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