crypto/chacha/asm/chacha-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 # January 2015
  18 #
  19 # ChaCha20 for x86.
  20 #
  21 # Performance in cycles per byte out of large buffer.
  22 #
  23 #               1xIALU/gcc      4xSSSE3
  24 # Pentium       17.5/+80%
  25 # PIII          14.2/+60%
  26 # P4            18.6/+84%
  27 # Core2         9.56/+89%       4.83
  28 # Westmere      9.50/+45%       3.35
  29 # Sandy Bridge  10.5/+47%       3.20
  30 # Haswell       8.15/+50%       2.83
  31 # Silvermont    17.4/+36%       8.35
  32 # Sledgehammer  10.2/+54%
  33 # Bulldozer     13.4/+50%       4.38(*)
  34 #
  35 # (*)   Bulldozer actually executes 4xXOP code path that delivers 3.55;
  36
  37 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  38 push(@INC,"${dir}","${dir}../../perlasm");
  39 require "x86asm.pl";
  40
  41 $output=pop;
  42 open STDOUT,">$output";
  43
  44 &asm_init($ARGV[0],"chacha-x86.pl",$ARGV[$#ARGV] eq "386");
  45
  46 $xmm=$ymm=0;
  47 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
  48
  49 $ymm=1 if ($xmm &&
  50                 `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
  51                         =~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
  52                 $1>=2.19);      # first version supporting AVX
  53
  54 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32n" &&
  55                 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
  56                 $1>=2.03);      # first version supporting AVX
  57
  58 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32" &&
  59                 `ml 2>&1` =~ /Version ([0-9]+)\./ &&
  60                 $1>=10);        # first version supporting AVX
  61
  62 $ymm=1 if ($xmm && !$ymm &&
  63                 `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/ &&
  64                 $2>=3.0);       # first version supporting AVX
  65
  66 $a="eax";
  67 ($b,$b_)=("ebx","ebp");
  68 ($c,$c_)=("ecx","esi");
  69 ($d,$d_)=("edx","edi");
  70
  71 sub QUARTERROUND {
  72 my ($ai,$bi,$ci,$di,$i)=@_;
  73 my ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+1)&3),($ai,$bi,$ci,$di)); # next
  74 my ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-1)&3),($ai,$bi,$ci,$di)); # previous
  75
  76         #       a   b   c   d
  77         #
  78         #       0   4   8  12 < even round
  79         #       1   5   9  13
  80         #       2   6  10  14
  81         #       3   7  11  15
  82         #       0   5  10  15 < odd round
  83         #       1   6  11  12
  84         #       2   7   8  13
  85         #       3   4   9  14
  86
  87         if ($i==0) {
  88             my $j=4;
  89             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-$j--)&3),($ap,$bp,$cp,$dp));
  90         } elsif ($i==3) {
  91             my $j=0;
  92             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+$j++)&3),($an,$bn,$cn,$dn));
  93         } elsif ($i==4) {
  94             my $j=4;
  95             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_+$j--)&3),($ap,$bp,$cp,$dp));
  96         } elsif ($i==7) {
  97             my $j=0;
  98             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_-$j++)&3),($an,$bn,$cn,$dn));
  99         }
 100
 101         #&add   ($a,$b);                        # see elsewhere
 102         &xor    ($d,$a);
 103          &mov   (&DWP(4*$cp,"esp"),$c_)         if ($ai>0 && $ai<3);
 104         &rol    ($d,16);
 105          &mov   (&DWP(4*$bp,"esp"),$b_)         if ($i!=0);
 106         &add    ($c,$d);
 107          &mov   ($c_,&DWP(4*$cn,"esp"))         if ($ai>0 && $ai<3);
 108         &xor    ($b,$c);
 109          &mov   ($d_,&DWP(4*$dn,"esp"))         if ($di!=$dn);
 110         &rol    ($b,12);
 111          &mov   ($b_,&DWP(4*$bn,"esp"))         if ($i<7);
 112          &mov   ($b_,&DWP(128,"esp"))           if ($i==7);     # loop counter
 113         &add    ($a,$b);
 114         &xor    ($d,$a);
 115         &mov    (&DWP(4*$ai,"esp"),$a);
 116         &rol    ($d,8);
 117         &mov    ($a,&DWP(4*$an,"esp"));
 118         &add    ($c,$d);
 119         &mov    (&DWP(4*$di,"esp"),$d)          if ($di!=$dn);
 120         &mov    ($d_,$d)                        if ($di==$dn);
 121         &xor    ($b,$c);
 122          &add   ($a,$b_)                        if ($i<7);      # elsewhere
 123         &rol    ($b,7);
 124
 125         ($b,$b_)=($b_,$b);
 126         ($c,$c_)=($c_,$c);
 127         ($d,$d_)=($d_,$d);
 128 }
 129
 130 &static_label("ssse3_shortcut");
 131 &static_label("xop_shortcut");
 132 &static_label("ssse3_data");
 133 &static_label("pic_point");
 134
 135 &function_begin("ChaCha20_ctr32");
 136         &xor    ("eax","eax");
 137         &cmp    ("eax",&wparam(2));             # len==0?
 138         &je     (&label("no_data"));
 139 if ($xmm) {
 140         &call   (&label("pic_point"));
 141 &set_label("pic_point");
 142         &blindpop("eax");
 143         &picmeup("ebp","OPENSSL_ia32cap_P","eax",&label("pic_point"));
 144         &test   (&DWP(0,"ebp"),1<<24);          # test FXSR bit
 145         &jz     (&label("x86"));
 146         &test   (&DWP(4,"ebp"),1<<9);           # test SSSE3 bit
 147         &jz     (&label("x86"));
 148         &jmp    (&label("ssse3_shortcut"));
 149 &set_label("x86");
 150 }
 151         &mov    ("esi",&wparam(3));             # key
 152         &mov    ("edi",&wparam(4));             # counter and nonce
 153
 154         &stack_push(33);
 155
 156         &mov    ("eax",&DWP(4*0,"esi"));        # copy key
 157         &mov    ("ebx",&DWP(4*1,"esi"));
 158         &mov    ("ecx",&DWP(4*2,"esi"));
 159         &mov    ("edx",&DWP(4*3,"esi"));
 160         &mov    (&DWP(64+4*4,"esp"),"eax");
 161         &mov    (&DWP(64+4*5,"esp"),"ebx");
 162         &mov    (&DWP(64+4*6,"esp"),"ecx");
 163         &mov    (&DWP(64+4*7,"esp"),"edx");
 164         &mov    ("eax",&DWP(4*4,"esi"));
 165         &mov    ("ebx",&DWP(4*5,"esi"));
 166         &mov    ("ecx",&DWP(4*6,"esi"));
 167         &mov    ("edx",&DWP(4*7,"esi"));
 168         &mov    (&DWP(64+4*8,"esp"),"eax");
 169         &mov    (&DWP(64+4*9,"esp"),"ebx");
 170         &mov    (&DWP(64+4*10,"esp"),"ecx");
