pan/midgard: Use new scheduler
[mesa.git] / 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 /* Generic in-memory data type repesenting a single logical instruction, rather
71 * than a single instruction group. This is the preferred form for code gen.
72 * Multiple midgard_insturctions will later be combined during scheduling,
73 * though this is not represented in this structure. Its format bridges
74 * the low-level binary representation with the higher level semantic meaning.
75 *
76 * Notably, it allows registers to be specified as block local SSA, for code
77 * emitted before the register allocation pass.
78 */
79
80 typedef struct midgard_instruction {
81 /* Must be first for casting */
82 struct list_head link;
83
84 unsigned type; /* ALU, load/store, texture */
85
86 /* Instruction arguments represented as block-local SSA
87 * indices, rather than registers. ~0 means unused. */
88 unsigned src[3];
89 unsigned dest;
90
91 /* Swizzle for the conditional for a csel/branch */
92 unsigned cond_swizzle;
93
94 /* Special fields for an ALU instruction */
95 midgard_reg_info registers;
96
97 /* I.e. (1 << alu_bit) */
98 int unit;
99
100 bool has_constants;
101 uint32_t constants[4];
102 uint16_t inline_constant;
103 bool has_blend_constant;
104 bool has_inline_constant;
105
106 bool compact_branch;
107 bool writeout;
108 bool prepacked_branch;
109
110 /* Kind of a hack, but hint against aggressive DCE */
111 bool dont_eliminate;
112
113 /* Masks in a saneish format. One bit per channel, not packed fancy.
114 * Use this instead of the op specific ones, and switch over at emit
115 * time */
116
117 uint16_t mask;
118
119 /* For ALU ops only: set to true to invert (bitwise NOT) the
120 * destination of an integer-out op. Not imeplemented in hardware but
121 * allows more optimizations */
122
123 bool invert;
124
125 /* Hint for the register allocator not to spill the destination written
126 * from this instruction (because it is a spill/unspill node itself) */
127
128 bool no_spill;
129
130 /* Generic hint for intra-pass use */
131 bool hint;
132
133 /* During scheduling, the backwards dependency graph
134 * (DAG). nr_dependencies is the number of unscheduled
135 * instructions that must still be scheduled after
136 * (before) this instruction. dependents are which
137 * instructions need to be scheduled before (after) this
138 * instruction. */
139
140 unsigned nr_dependencies;
141 BITSET_WORD *dependents;
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 /* Index of the block in source order */
164 unsigned source_id;
165
166 bool is_scheduled;
167
168 /* List of midgard_bundles emitted (after the scheduler has run) */
169 struct util_dynarray bundles;
170
171 /* Number of quadwords _actually_ emitted, as determined after scheduling */
172 unsigned quadword_count;
173
174 /* Succeeding blocks. The compiler should not necessarily rely on
175 * source-order traversal */
176 struct midgard_block *successors[2];
177 unsigned nr_successors;
178
179 struct set *predecessors;
180
181 /* The successors pointer form a graph, and in the case of
182 * complex control flow, this graph has a cycles. To aid
183 * traversal during liveness analysis, we have a visited?
