crypto/modes/asm/ghash-x86_64.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 # March, June 2010
  11 #
  12 # The module implements "4-bit" GCM GHASH function and underlying
  13 # single multiplication operation in GF(2^128). "4-bit" means that
  14 # it uses 256 bytes per-key table [+128 bytes shared table]. GHASH
  15 # function features so called "528B" variant utilizing additional
  16 # 256+16 bytes of per-key storage [+512 bytes shared table].
  17 # Performance results are for this streamed GHASH subroutine and are
  18 # expressed in cycles per processed byte, less is better:
  19 #
  20 #               gcc 3.4.x(*)    assembler
  21 #
  22 # P4            28.6            14.0            +100%
  23 # Opteron       19.3            7.7             +150%
  24 # Core2         17.8            8.1(**)         +120%
  25 #
  26 # (*)   comparison is not completely fair, because C results are
  27 #       for vanilla "256B" implementation, while assembler results
  28 #       are for "528B";-)
  29 # (**)  it's mystery [to me] why Core2 result is not same as for
  30 #       Opteron;
  31
  32 # May 2010
  33 #
  34 # Add PCLMULQDQ version performing at 2.02 cycles per processed byte.
  35 # See ghash-x86.pl for background information and details about coding
  36 # techniques.
  37 #
  38 # Special thanks to David Woodhouse <dwmw2@infradead.org> for
  39 # providing access to a Westmere-based system on behalf of Intel
  40 # Open Source Technology Centre.
  41
  42 $flavour = shift;
  43 $output  = shift;
  44 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  45
  46 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
  47
  48 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  49 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  50 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
  51 die "can't locate x86_64-xlate.pl";
  52
  53 open STDOUT,"| $^X $xlate $flavour $output";
  54
  55 # common register layout
  56 $nlo="%rax";
  57 $nhi="%rbx";
  58 $Zlo="%r8";
  59 $Zhi="%r9";
  60 $tmp="%r10";
  61 $rem_4bit = "%r11";
  62
  63 $Xi="%rdi";
  64 $Htbl="%rsi";
  65
  66 # per-function register layout
  67 $cnt="%rcx";
  68 $rem="%rdx";
  69
  70 sub LB() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/     or
  71                         $r =~ s/%[er]([sd]i)/%\1l/      or
  72                         $r =~ s/%[er](bp)/%\1l/         or
  73                         $r =~ s/%(r[0-9]+)[d]?/%\1b/;   $r; }
  74
  75 sub AUTOLOAD()          # thunk [simplified] 32-bit style perlasm
  76 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
  77   my $arg = pop;
  78     $arg = "\$$arg" if ($arg*1 eq $arg);
  79     $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
  80 }
  81 \f
  82 { my $N;
  83   sub loop() {
  84   my $inp = shift;
  85
  86         $N++;
  87 $code.=<<___;
  88         xor     $nlo,$nlo
  89         xor     $nhi,$nhi
  90         mov     `&LB("$Zlo")`,`&LB("$nlo")`
  91         mov     `&LB("$Zlo")`,`&LB("$nhi")`
  92         shl     \$4,`&LB("$nlo")`
  93         mov     \$14,$cnt
  94         mov     8($Htbl,$nlo),$Zlo
  95         mov     ($Htbl,$nlo),$Zhi
  96         and     \$0xf0,`&LB("$nhi")`
  97         mov     $Zlo,$rem
  98         jmp     .Loop$N
  99
 100 .align  16
 101 .Loop$N:
 102         shr     \$4,$Zlo
 103         and     \$0xf,$rem
 104         mov     $Zhi,$tmp
 105         mov     ($inp,$cnt),`&LB("$nlo")`
 106         shr     \$4,$Zhi
 107         xor     8($Htbl,$nhi),$Zlo
 108         shl     \$60,$tmp
 109         xor     ($Htbl,$nhi),$Zhi
 110         mov     `&LB("$nlo")`,`&LB("$nhi")`
 111         xor     ($rem_4bit,$rem,8),$Zhi
 112         mov     $Zlo,$rem
 113         shl     \$4,`&LB("$nlo")`
 114         xor     $tmp,$Zlo
 115         dec     $cnt
 116         js      .Lbreak$N
 117
 118         shr     \$4,$Zlo
 119         and     \$0xf,$rem
 120         mov     $Zhi,$tmp
