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