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      1.78(+13%)
  63 # Sandy Bridge  1.80(+8%)
  64 # Ivy Bridge    1.80(+7%)
  65 # Haswell       0.55(+93%) (if system doesn't support AVX)
  66 # Bulldozer     1.49(+27%)
  67
  68 # March 2013
  69 #
  70 # ... 8x aggregate factor AVX code path is using reduction algorithm
  71 # suggested by Shay Gueron[1]. Even though contemporary AVX-capable
  72 # CPUs such as Sandy and Ivy Bridge can execute it, the code performs
  73 # sub-optimally in comparison to above mentioned version. But thanks
  74 # to Ilya Albrekht and Max Locktyukhin of Intel Corp. we knew that
  75 # it performs in 0.41 cycles per byte on Haswell processor.
  76 #
  77 # [1] http://rt.openssl.org/Ticket/Display.html?id=2900&user=guest&pass=guest
  78
  79 $flavour = shift;
  80 $output  = shift;
  81 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  82
  83 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
  84
  85 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  86 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  87 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
  88 die "can't locate x86_64-xlate.pl";
  89
  90 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
  91                 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
  92         $avx = ($1>=2.19) + ($1>=2.22);
  93 }
  94
  95 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
  96             `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
  97         $avx = ($1>=2.09) + ($1>=2.10);
  98 }
  99
 100 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
 101             `ml64 2>&1` =~ /Version ([0-9]+)\./) {
 102         $avx = ($1>=10) + ($1>=11);
 103 }
 104
 105 if (!$avx && `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/) {
 106         $avx = ($2>=3.0) + ($2>3.0);
 107 }
 108
 109 open OUT,"| \"$^X\" $xlate $flavour $output";
 110 *STDOUT=*OUT;
 111
 112 $do4xaggr=1;
 113
 114 # common register layout
 115 $nlo="%rax";
 116 $nhi="%rbx";
 117 $Zlo="%r8";
 118 $Zhi="%r9";
 119 $tmp="%r10";
 120 $rem_4bit = "%r11";
 121
 122 $Xi="%rdi";
 123 $Htbl="%rsi";
 124
 125 # per-function register layout
 126 $cnt="%rcx";
 127 $rem="%rdx";
 128
 129 sub LB() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/     or
 130                         $r =~ s/%[er]([sd]i)/%\1l/      or
 131                         $r =~ s/%[er](bp)/%\1l/         or
 132                         $r =~ s/%(r[0-9]+)[d]?/%\1b/;   $r; }
 133
 134 sub AUTOLOAD()          # thunk [simplified] 32-bit style perlasm
 135 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
 136   my $arg = pop;
 137     $arg = "\$$arg" if ($arg*1 eq $arg);
 138     $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
 139 }
 140 \f
 141 { my $N;
 142   sub loop() {
 143   my $inp = shift;
 144
 145         $N++;
 146 $code.=<<___;
 147         xor     $nlo,$nlo
 148         xor     $nhi,$nhi
 149         mov     `&LB("$Zlo")`,`&LB("$nlo")`
 150         mov     `&LB("$Zlo")`,`&LB("$nhi")`
 151         shl     \$4,`&LB("$nlo")`
 152         mov     \$14,$cnt
 153         mov     8($Htbl,$nlo),$Zlo
 154         mov     ($Htbl,$nlo),$Zhi
 155         and     \$0xf0,`&LB("$nhi")`
 156         mov     $Zlo,$rem
 157         jmp     .Loop$N
 158
 159 .align  16
 160 .Loop$N:
 161         shr     \$4,$Zlo
 162         and     \$0xf,$rem
 163         mov     $Zhi,$tmp
 164         mov     ($inp,$cnt),`&LB("$nlo")`
 165         shr     \$4,$Zhi
 166         xor     8($Htbl,$nhi),$Zlo
 167         shl     \$60,$tmp
 168         xor     ($Htbl,$nhi),$Zhi
 169         mov     `&LB("$nlo")`,`&LB("$nhi")`
 170         xor     ($rem_4bit,$rem,8),$Zhi
 171         mov     $Zlo,$rem
 172         shl     \$4,`&LB("$nlo")`
 173         xor     $tmp,$Zlo
 174         dec     $cnt
 175         js      .Lbreak$N
 176
 177         shr     \$4,$Zlo
 178         and     \$0xf,$rem
 179         mov     $Zhi,$tmp
 180         shr     \$4,$Zhi
 181         xor     8($Htbl,$nlo),$Zlo
 182         shl     \$60,$tmp
 183         xor     ($Htbl,$nlo),$Zhi
 184         and     \$0xf0,`&LB("$nhi")`
 185         xor     ($rem_4bit,$rem,8),$Zhi
 186         mov     $Zlo,$rem
 187         xor     $tmp,$Zlo
 188         jmp     .Loop$N
 189
 190 .align  16
 191 .Lbreak$N:
 192         shr     \$4,$Zlo
 193         and     \$0xf,$rem
 194         mov     $Zhi,$tmp
 195         shr     \$4,$Zhi
 196         xor     8($Htbl,$nlo),$Zlo
 197         shl     \$60,$tmp
 198         xor     ($Htbl,$nlo),$Zhi
 199         and     \$0xf0,`&LB("$nhi")`
 200         xor     ($rem_4bit,$rem,8),$Zhi
 201         mov     $Zlo,$rem
 202         xor     $tmp,$Zlo
 203
 204         shr     \$4,$Zlo
 205         and     \$0xf,$rem
 206         mov     $Zhi,$tmp
 207         shr     \$4,$Zhi
 208         xor     8($Htbl,$nhi),$Zlo
 209         shl     \$60,$tmp
 210         xor     ($Htbl,$nhi),$Zhi
 211         xor     $tmp,$Zlo
 212         xor     ($rem_4bit,$rem,8),$Zhi
 213
 214         bswap   $Zlo
 215         bswap   $Zhi
 216 ___
 217 }}
 218
 219 $code=<<___;
 220 .text
 221 .extern OPENSSL_ia32cap_P
 222
 223 .globl  gcm_gmult_4bit
 224 .type   gcm_gmult_4bit,\@function,2
 225 .align  16
 226 gcm_gmult_4bit:
 227         push    %rbx
 228         push    %rbp            # %rbp and %r12 are pushed exclusively in
 229         push    %r12            # order to reuse Win64 exception handler...
 230 .Lgmult_prologue:
 231
 232         movzb   15($Xi),$Zlo
 233         lea     .Lrem_4bit(%rip),$rem_4bit
 234 ___
 235         &loop   ($Xi);
 236 $code.=<<___;
 237         mov     $Zlo,8($Xi)
 238         mov     $Zhi,($Xi)
 239
 240         mov     16(%rsp),%rbx
 241         lea     24(%rsp),%rsp
 242 .Lgmult_epilogue:
 243         ret
 244 .size   gcm_gmult_4bit,.-gcm_gmult_4bit
 245 ___
 246 \f
 247 # per-function register layout
 248 $inp="%rdx";
 249 $len="%rcx";
 250 $rem_8bit=$rem_4bit;
 251
 252 $code.=<<___;
 253 .globl  gcm_ghash_4bit
 254 .type   gcm_ghash_4bit,\@function,4
 255 .align  16
 256 gcm_ghash_4bit:
 257         push    %rbx
 258         push    %rbp
 259         push    %r12
 260         push    %r13
 261         push    %r14
 262         push    %r15
 263         sub     \$280,%rsp
 264 .Lghash_prologue:
 265         mov     $inp,%r14               # reassign couple of args
 266         mov     $len,%r15
 267 ___
 268 { my $inp="%r14";
 269   my $dat="%edx";
 270   my $len="%r15";
 271   my @nhi=("%ebx","%ecx");
 272   my @rem=("%r12","%r13");
 273   my $Hshr4="%rbp";
 274
 275         &sub    ($Htbl,-128);           # size optimization
 276         &lea    ($Hshr4,"16+128(%rsp)");
 277         { my @lo =($nlo,$nhi);
 278           my @hi =($Zlo,$Zhi);
 279
 280           &xor  ($dat,$dat);
 281           for ($i=0,$j=-2;$i<18;$i++,$j++) {
 282             &mov        ("$j(%rsp)",&LB($dat))          if ($i>1);
 283             &or         ($lo[0],$tmp)                   if ($i>1);
 284             &mov        (&LB($dat),&LB($lo[1]))         if ($i>0 && $i<17);
 285             &shr        ($lo[1],4)                      if ($i>0 && $i<17);
 286             &mov        ($tmp,$hi[1])                   if ($i>0 && $i<17);
 287             &shr        ($hi[1],4)                      if ($i>0 && $i<17);
 288             &mov        ("8*$j($Hshr4)",$hi[0])         if ($i>1);
 289             &mov        ($hi[0],"16*$i+0-128($Htbl)")   if ($i<16);
 290             &shl        (&LB($dat),4)                   if ($i>0 && $i<17);
 291             &mov        ("8*$j-128($Hshr4)",$lo[0])     if ($i>1);
 292             &mov        ($lo[0],"16*$i+8-128($Htbl)")   if ($i<16);
 293             &shl        ($tmp,60)                       if ($i>0 && $i<17);
 294
 295             push        (@lo,shift(@lo));
 296             push        (@hi,shift(@hi));
 297           }
 298         }
 299         &add    ($Htbl,-128);
 300         &mov    ($Zlo,"8($Xi)");
 301         &mov    ($Zhi,"0($Xi)");
 302         &add    ($len,$inp);            # pointer to the end of data
 303         &lea    ($rem_8bit,".Lrem_8bit(%rip)");
 304         &jmp    (".Louter_loop");
 305
 306 $code.=".align  16\n.Louter_loop:\n";
 307         &xor    ($Zhi,"($inp)");
