crypto/poly1305/asm/poly1305-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 # This module implements Poly1305 hash for x86.
  11 #
  12 # April 2015
  13 #
  14 # Numbers are cycles per processed byte with poly1305_blocks alone,
  15 # measured with rdtsc at fixed clock frequency.
  16 #
  17 #               IALU/gcc-3.4(*) SSE2(**)        AVX2
  18 # Pentium       15.7/+80%       -
  19 # PIII          6.21/+90%       -
  20 # P4            19.8/+40%       3.24
  21 # Core 2        4.85/+90%       1.80
  22 # Westmere      4.58/+100%      1.43
  23 # Sandy Bridge  3.90/+100%      1.36
  24 # Haswell       3.88/+70%       1.18            0.72
  25 # Silvermont    11.0/+40%       4.80
  26 # VIA Nano      6.71/+90%       2.47
  27 # Sledgehammer  3.51/+180%      4.27
  28 # Bulldozer     4.53/+140%      1.31
  29 #
  30 # (*)   gcc 4.8 for some reason generated worse code;
  31 # (**)  besides SSE2 there are floating-point and AVX options; FP
  32 #       is deemed unnecessary, because pre-SSE2 processor are too
  33 #       old to care about, while it's not the fastest option on
  34 #       SSE2-capable ones; AVX is omitted, because it doesn't give
  35 #       a lot of improvement, 5-10% depending on processor;
  36
  37 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  38 push(@INC,"${dir}","${dir}../../perlasm");
  39 require "x86asm.pl";
  40
  41 &asm_init($ARGV[0],"poly1305-x86.pl",$ARGV[$#ARGV] eq "386");
  42
  43 $sse2=$avx=0;
  44 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
  45
  46 if ($sse2) {
  47         &static_label("const_sse2");
  48         &static_label("enter_blocks");
  49         &static_label("enter_emit");
  50         &external_label("OPENSSL_ia32cap_P");
  51
  52         if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
  53                         =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
  54                 $avx = ($1>=2.19) + ($1>=2.22);
  55         }
  56
  57         if (!$avx && $ARGV[0] eq "win32n" &&
  58            `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
  59         $avx = ($1>=2.09) + ($1>=2.10);
  60         }
  61
  62         if (!$avx && `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/) {
  63                 $avx = ($2>=3.0) + ($2>3.0);
  64         }
  65 }
  66
  67 ########################################################################
  68 # Layout of opaque area is following.
  69 #
  70 #       unsigned __int32 h[5];          # current hash value base 2^32
  71 #       unsigned __int32 pad;           # is_base2_26 in vector context
  72 #       unsigned __int32 r[4];          # key value base 2^32
  73
  74 &align(64);
  75 &function_begin("poly1305_init");
  76         &mov    ("edi",&wparam(0));             # context
  77         &mov    ("esi",&wparam(1));             # key
  78         &mov    ("ebp",&wparam(2));             # function table
  79
  80         &xor    ("eax","eax");
  81         &mov    (&DWP(4*0,"edi"),"eax");        # zero hash value
  82         &mov    (&DWP(4*1,"edi"),"eax");
  83         &mov    (&DWP(4*2,"edi"),"eax");
  84         &mov    (&DWP(4*3,"edi"),"eax");
  85         &mov    (&DWP(4*4,"edi"),"eax");
  86         &mov    (&DWP(4*5,"edi"),"eax");        # is_base2_26
  87
  88         &cmp    ("esi",0);
  89         &je     (&label("nokey"));
  90
  91     if ($sse2) {
  92         &call   (&label("pic_point"));
  93     &set_label("pic_point");
  94         &blindpop("ebx");
  95
  96         &lea    ("eax",&DWP("poly1305_blocks-".&label("pic_point"),"ebx"));
  97         &lea    ("edx",&DWP("poly1305_emit-".&label("pic_point"),"ebx"));
  98
  99         &picmeup("edi","OPENSSL_ia32cap_P","ebx",&label("pic_point"));
 100         &mov    ("ecx",&DWP(0,"edi"));
 101         &and    ("ecx",1<<26|1<<24);
 102         &cmp    ("ecx",1<<26|1<<24);            # SSE2 and XMM?
 103         &jne    (&label("no_sse2"));
 104
 105         &lea    ("eax",&DWP("_poly1305_blocks_sse2-".&label("pic_point"),"ebx"));
 106         &lea    ("edx",&DWP("_poly1305_emit_sse2-".&label("pic_point"),"ebx"));
 107
 108       if ($avx>1) {
 109         &mov    ("ecx",&DWP(8,"edi"));
 110         &test   ("ecx",1<<5);                   # AVX2?
 111         &jz     (&label("no_sse2"));
 112
 113         &lea    ("eax",&DWP("_poly1305_blocks_avx2-".&label("pic_point"),"ebx"));
 114       }
 115     &set_label("no_sse2");
 116         &mov    ("edi",&wparam(0));             # reload context
 117         &mov    (&DWP(0,"ebp"),"eax");          # fill function table
 118         &mov    (&DWP(4,"ebp"),"edx");
 119     }
 120
 121         &mov    ("eax",&DWP(4*0,"esi"));        # load input key
 122         &mov    ("ebx",&DWP(4*1,"esi"));
 123         &mov    ("ecx",&DWP(4*2,"esi"));
 124         &mov    ("edx",&DWP(4*3,"esi"));
 125         &and    ("eax",0x0fffffff);
 126         &and    ("ebx",0x0ffffffc);
 127         &and    ("ecx",0x0ffffffc);
 128         &and    ("edx",0x0ffffffc);
 129         &mov    (&DWP(4*6,"edi"),"eax");
 130         &mov    (&DWP(4*7,"edi"),"ebx");
 131         &mov    (&DWP(4*8,"edi"),"ecx");
 132         &mov    (&DWP(4*9,"edi"),"edx");
 133
 134         &mov    ("eax",$sse2);
 135 &set_label("nokey");
 136 &function_end("poly1305_init");
 137
 138 ($h0,$h1,$h2,$h3,$h4,
 139  $d0,$d1,$d2,$d3,
 140  $r0,$r1,$r2,$r3,
 141      $s1,$s2,$s3)=map(4*$_,(0..15));
 142
 143 &function_begin("poly1305_blocks");
 144         &mov    ("edi",&wparam(0));             # ctx
 145         &mov    ("esi",&wparam(1));             # inp
 146         &mov    ("ecx",&wparam(2));             # len
 147 &set_label("enter_blocks");
 148         &and    ("ecx",-15);
 149         &jz     (&label("nodata"));
 150
 151         &stack_push(16);
 152         &mov    ("eax",&DWP(4*6,"edi"));        # r0
 153         &mov    ("ebx",&DWP(4*7,"edi"));        # r1
 154          &lea   ("ebp",&DWP(0,"esi","ecx"));    # end of input
 155         &mov    ("ecx",&DWP(4*8,"edi"));        # r2
 156         &mov    ("edx",&DWP(4*9,"edi"));        # r3
 157
 158         &mov    (&wparam(2),"ebp");
 159         &mov    ("ebp","esi");
 160
 161         &mov    (&DWP($r0,"esp"),"eax");        # r0
 162         &mov    ("eax","ebx");
 163         &shr    ("eax",2);
 164         &mov    (&DWP($r1,"esp"),"ebx");        # r1
 165         &add    ("eax","ebx");                  # s1
 166         &mov    ("ebx","ecx");
 167         &shr    ("ebx",2);
 168         &mov    (&DWP($r2,"esp"),"ecx");        # r2
 169         &add    ("ebx","ecx");                  # s2
 170         &mov    ("ecx","edx");
 171         &shr    ("ecx",2);
 172         &mov    (&DWP($r3,"esp"),"edx");        # r3
 173         &add    ("ecx","edx");                  # s3
 174         &mov    (&DWP($s1,"esp"),"eax");        # s1
 175         &mov    (&DWP($s2,"esp"),"ebx");        # s2
 176         &mov    (&DWP($s3,"esp"),"ecx");        # s3
 177
 178         &mov    ("eax",&DWP(4*0,"edi"));        # load hash value
 179         &mov    ("ebx",&DWP(4*1,"edi"));
 180         &mov    ("ecx",&DWP(4*2,"edi"));
 181         &mov    ("esi",&DWP(4*3,"edi"));
 182         &mov    ("edi",&DWP(4*4,"edi"));
 183         &jmp    (&label("loop"));
 184
 185 &set_label("loop",32);
 186         &add    ("eax",&DWP(4*0,"ebp"));        # accumulate input
 187         &adc    ("ebx",&DWP(4*1,"ebp"));
 188         &adc    ("ecx",&DWP(4*2,"ebp"));
 189         &adc    ("esi",&DWP(4*3,"ebp"));
 190         &lea    ("ebp",&DWP(4*4,"ebp"));
 191         &adc    ("edi",&wparam(3));             # padbit
 192
 193         &mov    (&DWP($h0,"esp"),"eax");        # put aside hash[+inp]
 194         &mov    (&DWP($h3,"esp"),"esi");
 195
 196         &mul    (&DWP($r0,"esp"));              # h0*r0
 197          &mov   (&DWP($h4,"esp"),"edi");
 198         &mov    ("edi","eax");
 199         &mov    ("eax","ebx");                  # h1
 200         &mov    ("esi","edx");
 201         &mul    (&DWP($s3,"esp"));              # h1*s3
 202         &add    ("edi","eax");
 203         &mov    ("eax","ecx");                  # h2
 204         &adc    ("esi","edx");
 205         &mul    (&DWP($s2,"esp"));              # h2*s2
 206         &add    ("edi","eax");
 207         &mov    ("eax",&DWP($h3,"esp"));
 208         &adc    ("esi","edx");
 209         &mul    (&DWP($s1,"esp"));              # h3*s1
 210         &add    ("edi","eax");
 211          &mov   ("eax",&DWP($h0,"esp"));
 212         &adc    ("esi","edx");
 213
 214         &mul    (&DWP($r1,"esp"));              # h0*r1
 215          &mov   (&DWP($d0,"esp"),"edi");
 216         &xor    ("edi","edi");
 217         &add    ("esi","eax");
 218         &mov    ("eax","ebx");                  # h1
 219         &adc    ("edi","edx");
 220         &mul    (&DWP($r0,"esp"));              # h1*r0
 221         &add    ("esi","eax");
 222         &mov    ("eax","ecx");                  # h2
 223         &adc    ("edi","edx");
 224         &mul    (&DWP($s3,"esp"));              # h2*s3
 225         &add    ("esi","eax");
 226         &mov    ("eax",&DWP($h3,"esp"));
 227         &adc    ("edi","edx");
 228         &mul    (&DWP($s2,"esp"));              # h3*s2
 229         &add    ("esi","eax");
 230         &mov    ("eax",&DWP($h4,"esp"));
 231         &adc    ("edi","edx");
 232         &imul   ("eax",&DWP($s1,"esp"));        # h4*s1
 233         &add    ("esi","eax");
 234          &mov   ("eax",&DWP($h0,"esp"));
 235         &adc    ("edi",0);
 236
 237         &mul    (&DWP($r2,"esp"));              # h0*r2
 238          &mov   (&DWP($d1,"esp"),"esi");
 239         &xor    ("esi","esi");
 240         &add    ("edi","eax");
 241         &mov    ("eax","ebx");                  # h1
 242         &adc    ("esi","edx");
 243         &mul    (&DWP($r1,"esp"));              # h1*r1
 244         &add    ("edi","eax");
 245         &mov    ("eax","ecx");                  # h2
 246         &adc    ("esi","edx");
 247         &mul    (&DWP($r0,"esp"));              # h2*r0
 248         &add    ("edi","eax");
 249         &mov    ("eax",&DWP($h3,"esp"));
 250         &adc    ("esi","edx");
 251         &mul    (&DWP($s3,"esp"));              # h3*s3
 252         &add    ("edi","eax");
 253         &mov    ("eax",&DWP($h4,"esp"));
 254         &adc    ("esi","edx");
 255         &imul   ("eax",&DWP($s2,"esp"));        # h4*s2
 256         &add    ("edi","eax");
 257          &mov   ("eax",&DWP($h0,"esp"));
 258         &adc    ("esi",0);
 259
 260         &mul    (&DWP($r3,"esp"));              # h0*r3
 261          &mov   (&DWP($d2,"esp"),"edi");
 262         &xor    ("edi","edi");
 263         &add    ("esi","eax");
 264         &mov    ("eax","ebx");                  # h1
 265         &adc    ("edi","edx");
 266         &mul    (&DWP($r2,"esp"));              # h1*r2
 267         &add    ("esi","eax");
 268         &mov    ("eax","ecx");                  # h2
 269         &adc    ("edi","edx");
 270         &mul    (&DWP($r1,"esp"));              # h2*r1
 271         &add    ("esi","eax");
 272         &mov    ("eax",&DWP($h3,"esp"));
 273         &adc    ("edi","edx");
 274         &mul    (&DWP($r0,"esp"));              # h3*r0
 275         &add    ("esi","eax");
 276          &mov   ("ecx",&DWP($h4,"esp"));
 277         &adc    ("edi","edx");
 278
 279         &mov    ("edx","ecx");
 280         &imul   ("ecx",&DWP($s3,"esp"));        # h4*s3
 281         &add    ("esi","ecx");
 282          &mov   ("eax",&DWP($d0,"esp"));
 283         &adc    ("edi",0);
 284
 285         &imul   ("edx",&DWP($r0,"esp"));        # h4*r0
 286         &add    ("edx","edi");
 287
 288         &mov    ("ebx",&DWP($d1,"esp"));
 289         &mov    ("ecx",&DWP($d2,"esp"));
 290
 291         &mov    ("edi","edx");                  # last reduction step
 292         &shr    ("edx",2);
 293         &and    ("edi",3);
 294         &lea    ("edx",&DWP(0,"edx","edx",4));  # *5
 295         &add    ("eax","edx");
 296         &adc    ("ebx",0);
 297         &adc    ("ecx",0);
 298         &adc    ("esi",0);
 299
 300         &cmp    ("ebp",&wparam(2));             # done yet?
