2 * Copyright (C) 2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
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:
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
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
24 #ifndef _MDG_COMPILER_H
25 #define _MDG_COMPILER_H
29 #include "midgard_compile.h"
32 #include "util/hash_table.h"
33 #include "util/u_dynarray.h"
35 #include "util/list.h"
37 #include "main/mtypes.h"
38 #include "compiler/nir_types.h"
39 #include "compiler/nir/nir.h"
44 /* Target types. Defaults to TARGET_GOTO (the type corresponding directly to
45 * the hardware), hence why that must be zero. TARGET_DISCARD signals this
46 * instruction is actually a discard op. */
49 #define TARGET_BREAK 1
50 #define TARGET_CONTINUE 2
51 #define TARGET_DISCARD 3
53 typedef struct midgard_branch
{
54 /* If conditional, the condition is specified in r31.w */
57 /* For conditionals, if this is true, we branch on FALSE. If false, we branch on TRUE. */
58 bool invert_conditional
;
60 /* 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 /* The actual target */
71 /* Generic in-memory data type repesenting a single logical instruction, rather
72 * than a single instruction group. This is the preferred form for code gen.
73 * Multiple midgard_insturctions will later be combined during scheduling,
74 * though this is not represented in this structure. Its format bridges
75 * the low-level binary representation with the higher level semantic meaning.
77 * Notably, it allows registers to be specified as block local SSA, for code
78 * emitted before the register allocation pass.
81 #define MIR_SRC_COUNT 3
82 #define MIR_VEC_COMPONENTS 16
84 typedef struct midgard_instruction
{
85 /* Must be first for casting */
86 struct list_head link
;
88 unsigned type
; /* ALU, load/store, texture */
90 /* Instruction arguments represented as block-local SSA
91 * indices, rather than registers. ~0 means unused. */
95 /* vec16 swizzle, unpacked, per source */
96 unsigned swizzle
[MIR_SRC_COUNT
][MIR_VEC_COMPONENTS
];
98 /* Special fields for an ALU instruction */
99 midgard_reg_info registers
;
101 /* I.e. (1 << alu_bit) */
105 uint32_t constants
[4];
106 uint16_t inline_constant
;
107 bool has_blend_constant
;
108 bool has_inline_constant
;
112 bool prepacked_branch
;
114 /* Kind of a hack, but hint against aggressive DCE */
117 /* Masks in a saneish format. One bit per channel, not packed fancy.
118 * Use this instead of the op specific ones, and switch over at emit
123 /* For ALU ops only: set to true to invert (bitwise NOT) the
124 * destination of an integer-out op. Not imeplemented in hardware but
125 * allows more optimizations */
129 /* Hint for the register allocator not to spill the destination written
130 * from this instruction (because it is a spill/unspill node itself).
131 * Bitmask of spilled classes */
135 /* Generic hint for intra-pass use */
138 /* During scheduling, the backwards dependency graph
139 * (DAG). nr_dependencies is the number of unscheduled
140 * instructions that must still be scheduled after
141 * (before) this instruction. dependents are which
142 * instructions need to be scheduled before (after) this
145 unsigned nr_dependencies
;
146 BITSET_WORD
*dependents
;
148 /* For load/store ops.. force 64-bit destination */
152 midgard_load_store_word load_store
;
153 midgard_vector_alu alu
;
154 midgard_texture_word texture
;
155 midgard_branch_extended branch_extended
;
158 /* General branch, rather than packed br_compact. Higher level
159 * than the other components */
160 midgard_branch branch
;
162 } midgard_instruction
;
164 typedef struct midgard_block
{
165 /* Link to next block. Must be first for mir_get_block */
166 struct list_head link
;
168 /* List of midgard_instructions emitted for the current block */
169 struct list_head instructions
;
171 /* Index of the block in source order */
176 /* List of midgard_bundles emitted (after the scheduler has run) */
177 struct util_dynarray bundles
;
179 /* Number of quadwords _actually_ emitted, as determined after scheduling */
180 unsigned quadword_count
;
182 /* Succeeding blocks. The compiler should not necessarily rely on
183 * source-order traversal */
184 struct midgard_block
*successors
[2];
185 unsigned nr_successors
;
187 struct set
*predecessors
;
189 /* The successors pointer form a graph, and in the case of
190 * complex control flow, this graph has a cycles. To aid
191 * traversal during liveness analysis, we have a visited?
