1/* 2 * Copyright (c) 2013 Rob Clark <robdclark@gmail.com> 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 * SOFTWARE. 22 */ 23 24#ifndef INSTR_A3XX_H_ 25#define INSTR_A3XX_H_ 26 27#define PACKED __attribute__((__packed__)) 28 29#include <stdint.h> 30#include <assert.h> 31 32/* size of largest OPC field of all the instruction categories: */ 33#define NOPC_BITS 6 34 35#define _OPC(cat, opc) (((cat) << NOPC_BITS) | opc) 36 37typedef enum { 38 /* category 0: */ 39 OPC_NOP = _OPC(0, 0), 40 OPC_BR = _OPC(0, 1), 41 OPC_JUMP = _OPC(0, 2), 42 OPC_CALL = _OPC(0, 3), 43 OPC_RET = _OPC(0, 4), 44 OPC_KILL = _OPC(0, 5), 45 OPC_END = _OPC(0, 6), 46 OPC_EMIT = _OPC(0, 7), 47 OPC_CUT = _OPC(0, 8), 48 OPC_CHMASK = _OPC(0, 9), 49 OPC_CHSH = _OPC(0, 10), 50 OPC_FLOW_REV = _OPC(0, 11), 51 52 /* category 1: */ 53 OPC_MOV = _OPC(1, 0), 54 55 /* category 2: */ 56 OPC_ADD_F = _OPC(2, 0), 57 OPC_MIN_F = _OPC(2, 1), 58 OPC_MAX_F = _OPC(2, 2), 59 OPC_MUL_F = _OPC(2, 3), 60 OPC_SIGN_F = _OPC(2, 4), 61 OPC_CMPS_F = _OPC(2, 5), 62 OPC_ABSNEG_F = _OPC(2, 6), 63 OPC_CMPV_F = _OPC(2, 7), 64 /* 8 - invalid */ 65 OPC_FLOOR_F = _OPC(2, 9), 66 OPC_CEIL_F = _OPC(2, 10), 67 OPC_RNDNE_F = _OPC(2, 11), 68 OPC_RNDAZ_F = _OPC(2, 12), 69 OPC_TRUNC_F = _OPC(2, 13), 70 /* 14-15 - invalid */ 71 OPC_ADD_U = _OPC(2, 16), 72 OPC_ADD_S = _OPC(2, 17), 73 OPC_SUB_U = _OPC(2, 18), 74 OPC_SUB_S = _OPC(2, 19), 75 OPC_CMPS_U = _OPC(2, 20), 76 OPC_CMPS_S = _OPC(2, 21), 77 OPC_MIN_U = _OPC(2, 22), 78 OPC_MIN_S = _OPC(2, 23), 79 OPC_MAX_U = _OPC(2, 24), 80 OPC_MAX_S = _OPC(2, 25), 81 OPC_ABSNEG_S = _OPC(2, 26), 82 /* 27 - invalid */ 83 OPC_AND_B = _OPC(2, 28), 84 OPC_OR_B = _OPC(2, 29), 85 OPC_NOT_B = _OPC(2, 30), 86 OPC_XOR_B = _OPC(2, 31), 87 /* 32 - invalid */ 88 OPC_CMPV_U = _OPC(2, 33), 89 OPC_CMPV_S = _OPC(2, 34), 90 /* 35-47 - invalid */ 91 OPC_MUL_U = _OPC(2, 48), 92 OPC_MUL_S = _OPC(2, 49), 93 OPC_MULL_U = _OPC(2, 50), 94 OPC_BFREV_B = _OPC(2, 51), 95 OPC_CLZ_S = _OPC(2, 52), 96 OPC_CLZ_B = _OPC(2, 53), 97 OPC_SHL_B = _OPC(2, 54), 98 OPC_SHR_B = _OPC(2, 55), 99 OPC_ASHR_B = _OPC(2, 56), 100 OPC_BARY_F = _OPC(2, 57), 101 OPC_MGEN_B = _OPC(2, 58), 102 OPC_GETBIT_B = _OPC(2, 59), 103 OPC_SETRM = _OPC(2, 60), 104 OPC_CBITS_B = _OPC(2, 61), 105 OPC_SHB = _OPC(2, 62), 106 OPC_MSAD = _OPC(2, 63), 107 108 /* category 3: */ 109 OPC_MAD_U16 = _OPC(3, 0), 110 OPC_MADSH_U16 = _OPC(3, 1), 111 OPC_MAD_S16 = _OPC(3, 2), 112 OPC_MADSH_M16 = _OPC(3, 3), /* should this be .s16? */ 113 OPC_MAD_U24 = _OPC(3, 4), 114 