crypto/modes/asm/ghash-x86.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, May, 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 it
  14 # uses 256 bytes per-key table [+64/128 bytes fixed table]. It has two
  15 # code paths: vanilla x86 and vanilla MMX. Former will be executed on
  16 # 486 and Pentium, latter on all others. MMX GHASH features so called
  17 # "528B" variant of "4-bit" method utilizing additional 256+16 bytes
  18 # of per-key storage [+512 bytes shared table]. Performance results
  19 # are for streamed GHASH subroutine and are expressed in cycles per
  20 # processed byte, less is better:
  21 #
  22 #               gcc 2.95.3(*)   MMX assembler   x86 assembler
  23 #
  24 # Pentium       100/112(**)     -               50
  25 # PIII          63 /77          12.2            24
  26 # P4            96 /122         18.0            84(***)
  27 # Opteron       50 /71          10.1            30
  28 # Core2         54 /68          8.6             18
  29 #
  30 # (*)   gcc 3.4.x was observed to generate few percent slower code,
  31 #       which is one of reasons why 2.95.3 results were chosen,
  32 #       another reason is lack of 3.4.x results for older CPUs;
  33 # (**)  second number is result for code compiled with -fPIC flag,
  34 #       which is actually more relevant, because assembler code is
  35 #       position-independent;
  36 # (***) see comment in non-MMX routine for further details;
  37 #
  38 # To summarize, it's >2-5 times faster than gcc-generated code. To
  39 # anchor it to something else SHA1 assembler processes one byte in
  40 # 11-13 cycles on contemporary x86 cores. As for choice of MMX in
  41 # particular, see comment at the end of the file...
  42
  43 # May 2010
  44 #
  45 # Add PCLMULQDQ version performing at 2.13 cycles per processed byte.
  46 # The question is how close is it to theoretical limit? The pclmulqdq
  47 # instruction latency appears to be 14 cycles and there can't be more
  48 # than 2 of them executing at any given time. This means that single
  49 # Karatsuba multiplication would take 28 cycles *plus* few cycles for
  50 # pre- and post-processing. Then multiplication has to be followed by
  51 # modulo-reduction. Given that aggregated reduction method [see
  52 # "Carry-less Multiplication and Its Usage for Computing the GCM Mode"
  53 # white paper by Intel] allows you to perform reduction only once in
  54 # a while we can assume that asymptotic performance can be estimated
  55 # as (28+Tmod/Naggr)/16, where Tmod is time to perform reduction
  56 # and Naggr is the aggregation factor.
  57 #
  58 # Before we proceed to this implementation let's have closer look at
  59 # the best-performing code suggested by Intel in their white paper.
  60 # By tracing inter-register dependencies Tmod is estimated as ~19
  61 # cycles and Naggr is 4, resulting in 2.05 cycles per processed byte.
  62 # As implied, this is quite optimistic estimate, because it does not
  63 # account for Karatsuba pre- and post-processing, which for a single
  64 # multiplication is ~5 cycles. Unfortunately Intel does not provide
  65 # performance data for GHASH alone, only for fused GCM mode. But
  66 # we can estimate it by subtracting CTR performance result provided
  67 # in "AES Instruction Set" white paper: 3.54-1.38=2.16 cycles per
  68 # processed byte or 5% off the estimate. It should be noted though
  69 # that 3.54 is GCM result for 16KB block size, while 1.38 is CTR for
  70 # 1KB block size, meaning that real number is likely to be a bit
  71 # further from estimate.
  72 #
  73 # Moving on to the implementation in question. Tmod is estimated as
  74 # ~13 cycles and Naggr is 2, giving asymptotic performance of ...
  75 # 2.16. How is it possible that measured performance is better than
  76 # optimistic theoretical estimate? There is one thing Intel failed
  77 # to recognize. By fusing GHASH with CTR former's performance is
  78 # really limited to above (Tmul + Tmod/Naggr) equation. But if GHASH
  79 # procedure is detached, the modulo-reduction can be interleaved with
  80 # Naggr-1 multiplications and under ideal conditions even disappear
  81 # from the equation. So that optimistic theoretical estimate for this
  82 # implementation is ... 28/16=1.75, and not 2.16. Well, it's probably
  83 # way too optimistic, at least for such small Naggr. I'd argue that
  84 # (28+Tproc/Naggr), where Tproc is time required for Karatsuba pre-
  85 # and post-processing, is more realistic estimate. In this case it
  86 # gives ... 1.91 cycles per processed byte. Or in other words,
  87 # depending on how well we can interleave reduction and one of the
  88 # two multiplications the performance should be betwen 1.91 and 2.16.
  89 # As already mentioned, this implementation processes one byte [out
  90 # of 1KB buffer] in 2.13 cycles, while x86_64 counterpart - in 2.07.
  91 # x86_64 performance is better, because larger register bank allows
  92 # to interleave reduction and multiplication better.
  93 #
  94 # Does it make sense to increase Naggr? To start with it's virtually
  95 # impossible in 32-bit mode, because of limited register bank
  96 # capacity. Otherwise improvement has to be weighed agiainst slower
  97 # setup, as well as code size and complexity increase. As even
  98 # optimistic estimate doesn't promise 30% performance improvement,
  99 # there are currently no plans to increase Naggr.
 100 #
 101 # Special thanks to David Woodhouse <dwmw2@infradead.org> for
 102 # providing access to a Westmere-based system on behalf of Intel
 103 # Open Source Technology Centre.
 104
 105 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 106 push(@INC,"${dir}","${dir}../../perlasm");
 107 require "x86asm.pl";
 108
 109 &asm_init($ARGV[0],"ghash-x86.pl",$x86only = $ARGV[$#ARGV] eq "386");
 110
 111 $sse2=0;
 112 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
 113
 114 ($Zhh,$Zhl,$Zlh,$Zll) = ("ebp","edx","ecx","ebx");
 115 $inp  = "edi";
 116 $Htbl = "esi";
 117 \f
 118 $unroll = 0;    # Affects x86 loop. Folded loop performs ~7% worse
 119                 # than unrolled, which has to be weighted against
 120                 # 2.5x x86-specific code size reduction.
 121
 122 sub x86_loop {
 123     my $off = shift;
 124     my $rem = "eax";
 125
 126         &mov    ($Zhh,&DWP(4,$Htbl,$Zll));
 127         &mov    ($Zhl,&DWP(0,$Htbl,$Zll));
 128         &mov    ($Zlh,&DWP(12,$Htbl,$Zll));
 129         &mov    ($Zll,&DWP(8,$Htbl,$Zll));
 130         &xor    ($rem,$rem);    # avoid partial register stalls on PIII
 131
 132         # shrd practically kills P4, 2.5x deterioration, but P4 has
 133         # MMX code-path to execute. shrd runs tad faster [than twice
 134         # the shifts, move's and or's] on pre-MMX Pentium (as well as
 135         # PIII and Core2), *but* minimizes code size, spares register
 136         # and thus allows to fold the loop...
