0],$TMP);
 551         &movdqu         (@MSG[3],&QWP(0x30,$inp));
 552         &movdqa         (&QWP(16,"esp"),$CDGH);         # offload
 553
 554         &movdqa         ($Wi,&QWP(0*16-0x80,$K256));
 555         &paddd          ($Wi,@MSG[0]);
 556         &pshufb         (@MSG[1],$TMP);
 557         &sha256rnds2    ($CDGH,$ABEF);                  # 0-3
 558         &pshufd         ($Wi,$Wi,0x0e);
 559         &nop            ();
 560         &movdqa         (&QWP(0,"esp"),$ABEF);          # offload
 561         &sha256rnds2    ($ABEF,$CDGH);
 562
 563         &movdqa         ($Wi,&QWP(1*16-0x80,$K256));
 564         &paddd          ($Wi,@MSG[1]);
 565         &pshufb         (@MSG[2],$TMP);
 566         &sha256rnds2    ($CDGH,$ABEF);                  # 4-7
 567         &pshufd         ($Wi,$Wi,0x0e);
 568         &lea            ($inp,&DWP(0x40,$inp));
 569         &sha256msg1     (@MSG[0],@MSG[1]);
 570         &sha256rnds2    ($ABEF,$CDGH);
 571
 572         &movdqa         ($Wi,&QWP(2*16-0x80,$K256));
 573         &paddd          ($Wi,@MSG[2]);
 574         &pshufb         (@MSG[3],$TMP);
 575         &sha256rnds2    ($CDGH,$ABEF);                  # 8-11
 576         &pshufd         ($Wi,$Wi,0x0e);
 577         &movdqa         ($TMP,@MSG[3]);
 578         &palignr        ($TMP,@MSG[2],4);
 579         &nop            ();
 580         &paddd          (@MSG[0],$TMP);
 581         &sha256msg1     (@MSG[1],@MSG[2]);
 582         &sha256rnds2    ($ABEF,$CDGH);
 583
 584         &movdqa         ($Wi,&QWP(3*16-0x80,$K256));
 585         &paddd          ($Wi,@MSG[3]);
 586         &sha256msg2     (@MSG[0],@MSG[3]);
 587         &sha256rnds2    ($CDGH,$ABEF);                  # 12-15
 588         &pshufd         ($Wi,$Wi,0x0e);
 589         &movdqa         ($TMP,@MSG[0]);
 590         &palignr        ($TMP,@MSG[3],4);
 591         &nop            ();
 592         &paddd          (@MSG[1],$TMP);
 593         &sha256msg1     (@MSG[2],@MSG[3]);
 594         &sha256rnds2    ($ABEF,$CDGH);
 595
 596 for($i=4;$i<16-3;$i++) {
 597         &movdqa         ($Wi,&QWP($i*16-0x80,$K256));
 598         &paddd          ($Wi,@MSG[0]);
 599         &sha256msg2     (@MSG[1],@MSG[0]);
 600         &sha256rnds2    ($CDGH,$ABEF);                  # 16-19...
 601         &pshufd         ($Wi,$Wi,0x0e);
 602         &movdqa         ($TMP,@MSG[1]);
 603         &palignr        ($TMP,@MSG[0],4);
 604         &nop            ();
 605         &paddd          (@MSG[2],$TMP);
 606         &sha256msg1     (@MSG[3],@MSG[0]);
 607         &sha256rnds2    ($ABEF,$CDGH);
 608
 609         push(@MSG,shift(@MSG));
 610 }
 611         &movdqa         ($Wi,&QWP(13*16-0x80,$K256));
 612         &paddd          ($Wi,@MSG[0]);
 613         &sha256msg2     (@MSG[1],@MSG[0]);
 614         &sha256rnds2    ($CDGH,$ABEF);                  # 52-55
 615         &pshufd         ($Wi,$Wi,0x0e);
 616         &movdqa         ($TMP,@MSG[1])
 617         &palignr        ($TMP,@MSG[0],4);
 618         &sha256rnds2    ($ABEF,$CDGH);
 619         &paddd          (@MSG[2],$TMP);
 620
 621         &movdqa         ($Wi,&QWP(14*16-0x80,$K256));
 622         &paddd          ($Wi,@MSG[1]);
 623         &sha256rnds2    ($CDGH,$ABEF);                  # 56-59
 624         &pshufd         ($Wi,$Wi,0x0e);
 625         &sha256msg2     (@MSG[2],@MSG[1]);
 626         &movdqa         ($TMP,&QWP(0x100-0x80,$K256));  # byte swap mask
 627         &sha256rnds2    ($ABEF,$CDGH);
 628
 629         &movdqa         ($Wi,&QWP(15*16-0x80,$K256));
 630         &paddd          ($Wi,@MSG[2]);
 631         &nop            ();
 632         &sha256rnds2    ($CDGH,$ABEF);                  # 60-63
 633         &pshufd         ($Wi,$Wi,0x0e);
 634         &cmp            ($end,$inp);
 635         &nop            ();
 636         &sha256rnds2    ($ABEF,$CDGH);
 637
 638         &paddd          ($CDGH,&QWP(16,"esp"));
 639         &paddd          ($ABEF,&QWP(0,"esp"));
