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 midgard_block
{
169 /* Link to next block. Must be first for mir_get_block */
170 struct list_head link
;
172 /* List of midgard_instructions emitted for the current block */
173 struct list_head instructions
;
175 /* Index of the block in source order */
180 /* List of midgard_bundles emitted (after the scheduler has run) */
181 struct util_dynarray bundles
;
183 /* Number of quadwords _actually_ emitted, as determined after scheduling */
184 unsigned quadword_count
;
186 /* Succeeding blocks. The compiler should not necessarily rely on
187 * source-order traversal */
188 struct midgard_block
*successors
[2];
189 unsigned nr_successors
;
191 struct set
*predecessors
;
193 /* The successors pointer form a graph, and in the case of
194 * complex control flow, this graph has a cycles. To aid
195 * traversal during liveness analysis, we have a visited?
196 * boolean for passes to use as they see fit, provided they
200 /* In liveness analysis, these are live masks (per-component) for
201 * indices for the block. Scalar compilers have the luxury of using
202 * simple bit fields, but for us, liveness is a vector idea. */
206 /* Indicates this is a fixed-function fragment epilogue block */
210 typedef struct midgard_bundle
{
211 /* Tag for the overall bundle */
214 /* Instructions contained by the bundle. instruction_count <= 6 (vmul,
215 * sadd, vadd, smul, vlut, branch) */
216 int instruction_count
;
217 midgard_instruction
*instructions
[6];
219 /* Bundle-wide ALU configuration */
222 bool has_embedded_constants
;
223 midgard_constants constants
;
224 bool has_blend_constant
;
237 struct panfrost_sysvals
{
238 /* The mapping of sysvals to uniforms, the count, and the off-by-one inverse */
239 unsigned sysvals
[MAX_SYSVAL_COUNT
];
240 unsigned sysval_count
;
241 struct hash_table_u64
*sysval_to_id
;
244 typedef struct compiler_context
{
246 gl_shader_stage stage
;
248 /* Is internally a blend shader? Depends on stage == FRAGMENT */
251 /* Render target number for a keyed blend shader. Depends on is_blend */
254 /* Tracking for blend constant patching */
255 int blend_constant_offset
;
257 /* Number of bytes used for Thread Local Storage */
260 /* Count of spills and fills for shaderdb */
264 /* Current NIR function */
267 /* Allocated compiler temporary counter */
270 /* Unordered list of midgard_blocks */
272 struct list_head blocks
;
274 /* TODO merge with block_count? */
275 unsigned block_source_count
;
277 /* List of midgard_instructions emitted for the current block */
278 midgard_block
*current_block
;
280 /* If there is a preset after block, use this, otherwise emit_block will create one if NULL */
281 midgard_block
*after_block
;
283 /* The current "depth" of the loop, for disambiguating breaks/continues
284 * when using nested loops */
285 int current_loop_depth
;
287 /* Total number of loops for shader-db */
290 /* Constants which have been loaded, for later inlining */
291 struct hash_table_u64
*ssa_constants
;
293 /* Mapping of hashes computed from NIR indices to the sequential temp indices ultimately used in MIR */
294 struct hash_table_u64
*hash_to_temp
;
298 /* Just the count of the max register used. Higher count => higher
299 * register pressure */
302 /* Used for cont/last hinting. Increase when a tex op is added.
