crypto/modes/asm/ghash-x86_64.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2010-2016 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the OpenSSL license (the "License").  You may not use
   5 # this file except in compliance with the License.  You can obtain a copy
   6 # in the file LICENSE in the source distribution or at
   7 # https://www.openssl.org/source/license.html
   8
   9 #
  10 # ====================================================================
  11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
  12 # project. The module is, however, dual licensed under OpenSSL and
  13 # CRYPTOGAMS licenses depending on where you obtain it. For further
  14 # details see http://www.openssl.org/~appro/cryptogams/.
  15 # ====================================================================
  16 #
  17 # March, June 2010
  18 #
  19 # The module implements "4-bit" GCM GHASH function and underlying
  20 # single multiplication operation in GF(2^128). "4-bit" means that
  21 # it uses 256 bytes per-key table [+128 bytes shared table]. GHASH
  22 # function features so called "528B" variant utilizing additional
  23 # 256+16 bytes of per-key storage [+512 bytes shared table].
  24 # Performance results are for this streamed GHASH subroutine and are
  25 # expressed in cycles per processed byte, less is better:
  26 #
  27 #               gcc 3.4.x(*)    assembler
  28 #
  29 # P4            28.6            14.0            +100%
  30 # Opteron       19.3            7.7             +150%
  31 # Core2         17.8            8.1(**)         +120%
  32 # Atom          31.6            16.8            +88%
  33 # VIA Nano      21.8            10.1            +115%
  34 #
  35 # (*)   comparison is not completely fair, because C results are
  36 #       for vanilla "256B" implementation, while assembler results
  37 #       are for "528B";-)
  38 # (**)  it's mystery [to me] why Core2 result is not same as for
  39 #       Opteron;
  40
  41 # May 2010
  42 #
  43 # Add PCLMULQDQ version performing at 2.02 cycles per processed byte.
  44 # See ghash-x86.pl for background information and details about coding
  45 # techniques.
  46 #
  47 # Special thanks to David Woodhouse <dwmw2@infradead.org> for
  48 # providing access to a Westmere-based system on behalf of Intel
  49 # Open Source Technology Centre.
  50
  51 # December 2012
  52 #
  53 # Overhaul: aggregate Karatsuba post-processing, improve ILP in
  54 # reduction_alg9, increase reduction aggregate factor to 4x. As for
  55 # the latter. ghash-x86.pl discusses that it makes lesser sense to
  56 # increase aggregate factor. Then why increase here? Critical path
  57 # consists of 3 independent pclmulqdq instructions, Karatsuba post-
  58 # processing and reduction. "On top" of this we lay down aggregated
  59 # multiplication operations, triplets of independent pclmulqdq's. As
  60 # issue rate for pclmulqdq is limited, it makes lesser sense to
  61 # aggregate more multiplications than it takes to perform remaining
  62 # non-multiplication operations. 2x is near-optimal coefficient for
  63 # contemporary Intel CPUs (therefore modest improvement coefficient),
  64 # but not for Bulldozer. Latter is because logical SIMD operations
  65 # are twice as slow in comparison to Intel, so that critical path is
  66 # longer. A CPU with higher pclmulqdq issue rate would also benefit
  67 # from higher aggregate factor...
  68 #
  69 # Westmere      1.78(+13%)
  70 # Sandy Bridge  1.80(+8%)
  71 # Ivy Bridge    1.80(+7%)
  72 # Haswell       0.55(+93%) (if system doesn't support AVX)
  73 # Broadwell     0.45(+110%)(if system doesn't support AVX)
  74 # Skylake       0.44(+110%)(if system doesn't support AVX)
  75 # Bulldozer     1.49(+27%)
  76 # Silvermont    2.88(+13%)
  77 # Goldmont      1.08(+24%)
  78
  79 # March 2013
  80 #
  81 # ... 8x aggregate factor AVX code path is using reduction algorithm
  82 # suggested by Shay Gueron[1]. Even though contemporary AVX-capable
  83 # CPUs such as Sandy and Ivy Bridge can execute it, the code performs
  84 # sub-optimally in comparison to above mentioned version. But thanks
  85 # to Ilya Albrekht and Max Locktyukhin of Intel Corp. we knew that
  86 # it performs in 0.41 cycles per byte on Haswell processor, in
  87 # 0.29 on Broadwell, and in 0.36 on Skylake.
  88 #
  89 # [1] http://rt.openssl.org/Ticket/Display.html?id=2900&user=guest&pass=guest
  90
  91 $flavour = shift;
  92 $output  = shift;
  93 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  94
  95 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
  96
  97 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  98 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  99 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
 100 die "can't locate x86_64-xlate.pl";
 101
 102 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
 103                 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
 104         $avx = ($1>=2.20) + ($1>=2.22);
 105 }
 106
 107 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
 108             `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
 109         $avx = ($1>=2.09) + ($1>=2.10);
 110 }
 111
 112 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
 113             `ml64 2>&1` =~ /Version ([0-9]+)\./) {
 114         $avx = ($1>=10) + ($1>=11);
 115 }
 116
 117 if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([3-9]\.[0-9]+)/) {
 118         $avx = ($2>=3.0) + ($2>3.0);
 119 }
 120
 121 open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"";
 122 *STDOUT=*OUT;
 123
 124 $do4xaggr=1;
 125
 126 # common register layout
 127 $nlo="%rax";
 128 $nhi="%rbx";
 129 $Zlo="%r8";
 130 $Zhi="%r9";
 131 $tmp="%r10";
 132 $rem_4bit = "%r11";
 133
 134 $Xi="%rdi";
 135 $Htbl="%rsi";
 136
 137 # per-function register layout
 138 $cnt="%rcx";
 139 $rem="%rdx";
 140
 141 sub LB() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/     or
 142                         $r =~ s/%[er]([sd]i)/%\1l/      or
 143                         $r =~ s/%[er](bp)/%\1l/         or
 144                         $r =~ s/%(r[0-9]+)[d]?/%\1b/;   $r; }
 145
 146 sub AUTOLOAD()          # thunk [simplified] 32-bit style perlasm
 147 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
 148   my $arg = pop;
 149     $arg = "\$$arg" if ($arg*1 eq $arg);
 150     $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
 151 }
 152 \f
 153 { my $N;
 154   sub loop() {
 155   my $inp = shift;
 156
 157         $N++;
 158 $code.=<<___;
 159         xor     $nlo,$nlo
 160         xor     $nhi,$nhi
 161         mov     `&LB("$Zlo")`,`&LB("$nlo")`
 162         mov     `&LB("$Zlo")`,`&LB("$nhi")`
 163         shl     \$4,`&LB("$nlo")`
 164         mov     \$14,$cnt
 165         mov     8($Htbl,$nlo),$Zlo
 166         mov     ($Htbl,$nlo),$Zhi
 167         and     \$0xf0,`&LB("$nhi")`
 168         mov     $Zlo,$rem
 169         jmp     .Loop$N
 170
 171 .align  16
 172 .Loop$N:
 173         shr     \$4,$Zlo
 174         and     \$0xf,$rem
 175         mov     $Zhi,$tmp
 176         mov     ($inp,$cnt),`&LB("$nlo")`
 177         shr     \$4,$Zhi
 178         xor     8($Htbl,$nhi),$Zlo
 179         shl     \$60,$tmp
 180         xor     ($Htbl,$nhi),$Zhi
 181         mov     `&LB("$nlo")`,`&LB("$nhi")`
 182         xor     ($rem_4bit,$rem,8),$Zhi
 183         mov     $Zlo,$rem
 184         shl     \$4,`&LB("$nlo")`
 185         xor     $tmp,$Zlo
 186         dec     $cnt
 187         js      .Lbreak$N
 188
 189         shr     \$4,$Zlo
 190         and     \$0xf,$rem
 191         mov     $Zhi,$tmp
 192         shr     \$4,$Zhi
 193         xor     8($Htbl,$nlo),$Zlo
 194         shl     \$60,$tmp
 195         xor     ($Htbl,$nlo),$Zhi
 196         and     \$0xf0,`&LB("$nhi")`
 197         xor     ($rem_4bit,$rem,8),$Zhi
 198         mov     $Zlo,$rem
 199         xor     $tmp,$Zlo
 200         jmp     .Loop$N
 201
 202 .align  16
 203 .Lbreak$N:
 204         shr     \$4,$Zlo
 205         and     \$0xf,$rem
 206         mov     $Zhi,$tmp
 207         shr     \$4,$Zhi
 208         xor     8($Htbl,$nlo),$Zlo
 209         shl     \$60,$tmp
 210         xor     ($Htbl,$nlo),$Zhi
 211         and     \$0xf0,`&LB("$nhi")`
 212         xor     ($rem_4bit,$rem,8),$Zhi
 213         mov     $Zlo,$rem
 214         xor     $tmp,$Zlo
 215
 216         shr     \$4,$Zlo
 217         and     \$0xf,$rem
 218         mov     $Zhi,$tmp
 219         shr     \$4,$Zhi
 220         xor     8($Htbl,$nhi),$Zlo
 221         shl     \$60,$tmp
 222         xor     ($Htbl,$nhi),$Zhi
 223         xor     $tmp,$Zlo
 224         xor     ($rem_4bit,$rem,8),$Zhi
 225
 226         bswap   $Zlo
 227         bswap   $Zhi
 228 ___
 229 }}
 230
 231 $code=<<___;
 232 .text
 233 .extern OPENSSL_ia32cap_P
 234
 235 .globl  gcm_gmult_4bit
 236 .type   gcm_gmult_4bit,\@function,2
 237 .align  16
 238 gcm_gmult_4bit:
 239 .cfi_startproc
 240         push    %rbx
 241 .cfi_push       %rbx
 242         push    %rbp            # %rbp and others are pushed exclusively in
 243 .cfi_push       %rbp
 244         push    %r12            # order to reuse Win64 exception handler...