184 * boolean for passes to use as they see fit, provided they
185 * clean up later */
186 bool visited;
187
188 /* In liveness analysis, these are live masks (per-component) for
189 * indices for the block. Scalar compilers have the luxury of using
190 * simple bit fields, but for us, liveness is a vector idea. We use
191 * 8-bit to allow finegrained tracking up to vec8. If you're
192 * implementing vec16 on Panfrost... I'm sorry. */
193 uint8_t *live_in;
194 uint8_t *live_out;
195 } midgard_block;
196
197 typedef struct midgard_bundle {
198 /* Tag for the overall bundle */
199 int tag;
200
201 /* Instructions contained by the bundle */
202 int instruction_count;
203 midgard_instruction *instructions[5];
204
205 /* Bundle-wide ALU configuration */
206 int padding;
207 int control;
208 bool has_embedded_constants;
209 float constants[4];
210 bool has_blend_constant;
211 } midgard_bundle;
212
213 typedef struct compiler_context {
214 nir_shader *nir;
215 gl_shader_stage stage;
216
217 /* The screen we correspond to */
218 struct midgard_screen *screen;
219
220 /* Is internally a blend shader? Depends on stage == FRAGMENT */
221 bool is_blend;
222
223 /* Tracking for blend constant patching */
224 int blend_constant_offset;
225
226 /* Number of bytes used for Thread Local Storage */
227 unsigned tls_size;
228
229 /* Count of spills and fills for shaderdb */
230 unsigned spills;
231 unsigned fills;
232
233 /* Current NIR function */
234 nir_function *func;
235
236 /* Allocated compiler temporary counter */
237 unsigned temp_alloc;
238
239 /* Unordered list of midgard_blocks */
240 int block_count;
241 struct list_head blocks;
242
243 /* TODO merge with block_count? */
244 unsigned block_source_count;
245
246 /* List of midgard_instructions emitted for the current block */
247 midgard_block *current_block;
248
249 /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
250 midgard_block *after_block;
251
252 /* The current "depth" of the loop, for disambiguating breaks/continues
253 * when using nested loops */
254 int current_loop_depth;
255
256 /* Total number of loops for shader-db */
257 unsigned loop_count;
258
259 /* Constants which have been loaded, for later inlining */
260 struct hash_table_u64 *ssa_constants;
261
262 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
263 struct hash_table_u64 *hash_to_temp;
264 int temp_count;
265 int max_hash;
266
267 /* Just the count of the max register used. Higher count => higher
268 * register pressure */
269 int work_registers;
270
271 /* Used for cont/last hinting. Increase when a tex op is added.
272 * Decrease when a tex op is removed. */
273 int texture_op_count;
274
275 /* The number of uniforms allowable for the fast path */
276 int uniform_cutoff;
277
278 /* Count of instructions emitted from NIR overall, across all blocks */
279 int instruction_count;
280
281 /* Alpha ref value passed in */
282 float alpha_ref;
283
284 unsigned quadword_count;
285
286 /* The mapping of sysvals to uniforms, the count, and the off-by-one inverse */
287 unsigned sysvals[MAX_SYSVAL_COUNT];
288 unsigned sysval_count;
289 struct hash_table_u64 *sysval_to_id;
290 } compiler_context;
291
292 /* Helpers for manipulating the above structures (forming the driver IR) */
293
294 /* Append instruction to end of current block */
295
296 static inline midgard_instruction *
297 mir_upload_ins(struct compiler_context *ctx, struct midgard_instruction ins)
298 {
299 midgard_instruction *heap = ralloc(ctx, struct midgard_instruction);
300 memcpy(heap, &ins, sizeof(ins));
301 return heap;
302 }
303
304 static inline midgard_instruction *
305 emit_mir_instruction(struct compiler_context *ctx, struct midgard_instruction ins)
306 {
307 midgard_instruction *u = mir_upload_ins(ctx, ins);
308 list_addtail(&u->link, &ctx->current_block->instructions);
309 return u;
310 }
311
312 static inline struct midgard_instruction *