303 * Decrease when a tex op is removed. */
304 int texture_op_count
;
306 /* The number of uniforms allowable for the fast path */
309 /* Count of instructions emitted from NIR overall, across all blocks */
310 int instruction_count
;
312 /* Alpha ref value passed in */
315 unsigned quadword_count
;
317 /* Bitmask of valid metadata */
320 /* Model-specific quirk set */
323 /* Writeout instructions for each render target */
324 midgard_instruction
*writeout_branch
[MIDGARD_NUM_RTS
];
326 struct panfrost_sysvals sysvals
;
329 /* Per-block live_in/live_out */
330 #define MIDGARD_METADATA_LIVENESS (1 << 0)
332 /* Helpers for manipulating the above structures (forming the driver IR) */
334 /* Append instruction to end of current block */
336 static inline midgard_instruction
*
337 mir_upload_ins(struct compiler_context
*ctx
, struct midgard_instruction ins
)
339 midgard_instruction
*heap
= ralloc(ctx
, struct midgard_instruction
);
340 memcpy(heap
, &ins
, sizeof(ins
));
344 static inline midgard_instruction
*
345 emit_mir_instruction(struct compiler_context
*ctx
, struct midgard_instruction ins
)
347 midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
348 list_addtail(&u
->link
, &ctx
->current_block
->instructions
);
352 static inline struct midgard_instruction
*
353 mir_insert_instruction_before(struct compiler_context
*ctx
,
354 struct midgard_instruction
*tag
,
355 struct midgard_instruction ins
)
357 struct midgard_instruction
*u
= mir_upload_ins(ctx
, ins
);
358 list_addtail(&u
->link
, &tag
->link
);
363 mir_remove_instruction(struct midgard_instruction
*ins
)
365 list_del(&ins
->link
);
368 static inline midgard_instruction
*
369 mir_prev_op(struct midgard_instruction
*ins
)
371 return list_last_entry(&(ins
->link
), midgard_instruction
, link
);
374 static inline midgard_instruction
*
375 mir_next_op(struct midgard_instruction
*ins
)
377 return list_first_entry(&(ins
->link
), midgard_instruction
, link
);
380 #define mir_foreach_block(ctx, v) \
381 list_for_each_entry(struct midgard_block, v, &ctx->blocks, link)
383 #define mir_foreach_block_from(ctx, from, v) \
384 list_for_each_entry_from(struct midgard_block, v, from, &ctx->blocks, link)
386 #define mir_foreach_instr_in_block(block, v) \
387 list_for_each_entry(struct midgard_instruction, v, &block->instructions, link)
388 #define mir_foreach_instr_in_block_rev(block, v) \
389 list_for_each_entry_rev(struct midgard_instruction, v, &block->instructions, link)
391 #define mir_foreach_instr_in_block_safe(block, v) \
392 list_for_each_entry_safe(struct midgard_instruction, v, &block->instructions, link)
394 #define mir_foreach_instr_in_block_safe_rev(block, v) \
395 list_for_each_entry_safe_rev(struct midgard_instruction, v, &block->instructions, link)
397 #define mir_foreach_instr_in_block_from(block, v, from) \
398 list_for_each_entry_from(struct midgard_instruction, v, from, &block->instructions, link)
400 #define mir_foreach_instr_in_block_from_rev(block, v, from) \
401 list_for_each_entry_from_rev(struct midgard_instruction, v, from, &block->instructions, link)
403 #define mir_foreach_bundle_in_block(block, v) \
404 util_dynarray_foreach(&block->bundles, midgard_bundle, v)
406 #define mir_foreach_bundle_in_block_rev(block, v) \
407 util_dynarray_foreach_reverse(&block->bundles, midgard_bundle, v)
409 #define mir_foreach_instr_in_block_scheduled_rev(block, v) \
410 midgard_instruction* v; \
412 mir_foreach_bundle_in_block_rev(block, _bundle) \
413 for (i = (_bundle->instruction_count - 1), v = _bundle->instructions[i]; \
415 --i, v = (i >= 0) ? _bundle->instructions[i] : NULL) \
417 #define mir_foreach_instr_global(ctx, v) \
418 mir_foreach_block(ctx, v_block) \
419 mir_foreach_instr_in_block(v_block, v)
421 #define mir_foreach_instr_global_safe(ctx, v) \
422 mir_foreach_block(ctx, v_block) \
423 mir_foreach_instr_in_block_safe(v_block, v)
425 #define mir_foreach_successor(blk, v) \
426 struct midgard_block *v; \
427 struct midgard_block **_v; \
428 for (_v = &blk->successors[0], \
