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         case nir_op_fsub:
 235                 return BI_ADD;
 236         case nir_op_isub:
 237                 return BI_ISUB;
 238
 239         case nir_op_i2i8:
 240         case nir_op_i2i16:
 241         case nir_op_i2i32:
 242         case nir_op_i2i64:
 243         case nir_op_u2u8:
 244         case nir_op_u2u16:
 245         case nir_op_u2u32:
 246         case nir_op_u2u64:
 247         case nir_op_f2i16:
 248         case nir_op_f2i32:
 249         case nir_op_f2i64:
 250         case nir_op_f2u16:
 251         case nir_op_f2u32:
 252         case nir_op_f2u64:
 253         case nir_op_i2f16:
 254         case nir_op_i2f32:
 255         case nir_op_i2f64:
 256         case nir_op_u2f16:
 257         case nir_op_u2f32:
 258         case nir_op_u2f64:
 259                 return BI_CONVERT;
 260
 261         case nir_op_fmul:
 262                 return BI_FMA;
 263
 264         case nir_op_imin:
 265         case nir_op_imax:
 266         case nir_op_umin:
 267         case nir_op_umax:
 268         case nir_op_fmin:
 269         case nir_op_fmax:
 270                 return BI_MINMAX;
 271
 272         case nir_op_fsat:
 273         case nir_op_fneg:
 274         case nir_op_fabs:
 275         case nir_op_mov:
 276                 return BI_MOV;
 277
 278         case nir_op_frcp:
 279         case nir_op_frsq:
 280         case nir_op_fsin:
 281         case nir_op_fcos:
 282                 return BI_SPECIAL;
 283
 284         default:
 285                 unreachable("Unknown ALU op");
 286         }
 287 }
 288
 289 static void
 290 emit_alu(bi_context *ctx, nir_alu_instr *instr)
 291 {
 292         /* Assume it's something we can handle normally */
 293         bi_instruction alu = {
 294                 .type = bi_class_for_nir_alu(instr->op),
 295                 .dest = bir_dest_index(&instr->dest.dest),
 296                 .dest_type = nir_op_infos[instr->op].output_type
 297                         | nir_dest_bit_size(instr->dest.dest),
 298         };
 299
 300         /* TODO: Implement lowering of special functions for older Bifrost */
 301         assert((alu.type != BI_SPECIAL) || !(ctx->quirks & BIFROST_NO_FAST_OP));
 302
 303         if (instr->dest.dest.is_ssa) {
 304                 /* Construct a writemask */
 305                 unsigned bits_per_comp = instr->dest.dest.ssa.bit_size;
 306                 unsigned comps = instr->dest.dest.ssa.num_components;
 307                 assert(comps == 1);
 308                 unsigned bits = bits_per_comp * comps;
 309                 unsigned bytes = MAX2(bits / 8, 1);
 310                 alu.writemask = (1 << bytes) - 1;
 311         } else {
 312                 unsigned comp_mask = instr->dest.write_mask;
 313
 314                 alu.writemask = pan_to_bytemask(nir_dest_bit_size(instr->dest.dest),
 315                                 comp_mask);
 316         }
 317
 318         /* We inline constants as we go. This tracks how many constants have
 319          * been inlined, since we're limited to 64-bits of constants per
 320          * instruction */
 321
 322         unsigned dest_bits = nir_dest_bit_size(instr->dest.dest);
 323         unsigned constants_left = (64 / dest_bits);
 324         unsigned constant_shift = 0;
 325
 326         /* Copy sources */
 327
 328         unsigned num_inputs = nir_op_infos[instr->op].num_inputs;
 329         assert(num_inputs <= ARRAY_SIZE(alu.src));
 330
 331         for (unsigned i = 0; i < num_inputs; ++i) {
 332                 unsigned bits = nir_src_bit_size(instr->src[i].src);
 333                 alu.src_types[i] = nir_op_infos[instr->op].input_types[i]
 334                         | bits;
 335
 336                 /* Try to inline a constant */
 337                 if (nir_src_is_const(instr->src[i].src) && constants_left && (dest_bits == bits)) {
 338                         alu.constant.u64 |=
 339                                 (nir_src_as_uint(instr->src[i].src)) << constant_shift;
 340
 341                         alu.src[i] = BIR_INDEX_CONSTANT | constant_shift;
 342                         --constants_left;
 343                         constant_shift += dest_bits;
 344                         continue;
 345                 }
 346
 347                 alu.src[i] = bir_src_index(&instr->src[i].src);
 348
 349                 /* We assert scalarization above */
 350                 alu.swizzle[i][0] = instr->src[i].swizzle[0];
 351         }
 352
 353         /* Op-specific fixup */
 354         switch (instr->op) {
 355         case nir_op_fmul:
 356                 alu.src[2] = BIR_INDEX_ZERO; /* FMA */
 357                 break;
 358         case nir_op_fsat:
 359                 alu.outmod = BIFROST_SAT; /* MOV */
 360                 break;
 361         case nir_op_fneg:
 362                 alu.src_neg[0] = true; /* MOV */
 363                 break;
 364         case nir_op_fabs:
 365                 alu.src_abs[0] = true; /* MOV */
 366                 break;
