crypto/x86cpuid.pl

   1 #!/usr/bin/env perl
   2
   3 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
   4 push(@INC, "${dir}perlasm", "perlasm");
   5 require "x86asm.pl";
   6
   7 &asm_init($ARGV[0],"x86cpuid");
   8
   9 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
  10
  11 &function_begin("OPENSSL_ia32_cpuid");
  12         &xor    ("edx","edx");
  13         &pushf  ();
  14         &pop    ("eax");
  15         &mov    ("ecx","eax");
  16         &xor    ("eax",1<<21);
  17         &push   ("eax");
  18         &popf   ();
  19         &pushf  ();
  20         &pop    ("eax");
  21         &xor    ("ecx","eax");
  22         &xor    ("eax","eax");
  23         &bt     ("ecx",21);
  24         &jnc    (&label("nocpuid"));
  25         &mov    ("esi",&wparam(0));
  26         &mov    (&DWP(8,"esi"),"eax");  # clear 3rd word
  27         &cpuid  ();
  28         &mov    ("edi","eax");          # max value for standard query level
  29
  30         &xor    ("eax","eax");
  31         &cmp    ("ebx",0x756e6547);     # "Genu"
  32         &setne  (&LB("eax"));
  33         &mov    ("ebp","eax");
  34         &cmp    ("edx",0x49656e69);     # "ineI"
  35         &setne  (&LB("eax"));
  36         &or     ("ebp","eax");
  37         &cmp    ("ecx",0x6c65746e);     # "ntel"
  38         &setne  (&LB("eax"));
  39         &or     ("ebp","eax");          # 0 indicates Intel CPU
  40         &jz     (&label("intel"));
  41
  42         &cmp    ("ebx",0x68747541);     # "Auth"
  43         &setne  (&LB("eax"));
  44         &mov    ("esi","eax");
  45         &cmp    ("edx",0x69746E65);     # "enti"
  46         &setne  (&LB("eax"));
  47         &or     ("esi","eax");
  48         &cmp    ("ecx",0x444D4163);     # "cAMD"
  49         &setne  (&LB("eax"));
  50         &or     ("esi","eax");          # 0 indicates AMD CPU
  51         &jnz    (&label("intel"));
  52
  53         # AMD specific
  54         &mov    ("eax",0x80000000);
  55         &cpuid  ();
  56         &cmp    ("eax",0x80000001);
  57         &jb     (&label("intel"));
  58         &mov    ("esi","eax");
  59         &mov    ("eax",0x80000001);
  60         &cpuid  ();
  61         &or     ("ebp","ecx");
  62         &and    ("ebp",1<<11|1);        # isolate XOP bit
  63         &cmp    ("esi",0x80000008);
  64         &jb     (&label("intel"));
  65
  66         &mov    ("eax",0x80000008);
  67         &cpuid  ();
  68         &movz   ("esi",&LB("ecx"));     # number of cores - 1
  69         &inc    ("esi");                # number of cores
  70
  71         &mov    ("eax",1);
  72         &xor    ("ecx","ecx");
  73         &cpuid  ();
  74         &bt     ("edx",28);
  75         &jnc    (&label("generic"));
  76         &shr    ("ebx",16);
  77         &and    ("ebx",0xff);
  78         &cmp    ("ebx","esi");
  79         &ja     (&label("generic"));
  80         &and    ("edx",0xefffffff);     # clear hyper-threading bit
  81         &jmp    (&label("generic"));
  82
  83 &set_label("intel");
  84         &cmp    ("edi",4);
  85         &mov    ("edi",-1);
  86         &jb     (&label("nocacheinfo"));
  87
  88         &mov    ("eax",4);
  89         &mov    ("ecx",0);              # query L1D
  90         &cpuid  ();
  91         &mov    ("edi","eax");
  92         &shr    ("edi",14);
  93         &and    ("edi",0xfff);          # number of cores -1 per L1D
  94
  95         &cmp    ("edi",7);
  96         &jb     (&label("nocacheinfo"));
  97
  98         &mov    ("esi",&wparam(0));
  99         &mov    ("eax",7);
 100         &xor    ("ecx","ecx");
 101         &cpuid  ();
 102         &mov    (&DWP(8,"esi"),"ebx");
 103
 104 &set_label("nocacheinfo");
 105         &mov    ("eax",1);
 106         &xor    ("ecx","ecx");
 107         &cpuid  ();
 108         &and    ("edx",0xbfefffff);     # force reserved bits #20, #30 to 0
 109         &cmp    ("ebp",0);
 110         &jne    (&label("notintel"));
 111         &or     ("edx",1<<30);          # set reserved bit#30 on Intel CPUs
 112         &and    (&HB("eax"),15);        # familiy ID
 113         &cmp    (&HB("eax"),15);        # P4?