 245 .cfi_push       %r12
 246         push    %r13
 247 .cfi_push       %r13
 248         push    %r14
 249 .cfi_push       %r14
 250         push    %r15
 251 .cfi_push       %r15
 252         sub     \$280,%rsp
 253 .cfi_adjust_cfa_offset  280
 254 .Lgmult_prologue:
 255
 256         movzb   15($Xi),$Zlo
 257         lea     .Lrem_4bit(%rip),$rem_4bit
 258 ___
 259         &loop   ($Xi);
 260 $code.=<<___;
 261         mov     $Zlo,8($Xi)
 262         mov     $Zhi,($Xi)
 263
 264         lea     280+48(%rsp),%rsi
 265 .cfi_def_cfa    %rsi,8
 266         mov     -8(%rsi),%rbx
 267 .cfi_restore    %rbx
 268         lea     (%rsi),%rsp
 269 .cfi_def_cfa_register   %rsp
 270 .Lgmult_epilogue:
 271         ret
 272 .cfi_endproc
 273 .size   gcm_gmult_4bit,.-gcm_gmult_4bit
 274 ___
 275 \f
 276 # per-function register layout
 277 $inp="%rdx";
 278 $len="%rcx";
 279 $rem_8bit=$rem_4bit;
 280
 281 $code.=<<___;
 282 .globl  gcm_ghash_4bit
 283 .type   gcm_ghash_4bit,\@function,4
 284 .align  16
 285 gcm_ghash_4bit:
 286 .cfi_startproc
 287         push    %rbx
 288 .cfi_push       %rbx
 289         push    %rbp
 290 .cfi_push       %rbp
 291         push    %r12
 292 .cfi_push       %r12
 293         push    %r13
 294 .cfi_push       %r13
 295         push    %r14
 296 .cfi_push       %r14
 297         push    %r15
 298 .cfi_push       %r15
 299         sub     \$280,%rsp
 300 .cfi_adjust_cfa_offset  280
 301 .Lghash_prologue:
 302         mov     $inp,%r14               # reassign couple of args
 303         mov     $len,%r15
 304 ___
 305 { my $inp="%r14";
 306   my $dat="%edx";
 307   my $len="%r15";
 308   my @nhi=("%ebx","%ecx");
 309   my @rem=("%r12","%r13");
 310   my $Hshr4="%rbp";
 311
 312         &sub    ($Htbl,-128);           # size optimization
 313         &lea    ($Hshr4,"16+128(%rsp)");
 314         { my @lo =($nlo,$nhi);
 315           my @hi =($Zlo,$Zhi);
 316
 317           &xor  ($dat,$dat);
 318           for ($i=0,$j=-2;$i<18;$i++,$j++) {
 319             &mov        ("$j(%rsp)",&LB($dat))          if ($i>1);
 320             &or         ($lo[0],$tmp)                   if ($i>1);
 321             &mov        (&LB($dat),&LB($lo[1]))         if ($i>0 && $i<17);
 322             &shr        ($lo[1],4)                      if ($i>0 && $i<17);
 323             &mov        ($tmp,$hi[1])                   if ($i>0 && $i<17);
 324             &shr        ($hi[1],4)                      if ($i>0 && $i<17);
 325             &mov        ("8*$j($Hshr4)",$hi[0])         if ($i>1);
 326             &mov        ($hi[0],"16*$i+0-128($Htbl)")   if ($i<16);
 327             &shl        (&LB($dat),4)                   if ($i>0 && $i<17);
 328             &mov        ("8*$j-128($Hshr4)",$lo[0])     if ($i>1);
 329             &mov        ($lo[0],"16*$i+8-128($Htbl)")   if ($i<16);
 330             &shl        ($tmp,60)                       if ($i>0 && $i<17);
 331
 332             push        (@lo,shift(@lo));
 333             push        (@hi,shift(@hi));
 334           }
 335         }
 336         &add    ($Htbl,-128);
 337         &mov    ($Zlo,"8($Xi)");
 338         &mov    ($Zhi,"0($Xi)");
 339         &add    ($len,$inp);            # pointer to the end of data
 340         &lea    ($rem_8bit,".Lrem_8bit(%rip)");
 341         &jmp    (".Louter_loop");
 342
 343 $code.=".align  16\n.Louter_loop:\n";
 344         &xor    ($Zhi,"($inp)");
 345         &mov    ("%rdx","8($inp)");
 346         &lea    ($inp,"16($inp)");
 347         &xor    ("%rdx",$Zlo);
 348         &mov    ("($Xi)",$Zhi);
 349         &mov    ("8($Xi)","%rdx");
 350         &shr    ("%rdx",32);
 351
 352         &xor    ($nlo,$nlo);
 353         &rol    ($dat,8);
 354         &mov    (&LB($nlo),&LB($dat));
 355         &movz   ($nhi[0],&LB($dat));
 356         &shl    (&LB($nlo),4);
 357         &shr    ($nhi[0],4);
 358
 359         for ($j=11,$i=0;$i<15;$i++) {
 360             &rol        ($dat,8);
 361             &xor        ($Zlo,"8($Htbl,$nlo)")                  if ($i>0);
 362             &xor        ($Zhi,"($Htbl,$nlo)")                   if ($i>0);
 363             &mov        ($Zlo,"8($Htbl,$nlo)")                  if ($i==0);
 364             &mov        ($Zhi,"($Htbl,$nlo)")                   if ($i==0);
 365
 366             &mov        (&LB($nlo),&LB($dat));
 367             &xor        ($Zlo,$tmp)                             if ($i>0);
 368             &movzw      ($rem[1],"($rem_8bit,$rem[1],2)")       if ($i>0);
 369
 370             &movz       ($nhi[1],&LB($dat));
 371             &shl        (&LB($nlo),4);
 372             &movzb      ($rem[0],"(%rsp,$nhi[0])");
 373
 374             &shr        ($nhi[1],4)                             if ($i<14);
 375             &and        ($nhi[1],0xf0)                          if ($i==14);
 376             &shl        ($rem[1],48)                            if ($i>0);
 377             &xor        ($rem[0],$Zlo);
 378
 379             &mov        ($tmp,$Zhi);
 380             &xor        ($Zhi,$rem[1])                          if ($i>0);
 381             &shr        ($Zlo,8);
 382
 383             &movz       ($rem[0],&LB($rem[0]));
 384             &mov        ($dat,"$j($Xi)")                        if (--$j%4==0);
 385             &shr        ($Zhi,8);
 386
 387             &xor        ($Zlo,"-128($Hshr4,$nhi[0],8)");
 388             &shl        ($tmp,56);
 389             &xor        ($Zhi,"($Hshr4,$nhi[0],8)");
 390
 391             unshift     (@nhi,pop(@nhi));               # "rotate" registers
 392             unshift     (@rem,pop(@rem));
 393         }
 394         &movzw  ($rem[1],"($rem_8bit,$rem[1],2)");
 395         &xor    ($Zlo,"8($Htbl,$nlo)");
 396         &xor    ($Zhi,"($Htbl,$nlo)");
 397
 398         &shl    ($rem[1],48);
 399         &xor    ($Zlo,$tmp);
 400
 401         &xor    ($Zhi,$rem[1]);
 402         &movz   ($rem[0],&LB($Zlo));
 403         &shr    ($Zlo,4);
 404
 405         &mov    ($tmp,$Zhi);
 406         &shl    (&LB($rem[0]),4);
 407         &shr    ($Zhi,4);
 408
 409         &xor    ($Zlo,"8($Htbl,$nhi[0])");
 410         &movzw  ($rem[0],"($rem_8bit,$rem[0],2)");
 411         &shl    ($tmp,60);
 412
 413         &xor    ($Zhi,"($Htbl,$nhi[0])");
 414         &xor    ($Zlo,$tmp);
 415         &shl    ($rem[0],48);
 416
 417         &bswap  ($Zlo);
 418         &xor    ($Zhi,$rem[0]);
 419
 420         &bswap  ($Zhi);
 421         &cmp    ($inp,$len);
 422         &jb     (".Louter_loop");
 423 }
 424 $code.=<<___;
 425         mov     $Zlo,8($Xi)
 426         mov     $Zhi,($Xi)
 427
 428         lea     280+48(%rsp),%rsi
 429 .cfi_def_cfa    %rsi,8
 430         mov     -48(%rsi),%r15
 431 .cfi_restore    %r15
 432         mov     -40(%rsi),%r14
 433 .cfi_restore    %r14
 434         mov     -32(%rsi),%r13
 435 .cfi_restore    %r13
 436         mov     -24(%rsi),%r12
 437 .cfi_restore    %r12
 438         mov     -16(%rsi),%rbp
 439 .cfi_restore    %rbp
 440         mov     -8(%rsi),%rbx
 441 .cfi_restore    %rbx
 442         lea     0(%rsi),%rsp
 443 .cfi_def_cfa_register   %rsp
 444 .Lghash_epilogue:
 445         ret
 446 .cfi_endproc
 447 .size   gcm_ghash_4bit,.-gcm_ghash_4bit
 448 ___
 449 \f
 450 ######################################################################
 451 # PCLMULQDQ version.
