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 # Atom          31.6            16.8            +88%
  26 # VIA Nano      21.8            10.1            +115%
  27 #
  28 # (*)   comparison is not completely fair, because C results are
  29 #       for vanilla "256B" implementation, while assembler results
  30 #       are for "528B";-)
  31 # (**)  it's mystery [to me] why Core2 result is not same as for
  32 #       Opteron;
  33
  34 # May 2010
  35 #
  36 # Add PCLMULQDQ version performing at 2.02 cycles per processed byte.
  37 # See ghash-x86.pl for background information and details about coding
  38 # techniques.
  39 #
  40 # Special thanks to David Woodhouse <dwmw2@infradead.org> for
  41 # providing access to a Westmere-based system on behalf of Intel
  42 # Open Source Technology Centre.
  43
  44 $flavour = shift;
  45 $output  = shift;
  46 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  47
  48 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
  49
  50 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  51 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  52 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
  53 die "can't locate x86_64-xlate.pl";
  54
  55 open STDOUT,"| $^X $xlate $flavour $output";
  56
  57 # common register layout
  58 $nlo="%rax";
  59 $nhi="%rbx";
  60 $Zlo="%r8";
  61 $Zhi="%r9";
  62 $tmp="%r10";
  63 $rem_4bit = "%r11";
  64
  65 $Xi="%rdi";
  66 $Htbl="%rsi";
  67
  68 # per-function register layout
  69 $cnt="%rcx";
  70 $rem="%rdx";
  71
  72 sub LB() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/     or
  73                         $r =~ s/%[er]([sd]i)/%\1l/      or
  74                         $r =~ s/%[er](bp)/%\1l/         or
  75                         $r =~ s/%(r[0-9]+)[d]?/%\1b/;   $r; }
  76
  77 sub AUTOLOAD()          # thunk [simplified] 32-bit style perlasm
  78 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
  79   my $arg = pop;
  80     $arg = "\$$arg" if ($arg*1 eq $arg);
  81     $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
  82 }
  83 \f
  84 { my $N;
  85   sub loop() {
  86   my $inp = shift;
  87
  88         $N++;
  89 $code.=<<___;
  90         xor     $nlo,$nlo
  91         xor     $nhi,$nhi
  92         mov     `&LB("$Zlo")`,`&LB("$nlo")`
  93         mov     `&LB("$Zlo")`,`&LB("$nhi")`
  94         shl     \$4,`&LB("$nlo")`
  95         mov     \$14,$cnt
  96         mov     8($Htbl,$nlo),$Zlo
  97         mov     ($Htbl,$nlo),$Zhi
  98         and     \$0xf0,`&LB("$nhi")`
  99         mov     $Zlo,$rem
 100         jmp     .Loop$N
 101
 102 .align  16
 103 .Loop$N:
 104         shr     \$4,$Zlo
 105         and     \$0xf,$rem
 106         mov     $Zhi,$tmp
 107         mov     ($inp,$cnt),`&LB("$nlo")`
 108         shr     \$4,$Zhi
 109         xor     8($Htbl,$nhi),$Zlo
 110         shl     \$60,$tmp
 111         xor     ($Htbl,$nhi),$Zhi
 112         mov     `&LB("$nlo")`,`&LB("$nhi")`
 113         xor     ($rem_4bit,$rem,8),$Zhi
 114         mov     $Zlo,$rem
 115         shl     \$4,`&LB("$nlo")`
 116         xor     $tmp,$Zlo
 117         dec     $cnt
 118         js      .Lbreak$N
 119
 120         shr     \$4,$Zlo
 121         and     \$0xf,$rem
 122         mov     $Zhi,$tmp
 123         shr     \$4,$Zhi
 124         xor     8($Htbl,$nlo),$Zlo
 125         shl     \$60,$tmp
 126         xor     ($Htbl,$nlo),$Zhi
 127         and     \$0xf0,`&LB("$nhi")`
 128         xor     ($rem_4bit,$rem,8),$Zhi
 129         mov     $Zlo,$rem
 130         xor     $tmp,$Zlo
 131         jmp     .Loop$N
 132
 133 .align  16
 134 .Lbreak$N:
 135         shr     \$4,$Zlo
 136         and     \$0xf,$rem
 137         mov     $Zhi,$tmp
