crypto/sha/asm/sha512-586.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 # SHA512 block transform for x86. September 2007.
  11 #
  12 # May 2013.
  13 #
  14 # Add SSSE3 code path, 20-25% improvement [over original SSE2 code].
  15 #
  16 # Performance in clock cycles per processed byte (less is better):
  17 #
  18 #               gcc     icc     x86 asm SIMD(*) x86_64(**)
  19 # Pentium       100     97      61      -       -
  20 # PIII          75      77      56      -       -
  21 # P4            116     95      82      34.6    30.8
  22 # AMD K8        54      55      36      20.7    9.57
  23 # Core2         66      57      40      15.9    9.97
  24 # Westmere      70      -       38      12.2    9.58
  25 # Sandy Bridge  58      -       35      11.9    11.2
  26 # Ivy Bridge    50      -       33      11.5    8.17
  27 # Haswell       46      -       29      11.3    7.66
  28 # Bulldozer     121     -       50      14.0    13.5
  29 # VIA Nano      91      -       52      33      14.7
  30 # Atom          126     -       68      48(***) 14.7
  31 # Silvermont    97      -       58      42(***) 17.5
  32 #
  33 # (*)   whichever best applicable.
  34 # (**)  x86_64 assembler performance is presented for reference
  35 #       purposes, the results are for integer-only code.
  36 # (***) paddq is increadibly slow on Atom.
  37 #
  38 # IALU code-path is optimized for elder Pentiums. On vanilla Pentium
  39 # performance improvement over compiler generated code reaches ~60%,
  40 # while on PIII - ~35%. On newer µ-archs improvement varies from 15%
  41 # to 50%, but it's less important as they are expected to execute SSE2
  42 # code-path, which is commonly ~2-3x faster [than compiler generated
  43 # code]. SSE2 code-path is as fast as original sha512-sse2.pl, even
  44 # though it does not use 128-bit operations. The latter means that
  45 # SSE2-aware kernel is no longer required to execute the code. Another
  46 # difference is that new code optimizes amount of writes, but at the
  47 # cost of increased data cache "footprint" by 1/2KB.
  48
  49 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  50 push(@INC,"${dir}","${dir}../../perlasm");
  51 require "x86asm.pl";
  52
  53 $output=pop;
  54 open STDOUT,">$output";
  55
  56 &asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");
  57
  58 $sse2=0;
  59 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
  60
  61 &external_label("OPENSSL_ia32cap_P") if ($sse2);
  62
  63 $Tlo=&DWP(0,"esp");     $Thi=&DWP(4,"esp");
  64 $Alo=&DWP(8,"esp");     $Ahi=&DWP(8+4,"esp");
  65 $Blo=&DWP(16,"esp");    $Bhi=&DWP(16+4,"esp");
  66 $Clo=&DWP(24,"esp");    $Chi=&DWP(24+4,"esp");
  67 $Dlo=&DWP(32,"esp");    $Dhi=&DWP(32+4,"esp");
  68 $Elo=&DWP(40,"esp");    $Ehi=&DWP(40+4,"esp");
  69 $Flo=&DWP(48,"esp");    $Fhi=&DWP(48+4,"esp");
  70 $Glo=&DWP(56,"esp");    $Ghi=&DWP(56+4,"esp");
  71 $Hlo=&DWP(64,"esp");    $Hhi=&DWP(64+4,"esp");
  72 $K512="ebp";
  73
  74 $Asse2=&QWP(0,"esp");
  75 $Bsse2=&QWP(8,"esp");
  76 $Csse2=&QWP(16,"esp");
  77 $Dsse2=&QWP(24,"esp");
  78 $Esse2=&QWP(32,"esp");
  79 $Fsse2=&QWP(40,"esp");
  80 $Gsse2=&QWP(48,"esp");
  81 $Hsse2=&QWP(56,"esp");
  82
  83 $A="mm0";       # B-D and
  84 $E="mm4";       # F-H are commonly loaded to respectively mm1-mm3 and
  85                 # mm5-mm7, but it's done on on-demand basis...
  86 $BxC="mm2";     # ... except for B^C
  87
  88 sub BODY_00_15_sse2 {
  89     my $phase=shift;
  90
  91         #&movq  ("mm5",$Fsse2);                 # load f
  92         #&movq  ("mm6",$Gsse2);                 # load g
  93
  94         &movq   ("mm1",$E);                     # %mm1 is sliding right
  95          &pxor  ("mm5","mm6");                  # f^=g
  96         &psrlq  ("mm1",14);
  97          &movq  ($Esse2,$E);                    # modulo-scheduled save e
  98          &pand  ("mm5",$E);                     # f&=e
  99         &psllq  ($E,23);                        # $E is sliding left
 100          &movq  ($A,"mm3")                      if ($phase<2);
 101          &movq  (&QWP(8*9,"esp"),"mm7")         # save X[i]
 102         &movq   ("mm3","mm1");                  # %mm3 is T1
 103          &psrlq ("mm1",4);
 104          &pxor  ("mm5","mm6");                  # Ch(e,f,g)
 105         &pxor   ("mm3",$E);
 106          &psllq ($E,23);
 107         &pxor   ("mm3","mm1");
 108          &movq  ($Asse2,$A);                    # modulo-scheduled save a
 109          &paddq ("mm7","mm5");                  # X[i]+=Ch(e,f,g)
 110         &pxor   ("mm3",$E);
 111          &psrlq ("mm1",23);
 112          &paddq ("mm7",$Hsse2);                 # X[i]+=h
 113         &pxor   ("mm3","mm1");
 114          &psllq ($E,4);
 115          &paddq ("mm7",QWP(0,$K512));           # X[i]+=K512[i]