 301         &jne    (&label("loop"));
 302
 303         &mov    ("edx",&wparam(0));             # ctx
 304         &stack_pop(16);
 305         &mov    (&DWP(4*0,"edx"),"eax");        # store hash value
 306         &mov    (&DWP(4*1,"edx"),"ebx");
 307         &mov    (&DWP(4*2,"edx"),"ecx");
 308         &mov    (&DWP(4*3,"edx"),"esi");
 309         &mov    (&DWP(4*4,"edx"),"edi");
 310 &set_label("nodata");
 311 &function_end("poly1305_blocks");
 312
 313 &function_begin("poly1305_emit");
 314         &mov    ("ebp",&wparam(0));             # context
 315 &set_label("enter_emit");
 316         &mov    ("edi",&wparam(1));             # output
 317         &mov    ("eax",&DWP(4*0,"ebp"));        # load hash value
 318         &mov    ("ebx",&DWP(4*1,"ebp"));
 319         &mov    ("ecx",&DWP(4*2,"ebp"));
 320         &mov    ("edx",&DWP(4*3,"ebp"));
 321         &mov    ("esi",&DWP(4*4,"ebp"));
 322
 323         &add    ("eax",5);                      # compare to modulus
 324         &adc    ("ebx",0);
 325         &adc    ("ecx",0);
 326         &adc    ("edx",0);
 327         &adc    ("esi",0);
 328         &shr    ("esi",2);                      # did it carry/borrow?
 329         &neg    ("esi");                        # do we choose hash-modulus?
 330
 331         &and    ("eax","esi");
 332         &and    ("ebx","esi");
 333         &and    ("ecx","esi");
 334         &and    ("edx","esi");
 335         &mov    (&DWP(4*0,"edi"),"eax");
 336         &mov    (&DWP(4*1,"edi"),"ebx");
 337         &mov    (&DWP(4*2,"edi"),"ecx");
 338         &mov    (&DWP(4*3,"edi"),"edx");
 339
 340         &not    ("esi");                        # or original hash value?
 341         &mov    ("eax",&DWP(4*0,"ebp"));
 342         &mov    ("ebx",&DWP(4*1,"ebp"));
 343         &mov    ("ecx",&DWP(4*2,"ebp"));
 344         &mov    ("edx",&DWP(4*3,"ebp"));
 345         &mov    ("ebp",&wparam(2));
 346         &and    ("eax","esi");
 347         &and    ("ebx","esi");
 348         &and    ("ecx","esi");
 349         &and    ("edx","esi");
 350         &or     ("eax",&DWP(4*0,"edi"));
 351         &or     ("ebx",&DWP(4*1,"edi"));
 352         &or     ("ecx",&DWP(4*2,"edi"));
 353         &or     ("edx",&DWP(4*3,"edi"));
 354
 355         &add    ("eax",&DWP(4*0,"ebp"));        # accumulate key
 356         &adc    ("ebx",&DWP(4*1,"ebp"));
 357         &adc    ("ecx",&DWP(4*2,"ebp"));
 358         &adc    ("edx",&DWP(4*3,"ebp"));
 359
 360         &mov    (&DWP(4*0,"edi"),"eax");
 361         &mov    (&DWP(4*1,"edi"),"ebx");
 362         &mov    (&DWP(4*2,"edi"),"ecx");
 363         &mov    (&DWP(4*3,"edi"),"edx");
 364 &function_end("poly1305_emit");
 365
 366 if ($sse2) {
 367 ########################################################################
 368 # Layout of opaque area is following.
 369 #
 370 #       unsigned __int32 h[5];          # current hash value base 2^26
 371 #       unsigned __int32 is_base2_26;
 372 #       unsigned __int32 r[4];          # key value base 2^32
 373 #       unsigned __int32 pad[2];
 374 #       struct { unsigned __int32 r^4, r^3, r^2, r^1; } r[9];
 375 #
 376 # where r^n are base 2^26 digits of degrees of multiplier key. There are
 377 # 5 digits, but last four are interleaved with multiples of 5, totalling
 378 # in 9 elements: r0, r1, 5*r1, r2, 5*r2, r3, 5*r3, r4, 5*r4.
 379
 380 my ($D0,$D1,$D2,$D3,$D4,$T0,$T1,$T2)=map("xmm$_",(0..7));
 381 my $MASK=$T2;   # borrow and keep in mind
 382
 383 &align  (32);
 384 &function_begin_B("_poly1305_init_sse2");
 385         &movdqu         ($D4,&QWP(4*6,"edi"));          # key base 2^32
 386         &lea            ("edi",&DWP(16*3,"edi"));       # size optimization
 387         &mov            ("ebp","esp");
 388         &sub            ("esp",16*(9+5));
 389         &and            ("esp",-16);
 390
 391         #&pand          ($D4,&QWP(96,"ebx"));           # magic mask
 392         &movq           ($MASK,&QWP(64,"ebx"));
 393
 394         &movdqa         ($D0,$D4);
 395         &movdqa         ($D1,$D4);
 396         &movdqa         ($D2,$D4);
 397
 398         &pand           ($D0,$MASK);                    # -> base 2^26
 399         &psrlq          ($D1,26);
 400         &psrldq         ($D2,6);
 401         &pand           ($D1,$MASK);
 402         &movdqa         ($D3,$D2);
 403         &psrlq          ($D2,4)
 404         &psrlq          ($D3,30);
 405         &pand           ($D2,$MASK);
 406         &pand           ($D3,$MASK);
 407         &psrldq         ($D4,13);
 408
 409         &lea            ("edx",&DWP(16*9,"esp"));       # size optimization
 410         &mov            ("ecx",2);
 411 &set_label("square");
 412         &movdqa         (&QWP(16*0,"esp"),$D0);
 413         &movdqa         (&QWP(16*1,"esp"),$D1);
 414         &movdqa         (&QWP(16*2,"esp"),$D2);
 415         &movdqa         (&QWP(16*3,"esp"),$D3);
 416         &movdqa         (&QWP(16*4,"esp"),$D4);
 417
 418         &movdqa         ($T1,$D1);
 419         &movdqa         ($T0,$D2);
 420         &pslld          ($T1,2);
 421         &pslld          ($T0,2);
 422         &paddd          ($T1,$D1);                      # *5
 423         &paddd          ($T0,$D2);                      # *5
 424         &movdqa         (&QWP(16*5,"esp"),$T1);
 425         &movdqa         (&QWP(16*6,"esp"),$T0);
 426         &movdqa         ($T1,$D3);
 427         &movdqa         ($T0,$D4);
 428         &pslld          ($T1,2);
 429         &pslld          ($T0,2);
 430         &paddd          ($T1,$D3);                      # *5
 431         &paddd          ($T0,$D4);                      # *5
 432         &movdqa         (&QWP(16*7,"esp"),$T1);
 433         &movdqa         (&QWP(16*8,"esp"),$T0);
 434
 435         &pshufd         ($T1,$D0,0b01000100);
 436         &movdqa         ($T0,$D1);
 437         &pshufd         ($D1,$D1,0b01000100);
 438         &pshufd         ($D2,$D2,0b01000100);
 439         &pshufd         ($D3,$D3,0b01000100);
 440         &pshufd         ($D4,$D4,0b01000100);
 441         &movdqa         (&QWP(16*0,"edx"),$T1);
 442         &movdqa         (&QWP(16*1,"edx"),$D1);
 443         &movdqa         (&QWP(16*2,"edx"),$D2);
 444         &movdqa         (&QWP(16*3,"edx"),$D3);
 445         &movdqa         (&QWP(16*4,"edx"),$D4);
 446
 447         ################################################################
 448         # d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
 449         # d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
 450         # d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
 451         # d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
 452         # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
 453
 454         &pmuludq        ($D4,$D0);                      # h4*r0
 455         &pmuludq        ($D3,$D0);                      # h3*r0
 456         &pmuludq        ($D2,$D0);                      # h2*r0
 457         &pmuludq        ($D1,$D0);                      # h1*r0
 458         &pmuludq        ($D0,$T1);                      # h0*r0
 459
 460 sub pmuladd {
 461 my $load = shift;
 462 my $base = shift; $base = "esp" if (!defined($base));
 463
 464         ################################################################
 465         # As for choice to "rotate" $T0-$T2 in order to move paddq
 466         # past next multiplication. While it makes code harder to read
 467         # and doesn't have significant effect on most processors, it
 468         # makes a lot of difference on Atom, up to 30% improvement.
