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 # Skylake       7.53/+22%       2.75
  32 # Silvermont    17.4/+36%       8.35
  33 # Goldmont      13.4/+40%       4.36
  34 # Sledgehammer  10.2/+54%
  35 # Bulldozer     13.4/+50%       4.38(*)
  36 #
  37 # (*)   Bulldozer actually executes 4xXOP code path that delivers 3.55;
  38
  39 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  40 push(@INC,"${dir}","${dir}../../perlasm");
  41 require "x86asm.pl";
  42
  43 $output=pop;
  44 open STDOUT,">$output";
  45
  46 &asm_init($ARGV[0],$ARGV[$#ARGV] eq "386");
  47
  48 $xmm=$ymm=0;
  49 for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
  50
  51 $ymm=1 if ($xmm &&
  52                 `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
  53                         =~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
  54                 ($gasver=$1)>=2.19);    # first version supporting AVX
  55
  56 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32n" &&
  57                 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
  58                 $1>=2.03);      # first version supporting AVX
  59
  60 $ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32" &&
  61                 `ml 2>&1` =~ /Version ([0-9]+)\./ &&
  62                 $1>=10);        # first version supporting AVX
  63
  64 $ymm=1 if ($xmm && !$ymm &&
  65                 `$ENV{CC} -v 2>&1` =~ /(^clang version|based on LLVM) ([3-9]\.[0-9]+)/ &&
  66                 $2>=3.0);       # first version supporting AVX
  67
  68 $a="eax";
  69 ($b,$b_)=("ebx","ebp");
  70 ($c,$c_)=("ecx","esi");
  71 ($d,$d_)=("edx","edi");
  72
  73 sub QUARTERROUND {
  74 my ($ai,$bi,$ci,$di,$i)=@_;
  75 my ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+1)&3),($ai,$bi,$ci,$di)); # next
  76 my ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-1)&3),($ai,$bi,$ci,$di)); # previous
  77
  78         #       a   b   c   d
  79         #
  80         #       0   4   8  12 < even round
  81         #       1   5   9  13
  82         #       2   6  10  14
  83         #       3   7  11  15
  84         #       0   5  10  15 < odd round
  85         #       1   6  11  12
  86         #       2   7   8  13
  87         #       3   4   9  14
  88
  89         if ($i==0) {
  90             my $j=4;
  91             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-$j--)&3),($ap,$bp,$cp,$dp));
  92         } elsif ($i==3) {
  93             my $j=0;
  94             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+$j++)&3),($an,$bn,$cn,$dn));
  95         } elsif ($i==4) {
  96             my $j=4;
  97             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_+$j--)&3),($ap,$bp,$cp,$dp));
  98         } elsif ($i==7) {
  99             my $j=0;
 100             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_-$j++)&3),($an,$bn,$cn,$dn));
 101         }
 102
 103         #&add   ($a,$b);                        # see elsewhere
 104         &xor    ($d,$a);
 105          &mov   (&DWP(4*$cp,"esp"),$c_)         if ($ai>0 && $ai<3);
 106         &rol    ($d,16);
 107          &mov   (&DWP(4*$bp,"esp"),$b_)         if ($i!=0);
 108         &add    ($c,$d);
 109          &mov   ($c_,&DWP(4*$cn,"esp"))         if ($ai>0 && $ai<3);
 110         &xor    ($b,$c);
 111          &mov   ($d_,&DWP(4*$dn,"esp"))         if ($di!=$dn);
 112         &rol    ($b,12);
 113          &mov   ($b_,&DWP(4*$bn,"esp"))         if ($i<7);
 114          &mov   ($b_,&DWP(128,"esp"))           if ($i==7);     # loop counter
 115         &add    ($a,$b);
 116         &xor    ($d,$a);
 117         &mov    (&DWP(4*$ai,"esp"),$a);
 118         &rol    ($d,8);
 119         &mov    ($a,&DWP(4*$an,"esp"));
 120         &add    ($c,$d);
 121         &mov    (&DWP(4*$di,"esp"),$d)          if ($di!=$dn);
 122         &mov    ($d_,$d)                        if ($di==$dn);
 123         &xor    ($b,$c);
 124          &add   ($a,$b_)                        if ($i<7);      # elsewhere
 125         &rol    ($b,7);
 126
 127         ($b,$b_)=($b_,$b);
 128         ($c,$c_)=($c_,$c);
 129         ($d,$d_)=($d_,$d);
 130 }
 131
 132 &static_label("ssse3_shortcut");
 133 &static_label("xop_shortcut");
 134 &static_label("ssse3_data");
 135 &static_label("pic_point");
 136
 137 &function_begin("ChaCha20_ctr32");
 138         &xor    ("eax","eax");
 139         &cmp    ("eax",&wparam(2));             # len==0?
 140         &je     (&label("no_data"));
 141 if ($xmm) {
 142         &call   (&label("pic_point"));
 143 &set_label("pic_point");
 144         &blindpop("eax");
 145         &picmeup("ebp","OPENSSL_ia32cap_P","eax",&label("pic_point"));
 146         &test   (&DWP(0,"ebp"),1<<24);          # test FXSR bit
 147         &jz     (&label("x86"));
 148         &test   (&DWP(4,"ebp"),1<<9);           # test SSSE3 bit
 149         &jz     (&label("x86"));
 150         &jmp    (&label("ssse3_shortcut"));
 151 &set_label("x86");
 152 }
 153         &mov    ("esi",&wparam(3));             # key
 154         &mov    ("edi",&wparam(4));             # counter and nonce
 155
 156         &stack_push(33);
 157
 158         &mov    ("eax",&DWP(4*0,"esi"));        # copy key
 159         &mov    ("ebx",&DWP(4*1,"esi"));
 160         &mov    ("ecx",&DWP(4*2,"esi"));
 161         &mov    ("edx",&DWP(4*3,"esi"));
 162         &mov    (&DWP(64+4*4,"esp"),"eax");
 163         &mov    (&DWP(64+4*5,"esp"),"ebx");
 164         &mov    (&DWP(64+4*6,"esp"),"ecx");
 165         &mov    (&DWP(64+4*7,"esp"),"edx");
 166         &mov    ("eax",&DWP(4*4,"esi"));
 167         &mov    ("ebx",&DWP(4*5,"esi"));
 168         &mov    ("ecx",&DWP(4*6,"esi"));
 169         &mov    ("edx",&DWP(4*7,"esi"));
 170         &mov    (&DWP(64+4*8,"esp"),"eax");
 171         &mov    (&DWP(64+4*9,"esp"),"ebx");
 172         &mov    (&DWP(64+4*10,"esp"),"ecx");
 173         &mov    (&DWP(64+4*11,"esp"),"edx");
 174         &mov    ("eax",&DWP(4*0,"edi"));        # copy counter and nonce
 175         &mov    ("ebx",&DWP(4*1,"edi"));
 176         &mov    ("ecx",&DWP(4*2,"edi"));
 177         &mov    ("edx",&DWP(4*3,"edi"));
 178         &sub    ("eax",1);
 179         &mov    (&DWP(64+4*12,"esp"),"eax");
 180         &mov    (&DWP(64+4*13,"esp"),"ebx");
 181         &mov    (&DWP(64+4*14,"esp"),"ecx");