 452
 453 @_4args=$win64? ("%rcx","%rdx","%r8", "%r9") :  # Win64 order
 454                 ("%rdi","%rsi","%rdx","%rcx");  # Unix order
 455
 456 ($Xi,$Xhi)=("%xmm0","%xmm1");   $Hkey="%xmm2";
 457 ($T1,$T2,$T3)=("%xmm3","%xmm4","%xmm5");
 458
 459 sub clmul64x64_T2 {     # minimal register pressure
 460 my ($Xhi,$Xi,$Hkey,$HK)=@_;
 461
 462 if (!defined($HK)) {    $HK = $T2;
 463 $code.=<<___;
 464         movdqa          $Xi,$Xhi                #
 465         pshufd          \$0b01001110,$Xi,$T1
 466         pshufd          \$0b01001110,$Hkey,$T2
 467         pxor            $Xi,$T1                 #
 468         pxor            $Hkey,$T2
 469 ___
 470 } else {
 471 $code.=<<___;
 472         movdqa          $Xi,$Xhi                #
 473         pshufd          \$0b01001110,$Xi,$T1
 474         pxor            $Xi,$T1                 #
 475 ___
 476 }
 477 $code.=<<___;
 478         pclmulqdq       \$0x00,$Hkey,$Xi        #######
 479         pclmulqdq       \$0x11,$Hkey,$Xhi       #######
 480         pclmulqdq       \$0x00,$HK,$T1          #######
 481         pxor            $Xi,$T1                 #
 482         pxor            $Xhi,$T1                #
 483
 484         movdqa          $T1,$T2                 #
 485         psrldq          \$8,$T1
 486         pslldq          \$8,$T2                 #
 487         pxor            $T1,$Xhi
 488         pxor            $T2,$Xi                 #
 489 ___
 490 }
 491
 492 sub reduction_alg9 {    # 17/11 times faster than Intel version
 493 my ($Xhi,$Xi) = @_;
 494
 495 $code.=<<___;
 496         # 1st phase
 497         movdqa          $Xi,$T2                 #
 498         movdqa          $Xi,$T1
 499         psllq           \$5,$Xi
 500         pxor            $Xi,$T1                 #
 501         psllq           \$1,$Xi
 502         pxor            $T1,$Xi                 #
 503         psllq           \$57,$Xi                #
 504         movdqa          $Xi,$T1                 #
 505         pslldq          \$8,$Xi
 506         psrldq          \$8,$T1                 #
 507         pxor            $T2,$Xi
 508         pxor            $T1,$Xhi                #
 509
 510         # 2nd phase
 511         movdqa          $Xi,$T2
 512         psrlq           \$1,$Xi
 513         pxor            $T2,$Xhi                #
 514         pxor            $Xi,$T2
 515         psrlq           \$5,$Xi
 516         pxor            $T2,$Xi                 #
 517         psrlq           \$1,$Xi                 #
 518         pxor            $Xhi,$Xi                #
 519 ___
 520 }
 521 \f
 522 { my ($Htbl,$Xip)=@_4args;
 523   my $HK="%xmm6";
 524
 525 $code.=<<___;
 526 .globl  gcm_init_clmul
 527 .type   gcm_init_clmul,\@abi-omnipotent
 528 .align  16
 529 gcm_init_clmul:
 530 .L_init_clmul:
 531 ___
 532 $code.=<<___ if ($win64);
 533 .LSEH_begin_gcm_init_clmul:
 534         # I can't trust assembler to use specific encoding:-(
 535         .byte   0x48,0x83,0xec,0x18             #sub    $0x18,%rsp
 536         .byte   0x0f,0x29,0x34,0x24             #movaps %xmm6,(%rsp)
 537 ___
 538 $code.=<<___;
 539         movdqu          ($Xip),$Hkey
 540         pshufd          \$0b01001110,$Hkey,$Hkey        # dword swap
 541
 542         # <<1 twist
 543         pshufd          \$0b11111111,$Hkey,$T2  # broadcast uppermost dword
 544         movdqa          $Hkey,$T1
 545         psllq           \$1,$Hkey
 546         pxor            $T3,$T3                 #
 547         psrlq           \$63,$T1
 548         pcmpgtd         $T2,$T3                 # broadcast carry bit
 549         pslldq          \$8,$T1
 550         por             $T1,$Hkey               # H<<=1
 551
 552         # magic reduction
 553         pand            .L0x1c2_polynomial(%rip),$T3
 554         pxor            $T3,$Hkey               # if(carry) H^=0x1c2_polynomial
 555
 556         # calculate H^2
 557         pshufd          \$0b01001110,$Hkey,$HK
 558         movdqa          $Hkey,$Xi
 559         pxor            $Hkey,$HK
 560 ___
 561         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);
 562         &reduction_alg9 ($Xhi,$Xi);
 563 $code.=<<___;
 564         pshufd          \$0b01001110,$Hkey,$T1
 565         pshufd          \$0b01001110,$Xi,$T2
 566         pxor            $Hkey,$T1               # Karatsuba pre-processing
 567         movdqu          $Hkey,0x00($Htbl)       # save H
 568         pxor            $Xi,$T2                 # Karatsuba pre-processing
 569         movdqu          $Xi,0x10($Htbl)         # save H^2
 570         palignr         \$8,$T1,$T2             # low part is H.lo^H.hi...
 571         movdqu          $T2,0x20($Htbl)         # save Karatsuba "salt"
 572 ___
 573 if ($do4xaggr) {
 574         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);   # H^3
 575         &reduction_alg9 ($Xhi,$Xi);
 576 $code.=<<___;
 577         movdqa          $Xi,$T3
 578 ___
 579         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);   # H^4
 580         &reduction_alg9 ($Xhi,$Xi);
 581 $code.=<<___;
 582         pshufd          \$0b01001110,$T3,$T1
 583         pshufd          \$0b01001110,$Xi,$T2
 584         pxor            $T3,$T1                 # Karatsuba pre-processing
 585         movdqu          $T3,0x30($Htbl)         # save H^3
 586         pxor            $Xi,$T2                 # Karatsuba pre-processing
 587         movdqu          $Xi,0x40($Htbl)         # save H^4
 588         palignr         \$8,$T1,$T2             # low part is H^3.lo^H^3.hi...
 589         movdqu          $T2,0x50($Htbl)         # save Karatsuba "salt"
 590 ___
 591 }
 592 $code.=<<___ if ($win64);
 593         movaps  (%rsp),%xmm6
 594         lea     0x18(%rsp),%rsp
 595 .LSEH_end_gcm_init_clmul:
 596 ___
 597 $code.=<<___;
 598         ret
 599 .size   gcm_init_clmul,.-gcm_init_clmul
 600 ___
 601 }
 602
 603 { my ($Xip,$Htbl)=@_4args;
 604
 605 $code.=<<___;
 606 .globl  gcm_gmult_clmul
 607 .type   gcm_gmult_clmul,\@abi-omnipotent
 608 .align  16
 609 gcm_gmult_clmul:
 610 .L_gmult_clmul:
 611         movdqu          ($Xip),$Xi
 612         movdqa          .Lbswap_mask(%rip),$T3
 613         movdqu          ($Htbl),$Hkey
 614         movdqu          0x20($Htbl),$T2
 615         pshufb          $T3,$Xi
 616 ___
 617         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$T2);
 618 $code.=<<___ if (0 || (&reduction_alg9($Xhi,$Xi)&&0));
 619         # experimental alternative. special thing about is that there
 620         # no dependency between the two multiplications...
 621         mov             \$`0xE1<<1`,%eax
 622         mov             \$0xA040608020C0E000,%r10       # ((7..0)·0xE0)&0xff
 623         mov             \$0x07,%r11d
 624         movq            %rax,$T1
 625         movq            %r10,$T2
 626         movq            %r11,$T3                # borrow $T3
 627         pand            $Xi,$T3
 628         pshufb          $T3,$T2                 # ($Xi&7)·0xE0
 629         movq            %rax,$T3
 630         pclmulqdq       \$0x00,$Xi,$T1          # ·(0xE1<<1)
 631         pxor            $Xi,$T2
 632         pslldq          \$15,$T2
 633         paddd           $T2,$T2                 # <<(64+56+1)
 634         pxor            $T2,$Xi
 635         pclmulqdq       \$0x01,$T3,$Xi
 636         movdqa          .Lbswap_mask(%rip),$T3  # reload $T3
 637         psrldq          \$1,$T1
 638         pxor            $T1,$Xhi
 639         pslldq          \$7,$Xi
 640         pxor            $Xhi,$Xi
 641 ___
 642 $code.=<<___;
 643         pshufb          $T3,$Xi
 644         movdqu          $Xi,($Xip)
 645         ret
 646 .size   gcm_gmult_clmul,.-gcm_gmult_clmul
 647 ___
 648 }
 649 \f
 650 { my ($Xip,$Htbl,$inp,$len)=@_4args;
 651   my ($Xln,$Xmn,$Xhn,$Hkey2,$HK) = map("%xmm$_",(3..7));
 652   my ($T1,$T2,$T3)=map("%xmm$_",(8..10));
 653
 654 $code.=<<___;
 655 .globl  gcm_ghash_clmul
 656 .type   gcm_ghash_clmul,\@abi-omnipotent
 657 .align  32
 658 gcm_ghash_clmul:
 659 .L_ghash_clmul:
 660 ___
 661 $code.=<<___ if ($win64);
 662         lea     -0x88(%rsp),%rax
 663 .LSEH_begin_gcm_ghash_clmul:
 664         # I can't trust assembler to use specific encoding:-(
 665         .byte   0x48,0x8d,0x60,0xe0             #lea    -0x20(%rax),%rsp
 666         .byte   0x0f,0x29,0x70,0xe0             #movaps %xmm6,-0x20(%rax)
 667         .byte   0x0f,0x29,0x78,0xf0             #movaps %xmm7,-0x10(%rax)
 668         .byte   0x44,0x0f,0x29,0x00             #movaps %xmm8,0(%rax)
