src/panfrost/midgard/compiler.h

   1 /*
   2  * Copyright (C) 2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
   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
  24 #ifndef _MDG_COMPILER_H
  25 #define _MDG_COMPILER_H
  26
  27 #include "midgard.h"
  28 #include "helpers.h"
  29 #include "midgard_compile.h"
  30 #include "midgard_ops.h"
  31 #include "lcra.h"
  32
  33 #include "util/hash_table.h"
  34 #include "util/u_dynarray.h"
  35 #include "util/set.h"
  36 #include "util/list.h"
  37
  38 #include "main/mtypes.h"
  39 #include "compiler/nir_types.h"
  40 #include "compiler/nir/nir.h"
  41 #include "panfrost/util/pan_ir.h"
  42
  43 /* Forward declare */
  44 struct midgard_block;
  45
  46 /* Target types. Defaults to TARGET_GOTO (the type corresponding directly to
  47  * the hardware), hence why that must be zero. TARGET_DISCARD signals this
  48  * instruction is actually a discard op. */
  49
  50 #define TARGET_GOTO 0
  51 #define TARGET_BREAK 1
  52 #define TARGET_CONTINUE 2
  53 #define TARGET_DISCARD 3
  54
  55 typedef struct midgard_branch {
  56         /* If conditional, the condition is specified in r31.w */
  57         bool conditional;
  58
  59         /* For conditionals, if this is true, we branch on FALSE. If false, we  branch on TRUE. */
  60         bool invert_conditional;
  61
  62         /* Branch targets: the start of a block, the start of a loop (continue), the end of a loop (break). Value is one of TARGET_ */
  63         unsigned target_type;
  64
  65         /* The actual target */
  66         union {
  67                 int target_block;
  68                 int target_break;
  69                 int target_continue;
  70         };
  71 } midgard_branch;
  72
  73 /* Generic in-memory data type repesenting a single logical instruction, rather
  74  * than a single instruction group. This is the preferred form for code gen.
  75  * Multiple midgard_insturctions will later be combined during scheduling,
  76  * though this is not represented in this structure.  Its format bridges
  77  * the low-level binary representation with the higher level semantic meaning.
  78  *
  79  * Notably, it allows registers to be specified as block local SSA, for code
  80  * emitted before the register allocation pass.
  81  */
  82
  83 #define MIR_SRC_COUNT 4
  84 #define MIR_VEC_COMPONENTS 16
  85
  86 typedef struct midgard_instruction {
  87         /* Must be first for casting */
  88         struct list_head link;
  89
  90         unsigned type; /* ALU, load/store, texture */
  91
  92         /* Instruction arguments represented as block-local SSA
  93          * indices, rather than registers. ~0 means unused. */
  94         unsigned src[MIR_SRC_COUNT];
  95         unsigned dest;
  96
  97         /* vec16 swizzle, unpacked, per source */
  98         unsigned swizzle[MIR_SRC_COUNT][MIR_VEC_COMPONENTS];
  99
 100         /* Special fields for an ALU instruction */
 101         midgard_reg_info registers;
 102
 103         /* I.e. (1 << alu_bit) */
 104         int unit;
 105
 106         bool has_constants;
 107         midgard_constants constants;
 108         uint16_t inline_constant;
 109         bool has_blend_constant;
 110         bool has_inline_constant;
 111
 112         bool compact_branch;
 113         bool writeout;
 114         bool writeout_depth;
 115         bool writeout_stencil;
 116         bool last_writeout;
 117
 118         /* Kind of a hack, but hint against aggressive DCE */
 119         bool dont_eliminate;
 120
 121         /* Masks in a saneish format. One bit per channel, not packed fancy.
