crypto/poly1305/asm/poly1305-x86.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2016 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the OpenSSL license (the "License").  You may not use
   5 # this file except in compliance with the License.  You can obtain a copy
   6 # in the file LICENSE in the source distribution or at
   7 # https://www.openssl.org/source/license.html
   8
   9 #
  10 # ====================================================================
  11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
  12 # project. The module is, however, dual licensed under OpenSSL and
  13 # CRYPTOGAMS licenses depending on where you obtain it. For further
  14 # details see http://www.openssl.org/~appro/cryptogams/.
  15 # ====================================================================
  16 #
  17 # This module implements Poly1305 hash for x86.
  18 #
  19 # April 2015
  20 #
  21 # Numbers are cycles per processed byte with poly1305_blocks alone,
  22 # measured with rdtsc at fixed clock frequency.
  23 #
  24 #               IALU/gcc-3.4(*) SSE2(**)        AVX2
  25 # Pentium       15.7/+80%       -
  26 # PIII          6.21/+90%       -
  27 # P4            19.8/+40%       3.24
  28 # Core 2        4.85/+90%       1.80
  29 # Westmere      4.58/+100%      1.43
  30 # Sandy Bridge  3.90/+100%      1.36
  31 # Haswell       3.88/+70%       1.18            0.72
  32 # Skylake       3.10/+60%       1.14            0.62
  33 # Silvermont    11.0/+40%       4.80
  34 # Goldmont      4.10/+200%      2.10
  35 # VIA Nano      6.71/+90%       2.47
  36 # Sledgehammer  3.51/+180%      4.27
  37 # Bulldozer     4.53/+140%      1.31
  38 #
  39 # (*)   gcc 4.8 for some reason generated worse code;
  40 # (**)  besides SSE2 there are floating-point and AVX options; FP
  41 #       is deemed unnecessary, because pre-SSE2 processor are too
  42 #       old to care about, while it's not the fastest option on
  43 #       SSE2-capable ones; AVX is omitted, because it doesn't give
  44 #       a lot of improvement, 5-10% depending on processor;
  45
  46 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  47 push(@INC,"${dir}","${dir}../../perlasm");
  48 require "x86asm.pl";
  49
  50 $output=pop;
  51 open STDOUT,">$output";
  52
  53 &asm_init($ARGV[0],$ARGV[$#ARGV] eq "386");
  54
  55 $sse2=$avx=0;
  56 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
  57
  58 if ($sse2) {
  59         &static_label("const_sse2");
  60         &static_label("enter_blocks");
  61         &static_label("enter_emit");
  62         &external_label("OPENSSL_ia32cap_P");
  63
  64         if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
  65                         =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
  66                 $avx = ($1>=2.19) + ($1>=2.22);
  67         }
  68
  69         if (!$avx && $ARGV[0] eq "win32n" &&
  70            `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
  71         $avx = ($1>=2.09) + ($1>=2.10);
  72         }
  73
  74         if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|based on LLVM) ([3-9]\.[0-9]+)/) {
  75                 $avx = ($2>=3.0) + ($2>3.0);
  76         }
  77 }
  78
  79 ########################################################################
  80 # Layout of opaque area is following.
  81 #
  82 #       unsigned __int32 h[5];          # current hash value base 2^32
  83 #       unsigned __int32 pad;           # is_base2_26 in vector context
  84 #       unsigned __int32 r[4];          # key value base 2^32
  85
  86 &align(64);
  87 &function_begin("poly1305_init");
  88         &mov    ("edi",&wparam(0));             # context
  89         &mov    ("esi",&wparam(1));             # key
  90         &mov    ("ebp",&wparam(2));             # function table
  91
  92         &xor    ("eax","eax");
  93         &mov    (&DWP(4*0,"edi"),"eax");        # zero hash value
  94         &mov    (&DWP(4*1,"edi"),"eax");
  95         &mov    (&DWP(4*2,"edi"),"eax");
  96         &mov    (&DWP(4*3,"edi"),"eax");
  97         &mov    (&DWP(4*4,"edi"),"eax");
  98         &mov    (&DWP(4*5,"edi"),"eax");        # is_base2_26
  99
 100         &cmp    ("esi",0);
 101         &je     (&label("nokey"));
 102
 103     if ($sse2) {
 104         &call   (&label("pic_point"));
 105     &set_label("pic_point");
 106         &blindpop("ebx");
 107
 108         &lea    ("eax",&DWP("poly1305_blocks-".&label("pic_point"),"ebx"));
 109         &lea    ("edx",&DWP("poly1305_emit-".&label("pic_point"),"ebx"));
 110
 111         &picmeup("edi","OPENSSL_ia32cap_P","ebx",&label("pic_point"));
 112         &mov    ("ecx",&DWP(0,"edi"));
 113         &and    ("ecx",1<<26|1<<24);
 114         &cmp    ("ecx",1<<26|1<<24);            # SSE2 and XMM?
 115         &jne    (&label("no_sse2"));
 116
 117         &lea    ("eax",&DWP("_poly1305_blocks_sse2-".&label("pic_point"),"ebx"));
 118         &lea    ("edx",&DWP("_poly1305_emit_sse2-".&label("pic_point"),"ebx"));
 119
 120       if ($avx>1) {
 121         &mov    ("ecx",&DWP(8,"edi"));
 122         &test   ("ecx",1<<5);                   # AVX2?
 123         &jz     (&label("no_sse2"));
 124
 125         &lea    ("eax",&DWP("_poly1305_blocks_avx2-".&label("pic_point"),"ebx"));
 126       }
 127     &set_label("no_sse2");
 128         &mov    ("edi",&wparam(0));             # reload context
 129         &mov    (&DWP(0,"ebp"),"eax");          # fill function table
 130         &mov    (&DWP(4,"ebp"),"edx");
 131     }
 132
 133         &mov    ("eax",&DWP(4*0,"esi"));        # load input key
 134         &mov    ("ebx",&DWP(4*1,"esi"));
 135         &mov    ("ecx",&DWP(4*2,"esi"));
 136         &mov    ("edx",&DWP(4*3,"esi"));
 137         &and    ("eax",0x0fffffff);
 138         &and    ("ebx",0x0ffffffc);
 139         &and    ("ecx",0x0ffffffc);
 140         &and    ("edx",0x0ffffffc);
 141         &mov    (&DWP(4*6,"edi"),"eax");
 142         &mov    (&DWP(4*7,"edi"),"ebx");
 143         &mov    (&DWP(4*8,"edi"),"ecx");
 144         &mov    (&DWP(4*9,"edi"),"edx");
 145
 146         &mov    ("eax",$sse2);
 147 &set_label("nokey");
 148 &function_end("poly1305_init");
 149
 150 ($h0,$h1,$h2,$h3,$h4,
 151  $d0,$d1,$d2,$d3,
 152  $r0,$r1,$r2,$r3,
 153      $s1,$s2,$s3)=map(4*$_,(0..15));
 154
 155 &function_begin("poly1305_blocks");
 156         &mov    ("edi",&wparam(0));             # ctx
 157         &mov    ("esi",&wparam(1));             # inp
 158         &mov    ("ecx",&wparam(2));             # len
 159 &set_label("enter_blocks");
 160         &and    ("ecx",-15);
 161         &jz     (&label("nodata"));
 162
 163         &stack_push(16);
 164         &mov    ("eax",&DWP(4*6,"edi"));        # r0
 165         &mov    ("ebx",&DWP(4*7,"edi"));        # r1
 166          &lea   ("ebp",&DWP(0,"esi","ecx"));    # end of input
 167         &mov    ("ecx",&DWP(4*8,"edi"));        # r2
 168         &mov    ("edx",&DWP(4*9,"edi"));        # r3
 169
 170         &mov    (&wparam(2),"ebp");
 171         &mov    ("ebp","esi");
 172
 173         &mov    (&DWP($r0,"esp"),"eax");        # r0
 174         &mov    ("eax","ebx");
 175         &shr    ("eax",2);
 176         &mov    (&DWP($r1,"esp"),"ebx");        # r1
 177         &add    ("eax","ebx");                  # s1
 178         &mov    ("ebx","ecx");
 179         &shr    ("ebx",2);
 180         &mov    (&DWP($r2,"esp"),"ecx");        # r2
 181         &add    ("ebx","ecx");                  # s2
 182         &mov    ("ecx","edx");
 183         &shr    ("ecx",2);
 184         &mov    (&DWP($r3,"esp"),"edx");        # r3
 185         &add    ("ecx","edx");                  # s3
 186         &mov    (&DWP($s1,"esp"),"eax");        # s1
 187         &mov    (&DWP($s2,"esp"),"ebx");        # s2
 188         &mov    (&DWP($s3,"esp"),"ecx");        # s3
 189
 190         &mov    ("eax",&DWP(4*0,"edi"));        # load hash value
 191         &mov    ("ebx",&DWP(4*1,"edi"));
 192         &mov    ("ecx",&DWP(4*2,"edi"));
 193         &mov    ("esi",&DWP(4*3,"edi"));
 194         &mov    ("edi",&DWP(4*4,"edi"));
 195         &jmp    (&label("loop"));
 196
 197 &set_label("loop",32);
 198         &add    ("eax",&DWP(4*0,"ebp"));        # accumulate input
 199         &adc    ("ebx",&DWP(4*1,"ebp"));
 200         &adc    ("ecx",&DWP(4*2,"ebp"));
 201         &adc    ("esi",&DWP(4*3,"ebp"));
 202         &lea    ("ebp",&DWP(4*4,"ebp"));
 203         &adc    ("edi",&wparam(3));             # padbit
 204
 205         &mov    (&DWP($h0,"esp"),"eax");        # put aside hash[+inp]
 206         &mov    (&DWP($h3,"esp"),"esi");
 207
 208         &mul    (&DWP($r0,"esp"));              # h0*r0
 209          &mov   (&DWP($h4,"esp"),"edi");
 210         &mov    ("edi","eax");
 211         &mov    ("eax","ebx");                  # h1
 212         &mov    ("esi","edx");
 213         &mul    (&DWP($s3,"esp"));              # h1*s3
 214         &add    ("edi","eax");
 215         &mov    ("eax","ecx");                  # h2
 216         &adc    ("esi","edx");
 217         &mul    (&DWP($s2,"esp"));              # h2*s2
 218         &add    ("edi","eax");
 219         &mov    ("eax",&DWP($h3,"esp"));
 220         &adc    ("esi","edx");
 221         &mul    (&DWP($s1,"esp"));              # h3*s1
 222         &add    ("edi","eax");
 223          &mov   ("eax",&DWP($h0,"esp"));
 224         &adc    ("esi","edx");
 225
 226         &mul    (&DWP($r1,"esp"));              # h0*r1
 227          &mov   (&DWP($d0,"esp"),"edi");
 228         &xor    ("edi","edi");
 229         &add    ("esi","eax");
 230         &mov    ("eax","ebx");                  # h1
 231         &adc    ("edi","edx");
 232         &mul    (&DWP($r0,"esp"));              # h1*r0
 233         &add    ("esi","eax");
 234         &mov    ("eax","ecx");                  # h2
 235         &adc    ("edi","edx");
 236         &mul    (&DWP($s3,"esp"));              # h2*s3
 237         &add    ("esi","eax");
 238         &mov    ("eax",&DWP($h3,"esp"));
 239         &adc    ("edi","edx");
 240         &mul    (&DWP($s2,"esp"));              # h3*s2
 241         &add    ("esi","eax");
 242         &mov    ("eax",&DWP($h4,"esp"));
 243         &adc    ("edi","edx");
 244         &imul   ("eax",&DWP($s1,"esp"));        # h4*s1
 245         &add    ("esi","eax");
 246          &mov   ("eax",&DWP($h0,"esp"));
 247         &adc    ("edi",0);
 248
 249         &mul    (&DWP($r2,"esp"));              # h0*r2
 250          &mov   (&DWP($d1,"esp"),"esi");
 251         &xor    ("esi","esi");
 252         &add    ("edi","eax");
 253         &mov    ("eax","ebx");                  # h1
 254         &adc    ("esi","edx");
 255         &mul    (&DWP($r1,"esp"));              # h1*r1
 256         &add    ("edi","eax");
 257         &mov    ("eax","ecx");                  # h2
 258         &adc    ("esi","edx");
 259         &mul    (&DWP($r0,"esp"));              # h2*r0
 260         &add    ("edi","eax");
 261         &mov    ("eax",&DWP($h3,"esp"));
 262         &adc    ("esi","edx");
 263         &mul    (&DWP($s3,"esp"));              # h3*s3
 264         &add    ("edi","eax");
 265         &mov    ("eax",&DWP($h4,"esp"));
 266         &adc    ("esi","edx");
 267         &imul   ("eax",&DWP($s2,"esp"));        # h4*s2
 268         &add    ("edi","eax");
 269          &mov   ("eax",&DWP($h0,"esp"));
 270         &adc    ("esi",0);
 271
 272         &mul    (&DWP($r3,"esp"));              # h0*r3
 273          &mov   (&DWP($d2,"esp"),"edi");
 274         &xor    ("edi","edi");
 275         &add    ("esi","eax");
 276         &mov    ("eax","ebx");                  # h1
 277         &adc    ("edi","edx");
 278         &mul    (&DWP($r2,"esp"));              # h1*r2
 279         &add    ("esi","eax");
 280         &mov    ("eax","ecx");                  # h2
 281         &adc    ("edi","edx");
 282         &mul    (&DWP($r1,"esp"));              # h2*r1
 283         &add    ("esi","eax");
 284         &mov    ("eax",&DWP($h3,"esp"));
 285         &adc    ("edi","edx");
 286         &mul    (&DWP($r0,"esp"));              # h3*r0
 287         &add    ("esi","eax");
 288          &mov   ("ecx",&DWP($h4,"esp"));
 289         &adc    ("edi","edx");
 290
 291         &mov    ("edx","ecx");
 292         &imul   ("ecx",&DWP($s3,"esp"));        # h4*s3
 293         &add    ("esi","ecx");
 294          &mov   ("eax",&DWP($d0,"esp"));
 295         &adc    ("edi",0);
 296
 297         &imul   ("edx",&DWP($r0,"esp"));        # h4*r0
 298         &add    ("edx","edi");
 299
 300         &mov    ("ebx",&DWP($d1,"esp"));
 301         &mov    ("ecx",&DWP($d2,"esp"));
 302
 303         &mov    ("edi","edx");                  # last reduction step
 304         &shr    ("edx",2);
 305         &and    ("edi",3);
 306         &lea    ("edx",&DWP(0,"edx","edx",4));  # *5
 307         &add    ("eax","edx");
 308         &adc    ("ebx",0);
 309         &adc    ("ecx",0);
 310         &adc    ("esi",0);
 311         &adc    ("edi",0);
 312
 313         &cmp    ("ebp",&wparam(2));             # done yet?
