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 # Skylake       40      -       26      13.3    7.25
  36 # Bulldozer     121     -       50      14.0    13.5
  37 # VIA Nano      91      -       52      33      14.7
  38 # Atom          126     -       68      48(***) 14.7
  39 # Silvermont    97      -       58      42(***) 17.5
  40 # Goldmont      80      -       48      19.5    12.0
  41 #
  42 # (*)   whichever best applicable.
  43 # (**)  x86_64 assembler performance is presented for reference
  44 #       purposes, the results are for integer-only code.
  45 # (***) paddq is increadibly slow on Atom.
  46 #
  47 # IALU code-path is optimized for elder Pentiums. On vanilla Pentium
  48 # performance improvement over compiler generated code reaches ~60%,
  49 # while on PIII - ~35%. On newer µ-archs improvement varies from 15%
  50 # to 50%, but it's less important as they are expected to execute SSE2
  51 # code-path, which is commonly ~2-3x faster [than compiler generated
  52 # code]. SSE2 code-path is as fast as original sha512-sse2.pl, even
  53 # though it does not use 128-bit operations. The latter means that
  54 # SSE2-aware kernel is no longer required to execute the code. Another
  55 # difference is that new code optimizes amount of writes, but at the
  56 # cost of increased data cache "footprint" by 1/2KB.
  57
  58 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  59 push(@INC,"${dir}","${dir}../../perlasm");
  60 require "x86asm.pl";
  61
  62 $output=pop;
  63 open STDOUT,">$output";
  64
  65 &asm_init($ARGV[0],$ARGV[$#ARGV] eq "386");
  66
  67 $sse2=0;
  68 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
  69
  70 &external_label("OPENSSL_ia32cap_P") if ($sse2);
  71
  72 $Tlo=&DWP(0,"esp");     $Thi=&DWP(4,"esp");
  73 $Alo=&DWP(8,"esp");     $Ahi=&DWP(8+4,"esp");
  74 $Blo=&DWP(16,"esp");    $Bhi=&DWP(16+4,"esp");
  75 $Clo=&DWP(24,"esp");    $Chi=&DWP(24+4,"esp");
  76 $Dlo=&DWP(32,"esp");    $Dhi=&DWP(32+4,"esp");
  77 $Elo=&DWP(40,"esp");    $Ehi=&DWP(40+4,"esp");
  78 $Flo=&DWP(48,"esp");    $Fhi=&DWP(48+4,"esp");
  79 $Glo=&DWP(56,"esp");    $Ghi=&DWP(56+4,"esp");
  80 $Hlo=&DWP(64,"esp");    $Hhi=&DWP(64+4,"esp");
  81 $K512="ebp";
  82
  83 $Asse2=&QWP(0,"esp");
  84 $Bsse2=&QWP(8,"esp");
  85 $Csse2=&QWP(16,"esp");
  86 $Dsse2=&QWP(24,"esp");
  87 $Esse2=&QWP(32,"esp");
  88 $Fsse2=&QWP(40,"esp");
  89 $Gsse2=&QWP(48,"esp");
  90 $Hsse2=&QWP(56,"esp");
  91
  92 $A="mm0";       # B-D and
  93 $E="mm4";       # F-H are commonly loaded to respectively mm1-mm3 and
  94                 # mm5-mm7, but it's done on on-demand basis...
  95 $BxC="mm2";     # ... except for B^C
  96
  97 sub BODY_00_15_sse2 {
  98     my $phase=shift;
  99
 100         #&movq  ("mm5",$Fsse2);                 # load f
 101         #&movq  ("mm6",$Gsse2);                 # load g
 102
 103         &movq   ("mm1",$E);                     # %mm1 is sliding right
 104          &pxor  ("mm5","mm6");                  # f^=g
 105         &psrlq  ("mm1",14);
 106          &movq  ($Esse2,$E);                    # modulo-scheduled save e
 107          &pand  ("mm5",$E);                     # f&=e
 108         &psllq  ($E,23);                        # $E is sliding left
 109          &movq  ($A,"mm3")                      if ($phase<2);
 110          &movq  (&QWP(8*9,"esp"),"mm7")         # save X[i]
 111         &movq   ("mm3","mm1");                  # %mm3 is T1
 112          &psrlq ("mm1",4);
 113          &pxor  ("mm5","mm6");                  # Ch(e,f,g)
 114         &pxor   ("mm3",$E);
 115          &psllq ($E,23);
 116         &pxor   ("mm3","mm1");
 117          &movq  ($Asse2,$A);                    # modulo-scheduled save a
 118          &paddq ("mm7","mm5");                  # X[i]+=Ch(e,f,g)
 119         &pxor   ("mm3",$E);
 120          &psrlq ("mm1",23);
 121          &paddq ("mm7",$Hsse2);                 # X[i]+=h
 122         &pxor   ("mm3","mm1");
 123          &psllq ($E,4);
 124          &paddq ("mm7",QWP(0,$K512));           # X[i]+=K512[i]
 125         &pxor   ("mm3",$E);                     # T1=Sigma1_512(e)
 126
 127          &movq  ($E,$Dsse2);                    # e = load d, e in next round
 128         &paddq  ("mm3","mm7");                  # T1+=X[i]
 129          &movq  ("mm5",$A);                     # %mm5 is sliding right
 130          &psrlq ("mm5",28);
 131         &paddq  ($E,"mm3");                     # d += T1
 132          &movq  ("mm6",$A);                     # %mm6 is sliding left
 133          &movq  ("mm7","mm5");
 134          &psllq ("mm6",25);
 135         &movq   ("mm1",$Bsse2);                 # load b
 136          &psrlq ("mm5",6);
 137          &pxor  ("mm7","mm6");
 138         &sub    ("esp",8);
 139          &psllq ("mm6",5);
 140          &pxor  ("mm7","mm5");
 141         &pxor   ($A,"mm1");                     # a^b, b^c in next round