313 mir_insert_instruction_before(struct compiler_context *ctx,
314 struct midgard_instruction *tag,
315 struct midgard_instruction ins)
316 {
317 struct midgard_instruction *u = mir_upload_ins(ctx, ins);
318 list_addtail(&u->link, &tag->link);
319 return u;
320 }
321
322 static inline void
323 mir_remove_instruction(struct midgard_instruction *ins)
324 {
325 list_del(&ins->link);
326 }
327
328 static inline midgard_instruction*
329 mir_prev_op(struct midgard_instruction *ins)
330 {
331 return list_last_entry(&(ins->link), midgard_instruction, link);
332 }
333
334 static inline midgard_instruction*
335 mir_next_op(struct midgard_instruction *ins)
336 {
337 return list_first_entry(&(ins->link), midgard_instruction, link);
338 }
339
340 #define mir_foreach_block(ctx, v) \
341 list_for_each_entry(struct midgard_block, v, &ctx->blocks, link)
342
343 #define mir_foreach_block_from(ctx, from, v) \
344 list_for_each_entry_from(struct midgard_block, v, from, &ctx->blocks, link)
345
346 #define mir_foreach_instr(ctx, v) \
347 list_for_each_entry(struct midgard_instruction, v, &ctx->current_block->instructions, link)
348
349 #define mir_foreach_instr_safe(ctx, v) \
350 list_for_each_entry_safe(struct midgard_instruction, v, &ctx->current_block->instructions, link)
351
352 #define mir_foreach_instr_in_block(block, v) \
353 list_for_each_entry(struct midgard_instruction, v, &block->instructions, link)
354 #define mir_foreach_instr_in_block_rev(block, v) \
355 list_for_each_entry_rev(struct midgard_instruction, v, &block->instructions, link)
356
357 #define mir_foreach_instr_in_block_safe(block, v) \
358 list_for_each_entry_safe(struct midgard_instruction, v, &block->instructions, link)
359
360 #define mir_foreach_instr_in_block_safe_rev(block, v) \
361 list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->instructions, link)
362
363 #define mir_foreach_instr_in_block_from(block, v, from) \
364 list_for_each_entry_from(struct midgard_instruction, v, from, &block->instructions, link)
365
366 #define mir_foreach_instr_in_block_from_rev(block, v, from) \
367 list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->instructions, link)
368
369 #define mir_foreach_bundle_in_block(block, v) \
370 util_dynarray_foreach(&block->bundles, midgard_bundle, v)
371
372 #define mir_foreach_bundle_in_block_rev(block, v) \
373 util_dynarray_foreach_reverse(&block->bundles, midgard_bundle, v)
374
375 #define mir_foreach_instr_in_block_scheduled_rev(block, v) \
376 midgard_instruction* v; \
377 signed i = 0; \
378 mir_foreach_bundle_in_block_rev(block, _bundle) \
379 for (i = (_bundle->instruction_count - 1), v = _bundle->instructions[i]; \
380 i >= 0; \
381 --i, v = _bundle->instructions[i]) \
382
383 #define mir_foreach_instr_global(ctx, v) \
384 mir_foreach_block(ctx, v_block) \
385 mir_foreach_instr_in_block(v_block, v)
386
387 #define mir_foreach_instr_global_safe(ctx, v) \
388 mir_foreach_block(ctx, v_block) \
389 mir_foreach_instr_in_block_safe(v_block, v)
390
391 #define mir_foreach_successor(blk, v) \
392 struct midgard_block *v; \
393 struct midgard_block **_v; \
394 for (_v = &blk->successors[0], \
395 v = *_v; \
396 v != NULL && _v < &blk->successors[2]; \
397 _v++, v = *_v) \
398
399 /* Based on set_foreach, expanded with automatic type casts */
400
401 #define mir_foreach_predecessor(blk, v) \
402 struct set_entry *_entry_##v; \
403 struct midgard_block *v; \
404 for (_entry_##v = _mesa_set_next_entry(blk->predecessors, NULL), \
405 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL); \
406 _entry_##v != NULL; \
407 _entry_##v = _mesa_set_next_entry(blk->predecessors, _entry_##v), \
408 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
409
410 #define mir_foreach_src(ins, v) \
411 for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
412
413 static inline midgard_instruction *
414 mir_last_in_block(struct midgard_block *block)
415 {
416 return list_last_entry(&block->instructions, struct midgard_instruction, link);