 182         &mov    (&DWP(64+4*15,"esp"),"edx");
 183         &jmp    (&label("entry"));
 184
 185 &set_label("outer_loop",16);
 186         &mov    (&wparam(1),$b);                # save input
 187         &mov    (&wparam(0),$a);                # save output
 188         &mov    (&wparam(2),$c);                # save len
 189 &set_label("entry");
 190         &mov    ($a,0x61707865);
 191         &mov    (&DWP(4*1,"esp"),0x3320646e);
 192         &mov    (&DWP(4*2,"esp"),0x79622d32);
 193         &mov    (&DWP(4*3,"esp"),0x6b206574);
 194
 195         &mov    ($b, &DWP(64+4*5,"esp"));       # copy key material
 196         &mov    ($b_,&DWP(64+4*6,"esp"));
 197         &mov    ($c, &DWP(64+4*10,"esp"));
 198         &mov    ($c_,&DWP(64+4*11,"esp"));
 199         &mov    ($d, &DWP(64+4*13,"esp"));
 200         &mov    ($d_,&DWP(64+4*14,"esp"));
 201         &mov    (&DWP(4*5,"esp"),$b);
 202         &mov    (&DWP(4*6,"esp"),$b_);
 203         &mov    (&DWP(4*10,"esp"),$c);
 204         &mov    (&DWP(4*11,"esp"),$c_);
 205         &mov    (&DWP(4*13,"esp"),$d);
 206         &mov    (&DWP(4*14,"esp"),$d_);
 207
 208         &mov    ($b, &DWP(64+4*7,"esp"));
 209         &mov    ($d_,&DWP(64+4*15,"esp"));
 210         &mov    ($d, &DWP(64+4*12,"esp"));
 211         &mov    ($b_,&DWP(64+4*4,"esp"));
 212         &mov    ($c, &DWP(64+4*8,"esp"));
 213         &mov    ($c_,&DWP(64+4*9,"esp"));
 214         &add    ($d,1);                         # counter value
 215         &mov    (&DWP(4*7,"esp"),$b);
 216         &mov    (&DWP(4*15,"esp"),$d_);
 217         &mov    (&DWP(64+4*12,"esp"),$d);       # save counter value
 218
 219         &mov    ($b,10);                        # loop counter
 220         &jmp    (&label("loop"));
 221
 222 &set_label("loop",16);
 223         &add    ($a,$b_);                       # elsewhere
 224         &mov    (&DWP(128,"esp"),$b);           # save loop counter
 225         &mov    ($b,$b_);
 226         &QUARTERROUND(0, 4, 8, 12, 0);
 227         &QUARTERROUND(1, 5, 9, 13, 1);
 228         &QUARTERROUND(2, 6,10, 14, 2);
 229         &QUARTERROUND(3, 7,11, 15, 3);
 230         &QUARTERROUND(0, 5,10, 15, 4);
 231         &QUARTERROUND(1, 6,11, 12, 5);
 232         &QUARTERROUND(2, 7, 8, 13, 6);
 233         &QUARTERROUND(3, 4, 9, 14, 7);
 234         &dec    ($b);
 235         &jnz    (&label("loop"));
 236
 237         &mov    ($b,&wparam(2));                # load len
 238
 239         &add    ($a,0x61707865);                # accumulate key material
 240         &add    ($b_,&DWP(64+4*4,"esp"));
 241         &add    ($c, &DWP(64+4*8,"esp"));
 242         &add    ($c_,&DWP(64+4*9,"esp"));
 243
 244         &cmp    ($b,64);
 245         &jb     (&label("tail"));
 246
 247         &mov    ($b,&wparam(1));                # load input pointer
 248         &add    ($d, &DWP(64+4*12,"esp"));
 249         &add    ($d_,&DWP(64+4*14,"esp"));
 250
 251         &xor    ($a, &DWP(4*0,$b));             # xor with input
 252         &xor    ($b_,&DWP(4*4,$b));
 253         &mov    (&DWP(4*0,"esp"),$a);
 254         &mov    ($a,&wparam(0));                # load output pointer
 255         &xor    ($c, &DWP(4*8,$b));
 256         &xor    ($c_,&DWP(4*9,$b));
 257         &xor    ($d, &DWP(4*12,$b));
 258         &xor    ($d_,&DWP(4*14,$b));
 259         &mov    (&DWP(4*4,$a),$b_);             # write output
 260         &mov    (&DWP(4*8,$a),$c);
 261         &mov    (&DWP(4*9,$a),$c_);
 262         &mov    (&DWP(4*12,$a),$d);
 263         &mov    (&DWP(4*14,$a),$d_);
 264
 265         &mov    ($b_,&DWP(4*1,"esp"));
 266         &mov    ($c, &DWP(4*2,"esp"));
 267         &mov    ($c_,&DWP(4*3,"esp"));
 268         &mov    ($d, &DWP(4*5,"esp"));
 269         &mov    ($d_,&DWP(4*6,"esp"));
 270         &add    ($b_,0x3320646e);               # accumulate key material
 271         &add    ($c, 0x79622d32);
 272         &add    ($c_,0x6b206574);
 273         &add    ($d, &DWP(64+4*5,"esp"));
 274         &add    ($d_,&DWP(64+4*6,"esp"));
 275         &xor    ($b_,&DWP(4*1,$b));
 276         &xor    ($c, &DWP(4*2,$b));
 277         &xor    ($c_,&DWP(4*3,$b));
 278         &xor    ($d, &DWP(4*5,$b));
 279         &xor    ($d_,&DWP(4*6,$b));
 280         &mov    (&DWP(4*1,$a),$b_);
 281         &mov    (&DWP(4*2,$a),$c);
 282         &mov    (&DWP(4*3,$a),$c_);
 283         &mov    (&DWP(4*5,$a),$d);
 284         &mov    (&DWP(4*6,$a),$d_);
 285
 286         &mov    ($b_,&DWP(4*7,"esp"));
 287         &mov    ($c, &DWP(4*10,"esp"));
 288         &mov    ($c_,&DWP(4*11,"esp"));
 289         &mov    ($d, &DWP(4*13,"esp"));
 290         &mov    ($d_,&DWP(4*15,"esp"));
 291         &add    ($b_,&DWP(64+4*7,"esp"));
 292         &add    ($c, &DWP(64+4*10,"esp"));
 293         &add    ($c_,&DWP(64+4*11,"esp"));
 294         &add    ($d, &DWP(64+4*13,"esp"));
 295         &add    ($d_,&DWP(64+4*15,"esp"));
 296         &xor    ($b_,&DWP(4*7,$b));
 297         &xor    ($c, &DWP(4*10,$b));
 298         &xor    ($c_,&DWP(4*11,$b));
 299         &xor    ($d, &DWP(4*13,$b));
 300         &xor    ($d_,&DWP(4*15,$b));
 301         &lea    ($b,&DWP(4*16,$b));
 302         &mov    (&DWP(4*7,$a),$b_);
 303         &mov    ($b_,&DWP(4*0,"esp"));
 304         &mov    (&DWP(4*10,$a),$c);
 305         &mov    ($c,&wparam(2));                # len
 306         &mov    (&DWP(4*11,$a),$c_);
 307         &mov    (&DWP(4*13,$a),$d);
 308         &mov    (&DWP(4*15,$a),$d_);
 309         &mov    (&DWP(4*0,$a),$b_);
 310         &lea    ($a,&DWP(4*16,$a));
 311         &sub    ($c,64);
 312         &jnz    (&label("outer_loop"));
 313
 314         &jmp    (&label("done"));
 315
 316 &set_label("tail");
 317         &add    ($d, &DWP(64+4*12,"esp"));
 318         &add    ($d_,&DWP(64+4*14,"esp"));
 319         &mov    (&DWP(4*0,"esp"),$a);
 320         &mov    (&DWP(4*4,"esp"),$b_);
 321         &mov    (&DWP(4*8,"esp"),$c);
 322         &mov    (&DWP(4*9,"esp"),$c_);
 323         &mov    (&DWP(4*12,"esp"),$d);
 324         &mov    (&DWP(4*14,"esp"),$d_);
 325
 326         &mov    ($b_,&DWP(4*1,"esp"));
 327         &mov    ($c, &DWP(4*2,"esp"));
 328         &mov    ($c_,&DWP(4*3,"esp"));
 329         &mov    ($d, &DWP(4*5,"esp"));
 330         &mov    ($d_,&DWP(4*6,"esp"));
 331         &add    ($b_,0x3320646e);               # accumulate key material
