src/broadcom/compiler/vir.c

   1 /*
   2  * Copyright © 2016-2017 Broadcom
   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
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "broadcom/common/v3d_device_info.h"
  25 #include "v3d_compiler.h"
  26
  27 int
  28 vir_get_non_sideband_nsrc(struct qinst *inst)
  29 {
  30         switch (inst->qpu.type) {
  31         case V3D_QPU_INSTR_TYPE_BRANCH:
  32                 return 0;
  33         case V3D_QPU_INSTR_TYPE_ALU:
  34                 if (inst->qpu.alu.add.op != V3D_QPU_A_NOP)
  35                         return v3d_qpu_add_op_num_src(inst->qpu.alu.add.op);
  36                 else
  37                         return v3d_qpu_mul_op_num_src(inst->qpu.alu.mul.op);
  38         }
  39
  40         return 0;
  41 }
  42
  43 int
  44 vir_get_nsrc(struct qinst *inst)
  45 {
  46         int nsrc = vir_get_non_sideband_nsrc(inst);
  47
  48         if (vir_has_implicit_uniform(inst))
  49                 nsrc++;
  50
  51         return nsrc;
  52 }
  53
  54 bool
  55 vir_has_implicit_uniform(struct qinst *inst)
  56 {
  57         switch (inst->qpu.type) {
  58         case V3D_QPU_INSTR_TYPE_BRANCH:
  59                 return true;
  60         case V3D_QPU_INSTR_TYPE_ALU:
  61                 switch (inst->dst.file) {
  62                 case QFILE_TLBU:
  63                         return true;
  64                 default:
  65                         return inst->has_implicit_uniform;
  66                 }
  67         }
  68         return false;
  69 }
  70
  71 /* The sideband uniform for textures gets stored after the normal ALU
  72  * arguments.
  73  */
  74 int
  75 vir_get_implicit_uniform_src(struct qinst *inst)
  76 {
  77         return vir_get_nsrc(inst) - 1;
  78 }
  79
  80 /**
  81  * Returns whether the instruction has any side effects that must be
  82  * preserved.
  83  */
  84 bool
  85 vir_has_side_effects(struct v3d_compile *c, struct qinst *inst)
  86 {
  87         switch (inst->qpu.type) {
  88         case V3D_QPU_INSTR_TYPE_BRANCH:
  89                 return true;
  90         case V3D_QPU_INSTR_TYPE_ALU:
  91                 switch (inst->qpu.alu.add.op) {
  92                 case V3D_QPU_A_SETREVF:
  93                 case V3D_QPU_A_SETMSF:
  94                 case V3D_QPU_A_VPMSETUP:
  95                 case V3D_QPU_A_STVPMV:
  96                 case V3D_QPU_A_STVPMD:
  97                 case V3D_QPU_A_STVPMP:
  98                 case V3D_QPU_A_VPMWT:
  99                         return true;
 100                 default:
 101                         break;
 102                 }
 103
 104                 switch (inst->qpu.alu.mul.op) {
 105                 case V3D_QPU_M_MULTOP:
 106                         return true;
 107                 default:
 108                         break;
 109                 }
 110         }
 111
 112         if (inst->qpu.sig.ldtmu ||
 113             inst->qpu.sig.ldvary ||
 114             inst->qpu.sig.wrtmuc ||
 115             inst->qpu.sig.thrsw) {
 116                 return true;
 117         }
 118
 119         return false;
 120 }
 121
 122 bool
 123 vir_is_float_input(struct qinst *inst)
 124 {
 125         /* XXX: More instrs */
 126         switch (inst->qpu.type) {
 127         case V3D_QPU_INSTR_TYPE_BRANCH:
 128                 return false;
 129         case V3D_QPU_INSTR_TYPE_ALU:
 130                 switch (inst->qpu.alu.add.op) {
 131                 case V3D_QPU_A_FADD:
 132                 case V3D_QPU_A_FSUB:
 133                 case V3D_QPU_A_FMIN:
 134                 case V3D_QPU_A_FMAX:
 135                 case V3D_QPU_A_FTOIN:
 136                         return true;
 137                 default:
 138                         break;
 139                 }
 140
 141                 switch (inst->qpu.alu.mul.op) {
