1 /* 2 * arch/sparc64/math-emu/math.c 3 * 4 * Copyright (C) 1997,1999 Jakub Jelinek (jj@ultra.linux.cz) 5 * Copyright (C) 1999 David S. Miller (davem@redhat.com) 6 * 7 * Emulation routines originate from soft-fp package, which is part 8 * of glibc and has appropriate copyrights in it. 9 */ 10 11 #include <linux/types.h> 12 #include <linux/sched.h> 13 #include <linux/errno.h> 14 #include <linux/perf_event.h> 15 16 #include <asm/fpumacro.h> 17 #include <asm/ptrace.h> 18 #include <linux/uaccess.h> 19 #include <asm/cacheflush.h> 20 21 #include "sfp-util_64.h" 22 #include <math-emu/soft-fp.h> 23 #include <math-emu/single.h> 24 #include <math-emu/double.h> 25 #include <math-emu/quad.h> 26 27 /* QUAD - ftt == 3 */ 28 #define FMOVQ 0x003 29 #define FNEGQ 0x007 30 #define FABSQ 0x00b 31 #define FSQRTQ 0x02b 32 #define FADDQ 0x043 33 #define FSUBQ 0x047 34 #define FMULQ 0x04b 35 #define FDIVQ 0x04f 36 #define FDMULQ 0x06e 37 #define FQTOX 0x083 38 #define FXTOQ 0x08c 39 #define FQTOS 0x0c7 40 #define FQTOD 0x0cb 41 #define FITOQ 0x0cc 42 #define FSTOQ 0x0cd 43 #define FDTOQ 0x0ce 44 #define FQTOI 0x0d3 45 /* SUBNORMAL - ftt == 2 */ 46 #define FSQRTS 0x029 47 #define FSQRTD 0x02a 48 #define FADDS 0x041 49 #define FADDD 0x042 50 #define FSUBS 0x045 51 #define FSUBD 0x046 52 #define FMULS 0x049 53 #define FMULD 0x04a 54 #define FDIVS 0x04d 55 #define FDIVD 0x04e 56 #define FSMULD 0x069 57 #define FSTOX 0x081 58 #define FDTOX 0x082 59 #define FDTOS 0x0c6 60 #define FSTOD 0x0c9 61 #define FSTOI 0x0d1 62 #define FDTOI 0x0d2 63 #define FXTOS 0x084 /* Only Ultra-III generates this. */ 64 #define FXTOD 0x088 /* Only Ultra-III generates this. */ 65 #if 0 /* Optimized inline in sparc64/kernel/entry.S */ 66 #define FITOS 0x0c4 /* Only Ultra-III generates this. */ 67 #endif 68 #define FITOD 0x0c8 /* Only Ultra-III generates this. */ 69 /* FPOP2 */ 70 #define FCMPQ 0x053 71 #define FCMPEQ 0x057 72 #define FMOVQ0 0x003 73 #define FMOVQ1 0x043 74 #define FMOVQ2 0x083 75 #define FMOVQ3 0x0c3 76 #define FMOVQI 0x103 77 #define FMOVQX 0x183 78 #define FMOVQZ 0x027 79 #define FMOVQLE 0x047 80 #define FMOVQLZ 0x067 81 #define FMOVQNZ 0x0a7 82 #define FMOVQGZ 0x0c7 83 #define FMOVQGE 0x0e7 84 85 #define FSR_TEM_SHIFT 23UL 86 #define FSR_TEM_MASK (0x1fUL << FSR_TEM_SHIFT) 87 #define FSR_AEXC_SHIFT 5UL 88 #define FSR_AEXC_MASK (0x1fUL << FSR_AEXC_SHIFT) 89 #define FSR_CEXC_SHIFT 0UL 90 #define FSR_CEXC_MASK (0x1fUL << FSR_CEXC_SHIFT) 91 92 /* All routines returning an exception to raise should detect 93 * such exceptions _before_ rounding to be consistent with 94 * the behavior of the hardware in the implemented cases 95 * (and thus with the recommendations in the V9 architecture 96 * manual). 97 * 98 * We return 0 if a SIGFPE should be sent, 1 otherwise. 