 367         case nir_op_fsub:
 368                 alu.src_neg[1] = true; /* ADD */
 369                 break;
 370         case nir_op_fmax:
 371         case nir_op_imax:
 372         case nir_op_umax:
 373                 alu.op.minmax = BI_MINMAX_MAX; /* MINMAX */
 374                 break;
 375         case nir_op_frcp:
 376                 alu.op.special = BI_SPECIAL_FRCP;
 377                 break;
 378         case nir_op_frsq:
 379                 alu.op.special = BI_SPECIAL_FRSQ;
 380                 break;
 381         case nir_op_fsin:
 382                 alu.op.special = BI_SPECIAL_FSIN;
 383                 break;
 384         case nir_op_fcos:
 385                 alu.op.special = BI_SPECIAL_FCOS;
 386                 break;
 387         default:
 388                 break;
 389         }
 390
 391         bi_emit(ctx, alu);
 392 }
 393
 394 static void
 395 emit_instr(bi_context *ctx, struct nir_instr *instr)
 396 {
 397         switch (instr->type) {
 398         case nir_instr_type_load_const:
 399                 emit_load_const(ctx, nir_instr_as_load_const(instr));
 400                 break;
 401
 402         case nir_instr_type_intrinsic:
 403                 emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 404                 break;
 405
 406         case nir_instr_type_alu:
 407                 emit_alu(ctx, nir_instr_as_alu(instr));
 408                 break;
 409
 410 #if 0
 411         case nir_instr_type_tex:
 412                 emit_tex(ctx, nir_instr_as_tex(instr));
 413                 break;
 414 #endif
 415
 416         case nir_instr_type_jump:
 417                 emit_jump(ctx, nir_instr_as_jump(instr));
 418                 break;
 419
 420         case nir_instr_type_ssa_undef:
 421                 /* Spurious */
 422                 break;
 423
 424         default:
 425                 //unreachable("Unhandled instruction type");
 426                 break;
 427         }
 428 }
 429
 430
 431
 432 static bi_block *
 433 create_empty_block(bi_context *ctx)
 434 {
 435         bi_block *blk = rzalloc(ctx, bi_block);
 436
 437         blk->predecessors = _mesa_set_create(blk,
 438                         _mesa_hash_pointer,
 439                         _mesa_key_pointer_equal);
 440
 441         blk->name = ctx->block_name_count++;
 442
 443         return blk;
 444 }
 445
 446 static void
 447 bi_block_add_successor(bi_block *block, bi_block *successor)
 448 {
 449         assert(block);
 450         assert(successor);
 451
 452         for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) {
 453                 if (block->successors[i]) {
 454                        if (block->successors[i] == successor)
 455                                return;
 456                        else
 457                                continue;
 458                 }
 459
 460                 block->successors[i] = successor;
 461                 _mesa_set_add(successor->predecessors, block);
 462                 return;
 463         }
 464
 465         unreachable("Too many successors");
 466 }
 467
 468 static void
 469 bi_schedule_barrier(bi_context *ctx)
 470 {
 471         bi_block *temp = ctx->after_block;
 472         ctx->after_block = create_empty_block(ctx);
 473         list_addtail(&ctx->after_block->link, &ctx->blocks);
 474         list_inithead(&ctx->after_block->instructions);
 475         bi_block_add_successor(ctx->current_block, ctx->after_block);
 476         ctx->current_block = ctx->after_block;
 477         ctx->after_block = temp;
 478 }
 479
 480 static bi_block *
 481 emit_block(bi_context *ctx, nir_block *block)
 482 {
 483         if (ctx->after_block) {
 484                 ctx->current_block = ctx->after_block;
 485                 ctx->after_block = NULL;
 486         } else {
 487                 ctx->current_block = create_empty_block(ctx);
 488         }
 489
 490         list_addtail(&ctx->current_block->link, &ctx->blocks);
 491         list_inithead(&ctx->current_block->instructions);
 492
 493         nir_foreach_instr(instr, block) {
 494                 emit_instr(ctx, instr);
 495                 ++ctx->instruction_count;
 496         }
 497
 498         return ctx->current_block;
 499 }
 500
 501 /* Emits an unconditional branch to the end of the current block, returning a
 502  * pointer so the user can fill in details */
 503
 504 static bi_instruction *
 505 bi_emit_branch(bi_context *ctx)
 506 {
 507         bi_instruction branch = {
 508                 .type = BI_BRANCH,
 509                 .branch = {
 510                         .cond = BI_COND_ALWAYS
 511                 }
 512         };
 513
 514         return bi_emit(ctx, branch);
 515 }
 516
 517 /* Sets a condition for a branch by examing the NIR condition. If we're
 518  * familiar with the condition, we unwrap it to fold it into the branch
 519  * instruction. Otherwise, we consume the condition directly. We
 520  * generally use 1-bit booleans which allows us to use small types for
 521  * the conditions.