 142                 case V3D_QPU_M_FMOV:
 143                 case V3D_QPU_M_VFMUL:
 144                 case V3D_QPU_M_FMUL:
 145                         return true;
 146                 default:
 147                         break;
 148                 }
 149         }
 150
 151         return false;
 152 }
 153
 154 bool
 155 vir_is_raw_mov(struct qinst *inst)
 156 {
 157         if (inst->qpu.type != V3D_QPU_INSTR_TYPE_ALU ||
 158             (inst->qpu.alu.mul.op != V3D_QPU_M_FMOV &&
 159              inst->qpu.alu.mul.op != V3D_QPU_M_MOV)) {
 160                 return false;
 161         }
 162
 163         if (inst->qpu.alu.add.output_pack != V3D_QPU_PACK_NONE ||
 164             inst->qpu.alu.mul.output_pack != V3D_QPU_PACK_NONE) {
 165                 return false;
 166         }
 167
 168         if (inst->qpu.flags.ac != V3D_QPU_COND_NONE ||
 169             inst->qpu.flags.mc != V3D_QPU_COND_NONE)
 170                 return false;
 171
 172         return true;
 173 }
 174
 175 bool
 176 vir_is_add(struct qinst *inst)
 177 {
 178         return (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
 179                 inst->qpu.alu.add.op != V3D_QPU_A_NOP);
 180 }
 181
 182 bool
 183 vir_is_mul(struct qinst *inst)
 184 {
 185         return (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
 186                 inst->qpu.alu.mul.op != V3D_QPU_M_NOP);
 187 }
 188
 189 bool
 190 vir_is_tex(struct qinst *inst)
 191 {
 192         if (inst->dst.file == QFILE_MAGIC)
 193                 return v3d_qpu_magic_waddr_is_tmu(inst->dst.index);
 194
 195         return false;
 196 }
 197
 198 bool
 199 vir_depends_on_flags(struct qinst *inst)
 200 {
 201         if (inst->qpu.type == V3D_QPU_INSTR_TYPE_BRANCH) {
 202                 return (inst->qpu.branch.cond != V3D_QPU_BRANCH_COND_ALWAYS);
 203         } else {
 204                 return (inst->qpu.flags.ac != V3D_QPU_COND_NONE &&
 205                         inst->qpu.flags.mc != V3D_QPU_COND_NONE);
 206         }
 207 }
 208
 209 bool
 210 vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst)
 211 {
 212         for (int i = 0; i < vir_get_nsrc(inst); i++) {
 213                 switch (inst->src[i].file) {
 214                 case QFILE_VPM:
 215                         return true;
 216                 default:
 217                         break;
 218                 }
 219         }
 220
 221         if (devinfo->ver < 41 && (inst->qpu.sig.ldvary ||
 222                                   inst->qpu.sig.ldtlb ||
 223                                   inst->qpu.sig.ldtlbu ||
 224                                   inst->qpu.sig.ldvpm)) {
 225                 return true;
 226         }
 227
 228         return false;
 229 }
 230
 231 bool
 232 vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst)
 233 {
 234         switch (inst->dst.file) {
 235         case QFILE_MAGIC:
 236                 switch (inst->dst.index) {
 237                 case V3D_QPU_WADDR_RECIP:
 238                 case V3D_QPU_WADDR_RSQRT:
 239                 case V3D_QPU_WADDR_EXP:
 240                 case V3D_QPU_WADDR_LOG:
 241                 case V3D_QPU_WADDR_SIN:
 242                         return true;
 243                 }
 244                 break;
 245         default:
 246                 break;
 247         }
 248
 249         if (devinfo->ver < 41 && inst->qpu.sig.ldtmu)
 250                 return true;
 251
 252         return false;
 253 }
 254
 255 void
 256 vir_set_unpack(struct qinst *inst, int src,
 257                enum v3d_qpu_input_unpack unpack)
 258 {
 259         assert(src == 0 || src == 1);
 260
 261         if (vir_is_add(inst)) {
 262                 if (src == 0)
 263                         inst->qpu.alu.add.a_unpack = unpack;
 264                 else
