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"
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"
40 #include "panfrost/util/pan_ir.h"
41 #include "panfrost/util/lcra.h"
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. */
51 #define TARGET_BREAK 1
52 #define TARGET_CONTINUE 2
53 #define TARGET_DISCARD 3
55 typedef struct midgard_branch
{
56 /* If conditional, the condition is specified in r31.w */
59 /* For conditionals, if this is true, we branch on FALSE. If false, we branch on TRUE. */
60 bool invert_conditional
;
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_ */
65 /* The actual target */
73 /* Generic in-memory data type repesenting a single logical instruction, rather
74 * than a single instruction group. This is the preferred form for code gen.
75 * Multiple midgard_insturctions will later be combined during scheduling,
76 * though this is not represented in this structure. Its format bridges
77 * the low-level binary representation with the higher level semantic meaning.
79 * Notably, it allows registers to be specified as block local SSA, for code
80 * emitted before the register allocation pass.
83 #define MIR_SRC_COUNT 4
84 #define MIR_VEC_COMPONENTS 16
86 typedef struct midgard_instruction
{
87 /* Must be first for casting */
88 struct list_head link
;
90 unsigned type
; /* ALU, load/store, texture */
92 /* Instruction arguments represented as block-local SSA
93 * indices, rather than registers. ~0 means unused. */
94 unsigned src
[MIR_SRC_COUNT
];
97 /* vec16 swizzle, unpacked, per source */
98 unsigned swizzle
[MIR_SRC_COUNT
][MIR_VEC_COMPONENTS
];
101 nir_alu_type src_types
[MIR_SRC_COUNT
];
102 nir_alu_type dest_type
;
104 /* Special fields for an ALU instruction */
105 midgard_reg_info registers
;
107 /* For textures: should helpers execute this instruction (instead of
108 * just helping with derivatives)? Should helpers terminate after? */
109 bool helper_terminate
;
112 /* I.e. (1 << alu_bit) */
116 midgard_constants constants
;
117 uint16_t inline_constant
;
118 bool has_blend_constant
;
119 bool has_inline_constant
;
124 bool writeout_stencil
;
127 /* Masks in a saneish format. One bit per channel, not packed fancy.
128 * Use this instead of the op specific ones, and switch over at emit
133 /* For accepting ALU ops - invert the nth source */
134 bool src_invert
[MIR_SRC_COUNT
];
136 /* Hint for the register allocator not to spill the destination written
137 * from this instruction (because it is a spill/unspill node itself).
138 * Bitmask of spilled classes */
142 /* Generic hint for intra-pass use */
145 /* During scheduling, the backwards dependency graph
146 * (DAG). nr_dependencies is the number of unscheduled
147 * instructions that must still be scheduled after
148 * (before) this instruction. dependents are which
149 * instructions need to be scheduled before (after) this
152 unsigned nr_dependencies
;
153 BITSET_WORD
*dependents
;
155 /* For load/store ops.. force 64-bit destination */
159 midgard_load_store_word load_store
;
160 midgard_vector_alu alu
;
161 midgard_texture_word texture
;
162 midgard_branch_extended branch_extended
;
165 /* General branch, rather than packed br_compact. Higher level
166 * than the other components */
167 midgard_branch branch
;
169 } midgard_instruction
;
171 typedef struct midgard_block
{
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 /* Indicates this is a fixed-function fragment epilogue block */
185 /* Are helper invocations required by this block? */
189 typedef struct midgard_bundle
{
190 /* Tag for the overall bundle */
193 /* Instructions contained by the bundle. instruction_count <= 6 (vmul,
194 * sadd, vadd, smul, vlut, branch) */
195 int instruction_count
;
196 midgard_instruction
*instructions
[6];
198 /* Bundle-wide ALU configuration */
201 bool has_embedded_constants
;
202 midgard_constants constants
;
203 bool has_blend_constant
;
216 typedef struct compiler_context
{
218 gl_shader_stage stage
;
220 /* Is internally a blend shader? Depends on stage == FRAGMENT */
223 /* Render target number for a keyed blend shader. Depends on is_blend */
