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 # Silvermont    40      -       34/31           22.9    20.6
  57 #
  58 # (*)   numbers after slash are for unrolled loop, where applicable;
  59 # (**)  x86_64 assembly performance is presented for reference
  60 #       purposes, results are best-available;
  61
  62 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  63 push(@INC,"${dir}","${dir}../../perlasm");
  64 require "x86asm.pl";
  65
  66 $output=pop;
  67 open STDOUT,">$output";
  68
  69 &asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");
  70
  71 $xmm=$avx=0;
  72 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
  73
  74 if ($xmm &&     `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
  75                         =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
  76         $avx = ($1>=2.19) + ($1>=2.22);
  77 }
  78
  79 if ($xmm && !$avx && $ARGV[0] eq "win32n" &&
  80                 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
  81         $avx = ($1>=2.03) + ($1>=2.10);
  82 }
  83
  84 if ($xmm && !$avx && $ARGV[0] eq "win32" &&
  85                 `ml 2>&1` =~ /Version ([0-9]+)\./) {
  86         $avx = ($1>=10) + ($1>=11);
  87 }
  88
  89 if ($xmm && !$avx && `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/) {
  90         $avx = ($2>=3.0) + ($2>3.0);
  91 }
  92
  93 $shaext=$xmm;   ### set to zero if compiling for 1.0.1
  94
  95 $unroll_after = 64*4;   # If pre-evicted from L1P cache first spin of
  96                         # fully unrolled loop was measured to run about
  97                         # 3-4x slower. If slowdown coefficient is N and
  98                         # unrolled loop is m times faster, then you break
  99                         # even at (N-1)/(m-1) blocks. Then it needs to be
 100                         # adjusted for probability of code being evicted,
 101                         # code size/cache size=1/4. Typical m is 1.15...
 102
 103 $A="eax";
 104 $E="edx";
 105 $T="ebx";
 106 $Aoff=&DWP(4,"esp");
 107 $Boff=&DWP(8,"esp");
 108 $Coff=&DWP(12,"esp");
 109 $Doff=&DWP(16,"esp");
 110 $Eoff=&DWP(20,"esp");
 111 $Foff=&DWP(24,"esp");
 112 $Goff=&DWP(28,"esp");
 113 $Hoff=&DWP(32,"esp");
 114 $Xoff=&DWP(36,"esp");
 115 $K256="ebp";
 116
 117 sub BODY_16_63() {
 118         &mov    ($T,"ecx");                     # "ecx" is preloaded
 119          &mov   ("esi",&DWP(4*(9+15+16-14),"esp"));
 120         &ror    ("ecx",18-7);
 121          &mov   ("edi","esi");
 122         &ror    ("esi",19-17);
 123          &xor   ("ecx",$T);
 124          &shr   ($T,3);
 125         &ror    ("ecx",7);
 126          &xor   ("esi","edi");
 127          &xor   ($T,"ecx");                     # T = sigma0(X[-15])
 128         &ror    ("esi",17);
 129          &add   ($T,&DWP(4*(9+15+16),"esp"));   # T += X[-16]
 130         &shr    ("edi",10);
 131          &add   ($T,&DWP(4*(9+15+16-9),"esp")); # T += X[-7]
 132         #&xor   ("edi","esi")                   # sigma1(X[-2])
 133         # &add  ($T,"edi");                     # T += sigma1(X[-2])
 134         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
 135
 136         &BODY_00_15(1);
 137 }
 138 sub BODY_00_15() {
 139     my $in_16_63=shift;
 140
 141         &mov    ("ecx",$E);
 142          &xor   ("edi","esi")                   if ($in_16_63); # sigma1(X[-2])
 143          &mov   ("esi",$Foff);
 144         &ror    ("ecx",25-11);
 145          &add   ($T,"edi")                      if ($in_16_63); # T += sigma1(X[-2])
 146          &mov   ("edi",$Goff);
 147         &xor    ("ecx",$E);
 148          &xor   ("esi","edi");
 149          &mov   ($T,&DWP(4*(9+15),"esp"))       if (!$in_16_63);
 150          &mov   (&DWP(4*(9+15),"esp"),$T)       if ($in_16_63); # save X[0]
 151         &ror    ("ecx",11-6);
 152          &and   ("esi",$E);
 153          &mov   ($Eoff,$E);             # modulo-scheduled
 154         &xor    ($E,"ecx");
 155          &add   ($T,$Hoff);             # T += h
 156          &xor   ("esi","edi");          # Ch(e,f,g)
 157         &ror    ($E,6);                 # Sigma1(e)
 158          &mov   ("ecx",$A);
 159          &add   ($T,"esi");             # T += Ch(e,f,g)
 160
 161         &ror    ("ecx",22-13);
 162          &add   ($T,$E);                # T += Sigma1(e)
 163          &mov   ("edi",$Boff);
 164         &xor    ("ecx",$A);
 165          &mov   ($Aoff,$A);             # modulo-scheduled
 166          &lea   ("esp",&DWP(-4,"esp"));
 167         &ror    ("ecx",13-2);
 168          &mov   ("esi",&DWP(0,$K256));
 169         &xor    ("ecx",$A);
 170          &mov   ($E,$Eoff);             # e in next iteration, d in this one
 171          &xor   ($A,"edi");             # a ^= b
 172         &ror    ("ecx",2);              # Sigma0(a)
 173
 174          &add   ($T,"esi");             # T+= K[i]
 175          &mov   (&DWP(0,"esp"),$A);     # (b^c) in next round
 176         &add    ($E,$T);                # d += T
 177          &and   ($A,&DWP(4,"esp"));     # a &= (b^c)
 178         &add    ($T,"ecx");             # T += Sigma0(a)
 179          &xor   ($A,"edi");             # h = Maj(a,b,c) = Ch(a^b,c,b)
 180          &mov   ("ecx",&DWP(4*(9+15+16-1),"esp"))       if ($in_16_63); # preload T
 181         &add    ($K256,4);
 182          &add   ($A,$T);                # h += T
 183 }
 184
 185 &external_label("OPENSSL_ia32cap_P")            if (!$i386);
 186
 187 &function_begin("sha256_block_data_order");
 188         &mov    ("esi",wparam(0));      # ctx
 189         &mov    ("edi",wparam(1));      # inp
 190         &mov    ("eax",wparam(2));      # num
 191         &mov    ("ebx","esp");          # saved sp
 192
 193         &call   (&label("pic_point"));  # make it PIC!