 332         &add    ($c, 0x79622d32);
 333         &add    ($c_,0x6b206574);
 334         &add    ($d, &DWP(64+4*5,"esp"));
 335         &add    ($d_,&DWP(64+4*6,"esp"));
 336         &mov    (&DWP(4*1,"esp"),$b_);
 337         &mov    (&DWP(4*2,"esp"),$c);
 338         &mov    (&DWP(4*3,"esp"),$c_);
 339         &mov    (&DWP(4*5,"esp"),$d);
 340         &mov    (&DWP(4*6,"esp"),$d_);
 341
 342         &mov    ($b_,&DWP(4*7,"esp"));
 343         &mov    ($c, &DWP(4*10,"esp"));
 344         &mov    ($c_,&DWP(4*11,"esp"));
 345         &mov    ($d, &DWP(4*13,"esp"));
 346         &mov    ($d_,&DWP(4*15,"esp"));
 347         &add    ($b_,&DWP(64+4*7,"esp"));
 348         &add    ($c, &DWP(64+4*10,"esp"));
 349         &add    ($c_,&DWP(64+4*11,"esp"));
 350         &add    ($d, &DWP(64+4*13,"esp"));
 351         &add    ($d_,&DWP(64+4*15,"esp"));
 352         &mov    (&DWP(4*7,"esp"),$b_);
 353         &mov    ($b_,&wparam(1));               # load input
 354         &mov    (&DWP(4*10,"esp"),$c);
 355         &mov    ($c,&wparam(0));                # load output
 356         &mov    (&DWP(4*11,"esp"),$c_);
 357         &xor    ($c_,$c_);
 358         &mov    (&DWP(4*13,"esp"),$d);
 359         &mov    (&DWP(4*15,"esp"),$d_);
 360
 361         &xor    ("eax","eax");
 362         &xor    ("edx","edx");
 363 &set_label("tail_loop");
 364         &movb   ("al",&BP(0,$c_,$b_));
 365         &movb   ("dl",&BP(0,"esp",$c_));
 366         &lea    ($c_,&DWP(1,$c_));
 367         &xor    ("al","dl");
 368         &mov    (&BP(-1,$c,$c_),"al");
 369         &dec    ($b);
 370         &jnz    (&label("tail_loop"));
 371
 372 &set_label("done");
 373         &stack_pop(33);
 374 &set_label("no_data");
 375 &function_end("ChaCha20_ctr32");
 376
 377 if ($xmm) {
 378 my ($xa,$xa_,$xb,$xb_,$xc,$xc_,$xd,$xd_)=map("xmm$_",(0..7));
 379 my ($out,$inp,$len)=("edi","esi","ecx");
 380
 381 sub QUARTERROUND_SSSE3 {
 382 my ($ai,$bi,$ci,$di,$i)=@_;
 383 my ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+1)&3),($ai,$bi,$ci,$di)); # next
 384 my ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-1)&3),($ai,$bi,$ci,$di)); # previous
 385
 386         #       a   b   c   d
 387         #
 388         #       0   4   8  12 < even round
 389         #       1   5   9  13
 390         #       2   6  10  14
 391         #       3   7  11  15
 392         #       0   5  10  15 < odd round
 393         #       1   6  11  12
 394         #       2   7   8  13
 395         #       3   4   9  14
 396
 397         if ($i==0) {
 398             my $j=4;
 399             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-$j--)&3),($ap,$bp,$cp,$dp));
 400         } elsif ($i==3) {
 401             my $j=0;
 402             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+$j++)&3),($an,$bn,$cn,$dn));
 403         } elsif ($i==4) {
 404             my $j=4;
 405             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_+$j--)&3),($ap,$bp,$cp,$dp));
 406         } elsif ($i==7) {
 407             my $j=0;
 408             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_-$j++)&3),($an,$bn,$cn,$dn));
 409         }
 410
 411         #&paddd ($xa,$xb);                      # see elsewhere
 412         #&pxor  ($xd,$xa);                      # see elsewhere
 413          &movdqa(&QWP(16*$cp-128,"ebx"),$xc_)   if ($ai>0 && $ai<3);
 414         &pshufb ($xd,&QWP(0,"eax"));            # rot16
 415          &movdqa(&QWP(16*$bp-128,"ebx"),$xb_)   if ($i!=0);
 416         &paddd  ($xc,$xd);
 417          &movdqa($xc_,&QWP(16*$cn-128,"ebx"))   if ($ai>0 && $ai<3);
 418         &pxor   ($xb,$xc);
 419          &movdqa($xb_,&QWP(16*$bn-128,"ebx"))   if ($i<7);
 420         &movdqa ($xa_,$xb);                     # borrow as temporary
 421         &pslld  ($xb,12);
 422         &psrld  ($xa_,20);
 423         &por    ($xb,$xa_);
 424          &movdqa($xa_,&QWP(16*$an-128,"ebx"));
 425         &paddd  ($xa,$xb);
 426          &movdqa($xd_,&QWP(16*$dn-128,"ebx"))   if ($di!=$dn);
 427         &pxor   ($xd,$xa);
 428         &movdqa (&QWP(16*$ai-128,"ebx"),$xa);
 429         &pshufb ($xd,&QWP(16,"eax"));           # rot8
 430         &paddd  ($xc,$xd);
 431         &movdqa (&QWP(16*$di-128,"ebx"),$xd)    if ($di!=$dn);
 432         &movdqa ($xd_,$xd)                      if ($di==$dn);
 433         &pxor   ($xb,$xc);
 434          &paddd ($xa_,$xb_)                     if ($i<7);      # elsewhere
 435         &movdqa ($xa,$xb);                      # borrow as temporary
 436         &pslld  ($xb,7);
 437         &psrld  ($xa,25);
 438          &pxor  ($xd_,$xa_)                     if ($i<7);      # elsewhere
 439         &por    ($xb,$xa);
 440
 441         ($xa,$xa_)=($xa_,$xa);
 442         ($xb,$xb_)=($xb_,$xb);
 443         ($xc,$xc_)=($xc_,$xc);
 444         ($xd,$xd_)=($xd_,$xd);
 445 }
 446
 447 &function_begin("ChaCha20_ssse3");
 448 &set_label("ssse3_shortcut");
 449 if ($ymm) {
 450         &test           (&DWP(4,"ebp"),1<<11);          # test XOP bit
 451         &jnz            (&label("xop_shortcut"));
 452 }
 453
 454         &mov            ($out,&wparam(0));
 455         &mov            ($inp,&wparam(1));
 456         &mov            ($len,&wparam(2));
 457         &mov            ("edx",&wparam(3));             # key
 458         &mov            ("ebx",&wparam(4));             # counter and nonce
 459
 460         &mov            ("ebp","esp");
 461         &stack_push     (131);
 462         &and            ("esp",-64);
 463         &mov            (&DWP(512,"esp"),"ebp");
 464
 465         &lea            ("eax",&DWP(&label("ssse3_data")."-".
 466                                     &label("pic_point"),"eax"));
 467         &movdqu         ("xmm3",&QWP(0,"ebx"));         # counter and nonce
 468
 469 if (defined($gasver) && $gasver>=2.17) {                # even though we encode
 470                                                         # pshufb manually, we
 471                                                         # handle only register
 472                                                         # operands, while this
 473                                                         # segment uses memory
 474                                                         # operand...