430 v != NULL && _v < &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->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->predecessors, _entry_##v), \
442 v = (struct midgard_block *) (_entry_##v ? _entry_##v->key : NULL))
444 #define mir_foreach_src(ins, v) \
445 for (unsigned v = 0; v < ARRAY_SIZE(ins->src); ++v)
447 static inline midgard_instruction
*
448 mir_last_in_block(struct midgard_block
*block
)
450 return list_last_entry(&block
->instructions
, struct midgard_instruction
, link
);
453 static inline midgard_block
*
454 mir_get_block(compiler_context
*ctx
, int idx
)
456 struct list_head
*lst
= &ctx
->blocks
;
461 return (struct midgard_block
*) lst
;
464 static inline midgard_block
*
465 mir_exit_block(struct compiler_context
*ctx
)
467 midgard_block
*last
= list_last_entry(&ctx
->blocks
,
468 struct midgard_block
, link
);
470 /* The last block must be empty logically but contains branch writeout
471 * for fragment shaders */
473 assert(last
->nr_successors
== 0);
479 mir_is_alu_bundle(midgard_bundle
*bundle
)
481 return IS_ALU(bundle
->tag
);
484 /* Registers/SSA are distinguish in the backend by the bottom-most bit */
488 static inline unsigned
489 make_compiler_temp(compiler_context
*ctx
)
491 return (ctx
->func
->impl
->ssa_alloc
+ ctx
->temp_alloc
++) << 1;
494 static inline unsigned
495 make_compiler_temp_reg(compiler_context
*ctx
)
497 return ((ctx
->func
->impl
->reg_alloc
+ ctx
->temp_alloc
++) << 1) | IS_REG
;
500 static inline unsigned
501 nir_ssa_index(nir_ssa_def
*ssa
)
503 return (ssa
->index
<< 1) | 0;
506 static inline unsigned
507 nir_src_index(compiler_context
*ctx
, nir_src
*src
)
510 return nir_ssa_index(src
->ssa
);
512 assert(!src
->reg
.indirect
);
513 return (src
->reg
.reg
->index
<< 1) | IS_REG
;
517 static inline unsigned
518 nir_alu_src_index(compiler_context
*ctx
, nir_alu_src
*src
)
520 return nir_src_index(ctx
, &src
->src
);
523 static inline unsigned
524 nir_dest_index(nir_dest
*dst
)
527 return (dst
->ssa
.index
<< 1) | 0;
529 assert(!dst
->reg
.indirect
);
530 return (dst
->reg
.reg
->index
<< 1) | IS_REG
;
536 /* MIR manipulation */
538 void mir_rewrite_index(compiler_context
*ctx
, unsigned old
, unsigned new);
539 void mir_rewrite_index_src(compiler_context
*ctx
, unsigned old
, unsigned new);
540 void mir_rewrite_index_dst(compiler_context
*ctx
, unsigned old
, unsigned new);
541 void mir_rewrite_index_dst_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
542 void mir_rewrite_index_src_single(midgard_instruction
*ins
, unsigned old
, unsigned new);
543 void mir_rewrite_index_src_swizzle(compiler_context
*ctx
, unsigned old
, unsigned new, unsigned *swizzle
);
544 bool mir_single_use(compiler_context
*ctx
, unsigned value
);
545 bool mir_special_index(compiler_context
*ctx
, unsigned idx
);
546 unsigned mir_use_count(compiler_context
*ctx
, unsigned value
);
547 bool mir_is_written_before(compiler_context
*ctx
, midgard_instruction
*ins
, unsigned node
);
548 uint16_t mir_bytemask_of_read_components(midgard_instruction
*ins
, unsigned node
);
549 uint16_t mir_bytemask_of_read_components_index(midgard_instruction
*ins
, unsigned i
);
550 midgard_reg_mode
mir_typesize(midgard_instruction
*ins
);
551 midgard_reg_mode
mir_srcsize(midgard_instruction
*ins
, unsigned i
);
552 unsigned mir_bytes_for_mode(midgard_reg_mode mode
);
553 midgard_reg_mode
mir_mode_for_destsize(unsigned size
);
554 uint16_t mir_from_bytemask(uint16_t bytemask
, midgard_reg_mode mode
);
555 uint16_t mir_bytemask(midgard_instruction
*ins
);
556 uint16_t mir_round_bytemask_up(uint16_t mask
, midgard_reg_mode mode
);
557 void mir_set_bytemask(midgard_instruction
*ins
, uint16_t bytemask
);
558 unsigned mir_upper_override(midgard_instruction
*ins
);
562 void mir_print_instruction(midgard_instruction
*ins
);
563 void mir_print_bundle(midgard_bundle
*ctx
);
564 void mir_print_block(midgard_block
*block
);
565 void mir_print_shader(compiler_context
*ctx
);
566 bool mir_nontrivial_source2_mod(midgard_instruction
*ins
);
567 bool mir_nontrivial_source2_mod_simple(midgard_instruction
*ins
);
568 bool mir_nontrivial_outmod(midgard_instruction
*ins
);