 308         &mov    ("%rdx","8($inp)");
 309         &lea    ($inp,"16($inp)");
 310         &xor    ("%rdx",$Zlo);
 311         &mov    ("($Xi)",$Zhi);
 312         &mov    ("8($Xi)","%rdx");
 313         &shr    ("%rdx",32);
 314
 315         &xor    ($nlo,$nlo);
 316         &rol    ($dat,8);
 317         &mov    (&LB($nlo),&LB($dat));
 318         &movz   ($nhi[0],&LB($dat));
 319         &shl    (&LB($nlo),4);
 320         &shr    ($nhi[0],4);
 321
 322         for ($j=11,$i=0;$i<15;$i++) {
 323             &rol        ($dat,8);
 324             &xor        ($Zlo,"8($Htbl,$nlo)")                  if ($i>0);
 325             &xor        ($Zhi,"($Htbl,$nlo)")                   if ($i>0);
 326             &mov        ($Zlo,"8($Htbl,$nlo)")                  if ($i==0);
 327             &mov        ($Zhi,"($Htbl,$nlo)")                   if ($i==0);
 328
 329             &mov        (&LB($nlo),&LB($dat));
 330             &xor        ($Zlo,$tmp)                             if ($i>0);
 331             &movzw      ($rem[1],"($rem_8bit,$rem[1],2)")       if ($i>0);
 332
 333             &movz       ($nhi[1],&LB($dat));
 334             &shl        (&LB($nlo),4);
 335             &movzb      ($rem[0],"(%rsp,$nhi[0])");
 336
 337             &shr        ($nhi[1],4)                             if ($i<14);
 338             &and        ($nhi[1],0xf0)                          if ($i==14);
 339             &shl        ($rem[1],48)                            if ($i>0);
 340             &xor        ($rem[0],$Zlo);
 341
 342             &mov        ($tmp,$Zhi);
 343             &xor        ($Zhi,$rem[1])                          if ($i>0);
 344             &shr        ($Zlo,8);
 345
 346             &movz       ($rem[0],&LB($rem[0]));
 347             &mov        ($dat,"$j($Xi)")                        if (--$j%4==0);
 348             &shr        ($Zhi,8);
 349
 350             &xor        ($Zlo,"-128($Hshr4,$nhi[0],8)");
 351             &shl        ($tmp,56);
 352             &xor        ($Zhi,"($Hshr4,$nhi[0],8)");
 353
 354             unshift     (@nhi,pop(@nhi));               # "rotate" registers
 355             unshift     (@rem,pop(@rem));
 356         }
 357         &movzw  ($rem[1],"($rem_8bit,$rem[1],2)");
 358         &xor    ($Zlo,"8($Htbl,$nlo)");
 359         &xor    ($Zhi,"($Htbl,$nlo)");
 360
 361         &shl    ($rem[1],48);
 362         &xor    ($Zlo,$tmp);
 363
 364         &xor    ($Zhi,$rem[1]);
 365         &movz   ($rem[0],&LB($Zlo));
 366         &shr    ($Zlo,4);
 367
 368         &mov    ($tmp,$Zhi);
 369         &shl    (&LB($rem[0]),4);
 370         &shr    ($Zhi,4);
 371
 372         &xor    ($Zlo,"8($Htbl,$nhi[0])");
 373         &movzw  ($rem[0],"($rem_8bit,$rem[0],2)");
 374         &shl    ($tmp,60);
 375
 376         &xor    ($Zhi,"($Htbl,$nhi[0])");
 377         &xor    ($Zlo,$tmp);
 378         &shl    ($rem[0],48);
 379
 380         &bswap  ($Zlo);
 381         &xor    ($Zhi,$rem[0]);
 382
 383         &bswap  ($Zhi);
 384         &cmp    ($inp,$len);
 385         &jb     (".Louter_loop");
 386 }
 387 $code.=<<___;
 388         mov     $Zlo,8($Xi)
 389         mov     $Zhi,($Xi)
 390
 391         lea     280(%rsp),%rsi
 392         mov     0(%rsi),%r15
 393         mov     8(%rsi),%r14
 394         mov     16(%rsi),%r13
 395         mov     24(%rsi),%r12
 396         mov     32(%rsi),%rbp
 397         mov     40(%rsi),%rbx
 398         lea     48(%rsi),%rsp
 399 .Lghash_epilogue:
 400         ret
 401 .size   gcm_ghash_4bit,.-gcm_ghash_4bit
 402 ___
 403 \f
 404 ######################################################################
 405 # PCLMULQDQ version.
 406
 407 @_4args=$win64? ("%rcx","%rdx","%r8", "%r9") :  # Win64 order
 408                 ("%rdi","%rsi","%rdx","%rcx");  # Unix order
 409
 410 ($Xi,$Xhi)=("%xmm0","%xmm1");   $Hkey="%xmm2";
 411 ($T1,$T2,$T3)=("%xmm3","%xmm4","%xmm5");
 412
 413 sub clmul64x64_T2 {     # minimal register pressure
 414 my ($Xhi,$Xi,$Hkey,$HK)=@_;
 415
 416 if (!defined($HK)) {    $HK = $T2;
 417 $code.=<<___;
 418         movdqa          $Xi,$Xhi                #
 419         pshufd          \$0b01001110,$Xi,$T1
 420         pshufd          \$0b01001110,$Hkey,$T2
 421         pxor            $Xi,$T1                 #
 422         pxor            $Hkey,$T2
 423 ___
 424 } else {
 425 $code.=<<___;
 426         movdqa          $Xi,$Xhi                #
 427         pshufd          \$0b01001110,$Xi,$T1
 428         pxor            $Xi,$T1                 #
 429 ___
 430 }
 431 $code.=<<___;
 432         pclmulqdq       \$0x00,$Hkey,$Xi        #######
 433         pclmulqdq       \$0x11,$Hkey,$Xhi       #######
 434         pclmulqdq       \$0x00,$HK,$T1          #######
 435         pxor            $Xi,$T1                 #
 436         pxor            $Xhi,$T1                #
 437
 438         movdqa          $T1,$T2                 #
 439         psrldq          \$8,$T1
 440         pslldq          \$8,$T2                 #
 441         pxor            $T1,$Xhi
 442         pxor            $T2,$Xi                 #
 443 ___
 444 }
 445
 446 sub reduction_alg9 {    # 17/11 times faster than Intel version
 447 my ($Xhi,$Xi) = @_;
 448
 449 $code.=<<___;
 450         # 1st phase
 451         movdqa          $Xi,$T2                 #
 452         movdqa          $Xi,$T1
 453         psllq           \$5,$Xi
 454         pxor            $Xi,$T1                 #
 455         psllq           \$1,$Xi
 456         pxor            $T1,$Xi                 #
 457         psllq           \$57,$Xi                #
 458         movdqa          $Xi,$T1                 #
 459         pslldq          \$8,$Xi
 460         psrldq          \$8,$T1                 #
 461         pxor            $T2,$Xi
 462         pxor            $T1,$Xhi                #
 463
 464         # 2nd phase
 465         movdqa          $Xi,$T2
 466         psrlq           \$1,$Xi
 467         pxor            $T2,$Xhi                #
 468         pxor            $Xi,$T2
 469         psrlq           \$5,$Xi
 470         pxor            $T2,$Xi                 #
 471         psrlq           \$1,$Xi                 #
 472         pxor            $Xhi,$Xi                #
 473 ___
 474 }
 475 \f
 476 { my ($Htbl,$Xip)=@_4args;
 477   my $HK="%xmm6";
 478
 479 $code.=<<___;
 480 .globl  gcm_init_clmul
 481 .type   gcm_init_clmul,\@abi-omnipotent
 482 .align  16
 483 gcm_init_clmul:
 484 .L_init_clmul:
 485 ___
 486 $code.=<<___ if ($win64);
 487 .LSEH_begin_gcm_init_clmul:
 488         # I can't trust assembler to use specific encoding:-(
 489         .byte   0x48,0x83,0xec,0x18             #sub    $0x18,%rsp
 490         .byte   0x0f,0x29,0x34,0x24             #movaps %xmm6,(%rsp)
 491 ___
 492 $code.=<<___;
 493         movdqu          ($Xip),$Hkey
 494         pshufd          \$0b01001110,$Hkey,$Hkey        # dword swap
 495
 496         # <<1 twist
 497         pshufd          \$0b11111111,$Hkey,$T2  # broadcast uppermost dword
 498         movdqa          $Hkey,$T1
 499         psllq           \$1,$Hkey
 500         pxor            $T3,$T3                 #
 501         psrlq           \$63,$T1
 502         pcmpgtd         $T2,$T3                 # broadcast carry bit
 503         pslldq          \$8,$T1
 504         por             $T1,$Hkey               # H<<=1
 505
 506         # magic reduction
 507         pand            .L0x1c2_polynomial(%rip),$T3
 508         pxor            $T3,$Hkey               # if(carry) H^=0x1c2_polynomial
 509
 510         # calculate H^2
 511         pshufd          \$0b01001110,$Hkey,$HK
 512         movdqa          $Hkey,$Xi
 513         pxor            $Hkey,$HK
 514 ___
 515         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);
 516         &reduction_alg9 ($Xhi,$Xi);
 517 $code.=<<___;
 518         pshufd          \$0b01001110,$Hkey,$T1
 519         pshufd          \$0b01001110,$Xi,$T2
 520         pxor            $Hkey,$T1               # Karatsuba pre-processing
 521         movdqu          $Hkey,0x00($Htbl)       # save H
 522         pxor            $Xi,$T2                 # Karatsuba pre-processing
 523         movdqu          $Xi,0x10($Htbl)         # save H^2
 524         palignr         \$8,$T1,$T2             # low part is H.lo^H.hi...
