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
54 #define TARGET_TILEBUF_WAIT 4
56 typedef struct midgard_branch
{
57 /* If conditional, the condition is specified in r31.w */
60 /* For conditionals, if this is true, we branch on FALSE. If false, we branch on TRUE. */
61 bool invert_conditional
;
63 /* Branch targets: the start of a block, the start of a loop (continue), the end of a loop (break). Value is one of TARGET_ */
66 /* The actual target */
74 #define PAN_WRITEOUT_C 1
75 #define PAN_WRITEOUT_Z 2
76 #define PAN_WRITEOUT_S 4
78 /* Generic in-memory data type repesenting a single logical instruction, rather
79 * than a single instruction group. This is the preferred form for code gen.
80 * Multiple midgard_insturctions will later be combined during scheduling,
81 * though this is not represented in this structure. Its format bridges
82 * the low-level binary representation with the higher level semantic meaning.
84 * Notably, it allows registers to be specified as block local SSA, for code
85 * emitted before the register allocation pass.
88 #define MIR_SRC_COUNT 4
89 #define MIR_VEC_COMPONENTS 16
91 typedef struct midgard_instruction
{
92 /* Must be first for casting */
93 struct list_head link
;
95 unsigned type
; /* ALU, load/store, texture */
97 /* Instruction arguments represented as block-local SSA
98 * indices, rather than registers. ~0 means unused. */
99 unsigned src
[MIR_SRC_COUNT
];
102 /* vec16 swizzle, unpacked, per source */
103 unsigned swizzle
[MIR_SRC_COUNT
][MIR_VEC_COMPONENTS
];
106 nir_alu_type src_types
[MIR_SRC_COUNT
];
107 nir_alu_type dest_type
;
109 /* Packing ops have non-32-bit dest types even though they functionally
110 * work at the 32-bit level, use this as a signal to disable copyprop.
111 * We maybe need synthetic pack ops instead. */
114 /* Modifiers, depending on type */
117 bool src_abs
[MIR_SRC_COUNT
];
118 bool src_neg
[MIR_SRC_COUNT
];
122 bool src_shift
[MIR_SRC_COUNT
];
126 /* Out of the union for csel (could maybe be fixed..) */
127 bool src_invert
[MIR_SRC_COUNT
];
129 /* If the op supports it */
130 enum midgard_roundmode roundmode
;
132 /* Special fields for an ALU instruction */
133 midgard_reg_info registers
;
135 /* For textures: should helpers execute this instruction (instead of
136 * just helping with derivatives)? Should helpers terminate after? */
137 bool helper_terminate
;
140 /* I.e. (1 << alu_bit) */
144 midgard_constants constants
;
145 uint16_t inline_constant
;
146 bool has_blend_constant
;
147 bool has_inline_constant
;
153 /* Masks in a saneish format. One bit per channel, not packed fancy.
154 * Use this instead of the op specific ones, and switch over at emit
159 /* Hint for the register allocator not to spill the destination written
160 * from this instruction (because it is a spill/unspill node itself).
161 * Bitmask of spilled classes */
165 /* Generic hint for intra-pass use */
168 /* During scheduling, the backwards dependency graph
169 * (DAG). nr_dependencies is the number of unscheduled
170 * instructions that must still be scheduled after
171 * (before) this instruction. dependents are which
172 * instructions need to be scheduled before (after) this
175 unsigned nr_dependencies
;
176 BITSET_WORD
*dependents
;
178 /* Use this in conjunction with `type` */
182 midgard_load_store_word load_store
;
183 midgard_vector_alu alu
;
184 midgard_texture_word texture
;
185 midgard_branch_extended branch_extended
;
188 /* General branch, rather than packed br_compact. Higher level
189 * than the other components */
190 midgard_branch branch
;
192 } midgard_instruction
;
194 typedef struct midgard_block
{
199 /* List of midgard_bundles emitted (after the scheduler has run) */
200 struct util_dynarray bundles
;
202 /* Number of quadwords _actually_ emitted, as determined after scheduling */
203 unsigned quadword_count
;
205 /* Indicates this is a fixed-function fragment epilogue block */
208 /* Are helper invocations required by this block? */
212 typedef struct midgard_bundle
{
213 /* Tag for the overall bundle */
216 /* Instructions contained by the bundle. instruction_count <= 6 (vmul,
217 * sadd, vadd, smul, vlut, branch) */