 640         &jnz            (&label("loop_shaext"));
 641
 642         &pshufd         ($CDGH,$CDGH,0xb1);             # DCHG
 643         &pshufd         ($TMP,$ABEF,0x1b);              # FEBA
 644         &pshufd         ($ABEF,$ABEF,0xb1);             # BAFE
 645         &punpckhqdq     ($ABEF,$CDGH);                  # DCBA
 646         &palignr        ($CDGH,$TMP,8);                 # HGFE
 647
 648         &mov            ("esp",&DWP(32+12,"esp"));
 649         &movdqu         (&QWP(0,$ctx),$ABEF);
 650         &movdqu         (&QWP(16,$ctx),$CDGH);
 651 &function_end_A();
 652 }
 653
 654 my @X = map("xmm$_",(0..3));
 655 my ($t0,$t1,$t2,$t3) = map("xmm$_",(4..7));
 656 my @AH = ($A,$T);
 657
 658 &set_label("SSSE3",32);
 659         &lea    ("esp",&DWP(-96,"esp"));
 660         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 661         &mov    ($AH[0],&DWP(0,"esi"));
 662         &mov    ($AH[1],&DWP(4,"esi"));
 663         &mov    ("ecx",&DWP(8,"esi"));
 664         &mov    ("edi",&DWP(12,"esi"));
 665         #&mov   (&DWP(0,"esp"),$AH[0]);
 666         &mov    (&DWP(4,"esp"),$AH[1]);
 667         &xor    ($AH[1],"ecx");                 # magic
 668         &mov    (&DWP(8,"esp"),"ecx");
 669         &mov    (&DWP(12,"esp"),"edi");
 670         &mov    ($E,&DWP(16,"esi"));
 671         &mov    ("edi",&DWP(20,"esi"));
 672         &mov    ("ecx",&DWP(24,"esi"));
 673         &mov    ("esi",&DWP(28,"esi"));
 674         #&mov   (&DWP(16,"esp"),$E);
 675         &mov    (&DWP(20,"esp"),"edi");
 676         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 677         &mov    (&DWP(24,"esp"),"ecx");
 678         &mov    (&DWP(28,"esp"),"esi");
 679         &movdqa ($t3,&QWP(256,$K256));
 680         &jmp    (&label("grand_ssse3"));
 681
 682 &set_label("grand_ssse3",16);
 683         # load input, reverse byte order, add K256[0..15], save to stack
 684         &movdqu (@X[0],&QWP(0,"edi"));
 685         &movdqu (@X[1],&QWP(16,"edi"));
 686         &movdqu (@X[2],&QWP(32,"edi"));
 687         &movdqu (@X[3],&QWP(48,"edi"));
 688         &add    ("edi",64);
 689         &pshufb (@X[0],$t3);
 690         &mov    (&DWP(96+4,"esp"),"edi");
 691         &pshufb (@X[1],$t3);
 692         &movdqa ($t0,&QWP(0,$K256));
 693         &pshufb (@X[2],$t3);
 694         &movdqa ($t1,&QWP(16,$K256));
 695         &paddd  ($t0,@X[0]);
 696         &pshufb (@X[3],$t3);
 697         &movdqa ($t2,&QWP(32,$K256));
 698         &paddd  ($t1,@X[1]);
 699         &movdqa ($t3,&QWP(48,$K256));
 700         &movdqa (&QWP(32+0,"esp"),$t0);
 701         &paddd  ($t2,@X[2]);
 702         &movdqa (&QWP(32+16,"esp"),$t1);
 703         &paddd  ($t3,@X[3]);
 704         &movdqa (&QWP(32+32,"esp"),$t2);
 705         &movdqa (&QWP(32+48,"esp"),$t3);
 706         &jmp    (&label("ssse3_00_47"));
 707
 708 &set_label("ssse3_00_47",16);
 709         &add            ($K256,64);
 710
 711 sub SSSE3_00_47 () {
 712 my $j = shift;
 713 my $body = shift;
 714 my @X = @_;
 715 my @insns = (&$body,&$body,&$body,&$body);      # 120 instructions
 716
 717           eval(shift(@insns));
 718         &movdqa         ($t0,@X[1]);
 719           eval(shift(@insns));                  # @
 720           eval(shift(@insns));
 721         &movdqa         ($t3,@X[3]);
 722           eval(shift(@insns));
 723           eval(shift(@insns));
 724         &palignr        ($t0,@X[0],4);          # X[1..4]
 725           eval(shift(@insns));
 726           eval(shift(@insns));                  # @
 727           eval(shift(@insns));
 728          &palignr       ($t3,@X[2],4);          # X[9..12]
 729           eval(shift(@insns));
 730           eval(shift(@insns));
 731           eval(shift(@insns));
 732         &movdqa         ($t1,$t0);
 733           eval(shift(@insns));                  # @
 734           eval(shift(@insns));
 735         &movdqa         ($t2,$t0);