 265                         inst->qpu.alu.add.b_unpack = unpack;
 266         } else {
 267                 assert(vir_is_mul(inst));
 268                 if (src == 0)
 269                         inst->qpu.alu.mul.a_unpack = unpack;
 270                 else
 271                         inst->qpu.alu.mul.b_unpack = unpack;
 272         }
 273 }
 274
 275 void
 276 vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond)
 277 {
 278         if (vir_is_add(inst)) {
 279                 inst->qpu.flags.ac = cond;
 280         } else {
 281                 assert(vir_is_mul(inst));
 282                 inst->qpu.flags.mc = cond;
 283         }
 284 }
 285
 286 void
 287 vir_set_pf(struct qinst *inst, enum v3d_qpu_pf pf)
 288 {
 289         if (vir_is_add(inst)) {
 290                 inst->qpu.flags.apf = pf;
 291         } else {
 292                 assert(vir_is_mul(inst));
 293                 inst->qpu.flags.mpf = pf;
 294         }
 295 }
 296
 297 #if 0
 298 uint8_t
 299 vir_channels_written(struct qinst *inst)
 300 {
 301         if (vir_is_mul(inst)) {
 302                 switch (inst->dst.pack) {
 303                 case QPU_PACK_MUL_NOP:
 304                 case QPU_PACK_MUL_8888:
 305                         return 0xf;
 306                 case QPU_PACK_MUL_8A:
 307                         return 0x1;
 308                 case QPU_PACK_MUL_8B:
 309                         return 0x2;
 310                 case QPU_PACK_MUL_8C:
 311                         return 0x4;
 312                 case QPU_PACK_MUL_8D:
 313                         return 0x8;
 314                 }
 315         } else {
 316                 switch (inst->dst.pack) {
 317                 case QPU_PACK_A_NOP:
 318                 case QPU_PACK_A_8888:
 319                 case QPU_PACK_A_8888_SAT:
 320                 case QPU_PACK_A_32_SAT:
 321                         return 0xf;
 322                 case QPU_PACK_A_8A:
 323                 case QPU_PACK_A_8A_SAT:
 324                         return 0x1;
 325                 case QPU_PACK_A_8B:
 326                 case QPU_PACK_A_8B_SAT:
 327                         return 0x2;
 328                 case QPU_PACK_A_8C:
 329                 case QPU_PACK_A_8C_SAT:
 330                         return 0x4;
 331                 case QPU_PACK_A_8D:
 332                 case QPU_PACK_A_8D_SAT:
 333                         return 0x8;
 334                 case QPU_PACK_A_16A:
 335                 case QPU_PACK_A_16A_SAT:
 336                         return 0x3;
 337                 case QPU_PACK_A_16B:
 338                 case QPU_PACK_A_16B_SAT:
 339                         return 0xc;
 340                 }
 341         }
 342         unreachable("Bad pack field");
 343 }
 344 #endif
 345
 346 struct qreg
 347 vir_get_temp(struct v3d_compile *c)
 348 {
 349         struct qreg reg;
 350
 351         reg.file = QFILE_TEMP;
 352         reg.index = c->num_temps++;
 353
 354         if (c->num_temps > c->defs_array_size) {
 355                 uint32_t old_size = c->defs_array_size;
 356                 c->defs_array_size = MAX2(old_size * 2, 16);
 357                 c->defs = reralloc(c, c->defs, struct qinst *,
 358                                    c->defs_array_size);
 359                 memset(&c->defs[old_size], 0,
 360                        sizeof(c->defs[0]) * (c->defs_array_size - old_size));
 361         }
 362
 363         return reg;
 364 }
 365
 366 struct qinst *
 367 vir_add_inst(enum v3d_qpu_add_op op, struct qreg dst, struct qreg src0, struct qreg src1)
 368 {
 369         struct qinst *inst = calloc(1, sizeof(*inst));
 370
 371         inst->qpu = v3d_qpu_nop();
 372         inst->qpu.alu.add.op = op;
 373
 374         inst->dst = dst;
 375         inst->src[0] = src0;
 376         inst->src[1] = src1;
 377         inst->uniform = ~0;
 378
