crypto/sha/asm/sha512-586.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2007-2016 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the OpenSSL license (the "License").  You may not use
   5 # this file except in compliance with the License.  You can obtain a copy
   6 # in the file LICENSE in the source distribution or at
   7 # https://www.openssl.org/source/license.html
   8
   9 #
  10 # ====================================================================
  11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
  12 # project. The module is, however, dual licensed under OpenSSL and
  13 # CRYPTOGAMS licenses depending on where you obtain it. For further
  14 # details see http://www.openssl.org/~appro/cryptogams/.
  15 # ====================================================================
  16 #
  17 # SHA512 block transform for x86. September 2007.
  18 #
  19 # May 2013.
  20 #
  21 # Add SSSE3 code path, 20-25% improvement [over original SSE2 code].
  22 #
  23 # Performance in clock cycles per processed byte (less is better):
  24 #
  25 #               gcc     icc     x86 asm SIMD(*) x86_64(**)
  26 # Pentium       100     97      61      -       -
  27 # PIII          75      77      56      -       -
  28 # P4            116     95      82      34.6    30.8
  29 # AMD K8        54      55      36      20.7    9.57
  30 # Core2         66      57      40      15.9    9.97
  31 # Westmere      70      -       38      12.2    9.58
  32 # Sandy Bridge  58      -       35      11.9    11.2
  33 # Ivy Bridge    50      -       33      11.5    8.17
  34 # Haswell       46      -       29      11.3    7.66
  35 # Bulldozer     121     -       50      14.0    13.5
  36 # VIA Nano      91      -       52      33      14.7
  37 # Atom          126     -       68      48(***) 14.7
  38 # Silvermont    97      -       58      42(***) 17.5
  39 #
  40 # (*)   whichever best applicable.
  41 # (**)  x86_64 assembler performance is presented for reference
  42 #       purposes, the results are for integer-only code.
  43 # (***) paddq is increadibly slow on Atom.
  44 #
  45 # IALU code-path is optimized for elder Pentiums. On vanilla Pentium
  46 # performance improvement over compiler generated code reaches ~60%,
  47 # while on PIII - ~35%. On newer µ-archs improvement varies from 15%
  48 # to 50%, but it's less important as they are expected to execute SSE2
  49 # code-path, which is commonly ~2-3x faster [than compiler generated
  50 # code]. SSE2 code-path is as fast as original sha512-sse2.pl, even
  51 # though it does not use 128-bit operations. The latter means that
  52 # SSE2-aware kernel is no longer required to execute the code. Another
  53 # difference is that new code optimizes amount of writes, but at the
  54 # cost of increased data cache "footprint" by 1/2KB.
  55
  56 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  57 push(@INC,"${dir}","${dir}../../perlasm");
  58 require "x86asm.pl";
  59
  60 $output=pop;
  61 open STDOUT,">$output";
  62
  63 &asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");
  64
  65 $sse2=0;
  66 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
  67
  68 &external_label("OPENSSL_ia32cap_P") if ($sse2);
  69
  70 $Tlo=&DWP(0,"esp");     $Thi=&DWP(4,"esp");
  71 $Alo=&DWP(8,"esp");     $Ahi=&DWP(8+4,"esp");
  72 $Blo=&DWP(16,"esp");    $Bhi=&DWP(16+4,"esp");
  73 $Clo=&DWP(24,"esp");    $Chi=&DWP(24+4,"esp");
  74 $Dlo=&DWP(32,"esp");    $Dhi=&DWP(32+4,"esp");
  75 $Elo=&DWP(40,"esp");    $Ehi=&DWP(40+4,"esp");
  76 $Flo=&DWP(48,"esp");    $Fhi=&DWP(48+4,"esp");
  77 $Glo=&DWP(56,"esp");    $Ghi=&DWP(56+4,"esp");
  78 $Hlo=&DWP(64,"esp");    $Hhi=&DWP(64+4,"esp");
  79 $K512="ebp";
  80
  81 $Asse2=&QWP(0,"esp");
  82 $Bsse2=&QWP(8,"esp");
  83 $Csse2=&QWP(16,"esp");
  84 $Dsse2=&QWP(24,"esp");
  85 $Esse2=&QWP(32,"esp");
  86 $Fsse2=&QWP(40,"esp");
  87 $Gsse2=&QWP(48,"esp");
  88 $Hsse2=&QWP(56,"esp");
  89
  90 $A="mm0";       # B-D and
  91 $E="mm4";       # F-H are commonly loaded to respectively mm1-mm3 and
  92                 # mm5-mm7, but it's done on on-demand basis...
