1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * arch/powerpc/math-emu/math_efp.c 4 * 5 * Copyright (C) 2006-2008, 2010 Freescale Semiconductor, Inc. 6 * 7 * Author: Ebony Zhu, <ebony.zhu@freescale.com> 8 * Yu Liu, <yu.liu@freescale.com> 9 * 10 * Derived from arch/alpha/math-emu/math.c 11 * arch/powerpc/math-emu/math.c 12 * 13 * Description: 14 * This file is the exception handler to make E500 SPE instructions 15 * fully comply with IEEE-754 floating point standard. 16 */ 17 18 #include <linux/types.h> 19 #include <linux/prctl.h> 20 21 #include <linux/uaccess.h> 22 #include <asm/reg.h> 23 24 #define FP_EX_BOOKE_E500_SPE 25 #include <asm/sfp-machine.h> 26 27 #include <math-emu/soft-fp.h> 28 #include <math-emu/single.h> 29 #include <math-emu/double.h> 30 31 #define EFAPU 0x4 32 33 #define VCT 0x4 34 #define SPFP 0x6 35 #define DPFP 0x7 36 37 #define EFSADD 0x2c0 38 #define EFSSUB 0x2c1 39 #define EFSABS 0x2c4 40 #define EFSNABS 0x2c5 41 #define EFSNEG 0x2c6 42 #define EFSMUL 0x2c8 43 #define EFSDIV 0x2c9 44 #define EFSCMPGT 0x2cc 45 #define EFSCMPLT 0x2cd 46 #define EFSCMPEQ 0x2ce 47 #define EFSCFD 0x2cf 48 #define EFSCFSI 0x2d1 49 #define EFSCTUI 0x2d4 50 #define EFSCTSI 0x2d5 51 #define EFSCTUF 0x2d6 52 #define EFSCTSF 0x2d7 53 #define EFSCTUIZ 0x2d8 54 #define EFSCTSIZ 0x2da 55 56 #define EVFSADD 0x280 57 #define EVFSSUB 0x281 58 #define EVFSABS 0x284 59 #define EVFSNABS 0x285 60 #define EVFSNEG 0x286 61 #define EVFSMUL 0x288 62 #define EVFSDIV 0x289 63 #define EVFSCMPGT 0x28c 64 #define EVFSCMPLT 0x28d 65 #define EVFSCMPEQ 0x28e 66 #define EVFSCTUI 0x294 67 #define EVFSCTSI 0x295 68 #define EVFSCTUF 0x296 69 #define EVFSCTSF 0x297 70 #define EVFSCTUIZ 0x298 71 #define EVFSCTSIZ 0x29a 72 73 #define EFDADD 0x2e0 74 #define EFDSUB 0x2e1 75 #define EFDABS 0x2e4 76 #define EFDNABS 0x2e5 77 #define EFDNEG 0x2e6 78 #define EFDMUL 0x2e8 79 #define EFDDIV 0x2e9 80 #define EFDCTUIDZ 0x2ea 81 #define EFDCTSIDZ 0x2eb 82 #define EFDCMPGT 0x2ec 83 #define EFDCMPLT 0x2ed 84 #define EFDCMPEQ 0x2ee 85 #define EFDCFS 0x2ef 86 #define EFDCTUI 0x2f4 87 #define EFDCTSI 0x2f5 88 #define EFDCTUF 0x2f6 89 #define EFDCTSF 0x2f7 90 #define EFDCTUIZ 0x2f8 91 #define EFDCTSIZ 0x2fa 92 93 #define AB 2 94 #define XA 3 95 #define XB 4 96 #define XCR 5 97 #define NOTYPE 0 98 99 #define SIGN_BIT_S (1UL << 31) 100 #define SIGN_BIT_D (1ULL << 63) 101 #define FP_EX_MASK (FP_EX_INEXACT | FP_EX_INVALID | FP_EX_DIVZERO | \ 102 FP_EX_UNDERFLOW | FP_EX_OVERFLOW) 103 104 static int have_e500_cpu_a005_erratum; 105 106 union dw_union { 107 u64 dp[1]; 108 u32 wp[2]; 109 }; 110 111 static unsigned long insn_type(unsigned long speinsn) 112 { 113 unsigned long ret = NOTYPE; 114 115 switch (speinsn & 0x7ff) { 116 case EFSABS: ret = XA; break; 117 case EFSADD: ret = AB; break; 118 case