 669         .byte   0x44,0x0f,0x29,0x48,0x10        #movaps %xmm9,0x10(%rax)
 670         .byte   0x44,0x0f,0x29,0x50,0x20        #movaps %xmm10,0x20(%rax)
 671         .byte   0x44,0x0f,0x29,0x58,0x30        #movaps %xmm11,0x30(%rax)
 672         .byte   0x44,0x0f,0x29,0x60,0x40        #movaps %xmm12,0x40(%rax)
 673         .byte   0x44,0x0f,0x29,0x68,0x50        #movaps %xmm13,0x50(%rax)
 674         .byte   0x44,0x0f,0x29,0x70,0x60        #movaps %xmm14,0x60(%rax)
 675         .byte   0x44,0x0f,0x29,0x78,0x70        #movaps %xmm15,0x70(%rax)
 676 ___
 677 $code.=<<___;
 678         movdqa          .Lbswap_mask(%rip),$T3
 679
 680         movdqu          ($Xip),$Xi
 681         movdqu          ($Htbl),$Hkey
 682         movdqu          0x20($Htbl),$HK
 683         pshufb          $T3,$Xi
 684
 685         sub             \$0x10,$len
 686         jz              .Lodd_tail
 687
 688         movdqu          0x10($Htbl),$Hkey2
 689 ___
 690 if ($do4xaggr) {
 691 my ($Xl,$Xm,$Xh,$Hkey3,$Hkey4)=map("%xmm$_",(11..15));
 692
 693 $code.=<<___;
 694         mov             OPENSSL_ia32cap_P+4(%rip),%eax
 695         cmp             \$0x30,$len
 696         jb              .Lskip4x
 697
 698         and             \$`1<<26|1<<22`,%eax    # isolate MOVBE+XSAVE
 699         cmp             \$`1<<22`,%eax          # check for MOVBE without XSAVE
 700         je              .Lskip4x
 701
 702         sub             \$0x30,$len
 703         mov             \$0xA040608020C0E000,%rax       # ((7..0)·0xE0)&0xff
 704         movdqu          0x30($Htbl),$Hkey3
 705         movdqu          0x40($Htbl),$Hkey4
 706
 707         #######
 708         # Xi+4 =[(H*Ii+3) + (H^2*Ii+2) + (H^3*Ii+1) + H^4*(Ii+Xi)] mod P
 709         #
 710         movdqu          0x30($inp),$Xln
 711          movdqu         0x20($inp),$Xl
 712         pshufb          $T3,$Xln
 713          pshufb         $T3,$Xl
 714         movdqa          $Xln,$Xhn
 715         pshufd          \$0b01001110,$Xln,$Xmn
 716         pxor            $Xln,$Xmn
 717         pclmulqdq       \$0x00,$Hkey,$Xln
 718         pclmulqdq       \$0x11,$Hkey,$Xhn
 719         pclmulqdq       \$0x00,$HK,$Xmn
 720
 721         movdqa          $Xl,$Xh
 722         pshufd          \$0b01001110,$Xl,$Xm
 723         pxor            $Xl,$Xm
 724         pclmulqdq       \$0x00,$Hkey2,$Xl
 725         pclmulqdq       \$0x11,$Hkey2,$Xh
 726         pclmulqdq       \$0x10,$HK,$Xm
 727         xorps           $Xl,$Xln
 728         xorps           $Xh,$Xhn
 729         movups          0x50($Htbl),$HK
 730         xorps           $Xm,$Xmn
 731
 732         movdqu          0x10($inp),$Xl
 733          movdqu         0($inp),$T1
 734         pshufb          $T3,$Xl
 735          pshufb         $T3,$T1
 736         movdqa          $Xl,$Xh
 737         pshufd          \$0b01001110,$Xl,$Xm
 738          pxor           $T1,$Xi
 739         pxor            $Xl,$Xm
 740         pclmulqdq       \$0x00,$Hkey3,$Xl
 741          movdqa         $Xi,$Xhi
 742          pshufd         \$0b01001110,$Xi,$T1
 743          pxor           $Xi,$T1
 744         pclmulqdq       \$0x11,$Hkey3,$Xh
 745         pclmulqdq       \$0x00,$HK,$Xm
 746         xorps           $Xl,$Xln
 747         xorps           $Xh,$Xhn
 748
 749         lea     0x40($inp),$inp
 750         sub     \$0x40,$len
 751         jc      .Ltail4x
 752
 753         jmp     .Lmod4_loop
 754 .align  32
 755 .Lmod4_loop:
 756         pclmulqdq       \$0x00,$Hkey4,$Xi
 757         xorps           $Xm,$Xmn
 758          movdqu         0x30($inp),$Xl
 759          pshufb         $T3,$Xl
 760         pclmulqdq       \$0x11,$Hkey4,$Xhi
 761         xorps           $Xln,$Xi
 762          movdqu         0x20($inp),$Xln
 763          movdqa         $Xl,$Xh
 764         pclmulqdq       \$0x10,$HK,$T1
 765          pshufd         \$0b01001110,$Xl,$Xm
 766         xorps           $Xhn,$Xhi
 767          pxor           $Xl,$Xm
 768          pshufb         $T3,$Xln
 769         movups          0x20($Htbl),$HK
 770         xorps           $Xmn,$T1
 771          pclmulqdq      \$0x00,$Hkey,$Xl
 772          pshufd         \$0b01001110,$Xln,$Xmn
 773
 774         pxor            $Xi,$T1                 # aggregated Karatsuba post-processing
 775          movdqa         $Xln,$Xhn
 776         pxor            $Xhi,$T1                #
 777          pxor           $Xln,$Xmn
 778         movdqa          $T1,$T2                 #
 779          pclmulqdq      \$0x11,$Hkey,$Xh
 780         pslldq          \$8,$T1
 781         psrldq          \$8,$T2                 #
 782         pxor            $T1,$Xi
 783         movdqa          .L7_mask(%rip),$T1
 784         pxor            $T2,$Xhi                #
 785         movq            %rax,$T2
 786
 787         pand            $Xi,$T1                 # 1st phase
 788         pshufb          $T1,$T2                 #
 789         pxor            $Xi,$T2                 #
 790          pclmulqdq      \$0x00,$HK,$Xm
 791         psllq           \$57,$T2                #
 792         movdqa          $T2,$T1                 #
 793         pslldq          \$8,$T2
 794          pclmulqdq      \$0x00,$Hkey2,$Xln
 795         psrldq          \$8,$T1                 #
 796         pxor            $T2,$Xi
 797         pxor            $T1,$Xhi                #
 798         movdqu          0($inp),$T1
 799
 800         movdqa          $Xi,$T2                 # 2nd phase
 801         psrlq           \$1,$Xi
 802          pclmulqdq      \$0x11,$Hkey2,$Xhn
 803          xorps          $Xl,$Xln
 804          movdqu         0x10($inp),$Xl
 805          pshufb         $T3,$Xl
 806          pclmulqdq      \$0x10,$HK,$Xmn
 807          xorps          $Xh,$Xhn
 808          movups         0x50($Htbl),$HK
 809         pshufb          $T3,$T1
 810         pxor            $T2,$Xhi                #
 811         pxor            $Xi,$T2
 812         psrlq           \$5,$Xi
 813
 814          movdqa         $Xl,$Xh
 815          pxor           $Xm,$Xmn
 816          pshufd         \$0b01001110,$Xl,$Xm
 817         pxor            $T2,$Xi                 #
 818         pxor            $T1,$Xhi
 819          pxor           $Xl,$Xm
 820          pclmulqdq      \$0x00,$Hkey3,$Xl
 821         psrlq           \$1,$Xi                 #
 822         pxor            $Xhi,$Xi                #
 823         movdqa          $Xi,$Xhi
 824          pclmulqdq      \$0x11,$Hkey3,$Xh
 825          xorps          $Xl,$Xln
 826         pshufd          \$0b01001110,$Xi,$T1
 827         pxor            $Xi,$T1
 828
 829          pclmulqdq      \$0x00,$HK,$Xm
 830          xorps          $Xh,$Xhn
 831
 832         lea     0x40($inp),$inp
 833         sub     \$0x40,$len
 834         jnc     .Lmod4_loop
 835
 836 .Ltail4x:
 837         pclmulqdq       \$0x00,$Hkey4,$Xi
 838         pclmulqdq       \$0x11,$Hkey4,$Xhi
 839         pclmulqdq       \$0x10,$HK,$T1
 840         xorps           $Xm,$Xmn
 841         xorps           $Xln,$Xi
 842         xorps           $Xhn,$Xhi
 843         pxor            $Xi,$Xhi                # aggregated Karatsuba post-processing
 844         pxor            $Xmn,$T1
 845
 846         pxor            $Xhi,$T1                #
 847         pxor            $Xi,$Xhi
 848
 849         movdqa          $T1,$T2                 #
 850         psrldq          \$8,$T1
 851         pslldq          \$8,$T2                 #
 852         pxor            $T1,$Xhi
 853         pxor            $T2,$Xi                 #
 854 ___
 855         &reduction_alg9($Xhi,$Xi);
 856 $code.=<<___;
 857         add     \$0x40,$len
 858         jz      .Ldone
 859         movdqu  0x20($Htbl),$HK
 860         sub     \$0x10,$len
 861         jz      .Lodd_tail
 862 .Lskip4x:
 863 ___
 864 }
 865 $code.=<<___;
 866         #######
 867         # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
 868         #       [(H*Ii+1) + (H*Xi+1)] mod P =
 869         #       [(H*Ii+1) + H^2*(Ii+Xi)] mod P
 870         #
 871         movdqu          ($inp),$T1              # Ii
 872         movdqu          16($inp),$Xln           # Ii+1
 873         pshufb          $T3,$T1
 874         pshufb          $T3,$Xln
 875         pxor            $T1,$Xi                 # Ii+Xi
 876
 877         movdqa          $Xln,$Xhn
 878         pshufd          \$0b01001110,$Xln,$Xmn
 879         pxor            $Xln,$Xmn
 880         pclmulqdq       \$0x00,$Hkey,$Xln
 881         pclmulqdq       \$0x11,$Hkey,$Xhn