 475         &cmp            ($len,64*4);
 476         &jb             (&label("1x"));
 477
 478         &mov            (&DWP(512+4,"esp"),"edx");      # offload pointers
 479         &mov            (&DWP(512+8,"esp"),"ebx");
 480         &sub            ($len,64*4);                    # bias len
 481         &lea            ("ebp",&DWP(256+128,"esp"));    # size optimization
 482
 483         &movdqu         ("xmm7",&QWP(0,"edx"));         # key
 484         &pshufd         ("xmm0","xmm3",0x00);
 485         &pshufd         ("xmm1","xmm3",0x55);
 486         &pshufd         ("xmm2","xmm3",0xaa);
 487         &pshufd         ("xmm3","xmm3",0xff);
 488          &paddd         ("xmm0",&QWP(16*3,"eax"));      # fix counters
 489         &pshufd         ("xmm4","xmm7",0x00);
 490         &pshufd         ("xmm5","xmm7",0x55);
 491          &psubd         ("xmm0",&QWP(16*4,"eax"));
 492         &pshufd         ("xmm6","xmm7",0xaa);
 493         &pshufd         ("xmm7","xmm7",0xff);
 494         &movdqa         (&QWP(16*12-128,"ebp"),"xmm0");
 495         &movdqa         (&QWP(16*13-128,"ebp"),"xmm1");
 496         &movdqa         (&QWP(16*14-128,"ebp"),"xmm2");
 497         &movdqa         (&QWP(16*15-128,"ebp"),"xmm3");
 498          &movdqu        ("xmm3",&QWP(16,"edx"));        # key
 499         &movdqa         (&QWP(16*4-128,"ebp"),"xmm4");
 500         &movdqa         (&QWP(16*5-128,"ebp"),"xmm5");
 501         &movdqa         (&QWP(16*6-128,"ebp"),"xmm6");
 502         &movdqa         (&QWP(16*7-128,"ebp"),"xmm7");
 503          &movdqa        ("xmm7",&QWP(16*2,"eax"));      # sigma
 504          &lea           ("ebx",&DWP(128,"esp"));        # size optimization
 505
 506         &pshufd         ("xmm0","xmm3",0x00);
 507         &pshufd         ("xmm1","xmm3",0x55);
 508         &pshufd         ("xmm2","xmm3",0xaa);
 509         &pshufd         ("xmm3","xmm3",0xff);
 510         &pshufd         ("xmm4","xmm7",0x00);
 511         &pshufd         ("xmm5","xmm7",0x55);
 512         &pshufd         ("xmm6","xmm7",0xaa);
 513         &pshufd         ("xmm7","xmm7",0xff);
 514         &movdqa         (&QWP(16*8-128,"ebp"),"xmm0");
 515         &movdqa         (&QWP(16*9-128,"ebp"),"xmm1");
 516         &movdqa         (&QWP(16*10-128,"ebp"),"xmm2");
 517         &movdqa         (&QWP(16*11-128,"ebp"),"xmm3");
 518         &movdqa         (&QWP(16*0-128,"ebp"),"xmm4");
 519         &movdqa         (&QWP(16*1-128,"ebp"),"xmm5");
 520         &movdqa         (&QWP(16*2-128,"ebp"),"xmm6");
 521         &movdqa         (&QWP(16*3-128,"ebp"),"xmm7");
 522
 523         &lea            ($inp,&DWP(128,$inp));          # size optimization
 524         &lea            ($out,&DWP(128,$out));          # size optimization
 525         &jmp            (&label("outer_loop"));
 526
 527 &set_label("outer_loop",16);
 528         #&movdqa        ("xmm0",&QWP(16*0-128,"ebp"));  # copy key material
 529         &movdqa         ("xmm1",&QWP(16*1-128,"ebp"));
 530         &movdqa         ("xmm2",&QWP(16*2-128,"ebp"));
 531         &movdqa         ("xmm3",&QWP(16*3-128,"ebp"));
 532         #&movdqa        ("xmm4",&QWP(16*4-128,"ebp"));
 533         &movdqa         ("xmm5",&QWP(16*5-128,"ebp"));
 534         &movdqa         ("xmm6",&QWP(16*6-128,"ebp"));
 535         &movdqa         ("xmm7",&QWP(16*7-128,"ebp"));
 536         #&movdqa        (&QWP(16*0-128,"ebx"),"xmm0");
 537         &movdqa         (&QWP(16*1-128,"ebx"),"xmm1");
 538         &movdqa         (&QWP(16*2-128,"ebx"),"xmm2");
 539         &movdqa         (&QWP(16*3-128,"ebx"),"xmm3");
 540         #&movdqa        (&QWP(16*4-128,"ebx"),"xmm4");
 541         &movdqa         (&QWP(16*5-128,"ebx"),"xmm5");
 542         &movdqa         (&QWP(16*6-128,"ebx"),"xmm6");
 543         &movdqa         (&QWP(16*7-128,"ebx"),"xmm7");
 544         #&movdqa        ("xmm0",&QWP(16*8-128,"ebp"));
 545         #&movdqa        ("xmm1",&QWP(16*9-128,"ebp"));
 546         &movdqa         ("xmm2",&QWP(16*10-128,"ebp"));
 547         &movdqa         ("xmm3",&QWP(16*11-128,"ebp"));
 548         &movdqa         ("xmm4",&QWP(16*12-128,"ebp"));
 549         &movdqa         ("xmm5",&QWP(16*13-128,"ebp"));
 550         &movdqa         ("xmm6",&QWP(16*14-128,"ebp"));
 551         &movdqa         ("xmm7",&QWP(16*15-128,"ebp"));
 552         &paddd          ("xmm4",&QWP(16*4,"eax"));      # counter value
 553         #&movdqa        (&QWP(16*8-128,"ebx"),"xmm0");
 554         #&movdqa        (&QWP(16*9-128,"ebx"),"xmm1");
 555         &movdqa         (&QWP(16*10-128,"ebx"),"xmm2");
 556         &movdqa         (&QWP(16*11-128,"ebx"),"xmm3");
 557         &movdqa         (&QWP(16*12-128,"ebx"),"xmm4");
 558         &movdqa         (&QWP(16*13-128,"ebx"),"xmm5");
 559         &movdqa         (&QWP(16*14-128,"ebx"),"xmm6");
 560         &movdqa         (&QWP(16*15-128,"ebx"),"xmm7");
 561         &movdqa         (&QWP(16*12-128,"ebp"),"xmm4"); # save counter value
 562
 563         &movdqa         ($xa, &QWP(16*0-128,"ebp"));
 564         &movdqa         ($xd, "xmm4");
 565         &movdqa         ($xb_,&QWP(16*4-128,"ebp"));
 566         &movdqa         ($xc, &QWP(16*8-128,"ebp"));
 567         &movdqa         ($xc_,&QWP(16*9-128,"ebp"));
 568
 569         &mov            ("edx",10);                     # loop counter
 570         &nop            ();
 571
 572 &set_label("loop",16);
 573         &paddd          ($xa,$xb_);                     # elsewhere
 574         &movdqa         ($xb,$xb_);
 575         &pxor           ($xd,$xa);                      # elsewhere
 576         &QUARTERROUND_SSSE3(0, 4, 8, 12, 0);
 577         &QUARTERROUND_SSSE3(1, 5, 9, 13, 1);
 578         &QUARTERROUND_SSSE3(2, 6,10, 14, 2);
 579         &QUARTERROUND_SSSE3(3, 7,11, 15, 3);
 580         &QUARTERROUND_SSSE3(0, 5,10, 15, 4);
 581         &QUARTERROUND_SSSE3(1, 6,11, 12, 5);
 582         &QUARTERROUND_SSSE3(2, 7, 8, 13, 6);
 583         &QUARTERROUND_SSSE3(3, 4, 9, 14, 7);
 584         &dec            ("edx");
 585         &jnz            (&label("loop"));
 586