 525         movdqu          $T2,0x20($Htbl)         # save Karatsuba "salt"
 526 ___
 527 if ($do4xaggr) {
 528         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);   # H^3
 529         &reduction_alg9 ($Xhi,$Xi);
 530 $code.=<<___;
 531         movdqa          $Xi,$T3
 532 ___
 533         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);   # H^4
 534         &reduction_alg9 ($Xhi,$Xi);
 535 $code.=<<___;
 536         pshufd          \$0b01001110,$T3,$T1
 537         pshufd          \$0b01001110,$Xi,$T2
 538         pxor            $T3,$T1                 # Karatsuba pre-processing
 539         movdqu          $T3,0x30($Htbl)         # save H^3
 540         pxor            $Xi,$T2                 # Karatsuba pre-processing
 541         movdqu          $Xi,0x40($Htbl)         # save H^4
 542         palignr         \$8,$T1,$T2             # low part is H^3.lo^H^3.hi...
 543         movdqu          $T2,0x50($Htbl)         # save Karatsuba "salt"
 544 ___
 545 }
 546 $code.=<<___ if ($win64);
 547         movaps  (%rsp),%xmm6
 548         lea     0x18(%rsp),%rsp
 549 .LSEH_end_gcm_init_clmul:
 550 ___
 551 $code.=<<___;
 552         ret
 553 .size   gcm_init_clmul,.-gcm_init_clmul
 554 ___
 555 }
 556
 557 { my ($Xip,$Htbl)=@_4args;
 558
 559 $code.=<<___;
 560 .globl  gcm_gmult_clmul
 561 .type   gcm_gmult_clmul,\@abi-omnipotent
 562 .align  16
 563 gcm_gmult_clmul:
 564 .L_gmult_clmul:
 565         movdqu          ($Xip),$Xi
 566         movdqa          .Lbswap_mask(%rip),$T3
 567         movdqu          ($Htbl),$Hkey
 568         movdqu          0x20($Htbl),$T2
 569         pshufb          $T3,$Xi
 570 ___
 571         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$T2);
 572 $code.=<<___ if (0 || (&reduction_alg9($Xhi,$Xi)&&0));
 573         # experimental alternative. special thing about is that there
 574         # no dependency between the two multiplications...
 575         mov             \$`0xE1<<1`,%eax
 576         mov             \$0xA040608020C0E000,%r10       # ((7..0)·0xE0)&0xff
 577         mov             \$0x07,%r11d
 578         movq            %rax,$T1
 579         movq            %r10,$T2
 580         movq            %r11,$T3                # borrow $T3
 581         pand            $Xi,$T3
 582         pshufb          $T3,$T2                 # ($Xi&7)·0xE0
 583         movq            %rax,$T3
 584         pclmulqdq       \$0x00,$Xi,$T1          # ·(0xE1<<1)
 585         pxor            $Xi,$T2
 586         pslldq          \$15,$T2
 587         paddd           $T2,$T2                 # <<(64+56+1)
 588         pxor            $T2,$Xi
 589         pclmulqdq       \$0x01,$T3,$Xi
 590         movdqa          .Lbswap_mask(%rip),$T3  # reload $T3
 591         psrldq          \$1,$T1
 592         pxor            $T1,$Xhi
 593         pslldq          \$7,$Xi
 594         pxor            $Xhi,$Xi
 595 ___
 596 $code.=<<___;
 597         pshufb          $T3,$Xi
 598         movdqu          $Xi,($Xip)
 599         ret
 600 .size   gcm_gmult_clmul,.-gcm_gmult_clmul
 601 ___
 602 }
 603 \f
 604 { my ($Xip,$Htbl,$inp,$len)=@_4args;
 605   my ($Xln,$Xmn,$Xhn,$Hkey2,$HK) = map("%xmm$_",(3..7));
 606   my ($T1,$T2,$T3)=map("%xmm$_",(8..10));
 607
 608 $code.=<<___;
 609 .globl  gcm_ghash_clmul
 610 .type   gcm_ghash_clmul,\@abi-omnipotent
 611 .align  32
 612 gcm_ghash_clmul:
 613 .L_ghash_clmul:
 614 ___
 615 $code.=<<___ if ($win64);
 616         lea     -0x88(%rsp),%rax
 617 .LSEH_begin_gcm_ghash_clmul:
 618         # I can't trust assembler to use specific encoding:-(
 619         .byte   0x48,0x8d,0x60,0xe0             #lea    -0x20(%rax),%rsp
 620         .byte   0x0f,0x29,0x70,0xe0             #movaps %xmm6,-0x20(%rax)
 621         .byte   0x0f,0x29,0x78,0xf0             #movaps %xmm7,-0x10(%rax)
 622         .byte   0x44,0x0f,0x29,0x00             #movaps %xmm8,0(%rax)
 623         .byte   0x44,0x0f,0x29,0x48,0x10        #movaps %xmm9,0x10(%rax)
 624         .byte   0x44,0x0f,0x29,0x50,0x20        #movaps %xmm10,0x20(%rax)
 625         .byte   0x44,0x0f,0x29,0x58,0x30        #movaps %xmm11,0x30(%rax)
 626         .byte   0x44,0x0f,0x29,0x60,0x40        #movaps %xmm12,0x40(%rax)
 627         .byte   0x44,0x0f,0x29,0x68,0x50        #movaps %xmm13,0x50(%rax)
 628         .byte   0x44,0x0f,0x29,0x70,0x60        #movaps %xmm14,0x60(%rax)
 629         .byte   0x44,0x0f,0x29,0x78,0x70        #movaps %xmm15,0x70(%rax)
 630 ___
 631 $code.=<<___;
 632         movdqa          .Lbswap_mask(%rip),$T3
 633
 634         movdqu          ($Xip),$Xi
 635         movdqu          ($Htbl),$Hkey
 636         movdqu          0x20($Htbl),$HK
 637         pshufb          $T3,$Xi
 638
 639         sub             \$0x10,$len
 640         jz              .Lodd_tail
 641
 642         movdqu          0x10($Htbl),$Hkey2
 643 ___
 644 if ($do4xaggr) {
 645 my ($Xl,$Xm,$Xh,$Hkey3,$Hkey4)=map("%xmm$_",(11..15));
 646
 647 $code.=<<___;
 648         mov             OPENSSL_ia32cap_P+4(%rip),%eax
 649         cmp             \$0x30,$len
 650         jb              .Lskip4x
 651
 652         and             \$`1<<26|1<<22`,%eax    # isolate MOVBE+XSAVE
 653         cmp             \$`1<<22`,%eax          # check for MOVBE without XSAVE
 654         je              .Lskip4x
 655
 656         sub             \$0x30,$len
 657         mov             \$0xA040608020C0E000,%rax       # ((7..0)·0xE0)&0xff
 658         movdqu          0x30($Htbl),$Hkey3
 659         movdqu          0x40($Htbl),$Hkey4
 660
 661         #######
 662         # Xi+4 =[(H*Ii+3) + (H^2*Ii+2) + (H^3*Ii+1) + H^4*(Ii+Xi)] mod P
 663         #
 664         movdqu          0x30($inp),$Xln
 665          movdqu         0x20($inp),$Xl
 666         pshufb          $T3,$Xln
 667          pshufb         $T3,$Xl
 668         movdqa          $Xln,$Xhn
 669         pshufd          \$0b01001110,$Xln,$Xmn
 670         pxor            $Xln,$Xmn
 671         pclmulqdq       \$0x00,$Hkey,$Xln
 672         pclmulqdq       \$0x11,$Hkey,$Xhn
 673         pclmulqdq       \$0x00,$HK,$Xmn
 674
 675         movdqa          $Xl,$Xh
 676         pshufd          \$0b01001110,$Xl,$Xm
 677         pxor            $Xl,$Xm
 678         pclmulqdq       \$0x00,$Hkey2,$Xl
 679         pclmulqdq       \$0x11,$Hkey2,$Xh
 680         pclmulqdq       \$0x10,$HK,$Xm
 681         xorps           $Xl,$Xln
 682         xorps           $Xh,$Xhn
 683         movups          0x50($Htbl),$HK
 684         xorps           $Xm,$Xmn
 685
 686         movdqu          0x10($inp),$Xl
 687          movdqu         0($inp),$T1
 688         pshufb          $T3,$Xl
 689          pshufb         $T3,$T1
 690         movdqa          $Xl,$Xh
 691         pshufd          \$0b01001110,$Xl,$Xm
 692          pxor           $T1,$Xi
 693         pxor            $Xl,$Xm
 694         pclmulqdq       \$0x00,$Hkey3,$Xl
 695          movdqa         $Xi,$Xhi
 696          pshufd         \$0b01001110,$Xi,$T1
 697          pxor           $Xi,$T1
 698         pclmulqdq       \$0x11,$Hkey3,$Xh
 699         pclmulqdq       \$0x00,$HK,$Xm
 700         xorps           $Xl,$Xln
 701         xorps           $Xh,$Xhn
 702
 703         lea     0x40($inp),$inp
 704         sub     \$0x40,$len
 705         jc      .Ltail4x
 706
 707         jmp     .Lmod4_loop
 708 .align  32
 709 .Lmod4_loop:
 710         pclmulqdq       \$0x00,$Hkey4,$Xi
 711         xorps           $Xm,$Xmn
 712          movdqu         0x30($inp),$Xl
 713          pshufb         $T3,$Xl
 714         pclmulqdq       \$0x11,$Hkey4,$Xhi
 715         xorps           $Xln,$Xi
 716          movdqu         0x20($inp),$Xln
 717          movdqa         $Xl,$Xh
 718         pclmulqdq       \$0x10,$HK,$T1
 719          pshufd         \$0b01001110,$Xl,$Xm
 720         xorps           $Xhn,$Xhi
 721          pxor           $Xl,$Xm
 722          pshufb         $T3,$Xln
 723         movups          0x20($Htbl),$HK
 724         xorps           $Xmn,$T1
 725          pclmulqdq      \$0x00,$Hkey,$Xl
 726          pshufd         \$0b01001110,$Xln,$Xmn
 727
 728         pxor            $Xi,$T1                 # aggregated Karatsuba post-processing
 729          movdqa         $Xln,$Xhn
 730         pxor            $Xhi,$T1                #
 731          pxor           $Xln,$Xmn
 732         movdqa          $T1,$T2                 #
 733          pclmulqdq      \$0x11,$Hkey,$Xh
 734         pslldq          \$8,$T1
 735         psrldq          \$8,$T2                 #