 137         if (!$unroll) {
 138         my $cnt = $inp;
 139         &mov    ($cnt,15);
 140         &jmp    (&label("x86_loop"));
 141         &set_label("x86_loop",16);
 142             for($i=1;$i<=2;$i++) {
 143                 &mov    (&LB($rem),&LB($Zll));
 144                 &shrd   ($Zll,$Zlh,4);
 145                 &and    (&LB($rem),0xf);
 146                 &shrd   ($Zlh,$Zhl,4);
 147                 &shrd   ($Zhl,$Zhh,4);
 148                 &shr    ($Zhh,4);
 149                 &xor    ($Zhh,&DWP($off+16,"esp",$rem,4));
 150
 151                 &mov    (&LB($rem),&BP($off,"esp",$cnt));
 152                 if ($i&1) {
 153                         &and    (&LB($rem),0xf0);
 154                 } else {
 155                         &shl    (&LB($rem),4);
 156                 }
 157
 158                 &xor    ($Zll,&DWP(8,$Htbl,$rem));
 159                 &xor    ($Zlh,&DWP(12,$Htbl,$rem));
 160                 &xor    ($Zhl,&DWP(0,$Htbl,$rem));
 161                 &xor    ($Zhh,&DWP(4,$Htbl,$rem));
 162
 163                 if ($i&1) {
 164                         &dec    ($cnt);
 165                         &js     (&label("x86_break"));
 166                 } else {
 167                         &jmp    (&label("x86_loop"));
 168                 }
 169             }
 170         &set_label("x86_break",16);
 171         } else {
 172             for($i=1;$i<32;$i++) {
 173                 &comment($i);
 174                 &mov    (&LB($rem),&LB($Zll));
 175                 &shrd   ($Zll,$Zlh,4);
 176                 &and    (&LB($rem),0xf);
 177                 &shrd   ($Zlh,$Zhl,4);
 178                 &shrd   ($Zhl,$Zhh,4);
 179                 &shr    ($Zhh,4);
 180                 &xor    ($Zhh,&DWP($off+16,"esp",$rem,4));
 181
 182                 if ($i&1) {
 183                         &mov    (&LB($rem),&BP($off+15-($i>>1),"esp"));
 184                         &and    (&LB($rem),0xf0);
 185                 } else {
 186                         &mov    (&LB($rem),&BP($off+15-($i>>1),"esp"));
 187                         &shl    (&LB($rem),4);
 188                 }
 189
 190                 &xor    ($Zll,&DWP(8,$Htbl,$rem));
 191                 &xor    ($Zlh,&DWP(12,$Htbl,$rem));
 192                 &xor    ($Zhl,&DWP(0,$Htbl,$rem));
 193                 &xor    ($Zhh,&DWP(4,$Htbl,$rem));
 194             }
 195         }
 196         &bswap  ($Zll);
 197         &bswap  ($Zlh);
 198         &bswap  ($Zhl);
 199         if (!$x86only) {
 200                 &bswap  ($Zhh);
 201         } else {
 202                 &mov    ("eax",$Zhh);
 203                 &bswap  ("eax");
 204                 &mov    ($Zhh,"eax");
 205         }
 206 }
 207
 208 if ($unroll) {
 209     &function_begin_B("_x86_gmult_4bit_inner");
 210         &x86_loop(4);
 211         &ret    ();
 212     &function_end_B("_x86_gmult_4bit_inner");
 213 }
 214
 215 sub deposit_rem_4bit {
 216     my $bias = shift;
 217
 218         &mov    (&DWP($bias+0, "esp"),0x0000<<16);
 219         &mov    (&DWP($bias+4, "esp"),0x1C20<<16);
 220         &mov    (&DWP($bias+8, "esp"),0x3840<<16);
 221         &mov    (&DWP($bias+12,"esp"),0x2460<<16);
 222         &mov    (&DWP($bias+16,"esp"),0x7080<<16);
 223         &mov    (&DWP($bias+20,"esp"),0x6CA0<<16);
 224         &mov    (&DWP($bias+24,"esp"),0x48C0<<16);
 225         &mov    (&DWP($bias+28,"esp"),0x54E0<<16);
 226         &mov    (&DWP($bias+32,"esp"),0xE100<<16);
 227         &mov    (&DWP($bias+36,"esp"),0xFD20<<16);
 228         &mov    (&DWP($bias+40,"esp"),0xD940<<16);
 229         &mov    (&DWP($bias+44,"esp"),0xC560<<16);
 230         &mov    (&DWP($bias+48,"esp"),0x9180<<16);
 231         &mov    (&DWP($bias+52,"esp"),0x8DA0<<16);
 232         &mov    (&DWP($bias+56,"esp"),0xA9C0<<16);
 233         &mov    (&DWP($bias+60,"esp"),0xB5E0<<16);
 234 }
 235 \f
 236 $suffix = $x86only ? "" : "_x86";
 237
 238 &function_begin("gcm_gmult_4bit".$suffix);
 239         &stack_push(16+4+1);                    # +1 for stack alignment
 240         &mov    ($inp,&wparam(0));              # load Xi
 241         &mov    ($Htbl,&wparam(1));             # load Htable
 242
 243         &mov    ($Zhh,&DWP(0,$inp));            # load Xi[16]
 244         &mov    ($Zhl,&DWP(4,$inp));
 245         &mov    ($Zlh,&DWP(8,$inp));
 246         &mov    ($Zll,&DWP(12,$inp));
 247
 248         &deposit_rem_4bit(16);
 249
 250         &mov    (&DWP(0,"esp"),$Zhh);           # copy Xi[16] on stack
 251         &mov    (&DWP(4,"esp"),$Zhl);
 252         &mov    (&DWP(8,"esp"),$Zlh);
 253         &mov    (&DWP(12,"esp"),$Zll);
 254         &shr    ($Zll,20);
 255         &and    ($Zll,0xf0);
 256
 257         if ($unroll) {
 258                 &call   ("_x86_gmult_4bit_inner");
 259         } else {
 260                 &x86_loop(0);
 261                 &mov    ($inp,&wparam(0));
 262         }
 263
 264         &mov    (&DWP(12,$inp),$Zll);
 265         &mov    (&DWP(8,$inp),$Zlh);
 266         &mov    (&DWP(4,$inp),$Zhl);
 267         &mov    (&DWP(0,$inp),$Zhh);
 268         &stack_pop(16+4+1);
 269 &function_end("gcm_gmult_4bit".$suffix);
 270
 271 &function_begin("gcm_ghash_4bit".$suffix);
 272         &stack_push(16+4+1);                    # +1 for 64-bit alignment
 273         &mov    ($Zll,&wparam(0));              # load Xi
 274         &mov    ($Htbl,&wparam(1));             # load Htable
 275         &mov    ($inp,&wparam(2));              # load in
 276         &mov    ("ecx",&wparam(3));             # load len
 277         &add    ("ecx",$inp);
 278         &mov    (&wparam(3),"ecx");
 279
 280         &mov    ($Zhh,&DWP(0,$Zll));            # load Xi[16]
 281         &mov    ($Zhl,&DWP(4,$Zll));
 282         &mov    ($Zlh,&DWP(8,$Zll));
 283         &mov    ($Zll,&DWP(12,$Zll));
 284
 285         &deposit_rem_4bit(16);
 286
 287     &set_label("x86_outer_loop",16);
 288         &xor    ($Zll,&DWP(12,$inp));           # xor with input
 289         &xor    ($Zlh,&DWP(8,$inp));
 290         &xor    ($Zhl,&DWP(4,$inp));
 291         &xor    ($Zhh,&DWP(0,$inp));
 292         &mov    (&DWP(12,"esp"),$Zll);          # dump it on stack
 293         &mov    (&DWP(8,"esp"),$Zlh);
 294         &mov    (&DWP(4,"esp"),$Zhl);
 295         &mov    (&DWP(0,"esp"),$Zhh);
 296
 297         &shr    ($Zll,20);
 298         &and    ($Zll,0xf0);
 299
 300         if ($unroll) {
 301                 &call   ("_x86_gmult_4bit_inner");
 302         } else {
 303                 &x86_loop(0);
 304                 &mov    ($inp,&wparam(2));
 305         }
 306         &lea    ($inp,&DWP(16,$inp));
 307         &cmp    ($inp,&wparam(3));
 308         &mov    (&wparam(2),$inp)       if (!$unroll);
 309         &jb     (&label("x86_outer_loop"));
 310
 311         &mov    ($inp,&wparam(0));      # load Xi
 312         &mov    (&DWP(12,$inp),$Zll);
 313         &mov    (&DWP(8,$inp),$Zlh);
 314         &mov    (&DWP(4,$inp),$Zhl);
 315         &mov    (&DWP(0,$inp),$Zhh);
 316         &stack_pop(16+4+1);
 317 &function_end("gcm_ghash_4bit".$suffix);
 318 \f
 319 if (!$x86only) {{{
 320
 321 &static_label("rem_4bit");
 322
 323 if (0) {{       # "May" MMX version is kept for reference...
 324
 325 $S=12;          # shift factor for rem_4bit
 326
 327 &function_begin_B("_mmx_gmult_4bit_inner");
 328 # MMX version performs 3.5 times better on P4 (see comment in non-MMX
 329 # routine for further details), 100% better on Opteron, ~70% better
 330 # on Core2 and PIII... In other words effort is considered to be well
 331 # spent... Since initial release the loop was unrolled in order to
 332 # "liberate" register previously used as loop counter. Instead it's
 333 # used to optimize critical path in 'Z.hi ^= rem_4bit[Z.lo&0xf]'.
