pan/mdg: Replace writeout booleans with a single value
[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 #include "midgard_ops.h"
31
32 #include "util/hash_table.h"
33 #include "util/u_dynarray.h"
34 #include "util/set.h"
35 #include "util/list.h"
36
37 #include "main/mtypes.h"
38 #include "compiler/nir_types.h"
39 #include "compiler/nir/nir.h"
40 #include "panfrost/util/pan_ir.h"
41 #include "panfrost/util/lcra.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 #define PAN_WRITEOUT_C 1
74 #define PAN_WRITEOUT_Z 2
75 #define PAN_WRITEOUT_S 4
76
77 /* Generic in-memory data type repesenting a single logical instruction, rather
78 * than a single instruction group. This is the preferred form for code gen.
79 * Multiple midgard_insturctions will later be combined during scheduling,
80 * though this is not represented in this structure. Its format bridges
81 * the low-level binary representation with the higher level semantic meaning.
82 *
83 * Notably, it allows registers to be specified as block local SSA, for code
84 * emitted before the register allocation pass.
85 */
86
87 #define MIR_SRC_COUNT 4
88 #define MIR_VEC_COMPONENTS 16
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 /* Instruction arguments represented as block-local SSA
97 * indices, rather than registers. ~0 means unused. */
98 unsigned src[MIR_SRC_COUNT];
99 unsigned dest;
100
101 /* vec16 swizzle, unpacked, per source */
102 unsigned swizzle[MIR_SRC_COUNT][MIR_VEC_COMPONENTS];
103
104 /* Types! */
105 nir_alu_type src_types[MIR_SRC_COUNT];
106 nir_alu_type dest_type;
107
108 /* Packing ops have non-32-bit dest types even though they functionally
109 * work at the 32-bit level, use this as a signal to disable copyprop.
110 * We maybe need synthetic pack ops instead. */
111 bool is_pack;
112
113 /* Modifiers, depending on type */
114 union {
115 struct {
116 bool src_abs[MIR_SRC_COUNT];
117 bool src_neg[MIR_SRC_COUNT];
118 };
119
120 struct {
121 bool src_shift[MIR_SRC_COUNT];
122 };
123 };
124
125 /* Out of the union for csel (could maybe be fixed..) */
126 bool src_invert[MIR_SRC_COUNT];
127
128 /* If the op supports it */
129 enum midgard_roundmode roundmode;
130
131 /* Special fields for an ALU instruction */
132 midgard_reg_info registers;
133
134 /* For textures: should helpers execute this instruction (instead of
135 * just helping with derivatives)? Should helpers terminate after? */
136 bool helper_terminate;
137 bool helper_execute;
138
139 /* I.e. (1 << alu_bit) */
140 int unit;
141
142 bool has_constants;
143 midgard_constants constants;
144 uint16_t inline_constant;
145 bool has_blend_constant;
146 bool has_inline_constant;
147
148 bool compact_branch;
149 uint8_t writeout;
150 bool last_writeout;
151
152 /* Masks in a saneish format. One bit per channel, not packed fancy.
153 * Use this instead of the op specific ones, and switch over at emit
154 * time */
155
156 uint16_t mask;
157
158 /* Hint for the register allocator not to spill the destination written
159 * from this instruction (because it is a spill/unspill node itself).
