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 # Performance in clock cycles per processed byte (less is better):
  32 #
  33 #               PIII    P4      AMD K8  Core2   SB      Atom    Bldzr
  34 # gcc           36      41      27      26      25      50      36
  35 # icc           33      38      25      23      -       -       -
  36 # x86 asm(*)    27/24   28      19/15.5 18/15.6 12.5    30/25   16.6
  37 # x86_64 asm(**)        17.5    15.1    13.9    11.6    22      13.7
  38 #
  39 # (*)   numbers after slash are for unrolled loop, where available;
  40 # (**)  x86_64 assembly performance is presented for reference
  41 #       purposes.
  42
  43 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  44 push(@INC,"${dir}","${dir}../../perlasm");
  45 require "x86asm.pl";
  46
  47 &asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");
  48
  49 $xmm=$ymm=0;
  50 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
  51
  52 $ymm=1 if ($xmm &&
  53                 `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
  54                         =~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
  55                 $1>=2.19);      # first version supporting AVX
  56
  57 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32n" &&
  58                 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
  59                 $1>=2.03);      # first version supporting AVX
  60
  61 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32" &&
  62                 `ml 2>&1` =~ /Version ([0-9]+)\./ &&
  63                 $1>=10);        # first version supporting AVX
  64
  65 $unroll_after = 64*4;   # If pre-evicted from L1P cache first spin of
  66                         # fully unrolled loop was measured to run about
  67                         # 3-4x slower. If slowdown coefficient is N and
  68                         # unrolled loop is m times faster, then you break
  69                         # even at (N-1)/(m-1) blocks. Then it needs to be
  70                         # adjusted for probability of code being evicted,
  71                         # code size/cache size=1/4. Typical m is 1.15...
  72
  73 $A="eax";
  74 $E="edx";
  75 $T="ebx";
  76 $Aoff=&DWP(4,"esp");
  77 $Boff=&DWP(8,"esp");
  78 $Coff=&DWP(12,"esp");
  79 $Doff=&DWP(16,"esp");
  80 $Eoff=&DWP(20,"esp");
  81 $Foff=&DWP(24,"esp");
  82 $Goff=&DWP(28,"esp");
  83 $Hoff=&DWP(32,"esp");
  84 $Xoff=&DWP(36,"esp");
  85 $K256="ebp";
  86
  87 sub BODY_16_63() {
  88         &mov    ($T,"ecx");                     # "ecx" is preloaded
  89          &mov   ("esi",&DWP(4*(9+15+16-14),"esp"));
  90         &ror    ("ecx",18-7);
  91          &mov   ("edi","esi");
  92         &ror    ("esi",19-17);
  93          &xor   ("ecx",$T);
  94          &shr   ($T,3);
  95         &ror    ("ecx",7);
  96          &xor   ("esi","edi");
  97          &xor   ($T,"ecx");                     # T = sigma0(X[-15])
  98         &ror    ("esi",17);
  99          &add   ($T,&DWP(4*(9+15+16),"esp"));   # T += X[-16]
 100         &shr    ("edi",10);
 101          &add   ($T,&DWP(4*(9+15+16-9),"esp")); # T += X[-7]
 102         #&xor   ("edi","esi")                   # sigma1(X[-2])
 103         # &add  ($T,"edi");                     # T += sigma1(X[-2])
 104         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
 105
 106         &BODY_00_15(1);
 107 }
 108 sub BODY_00_15() {
 109     my $in_16_63=shift;
 110
 111         &mov    ("ecx",$E);
 112          &xor   ("edi","esi")                   if ($in_16_63); # sigma1(X[-2])
 113          &mov   ("esi",$Foff);
 114         &ror    ("ecx",25-11);
 115          &add   ($T,"edi")                      if ($in_16_63); # T += sigma1(X[-2])
 116          &mov   ("edi",$Goff);
 117         &xor    ("ecx",$E);
 118          &xor   ("esi","edi");
 119          &mov   ($T,&DWP(4*(9+15),"esp"))       if (!$in_16_63);
 120          &mov   (&DWP(4*(9+15),"esp"),$T)       if ($in_16_63); # save X[0]
 121         &ror    ("ecx",11-6);
 122          &and   ("esi",$E);
 123          &mov   ($Eoff,$E);             # modulo-scheduled
 124         &xor    ($E,"ecx");
 125          &add   ($T,$Hoff);             # T += h
 126          &xor   ("esi","edi");          # Ch(e,f,g)
 127         &ror    ($E,6);                 # Sigma1(e)
 128          &mov   ("ecx",$A);
 129          &add   ($T,"esi");             # T += Ch(e,f,g)
 130
 131         &ror    ("ecx",22-13);
 132          &add   ($T,$E);                # T += Sigma1(e)
 133          &mov   ("edi",$Boff);
 134         &xor    ("ecx",$A);
 135          &mov   ($Aoff,$A);             # modulo-scheduled
 136          &lea   ("esp",&DWP(-4,"esp"));
 137         &ror    ("ecx",13-2);
 138          &mov   ("esi",&DWP(0,$K256));
 139         &xor    ("ecx",$A);
 140          &mov   ($E,$Eoff);             # e in next iteration, d in this one
 141          &xor   ($A,"edi");             # a ^= b
 142         &ror    ("ecx",2);              # Sigma0(a)
 143
 144          &add   ($T,"esi");             # T+= K[i]
 145          &mov   (&DWP(0,"esp"),$A);     # (b^c) in next round
 146         &add    ($E,$T);                # d += T
 147          &and   ($A,&DWP(4,"esp"));     # a &= (b^c)
 148         &add    ($T,"ecx");             # T += Sigma0(a)
 149          &xor   ($A,"edi");             # h = Maj(a,b,c) = Ch(a^b,c,b)
 150          &mov   ("ecx",&DWP(4*(9+15+16-1),"esp"))       if ($in_16_63); # preload T
 151         &add    ($K256,4);
 152          &add   ($A,$T);                # h += T
 153 }
 154
 155 &external_label("OPENSSL_ia32cap_P")            if (!$i386);
 156
 157 &function_begin("sha256_block_data_order");
 158         &mov    ("esi",wparam(0));      # ctx
 159         &mov    ("edi",wparam(1));      # inp
 160         &mov    ("eax",wparam(2));      # num
 161         &mov    ("ebx","esp");          # saved sp
 162
 163         &call   (&label("pic_point"));  # make it PIC!