 171         &mov    (&DWP(64+4*11,"esp"),"edx");
 172         &mov    ("eax",&DWP(4*0,"edi"));        # copy counter and nonce
 173         &mov    ("ebx",&DWP(4*1,"edi"));
 174         &mov    ("ecx",&DWP(4*2,"edi"));
 175         &mov    ("edx",&DWP(4*3,"edi"));
 176         &sub    ("eax",1);
 177         &mov    (&DWP(64+4*12,"esp"),"eax");
 178         &mov    (&DWP(64+4*13,"esp"),"ebx");
 179         &mov    (&DWP(64+4*14,"esp"),"ecx");
 180         &mov    (&DWP(64+4*15,"esp"),"edx");
 181         &jmp    (&label("entry"));
 182
 183 &set_label("outer_loop",16);
 184         &mov    (&wparam(1),$b);                # save input
 185         &mov    (&wparam(0),$a);                # save output
 186         &mov    (&wparam(2),$c);                # save len
 187 &set_label("entry");
 188         &mov    ($a,0x61707865);
 189         &mov    (&DWP(4*1,"esp"),0x3320646e);
 190         &mov    (&DWP(4*2,"esp"),0x79622d32);
 191         &mov    (&DWP(4*3,"esp"),0x6b206574);
 192
 193         &mov    ($b, &DWP(64+4*5,"esp"));       # copy key material
 194         &mov    ($b_,&DWP(64+4*6,"esp"));
 195         &mov    ($c, &DWP(64+4*10,"esp"));
 196         &mov    ($c_,&DWP(64+4*11,"esp"));
 197         &mov    ($d, &DWP(64+4*13,"esp"));
 198         &mov    ($d_,&DWP(64+4*14,"esp"));
 199         &mov    (&DWP(4*5,"esp"),$b);
 200         &mov    (&DWP(4*6,"esp"),$b_);
 201         &mov    (&DWP(4*10,"esp"),$c);
 202         &mov    (&DWP(4*11,"esp"),$c_);
 203         &mov    (&DWP(4*13,"esp"),$d);
 204         &mov    (&DWP(4*14,"esp"),$d_);
 205
 206         &mov    ($b, &DWP(64+4*7,"esp"));
 207         &mov    ($d_,&DWP(64+4*15,"esp"));
 208         &mov    ($d, &DWP(64+4*12,"esp"));
 209         &mov    ($b_,&DWP(64+4*4,"esp"));
 210         &mov    ($c, &DWP(64+4*8,"esp"));
 211         &mov    ($c_,&DWP(64+4*9,"esp"));
 212         &add    ($d,1);                         # counter value
 213         &mov    (&DWP(4*7,"esp"),$b);
 214         &mov    (&DWP(4*15,"esp"),$d_);
 215         &mov    (&DWP(64+4*12,"esp"),$d);       # save counter value
 216
 217         &mov    ($b,10);                        # loop counter
 218         &jmp    (&label("loop"));
 219
 220 &set_label("loop",16);
 221         &add    ($a,$b_);                       # elsewhere
 222         &mov    (&DWP(128,"esp"),$b);           # save loop counter
 223         &mov    ($b,$b_);
 224         &QUARTERROUND(0, 4, 8, 12, 0);
 225         &QUARTERROUND(1, 5, 9, 13, 1);
 226         &QUARTERROUND(2, 6,10, 14, 2);
 227         &QUARTERROUND(3, 7,11, 15, 3);
 228         &QUARTERROUND(0, 5,10, 15, 4);
 229         &QUARTERROUND(1, 6,11, 12, 5);
 230         &QUARTERROUND(2, 7, 8, 13, 6);
 231         &QUARTERROUND(3, 4, 9, 14, 7);
 232         &dec    ($b);
 233         &jnz    (&label("loop"));
 234
 235         &mov    ($b,&wparam(2));                # load len
 236
 237         &add    ($a,0x61707865);                # accumulate key material
 238         &add    ($b_,&DWP(64+4*4,"esp"));
 239         &add    ($c, &DWP(64+4*8,"esp"));
 240         &add    ($c_,&DWP(64+4*9,"esp"));
 241
 242         &cmp    ($b,64);
 243         &jb     (&label("tail"));
 244
 245         &mov    ($b,&wparam(1));                # load input pointer
 246         &add    ($d, &DWP(64+4*12,"esp"));
 247         &add    ($d_,&DWP(64+4*14,"esp"));
 248
 249         &xor    ($a, &DWP(4*0,$b));             # xor with input
 250         &xor    ($b_,&DWP(4*4,$b));
 251         &mov    (&DWP(4*0,"esp"),$a);
 252         &mov    ($a,&wparam(0));                # load output pointer
 253         &xor    ($c, &DWP(4*8,$b));
 254         &xor    ($c_,&DWP(4*9,$b));
 255         &xor    ($d, &DWP(4*12,$b));
 256         &xor    ($d_,&DWP(4*14,$b));
 257         &mov    (&DWP(4*4,$a),$b_);             # write output
 258         &mov    (&DWP(4*8,$a),$c);
 259         &mov    (&DWP(4*9,$a),$c_);
 260         &mov    (&DWP(4*12,$a),$d);
 261         &mov    (&DWP(4*14,$a),$d_);
 262
 263         &mov    ($b_,&DWP(4*1,"esp"));
 264         &mov    ($c, &DWP(4*2,"esp"));
 265         &mov    ($c_,&DWP(4*3,"esp"));
 266         &mov    ($d, &DWP(4*5,"esp"));
 267         &mov    ($d_,&DWP(4*6,"esp"));
 268         &add    ($b_,0x3320646e);               # accumulate key material
 269         &add    ($c, 0x79622d32);
 270         &add    ($c_,0x6b206574);
 271         &add    ($d, &DWP(64+4*5,"esp"));
 272         &add    ($d_,&DWP(64+4*6,"esp"));
 273         &xor    ($b_,&DWP(4*1,$b));
 274         &xor    ($c, &DWP(4*2,$b));
 275         &xor    ($c_,&DWP(4*3,$b));
 276         &xor    ($d, &DWP(4*5,$b));
 277         &xor    ($d_,&DWP(4*6,$b));
 278         &mov    (&DWP(4*1,$a),$b_);
 279         &mov    (&DWP(4*2,$a),$c);
 280         &mov    (&DWP(4*3,$a),$c_);
 281         &mov    (&DWP(4*5,$a),$d);
 282         &mov    (&DWP(4*6,$a),$d_);
 283
 284         &mov    ($b_,&DWP(4*7,"esp"));
 285         &mov    ($c, &DWP(4*10,"esp"));
 286         &mov    ($c_,&DWP(4*11,"esp"));
 287         &mov    ($d, &DWP(4*13,"esp"));
 288         &mov    ($d_,&DWP(4*15,"esp"));
 289         &add    ($b_,&DWP(64+4*7,"esp"));
 290         &add    ($c, &DWP(64+4*10,"esp"));
 291         &add    ($c_,&DWP(64+4*11,"esp"));
 292         &add    ($d, &DWP(64+4*13,"esp"));
 293         &add    ($d_,&DWP(64+4*15,"esp"));
 294         &xor    ($b_,&DWP(4*7,$b));
 295         &xor    ($c, &DWP(4*10,$b));
 296         &xor    ($c_,&DWP(4*11,$b));
 297         &xor    ($d, &DWP(4*13,$b));
 298         &xor    ($d_,&DWP(4*15,$b));
 299         &lea    ($b,&DWP(4*16,$b));
 300         &mov    (&DWP(4*7,$a),$b_);
 301         &mov    ($b_,&DWP(4*0,"esp"));
 302         &mov    (&DWP(4*10,$a),$c);
 303         &mov    ($c,&wparam(2));                # len
 304         &mov    (&DWP(4*11,$a),$c_);
 305         &mov    (&DWP(4*13,$a),$d);
 306         &mov    (&DWP(4*15,$a),$d_);
 307         &mov    (&DWP(4*0,$a),$b_);
 308         &lea    ($a,&DWP(4*16,$a));
 309         &sub    ($c,64);
 310         &jnz    (&label("outer_loop"));
 311
 312         &jmp    (&label("done"));
 313
 314 &set_label("tail");
 315         &add    ($d, &DWP(64+4*12,"esp"));
 316         &add    ($d_,&DWP(64+4*14,"esp"));
