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 # December 2012
  45 #
  46 # Overhaul: aggregate Karatsuba post-processing, improve ILP in
  47 # reduction_alg9, increase reduction aggregate factor to 4x. As for
  48 # the latter. ghash-x86.pl discusses that it makes lesser sense to
  49 # increase aggregate factor. Then why increase here? Critical path
  50 # consists of 3 independent pclmulqdq instructions, Karatsuba post-
  51 # processing and reduction. "On top" of this we lay down aggregated
  52 # multiplication operations, triplets of independent pclmulqdq's. As
  53 # issue rate for pclmulqdq is limited, it makes lesser sense to
  54 # aggregate more multiplications than it takes to perform remaining
  55 # non-multiplication operations. 2x is near-optimal coefficient for
  56 # contemporary Intel CPUs (therefore modest improvement coefficient),
  57 # but not for Bulldozer. Latter is because logical SIMD operations
  58 # are twice as slow in comparison to Intel, so that critical path is
  59 # longer. A CPU with higher pclmulqdq issue rate would also benefit
  60 # from higher aggregate factor...
  61 #
  62 # Westmere      1.76(+14%)
  63 # Sandy Bridge  1.79(+9%)
  64 # Ivy Bridge    1.79(+8%)
  65 # Bulldozer     1.52(+25%)
  66
  67 $flavour = shift;
  68 $output  = shift;
  69 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  70
  71 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
  72
  73 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  74 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  75 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
  76 die "can't locate x86_64-xlate.pl";
  77
  78 open OUT,"| \"$^X\" $xlate $flavour $output";
  79 *STDOUT=*OUT;
  80
  81 $do4xaggr=1;
  82
  83 # common register layout
  84 $nlo="%rax";
  85 $nhi="%rbx";
  86 $Zlo="%r8";
  87 $Zhi="%r9";
  88 $tmp="%r10";
  89 $rem_4bit = "%r11";
  90
  91 $Xi="%rdi";
  92 $Htbl="%rsi";
  93
  94 # per-function register layout
  95 $cnt="%rcx";
  96 $rem="%rdx";
  97
  98 sub LB() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/     or
  99                         $r =~ s/%[er]([sd]i)/%\1l/      or
 100                         $r =~ s/%[er](bp)/%\1l/         or
 101                         $r =~ s/%(r[0-9]+)[d]?/%\1b/;   $r; }
 102
 103 sub AUTOLOAD()          # thunk [simplified] 32-bit style perlasm
 104 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
 105   my $arg = pop;
 106     $arg = "\$$arg" if ($arg*1 eq $arg);
 107     $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
 108 }
 109 \f
 110 { my $N;
 111   sub loop() {
 112   my $inp = shift;
 113
 114         $N++;
 115 $code.=<<___;
 116         xor     $nlo,$nlo
 117         xor     $nhi,$nhi
 118         mov     `&LB("$Zlo")`,`&LB("$nlo")`
 119         mov     `&LB("$Zlo")`,`&LB("$nhi")`
 120         shl     \$4,`&LB("$nlo")`
 121         mov     \$14,$cnt
 122         mov     8($Htbl,$nlo),$Zlo
 123         mov     ($Htbl,$nlo),$Zhi
 124         and     \$0xf0,`&LB("$nhi")`
 125         mov     $Zlo,$rem
 126         jmp     .Loop$N
 127
 128 .align  16
 129 .Loop$N:
 130         shr     \$4,$Zlo
 131         and     \$0xf,$rem
 132         mov     $Zhi,$tmp
 133         mov     ($inp,$cnt),`&LB("$nlo")`
 134         shr     \$4,$Zhi
 135         xor     8($Htbl,$nhi),$Zlo
 136         shl     \$60,$tmp
 137         xor     ($Htbl,$nhi),$Zhi
 138         mov     `&LB("$nlo")`,`&LB("$nhi")`
 139         xor     ($rem_4bit,$rem,8),$Zhi
 140         mov     $Zlo,$rem
 141         shl     \$4,`&LB("$nlo")`
 142         xor     $tmp,$Zlo
 143         dec     $cnt
 144         js      .Lbreak$N
 145
 146         shr     \$4,$Zlo
 147         and     \$0xf,$rem
 148         mov     $Zhi,$tmp
 149         shr     \$4,$Zhi
 150         xor     8($Htbl,$nlo),$Zlo
 151         shl     \$60,$tmp
 152         xor     ($Htbl,$nlo),$Zhi
 153         and     \$0xf0,`&LB("$nhi")`
 154         xor     ($rem_4bit,$rem,8),$Zhi
 155         mov     $Zlo,$rem
 156         xor     $tmp,$Zlo
 157         jmp     .Loop$N
 158
 159 .align  16
 160 .Lbreak$N:
 161         shr     \$4,$Zlo
 162         and     \$0xf,$rem
 163         mov     $Zhi,$tmp
 164         shr     \$4,$Zhi
 165         xor     8($Htbl,$nlo),$Zlo
 166         shl     \$60,$tmp
 167         xor     ($Htbl,$nlo),$Zhi
 168         and     \$0xf0,`&LB("$nhi")`
 169         xor     ($rem_4bit,$rem,8),$Zhi
 170         mov     $Zlo,$rem
 171         xor     $tmp,$Zlo
 172
 173         shr     \$4,$Zlo
 174         and     \$0xf,$rem
 175         mov     $Zhi,$tmp
 176         shr     \$4,$Zhi
 177         xor     8($Htbl,$nhi),$Zlo
 178         shl     \$60,$tmp
 179         xor     ($Htbl,$nhi),$Zhi
 180         xor     $tmp,$Zlo
 181         xor     ($rem_4bit,$rem,8),$Zhi
 182
 183         bswap   $Zlo
 184         bswap   $Zhi
 185 ___
 186 }}
 187
 188 $code=<<___;
 189 .text
 190
 191 .globl  gcm_gmult_4bit
 192 .type   gcm_gmult_4bit,\@function,2
 193 .align  16
 194 gcm_gmult_4bit:
 195         push    %rbx
 196         push    %rbp            # %rbp and %r12 are pushed exclusively in
 197         push    %r12            # order to reuse Win64 exception handler...