 379         return inst;
 380 }
 381
 382 struct qinst *
 383 vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst, struct qreg src0, struct qreg src1)
 384 {
 385         struct qinst *inst = calloc(1, sizeof(*inst));
 386
 387         inst->qpu = v3d_qpu_nop();
 388         inst->qpu.alu.mul.op = op;
 389
 390         inst->dst = dst;
 391         inst->src[0] = src0;
 392         inst->src[1] = src1;
 393         inst->uniform = ~0;
 394
 395         return inst;
 396 }
 397
 398 struct qinst *
 399 vir_branch_inst(enum v3d_qpu_branch_cond cond, struct qreg src)
 400 {
 401         struct qinst *inst = calloc(1, sizeof(*inst));
 402
 403         inst->qpu = v3d_qpu_nop();
 404         inst->qpu.type = V3D_QPU_INSTR_TYPE_BRANCH;
 405         inst->qpu.branch.cond = cond;
 406         inst->qpu.branch.msfign = V3D_QPU_MSFIGN_NONE;
 407         inst->qpu.branch.bdi = V3D_QPU_BRANCH_DEST_REL;
 408         inst->qpu.branch.ub = true;
 409         inst->qpu.branch.bdu = V3D_QPU_BRANCH_DEST_REL;
 410
 411         inst->dst = vir_reg(QFILE_NULL, 0);
 412         inst->src[0] = src;
 413         inst->uniform = ~0;
 414
 415         return inst;
 416 }
 417
 418 static void
 419 vir_emit(struct v3d_compile *c, struct qinst *inst)
 420 {
 421         switch (c->cursor.mode) {
 422         case vir_cursor_add:
 423                 list_add(&inst->link, c->cursor.link);
 424                 break;
 425         case vir_cursor_addtail:
 426                 list_addtail(&inst->link, c->cursor.link);
 427                 break;
 428         }
 429
 430         c->cursor = vir_after_inst(inst);
 431 }
 432
 433 /* Updates inst to write to a new temporary, emits it, and notes the def. */
 434 struct qreg
 435 vir_emit_def(struct v3d_compile *c, struct qinst *inst)
 436 {
 437         assert(inst->dst.file == QFILE_NULL);
 438
 439         inst->dst = vir_get_temp(c);
 440
 441         if (inst->dst.file == QFILE_TEMP)
 442                 c->defs[inst->dst.index] = inst;
 443
 444         vir_emit(c, inst);
 445
 446         return inst->dst;
 447 }
 448
 449 struct qinst *
 450 vir_emit_nondef(struct v3d_compile *c, struct qinst *inst)
 451 {
 452         if (inst->dst.file == QFILE_TEMP)
 453                 c->defs[inst->dst.index] = NULL;
 454
 455         vir_emit(c, inst);
 456
 457         return inst;
 458 }
 459
 460 struct qblock *
 461 vir_new_block(struct v3d_compile *c)
 462 {
 463         struct qblock *block = rzalloc(c, struct qblock);
 464
 465         list_inithead(&block->instructions);
 466
 467         block->predecessors = _mesa_set_create(block,
 468                                                _mesa_hash_pointer,
 469                                                _mesa_key_pointer_equal);
 470
 471         block->index = c->next_block_index++;
 472
 473         return block;
 474 }
 475
 476 void
 477 vir_set_emit_block(struct v3d_compile *c, struct qblock *block)
 478 {
 479         c->cur_block = block;
 480         c->cursor = vir_after_block(block);
 481         list_addtail(&block->link, &c->blocks);
 482 }
 483
 484 struct qblock *
 485 vir_entry_block(struct v3d_compile *c)
 486 {
 487         return list_first_entry(&c->blocks, struct qblock, link);
 488 }
 489
 490 struct qblock *
 491 vir_exit_block(struct v3d_compile *c)
 492 {
 493         return list_last_entry(&c->blocks, struct qblock, link);
 494 }
 495
 496 void
 497 vir_link_blocks(struct qblock *predecessor, struct qblock *successor)
 498 {
 499         _mesa_set_add(successor->predecessors, predecessor);
 500         if (predecessor->successors[0]) {
 501                 assert(!predecessor->successors[1]);
 502                 predecessor->successors[1] = successor;
 503         } else {
 504                 predecessor->successors[0] = successor;
 505         }
 506 }
 507