 882         pclmulqdq       \$0x00,$HK,$Xmn
 883
 884         lea             32($inp),$inp           # i+=2
 885         nop
 886         sub             \$0x20,$len
 887         jbe             .Leven_tail
 888         nop
 889         jmp             .Lmod_loop
 890
 891 .align  32
 892 .Lmod_loop:
 893         movdqa          $Xi,$Xhi
 894         movdqa          $Xmn,$T1
 895         pshufd          \$0b01001110,$Xi,$Xmn   #
 896         pxor            $Xi,$Xmn                #
 897
 898         pclmulqdq       \$0x00,$Hkey2,$Xi
 899         pclmulqdq       \$0x11,$Hkey2,$Xhi
 900         pclmulqdq       \$0x10,$HK,$Xmn
 901
 902         pxor            $Xln,$Xi                # (H*Ii+1) + H^2*(Ii+Xi)
 903         pxor            $Xhn,$Xhi
 904           movdqu        ($inp),$T2              # Ii
 905         pxor            $Xi,$T1                 # aggregated Karatsuba post-processing
 906           pshufb        $T3,$T2
 907           movdqu        16($inp),$Xln           # Ii+1
 908
 909         pxor            $Xhi,$T1
 910           pxor          $T2,$Xhi                # "Ii+Xi", consume early
 911         pxor            $T1,$Xmn
 912          pshufb         $T3,$Xln
 913         movdqa          $Xmn,$T1                #
 914         psrldq          \$8,$T1
 915         pslldq          \$8,$Xmn                #
 916         pxor            $T1,$Xhi
 917         pxor            $Xmn,$Xi                #
 918
 919         movdqa          $Xln,$Xhn               #
 920
 921           movdqa        $Xi,$T2                 # 1st phase
 922           movdqa        $Xi,$T1
 923           psllq         \$5,$Xi
 924           pxor          $Xi,$T1                 #
 925         pclmulqdq       \$0x00,$Hkey,$Xln       #######
 926           psllq         \$1,$Xi
 927           pxor          $T1,$Xi                 #
 928           psllq         \$57,$Xi                #
 929           movdqa        $Xi,$T1                 #
 930           pslldq        \$8,$Xi
 931           psrldq        \$8,$T1                 #
 932           pxor          $T2,$Xi
 933         pshufd          \$0b01001110,$Xhn,$Xmn
 934           pxor          $T1,$Xhi                #
 935         pxor            $Xhn,$Xmn               #
 936
 937           movdqa        $Xi,$T2                 # 2nd phase
 938           psrlq         \$1,$Xi
 939         pclmulqdq       \$0x11,$Hkey,$Xhn       #######
 940           pxor          $T2,$Xhi                #
 941           pxor          $Xi,$T2
 942           psrlq         \$5,$Xi
 943           pxor          $T2,$Xi                 #
 944         lea             32($inp),$inp
 945           psrlq         \$1,$Xi                 #
 946         pclmulqdq       \$0x00,$HK,$Xmn         #######
 947           pxor          $Xhi,$Xi                #
 948
 949         sub             \$0x20,$len
 950         ja              .Lmod_loop
 951
 952 .Leven_tail:
 953          movdqa         $Xi,$Xhi
 954          movdqa         $Xmn,$T1
 955          pshufd         \$0b01001110,$Xi,$Xmn   #
 956          pxor           $Xi,$Xmn                #
 957
 958         pclmulqdq       \$0x00,$Hkey2,$Xi
 959         pclmulqdq       \$0x11,$Hkey2,$Xhi
 960         pclmulqdq       \$0x10,$HK,$Xmn
 961
 962         pxor            $Xln,$Xi                # (H*Ii+1) + H^2*(Ii+Xi)
 963         pxor            $Xhn,$Xhi
 964         pxor            $Xi,$T1
 965         pxor            $Xhi,$T1
 966         pxor            $T1,$Xmn
 967         movdqa          $Xmn,$T1                #
 968         psrldq          \$8,$T1
 969         pslldq          \$8,$Xmn                #
 970         pxor            $T1,$Xhi
 971         pxor            $Xmn,$Xi                #
 972 ___
 973         &reduction_alg9 ($Xhi,$Xi);
 974 $code.=<<___;
 975         test            $len,$len
 976         jnz             .Ldone
 977
 978 .Lodd_tail:
 979         movdqu          ($inp),$T1              # Ii
 980         pshufb          $T3,$T1
 981         pxor            $T1,$Xi                 # Ii+Xi
 982 ___
 983         &clmul64x64_T2  ($Xhi,$Xi,$Hkey,$HK);   # H*(Ii+Xi)
 984         &reduction_alg9 ($Xhi,$Xi);
 985 $code.=<<___;
 986 .Ldone:
 987         pshufb          $T3,$Xi
 988         movdqu          $Xi,($Xip)
 989 ___
 990 $code.=<<___ if ($win64);
 991         movaps  (%rsp),%xmm6
 992         movaps  0x10(%rsp),%xmm7
 993         movaps  0x20(%rsp),%xmm8
 994         movaps  0x30(%rsp),%xmm9
 995         movaps  0x40(%rsp),%xmm10
 996         movaps  0x50(%rsp),%xmm11
 997         movaps  0x60(%rsp),%xmm12
 998         movaps  0x70(%rsp),%xmm13
 999         movaps  0x80(%rsp),%xmm14
1000         movaps  0x90(%rsp),%xmm15
1001         lea     0xa8(%rsp),%rsp
1002 .LSEH_end_gcm_ghash_clmul:
1003 ___
1004 $code.=<<___;
1005         ret
1006 .size   gcm_ghash_clmul,.-gcm_ghash_clmul
1007 ___
1008 }
1009 \f
1010 $code.=<<___;
1011 .globl  gcm_init_avx
1012 .type   gcm_init_avx,\@abi-omnipotent
1013 .align  32
1014 gcm_init_avx:
1015 ___
1016 if ($avx) {
1017 my ($Htbl,$Xip)=@_4args;
1018 my $HK="%xmm6";
1019
1020 $code.=<<___ if ($win64);
1021 .LSEH_begin_gcm_init_avx:
1022         # I can't trust assembler to use specific encoding:-(
1023         .byte   0x48,0x83,0xec,0x18             #sub    $0x18,%rsp
1024         .byte   0x0f,0x29,0x34,0x24             #movaps %xmm6,(%rsp)
1025 ___
1026 $code.=<<___;
1027         vzeroupper
1028
1029         vmovdqu         ($Xip),$Hkey
1030         vpshufd         \$0b01001110,$Hkey,$Hkey        # dword swap
1031
1032         # <<1 twist
1033         vpshufd         \$0b11111111,$Hkey,$T2  # broadcast uppermost dword
1034         vpsrlq          \$63,$Hkey,$T1
1035         vpsllq          \$1,$Hkey,$Hkey
1036         vpxor           $T3,$T3,$T3             #
1037         vpcmpgtd        $T2,$T3,$T3             # broadcast carry bit
1038         vpslldq         \$8,$T1,$T1
1039         vpor            $T1,$Hkey,$Hkey         # H<<=1
1040
1041         # magic reduction
1042         vpand           .L0x1c2_polynomial(%rip),$T3,$T3
1043         vpxor           $T3,$Hkey,$Hkey         # if(carry) H^=0x1c2_polynomial
1044
1045         vpunpckhqdq     $Hkey,$Hkey,$HK
1046         vmovdqa         $Hkey,$Xi
1047         vpxor           $Hkey,$HK,$HK
1048         mov             \$4,%r10                # up to H^8
1049         jmp             .Linit_start_avx
1050 ___
1051
1052 sub clmul64x64_avx {
1053 my ($Xhi,$Xi,$Hkey,$HK)=@_;
1054
1055 if (!defined($HK)) {    $HK = $T2;
1056 $code.=<<___;
1057         vpunpckhqdq     $Xi,$Xi,$T1
1058         vpunpckhqdq     $Hkey,$Hkey,$T2
1059         vpxor           $Xi,$T1,$T1             #
1060         vpxor           $Hkey,$T2,$T2
1061 ___
1062 } else {
1063 $code.=<<___;
1064         vpunpckhqdq     $Xi,$Xi,$T1
1065         vpxor           $Xi,$T1,$T1             #
1066 ___
1067 }
1068 $code.=<<___;
1069         vpclmulqdq      \$0x11,$Hkey,$Xi,$Xhi   #######
1070         vpclmulqdq      \$0x00,$Hkey,$Xi,$Xi    #######
1071         vpclmulqdq      \$0x00,$HK,$T1,$T1      #######
1072         vpxor           $Xi,$Xhi,$T2            #
1073         vpxor           $T2,$T1,$T1             #
1074
1075         vpslldq         \$8,$T1,$T2             #
1076         vpsrldq         \$8,$T1,$T1
1077         vpxor           $T2,$Xi,$Xi             #
1078         vpxor           $T1,$Xhi,$Xhi
1079 ___
1080 }
1081
1082 sub reduction_avx {
1083 my ($Xhi,$Xi) = @_;
1084
1085 $code.=<<___;
1086         vpsllq          \$57,$Xi,$T1            # 1st phase
1087         vpsllq          \$62,$Xi,$T2
1088         vpxor           $T1,$T2,$T2             #
1089         vpsllq          \$63,$Xi,$T1
1090         vpxor           $T1,$T2,$T2             #
1091         vpslldq         \$8,$T2,$T1             #
1092         vpsrldq         \$8,$T2,$T2
1093         vpxor           $T1,$Xi,$Xi             #
1094         vpxor           $T2,$Xhi,$Xhi
1095
1096         vpsrlq          \$1,$Xi,$T2             # 2nd phase
1097         vpxor           $Xi,$Xhi,$Xhi
1098         vpxor           $T2,$Xi,$Xi             #
1099         vpsrlq          \$5,$T2,$T2
1100         vpxor           $T2,$Xi,$Xi             #
1101         vpsrlq          \$1,$Xi,$Xi             #
1102         vpxor           $Xhi,$Xi,$Xi            #
1103 ___
1104 }
1105
1106 $code.=<<___;
1107 .align  32
1108 .Linit_loop_avx:
1109         vpalignr        \$8,$T1,$T2,$T3         # low part is H.lo^H.hi...