 122          * Use this instead of the op specific ones, and switch over at emit
 123          * time */
 124
 125         uint16_t mask;
 126
 127         /* For ALU ops only: set to true to invert (bitwise NOT) the
 128          * destination of an integer-out op. Not implemented in hardware but
 129          * allows more optimizations */
 130
 131         bool invert;
 132
 133         /* Hint for the register allocator not to spill the destination written
 134          * from this instruction (because it is a spill/unspill node itself).
 135          * Bitmask of spilled classes */
 136
 137         unsigned no_spill;
 138
 139         /* Generic hint for intra-pass use */
 140         bool hint;
 141
 142         /* During scheduling, the backwards dependency graph
 143          * (DAG). nr_dependencies is the number of unscheduled
 144          * instructions that must still be scheduled after
 145          * (before) this instruction. dependents are which
 146          * instructions need to be scheduled before (after) this
 147          * instruction. */
 148
 149         unsigned nr_dependencies;
 150         BITSET_WORD *dependents;
 151
 152         /* For load/store ops.. force 64-bit destination */
 153         bool load_64;
 154
 155         union {
 156                 midgard_load_store_word load_store;
 157                 midgard_vector_alu alu;
 158                 midgard_texture_word texture;
 159                 midgard_branch_extended branch_extended;
 160                 uint16_t br_compact;
 161
 162                 /* General branch, rather than packed br_compact. Higher level
 163                  * than the other components */
 164                 midgard_branch branch;
 165         };
 166 } midgard_instruction;
 167
 168 typedef struct midgard_block {
 169         /* Link to next block. Must be first for mir_get_block */
 170         struct list_head link;
 171
 172         /* List of midgard_instructions emitted for the current block */
 173         struct list_head instructions;
 174
 175         /* Index of the block in source order */
 176         unsigned source_id;
 177
 178         bool is_scheduled;
 179
 180         /* List of midgard_bundles emitted (after the scheduler has run) */
 181         struct util_dynarray bundles;
 182
 183         /* Number of quadwords _actually_ emitted, as determined after scheduling */
 184         unsigned quadword_count;
 185
 186         /* Succeeding blocks. The compiler should not necessarily rely on
 187          * source-order traversal */
 188         struct midgard_block *successors[2];
 189         unsigned nr_successors;
 190
 191         struct set *predecessors;
 192
 193         /* The successors pointer form a graph, and in the case of
 194          * complex control flow, this graph has a cycles. To aid
 195          * traversal during liveness analysis, we have a visited?
 196          * boolean for passes to use as they see fit, provided they
 197          * clean up later */
 198         bool visited;
 199
 200         /* In liveness analysis, these are live masks (per-component) for
 201          * indices for the block. Scalar compilers have the luxury of using
 202          * simple bit fields, but for us, liveness is a vector idea. */
 203         uint16_t *live_in;
 204         uint16_t *live_out;
 205
 206         /* Indicates this is a fixed-function fragment epilogue block */
 207         bool epilogue;
 208 } midgard_block;
 209
 210 typedef struct midgard_bundle {
 211         /* Tag for the overall bundle */
 212         int tag;
 213
 214         /* Instructions contained by the bundle. instruction_count <= 6 (vmul,
 215          * sadd, vadd, smul, vlut, branch) */
 216         int instruction_count;
 217         midgard_instruction *instructions[6];