192 * boolean for passes to use as they see fit, provided they
196 /* In liveness analysis, these are live masks (per-component) for
197 * indices for the block. Scalar compilers have the luxury of using
198 * simple bit fields, but for us, liveness is a vector idea. */
203 typedef struct midgard_bundle
{
204 /* Tag for the overall bundle */
207 /* Instructions contained by the bundle. instruction_count <= 6 (vmul,
208 * sadd, vadd, smul, vlut, branch) */
209 int instruction_count
;
210 midgard_instruction
*instructions
[6];
212 /* Bundle-wide ALU configuration */
215 bool has_embedded_constants
;
217 bool has_blend_constant
;
220 typedef struct compiler_context
{
222 gl_shader_stage stage
;
224 /* Is internally a blend shader? Depends on stage == FRAGMENT */
227 /* Render target number for a keyed blend shader. Depends on is_blend */
230 /* Tracking for blend constant patching */
231 int blend_constant_offset
;
233 /* Number of bytes used for Thread Local Storage */
236 /* Count of spills and fills for shaderdb */
240 /* Current NIR function */
243 /* Allocated compiler temporary counter */
246 /* Unordered list of midgard_blocks */
248 struct list_head blocks
;
250 /* TODO merge with block_count? */
251 unsigned block_source_count
;
253 /* List of midgard_instructions emitted for the current block */
254 midgard_block
*current_block
;
256 /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
257 midgard_block
*after_block
;
259 /* The current "depth" of the loop, for disambiguating breaks/continues
260 * when using nested loops */
261 int current_loop_depth
;
263 /* Total number of loops for shader-db */
266 /* Constants which have been loaded, for later inlining */
267 struct hash_table_u64
*ssa_constants
;
269 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
270 struct hash_table_u64
*hash_to_temp
;
274 /* Just the count of the max register used. Higher count => higher
275 * register pressure */
278 /* Used for cont/last hinting. Increase when a tex op is added.
279 * Decrease when a tex op is removed. */
280 int texture_op_count
;
282 /* The number of uniforms allowable for the fast path */
285 /* Count of instructions emitted from NIR overall, across all blocks */
286 int instruction_count
;
288 /* Alpha ref value passed in */
291 unsigned quadword_count
;
293 /* The mapping of sysvals to uniforms, the count, and the off-by-one inverse */
294 unsigned sysvals
[MAX_SYSVAL_COUNT
];
295 unsigned sysval_count
;
296 struct hash_table_u64
*sysval_to_id
;
298 /* Bitmask of valid metadata */
301 /* Model-specific quirk set */
305 /* Per-block live_in/live_out */
306 #define MIDGARD_METADATA_LIVENESS (1 << 0)
308 /* Helpers for manipulating the above structures (forming the driver IR) */
310 /* Append instruction to end of current block */
312 static inline midgard_instruction
*
313 mir_upload_ins(struct compiler_context
*ctx
, struct midgard_instruction ins
)
315 midgard_instruction
*heap
= ralloc(ctx
, struct midgard_instruction
);
316 memcpy(heap
, &ins
, sizeof(ins
));
320 static inline midgard_instruction
*
321 emit_mir_instruction(struct compiler_context
*ctx
, struct midgard_instruction ins
)
323 midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
324 list_addtail(&u
->link
, &ctx
->current_block
->instructions
);
328 static inline struct midgard_instruction
*
329 mir_insert_instruction_before(struct compiler_context
*ctx
,
330 struct midgard_instruction
*tag
,
331 struct midgard_instruction ins
)
333 struct midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
334 list_addtail(&u
->link
, &tag
->link
);