1110         vmovdqu         $T3,-0x10($Htbl)        # save Karatsuba "salt"
1111 ___
1112         &clmul64x64_avx ($Xhi,$Xi,$Hkey,$HK);   # calculate H^3,5,7
1113         &reduction_avx  ($Xhi,$Xi);
1114 $code.=<<___;
1115 .Linit_start_avx:
1116         vmovdqa         $Xi,$T3
1117 ___
1118         &clmul64x64_avx ($Xhi,$Xi,$Hkey,$HK);   # calculate H^2,4,6,8
1119         &reduction_avx  ($Xhi,$Xi);
1120 $code.=<<___;
1121         vpshufd         \$0b01001110,$T3,$T1
1122         vpshufd         \$0b01001110,$Xi,$T2
1123         vpxor           $T3,$T1,$T1             # Karatsuba pre-processing
1124         vmovdqu         $T3,0x00($Htbl)         # save H^1,3,5,7
1125         vpxor           $Xi,$T2,$T2             # Karatsuba pre-processing
1126         vmovdqu         $Xi,0x10($Htbl)         # save H^2,4,6,8
1127         lea             0x30($Htbl),$Htbl
1128         sub             \$1,%r10
1129         jnz             .Linit_loop_avx
1130
1131         vpalignr        \$8,$T2,$T1,$T3         # last "salt" is flipped
1132         vmovdqu         $T3,-0x10($Htbl)
1133
1134         vzeroupper
1135 ___
1136 $code.=<<___ if ($win64);
1137         movaps  (%rsp),%xmm6
1138         lea     0x18(%rsp),%rsp
1139 .LSEH_end_gcm_init_avx:
1140 ___
1141 $code.=<<___;
1142         ret
1143 .size   gcm_init_avx,.-gcm_init_avx
1144 ___
1145 } else {
1146 $code.=<<___;
1147         jmp     .L_init_clmul
1148 .size   gcm_init_avx,.-gcm_init_avx
1149 ___
1150 }
1151
1152 $code.=<<___;
1153 .globl  gcm_gmult_avx
1154 .type   gcm_gmult_avx,\@abi-omnipotent
1155 .align  32
1156 gcm_gmult_avx:
1157         jmp     .L_gmult_clmul
1158 .size   gcm_gmult_avx,.-gcm_gmult_avx
1159 ___
1160 \f
1161 $code.=<<___;
1162 .globl  gcm_ghash_avx
1163 .type   gcm_ghash_avx,\@abi-omnipotent
1164 .align  32
1165 gcm_ghash_avx:
1166 ___
1167 if ($avx) {
1168 my ($Xip,$Htbl,$inp,$len)=@_4args;
1169 my ($Xlo,$Xhi,$Xmi,
1170     $Zlo,$Zhi,$Zmi,
1171     $Hkey,$HK,$T1,$T2,
1172     $Xi,$Xo,$Tred,$bswap,$Ii,$Ij) = map("%xmm$_",(0..15));
1173
1174 $code.=<<___ if ($win64);
1175         lea     -0x88(%rsp),%rax
1176 .LSEH_begin_gcm_ghash_avx:
1177         # I can't trust assembler to use specific encoding:-(
1178         .byte   0x48,0x8d,0x60,0xe0             #lea    -0x20(%rax),%rsp
1179         .byte   0x0f,0x29,0x70,0xe0             #movaps %xmm6,-0x20(%rax)
1180         .byte   0x0f,0x29,0x78,0xf0             #movaps %xmm7,-0x10(%rax)
1181         .byte   0x44,0x0f,0x29,0x00             #movaps %xmm8,0(%rax)
1182         .byte   0x44,0x0f,0x29,0x48,0x10        #movaps %xmm9,0x10(%rax)
1183         .byte   0x44,0x0f,0x29,0x50,0x20        #movaps %xmm10,0x20(%rax)
1184         .byte   0x44,0x0f,0x29,0x58,0x30        #movaps %xmm11,0x30(%rax)
1185         .byte   0x44,0x0f,0x29,0x60,0x40        #movaps %xmm12,0x40(%rax)
1186         .byte   0x44,0x0f,0x29,0x68,0x50        #movaps %xmm13,0x50(%rax)
1187         .byte   0x44,0x0f,0x29,0x70,0x60        #movaps %xmm14,0x60(%rax)
1188         .byte   0x44,0x0f,0x29,0x78,0x70        #movaps %xmm15,0x70(%rax)
1189 ___
1190 $code.=<<___;
1191         vzeroupper
1192
1193         vmovdqu         ($Xip),$Xi              # load $Xi
1194         lea             .L0x1c2_polynomial(%rip),%r10
1195         lea             0x40($Htbl),$Htbl       # size optimization
1196         vmovdqu         .Lbswap_mask(%rip),$bswap
1197         vpshufb         $bswap,$Xi,$Xi
1198         cmp             \$0x80,$len
1199         jb              .Lshort_avx
1200         sub             \$0x80,$len
1201
1202         vmovdqu         0x70($inp),$Ii          # I[7]
1203         vmovdqu         0x00-0x40($Htbl),$Hkey  # $Hkey^1
1204         vpshufb         $bswap,$Ii,$Ii
1205         vmovdqu         0x20-0x40($Htbl),$HK
1206
1207         vpunpckhqdq     $Ii,$Ii,$T2
1208          vmovdqu        0x60($inp),$Ij          # I[6]
1209         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1210         vpxor           $Ii,$T2,$T2
1211          vpshufb        $bswap,$Ij,$Ij
1212         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1213          vmovdqu        0x10-0x40($Htbl),$Hkey  # $Hkey^2
1214          vpunpckhqdq    $Ij,$Ij,$T1
1215          vmovdqu        0x50($inp),$Ii          # I[5]
1216         vpclmulqdq      \$0x00,$HK,$T2,$Xmi
1217          vpxor          $Ij,$T1,$T1
1218
1219          vpshufb        $bswap,$Ii,$Ii
1220         vpclmulqdq      \$0x00,$Hkey,$Ij,$Zlo
1221          vpunpckhqdq    $Ii,$Ii,$T2
1222         vpclmulqdq      \$0x11,$Hkey,$Ij,$Zhi
1223          vmovdqu        0x30-0x40($Htbl),$Hkey  # $Hkey^3
1224          vpxor          $Ii,$T2,$T2
1225          vmovdqu        0x40($inp),$Ij          # I[4]
1226         vpclmulqdq      \$0x10,$HK,$T1,$Zmi
1227          vmovdqu        0x50-0x40($Htbl),$HK
1228
1229          vpshufb        $bswap,$Ij,$Ij
1230         vpxor           $Xlo,$Zlo,$Zlo
1231         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1232         vpxor           $Xhi,$Zhi,$Zhi
1233          vpunpckhqdq    $Ij,$Ij,$T1
1234         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1235          vmovdqu        0x40-0x40($Htbl),$Hkey  # $Hkey^4
1236         vpxor           $Xmi,$Zmi,$Zmi
1237         vpclmulqdq      \$0x00,$HK,$T2,$Xmi
1238          vpxor          $Ij,$T1,$T1
1239
1240          vmovdqu        0x30($inp),$Ii          # I[3]
1241         vpxor           $Zlo,$Xlo,$Xlo
1242         vpclmulqdq      \$0x00,$Hkey,$Ij,$Zlo
1243         vpxor           $Zhi,$Xhi,$Xhi
1244          vpshufb        $bswap,$Ii,$Ii
1245         vpclmulqdq      \$0x11,$Hkey,$Ij,$Zhi
1246          vmovdqu        0x60-0x40($Htbl),$Hkey  # $Hkey^5
1247         vpxor           $Zmi,$Xmi,$Xmi
1248          vpunpckhqdq    $Ii,$Ii,$T2
1249         vpclmulqdq      \$0x10,$HK,$T1,$Zmi
1250          vmovdqu        0x80-0x40($Htbl),$HK
1251          vpxor          $Ii,$T2,$T2
1252
1253          vmovdqu        0x20($inp),$Ij          # I[2]
1254         vpxor           $Xlo,$Zlo,$Zlo
1255         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1256         vpxor           $Xhi,$Zhi,$Zhi
1257          vpshufb        $bswap,$Ij,$Ij
1258         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1259          vmovdqu        0x70-0x40($Htbl),$Hkey  # $Hkey^6
1260         vpxor           $Xmi,$Zmi,$Zmi
1261          vpunpckhqdq    $Ij,$Ij,$T1
1262         vpclmulqdq      \$0x00,$HK,$T2,$Xmi
1263          vpxor          $Ij,$T1,$T1
1264
1265          vmovdqu        0x10($inp),$Ii          # I[1]
1266         vpxor           $Zlo,$Xlo,$Xlo
1267         vpclmulqdq      \$0x00,$Hkey,$Ij,$Zlo
1268         vpxor           $Zhi,$Xhi,$Xhi
1269          vpshufb        $bswap,$Ii,$Ii
1270         vpclmulqdq      \$0x11,$Hkey,$Ij,$Zhi
1271          vmovdqu        0x90-0x40($Htbl),$Hkey  # $Hkey^7
1272         vpxor           $Zmi,$Xmi,$Xmi
1273          vpunpckhqdq    $Ii,$Ii,$T2
1274         vpclmulqdq      \$0x10,$HK,$T1,$Zmi
1275          vmovdqu        0xb0-0x40($Htbl),$HK
1276          vpxor          $Ii,$T2,$T2
1277
1278          vmovdqu        ($inp),$Ij              # I[0]
1279         vpxor           $Xlo,$Zlo,$Zlo
1280         vpclmulqdq      \$0x00,$Hkey,$Ii,$Xlo
1281         vpxor           $Xhi,$Zhi,$Zhi
1282          vpshufb        $bswap,$Ij,$Ij
1283         vpclmulqdq      \$0x11,$Hkey,$Ii,$Xhi
1284          vmovdqu        0xa0-0x40($Htbl),$Hkey  # $Hkey^8
1285         vpxor           $Xmi,$Zmi,$Zmi
1286         vpclmulqdq      \$0x10,$HK,$T2,$Xmi
1287
1288         lea             0x80($inp),$inp
1289         cmp             \$0x80,$len
1290         jb              .Ltail_avx
1291
1292         vpxor           $Xi,$Ij,$Ij             # accumulate $Xi
1293         sub             \$0x80,$len
1294         jmp             .Loop8x_avx
1295
1296 .align  32
1297 .Loop8x_avx:
1298         vpunpckhqdq     $Ij,$Ij,$T1
1299          vmovdqu        0x70($inp),$Ii          # I[7]
1300         vpxor           $Xlo,$Zlo,$Zlo
1301         vpxor           $Ij,$T1,$T1
1302         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xi
1303          vpshufb        $bswap,$Ii,$Ii
1304         vpxor           $Xhi,$Zhi,$Zhi
1305         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xo
1306          vmovdqu        0x00-0x40($Htbl),$Hkey  # $Hkey^1
1307          vpunpckhqdq    $Ii,$Ii,$T2
1308         vpxor           $Xmi,$Zmi,$Zmi
1309         vpclmulqdq      \$0x00,$HK,$T1,$Tred
1310          vmovdqu        0x20-0x40($Htbl),$HK
1311          vpxor          $Ii,$T2,$T2
1312
1313           vmovdqu       0x60($inp),$Ij          # I[6]
1314          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1315         vpxor           $Zlo,$Xi,$Xi            # collect result
1316           vpshufb       $bswap,$Ij,$Ij
1317          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1318         vxorps          $Zhi,$Xo,$Xo
1319           vmovdqu       0x10-0x40($Htbl),$Hkey  # $Hkey^2
1320          vpunpckhqdq    $Ij,$Ij,$T1
1321          vpclmulqdq     \$0x00,$HK,  $T2,$Xmi
1322         vpxor           $Zmi,$Tred,$Tred
1323          vxorps         $Ij,$T1,$T1
1324
1325           vmovdqu       0x50($inp),$Ii          # I[5]
1326         vpxor           $Xi,$Tred,$Tred         # aggregated Karatsuba post-processing
1327          vpclmulqdq     \$0x00,$Hkey,$Ij,$Zlo
1328         vpxor           $Xo,$Tred,$Tred
1329         vpslldq         \$8,$Tred,$T2
1330          vpxor          $Xlo,$Zlo,$Zlo
1331          vpclmulqdq     \$0x11,$Hkey,$Ij,$Zhi
1332         vpsrldq         \$8,$Tred,$Tred
1333         vpxor           $T2, $Xi, $Xi
1334           vmovdqu       0x30-0x40($Htbl),$Hkey  # $Hkey^3
1335           vpshufb       $bswap,$Ii,$Ii
1336         vxorps          $Tred,$Xo, $Xo
1337          vpxor          $Xhi,$Zhi,$Zhi
1338          vpunpckhqdq    $Ii,$Ii,$T2
1339          vpclmulqdq     \$0x10,$HK,  $T1,$Zmi
1340           vmovdqu       0x50-0x40($Htbl),$HK
1341          vpxor          $Ii,$T2,$T2
1342          vpxor          $Xmi,$Zmi,$Zmi