 314         &jne    (&label("loop"));
 315
 316         &mov    ("edx",&wparam(0));             # ctx
 317         &stack_pop(16);
 318         &mov    (&DWP(4*0,"edx"),"eax");        # store hash value
 319         &mov    (&DWP(4*1,"edx"),"ebx");
 320         &mov    (&DWP(4*2,"edx"),"ecx");
 321         &mov    (&DWP(4*3,"edx"),"esi");
 322         &mov    (&DWP(4*4,"edx"),"edi");
 323 &set_label("nodata");
 324 &function_end("poly1305_blocks");
 325
 326 &function_begin("poly1305_emit");
 327         &mov    ("ebp",&wparam(0));             # context
 328 &set_label("enter_emit");
 329         &mov    ("edi",&wparam(1));             # output
 330         &mov    ("eax",&DWP(4*0,"ebp"));        # load hash value
 331         &mov    ("ebx",&DWP(4*1,"ebp"));
 332         &mov    ("ecx",&DWP(4*2,"ebp"));
 333         &mov    ("edx",&DWP(4*3,"ebp"));
 334         &mov    ("esi",&DWP(4*4,"ebp"));
 335
 336         &add    ("eax",5);                      # compare to modulus
 337         &adc    ("ebx",0);
 338         &adc    ("ecx",0);
 339         &adc    ("edx",0);
 340         &adc    ("esi",0);
 341         &shr    ("esi",2);                      # did it carry/borrow?
 342         &neg    ("esi");                        # do we choose hash-modulus?
 343
 344         &and    ("eax","esi");
 345         &and    ("ebx","esi");
 346         &and    ("ecx","esi");
 347         &and    ("edx","esi");
 348         &mov    (&DWP(4*0,"edi"),"eax");
 349         &mov    (&DWP(4*1,"edi"),"ebx");
 350         &mov    (&DWP(4*2,"edi"),"ecx");
 351         &mov    (&DWP(4*3,"edi"),"edx");
 352
 353         &not    ("esi");                        # or original hash value?
 354         &mov    ("eax",&DWP(4*0,"ebp"));
 355         &mov    ("ebx",&DWP(4*1,"ebp"));
 356         &mov    ("ecx",&DWP(4*2,"ebp"));
 357         &mov    ("edx",&DWP(4*3,"ebp"));
 358         &mov    ("ebp",&wparam(2));
 359         &and    ("eax","esi");
 360         &and    ("ebx","esi");
 361         &and    ("ecx","esi");
 362         &and    ("edx","esi");
 363         &or     ("eax",&DWP(4*0,"edi"));
 364         &or     ("ebx",&DWP(4*1,"edi"));
 365         &or     ("ecx",&DWP(4*2,"edi"));
 366         &or     ("edx",&DWP(4*3,"edi"));
 367
 368         &add    ("eax",&DWP(4*0,"ebp"));        # accumulate key
 369         &adc    ("ebx",&DWP(4*1,"ebp"));
 370         &adc    ("ecx",&DWP(4*2,"ebp"));
 371         &adc    ("edx",&DWP(4*3,"ebp"));
 372
 373         &mov    (&DWP(4*0,"edi"),"eax");
 374         &mov    (&DWP(4*1,"edi"),"ebx");
 375         &mov    (&DWP(4*2,"edi"),"ecx");
 376         &mov    (&DWP(4*3,"edi"),"edx");
 377 &function_end("poly1305_emit");
 378
 379 if ($sse2) {
 380 ########################################################################
 381 # Layout of opaque area is following.
 382 #
 383 #       unsigned __int32 h[5];          # current hash value base 2^26
 384 #       unsigned __int32 is_base2_26;
 385 #       unsigned __int32 r[4];          # key value base 2^32
 386 #       unsigned __int32 pad[2];
 387 #       struct { unsigned __int32 r^4, r^3, r^2, r^1; } r[9];
 388 #
 389 # where r^n are base 2^26 digits of degrees of multiplier key. There are
 390 # 5 digits, but last four are interleaved with multiples of 5, totalling
 391 # in 9 elements: r0, r1, 5*r1, r2, 5*r2, r3, 5*r3, r4, 5*r4.
 392
 393 my ($D0,$D1,$D2,$D3,$D4,$T0,$T1,$T2)=map("xmm$_",(0..7));
 394 my $MASK=$T2;   # borrow and keep in mind
 395
 396 &align  (32);
 397 &function_begin_B("_poly1305_init_sse2");
 398         &movdqu         ($D4,&QWP(4*6,"edi"));          # key base 2^32
 399         &lea            ("edi",&DWP(16*3,"edi"));       # size optimization
 400         &mov            ("ebp","esp");
 401         &sub            ("esp",16*(9+5));
 402         &and            ("esp",-16);
 403
 404         #&pand          ($D4,&QWP(96,"ebx"));           # magic mask
 405         &movq           ($MASK,&QWP(64,"ebx"));
 406
 407         &movdqa         ($D0,$D4);
 408         &movdqa         ($D1,$D4);
 409         &movdqa         ($D2,$D4);
 410
 411         &pand           ($D0,$MASK);                    # -> base 2^26
 412         &psrlq          ($D1,26);
 413         &psrldq         ($D2,6);
 414         &pand           ($D1,$MASK);
 415         &movdqa         ($D3,$D2);
 416         &psrlq          ($D2,4)
 417         &psrlq          ($D3,30);
 418         &pand           ($D2,$MASK);
 419         &pand           ($D3,$MASK);
 420         &psrldq         ($D4,13);
 421
 422         &lea            ("edx",&DWP(16*9,"esp"));       # size optimization
 423         &mov            ("ecx",2);
 424 &set_label("square");
 425         &movdqa         (&QWP(16*0,"esp"),$D0);
 426         &movdqa         (&QWP(16*1,"esp"),$D1);
 427         &movdqa         (&QWP(16*2,"esp"),$D2);
 428         &movdqa         (&QWP(16*3,"esp"),$D3);
 429         &movdqa         (&QWP(16*4,"esp"),$D4);
 430
 431         &movdqa         ($T1,$D1);
 432         &movdqa         ($T0,$D2);
 433         &pslld          ($T1,2);
 434         &pslld          ($T0,2);
 435         &paddd          ($T1,$D1);                      # *5
 436         &paddd          ($T0,$D2);                      # *5
 437         &movdqa         (&QWP(16*5,"esp"),$T1);
 438         &movdqa         (&QWP(16*6,"esp"),$T0);
 439         &movdqa         ($T1,$D3);
 440         &movdqa         ($T0,$D4);
 441         &pslld          ($T1,2);
 442         &pslld          ($T0,2);
 443         &paddd          ($T1,$D3);                      # *5
 444         &paddd          ($T0,$D4);                      # *5
 445         &movdqa         (&QWP(16*7,"esp"),$T1);
 446         &movdqa         (&QWP(16*8,"esp"),$T0);
 447
 448         &pshufd         ($T1,$D0,0b01000100);
 449         &movdqa         ($T0,$D1);
 450         &pshufd         ($D1,$D1,0b01000100);
 451         &pshufd         ($D2,$D2,0b01000100);
 452         &pshufd         ($D3,$D3,0b01000100);
 453         &pshufd         ($D4,$D4,0b01000100);
 454         &movdqa         (&QWP(16*0,"edx"),$T1);
 455         &movdqa         (&QWP(16*1,"edx"),$D1);
 456         &movdqa         (&QWP(16*2,"edx"),$D2);
 457         &movdqa         (&QWP(16*3,"edx"),$D3);
 458         &movdqa         (&QWP(16*4,"edx"),$D4);
 459
 460         ################################################################
 461         # d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
 462         # d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
 463         # d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
 464         # d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
 465         # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
 466
 467         &pmuludq        ($D4,$D0);                      # h4*r0
 468         &pmuludq        ($D3,$D0);                      # h3*r0
 469         &pmuludq        ($D2,$D0);                      # h2*r0
 470         &pmuludq        ($D1,$D0);                      # h1*r0
 471         &pmuludq        ($D0,$T1);                      # h0*r0
 472
 473 sub pmuladd {
 474 my $load = shift;
 475 my $base = shift; $base = "esp" if (!defined($base));
 476
 477         ################################################################
 478         # As for choice to "rotate" $T0-$T2 in order to move paddq
 479         # past next multiplication. While it makes code harder to read
 480         # and doesn't have significant effect on most processors, it
 481         # makes a lot of difference on Atom, up to 30% improvement.