 587         &movdqa         (&QWP(16*4-128,"ebx"),$xb_);
 588         &movdqa         (&QWP(16*8-128,"ebx"),$xc);
 589         &movdqa         (&QWP(16*9-128,"ebx"),$xc_);
 590         &movdqa         (&QWP(16*12-128,"ebx"),$xd);
 591         &movdqa         (&QWP(16*14-128,"ebx"),$xd_);
 592
 593     my ($xa0,$xa1,$xa2,$xa3,$xt0,$xt1,$xt2,$xt3)=map("xmm$_",(0..7));
 594
 595         #&movdqa        ($xa0,&QWP(16*0-128,"ebx"));    # it's there
 596         &movdqa         ($xa1,&QWP(16*1-128,"ebx"));
 597         &movdqa         ($xa2,&QWP(16*2-128,"ebx"));
 598         &movdqa         ($xa3,&QWP(16*3-128,"ebx"));
 599
 600     for($i=0;$i<256;$i+=64) {
 601         &paddd          ($xa0,&QWP($i+16*0-128,"ebp")); # accumulate key material
 602         &paddd          ($xa1,&QWP($i+16*1-128,"ebp"));
 603         &paddd          ($xa2,&QWP($i+16*2-128,"ebp"));
 604         &paddd          ($xa3,&QWP($i+16*3-128,"ebp"));
 605
 606         &movdqa         ($xt2,$xa0);            # "de-interlace" data
 607         &punpckldq      ($xa0,$xa1);
 608         &movdqa         ($xt3,$xa2);
 609         &punpckldq      ($xa2,$xa3);
 610         &punpckhdq      ($xt2,$xa1);
 611         &punpckhdq      ($xt3,$xa3);
 612         &movdqa         ($xa1,$xa0);
 613         &punpcklqdq     ($xa0,$xa2);            # "a0"
 614         &movdqa         ($xa3,$xt2);
 615         &punpcklqdq     ($xt2,$xt3);            # "a2"
 616         &punpckhqdq     ($xa1,$xa2);            # "a1"
 617         &punpckhqdq     ($xa3,$xt3);            # "a3"
 618
 619         #($xa2,$xt2)=($xt2,$xa2);
 620
 621         &movdqu         ($xt0,&QWP(64*0-128,$inp));     # load input
 622         &movdqu         ($xt1,&QWP(64*1-128,$inp));
 623         &movdqu         ($xa2,&QWP(64*2-128,$inp));
 624         &movdqu         ($xt3,&QWP(64*3-128,$inp));
 625         &lea            ($inp,&QWP($i<192?16:(64*4-16*3),$inp));
 626         &pxor           ($xt0,$xa0);
 627         &movdqa         ($xa0,&QWP($i+16*4-128,"ebx"))  if ($i<192);
 628         &pxor           ($xt1,$xa1);
 629         &movdqa         ($xa1,&QWP($i+16*5-128,"ebx"))  if ($i<192);
 630         &pxor           ($xt2,$xa2);
 631         &movdqa         ($xa2,&QWP($i+16*6-128,"ebx"))  if ($i<192);
 632         &pxor           ($xt3,$xa3);
 633         &movdqa         ($xa3,&QWP($i+16*7-128,"ebx"))  if ($i<192);
 634         &movdqu         (&QWP(64*0-128,$out),$xt0);     # store output
 635         &movdqu         (&QWP(64*1-128,$out),$xt1);
 636         &movdqu         (&QWP(64*2-128,$out),$xt2);
 637         &movdqu         (&QWP(64*3-128,$out),$xt3);
 638         &lea            ($out,&QWP($i<192?16:(64*4-16*3),$out));
 639     }
 640         &sub            ($len,64*4);
 641         &jnc            (&label("outer_loop"));
 642
 643         &add            ($len,64*4);
 644         &jz             (&label("done"));
 645
 646         &mov            ("ebx",&DWP(512+8,"esp"));      # restore pointers
 647         &lea            ($inp,&DWP(-128,$inp));
 648         &mov            ("edx",&DWP(512+4,"esp"));
 649         &lea            ($out,&DWP(-128,$out));
 650
 651         &movd           ("xmm2",&DWP(16*12-128,"ebp")); # counter value
 652         &movdqu         ("xmm3",&QWP(0,"ebx"));
 653         &paddd          ("xmm2",&QWP(16*6,"eax"));      # +four
 654         &pand           ("xmm3",&QWP(16*7,"eax"));
 655         &por            ("xmm3","xmm2");                # counter value
 656 }
 657 {
 658 my ($a,$b,$c,$d,$t,$t1,$rot16,$rot24)=map("xmm$_",(0..7));
 659
 660 sub SSSE3ROUND {        # critical path is 20 "SIMD ticks" per round
 661         &paddd          ($a,$b);
 662         &pxor           ($d,$a);
 663         &pshufb         ($d,$rot16);
 664
 665         &paddd          ($c,$d);
 666         &pxor           ($b,$c);
 667         &movdqa         ($t,$b);
 668         &psrld          ($b,20);
 669         &pslld          ($t,12);
 670         &por            ($b,$t);
 671
 672         &paddd          ($a,$b);
 673         &pxor           ($d,$a);
 674         &pshufb         ($d,$rot24);
 675
 676         &paddd          ($c,$d);
 677         &pxor           ($b,$c);
 678         &movdqa         ($t,$b);
 679         &psrld          ($b,25);
 680         &pslld          ($t,7);
 681         &por            ($b,$t);
 682 }
 683
 684 &set_label("1x");
 685         &movdqa         ($a,&QWP(16*2,"eax"));          # sigma
 686         &movdqu         ($b,&QWP(0,"edx"));
 687         &movdqu         ($c,&QWP(16,"edx"));
 688         #&movdqu        ($d,&QWP(0,"ebx"));             # already loaded
 689         &movdqa         ($rot16,&QWP(0,"eax"));
 690         &movdqa         ($rot24,&QWP(16,"eax"));
 691         &mov            (&DWP(16*3,"esp"),"ebp");
 692
 693         &movdqa         (&QWP(16*0,"esp"),$a);
 694         &movdqa         (&QWP(16*1,"esp"),$b);
 695         &movdqa         (&QWP(16*2,"esp"),$c);
 696         &movdqa         (&QWP(16*3,"esp"),$d);
 697         &mov            ("edx",10);
 698         &jmp            (&label("loop1x"));
 699
 700 &set_label("outer1x",16);
 701         &movdqa         ($d,&QWP(16*5,"eax"));          # one
 702         &movdqa         ($a,&QWP(16*0,"esp"));
 703         &movdqa         ($b,&QWP(16*1,"esp"));
 704         &movdqa         ($c,&QWP(16*2,"esp"));
 705         &paddd          ($d,&QWP(16*3,"esp"));
 706         &mov            ("edx",10);
 707         &movdqa         (&QWP(16*3,"esp"),$d);
 708         &jmp            (&label("loop1x"));
 709
 710 &set_label("loop1x",16);
 711         &SSSE3ROUND();
 712         &pshufd ($c,$c,0b01001110);
 713         &pshufd ($b,$b,0b00111001);
 714         &pshufd ($d,$d,0b10010011);
 715         &nop    ();
 716
 717         &SSSE3ROUND();
 718         &pshufd ($c,$c,0b01001110);
 719         &pshufd ($b,$b,0b10010011);
 720         &pshufd ($d,$d,0b00111001);
 721
 722         &dec            ("edx");
 723         &jnz            (&label("loop1x"));
 724
 725         &paddd          ($a,&QWP(16*0,"esp"));