  93 $BxC="mm2";     # ... except for B^C
  94
  95 sub BODY_00_15_sse2 {
  96     my $phase=shift;
  97
  98         #&movq  ("mm5",$Fsse2);                 # load f
  99         #&movq  ("mm6",$Gsse2);                 # load g
 100
 101         &movq   ("mm1",$E);                     # %mm1 is sliding right
 102          &pxor  ("mm5","mm6");                  # f^=g
 103         &psrlq  ("mm1",14);
 104          &movq  ($Esse2,$E);                    # modulo-scheduled save e
 105          &pand  ("mm5",$E);                     # f&=e
 106         &psllq  ($E,23);                        # $E is sliding left
 107          &movq  ($A,"mm3")                      if ($phase<2);
 108          &movq  (&QWP(8*9,"esp"),"mm7")         # save X[i]
 109         &movq   ("mm3","mm1");                  # %mm3 is T1
 110          &psrlq ("mm1",4);
 111          &pxor  ("mm5","mm6");                  # Ch(e,f,g)
 112         &pxor   ("mm3",$E);
 113          &psllq ($E,23);
 114         &pxor   ("mm3","mm1");
 115          &movq  ($Asse2,$A);                    # modulo-scheduled save a
 116          &paddq ("mm7","mm5");                  # X[i]+=Ch(e,f,g)
 117         &pxor   ("mm3",$E);
 118          &psrlq ("mm1",23);
 119          &paddq ("mm7",$Hsse2);                 # X[i]+=h
 120         &pxor   ("mm3","mm1");
 121          &psllq ($E,4);
 122          &paddq ("mm7",QWP(0,$K512));           # X[i]+=K512[i]
 123         &pxor   ("mm3",$E);                     # T1=Sigma1_512(e)
 124
 125          &movq  ($E,$Dsse2);                    # e = load d, e in next round
 126         &paddq  ("mm3","mm7");                  # T1+=X[i]
 127          &movq  ("mm5",$A);                     # %mm5 is sliding right
 128          &psrlq ("mm5",28);
 129         &paddq  ($E,"mm3");                     # d += T1
 130          &movq  ("mm6",$A);                     # %mm6 is sliding left
 131          &movq  ("mm7","mm5");
 132          &psllq ("mm6",25);
 133         &movq   ("mm1",$Bsse2);                 # load b
 134          &psrlq ("mm5",6);
 135          &pxor  ("mm7","mm6");
 136         &sub    ("esp",8);
 137          &psllq ("mm6",5);
 138          &pxor  ("mm7","mm5");
 139         &pxor   ($A,"mm1");                     # a^b, b^c in next round
 140          &psrlq ("mm5",5);
 141          &pxor  ("mm7","mm6");
 142         &pand   ($BxC,$A);                      # (b^c)&(a^b)
 143          &psllq ("mm6",6);
 144          &pxor  ("mm7","mm5");
 145         &pxor   ($BxC,"mm1");                   # [h=]Maj(a,b,c)
 146          &pxor  ("mm6","mm7");                  # Sigma0_512(a)
 147          &movq  ("mm7",&QWP(8*(9+16-1),"esp"))  if ($phase!=0); # pre-fetch
 148          &movq  ("mm5",$Fsse2)                  if ($phase==0); # load f
 149
 150     if ($phase>1) {
 151         &paddq  ($BxC,"mm6");                   # h+=Sigma0(a)
 152          &add   ($K512,8);
 153         #&paddq ($BxC,"mm3");                   # h+=T1
 154
 155         ($A,$BxC) = ($BxC,$A);                  # rotate registers
 156     } else {
 157         &paddq  ("mm3",$BxC);                   # T1+=Maj(a,b,c)
 158          &movq  ($BxC,$A);
 159          &add   ($K512,8);
 160         &paddq  ("mm3","mm6");                  # T1+=Sigma0(a)
 161          &movq  ("mm6",$Gsse2)                  if ($phase==0); # load g
 162         #&movq  ($A,"mm3");                     # h=T1
 163     }
 164 }
 165
 166 sub BODY_00_15_x86 {
 167         #define Sigma1(x)       (ROTR((x),14) ^ ROTR((x),18)  ^ ROTR((x),41))
 168         #       LO              lo>>14^hi<<18 ^ lo>>18^hi<<14 ^ hi>>9^lo<<23
 169         #       HI              hi>>14^lo<<18 ^ hi>>18^lo<<14 ^ lo>>9^hi<<23
 170         &mov    ("ecx",$Elo);
 171         &mov    ("edx",$Ehi);
 172         &mov    ("esi","ecx");
 173
 174         &shr    ("ecx",9);      # lo>>9
 175         &mov    ("edi","edx");
 176         &shr    ("edx",9);      # hi>>9
 177         &mov    ("ebx","ecx");
 178         &shl    ("esi",14);     # lo<<14
 179         &mov    ("eax","edx");
 180         &shl    ("edi",14);     # hi<<14
 181         &xor    ("ebx","esi");
 182
 183         &shr    ("ecx",14-9);   # lo>>14
 184         &xor    ("eax","edi");
 185         &shr    ("edx",14-9);   # hi>>14
 186         &xor    ("eax","ecx");
 187         &shl    ("esi",18-14);  # lo<<18
 188         &xor    ("ebx","edx");
 189         &shl    ("edi",18-14);  # hi<<18
 190         &xor    ("ebx","esi");
 191
 192         &shr    ("ecx",18-14);  # lo>>18
 193         &xor    ("eax","edi");
 194         &shr    ("edx",18-14);  # hi>>18
 195         &xor    ("eax","ecx");
 196         &shl    ("esi",23-18);  # lo<<23
 197         &xor    ("ebx","edx");
 198         &shl    ("edi",23-18);  # hi<<23
 199         &xor    ("eax","esi");
 200         &xor    ("ebx","edi");                  # T1 = Sigma1(e)
 201
 202         &mov    ("ecx",$Flo);
 203         &mov    ("edx",$Fhi);
 204         &mov    ("esi",$Glo);
 205         &mov    ("edi",$Ghi);
 206          &add   ("eax",$Hlo);
 207          &adc   ("ebx",$Hhi);                   # T1 += h