 164 &set_label("pic_point");
 165         &blindpop($K256);
 166         &lea    ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256));
 167
 168         &sub    ("esp",16);
 169         &and    ("esp",-64);
 170
 171         &shl    ("eax",6);
 172         &add    ("eax","edi");
 173         &mov    (&DWP(0,"esp"),"esi");  # ctx
 174         &mov    (&DWP(4,"esp"),"edi");  # inp
 175         &mov    (&DWP(8,"esp"),"eax");  # inp+num*128
 176         &mov    (&DWP(12,"esp"),"ebx"); # saved sp
 177                                                 if (!$i386) {
 178         &picmeup("edx","OPENSSL_ia32cap_P",$K256,&label("K256"));
 179         &mov    ("ecx",&DWP(0,"edx"));
 180         &mov    ("edx",&DWP(4,"edx"));
 181         &test   ("ecx",1<<20);          # check for P4
 182         &jnz    (&label("loop"));
 183         &test   ("edx",1<<11);          # check for XOP
 184         &jnz    (&label("XOP"))         if ($ymm);
 185         &and    ("ecx",1<<30);          # mask "Intel CPU" bit
 186         &and    ("edx",1<<28);          # mask AVX bit
 187         &or     ("ecx","edx");
 188         &cmp    ("ecx",1<<28|1<<30);
 189         &je     (&label("AVX"))         if ($ymm);
 190         &je     (&label("loop_shrd"))   if (!$ymm);
 191                                                 if ($unroll_after) {
 192         &sub    ("eax","edi");
 193         &cmp    ("eax",$unroll_after);
 194         &jae    (&label("unrolled"));
 195                                                 } }
 196         &jmp    (&label("loop"));
 197
 198 sub COMPACT_LOOP() {
 199 my $suffix=shift;
 200
 201 &set_label("loop$suffix",16);
 202     # copy input block to stack reversing byte and dword order
 203     for($i=0;$i<4;$i++) {
 204         &mov    ("eax",&DWP($i*16+0,"edi"));
 205         &mov    ("ebx",&DWP($i*16+4,"edi"));
 206         &mov    ("ecx",&DWP($i*16+8,"edi"));
 207         &bswap  ("eax");
 208         &mov    ("edx",&DWP($i*16+12,"edi"));
 209         &bswap  ("ebx");
 210         &push   ("eax");
 211         &bswap  ("ecx");
 212         &push   ("ebx");
 213         &bswap  ("edx");
 214         &push   ("ecx");
 215         &push   ("edx");
 216     }
 217         &add    ("edi",64);
 218         &lea    ("esp",&DWP(-4*9,"esp"));# place for A,B,C,D,E,F,G,H
 219         &mov    (&DWP(4*(9+16)+4,"esp"),"edi");
 220
 221         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 222         &mov    ($A,&DWP(0,"esi"));
 223         &mov    ("ebx",&DWP(4,"esi"));
 224         &mov    ("ecx",&DWP(8,"esi"));
 225         &mov    ("edi",&DWP(12,"esi"));
 226         # &mov  ($Aoff,$A);
 227         &mov    ($Boff,"ebx");
 228         &xor    ("ebx","ecx");
 229         &mov    ($Coff,"ecx");
 230         &mov    ($Doff,"edi");
 231         &mov    (&DWP(0,"esp"),"ebx");  # magic
 232         &mov    ($E,&DWP(16,"esi"));
 233         &mov    ("ebx",&DWP(20,"esi"));
 234         &mov    ("ecx",&DWP(24,"esi"));
 235         &mov    ("edi",&DWP(28,"esi"));
 236         # &mov  ($Eoff,$E);
 237         &mov    ($Foff,"ebx");
 238         &mov    ($Goff,"ecx");
 239         &mov    ($Hoff,"edi");
 240
 241 &set_label("00_15$suffix",16);
 242
 243         &BODY_00_15();
 244
 245         &cmp    ("esi",0xc19bf174);
 246         &jne    (&label("00_15$suffix"));
 247
 248         &mov    ("ecx",&DWP(4*(9+15+16-1),"esp"));      # preloaded in BODY_00_15(1)
 249         &jmp    (&label("16_63$suffix"));
 250
 251 &set_label("16_63$suffix",16);
 252
 253         &BODY_16_63();
 254
 255         &cmp    ("esi",0xc67178f2);
 256         &jne    (&label("16_63$suffix"));
 257
 258         &mov    ("esi",&DWP(4*(9+16+64)+0,"esp"));#ctx
 259         # &mov  ($A,$Aoff);
 260         &mov    ("ebx",$Boff);
 261         # &mov  ("edi",$Coff);
 262         &mov    ("ecx",$Doff);
 263         &add    ($A,&DWP(0,"esi"));
 264         &add    ("ebx",&DWP(4,"esi"));
 265         &add    ("edi",&DWP(8,"esi"));
 266         &add    ("ecx",&DWP(12,"esi"));
 267         &mov    (&DWP(0,"esi"),$A);
 268         &mov    (&DWP(4,"esi"),"ebx");
 269         &mov    (&DWP(8,"esi"),"edi");
 270         &mov    (&DWP(12,"esi"),"ecx");
 271         # &mov  ($E,$Eoff);
 272         &mov    ("eax",$Foff);
 273         &mov    ("ebx",$Goff);
 274         &mov    ("ecx",$Hoff);
 275         &mov    ("edi",&DWP(4*(9+16+64)+4,"esp"));#inp
 276         &add    ($E,&DWP(16,"esi"));
 277         &add    ("eax",&DWP(20,"esi"));
 278         &add    ("ebx",&DWP(24,"esi"));
 279         &add    ("ecx",&DWP(28,"esi"));
 280         &mov    (&DWP(16,"esi"),$E);
 281         &mov    (&DWP(20,"esi"),"eax");
 282         &mov    (&DWP(24,"esi"),"ebx");
 283         &mov    (&DWP(28,"esi"),"ecx");
 284
 285         &lea    ("esp",&DWP(4*(9+16+64),"esp"));# destroy frame
 286         &sub    ($K256,4*64);                   # rewind K
 287
 288         &cmp    ("edi",&DWP(8,"esp"));          # are we done yet?