 726         &paddd          ($b,&QWP(16*1,"esp"));
 727         &paddd          ($c,&QWP(16*2,"esp"));
 728         &paddd          ($d,&QWP(16*3,"esp"));
 729
 730         &cmp            ($len,64);
 731         &jb             (&label("tail"));
 732
 733         &movdqu         ($t,&QWP(16*0,$inp));
 734         &movdqu         ($t1,&QWP(16*1,$inp));
 735         &pxor           ($a,$t);                # xor with input
 736         &movdqu         ($t,&QWP(16*2,$inp));
 737         &pxor           ($b,$t1);
 738         &movdqu         ($t1,&QWP(16*3,$inp));
 739         &pxor           ($c,$t);
 740         &pxor           ($d,$t1);
 741         &lea            ($inp,&DWP(16*4,$inp)); # inp+=64
 742
 743         &movdqu         (&QWP(16*0,$out),$a);   # write output
 744         &movdqu         (&QWP(16*1,$out),$b);
 745         &movdqu         (&QWP(16*2,$out),$c);
 746         &movdqu         (&QWP(16*3,$out),$d);
 747         &lea            ($out,&DWP(16*4,$out)); # inp+=64
 748
 749         &sub            ($len,64);
 750         &jnz            (&label("outer1x"));
 751
 752         &jmp            (&label("done"));
 753
 754 &set_label("tail");
 755         &movdqa         (&QWP(16*0,"esp"),$a);
 756         &movdqa         (&QWP(16*1,"esp"),$b);
 757         &movdqa         (&QWP(16*2,"esp"),$c);
 758         &movdqa         (&QWP(16*3,"esp"),$d);
 759
 760         &xor            ("eax","eax");
 761         &xor            ("edx","edx");
 762         &xor            ("ebp","ebp");
 763
 764 &set_label("tail_loop");
 765         &movb           ("al",&BP(0,"esp","ebp"));
 766         &movb           ("dl",&BP(0,$inp,"ebp"));
 767         &lea            ("ebp",&DWP(1,"ebp"));
 768         &xor            ("al","dl");
 769         &movb           (&BP(-1,$out,"ebp"),"al");
 770         &dec            ($len);
 771         &jnz            (&label("tail_loop"));
 772 }
 773 &set_label("done");
 774         &mov            ("esp",&DWP(512,"esp"));
 775 &function_end("ChaCha20_ssse3");
 776
 777 &align  (64);
 778 &set_label("ssse3_data");
 779 &data_byte(0x2,0x3,0x0,0x1, 0x6,0x7,0x4,0x5, 0xa,0xb,0x8,0x9, 0xe,0xf,0xc,0xd);
 780 &data_byte(0x3,0x0,0x1,0x2, 0x7,0x4,0x5,0x6, 0xb,0x8,0x9,0xa, 0xf,0xc,0xd,0xe);
 781 &data_word(0x61707865,0x3320646e,0x79622d32,0x6b206574);
 782 &data_word(0,1,2,3);
 783 &data_word(4,4,4,4);
 784 &data_word(1,0,0,0);
 785 &data_word(4,0,0,0);
 786 &data_word(0,-1,-1,-1);
 787 &align  (64);
 788 }
 789 &asciz  ("ChaCha20 for x86, CRYPTOGAMS by <appro\@openssl.org>");
 790
 791 if ($ymm) {
 792 my ($xa,$xa_,$xb,$xb_,$xc,$xc_,$xd,$xd_)=map("xmm$_",(0..7));
 793 my ($out,$inp,$len)=("edi","esi","ecx");
 794
 795 sub QUARTERROUND_XOP {
 796 my ($ai,$bi,$ci,$di,$i)=@_;
 797 my ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+1)&3),($ai,$bi,$ci,$di)); # next
 798 my ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-1)&3),($ai,$bi,$ci,$di)); # previous
 799
 800         #       a   b   c   d
 801         #
 802         #       0   4   8  12 < even round
 803         #       1   5   9  13
 804         #       2   6  10  14
 805         #       3   7  11  15
 806         #       0   5  10  15 < odd round
 807         #       1   6  11  12
 808         #       2   7   8  13
 809         #       3   4   9  14
 810
 811         if ($i==0) {
 812             my $j=4;
 813             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_-$j--)&3),($ap,$bp,$cp,$dp));
 814         } elsif ($i==3) {
 815             my $j=0;
 816             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_+$j++)&3),($an,$bn,$cn,$dn));
 817         } elsif ($i==4) {
 818             my $j=4;
 819             ($ap,$bp,$cp,$dp)=map(($_&~3)+(($_+$j--)&3),($ap,$bp,$cp,$dp));
 820         } elsif ($i==7) {
 821             my $j=0;
 822             ($an,$bn,$cn,$dn)=map(($_&~3)+(($_-$j++)&3),($an,$bn,$cn,$dn));
 823         }
 824
 825         #&vpaddd        ($xa,$xa,$xb);                  # see elsewhere
 826         #&vpxor         ($xd,$xd,$xa);                  # see elsewhere
 827          &vmovdqa       (&QWP(16*$cp-128,"ebx"),$xc_)   if ($ai>0 && $ai<3);
 828         &vprotd         ($xd,$xd,16);
 829          &vmovdqa       (&QWP(16*$bp-128,"ebx"),$xb_)   if ($i!=0);
 830         &vpaddd         ($xc,$xc,$xd);
 831          &vmovdqa       ($xc_,&QWP(16*$cn-128,"ebx"))   if ($ai>0 && $ai<3);
 832         &vpxor          ($xb,$i!=0?$xb:$xb_,$xc);
 833          &vmovdqa       ($xa_,&QWP(16*$an-128,"ebx"));
 834         &vprotd         ($xb,$xb,12);
 835          &vmovdqa       ($xb_,&QWP(16*$bn-128,"ebx"))   if ($i<7);
 836         &vpaddd         ($xa,$xa,$xb);
 837          &vmovdqa       ($xd_,&QWP(16*$dn-128,"ebx"))   if ($di!=$dn);
 838         &vpxor          ($xd,$xd,$xa);
 839          &vpaddd        ($xa_,$xa_,$xb_)                if ($i<7);      # elsewhere
 840         &vprotd         ($xd,$xd,8);
 841         &vmovdqa        (&QWP(16*$ai-128,"ebx"),$xa);
 842         &vpaddd         ($xc,$xc,$xd);
 843         &vmovdqa        (&QWP(16*$di-128,"ebx"),$xd)    if ($di!=$dn);
 844         &vpxor          ($xb,$xb,$xc);
 845          &vpxor         ($xd_,$di==$dn?$xd:$xd_,$xa_)   if ($i<7);      # elsewhere
 846         &vprotd         ($xb,$xb,7);
 847
 848         ($xa,$xa_)=($xa_,$xa);
 849         ($xb,$xb_)=($xb_,$xb);
 850         ($xc,$xc_)=($xc_,$xc);
 851         ($xd,$xd_)=($xd_,$xd);
 852 }
 853
 854 &function_begin("ChaCha20_xop");
 855 &set_label("xop_shortcut");
 856         &mov            ($out,&wparam(0));
 857         &mov            ($inp,&wparam(1));
 858         &mov            ($len,&wparam(2));
 859         &mov            ("edx",&wparam(3));             # key
 860         &mov            ("ebx",&wparam(4));             # counter and nonce
 861         &vzeroupper     ();
 862
 863         &mov            ("ebp","esp");
 864         &stack_push     (131);
 865         &and            ("esp",-64);
 866         &mov            (&DWP(512,"esp"),"ebp");
 867
 868         &lea            ("eax",&DWP(&label("ssse3_data")."-".