 218
 219         /* Bundle-wide ALU configuration */
 220         int padding;
 221         int control;
 222         bool has_embedded_constants;
 223         midgard_constants constants;
 224         bool has_blend_constant;
 225         bool last_writeout;
 226 } midgard_bundle;
 227
 228 enum midgard_rt_id {
 229         MIDGARD_COLOR_RT0,
 230         MIDGARD_COLOR_RT1,
 231         MIDGARD_COLOR_RT2,
 232         MIDGARD_COLOR_RT3,
 233         MIDGARD_ZS_RT,
 234         MIDGARD_NUM_RTS,
 235 };
 236
 237 typedef struct compiler_context {
 238         nir_shader *nir;
 239         gl_shader_stage stage;
 240
 241         /* Is internally a blend shader? Depends on stage == FRAGMENT */
 242         bool is_blend;
 243
 244         /* Render target number for a keyed blend shader. Depends on is_blend */
 245         unsigned blend_rt;
 246
 247         /* Tracking for blend constant patching */
 248         int blend_constant_offset;
 249
 250         /* Number of bytes used for Thread Local Storage */
 251         unsigned tls_size;
 252
 253         /* Count of spills and fills for shaderdb */
 254         unsigned spills;
 255         unsigned fills;
 256
 257         /* Current NIR function */
 258         nir_function *func;
 259
 260         /* Allocated compiler temporary counter */
 261         unsigned temp_alloc;
 262
 263         /* Unordered list of midgard_blocks */
 264         int block_count;
 265         struct list_head blocks;
 266
 267         /* TODO merge with block_count? */
 268         unsigned block_source_count;
 269
 270         /* List of midgard_instructions emitted for the current block */
 271         midgard_block *current_block;
 272
 273         /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
 274         midgard_block *after_block;
 275
 276         /* The current "depth" of the loop, for disambiguating breaks/continues
 277          * when using nested loops */
 278         int current_loop_depth;
 279
 280         /* Total number of loops for shader-db */
 281         unsigned loop_count;
 282
 283         /* Constants which have been loaded, for later inlining */
 284         struct hash_table_u64 *ssa_constants;
 285
 286         /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
 287         struct hash_table_u64 *hash_to_temp;
 288         int temp_count;
 289         int max_hash;
 290
 291         /* Just the count of the max register used. Higher count => higher
 292          * register pressure */
 293         int work_registers;
 294
 295         /* Used for cont/last hinting. Increase when a tex op is added.
 296          * Decrease when a tex op is removed. */
 297         int texture_op_count;
 298
 299         /* The number of uniforms allowable for the fast path */
 300         int uniform_cutoff;
 301
 302         /* Count of instructions emitted from NIR overall, across all blocks */
 303         int instruction_count;
 304
 305         /* Alpha ref value passed in */
 306         float alpha_ref;
 307
 308         unsigned quadword_count;
 309
 310         /* Bitmask of valid metadata */
 311         unsigned metadata;
 312
 313         /* Model-specific quirk set */
 314         uint32_t quirks;
 315
 316         /* Writeout instructions for each render target */
 317         midgard_instruction *writeout_branch[MIDGARD_NUM_RTS];
 318
 319         struct panfrost_sysvals sysvals;
 320 } compiler_context;
 321
 322 /* Per-block live_in/live_out */
 323 #define MIDGARD_METADATA_LIVENESS (1 << 0)
 324
 325 /* Helpers for manipulating the above structures (forming the driver IR) */
 326
 327 /* Append instruction to end of current block */
 328
 329 static inline midgard_instruction *
 330 mir_upload_ins(struct compiler_context *ctx, struct midgard_instruction ins)