 508 const struct v3d_compiler *
 509 v3d_compiler_init(const struct v3d_device_info *devinfo)
 510 {
 511         struct v3d_compiler *compiler = rzalloc(NULL, struct v3d_compiler);
 512         if (!compiler)
 513                 return NULL;
 514
 515         compiler->devinfo = devinfo;
 516
 517         if (!vir_init_reg_sets(compiler)) {
 518                 ralloc_free(compiler);
 519                 return NULL;
 520         }
 521
 522         return compiler;
 523 }
 524
 525 void
 526 v3d_compiler_free(const struct v3d_compiler *compiler)
 527 {
 528         ralloc_free((void *)compiler);
 529 }
 530
 531 static struct v3d_compile *
 532 vir_compile_init(const struct v3d_compiler *compiler,
 533                  struct v3d_key *key,
 534                  nir_shader *s,
 535                  int program_id, int variant_id)
 536 {
 537         struct v3d_compile *c = rzalloc(NULL, struct v3d_compile);
 538
 539         c->compiler = compiler;
 540         c->devinfo = compiler->devinfo;
 541         c->key = key;
 542         c->program_id = program_id;
 543         c->variant_id = variant_id;
 544         c->threads = 4;
 545
 546         s = nir_shader_clone(c, s);
 547         c->s = s;
 548
 549         list_inithead(&c->blocks);
 550         vir_set_emit_block(c, vir_new_block(c));
 551
 552         c->output_position_index = -1;
 553         c->output_point_size_index = -1;
 554         c->output_sample_mask_index = -1;
 555
 556         c->def_ht = _mesa_hash_table_create(c, _mesa_hash_pointer,
 557                                             _mesa_key_pointer_equal);
 558
 559         return c;
 560 }
 561
 562 static void
 563 v3d_lower_nir(struct v3d_compile *c)
 564 {
 565         struct nir_lower_tex_options tex_options = {
 566                 .lower_txd = true,
 567                 .lower_rect = false, /* XXX */
 568                 .lower_txp = ~0,
 569                 /* Apply swizzles to all samplers. */
 570                 .swizzle_result = ~0,
 571         };
 572
 573         /* Lower the format swizzle and (for 32-bit returns)
 574          * ARB_texture_swizzle-style swizzle.
 575          */
 576         for (int i = 0; i < ARRAY_SIZE(c->key->tex); i++) {
 577                 for (int j = 0; j < 4; j++)
 578                         tex_options.swizzles[i][j] = c->key->tex[i].swizzle[j];
 579
 580                 if (c->key->tex[i].clamp_s)
 581                         tex_options.saturate_s |= 1 << i;
 582                 if (c->key->tex[i].clamp_t)
 583                         tex_options.saturate_t |= 1 << i;
 584                 if (c->key->tex[i].clamp_r)
 585                         tex_options.saturate_r |= 1 << i;
 586         }
 587
 588         NIR_PASS_V(c->s, nir_lower_tex, &tex_options);
 589 }
 590
 591 static void
 592 v3d_lower_nir_late(struct v3d_compile *c)
 593 {
 594         NIR_PASS_V(c->s, v3d_nir_lower_io, c);
 595         NIR_PASS_V(c->s, v3d_nir_lower_txf_ms, c);
 596         NIR_PASS_V(c->s, nir_lower_idiv);
 597 }
 598
 599 static void
 600 v3d_set_prog_data_uniforms(struct v3d_compile *c,
 601                            struct v3d_prog_data *prog_data)
 602 {
 603         int count = c->num_uniforms;
 604         struct v3d_uniform_list *ulist = &prog_data->uniforms;
 605
 606         ulist->count = count;
 607         ulist->data = ralloc_array(prog_data, uint32_t, count);
 608         memcpy(ulist->data, c->uniform_data,
 609                count * sizeof(*ulist->data));
 610         ulist->contents = ralloc_array(prog_data, enum quniform_contents, count);
 611         memcpy(ulist->contents, c->uniform_contents,
 612                count * sizeof(*ulist->contents));
 613 }
 614
 615 /* Copy the compiler UBO range state to the compiled shader, dropping out
 616  * arrays that were never referenced by an indirect load.
