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"
41 #include "panfrost/util/pan_ir.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
];
100 /* Special fields for an ALU instruction */
101 midgard_reg_info registers
;
103 /* I.e. (1 << alu_bit) */
107 midgard_constants constants
;
108 uint16_t inline_constant
;
109 bool has_blend_constant
;
110 bool has_inline_constant
;
115 bool writeout_stencil
;
118 /* Kind of a hack, but hint against aggressive DCE */
121 /* Masks in a saneish format. One bit per channel, not packed fancy.
122 * Use this instead of the op specific ones, and switch over at emit
127 /* For ALU ops only: set to true to invert (bitwise NOT) the
128 * destination of an integer-out op. Not implemented in hardware but
129 * allows more optimizations */
133 /* Hint for the register allocator not to spill the destination written
134 * from this instruction (because it is a spill/unspill node itself).
135 * Bitmask of spilled classes */
139 /* Generic hint for intra-pass use */
142 /* During scheduling, the backwards dependency graph
143 * (DAG). nr_dependencies is the number of unscheduled
144 * instructions that must still be scheduled after
145 * (before) this instruction. dependents are which
146 * instructions need to be scheduled before (after) this
149 unsigned nr_dependencies
;
150 BITSET_WORD
*dependents
;
152 /* For load/store ops.. force 64-bit destination */
156 midgard_load_store_word load_store
;
157 midgard_vector_alu alu
;
158 midgard_texture_word texture
;
159 midgard_branch_extended branch_extended
;
162 /* General branch, rather than packed br_compact. Higher level
163 * than the other components */
164 midgard_branch branch
;
166 } midgard_instruction
;
168 typedef struct pan_block
{
169 /* Link to next block. Must be first for mir_get_block */
170 struct list_head link
;
172 /* List of instructions emitted for the current block */
173 struct list_head instructions
;
175 /* Index of the block in source order */
178 /* Control flow graph */
179 struct pan_block
*successors
[2];
180 unsigned nr_successors
;
182 struct set
*predecessors
;
184 /* In liveness analysis, these are live masks (per-component) for
185 * indices for the block. Scalar compilers have the luxury of using
186 * simple bit fields, but for us, liveness is a vector idea. */
191 typedef struct midgard_block
{
196 /* List of midgard_bundles emitted (after the scheduler has run) */
197 struct util_dynarray bundles
;
199 /* Number of quadwords _actually_ emitted, as determined after scheduling */
200 unsigned quadword_count
;
202 /* Indicates this is a fixed-function fragment epilogue block */
206 typedef struct midgard_bundle
{
207 /* Tag for the overall bundle */
210 /* Instructions contained by the bundle. instruction_count <= 6 (vmul,
211 * sadd, vadd, smul, vlut, branch) */
212 int instruction_count
;
213 midgard_instruction
*instructions
[6];
215 /* Bundle-wide ALU configuration */
218 bool has_embedded_constants
;
219 midgard_constants constants
;
220 bool has_blend_constant
;
233 typedef struct compiler_context
{
235 gl_shader_stage stage
;
237 /* Is internally a blend shader? Depends on stage == FRAGMENT */
240 /* Render target number for a keyed blend shader. Depends on is_blend */
243 /* Tracking for blend constant patching */
244 int blend_constant_offset
;
246 /* Number of bytes used for Thread Local Storage */
249 /* Count of spills and fills for shaderdb */
253 /* Current NIR function */
256 /* Allocated compiler temporary counter */
259 /* Unordered list of midgard_blocks */
261 struct list_head blocks
;
263 /* TODO merge with block_count? */
264 unsigned block_source_count
;
266 /* List of midgard_instructions emitted for the current block */
267 midgard_block
*current_block
;
269 /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
270 midgard_block
*after_block
;
272 /* The current "depth" of the loop, for disambiguating breaks/continues
273 * when using nested loops */
274 int current_loop_depth
;
276 /* Total number of loops for shader-db */
279 /* Constants which have been loaded, for later inlining */
280 struct hash_table_u64
*ssa_constants
;
282 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
283 struct hash_table_u64
*hash_to_temp
;
287 /* Just the count of the max register used. Higher count => higher
288 * register pressure */
291 /* Used for cont/last hinting. Increase when a tex op is added.