218 int instruction_count
;
219 midgard_instruction
*instructions
[6];
221 /* Bundle-wide ALU configuration */
224 bool has_embedded_constants
;
225 midgard_constants constants
;
226 bool has_blend_constant
;
231 MIDGARD_COLOR_RT0
= 0,
243 typedef struct compiler_context
{
245 gl_shader_stage stage
;
247 /* Is internally a blend shader? Depends on stage == FRAGMENT */
250 /* Render target number for a keyed blend shader. Depends on is_blend */
253 /* Index to precolour to r0 for an input blend colour */
254 unsigned blend_input
;
256 /* Index to precolour to r2 for a dual-source blend colour */
259 /* Tracking for blend constant patching */
260 int blend_constant_offset
;
262 /* Number of bytes used for Thread Local Storage */
265 /* Count of spills and fills for shaderdb */
269 /* Current NIR function */
272 /* Allocated compiler temporary counter */
275 /* Unordered list of midgard_blocks */
277 struct list_head blocks
;
279 /* TODO merge with block_count? */
280 unsigned block_source_count
;
282 /* List of midgard_instructions emitted for the current block */
283 midgard_block
*current_block
;
285 /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
286 midgard_block
*after_block
;
288 /* The current "depth" of the loop, for disambiguating breaks/continues
289 * when using nested loops */
290 int current_loop_depth
;
292 /* Total number of loops for shader-db */
295 /* Constants which have been loaded, for later inlining */
296 struct hash_table_u64
*ssa_constants
;
298 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
299 struct hash_table_u64
*hash_to_temp
;
303 /* Set of NIR indices that were already emitted as outmods */
304 BITSET_WORD
*already_emitted
;
306 /* Just the count of the max register used. Higher count => higher
307 * register pressure */
310 /* The number of uniforms allowable for the fast path */
313 /* Count of instructions emitted from NIR overall, across all blocks */
314 int instruction_count
;
316 /* Alpha ref value passed in */
319 unsigned quadword_count
;
321 /* Bitmask of valid metadata */
324 /* Model-specific quirk set */
327 /* Writeout instructions for each render target */
328 midgard_instruction
*writeout_branch
[MIDGARD_NUM_RTS
];
330 struct panfrost_sysvals sysvals
;
333 /* Per-block live_in/live_out */
334 #define MIDGARD_METADATA_LIVENESS (1 << 0)
336 /* Helpers for manipulating the above structures (forming the driver IR) */
338 /* Append instruction to end of current block */
340 static inline midgard_instruction
*
341 mir_upload_ins(struct compiler_context
*ctx
, struct midgard_instruction ins
)
343 midgard_instruction
*heap
= ralloc(ctx
, struct midgard_instruction
);
344 memcpy(heap
, &ins
, sizeof(ins
));
348 static inline midgard_instruction
*
349 emit_mir_instruction(struct compiler_context
*ctx
, struct midgard_instruction ins
)
351 midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
352 list_addtail(&u
->link
, &ctx
->current_block
->base
.instructions
);
356 static inline struct midgard_instruction
*
357 mir_insert_instruction_before(struct compiler_context
*ctx
,
358 struct midgard_instruction
*tag
,
359 struct midgard_instruction ins
)
361 struct midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
362 list_addtail(&u
->link
, &tag
->link
);
367 mir_remove_instruction(struct midgard_instruction
*ins
)
369 list_del(&ins
->link
);
372 static inline midgard_instruction
*
373 mir_prev_op(struct midgard_instruction
*ins
)
375 return list_last_entry(&(ins
->link
), midgard_instruction
, link
);
378 static inline midgard_instruction
*
379 mir_next_op(struct midgard_instruction
*ins
)
381 return list_first_entry(&(ins
->link
), midgard_instruction
, link
);
384 #define mir_foreach_block(ctx, v) \
385 list_for_each_entry(pan_block, v, &ctx->blocks, link)
387 #define mir_foreach_block_from(ctx, from, v) \
388 list_for_each_entry_from(pan_block, v, &from->base, &ctx->blocks, link)
390 #define mir_foreach_instr_in_block(block, v) \
391 list_for_each_entry(struct midgard_instruction, v, &block->base.instructions, link)
392 #define mir_foreach_instr_in_block_rev(block, v) \
393 list_for_each_entry_rev(struct midgard_instruction, v, &block->base.instructions, link)