 331 {
 332         midgard_instruction *heap = ralloc(ctx, struct midgard_instruction);
 333         memcpy(heap, &ins, sizeof(ins));
 334         return heap;
 335 }
 336
 337 static inline midgard_instruction *
 338 emit_mir_instruction(struct compiler_context *ctx, struct midgard_instruction ins)
 339 {
 340         midgard_instruction *u = mir_upload_ins(ctx, ins);
 341         list_addtail(&u->link, &ctx->current_block->instructions);
 342         return u;
 343 }
 344
 345 static inline struct midgard_instruction *
 346 mir_insert_instruction_before(struct compiler_context *ctx,
 347                               struct midgard_instruction *tag,
 348                               struct midgard_instruction ins)
 349 {
 350         struct midgard_instruction *u = mir_upload_ins(ctx, ins);
 351         list_addtail(&u->link, &tag->link);
 352         return u;
 353 }
 354
 355 static inline void
 356 mir_remove_instruction(struct midgard_instruction *ins)
 357 {
 358         list_del(&ins->link);
 359 }
 360
 361 static inline midgard_instruction*
 362 mir_prev_op(struct midgard_instruction *ins)
 363 {
 364         return list_last_entry(&(ins->link), midgard_instruction, link);
 365 }
 366
 367 static inline midgard_instruction*
 368 mir_next_op(struct midgard_instruction *ins)
 369 {
 370         return list_first_entry(&(ins->link), midgard_instruction, link);
 371 }
 372
 373 #define mir_foreach_block(ctx, v) \
 374         list_for_each_entry(struct midgard_block, v, &ctx->blocks, link)
 375
 376 #define mir_foreach_block_from(ctx, from, v) \
 377         list_for_each_entry_from(struct midgard_block, v, from, &ctx->blocks, link)
 378
 379 #define mir_foreach_instr_in_block(block, v) \
 380         list_for_each_entry(struct midgard_instruction, v, &block->instructions, link)
 381 #define mir_foreach_instr_in_block_rev(block, v) \
 382         list_for_each_entry_rev(struct midgard_instruction, v, &block->instructions, link)
 383
 384 #define mir_foreach_instr_in_block_safe(block, v) \
 385         list_for_each_entry_safe(struct midgard_instruction, v, &block->instructions, link)
 386
 387 #define mir_foreach_instr_in_block_safe_rev(block, v) \
 388         list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->instructions, link)
 389
 390 #define mir_foreach_instr_in_block_from(block, v, from) \
 391         list_for_each_entry_from(struct midgard_instruction, v, from, &block->instructions, link)
 392
 393 #define mir_foreach_instr_in_block_from_rev(block, v, from) \
 394         list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->instructions, link)
 395
 396 #define mir_foreach_bundle_in_block(block, v) \
 397         util_dynarray_foreach(&block->bundles, midgard_bundle, v)
 398
 399 #define mir_foreach_bundle_in_block_rev(block, v) \
 400         util_dynarray_foreach_reverse(&block->bundles, midgard_bundle, v)
 401
 402 #define mir_foreach_instr_in_block_scheduled_rev(block, v) \
 403         midgard_instruction* v; \
 404         signed i = 0; \
 405         mir_foreach_bundle_in_block_rev(block, _bundle) \
 406                 for (i = (_bundle->instruction_count - 1), v = _bundle->instructions[i]; \
 407                                 i >= 0; \
 408                                 --i, v = (i >= 0) ? _bundle->instructions[i] : NULL) \
 409
 410 #define mir_foreach_instr_global(ctx, v) \
 411         mir_foreach_block(ctx, v_block) \
 412                 mir_foreach_instr_in_block(v_block, v)
 413
 414 #define mir_foreach_instr_global_safe(ctx, v) \
 415         mir_foreach_block(ctx, v_block) \
 416                 mir_foreach_instr_in_block_safe(v_block, v)
 417
 418 #define mir_foreach_successor(blk, v) \
 419         struct midgard_block *v; \
 420         struct midgard_block **_v; \