 317         &mov    (&DWP(4*0,"esp"),$a);
 318         &mov    (&DWP(4*4,"esp"),$b_);
 319         &mov    (&DWP(4*8,"esp"),$c);
 320         &mov    (&DWP(4*9,"esp"),$c_);
 321         &mov    (&DWP(4*12,"esp"),$d);
 322         &mov    (&DWP(4*14,"esp"),$d_);
 323
 324         &mov    ($b_,&DWP(4*1,"esp"));
 325         &mov    ($c, &DWP(4*2,"esp"));
 326         &mov    ($c_,&DWP(4*3,"esp"));
 327         &mov    ($d, &DWP(4*5,"esp"));
 328         &mov    ($d_,&DWP(4*6,"esp"));
 329         &add    ($b_,0x3320646e);               # accumulate key material
 330         &add    ($c, 0x79622d32);
 331         &add    ($c_,0x6b206574);
 332         &add    ($d, &DWP(64+4*5,"esp"));
 333         &add    ($d_,&DWP(64+4*6,"esp"));
 334         &mov    (&DWP(4*1,"esp"),$b_);
 335         &mov    (&DWP(4*2,"esp"),$c);
 336         &mov    (&DWP(4*3,"esp"),$c_);
 337         &mov    (&DWP(4*5,"esp"),$d);
 338         &mov    (&DWP(4*6,"esp"),$d_);
 339
 340         &mov    ($b_,&DWP(4*7,"esp"));
 341         &mov    ($c, &DWP(4*10,"esp"));
 342         &mov    ($c_,&DWP(4*11,"esp"));
 343         &mov    ($d, &DWP(4*13,"esp"));
 344         &mov    ($d_,&DWP(4*15,"esp"));
 345         &add    ($b_,&DWP(64+4*7,"esp"));
 346         &add    ($c, &DWP(64+4*10,"esp"));
 347         &add    ($c_,&DWP(64+4*11,"esp"));
 348         &add    ($d, &DWP(64+4*13,"esp"));
 349         &add    ($d_,&DWP(64+4*15,"esp"));
 350         &mov    (&DWP(4*7,"esp"),$b_);
 351         &mov    ($b_,&wparam(1));               # load input
 352         &mov    (&DWP(4*10,"esp"),$c);
 353         &mov    ($c,&wparam(0));                # load output
 354         &mov    (&DWP(4*11,"esp"),$c_);
 355         &xor    ($c_,$c_);
 356         &mov    (&DWP(4*13,"esp"),$d);
 357         &mov    (&DWP(4*15,"esp"),$d_);
 358
 359         &xor    ("eax","eax");
 360         &xor    ("edx","edx");
 361 &set_label("tail_loop");
 362         &movb   ("al",&BP(0,$c_,$b_));
 363         &movb   ("dl",&BP(0,"esp",$c_));
 364         &lea    ($c_,&DWP(1,$c_));
 365         &xor    ("al","dl");
 366         &mov    (&BP(-1,$c,$c_),"al");
 367         &dec    ($b);
 368         &jnz    (&label("tail_loop"));
 369
 370 &set_label("done");
 371         &stack_pop(33);
 372 &set_label("no_data");
 373 &function_end("ChaCha20_ctr32");
 374
 375 if ($xmm) {
 376 my ($xa,$xa_,$xb,$xb_,$xc,$xc_,$xd,$xd_)=map("xmm$_",(0..7));
 377 my ($out,$inp,$len)=("edi","esi","ecx");
 378
 379 sub QUARTERROUND_SSSE3 {
 380 my ($ai,$bi,$ci,$di,$i)=@_;
 381 my ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+1)&3),($ai,$bi,$ci,$di)); # next
 382 my ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-1)&3),($ai,$bi,$ci,$di)); # previous
 383
 384         #       a   b   c   d
 385         #
 386         #       0   4   8  12 < even round
 387         #       1   5   9  13
 388         #       2   6  10  14
 389         #       3   7  11  15
 390         #       0   5  10  15 < odd round
 391         #       1   6  11  12
 392         #       2   7   8  13
 393         #       3   4   9  14
 394
 395         if ($i==0) {
 396             my $j=4;
 397             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-$j--)&3),($ap,$bp,$cp,$dp));
 398         } elsif ($i==3) {
 399             my $j=0;
 400             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+$j++)&3),($an,$bn,$cn,$dn));
 401         } elsif ($i==4) {
 402             my $j=4;
 403             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_+$j--)&3),($ap,$bp,$cp,$dp));
 404         } elsif ($i==7) {
 405             my $j=0;
 406             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_-$j++)&3),($an,$bn,$cn,$dn));
 407         }
 408
 409         #&paddd ($xa,$xb);                      # see elsewhere
 410         #&pxor  ($xd,$xa);                      # see elsewhere
 411          &movdqa(&QWP(16*$cp-128,"ebx"),$xc_)   if ($ai>0 && $ai<3);
 412         &pshufb ($xd,&QWP(0,"eax"));            # rot16
 413          &movdqa(&QWP(16*$bp-128,"ebx"),$xb_)   if ($i!=0);
 414         &paddd  ($xc,$xd);
 415          &movdqa($xc_,&QWP(16*$cn-128,"ebx"))   if ($ai>0 && $ai<3);
 416         &pxor   ($xb,$xc);
 417          &movdqa($xb_,&QWP(16*$bn-128,"ebx"))   if ($i<7);
 418         &movdqa ($xa_,$xb);                     # borrow as temporary
 419         &pslld  ($xb,12);
 420         &psrld  ($xa_,20);
 421         &por    ($xb,$xa_);
 422          &movdqa($xa_,&QWP(16*$an-128,"ebx"));
 423         &paddd  ($xa,$xb);
 424          &movdqa($xd_,&QWP(16*$dn-128,"ebx"))   if ($di!=$dn);
 425         &pxor   ($xd,$xa);
 426         &movdqa (&QWP(16*$ai-128,"ebx"),$xa);
 427         &pshufb ($xd,&QWP(16,"eax"));           # rot8
 428         &paddd  ($xc,$xd);
 429         &movdqa (&QWP(16*$di-128,"ebx"),$xd)    if ($di!=$dn);
 430         &movdqa ($xd_,$xd)                      if ($di==$dn);
 431         &pxor   ($xb,$xc);
 432          &paddd ($xa_,$xb_)                     if ($i<7);      # elsewhere
 433         &movdqa ($xa,$xb);                      # borrow as temporary
 434         &pslld  ($xb,7);
 435         &psrld  ($xa,25);
 436          &pxor  ($xd_,$xa_)                     if ($i<7);      # elsewhere
 437         &por    ($xb,$xa);
 438
 439         ($xa,$xa_)=($xa_,$xa);
 440         ($xb,$xb_)=($xb_,$xb);
 441         ($xc,$xc_)=($xc_,$xc);
 442         ($xd,$xd_)=($xd_,$xd);
 443 }
 444
 445 &function_begin("ChaCha20_ssse3");
 446 &set_label("ssse3_shortcut");
 447 if ($ymm) {
 448         &test           (&DWP(4,"ebp"),1<<11);          # test XOP bit
 449         &jnz            (&label("xop_shortcut"));
 450 }
 451
 452         &mov            ($out,&wparam(0));
 453         &mov            ($inp,&wparam(1));
 454         &mov            ($len,&wparam(2));
 455         &mov            ("edx",&wparam(3));             # key
 456         &mov            ("ebx",&wparam(4));             # counter and nonce
 457
 458         &mov            ("ebp","esp");
 459         &stack_push     (131);
 460         &and            ("esp",-64);
 461         &mov            (&DWP(512,"esp"),"ebp");
 462
 463         &lea            ("eax",&DWP(&label("ssse3_data")."-".