OPC_MAD_S24 = _OPC(3, 5), 115 OPC_MAD_F16 = _OPC(3, 6), 116 OPC_MAD_F32 = _OPC(3, 7), 117 OPC_SEL_B16 = _OPC(3, 8), 118 OPC_SEL_B32 = _OPC(3, 9), 119 OPC_SEL_S16 = _OPC(3, 10), 120 OPC_SEL_S32 = _OPC(3, 11), 121 OPC_SEL_F16 = _OPC(3, 12), 122 OPC_SEL_F32 = _OPC(3, 13), 123 OPC_SAD_S16 = _OPC(3, 14), 124 OPC_SAD_S32 = _OPC(3, 15), 125 126 /* category 4: */ 127 OPC_RCP = _OPC(4, 0), 128 OPC_RSQ = _OPC(4, 1), 129 OPC_LOG2 = _OPC(4, 2), 130 OPC_EXP2 = _OPC(4, 3), 131 OPC_SIN = _OPC(4, 4), 132 OPC_COS = _OPC(4, 5), 133 OPC_SQRT = _OPC(4, 6), 134 // 7-63 - invalid 135 136 /* category 5: */ 137 OPC_ISAM = _OPC(5, 0), 138 OPC_ISAML = _OPC(5, 1), 139 OPC_ISAMM = _OPC(5, 2), 140 OPC_SAM = _OPC(5, 3), 141 OPC_SAMB = _OPC(5, 4), 142 OPC_SAML = _OPC(5, 5), 143 OPC_SAMGQ = _OPC(5, 6), 144 OPC_GETLOD = _OPC(5, 7), 145 OPC_CONV = _OPC(5, 8), 146 OPC_CONVM = _OPC(5, 9), 147 OPC_GETSIZE = _OPC(5, 10), 148 OPC_GETBUF = _OPC(5, 11), 149 OPC_GETPOS = _OPC(5, 12), 150 OPC_GETINFO = _OPC(5, 13), 151 OPC_DSX = _OPC(5, 14), 152 OPC_DSY = _OPC(5, 15), 153 OPC_GATHER4R = _OPC(5, 16), 154 OPC_GATHER4G = _OPC(5, 17), 155 OPC_GATHER4B = _OPC(5, 18), 156 OPC_GATHER4A = _OPC(5, 19), 157 OPC_SAMGP0 = _OPC(5, 20), 158 OPC_SAMGP1 = _OPC(5, 21), 159 OPC_SAMGP2 = _OPC(5, 22), 160 OPC_SAMGP3 = _OPC(5, 23), 161 OPC_DSXPP_1 = _OPC(5, 24), 162 OPC_DSYPP_1 = _OPC(5, 25), 163 OPC_RGETPOS = _OPC(5, 26), 164 OPC_RGETINFO = _OPC(5, 27), 165 166 /* category 6: */ 167 OPC_LDG = _OPC(6, 0), /* load-global */ 168 OPC_LDL = _OPC(6, 1), 169 OPC_LDP = _OPC(6, 2), 170 OPC_STG = _OPC(6, 3), /* store-global */ 171 OPC_STL = _OPC(6, 4), 172 OPC_STP = _OPC(6, 5), 173 OPC_STI = _OPC(6, 6), 174 OPC_G2L = _OPC(6, 7), 175 OPC_L2G = _OPC(6, 8), 176 OPC_PREFETCH = _OPC(6, 9), 177 OPC_LDLW = _OPC(6, 10), 178 OPC_STLW = _OPC(6, 11), 179 OPC_RESFMT = _OPC(6, 14), 180 OPC_RESINFO = _OPC(6, 15), 181 OPC_ATOMIC_ADD = _OPC(6, 16), 182 OPC_ATOMIC_SUB = _OPC(6, 17), 183 OPC_ATOMIC_XCHG = _OPC(6, 18), 184 OPC_ATOMIC_INC = _OPC(6, 19), 185 OPC_ATOMIC_DEC = _OPC(6, 20), 186 OPC_ATOMIC_CMPXCHG = _OPC(6, 21), 187 OPC_ATOMIC_MIN = _OPC(6, 22), 188 OPC_ATOMIC_MAX = _OPC(6, 23), 189 OPC_ATOMIC_AND = _OPC(6, 24), 190 OPC_ATOMIC_OR = _OPC(6, 25), 191 OPC_ATOMIC_XOR = _OPC(6, 26), 192 OPC_LDGB_TYPED_4D = _OPC(6, 27), 193 OPC_STGB_4D_4 = _OPC(6, 28), 194 OPC_STIB = _OPC(6, 29), 195 OPC_LDC_4 = _OPC(6, 30), 196 OPC_LDLV = _OPC(6, 31), 197 198 /* meta instructions (category -1): */ 199 /* placeholder instr to mark shader inputs: */ 200 OPC_META_INPUT = _OPC(-1, 0), 201 OPC_META_PHI = _OPC(-1, 1), 202 /* The "fan-in" and "fan-out" instructions are used for keeping 203 * track of instructions that write to multiple dst registers 204 * (fan-out) like texture sample instructions, or read multiple 205 * consecutive scalar registers (fan-in) (bary.f, texture samp) 206 */ 207 OPC_META_FO = _OPC(-1, 2), 208 OPC_META_FI = _OPC(-1, 3), 209 210} opc_t; 211 212#define opc_cat(opc) ((int)((opc) >> NOPC_BITS)) 213#define opc_op(opc) ((unsigned)((opc) & ((1 << NOPC_BITS) - 1))) 214 215typedef enum { 216 TYPE_F16 = 0, 217 TYPE_F32 = 1, 218 TYPE_U16 = 2, 219 TYPE_U32 = 3, 220 TYPE_S16 = 4, 221 TYPE_S32 = 5, 222 TYPE_U8 = 6, 223 TYPE_S8 = 7, // XXX I assume? 224} type_t; 225 226static inline uint32_t type_size(type_t type) 227{ 228 switch (type) { 229 case TYPE_F32: 230 case TYPE_U32: 231 case TYPE_S32: 232 return 32; 233 case TYPE_F16: 234 case TYPE_U16: 235 case TYPE_S16: 236 return 16; 237 case TYPE_U8: 238 case TYPE_S8: 239 return 8; 240 default: 241 assert(0); /* invalid type */ 242 return 0; 243 } 244} 245 246static inline int type_float(type_t type) 247{ 248 return (type == TYPE_F32) || (type == TYPE_F16); 249} 250 251static inline int type_uint(type_t type) 252{ 253 return (type == TYPE_U32) || (type == TYPE_U16) || (type == TYPE_U8); 254} 255 256static inline int type_sint(type_t type) 257{ 258 return (type == TYPE_S32) || (type == TYPE_S16) || (type == TYPE_S8); 259} 260 261typedef union PACKED { 262 /* normal gpr or const src register: */ 263 struct PACKED { 264 uint32_t comp : 2; 265 uint32_t num : 10; 266 }; 267 /* for immediate val: */ 268 int32_t iim_val : 11; 269 /* to make compiler happy: */ 270 uint32_t dummy32; 271 uint32_t dummy10 : 10; 272 int32_t idummy10 : 10; 273 uint32_t dummy11 : 11; 274 uint32_t dummy12 : 12; 275 uint32_t dummy13 : 13; 276 uint32_t dummy8 : 8; 277} reg_t; 278 279/* special registers: */ 280#define REG_A0 61 /* address register */ 281#define REG_P0 62 /* predicate register */ 282 283static inline int reg_special(reg_t reg) 284{ 285 return (reg.num == REG_A0) || (reg.num == REG_P0); 286} 287 288typedef struct PACKED { 289 /* dword0: */ 290 union PACKED { 291 struct PACKED { 292 int16_t immed : 16; 293 uint32_t dummy1 : 16; 294 } a3xx; 295 struct PACKED { 296 int32_t immed : 20; 297 uint32_t dummy1 : 12; 298 } a4xx; 299 struct PACKED { 300 uint32_t immed : 32; 301 } a5xx; 302 }; 303 304 /* dword1: */ 305 uint32_t dummy2 : 8; 306 uint32_t repeat : 3; 307 uint32_t dummy3 : 1; 308 