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_bcsel:
 241                 return BI_CSEL;
 242
 243         case nir_op_i2i8:
 244         case nir_op_i2i16:
 245         case nir_op_i2i32:
 246         case nir_op_i2i64:
 247         case nir_op_u2u8:
 248         case nir_op_u2u16:
 249         case nir_op_u2u32:
 250         case nir_op_u2u64:
 251         case nir_op_f2i16:
 252         case nir_op_f2i32:
 253         case nir_op_f2i64:
 254         case nir_op_f2u16:
 255         case nir_op_f2u32:
 256         case nir_op_f2u64:
 257         case nir_op_i2f16:
 258         case nir_op_i2f32:
 259         case nir_op_i2f64:
 260         case nir_op_u2f16:
 261         case nir_op_u2f32:
 262         case nir_op_u2f64:
 263                 return BI_CONVERT;
 264
 265         case nir_op_fmul:
 266                 return BI_FMA;
 267
 268         case nir_op_imin:
 269         case nir_op_imax:
 270         case nir_op_umin:
 271         case nir_op_umax:
 272         case nir_op_fmin:
 273         case nir_op_fmax:
 274                 return BI_MINMAX;
 275
 276         case nir_op_fsat:
 277         case nir_op_fneg:
 278         case nir_op_fabs:
 279         case nir_op_mov:
 280                 return BI_MOV;
 281
 282         case nir_op_frcp:
 283         case nir_op_frsq:
 284         case nir_op_fsin:
 285         case nir_op_fcos:
 286                 return BI_SPECIAL;
 287
 288         default:
 289                 unreachable("Unknown ALU op");
 290         }
 291 }
 292
 293 static void
 294 emit_alu(bi_context *ctx, nir_alu_instr *instr)
 295 {
 296         /* Assume it's something we can handle normally */
 297         bi_instruction alu = {
 298                 .type = bi_class_for_nir_alu(instr->op),
 299                 .dest = bir_dest_index(&instr->dest.dest),
 300                 .dest_type = nir_op_infos[instr->op].output_type
 301                         | nir_dest_bit_size(instr->dest.dest),
 302         };
 303
 304         /* TODO: Implement lowering of special functions for older Bifrost */
 305         assert((alu.type != BI_SPECIAL) || !(ctx->quirks & BIFROST_NO_FAST_OP));
 306
 307         if (instr->dest.dest.is_ssa) {
 308                 /* Construct a writemask */
 309                 unsigned bits_per_comp = instr->dest.dest.ssa.bit_size;
 310                 unsigned comps = instr->dest.dest.ssa.num_components;
 311                 assert(comps == 1);
 312                 unsigned bits = bits_per_comp * comps;
 313                 unsigned bytes = MAX2(bits / 8, 1);
 314                 alu.writemask = (1 << bytes) - 1;
 315         } else {
 316                 unsigned comp_mask = instr->dest.write_mask;
 317
 318                 alu.writemask = pan_to_bytemask(nir_dest_bit_size(instr->dest.dest),
 319                                 comp_mask);
 320         }
 321
 322         /* We inline constants as we go. This tracks how many constants have
 323          * been inlined, since we're limited to 64-bits of constants per
 324          * instruction */
 325
 326         unsigned dest_bits = nir_dest_bit_size(instr->dest.dest);
 327         unsigned constants_left = (64 / dest_bits);
 328         unsigned constant_shift = 0;
 329
 330         /* Copy sources */
 331
 332         unsigned num_inputs = nir_op_infos[instr->op].num_inputs;
 333         assert(num_inputs <= ARRAY_SIZE(alu.src));
 334
 335         for (unsigned i = 0; i < num_inputs; ++i) {
 336                 unsigned bits = nir_src_bit_size(instr->src[i].src);
 337                 alu.src_types[i] = nir_op_infos[instr->op].input_types[i]
 338                         | bits;
 339
 340                 /* Try to inline a constant */
 341                 if (nir_src_is_const(instr->src[i].src) && constants_left && (dest_bits == bits)) {
 342                         alu.constant.u64 |=
 343                                 (nir_src_as_uint(instr->src[i].src)) << constant_shift;
 344
 345                         alu.src[i] = BIR_INDEX_CONSTANT | constant_shift;
 346                         --constants_left;
 347                         constant_shift += dest_bits;
 348                         continue;
 349                 }
 350
 351                 alu.src[i] = bir_src_index(&instr->src[i].src);
