arch/alpha/math-emu/math.c

   1 #include <linux/module.h>
   2 #include <linux/types.h>
   3 #include <linux/kernel.h>
   4 #include <linux/sched.h>
   5
   6 #include <asm/uaccess.h>
   7
   8 #include "sfp-util.h"
   9 #include <math-emu/soft-fp.h>
  10 #include <math-emu/single.h>
  11 #include <math-emu/double.h>
  12
  13 #define OPC_PAL         0x00
  14 #define OPC_INTA        0x10
  15 #define OPC_INTL        0x11
  16 #define OPC_INTS        0x12
  17 #define OPC_INTM        0x13
  18 #define OPC_FLTC        0x14
  19 #define OPC_FLTV        0x15
  20 #define OPC_FLTI        0x16
  21 #define OPC_FLTL        0x17
  22 #define OPC_MISC        0x18
  23 #define OPC_JSR         0x1a
  24
  25 #define FOP_SRC_S       0
  26 #define FOP_SRC_T       2
  27 #define FOP_SRC_Q       3
  28
  29 #define FOP_FNC_ADDx    0
  30 #define FOP_FNC_CVTQL   0
  31 #define FOP_FNC_SUBx    1
  32 #define FOP_FNC_MULx    2
  33 #define FOP_FNC_DIVx    3
  34 #define FOP_FNC_CMPxUN  4
  35 #define FOP_FNC_CMPxEQ  5
  36 #define FOP_FNC_CMPxLT  6
  37 #define FOP_FNC_CMPxLE  7
  38 #define FOP_FNC_SQRTx   11
  39 #define FOP_FNC_CVTxS   12
  40 #define FOP_FNC_CVTxT   14
  41 #define FOP_FNC_CVTxQ   15
  42
  43 #define MISC_TRAPB      0x0000
  44 #define MISC_EXCB       0x0400
  45
  46 extern unsigned long alpha_read_fp_reg (unsigned long reg);
  47 extern void alpha_write_fp_reg (unsigned long reg, unsigned long val);
  48 extern unsigned long alpha_read_fp_reg_s (unsigned long reg);
  49 extern void alpha_write_fp_reg_s (unsigned long reg, unsigned long val);
  50
  51
  52 #ifdef MODULE
  53
  54 MODULE_DESCRIPTION("FP Software completion module");
  55
  56 extern long (*alpha_fp_emul_imprecise)(struct pt_regs *, unsigned long);
  57 extern long (*alpha_fp_emul) (unsigned long pc);
  58
  59 static long (*save_emul_imprecise)(struct pt_regs *, unsigned long);
  60 static long (*save_emul) (unsigned long pc);
  61
  62 long do_alpha_fp_emul_imprecise(struct pt_regs *, unsigned long);
  63 long do_alpha_fp_emul(unsigned long);
  64
  65 int init_module(void)
  66 {
  67         save_emul_imprecise = alpha_fp_emul_imprecise;
  68         save_emul = alpha_fp_emul;
  69         alpha_fp_emul_imprecise = do_alpha_fp_emul_imprecise;
  70         alpha_fp_emul = do_alpha_fp_emul;
  71         return 0;
  72 }
  73
  74 void cleanup_module(void)
  75 {
  76         alpha_fp_emul_imprecise = save_emul_imprecise;
  77         alpha_fp_emul = save_emul;
  78 }
  79
  80 #undef  alpha_fp_emul_imprecise
  81 #define alpha_fp_emul_imprecise         do_alpha_fp_emul_imprecise
  82 #undef  alpha_fp_emul
  83 #define alpha_fp_emul                   do_alpha_fp_emul
  84
  85 #endif /* MODULE */
  86
  87
  88 /*
  89  * Emulate the floating point instruction at address PC.  Returns -1 if the
  90  * instruction to be emulated is illegal (such as with the opDEC trap), else
  91  * the SI_CODE for a SIGFPE signal, else 0 if everything's ok.
  92  *
  93  * Notice that the kernel does not and cannot use FP regs.  This is good
  94  * because it means that instead of saving/restoring all fp regs, we simply
  95  * stick the result of the operation into the appropriate register.