160 * Bitmask of spilled classes */
161
162 unsigned no_spill;
163
164 /* Generic hint for intra-pass use */
165 bool hint;
166
167 /* During scheduling, the backwards dependency graph
168 * (DAG). nr_dependencies is the number of unscheduled
169 * instructions that must still be scheduled after
170 * (before) this instruction. dependents are which
171 * instructions need to be scheduled before (after) this
172 * instruction. */
173
174 unsigned nr_dependencies;
175 BITSET_WORD *dependents;
176
177 union {
178 midgard_load_store_word load_store;
179 midgard_vector_alu alu;
180 midgard_texture_word texture;
181 midgard_branch_extended branch_extended;
182 uint16_t br_compact;
183
184 /* General branch, rather than packed br_compact. Higher level
185 * than the other components */
186 midgard_branch branch;
187 };
188 } midgard_instruction;
189
190 typedef struct midgard_block {
191 pan_block base;
192
193 bool scheduled;
194
195 /* List of midgard_bundles emitted (after the scheduler has run) */
196 struct util_dynarray bundles;
197
198 /* Number of quadwords _actually_ emitted, as determined after scheduling */
199 unsigned quadword_count;
200
201 /* Indicates this is a fixed-function fragment epilogue block */
202 bool epilogue;
203
204 /* Are helper invocations required by this block? */
205 bool helpers_in;
206 } midgard_block;
207
208 typedef struct midgard_bundle {
209 /* Tag for the overall bundle */
210 int tag;
211
212 /* Instructions contained by the bundle. instruction_count <= 6 (vmul,
213 * sadd, vadd, smul, vlut, branch) */
214 int instruction_count;
215 midgard_instruction *instructions[6];
216
217 /* Bundle-wide ALU configuration */
218 int padding;
219 int control;
220 bool has_embedded_constants;
221 midgard_constants constants;
222 bool has_blend_constant;
223 bool last_writeout;
224 } midgard_bundle;
225
226 enum midgard_rt_id {
227 MIDGARD_COLOR_RT0,
228 MIDGARD_COLOR_RT1,
229 MIDGARD_COLOR_RT2,
230 MIDGARD_COLOR_RT3,
231 MIDGARD_ZS_RT,
232 MIDGARD_NUM_RTS,
233 };
234
235 typedef struct compiler_context {
236 nir_shader *nir;
237 gl_shader_stage stage;
238
239 /* Is internally a blend shader? Depends on stage == FRAGMENT */
240 bool is_blend;
241
242 /* Render target number for a keyed blend shader. Depends on is_blend */
243 unsigned blend_rt;
244
245 /* Tracking for blend constant patching */
246 int blend_constant_offset;
247
248 /* Number of bytes used for Thread Local Storage */
249 unsigned tls_size;
250
251 /* Count of spills and fills for shaderdb */
252 unsigned spills;
253 unsigned fills;
254
255 /* Current NIR function */
256 nir_function *func;
257
258 /* Allocated compiler temporary counter */
259 unsigned temp_alloc;
260
261 /* Unordered list of midgard_blocks */
262 int block_count;
263 struct list_head blocks;
264
265 /* TODO merge with block_count? */
266 unsigned block_source_count;
267
268 /* List of midgard_instructions emitted for the current block */
269 midgard_block *current_block;
270
271 /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
272 midgard_block *after_block;
273
274 /* The current "depth" of the loop, for disambiguating breaks/continues
275 * when using nested loops */
276 int current_loop_depth;
277
278 /* Total number of loops for shader-db */
279 unsigned loop_count;
280
281 /* Constants which have been loaded, for later inlining */
282 struct hash_table_u64 *ssa_constants;
283
284 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
285 struct hash_table_u64 *hash_to_temp;
286 int temp_count;
287 int max_hash;
288
289 /* Set of NIR indices that were already emitted as outmods */
290 BITSET_WORD *already_emitted;
291
292 /* Just the count of the max register used. Higher count => higher
293 * register pressure */
294 int work_registers;
295
296 /* The number of uniforms allowable for the fast path */
297 int uniform_cutoff;
298
299 /* Count of instructions emitted from NIR overall, across all blocks */
300 int instruction_count;
301
302 /* Alpha ref value passed in */