 352
 353                 /* We assert scalarization above */
 354                 alu.swizzle[i][0] = instr->src[i].swizzle[0];
 355         }
 356
 357         /* Op-specific fixup */
 358         switch (instr->op) {
 359         case nir_op_fmul:
 360                 alu.src[2] = BIR_INDEX_ZERO; /* FMA */
 361                 break;
 362         case nir_op_fsat:
 363                 alu.outmod = BIFROST_SAT; /* MOV */
 364                 break;
 365         case nir_op_fneg:
 366                 alu.src_neg[0] = true; /* MOV */
 367                 break;
 368         case nir_op_fabs:
 369                 alu.src_abs[0] = true; /* MOV */
 370                 break;
 371         case nir_op_fsub:
 372                 alu.src_neg[1] = true; /* ADD */
 373                 break;
 374         case nir_op_fmax:
 375         case nir_op_imax:
 376         case nir_op_umax:
 377                 alu.op.minmax = BI_MINMAX_MAX; /* MINMAX */
 378                 break;
 379         case nir_op_frcp:
 380                 alu.op.special = BI_SPECIAL_FRCP;
 381                 break;
 382         case nir_op_frsq:
 383                 alu.op.special = BI_SPECIAL_FRSQ;
 384                 break;
 385         case nir_op_fsin:
 386                 alu.op.special = BI_SPECIAL_FSIN;
 387                 break;
 388         case nir_op_fcos:
 389                 alu.op.special = BI_SPECIAL_FCOS;
 390                 break;
 391         default:
 392                 break;
 393         }
 394
 395         bi_emit(ctx, alu);
 396 }
 397
 398 static void
 399 emit_instr(bi_context *ctx, struct nir_instr *instr)
 400 {
 401         switch (instr->type) {
 402         case nir_instr_type_load_const:
 403                 emit_load_const(ctx, nir_instr_as_load_const(instr));
 404                 break;
 405
 406         case nir_instr_type_intrinsic:
 407                 emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
 408                 break;
 409
 410         case nir_instr_type_alu:
 411                 emit_alu(ctx, nir_instr_as_alu(instr));
 412                 break;
 413
 414 #if 0
 415         case nir_instr_type_tex:
 416                 emit_tex(ctx, nir_instr_as_tex(instr));
 417                 break;
 418 #endif
 419
 420         case nir_instr_type_jump:
 421                 emit_jump(ctx, nir_instr_as_jump(instr));
 422                 break;
 423
 424         case nir_instr_type_ssa_undef:
 425                 /* Spurious */
 426                 break;
 427
 428         default:
 429                 //unreachable("Unhandled instruction type");
 430                 break;
 431         }
 432 }
 433
 434
 435
 436 static bi_block *
 437 create_empty_block(bi_context *ctx)
 438 {
 439         bi_block *blk = rzalloc(ctx, bi_block);
 440
 441         blk->predecessors = _mesa_set_create(blk,
 442                         _mesa_hash_pointer,
 443                         _mesa_key_pointer_equal);
 444
 445         blk->name = ctx->block_name_count++;
 446
 447         return blk;
 448 }
 449
 450 static void
 451 bi_block_add_successor(bi_block *block, bi_block *successor)
 452 {
 453         assert(block);
 454         assert(successor);
 455
 456         for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) {
 457                 if (block->successors[i]) {
 458                        if (block->successors[i] == successor)
 459                                return;
 460                        else
 461                                continue;
 462                 }
 463
 464                 block->successors[i] = successor;
 465                 _mesa_set_add(successor->predecessors, block);
 466                 return;
 467         }
 468
 469         unreachable("Too many successors");
 470 }
 471
 472 static void
 473 bi_schedule_barrier(bi_context *ctx)
 474 {
 475         bi_block *temp = ctx->after_block;
 476         ctx->after_block = create_empty_block(ctx);
 477         list_addtail(&ctx->after_block->link, &ctx->blocks);
 478         list_inithead(&ctx->after_block->instructions);
 479         bi_block_add_successor(ctx->current_block, ctx->after_block);
 480         ctx->current_block = ctx->after_block;
 481         ctx->after_block = temp;
 482 }
 483