 194 &set_label("pic_point");
 195         &blindpop($K256);
 196         &lea    ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256));
 197
 198         &sub    ("esp",16);
 199         &and    ("esp",-64);
 200
 201         &shl    ("eax",6);
 202         &add    ("eax","edi");
 203         &mov    (&DWP(0,"esp"),"esi");  # ctx
 204         &mov    (&DWP(4,"esp"),"edi");  # inp
 205         &mov    (&DWP(8,"esp"),"eax");  # inp+num*128
 206         &mov    (&DWP(12,"esp"),"ebx"); # saved sp
 207                                                 if (!$i386 && $xmm) {
 208         &picmeup("edx","OPENSSL_ia32cap_P",$K256,&label("K256"));
 209         &mov    ("ecx",&DWP(0,"edx"));
 210         &mov    ("ebx",&DWP(4,"edx"));
 211         &test   ("ecx",1<<20);          # check for P4
 212         &jnz    (&label("loop"));
 213         &mov    ("edx",&DWP(8,"edx"))   if ($xmm);
 214         &test   ("ecx",1<<24);          # check for FXSR
 215         &jz     ($unroll_after?&label("no_xmm"):&label("loop"));
 216         &and    ("ecx",1<<30);          # mask "Intel CPU" bit
 217         &and    ("ebx",1<<28|1<<9);     # mask AVX and SSSE3 bits
 218         &test   ("edx",1<<29)           if ($shaext);   # check for SHA
 219         &jnz    (&label("shaext"))      if ($shaext);
 220         &or     ("ecx","ebx");
 221         &and    ("ecx",1<<28|1<<30);
 222         &cmp    ("ecx",1<<28|1<<30);
 223                                         if ($xmm) {
 224         &je     (&label("AVX"))         if ($avx);
 225         &test   ("ebx",1<<9);           # check for SSSE3
 226         &jnz    (&label("SSSE3"));
 227                                         } else {
 228         &je     (&label("loop_shrd"));
 229                                         }
 230                                                 if ($unroll_after) {
 231 &set_label("no_xmm");
 232         &sub    ("eax","edi");
 233         &cmp    ("eax",$unroll_after);
 234         &jae    (&label("unrolled"));
 235                                                 } }
 236         &jmp    (&label("loop"));
 237
 238 sub COMPACT_LOOP() {
 239 my $suffix=shift;
 240
 241 &set_label("loop$suffix",$suffix?32:16);
 242     # copy input block to stack reversing byte and dword order
 243     for($i=0;$i<4;$i++) {
 244         &mov    ("eax",&DWP($i*16+0,"edi"));
 245         &mov    ("ebx",&DWP($i*16+4,"edi"));
 246         &mov    ("ecx",&DWP($i*16+8,"edi"));
 247         &bswap  ("eax");
 248         &mov    ("edx",&DWP($i*16+12,"edi"));
 249         &bswap  ("ebx");
 250         &push   ("eax");
 251         &bswap  ("ecx");
 252         &push   ("ebx");
 253         &bswap  ("edx");
 254         &push   ("ecx");
 255         &push   ("edx");
 256     }
 257         &add    ("edi",64);
 258         &lea    ("esp",&DWP(-4*9,"esp"));# place for A,B,C,D,E,F,G,H
 259         &mov    (&DWP(4*(9+16)+4,"esp"),"edi");
 260
 261         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 262         &mov    ($A,&DWP(0,"esi"));
 263         &mov    ("ebx",&DWP(4,"esi"));
 264         &mov    ("ecx",&DWP(8,"esi"));
 265         &mov    ("edi",&DWP(12,"esi"));
 266         # &mov  ($Aoff,$A);
 267         &mov    ($Boff,"ebx");
 268         &xor    ("ebx","ecx");
 269         &mov    ($Coff,"ecx");
 270         &mov    ($Doff,"edi");
 271         &mov    (&DWP(0,"esp"),"ebx");  # magic
 272         &mov    ($E,&DWP(16,"esi"));
 273         &mov    ("ebx",&DWP(20,"esi"));
 274         &mov    ("ecx",&DWP(24,"esi"));
 275         &mov    ("edi",&DWP(28,"esi"));
 276         # &mov  ($Eoff,$E);
 277         &mov    ($Foff,"ebx");
 278         &mov    ($Goff,"ecx");
 279         &mov    ($Hoff,"edi");
 280
 281 &set_label("00_15$suffix",16);
 282
 283         &BODY_00_15();
 284
 285         &cmp    ("esi",0xc19bf174);
 286         &jne    (&label("00_15$suffix"));
 287
 288         &mov    ("ecx",&DWP(4*(9+15+16-1),"esp"));      # preloaded in BODY_00_15(1)
 289         &jmp    (&label("16_63$suffix"));
 290
 291 &set_label("16_63$suffix",16);
 292
 293         &BODY_16_63();
 294
 295         &cmp    ("esi",0xc67178f2);
 296         &jne    (&label("16_63$suffix"));
 297
 298         &mov    ("esi",&DWP(4*(9+16+64)+0,"esp"));#ctx
 299         # &mov  ($A,$Aoff);
 300         &mov    ("ebx",$Boff);
 301         # &mov  ("edi",$Coff);
 302         &mov    ("ecx",$Doff);
 303         &add    ($A,&DWP(0,"esi"));
 304         &add    ("ebx",&DWP(4,"esi"));
 305         &add    ("edi",&DWP(8,"esi"));
 306         &add    ("ecx",&DWP(12,"esi"));
 307         &mov    (&DWP(0,"esi"),$A);
 308         &mov    (&DWP(4,"esi"),"ebx");
 309         &mov    (&DWP(8,"esi"),"edi");
 310         &mov    (&DWP(12,"esi"),"ecx");
 311         # &mov  ($E,$Eoff);
 312         &mov    ("eax",$Foff);
 313         &mov    ("ebx",$Goff);
 314         &mov    ("ecx",$Hoff);
 315         &mov    ("edi",&DWP(4*(9+16+64)+4,"esp"));#inp
 316         &add    ($E,&DWP(16,"esi"));
 317         &add    ("eax",&DWP(20,"esi"));
 318         &add    ("ebx",&DWP(24,"esi"));
 319         &add    ("ecx",&DWP(28,"esi"));
 320         &mov    (&DWP(16,"esi"),$E);
 321         &mov    (&DWP(20,"esi"),"eax");
 322         &mov    (&DWP(24,"esi"),"ebx");
 323         &mov    (&DWP(28,"esi"),"ecx");
 324
 325         &lea    ("esp",&DWP(4*(9+16+64),"esp"));# destroy frame
 326         &sub    ($K256,4*64);                   # rewind K
 327
 328         &cmp    ("edi",&DWP(8,"esp"));          # are we done yet?
 329         &jb     (&label("loop$suffix"));
 330 }
 331         &COMPACT_LOOP();
 332         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 333 &function_end_A();
 334                                                 if (!$i386 && !$xmm) {
 335         # ~20% improvement on Sandy Bridge
 336         local *ror = sub { &shrd(@_[0],@_) };
 337         &COMPACT_LOOP("_shrd");
 338         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 339 &function_end_A();
 340                                                 }
 341
 342 &set_label("K256",64);  # Yes! I keep it in the code segment!