339 mir_remove_instruction(struct midgard_instruction
*ins
)
341 list_del(&ins
->link
);
344 static inline midgard_instruction
*
345 mir_prev_op(struct midgard_instruction
*ins
)
347 return list_last_entry(&(ins
->link
), midgard_instruction
, link
);
350 static inline midgard_instruction
*
351 mir_next_op(struct midgard_instruction
*ins
)
353 return list_first_entry(&(ins
->link
), midgard_instruction
, link
);
356 #define mir_foreach_block(ctx, v) \
357 list_for_each_entry(struct midgard_block, v, &ctx->blocks, link)
359 #define mir_foreach_block_from(ctx, from, v) \
360 list_for_each_entry_from(struct midgard_block, v, from, &ctx->blocks, link)
362 #define mir_foreach_instr(ctx, v) \
363 list_for_each_entry(struct midgard_instruction, v, &ctx->current_block->instructions, link)
365 #define mir_foreach_instr_safe(ctx, v) \
366 list_for_each_entry_safe(struct midgard_instruction, v, &ctx->current_block->instructions, link)
368 #define mir_foreach_instr_in_block(block, v) \
369 list_for_each_entry(struct midgard_instruction, v, &block->instructions, link)
370 #define mir_foreach_instr_in_block_rev(block, v) \
371 list_for_each_entry_rev(struct midgard_instruction, v, &block->instructions, link)
373 #define mir_foreach_instr_in_block_safe(block, v) \
374 list_for_each_entry_safe(struct midgard_instruction, v, &block->instructions, link)
376 #define mir_foreach_instr_in_block_safe_rev(block, v) \
377 list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->instructions, link)
379 #define mir_foreach_instr_in_block_from(block, v, from) \
380 list_for_each_entry_from(struct midgard_instruction, v, from, &block->instructions, link)
382 #define mir_foreach_instr_in_block_from_rev(block, v, from) \
383 list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->instructions, link)
385 #define mir_foreach_bundle_in_block(block, v) \
386 util_dynarray_foreach(&block->bundles, midgard_bundle, v)
388 #define mir_foreach_bundle_in_block_rev(block, v) \
389 util_dynarray_foreach_reverse(&block->bundles, midgard_bundle, v)
391 #define mir_foreach_instr_in_block_scheduled_rev(block, v) \
392 midgard_instruction* v; \
394 mir_foreach_bundle_in_block_rev(block, _bundle) \
395 for (i = (_bundle->instruction_count - 1), v = _bundle->instructions[i]; \
397 --i, v = (i >= 0) ? _bundle->instructions[i] : NULL) \
399 #define mir_foreach_instr_global(ctx, v) \
400 mir_foreach_block(ctx, v_block) \
401 mir_foreach_instr_in_block(v_block, v)
403 #define mir_foreach_instr_global_safe(ctx, v) \
404 mir_foreach_block(ctx, v_block) \
405 mir_foreach_instr_in_block_safe(v_block, v)
407 #define mir_foreach_successor(blk, v) \
408 struct midgard_block *v; \
409 struct midgard_block **_v; \
410 for (_v = &blk->successors[0], \
412 v != NULL && _v < &blk->successors[2]; \
415 /* Based on set_foreach, expanded with automatic type casts */
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->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->predecessors, _entry_##v), \
424 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
426 #define mir_foreach_src(ins, v) \
427 for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
429 static inline midgard_instruction
*
430 mir_last_in_block(struct midgard_block
*block
)
432 return list_last_entry(&block
->instructions
, struct midgard_instruction
, link
);
435 static inline midgard_block
*
436 mir_get_block(compiler_context
*ctx
, int idx
)
438 struct list_head
*lst
= &ctx
->blocks
;
443 return (struct midgard_block
*) lst
;
446 static inline midgard_block
*
447 mir_exit_block(struct compiler_context
*ctx
)
449 midgard_block
*last
