1/* 2 * Copyright (C) 2020 Collabora, Ltd. 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#include "compiler.h" 25 26/* This file contains the final passes of the compiler. Running after 27 * scheduling and RA, the IR is now finalized, so we need to emit it to actual 28 * bits on the wire (as well as fixup branches) */ 29 30static uint64_t 31bi_pack_header(bi_clause *clause, bi_clause *next_1, bi_clause *next_2) 32{ 33 /* next_dependencies are the union of the dependencies of successors' 34 * dependencies */ 35 36 unsigned dependency_wait = next_1 ? next_1->dependencies : 0; 37 dependency_wait |= next_2 ? next_2->dependencies : 0; 38 39 bool staging_barrier = next_1 ? next_1->staging_barrier : false; 40 staging_barrier |= next_2 ? next_2->staging_barrier : 0; 41 42 struct bifrost_header header = { 43 .flow_control = 44 (next_1 == NULL && next_2 == NULL) ? 45 BIFROST_FLOW_END : clause->flow_control, 46 .terminate_discarded_threads = clause->td, 47 .next_clause_prefetch = clause->next_clause_prefetch && next_1, 48 .staging_barrier = staging_barrier, 49 .staging_register = clause->staging_register, 50 .dependency_wait = dependency_wait, 51 .dependency_slot = clause->scoreboard_id, 52 .message_type = clause->message_type, 53 .next_message_type = next_1 ? next_1->message_type : 0, 54 }; 55 56 uint64_t u = 0; 57 memcpy(&u, &header, sizeof(header)); 58 return u; 59} 60 61/* Assigns a slot for reading, before anything is written */ 62 63static void 64bi_assign_slot_read(bi_registers *regs, bi_index src) 65{ 66 /* We only assign for registers */ 67 if (src.type != BI_INDEX_REGISTER) 68 return; 69 70 /* Check if we already assigned the slot */ 71 for (unsigned i = 0; i <= 1; ++i) { 72 if (regs->slot[i] == src.value && regs->enabled[i]) 73 return; 74 } 75 76 if (regs->slot[2] == src.value && regs->slot23.slot2 == BIFROST_OP_READ) 77 return; 78 79 /* Assign it now */ 80 81 for (unsigned i = 0; i <= 1; ++i) { 82 if (!regs->enabled[i]) { 83 regs->slot[i] = src.value; 84 regs->enabled[i] = true; 85 return; 86 } 87 } 88 89 if (!regs->slot23.slot3) { 90 regs->slot[2] = src.value; 91 regs->slot23.slot2 = BIFROST_OP_READ; 92 return; 93 } 94 95 bi_print_slots(regs, stderr); 96 unreachable("Failed to find a free slot for src"); 97} 98 99static bi_registers 100bi_assign_slots(bi_tuple *now, bi_tuple *prev) 101{ 102 /* We assign slots for the main register mechanism. Special ops 103 * use the data registers, which has its own mechanism entirely 104 * and thus gets skipped over here. */ 105 106 bool read_dreg = now->add && bi_opcode_props[now->add->op].sr_read; 107 bool write_dreg = prev->add && bi_opcode_props[prev->add->op].sr_write; 108 109 /* First, assign reads */ 110 111 if (now->fma) 112 bi_foreach_src(now->fma, src) 113 bi_assign_slot_read(&now->regs, (now->fma)->src[src]); 114 115 if (now->add) { 116 bi_foreach_src(now->add, src) { 117 if (!