292 * Decrease when a tex op is removed. */
293 int texture_op_count
;
295 /* The number of uniforms allowable for the fast path */
298 /* Count of instructions emitted from NIR overall, across all blocks */
299 int instruction_count
;
301 /* Alpha ref value passed in */
304 unsigned quadword_count
;
306 /* Bitmask of valid metadata */
309 /* Model-specific quirk set */
312 /* Writeout instructions for each render target */
313 midgard_instruction
*writeout_branch
[MIDGARD_NUM_RTS
];
315 struct panfrost_sysvals sysvals
;
318 /* Per-block live_in/live_out */
319 #define MIDGARD_METADATA_LIVENESS (1 << 0)
321 /* Helpers for manipulating the above structures (forming the driver IR) */
323 /* Append instruction to end of current block */
325 static inline midgard_instruction
*
326 mir_upload_ins(struct compiler_context
*ctx
, struct midgard_instruction ins
)
328 midgard_instruction
*heap
= ralloc(ctx
, struct midgard_instruction
);
329 memcpy(heap
, &ins
, sizeof(ins
));
333 static inline midgard_instruction
*
334 emit_mir_instruction(struct compiler_context
*ctx
, struct midgard_instruction ins
)
336 midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
337 list_addtail(&u
->link
, &ctx
->current_block
->base
.instructions
);
341 static inline struct midgard_instruction
*
342 mir_insert_instruction_before(struct compiler_context
*ctx
,
343 struct midgard_instruction
*tag
,
344 struct midgard_instruction ins
)
346 struct midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
347 list_addtail(&u
->link
, &tag
->link
);
352 mir_remove_instruction(struct midgard_instruction
*ins
)
354 list_del(&ins
->link
);
357 static inline midgard_instruction
*
358 mir_prev_op(struct midgard_instruction
*ins
)
360 return list_last_entry(&(ins
->link
), midgard_instruction
, link
);
363 static inline midgard_instruction
*
364 mir_next_op(struct midgard_instruction
*ins
)
366 return list_first_entry(&(ins
->link
), midgard_instruction
, link
);
369 #define mir_foreach_block(ctx, v) \
370 list_for_each_entry(pan_block, v, &ctx->blocks, link)
372 #define mir_foreach_block_from(ctx, from, v) \
373 list_for_each_entry_from(pan_block, v, &from->base, &ctx->blocks, link)
375 #define mir_foreach_instr_in_block(block, v) \
376 list_for_each_entry(struct midgard_instruction, v, &block->base.instructions, link)
377 #define mir_foreach_instr_in_block_rev(block, v) \
378 list_for_each_entry_rev(struct midgard_instruction, v, &block->base.instructions, link)
380 #define pan_foreach_instr_in_block_rev(block, v) \
381 list_for_each_entry_rev(struct midgard_instruction, v, &block->instructions, link)
383 #define mir_foreach_instr_in_block_safe(block, v) \
384 list_for_each_entry_safe(struct midgard_instruction, v, &block->base.instructions, link)
386 #define mir_foreach_instr_in_block_safe_rev(block, v) \
387 list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->base.instructions, link)
389 #define mir_foreach_instr_in_block_from(block, v, from) \
390 list_for_each_entry_from(struct midgard_instruction, v, from, &block->base.instructions, link)
392 #define mir_foreach_instr_in_block_from_rev(block, v, from) \
393 list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->base.instructions, link)
395 #define mir_foreach_bundle_in_block(block, v) \
396 util_dynarray_foreach(&block->bundles, midgard_bundle, v)