99 */ 100 static inline int record_exception(struct pt_regs *regs, int eflag) 101 { 102 u64 fsr = current_thread_info()->xfsr[0]; 103 int would_trap; 104 105 /* Determine if this exception would have generated a trap. */ 106 would_trap = (fsr & ((long)eflag << FSR_TEM_SHIFT)) != 0UL; 107 108 /* If trapping, we only want to signal one bit. */ 109 if(would_trap != 0) { 110 eflag &= ((fsr & FSR_TEM_MASK) >> FSR_TEM_SHIFT); 111 if((eflag & (eflag - 1)) != 0) { 112 if(eflag & FP_EX_INVALID) 113 eflag = FP_EX_INVALID; 114 else if(eflag & FP_EX_OVERFLOW) 115 eflag = FP_EX_OVERFLOW; 116 else if(eflag & FP_EX_UNDERFLOW) 117 eflag = FP_EX_UNDERFLOW; 118 else if(eflag & FP_EX_DIVZERO) 119 eflag = FP_EX_DIVZERO; 120 else if(eflag & FP_EX_INEXACT) 121 eflag = FP_EX_INEXACT; 122 } 123 } 124 125 /* Set CEXC, here is the rule: 126 * 127 * In general all FPU ops will set one and only one 128 * bit in the CEXC field, this is always the case 129 * when the IEEE exception trap is enabled in TEM. 130 */ 131 fsr &= ~(FSR_CEXC_MASK); 132 fsr |= ((long)eflag << FSR_CEXC_SHIFT); 133 134 /* Set the AEXC field, rule is: 135 * 136 * If a trap would not be generated, the 137 * CEXC just generated is OR'd into the 138 * existing value of AEXC. 139 */ 140 if(would_trap == 0) 141 fsr |= ((long)eflag << FSR_AEXC_SHIFT); 142 143 /* If trapping, indicate fault trap type IEEE. */ 144 if(would_trap != 0) 145 fsr |= (1UL << 14); 146 147 current_thread_info()->xfsr[0] = fsr; 148 149 /* If we will not trap, advance the program counter over 150 * the instruction being handled. 151 */ 152 if(would_trap == 0) { 153 regs->tpc = regs->tnpc; 154 regs->tnpc += 4; 155 } 156 157 return (would_trap ? 0 : 1); 158 } 159 160 typedef union { 161 u32 s; 162 u64 d; 163 u64 q[2]; 164 } *argp; 165 166 int do_mathemu(struct pt_regs *regs, struct fpustate *f, bool illegal_insn_trap) 167 { 168 unsigned long pc = regs->tpc; 169 unsigned long tstate = regs->tstate; 170 u32 insn = 0; 171 int type = 0; 172 /* ftt tells which ftt it may happen in, r is rd, b is rs2 and a is rs1. The *u arg tells 173 whether the argument should be packed/unpacked (0 - do not unpack/pack, 1 - unpack/pack) 174 non-u args tells the size of the argument (0 - no argument, 1 - single, 2 - double, 3 - quad */ 175 #define TYPE(ftt, r, ru, b, bu, a, au) type = (au << 2) | (a << 0) | (bu << 5) | (b << 3) | (ru << 8) | (r << 6) | (ftt << 9) 176 int freg; 177 static u64 zero[2] = { 0L, 0L }; 178 int flags; 179 FP_DECL_EX; 180 FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR); 181 FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR); 182 FP_DECL_Q(QA); FP_DECL_Q(QB); FP_DECL_Q(QR); 183 int IR; 184 long XR, xfsr; 185 186 if (tstate & TSTATE_PRIV) 187 die_if_kernel("unfinished/unimplemented FPop from kernel", regs); 188 perf_sw_event(PERF_COUNT_SW_EMULATION_FAULTS, 1, regs, 0); 189 if (test_thread_flag(TIF_32BIT)) 190 