(src == 0 && read_dreg)) 118 bi_assign_slot_read(&now->regs, (now->add)->src[src]); 119 } 120 } 121 122 /* Next, assign writes. Staging writes are assigned separately, but 123 * +ATEST wants its destination written to both a staging register 124 * _and_ a regular write, because it may not generate a message */ 125 126 if (prev->add && (!write_dreg || prev->add->op == BI_OPCODE_ATEST)) { 127 bi_index idx = prev->add->dest[0]; 128 129 if (idx.type == BI_INDEX_REGISTER) { 130 now->regs.slot[3] = idx.value; 131 now->regs.slot23.slot3 = BIFROST_OP_WRITE; 132 } 133 } 134 135 if (prev->fma) { 136 bi_index idx = (prev->fma)->dest[0]; 137 138 if (idx.type == BI_INDEX_REGISTER) { 139 if (now->regs.slot23.slot3) { 140 /* Scheduler constraint: cannot read 3 and write 2 */ 141 assert(!now->regs.slot23.slot2); 142 now->regs.slot[2] = idx.value; 143 now->regs.slot23.slot2 = BIFROST_OP_WRITE; 144 } else { 145 now->regs.slot[3] = idx.value; 146 now->regs.slot23.slot3 = BIFROST_OP_WRITE; 147 now->regs.slot23.slot3_fma = true; 148 } 149 } 150 } 151 152 return now->regs; 153} 154 155static enum bifrost_reg_mode 156bi_pack_register_mode(bi_registers r) 157{ 158 /* Handle idle as a special case */ 159 if (!(r.slot23.slot2 | r.slot23.slot3)) 160 return r.first_instruction ? BIFROST_IDLE_1 : BIFROST_IDLE; 161 162 /* Otherwise, use the LUT */ 163 for (unsigned i = 0; i < ARRAY_SIZE(bifrost_reg_ctrl_lut); ++i) { 164 if (memcmp(bifrost_reg_ctrl_lut + i, &r.slot23, sizeof(r.slot23)) == 0) 165 return i; 166 } 167 168 bi_print_slots(&r, stderr); 169 unreachable("Invalid slot assignment"); 170} 171 172static uint64_t 173bi_pack_registers(bi_registers regs) 174{ 175 enum bifrost_reg_mode mode = bi_pack_register_mode(regs); 176 struct bifrost_regs s = { 0 }; 177 uint64_t packed = 0; 178 179 /* Need to pack 5-bit mode as a 4-bit field. The decoder moves bit 3 to bit 4 for 180 * first instruction and adds 16 when reg 2 == reg 3 */ 181 182 unsigned ctrl; 183 bool r2_equals_r3 = false; 184 185 if (regs.first_instruction) { 186 /* Bit 3 implicitly must be clear for first instructions. 187 * The affected patterns all write both ADD/FMA, but that 188 * is forbidden for the last instruction (whose writes are 189 * encoded by the first), so this does not add additional 190 * encoding constraints */ 191 assert(!(mode & 0x8)); 192 193 /* Move bit 4 to bit 3, since bit 3 is clear */ 194 ctrl = (mode & 0x7) | ((mode & 0x10) >> 1); 195 196 /* If we can let r2 equal r3, we have to or the hardware raises 197 * INSTR_INVALID_ENC (it's unclear why). */ 198 if (!(regs.slot23.slot2 && regs.slot23.slot3)) 199 r2_equals_r3 = true; 200 } else { 201 /* We force r2=r3 or not for the upper bit */ 202 ctrl = (mode & 0xF); 203 r2_equals_r3 = (mode & 0x10); 204 } 205 206 if (regs.enabled[1]) { 207 /* Gotta save that bit!~ Required by the 63-x trick */ 208 assert(regs.slot[1] > regs.slot[0]); 209 assert(regs.enabled[0]); 210 211 /* Do the 63-x trick, see docs/disasm */ 212 if (regs.slot[0] > 31) { 213 regs.slot[0] = 63 - regs.slot[0]; 214 regs.slot[1] = 63 - regs.slot[1]; 215 } 216 217 assert(regs.slot[0] <= 31); 218 assert(regs.slot[1] <= 63); 219 220 s.ctrl = ctrl; 221 s.reg1 = regs.slot[1]; 222 s.reg0 = regs.slot[0]; 223 } else { 224 /* slot 1 disabled, so set to zero and use slot 1 for ctrl */ 225 s.ctrl = 0; 226 s.reg1 = ctrl << 2; 227 228 if (regs.enabled[0]) { 229 /* Bit 0 upper bit of slot 0 */ 230 s.reg1 |= (regs.slot[0] >> 5); 231 232 /* Rest of slot 0 in usual spot */ 233 s.reg0 = (regs.slot[0] & 0b11111); 234 } else { 235 /* Bit 1 set if slot 0 also disabled */ 236 s.reg1 |= (1 << 1); 237 } 238 } 239 240 /* Force r2 =/!