 121         shr     \$4,$Zhi
 122         xor     8($Htbl,$nlo),$Zlo
 123         shl     \$60,$tmp
 124         xor     ($Htbl,$nlo),$Zhi
 125         and     \$0xf0,`&LB("$nhi")`
 126         xor     ($rem_4bit,$rem,8),$Zhi
 127         mov     $Zlo,$rem
 128         xor     $tmp,$Zlo
 129         jmp     .Loop$N
 130
 131 .align  16
 132 .Lbreak$N:
 133         shr     \$4,$Zlo
 134         and     \$0xf,$rem
 135         mov     $Zhi,$tmp
 136         shr     \$4,$Zhi
 137         xor     8($Htbl,$nlo),$Zlo
 138         shl     \$60,$tmp
 139         xor     ($Htbl,$nlo),$Zhi
 140         and     \$0xf0,`&LB("$nhi")`
 141         xor     ($rem_4bit,$rem,8),$Zhi
 142         mov     $Zlo,$rem
 143         xor     $tmp,$Zlo
 144
 145         shr     \$4,$Zlo
 146         and     \$0xf,$rem
 147         mov     $Zhi,$tmp
 148         shr     \$4,$Zhi
 149         xor     8($Htbl,$nhi),$Zlo
 150         shl     \$60,$tmp
 151         xor     ($Htbl,$nhi),$Zhi
 152         xor     $tmp,$Zlo
 153         xor     ($rem_4bit,$rem,8),$Zhi
 154
 155         bswap   $Zlo
 156         bswap   $Zhi
 157 ___
 158 }}
 159
 160 $code=<<___;
 161 .text
 162
 163 .globl  gcm_gmult_4bit
 164 .type   gcm_gmult_4bit,\@function,2
 165 .align  16
 166 gcm_gmult_4bit:
 167         push    %rbx
 168         push    %rbp            # %rbp and %r12 are pushed exclusively in
 169         push    %r12            # order to reuse Win64 exception handler...
 170 .Lgmult_prologue:
 171
 172         movzb   15($Xi),$Zlo
 173         lea     .Lrem_4bit(%rip),$rem_4bit
 174 ___
 175         &loop   ($Xi);
 176 $code.=<<___;
 177         mov     $Zlo,8($Xi)
 178         mov     $Zhi,($Xi)
 179
 180         mov     16(%rsp),%rbx
 181         lea     24(%rsp),%rsp
 182 .Lgmult_epilogue:
 183         ret
 184 .size   gcm_gmult_4bit,.-gcm_gmult_4bit
 185 ___
 186 \f
 187 # per-function register layout
 188 $inp="%rdx";
 189 $len="%rcx";
 190 $rem_8bit=$rem_4bit;
 191
 192 $code.=<<___;
 193 .globl  gcm_ghash_4bit
 194 .type   gcm_ghash_4bit,\@function,4
 195 .align  16
 196 gcm_ghash_4bit:
 197         push    %rbx
 198         push    %rbp
 199         push    %r12
 200         push    %r13
 201         push    %r14
 202         push    %r15
 203         sub     \$280,%rsp
 204 .Lghash_prologue:
 205         mov     $inp,%r14               # reassign couple of args
 206         mov     $len,%r15
 207 ___
 208 { my $inp="%r14";
 209   my $dat="%edx";
 210   my $len="%r15";
 211   my @nhi=("%ebx","%ecx");
 212   my @rem=("%r12","%r13");
 213   my $Hshr4="%rbp";
 214
 215         &sub    ($Htbl,-128);           # size optimization
 216         &lea    ($Hshr4,"16+128(%rsp)");
 217         { my @lo =($nlo,$nhi);
 218           my @hi =($Zlo,$Zhi);
 219
 220           &xor  ($dat,$dat);
 221           for ($i=0,$j=-2;$i<18;$i++,$j++) {
 222             &mov        ("$j(%rsp)",&LB($dat))          if ($i>1);
 223             &or         ($lo[0],$tmp)                   if ($i>1);
 224             &mov        (&LB($dat),&LB($lo[1]))         if ($i>0 && $i<17);
 225             &shr        ($lo[1],4)                      if ($i>0 && $i<17);
 226             &mov        ($tmp,$hi[1])                   if ($i>0 && $i<17);
 227             &shr        ($hi[1],4)                      if ($i>0 && $i<17);
 228             &mov        ("8*$j($Hshr4)",$hi[0])         if ($i>1);
 229             &mov        ($hi[0],"16*$i+0-128($Htbl)")   if ($i<16);
 230             &shl        (&LB($dat),4)                   if ($i>0 && $i<17);
 231             &mov        ("8*$j-128($Hshr4)",$lo[0])     if ($i>1);
 232             &mov        ($lo[0],"16*$i+8-128($Htbl)")   if ($i<16);
 233             &shl        ($tmp,60)                       if ($i>0 && $i<17);
 234
 235             push        (@lo,shift(@lo));
 236             push        (@hi,shift(@hi));
 237           }
 238         }
 239         &add    ($Htbl,-128);
 240         &mov    ($Zlo,"8($Xi)");
