crypto/x86cpuid.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2004-2016 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 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  10 push(@INC, "${dir}perlasm", "perlasm");
  11 require "x86asm.pl";
  12
  13 $output = pop;
  14 open OUT,">$output";
  15 *STDOUT=*OUT;
  16
  17 &asm_init($ARGV[0]);
  18
  19 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
  20
  21 &function_begin("OPENSSL_ia32_cpuid");
  22         &xor    ("edx","edx");
  23         &pushf  ();
  24         &pop    ("eax");
  25         &mov    ("ecx","eax");
  26         &xor    ("eax",1<<21);
  27         &push   ("eax");
  28         &popf   ();
  29         &pushf  ();
  30         &pop    ("eax");
  31         &xor    ("ecx","eax");
  32         &xor    ("eax","eax");
  33         &mov    ("esi",&wparam(0));
  34         &mov    (&DWP(8,"esi"),"eax");  # clear extended feature flags
  35         &bt     ("ecx",21);
  36         &jnc    (&label("nocpuid"));
  37         &cpuid  ();
  38         &mov    ("edi","eax");          # max value for standard query level
  39
  40         &xor    ("eax","eax");
  41         &cmp    ("ebx",0x756e6547);     # "Genu"
  42         &setne  (&LB("eax"));
  43         &mov    ("ebp","eax");
  44         &cmp    ("edx",0x49656e69);     # "ineI"
  45         &setne  (&LB("eax"));
  46         &or     ("ebp","eax");
  47         &cmp    ("ecx",0x6c65746e);     # "ntel"
  48         &setne  (&LB("eax"));
  49         &or     ("ebp","eax");          # 0 indicates Intel CPU
  50         &jz     (&label("intel"));
  51
  52         &cmp    ("ebx",0x68747541);     # "Auth"
  53         &setne  (&LB("eax"));
  54         &mov    ("esi","eax");
  55         &cmp    ("edx",0x69746E65);     # "enti"
  56         &setne  (&LB("eax"));
  57         &or     ("esi","eax");
  58         &cmp    ("ecx",0x444D4163);     # "cAMD"
  59         &setne  (&LB("eax"));
  60         &or     ("esi","eax");          # 0 indicates AMD CPU
  61         &jnz    (&label("intel"));
  62
  63         # AMD specific
  64         &mov    ("eax",0x80000000);
  65         &cpuid  ();
  66         &cmp    ("eax",0x80000001);
  67         &jb     (&label("intel"));
  68         &mov    ("esi","eax");
  69         &mov    ("eax",0x80000001);
  70         &cpuid  ();
  71         &or     ("ebp","ecx");
  72         &and    ("ebp",1<<11|1);        # isolate XOP bit
  73         &cmp    ("esi",0x80000008);
  74         &jb     (&label("intel"));
  75
  76         &mov    ("eax",0x80000008);
  77         &cpuid  ();
  78         &movz   ("esi",&LB("ecx"));     # number of cores - 1
  79         &inc    ("esi");                # number of cores
  80
  81         &mov    ("eax",1);
  82         &xor    ("ecx","ecx");
  83         &cpuid  ();
  84         &bt     ("edx",28);
  85         &jnc    (&label("generic"));
  86         &shr    ("ebx",16);
  87         &and    ("ebx",0xff);
  88         &cmp    ("ebx","esi");
  89         &ja     (&label("generic"));
  90         &and    ("edx",0xefffffff);     # clear hyper-threading bit
  91         &jmp    (&label("generic"));
  92
  93 &set_label("intel");
  94         &cmp    ("edi",4);
  95         &mov    ("esi",-1);
  96         &jb     (&label("nocacheinfo"));
  97
  98         &mov    ("eax",4);
  99         &mov    ("ecx",0);              # query L1D
 100         &cpuid  ();
 101         &mov    ("esi","eax");
 102         &shr    ("esi",14);
 103         &and    ("esi",0xfff);          # number of cores -1 per L1D
 104
 105 &set_label("nocacheinfo");
 106         &mov    ("eax",1);
 107         &xor    ("ecx","ecx");
 108         &cpuid  ();
 109         &and    ("edx",0xbfefffff);     # force reserved bits #20, #30 to 0
 110         &cmp    ("ebp",0);
 111         &jne    (&label("notintel"));
 112         &or     ("edx",1<<30);          # set reserved bit#30 on Intel CPUs
 113         &and    (&HB("eax"),15);        # familiy ID
 114         &cmp    (&HB("eax"),15);        # P4?