 469
 470         &movdqa         ($T1,$T0);
 471         &pmuludq        ($T0,&QWP(16*3,$base));         # r1*h3
 472         &movdqa         ($T2,$T1);
 473         &pmuludq        ($T1,&QWP(16*2,$base));         # r1*h2
 474         &paddq          ($D4,$T0);
 475         &movdqa         ($T0,$T2);
 476         &pmuludq        ($T2,&QWP(16*1,$base));         # r1*h1
 477         &paddq          ($D3,$T1);
 478         &$load          ($T1,5);                        # s1
 479         &pmuludq        ($T0,&QWP(16*0,$base));         # r1*h0
 480         &paddq          ($D2,$T2);
 481         &pmuludq        ($T1,&QWP(16*4,$base));         # s1*h4
 482          &$load         ($T2,2);                        # r2^n
 483         &paddq          ($D1,$T0);
 484
 485         &movdqa         ($T0,$T2);
 486         &pmuludq        ($T2,&QWP(16*2,$base));         # r2*h2
 487          &paddq         ($D0,$T1);
 488         &movdqa         ($T1,$T0);
 489         &pmuludq        ($T0,&QWP(16*1,$base));         # r2*h1
 490         &paddq          ($D4,$T2);
 491         &$load          ($T2,6);                        # s2^n
 492         &pmuludq        ($T1,&QWP(16*0,$base));         # r2*h0
 493         &paddq          ($D3,$T0);
 494         &movdqa         ($T0,$T2);
 495         &pmuludq        ($T2,&QWP(16*4,$base));         # s2*h4
 496         &paddq          ($D2,$T1);
 497         &pmuludq        ($T0,&QWP(16*3,$base));         # s2*h3
 498          &$load         ($T1,3);                        # r3^n
 499         &paddq          ($D1,$T2);
 500
 501         &movdqa         ($T2,$T1);
 502         &pmuludq        ($T1,&QWP(16*1,$base));         # r3*h1
 503          &paddq         ($D0,$T0);
 504         &$load          ($T0,7);                        # s3^n
 505         &pmuludq        ($T2,&QWP(16*0,$base));         # r3*h0
 506         &paddq          ($D4,$T1);
 507         &movdqa         ($T1,$T0);
 508         &pmuludq        ($T0,&QWP(16*4,$base));         # s3*h4
 509         &paddq          ($D3,$T2);
 510         &movdqa         ($T2,$T1);
 511         &pmuludq        ($T1,&QWP(16*3,$base));         # s3*h3
 512         &paddq          ($D2,$T0);
 513         &pmuludq        ($T2,&QWP(16*2,$base));         # s3*h2
 514          &$load         ($T0,4);                        # r4^n
 515         &paddq          ($D1,$T1);
 516
 517         &$load          ($T1,8);                        # s4^n
 518         &pmuludq        ($T0,&QWP(16*0,$base));         # r4*h0
 519          &paddq         ($D0,$T2);
 520         &movdqa         ($T2,$T1);
 521         &pmuludq        ($T1,&QWP(16*4,$base));         # s4*h4
 522         &paddq          ($D4,$T0);
 523         &movdqa         ($T0,$T2);
 524         &pmuludq        ($T2,&QWP(16*1,$base));         # s4*h1
 525         &paddq          ($D3,$T1);
 526         &movdqa         ($T1,$T0);
 527         &pmuludq        ($T0,&QWP(16*2,$base));         # s4*h2
 528         &paddq          ($D0,$T2);
 529         &pmuludq        ($T1,&QWP(16*3,$base));         # s4*h3
 530          &movdqa        ($MASK,&QWP(64,"ebx"));
 531         &paddq          ($D1,$T0);
 532         &paddq          ($D2,$T1);
 533 }
 534         &pmuladd        (sub {  my ($reg,$i)=@_;
 535                                 &movdqa ($reg,&QWP(16*$i,"esp"));
 536                              },"edx");
 537
 538 sub lazy_reduction {
 539         ################################################################
 540         # lazy reduction as discussed in "NEON crypto" by D.J. Bernstein
 541         # and P. Schwabe
 542
 543          &movdqa        ($T0,$D3);
 544          &pand          ($D3,$MASK);
 545          &psrlq         ($T0,26);
 546          &paddq         ($T0,$D4);                      # h3 -> h4
 547         &movdqa         ($T1,$D0);
 548         &pand           ($D0,$MASK);
 549         &psrlq          ($T1,26);
 550          &movdqa        ($D4,$T0);
 551         &paddq          ($T1,$D1);                      # h0 -> h1
 552          &psrlq         ($T0,26);
 553          &pand          ($D4,$MASK);
 554         &movdqa         ($D1,$T1);
 555         &psrlq          ($T1,26);
 556          &paddd         ($D0,$T0);                      # favour paddd when
 557                                                         # possible, because
 558                                                         # paddq is "broken"
 559                                                         # on Atom
 560         &pand           ($D1,$MASK);
 561         &paddq          ($T1,$D2);                      # h1 -> h2
 562          &psllq         ($T0,2);
 563         &movdqa         ($D2,$T1);
 564         &psrlq          ($T1,26);
 565          &paddd         ($T0,$D0);                      # h4 -> h0
 566         &pand           ($D2,$MASK);
 567         &paddd          ($T1,$D3);                      # h2 -> h3
 568          &movdqa        ($D0,$T0);
 569          &psrlq         ($T0,26);
 570         &movdqa         ($D3,$T1);
 571         &psrlq          ($T1,26);
 572          &pand          ($D0,$MASK);
 573          &paddd         ($D1,$T0);                      # h0 -> h1
 574         &pand           ($D3,$MASK);
 575         &paddd          ($D4,$T1);                      # h3 -> h4
 576 }
 577         &lazy_reduction ();
 578
 579         &dec            ("ecx");
 580         &jz             (&label("square_break"));
 581
 582         &punpcklqdq     ($D0,&QWP(16*0,"esp"));         # 0:r^1:0:r^2
 583         &punpcklqdq     ($D1,&QWP(16*1,"esp"));
 584         &punpcklqdq     ($D2,&QWP(16*2,"esp"));
 585         &punpcklqdq     ($D3,&QWP(16*3,"esp"));
 586         &punpcklqdq     ($D4,&QWP(16*4,"esp"));
 587         &jmp            (&label("square"));
 588
 589 &set_label("square_break");
 590         &psllq          ($D0,32);                       # -> r^3:0:r^4:0
 591         &psllq          ($D1,32);
 592         &psllq          ($D2,32);
 593         &psllq          ($D3,32);
 594         &psllq          ($D4,32);
 595         &por            ($D0,&QWP(16*0,"esp"));         # r^3:r^1:r^4:r^2
 596         &por            ($D1,&QWP(16*1,"esp"));
 597         &por            ($D2,&QWP(16*2,"esp"));
 598         &por            ($D3,&QWP(16*3,"esp"));
 599         &por            ($D4,&QWP(16*4,"esp"));
 600
 601         &pshufd         ($D0,$D0,0b10001101);           # -> r^1:r^2:r^3:r^4
 602         &pshufd         ($D1,$D1,0b10001101);
 603         &pshufd         ($D2,$D2,0b10001101);
 604         &pshufd         ($D3,$D3,0b10001101);
 605         &pshufd         ($D4,$D4,0b10001101);
 606
 607         &movdqu         (&QWP(16*0,"edi"),$D0);         # save the table
 608         &movdqu         (&QWP(16*1,"edi"),$D1);
 609         &movdqu         (&QWP(16*2,"edi"),$D2);
 610         &movdqu         (&QWP(16*3,"edi"),$D3);
 611         &movdqu         (&QWP(16*4,"edi"),$D4);
 612
 613         &movdqa         ($T1,$D1);
 614         &movdqa         ($T0,$D2);
 615         &pslld          ($T1,2);
 616         &pslld          ($T0,2);
 617         &paddd          ($T1,$D1);                      # *5
 618         &paddd          ($T0,$D2);                      # *5
 619         &movdqu         (&QWP(16*5,"edi"),$T1);
 620         &movdqu         (&QWP(16*6,"edi"),$T0);
 621         &movdqa         ($T1,$D3);
 622         &movdqa         ($T0,$D4);
 623         &pslld          ($T1,2);
 624         &pslld          ($T0,2);
 625         &paddd          ($T1,$D3);                      # *5
 626         &paddd          ($T0,$D4);                      # *5
 627         &movdqu         (&QWP(16*7,"edi"),$T1);
 628         &movdqu         (&QWP(16*8,"edi"),$T0);
 629
 630         &mov            ("esp","ebp");
 631         &lea            ("edi",&DWP(-16*3,"edi"));      # size de-optimization
 632         &ret            ();
 633 &function_end_B("_poly1305_init_sse2");
 634
 635 &align  (32);
 636 &function_begin("_poly1305_blocks_sse2");
 637         &mov    ("edi",&wparam(0));                     # ctx
 638         &mov    ("esi",&wparam(1));                     # inp
 639         &mov    ("ecx",&wparam(2));                     # len
 640
 641         &mov    ("eax",&DWP(4*5,"edi"));                # is_base2_26
 642         &and    ("ecx",-16);
 643         &jz     (&label("nodata"));
 644         &cmp    ("ecx",64);
 645         &jae    (&label("enter_sse2"));
 646         &test   ("eax","eax");                          # is_base2_26?
 647         &jz     (&label("enter_blocks"));
 648
 649 &set_label("enter_sse2",16);
 650         &call   (&label("pic_point"));
 651 &set_label("pic_point");
 652         &blindpop("ebx");
 653         &lea    ("ebx",&DWP(&label("const_sse2")."-".&label("pic_point"),"ebx"));
 654
 655         &test   ("eax","eax");                          # is_base2_26?
 656         &jnz    (&label("base2_26"));
 657
 658         &call   ("_poly1305_init_sse2");
 659
 660         ################################################# base 2^32 -> base 2^26
 661         &mov    ("eax",&DWP(0,"edi"));
 662         &mov    ("ecx",&DWP(3,"edi"));
 663         &mov    ("edx",&DWP(6,"edi"));
 664         &mov    ("esi",&DWP(9,"edi"));
 665         &mov    ("ebp",&DWP(13,"edi"));
 666         &mov    (&DWP(4*5,"edi"),1);                    # is_base2_26
 667
 668         &shr    ("ecx",2);
 669         &and    ("eax",0x3ffffff);
 670         &shr    ("edx",4);
 671         &and    ("ecx",0x3ffffff);
 672         &shr    ("esi",6);
 673         &and    ("edx",0x3ffffff);
 674
 675         &movd   ($D0,"eax");
 676         &movd   ($D1,"ecx");
 677         &movd   ($D2,"edx");
 678         &movd   ($D3,"esi");
 679         &movd   ($D4,"ebp");
 680
 681         &mov    ("esi",&wparam(1));                     # [reload] inp
 682         &mov    ("ecx",&wparam(2));                     # [reload] len
 683         &jmp    (&label("base2_32"));
 684
 685 &set_label("base2_26",16);
 686         &movd   ($D0,&DWP(4*0,"edi"));                  # load hash value
 687         &movd   ($D1,&DWP(4*1,"edi"));
 688         &movd   ($D2,&DWP(4*2,"edi"));
 689         &movd   ($D3,&DWP(4*3,"edi"));
 690         &movd   ($D4,&DWP(4*4,"edi"));
 691         &movdqa ($MASK,&QWP(64,"ebx"));
 692
 693 &set_label("base2_32");
 694         &mov    ("eax",&wparam(3));                     # padbit
 695         &mov    ("ebp","esp");
 696
 697         &sub    ("esp",16*(5+5+5+9+9));
 698         &and    ("esp",-16);
 699
 700         &lea    ("edi",&DWP(16*3,"edi"));               # size optimization
 701         &shl    ("eax",24);                             # padbit
 702
 703         &test   ("ecx",31);
 704         &jz     (&label("even"));
 705
 706         ################################################################
 707         # process single block, with SSE2, because it's still faster
 708         # even though half of result is discarded
 709
 710         &movdqu         ($T1,&QWP(0,"esi"));            # input
 711         &lea            ("esi",&DWP(16,"esi"));
 712
 713         &movdqa         ($T0,$T1);                      # -> base 2^26 ...