 116         &pxor   ("mm3",$E);                     # T1=Sigma1_512(e)
 117
 118          &movq  ($E,$Dsse2);                    # e = load d, e in next round
 119         &paddq  ("mm3","mm7");                  # T1+=X[i]
 120          &movq  ("mm5",$A);                     # %mm5 is sliding right
 121          &psrlq ("mm5",28);
 122         &paddq  ($E,"mm3");                     # d += T1
 123          &movq  ("mm6",$A);                     # %mm6 is sliding left
 124          &movq  ("mm7","mm5");
 125          &psllq ("mm6",25);
 126         &movq   ("mm1",$Bsse2);                 # load b
 127          &psrlq ("mm5",6);
 128          &pxor  ("mm7","mm6");
 129         &sub    ("esp",8);
 130          &psllq ("mm6",5);
 131          &pxor  ("mm7","mm5");
 132         &pxor   ($A,"mm1");                     # a^b, b^c in next round
 133          &psrlq ("mm5",5);
 134          &pxor  ("mm7","mm6");
 135         &pand   ($BxC,$A);                      # (b^c)&(a^b)
 136          &psllq ("mm6",6);
 137          &pxor  ("mm7","mm5");
 138         &pxor   ($BxC,"mm1");                   # [h=]Maj(a,b,c)
 139          &pxor  ("mm6","mm7");                  # Sigma0_512(a)
 140          &movq  ("mm7",&QWP(8*(9+16-1),"esp"))  if ($phase!=0); # pre-fetch
 141          &movq  ("mm5",$Fsse2)                  if ($phase==0); # load f
 142
 143     if ($phase>1) {
 144         &paddq  ($BxC,"mm6");                   # h+=Sigma0(a)
 145          &add   ($K512,8);
 146         #&paddq ($BxC,"mm3");                   # h+=T1
 147
 148         ($A,$BxC) = ($BxC,$A);                  # rotate registers
 149     } else {
 150         &paddq  ("mm3",$BxC);                   # T1+=Maj(a,b,c)
 151          &movq  ($BxC,$A);
 152          &add   ($K512,8);
 153         &paddq  ("mm3","mm6");                  # T1+=Sigma0(a)
 154          &movq  ("mm6",$Gsse2)                  if ($phase==0); # load g
 155         #&movq  ($A,"mm3");                     # h=T1
 156     }
 157 }
 158
 159 sub BODY_00_15_x86 {
 160         #define Sigma1(x)       (ROTR((x),14) ^ ROTR((x),18)  ^ ROTR((x),41))
 161         #       LO              lo>>14^hi<<18 ^ lo>>18^hi<<14 ^ hi>>9^lo<<23
 162         #       HI              hi>>14^lo<<18 ^ hi>>18^lo<<14 ^ lo>>9^hi<<23
 163         &mov    ("ecx",$Elo);
 164         &mov    ("edx",$Ehi);
 165         &mov    ("esi","ecx");
 166
 167         &shr    ("ecx",9);      # lo>>9
 168         &mov    ("edi","edx");
 169         &shr    ("edx",9);      # hi>>9
 170         &mov    ("ebx","ecx");
 171         &shl    ("esi",14);     # lo<<14
 172         &mov    ("eax","edx");
 173         &shl    ("edi",14);     # hi<<14
 174         &xor    ("ebx","esi");
 175
 176         &shr    ("ecx",14-9);   # lo>>14
 177         &xor    ("eax","edi");
 178         &shr    ("edx",14-9);   # hi>>14
 179         &xor    ("eax","ecx");
 180         &shl    ("esi",18-14);  # lo<<18
 181         &xor    ("ebx","edx");
 182         &shl    ("edi",18-14);  # hi<<18
 183         &xor    ("ebx","esi");
 184
 185         &shr    ("ecx",18-14);  # lo>>18
 186         &xor    ("eax","edi");
 187         &shr    ("edx",18-14);  # hi>>18
 188         &xor    ("eax","ecx");
 189         &shl    ("esi",23-18);  # lo<<23
 190         &xor    ("ebx","edx");
 191         &shl    ("edi",23-18);  # hi<<23
 192         &xor    ("eax","esi");
 193         &xor    ("ebx","edi");                  # T1 = Sigma1(e)
 194
 195         &mov    ("ecx",$Flo);
 196         &mov    ("edx",$Fhi);
 197         &mov    ("esi",$Glo);
 198         &mov    ("edi",$Ghi);
 199          &add   ("eax",$Hlo);
 200          &adc   ("ebx",$Hhi);                   # T1 += h
 201         &xor    ("ecx","esi");
 202         &xor    ("edx","edi");
 203         &and    ("ecx",$Elo);
 204         &and    ("edx",$Ehi);
 205          &add   ("eax",&DWP(8*(9+15)+0,"esp"));
 206          &adc   ("ebx",&DWP(8*(9+15)+4,"esp")); # T1 += X[0]
 207         &xor    ("ecx","esi");
 208         &xor    ("edx","edi");                  # Ch(e,f,g) = (f^g)&e)^g
 209
 210         &mov    ("esi",&DWP(0,$K512));
 211         &mov    ("edi",&DWP(4,$K512));          # K[i]
 212         &add    ("eax","ecx");
 213         &adc    ("ebx","edx");                  # T1 += Ch(e,f,g)
 214         &mov    ("ecx",$Dlo);
 215         &mov    ("edx",$Dhi);
 216         &add    ("eax","esi");
 217         &adc    ("ebx","edi");                  # T1 += K[i]
 218         &mov    ($Tlo,"eax");
 219         &mov    ($Thi,"ebx");                   # put T1 away
 220         &add    ("eax","ecx");
 221         &adc    ("ebx","edx");                  # d += T1
 222
 223         #define Sigma0(x)       (ROTR((x),28) ^ ROTR((x),34) ^ ROTR((x),39))
 224         #       LO              lo>>28^hi<<4  ^ hi>>2^lo<<30 ^ hi>>7^lo<<25