 334 # The path involves move of Z.lo from MMX to integer register,
 335 # effective address calculation and finally merge of value to Z.hi.
 336 # Reference to rem_4bit is scheduled so late that I had to >>4
 337 # rem_4bit elements. This resulted in 20-45% procent improvement
 338 # on contemporary µ-archs.
 339 {
 340     my $cnt;
 341     my $rem_4bit = "eax";
 342     my @rem = ($Zhh,$Zll);
 343     my $nhi = $Zhl;
 344     my $nlo = $Zlh;
 345
 346     my ($Zlo,$Zhi) = ("mm0","mm1");
 347     my $tmp = "mm2";
 348
 349         &xor    ($nlo,$nlo);    # avoid partial register stalls on PIII
 350         &mov    ($nhi,$Zll);
 351         &mov    (&LB($nlo),&LB($nhi));
 352         &shl    (&LB($nlo),4);
 353         &and    ($nhi,0xf0);
 354         &movq   ($Zlo,&QWP(8,$Htbl,$nlo));
 355         &movq   ($Zhi,&QWP(0,$Htbl,$nlo));
 356         &movd   ($rem[0],$Zlo);
 357
 358         for ($cnt=28;$cnt>=-2;$cnt--) {
 359             my $odd = $cnt&1;
 360             my $nix = $odd ? $nlo : $nhi;
 361
 362                 &shl    (&LB($nlo),4)                   if ($odd);
 363                 &psrlq  ($Zlo,4);
 364                 &movq   ($tmp,$Zhi);
 365                 &psrlq  ($Zhi,4);
 366                 &pxor   ($Zlo,&QWP(8,$Htbl,$nix));
 367                 &mov    (&LB($nlo),&BP($cnt/2,$inp))    if (!$odd && $cnt>=0);
 368                 &psllq  ($tmp,60);
 369                 &and    ($nhi,0xf0)                     if ($odd);
 370                 &pxor   ($Zhi,&QWP(0,$rem_4bit,$rem[1],8)) if ($cnt<28);
 371                 &and    ($rem[0],0xf);
 372                 &pxor   ($Zhi,&QWP(0,$Htbl,$nix));
 373                 &mov    ($nhi,$nlo)                     if (!$odd && $cnt>=0);
 374                 &movd   ($rem[1],$Zlo);
 375                 &pxor   ($Zlo,$tmp);
 376
 377                 push    (@rem,shift(@rem));             # "rotate" registers
 378         }
 379
 380         &mov    ($inp,&DWP(4,$rem_4bit,$rem[1],8));     # last rem_4bit[rem]
 381
 382         &psrlq  ($Zlo,32);      # lower part of Zlo is already there
 383         &movd   ($Zhl,$Zhi);
 384         &psrlq  ($Zhi,32);
 385         &movd   ($Zlh,$Zlo);
 386         &movd   ($Zhh,$Zhi);
 387         &shl    ($inp,4);       # compensate for rem_4bit[i] being >>4
 388
 389         &bswap  ($Zll);
 390         &bswap  ($Zhl);
 391         &bswap  ($Zlh);
 392         &xor    ($Zhh,$inp);
 393         &bswap  ($Zhh);
 394
 395         &ret    ();
 396 }
 397 &function_end_B("_mmx_gmult_4bit_inner");
 398
 399 &function_begin("gcm_gmult_4bit_mmx");
 400         &mov    ($inp,&wparam(0));      # load Xi
 401         &mov    ($Htbl,&wparam(1));     # load Htable
 402
 403         &call   (&label("pic_point"));
 404         &set_label("pic_point");
 405         &blindpop("eax");
 406         &lea    ("eax",&DWP(&label("rem_4bit")."-".&label("pic_point"),"eax"));
 407
 408         &movz   ($Zll,&BP(15,$inp));
 409
 410         &call   ("_mmx_gmult_4bit_inner");
 411
 412         &mov    ($inp,&wparam(0));      # load Xi
 413         &emms   ();
 414         &mov    (&DWP(12,$inp),$Zll);
 415         &mov    (&DWP(4,$inp),$Zhl);
 416         &mov    (&DWP(8,$inp),$Zlh);
 417         &mov    (&DWP(0,$inp),$Zhh);
 418 &function_end("gcm_gmult_4bit_mmx");
 419 \f
 420 # Streamed version performs 20% better on P4, 7% on Opteron,
 421 # 10% on Core2 and PIII...
 422 &function_begin("gcm_ghash_4bit_mmx");
 423         &mov    ($Zhh,&wparam(0));      # load Xi
 424         &mov    ($Htbl,&wparam(1));     # load Htable
 425         &mov    ($inp,&wparam(2));      # load in
 426         &mov    ($Zlh,&wparam(3));      # load len
 427
 428         &call   (&label("pic_point"));
 429         &set_label("pic_point");
 430         &blindpop("eax");
 431         &lea    ("eax",&DWP(&label("rem_4bit")."-".&label("pic_point"),"eax"));
 432
 433         &add    ($Zlh,$inp);
 434         &mov    (&wparam(3),$Zlh);      # len to point at the end of input
 435         &stack_push(4+1);               # +1 for stack alignment
 436
 437         &mov    ($Zll,&DWP(12,$Zhh));   # load Xi[16]
 438         &mov    ($Zhl,&DWP(4,$Zhh));
 439         &mov    ($Zlh,&DWP(8,$Zhh));
 440         &mov    ($Zhh,&DWP(0,$Zhh));
 441         &jmp    (&label("mmx_outer_loop"));
 442
 443     &set_label("mmx_outer_loop",16);
 444         &xor    ($Zll,&DWP(12,$inp));
 445         &xor    ($Zhl,&DWP(4,$inp));
 446         &xor    ($Zlh,&DWP(8,$inp));
 447         &xor    ($Zhh,&DWP(0,$inp));
 448         &mov    (&wparam(2),$inp);
 449         &mov    (&DWP(12,"esp"),$Zll);
 450         &mov    (&DWP(4,"esp"),$Zhl);
 451         &mov    (&DWP(8,"esp"),$Zlh);
 452         &mov    (&DWP(0,"esp"),$Zhh);
 453
 454         &mov    ($inp,"esp");
 455         &shr    ($Zll,24);
 456
 457         &call   ("_mmx_gmult_4bit_inner");
 458
 459         &mov    ($inp,&wparam(2));
 460         &lea    ($inp,&DWP(16,$inp));
 461         &cmp    ($inp,&wparam(3));
 462         &jb     (&label("mmx_outer_loop"));
 463
 464         &mov    ($inp,&wparam(0));      # load Xi
 465         &emms   ();
 466         &mov    (&DWP(12,$inp),$Zll);
 467         &mov    (&DWP(4,$inp),$Zhl);
 468         &mov    (&DWP(8,$inp),$Zlh);
 469         &mov    (&DWP(0,$inp),$Zhh);
 470
 471         &stack_pop(4+1);
 472 &function_end("gcm_ghash_4bit_mmx");
 473 \f
 474 }} else {{      # "June" MMX version...
 475                 # ... has "April" gcm_gmult_4bit_mmx with folded loop.
 476                 # This is done to conserve code size...
 477 $S=16;          # shift factor for rem_4bit
 478
 479 sub mmx_loop() {
 480 # MMX version performs 2.8 times better on P4 (see comment in non-MMX
 481 # routine for further details), 40% better on Opteron and Core2, 50%
 482 # better on PIII... In other words effort is considered to be well
 483 # spent...