 736         pxor            $T1,$Xi
 737         movdqa          .L7_mask(%rip),$T1
 738         pxor            $T2,$Xhi                #
 739         movq            %rax,$T2
 740
 741         pand            $Xi,$T1                 # 1st phase
 742         pshufb          $T1,$T2                 #
 743         pxor            $Xi,$T2                 #
 744          pclmulqdq      \$0x00,$HK,$Xm
 745         psllq           \$57,$T2                #
 746         movdqa          $T2,$T1                 #
 747         pslldq          \$8,$T2
 748          pclmulqdq      \$0x00,$Hkey2,$Xln
 749         psrldq          \$8,$T1                 #
 750         pxor            $T2,$Xi
 751         pxor            $T1,$Xhi                #
 752         movdqu          0($inp),$T1
 753
 754         movdqa          $Xi,$T2                 # 2nd phase
 755         psrlq           \$1,$Xi
 756          pclmulqdq      \$0x11,$Hkey2,$Xhn
 757          xorps          $Xl,$Xln
 758          movdqu         0x10($inp),$Xl
 759          pshufb         $T3,$Xl
 760          pclmulqdq      \$0x10,$HK,$Xmn
 761          xorps          $Xh,$Xhn
 762          movups         0x50($Htbl),$HK
 763         pshufb          $T3,$T1
 764         pxor            $T2,$Xhi                #
 765         pxor            $Xi,$T2
 766         psrlq           \$5,$Xi
 767
 768          movdqa         $Xl,$Xh
 769          pxor           $Xm,$Xmn
 770          pshufd         \$0b01001110,$Xl,$Xm
 771         pxor            $T2,$Xi                 #
 772         pxor            $T1,$Xhi
 773          pxor           $Xl,$Xm
 774          pclmulqdq      \$0x00,$Hkey3,$Xl
 775         psrlq           \$1,$Xi                 #
 776         pxor            $Xhi,$Xi                #
 777         movdqa          $Xi,$Xhi
 778          pclmulqdq      \$0x11,$Hkey3,$Xh
 779          xorps          $Xl,$Xln
 780         pshufd          \$0b01001110,$Xi,$T1
 781         pxor            $Xi,$T1
 782
 783          pclmulqdq      \$0x00,$HK,$Xm
 784          xorps          $Xh,$Xhn
 785
 786         lea     0x40($inp),$inp
 787         sub     \$0x40,$len
 788         jnc     .Lmod4_loop
 789
 790 .Ltail4x:
 791         pclmulqdq       \$0x00,$Hkey4,$Xi
 792         pclmulqdq       \$0x11,$Hkey4,$Xhi
 793         pclmulqdq       \$0x10,$HK,$T1
 794         xorps           $Xm,$Xmn
 795         xorps           $Xln,$Xi
 796         xorps           $Xhn,$Xhi
 797         pxor            $Xi,$Xhi                # aggregated Karatsuba post-processing
 798         pxor            $Xmn,$T1
 799
 800         pxor            $Xhi,$T1                #
 801         pxor            $Xi,$Xhi
 802
 803         movdqa          $T1,$T2                 #
 804         psrldq          \$8,$T1
 805         pslldq          \$8,$T2                 #
 806         pxor            $T1,$Xhi
 807         pxor            $T2,$Xi                 #
 808 ___
 809         &reduction_alg9($Xhi,$Xi);
 810 $code.=<<___;
 811         add     \$0x40,$len
 812         jz      .Ldone
 813         movdqu  0x20($Htbl),$HK
 814         sub     \$0x10,$len
 815         jz      .Lodd_tail
 816 .Lskip4x:
 817 ___
 818 }
 819 $code.=<<___;
 820         #######
 821         # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
 822         #       [(H*Ii+1) + (H*Xi+1)] mod P =
 823         #       [(H*Ii+1) + H^2*(Ii+Xi)] mod P
 824         #
 825         movdqu          ($inp),$T1              # Ii
 826         movdqu          16($inp),$Xln           # Ii+1
 827         pshufb          $T3,$T1
 828         pshufb          $T3,$Xln
 829         pxor            $T1,$Xi                 # Ii+Xi
 830
 831         movdqa          $Xln,$Xhn
 832         pshufd          \$0b01001110,$Xln,$Xmn
 833         pxor            $Xln,$Xmn
 834         pclmulqdq       \$0x00,$Hkey,$Xln
 835         pclmulqdq       \$0x11,$Hkey,$Xhn
 836         pclmulqdq       \$0x00,$HK,$Xmn
 837
 838         lea             32($inp),$inp           # i+=2
 839         nop
 840         sub             \$0x20,$len
 841         jbe             .Leven_tail
 842         nop
 843         jmp             .Lmod_loop
 844
 845 .align  32
 846 .Lmod_loop:
 847         movdqa          $Xi,$Xhi
 848         movdqa          $Xmn,$T1
 849         pshufd          \$0b01001110,$Xi,$Xmn   #
 850         pxor            $Xi,$Xmn                #
 851
 852         pclmulqdq       \$0x00,$Hkey2,$Xi
 853         pclmulqdq       \$0x11,$Hkey2,$Xhi
 854         pclmulqdq       \$0x10,$HK,$Xmn
 855
 856         pxor            $Xln,$Xi                # (H*Ii+1) + H^2*(Ii+Xi)
 857         pxor            $Xhn,$Xhi
 858           movdqu        ($inp),$T2              # Ii
 859         pxor            $Xi,$T1                 # aggregated Karatsuba post-processing
 860           pshufb        $T3,$T2
 861           movdqu        16($inp),$Xln           # Ii+1
 862
 863         pxor            $Xhi,$T1
 864           pxor          $T2,$Xhi                # "Ii+Xi", consume early
 865         pxor            $T1,$Xmn
 866          pshufb         $T3,$Xln
 867         movdqa          $Xmn,$T1                #
 868         psrldq          \$8,$T1
 869         pslldq          \$8,$Xmn                #
 870         pxor            $T1,$Xhi
 871         pxor            $Xmn,$Xi                #
 872
 873         movdqa          $Xln,$Xhn               #
 874
 875           movdqa        $Xi,$T2                 # 1st phase
 876           movdqa        $Xi,$T1
 877           psllq         \$5,$Xi
 878           pxor          $Xi,$T1                 #
 879         pclmulqdq       \$0x00,$Hkey,$Xln       #######
 880           psllq         \$1,$Xi
 881           pxor          $T1,$Xi                 #
 882           psllq         \$57,$Xi                #
 883           movdqa        $Xi,$T1                 #
 884           pslldq        \$8,$Xi
 885           psrldq        \$8,$T1                 #
 886           pxor          $T2,$Xi
 887         pshufd          \$0b01001110,$Xhn,$Xmn
 888           pxor          $T1,$Xhi                #
 889         pxor            $Xhn,$Xmn               #
 890
 891           movdqa        $Xi,$T2                 # 2nd phase
 892           psrlq         \$1,$Xi
 893         pclmulqdq       \$0x11,$Hkey,$Xhn       #######
 894           pxor          $T2,$Xhi                #
 895           pxor          $Xi,$T2
 896           psrlq         \$5,$Xi
 897           pxor          $T2,$Xi                 #
 898         lea             32($inp),$inp
 899           psrlq         \$1,$Xi                 #
 900         pclmulqdq       \$0x00,$HK,$Xmn         #######
 901           pxor          $Xhi,$Xi                #
 902
 903         sub             \$0x20,$len
 904         ja              .Lmod_loop
 905
 906 .Leven_tail:
 907          movdqa         $Xi,$Xhi
 908          movdqa         $Xmn,$T1
 909          pshufd         \$0b01001110,$Xi,$Xmn   #
 910          pxor           $Xi,$Xmn                #
 911
 912         pclmulqdq       \$0x00,$Hkey2,$Xi
 913         pclmulqdq       \$0x11,$Hkey2,$Xhi
 914         pclmulqdq       \$0x10,$HK,$Xmn
 915
 916         pxor            $Xln,$Xi                # (H*Ii+1) + H^2*(Ii+Xi)
 917         pxor            $Xhn,$Xhi
 918         pxor            $Xi,$T1
 919         pxor            $Xhi,$T1
 920         pxor            $T1,$Xmn
 921         movdqa          $Xmn,$T1                #
 922         psrldq          \$8,$T1
 923         pslldq          \$8,$Xmn                #
 924         pxor            $T1,$Xhi
 925         pxor            $Xmn,$Xi                #
 926 ___
 927         &reduction_alg9 ($Xhi,$Xi);
 928 $code.=<<___;
 929         test            $len,$len
 930         jnz             .Ldone
 931
 932 .Lodd_tail:
 933         movdqu          ($inp),$T1              # Ii
 934         pshufb          $T3,$T1
 935         pxor            $T1,$Xi                 # Ii+Xi
 936 ___
 937         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);   # H*(Ii+Xi)
 938         &reduction_alg9 ($Xhi,$Xi);
 939 $code.=<<___;
 940 .Ldone:
 941         pshufb          $T3,$Xi
 942         movdqu          $Xi,($Xip)
 943 ___
 944 $code.=<<___ if ($win64);
 945         movaps  (%rsp),%xmm6
 946         movaps  0x10(%rsp),%xmm7
 947         movaps  0x20(%rsp),%xmm8
 948         movaps  0x30(%rsp),%xmm9
 949         movaps  0x40(%rsp),%xmm10
