src/panfrost/bifrost/bifrost_compile.c

   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  * Authors (Collabora):
  24  *      Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
  25  */
  26
  27 #include "main/mtypes.h"
  28 #include "compiler/glsl/glsl_to_nir.h"
  29 #include "compiler/nir_types.h"
  30 #include "main/imports.h"
  31 #include "compiler/nir/nir_builder.h"
  32
  33 #include "disassemble.h"
  34 #include "bifrost_compile.h"
  35 #include "bifrost_nir.h"
  36 #include "compiler.h"
  37 #include "bi_quirks.h"
  38 #include "bi_print.h"
  39
  40 static bi_block *emit_cf_list(bi_context *ctx, struct exec_list *list);
  41 static bi_instruction *bi_emit_branch(bi_context *ctx);
  42 static void bi_block_add_successor(bi_block *block, bi_block *successor);
  43 static void bi_schedule_barrier(bi_context *ctx);
  44
  45 static void
  46 emit_jump(bi_context *ctx, nir_jump_instr *instr)
  47 {
  48         bi_instruction *branch = bi_emit_branch(ctx);
  49
  50         switch (instr->type) {
  51         case nir_jump_break:
  52                 branch->branch.target = ctx->break_block;
  53                 break;
  54         case nir_jump_continue:
  55                 branch->branch.target = ctx->continue_block;
  56                 break;
  57         default:
  58                 unreachable("Unhandled jump type");
  59         }
  60
  61         bi_block_add_successor(ctx->current_block, branch->branch.target);
  62 }
  63
  64 /* Gets a bytemask for a complete vecN write */
  65 static unsigned
  66 bi_mask_for_channels_32(unsigned i)
  67 {
  68         return (1 << (4 * i)) - 1;
  69 }
  70
  71 static bi_instruction
  72 bi_load(enum bi_class T, nir_intrinsic_instr *instr)
  73 {
  74         bi_instruction load = {
  75                 .type = T,
  76                 .writemask = bi_mask_for_channels_32(instr->num_components),
  77                 .src = { BIR_INDEX_CONSTANT },
  78                 .constant = { .u64 = nir_intrinsic_base(instr) },
  79         };
  80
  81         const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic];
  82
  83         if (info->has_dest)
  84                 load.dest = bir_dest_index(&instr->dest);
  85
  86         if (info->has_dest && info->index_map[NIR_INTRINSIC_TYPE] > 0)
  87                 load.dest_type = nir_intrinsic_type(instr);
  88
  89         nir_src *offset = nir_get_io_offset_src(instr);
  90
  91         if (nir_src_is_const(*offset))
  92                 load.constant.u64 += nir_src_as_uint(*offset);
  93         else
  94                 load.src[0] = bir_src_index(offset);
  95
  96         return load;
  97 }
  98
  99 static void
 100 bi_emit_ld_vary(bi_context *ctx, nir_intrinsic_instr *instr)
 101 {
 102         bi_instruction ins = bi_load(BI_LOAD_VAR, instr);
 103         ins.load_vary.interp_mode = BIFROST_INTERP_DEFAULT; /* TODO */
 104         ins.load_vary.reuse = false; /* TODO */
 105         ins.load_vary.flat = instr->intrinsic != nir_intrinsic_load_interpolated_input;
 106         ins.dest_type = nir_type_float | nir_dest_bit_size(instr->dest),
 107         bi_emit(ctx, ins);
 108 }
 109
 110 static void
 111 bi_emit_frag_out(bi_context *ctx, nir_intrinsic_instr *instr)
 112 {
 113         if (!ctx->emitted_atest) {
 114                 bi_instruction ins = {
 115                         .type = BI_ATEST
 116                 };
 117
 118                 bi_emit(ctx, ins);
 119                 bi_schedule_barrier(ctx);
 120                 ctx->emitted_atest = true;
 121         }
 122
 123         bi_instruction blend = {
 124                 .type = BI_BLEND,
 125                 .blend_location = nir_intrinsic_base(instr),
 126                 .src = {
 127                         bir_src_index(&instr->src[0])
 128                 },
 129                 .swizzle = {
 130                         { 0, 1, 2, 3 }
 131                 }
 132         };
 133
 134         bi_emit(ctx, blend);
 135         bi_schedule_barrier(ctx);
 136 }
 137
 138 static void
 139 bi_emit_st_vary(bi_context *ctx, nir_intrinsic_instr *instr)
 140 {
 141         bi_instruction address = bi_load(BI_LOAD_VAR_ADDRESS, instr);
 142         address.dest = bi_make_temp(ctx);
 143         address.dest_type = nir_type_uint64;
 144         address.writemask = (1 << 8) - 1;
 145
 146         bi_instruction st = {