 142          &psrlq ("mm5",5);
 143          &pxor  ("mm7","mm6");
 144         &pand   ($BxC,$A);                      # (b^c)&(a^b)
 145          &psllq ("mm6",6);
 146          &pxor  ("mm7","mm5");
 147         &pxor   ($BxC,"mm1");                   # [h=]Maj(a,b,c)
 148          &pxor  ("mm6","mm7");                  # Sigma0_512(a)
 149          &movq  ("mm7",&QWP(8*(9+16-1),"esp"))  if ($phase!=0); # pre-fetch
 150          &movq  ("mm5",$Fsse2)                  if ($phase==0); # load f
 151
 152     if ($phase>1) {
 153         &paddq  ($BxC,"mm6");                   # h+=Sigma0(a)
 154          &add   ($K512,8);
 155         #&paddq ($BxC,"mm3");                   # h+=T1
 156
 157         ($A,$BxC) = ($BxC,$A);                  # rotate registers
 158     } else {
 159         &paddq  ("mm3",$BxC);                   # T1+=Maj(a,b,c)
 160          &movq  ($BxC,$A);
 161          &add   ($K512,8);
 162         &paddq  ("mm3","mm6");                  # T1+=Sigma0(a)
 163          &movq  ("mm6",$Gsse2)                  if ($phase==0); # load g
 164         #&movq  ($A,"mm3");                     # h=T1
 165     }
 166 }
 167
 168 sub BODY_00_15_x86 {
 169         #define Sigma1(x)       (ROTR((x),14) ^ ROTR((x),18)  ^ ROTR((x),41))
 170         #       LO              lo>>14^hi<<18 ^ lo>>18^hi<<14 ^ hi>>9^lo<<23
 171         #       HI              hi>>14^lo<<18 ^ hi>>18^lo<<14 ^ lo>>9^hi<<23
 172         &mov    ("ecx",$Elo);
 173         &mov    ("edx",$Ehi);
 174         &mov    ("esi","ecx");
 175
 176         &shr    ("ecx",9);      # lo>>9
 177         &mov    ("edi","edx");
 178         &shr    ("edx",9);      # hi>>9
 179         &mov    ("ebx","ecx");
 180         &shl    ("esi",14);     # lo<<14
 181         &mov    ("eax","edx");
 182         &shl    ("edi",14);     # hi<<14
 183         &xor    ("ebx","esi");
 184
 185         &shr    ("ecx",14-9);   # lo>>14
 186         &xor    ("eax","edi");
 187         &shr    ("edx",14-9);   # hi>>14
 188         &xor    ("eax","ecx");
 189         &shl    ("esi",18-14);  # lo<<18
 190         &xor    ("ebx","edx");
 191         &shl    ("edi",18-14);  # hi<<18
 192         &xor    ("ebx","esi");
 193
 194         &shr    ("ecx",18-14);  # lo>>18
 195         &xor    ("eax","edi");
 196         &shr    ("edx",18-14);  # hi>>18
 197         &xor    ("eax","ecx");
 198         &shl    ("esi",23-18);  # lo<<23
 199         &xor    ("ebx","edx");
 200         &shl    ("edi",23-18);  # hi<<23
 201         &xor    ("eax","esi");
 202         &xor    ("ebx","edi");                  # T1 = Sigma1(e)
 203
 204         &mov    ("ecx",$Flo);
 205         &mov    ("edx",$Fhi);
 206         &mov    ("esi",$Glo);
 207         &mov    ("edi",$Ghi);
 208          &add   ("eax",$Hlo);
 209          &adc   ("ebx",$Hhi);                   # T1 += h
 210         &xor    ("ecx","esi");
 211         &xor    ("edx","edi");
 212         &and    ("ecx",$Elo);
 213         &and    ("edx",$Ehi);
 214          &add   ("eax",&DWP(8*(9+15)+0,"esp"));
 215          &adc   ("ebx",&DWP(8*(9+15)+4,"esp")); # T1 += X[0]
 216         &xor    ("ecx","esi");
 217         &xor    ("edx","edi");                  # Ch(e,f,g) = (f^g)&e)^g
 218
 219         &mov    ("esi",&DWP(0,$K512));
 220         &mov    ("edi",&DWP(4,$K512));          # K[i]
 221         &add    ("eax","ecx");
 222         &adc    ("ebx","edx");                  # T1 += Ch(e,f,g)
 223         &mov    ("ecx",$Dlo);
 224         &mov    ("edx",$Dhi);
 225         &add    ("eax","esi");
 226         &adc    ("ebx","edi");                  # T1 += K[i]
 227         &mov    ($Tlo,"eax");
 228         &mov    ($Thi,"ebx");                   # put T1 away
 229         &add    ("eax","ecx");
 230         &adc    ("ebx","edx");                  # d += T1
 231
 232         #define Sigma0(x)       (ROTR((x),28) ^ ROTR((x),34) ^ ROTR((x),39))
 233         #       LO              lo>>28^hi<<4  ^ hi>>2^lo<<30 ^ hi>>7^lo<<25
 234         #       HI              hi>>28^lo<<4  ^ lo>>2^hi<<30 ^ lo>>7^hi<<25
 235         &mov    ("ecx",$Alo);
 236         &mov    ("edx",$Ahi);
 237         &mov    ($Dlo,"eax");
 238         &mov    ($Dhi,"ebx");
 239         &mov    ("esi","ecx");
 240
 241         &shr    ("ecx",2);      # lo>>2
 242         &mov    ("edi","edx");
 243         &shr    ("edx",2);      # hi>>2
 244         &mov    ("ebx","ecx");
 245         &shl    ("esi",4);      # lo<<4
 246         &mov    ("eax","edx");
 247         &shl    ("edi",4);      # hi<<4
 248         &xor    ("ebx","esi");
 249
 250         &shr    ("ecx",7-2);    # lo>>7
 251         &xor    ("eax","edi");
 252         &shr    ("edx",7-2);    # hi>>7
 253         &xor    ("ebx","ecx");
 254         &shl    ("esi",25-4);   # lo<<25
 255         &xor    ("eax","edx");
 256         &shl    ("edi",25-4);   # hi<<25
 257         &xor    ("eax","esi");
 258