uint32_t ss : 1; 309 uint32_t dummy4 : 7; 310 uint32_t inv : 1; 311 uint32_t comp : 2; 312 uint32_t opc : 4; 313 uint32_t jmp_tgt : 1; 314 uint32_t sync : 1; 315 uint32_t opc_cat : 3; 316} instr_cat0_t; 317 318typedef struct PACKED { 319 /* dword0: */ 320 union PACKED { 321 /* for normal src register: */ 322 struct PACKED { 323 uint32_t src : 11; 324 /* at least low bit of pad must be zero or it will 325 * look like a address relative src 326 */ 327 uint32_t pad : 21; 328 }; 329 /* for address relative: */ 330 struct PACKED { 331 int32_t off : 10; 332 uint32_t src_rel_c : 1; 333 uint32_t src_rel : 1; 334 uint32_t unknown : 20; 335 }; 336 /* for immediate: */ 337 int32_t iim_val; 338 uint32_t uim_val; 339 float fim_val; 340 }; 341 342 /* dword1: */ 343 uint32_t dst : 8; 344 uint32_t repeat : 3; 345 uint32_t src_r : 1; 346 uint32_t ss : 1; 347 uint32_t ul : 1; 348 uint32_t dst_type : 3; 349 uint32_t dst_rel : 1; 350 uint32_t src_type : 3; 351 uint32_t src_c : 1; 352 uint32_t src_im : 1; 353 uint32_t even : 1; 354 uint32_t pos_inf : 1; 355 uint32_t must_be_0 : 2; 356 uint32_t jmp_tgt : 1; 357 uint32_t sync : 1; 358 uint32_t opc_cat : 3; 359} instr_cat1_t; 360 361typedef struct PACKED { 362 /* dword0: */ 363 union PACKED { 364 struct PACKED { 365 uint32_t src1 : 11; 366 uint32_t must_be_zero1: 2; 367 uint32_t src1_im : 1; /* immediate */ 368 uint32_t src1_neg : 1; /* negate */ 369 uint32_t src1_abs : 1; /* absolute value */ 370 }; 371 struct PACKED { 372 uint32_t src1 : 10; 373 uint32_t src1_c : 1; /* relative-const */ 374 uint32_t src1_rel : 1; /* relative address */ 375 uint32_t must_be_zero : 1; 376 uint32_t dummy : 3; 377 } rel1; 378 struct PACKED { 379 uint32_t src1 : 12; 380 uint32_t src1_c : 1; /* const */ 381 uint32_t dummy : 3; 382 } c1; 383 }; 384 385 union PACKED { 386 struct PACKED { 387 uint32_t src2 : 11; 388 uint32_t must_be_zero2: 2; 389 uint32_t src2_im : 1; /* immediate */ 390 uint32_t src2_neg : 1; /* negate */ 391 uint32_t src2_abs : 1; /* absolute value */ 392 }; 393 struct PACKED { 394 uint32_t src2 : 10; 395 uint32_t src2_c : 1; /* relative-const */ 396 uint32_t src2_rel : 1; /* relative address */ 397 uint32_t must_be_zero : 1; 398 uint32_t dummy : 3; 399 } rel2; 400 struct PACKED { 401 uint32_t src2 : 12; 402 uint32_t src2_c : 1; /* const */ 403 uint32_t dummy : 3; 404 } c2; 405 }; 406 407 /* dword1: */ 408 uint32_t dst : 8; 409 uint32_t repeat : 3; 410 uint32_t src1_r : 1; 411 uint32_t ss : 1; 412 uint32_t ul : 1; /* dunno */ 413 uint32_t