= r3 as needed */ 241 if (r2_equals_r3) { 242 assert(regs.slot[3] == regs.slot[2] || !(regs.slot23.slot2 && regs.slot23.slot3)); 243 244 if (regs.slot23.slot2) 245 regs.slot[3] = regs.slot[2]; 246 else 247 regs.slot[2] = regs.slot[3]; 248 } else if (!regs.first_instruction) { 249 /* Enforced by the encoding anyway */ 250 assert(regs.slot[2] != regs.slot[3]); 251 } 252 253 s.reg2 = regs.slot[2]; 254 s.reg3 = regs.slot[3]; 255 s.fau_idx = regs.fau_idx; 256 257 memcpy(&packed, &s, sizeof(s)); 258 return packed; 259} 260 261/* We must ensure slot 1 > slot 0 for the 63-x trick to function, so we fix 262 * this up at pack time. (Scheduling doesn't care.) */ 263 264static void 265bi_flip_slots(bi_registers *regs) 266{ 267 if (regs->enabled[0] && regs->enabled[1] && regs->slot[1] < regs->slot[0]) { 268 unsigned temp = regs->slot[0]; 269 regs->slot[0] = regs->slot[1]; 270 regs->slot[1] = temp; 271 } 272 273} 274 275static inline enum bifrost_packed_src 276bi_get_src_slot(bi_registers *regs, unsigned reg) 277{ 278 if (regs->slot[0] == reg && regs->enabled[0]) 279 return BIFROST_SRC_PORT0; 280 else if (regs->slot[1] == reg && regs->enabled[1]) 281 return BIFROST_SRC_PORT1; 282 else if (regs->slot[2] == reg && regs->slot23.slot2 == BIFROST_OP_READ) 283 return BIFROST_SRC_PORT2; 284 else 285 unreachable("Tried to access register with no port"); 286} 287 288static inline enum bifrost_packed_src 289bi_get_src_new(bi_instr *ins, bi_registers *regs, unsigned s) 290{ 291 if (!ins) 292 return 0; 293 294 bi_index src = ins->src[s]; 295 296 if (src.type == BI_INDEX_REGISTER) 297 return bi_get_src_slot(regs, src.value); 298 else if (src.type == BI_INDEX_PASS) 299 return src.value; 300 else if (bi_is_null(src) && ins->op == BI_OPCODE_ZS_EMIT && s < 2) 301 return BIFROST_SRC_STAGE; 302 else { 303 /* TODO make safer */ 304 return BIFROST_SRC_STAGE; 305 } 306} 307 308static struct bi_packed_tuple 309bi_pack_tuple(bi_clause *clause, bi_tuple *tuple, bi_tuple *prev, bool first_tuple, gl_shader_stage stage) 310{ 311 bi_assign_slots(tuple, prev); 312 tuple->regs.fau_idx = tuple->fau_idx; 313 tuple->regs.first_instruction = first_tuple; 314 315 bi_flip_slots(&tuple->regs); 316 317 bool sr_read = tuple->add && 318 bi_opcode_props[(tuple->add)->op].sr_read; 319 320 uint64_t reg = bi_pack_registers(tuple->regs); 321 uint64_t fma = bi_pack_fma(tuple->fma, 322 bi_get_src_new(tuple->fma, &tuple->regs, 0), 323 bi_get_src_new(tuple->fma, &tuple->regs, 1), 324 bi_get_src_new(tuple->fma, &tuple->regs, 2), 325 bi_get_src_new(tuple->fma, &tuple->regs, 3)); 326 327 uint64_t add = bi_pack_add(tuple->add, 328 bi_get_src_new(tuple->add, &tuple->regs, sr_read + 0), 329 bi_get_src_new(tuple->add, &tuple->regs, sr_read + 1), 330 bi_get_src_new(tuple->add, &tuple->regs, sr_read + 2), 331 0); 332 333 if (tuple->add) { 334 bi_instr *add = tuple->add; 335 336 bool sr_write = bi_opcode_props[add->op].sr_write && 337 !bi_is_null(add->dest[0]); 338 339 if (sr_read && !bi_is_null(add->src[0])) { 340 assert(add->src[0].type == BI_INDEX_REGISTER); 341 clause->staging_register = add->src[0].value; 342 343 if (sr_write) 344 assert(bi_is_equiv(add->src[0], add->dest[0])); 345 } else if (sr_write) { 346 assert(add->dest[0].type == BI_INDEX_REGISTER); 347 clause->staging_register = add->dest[0].value; 348 } 349 } 350 351 struct bi_packed_tuple packed = { 352 .lo = reg | (fma << 35) | ((add & 0b111111) << 58), 353 .hi = add >> 6 354 }; 355 356 return packed; 357} 358 359/* A block contains at most one PC-relative constant, from a terminal branch. 360 * Find the last instruction and if it is a relative branch, fix up the 361 * PC-relative constant to contain the absolute offset. This occurs at pack 362 * time instead of schedule time because the number of quadwords between each 363 * block is not known until after all other passes have finished. 364 */ 365 366static void 367bi_assign_branch_offset(bi_context *ctx, bi_block *block) 368{ 369 if (list_is_empty(&block->clauses)) 370 return; 371 372 bi_clause *clause = list_last_entry(&block->clauses, bi_clause, link); 373 bi_instr *br = bi_last_instr_in_clause(clause); 374 375 if (!br->branch_target) 376 return; 377 378 /* Put it in the high place */ 379 int32_t qwords = bi_block_offset(ctx, clause, br->branch_target); 380 int32_t bytes = qwords * 16; 381 382 /* Copy so we can toy with the sign without undefined behaviour */ 383 uint32_t raw = 0; 384 memcpy(&raw, &bytes, sizeof(raw)); 385 386 /* Clear off top bits for A1/B1 bits */ 387 raw &= ~0xF0000000; 388 389 /* Put in top 32-bits */ 390 assert(clause->pcrel_idx < 8); 391 clause->constants[clause->pcrel_idx] |= ((uint64_t) raw) << 32ull; 392} 393 394static void 395bi_pack_constants(unsigned tuple_count, uint64_t *constants, 396 unsigned word_idx, unsigned constant_words, bool ec0_packed, 397 struct util_dynarray *emission) 398{ 399 unsigned index = (word_idx << 1) + ec0_packed; 400 401 /* Do more constants follow */ 402 bool more = (word_idx + 1) < constant_words; 403 404 /* Indexed first by tuple count and second by constant word number, 405 * indicates the position in the clause */ 406 unsigned pos_lookup[8][3] = { 407 { 0 }, 408 { 1 }, 409 { 3 }, 410 { 2, 5 }, 411 { 4, 8 }, 412 { 7, 11, 14 }, 413 { 6, 10, 13 }, 414 { 9, 12 } 415 }; 416 417 /* Compute the pos, and check everything is reasonable */ 418 assert((tuple_count - 1) < 8); 419 assert(word_idx < 3); 420 unsigned pos = pos_lookup[tuple_count - 1][word_idx]; 421 assert(pos != 0 || (tuple_count == 1 && word_idx == 0)); 422 423 struct bifrost_fmt_constant quad = { 424 .pos = pos, 425 .tag = more ? BIFROST_FMTC_CONSTANTS : BIFROST_FMTC_FINAL, 426 .imm_1 = constants[index + 0] >> 4, 427 .imm_2 = constants[index + 1] >> 4, 428 }; 429 430 util_dynarray_append(emission, struct bifrost_fmt_constant, quad); 431} 432 433uint8_t 434bi_pack_literal(enum bi_clause_subword literal) 435{ 436 assert(literal >= BI_CLAUSE_SUBWORD_LITERAL_0); 437 assert(literal <= BI_CLAUSE_SUBWORD_LITERAL_7); 438 439 return (literal - BI_CLAUSE_SUBWORD_LITERAL_0); 440} 441 