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"
31 #include "util/hash_table.h"
32 #include "util/u_dynarray.h"
34 #include "util/list.h"
36 #include "main/mtypes.h"
37 #include "compiler/nir_types.h"
38 #include "compiler/nir/nir.h"
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. */
48 #define TARGET_BREAK 1
49 #define TARGET_CONTINUE 2
50 #define TARGET_DISCARD 3
52 typedef struct midgard_branch
{
53 /* If conditional, the condition is specified in r31.w */
56 /* For conditionals, if this is true, we branch on FALSE. If false, we branch on TRUE. */
57 bool invert_conditional
;
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_ */
62 /* The actual target */
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.
76 * Notably, it allows registers to be specified as block local SSA, for code
77 * emitted before the register allocation pass.
80 #define MIR_SRC_COUNT 3
81 #define MIR_VEC_COMPONENTS 16
83 typedef struct midgard_instruction
{
84 /* Must be first for casting */
85 struct list_head link
;
87 unsigned type
; /* ALU, load/store, texture */
89 /* Instruction arguments represented as block-local SSA
90 * indices, rather than registers. ~0 means unused. */
94 /* vec16 swizzle, unpacked, per source */
95 unsigned swizzle
[MIR_SRC_COUNT
][MIR_VEC_COMPONENTS
];
97 /* Special fields for an ALU instruction */
98 midgard_reg_info registers
;
100 /* I.e. (1 << alu_bit) */
104 uint32_t constants
[4];
105 uint16_t inline_constant
;
106 bool has_blend_constant
;
107 bool has_inline_constant
;
111 bool prepacked_branch
;
113 /* Kind of a hack, but hint against aggressive DCE */
116 /* Masks in a saneish format. One bit per channel, not packed fancy.
117 * Use this instead of the op specific ones, and switch over at emit
122 /* For ALU ops only: set to true to invert (bitwise NOT) the
123 * destination of an integer-out op. Not imeplemented in hardware but
124 * allows more optimizations */
128 /* Hint for the register allocator not to spill the destination written
129 * from this instruction (because it is a spill/unspill node itself) */
133 /* Generic hint for intra-pass use */
136 /* During scheduling, the backwards dependency graph
137 * (DAG). nr_dependencies is the number of unscheduled
138 * instructions that must still be scheduled after
139 * (before) this instruction. dependents are which
140 * instructions need to be scheduled before (after) this
143 unsigned nr_dependencies
;
144 BITSET_WORD
*dependents
;
147 midgard_load_store_word load_store
;
148 midgard_vector_alu alu
;
149 midgard_texture_word texture
;
150 midgard_branch_extended branch_extended
;
153 /* General branch, rather than packed br_compact. Higher level
154 * than the other components */
155 midgard_branch branch
;
157 } midgard_instruction
;
159 typedef struct midgard_block
{
160 /* Link to next block. Must be first for mir_get_block */
161 struct list_head link
;
163 /* List of midgard_instructions emitted for the current block */
164 struct list_head instructions
;
166 /* Index of the block in source order */
171 /* List of midgard_bundles emitted (after the scheduler has run) */
172 struct util_dynarray bundles
;
174 /* Number of quadwords _actually_ emitted, as determined after scheduling */
175 unsigned quadword_count
;
177 /* Succeeding blocks. The compiler should not necessarily rely on
178 * source-order traversal */
179 struct midgard_block
*successors
[2];
180 unsigned nr_successors
;
182 struct set
*predecessors
;
184 /* The successors pointer form a graph, and in the case of
185 * complex control flow, this graph has a cycles. To aid
186 * traversal during liveness analysis, we have a visited?
