2 * Copyright © 2016 Intel Corporation
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 #include "nir_builder.h"
26 #include "nir_deref.h"
28 #include "util/bitscan.h"
29 #include "util/u_dynarray.h"
32 * Variable-based copy propagation
34 * Normally, NIR trusts in SSA form for most of its copy-propagation needs.
35 * However, there are cases, especially when dealing with indirects, where SSA
36 * won't help you. This pass is for those times. Specifically, it handles
37 * the following things that the rest of NIR can't:
39 * 1) Copy-propagation on variables that have indirect access. This includes
40 * propagating from indirect stores into indirect loads.
42 * 2) Removal of redundant load_deref intrinsics. We can't trust regular CSE
43 * to do this because it isn't aware of variable writes that may alias the
44 * value and make the former load invalid.
46 * This pass uses an intermediate solution between being local / "per-block"
47 * and a complete data-flow analysis. It follows the control flow graph, and
48 * propagate the available copy information forward, invalidating data at each
51 * Removal of dead writes to variables is handled by another pass.
55 nir_variable_mode modes
;
57 /* Key is deref and value is the uintptr_t with the write mask. */
58 struct hash_table
*derefs
;
65 nir_deref_instr
*deref
;
75 struct copy_prop_var_state
{
76 nir_function_impl
*impl
;
81 /* Maps nodes to vars_written. Used to invalidate copy entries when
84 struct hash_table
*vars_written_map
;
90 value_equals_store_src(struct value
*value
, nir_intrinsic_instr
*intrin
)
92 assert(intrin
->intrinsic
== nir_intrinsic_store_deref
);
93 uintptr_t write_mask
= nir_intrinsic_write_mask(intrin
);
95 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
96 if ((write_mask
& (1 << i
)) &&
97 value
->ssa
[i
] != intrin
->src
[1].ssa
)
104 static struct vars_written
*
105 create_vars_written(struct copy_prop_var_state
*state
)
107 struct vars_written
*written
=
108 linear_zalloc_child(state
->lin_ctx
, sizeof(struct vars_written
));
109 written
->derefs
= _mesa_pointer_hash_table_create(state
->mem_ctx
);
114 gather_vars_written(struct copy_prop_var_state
*state
,
115 struct vars_written
*written
,
116 nir_cf_node
*cf_node
)
118 struct vars_written
*new_written
= NULL
;
120 switch (cf_node
->type
) {
121 case nir_cf_node_function
: {
122 nir_function_impl
*impl
= nir_cf_node_as_function(cf_node
);
123 foreach_list_typed_safe(nir_cf_node
, cf_node
, node
, &impl
->body
)
124 gather_vars_written(state
, NULL
, cf_node
);
128 case nir_cf_node_block
: {
132 nir_block
*block
= nir_cf_node_as_block(cf_node
);
133 nir_foreach_instr(instr
, block
) {
134 if (instr
->type
== nir_instr_type_call
) {
135 written
->modes
|= nir_var_shader_out
|
143 if (instr
->type
!= nir_instr_type_intrinsic
)
146 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
147 switch (intrin
->intrinsic
) {
148 case nir_intrinsic_barrier
:
149 case nir_intrinsic_memory_barrier
:
150 written
->modes
|= nir_var_shader_out
|
155 case nir_intrinsic_emit_vertex
:
156 case nir_intrinsic_emit_vertex_with_counter
:
157 written
->modes
= nir_var_shader_out
;
160 case nir_intrinsic_store_deref
:
161 case nir_intrinsic_copy_deref
: {
162 /* Destination in _both_ store_deref and copy_deref is src[0]. */
163 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
165 uintptr_t mask
= intrin
->intrinsic
== nir_intrinsic_store_deref
?