 343 @K256=( 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5,
 344         0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5,
 345         0xd807aa98,0x12835b01,0x243185be,0x550c7dc3,
 346         0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174,
 347         0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc,
 348         0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da,
 349         0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7,
 350         0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967,
 351         0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13,
 352         0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85,
 353         0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3,
 354         0xd192e819,0xd6990624,0xf40e3585,0x106aa070,
 355         0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5,
 356         0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3,
 357         0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208,
 358         0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2     );
 359 &data_word(@K256);
 360 &data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f);        # byte swap mask
 361 &asciz("SHA256 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
 362
 363 ($a,$b,$c,$d,$e,$f,$g,$h)=(0..7);       # offsets
 364 sub off { &DWP(4*(((shift)-$i)&7),"esp"); }
 365
 366 if (!$i386 && $unroll_after) {
 367 my @AH=($A,$K256);
 368
 369 &set_label("unrolled",16);
 370         &lea    ("esp",&DWP(-96,"esp"));
 371         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 372         &mov    ($AH[0],&DWP(0,"esi"));
 373         &mov    ($AH[1],&DWP(4,"esi"));
 374         &mov    ("ecx",&DWP(8,"esi"));
 375         &mov    ("ebx",&DWP(12,"esi"));
 376         #&mov   (&DWP(0,"esp"),$AH[0]);
 377         &mov    (&DWP(4,"esp"),$AH[1]);
 378         &xor    ($AH[1],"ecx");         # magic
 379         &mov    (&DWP(8,"esp"),"ecx");
 380         &mov    (&DWP(12,"esp"),"ebx");
 381         &mov    ($E,&DWP(16,"esi"));
 382         &mov    ("ebx",&DWP(20,"esi"));
 383         &mov    ("ecx",&DWP(24,"esi"));
 384         &mov    ("esi",&DWP(28,"esi"));
 385         #&mov   (&DWP(16,"esp"),$E);
 386         &mov    (&DWP(20,"esp"),"ebx");
 387         &mov    (&DWP(24,"esp"),"ecx");
 388         &mov    (&DWP(28,"esp"),"esi");
 389         &jmp    (&label("grand_loop"));
 390
 391 &set_label("grand_loop",16);
 392     # copy input block to stack reversing byte order
 393     for($i=0;$i<5;$i++) {
 394         &mov    ("ebx",&DWP(12*$i+0,"edi"));
 395         &mov    ("ecx",&DWP(12*$i+4,"edi"));
 396         &bswap  ("ebx");
 397         &mov    ("esi",&DWP(12*$i+8,"edi"));
 398         &bswap  ("ecx");
 399         &mov    (&DWP(32+12*$i+0,"esp"),"ebx");
 400         &bswap  ("esi");
 401         &mov    (&DWP(32+12*$i+4,"esp"),"ecx");
 402         &mov    (&DWP(32+12*$i+8,"esp"),"esi");
 403     }
 404         &mov    ("ebx",&DWP($i*12,"edi"));
 405         &add    ("edi",64);
 406         &bswap  ("ebx");
 407         &mov    (&DWP(96+4,"esp"),"edi");
 408         &mov    (&DWP(32+12*$i,"esp"),"ebx");
 409
 410     my ($t1,$t2) = ("ecx","esi");
 411
 412     for ($i=0;$i<64;$i++) {
 413
 414       if ($i>=16) {
 415         &mov    ($T,$t1);                       # $t1 is preloaded
 416         # &mov  ($t2,&DWP(32+4*(($i+14)&15),"esp"));
 417         &ror    ($t1,18-7);
 418          &mov   ("edi",$t2);
 419         &ror    ($t2,19-17);
 420          &xor   ($t1,$T);
 421          &shr   ($T,3);
 422         &ror    ($t1,7);
 423          &xor   ($t2,"edi");
 424          &xor   ($T,$t1);                       # T = sigma0(X[-15])
 425         &ror    ($t2,17);
 426          &add   ($T,&DWP(32+4*($i&15),"esp"));  # T += X[-16]
 427         &shr    ("edi",10);
 428          &add   ($T,&DWP(32+4*(($i+9)&15),"esp"));      # T += X[-7]
 429         #&xor   ("edi",$t2)                     # sigma1(X[-2])
 430         # &add  ($T,"edi");                     # T += sigma1(X[-2])
 431         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
 432       }
 433         &mov    ($t1,$E);
 434          &xor   ("edi",$t2)                     if ($i>=16);    # sigma1(X[-2])
 435          &mov   ($t2,&off($f));
 436         &ror    ($E,25-11);
 437          &add   ($T,"edi")                      if ($i>=16);    # T += sigma1(X[-2])
 438          &mov   ("edi",&off($g));
 439         &xor    ($E,$t1);
 440          &mov   ($T,&DWP(32+4*($i&15),"esp"))   if ($i<16);     # X[i]
 441          &mov   (&DWP(32+4*($i&15),"esp"),$T)   if ($i>=16 && $i<62);   # save X[0]
 442          &xor   ($t2,"edi");
 443         &ror    ($E,11-6);
 444          &and   ($t2,$t1);
 445          &mov   (&off($e),$t1);         # save $E, modulo-scheduled
 446         &xor    ($E,$t1);
 447          &add   ($T,&off($h));          # T += h
 448          &xor   ("edi",$t2);            # Ch(e,f,g)
 449         &ror    ($E,6);                 # Sigma1(e)
 450          &mov   ($t1,$AH[0]);
 451          &add   ($T,"edi");             # T += Ch(e,f,g)
 452
 453         &ror    ($t1,22-13);
 454          &mov   ($t2,$AH[0]);
 455          &mov   ("edi",&off($b));
 456         &xor    ($t1,$AH[0]);
 457          &mov   (&off($a),$AH[0]);      # save $A, modulo-scheduled
 458          &xor   ($AH[0],"edi");         # a ^= b, (b^c) in next round
 459         &ror    ($t1,13-2);
 460          &and   ($AH[1],$AH[0]);        # (b^c) &= (a^b)
 461          &lea   ($E,&DWP(@K256[$i],$T,$E));     # T += Sigma1(1)+K[i]
 462         &xor    ($t1,$t2);
 463          &xor   ($AH[1],"edi");         # h = Maj(a,b,c) = Ch(a^b,c,b)
 464          &mov   ($t2,&DWP(32+4*(($i+2)&15),"esp"))      if ($i>=15 && $i<63);
 465         &ror    ($t1,2);                # Sigma0(a)
 466
 467          &add   ($AH[1],$E);            # h += T
 468          &add   ($E,&off($d));          # d += T
 469         &add    ($AH[1],$t1);           # h += Sigma0(a)
 470          &mov   ($t1,&DWP(32+4*(($i+15)&15),"esp"))     if ($i>=15 && $i<63);
 471
 472         @AH = reverse(@AH);             # rotate(a,h)
 473         ($t1,$t2) = ($t2,$t1);          # rotate(t1,t2)
 474     }
 475         &mov    ("esi",&DWP(96,"esp")); #ctx
 476                                         #&mov   ($AH[0],&DWP(0,"esp"));
 477         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
 478                                         #&mov   ("edi", &DWP(8,"esp"));
 479         &mov    ("ecx",&DWP(12,"esp"));
 480         &add    ($AH[0],&DWP(0,"esi"));
 481         &add    ($AH[1],&DWP(4,"esi"));
 482         &add    ("edi",&DWP(8,"esi"));
 483         &add    ("ecx",&DWP(12,"esi"));
 484         &mov    (&DWP(0,"esi"),$AH[0]);
 485         &mov    (&DWP(4,"esi"),$AH[1]);
 486         &mov    (&DWP(8,"esi"),"edi");
 487         &mov    (&DWP(12,"esi"),"ecx");
 488          #&mov  (&DWP(0,"esp"),$AH[0]);
 489          &mov   (&DWP(4,"esp"),$AH[1]);
 490          &xor   ($AH[1],"edi");         # magic
 491          &mov   (&DWP(8,"esp"),"edi");
 492          &mov   (&DWP(12,"esp"),"ecx");
 493         #&mov   ($E,&DWP(16,"esp"));
 494         &mov    ("edi",&DWP(20,"esp"));
 495         &mov    ("ebx",&DWP(24,"esp"));
 496         &mov    ("ecx",&DWP(28,"esp"));
 497         &add    ($E,&DWP(16,"esi"));
 498         &add    ("edi",&DWP(20,"esi"));
 499         &add    ("ebx",&DWP(24,"esi"));
 500         &add    ("ecx",&DWP(28,"esi"));
 501         &mov    (&DWP(16,"esi"),$E);
 502         &mov    (&DWP(20,"esi"),"edi");
 503         &mov    (&DWP(24,"esi"),"ebx");
 504         &mov    (&DWP(28,"esi"),"ecx");
 505          #&mov  (&DWP(16,"esp"),$E);
 506          &mov   (&DWP(20,"esp"),"edi");
 507         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 508          &mov   (&DWP(24,"esp"),"ebx");
 509          &mov   (&DWP(28,"esp"),"ecx");
 510
 511         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
 512         &jb     (&label("grand_loop"));
 513
 514         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
 515 &function_end_A();
 516 }
 517                                                 if (!$i386 && $xmm) {{{
 518 if ($shaext) {
 519 ######################################################################
 520 # Intel SHA Extensions implementation of SHA256 update function.