 950         movaps  0x50(%rsp),%xmm11
 951         movaps  0x60(%rsp),%xmm12
 952         movaps  0x70(%rsp),%xmm13
 953         movaps  0x80(%rsp),%xmm14
 954         movaps  0x90(%rsp),%xmm15
 955         lea     0xa8(%rsp),%rsp
 956 .LSEH_end_gcm_ghash_clmul:
 957 ___
 958 $code.=<<___;
 959         ret
 960 .size   gcm_ghash_clmul,.-gcm_ghash_clmul
 961 ___
 962 }
 963 \f
 964 $code.=<<___;
 965 .globl  gcm_init_avx
 966 .type   gcm_init_avx,\@abi-omnipotent
 967 .align  32
 968 gcm_init_avx:
 969 ___
 970 if ($avx) {
 971 my ($Htbl,$Xip)=@_4args;
 972 my $HK="%xmm6";
 973
 974 $code.=<<___ if ($win64);
 975 .LSEH_begin_gcm_init_avx:
 976         # I can't trust assembler to use specific encoding:-(
 977         .byte   0x48,0x83,0xec,0x18             #sub    $0x18,%rsp
 978         .byte   0x0f,0x29,0x34,0x24             #movaps %xmm6,(%rsp)
 979 ___
 980 $code.=<<___;
 981         vzeroupper
 982
 983         vmovdqu         ($Xip),$Hkey
 984         vpshufd         \$0b01001110,$Hkey,$Hkey        # dword swap
 985
 986         # <<1 twist
 987         vpshufd         \$0b11111111,$Hkey,$T2  # broadcast uppermost dword
 988         vpsrlq          \$63,$Hkey,$T1
 989         vpsllq          \$1,$Hkey,$Hkey
 990         vpxor           $T3,$T3,$T3             #
 991         vpcmpgtd        $T2,$T3,$T3             # broadcast carry bit
 992         vpslldq         \$8,$T1,$T1
 993         vpor            $T1,$Hkey,$Hkey         # H<<=1
 994
 995         # magic reduction
 996         vpand           .L0x1c2_polynomial(%rip),$T3,$T3
 997         vpxor           $T3,$Hkey,$Hkey         # if(carry) H^=0x1c2_polynomial
 998
 999         vpunpckhqdq     $Hkey,$Hkey,$HK
1000         vmovdqa         $Hkey,$Xi
1001         vpxor           $Hkey,$HK,$HK
1002         mov             \$4,%r10                # up to H^8
1003         jmp             .Linit_start_avx
1004 ___
1005
1006 sub clmul64x64_avx {
1007 my ($Xhi,$Xi,$Hkey,$HK)=@_;
1008
1009 if (!defined($HK)) {    $HK = $T2;
1010 $code.=<<___;
1011         vpunpckhqdq     $Xi,$Xi,$T1
1012         vpunpckhqdq     $Hkey,$Hkey,$T2
1013         vpxor           $Xi,$T1,$T1             #
1014         vpxor           $Hkey,$T2,$T2
1015 ___
1016 } else {
1017 $code.=<<___;
1018         vpunpckhqdq     $Xi,$Xi,$T1
1019         vpxor           $Xi,$T1,$T1             #
1020 ___
1021 }
1022 $code.=<<___;
1023         vpclmulqdq      \$0x11,$Hkey,$Xi,$Xhi   #######
1024         vpclmulqdq      \$0x00,$Hkey,$Xi,$Xi    #######
1025         vpclmulqdq      \$0x00,$HK,$T1,$T1      #######
1026         vpxor           $Xi,$Xhi,$T2            #
1027         vpxor           $T2,$T1,$T1             #
1028
1029         vpslldq         \$8,$T1,$T2             #
1030         vpsrldq         \$8,$T1,$T1
1031         vpxor           $T2,$Xi,$Xi             #
1032         vpxor           $T1,$Xhi,$Xhi
1033 ___
1034 }
1035
1036 sub reduction_avx {
1037 my ($Xhi,$Xi) = @_;
1038
1039 $code.=<<___;
1040         vpsllq          \$57,$Xi,$T1            # 1st phase
1041         vpsllq          \$62,$Xi,$T2
1042         vpxor           $T1,$T2,$T2             #
1043         vpsllq          \$63,$Xi,$T1
1044         vpxor           $T1,$T2,$T2             #
1045         vpslldq         \$8,$T2,$T1             #
1046         vpsrldq         \$8,$T2,$T2
1047         vpxor           $T1,$Xi,$Xi             #
1048         vpxor           $T2,$Xhi,$Xhi
1049
1050         vpsrlq          \$1,$Xi,$T2             # 2nd phase
1051         vpxor           $Xi,$Xhi,$Xhi
1052         vpxor           $T2,$Xi,$Xi             #
1053         vpsrlq          \$5,$T2,$T2
1054         vpxor           $T2,$Xi,$Xi             #
1055         vpsrlq          \$1,$Xi,$Xi             #
1056         vpxor           $Xhi,$Xi,$Xi            #
1057 ___
1058 }
1059
1060 $code.=<<___;
1061 .align  32
1062 .Linit_loop_avx:
1063         vpalignr        \$8,$T1,$T2,$T3         # low part is H.lo^H.hi...
1064         vmovdqu         $T3,-0x10($Htbl)        # save Karatsuba "salt"
1065 ___
1066         &clmul64x64_avx ($Xhi,$Xi,$Hkey,$HK);   # calculate H^3,5,7
1067         &reduction_avx  ($Xhi,$Xi);
1068 $code.=<<___;
1069 .Linit_start_avx:
1070         vmovdqa         $Xi,$T3
1071 ___
1072         &clmul64x64_avx ($Xhi,$Xi,$Hkey,$HK);   # calculate H^2,4,6,8
1073         &reduction_avx  ($Xhi,$Xi);
1074 $code.=<<___;
1075         vpshufd         \$0b01001110,$T3,$T1
1076         vpshufd         \$0b01001110,$Xi,$T2
1077         vpxor           $T3,$T1,$T1             # Karatsuba pre-processing
1078         vmovdqu         $T3,0x00($Htbl)         # save H^1,3,5,7
1079         vpxor           $Xi,$T2,$T2             # Karatsuba pre-processing
1080         vmovdqu         $Xi,0x10($Htbl)         # save H^2,4,6,8
1081         lea             0x30($Htbl),$Htbl
1082         sub             \$1,%r10
1083         jnz             .Linit_loop_avx
1084
1085         vpalignr        \$8,$T2,$T1,$T3         # last "salt" is flipped
1086         vmovdqu         $T3,-0x10($Htbl)
1087
1088         vzeroupper
1089 ___
1090 $code.=<<___ if ($win64);
1091         movaps  (%rsp),%xmm6
1092         lea     0x18(%rsp),%rsp
1093 .LSEH_end_gcm_init_avx:
1094 ___
1095 $code.=<<___;
1096         ret
1097 .size   gcm_init_avx,.-gcm_init_avx
1098 ___
1099 } else {
1100 $code.=<<___;
1101         jmp     .L_init_clmul
1102 .size   gcm_init_avx,.-gcm_init_avx
1103 ___
1104 }
1105
1106 $code.=<<___;
1107 .globl  gcm_gmult_avx
1108 .type   gcm_gmult_avx,\@abi-omnipotent
1109 .align  32
1110 gcm_gmult_avx:
1111         jmp     .L_gmult_clmul
1112 .size   gcm_gmult_avx,.-gcm_gmult_avx
1113 ___
1114 \f
1115 $code.=<<___;
1116 .globl  gcm_ghash_avx
1117 .type   gcm_ghash_avx,\@abi-omnipotent
1118 .align  32
1119 gcm_ghash_avx:
1120 ___
1121 if ($avx) {
1122 my ($Xip,$Htbl,$inp,$len)=@_4args;
1123 my ($Xlo,$Xhi,$Xmi,
1124     $Zlo,$Zhi,$Zmi,
1125     $Hkey,$HK,$T1,$T2,
1126     $Xi,$Xo,$Tred,$bswap,$Ii,$Ij) = map("%xmm$_",(0..15));
1127
1128 $code.=<<___ if ($win64);
1129         lea     -0x88(%rsp),%rax
1130 .LSEH_begin_gcm_ghash_avx:
1131         # I can't trust assembler to use specific encoding:-(
1132         .byte   0x48,0x8d,0x60,0xe0             #lea    -0x20(%rax),%rsp
1133         .byte   0x0f,0x29,0x70,0xe0             #movaps %xmm6,-0x20(%rax)
1134         .byte   0x0f,0x29,0x78,0xf0             #movaps %xmm7,-0x10(%rax)
1135         .byte   0x44,0x0f,0x29,0x00             #movaps %xmm8,0(%rax)
1136         .byte   0x44,0x0f,0x29,0x48,0x10        #movaps %xmm9,0x10(%rax)
1137         .byte   0x44,0x0f,0x29,0x50,0x20        #movaps %xmm10,0x20(%rax)
1138         .byte   0x44,0x0f,0x29,0x58,0x30        #movaps %xmm11,0x30(%rax)
1139         .byte   0x44,0x0f,0x29,0x60,0x40        #movaps %xmm12,0x40(%rax)
1140         .byte   0x44,0x0f,0x29,0x68,0x50        #movaps %xmm13,0x50(%rax)
1141         .byte   0x44,0x0f,0x29,0x70,0x60        #movaps %xmm14,0x60(%rax)
1142         .byte   0x44,0x0f,0x29,0x78,0x70        #movaps %xmm15,0x70(%rax)
1143 ___
1144 $code.=<<___;
1145         vzeroupper
1146
1147         vmovdqu         ($Xip),$Xi              # load $Xi
1148         lea             .L0x1c2_polynomial(%rip),%r10
1149         lea             0x40($Htbl),$Htbl       # size optimization
1150         vmovdqu         .Lbswap_mask(%rip),$bswap
1151         vpshufb         $bswap,$Xi,$Xi
1152         cmp             \$0x80,$len
1153         jb              .Lshort_avx
1154         sub             \$0x80,$len
1155
1156         vmovdqu         0x70($inp),$Ii          # I[7]
1157         vmovdqu         0x00-0x40($Htbl),$Hkey  # $Hkey^1
1158         vpshufb         $bswap,$Ii,$Ii
1159         vmovdqu         0x20-0x40($Htbl),$HK
1160
1161         vpunpckhqdq     $Ii,$Ii,$T2
1162          vmovdqu        0x60($inp),$Ij          # I[6]
1163         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1164         vpxor           $Ii,$T2,$T2
1165          vpshufb        $bswap,$Ij,$Ij
1166         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1167          vmovdqu        0x10-0x40($Htbl),$Hkey  # $Hkey^2
1168          vpunpckhqdq    $Ij,$Ij,$T1
1169          vmovdqu        0x50($inp),$Ii          # I[5]
1170         vpclmulqdq      \$0x00,$HK,$T2,$Xmi
1171          vpxor          $Ij,$T1,$T1
1172
1173          vpshufb        $bswap,$Ii,$Ii
1174         vpclmulqdq      \$0x00,$Hkey,$Ij,$Zlo
1175          vpunpckhqdq    $Ii,$Ii,$T2