 225         #       HI              hi>>28^lo<<4  ^ lo>>2^hi<<30 ^ lo>>7^hi<<25
 226         &mov    ("ecx",$Alo);
 227         &mov    ("edx",$Ahi);
 228         &mov    ($Dlo,"eax");
 229         &mov    ($Dhi,"ebx");
 230         &mov    ("esi","ecx");
 231
 232         &shr    ("ecx",2);      # lo>>2
 233         &mov    ("edi","edx");
 234         &shr    ("edx",2);      # hi>>2
 235         &mov    ("ebx","ecx");
 236         &shl    ("esi",4);      # lo<<4
 237         &mov    ("eax","edx");
 238         &shl    ("edi",4);      # hi<<4
 239         &xor    ("ebx","esi");
 240
 241         &shr    ("ecx",7-2);    # lo>>7
 242         &xor    ("eax","edi");
 243         &shr    ("edx",7-2);    # hi>>7
 244         &xor    ("ebx","ecx");
 245         &shl    ("esi",25-4);   # lo<<25
 246         &xor    ("eax","edx");
 247         &shl    ("edi",25-4);   # hi<<25
 248         &xor    ("eax","esi");
 249
 250         &shr    ("ecx",28-7);   # lo>>28
 251         &xor    ("ebx","edi");
 252         &shr    ("edx",28-7);   # hi>>28
 253         &xor    ("eax","ecx");
 254         &shl    ("esi",30-25);  # lo<<30
 255         &xor    ("ebx","edx");
 256         &shl    ("edi",30-25);  # hi<<30
 257         &xor    ("eax","esi");
 258         &xor    ("ebx","edi");                  # Sigma0(a)
 259
 260         &mov    ("ecx",$Alo);
 261         &mov    ("edx",$Ahi);
 262         &mov    ("esi",$Blo);
 263         &mov    ("edi",$Bhi);
 264         &add    ("eax",$Tlo);
 265         &adc    ("ebx",$Thi);                   # T1 = Sigma0(a)+T1
 266         &or     ("ecx","esi");
 267         &or     ("edx","edi");
 268         &and    ("ecx",$Clo);
 269         &and    ("edx",$Chi);
 270         &and    ("esi",$Alo);
 271         &and    ("edi",$Ahi);
 272         &or     ("ecx","esi");
 273         &or     ("edx","edi");                  # Maj(a,b,c) = ((a|b)&c)|(a&b)
 274
 275         &add    ("eax","ecx");
 276         &adc    ("ebx","edx");                  # T1 += Maj(a,b,c)
 277         &mov    ($Tlo,"eax");
 278         &mov    ($Thi,"ebx");
 279
 280         &mov    (&LB("edx"),&BP(0,$K512));      # pre-fetch LSB of *K
 281         &sub    ("esp",8);
 282         &lea    ($K512,&DWP(8,$K512));          # K++
 283 }
 284
 285
 286 &function_begin("sha512_block_data_order");
 287         &mov    ("esi",wparam(0));      # ctx
 288         &mov    ("edi",wparam(1));      # inp
 289         &mov    ("eax",wparam(2));      # num
 290         &mov    ("ebx","esp");          # saved sp
 291
 292         &call   (&label("pic_point"));  # make it PIC!
 293 &set_label("pic_point");
 294         &blindpop($K512);
 295         &lea    ($K512,&DWP(&label("K512")."-".&label("pic_point"),$K512));
 296
 297         &sub    ("esp",16);
 298         &and    ("esp",-64);
 299
 300         &shl    ("eax",7);
 301         &add    ("eax","edi");
 302         &mov    (&DWP(0,"esp"),"esi");  # ctx
 303         &mov    (&DWP(4,"esp"),"edi");  # inp
 304         &mov    (&DWP(8,"esp"),"eax");  # inp+num*128
 305         &mov    (&DWP(12,"esp"),"ebx"); # saved sp
 306
 307 if ($sse2) {
 308         &picmeup("edx","OPENSSL_ia32cap_P",$K512,&label("K512"));
 309         &mov    ("ecx",&DWP(0,"edx"));
 310         &test   ("ecx",1<<26);
 311         &jz     (&label("loop_x86"));
 312
 313         &mov    ("edx",&DWP(4,"edx"));
 314
 315         # load ctx->h[0-7]
 316         &movq   ($A,&QWP(0,"esi"));
 317          &and   ("ecx",1<<24);          # XMM registers availability
 318         &movq   ("mm1",&QWP(8,"esi"));
 319          &and   ("edx",1<<9);           # SSSE3 bit
 320         &movq   ($BxC,&QWP(16,"esi"));
 321          &or    ("ecx","edx");
 322         &movq   ("mm3",&QWP(24,"esi"));
 323         &movq   ($E,&QWP(32,"esi"));
 324         &movq   ("mm5",&QWP(40,"esi"));
 325         &movq   ("mm6",&QWP(48,"esi"));
 326         &movq   ("mm7",&QWP(56,"esi"));
 327         &cmp    ("ecx",1<<24|1<<9);
 328         &je     (&label("SSSE3"));
 329         &sub    ("esp",8*10);
 330         &jmp    (&label("loop_sse2"));
 331
 332 &set_label("loop_sse2",16);
 333         #&movq  ($Asse2,$A);
 334         &movq   ($Bsse2,"mm1");
 335         &movq   ($Csse2,$BxC);
 336         &movq   ($Dsse2,"mm3");
 337         #&movq  ($Esse2,$E);
 338         &movq   ($Fsse2,"mm5");
 339         &movq   ($Gsse2,"mm6");
 340         &pxor   ($BxC,"mm1");                   # magic
 341         &movq   ($Hsse2,"mm7");
 342         &movq   ("mm3",$A);                     # magic
 343
 344         &mov    ("eax",&DWP(0,"edi"));
 345         &mov    ("ebx",&DWP(4,"edi"));
 346         &add    ("edi",8);
 347         &mov    ("edx",15);                     # counter
 348         &bswap  ("eax");
 349         &bswap  ("ebx");
 350         &jmp    (&label("00_14_sse2"));