442static inline uint8_t 443bi_clause_upper(unsigned val, 444 struct bi_packed_tuple *tuples, 445 ASSERTED unsigned tuple_count) 446{ 447 assert(val < tuple_count); 448 449 /* top 3-bits of 78-bits is tuple >> 75 == (tuple >> 64) >> 11 */ 450 struct bi_packed_tuple tuple = tuples[val]; 451 return (tuple.hi >> 11); 452} 453 454uint8_t 455bi_pack_upper(enum bi_clause_subword upper, 456 struct bi_packed_tuple *tuples, 457 ASSERTED unsigned tuple_count) 458{ 459 assert(upper >= BI_CLAUSE_SUBWORD_UPPER_0); 460 assert(upper <= BI_CLAUSE_SUBWORD_UPPER_7); 461 462 return bi_clause_upper(upper - BI_CLAUSE_SUBWORD_UPPER_0, tuples, 463 tuple_count); 464} 465 466uint64_t 467bi_pack_tuple_bits(enum bi_clause_subword idx, 468 struct bi_packed_tuple *tuples, 469 ASSERTED unsigned tuple_count, 470 unsigned offset, unsigned nbits) 471{ 472 assert(idx >= BI_CLAUSE_SUBWORD_TUPLE_0); 473 assert(idx <= BI_CLAUSE_SUBWORD_TUPLE_7); 474 475 unsigned val = (idx - BI_CLAUSE_SUBWORD_TUPLE_0); 476 assert(val < tuple_count); 477 478 struct bi_packed_tuple tuple = tuples[val]; 479 480 assert(offset + nbits < 78); 481 assert(nbits <= 64); 482 483 /* (X >> start) & m 484 * = (((hi << 64) | lo) >> start) & m 485 * = (((hi << 64) >> start) | (lo >> start)) & m 486 * = { ((hi << (64 - start)) | (lo >> start)) & m if start <= 64 487 * { ((hi >> (start - 64)) | (lo >> start)) & m if start >= 64 488 * = { ((hi << (64 - start)) & m) | ((lo >> start) & m) if start <= 64 489 * { ((hi >> (start - 64)) & m) | ((lo >> start) & m) if start >= 64 490 * 491 * By setting m = 2^64 - 1, we justify doing the respective shifts as 492 * 64-bit integers. Zero special cased to avoid undefined behaviour. 493 */ 494 495 uint64_t lo = (tuple.lo >> offset); 496 uint64_t hi = (offset == 0) ? 0 497 : (offset > 64) ? (tuple.hi >> (offset - 64)) 498 : (tuple.hi << (64 - offset)); 499 500 return (lo | hi) & ((1ULL << nbits) - 1); 501} 502 503static inline uint16_t 504bi_pack_lu(enum bi_clause_subword word, 505 struct bi_packed_tuple *tuples, 506 ASSERTED unsigned tuple_count) 507{ 508 return (word >= BI_CLAUSE_SUBWORD_UPPER_0) ? 509 bi_pack_upper(word, tuples, tuple_count) : 510 bi_pack_literal(word); 511} 512 513uint8_t 514bi_pack_sync(enum bi_clause_subword t1, 515 enum bi_clause_subword t2, 516 enum bi_clause_subword t3, 517 struct bi_packed_tuple *tuples, 518 ASSERTED unsigned tuple_count, 519 bool z) 520{ 521 uint8_t sync = 522 (bi_pack_lu(t3, tuples, tuple_count) << 0) | 523 (bi_pack_lu(t2, tuples, tuple_count) << 3); 524 525 if (t1 == BI_CLAUSE_SUBWORD_Z) 526 sync |= z << 6; 527 else 528 sync |= bi_pack_literal(t1) << 6; 529 530 return sync; 531} 532 533static inline uint64_t 534bi_pack_t_ec(enum bi_clause_subword word, 535 struct bi_packed_tuple *tuples, 536 ASSERTED unsigned tuple_count, 537 uint64_t ec0) 538{ 539 if (word == BI_CLAUSE_SUBWORD_CONSTANT) 540 return ec0; 541 else 542 return bi_pack_tuple_bits(word, tuples, tuple_count, 0, 60); 543} 544 545static uint32_t 546bi_pack_subwords_56(enum bi_clause_subword t, 547 struct bi_packed_tuple *tuples, 548 ASSERTED unsigned tuple_count, 549 uint64_t header, uint64_t ec0, 550 unsigned tuple_subword) 