 714         &pand           ($T1,$MASK);
 715         &paddd          ($D0,$T1);                      # ... and accumuate
 716
 717         &movdqa         ($T1,$T0);
 718         &psrlq          ($T0,26);
 719         &psrldq         ($T1,6);
 720         &pand           ($T0,$MASK);
 721         &paddd          ($D1,$T0);
 722
 723         &movdqa         ($T0,$T1);
 724         &psrlq          ($T1,4);
 725         &pand           ($T1,$MASK);
 726         &paddd          ($D2,$T1);
 727
 728         &movdqa         ($T1,$T0);
 729         &psrlq          ($T0,30);
 730         &pand           ($T0,$MASK);
 731         &psrldq         ($T1,7);
 732         &paddd          ($D3,$T0);
 733
 734         &movd           ($T0,"eax");                    # padbit
 735         &paddd          ($D4,$T1);
 736          &movd          ($T1,&DWP(16*0+12,"edi"));      # r0
 737         &paddd          ($D4,$T0);
 738
 739         &movdqa         (&QWP(16*0,"esp"),$D0);
 740         &movdqa         (&QWP(16*1,"esp"),$D1);
 741         &movdqa         (&QWP(16*2,"esp"),$D2);
 742         &movdqa         (&QWP(16*3,"esp"),$D3);
 743         &movdqa         (&QWP(16*4,"esp"),$D4);
 744
 745         ################################################################
 746         # d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
 747         # d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
 748         # d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
 749         # d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
 750         # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
 751
 752         &pmuludq        ($D0,$T1);                      # h4*r0
 753         &pmuludq        ($D1,$T1);                      # h3*r0
 754         &pmuludq        ($D2,$T1);                      # h2*r0
 755          &movd          ($T0,&DWP(16*1+12,"edi"));      # r1
 756         &pmuludq        ($D3,$T1);                      # h1*r0
 757         &pmuludq        ($D4,$T1);                      # h0*r0
 758
 759         &pmuladd        (sub {  my ($reg,$i)=@_;
 760                                 &movd ($reg,&DWP(16*$i+12,"edi"));
 761                              });
 762
 763         &lazy_reduction ();
 764
 765         &sub            ("ecx",16);
 766         &jz             (&label("done"));
 767
 768 &set_label("even");
 769         &lea            ("edx",&DWP(16*(5+5+5+9),"esp"));# size optimization
 770         &lea            ("eax",&DWP(-16*2,"esi"));
 771         &sub            ("ecx",64);
 772
 773         ################################################################
 774         # expand and copy pre-calculated table to stack
 775
 776         &movdqu         ($T0,&QWP(16*0,"edi"));         # r^1:r^2:r^3:r^4
 777         &pshufd         ($T1,$T0,0b01000100);           # duplicate r^3:r^4
 778         &cmovb          ("esi","eax");
 779         &pshufd         ($T0,$T0,0b11101110);           # duplicate r^1:r^2
 780         &movdqa         (&QWP(16*0,"edx"),$T1);
 781         &lea            ("eax",&DWP(16*10,"esp"));
 782         &movdqu         ($T1,&QWP(16*1,"edi"));
 783         &movdqa         (&QWP(16*(0-9),"edx"),$T0);
 784         &pshufd         ($T0,$T1,0b01000100);
 785         &pshufd         ($T1,$T1,0b11101110);
 786         &movdqa         (&QWP(16*1,"edx"),$T0);
 787         &movdqu         ($T0,&QWP(16*2,"edi"));
 788         &movdqa         (&QWP(16*(1-9),"edx"),$T1);
 789         &pshufd         ($T1,$T0,0b01000100);
 790         &pshufd         ($T0,$T0,0b11101110);
 791         &movdqa         (&QWP(16*2,"edx"),$T1);
 792         &movdqu         ($T1,&QWP(16*3,"edi"));
 793         &movdqa         (&QWP(16*(2-9),"edx"),$T0);
 794         &pshufd         ($T0,$T1,0b01000100);
 795         &pshufd         ($T1,$T1,0b11101110);
 796         &movdqa         (&QWP(16*3,"edx"),$T0);
 797         &movdqu         ($T0,&QWP(16*4,"edi"));
 798         &movdqa         (&QWP(16*(3-9),"edx"),$T1);
 799         &pshufd         ($T1,$T0,0b01000100);
 800         &pshufd         ($T0,$T0,0b11101110);
 801         &movdqa         (&QWP(16*4,"edx"),$T1);
 802         &movdqu         ($T1,&QWP(16*5,"edi"));
 803         &movdqa         (&QWP(16*(4-9),"edx"),$T0);
 804         &pshufd         ($T0,$T1,0b01000100);
 805         &pshufd         ($T1,$T1,0b11101110);
 806         &movdqa         (&QWP(16*5,"edx"),$T0);
 807         &movdqu         ($T0,&QWP(16*6,"edi"));
 808         &movdqa         (&QWP(16*(5-9),"edx"),$T1);
 809         &pshufd         ($T1,$T0,0b01000100);
 810         &pshufd         ($T0,$T0,0b11101110);
 811         &movdqa         (&QWP(16*6,"edx"),$T1);
 812         &movdqu         ($T1,&QWP(16*7,"edi"));
 813         &movdqa         (&QWP(16*(6-9),"edx"),$T0);
 814         &pshufd         ($T0,$T1,0b01000100);
 815         &pshufd         ($T1,$T1,0b11101110);
 816         &movdqa         (&QWP(16*7,"edx"),$T0);
 817         &movdqu         ($T0,&QWP(16*8,"edi"));
 818         &movdqa         (&QWP(16*(7-9),"edx"),$T1);
 819         &pshufd         ($T1,$T0,0b01000100);
 820         &pshufd         ($T0,$T0,0b11101110);
 821         &movdqa         (&QWP(16*8,"edx"),$T1);
 822         &movdqa         (&QWP(16*(8-9),"edx"),$T0);
 823
 824 sub load_input {
 825 my ($inpbase,$offbase)=@_;
 826
 827         &movdqu         ($T0,&QWP($inpbase+0,"esi"));   # load input
 828         &movdqu         ($T1,&QWP($inpbase+16,"esi"));
 829         &lea            ("esi",&DWP(16*2,"esi"));
 830
 831         &movdqa         (&QWP($offbase+16*2,"esp"),$D2);
 832         &movdqa         (&QWP($offbase+16*3,"esp"),$D3);
 833         &movdqa         (&QWP($offbase+16*4,"esp"),$D4);
 834
 835         &movdqa         ($D2,$T0);                      # splat input
 836         &movdqa         ($D3,$T1);
 837         &psrldq         ($D2,6);
 838         &psrldq         ($D3,6);
 839         &movdqa         ($D4,$T0);
 840         &punpcklqdq     ($D2,$D3);                      # 2:3
 841         &punpckhqdq     ($D4,$T1);                      # 4
 842         &punpcklqdq     ($T0,$T1);                      # 0:1
 843
 844         &movdqa         ($D3,$D2);
 845         &psrlq          ($D2,4);
 846         &psrlq          ($D3,30);
 847         &movdqa         ($T1,$T0);
 848         &psrlq          ($D4,40);                       # 4
 849         &psrlq          ($T1,26);
 850         &pand           ($T0,$MASK);                    # 0
 851         &pand           ($T1,$MASK);                    # 1
 852         &pand           ($D2,$MASK);                    # 2
 853         &pand           ($D3,$MASK);                    # 3
 854         &por            ($D4,&QWP(0,"ebx"));            # padbit, yes, always
 855
 856         &movdqa         (&QWP($offbase+16*0,"esp"),$D0) if ($offbase);
 857         &movdqa         (&QWP($offbase+16*1,"esp"),$D1) if ($offbase);
 858 }
 859         &load_input     (16*2,16*5);
 860
 861         &jbe            (&label("skip_loop"));
 862         &jmp            (&label("loop"));
 863
 864 &set_label("loop",32);
 865         ################################################################
 866         # ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2
 867         # ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^3+inp[7]*r
 868         #   \___________________/
 869         # ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2+inp[8])*r^2
 870         # ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^4+inp[7]*r^2+inp[9])*r
 871         #   \___________________/ \____________________/
 872         ################################################################
 873
 874         &movdqa         ($T2,&QWP(16*(0-9),"edx"));     # r0^2
 875         &movdqa         (&QWP(16*1,"eax"),$T1);
 876         &movdqa         (&QWP(16*2,"eax"),$D2);
 877         &movdqa         (&QWP(16*3,"eax"),$D3);
 878         &movdqa         (&QWP(16*4,"eax"),$D4);
 879
 880         ################################################################
 881         # d4 = h4*r0 + h0*r4   + h1*r3   + h2*r2   + h3*r1
 882         # d3 = h3*r0 + h0*r3   + h1*r2   + h2*r1   + h4*5*r4
 883         # d2 = h2*r0 + h0*r2   + h1*r1   + h3*5*r4 + h4*5*r3
 884         # d1 = h1*r0 + h0*r1   + h2*5*r4 + h3*5*r3 + h4*5*r2
 885         # d0 = h0*r0 + h1*5*r4 + h2*5*r3 + h3*5*r2 + h4*5*r1
 886
 887         &movdqa         ($D1,$T0);
 888         &pmuludq        ($T0,$T2);                      # h0*r0
 889         &movdqa         ($D0,$T1);
 890         &pmuludq        ($T1,$T2);                      # h1*r0
 891         &pmuludq        ($D2,$T2);                      # h2*r0
 892         &pmuludq        ($D3,$T2);                      # h3*r0
 893         &pmuludq        ($D4,$T2);                      # h4*r0
 894
 895 sub pmuladd_alt {
 896 my $addr = shift;
 897
 898         &pmuludq        ($D0,&$addr(8));                # h1*s4
 899         &movdqa         ($T2,$D1);
 900         &pmuludq        ($D1,&$addr(1));                # h0*r1
 901         &paddq          ($D0,$T0);
 902         &movdqa         ($T0,$T2);
 903         &pmuludq        ($T2,&$addr(2));                # h0*r2
 904         &paddq          ($D1,$T1);
 905         &movdqa         ($T1,$T0);
 906         &pmuludq        ($T0,&$addr(3));                # h0*r3
 907         &paddq          ($D2,$T2);
 908          &movdqa        ($T2,&QWP(16*1,"eax"));         # pull h1
 909         &pmuludq        ($T1,&$addr(4));                # h0*r4
 910         &paddq          ($D3,$T0);
 911
 912         &movdqa         ($T0,$T2);
 913         &pmuludq        ($T2,&$addr(1));                # h1*r1
 914          &paddq         ($D4,$T1);
 915         &movdqa         ($T1,$T0);
 916         &pmuludq        ($T0,&$addr(2));                # h1*r2
 917         &paddq          ($D2,$T2);
 918         &movdqa         ($T2,&QWP(16*2,"eax"));         # pull h2
 919         &pmuludq        ($T1,&$addr(3));                # h1*r3
 920         &paddq          ($D3,$T0);
 921         &movdqa         ($T0,$T2);
 922         &pmuludq        ($T2,&$addr(7));                # h2*s3
 923         &paddq          ($D4,$T1);
 924         &movdqa         ($T1,$T0);
 925         &pmuludq        ($T0,&$addr(8));                # h2*s4
 926         &paddq          ($D0,$T2);
 927
 928         &movdqa         ($T2,$T1);
 929         &pmuludq        ($T1,&$addr(1));                # h2*r1
 930          &paddq         ($D1,$T0);
 931         &movdqa         ($T0,&QWP(16*3,"eax"));         # pull h3
 932         &pmuludq        ($T2,&$addr(2));                # h2*r2
 933         &paddq          ($D3,$T1);
 934         &movdqa         ($T1,$T0);
 935         &pmuludq        ($T0,&$addr(6));                # h3*s2
 936         &paddq          ($D4,$T2);
 937         &movdqa         ($T2,$T1);
 938         &pmuludq        ($T1,&$addr(7));                # h3*s3
 939         &paddq          ($D0,$T0);
 940         &movdqa         ($T0,$T2);