 147                 .type = BI_STORE_VAR,
 148                 .src = {
 149                         address.dest,
 150                         bir_src_index(&instr->src[0])
 151                 },
 152                 .swizzle = {
 153                         { 0, 1, 2, 3 }
 154                 }
 155         };
 156
 157         bi_emit(ctx, address);
 158         bi_emit(ctx, st);
 159 }
 160
 161 static void
 162 bi_emit_ld_uniform(bi_context *ctx, nir_intrinsic_instr *instr)
 163 {
 164         bi_instruction ld = bi_load(BI_LOAD_UNIFORM, instr);
 165         ld.src[1] = BIR_INDEX_ZERO; /* TODO: UBO index */
 166         bi_emit(ctx, ld);
 167 }
 168
 169 static void
 170 emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr)
 171 {
 172
 173         switch (instr->intrinsic) {
 174         case nir_intrinsic_load_barycentric_pixel:
 175                 /* stub */
 176                 break;
 177         case nir_intrinsic_load_interpolated_input:
 178         case nir_intrinsic_load_input:
 179                 if (ctx->stage == MESA_SHADER_FRAGMENT)
 180                         bi_emit_ld_vary(ctx, instr);
 181                 else if (ctx->stage == MESA_SHADER_VERTEX)
 182                         bi_emit(ctx, bi_load(BI_LOAD_ATTR, instr));
 183                 else {
 184                         unreachable("Unsupported shader stage");
 185                 }
 186                 break;
 187
 188         case nir_intrinsic_store_output:
 189                 if (ctx->stage == MESA_SHADER_FRAGMENT)
 190                         bi_emit_frag_out(ctx, instr);
 191                 else if (ctx->stage == MESA_SHADER_VERTEX)
 192                         bi_emit_st_vary(ctx, instr);
 193                 else
 194                         unreachable("Unsupported shader stage");
 195                 break;
 196
 197         case nir_intrinsic_load_uniform:
 198                 bi_emit_ld_uniform(ctx, instr);
 199                 break;
 200
 201         default:
 202                 /* todo */
 203                 break;
 204         }
 205 }
 206
 207 static void
 208 emit_load_const(bi_context *ctx, nir_load_const_instr *instr)
 209 {
 210         /* Make sure we've been lowered */
 211         assert(instr->def.num_components == 1);
 212
 213         bi_instruction move = {
 214                 .type = BI_MOV,
 215                 .dest = bir_ssa_index(&instr->def),
 216                 .dest_type = instr->def.bit_size | nir_type_uint,
 217                 .writemask = (1 << (instr->def.bit_size / 8)) - 1,
 218                 .src = {
 219                         BIR_INDEX_CONSTANT
 220                 },
 221                 .constant = {
 222                         .u64 = nir_const_value_as_uint(instr->value[0], instr->def.bit_size)
 223                 }
 224         };
 225
 226         bi_emit(ctx, move);
 227 }
 228
 229 static enum bi_class
 230 bi_class_for_nir_alu(nir_op op)
 231 {
 232         switch (op) {
 233         case nir_op_iadd:
 234         case nir_op_fadd:
 235         case nir_op_fsub:
 236                 return BI_ADD;
 237         case nir_op_isub:
 238                 return BI_ISUB;
 239
 240         case nir_op_i2i8:
 241         case nir_op_i2i16:
 242         case nir_op_i2i32:
 243         case nir_op_i2i64:
 244         case nir_op_u2u8:
 245         case nir_op_u2u16:
 246         case nir_op_u2u32:
 247         case nir_op_u2u64:
 248         case nir_op_f2i16:
 249         case nir_op_f2i32:
 250         case nir_op_f2i64:
 251         case nir_op_f2u16:
 252         case nir_op_f2u32:
 253         case nir_op_f2u64:
 254         case nir_op_i2f16:
 255         case nir_op_i2f32:
 256         case nir_op_i2f64:
 257         case nir_op_u2f16:
 258         case nir_op_u2f32:
 259         case nir_op_u2f64:
 260                 return BI_CONVERT;
 261
 262         case nir_op_fmul:
 263                 return BI_FMA;
 264
 265         case nir_op_imin:
 266         case nir_op_imax:
 267         case nir_op_umin:
 268         case nir_op_umax:
 269         case nir_op_fmin:
 270         case nir_op_fmax:
 271                 return BI_MINMAX;
 272
 273         case nir_op_fsat:
 274         case nir_op_fneg:
 275         case nir_op_fabs:
 276         case nir_op_mov:
 277                 return BI_MOV;
 278
 279         case nir_op_frcp:
 280         case nir_op_frsq:
 281         case nir_op_fsin:
 282         case nir_op_fcos:
 283                 return BI_SPECIAL;
 284
 285         default:
 286                 unreachable("Unknown ALU op");
 287         }
 288 }
 289
 290 static void
 291 emit_alu(bi_context *ctx, nir_alu_instr *instr)
 292 {
 293         /* Assume it's something we can handle normally */
 294         bi_instruction alu = {
 295                 .type = bi_class_for_nir_alu(instr->op),
 296                 .dest = bir_dest_index(&instr->dest.dest),
 297                 .dest_type = nir_op_infos[instr->op].output_type
 298                         | nir_dest_bit_size(instr->dest.dest),
 299         };
 300
 301         /* TODO: Implement lowering of special functions for older Bifrost */
 302         assert((alu.type != BI_SPECIAL) || !(ctx->quirks & BIFROST_NO_FAST_OP));
 303
 304         if (instr->dest.dest.is_ssa) {
 305                 /* Construct a writemask */
 306                 unsigned bits_per_comp = instr->dest.dest.ssa.bit_size;
 307                 unsigned comps = instr->dest.dest.ssa.num_components;
 308                 assert(comps == 1);
 309                 unsigned bits = bits_per_comp * comps;
 310                 unsigned bytes = MAX2(bits / 8, 1);
 311                 alu.writemask = (1 << bytes) - 1;
 312         } else {
 313                 unsigned comp_mask = instr->dest.write_mask;
 314
 315                 alu.writemask = pan_to_bytemask(nir_dest_bit_size(instr->dest.dest),
 316                                 comp_mask);
 317         }
 318
 319         /* We inline constants as we go. This tracks how many constants have
 320          * been inlined, since we're limited to 64-bits of constants per
 321          * instruction */
 322
 323         unsigned dest_bits = nir_dest_bit_size(instr->dest.dest);
 324         unsigned constants_left = (64 / dest_bits);
 325         unsigned constant_shift = 0;
 326
 327         /* Copy sources */
 328
 329         unsigned num_inputs = nir_op_infos[instr->op].num_inputs;
 330         assert(num_inputs <= ARRAY_SIZE(alu.src));
 331
 332         for (unsigned i = 0; i < num_inputs; ++i) {
 333                 unsigned bits = nir_src_bit_size(instr->src[i].src);
 334                 alu.src_types[i] = nir_op_infos[instr->op].input_types[i]
 335                         | bits;
 336
 337                 /* Try to inline a constant */
 338                 if (nir_src_is_const(instr->src[i].src) && constants_left && (dest_bits == bits)) {
 339                         alu.constant.u64 |=
 340                                 (nir_src_as_uint(instr->src[i].src)) << constant_shift;
 341
 342                         alu.src[i] = BIR_INDEX_CONSTANT | constant_shift;
 343                         --constants_left;
 344                         constant_shift += dest_bits;
 345                         continue;
 346                 }
 347
 348                 alu.src[i] = bir_src_index(&instr->src[i].src);
 349
 350                 /* We assert scalarization above */
 351                 alu.swizzle[i][0] = instr->src[i].swizzle[0];
 352         }
 353
 354         /* Op-specific fixup */
 355         switch (instr->op) {
 356         case nir_op_fmul:
 357                 alu.src[2] = BIR_INDEX_ZERO; /* FMA */
 358                 break;
 359         case nir_op_fsat:
 360                 alu.outmod = BIFROST_SAT; /* MOV */
 361                 break;
 362         case nir_op_fneg:
 363                 alu.src_neg[0] = true; /* MOV */
 364                 break;
 365         case nir_op_fabs:
 366                 alu.src_abs[0] = true; /* MOV */
 367                 break;
 368         case nir_op_fsub:
 369                 alu.src_neg[1] = true; /* ADD */
 370                 break;
 371         case nir_op_fmax:
 372         case nir_op_imax:
 373         case nir_op_umax:
 374                 alu.op.minmax = BI_MINMAX_MAX; /* MINMAX */
 375                 break;
 376         case nir_op_frcp:
 377                 alu.op.special = BI_SPECIAL_FRCP;
 378                 break;
 379         case nir_op_frsq:
 380                 alu.op.special = BI_SPECIAL_FRSQ;
 381                 break;
 382         case nir_op_fsin:
 383                 alu.op.special = BI_SPECIAL_FSIN;
 384                 break;
 385         case nir_op_fcos:
 386                 alu.op.special = BI_SPECIAL_FCOS;
 387                 break;
 388         default:
 389                 break;
 390         }
 391
 392         bi_emit(ctx, alu);
 393 }
 394
 395 static void
 396 emit_instr(bi_context *ctx, struct nir_instr *instr)
 397 {
 398         switch (instr->type) {
 399         case nir_instr_type_load_const:
 400                 emit_load_const(ctx, nir_instr_as_load_const(instr));
 401                 break;
 402
 403         case nir_instr_type_intrinsic:
 404                 emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 405                 break;
 406
 407         case nir_instr_type_alu:
 408                 emit_alu(ctx, nir_instr_as_alu(instr));
 409                 break;
 410
 411 #if 0
 412         case nir_instr_type_tex:
 413                 emit_tex(ctx, nir_instr_as_tex(instr));
 414                 break;
 415 #endif
 416
 417         case nir_instr_type_jump:
 418                 emit_jump(ctx, nir_instr_as_jump(instr));
 419                 break;
 420
 421         case nir_instr_type_ssa_undef:
 422                 /* Spurious */
 423                 break;
 424
 425         default:
 426                 //unreachable("Unhandled instruction type");
 427                 break;
 428         }
 429 }
 430
 431
 432
 433 static bi_block *
 434 create_empty_block(bi_context *ctx)
 435 {
 436         bi_block *blk = rzalloc(ctx, bi_block);
 437
 438         blk->predecessors = _mesa_set_create(blk,
 439                         _mesa_hash_pointer,
 440                         _mesa_key_pointer_equal);
 441
 442         blk->name = ctx->block_name_count++;
 443
 444         return blk;
 445 }
 446
 447 static void
 448 bi_block_add_successor(bi_block *block, bi_block *successor)
 449 {
 450         assert(block);
 451         assert(successor);
 452
 453         for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) {
 454                 if (block->successors[i]) {
 455                        if (block->successors[i] == successor)
 456                                return;
 457                        else
 458                                continue;
 459                 }
 460
 461                 block->successors[i] = successor;
 462                 _mesa_set_add(successor->predecessors, block);
 463                 return;
 464         }
 465
 466         unreachable("Too many successors");
 467 }
 468
 469 static void
 470 bi_schedule_barrier(bi_context *ctx)
 471 {
 472         bi_block *temp = ctx->after_block;
 473         ctx->after_block = create_empty_block(ctx);
 474         list_addtail(&ctx->after_block->link, &ctx->blocks);
 475         list_inithead(&ctx->after_block->instructions);
 476         bi_block_add_successor(ctx->current_block, ctx->after_block);
 477         ctx->current_block = ctx->after_block;
 478         ctx->after_block = temp;
 479 }
 480
 481 static bi_block *
 482 emit_block(bi_context *ctx, nir_block *block)
 483 {
 484         if (ctx->after_block) {
 485                 ctx->current_block = ctx->after_block;
 486                 ctx->after_block = NULL;
 487         } else {
 488                 ctx->current_block = create_empty_block(ctx);
 489         }
 490
 491         list_addtail(&ctx->current_block->link, &ctx->blocks);
 492         list_inithead(&ctx->current_block->instructions);
 493
 494         nir_foreach_instr(instr, block) {
 495                 emit_instr(ctx, instr);
 496                 ++ctx->instruction_count;
 497         }
 498
 499         return ctx->current_block;
 500 }
 501
 502 /* Emits an unconditional branch to the end of the current block, returning a
 503  * pointer so the user can fill in details */
 504
 505 static bi_instruction *
 506 bi_emit_branch(bi_context *ctx)
 507 {
 508         bi_instruction branch = {
 509                 .type = BI_BRANCH,
 510                 .branch = {
 511                         .cond = BI_COND_ALWAYS
 512                 }
 513         };
 514
 515         return bi_emit(ctx, branch);
 516 }
 517
 518 /* Sets a condition for a branch by examing the NIR condition. If we're
 519  * familiar with the condition, we unwrap it to fold it into the branch
 520  * instruction. Otherwise, we consume the condition directly. We
 521  * generally use 1-bit booleans which allows us to use small types for
 522  * the conditions.