 482
 483         &movdqa         ($T1,$T0);
 484         &pmuludq        ($T0,&QWP(16*3,$base));         # r1*h3
 485         &movdqa         ($T2,$T1);
 486         &pmuludq        ($T1,&QWP(16*2,$base));         # r1*h2
 487         &paddq          ($D4,$T0);
 488         &movdqa         ($T0,$T2);
 489         &pmuludq        ($T2,&QWP(16*1,$base));         # r1*h1
 490         &paddq          ($D3,$T1);
 491         &$load          ($T1,5);                        # s1
 492         &pmuludq        ($T0,&QWP(16*0,$base));         # r1*h0
 493         &paddq          ($D2,$T2);
 494         &pmuludq        ($T1,&QWP(16*4,$base));         # s1*h4
 495          &$load         ($T2,2);                        # r2^n
 496         &paddq          ($D1,$T0);
 497
 498         &movdqa         ($T0,$T2);
 499         &pmuludq        ($T2,&QWP(16*2,$base));         # r2*h2
 500          &paddq         ($D0,$T1);
 501         &movdqa         ($T1,$T0);
 502         &pmuludq        ($T0,&QWP(16*1,$base));         # r2*h1
 503         &paddq          ($D4,$T2);
 504         &$load          ($T2,6);                        # s2^n
 505         &pmuludq        ($T1,&QWP(16*0,$base));         # r2*h0
 506         &paddq          ($D3,$T0);
 507         &movdqa         ($T0,$T2);
 508         &pmuludq        ($T2,&QWP(16*4,$base));         # s2*h4
 509         &paddq          ($D2,$T1);
 510         &pmuludq        ($T0,&QWP(16*3,$base));         # s2*h3
 511          &$load         ($T1,3);                        # r3^n
 512         &paddq          ($D1,$T2);
 513
 514         &movdqa         ($T2,$T1);
 515         &pmuludq        ($T1,&QWP(16*1,$base));         # r3*h1
 516          &paddq         ($D0,$T0);
 517         &$load          ($T0,7);                        # s3^n
 518         &pmuludq        ($T2,&QWP(16*0,$base));         # r3*h0
 519         &paddq          ($D4,$T1);
 520         &movdqa         ($T1,$T0);
 521         &pmuludq        ($T0,&QWP(16*4,$base));         # s3*h4
 522         &paddq          ($D3,$T2);
 523         &movdqa         ($T2,$T1);
 524         &pmuludq        ($T1,&QWP(16*3,$base));         # s3*h3
 525         &paddq          ($D2,$T0);
 526         &pmuludq        ($T2,&QWP(16*2,$base));         # s3*h2
 527          &$load         ($T0,4);                        # r4^n
 528         &paddq          ($D1,$T1);
 529
 530         &$load          ($T1,8);                        # s4^n
 531         &pmuludq        ($T0,&QWP(16*0,$base));         # r4*h0
 532          &paddq         ($D0,$T2);
 533         &movdqa         ($T2,$T1);
 534         &pmuludq        ($T1,&QWP(16*4,$base));         # s4*h4
 535         &paddq          ($D4,$T0);
 536         &movdqa         ($T0,$T2);
 537         &pmuludq        ($T2,&QWP(16*1,$base));         # s4*h1
 538         &paddq          ($D3,$T1);
 539         &movdqa         ($T1,$T0);
 540         &pmuludq        ($T0,&QWP(16*2,$base));         # s4*h2
 541         &paddq          ($D0,$T2);
 542         &pmuludq        ($T1,&QWP(16*3,$base));         # s4*h3
 543          &movdqa        ($MASK,&QWP(64,"ebx"));
 544         &paddq          ($D1,$T0);
 545         &paddq          ($D2,$T1);
 546 }
 547         &pmuladd        (sub {  my ($reg,$i)=@_;
 548                                 &movdqa ($reg,&QWP(16*$i,"esp"));
 549                              },"edx");
 550
 551 sub lazy_reduction {
 552 my $extra = shift;
 553
 554         ################################################################
 555         # lazy reduction as discussed in "NEON crypto" by D.J. Bernstein
 556         # and P. Schwabe
 557         #
 558         # [(*) see discussion in poly1305-armv4 module]
 559
 560          &movdqa        ($T0,$D3);
 561          &pand          ($D3,$MASK);
 562          &psrlq         ($T0,26);
 563          &$extra        ()                              if (defined($extra));
 564          &paddq         ($T0,$D4);                      # h3 -> h4
 565         &movdqa         ($T1,$D0);
 566         &pand           ($D0,$MASK);
 567         &psrlq          ($T1,26);
 568          &movdqa        ($D4,$T0);
 569         &paddq          ($T1,$D1);                      # h0 -> h1
 570          &psrlq         ($T0,26);
 571          &pand          ($D4,$MASK);
 572         &movdqa         ($D1,$T1);
 573         &psrlq          ($T1,26);
 574          &paddd         ($D0,$T0);                      # favour paddd when
 575                                                         # possible, because
 576                                                         # paddq is "broken"
 577                                                         # on Atom
 578          &psllq         ($T0,2);
 579         &paddq          ($T1,$D2);                      # h1 -> h2
 580          &paddq         ($T0,$D0);                      # h4 -> h0 (*)
 581         &pand           ($D1,$MASK);
 582         &movdqa         ($D2,$T1);
 583         &psrlq          ($T1,26);
 584         &pand           ($D2,$MASK);
 585         &paddd          ($T1,$D3);                      # h2 -> h3
 586          &movdqa        ($D0,$T0);
 587          &psrlq         ($T0,26);
 588         &movdqa         ($D3,$T1);
 589         &psrlq          ($T1,26);
 590          &pand          ($D0,$MASK);
 591          &paddd         ($D1,$T0);                      # h0 -> h1
 592         &pand           ($D3,$MASK);
 593         &paddd          ($D4,$T1);                      # h3 -> h4
 594 }
 595         &lazy_reduction ();
 596
 597         &dec            ("ecx");
 598         &jz             (&label("square_break"));
 599
 600         &punpcklqdq     ($D0,&QWP(16*0,"esp"));         # 0:r^1:0:r^2
 601         &punpcklqdq     ($D1,&QWP(16*1,"esp"));
 602         &punpcklqdq     ($D2,&QWP(16*2,"esp"));
 603         &punpcklqdq     ($D3,&QWP(16*3,"esp"));
 604         &punpcklqdq     ($D4,&QWP(16*4,"esp"));
 605         &jmp            (&label("square"));
 606
 607 &set_label("square_break");
 608         &psllq          ($D0,32);                       # -> r^3:0:r^4:0
 609         &psllq          ($D1,32);
 610         &psllq          ($D2,32);
 611         &psllq          ($D3,32);
 612         &psllq          ($D4,32);
 613         &por            ($D0,&QWP(16*0,"esp"));         # r^3:r^1:r^4:r^2
 614         &por            ($D1,&QWP(16*1,"esp"));
 615         &por            ($D2,&QWP(16*2,"esp"));
 616         &por            ($D3,&QWP(16*3,"esp"));
 617         &por            ($D4,&QWP(16*4,"esp"));
 618
 619         &pshufd         ($D0,$D0,0b10001101);           # -> r^1:r^2:r^3:r^4
 620         &pshufd         ($D1,$D1,0b10001101);
 621         &pshufd         ($D2,$D2,0b10001101);
 622         &pshufd         ($D3,$D3,0b10001101);
 623         &pshufd         ($D4,$D4,0b10001101);
 624
 625         &movdqu         (&QWP(16*0,"edi"),$D0);         # save the table
 626         &movdqu         (&QWP(16*1,"edi"),$D1);
 627         &movdqu         (&QWP(16*2,"edi"),$D2);
 628         &movdqu         (&QWP(16*3,"edi"),$D3);
 629         &movdqu         (&QWP(16*4,"edi"),$D4);
 630
 631         &movdqa         ($T1,$D1);
 632         &movdqa         ($T0,$D2);
 633         &pslld          ($T1,2);
 634         &pslld          ($T0,2);
 635         &paddd          ($T1,$D1);                      # *5
 636         &paddd          ($T0,$D2);                      # *5
 637         &movdqu         (&QWP(16*5,"edi"),$T1);
 638         &movdqu         (&QWP(16*6,"edi"),$T0);
 639         &movdqa         ($T1,$D3);
 640         &movdqa         ($T0,$D4);
 641         &pslld          ($T1,2);
 642         &pslld          ($T0,2);
 643         &paddd          ($T1,$D3);                      # *5
 644         &paddd          ($T0,$D4);                      # *5
 645         &movdqu         (&QWP(16*7,"edi"),$T1);
 646         &movdqu         (&QWP(16*8,"edi"),$T0);
 647
 648         &mov            ("esp","ebp");
 649         &lea            ("edi",&DWP(-16*3,"edi"));      # size de-optimization
 650         &ret            ();
 651 &function_end_B("_poly1305_init_sse2");
 652
 653 &align  (32);
 654 &function_begin("_poly1305_blocks_sse2");
 655         &mov    ("edi",&wparam(0));                     # ctx
 656         &mov    ("esi",&wparam(1));                     # inp
 657         &mov    ("ecx",&wparam(2));                     # len
 658
 659         &mov    ("eax",&DWP(4*5,"edi"));                # is_base2_26
 660         &and    ("ecx",-16);
 661         &jz     (&label("nodata"));
 662         &cmp    ("ecx",64);
 663         &jae    (&label("enter_sse2"));
 664         &test   ("eax","eax");                          # is_base2_26?
 665         &jz     (&label("enter_blocks"));
 666
 667 &set_label("enter_sse2",16);
 668         &call   (&label("pic_point"));
 669 &set_label("pic_point");
 670         &blindpop("ebx");
 671         &lea    ("ebx",&DWP(&label("const_sse2")."-".&label("pic_point"),"ebx"));
 672
 673         &test   ("eax","eax");                          # is_base2_26?
 674         &jnz    (&label("base2_26"));
 675
 676         &call   ("_poly1305_init_sse2");
 677
 678         ################################################# base 2^32 -> base 2^26
 679         &mov    ("eax",&DWP(0,"edi"));
 680         &mov    ("ecx",&DWP(3,"edi"));
 681         &mov    ("edx",&DWP(6,"edi"));
 682         &mov    ("esi",&DWP(9,"edi"));
 683         &mov    ("ebp",&DWP(13,"edi"));
 684         &mov    (&DWP(4*5,"edi"),1);                    # is_base2_26
 685
 686         &shr    ("ecx",2);
 687         &and    ("eax",0x3ffffff);
 688         &shr    ("edx",4);
 689         &and    ("ecx",0x3ffffff);
 690         &shr    ("esi",6);
 691         &and    ("edx",0x3ffffff);
 692
 693         &movd   ($D0,"eax");
 694         &movd   ($D1,"ecx");
 695         &movd   ($D2,"edx");
 696         &movd   ($D3,"esi");
 697         &movd   ($D4,"ebp");
 698
 699         &mov    ("esi",&wparam(1));                     # [reload] inp
 700         &mov    ("ecx",&wparam(2));                     # [reload] len
 701         &jmp    (&label("base2_32"));
 702
 703 &set_label("base2_26",16);
 704         &movd   ($D0,&DWP(4*0,"edi"));                  # load hash value
 705         &movd   ($D1,&DWP(4*1,"edi"));
 706         &movd   ($D2,&DWP(4*2,"edi"));
 707         &movd   ($D3,&DWP(4*3,"edi"));
 708         &movd   ($D4,&DWP(4*4,"edi"));
 709         &movdqa ($MASK,&QWP(64,"ebx"));
 710
 711 &set_label("base2_32");
 712         &mov    ("eax",&wparam(3));                     # padbit
 713         &mov    ("ebp","esp");
 714
 715         &sub    ("esp",16*(5+5+5+9+9));
 716         &and    ("esp",-16);
 717
 718         &lea    ("edi",&DWP(16*3,"edi"));               # size optimization
 719         &shl    ("eax",24);                             # padbit
 720
 721         &test   ("ecx",31);
 722         &jz     (&label("even"));
 723
 724         ################################################################
 725         # process single block, with SSE2, because it's still faster
 726         # even though half of result is discarded
 727
 728         &movdqu         ($T1,&QWP(0,"esi"));            # input
 729         &lea            ("esi",&DWP(16,"esi"));
 730
 731         &movdqa         ($T0,$T1);                      # -> base 2^26 ...