226 /* Tracking for blend constant patching */
227 int blend_constant_offset
;
229 /* Number of bytes used for Thread Local Storage */
232 /* Count of spills and fills for shaderdb */
236 /* Current NIR function */
239 /* Allocated compiler temporary counter */
242 /* Unordered list of midgard_blocks */
244 struct list_head blocks
;
246 /* TODO merge with block_count? */
247 unsigned block_source_count
;
249 /* List of midgard_instructions emitted for the current block */
250 midgard_block
*current_block
;
252 /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
253 midgard_block
*after_block
;
255 /* The current "depth" of the loop, for disambiguating breaks/continues
256 * when using nested loops */
257 int current_loop_depth
;
259 /* Total number of loops for shader-db */
262 /* Constants which have been loaded, for later inlining */
263 struct hash_table_u64
*ssa_constants
;
265 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
266 struct hash_table_u64
*hash_to_temp
;
270 /* Set of NIR indices that were already emitted as outmods */
271 BITSET_WORD
*already_emitted
;
273 /* Just the count of the max register used. Higher count => higher
274 * register pressure */
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 /* Bitmask of valid metadata */
291 /* Model-specific quirk set */
294 /* Writeout instructions for each render target */
295 midgard_instruction
*writeout_branch
[MIDGARD_NUM_RTS
];
297 struct panfrost_sysvals sysvals
;
300 /* Per-block live_in/live_out */
301 #define MIDGARD_METADATA_LIVENESS (1 << 0)
303 /* Helpers for manipulating the above structures (forming the driver IR) */
305 /* Append instruction to end of current block */
307 static inline midgard_instruction
*
308 mir_upload_ins(struct compiler_context
*ctx
, struct midgard_instruction ins
)
310 midgard_instruction
*heap
= ralloc(ctx
, struct midgard_instruction
);
311 memcpy(heap
, &ins
, sizeof(ins
));
315 static inline midgard_instruction
*
316 emit_mir_instruction(struct compiler_context
*ctx
, struct midgard_instruction ins
)
318 midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
319 list_addtail(&u
->link
, &ctx
->current_block
->base
.instructions
);
323 static inline struct midgard_instruction
*
324 mir_insert_instruction_before(struct compiler_context
*ctx
,
325 struct midgard_instruction
*tag
,
326 struct midgard_instruction ins
)
328 struct midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
329 list_addtail(&u
->link
, &tag
->link
);
334 mir_remove_instruction(struct midgard_instruction
*ins
)
336 list_del(&ins
->link
);
339 static inline midgard_instruction
*
340 mir_prev_op(struct midgard_instruction
*ins
)
342 return list_last_entry(&(ins
->link
), midgard_instruction
, link
);
345 static inline midgard_instruction
*
346 mir_next_op(struct midgard_instruction
*ins
)
348 return list_first_entry(&(ins
->link
), midgard_instruction
, link
);
351 #define mir_foreach_block(ctx, v) \
352 list_for_each_entry(pan_block, v, &ctx->blocks, link)
354 #define mir_foreach_block_from(ctx, from, v) \
355 list_for_each_entry_from(pan_block, v, &from->base, &ctx->blocks, link)
357 #define mir_foreach_instr_in_block(block, v) \
358 list_for_each_entry(struct midgard_instruction, v, &block->base.instructions, link)
359 #define mir_foreach_instr_in_block_rev(block, v) \
360 list_for_each_entry_rev(struct midgard_instruction, v, &block->base.instructions, link)
362 #define mir_foreach_instr_in_block_safe(block, v) \
363 list_for_each_entry_safe(struct midgard_instruction, v, &block->base.instructions, link)
365 #define mir_foreach_instr_in_block_safe_rev(block, v) \
366 list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->base.instructions, link)
368 #define mir_foreach_instr_in_block_from(block, v, from) \
369 list_for_each_entry_from(struct midgard_instruction, v, from, &block->base.instructions, link)
371 #define mir_foreach_instr_in_block_from_rev(block, v, from) \
372 list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->base.instructions, link)
374 #define mir_foreach_bundle_in_block(block, v) \