EFSCFD: ret = XB; break; 119 case EFSCMPEQ: ret = XCR; break; 120 case EFSCMPGT: ret = XCR; break; 121 case EFSCMPLT: ret = XCR; break; 122 case EFSCTSF: ret = XB; break; 123 case EFSCTSI: ret = XB; break; 124 case EFSCTSIZ: ret = XB; break; 125 case EFSCTUF: ret = XB; break; 126 case EFSCTUI: ret = XB; break; 127 case EFSCTUIZ: ret = XB; break; 128 case EFSDIV: ret = AB; break; 129 case EFSMUL: ret = AB; break; 130 case EFSNABS: ret = XA; break; 131 case EFSNEG: ret = XA; break; 132 case EFSSUB: ret = AB; break; 133 case EFSCFSI: ret = XB; break; 134 135 case EVFSABS: ret = XA; break; 136 case EVFSADD: ret = AB; break; 137 case EVFSCMPEQ: ret = XCR; break; 138 case EVFSCMPGT: ret = XCR; break; 139 case EVFSCMPLT: ret = XCR; break; 140 case EVFSCTSF: ret = XB; break; 141 case EVFSCTSI: ret = XB; break; 142 case EVFSCTSIZ: ret = XB; break; 143 case EVFSCTUF: ret = XB; break; 144 case EVFSCTUI: ret = XB; break; 145 case EVFSCTUIZ: ret = XB; break; 146 case EVFSDIV: ret = AB; break; 147 case EVFSMUL: ret = AB; break; 148 case EVFSNABS: ret = XA; break; 149 case EVFSNEG: ret = XA; break; 150 case EVFSSUB: ret = AB; break; 151 152 case EFDABS: ret = XA; break; 153 case EFDADD: ret = AB; break; 154 case EFDCFS: ret = XB; break; 155 case EFDCMPEQ: ret = XCR; break; 156 case EFDCMPGT: ret = XCR; break; 157 case EFDCMPLT: ret = XCR; break; 158 case EFDCTSF: ret = XB; break; 159 case EFDCTSI: ret = XB; break; 160 case EFDCTSIDZ: ret = XB; break; 161 case EFDCTSIZ: ret = XB; break; 162 case EFDCTUF: ret = XB; break; 163 case EFDCTUI: ret = XB; break; 164 case EFDCTUIDZ: ret = XB; break; 165 case EFDCTUIZ: ret = XB; break; 166 case EFDDIV: ret = AB; break; 167 case EFDMUL: ret = AB; break; 168 case EFDNABS: ret = XA; break; 169 case EFDNEG: ret = XA; break; 170 case EFDSUB: ret = AB; break; 171 } 172 173 return ret; 174 } 175 176 int do_spe_mathemu(struct pt_regs *regs) 177 { 178 FP_DECL_EX; 179 int IR, cmp; 180 181 unsigned long type, func, fc, fa, fb, src, speinsn; 182 union dw_union vc, va, vb; 183 184 if (get_user(speinsn, (unsigned int __user *) regs->nip)) 185 return -EFAULT; 186 if ((speinsn >> 26) != EFAPU) 187 return -EINVAL; /* not an spe instruction */ 188 189 type = insn_type(speinsn); 190 if (type == NOTYPE) 191 goto illegal; 192 193 func = speinsn & 0x7ff; 194 fc = (speinsn >> 21) & 0x1f; 195 fa = (speinsn >> 16) & 0x1f; 196 fb = (speinsn >> 11) & 0x1f; 197 src = (speinsn >> 5) & 0x7; 198 199 vc.wp[0] = current->thread.evr[fc]; 200 vc.wp[1] = regs->gpr[fc]; 201 va.wp[0] = current->thread.evr[fa]; 202 va.wp[1] = regs->gpr[fa]; 203 vb.wp[0] = current->thread.evr[fb]; 204 vb.wp[1] = regs->gpr[fb]; 205 206 __FPU_FPSCR = mfspr(SPRN_SPEFSCR); 207 208 pr_debug("speinsn:%08lx spefscr:%08lx\n", speinsn, __FPU_FPSCR); 209 pr_debug("vc: %08x %08x\n", vc.wp[0], vc.wp[1]); 210 pr_debug("va: %08x %08x\n", va.wp[0], va.wp[1]); 211 pr_debug("vb: %08x %08x\n", vb.wp[0], vb.wp[1]); 212 213 switch (src) { 214 case SPFP: { 215 FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR); 216 217 switch (type) { 218 case AB: 219 case XCR: 220 FP_UNPACK_SP(SA, va.wp + 1); 221 case XB: 222 FP_UNPACK_SP(SB, vb.wp + 1); 223 break; 224 case XA: 225 FP_UNPACK_SP(SA, va.wp + 1); 226 break; 227 } 228 229 pr_debug("SA: %ld %08lx %ld (%ld)\n", SA_s, SA_f, SA_e, SA_c); 230 pr_debug("SB: %ld %08lx %ld (%ld)\n", SB_s, SB_f, SB_e, SB_c); 231 232 switch (func) { 233 case EFSABS: 234 vc.wp[1] = va.wp[1] & ~SIGN_BIT_S; 235 goto update_regs; 236 237 case EFSNABS: 238 vc.wp[1] = va.wp[1] | SIGN_BIT_S; 239 goto update_regs; 240 241 case EFSNEG: 242 vc.wp[1] = va.wp[1] ^ SIGN_BIT_S; 243 goto update_regs; 244 245 case EFSADD: 246 FP_ADD_S(SR, SA, SB); 247 goto pack_s; 248 249 case EFSSUB: 250 FP_SUB_S(SR, SA, SB); 251 goto pack_s; 252 253 case EFSMUL: 254 FP_MUL_S(SR, SA, SB); 255 goto pack_s; 256 257 case EFSDIV: 258 FP_DIV_S(SR, SA, SB); 259 goto pack_s; 260 261 case EFSCMPEQ: 262 cmp = 0; 263 goto cmp_s; 264 265 case EFSCMPGT: 266 cmp = 1; 267 goto cmp_s; 268 269 case EFSCMPLT: 270 cmp = -1; 271 goto cmp_s; 272 273 case EFSCTSF: 274 case EFSCTUF: 275 if (SB_c == FP_CLS_NAN) { 276 vc.wp[1] = 0; 277 FP_SET_EXCEPTION(FP_EX_INVALID); 278 } else { 279 SB_e += (func == EFSCTSF ? 31 : 32); 280 FP_TO_INT_ROUND_S(vc.wp[1], SB, 32, 281 (func == EFSCTSF)); 282 } 283 goto update_regs; 284 285 case EFSCFD: { 286 FP_DECL_D(DB); 287 FP_CLEAR_EXCEPTIONS; 288 FP_UNPACK_DP(DB, vb.dp); 289 290 pr_debug("DB: %ld %08lx %08lx %ld (%ld)\n", 291 DB_s, DB_f1, DB_f0, DB_e, DB_c); 292 293 FP_CONV(S, D, 1, 2, SR, DB); 294 goto pack_s; 295 } 296 297 case EFSCTSI: 298 case EFSCTUI: 299 if (SB_c == FP_CLS_NAN) { 300 vc.wp[1] = 0; 301 FP_SET_EXCEPTION(FP_EX_INVALID); 302 } else { 303 FP_TO_INT_ROUND_S(vc.wp[1], SB, 32, 304 ((func & 0x3) != 0)); 305 } 306 goto update_regs; 307 308 case EFSCTSIZ: 309 case EFSCTUIZ: 310 if (SB_c == FP_CLS_NAN) { 311 vc.wp[1] = 0; 312 FP_SET_EXCEPTION(FP_EX_INVALID); 313 } else { 314 FP_TO_INT_S(vc.wp[1], SB, 32, 315 ((func & 0x3) != 0)); 316 } 317 goto update_regs; 318 319 default: 320 goto illegal; 321 } 322 break; 323 324 pack_s: 325 pr_debug("SR: %ld %08lx %ld (%ld)\n", SR_s, SR_f, SR_e, SR_c); 326 327 FP_PACK_SP(vc.wp + 1, SR); 328 goto update_regs; 329 330 cmp_s: 331 FP_CMP_S(IR, SA, SB, 3); 332 if (IR == 3 && (FP_ISSIGNAN_S(SA) || FP_ISSIGNAN_S(SB))) 333 FP_SET_EXCEPTION(FP_EX_INVALID); 334 if (IR == cmp) { 335 IR = 0x4; 336 } else { 337 IR = 0; 338 } 339 goto update_ccr; 340 } 341 342 case DPFP: { 343 FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR); 344 345 switch (type) { 346 case AB: 347 case XCR: 348 