crypto/sparccpuid.S

   1 #ifdef OPENSSL_FIPSCANISTER
   2 #include <openssl/fipssyms.h>
   3 #endif
   4
   5 #if defined(__SUNPRO_C) && defined(__sparcv9)
   6 # define ABI64  /* They've said -xarch=v9 at command line */
   7 #elif defined(__GNUC__) && defined(__arch64__)
   8 # define ABI64  /* They've said -m64 at command line */
   9 #endif
  10
  11 #ifdef ABI64
  12   .register     %g2,#scratch
  13   .register     %g3,#scratch
  14 # define        FRAME   -192
  15 # define        BIAS    2047
  16 #else
  17 # define        FRAME   -96
  18 # define        BIAS    0
  19 #endif
  20
  21 .text
  22 .align  32
  23 .global OPENSSL_wipe_cpu
  24 .type   OPENSSL_wipe_cpu,#function
  25 ! Keep in mind that this does not excuse us from wiping the stack!
  26 ! This routine wipes registers, but not the backing store [which
  27 ! resides on the stack, toward lower addresses]. To facilitate for
  28 ! stack wiping I return pointer to the top of stack of the *caller*.
  29 OPENSSL_wipe_cpu:
  30         save    %sp,FRAME,%sp
  31         nop
  32 #ifdef __sun
  33 #include <sys/trap.h>
  34         ta      ST_CLEAN_WINDOWS
  35 #else
  36         call    .walk.reg.wins
  37 #endif
  38         nop
  39         call    .PIC.zero.up
  40         mov     .zero-(.-4),%o0
  41         ld      [%o0],%f0
  42         ld      [%o0],%f1
  43
  44         subcc   %g0,1,%o0
  45         ! Following is V9 "rd %ccr,%o0" instruction. However! V8
  46         ! specification says that it ("rd %asr2,%o0" in V8 terms) does
  47         ! not cause illegal_instruction trap. It therefore can be used
  48         ! to determine if the CPU the code is executing on is V8- or
  49         ! V9-compliant, as V9 returns a distinct value of 0x99,
  50         ! "negative" and "borrow" bits set in both %icc and %xcc.
  51         .word   0x91408000      !rd     %ccr,%o0
  52         cmp     %o0,0x99
  53         bne     .v8
  54         nop
  55                         ! Even though we do not use %fp register bank,
  56                         ! we wipe it as memcpy might have used it...
  57                         .word   0xbfa00040      !fmovd  %f0,%f62
  58                         .word   0xbba00040      !...
  59                         .word   0xb7a00040
  60                         .word   0xb3a00040
  61                         .word   0xafa00040
  62                         .word   0xaba00040
  63                         .word   0xa7a00040
  64                         .word   0xa3a00040
  65                         .word   0x9fa00040
  66                         .word   0x9ba00040
  67                         .word   0x97a00040
  68                         .word   0x93a00040
  69                         .word   0x8fa00040
  70                         .word   0x8ba00040
  71                         .word   0x87a00040
  72                         .word   0x83a00040      !fmovd  %f0,%f32
  73 .v8:                    fmovs   %f1,%f31
  74         clr     %o0
  75                         fmovs   %f0,%f30
  76         clr     %o1
  77                         fmovs   %f1,%f29
  78         clr     %o2
  79                         fmovs   %f0,%f28
  80         clr     %o3
  81                         fmovs   %f1,%f27
  82         clr     %o4
  83                         fmovs   %f0,%f26
  84         clr     %o5
  85                         fmovs   %f1,%f25
  86         clr     %o7
  87                         fmovs   %f0,%f24
  88         clr     %l0
  89                         fmovs   %f1,%f23
  90         clr     %l1
  91                         fmovs   %f0,%f22
  92         clr     %l2
  93                         fmovs   %f1,%f21
  94         clr     %l3
  95                         fmovs   %f0,%f20
  96         clr     %l4
  97                         fmovs   %f1,%f19
  98         clr     %l5
  99                         fmovs   %f0,%f18