166 nir_intrinsic_write_mask(intrin
) : (1 << glsl_get_vector_elements(dst
->type
)) - 1;
168 struct hash_entry
*ht_entry
= _mesa_hash_table_search(written
->derefs
, dst
);
170 ht_entry
->data
= (void *)(mask
| (uintptr_t)ht_entry
->data
);
172 _mesa_hash_table_insert(written
->derefs
, dst
, (void *)mask
);
185 case nir_cf_node_if
: {
186 nir_if
*if_stmt
= nir_cf_node_as_if(cf_node
);
188 new_written
= create_vars_written(state
);
190 foreach_list_typed_safe(nir_cf_node
, cf_node
, node
, &if_stmt
->then_list
)
191 gather_vars_written(state
, new_written
, cf_node
);
193 foreach_list_typed_safe(nir_cf_node
, cf_node
, node
, &if_stmt
->else_list
)
194 gather_vars_written(state
, new_written
, cf_node
);
199 case nir_cf_node_loop
: {
200 nir_loop
*loop
= nir_cf_node_as_loop(cf_node
);
202 new_written
= create_vars_written(state
);
204 foreach_list_typed_safe(nir_cf_node
, cf_node
, node
, &loop
->body
)
205 gather_vars_written(state
, new_written
, cf_node
);
211 unreachable("Invalid CF node type");
215 /* Merge new information to the parent control flow node. */
217 written
->modes
|= new_written
->modes
;
218 hash_table_foreach(new_written
->derefs
, new_entry
) {
219 struct hash_entry
*old_entry
=
220 _mesa_hash_table_search_pre_hashed(written
->derefs
, new_entry
->hash
,
223 nir_component_mask_t merged
= (uintptr_t) new_entry
->data
|
224 (uintptr_t) old_entry
->data
;
225 old_entry
->data
= (void *) ((uintptr_t) merged
);
227 _mesa_hash_table_insert_pre_hashed(written
->derefs
, new_entry
->hash
,
228 new_entry
->key
, new_entry
->data
);
232 _mesa_hash_table_insert(state
->vars_written_map
, cf_node
, new_written
);
236 static struct copy_entry
*
237 copy_entry_create(struct util_dynarray
*copies
,
238 nir_deref_instr
*dst_deref
)
240 struct copy_entry new_entry
= {
243 util_dynarray_append(copies
, struct copy_entry
, new_entry
);
244 return util_dynarray_top_ptr(copies
, struct copy_entry
);
247 /* Remove copy entry by swapping it with the last element and reducing the
248 * size. If used inside an iteration on copies, it must be a reverse
249 * (backwards) iteration. It is safe to use in those cases because the swap
250 * will not affect the rest of the iteration.
253 copy_entry_remove(struct util_dynarray
*copies
,
254 struct copy_entry
*entry
)
256 /* This also works when removing the last element since pop don't shrink
257 * the memory used by the array, so the swap is useless but not invalid.
259 *entry
= util_dynarray_pop(copies
, struct copy_entry
);
262 static struct copy_entry
*
263 lookup_entry_for_deref(struct util_dynarray
*copies
,
264 nir_deref_instr
*deref
,
265 nir_deref_compare_result allowed_comparisons
)
267 util_dynarray_foreach(copies
, struct copy_entry
, iter
) {
268 if (nir_compare_derefs(iter
->dst
, deref
) & allowed_comparisons
)
275 static struct copy_entry
*
276 lookup_entry_and_kill_aliases(struct util_dynarray
*copies
,
277 nir_deref_instr
*deref
,
280 /* TODO: Take into account the write_mask. */
282 nir_deref_instr
*dst_match
= NULL
;
283 util_dynarray_foreach_reverse(copies
, struct copy_entry
, iter
) {
284 if (!iter
->src
.is_ssa
) {
285 /* If this write aliases the source of some entry, get rid of it */
286 if (nir_compare_derefs(iter
->src
.deref
, deref
) & nir_derefs_may_alias_bit
) {
287 copy_entry_remove(copies
, iter
);
292 nir_deref_compare_result comp
= nir_compare_derefs(iter
->dst
, deref
);
294 if (comp
& nir_derefs_equal_bit
) {
295 /* Removing entries invalidate previous iter pointers, so we'll
296 * collect the matching entry later. Just make sure it is unique.