  96  */
  97 long
  98 alpha_fp_emul (unsigned long pc)
  99 {
 100         FP_DECL_EX;
 101         FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR);
 102         FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR);
 103
 104         unsigned long fa, fb, fc, func, mode, src;
 105         unsigned long res, va, vb, vc, swcr, fpcr;
 106         __u32 insn;
 107         long si_code;
 108
 109         get_user(insn, (__u32 __user *)pc);
 110         fc     = (insn >>  0) & 0x1f;   /* destination register */
 111         fb     = (insn >> 16) & 0x1f;
 112         fa     = (insn >> 21) & 0x1f;
 113         func   = (insn >>  5) & 0xf;
 114         src    = (insn >>  9) & 0x3;
 115         mode   = (insn >> 11) & 0x3;
 116
 117         fpcr = rdfpcr();
 118         swcr = swcr_update_status(current_thread_info()->ieee_state, fpcr);
 119
 120         if (mode == 3) {
 121                 /* Dynamic -- get rounding mode from fpcr.  */
 122                 mode = (fpcr >> FPCR_DYN_SHIFT) & 3;
 123         }
 124
 125         switch (src) {
 126         case FOP_SRC_S:
 127                 va = alpha_read_fp_reg_s(fa);
 128                 vb = alpha_read_fp_reg_s(fb);
 129
 130                 FP_UNPACK_SP(SA, &va);
 131                 FP_UNPACK_SP(SB, &vb);
 132
 133                 switch (func) {
 134                 case FOP_FNC_SUBx:
 135                         FP_SUB_S(SR, SA, SB);
 136                         goto pack_s;
 137
 138                 case FOP_FNC_ADDx:
 139                         FP_ADD_S(SR, SA, SB);
 140                         goto pack_s;
 141
 142                 case FOP_FNC_MULx:
 143                         FP_MUL_S(SR, SA, SB);
 144                         goto pack_s;
 145
 146                 case FOP_FNC_DIVx:
 147                         FP_DIV_S(SR, SA, SB);
 148                         goto pack_s;
 149
 150                 case FOP_FNC_SQRTx:
 151                         FP_SQRT_S(SR, SB);
 152                         goto pack_s;
 153                 }
 154                 goto bad_insn;
 155
 156         case FOP_SRC_T:
 157                 va = alpha_read_fp_reg(fa);
 158                 vb = alpha_read_fp_reg(fb);
 159
 160                 if ((func & ~3) == FOP_FNC_CMPxUN) {
 161                         FP_UNPACK_RAW_DP(DA, &va);
 162                         FP_UNPACK_RAW_DP(DB, &vb);
 163                         if (!DA_e && !_FP_FRAC_ZEROP_1(DA)) {
 164                                 FP_SET_EXCEPTION(FP_EX_DENORM);
 165                                 if (FP_DENORM_ZERO)
 166                                         _FP_FRAC_SET_1(DA, _FP_ZEROFRAC_1);
 167                         }
 168                         if (!DB_e && !_FP_FRAC_ZEROP_1(DB)) {
 169                                 FP_SET_EXCEPTION(FP_EX_DENORM);
 170                                 if (FP_DENORM_ZERO)
 171                                         _FP_FRAC_SET_1(DB, _FP_ZEROFRAC_1);
 172                         }
 173                         FP_CMP_D(res, DA, DB, 3);
 174                         vc = 0x4000000000000000UL;
 175                         /* CMPTEQ, CMPTUN don't trap on QNaN,
 176                            while CMPTLT and CMPTLE do */
 177                         if (res == 3
 178                             && ((func & 3) >= 2
 179                                 || FP_ISSIGNAN_D(DA)
 180                                 || FP_ISSIGNAN_D(DB))) {
 181                                 FP_SET_EXCEPTION(FP_EX_INVALID);
 182                         }
 183                         switch (func) {
 184                         case FOP_FNC_CMPxUN: if (res != 3) vc = 0; break;
 185                         case FOP_FNC_CMPxEQ: if (res) vc = 0; break;
 186                         case FOP_FNC_CMPxLT: if (res != -1) vc = 0; break;
 187                         case FOP_FNC_CMPxLE: if ((long)res > 0) vc = 0; break;
 188                         }
 189                         goto done_d;
 190                 }
 191
 192                 FP_UNPACK_DP(DA, &va);
 193                 FP_UNPACK_DP(DB, &vb);
 194
 195                 switch (func) {
 196                 case FOP_FNC_SUBx:
 197                         FP_SUB_D(DR, DA, DB);