 484     my $inp = shift;
 485     my $rem_4bit = shift;
 486     my $cnt = $Zhh;
 487     my $nhi = $Zhl;
 488     my $nlo = $Zlh;
 489     my $rem = $Zll;
 490
 491     my ($Zlo,$Zhi) = ("mm0","mm1");
 492     my $tmp = "mm2";
 493
 494         &xor    ($nlo,$nlo);    # avoid partial register stalls on PIII
 495         &mov    ($nhi,$Zll);
 496         &mov    (&LB($nlo),&LB($nhi));
 497         &mov    ($cnt,14);
 498         &shl    (&LB($nlo),4);
 499         &and    ($nhi,0xf0);
 500         &movq   ($Zlo,&QWP(8,$Htbl,$nlo));
 501         &movq   ($Zhi,&QWP(0,$Htbl,$nlo));
 502         &movd   ($rem,$Zlo);
 503         &jmp    (&label("mmx_loop"));
 504
 505     &set_label("mmx_loop",16);
 506         &psrlq  ($Zlo,4);
 507         &and    ($rem,0xf);
 508         &movq   ($tmp,$Zhi);
 509         &psrlq  ($Zhi,4);
 510         &pxor   ($Zlo,&QWP(8,$Htbl,$nhi));
 511         &mov    (&LB($nlo),&BP(0,$inp,$cnt));
 512         &psllq  ($tmp,60);
 513         &pxor   ($Zhi,&QWP(0,$rem_4bit,$rem,8));
 514         &dec    ($cnt);
 515         &movd   ($rem,$Zlo);
 516         &pxor   ($Zhi,&QWP(0,$Htbl,$nhi));
 517         &mov    ($nhi,$nlo);
 518         &pxor   ($Zlo,$tmp);
 519         &js     (&label("mmx_break"));
 520
 521         &shl    (&LB($nlo),4);
 522         &and    ($rem,0xf);
 523         &psrlq  ($Zlo,4);
 524         &and    ($nhi,0xf0);
 525         &movq   ($tmp,$Zhi);
 526         &psrlq  ($Zhi,4);
 527         &pxor   ($Zlo,&QWP(8,$Htbl,$nlo));
 528         &psllq  ($tmp,60);
 529         &pxor   ($Zhi,&QWP(0,$rem_4bit,$rem,8));
 530         &movd   ($rem,$Zlo);
 531         &pxor   ($Zhi,&QWP(0,$Htbl,$nlo));
 532         &pxor   ($Zlo,$tmp);
 533         &jmp    (&label("mmx_loop"));
 534
 535     &set_label("mmx_break",16);
 536         &shl    (&LB($nlo),4);
 537         &and    ($rem,0xf);
 538         &psrlq  ($Zlo,4);
 539         &and    ($nhi,0xf0);
 540         &movq   ($tmp,$Zhi);
 541         &psrlq  ($Zhi,4);
 542         &pxor   ($Zlo,&QWP(8,$Htbl,$nlo));
 543         &psllq  ($tmp,60);
 544         &pxor   ($Zhi,&QWP(0,$rem_4bit,$rem,8));
 545         &movd   ($rem,$Zlo);
 546         &pxor   ($Zhi,&QWP(0,$Htbl,$nlo));
 547         &pxor   ($Zlo,$tmp);
 548
 549         &psrlq  ($Zlo,4);
 550         &and    ($rem,0xf);
 551         &movq   ($tmp,$Zhi);
 552         &psrlq  ($Zhi,4);
 553         &pxor   ($Zlo,&QWP(8,$Htbl,$nhi));
 554         &psllq  ($tmp,60);
 555         &pxor   ($Zhi,&QWP(0,$rem_4bit,$rem,8));
 556         &movd   ($rem,$Zlo);
 557         &pxor   ($Zhi,&QWP(0,$Htbl,$nhi));
 558         &pxor   ($Zlo,$tmp);
 559
 560         &psrlq  ($Zlo,32);      # lower part of Zlo is already there
 561         &movd   ($Zhl,$Zhi);
 562         &psrlq  ($Zhi,32);
 563         &movd   ($Zlh,$Zlo);
 564         &movd   ($Zhh,$Zhi);
 565
 566         &bswap  ($Zll);
 567         &bswap  ($Zhl);
 568         &bswap  ($Zlh);
 569         &bswap  ($Zhh);
 570 }
 571
 572 &function_begin("gcm_gmult_4bit_mmx");
 573         &mov    ($inp,&wparam(0));      # load Xi
 574         &mov    ($Htbl,&wparam(1));     # load Htable
 575
 576         &call   (&label("pic_point"));
 577         &set_label("pic_point");
 578         &blindpop("eax");
 579         &lea    ("eax",&DWP(&label("rem_4bit")."-".&label("pic_point"),"eax"));
 580
 581         &movz   ($Zll,&BP(15,$inp));
 582
 583         &mmx_loop($inp,"eax");
 584
 585         &emms   ();
 586         &mov    (&DWP(12,$inp),$Zll);
 587         &mov    (&DWP(4,$inp),$Zhl);
 588         &mov    (&DWP(8,$inp),$Zlh);
 589         &mov    (&DWP(0,$inp),$Zhh);
 590 &function_end("gcm_gmult_4bit_mmx");
 591 \f
 592 ######################################################################
 593 # Below subroutine is "528B" variant of "4-bit" GCM GHASH function
 594 # (see gcm128.c for details). It provides further 20-40% performance
 595 # improvement over *previous* version of this module.
 596
 597 &static_label("rem_8bit");
 598
 599 &function_begin("gcm_ghash_4bit_mmx");
 600 { my ($Zlo,$Zhi) = ("mm7","mm6");
 601   my $rem_8bit = "esi";
 602   my $Htbl = "ebx";
 603
 604     # parameter block
 605     &mov        ("eax",&wparam(0));             # Xi
 606     &mov        ("ebx",&wparam(1));             # Htable
 607     &mov        ("ecx",&wparam(2));             # inp
 608     &mov        ("edx",&wparam(3));             # len
 609     &mov        ("ebp","esp");                  # original %esp
 610     &call       (&label("pic_point"));
 611     &set_label  ("pic_point");
 612     &blindpop   ($rem_8bit);
 613     &lea        ($rem_8bit,&DWP(&label("rem_8bit")."-".&label("pic_point"),$rem_8bit));
 614
 615     &sub        ("esp",512+16+16);              # allocate stack frame...
 616     &and        ("esp",-64);                    # ...and align it
 617     &sub        ("esp",16);                     # place for (u8)(H[]<<4)
 618
 619     &add        ("edx","ecx");                  # pointer to the end of input
 620     &mov        (&DWP(528+16+0,"esp"),"eax");   # save Xi
 621     &mov        (&DWP(528+16+8,"esp"),"edx");   # save inp+len
 622     &mov        (&DWP(528+16+12,"esp"),"ebp");  # save original %esp
 623
 624     { my @lo  = ("mm0","mm1","mm2");
 625       my @hi  = ("mm3","mm4","mm5");
 626       my @tmp = ("mm6","mm7");
 627       my $off1=0,$off2=0,$i;
 628
 629       &add      ($Htbl,128);                    # optimize for size
 630       &lea      ("edi",&DWP(16+128,"esp"));
 631       &lea      ("ebp",&DWP(16+256+128,"esp"));
 632
 633       # decompose Htable (low and high parts are kept separately),
 634       # generate Htable>>4, save to stack...