 208         &xor    ("ecx","esi");
 209         &xor    ("edx","edi");
 210         &and    ("ecx",$Elo);
 211         &and    ("edx",$Ehi);
 212          &add   ("eax",&DWP(8*(9+15)+0,"esp"));
 213          &adc   ("ebx",&DWP(8*(9+15)+4,"esp")); # T1 += X[0]
 214         &xor    ("ecx","esi");
 215         &xor    ("edx","edi");                  # Ch(e,f,g) = (f^g)&e)^g
 216
 217         &mov    ("esi",&DWP(0,$K512));
 218         &mov    ("edi",&DWP(4,$K512));          # K[i]
 219         &add    ("eax","ecx");
 220         &adc    ("ebx","edx");                  # T1 += Ch(e,f,g)
 221         &mov    ("ecx",$Dlo);
 222         &mov    ("edx",$Dhi);
 223         &add    ("eax","esi");
 224         &adc    ("ebx","edi");                  # T1 += K[i]
 225         &mov    ($Tlo,"eax");
 226         &mov    ($Thi,"ebx");                   # put T1 away
 227         &add    ("eax","ecx");
 228         &adc    ("ebx","edx");                  # d += T1
 229
 230         #define Sigma0(x)       (ROTR((x),28) ^ ROTR((x),34) ^ ROTR((x),39))
 231         #       LO              lo>>28^hi<<4  ^ hi>>2^lo<<30 ^ hi>>7^lo<<25
 232         #       HI              hi>>28^lo<<4  ^ lo>>2^hi<<30 ^ lo>>7^hi<<25
 233         &mov    ("ecx",$Alo);
 234         &mov    ("edx",$Ahi);
 235         &mov    ($Dlo,"eax");
 236         &mov    ($Dhi,"ebx");
 237         &mov    ("esi","ecx");
 238
 239         &shr    ("ecx",2);      # lo>>2
 240         &mov    ("edi","edx");
 241         &shr    ("edx",2);      # hi>>2
 242         &mov    ("ebx","ecx");
 243         &shl    ("esi",4);      # lo<<4
 244         &mov    ("eax","edx");
 245         &shl    ("edi",4);      # hi<<4
 246         &xor    ("ebx","esi");
 247
 248         &shr    ("ecx",7-2);    # lo>>7
 249         &xor    ("eax","edi");
 250         &shr    ("edx",7-2);    # hi>>7
 251         &xor    ("ebx","ecx");
 252         &shl    ("esi",25-4);   # lo<<25
 253         &xor    ("eax","edx");
 254         &shl    ("edi",25-4);   # hi<<25
 255         &xor    ("eax","esi");
 256
 257         &shr    ("ecx",28-7);   # lo>>28
 258         &xor    ("ebx","edi");
 259         &shr    ("edx",28-7);   # hi>>28
 260         &xor    ("eax","ecx");
 261         &shl    ("esi",30-25);  # lo<<30
 262         &xor    ("ebx","edx");
 263         &shl    ("edi",30-25);  # hi<<30
 264         &xor    ("eax","esi");
 265         &xor    ("ebx","edi");                  # Sigma0(a)
 266
 267         &mov    ("ecx",$Alo);
 268         &mov    ("edx",$Ahi);
 269         &mov    ("esi",$Blo);
 270         &mov    ("edi",$Bhi);
 271         &add    ("eax",$Tlo);
 272         &adc    ("ebx",$Thi);                   # T1 = Sigma0(a)+T1
 273         &or     ("ecx","esi");
 274         &or     ("edx","edi");
 275         &and    ("ecx",$Clo);
 276         &and    ("edx",$Chi);
 277         &and    ("esi",$Alo);
 278         &and    ("edi",$Ahi);
 279         &or     ("ecx","esi");
 280         &or     ("edx","edi");                  # Maj(a,b,c) = ((a|b)&c)|(a&b)
 281
 282         &add    ("eax","ecx");
 283         &adc    ("ebx","edx");                  # T1 += Maj(a,b,c)
 284         &mov    ($Tlo,"eax");
 285         &mov    ($Thi,"ebx");
 286
 287         &mov    (&LB("edx"),&BP(0,$K512));      # pre-fetch LSB of *K
 288         &sub    ("esp",8);
 289         &lea    ($K512,&DWP(8,$K512));          # K++
 290 }
 291
 292
 293 &function_begin("sha512_block_data_order");
 294         &mov    ("esi",wparam(0));      # ctx
 295         &mov    ("edi",wparam(1));      # inp
 296         &mov    ("eax",wparam(2));      # num
 297         &mov    ("ebx","esp");          # saved sp
 298
 299         &call   (&label("pic_point"));  # make it PIC!
 300 &set_label("pic_point");
 301         &blindpop($K512);
 302         &lea    ($K512,&DWP(&label("K512")."-".&label("pic_point"),$K512));
 303
 304         &sub    ("esp",16);
 305         &and    ("esp",-64);
 306
 307         &shl    ("eax",7);
 308         &add    ("eax","edi");
 309         &mov    (&DWP(0,"esp"),"esi");  # ctx
 310         &mov    (&DWP(4,"esp"),"edi");  # inp
 311         &mov    (&DWP(8,"esp"),"eax");  # inp+num*128
 312         &mov    (&DWP(12,"esp"),"ebx"); # saved sp
 313
 314 if ($sse2) {
 315         &picmeup("edx","OPENSSL_ia32cap_P",$K512,&label("K512"));
 316         &mov    ("ecx",&DWP(0,"edx"));
 317         &test   ("ecx",1<<26);
 318         &jz     (&label("loop_x86"));
 319
 320         &mov    ("edx",&DWP(4,"edx"));
 321
 322         # load ctx->h[0-7]
 323         &movq   ($A,&QWP(0,"esi"));
 324          &and   ("ecx",1<<24);          # XMM registers availability
 325         &movq   ("mm1",&QWP(8,"esi"));
 326          &and   ("edx",1<<9);           # SSSE3 bit
 327         &movq   ($BxC,&QWP(16,"esi"));
 328          &or    ("ecx","edx");
 329         &movq   ("mm3",&QWP(24,"esi"));
 330         &movq   ($E,&QWP(32,"esi"));
 331         &movq   ("mm5",&QWP(40,"esi"));
 332         &movq   ("mm6",&QWP(48,"esi"));
 333         &movq   ("mm7",&QWP(56,"esi"));
 334         &cmp    ("ecx",1<<24|1<<9);
