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