 289         &jb     (&label("loop$suffix"));
 290 }
 291         &COMPACT_LOOP();
 292         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 293 &function_end_A();
 294                                                 if (!$i386 && !$ymm) {
 295         # ~20% improvement on Sandy Bridge
 296         local *ror = sub { &shrd(@_[0],@_) };
 297         &COMPACT_LOOP("_shrd");
 298         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 299 &function_end_A();
 300                                                 }
 301
 302 &set_label("K256",64);  # Yes! I keep it in the code segment!
 303 @K256=( 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5,
 304         0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5,
 305         0xd807aa98,0x12835b01,0x243185be,0x550c7dc3,
 306         0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174,
 307         0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc,
 308         0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da,
 309         0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7,
 310         0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967,
 311         0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13,
 312         0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85,
 313         0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3,
 314         0xd192e819,0xd6990624,0xf40e3585,0x106aa070,
 315         0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5,
 316         0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3,
 317         0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208,
 318         0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2     );
 319 &data_word(@K256);
 320 &data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f);
 321
 322 if (!$i386 && $unroll_after) {
 323 my @AH=($A,$K256);
 324
 325 &set_label("unrolled",16);
 326         &lea    ("esp",&DWP(-96,"esp"));
 327         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 328         &mov    ($AH[0],&DWP(0,"esi"));
 329         &mov    ($AH[1],&DWP(4,"esi"));
 330         &mov    ("ecx",&DWP(8,"esi"));
 331         &mov    ("ebx",&DWP(12,"esi"));
 332         #&mov   (&DWP(0,"esp"),$AH[0]);
 333         &mov    (&DWP(4,"esp"),$AH[1]);
 334         &xor    ($AH[1],"ecx");         # magic
 335         &mov    (&DWP(8,"esp"),"ecx");
 336         &mov    (&DWP(12,"esp"),"ebx");
 337         &mov    ($E,&DWP(16,"esi"));
 338         &mov    ("ebx",&DWP(20,"esi"));
 339         &mov    ("ecx",&DWP(24,"esi"));
 340         &mov    ("esi",&DWP(28,"esi"));
 341         #&mov   (&DWP(16,"esp"),$E);
 342         &mov    (&DWP(20,"esp"),"ebx");
 343         &mov    (&DWP(24,"esp"),"ecx");
 344         &mov    (&DWP(28,"esp"),"esi");
 345         &jmp    (&label("grand_loop"));
 346
 347 &set_label("grand_loop",16);
 348     # copy input block to stack reversing byte order
 349     for($i=0;$i<5;$i++) {
 350         &mov    ("ebx",&DWP(12*$i+0,"edi"));
 351         &mov    ("ecx",&DWP(12*$i+4,"edi"));
 352         &bswap  ("ebx");
 353         &mov    ("esi",&DWP(12*$i+8,"edi"));
 354         &bswap  ("ecx");
 355         &mov    (&DWP(32+12*$i+0,"esp"),"ebx");
 356         &bswap  ("esi");
 357         &mov    (&DWP(32+12*$i+4,"esp"),"ecx");
 358         &mov    (&DWP(32+12*$i+8,"esp"),"esi");
 359     }
 360         &mov    ("ebx",&DWP($i*12,"edi"));
 361         &add    ("edi",64);
 362         &bswap  ("ebx");
 363         &mov    (&DWP(96+4,"esp"),"edi");
 364         &mov    (&DWP(32+12*$i,"esp"),"ebx");
 365
 366     my ($t1,$t2) = ("ecx","esi");
 367     my ($a,$b,$c,$d,$e,$f,$g,$h)=(0..7);        # offsets
 368     sub off { &DWP(4*(((shift)-$i)&7),"esp"); }
 369
 370     for ($i=0;$i<64;$i++) {
 371
 372       if ($i>=16) {
 373         &mov    ($T,$t1);                       # $t1 is preloaded
 374         # &mov  ($t2,&DWP(32+4*(($i+14)&15),"esp"));
 375         &ror    ($t1,18-7);
 376          &mov   ("edi",$t2);
 377         &ror    ($t2,19-17);
 378          &xor   ($t1,$T);
 379          &shr   ($T,3);
 380         &ror    ($t1,7);
 381          &xor   ($t2,"edi");
 382          &xor   ($T,$t1);                       # T = sigma0(X[-15])
 383         &ror    ($t2,17);
 384          &add   ($T,&DWP(32+4*($i&15),"esp"));  # T += X[-16]
 385         &shr    ("edi",10);
 386          &add   ($T,&DWP(32+4*(($i+9)&15),"esp"));      # T += X[-7]