 335         &je     (&label("SSSE3"));
 336         &sub    ("esp",8*10);
 337         &jmp    (&label("loop_sse2"));
 338
 339 &set_label("loop_sse2",16);
 340         #&movq  ($Asse2,$A);
 341         &movq   ($Bsse2,"mm1");
 342         &movq   ($Csse2,$BxC);
 343         &movq   ($Dsse2,"mm3");
 344         #&movq  ($Esse2,$E);
 345         &movq   ($Fsse2,"mm5");
 346         &movq   ($Gsse2,"mm6");
 347         &pxor   ($BxC,"mm1");                   # magic
 348         &movq   ($Hsse2,"mm7");
 349         &movq   ("mm3",$A);                     # magic
 350
 351         &mov    ("eax",&DWP(0,"edi"));
 352         &mov    ("ebx",&DWP(4,"edi"));
 353         &add    ("edi",8);
 354         &mov    ("edx",15);                     # counter
 355         &bswap  ("eax");
 356         &bswap  ("ebx");
 357         &jmp    (&label("00_14_sse2"));
 358
 359 &set_label("00_14_sse2",16);
 360         &movd   ("mm1","eax");
 361         &mov    ("eax",&DWP(0,"edi"));
 362         &movd   ("mm7","ebx");
 363         &mov    ("ebx",&DWP(4,"edi"));
 364         &add    ("edi",8);
 365         &bswap  ("eax");
 366         &bswap  ("ebx");
 367         &punpckldq("mm7","mm1");
 368
 369         &BODY_00_15_sse2();
 370
 371         &dec    ("edx");
 372         &jnz    (&label("00_14_sse2"));
 373
 374         &movd   ("mm1","eax");
 375         &movd   ("mm7","ebx");
 376         &punpckldq("mm7","mm1");
 377
 378         &BODY_00_15_sse2(1);
 379
 380         &pxor   ($A,$A);                        # A is in %mm3
 381         &mov    ("edx",32);                     # counter
 382         &jmp    (&label("16_79_sse2"));
 383
 384 &set_label("16_79_sse2",16);
 385     for ($j=0;$j<2;$j++) {                      # 2x unroll
 386         #&movq  ("mm7",&QWP(8*(9+16-1),"esp")); # prefetched in BODY_00_15
 387         &movq   ("mm5",&QWP(8*(9+16-14),"esp"));
 388         &movq   ("mm1","mm7");
 389         &psrlq  ("mm7",1);
 390          &movq  ("mm6","mm5");
 391          &psrlq ("mm5",6);
 392         &psllq  ("mm1",56);
 393          &paddq ($A,"mm3");                     # from BODY_00_15
 394          &movq  ("mm3","mm7");
 395         &psrlq  ("mm7",7-1);
 396          &pxor  ("mm3","mm1");
 397          &psllq ("mm1",63-56);
 398         &pxor   ("mm3","mm7");
 399          &psrlq ("mm7",8-7);
 400         &pxor   ("mm3","mm1");
 401          &movq  ("mm1","mm5");
 402          &psrlq ("mm5",19-6);
 403         &pxor   ("mm7","mm3");                  # sigma0
 404
 405          &psllq ("mm6",3);
 406          &pxor  ("mm1","mm5");
 407         &paddq  ("mm7",&QWP(8*(9+16),"esp"));
 408          &pxor  ("mm1","mm6");
 409          &psrlq ("mm5",61-19);
 410         &paddq  ("mm7",&QWP(8*(9+16-9),"esp"));
 411          &pxor  ("mm1","mm5");
 412          &psllq ("mm6",45-3);
 413         &movq   ("mm5",$Fsse2);                 # load f
 414          &pxor  ("mm1","mm6");                  # sigma1
 415         &movq   ("mm6",$Gsse2);                 # load g
 416
 417         &paddq  ("mm7","mm1");                  # X[i]
 418         #&movq  (&QWP(8*9,"esp"),"mm7");        # moved to BODY_00_15
 419
 420         &BODY_00_15_sse2(2);
 421     }
 422         &dec    ("edx");
 423         &jnz    (&label("16_79_sse2"));
 424
 425         #&movq  ($A,$Asse2);
 426         &paddq  ($A,"mm3");                     # from BODY_00_15
 427         &movq   ("mm1",$Bsse2);
 428         #&movq  ($BxC,$Csse2);
 429         &movq   ("mm3",$Dsse2);
 430         #&movq  ($E,$Esse2);
 431         &movq   ("mm5",$Fsse2);
 432         &movq   ("mm6",$Gsse2);
 433         &movq   ("mm7",$Hsse2);
 434
 435         &pxor   ($BxC,"mm1");                   # de-magic
 436         &paddq  ($A,&QWP(0,"esi"));
 437         &paddq  ("mm1",&QWP(8,"esi"));
 438         &paddq  ($BxC,&QWP(16,"esi"));
 439         &paddq  ("mm3",&QWP(24,"esi"));
 440         &paddq  ($E,&QWP(32,"esi"));
 441         &paddq  ("mm5",&QWP(40,"esi"));
 442         &paddq  ("mm6",&QWP(48,"esi"));
 443         &paddq  ("mm7",&QWP(56,"esi"));
 444
 445         &mov    ("eax",8*80);
 446         &movq   (&QWP(0,"esi"),$A);
 447         &movq   (&QWP(8,"esi"),"mm1");
 448         &movq   (&QWP(16,"esi"),$BxC);
 449         &movq   (&QWP(24,"esi"),"mm3");
 450         &movq   (&QWP(32,"esi"),$E);
 451         &movq   (&QWP(40,"esi"),"mm5");
 452         &movq   (&QWP(48,"esi"),"mm6");
 453         &movq   (&QWP(56,"esi"),"mm7");
 454
 455         &lea    ("esp",&DWP(0,"esp","eax"));    # destroy frame
 456         &sub    ($K512,"eax");                  # rewind K
 457
 458         &cmp    ("edi",&DWP(8*10+8,"esp"));     # are we done yet?
 459         &jb     (&label("loop_sse2"));
 460
 461         &mov    ("esp",&DWP(8*10+12,"esp"));    # restore sp
 462         &emms   ();
 463 &function_end_A();
 464
 465 &set_label("SSSE3",32);
 466 { my ($cnt,$frame)=("ecx","edx");
 467   my @X=map("xmm$_",(0..7));
 468   my $j;
 469   my $i=0;
 470
 471         &lea    ($frame,&DWP(-64,"esp"));
 472         &sub    ("esp",256);
 473
 474         # fixed stack frame layout