 114         &jne    (&label("notintel"));
 115         &or     ("edx",1<<20);          # set reserved bit#20 to engage RC4_CHAR
 116 &set_label("notintel");
 117         &bt     ("edx",28);             # test hyper-threading bit
 118         &jnc    (&label("generic"));
 119         &and    ("edx",0xefffffff);
 120         &cmp    ("edi",0);
 121         &je     (&label("generic"));
 122
 123         &or     ("edx",0x10000000);
 124         &shr    ("ebx",16);
 125         &cmp    (&LB("ebx"),1);
 126         &ja     (&label("generic"));
 127         &and    ("edx",0xefffffff);     # clear hyper-threading bit if not
 128
 129 &set_label("generic");
 130         &and    ("ebp",1<<11);          # isolate AMD XOP flag
 131         &and    ("ecx",0xfffff7ff);     # force 11th bit to 0
 132         &mov    ("esi","edx");
 133         &or     ("ebp","ecx");          # merge AMD XOP flag
 134
 135         &bt     ("ecx",27);             # check OSXSAVE bit
 136         &jnc    (&label("clear_avx"));
 137         &xor    ("ecx","ecx");
 138         &data_byte(0x0f,0x01,0xd0);     # xgetbv
 139         &and    ("eax",6);
 140         &cmp    ("eax",6);
 141         &je     (&label("done"));
 142         &cmp    ("eax",2);
 143         &je     (&label("clear_avx"));
 144 &set_label("clear_xmm");
 145         &and    ("ebp",0xfdfffffd);     # clear AESNI and PCLMULQDQ bits
 146         &and    ("esi",0xfeffffff);     # clear FXSR
 147 &set_label("clear_avx");
 148         &and    ("ebp",0xefffe7ff);     # clear AVX, FMA and AMD XOP bits
 149         &mov    ("edi",&wparam(0));
 150         &and    (&DWP(8,"edi"),0xffffffdf);     # clear AVX2
 151 &set_label("done");
 152         &mov    ("eax","esi");
 153         &mov    ("edx","ebp");
 154 &set_label("nocpuid");
 155 &function_end("OPENSSL_ia32_cpuid");
 156
 157 &external_label("OPENSSL_ia32cap_P");
 158
 159 &function_begin_B("OPENSSL_rdtsc","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
 160         &xor    ("eax","eax");
 161         &xor    ("edx","edx");
 162         &picmeup("ecx","OPENSSL_ia32cap_P");
 163         &bt     (&DWP(0,"ecx"),4);
 164         &jnc    (&label("notsc"));
 165         &rdtsc  ();
 166 &set_label("notsc");
 167         &ret    ();
 168 &function_end_B("OPENSSL_rdtsc");
 169
 170 # This works in Ring 0 only [read DJGPP+MS-DOS+privileged DPMI host],
 171 # but it's safe to call it on any [supported] 32-bit platform...
 172 # Just check for [non-]zero return value...
 173 &function_begin_B("OPENSSL_instrument_halt","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
 174         &picmeup("ecx","OPENSSL_ia32cap_P");
 175         &bt     (&DWP(0,"ecx"),4);
 176         &jnc    (&label("nohalt"));     # no TSC
 177
 178         &data_word(0x9058900e);         # push %cs; pop %eax
 179         &and    ("eax",3);
 180         &jnz    (&label("nohalt"));     # not enough privileges
 181
 182         &pushf  ();
 183         &pop    ("eax");
 184         &bt     ("eax",9);
 185         &jnc    (&label("nohalt"));     # interrupts are disabled
 186
 187         &rdtsc  ();
 188         &push   ("edx");
 189         &push   ("eax");
 190         &halt   ();
 191         &rdtsc  ();
 192
 193         &sub    ("eax",&DWP(0,"esp"));
 194         &sbb    ("edx",&DWP(4,"esp"));
 195         &add    ("esp",8);
 196         &ret    ();
 197
 198 &set_label("nohalt");
 199         &xor    ("eax","eax");
 200         &xor    ("edx","edx");
 201         &ret    ();
 202 &function_end_B("OPENSSL_instrument_halt");
 203
 204 # Essentially there is only one use for this function. Under DJGPP:
 205 #
 206 #       #include <go32.h>
 207 #       ...
