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