 351
 352 &set_label("00_14_sse2",16);
 353         &movd   ("mm1","eax");
 354         &mov    ("eax",&DWP(0,"edi"));
 355         &movd   ("mm7","ebx");
 356         &mov    ("ebx",&DWP(4,"edi"));
 357         &add    ("edi",8);
 358         &bswap  ("eax");
 359         &bswap  ("ebx");
 360         &punpckldq("mm7","mm1");
 361
 362         &BODY_00_15_sse2();
 363
 364         &dec    ("edx");
 365         &jnz    (&label("00_14_sse2"));
 366
 367         &movd   ("mm1","eax");
 368         &movd   ("mm7","ebx");
 369         &punpckldq("mm7","mm1");
 370
 371         &BODY_00_15_sse2(1);
 372
 373         &pxor   ($A,$A);                        # A is in %mm3
 374         &mov    ("edx",32);                     # counter
 375         &jmp    (&label("16_79_sse2"));
 376
 377 &set_label("16_79_sse2",16);
 378     for ($j=0;$j<2;$j++) {                      # 2x unroll
 379         #&movq  ("mm7",&QWP(8*(9+16-1),"esp")); # prefetched in BODY_00_15
 380         &movq   ("mm5",&QWP(8*(9+16-14),"esp"));
 381         &movq   ("mm1","mm7");
 382         &psrlq  ("mm7",1);
 383          &movq  ("mm6","mm5");
 384          &psrlq ("mm5",6);
 385         &psllq  ("mm1",56);
 386          &paddq ($A,"mm3");                     # from BODY_00_15
 387          &movq  ("mm3","mm7");
 388         &psrlq  ("mm7",7-1);
 389          &pxor  ("mm3","mm1");
 390          &psllq ("mm1",63-56);
 391         &pxor   ("mm3","mm7");
 392          &psrlq ("mm7",8-7);
 393         &pxor   ("mm3","mm1");
 394          &movq  ("mm1","mm5");
 395          &psrlq ("mm5",19-6);
 396         &pxor   ("mm7","mm3");                  # sigma0
 397
 398          &psllq ("mm6",3);
 399          &pxor  ("mm1","mm5");
 400         &paddq  ("mm7",&QWP(8*(9+16),"esp"));
 401          &pxor  ("mm1","mm6");
 402          &psrlq ("mm5",61-19);
 403         &paddq  ("mm7",&QWP(8*(9+16-9),"esp"));
 404          &pxor  ("mm1","mm5");
 405          &psllq ("mm6",45-3);
 406         &movq   ("mm5",$Fsse2);                 # load f
 407          &pxor  ("mm1","mm6");                  # sigma1
 408         &movq   ("mm6",$Gsse2);                 # load g
 409
 410         &paddq  ("mm7","mm1");                  # X[i]
 411         #&movq  (&QWP(8*9,"esp"),"mm7");        # moved to BODY_00_15
 412
 413         &BODY_00_15_sse2(2);
 414     }
 415         &dec    ("edx");
 416         &jnz    (&label("16_79_sse2"));
 417
 418         #&movq  ($A,$Asse2);
 419         &paddq  ($A,"mm3");                     # from BODY_00_15
 420         &movq   ("mm1",$Bsse2);
 421         #&movq  ($BxC,$Csse2);
 422         &movq   ("mm3",$Dsse2);
 423         #&movq  ($E,$Esse2);
 424         &movq   ("mm5",$Fsse2);
 425         &movq   ("mm6",$Gsse2);
 426         &movq   ("mm7",$Hsse2);
 427
 428         &pxor   ($BxC,"mm1");                   # de-magic
 429         &paddq  ($A,&QWP(0,"esi"));
 430         &paddq  ("mm1",&QWP(8,"esi"));
 431         &paddq  ($BxC,&QWP(16,"esi"));
 432         &paddq  ("mm3",&QWP(24,"esi"));
 433         &paddq  ($E,&QWP(32,"esi"));
 434         &paddq  ("mm5",&QWP(40,"esi"));
 435         &paddq  ("mm6",&QWP(48,"esi"));
 436         &paddq  ("mm7",&QWP(56,"esi"));
 437
 438         &mov    ("eax",8*80);
 439         &movq   (&QWP(0,"esi"),$A);
 440         &movq   (&QWP(8,"esi"),"mm1");
 441         &movq   (&QWP(16,"esi"),$BxC);
 442         &movq   (&QWP(24,"esi"),"mm3");
 443         &movq   (&QWP(32,"esi"),$E);
 444         &movq   (&QWP(40,"esi"),"mm5");
 445         &movq   (&QWP(48,"esi"),"mm6");
 446         &movq   (&QWP(56,"esi"),"mm7");
 447
 448         &lea    ("esp",&DWP(0,"esp","eax"));    # destroy frame
 449         &sub    ($K512,"eax");                  # rewind K
 450
 451         &cmp    ("edi",&DWP(8*10+8,"esp"));     # are we done yet?
 452         &jb     (&label("loop_sse2"));
 453
 454         &mov    ("esp",&DWP(8*10+12,"esp"));    # restore sp
 455         &emms   ();
 456 &function_end_A();
 457
 458 &set_label("SSSE3",32);
 459 { my ($cnt,$frame)=("ecx","edx");
 460   my @X=map("xmm$_",(0..7));
 461   my $j;
 462   my $i=0;
 463
 464         &lea    ($frame,&DWP(-64,"esp"));
 465         &sub    ("esp",256);
 466
 467         # fixed stack frame layout
 468         #
 469         # +0    A B C D E F G H         # backing store
 470         # +64   X[0]+K[i] .. X[15]+K[i] # XMM->MM xfer area
 471         # +192                          # XMM off-load ring buffer
 472         # +256                          # saved parameters
 473
 474         &movdqa         (@X[1],&QWP(80*8,$K512));               # byte swap mask
 475         &movdqu         (@X[0],&QWP(0,"edi"));
 476         &pshufb         (@X[0],@X[1]);
 477     for ($j=0;$j<8;$j++) {
 478         &movdqa         (&QWP(16*(($j-1)%4),$frame),@X[3])      if ($j>4); # off-load