 259         &shr    ("ecx",28-7);   # lo>>28
 260         &xor    ("ebx","edi");
 261         &shr    ("edx",28-7);   # hi>>28
 262         &xor    ("eax","ecx");
 263         &shl    ("esi",30-25);  # lo<<30
 264         &xor    ("ebx","edx");
 265         &shl    ("edi",30-25);  # hi<<30
 266         &xor    ("eax","esi");
 267         &xor    ("ebx","edi");                  # Sigma0(a)
 268
 269         &mov    ("ecx",$Alo);
 270         &mov    ("edx",$Ahi);
 271         &mov    ("esi",$Blo);
 272         &mov    ("edi",$Bhi);
 273         &add    ("eax",$Tlo);
 274         &adc    ("ebx",$Thi);                   # T1 = Sigma0(a)+T1
 275         &or     ("ecx","esi");
 276         &or     ("edx","edi");
 277         &and    ("ecx",$Clo);
 278         &and    ("edx",$Chi);
 279         &and    ("esi",$Alo);
 280         &and    ("edi",$Ahi);
 281         &or     ("ecx","esi");
 282         &or     ("edx","edi");                  # Maj(a,b,c) = ((a|b)&c)|(a&b)
 283
 284         &add    ("eax","ecx");
 285         &adc    ("ebx","edx");                  # T1 += Maj(a,b,c)
 286         &mov    ($Tlo,"eax");
 287         &mov    ($Thi,"ebx");
 288
 289         &mov    (&LB("edx"),&BP(0,$K512));      # pre-fetch LSB of *K
 290         &sub    ("esp",8);
 291         &lea    ($K512,&DWP(8,$K512));          # K++
 292 }
 293
 294
 295 &function_begin("sha512_block_data_order");
 296         &mov    ("esi",wparam(0));      # ctx
 297         &mov    ("edi",wparam(1));      # inp
 298         &mov    ("eax",wparam(2));      # num
 299         &mov    ("ebx","esp");          # saved sp
 300
 301         &call   (&label("pic_point"));  # make it PIC!
 302 &set_label("pic_point");
 303         &blindpop($K512);
 304         &lea    ($K512,&DWP(&label("K512")."-".&label("pic_point"),$K512));
 305
 306         &sub    ("esp",16);
 307         &and    ("esp",-64);
 308
 309         &shl    ("eax",7);
 310         &add    ("eax","edi");
 311         &mov    (&DWP(0,"esp"),"esi");  # ctx
 312         &mov    (&DWP(4,"esp"),"edi");  # inp
 313         &mov    (&DWP(8,"esp"),"eax");  # inp+num*128
 314         &mov    (&DWP(12,"esp"),"ebx"); # saved sp
 315
 316 if ($sse2) {
 317         &picmeup("edx","OPENSSL_ia32cap_P",$K512,&label("K512"));
 318         &mov    ("ecx",&DWP(0,"edx"));
 319         &test   ("ecx",1<<26);
 320         &jz     (&label("loop_x86"));
 321
 322         &mov    ("edx",&DWP(4,"edx"));
 323
 324         # load ctx->h[0-7]
 325         &movq   ($A,&QWP(0,"esi"));
 326          &and   ("ecx",1<<24);          # XMM registers availability
 327         &movq   ("mm1",&QWP(8,"esi"));
 328          &and   ("edx",1<<9);           # SSSE3 bit
 329         &movq   ($BxC,&QWP(16,"esi"));
 330          &or    ("ecx","edx");
 331         &movq   ("mm3",&QWP(24,"esi"));
 332         &movq   ($E,&QWP(32,"esi"));
 333         &movq   ("mm5",&QWP(40,"esi"));
 334         &movq   ("mm6",&QWP(48,"esi"));
 335         &movq   ("mm7",&QWP(56,"esi"));
 336         &cmp    ("ecx",1<<24|1<<9);
 337         &je     (&label("SSSE3"));
 338         &sub    ("esp",8*10);
 339         &jmp    (&label("loop_sse2"));
 340
 341 &set_label("loop_sse2",16);
 342         #&movq  ($Asse2,$A);
 343         &movq   ($Bsse2,"mm1");
 344         &movq   ($Csse2,$BxC);
 345         &movq   ($Dsse2,"mm3");
 346         #&movq  ($Esse2,$E);
 347         &movq   ($Fsse2,"mm5");
 348         &movq   ($Gsse2,"mm6");
 349         &pxor   ($BxC,"mm1");                   # magic
 350         &movq   ($Hsse2,"mm7");
 351         &movq   ("mm3",$A);                     # magic
 352
 353         &mov    ("eax",&DWP(0,"edi"));
 354         &mov    ("ebx",&DWP(4,"edi"));
 355         &add    ("edi",8);
 356         &mov    ("edx",15);                     # counter
 357         &bswap  ("eax");
 358         &bswap  ("ebx");
 359         &jmp    (&label("00_14_sse2"));
 360
 361 &set_label("00_14_sse2",16);
 362         &movd   ("mm1","eax");
 363         &mov    ("eax",&DWP(0,"edi"));
 364         &movd   ("mm7","ebx");
 365         &mov    ("ebx",&DWP(4,"edi"));
 366         &add    ("edi",8);
 367         &bswap  ("eax");
 368         &bswap  ("ebx");
 369         &punpckldq("mm7","mm1");
 370
 371         &BODY_00_15_sse2();
 372
 373         &dec    ("edx");
 374         &jnz    (&label("00_14_sse2"));
 375
 376         &movd   ("mm1","eax");
 377         &movd   ("mm7","ebx");
 378         &punpckldq("mm7","mm1");
 379
 380         &BODY_00_15_sse2(1);
 381
 382         &pxor   ($A,$A);                        # A is in %mm3
 383         &mov    ("edx",32);                     # counter
 384         &jmp    (&label("16_79_sse2"));
 385
 386 &set_label("16_79_sse2",16);
 387     for ($j=0;$j<2;$j++) {                      # 2x unroll
 388         #&movq  ("mm7",&QWP(8*(9+16-1),"esp")); # prefetched in BODY_00_15
 389         &movq   ("mm5",&QWP(8*(9+16-14),"esp"));