dst_half : 1; /* or widen/narrow.. ie. dst hrN <-> rN */ 414 uint32_t ei : 1; 415 uint32_t cond : 3; 416 uint32_t src2_r : 1; 417 uint32_t full : 1; /* not half */ 418 uint32_t opc : 6; 419 uint32_t jmp_tgt : 1; 420 uint32_t sync : 1; 421 uint32_t opc_cat : 3; 422} instr_cat2_t; 423 424typedef struct PACKED { 425 /* dword0: */ 426 union PACKED { 427 struct PACKED { 428 uint32_t src1 : 11; 429 uint32_t must_be_zero1: 2; 430 uint32_t src2_c : 1; 431 uint32_t src1_neg : 1; 432 uint32_t src2_r : 1; 433 }; 434 struct PACKED { 435 uint32_t src1 : 10; 436 uint32_t src1_c : 1; 437 uint32_t src1_rel : 1; 438 uint32_t must_be_zero : 1; 439 uint32_t dummy : 3; 440 } rel1; 441 struct PACKED { 442 uint32_t src1 : 12; 443 uint32_t src1_c : 1; 444 uint32_t dummy : 3; 445 } c1; 446 }; 447 448 union PACKED { 449 struct PACKED { 450 uint32_t src3 : 11; 451 uint32_t must_be_zero2: 2; 452 uint32_t src3_r : 1; 453 uint32_t src2_neg : 1; 454 uint32_t src3_neg : 1; 455 }; 456 struct PACKED { 457 uint32_t src3 : 10; 458 uint32_t src3_c : 1; 459 uint32_t src3_rel : 1; 460 uint32_t must_be_zero : 1; 461 uint32_t dummy : 3; 462 } rel2; 463 struct PACKED { 464 uint32_t src3 : 12; 465 uint32_t src3_c : 1; 466 uint32_t dummy : 3; 467 } c2; 468 }; 469 470 /* dword1: */ 471 uint32_t dst : 8; 472 uint32_t repeat : 3; 473 uint32_t src1_r : 1; 474 uint32_t ss : 1; 475 uint32_t ul : 1; 476 uint32_t dst_half : 1; /* or widen/narrow.. ie. dst hrN <-> rN */ 477 uint32_t src2 : 8; 478 uint32_t opc : 4; 479 uint32_t jmp_tgt : 1; 480 uint32_t sync : 1; 481 uint32_t opc_cat : 3; 482} instr_cat3_t; 483 484static inline bool instr_cat3_full(instr_cat3_t *cat3) 485{ 486 switch (_OPC(3, cat3->opc)) { 487 case OPC_MAD_F16: 488 case OPC_MAD_U16: 489 case OPC_MAD_S16: 490 case OPC_SEL_B16: 491 case OPC_SEL_S16: 492 case OPC_SEL_F16: 493 case OPC_SAD_S16: 494 case OPC_SAD_S32: // really?? 495 return false; 496 default: 497 return true; 498 } 499} 500 501typedef struct PACKED { 502 /* dword0: */ 503 union PACKED { 504 struct PACKED { 505 uint32_t src : 11; 506 uint32_t must_be_zero1: 2; 507 uint32_t src_im : 1; /* immediate */ 508 uint32_t src_neg : 1; /* negate */ 509 uint32_t src_abs : 1; /* absolute value */ 510 }; 511 struct PACKED { 512 uint32_t src : 10; 513 uint32_t src_c : 1; /* relative-const */ 514 uint32_t src_rel : 1; /* relative address */ 515 uint32_t must_be_zero : 1; 516 uint32_t dummy : 3; 517 } rel; 518 struct PACKED { 519 uint32_t src : 12; 520 uint32_t src_c : 1; /* const */ 521 uint32_t