 522  */
 523
 524 static void
 525 bi_set_branch_cond(bi_instruction *branch, nir_src *cond, bool invert)
 526 {
 527         /* TODO: Try to unwrap instead of always bailing */
 528         branch->src[0] = bir_src_index(cond);
 529         branch->src[1] = BIR_INDEX_ZERO;
 530         branch->src_types[0] = branch->src_types[1] = nir_type_uint16;
 531         branch->branch.cond = invert ? BI_COND_EQ : BI_COND_NE;
 532 }
 533
 534 static void
 535 emit_if(bi_context *ctx, nir_if *nif)
 536 {
 537         bi_block *before_block = ctx->current_block;
 538
 539         /* Speculatively emit the branch, but we can't fill it in until later */
 540         bi_instruction *then_branch = bi_emit_branch(ctx);
 541         bi_set_branch_cond(then_branch, &nif->condition, true);
 542
 543         /* Emit the two subblocks. */
 544         bi_block *then_block = emit_cf_list(ctx, &nif->then_list);
 545         bi_block *end_then_block = ctx->current_block;
 546
 547         /* Emit a jump from the end of the then block to the end of the else */
 548         bi_instruction *then_exit = bi_emit_branch(ctx);
 549
 550         /* Emit second block, and check if it's empty */
 551
 552         int count_in = ctx->instruction_count;
 553         bi_block *else_block = emit_cf_list(ctx, &nif->else_list);
 554         bi_block *end_else_block = ctx->current_block;
 555         ctx->after_block = create_empty_block(ctx);
 556
 557         /* Now that we have the subblocks emitted, fix up the branches */
 558
 559         assert(then_block);
 560         assert(else_block);
 561
 562         if (ctx->instruction_count == count_in) {
 563                 /* The else block is empty, so don't emit an exit jump */
 564                 bi_remove_instruction(then_exit);
 565                 then_branch->branch.target = ctx->after_block;
 566         } else {
 567                 then_branch->branch.target = else_block;
 568                 then_exit->branch.target = ctx->after_block;
 569                 bi_block_add_successor(end_then_block, then_exit->branch.target);
 570         }
 571
 572         /* Wire up the successors */
 573
 574         bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */
 575
 576         bi_block_add_successor(before_block, then_block); /* fallthrough */
 577         bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
 578 }
 579
 580 static void
 581 emit_loop(bi_context *ctx, nir_loop *nloop)
 582 {
 583         /* Remember where we are */
 584         bi_block *start_block = ctx->current_block;
 585
 586         bi_block *saved_break = ctx->break_block;
 587         bi_block *saved_continue = ctx->continue_block;
 588
 589         ctx->continue_block = create_empty_block(ctx);
 590         ctx->break_block = create_empty_block(ctx);
 591         ctx->after_block = ctx->continue_block;
 592
 593         /* Emit the body itself */
 594         emit_cf_list(ctx, &nloop->body);
 595
 596         /* Branch back to loop back */
 597         bi_instruction *br_back = bi_emit_branch(ctx);
 598         br_back->branch.target = ctx->continue_block;
 599         bi_block_add_successor(start_block, ctx->continue_block);
 600         bi_block_add_successor(ctx->current_block, ctx->continue_block);
 601
 602         ctx->after_block = ctx->break_block;
 603
 604         /* Pop off */
 605         ctx->break_block = saved_break;
 606         ctx->continue_block = saved_continue;
 607         ++ctx->loop_count;
 608 }
 609
 610 static bi_block *
 611 emit_cf_list(bi_context *ctx, struct exec_list *list)
 612 {
 613         bi_block *start_block = NULL;
 614
 615         foreach_list_typed(nir_cf_node, node, node, list) {
 616                 switch (node->type) {
 617                 case nir_cf_node_block: {
 618                         bi_block *block = emit_block(ctx, nir_cf_node_as_block(node));
 619
 620                         if (!start_block)
 621                                 start_block = block;
 622
 623                         break;
 624                 }
 625
 626                 case nir_cf_node_if:
 627                         emit_if(ctx, nir_cf_node_as_if(node));
 628                         break;
 629
 630                 case nir_cf_node_loop:
 631                         emit_loop(ctx, nir_cf_node_as_loop(node));
 632                         break;
 633
 634                 default:
 635                         unreachable("Unknown control flow");
 636                 }
 637         }
 638
 639         return start_block;
 640 }
 641
 642 static int
 643 glsl_type_size(const struct glsl_type *type, bool bindless)
 644 {
 645         return glsl_count_attribute_slots(type, false);
 646 }
 647
 648 static void
 649 bi_optimize_nir(nir_shader *nir)
 650 {
 651         bool progress;
 652         unsigned lower_flrp = 16 | 32 | 64;
 653
 654         NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
 655         NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);