FP_UNPACK_DP(DA, va.dp); 349 case XB: 350 FP_UNPACK_DP(DB, vb.dp); 351 break; 352 case XA: 353 FP_UNPACK_DP(DA, va.dp); 354 break; 355 } 356 357 pr_debug("DA: %ld %08lx %08lx %ld (%ld)\n", 358 DA_s, DA_f1, DA_f0, DA_e, DA_c); 359 pr_debug("DB: %ld %08lx %08lx %ld (%ld)\n", 360 DB_s, DB_f1, DB_f0, DB_e, DB_c); 361 362 switch (func) { 363 case EFDABS: 364 vc.dp[0] = va.dp[0] & ~SIGN_BIT_D; 365 goto update_regs; 366 367 case EFDNABS: 368 vc.dp[0] = va.dp[0] | SIGN_BIT_D; 369 goto update_regs; 370 371 case EFDNEG: 372 vc.dp[0] = va.dp[0] ^ SIGN_BIT_D; 373 goto update_regs; 374 375 case EFDADD: 376 FP_ADD_D(DR, DA, DB); 377 goto pack_d; 378 379 case EFDSUB: 380 FP_SUB_D(DR, DA, DB); 381 goto pack_d; 382 383 case EFDMUL: 384 FP_MUL_D(DR, DA, DB); 385 goto pack_d; 386 387 case EFDDIV: 388 FP_DIV_D(DR, DA, DB); 389 goto pack_d; 390 391 case EFDCMPEQ: 392 cmp = 0; 393 goto cmp_d; 394 395 case EFDCMPGT: 396 cmp = 1; 397 goto cmp_d; 398 399 case EFDCMPLT: 400 cmp = -1; 401 goto cmp_d; 402 403 case EFDCTSF: 404 case EFDCTUF: 405 if (DB_c == FP_CLS_NAN) { 406 vc.wp[1] = 0; 407 FP_SET_EXCEPTION(FP_EX_INVALID); 408 } else { 409 DB_e += (func == EFDCTSF ? 31 : 32); 410 FP_TO_INT_ROUND_D(vc.wp[1], DB, 32, 411 (func == EFDCTSF)); 412 } 413 goto update_regs; 414 415 case EFDCFS: { 416 FP_DECL_S(SB); 417 FP_CLEAR_EXCEPTIONS; 418 FP_UNPACK_SP(SB, vb.wp + 1); 419 420 pr_debug("SB: %ld %08lx %ld (%ld)\n", 421 SB_s, SB_f, SB_e, SB_c); 422 423 FP_CONV(D, S, 2, 1, DR, SB); 424 goto pack_d; 425 } 426 427 case EFDCTUIDZ: 428 case EFDCTSIDZ: 429 if (DB_c == FP_CLS_NAN) { 430 vc.dp[0] = 0; 431 FP_SET_EXCEPTION(FP_EX_INVALID); 432 } else { 433 FP_TO_INT_D(vc.dp[0], DB, 64, 434 ((func & 0x1) == 0)); 435 } 436 goto update_regs; 437 438 case EFDCTUI: 439 case EFDCTSI: 440 if (DB_c == FP_CLS_NAN) { 441 vc.wp[1] = 0; 442 FP_SET_EXCEPTION(FP_EX_INVALID); 443 } else { 444 FP_TO_INT_ROUND_D(vc.wp[1], DB, 32, 445 ((func & 0x3) != 0)); 446 } 447 goto update_regs; 448 449 case EFDCTUIZ: 450 case EFDCTSIZ: 451 if (DB_c == FP_CLS_NAN) { 452 vc.wp[1] = 0; 453 FP_SET_EXCEPTION(FP_EX_INVALID); 454 } else { 455 FP_TO_INT_D(vc.wp[1], DB, 32, 456 ((func & 0x3) != 0)); 457 } 458 goto update_regs; 459 460 default: 461 goto illegal; 462 } 463 break; 464 465 pack_d: 466 pr_debug("DR: %ld %08lx %08lx %ld (%ld)\n", 467 DR_s, DR_f1, DR_f0, DR_e, DR_c); 468 469 FP_PACK_DP(vc.dp, DR); 470 goto update_regs; 471 472 cmp_d: 473 FP_CMP_D(IR, DA, DB, 3); 474 if (IR == 3 && (FP_ISSIGNAN_D(DA) || FP_ISSIGNAN_D(DB))) 475 FP_SET_EXCEPTION(FP_EX_INVALID); 476 if (IR == cmp) { 477 IR = 0x4; 478 } else { 479 IR = 0; 480 } 481 goto update_ccr; 482 483 } 484 485 case VCT: { 486 FP_DECL_S(SA0); FP_DECL_S(SB0); FP_DECL_S(SR0); 487 FP_DECL_S(SA1); FP_DECL_S(SB1); FP_DECL_S(SR1); 488 int IR0, IR1; 489 490 switch (type) { 491 case