 635       for ($i=0;$i<18;$i++) {
 636
 637         &mov    ("edx",&DWP(16*$i+8-128,$Htbl))         if ($i<16);
 638         &movq   ($lo[0],&QWP(16*$i+8-128,$Htbl))        if ($i<16);
 639         &psllq  ($tmp[1],60)                            if ($i>1);
 640         &movq   ($hi[0],&QWP(16*$i+0-128,$Htbl))        if ($i<16);
 641         &por    ($lo[2],$tmp[1])                        if ($i>1);
 642         &movq   (&QWP($off1-128,"edi"),$lo[1])          if ($i>0 && $i<17);
 643         &psrlq  ($lo[1],4)                              if ($i>0 && $i<17);
 644         &movq   (&QWP($off1,"edi"),$hi[1])              if ($i>0 && $i<17);
 645         &movq   ($tmp[0],$hi[1])                        if ($i>0 && $i<17);
 646         &movq   (&QWP($off2-128,"ebp"),$lo[2])          if ($i>1);
 647         &psrlq  ($hi[1],4)                              if ($i>0 && $i<17);
 648         &movq   (&QWP($off2,"ebp"),$hi[2])              if ($i>1);
 649         &shl    ("edx",4)                               if ($i<16);
 650         &mov    (&BP($i,"esp"),&LB("edx"))              if ($i<16);
 651
 652         unshift (@lo,pop(@lo));                 # "rotate" registers
 653         unshift (@hi,pop(@hi));
 654         unshift (@tmp,pop(@tmp));
 655         $off1 += 8      if ($i>0);
 656         $off2 += 8      if ($i>1);
 657       }
 658     }
 659
 660     &movq       ($Zhi,&QWP(0,"eax"));
 661     &mov        ("ebx",&DWP(8,"eax"));
 662     &mov        ("edx",&DWP(12,"eax"));         # load Xi
 663
 664 &set_label("outer",16);
 665   { my $nlo = "eax";
 666     my $dat = "edx";
 667     my @nhi = ("edi","ebp");
 668     my @rem = ("ebx","ecx");
 669     my @red = ("mm0","mm1","mm2");
 670     my $tmp = "mm3";
 671
 672     &xor        ($dat,&DWP(12,"ecx"));          # merge input
 673     &xor        ("ebx",&DWP(8,"ecx"));
 674     &pxor       ($Zhi,&QWP(0,"ecx"));
 675     &lea        ("ecx",&DWP(16,"ecx"));         # inp+=16
 676     #&mov       (&DWP(528+12,"esp"),$dat);      # save inp^Xi
 677     &mov        (&DWP(528+8,"esp"),"ebx");
 678     &movq       (&QWP(528+0,"esp"),$Zhi);
 679     &mov        (&DWP(528+16+4,"esp"),"ecx");   # save inp
 680
 681     &xor        ($nlo,$nlo);
 682     &rol        ($dat,8);
 683     &mov        (&LB($nlo),&LB($dat));
 684     &mov        ($nhi[1],$nlo);
 685     &and        (&LB($nlo),0x0f);
 686     &shr        ($nhi[1],4);
 687     &pxor       ($red[0],$red[0]);
 688     &rol        ($dat,8);                               # next byte
 689     &pxor       ($red[1],$red[1]);
 690     &pxor       ($red[2],$red[2]);
 691
 692     # Just like in "May" verson modulo-schedule for critical path in
 693     # 'Z.hi ^= rem_8bit[Z.lo&0xff^((u8)H[nhi]<<4)]<<48'. Final xor
 694     # is scheduled so late that rem_8bit is shifted *right* by 16,
 695     # which is why last argument to pinsrw is 2, which corresponds to
 696     # <<32...
 697     for ($j=11,$i=0;$i<15;$i++) {
 698
 699       if ($i>0) {
 700         &pxor   ($Zlo,&QWP(16,"esp",$nlo,8));           # Z^=H[nlo]
 701         &rol    ($dat,8);                               # next byte
 702         &pxor   ($Zhi,&QWP(16+128,"esp",$nlo,8));
 703
 704         &pxor   ($Zlo,$tmp);
 705         &pxor   ($Zhi,&QWP(16+256+128,"esp",$nhi[0],8));
 706         &xor    (&LB($rem[1]),&BP(0,"esp",$nhi[0]));    # rem^H[nhi]<<4
 707       } else {
 708         &movq   ($Zlo,&QWP(16,"esp",$nlo,8));
 709         &movq   ($Zhi,&QWP(16+128,"esp",$nlo,8));
 710       }
 711
 712         &mov    (&LB($nlo),&LB($dat));
 713         &mov    ($dat,&DWP(528+$j,"esp"))       if (--$j%4==0);
 714
 715         &movd   ($rem[0],$Zlo);
 716         &movz   ($rem[1],&LB($rem[1]))          if ($i>0);
 717         &psrlq  ($Zlo,8);
 718
 719         &movq   ($tmp,$Zhi);
 720         &mov    ($nhi[0],$nlo);
 721         &psrlq  ($Zhi,8);
 722
 723         &pxor   ($Zlo,&QWP(16+256+0,"esp",$nhi[1],8));  # Z^=H[nhi]>>4
 724         &and    (&LB($nlo),0x0f);
 725         &psllq  ($tmp,56);
 726
 727         &pxor   ($Zhi,$red[1])                          if ($i>1);
 728         &shr    ($nhi[0],4);
 729         &pinsrw ($red[0],&WP(0,$rem_8bit,$rem[1],2),2)  if ($i>0);
 730
 731         unshift (@red,pop(@red));                       # "rotate" registers
 732         unshift (@rem,pop(@rem));
 733         unshift (@nhi,pop(@nhi));
 734     }
 735
 736     &pxor       ($Zlo,&QWP(16,"esp",$nlo,8));           # Z^=H[nlo]
 737     &pxor       ($Zhi,&QWP(16+128,"esp",$nlo,8));
 738     &xor        (&LB($rem[1]),&BP(0,"esp",$nhi[0]));    #$rem[0]);                      # rem^H[nhi]<<4
 739
 740     &pxor       ($Zlo,$tmp);
 741     &pxor       ($Zhi,&QWP(16+256+128,"esp",$nhi[0],8));
 742     &movz       ($rem[1],&LB($rem[1]));
 743
 744     &pxor       ($red[2],$red[2]);                      # clear 2nd word
 745     &psllq      ($red[1],4);
 746
 747     &movd       ($rem[0],$Zlo);
 748     &psrlq      ($Zlo,4);
 749
 750     &movq       ($tmp,$Zhi);
 751     &psrlq      ($Zhi,4);
 752     &shl        ($rem[0],4);
 753
 754     &pxor       ($Zlo,&QWP(16,"esp",$nhi[1],8));        # Z^=H[nhi]
 755     &psllq      ($tmp,60);
 756     &movz       ($rem[0],&LB($rem[0]));
 757
 758     &pxor       ($Zlo,$tmp);
 759     &pxor       ($Zhi,&QWP(16+128,"esp",$nhi[1],8));
 760
 761     &pinsrw     ($red[0],&WP(0,$rem_8bit,$rem[1],2),2);
 762     &pxor       ($Zhi,$red[1]);
 763
 764     &movd       ($dat,$Zlo);
 765     &pinsrw     ($red[2],&WP(0,$rem_8bit,$rem[0],2),3);
 766
 767     &psllq      ($red[0],12);
 768     &pxor       ($Zhi,$red[0]);
 769     &psrlq      ($Zlo,32);
 770     &pxor       ($Zhi,$red[2]);
 771
 772     &mov        ("ecx",&DWP(528+16+4,"esp"));   # restore inp
 773     &movd       ("ebx",$Zlo);
 774     &movq       ($tmp,$Zhi);                    # 01234567
 775     &psllw      ($Zhi,8);                       # 1.3.5.7.
 776     &psrlw      ($tmp,8);                       # .0.2.4.6
 777     &por        ($Zhi,$tmp);                    # 10325476
 778     &bswap      ($dat);
 779     &pshufw     ($Zhi,$Zhi,0b00011011);         # 76543210
 780     &bswap      ("ebx");
 781
 782     &cmp        ("ecx",&DWP(528+16+8,"esp"));   # are we done?
 783     &jne        (&label("outer"));
 784   }
 785
 786     &mov        ("eax",&DWP(528+16+0,"esp"));   # restore Xi
 787     &mov        (&DWP(12,"eax"),"edx");
 788     &mov        (&DWP(8,"eax"),"ebx");
 789     &movq       (&QWP(0,"eax"),$Zhi);
 790
 791     &mov        ("esp",&DWP(528+16+12,"esp"));  # restore original %esp
 792     &emms       ();
 793 }
 794 &function_end("gcm_ghash_4bit_mmx");
 795 }}
 796 \f
 797 if ($sse2) {{
 798 ######################################################################
 799 # PCLMULQDQ version.