 115         &jne    (&label("notintel"));
 116         &or     ("edx",1<<20);          # set reserved bit#20 to engage RC4_CHAR
 117 &set_label("notintel");
 118         &bt     ("edx",28);             # test hyper-threading bit
 119         &jnc    (&label("generic"));
 120         &and    ("edx",0xefffffff);
 121         &cmp    ("esi",0);
 122         &je     (&label("generic"));
 123
 124         &or     ("edx",0x10000000);
 125         &shr    ("ebx",16);
 126         &cmp    (&LB("ebx"),1);
 127         &ja     (&label("generic"));
 128         &and    ("edx",0xefffffff);     # clear hyper-threading bit if not
 129
 130 &set_label("generic");
 131         &and    ("ebp",1<<11);          # isolate AMD XOP flag
 132         &and    ("ecx",0xfffff7ff);     # force 11th bit to 0
 133         &mov    ("esi","edx");          # %ebp:%esi is copy of %ecx:%edx
 134         &or     ("ebp","ecx");          # merge AMD XOP flag
 135
 136         &cmp    ("edi",7);
 137         &mov    ("edi",&wparam(0));
 138         &jb     (&label("no_extended_info"));
 139         &mov    ("eax",7);
 140         &xor    ("ecx","ecx");
 141         &cpuid  ();
 142         &mov    (&DWP(8,"edi"),"ebx");  # save extended feature flag
 143 &set_label("no_extended_info");
 144
 145         &bt     ("ebp",27);             # check OSXSAVE bit
 146         &jnc    (&label("clear_avx"));
 147         &xor    ("ecx","ecx");
 148         &data_byte(0x0f,0x01,0xd0);     # xgetbv
 149         &and    ("eax",6);
 150         &cmp    ("eax",6);
 151         &je     (&label("done"));
 152         &cmp    ("eax",2);
 153         &je     (&label("clear_avx"));
 154 &set_label("clear_xmm");
 155         &and    ("ebp",0xfdfffffd);     # clear AESNI and PCLMULQDQ bits
 156         &and    ("esi",0xfeffffff);     # clear FXSR
 157 &set_label("clear_avx");
 158         &and    ("ebp",0xefffe7ff);     # clear AVX, FMA and AMD XOP bits
 159         &and    (&DWP(8,"edi"),0xffffffdf);     # clear AVX2
 160 &set_label("done");
 161         &mov    ("eax","esi");
 162         &mov    ("edx","ebp");
 163 &set_label("nocpuid");
 164 &function_end("OPENSSL_ia32_cpuid");
 165
 166 &external_label("OPENSSL_ia32cap_P");
 167
 168 &function_begin_B("OPENSSL_rdtsc","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
 169         &xor    ("eax","eax");
 170         &xor    ("edx","edx");
 171         &picmeup("ecx","OPENSSL_ia32cap_P");
 172         &bt     (&DWP(0,"ecx"),4);
 173         &jnc    (&label("notsc"));
 174         &rdtsc  ();
 175 &set_label("notsc");
 176         &ret    ();
 177 &function_end_B("OPENSSL_rdtsc");
 178
 179 # This works in Ring 0 only [read DJGPP+MS-DOS+privileged DPMI host],
 180 # but it's safe to call it on any [supported] 32-bit platform...
 181 # Just check for [non-]zero return value...
 182 &function_begin_B("OPENSSL_instrument_halt","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
 183         &picmeup("ecx","OPENSSL_ia32cap_P");
 184         &bt     (&DWP(0,"ecx"),4);
 185         &jnc    (&label("nohalt"));     # no TSC
 186
 187         &data_word(0x9058900e);         # push %cs; pop %eax
 188         &and    ("eax",3);
 189         &jnz    (&label("nohalt"));     # not enough privileges
 190
 191         &pushf  ();
 192         &pop    ("eax");
 193         &bt     ("eax",9);
 194         &jnc    (&label("nohalt"));     # interrupts are disabled
 195
 196         &rdtsc  ();
 197         &push   ("edx");
 198         &push   ("eax");
 199         &halt   ();
 200         &rdtsc  ();
 201
 202         &sub    ("eax",&DWP(0,"esp"));
 203         &sbb    ("edx",&DWP(4,"esp"));
 204         &add    ("esp",8);
 205         &ret    ();
 206
 207 &set_label("nohalt");
 208         &xor    ("eax","eax");
 209         &xor    ("edx","edx");
 210         &ret    ();
 211 &function_end_B("OPENSSL_instrument_halt");
 212
 213 # Essentially there is only one use for this function. Under DJGPP:
 214 #
 215 #       #include <go32.h>
 216 #       ...