pc = (u32)pc; 191 if (get_user(insn, (u32 __user *) pc) != -EFAULT) { 192 if ((insn & 0xc1f80000) == 0x81a00000) /* FPOP1 */ { 193 switch ((insn >> 5) & 0x1ff) { 194 /* QUAD - ftt == 3 */ 195 case FMOVQ: 196 case FNEGQ: 197 case FABSQ: TYPE(3,3,0,3,0,0,0); break; 198 case FSQRTQ: TYPE(3,3,1,3,1,0,0); break; 199 case FADDQ: 200 case FSUBQ: 201 case FMULQ: 202 case FDIVQ: TYPE(3,3,1,3,1,3,1); break; 203 case FDMULQ: TYPE(3,3,1,2,1,2,1); break; 204 case FQTOX: TYPE(3,2,0,3,1,0,0); break; 205 case FXTOQ: TYPE(3,3,1,2,0,0,0); break; 206 case FQTOS: TYPE(3,1,1,3,1,0,0); break; 207 case FQTOD: TYPE(3,2,1,3,1,0,0); break; 208 case FITOQ: TYPE(3,3,1,1,0,0,0); break; 209 case FSTOQ: TYPE(3,3,1,1,1,0,0); break; 210 case FDTOQ: TYPE(3,3,1,2,1,0,0); break; 211 case FQTOI: TYPE(3,1,0,3,1,0,0); break; 212 213 /* We can get either unimplemented or unfinished 214 * for these cases. Pre-Niagara systems generate 215 * unfinished fpop for SUBNORMAL cases, and Niagara 216 * always gives unimplemented fpop for fsqrt{s,d}. 217 */ 218 case FSQRTS: { 219 unsigned long x = current_thread_info()->xfsr[0]; 220 221 x = (x >> 14) & 0x7; 222 TYPE(x,1,1,1,1,0,0); 223 break; 224 } 225 226 case FSQRTD: { 227 unsigned long x = current_thread_info()->xfsr[0]; 228 229 x = (x >> 14) & 0x7; 230 TYPE(x,2,1,2,1,0,0); 231 break; 232 } 233 234 /* SUBNORMAL - ftt == 2 */ 235 case FADDD: 236 case FSUBD: 237 case FMULD: 238 case FDIVD: TYPE(2,2,1,2,1,2,1); break; 239 case FADDS: 240 case FSUBS: 241 case FMULS: 242 case FDIVS: TYPE(2,1,1,1,1,1,1); break; 243 case FSMULD: TYPE(2,2,1,1,1,1,1); break; 244 case FSTOX: TYPE(2,2,0,1,1,0,0); break; 245 case FDTOX: TYPE(2,2,0,2,1,0,0); break; 246 case FDTOS: TYPE(2,1,1,2,1,0,0); break; 247 case FSTOD: TYPE(2,2,1,1,1,0,0); break; 248 case FSTOI: TYPE(2,1,0,1,1,0,0); break; 249 case FDTOI: TYPE(2,1,0,2,1,0,0); break; 250 251 /* Only Ultra-III generates these */ 252 case FXTOS: TYPE(2,1,1,2,0,0,0); break; 253 case FXTOD: TYPE(2,2,1,2,0,0,0); break; 254 #if 0 /* Optimized inline in sparc64/kernel/entry.S */ 255 case FITOS: TYPE(2,1,1,1,0,0,0); break; 256 #endif 257 case FITOD: TYPE(2,2,1,1,0,0,0); break; 258 } 259 } 260 else if ((insn & 0xc1f80000) == 0x81a80000) /* FPOP2 */ { 261 IR = 2; 262 switch ((insn >> 5) & 0x1ff) { 263 case FCMPQ: TYPE(3,0,0,3,1,3,1); break; 264 case FCMPEQ: TYPE(3,0,0,3,1,3,1); break; 265 /* Now the conditional fmovq support */ 266 case FMOVQ0: 267 case FMOVQ1: 268 case FMOVQ2: 269 case FMOVQ3: 270 /* fmovq %fccX, %fY, %fZ */ 271 if (!