398 #define mir_foreach_bundle_in_block_rev(block, v) \
399 util_dynarray_foreach_reverse(&block->bundles, midgard_bundle, v)
401 #define mir_foreach_instr_in_block_scheduled_rev(block, v) \
402 midgard_instruction* v; \
404 mir_foreach_bundle_in_block_rev(block, _bundle) \
405 for (i = (_bundle->instruction_count - 1), v = _bundle->instructions[i]; \
407 --i, v = (i >= 0) ? _bundle->instructions[i] : NULL) \
409 #define mir_foreach_instr_global(ctx, v) \
410 mir_foreach_block(ctx, v_block) \
411 mir_foreach_instr_in_block(((midgard_block *) v_block), v)
413 #define mir_foreach_instr_global_safe(ctx, v) \
414 mir_foreach_block(ctx, v_block) \
415 mir_foreach_instr_in_block_safe(((midgard_block *) v_block), v)
417 #define mir_foreach_successor(blk, v) \
418 struct midgard_block *v; \
419 struct midgard_block **_v; \
420 for (_v = &blk->base.successors[0], \
422 v != NULL && _v < &blk->base.successors[2]; \
425 #define pan_foreach_successor(blk, v) \
428 for (_v = (pan_block **) &blk->successors[0], \
430 v != NULL && _v < (pan_block **) &blk->successors[2]; \
433 /* Based on set_foreach, expanded with automatic type casts */
435 #define mir_foreach_predecessor(blk, v) \
436 struct set_entry *_entry_##v; \
437 struct midgard_block *v; \
438 for (_entry_##v = _mesa_set_next_entry(blk->base.predecessors, NULL), \
439 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL); \
440 _entry_##v != NULL; \
441 _entry_##v = _mesa_set_next_entry(blk->base.predecessors, _entry_##v), \
442 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
444 #define pan_foreach_predecessor(blk, v) \
445 struct set_entry *_entry_##v; \
446 struct pan_block *v; \
447 for (_entry_##v = _mesa_set_next_entry(blk->predecessors, NULL), \
448 v = (struct pan_block *) (_entry_##v ? _entry_##v->key : NULL); \
449 _entry_##v != NULL; \
450 _entry_##v = _mesa_set_next_entry(blk->predecessors, _entry_##v), \
451 v = (struct pan_block *) (_entry_##v ? _entry_##v->key : NULL))
453 #define mir_foreach_src(ins, v) \
454 for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
456 static inline midgard_instruction
*
457 mir_last_in_block(struct midgard_block
*block
)
459 return list_last_entry(&block
->base
.instructions
, struct midgard_instruction
, link
);
462 static inline midgard_block
*
463 mir_get_block(compiler_context
*ctx
, int idx
)
465 struct list_head
*lst
= &ctx
->blocks
;
470 return (struct midgard_block
*) lst
;
473 static inline midgard_block
*
474 mir_exit_block(struct compiler_context
*ctx
)
476 pan_block
*last
= list_last_entry(&ctx
->blocks
, pan_block
, link
);
478 /* The last block must be empty logically but contains branch writeout
479 * for fragment shaders */
481 assert(last
->nr_successors
== 0);
483 return (midgard_block
*) last
;
487 mir_is_alu_bundle(midgard_bundle
*bundle
)
489 return IS_ALU(bundle
->tag
);
492 /* Registers/SSA are distinguish in the backend by the bottom-most bit */
496 static inline unsigned
497 make_compiler_temp(compiler_context
*ctx
)
499 return (ctx
->func
->impl
->ssa_alloc
+ ctx
->temp_alloc
++) << 1;
502 static inline unsigned
503 make_compiler_temp_reg(compiler_context
*ctx
)
505 return ((ctx
->func
->impl
->reg_alloc
+ ctx
->temp_alloc
++) << 1) | IS_REG
;
508 static inline unsigned
509 nir_ssa_index(nir_ssa_def
*ssa
)
511 return (ssa
->index
<< 1) | 0;