 464                                     &label("pic_point"),"eax"));
 465         &movdqu         ("xmm3",&QWP(0,"ebx"));         # counter and nonce
 466
 467         &cmp            ($len,64*4);
 468         &jb             (&label("1x"));
 469
 470         &mov            (&DWP(512+4,"esp"),"edx");      # offload pointers
 471         &mov            (&DWP(512+8,"esp"),"ebx");
 472         &sub            ($len,64*4);                    # bias len
 473         &lea            ("ebp",&DWP(256+128,"esp"));    # size optimization
 474
 475         &movdqu         ("xmm7",&QWP(0,"edx"));         # key
 476         &pshufd         ("xmm0","xmm3",0x00);
 477         &pshufd         ("xmm1","xmm3",0x55);
 478         &pshufd         ("xmm2","xmm3",0xaa);
 479         &pshufd         ("xmm3","xmm3",0xff);
 480          &paddd         ("xmm0",&QWP(16*3,"eax"));      # fix counters
 481         &pshufd         ("xmm4","xmm7",0x00);
 482         &pshufd         ("xmm5","xmm7",0x55);
 483          &psubd         ("xmm0",&QWP(16*4,"eax"));
 484         &pshufd         ("xmm6","xmm7",0xaa);
 485         &pshufd         ("xmm7","xmm7",0xff);
 486         &movdqa         (&QWP(16*12-128,"ebp"),"xmm0");
 487         &movdqa         (&QWP(16*13-128,"ebp"),"xmm1");
 488         &movdqa         (&QWP(16*14-128,"ebp"),"xmm2");
 489         &movdqa         (&QWP(16*15-128,"ebp"),"xmm3");
 490          &movdqu        ("xmm3",&QWP(16,"edx"));        # key
 491         &movdqa         (&QWP(16*4-128,"ebp"),"xmm4");
 492         &movdqa         (&QWP(16*5-128,"ebp"),"xmm5");
 493         &movdqa         (&QWP(16*6-128,"ebp"),"xmm6");
 494         &movdqa         (&QWP(16*7-128,"ebp"),"xmm7");
 495          &movdqa        ("xmm7",&QWP(16*2,"eax"));      # sigma
 496          &lea           ("ebx",&DWP(128,"esp"));        # size optimization
 497
 498         &pshufd         ("xmm0","xmm3",0x00);
 499         &pshufd         ("xmm1","xmm3",0x55);
 500         &pshufd         ("xmm2","xmm3",0xaa);
 501         &pshufd         ("xmm3","xmm3",0xff);
 502         &pshufd         ("xmm4","xmm7",0x00);
 503         &pshufd         ("xmm5","xmm7",0x55);
 504         &pshufd         ("xmm6","xmm7",0xaa);
 505         &pshufd         ("xmm7","xmm7",0xff);
 506         &movdqa         (&QWP(16*8-128,"ebp"),"xmm0");
 507         &movdqa         (&QWP(16*9-128,"ebp"),"xmm1");
 508         &movdqa         (&QWP(16*10-128,"ebp"),"xmm2");
 509         &movdqa         (&QWP(16*11-128,"ebp"),"xmm3");
 510         &movdqa         (&QWP(16*0-128,"ebp"),"xmm4");
 511         &movdqa         (&QWP(16*1-128,"ebp"),"xmm5");
 512         &movdqa         (&QWP(16*2-128,"ebp"),"xmm6");
 513         &movdqa         (&QWP(16*3-128,"ebp"),"xmm7");
 514
 515         &lea            ($inp,&DWP(128,$inp));          # size optimization
 516         &lea            ($out,&DWP(128,$out));          # size optimization
 517         &jmp            (&label("outer_loop"));
 518
 519 &set_label("outer_loop",16);
 520         #&movdqa        ("xmm0",&QWP(16*0-128,"ebp"));  # copy key material
 521         &movdqa         ("xmm1",&QWP(16*1-128,"ebp"));
 522         &movdqa         ("xmm2",&QWP(16*2-128,"ebp"));
 523         &movdqa         ("xmm3",&QWP(16*3-128,"ebp"));
 524         #&movdqa        ("xmm4",&QWP(16*4-128,"ebp"));
 525         &movdqa         ("xmm5",&QWP(16*5-128,"ebp"));
 526         &movdqa         ("xmm6",&QWP(16*6-128,"ebp"));
 527         &movdqa         ("xmm7",&QWP(16*7-128,"ebp"));
 528         #&movdqa        (&QWP(16*0-128,"ebx"),"xmm0");
 529         &movdqa         (&QWP(16*1-128,"ebx"),"xmm1");
 530         &movdqa         (&QWP(16*2-128,"ebx"),"xmm2");
 531         &movdqa         (&QWP(16*3-128,"ebx"),"xmm3");
 532         #&movdqa        (&QWP(16*4-128,"ebx"),"xmm4");
 533         &movdqa         (&QWP(16*5-128,"ebx"),"xmm5");
 534         &movdqa         (&QWP(16*6-128,"ebx"),"xmm6");
 535         &movdqa         (&QWP(16*7-128,"ebx"),"xmm7");
 536         #&movdqa        ("xmm0",&QWP(16*8-128,"ebp"));
 537         #&movdqa        ("xmm1",&QWP(16*9-128,"ebp"));
 538         &movdqa         ("xmm2",&QWP(16*10-128,"ebp"));
 539         &movdqa         ("xmm3",&QWP(16*11-128,"ebp"));
 540         &movdqa         ("xmm4",&QWP(16*12-128,"ebp"));
 541         &movdqa         ("xmm5",&QWP(16*13-128,"ebp"));
 542         &movdqa         ("xmm6",&QWP(16*14-128,"ebp"));
 543         &movdqa         ("xmm7",&QWP(16*15-128,"ebp"));
 544         &paddd          ("xmm4",&QWP(16*4,"eax"));      # counter value
 545         #&movdqa        (&QWP(16*8-128,"ebx"),"xmm0");
 546         #&movdqa        (&QWP(16*9-128,"ebx"),"xmm1");
 547         &movdqa         (&QWP(16*10-128,"ebx"),"xmm2");
 548         &movdqa         (&QWP(16*11-128,"ebx"),"xmm3");
 549         &movdqa         (&QWP(16*12-128,"ebx"),"xmm4");
 550         &movdqa         (&QWP(16*13-128,"ebx"),"xmm5");
 551         &movdqa         (&QWP(16*14-128,"ebx"),"xmm6");
 552         &movdqa         (&QWP(16*15-128,"ebx"),"xmm7");
 553         &movdqa         (&QWP(16*12-128,"ebp"),"xmm4"); # save counter value
 554
 555         &movdqa         ($xa, &QWP(16*0-128,"ebp"));
 556         &movdqa         ($xd, "xmm4");
 557         &movdqa         ($xb_,&QWP(16*4-128,"ebp"));
 558         &movdqa         ($xc, &QWP(16*8-128,"ebp"));
 559         &movdqa         ($xc_,&QWP(16*9-128,"ebp"));
 560
 561         &mov            ("edx",10);                     # loop counter
 562         &nop            ();
 563
 564 &set_label("loop",16);
 565         &paddd          ($xa,$xb_);                     # elsewhere
 566         &movdqa         ($xb,$xb_);
 567         &pxor           ($xd,$xa);                      # elsewhere
 568         &QUARTERROUND_SSSE3(0, 4, 8, 12, 0);
 569         &QUARTERROUND_SSSE3(1, 5, 9, 13, 1);
 570         &QUARTERROUND_SSSE3(2, 6,10, 14, 2);
 571         &QUARTERROUND_SSSE3(3, 7,11, 15, 3);
 572         &QUARTERROUND_SSSE3(0, 5,10, 15, 4);
 573         &QUARTERROUND_SSSE3(1, 6,11, 12, 5);
 574         &QUARTERROUND_SSSE3(2, 7, 8, 13, 6);
 575         &QUARTERROUND_SSSE3(3, 4, 9, 14, 7);