AB: 492 case XCR: 493 FP_UNPACK_SP(SA0, va.wp); 494 FP_UNPACK_SP(SA1, va.wp + 1); 495 case XB: 496 FP_UNPACK_SP(SB0, vb.wp); 497 FP_UNPACK_SP(SB1, vb.wp + 1); 498 break; 499 case XA: 500 FP_UNPACK_SP(SA0, va.wp); 501 FP_UNPACK_SP(SA1, va.wp + 1); 502 break; 503 } 504 505 pr_debug("SA0: %ld %08lx %ld (%ld)\n", 506 SA0_s, SA0_f, SA0_e, SA0_c); 507 pr_debug("SA1: %ld %08lx %ld (%ld)\n", 508 SA1_s, SA1_f, SA1_e, SA1_c); 509 pr_debug("SB0: %ld %08lx %ld (%ld)\n", 510 SB0_s, SB0_f, SB0_e, SB0_c); 511 pr_debug("SB1: %ld %08lx %ld (%ld)\n", 512 SB1_s, SB1_f, SB1_e, SB1_c); 513 514 switch (func) { 515 case EVFSABS: 516 vc.wp[0] = va.wp[0] & ~SIGN_BIT_S; 517 vc.wp[1] = va.wp[1] & ~SIGN_BIT_S; 518 goto update_regs; 519 520 case EVFSNABS: 521 vc.wp[0] = va.wp[0] | SIGN_BIT_S; 522 vc.wp[1] = va.wp[1] | SIGN_BIT_S; 523 goto update_regs; 524 525 case EVFSNEG: 526 vc.wp[0] = va.wp[0] ^ SIGN_BIT_S; 527 vc.wp[1] = va.wp[1] ^ SIGN_BIT_S; 528 goto update_regs; 529 530 case EVFSADD: 531 FP_ADD_S(SR0, SA0, SB0); 532 FP_ADD_S(SR1, SA1, SB1); 533 goto pack_vs; 534 535 case EVFSSUB: 536 FP_SUB_S(SR0, SA0, SB0); 537 FP_SUB_S(SR1, SA1, SB1); 538 goto pack_vs; 539 540 case EVFSMUL: 541 FP_MUL_S(SR0, SA0, SB0); 542 FP_MUL_S(SR1, SA1, SB1); 543 goto pack_vs; 544 545 case EVFSDIV: 546 FP_DIV_S(SR0, SA0, SB0); 547 FP_DIV_S(SR1, SA1, SB1); 548 goto pack_vs; 549 550 case EVFSCMPEQ: 551 cmp = 0; 552 goto cmp_vs; 553 554 case EVFSCMPGT: 555 cmp = 1; 556 goto cmp_vs; 557 558 case EVFSCMPLT: 559 cmp = -1; 560 goto cmp_vs; 561 562 case EVFSCTUF: 563 case EVFSCTSF: 564 if (SB0_c == FP_CLS_NAN) { 565 vc.wp[0] = 0; 566 FP_SET_EXCEPTION(FP_EX_INVALID); 567 } else { 568 SB0_e += (func == EVFSCTSF ? 31 : 32); 569 FP_TO_INT_ROUND_S(vc.wp[0], SB0, 32, 570 (func == EVFSCTSF)); 571 } 572 if (SB1_c == FP_CLS_NAN) { 573 vc.wp[1] = 0; 574 FP_SET_EXCEPTION(FP_EX_INVALID); 575 } else { 576 SB1_e += (func == EVFSCTSF ? 31 : 32); 577 FP_TO_INT_ROUND_S(vc.wp[1], SB1, 32, 578 (func == EVFSCTSF)); 579 } 580 goto update_regs; 581 582 case EVFSCTUI: 583 case EVFSCTSI: 584 if (SB0_c == FP_CLS_NAN) { 585 vc.wp[0] = 0; 586 FP_SET_EXCEPTION(FP_EX_INVALID); 587 } else { 588 FP_TO_INT_ROUND_S(vc.wp[0], SB0, 32, 589 ((func & 0x3) != 0)); 590 } 591 if (SB1_c == FP_CLS_NAN) { 592 vc.wp[1] = 0; 593 FP_SET_EXCEPTION(FP_EX_INVALID); 594 } else { 595 FP_TO_INT_ROUND_S(vc.wp[1], SB1, 32, 596 ((func & 0x3) != 0)); 597 } 598 goto update_regs; 599 600 case EVFSCTUIZ: 601 case EVFSCTSIZ: 602 if (SB0_c == FP_CLS_NAN) { 603 vc.wp[0] = 0; 604 FP_SET_EXCEPTION(FP_EX_INVALID); 605 } else { 606 FP_TO_INT_S(vc.wp[0], SB0, 32, 607 ((func & 0x3) != 0)); 608 } 609 if (SB1_c == FP_CLS_NAN) { 610 vc.wp[1] = 0; 611 FP_SET_EXCEPTION(FP_EX_INVALID); 612 } else { 613 FP_TO_INT_S(vc.wp[1], SB1, 32, 614 ((func & 0x3) != 0)); 615 } 616 goto