 941         &pmuludq        ($T2,&$addr(8));                # h3*s4
 942         &paddq          ($D1,$T1);
 943
 944         &movdqa         ($T1,&QWP(16*4,"eax"));         # pull h4
 945         &pmuludq        ($T0,&$addr(1));                # h3*r1
 946          &paddq         ($D2,$T2);
 947         &movdqa         ($T2,$T1);
 948         &pmuludq        ($T1,&$addr(8));                # h4*s4
 949         &paddq          ($D4,$T0);
 950         &movdqa         ($T0,$T2);
 951         &pmuludq        ($T2,&$addr(5));                # h4*s1
 952         &paddq          ($D3,$T1);
 953         &movdqa         ($T1,$T0);
 954         &pmuludq        ($T0,&$addr(6));                # h4*s2
 955         &paddq          ($D0,$T2);
 956          &movdqa        ($MASK,&QWP(64,"ebx"));
 957         &pmuludq        ($T1,&$addr(7));                # h4*s3
 958         &paddq          ($D1,$T0);
 959         &paddq          ($D2,$T1);
 960 }
 961         &pmuladd_alt    (sub {  my $i=shift; &QWP(16*($i-9),"edx");     });
 962
 963         &load_input     (-16*2,0);
 964         &lea            ("eax",&DWP(-16*2,"esi"));
 965         &sub            ("ecx",64);
 966
 967         &paddd          ($T0,&QWP(16*(5+0),"esp"));     # add hash value
 968         &paddd          ($T1,&QWP(16*(5+1),"esp"));
 969         &paddd          ($D2,&QWP(16*(5+2),"esp"));
 970         &paddd          ($D3,&QWP(16*(5+3),"esp"));
 971         &paddd          ($D4,&QWP(16*(5+4),"esp"));
 972
 973         &cmovb          ("esi","eax");
 974         &lea            ("eax",&DWP(16*10,"esp"));
 975
 976         &movdqa         ($T2,&QWP(16*0,"edx"));         # r0^4
 977         &movdqa         (&QWP(16*1,"esp"),$D1);
 978         &movdqa         (&QWP(16*1,"eax"),$T1);
 979         &movdqa         (&QWP(16*2,"eax"),$D2);
 980         &movdqa         (&QWP(16*3,"eax"),$D3);
 981         &movdqa         (&QWP(16*4,"eax"),$D4);
 982
 983         ################################################################
 984         # d4 += h4*r0 + h0*r4   + h1*r3   + h2*r2   + h3*r1
 985         # d3 += h3*r0 + h0*r3   + h1*r2   + h2*r1   + h4*5*r4
 986         # d2 += h2*r0 + h0*r2   + h1*r1   + h3*5*r4 + h4*5*r3
 987         # d1 += h1*r0 + h0*r1   + h2*5*r4 + h3*5*r3 + h4*5*r2
 988         # d0 += h0*r0 + h1*5*r4 + h2*5*r3 + h3*5*r2 + h4*5*r1
 989
 990         &movdqa         ($D1,$T0);
 991         &pmuludq        ($T0,$T2);                      # h0*r0
 992         &paddq          ($T0,$D0);
 993         &movdqa         ($D0,$T1);
 994         &pmuludq        ($T1,$T2);                      # h1*r0
 995         &pmuludq        ($D2,$T2);                      # h2*r0
 996         &pmuludq        ($D3,$T2);                      # h3*r0
 997         &pmuludq        ($D4,$T2);                      # h4*r0
 998
 999         &paddq          ($T1,&QWP(16*1,"esp"));
1000         &paddq          ($D2,&QWP(16*2,"esp"));
1001         &paddq          ($D3,&QWP(16*3,"esp"));
1002         &paddq          ($D4,&QWP(16*4,"esp"));
1003
1004         &pmuladd_alt    (sub {  my $i=shift; &QWP(16*$i,"edx"); });
1005
1006         &lazy_reduction ();
1007
1008         &load_input     (16*2,16*5);
1009
1010         &ja             (&label("loop"));
1011
1012 &set_label("skip_loop");
1013         ################################################################
1014         # multiply (inp[0:1]+hash) or inp[2:3] by r^2:r^1
1015
1016          &pshufd        ($T2,&QWP(16*(0-9),"edx"),0x10);# r0^n
1017         &add            ("ecx",32);
1018         &jnz            (&label("long_tail"));
1019
1020         &paddd          ($T0,$D0);                      # add hash value
1021         &paddd          ($T1,$D1);
1022         &paddd          ($D2,&QWP(16*7,"esp"));
1023         &paddd          ($D3,&QWP(16*8,"esp"));
1024         &paddd          ($D4,&QWP(16*9,"esp"));
1025
1026 &set_label("long_tail");
1027
1028         &movdqa         (&QWP(16*0,"eax"),$T0);
1029         &movdqa         (&QWP(16*1,"eax"),$T1);
1030         &movdqa         (&QWP(16*2,"eax"),$D2);
1031         &movdqa         (&QWP(16*3,"eax"),$D3);
1032         &movdqa         (&QWP(16*4,"eax"),$D4);
1033
1034         ################################################################
1035         # d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
1036         # d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
1037         # d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
1038         # d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
1039         # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1040
1041         &pmuludq        ($T0,$T2);                      # h0*r0
1042         &pmuludq        ($T1,$T2);                      # h1*r0
1043         &pmuludq        ($D2,$T2);                      # h2*r0
1044         &movdqa         ($D0,$T0);
1045          &pshufd        ($T0,&QWP(16*(1-9),"edx"),0x10);# r1^n
1046         &pmuludq        ($D3,$T2);                      # h3*r0
1047         &movdqa         ($D1,$T1);
1048         &pmuludq        ($D4,$T2);                      # h4*r0
1049
1050         &pmuladd        (sub {  my ($reg,$i)=@_;
1051                                 &pshufd ($reg,&QWP(16*($i-9),"edx"),0x10);
1052                              },"eax");
1053
1054         &jz             (&label("short_tail"));
1055
1056         &load_input     (-16*2,0);
1057
1058          &pshufd        ($T2,&QWP(16*0,"edx"),0x10);    # r0^n
1059         &paddd          ($T0,&QWP(16*5,"esp"));         # add hash value
1060         &paddd          ($T1,&QWP(16*6,"esp"));
1061         &paddd          ($D2,&QWP(16*7,"esp"));
1062         &paddd          ($D3,&QWP(16*8,"esp"));
1063         &paddd          ($D4,&QWP(16*9,"esp"));
1064
1065         ################################################################
1066         # multiply inp[0:1] by r^4:r^3 and accumulate
1067
1068         &movdqa         (&QWP(16*0,"esp"),$T0);
1069         &pmuludq        ($T0,$T2);                      # h0*r0
1070         &movdqa         (&QWP(16*1,"esp"),$T1);
1071         &pmuludq        ($T1,$T2);                      # h1*r0
1072         &paddq          ($D0,$T0);
1073         &movdqa         ($T0,$D2);
1074         &pmuludq        ($D2,$T2);                      # h2*r0
1075         &paddq          ($D1,$T1);
1076         &movdqa         ($T1,$D3);
1077         &pmuludq        ($D3,$T2);                      # h3*r0
1078         &paddq          ($D2,&QWP(16*2,"esp"));
1079         &movdqa         (&QWP(16*2,"esp"),$T0);
1080          &pshufd        ($T0,&QWP(16*1,"edx"),0x10);    # r1^n
1081         &paddq          ($D3,&QWP(16*3,"esp"));
1082         &movdqa         (&QWP(16*3,"esp"),$T1);
1083         &movdqa         ($T1,$D4);
1084         &pmuludq        ($D4,$T2);                      # h4*r0
1085         &paddq          ($D4,&QWP(16*4,"esp"));
1086         &movdqa         (&QWP(16*4,"esp"),$T1);
1087
1088         &pmuladd        (sub {  my ($reg,$i)=@_;
1089                                 &pshufd ($reg,&QWP(16*$i,"edx"),0x10);
1090                              });
1091
1092 &set_label("short_tail");
1093
1094         &lazy_reduction ();
1095
1096         ################################################################
1097         # horizontal addition
1098
1099         &pshufd         ($T1,$D0,0b01001110);
1100         &pshufd         ($T0,$D1,0b01001110);
1101         &paddd          ($D0,$T1);
1102         &pshufd         ($T1,$D2,0b01001110);
1103         &paddd          ($D1,$T0);
1104         &pshufd         ($T0,$D3,0b01001110);
1105         &paddd          ($D2,$T1);
1106         &pshufd         ($T1,$D4,0b01001110);
1107         &paddd          ($D3,$T0);
1108         &paddd          ($D4,$T1);
1109
1110 &set_label("done");
1111         &movd           (&DWP(-16*3+4*0,"edi"),$D0);    # store hash value
1112         &movd           (&DWP(-16*3+4*1,"edi"),$D1);
1113         &movd           (&DWP(-16*3+4*2,"edi"),$D2);
1114         &movd           (&DWP(-16*3+4*3,"edi"),$D3);
1115         &movd           (&DWP(-16*3+4*4,"edi"),$D4);
1116 &set_label("nodata");
1117         &mov    ("esp","ebp");
1118 &function_end("_poly1305_blocks_sse2");
1119
1120 &align  (32);
1121 &function_begin("_poly1305_emit_sse2");
1122         &mov    ("ebp",&wparam(0));             # context
1123
1124         &cmp    (&DWP(4*5,"ebp"),0);            # is_base2_26?
1125         &je     (&label("enter_emit"));
1126
1127         &mov    ("eax",&DWP(4*0,"ebp"));        # load hash value
1128         &mov    ("edi",&DWP(4*1,"ebp"));
1129         &mov    ("ecx",&DWP(4*2,"ebp"));
1130         &mov    ("edx",&DWP(4*3,"ebp"));
1131         &mov    ("esi",&DWP(4*4,"ebp"));
1132
1133         &mov    ("ebx","edi");                  # base 2^26 -> base 2^32
1134         &shl    ("edi",26);
1135         &shr    ("ebx",6);
1136         &add    ("eax","edi");
1137         &mov    ("edi","ecx");
1138         &adc    ("ebx",0);
1139
1140         &shl    ("edi",20);
1141         &shr    ("ecx",12);
1142         &add    ("ebx","edi");
1143         &mov    ("edi","edx");
1144         &adc    ("ecx",0);
1145
1146         &shl    ("edi",14);
1147         &shr    ("edx",18);
1148         &add    ("ecx","edi");
1149         &mov    ("edi","esi");
1150         &adc    ("edx",0);
1151
1152         &shl    ("edi",8);
1153         &shr    ("esi",24);
1154         &add    ("edx","edi");
1155         &adc    ("esi",0);                      # can be partially reduced
1156
1157         &mov    ("edi","esi");                  # final reduction
1158         &and    ("esi",3);
1159         &shr    ("edi",2);
1160         &lea    ("ebp",&DWP(0,"edi","edi",4));  # *5
1161          &mov   ("edi",&wparam(1));             # output
1162         add     ("eax","ebp");
1163          &mov   ("ebp",&wparam(2));             # key
1164         adc     ("ebx",0);
1165         adc     ("ecx",0);
1166         adc     ("edx",0);
1167
1168         &movd   ($D0,"eax");                    # offload original hash value
1169         &add    ("eax",5);                      # compare to modulus
1170         &movd   ($D1,"ebx");
1171         &adc    ("ebx",0);
1172         &movd   ($D2,"ecx");
1173         &adc    ("ecx",0);
1174         &movd   ($D3,"edx");
1175         &adc    ("edx",0);
1176         &adc    ("esi",0);
1177         &shr    ("esi",2);                      # did it carry/borrow?
1178
1179         &neg    ("esi");                        # do we choose (hash-modulus) ...
1180         &and    ("eax","esi");
1181         &and    ("ebx","esi");
1182         &and    ("ecx","esi");
1183         &and    ("edx","esi");
1184         &mov    (&DWP(4*0,"edi"),"eax");
1185         &movd   ("eax",$D0);
1186         &mov    (&DWP(4*1,"edi"),"ebx");
1187         &movd   ("ebx",$D1);
1188         &mov    (&DWP(4*2,"edi"),"ecx");
1189         &movd   ("ecx",$D2);
1190         &mov    (&DWP(4*3,"edi"),"edx");
1191         &movd   ("edx",$D3);
1192
1193         &not    ("esi");                        # ... or original hash value?