 387         #&xor   ("edi",$t2)                     # sigma1(X[-2])
 388         # &add  ($T,"edi");                     # T += sigma1(X[-2])
 389         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
 390       }
 391         &mov    ($t1,$E);
 392          &xor   ("edi",$t2)                     if ($i>=16);    # sigma1(X[-2])
 393          &mov   ($t2,&off($f));
 394         &ror    ($E,25-11);
 395          &add   ($T,"edi")                      if ($i>=16);    # T += sigma1(X[-2])
 396          &mov   ("edi",&off($g));
 397         &xor    ($E,$t1);
 398          &mov   ($T,&DWP(32+4*($i&15),"esp"))   if ($i<16);     # X[i]
 399          &mov   (&DWP(32+4*($i&15),"esp"),$T)   if ($i>=16 && $i<62);   # save X[0]
 400          &xor   ($t2,"edi");
 401         &ror    ($E,11-6);
 402          &and   ($t2,$t1);
 403          &mov   (&off($e),$t1);         # save $E, modulo-scheduled
 404         &xor    ($E,$t1);
 405          &add   ($T,&off($h));          # T += h
 406          &xor   ("edi",$t2);            # Ch(e,f,g)
 407         &ror    ($E,6);                 # Sigma1(e)
 408          &mov   ($t1,$AH[0]);
 409          &add   ($T,"edi");             # T += Ch(e,f,g)
 410
 411         &ror    ($t1,22-13);
 412          &mov   ($t2,$AH[0]);
 413          &mov   ("edi",&off($b));
 414         &xor    ($t1,$AH[0]);
 415          &mov   (&off($a),$AH[0]);      # save $A, modulo-scheduled
 416          &xor   ($AH[0],"edi");         # a ^= b, (b^c) in next round
 417         &ror    ($t1,13-2);
 418          &and   ($AH[1],$AH[0]);        # (b^c) &= (a^b)
 419          &lea   ($E,&DWP(@K256[$i],$T,$E));     # T += Sigma1(1)+K[i]
 420         &xor    ($t1,$t2);
 421          &xor   ($AH[1],"edi");         # h = Maj(a,b,c) = Ch(a^b,c,b)
 422          &mov   ($t2,&DWP(32+4*(($i+2)&15),"esp"))      if ($i>=15 && $i<63);
 423         &ror    ($t1,2);                # Sigma0(a)
 424
 425          &add   ($AH[1],$E);            # h += T
 426          &add   ($E,&off($d));          # d += T
 427         &add    ($AH[1],$t1);           # h += Sigma0(a)
 428          &mov   ($t1,&DWP(32+4*(($i+15)&15),"esp"))     if ($i>=15 && $i<63);
 429
 430         @AH = reverse(@AH);             # rotate(a,h)
 431         ($t1,$t2) = ($t2,$t1);          # rotate(t1,t2)
 432     }
 433         &mov    ("esi",&DWP(96,"esp")); #ctx
 434                                         #&mov   ($AH[0],&DWP(0,"esp"));
 435         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
 436                                         #&mov   ("edi", &DWP(8,"esp"));
 437         &mov    ("ecx",&DWP(12,"esp"));
 438         &add    ($AH[0],&DWP(0,"esi"));
 439         &add    ($AH[1],&DWP(4,"esi"));
 440         &add    ("edi",&DWP(8,"esi"));
 441         &add    ("ecx",&DWP(12,"esi"));
 442         &mov    (&DWP(0,"esi"),$AH[0]);
 443         &mov    (&DWP(4,"esi"),$AH[1]);
 444         &mov    (&DWP(8,"esi"),"edi");
 445         &mov    (&DWP(12,"esi"),"ecx");
 446          #&mov  (&DWP(0,"esp"),$AH[0]);
 447          &mov   (&DWP(4,"esp"),$AH[1]);
 448          &xor   ($AH[1],"edi");         # magic
 449          &mov   (&DWP(8,"esp"),"edi");
 450          &mov   (&DWP(12,"esp"),"ecx");
 451         #&mov   ($E,&DWP(16,"esp"));
 452         &mov    ("edi",&DWP(20,"esp"));
 453         &mov    ("ebx",&DWP(24,"esp"));
 454         &mov    ("ecx",&DWP(28,"esp"));
 455         &add    ($E,&DWP(16,"esi"));
 456         &add    ("edi",&DWP(20,"esi"));
 457         &add    ("ebx",&DWP(24,"esi"));
 458         &add    ("ecx",&DWP(28,"esi"));
 459         &mov    (&DWP(16,"esi"),$E);
 460         &mov    (&DWP(20,"esi"),"edi");
 461         &mov    (&DWP(24,"esi"),"ebx");
 462         &mov    (&DWP(28,"esi"),"ecx");
 463          #&mov  (&DWP(16,"esp"),$E);
 464          &mov   (&DWP(20,"esp"),"edi");
 465         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 466          &mov   (&DWP(24,"esp"),"ebx");
 467          &mov   (&DWP(28,"esp"),"ecx");
 468
 469         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
 470         &jb     (&label("grand_loop"));
 471
 472         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
 473 &function_end_A();
 474
 475                                                 if ($ymm) {{{
 476 my @X = map("xmm$_",(0..3));
 477 my ($t0,$t1,$t2,$t3) = map("xmm$_",(4..7));
 478 my @AH = ($A,$T);