dummy : 3; 522 } c; 523 }; 524 uint32_t dummy1 : 16; /* seem to be ignored */ 525 526 /* dword1: */ 527 uint32_t dst : 8; 528 uint32_t repeat : 3; 529 uint32_t src_r : 1; 530 uint32_t ss : 1; 531 uint32_t ul : 1; 532 uint32_t dst_half : 1; /* or widen/narrow.. ie. dst hrN <-> rN */ 533 uint32_t dummy2 : 5; /* seem to be ignored */ 534 uint32_t full : 1; /* not half */ 535 uint32_t opc : 6; 536 uint32_t jmp_tgt : 1; 537 uint32_t sync : 1; 538 uint32_t opc_cat : 3; 539} instr_cat4_t; 540 541typedef struct PACKED { 542 /* dword0: */ 543 union PACKED { 544 /* normal case: */ 545 struct PACKED { 546 uint32_t full : 1; /* not half */ 547 uint32_t src1 : 8; 548 uint32_t src2 : 8; 549 uint32_t dummy1 : 4; /* seem to be ignored */ 550 uint32_t samp : 4; 551 uint32_t tex : 7; 552 } norm; 553 /* s2en case: */ 554 struct PACKED { 555 uint32_t full : 1; /* not half */ 556 uint32_t src1 : 8; 557 uint32_t src2 : 11; 558 uint32_t dummy1 : 1; 559 uint32_t src3 : 8; 560 uint32_t dummy2 : 3; 561 } s2en; 562 /* same in either case: */ 563 // XXX I think, confirm this 564 struct PACKED { 565 uint32_t full : 1; /* not half */ 566 uint32_t src1 : 8; 567 uint32_t pad : 23; 568 }; 569 }; 570 571 /* dword1: */ 572 uint32_t dst : 8; 573 uint32_t wrmask : 4; /* write-mask */ 574 uint32_t type : 3; 575 uint32_t dummy2 : 1; /* seems to be ignored */ 576 uint32_t is_3d : 1; 577 578 uint32_t is_a : 1; 579 uint32_t is_s : 1; 580 uint32_t is_s2en : 1; 581 uint32_t is_o : 1; 582 uint32_t is_p : 1; 583 584 uint32_t opc : 5; 585 uint32_t jmp_tgt : 1; 586 uint32_t sync : 1; 587 uint32_t opc_cat : 3; 588} instr_cat5_t; 589 590/* dword0 encoding for src_off: [src1 + off], src2: */ 591typedef struct PACKED { 592 /* dword0: */ 593 uint32_t mustbe1 : 1; 594 int32_t off : 13; 595 uint32_t src1 : 8; 596 uint32_t src1_im : 1; 597 uint32_t src2_im : 1; 598 uint32_t src2 : 8; 599 600 /* dword1: */ 601 uint32_t dword1; 602} instr_cat6a_t; 603 604/* dword0 encoding for !src_off: [src1], src2 */ 605typedef struct PACKED { 606 /* dword0: */ 607 uint32_t mustbe0 : 1; 608 uint32_t src1 : 13; 609 uint32_t ignore0 : 8; 610 uint32_t src1_im : 1; 611 uint32_t src2_im : 1; 612 uint32_t src2 : 8; 613 614 /* dword1: */ 615 uint32_t dword1; 616} instr_cat6b_t; 617 618/* dword1 encoding for dst_off: */ 619typedef struct PACKED { 620 /* dword0: */ 621 uint32_t dword0; 622 623 /* note: there is some weird stuff going on where sometimes 624 * cat6->a.off is involved.. but that seems like a bug in 625 * the blob, since