 732         &pand           ($T1,$MASK);
 733         &paddd          ($D0,$T1);                      # ... and accumulate
 734
 735         &movdqa         ($T1,$T0);
 736         &psrlq          ($T0,26);
 737         &psrldq         ($T1,6);
 738         &pand           ($T0,$MASK);
 739         &paddd          ($D1,$T0);
 740
 741         &movdqa         ($T0,$T1);
 742         &psrlq          ($T1,4);
 743         &pand           ($T1,$MASK);
 744         &paddd          ($D2,$T1);
 745
 746         &movdqa         ($T1,$T0);
 747         &psrlq          ($T0,30);
 748         &pand           ($T0,$MASK);
 749         &psrldq         ($T1,7);
 750         &paddd          ($D3,$T0);
 751
 752         &movd           ($T0,"eax");                    # padbit
 753         &paddd          ($D4,$T1);
 754          &movd          ($T1,&DWP(16*0+12,"edi"));      # r0
 755         &paddd          ($D4,$T0);
 756
 757         &movdqa         (&QWP(16*0,"esp"),$D0);
 758         &movdqa         (&QWP(16*1,"esp"),$D1);
 759         &movdqa         (&QWP(16*2,"esp"),$D2);
 760         &movdqa         (&QWP(16*3,"esp"),$D3);
 761         &movdqa         (&QWP(16*4,"esp"),$D4);
 762
 763         ################################################################
 764         # d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
 765         # d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
 766         # d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
 767         # d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
 768         # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
 769
 770         &pmuludq        ($D0,$T1);                      # h4*r0
 771         &pmuludq        ($D1,$T1);                      # h3*r0
 772         &pmuludq        ($D2,$T1);                      # h2*r0
 773          &movd          ($T0,&DWP(16*1+12,"edi"));      # r1
 774         &pmuludq        ($D3,$T1);                      # h1*r0
 775         &pmuludq        ($D4,$T1);                      # h0*r0
 776
 777         &pmuladd        (sub {  my ($reg,$i)=@_;
 778                                 &movd ($reg,&DWP(16*$i+12,"edi"));
 779                              });
 780
 781         &lazy_reduction ();
 782
 783         &sub            ("ecx",16);
 784         &jz             (&label("done"));
 785
 786 &set_label("even");
 787         &lea            ("edx",&DWP(16*(5+5+5+9),"esp"));# size optimization
 788         &lea            ("eax",&DWP(-16*2,"esi"));
 789         &sub            ("ecx",64);
 790
 791         ################################################################
 792         # expand and copy pre-calculated table to stack
 793
 794         &movdqu         ($T0,&QWP(16*0,"edi"));         # r^1:r^2:r^3:r^4
 795         &pshufd         ($T1,$T0,0b01000100);           # duplicate r^3:r^4
 796         &cmovb          ("esi","eax");
 797         &pshufd         ($T0,$T0,0b11101110);           # duplicate r^1:r^2
 798         &movdqa         (&QWP(16*0,"edx"),$T1);
 799         &lea            ("eax",&DWP(16*10,"esp"));
 800         &movdqu         ($T1,&QWP(16*1,"edi"));
 801         &movdqa         (&QWP(16*(0-9),"edx"),$T0);
 802         &pshufd         ($T0,$T1,0b01000100);
 803         &pshufd         ($T1,$T1,0b11101110);
 804         &movdqa         (&QWP(16*1,"edx"),$T0);
 805         &movdqu         ($T0,&QWP(16*2,"edi"));
 806         &movdqa         (&QWP(16*(1-9),"edx"),$T1);
 807         &pshufd         ($T1,$T0,0b01000100);
 808         &pshufd         ($T0,$T0,0b11101110);
 809         &movdqa         (&QWP(16*2,"edx"),$T1);
 810         &movdqu         ($T1,&QWP(16*3,"edi"));
 811         &movdqa         (&QWP(16*(2-9),"edx"),$T0);
 812         &pshufd         ($T0,$T1,0b01000100);
 813         &pshufd         ($T1,$T1,0b11101110);
 814         &movdqa         (&QWP(16*3,"edx"),$T0);
 815         &movdqu         ($T0,&QWP(16*4,"edi"));
 816         &movdqa         (&QWP(16*(3-9),"edx"),$T1);
 817         &pshufd         ($T1,$T0,0b01000100);
 818         &pshufd         ($T0,$T0,0b11101110);
 819         &movdqa         (&QWP(16*4,"edx"),$T1);
 820         &movdqu         ($T1,&QWP(16*5,"edi"));
 821         &movdqa         (&QWP(16*(4-9),"edx"),$T0);
 822         &pshufd         ($T0,$T1,0b01000100);
 823         &pshufd         ($T1,$T1,0b11101110);
 824         &movdqa         (&QWP(16*5,"edx"),$T0);
 825         &movdqu         ($T0,&QWP(16*6,"edi"));
 826         &movdqa         (&QWP(16*(5-9),"edx"),$T1);
 827         &pshufd         ($T1,$T0,0b01000100);
 828         &pshufd         ($T0,$T0,0b11101110);
 829         &movdqa         (&QWP(16*6,"edx"),$T1);
 830         &movdqu         ($T1,&QWP(16*7,"edi"));
 831         &movdqa         (&QWP(16*(6-9),"edx"),$T0);
 832         &pshufd         ($T0,$T1,0b01000100);
 833         &pshufd         ($T1,$T1,0b11101110);
 834         &movdqa         (&QWP(16*7,"edx"),$T0);
 835         &movdqu         ($T0,&QWP(16*8,"edi"));
 836         &movdqa         (&QWP(16*(7-9),"edx"),$T1);
 837         &pshufd         ($T1,$T0,0b01000100);
 838         &pshufd         ($T0,$T0,0b11101110);
 839         &movdqa         (&QWP(16*8,"edx"),$T1);
 840         &movdqa         (&QWP(16*(8-9),"edx"),$T0);
 841
 842 sub load_input {
 843 my ($inpbase,$offbase)=@_;
 844
 845         &movdqu         ($T0,&QWP($inpbase+0,"esi"));   # load input
 846         &movdqu         ($T1,&QWP($inpbase+16,"esi"));
 847         &lea            ("esi",&DWP(16*2,"esi"));
 848
 849         &movdqa         (&QWP($offbase+16*2,"esp"),$D2);
 850         &movdqa         (&QWP($offbase+16*3,"esp"),$D3);
 851         &movdqa         (&QWP($offbase+16*4,"esp"),$D4);
 852
 853         &movdqa         ($D2,$T0);                      # splat input
 854         &movdqa         ($D3,$T1);
 855         &psrldq         ($D2,6);
 856         &psrldq         ($D3,6);
 857         &movdqa         ($D4,$T0);
 858         &punpcklqdq     ($D2,$D3);                      # 2:3
 859         &punpckhqdq     ($D4,$T1);                      # 4
 860         &punpcklqdq     ($T0,$T1);                      # 0:1
 861
 862         &movdqa         ($D3,$D2);
 863         &psrlq          ($D2,4);
 864         &psrlq          ($D3,30);
 865         &movdqa         ($T1,$T0);
 866         &psrlq          ($D4,40);                       # 4
 867         &psrlq          ($T1,26);
 868         &pand           ($T0,$MASK);                    # 0
 869         &pand           ($T1,$MASK);                    # 1
 870         &pand           ($D2,$MASK);                    # 2
 871         &pand           ($D3,$MASK);                    # 3
 872         &por            ($D4,&QWP(0,"ebx"));            # padbit, yes, always
 873
 874         &movdqa         (&QWP($offbase+16*0,"esp"),$D0) if ($offbase);
 875         &movdqa         (&QWP($offbase+16*1,"esp"),$D1) if ($offbase);
 876 }
 877         &load_input     (16*2,16*5);
 878
 879         &jbe            (&label("skip_loop"));
 880         &jmp            (&label("loop"));
 881
 882 &set_label("loop",32);
 883         ################################################################
 884         # ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2
 885         # ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^3+inp[7]*r
 886         #   \___________________/
 887         # ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2+inp[8])*r^2
 888         # ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^4+inp[7]*r^2+inp[9])*r
 889         #   \___________________/ \____________________/
 890         ################################################################
 891
 892         &movdqa         ($T2,&QWP(16*(0-9),"edx"));     # r0^2
 893         &movdqa         (&QWP(16*1,"eax"),$T1);
 894         &movdqa         (&QWP(16*2,"eax"),$D2);
 895         &movdqa         (&QWP(16*3,"eax"),$D3);
 896         &movdqa         (&QWP(16*4,"eax"),$D4);
 897
 898         ################################################################
 899         # d4 = h4*r0 + h0*r4   + h1*r3   + h2*r2   + h3*r1
 900         # d3 = h3*r0 + h0*r3   + h1*r2   + h2*r1   + h4*5*r4
 901         # d2 = h2*r0 + h0*r2   + h1*r1   + h3*5*r4 + h4*5*r3
 902         # d1 = h1*r0 + h0*r1   + h2*5*r4 + h3*5*r3 + h4*5*r2
 903         # d0 = h0*r0 + h1*5*r4 + h2*5*r3 + h3*5*r2 + h4*5*r1
 904
 905         &movdqa         ($D1,$T0);
 906         &pmuludq        ($T0,$T2);                      # h0*r0
 907         &movdqa         ($D0,$T1);
 908         &pmuludq        ($T1,$T2);                      # h1*r0
 909         &pmuludq        ($D2,$T2);                      # h2*r0
 910         &pmuludq        ($D3,$T2);                      # h3*r0
 911         &pmuludq        ($D4,$T2);                      # h4*r0
 912
 913 sub pmuladd_alt {
 914 my $addr = shift;
 915
 916         &pmuludq        ($D0,&$addr(8));                # h1*s4
 917         &movdqa         ($T2,$D1);
 918         &pmuludq        ($D1,&$addr(1));                # h0*r1
 919         &paddq          ($D0,$T0);
 920         &movdqa         ($T0,$T2);
 921         &pmuludq        ($T2,&$addr(2));                # h0*r2
 922         &paddq          ($D1,$T1);
 923         &movdqa         ($T1,$T0);
 924         &pmuludq        ($T0,&$addr(3));                # h0*r3
 925         &paddq          ($D2,$T2);
 926          &movdqa        ($T2,&QWP(16*1,"eax"));         # pull h1
 927         &pmuludq        ($T1,&$addr(4));                # h0*r4
 928         &paddq          ($D3,$T0);
 929
 930         &movdqa         ($T0,$T2);
 931         &pmuludq        ($T2,&$addr(1));                # h1*r1
 932          &paddq         ($D4,$T1);
 933         &movdqa         ($T1,$T0);
 934         &pmuludq        ($T0,&$addr(2));                # h1*r2
 935         &paddq          ($D2,$T2);
 936         &movdqa         ($T2,&QWP(16*2,"eax"));         # pull h2
 937         &pmuludq        ($T1,&$addr(3));                # h1*r3
 938         &paddq          ($D3,$T0);
 939         &movdqa         ($T0,$T2);
 940         &pmuludq        ($T2,&$addr(7));                # h2*s3
 941         &paddq          ($D4,$T1);
 942         &movdqa         ($T1,$T0);
 943         &pmuludq        ($T0,&$addr(8));                # h2*s4
 944         &paddq          ($D0,$T2);
 945
 946         &movdqa         ($T2,$T1);
 947         &pmuludq        ($T1,&$addr(1));                # h2*r1
 948          &paddq         ($D1,$T0);
 949         &movdqa         ($T0,&QWP(16*3,"eax"));         # pull h3
 950         &pmuludq        ($T2,&$addr(2));                # h2*r2
 951         &paddq          ($D3,$T1);
 952         &movdqa         ($T1,$T0);
 953         &pmuludq        ($T0,&$addr(6));                # h3*s2
 954         &paddq          ($D4,$T2);
 955         &movdqa         ($T2,$T1);
 956         &pmuludq        ($T1,&$addr(7));                # h3*s3
 957         &paddq          ($D0,$T0);
 958         &movdqa         ($T0,$T2);
 959         &pmuludq        ($T2,&$addr(8));                # h3*s4