 576         &dec            ("edx");
 577         &jnz            (&label("loop"));
 578
 579         &movdqa         (&QWP(16*4-128,"ebx"),$xb_);
 580         &movdqa         (&QWP(16*8-128,"ebx"),$xc);
 581         &movdqa         (&QWP(16*9-128,"ebx"),$xc_);
 582         &movdqa         (&QWP(16*12-128,"ebx"),$xd);
 583         &movdqa         (&QWP(16*14-128,"ebx"),$xd_);
 584
 585     my ($xa0,$xa1,$xa2,$xa3,$xt0,$xt1,$xt2,$xt3)=map("xmm$_",(0..7));
 586
 587         #&movdqa        ($xa0,&QWP(16*0-128,"ebx"));    # it's there
 588         &movdqa         ($xa1,&QWP(16*1-128,"ebx"));
 589         &movdqa         ($xa2,&QWP(16*2-128,"ebx"));
 590         &movdqa         ($xa3,&QWP(16*3-128,"ebx"));
 591
 592     for($i=0;$i<256;$i+=64) {
 593         &paddd          ($xa0,&QWP($i+16*0-128,"ebp")); # accumulate key material
 594         &paddd          ($xa1,&QWP($i+16*1-128,"ebp"));
 595         &paddd          ($xa2,&QWP($i+16*2-128,"ebp"));
 596         &paddd          ($xa3,&QWP($i+16*3-128,"ebp"));
 597
 598         &movdqa         ($xt2,$xa0);            # "de-interlace" data
 599         &punpckldq      ($xa0,$xa1);
 600         &movdqa         ($xt3,$xa2);
 601         &punpckldq      ($xa2,$xa3);
 602         &punpckhdq      ($xt2,$xa1);
 603         &punpckhdq      ($xt3,$xa3);
 604         &movdqa         ($xa1,$xa0);
 605         &punpcklqdq     ($xa0,$xa2);            # "a0"
 606         &movdqa         ($xa3,$xt2);
 607         &punpcklqdq     ($xt2,$xt3);            # "a2"
 608         &punpckhqdq     ($xa1,$xa2);            # "a1"
 609         &punpckhqdq     ($xa3,$xt3);            # "a3"
 610
 611         #($xa2,$xt2)=($xt2,$xa2);
 612
 613         &movdqu         ($xt0,&QWP(64*0-128,$inp));     # load input
 614         &movdqu         ($xt1,&QWP(64*1-128,$inp));
 615         &movdqu         ($xa2,&QWP(64*2-128,$inp));
 616         &movdqu         ($xt3,&QWP(64*3-128,$inp));
 617         &lea            ($inp,&QWP($i<192?16:(64*4-16*3),$inp));
 618         &pxor           ($xt0,$xa0);
 619         &movdqa         ($xa0,&QWP($i+16*4-128,"ebx"))  if ($i<192);
 620         &pxor           ($xt1,$xa1);
 621         &movdqa         ($xa1,&QWP($i+16*5-128,"ebx"))  if ($i<192);
 622         &pxor           ($xt2,$xa2);
 623         &movdqa         ($xa2,&QWP($i+16*6-128,"ebx"))  if ($i<192);
 624         &pxor           ($xt3,$xa3);
 625         &movdqa         ($xa3,&QWP($i+16*7-128,"ebx"))  if ($i<192);
 626         &movdqu         (&QWP(64*0-128,$out),$xt0);     # store output
 627         &movdqu         (&QWP(64*1-128,$out),$xt1);
 628         &movdqu         (&QWP(64*2-128,$out),$xt2);
 629         &movdqu         (&QWP(64*3-128,$out),$xt3);
 630         &lea            ($out,&QWP($i<192?16:(64*4-16*3),$out));
 631     }
 632         &sub            ($len,64*4);
 633         &jnc            (&label("outer_loop"));
 634
 635         &add            ($len,64*4);
 636         &jz             (&label("done"));
 637
 638         &mov            ("ebx",&DWP(512+8,"esp"));      # restore pointers
 639         &lea            ($inp,&DWP(-128,$inp));
 640         &mov            ("edx",&DWP(512+4,"esp"));
 641         &lea            ($out,&DWP(-128,$out));
 642
 643         &movd           ("xmm2",&DWP(16*12-128,"ebp")); # counter value
 644         &movdqu         ("xmm3",&QWP(0,"ebx"));
 645         &paddd          ("xmm2",&QWP(16*6,"eax"));      # +four
 646         &pand           ("xmm3",&QWP(16*7,"eax"));
 647         &por            ("xmm3","xmm2");                # counter value
 648 {
 649 my ($a,$b,$c,$d,$t,$t1,$rot16,$rot24)=map("xmm$_",(0..7));
 650
 651 sub SSSE3ROUND {        # critical path is 20 "SIMD ticks" per round
 652         &paddd          ($a,$b);
 653         &pxor           ($d,$a);
 654         &pshufb         ($d,$rot16);
 655
 656         &paddd          ($c,$d);
 657         &pxor           ($b,$c);
 658         &movdqa         ($t,$b);
 659         &psrld          ($b,20);
 660         &pslld          ($t,12);
 661         &por            ($b,$t);
 662
 663         &paddd          ($a,$b);
 664         &pxor           ($d,$a);
 665         &pshufb         ($d,$rot24);
 666
 667         &paddd          ($c,$d);
 668         &pxor           ($b,$c);
 669         &movdqa         ($t,$b);
 670         &psrld          ($b,25);
 671         &pslld          ($t,7);
 672         &por            ($b,$t);
 673 }
 674
 675 &set_label("1x");
 676         &movdqa         ($a,&QWP(16*2,"eax"));          # sigma
 677         &movdqu         ($b,&QWP(0,"edx"));
 678         &movdqu         ($c,&QWP(16,"edx"));
 679         #&movdqu        ($d,&QWP(0,"ebx"));             # already loaded
 680         &movdqa         ($rot16,&QWP(0,"eax"));
 681         &movdqa         ($rot24,&QWP(16,"eax"));
 682         &mov            (&DWP(16*3,"esp"),"ebp");
 683
 684         &movdqa         (&QWP(16*0,"esp"),$a);
 685         &movdqa         (&QWP(16*1,"esp"),$b);
 686         &movdqa         (&QWP(16*2,"esp"),$c);
 687         &movdqa         (&QWP(16*3,"esp"),$d);
 688         &mov            ("edx",10);
 689         &jmp            (&label("loop1x"));
 690
 691 &set_label("outer1x",16);
 692         &movdqa         ($d,&QWP(16*5,"eax"));          # one
 693         &movdqa         ($a,&QWP(16*0,"esp"));
 694         &movdqa         ($b,&QWP(16*1,"esp"));
 695         &movdqa         ($c,&QWP(16*2,"esp"));
 696         &paddd          ($d,&QWP(16*3,"esp"));
 697         &mov            ("edx",10);
 698         &movdqa         (&QWP(16*3,"esp"),$d);
 699         &jmp            (&label("loop1x"));
 700
 701 &set_label("loop1x",16);
 702         &SSSE3ROUND();
 703         &pshufd ($c,$c,0b01001110);
 704         &pshufd ($b,$b,0b00111001);
 705         &pshufd ($d,$d,0b10010011);
 706         &nop    ();
 707
 708         &SSSE3ROUND();
 709         &pshufd ($c,$c,0b01001110);
 710         &pshufd ($b,$b,0b10010011);
 711         &pshufd ($d,$d,0b00111001);
 712
 713         &dec            ("edx");
 714         &jnz            (&label("loop1x"));
 715
 716         &paddd          ($a,&QWP(16*0,"esp"));