 100         clr     %l6
 101                         fmovs   %f1,%f17
 102         clr     %l7
 103                         fmovs   %f0,%f16
 104         clr     %i0
 105                         fmovs   %f1,%f15
 106         clr     %i1
 107                         fmovs   %f0,%f14
 108         clr     %i2
 109                         fmovs   %f1,%f13
 110         clr     %i3
 111                         fmovs   %f0,%f12
 112         clr     %i4
 113                         fmovs   %f1,%f11
 114         clr     %i5
 115                         fmovs   %f0,%f10
 116         clr     %g1
 117                         fmovs   %f1,%f9
 118         clr     %g2
 119                         fmovs   %f0,%f8
 120         clr     %g3
 121                         fmovs   %f1,%f7
 122         clr     %g4
 123                         fmovs   %f0,%f6
 124         clr     %g5
 125                         fmovs   %f1,%f5
 126                         fmovs   %f0,%f4
 127                         fmovs   %f1,%f3
 128                         fmovs   %f0,%f2
 129
 130         add     %fp,BIAS,%i0    ! return pointer to caller´s top of stack
 131
 132         ret
 133         restore
 134
 135 .zero:  .long   0x0,0x0
 136 .PIC.zero.up:
 137         retl
 138         add     %o0,%o7,%o0
 139 #ifdef DEBUG
 140 .global walk_reg_wins
 141 .type   walk_reg_wins,#function
 142 walk_reg_wins:
 143 #endif
 144 .walk.reg.wins:
 145         save    %sp,FRAME,%sp
 146         cmp     %i7,%o7
 147         be      2f
 148         clr     %o0
 149         cmp     %o7,0   ! compiler never cleans %o7...
 150         be      1f      ! could have been a leaf function...
 151         clr     %o1
 152         call    .walk.reg.wins
 153         nop
 154 1:      clr     %o2
 155         clr     %o3
 156         clr     %o4
 157         clr     %o5
 158         clr     %o7
 159         clr     %l0
 160         clr     %l1
 161         clr     %l2
 162         clr     %l3
 163         clr     %l4
 164         clr     %l5
 165         clr     %l6
 166         clr     %l7
 167         add     %o0,1,%i0       ! used for debugging
 168 2:      ret
 169         restore
 170 .size   OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
 171
 172 .global OPENSSL_atomic_add
 173 .type   OPENSSL_atomic_add,#function
 174 .align  32
 175 OPENSSL_atomic_add:
 176 #ifndef ABI64
 177         subcc   %g0,1,%o2
 178         .word   0x95408000      !rd     %ccr,%o2, see comment above
 179         cmp     %o2,0x99
 180         be      .v9
 181         nop
 182         save    %sp,FRAME,%sp
 183         ba      .enter
 184         nop
 185 #ifdef __sun
 186 ! Note that you do not have to link with libthread to call thr_yield,
 187 ! as libc provides a stub, which is overloaded the moment you link
 188 ! with *either* libpthread or libthread...
 189 #define YIELD_CPU       thr_yield
 190 #else
 191 ! applies at least to Linux and FreeBSD... Feedback expected...
 192 #define YIELD_CPU       sched_yield
 193 #endif
 194 .spin:  call    YIELD_CPU
 195         nop
 196 .enter: ld      [%i0],%i2
 197         cmp     %i2,-4096
 198         be      .spin
 199         mov     -1,%i2
 200         swap    [%i0],%i2
 201         cmp     %i2,-1
 202         be      .spin
 203         add     %i2,%i1,%i2
 204         stbar
 205         st      %i2,[%i0]
 206         sra     %i2,%g0,%i0
 207         ret
 208         restore
 209 .v9:
 210 #endif
 211         ld      [%o0],%o2
 212 1:      add     %o1,%o2,%o3
 213         .word   0xd7e2100a      !cas [%o0],%o2,%o3, compare [%o0] with %o2 and swap %o3
 214         cmp     %o2,%o3
 215         bne     1b
 216         mov     %o3,%o2         ! cas is always fetching to dest. register
 217         add     %o1,%o2,%o0     ! OpenSSL expects the new value
 218         retl
 219         sra     %o0,%g0,%o0     ! we return signed int, remember?