375 util_dynarray_foreach(&block->bundles, midgard_bundle, v)
377 #define mir_foreach_bundle_in_block_rev(block, v) \
378 util_dynarray_foreach_reverse(&block->bundles, midgard_bundle, v)
380 #define mir_foreach_instr_in_block_scheduled_rev(block, v) \
381 midgard_instruction* v; \
383 mir_foreach_bundle_in_block_rev(block, _bundle) \
384 for (i = (_bundle->instruction_count - 1), v = _bundle->instructions[i]; \
386 --i, v = (i >= 0) ? _bundle->instructions[i] : NULL) \
388 #define mir_foreach_instr_global(ctx, v) \
389 mir_foreach_block(ctx, v_block) \
390 mir_foreach_instr_in_block(((midgard_block *) v_block), v)
392 #define mir_foreach_instr_global_safe(ctx, v) \
393 mir_foreach_block(ctx, v_block) \
394 mir_foreach_instr_in_block_safe(((midgard_block *) v_block), v)
396 /* Based on set_foreach, expanded with automatic type casts */
398 #define mir_foreach_predecessor(blk, v) \
399 struct set_entry *_entry_##v; \
400 struct midgard_block *v; \
401 for (_entry_##v = _mesa_set_next_entry(blk->base.predecessors, NULL), \
402 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL); \
403 _entry_##v != NULL; \
404 _entry_##v = _mesa_set_next_entry(blk->base.predecessors, _entry_##v), \
405 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
407 #define mir_foreach_src(ins, v) \
408 for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
410 static inline midgard_instruction
*
411 mir_last_in_block(struct midgard_block
*block
)
413 return list_last_entry(&block
->base
.instructions
, struct midgard_instruction
, link
);
416 static inline midgard_block
*
417 mir_get_block(compiler_context
*ctx
, int idx
)
419 struct list_head
*lst
= &ctx
->blocks
;
424 return (struct midgard_block
*) lst
;
428 mir_is_alu_bundle(midgard_bundle
*bundle
)
430 return IS_ALU(bundle
->tag
);
433 static inline unsigned
434 make_compiler_temp(compiler_context
*ctx
)
436 return (ctx
->func
->impl
->ssa_alloc
+ ctx
->temp_alloc
++) << 1;
439 static inline unsigned
440 make_compiler_temp_reg(compiler_context
*ctx
)
442 return ((ctx
->func
->impl
->reg_alloc
+ ctx
->temp_alloc
++) << 1) | PAN_IS_REG
;
445 static inline unsigned
446 nir_ssa_index(nir_ssa_def
*ssa
)
448 return (ssa
->index
<< 1) | 0;
451 static inline unsigned
452 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
455 return nir_ssa_index(src
->ssa
);
457 assert(!src
->reg
.indirect
);
458 return (src
->reg
.reg
->index
<< 1) | PAN_IS_REG
;
462 static inline unsigned
463 nir_dest_index(nir_dest
*dst
)
466 return (dst
->ssa
.index
<< 1) | 0;
468 assert(!dst
->reg
.indirect
);
469 return (dst
->reg
.reg
->index
<< 1) | PAN_IS_REG
;
475 /* MIR manipulation */
477 void mir_rewrite_index(compiler_context
*ctx
, unsigned old
, unsigned new);
478 void mir_rewrite_index_src(compiler_context
*ctx
, unsigned old
, unsigned new);
479 void mir_rewrite_index_dst(compiler_context
*ctx
, unsigned old
, unsigned new);
480 void mir_rewrite_index_dst_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
481 void mir_rewrite_index_src_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
482 void mir_rewrite_index_src_swizzle(compiler_context
*ctx
, unsigned old
, unsigned new, unsigned *swizzle
);
483 bool mir_single_use(compiler_context
*ctx
, unsigned value
);
484 unsigned mir_use_count(compiler_context
*ctx
, unsigned value
);
485 uint16_t mir_bytemask_of_read_components(midgard_instruction
*ins
, unsigned node
);
486 uint16_t mir_bytemask_of_read_components_index(midgard_instruction
*ins
, unsigned i
);
487 midgard_reg_mode
mir_typesize(midgard_instruction
*ins
);
488 midgard_reg_mode
mir_srcsize(midgard_instruction
*ins
, unsigned i
);
489 unsigned mir_bytes_for_mode(midgard_reg_mode mode
);
490 midgard_reg_mode
mir_mode_for_destsize(unsigned size
);
491 uint16_t mir_from_bytemask(uint16_t bytemask
, midgard_reg_mode mode
);
492 uint16_t mir_bytemask(midgard_instruction
*ins
);
493 uint16_t mir_round_bytemask_up(uint16_t mask
, midgard_reg_mode mode
);
494 void mir_set_bytemask(midgard_instruction
*ins
, uint16_t bytemask
);
495 unsigned mir_upper_override(midgard_instruction