187 * boolean for passes to use as they see fit, provided they
191 /* In liveness analysis, these are live masks (per-component) for
192 * indices for the block. Scalar compilers have the luxury of using
193 * simple bit fields, but for us, liveness is a vector idea. */
198 typedef struct midgard_bundle
{
199 /* Tag for the overall bundle */
202 /* Instructions contained by the bundle. instruction_count <= 6 (vmul,
203 * sadd, vadd, smul, vlut, branch) */
204 int instruction_count
;
205 midgard_instruction
*instructions
[6];
207 /* Bundle-wide ALU configuration */
210 bool has_embedded_constants
;
212 bool has_blend_constant
;
215 typedef struct compiler_context
{
217 gl_shader_stage stage
;
219 /* The screen we correspond to */
220 struct midgard_screen
*screen
;
222 /* Is internally a blend shader? Depends on stage == FRAGMENT */
225 /* Tracking for blend constant patching */
226 int blend_constant_offset
;
228 /* Number of bytes used for Thread Local Storage */
231 /* Count of spills and fills for shaderdb */
235 /* Current NIR function */
238 /* Allocated compiler temporary counter */
241 /* Unordered list of midgard_blocks */
243 struct list_head blocks
;
245 /* TODO merge with block_count? */
246 unsigned block_source_count
;
248 /* List of midgard_instructions emitted for the current block */
249 midgard_block
*current_block
;
251 /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
252 midgard_block
*after_block
;
254 /* The current "depth" of the loop, for disambiguating breaks/continues
255 * when using nested loops */
256 int current_loop_depth
;
258 /* Total number of loops for shader-db */
261 /* Constants which have been loaded, for later inlining */
262 struct hash_table_u64
*ssa_constants
;
264 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
265 struct hash_table_u64
*hash_to_temp
;
269 /* Just the count of the max register used. Higher count => higher
270 * register pressure */
273 /* Used for cont/last hinting. Increase when a tex op is added.
274 * Decrease when a tex op is removed. */
275 int texture_op_count
;
277 /* The number of uniforms allowable for the fast path */
280 /* Count of instructions emitted from NIR overall, across all blocks */
281 int instruction_count
;
283 /* Alpha ref value passed in */
286 unsigned quadword_count
;
288 /* The mapping of sysvals to uniforms, the count, and the off-by-one inverse */
289 unsigned sysvals
[MAX_SYSVAL_COUNT
];
290 unsigned sysval_count
;
291 struct hash_table_u64
*sysval_to_id
;
293 /* Bitmask of valid metadata */
299 /* Per-block live_in/live_out */
300 #define MIDGARD_METADATA_LIVENESS (1 << 0)
302 /* Helpers for manipulating the above structures (forming the driver IR) */
304 /* Append instruction to end of current block */
306 static inline midgard_instruction
*
307 mir_upload_ins(struct compiler_context
*ctx
, struct midgard_instruction ins
)
309 midgard_instruction
*heap
= ralloc(ctx
, struct midgard_instruction
);
310 memcpy(heap
, &ins
, sizeof(ins
));
314 static inline midgard_instruction
*
315 emit_mir_instruction(struct compiler_context
*ctx
, struct midgard_instruction ins
)
317 midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
318 list_addtail(&u
->link
, &ctx
->current_block
->instructions
);
322 static inline struct midgard_instruction
*
323 mir_insert_instruction_before(struct compiler_context
*ctx
,
324 struct midgard_instruction
*tag
,
325 struct midgard_instruction ins
)
327 struct midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
328 list_addtail(&u
->link
, &tag
->link
);
333 mir_remove_instruction(struct midgard_instruction
*ins
)
335 list_del(&ins
->link
);
338 static inline midgard_instruction
*
339 mir_prev_op(struct midgard_instruction
*ins
)
341 return list_last_entry(&(ins
->link
), midgard_instruction
, link
);
344 static inline midgard_instruction
*
345 mir_next_op(struct midgard_instruction
*ins
)
347 return list_first_entry(&(ins
->link
), midgard_instruction
, link
);
350 #define mir_foreach_block(ctx, v) \
351 list_for_each_entry(struct midgard_block, v, &ctx->blocks, link)
353 #define mir_foreach_block_from(ctx, from, v) \
354 list_for_each_entry_from(struct midgard_block, v, from, &ctx->blocks, link)
356 #define mir_foreach_instr(ctx, v) \
357 list_for_each_entry(struct midgard_instruction, v, &ctx->current_block->instructions, link)