 198 .Lgmult_prologue:
 199
 200         movzb   15($Xi),$Zlo
 201         lea     .Lrem_4bit(%rip),$rem_4bit
 202 ___
 203         &loop   ($Xi);
 204 $code.=<<___;
 205         mov     $Zlo,8($Xi)
 206         mov     $Zhi,($Xi)
 207
 208         mov     16(%rsp),%rbx
 209         lea     24(%rsp),%rsp
 210 .Lgmult_epilogue:
 211         ret
 212 .size   gcm_gmult_4bit,.-gcm_gmult_4bit
 213 ___
 214 \f
 215 # per-function register layout
 216 $inp="%rdx";
 217 $len="%rcx";
 218 $rem_8bit=$rem_4bit;
 219
 220 $code.=<<___;
 221 .globl  gcm_ghash_4bit
 222 .type   gcm_ghash_4bit,\@function,4
 223 .align  16
 224 gcm_ghash_4bit:
 225         push    %rbx
 226         push    %rbp
 227         push    %r12
 228         push    %r13
 229         push    %r14
 230         push    %r15
 231         sub     \$280,%rsp
 232 .Lghash_prologue:
 233         mov     $inp,%r14               # reassign couple of args
 234         mov     $len,%r15
 235 ___
 236 { my $inp="%r14";
 237   my $dat="%edx";
 238   my $len="%r15";
 239   my @nhi=("%ebx","%ecx");
 240   my @rem=("%r12","%r13");
 241   my $Hshr4="%rbp";
 242
 243         &sub    ($Htbl,-128);           # size optimization
 244         &lea    ($Hshr4,"16+128(%rsp)");
 245         { my @lo =($nlo,$nhi);
 246           my @hi =($Zlo,$Zhi);
 247
 248           &xor  ($dat,$dat);
 249           for ($i=0,$j=-2;$i<18;$i++,$j++) {
 250             &mov        ("$j(%rsp)",&LB($dat))          if ($i>1);
 251             &or         ($lo[0],$tmp)                   if ($i>1);
 252             &mov        (&LB($dat),&LB($lo[1]))         if ($i>0 && $i<17);
 253             &shr        ($lo[1],4)                      if ($i>0 && $i<17);
 254             &mov        ($tmp,$hi[1])                   if ($i>0 && $i<17);
 255             &shr        ($hi[1],4)                      if ($i>0 && $i<17);
 256             &mov        ("8*$j($Hshr4)",$hi[0])         if ($i>1);
 257             &mov        ($hi[0],"16*$i+0-128($Htbl)")   if ($i<16);
 258             &shl        (&LB($dat),4)                   if ($i>0 && $i<17);
 259             &mov        ("8*$j-128($Hshr4)",$lo[0])     if ($i>1);
 260             &mov        ($lo[0],"16*$i+8-128($Htbl)")   if ($i<16);
 261             &shl        ($tmp,60)                       if ($i>0 && $i<17);
 262
 263             push        (@lo,shift(@lo));
 264             push        (@hi,shift(@hi));
 265           }
 266         }
 267         &add    ($Htbl,-128);
 268         &mov    ($Zlo,"8($Xi)");
 269         &mov    ($Zhi,"0($Xi)");
 270         &add    ($len,$inp);            # pointer to the end of data
 271         &lea    ($rem_8bit,".Lrem_8bit(%rip)");
 272         &jmp    (".Louter_loop");
 273
 274 $code.=".align  16\n.Louter_loop:\n";
 275         &xor    ($Zhi,"($inp)");
 276         &mov    ("%rdx","8($inp)");
 277         &lea    ($inp,"16($inp)");
 278         &xor    ("%rdx",$Zlo);
 279         &mov    ("($Xi)",$Zhi);
 280         &mov    ("8($Xi)","%rdx");
 281         &shr    ("%rdx",32);
 282
 283         &xor    ($nlo,$nlo);
 284         &rol    ($dat,8);
 285         &mov    (&LB($nlo),&LB($dat));
 286         &movz   ($nhi[0],&LB($dat));
 287         &shl    (&LB($nlo),4);
 288         &shr    ($nhi[0],4);
 289
 290         for ($j=11,$i=0;$i<15;$i++) {
 291             &rol        ($dat,8);
 292             &xor        ($Zlo,"8($Htbl,$nlo)")                  if ($i>0);
 293             &xor        ($Zhi,"($Htbl,$nlo)")                   if ($i>0);
 294             &mov        ($Zlo,"8($Htbl,$nlo)")                  if ($i==0);
 295             &mov        ($Zhi,"($Htbl,$nlo)")                   if ($i==0);
 296
 297             &mov        (&LB($nlo),&LB($dat));
 298             &xor        ($Zlo,$tmp)                             if ($i>0);
 299             &movzw      ($rem[1],"($rem_8bit,$rem[1],2)")       if ($i>0);
 300
 301             &movz       ($nhi[1],&LB($dat));
 302             &shl        (&LB($nlo),4);
 303             &movzb      ($rem[0],"(%rsp,$nhi[0])");
 304
 305             &shr        ($nhi[1],4)                             if ($i<14);
 306             &and        ($nhi[1],0xf0)                          if ($i==14);
 307             &shl        ($rem[1],48)                            if ($i>0);
 308             &xor        ($rem[0],$Zlo);
 309
 310             &mov        ($tmp,$Zhi);
 311             &xor        ($Zhi,$rem[1])                          if ($i>0);
 312             &shr        ($Zlo,8);
 313
 314             &movz       ($rem[0],&LB($rem[0]));
 315             &mov        ($dat,"$j($Xi)")                        if (--$j%4==0);
 316             &shr        ($Zhi,8);
 317
 318             &xor        ($Zlo,"-128($Hshr4,$nhi[0],8)");
 319             &shl        ($tmp,56);
 320             &xor        ($Zhi,"($Hshr4,$nhi[0],8)");
 321
 322             unshift     (@nhi,pop(@nhi));               # "rotate" registers
 323             unshift     (@rem,pop(@rem));
 324         }
 325         &movzw  ($rem[1],"($rem_8bit,$rem[1],2)");
 326         &xor    ($Zlo,"8($Htbl,$nlo)");
 327         &xor    ($Zhi,"($Htbl,$nlo)");
 328
 329         &shl    ($rem[1],48);
 330         &xor    ($Zlo,$tmp);
 331
 332         &xor    ($Zhi,$rem[1]);
 333         &movz   ($rem[0],&LB($Zlo));
 334         &shr    ($Zlo,4);