395 #define mir_foreach_instr_in_block_safe(block, v) \
396 list_for_each_entry_safe(struct midgard_instruction, v, &block->base.instructions, link)
398 #define mir_foreach_instr_in_block_safe_rev(block, v) \
399 list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->base.instructions, link)
401 #define mir_foreach_instr_in_block_from(block, v, from) \
402 list_for_each_entry_from(struct midgard_instruction, v, from, &block->base.instructions, link)
404 #define mir_foreach_instr_in_block_from_rev(block, v, from) \
405 list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->base.instructions, link)
407 #define mir_foreach_bundle_in_block(block, v) \
408 util_dynarray_foreach(&block->bundles, midgard_bundle, v)
410 #define mir_foreach_bundle_in_block_rev(block, v) \
411 util_dynarray_foreach_reverse(&block->bundles, midgard_bundle, v)
413 #define mir_foreach_instr_in_block_scheduled_rev(block, v) \
414 midgard_instruction* v; \
416 mir_foreach_bundle_in_block_rev(block, _bundle) \
417 for (i = (_bundle->instruction_count - 1), v = _bundle->instructions[i]; \
419 --i, v = (i >= 0) ? _bundle->instructions[i] : NULL) \
421 #define mir_foreach_instr_global(ctx, v) \
422 mir_foreach_block(ctx, v_block) \
423 mir_foreach_instr_in_block(((midgard_block *) v_block), v)
425 #define mir_foreach_instr_global_safe(ctx, v) \
426 mir_foreach_block(ctx, v_block) \
427 mir_foreach_instr_in_block_safe(((midgard_block *) v_block), v)
429 /* Based on set_foreach, expanded with automatic type casts */
431 #define mir_foreach_predecessor(blk, v) \
432 struct set_entry *_entry_##v; \
433 struct midgard_block *v; \
434 for (_entry_##v = _mesa_set_next_entry(blk->base.predecessors, NULL), \
435 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL); \
436 _entry_##v != NULL; \
437 _entry_##v = _mesa_set_next_entry(blk->base.predecessors, _entry_##v), \
438 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
440 #define mir_foreach_src(ins, v) \
441 for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
443 static inline midgard_instruction
*
444 mir_last_in_block(struct midgard_block
*block
)
446 return list_last_entry(&block
->base
.instructions
, struct midgard_instruction
, link
);
449 static inline midgard_block
*
450 mir_get_block(compiler_context
*ctx
, int idx
)
452 struct list_head
*lst
= &ctx
->blocks
;
457 return (struct midgard_block
*) lst
;
461 mir_is_alu_bundle(midgard_bundle
*bundle
)
463 return IS_ALU(bundle
->tag
);
466 static inline unsigned
467 make_compiler_temp(compiler_context
*ctx
)
469 return (ctx
->func
->impl
->ssa_alloc
+ ctx
->temp_alloc
++) << 1;
472 static inline unsigned
473 make_compiler_temp_reg(compiler_context
*ctx
)
475 return ((ctx
->func
->impl
->reg_alloc
+ ctx
->temp_alloc
++) << 1) | PAN_IS_REG
;
478 static inline unsigned
479 nir_ssa_index(nir_ssa_def
*ssa
)
481 return (ssa
->index
<< 1) | 0;
484 static inline unsigned
485 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
488 return nir_ssa_index(src
->ssa
);
490 assert(!src
->reg
.indirect
);
491 return (src
->reg
.reg
->index
<< 1) | PAN_IS_REG
;
495 static inline unsigned
496 nir_dest_index(nir_dest
*dst
)
499 return (dst
->ssa
.index
<< 1) | 0;
501 assert(!dst
->reg
.indirect
);
502 return (dst
->reg
.reg
->index
<< 1) | PAN_IS_REG
;
508 /* MIR manipulation */
510 void mir_rewrite_index(compiler_context
*ctx
, unsigned old
, unsigned new);
511 void mir_rewrite_index_src(compiler_context
*ctx
, unsigned old
, unsigned new);
512 void mir_rewrite_index_dst(compiler_context
*ctx
, unsigned old
, unsigned new);
513 void mir_rewrite_index_dst_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
514 void mir_rewrite_index_src_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
515 void mir_rewrite_index_src_swizzle(compiler_context
*ctx
, unsigned old
, unsigned new, unsigned *swizzle
);
516 bool mir_single_use(compiler_context
*ctx
, unsigned value
);
517 unsigned mir_use_count(compiler_context
*ctx
, unsigned value
);
518 uint16_t mir_bytemask_of_read_components(midgard_instruction
*ins
, unsigned node
);
519 uint16_t mir_bytemask_of_read_components_index(midgard_instruction
*ins
, unsigned i
);