 390         &movq   ("mm1","mm7");
 391         &psrlq  ("mm7",1);
 392          &movq  ("mm6","mm5");
 393          &psrlq ("mm5",6);
 394         &psllq  ("mm1",56);
 395          &paddq ($A,"mm3");                     # from BODY_00_15
 396          &movq  ("mm3","mm7");
 397         &psrlq  ("mm7",7-1);
 398          &pxor  ("mm3","mm1");
 399          &psllq ("mm1",63-56);
 400         &pxor   ("mm3","mm7");
 401          &psrlq ("mm7",8-7);
 402         &pxor   ("mm3","mm1");
 403          &movq  ("mm1","mm5");
 404          &psrlq ("mm5",19-6);
 405         &pxor   ("mm7","mm3");                  # sigma0
 406
 407          &psllq ("mm6",3);
 408          &pxor  ("mm1","mm5");
 409         &paddq  ("mm7",&QWP(8*(9+16),"esp"));
 410          &pxor  ("mm1","mm6");
 411          &psrlq ("mm5",61-19);
 412         &paddq  ("mm7",&QWP(8*(9+16-9),"esp"));
 413          &pxor  ("mm1","mm5");
 414          &psllq ("mm6",45-3);
 415         &movq   ("mm5",$Fsse2);                 # load f
 416          &pxor  ("mm1","mm6");                  # sigma1
 417         &movq   ("mm6",$Gsse2);                 # load g
 418
 419         &paddq  ("mm7","mm1");                  # X[i]
 420         #&movq  (&QWP(8*9,"esp"),"mm7");        # moved to BODY_00_15
 421
 422         &BODY_00_15_sse2(2);
 423     }
 424         &dec    ("edx");
 425         &jnz    (&label("16_79_sse2"));
 426
 427         #&movq  ($A,$Asse2);
 428         &paddq  ($A,"mm3");                     # from BODY_00_15
 429         &movq   ("mm1",$Bsse2);
 430         #&movq  ($BxC,$Csse2);
 431         &movq   ("mm3",$Dsse2);
 432         #&movq  ($E,$Esse2);
 433         &movq   ("mm5",$Fsse2);
 434         &movq   ("mm6",$Gsse2);
 435         &movq   ("mm7",$Hsse2);
 436
 437         &pxor   ($BxC,"mm1");                   # de-magic
 438         &paddq  ($A,&QWP(0,"esi"));
 439         &paddq  ("mm1",&QWP(8,"esi"));
 440         &paddq  ($BxC,&QWP(16,"esi"));
 441         &paddq  ("mm3",&QWP(24,"esi"));
 442         &paddq  ($E,&QWP(32,"esi"));
 443         &paddq  ("mm5",&QWP(40,"esi"));
 444         &paddq  ("mm6",&QWP(48,"esi"));
 445         &paddq  ("mm7",&QWP(56,"esi"));
 446
 447         &mov    ("eax",8*80);
 448         &movq   (&QWP(0,"esi"),$A);
 449         &movq   (&QWP(8,"esi"),"mm1");
 450         &movq   (&QWP(16,"esi"),$BxC);
 451         &movq   (&QWP(24,"esi"),"mm3");
 452         &movq   (&QWP(32,"esi"),$E);
 453         &movq   (&QWP(40,"esi"),"mm5");
 454         &movq   (&QWP(48,"esi"),"mm6");
 455         &movq   (&QWP(56,"esi"),"mm7");
 456
 457         &lea    ("esp",&DWP(0,"esp","eax"));    # destroy frame
 458         &sub    ($K512,"eax");                  # rewind K
 459
 460         &cmp    ("edi",&DWP(8*10+8,"esp"));     # are we done yet?
 461         &jb     (&label("loop_sse2"));
 462
 463         &mov    ("esp",&DWP(8*10+12,"esp"));    # restore sp
 464         &emms   ();
 465 &function_end_A();
 466
 467 &set_label("SSSE3",32);
 468 { my ($cnt,$frame)=("ecx","edx");
 469   my @X=map("xmm$_",(0..7));
 470   my $j;
 471   my $i=0;
 472
 473         &lea    ($frame,&DWP(-64,"esp"));
 474         &sub    ("esp",256);
 475
 476         # fixed stack frame layout
 477         #
 478         # +0    A B C D E F G H         # backing store
 479         # +64   X[0]+K[i] .. X[15]+K[i] # XMM->MM xfer area
 480         # +192                          # XMM off-load ring buffer
 481         # +256                          # saved parameters
 482
 483         &movdqa         (@X[1],&QWP(80*8,$K512));               # byte swap mask
 484         &movdqu         (@X[0],&QWP(0,"edi"));
 485         &pshufb         (@X[0],@X[1]);
 486     for ($j=0;$j<8;$j++) {
 487         &movdqa         (&QWP(16*(($j-1)%4),$frame),@X[3])      if ($j>4); # off-load
 488         &movdqa         (@X[3],&QWP(16*($j%8),$K512));
 489         &movdqa         (@X[2],@X[1])                           if ($j<7); # perpetuate byte swap mask
 490         &movdqu         (@X[1],&QWP(16*($j+1),"edi"))           if ($j<7); # next input
 491         &movdqa         (@X[1],&QWP(16*(($j+1)%4),$frame))      if ($j==7);# restore @X[0]
 492         &paddq          (@X[3],@X[0]);
 493         &pshufb         (@X[1],@X[2])                           if ($j<7);
 494         &movdqa         (&QWP(16*($j%8)-128,$frame),@X[3]);     # xfer X[i]+K[i]
 495
 496         push(@X,shift(@X));                                     # rotate(@X)
 497     }
 498         #&jmp           (&label("loop_ssse3"));
 499         &nop            ();
 500
 501 &set_label("loop_ssse3",32);
 502         &movdqa         (@X[2],&QWP(16*(($j+1)%4),$frame));     # pre-restore @X[1]
 503         &movdqa         (&QWP(16*(($j-1)%4),$frame),@X[3]);     # off-load @X[3]
 504         &lea            ($K512,&DWP(16*8,$K512));
 505
 506         #&movq  ($Asse2,$A);                    # off-load A-H