359 #define mir_foreach_instr_safe(ctx, v) \
360 list_for_each_entry_safe(struct midgard_instruction, v, &ctx->current_block->instructions, link)
362 #define mir_foreach_instr_in_block(block, v) \
363 list_for_each_entry(struct midgard_instruction, v, &block->instructions, link)
364 #define mir_foreach_instr_in_block_rev(block, v) \
365 list_for_each_entry_rev(struct midgard_instruction, v, &block->instructions, link)
367 #define mir_foreach_instr_in_block_safe(block, v) \
368 list_for_each_entry_safe(struct midgard_instruction, v, &block->instructions, link)
370 #define mir_foreach_instr_in_block_safe_rev(block, v) \
371 list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->instructions, link)
373 #define mir_foreach_instr_in_block_from(block, v, from) \
374 list_for_each_entry_from(struct midgard_instruction, v, from, &block->instructions, link)
376 #define mir_foreach_instr_in_block_from_rev(block, v, from) \
377 list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->instructions, link)
379 #define mir_foreach_bundle_in_block(block, v) \
380 util_dynarray_foreach(&block->bundles, midgard_bundle, v)
382 #define mir_foreach_bundle_in_block_rev(block, v) \
383 util_dynarray_foreach_reverse(&block->bundles, midgard_bundle, v)
385 #define mir_foreach_instr_in_block_scheduled_rev(block, v) \
386 midgard_instruction* v; \
388 mir_foreach_bundle_in_block_rev(block, _bundle) \
389 for (i = (_bundle->instruction_count - 1), v = _bundle->instructions[i]; \
391 --i, v = _bundle->instructions[i]) \
393 #define mir_foreach_instr_global(ctx, v) \
394 mir_foreach_block(ctx, v_block) \
395 mir_foreach_instr_in_block(v_block, v)
397 #define mir_foreach_instr_global_safe(ctx, v) \
398 mir_foreach_block(ctx, v_block) \
399 mir_foreach_instr_in_block_safe(v_block, v)
401 #define mir_foreach_successor(blk, v) \
402 struct midgard_block *v; \
403 struct midgard_block **_v; \
404 for (_v = &blk->successors[0], \
406 v != NULL && _v < &blk->successors[2]; \
409 /* Based on set_foreach, expanded with automatic type casts */
411 #define mir_foreach_predecessor(blk, v) \
412 struct set_entry *_entry_##v; \
413 struct midgard_block *v; \
414 for (_entry_##v = _mesa_set_next_entry(blk->predecessors, NULL), \
415 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL); \
416 _entry_##v != NULL; \
417 _entry_##v = _mesa_set_next_entry(blk->predecessors, _entry_##v), \
418 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
420 #define mir_foreach_src(ins, v) \
421 for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
423 static inline midgard_instruction
*
424 mir_last_in_block(struct midgard_block
*block
)
426 return list_last_entry(&block
->instructions
, struct midgard_instruction
, link
);
429 static inline midgard_block
*
430 mir_get_block(compiler_context
*ctx
, int idx
)
432 struct list_head
*lst
= &ctx
->blocks
;
437 return (struct midgard_block
*) lst
;
440 static inline midgard_block
*
441 mir_exit_block(struct compiler_context
*ctx
)
443 midgard_block
*last
= list_last_entry(&ctx
->blocks
,
444 struct midgard_block
, link
);
446 /* The last block must be empty logically but contains branch writeout
447 * for fragment shaders */
449 assert(last
->nr_successors
== 0);
455 mir_is_alu_bundle(midgard_bundle
*bundle
)
457 return IS_ALU(bundle
->tag
);
460 /* Registers/SSA are distinguish in the backend by the bottom-most bit */
464 static inline unsigned
465 make_compiler_temp(compiler_context
*ctx
)
467 return (ctx
->func
->impl
->ssa_alloc
+ ctx
->temp_alloc
++) << 1;
470 static inline unsigned
471 make_compiler_temp_reg(compiler_context
*ctx
)
473 return ((ctx
->func
->impl
->reg_alloc
+ ctx
->temp_alloc
++) << 1) | IS_REG
;
476 static inline unsigned
477 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
480 return (src
->ssa
->index
<< 1) | 0;
482 assert(!src
->reg
.indirect
);
483 return (src
->reg
.reg
->index
<< 1) | IS_REG
;
487 static inline unsigned
488 nir_alu_src_index(compiler_context
*ctx
, nir_alu_src
*src
)
490 return nir_src_index(ctx
, &src
->src
);
493 static inline unsigned
494 nir_dest_index(compiler_context
*ctx
, nir_dest
*dst
)
497 return (dst
->ssa
.index
<< 1) | 0;
499 assert(!dst
->reg
.indirect
);
500 return (dst
->reg
.reg
->index
<< 1) | IS_REG
;
506 /* MIR manipulation */
508 void mir_rewrite_index(compiler_context
*ctx
, unsigned old
, unsigned new);