417 }
418
419 static inline midgard_block *
420 mir_get_block(compiler_context *ctx, int idx)
421 {
422 struct list_head *lst = &ctx->blocks;
423
424 while ((idx--) + 1)
425 lst = lst->next;
426
427 return (struct midgard_block *) lst;
428 }
429
430 static inline midgard_block *
431 mir_exit_block(struct compiler_context *ctx)
432 {
433 midgard_block *last = list_last_entry(&ctx->blocks,
434 struct midgard_block, link);
435
436 /* The last block must be empty logically but contains branch writeout
437 * for fragment shaders */
438
439 assert(last->nr_successors == 0);
440
441 return last;
442 }
443
444 static inline bool
445 mir_is_alu_bundle(midgard_bundle *bundle)
446 {
447 return IS_ALU(bundle->tag);
448 }
449
450 /* Registers/SSA are distinguish in the backend by the bottom-most bit */
451
452 #define IS_REG (1)
453
454 static inline unsigned
455 make_compiler_temp(compiler_context *ctx)
456 {
457 return (ctx->func->impl->ssa_alloc + ctx->temp_alloc++) << 1;
458 }
459
460 static inline unsigned
461 make_compiler_temp_reg(compiler_context *ctx)
462 {
463 return ((ctx->func->impl->reg_alloc + ctx->temp_alloc++) << 1) | IS_REG;
464 }
465
466 static inline unsigned
467 nir_src_index(compiler_context *ctx, nir_src *src)
468 {
469 if (src->is_ssa)
470 return (src->ssa->index << 1) | 0;
471 else {
472 assert(!src->reg.indirect);
473 return (src->reg.reg->index << 1) | IS_REG;
474 }
475 }
476
477 static inline unsigned
478 nir_alu_src_index(compiler_context *ctx, nir_alu_src *src)
479 {
480 return nir_src_index(ctx, &src->src);
481 }
482
483 static inline unsigned
484 nir_dest_index(compiler_context *ctx, nir_dest *dst)
485 {
486 if (dst->is_ssa)
487 return (dst->ssa.index << 1) | 0;
488 else {
489 assert(!dst->reg.indirect);
490 return (dst->reg.reg->index << 1) | IS_REG;
491 }
492 }
493
494
495
496 /* MIR manipulation */
497
498 unsigned mir_get_swizzle(midgard_instruction *ins, unsigned idx);
499 void mir_set_swizzle(midgard_instruction *ins, unsigned idx, unsigned new);
500 void mir_rewrite_index(compiler_context *ctx, unsigned old, unsigned new);
501 void mir_rewrite_index_src(compiler_context *ctx, unsigned old, unsigned new);
502 void mir_rewrite_index_dst(compiler_context *ctx, unsigned old, unsigned new);
503 void mir_rewrite_index_dst_single(midgard_instruction *ins, unsigned old, unsigned new);
504 void mir_rewrite_index_src_single(midgard_instruction *ins, unsigned old, unsigned new);
505 void mir_rewrite_index_src_swizzle(compiler_context *ctx, unsigned old, unsigned new, unsigned swizzle);
506 bool mir_single_use(compiler_context *ctx, unsigned value);
507 bool mir_special_index(compiler_context *ctx, unsigned idx);
508 unsigned mir_use_count(compiler_context *ctx, unsigned value);
509 bool mir_is_written_before(compiler_context *ctx, midgard_instruction *ins, unsigned node);
510 unsigned mir_mask_of_read_components(midgard_instruction *ins, unsigned node);
511 unsigned mir_ubo_shift(midgard_load_store_op op);
512
513 /* MIR printing */
514
515 void mir_print_instruction(midgard_instruction *ins);
516 void mir_print_bundle(midgard_bundle *ctx);
517 void mir_print_block(midgard_block *block);
518 void mir_print_shader(compiler_context *ctx);
519 bool mir_nontrivial_source2_mod(midgard_instruction *ins);
520 bool mir_nontrivial_source2_mod_simple(midgard_instruction *ins);
521 bool mir_nontrivial_mod(midgard_vector_alu_src src, bool is_int, unsigned mask);
522 bool mir_nontrivial_outmod(midgard_instruction *ins);
523
524 void mir_insert_instruction_before_scheduled(compiler_context *ctx, midgard_block *block, midgard_instruction *tag, midgard_instruction ins);
525 void mir_insert_instruction_after_scheduled(compiler_context *ctx, midgard_block *block, midgard_instruction *tag, midgard_instruction ins);
526
527 /* MIR goodies */
528
529 static const midgard_vector_alu_src blank_alu_src = {