1176         vpclmulqdq      \$0x11,$Hkey,$Ij,$Zhi
1177          vmovdqu        0x30-0x40($Htbl),$Hkey  # $Hkey^3
1178          vpxor          $Ii,$T2,$T2
1179          vmovdqu        0x40($inp),$Ij          # I[4]
1180         vpclmulqdq      \$0x10,$HK,$T1,$Zmi
1181          vmovdqu        0x50-0x40($Htbl),$HK
1182
1183          vpshufb        $bswap,$Ij,$Ij
1184         vpxor           $Xlo,$Zlo,$Zlo
1185         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1186         vpxor           $Xhi,$Zhi,$Zhi
1187          vpunpckhqdq    $Ij,$Ij,$T1
1188         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1189          vmovdqu        0x40-0x40($Htbl),$Hkey  # $Hkey^4
1190         vpxor           $Xmi,$Zmi,$Zmi
1191         vpclmulqdq      \$0x00,$HK,$T2,$Xmi
1192          vpxor          $Ij,$T1,$T1
1193
1194          vmovdqu        0x30($inp),$Ii          # I[3]
1195         vpxor           $Zlo,$Xlo,$Xlo
1196         vpclmulqdq      \$0x00,$Hkey,$Ij,$Zlo
1197         vpxor           $Zhi,$Xhi,$Xhi
1198          vpshufb        $bswap,$Ii,$Ii
1199         vpclmulqdq      \$0x11,$Hkey,$Ij,$Zhi
1200          vmovdqu        0x60-0x40($Htbl),$Hkey  # $Hkey^5
1201         vpxor           $Zmi,$Xmi,$Xmi
1202          vpunpckhqdq    $Ii,$Ii,$T2
1203         vpclmulqdq      \$0x10,$HK,$T1,$Zmi
1204          vmovdqu        0x80-0x40($Htbl),$HK
1205          vpxor          $Ii,$T2,$T2
1206
1207          vmovdqu        0x20($inp),$Ij          # I[2]
1208         vpxor           $Xlo,$Zlo,$Zlo
1209         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1210         vpxor           $Xhi,$Zhi,$Zhi
1211          vpshufb        $bswap,$Ij,$Ij
1212         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1213          vmovdqu        0x70-0x40($Htbl),$Hkey  # $Hkey^6
1214         vpxor           $Xmi,$Zmi,$Zmi
1215          vpunpckhqdq    $Ij,$Ij,$T1
1216         vpclmulqdq      \$0x00,$HK,$T2,$Xmi
1217          vpxor          $Ij,$T1,$T1
1218
1219          vmovdqu        0x10($inp),$Ii          # I[1]
1220         vpxor           $Zlo,$Xlo,$Xlo
1221         vpclmulqdq      \$0x00,$Hkey,$Ij,$Zlo
1222         vpxor           $Zhi,$Xhi,$Xhi
1223          vpshufb        $bswap,$Ii,$Ii
1224         vpclmulqdq      \$0x11,$Hkey,$Ij,$Zhi
1225          vmovdqu        0x90-0x40($Htbl),$Hkey  # $Hkey^7
1226         vpxor           $Zmi,$Xmi,$Xmi
1227          vpunpckhqdq    $Ii,$Ii,$T2
1228         vpclmulqdq      \$0x10,$HK,$T1,$Zmi
1229          vmovdqu        0xb0-0x40($Htbl),$HK
1230          vpxor          $Ii,$T2,$T2
1231
1232          vmovdqu        ($inp),$Ij              # I[0]
1233         vpxor           $Xlo,$Zlo,$Zlo
1234         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1235         vpxor           $Xhi,$Zhi,$Zhi
1236          vpshufb        $bswap,$Ij,$Ij
1237         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1238          vmovdqu        0xa0-0x40($Htbl),$Hkey  # $Hkey^8
1239         vpxor           $Xmi,$Zmi,$Zmi
1240         vpclmulqdq      \$0x10,$HK,$T2,$Xmi
1241
1242         lea             0x80($inp),$inp
1243         cmp             \$0x80,$len
1244         jb              .Ltail_avx
1245
1246         vpxor           $Xi,$Ij,$Ij             # accumulate $Xi
1247         sub             \$0x80,$len
1248         jmp             .Loop8x_avx
1249
1250 .align  32
1251 .Loop8x_avx:
1252         vpunpckhqdq     $Ij,$Ij,$T1
1253          vmovdqu        0x70($inp),$Ii          # I[7]
1254         vpxor           $Xlo,$Zlo,$Zlo
1255         vpxor           $Ij,$T1,$T1
1256         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xi
1257          vpshufb        $bswap,$Ii,$Ii
1258         vpxor           $Xhi,$Zhi,$Zhi
1259         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xo
1260          vmovdqu        0x00-0x40($Htbl),$Hkey  # $Hkey^1
1261          vpunpckhqdq    $Ii,$Ii,$T2
1262         vpxor           $Xmi,$Zmi,$Zmi
1263         vpclmulqdq      \$0x00,$HK,$T1,$Tred
1264          vmovdqu        0x20-0x40($Htbl),$HK
1265          vpxor          $Ii,$T2,$T2
1266
1267           vmovdqu       0x60($inp),$Ij          # I[6]
1268          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1269         vpxor           $Zlo,$Xi,$Xi            # collect result
1270           vpshufb       $bswap,$Ij,$Ij
1271          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1272         vxorps          $Zhi,$Xo,$Xo
1273           vmovdqu       0x10-0x40($Htbl),$Hkey  # $Hkey^2
1274          vpunpckhqdq    $Ij,$Ij,$T1
1275          vpclmulqdq     \$0x00,$HK,  $T2,$Xmi
1276         vpxor           $Zmi,$Tred,$Tred
1277          vxorps         $Ij,$T1,$T1
1278
1279           vmovdqu       0x50($inp),$Ii          # I[5]
1280         vpxor           $Xi,$Tred,$Tred         # aggregated Karatsuba post-processing
1281          vpclmulqdq     \$0x00,$Hkey,$Ij,$Zlo
1282         vpxor           $Xo,$Tred,$Tred
1283         vpslldq         \$8,$Tred,$T2
1284          vpxor          $Xlo,$Zlo,$Zlo
1285          vpclmulqdq     \$0x11,$Hkey,$Ij,$Zhi
1286         vpsrldq         \$8,$Tred,$Tred
1287         vpxor           $T2, $Xi, $Xi
1288           vmovdqu       0x30-0x40($Htbl),$Hkey  # $Hkey^3
1289           vpshufb       $bswap,$Ii,$Ii
1290         vxorps          $Tred,$Xo, $Xo
1291          vpxor          $Xhi,$Zhi,$Zhi
1292          vpunpckhqdq    $Ii,$Ii,$T2
1293          vpclmulqdq     \$0x10,$HK,  $T1,$Zmi
1294           vmovdqu       0x50-0x40($Htbl),$HK
1295          vpxor          $Ii,$T2,$T2
1296          vpxor          $Xmi,$Zmi,$Zmi
1297
1298           vmovdqu       0x40($inp),$Ij          # I[4]
1299         vpalignr        \$8,$Xi,$Xi,$Tred       # 1st phase
1300          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1301           vpshufb       $bswap,$Ij,$Ij
1302          vpxor          $Zlo,$Xlo,$Xlo
1303          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1304           vmovdqu       0x40-0x40($Htbl),$Hkey  # $Hkey^4
1305          vpunpckhqdq    $Ij,$Ij,$T1
1306          vpxor          $Zhi,$Xhi,$Xhi
1307          vpclmulqdq     \$0x00,$HK,  $T2,$Xmi
1308          vxorps         $Ij,$T1,$T1
1309          vpxor          $Zmi,$Xmi,$Xmi
1310
1311           vmovdqu       0x30($inp),$Ii          # I[3]
1312         vpclmulqdq      \$0x10,(%r10),$Xi,$Xi
1313          vpclmulqdq     \$0x00,$Hkey,$Ij,$Zlo
1314           vpshufb       $bswap,$Ii,$Ii
1315          vpxor          $Xlo,$Zlo,$Zlo
1316          vpclmulqdq     \$0x11,$Hkey,$Ij,$Zhi
1317           vmovdqu       0x60-0x40($Htbl),$Hkey  # $Hkey^5
1318          vpunpckhqdq    $Ii,$Ii,$T2
1319          vpxor          $Xhi,$Zhi,$Zhi
1320          vpclmulqdq     \$0x10,$HK,  $T1,$Zmi
1321           vmovdqu       0x80-0x40($Htbl),$HK
1322          vpxor          $Ii,$T2,$T2
1323          vpxor          $Xmi,$Zmi,$Zmi
1324
1325           vmovdqu       0x20($inp),$Ij          # I[2]
1326          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1327           vpshufb       $bswap,$Ij,$Ij
1328          vpxor          $Zlo,$Xlo,$Xlo
1329          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1330           vmovdqu       0x70-0x40($Htbl),$Hkey  # $Hkey^6
1331          vpunpckhqdq    $Ij,$Ij,$T1
1332          vpxor          $Zhi,$Xhi,$Xhi
1333          vpclmulqdq     \$0x00,$HK,  $T2,$Xmi
1334          vpxor          $Ij,$T1,$T1
1335          vpxor          $Zmi,$Xmi,$Xmi
1336         vxorps          $Tred,$Xi,$Xi
1337
1338           vmovdqu       0x10($inp),$Ii          # I[1]
1339         vpalignr        \$8,$Xi,$Xi,$Tred       # 2nd phase
1340          vpclmulqdq     \$0x00,$Hkey,$Ij,$Zlo
1341           vpshufb       $bswap,$Ii,$Ii
1342          vpxor          $Xlo,$Zlo,$Zlo
1343          vpclmulqdq     \$0x11,$Hkey,$Ij,$Zhi
1344           vmovdqu       0x90-0x40($Htbl),$Hkey  # $Hkey^7
1345         vpclmulqdq      \$0x10,(%r10),$Xi,$Xi
1346         vxorps          $Xo,$Tred,$Tred
1347          vpunpckhqdq    $Ii,$Ii,$T2
1348          vpxor          $Xhi,$Zhi,$Zhi
1349          vpclmulqdq     \$0x10,$HK,  $T1,$Zmi
1350           vmovdqu       0xb0-0x40($Htbl),$HK
1351          vpxor          $Ii,$T2,$T2
1352          vpxor          $Xmi,$Zmi,$Zmi
1353
1354           vmovdqu       ($inp),$Ij              # I[0]
1355          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1356           vpshufb       $bswap,$Ij,$Ij
1357          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1358           vmovdqu       0xa0-0x40($Htbl),$Hkey  # $Hkey^8