 208 #       i=OPENSSL_far_spin(_dos_ds,0x46c);
 209 #       ...
 210 # to obtain the number of spins till closest timer interrupt.
 211
 212 &function_begin_B("OPENSSL_far_spin");
 213         &pushf  ();
 214         &pop    ("eax");
 215         &bt     ("eax",9);
 216         &jnc    (&label("nospin"));     # interrupts are disabled
 217
 218         &mov    ("eax",&DWP(4,"esp"));
 219         &mov    ("ecx",&DWP(8,"esp"));
 220         &data_word (0x90d88e1e);        # push %ds, mov %eax,%ds
 221         &xor    ("eax","eax");
 222         &mov    ("edx",&DWP(0,"ecx"));
 223         &jmp    (&label("spin"));
 224
 225         &align  (16);
 226 &set_label("spin");
 227         &inc    ("eax");
 228         &cmp    ("edx",&DWP(0,"ecx"));
 229         &je     (&label("spin"));
 230
 231         &data_word (0x1f909090);        # pop   %ds
 232         &ret    ();
 233
 234 &set_label("nospin");
 235         &xor    ("eax","eax");
 236         &xor    ("edx","edx");
 237         &ret    ();
 238 &function_end_B("OPENSSL_far_spin");
 239
 240 &function_begin_B("OPENSSL_wipe_cpu","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
 241         &xor    ("eax","eax");
 242         &xor    ("edx","edx");
 243         &picmeup("ecx","OPENSSL_ia32cap_P");
 244         &mov    ("ecx",&DWP(0,"ecx"));
 245         &bt     (&DWP(0,"ecx"),1);
 246         &jnc    (&label("no_x87"));
 247         if ($sse2) {
 248                 &and    ("ecx",1<<26|1<<24);    # check SSE2 and FXSR bits
 249                 &cmp    ("ecx",1<<26|1<<24);
 250                 &jne    (&label("no_sse2"));
 251                 &pxor   ("xmm0","xmm0");
 252                 &pxor   ("xmm1","xmm1");
 253                 &pxor   ("xmm2","xmm2");
 254                 &pxor   ("xmm3","xmm3");
 255                 &pxor   ("xmm4","xmm4");
 256                 &pxor   ("xmm5","xmm5");
 257                 &pxor   ("xmm6","xmm6");
 258                 &pxor   ("xmm7","xmm7");
 259         &set_label("no_sse2");
 260         }
 261         # just a bunch of fldz to zap the fp/mm bank followed by finit...
 262         &data_word(0xeed9eed9,0xeed9eed9,0xeed9eed9,0xeed9eed9,0x90e3db9b);
 263 &set_label("no_x87");
 264         &lea    ("eax",&DWP(4,"esp"));
 265         &ret    ();
 266 &function_end_B("OPENSSL_wipe_cpu");
 267
 268 &function_begin_B("OPENSSL_atomic_add");
 269         &mov    ("edx",&DWP(4,"esp"));  # fetch the pointer, 1st arg
 270         &mov    ("ecx",&DWP(8,"esp"));  # fetch the increment, 2nd arg
 271         &push   ("ebx");
 272         &nop    ();
 273         &mov    ("eax",&DWP(0,"edx"));
 274 &set_label("spin");
 275         &lea    ("ebx",&DWP(0,"eax","ecx"));
 276         &nop    ();
 277         &data_word(0x1ab10ff0); # lock; cmpxchg %ebx,(%edx)     # %eax is envolved and is always reloaded
 278         &jne    (&label("spin"));
 279         &mov    ("eax","ebx");  # OpenSSL expects the new value
 280         &pop    ("ebx");
 281         &ret    ();
 282 &function_end_B("OPENSSL_atomic_add");
 283
 284 # This function can become handy under Win32 in situations when
 285 # we don't know which calling convention, __stdcall or __cdecl(*),
 286 # indirect callee is using. In C it can be deployed as
 287 #
 288 #ifdef OPENSSL_CPUID_OBJ
 289 #       type OPENSSL_indirect_call(void *f,...);
 290 #       ...