 736           eval(shift(@insns));
 737           eval(shift(@insns));
 738         &psrld          ($t0,3);
 739           eval(shift(@insns));
 740           eval(shift(@insns));                  # @
 741          &paddd         (@X[0],$t3);            # X[0..3] += X[9..12]
 742           eval(shift(@insns));
 743           eval(shift(@insns));
 744         &psrld          ($t2,7);
 745           eval(shift(@insns));
 746           eval(shift(@insns));
 747           eval(shift(@insns));                  # @
 748           eval(shift(@insns));
 749          &pshufd        ($t3,@X[3],0b11111010); # X[14..15]
 750           eval(shift(@insns));
 751           eval(shift(@insns));
 752         &pslld          ($t1,32-18);
 753           eval(shift(@insns));
 754           eval(shift(@insns));                  # @
 755         &pxor           ($t0,$t2);
 756           eval(shift(@insns));
 757           eval(shift(@insns));
 758         &psrld          ($t2,18-7);
 759           eval(shift(@insns));
 760           eval(shift(@insns));
 761           eval(shift(@insns));                  # @
 762         &pxor           ($t0,$t1);
 763           eval(shift(@insns));
 764           eval(shift(@insns));
 765         &pslld          ($t1,18-7);
 766           eval(shift(@insns));
 767           eval(shift(@insns));
 768           eval(shift(@insns));                  # @
 769         &pxor           ($t0,$t2);
 770           eval(shift(@insns));
 771           eval(shift(@insns));
 772          &movdqa        ($t2,$t3);
 773           eval(shift(@insns));
 774           eval(shift(@insns));
 775           eval(shift(@insns));                  # @
 776         &pxor           ($t0,$t1);              # sigma0(X[1..4])
 777           eval(shift(@insns));
 778           eval(shift(@insns));
 779          &psrld         ($t3,10);
 780           eval(shift(@insns));
 781           eval(shift(@insns));
 782           eval(shift(@insns));                  # @
 783         &paddd          (@X[0],$t0);            # X[0..3] += sigma0(X[1..4])
 784           eval(shift(@insns));
 785           eval(shift(@insns));
 786          &psrlq         ($t2,17);
 787           eval(shift(@insns));
 788           eval(shift(@insns));
 789           eval(shift(@insns));                  # @
 790          &pxor          ($t3,$t2);
 791           eval(shift(@insns));
 792           eval(shift(@insns));
 793          &psrlq         ($t2,19-17);
 794           eval(shift(@insns));
 795           eval(shift(@insns));
 796           eval(shift(@insns));                  # @
 797          &pxor          ($t3,$t2);
 798           eval(shift(@insns));
 799           eval(shift(@insns));
 800          &pshufd        ($t3,$t3,0b10000000);
 801           eval(shift(@insns));
 802           eval(shift(@insns));
 803           eval(shift(@insns));                  # @
 804           eval(shift(@insns));
 805           eval(shift(@insns));
 806           eval(shift(@insns));
 807           eval(shift(@insns));
 808           eval(shift(@insns));                  # @
 809           eval(shift(@insns));
 810          &psrldq        ($t3,8);
 811           eval(shift(@insns));
 812           eval(shift(@insns));
 813           eval(shift(@insns));
 814         &paddd          (@X[0],$t3);            # X[0..1] += sigma1(X[14..15])
 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          &pshufd        ($t3,@X[0],0b01010000); # X[16..17]
 823           eval(shift(@insns));
 824           eval(shift(@insns));
 825           eval(shift(@insns));
 826          &movdqa        ($t2,$t3);
 827           eval(shift(@insns));                  # @
 828          &psrld         ($t3,10);
 829           eval(shift(@insns));
 830          &psrlq         ($t2,17);
 831           eval(shift(@insns));
 832           eval(shift(@insns));
 833           eval(shift(@insns));
 834           eval(shift(@insns));                  # @