 421         for (_v = &blk->successors[0], \
 422                 v = *_v; \
 423                 v != NULL && _v < &blk->successors[2]; \
 424                 _v++, v = *_v) \
 425
 426 /* Based on set_foreach, expanded with automatic type casts */
 427
 428 #define mir_foreach_predecessor(blk, v) \
 429         struct set_entry *_entry_##v; \
 430         struct midgard_block *v; \
 431         for (_entry_##v = _mesa_set_next_entry(blk->predecessors, NULL), \
 432                 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL);  \
 433                 _entry_##v != NULL; \
 434                 _entry_##v = _mesa_set_next_entry(blk->predecessors, _entry_##v), \
 435                 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
 436
 437 #define mir_foreach_src(ins, v) \
 438         for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
 439
 440 static inline midgard_instruction *
 441 mir_last_in_block(struct midgard_block *block)
 442 {
 443         return list_last_entry(&block->instructions, struct midgard_instruction, link);
 444 }
 445
 446 static inline midgard_block *
 447 mir_get_block(compiler_context *ctx, int idx)
 448 {
 449         struct list_head *lst = &ctx->blocks;
 450
 451         while ((idx--) + 1)
 452                 lst = lst->next;
 453
 454         return (struct midgard_block *) lst;
 455 }
 456
 457 static inline midgard_block *
 458 mir_exit_block(struct compiler_context *ctx)
 459 {
 460         midgard_block *last = list_last_entry(&ctx->blocks,
 461                         struct midgard_block, link);
 462
 463         /* The last block must be empty logically but contains branch writeout
 464          * for fragment shaders */
 465
 466         assert(last->nr_successors == 0);
 467
 468         return last;
 469 }
 470
 471 static inline bool
 472 mir_is_alu_bundle(midgard_bundle *bundle)
 473 {
 474         return IS_ALU(bundle->tag);
 475 }
 476
 477 /* Registers/SSA are distinguish in the backend by the bottom-most bit */
 478
 479 #define IS_REG (1)
 480
 481 static inline unsigned
 482 make_compiler_temp(compiler_context *ctx)
 483 {
 484         return (ctx->func->impl->ssa_alloc + ctx->temp_alloc++) << 1;
 485 }
 486
 487 static inline unsigned
 488 make_compiler_temp_reg(compiler_context *ctx)
 489 {
 490         return ((ctx->func->impl->reg_alloc + ctx->temp_alloc++) << 1) | IS_REG;
 491 }
 492
 493 static inline unsigned
 494 nir_ssa_index(nir_ssa_def *ssa)
 495 {
 496         return (ssa->index << 1) | 0;
 497 }
 498
 499 static inline unsigned
 500 nir_src_index(compiler_context *ctx, nir_src *src)
 501 {
 502         if (src->is_ssa)
 503                 return nir_ssa_index(src->ssa);
 504         else {
 505                 assert(!src->reg.indirect);
 506                 return (src->reg.reg->index << 1) | IS_REG;
 507         }
 508 }
 509
 510 static inline unsigned
 511 nir_alu_src_index(compiler_context *ctx, nir_alu_src *src)
 512 {
 513         return nir_src_index(ctx, &src->src);
 514 }
 515
 516 static inline unsigned
 517 nir_dest_index(nir_dest *dst)
 518 {
 519         if (dst->is_ssa)
 520                 return (dst->ssa.index << 1) | 0;
 521         else {
 522                 assert(!dst->reg.indirect);
 523                 return (dst->reg.reg->index << 1) | IS_REG;
 524         }
 525 }
 526
 527
 528
 529 /* MIR manipulation */
 530
 531 void mir_rewrite_index(compiler_context *ctx, unsigned old, unsigned new);
 532 void mir_rewrite_index_src(compiler_context *ctx, unsigned old, unsigned new);
 533 void mir_rewrite_index_dst(compiler_context *ctx, unsigned old, unsigned new);
 534 void mir_rewrite_index_dst_single(midgard_instruction *ins, unsigned old, unsigned new);
 535 void mir_rewrite_index_src_single(midgard_instruction *ins, unsigned old, unsigned new);