 335
 336         &mov    ($tmp,$Zhi);
 337         &shl    (&LB($rem[0]),4);
 338         &shr    ($Zhi,4);
 339
 340         &xor    ($Zlo,"8($Htbl,$nhi[0])");
 341         &movzw  ($rem[0],"($rem_8bit,$rem[0],2)");
 342         &shl    ($tmp,60);
 343
 344         &xor    ($Zhi,"($Htbl,$nhi[0])");
 345         &xor    ($Zlo,$tmp);
 346         &shl    ($rem[0],48);
 347
 348         &bswap  ($Zlo);
 349         &xor    ($Zhi,$rem[0]);
 350
 351         &bswap  ($Zhi);
 352         &cmp    ($inp,$len);
 353         &jb     (".Louter_loop");
 354 }
 355 $code.=<<___;
 356         mov     $Zlo,8($Xi)
 357         mov     $Zhi,($Xi)
 358
 359         lea     280(%rsp),%rsi
 360         mov     0(%rsi),%r15
 361         mov     8(%rsi),%r14
 362         mov     16(%rsi),%r13
 363         mov     24(%rsi),%r12
 364         mov     32(%rsi),%rbp
 365         mov     40(%rsi),%rbx
 366         lea     48(%rsi),%rsp
 367 .Lghash_epilogue:
 368         ret
 369 .size   gcm_ghash_4bit,.-gcm_ghash_4bit
 370 ___
 371 \f
 372 ######################################################################
 373 # PCLMULQDQ version.
 374
 375 @_4args=$win64? ("%rcx","%rdx","%r8", "%r9") :  # Win64 order
 376                 ("%rdi","%rsi","%rdx","%rcx");  # Unix order
 377
 378 ($Xi,$Xhi)=("%xmm0","%xmm1");   $Hkey="%xmm2";
 379 ($T1,$T2,$T3)=("%xmm3","%xmm4","%xmm5");
 380
 381 sub clmul64x64_T2 {     # minimal register pressure
 382 my ($Xhi,$Xi,$Hkey,$HK)=@_;
 383
 384 if (!defined($HK)) {    $HK = $T2;
 385 $code.=<<___;
 386         movdqa          $Xi,$Xhi                #
 387         pshufd          \$0b01001110,$Xi,$T1
 388         pshufd          \$0b01001110,$Hkey,$T2
 389         pxor            $Xi,$T1                 #
 390         pxor            $Hkey,$T2
 391 ___
 392 } else {
 393 $code.=<<___;
 394         movdqa          $Xi,$Xhi                #
 395         pshufd          \$0b01001110,$Xi,$T1
 396         pxor            $Xi,$T1                 #
 397 ___
 398 }
 399 $code.=<<___;
 400         pclmulqdq       \$0x00,$Hkey,$Xi        #######
 401         pclmulqdq       \$0x11,$Hkey,$Xhi       #######
 402         pclmulqdq       \$0x00,$HK,$T1          #######
 403         pxor            $Xi,$T1                 #
 404         pxor            $Xhi,$T1                #
 405
 406         movdqa          $T1,$T2                 #
 407         psrldq          \$8,$T1
 408         pslldq          \$8,$T2                 #
 409         pxor            $T1,$Xhi
 410         pxor            $T2,$Xi                 #
 411 ___
 412 }
 413
 414 sub reduction_alg9 {    # 17/11 times faster than Intel version
 415 my ($Xhi,$Xi) = @_;
 416
 417 $code.=<<___;
 418         # 1st phase
 419         movdqa          $Xi,$T2                 #
 420         movdqa          $Xi,$T1
 421         psllq           \$5,$Xi
 422         pxor            $Xi,$T1                 #
 423         psllq           \$1,$Xi
 424         pxor            $T1,$Xi                 #
 425         psllq           \$57,$Xi                #
 426         movdqa          $Xi,$T1                 #
 427         pslldq          \$8,$Xi
 428         psrldq          \$8,$T1                 #
 429         pxor            $T2,$Xi
 430         pxor            $T1,$Xhi                #
 431
 432         # 2nd phase
 433         movdqa          $Xi,$T2
 434         psrlq           \$1,$Xi
 435         pxor            $T2,$Xhi                #
 436         pxor            $Xi,$T2
 437         psrlq           \$5,$Xi
 438         pxor            $T2,$Xi                 #
 439         psrlq           \$1,$Xi                 #
 440         pxor            $Xhi,$Xi                #
 441 ___
 442 }
 443 \f
 444 { my ($Htbl,$Xip)=@_4args;
 445
 446 $code.=<<___;
 447 .globl  gcm_init_clmul
 448 .type   gcm_init_clmul,\@abi-omnipotent
 449 .align  16
 450 gcm_init_clmul:
 451         movdqu          ($Xip),$Hkey
 452         pshufd          \$0b01001110,$Hkey,$Hkey        # dword swap
 453
 454         # <<1 twist
 455         pshufd          \$0b11111111,$Hkey,$T2  # broadcast uppermost dword
 456         movdqa          $Hkey,$T1
 457         psllq           \$1,$Hkey
 458         pxor            $T3,$T3                 #
 459         psrlq           \$63,$T1
 460         pcmpgtd         $T2,$T3                 # broadcast carry bit
 461         pslldq          \$8,$T1
 462         por             $T1,$Hkey               # H<<=1
 463
 464         # magic reduction
 465         pand            .L0x1c2_polynomial(%rip),$T3
 466         pxor            $T3,$Hkey               # if(carry) H^=0x1c2_polynomial
 467
 468         # calculate H^2
 469         movdqa          $Hkey,$Xi
 470 ___
 471         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);
 472         &reduction_alg9 ($Xhi,$Xi);
 473 $code.=<<___;
 474         pshufd          \$0b01001110,$Hkey,$T1
 475         pshufd          \$0b01001110,$Xi,$T2
 476         pxor            $Hkey,$T1               # Karatsuba pre-processing
 477         movdqu          $Hkey,0x00($Htbl)       # save H
 478         pxor            $Xi,$T2                 # Karatsuba pre-processing
 479         movdqu          $Xi,0x10($Htbl)         # save H^2
 480         palignr         \$8,$T1,$T2             # low part is H.lo^H.hi...