 217 #       i=OPENSSL_far_spin(_dos_ds,0x46c);
 218 #       ...
 219 # to obtain the number of spins till closest timer interrupt.
 220
 221 &function_begin_B("OPENSSL_far_spin");
 222         &pushf  ();
 223         &pop    ("eax");
 224         &bt     ("eax",9);
 225         &jnc    (&label("nospin"));     # interrupts are disabled
 226
 227         &mov    ("eax",&DWP(4,"esp"));
 228         &mov    ("ecx",&DWP(8,"esp"));
 229         &data_word (0x90d88e1e);        # push %ds, mov %eax,%ds
 230         &xor    ("eax","eax");
 231         &mov    ("edx",&DWP(0,"ecx"));
 232         &jmp    (&label("spin"));
 233
 234         &align  (16);
 235 &set_label("spin");
 236         &inc    ("eax");
 237         &cmp    ("edx",&DWP(0,"ecx"));
 238         &je     (&label("spin"));
 239
 240         &data_word (0x1f909090);        # pop   %ds
 241         &ret    ();
 242
 243 &set_label("nospin");
 244         &xor    ("eax","eax");
 245         &xor    ("edx","edx");
 246         &ret    ();
 247 &function_end_B("OPENSSL_far_spin");
 248
 249 &function_begin_B("OPENSSL_wipe_cpu","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
 250         &xor    ("eax","eax");
 251         &xor    ("edx","edx");
 252         &picmeup("ecx","OPENSSL_ia32cap_P");
 253         &mov    ("ecx",&DWP(0,"ecx"));
 254         &bt     (&DWP(0,"ecx"),1);
 255         &jnc    (&label("no_x87"));
 256         if ($sse2) {
 257                 &and    ("ecx",1<<26|1<<24);    # check SSE2 and FXSR bits
 258                 &cmp    ("ecx",1<<26|1<<24);
 259                 &jne    (&label("no_sse2"));
 260                 &pxor   ("xmm0","xmm0");
 261                 &pxor   ("xmm1","xmm1");
 262                 &pxor   ("xmm2","xmm2");
 263                 &pxor   ("xmm3","xmm3");
 264                 &pxor   ("xmm4","xmm4");
 265                 &pxor   ("xmm5","xmm5");
 266                 &pxor   ("xmm6","xmm6");
 267                 &pxor   ("xmm7","xmm7");
 268         &set_label("no_sse2");
 269         }
 270         # just a bunch of fldz to zap the fp/mm bank followed by finit...
 271         &data_word(0xeed9eed9,0xeed9eed9,0xeed9eed9,0xeed9eed9,0x90e3db9b);
 272 &set_label("no_x87");
 273         &lea    ("eax",&DWP(4,"esp"));
 274         &ret    ();
 275 &function_end_B("OPENSSL_wipe_cpu");
 276
 277 &function_begin_B("OPENSSL_atomic_add");
 278         &mov    ("edx",&DWP(4,"esp"));  # fetch the pointer, 1st arg
 279         &mov    ("ecx",&DWP(8,"esp"));  # fetch the increment, 2nd arg
 280         &push   ("ebx");
 281         &nop    ();
 282         &mov    ("eax",&DWP(0,"edx"));
 283 &set_label("spin");
 284         &lea    ("ebx",&DWP(0,"eax","ecx"));
 285         &nop    ();
 286         &data_word(0x1ab10ff0); # lock; cmpxchg %ebx,(%edx)     # %eax is involved and is always reloaded
 287         &jne    (&label("spin"));
 288         &mov    ("eax","ebx");  # OpenSSL expects the new value
 289         &pop    ("ebx");
 290         &ret    ();
 291 &function_end_B("OPENSSL_atomic_add");
 292
 293 &function_begin_B("OPENSSL_cleanse");
 294         &mov    ("edx",&wparam(0));
 295         &mov    ("ecx",&wparam(1));
 296         &xor    ("eax","eax");
 297         &cmp    ("ecx",7);
 298         &jae    (&label("lot"));
 299         &cmp    ("ecx",0);
 300         &je     (&label("ret"));
 301 &set_label("little");
 302         &mov    (&BP(0,"edx"),"al");
 303         &sub    ("ecx",1);
 304         &lea    ("edx",&DWP(1,"edx"));
 305         &jnz    (&label("little"));
 306 &set_label("ret");
 307         &ret    ();
 308
 309 &set_label("lot",16);
 310         &test   ("edx",3);
 311         &jz     (&label("aligned"));