 479
 480 &set_label("XOP",16);
 481         &lea    ("esp",&DWP(-96,"esp"));
 482         &vzeroall       ();
 483         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 484         &mov    ($AH[0],&DWP(0,"esi"));
 485         &mov    ($AH[1],&DWP(4,"esi"));
 486         &mov    ("ecx",&DWP(8,"esi"));
 487         &mov    ("edi",&DWP(12,"esi"));
 488         #&mov   (&DWP(0,"esp"),$AH[0]);
 489         &mov    (&DWP(4,"esp"),$AH[1]);
 490         &xor    ($AH[1],"ecx");                 # magic
 491         &mov    (&DWP(8,"esp"),"ecx");
 492         &mov    (&DWP(12,"esp"),"edi");
 493         &mov    ($E,&DWP(16,"esi"));
 494         &mov    ("edi",&DWP(20,"esi"));
 495         &mov    ("ecx",&DWP(24,"esi"));
 496         &mov    ("esi",&DWP(28,"esi"));
 497         #&mov   (&DWP(16,"esp"),$E);
 498         &mov    (&DWP(20,"esp"),"edi");
 499         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 500         &mov    (&DWP(24,"esp"),"ecx");
 501         &mov    (&DWP(28,"esp"),"esi");
 502         &vmovdqa        ($t3,&QWP(256,$K256));
 503         &jmp    (&label("grand_xop"));
 504
 505 &set_label("grand_xop",16);
 506         # load input, reverse byte order, add K256[0..15], save to stack
 507         &vmovdqu        (@X[0],&QWP(0,"edi"));
 508         &vmovdqu        (@X[1],&QWP(16,"edi"));
 509         &vmovdqu        (@X[2],&QWP(32,"edi"));
 510         &vmovdqu        (@X[3],&QWP(48,"edi"));
 511         &add            ("edi",64);
 512         &vpshufb        (@X[0],@X[0],$t3);
 513         &mov            (&DWP(96+4,"esp"),"edi");
 514         &vpshufb        (@X[1],@X[1],$t3);
 515         &vpshufb        (@X[2],@X[2],$t3);
 516         &vpaddd         ($t0,@X[0],&QWP(0,$K256));
 517         &vpshufb        (@X[3],@X[3],$t3);
 518         &vpaddd         ($t1,@X[1],&QWP(16,$K256));
 519         &vpaddd         ($t2,@X[2],&QWP(32,$K256));
 520         &vpaddd         ($t3,@X[3],&QWP(48,$K256));
 521         &vmovdqa        (&QWP(32+0,"esp"),$t0);
 522         &vmovdqa        (&QWP(32+16,"esp"),$t1);
 523         &vmovdqa        (&QWP(32+32,"esp"),$t2);
 524         &vmovdqa        (&QWP(32+48,"esp"),$t3);
 525         &jmp            (&label("xop_00_47"));
 526
 527 &set_label("xop_00_47",16);
 528         &add            ($K256,64);
 529
 530 sub XOP_00_47 () {
 531 my $j = shift;
 532 my $body = shift;
 533 my @X = @_;
 534 my @insns = (&$body,&$body,&$body,&$body);      # 120 instructions
 535
 536         &vpalignr       ($t0,@X[1],@X[0],4);    # X[1..4]
 537           eval(shift(@insns));
 538           eval(shift(@insns));
 539          &vpalignr      ($t3,@X[3],@X[2],4);    # X[9..12]
 540           eval(shift(@insns));
 541           eval(shift(@insns));
 542         &vprotd         ($t1,$t0,14);
 543           eval(shift(@insns));
 544           eval(shift(@insns));
 545         &vpsrld         ($t0,$t0,3);
 546          &vpaddd        (@X[0],@X[0],$t3);      # X[0..3] += X[9..12]
 547           eval(shift(@insns));
 548           eval(shift(@insns));
 549           eval(shift(@insns));
 550           eval(shift(@insns));
 551         &vprotd         ($t2,$t1,25-14);
 552         &vpxor          ($t0,$t0,$t1);
 553           eval(shift(@insns));
 554           eval(shift(@insns));
 555           eval(shift(@insns));
 556           eval(shift(@insns));
 557          &vprotd        ($t3,@X[3],13);
 558         &vpxor          ($t0,$t0,$t2);          # sigma0(X[1..4])
 559           eval(shift(@insns));
 560           eval(shift(@insns));
 561          &vpsrld        ($t2,@X[3],10);
 562           eval(shift(@insns));
 563           eval(shift(@insns));
 564         &vpaddd         (@X[0],@X[0],$t0);      # X[0..3] += sigma0(X[1..4])
 565           eval(shift(@insns));
 566           eval(shift(@insns));
 567          &vprotd        ($t1,$t3,15-13);
 568          &vpxor         ($t3,$t3,$t2);
 569           eval(shift(@insns));
 570           eval(shift(@insns));
 571           eval(shift(@insns));
 572           eval(shift(@insns));
 573          &vpxor         ($t3,$t3,$t1);          # sigma1(X[14..15])
 574           eval(shift(@insns));
 575           eval(shift(@insns));
 576           eval(shift(@insns));
 577           eval(shift(@insns));
 578         &vpsrldq        ($t3,$t3,8);
 579           eval(shift(@insns));
 580           eval(shift(@insns));
 581           eval(shift(@insns));