 138         shr     \$4,$Zhi
 139         xor     8($Htbl,$nlo),$Zlo
 140         shl     \$60,$tmp
 141         xor     ($Htbl,$nlo),$Zhi
 142         and     \$0xf0,`&LB("$nhi")`
 143         xor     ($rem_4bit,$rem,8),$Zhi
 144         mov     $Zlo,$rem
 145         xor     $tmp,$Zlo
 146
 147         shr     \$4,$Zlo
 148         and     \$0xf,$rem
 149         mov     $Zhi,$tmp
 150         shr     \$4,$Zhi
 151         xor     8($Htbl,$nhi),$Zlo
 152         shl     \$60,$tmp
 153         xor     ($Htbl,$nhi),$Zhi
 154         xor     $tmp,$Zlo
 155         xor     ($rem_4bit,$rem,8),$Zhi
 156
 157         bswap   $Zlo
 158         bswap   $Zhi
 159 ___
 160 }}
 161
 162 $code=<<___;
 163 .text
 164
 165 .globl  gcm_gmult_4bit
 166 .type   gcm_gmult_4bit,\@function,2
 167 .align  16
 168 gcm_gmult_4bit:
 169         push    %rbx
 170         push    %rbp            # %rbp and %r12 are pushed exclusively in
 171         push    %r12            # order to reuse Win64 exception handler...
 172 .Lgmult_prologue:
 173
 174         movzb   15($Xi),$Zlo
 175         lea     .Lrem_4bit(%rip),$rem_4bit
 176 ___
 177         &loop   ($Xi);
 178 $code.=<<___;
 179         mov     $Zlo,8($Xi)
 180         mov     $Zhi,($Xi)
 181
 182         mov     16(%rsp),%rbx
 183         lea     24(%rsp),%rsp
 184 .Lgmult_epilogue:
 185         ret
 186 .size   gcm_gmult_4bit,.-gcm_gmult_4bit
 187 ___
 188 \f
 189 # per-function register layout
 190 $inp="%rdx";
 191 $len="%rcx";
 192 $rem_8bit=$rem_4bit;
 193
 194 $code.=<<___;
 195 .globl  gcm_ghash_4bit
 196 .type   gcm_ghash_4bit,\@function,4
 197 .align  16
 198 gcm_ghash_4bit:
 199         push    %rbx
 200         push    %rbp
 201         push    %r12
 202         push    %r13
 203         push    %r14
 204         push    %r15
 205         sub     \$280,%rsp
 206 .Lghash_prologue:
 207         mov     $inp,%r14               # reassign couple of args
 208         mov     $len,%r15
 209 ___
 210 { my $inp="%r14";
 211   my $dat="%edx";
 212   my $len="%r15";
 213   my @nhi=("%ebx","%ecx");
 214   my @rem=("%r12","%r13");
 215   my $Hshr4="%rbp";
 216
 217         &sub    ($Htbl,-128);           # size optimization
 218         &lea    ($Hshr4,"16+128(%rsp)");
 219         { my @lo =($nlo,$nhi);
 220           my @hi =($Zlo,$Zhi);
 221
 222           &xor  ($dat,$dat);
 223           for ($i=0,$j=-2;$i<18;$i++,$j++) {
 224             &mov        ("$j(%rsp)",&LB($dat))          if ($i>1);
 225             &or         ($lo[0],$tmp)                   if ($i>1);
 226             &mov        (&LB($dat),&LB($lo[1]))         if ($i>0 && $i<17);
 227             &shr        ($lo[1],4)                      if ($i>0 && $i<17);
 228             &mov        ($tmp,$hi[1])                   if ($i>0 && $i<17);
 229             &shr        ($hi[1],4)                      if ($i>0 && $i<17);
 230             &mov        ("8*$j($Hshr4)",$hi[0])         if ($i>1);
 231             &mov        ($hi[0],"16*$i+0-128($Htbl)")   if ($i<16);
 232             &shl        (&LB($dat),4)                   if ($i>0 && $i<17);
 233             &mov        ("8*$j-128($Hshr4)",$lo[0])     if ($i>1);
 234             &mov        ($lo[0],"16*$i+8-128($Htbl)")   if ($i<16);
 235             &shl        ($tmp,60)                       if ($i>0 && $i<17);
 236
 237             push        (@lo,shift(@lo));
 238             push        (@hi,shift(@hi));
 239           }
 240         }
 241         &add    ($Htbl,-128);
 242         &mov    ($Zlo,"8($Xi)");
 243         &mov    ($Zhi,"0($Xi)");
 244         &add    ($len,$inp);            # pointer to the end of data
 245         &lea    ($rem_8bit,".Lrem_8bit(%rip)");
 246         &jmp    (".Louter_loop");
 247