 220 .size   OPENSSL_atomic_add,.-OPENSSL_atomic_add
 221
 222 .global _sparcv9_rdtick
 223 .align  32
 224 _sparcv9_rdtick:
 225         subcc   %g0,1,%o0
 226         .word   0x91408000      !rd     %ccr,%o0
 227         cmp     %o0,0x99
 228         bne     .notick
 229         xor     %o0,%o0,%o0
 230         .word   0x91410000      !rd     %tick,%o0
 231         retl
 232         .word   0x93323020      !srlx   %o0,32,%o1
 233 .notick:
 234         retl
 235         xor     %o1,%o1,%o1
 236 .type   _sparcv9_rdtick,#function
 237 .size   _sparcv9_rdtick,.-_sparcv9_rdtick
 238
 239 .global _sparcv9_vis1_probe
 240 .align  8
 241 _sparcv9_vis1_probe:
 242         .word   0x81b00d80      !fxor   %f0,%f0,%f0
 243         add     %sp,BIAS+2,%o1
 244         retl
 245         .word   0xc19a5a40      !ldda   [%o1]ASI_FP16_P,%f0
 246 .type   _sparcv9_vis1_probe,#function
 247 .size   _sparcv9_vis1_probe,.-_sparcv9_vis1_probe
 248
 249 ! Probe and instrument VIS1 instruction. Output is number of cycles it
 250 ! takes to execute rdtick and pair of VIS1 instructions. US-Tx VIS unit
 251 ! is slow (documented to be 6 cycles on T2) and the core is in-order
 252 ! single-issue, it should be possible to distinguish Tx reliably...
 253 ! Observed return values are:
 254 !
 255 !       UltraSPARC IIe          7
 256 !       UltraSPARC III          7
 257 !       UltraSPARC T1           24
 258 !
 259 ! Numbers for T2 and SPARC64 V-VII are more than welcomed.
 260 !
 261 ! It would be possible to detect specifically US-T1 by instrumenting
 262 ! fmul8ulx16, which is emulated on T1 and as such accounts for quite
 263 ! a lot of %tick-s, couple of thousand on Linux...
 264 .global _sparcv9_vis1_instrument
 265 .align  8
 266 _sparcv9_vis1_instrument:
 267         .word   0x91410000      !rd     %tick,%o0
 268         .word   0x81b00d80      !fxor   %f0,%f0,%f0
 269         .word   0x85b08d82      !fxor   %f2,%f2,%f2
 270         .word   0x93410000      !rd     %tick,%o1
 271         .word   0x81b00d80      !fxor   %f0,%f0,%f0
 272         .word   0x85b08d82      !fxor   %f2,%f2,%f2
 273         .word   0x95410000      !rd     %tick,%o2
 274         .word   0x81b00d80      !fxor   %f0,%f0,%f0
 275         .word   0x85b08d82      !fxor   %f2,%f2,%f2
 276         .word   0x97410000      !rd     %tick,%o3
 277         .word   0x81b00d80      !fxor   %f0,%f0,%f0
 278         .word   0x85b08d82      !fxor   %f2,%f2,%f2
 279         .word   0x99410000      !rd     %tick,%o4
 280
 281         ! calculate intervals
 282         sub     %o1,%o0,%o0
 283         sub     %o2,%o1,%o1
 284         sub     %o3,%o2,%o2
 285         sub     %o4,%o3,%o3
 286
 287         ! find minumum value
 288         cmp     %o0,%o1
 289         .word   0x38680002      !bgu,a  %xcc,.+8
 290         mov     %o1,%o0
 291         cmp     %o0,%o2
 292         .word   0x38680002      !bgu,a  %xcc,.+8
 293         mov     %o2,%o0
 294         cmp     %o0,%o3
 295         .word   0x38680002      !bgu,a  %xcc,.+8
 296         mov     %o3,%o0
 297
 298         retl
 299         nop
 300 .type   _sparcv9_vis1_instrument,#function
 301 .size   _sparcv9_vis1_instrument,.-_sparcv9_vis1_instrument
 302
 303 .global _sparcv9_vis2_probe
 304 .align  8
 305 _sparcv9_vis2_probe:
 306         retl
 307         .word   0x81b00980      !bshuffle       %f0,%f0,%f0
 308 .type   _sparcv9_vis2_probe,#function
 309 .size   _sparcv9_vis2_probe,.-_sparcv9_vis2_probe
 310
 311 .global _sparcv9_fmadd_probe
 312 .align  8