 582           eval(shift(@insns));
 583         &vpaddd         (@X[0],@X[0],$t3);      # X[0..1] += sigma1(X[14..15])
 584           eval(shift(@insns));
 585           eval(shift(@insns));
 586           eval(shift(@insns));
 587           eval(shift(@insns));
 588          &vprotd        ($t3,@X[0],13);
 589           eval(shift(@insns));
 590           eval(shift(@insns));
 591          &vpsrld        ($t2,@X[0],10);
 592           eval(shift(@insns));
 593           eval(shift(@insns));
 594          &vprotd        ($t1,$t3,15-13);
 595           eval(shift(@insns));
 596           eval(shift(@insns));
 597          &vpxor         ($t3,$t3,$t2);
 598           eval(shift(@insns));
 599           eval(shift(@insns));
 600           eval(shift(@insns));
 601           eval(shift(@insns));
 602          &vpxor         ($t3,$t3,$t1);          # sigma1(X[16..17])
 603           eval(shift(@insns));
 604           eval(shift(@insns));
 605           eval(shift(@insns));
 606           eval(shift(@insns));
 607         &vpslldq        ($t3,$t3,8);            # 22 instructions
 608           eval(shift(@insns));
 609           eval(shift(@insns));
 610           eval(shift(@insns));
 611           eval(shift(@insns));
 612         &vpaddd         (@X[0],@X[0],$t3);      # X[2..3] += sigma1(X[16..17])
 613           eval(shift(@insns));
 614           eval(shift(@insns));
 615           eval(shift(@insns));
 616           eval(shift(@insns));
 617         &vpaddd         ($t2,@X[0],&QWP(16*$j,$K256));
 618
 619         foreach (@insns) { eval; }              # remaining instructions
 620
 621         &vmovdqa        (&QWP(32+16*$j,"esp"),$t2);
 622 }
 623
 624 sub body_00_15 () {
 625         (
 626         '&mov   ("ecx",$E);',
 627          '&mov  ("esi",&off($f));',
 628         '&ror   ($E,25-11);',
 629          '&mov  ("edi",&off($g));',
 630         '&xor   ($E,"ecx");',
 631          '&xor  ("esi","edi");',
 632         '&ror   ($E,11-6);',
 633          '&and  ("esi","ecx");',
 634          '&mov  (&off($e),"ecx");',     # save $E, modulo-scheduled
 635         '&xor   ($E,"ecx");',
 636          '&xor  ("edi","esi");',        # Ch(e,f,g)
 637         '&ror   ($E,6);',               # T = Sigma1(e)
 638          '&mov  ("ecx",$AH[0]);',
 639         '&mov   ("esi",$AH[0]);',
 640          '&add  ($E,&off($h));',        # T += h
 641
 642         '&ror   ("ecx",22-13);',
 643          '&add  ($E,"edi");',           # T += Ch(e,f,g)
 644          '&mov  ("edi",&off($b));',
 645         '&xor   ("ecx",$AH[0]);',
 646          '&mov  (&off($a),$AH[0]);',    # save $A, modulo-scheduled
 647          '&xor  ($AH[0],"edi");',       # a ^= b, (b^c) in next round
 648         '&ror   ("ecx",13-2);',
 649          '&and  ($AH[1],$AH[0]);',      # (b^c) &= (a^b)
 650          '&add  ($E,&DWP(32+4*($i&15),"esp"));',        # T += K[i]+X[i]
 651         '&xor   ("ecx","esi");',
 652          '&xor  ($AH[1],"edi");',       # h = Maj(a,b,c) = Ch(a^b,c,b)
 653         '&ror   ("ecx",2);',            # Sigma0(a)
 654
 655          '&add  ($AH[1],$E);',          # h += T
 656          '&add  ($E,&off($d));',        # d += T
 657         '&add   ($AH[1],"ecx");'.       # h += Sigma0(a)
 658
 659         '@AH = reverse(@AH); $i++;'     # rotate(a,h)
 660         );
 661 }
 662
 663     for ($i=0,$j=0; $j<4; $j++) {
 664         &XOP_00_47($j,\&body_00_15,@X);
 665         push(@X,shift(@X));             # rotate(@X)
 666     }
 667         &cmp    (&DWP(16*$j,$K256),0x00010203);
 668         &jne    (&label("xop_00_47"));
 669
 670     for ($i=0; $i<16; ) {
 671         foreach(body_00_15()) { eval; }
 672     }
 673
 674         &mov    ("esi",&DWP(96,"esp")); #ctx
 675                                         #&mov   ($AH[0],&DWP(0,"esp"));
 676         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
 677                                         #&mov   ("edi", &DWP(8,"esp"));
 678         &mov    ("ecx",&DWP(12,"esp"));
 679         &add    ($AH[0],&DWP(0,"esi"));
 680         &add    ($AH[1],&DWP(4,"esi"));
 681         &add    ("edi",&DWP(8,"esi"));
 682         &add    ("ecx",&DWP(12,"esi"));
 683         &mov    (&DWP(0,"esi"),$AH[0]);
 684         &mov    (&DWP(4,"esi"),$AH[1]);
 685         &mov    (&DWP(8,"esi"),"edi");
 686         &mov    (&DWP(12,"esi"),"ecx");
 687          #&mov  (&DWP(0,"esp"),$AH[0]);