 248 $code.=".align  16\n.Louter_loop:\n";
 249         &xor    ($Zhi,"($inp)");
 250         &mov    ("%rdx","8($inp)");
 251         &lea    ($inp,"16($inp)");
 252         &xor    ("%rdx",$Zlo);
 253         &mov    ("($Xi)",$Zhi);
 254         &mov    ("8($Xi)","%rdx");
 255         &shr    ("%rdx",32);
 256
 257         &xor    ($nlo,$nlo);
 258         &rol    ($dat,8);
 259         &mov    (&LB($nlo),&LB($dat));
 260         &movz   ($nhi[0],&LB($dat));
 261         &shl    (&LB($nlo),4);
 262         &shr    ($nhi[0],4);
 263
 264         for ($j=11,$i=0;$i<15;$i++) {
 265             &rol        ($dat,8);
 266             &xor        ($Zlo,"8($Htbl,$nlo)")                  if ($i>0);
 267             &xor        ($Zhi,"($Htbl,$nlo)")                   if ($i>0);
 268             &mov        ($Zlo,"8($Htbl,$nlo)")                  if ($i==0);
 269             &mov        ($Zhi,"($Htbl,$nlo)")                   if ($i==0);
 270
 271             &mov        (&LB($nlo),&LB($dat));
 272             &xor        ($Zlo,$tmp)                             if ($i>0);
 273             &movzw      ($rem[1],"($rem_8bit,$rem[1],2)")       if ($i>0);
 274
 275             &movz       ($nhi[1],&LB($dat));
 276             &shl        (&LB($nlo),4);
 277             &movzb      ($rem[0],"(%rsp,$nhi[0])");
 278
 279             &shr        ($nhi[1],4)                             if ($i<14);
 280             &and        ($nhi[1],0xf0)                          if ($i==14);
 281             &shl        ($rem[1],48)                            if ($i>0);
 282             &xor        ($rem[0],$Zlo);
 283
 284             &mov        ($tmp,$Zhi);
 285             &xor        ($Zhi,$rem[1])                          if ($i>0);
 286             &shr        ($Zlo,8);
 287
 288             &movz       ($rem[0],&LB($rem[0]));
 289             &mov        ($dat,"$j($Xi)")                        if (--$j%4==0);
 290             &shr        ($Zhi,8);
 291
 292             &xor        ($Zlo,"-128($Hshr4,$nhi[0],8)");
 293             &shl        ($tmp,56);
 294             &xor        ($Zhi,"($Hshr4,$nhi[0],8)");
 295
 296             unshift     (@nhi,pop(@nhi));               # "rotate" registers
 297             unshift     (@rem,pop(@rem));
 298         }
 299         &movzw  ($rem[1],"($rem_8bit,$rem[1],2)");
 300         &xor    ($Zlo,"8($Htbl,$nlo)");
 301         &xor    ($Zhi,"($Htbl,$nlo)");
 302
 303         &shl    ($rem[1],48);
 304         &xor    ($Zlo,$tmp);
 305
 306         &xor    ($Zhi,$rem[1]);
 307         &movz   ($rem[0],&LB($Zlo));
 308         &shr    ($Zlo,4);
 309
 310         &mov    ($tmp,$Zhi);
 311         &shl    (&LB($rem[0]),4);
 312         &shr    ($Zhi,4);
 313
 314         &xor    ($Zlo,"8($Htbl,$nhi[0])");
 315         &movzw  ($rem[0],"($rem_8bit,$rem[0],2)");
 316         &shl    ($tmp,60);
 317
 318         &xor    ($Zhi,"($Htbl,$nhi[0])");
 319         &xor    ($Zlo,$tmp);
 320         &shl    ($rem[0],48);
 321
 322         &bswap  ($Zlo);
 323         &xor    ($Zhi,$rem[0]);
 324
 325         &bswap  ($Zhi);
 326         &cmp    ($inp,$len);
 327         &jb     (".Louter_loop");
 328 }
 329 $code.=<<___;
 330         mov     $Zlo,8($Xi)
 331         mov     $Zhi,($Xi)
 332
 333         lea     280(%rsp),%rsi
 334         mov     0(%rsi),%r15
 335         mov     8(%rsi),%r14
 336         mov     16(%rsi),%r13
 337         mov     24(%rsi),%r12
 338         mov     32(%rsi),%rbp
 339         mov     40(%rsi),%rbx
 340         lea     48(%rsi),%rsp
 341 .Lghash_epilogue:
 342         ret
 343 .size   gcm_ghash_4bit,.-gcm_ghash_4bit
 344 ___
 345 \f
 346 ######################################################################
 347 # PCLMULQDQ version.