 241         &mov    ($Zhi,"0($Xi)");
 242         &add    ($len,$inp);            # pointer to the end of data
 243         &lea    ($rem_8bit,".Lrem_8bit(%rip)");
 244         &jmp    (".Louter_loop");
 245
 246 $code.=".align  16\n.Louter_loop:\n";
 247         &xor    ($Zhi,"($inp)");
 248         &mov    ("%rdx","8($inp)");
 249         &lea    ($inp,"16($inp)");
 250         &xor    ("%rdx",$Zlo);
 251         &mov    ("($Xi)",$Zhi);
 252         &mov    ("8($Xi)","%rdx");
 253         &shr    ("%rdx",32);
 254
 255         &xor    ($nlo,$nlo);
 256         &rol    ($dat,8);
 257         &mov    (&LB($nlo),&LB($dat));
 258         &movz   ($nhi[0],&LB($dat));
 259         &shl    (&LB($nlo),4);
 260         &shr    ($nhi[0],4);
 261
 262         for ($j=11,$i=0;$i<15;$i++) {
 263             &rol        ($dat,8);
 264             &xor        ($Zlo,"8($Htbl,$nlo)")                  if ($i>0);
 265             &xor        ($Zhi,"($Htbl,$nlo)")                   if ($i>0);
 266             &mov        ($Zlo,"8($Htbl,$nlo)")                  if ($i==0);
 267             &mov        ($Zhi,"($Htbl,$nlo)")                   if ($i==0);
 268
 269             &mov        (&LB($nlo),&LB($dat));
 270             &xor        ($Zlo,$tmp)                             if ($i>0);
 271             &movzw      ($rem[1],"($rem_8bit,$rem[1],2)")       if ($i>0);
 272
 273             &movz       ($nhi[1],&LB($dat));
 274             &shl        (&LB($nlo),4);
 275             &movzb      ($rem[0],"(%rsp,$nhi[0])");
 276
 277             &shr        ($nhi[1],4)                             if ($i<14);
 278             &and        ($nhi[1],0xf0)                          if ($i==14);
 279             &shl        ($rem[1],48)                            if ($i>0);
 280             &xor        ($rem[0],$Zlo);
 281
 282             &mov        ($tmp,$Zhi);
 283             &xor        ($Zhi,$rem[1])                          if ($i>0);
 284             &shr        ($Zlo,8);
 285
 286             &movz       ($rem[0],&LB($rem[0]));
 287             &mov        ($dat,"$j($Xi)")                        if (--$j%4==0);
 288             &shr        ($Zhi,8);
 289
 290             &xor        ($Zlo,"-128($Hshr4,$nhi[0],8)");
 291             &shl        ($tmp,56);
 292             &xor        ($Zhi,"($Hshr4,$nhi[0],8)");
 293
 294             unshift     (@nhi,pop(@nhi));               # "rotate" registers
 295             unshift     (@rem,pop(@rem));
 296         }
 297         &movzw  ($rem[1],"($rem_8bit,$rem[1],2)");
 298         &xor    ($Zlo,"8($Htbl,$nlo)");
 299         &xor    ($Zhi,"($Htbl,$nlo)");
 300
 301         &shl    ($rem[1],48);
 302         &xor    ($Zlo,$tmp);
 303
 304         &xor    ($Zhi,$rem[1]);
 305         &movz   ($rem[0],&LB($Zlo));
 306         &shr    ($Zlo,4);
 307
 308         &mov    ($tmp,$Zhi);
 309         &shl    (&LB($rem[0]),4);
 310         &shr    ($Zhi,4);
 311
 312         &xor    ($Zlo,"8($Htbl,$nhi[0])");
 313         &movzw  ($rem[0],"($rem_8bit,$rem[0],2)");
 314         &shl    ($tmp,60);
 315
 316         &xor    ($Zhi,"($Htbl,$nhi[0])");
 317         &xor    ($Zlo,$tmp);
 318         &shl    ($rem[0],48);
 319
 320         &bswap  ($Zlo);
 321         &xor    ($Zhi,$rem[0]);
 322
 323         &bswap  ($Zhi);
 324         &cmp    ($inp,$len);
 325         &jb     (".Louter_loop");
 326 }
 327 $code.=<<___;
 328         mov     $Zlo,8($Xi)
 329         mov     $Zhi,($Xi)
 330
 331         lea     280(%rsp),%rsi
 332         mov     0(%rsi),%r15
 333         mov     8(%rsi),%r14
 334         mov     16(%rsi),%r13
 335         mov     24(%rsi),%r12
 336         mov     32(%rsi),%rbp
 337         mov     40(%rsi),%rbx
 338         lea     48(%rsi),%rsp
 339 .Lghash_epilogue:
 340         ret
 341 .size   gcm_ghash_4bit,.-gcm_ghash_4bit
 342 ___
 343 \f
 344 ######################################################################
 345 # PCLMULQDQ version.