299 dst_match
= iter
->dst
;
300 } else if (comp
& nir_derefs_may_alias_bit
) {
301 copy_entry_remove(copies
, iter
);
305 struct copy_entry
*entry
= NULL
;
307 util_dynarray_foreach(copies
, struct copy_entry
, iter
) {
308 if (iter
->dst
== dst_match
) {
319 kill_aliases(struct util_dynarray
*copies
,
320 nir_deref_instr
*deref
,
323 /* TODO: Take into account the write_mask. */
325 struct copy_entry
*entry
=
326 lookup_entry_and_kill_aliases(copies
, deref
, write_mask
);
328 copy_entry_remove(copies
, entry
);
331 static struct copy_entry
*
332 get_entry_and_kill_aliases(struct util_dynarray
*copies
,
333 nir_deref_instr
*deref
,
336 /* TODO: Take into account the write_mask. */
338 struct copy_entry
*entry
=
339 lookup_entry_and_kill_aliases(copies
, deref
, write_mask
);
342 entry
= copy_entry_create(copies
, deref
);
348 apply_barrier_for_modes(struct util_dynarray
*copies
,
349 nir_variable_mode modes
)
351 util_dynarray_foreach_reverse(copies
, struct copy_entry
, iter
) {
352 if ((iter
->dst
->mode
& modes
) ||
353 (!iter
->src
.is_ssa
&& (iter
->src
.deref
->mode
& modes
)))
354 copy_entry_remove(copies
, iter
);
359 store_to_entry(struct copy_prop_var_state
*state
, struct copy_entry
*entry
,
360 const struct value
*value
, unsigned write_mask
)
363 /* Clear src if it was being used as non-SSA. */
364 if (!entry
->src
.is_ssa
)
365 memset(entry
->src
.ssa
, 0, sizeof(entry
->src
.ssa
));
366 entry
->src
.is_ssa
= true;
367 /* Only overwrite the written components */
368 for (unsigned i
= 0; i
< 4; i
++) {
369 if (write_mask
& (1 << i
))
370 entry
->src
.ssa
[i
] = value
->ssa
[i
];
373 /* Non-ssa stores always write everything */
374 entry
->src
.is_ssa
= false;
375 entry
->src
.deref
= value
->deref
;
379 /* Do a "load" from an SSA-based entry return it in "value" as a value with a
380 * single SSA def. Because an entry could reference up to 4 different SSA
381 * defs, a vecN operation may be inserted to combine them into a single SSA
382 * def before handing it back to the caller. If the load instruction is no
383 * longer needed, it is removed and nir_instr::block is set to NULL. (It is
384 * possible, in some cases, for the load to be used in the vecN operation in
385 * which case it isn't deleted.)
388 load_from_ssa_entry_value(struct copy_prop_var_state
*state
,
389 struct copy_entry
*entry
,
390 nir_builder
*b
, nir_intrinsic_instr
*intrin
,
394 assert(value
->is_ssa
);
396 const struct glsl_type
*type
= entry
->dst
->type
;
397 unsigned num_components
= glsl_get_vector_elements(type
);
399 nir_component_mask_t available
= 0;
400 bool all_same
= true;
401 for (unsigned i
= 0; i
< num_components
; i
++) {
403 available
|= (1 << i
);
405 if (value
->ssa
[i
] != value
->ssa
[0])
410 /* Our work here is done */
411 b
->cursor
= nir_instr_remove(&intrin
->instr
);
412 intrin
->instr
.block
= NULL
;
416 if (available
!= (1 << num_components
) - 1 &&
417 intrin
->intrinsic
== nir_intrinsic_load_deref
&&
418 (available
& nir_ssa_def_components_read(&intrin
->dest
.ssa
)) == 0) {
419 /* If none of the components read are available as SSA values, then we
420 * should just bail. Otherwise, we would end up replacing the uses of
421 * the load_deref a vecN() that just gathers up its components.
426 b
->cursor
= nir_after_instr(&intrin
->instr
);
428 nir_ssa_def
*load_def
=
429 intrin
->intrinsic
== nir_intrinsic_load_deref
? &intrin
->dest
.ssa
: NULL
;
431 bool keep_intrin
= false;
432 nir_ssa_def
*comps
[NIR_MAX_VEC_COMPONENTS
];
433 for (unsigned i
= 0; i
< num_components
; i
++) {
435 comps
[i
] = nir_channel(b
, value
->ssa
[i
], i
);
437 /* We don't have anything for this component in our
438 * list. Just re-use a channel from the load.