 617  *
 618  * (Note that QIR dead code elimination of an array access still leaves that
 619  * array alive, though)
 620  */
 621 static void
 622 v3d_set_prog_data_ubo(struct v3d_compile *c,
 623                       struct v3d_prog_data *prog_data)
 624 {
 625         if (!c->num_ubo_ranges)
 626                 return;
 627
 628         prog_data->num_ubo_ranges = 0;
 629         prog_data->ubo_ranges = ralloc_array(prog_data, struct v3d_ubo_range,
 630                                              c->num_ubo_ranges);
 631         for (int i = 0; i < c->num_ubo_ranges; i++) {
 632                 if (!c->ubo_range_used[i])
 633                         continue;
 634
 635                 struct v3d_ubo_range *range = &c->ubo_ranges[i];
 636                 prog_data->ubo_ranges[prog_data->num_ubo_ranges++] = *range;
 637                 prog_data->ubo_size += range->size;
 638         }
 639
 640         if (prog_data->ubo_size) {
 641                 if (V3D_DEBUG & V3D_DEBUG_SHADERDB) {
 642                         fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d UBO uniforms\n",
 643                                 vir_get_stage_name(c),
 644                                 c->program_id, c->variant_id,
 645                                 prog_data->ubo_size / 4);
 646                 }
 647         }
 648 }
 649
 650 static void
 651 v3d_set_prog_data(struct v3d_compile *c,
 652                   struct v3d_prog_data *prog_data)
 653 {
 654         prog_data->threads = c->threads;
 655         prog_data->single_seg = !c->last_thrsw;
 656
 657         v3d_set_prog_data_uniforms(c, prog_data);
 658         v3d_set_prog_data_ubo(c, prog_data);
 659 }
 660
 661 static uint64_t *
 662 v3d_return_qpu_insts(struct v3d_compile *c, uint32_t *final_assembly_size)
 663 {
 664         *final_assembly_size = c->qpu_inst_count * sizeof(uint64_t);
 665
 666         uint64_t *qpu_insts = malloc(*final_assembly_size);
 667         if (!qpu_insts)
 668                 return NULL;
 669
 670         memcpy(qpu_insts, c->qpu_insts, *final_assembly_size);
 671
 672         vir_compile_destroy(c);
 673
 674         return qpu_insts;
 675 }
 676
 677 uint64_t *v3d_compile_vs(const struct v3d_compiler *compiler,
 678                          struct v3d_vs_key *key,
 679                          struct v3d_vs_prog_data *prog_data,
 680                          nir_shader *s,
 681                          int program_id, int variant_id,
 682                          uint32_t *final_assembly_size)
 683 {
 684         struct v3d_compile *c = vir_compile_init(compiler, &key->base, s,
 685                                                  program_id, variant_id);
 686
 687         c->vs_key = key;
 688
 689         v3d_lower_nir(c);
 690
 691         if (key->clamp_color)
 692                 NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
 693
 694         if (key->base.ucp_enables) {
 695                 NIR_PASS_V(c->s, nir_lower_clip_vs, key->base.ucp_enables);
 696                 NIR_PASS_V(c->s, nir_lower_io_to_scalar,
 697                            nir_var_shader_out);
 698         }
 699
 700         /* Note: VS output scalarizing must happen after nir_lower_clip_vs. */
 701         NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_out);
 702
 703         v3d_lower_nir_late(c);
 704         v3d_optimize_nir(c->s);
 705         NIR_PASS_V(c->s, nir_convert_from_ssa, true);
 706
 707         v3d_nir_to_vir(c);
 708
 709         v3d_set_prog_data(c, &prog_data->base);
 710
 711         prog_data->base.num_inputs = c->num_inputs;
 712
 713         /* The vertex data gets format converted by the VPM so that
 714          * each attribute channel takes up a VPM column.  Precompute
 715          * the sizes for the shader record.