 484 static bi_block *
 485 emit_block(bi_context *ctx, nir_block *block)
 486 {
 487         if (ctx->after_block) {
 488                 ctx->current_block = ctx->after_block;
 489                 ctx->after_block = NULL;
 490         } else {
 491                 ctx->current_block = create_empty_block(ctx);
 492         }
 493
 494         list_addtail(&ctx->current_block->link, &ctx->blocks);
 495         list_inithead(&ctx->current_block->instructions);
 496
 497         nir_foreach_instr(instr, block) {
 498                 emit_instr(ctx, instr);
 499                 ++ctx->instruction_count;
 500         }
 501
 502         return ctx->current_block;
 503 }
 504
 505 /* Emits an unconditional branch to the end of the current block, returning a
 506  * pointer so the user can fill in details */
 507
 508 static bi_instruction *
 509 bi_emit_branch(bi_context *ctx)
 510 {
 511         bi_instruction branch = {
 512                 .type = BI_BRANCH,
 513                 .branch = {
 514                         .cond = BI_COND_ALWAYS
 515                 }
 516         };
 517
 518         return bi_emit(ctx, branch);
 519 }
 520
 521 /* Sets a condition for a branch by examing the NIR condition. If we're
 522  * familiar with the condition, we unwrap it to fold it into the branch
 523  * instruction. Otherwise, we consume the condition directly. We
 524  * generally use 1-bit booleans which allows us to use small types for
 525  * the conditions.
 526  */
 527
 528 static void
 529 bi_set_branch_cond(bi_instruction *branch, nir_src *cond, bool invert)
 530 {
 531         /* TODO: Try to unwrap instead of always bailing */
 532         branch->src[0] = bir_src_index(cond);
 533         branch->src[1] = BIR_INDEX_ZERO;
 534         branch->src_types[0] = branch->src_types[1] = nir_type_uint16;
 535         branch->branch.cond = invert ? BI_COND_EQ : BI_COND_NE;
 536 }
 537
 538 static void
 539 emit_if(bi_context *ctx, nir_if *nif)
 540 {
 541         bi_block *before_block = ctx->current_block;
 542
 543         /* Speculatively emit the branch, but we can't fill it in until later */
 544         bi_instruction *then_branch = bi_emit_branch(ctx);
 545         bi_set_branch_cond(then_branch, &nif->condition, true);
 546
 547         /* Emit the two subblocks. */
 548         bi_block *then_block = emit_cf_list(ctx, &nif->then_list);
 549         bi_block *end_then_block = ctx->current_block;
 550
 551         /* Emit a jump from the end of the then block to the end of the else */
 552         bi_instruction *then_exit = bi_emit_branch(ctx);
 553
 554         /* Emit second block, and check if it's empty */
 555
 556         int count_in = ctx->instruction_count;
 557         bi_block *else_block = emit_cf_list(ctx, &nif->else_list);
 558         bi_block *end_else_block = ctx->current_block;
 559         ctx->after_block = create_empty_block(ctx);
 560
 561         /* Now that we have the subblocks emitted, fix up the branches */
 562
 563         assert(then_block);
 564         assert(else_block);
 565
 566         if (ctx->instruction_count == count_in) {
 567                 /* The else block is empty, so don't emit an exit jump */
 568                 bi_remove_instruction(then_exit);
 569                 then_branch->branch.target = ctx->after_block;
 570         } else {
 571                 then_branch->branch.target = else_block;
 572                 then_exit->branch.target = ctx->after_block;
 573                 bi_block_add_successor(end_then_block, then_exit->branch.target);
 574         }
 575
 576         /* Wire up the successors */
 577
 578         bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */
 579
 580         bi_block_add_successor(before_block, then_block); /* fallthrough */
 581         bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */
 582 }
 583
 584 static void
 585 emit_loop(bi_context *ctx, nir_loop *nloop)
 586 {
 587         /* Remember where we are */
 588         bi_block *start_block = ctx->current_block;
 589
 590         bi_block *saved_break = ctx->break_block;
 591         bi_block *saved_continue = ctx->continue_block;
 592
 593         ctx->continue_block = create_empty_block(ctx);
 594         ctx->break_block = create_empty_block(ctx);