570 void mir_insert_instruction_before_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
571 void mir_insert_instruction_after_scheduled(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*tag
, midgard_instruction ins
);
572 void mir_flip(midgard_instruction
*ins
);
573 void mir_compute_temp_count(compiler_context
*ctx
);
575 void mir_set_offset(compiler_context
*ctx
, midgard_instruction
*ins
, nir_src
*offset
, bool is_shared
);
577 /* 'Intrinsic' move for aliasing */
579 static inline midgard_instruction
580 v_mov(unsigned src
, unsigned dest
)
582 midgard_instruction ins
= {
585 .src
= { ~0, src
, ~0, ~0 },
586 .swizzle
= SWIZZLE_IDENTITY
,
589 .op
= midgard_alu_op_imov
,
590 .reg_mode
= midgard_reg_mode_32
,
591 .dest_override
= midgard_dest_override_none
,
592 .outmod
= midgard_outmod_int_wrap
599 /* Broad types of register classes so we can handle special
602 #define REG_CLASS_WORK 0
603 #define REG_CLASS_LDST 1
604 #define REG_CLASS_TEXR 3
605 #define REG_CLASS_TEXW 4
607 /* Like a move, but to thread local storage! */
609 static inline midgard_instruction
610 v_load_store_scratch(
616 /* We index by 32-bit vec4s */
617 unsigned byte
= (index
* 4 * 4);
619 midgard_instruction ins
= {
620 .type
= TAG_LOAD_STORE_4
,
623 .src
= { ~0, ~0, ~0, ~0 },
624 .swizzle
= SWIZZLE_IDENTITY_4
,
626 .op
= is_store
? midgard_op_st_int4
: midgard_op_ld_int4
,
628 /* For register spilling - to thread local storage */
633 /* If we spill an unspill, RA goes into an infinite loop */
634 .no_spill
= (1 << REG_CLASS_WORK
)
637 ins
.constants
.u32
[0] = byte
;
640 ins
.src
[0] = srcdest
;
642 /* Ensure we are tightly swizzled so liveness analysis is
645 for (unsigned i
= 0; i
< 4; ++i
) {
646 if (!(mask
& (1 << i
)))
647 ins
.swizzle
[0][i
] = COMPONENT_X
;
656 mir_has_arg(midgard_instruction
*ins
, unsigned arg
)
661 for (unsigned i
= 0; i
< ARRAY_SIZE(ins
->src
); ++i
) {
662 if (ins
->src
[i
] == arg
)
671 void midgard_schedule_program(compiler_context
*ctx
);
673 void mir_ra(compiler_context
*ctx
);
674 void mir_squeeze_index(compiler_context
*ctx
);
675 void mir_lower_special_reads(compiler_context
*ctx
);
676 void mir_liveness_ins_update(uint16_t *live
, midgard_instruction
*ins
, unsigned max
);
677 void mir_compute_liveness(compiler_context
*ctx
);
678 void mir_invalidate_liveness(compiler_context
*ctx
);
679 bool mir_is_live_after(compiler_context
*ctx
, midgard_block
*block
, midgard_instruction
*start
, int src
);
681 void mir_create_pipeline_registers(compiler_context
*ctx
);
682 void midgard_promote_uniforms(compiler_context
*ctx
);
685 midgard_emit_derivatives(compiler_context
*ctx
, nir_alu_instr
*instr
);
688 midgard_lower_derivatives(compiler_context
*ctx
, midgard_block
*block
);
690 bool mir_op_computes_derivatives(gl_shader_stage stage
, unsigned op
);
694 void emit_binary_bundle(
695 compiler_context
*ctx
,
696 midgard_bundle
*bundle
,
697 struct util_dynarray
*emission
,
701 nir_undef_to_zero(nir_shader
*shader
);
703 void midgard_nir_lod_errata(nir_shader
*shader
);
707 bool midgard_opt_copy_prop(compiler_context
*ctx
, midgard_block
*block
);
708 bool midgard_opt_combine_projection(compiler_context
*ctx
, midgard_block
*block
);
709 bool midgard_opt_varying_projection(compiler_context
*ctx
, midgard_block
*block
);
710 bool midgard_opt_dead_code_eliminate(compiler_context
*ctx
, midgard_block
*block
);
711 bool midgard_opt_dead_move_eliminate(compiler_context
*ctx
, midgard_block
*block
);
713 void midgard_lower_invert(compiler_context
*ctx
, midgard_block
*block
);
714 bool midgard_opt_not_propagate(compiler_context
*ctx
, midgard_block
*block
);
715 bool midgard_opt_fuse_src_invert(compiler_context
*ctx
, midgard_block
*block
);
716 bool midgard_opt_fuse_dest_invert(compiler_context
*ctx
, midgard_block
*block
);
717 bool midgard_opt_csel_invert(compiler_context
*ctx
, midgard_block
*block
);
718 bool midgard_opt_promote_fmov(compiler_context
*ctx
, midgard_block
*block
);
719 bool midgard_opt_drop_cmp_invert(compiler_context
*ctx
, midgard_block
*block
);
720 bool midgard_opt_invert_branch(compiler_context
*ctx
, midgard_block
*block
);