1194         &and    ("eax","esi");
1195         &and    ("ebx","esi");
1196         &or     ("eax",&DWP(4*0,"edi"));
1197         &and    ("ecx","esi");
1198         &or     ("ebx",&DWP(4*1,"edi"));
1199         &and    ("edx","esi");
1200         &or     ("ecx",&DWP(4*2,"edi"));
1201         &or     ("edx",&DWP(4*3,"edi"));
1202
1203         &add    ("eax",&DWP(4*0,"ebp"));        # accumulate key
1204         &adc    ("ebx",&DWP(4*1,"ebp"));
1205         &mov    (&DWP(4*0,"edi"),"eax");
1206         &adc    ("ecx",&DWP(4*2,"ebp"));
1207         &mov    (&DWP(4*1,"edi"),"ebx");
1208         &adc    ("edx",&DWP(4*3,"ebp"));
1209         &mov    (&DWP(4*2,"edi"),"ecx");
1210         &mov    (&DWP(4*3,"edi"),"edx");
1211 &function_end("_poly1305_emit_sse2");
1212
1213 if ($avx>1) {
1214 ########################################################################
1215 # Note that poly1305_init_avx2 operates on %xmm, I could have used
1216 # poly1305_init_sse2...
1217
1218 &align  (32);
1219 &function_begin_B("_poly1305_init_avx2");
1220         &vmovdqu        ($D4,&QWP(4*6,"edi"));          # key base 2^32
1221         &lea            ("edi",&DWP(16*3,"edi"));       # size optimization
1222         &mov            ("ebp","esp");
1223         &sub            ("esp",16*(9+5));
1224         &and            ("esp",-16);
1225
1226         #&vpand         ($D4,$D4,&QWP(96,"ebx"));       # magic mask
1227         &vmovdqa        ($MASK,&QWP(64,"ebx"));
1228
1229         &vpand          ($D0,$D4,$MASK);                # -> base 2^26
1230         &vpsrlq         ($D1,$D4,26);
1231         &vpsrldq        ($D3,$D4,6);
1232         &vpand          ($D1,$D1,$MASK);
1233         &vpsrlq         ($D2,$D3,4)
1234         &vpsrlq         ($D3,$D3,30);
1235         &vpand          ($D2,$D2,$MASK);
1236         &vpand          ($D3,$D3,$MASK);
1237         &vpsrldq        ($D4,$D4,13);
1238
1239         &lea            ("edx",&DWP(16*9,"esp"));       # size optimization
1240         &mov            ("ecx",2);
1241 &set_label("square");
1242         &vmovdqa        (&QWP(16*0,"esp"),$D0);
1243         &vmovdqa        (&QWP(16*1,"esp"),$D1);
1244         &vmovdqa        (&QWP(16*2,"esp"),$D2);
1245         &vmovdqa        (&QWP(16*3,"esp"),$D3);
1246         &vmovdqa        (&QWP(16*4,"esp"),$D4);
1247
1248         &vpslld         ($T1,$D1,2);
1249         &vpslld         ($T0,$D2,2);
1250         &vpaddd         ($T1,$T1,$D1);                  # *5
1251         &vpaddd         ($T0,$T0,$D2);                  # *5
1252         &vmovdqa        (&QWP(16*5,"esp"),$T1);
1253         &vmovdqa        (&QWP(16*6,"esp"),$T0);
1254         &vpslld         ($T1,$D3,2);
1255         &vpslld         ($T0,$D4,2);
1256         &vpaddd         ($T1,$T1,$D3);                  # *5
1257         &vpaddd         ($T0,$T0,$D4);                  # *5
1258         &vmovdqa        (&QWP(16*7,"esp"),$T1);
1259         &vmovdqa        (&QWP(16*8,"esp"),$T0);
1260
1261         &vpshufd        ($T0,$D0,0b01000100);
1262         &vmovdqa        ($T1,$D1);
1263         &vpshufd        ($D1,$D1,0b01000100);
1264         &vpshufd        ($D2,$D2,0b01000100);
1265         &vpshufd        ($D3,$D3,0b01000100);
1266         &vpshufd        ($D4,$D4,0b01000100);
1267         &vmovdqa        (&QWP(16*0,"edx"),$T0);
1268         &vmovdqa        (&QWP(16*1,"edx"),$D1);
1269         &vmovdqa        (&QWP(16*2,"edx"),$D2);
1270         &vmovdqa        (&QWP(16*3,"edx"),$D3);
1271         &vmovdqa        (&QWP(16*4,"edx"),$D4);
1272
1273         ################################################################
1274         # d4 = h4*r0 + h3*r1   + h2*r2   + h1*r3   + h0*r4
1275         # d3 = h3*r0 + h2*r1   + h1*r2   + h0*r3   + h4*5*r4
1276         # d2 = h2*r0 + h1*r1   + h0*r2   + h4*5*r3 + h3*5*r4
1277         # d1 = h1*r0 + h0*r1   + h4*5*r2 + h3*5*r3 + h2*5*r4
1278         # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1279
1280         &vpmuludq       ($D4,$D4,$D0);                  # h4*r0
1281         &vpmuludq       ($D3,$D3,$D0);                  # h3*r0
1282         &vpmuludq       ($D2,$D2,$D0);                  # h2*r0
1283         &vpmuludq       ($D1,$D1,$D0);                  # h1*r0
1284         &vpmuludq       ($D0,$T0,$D0);                  # h0*r0
1285
1286         &vpmuludq       ($T0,$T1,&QWP(16*3,"edx"));     # r1*h3
1287         &vpaddq         ($D4,$D4,$T0);
1288         &vpmuludq       ($T2,$T1,&QWP(16*2,"edx"));     # r1*h2
1289         &vpaddq         ($D3,$D3,$T2);
1290         &vpmuludq       ($T0,$T1,&QWP(16*1,"edx"));     # r1*h1
1291         &vpaddq         ($D2,$D2,$T0);
1292         &vmovdqa        ($T2,&QWP(16*5,"esp"));         # s1
1293         &vpmuludq       ($T1,$T1,&QWP(16*0,"edx"));     # r1*h0
1294         &vpaddq         ($D1,$D1,$T1);
1295          &vmovdqa       ($T0,&QWP(16*2,"esp"));         # r2
1296         &vpmuludq       ($T2,$T2,&QWP(16*4,"edx"));     # s1*h4
1297         &vpaddq         ($D0,$D0,$T2);
1298
1299         &vpmuludq       ($T1,$T0,&QWP(16*2,"edx"));     # r2*h2
1300         &vpaddq         ($D4,$D4,$T1);
1301         &vpmuludq       ($T2,$T0,&QWP(16*1,"edx"));     # r2*h1
1302         &vpaddq         ($D3,$D3,$T2);
1303         &vmovdqa        ($T1,&QWP(16*6,"esp"));         # s2
1304         &vpmuludq       ($T0,$T0,&QWP(16*0,"edx"));     # r2*h0
1305         &vpaddq         ($D2,$D2,$T0);
1306         &vpmuludq       ($T2,$T1,&QWP(16*4,"edx"));     # s2*h4
1307         &vpaddq         ($D1,$D1,$T2);
1308          &vmovdqa       ($T0,&QWP(16*3,"esp"));         # r3
1309         &vpmuludq       ($T1,$T1,&QWP(16*3,"edx"));     # s2*h3
1310         &vpaddq         ($D0,$D0,$T1);
1311
1312         &vpmuludq       ($T2,$T0,&QWP(16*1,"edx"));     # r3*h1
1313         &vpaddq         ($D4,$D4,$T2);
1314         &vmovdqa        ($T1,&QWP(16*7,"esp"));         # s3
1315         &vpmuludq       ($T0,$T0,&QWP(16*0,"edx"));     # r3*h0
1316         &vpaddq         ($D3,$D3,$T0);
1317         &vpmuludq       ($T2,$T1,&QWP(16*4,"edx"));     # s3*h4
1318         &vpaddq         ($D2,$D2,$T2);
1319         &vpmuludq       ($T0,$T1,&QWP(16*3,"edx"));     # s3*h3
1320         &vpaddq         ($D1,$D1,$T0);
1321          &vmovdqa       ($T2,&QWP(16*4,"esp"));         # r4
1322         &vpmuludq       ($T1,$T1,&QWP(16*2,"edx"));     # s3*h2
1323         &vpaddq         ($D0,$D0,$T1);
1324
1325         &vmovdqa        ($T0,&QWP(16*8,"esp"));         # s4
1326         &vpmuludq       ($T2,$T2,&QWP(16*0,"edx"));     # r4*h0
1327         &vpaddq         ($D4,$D4,$T2);
1328         &vpmuludq       ($T1,$T0,&QWP(16*4,"edx"));     # s4*h4
1329         &vpaddq         ($D3,$D3,$T1);
1330         &vpmuludq       ($T2,$T0,&QWP(16*1,"edx"));     # s4*h1
1331         &vpaddq         ($D0,$D0,$T2);
1332         &vpmuludq       ($T1,$T0,&QWP(16*2,"edx"));     # s4*h2
1333         &vpaddq         ($D1,$D1,$T1);
1334          &vmovdqa       ($MASK,&QWP(64,"ebx"));
1335         &vpmuludq       ($T0,$T0,&QWP(16*3,"edx"));     # s4*h3
1336         &vpaddq         ($D2,$D2,$T0);
1337
1338         ################################################################
1339         # lazy reduction
1340          &vpsrlq        ($T0,$D3,26);
1341          &vpand         ($D3,$D3,$MASK);
1342         &vpsrlq         ($T1,$D0,26);
1343         &vpand          ($D0,$D0,$MASK);
1344          &vpaddq        ($D4,$D4,$T0);                  # h3 -> h4
1345         &vpaddq         ($D1,$D1,$T1);                  # h0 -> h1
1346          &vpsrlq        ($T0,$D4,26);
1347          &vpand         ($D4,$D4,$MASK);
1348         &vpsrlq         ($T1,$D1,26);
1349         &vpand          ($D1,$D1,$MASK);
1350         &vpaddq         ($D2,$D2,$T1);                  # h1 -> h2
1351          &vpaddd        ($D0,$D0,$T0);
1352          &vpsllq        ($T0,$T0,2);
1353         &vpsrlq         ($T1,$D2,26);
1354         &vpand          ($D2,$D2,$MASK);
1355          &vpaddd        ($D0,$D0,$T0);                  # h4 -> h0
1356         &vpaddd         ($D3,$D3,$T1);                  # h2 -> h3
1357         &vpsrlq         ($T1,$D3,26);
1358          &vpsrlq        ($T0,$D0,26);
1359          &vpand         ($D0,$D0,$MASK);
1360         &vpand          ($D3,$D3,$MASK);
1361          &vpaddd        ($D1,$D1,$T0);                  # h0 -> h1
1362         &vpaddd         ($D4,$D4,$T1);                  # h3 -> h4
1363
1364         &dec            ("ecx");
1365         &jz             (&label("square_break"));
1366
1367         &vpunpcklqdq    ($D0,$D0,&QWP(16*0,"esp"));     # 0:r^1:0:r^2
1368         &vpunpcklqdq    ($D1,$D1,&QWP(16*1,"esp"));
1369         &vpunpcklqdq    ($D2,$D2,&QWP(16*2,"esp"));
1370         &vpunpcklqdq    ($D3,$D3,&QWP(16*3,"esp"));
1371         &vpunpcklqdq    ($D4,$D4,&QWP(16*4,"esp"));
1372         &jmp            (&label("square"));
1373
1374 &set_label("square_break");
1375         &vpsllq         ($D0,$D0,32);                   # -> r^3:0:r^4:0
1376         &vpsllq         ($D1,$D1,32);
1377         &vpsllq         ($D2,$D2,32);
1378         &vpsllq         ($D3,$D3,32);
1379         &vpsllq         ($D4,$D4,32);
1380         &vpor           ($D0,$D0,&QWP(16*0,"esp"));     # r^3:r^1:r^4:r^2
1381         &vpor           ($D1,$D1,&QWP(16*1,"esp"));
1382         &vpor           ($D2,$D2,&QWP(16*2,"esp"));
1383         &vpor           ($D3,$D3,&QWP(16*3,"esp"));
1384         &vpor           ($D4,$D4,&QWP(16*4,"esp"));
1385
1386         &vpshufd        ($D0,$D0,0b10001101);           # -> r^1:r^2:r^3:r^4
1387         &vpshufd        ($D1,$D1,0b10001101);
1388         &vpshufd        ($D2,$D2,0b10001101);
1389         &vpshufd        ($D3,$D3,0b10001101);
1390         &vpshufd        ($D4,$D4,0b10001101);
1391
1392         &vmovdqu        (&QWP(16*0,"edi"),$D0);         # save the table
1393         &vmovdqu        (&QWP(16*1,"edi"),$D1);
1394         &vmovdqu        (&QWP(16*2,"edi"),$D2);
1395         &vmovdqu        (&QWP(16*3,"edi"),$D3);
1396         &vmovdqu        (&QWP(16*4,"edi"),$D4);
1397
1398         &vpslld         ($T1,$D1,2);
1399         &vpslld         ($T0,$D2,2);
1400         &vpaddd         ($T1,$T1,$D1);                  # *5
1401         &vpaddd         ($T0,$T0,$D2);                  # *5
1402         &vmovdqu        (&QWP(16*5,"edi"),$T1);
1403         &vmovdqu        (&QWP(16*6,"edi"),$T0);
1404         &vpslld         ($T1,$D3,2);
1405         &vpslld         ($T0,$D4,2);
1406         &vpaddd         ($T1,$T1,$D3);                  # *5
1407         &vpaddd         ($T0,$T0,$D4);                  # *5
1408         &vmovdqu        (&QWP(16*7,"edi"),$T1);
1409         &vmovdqu        (&QWP(16*8,"edi"),$T0);
1410
1411         &mov            ("esp","ebp");
1412         &lea            ("edi",&DWP(-16*3,"edi"));      # size de-optimization
1413         &ret            ();
1414 &function_end_B("_poly1305_init_avx2");
1415
1416 ########################################################################
1417 # now it's time to switch to %ymm
1418
1419 my ($D0,$D1,$D2,$D3,$D4,$T0,$T1,$T2)=map("ymm$_",(0..7));
1420 my $MASK=$T2;
1421
1422 sub X { my $reg=shift; $reg=~s/^ymm/xmm/; $reg; }
1423
1424 &align  (32);
1425 &function_begin("_poly1305_blocks_avx2");
1426         &mov    ("edi",&wparam(0));                     # ctx
1427         &mov    ("esi",&wparam(1));                     # inp
1428         &mov    ("ecx",&wparam(2));                     # len
1429
1430         &mov    ("eax",&DWP(4*5,"edi"));                # is_base2_26
1431         &and    ("ecx",-16);
1432         &jz     (&label("nodata"));
1433         &cmp    ("ecx",64);
1434         &jae    (&label("enter_avx2"));
1435         &test   ("eax","eax");                          # is_base2_26?