 717         &paddd          ($b,&QWP(16*1,"esp"));
 718         &paddd          ($c,&QWP(16*2,"esp"));
 719         &paddd          ($d,&QWP(16*3,"esp"));
 720
 721         &cmp            ($len,64);
 722         &jb             (&label("tail"));
 723
 724         &movdqu         ($t,&QWP(16*0,$inp));
 725         &movdqu         ($t1,&QWP(16*1,$inp));
 726         &pxor           ($a,$t);                # xor with input
 727         &movdqu         ($t,&QWP(16*2,$inp));
 728         &pxor           ($b,$t1);
 729         &movdqu         ($t1,&QWP(16*3,$inp));
 730         &pxor           ($c,$t);
 731         &pxor           ($d,$t1);
 732         &lea            ($inp,&DWP(16*4,$inp)); # inp+=64
 733
 734         &movdqu         (&QWP(16*0,$out),$a);   # write output
 735         &movdqu         (&QWP(16*1,$out),$b);
 736         &movdqu         (&QWP(16*2,$out),$c);
 737         &movdqu         (&QWP(16*3,$out),$d);
 738         &lea            ($out,&DWP(16*4,$out)); # inp+=64
 739
 740         &sub            ($len,64);
 741         &jnz            (&label("outer1x"));
 742
 743         &jmp            (&label("done"));
 744
 745 &set_label("tail");
 746         &movdqa         (&QWP(16*0,"esp"),$a);
 747         &movdqa         (&QWP(16*1,"esp"),$b);
 748         &movdqa         (&QWP(16*2,"esp"),$c);
 749         &movdqa         (&QWP(16*3,"esp"),$d);
 750
 751         &xor            ("eax","eax");
 752         &xor            ("edx","edx");
 753         &xor            ("ebp","ebp");
 754
 755 &set_label("tail_loop");
 756         &movb           ("al",&BP(0,"esp","ebp"));
 757         &movb           ("dl",&BP(0,$inp,"ebp"));
 758         &lea            ("ebp",&DWP(1,"ebp"));
 759         &xor            ("al","dl");
 760         &movb           (&BP(-1,$out,"ebp"),"al");
 761         &dec            ($len);
 762         &jnz            (&label("tail_loop"));
 763 }
 764 &set_label("done");
 765         &mov            ("esp",&DWP(512,"esp"));
 766 &function_end("ChaCha20_ssse3");
 767
 768 &align  (64);
 769 &set_label("ssse3_data");
 770 &data_byte(0x2,0x3,0x0,0x1, 0x6,0x7,0x4,0x5, 0xa,0xb,0x8,0x9, 0xe,0xf,0xc,0xd);
 771 &data_byte(0x3,0x0,0x1,0x2, 0x7,0x4,0x5,0x6, 0xb,0x8,0x9,0xa, 0xf,0xc,0xd,0xe);
 772 &data_word(0x61707865,0x3320646e,0x79622d32,0x6b206574);
 773 &data_word(0,1,2,3);
 774 &data_word(4,4,4,4);
 775 &data_word(1,0,0,0);
 776 &data_word(4,0,0,0);
 777 &data_word(0,-1,-1,-1);
 778 &align  (64);
 779 }
 780 &asciz  ("ChaCha20 for x86, CRYPTOGAMS by <appro\@openssl.org>");
 781
 782 if ($ymm) {
 783 my ($xa,$xa_,$xb,$xb_,$xc,$xc_,$xd,$xd_)=map("xmm$_",(0..7));
 784 my ($out,$inp,$len)=("edi","esi","ecx");
 785
 786 sub QUARTERROUND_XOP {
 787 my ($ai,$bi,$ci,$di,$i)=@_;
 788 my ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+1)&3),($ai,$bi,$ci,$di)); # next
 789 my ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-1)&3),($ai,$bi,$ci,$di)); # previous
 790
 791         #       a   b   c   d
 792         #
 793         #       0   4   8  12 < even round
 794         #       1   5   9  13
 795         #       2   6  10  14
 796         #       3   7  11  15
 797         #       0   5  10  15 < odd round
 798         #       1   6  11  12
 799         #       2   7   8  13
 800         #       3   4   9  14
 801
 802         if ($i==0) {
 803             my $j=4;
 804             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-$j--)&3),($ap,$bp,$cp,$dp));
 805         } elsif ($i==3) {
 806             my $j=0;
 807             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+$j++)&3),($an,$bn,$cn,$dn));
 808         } elsif ($i==4) {
 809             my $j=4;
 810             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_+$j--)&3),($ap,$bp,$cp,$dp));
 811         } elsif ($i==7) {
 812             my $j=0;
 813             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_-$j++)&3),($an,$bn,$cn,$dn));
 814         }
 815
 816         #&vpaddd        ($xa,$xa,$xb);                  # see elsewhere
 817         #&vpxor         ($xd,$xd,$xa);                  # see elsewhere
 818          &vmovdqa       (&QWP(16*$cp-128,"ebx"),$xc_)   if ($ai>0 && $ai<3);
 819         &vprotd         ($xd,$xd,16);
 820          &vmovdqa       (&QWP(16*$bp-128,"ebx"),$xb_)   if ($i!=0);
 821         &vpaddd         ($xc,$xc,$xd);
 822          &vmovdqa       ($xc_,&QWP(16*$cn-128,"ebx"))   if ($ai>0 && $ai<3);
 823         &vpxor          ($xb,$i!=0?$xb:$xb_,$xc);
 824          &vmovdqa       ($xa_,&QWP(16*$an-128,"ebx"));
 825         &vprotd         ($xb,$xb,12);
 826          &vmovdqa       ($xb_,&QWP(16*$bn-128,"ebx"))   if ($i<7);
 827         &vpaddd         ($xa,$xa,$xb);
 828          &vmovdqa       ($xd_,&QWP(16*$dn-128,"ebx"))   if ($di!=$dn);
 829         &vpxor          ($xd,$xd,$xa);
 830          &vpaddd        ($xa_,$xa_,$xb_)                if ($i<7);      # elsewhere
 831         &vprotd         ($xd,$xd,8);
 832         &vmovdqa        (&QWP(16*$ai-128,"ebx"),$xa);
 833         &vpaddd         ($xc,$xc,$xd);
 834         &vmovdqa        (&QWP(16*$di-128,"ebx"),$xd)    if ($di!=$dn);
 835         &vpxor          ($xb,$xb,$xc);
 836          &vpxor         ($xd_,$di==$dn?$xd:$xd_,$xa_)   if ($i<7);      # elsewhere
 837         &vprotd         ($xb,$xb,7);
 838
 839         ($xa,$xa_)=($xa_,$xa);
 840         ($xb,$xb_)=($xb_,$xb);
 841         ($xc,$xc_)=($xc_,$xc);
 842         ($xd,$xd_)=($xd_,$xd);
 843 }
 844
 845 &function_begin("ChaCha20_xop");
 846 &set_label("xop_shortcut");
 847         &mov            ($out,&wparam(0));
 848         &mov            ($inp,&wparam(1));
 849         &mov            ($len,&wparam(2));
 850         &mov            ("edx",&wparam(3));             # key
 851         &mov            ("ebx",&wparam(4));             # counter and nonce
 852         &vzeroupper     ();
 853
 854         &mov            ("ebp","esp");
 855         &stack_push     (131);
 856         &and            ("esp",-64);
 857         &mov            (&DWP(512,"esp"),"ebp");
 858
 859         &lea            ("eax",&DWP(&label("ssse3_data")."-".
