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
  31 #include "util/hash_table.h"
  32 #include "util/u_dynarray.h"
  33 #include "util/set.h"
  34 #include "util/list.h"
  35
  36 #include "main/mtypes.h"
  37 #include "compiler/nir_types.h"
  38 #include "compiler/nir/nir.h"
  39
  40 /* Forward declare */
  41 struct midgard_block;
  42
  43 /* Target types. Defaults to TARGET_GOTO (the type corresponding directly to
  44  * the hardware), hence why that must be zero. TARGET_DISCARD signals this
  45  * instruction is actually a discard op. */
  46
  47 #define TARGET_GOTO 0
  48 #define TARGET_BREAK 1
  49 #define TARGET_CONTINUE 2
  50 #define TARGET_DISCARD 3
  51
  52 typedef struct midgard_branch {
  53         /* If conditional, the condition is specified in r31.w */
  54         bool conditional;
  55
  56         /* For conditionals, if this is true, we branch on FALSE. If false, we  branch on TRUE. */
  57         bool invert_conditional;
  58
  59         /* Branch targets: the start of a block, the start of a loop (continue), the end of a loop (break). Value is one of TARGET_ */
  60         unsigned target_type;
  61
  62         /* The actual target */
  63         union {
  64                 int target_block;
  65                 int target_break;
  66                 int target_continue;
  67         };
  68 } midgard_branch;
  69
  70 /* Instruction arguments represented as block-local SSA indices, rather than
  71  * registers. Negative values mean unused. */
  72
  73 typedef struct {
  74         int src[3];
  75         int dest;
  76
  77         bool inline_constant;
  78 } ssa_args;
  79
  80 /* Generic in-memory data type repesenting a single logical instruction, rather
  81  * than a single instruction group. This is the preferred form for code gen.
  82  * Multiple midgard_insturctions will later be combined during scheduling,
  83  * though this is not represented in this structure.  Its format bridges
  84  * the low-level binary representation with the higher level semantic meaning.
  85  *
  86  * Notably, it allows registers to be specified as block local SSA, for code
  87  * emitted before the register allocation pass.
  88  */
  89
  90 typedef struct midgard_instruction {
  91         /* Must be first for casting */
  92         struct list_head link;
  93
  94         unsigned type; /* ALU, load/store, texture */
  95
  96         /* If the register allocator has not run yet... */
  97         ssa_args ssa_args;
  98
  99         /* Special fields for an ALU instruction */
 100         midgard_reg_info registers;
 101
 102         /* I.e. (1 << alu_bit) */
 103         int unit;
 104
 105         /* When emitting bundle, should this instruction have a break forced
 106          * before it? Used for r31 writes which are valid only within a single
 107          * bundle and *need* to happen as early as possible... this is a hack,
 108          * TODO remove when we have a scheduler */
 109         bool precede_break;
 110
 111         bool has_constants;
 112         float constants[4];
 113         uint16_t inline_constant;
 114         bool has_blend_constant;
 115
 116         bool compact_branch;
 117         bool writeout;
 118         bool prepacked_branch;
 119
 120         /* Kind of a hack, but hint against aggressive DCE */
 121         bool dont_eliminate;
 122
 123         /* Masks in a saneish format. One bit per channel, not packed fancy.
 124          * Use this instead of the op specific ones, and switch over at emit
 125          * time */
 126
 127         uint16_t mask;
 128
 129         /* For ALU ops only: set to true to invert (bitwise NOT) the
 130          * destination of an integer-out op. Not imeplemented in hardware but
 131          * allows more optimizations */
 132
 133         bool invert;
 134
 135         /* Hint for the register allocator not to spill the destination written
 136          * from this instruction (because it is a spill/unspill node itself) */
 137
 138         bool no_spill;
 139
 140         /* Generic hint for intra-pass use */
 141         bool hint;
 142
 143         union {
 144                 midgard_load_store_word load_store;
 145                 midgard_vector_alu alu;
 146                 midgard_texture_word texture;
 147                 midgard_branch_extended branch_extended;
 148                 uint16_t br_compact;
 149
 150                 /* General branch, rather than packed br_compact. Higher level
 151                  * than the other components */
 152                 midgard_branch branch;
 153         };
 154 } midgard_instruction;
 155
 156 typedef struct midgard_block {
 157         /* Link to next block. Must be first for mir_get_block */
 158         struct list_head link;
 159
 160         /* List of midgard_instructions emitted for the current block */
 161         struct list_head instructions;
 162
 163         bool is_scheduled;
 164
 165         /* List of midgard_bundles emitted (after the scheduler has run) */
 166         struct util_dynarray bundles;
 167
 168         /* Number of quadwords _actually_ emitted, as determined after scheduling */
 169         unsigned quadword_count;
 170
 171         /* Succeeding blocks. The compiler should not necessarily rely on
 172          * source-order traversal */
 173         struct midgard_block *successors[2];
 174         unsigned nr_successors;
 175
 176         struct set *predecessors;
 177
 178         /* The successors pointer form a graph, and in the case of
 179          * complex control flow, this graph has a cycles. To aid
 180          * traversal during liveness analysis, we have a visited?