 348
 349 @_4args=$win64? ("%rcx","%rdx","%r8", "%r9") :  # Win64 order
 350                 ("%rdi","%rsi","%rdx","%rcx");  # Unix order
 351
 352 ($Xi,$Xhi)=("%xmm0","%xmm1");   $Hkey="%xmm2";
 353 ($T1,$T2,$T3)=("%xmm3","%xmm4","%xmm5");
 354
 355 sub clmul64x64_T2 {     # minimal register pressure
 356 my ($Xhi,$Xi,$Hkey,$modulo)=@_;
 357
 358 $code.=<<___ if (!defined($modulo));
 359         movdqa          $Xi,$Xhi                #
 360         pshufd          \$0b01001110,$Xi,$T1
 361         pshufd          \$0b01001110,$Hkey,$T2
 362         pxor            $Xi,$T1                 #
 363         pxor            $Hkey,$T2
 364 ___
 365 $code.=<<___;
 366         pclmulqdq       \$0x00,$Hkey,$Xi        #######
 367         pclmulqdq       \$0x11,$Hkey,$Xhi       #######
 368         pclmulqdq       \$0x00,$T2,$T1          #######
 369         pxor            $Xi,$T1                 #
 370         pxor            $Xhi,$T1                #
 371
 372         movdqa          $T1,$T2                 #
 373         psrldq          \$8,$T1
 374         pslldq          \$8,$T2                 #
 375         pxor            $T1,$Xhi
 376         pxor            $T2,$Xi                 #
 377 ___
 378 }
 379
 380 sub reduction_alg9 {    # 17/13 times faster than Intel version
 381 my ($Xhi,$Xi) = @_;
 382
 383 $code.=<<___;
 384         # 1st phase
 385         movdqa          $Xi,$T1                 #
 386         psllq           \$1,$Xi
 387         pxor            $T1,$Xi                 #
 388         psllq           \$5,$Xi                 #
 389         pxor            $T1,$Xi                 #
 390         psllq           \$57,$Xi                #
 391         movdqa          $Xi,$T2                 #
 392         pslldq          \$8,$Xi
 393         psrldq          \$8,$T2                 #
 394         pxor            $T1,$Xi
 395         pxor            $T2,$Xhi                #
 396
 397         # 2nd phase
 398         movdqa          $Xi,$T2
 399         psrlq           \$5,$Xi
 400         pxor            $T2,$Xi                 #
 401         psrlq           \$1,$Xi                 #
 402         pxor            $T2,$Xi                 #
 403         pxor            $Xhi,$T2
 404         psrlq           \$1,$Xi                 #
 405         pxor            $T2,$Xi                 #
 406 ___
 407 }
 408 \f
 409 { my ($Htbl,$Xip)=@_4args;
 410
 411 $code.=<<___;
 412 .globl  gcm_init_clmul
 413 .type   gcm_init_clmul,\@abi-omnipotent
 414 .align  16
 415 gcm_init_clmul:
 416         movdqu          ($Xip),$Hkey
 417         pshufd          \$0b01001110,$Hkey,$Hkey        # dword swap
 418
 419         # <<1 twist
 420         pshufd          \$0b11111111,$Hkey,$T2  # broadcast uppermost dword
 421         movdqa          $Hkey,$T1
 422         psllq           \$1,$Hkey
 423         pxor            $T3,$T3                 #
 424         psrlq           \$63,$T1
 425         pcmpgtd         $T2,$T3                 # broadcast carry bit
 426         pslldq          \$8,$T1
 427         por             $T1,$Hkey               # H<<=1
 428
 429         # magic reduction
 430         pand            .L0x1c2_polynomial(%rip),$T3
 431         pxor            $T3,$Hkey               # if(carry) H^=0x1c2_polynomial
 432
 433         # calculate H^2
 434         movdqa          $Hkey,$Xi
 435 ___
 436         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);
 437         &reduction_alg9 ($Xhi,$Xi);
 438 $code.=<<___;
 439         movdqu          $Hkey,($Htbl)           # save H
 440         movdqu          $Xi,16($Htbl)           # save H^2
 441         ret
 442 .size   gcm_init_clmul,.-gcm_init_clmul
 443 ___
 444 }
 445
 446 { my ($Xip,$Htbl)=@_4args;
 447