 800
 801 $Xip="eax";
 802 $Htbl="edx";
 803 $const="ecx";
 804 $inp="esi";
 805 $len="ebx";
 806
 807 ($Xi,$Xhi)=("xmm0","xmm1");     $Hkey="xmm2";
 808 ($T1,$T2,$T3)=("xmm3","xmm4","xmm5");
 809 ($Xn,$Xhn)=("xmm6","xmm7");
 810
 811 &static_label("bswap");
 812
 813 sub clmul64x64_T2 {     # minimal "register" pressure
 814 my ($Xhi,$Xi,$Hkey)=@_;
 815
 816         &movdqa         ($Xhi,$Xi);             #
 817         &pshufd         ($T1,$Xi,0b01001110);
 818         &pshufd         ($T2,$Hkey,0b01001110);
 819         &pxor           ($T1,$Xi);              #
 820         &pxor           ($T2,$Hkey);
 821
 822         &pclmulqdq      ($Xi,$Hkey,0x00);       #######
 823         &pclmulqdq      ($Xhi,$Hkey,0x11);      #######
 824         &pclmulqdq      ($T1,$T2,0x00);         #######
 825         &pxor           ($T1,$Xi);              #
 826         &pxor           ($T1,$Xhi);             #
 827
 828         &movdqa         ($T2,$T1);              #
 829         &psrldq         ($T1,8);
 830         &pslldq         ($T2,8);                #
 831         &pxor           ($Xhi,$T1);
 832         &pxor           ($Xi,$T2);              #
 833 }
 834
 835 sub clmul64x64_T3 {
 836 # Even though this subroutine offers visually better ILP, it
 837 # was empirically found to be a tad slower than above version.
 838 # At least in gcm_ghash_clmul context. But it's just as well,
 839 # because loop modulo-scheduling is possible only thanks to
 840 # minimized "register" pressure...
 841 my ($Xhi,$Xi,$Hkey)=@_;
 842
 843         &movdqa         ($T1,$Xi);              #
 844         &movdqa         ($Xhi,$Xi);
 845         &pclmulqdq      ($Xi,$Hkey,0x00);       #######
 846         &pclmulqdq      ($Xhi,$Hkey,0x11);      #######
 847         &pshufd         ($T2,$T1,0b01001110);   #
 848         &pshufd         ($T3,$Hkey,0b01001110);
 849         &pxor           ($T2,$T1);              #
 850         &pxor           ($T3,$Hkey);
 851         &pclmulqdq      ($T2,$T3,0x00);         #######
 852         &pxor           ($T2,$Xi);              #
 853         &pxor           ($T2,$Xhi);             #
 854
 855         &movdqa         ($T3,$T2);              #
 856         &psrldq         ($T2,8);
 857         &pslldq         ($T3,8);                #
 858         &pxor           ($Xhi,$T2);
 859         &pxor           ($Xi,$T3);              #
 860 }
 861 \f
 862 if (1) {                # Algorithm 9 with <<1 twist.
 863                         # Reduction is shorter and uses only two
 864                         # temporary registers, which makes it better
 865                         # candidate for interleaving with 64x64
 866                         # multiplication. Pre-modulo-scheduled loop
 867                         # was found to be ~20% faster than Algorithm 5
 868                         # below. Algorithm 9 was therefore chosen for
 869                         # further optimization...
 870
 871 sub reduction_alg9 {    # 17/13 times faster than Intel version
 872 my ($Xhi,$Xi) = @_;
 873
 874         # 1st phase
 875         &movdqa         ($T1,$Xi)               #
 876         &psllq          ($Xi,1);
 877         &pxor           ($Xi,$T1);              #
 878         &psllq          ($Xi,5);                #
 879         &pxor           ($Xi,$T1);              #
 880         &psllq          ($Xi,57);               #
 881         &movdqa         ($T2,$Xi);              #
 882         &pslldq         ($Xi,8);
 883         &psrldq         ($T2,8);                #
 884         &pxor           ($Xi,$T1);
 885         &pxor           ($Xhi,$T2);             #
 886
 887         # 2nd phase
 888         &movdqa         ($T2,$Xi);
 889         &psrlq          ($Xi,5);
 890         &pxor           ($Xi,$T2);              #
 891         &psrlq          ($Xi,1);                #
 892         &pxor           ($Xi,$T2);              #
 893         &pxor           ($T2,$Xhi);
 894         &psrlq          ($Xi,1);                #
 895         &pxor           ($Xi,$T2);              #
 896 }
 897
 898 &function_begin_B("gcm_init_clmul");
 899         &mov            ($Htbl,&wparam(0));
 900         &mov            ($Xip,&wparam(1));
 901
 902         &call           (&label("pic"));
 903 &set_label("pic");
 904         &blindpop       ($const);
 905         &lea            ($const,&DWP(&label("bswap")."-".&label("pic"),$const));
 906
 907         &movdqu         ($Hkey,&QWP(0,$Xip));
 908         &pshufd         ($Hkey,$Hkey,0b01001110);# dword swap
 909
 910         # <<1 twist
 911         &pshufd         ($T2,$Hkey,0b11111111); # broadcast uppermost dword
 912         &movdqa         ($T1,$Hkey);
 913         &psllq          ($Hkey,1);
 914         &pxor           ($T3,$T3);              #
 915         &psrlq          ($T1,63);
 916         &pcmpgtd        ($T3,$T2);              # broadcast carry bit
 917         &pslldq         ($T1,8);
 918         &por            ($Hkey,$T1);            # H<<=1
 919
 920         # magic reduction
 921         &pand           ($T3,&QWP(16,$const));  # 0x1c2_polynomial
 922         &pxor           ($Hkey,$T3);            # if(carry) H^=0x1c2_polynomial
 923
 924         # calculate H^2
 925         &movdqa         ($Xi,$Hkey);
 926         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);
 927         &reduction_alg9 ($Xhi,$Xi);
 928
 929         &movdqu         (&QWP(0,$Htbl),$Hkey);  # save H
 930         &movdqu         (&QWP(16,$Htbl),$Xi);   # save H^2
 931
 932         &ret            ();
 933 &function_end_B("gcm_init_clmul");
 934
 935 &function_begin_B("gcm_gmult_clmul");
 936         &mov            ($Xip,&wparam(0));
 937         &mov            ($Htbl,&wparam(1));
 938
 939         &call           (&label("pic"));
 940 &set_label("pic");
 941         &blindpop       ($const);
 942         &lea            ($const,&DWP(&label("bswap")."-".&label("pic"),$const));
 943
 944         &movdqu         ($Xi,&QWP(0,$Xip));
 945         &movdqa         ($T3,&QWP(0,$const));
 946         &movdqu         ($Hkey,&QWP(0,$Htbl));
 947         &pshufb         ($Xi,$T3);
 948
 949         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);
 950         &reduction_alg9 ($Xhi,$Xi);
 951
 952         &pshufb         ($Xi,$T3);
 953         &movdqu         (&QWP(0,$Xip),$Xi);
 954
 955         &ret    ();
 956 &function_end_B("gcm_gmult_clmul");
 957
 958 &function_begin("gcm_ghash_clmul");
 959         &mov            ($Xip,&wparam(0));
 960         &mov            ($Htbl,&wparam(1));
 961         &mov            ($inp,&wparam(2));
 962         &mov            ($len,&wparam(3));
 963
 964         &call           (&label("pic"));
 965 &set_label("pic");
 966         &blindpop       ($const);
 967         &lea            ($const,&DWP(&label("bswap")."-".&label("pic"),$const));
 968
 969         &movdqu         ($Xi,&QWP(0,$Xip));
 970         &movdqa         ($T3,&QWP(0,$const));
 971         &movdqu         ($Hkey,&QWP(0,$Htbl));
 972         &pshufb         ($Xi,$T3);
 973
 974         &sub            ($len,0x10);
 975         &jz             (&label("odd_tail"));
 976
 977         #######
 978         # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
 979         #       [(H*Ii+1) + (H*Xi+1)] mod P =
 980         #       [(H*Ii+1) + H^2*(Ii+Xi)] mod P
 981         #
 982         &movdqu         ($T1,&QWP(0,$inp));     # Ii
 983         &movdqu         ($Xn,&QWP(16,$inp));    # Ii+1
 984         &pshufb         ($T1,$T3);
 985         &pshufb         ($Xn,$T3);
 986         &pxor           ($Xi,$T1);              # Ii+Xi
 987
 988         &clmul64x64_T2  ($Xhn,$Xn,$Hkey);       # H*Ii+1
 989         &movdqu         ($Hkey,&QWP(16,$Htbl)); # load H^2
 990
 991         &lea            ($inp,&DWP(32,$inp));   # i+=2
 992         &sub            ($len,0x20);