 716          */
 717         for (int i = 0; i < ARRAY_SIZE(prog_data->vattr_sizes); i++) {
 718                 prog_data->vattr_sizes[i] = c->vattr_sizes[i];
 719                 prog_data->vpm_input_size += c->vattr_sizes[i];
 720         }
 721
 722         /* Input/output segment size are in 8x32-bit multiples. */
 723         prog_data->vpm_input_size = align(prog_data->vpm_input_size, 8) / 8;
 724         prog_data->vpm_output_size = align(c->num_vpm_writes, 8) / 8;
 725
 726         prog_data->uses_vid = (s->info.system_values_read &
 727                                (1ull << SYSTEM_VALUE_VERTEX_ID));
 728         prog_data->uses_iid = (s->info.system_values_read &
 729                                (1ull << SYSTEM_VALUE_INSTANCE_ID));
 730
 731         return v3d_return_qpu_insts(c, final_assembly_size);
 732 }
 733
 734 static void
 735 v3d_set_fs_prog_data_inputs(struct v3d_compile *c,
 736                             struct v3d_fs_prog_data *prog_data)
 737 {
 738         prog_data->base.num_inputs = c->num_inputs;
 739         memcpy(prog_data->input_slots, c->input_slots,
 740                c->num_inputs * sizeof(*c->input_slots));
 741
 742         STATIC_ASSERT(ARRAY_SIZE(prog_data->flat_shade_flags) >
 743                       (V3D_MAX_FS_INPUTS - 1) / 24);
 744         for (int i = 0; i < V3D_MAX_FS_INPUTS; i++) {
 745                 if (BITSET_TEST(c->flat_shade_flags, i))
 746                         prog_data->flat_shade_flags[i / 24] |= 1 << (i % 24);
 747         }
 748 }
 749
 750 uint64_t *v3d_compile_fs(const struct v3d_compiler *compiler,
 751                          struct v3d_fs_key *key,
 752                          struct v3d_fs_prog_data *prog_data,
 753                          nir_shader *s,
 754                          int program_id, int variant_id,
 755                          uint32_t *final_assembly_size)
 756 {
 757         struct v3d_compile *c = vir_compile_init(compiler, &key->base, s,
 758                                                  program_id, variant_id);
 759
 760         c->fs_key = key;
 761
 762         v3d_lower_nir(c);
 763
 764         if (key->light_twoside)
 765                 NIR_PASS_V(c->s, nir_lower_two_sided_color);
 766
 767         if (key->clamp_color)
 768                 NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
 769
 770         if (key->alpha_test) {
 771                 NIR_PASS_V(c->s, nir_lower_alpha_test, key->alpha_test_func,
 772                            false);
 773         }
 774
 775         if (key->base.ucp_enables)
 776                 NIR_PASS_V(c->s, nir_lower_clip_fs, key->base.ucp_enables);
 777
 778         /* Note: FS input scalarizing must happen after
 779          * nir_lower_two_sided_color, which only handles a vec4 at a time.
 780          */
 781         NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_in);
 782
 783         v3d_lower_nir_late(c);
 784         v3d_optimize_nir(c->s);
 785         NIR_PASS_V(c->s, nir_convert_from_ssa, true);
 786
 787         v3d_nir_to_vir(c);
 788
 789         v3d_set_prog_data(c, &prog_data->base);
 790         v3d_set_fs_prog_data_inputs(c, prog_data);
 791         prog_data->writes_z = (c->s->info.outputs_written &
 792                                (1 << FRAG_RESULT_DEPTH));
 793         prog_data->discard = c->s->info.fs.uses_discard;
 794
 795         return v3d_return_qpu_insts(c, final_assembly_size);
 796 }
 797
 798 void
 799 vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst)
 800 {
 801         if (qinst->dst.file == QFILE_TEMP)
 802                 c->defs[qinst->dst.index] = NULL;
 803
 804         assert(&qinst->link != c->cursor.link);
 805
 806         list_del(&qinst->link);
 807         free(qinst);
 808 }
 809
 810 struct qreg
 811 vir_follow_movs(struct v3d_compile *c, struct qreg reg)
 812 {
 813         /* XXX
 814         int pack = reg.pack;
 815
 816         while (reg.file == QFILE_TEMP &&
 817                c->defs[reg.index] &&
 818                (c->defs[reg.index]->op == QOP_MOV ||
 819                 c->defs[reg.index]->op == QOP_FMOV) &&
 820                !c->defs[reg.index]->dst.pack &&
 821                !c->defs[reg.index]->src[0].pack) {
 822                 reg = c->defs[reg.index]->src[0];
 823         }
 824
 825         reg.pack = pack;
 826         */
 827         return reg;
 828 }
 829
 830 void
 831 vir_compile_destroy(struct v3d_compile *c)
 832 {
 833         /* Defuse the assert that we aren't removing the cursor's instruction.