 475         #
 476         # +0    A B C D E F G H         # backing store
 477         # +64   X[0]+K[i] .. X[15]+K[i] # XMM->MM xfer area
 478         # +192                          # XMM off-load ring buffer
 479         # +256                          # saved parameters
 480
 481         &movdqa         (@X[1],&QWP(80*8,$K512));               # byte swap mask
 482         &movdqu         (@X[0],&QWP(0,"edi"));
 483         &pshufb         (@X[0],@X[1]);
 484     for ($j=0;$j<8;$j++) {
 485         &movdqa         (&QWP(16*(($j-1)%4),$frame),@X[3])      if ($j>4); # off-load
 486         &movdqa         (@X[3],&QWP(16*($j%8),$K512));
 487         &movdqa         (@X[2],@X[1])                           if ($j<7); # perpetuate byte swap mask
 488         &movdqu         (@X[1],&QWP(16*($j+1),"edi"))           if ($j<7); # next input
 489         &movdqa         (@X[1],&QWP(16*(($j+1)%4),$frame))      if ($j==7);# restore @X[0]
 490         &paddq          (@X[3],@X[0]);
 491         &pshufb         (@X[1],@X[2])                           if ($j<7);
 492         &movdqa         (&QWP(16*($j%8)-128,$frame),@X[3]);     # xfer X[i]+K[i]
 493
 494         push(@X,shift(@X));                                     # rotate(@X)
 495     }
 496         #&jmp           (&label("loop_ssse3"));
 497         &nop            ();
 498
 499 &set_label("loop_ssse3",32);
 500         &movdqa         (@X[2],&QWP(16*(($j+1)%4),$frame));     # pre-restore @X[1]
 501         &movdqa         (&QWP(16*(($j-1)%4),$frame),@X[3]);     # off-load @X[3]
 502         &lea            ($K512,&DWP(16*8,$K512));
 503
 504         #&movq  ($Asse2,$A);                    # off-load A-H
 505         &movq   ($Bsse2,"mm1");
 506          &mov   ("ebx","edi");
 507         &movq   ($Csse2,$BxC);
 508          &lea   ("edi",&DWP(128,"edi"));        # advance input
 509         &movq   ($Dsse2,"mm3");
 510          &cmp   ("edi","eax");
 511         #&movq  ($Esse2,$E);
 512         &movq   ($Fsse2,"mm5");
 513          &cmovb ("ebx","edi");
 514         &movq   ($Gsse2,"mm6");
 515          &mov   ("ecx",4);                      # loop counter
 516         &pxor   ($BxC,"mm1");                   # magic
 517         &movq   ($Hsse2,"mm7");
 518         &pxor   ("mm3","mm3");                  # magic
 519
 520         &jmp            (&label("00_47_ssse3"));
 521
 522 sub BODY_00_15_ssse3 {          # "phase-less" copy of BODY_00_15_sse2
 523         (
 524         '&movq  ("mm1",$E)',                            # %mm1 is sliding right
 525         '&movq  ("mm7",&QWP(((-8*$i)%128)-128,$frame))',# X[i]+K[i]
 526          '&pxor ("mm5","mm6")',                         # f^=g
 527         '&psrlq ("mm1",14)',
 528          '&movq (&QWP(8*($i+4)%64,"esp"),$E)',          # modulo-scheduled save e
 529          '&pand ("mm5",$E)',                            # f&=e
 530         '&psllq ($E,23)',                               # $E is sliding left
 531         '&paddq ($A,"mm3")',                            # [h+=Maj(a,b,c)]
 532         '&movq  ("mm3","mm1")',                         # %mm3 is T1
 533          '&psrlq("mm1",4)',
 534          '&pxor ("mm5","mm6")',                         # Ch(e,f,g)
 535         '&pxor  ("mm3",$E)',
 536          '&psllq($E,23)',
 537         '&pxor  ("mm3","mm1")',
 538          '&movq (&QWP(8*$i%64,"esp"),$A)',              # modulo-scheduled save a
 539          '&paddq("mm7","mm5")',                         # X[i]+=Ch(e,f,g)
 540         '&pxor  ("mm3",$E)',
 541          '&psrlq("mm1",23)',
 542          '&paddq("mm7",&QWP(8*($i+7)%64,"esp"))',       # X[i]+=h
 543         '&pxor  ("mm3","mm1")',
 544          '&psllq($E,4)',
 545         '&pxor  ("mm3",$E)',                            # T1=Sigma1_512(e)
 546
 547          '&movq ($E,&QWP(8*($i+3)%64,"esp"))',          # e = load d, e in next round
 548         '&paddq ("mm3","mm7")',                         # T1+=X[i]
 549          '&movq ("mm5",$A)',                            # %mm5 is sliding right
 550          '&psrlq("mm5",28)',
 551         '&paddq ($E,"mm3")',                            # d += T1
 552          '&movq ("mm6",$A)',                            # %mm6 is sliding left
 553          '&movq ("mm7","mm5")',
 554          '&psllq("mm6",25)',
 555         '&movq  ("mm1",&QWP(8*($i+1)%64,"esp"))',       # load b
 556          '&psrlq("mm5",6)',
 557          '&pxor ("mm7","mm6")',
 558          '&psllq("mm6",5)',
 559          '&pxor ("mm7","mm5")',
 560         '&pxor  ($A,"mm1")',                            # a^b, b^c in next round
 561          '&psrlq("mm5",5)',
 562          '&pxor ("mm7","mm6")',
 563         '&pand  ($BxC,$A)',                             # (b^c)&(a^b)
 564          '&psllq("mm6",6)',
 565          '&pxor ("mm7","mm5")',
 566         '&pxor  ($BxC,"mm1")',                          # [h=]Maj(a,b,c)
 567          '&pxor ("mm6","mm7")',                         # Sigma0_512(a)
 568          '&movq ("mm5",&QWP(8*($i+5-1)%64,"esp"))',     # pre-load f
 569         '&paddq ($BxC,"mm6")',                          # h+=Sigma0(a)