 448 $code.=<<___;
 449 .globl  gcm_gmult_clmul
 450 .type   gcm_gmult_clmul,\@abi-omnipotent
 451 .align  16
 452 gcm_gmult_clmul:
 453         movdqu          ($Xip),$Xi
 454         movdqa          .Lbswap_mask(%rip),$T3
 455         movdqu          ($Htbl),$Hkey
 456         pshufb          $T3,$Xi
 457 ___
 458         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);
 459         &reduction_alg9 ($Xhi,$Xi);
 460 $code.=<<___;
 461         pshufb          $T3,$Xi
 462         movdqu          $Xi,($Xip)
 463         ret
 464 .size   gcm_gmult_clmul,.-gcm_gmult_clmul
 465 ___
 466 }
 467 \f
 468 { my ($Xip,$Htbl,$inp,$len)=@_4args;
 469   my $Xn="%xmm6";
 470   my $Xhn="%xmm7";
 471   my $Hkey2="%xmm8";
 472   my $T1n="%xmm9";
 473   my $T2n="%xmm10";
 474
 475 $code.=<<___;
 476 .globl  gcm_ghash_clmul
 477 .type   gcm_ghash_clmul,\@abi-omnipotent
 478 .align  16
 479 gcm_ghash_clmul:
 480 ___
 481 $code.=<<___ if ($win64);
 482 .LSEH_begin_gcm_ghash_clmul:
 483         # I can't trust assembler to use specific encoding:-(
 484         .byte   0x48,0x83,0xec,0x58             #sub    \$0x58,%rsp
 485         .byte   0x0f,0x29,0x34,0x24             #movaps %xmm6,(%rsp)
 486         .byte   0x0f,0x29,0x7c,0x24,0x10        #movdqa %xmm7,0x10(%rsp)
 487         .byte   0x44,0x0f,0x29,0x44,0x24,0x20   #movaps %xmm8,0x20(%rsp)
 488         .byte   0x44,0x0f,0x29,0x4c,0x24,0x30   #movaps %xmm9,0x30(%rsp)
 489         .byte   0x44,0x0f,0x29,0x54,0x24,0x40   #movaps %xmm10,0x40(%rsp)
 490 ___
 491 $code.=<<___;
 492         movdqa          .Lbswap_mask(%rip),$T3
 493
 494         movdqu          ($Xip),$Xi
 495         movdqu          ($Htbl),$Hkey
 496         pshufb          $T3,$Xi
 497
 498         sub             \$0x10,$len
 499         jz              .Lodd_tail
 500
 501         movdqu          16($Htbl),$Hkey2
 502         #######
 503         # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
 504         #       [(H*Ii+1) + (H*Xi+1)] mod P =
 505         #       [(H*Ii+1) + H^2*(Ii+Xi)] mod P
 506         #
 507         movdqu          ($inp),$T1              # Ii
 508         movdqu          16($inp),$Xn            # Ii+1
 509         pshufb          $T3,$T1
 510         pshufb          $T3,$Xn
 511         pxor            $T1,$Xi                 # Ii+Xi
 512 ___
 513         &clmul64x64_T2  ($Xhn,$Xn,$Hkey);       # H*Ii+1
 514 $code.=<<___;
 515         movdqa          $Xi,$Xhi                #
 516         pshufd          \$0b01001110,$Xi,$T1
 517         pshufd          \$0b01001110,$Hkey2,$T2
 518         pxor            $Xi,$T1                 #
 519         pxor            $Hkey2,$T2
 520
 521         lea             32($inp),$inp           # i+=2
 522         sub             \$0x20,$len
 523         jbe             .Leven_tail
 524
 525 .Lmod_loop:
 526 ___
 527         &clmul64x64_T2  ($Xhi,$Xi,$Hkey2,1);    # H^2*(Ii+Xi)
 528 $code.=<<___;
 529         movdqu          ($inp),$T1              # Ii
 530         pxor            $Xn,$Xi                 # (H*Ii+1) + H^2*(Ii+Xi)
 531         pxor            $Xhn,$Xhi
 532
 533         movdqu          16($inp),$Xn            # Ii+1
 534         pshufb          $T3,$T1
 535         pshufb          $T3,$Xn
 536
 537         movdqa          $Xn,$Xhn                #
 538         pshufd          \$0b01001110,$Xn,$T1n
 539         pshufd          \$0b01001110,$Hkey,$T2n
 540         pxor            $Xn,$T1n                #
 541         pxor            $Hkey,$T2n
 542          pxor           $T1,$Xhi                # "Ii+Xi", consume early
 543
 544           movdqa        $Xi,$T1                 # 1st phase
 545           psllq         \$1,$Xi
 546           pxor          $T1,$Xi                 #