1343
1344           vmovdqu       0x40($inp),$Ij          # I[4]
1345         vpalignr        \$8,$Xi,$Xi,$Tred       # 1st phase
1346          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1347           vpshufb       $bswap,$Ij,$Ij
1348          vpxor          $Zlo,$Xlo,$Xlo
1349          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1350           vmovdqu       0x40-0x40($Htbl),$Hkey  # $Hkey^4
1351          vpunpckhqdq    $Ij,$Ij,$T1
1352          vpxor          $Zhi,$Xhi,$Xhi
1353          vpclmulqdq     \$0x00,$HK,  $T2,$Xmi
1354          vxorps         $Ij,$T1,$T1
1355          vpxor          $Zmi,$Xmi,$Xmi
1356
1357           vmovdqu       0x30($inp),$Ii          # I[3]
1358         vpclmulqdq      \$0x10,(%r10),$Xi,$Xi
1359          vpclmulqdq     \$0x00,$Hkey,$Ij,$Zlo
1360           vpshufb       $bswap,$Ii,$Ii
1361          vpxor          $Xlo,$Zlo,$Zlo
1362          vpclmulqdq     \$0x11,$Hkey,$Ij,$Zhi
1363           vmovdqu       0x60-0x40($Htbl),$Hkey  # $Hkey^5
1364          vpunpckhqdq    $Ii,$Ii,$T2
1365          vpxor          $Xhi,$Zhi,$Zhi
1366          vpclmulqdq     \$0x10,$HK,  $T1,$Zmi
1367           vmovdqu       0x80-0x40($Htbl),$HK
1368          vpxor          $Ii,$T2,$T2
1369          vpxor          $Xmi,$Zmi,$Zmi
1370
1371           vmovdqu       0x20($inp),$Ij          # I[2]
1372          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1373           vpshufb       $bswap,$Ij,$Ij
1374          vpxor          $Zlo,$Xlo,$Xlo
1375          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1376           vmovdqu       0x70-0x40($Htbl),$Hkey  # $Hkey^6
1377          vpunpckhqdq    $Ij,$Ij,$T1
1378          vpxor          $Zhi,$Xhi,$Xhi
1379          vpclmulqdq     \$0x00,$HK,  $T2,$Xmi
1380          vpxor          $Ij,$T1,$T1
1381          vpxor          $Zmi,$Xmi,$Xmi
1382         vxorps          $Tred,$Xi,$Xi
1383
1384           vmovdqu       0x10($inp),$Ii          # I[1]
1385         vpalignr        \$8,$Xi,$Xi,$Tred       # 2nd phase
1386          vpclmulqdq     \$0x00,$Hkey,$Ij,$Zlo
1387           vpshufb       $bswap,$Ii,$Ii
1388          vpxor          $Xlo,$Zlo,$Zlo
1389          vpclmulqdq     \$0x11,$Hkey,$Ij,$Zhi
1390           vmovdqu       0x90-0x40($Htbl),$Hkey  # $Hkey^7
1391         vpclmulqdq      \$0x10,(%r10),$Xi,$Xi
1392         vxorps          $Xo,$Tred,$Tred
1393          vpunpckhqdq    $Ii,$Ii,$T2
1394          vpxor          $Xhi,$Zhi,$Zhi
1395          vpclmulqdq     \$0x10,$HK,  $T1,$Zmi
1396           vmovdqu       0xb0-0x40($Htbl),$HK
1397          vpxor          $Ii,$T2,$T2
1398          vpxor          $Xmi,$Zmi,$Zmi
1399
1400           vmovdqu       ($inp),$Ij              # I[0]
1401          vpclmulqdq     \$0x00,$Hkey,$Ii,$Xlo
1402           vpshufb       $bswap,$Ij,$Ij
1403          vpclmulqdq     \$0x11,$Hkey,$Ii,$Xhi
1404           vmovdqu       0xa0-0x40($Htbl),$Hkey  # $Hkey^8
1405         vpxor           $Tred,$Ij,$Ij
1406          vpclmulqdq     \$0x10,$HK,  $T2,$Xmi
1407         vpxor           $Xi,$Ij,$Ij             # accumulate $Xi
1408
1409         lea             0x80($inp),$inp
1410         sub             \$0x80,$len
1411         jnc             .Loop8x_avx
1412
1413         add             \$0x80,$len
1414         jmp             .Ltail_no_xor_avx
1415
1416 .align  32
1417 .Lshort_avx:
1418         vmovdqu         -0x10($inp,$len),$Ii    # very last word
1419         lea             ($inp,$len),$inp
1420         vmovdqu         0x00-0x40($Htbl),$Hkey  # $Hkey^1
1421         vmovdqu         0x20-0x40($Htbl),$HK
1422         vpshufb         $bswap,$Ii,$Ij
1423
1424         vmovdqa         $Xlo,$Zlo               # subtle way to zero $Zlo,
1425         vmovdqa         $Xhi,$Zhi               # $Zhi and
1426         vmovdqa         $Xmi,$Zmi               # $Zmi
1427         sub             \$0x10,$len
1428         jz              .Ltail_avx
1429
1430         vpunpckhqdq     $Ij,$Ij,$T1
1431         vpxor           $Xlo,$Zlo,$Zlo
1432         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1433         vpxor           $Ij,$T1,$T1
1434          vmovdqu        -0x20($inp),$Ii
1435         vpxor           $Xhi,$Zhi,$Zhi
1436         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1437         vmovdqu         0x10-0x40($Htbl),$Hkey  # $Hkey^2
1438          vpshufb        $bswap,$Ii,$Ij
1439         vpxor           $Xmi,$Zmi,$Zmi
1440         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1441         vpsrldq         \$8,$HK,$HK
1442         sub             \$0x10,$len
1443         jz              .Ltail_avx
1444
1445         vpunpckhqdq     $Ij,$Ij,$T1
1446         vpxor           $Xlo,$Zlo,$Zlo
1447         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1448         vpxor           $Ij,$T1,$T1
1449          vmovdqu        -0x30($inp),$Ii
1450         vpxor           $Xhi,$Zhi,$Zhi
1451         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1452         vmovdqu         0x30-0x40($Htbl),$Hkey  # $Hkey^3
1453          vpshufb        $bswap,$Ii,$Ij
1454         vpxor           $Xmi,$Zmi,$Zmi
1455         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1456         vmovdqu         0x50-0x40($Htbl),$HK
1457         sub             \$0x10,$len
1458         jz              .Ltail_avx
1459
1460         vpunpckhqdq     $Ij,$Ij,$T1
1461         vpxor           $Xlo,$Zlo,$Zlo
1462         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1463         vpxor           $Ij,$T1,$T1
1464          vmovdqu        -0x40($inp),$Ii
1465         vpxor           $Xhi,$Zhi,$Zhi
1466         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1467         vmovdqu         0x40-0x40($Htbl),$Hkey  # $Hkey^4
1468          vpshufb        $bswap,$Ii,$Ij
1469         vpxor           $Xmi,$Zmi,$Zmi
1470         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1471         vpsrldq         \$8,$HK,$HK
1472         sub             \$0x10,$len
1473         jz              .Ltail_avx
1474
1475         vpunpckhqdq     $Ij,$Ij,$T1
1476         vpxor           $Xlo,$Zlo,$Zlo
1477         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1478         vpxor           $Ij,$T1,$T1
1479          vmovdqu        -0x50($inp),$Ii
1480         vpxor           $Xhi,$Zhi,$Zhi
1481         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1482         vmovdqu         0x60-0x40($Htbl),$Hkey  # $Hkey^5
1483          vpshufb        $bswap,$Ii,$Ij
1484         vpxor           $Xmi,$Zmi,$Zmi
1485         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1486         vmovdqu         0x80-0x40($Htbl),$HK
1487         sub             \$0x10,$len
1488         jz              .Ltail_avx
1489
1490         vpunpckhqdq     $Ij,$Ij,$T1
1491         vpxor           $Xlo,$Zlo,$Zlo
1492         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1493         vpxor           $Ij,$T1,$T1
1494          vmovdqu        -0x60($inp),$Ii
1495         vpxor           $Xhi,$Zhi,$Zhi
1496         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1497         vmovdqu         0x70-0x40($Htbl),$Hkey  # $Hkey^6
1498          vpshufb        $bswap,$Ii,$Ij
1499         vpxor           $Xmi,$Zmi,$Zmi
1500         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1501         vpsrldq         \$8,$HK,$HK
1502         sub             \$0x10,$len
1503         jz              .Ltail_avx
1504
1505         vpunpckhqdq     $Ij,$Ij,$T1
1506         vpxor           $Xlo,$Zlo,$Zlo
1507         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1508         vpxor           $Ij,$T1,$T1
1509          vmovdqu        -0x70($inp),$Ii
1510         vpxor           $Xhi,$Zhi,$Zhi
1511         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1512         vmovdqu         0x90-0x40($Htbl),$Hkey  # $Hkey^7
1513          vpshufb        $bswap,$Ii,$Ij
1514         vpxor           $Xmi,$Zmi,$Zmi
1515         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1516         vmovq           0xb8-0x40($Htbl),$HK
1517         sub             \$0x10,$len
1518         jmp             .Ltail_avx
1519
1520 .align  32
1521 .Ltail_avx:
1522         vpxor           $Xi,$Ij,$Ij             # accumulate $Xi
1523 .Ltail_no_xor_avx:
1524         vpunpckhqdq     $Ij,$Ij,$T1
1525         vpxor           $Xlo,$Zlo,$Zlo
1526         vpclmulqdq      \$0x00,$Hkey,$Ij,$Xlo
1527         vpxor           $Ij,$T1,$T1
1528         vpxor           $Xhi,$Zhi,$Zhi
1529         vpclmulqdq      \$0x11,$Hkey,$Ij,$Xhi
1530         vpxor           $Xmi,$Zmi,$Zmi
1531         vpclmulqdq      \$0x00,$HK,$T1,$Xmi
1532
1533         vmovdqu         (%r10),$Tred
1534
1535         vpxor           $Xlo,$Zlo,$Xi
1536         vpxor           $Xhi,$Zhi,$Xo
1537         vpxor           $Xmi,$Zmi,$Zmi
1538
1539         vpxor           $Xi, $Zmi,$Zmi          # aggregated Karatsuba post-processing
1540         vpxor           $Xo, $Zmi,$Zmi
1541         vpslldq         \$8, $Zmi,$T2
1542         vpsrldq         \$8, $Zmi,$Zmi
1543         vpxor           $T2, $Xi, $Xi
1544         vpxor           $Zmi,$Xo, $Xo
1545
1546         vpclmulqdq      \$0x10,$Tred,$Xi,$T2    # 1st phase
1547         vpalignr        \$8,$Xi,$Xi,$Xi
1548         vpxor           $T2,$Xi,$Xi
1549
1550         vpclmulqdq      \$0x10,$Tred,$Xi,$T2    # 2nd phase
1551         vpalignr        \$8,$Xi,$Xi,$Xi
1552         vpxor           $Xo,$Xi,$Xi
1553         vpxor           $T2,$Xi,$Xi
1554
1555         cmp             \$0,$len
1556         jne             .Lshort_avx
1557
1558         vpshufb         $bswap,$Xi,$Xi
1559         vmovdqu         $Xi,($Xip)
1560         vzeroupper
1561 ___
1562 $code.=<<___ if ($win64);
1563         movaps  (%rsp),%xmm6
1564         movaps  0x10(%rsp),%xmm7
1565         movaps  0x20(%rsp),%xmm8
1566         movaps  0x30(%rsp),%xmm9
1567         movaps  0x40(%rsp),%xmm10
1568         movaps  0x50(%rsp),%xmm11
1569         movaps  0x60(%rsp),%xmm12
1570         movaps  0x70(%rsp),%xmm13
1571         movaps  0x80(%rsp),%xmm14