 688          &mov   (&DWP(4,"esp"),$AH[1]);
 689          &xor   ($AH[1],"edi");                 # magic
 690          &mov   (&DWP(8,"esp"),"edi");
 691          &mov   (&DWP(12,"esp"),"ecx");
 692         #&mov   ($E,&DWP(16,"esp"));
 693         &mov    ("edi",&DWP(20,"esp"));
 694         &mov    ("ecx",&DWP(24,"esp"));
 695         &add    ($E,&DWP(16,"esi"));
 696         &add    ("edi",&DWP(20,"esi"));
 697         &add    ("ecx",&DWP(24,"esi"));
 698         &mov    (&DWP(16,"esi"),$E);
 699         &mov    (&DWP(20,"esi"),"edi");
 700          &mov   (&DWP(20,"esp"),"edi");
 701         &mov    ("edi",&DWP(28,"esp"));
 702         &mov    (&DWP(24,"esi"),"ecx");
 703          #&mov  (&DWP(16,"esp"),$E);
 704         &add    ("edi",&DWP(28,"esi"));
 705          &mov   (&DWP(24,"esp"),"ecx");
 706         &mov    (&DWP(28,"esi"),"edi");
 707          &mov   (&DWP(28,"esp"),"edi");
 708         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 709
 710         &vmovdqa        ($t3,&QWP(64,$K256));
 711         &sub    ($K256,3*64);                   # rewind K
 712         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
 713         &jb     (&label("grand_xop"));
 714
 715         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
 716         &vzeroall       ();
 717 &function_end_A();
 718
 719 &set_label("AVX",16);
 720         &lea    ("esp",&DWP(-96,"esp"));
 721         &vzeroall       ();
 722         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 723         &mov    ($AH[0],&DWP(0,"esi"));
 724         &mov    ($AH[1],&DWP(4,"esi"));
 725         &mov    ("ecx",&DWP(8,"esi"));
 726         &mov    ("edi",&DWP(12,"esi"));
 727         #&mov   (&DWP(0,"esp"),$AH[0]);
 728         &mov    (&DWP(4,"esp"),$AH[1]);
 729         &xor    ($AH[1],"ecx");                 # magic
 730         &mov    (&DWP(8,"esp"),"ecx");
 731         &mov    (&DWP(12,"esp"),"edi");
 732         &mov    ($E,&DWP(16,"esi"));
 733         &mov    ("edi",&DWP(20,"esi"));
 734         &mov    ("ecx",&DWP(24,"esi"));
 735         &mov    ("esi",&DWP(28,"esi"));
 736         #&mov   (&DWP(16,"esp"),$E);
 737         &mov    (&DWP(20,"esp"),"edi");
 738         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 739         &mov    (&DWP(24,"esp"),"ecx");
 740         &mov    (&DWP(28,"esp"),"esi");
 741         &vmovdqa        ($t3,&QWP(256,$K256));
 742         &jmp    (&label("grand_avx"));
 743
 744 &set_label("grand_avx",16);
 745         # load input, reverse byte order, add K256[0..15], save to stack
 746         &vmovdqu        (@X[0],&QWP(0,"edi"));
 747         &vmovdqu        (@X[1],&QWP(16,"edi"));
 748         &vmovdqu        (@X[2],&QWP(32,"edi"));
 749         &vmovdqu        (@X[3],&QWP(48,"edi"));
 750         &add            ("edi",64);
 751         &vpshufb        (@X[0],@X[0],$t3);
 752         &mov            (&DWP(96+4,"esp"),"edi");
 753         &vpshufb        (@X[1],@X[1],$t3);
 754         &vpshufb        (@X[2],@X[2],$t3);
 755         &vpaddd         ($t0,@X[0],&QWP(0,$K256));
 756         &vpshufb        (@X[3],@X[3],$t3);
 757         &vpaddd         ($t1,@X[1],&QWP(16,$K256));
 758         &vpaddd         ($t2,@X[2],&QWP(32,$K256));
 759         &vpaddd         ($t3,@X[3],&QWP(48,$K256));
 760         &vmovdqa        (&QWP(32+0,"esp"),$t0);
 761         &vmovdqa        (&QWP(32+16,"esp"),$t1);
 762         &vmovdqa        (&QWP(32+32,"esp"),$t2);
 763         &vmovdqa        (&QWP(32+48,"esp"),$t3);
 764         &jmp            (&label("avx_00_47"));
 765
 766 &set_label("avx_00_47",16);
 767         &add            ($K256,64);
 768
 769 sub Xupdate_AVX () {
 770         (
 771         '&vpalignr      ($t0,@X[1],@X[0],4);',  # X[1..4]
 772          '&vpalignr     ($t3,@X[3],@X[2],4);',  # X[9..12]
 773         '&vpsrld        ($t2,$t0,7);',
 774          '&vpaddd       (@X[0],@X[0],$t3);',    # X[0..3] += X[9..16]
 775         '&vpsrld        ($t3,$t0,3);',
 776         '&vpslld        ($t1,$t0,14);',
 777         '&vpxor         ($t0,$t3,$t2);',
 778         '&vpsrld        ($t2,$t2,18-7);',
 779         '&vpxor         ($t0,$t0,$t1);',
 780         '&vpslld        ($t1,$t1,25-14);',
 781         '&vpxor         ($t0,$t0,$t2);',
 782          '&vpsrld       ($t3,@X[3],10);',
 783         '&vpxor         ($t0,$t0,$t1);',        # sigma0(X[1..4])
 784          '&vpslld       ($t2,@X[3],13);',