 481         movdqu          $T2,0x20($Htbl)         # save Karatsuba "salt"
 482 ___
 483 if ($do4xaggr) {
 484         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);       # H^3
 485         &reduction_alg9 ($Xhi,$Xi);
 486 $code.=<<___;
 487         movdqa          $Xi,$T3
 488 ___
 489         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);       # H^4
 490         &reduction_alg9 ($Xhi,$Xi);
 491 $code.=<<___;
 492         pshufd          \$0b01001110,$T3,$T1
 493         pshufd          \$0b01001110,$Xi,$T2
 494         pxor            $T3,$T1                 # Karatsuba pre-processing
 495         movdqu          $T3,0x30($Htbl)         # save H^3
 496         pxor            $Xi,$T2                 # Karatsuba pre-processing
 497         movdqu          $Xi,0x40($Htbl)         # save H^4
 498         palignr         \$8,$T1,$T2             # low part is H.lo^H.hi...
 499         movdqu          $T2,0x50($Htbl)         # save Karatsuba "salt"
 500 ___
 501 }
 502 $code.=<<___;
 503         ret
 504 .size   gcm_init_clmul,.-gcm_init_clmul
 505 ___
 506 }
 507
 508 { my ($Xip,$Htbl)=@_4args;
 509
 510 $code.=<<___;
 511 .globl  gcm_gmult_clmul
 512 .type   gcm_gmult_clmul,\@abi-omnipotent
 513 .align  16
 514 gcm_gmult_clmul:
 515         movdqu          ($Xip),$Xi
 516         movdqa          .Lbswap_mask(%rip),$T3
 517         movdqu          ($Htbl),$Hkey
 518         movdqu          0x20($Htbl),$T2
 519         pshufb          $T3,$Xi
 520 ___
 521         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$T2);
 522 $code.=<<___ if (0 || (&reduction_alg9($Xhi,$Xi)&&0));
 523         # experimental alternative. special thing about is that there
 524         # no dependency between the two multiplications...
 525         mov             \$`0xE1<<1`,%eax
 526         mov             \$0xA040608020C0E000,%r10       # ((7..0)·0xE0)&0xff
 527         mov             \$0x07,%r11d
 528         movq            %rax,$T1
 529         movq            %r10,$T2
 530         movq            %r11,$T3                # borrow $T3
 531         pand            $Xi,$T3
 532         pshufb          $T3,$T2                 # ($Xi&7)·0xE0
 533         movq            %rax,$T3
 534         pclmulqdq       \$0x00,$Xi,$T1          # ·(0xE1<<1)
 535         pxor            $Xi,$T2
 536         pslldq          \$15,$T2
 537         paddd           $T2,$T2                 # <<(64+56+1)
 538         pxor            $T2,$Xi
 539         pclmulqdq       \$0x01,$T3,$Xi
 540         movdqa          .Lbswap_mask(%rip),$T3  # reload $T3
 541         psrldq          \$1,$T1
 542         pxor            $T1,$Xhi
 543         pslldq          \$7,$Xi
 544         pxor            $Xhi,$Xi
 545 ___
 546 $code.=<<___;
 547         pshufb          $T3,$Xi
 548         movdqu          $Xi,($Xip)
 549         ret
 550 .size   gcm_gmult_clmul,.-gcm_gmult_clmul
 551 ___
 552 }
 553 \f
 554 { my ($Xip,$Htbl,$inp,$len)=@_4args;
 555   my ($Xln,$Xmn,$Xhn,$Hkey2,$HK) = map("%xmm$_",(6..10));
 556
 557 $code.=<<___;
 558 .globl  gcm_ghash_clmul
 559 .type   gcm_ghash_clmul,\@abi-omnipotent
 560 .align  32
 561 gcm_ghash_clmul:
 562 ___
 563 $code.=<<___ if ($win64);
 564         lea     -0x88(%rsp),%rax
 565 .LSEH_begin_gcm_ghash_clmul:
 566         # I can't trust assembler to use specific encoding:-(
 567         .byte   0x48,0x8d,0x60,0xe0             #lea    -0x20(%rax),%rsp
 568         .byte   0x0f,0x29,0x70,0xe0             #movaps %xmm6,-0x20(%rax)
 569         .byte   0x0f,0x29,0x78,0xf0             #movaps %xmm7,-0x10(%rax)
 570         .byte   0x44,0x0f,0x29,0x00             #movaps %xmm8,0(%rax)
 571         .byte   0x44,0x0f,0x29,0x48,0x10        #movaps %xmm9,0x10(%rax)
 572         .byte   0x44,0x0f,0x29,0x50,0x20        #movaps %xmm10,0x20(%rax)
 573         .byte   0x44,0x0f,0x29,0x58,0x30        #movaps %xmm11,0x30(%rax)
 574         .byte   0x44,0x0f,0x29,0x60,0x40        #movaps %xmm12,0x40(%rax)
 575         .byte   0x44,0x0f,0x29,0x68,0x50        #movaps %xmm13,0x50(%rax)
 576         .byte   0x44,0x0f,0x29,0x70,0x60        #movaps %xmm14,0x60(%rax)
 577         .byte   0x44,0x0f,0x29,0x78,0x70        #movaps %xmm15,0x70(%rax)
 578 ___
 579 $code.=<<___;
 580         movdqa          .Lbswap_mask(%rip),$T3
 581         mov             \$0xA040608020C0E000,%rax       # ((7..0)·0xE0)&0xff
 582
 583         movdqu          ($Xip),$Xi
 584         movdqu          ($Htbl),$Hkey
 585         movdqu          0x20($Htbl),$HK
 586         pshufb          $T3,$Xi
 587
 588         sub             \$0x10,$len
 589         jz              .Lodd_tail
 590
 591         movdqu          0x10($Htbl),$Hkey2
 592 ___
 593 if ($do4xaggr) {
 594 my ($Xl,$Xm,$Xh,$Hkey3,$Hkey4)=map("%xmm$_",(11..15));
 595
 596 $code.=<<___;
 597         cmp             \$0x30,$len
 598         jb              .Lskip4x
 599
 600         sub             \$0x30,$len
 601         movdqu          0x30($Htbl),$Hkey3
 602         movdqu          0x40($Htbl),$Hkey4
 603
 604         #######