= list_last_entry(&ctx
->blocks
,
450 struct midgard_block
, link
);
452 /* The last block must be empty logically but contains branch writeout
453 * for fragment shaders */
455 assert(last
->nr_successors
== 0);
461 mir_is_alu_bundle(midgard_bundle
*bundle
)
463 return IS_ALU(bundle
->tag
);
466 /* Registers/SSA are distinguish in the backend by the bottom-most bit */
470 static inline unsigned
471 make_compiler_temp(compiler_context
*ctx
)
473 return (ctx
->func
->impl
->ssa_alloc
+ ctx
->temp_alloc
++) << 1;
476 static inline unsigned
477 make_compiler_temp_reg(compiler_context
*ctx
)
479 return ((ctx
->func
->impl
->reg_alloc
+ ctx
->temp_alloc
++) << 1) | IS_REG
;
482 static inline unsigned
483 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
486 return (src
->ssa
->index
<< 1) | 0;
488 assert(!src
->reg
.indirect
);
489 return (src
->reg
.reg
->index
<< 1) | IS_REG
;
493 static inline unsigned
494 nir_alu_src_index(compiler_context
*ctx
, nir_alu_src
*src
)
496 return nir_src_index(ctx
, &src
->src
);
499 static inline unsigned
500 nir_dest_index(compiler_context
*ctx
, nir_dest
*dst
)
503 return (dst
->ssa
.index
<< 1) | 0;
505 assert(!dst
->reg
.indirect
);
506 return (dst
->reg
.reg
->index
<< 1) | IS_REG
;
512 /* MIR manipulation */
514 void mir_rewrite_index(compiler_context
*ctx
, unsigned old
, unsigned new);
515 void mir_rewrite_index_src(compiler_context
*ctx
, unsigned old
, unsigned new);
516 void mir_rewrite_index_dst(compiler_context
*ctx
, unsigned old
, unsigned new);
517 void mir_rewrite_index_dst_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
518 void mir_rewrite_index_src_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
519 void mir_rewrite_index_src_swizzle(compiler_context
*ctx
, unsigned old
, unsigned new, unsigned *swizzle
);
520 bool mir_single_use(compiler_context
*ctx
, unsigned value
);
521 bool mir_special_index(compiler_context
*ctx
, unsigned idx
);
522 unsigned mir_use_count(compiler_context
*ctx
, unsigned value
);
523 bool mir_is_written_before(compiler_context
*ctx
, midgard_instruction
*ins
, unsigned node
);
524 uint16_t mir_bytemask_of_read_components(midgard_instruction
*ins
, unsigned node
);
525 midgard_reg_mode
mir_typesize(midgard_instruction
*ins
);
526 midgard_reg_mode
mir_srcsize(midgard_instruction
*ins
, unsigned i
);
527 unsigned mir_bytes_for_mode(midgard_reg_mode mode
);
528 midgard_reg_mode
mir_mode_for_destsize(unsigned size
);
529 uint16_t mir_from_bytemask(uint16_t bytemask
, midgard_reg_mode mode
);
530 uint16_t mir_to_bytemask(midgard_reg_mode mode
, unsigned mask
);
531 uint16_t mir_bytemask(midgard_instruction
*ins
);
532 uint16_t mir_round_bytemask_down(uint16_t mask
, midgard_reg_mode mode
);
533 void mir_set_bytemask(midgard_instruction
*ins
, uint16_t bytemask
);
537 void mir_print_instruction(midgard_instruction
*ins
);
538 void mir_print_bundle(midgard_bundle
*ctx
);
539 void mir_print_block(midgard_block
*block
);
540 void mir_print_shader(compiler_context
*ctx
);
541 bool mir_nontrivial_source2_mod(midgard_instruction
*ins
);
542 bool mir_nontrivial_source2_mod_simple(midgard_instruction
*ins
);
543 bool mir_nontrivial_outmod(midgard_instruction
*ins
);
545 void mir_insert_instruction_before_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
546 void mir_insert_instruction_after_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
547 void mir_flip(midgard_instruction
*ins
);
548 void mir_compute_temp_count(compiler_context
*ctx
);
550 /* 'Intrinsic' move for aliasing */
552 static inline midgard_instruction
553 v_mov(unsigned src
, unsigned dest
)