1359         vpxor           $Tred,$Ij,$Ij
1360          vpclmulqdq     \$0x10,$HK,  $T2,$Xmi
1361         vpxor           $Xi,$Ij,$Ij             # accumulate $Xi
1362
1363         lea             0x80($inp),$inp
1364         sub             \$0x80,$len
1365         jnc             .Loop8x_avx
1366
1367         add             \$0x80,$len
1368         jmp             .Ltail_no_xor_avx
1369
1370 .align  32
1371 .Lshort_avx:
1372         vmovdqu         -0x10($inp,$len),$Ii    # very last word
1373         lea             ($inp,$len),$inp
1374         vmovdqu         0x00-0x40($Htbl),$Hkey  # $Hkey^1
1375         vmovdqu         0x20-0x40($Htbl),$HK
1376         vpshufb         $bswap,$Ii,$Ij
1377
1378         vmovdqa         $Xlo,$Zlo               # subtle way to zero $Zlo,
1379         vmovdqa         $Xhi,$Zhi               # $Zhi and
1380         vmovdqa         $Xmi,$Zmi               # $Zmi
1381         sub             \$0x10,$len
1382         jz              .Ltail_avx
1383
1384         vpunpckhqdq     $Ij,$Ij,$T1
1385         vpxor           $Xlo,$Zlo,$Zlo
1386         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1387         vpxor           $Ij,$T1,$T1
1388          vmovdqu        -0x20($inp),$Ii
1389         vpxor           $Xhi,$Zhi,$Zhi
1390         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1391         vmovdqu         0x10-0x40($Htbl),$Hkey  # $Hkey^2
1392          vpshufb        $bswap,$Ii,$Ij
1393         vpxor           $Xmi,$Zmi,$Zmi
1394         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1395         vpsrldq         \$8,$HK,$HK
1396         sub             \$0x10,$len
1397         jz              .Ltail_avx
1398
1399         vpunpckhqdq     $Ij,$Ij,$T1
1400         vpxor           $Xlo,$Zlo,$Zlo
1401         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1402         vpxor           $Ij,$T1,$T1
1403          vmovdqu        -0x30($inp),$Ii
1404         vpxor           $Xhi,$Zhi,$Zhi
1405         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1406         vmovdqu         0x30-0x40($Htbl),$Hkey  # $Hkey^3
1407          vpshufb        $bswap,$Ii,$Ij
1408         vpxor           $Xmi,$Zmi,$Zmi
1409         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1410         vmovdqu         0x50-0x40($Htbl),$HK
1411         sub             \$0x10,$len
1412         jz              .Ltail_avx
1413
1414         vpunpckhqdq     $Ij,$Ij,$T1
1415         vpxor           $Xlo,$Zlo,$Zlo
1416         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1417         vpxor           $Ij,$T1,$T1
1418          vmovdqu        -0x40($inp),$Ii
1419         vpxor           $Xhi,$Zhi,$Zhi
1420         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1421         vmovdqu         0x40-0x40($Htbl),$Hkey  # $Hkey^4
1422          vpshufb        $bswap,$Ii,$Ij
1423         vpxor           $Xmi,$Zmi,$Zmi
1424         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1425         vpsrldq         \$8,$HK,$HK
1426         sub             \$0x10,$len
1427         jz              .Ltail_avx
1428
1429         vpunpckhqdq     $Ij,$Ij,$T1
1430         vpxor           $Xlo,$Zlo,$Zlo
1431         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1432         vpxor           $Ij,$T1,$T1
1433          vmovdqu        -0x50($inp),$Ii
1434         vpxor           $Xhi,$Zhi,$Zhi
1435         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1436         vmovdqu         0x60-0x40($Htbl),$Hkey  # $Hkey^5
1437          vpshufb        $bswap,$Ii,$Ij
1438         vpxor           $Xmi,$Zmi,$Zmi
1439         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1440         vmovdqu         0x80-0x40($Htbl),$HK
1441         sub             \$0x10,$len
1442         jz              .Ltail_avx
1443
1444         vpunpckhqdq     $Ij,$Ij,$T1
1445         vpxor           $Xlo,$Zlo,$Zlo
1446         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1447         vpxor           $Ij,$T1,$T1
1448          vmovdqu        -0x60($inp),$Ii
1449         vpxor           $Xhi,$Zhi,$Zhi
1450         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1451         vmovdqu         0x70-0x40($Htbl),$Hkey  # $Hkey^6
1452          vpshufb        $bswap,$Ii,$Ij
1453         vpxor           $Xmi,$Zmi,$Zmi
1454         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1455         vpsrldq         \$8,$HK,$HK
1456         sub             \$0x10,$len
1457         jz              .Ltail_avx
1458
1459         vpunpckhqdq     $Ij,$Ij,$T1
1460         vpxor           $Xlo,$Zlo,$Zlo
1461         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1462         vpxor           $Ij,$T1,$T1
1463          vmovdqu        -0x70($inp),$Ii
1464         vpxor           $Xhi,$Zhi,$Zhi
1465         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1466         vmovdqu         0x90-0x40($Htbl),$Hkey  # $Hkey^7
1467          vpshufb        $bswap,$Ii,$Ij
1468         vpxor           $Xmi,$Zmi,$Zmi
1469         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1470         vmovq           0xb8-0x40($Htbl),$HK
1471         sub             \$0x10,$len
1472         jmp             .Ltail_avx
1473
1474 .align  32
1475 .Ltail_avx:
1476         vpxor           $Xi,$Ij,$Ij             # accumulate $Xi
1477 .Ltail_no_xor_avx:
1478         vpunpckhqdq     $Ij,$Ij,$T1
1479         vpxor           $Xlo,$Zlo,$Zlo
1480         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1481         vpxor           $Ij,$T1,$T1
1482         vpxor           $Xhi,$Zhi,$Zhi
1483         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1484         vpxor           $Xmi,$Zmi,$Zmi
1485         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1486
1487         vmovdqu         (%r10),$Tred
1488
1489         vpxor           $Xlo,$Zlo,$Xi
1490         vpxor           $Xhi,$Zhi,$Xo
1491         vpxor           $Xmi,$Zmi,$Zmi
1492
1493         vpxor           $Xi, $Zmi,$Zmi          # aggregated Karatsuba post-processing
1494         vpxor           $Xo, $Zmi,$Zmi
1495         vpslldq         \$8, $Zmi,$T2
1496         vpsrldq         \$8, $Zmi,$Zmi
1497         vpxor           $T2, $Xi, $Xi
1498         vpxor           $Zmi,$Xo, $Xo
1499
1500         vpclmulqdq      \$0x10,$Tred,$Xi,$T2    # 1st phase
1501         vpalignr        \$8,$Xi,$Xi,$Xi
1502         vpxor           $T2,$Xi,$Xi
1503
1504         vpclmulqdq      \$0x10,$Tred,$Xi,$T2    # 2nd phase
1505         vpalignr        \$8,$Xi,$Xi,$Xi
1506         vpxor           $Xo,$Xi,$Xi
1507         vpxor           $T2,$Xi,$Xi
1508
1509         cmp             \$0,$len
1510         jne             .Lshort_avx
1511
1512         vpshufb         $bswap,$Xi,$Xi
1513         vmovdqu         $Xi,($Xip)
1514         vzeroupper
1515 ___
1516 $code.=<<___ if ($win64);
1517         movaps  (%rsp),%xmm6
1518         movaps  0x10(%rsp),%xmm7
1519         movaps  0x20(%rsp),%xmm8
1520         movaps  0x30(%rsp),%xmm9
1521         movaps  0x40(%rsp),%xmm10
1522         movaps  0x50(%rsp),%xmm11
1523         movaps  0x60(%rsp),%xmm12
1524         movaps  0x70(%rsp),%xmm13
1525         movaps  0x80(%rsp),%xmm14
1526         movaps  0x90(%rsp),%xmm15
1527         lea     0xa8(%rsp),%rsp
1528 .LSEH_end_gcm_ghash_avx:
1529 ___
1530 $code.=<<___;
1531         ret
1532 .size   gcm_ghash_avx,.-gcm_ghash_avx
1533 ___
1534 } else {
1535 $code.=<<___;
1536         jmp     .L_ghash_clmul
1537 .size   gcm_ghash_avx,.-gcm_ghash_avx
1538 ___
1539 }
1540 \f
1541 $code.=<<___;
1542 .align  64
1543 .Lbswap_mask:
1544         .byte   15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
1545 .L0x1c2_polynomial:
1546         .byte   1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
1547 .L7_mask:
1548         .long   7,0,7,0
1549 .L7_mask_poly:
1550         .long   7,0,`0xE1<<1`,0
1551 .align  64
1552 .type   .Lrem_4bit,\@object
1553 .Lrem_4bit:
1554         .long   0,`0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`
1555         .long   0,`0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`
1556         .long   0,`0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`
1557         .long   0,`0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`
1558 .type   .Lrem_8bit,\@object
1559 .Lrem_8bit:
1560         .value  0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E
1561         .value  0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E