 521 #
 522 my ($ctx,$inp,$end)=("esi","edi","eax");
 523 my ($Wi,$ABEF,$CDGH,$TMP)=map("xmm$_",(0..2,7));
 524 my @MSG=map("xmm$_",(3..6));
 525
 526 sub sha256op38 {
 527  my ($opcodelet,$dst,$src)=@_;
 528     if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
 529     {   &data_byte(0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2);       }
 530 }
 531 sub sha256rnds2 { sha256op38(0xcb,@_); }
 532 sub sha256msg1  { sha256op38(0xcc,@_); }
 533 sub sha256msg2  { sha256op38(0xcd,@_); }
 534
 535 &set_label("shaext",32);
 536         &sub            ("esp",32);
 537
 538         &movdqu         ($ABEF,&QWP(0,$ctx));           # DCBA
 539         &lea            ($K256,&DWP(0x80,$K256));
 540         &movdqu         ($CDGH,&QWP(16,$ctx));          # HGFE
 541         &movdqa         ($TMP,&QWP(0x100-0x80,$K256));  # byte swap mask
 542
 543         &pshufd         ($Wi,$ABEF,0x1b);               # ABCD
 544         &pshufd         ($ABEF,$ABEF,0xb1);             # CDAB
 545         &pshufd         ($CDGH,$CDGH,0x1b);             # EFGH
 546         &palignr        ($ABEF,$CDGH,8);                # ABEF
 547         &punpcklqdq     ($CDGH,$Wi);                    # CDGH
 548         &jmp            (&label("loop_shaext"));
 549
 550 &set_label("loop_shaext",16);
 551         &movdqu         (@MSG[0],&QWP(0,$inp));
 552         &movdqu         (@MSG[1],&QWP(0x10,$inp));
 553         &movdqu         (@MSG[2],&QWP(0x20,$inp));
 554         &pshufb         (@MSG[0],$TMP);
 555         &movdqu         (@MSG[3],&QWP(0x30,$inp));
 556         &movdqa         (&QWP(16,"esp"),$CDGH);         # offload
 557
 558         &movdqa         ($Wi,&QWP(0*16-0x80,$K256));
 559         &paddd          ($Wi,@MSG[0]);
 560         &pshufb         (@MSG[1],$TMP);
 561         &sha256rnds2    ($CDGH,$ABEF);                  # 0-3
 562         &pshufd         ($Wi,$Wi,0x0e);
 563         &nop            ();
 564         &movdqa         (&QWP(0,"esp"),$ABEF);          # offload
 565         &sha256rnds2    ($ABEF,$CDGH);
 566
 567         &movdqa         ($Wi,&QWP(1*16-0x80,$K256));
 568         &paddd          ($Wi,@MSG[1]);
 569         &pshufb         (@MSG[2],$TMP);
 570         &sha256rnds2    ($CDGH,$ABEF);                  # 4-7
 571         &pshufd         ($Wi,$Wi,0x0e);
 572         &lea            ($inp,&DWP(0x40,$inp));
 573         &sha256msg1     (@MSG[0],@MSG[1]);
 574         &sha256rnds2    ($ABEF,$CDGH);
 575
 576         &movdqa         ($Wi,&QWP(2*16-0x80,$K256));
 577         &paddd          ($Wi,@MSG[2]);
 578         &pshufb         (@MSG[3],$TMP);
 579         &sha256rnds2    ($CDGH,$ABEF);                  # 8-11
 580         &pshufd         ($Wi,$Wi,0x0e);
 581         &movdqa         ($TMP,@MSG[3]);
 582         &palignr        ($TMP,@MSG[2],4);
 583         &nop            ();
 584         &paddd          (@MSG[0],$TMP);
 585         &sha256msg1     (@MSG[1],@MSG[2]);
 586         &sha256rnds2    ($ABEF,$CDGH);
 587
 588         &movdqa         ($Wi,&QWP(3*16-0x80,$K256));
 589         &paddd          ($Wi,@MSG[3]);
 590         &sha256msg2     (@MSG[0],@MSG[3]);
 591         &sha256rnds2    ($CDGH,$ABEF);                  # 12-15
 592         &pshufd         ($Wi,$Wi,0x0e);
 593         &movdqa         ($TMP,@MSG[0]);
 594         &palignr        ($TMP,@MSG[3],4);
 595         &nop            ();
 596         &paddd          (@MSG[1],$TMP);
 597         &sha256msg1     (@MSG[2],@MSG[3]);
 598         &sha256rnds2    ($ABEF,$CDGH);
 599
 600 for($i=4;$i<16-3;$i++) {
 601         &movdqa         ($Wi,&QWP($i*16-0x80,$K256));
 602         &paddd          ($Wi,@MSG[0]);
 603         &sha256msg2     (@MSG[1],@MSG[0]);
 604         &sha256rnds2    ($CDGH,$ABEF);                  # 16-19...
 605         &pshufd         ($Wi,$Wi,0x0e);
 606         &movdqa         ($TMP,@MSG[1]);
 607         &palignr        ($TMP,@MSG[0],4);
 608         &nop            ();
 609         &paddd          (@MSG[2],$TMP);
 610         &sha256msg1     (@MSG[3],@MSG[0]);
 611         &sha256rnds2    ($ABEF,$CDGH);
 612
 613         push(@MSG,shift(@MSG));
 614 }
 615         &movdqa         ($Wi,&QWP(13*16-0x80,$K256));
 616         &paddd          ($Wi,@MSG[0]);
 617         &sha256msg2     (@MSG[1],@MSG[0]);
 618         &sha256rnds2    ($CDGH,$ABEF);                  # 52-55
 619         &pshufd         ($Wi,$Wi,0x0e);
 620         &movdqa         ($TMP,@MSG[1])
 621         &palignr        ($TMP,@MSG[0],4);
 622         &sha256rnds2    ($ABEF,$CDGH);
 623         &paddd          (@MSG[2],$TMP);
 624
 625         &movdqa         ($Wi,&QWP(14*16-0x80,$K256));
 626         &paddd          ($Wi,@MSG[1]);
 627         &sha256rnds2    ($CDGH,$ABEF);                  # 56-59
 628         &pshufd         ($Wi,$Wi,0x0e);
 629         &sha256msg2     (@MSG[2],@MSG[1]);
 630         &movdqa         ($TMP,&QWP(0x100-0x80,$K256));  # byte swap mask
 631         &sha256rnds2    ($ABEF,$CDGH);
 632
 633         &movdqa         ($Wi,&QWP(15*16-0x80,$K256));
 634         &paddd          ($Wi,@MSG[2]);
 635         &nop            ();
 636         &sha256rnds2    ($CDGH,$ABEF);                  # 60-63
 637         &pshufd         ($Wi,$Wi,0x0e);
 638         &cmp            ($end,$inp);
 639         &nop            ();
 640         &sha256rnds2    ($ABEF,$CDGH);
 641
 642         &paddd          ($CDGH,&QWP(16,"esp"));
 643         &paddd          ($ABEF,&QWP(0,"esp"));
 644         &jnz            (&label("loop_shaext"));
 645
 646         &pshufd         ($CDGH,$CDGH,0xb1);             # DCHG
 647         &pshufd         ($TMP,$ABEF,0x1b);              # FEBA
 648         &pshufd         ($ABEF,$ABEF,0xb1);             # BAFE
 649         &punpckhqdq     ($ABEF,$CDGH);                  # DCBA
 650         &palignr        ($CDGH,$TMP,8);                 # HGFE
 651
 652         &mov            ("esp",&DWP(32+12,"esp"));
 653         &movdqu         (&QWP(0,$ctx),$ABEF);
 654         &movdqu         (&QWP(16,$ctx),$CDGH);
 655 &function_end_A();
 656 }
 657
 658 my @X = map("xmm$_",(0..3));
 659 my ($t0,$t1,$t2,$t3) = map("xmm$_",(4..7));
 660 my @AH = ($A,$T);
 661
 662 &set_label("SSSE3",32);
 663         &lea    ("esp",&DWP(-96,"esp"));
 664         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 665         &mov    ($AH[0],&DWP(0,"esi"));
 666         &mov    ($AH[1],&DWP(4,"esi"));
 667         &mov    ("ecx",&DWP(8,"esi"));
 668         &mov    ("edi",&DWP(12,"esi"));
 669         #&mov   (&DWP(0,"esp"),$AH[0]);
 670         &mov    (&DWP(4,"esp"),$AH[1]);
 671         &xor    ($AH[1],"ecx");                 # magic