555 midgard_instruction ins
= {
558 .src
= { SSA_UNUSED
, src
, SSA_UNUSED
},
559 .swizzle
= SWIZZLE_IDENTITY
,
562 .op
= midgard_alu_op_imov
,
563 .reg_mode
= midgard_reg_mode_32
,
564 .dest_override
= midgard_dest_override_none
,
565 .outmod
= midgard_outmod_int_wrap
572 /* Broad types of register classes so we can handle special
575 #define REG_CLASS_WORK 0
576 #define REG_CLASS_LDST 1
577 #define REG_CLASS_TEXR 3
578 #define REG_CLASS_TEXW 4
580 /* Like a move, but to thread local storage! */
582 static inline midgard_instruction
583 v_load_store_scratch(
589 /* We index by 32-bit vec4s */
590 unsigned byte
= (index
* 4 * 4);
592 midgard_instruction ins
= {
593 .type
= TAG_LOAD_STORE_4
,
596 .src
= { ~0, ~0, ~0 },
597 .swizzle
= SWIZZLE_IDENTITY_4
,
599 .op
= is_store
? midgard_op_st_int4
: midgard_op_ld_int4
,
601 /* For register spilling - to thread local storage */
606 /* If we spill an unspill, RA goes into an infinite loop */
607 .no_spill
= (1 << REG_CLASS_WORK
)
610 ins
.constants
[0] = byte
;
613 ins
.src
[0] = srcdest
;
615 /* Ensure we are tightly swizzled so liveness analysis is
618 for (unsigned i
= 0; i
< 4; ++i
) {
619 if (!(mask
& (1 << i
)))
620 ins
.swizzle
[0][i
] = COMPONENT_X
;
629 mir_has_arg(midgard_instruction
*ins
, unsigned arg
)
634 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
->src
); ++i
) {
635 if (ins
->src
[i
] == arg
)
644 void schedule_program(compiler_context
*ctx
);
646 void mir_ra(compiler_context
*ctx
);
647 void mir_squeeze_index(compiler_context
*ctx
);
648 void mir_lower_special_reads(compiler_context
*ctx
);
649 void mir_liveness_ins_update(uint16_t *live
, midgard_instruction
*ins
, unsigned max
);
650 void mir_compute_liveness(compiler_context
*ctx
);
651 void mir_invalidate_liveness(compiler_context
*ctx
);
652 bool mir_is_live_after(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*start
, int src
);
654 void mir_create_pipeline_registers(compiler_context
*ctx
);
657 midgard_promote_uniforms(compiler_context
*ctx
, unsigned promoted_count
);
659 midgard_instruction
*
661 compiler_context
*ctx
,
665 nir_src
*indirect_offset
,
669 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
, signed dest_override
, unsigned nr_components
);
672 midgard_emit_derivatives(compiler_context
*ctx
, nir_alu_instr
*instr
);
675 midgard_lower_derivatives(compiler_context
*ctx
, midgard_block
*block
);
677 bool mir_op_computes_derivatives(gl_shader_stage stage
, unsigned op
);
681 void emit_binary_bundle(
682 compiler_context
*ctx
,
683 midgard_bundle
*bundle
,
684 struct util_dynarray
*emission
,
688 nir_undef_to_zero(nir_shader
*shader
);
690 void midgard_nir_lod_errata(nir_shader
*shader
);
694 bool midgard_opt_copy_prop(compiler_context
*ctx
, midgard_block
*block
);
695 bool midgard_opt_combine_projection(compiler_context
*ctx
, midgard_block
*block
);
696 bool midgard_opt_varying_projection(compiler_context
*ctx
, midgard_block
*block
);
697 bool midgard_opt_dead_code_eliminate(compiler_context
*ctx
, midgard_block
*block
);
698 bool midgard_opt_dead_move_eliminate(compiler_context
*ctx
, midgard_block
*block
);
700 void midgard_lower_invert(compiler_context
*ctx
, midgard_block
*block
);
701 bool midgard_opt_not_propagate(compiler_context
*ctx
, midgard_block
*block
);
702 bool midgard_opt_fuse_src_invert(compiler_context
*ctx
, midgard_block
*block
);
703 bool midgard_opt_fuse_dest_invert(compiler_context
*ctx
, midgard_block
*block
);
704 bool midgard_opt_csel_invert(compiler_context
*ctx
, midgard_block
*block
);
705 bool midgard_opt_promote_fmov(compiler_context
*ctx
, midgard_block
*block
);