 291 #       OPENSSL_indirect_call(func,[up to $max arguments]);
 292 #endif
 293 #
 294 # (*)   it's designed to work even for __fastcall if number of
 295 #       arguments is 1 or 2!
 296 &function_begin_B("OPENSSL_indirect_call");
 297         {
 298         my ($max,$i)=(7,);      # $max has to be chosen as 4*n-1
 299                                 # in order to preserve eventual
 300                                 # stack alignment
 301         &push   ("ebp");
 302         &mov    ("ebp","esp");
 303         &sub    ("esp",$max*4);
 304         &mov    ("ecx",&DWP(12,"ebp"));
 305         &mov    (&DWP(0,"esp"),"ecx");
 306         &mov    ("edx",&DWP(16,"ebp"));
 307         &mov    (&DWP(4,"esp"),"edx");
 308         for($i=2;$i<$max;$i++)
 309                 {
 310                 # Some copies will be redundant/bogus...
 311                 &mov    ("eax",&DWP(12+$i*4,"ebp"));
 312                 &mov    (&DWP(0+$i*4,"esp"),"eax");
 313                 }
 314         &call_ptr       (&DWP(8,"ebp"));# make the call...
 315         &mov    ("esp","ebp");  # ... and just restore the stack pointer
 316                                 # without paying attention to what we called,
 317                                 # (__cdecl *func) or (__stdcall *one).
 318         &pop    ("ebp");
 319         &ret    ();
 320         }
 321 &function_end_B("OPENSSL_indirect_call");
 322
 323 &function_begin_B("OPENSSL_cleanse");
 324         &mov    ("edx",&wparam(0));
 325         &mov    ("ecx",&wparam(1));
 326         &xor    ("eax","eax");
 327         &cmp    ("ecx",7);
 328         &jae    (&label("lot"));
 329         &cmp    ("ecx",0);
 330         &je     (&label("ret"));
 331 &set_label("little");
 332         &mov    (&BP(0,"edx"),"al");
 333         &sub    ("ecx",1);
 334         &lea    ("edx",&DWP(1,"edx"));
 335         &jnz    (&label("little"));
 336 &set_label("ret");
 337         &ret    ();
 338
 339 &set_label("lot",16);
 340         &test   ("edx",3);
 341         &jz     (&label("aligned"));
 342         &mov    (&BP(0,"edx"),"al");
 343         &lea    ("ecx",&DWP(-1,"ecx"));
 344         &lea    ("edx",&DWP(1,"edx"));
 345         &jmp    (&label("lot"));
 346 &set_label("aligned");
 347         &mov    (&DWP(0,"edx"),"eax");
 348         &lea    ("ecx",&DWP(-4,"ecx"));
 349         &test   ("ecx",-4);
 350         &lea    ("edx",&DWP(4,"edx"));
 351         &jnz    (&label("aligned"));
 352         &cmp    ("ecx",0);
 353         &jne    (&label("little"));
 354         &ret    ();
 355 &function_end_B("OPENSSL_cleanse");
 356
 357 {
 358 my $lasttick = "esi";
 359 my $lastdiff = "ebx";
 360 my $out = "edi";
 361 my $cnt = "ecx";
 362 my $max = "ebp";
 363
 364 &function_begin("OPENSSL_instrument_bus");
 365     &mov        ("eax",0);
 366     if ($sse2) {
 367         &picmeup("edx","OPENSSL_ia32cap_P");
 368         &bt     (&DWP(0,"edx"),4);
 369         &jnc    (&label("nogo"));       # no TSC
 370         &bt     (&DWP(0,"edx"),19);
 371         &jnc    (&label("nogo"));       # no CLFLUSH
 372
 373         &mov    ($out,&wparam(0));      # load arguments
 374         &mov    ($cnt,&wparam(1));
 375
 376         # collect 1st tick
 377         &rdtsc  ();
 378         &mov    ($lasttick,"eax");      # lasttick = tick
 379         &mov    ($lastdiff,0);          # lastdiff = 0
 380         &clflush(&DWP(0,$out));
 381         &data_byte(0xf0);               # lock
 382         &add    (&DWP(0,$out),$lastdiff);
 383         &jmp    (&label("loop"));