 993         &jbe            (&label("even_tail"));
 994
 995 &set_label("mod_loop");
 996         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);       # H^2*(Ii+Xi)
 997         &movdqu         ($T1,&QWP(0,$inp));     # Ii
 998         &movdqu         ($Hkey,&QWP(0,$Htbl));  # load H
 999
1000         &pxor           ($Xi,$Xn);              # (H*Ii+1) + H^2*(Ii+Xi)
1001         &pxor           ($Xhi,$Xhn);
1002
1003         &movdqu         ($Xn,&QWP(16,$inp));    # Ii+1
1004         &pshufb         ($T1,$T3);
1005         &pshufb         ($Xn,$T3);
1006
1007         &movdqa         ($T3,$Xn);              #&clmul64x64_TX ($Xhn,$Xn,$Hkey); H*Ii+1
1008         &movdqa         ($Xhn,$Xn);
1009          &pxor          ($Xhi,$T1);             # "Ii+Xi", consume early
1010
1011           &movdqa       ($T1,$Xi)               #&reduction_alg9($Xhi,$Xi); 1st phase
1012           &psllq        ($Xi,1);
1013           &pxor         ($Xi,$T1);              #
1014           &psllq        ($Xi,5);                #
1015           &pxor         ($Xi,$T1);              #
1016         &pclmulqdq      ($Xn,$Hkey,0x00);       #######
1017           &psllq        ($Xi,57);               #
1018           &movdqa       ($T2,$Xi);              #
1019           &pslldq       ($Xi,8);
1020           &psrldq       ($T2,8);                #
1021           &pxor         ($Xi,$T1);
1022         &pshufd         ($T1,$T3,0b01001110);
1023           &pxor         ($Xhi,$T2);             #
1024         &pxor           ($T1,$T3);
1025         &pshufd         ($T3,$Hkey,0b01001110);
1026         &pxor           ($T3,$Hkey);            #
1027
1028         &pclmulqdq      ($Xhn,$Hkey,0x11);      #######
1029           &movdqa       ($T2,$Xi);              # 2nd phase
1030           &psrlq        ($Xi,5);
1031           &pxor         ($Xi,$T2);              #
1032           &psrlq        ($Xi,1);                #
1033           &pxor         ($Xi,$T2);              #
1034           &pxor         ($T2,$Xhi);
1035           &psrlq        ($Xi,1);                #
1036           &pxor         ($Xi,$T2);              #
1037
1038         &pclmulqdq      ($T1,$T3,0x00);         #######
1039         &movdqu         ($Hkey,&QWP(16,$Htbl)); # load H^2
1040         &pxor           ($T1,$Xn);              #
1041         &pxor           ($T1,$Xhn);             #
1042
1043         &movdqa         ($T3,$T1);              #
1044         &psrldq         ($T1,8);
1045         &pslldq         ($T3,8);                #
1046         &pxor           ($Xhn,$T1);
1047         &pxor           ($Xn,$T3);              #
1048         &movdqa         ($T3,&QWP(0,$const));
1049
1050         &lea            ($inp,&DWP(32,$inp));
1051         &sub            ($len,0x20);
1052         &ja             (&label("mod_loop"));
1053
1054 &set_label("even_tail");
1055         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);       # H^2*(Ii+Xi)
1056
1057         &pxor           ($Xi,$Xn);              # (H*Ii+1) + H^2*(Ii+Xi)
1058         &pxor           ($Xhi,$Xhn);
1059
1060         &reduction_alg9 ($Xhi,$Xi);
1061
1062         &test           ($len,$len);
1063         &jnz            (&label("done"));
1064
1065         &movdqu         ($Hkey,&QWP(0,$Htbl));  # load H
1066 &set_label("odd_tail");
1067         &movdqu         ($T1,&QWP(0,$inp));     # Ii
1068         &pshufb         ($T1,$T3);
1069         &pxor           ($Xi,$T1);              # Ii+Xi
1070
1071         &clmul64x64_T2  ($Xhi,$Xi,$Hkey);       # H*(Ii+Xi)
1072         &reduction_alg9 ($Xhi,$Xi);
1073
1074 &set_label("done");
1075         &pshufb         ($Xi,$T3);
1076         &movdqu         (&QWP(0,$Xip),$Xi);
1077 &function_end("gcm_ghash_clmul");
1078 \f
1079 } else {                # Algorith 5. Kept for reference purposes.
1080
1081 sub reduction_alg5 {    # 19/16 times faster than Intel version
1082 my ($Xhi,$Xi)=@_;
1083
1084         # <<1
1085         &movdqa         ($T1,$Xi);              #
1086         &movdqa         ($T2,$Xhi);
1087         &pslld          ($Xi,1);
1088         &pslld          ($Xhi,1);               #
1089         &psrld          ($T1,31);
1090         &psrld          ($T2,31);               #
1091         &movdqa         ($T3,$T1);
1092         &pslldq         ($T1,4);
1093         &psrldq         ($T3,12);               #
1094         &pslldq         ($T2,4);
1095         &por            ($Xhi,$T3);             #
1096         &por            ($Xi,$T1);
1097         &por            ($Xhi,$T2);             #
1098
1099         # 1st phase
1100         &movdqa         ($T1,$Xi);
1101         &movdqa         ($T2,$Xi);
1102         &movdqa         ($T3,$Xi);              #
1103         &pslld          ($T1,31);
1104         &pslld          ($T2,30);
1105         &pslld          ($Xi,25);               #
1106         &pxor           ($T1,$T2);
1107         &pxor           ($T1,$Xi);              #
1108         &movdqa         ($T2,$T1);              #
1109         &pslldq         ($T1,12);
1110         &psrldq         ($T2,4);                #
1111         &pxor           ($T3,$T1);
1112
1113         # 2nd phase
1114         &pxor           ($Xhi,$T3);             #
1115         &movdqa         ($Xi,$T3);
1116         &movdqa         ($T1,$T3);
1117         &psrld          ($Xi,1);                #
1118         &psrld          ($T1,2);
1119         &psrld          ($T3,7);                #
1120         &pxor           ($Xi,$T1);
1121         &pxor           ($Xhi,$T2);
1122         &pxor           ($Xi,$T3);              #
1123         &pxor           ($Xi,$Xhi);             #
1124 }
1125
1126 &function_begin_B("gcm_init_clmul");
1127         &mov            ($Htbl,&wparam(0));
1128         &mov            ($Xip,&wparam(1));
1129
1130         &call           (&label("pic"));
1131 &set_label("pic");
1132         &blindpop       ($const);
1133         &lea            ($const,&DWP(&label("bswap")."-".&label("pic"),$const));
1134
1135         &movdqu         ($Hkey,&QWP(0,$Xip));
1136         &pshufd         ($Hkey,$Hkey,0b01001110);# dword swap
1137
1138         # calculate H^2
1139         &movdqa         ($Xi,$Hkey);
1140         &clmul64x64_T3  ($Xhi,$Xi,$Hkey);
1141         &reduction_alg5 ($Xhi,$Xi);
1142
1143         &movdqu         (&QWP(0,$Htbl),$Hkey);  # save H
1144         &movdqu         (&QWP(16,$Htbl),$Xi);   # save H^2
1145
1146         &ret            ();
1147 &function_end_B("gcm_init_clmul");
1148
1149 &function_begin_B("gcm_gmult_clmul");
1150         &mov            ($Xip,&wparam(0));
1151         &mov            ($Htbl,&wparam(1));
1152
1153         &call           (&label("pic"));
1154 &set_label("pic");
1155         &blindpop       ($const);
1156         &lea            ($const,&DWP(&label("bswap")."-".&label("pic"),$const));
1157
1158         &movdqu         ($Xi,&QWP(0,$Xip));
1159         &movdqa         ($Xn,&QWP(0,$const));
1160         &movdqu         ($Hkey,&QWP(0,$Htbl));
1161         &pshufb         ($Xi,$Xn);
1162
1163         &clmul64x64_T3  ($Xhi,$Xi,$Hkey);
1164         &reduction_alg5 ($Xhi,$Xi);
1165
1166         &pshufb         ($Xi,$Xn);
1167         &movdqu         (&QWP(0,$Xip),$Xi);
1168
1169         &ret    ();
1170 &function_end_B("gcm_gmult_clmul");
1171
1172 &function_begin("gcm_ghash_clmul");
1173         &mov            ($Xip,&wparam(0));
1174         &mov            ($Htbl,&wparam(1));
1175         &mov            ($inp,&wparam(2));
1176         &mov            ($len,&wparam(3));
1177
1178         &call           (&label("pic"));
1179 &set_label("pic");
1180         &blindpop       ($const);
1181         &lea            ($const,&DWP(&label("bswap")."-".&label("pic"),$const));
1182
1183         &movdqu         ($Xi,&QWP(0,$Xip));
1184         &movdqa         ($T3,&QWP(0,$const));
1185         &movdqu         ($Hkey,&QWP(0,$Htbl));