 479         &movdqa         (@X[3],&QWP(16*($j%8),$K512));
 480         &movdqa         (@X[2],@X[1])                           if ($j<7); # perpetuate byte swap mask
 481         &movdqu         (@X[1],&QWP(16*($j+1),"edi"))           if ($j<7); # next input
 482         &movdqa         (@X[1],&QWP(16*(($j+1)%4),$frame))      if ($j==7);# restore @X[0]
 483         &paddq          (@X[3],@X[0]);
 484         &pshufb         (@X[1],@X[2])                           if ($j<7);
 485         &movdqa         (&QWP(16*($j%8)-128,$frame),@X[3]);     # xfer X[i]+K[i]
 486
 487         push(@X,shift(@X));                                     # rotate(@X)
 488     }
 489         #&jmp           (&label("loop_ssse3"));
 490         &nop            ();
 491
 492 &set_label("loop_ssse3",32);
 493         &movdqa         (@X[2],&QWP(16*(($j+1)%4),$frame));     # pre-restore @X[1]
 494         &movdqa         (&QWP(16*(($j-1)%4),$frame),@X[3]);     # off-load @X[3]
 495         &lea            ($K512,&DWP(16*8,$K512));
 496
 497         #&movq  ($Asse2,$A);                    # off-load A-H
 498         &movq   ($Bsse2,"mm1");
 499          &mov   ("ebx","edi");
 500         &movq   ($Csse2,$BxC);
 501          &lea   ("edi",&DWP(128,"edi"));        # advance input
 502         &movq   ($Dsse2,"mm3");
 503          &cmp   ("edi","eax");
 504         #&movq  ($Esse2,$E);
 505         &movq   ($Fsse2,"mm5");
 506          &cmovb ("ebx","edi");
 507         &movq   ($Gsse2,"mm6");
 508          &mov   ("ecx",4);                      # loop counter
 509         &pxor   ($BxC,"mm1");                   # magic
 510         &movq   ($Hsse2,"mm7");
 511         &pxor   ("mm3","mm3");                  # magic
 512
 513         &jmp            (&label("00_47_ssse3"));
 514
 515 sub BODY_00_15_ssse3 {          # "phase-less" copy of BODY_00_15_sse2
 516         (
 517         '&movq  ("mm1",$E)',                            # %mm1 is sliding right
 518         '&movq  ("mm7",&QWP(((-8*$i)%128)-128,$frame))',# X[i]+K[i]
 519          '&pxor ("mm5","mm6")',                         # f^=g
 520         '&psrlq ("mm1",14)',
 521          '&movq (&QWP(8*($i+4)%64,"esp"),$E)',          # modulo-scheduled save e
 522          '&pand ("mm5",$E)',                            # f&=e
 523         '&psllq ($E,23)',                               # $E is sliding left
 524         '&paddq ($A,"mm3")',                            # [h+=Maj(a,b,c)]
 525         '&movq  ("mm3","mm1")',                         # %mm3 is T1
 526          '&psrlq("mm1",4)',
 527          '&pxor ("mm5","mm6")',                         # Ch(e,f,g)
 528         '&pxor  ("mm3",$E)',
 529          '&psllq($E,23)',
 530         '&pxor  ("mm3","mm1")',
 531          '&movq (&QWP(8*$i%64,"esp"),$A)',              # modulo-scheduled save a
 532          '&paddq("mm7","mm5")',                         # X[i]+=Ch(e,f,g)
 533         '&pxor  ("mm3",$E)',
 534          '&psrlq("mm1",23)',
 535          '&paddq("mm7",&QWP(8*($i+7)%64,"esp"))',       # X[i]+=h
 536         '&pxor  ("mm3","mm1")',
 537          '&psllq($E,4)',
 538         '&pxor  ("mm3",$E)',                            # T1=Sigma1_512(e)
 539
 540          '&movq ($E,&QWP(8*($i+3)%64,"esp"))',          # e = load d, e in next round
 541         '&paddq ("mm3","mm7")',                         # T1+=X[i]
 542          '&movq ("mm5",$A)',                            # %mm5 is sliding right
 543          '&psrlq("mm5",28)',
 544         '&paddq ($E,"mm3")',                            # d += T1
 545          '&movq ("mm6",$A)',                            # %mm6 is sliding left
 546          '&movq ("mm7","mm5")',
 547          '&psllq("mm6",25)',
 548         '&movq  ("mm1",&QWP(8*($i+1)%64,"esp"))',       # load b
 549          '&psrlq("mm5",6)',
 550          '&pxor ("mm7","mm6")',
 551          '&psllq("mm6",5)',
 552          '&pxor ("mm7","mm5")',
 553         '&pxor  ($A,"mm1")',                            # a^b, b^c in next round
 554          '&psrlq("mm5",5)',
 555          '&pxor ("mm7","mm6")',
 556         '&pand  ($BxC,$A)',                             # (b^c)&(a^b)
 557          '&psllq("mm6",6)',
 558          '&pxor ("mm7","mm5")',
 559         '&pxor  ($BxC,"mm1")',                          # [h=]Maj(a,b,c)
 560          '&pxor ("mm6","mm7")',                         # Sigma0_512(a)
 561          '&movq ("mm5",&QWP(8*($i+5-1)%64,"esp"))',     # pre-load f
 562         '&paddq ($BxC,"mm6")',                          # h+=Sigma0(a)
 563          '&movq ("mm6",&QWP(8*($i+6-1)%64,"esp"))',     # pre-load g
 564
 565         '($A,$BxC) = ($BxC,$A); $i--;'
 566         );
 567 }
 568
 569 &set_label("00_47_ssse3",32);
 570
 571     for(;$j<16;$j++) {
 572         my ($t0,$t2,$t1)=@X[2..4];