 656
 657         nir_lower_tex_options lower_tex_options = {
 658                 .lower_txs_lod = true,
 659                 .lower_txp = ~0,
 660                 .lower_tex_without_implicit_lod = true,
 661                 .lower_txd = true,
 662         };
 663
 664         NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
 665         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 666         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 667
 668         do {
 669                 progress = false;
 670
 671                 NIR_PASS(progress, nir, nir_lower_var_copies);
 672                 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
 673
 674                 NIR_PASS(progress, nir, nir_copy_prop);
 675                 NIR_PASS(progress, nir, nir_opt_remove_phis);
 676                 NIR_PASS(progress, nir, nir_opt_dce);
 677                 NIR_PASS(progress, nir, nir_opt_dead_cf);
 678                 NIR_PASS(progress, nir, nir_opt_cse);
 679                 NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
 680                 NIR_PASS(progress, nir, nir_opt_algebraic);
 681                 NIR_PASS(progress, nir, nir_opt_constant_folding);
 682
 683                 if (lower_flrp != 0) {
 684                         bool lower_flrp_progress = false;
 685                         NIR_PASS(lower_flrp_progress,
 686                                  nir,
 687                                  nir_lower_flrp,
 688                                  lower_flrp,
 689                                  false /* always_precise */,
 690                                  nir->options->lower_ffma);
 691                         if (lower_flrp_progress) {
 692                                 NIR_PASS(progress, nir,
 693                                          nir_opt_constant_folding);
 694                                 progress = true;
 695                         }
 696
 697                         /* Nothing should rematerialize any flrps, so we only
 698                          * need to do this lowering once.
 699                          */
 700                         lower_flrp = 0;
 701                 }
 702
 703                 NIR_PASS(progress, nir, nir_opt_undef);
 704                 NIR_PASS(progress, nir, nir_opt_loop_unroll,
 705                          nir_var_shader_in |
 706                          nir_var_shader_out |
 707                          nir_var_function_temp);
 708         } while (progress);
 709
 710         NIR_PASS(progress, nir, nir_opt_algebraic_late);
 711         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 712         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 713
 714         /* Take us out of SSA */
 715         NIR_PASS(progress, nir, nir_lower_locals_to_regs);
 716         NIR_PASS(progress, nir, nir_convert_from_ssa, true);
 717
 718         /* We're a primary scalar architecture but there's enough vector that
 719          * we use a vector IR so let's not also deal with scalar hacks on top
 720          * of the vector hacks */
 721
 722         NIR_PASS(progress, nir, nir_move_vec_src_uses_to_dest);
 723         NIR_PASS(progress, nir, nir_lower_vec_to_movs);
 724         NIR_PASS(progress, nir, nir_opt_dce);
 725 }
 726
 727 void
 728 bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned product_id)
 729 {
 730         bi_context *ctx = rzalloc(NULL, bi_context);
 731         ctx->nir = nir;
 732         ctx->stage = nir->info.stage;
 733         ctx->quirks = bifrost_get_quirks(product_id);
 734         list_inithead(&ctx->blocks);
 735
 736         /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
 737          * (so we don't accidentally duplicate the epilogue since mesa/st has
 738          * messed with our I/O quite a bit already) */
 739
 740         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 741
 742         if (ctx->stage == MESA_SHADER_VERTEX) {
 743                 NIR_PASS_V(nir, nir_lower_viewport_transform);
 744                 NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
 745         }
 746
 747         NIR_PASS_V(nir, nir_split_var_copies);
 748         NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 749         NIR_PASS_V(nir, nir_lower_var_copies);
 750         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 751         NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
 752         NIR_PASS_V(nir, nir_lower_ssbo);
 753
 754         bi_optimize_nir(nir);
 755         nir_print_shader(nir, stdout);
 756
 757         nir_foreach_function(func, nir) {
 758                 if (!func->impl)
 759                         continue;
 760
 761                 ctx->impl = func->impl;
 762                 emit_cf_list(ctx, &func->impl->body);
 763                 break; /* TODO: Multi-function shaders */
 764         }
 765
 766         bi_print_shader(ctx, stdout);
 767         bi_schedule(ctx);
 768
 769         ralloc_free(ctx);
 770 }