514 static inline unsigned
515 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
518 return nir_ssa_index(src
->ssa
);
520 assert(!src
->reg
.indirect
);
521 return (src
->reg
.reg
->index
<< 1) | IS_REG
;
525 static inline unsigned
526 nir_alu_src_index(compiler_context
*ctx
, nir_alu_src
*src
)
528 return nir_src_index(ctx
, &src
->src
);
531 static inline unsigned
532 nir_dest_index(nir_dest
*dst
)
535 return (dst
->ssa
.index
<< 1) | 0;
537 assert(!dst
->reg
.indirect
);
538 return (dst
->reg
.reg
->index
<< 1) | IS_REG
;
544 /* MIR manipulation */
546 void mir_rewrite_index(compiler_context
*ctx
, unsigned old
, unsigned new);
547 void mir_rewrite_index_src(compiler_context
*ctx
, unsigned old
, unsigned new);
548 void mir_rewrite_index_dst(compiler_context
*ctx
, unsigned old
, unsigned new);
549 void mir_rewrite_index_dst_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
550 void mir_rewrite_index_src_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
551 void mir_rewrite_index_src_swizzle(compiler_context
*ctx
, unsigned old
, unsigned new, unsigned *swizzle
);
552 bool mir_single_use(compiler_context
*ctx
, unsigned value
);
553 bool mir_special_index(compiler_context
*ctx
, unsigned idx
);
554 unsigned mir_use_count(compiler_context
*ctx
, unsigned value
);
555 bool mir_is_written_before(compiler_context
*ctx
, midgard_instruction
*ins
, unsigned node
);
556 uint16_t mir_bytemask_of_read_components(midgard_instruction
*ins
, unsigned node
);
557 uint16_t mir_bytemask_of_read_components_index(midgard_instruction
*ins
, unsigned i
);
558 midgard_reg_mode
mir_typesize(midgard_instruction
*ins
);
559 midgard_reg_mode
mir_srcsize(midgard_instruction
*ins
, unsigned i
);
560 unsigned mir_bytes_for_mode(midgard_reg_mode mode
);
561 midgard_reg_mode
mir_mode_for_destsize(unsigned size
);
562 uint16_t mir_from_bytemask(uint16_t bytemask
, midgard_reg_mode mode
);
563 uint16_t mir_bytemask(midgard_instruction
*ins
);
564 uint16_t mir_round_bytemask_up(uint16_t mask
, midgard_reg_mode mode
);
565 void mir_set_bytemask(midgard_instruction
*ins
, uint16_t bytemask
);
566 unsigned mir_upper_override(midgard_instruction
*ins
);
570 void mir_print_instruction(midgard_instruction
*ins
);
571 void mir_print_bundle(midgard_bundle
*ctx
);
572 void mir_print_block(midgard_block
*block
);
573 void mir_print_shader(compiler_context
*ctx
);
574 bool mir_nontrivial_source2_mod(midgard_instruction
*ins
);
575 bool mir_nontrivial_source2_mod_simple(midgard_instruction
*ins
);
576 bool mir_nontrivial_outmod(midgard_instruction
*ins
);
578 void mir_insert_instruction_before_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
579 void mir_insert_instruction_after_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
580 void mir_flip(midgard_instruction
*ins
);
581 void mir_compute_temp_count(compiler_context
*ctx
);
583 void mir_set_offset(compiler_context
*ctx
, midgard_instruction
*ins
, nir_src
*offset
, bool is_shared
);
585 /* 'Intrinsic' move for aliasing */
587 static inline midgard_instruction
588 v_mov(unsigned src
, unsigned dest
)
590 midgard_instruction ins
= {
593 .src
= { ~0, src
, ~0, ~0 },
594 .swizzle
= SWIZZLE_IDENTITY
,
597 .op
= midgard_alu_op_imov
,
598 .reg_mode
= midgard_reg_mode_32
,
599 .dest_override
= midgard_dest_override_none
,
600 .outmod