 595         ctx->after_block = ctx->continue_block;
 596
 597         /* Emit the body itself */
 598         emit_cf_list(ctx, &nloop->body);
 599
 600         /* Branch back to loop back */
 601         bi_instruction *br_back = bi_emit_branch(ctx);
 602         br_back->branch.target = ctx->continue_block;
 603         bi_block_add_successor(start_block, ctx->continue_block);
 604         bi_block_add_successor(ctx->current_block, ctx->continue_block);
 605
 606         ctx->after_block = ctx->break_block;
 607
 608         /* Pop off */
 609         ctx->break_block = saved_break;
 610         ctx->continue_block = saved_continue;
 611         ++ctx->loop_count;
 612 }
 613
 614 static bi_block *
 615 emit_cf_list(bi_context *ctx, struct exec_list *list)
 616 {
 617         bi_block *start_block = NULL;
 618
 619         foreach_list_typed(nir_cf_node, node, node, list) {
 620                 switch (node->type) {
 621                 case nir_cf_node_block: {
 622                         bi_block *block = emit_block(ctx, nir_cf_node_as_block(node));
 623
 624                         if (!start_block)
 625                                 start_block = block;
 626
 627                         break;
 628                 }
 629
 630                 case nir_cf_node_if:
 631                         emit_if(ctx, nir_cf_node_as_if(node));
 632                         break;
 633
 634                 case nir_cf_node_loop:
 635                         emit_loop(ctx, nir_cf_node_as_loop(node));
 636                         break;
 637
 638                 default:
 639                         unreachable("Unknown control flow");
 640                 }
 641         }
 642
 643         return start_block;
 644 }
 645
 646 static int
 647 glsl_type_size(const struct glsl_type *type, bool bindless)
 648 {
 649         return glsl_count_attribute_slots(type, false);
 650 }
 651
 652 static void
 653 bi_optimize_nir(nir_shader *nir)
 654 {
 655         bool progress;
 656         unsigned lower_flrp = 16 | 32 | 64;
 657
 658         NIR_PASS(progress, nir, nir_lower_regs_to_ssa);
 659         NIR_PASS(progress, nir, nir_lower_idiv, nir_lower_idiv_fast);
 660
 661         nir_lower_tex_options lower_tex_options = {
 662                 .lower_txs_lod = true,
 663                 .lower_txp = ~0,
 664                 .lower_tex_without_implicit_lod = true,
 665                 .lower_txd = true,
 666         };
 667
 668         NIR_PASS(progress, nir, nir_lower_tex, &lower_tex_options);
 669         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 670         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 671
 672         do {
 673                 progress = false;
 674
 675                 NIR_PASS(progress, nir, nir_lower_var_copies);
 676                 NIR_PASS(progress, nir, nir_lower_vars_to_ssa);
 677
 678                 NIR_PASS(progress, nir, nir_copy_prop);
 679                 NIR_PASS(progress, nir, nir_opt_remove_phis);
 680                 NIR_PASS(progress, nir, nir_opt_dce);
 681                 NIR_PASS(progress, nir, nir_opt_dead_cf);
 682                 NIR_PASS(progress, nir, nir_opt_cse);
 683                 NIR_PASS(progress, nir, nir_opt_peephole_select, 64, false, true);
 684                 NIR_PASS(progress, nir, nir_opt_algebraic);
 685                 NIR_PASS(progress, nir, nir_opt_constant_folding);
 686
 687                 if (lower_flrp != 0) {
 688                         bool lower_flrp_progress = false;
 689                         NIR_PASS(lower_flrp_progress,
 690                                  nir,
 691                                  nir_lower_flrp,
 692                                  lower_flrp,
 693                                  false /* always_precise */,
 694                                  nir->options->lower_ffma);
 695                         if (lower_flrp_progress) {
 696                                 NIR_PASS(progress, nir,
 697                                          nir_opt_constant_folding);
 698                                 progress = true;
 699                         }
 700
 701                         /* Nothing should rematerialize any flrps, so we only
 702                          * need to do this lowering once.