((insn >> 11) & 3)) 272 XR = current_thread_info()->xfsr[0] >> 10; 273 else 274 XR = current_thread_info()->xfsr[0] >> (30 + ((insn >> 10) & 0x6)); 275 XR &= 3; 276 IR = 0; 277 switch ((insn >> 14) & 0x7) { 278 /* case 0: IR = 0; break; */ /* Never */ 279 case 1: if (XR) IR = 1; break; /* Not Equal */ 280 case 2: if (XR == 1 || XR == 2) IR = 1; break; /* Less or Greater */ 281 case 3: if (XR & 1) IR = 1; break; /* Unordered or Less */ 282 case 4: if (XR == 1) IR = 1; break; /* Less */ 283 case 5: if (XR & 2) IR = 1; break; /* Unordered or Greater */ 284 case 6: if (XR == 2) IR = 1; break; /* Greater */ 285 case 7: if (XR == 3) IR = 1; break; /* Unordered */ 286 } 287 if ((insn >> 14) & 8) 288 IR ^= 1; 289 break; 290 case FMOVQI: 291 case FMOVQX: 292 /* fmovq %[ix]cc, %fY, %fZ */ 293 XR = regs->tstate >> 32; 294 if ((insn >> 5) & 0x80) 295 XR >>= 4; 296 XR &= 0xf; 297 IR = 0; 298 freg = ((XR >> 2) ^ XR) & 2; 299 switch ((insn >> 14) & 0x7) { 300 /* case 0: IR = 0; break; */ /* Never */ 301 case 1: if (XR & 4) IR = 1; break; /* Equal */ 302 case 2: if ((XR & 4) || freg) IR = 1; break; /* Less or Equal */ 303 case 3: if (freg) IR = 1; break; /* Less */ 304 case 4: if (XR & 5) IR = 1; break; /* Less or Equal Unsigned */ 305 case 5: if (XR & 1) IR = 1; break; /* Carry Set */ 306 case 6: if (XR & 8) IR = 1; break; /* Negative */ 307 case 7: if (XR & 2) IR = 1; break; /* Overflow Set */ 308 } 309 if ((insn >> 14) & 8) 310 IR ^= 1; 311 break; 312 case FMOVQZ: 313 case FMOVQLE: 314 case FMOVQLZ: 315 case FMOVQNZ: 316 case FMOVQGZ: 317 case FMOVQGE: 318 freg = (insn >> 14) & 0x1f; 319 if (!freg) 320 XR = 0; 321 else if (freg < 16) 322 XR = regs->u_regs[freg]; 323 else if (!test_thread_64bit_stack(regs->u_regs[UREG_FP])) { 324 struct reg_window32 __user *win32; 325 flushw_user (); 326 win32 = (struct reg_window32 __user *)((unsigned long)((u32)regs->u_regs[UREG_FP])); 327 get_user(XR, &win32->locals[freg - 16]); 328 } else { 329 struct reg_window __user *win; 330 flushw_user (); 331 win = (struct reg_window __user *)(regs->u_regs[UREG_FP] + STACK_BIAS); 332 get_user(XR, &win->locals[freg - 16]); 333 } 334 IR = 0; 335 switch ((insn >> 10) & 3) { 336 case 1: if (!XR) IR = 1; break; /* Register Zero */ 337 case 2: if (XR <= 0) IR = 1; break; /* Register Less Than or Equal to Zero */ 338 case 3: if (XR < 0) IR = 1; break; /* Register Less Than Zero */ 339 } 340 if ((insn >> 10) & 4) 341 IR ^= 1; 342 break; 343 } 344 if (IR == 0) { 345 /* The fmov test was false. Do a nop instead */ 346 current_thread_info()->xfsr[0] &= ~(FSR_CEXC_MASK); 347 regs->tpc = regs->tnpc; 348 regs->tnpc += 4; 349 return 1; 350 } else if (IR == 1) { 351 /* Change the instruction into plain fmovq */ 352 insn = (insn & 0x3e00001f) | 0x81a00060; 353 TYPE(3,3,0,3,0,0,0); 354 } 355 } 356 } 357 if (type) { 358 argp rs1 = NULL, rs2 = NULL, rd = NULL; 359 360 /* Starting with UltraSPARC-T2, the cpu does not set the FP Trap 361 * Type field in the %fsr to unimplemented_FPop. Nor does it 362 * use the fp_exception_other trap. Instead it signals an 363 * illegal instruction and leaves the FP trap type field of 364 * the %fsr unchanged. 