= midgard_outmod_int_wrap
607 /* Broad types of register classes so we can handle special
610 #define REG_CLASS_WORK 0
611 #define REG_CLASS_LDST 1
612 #define REG_CLASS_TEXR 3
613 #define REG_CLASS_TEXW 4
615 /* Like a move, but to thread local storage! */
617 static inline midgard_instruction
618 v_load_store_scratch(
624 /* We index by 32-bit vec4s */
625 unsigned byte
= (index
* 4 * 4);
627 midgard_instruction ins
= {
628 .type
= TAG_LOAD_STORE_4
,
631 .src
= { ~0, ~0, ~0, ~0 },
632 .swizzle
= SWIZZLE_IDENTITY_4
,
634 .op
= is_store
? midgard_op_st_int4
: midgard_op_ld_int4
,
636 /* For register spilling - to thread local storage */
641 /* If we spill an unspill, RA goes into an infinite loop */
642 .no_spill
= (1 << REG_CLASS_WORK
)
645 ins
.constants
.u32
[0] = byte
;
648 ins
.src
[0] = srcdest
;
650 /* Ensure we are tightly swizzled so liveness analysis is
653 for (unsigned i
= 0; i
< 4; ++i
) {
654 if (!(mask
& (1 << i
)))
655 ins
.swizzle
[0][i
] = COMPONENT_X
;
664 mir_has_arg(midgard_instruction
*ins
, unsigned arg
)
669 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
->src
); ++i
) {
670 if (ins
->src
[i
] == arg
)
679 void midgard_schedule_program(compiler_context
*ctx
);
681 void mir_ra(compiler_context
*ctx
);
682 void mir_squeeze_index(compiler_context
*ctx
);
683 void mir_lower_special_reads(compiler_context
*ctx
);
684 void mir_liveness_ins_update(uint16_t *live
, midgard_instruction
*ins
, unsigned max
);
685 void mir_compute_liveness(compiler_context
*ctx
);
686 void mir_invalidate_liveness(compiler_context
*ctx
);
687 bool mir_is_live_after(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*start
, int src
);
689 void mir_create_pipeline_registers(compiler_context
*ctx
);
690 void midgard_promote_uniforms(compiler_context
*ctx
);
693 midgard_emit_derivatives(compiler_context
*ctx
, nir_alu_instr
*instr
);
696 midgard_lower_derivatives(compiler_context
*ctx
, midgard_block
*block
);
698 bool mir_op_computes_derivatives(gl_shader_stage stage
, unsigned op
);
702 void emit_binary_bundle(
703 compiler_context
*ctx
,
704 midgard_bundle
*bundle
,
705 struct util_dynarray
*emission
,
709 nir_undef_to_zero(nir_shader
*shader
);
711 void midgard_nir_lod_errata(nir_shader
*shader
);
715 bool midgard_opt_copy_prop(compiler_context
*ctx
, midgard_block
*block
);
716 bool midgard_opt_combine_projection(compiler_context
*ctx
, midgard_block
*block
);
717 bool midgard_opt_varying_projection(compiler_context
*ctx
, midgard_block
*block
);
718 bool midgard_opt_dead_code_eliminate(compiler_context
*ctx
, midgard_block
*block
);
719 bool midgard_opt_dead_move_eliminate(compiler_context
*ctx
, midgard_block
*block
);
721 void midgard_lower_invert(compiler_context
*ctx
, midgard_block
*block
);
722 bool midgard_opt_not_propagate(compiler_context
*ctx
, midgard_block
*block
);
723 bool midgard_opt_fuse_src_invert(compiler_context
*ctx
, midgard_block
*block
);
724 bool midgard_opt_fuse_dest_invert(compiler_context
*ctx
, midgard_block
*block
);
725 bool midgard_opt_csel_invert(compiler_context
*ctx
, midgard_block
*block
);
726 bool midgard_opt_promote_fmov(compiler_context
*ctx
, midgard_block
*block
);
727 bool midgard_opt_drop_cmp_invert(compiler_context
*ctx
, midgard_block
*block
);
728 bool midgard_opt_invert_branch(compiler_context
*ctx
, midgard_block
*block
);