 346
 347 @_4args=$win64? ("%rcx","%rdx","%r8", "%r9") :  # Win64 order
 348                 ("%rdi","%rsi","%rdx","%rcx");  # Unix order
 349
 350 ($Xi,$Xhi)=("%xmm0","%xmm1");   $Hkey="%xmm2";
 351 ($T1,$T2,$T3)=("%xmm3","%xmm4","%xmm5");
 352
 353 sub clmul64x64_T2 {     # minimal register pressure
 354 my ($Xhi,$Xi,$Hkey,$modulo)=@_;
 355
 356 $code.=<<___ if (!defined($modulo));
 357         movdqa          $Xi,$Xhi                #
 358         pshufd          \$0b01001110,$Xi,$T1
 359         pshufd          \$0b01001110,$Hkey,$T2
 360         pxor            $Xi,$T1                 #
 361         pxor            $Hkey,$T2
 362 ___
 363 $code.=<<___;
 364         pclmulqdq       \$0x00,$Hkey,$Xi        #######
 365         pclmulqdq       \$0x11,$Hkey,$Xhi       #######
 366         pclmulqdq       \$0x00,$T2,$T1          #######
 367         pxor            $Xi,$T1                 #
 368         pxor            $Xhi,$T1                #
 369
 370         movdqa          $T1,$T2                 #
 371         psrldq          \$8,$T1
 372         pslldq          \$8,$T2                 #
 373         pxor            $T1,$Xhi
 374         pxor            $T2,$Xi                 #
 375 ___
 376 }
 377
 378 sub reduction_alg9 {    # 17/13 times faster than Intel version
 379 my ($Xhi,$Xi) = @_;
 380
 381 $code.=<<___;
 382         # 1st phase
 383         movdqa          $Xi,$T1                 #
 384         psllq           \$1,$Xi
 385         pxor            $T1,$Xi                 #
 386         psllq           \$5,$Xi                 #
 387         pxor            $T1,$Xi                 #
 388         psllq           \$57,$Xi                #
 389         movdqa          $Xi,$T2                 #
 390         pslldq          \$8,$Xi
 391         psrldq          \$8,$T2                 #
 392         pxor            $T1,$Xi
 393         pxor            $T2,$Xhi                #
 394
 395         # 2nd phase
 396         movdqa          $Xi,$T2
 397         psrlq           \$5,$Xi
 398         pxor            $T2,$Xi                 #
 399         psrlq           \$1,$Xi                 #
 400         pxor            $T2,$Xi                 #
 401         pxor            $Xhi,$T2
 402         psrlq           \$1,$Xi                 #
 403         pxor            $T2,$Xi                 #
 404 ___
 405 }
 406 \f
 407 { my ($Htbl,$Xip)=@_4args;
 408
 409 $code.=<<___;
 410 .globl  gcm_init_clmul
 411 .type   gcm_init_clmul,\@abi-omnipotent
 412 .align  16
 413 gcm_init_clmul:
 414         movdqu          ($Xip),$Hkey
 415         pshufd          \$0b01001110,$Hkey,$Hkey        # dword swap
 416
 417         # <<1 twist
 418         pshufd          \$0b11111111,$Hkey,$T2  # broadcast uppermost dword
 419         movdqa          $Hkey,$T1
 420         psllq           \$1,$Hkey
 421         pxor            $T3,$T3                 #
 422         psrlq           \$63,$T1
 423         pcmpgtd         $T2,$T3                 # broadcast carry bit
 424         pslldq          \$8,$T1
 425         por             $T1,$Hkey               # H<<=1
 426
 427         # magic reduction
 428         pand            .L0x1c2_polynomial(%rip),$T3
 429         pxor            $T3,$Hkey               # if(carry) H^=0x1c2_polynomial
 430
 431         # calculate H^2
 432         movdqa          $Hkey,$Xi
 433 ___
 434         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);
 435         &reduction_alg9 ($Xhi,$Xi);
 436 $code.=<<___;
 437         movdqu          $Hkey,($Htbl)           # save H
 438         movdqu          $Xi,16($Htbl)           # save H^2
 439         ret
 440 .size   gcm_init_clmul,.-gcm_init_clmul
 441 ___
 442 }
 443
 444 { my ($Xip,$Htbl)=@_4args;
 445
 446 $code.=<<___;
 447 .globl  gcm_gmult_clmul
 448 .type   gcm_gmult_clmul,\@abi-omnipotent
 449 .align  16
 450 gcm_gmult_clmul:
 451         movdqu          ($Xip),$Xi