 605         # Xi+4 =[(H*Ii+3) + (H^2*Ii+2) + (H^3*Ii+1) + H^4*(Ii+Xi)] mod P
 606         #
 607         movdqu          0x30($inp),$Xln
 608          movdqu         0x20($inp),$Xl
 609         pshufb          $T3,$Xln
 610          pshufb         $T3,$Xl
 611         movdqa          $Xln,$Xhn
 612         pshufd          \$0b01001110,$Xln,$Xmn
 613         pxor            $Xln,$Xmn
 614         pclmulqdq       \$0x00,$Hkey,$Xln
 615         pclmulqdq       \$0x11,$Hkey,$Xhn
 616         pclmulqdq       \$0x00,$HK,$Xmn
 617
 618         movdqa          $Xl,$Xh
 619         pshufd          \$0b01001110,$Xl,$Xm
 620         pxor            $Xl,$Xm
 621         pclmulqdq       \$0x00,$Hkey2,$Xl
 622         pclmulqdq       \$0x11,$Hkey2,$Xh
 623         xorps           $Xl,$Xln
 624         pclmulqdq       \$0x10,$HK,$Xm
 625         xorps           $Xh,$Xhn
 626         movups          0x50($Htbl),$HK
 627         xorps           $Xm,$Xmn
 628
 629         movdqu          0x10($inp),$Xl
 630          movdqu         0($inp),$T1
 631         pshufb          $T3,$Xl
 632          pshufb         $T3,$T1
 633         movdqa          $Xl,$Xh
 634         pshufd          \$0b01001110,$Xl,$Xm
 635          pxor           $T1,$Xi
 636         pxor            $Xl,$Xm
 637         pclmulqdq       \$0x00,$Hkey3,$Xl
 638          movdqa         $Xi,$Xhi
 639          pshufd         \$0b01001110,$Xi,$T1
 640          pxor           $Xi,$T1
 641         pclmulqdq       \$0x11,$Hkey3,$Xh
 642         xorps           $Xl,$Xln
 643         pclmulqdq       \$0x00,$HK,$Xm
 644         xorps           $Xh,$Xhn
 645
 646         lea     0x40($inp),$inp
 647         sub     \$0x40,$len
 648         jc      .Ltail4x
 649
 650         jmp     .Lmod4_loop
 651 .align  32
 652 .Lmod4_loop:
 653         pclmulqdq       \$0x00,$Hkey4,$Xi
 654         xorps           $Xm,$Xmn
 655          movdqu         0x30($inp),$Xl
 656          pshufb         $T3,$Xl
 657         pclmulqdq       \$0x11,$Hkey4,$Xhi
 658         xorps           $Xln,$Xi
 659          movdqu         0x20($inp),$Xln
 660          movdqa         $Xl,$Xh
 661          pshufd         \$0b01001110,$Xl,$Xm
 662         pclmulqdq       \$0x10,$HK,$T1
 663         xorps           $Xhn,$Xhi
 664          pxor           $Xl,$Xm
 665          pshufb         $T3,$Xln
 666         movups          0x20($Htbl),$HK
 667          pclmulqdq      \$0x00,$Hkey,$Xl
 668         xorps           $Xmn,$T1
 669          movdqa         $Xln,$Xhn
 670          pshufd         \$0b01001110,$Xln,$Xmn
 671
 672         pxor            $Xi,$T1                 # aggregated Karatsuba post-processing
 673          pxor           $Xln,$Xmn
 674         pxor            $Xhi,$T1                #
 675         movdqa          $T1,$T2                 #
 676         pslldq          \$8,$T1
 677          pclmulqdq      \$0x11,$Hkey,$Xh
 678         psrldq          \$8,$T2                 #
 679         pxor            $T1,$Xi
 680         movdqa          .L7_mask(%rip),$T1
 681         pxor            $T2,$Xhi                #
 682         movq            %rax,$T2
 683
 684         pand            $Xi,$T1                 # 1st phase
 685         pshufb          $T1,$T2                 #
 686          pclmulqdq      \$0x00,$HK,$Xm
 687         pxor            $Xi,$T2                 #
 688         psllq           \$57,$T2                #
 689         movdqa          $T2,$T1                 #
 690         pslldq          \$8,$T2
 691          pclmulqdq      \$0x00,$Hkey2,$Xln
 692         psrldq          \$8,$T1                 #
 693         pxor            $T2,$Xi
 694         pxor            $T1,$Xhi                #
 695         movdqu          0($inp),$T1
 696
 697         movdqa          $Xi,$T2                 # 2nd phase
 698         psrlq           \$1,$Xi
 699          pclmulqdq      \$0x11,$Hkey2,$Xhn
 700          xorps          $Xl,$Xln
 701          movdqu         0x10($inp),$Xl
 702          pshufb         $T3,$Xl
 703          pclmulqdq      \$0x10,$HK,$Xmn
 704          xorps          $Xh,$Xhn
 705          movups         0x50($Htbl),$HK
 706         pshufb          $T3,$T1
 707         pxor            $T2,$Xhi                #
 708         pxor            $Xi,$T2
 709         psrlq           \$5,$Xi
 710
 711          movdqa         $Xl,$Xh
 712          pxor           $Xm,$Xmn
 713          pshufd         \$0b01001110,$Xl,$Xm
 714          pxor           $Xl,$Xm
 715          pclmulqdq      \$0x00,$Hkey3,$Xl
 716         pxor            $T2,$Xi                 #
 717         pxor            $T1,$Xhi
 718         psrlq           \$1,$Xi                 #
 719          pclmulqdq      \$0x11,$Hkey3,$Xh
 720          xorps          $Xl,$Xln
 721         pxor            $Xhi,$Xi                #
 722
 723          pclmulqdq      \$0x00,$HK,$Xm
 724          xorps          $Xh,$Xhn
 725
 726         movdqa          $Xi,$Xhi
 727         pshufd          \$0b01001110,$Xi,$T1
 728         pxor            $Xi,$T1
 729
 730         lea     0x40($inp),$inp
 731         sub     \$0x40,$len