 672         &mov    (&DWP(8,"esp"),"ecx");
 673         &mov    (&DWP(12,"esp"),"edi");
 674         &mov    ($E,&DWP(16,"esi"));
 675         &mov    ("edi",&DWP(20,"esi"));
 676         &mov    ("ecx",&DWP(24,"esi"));
 677         &mov    ("esi",&DWP(28,"esi"));
 678         #&mov   (&DWP(16,"esp"),$E);
 679         &mov    (&DWP(20,"esp"),"edi");
 680         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 681         &mov    (&DWP(24,"esp"),"ecx");
 682         &mov    (&DWP(28,"esp"),"esi");
 683         &movdqa ($t3,&QWP(256,$K256));
 684         &jmp    (&label("grand_ssse3"));
 685
 686 &set_label("grand_ssse3",16);
 687         # load input, reverse byte order, add K256[0..15], save to stack
 688         &movdqu (@X[0],&QWP(0,"edi"));
 689         &movdqu (@X[1],&QWP(16,"edi"));
 690         &movdqu (@X[2],&QWP(32,"edi"));
 691         &movdqu (@X[3],&QWP(48,"edi"));
 692         &add    ("edi",64);
 693         &pshufb (@X[0],$t3);
 694         &mov    (&DWP(96+4,"esp"),"edi");
 695         &pshufb (@X[1],$t3);
 696         &movdqa ($t0,&QWP(0,$K256));
 697         &pshufb (@X[2],$t3);
 698         &movdqa ($t1,&QWP(16,$K256));
 699         &paddd  ($t0,@X[0]);
 700         &pshufb (@X[3],$t3);
 701         &movdqa ($t2,&QWP(32,$K256));
 702         &paddd  ($t1,@X[1]);
 703         &movdqa ($t3,&QWP(48,$K256));
 704         &movdqa (&QWP(32+0,"esp"),$t0);
 705         &paddd  ($t2,@X[2]);
 706         &movdqa (&QWP(32+16,"esp"),$t1);
 707         &paddd  ($t3,@X[3]);
 708         &movdqa (&QWP(32+32,"esp"),$t2);
 709         &movdqa (&QWP(32+48,"esp"),$t3);
 710         &jmp    (&label("ssse3_00_47"));
 711
 712 &set_label("ssse3_00_47",16);
 713         &add            ($K256,64);
 714
 715 sub SSSE3_00_47 () {
 716 my $j = shift;
 717 my $body = shift;
 718 my @X = @_;
 719 my @insns = (&$body,&$body,&$body,&$body);      # 120 instructions
 720
 721           eval(shift(@insns));
 722         &movdqa         ($t0,@X[1]);
 723           eval(shift(@insns));                  # @
 724           eval(shift(@insns));
 725         &movdqa         ($t3,@X[3]);
 726           eval(shift(@insns));
 727           eval(shift(@insns));
 728         &palignr        ($t0,@X[0],4);          # X[1..4]
 729           eval(shift(@insns));
 730           eval(shift(@insns));                  # @
 731           eval(shift(@insns));
 732          &palignr       ($t3,@X[2],4);          # X[9..12]
 733           eval(shift(@insns));
 734           eval(shift(@insns));
 735           eval(shift(@insns));
 736         &movdqa         ($t1,$t0);
 737           eval(shift(@insns));                  # @
 738           eval(shift(@insns));
 739         &movdqa         ($t2,$t0);
 740           eval(shift(@insns));
 741           eval(shift(@insns));
 742         &psrld          ($t0,3);
 743           eval(shift(@insns));
 744           eval(shift(@insns));                  # @
 745          &paddd         (@X[0],$t3);            # X[0..3] += X[9..12]
 746           eval(shift(@insns));
 747           eval(shift(@insns));
 748         &psrld          ($t2,7);
 749           eval(shift(@insns));
 750           eval(shift(@insns));
 751           eval(shift(@insns));                  # @
 752           eval(shift(@insns));
 753          &pshufd        ($t3,@X[3],0b11111010); # X[14..15]
 754           eval(shift(@insns));
 755           eval(shift(@insns));
 756         &pslld          ($t1,32-18);
 757           eval(shift(@insns));
 758           eval(shift(@insns));                  # @
 759         &pxor           ($t0,$t2);
 760           eval(shift(@insns));
 761           eval(shift(@insns));
 762         &psrld          ($t2,18-7);
 763           eval(shift(@insns));
 764           eval(shift(@insns));
 765           eval(shift(@insns));                  # @
 766         &pxor           ($t0,$t1);
 767           eval(shift(@insns));
 768           eval(shift(@insns));
 769         &pslld          ($t1,18-7);
 770           eval(shift(@insns));
 771           eval(shift(@insns));
 772           eval(shift(@insns));                  # @
 773         &pxor           ($t0,$t2);
 774           eval(shift(@insns));
 775           eval(shift(@insns));
 776          &movdqa        ($t2,$t3);
 777           eval(shift(@insns));
 778           eval(shift(@insns));
 779           eval(shift(@insns));                  # @
 780         &pxor           ($t0,$t1);              # sigma0(X[1..4])
 781           eval(shift(@insns));
 782           eval(shift(@insns));
 783          &psrld         ($t3,10);
 784           eval(shift(@insns));
 785           eval(shift(@insns));
 786           eval(shift(@insns));                  # @
 787         &paddd          (@X[0],$t0);            # X[0..3] += sigma0(X[1..4])
 788           eval(shift(@insns));
 789           eval(shift(@insns));
 790          &psrlq         ($t2,17);
 791           eval(shift(@insns));
 792           eval(shift(@insns));
 793           eval(shift(@insns));                  # @
 794          &pxor          ($t3,$t2);
 795           eval(shift(@insns));
 796           eval(shift(@insns));
 797          &psrlq         ($t2,19-17);
 798           eval(shift(@insns));
 799           eval(shift(@insns));
 800           eval(shift(@insns));                  # @
 801          &pxor          ($t3,$t2);
 802           eval(shift(@insns));
 803           eval(shift(@insns));
 804          &pshufd        ($t3,$t3,0b10000000);
 805           eval(shift(@insns));
 806           eval(shift(@insns));
 807           eval(shift(@insns));                  # @
 808           eval(shift(@insns));
 809           eval(shift(@insns));
 810           eval(shift(@insns));
 811           eval(shift(@insns));
 812           eval(shift(@insns));                  # @
 813           eval(shift(@insns));
 814          &psrldq        ($t3,8);
 815           eval(shift(@insns));
 816           eval(shift(@insns));
 817           eval(shift(@insns));
 818         &paddd          (@X[0],$t3);            # X[0..1] += sigma1(X[14..15])
 819           eval(shift(@insns));                  # @
 820           eval(shift(@insns));
 821           eval(shift(@insns));
 822           eval(shift(@insns));
 823           eval(shift(@insns));
 824           eval(shift(@insns));                  # @
 825           eval(shift(@insns));
 826          &pshufd        ($t3,@X[0],0b01010000); # X[16..17]
 827           eval(shift(@insns));
 828           eval(shift(@insns));
 829           eval(shift(@insns));
 830          &movdqa        ($t2,$t3);
 831           eval(shift(@insns));                  # @
 832          &psrld         ($t3,10);