 507         &movq   ($Bsse2,"mm1");
 508          &mov   ("ebx","edi");
 509         &movq   ($Csse2,$BxC);
 510          &lea   ("edi",&DWP(128,"edi"));        # advance input
 511         &movq   ($Dsse2,"mm3");
 512          &cmp   ("edi","eax");
 513         #&movq  ($Esse2,$E);
 514         &movq   ($Fsse2,"mm5");
 515          &cmovb ("ebx","edi");
 516         &movq   ($Gsse2,"mm6");
 517          &mov   ("ecx",4);                      # loop counter
 518         &pxor   ($BxC,"mm1");                   # magic
 519         &movq   ($Hsse2,"mm7");
 520         &pxor   ("mm3","mm3");                  # magic
 521
 522         &jmp            (&label("00_47_ssse3"));
 523
 524 sub BODY_00_15_ssse3 {          # "phase-less" copy of BODY_00_15_sse2
 525         (
 526         '&movq  ("mm1",$E)',                            # %mm1 is sliding right
 527         '&movq  ("mm7",&QWP(((-8*$i)%128)-128,$frame))',# X[i]+K[i]
 528          '&pxor ("mm5","mm6")',                         # f^=g
 529         '&psrlq ("mm1",14)',
 530          '&movq (&QWP(8*($i+4)%64,"esp"),$E)',          # modulo-scheduled save e
 531          '&pand ("mm5",$E)',                            # f&=e
 532         '&psllq ($E,23)',                               # $E is sliding left
 533         '&paddq ($A,"mm3")',                            # [h+=Maj(a,b,c)]
 534         '&movq  ("mm3","mm1")',                         # %mm3 is T1
 535          '&psrlq("mm1",4)',
 536          '&pxor ("mm5","mm6")',                         # Ch(e,f,g)
 537         '&pxor  ("mm3",$E)',
 538          '&psllq($E,23)',
 539         '&pxor  ("mm3","mm1")',
 540          '&movq (&QWP(8*$i%64,"esp"),$A)',              # modulo-scheduled save a
 541          '&paddq("mm7","mm5")',                         # X[i]+=Ch(e,f,g)
 542         '&pxor  ("mm3",$E)',
 543          '&psrlq("mm1",23)',
 544          '&paddq("mm7",&QWP(8*($i+7)%64,"esp"))',       # X[i]+=h
 545         '&pxor  ("mm3","mm1")',
 546          '&psllq($E,4)',
 547         '&pxor  ("mm3",$E)',                            # T1=Sigma1_512(e)
 548
 549          '&movq ($E,&QWP(8*($i+3)%64,"esp"))',          # e = load d, e in next round
 550         '&paddq ("mm3","mm7")',                         # T1+=X[i]
 551          '&movq ("mm5",$A)',                            # %mm5 is sliding right
 552          '&psrlq("mm5",28)',
 553         '&paddq ($E,"mm3")',                            # d += T1
 554          '&movq ("mm6",$A)',                            # %mm6 is sliding left
 555          '&movq ("mm7","mm5")',
 556          '&psllq("mm6",25)',
 557         '&movq  ("mm1",&QWP(8*($i+1)%64,"esp"))',       # load b
 558          '&psrlq("mm5",6)',
 559          '&pxor ("mm7","mm6")',
 560          '&psllq("mm6",5)',
 561          '&pxor ("mm7","mm5")',
 562         '&pxor  ($A,"mm1")',                            # a^b, b^c in next round
 563          '&psrlq("mm5",5)',
 564          '&pxor ("mm7","mm6")',
 565         '&pand  ($BxC,$A)',                             # (b^c)&(a^b)
 566          '&psllq("mm6",6)',
 567          '&pxor ("mm7","mm5")',
 568         '&pxor  ($BxC,"mm1")',                          # [h=]Maj(a,b,c)
 569          '&pxor ("mm6","mm7")',                         # Sigma0_512(a)
 570          '&movq ("mm5",&QWP(8*($i+5-1)%64,"esp"))',     # pre-load f
 571         '&paddq ($BxC,"mm6")',                          # h+=Sigma0(a)
 572          '&movq ("mm6",&QWP(8*($i+6-1)%64,"esp"))',     # pre-load g
 573
 574         '($A,$BxC) = ($BxC,$A); $i--;'
 575         );
 576 }
 577
 578 &set_label("00_47_ssse3",32);
 579
 580     for(;$j<16;$j++) {
 581         my ($t0,$t2,$t1)=@X[2..4];
 582         my @insns = (&BODY_00_15_ssse3(),&BODY_00_15_ssse3());
 583
 584         &movdqa         ($t2,@X[5]);
 585         &movdqa         (@X[1],$t0);                    # restore @X[1]
 586         &palignr        ($t0,@X[0],8);                  # X[1..2]
 587         &movdqa         (&QWP(16*($j%4),$frame),@X[4]); # off-load @X[4]
 588          &palignr       ($t2,@X[4],8);                  # X[9..10]
 589
 590         &movdqa         ($t1,$t0);
 591         &psrlq          ($t0,7);
 592          &paddq         (@X[0],$t2);                    # X[0..1] += X[9..10]
 593         &movdqa         ($t2,$t1);
 594         &psrlq          ($t1,1);
 595         &psllq          ($t2,64-8);
 596         &pxor           ($t0,$t1);
 597         &psrlq          ($t1,8-1);
 598         &pxor           ($t0,$t2);
 599         &psllq          ($t2,8-1);
 600         &pxor           ($t0,$t1);
 601          &movdqa        ($t1,@X[7]);
 602         &pxor           ($t0,$t2);                      # sigma0(X[1..2])
 603          &movdqa        ($t2,@X[7]);
 604          &psrlq         ($t1,6);
 605         &paddq          (@X[0],$t0);                    # X[0..1] += sigma0(X[1..2])
 606
 607         &movdqa         ($t0,@X[7]);