1186         &pshufb         ($Xi,$T3);
1187
1188         &sub            ($len,0x10);
1189         &jz             (&label("odd_tail"));
1190
1191         #######
1192         # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
1193         #       [(H*Ii+1) + (H*Xi+1)] mod P =
1194         #       [(H*Ii+1) + H^2*(Ii+Xi)] mod P
1195         #
1196         &movdqu         ($T1,&QWP(0,$inp));     # Ii
1197         &movdqu         ($Xn,&QWP(16,$inp));    # Ii+1
1198         &pshufb         ($T1,$T3);
1199         &pshufb         ($Xn,$T3);
1200         &pxor           ($Xi,$T1);              # Ii+Xi
1201
1202         &clmul64x64_T3  ($Xhn,$Xn,$Hkey);       # H*Ii+1
1203         &movdqu         ($Hkey,&QWP(16,$Htbl)); # load H^2
1204
1205         &sub            ($len,0x20);
1206         &lea            ($inp,&DWP(32,$inp));   # i+=2
1207         &jbe            (&label("even_tail"));
1208
1209 &set_label("mod_loop");
1210         &clmul64x64_T3  ($Xhi,$Xi,$Hkey);       # H^2*(Ii+Xi)
1211         &movdqu         ($Hkey,&QWP(0,$Htbl));  # load H
1212
1213         &pxor           ($Xi,$Xn);              # (H*Ii+1) + H^2*(Ii+Xi)
1214         &pxor           ($Xhi,$Xhn);
1215
1216         &reduction_alg5 ($Xhi,$Xi);
1217
1218         #######
1219         &movdqa         ($T3,&QWP(0,$const));
1220         &movdqu         ($T1,&QWP(0,$inp));     # Ii
1221         &movdqu         ($Xn,&QWP(16,$inp));    # Ii+1
1222         &pshufb         ($T1,$T3);
1223         &pshufb         ($Xn,$T3);
1224         &pxor           ($Xi,$T1);              # Ii+Xi
1225
1226         &clmul64x64_T3  ($Xhn,$Xn,$Hkey);       # H*Ii+1
1227         &movdqu         ($Hkey,&QWP(16,$Htbl)); # load H^2
1228
1229         &sub            ($len,0x20);
1230         &lea            ($inp,&DWP(32,$inp));
1231         &ja             (&label("mod_loop"));
1232
1233 &set_label("even_tail");
1234         &clmul64x64_T3  ($Xhi,$Xi,$Hkey);       # H^2*(Ii+Xi)
1235
1236         &pxor           ($Xi,$Xn);              # (H*Ii+1) + H^2*(Ii+Xi)
1237         &pxor           ($Xhi,$Xhn);
1238
1239         &reduction_alg5 ($Xhi,$Xi);
1240
1241         &movdqa         ($T3,&QWP(0,$const));
1242         &test           ($len,$len);
1243         &jnz            (&label("done"));
1244
1245         &movdqu         ($Hkey,&QWP(0,$Htbl));  # load H
1246 &set_label("odd_tail");
1247         &movdqu         ($T1,&QWP(0,$inp));     # Ii
1248         &pshufb         ($T1,$T3);
1249         &pxor           ($Xi,$T1);              # Ii+Xi
1250
1251         &clmul64x64_T3  ($Xhi,$Xi,$Hkey);       # H*(Ii+Xi)
1252         &reduction_alg5 ($Xhi,$Xi);
1253
1254         &movdqa         ($T3,&QWP(0,$const));
1255 &set_label("done");
1256         &pshufb         ($Xi,$T3);
1257         &movdqu         (&QWP(0,$Xip),$Xi);
1258 &function_end("gcm_ghash_clmul");
1259
1260 }
1261 \f
1262 &set_label("bswap",64);
1263         &data_byte(15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0);
1264         &data_byte(1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2); # 0x1c2_polynomial
1265 }}      # $sse2
1266
1267 &set_label("rem_4bit",64);
1268         &data_word(0,0x0000<<$S,0,0x1C20<<$S,0,0x3840<<$S,0,0x2460<<$S);
1269         &data_word(0,0x7080<<$S,0,0x6CA0<<$S,0,0x48C0<<$S,0,0x54E0<<$S);
1270         &data_word(0,0xE100<<$S,0,0xFD20<<$S,0,0xD940<<$S,0,0xC560<<$S);
1271         &data_word(0,0x9180<<$S,0,0x8DA0<<$S,0,0xA9C0<<$S,0,0xB5E0<<$S);
1272 &set_label("rem_8bit",64);
1273         &data_short(0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E);
1274         &data_short(0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E);
1275         &data_short(0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E);
1276         &data_short(0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E);
1277         &data_short(0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E);
1278         &data_short(0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E);
1279         &data_short(0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E);
1280         &data_short(0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E);
1281         &data_short(0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE);
1282         &data_short(0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE);
1283         &data_short(0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE);
1284         &data_short(0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE);
1285         &data_short(0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E);
1286         &data_short(0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E);
1287         &data_short(0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE);
1288         &data_short(0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE);
1289         &data_short(0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E);
1290         &data_short(0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E);
1291         &data_short(0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E);
1292         &data_short(0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E);
1293         &data_short(0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E);
1294         &data_short(0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E);
1295         &data_short(0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E);
1296         &data_short(0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E);
1297         &data_short(0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE);
1298         &data_short(0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE);
1299         &data_short(0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE);
1300         &data_short(0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE);
1301         &data_short(0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E);
1302         &data_short(0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E);
1303         &data_short(0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE);
1304         &data_short(0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE);
1305 }}}     # !$x86only
1306
1307 &asciz("GHASH for x86, CRYPTOGAMS by <appro\@openssl.org>");
1308 &asm_finish();
1309
1310 # A question was risen about choice of vanilla MMX. Or rather why wasn't
1311 # SSE2 chosen instead? In addition to the fact that MMX runs on legacy
1312 # CPUs such as PIII, "4-bit" MMX version was observed to provide better
1313 # performance than *corresponding* SSE2 one even on contemporary CPUs.
1314 # SSE2 results were provided by Peter-Michael Hager. He maintains SSE2
1315 # implementation featuring full range of lookup-table sizes, but with
1316 # per-invocation lookup table setup. Latter means that table size is
1317 # chosen depending on how much data is to be hashed in every given call,
1318 # more data - larger table. Best reported result for Core2 is ~4 cycles
1319 # per processed byte out of 64KB block. Recall that this number accounts
1320 # even for 64KB table setup overhead. As discussed in gcm128.c we choose
1321 # to be more conservative in respect to lookup table sizes, but how
1322 # do the results compare? Minimalistic "256B" MMX version delivers ~11
1323 # cycles on same platform. As also discussed in gcm128.c, next in line
1324 # "8-bit Shoup's" method should deliver twice the performance of "4-bit"
1325 # one. It should be also be noted that in SSE2 case improvement can be
1326 # "super-linear," i.e. more than twice, mostly because >>8 maps to
1327 # single instruction on SSE2 register. This is unlike "4-bit" case when
1328 # >>4 maps to same amount of instructions in both MMX and SSE2 cases.
1329 # Bottom line is that switch to SSE2 is considered to be justifiable
1330 # only in case we choose to implement "8-bit" method...