440 if (load_def
== NULL
)
441 load_def
= nir_load_deref(b
, entry
->dst
);
443 if (load_def
->parent_instr
== &intrin
->instr
)
446 comps
[i
] = nir_channel(b
, load_def
, i
);
450 nir_ssa_def
*vec
= nir_vec(b
, comps
, num_components
);
451 for (unsigned i
= 0; i
< num_components
; i
++)
455 /* Removing this instruction should not touch the cursor because we
456 * created the cursor after the intrinsic and have added at least one
457 * instruction (the vec) since then.
459 assert(b
->cursor
.instr
!= &intrin
->instr
);
460 nir_instr_remove(&intrin
->instr
);
461 intrin
->instr
.block
= NULL
;
468 * Specialize the wildcards in a deref chain
470 * This function returns a deref chain identical to \param deref except that
471 * some of its wildcards are replaced with indices from \param specific. The
472 * process is guided by \param guide which references the same type as \param
473 * specific but has the same wildcard array lengths as \param deref.
475 static nir_deref_instr
*
476 specialize_wildcards(nir_builder
*b
,
477 nir_deref_path
*deref
,
478 nir_deref_path
*guide
,
479 nir_deref_path
*specific
)
481 nir_deref_instr
**deref_p
= &deref
->path
[1];
482 nir_deref_instr
**guide_p
= &guide
->path
[1];
483 nir_deref_instr
**spec_p
= &specific
->path
[1];
484 nir_deref_instr
*ret_tail
= deref
->path
[0];
485 for (; *deref_p
; deref_p
++) {
486 if ((*deref_p
)->deref_type
== nir_deref_type_array_wildcard
) {
487 /* This is where things get tricky. We have to search through
488 * the entry deref to find its corresponding wildcard and fill
489 * this slot in with the value from the src.
492 (*guide_p
)->deref_type
!= nir_deref_type_array_wildcard
) {
496 assert(*guide_p
&& *spec_p
);
498 ret_tail
= nir_build_deref_follower(b
, ret_tail
, *spec_p
);
503 ret_tail
= nir_build_deref_follower(b
, ret_tail
, *deref_p
);
510 /* Do a "load" from an deref-based entry return it in "value" as a value. The
511 * deref returned in "value" will always be a fresh copy so the caller can
512 * steal it and assign it to the instruction directly without copying it
516 load_from_deref_entry_value(struct copy_prop_var_state
*state
,
517 struct copy_entry
*entry
,
518 nir_builder
*b
, nir_intrinsic_instr
*intrin
,
519 nir_deref_instr
*src
, struct value
*value
)
523 b
->cursor
= nir_instr_remove(&intrin
->instr
);
525 nir_deref_path entry_dst_path
, src_path
;
526 nir_deref_path_init(&entry_dst_path
, entry
->dst
, state
->mem_ctx
);
527 nir_deref_path_init(&src_path
, src
, state
->mem_ctx
);
529 bool need_to_specialize_wildcards
= false;
530 nir_deref_instr
**entry_p
= &entry_dst_path
.path
[1];
531 nir_deref_instr
**src_p
= &src_path
.path
[1];
532 while (*entry_p
&& *src_p
) {
533 nir_deref_instr
*entry_tail
= *entry_p
++;
534 nir_deref_instr
*src_tail
= *src_p
++;
536 if (src_tail
->deref_type
== nir_deref_type_array
&&
537 entry_tail
->deref_type
== nir_deref_type_array_wildcard
)
538 need_to_specialize_wildcards
= true;
541 /* If the entry deref is longer than the source deref then it refers to a
542 * smaller type and we can't source from it.
544 assert(*entry_p
== NULL
);
546 if (need_to_specialize_wildcards
) {
547 /* The entry has some wildcards that are not in src. This means we need
548 * to construct a new deref based on the entry but using the wildcards
549 * from the source and guided by the entry dst. Oof.