 573         my @insns = (&BODY_00_15_ssse3(),&BODY_00_15_ssse3());
 574
 575         &movdqa         ($t2,@X[5]);
 576         &movdqa         (@X[1],$t0);                    # restore @X[1]
 577         &palignr        ($t0,@X[0],8);                  # X[1..2]
 578         &movdqa         (&QWP(16*($j%4),$frame),@X[4]); # off-load @X[4]
 579          &palignr       ($t2,@X[4],8);                  # X[9..10]
 580
 581         &movdqa         ($t1,$t0);
 582         &psrlq          ($t0,7);
 583          &paddq         (@X[0],$t2);                    # X[0..1] += X[9..10]
 584         &movdqa         ($t2,$t1);
 585         &psrlq          ($t1,1);
 586         &psllq          ($t2,64-8);
 587         &pxor           ($t0,$t1);
 588         &psrlq          ($t1,8-1);
 589         &pxor           ($t0,$t2);
 590         &psllq          ($t2,8-1);
 591         &pxor           ($t0,$t1);
 592          &movdqa        ($t1,@X[7]);
 593         &pxor           ($t0,$t2);                      # sigma0(X[1..2])
 594          &movdqa        ($t2,@X[7]);
 595          &psrlq         ($t1,6);
 596         &paddq          (@X[0],$t0);                    # X[0..1] += sigma0(X[1..2])
 597
 598         &movdqa         ($t0,@X[7]);
 599         &psrlq          ($t2,19);
 600         &psllq          ($t0,64-61);
 601         &pxor           ($t1,$t2);
 602         &psrlq          ($t2,61-19);
 603         &pxor           ($t1,$t0);
 604         &psllq          ($t0,61-19);
 605         &pxor           ($t1,$t2);
 606         &movdqa         ($t2,&QWP(16*(($j+2)%4),$frame));# pre-restore @X[1]
 607         &pxor           ($t1,$t0);                      # sigma0(X[1..2])
 608         &movdqa         ($t0,&QWP(16*($j%8),$K512));
 609          eval(shift(@insns));
 610         &paddq          (@X[0],$t1);                    # X[0..1] += sigma0(X[14..15])
 611          eval(shift(@insns));
 612          eval(shift(@insns));
 613          eval(shift(@insns));
 614          eval(shift(@insns));
 615         &paddq          ($t0,@X[0]);
 616          foreach(@insns) { eval; }
 617         &movdqa         (&QWP(16*($j%8)-128,$frame),$t0);# xfer X[i]+K[i]
 618
 619         push(@X,shift(@X));                             # rotate(@X)
 620     }
 621         &lea            ($K512,&DWP(16*8,$K512));
 622         &dec            ("ecx");
 623         &jnz            (&label("00_47_ssse3"));
 624
 625         &movdqa         (@X[1],&QWP(0,$K512));          # byte swap mask
 626         &lea            ($K512,&DWP(-80*8,$K512));      # rewind
 627         &movdqu         (@X[0],&QWP(0,"ebx"));
 628         &pshufb         (@X[0],@X[1]);
 629
 630     for ($j=0;$j<8;$j++) {      # load next or same block
 631         my @insns = (&BODY_00_15_ssse3(),&BODY_00_15_ssse3());
 632
 633         &movdqa         (&QWP(16*(($j-1)%4),$frame),@X[3])      if ($j>4); # off-load
 634         &movdqa         (@X[3],&QWP(16*($j%8),$K512));
 635         &movdqa         (@X[2],@X[1])                           if ($j<7); # perpetuate byte swap mask
 636         &movdqu         (@X[1],&QWP(16*($j+1),"ebx"))           if ($j<7); # next input
 637         &movdqa         (@X[1],&QWP(16*(($j+1)%4),$frame))      if ($j==7);# restore @X[0]
 638         &paddq          (@X[3],@X[0]);
 639         &pshufb         (@X[1],@X[2])                           if ($j<7);
 640          foreach(@insns) { eval; }
 641         &movdqa         (&QWP(16*($j%8)-128,$frame),@X[3]);# xfer X[i]+K[i]
 642
 643         push(@X,shift(@X));                             # rotate(@X)
 644     }
 645
 646         #&movq  ($A,$Asse2);                    # load A-H
 647         &movq   ("mm1",$Bsse2);
 648         &paddq  ($A,"mm3");                     # from BODY_00_15
 649         #&movq  ($BxC,$Csse2);
 650         &movq   ("mm3",$Dsse2);
 651         #&movq  ($E,$Esse2);
 652         #&movq  ("mm5",$Fsse2);
 653         #&movq  ("mm6",$Gsse2);
 654         &movq   ("mm7",$Hsse2);
 655
 656         &pxor   ($BxC,"mm1");                   # de-magic
 657         &paddq  ($A,&QWP(0,"esi"));
 658         &paddq  ("mm1",&QWP(8,"esi"));
 659         &paddq  ($BxC,&QWP(16,"esi"));
 660         &paddq  ("mm3",&QWP(24,"esi"));
 661         &paddq  ($E,&QWP(32,"esi"));
 662         &paddq  ("mm5",&QWP(40,"esi"));
 663         &paddq  ("mm6",&QWP(48,"esi"));
 664         &paddq  ("mm7",&QWP(56,"esi"));
 665
 666         &movq   (&QWP(0,"esi"),$A);
 667         &movq   (&QWP(8,"esi"),"mm1");
 668         &movq   (&QWP(16,"esi"),$BxC);
 669         &movq   (&QWP(24,"esi"),"mm3");
 670         &movq   (&QWP(32,"esi"),$E);
 671         &movq   (&QWP(40,"esi"),"mm5");
 672         &movq   (&QWP(48,"esi"),"mm6");
 673         &movq   (&QWP(56,"esi"),"mm7");
 674
 675         &cmp    ("edi","eax")                   # are we done yet?