 452         movdqa          .Lbswap_mask(%rip),$T3
 453         movdqu          ($Htbl),$Hkey
 454         pshufb          $T3,$Xi
 455 ___
 456         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);
 457         &reduction_alg9 ($Xhi,$Xi);
 458 $code.=<<___;
 459         pshufb          $T3,$Xi
 460         movdqu          $Xi,($Xip)
 461         ret
 462 .size   gcm_gmult_clmul,.-gcm_gmult_clmul
 463 ___
 464 }
 465 \f
 466 { my ($Xip,$Htbl,$inp,$len)=@_4args;
 467   my $Xn="%xmm6";
 468   my $Xhn="%xmm7";
 469   my $Hkey2="%xmm8";
 470   my $T1n="%xmm9";
 471   my $T2n="%xmm10";
 472
 473 $code.=<<___;
 474 .globl  gcm_ghash_clmul
 475 .type   gcm_ghash_clmul,\@abi-omnipotent
 476 .align  16
 477 gcm_ghash_clmul:
 478 ___
 479 $code.=<<___ if ($win64);
 480 .LSEH_begin_gcm_ghash_clmul:
 481         # I can't trust assembler to use specific encoding:-(
 482         .byte   0x48,0x83,0xec,0x58             #sub    \$0x58,%rsp
 483         .byte   0x0f,0x29,0x34,0x24             #movaps %xmm6,(%rsp)
 484         .byte   0x0f,0x29,0x7c,0x24,0x10        #movdqa %xmm7,0x10(%rsp)
 485         .byte   0x44,0x0f,0x29,0x44,0x24,0x20   #movaps %xmm8,0x20(%rsp)
 486         .byte   0x44,0x0f,0x29,0x4c,0x24,0x30   #movaps %xmm9,0x30(%rsp)
 487         .byte   0x44,0x0f,0x29,0x54,0x24,0x40   #movaps %xmm10,0x40(%rsp)
 488 ___
 489 $code.=<<___;
 490         movdqa          .Lbswap_mask(%rip),$T3
 491
 492         movdqu          ($Xip),$Xi
 493         movdqu          ($Htbl),$Hkey
 494         pshufb          $T3,$Xi
 495
 496         sub             \$0x10,$len
 497         jz              .Lodd_tail
 498
 499         movdqu          16($Htbl),$Hkey2
 500         #######
 501         # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
 502         #       [(H*Ii+1) + (H*Xi+1)] mod P =
 503         #       [(H*Ii+1) + H^2*(Ii+Xi)] mod P
 504         #
 505         movdqu          ($inp),$T1              # Ii
 506         movdqu          16($inp),$Xn            # Ii+1
 507         pshufb          $T3,$T1
 508         pshufb          $T3,$Xn
 509         pxor            $T1,$Xi                 # Ii+Xi
 510 ___
 511         &clmul64x64_T2  ($Xhn,$Xn,$Hkey);       # H*Ii+1
 512 $code.=<<___;
 513         movdqa          $Xi,$Xhi                #
 514         pshufd          \$0b01001110,$Xi,$T1
 515         pshufd          \$0b01001110,$Hkey2,$T2
 516         pxor            $Xi,$T1                 #
 517         pxor            $Hkey2,$T2
 518
 519         lea             32($inp),$inp           # i+=2
 520         sub             \$0x20,$len
 521         jbe             .Leven_tail
 522
 523 .Lmod_loop:
 524 ___
 525         &clmul64x64_T2  ($Xhi,$Xi,$Hkey2,1);    # H^2*(Ii+Xi)
 526 $code.=<<___;
 527         movdqu          ($inp),$T1              # Ii
 528         pxor            $Xn,$Xi                 # (H*Ii+1) + H^2*(Ii+Xi)
 529         pxor            $Xhn,$Xhi
 530
 531         movdqu          16($inp),$Xn            # Ii+1
 532         pshufb          $T3,$T1
 533         pshufb          $T3,$Xn
 534
 535         movdqa          $Xn,$Xhn                #
 536         pshufd          \$0b01001110,$Xn,$T1n
 537         pshufd          \$0b01001110,$Hkey,$T2n
 538         pxor            $Xn,$T1n                #
 539         pxor            $Hkey,$T2n
 540          pxor           $T1,$Xhi                # "Ii+Xi", consume early
 541
 542           movdqa        $Xi,$T1                 # 1st phase
 543           psllq         \$1,$Xi
 544           pxor          $T1,$Xi                 #
 545           psllq         \$5,$Xi                 #
 546           pxor          $T1,$Xi                 #
 547         pclmulqdq       \$0x00,$Hkey,$Xn        #######
 548           psllq         \$57,$Xi                #
 549           movdqa        $Xi,$T2                 #
 550           pslldq        \$8,$Xi