303 float alpha_ref;
304
305 unsigned quadword_count;
306
307 /* Bitmask of valid metadata */
308 unsigned metadata;
309
310 /* Model-specific quirk set */
311 uint32_t quirks;
312
313 /* Writeout instructions for each render target */
314 midgard_instruction *writeout_branch[MIDGARD_NUM_RTS];
315
316 struct panfrost_sysvals sysvals;
317 } compiler_context;
318
319 /* Per-block live_in/live_out */
320 #define MIDGARD_METADATA_LIVENESS (1 << 0)
321
322 /* Helpers for manipulating the above structures (forming the driver IR) */
323
324 /* Append instruction to end of current block */
325
326 static inline midgard_instruction *
327 mir_upload_ins(struct compiler_context *ctx, struct midgard_instruction ins)
328 {
329 midgard_instruction *heap = ralloc(ctx, struct midgard_instruction);
330 memcpy(heap, &ins, sizeof(ins));
331 return heap;
332 }
333
334 static inline midgard_instruction *
335 emit_mir_instruction(struct compiler_context *ctx, struct midgard_instruction ins)
336 {
337 midgard_instruction *u = mir_upload_ins(ctx, ins);
338 list_addtail(&u->link, &ctx->current_block->base.instructions);
339 return u;
340 }
341
342 static inline struct midgard_instruction *
343 mir_insert_instruction_before(struct compiler_context *ctx,
344 struct midgard_instruction *tag,
345 struct midgard_instruction ins)
346 {
347 struct midgard_instruction *u = mir_upload_ins(ctx, ins);
348 list_addtail(&u->link, &tag->link);
349 return u;
350 }
351
352 static inline void
353 mir_remove_instruction(struct midgard_instruction *ins)
354 {
355 list_del(&ins->link);
356 }
357
358 static inline midgard_instruction*
359 mir_prev_op(struct midgard_instruction *ins)
360 {
361 return list_last_entry(&(ins->link), midgard_instruction, link);
362 }
363
364 static inline midgard_instruction*
365 mir_next_op(struct midgard_instruction *ins)
366 {
367 return list_first_entry(&(ins->link), midgard_instruction, link);
368 }
369
370 #define mir_foreach_block(ctx, v) \
371 list_for_each_entry(pan_block, v, &ctx->blocks, link)
372
373 #define mir_foreach_block_from(ctx, from, v) \
374 list_for_each_entry_from(pan_block, v, &from->base, &ctx->blocks, link)
375
376 #define mir_foreach_instr_in_block(block, v) \
377 list_for_each_entry(struct midgard_instruction, v, &block->base.instructions, link)
378 #define mir_foreach_instr_in_block_rev(block, v) \
379 list_for_each_entry_rev(struct midgard_instruction, v, &block->base.instructions, link)
380
381 #define mir_foreach_instr_in_block_safe(block, v) \
382 list_for_each_entry_safe(struct midgard_instruction, v, &block->base.instructions, link)
383
384 #define mir_foreach_instr_in_block_safe_rev(block, v) \
385 list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->base.instructions, link)
386
387 #define mir_foreach_instr_in_block_from(block, v, from) \
388 list_for_each_entry_from(struct midgard_instruction, v, from, &block->base.instructions, link)
389
390 #define mir_foreach_instr_in_block_from_rev(block, v, from) \
391 list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->base.instructions, link)
392
393 #define mir_foreach_bundle_in_block(block, v) \
394 util_dynarray_foreach(&block->bundles, midgard_bundle, v)
395
396 #define mir_foreach_bundle_in_block_rev(block, v) \
397 util_dynarray_foreach_reverse(&block->bundles, midgard_bundle, v)
398
399 #define mir_foreach_instr_in_block_scheduled_rev(block, v) \
400 midgard_instruction* v; \
401 signed i = 0; \
402 mir_foreach_bundle_in_block_rev(block, _bundle) \
403 for (i = (_bundle->instruction_count - 1), v = _bundle->instructions[i]; \
404 i >= 0; \
405 --i, v = (i >= 0) ? _bundle->instructions[i] : NULL) \
406
407 #define mir_foreach_instr_global(ctx, v) \
408 mir_foreach_block(ctx, v_block) \
409 mir_foreach_instr_in_block(((midgard_block *) v_block), v)
410
411 #define mir_foreach_instr_global_safe(ctx, v) \
412 mir_foreach_block(ctx, v_block) \
413 mir_foreach_instr_in_block_safe(((midgard_block *) v_block), v)
414
