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"
30 #include "midgard_ops.h"
33 #include "util/hash_table.h"
34 #include "util/u_dynarray.h"
36 #include "util/list.h"
38 #include "main/mtypes.h"
39 #include "compiler/nir_types.h"
40 #include "compiler/nir/nir.h"
45 /* Target types. Defaults to TARGET_GOTO (the type corresponding directly to
46 * the hardware), hence why that must be zero. TARGET_DISCARD signals this
47 * instruction is actually a discard op. */
50 #define TARGET_BREAK 1
51 #define TARGET_CONTINUE 2
52 #define TARGET_DISCARD 3
54 typedef struct midgard_branch
{
55 /* If conditional, the condition is specified in r31.w */
58 /* For conditionals, if this is true, we branch on FALSE. If false, we branch on TRUE. */
59 bool invert_conditional
;
61 /* Branch targets: the start of a block, the start of a loop (continue), the end of a loop (break). Value is one of TARGET_ */
64 /* The actual target */
72 /* Generic in-memory data type repesenting a single logical instruction, rather
73 * than a single instruction group. This is the preferred form for code gen.
74 * Multiple midgard_insturctions will later be combined during scheduling,
75 * though this is not represented in this structure. Its format bridges
76 * the low-level binary representation with the higher level semantic meaning.
78 * Notably, it allows registers to be specified as block local SSA, for code
79 * emitted before the register allocation pass.
82 #define MIR_SRC_COUNT 4
83 #define MIR_VEC_COMPONENTS 16
85 typedef struct midgard_instruction
{
86 /* Must be first for casting */
87 struct list_head link
;
89 unsigned type
; /* ALU, load/store, texture */
91 /* Instruction arguments represented as block-local SSA
92 * indices, rather than registers. ~0 means unused. */
93 unsigned src
[MIR_SRC_COUNT
];
96 /* vec16 swizzle, unpacked, per source */
97 unsigned swizzle
[MIR_SRC_COUNT
][MIR_VEC_COMPONENTS
];
99 /* Special fields for an ALU instruction */
100 midgard_reg_info registers
;
102 /* I.e. (1 << alu_bit) */
106 midgard_constants constants
;
107 uint16_t inline_constant
;
108 bool has_blend_constant
;
109 bool has_inline_constant
;
115 /* Kind of a hack, but hint against aggressive DCE */
118 /* Masks in a saneish format. One bit per channel, not packed fancy.
119 * Use this instead of the op specific ones, and switch over at emit
124 /* For ALU ops only: set to true to invert (bitwise NOT) the
125 * destination of an integer-out op. Not implemented in hardware but
126 * allows more optimizations */
130 /* Hint for the register allocator not to spill the destination written
131 * from this instruction (because it is a spill/unspill node itself).
132 * Bitmask of spilled classes */
136 /* Generic hint for intra-pass use */
139 /* During scheduling, the backwards dependency graph
140 * (DAG). nr_dependencies is the number of unscheduled
141 * instructions that must still be scheduled after
142 * (before) this instruction. dependents are which
143 * instructions need to be scheduled before (after) this
146 unsigned nr_dependencies
;
147 BITSET_WORD
*dependents
;
149 /* For load/store ops.. force 64-bit destination */
153 midgard_load_store_word load_store
;
154 midgard_vector_alu alu
;
155 midgard_texture_word texture
;
156 midgard_branch_extended branch_extended
;
159 /* General branch, rather than packed br_compact. Higher level
160 * than the other components */
161 midgard_branch branch
;
163 } midgard_instruction
;
165 typedef struct midgard_block
{
166 /* Link to next block. Must be first for mir_get_block */
167 struct list_head link
;
169 /* List of midgard_instructions emitted for the current block */
170 struct list_head instructions
;
172 /* Index of the block in source order */
177 /* List of midgard_bundles emitted (after the scheduler has run) */
178 struct util_dynarray bundles
;
180 /* Number of quadwords _actually_ emitted, as determined after scheduling */
181 unsigned quadword_count
;
183 /* Succeeding blocks. The compiler should not necessarily rely on
184 * source-order traversal */
185 struct midgard_block
*successors
[2];
186 unsigned nr_successors
;
188 struct set
*predecessors
;
190 /* The successors pointer form a graph, and in the case of
191 * complex control flow, this graph has a cycles. To aid
192 * traversal during liveness analysis, we have a visited?