 181          * boolean for passes to use as they see fit, provided they
 182          * clean up later */
 183         bool visited;
 184 } midgard_block;
 185
 186 typedef struct midgard_bundle {
 187         /* Tag for the overall bundle */
 188         int tag;
 189
 190         /* Instructions contained by the bundle */
 191         int instruction_count;
 192         midgard_instruction *instructions[5];
 193
 194         /* Bundle-wide ALU configuration */
 195         int padding;
 196         int control;
 197         bool has_embedded_constants;
 198         float constants[4];
 199         bool has_blend_constant;
 200 } midgard_bundle;
 201
 202 typedef struct compiler_context {
 203         nir_shader *nir;
 204         gl_shader_stage stage;
 205
 206         /* The screen we correspond to */
 207         struct midgard_screen *screen;
 208
 209         /* Is internally a blend shader? Depends on stage == FRAGMENT */
 210         bool is_blend;
 211
 212         /* Tracking for blend constant patching */
 213         int blend_constant_offset;
 214
 215         /* Number of bytes used for Thread Local Storage */
 216         unsigned tls_size;
 217
 218         /* Count of spills and fills for shaderdb */
 219         unsigned spills;
 220         unsigned fills;
 221
 222         /* Current NIR function */
 223         nir_function *func;
 224
 225         /* Allocated compiler temporary counter */
 226         unsigned temp_alloc;
 227
 228         /* Unordered list of midgard_blocks */
 229         int block_count;
 230         struct list_head blocks;
 231
 232         /* List of midgard_instructions emitted for the current block */
 233         midgard_block *current_block;
 234
 235         /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
 236         midgard_block *after_block;
 237
 238         /* The current "depth" of the loop, for disambiguating breaks/continues
 239          * when using nested loops */
 240         int current_loop_depth;
 241
 242         /* Total number of loops for shader-db */
 243         unsigned loop_count;
 244
 245         /* Constants which have been loaded, for later inlining */
 246         struct hash_table_u64 *ssa_constants;
 247
 248         /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
 249         struct hash_table_u64 *hash_to_temp;
 250         int temp_count;
 251         int max_hash;
 252
 253         /* Just the count of the max register used. Higher count => higher
 254          * register pressure */
 255         int work_registers;
 256
 257         /* Used for cont/last hinting. Increase when a tex op is added.
 258          * Decrease when a tex op is removed. */
 259         int texture_op_count;
 260
 261         /* Mapping of texture register -> SSA index for unaliasing */
 262         int texture_index[2];
 263
 264         /* The number of uniforms allowable for the fast path */
 265         int uniform_cutoff;
 266
 267         /* Count of instructions emitted from NIR overall, across all blocks */
 268         int instruction_count;
 269
 270         /* Alpha ref value passed in */
 271         float alpha_ref;
 272
 273         /* The mapping of sysvals to uniforms, the count, and the off-by-one inverse */
 274         unsigned sysvals[MAX_SYSVAL_COUNT];
 275         unsigned sysval_count;
 276         struct hash_table_u64 *sysval_to_id;
 277 } compiler_context;
 278
 279 /* Helpers for manipulating the above structures (forming the driver IR) */
 280
 281 /* Append instruction to end of current block */
 282
 283 static inline midgard_instruction *
 284 mir_upload_ins(struct midgard_instruction ins)
 285 {
 286         midgard_instruction *heap = malloc(sizeof(ins));
 287         memcpy(heap, &ins, sizeof(ins));
 288         return heap;
 289 }
 290
 291 static inline midgard_instruction *
 292 emit_mir_instruction(struct compiler_context *ctx, struct midgard_instruction ins)
 293 {
 294         midgard_instruction *u = mir_upload_ins(ins);
 295         list_addtail(&u->link, &ctx->current_block->instructions);