 313 _sparcv9_fmadd_probe:
 314         .word   0x81b00d80      !fxor   %f0,%f0,%f0
 315         .word   0x85b08d82      !fxor   %f2,%f2,%f2
 316         retl
 317         .word   0x81b80440      !fmaddd %f0,%f0,%f2,%f0
 318 .type   _sparcv9_fmadd_probe,#function
 319 .size   _sparcv9_fmadd_probe,.-_sparcv9_fmadd_probe
 320
 321 .global OPENSSL_cleanse
 322 .align  32
 323 OPENSSL_cleanse:
 324         cmp     %o1,14
 325         nop
 326 #ifdef ABI64
 327         bgu     %xcc,.Lot
 328 #else
 329         bgu     .Lot
 330 #endif
 331         cmp     %o1,0
 332         bne     .Little
 333         nop
 334         retl
 335         nop
 336
 337 .Little:
 338         stb     %g0,[%o0]
 339         subcc   %o1,1,%o1
 340         bnz     .Little
 341         add     %o0,1,%o0
 342         retl
 343         nop
 344 .align  32
 345 .Lot:
 346 #ifndef ABI64
 347         subcc   %g0,1,%g1
 348         ! see above for explanation
 349         .word   0x83408000      !rd     %ccr,%g1
 350         cmp     %g1,0x99
 351         bne     .v8lot
 352         nop
 353 #endif
 354
 355 .v9lot: andcc   %o0,7,%g0
 356         bz      .v9aligned
 357         nop
 358         stb     %g0,[%o0]
 359         sub     %o1,1,%o1
 360         ba      .v9lot
 361         add     %o0,1,%o0
 362 .align  16,0x01000000
 363 .v9aligned:
 364         .word   0xc0720000      !stx    %g0,[%o0]
 365         sub     %o1,8,%o1
 366         andcc   %o1,-8,%g0
 367 #ifdef ABI64
 368         .word   0x126ffffd      !bnz    %xcc,.v9aligned
 369 #else
 370         .word   0x124ffffd      !bnz    %icc,.v9aligned
 371 #endif
 372         add     %o0,8,%o0
 373
 374         cmp     %o1,0
 375         bne     .Little
 376         nop
 377         retl
 378         nop
 379 #ifndef ABI64
 380 .v8lot: andcc   %o0,3,%g0
 381         bz      .v8aligned
 382         nop
 383         stb     %g0,[%o0]
 384         sub     %o1,1,%o1
 385         ba      .v8lot
 386         add     %o0,1,%o0
 387         nop
 388 .v8aligned:
 389         st      %g0,[%o0]
 390         sub     %o1,4,%o1
 391         andcc   %o1,-4,%g0
 392         bnz     .v8aligned
 393         add     %o0,4,%o0
 394
 395         cmp     %o1,0
 396         bne     .Little
 397         nop
 398         retl
 399         nop
 400 #endif
 401 .type   OPENSSL_cleanse,#function
 402 .size   OPENSSL_cleanse,.-OPENSSL_cleanse
 403
 404 .global _sparcv9_vis1_instrument_bus
 405 .align  8
 406 _sparcv9_vis1_instrument_bus:
 407         mov     %o1,%o3                                 ! save cnt
 408         .word   0x99410000      !rd     %tick,%o4       ! tick
 409         mov     %o4,%o5                                 ! lasttick = tick
 410         set     0,%g4                                   ! diff
 411
 412         andn    %o0,63,%g1
 413         .word   0xc1985e00      !ldda   [%g1]0xf0,%f0   ! block load
 414         .word   0x8143e040      !membar #Sync
 415         .word   0xc1b85c00      !stda   %f0,[%g1]0xe0   ! block store and commit
 416         .word   0x8143e040      !membar #Sync
 417         ld      [%o0],%o4
 418         add     %o4,%g4,%g4
 419         .word   0xc9e2100c      !cas    [%o0],%o4,%g4
 420
 421 .Loop:  .word   0x99410000      !rd     %tick,%o4
 422         sub     %o4,%o5,%g4                             ! diff=tick-lasttick
 423         mov     %o4,%o5                                 ! lasttick=tick
 424
 425         andn    %o0,63,%g1
 426         .word   0xc1985e00      !ldda   [%g1]0xf0,%f0   ! block load