551 nir_deref_path entry_src_path
;
552 nir_deref_path_init(&entry_src_path
, entry
->src
.deref
, state
->mem_ctx
);
553 value
->deref
= specialize_wildcards(b
, &entry_src_path
,
554 &entry_dst_path
, &src_path
);
555 nir_deref_path_finish(&entry_src_path
);
558 /* If our source deref is longer than the entry deref, that's ok because
559 * it just means the entry deref needs to be extended a bit.
562 nir_deref_instr
*src_tail
= *src_p
++;
563 value
->deref
= nir_build_deref_follower(b
, value
->deref
, src_tail
);
566 nir_deref_path_finish(&entry_dst_path
);
567 nir_deref_path_finish(&src_path
);
573 try_load_from_entry(struct copy_prop_var_state
*state
, struct copy_entry
*entry
,
574 nir_builder
*b
, nir_intrinsic_instr
*intrin
,
575 nir_deref_instr
*src
, struct value
*value
)
580 if (entry
->src
.is_ssa
) {
581 return load_from_ssa_entry_value(state
, entry
, b
, intrin
, value
);
583 return load_from_deref_entry_value(state
, entry
, b
, intrin
, src
, value
);
588 invalidate_copies_for_cf_node(struct copy_prop_var_state
*state
,
589 struct util_dynarray
*copies
,
590 nir_cf_node
*cf_node
)
592 struct hash_entry
*ht_entry
= _mesa_hash_table_search(state
->vars_written_map
, cf_node
);
595 struct vars_written
*written
= ht_entry
->data
;
596 if (written
->modes
) {
597 util_dynarray_foreach_reverse(copies
, struct copy_entry
, entry
) {
598 if (entry
->dst
->mode
& written
->modes
)
599 copy_entry_remove(copies
, entry
);
603 hash_table_foreach (written
->derefs
, entry
) {
604 nir_deref_instr
*deref_written
= (nir_deref_instr
*)entry
->key
;
605 kill_aliases(copies
, deref_written
, (uintptr_t)entry
->data
);
610 copy_prop_vars_block(struct copy_prop_var_state
*state
,
611 nir_builder
*b
, nir_block
*block
,
612 struct util_dynarray
*copies
)
614 nir_foreach_instr_safe(instr
, block
) {
615 if (instr
->type
== nir_instr_type_call
) {
616 apply_barrier_for_modes(copies
, nir_var_shader_out
|
624 if (instr
->type
!= nir_instr_type_intrinsic
)
627 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
628 switch (intrin
->intrinsic
) {
629 case nir_intrinsic_barrier
:
630 case nir_intrinsic_memory_barrier
:
631 apply_barrier_for_modes(copies
, nir_var_shader_out
|
636 case nir_intrinsic_emit_vertex
:
637 case nir_intrinsic_emit_vertex_with_counter
:
638 apply_barrier_for_modes(copies
, nir_var_shader_out
);
641 case nir_intrinsic_load_deref
: {
642 nir_deref_instr
*src
= nir_src_as_deref(intrin
->src
[0]);
644 struct copy_entry
*src_entry
=
645 lookup_entry_for_deref(copies
, src
, nir_derefs_a_contains_b_bit
);
647 if (try_load_from_entry(state
, src_entry
, b
, intrin
, src
, &value
)) {
649 /* lookup_load has already ensured that we get a single SSA
650 * value that has all of the channels. We just have to do the
653 if (intrin
->instr
.block
) {
654 /* The lookup left our instruction in-place. This means it
655 * must have used it to vec up a bunch of different sources.
656 * We need to be careful when rewriting uses so we don't
657 * rewrite the vecN itself.
659 nir_ssa_def_rewrite_uses_after(&intrin
->dest
.ssa
,
660 nir_src_for_ssa(value
.ssa
[0]),
661 value
.ssa
[0]->parent_instr
);
663 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
664 nir_src_for_ssa(value
.ssa
[0]));
667 /* We're turning it into a load of a different variable */
668 intrin
->src
[0] = nir_src_for_ssa(&value
.deref
->dest
.ssa
);
670 /* Put it back in again. */
671 nir_builder_instr_insert(b
, instr
);
674 for (unsigned i
= 0; i
< intrin
->num_components
; i
++)
675 value
.ssa
[i
] = &intrin
->dest
.ssa
;
677 state
->progress
= true;
680 for (unsigned i
= 0; i
< intrin
->num_components
; i
++)
681 value
.ssa
[i
] = &intrin
->dest
.ssa
;
684 /* Now that we have a value, we're going to store it back so that we
685 * have the right value next time we come looking for it. In order
686 * to do this, we need an exact match, not just something that
687 * contains what we're looking for.