 523  */
 524
 525 static void
 526 bi_set_branch_cond(bi_instruction *branch, nir_src *cond, bool invert)
 527 {
 528         /* TODO: Try to unwrap instead of always bailing */
 529         branch->src[0] = bir_src_index(cond);
 530         branch->src[1] = BIR_INDEX_ZERO;
 531         branch->src_types[0] = branch->src_types[1] = nir_type_uint16;
 532         branch->branch.cond = invert ? BI_COND_EQ : BI_COND_NE;
 533 }
 534
 535 static void
 536 emit_if(bi_context *ctx, nir_if *nif)
 537 {
 538         bi_block *before_block = ctx->current_block;
 539
 540         /* Speculatively emit the branch, but we can't fill it in until later */
 541         bi_instruction *then_branch = bi_emit_branch(ctx);
 542         bi_set_branch_cond(then_branch, &nif->condition, true);
 543
 544         /* Emit the two subblocks. */
 545         bi_block *then_block = emit_cf_list(ctx, &nif->then_list);
 546         bi_block *end_then_block = ctx->current_block;
 547
 548         /* Emit a jump from the end of the then block to the end of the else */
 549         bi_instruction *then_exit = bi_emit_branch(ctx);
 550
 551         /* Emit second block, and check if it's empty */
 552
 553         int count_in = ctx->instruction_count;
 554         bi_block *else_block = emit_cf_list(ctx, &nif->else_list);
 555         bi_block *end_else_block = ctx->current_block;
 556         ctx->after_block = create_empty_block(ctx);
 557
 558         /* Now that we have the subblocks emitted, fix up the branches */
 559
 560         assert(then_block);
 561         assert(else_block);
 562
 563         if (ctx->instruction_count == count_in) {
 564                 /* The else block is empty, so don't emit an exit jump */
 565                 bi_remove_instruction(then_exit);
 566                 then_branch->branch.target = ctx->after_block;
 567         } else {
 568                 then_branch->branch.target = else_block;
 569                 then_exit->branch.target = ctx->after_block;
 570                 bi_block_add_successor(end_then_block, then_exit->branch.target);
 571         }
 572
 573         /* Wire up the successors */
 574
 575         bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */
 576
 577         bi_block_add_successor(before_block, then_block); /* fallthrough */
 578         bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
 579 }
 580
 581 static void
 582 emit_loop(bi_context *ctx, nir_loop *nloop)
 583 {
 584         /* Remember where we are */
 585         bi_block *start_block = ctx->current_block;
 586
 587         bi_block *saved_break = ctx->break_block;
 588         bi_block *saved_continue = ctx->continue_block;
 589
 590         ctx->continue_block = create_empty_block(ctx);
 591         ctx->break_block = create_empty_block(ctx);
 592         ctx->after_block = ctx->continue_block;
 593
 594         /* Emit the body itself */
 595         emit_cf_list(ctx, &nloop->body);
 596
 597         /* Branch back to loop back */
 598         bi_instruction *br_back = bi_emit_branch(ctx);
 599         br_back->branch.target = ctx->continue_block;
 600         bi_block_add_successor(start_block, ctx->continue_block);
 601         bi_block_add_successor(ctx->current_block, ctx->continue_block);
 602
 603         ctx->after_block = ctx->break_block;
 604
 605         /* Pop off */
 606         ctx->break_block = saved_break;
 607         ctx->continue_block = saved_continue;
 608         ++ctx->loop_count;
 609 }
 610
 611 static bi_block *
 612 emit_cf_list(bi_context *ctx, struct exec_list *list)
 613 {
 614         bi_block *start_block = NULL;
 615
 616         foreach_list_typed(nir_cf_node, node, node, list) {
 617                 switch (node->type) {
 618                 case nir_cf_node_block: {
 619                         bi_block *block = emit_block(ctx, nir_cf_node_as_block(node));
 620
 621                         if (!start_block)
 622                                 start_block = block;
 623
 624                         break;
 625                 }
 626
 627                 case nir_cf_node_if:
 628                         emit_if(ctx, nir_cf_node_as_if(node));
 629                         break;
 630
 631                 case nir_cf_node_loop:
 632                         emit_loop(ctx, nir_cf_node_as_loop(node));
 633                         break;
 634
 635                 default:
 636                         unreachable("Unknown control flow");
 637                 }
 638         }
 639
 640         return start_block;
 641 }
 642
 643 static int
 644 glsl_type_size(const struct glsl_type *type, bool bindless)
 645 {
 646         return glsl_count_attribute_slots(type, false);
 647 }
 648
 649 static void
 650 bi_optimize_nir(nir_shader *nir)
 651 {
 652         bool progress;
 653         unsigned lower_flrp = 16 | 32 | 64;
 654
 655         NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
 656         NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);
 657
 658         nir_lower_tex_options lower_tex_options = {