 547           psllq         \$5,$Xi                 #
 548           pxor          $T1,$Xi                 #
 549         pclmulqdq       \$0x00,$Hkey,$Xn        #######
 550           psllq         \$57,$Xi                #
 551           movdqa        $Xi,$T2                 #
 552           pslldq        \$8,$Xi
 553           psrldq        \$8,$T2                 #
 554           pxor          $T1,$Xi
 555           pxor          $T2,$Xhi                #
 556
 557         pclmulqdq       \$0x11,$Hkey,$Xhn       #######
 558           movdqa        $Xi,$T2                 # 2nd phase
 559           psrlq         \$5,$Xi
 560           pxor          $T2,$Xi                 #
 561           psrlq         \$1,$Xi                 #
 562           pxor          $T2,$Xi                 #
 563           pxor          $Xhi,$T2
 564           psrlq         \$1,$Xi                 #
 565           pxor          $T2,$Xi                 #
 566
 567         pclmulqdq       \$0x00,$T2n,$T1n        #######
 568          movdqa         $Xi,$Xhi                #
 569          pshufd         \$0b01001110,$Xi,$T1
 570          pshufd         \$0b01001110,$Hkey2,$T2
 571          pxor           $Xi,$T1                 #
 572          pxor           $Hkey2,$T2
 573
 574         pxor            $Xn,$T1n                #
 575         pxor            $Xhn,$T1n               #
 576         movdqa          $T1n,$T2n               #
 577         psrldq          \$8,$T1n
 578         pslldq          \$8,$T2n                #
 579         pxor            $T1n,$Xhn
 580         pxor            $T2n,$Xn                #
 581
 582         lea             32($inp),$inp
 583         sub             \$0x20,$len
 584         ja              .Lmod_loop
 585
 586 .Leven_tail:
 587 ___
 588         &clmul64x64_T2  ($Xhi,$Xi,$Hkey2,1);    # H^2*(Ii+Xi)
 589 $code.=<<___;
 590         pxor            $Xn,$Xi                 # (H*Ii+1) + H^2*(Ii+Xi)
 591         pxor            $Xhn,$Xhi
 592 ___
 593         &reduction_alg9 ($Xhi,$Xi);
 594 $code.=<<___;
 595         test            $len,$len
 596         jnz             .Ldone
 597
 598 .Lodd_tail:
 599         movdqu          ($inp),$T1              # Ii
 600         pshufb          $T3,$T1
 601         pxor            $T1,$Xi                 # Ii+Xi
 602 ___
 603         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);       # H*(Ii+Xi)
 604         &reduction_alg9 ($Xhi,$Xi);
 605 $code.=<<___;
 606 .Ldone:
 607         pshufb          $T3,$Xi
 608         movdqu          $Xi,($Xip)
 609 ___
 610 $code.=<<___ if ($win64);
 611         movaps  (%rsp),%xmm6
 612         movaps  0x10(%rsp),%xmm7
 613         movaps  0x20(%rsp),%xmm8
 614         movaps  0x30(%rsp),%xmm9
 615         movaps  0x40(%rsp),%xmm10
 616         add     \$0x58,%rsp
 617 ___
 618 $code.=<<___;
 619         ret
 620 .LSEH_end_gcm_ghash_clmul:
 621 .size   gcm_ghash_clmul,.-gcm_ghash_clmul
 622 ___
 623 }
 624
 625 $code.=<<___;
 626 .align  64
 627 .Lbswap_mask:
 628         .byte   15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
 629 .L0x1c2_polynomial:
 630         .byte   1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
 631 .align  64
 632 .type   .Lrem_4bit,\@object
 633 .Lrem_4bit:
 634         .long   0,`0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`
 635         .long   0,`0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`
 636         .long   0,`0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`
 637         .long   0,`0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`
 638 .type   .Lrem_8bit,\@object
 639 .Lrem_8bit:
 640         .value  0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E
