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