 869                                     &label("pic_point"),"eax"));
 870         &vmovdqu        ("xmm3",&QWP(0,"ebx"));         # counter and nonce
 871
 872         &cmp            ($len,64*4);
 873         &jb             (&label("1x"));
 874
 875         &mov            (&DWP(512+4,"esp"),"edx");      # offload pointers
 876         &mov            (&DWP(512+8,"esp"),"ebx");
 877         &sub            ($len,64*4);                    # bias len
 878         &lea            ("ebp",&DWP(256+128,"esp"));    # size optimization
 879
 880         &vmovdqu        ("xmm7",&QWP(0,"edx"));         # key
 881         &vpshufd        ("xmm0","xmm3",0x00);
 882         &vpshufd        ("xmm1","xmm3",0x55);
 883         &vpshufd        ("xmm2","xmm3",0xaa);
 884         &vpshufd        ("xmm3","xmm3",0xff);
 885          &vpaddd        ("xmm0","xmm0",&QWP(16*3,"eax"));       # fix counters
 886         &vpshufd        ("xmm4","xmm7",0x00);
 887         &vpshufd        ("xmm5","xmm7",0x55);
 888          &vpsubd        ("xmm0","xmm0",&QWP(16*4,"eax"));
 889         &vpshufd        ("xmm6","xmm7",0xaa);
 890         &vpshufd        ("xmm7","xmm7",0xff);
 891         &vmovdqa        (&QWP(16*12-128,"ebp"),"xmm0");
 892         &vmovdqa        (&QWP(16*13-128,"ebp"),"xmm1");
 893         &vmovdqa        (&QWP(16*14-128,"ebp"),"xmm2");
 894         &vmovdqa        (&QWP(16*15-128,"ebp"),"xmm3");
 895          &vmovdqu       ("xmm3",&QWP(16,"edx"));        # key
 896         &vmovdqa        (&QWP(16*4-128,"ebp"),"xmm4");
 897         &vmovdqa        (&QWP(16*5-128,"ebp"),"xmm5");
 898         &vmovdqa        (&QWP(16*6-128,"ebp"),"xmm6");
 899         &vmovdqa        (&QWP(16*7-128,"ebp"),"xmm7");
 900          &vmovdqa       ("xmm7",&QWP(16*2,"eax"));      # sigma
 901          &lea           ("ebx",&DWP(128,"esp"));        # size optimization
 902
 903         &vpshufd        ("xmm0","xmm3",0x00);
 904         &vpshufd        ("xmm1","xmm3",0x55);
 905         &vpshufd        ("xmm2","xmm3",0xaa);
 906         &vpshufd        ("xmm3","xmm3",0xff);
 907         &vpshufd        ("xmm4","xmm7",0x00);
 908         &vpshufd        ("xmm5","xmm7",0x55);
 909         &vpshufd        ("xmm6","xmm7",0xaa);
 910         &vpshufd        ("xmm7","xmm7",0xff);
 911         &vmovdqa        (&QWP(16*8-128,"ebp"),"xmm0");
 912         &vmovdqa        (&QWP(16*9-128,"ebp"),"xmm1");
 913         &vmovdqa        (&QWP(16*10-128,"ebp"),"xmm2");
 914         &vmovdqa        (&QWP(16*11-128,"ebp"),"xmm3");
 915         &vmovdqa        (&QWP(16*0-128,"ebp"),"xmm4");
 916         &vmovdqa        (&QWP(16*1-128,"ebp"),"xmm5");
 917         &vmovdqa        (&QWP(16*2-128,"ebp"),"xmm6");
 918         &vmovdqa        (&QWP(16*3-128,"ebp"),"xmm7");
 919
 920         &lea            ($inp,&DWP(128,$inp));          # size optimization
 921         &lea            ($out,&DWP(128,$out));          # size optimization
 922         &jmp            (&label("outer_loop"));
 923
 924 &set_label("outer_loop",32);
 925         #&vmovdqa       ("xmm0",&QWP(16*0-128,"ebp"));  # copy key material
 926         &vmovdqa        ("xmm1",&QWP(16*1-128,"ebp"));
 927         &vmovdqa        ("xmm2",&QWP(16*2-128,"ebp"));
 928         &vmovdqa        ("xmm3",&QWP(16*3-128,"ebp"));
 929         #&vmovdqa       ("xmm4",&QWP(16*4-128,"ebp"));
 930         &vmovdqa        ("xmm5",&QWP(16*5-128,"ebp"));
 931         &vmovdqa        ("xmm6",&QWP(16*6-128,"ebp"));
 932         &vmovdqa        ("xmm7",&QWP(16*7-128,"ebp"));
 933         #&vmovdqa       (&QWP(16*0-128,"ebx"),"xmm0");
 934         &vmovdqa        (&QWP(16*1-128,"ebx"),"xmm1");
 935         &vmovdqa        (&QWP(16*2-128,"ebx"),"xmm2");
 936         &vmovdqa        (&QWP(16*3-128,"ebx"),"xmm3");
 937         #&vmovdqa       (&QWP(16*4-128,"ebx"),"xmm4");
 938         &vmovdqa        (&QWP(16*5-128,"ebx"),"xmm5");
 939         &vmovdqa        (&QWP(16*6-128,"ebx"),"xmm6");
 940         &vmovdqa        (&QWP(16*7-128,"ebx"),"xmm7");
 941         #&vmovdqa       ("xmm0",&QWP(16*8-128,"ebp"));
 942         #&vmovdqa       ("xmm1",&QWP(16*9-128,"ebp"));
 943         &vmovdqa        ("xmm2",&QWP(16*10-128,"ebp"));
 944         &vmovdqa        ("xmm3",&QWP(16*11-128,"ebp"));
 945         &vmovdqa        ("xmm4",&QWP(16*12-128,"ebp"));
 946         &vmovdqa        ("xmm5",&QWP(16*13-128,"ebp"));
 947         &vmovdqa        ("xmm6",&QWP(16*14-128,"ebp"));
 948         &vmovdqa        ("xmm7",&QWP(16*15-128,"ebp"));
 949         &vpaddd         ("xmm4","xmm4",&QWP(16*4,"eax"));       # counter value
 950         #&vmovdqa       (&QWP(16*8-128,"ebx"),"xmm0");
 951         #&vmovdqa       (&QWP(16*9-128,"ebx"),"xmm1");
 952         &vmovdqa        (&QWP(16*10-128,"ebx"),"xmm2");
 953         &vmovdqa        (&QWP(16*11-128,"ebx"),"xmm3");
 954         &vmovdqa        (&QWP(16*12-128,"ebx"),"xmm4");
 955         &vmovdqa        (&QWP(16*13-128,"ebx"),"xmm5");
 956         &vmovdqa        (&QWP(16*14-128,"ebx"),"xmm6");
 957         &vmovdqa        (&QWP(16*15-128,"ebx"),"xmm7");
 958         &vmovdqa        (&QWP(16*12-128,"ebp"),"xmm4"); # save counter value
 959
 960         &vmovdqa        ($xa, &QWP(16*0-128,"ebp"));
 961         &vmovdqa        ($xd, "xmm4");
 962         &vmovdqa        ($xb_,&QWP(16*4-128,"ebp"));
 963         &vmovdqa        ($xc, &QWP(16*8-128,"ebp"));
 964         &vmovdqa        ($xc_,&QWP(16*9-128,"ebp"));
 965
 966         &mov            ("edx",10);                     # loop counter
 967         &nop            ();
 968
 969 &set_label("loop",32);
 970         &vpaddd         ($xa,$xa,$xb_);                 # elsewhere
 971         &vpxor          ($xd,$xd,$xa);                  # elsewhere
 972         &QUARTERROUND_XOP(0, 4, 8, 12, 0);
 973         &QUARTERROUND_XOP(1, 5, 9, 13, 1);
 974         &QUARTERROUND_XOP(2, 6,10, 14, 2);
 975         &QUARTERROUND_XOP(3, 7,11, 15, 3);
 976         &QUARTERROUND_XOP(0, 5,10, 15, 4);
 977         &QUARTERROUND_XOP(1, 6,11, 12, 5);
 978         &QUARTERROUND_XOP(2, 7, 8, 13, 6);
 979         &QUARTERROUND_XOP(3, 4, 9, 14, 7);
 980         &dec            ("edx");
 981         &jnz            (&label("loop"));
 982
 983         &vmovdqa        (&QWP(16*4-128,"ebx"),$xb_);
 984         &vmovdqa        (&QWP(16*8-128,"ebx"),$xc);
 985         &vmovdqa        (&QWP(16*9-128,"ebx"),$xc_);
 986         &vmovdqa        (&QWP(16*12-128,"ebx"),$xd);
 987         &vmovdqa        (&QWP(16*14-128,"ebx"),$xd_);
 988
 989     my ($xa0,$xa1,$xa2,$xa3,$xt0,$xt1,$xt2,$xt3)=map("xmm$_",(0..7));
 990
 991         #&vmovdqa       ($xa0,&QWP(16*0-128,"ebx"));    # it's there
 992         &vmovdqa        ($xa1,&QWP(16*1-128,"ebx"));
 993         &vmovdqa        ($xa2,&QWP(16*2-128,"ebx"));
 994         &vmovdqa        ($xa3,&QWP(16*3-128,"ebx"));
 995
 996     for($i=0;$i<256;$i+=64) {