 570          '&movq ("mm6",&QWP(8*($i+6-1)%64,"esp"))',     # pre-load g
 571
 572         '($A,$BxC) = ($BxC,$A); $i--;'
 573         );
 574 }
 575
 576 &set_label("00_47_ssse3",32);
 577
 578     for(;$j<16;$j++) {
 579         my ($t0,$t2,$t1)=@X[2..4];
 580         my @insns = (&BODY_00_15_ssse3(),&BODY_00_15_ssse3());
 581
 582         &movdqa         ($t2,@X[5]);
 583         &movdqa         (@X[1],$t0);                    # restore @X[1]
 584         &palignr        ($t0,@X[0],8);                  # X[1..2]
 585         &movdqa         (&QWP(16*($j%4),$frame),@X[4]); # off-load @X[4]
 586          &palignr       ($t2,@X[4],8);                  # X[9..10]
 587
 588         &movdqa         ($t1,$t0);
 589         &psrlq          ($t0,7);
 590          &paddq         (@X[0],$t2);                    # X[0..1] += X[9..10]
 591         &movdqa         ($t2,$t1);
 592         &psrlq          ($t1,1);
 593         &psllq          ($t2,64-8);
 594         &pxor           ($t0,$t1);
 595         &psrlq          ($t1,8-1);
 596         &pxor           ($t0,$t2);
 597         &psllq          ($t2,8-1);
 598         &pxor           ($t0,$t1);
 599          &movdqa        ($t1,@X[7]);
 600         &pxor           ($t0,$t2);                      # sigma0(X[1..2])
 601          &movdqa        ($t2,@X[7]);
 602          &psrlq         ($t1,6);
 603         &paddq          (@X[0],$t0);                    # X[0..1] += sigma0(X[1..2])
 604
 605         &movdqa         ($t0,@X[7]);
 606         &psrlq          ($t2,19);
 607         &psllq          ($t0,64-61);
 608         &pxor           ($t1,$t2);
 609         &psrlq          ($t2,61-19);
 610         &pxor           ($t1,$t0);
 611         &psllq          ($t0,61-19);
 612         &pxor           ($t1,$t2);
 613         &movdqa         ($t2,&QWP(16*(($j+2)%4),$frame));# pre-restore @X[1]
 614         &pxor           ($t1,$t0);                      # sigma0(X[1..2])
 615         &movdqa         ($t0,&QWP(16*($j%8),$K512));
 616          eval(shift(@insns));
 617         &paddq          (@X[0],$t1);                    # X[0..1] += sigma0(X[14..15])
 618          eval(shift(@insns));
 619          eval(shift(@insns));
 620          eval(shift(@insns));
 621          eval(shift(@insns));
 622         &paddq          ($t0,@X[0]);
 623          foreach(@insns) { eval; }
 624         &movdqa         (&QWP(16*($j%8)-128,$frame),$t0);# xfer X[i]+K[i]
 625
 626         push(@X,shift(@X));                             # rotate(@X)
 627     }
 628         &lea            ($K512,&DWP(16*8,$K512));
 629         &dec            ("ecx");
 630         &jnz            (&label("00_47_ssse3"));
 631
 632         &movdqa         (@X[1],&QWP(0,$K512));          # byte swap mask
 633         &lea            ($K512,&DWP(-80*8,$K512));      # rewind
 634         &movdqu         (@X[0],&QWP(0,"ebx"));
 635         &pshufb         (@X[0],@X[1]);
 636
 637     for ($j=0;$j<8;$j++) {      # load next or same block
 638         my @insns = (&BODY_00_15_ssse3(),&BODY_00_15_ssse3());
 639
 640         &movdqa         (&QWP(16*(($j-1)%4),$frame),@X[3])      if ($j>4); # off-load
 641         &movdqa         (@X[3],&QWP(16*($j%8),$K512));
 642         &movdqa         (@X[2],@X[1])                           if ($j<7); # perpetuate byte swap mask
 643         &movdqu         (@X[1],&QWP(16*($j+1),"ebx"))           if ($j<7); # next input
 644         &movdqa         (@X[1],&QWP(16*(($j+1)%4),$frame))      if ($j==7);# restore @X[0]
 645         &paddq          (@X[3],@X[0]);
 646         &pshufb         (@X[1],@X[2])                           if ($j<7);
 647          foreach(@insns) { eval; }
 648         &movdqa         (&QWP(16*($j%8)-128,$frame),@X[3]);# xfer X[i]+K[i]
 649
 650         push(@X,shift(@X));                             # rotate(@X)
 651     }
 652
 653         #&movq  ($A,$Asse2);                    # load A-H
 654         &movq   ("mm1",$Bsse2);
 655         &paddq  ($A,"mm3");                     # from BODY_00_15
 656         #&movq  ($BxC,$Csse2);
 657         &movq   ("mm3",$Dsse2);
 658         #&movq  ($E,$Esse2);
 659         #&movq  ("mm5",$Fsse2);
 660         #&movq  ("mm6",$Gsse2);
 661         &movq   ("mm7",$Hsse2);
 662
 663         &pxor   ($BxC,"mm1");                   # de-magic
 664         &paddq  ($A,&QWP(0,"esi"));
 665         &paddq  ("mm1",&QWP(8,"esi"));
 666         &paddq  ($BxC,&QWP(16,"esi"));
 667         &paddq  ("mm3",&QWP(24,"esi"));
 668         &paddq  ($E,&QWP(32,"esi"));
 669         &paddq  ("mm5",&QWP(40,"esi"));
 670         &paddq  ("mm6",&QWP(48,"esi"));
 671         &paddq  ("mm7",&QWP(56,"esi"));
 672
 673         &movq   (&QWP(0,"esi"),$A);
 674         &movq   (&QWP(8,"esi"),"mm1");
 675         &movq   (&QWP(16,"esi"),$BxC);
 676         &movq   (&QWP(24,"esi"),"mm3");
 677         &movq   (&QWP(32,"esi"),$E);
 678         &movq   (&QWP(40,"esi"),"mm5");
 679         &movq   (&QWP(48,"esi"),"mm6");
 680         &movq   (&QWP(56,"esi"),"mm7");
 681
 682         &cmp    ("edi","eax")                   # are we done yet?