 551           psrldq        \$8,$T2                 #
 552           pxor          $T1,$Xi
 553           pxor          $T2,$Xhi                #
 554
 555         pclmulqdq       \$0x11,$Hkey,$Xhn       #######
 556           movdqa        $Xi,$T2                 # 2nd phase
 557           psrlq         \$5,$Xi
 558           pxor          $T2,$Xi                 #
 559           psrlq         \$1,$Xi                 #
 560           pxor          $T2,$Xi                 #
 561           pxor          $Xhi,$T2
 562           psrlq         \$1,$Xi                 #
 563           pxor          $T2,$Xi                 #
 564
 565         pclmulqdq       \$0x00,$T2n,$T1n        #######
 566          movdqa         $Xi,$Xhi                #
 567          pshufd         \$0b01001110,$Xi,$T1
 568          pshufd         \$0b01001110,$Hkey2,$T2
 569          pxor           $Xi,$T1                 #
 570          pxor           $Hkey2,$T2
 571
 572         pxor            $Xn,$T1n                #
 573         pxor            $Xhn,$T1n               #
 574         movdqa          $T1n,$T2n               #
 575         psrldq          \$8,$T1n
 576         pslldq          \$8,$T2n                #
 577         pxor            $T1n,$Xhn
 578         pxor            $T2n,$Xn                #
 579
 580         lea             32($inp),$inp
 581         sub             \$0x20,$len
 582         ja              .Lmod_loop
 583
 584 .Leven_tail:
 585 ___
 586         &clmul64x64_T2  ($Xhi,$Xi,$Hkey2,1);    # H^2*(Ii+Xi)
 587 $code.=<<___;
 588         pxor            $Xn,$Xi                 # (H*Ii+1) + H^2*(Ii+Xi)
 589         pxor            $Xhn,$Xhi
 590 ___
 591         &reduction_alg9 ($Xhi,$Xi);
 592 $code.=<<___;
 593         test            $len,$len
 594         jnz             .Ldone
 595
 596 .Lodd_tail:
 597         movdqu          ($inp),$T1              # Ii
 598         pshufb          $T3,$T1
 599         pxor            $T1,$Xi                 # Ii+Xi
 600 ___
 601         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);       # H*(Ii+Xi)
 602         &reduction_alg9 ($Xhi,$Xi);
 603 $code.=<<___;
 604 .Ldone:
 605         pshufb          $T3,$Xi
 606         movdqu          $Xi,($Xip)
 607 ___
 608 $code.=<<___ if ($win64);
 609         movaps  (%rsp),%xmm6
 610         movaps  0x10(%rsp),%xmm7
 611         movaps  0x20(%rsp),%xmm8
 612         movaps  0x30(%rsp),%xmm9
 613         movaps  0x40(%rsp),%xmm10
 614         add     \$0x58,%rsp
 615 ___
 616 $code.=<<___;
 617         ret
 618 .LSEH_end_gcm_ghash_clmul:
 619 .size   gcm_ghash_clmul,.-gcm_ghash_clmul
 620 ___
 621 }
 622
 623 $code.=<<___;
 624 .align  64
 625 .Lbswap_mask:
 626         .byte   15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
 627 .L0x1c2_polynomial:
 628         .byte   1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
 629 .align  64
 630 .type   .Lrem_4bit,\@object
 631 .Lrem_4bit:
 632         .long   0,`0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`
 633         .long   0,`0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`
 634         .long   0,`0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`
 635         .long   0,`0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`
 636 .type   .Lrem_8bit,\@object
 637 .Lrem_8bit:
 638         .value  0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E
 639         .value  0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E
 640         .value  0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E
 641         .value  0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E
 642         .value  0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E
 643         .value  0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E
 644         .value  0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E