 676         &jb     (&label("loop_ssse3"));
 677
 678         &mov    ("esp",&DWP(64+12,$frame));     # restore sp
 679         &emms   ();
 680 }
 681 &function_end_A();
 682 }
 683 &set_label("loop_x86",16);
 684     # copy input block to stack reversing byte and qword order
 685     for ($i=0;$i<8;$i++) {
 686         &mov    ("eax",&DWP($i*16+0,"edi"));
 687         &mov    ("ebx",&DWP($i*16+4,"edi"));
 688         &mov    ("ecx",&DWP($i*16+8,"edi"));
 689         &mov    ("edx",&DWP($i*16+12,"edi"));
 690         &bswap  ("eax");
 691         &bswap  ("ebx");
 692         &bswap  ("ecx");
 693         &bswap  ("edx");
 694         &push   ("eax");
 695         &push   ("ebx");
 696         &push   ("ecx");
 697         &push   ("edx");
 698     }
 699         &add    ("edi",128);
 700         &sub    ("esp",9*8);            # place for T,A,B,C,D,E,F,G,H
 701         &mov    (&DWP(8*(9+16)+4,"esp"),"edi");
 702
 703         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 704         &lea    ("edi",&DWP(8,"esp"));
 705         &mov    ("ecx",16);
 706         &data_word(0xA5F3F689);         # rep movsd
 707
 708 &set_label("00_15_x86",16);
 709         &BODY_00_15_x86();
 710
 711         &cmp    (&LB("edx"),0x94);
 712         &jne    (&label("00_15_x86"));
 713
 714 &set_label("16_79_x86",16);
 715         #define sigma0(x)       (ROTR((x),1)  ^ ROTR((x),8)  ^ ((x)>>7))
 716         #       LO              lo>>1^hi<<31  ^ lo>>8^hi<<24 ^ lo>>7^hi<<25
 717         #       HI              hi>>1^lo<<31  ^ hi>>8^lo<<24 ^ hi>>7
 718         &mov    ("ecx",&DWP(8*(9+15+16-1)+0,"esp"));
 719         &mov    ("edx",&DWP(8*(9+15+16-1)+4,"esp"));
 720         &mov    ("esi","ecx");
 721
 722         &shr    ("ecx",1);      # lo>>1
 723         &mov    ("edi","edx");
 724         &shr    ("edx",1);      # hi>>1
 725         &mov    ("eax","ecx");
 726         &shl    ("esi",24);     # lo<<24
 727         &mov    ("ebx","edx");
 728         &shl    ("edi",24);     # hi<<24
 729         &xor    ("ebx","esi");
 730
 731         &shr    ("ecx",7-1);    # lo>>7
 732         &xor    ("eax","edi");
 733         &shr    ("edx",7-1);    # hi>>7
 734         &xor    ("eax","ecx");
 735         &shl    ("esi",31-24);  # lo<<31
 736         &xor    ("ebx","edx");
 737         &shl    ("edi",25-24);  # hi<<25
 738         &xor    ("ebx","esi");
 739
 740         &shr    ("ecx",8-7);    # lo>>8
 741         &xor    ("eax","edi");
 742         &shr    ("edx",8-7);    # hi>>8
 743         &xor    ("eax","ecx");
 744         &shl    ("edi",31-25);  # hi<<31
 745         &xor    ("ebx","edx");
 746         &xor    ("eax","edi");                  # T1 = sigma0(X[-15])
 747
 748         &mov    (&DWP(0,"esp"),"eax");
 749         &mov    (&DWP(4,"esp"),"ebx");          # put T1 away
 750
 751         #define sigma1(x)       (ROTR((x),19) ^ ROTR((x),61) ^ ((x)>>6))
 752         #       LO              lo>>19^hi<<13 ^ hi>>29^lo<<3 ^ lo>>6^hi<<26
 753         #       HI              hi>>19^lo<<13 ^ lo>>29^hi<<3 ^ hi>>6
 754         &mov    ("ecx",&DWP(8*(9+15+16-14)+0,"esp"));
 755         &mov    ("edx",&DWP(8*(9+15+16-14)+4,"esp"));
 756         &mov    ("esi","ecx");
 757
 758         &shr    ("ecx",6);      # lo>>6
 759         &mov    ("edi","edx");
 760         &shr    ("edx",6);      # hi>>6
 761         &mov    ("eax","ecx");
 762         &shl    ("esi",3);      # lo<<3
 763         &mov    ("ebx","edx");
 764         &shl    ("edi",3);      # hi<<3
 765         &xor    ("eax","esi");
 766
 767         &shr    ("ecx",19-6);   # lo>>19
 768         &xor    ("ebx","edi");
 769         &shr    ("edx",19-6);   # hi>>19
 770         &xor    ("eax","ecx");
 771         &shl    ("esi",13-3);   # lo<<13
 772         &xor    ("ebx","edx");
 773         &shl    ("edi",13-3);   # hi<<13
 774         &xor    ("ebx","esi");
 775
 776         &shr    ("ecx",29-19);  # lo>>29
 777         &xor    ("eax","edi");
 778         &shr    ("edx",29-19);  # hi>>29
 779         &xor    ("ebx","ecx");
 780         &shl    ("edi",26-13);  # hi<<26
 781         &xor    ("eax","edx");
 782         &xor    ("eax","edi");                  # sigma1(X[-2])
 783
 784         &mov    ("ecx",&DWP(8*(9+15+16)+0,"esp"));
 785         &mov    ("edx",&DWP(8*(9+15+16)+4,"esp"));
 786         &add    ("eax",&DWP(0,"esp"));
 787         &adc    ("ebx",&DWP(4,"esp"));          # T1 = sigma1(X[-2])+T1
 788         &mov    ("esi",&DWP(8*(9+15+16-9)+0,"esp"));
 789         &mov    ("edi",&DWP(8*(9+15+16-9)+4,"esp"));
 790         &add    ("eax","ecx");
 791         &adc    ("ebx","edx");                  # T1 += X[-16]
 792         &add    ("eax","esi");
 793         &adc    ("ebx","edi");                  # T1 += X[-7]
 794         &mov    (&DWP(8*(9+15)+0,"esp"),"eax");
 795         &mov    (&DWP(8*(9+15)+4,"esp"),"ebx"); # save X[0]
 796
 797         &BODY_00_15_x86();
 798
 799         &cmp    (&LB("edx"),0x17);
 800         &jne    (&label("16_79_x86"));
 801
 802         &mov    ("esi",&DWP(8*(9+16+80)+0,"esp"));# ctx
 803         &mov    ("edi",&DWP(8*(9+16+80)+4,"esp"));# inp
 804     for($i=0;$i<4;$i++) {
 805         &mov    ("eax",&DWP($i*16+0,"esi"));
 806         &mov    ("ebx",&DWP($i*16+4,"esi"));
 807         &mov    ("ecx",&DWP($i*16+8,"esi"));
 808         &mov    ("edx",&DWP($i*16+12,"esi"));
 809         &add    ("eax",&DWP(8+($i*16)+0,"esp"));
 810         &adc    ("ebx",&DWP(8+($i*16)+4,"esp"));
 811         &mov    (&DWP($i*16+0,"esi"),"eax");
 812         &mov    (&DWP($i*16+4,"esi"),"ebx");
 813         &add    ("ecx",&DWP(8+($i*16)+8,"esp"));
 814         &adc    ("edx",&DWP(8+($i*16)+12,"esp"));
 815         &mov    (&DWP($i*16+8,"esi"),"ecx");
 816         &mov    (&DWP($i*16+12,"esi"),"edx");
 817     }
 818         &add    ("esp",8*(9+16+80));            # destroy frame
 819         &sub    ($K512,8*80);                   # rewind K
 820
 821         &cmp    ("edi",&DWP(8,"esp"));          # are we done yet?
 822         &jb     (&label("loop_x86"));
 823
 824         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 825 &function_end_A();
 826
 827 &set_label("K512",64);  # Yes! I keep it in the code segment!