 536 void mir_rewrite_index_src_swizzle(compiler_context *ctx, unsigned old, unsigned new, unsigned *swizzle);
 537 bool mir_single_use(compiler_context *ctx, unsigned value);
 538 bool mir_special_index(compiler_context *ctx, unsigned idx);
 539 unsigned mir_use_count(compiler_context *ctx, unsigned value);
 540 bool mir_is_written_before(compiler_context *ctx, midgard_instruction *ins, unsigned node);
 541 uint16_t mir_bytemask_of_read_components(midgard_instruction *ins, unsigned node);
 542 uint16_t mir_bytemask_of_read_components_index(midgard_instruction *ins, unsigned i);
 543 midgard_reg_mode mir_typesize(midgard_instruction *ins);
 544 midgard_reg_mode mir_srcsize(midgard_instruction *ins, unsigned i);
 545 unsigned mir_bytes_for_mode(midgard_reg_mode mode);
 546 midgard_reg_mode mir_mode_for_destsize(unsigned size);
 547 uint16_t mir_from_bytemask(uint16_t bytemask, midgard_reg_mode mode);
 548 uint16_t mir_bytemask(midgard_instruction *ins);
 549 uint16_t mir_round_bytemask_up(uint16_t mask, midgard_reg_mode mode);
 550 void mir_set_bytemask(midgard_instruction *ins, uint16_t bytemask);
 551 unsigned mir_upper_override(midgard_instruction *ins);
 552
 553 /* MIR printing */
 554
 555 void mir_print_instruction(midgard_instruction *ins);
 556 void mir_print_bundle(midgard_bundle *ctx);
 557 void mir_print_block(midgard_block *block);
 558 void mir_print_shader(compiler_context *ctx);
 559 bool mir_nontrivial_source2_mod(midgard_instruction *ins);
 560 bool mir_nontrivial_source2_mod_simple(midgard_instruction *ins);
 561 bool mir_nontrivial_outmod(midgard_instruction *ins);
 562
 563 void mir_insert_instruction_before_scheduled(compiler_context *ctx, midgard_block *block, midgard_instruction *tag, midgard_instruction ins);
 564 void mir_insert_instruction_after_scheduled(compiler_context *ctx, midgard_block *block, midgard_instruction *tag, midgard_instruction ins);
 565 void mir_flip(midgard_instruction *ins);
 566 void mir_compute_temp_count(compiler_context *ctx);
 567
 568 void mir_set_offset(compiler_context *ctx, midgard_instruction *ins, nir_src *offset, bool is_shared);
 569
 570 /* 'Intrinsic' move for aliasing */
 571
 572 static inline midgard_instruction
 573 v_mov(unsigned src, unsigned dest)
 574 {
 575         midgard_instruction ins = {
 576                 .type = TAG_ALU_4,
 577                 .mask = 0xF,
 578                 .src = { ~0, src, ~0, ~0 },
 579                 .swizzle = SWIZZLE_IDENTITY,
 580                 .dest = dest,
 581                 .alu = {
 582                         .op = midgard_alu_op_imov,
 583                         .reg_mode = midgard_reg_mode_32,
 584                         .dest_override = midgard_dest_override_none,
 585                         .outmod = midgard_outmod_int_wrap
 586                 },
 587         };
 588
 589         return ins;
 590 }
 591
 592 /* Broad types of register classes so we can handle special
 593  * registers */
 594
 595 #define REG_CLASS_WORK          0
 596 #define REG_CLASS_LDST          1
 597 #define REG_CLASS_TEXR          3
 598 #define REG_CLASS_TEXW          4
 599
 600 /* Like a move, but to thread local storage! */
 601
 602 static inline midgard_instruction
 603 v_load_store_scratch(
 604                 unsigned srcdest,
 605                 unsigned index,
 606                 bool is_store,
 607                 unsigned mask)
 608 {
 609         /* We index by 32-bit vec4s */
 610         unsigned byte = (index * 4 * 4);
 611
 612         midgard_instruction ins = {
 613                 .type = TAG_LOAD_STORE_4,
 614                 .mask = mask,
 615                 .dest = ~0,
 616                 .src = { ~0, ~0, ~0, ~0 },
 617                 .swizzle = SWIZZLE_IDENTITY_4,