1436         &jz     (&label("enter_blocks"));
1437
1438 &set_label("enter_avx2",16);
1439         &vzeroupper     ();
1440
1441         &call   (&label("pic_point"));
1442 &set_label("pic_point");
1443         &blindpop("ebx");
1444         &lea    ("ebx",&DWP(&label("const_sse2")."-".&label("pic_point"),"ebx"));
1445
1446         &test   ("eax","eax");                          # is_base2_26?
1447         &jnz    (&label("base2_26"));
1448
1449         &call   ("_poly1305_init_avx2");
1450
1451         ################################################# base 2^32 -> base 2^26
1452         &mov    ("eax",&DWP(0,"edi"));
1453         &mov    ("ecx",&DWP(3,"edi"));
1454         &mov    ("edx",&DWP(6,"edi"));
1455         &mov    ("esi",&DWP(9,"edi"));
1456         &mov    ("ebp",&DWP(13,"edi"));
1457
1458         &shr    ("ecx",2);
1459         &and    ("eax",0x3ffffff);
1460         &shr    ("edx",4);
1461         &and    ("ecx",0x3ffffff);
1462         &shr    ("esi",6);
1463         &and    ("edx",0x3ffffff);
1464
1465         &mov    (&DWP(4*0,"edi"),"eax");
1466         &mov    (&DWP(4*1,"edi"),"ecx");
1467         &mov    (&DWP(4*2,"edi"),"edx");
1468         &mov    (&DWP(4*3,"edi"),"esi");
1469         &mov    (&DWP(4*4,"edi"),"ebp");
1470         &mov    (&DWP(4*5,"edi"),1);                    # is_base2_26
1471
1472         &mov    ("esi",&wparam(1));                     # [reload] inp
1473         &mov    ("ecx",&wparam(2));                     # [reload] len
1474
1475 &set_label("base2_26");
1476         &mov    ("eax",&wparam(3));                     # padbit
1477         &mov    ("ebp","esp");
1478
1479         &sub    ("esp",32*(5+9));
1480         &and    ("esp",-512);                           # ensure that frame
1481                                                         # doesn't cross page
1482                                                         # boundary, which is
1483                                                         # essential for
1484                                                         # misaligned 32-byte
1485                                                         # loads
1486
1487         ################################################################
1488         # expand and copy pre-calculated table to stack
1489
1490         &vmovdqu        (&X($D0),&QWP(16*(3+0),"edi"));
1491         &lea            ("edx",&DWP(32*5+128,"esp"));   # +128 size optimization
1492         &vmovdqu        (&X($D1),&QWP(16*(3+1),"edi"));
1493         &vmovdqu        (&X($D2),&QWP(16*(3+2),"edi"));
1494         &vmovdqu        (&X($D3),&QWP(16*(3+3),"edi"));
1495         &vmovdqu        (&X($D4),&QWP(16*(3+4),"edi"));
1496         &lea            ("edi",&DWP(16*3,"edi"));       # size optimization
1497         &vpermq         ($D0,$D0,0b01000000);           # 00001234 -> 12343434
1498         &vpermq         ($D1,$D1,0b01000000);
1499         &vpermq         ($D2,$D2,0b01000000);
1500         &vpermq         ($D3,$D3,0b01000000);
1501         &vpermq         ($D4,$D4,0b01000000);
1502         &vpshufd        ($D0,$D0,0b11001000);           # 12343434 -> 14243444
1503         &vpshufd        ($D1,$D1,0b11001000);
1504         &vpshufd        ($D2,$D2,0b11001000);
1505         &vpshufd        ($D3,$D3,0b11001000);
1506         &vpshufd        ($D4,$D4,0b11001000);
1507         &vmovdqa        (&QWP(32*0-128,"edx"),$D0);
1508         &vmovdqu        (&X($D0),&QWP(16*5,"edi"));
1509         &vmovdqa        (&QWP(32*1-128,"edx"),$D1);
1510         &vmovdqu        (&X($D1),&QWP(16*6,"edi"));
1511         &vmovdqa        (&QWP(32*2-128,"edx"),$D2);
1512         &vmovdqu        (&X($D2),&QWP(16*7,"edi"));
1513         &vmovdqa        (&QWP(32*3-128,"edx"),$D3);
1514         &vmovdqu        (&X($D3),&QWP(16*8,"edi"));
1515         &vmovdqa        (&QWP(32*4-128,"edx"),$D4);
1516         &vpermq         ($D0,$D0,0b01000000);
1517         &vpermq         ($D1,$D1,0b01000000);
1518         &vpermq         ($D2,$D2,0b01000000);
1519         &vpermq         ($D3,$D3,0b01000000);
1520         &vpshufd        ($D0,$D0,0b11001000);
1521         &vpshufd        ($D1,$D1,0b11001000);
1522         &vpshufd        ($D2,$D2,0b11001000);
1523         &vpshufd        ($D3,$D3,0b11001000);
1524         &vmovdqa        (&QWP(32*5-128,"edx"),$D0);
1525         &vmovd          (&X($D0),&DWP(-16*3+4*0,"edi"));# load hash value
1526         &vmovdqa        (&QWP(32*6-128,"edx"),$D1);
1527         &vmovd          (&X($D1),&DWP(-16*3+4*1,"edi"));
1528         &vmovdqa        (&QWP(32*7-128,"edx"),$D2);
1529         &vmovd          (&X($D2),&DWP(-16*3+4*2,"edi"));
1530         &vmovdqa        (&QWP(32*8-128,"edx"),$D3);
1531         &vmovd          (&X($D3),&DWP(-16*3+4*3,"edi"));
1532         &vmovd          (&X($D4),&DWP(-16*3+4*4,"edi"));
1533         &vmovdqa        ($MASK,&QWP(64,"ebx"));
1534         &neg            ("eax");                        # padbit
1535
1536         &test           ("ecx",63);
1537         &jz             (&label("even"));
1538
1539         &mov            ("edx","ecx");
1540         &and            ("ecx",-64);
1541         &and            ("edx",63);
1542
1543         &vmovdqu        (&X($T0),&QWP(16*0,"esi"));
1544         &cmp            ("edx",32);
1545         &jb             (&label("one"));
1546
1547         &vmovdqu        (&X($T1),&QWP(16*1,"esi"));
1548         &je             (&label("two"));
1549
1550         &vinserti128    ($T0,$T0,&QWP(16*2,"esi"),1);
1551         &lea            ("esi",&DWP(16*3,"esi"));
1552         &lea            ("ebx",&DWP(8,"ebx"));          # three padbits
1553         &lea            ("edx",&DWP(32*5+128+8,"esp")); # --:r^1:r^2:r^3 (*)
1554         &jmp            (&label("tail"));
1555
1556 &set_label("two");
1557         &lea            ("esi",&DWP(16*2,"esi"));
1558         &lea            ("ebx",&DWP(16,"ebx"));         # two padbits
1559         &lea            ("edx",&DWP(32*5+128+16,"esp"));# --:--:r^1:r^2 (*)
1560         &jmp            (&label("tail"));
1561
1562 &set_label("one");
1563         &lea            ("esi",&DWP(16*1,"esi"));
1564         &vpxor          ($T1,$T1,$T1);
1565         &lea            ("ebx",&DWP(32,"ebx","eax",8)); # one or no padbits
1566         &lea            ("edx",&DWP(32*5+128+24,"esp"));# --:--:--:r^1 (*)
1567         &jmp            (&label("tail"));
1568
1569 # (*)   spots marked with '--' are data from next table entry, but they
1570 #       are multiplied by 0 and therefore rendered insignificant
1571
1572 &set_label("even",32);
1573         &vmovdqu        (&X($T0),&QWP(16*0,"esi"));     # load input
1574         &vmovdqu        (&X($T1),&QWP(16*1,"esi"));
1575         &vinserti128    ($T0,$T0,&QWP(16*2,"esi"),1);
1576         &vinserti128    ($T1,$T1,&QWP(16*3,"esi"),1);
1577         &lea            ("esi",&DWP(16*4,"esi"));
1578         &sub            ("ecx",64);
1579         &jz             (&label("tail"));
1580
1581 &set_label("loop");
1582         ################################################################
1583         # ((inp[0]*r^4+r[4])*r^4+r[8])*r^4
1584         # ((inp[1]*r^4+r[5])*r^4+r[9])*r^3
1585         # ((inp[2]*r^4+r[6])*r^4+r[10])*r^2
1586         # ((inp[3]*r^4+r[7])*r^4+r[11])*r^1
1587         #   \________/ \_______/
1588         ################################################################
1589
1590 sub vsplat_input {
1591         &vmovdqa        (&QWP(32*2,"esp"),$D2);
1592         &vpsrldq        ($D2,$T0,6);                    # splat input
1593         &vmovdqa        (&QWP(32*0,"esp"),$D0);
1594         &vpsrldq        ($D0,$T1,6);
1595         &vmovdqa        (&QWP(32*1,"esp"),$D1);
1596         &vpunpckhqdq    ($D1,$T0,$T1);                  # 4
1597         &vpunpcklqdq    ($T0,$T0,$T1);                  # 0:1
1598         &vpunpcklqdq    ($D2,$D2,$D0);                  # 2:3
1599
1600         &vpsrlq         ($D0,$D2,30);
1601         &vpsrlq         ($D2,$D2,4);
1602         &vpsrlq         ($T1,$T0,26);
1603         &vpsrlq         ($D1,$D1,40);                   # 4
1604         &vpand          ($D2,$D2,$MASK);                # 2
1605         &vpand          ($T0,$T0,$MASK);                # 0
1606         &vpand          ($T1,$T1,$MASK);                # 1
1607         &vpand          ($D0,$D0,$MASK);                # 3 (*)
1608         &vpor           ($D1,$D1,&QWP(0,"ebx"));        # padbit, yes, always
1609
1610         # (*)   note that output is counterintuitive, inp[3:4] is
1611         #       returned in $D1-2, while $D3-4 are preserved;
1612 }
1613         &vsplat_input   ();
1614
1615 sub vpmuladd {
1616 my $addr = shift;
1617
1618         &vpaddq         ($D2,$D2,&QWP(32*2,"esp"));     # add hash value
1619         &vpaddq         ($T0,$T0,&QWP(32*0,"esp"));
1620         &vpaddq         ($T1,$T1,&QWP(32*1,"esp"));
1621         &vpaddq         ($D0,$D0,$D3);