 833           eval(shift(@insns));
 834          &psrlq         ($t2,17);
 835           eval(shift(@insns));
 836           eval(shift(@insns));
 837           eval(shift(@insns));
 838           eval(shift(@insns));                  # @
 839          &pxor          ($t3,$t2);
 840           eval(shift(@insns));
 841           eval(shift(@insns));
 842          &psrlq         ($t2,19-17);
 843           eval(shift(@insns));
 844           eval(shift(@insns));
 845           eval(shift(@insns));                  # @
 846          &pxor          ($t3,$t2);
 847           eval(shift(@insns));
 848           eval(shift(@insns));
 849           eval(shift(@insns));
 850          &pshufd        ($t3,$t3,0b00001000);
 851           eval(shift(@insns));
 852           eval(shift(@insns));                  # @
 853         &movdqa         ($t2,&QWP(16*$j,$K256));
 854           eval(shift(@insns));
 855           eval(shift(@insns));
 856          &pslldq        ($t3,8);
 857           eval(shift(@insns));
 858           eval(shift(@insns));
 859           eval(shift(@insns));                  # @
 860           eval(shift(@insns));
 861           eval(shift(@insns));
 862           eval(shift(@insns));
 863           eval(shift(@insns));
 864           eval(shift(@insns));                  # @
 865         &paddd          (@X[0],$t3);            # X[2..3] += sigma1(X[16..17])
 866           eval(shift(@insns));
 867           eval(shift(@insns));
 868           eval(shift(@insns));
 869           eval(shift(@insns));
 870         &paddd          ($t2,@X[0]);
 871           eval(shift(@insns));                  # @
 872
 873         foreach (@insns) { eval; }              # remaining instructions
 874
 875         &movdqa         (&QWP(32+16*$j,"esp"),$t2);
 876 }
 877
 878 sub body_00_15 () {
 879         (
 880         '&mov   ("ecx",$E);',
 881         '&ror   ($E,25-11);',
 882          '&mov  ("esi",&off($f));',
 883         '&xor   ($E,"ecx");',
 884          '&mov  ("edi",&off($g));',
 885          '&xor  ("esi","edi");',
 886         '&ror   ($E,11-6);',
 887          '&and  ("esi","ecx");',
 888          '&mov  (&off($e),"ecx");',     # save $E, modulo-scheduled
 889         '&xor   ($E,"ecx");',
 890          '&xor  ("edi","esi");',        # Ch(e,f,g)
 891         '&ror   ($E,6);',               # T = Sigma1(e)
 892          '&mov  ("ecx",$AH[0]);',
 893          '&add  ($E,"edi");',           # T += Ch(e,f,g)
 894          '&mov  ("edi",&off($b));',
 895         '&mov   ("esi",$AH[0]);',
 896
 897         '&ror   ("ecx",22-13);',
 898          '&mov  (&off($a),$AH[0]);',    # save $A, modulo-scheduled
 899         '&xor   ("ecx",$AH[0]);',
 900          '&xor  ($AH[0],"edi");',       # a ^= b, (b^c) in next round
 901          '&add  ($E,&off($h));',        # T += h
 902         '&ror   ("ecx",13-2);',
 903          '&and  ($AH[1],$AH[0]);',      # (b^c) &= (a^b)
 904         '&xor   ("ecx","esi");',
 905          '&add  ($E,&DWP(32+4*($i&15),"esp"));',        # T += K[i]+X[i]
 906          '&xor  ($AH[1],"edi");',       # h = Maj(a,b,c) = Ch(a^b,c,b)
 907         '&ror   ("ecx",2);',            # Sigma0(a)
 908
 909          '&add  ($AH[1],$E);',          # h += T
 910          '&add  ($E,&off($d));',        # d += T
 911         '&add   ($AH[1],"ecx");'.       # h += Sigma0(a)
 912
 913         '@AH = reverse(@AH); $i++;'     # rotate(a,h)
 914         );
 915 }
 916
 917     for ($i=0,$j=0; $j<4; $j++) {
 918         &SSSE3_00_47($j,\&body_00_15,@X);
 919         push(@X,shift(@X));             # rotate(@X)
 920     }
 921         &cmp    (&DWP(16*$j,$K256),0x00010203);
 922         &jne    (&label("ssse3_00_47"));
 923
 924     for ($i=0; $i<16; ) {
 925         foreach(body_00_15()) { eval; }
 926     }
 927
 928         &mov    ("esi",&DWP(96,"esp")); #ctx
 929                                         #&mov   ($AH[0],&DWP(0,"esp"));
 930         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
 931                                         #&mov   ("edi", &DWP(8,"esp"));
 932         &mov    ("ecx",&DWP(12,"esp"));
 933         &add    ($AH[0],&DWP(0,"esi"));
 934         &add    ($AH[1],&DWP(4,"esi"));
 935         &add    ("edi",&DWP(8,"esi"));
 936         &add    ("ecx",&DWP(12,"esi"));
 937         &mov    (&DWP(0,"esi"),$AH[0]);
 938         &mov    (&DWP(4,"esi"),$AH[1]);
 939         &mov    (&DWP(8,"esi"),"edi");
 940         &mov    (&DWP(12,"esi"),"ecx");
 941          #&mov  (&DWP(0,"esp"),$AH[0]);
 942          &mov   (&DWP(4,"esp"),$AH[1]);
 943          &xor   ($AH[1],"edi");                 # magic
 944          &mov   (&DWP(8,"esp"),"edi");
 945          &mov   (&DWP(12,"esp"),"ecx");
 946         #&mov   ($E,&DWP(16,"esp"));
 947         &mov    ("edi",&DWP(20,"esp"));
 948         &mov    ("ecx",&DWP(24,"esp"));
 949         &add    ($E,&DWP(16,"esi"));
 950         &add    ("edi",&DWP(20,"esi"));
 951         &add    ("ecx",&DWP(24,"esi"));
 952         &mov    (&DWP(16,"esi"),$E);
 953         &mov    (&DWP(20,"esi"),"edi");
 954          &mov   (&DWP(20,"esp"),"edi");
 955         &mov    ("edi",&DWP(28,"esp"));
 956         &mov    (&DWP(24,"esi"),"ecx");
 957          #&mov  (&DWP(16,"esp"),$E);
 958         &add    ("edi",&DWP(28,"esi"));
 959          &mov   (&DWP(24,"esp"),"ecx");
 960         &mov    (&DWP(28,"esi"),"edi");
 961          &mov   (&DWP(28,"esp"),"edi");
 962         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 963
 964         &movdqa ($t3,&QWP(64,$K256));
 965         &sub    ($K256,3*64);                   # rewind K
 966         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
 967         &jb     (&label("grand_ssse3"));
 968
 969         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
 970 &function_end_A();
 971                                                 if ($avx) {
 972 &set_label("AVX",32);
 973                                                 if ($avx>1) {
 974         &and    ("edx",1<<8|1<<3);              # check for BMI2+BMI1
 975         &cmp    ("edx",1<<8|1<<3);
 976         &je     (&label("AVX_BMI"));
 977                                                 }
 978         &lea    ("esp",&DWP(-96,"esp"));
 979         &vzeroall       ();
 980         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 981         &mov    ($AH[0],&DWP(0,"esi"));
 982         &mov    ($AH[1],&DWP(4,"esi"));
 983         &mov    ("ecx",&DWP(8,"esi"));
 984         &mov    ("edi",&DWP(12,"esi"));
 985         #&mov   (&DWP(0,"esp"),$AH[0]);
 986         &mov    (&DWP(4,"esp"),$AH[1]);
 987         &xor    ($AH[1],"ecx");                 # magic