 645         .value  0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E
 646         .value  0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE
 647         .value  0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE
 648         .value  0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE
 649         .value  0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE
 650         .value  0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E
 651         .value  0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E
 652         .value  0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE
 653         .value  0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE
 654         .value  0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E
 655         .value  0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E
 656         .value  0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E
 657         .value  0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E
 658         .value  0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E
 659         .value  0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E
 660         .value  0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E
 661         .value  0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E
 662         .value  0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE
 663         .value  0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE
 664         .value  0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE
 665         .value  0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE
 666         .value  0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E
 667         .value  0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E
 668         .value  0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE
 669         .value  0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE
 670
 671 .asciz  "GHASH for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 672 .align  64
 673 ___
 674 \f
 675 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
 676 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
 677 if ($win64) {
 678 $rec="%rcx";
 679 $frame="%rdx";
 680 $context="%r8";
 681 $disp="%r9";
 682
 683 $code.=<<___;
 684 .extern __imp_RtlVirtualUnwind
 685 .type   se_handler,\@abi-omnipotent
 686 .align  16
 687 se_handler:
 688         push    %rsi
 689         push    %rdi
 690         push    %rbx
 691         push    %rbp
 692         push    %r12
 693         push    %r13
 694         push    %r14
 695         push    %r15
 696         pushfq
 697         sub     \$64,%rsp
 698
 699         mov     120($context),%rax      # pull context->Rax
 700         mov     248($context),%rbx      # pull context->Rip
 701
 702         mov     8($disp),%rsi           # disp->ImageBase
 703         mov     56($disp),%r11          # disp->HandlerData
 704
 705         mov     0(%r11),%r10d           # HandlerData[0]
 706         lea     (%rsi,%r10),%r10        # prologue label
 707         cmp     %r10,%rbx               # context->Rip<prologue label
 708         jb      .Lin_prologue
 709
 710         mov     152($context),%rax      # pull context->Rsp
 711
 712         mov     4(%r11),%r10d           # HandlerData[1]
 713         lea     (%rsi,%r10),%r10        # epilogue label
 714         cmp     %r10,%rbx               # context->Rip>=epilogue label
 715         jae     .Lin_prologue
 716
 717         lea     24(%rax),%rax           # adjust "rsp"
 718
 719         mov     -8(%rax),%rbx
 720         mov     -16(%rax),%rbp
 721         mov     -24(%rax),%r12
 722         mov     %rbx,144($context)      # restore context->Rbx
 723         mov     %rbp,160($context)      # restore context->Rbp
 724         mov     %r12,216($context)      # restore context->R12
 725
 726 .Lin_prologue:
 727         mov     8(%rax),%rdi
 728         mov     16(%rax),%rsi