 997         &vpaddd         ($xa0,$xa0,&QWP($i+16*0-128,"ebp"));    # accumulate key material
 998         &vpaddd         ($xa1,$xa1,&QWP($i+16*1-128,"ebp"));
 999         &vpaddd         ($xa2,$xa2,&QWP($i+16*2-128,"ebp"));
1000         &vpaddd         ($xa3,$xa3,&QWP($i+16*3-128,"ebp"));
1001
1002         &vpunpckldq     ($xt2,$xa0,$xa1);       # "de-interlace" data
1003         &vpunpckldq     ($xt3,$xa2,$xa3);
1004         &vpunpckhdq     ($xa0,$xa0,$xa1);
1005         &vpunpckhdq     ($xa2,$xa2,$xa3);
1006         &vpunpcklqdq    ($xa1,$xt2,$xt3);       # "a0"
1007         &vpunpckhqdq    ($xt2,$xt2,$xt3);       # "a1"
1008         &vpunpcklqdq    ($xt3,$xa0,$xa2);       # "a2"
1009         &vpunpckhqdq    ($xa3,$xa0,$xa2);       # "a3"
1010
1011         &vpxor          ($xt0,$xa1,&QWP(64*0-128,$inp));
1012         &vpxor          ($xt1,$xt2,&QWP(64*1-128,$inp));
1013         &vpxor          ($xt2,$xt3,&QWP(64*2-128,$inp));
1014         &vpxor          ($xt3,$xa3,&QWP(64*3-128,$inp));
1015         &lea            ($inp,&QWP($i<192?16:(64*4-16*3),$inp));
1016         &vmovdqa        ($xa0,&QWP($i+16*4-128,"ebx"))  if ($i<192);
1017         &vmovdqa        ($xa1,&QWP($i+16*5-128,"ebx"))  if ($i<192);
1018         &vmovdqa        ($xa2,&QWP($i+16*6-128,"ebx"))  if ($i<192);
1019         &vmovdqa        ($xa3,&QWP($i+16*7-128,"ebx"))  if ($i<192);
1020         &vmovdqu        (&QWP(64*0-128,$out),$xt0);     # store output
1021         &vmovdqu        (&QWP(64*1-128,$out),$xt1);
1022         &vmovdqu        (&QWP(64*2-128,$out),$xt2);
1023         &vmovdqu        (&QWP(64*3-128,$out),$xt3);
1024         &lea            ($out,&QWP($i<192?16:(64*4-16*3),$out));
1025     }
1026         &sub            ($len,64*4);
1027         &jnc            (&label("outer_loop"));
1028
1029         &add            ($len,64*4);
1030         &jz             (&label("done"));
1031
1032         &mov            ("ebx",&DWP(512+8,"esp"));      # restore pointers
1033         &lea            ($inp,&DWP(-128,$inp));
1034         &mov            ("edx",&DWP(512+4,"esp"));
1035         &lea            ($out,&DWP(-128,$out));
1036
1037         &vmovd          ("xmm2",&DWP(16*12-128,"ebp")); # counter value
1038         &vmovdqu        ("xmm3",&QWP(0,"ebx"));
1039         &vpaddd         ("xmm2","xmm2",&QWP(16*6,"eax"));# +four
1040         &vpand          ("xmm3","xmm3",&QWP(16*7,"eax"));
1041         &vpor           ("xmm3","xmm3","xmm2");         # counter value
1042 {
1043 my ($a,$b,$c,$d,$t,$t1,$rot16,$rot24)=map("xmm$_",(0..7));
1044
1045 sub XOPROUND {
1046         &vpaddd         ($a,$a,$b);
1047         &vpxor          ($d,$d,$a);
1048         &vprotd         ($d,$d,16);
1049
1050         &vpaddd         ($c,$c,$d);
1051         &vpxor          ($b,$b,$c);
1052         &vprotd         ($b,$b,12);
1053
1054         &vpaddd         ($a,$a,$b);
1055         &vpxor          ($d,$d,$a);
1056         &vprotd         ($d,$d,8);
1057
1058         &vpaddd         ($c,$c,$d);
1059         &vpxor          ($b,$b,$c);
1060         &vprotd         ($b,$b,7);
1061 }
1062
1063 &set_label("1x");
1064         &vmovdqa        ($a,&QWP(16*2,"eax"));          # sigma
1065         &vmovdqu        ($b,&QWP(0,"edx"));
1066         &vmovdqu        ($c,&QWP(16,"edx"));
1067         #&vmovdqu       ($d,&QWP(0,"ebx"));             # already loaded
1068         &vmovdqa        ($rot16,&QWP(0,"eax"));
1069         &vmovdqa        ($rot24,&QWP(16,"eax"));
1070         &mov            (&DWP(16*3,"esp"),"ebp");
1071
1072         &vmovdqa        (&QWP(16*0,"esp"),$a);
1073         &vmovdqa        (&QWP(16*1,"esp"),$b);
1074         &vmovdqa        (&QWP(16*2,"esp"),$c);
1075         &vmovdqa        (&QWP(16*3,"esp"),$d);
1076         &mov            ("edx",10);
1077         &jmp            (&label("loop1x"));
1078
1079 &set_label("outer1x",16);
1080         &vmovdqa        ($d,&QWP(16*5,"eax"));          # one
1081         &vmovdqa        ($a,&QWP(16*0,"esp"));
1082         &vmovdqa        ($b,&QWP(16*1,"esp"));
1083         &vmovdqa        ($c,&QWP(16*2,"esp"));
1084         &vpaddd         ($d,$d,&QWP(16*3,"esp"));
1085         &mov            ("edx",10);
1086         &vmovdqa        (&QWP(16*3,"esp"),$d);
1087         &jmp            (&label("loop1x"));
1088
1089 &set_label("loop1x",16);
1090         &XOPROUND();
1091         &vpshufd        ($c,$c,0b01001110);
1092         &vpshufd        ($b,$b,0b00111001);
1093         &vpshufd        ($d,$d,0b10010011);
1094
1095         &XOPROUND();
1096         &vpshufd        ($c,$c,0b01001110);
1097         &vpshufd        ($b,$b,0b10010011);
1098         &vpshufd        ($d,$d,0b00111001);
1099
1100         &dec            ("edx");
1101         &jnz            (&label("loop1x"));
1102
1103         &vpaddd         ($a,$a,&QWP(16*0,"esp"));
1104         &vpaddd         ($b,$b,&QWP(16*1,"esp"));
1105         &vpaddd         ($c,$c,&QWP(16*2,"esp"));
1106         &vpaddd         ($d,$d,&QWP(16*3,"esp"));
1107
1108         &cmp            ($len,64);
1109         &jb             (&label("tail"));
1110
1111         &vpxor          ($a,$a,&QWP(16*0,$inp));        # xor with input
1112         &vpxor          ($b,$b,&QWP(16*1,$inp));
1113         &vpxor          ($c,$c,&QWP(16*2,$inp));
1114         &vpxor          ($d,$d,&QWP(16*3,$inp));
1115         &lea            ($inp,&DWP(16*4,$inp));         # inp+=64
1116
1117         &vmovdqu        (&QWP(16*0,$out),$a);           # write output
1118         &vmovdqu        (&QWP(16*1,$out),$b);
1119         &vmovdqu        (&QWP(16*2,$out),$c);
1120         &vmovdqu        (&QWP(16*3,$out),$d);
1121         &lea            ($out,&DWP(16*4,$out));         # inp+=64
1122
1123         &sub            ($len,64);
1124         &jnz            (&label("outer1x"));
1125
1126         &jmp            (&label("done"));
1127
1128 &set_label("tail");
1129         &vmovdqa        (&QWP(16*0,"esp"),$a);
1130         &vmovdqa        (&QWP(16*1,"esp"),$b);
1131         &vmovdqa        (&QWP(16*2,"esp"),$c);
1132         &vmovdqa        (&QWP(16*3,"esp"),$d);
1133
1134         &xor            ("eax","eax");
1135         &xor            ("edx","edx");
1136         &xor            ("ebp","ebp");
1137
1138 &set_label("tail_loop");
1139         &movb           ("al",&BP(0,"esp","ebp"));
1140         &movb           ("dl",&BP(0,$inp,"ebp"));
1141         &lea            ("ebp",&DWP(1,"ebp"));
1142         &xor            ("al","dl");
1143         &movb           (&BP(-1,$out,"ebp"),"al");
1144         &dec            ($len);
1145         &jnz            (&label("tail_loop"));
1146 }
1147 &set_label("done");
1148         &vzeroupper     ();
1149         &mov            ("esp",&DWP(512,"esp"));
1150 &function_end("ChaCha20_xop");
1151 }
1152
1153 &asm_finish();
1154
1155 close STDOUT;