 988         &mov    (&DWP(8,"esp"),"ecx");
 989         &mov    (&DWP(12,"esp"),"edi");
 990         &mov    ($E,&DWP(16,"esi"));
 991         &mov    ("edi",&DWP(20,"esi"));
 992         &mov    ("ecx",&DWP(24,"esi"));
 993         &mov    ("esi",&DWP(28,"esi"));
 994         #&mov   (&DWP(16,"esp"),$E);
 995         &mov    (&DWP(20,"esp"),"edi");
 996         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 997         &mov    (&DWP(24,"esp"),"ecx");
 998         &mov    (&DWP(28,"esp"),"esi");
 999         &vmovdqa        ($t3,&QWP(256,$K256));
1000         &jmp    (&label("grand_avx"));
1001
1002 &set_label("grand_avx",32);
1003         # load input, reverse byte order, add K256[0..15], save to stack
1004         &vmovdqu        (@X[0],&QWP(0,"edi"));
1005         &vmovdqu        (@X[1],&QWP(16,"edi"));
1006         &vmovdqu        (@X[2],&QWP(32,"edi"));
1007         &vmovdqu        (@X[3],&QWP(48,"edi"));
1008         &add            ("edi",64);
1009         &vpshufb        (@X[0],@X[0],$t3);
1010         &mov            (&DWP(96+4,"esp"),"edi");
1011         &vpshufb        (@X[1],@X[1],$t3);
1012         &vpshufb        (@X[2],@X[2],$t3);
1013         &vpaddd         ($t0,@X[0],&QWP(0,$K256));
1014         &vpshufb        (@X[3],@X[3],$t3);
1015         &vpaddd         ($t1,@X[1],&QWP(16,$K256));
1016         &vpaddd         ($t2,@X[2],&QWP(32,$K256));
1017         &vpaddd         ($t3,@X[3],&QWP(48,$K256));
1018         &vmovdqa        (&QWP(32+0,"esp"),$t0);
1019         &vmovdqa        (&QWP(32+16,"esp"),$t1);
1020         &vmovdqa        (&QWP(32+32,"esp"),$t2);
1021         &vmovdqa        (&QWP(32+48,"esp"),$t3);
1022         &jmp            (&label("avx_00_47"));
1023
1024 &set_label("avx_00_47",16);
1025         &add            ($K256,64);
1026
1027 sub Xupdate_AVX () {
1028         (
1029         '&vpalignr      ($t0,@X[1],@X[0],4);',  # X[1..4]
1030          '&vpalignr     ($t3,@X[3],@X[2],4);',  # X[9..12]
1031         '&vpsrld        ($t2,$t0,7);',
1032          '&vpaddd       (@X[0],@X[0],$t3);',    # X[0..3] += X[9..16]
1033         '&vpsrld        ($t3,$t0,3);',
1034         '&vpslld        ($t1,$t0,14);',
1035         '&vpxor         ($t0,$t3,$t2);',
1036          '&vpshufd      ($t3,@X[3],0b11111010)',# X[14..15]
1037         '&vpsrld        ($t2,$t2,18-7);',
1038         '&vpxor         ($t0,$t0,$t1);',
1039         '&vpslld        ($t1,$t1,25-14);',
1040         '&vpxor         ($t0,$t0,$t2);',
1041          '&vpsrld       ($t2,$t3,10);',
1042         '&vpxor         ($t0,$t0,$t1);',        # sigma0(X[1..4])
1043          '&vpsrlq       ($t1,$t3,17);',
1044         '&vpaddd        (@X[0],@X[0],$t0);',    # X[0..3] += sigma0(X[1..4])
1045          '&vpxor        ($t2,$t2,$t1);',
1046          '&vpsrlq       ($t3,$t3,19);',
1047          '&vpxor        ($t2,$t2,$t3);',        # sigma1(X[14..15]
1048          '&vpshufd      ($t3,$t2,0b10000100);',
1049         '&vpsrldq       ($t3,$t3,8);',
1050         '&vpaddd        (@X[0],@X[0],$t3);',    # X[0..1] += sigma1(X[14..15])
1051          '&vpshufd      ($t3,@X[0],0b01010000)',# X[16..17]
1052          '&vpsrld       ($t2,$t3,10);',
1053          '&vpsrlq       ($t1,$t3,17);',
1054          '&vpxor        ($t2,$t2,$t1);',
1055          '&vpsrlq       ($t3,$t3,19);',
1056          '&vpxor        ($t2,$t2,$t3);',        # sigma1(X[16..17]
1057          '&vpshufd      ($t3,$t2,0b11101000);',
1058         '&vpslldq       ($t3,$t3,8);',
1059         '&vpaddd        (@X[0],@X[0],$t3);'     # X[2..3] += sigma1(X[16..17])
1060         );
1061 }
1062
1063 local *ror = sub { &shrd(@_[0],@_) };
1064 sub AVX_00_47 () {
1065 my $j = shift;
1066 my $body = shift;
1067 my @X = @_;
1068 my @insns = (&$body,&$body,&$body,&$body);      # 120 instructions
1069 my $insn;
1070
1071         foreach (Xupdate_AVX()) {               # 31 instructions
1072             eval;
1073             eval(shift(@insns));
1074             eval(shift(@insns));
1075             eval($insn = shift(@insns));
1076             eval(shift(@insns)) if ($insn =~ /rorx/ && @insns[0] =~ /rorx/);
1077         }
1078         &vpaddd         ($t2,@X[0],&QWP(16*$j,$K256));
1079         foreach (@insns) { eval; }              # remaining instructions
1080         &vmovdqa        (&QWP(32+16*$j,"esp"),$t2);
1081 }
1082
1083     for ($i=0,$j=0; $j<4; $j++) {
1084         &AVX_00_47($j,\&body_00_15,@X);
1085         push(@X,shift(@X));             # rotate(@X)
1086     }
1087         &cmp    (&DWP(16*$j,$K256),0x00010203);
1088         &jne    (&label("avx_00_47"));
1089
1090     for ($i=0; $i<16; ) {
1091         foreach(body_00_15()) { eval; }
1092     }
1093
1094         &mov    ("esi",&DWP(96,"esp")); #ctx
1095                                         #&mov   ($AH[0],&DWP(0,"esp"));
1096         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
1097                                         #&mov   ("edi", &DWP(8,"esp"));
1098         &mov    ("ecx",&DWP(12,"esp"));
1099         &add    ($AH[0],&DWP(0,"esi"));
1100         &add    ($AH[1],&DWP(4,"esi"));
1101         &add    ("edi",&DWP(8,"esi"));
1102         &add    ("ecx",&DWP(12,"esi"));
1103         &mov    (&DWP(0,"esi"),$AH[0]);
1104         &mov    (&DWP(4,"esi"),$AH[1]);
1105         &mov    (&DWP(8,"esi"),"edi");
1106         &mov    (&DWP(12,"esi"),"ecx");
1107          #&mov  (&DWP(0,"esp"),$AH[0]);
1108          &mov   (&DWP(4,"esp"),$AH[1]);
1109          &xor   ($AH[1],"edi");                 # magic
1110          &mov   (&DWP(8,"esp"),"edi");
1111          &mov   (&DWP(12,"esp"),"ecx");
1112         #&mov   ($E,&DWP(16,"esp"));
1113         &mov    ("edi",&DWP(20,"esp"));
1114         &mov    ("ecx",&DWP(24,"esp"));
1115         &add    ($E,&DWP(16,"esi"));
1116         &add    ("edi",&DWP(20,"esi"));
1117         &add    ("ecx",&DWP(24,"esi"));
1118         &mov    (&DWP(16,"esi"),$E);
1119         &mov    (&DWP(20,"esi"),"edi");
1120          &mov   (&DWP(20,"esp"),"edi");
1121         &mov    ("edi",&DWP(28,"esp"));
1122         &mov    (&DWP(24,"esi"),"ecx");
1123          #&mov  (&DWP(16,"esp"),$E);
1124         &add    ("edi",&DWP(28,"esi"));
1125          &mov   (&DWP(24,"esp"),"ecx");
1126         &mov    (&DWP(28,"esi"),"edi");
1127          &mov   (&DWP(28,"esp"),"edi");
1128         &mov    ("edi",&DWP(96+4,"esp"));       # inp
1129
1130         &vmovdqa        ($t3,&QWP(64,$K256));
1131         &sub    ($K256,3*64);                   # rewind K
1132         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
1133         &jb     (&label("grand_avx"));
1134
1135         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
