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