365 */ 366 if (!illegal_insn_trap) { 367 int ftt = (current_thread_info()->xfsr[0] >> 14) & 0x7; 368 if (ftt != (type >> 9)) 369 goto err; 370 } 371 current_thread_info()->xfsr[0] &= ~0x1c000; 372 freg = ((insn >> 14) & 0x1f); 373 switch (type & 0x3) { 374 case 3: if (freg & 2) { 375 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */; 376 goto err; 377 } 378 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e); 379 case 1: rs1 = (argp)&f->regs[freg]; 380 flags = (freg < 32) ? FPRS_DL : FPRS_DU; 381 if (!(current_thread_info()->fpsaved[0] & flags)) 382 rs1 = (argp)&zero; 383 break; 384 } 385 switch (type & 0x7) { 386 case 7: FP_UNPACK_QP (QA, rs1); break; 387 case 6: FP_UNPACK_DP (DA, rs1); break; 388 case 5: FP_UNPACK_SP (SA, rs1); break; 389 } 390 freg = (insn & 0x1f); 391 switch ((type >> 3) & 0x3) { 392 case 3: if (freg & 2) { 393 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */; 394 goto err; 395 } 396 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e); 397 case 1: rs2 = (argp)&f->regs[freg]; 398 flags = (freg < 32) ? FPRS_DL : FPRS_DU; 399 if (!(current_thread_info()->fpsaved[0] & flags)) 400 rs2 = (argp)&zero; 401 break; 402 } 403 switch ((type >> 3) & 0x7) { 404 case 7: FP_UNPACK_QP (QB, rs2); break; 405 case 6: FP_UNPACK_DP (DB, rs2); break; 406 case 5: FP_UNPACK_SP (SB, rs2); break; 407 } 408 freg = ((insn >> 25) & 0x1f); 409 switch ((type >> 6) & 0x3) { 410 case 3: if (freg & 2) { 411 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */; 412 goto err; 413 } 414 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e); 415 case 1: rd = (argp)&f->regs[freg]; 416 flags = (freg < 32) ? FPRS_DL : FPRS_DU; 417 if (!(current_thread_info()->fpsaved[0] & FPRS_FEF)) { 418 current_thread_info()->fpsaved[0] = FPRS_FEF; 419 current_thread_info()->gsr[0] = 0; 420 } 421 if (!(current_thread_info()->fpsaved[0] & flags)) { 422 if (freg < 32) 423 memset(f->regs, 0, 32*sizeof(u32)); 424 else 425 memset(f->regs+32, 0, 32*sizeof(u32)); 426 } 427 current_thread_info()->fpsaved[0] |= flags; 428 break; 429 } 430 switch ((insn >> 5) & 0x1ff) { 431 /* + */ 432 case FADDS: FP_ADD_S (SR, SA, SB); break; 433 case FADDD: FP_ADD_D (DR, DA, DB); break; 434 case FADDQ: FP_ADD_Q (QR, QA, QB); break; 435 /* - */ 436 case FSUBS: FP_SUB_S (SR, SA, SB); break; 437 case FSUBD: FP_SUB_D (DR, DA, DB); break; 438 case FSUBQ: FP_SUB_Q (QR, QA, QB); break; 439 /* * */ 440 case FMULS: FP_MUL_S (SR, SA, SB); break; 441 case FSMULD: FP_CONV (D, S, 1, 1, DA, SA); 442 FP_CONV (D, S, 1, 1, DB, SB); 443 case FMULD: FP_MUL_D (DR, DA, DB); break; 444 case FDMULQ: FP_CONV (Q, D, 2, 1, QA, DA); 445 FP_CONV (Q, D, 2, 1, QB, DB); 446 case FMULQ: FP_MUL_Q (QR, QA, QB); break; 447 /* / */ 448 case FDIVS: FP_DIV_S (SR, SA, SB); break; 449 case FDIVD: FP_DIV_D (DR, DA, DB); break; 450 case FDIVQ: FP_DIV_Q (QR, QA, QB); break; 