 312         &mov    (&BP(0,"edx"),"al");
 313         &lea    ("ecx",&DWP(-1,"ecx"));
 314         &lea    ("edx",&DWP(1,"edx"));
 315         &jmp    (&label("lot"));
 316 &set_label("aligned");
 317         &mov    (&DWP(0,"edx"),"eax");
 318         &lea    ("ecx",&DWP(-4,"ecx"));
 319         &test   ("ecx",-4);
 320         &lea    ("edx",&DWP(4,"edx"));
 321         &jnz    (&label("aligned"));
 322         &cmp    ("ecx",0);
 323         &jne    (&label("little"));
 324         &ret    ();
 325 &function_end_B("OPENSSL_cleanse");
 326
 327 &function_begin_B("CRYPTO_memcmp");
 328         &push   ("esi");
 329         &push   ("edi");
 330         &mov    ("esi",&wparam(0));
 331         &mov    ("edi",&wparam(1));
 332         &mov    ("ecx",&wparam(2));
 333         &xor    ("eax","eax");
 334         &xor    ("edx","edx");
 335         &cmp    ("ecx",0);
 336         &je     (&label("no_data"));
 337 &set_label("loop");
 338         &mov    ("dl",&BP(0,"esi"));
 339         &lea    ("esi",&DWP(1,"esi"));
 340         &xor    ("dl",&BP(0,"edi"));
 341         &lea    ("edi",&DWP(1,"edi"));
 342         &or     ("al","dl");
 343         &dec    ("ecx");
 344         &jnz    (&label("loop"));
 345         &neg    ("eax");
 346         &shr    ("eax",31);
 347 &set_label("no_data");
 348         &pop    ("edi");
 349         &pop    ("esi");
 350         &ret    ();
 351 &function_end_B("CRYPTO_memcmp");
 352 {
 353 my $lasttick = "esi";
 354 my $lastdiff = "ebx";
 355 my $out = "edi";
 356 my $cnt = "ecx";
 357 my $max = "ebp";
 358
 359 &function_begin("OPENSSL_instrument_bus");
 360     &mov        ("eax",0);
 361     if ($sse2) {
 362         &picmeup("edx","OPENSSL_ia32cap_P");
 363         &bt     (&DWP(0,"edx"),4);
 364         &jnc    (&label("nogo"));       # no TSC
 365         &bt     (&DWP(0,"edx"),19);
 366         &jnc    (&label("nogo"));       # no CLFLUSH
 367
 368         &mov    ($out,&wparam(0));      # load arguments
 369         &mov    ($cnt,&wparam(1));
 370
 371         # collect 1st tick
 372         &rdtsc  ();
 373         &mov    ($lasttick,"eax");      # lasttick = tick
 374         &mov    ($lastdiff,0);          # lastdiff = 0
 375         &clflush(&DWP(0,$out));
 376         &data_byte(0xf0);               # lock
 377         &add    (&DWP(0,$out),$lastdiff);
 378         &jmp    (&label("loop"));
 379
 380 &set_label("loop",16);
 381         &rdtsc  ();
 382         &mov    ("edx","eax");          # put aside tick (yes, I neglect edx)
 383         &sub    ("eax",$lasttick);      # diff
 384         &mov    ($lasttick,"edx");      # lasttick = tick
 385         &mov    ($lastdiff,"eax");      # lastdiff = diff
 386         &clflush(&DWP(0,$out));
 387         &data_byte(0xf0);               # lock
 388         &add    (&DWP(0,$out),"eax");   # accumulate diff
 389         &lea    ($out,&DWP(4,$out));    # ++$out
 390         &sub    ($cnt,1);               # --$cnt
 391         &jnz    (&label("loop"));
 392
 393         &mov    ("eax",&wparam(1));
 394 &set_label("nogo");
 395     }
 396 &function_end("OPENSSL_instrument_bus");
 397
 398 &function_begin("OPENSSL_instrument_bus2");
 399     &mov        ("eax",0);
 400     if ($sse2) {
 401         &picmeup("edx","OPENSSL_ia32cap_P");
 402         &bt     (&DWP(0,"edx"),4);
 403         &jnc    (&label("nogo"));       # no TSC
 404         &bt     (&DWP(0,"edx"),19);
 405         &jnc    (&label("nogo"));       # no CLFLUSH
 406
 407         &mov    ($out,&wparam(0));      # load arguments
 408         &mov    ($cnt,&wparam(1));
 409         &mov    ($max,&wparam(2));
 410
 411         &rdtsc  ();                     # collect 1st tick