 641         .value  0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E
 642         .value  0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E
 643         .value  0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E
 644         .value  0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E
 645         .value  0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E
 646         .value  0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E
 647         .value  0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E
 648         .value  0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE
 649         .value  0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE
 650         .value  0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE
 651         .value  0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE
 652         .value  0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E
 653         .value  0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E
 654         .value  0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE
 655         .value  0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE
 656         .value  0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E
 657         .value  0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E
 658         .value  0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E
 659         .value  0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E
 660         .value  0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E
 661         .value  0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E
 662         .value  0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E
 663         .value  0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E
 664         .value  0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE
 665         .value  0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE
 666         .value  0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE
 667         .value  0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE
 668         .value  0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E
 669         .value  0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E
 670         .value  0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE
 671         .value  0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE
 672
 673 .asciz  "GHASH for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 674 .align  64
 675 ___
 676 \f
 677 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
 678 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
 679 if ($win64) {
 680 $rec="%rcx";
 681 $frame="%rdx";
 682 $context="%r8";
 683 $disp="%r9";
 684
 685 $code.=<<___;
 686 .extern __imp_RtlVirtualUnwind
 687 .type   se_handler,\@abi-omnipotent
 688 .align  16
 689 se_handler:
 690         push    %rsi
 691         push    %rdi
 692         push    %rbx
 693         push    %rbp
 694         push    %r12
 695         push    %r13
 696         push    %r14
 697         push    %r15
 698         pushfq
 699         sub     \$64,%rsp
 700
 701         mov     120($context),%rax      # pull context->Rax
 702         mov     248($context),%rbx      # pull context->Rip
 703
 704         mov     8($disp),%rsi           # disp->ImageBase
 705         mov     56($disp),%r11          # disp->HandlerData
 706
 707         mov     0(%r11),%r10d           # HandlerData[0]
 708         lea     (%rsi,%r10),%r10        # prologue label
 709         cmp     %r10,%rbx               # context->Rip<prologue label
 710         jb      .Lin_prologue
 711
 712         mov     152($context),%rax      # pull context->Rsp
 713
 714         mov     4(%r11),%r10d           # HandlerData[1]