 785         '&vpaddd        (@X[0],@X[0],$t0);',    # X[0..3] += sigma0(X[1..4])
 786          '&vpsrld       ($t1,@X[3],17);',
 787          '&vpxor        ($t3,$t3,$t2);',
 788          '&vpslld       ($t2,$t2,15-13);',
 789          '&vpxor        ($t3,$t3,$t1);',
 790          '&vpsrld       ($t1,$t1,19-17);',
 791          '&vpxor        ($t3,$t3,$t2);',
 792          '&vpxor        ($t3,$t3,$t1);',        # sigma1(X[14..15])
 793         '&vpsrldq       ($t3,$t3,8);',
 794         '&vpaddd        (@X[0],@X[0],$t3);',    # X[0..1] += sigma1(X[14..15])
 795          '&vpsrld       ($t3,@X[0],10);',
 796          '&vpslld       ($t2,@X[0],13);',
 797          '&vpsrld       ($t1,@X[0],17);',
 798          '&vpxor        ($t3,$t3,$t2);',
 799          '&vpslld       ($t2,$t2,15-13);',
 800          '&vpxor        ($t3,$t3,$t1);',
 801          '&vpsrld       ($t1,$t1,19-17);',
 802          '&vpxor        ($t3,$t3,$t2);',
 803          '&vpxor        ($t3,$t3,$t1);',        # sigma1(X[16..17])
 804         '&vpslldq       ($t3,$t3,8);',
 805         '&vpaddd        (@X[0],@X[0],$t3);'     # X[2..3] += sigma1(X[16..17])
 806         );
 807 }
 808
 809 local *ror = sub { &shrd(@_[0],@_) };
 810 sub AVX_00_47 () {
 811 my $j = shift;
 812 my $body = shift;
 813 my @X = @_;
 814 my @insns = (&$body,&$body,&$body,&$body);      # 120 instructions
 815
 816         foreach (Xupdate_AVX()) {               # 35 instructions
 817             eval;
 818             eval(shift(@insns));
 819             eval(shift(@insns));
 820             eval(shift(@insns));
 821         }
 822         &vpaddd         ($t2,@X[0],&QWP(16*$j,$K256));
 823         foreach (@insns) { eval; }              # remaining instructions
 824         &vmovdqa        (&QWP(32+16*$j,"esp"),$t2);
 825 }
 826
 827     for ($i=0,$j=0; $j<4; $j++) {
 828         &AVX_00_47($j,\&body_00_15,@X);
 829         push(@X,shift(@X));             # rotate(@X)
 830     }
 831         &cmp    (&DWP(16*$j,$K256),0x00010203);
 832         &jne    (&label("avx_00_47"));
 833
 834     for ($i=0; $i<16; ) {
 835         foreach(body_00_15()) { eval; }
 836     }
 837
 838         &mov    ("esi",&DWP(96,"esp")); #ctx
 839                                         #&mov   ($AH[0],&DWP(0,"esp"));
 840         &xor    ($AH[1],"edi");         #&mov   ($AH[1],&DWP(4,"esp"));
 841                                         #&mov   ("edi", &DWP(8,"esp"));
 842         &mov    ("ecx",&DWP(12,"esp"));
 843         &add    ($AH[0],&DWP(0,"esi"));
 844         &add    ($AH[1],&DWP(4,"esi"));
 845         &add    ("edi",&DWP(8,"esi"));
 846         &add    ("ecx",&DWP(12,"esi"));
 847         &mov    (&DWP(0,"esi"),$AH[0]);
 848         &mov    (&DWP(4,"esi"),$AH[1]);
 849         &mov    (&DWP(8,"esi"),"edi");
 850         &mov    (&DWP(12,"esi"),"ecx");
 851          #&mov  (&DWP(0,"esp"),$AH[0]);
 852          &mov   (&DWP(4,"esp"),$AH[1]);
 853          &xor   ($AH[1],"edi");                 # magic
 854          &mov   (&DWP(8,"esp"),"edi");
 855          &mov   (&DWP(12,"esp"),"ecx");
 856         #&mov   ($E,&DWP(16,"esp"));
 857         &mov    ("edi",&DWP(20,"esp"));
 858         &mov    ("ecx",&DWP(24,"esp"));
 859         &add    ($E,&DWP(16,"esi"));
 860         &add    ("edi",&DWP(20,"esi"));
 861         &add    ("ecx",&DWP(24,"esi"));
 862         &mov    (&DWP(16,"esi"),$E);
 863         &mov    (&DWP(20,"esi"),"edi");
 864          &mov   (&DWP(20,"esp"),"edi");
 865         &mov    ("edi",&DWP(28,"esp"));
 866         &mov    (&DWP(24,"esi"),"ecx");
 867          #&mov  (&DWP(16,"esp"),$E);
 868         &add    ("edi",&DWP(28,"esi"));
 869          &mov   (&DWP(24,"esp"),"ecx");
 870         &mov    (&DWP(28,"esi"),"edi");
 871          &mov   (&DWP(28,"esp"),"edi");
 872         &mov    ("edi",&DWP(96+4,"esp"));       # inp
 873
 874         &vmovdqa        ($t3,&QWP(64,$K256));
 875         &sub    ($K256,3*64);                   # rewind K
 876         &cmp    ("edi",&DWP(96+8,"esp"));       # are we done yet?
 877         &jb     (&label("grand_avx"));
 878
 879         &mov    ("esp",&DWP(96+12,"esp"));      # restore sp
 880         &vzeroall       ();
 881 &function_end_A();
 882                                                 }}}
 883 }
 884 &function_end_B("sha256_block_data_order");
 885 &asciz("SHA256 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
 886
 887 &asm_finish();