1562         .value  0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E
1563         .value  0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E
1564         .value  0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E
1565         .value  0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E
1566         .value  0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E
1567         .value  0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E
1568         .value  0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE
1569         .value  0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE
1570         .value  0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE
1571         .value  0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE
1572         .value  0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E
1573         .value  0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E
1574         .value  0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE
1575         .value  0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE
1576         .value  0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E
1577         .value  0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E
1578         .value  0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E
1579         .value  0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E
1580         .value  0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E
1581         .value  0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E
1582         .value  0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E
1583         .value  0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E
1584         .value  0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE
1585         .value  0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE
1586         .value  0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE
1587         .value  0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE
1588         .value  0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E
1589         .value  0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E
1590         .value  0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE
1591         .value  0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE
1592
1593 .asciz  "GHASH for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1594 .align  64
1595 ___
1596 \f
1597 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1598 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
1599 if ($win64) {
1600 $rec="%rcx";
1601 $frame="%rdx";
1602 $context="%r8";
1603 $disp="%r9";
1604
1605 $code.=<<___;
1606 .extern __imp_RtlVirtualUnwind
1607 .type   se_handler,\@abi-omnipotent
1608 .align  16
1609 se_handler:
1610         push    %rsi
1611         push    %rdi
1612         push    %rbx
1613         push    %rbp
1614         push    %r12
1615         push    %r13
1616         push    %r14
1617         push    %r15
1618         pushfq
1619         sub     \$64,%rsp
1620
1621         mov     120($context),%rax      # pull context->Rax
1622         mov     248($context),%rbx      # pull context->Rip
1623
1624         mov     8($disp),%rsi           # disp->ImageBase
1625         mov     56($disp),%r11          # disp->HandlerData
1626
1627         mov     0(%r11),%r10d           # HandlerData[0]
1628         lea     (%rsi,%r10),%r10        # prologue label
1629         cmp     %r10,%rbx               # context->Rip<prologue label
1630         jb      .Lin_prologue
1631
1632         mov     152($context),%rax      # pull context->Rsp
1633
1634         mov     4(%r11),%r10d           # HandlerData[1]
1635         lea     (%rsi,%r10),%r10        # epilogue label
1636         cmp     %r10,%rbx               # context->Rip>=epilogue label
1637         jae     .Lin_prologue
1638
1639         lea     24(%rax),%rax           # adjust "rsp"
1640
1641         mov     -8(%rax),%rbx
1642         mov     -16(%rax),%rbp
1643         mov     -24(%rax),%r12
1644         mov     %rbx,144($context)      # restore context->Rbx
1645         mov     %rbp,160($context)      # restore context->Rbp
1646         mov     %r12,216($context)      # restore context->R12
1647
1648 .Lin_prologue:
1649         mov     8(%rax),%rdi
1650         mov     16(%rax),%rsi
1651         mov     %rax,152($context)      # restore context->Rsp
1652         mov     %rsi,168($context)      # restore context->Rsi
1653         mov     %rdi,176($context)      # restore context->Rdi
1654
1655         mov     40($disp),%rdi          # disp->ContextRecord
1656         mov     $context,%rsi           # context
1657         mov     \$`1232/8`,%ecx         # sizeof(CONTEXT)
1658         .long   0xa548f3fc              # cld; rep movsq
1659
1660         mov     $disp,%rsi
1661         xor     %rcx,%rcx               # arg1, UNW_FLAG_NHANDLER
1662         mov     8(%rsi),%rdx            # arg2, disp->ImageBase
1663         mov     0(%rsi),%r8             # arg3, disp->ControlPc
1664         mov     16(%rsi),%r9            # arg4, disp->FunctionEntry
1665         mov     40(%rsi),%r10           # disp->ContextRecord
1666         lea     56(%rsi),%r11           # &disp->HandlerData
1667         lea     24(%rsi),%r12           # &disp->EstablisherFrame
1668         mov     %r10,32(%rsp)           # arg5
1669         mov     %r11,40(%rsp)           # arg6
1670         mov     %r12,48(%rsp)           # arg7
1671         mov     %rcx,56(%rsp)           # arg8, (NULL)
1672         call    *__imp_RtlVirtualUnwind(%rip)
1673
1674         mov     \$1,%eax                # ExceptionContinueSearch
1675         add     \$64,%rsp
1676         popfq
1677         pop     %r15
1678         pop     %r14
1679         pop     %r13
1680         pop     %r12
1681         pop     %rbp
1682         pop     %rbx
1683         pop     %rdi
1684         pop     %rsi
1685         ret
1686 .size   se_handler,.-se_handler
1687
1688 .section        .pdata
1689 .align  4
1690         .rva    .LSEH_begin_gcm_gmult_4bit
1691         .rva    .LSEH_end_gcm_gmult_4bit
1692         .rva    .LSEH_info_gcm_gmult_4bit
1693
1694         .rva    .LSEH_begin_gcm_ghash_4bit
1695         .rva    .LSEH_end_gcm_ghash_4bit
1696         .rva    .LSEH_info_gcm_ghash_4bit
1697
1698         .rva    .LSEH_begin_gcm_init_clmul
1699         .rva    .LSEH_end_gcm_init_clmul
1700         .rva    .LSEH_info_gcm_init_clmul
1701
1702         .rva    .LSEH_begin_gcm_ghash_clmul
1703         .rva    .LSEH_end_gcm_ghash_clmul
1704         .rva    .LSEH_info_gcm_ghash_clmul
1705 ___
1706 $code.=<<___    if ($avx);
1707         .rva    .LSEH_begin_gcm_init_avx
1708         .rva    .LSEH_end_gcm_init_avx
1709         .rva    .LSEH_info_gcm_init_clmul
1710
1711         .rva    .LSEH_begin_gcm_ghash_avx
1712         .rva    .LSEH_end_gcm_ghash_avx
1713         .rva    .LSEH_info_gcm_ghash_clmul
1714 ___
1715 $code.=<<___;
1716 .section        .xdata
1717 .align  8
1718 .LSEH_info_gcm_gmult_4bit:
1719         .byte   9,0,0,0
1720         .rva    se_handler
1721         .rva    .Lgmult_prologue,.Lgmult_epilogue       # HandlerData
1722 .LSEH_info_gcm_ghash_4bit:
1723         .byte   9,0,0,0
1724         .rva    se_handler
1725         .rva    .Lghash_prologue,.Lghash_epilogue       # HandlerData
1726 .LSEH_info_gcm_init_clmul:
1727         .byte   0x01,0x08,0x03,0x00
1728         .byte   0x08,0x68,0x00,0x00     #movaps 0x00(rsp),xmm6
1729         .byte   0x04,0x22,0x00,0x00     #sub    rsp,0x18
1730 .LSEH_info_gcm_ghash_clmul:
1731         .byte   0x01,0x33,0x16,0x00
1732         .byte   0x33,0xf8,0x09,0x00     #movaps 0x90(rsp),xmm15
1733         .byte   0x2e,0xe8,0x08,0x00     #movaps 0x80(rsp),xmm14
1734         .byte   0x29,0xd8,0x07,0x00     #movaps 0x70(rsp),xmm13
1735         .byte   0x24,0xc8,0x06,0x00     #movaps 0x60(rsp),xmm12
1736         .byte   0x1f,0xb8,0x05,0x00     #movaps 0x50(rsp),xmm11
1737         .byte   0x1a,0xa8,0x04,0x00     #movaps 0x40(rsp),xmm10
1738         .byte   0x15,0x98,0x03,0x00     #movaps 0x30(rsp),xmm9
1739         .byte   0x10,0x88,0x02,0x00     #movaps 0x20(rsp),xmm8
1740         .byte   0x0c,0x78,0x01,0x00     #movaps 0x10(rsp),xmm7
1741         .byte   0x08,0x68,0x00,0x00     #movaps 0x00(rsp),xmm6
1742         .byte   0x04,0x01,0x15,0x00     #sub    rsp,0xa8
1743 ___
1744 }
1745 \f
1746 $code =~ s/\`([^\`]*)\`/eval($1)/gem;
1747
1748 print $code;
1749
1750 close STDOUT;