it is used even if !cat6->src_off 626 * It would make sense for there to be some more bits to 627 * bring us to 11 bits worth of offset, but not sure.. 628 */ 629 int32_t off : 8; 630 uint32_t mustbe1 : 1; 631 uint32_t dst : 8; 632 uint32_t pad1 : 15; 633} instr_cat6c_t; 634 635/* dword1 encoding for !dst_off: */ 636typedef struct PACKED { 637 /* dword0: */ 638 uint32_t dword0; 639 640 uint32_t dst : 8; 641 uint32_t mustbe0 : 1; 642 uint32_t pad0 : 23; 643} instr_cat6d_t; 644 645/* I think some of the other cat6 instructions use additional 646 * sub-encodings.. 647 */ 648 649typedef union PACKED { 650 instr_cat6a_t a; 651 instr_cat6b_t b; 652 instr_cat6c_t c; 653 instr_cat6d_t d; 654 struct PACKED { 655 /* dword0: */ 656 uint32_t src_off : 1; 657 uint32_t pad1 : 31; 658 659 /* dword1: */ 660 uint32_t pad2 : 8; 661 uint32_t dst_off : 1; 662 uint32_t pad3 : 8; 663 uint32_t type : 3; 664 uint32_t g : 1; /* or in some cases it means dst immed */ 665 uint32_t pad4 : 1; 666 uint32_t opc : 5; 667 uint32_t jmp_tgt : 1; 668 uint32_t sync : 1; 669 uint32_t opc_cat : 3; 670 }; 671} instr_cat6_t; 672 673typedef union PACKED { 674 instr_cat0_t cat0; 675 instr_cat1_t cat1; 676 instr_cat2_t cat2; 677 instr_cat3_t cat3; 678 instr_cat4_t cat4; 679 instr_cat5_t cat5; 680 instr_cat6_t cat6; 681 struct PACKED { 682 /* dword0: */ 683 uint64_t pad1 : 40; 684 uint32_t repeat : 3; /* cat0-cat4 */ 685 uint32_t pad2 : 1; 686 uint32_t ss : 1; /* cat1-cat4 (cat0??) */ 687 uint32_t ul : 1; /* cat2-cat4 (and cat1 in blob.. which may be bug??) */ 688 uint32_t pad3 : 13; 689 uint32_t jmp_tgt : 1; 690 uint32_t sync : 1; 691 uint32_t opc_cat : 3; 692 693 }; 694} instr_t; 695 696static inline uint32_t instr_opc(instr_t *instr) 697{ 698 switch (instr->opc_cat) { 699 case 0: return instr->cat0.opc; 700 case 1: return 0; 701 case 2: return instr->cat2.opc; 702 case 3: return instr->cat3.opc; 703 case 4: return instr->cat4.opc; 704 case 5: return instr->cat5.opc; 705 case 6: return instr->cat6.opc; 706 default: return 0; 707 } 708} 709 710static inline bool is_mad(opc_t opc) 711{ 712 switch (opc) { 713 case OPC_MAD_U16: 714 case OPC_MAD_S16: 715 case OPC_MAD_U24: 716 case OPC_MAD_S24: 717 case OPC_MAD_F16: 718 case OPC_MAD_F32: 719 return true; 720 default: 721 return false; 722 } 723} 724 725static inline bool is_madsh(opc_t opc) 726{ 727 switch (opc) { 728 case OPC_MADSH_U16: 729 case OPC_MADSH_M16: 730 return true; 731 default: 732 return false; 733 } 734} 735 736#endif /* INSTR_A3XX_H_ */ 737