 960         &paddq          ($D1,$T1);
 961
 962         &movdqa         ($T1,&QWP(16*4,"eax"));         # pull h4
 963         &pmuludq        ($T0,&$addr(1));                # h3*r1
 964          &paddq         ($D2,$T2);
 965         &movdqa         ($T2,$T1);
 966         &pmuludq        ($T1,&$addr(8));                # h4*s4
 967         &paddq          ($D4,$T0);
 968         &movdqa         ($T0,$T2);
 969         &pmuludq        ($T2,&$addr(5));                # h4*s1
 970         &paddq          ($D3,$T1);
 971         &movdqa         ($T1,$T0);
 972         &pmuludq        ($T0,&$addr(6));                # h4*s2
 973         &paddq          ($D0,$T2);
 974          &movdqa        ($MASK,&QWP(64,"ebx"));
 975         &pmuludq        ($T1,&$addr(7));                # h4*s3
 976         &paddq          ($D1,$T0);
 977         &paddq          ($D2,$T1);
 978 }
 979         &pmuladd_alt    (sub {  my $i=shift; &QWP(16*($i-9),"edx");     });
 980
 981         &load_input     (-16*2,0);
 982         &lea            ("eax",&DWP(-16*2,"esi"));
 983         &sub            ("ecx",64);
 984
 985         &paddd          ($T0,&QWP(16*(5+0),"esp"));     # add hash value
 986         &paddd          ($T1,&QWP(16*(5+1),"esp"));
 987         &paddd          ($D2,&QWP(16*(5+2),"esp"));
 988         &paddd          ($D3,&QWP(16*(5+3),"esp"));
 989         &paddd          ($D4,&QWP(16*(5+4),"esp"));
 990
 991         &cmovb          ("esi","eax");
 992         &lea            ("eax",&DWP(16*10,"esp"));
 993
 994         &movdqa         ($T2,&QWP(16*0,"edx"));         # r0^4
 995         &movdqa         (&QWP(16*1,"esp"),$D1);
 996         &movdqa         (&QWP(16*1,"eax"),$T1);
 997         &movdqa         (&QWP(16*2,"eax"),$D2);
 998         &movdqa         (&QWP(16*3,"eax"),$D3);
 999         &movdqa         (&QWP(16*4,"eax"),$D4);
1000
1001         ################################################################
1002         # d4 += h4*r0 + h0*r4   + h1*r3   + h2*r2   + h3*r1
1003         # d3 += h3*r0 + h0*r3   + h1*r2   + h2*r1   + h4*5*r4
1004         # d2 += h2*r0 + h0*r2   + h1*r1   + h3*5*r4 + h4*5*r3
1005         # d1 += h1*r0 + h0*r1   + h2*5*r4 + h3*5*r3 + h4*5*r2
1006         # d0 += h0*r0 + h1*5*r4 + h2*5*r3 + h3*5*r2 + h4*5*r1
1007
1008         &movdqa         ($D1,$T0);
1009         &pmuludq        ($T0,$T2);                      # h0*r0
1010         &paddq          ($T0,$D0);
1011         &movdqa         ($D0,$T1);
1012         &pmuludq        ($T1,$T2);                      # h1*r0
1013         &pmuludq        ($D2,$T2);                      # h2*r0
1014         &pmuludq        ($D3,$T2);                      # h3*r0
1015         &pmuludq        ($D4,$T2);                      # h4*r0
1016
1017         &paddq          ($T1,&QWP(16*1,"esp"));
1018         &paddq          ($D2,&QWP(16*2,"esp"));
1019         &paddq          ($D3,&QWP(16*3,"esp"));
1020         &paddq          ($D4,&QWP(16*4,"esp"));
1021
1022         &pmuladd_alt    (sub {  my $i=shift; &QWP(16*$i,"edx"); });
1023
1024         &lazy_reduction ();
1025
1026         &load_input     (16*2,16*5);
1027
1028         &ja             (&label("loop"));
1029
1030 &set_label("skip_loop");
1031         ################################################################
1032         # multiply (inp[0:1]+hash) or inp[2:3] by r^2:r^1
1033
1034          &pshufd        ($T2,&QWP(16*(0-9),"edx"),0x10);# r0^n
1035         &add            ("ecx",32);
1036         &jnz            (&label("long_tail"));
1037
1038         &paddd          ($T0,$D0);                      # add hash value
1039         &paddd          ($T1,$D1);
1040         &paddd          ($D2,&QWP(16*7,"esp"));
1041         &paddd          ($D3,&QWP(16*8,"esp"));
1042         &paddd          ($D4,&QWP(16*9,"esp"));
1043
1044 &set_label("long_tail");
1045
1046         &movdqa         (&QWP(16*0,"eax"),$T0);
1047         &movdqa         (&QWP(16*1,"eax"),$T1);
1048         &movdqa         (&QWP(16*2,"eax"),$D2);
1049         &movdqa         (&QWP(16*3,"eax"),$D3);
1050         &movdqa         (&QWP(16*4,"eax"),$D4);
1051
1052         ################################################################
1053         # d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
1054         # d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
1055         # d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
1056         # d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
1057         # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1058
1059         &pmuludq        ($T0,$T2);                      # h0*r0
1060         &pmuludq        ($T1,$T2);                      # h1*r0
1061         &pmuludq        ($D2,$T2);                      # h2*r0
1062         &movdqa         ($D0,$T0);
1063          &pshufd        ($T0,&QWP(16*(1-9),"edx"),0x10);# r1^n
1064         &pmuludq        ($D3,$T2);                      # h3*r0
1065         &movdqa         ($D1,$T1);
1066         &pmuludq        ($D4,$T2);                      # h4*r0
1067
1068         &pmuladd        (sub {  my ($reg,$i)=@_;
1069                                 &pshufd ($reg,&QWP(16*($i-9),"edx"),0x10);
1070                              },"eax");
1071
1072         &jz             (&label("short_tail"));
1073
1074         &load_input     (-16*2,0);
1075
1076          &pshufd        ($T2,&QWP(16*0,"edx"),0x10);    # r0^n
1077         &paddd          ($T0,&QWP(16*5,"esp"));         # add hash value
1078         &paddd          ($T1,&QWP(16*6,"esp"));
1079         &paddd          ($D2,&QWP(16*7,"esp"));
1080         &paddd          ($D3,&QWP(16*8,"esp"));
1081         &paddd          ($D4,&QWP(16*9,"esp"));
1082
1083         ################################################################
1084         # multiply inp[0:1] by r^4:r^3 and accumulate
1085
1086         &movdqa         (&QWP(16*0,"esp"),$T0);
1087         &pmuludq        ($T0,$T2);                      # h0*r0
1088         &movdqa         (&QWP(16*1,"esp"),$T1);
1089         &pmuludq        ($T1,$T2);                      # h1*r0
1090         &paddq          ($D0,$T0);
1091         &movdqa         ($T0,$D2);
1092         &pmuludq        ($D2,$T2);                      # h2*r0
1093         &paddq          ($D1,$T1);
1094         &movdqa         ($T1,$D3);
1095         &pmuludq        ($D3,$T2);                      # h3*r0
1096         &paddq          ($D2,&QWP(16*2,"esp"));
1097         &movdqa         (&QWP(16*2,"esp"),$T0);
1098          &pshufd        ($T0,&QWP(16*1,"edx"),0x10);    # r1^n
1099         &paddq          ($D3,&QWP(16*3,"esp"));
1100         &movdqa         (&QWP(16*3,"esp"),$T1);
1101         &movdqa         ($T1,$D4);
1102         &pmuludq        ($D4,$T2);                      # h4*r0
1103         &paddq          ($D4,&QWP(16*4,"esp"));
1104         &movdqa         (&QWP(16*4,"esp"),$T1);
1105
1106         &pmuladd        (sub {  my ($reg,$i)=@_;
1107                                 &pshufd ($reg,&QWP(16*$i,"edx"),0x10);
1108                              });
1109
1110 &set_label("short_tail");
1111
1112         ################################################################
1113         # horizontal addition
1114
1115         &pshufd         ($T1,$D4,0b01001110);
1116         &pshufd         ($T0,$D3,0b01001110);
1117         &paddq          ($D4,$T1);
1118         &paddq          ($D3,$T0);
1119         &pshufd         ($T1,$D0,0b01001110);
1120         &pshufd         ($T0,$D1,0b01001110);
1121         &paddq          ($D0,$T1);
1122         &paddq          ($D1,$T0);
1123         &pshufd         ($T1,$D2,0b01001110);
1124         #&paddq         ($D2,$T1);
1125
1126         &lazy_reduction (sub { &paddq ($D2,$T1) });
1127
1128 &set_label("done");
1129         &movd           (&DWP(-16*3+4*0,"edi"),$D0);    # store hash value
1130         &movd           (&DWP(-16*3+4*1,"edi"),$D1);
1131         &movd           (&DWP(-16*3+4*2,"edi"),$D2);
1132         &movd           (&DWP(-16*3+4*3,"edi"),$D3);
1133         &movd           (&DWP(-16*3+4*4,"edi"),$D4);
1134         &mov    ("esp","ebp");
1135 &set_label("nodata");
1136 &function_end("_poly1305_blocks_sse2");
1137
1138 &align  (32);
1139 &function_begin("_poly1305_emit_sse2");
1140         &mov    ("ebp",&wparam(0));             # context
1141
1142         &cmp    (&DWP(4*5,"ebp"),0);            # is_base2_26?
1143         &je     (&label("enter_emit"));
1144
1145         &mov    ("eax",&DWP(4*0,"ebp"));        # load hash value
1146         &mov    ("edi",&DWP(4*1,"ebp"));
1147         &mov    ("ecx",&DWP(4*2,"ebp"));
1148         &mov    ("edx",&DWP(4*3,"ebp"));
1149         &mov    ("esi",&DWP(4*4,"ebp"));
1150
1151         &mov    ("ebx","edi");                  # base 2^26 -> base 2^32
1152         &shl    ("edi",26);
1153         &shr    ("ebx",6);
1154         &add    ("eax","edi");
1155         &mov    ("edi","ecx");
1156         &adc    ("ebx",0);
1157
1158         &shl    ("edi",20);
1159         &shr    ("ecx",12);
1160         &add    ("ebx","edi");
1161         &mov    ("edi","edx");
1162         &adc    ("ecx",0);
1163
1164         &shl    ("edi",14);
1165         &shr    ("edx",18);
1166         &add    ("ecx","edi");
1167         &mov    ("edi","esi");
1168         &adc    ("edx",0);
1169
1170         &shl    ("edi",8);
1171         &shr    ("esi",24);
1172         &add    ("edx","edi");
1173         &adc    ("esi",0);                      # can be partially reduced
1174
1175         &mov    ("edi","esi");                  # final reduction
1176         &and    ("esi",3);
1177         &shr    ("edi",2);
1178         &lea    ("ebp",&DWP(0,"edi","edi",4));  # *5
1179          &mov   ("edi",&wparam(1));             # output
1180         &add    ("eax","ebp");
1181          &mov   ("ebp",&wparam(2));             # key
1182         &adc    ("ebx",0);
1183         &adc    ("ecx",0);
1184         &adc    ("edx",0);
1185         &adc    ("esi",0);
1186
1187         &movd   ($D0,"eax");                    # offload original hash value
1188         &add    ("eax",5);                      # compare to modulus
1189         &movd   ($D1,"ebx");
1190         &adc    ("ebx",0);
1191         &movd   ($D2,"ecx");
1192         &adc    ("ecx",0);
1193         &movd   ($D3,"edx");
1194         &adc    ("edx",0);
1195         &adc    ("esi",0);
1196         &shr    ("esi",2);                      # did it carry/borrow?
1197
1198         &neg    ("esi");                        # do we choose (hash-modulus) ...
1199         &and    ("eax","esi");
1200         &and    ("ebx","esi");
1201         &and    ("ecx","esi");
1202         &and    ("edx","esi");
1203         &mov    (&DWP(4*0,"edi"),"eax");
1204         &movd   ("eax",$D0);
1205         &mov    (&DWP(4*1,"edi"),"ebx");
1206         &movd   ("ebx",$D1);
1207         &mov    (&DWP(4*2,"edi"),"ecx");
1208         &movd   ("ecx",$D2);
1209         &mov    (&DWP(4*3,"edi"),"edx");
1210         &movd   ("edx",$D3);
1211
1212         &not    ("esi");                        # ... or original hash value?