 729         mov     %rax,152($context)      # restore context->Rsp
 730         mov     %rsi,168($context)      # restore context->Rsi
 731         mov     %rdi,176($context)      # restore context->Rdi
 732
 733         mov     40($disp),%rdi          # disp->ContextRecord
 734         mov     $context,%rsi           # context
 735         mov     \$`1232/8`,%ecx         # sizeof(CONTEXT)
 736         .long   0xa548f3fc              # cld; rep movsq
 737
 738         mov     $disp,%rsi
 739         xor     %rcx,%rcx               # arg1, UNW_FLAG_NHANDLER
 740         mov     8(%rsi),%rdx            # arg2, disp->ImageBase
 741         mov     0(%rsi),%r8             # arg3, disp->ControlPc
 742         mov     16(%rsi),%r9            # arg4, disp->FunctionEntry
 743         mov     40(%rsi),%r10           # disp->ContextRecord
 744         lea     56(%rsi),%r11           # &disp->HandlerData
 745         lea     24(%rsi),%r12           # &disp->EstablisherFrame
 746         mov     %r10,32(%rsp)           # arg5
 747         mov     %r11,40(%rsp)           # arg6
 748         mov     %r12,48(%rsp)           # arg7
 749         mov     %rcx,56(%rsp)           # arg8, (NULL)
 750         call    *__imp_RtlVirtualUnwind(%rip)
 751
 752         mov     \$1,%eax                # ExceptionContinueSearch
 753         add     \$64,%rsp
 754         popfq
 755         pop     %r15
 756         pop     %r14
 757         pop     %r13
 758         pop     %r12
 759         pop     %rbp
 760         pop     %rbx
 761         pop     %rdi
 762         pop     %rsi
 763         ret
 764 .size   se_handler,.-se_handler
 765
 766 .section        .pdata
 767 .align  4
 768         .rva    .LSEH_begin_gcm_gmult_4bit
 769         .rva    .LSEH_end_gcm_gmult_4bit
 770         .rva    .LSEH_info_gcm_gmult_4bit
 771
 772         .rva    .LSEH_begin_gcm_ghash_4bit
 773         .rva    .LSEH_end_gcm_ghash_4bit
 774         .rva    .LSEH_info_gcm_ghash_4bit
 775
 776         .rva    .LSEH_begin_gcm_ghash_clmul
 777         .rva    .LSEH_end_gcm_ghash_clmul
 778         .rva    .LSEH_info_gcm_ghash_clmul
 779
 780 .section        .xdata
 781 .align  8
 782 .LSEH_info_gcm_gmult_4bit:
 783         .byte   9,0,0,0
 784         .rva    se_handler
 785         .rva    .Lgmult_prologue,.Lgmult_epilogue       # HandlerData
 786 .LSEH_info_gcm_ghash_4bit:
 787         .byte   9,0,0,0
 788         .rva    se_handler
 789         .rva    .Lghash_prologue,.Lghash_epilogue       # HandlerData
 790 .LSEH_info_gcm_ghash_clmul:
 791         .byte   0x01,0x1f,0x0b,0x00
 792         .byte   0x1f,0xa8,0x04,0x00     #movaps 0x40(rsp),xmm10
 793         .byte   0x19,0x98,0x03,0x00     #movaps 0x30(rsp),xmm9
 794         .byte   0x13,0x88,0x02,0x00     #movaps 0x20(rsp),xmm8
 795         .byte   0x0d,0x78,0x01,0x00     #movaps 0x10(rsp),xmm7
 796         .byte   0x08,0x68,0x00,0x00     #movaps (rsp),xmm6
 797         .byte   0x04,0xa2,0x00,0x00     #sub    rsp,0x58
 798 ___
 799 }
 800 \f
 801 sub rex {
 802  local *opcode=shift;
 803  my ($dst,$src)=@_;
 804
 805    if ($dst>=8 || $src>=8) {
 806         $rex=0x40;
 807         $rex|=0x04 if($dst>=8);
 808         $rex|=0x01 if($src>=8);
 809         push @opcode,$rex;
 810    }
 811 }
 812
 813 sub pclmulqdq {
 814   my $arg=shift;
 815   my @opcode=(0x66);
 816
 817     if ($arg=~/\$([x0-9a-f]+),\s*%xmm([0-9]+),\s*%xmm([0-9]+)/) {
 818         rex(\@opcode,$3,$2);
 819         push @opcode,0x0f,0x3a,0x44;
 820         push @opcode,0xc0|($2&7)|(($3&7)<<3);   # ModR/M
 821         my $c=$1;
 822         push @opcode,$c=~/^0/?oct($c):$c;
 823         return ".byte\t".join(',',@opcode);
 824     }
 825     return "pclmulqdq\t".$arg;
 826 }
 827
 828 $code =~ s/\`([^\`]*)\`/eval($1)/gem;
 829 $code =~ s/\bpclmulqdq\s+(\$.*%xmm[0-9]+).*$/pclmulqdq($1)/gem;
 830
 831 print $code;
 832
 833 close STDOUT;