 835          &pxor          ($t3,$t2);
 836           eval(shift(@insns));
 837           eval(shift(@insns));
 838          &psrlq         ($t2,19-17);
 839           eval(shift(@insns));
 840           eval(shift(@insns));
 841           eval(shift(@insns));                  # @
 842          &pxor          ($t3,$t2);
 843           eval(shift(@insns));
 844           eval(shift(@insns));
 845           eval(shift(@insns));
 846          &pshufd        ($t3,$t3,0b00001000);
 847           eval(shift(@insns));
 848           eval(shift(@insns));                  # @
 849         &movdqa         ($t2,&QWP(16*$j,$K256));
 850           eval(shift(@insns));
 851           eval(shift(@insns));
 852          &pslldq        ($t3,8);
 853           eval(shift(@insns));
 854           eval(shift(@insns));
 855           eval(shift(@insns));                  # @
 856           eval(shift(@insns));
 857           eval(shift(@insns));
 858           eval(shift(@insns));
 859           eval(shift(@insns));
 860           eval(shift(@insns));                  # @
 861         &paddd          (@X[0],$t3);            # X[2..3] += sigma1(X[16..17])
 862           eval(shift(@insns));
 863           eval(shift(@insns));
 864           eval(shift(@insns));
 865           eval(shift(@insns));
 866         &paddd          ($t2,@X[0]);
 867           eval(shift(@insns));                  # @
 868
 869         foreach (@insns) { eval; }              # remaining instructions
 870
 871         &movdqa         (&QWP(32+16*$j,"esp"),$t2);
 872 }
 873
 874 sub body_00_15 () {
 875         (
 876         '&mov   ("ecx",$E);',
 877         '&ror   ($E,25-11);',
 878          '&mov  ("esi",&off($f));',
 879         '&xor   ($E,"ecx");',
 880          '&mov  ("edi",&off($g));',
 881          '&xor  ("esi","edi");',
 882         '&ror   ($E,11-6);',
 883          '&and  ("esi","ecx");',
 884          '&mov  (&off($e),"ecx");',     # save $E, modulo-scheduled
 885         '&xor   ($E,"ecx");',
 886          '&xor  ("edi","esi");',        # Ch(e,f,g)
 887         '&ror   ($E,6);',               # T = Sigma1(e)
 888          '&mov  ("ecx",$AH[0]);',
 889          '&add  ($E,"edi");',           # T += Ch(e,f,g)
 890          '&mov  ("edi",&off($b));',
 891         '&mov   ("esi",$AH[0]);',
 892
 893         '&ror   ("ecx",22-13);',
 894          '&mov  (&off($a),$AH[0]);',    # save $A, modulo-scheduled
 895         '&xor   ("ecx",$AH[0]);',
 896          '&xor  ($AH[0],"edi");',       # a ^= b, (b^c) in next round
 897          '&add  ($E,&off($h));',        # T += h
 898         '&ror   ("ecx",13-2);',
 899          '&and  ($AH[1],$AH[0]);',      # (b^c) &= (a^b)
 900         '&xor   ("ecx","esi");',
 901          '&add  ($E,&DWP(32+4*($i&15),"esp"));',        # T += K[i]+X[i]
 902          '&xor  ($AH[1],"edi");',       # h = Maj(a,b,c) = Ch(a^b,c,b)
 903         '&ror   ("ecx",2);',            # Sigma0(a)
 904
 905          '&add  ($AH[1],$E);',          # h += T
 906          '&add  ($E,&off($d));',        # d += T
 907         '&add   ($AH[1],"ecx");'.       # h += Sigma0(a)
 908
 909         '@AH = reverse(@AH); $i++;'     # rotate(a,h)
 910         );
 911 }
 912
 913     for ($i=0,$j=0; $j<4; $j++) {
 914         &SSSE3_00_47($j,\&body_00_15,@X);
 915         push(@X,shift(@X));             # rotate(@X)
 916     }
 917         &cmp    (&DWP(16*$j,$K256),0x00010203);
 918         &jne    (&label("ssse3_00_47"));
 919
 920     for ($i=0; $i<16; ) {
 921         foreach(body_00_15()) { eval; }
 922     }
 923
 924         &mov    ("esi",&DWP(96,"esp")); #ctx
 925                                         #&mov   ($AH[0],&DWP(0,"esp"));
 926         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
 927                                         #&mov   ("edi", &DWP(8,"esp"));
 928         &mov    ("ecx",&DWP(12,"esp"));
 929         &add    ($AH[0],&DWP(0,"esi"));