193 * boolean for passes to use as they see fit, provided they
197 /* In liveness analysis, these are live masks (per-component) for
198 * indices for the block. Scalar compilers have the luxury of using
199 * simple bit fields, but for us, liveness is a vector idea. */
203 /* Indicates this is a fixed-function fragment epilogue block */
207 typedef struct midgard_bundle
{
208 /* Tag for the overall bundle */
211 /* Instructions contained by the bundle. instruction_count <= 6 (vmul,
212 * sadd, vadd, smul, vlut, branch) */
213 int instruction_count
;
214 midgard_instruction
*instructions
[6];
216 /* Bundle-wide ALU configuration */
219 bool has_embedded_constants
;
220 midgard_constants constants
;
221 bool has_blend_constant
;
225 typedef struct compiler_context
{
227 gl_shader_stage stage
;
229 /* Is internally a blend shader? Depends on stage == FRAGMENT */
232 /* Render target number for a keyed blend shader. Depends on is_blend */
235 /* Tracking for blend constant patching */
236 int blend_constant_offset
;
238 /* Number of bytes used for Thread Local Storage */
241 /* Count of spills and fills for shaderdb */
245 /* Current NIR function */
248 /* Allocated compiler temporary counter */
251 /* Unordered list of midgard_blocks */
253 struct list_head blocks
;
255 /* TODO merge with block_count? */
256 unsigned block_source_count
;
258 /* List of midgard_instructions emitted for the current block */
259 midgard_block
*current_block
;
261 /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
262 midgard_block
*after_block
;
264 /* The current "depth" of the loop, for disambiguating breaks/continues
265 * when using nested loops */
266 int current_loop_depth
;
268 /* Total number of loops for shader-db */
271 /* Constants which have been loaded, for later inlining */
272 struct hash_table_u64
*ssa_constants
;
274 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
275 struct hash_table_u64
*hash_to_temp
;
279 /* Just the count of the max register used. Higher count => higher
280 * register pressure */
283 /* Used for cont/last hinting. Increase when a tex op is added.
284 * Decrease when a tex op is removed. */
285 int texture_op_count
;
287 /* The number of uniforms allowable for the fast path */
290 /* Count of instructions emitted from NIR overall, across all blocks */
291 int instruction_count
;
293 /* Alpha ref value passed in */
296 unsigned quadword_count
;
298 /* The mapping of sysvals to uniforms, the count, and the off-by-one inverse */
299 unsigned sysvals
[MAX_SYSVAL_COUNT
];
300 unsigned sysval_count
;
301 struct hash_table_u64
*sysval_to_id
;
303 /* Bitmask of valid metadata */
306 /* Model-specific quirk set */
309 /* Writeout instructions for each render target */
310 midgard_instruction
*writeout_branch
[4];
313 /* Per-block live_in/live_out */
314 #define MIDGARD_METADATA_LIVENESS (1 << 0)
316 /* Helpers for manipulating the above structures (forming the driver IR) */
318 /* Append instruction to end of current block */
320 static inline midgard_instruction
*
321 mir_upload_ins(struct compiler_context
*ctx
, struct midgard_instruction ins
)
323 midgard_instruction
*heap
= ralloc(ctx
, struct midgard_instruction
);
324 memcpy(heap
, &ins
, sizeof(ins
));
328 static inline midgard_instruction
*
329 emit_mir_instruction(struct compiler_context
*ctx
, struct midgard_instruction ins
)
331 midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
332 list_addtail(&u
->link
, &ctx
->current_block
->instructions
);
336 static inline struct midgard_instruction
*
337 mir_insert_instruction_before(struct compiler_context
*ctx
,
338 struct midgard_instruction
*tag
,
339 struct midgard_instruction ins
)
341 struct midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
