crypto/chacha/asm/chacha-x86.pl

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