1213         &and    ("eax","esi");
1214         &and    ("ebx","esi");
1215         &or     ("eax",&DWP(4*0,"edi"));
1216         &and    ("ecx","esi");
1217         &or     ("ebx",&DWP(4*1,"edi"));
1218         &and    ("edx","esi");
1219         &or     ("ecx",&DWP(4*2,"edi"));
1220         &or     ("edx",&DWP(4*3,"edi"));
1221
1222         &add    ("eax",&DWP(4*0,"ebp"));        # accumulate key
1223         &adc    ("ebx",&DWP(4*1,"ebp"));
1224         &mov    (&DWP(4*0,"edi"),"eax");
1225         &adc    ("ecx",&DWP(4*2,"ebp"));
1226         &mov    (&DWP(4*1,"edi"),"ebx");
1227         &adc    ("edx",&DWP(4*3,"ebp"));
1228         &mov    (&DWP(4*2,"edi"),"ecx");
1229         &mov    (&DWP(4*3,"edi"),"edx");
1230 &function_end("_poly1305_emit_sse2");
1231
1232 if ($avx>1) {
1233 ########################################################################
1234 # Note that poly1305_init_avx2 operates on %xmm, I could have used
1235 # poly1305_init_sse2...
1236
1237 &align  (32);
1238 &function_begin_B("_poly1305_init_avx2");
1239         &vmovdqu        ($D4,&QWP(4*6,"edi"));          # key base 2^32
1240         &lea            ("edi",&DWP(16*3,"edi"));       # size optimization
1241         &mov            ("ebp","esp");
1242         &sub            ("esp",16*(9+5));
1243         &and            ("esp",-16);
1244
1245         #&vpand         ($D4,$D4,&QWP(96,"ebx"));       # magic mask
1246         &vmovdqa        ($MASK,&QWP(64,"ebx"));
1247
1248         &vpand          ($D0,$D4,$MASK);                # -> base 2^26
1249         &vpsrlq         ($D1,$D4,26);
1250         &vpsrldq        ($D3,$D4,6);
1251         &vpand          ($D1,$D1,$MASK);
1252         &vpsrlq         ($D2,$D3,4)
1253         &vpsrlq         ($D3,$D3,30);
1254         &vpand          ($D2,$D2,$MASK);
1255         &vpand          ($D3,$D3,$MASK);
1256         &vpsrldq        ($D4,$D4,13);
1257
1258         &lea            ("edx",&DWP(16*9,"esp"));       # size optimization
1259         &mov            ("ecx",2);
1260 &set_label("square");
1261         &vmovdqa        (&QWP(16*0,"esp"),$D0);
1262         &vmovdqa        (&QWP(16*1,"esp"),$D1);
1263         &vmovdqa        (&QWP(16*2,"esp"),$D2);
1264         &vmovdqa        (&QWP(16*3,"esp"),$D3);
1265         &vmovdqa        (&QWP(16*4,"esp"),$D4);
1266
1267         &vpslld         ($T1,$D1,2);
1268         &vpslld         ($T0,$D2,2);
1269         &vpaddd         ($T1,$T1,$D1);                  # *5
1270         &vpaddd         ($T0,$T0,$D2);                  # *5
1271         &vmovdqa        (&QWP(16*5,"esp"),$T1);
1272         &vmovdqa        (&QWP(16*6,"esp"),$T0);
1273         &vpslld         ($T1,$D3,2);
1274         &vpslld         ($T0,$D4,2);
1275         &vpaddd         ($T1,$T1,$D3);                  # *5
1276         &vpaddd         ($T0,$T0,$D4);                  # *5
1277         &vmovdqa        (&QWP(16*7,"esp"),$T1);
1278         &vmovdqa        (&QWP(16*8,"esp"),$T0);
1279
1280         &vpshufd        ($T0,$D0,0b01000100);
1281         &vmovdqa        ($T1,$D1);
1282         &vpshufd        ($D1,$D1,0b01000100);
1283         &vpshufd        ($D2,$D2,0b01000100);
1284         &vpshufd        ($D3,$D3,0b01000100);
1285         &vpshufd        ($D4,$D4,0b01000100);
1286         &vmovdqa        (&QWP(16*0,"edx"),$T0);
1287         &vmovdqa        (&QWP(16*1,"edx"),$D1);
1288         &vmovdqa        (&QWP(16*2,"edx"),$D2);
1289         &vmovdqa        (&QWP(16*3,"edx"),$D3);
1290         &vmovdqa        (&QWP(16*4,"edx"),$D4);
1291
1292         ################################################################
1293         # d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
1294         # d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
1295         # d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
1296         # d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
1297         # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1298
1299         &vpmuludq       ($D4,$D4,$D0);                  # h4*r0
1300         &vpmuludq       ($D3,$D3,$D0);                  # h3*r0
1301         &vpmuludq       ($D2,$D2,$D0);                  # h2*r0
1302         &vpmuludq       ($D1,$D1,$D0);                  # h1*r0
1303         &vpmuludq       ($D0,$T0,$D0);                  # h0*r0
1304
1305         &vpmuludq       ($T0,$T1,&QWP(16*3,"edx"));     # r1*h3
1306         &vpaddq         ($D4,$D4,$T0);
1307         &vpmuludq       ($T2,$T1,&QWP(16*2,"edx"));     # r1*h2
1308         &vpaddq         ($D3,$D3,$T2);
1309         &vpmuludq       ($T0,$T1,&QWP(16*1,"edx"));     # r1*h1
1310         &vpaddq         ($D2,$D2,$T0);
1311         &vmovdqa        ($T2,&QWP(16*5,"esp"));         # s1
1312         &vpmuludq       ($T1,$T1,&QWP(16*0,"edx"));     # r1*h0
1313         &vpaddq         ($D1,$D1,$T1);
1314          &vmovdqa       ($T0,&QWP(16*2,"esp"));         # r2
1315         &vpmuludq       ($T2,$T2,&QWP(16*4,"edx"));     # s1*h4
1316         &vpaddq         ($D0,$D0,$T2);
1317
1318         &vpmuludq       ($T1,$T0,&QWP(16*2,"edx"));     # r2*h2
1319         &vpaddq         ($D4,$D4,$T1);
1320         &vpmuludq       ($T2,$T0,&QWP(16*1,"edx"));     # r2*h1
1321         &vpaddq         ($D3,$D3,$T2);
1322         &vmovdqa        ($T1,&QWP(16*6,"esp"));         # s2
1323         &vpmuludq       ($T0,$T0,&QWP(16*0,"edx"));     # r2*h0
1324         &vpaddq         ($D2,$D2,$T0);
1325         &vpmuludq       ($T2,$T1,&QWP(16*4,"edx"));     # s2*h4
1326         &vpaddq         ($D1,$D1,$T2);
1327          &vmovdqa       ($T0,&QWP(16*3,"esp"));         # r3
1328         &vpmuludq       ($T1,$T1,&QWP(16*3,"edx"));     # s2*h3
1329         &vpaddq         ($D0,$D0,$T1);
1330
1331         &vpmuludq       ($T2,$T0,&QWP(16*1,"edx"));     # r3*h1
1332         &vpaddq         ($D4,$D4,$T2);
1333         &vmovdqa        ($T1,&QWP(16*7,"esp"));         # s3
1334         &vpmuludq       ($T0,$T0,&QWP(16*0,"edx"));     # r3*h0
1335         &vpaddq         ($D3,$D3,$T0);
1336         &vpmuludq       ($T2,$T1,&QWP(16*4,"edx"));     # s3*h4
1337         &vpaddq         ($D2,$D2,$T2);
1338         &vpmuludq       ($T0,$T1,&QWP(16*3,"edx"));     # s3*h3
1339         &vpaddq         ($D1,$D1,$T0);
1340          &vmovdqa       ($T2,&QWP(16*4,"esp"));         # r4
1341         &vpmuludq       ($T1,$T1,&QWP(16*2,"edx"));     # s3*h2
1342         &vpaddq         ($D0,$D0,$T1);
1343
1344         &vmovdqa        ($T0,&QWP(16*8,"esp"));         # s4
1345         &vpmuludq       ($T2,$T2,&QWP(16*0,"edx"));     # r4*h0
1346         &vpaddq         ($D4,$D4,$T2);
1347         &vpmuludq       ($T1,$T0,&QWP(16*4,"edx"));     # s4*h4
1348         &vpaddq         ($D3,$D3,$T1);
1349         &vpmuludq       ($T2,$T0,&QWP(16*1,"edx"));     # s4*h1
1350         &vpaddq         ($D0,$D0,$T2);
1351         &vpmuludq       ($T1,$T0,&QWP(16*2,"edx"));     # s4*h2
1352         &vpaddq         ($D1,$D1,$T1);
1353          &vmovdqa       ($MASK,&QWP(64,"ebx"));
1354         &vpmuludq       ($T0,$T0,&QWP(16*3,"edx"));     # s4*h3
1355         &vpaddq         ($D2,$D2,$T0);
1356
1357         ################################################################
1358         # lazy reduction
1359          &vpsrlq        ($T0,$D3,26);
1360          &vpand         ($D3,$D3,$MASK);
1361         &vpsrlq         ($T1,$D0,26);
1362         &vpand          ($D0,$D0,$MASK);
1363          &vpaddq        ($D4,$D4,$T0);                  # h3 -> h4
1364         &vpaddq         ($D1,$D1,$T1);                  # h0 -> h1
1365          &vpsrlq        ($T0,$D4,26);
1366          &vpand         ($D4,$D4,$MASK);
1367         &vpsrlq         ($T1,$D1,26);
1368         &vpand          ($D1,$D1,$MASK);
1369         &vpaddq         ($D2,$D2,$T1);                  # h1 -> h2
1370          &vpaddd        ($D0,$D0,$T0);
1371          &vpsllq        ($T0,$T0,2);
1372         &vpsrlq         ($T1,$D2,26);
1373         &vpand          ($D2,$D2,$MASK);
1374          &vpaddd        ($D0,$D0,$T0);                  # h4 -> h0
1375         &vpaddd         ($D3,$D3,$T1);                  # h2 -> h3
1376         &vpsrlq         ($T1,$D3,26);
1377          &vpsrlq        ($T0,$D0,26);
1378          &vpand         ($D0,$D0,$MASK);
1379         &vpand          ($D3,$D3,$MASK);
1380          &vpaddd        ($D1,$D1,$T0);                  # h0 -> h1
1381         &vpaddd         ($D4,$D4,$T1);                  # h3 -> h4
1382
1383         &dec            ("ecx");
1384         &jz             (&label("square_break"));
1385
1386         &vpunpcklqdq    ($D0,$D0,&QWP(16*0,"esp"));     # 0:r^1:0:r^2
1387         &vpunpcklqdq    ($D1,$D1,&QWP(16*1,"esp"));
1388         &vpunpcklqdq    ($D2,$D2,&QWP(16*2,"esp"));
1389         &vpunpcklqdq    ($D3,$D3,&QWP(16*3,"esp"));
1390         &vpunpcklqdq    ($D4,$D4,&QWP(16*4,"esp"));
1391         &jmp            (&label("square"));
1392
1393 &set_label("square_break");
1394         &vpsllq         ($D0,$D0,32);                   # -> r^3:0:r^4:0
1395         &vpsllq         ($D1,$D1,32);
1396         &vpsllq         ($D2,$D2,32);
1397         &vpsllq         ($D3,$D3,32);
1398         &vpsllq         ($D4,$D4,32);
1399         &vpor           ($D0,$D0,&QWP(16*0,"esp"));     # r^3:r^1:r^4:r^2
1400         &vpor           ($D1,$D1,&QWP(16*1,"esp"));
1401         &vpor           ($D2,$D2,&QWP(16*2,"esp"));
1402         &vpor           ($D3,$D3,&QWP(16*3,"esp"));
1403         &vpor           ($D4,$D4,&QWP(16*4,"esp"));
1404
1405         &vpshufd        ($D0,$D0,0b10001101);           # -> r^1:r^2:r^3:r^4
1406         &vpshufd        ($D1,$D1,0b10001101);
1407         &vpshufd        ($D2,$D2,0b10001101);
1408         &vpshufd        ($D3,$D3,0b10001101);
1409         &vpshufd        ($D4,$D4,0b10001101);
1410
1411         &vmovdqu        (&QWP(16*0,"edi"),$D0);         # save the table
1412         &vmovdqu        (&QWP(16*1,"edi"),$D1);
1413         &vmovdqu        (&QWP(16*2,"edi"),$D2);
1414         &vmovdqu        (&QWP(16*3,"edi"),$D3);
1415         &vmovdqu        (&QWP(16*4,"edi"),$D4);
1416
1417         &vpslld         ($T1,$D1,2);
1418         &vpslld         ($T0,$D2,2);
1419         &vpaddd         ($T1,$T1,$D1);                  # *5
1420         &vpaddd         ($T0,$T0,$D2);                  # *5
1421         &vmovdqu        (&QWP(16*5,"edi"),$T1);
1422         &vmovdqu        (&QWP(16*6,"edi"),$T0);
1423         &vpslld         ($T1,$D3,2);
1424         &vpslld         ($T0,$D4,2);
1425         &vpaddd         ($T1,$T1,$D3);                  # *5
1426         &vpaddd         ($T0,$T0,$D4);                  # *5
1427         &vmovdqu        (&QWP(16*7,"edi"),$T1);
1428         &vmovdqu        (&QWP(16*8,"edi"),$T0);
1429
1430         &mov            ("esp","ebp");
1431         &lea            ("edi",&DWP(-16*3,"edi"));      # size de-optimization
1432         &ret            ();
1433 &function_end_B("_poly1305_init_avx2");
1434
1435 ########################################################################
1436 # now it's time to switch to %ymm
1437
1438 my ($D0,$D1,$D2,$D3,$D4,$T0,$T1,$T2)=map("ymm$_",(0..7));
1439 my $MASK=$T2;
1440
1441 sub X { my $reg=shift; $reg=~s/^ymm/xmm/; $reg; }
1442
1443 &align  (32);
1444 &function_begin("_poly1305_blocks_avx2");
1445         &mov    ("edi",&wparam(0));                     # ctx
1446         &mov    ("esi",&wparam(1));                     # inp
1447         &mov    ("ecx",&wparam(2));                     # len
1448
1449         &mov    ("eax",&DWP(4*5,"edi"));                # is_base2_26
1450         &and    ("ecx",-16);
1451         &jz     (&label("nodata"));
1452         &cmp    ("ecx",64);
1453         &jae    (&label("enter_avx2"));
1454         &test   ("eax","eax");                          # is_base2_26?