 659                 .lower_txs_lod = true,
 660                 .lower_txp = ~0,
 661                 .lower_tex_without_implicit_lod = true,
 662                 .lower_txd = true,
 663         };
 664
 665         NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
 666         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 667         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 668
 669         do {
 670                 progress = false;
 671
 672                 NIR_PASS(progress, nir, nir_lower_var_copies);
 673                 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
 674
 675                 NIR_PASS(progress, nir, nir_copy_prop);
 676                 NIR_PASS(progress, nir, nir_opt_remove_phis);
 677                 NIR_PASS(progress, nir, nir_opt_dce);
 678                 NIR_PASS(progress, nir, nir_opt_dead_cf);
 679                 NIR_PASS(progress, nir, nir_opt_cse);
 680                 NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
 681                 NIR_PASS(progress, nir, nir_opt_algebraic);
 682                 NIR_PASS(progress, nir, nir_opt_constant_folding);
 683
 684                 if (lower_flrp != 0) {
 685                         bool lower_flrp_progress = false;
 686                         NIR_PASS(lower_flrp_progress,
 687                                  nir,
 688                                  nir_lower_flrp,
 689                                  lower_flrp,
 690                                  false /* always_precise */,
 691                                  nir->options->lower_ffma);
 692                         if (lower_flrp_progress) {
 693                                 NIR_PASS(progress, nir,
 694                                          nir_opt_constant_folding);
 695                                 progress = true;
 696                         }
 697
 698                         /* Nothing should rematerialize any flrps, so we only
 699                          * need to do this lowering once.
 700                          */
 701                         lower_flrp = 0;
 702                 }
 703
 704                 NIR_PASS(progress, nir, nir_opt_undef);
 705                 NIR_PASS(progress, nir, nir_opt_loop_unroll,
 706                          nir_var_shader_in |
 707                          nir_var_shader_out |
 708                          nir_var_function_temp);
 709         } while (progress);
 710
 711         NIR_PASS(progress, nir, nir_opt_algebraic_late);
 712         NIR_PASS(progress, nir, bifrost_nir_lower_algebraic_late);
 713         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 714         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 715
 716         /* Take us out of SSA */
 717         NIR_PASS(progress, nir, nir_lower_locals_to_regs);
 718         NIR_PASS(progress, nir, nir_convert_from_ssa, true);
 719
 720         /* We're a primary scalar architecture but there's enough vector that
 721          * we use a vector IR so let's not also deal with scalar hacks on top
 722          * of the vector hacks */
 723
 724         NIR_PASS(progress, nir, nir_move_vec_src_uses_to_dest);
 725         NIR_PASS(progress, nir, nir_lower_vec_to_movs);
 726         NIR_PASS(progress, nir, nir_opt_dce);
 727 }
 728
 729 void
 730 bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned product_id)
 731 {
 732         bi_context *ctx = rzalloc(NULL, bi_context);
 733         ctx->nir = nir;
 734         ctx->stage = nir->info.stage;
 735         ctx->quirks = bifrost_get_quirks(product_id);
 736         list_inithead(&ctx->blocks);
 737
 738         /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
 739          * (so we don't accidentally duplicate the epilogue since mesa/st has
 740          * messed with our I/O quite a bit already) */
 741
 742         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 743
 744         if (ctx->stage == MESA_SHADER_VERTEX) {
 745                 NIR_PASS_V(nir, nir_lower_viewport_transform);
 746                 NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
 747         }
 748
 749         NIR_PASS_V(nir, nir_split_var_copies);
 750         NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 751         NIR_PASS_V(nir, nir_lower_var_copies);
 752         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 753         NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
 754         NIR_PASS_V(nir, nir_lower_ssbo);
 755
 756         bi_optimize_nir(nir);
 757         nir_print_shader(nir, stdout);
 758
 759         nir_foreach_function(func, nir) {
 760                 if (!func->impl)
 761                         continue;
 762
 763                 ctx->impl = func->impl;
 764                 emit_cf_list(ctx, &func->impl->body);
 765                 break; /* TODO: Multi-function shaders */
 766         }
 767
 768         bi_print_shader(ctx, stdout);
 769         bi_schedule(ctx);
 770
 771         ralloc_free(ctx);
 772 }