342 list_addtail(&u
->link
, &tag
->link
);
347 mir_remove_instruction(struct midgard_instruction
*ins
)
349 list_del(&ins
->link
);
352 static inline midgard_instruction
*
353 mir_prev_op(struct midgard_instruction
*ins
)
355 return list_last_entry(&(ins
->link
), midgard_instruction
, link
);
358 static inline midgard_instruction
*
359 mir_next_op(struct midgard_instruction
*ins
)
361 return list_first_entry(&(ins
->link
), midgard_instruction
, link
);
364 #define mir_foreach_block(ctx, v) \
365 list_for_each_entry(struct midgard_block, v, &ctx->blocks, link)
367 #define mir_foreach_block_from(ctx, from, v) \
368 list_for_each_entry_from(struct midgard_block, v, from, &ctx->blocks, link)
370 #define mir_foreach_instr(ctx, v) \
371 list_for_each_entry(struct midgard_instruction, v, &ctx->current_block->instructions, link)
373 #define mir_foreach_instr_safe(ctx, v) \
374 list_for_each_entry_safe(struct midgard_instruction, v, &ctx->current_block->instructions, link)
376 #define mir_foreach_instr_in_block(block, v) \
377 list_for_each_entry(struct midgard_instruction, v, &block->instructions, link)
378 #define mir_foreach_instr_in_block_rev(block, v) \
379 list_for_each_entry_rev(struct midgard_instruction, v, &block->instructions, link)
381 #define mir_foreach_instr_in_block_safe(block, v) \
382 list_for_each_entry_safe(struct midgard_instruction, v, &block->instructions, link)
384 #define mir_foreach_instr_in_block_safe_rev(block, v) \
385 list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->instructions, link)
387 #define mir_foreach_instr_in_block_from(block, v, from) \
388 list_for_each_entry_from(struct midgard_instruction, v, from, &block->instructions, link)
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->instructions, link)
393 #define mir_foreach_bundle_in_block(block, v) \
394 util_dynarray_foreach(&block->bundles, midgard_bundle, v)
396 #define mir_foreach_bundle_in_block_rev(block, v) \
397 util_dynarray_foreach_reverse(&block->bundles, midgard_bundle, v)
399 #define mir_foreach_instr_in_block_scheduled_rev(block, v) \
400 midgard_instruction* v; \
402 mir_foreach_bundle_in_block_rev(block, _bundle) \
403 for (i = (_bundle->instruction_count - 1), v = _bundle->instructions[i]; \
405 --i, v = (i >= 0) ? _bundle->instructions[i] : NULL) \
407 #define mir_foreach_instr_global(ctx, v) \
408 mir_foreach_block(ctx, v_block) \
409 mir_foreach_instr_in_block(v_block, v)
411 #define mir_foreach_instr_global_safe(ctx, v) \
412 mir_foreach_block(ctx, v_block) \
413 mir_foreach_instr_in_block_safe(v_block, v)
415 #define mir_foreach_successor(blk, v) \
416 struct midgard_block *v; \
417 struct midgard_block **_v; \
418 for (_v = &blk->successors[0], \
420 v != NULL && _v < &blk->successors[2]; \
423 /* Based on set_foreach, expanded with automatic type casts */
425 #define mir_foreach_predecessor(blk, v) \
426 struct set_entry *_entry_##v; \
427 struct midgard_block *v; \
428 for (_entry_##v = _mesa_set_next_entry(blk->predecessors, NULL), \
429 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL); \
430 _entry_##v != NULL; \
431 _entry_##v = _mesa_set_next_entry(blk->predecessors, _entry_##v), \
432 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
434 #define mir_foreach_src(ins, v) \
435 for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
437 static inline midgard_instruction
*
438 mir_last_in_block(struct midgard_block
*block
)
440 return list_last_entry(&block
->instructions
, struct midgard_instruction
, link
);
443 static inline midgard_block
*
444 mir_get_block(compiler_context
*ctx
, int idx
)
446 struct list_head
*lst
= &ctx
->blocks
;
451 return (struct midgard_block
*) lst
;
454 static inline midgard_block