 732         jnc     .Lmod4_loop
 733
 734 .Ltail4x:
 735         pclmulqdq       \$0x00,$Hkey4,$Xi
 736         xorps           $Xm,$Xmn
 737         pclmulqdq       \$0x11,$Hkey4,$Xhi
 738         xorps           $Xln,$Xi
 739         pclmulqdq       \$0x10,$HK,$T1
 740         xorps           $Xhn,$Xhi
 741         pxor            $Xi,$Xhi                # aggregated Karatsuba post-processing
 742         pxor            $Xmn,$T1
 743
 744         pxor            $Xhi,$T1                #
 745         pxor            $Xi,$Xhi
 746
 747         movdqa          $T1,$T2                 #
 748         psrldq          \$8,$T1
 749         pslldq          \$8,$T2                 #
 750         pxor            $T1,$Xhi
 751         pxor            $T2,$Xi                 #
 752 ___
 753         &reduction_alg9($Xhi,$Xi);
 754 $code.=<<___;
 755         add     \$0x40,$len
 756         jz      .Ldone
 757         movdqu  0x20($Htbl),$HK
 758         sub     \$0x10,$len
 759         jz      .Lodd_tail
 760 .Lskip4x:
 761 ___
 762 }
 763 $code.=<<___;
 764         #######
 765         # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
 766         #       [(H*Ii+1) + (H*Xi+1)] mod P =
 767         #       [(H*Ii+1) + H^2*(Ii+Xi)] mod P
 768         #
 769         movdqu          ($inp),$T1              # Ii
 770         movdqu          16($inp),$Xln           # Ii+1
 771         pshufb          $T3,$T1
 772         pshufb          $T3,$Xln
 773         pxor            $T1,$Xi                 # Ii+Xi
 774
 775         movdqa          $Xln,$Xhn
 776         pshufd          \$0b01001110,$Xln,$T1
 777         pxor            $Xln,$T1
 778         pclmulqdq       \$0x00,$Hkey,$Xln
 779         pclmulqdq       \$0x11,$Hkey,$Xhn
 780         pclmulqdq       \$0x00,$HK,$T1
 781
 782         lea             32($inp),$inp           # i+=2
 783         sub             \$0x20,$len
 784         jbe             .Leven_tail
 785         jmp             .Lmod_loop
 786
 787 .align  32
 788 .Lmod_loop:
 789         movdqa          $Xi,$Xhi
 790         pshufd          \$0b01001110,$Xi,$T2    #
 791         pxor            $Xi,$T2                 #
 792
 793         pclmulqdq       \$0x00,$Hkey2,$Xi
 794         pclmulqdq       \$0x11,$Hkey2,$Xhi
 795         pclmulqdq       \$0x10,$HK,$T2
 796
 797         pxor            $Xln,$Xi                # (H*Ii+1) + H^2*(Ii+Xi)
 798         pxor            $Xhn,$Xhi
 799           movdqu        ($inp),$Xhn             # Ii
 800           pshufb        $T3,$Xhn
 801           movdqu        16($inp),$Xln           # Ii+1
 802
 803         pxor            $Xi,$T1                 # aggregated Karatsuba post-processing
 804         pxor            $Xhi,$T1
 805           pxor          $Xhn,$Xhi               # "Ii+Xi", consume early
 806         pxor            $T1,$T2
 807          pshufb         $T3,$Xln
 808         movdqa          $T2,$T1                 #
 809         psrldq          \$8,$T1
 810         pslldq          \$8,$T2                 #
 811         pxor            $T1,$Xhi
 812         pxor            $T2,$Xi                 #
 813
 814         movdqa          $Xln,$Xhn               #
 815
 816           movdqa        $Xi,$T2                 # 1st phase
 817           movdqa        $Xi,$T1
 818           psllq         \$5,$Xi
 819         pclmulqdq       \$0x00,$Hkey,$Xln       #######
 820           pxor          $Xi,$T1                 #
 821           psllq         \$1,$Xi
 822           pxor          $T1,$Xi                 #
 823           psllq         \$57,$Xi                #
 824           movdqa        $Xi,$T1                 #
 825           pslldq        \$8,$Xi
 826           psrldq        \$8,$T1                 #
 827           pxor          $T2,$Xi
 828           pxor          $T1,$Xhi                #
 829         pshufd          \$0b01001110,$Xhn,$T1
 830         pxor            $Xhn,$T1                #
 831
 832         pclmulqdq       \$0x11,$Hkey,$Xhn       #######
 833           movdqa        $Xi,$T2                 # 2nd phase
 834           psrlq         \$1,$Xi
 835           pxor          $T2,$Xhi                #
 836           pxor          $Xi,$T2
 837           psrlq         \$5,$Xi
 838           pxor          $T2,$Xi                 #
 839           psrlq         \$1,$Xi                 #
 840         pclmulqdq       \$0x00,$HK,$T1          #######
 841           pxor          $Xhi,$Xi                #
 842
 843         lea             32($inp),$inp
 844         sub             \$0x20,$len
 845         ja              .Lmod_loop
 846
 847 .Leven_tail:
 848          movdqa         $Xi,$Xhi
 849          pshufd         \$0b01001110,$Xi,$T2    #
 850          pxor           $Xi,$T2                 #
 851
 852         pclmulqdq       \$0x00,$Hkey2,$Xi
 853         pclmulqdq       \$0x11,$Hkey2,$Xhi
 854         pclmulqdq       \$0x10,$HK,$T2
 855
 856         pxor            $Xln,$Xi                # (H*Ii+1) + H^2*(Ii+Xi)
 857         pxor            $Xhn,$Xhi
 858         pxor            $Xi,$T1
 859         pxor            $Xhi,$T1