520 uint16_t mir_from_bytemask(uint16_t bytemask
, unsigned bits
);
521 uint16_t mir_bytemask(midgard_instruction
*ins
);
522 uint16_t mir_round_bytemask_up(uint16_t mask
, unsigned bits
);
523 void mir_set_bytemask(midgard_instruction
*ins
, uint16_t bytemask
);
524 signed mir_upper_override(midgard_instruction
*ins
, unsigned inst_size
);
525 unsigned mir_components_for_type(nir_alu_type T
);
526 unsigned max_bitsize_for_alu(midgard_instruction
*ins
);
527 midgard_reg_mode
reg_mode_for_bitsize(unsigned bitsize
);
531 void mir_print_instruction(midgard_instruction
*ins
);
532 void mir_print_bundle(midgard_bundle
*ctx
);
533 void mir_print_block(midgard_block
*block
);
534 void mir_print_shader(compiler_context
*ctx
);
535 bool mir_nontrivial_mod(midgard_instruction
*ins
, unsigned i
, bool check_swizzle
);
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 void mir_set_offset(compiler_context
*ctx
, midgard_instruction
*ins
, nir_src
*offset
, bool is_shared
);
545 /* 'Intrinsic' move for aliasing */
547 static inline midgard_instruction
548 v_mov(unsigned src
, unsigned dest
)
550 midgard_instruction ins
= {
553 .src
= { ~0, src
, ~0, ~0 },
554 .src_types
= { 0, nir_type_uint32
},
555 .swizzle
= SWIZZLE_IDENTITY
,
557 .dest_type
= nir_type_uint32
,
558 .op
= midgard_alu_op_imov
,
560 .outmod
= midgard_outmod_int_wrap
567 /* Broad types of register classes so we can handle special
570 #define REG_CLASS_WORK 0
571 #define REG_CLASS_LDST 1
572 #define REG_CLASS_TEXR 3
573 #define REG_CLASS_TEXW 4
575 /* Like a move, but to thread local storage! */
577 static inline midgard_instruction
578 v_load_store_scratch(
584 /* We index by 32-bit vec4s */
585 unsigned byte
= (index
* 4 * 4);
587 midgard_instruction ins
= {
588 .type
= TAG_LOAD_STORE_4
,
590 .dest_type
= nir_type_uint32
,
592 .src
= { ~0, ~0, ~0, ~0 },
593 .swizzle
= SWIZZLE_IDENTITY_4
,
595 .op
= is_store
? midgard_op_st_int4
: midgard_op_ld_int4
,
597 /* For register spilling - to thread local storage */
602 /* If we spill an unspill, RA goes into an infinite loop */
603 .no_spill
= (1 << REG_CLASS_WORK
)
606 ins
.constants
.u32
[0] = byte
;
609 ins
.src
[0] = srcdest
;
610 ins
.src_types
[0] = nir_type_uint32
;
612 /* Ensure we are tightly swizzled so liveness analysis is
615 for (unsigned i
= 0; i
< 4; ++i
) {
616 if (!(mask
& (1 << i
)))
617 ins
.swizzle
[0][i
] = COMPONENT_X
;
626 mir_has_arg(midgard_instruction
*ins
, unsigned arg
)
631 mir_foreach_src(ins
, i
) {
632 if (ins
->src
[i
] == arg
)
641 void midgard_schedule_program(compiler_context
*ctx
);
643 void mir_ra(compiler_context
*ctx
);
644 void mir_squeeze_index(compiler_context
*ctx
);
645 void mir_lower_special_reads(compiler_context
*ctx
);
646 void mir_liveness_ins_update(uint16_t *live
, midgard_instruction
*ins
, unsigned max
);
647 void mir_compute_liveness(compiler_context
*ctx
);
648 void mir_invalidate_liveness(compiler_context
*ctx
);
649 bool mir_is_live_after(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*start
, int src
);
651 void mir_create_pipeline_registers(compiler_context
*ctx
);
652 void midgard_promote_uniforms(compiler_context
*ctx
);
655 midgard_emit_derivatives(compiler_context
*ctx
, nir_alu_instr
*instr
);
658 midgard_lower_derivatives(compiler_context
*ctx
, midgard_block
*block
);
660 bool mir_op_computes_derivatives(gl_shader_stage stage
, unsigned op
);
662 void mir_analyze_helper_terminate(compiler_context
*ctx
);
663 void mir_analyze_helper_requirements(compiler_context
*ctx
);
667 void emit_binary_bundle(
668 compiler_context
*ctx
,
669 midgard_block
*block
,
670 midgard_bundle
*bundle
,
671 struct util_dynarray
*emission
,
675 nir_undef_to_zero(nir_shader
*shader
);
676 bool nir_fuse_io_16(nir_shader
*shader
);
678 void midgard_nir_lod_errata(nir_shader
*shader
);
682 bool midgard_opt_copy_prop(compiler_context
*ctx
, midgard_block
*block
);
683 bool midgard_opt_combine_projection(compiler_context
*ctx
, midgard_block
*block
);
684 bool midgard_opt_varying_projection(compiler_context
*ctx
, midgard_block
*block
);
685 bool midgard_opt_dead_code_eliminate(compiler_context
*ctx
);
686 bool midgard_opt_dead_move_eliminate(compiler_context
*ctx
, midgard_block
*block
);