*ins
);
499 void mir_print_instruction(midgard_instruction
*ins
);
500 void mir_print_bundle(midgard_bundle
*ctx
);
501 void mir_print_block(midgard_block
*block
);
502 void mir_print_shader(compiler_context
*ctx
);
503 bool mir_nontrivial_source2_mod(midgard_instruction
*ins
);
504 bool mir_nontrivial_source2_mod_simple(midgard_instruction
*ins
);
505 bool mir_nontrivial_outmod(midgard_instruction
*ins
);
507 void mir_insert_instruction_before_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
508 void mir_insert_instruction_after_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
509 void mir_flip(midgard_instruction
*ins
);
510 void mir_compute_temp_count(compiler_context
*ctx
);
512 void mir_set_offset(compiler_context
*ctx
, midgard_instruction
*ins
, nir_src
*offset
, bool is_shared
);
514 /* 'Intrinsic' move for aliasing */
516 static inline midgard_instruction
517 v_mov(unsigned src
, unsigned dest
)
519 midgard_instruction ins
= {
522 .src
= { ~0, src
, ~0, ~0 },
523 .src_types
= { 0, nir_type_uint32
},
524 .swizzle
= SWIZZLE_IDENTITY
,
526 .dest_type
= nir_type_uint32
,
528 .op
= midgard_alu_op_imov
,
529 .reg_mode
= midgard_reg_mode_32
,
530 .dest_override
= midgard_dest_override_none
,
531 .outmod
= midgard_outmod_int_wrap
538 /* Broad types of register classes so we can handle special
541 #define REG_CLASS_WORK 0
542 #define REG_CLASS_LDST 1
543 #define REG_CLASS_TEXR 3
544 #define REG_CLASS_TEXW 4
546 /* Like a move, but to thread local storage! */
548 static inline midgard_instruction
549 v_load_store_scratch(
555 /* We index by 32-bit vec4s */
556 unsigned byte
= (index
* 4 * 4);
558 midgard_instruction ins
= {
559 .type
= TAG_LOAD_STORE_4
,
562 .src
= { ~0, ~0, ~0, ~0 },
563 .swizzle
= SWIZZLE_IDENTITY_4
,
565 .op
= is_store
? midgard_op_st_int4
: midgard_op_ld_int4
,
567 /* For register spilling - to thread local storage */
572 /* If we spill an unspill, RA goes into an infinite loop */
573 .no_spill
= (1 << REG_CLASS_WORK
)
576 ins
.constants
.u32
[0] = byte
;
579 ins
.src
[0] = srcdest
;
581 /* Ensure we are tightly swizzled so liveness analysis is
584 for (unsigned i
= 0; i
< 4; ++i
) {
585 if (!(mask
& (1 << i
)))
586 ins
.swizzle
[0][i
] = COMPONENT_X
;
595 mir_has_arg(midgard_instruction
*ins
, unsigned arg
)
600 mir_foreach_src(ins
, i
) {
601 if (ins
->src
[i
] == arg
)
610 void midgard_schedule_program(compiler_context
*ctx
);
612 void mir_ra(compiler_context
*ctx
);
613 void mir_squeeze_index(compiler_context
*ctx
);
614 void mir_lower_special_reads(compiler_context
*ctx
);
615 void mir_liveness_ins_update(uint16_t *live
, midgard_instruction
*ins
, unsigned max
);
616 void mir_compute_liveness(compiler_context
*ctx
);
617 void mir_invalidate_liveness(compiler_context
*ctx
);
618 bool mir_is_live_after(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*start
, int src
);
620 void mir_create_pipeline_registers(compiler_context
*ctx
);
621 void midgard_promote_uniforms(compiler_context
*ctx
);
624 midgard_emit_derivatives(compiler_context
*ctx
, nir_alu_instr
*instr
);
627 midgard_lower_derivatives(compiler_context
*ctx
, midgard_block
*block
);
629 bool mir_op_computes_derivatives(gl_shader_stage stage
, unsigned op
);
631 void mir_analyze_helper_terminate(compiler_context
*ctx
);
632 void mir_analyze_helper_requirements(compiler_context
*ctx
);
636 void emit_binary_bundle(
637 compiler_context
*ctx
,
638 midgard_bundle
*bundle
,
639 struct util_dynarray
*emission
,
643 nir_undef_to_zero(nir_shader
*shader
);
645 void midgard_nir_lod_errata(nir_shader
*shader
);
649 bool midgard_opt_copy_prop(compiler_context
*ctx
, midgard_block
*block
);
650 bool midgard_opt_combine_projection(compiler_context
*ctx
, midgard_block
*block
);
651 bool midgard_opt_varying_projection(compiler_context
*ctx
, midgard_block
*block
);
652 bool midgard_opt_dead_code_eliminate(compiler_context
*ctx
, midgard_block
*block
);
653 bool midgard_opt_dead_move_eliminate(compiler_context
*ctx
, midgard_block
*block
);
654 bool midgard_opt_promote_fmov(compiler_context
*ctx
, midgard_block
*block
);