 296         return u;
 297 }
 298
 299 static inline struct midgard_instruction *
 300 mir_insert_instruction_before(struct midgard_instruction *tag, struct midgard_instruction ins)
 301 {
 302         struct midgard_instruction *u = mir_upload_ins(ins);
 303         list_addtail(&u->link, &tag->link);
 304         return u;
 305 }
 306
 307 static inline void
 308 mir_remove_instruction(struct midgard_instruction *ins)
 309 {
 310         list_del(&ins->link);
 311 }
 312
 313 static inline midgard_instruction*
 314 mir_prev_op(struct midgard_instruction *ins)
 315 {
 316         return list_last_entry(&(ins->link), midgard_instruction, link);
 317 }
 318
 319 static inline midgard_instruction*
 320 mir_next_op(struct midgard_instruction *ins)
 321 {
 322         return list_first_entry(&(ins->link), midgard_instruction, link);
 323 }
 324
 325 #define mir_foreach_block(ctx, v) \
 326         list_for_each_entry(struct midgard_block, v, &ctx->blocks, link)
 327
 328 #define mir_foreach_block_from(ctx, from, v) \
 329         list_for_each_entry_from(struct midgard_block, v, from, &ctx->blocks, link)
 330
 331 #define mir_foreach_instr(ctx, v) \
 332         list_for_each_entry(struct midgard_instruction, v, &ctx->current_block->instructions, link)
 333
 334 #define mir_foreach_instr_safe(ctx, v) \
 335         list_for_each_entry_safe(struct midgard_instruction, v, &ctx->current_block->instructions, link)
 336
 337 #define mir_foreach_instr_in_block(block, v) \
 338         list_for_each_entry(struct midgard_instruction, v, &block->instructions, link)
 339
 340 #define mir_foreach_instr_in_block_safe(block, v) \
 341         list_for_each_entry_safe(struct midgard_instruction, v, &block->instructions, link)
 342
 343 #define mir_foreach_instr_in_block_safe_rev(block, v) \
 344         list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->instructions, link)
 345
 346 #define mir_foreach_instr_in_block_from(block, v, from) \
 347         list_for_each_entry_from(struct midgard_instruction, v, from, &block->instructions, link)
 348
 349 #define mir_foreach_instr_in_block_from_rev(block, v, from) \
 350         list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->instructions, link)
 351
 352 #define mir_foreach_bundle_in_block(block, v) \
 353         util_dynarray_foreach(&block->bundles, midgard_bundle, v)
 354
 355 #define mir_foreach_instr_global(ctx, v) \
 356         mir_foreach_block(ctx, v_block) \
 357                 mir_foreach_instr_in_block(v_block, v)
 358
 359 #define mir_foreach_instr_global_safe(ctx, v) \
 360         mir_foreach_block(ctx, v_block) \
 361                 mir_foreach_instr_in_block_safe(v_block, v)
 362
 363 /* Based on set_foreach, expanded with automatic type casts */
 364
 365 #define mir_foreach_predecessor(blk, v) \
 366         struct set_entry *_entry_##v; \
 367         struct midgard_block *v; \
 368         for (_entry_##v = _mesa_set_next_entry(blk->predecessors, NULL), \
 369                 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL);  \
 370                 _entry_##v != NULL; \
 371                 _entry_##v = _mesa_set_next_entry(blk->predecessors, _entry_##v), \
 372                 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
 373
 374 static inline midgard_instruction *
 375 mir_last_in_block(struct midgard_block *block)
 376 {
 377         return list_last_entry(&block->instructions, struct midgard_instruction, link);
 378 }
 379
 380 static inline midgard_block *
 381 mir_get_block(compiler_context *ctx, int idx)
 382 {
 383         struct list_head *lst = &ctx->blocks;
 384
 385         while ((idx--) + 1)
 386                 lst = lst->next;
 387
 388         return (struct midgard_block *) lst;
 389 }
 390
 391 static inline midgard_block *
 392 mir_exit_block(struct compiler_context *ctx)
 393 {
 394         midgard_block *last = list_last_entry(&ctx->blocks,
 395                         struct midgard_block, link);
 396
 397         /* The last block must be empty (the exit block) */