 930         &add    ($AH[1],&DWP(4,"esi"));
 931         &add    ("edi",&DWP(8,"esi"));
 932         &add    ("ecx",&DWP(12,"esi"));
 933         &mov    (&DWP(0,"esi"),$AH[0]);
 934         &mov    (&DWP(4,"esi"),$AH[1]);
 935         &mov    (&DWP(8,"esi"),"edi");
 936         &mov    (&DWP(12,"esi"),"ecx");
 937          #&mov  (&DWP(0,"esp"),$AH[0]);
 938          &mov   (&DWP(4,"esp"),$AH[1]);
 939          &xor   ($AH[1],"edi");                 # magic
 940          &mov   (&DWP(8,"esp"),"edi");
 941          &mov   (&DWP(12,"esp"),"ecx");
 942         #&mov   ($E,&DWP(16,"esp"));
 943         &mov    ("edi",&DWP(20,"esp"));
 944         &mov    ("ecx",&DWP(24,"esp"));
 945         &add    ($E,&DWP(16,"esi"));
 946         &add    ("edi",&DWP(20,"esi"));
 947         &add    ("ecx",&DWP(24,"esi"));
 948         &mov    (&DWP(16,"esi"),$E);
 949         &mov    (&DWP(20,"esi"),"edi");
 950          &mov   (&DWP(20,"esp"),"edi");
 951         &mov    ("edi",&DWP(28,"esp"));
 952         &mov    (&DWP(24,"esi"),"ecx");
 953          #&mov  (&DWP(16,"esp"),$E);
 954         &add    ("edi",&DWP(28,"esi"));
 955          &mov   (&DWP(24,"esp"),"ecx");
 956         &mov    (&DWP(28,"esi"),"edi");
 957          &mov   (&DWP(28,"esp"),"edi");
 958         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 959
 960         &movdqa ($t3,&QWP(64,$K256));
 961         &sub    ($K256,3*64);                   # rewind K
 962         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
 963         &jb     (&label("grand_ssse3"));
 964
 965         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
 966 &function_end_A();
 967                                                 if ($avx) {
 968 &set_label("AVX",32);
 969                                                 if ($avx>1) {
 970         &and    ("edx",1<<8|1<<3);              # check for BMI2+BMI1
 971         &cmp    ("edx",1<<8|1<<3);
 972         &je     (&label("AVX_BMI"));
 973                                                 }
 974         &lea    ("esp",&DWP(-96,"esp"));
 975         &vzeroall       ();
 976         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 977         &mov    ($AH[0],&DWP(0,"esi"));
 978         &mov    ($AH[1],&DWP(4,"esi"));
 979         &mov    ("ecx",&DWP(8,"esi"));
 980         &mov    ("edi",&DWP(12,"esi"));
 981         #&mov   (&DWP(0,"esp"),$AH[0]);
 982         &mov    (&DWP(4,"esp"),$AH[1]);
 983         &xor    ($AH[1],"ecx");                 # magic
 984         &mov    (&DWP(8,"esp"),"ecx");
 985         &mov    (&DWP(12,"esp"),"edi");
 986         &mov    ($E,&DWP(16,"esi"));
 987         &mov    ("edi",&DWP(20,"esi"));
 988         &mov    ("ecx",&DWP(24,"esi"));
 989         &mov    ("esi",&DWP(28,"esi"));
 990         #&mov   (&DWP(16,"esp"),$E);
 991         &mov    (&DWP(20,"esp"),"edi");
 992         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 993         &mov    (&DWP(24,"esp"),"ecx");
 994         &mov    (&DWP(28,"esp"),"esi");
 995         &vmovdqa        ($t3,&QWP(256,$K256));
 996         &jmp    (&label("grand_avx"));
 997
 998 &set_label("grand_avx",32);
 999         # load input, reverse byte order, add K256[0..15], save to stack
1000         &vmovdqu        (@X[0],&QWP(0,"edi"));
1001         &vmovdqu        (@X[1],&QWP(16,"edi"));
1002         &vmovdqu        (@X[2],&QWP(32,"edi"));
1003         &vmovdqu        (@X[3],&QWP(48,"edi"));
1004         &add            ("edi",64);
1005         &vpshufb        (@X[0],@X[0],$t3);
1006         &mov            (&DWP(96+4,"esp"),"edi");
1007         &vpshufb        (@X[1],@X[1],$t3);
1008         &vpshufb        (@X[2],@X[2],$t3);
1009         &vpaddd         ($t0,@X[0],&QWP(0,$K256));
1010         &vpshufb        (@X[3],@X[3],$t3);
1011         &vpaddd         ($t1,@X[1],&QWP(16,$K256));
1012         &vpaddd         ($t2,@X[2],&QWP(32,$K256));
1013         &vpaddd         ($t3,@X[3],&QWP(48,$K256));