551{ 552 switch (t) { 553 case BI_CLAUSE_SUBWORD_HEADER: 554 return (header & ((1 << 30) - 1)); 555 case BI_CLAUSE_SUBWORD_RESERVED: 556 return 0; 557 case BI_CLAUSE_SUBWORD_CONSTANT: 558 return (ec0 >> 15) & ((1 << 30) - 1); 559 default: 560 return bi_pack_tuple_bits(t, tuples, tuple_count, tuple_subword * 15, 30); 561 } 562} 563 564static uint16_t 565bi_pack_subword(enum bi_clause_subword t, unsigned format, 566 struct bi_packed_tuple *tuples, 567 ASSERTED unsigned tuple_count, 568 uint64_t header, uint64_t ec0, unsigned m0, 569 unsigned tuple_subword) 570{ 571 switch (t) { 572 case BI_CLAUSE_SUBWORD_HEADER: 573 return header >> 30; 574 case BI_CLAUSE_SUBWORD_M: 575 return m0; 576 case BI_CLAUSE_SUBWORD_CONSTANT: 577 return (format == 5 || format == 10) ? 578 (ec0 & ((1 << 15) - 1)) : 579 (ec0 >> (15 + 30)); 580 case BI_CLAUSE_SUBWORD_UPPER_23: 581 return (bi_clause_upper(2, tuples, tuple_count) << 12) | 582 (bi_clause_upper(3, tuples, tuple_count) << 9); 583 case BI_CLAUSE_SUBWORD_UPPER_56: 584 return (bi_clause_upper(5, tuples, tuple_count) << 12) | 585 (bi_clause_upper(6, tuples, tuple_count) << 9); 586 case BI_CLAUSE_SUBWORD_UPPER_0 ... BI_CLAUSE_SUBWORD_UPPER_7: 587 return bi_pack_upper(t, tuples, tuple_count) << 12; 588 default: 589 return bi_pack_tuple_bits(t, tuples, tuple_count, tuple_subword * 15, 15); 590 } 591} 592 593/* EC0 is 60-bits (bottom 4 already shifted off) */ 594void 595bi_pack_format(struct util_dynarray *emission, 596 unsigned index, 597 struct bi_packed_tuple *tuples, 598 ASSERTED unsigned tuple_count, 599 uint64_t header, uint64_t ec0, 600 unsigned m0, bool z) 601{ 602 struct bi_clause_format format = bi_clause_formats[index]; 603 604 uint8_t sync = bi_pack_sync(format.tag_1, format.tag_2, format.tag_3, 605 tuples, tuple_count, z); 606 607 uint64_t s0_s3 = bi_pack_t_ec(format.s0_s3, tuples, tuple_count, ec0); 608 609 uint16_t s4 = bi_pack_subword(format.s4, format.format, tuples, tuple_count, header, ec0, m0, 4); 610 611 uint32_t s5_s6 = bi_pack_subwords_56(format.s5_s6, 612 tuples, tuple_count, header, ec0, 613 (format.format == 2 || format.format == 7) ? 0 : 3); 614 615 uint64_t s7 = bi_pack_subword(format.s7, format.format, tuples, tuple_count, header, ec0, m0, 2); 616 617 /* Now that subwords are packed, split into 64-bit halves and emit */ 618 uint64_t lo = sync | ((s0_s3 & ((1ull << 56) - 1)) << 8); 619 uint64_t hi = (s0_s3 >> 56) | ((uint64_t) s4 << 4) | ((uint64_t) s5_s6 << 19) | ((uint64_t) s7 << 49); 620 621 util_dynarray_append(emission, uint64_t, lo); 622 util_dynarray_append(emission, uint64_t, hi); 623} 624 625static void 626bi_pack_clause(bi_context *ctx, bi_clause *clause, 627 bi_clause *next_1, bi_clause *next_2, 628 struct util_dynarray *emission, gl_shader_stage stage) 629{ 630 struct bi_packed_tuple ins[8] = { 0 }; 631 632 for (unsigned i = 0; i < clause->tuple_count; ++i) { 633 unsigned prev = ((i == 0) ? clause->tuple_count : i) - 1; 634 ins[i] = bi_pack_tuple(clause, &clause->tuples[i], 635 &clause->tuples[prev], i == 0, stage); 636 } 637 638 bool ec0_packed = bi_ec0_packed(clause->tuple_count); 639 640 if (ec0_packed) 641 clause->constant_count = MAX2(clause->constant_count, 1); 642 643 unsigned constant_quads = 644 DIV_ROUND_UP(clause->constant_count - (ec0_packed ? 