1572         movaps  0x90(%rsp),%xmm15
1573         lea     0xa8(%rsp),%rsp
1574 .LSEH_end_gcm_ghash_avx:
1575 ___
1576 $code.=<<___;
1577         ret
1578 .size   gcm_ghash_avx,.-gcm_ghash_avx
1579 ___
1580 } else {
1581 $code.=<<___;
1582         jmp     .L_ghash_clmul
1583 .size   gcm_ghash_avx,.-gcm_ghash_avx
1584 ___
1585 }
1586 \f
1587 $code.=<<___;
1588 .align  64
1589 .Lbswap_mask:
1590         .byte   15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
1591 .L0x1c2_polynomial:
1592         .byte   1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
1593 .L7_mask:
1594         .long   7,0,7,0
1595 .L7_mask_poly:
1596         .long   7,0,`0xE1<<1`,0
1597 .align  64
1598 .type   .Lrem_4bit,\@object
1599 .Lrem_4bit:
1600         .long   0,`0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`
1601         .long   0,`0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`
1602         .long   0,`0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`
1603         .long   0,`0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`
1604 .type   .Lrem_8bit,\@object
1605 .Lrem_8bit:
1606         .value  0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E
1607         .value  0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E
1608         .value  0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E
1609         .value  0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E
1610         .value  0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E
1611         .value  0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E
1612         .value  0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E
1613         .value  0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E
1614         .value  0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE
1615         .value  0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE
1616         .value  0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE
1617         .value  0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE
1618         .value  0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E
1619         .value  0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E
1620         .value  0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE
1621         .value  0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE
1622         .value  0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E
1623         .value  0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E
1624         .value  0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E
1625         .value  0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E
1626         .value  0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E
1627         .value  0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E
1628         .value  0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E
1629         .value  0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E
1630         .value  0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE
1631         .value  0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE
1632         .value  0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE
1633         .value  0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE
1634         .value  0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E
1635         .value  0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E
1636         .value  0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE
1637         .value  0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE
1638
1639 .asciz  "GHASH for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1640 .align  64
1641 ___
1642 \f
1643 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1644 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
1645 if ($win64) {
1646 $rec="%rcx";
1647 $frame="%rdx";
1648 $context="%r8";
1649 $disp="%r9";
1650
1651 $code.=<<___;
1652 .extern __imp_RtlVirtualUnwind
1653 .type   se_handler,\@abi-omnipotent
1654 .align  16
1655 se_handler:
1656         push    %rsi
1657         push    %rdi
1658         push    %rbx
1659         push    %rbp
1660         push    %r12
1661         push    %r13
1662         push    %r14
1663         push    %r15
1664         pushfq
1665         sub     \$64,%rsp
1666
1667         mov     120($context),%rax      # pull context->Rax
1668         mov     248($context),%rbx      # pull context->Rip
1669
1670         mov     8($disp),%rsi           # disp->ImageBase
1671         mov     56($disp),%r11          # disp->HandlerData
1672
1673         mov     0(%r11),%r10d           # HandlerData[0]
1674         lea     (%rsi,%r10),%r10        # prologue label
1675         cmp     %r10,%rbx               # context->Rip<prologue label
1676         jb      .Lin_prologue
1677
1678         mov     152($context),%rax      # pull context->Rsp
1679
1680         mov     4(%r11),%r10d           # HandlerData[1]
1681         lea     (%rsi,%r10),%r10        # epilogue label
1682         cmp     %r10,%rbx               # context->Rip>=epilogue label
1683         jae     .Lin_prologue
1684
1685         lea     48+280(%rax),%rax       # adjust "rsp"
1686
1687         mov     -8(%rax),%rbx
1688         mov     -16(%rax),%rbp
1689         mov     -24(%rax),%r12
1690         mov     -32(%rax),%r13
1691         mov     -40(%rax),%r14
1692         mov     -48(%rax),%r15
1693         mov     %rbx,144($context)      # restore context->Rbx
1694         mov     %rbp,160($context)      # restore context->Rbp
1695         mov     %r12,216($context)      # restore context->R12
1696         mov     %r13,224($context)      # restore context->R13
1697         mov     %r14,232($context)      # restore context->R14
1698         mov     %r15,240($context)      # restore context->R15
1699
1700 .Lin_prologue:
1701         mov     8(%rax),%rdi
1702         mov     16(%rax),%rsi
1703         mov     %rax,152($context)      # restore context->Rsp
1704         mov     %rsi,168($context)      # restore context->Rsi
1705         mov     %rdi,176($context)      # restore context->Rdi
1706
1707         mov     40($disp),%rdi          # disp->ContextRecord
1708         mov     $context,%rsi           # context
1709         mov     \$`1232/8`,%ecx         # sizeof(CONTEXT)
1710         .long   0xa548f3fc              # cld; rep movsq
1711
1712         mov     $disp,%rsi
1713         xor     %rcx,%rcx               # arg1, UNW_FLAG_NHANDLER
1714         mov     8(%rsi),%rdx            # arg2, disp->ImageBase
1715         mov     0(%rsi),%r8             # arg3, disp->ControlPc
1716         mov     16(%rsi),%r9            # arg4, disp->FunctionEntry
1717         mov     40(%rsi),%r10           # disp->ContextRecord
1718         lea     56(%rsi),%r11           # &disp->HandlerData
1719         lea     24(%rsi),%r12           # &disp->EstablisherFrame
1720         mov     %r10,32(%rsp)           # arg5
1721         mov     %r11,40(%rsp)           # arg6
1722         mov     %r12,48(%rsp)           # arg7
1723         mov     %rcx,56(%rsp)           # arg8, (NULL)
1724         call    *__imp_RtlVirtualUnwind(%rip)
1725
1726         mov     \$1,%eax                # ExceptionContinueSearch
1727         add     \$64,%rsp
1728         popfq
1729         pop     %r15
1730         pop     %r14
1731         pop     %r13
1732         pop     %r12
1733         pop     %rbp
1734         pop     %rbx
1735         pop     %rdi
1736         pop     %rsi
1737         ret
1738 .size   se_handler,.-se_handler
1739
1740 .section        .pdata
1741 .align  4
1742         .rva    .LSEH_begin_gcm_gmult_4bit
1743         .rva    .LSEH_end_gcm_gmult_4bit
1744         .rva    .LSEH_info_gcm_gmult_4bit
1745
1746         .rva    .LSEH_begin_gcm_ghash_4bit
1747         .rva    .LSEH_end_gcm_ghash_4bit
1748         .rva    .LSEH_info_gcm_ghash_4bit
1749
1750         .rva    .LSEH_begin_gcm_init_clmul
1751         .rva    .LSEH_end_gcm_init_clmul
1752         .rva    .LSEH_info_gcm_init_clmul
1753
1754         .rva    .LSEH_begin_gcm_ghash_clmul
1755         .rva    .LSEH_end_gcm_ghash_clmul
1756         .rva    .LSEH_info_gcm_ghash_clmul
1757 ___
1758 $code.=<<___    if ($avx);
1759         .rva    .LSEH_begin_gcm_init_avx
1760         .rva    .LSEH_end_gcm_init_avx
1761         .rva    .LSEH_info_gcm_init_clmul
1762
1763         .rva    .LSEH_begin_gcm_ghash_avx
1764         .rva    .LSEH_end_gcm_ghash_avx
1765         .rva    .LSEH_info_gcm_ghash_clmul
1766 ___
1767 $code.=<<___;
1768 .section        .xdata
1769 .align  8
1770 .LSEH_info_gcm_gmult_4bit:
1771         .byte   9,0,0,0
1772         .rva    se_handler
1773         .rva    .Lgmult_prologue,.Lgmult_epilogue       # HandlerData
1774 .LSEH_info_gcm_ghash_4bit:
1775         .byte   9,0,0,0
1776         .rva    se_handler
1777         .rva    .Lghash_prologue,.Lghash_epilogue       # HandlerData
1778 .LSEH_info_gcm_init_clmul:
1779         .byte   0x01,0x08,0x03,0x00
1780         .byte   0x08,0x68,0x00,0x00     #movaps 0x00(rsp),xmm6
1781         .byte   0x04,0x22,0x00,0x00     #sub    rsp,0x18
1782 .LSEH_info_gcm_ghash_clmul:
1783         .byte   0x01,0x33,0x16,0x00
1784         .byte   0x33,0xf8,0x09,0x00     #movaps 0x90(rsp),xmm15
1785         .byte   0x2e,0xe8,0x08,0x00     #movaps 0x80(rsp),xmm14
1786         .byte   0x29,0xd8,0x07,0x00     #movaps 0x70(rsp),xmm13
1787         .byte   0x24,0xc8,0x06,0x00     #movaps 0x60(rsp),xmm12
1788         .byte   0x1f,0xb8,0x05,0x00     #movaps 0x50(rsp),xmm11
1789         .byte   0x1a,0xa8,0x04,0x00     #movaps 0x40(rsp),xmm10
1790         .byte   0x15,0x98,0x03,0x00     #movaps 0x30(rsp),xmm9
1791         .byte   0x10,0x88,0x02,0x00     #movaps 0x20(rsp),xmm8
1792         .byte   0x0c,0x78,0x01,0x00     #movaps 0x10(rsp),xmm7
1793         .byte   0x08,0x68,0x00,0x00     #movaps 0x00(rsp),xmm6
1794         .byte   0x04,0x01,0x15,0x00     #sub    rsp,0xa8
1795 ___
1796 }
1797 \f
1798 $code =~ s/\`([^\`]*)\`/eval($1)/gem;
1799
1800 print $code;
1801
1802 close STDOUT;