*
455 mir_exit_block(struct compiler_context
*ctx
)
457 midgard_block
*last
= list_last_entry(&ctx
->blocks
,
458 struct midgard_block
, link
);
460 /* The last block must be empty logically but contains branch writeout
461 * for fragment shaders */
463 assert(last
->nr_successors
== 0);
469 mir_is_alu_bundle(midgard_bundle
*bundle
)
471 return midgard_word_types
[bundle
->tag
] == midgard_word_type_alu
;
474 /* Registers/SSA are distinguish in the backend by the bottom-most bit */
478 static inline unsigned
479 make_compiler_temp(compiler_context
*ctx
)
481 return (ctx
->func
->impl
->ssa_alloc
+ ctx
->temp_alloc
++) << 1;
484 static inline unsigned
485 make_compiler_temp_reg(compiler_context
*ctx
)
487 return ((ctx
->func
->impl
->reg_alloc
+ ctx
->temp_alloc
++) << 1) | IS_REG
;
490 static inline unsigned
491 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
494 return (src
->ssa
->index
<< 1) | 0;
496 assert(!src
->reg
.indirect
);
497 return (src
->reg
.reg
->index
<< 1) | IS_REG
;
501 static inline unsigned
502 nir_alu_src_index(compiler_context
*ctx
, nir_alu_src
*src
)
504 return nir_src_index(ctx
, &src
->src
);
507 static inline unsigned
508 nir_dest_index(compiler_context
*ctx
, nir_dest
*dst
)
511 return (dst
->ssa
.index
<< 1) | 0;
513 assert(!dst
->reg
.indirect
);
514 return (dst
->reg
.reg
->index
<< 1) | IS_REG
;
520 /* MIR manipulation */
522 void mir_rewrite_index(compiler_context
*ctx
, unsigned old
, unsigned new);
523 void mir_rewrite_index_src(compiler_context
*ctx
, unsigned old
, unsigned new);
524 void mir_rewrite_index_dst(compiler_context
*ctx
, unsigned old
, unsigned new);
525 void mir_rewrite_index_dst_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
526 void mir_rewrite_index_src_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
527 void mir_rewrite_index_src_swizzle(compiler_context
*ctx
, unsigned old
, unsigned new, unsigned *swizzle
);
528 bool mir_single_use(compiler_context
*ctx
, unsigned value
);
529 bool mir_special_index(compiler_context
*ctx
, unsigned idx
);
530 unsigned mir_use_count(compiler_context
*ctx
, unsigned value
);
531 bool mir_is_written_before(compiler_context
*ctx
, midgard_instruction
*ins
, unsigned node
);
532 uint16_t mir_bytemask_of_read_components(midgard_instruction
*ins
, unsigned node
);
533 midgard_reg_mode
mir_typesize(midgard_instruction
*ins
);
534 midgard_reg_mode
mir_srcsize(midgard_instruction
*ins
, unsigned i
);
535 unsigned mir_bytes_for_mode(midgard_reg_mode mode
);
536 midgard_reg_mode
mir_mode_for_destsize(unsigned size
);
537 uint16_t mir_from_bytemask(uint16_t bytemask
, midgard_reg_mode mode
);
538 uint16_t mir_to_bytemask(midgard_reg_mode mode
, unsigned mask
);
539 uint16_t mir_bytemask(midgard_instruction
*ins
);
540 uint16_t mir_round_bytemask_up(uint16_t mask
, midgard_reg_mode mode
);
541 void mir_set_bytemask(midgard_instruction
*ins
, uint16_t bytemask
);
542 unsigned mir_upper_override(midgard_instruction
*ins
);
546 void mir_print_instruction(midgard_instruction
*ins
);
547 void mir_print_bundle(midgard_bundle
*ctx
);
548 void mir_print_block(midgard_block
*block
);
549 void mir_print_shader(compiler_context
*ctx
);
550 bool mir_nontrivial_source2_mod(midgard_instruction
*ins
);
551 bool mir_nontrivial_source2_mod_simple(midgard_instruction
*ins
);
552 bool mir_nontrivial_outmod(midgard_instruction
*ins
);
554 void mir_insert_instruction_before_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
555 void mir_insert_instruction_after_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
556 void mir_flip(midgard_instruction
*ins
);
557 void mir_compute_temp_count(compiler_context
*ctx
);
559 /* 'Intrinsic' move for aliasing */
561 static inline midgard_instruction