1 : 0), 2); 645 646 uint64_t header = bi_pack_header(clause, next_1, next_2); 647 uint64_t ec0 = (clause->constants[0] >> 4); 648 unsigned m0 = (clause->pcrel_idx == 0) ? 4 : 0; 649 650 unsigned counts[8] = { 651 1, 2, 3, 3, 4, 5, 5, 6 652 }; 653 654 unsigned indices[8][6] = { 655 { 1 }, 656 { 0, 2 }, 657 { 0, 3, 4 }, 658 { 0, 3, 6 }, 659 { 0, 3, 7, 8 }, 660 { 0, 3, 5, 9, 10 }, 661 { 0, 3, 5, 9, 11 }, 662 { 0, 3, 5, 9, 12, 13 }, 663 }; 664 665 unsigned count = counts[clause->tuple_count - 1]; 666 667 for (unsigned pos = 0; pos < count; ++pos) { 668 ASSERTED unsigned idx = indices[clause->tuple_count - 1][pos]; 669 assert(bi_clause_formats[idx].pos == pos); 670 assert((bi_clause_formats[idx].tag_1 == BI_CLAUSE_SUBWORD_Z) == 671 (pos == count - 1)); 672 673 /* Whether to end the clause immediately after the last tuple */ 674 bool z = (constant_quads == 0); 675 676 bi_pack_format(emission, indices[clause->tuple_count - 1][pos], 677 ins, clause->tuple_count, header, ec0, m0, 678 z); 679 } 680 681 /* Pack the remaining constants */ 682 683 for (unsigned pos = 0; pos < constant_quads; ++pos) { 684 bi_pack_constants(clause->tuple_count, clause->constants, 685 pos, constant_quads, ec0_packed, emission); 686 } 687} 688 689static void 690bi_collect_blend_ret_addr(bi_context *ctx, struct util_dynarray *emission, 691 const bi_clause *clause) 692{ 693 /* No need to collect return addresses when we're in a blend shader. */ 694 if (ctx->inputs->is_blend) 695 return; 696 697 const bi_tuple *tuple = &clause->tuples[clause->tuple_count - 1]; 698 const bi_instr *ins = tuple->add; 699 700 if (!ins || ins->op != BI_OPCODE_BLEND) 701 return; 702 703 704 unsigned loc = tuple->regs.fau_idx - BIR_FAU_BLEND_0; 705 assert(loc < ARRAY_SIZE(ctx->info->bifrost.blend)); 706 assert(!ctx->info->bifrost.blend[loc].return_offset); 707 ctx->info->bifrost.blend[loc].return_offset = 708 util_dynarray_num_elements(emission, uint8_t); 709 assert(!(ctx->info->bifrost.blend[loc].return_offset & 0x7)); 710} 711 712unsigned 713bi_pack(bi_context *ctx, struct util_dynarray *emission) 714{ 715 unsigned previous_size = emission->size; 716 717 bi_foreach_block(ctx, block) { 718 bi_assign_branch_offset(ctx, block); 719 720 bi_foreach_clause_in_block(block, clause) { 721 bool is_last = (clause->link.next == &block->clauses); 722 723 /* Get the succeeding clauses, either two successors of 724 * the block for the last clause in the block or just 725 * the next clause within the block */ 726 727 bi_clause *next = NULL, *next_2 = NULL; 728 729 if (is_last) { 730 next = bi_next_clause(ctx, block->successors[0], NULL); 731 next_2 = bi_next_clause(ctx, block->successors[1], NULL); 732 } else { 733 next = bi_next_clause(ctx, block, clause); 734 } 735 736 737 previous_size = emission->size; 738 739 bi_pack_clause(ctx, clause, next, next_2, emission, ctx->stage); 740 741 if (!is_last) 742 bi_collect_blend_ret_addr(ctx, emission, clause); 743 } 744 } 745 746 return emission->size - previous_size; 747} 748