 608         &psrlq          ($t2,19);
 609         &psllq          ($t0,64-61);
 610         &pxor           ($t1,$t2);
 611         &psrlq          ($t2,61-19);
 612         &pxor           ($t1,$t0);
 613         &psllq          ($t0,61-19);
 614         &pxor           ($t1,$t2);
 615         &movdqa         ($t2,&QWP(16*(($j+2)%4),$frame));# pre-restore @X[1]
 616         &pxor           ($t1,$t0);                      # sigma0(X[1..2])
 617         &movdqa         ($t0,&QWP(16*($j%8),$K512));
 618          eval(shift(@insns));
 619         &paddq          (@X[0],$t1);                    # X[0..1] += sigma0(X[14..15])
 620          eval(shift(@insns));
 621          eval(shift(@insns));
 622          eval(shift(@insns));
 623          eval(shift(@insns));
 624         &paddq          ($t0,@X[0]);
 625          foreach(@insns) { eval; }
 626         &movdqa         (&QWP(16*($j%8)-128,$frame),$t0);# xfer X[i]+K[i]
 627
 628         push(@X,shift(@X));                             # rotate(@X)
 629     }
 630         &lea            ($K512,&DWP(16*8,$K512));
 631         &dec            ("ecx");
 632         &jnz            (&label("00_47_ssse3"));
 633
 634         &movdqa         (@X[1],&QWP(0,$K512));          # byte swap mask
 635         &lea            ($K512,&DWP(-80*8,$K512));      # rewind
 636         &movdqu         (@X[0],&QWP(0,"ebx"));
 637         &pshufb         (@X[0],@X[1]);
 638
 639     for ($j=0;$j<8;$j++) {      # load next or same block
 640         my @insns = (&BODY_00_15_ssse3(),&BODY_00_15_ssse3());
 641
 642         &movdqa         (&QWP(16*(($j-1)%4),$frame),@X[3])      if ($j>4); # off-load
 643         &movdqa         (@X[3],&QWP(16*($j%8),$K512));
 644         &movdqa         (@X[2],@X[1])                           if ($j<7); # perpetuate byte swap mask
 645         &movdqu         (@X[1],&QWP(16*($j+1),"ebx"))           if ($j<7); # next input
 646         &movdqa         (@X[1],&QWP(16*(($j+1)%4),$frame))      if ($j==7);# restore @X[0]
 647         &paddq          (@X[3],@X[0]);
 648         &pshufb         (@X[1],@X[2])                           if ($j<7);
 649          foreach(@insns) { eval; }
 650         &movdqa         (&QWP(16*($j%8)-128,$frame),@X[3]);# xfer X[i]+K[i]
 651
 652         push(@X,shift(@X));                             # rotate(@X)
 653     }
 654
 655         #&movq  ($A,$Asse2);                    # load A-H
 656         &movq   ("mm1",$Bsse2);
 657         &paddq  ($A,"mm3");                     # from BODY_00_15
 658         #&movq  ($BxC,$Csse2);
 659         &movq   ("mm3",$Dsse2);
 660         #&movq  ($E,$Esse2);
 661         #&movq  ("mm5",$Fsse2);
 662         #&movq  ("mm6",$Gsse2);
 663         &movq   ("mm7",$Hsse2);
 664
 665         &pxor   ($BxC,"mm1");                   # de-magic
 666         &paddq  ($A,&QWP(0,"esi"));
 667         &paddq  ("mm1",&QWP(8,"esi"));
 668         &paddq  ($BxC,&QWP(16,"esi"));
 669         &paddq  ("mm3",&QWP(24,"esi"));
 670         &paddq  ($E,&QWP(32,"esi"));
 671         &paddq  ("mm5",&QWP(40,"esi"));
 672         &paddq  ("mm6",&QWP(48,"esi"));
 673         &paddq  ("mm7",&QWP(56,"esi"));
 674
 675         &movq   (&QWP(0,"esi"),$A);
 676         &movq   (&QWP(8,"esi"),"mm1");
 677         &movq   (&QWP(16,"esi"),$BxC);
 678         &movq   (&QWP(24,"esi"),"mm3");
 679         &movq   (&QWP(32,"esi"),$E);
 680         &movq   (&QWP(40,"esi"),"mm5");
 681         &movq   (&QWP(48,"esi"),"mm6");
 682         &movq   (&QWP(56,"esi"),"mm7");
 683
 684         &cmp    ("edi","eax")                   # are we done yet?
 685         &jb     (&label("loop_ssse3"));
 686
 687         &mov    ("esp",&DWP(64+12,$frame));     # restore sp
 688         &emms   ();
 689 }
 690 &function_end_A();
 691 }
 692 &set_label("loop_x86",16);
 693     # copy input block to stack reversing byte and qword order
 694     for ($i=0;$i<8;$i++) {
 695         &mov    ("eax",&DWP($i*16+0,"edi"));
 696         &mov    ("ebx",&DWP($i*16+4,"edi"));
 697         &mov    ("ecx",&DWP($i*16+8,"edi"));
 698         &mov    ("edx",&DWP($i*16+12,"edi"));
 699         &bswap  ("eax");
 700         &bswap  ("ebx");
 701         &bswap  ("ecx");
 702         &bswap  ("edx");
 703         &push   ("eax");
 704         &push   ("ebx");
 705         &push   ("ecx");
 706         &push   ("edx");
 707     }
 708         &add    ("edi",128);
 709         &sub    ("esp",9*8);            # place for T,A,B,C,D,E,F,G,H
 710         &mov    (&DWP(8*(9+16)+4,"esp"),"edi");
 711
 712         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 713         &lea    ("edi",&DWP(8,"esp"));
 714         &mov    ("ecx",16);
 715         &data_word(0xA5F3F689);         # rep movsd
 716
 717 &set_label("00_15_x86",16);
 718         &BODY_00_15_x86();
 719
 720         &cmp    (&LB("edx"),0x94);
 721         &jne    (&label("00_15_x86"));
 722