530 .swizzle = SWIZZLE(COMPONENT_X, COMPONENT_Y, COMPONENT_Z, COMPONENT_W),
531 };
532
533 static const midgard_vector_alu_src blank_alu_src_xxxx = {
534 .swizzle = SWIZZLE(COMPONENT_X, COMPONENT_X, COMPONENT_X, COMPONENT_X),
535 };
536
537 static const midgard_scalar_alu_src blank_scalar_alu_src = {
538 .full = true
539 };
540
541 /* Used for encoding the unused source of 1-op instructions */
542 static const midgard_vector_alu_src zero_alu_src = { 0 };
543
544 /* 'Intrinsic' move for aliasing */
545
546 static inline midgard_instruction
547 v_mov(unsigned src, midgard_vector_alu_src mod, unsigned dest)
548 {
549 midgard_instruction ins = {
550 .type = TAG_ALU_4,
551 .mask = 0xF,
552 .src = { SSA_UNUSED, src, SSA_UNUSED },
553 .dest = dest,
554 .alu = {
555 .op = midgard_alu_op_imov,
556 .reg_mode = midgard_reg_mode_32,
557 .dest_override = midgard_dest_override_none,
558 .outmod = midgard_outmod_int_wrap,
559 .src1 = vector_alu_srco_unsigned(zero_alu_src),
560 .src2 = vector_alu_srco_unsigned(mod)
561 },
562 };
563
564 return ins;
565 }
566
567 static inline bool
568 mir_has_arg(midgard_instruction *ins, unsigned arg)
569 {
570 if (!ins)
571 return false;
572
573 for (unsigned i = 0; i < ARRAY_SIZE(ins->src); ++i) {
574 if (ins->src[i] == arg)
575 return true;
576 }
577
578 return false;
579 }
580
581 /* Scheduling */
582
583 void schedule_program(compiler_context *ctx);
584
585 /* Register allocation */
586
587 struct ra_graph;
588
589 /* Broad types of register classes so we can handle special
590 * registers */
591
592 #define NR_REG_CLASSES 6
593
594 #define REG_CLASS_WORK 0
595 #define REG_CLASS_LDST 1
596 #define REG_CLASS_LDST27 2
597 #define REG_CLASS_TEXR 3
598 #define REG_CLASS_TEXW 4
599 #define REG_CLASS_FRAGC 5
600
601 void mir_lower_special_reads(compiler_context *ctx);
602 struct ra_graph* allocate_registers(compiler_context *ctx, bool *spilled);
603 void install_registers(compiler_context *ctx, struct ra_graph *g);
604 bool mir_is_live_after(compiler_context *ctx, midgard_block *block, midgard_instruction *start, int src);
605 bool mir_has_multiple_writes(compiler_context *ctx, int src);
606
607 void mir_create_pipeline_registers(compiler_context *ctx);
608
609 void
610 midgard_promote_uniforms(compiler_context *ctx, unsigned promoted_count);
611
612 midgard_instruction *
613 emit_ubo_read(
614 compiler_context *ctx,
615 nir_instr *instr,
616 unsigned dest,
617 unsigned offset,
618 nir_src *indirect_offset,
619 unsigned index);
620
621 void
622 emit_sysval_read(compiler_context *ctx, nir_instr *instr, signed dest_override, unsigned nr_components);
623
624 void
625 midgard_emit_derivatives(compiler_context *ctx, nir_alu_instr *instr);
626
627 void
628 midgard_lower_derivatives(compiler_context *ctx, midgard_block *block);
629
630 bool mir_op_computes_derivatives(unsigned op);
631
632 /* Final emission */
633
634 void emit_binary_bundle(
635 compiler_context *ctx,
636 midgard_bundle *bundle,
637 struct util_dynarray *emission,
638 int next_tag);
639
640 bool
641 nir_undef_to_zero(nir_shader *shader);
642
643 /* Optimizations */
644
645 bool midgard_opt_copy_prop(compiler_context *ctx, midgard_block *block);
646 bool midgard_opt_combine_projection(compiler_context *ctx, midgard_block *block);
647 bool midgard_opt_varying_projection(compiler_context *ctx, midgard_block *block);
648 bool midgard_opt_dead_code_eliminate(compiler_context *ctx, midgard_block *block);
649 bool midgard_opt_dead_move_eliminate(compiler_context *ctx, midgard_block *block);
650
651 void midgard_lower_invert(compiler_context *ctx, midgard_block *block);
652 bool midgard_opt_not_propagate(compiler_context *ctx, midgard_block *block);
653 bool midgard_opt_fuse_src_invert(compiler_context *ctx, midgard_block *block);
654 bool midgard_opt_fuse_dest_invert(compiler_context *ctx, midgard_block *block);
655 bool midgard_opt_promote_fmov(compiler_context *ctx, midgard_block *block);
656
657 #endif