415 /* Based on set_foreach, expanded with automatic type casts */
416
417 #define mir_foreach_predecessor(blk, v) \
418 struct set_entry *_entry_##v; \
419 struct midgard_block *v; \
420 for (_entry_##v = _mesa_set_next_entry(blk->base.predecessors, NULL), \
421 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL); \
422 _entry_##v != NULL; \
423 _entry_##v = _mesa_set_next_entry(blk->base.predecessors, _entry_##v), \
424 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
425
426 #define mir_foreach_src(ins, v) \
427 for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
428
429 static inline midgard_instruction *
430 mir_last_in_block(struct midgard_block *block)
431 {
432 return list_last_entry(&block->base.instructions, struct midgard_instruction, link);
433 }
434
435 static inline midgard_block *
436 mir_get_block(compiler_context *ctx, int idx)
437 {
438 struct list_head *lst = &ctx->blocks;
439
440 while ((idx--) + 1)
441 lst = lst->next;
442
443 return (struct midgard_block *) lst;
444 }
445
446 static inline bool
447 mir_is_alu_bundle(midgard_bundle *bundle)
448 {
449 return IS_ALU(bundle->tag);
450 }
451
452 static inline unsigned
453 make_compiler_temp(compiler_context *ctx)
454 {
455 return (ctx->func->impl->ssa_alloc + ctx->temp_alloc++) << 1;
456 }
457
458 static inline unsigned
459 make_compiler_temp_reg(compiler_context *ctx)
460 {
461 return ((ctx->func->impl->reg_alloc + ctx->temp_alloc++) << 1) | PAN_IS_REG;
462 }
463
464 static inline unsigned
465 nir_ssa_index(nir_ssa_def *ssa)
466 {
467 return (ssa->index << 1) | 0;
468 }
469
470 static inline unsigned
471 nir_src_index(compiler_context *ctx, nir_src *src)
472 {
473 if (src->is_ssa)
474 return nir_ssa_index(src->ssa);
475 else {
476 assert(!src->reg.indirect);
477 return (src->reg.reg->index << 1) | PAN_IS_REG;
478 }
479 }
480
481 static inline unsigned
482 nir_dest_index(nir_dest *dst)
483 {
484 if (dst->is_ssa)
485 return (dst->ssa.index << 1) | 0;
486 else {
487 assert(!dst->reg.indirect);
488 return (dst->reg.reg->index << 1) | PAN_IS_REG;
489 }
490 }
491
492
493
494 /* MIR manipulation */
495
496 void mir_rewrite_index(compiler_context *ctx, unsigned old, unsigned new);
497 void mir_rewrite_index_src(compiler_context *ctx, unsigned old, unsigned new);
498 void mir_rewrite_index_dst(compiler_context *ctx, unsigned old, unsigned new);
499 void mir_rewrite_index_dst_single(midgard_instruction *ins, unsigned old, unsigned new);
500 void mir_rewrite_index_src_single(midgard_instruction *ins, unsigned old, unsigned new);
501 void mir_rewrite_index_src_swizzle(compiler_context *ctx, unsigned old, unsigned new, unsigned *swizzle);
502 bool mir_single_use(compiler_context *ctx, unsigned value);
503 unsigned mir_use_count(compiler_context *ctx, unsigned value);
504 uint16_t mir_bytemask_of_read_components(midgard_instruction *ins, unsigned node);
505 uint16_t mir_bytemask_of_read_components_index(midgard_instruction *ins, unsigned i);
506 uint16_t mir_from_bytemask(uint16_t bytemask, unsigned bits);
507 uint16_t mir_bytemask(midgard_instruction *ins);
508 uint16_t mir_round_bytemask_up(uint16_t mask, unsigned bits);
509 void mir_set_bytemask(midgard_instruction *ins, uint16_t bytemask);
510 signed mir_upper_override(midgard_instruction *ins, unsigned inst_size);
511 unsigned mir_components_for_type(nir_alu_type T);
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_mod(midgard_instruction *ins, unsigned i, bool check_swizzle);
520 bool mir_nontrivial_outmod(midgard_instruction *ins);
521
522 void mir_insert_instruction_before_scheduled(compiler_context *ctx, midgard_block *block, midgard_instruction *tag, midgard_instruction ins);
523 void mir_insert_instruction_after_scheduled(compiler_context *ctx, midgard_block *block, midgard_instruction *tag, midgard_instruction ins);
524 void mir_flip(midgard_instruction *ins);
525 void mir_compute_temp_count(compiler_context *ctx);
526
527 void mir_set_offset(compiler_context *ctx, midgard_instruction *ins, nir_src *offset, bool is_shared);