1622         &vpaddq         ($D1,$D1,$D4);
1623
1624         ################################################################
1625         # d3 = h2*r1   + h0*r3 + h1*r2   + h3*r0   + h4*5*r4
1626         # d4 = h2*r2   + h0*r4 + h1*r3   + h3*r1   + h4*r0
1627         # d0 = h2*5*r3 + h0*r0 + h1*5*r4 + h3*5*r2 + h4*5*r1
1628         # d1 = h2*5*r4 + h0*r1 + h1*r0   + h3*5*r3 + h4*5*r2
1629         # d2 = h2*r0   + h0*r2 + h1*r1   + h3*5*r4 + h4*5*r3
1630
1631         &vpmuludq       ($D3,$D2,&$addr(1));            # d3 = h2*r1
1632          &vmovdqa       (QWP(32*1,"esp"),$T1);
1633         &vpmuludq       ($D4,$D2,&$addr(2));            # d4 = h2*r2
1634          &vmovdqa       (QWP(32*3,"esp"),$D0);
1635         &vpmuludq       ($D0,$D2,&$addr(7));            # d0 = h2*s3
1636          &vmovdqa       (QWP(32*4,"esp"),$D1);
1637         &vpmuludq       ($D1,$D2,&$addr(8));            # d1 = h2*s4
1638         &vpmuludq       ($D2,$D2,&$addr(0));            # d2 = h2*r0
1639
1640         &vpmuludq       ($T2,$T0,&$addr(3));            # h0*r3
1641         &vpaddq         ($D3,$D3,$T2);                  # d3 += h0*r3
1642         &vpmuludq       ($T1,$T0,&$addr(4));            # h0*r4
1643         &vpaddq         ($D4,$D4,$T1);                  # d4 + h0*r4
1644         &vpmuludq       ($T2,$T0,&$addr(0));            # h0*r0
1645         &vpaddq         ($D0,$D0,$T2);                  # d0 + h0*r0
1646          &vmovdqa       ($T2,&QWP(32*1,"esp"));         # h1
1647         &vpmuludq       ($T1,$T0,&$addr(1));            # h0*r1
1648         &vpaddq         ($D1,$D1,$T1);                  # d1 += h0*r1
1649         &vpmuludq       ($T0,$T0,&$addr(2));            # h0*r2
1650         &vpaddq         ($D2,$D2,$T0);                  # d2 += h0*r2
1651
1652         &vpmuludq       ($T1,$T2,&$addr(2));            # h1*r2
1653         &vpaddq         ($D3,$D3,$T1);                  # d3 += h1*r2
1654         &vpmuludq       ($T0,$T2,&$addr(3));            # h1*r3
1655         &vpaddq         ($D4,$D4,$T0);                  # d4 += h1*r3
1656         &vpmuludq       ($T1,$T2,&$addr(8));            # h1*s4
1657         &vpaddq         ($D0,$D0,$T1);                  # d0 += h1*s4
1658          &vmovdqa       ($T1,&QWP(32*3,"esp"));         # h3
1659         &vpmuludq       ($T0,$T2,&$addr(0));            # h1*r0
1660         &vpaddq         ($D1,$D1,$T0);                  # d1 += h1*r0
1661         &vpmuludq       ($T2,$T2,&$addr(1));            # h1*r1
1662         &vpaddq         ($D2,$D2,$T2);                  # d2 += h1*r1
1663
1664         &vpmuludq       ($T0,$T1,&$addr(0));            # h3*r0
1665         &vpaddq         ($D3,$D3,$T0);                  # d3 += h3*r0
1666         &vpmuludq       ($T2,$T1,&$addr(1));            # h3*r1
1667         &vpaddq         ($D4,$D4,$T2);                  # d4 += h3*r1
1668         &vpmuludq       ($T0,$T1,&$addr(6));            # h3*s2
1669         &vpaddq         ($D0,$D0,$T0);                  # d0 += h3*s2
1670          &vmovdqa       ($T0,&QWP(32*4,"esp"));         # h4
1671         &vpmuludq       ($T2,$T1,&$addr(7));            # h3*s3
1672         &vpaddq         ($D1,$D1,$T2);                  # d1+= h3*s3
1673         &vpmuludq       ($T1,$T1,&$addr(8));            # h3*s4
1674         &vpaddq         ($D2,$D2,$T1);                  # d2 += h3*s4
1675
1676         &vpmuludq       ($T2,$T0,&$addr(8));            # h4*s4
1677         &vpaddq         ($D3,$D3,$T2);                  # d3 += h4*s4
1678         &vpmuludq       ($T1,$T0,&$addr(5));            # h4*s1
1679         &vpaddq         ($D0,$D0,$T1);                  # d0 += h4*s1
1680         &vpmuludq       ($T2,$T0,&$addr(0));            # h4*r0
1681         &vpaddq         ($D4,$D4,$T2);                  # d4 += h4*r0
1682          &vmovdqa       ($MASK,&QWP(64,"ebx"));
1683         &vpmuludq       ($T1,$T0,&$addr(6));            # h4*s2
1684         &vpaddq         ($D1,$D1,$T1);                  # d1 += h4*s2
1685         &vpmuludq       ($T0,$T0,&$addr(7));            # h4*s3
1686         &vpaddq         ($D2,$D2,$T0);                  # d2 += h4*s3
1687 }
1688         &vpmuladd       (sub {  my $i=shift; &QWP(32*$i-128,"edx");     });
1689
1690 sub vlazy_reduction {
1691         ################################################################
1692         # lazy reduction
1693
1694          &vpsrlq        ($T0,$D3,26);
1695          &vpand         ($D3,$D3,$MASK);
1696         &vpsrlq         ($T1,$D0,26);
1697         &vpand          ($D0,$D0,$MASK);
1698          &vpaddq        ($D4,$D4,$T0);                  # h3 -> h4
1699         &vpaddq         ($D1,$D1,$T1);                  # h0 -> h1
1700          &vpsrlq        ($T0,$D4,26);
1701          &vpand         ($D4,$D4,$MASK);
1702         &vpsrlq         ($T1,$D1,26);
1703         &vpand          ($D1,$D1,$MASK);
1704         &vpaddq         ($D2,$D2,$T1);                  # h1 -> h2
1705          &vpaddd        ($D0,$D0,$T0);
1706          &vpsllq        ($T0,$T0,2);
1707         &vpsrlq         ($T1,$D2,26);
1708         &vpand          ($D2,$D2,$MASK);
1709          &vpaddd        ($D0,$D0,$T0);                  # h4 -> h0
1710         &vpaddd         ($D3,$D3,$T1);                  # h2 -> h3
1711         &vpsrlq         ($T1,$D3,26);
1712          &vpsrlq        ($T0,$D0,26);
1713          &vpand         ($D0,$D0,$MASK);
1714         &vpand          ($D3,$D3,$MASK);
1715          &vpaddd        ($D1,$D1,$T0);                  # h0 -> h1
1716         &vpaddd         ($D4,$D4,$T1);                  # h3 -> h4
1717 }
1718         &vlazy_reduction();
1719
1720         &vmovdqu        (&X($T0),&QWP(16*0,"esi"));     # load input
1721         &vmovdqu        (&X($T1),&QWP(16*1,"esi"));
1722         &vinserti128    ($T0,$T0,&QWP(16*2,"esi"),1);
1723         &vinserti128    ($T1,$T1,&QWP(16*3,"esi"),1);
1724         &lea            ("esi",&DWP(16*4,"esi"));
1725         &sub            ("ecx",64);
1726         &jnz            (&label("loop"));
1727
1728 &set_label("tail");
1729         &vsplat_input   ();
1730         &and            ("ebx",-64);                    # restore pointer
1731
1732         &vpmuladd       (sub {  my $i=shift; &QWP(4+32*$i-128,"edx");   });
1733
1734         &vlazy_reduction();
1735
1736         ################################################################
1737         # horizontal addition
1738
1739         &vpsrldq        ($T0,$D0,8);
1740         &vpsrldq        ($T1,$D1,8);
1741         &vpaddq         ($D0,$D0,$T0);
1742         &vpsrldq        ($T0,$D2,8);
1743         &vpaddq         ($D1,$D1,$T1);
1744         &vpsrldq        ($T1,$D3,8);
1745         &vpaddq         ($D2,$D2,$T0);
1746         &vpsrldq        ($T0,$D4,8);
1747         &vpaddq         ($D3,$D3,$T1);
1748         &vpermq         ($T1,$D0,2);                    # keep folding
1749         &vpaddq         ($D4,$D4,$T0);
1750         &vpermq         ($T0,$D1,2);
1751         &vpaddq         ($D0,$D0,$T1);
1752         &vpermq         ($T1,$D2,2);
1753         &vpaddq         ($D1,$D1,$T0);
1754         &vpermq         ($T0,$D3,2);
1755         &vpaddq         ($D2,$D2,$T1);
1756         &vpermq         ($T1,$D4,2);
1757         &vpaddq         ($D3,$D3,$T0);
1758         &vpaddq         ($D4,$D4,$T1);
1759
1760         &cmp            ("ecx",0);
1761         &je             (&label("done"));
1762
1763         ################################################################
1764         # clear all but single word
1765
1766         &vpshufd        (&X($D0),&X($D0),0b11111100);
1767         &lea            ("edx",&DWP(32*5+128,"esp"));   # restore pointer
1768         &vpshufd        (&X($D1),&X($D1),0b11111100);
1769         &vpshufd        (&X($D2),&X($D2),0b11111100);
1770         &vpshufd        (&X($D3),&X($D3),0b11111100);
1771         &vpshufd        (&X($D4),&X($D4),0b11111100);
1772         &jmp            (&label("even"));
1773
1774 &set_label("done",16);
1775         &vmovd          (&DWP(-16*3+4*0,"edi"),"xmm0"); # store hash value
1776         &vmovd          (&DWP(-16*3+4*1,"edi"),"xmm1");
1777         &vmovd          (&DWP(-16*3+4*2,"edi"),"xmm2");
1778         &vmovd          (&DWP(-16*3+4*3,"edi"),"xmm3");
1779         &vmovd          (&DWP(-16*3+4*4,"edi"),"xmm4");
1780         &vzeroupper     ();
1781 &set_label("nodata");
1782         &mov    ("esp","ebp");
1783 &function_end("_poly1305_blocks_avx2");
1784 }
1785 &set_label("const_sse2",64);
1786         &data_word(1<<24,0,     1<<24,0,        1<<24,0,        1<<24,0);
1787         &data_word(0,0,         0,0,            0,0,            0,0);
1788         &data_word(0x03ffffff,0,0x03ffffff,0,   0x03ffffff,0,   0x03ffffff,0);
1789         &data_word(0x0fffffff,0x0ffffffc,0x0ffffffc,0x0ffffffc);
1790 }
1791 &asciz  ("Poly1305 for x86, CRYPTOGAMS by <appro\@openssl.org>");
1792 &align  (4);
1793
1794 &asm_finish();