 828         &data_word(0xd728ae22,0x428a2f98);      # u64
 829         &data_word(0x23ef65cd,0x71374491);      # u64
 830         &data_word(0xec4d3b2f,0xb5c0fbcf);      # u64
 831         &data_word(0x8189dbbc,0xe9b5dba5);      # u64
 832         &data_word(0xf348b538,0x3956c25b);      # u64
 833         &data_word(0xb605d019,0x59f111f1);      # u64
 834         &data_word(0xaf194f9b,0x923f82a4);      # u64
 835         &data_word(0xda6d8118,0xab1c5ed5);      # u64
 836         &data_word(0xa3030242,0xd807aa98);      # u64
 837         &data_word(0x45706fbe,0x12835b01);      # u64
 838         &data_word(0x4ee4b28c,0x243185be);      # u64
 839         &data_word(0xd5ffb4e2,0x550c7dc3);      # u64
 840         &data_word(0xf27b896f,0x72be5d74);      # u64
 841         &data_word(0x3b1696b1,0x80deb1fe);      # u64
 842         &data_word(0x25c71235,0x9bdc06a7);      # u64
 843         &data_word(0xcf692694,0xc19bf174);      # u64
 844         &data_word(0x9ef14ad2,0xe49b69c1);      # u64
 845         &data_word(0x384f25e3,0xefbe4786);      # u64
 846         &data_word(0x8b8cd5b5,0x0fc19dc6);      # u64
 847         &data_word(0x77ac9c65,0x240ca1cc);      # u64
 848         &data_word(0x592b0275,0x2de92c6f);      # u64
 849         &data_word(0x6ea6e483,0x4a7484aa);      # u64
 850         &data_word(0xbd41fbd4,0x5cb0a9dc);      # u64
 851         &data_word(0x831153b5,0x76f988da);      # u64
 852         &data_word(0xee66dfab,0x983e5152);      # u64
 853         &data_word(0x2db43210,0xa831c66d);      # u64
 854         &data_word(0x98fb213f,0xb00327c8);      # u64
 855         &data_word(0xbeef0ee4,0xbf597fc7);      # u64
 856         &data_word(0x3da88fc2,0xc6e00bf3);      # u64
 857         &data_word(0x930aa725,0xd5a79147);      # u64
 858         &data_word(0xe003826f,0x06ca6351);      # u64
 859         &data_word(0x0a0e6e70,0x14292967);      # u64
 860         &data_word(0x46d22ffc,0x27b70a85);      # u64
 861         &data_word(0x5c26c926,0x2e1b2138);      # u64
 862         &data_word(0x5ac42aed,0x4d2c6dfc);      # u64
 863         &data_word(0x9d95b3df,0x53380d13);      # u64
 864         &data_word(0x8baf63de,0x650a7354);      # u64
 865         &data_word(0x3c77b2a8,0x766a0abb);      # u64
 866         &data_word(0x47edaee6,0x81c2c92e);      # u64
 867         &data_word(0x1482353b,0x92722c85);      # u64
 868         &data_word(0x4cf10364,0xa2bfe8a1);      # u64
 869         &data_word(0xbc423001,0xa81a664b);      # u64
 870         &data_word(0xd0f89791,0xc24b8b70);      # u64
 871         &data_word(0x0654be30,0xc76c51a3);      # u64
 872         &data_word(0xd6ef5218,0xd192e819);      # u64
 873         &data_word(0x5565a910,0xd6990624);      # u64
 874         &data_word(0x5771202a,0xf40e3585);      # u64
 875         &data_word(0x32bbd1b8,0x106aa070);      # u64
 876         &data_word(0xb8d2d0c8,0x19a4c116);      # u64
 877         &data_word(0x5141ab53,0x1e376c08);      # u64
 878         &data_word(0xdf8eeb99,0x2748774c);      # u64
 879         &data_word(0xe19b48a8,0x34b0bcb5);      # u64
 880         &data_word(0xc5c95a63,0x391c0cb3);      # u64
 881         &data_word(0xe3418acb,0x4ed8aa4a);      # u64
 882         &data_word(0x7763e373,0x5b9cca4f);      # u64
 883         &data_word(0xd6b2b8a3,0x682e6ff3);      # u64
 884         &data_word(0x5defb2fc,0x748f82ee);      # u64
 885         &data_word(0x43172f60,0x78a5636f);      # u64
 886         &data_word(0xa1f0ab72,0x84c87814);      # u64
 887         &data_word(0x1a6439ec,0x8cc70208);      # u64
 888         &data_word(0x23631e28,0x90befffa);      # u64
 889         &data_word(0xde82bde9,0xa4506ceb);      # u64
 890         &data_word(0xb2c67915,0xbef9a3f7);      # u64
 891         &data_word(0xe372532b,0xc67178f2);      # u64
 892         &data_word(0xea26619c,0xca273ece);      # u64
 893         &data_word(0x21c0c207,0xd186b8c7);      # u64
 894         &data_word(0xcde0eb1e,0xeada7dd6);      # u64
 895         &data_word(0xee6ed178,0xf57d4f7f);      # u64
 896         &data_word(0x72176fba,0x06f067aa);      # u64
 897         &data_word(0xa2c898a6,0x0a637dc5);      # u64
 898         &data_word(0xbef90dae,0x113f9804);      # u64
 899         &data_word(0x131c471b,0x1b710b35);      # u64
 900         &data_word(0x23047d84,0x28db77f5);      # u64
 901         &data_word(0x40c72493,0x32caab7b);      # u64
 902         &data_word(0x15c9bebc,0x3c9ebe0a);      # u64
 903         &data_word(0x9c100d4c,0x431d67c4);      # u64
 904         &data_word(0xcb3e42b6,0x4cc5d4be);      # u64
 905         &data_word(0xfc657e2a,0x597f299c);      # u64
 906         &data_word(0x3ad6faec,0x5fcb6fab);      # u64
 907         &data_word(0x4a475817,0x6c44198c);      # u64
 908
 909         &data_word(0x04050607,0x00010203);      # byte swap
 910         &data_word(0x0c0d0e0f,0x08090a0b);      # mask
 911 &function_end_B("sha512_block_data_order");
 912 &asciz("SHA512 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
 913
 914 &asm_finish();
 915
 916 close STDOUT;