528
529 /* 'Intrinsic' move for aliasing */
530
531 static inline midgard_instruction
532 v_mov(unsigned src, unsigned dest)
533 {
534 midgard_instruction ins = {
535 .type = TAG_ALU_4,
536 .mask = 0xF,
537 .src = { ~0, src, ~0, ~0 },
538 .src_types = { 0, nir_type_uint32 },
539 .swizzle = SWIZZLE_IDENTITY,
540 .dest = dest,
541 .dest_type = nir_type_uint32,
542 .alu = {
543 .op = midgard_alu_op_imov,
544 .reg_mode = midgard_reg_mode_32,
545 .outmod = midgard_outmod_int_wrap
546 },
547 };
548
549 return ins;
550 }
551
552 /* Broad types of register classes so we can handle special
553 * registers */
554
555 #define REG_CLASS_WORK 0
556 #define REG_CLASS_LDST 1
557 #define REG_CLASS_TEXR 3
558 #define REG_CLASS_TEXW 4
559
560 /* Like a move, but to thread local storage! */
561
562 static inline midgard_instruction
563 v_load_store_scratch(
564 unsigned srcdest,
565 unsigned index,
566 bool is_store,
567 unsigned mask)
568 {
569 /* We index by 32-bit vec4s */
570 unsigned byte = (index * 4 * 4);
571
572 midgard_instruction ins = {
573 .type = TAG_LOAD_STORE_4,
574 .mask = mask,
575 .dest_type = nir_type_uint32,
576 .dest = ~0,
577 .src = { ~0, ~0, ~0, ~0 },
578 .swizzle = SWIZZLE_IDENTITY_4,
579 .load_store = {
580 .op = is_store ? midgard_op_st_int4 : midgard_op_ld_int4,
581
582 /* For register spilling - to thread local storage */
583 .arg_1 = 0xEA,
584 .arg_2 = 0x1E,
585 },
586
587 /* If we spill an unspill, RA goes into an infinite loop */
588 .no_spill = (1 << REG_CLASS_WORK)
589 };
590
591 ins.constants.u32[0] = byte;
592
593 if (is_store) {
594 ins.src[0] = srcdest;
595 ins.src_types[0] = nir_type_uint32;
596
597 /* Ensure we are tightly swizzled so liveness analysis is
598 * correct */
599
600 for (unsigned i = 0; i < 4; ++i) {
601 if (!(mask & (1 << i)))
602 ins.swizzle[0][i] = COMPONENT_X;
603 }
604 } else
605 ins.dest = srcdest;
606
607 return ins;
608 }
609
610 static inline bool
611 mir_has_arg(midgard_instruction *ins, unsigned arg)
612 {
613 if (!ins)
614 return false;
615
616 mir_foreach_src(ins, i) {
617 if (ins->src[i] == arg)
618 return true;
619 }
620
621 return false;
622 }
623
624 /* Scheduling */
625
626 void midgard_schedule_program(compiler_context *ctx);
627
628 void mir_ra(compiler_context *ctx);
629 void mir_squeeze_index(compiler_context *ctx);
630 void mir_lower_special_reads(compiler_context *ctx);
631 void mir_liveness_ins_update(uint16_t *live, midgard_instruction *ins, unsigned max);
632 void mir_compute_liveness(compiler_context *ctx);
633 void mir_invalidate_liveness(compiler_context *ctx);
634 bool mir_is_live_after(compiler_context *ctx, midgard_block *block, midgard_instruction *start, int src);
635
636 void mir_create_pipeline_registers(compiler_context *ctx);
637 void midgard_promote_uniforms(compiler_context *ctx);
638
639 void
640 midgard_emit_derivatives(compiler_context *ctx, nir_alu_instr *instr);
641
642 void
643 midgard_lower_derivatives(compiler_context *ctx, midgard_block *block);
644
645 bool mir_op_computes_derivatives(gl_shader_stage stage, unsigned op);
646
647 void mir_analyze_helper_terminate(compiler_context *ctx);
648 void mir_analyze_helper_requirements(compiler_context *ctx);
649
650 /* Final emission */
651
652 void emit_binary_bundle(
653 compiler_context *ctx,
654 midgard_block *block,
655 midgard_bundle *bundle,
656 struct util_dynarray *emission,
657 int next_tag);
658
659 bool
660 nir_undef_to_zero(nir_shader *shader);
661 bool nir_fuse_io_16(nir_shader *shader);
662
663 void midgard_nir_lod_errata(nir_shader *shader);
664
665 /* Optimizations */
666
667 bool midgard_opt_copy_prop(compiler_context *ctx, midgard_block *block);
668 bool midgard_opt_combine_projection(compiler_context *ctx, midgard_block *block);
669 bool midgard_opt_varying_projection(compiler_context *ctx, midgard_block *block);
670 bool midgard_opt_dead_code_eliminate(compiler_context *ctx);
671 bool midgard_opt_dead_move_eliminate(compiler_context *ctx, midgard_block *block);
672
673 #endif