1014         &vmovdqa        (&QWP(32+0,"esp"),$t0);
1015         &vmovdqa        (&QWP(32+16,"esp"),$t1);
1016         &vmovdqa        (&QWP(32+32,"esp"),$t2);
1017         &vmovdqa        (&QWP(32+48,"esp"),$t3);
1018         &jmp            (&label("avx_00_47"));
1019
1020 &set_label("avx_00_47",16);
1021         &add            ($K256,64);
1022
1023 sub Xupdate_AVX () {
1024         (
1025         '&vpalignr      ($t0,@X[1],@X[0],4);',  # X[1..4]
1026          '&vpalignr     ($t3,@X[3],@X[2],4);',  # X[9..12]
1027         '&vpsrld        ($t2,$t0,7);',
1028          '&vpaddd       (@X[0],@X[0],$t3);',    # X[0..3] += X[9..16]
1029         '&vpsrld        ($t3,$t0,3);',
1030         '&vpslld        ($t1,$t0,14);',
1031         '&vpxor         ($t0,$t3,$t2);',
1032          '&vpshufd      ($t3,@X[3],0b11111010)',# X[14..15]
1033         '&vpsrld        ($t2,$t2,18-7);',
1034         '&vpxor         ($t0,$t0,$t1);',
1035         '&vpslld        ($t1,$t1,25-14);',
1036         '&vpxor         ($t0,$t0,$t2);',
1037          '&vpsrld       ($t2,$t3,10);',
1038         '&vpxor         ($t0,$t0,$t1);',        # sigma0(X[1..4])
1039          '&vpsrlq       ($t1,$t3,17);',
1040         '&vpaddd        (@X[0],@X[0],$t0);',    # X[0..3] += sigma0(X[1..4])
1041          '&vpxor        ($t2,$t2,$t1);',
1042          '&vpsrlq       ($t3,$t3,19);',
1043          '&vpxor        ($t2,$t2,$t3);',        # sigma1(X[14..15]
1044          '&vpshufd      ($t3,$t2,0b10000100);',
1045         '&vpsrldq       ($t3,$t3,8);',
1046         '&vpaddd        (@X[0],@X[0],$t3);',    # X[0..1] += sigma1(X[14..15])
1047          '&vpshufd      ($t3,@X[0],0b01010000)',# X[16..17]
1048          '&vpsrld       ($t2,$t3,10);',
1049          '&vpsrlq       ($t1,$t3,17);',
1050          '&vpxor        ($t2,$t2,$t1);',
1051          '&vpsrlq       ($t3,$t3,19);',
1052          '&vpxor        ($t2,$t2,$t3);',        # sigma1(X[16..17]
1053          '&vpshufd      ($t3,$t2,0b11101000);',
1054         '&vpslldq       ($t3,$t3,8);',
1055         '&vpaddd        (@X[0],@X[0],$t3);'     # X[2..3] += sigma1(X[16..17])
1056         );
1057 }
1058
1059 local *ror = sub { &shrd(@_[0],@_) };
1060 sub AVX_00_47 () {
1061 my $j = shift;
1062 my $body = shift;
1063 my @X = @_;
1064 my @insns = (&$body,&$body,&$body,&$body);      # 120 instructions
1065 my $insn;
1066
1067         foreach (Xupdate_AVX()) {               # 31 instructions
1068             eval;
1069             eval(shift(@insns));
1070             eval(shift(@insns));
1071             eval($insn = shift(@insns));
1072             eval(shift(@insns)) if ($insn =~ /rorx/ && @insns[0] =~ /rorx/);
1073         }
1074         &vpaddd         ($t2,@X[0],&QWP(16*$j,$K256));
1075         foreach (@insns) { eval; }              # remaining instructions
1076         &vmovdqa        (&QWP(32+16*$j,"esp"),$t2);
1077 }
1078
1079     for ($i=0,$j=0; $j<4; $j++) {
1080         &AVX_00_47($j,\&body_00_15,@X);
1081         push(@X,shift(@X));             # rotate(@X)
1082     }
1083         &cmp    (&DWP(16*$j,$K256),0x00010203);
1084         &jne    (&label("avx_00_47"));
1085
1086     for ($i=0; $i<16; ) {
1087         foreach(body_00_15()) { eval; }
1088     }
1089
1090         &mov    ("esi",&DWP(96,"esp")); #ctx
1091                                         #&mov   ($AH[0],&DWP(0,"esp"));
1092         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
1093                                         #&mov   ("edi", &DWP(8,"esp"));
1094         &mov    ("ecx",&DWP(12,"esp"));
1095         &add    ($AH[0],&DWP(0,"esi"));
1096         &add    ($AH[1],&DWP(4,"esi"));
1097         &add    ("edi",&DWP(8,"esi"));
1098         &add    ("ecx",&DWP(12,"esi"));
1099         &mov    (&DWP(0,"esi"),$AH[0]);
1100         &mov    (&DWP(4,"esi"),$AH[1]);
1101         &mov    (&DWP(8,"esi"),"edi");
1102         &mov    (&DWP(12,"esi"),"ecx");
1103          #&mov  (&DWP(0,"esp"),$AH[0]);