1455         &jz     (&label("enter_blocks"));
1456
1457 &set_label("enter_avx2");
1458         &vzeroupper     ();
1459
1460         &call   (&label("pic_point"));
1461 &set_label("pic_point");
1462         &blindpop("ebx");
1463         &lea    ("ebx",&DWP(&label("const_sse2")."-".&label("pic_point"),"ebx"));
1464
1465         &test   ("eax","eax");                          # is_base2_26?
1466         &jnz    (&label("base2_26"));
1467
1468         &call   ("_poly1305_init_avx2");
1469
1470         ################################################# base 2^32 -> base 2^26
1471         &mov    ("eax",&DWP(0,"edi"));
1472         &mov    ("ecx",&DWP(3,"edi"));
1473         &mov    ("edx",&DWP(6,"edi"));
1474         &mov    ("esi",&DWP(9,"edi"));
1475         &mov    ("ebp",&DWP(13,"edi"));
1476
1477         &shr    ("ecx",2);
1478         &and    ("eax",0x3ffffff);
1479         &shr    ("edx",4);
1480         &and    ("ecx",0x3ffffff);
1481         &shr    ("esi",6);
1482         &and    ("edx",0x3ffffff);
1483
1484         &mov    (&DWP(4*0,"edi"),"eax");
1485         &mov    (&DWP(4*1,"edi"),"ecx");
1486         &mov    (&DWP(4*2,"edi"),"edx");
1487         &mov    (&DWP(4*3,"edi"),"esi");
1488         &mov    (&DWP(4*4,"edi"),"ebp");
1489         &mov    (&DWP(4*5,"edi"),1);                    # is_base2_26
1490
1491         &mov    ("esi",&wparam(1));                     # [reload] inp
1492         &mov    ("ecx",&wparam(2));                     # [reload] len
1493
1494 &set_label("base2_26");
1495         &mov    ("eax",&wparam(3));                     # padbit
1496         &mov    ("ebp","esp");
1497
1498         &sub    ("esp",32*(5+9));
1499         &and    ("esp",-512);                           # ensure that frame
1500                                                         # doesn't cross page
1501                                                         # boundary, which is
1502                                                         # essential for
1503                                                         # misaligned 32-byte
1504                                                         # loads
1505
1506         ################################################################
1507         # expand and copy pre-calculated table to stack
1508
1509         &vmovdqu        (&X($D0),&QWP(16*(3+0),"edi"));
1510         &lea            ("edx",&DWP(32*5+128,"esp"));   # +128 size optimization
1511         &vmovdqu        (&X($D1),&QWP(16*(3+1),"edi"));
1512         &vmovdqu        (&X($D2),&QWP(16*(3+2),"edi"));
1513         &vmovdqu        (&X($D3),&QWP(16*(3+3),"edi"));
1514         &vmovdqu        (&X($D4),&QWP(16*(3+4),"edi"));
1515         &lea            ("edi",&DWP(16*3,"edi"));       # size optimization
1516         &vpermq         ($D0,$D0,0b01000000);           # 00001234 -> 12343434
1517         &vpermq         ($D1,$D1,0b01000000);
1518         &vpermq         ($D2,$D2,0b01000000);
1519         &vpermq         ($D3,$D3,0b01000000);
1520         &vpermq         ($D4,$D4,0b01000000);
1521         &vpshufd        ($D0,$D0,0b11001000);           # 12343434 -> 14243444
1522         &vpshufd        ($D1,$D1,0b11001000);
1523         &vpshufd        ($D2,$D2,0b11001000);
1524         &vpshufd        ($D3,$D3,0b11001000);
1525         &vpshufd        ($D4,$D4,0b11001000);
1526         &vmovdqa        (&QWP(32*0-128,"edx"),$D0);
1527         &vmovdqu        (&X($D0),&QWP(16*5,"edi"));
1528         &vmovdqa        (&QWP(32*1-128,"edx"),$D1);
1529         &vmovdqu        (&X($D1),&QWP(16*6,"edi"));
1530         &vmovdqa        (&QWP(32*2-128,"edx"),$D2);
1531         &vmovdqu        (&X($D2),&QWP(16*7,"edi"));
1532         &vmovdqa        (&QWP(32*3-128,"edx"),$D3);
1533         &vmovdqu        (&X($D3),&QWP(16*8,"edi"));
1534         &vmovdqa        (&QWP(32*4-128,"edx"),$D4);
1535         &vpermq         ($D0,$D0,0b01000000);
1536         &vpermq         ($D1,$D1,0b01000000);
1537         &vpermq         ($D2,$D2,0b01000000);
1538         &vpermq         ($D3,$D3,0b01000000);
1539         &vpshufd        ($D0,$D0,0b11001000);
1540         &vpshufd        ($D1,$D1,0b11001000);
1541         &vpshufd        ($D2,$D2,0b11001000);
1542         &vpshufd        ($D3,$D3,0b11001000);
1543         &vmovdqa        (&QWP(32*5-128,"edx"),$D0);
1544         &vmovd          (&X($D0),&DWP(-16*3+4*0,"edi"));# load hash value
1545         &vmovdqa        (&QWP(32*6-128,"edx"),$D1);
1546         &vmovd          (&X($D1),&DWP(-16*3+4*1,"edi"));
1547         &vmovdqa        (&QWP(32*7-128,"edx"),$D2);
1548         &vmovd          (&X($D2),&DWP(-16*3+4*2,"edi"));
1549         &vmovdqa        (&QWP(32*8-128,"edx"),$D3);
1550         &vmovd          (&X($D3),&DWP(-16*3+4*3,"edi"));
1551         &vmovd          (&X($D4),&DWP(-16*3+4*4,"edi"));
1552         &vmovdqa        ($MASK,&QWP(64,"ebx"));
1553         &neg            ("eax");                        # padbit
1554
1555         &test           ("ecx",63);
1556         &jz             (&label("even"));
1557
1558         &mov            ("edx","ecx");
1559         &and            ("ecx",-64);
1560         &and            ("edx",63);
1561
1562         &vmovdqu        (&X($T0),&QWP(16*0,"esi"));
1563         &cmp            ("edx",32);
1564         &jb             (&label("one"));
1565
1566         &vmovdqu        (&X($T1),&QWP(16*1,"esi"));
1567         &je             (&label("two"));
1568
1569         &vinserti128    ($T0,$T0,&QWP(16*2,"esi"),1);
1570         &lea            ("esi",&DWP(16*3,"esi"));
1571         &lea            ("ebx",&DWP(8,"ebx"));          # three padbits
1572         &lea            ("edx",&DWP(32*5+128+8,"esp")); # --:r^1:r^2:r^3 (*)
1573         &jmp            (&label("tail"));
1574
1575 &set_label("two");
1576         &lea            ("esi",&DWP(16*2,"esi"));
1577         &lea            ("ebx",&DWP(16,"ebx"));         # two padbits
1578         &lea            ("edx",&DWP(32*5+128+16,"esp"));# --:--:r^1:r^2 (*)
1579         &jmp            (&label("tail"));
1580
1581 &set_label("one");
1582         &lea            ("esi",&DWP(16*1,"esi"));
1583         &vpxor          ($T1,$T1,$T1);
1584         &lea            ("ebx",&DWP(32,"ebx","eax",8)); # one or no padbits
1585         &lea            ("edx",&DWP(32*5+128+24,"esp"));# --:--:--:r^1 (*)
1586         &jmp            (&label("tail"));
1587
1588 # (*)   spots marked with '--' are data from next table entry, but they
1589 #       are multiplied by 0 and therefore rendered insignificant
1590
1591 &set_label("even",32);
1592         &vmovdqu        (&X($T0),&QWP(16*0,"esi"));     # load input
1593         &vmovdqu        (&X($T1),&QWP(16*1,"esi"));
1594         &vinserti128    ($T0,$T0,&QWP(16*2,"esi"),1);
1595         &vinserti128    ($T1,$T1,&QWP(16*3,"esi"),1);
1596         &lea            ("esi",&DWP(16*4,"esi"));
1597         &sub            ("ecx",64);
1598         &jz             (&label("tail"));
1599
1600 &set_label("loop");
1601         ################################################################
1602         # ((inp[0]*r^4+r[4])*r^4+r[8])*r^4
1603         # ((inp[1]*r^4+r[5])*r^4+r[9])*r^3
1604         # ((inp[2]*r^4+r[6])*r^4+r[10])*r^2
1605         # ((inp[3]*r^4+r[7])*r^4+r[11])*r^1
1606         #   \________/ \_______/
1607         ################################################################
1608
1609 sub vsplat_input {
1610         &vmovdqa        (&QWP(32*2,"esp"),$D2);
1611         &vpsrldq        ($D2,$T0,6);                    # splat input
1612         &vmovdqa        (&QWP(32*0,"esp"),$D0);
1613         &vpsrldq        ($D0,$T1,6);
1614         &vmovdqa        (&QWP(32*1,"esp"),$D1);
1615         &vpunpckhqdq    ($D1,$T0,$T1);                  # 4
1616         &vpunpcklqdq    ($T0,$T0,$T1);                  # 0:1
1617         &vpunpcklqdq    ($D2,$D2,$D0);                  # 2:3
1618
1619         &vpsrlq         ($D0,$D2,30);
1620         &vpsrlq         ($D2,$D2,4);
1621         &vpsrlq         ($T1,$T0,26);
1622         &vpsrlq         ($D1,$D1,40);                   # 4
1623         &vpand          ($D2,$D2,$MASK);                # 2
1624         &vpand          ($T0,$T0,$MASK);                # 0
1625         &vpand          ($T1,$T1,$MASK);                # 1
1626         &vpand          ($D0,$D0,$MASK);                # 3 (*)
1627         &vpor           ($D1,$D1,&QWP(0,"ebx"));        # padbit, yes, always
1628
1629         # (*)   note that output is counterintuitive, inp[3:4] is
1630         #       returned in $D1-2, while $D3-4 are preserved;
1631 }
1632         &vsplat_input   ();
1633
1634 sub vpmuladd {
1635 my $addr = shift;
1636
1637         &vpaddq         ($D2,$D2,&QWP(32*2,"esp"));     # add hash value
1638         &vpaddq         ($T0,$T0,&QWP(32*0,"esp"));
1639         &vpaddq         ($T1,$T1,&QWP(32*1,"esp"));
1640         &vpaddq         ($D0,$D0,$D3);
1641         &vpaddq         ($D1,$D1,$D4);