 860         pxor            $T1,$T2
 861         movdqa          $T2,$T1                 #
 862         psrldq          \$8,$T1
 863         pslldq          \$8,$T2                 #
 864         pxor            $T1,$Xhi
 865         pxor            $T2,$Xi                 #
 866 ___
 867         &reduction_alg9 ($Xhi,$Xi);
 868 $code.=<<___;
 869         test            $len,$len
 870         jnz             .Ldone
 871
 872 .Lodd_tail:
 873         movdqu          ($inp),$T1              # Ii
 874         pshufb          $T3,$T1
 875         pxor            $T1,$Xi                 # Ii+Xi
 876 ___
 877         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);   # H*(Ii+Xi)
 878         &reduction_alg9 ($Xhi,$Xi);
 879 $code.=<<___;
 880 .Ldone:
 881         pshufb          $T3,$Xi
 882         movdqu          $Xi,($Xip)
 883 ___
 884 $code.=<<___ if ($win64);
 885         movaps  (%rsp),%xmm6
 886         movaps  0x10(%rsp),%xmm7
 887         movaps  0x20(%rsp),%xmm8
 888         movaps  0x30(%rsp),%xmm9
 889         movaps  0x40(%rsp),%xmm10
 890         movaps  0x50(%rsp),%xmm11
 891         movaps  0x60(%rsp),%xmm12
 892         movaps  0x70(%rsp),%xmm13
 893         movaps  0x80(%rsp),%xmm14
 894         movaps  0x90(%rsp),%xmm15
 895         lea     0xa8(%rsp),%rsp
 896 ___
 897 $code.=<<___;
 898         ret
 899 .LSEH_end_gcm_ghash_clmul:
 900 .size   gcm_ghash_clmul,.-gcm_ghash_clmul
 901 ___
 902 }
 903
 904 $code.=<<___;
 905 .align  64
 906 .Lbswap_mask:
 907         .byte   15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
 908 .L0x1c2_polynomial:
 909         .byte   1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
 910 .L7_mask:
 911         .long   7,0,7,0
 912 .L7_mask_poly:
 913         .long   7,0,`0xE1<<1`,0
 914 .align  64
 915 .type   .Lrem_4bit,\@object
 916 .Lrem_4bit:
 917         .long   0,`0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`
 918         .long   0,`0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`
 919         .long   0,`0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`
 920         .long   0,`0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`
 921 .type   .Lrem_8bit,\@object
 922 .Lrem_8bit:
 923         .value  0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E
 924         .value  0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E
 925         .value  0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E
 926         .value  0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E
 927         .value  0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E
 928         .value  0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E
 929         .value  0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E
 930         .value  0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E
 931         .value  0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE
 932         .value  0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE
 933         .value  0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE
 934         .value  0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE
 935         .value  0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E
 936         .value  0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E
 937         .value  0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE
 938         .value  0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE
 939         .value  0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E
 940         .value  0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E
 941         .value  0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E
 942         .value  0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E
 943         .value  0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E
 944         .value  0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E
 945         .value  0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E
 946         .value  0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E
 947         .value  0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE
 948         .value  0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE
 949         .value  0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE
 950         .value  0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE
 951         .value  0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E
 952         .value  0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E
 953         .value  0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE
 954         .value  0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE
 955
 956 .asciz  "GHASH for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 957 .align  64
 958 ___
 959 \f
 960 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
 961 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
 962 if ($win64) {
 963 $rec="%rcx";
 964 $frame="%rdx";
 965 $context="%r8";
 966 $disp="%r9";
 967
 968 $code.=<<___;
 969 .extern __imp_RtlVirtualUnwind