 860                                     &label("pic_point"),"eax"));
 861         &vmovdqu        ("xmm3",&QWP(0,"ebx"));         # counter and nonce
 862
 863         &cmp            ($len,64*4);
 864         &jb             (&label("1x"));
 865
 866         &mov            (&DWP(512+4,"esp"),"edx");      # offload pointers
 867         &mov            (&DWP(512+8,"esp"),"ebx");
 868         &sub            ($len,64*4);                    # bias len
 869         &lea            ("ebp",&DWP(256+128,"esp"));    # size optimization
 870
 871         &vmovdqu        ("xmm7",&QWP(0,"edx"));         # key
 872         &vpshufd        ("xmm0","xmm3",0x00);
 873         &vpshufd        ("xmm1","xmm3",0x55);
 874         &vpshufd        ("xmm2","xmm3",0xaa);
 875         &vpshufd        ("xmm3","xmm3",0xff);
 876          &vpaddd        ("xmm0","xmm0",&QWP(16*3,"eax"));       # fix counters
 877         &vpshufd        ("xmm4","xmm7",0x00);
 878         &vpshufd        ("xmm5","xmm7",0x55);
 879          &vpsubd        ("xmm0","xmm0",&QWP(16*4,"eax"));
 880         &vpshufd        ("xmm6","xmm7",0xaa);
 881         &vpshufd        ("xmm7","xmm7",0xff);
 882         &vmovdqa        (&QWP(16*12-128,"ebp"),"xmm0");
 883         &vmovdqa        (&QWP(16*13-128,"ebp"),"xmm1");
 884         &vmovdqa        (&QWP(16*14-128,"ebp"),"xmm2");
 885         &vmovdqa        (&QWP(16*15-128,"ebp"),"xmm3");
 886          &vmovdqu       ("xmm3",&QWP(16,"edx"));        # key
 887         &vmovdqa        (&QWP(16*4-128,"ebp"),"xmm4");
 888         &vmovdqa        (&QWP(16*5-128,"ebp"),"xmm5");
 889         &vmovdqa        (&QWP(16*6-128,"ebp"),"xmm6");
 890         &vmovdqa        (&QWP(16*7-128,"ebp"),"xmm7");
 891          &vmovdqa       ("xmm7",&QWP(16*2,"eax"));      # sigma
 892          &lea           ("ebx",&DWP(128,"esp"));        # size optimization
 893
 894         &vpshufd        ("xmm0","xmm3",0x00);
 895         &vpshufd        ("xmm1","xmm3",0x55);
 896         &vpshufd        ("xmm2","xmm3",0xaa);
 897         &vpshufd        ("xmm3","xmm3",0xff);
 898         &vpshufd        ("xmm4","xmm7",0x00);
 899         &vpshufd        ("xmm5","xmm7",0x55);
 900         &vpshufd        ("xmm6","xmm7",0xaa);
 901         &vpshufd        ("xmm7","xmm7",0xff);
 902         &vmovdqa        (&QWP(16*8-128,"ebp"),"xmm0");
 903         &vmovdqa        (&QWP(16*9-128,"ebp"),"xmm1");
 904         &vmovdqa        (&QWP(16*10-128,"ebp"),"xmm2");
 905         &vmovdqa        (&QWP(16*11-128,"ebp"),"xmm3");
 906         &vmovdqa        (&QWP(16*0-128,"ebp"),"xmm4");
 907         &vmovdqa        (&QWP(16*1-128,"ebp"),"xmm5");
 908         &vmovdqa        (&QWP(16*2-128,"ebp"),"xmm6");
 909         &vmovdqa        (&QWP(16*3-128,"ebp"),"xmm7");
 910
 911         &lea            ($inp,&DWP(128,$inp));          # size optimization
 912         &lea            ($out,&DWP(128,$out));          # size optimization
 913         &jmp            (&label("outer_loop"));
 914
 915 &set_label("outer_loop",32);
 916         #&vmovdqa       ("xmm0",&QWP(16*0-128,"ebp"));  # copy key material
 917         &vmovdqa        ("xmm1",&QWP(16*1-128,"ebp"));
 918         &vmovdqa        ("xmm2",&QWP(16*2-128,"ebp"));
 919         &vmovdqa        ("xmm3",&QWP(16*3-128,"ebp"));
 920         #&vmovdqa       ("xmm4",&QWP(16*4-128,"ebp"));
 921         &vmovdqa        ("xmm5",&QWP(16*5-128,"ebp"));
 922         &vmovdqa        ("xmm6",&QWP(16*6-128,"ebp"));
 923         &vmovdqa        ("xmm7",&QWP(16*7-128,"ebp"));
 924         #&vmovdqa       (&QWP(16*0-128,"ebx"),"xmm0");
 925         &vmovdqa        (&QWP(16*1-128,"ebx"),"xmm1");
 926         &vmovdqa        (&QWP(16*2-128,"ebx"),"xmm2");
 927         &vmovdqa        (&QWP(16*3-128,"ebx"),"xmm3");
 928         #&vmovdqa       (&QWP(16*4-128,"ebx"),"xmm4");
 929         &vmovdqa        (&QWP(16*5-128,"ebx"),"xmm5");
 930         &vmovdqa        (&QWP(16*6-128,"ebx"),"xmm6");
 931         &vmovdqa        (&QWP(16*7-128,"ebx"),"xmm7");
 932         #&vmovdqa       ("xmm0",&QWP(16*8-128,"ebp"));
 933         #&vmovdqa       ("xmm1",&QWP(16*9-128,"ebp"));
 934         &vmovdqa        ("xmm2",&QWP(16*10-128,"ebp"));
 935         &vmovdqa        ("xmm3",&QWP(16*11-128,"ebp"));
 936         &vmovdqa        ("xmm4",&QWP(16*12-128,"ebp"));
 937         &vmovdqa        ("xmm5",&QWP(16*13-128,"ebp"));
 938         &vmovdqa        ("xmm6",&QWP(16*14-128,"ebp"));
 939         &vmovdqa        ("xmm7",&QWP(16*15-128,"ebp"));
 940         &vpaddd         ("xmm4","xmm4",&QWP(16*4,"eax"));       # counter value
 941         #&vmovdqa       (&QWP(16*8-128,"ebx"),"xmm0");
 942         #&vmovdqa       (&QWP(16*9-128,"ebx"),"xmm1");
 943         &vmovdqa        (&QWP(16*10-128,"ebx"),"xmm2");
 944         &vmovdqa        (&QWP(16*11-128,"ebx"),"xmm3");
 945         &vmovdqa        (&QWP(16*12-128,"ebx"),"xmm4");
 946         &vmovdqa        (&QWP(16*13-128,"ebx"),"xmm5");
 947         &vmovdqa        (&QWP(16*14-128,"ebx"),"xmm6");
 948         &vmovdqa        (&QWP(16*15-128,"ebx"),"xmm7");
 949         &vmovdqa        (&QWP(16*12-128,"ebp"),"xmm4"); # save counter value
 950
 951         &vmovdqa        ($xa, &QWP(16*0-128,"ebp"));
 952         &vmovdqa        ($xd, "xmm4");
 953         &vmovdqa        ($xb_,&QWP(16*4-128,"ebp"));
 954         &vmovdqa        ($xc, &QWP(16*8-128,"ebp"));
 955         &vmovdqa        ($xc_,&QWP(16*9-128,"ebp"));
 956
 957         &mov            ("edx",10);                     # loop counter
 958         &nop            ();
 959
 960 &set_label("loop",32);
 961         &vpaddd         ($xa,$xa,$xb_);                 # elsewhere
 962         &vpxor          ($xd,$xd,$xa);                  # elsewhere
 963         &QUARTERROUND_XOP(0, 4, 8, 12, 0);
 964         &QUARTERROUND_XOP(1, 5, 9, 13, 1);
 965         &QUARTERROUND_XOP(2, 6,10, 14, 2);
 966         &QUARTERROUND_XOP(3, 7,11, 15, 3);
 967         &QUARTERROUND_XOP(0, 5,10, 15, 4);
 968         &QUARTERROUND_XOP(1, 6,11, 12, 5);
 969         &QUARTERROUND_XOP(2, 7, 8, 13, 6);
 970         &QUARTERROUND_XOP(3, 4, 9, 14, 7);
 971         &dec            ("edx");
 972         &jnz            (&label("loop"));
 973
 974         &vmovdqa        (&QWP(16*4-128,"ebx"),$xb_);
 975         &vmovdqa        (&QWP(16*8-128,"ebx"),$xc);
 976         &vmovdqa        (&QWP(16*9-128,"ebx"),$xc_);
 977         &vmovdqa        (&QWP(16*12-128,"ebx"),$xd);
 978         &vmovdqa        (&QWP(16*14-128,"ebx"),$xd_);
 979
 980     my ($xa0,$xa1,$xa2,$xa3,$xt0,$xt1,$xt2,$xt3)=map("xmm$_",(0..7));
 981
 982         #&vmovdqa       ($xa0,&QWP(16*0-128,"ebx"));    # it's there
 983         &vmovdqa        ($xa1,&QWP(16*1-128,"ebx"));
 984         &vmovdqa        ($xa2,&QWP(16*2-128,"ebx"));
 985         &vmovdqa        ($xa3,&QWP(16*3-128,"ebx"));
 986
 987     for($i=0;$i<256;$i+=64) {