 384
 385 &set_label("loop",16);
 386         &rdtsc  ();
 387         &mov    ("edx","eax");          # put aside tick (yes, I neglect edx)
 388         &sub    ("eax",$lasttick);      # diff
 389         &mov    ($lasttick,"edx");      # lasttick = tick
 390         &mov    ($lastdiff,"eax");      # lastdiff = diff
 391         &clflush(&DWP(0,$out));
 392         &data_byte(0xf0);               # lock
 393         &add    (&DWP(0,$out),"eax");   # accumulate diff
 394         &lea    ($out,&DWP(4,$out));    # ++$out
 395         &sub    ($cnt,1);               # --$cnt
 396         &jnz    (&label("loop"));
 397
 398         &mov    ("eax",&wparam(1));
 399 &set_label("nogo");
 400     }
 401 &function_end("OPENSSL_instrument_bus");
 402
 403 &function_begin("OPENSSL_instrument_bus2");
 404     &mov        ("eax",0);
 405     if ($sse2) {
 406         &picmeup("edx","OPENSSL_ia32cap_P");
 407         &bt     (&DWP(0,"edx"),4);
 408         &jnc    (&label("nogo"));       # no TSC
 409         &bt     (&DWP(0,"edx"),19);
 410         &jnc    (&label("nogo"));       # no CLFLUSH
 411
 412         &mov    ($out,&wparam(0));      # load arguments
 413         &mov    ($cnt,&wparam(1));
 414         &mov    ($max,&wparam(2));
 415
 416         &rdtsc  ();                     # collect 1st tick
 417         &mov    ($lasttick,"eax");      # lasttick = tick
 418         &mov    ($lastdiff,0);          # lastdiff = 0
 419
 420         &clflush(&DWP(0,$out));
 421         &data_byte(0xf0);               # lock
 422         &add    (&DWP(0,$out),$lastdiff);
 423
 424         &rdtsc  ();                     # collect 1st diff
 425         &mov    ("edx","eax");          # put aside tick (yes, I neglect edx)
 426         &sub    ("eax",$lasttick);      # diff
 427         &mov    ($lasttick,"edx");      # lasttick = tick
 428         &mov    ($lastdiff,"eax");      # lastdiff = diff
 429         &jmp    (&label("loop2"));
 430
 431 &set_label("loop2",16);
 432         &clflush(&DWP(0,$out));
 433         &data_byte(0xf0);               # lock
 434         &add    (&DWP(0,$out),"eax");   # accumulate diff
 435
 436         &sub    ($max,1);
 437         &jz     (&label("done2"));
 438
 439         &rdtsc  ();
 440         &mov    ("edx","eax");          # put aside tick (yes, I neglect edx)
 441         &sub    ("eax",$lasttick);      # diff
 442         &mov    ($lasttick,"edx");      # lasttick = tick
 443         &cmp    ("eax",$lastdiff);
 444         &mov    ($lastdiff,"eax");      # lastdiff = diff
 445         &mov    ("edx",0);
 446         &setne  ("dl");
 447         &sub    ($cnt,"edx");           # conditional --$cnt
 448         &lea    ($out,&DWP(0,$out,"edx",4));    # conditional ++$out
 449         &jnz    (&label("loop2"));
 450
 451 &set_label("done2");
 452         &mov    ("eax",&wparam(1));
 453         &sub    ("eax",$cnt);
 454 &set_label("nogo");
 455     }
 456 &function_end("OPENSSL_instrument_bus2");
 457 }
 458
 459 &function_begin_B("OPENSSL_ia32_rdrand");
 460         &mov    ("ecx",8);
 461 &set_label("loop");
 462         &rdrand ("eax");
 463         &jc     (&label("break"));
 464         &loop   (&label("loop"));
 465 &set_label("break");
 466         &cmp    ("eax",0);
 467         &cmove  ("eax","ecx");
 468         &ret    ();
 469 &function_end_B("OPENSSL_ia32_rdrand");
 470
 471 &initseg("OPENSSL_cpuid_setup");
 472
 473 &hidden("OPENSSL_cpuid_setup");
 474 &hidden("OPENSSL_ia32cap_P");
 475
 476 &asm_finish();