 398         assert(list_empty(&last->instructions));
 399         assert(last->nr_successors == 0);
 400
 401         return last;
 402 }
 403
 404 static inline bool
 405 mir_is_alu_bundle(midgard_bundle *bundle)
 406 {
 407         return IS_ALU(bundle->tag);
 408 }
 409
 410 /* Registers/SSA are distinguish in the backend by the bottom-most bit */
 411
 412 #define IS_REG (1)
 413
 414 static inline unsigned
 415 make_compiler_temp(compiler_context *ctx)
 416 {
 417         return (ctx->func->impl->ssa_alloc + ctx->temp_alloc++) << 1;
 418 }
 419
 420 static inline unsigned
 421 make_compiler_temp_reg(compiler_context *ctx)
 422 {
 423         return ((ctx->func->impl->reg_alloc + ctx->temp_alloc++) << 1) | IS_REG;
 424 }
 425
 426 static inline unsigned
 427 nir_src_index(compiler_context *ctx, nir_src *src)
 428 {
 429         if (src->is_ssa)
 430                 return (src->ssa->index << 1) | 0;
 431         else {
 432                 assert(!src->reg.indirect);
 433                 return (src->reg.reg->index << 1) | IS_REG;
 434         }
 435 }
 436
 437 static inline unsigned
 438 nir_alu_src_index(compiler_context *ctx, nir_alu_src *src)
 439 {
 440         return nir_src_index(ctx, &src->src);
 441 }
 442
 443 static inline unsigned
 444 nir_dest_index(compiler_context *ctx, nir_dest *dst)
 445 {
 446         if (dst->is_ssa)
 447                 return (dst->ssa.index << 1) | 0;
 448         else {
 449                 assert(!dst->reg.indirect);
 450                 return (dst->reg.reg->index << 1) | IS_REG;
 451         }
 452 }
 453
 454
 455
 456 /* MIR manipulation */
 457
 458 void mir_rewrite_index(compiler_context *ctx, unsigned old, unsigned new);
 459 void mir_rewrite_index_src(compiler_context *ctx, unsigned old, unsigned new);
 460 void mir_rewrite_index_dst(compiler_context *ctx, unsigned old, unsigned new);
 461 void mir_rewrite_index_dst_tag(compiler_context *ctx, unsigned old, unsigned new, unsigned tag);
 462 void mir_rewrite_index_src_single(midgard_instruction *ins, unsigned old, unsigned new);
 463 void mir_rewrite_index_src_tag(compiler_context *ctx, unsigned old, unsigned new, unsigned tag);
 464 void mir_rewrite_index_src_swizzle(compiler_context *ctx, unsigned old, unsigned new, unsigned swizzle);
 465 bool mir_single_use(compiler_context *ctx, unsigned value);
 466 bool mir_special_index(compiler_context *ctx, unsigned idx);
 467 unsigned mir_use_count(compiler_context *ctx, unsigned value);
 468 bool mir_is_written_before(compiler_context *ctx, midgard_instruction *ins, unsigned node);
 469 unsigned mir_mask_of_read_components(midgard_instruction *ins, unsigned node);
 470 unsigned mir_ubo_shift(midgard_load_store_op op);
 471
 472 /* MIR printing */
 473
 474 void mir_print_instruction(midgard_instruction *ins);
 475 void mir_print_bundle(midgard_bundle *ctx);
 476 void mir_print_block(midgard_block *block);
 477 void mir_print_shader(compiler_context *ctx);
 478 bool mir_nontrivial_source2_mod(midgard_instruction *ins);
 479 bool mir_nontrivial_source2_mod_simple(midgard_instruction *ins);
 480 bool mir_nontrivial_mod(midgard_vector_alu_src src, bool is_int, unsigned mask);
 481 bool mir_nontrivial_outmod(midgard_instruction *ins);
 482
 483 /* MIR goodies */
 484
 485 static const midgard_vector_alu_src blank_alu_src = {
 486         .swizzle = SWIZZLE(COMPONENT_X, COMPONENT_Y, COMPONENT_Z, COMPONENT_W),
 487 };
 488
 489 static const midgard_vector_alu_src blank_alu_src_xxxx = {
 490         .swizzle = SWIZZLE(COMPONENT_X, COMPONENT_X, COMPONENT_X, COMPONENT_X),
 491 };
 492
 493 static const midgard_scalar_alu_src blank_scalar_alu_src = {
 494         .full = true
 495 };
 496
 497 /* Used for encoding the unused source of 1-op instructions */
 498 static const midgard_vector_alu_src zero_alu_src = { 0 };
 499
 500 /* 'Intrinsic' move for aliasing */
 501
 502 static inline midgard_instruction
 503 v_mov(unsigned src, midgard_vector_alu_src mod, unsigned dest)
 504 {
 505         midgard_instruction ins = {
