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