 683         &jb     (&label("loop_ssse3"));
 684
 685         &mov    ("esp",&DWP(64+12,$frame));     # restore sp
 686         &emms   ();
 687 }
 688 &function_end_A();
 689 }
 690 &set_label("loop_x86",16);
 691     # copy input block to stack reversing byte and qword order
 692     for ($i=0;$i<8;$i++) {
 693         &mov    ("eax",&DWP($i*16+0,"edi"));
 694         &mov    ("ebx",&DWP($i*16+4,"edi"));
 695         &mov    ("ecx",&DWP($i*16+8,"edi"));
 696         &mov    ("edx",&DWP($i*16+12,"edi"));
 697         &bswap  ("eax");
 698         &bswap  ("ebx");
 699         &bswap  ("ecx");
 700         &bswap  ("edx");
 701         &push   ("eax");
 702         &push   ("ebx");
 703         &push   ("ecx");
 704         &push   ("edx");
 705     }
 706         &add    ("edi",128);
 707         &sub    ("esp",9*8);            # place for T,A,B,C,D,E,F,G,H
 708         &mov    (&DWP(8*(9+16)+4,"esp"),"edi");
 709
 710         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 711         &lea    ("edi",&DWP(8,"esp"));
 712         &mov    ("ecx",16);
 713         &data_word(0xA5F3F689);         # rep movsd
 714
 715 &set_label("00_15_x86",16);
 716         &BODY_00_15_x86();
 717
 718         &cmp    (&LB("edx"),0x94);
 719         &jne    (&label("00_15_x86"));
 720
 721 &set_label("16_79_x86",16);
 722         #define sigma0(x)       (ROTR((x),1)  ^ ROTR((x),8)  ^ ((x)>>7))
 723         #       LO              lo>>1^hi<<31  ^ lo>>8^hi<<24 ^ lo>>7^hi<<25
 724         #       HI              hi>>1^lo<<31  ^ hi>>8^lo<<24 ^ hi>>7
 725         &mov    ("ecx",&DWP(8*(9+15+16-1)+0,"esp"));
 726         &mov    ("edx",&DWP(8*(9+15+16-1)+4,"esp"));
 727         &mov    ("esi","ecx");
 728
 729         &shr    ("ecx",1);      # lo>>1
 730         &mov    ("edi","edx");
 731         &shr    ("edx",1);      # hi>>1
 732         &mov    ("eax","ecx");
 733         &shl    ("esi",24);     # lo<<24
 734         &mov    ("ebx","edx");
 735         &shl    ("edi",24);     # hi<<24
 736         &xor    ("ebx","esi");
 737
 738         &shr    ("ecx",7-1);    # lo>>7
 739         &xor    ("eax","edi");
 740         &shr    ("edx",7-1);    # hi>>7
 741         &xor    ("eax","ecx");
 742         &shl    ("esi",31-24);  # lo<<31
 743         &xor    ("ebx","edx");
 744         &shl    ("edi",25-24);  # hi<<25
 745         &xor    ("ebx","esi");
 746
 747         &shr    ("ecx",8-7);    # lo>>8
 748         &xor    ("eax","edi");
 749         &shr    ("edx",8-7);    # hi>>8
 750         &xor    ("eax","ecx");
 751         &shl    ("edi",31-25);  # hi<<31
 752         &xor    ("ebx","edx");
 753         &xor    ("eax","edi");                  # T1 = sigma0(X[-15])
 754
 755         &mov    (&DWP(0,"esp"),"eax");
 756         &mov    (&DWP(4,"esp"),"ebx");          # put T1 away
 757
 758         #define sigma1(x)       (ROTR((x),19) ^ ROTR((x),61) ^ ((x)>>6))
 759         #       LO              lo>>19^hi<<13 ^ hi>>29^lo<<3 ^ lo>>6^hi<<26
 760         #       HI              hi>>19^lo<<13 ^ lo>>29^hi<<3 ^ hi>>6
 761         &mov    ("ecx",&DWP(8*(9+15+16-14)+0,"esp"));
 762         &mov    ("edx",&DWP(8*(9+15+16-14)+4,"esp"));
 763         &mov    ("esi","ecx");
 764
 765         &shr    ("ecx",6);      # lo>>6
 766         &mov    ("edi","edx");
 767         &shr    ("edx",6);      # hi>>6
 768         &mov    ("eax","ecx");
 769         &shl    ("esi",3);      # lo<<3
 770         &mov    ("ebx","edx");
 771         &shl    ("edi",3);      # hi<<3
 772         &xor    ("eax","esi");
 773
 774         &shr    ("ecx",19-6);   # lo>>19
 775         &xor    ("ebx","edi");
 776         &shr    ("edx",19-6);   # hi>>19
 777         &xor    ("eax","ecx");
 778         &shl    ("esi",13-3);   # lo<<13
 779         &xor    ("ebx","edx");
 780         &shl    ("edi",13-3);   # hi<<13
 781         &xor    ("ebx","esi");
 782
 783         &shr    ("ecx",29-19);  # lo>>29
 784         &xor    ("eax","edi");
 785         &shr    ("edx",29-19);  # hi>>29
 786         &xor    ("ebx","ecx");
 787         &shl    ("edi",26-13);  # hi<<26
 788         &xor    ("eax","edx");
 789         &xor    ("eax","edi");                  # sigma1(X[-2])
 790
 791         &mov    ("ecx",&DWP(8*(9+15+16)+0,"esp"));
 792         &mov    ("edx",&DWP(8*(9+15+16)+4,"esp"));
 793         &add    ("eax",&DWP(0,"esp"));
 794         &adc    ("ebx",&DWP(4,"esp"));          # T1 = sigma1(X[-2])+T1
 795         &mov    ("esi",&DWP(8*(9+15+16-9)+0,"esp"));
 796         &mov    ("edi",&DWP(8*(9+15+16-9)+4,"esp"));
 797         &add    ("eax","ecx");
 798         &adc    ("ebx","edx");                  # T1 += X[-16]
 799         &add    ("eax","esi");
 800         &adc    ("ebx","edi");                  # T1 += X[-7]
 801         &mov    (&DWP(8*(9+15)+0,"esp"),"eax");
 802         &mov    (&DWP(8*(9+15)+4,"esp"),"ebx"); # save X[0]
 803
 804         &BODY_00_15_x86();
 805
 806         &cmp    (&LB("edx"),0x17);
 807         &jne    (&label("16_79_x86"));
 808
 809         &mov    ("esi",&DWP(8*(9+16+80)+0,"esp"));# ctx
 810         &mov    ("edi",&DWP(8*(9+16+80)+4,"esp"));# inp
 811     for($i=0;$i<4;$i++) {
 812         &mov    ("eax",&DWP($i*16+0,"esi"));
 813         &mov    ("ebx",&DWP($i*16+4,"esi"));
 814         &mov    ("ecx",&DWP($i*16+8,"esi"));
 815         &mov    ("edx",&DWP($i*16+12,"esi"));
 816         &add    ("eax",&DWP(8+($i*16)+0,"esp"));
 817         &adc    ("ebx",&DWP(8+($i*16)+4,"esp"));
 818         &mov    (&DWP($i*16+0,"esi"),"eax");
 819         &mov    (&DWP($i*16+4,"esi"),"ebx");
 820         &add    ("ecx",&DWP(8+($i*16)+8,"esp"));
 821         &adc    ("edx",&DWP(8+($i*16)+12,"esp"));
 822         &mov    (&DWP($i*16+8,"esi"),"ecx");
 823         &mov    (&DWP($i*16+12,"esi"),"edx");
 824     }
 825         &add    ("esp",8*(9+16+80));            # destroy frame
 826         &sub    ($K512,8*80);                   # rewind K
 827
 828         &cmp    ("edi",&DWP(8,"esp"));          # are we done yet?
 829         &jb     (&label("loop_x86"));
 830
 831         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 832 &function_end_A();
 833
 834 &set_label("K512",64);  # Yes! I keep it in the code segment!