1104          &mov   (&DWP(4,"esp"),$AH[1]);
1105          &xor   ($AH[1],"edi");                 # magic
1106          &mov   (&DWP(8,"esp"),"edi");
1107          &mov   (&DWP(12,"esp"),"ecx");
1108         #&mov   ($E,&DWP(16,"esp"));
1109         &mov    ("edi",&DWP(20,"esp"));
1110         &mov    ("ecx",&DWP(24,"esp"));
1111         &add    ($E,&DWP(16,"esi"));
1112         &add    ("edi",&DWP(20,"esi"));
1113         &add    ("ecx",&DWP(24,"esi"));
1114         &mov    (&DWP(16,"esi"),$E);
1115         &mov    (&DWP(20,"esi"),"edi");
1116          &mov   (&DWP(20,"esp"),"edi");
1117         &mov    ("edi",&DWP(28,"esp"));
1118         &mov    (&DWP(24,"esi"),"ecx");
1119          #&mov  (&DWP(16,"esp"),$E);
1120         &add    ("edi",&DWP(28,"esi"));
1121          &mov   (&DWP(24,"esp"),"ecx");
1122         &mov    (&DWP(28,"esi"),"edi");
1123          &mov   (&DWP(28,"esp"),"edi");
1124         &mov    ("edi",&DWP(96+4,"esp"));       # inp
1125
1126         &vmovdqa        ($t3,&QWP(64,$K256));
1127         &sub    ($K256,3*64);                   # rewind K
1128         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
1129         &jb     (&label("grand_avx"));
1130
1131         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
1132         &vzeroall       ();
1133 &function_end_A();
1134                                                 if ($avx>1) {
1135 sub bodyx_00_15 () {                    # +10%
1136         (
1137         '&rorx  ("ecx",$E,6)',
1138         '&rorx  ("esi",$E,11)',
1139          '&mov  (&off($e),$E)',         # save $E, modulo-scheduled
1140         '&rorx  ("edi",$E,25)',
1141         '&xor   ("ecx","esi")',
1142          '&andn ("esi",$E,&off($g))',
1143         '&xor   ("ecx","edi")',         # Sigma1(e)
1144          '&and  ($E,&off($f))',
1145          '&mov  (&off($a),$AH[0]);',    # save $A, modulo-scheduled
1146          '&or   ($E,"esi")',            # T = Ch(e,f,g)
1147
1148         '&rorx  ("edi",$AH[0],2)',
1149         '&rorx  ("esi",$AH[0],13)',
1150          '&lea  ($E,&DWP(0,$E,"ecx"))', # T += Sigma1(e)
1151         '&rorx  ("ecx",$AH[0],22)',
1152         '&xor   ("esi","edi")',
1153          '&mov  ("edi",&off($b))',
1154         '&xor   ("ecx","esi")',         # Sigma0(a)
1155
1156          '&xor  ($AH[0],"edi")',        # a ^= b, (b^c) in next round
1157          '&add  ($E,&off($h))',         # T += h
1158          '&and  ($AH[1],$AH[0])',       # (b^c) &= (a^b)
1159          '&add  ($E,&DWP(32+4*($i&15),"esp"))', # T += K[i]+X[i]
1160          '&xor  ($AH[1],"edi")',        # h = Maj(a,b,c) = Ch(a^b,c,b)
1161
1162          '&add  ("ecx",$E)',            # h += T
1163          '&add  ($E,&off($d))',         # d += T
1164         '&lea   ($AH[1],&DWP(0,$AH[1],"ecx"));'.        # h += Sigma0(a)
1165
1166         '@AH = reverse(@AH); $i++;'     # rotate(a,h)
1167         );
1168 }
1169
1170 &set_label("AVX_BMI",32);
1171         &lea    ("esp",&DWP(-96,"esp"));
1172         &vzeroall       ();
1173         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
1174         &mov    ($AH[0],&DWP(0,"esi"));
1175         &mov    ($AH[1],&DWP(4,"esi"));
1176         &mov    ("ecx",&DWP(8,"esi"));
1177         &mov    ("edi",&DWP(12,"esi"));
1178         #&mov   (&DWP(0,"esp"),$AH[0]);
1179         &mov    (&DWP(4,"esp"),$AH[1]);
1180         &xor    ($AH[1],"ecx");                 # magic
1181         &mov    (&DWP(8,"esp"),"ecx");
1182         &mov    (&DWP(12,"esp"),"edi");
1183         &mov    ($E,&DWP(16,"esi"));
1184         &mov    ("edi",&DWP(20,"esi"));
1185         &mov    ("ecx",&DWP(24,"esi"));
1186         &mov    ("esi",&DWP(28,"esi"));
1187         #&mov   (&DWP(16,"esp"),$E);
1188         &mov    (&DWP(20,"esp"),"edi");
1189         &mov    ("edi",&DWP(96+4,"esp"));       # inp
1190         &mov    (&DWP(24,"esp"),"ecx");
1191         &mov    (&DWP(28,"esp"),"esi");