 703                          */
 704                         lower_flrp = 0;
 705                 }
 706
 707                 NIR_PASS(progress, nir, nir_opt_undef);
 708                 NIR_PASS(progress, nir, nir_opt_loop_unroll,
 709                          nir_var_shader_in |
 710                          nir_var_shader_out |
 711                          nir_var_function_temp);
 712         } while (progress);
 713
 714         NIR_PASS(progress, nir, nir_opt_algebraic_late);
 715         NIR_PASS(progress, nir, bifrost_nir_lower_algebraic_late);
 716         NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL);
 717         NIR_PASS(progress, nir, nir_lower_load_const_to_scalar);
 718
 719         /* Take us out of SSA */
 720         NIR_PASS(progress, nir, nir_lower_locals_to_regs);
 721         NIR_PASS(progress, nir, nir_convert_from_ssa, true);
 722
 723         /* We're a primary scalar architecture but there's enough vector that
 724          * we use a vector IR so let's not also deal with scalar hacks on top
 725          * of the vector hacks */
 726
 727         NIR_PASS(progress, nir, nir_move_vec_src_uses_to_dest);
 728         NIR_PASS(progress, nir, nir_lower_vec_to_movs);
 729         NIR_PASS(progress, nir, nir_opt_dce);
 730 }
 731
 732 void
 733 bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned product_id)
 734 {
 735         bi_context *ctx = rzalloc(NULL, bi_context);
 736         ctx->nir = nir;
 737         ctx->stage = nir->info.stage;
 738         ctx->quirks = bifrost_get_quirks(product_id);
 739         list_inithead(&ctx->blocks);
 740
 741         /* Lower gl_Position pre-optimisation, but after lowering vars to ssa
 742          * (so we don't accidentally duplicate the epilogue since mesa/st has
 743          * messed with our I/O quite a bit already) */
 744
 745         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 746
 747         if (ctx->stage == MESA_SHADER_VERTEX) {
 748                 NIR_PASS_V(nir, nir_lower_viewport_transform);
 749                 NIR_PASS_V(nir, nir_lower_point_size, 1.0, 1024.0);
 750         }
 751
 752         NIR_PASS_V(nir, nir_split_var_copies);
 753         NIR_PASS_V(nir, nir_lower_global_vars_to_local);
 754         NIR_PASS_V(nir, nir_lower_var_copies);
 755         NIR_PASS_V(nir, nir_lower_vars_to_ssa);
 756         NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
 757         NIR_PASS_V(nir, nir_lower_ssbo);
 758
 759         bi_optimize_nir(nir);
 760         nir_print_shader(nir, stdout);
 761
 762         nir_foreach_function(func, nir) {
 763                 if (!func->impl)
 764                         continue;
 765
 766                 ctx->impl = func->impl;
 767                 emit_cf_list(ctx, &func->impl->body);
 768                 break; /* TODO: Multi-function shaders */
 769         }
 770
 771         bi_print_shader(ctx, stdout);
 772         bi_schedule(ctx);
 773
 774         ralloc_free(ctx);
 775 }