509 void mir_rewrite_index_src(compiler_context
*ctx
, unsigned old
, unsigned new);
510 void mir_rewrite_index_dst(compiler_context
*ctx
, unsigned old
, unsigned new);
511 void mir_rewrite_index_dst_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
512 void mir_rewrite_index_src_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
513 void mir_rewrite_index_src_swizzle(compiler_context
*ctx
, unsigned old
, unsigned new, unsigned *swizzle
);
514 bool mir_single_use(compiler_context
*ctx
, unsigned value
);
515 bool mir_special_index(compiler_context
*ctx
, unsigned idx
);
516 unsigned mir_use_count(compiler_context
*ctx
, unsigned value
);
517 bool mir_is_written_before(compiler_context
*ctx
, midgard_instruction
*ins
, unsigned node
);
518 uint16_t mir_bytemask_of_read_components(midgard_instruction
*ins
, unsigned node
);
519 unsigned mir_ubo_shift(midgard_load_store_op op
);
520 midgard_reg_mode
mir_typesize(midgard_instruction
*ins
);
521 midgard_reg_mode
mir_srcsize(midgard_instruction
*ins
, unsigned i
);
522 unsigned mir_bytes_for_mode(midgard_reg_mode mode
);
523 uint16_t mir_from_bytemask(uint16_t bytemask
, midgard_reg_mode mode
);
524 uint16_t mir_bytemask(midgard_instruction
*ins
);
525 uint16_t mir_round_bytemask_down(uint16_t mask
, midgard_reg_mode mode
);
526 void mir_set_bytemask(midgard_instruction
*ins
, uint16_t bytemask
);
530 void mir_print_instruction(midgard_instruction
*ins
);
531 void mir_print_bundle(midgard_bundle
*ctx
);
532 void mir_print_block(midgard_block
*block
);
533 void mir_print_shader(compiler_context
*ctx
);
534 bool mir_nontrivial_source2_mod(midgard_instruction
*ins
);
535 bool mir_nontrivial_source2_mod_simple(midgard_instruction
*ins
);
536 bool mir_nontrivial_outmod(midgard_instruction
*ins
);
538 void mir_insert_instruction_before_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
539 void mir_insert_instruction_after_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
540 void mir_flip(midgard_instruction
*ins
);
541 void mir_compute_temp_count(compiler_context
*ctx
);
543 /* 'Intrinsic' move for aliasing */
545 static inline midgard_instruction
546 v_mov(unsigned src
, unsigned dest
)
548 midgard_instruction ins
= {
551 .src
= { SSA_UNUSED
, src
, SSA_UNUSED
},
552 .swizzle
= SWIZZLE_IDENTITY
,
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
566 mir_has_arg(midgard_instruction
*ins
, unsigned arg
)
571 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
->src
); ++i
) {
572 if (ins
->src
[i
] == arg
)
581 void schedule_program(compiler_context
*ctx
);
583 /* Register allocation */
587 /* Broad types of register classes so we can handle special
590 #define NR_REG_CLASSES 6
592 #define REG_CLASS_WORK 0
593 #define REG_CLASS_LDST 1
594 #define REG_CLASS_LDST27 2
595 #define REG_CLASS_TEXR 3
596 #define REG_CLASS_TEXW 4
597 #define REG_CLASS_FRAGC 5
599 void mir_lower_special_reads(compiler_context
*ctx
);
600 struct ra_graph
* allocate_registers(compiler_context
*ctx
, bool *spilled
);
601 void install_registers(compiler_context
*ctx
, struct ra_graph
*g
);
602 void mir_liveness_ins_update(uint16_t *live
, midgard_instruction
*ins
, unsigned max
);
603 void mir_compute_liveness(compiler_context
*ctx
);
604 void mir_invalidate_liveness(compiler_context
*ctx
);
605 bool mir_is_live_after(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*start
, int src
);
607 void mir_create_pipeline_registers(compiler_context
*ctx
);
610 midgard_promote_uniforms(compiler_context
*ctx
, unsigned promoted_count
);
612 midgard_instruction
*
614 compiler_context
*ctx
,
618 nir_src
*indirect_offset
,
622 emit_sysval_read(compiler_context
*ctx
, nir_instr
*instr
, signed dest_override
, unsigned nr_components
);
625 midgard_emit_derivatives(compiler_context
*ctx
, nir_alu_instr
*instr
);
628 midgard_lower_derivatives(compiler_context
*ctx
, midgard_block
*block
);
630 bool mir_op_computes_derivatives(unsigned op
);
634 void emit_binary_bundle(
635 compiler_context
*ctx
,
636 midgard_bundle
*bundle
,
637 struct util_dynarray
*emission
,
641 nir_undef_to_zero(nir_shader
*shader
);
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
);
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_csel_invert(compiler_context
*ctx
, midgard_block
*block
);
656 bool midgard_opt_promote_fmov(compiler_context
*ctx
, midgard_block
*block
);