 618                 .load_store = {
 619                         .op = is_store ? midgard_op_st_int4 : midgard_op_ld_int4,
 620
 621                         /* For register spilling - to thread local storage */
 622                         .arg_1 = 0xEA,
 623                         .arg_2 = 0x1E,
 624                 },
 625
 626                 /* If we spill an unspill, RA goes into an infinite loop */
 627                 .no_spill = (1 << REG_CLASS_WORK)
 628         };
 629
 630         ins.constants.u32[0] = byte;
 631
 632         if (is_store) {
 633                 ins.src[0] = srcdest;
 634
 635                 /* Ensure we are tightly swizzled so liveness analysis is
 636                  * correct */
 637
 638                 for (unsigned i = 0; i < 4; ++i) {
 639                         if (!(mask & (1 << i)))
 640                                 ins.swizzle[0][i] = COMPONENT_X;
 641                 }
 642         } else
 643                 ins.dest = srcdest;
 644
 645         return ins;
 646 }
 647
 648 static inline bool
 649 mir_has_arg(midgard_instruction *ins, unsigned arg)
 650 {
 651         if (!ins)
 652                 return false;
 653
 654         for (unsigned i = 0; i < ARRAY_SIZE(ins->src); ++i) {
 655                 if (ins->src[i] == arg)
 656                         return true;
 657         }
 658
 659         return false;
 660 }
 661
 662 /* Scheduling */
 663
 664 void midgard_schedule_program(compiler_context *ctx);
 665
 666 void mir_ra(compiler_context *ctx);
 667 void mir_squeeze_index(compiler_context *ctx);
 668 void mir_lower_special_reads(compiler_context *ctx);
 669 void mir_liveness_ins_update(uint16_t *live, midgard_instruction *ins, unsigned max);
 670 void mir_compute_liveness(compiler_context *ctx);
 671 void mir_invalidate_liveness(compiler_context *ctx);
 672 bool mir_is_live_after(compiler_context *ctx, midgard_block *block, midgard_instruction *start, int src);
 673
 674 void mir_create_pipeline_registers(compiler_context *ctx);
 675 void midgard_promote_uniforms(compiler_context *ctx);
 676
 677 void
 678 midgard_emit_derivatives(compiler_context *ctx, nir_alu_instr *instr);
 679
 680 void
 681 midgard_lower_derivatives(compiler_context *ctx, midgard_block *block);
 682
 683 bool mir_op_computes_derivatives(gl_shader_stage stage, unsigned op);
 684
 685 /* Final emission */
 686
 687 void emit_binary_bundle(
 688         compiler_context *ctx,
 689         midgard_bundle *bundle,
 690         struct util_dynarray *emission,
 691         int next_tag);
 692
 693 bool
 694 nir_undef_to_zero(nir_shader *shader);
 695
 696 void midgard_nir_lod_errata(nir_shader *shader);
 697
 698 /* Optimizations */
 699
 700 bool midgard_opt_copy_prop(compiler_context *ctx, midgard_block *block);
 701 bool midgard_opt_combine_projection(compiler_context *ctx, midgard_block *block);
 702 bool midgard_opt_varying_projection(compiler_context *ctx, midgard_block *block);
 703 bool midgard_opt_dead_code_eliminate(compiler_context *ctx, midgard_block *block);
 704 bool midgard_opt_dead_move_eliminate(compiler_context *ctx, midgard_block *block);
 705
 706 void midgard_lower_invert(compiler_context *ctx, midgard_block *block);
 707 bool midgard_opt_not_propagate(compiler_context *ctx, midgard_block *block);
 708 bool midgard_opt_fuse_src_invert(compiler_context *ctx, midgard_block *block);
 709 bool midgard_opt_fuse_dest_invert(compiler_context *ctx, midgard_block *block);
 710 bool midgard_opt_csel_invert(compiler_context *ctx, midgard_block *block);
 711 bool midgard_opt_promote_fmov(compiler_context *ctx, midgard_block *block);
 712 bool midgard_opt_drop_cmp_invert(compiler_context *ctx, midgard_block *block);
 713 bool midgard_opt_invert_branch(compiler_context *ctx, midgard_block *block);
 714
 715 #endif