 970 .type   se_handler,\@abi-omnipotent
 971 .align  16
 972 se_handler:
 973         push    %rsi
 974         push    %rdi
 975         push    %rbx
 976         push    %rbp
 977         push    %r12
 978         push    %r13
 979         push    %r14
 980         push    %r15
 981         pushfq
 982         sub     \$64,%rsp
 983
 984         mov     120($context),%rax      # pull context->Rax
 985         mov     248($context),%rbx      # pull context->Rip
 986
 987         mov     8($disp),%rsi           # disp->ImageBase
 988         mov     56($disp),%r11          # disp->HandlerData
 989
 990         mov     0(%r11),%r10d           # HandlerData[0]
 991         lea     (%rsi,%r10),%r10        # prologue label
 992         cmp     %r10,%rbx               # context->Rip<prologue label
 993         jb      .Lin_prologue
 994
 995         mov     152($context),%rax      # pull context->Rsp
 996
 997         mov     4(%r11),%r10d           # HandlerData[1]
 998         lea     (%rsi,%r10),%r10        # epilogue label
 999         cmp     %r10,%rbx               # context->Rip>=epilogue label
1000         jae     .Lin_prologue
1001
1002         lea     24(%rax),%rax           # adjust "rsp"
1003
1004         mov     -8(%rax),%rbx
1005         mov     -16(%rax),%rbp
1006         mov     -24(%rax),%r12
1007         mov     %rbx,144($context)      # restore context->Rbx
1008         mov     %rbp,160($context)      # restore context->Rbp
1009         mov     %r12,216($context)      # restore context->R12
1010
1011 .Lin_prologue:
1012         mov     8(%rax),%rdi
1013         mov     16(%rax),%rsi
1014         mov     %rax,152($context)      # restore context->Rsp
1015         mov     %rsi,168($context)      # restore context->Rsi
1016         mov     %rdi,176($context)      # restore context->Rdi
1017
1018         mov     40($disp),%rdi          # disp->ContextRecord
1019         mov     $context,%rsi           # context
1020         mov     \$`1232/8`,%ecx         # sizeof(CONTEXT)
1021         .long   0xa548f3fc              # cld; rep movsq
1022
1023         mov     $disp,%rsi
1024         xor     %rcx,%rcx               # arg1, UNW_FLAG_NHANDLER
1025         mov     8(%rsi),%rdx            # arg2, disp->ImageBase
1026         mov     0(%rsi),%r8             # arg3, disp->ControlPc
1027         mov     16(%rsi),%r9            # arg4, disp->FunctionEntry
1028         mov     40(%rsi),%r10           # disp->ContextRecord
1029         lea     56(%rsi),%r11           # &disp->HandlerData
1030         lea     24(%rsi),%r12           # &disp->EstablisherFrame
1031         mov     %r10,32(%rsp)           # arg5
1032         mov     %r11,40(%rsp)           # arg6
1033         mov     %r12,48(%rsp)           # arg7
1034         mov     %rcx,56(%rsp)           # arg8, (NULL)
1035         call    *__imp_RtlVirtualUnwind(%rip)
1036
1037         mov     \$1,%eax                # ExceptionContinueSearch
1038         add     \$64,%rsp
1039         popfq
1040         pop     %r15
1041         pop     %r14
1042         pop     %r13
1043         pop     %r12
1044         pop     %rbp
1045         pop     %rbx
1046         pop     %rdi
1047         pop     %rsi
1048         ret
1049 .size   se_handler,.-se_handler
1050
1051 .section        .pdata
1052 .align  4
1053         .rva    .LSEH_begin_gcm_gmult_4bit
1054         .rva    .LSEH_end_gcm_gmult_4bit
1055         .rva    .LSEH_info_gcm_gmult_4bit
1056
1057         .rva    .LSEH_begin_gcm_ghash_4bit
1058         .rva    .LSEH_end_gcm_ghash_4bit
1059         .rva    .LSEH_info_gcm_ghash_4bit
1060
1061         .rva    .LSEH_begin_gcm_ghash_clmul
1062         .rva    .LSEH_end_gcm_ghash_clmul
1063         .rva    .LSEH_info_gcm_ghash_clmul
1064
1065 .section        .xdata
1066 .align  8
1067 .LSEH_info_gcm_gmult_4bit:
1068         .byte   9,0,0,0
1069         .rva    se_handler
1070         .rva    .Lgmult_prologue,.Lgmult_epilogue       # HandlerData
1071 .LSEH_info_gcm_ghash_4bit:
1072         .byte   9,0,0,0
1073         .rva    se_handler
1074         .rva    .Lghash_prologue,.Lghash_epilogue       # HandlerData
1075 .LSEH_info_gcm_ghash_clmul:
1076         .byte   0x01,0x33,0x16,0x00
1077         .byte   0x33,0xf8,0x09,0x00     #movaps 0x90(rsp),xmm15
1078         .byte   0x2e,0xe8,0x08,0x00     #movaps 0x80(rsp),xmm14
1079         .byte   0x29,0xd8,0x07,0x00     #movaps 0x70(rsp),xmm13
1080         .byte   0x24,0xc8,0x06,0x00     #movaps 0x60(rsp),xmm12
1081         .byte   0x1f,0xb8,0x05,0x00     #movaps 0x50(rsp),xmm11
1082         .byte   0x1a,0xa8,0x04,0x00     #movaps 0x40(rsp),xmm10
1083         .byte   0x15,0x98,0x03,0x00     #movaps 0x30(rsp),xmm9
1084         .byte   0x10,0x88,0x02,0x00     #movaps 0x20(rsp),xmm8
1085         .byte   0x0c,0x78,0x01,0x00     #movaps 0x10(rsp),xmm7
1086         .byte   0x08,0x68,0x00,0x00     #movaps 0x00(rsp),xmm6
1087         .byte   0x04,0x01,0x15,0x00     #sub    0xa8,rsp
1088 ___
1089 }
1090 \f
1091 $code =~ s/\`([^\`]*)\`/eval($1)/gem;
1092
1093 print $code;
1094
1095 close STDOUT;