 715         lea     (%rsi,%r10),%r10        # epilogue label
 716         cmp     %r10,%rbx               # context->Rip>=epilogue label
 717         jae     .Lin_prologue
 718
 719         lea     24(%rax),%rax           # adjust "rsp"
 720
 721         mov     -8(%rax),%rbx
 722         mov     -16(%rax),%rbp
 723         mov     -24(%rax),%r12
 724         mov     %rbx,144($context)      # restore context->Rbx
 725         mov     %rbp,160($context)      # restore context->Rbp
 726         mov     %r12,216($context)      # restore context->R12
 727
 728 .Lin_prologue:
 729         mov     8(%rax),%rdi
 730         mov     16(%rax),%rsi
 731         mov     %rax,152($context)      # restore context->Rsp
 732         mov     %rsi,168($context)      # restore context->Rsi
 733         mov     %rdi,176($context)      # restore context->Rdi
 734
 735         mov     40($disp),%rdi          # disp->ContextRecord
 736         mov     $context,%rsi           # context
 737         mov     \$`1232/8`,%ecx         # sizeof(CONTEXT)
 738         .long   0xa548f3fc              # cld; rep movsq
 739
 740         mov     $disp,%rsi
 741         xor     %rcx,%rcx               # arg1, UNW_FLAG_NHANDLER
 742         mov     8(%rsi),%rdx            # arg2, disp->ImageBase
 743         mov     0(%rsi),%r8             # arg3, disp->ControlPc
 744         mov     16(%rsi),%r9            # arg4, disp->FunctionEntry
 745         mov     40(%rsi),%r10           # disp->ContextRecord
 746         lea     56(%rsi),%r11           # &disp->HandlerData
 747         lea     24(%rsi),%r12           # &disp->EstablisherFrame
 748         mov     %r10,32(%rsp)           # arg5
 749         mov     %r11,40(%rsp)           # arg6
 750         mov     %r12,48(%rsp)           # arg7
 751         mov     %rcx,56(%rsp)           # arg8, (NULL)
 752         call    *__imp_RtlVirtualUnwind(%rip)
 753
 754         mov     \$1,%eax                # ExceptionContinueSearch
 755         add     \$64,%rsp
 756         popfq
 757         pop     %r15
 758         pop     %r14
 759         pop     %r13
 760         pop     %r12
 761         pop     %rbp
 762         pop     %rbx
 763         pop     %rdi
 764         pop     %rsi
 765         ret
 766 .size   se_handler,.-se_handler
 767
 768 .section        .pdata
 769 .align  4
 770         .rva    .LSEH_begin_gcm_gmult_4bit
 771         .rva    .LSEH_end_gcm_gmult_4bit
 772         .rva    .LSEH_info_gcm_gmult_4bit
 773
 774         .rva    .LSEH_begin_gcm_ghash_4bit
 775         .rva    .LSEH_end_gcm_ghash_4bit
 776         .rva    .LSEH_info_gcm_ghash_4bit
 777
 778         .rva    .LSEH_begin_gcm_ghash_clmul
 779         .rva    .LSEH_end_gcm_ghash_clmul
 780         .rva    .LSEH_info_gcm_ghash_clmul
 781
 782 .section        .xdata
 783 .align  8
 784 .LSEH_info_gcm_gmult_4bit:
 785         .byte   9,0,0,0
 786         .rva    se_handler
 787         .rva    .Lgmult_prologue,.Lgmult_epilogue       # HandlerData
 788 .LSEH_info_gcm_ghash_4bit:
 789         .byte   9,0,0,0
 790         .rva    se_handler
 791         .rva    .Lghash_prologue,.Lghash_epilogue       # HandlerData
 792 .LSEH_info_gcm_ghash_clmul:
 793         .byte   0x01,0x1f,0x0b,0x00
 794         .byte   0x1f,0xa8,0x04,0x00     #movaps 0x40(rsp),xmm10
 795         .byte   0x19,0x98,0x03,0x00     #movaps 0x30(rsp),xmm9
 796         .byte   0x13,0x88,0x02,0x00     #movaps 0x20(rsp),xmm8
 797         .byte   0x0d,0x78,0x01,0x00     #movaps 0x10(rsp),xmm7
 798         .byte   0x08,0x68,0x00,0x00     #movaps (rsp),xmm6
 799         .byte   0x04,0xa2,0x00,0x00     #sub    rsp,0x58
 800 ___
 801 }
 802 \f
 803 $code =~ s/\`([^\`]*)\`/eval($1)/gem;
 804
 805 print $code;
 806
 807 close STDOUT;