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