451 /* sqrt */ 452 case FSQRTS: FP_SQRT_S (SR, SB); break; 453 case FSQRTD: FP_SQRT_D (DR, DB); break; 454 case FSQRTQ: FP_SQRT_Q (QR, QB); break; 455 /* mov */ 456 case FMOVQ: rd->q[0] = rs2->q[0]; rd->q[1] = rs2->q[1]; break; 457 case FABSQ: rd->q[0] = rs2->q[0] & 0x7fffffffffffffffUL; rd->q[1] = rs2->q[1]; break; 458 case FNEGQ: rd->q[0] = rs2->q[0] ^ 0x8000000000000000UL; rd->q[1] = rs2->q[1]; break; 459 /* float to int */ 460 case FSTOI: FP_TO_INT_S (IR, SB, 32, 1); break; 461 case FDTOI: FP_TO_INT_D (IR, DB, 32, 1); break; 462 case FQTOI: FP_TO_INT_Q (IR, QB, 32, 1); break; 463 case FSTOX: FP_TO_INT_S (XR, SB, 64, 1); break; 464 case FDTOX: FP_TO_INT_D (XR, DB, 64, 1); break; 465 case FQTOX: FP_TO_INT_Q (XR, QB, 64, 1); break; 466 /* int to float */ 467 case FITOQ: IR = rs2->s; FP_FROM_INT_Q (QR, IR, 32, int); break; 468 case FXTOQ: XR = rs2->d; FP_FROM_INT_Q (QR, XR, 64, long); break; 469 /* Only Ultra-III generates these */ 470 case FXTOS: XR = rs2->d; FP_FROM_INT_S (SR, XR, 64, long); break; 471 case FXTOD: XR = rs2->d; FP_FROM_INT_D (DR, XR, 64, long); break; 472 #if 0 /* Optimized inline in sparc64/kernel/entry.S */ 473 case FITOS: IR = rs2->s; FP_FROM_INT_S (SR, IR, 32, int); break; 474 #endif 475 case FITOD: IR = rs2->s; FP_FROM_INT_D (DR, IR, 32, int); break; 476 /* float to float */ 477 case FSTOD: FP_CONV (D, S, 1, 1, DR, SB); break; 478 case FSTOQ: FP_CONV (Q, S, 2, 1, QR, SB); break; 479 case FDTOQ: FP_CONV (Q, D, 2, 1, QR, DB); break; 480 case FDTOS: FP_CONV (S, D, 1, 1, SR, DB); break; 481 case FQTOS: FP_CONV (S, Q, 1, 2, SR, QB); break; 482 case FQTOD: FP_CONV (D, Q, 1, 2, DR, QB); break; 483 /* comparison */ 484 case FCMPQ: 485 case FCMPEQ: 486 FP_CMP_Q(XR, QB, QA, 3); 487 if (XR == 3 && 488 (((insn >> 5) & 0x1ff) == FCMPEQ || 489 FP_ISSIGNAN_Q(QA) || 490 FP_ISSIGNAN_Q(QB))) 491 FP_SET_EXCEPTION (FP_EX_INVALID); 492 } 493 if (!FP_INHIBIT_RESULTS) { 494 switch ((type >> 6) & 0x7) { 495 case 0: xfsr = current_thread_info()->xfsr[0]; 496 if (XR == -1) XR = 2; 497 switch (freg & 3) { 498 /* fcc0, 1, 2, 3 */ 499 case 0: xfsr &= ~0xc00; xfsr |= (XR << 10); break; 500 case 1: xfsr &= ~0x300000000UL; xfsr |= (XR << 32); break; 501 case 2: xfsr &= ~0xc00000000UL; xfsr |= (XR << 34); break; 502 case 3: xfsr &= ~0x3000000000UL; xfsr |= (XR << 36); break; 503 } 504 current_thread_info()->xfsr[0] = xfsr; 505 break; 506 case 1: rd->s = IR; break; 507 case 2: rd->d = XR; break; 508 case 5: FP_PACK_SP (rd, SR); break; 509 case 6: FP_PACK_DP (rd, DR); break; 510 case 7: FP_PACK_QP (rd, QR); break; 511 } 512 } 513 514 if(_fex != 0) 515 return record_exception(regs, _fex); 516 517 /* Success and no exceptions detected. */ 518 current_thread_info()->xfsr[0] &= ~(FSR_CEXC_MASK); 519 regs->tpc = regs->tnpc; 520 regs->tnpc += 4; 521 return 1; 522 } 523 err: return 0; 524 } 525