 834          */
 835         c->cursor.link = NULL;
 836
 837         vir_for_each_block(block, c) {
 838                 while (!list_empty(&block->instructions)) {
 839                         struct qinst *qinst =
 840                                 list_first_entry(&block->instructions,
 841                                                  struct qinst, link);
 842                         vir_remove_instruction(c, qinst);
 843                 }
 844         }
 845
 846         ralloc_free(c);
 847 }
 848
 849 struct qreg
 850 vir_uniform(struct v3d_compile *c,
 851             enum quniform_contents contents,
 852             uint32_t data)
 853 {
 854         for (int i = 0; i < c->num_uniforms; i++) {
 855                 if (c->uniform_contents[i] == contents &&
 856                     c->uniform_data[i] == data) {
 857                         return vir_reg(QFILE_UNIF, i);
 858                 }
 859         }
 860
 861         uint32_t uniform = c->num_uniforms++;
 862
 863         if (uniform >= c->uniform_array_size) {
 864                 c->uniform_array_size = MAX2(MAX2(16, uniform + 1),
 865                                              c->uniform_array_size * 2);
 866
 867                 c->uniform_data = reralloc(c, c->uniform_data,
 868                                            uint32_t,
 869                                            c->uniform_array_size);
 870                 c->uniform_contents = reralloc(c, c->uniform_contents,
 871                                                enum quniform_contents,
 872                                                c->uniform_array_size);
 873         }
 874
 875         c->uniform_contents[uniform] = contents;
 876         c->uniform_data[uniform] = data;
 877
 878         return vir_reg(QFILE_UNIF, uniform);
 879 }
 880
 881 void
 882 vir_PF(struct v3d_compile *c, struct qreg src, enum v3d_qpu_pf pf)
 883 {
 884         struct qinst *last_inst = NULL;
 885
 886         if (!list_empty(&c->cur_block->instructions)) {
 887                 last_inst = (struct qinst *)c->cur_block->instructions.prev;
 888
 889                 /* Can't stuff the PF into the last last inst if our cursor
 890                  * isn't pointing after it.
 891                  */
 892                 struct vir_cursor after_inst = vir_after_inst(last_inst);
 893                 if (c->cursor.mode != after_inst.mode ||
 894                     c->cursor.link != after_inst.link)
 895                         last_inst = NULL;
 896         }
 897
 898         if (src.file != QFILE_TEMP ||
 899             !c->defs[src.index] ||
 900             last_inst != c->defs[src.index]) {
 901                 /* XXX: Make the MOV be the appropriate type */
 902                 last_inst = vir_MOV_dest(c, vir_reg(QFILE_NULL, 0), src);
 903                 last_inst = (struct qinst *)c->cur_block->instructions.prev;
 904         }
 905
 906         vir_set_pf(last_inst, pf);
 907 }
 908
 909 #define OPTPASS(func)                                                   \
 910         do {                                                            \
 911                 bool stage_progress = func(c);                          \
 912                 if (stage_progress) {                                   \
 913                         progress = true;                                \
 914                         if (print_opt_debug) {                          \
 915                                 fprintf(stderr,                         \
 916                                         "VIR opt pass %2d: %s progress\n", \
 917                                         pass, #func);                   \
 918                         }                                               \
 919                         /*XXX vir_validate(c);*/                        \
 920                 }                                                       \
 921         } while (0)
 922
 923 void
 924 vir_optimize(struct v3d_compile *c)
 925 {
 926         bool print_opt_debug = false;
 927         int pass = 1;
 928
 929         while (true) {
 930                 bool progress = false;
 931
 932                 OPTPASS(vir_opt_copy_propagate);
 933                 OPTPASS(vir_opt_dead_code);
 934
 935                 if (!progress)
 936                         break;
 937
 938                 pass++;
 939         }
 940 }
 941
 942 const char *
 943 vir_get_stage_name(struct v3d_compile *c)
 944 {
 945         if (c->vs_key && c->vs_key->is_coord)
 946                 return "MESA_SHADER_COORD";
 947         else
 948                 return gl_shader_stage_name(c->s->info.stage);
 949 }