 835         &data_word(0xd728ae22,0x428a2f98);      # u64
 836         &data_word(0x23ef65cd,0x71374491);      # u64
 837         &data_word(0xec4d3b2f,0xb5c0fbcf);      # u64
 838         &data_word(0x8189dbbc,0xe9b5dba5);      # u64
 839         &data_word(0xf348b538,0x3956c25b);      # u64
 840         &data_word(0xb605d019,0x59f111f1);      # u64
 841         &data_word(0xaf194f9b,0x923f82a4);      # u64
 842         &data_word(0xda6d8118,0xab1c5ed5);      # u64
 843         &data_word(0xa3030242,0xd807aa98);      # u64
 844         &data_word(0x45706fbe,0x12835b01);      # u64
 845         &data_word(0x4ee4b28c,0x243185be);      # u64
 846         &data_word(0xd5ffb4e2,0x550c7dc3);      # u64
 847         &data_word(0xf27b896f,0x72be5d74);      # u64
 848         &data_word(0x3b1696b1,0x80deb1fe);      # u64
 849         &data_word(0x25c71235,0x9bdc06a7);      # u64
 850         &data_word(0xcf692694,0xc19bf174);      # u64
 851         &data_word(0x9ef14ad2,0xe49b69c1);      # u64
 852         &data_word(0x384f25e3,0xefbe4786);      # u64
 853         &data_word(0x8b8cd5b5,0x0fc19dc6);      # u64
 854         &data_word(0x77ac9c65,0x240ca1cc);      # u64
 855         &data_word(0x592b0275,0x2de92c6f);      # u64
 856         &data_word(0x6ea6e483,0x4a7484aa);      # u64
 857         &data_word(0xbd41fbd4,0x5cb0a9dc);      # u64
 858         &data_word(0x831153b5,0x76f988da);      # u64
 859         &data_word(0xee66dfab,0x983e5152);      # u64
 860         &data_word(0x2db43210,0xa831c66d);      # u64
 861         &data_word(0x98fb213f,0xb00327c8);      # u64
 862         &data_word(0xbeef0ee4,0xbf597fc7);      # u64
 863         &data_word(0x3da88fc2,0xc6e00bf3);      # u64
 864         &data_word(0x930aa725,0xd5a79147);      # u64
 865         &data_word(0xe003826f,0x06ca6351);      # u64
 866         &data_word(0x0a0e6e70,0x14292967);      # u64
 867         &data_word(0x46d22ffc,0x27b70a85);      # u64
 868         &data_word(0x5c26c926,0x2e1b2138);      # u64
 869         &data_word(0x5ac42aed,0x4d2c6dfc);      # u64
 870         &data_word(0x9d95b3df,0x53380d13);      # u64
 871         &data_word(0x8baf63de,0x650a7354);      # u64
 872         &data_word(0x3c77b2a8,0x766a0abb);      # u64
 873         &data_word(0x47edaee6,0x81c2c92e);      # u64
 874         &data_word(0x1482353b,0x92722c85);      # u64
 875         &data_word(0x4cf10364,0xa2bfe8a1);      # u64
 876         &data_word(0xbc423001,0xa81a664b);      # u64
 877         &data_word(0xd0f89791,0xc24b8b70);      # u64
 878         &data_word(0x0654be30,0xc76c51a3);      # u64
 879         &data_word(0xd6ef5218,0xd192e819);      # u64
 880         &data_word(0x5565a910,0xd6990624);      # u64
 881         &data_word(0x5771202a,0xf40e3585);      # u64
 882         &data_word(0x32bbd1b8,0x106aa070);      # u64
 883         &data_word(0xb8d2d0c8,0x19a4c116);      # u64
 884         &data_word(0x5141ab53,0x1e376c08);      # u64
 885         &data_word(0xdf8eeb99,0x2748774c);      # u64
 886         &data_word(0xe19b48a8,0x34b0bcb5);      # u64
 887         &data_word(0xc5c95a63,0x391c0cb3);      # u64
 888         &data_word(0xe3418acb,0x4ed8aa4a);      # u64
 889         &data_word(0x7763e373,0x5b9cca4f);      # u64
 890         &data_word(0xd6b2b8a3,0x682e6ff3);      # u64
 891         &data_word(0x5defb2fc,0x748f82ee);      # u64
 892         &data_word(0x43172f60,0x78a5636f);      # u64
 893         &data_word(0xa1f0ab72,0x84c87814);      # u64
 894         &data_word(0x1a6439ec,0x8cc70208);      # u64
 895         &data_word(0x23631e28,0x90befffa);      # u64
 896         &data_word(0xde82bde9,0xa4506ceb);      # u64
 897         &data_word(0xb2c67915,0xbef9a3f7);      # u64
 898         &data_word(0xe372532b,0xc67178f2);      # u64
 899         &data_word(0xea26619c,0xca273ece);      # u64
 900         &data_word(0x21c0c207,0xd186b8c7);      # u64
 901         &data_word(0xcde0eb1e,0xeada7dd6);      # u64
 902         &data_word(0xee6ed178,0xf57d4f7f);      # u64
 903         &data_word(0x72176fba,0x06f067aa);      # u64
 904         &data_word(0xa2c898a6,0x0a637dc5);      # u64
 905         &data_word(0xbef90dae,0x113f9804);      # u64
 906         &data_word(0x131c471b,0x1b710b35);      # u64
 907         &data_word(0x23047d84,0x28db77f5);      # u64
 908         &data_word(0x40c72493,0x32caab7b);      # u64
 909         &data_word(0x15c9bebc,0x3c9ebe0a);      # u64
 910         &data_word(0x9c100d4c,0x431d67c4);      # u64
 911         &data_word(0xcb3e42b6,0x4cc5d4be);      # u64
 912         &data_word(0xfc657e2a,0x597f299c);      # u64
 913         &data_word(0x3ad6faec,0x5fcb6fab);      # u64
 914         &data_word(0x4a475817,0x6c44198c);      # u64
 915
 916         &data_word(0x04050607,0x00010203);      # byte swap
 917         &data_word(0x0c0d0e0f,0x08090a0b);      # mask
 918 &function_end_B("sha512_block_data_order");
 919 &asciz("SHA512 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
 920
 921 &asm_finish();
 922
 923 close STDOUT;