 723 &set_label("16_79_x86",16);
 724         #define sigma0(x)       (ROTR((x),1)  ^ ROTR((x),8)  ^ ((x)>>7))
 725         #       LO              lo>>1^hi<<31  ^ lo>>8^hi<<24 ^ lo>>7^hi<<25
 726         #       HI              hi>>1^lo<<31  ^ hi>>8^lo<<24 ^ hi>>7
 727         &mov    ("ecx",&DWP(8*(9+15+16-1)+0,"esp"));
 728         &mov    ("edx",&DWP(8*(9+15+16-1)+4,"esp"));
 729         &mov    ("esi","ecx");
 730
 731         &shr    ("ecx",1);      # lo>>1
 732         &mov    ("edi","edx");
 733         &shr    ("edx",1);      # hi>>1
 734         &mov    ("eax","ecx");
 735         &shl    ("esi",24);     # lo<<24
 736         &mov    ("ebx","edx");
 737         &shl    ("edi",24);     # hi<<24
 738         &xor    ("ebx","esi");
 739
 740         &shr    ("ecx",7-1);    # lo>>7
 741         &xor    ("eax","edi");
 742         &shr    ("edx",7-1);    # hi>>7
 743         &xor    ("eax","ecx");
 744         &shl    ("esi",31-24);  # lo<<31
 745         &xor    ("ebx","edx");
 746         &shl    ("edi",25-24);  # hi<<25
 747         &xor    ("ebx","esi");
 748
 749         &shr    ("ecx",8-7);    # lo>>8
 750         &xor    ("eax","edi");
 751         &shr    ("edx",8-7);    # hi>>8
 752         &xor    ("eax","ecx");
 753         &shl    ("edi",31-25);  # hi<<31
 754         &xor    ("ebx","edx");
 755         &xor    ("eax","edi");                  # T1 = sigma0(X[-15])
 756
 757         &mov    (&DWP(0,"esp"),"eax");
 758         &mov    (&DWP(4,"esp"),"ebx");          # put T1 away
 759
 760         #define sigma1(x)       (ROTR((x),19) ^ ROTR((x),61) ^ ((x)>>6))
 761         #       LO              lo>>19^hi<<13 ^ hi>>29^lo<<3 ^ lo>>6^hi<<26
 762         #       HI              hi>>19^lo<<13 ^ lo>>29^hi<<3 ^ hi>>6
 763         &mov    ("ecx",&DWP(8*(9+15+16-14)+0,"esp"));
 764         &mov    ("edx",&DWP(8*(9+15+16-14)+4,"esp"));
 765         &mov    ("esi","ecx");
 766
 767         &shr    ("ecx",6);      # lo>>6
 768         &mov    ("edi","edx");
 769         &shr    ("edx",6);      # hi>>6
 770         &mov    ("eax","ecx");
 771         &shl    ("esi",3);      # lo<<3
 772         &mov    ("ebx","edx");
 773         &shl    ("edi",3);      # hi<<3
 774         &xor    ("eax","esi");
 775
 776         &shr    ("ecx",19-6);   # lo>>19
 777         &xor    ("ebx","edi");
 778         &shr    ("edx",19-6);   # hi>>19
 779         &xor    ("eax","ecx");
 780         &shl    ("esi",13-3);   # lo<<13
 781         &xor    ("ebx","edx");
 782         &shl    ("edi",13-3);   # hi<<13
 783         &xor    ("ebx","esi");
 784
 785         &shr    ("ecx",29-19);  # lo>>29
 786         &xor    ("eax","edi");
 787         &shr    ("edx",29-19);  # hi>>29
 788         &xor    ("ebx","ecx");
 789         &shl    ("edi",26-13);  # hi<<26
 790         &xor    ("eax","edx");
 791         &xor    ("eax","edi");                  # sigma1(X[-2])
 792
 793         &mov    ("ecx",&DWP(8*(9+15+16)+0,"esp"));
 794         &mov    ("edx",&DWP(8*(9+15+16)+4,"esp"));
 795         &add    ("eax",&DWP(0,"esp"));
 796         &adc    ("ebx",&DWP(4,"esp"));          # T1 = sigma1(X[-2])+T1
 797         &mov    ("esi",&DWP(8*(9+15+16-9)+0,"esp"));
 798         &mov    ("edi",&DWP(8*(9+15+16-9)+4,"esp"));
 799         &add    ("eax","ecx");
 800         &adc    ("ebx","edx");                  # T1 += X[-16]
 801         &add    ("eax","esi");
 802         &adc    ("ebx","edi");                  # T1 += X[-7]
 803         &mov    (&DWP(8*(9+15)+0,"esp"),"eax");
 804         &mov    (&DWP(8*(9+15)+4,"esp"),"ebx"); # save X[0]
 805
 806         &BODY_00_15_x86();
 807
 808         &cmp    (&LB("edx"),0x17);
 809         &jne    (&label("16_79_x86"));
 810
 811         &mov    ("esi",&DWP(8*(9+16+80)+0,"esp"));# ctx
 812         &mov    ("edi",&DWP(8*(9+16+80)+4,"esp"));# inp
 813     for($i=0;$i<4;$i++) {
 814         &mov    ("eax",&DWP($i*16+0,"esi"));
 815         &mov    ("ebx",&DWP($i*16+4,"esi"));
 816         &mov    ("ecx",&DWP($i*16+8,"esi"));
 817         &mov    ("edx",&DWP($i*16+12,"esi"));
 818         &add    ("eax",&DWP(8+($i*16)+0,"esp"));
 819         &adc    ("ebx",&DWP(8+($i*16)+4,"esp"));
 820         &mov    (&DWP($i*16+0,"esi"),"eax");
 821         &mov    (&DWP($i*16+4,"esi"),"ebx");
 822         &add    ("ecx",&DWP(8+($i*16)+8,"esp"));
 823         &adc    ("edx",&DWP(8+($i*16)+12,"esp"));
 824         &mov    (&DWP($i*16+8,"esi"),"ecx");
 825         &mov    (&DWP($i*16+12,"esi"),"edx");
 826     }
 827         &add    ("esp",8*(9+16+80));            # destroy frame
 828         &sub    ($K512,8*80);                   # rewind K
 829
 830         &cmp    ("edi",&DWP(8,"esp"));          # are we done yet?
 831         &jb     (&label("loop_x86"));
 832
 833         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 834 &function_end_A();
 835
 836 &set_label("K512",64);  # Yes! I keep it in the code segment!