562 v_mov(unsigned src
, unsigned dest
)
564 midgard_instruction ins
= {
567 .src
= { ~0, src
, ~0, ~0 },
568 .swizzle
= SWIZZLE_IDENTITY
,
571 .op
= midgard_alu_op_imov
,
572 .reg_mode
= midgard_reg_mode_32
,
573 .dest_override
= midgard_dest_override_none
,
574 .outmod
= midgard_outmod_int_wrap
581 /* Broad types of register classes so we can handle special
584 #define REG_CLASS_WORK 0
585 #define REG_CLASS_LDST 1
586 #define REG_CLASS_TEXR 3
587 #define REG_CLASS_TEXW 4
589 /* Like a move, but to thread local storage! */
591 static inline midgard_instruction
592 v_load_store_scratch(
598 /* We index by 32-bit vec4s */
599 unsigned byte
= (index
* 4 * 4);
601 midgard_instruction ins
= {
602 .type
= TAG_LOAD_STORE_4
,
605 .src
= { ~0, ~0, ~0, ~0 },
606 .swizzle
= SWIZZLE_IDENTITY_4
,
608 .op
= is_store
? midgard_op_st_int4
: midgard_op_ld_int4
,
610 /* For register spilling - to thread local storage */
615 /* If we spill an unspill, RA goes into an infinite loop */
616 .no_spill
= (1 << REG_CLASS_WORK
)
619 ins
.constants
.u32
[0] = byte
;
622 ins
.src
[0] = srcdest
;
624 /* Ensure we are tightly swizzled so liveness analysis is
627 for (unsigned i
= 0; i
< 4; ++i
) {
628 if (!(mask
& (1 << i
)))
629 ins
.swizzle
[0][i
] = COMPONENT_X
;
638 mir_has_arg(midgard_instruction
*ins
, unsigned arg
)
643 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
->src
); ++i
) {
644 if (ins
->src
[i
] == arg
)
653 void midgard_schedule_program(compiler_context
*ctx
);
655 void mir_ra(compiler_context
*ctx
);
656 void mir_squeeze_index(compiler_context
*ctx
);
657 void mir_lower_special_reads(compiler_context
*ctx
);
658 void mir_liveness_ins_update(uint16_t *live
, midgard_instruction
*ins
, unsigned max
);
659 void mir_compute_liveness(compiler_context
*ctx
);
660 void mir_invalidate_liveness(compiler_context
*ctx
);
661 bool mir_is_live_after(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*start
, int src
);
663 void mir_create_pipeline_registers(compiler_context
*ctx
);
664 void midgard_promote_uniforms(compiler_context
*ctx
);
667 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
, signed dest_override
, unsigned nr_components
);
670 midgard_emit_derivatives(compiler_context
*ctx
, nir_alu_instr
*instr
);
673 midgard_lower_derivatives(compiler_context
*ctx
, midgard_block
*block
);
675 bool mir_op_computes_derivatives(gl_shader_stage stage
, unsigned op
);
679 void emit_binary_bundle(
680 compiler_context
*ctx
,
681 midgard_bundle
*bundle
,
682 struct util_dynarray
*emission
,
686 nir_undef_to_zero(nir_shader
*shader
);
688 void midgard_nir_lod_errata(nir_shader
*shader
);
692 bool midgard_opt_copy_prop(compiler_context
*ctx
, midgard_block
*block
);
693 bool midgard_opt_combine_projection(compiler_context
*ctx
, midgard_block
*block
);
694 bool midgard_opt_varying_projection(compiler_context
*ctx
, midgard_block
*block
);
695 bool midgard_opt_dead_code_eliminate(compiler_context
*ctx
, midgard_block
*block
);
696 bool midgard_opt_dead_move_eliminate(compiler_context
*ctx
, midgard_block
*block
);
698 void midgard_lower_invert(compiler_context
*ctx
, midgard_block
*block
);
699 bool midgard_opt_not_propagate(compiler_context
*ctx
, midgard_block
*block
);
700 bool midgard_opt_fuse_src_invert(compiler_context
*ctx
, midgard_block
*block
);
701 bool midgard_opt_fuse_dest_invert(compiler_context
*ctx
, midgard_block
*block
);
702 bool midgard_opt_csel_invert(compiler_context
*ctx
, midgard_block
*block
);
703 bool midgard_opt_promote_fmov(compiler_context
*ctx
, midgard_block
*block
);
704 bool midgard_opt_drop_cmp_invert(compiler_context
*ctx
, midgard_block
*block
);
705 bool midgard_opt_invert_branch(compiler_context
*ctx
, midgard_block
*block
);