update_regs; 617 618 default: 619 goto illegal; 620 } 621 break; 622 623 pack_vs: 624 pr_debug("SR0: %ld %08lx %ld (%ld)\n", 625 SR0_s, SR0_f, SR0_e, SR0_c); 626 pr_debug("SR1: %ld %08lx %ld (%ld)\n", 627 SR1_s, SR1_f, SR1_e, SR1_c); 628 629 FP_PACK_SP(vc.wp, SR0); 630 FP_PACK_SP(vc.wp + 1, SR1); 631 goto update_regs; 632 633 cmp_vs: 634 { 635 int ch, cl; 636 637 FP_CMP_S(IR0, SA0, SB0, 3); 638 FP_CMP_S(IR1, SA1, SB1, 3); 639 if (IR0 == 3 && (FP_ISSIGNAN_S(SA0) || FP_ISSIGNAN_S(SB0))) 640 FP_SET_EXCEPTION(FP_EX_INVALID); 641 if (IR1 == 3 && (FP_ISSIGNAN_S(SA1) || FP_ISSIGNAN_S(SB1))) 642 FP_SET_EXCEPTION(FP_EX_INVALID); 643 ch = (IR0 == cmp) ? 1 : 0; 644 cl = (IR1 == cmp) ? 1 : 0; 645 IR = (ch << 3) | (cl << 2) | ((ch | cl) << 1) | 646 ((ch & cl) << 0); 647 goto update_ccr; 648 } 649 } 650 default: 651 return -EINVAL; 652 } 653 654 update_ccr: 655 regs->ccr &= ~(15 << ((7 - ((speinsn >> 23) & 0x7)) << 2)); 656 regs->ccr |= (IR << ((7 - ((speinsn >> 23) & 0x7)) << 2)); 657 658 update_regs: 659 /* 660 * If the "invalid" exception sticky bit was set by the 661 * processor for non-finite input, but was not set before the 662 * instruction being emulated, clear it. Likewise for the 663 * "underflow" bit, which may have been set by the processor 664 * for exact underflow, not just inexact underflow when the 665 * flag should be set for IEEE 754 semantics. Other sticky 666 * exceptions will only be set by the processor when they are 667 * correct according to IEEE 754 semantics, and we must not 668 * clear sticky bits that were already set before the emulated 669 * instruction as they represent the user-visible sticky 670 * exception status. "inexact" traps to kernel are not 671 * required for IEEE semantics and are not enabled by default, 672 * so the "inexact" sticky bit may have been set by a previous 673 * instruction without the kernel being aware of it. 674 */ 675 __FPU_FPSCR 676 &= ~(FP_EX_INVALID | FP_EX_UNDERFLOW) | current->thread.spefscr_last; 677 __FPU_FPSCR |= (FP_CUR_EXCEPTIONS & FP_EX_MASK); 678 mtspr(SPRN_SPEFSCR, __FPU_FPSCR); 679 current->thread.spefscr_last = __FPU_FPSCR; 680 681 current->thread.evr[fc] = vc.wp[0]; 682 regs->gpr[fc] = vc.wp[1]; 683 684 pr_debug("ccr = %08lx\n", regs->ccr); 685 pr_debug("cur exceptions = %08x spefscr = %08lx\n", 686 FP_CUR_EXCEPTIONS, __FPU_FPSCR); 687 pr_debug("vc: %08x %08x\n", vc.wp[0], vc.wp[1]); 688 pr_debug("va: %08x %08x\n", va.wp[0], va.wp[1]); 689 pr_debug("vb: %08x %08x\n", vb.wp[0], vb.wp[1]); 690 691 if (current->thread.fpexc_mode & PR_FP_EXC_SW_ENABLE) { 692 if ((FP_CUR_EXCEPTIONS & FP_EX_DIVZERO) 693 && (current->thread.fpexc_mode & PR_FP_EXC_DIV)) 694 return 1; 695 if ((FP_CUR_EXCEPTIONS & FP_EX_OVERFLOW) 696 && (current->thread.fpexc_mode & PR_FP_EXC_OVF)) 697 return 1; 698 if ((FP_CUR_EXCEPTIONS & FP_EX_UNDERFLOW) 699 && (current->thread.fpexc_mode & PR_FP_EXC_UND)) 700 return 1; 701 if ((FP_CUR_EXCEPTIONS & FP_EX_INEXACT) 702 && (current->thread.fpexc_mode & PR_FP_EXC_RES)) 703 return 1; 704 if ((FP_CUR_EXCEPTIONS & FP_EX_INVALID) 705 && (current->thread.fpexc_mode & PR_FP_EXC_INV)) 706 return 1; 707 } 708 return 0; 709 710 illegal: 711 if (have_e500_cpu_a005_erratum) { 712 /* according to e500 cpu a005 erratum, reissue efp inst */ 713 regs->nip -= 4; 714 pr_debug("re-issue efp inst: %08lx\n", speinsn); 715 return 0; 716 } 717 718 printk(KERN_ERR "\nOoops! IEEE-754 compliance handler encountered un-supported instruction.\ninst code: %08lx\n", speinsn); 719 return -ENOSYS; 720 } 721 722 int speround_handler(struct pt_regs *regs) 723 { 724 union dw_union fgpr; 725 int s_lo, s_hi; 726 int lo_inexact, hi_inexact; 727 int fp_result; 728 unsigned long speinsn, type, fb, fc, fptype, func; 729 730 if (get_user(speinsn, (unsigned int __user *) regs->nip)) 731 return -EFAULT; 732 if ((speinsn >> 26) != 4) 733 return -EINVAL; /* not an spe instruction */ 734 735 func = speinsn & 0x7ff; 736 type = insn_type(func); 737 if (type == XCR) return -ENOSYS; 738 739 __FPU_FPSCR = mfspr(SPRN_SPEFSCR); 740 pr_debug("speinsn:%08lx spefscr:%08lx\n", speinsn, __FPU_FPSCR); 741 742 fptype = (speinsn >> 5) & 0x7; 743 744 /* No need to round if the result is exact */ 745 lo_inexact = __FPU_FPSCR & (SPEFSCR_FG | SPEFSCR_FX); 746 hi_inexact = __FPU_FPSCR & (SPEFSCR_FGH | SPEFSCR_FXH); 747 if (!(lo_inexact || (hi_inexact && fptype == VCT))) 748 return 0; 749 750 fc = (speinsn >> 21) & 0x1f; 751 s_lo = regs->gpr[fc] & SIGN_BIT_S; 752 s_hi = current->thread.evr[fc] & SIGN_BIT_S; 753 fgpr.wp[0] = current->thread.evr[fc]; 754 fgpr.wp[1] = regs->gpr[fc]; 755 756 fb = (speinsn >> 11) & 0x1f; 757 switch (func) { 758 case EFSCTUIZ: 759 case EFSCTSIZ: 760 case EVFSCTUIZ: 761 case EVFSCTSIZ: 762 case EFDCTUIDZ: 763 case EFDCTSIDZ: 764 case EFDCTUIZ: 765 case EFDCTSIZ: 766 /* 767 * These instructions always round to zero, 768 * independent of the rounding mode. 769 */ 770 return 0; 771 772 case EFSCTUI: 773 case EFSCTUF: 774 case EVFSCTUI: 775 case EVFSCTUF: 776 case EFDCTUI: 777 case EFDCTUF: 778 fp_result = 0; 779 s_lo = 0; 780 s_hi = 0; 781 break; 782 783 case EFSCTSI: 784 case EFSCTSF: 785 fp_result = 0; 786 /* Recover the sign of a zero result if possible. */ 787 if (fgpr.wp[1] == 0) 788 s_lo = regs->gpr[fb] & SIGN_BIT_S; 789 break; 790 791 case EVFSCTSI: 792 case EVFSCTSF: 793 fp_result = 0; 794 /* Recover the sign of a zero result if possible. */ 795 if (fgpr.wp[1] == 0) 796 s_lo = regs->gpr[fb] & SIGN_BIT_S; 797 if (fgpr.wp[0] == 0) 798 s_hi = current->thread.evr[fb] & SIGN_BIT_S; 799 break; 800 801 case EFDCTSI: 802 case EFDCTSF: 803 fp_result = 0; 804 s_hi = s_lo; 805 /* Recover the sign of a zero result if possible. */ 806 if (fgpr.wp[1] == 0) 807 s_hi = current->thread.evr[fb] & SIGN_BIT_S; 808 break; 809 810 default: 811 fp_result = 1; 812 break; 813 } 814 815 pr_debug("round fgpr: %08x %08x\n", fgpr.wp[0], fgpr.wp[1]); 816 817 switch (fptype) { 818 /* Since SPE instructions on E500 core can handle round to nearest 819 * and round toward zero with IEEE-754 complied, we just need 820 * to handle round toward +Inf and round toward -Inf by software. 821 */ 822 case SPFP: 823 if ((FP_ROUNDMODE) == FP_RND_PINF) { 824 if (!s_lo) fgpr.wp[1]++; /* Z > 0, choose Z1 */ 825 } else { /* round to -Inf */ 826 if (s_lo) { 827 if (fp_result) 828 fgpr.wp[1]++; /* Z < 0, choose Z2 */ 829 else 830 fgpr.wp[1]--; /* Z < 0, choose Z2 */ 831 } 832 } 833 break; 834 835 case DPFP: 836 if (FP_ROUNDMODE == FP_RND_PINF) { 837 if (!s_hi) { 838 if (fp_result) 839 fgpr.dp[0]++; /* Z > 0, choose Z1 */ 840 else 841 fgpr.wp[1]++; /* Z > 0, choose Z1 */ 842 } 843 } else { /* round to -Inf */ 844 if (s_hi) { 845 if (fp_result) 846 fgpr.dp[0]++; /* Z < 0, choose Z2 */ 847 else 848 fgpr.wp[1]--; /* Z < 0, choose Z2 */ 849 } 850 } 851 break; 852 853 case VCT: 854 if (FP_ROUNDMODE == FP_RND_PINF) { 855 if (lo_inexact && !s_lo) 856 fgpr.wp[1]++; /* Z_low > 0, choose Z1 */ 857 if (hi_inexact && !s_hi) 858 fgpr.wp[0]++; /* Z_high word > 0, choose Z1 */ 859 } else { /* round to -Inf */ 860 if (lo_inexact && s_lo) { 861 if (fp_result) 862 fgpr.wp[1]++; /* Z_low < 0, choose Z2 */ 863 else 864 fgpr.wp[1]--; /* Z_low < 0, choose Z2 */ 865 } 866 if (hi_inexact && s_hi) { 867 if (fp_result) 868 fgpr.wp[0]++; /* Z_high < 0, choose Z2 */ 869 else 870 fgpr.wp[0]--; /* Z_high < 0, choose Z2 */ 871 } 872 } 873 break; 874 875 default: 876 return -EINVAL; 877 } 878 879 current->thread.evr[fc] = fgpr.wp[0]; 880 regs->gpr[fc] = fgpr.wp[1]; 881 882 pr_debug(" to fgpr: %08x %08x\n", fgpr.wp[0], fgpr.wp[1]); 883 884 if (current->thread.fpexc_mode & PR_FP_EXC_SW_ENABLE) 885 return (current->thread.fpexc_mode & PR_FP_EXC_RES) ? 1 : 0; 886 return 0; 887 } 888 889 int __init spe_mathemu_init(void) 890 { 891 u32 pvr, maj, min; 892 893 pvr = mfspr(SPRN_PVR); 894 895 if ((PVR_VER(pvr) == PVR_VER_E500V1) || 896 (PVR_VER(pvr) == PVR_VER_E500V2)) { 897 maj = PVR_MAJ(pvr); 898 min = PVR_MIN(pvr); 899 900 /* 901 * E500 revision below 1.1, 2.3, 3.1, 4.1, 5.1 902 * need cpu a005 errata workaround 903 */ 904 switch (maj) { 905 case 1: 906 if (min < 1) 907 have_e500_cpu_a005_erratum = 1; 908 break; 909 case 2: 910 if (min < 3) 911 have_e500_cpu_a005_erratum = 1; 912 break; 913 case 3: 914 case 4: 915 case 5: 916 if (min < 1) 917 have_e500_cpu_a005_erratum = 1; 918 break; 919 default: 920 break; 921 } 922 } 923 924 return 0; 925 } 926 927 module_init(spe_mathemu_init); 928