 198                         goto pack_d;
 199
 200                 case FOP_FNC_ADDx:
 201                         FP_ADD_D(DR, DA, DB);
 202                         goto pack_d;
 203
 204                 case FOP_FNC_MULx:
 205                         FP_MUL_D(DR, DA, DB);
 206                         goto pack_d;
 207
 208                 case FOP_FNC_DIVx:
 209                         FP_DIV_D(DR, DA, DB);
 210                         goto pack_d;
 211
 212                 case FOP_FNC_SQRTx:
 213                         FP_SQRT_D(DR, DB);
 214                         goto pack_d;
 215
 216                 case FOP_FNC_CVTxS:
 217                         /* It is irritating that DEC encoded CVTST with
 218                            SRC == T_floating.  It is also interesting that
 219                            the bit used to tell the two apart is /U... */
 220                         if (insn & 0x2000) {
 221                                 FP_CONV(S,D,1,1,SR,DB);
 222                                 goto pack_s;
 223                         } else {
 224                                 vb = alpha_read_fp_reg_s(fb);
 225                                 FP_UNPACK_SP(SB, &vb);
 226                                 DR_c = DB_c;
 227                                 DR_s = DB_s;
 228                                 DR_e = DB_e + (1024 - 128);
 229                                 DR_f = SB_f << (52 - 23);
 230                                 goto pack_d;
 231                         }
 232
 233                 case FOP_FNC_CVTxQ:
 234                         if (DB_c == FP_CLS_NAN
 235                             && (_FP_FRAC_HIGH_RAW_D(DB) & _FP_QNANBIT_D)) {
 236                           /* AAHB Table B-2 says QNaN should not trigger INV */
 237                                 vc = 0;
 238                         } else
 239                                 FP_TO_INT_ROUND_D(vc, DB, 64, 2);
 240                         goto done_d;
 241                 }
 242                 goto bad_insn;
 243
 244         case FOP_SRC_Q:
 245                 vb = alpha_read_fp_reg(fb);
 246
 247                 switch (func) {
 248                 case FOP_FNC_CVTQL:
 249                         /* Notice: We can get here only due to an integer
 250                            overflow.  Such overflows are reported as invalid
 251                            ops.  We return the result the hw would have
 252                            computed.  */
 253                         vc = ((vb & 0xc0000000) << 32 | /* sign and msb */
 254                               (vb & 0x3fffffff) << 29); /* rest of the int */
 255                         FP_SET_EXCEPTION (FP_EX_INVALID);
 256                         goto done_d;
 257
 258                 case FOP_FNC_CVTxS:
 259                         FP_FROM_INT_S(SR, ((long)vb), 64, long);
 260                         goto pack_s;
 261
 262                 case FOP_FNC_CVTxT:
 263                         FP_FROM_INT_D(DR, ((long)vb), 64, long);
 264                         goto pack_d;
 265                 }
 266                 goto bad_insn;
 267         }
 268         goto bad_insn;
 269
 270 pack_s:
 271         FP_PACK_SP(&vc, SR);
 272         if ((_fex & FP_EX_UNDERFLOW) && (swcr & IEEE_MAP_UMZ))
 273                 vc = 0;
 274         alpha_write_fp_reg_s(fc, vc);
 275         goto done;
 276
 277 pack_d:
 278         FP_PACK_DP(&vc, DR);
 279         if ((_fex & FP_EX_UNDERFLOW) && (swcr & IEEE_MAP_UMZ))
 280                 vc = 0;
 281 done_d:
 282         alpha_write_fp_reg(fc, vc);
 283         goto done;
 284
 285         /*
 286          * Take the appropriate action for each possible
 287          * floating-point result:
 288          *
 289          *      - Set the appropriate bits in the FPCR
 290          *      - If the specified exception is enabled in the FPCR,
 291          *        return.  The caller (entArith) will dispatch
 292          *        the appropriate signal to the translated program.
 293          *
 294          * In addition, properly track the exception state in software
 295          * as described in the Alpha Architecture Handbook section 4.7.7.3.