 412         &mov    ($lasttick,"eax");      # lasttick = tick
 413         &mov    ($lastdiff,0);          # lastdiff = 0
 414
 415         &clflush(&DWP(0,$out));
 416         &data_byte(0xf0);               # lock
 417         &add    (&DWP(0,$out),$lastdiff);
 418
 419         &rdtsc  ();                     # collect 1st diff
 420         &mov    ("edx","eax");          # put aside tick (yes, I neglect edx)
 421         &sub    ("eax",$lasttick);      # diff
 422         &mov    ($lasttick,"edx");      # lasttick = tick
 423         &mov    ($lastdiff,"eax");      # lastdiff = diff
 424         &jmp    (&label("loop2"));
 425
 426 &set_label("loop2",16);
 427         &clflush(&DWP(0,$out));
 428         &data_byte(0xf0);               # lock
 429         &add    (&DWP(0,$out),"eax");   # accumulate diff
 430
 431         &sub    ($max,1);
 432         &jz     (&label("done2"));
 433
 434         &rdtsc  ();
 435         &mov    ("edx","eax");          # put aside tick (yes, I neglect edx)
 436         &sub    ("eax",$lasttick);      # diff
 437         &mov    ($lasttick,"edx");      # lasttick = tick
 438         &cmp    ("eax",$lastdiff);
 439         &mov    ($lastdiff,"eax");      # lastdiff = diff
 440         &mov    ("edx",0);
 441         &setne  ("dl");
 442         &sub    ($cnt,"edx");           # conditional --$cnt
 443         &lea    ($out,&DWP(0,$out,"edx",4));    # conditional ++$out
 444         &jnz    (&label("loop2"));
 445
 446 &set_label("done2");
 447         &mov    ("eax",&wparam(1));
 448         &sub    ("eax",$cnt);
 449 &set_label("nogo");
 450     }
 451 &function_end("OPENSSL_instrument_bus2");
 452 }
 453
 454 sub gen_random {
 455 my $rdop = shift;
 456 &function_begin_B("OPENSSL_ia32_${rdop}");
 457         &mov    ("ecx",8);
 458 &set_label("loop");
 459         &${rdop}("eax");
 460         &jc     (&label("break"));
 461         &loop   (&label("loop"));
 462 &set_label("break");
 463         &cmp    ("eax",0);
 464         &cmove  ("eax","ecx");
 465         &ret    ();
 466 &function_end_B("OPENSSL_ia32_${rdop}");
 467
 468 &function_begin_B("OPENSSL_ia32_${rdop}_bytes");
 469         &push   ("edi");
 470         &push   ("ebx");
 471         &xor    ("eax","eax");          # return value
 472         &mov    ("edi",&wparam(0));
 473         &mov    ("ebx",&wparam(1));
 474
 475         &cmp    ("ebx",0);
 476         &je     (&label("done"));
 477
 478         &mov    ("ecx",8);
 479 &set_label("loop");
 480         &${rdop}("edx");
 481         &jc     (&label("break"));
 482         &loop   (&label("loop"));
 483         &jmp    (&label("done"));
 484
 485 &set_label("break",16);
 486         &cmp    ("ebx",4);
 487         &jb     (&label("tail"));
 488         &mov    (&DWP(0,"edi"),"edx");
 489         &lea    ("edi",&DWP(4,"edi"));
 490         &add    ("eax",4);
 491         &sub    ("ebx",4);
 492         &jz     (&label("done"));
 493         &mov    ("ecx",8);
 494         &jmp    (&label("loop"));
 495
 496 &set_label("tail",16);
 497         &mov    (&BP(0,"edi"),"dl");
 498         &lea    ("edi",&DWP(1,"edi"));
 499         &inc    ("eax");
 500         &shr    ("edx",8);
 501         &dec    ("ebx");
 502         &jnz    (&label("tail"));
 503
 504 &set_label("done");
 505         &pop    ("ebx");
 506         &pop    ("edi");
 507         &ret    ();
 508 &function_end_B("OPENSSL_ia32_${rdop}_bytes");
 509 }
 510 &gen_random("rdrand");
 511 &gen_random("rdseed");
 512
 513 &initseg("OPENSSL_cpuid_setup");
 514
 515 &hidden("OPENSSL_cpuid_setup");
 516 &hidden("OPENSSL_ia32cap_P");
 517
 518 &asm_finish();
 519
 520 close STDOUT;