1192         &vmovdqa        ($t3,&QWP(256,$K256));
1193         &jmp    (&label("grand_avx_bmi"));
1194
1195 &set_label("grand_avx_bmi",32);
1196         # load input, reverse byte order, add K256[0..15], save to stack
1197         &vmovdqu        (@X[0],&QWP(0,"edi"));
1198         &vmovdqu        (@X[1],&QWP(16,"edi"));
1199         &vmovdqu        (@X[2],&QWP(32,"edi"));
1200         &vmovdqu        (@X[3],&QWP(48,"edi"));
1201         &add            ("edi",64);
1202         &vpshufb        (@X[0],@X[0],$t3);
1203         &mov            (&DWP(96+4,"esp"),"edi");
1204         &vpshufb        (@X[1],@X[1],$t3);
1205         &vpshufb        (@X[2],@X[2],$t3);
1206         &vpaddd         ($t0,@X[0],&QWP(0,$K256));
1207         &vpshufb        (@X[3],@X[3],$t3);
1208         &vpaddd         ($t1,@X[1],&QWP(16,$K256));
1209         &vpaddd         ($t2,@X[2],&QWP(32,$K256));
1210         &vpaddd         ($t3,@X[3],&QWP(48,$K256));
1211         &vmovdqa        (&QWP(32+0,"esp"),$t0);
1212         &vmovdqa        (&QWP(32+16,"esp"),$t1);
1213         &vmovdqa        (&QWP(32+32,"esp"),$t2);
1214         &vmovdqa        (&QWP(32+48,"esp"),$t3);
1215         &jmp            (&label("avx_bmi_00_47"));
1216
1217 &set_label("avx_bmi_00_47",16);
1218         &add            ($K256,64);
1219
1220     for ($i=0,$j=0; $j<4; $j++) {
1221         &AVX_00_47($j,\&bodyx_00_15,@X);
1222         push(@X,shift(@X));             # rotate(@X)
1223     }
1224         &cmp    (&DWP(16*$j,$K256),0x00010203);
1225         &jne    (&label("avx_bmi_00_47"));
1226
1227     for ($i=0; $i<16; ) {
1228         foreach(bodyx_00_15()) { eval; }
1229     }
1230
1231         &mov    ("esi",&DWP(96,"esp")); #ctx
1232                                         #&mov   ($AH[0],&DWP(0,"esp"));
1233         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
1234                                         #&mov   ("edi", &DWP(8,"esp"));
1235         &mov    ("ecx",&DWP(12,"esp"));
1236         &add    ($AH[0],&DWP(0,"esi"));
1237         &add    ($AH[1],&DWP(4,"esi"));
1238         &add    ("edi",&DWP(8,"esi"));
1239         &add    ("ecx",&DWP(12,"esi"));
1240         &mov    (&DWP(0,"esi"),$AH[0]);
1241         &mov    (&DWP(4,"esi"),$AH[1]);
1242         &mov    (&DWP(8,"esi"),"edi");
1243         &mov    (&DWP(12,"esi"),"ecx");
1244          #&mov  (&DWP(0,"esp"),$AH[0]);
1245          &mov   (&DWP(4,"esp"),$AH[1]);
1246          &xor   ($AH[1],"edi");                 # magic
1247          &mov   (&DWP(8,"esp"),"edi");
1248          &mov   (&DWP(12,"esp"),"ecx");
1249         #&mov   ($E,&DWP(16,"esp"));
1250         &mov    ("edi",&DWP(20,"esp"));
1251         &mov    ("ecx",&DWP(24,"esp"));
1252         &add    ($E,&DWP(16,"esi"));
1253         &add    ("edi",&DWP(20,"esi"));
1254         &add    ("ecx",&DWP(24,"esi"));
1255         &mov    (&DWP(16,"esi"),$E);
1256         &mov    (&DWP(20,"esi"),"edi");
1257          &mov   (&DWP(20,"esp"),"edi");
1258         &mov    ("edi",&DWP(28,"esp"));
1259         &mov    (&DWP(24,"esi"),"ecx");
1260          #&mov  (&DWP(16,"esp"),$E);
1261         &add    ("edi",&DWP(28,"esi"));
1262          &mov   (&DWP(24,"esp"),"ecx");
1263         &mov    (&DWP(28,"esi"),"edi");
1264          &mov   (&DWP(28,"esp"),"edi");
1265         &mov    ("edi",&DWP(96+4,"esp"));       # inp
1266
1267         &vmovdqa        ($t3,&QWP(64,$K256));
1268         &sub    ($K256,3*64);                   # rewind K
1269         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
1270         &jb     (&label("grand_avx_bmi"));
1271
1272         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
1273         &vzeroall       ();
1274 &function_end_A();
1275                                                 }
1276                                                 }
1277                                                 }}}
1278 &function_end_B("sha256_block_data_order");
1279
1280 &asm_finish();