1642
1643         ################################################################
1644         # d3 = h2*r1   + h0*r3 + h1*r2   + h3*r0   + h4*5*r4
1645         # d4 = h2*r2   + h0*r4 + h1*r3   + h3*r1   + h4*r0
1646         # d0 = h2*5*r3 + h0*r0 + h1*5*r4 + h3*5*r2 + h4*5*r1
1647         # d1 = h2*5*r4 + h0*r1 + h1*r0   + h3*5*r3 + h4*5*r2
1648         # d2 = h2*r0   + h0*r2 + h1*r1   + h3*5*r4 + h4*5*r3
1649
1650         &vpmuludq       ($D3,$D2,&$addr(1));            # d3 = h2*r1
1651          &vmovdqa       (QWP(32*1,"esp"),$T1);
1652         &vpmuludq       ($D4,$D2,&$addr(2));            # d4 = h2*r2
1653          &vmovdqa       (QWP(32*3,"esp"),$D0);
1654         &vpmuludq       ($D0,$D2,&$addr(7));            # d0 = h2*s3
1655          &vmovdqa       (QWP(32*4,"esp"),$D1);
1656         &vpmuludq       ($D1,$D2,&$addr(8));            # d1 = h2*s4
1657         &vpmuludq       ($D2,$D2,&$addr(0));            # d2 = h2*r0
1658
1659         &vpmuludq       ($T2,$T0,&$addr(3));            # h0*r3
1660         &vpaddq         ($D3,$D3,$T2);                  # d3 += h0*r3
1661         &vpmuludq       ($T1,$T0,&$addr(4));            # h0*r4
1662         &vpaddq         ($D4,$D4,$T1);                  # d4 + h0*r4
1663         &vpmuludq       ($T2,$T0,&$addr(0));            # h0*r0
1664         &vpaddq         ($D0,$D0,$T2);                  # d0 + h0*r0
1665          &vmovdqa       ($T2,&QWP(32*1,"esp"));         # h1
1666         &vpmuludq       ($T1,$T0,&$addr(1));            # h0*r1
1667         &vpaddq         ($D1,$D1,$T1);                  # d1 += h0*r1
1668         &vpmuludq       ($T0,$T0,&$addr(2));            # h0*r2
1669         &vpaddq         ($D2,$D2,$T0);                  # d2 += h0*r2
1670
1671         &vpmuludq       ($T1,$T2,&$addr(2));            # h1*r2
1672         &vpaddq         ($D3,$D3,$T1);                  # d3 += h1*r2
1673         &vpmuludq       ($T0,$T2,&$addr(3));            # h1*r3
1674         &vpaddq         ($D4,$D4,$T0);                  # d4 += h1*r3
1675         &vpmuludq       ($T1,$T2,&$addr(8));            # h1*s4
1676         &vpaddq         ($D0,$D0,$T1);                  # d0 += h1*s4
1677          &vmovdqa       ($T1,&QWP(32*3,"esp"));         # h3
1678         &vpmuludq       ($T0,$T2,&$addr(0));            # h1*r0
1679         &vpaddq         ($D1,$D1,$T0);                  # d1 += h1*r0
1680         &vpmuludq       ($T2,$T2,&$addr(1));            # h1*r1
1681         &vpaddq         ($D2,$D2,$T2);                  # d2 += h1*r1
1682
1683         &vpmuludq       ($T0,$T1,&$addr(0));            # h3*r0
1684         &vpaddq         ($D3,$D3,$T0);                  # d3 += h3*r0
1685         &vpmuludq       ($T2,$T1,&$addr(1));            # h3*r1
1686         &vpaddq         ($D4,$D4,$T2);                  # d4 += h3*r1
1687         &vpmuludq       ($T0,$T1,&$addr(6));            # h3*s2
1688         &vpaddq         ($D0,$D0,$T0);                  # d0 += h3*s2
1689          &vmovdqa       ($T0,&QWP(32*4,"esp"));         # h4
1690         &vpmuludq       ($T2,$T1,&$addr(7));            # h3*s3
1691         &vpaddq         ($D1,$D1,$T2);                  # d1+= h3*s3
1692         &vpmuludq       ($T1,$T1,&$addr(8));            # h3*s4
1693         &vpaddq         ($D2,$D2,$T1);                  # d2 += h3*s4
1694
1695         &vpmuludq       ($T2,$T0,&$addr(8));            # h4*s4
1696         &vpaddq         ($D3,$D3,$T2);                  # d3 += h4*s4
1697         &vpmuludq       ($T1,$T0,&$addr(5));            # h4*s1
1698         &vpaddq         ($D0,$D0,$T1);                  # d0 += h4*s1
1699         &vpmuludq       ($T2,$T0,&$addr(0));            # h4*r0
1700         &vpaddq         ($D4,$D4,$T2);                  # d4 += h4*r0
1701          &vmovdqa       ($MASK,&QWP(64,"ebx"));
1702         &vpmuludq       ($T1,$T0,&$addr(6));            # h4*s2
1703         &vpaddq         ($D1,$D1,$T1);                  # d1 += h4*s2
1704         &vpmuludq       ($T0,$T0,&$addr(7));            # h4*s3
1705         &vpaddq         ($D2,$D2,$T0);                  # d2 += h4*s3
1706 }
1707         &vpmuladd       (sub {  my $i=shift; &QWP(32*$i-128,"edx");     });
1708
1709 sub vlazy_reduction {
1710         ################################################################
1711         # lazy reduction
1712
1713          &vpsrlq        ($T0,$D3,26);
1714          &vpand         ($D3,$D3,$MASK);
1715         &vpsrlq         ($T1,$D0,26);
1716         &vpand          ($D0,$D0,$MASK);
1717          &vpaddq        ($D4,$D4,$T0);                  # h3 -> h4
1718         &vpaddq         ($D1,$D1,$T1);                  # h0 -> h1
1719          &vpsrlq        ($T0,$D4,26);
1720          &vpand         ($D4,$D4,$MASK);
1721         &vpsrlq         ($T1,$D1,26);
1722         &vpand          ($D1,$D1,$MASK);
1723         &vpaddq         ($D2,$D2,$T1);                  # h1 -> h2
1724          &vpaddq        ($D0,$D0,$T0);
1725          &vpsllq        ($T0,$T0,2);
1726         &vpsrlq         ($T1,$D2,26);
1727         &vpand          ($D2,$D2,$MASK);
1728          &vpaddq        ($D0,$D0,$T0);                  # h4 -> h0
1729         &vpaddq         ($D3,$D3,$T1);                  # h2 -> h3
1730         &vpsrlq         ($T1,$D3,26);
1731          &vpsrlq        ($T0,$D0,26);
1732          &vpand         ($D0,$D0,$MASK);
1733         &vpand          ($D3,$D3,$MASK);
1734          &vpaddq        ($D1,$D1,$T0);                  # h0 -> h1
1735         &vpaddq         ($D4,$D4,$T1);                  # h3 -> h4
1736 }
1737         &vlazy_reduction();
1738
1739         &vmovdqu        (&X($T0),&QWP(16*0,"esi"));     # load input
1740         &vmovdqu        (&X($T1),&QWP(16*1,"esi"));
1741         &vinserti128    ($T0,$T0,&QWP(16*2,"esi"),1);
1742         &vinserti128    ($T1,$T1,&QWP(16*3,"esi"),1);
1743         &lea            ("esi",&DWP(16*4,"esi"));
1744         &sub            ("ecx",64);
1745         &jnz            (&label("loop"));
1746
1747 &set_label("tail");
1748         &vsplat_input   ();
1749         &and            ("ebx",-64);                    # restore pointer
1750
1751         &vpmuladd       (sub {  my $i=shift; &QWP(4+32*$i-128,"edx");   });
1752
1753         ################################################################
1754         # horizontal addition
1755
1756         &vpsrldq        ($T0,$D4,8);
1757         &vpsrldq        ($T1,$D3,8);
1758         &vpaddq         ($D4,$D4,$T0);
1759         &vpsrldq        ($T0,$D0,8);
1760         &vpaddq         ($D3,$D3,$T1);
1761         &vpsrldq        ($T1,$D1,8);
1762         &vpaddq         ($D0,$D0,$T0);
1763         &vpsrldq        ($T0,$D2,8);
1764         &vpaddq         ($D1,$D1,$T1);
1765         &vpermq         ($T1,$D4,2);                    # keep folding
1766         &vpaddq         ($D2,$D2,$T0);
1767         &vpermq         ($T0,$D3,2);
1768         &vpaddq         ($D4,$D4,$T1);
1769         &vpermq         ($T1,$D0,2);
1770         &vpaddq         ($D3,$D3,$T0);
1771         &vpermq         ($T0,$D1,2);
1772         &vpaddq         ($D0,$D0,$T1);
1773         &vpermq         ($T1,$D2,2);
1774         &vpaddq         ($D1,$D1,$T0);
1775         &vpaddq         ($D2,$D2,$T1);
1776
1777         &vlazy_reduction();
1778
1779         &cmp            ("ecx",0);
1780         &je             (&label("done"));
1781
1782         ################################################################
1783         # clear all but single word
1784
1785         &vpshufd        (&X($D0),&X($D0),0b11111100);
1786         &lea            ("edx",&DWP(32*5+128,"esp"));   # restore pointer
1787         &vpshufd        (&X($D1),&X($D1),0b11111100);
1788         &vpshufd        (&X($D2),&X($D2),0b11111100);
1789         &vpshufd        (&X($D3),&X($D3),0b11111100);
1790         &vpshufd        (&X($D4),&X($D4),0b11111100);
1791         &jmp            (&label("even"));
1792
1793 &set_label("done",16);
1794         &vmovd          (&DWP(-16*3+4*0,"edi"),&X($D0));# store hash value
1795         &vmovd          (&DWP(-16*3+4*1,"edi"),&X($D1));
1796         &vmovd          (&DWP(-16*3+4*2,"edi"),&X($D2));
1797         &vmovd          (&DWP(-16*3+4*3,"edi"),&X($D3));
1798         &vmovd          (&DWP(-16*3+4*4,"edi"),&X($D4));
1799         &vzeroupper     ();
1800         &mov    ("esp","ebp");
1801 &set_label("nodata");
1802 &function_end("_poly1305_blocks_avx2");
1803 }
1804 &set_label("const_sse2",64);
1805         &data_word(1<<24,0,     1<<24,0,        1<<24,0,        1<<24,0);
1806         &data_word(0,0,         0,0,            0,0,            0,0);
1807         &data_word(0x03ffffff,0,0x03ffffff,0,   0x03ffffff,0,   0x03ffffff,0);
1808         &data_word(0x0fffffff,0x0ffffffc,0x0ffffffc,0x0ffffffc);
1809 }
1810 &asciz  ("Poly1305 for x86, CRYPTOGAMS by <appro\@openssl.org>");
1811 &align  (4);
1812
1813 &asm_finish();
1814
1815 close STDOUT;