 988         &vpaddd         ($xa0,$xa0,&QWP($i+16*0-128,"ebp"));    # accumulate key material
 989         &vpaddd         ($xa1,$xa1,&QWP($i+16*1-128,"ebp"));
 990         &vpaddd         ($xa2,$xa2,&QWP($i+16*2-128,"ebp"));
 991         &vpaddd         ($xa3,$xa3,&QWP($i+16*3-128,"ebp"));
 992
 993         &vpunpckldq     ($xt2,$xa0,$xa1);       # "de-interlace" data
 994         &vpunpckldq     ($xt3,$xa2,$xa3);
 995         &vpunpckhdq     ($xa0,$xa0,$xa1);
 996         &vpunpckhdq     ($xa2,$xa2,$xa3);
 997         &vpunpcklqdq    ($xa1,$xt2,$xt3);       # "a0"
 998         &vpunpckhqdq    ($xt2,$xt2,$xt3);       # "a1"
 999         &vpunpcklqdq    ($xt3,$xa0,$xa2);       # "a2"
1000         &vpunpckhqdq    ($xa3,$xa0,$xa2);       # "a3"
1001
1002         &vpxor          ($xt0,$xa1,&QWP(64*0-128,$inp));
1003         &vpxor          ($xt1,$xt2,&QWP(64*1-128,$inp));
1004         &vpxor          ($xt2,$xt3,&QWP(64*2-128,$inp));
1005         &vpxor          ($xt3,$xa3,&QWP(64*3-128,$inp));
1006         &lea            ($inp,&QWP($i<192?16:(64*4-16*3),$inp));
1007         &vmovdqa        ($xa0,&QWP($i+16*4-128,"ebx"))  if ($i<192);
1008         &vmovdqa        ($xa1,&QWP($i+16*5-128,"ebx"))  if ($i<192);
1009         &vmovdqa        ($xa2,&QWP($i+16*6-128,"ebx"))  if ($i<192);
1010         &vmovdqa        ($xa3,&QWP($i+16*7-128,"ebx"))  if ($i<192);
1011         &vmovdqu        (&QWP(64*0-128,$out),$xt0);     # store output
1012         &vmovdqu        (&QWP(64*1-128,$out),$xt1);
1013         &vmovdqu        (&QWP(64*2-128,$out),$xt2);
1014         &vmovdqu        (&QWP(64*3-128,$out),$xt3);
1015         &lea            ($out,&QWP($i<192?16:(64*4-16*3),$out));
1016     }
1017         &sub            ($len,64*4);
1018         &jnc            (&label("outer_loop"));
1019
1020         &add            ($len,64*4);
1021         &jz             (&label("done"));
1022
1023         &mov            ("ebx",&DWP(512+8,"esp"));      # restore pointers
1024         &lea            ($inp,&DWP(-128,$inp));
1025         &mov            ("edx",&DWP(512+4,"esp"));
1026         &lea            ($out,&DWP(-128,$out));
1027
1028         &vmovd          ("xmm2",&DWP(16*12-128,"ebp")); # counter value
1029         &vmovdqu        ("xmm3",&QWP(0,"ebx"));
1030         &vpaddd         ("xmm2","xmm2",&QWP(16*6,"eax"));# +four
1031         &vpand          ("xmm3","xmm3",&QWP(16*7,"eax"));
1032         &vpor           ("xmm3","xmm3","xmm2");         # counter value
1033 {
1034 my ($a,$b,$c,$d,$t,$t1,$rot16,$rot24)=map("xmm$_",(0..7));
1035
1036 sub XOPROUND {
1037         &vpaddd         ($a,$a,$b);
1038         &vpxor          ($d,$d,$a);
1039         &vprotd         ($d,$d,16);
1040
1041         &vpaddd         ($c,$c,$d);
1042         &vpxor          ($b,$b,$c);
1043         &vprotd         ($b,$b,12);
1044
1045         &vpaddd         ($a,$a,$b);
1046         &vpxor          ($d,$d,$a);
1047         &vprotd         ($d,$d,8);
1048
1049         &vpaddd         ($c,$c,$d);
1050         &vpxor          ($b,$b,$c);
1051         &vprotd         ($b,$b,7);
1052 }
1053
1054 &set_label("1x");
1055         &vmovdqa        ($a,&QWP(16*2,"eax"));          # sigma
1056         &vmovdqu        ($b,&QWP(0,"edx"));
1057         &vmovdqu        ($c,&QWP(16,"edx"));
1058         #&vmovdqu       ($d,&QWP(0,"ebx"));             # already loaded
1059         &vmovdqa        ($rot16,&QWP(0,"eax"));
1060         &vmovdqa        ($rot24,&QWP(16,"eax"));
1061         &mov            (&DWP(16*3,"esp"),"ebp");
1062
1063         &vmovdqa        (&QWP(16*0,"esp"),$a);
1064         &vmovdqa        (&QWP(16*1,"esp"),$b);
1065         &vmovdqa        (&QWP(16*2,"esp"),$c);
1066         &vmovdqa        (&QWP(16*3,"esp"),$d);
1067         &mov            ("edx",10);
1068         &jmp            (&label("loop1x"));
1069
1070 &set_label("outer1x",16);
1071         &vmovdqa        ($d,&QWP(16*5,"eax"));          # one
1072         &vmovdqa        ($a,&QWP(16*0,"esp"));
1073         &vmovdqa        ($b,&QWP(16*1,"esp"));
1074         &vmovdqa        ($c,&QWP(16*2,"esp"));
1075         &vpaddd         ($d,$d,&QWP(16*3,"esp"));
1076         &mov            ("edx",10);
1077         &vmovdqa        (&QWP(16*3,"esp"),$d);
1078         &jmp            (&label("loop1x"));
1079
1080 &set_label("loop1x",16);
1081         &XOPROUND();
1082         &vpshufd        ($c,$c,0b01001110);
1083         &vpshufd        ($b,$b,0b00111001);
1084         &vpshufd        ($d,$d,0b10010011);
1085
1086         &XOPROUND();
1087         &vpshufd        ($c,$c,0b01001110);
1088         &vpshufd        ($b,$b,0b10010011);
1089         &vpshufd        ($d,$d,0b00111001);
1090
1091         &dec            ("edx");
1092         &jnz            (&label("loop1x"));
1093
1094         &vpaddd         ($a,$a,&QWP(16*0,"esp"));
1095         &vpaddd         ($b,$b,&QWP(16*1,"esp"));
1096         &vpaddd         ($c,$c,&QWP(16*2,"esp"));
1097         &vpaddd         ($d,$d,&QWP(16*3,"esp"));
1098
1099         &cmp            ($len,64);
1100         &jb             (&label("tail"));
1101
1102         &vpxor          ($a,$a,&QWP(16*0,$inp));        # xor with input
1103         &vpxor          ($b,$b,&QWP(16*1,$inp));
1104         &vpxor          ($c,$c,&QWP(16*2,$inp));
1105         &vpxor          ($d,$d,&QWP(16*3,$inp));
1106         &lea            ($inp,&DWP(16*4,$inp));         # inp+=64
1107
1108         &vmovdqu        (&QWP(16*0,$out),$a);           # write output
1109         &vmovdqu        (&QWP(16*1,$out),$b);
1110         &vmovdqu        (&QWP(16*2,$out),$c);
1111         &vmovdqu        (&QWP(16*3,$out),$d);
1112         &lea            ($out,&DWP(16*4,$out));         # inp+=64
1113
1114         &sub            ($len,64);
1115         &jnz            (&label("outer1x"));
1116
1117         &jmp            (&label("done"));
1118
1119 &set_label("tail");
1120         &vmovdqa        (&QWP(16*0,"esp"),$a);
1121         &vmovdqa        (&QWP(16*1,"esp"),$b);
1122         &vmovdqa        (&QWP(16*2,"esp"),$c);
1123         &vmovdqa        (&QWP(16*3,"esp"),$d);
1124
1125         &xor            ("eax","eax");
1126         &xor            ("edx","edx");
1127         &xor            ("ebp","ebp");
1128
1129 &set_label("tail_loop");
1130         &movb           ("al",&BP(0,"esp","ebp"));
1131         &movb           ("dl",&BP(0,$inp,"ebp"));
1132         &lea            ("ebp",&DWP(1,"ebp"));
1133         &xor            ("al","dl");
1134         &movb           (&BP(-1,$out,"ebp"),"al");
1135         &dec            ($len);
1136         &jnz            (&label("tail_loop"));
1137 }
1138 &set_label("done");
1139         &vzeroupper     ();
1140         &mov            ("esp",&DWP(512,"esp"));
1141 &function_end("ChaCha20_xop");
1142 }
1143
1144 &asm_finish();
1145
1146 close STDOUT;