 837         &data_word(0xd728ae22,0x428a2f98);      # u64
 838         &data_word(0x23ef65cd,0x71374491);      # u64
 839         &data_word(0xec4d3b2f,0xb5c0fbcf);      # u64
 840         &data_word(0x8189dbbc,0xe9b5dba5);      # u64
 841         &data_word(0xf348b538,0x3956c25b);      # u64
 842         &data_word(0xb605d019,0x59f111f1);      # u64
 843         &data_word(0xaf194f9b,0x923f82a4);      # u64
 844         &data_word(0xda6d8118,0xab1c5ed5);      # u64
 845         &data_word(0xa3030242,0xd807aa98);      # u64
 846         &data_word(0x45706fbe,0x12835b01);      # u64
 847         &data_word(0x4ee4b28c,0x243185be);      # u64
 848         &data_word(0xd5ffb4e2,0x550c7dc3);      # u64
 849         &data_word(0xf27b896f,0x72be5d74);      # u64
 850         &data_word(0x3b1696b1,0x80deb1fe);      # u64
 851         &data_word(0x25c71235,0x9bdc06a7);      # u64
 852         &data_word(0xcf692694,0xc19bf174);      # u64
 853         &data_word(0x9ef14ad2,0xe49b69c1);      # u64
 854         &data_word(0x384f25e3,0xefbe4786);      # u64
 855         &data_word(0x8b8cd5b5,0x0fc19dc6);      # u64
 856         &data_word(0x77ac9c65,0x240ca1cc);      # u64
 857         &data_word(0x592b0275,0x2de92c6f);      # u64
 858         &data_word(0x6ea6e483,0x4a7484aa);      # u64
 859         &data_word(0xbd41fbd4,0x5cb0a9dc);      # u64
 860         &data_word(0x831153b5,0x76f988da);      # u64
 861         &data_word(0xee66dfab,0x983e5152);      # u64
 862         &data_word(0x2db43210,0xa831c66d);      # u64
 863         &data_word(0x98fb213f,0xb00327c8);      # u64
 864         &data_word(0xbeef0ee4,0xbf597fc7);      # u64
 865         &data_word(0x3da88fc2,0xc6e00bf3);      # u64
 866         &data_word(0x930aa725,0xd5a79147);      # u64
 867         &data_word(0xe003826f,0x06ca6351);      # u64
 868         &data_word(0x0a0e6e70,0x14292967);      # u64
 869         &data_word(0x46d22ffc,0x27b70a85);      # u64
 870         &data_word(0x5c26c926,0x2e1b2138);      # u64
 871         &data_word(0x5ac42aed,0x4d2c6dfc);      # u64
 872         &data_word(0x9d95b3df,0x53380d13);      # u64
 873         &data_word(0x8baf63de,0x650a7354);      # u64
 874         &data_word(0x3c77b2a8,0x766a0abb);      # u64
 875         &data_word(0x47edaee6,0x81c2c92e);      # u64
 876         &data_word(0x1482353b,0x92722c85);      # u64
 877         &data_word(0x4cf10364,0xa2bfe8a1);      # u64
 878         &data_word(0xbc423001,0xa81a664b);      # u64
 879         &data_word(0xd0f89791,0xc24b8b70);      # u64
 880         &data_word(0x0654be30,0xc76c51a3);      # u64
 881         &data_word(0xd6ef5218,0xd192e819);      # u64
 882         &data_word(0x5565a910,0xd6990624);      # u64
 883         &data_word(0x5771202a,0xf40e3585);      # u64
 884         &data_word(0x32bbd1b8,0x106aa070);      # u64
 885         &data_word(0xb8d2d0c8,0x19a4c116);      # u64
 886         &data_word(0x5141ab53,0x1e376c08);      # u64
 887         &data_word(0xdf8eeb99,0x2748774c);      # u64
 888         &data_word(0xe19b48a8,0x34b0bcb5);      # u64
 889         &data_word(0xc5c95a63,0x391c0cb3);      # u64
 890         &data_word(0xe3418acb,0x4ed8aa4a);      # u64
 891         &data_word(0x7763e373,0x5b9cca4f);      # u64
 892         &data_word(0xd6b2b8a3,0x682e6ff3);      # u64
 893         &data_word(0x5defb2fc,0x748f82ee);      # u64
 894         &data_word(0x43172f60,0x78a5636f);      # u64
 895         &data_word(0xa1f0ab72,0x84c87814);      # u64
 896         &data_word(0x1a6439ec,0x8cc70208);      # u64
 897         &data_word(0x23631e28,0x90befffa);      # u64
 898         &data_word(0xde82bde9,0xa4506ceb);      # u64
 899         &data_word(0xb2c67915,0xbef9a3f7);      # u64
 900         &data_word(0xe372532b,0xc67178f2);      # u64
 901         &data_word(0xea26619c,0xca273ece);      # u64
 902         &data_word(0x21c0c207,0xd186b8c7);      # u64
 903         &data_word(0xcde0eb1e,0xeada7dd6);      # u64
 904         &data_word(0xee6ed178,0xf57d4f7f);      # u64
 905         &data_word(0x72176fba,0x06f067aa);      # u64
 906         &data_word(0xa2c898a6,0x0a637dc5);      # u64
 907         &data_word(0xbef90dae,0x113f9804);      # u64
 908         &data_word(0x131c471b,0x1b710b35);      # u64
 909         &data_word(0x23047d84,0x28db77f5);      # u64
 910         &data_word(0x40c72493,0x32caab7b);      # u64
 911         &data_word(0x15c9bebc,0x3c9ebe0a);      # u64
 912         &data_word(0x9c100d4c,0x431d67c4);      # u64
 913         &data_word(0xcb3e42b6,0x4cc5d4be);      # u64
 914         &data_word(0xfc657e2a,0x597f299c);      # u64
 915         &data_word(0x3ad6faec,0x5fcb6fab);      # u64
 916         &data_word(0x4a475817,0x6c44198c);      # u64
 917
 918         &data_word(0x04050607,0x00010203);      # byte swap
 919         &data_word(0x0c0d0e0f,0x08090a0b);      # mask
 920 &function_end_B("sha512_block_data_order");
 921 &asciz("SHA512 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
 922
 923 &asm_finish();
 924
 925 close STDOUT;