 506                 .type = TAG_ALU_4,
 507                 .mask = 0xF,
 508                 .ssa_args = {
 509                         .src = { SSA_UNUSED_1, src, -1 },
 510                         .dest = dest,
 511                 },
 512                 .alu = {
 513                         .op = midgard_alu_op_imov,
 514                         .reg_mode = midgard_reg_mode_32,
 515                         .dest_override = midgard_dest_override_none,
 516                         .outmod = midgard_outmod_int_wrap,
 517                         .src1 = vector_alu_srco_unsigned(zero_alu_src),
 518                         .src2 = vector_alu_srco_unsigned(mod)
 519                 },
 520         };
 521
 522         return ins;
 523 }
 524
 525 static inline bool
 526 mir_has_arg(midgard_instruction *ins, unsigned arg)
 527 {
 528         for (unsigned i = 0; i < ARRAY_SIZE(ins->ssa_args.src); ++i) {
 529                 if (ins->ssa_args.src[i] == arg)
 530                         return true;
 531         }
 532
 533         return false;
 534 }
 535
 536 /* Scheduling */
 537
 538 void schedule_program(compiler_context *ctx);
 539
 540 /* Register allocation */
 541
 542 struct ra_graph;
 543
 544 /* Broad types of register classes so we can handle special
 545  * registers */
 546
 547 #define NR_REG_CLASSES 5
 548
 549 #define REG_CLASS_WORK          0
 550 #define REG_CLASS_LDST          1
 551 #define REG_CLASS_LDST27        2
 552 #define REG_CLASS_TEXR          3
 553 #define REG_CLASS_TEXW          4
 554
 555 void mir_lower_special_reads(compiler_context *ctx);
 556 struct ra_graph* allocate_registers(compiler_context *ctx, bool *spilled);
 557 void install_registers(compiler_context *ctx, struct ra_graph *g);
 558 bool mir_is_live_after(compiler_context *ctx, midgard_block *block, midgard_instruction *start, int src);
 559 bool mir_has_multiple_writes(compiler_context *ctx, int src);
 560
 561 void mir_create_pipeline_registers(compiler_context *ctx);
 562
 563 void
 564 midgard_promote_uniforms(compiler_context *ctx, unsigned promoted_count);
 565
 566 midgard_instruction *
 567 emit_ubo_read(
 568         compiler_context *ctx,
 569         nir_instr *instr,
 570         unsigned dest,
 571         unsigned offset,
 572         nir_src *indirect_offset,
 573         unsigned index);
 574
 575 void
 576 emit_sysval_read(compiler_context *ctx, nir_instr *instr, signed dest_override, unsigned nr_components);
 577
 578 void
 579 midgard_emit_derivatives(compiler_context *ctx, nir_alu_instr *instr);
 580
 581 void
 582 midgard_lower_derivatives(compiler_context *ctx, midgard_block *block);
 583
 584 bool mir_op_computes_derivatives(unsigned op);
 585
 586 /* Final emission */
 587
 588 void emit_binary_bundle(
 589         compiler_context *ctx,
 590         midgard_bundle *bundle,
 591         struct util_dynarray *emission,
 592         int next_tag);
 593
 594 /* NIR stuff. TODO: Move? Share? Something? */
 595
 596 bool
 597 nir_undef_to_zero(nir_shader *shader);
 598
 599 void
 600 nir_clamp_psiz(nir_shader *shader, float min_size, float max_size);
 601
 602 /* Optimizations */
 603
 604 bool midgard_opt_copy_prop(compiler_context *ctx, midgard_block *block);
 605 bool midgard_opt_combine_projection(compiler_context *ctx, midgard_block *block);
 606 bool midgard_opt_varying_projection(compiler_context *ctx, midgard_block *block);
 607 bool midgard_opt_dead_code_eliminate(compiler_context *ctx, midgard_block *block);
 608 bool midgard_opt_dead_move_eliminate(compiler_context *ctx, midgard_block *block);
 609 void midgard_opt_post_move_eliminate(compiler_context *ctx, midgard_block *block, struct ra_graph *g);
 610
 611 void midgard_lower_invert(compiler_context *ctx, midgard_block *block);
 612 bool midgard_opt_not_propagate(compiler_context *ctx, midgard_block *block);
 613 bool midgard_opt_fuse_src_invert(compiler_context *ctx, midgard_block *block);
 614 bool midgard_opt_fuse_dest_invert(compiler_context *ctx, midgard_block *block);
 615
 616 #endif