 427         .word   0x8143e040      !membar #Sync
 428         .word   0xc1b85c00      !stda   %f0,[%g1]0xe0   ! block store and commit
 429         .word   0x8143e040      !membar #Sync
 430         ld      [%o0],%o4
 431         add     %o4,%g4,%g4
 432         .word   0xc9e2100c      !cas    [%o0],%o4,%g4
 433         subcc   %o1,1,%o1                               ! --$cnt
 434         bnz     .Loop
 435         add     %o0,4,%o0                               ! ++$out
 436
 437         retl
 438         mov     %o3,%o0
 439 .type   _sparcv9_vis1_instrument_bus,#function
 440 .size   _sparcv9_vis1_instrument_bus,.-_sparcv9_vis1_instrument_bus
 441
 442 .global _sparcv9_vis1_instrument_bus2
 443 .align  8
 444 _sparcv9_vis1_instrument_bus2:
 445         mov     %o1,%o3                                 ! save cnt
 446         sll     %o1,2,%o1                               ! cnt*=4
 447
 448         .word   0x99410000      !rd     %tick,%o4       ! tick
 449         mov     %o4,%o5                                 ! lasttick = tick
 450         set     0,%g4                                   ! diff
 451
 452         andn    %o0,63,%g1
 453         .word   0xc1985e00      !ldda   [%g1]0xf0,%f0   ! block load
 454         .word   0x8143e040      !membar #Sync
 455         .word   0xc1b85c00      !stda   %f0,[%g1]0xe0   ! block store and commit
 456         .word   0x8143e040      !membar #Sync
 457         ld      [%o0],%o4
 458         add     %o4,%g4,%g4
 459         .word   0xc9e2100c      !cas    [%o0],%o4,%g4
 460
 461         .word   0x99410000      !rd     %tick,%o4       ! tick
 462         sub     %o4,%o5,%g4                             ! diff=tick-lasttick
 463         mov     %o4,%o5                                 ! lasttick=tick
 464         mov     %g4,%g5                                 ! lastdiff=diff
 465 .Loop2:
 466         andn    %o0,63,%g1
 467         .word   0xc1985e00      !ldda   [%g1]0xf0,%f0   ! block load
 468         .word   0x8143e040      !membar #Sync
 469         .word   0xc1b85c00      !stda   %f0,[%g1]0xe0   ! block store and commit
 470         .word   0x8143e040      !membar #Sync
 471         ld      [%o0],%o4
 472         add     %o4,%g4,%g4
 473         .word   0xc9e2100c      !cas    [%o0],%o4,%g4
 474
 475         subcc   %o2,1,%o2                               ! --max
 476         bz      .Ldone2
 477         nop
 478
 479         .word   0x99410000      !rd     %tick,%o4       ! tick
 480         sub     %o4,%o5,%g4                             ! diff=tick-lasttick
 481         mov     %o4,%o5                                 ! lasttick=tick
 482         cmp     %g4,%g5
 483         mov     %g4,%g5                                 ! lastdiff=diff
 484
 485         .word   0x83408000      !rd     %ccr,%g1
 486         and     %g1,4,%g1                               ! isolate zero flag
 487         xor     %g1,4,%g1                               ! flip zero flag
 488
 489         subcc   %o1,%g1,%o1                             ! conditional --$cnt
 490         bnz     .Loop2
 491         add     %o0,%g1,%o0                             ! conditional ++$out
 492
 493 .Ldone2:
 494         srl     %o1,2,%o1
 495         retl
 496         sub     %o3,%o1,%o0
 497 .type   _sparcv9_vis1_instrument_bus2,#function
 498 .size   _sparcv9_vis1_instrument_bus2,.-_sparcv9_vis1_instrument_bus2
 499
 500 .section        ".init",#alloc,#execinstr
 501         call    OPENSSL_cpuid_setup
 502         nop