 296          */
 297 done:
 298         if (_fex) {
 299                 /* Record exceptions in software control word.  */
 300                 swcr |= (_fex << IEEE_STATUS_TO_EXCSUM_SHIFT);
 301                 current_thread_info()->ieee_state
 302                   |= (_fex << IEEE_STATUS_TO_EXCSUM_SHIFT);
 303
 304                 /* Update hardware control register.  */
 305                 fpcr &= (~FPCR_MASK | FPCR_DYN_MASK);
 306                 fpcr |= ieee_swcr_to_fpcr(swcr);
 307                 wrfpcr(fpcr);
 308
 309                 /* Do we generate a signal?  */
 310                 _fex = _fex & swcr & IEEE_TRAP_ENABLE_MASK;
 311                 si_code = 0;
 312                 if (_fex) {
 313                         if (_fex & IEEE_TRAP_ENABLE_DNO) si_code = FPE_FLTUND;
 314                         if (_fex & IEEE_TRAP_ENABLE_INE) si_code = FPE_FLTRES;
 315                         if (_fex & IEEE_TRAP_ENABLE_UNF) si_code = FPE_FLTUND;
 316                         if (_fex & IEEE_TRAP_ENABLE_OVF) si_code = FPE_FLTOVF;
 317                         if (_fex & IEEE_TRAP_ENABLE_DZE) si_code = FPE_FLTDIV;
 318                         if (_fex & IEEE_TRAP_ENABLE_INV) si_code = FPE_FLTINV;
 319                 }
 320
 321                 return si_code;
 322         }
 323
 324         /* We used to write the destination register here, but DEC FORTRAN
 325            requires that the result *always* be written... so we do the write
 326            immediately after the operations above.  */
 327
 328         return 0;
 329
 330 bad_insn:
 331         printk(KERN_ERR "alpha_fp_emul: Invalid FP insn %#x at %#lx\n",
 332                insn, pc);
 333         return -1;
 334 }
 335
 336 long
 337 alpha_fp_emul_imprecise (struct pt_regs *regs, unsigned long write_mask)
 338 {
 339         unsigned long trigger_pc = regs->pc - 4;
 340         unsigned long insn, opcode, rc, si_code = 0;
 341
 342         /*
 343          * Turn off the bits corresponding to registers that are the
 344          * target of instructions that set bits in the exception
 345          * summary register.  We have some slack doing this because a
 346          * register that is the target of a trapping instruction can
 347          * be written at most once in the trap shadow.
 348          *
 349          * Branches, jumps, TRAPBs, EXCBs and calls to PALcode all
 350          * bound the trap shadow, so we need not look any further than
 351          * up to the first occurrence of such an instruction.
 352          */
 353         while (write_mask) {
 354                 get_user(insn, (__u32 __user *)(trigger_pc));
 355                 opcode = insn >> 26;
 356                 rc = insn & 0x1f;
 357
 358                 switch (opcode) {
 359                       case OPC_PAL:
 360                       case OPC_JSR:
 361                       case 0x30 ... 0x3f:       /* branches */
 362                         goto egress;
 363
 364                       case OPC_MISC:
 365                         switch (insn & 0xffff) {
 366                               case MISC_TRAPB:
 367                               case MISC_EXCB:
 368                                 goto egress;
 369
 370                               default:
 371                                 break;
 372                         }
 373                         break;
 374
 375                       case OPC_INTA:
 376                       case OPC_INTL:
 377                       case OPC_INTS:
 378                       case OPC_INTM:
 379                         write_mask &= ~(1UL << rc);
 380                         break;
 381
 382                       case OPC_FLTC:
 383                       case OPC_FLTV:
 384                       case OPC_FLTI:
 385                       case OPC_FLTL:
 386                         write_mask &= ~(1UL << (rc + 32));
 387                         break;
 388                 }
 389                 if (!write_mask) {
 390                         /* Re-execute insns in the trap-shadow.  */
 391                         regs->pc = trigger_pc + 4;
 392                         si_code = alpha_fp_emul(trigger_pc);
 393                         goto egress;
 394                 }
 395                 trigger_pc -= 4;
 396         }
 397
 398 egress:
 399         return si_code;
 400 }