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