1136         &vzeroall       ();
1137 &function_end_A();
1138                                                 if ($avx>1) {
1139 sub bodyx_00_15 () {                    # +10%
1140         (
1141         '&rorx  ("ecx",$E,6)',
1142         '&rorx  ("esi",$E,11)',
1143          '&mov  (&off($e),$E)',         # save $E, modulo-scheduled
1144         '&rorx  ("edi",$E,25)',
1145         '&xor   ("ecx","esi")',
1146          '&andn ("esi",$E,&off($g))',
1147         '&xor   ("ecx","edi")',         # Sigma1(e)
1148          '&and  ($E,&off($f))',
1149          '&mov  (&off($a),$AH[0]);',    # save $A, modulo-scheduled
1150          '&or   ($E,"esi")',            # T = Ch(e,f,g)
1151
1152         '&rorx  ("edi",$AH[0],2)',
1153         '&rorx  ("esi",$AH[0],13)',
1154          '&lea  ($E,&DWP(0,$E,"ecx"))', # T += Sigma1(e)
1155         '&rorx  ("ecx",$AH[0],22)',
1156         '&xor   ("esi","edi")',
1157          '&mov  ("edi",&off($b))',
1158         '&xor   ("ecx","esi")',         # Sigma0(a)
1159
1160          '&xor  ($AH[0],"edi")',        # a ^= b, (b^c) in next round
1161          '&add  ($E,&off($h))',         # T += h
1162          '&and  ($AH[1],$AH[0])',       # (b^c) &= (a^b)
1163          '&add  ($E,&DWP(32+4*($i&15),"esp"))', # T += K[i]+X[i]
1164          '&xor  ($AH[1],"edi")',        # h = Maj(a,b,c) = Ch(a^b,c,b)
1165
1166          '&add  ("ecx",$E)',            # h += T
1167          '&add  ($E,&off($d))',         # d += T
1168         '&lea   ($AH[1],&DWP(0,$AH[1],"ecx"));'.        # h += Sigma0(a)
1169
1170         '@AH = reverse(@AH); $i++;'     # rotate(a,h)
1171         );
1172 }
1173
1174 &set_label("AVX_BMI",32);
1175         &lea    ("esp",&DWP(-96,"esp"));
1176         &vzeroall       ();
1177         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
1178         &mov    ($AH[0],&DWP(0,"esi"));
1179         &mov    ($AH[1],&DWP(4,"esi"));
1180         &mov    ("ecx",&DWP(8,"esi"));
1181         &mov    ("edi",&DWP(12,"esi"));
1182         #&mov   (&DWP(0,"esp"),$AH[0]);
1183         &mov    (&DWP(4,"esp"),$AH[1]);
1184         &xor    ($AH[1],"ecx");                 # magic
1185         &mov    (&DWP(8,"esp"),"ecx");
1186         &mov    (&DWP(12,"esp"),"edi");
1187         &mov    ($E,&DWP(16,"esi"));
1188         &mov    ("edi",&DWP(20,"esi"));
1189         &mov    ("ecx",&DWP(24,"esi"));
1190         &mov    ("esi",&DWP(28,"esi"));
1191         #&mov   (&DWP(16,"esp"),$E);
1192         &mov    (&DWP(20,"esp"),"edi");
1193         &mov    ("edi",&DWP(96+4,"esp"));       # inp
1194         &mov    (&DWP(24,"esp"),"ecx");
1195         &mov    (&DWP(28,"esp"),"esi");
1196         &vmovdqa        ($t3,&QWP(256,$K256));
1197         &jmp    (&label("grand_avx_bmi"));
1198
1199 &set_label("grand_avx_bmi",32);
1200         # load input, reverse byte order, add K256[0..15], save to stack
1201         &vmovdqu        (@X[0],&QWP(0,"edi"));
1202         &vmovdqu        (@X[1],&QWP(16,"edi"));
1203         &vmovdqu        (@X[2],&QWP(32,"edi"));
1204         &vmovdqu        (@X[3],&QWP(48,"edi"));
1205         &add            ("edi",64);
1206         &vpshufb        (@X[0],@X[0],$t3);
1207         &mov            (&DWP(96+4,"esp"),"edi");
1208         &vpshufb        (@X[1],@X[1],$t3);
1209         &vpshufb        (@X[2],@X[2],$t3);
1210         &vpaddd         ($t0,@X[0],&QWP(0,$K256));
1211         &vpshufb        (@X[3],@X[3],$t3);
1212         &vpaddd         ($t1,@X[1],&QWP(16,$K256));
1213         &vpaddd         ($t2,@X[2],&QWP(32,$K256));
1214         &vpaddd         ($t3,@X[3],&QWP(48,$K256));
1215         &vmovdqa        (&QWP(32+0,"esp"),$t0);
1216         &vmovdqa        (&QWP(32+16,"esp"),$t1);
1217         &vmovdqa        (&QWP(32+32,"esp"),$t2);
1218         &vmovdqa        (&QWP(32+48,"esp"),$t3);
1219         &jmp            (&label("avx_bmi_00_47"));
1220
1221 &set_label("avx_bmi_00_47",16);
1222         &add            ($K256,64);
1223
1224     for ($i=0,$j=0; $j<4; $j++) {
1225         &AVX_00_47($j,\&bodyx_00_15,@X);
1226         push(@X,shift(@X));             # rotate(@X)
1227     }
1228         &cmp    (&DWP(16*$j,$K256),0x00010203);
1229         &jne    (&label("avx_bmi_00_47"));
1230
1231     for ($i=0; $i<16; ) {
1232         foreach(bodyx_00_15()) { eval; }
1233     }
1234
1235         &mov    ("esi",&DWP(96,"esp")); #ctx
1236                                         #&mov   ($AH[0],&DWP(0,"esp"));
1237         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
1238                                         #&mov   ("edi", &DWP(8,"esp"));
1239         &mov    ("ecx",&DWP(12,"esp"));
1240         &add    ($AH[0],&DWP(0,"esi"));
1241         &add    ($AH[1],&DWP(4,"esi"));
1242         &add    ("edi",&DWP(8,"esi"));
1243         &add    ("ecx",&DWP(12,"esi"));
1244         &mov    (&DWP(0,"esi"),$AH[0]);
1245         &mov    (&DWP(4,"esi"),$AH[1]);
1246         &mov    (&DWP(8,"esi"),"edi");
1247         &mov    (&DWP(12,"esi"),"ecx");
1248          #&mov  (&DWP(0,"esp"),$AH[0]);
1249          &mov   (&DWP(4,"esp"),$AH[1]);
1250          &xor   ($AH[1],"edi");                 # magic
1251          &mov   (&DWP(8,"esp"),"edi");
1252          &mov   (&DWP(12,"esp"),"ecx");
1253         #&mov   ($E,&DWP(16,"esp"));
1254         &mov    ("edi",&DWP(20,"esp"));
1255         &mov    ("ecx",&DWP(24,"esp"));
1256         &add    ($E,&DWP(16,"esi"));
1257         &add    ("edi",&DWP(20,"esi"));
1258         &add    ("ecx",&DWP(24,"esi"));
1259         &mov    (&DWP(16,"esi"),$E);
1260         &mov    (&DWP(20,"esi"),"edi");
1261          &mov   (&DWP(20,"esp"),"edi");
1262         &mov    ("edi",&DWP(28,"esp"));
1263         &mov    (&DWP(24,"esi"),"ecx");
1264          #&mov  (&DWP(16,"esp"),$E);
1265         &add    ("edi",&DWP(28,"esi"));
1266          &mov   (&DWP(24,"esp"),"ecx");
1267         &mov    (&DWP(28,"esi"),"edi");
1268          &mov   (&DWP(28,"esp"),"edi");
1269         &mov    ("edi",&DWP(96+4,"esp"));       # inp
1270
1271         &vmovdqa        ($t3,&QWP(64,$K256));
1272         &sub    ($K256,3*64);                   # rewind K
1273         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
1274         &jb     (&label("grand_avx_bmi"));
1275
1276         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
1277         &vzeroall       ();
1278 &function_end_A();
1279                                                 }
1280                                                 }
1281                                                 }}}
1282 &function_end_B("sha256_block_data_order");
1283
1284 &asm_finish();
1285
1286 close STDOUT;