689 struct copy_entry
*store_entry
=
690 lookup_entry_for_deref(copies
, src
, nir_derefs_equal_bit
);
692 store_entry
= copy_entry_create(copies
, src
);
694 /* Set up a store to this entry with the value of the load. This way
695 * we can potentially remove subsequent loads. However, we use a
696 * NULL instruction so we don't try and delete the load on a
699 store_to_entry(state
, store_entry
, &value
,
700 ((1 << intrin
->num_components
) - 1));
704 case nir_intrinsic_store_deref
: {
705 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
706 struct copy_entry
*entry
=
707 lookup_entry_for_deref(copies
, dst
, nir_derefs_equal_bit
);
708 if (entry
&& value_equals_store_src(&entry
->src
, intrin
)) {
709 /* If we are storing the value from a load of the same var the
710 * store is redundant so remove it.
712 nir_instr_remove(instr
);
714 struct value value
= {
718 for (unsigned i
= 0; i
< intrin
->num_components
; i
++)
719 value
.ssa
[i
] = intrin
->src
[1].ssa
;
721 unsigned wrmask
= nir_intrinsic_write_mask(intrin
);
722 struct copy_entry
*entry
=
723 get_entry_and_kill_aliases(copies
, dst
, wrmask
);
724 store_to_entry(state
, entry
, &value
, wrmask
);
730 case nir_intrinsic_copy_deref
: {
731 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
732 nir_deref_instr
*src
= nir_src_as_deref(intrin
->src
[1]);
734 if (nir_compare_derefs(src
, dst
) & nir_derefs_equal_bit
) {
735 /* This is a no-op self-copy. Get rid of it */
736 nir_instr_remove(instr
);
740 struct copy_entry
*src_entry
=
741 lookup_entry_for_deref(copies
, src
, nir_derefs_a_contains_b_bit
);
743 if (try_load_from_entry(state
, src_entry
, b
, intrin
, src
, &value
)) {
744 /* If load works, intrin (the copy_deref) is removed. */
746 nir_store_deref(b
, dst
, value
.ssa
[0], 0xf);
748 /* If this would be a no-op self-copy, don't bother. */
749 if (nir_compare_derefs(value
.deref
, dst
) & nir_derefs_equal_bit
)
752 /* Just turn it into a copy of a different deref */
753 intrin
->src
[1] = nir_src_for_ssa(&value
.deref
->dest
.ssa
);
755 /* Put it back in again. */
756 nir_builder_instr_insert(b
, instr
);
759 state
->progress
= true;
761 value
= (struct value
) {
767 struct copy_entry
*dst_entry
=
768 get_entry_and_kill_aliases(copies
, dst
, 0xf);
769 store_to_entry(state
, dst_entry
, &value
, 0xf);
780 copy_prop_vars_cf_node(struct copy_prop_var_state
*state
,
781 struct util_dynarray
*copies
,
782 nir_cf_node
*cf_node
)
784 switch (cf_node
->type
) {
785 case nir_cf_node_function
: {
786 nir_function_impl
*impl
= nir_cf_node_as_function(cf_node
);
788 struct util_dynarray impl_copies
;
789 util_dynarray_init(&impl_copies
, state
->mem_ctx
);
791 foreach_list_typed_safe(nir_cf_node
, cf_node
, node
, &impl
->body
)
792 copy_prop_vars_cf_node(state
, &impl_copies
, cf_node
);
797 case nir_cf_node_block
: {
798 nir_block
*block
= nir_cf_node_as_block(cf_node
);
800 nir_builder_init(&b
, state
->impl
);
801 copy_prop_vars_block(state
, &b
, block
, copies
);
805 case nir_cf_node_if
: {
806 nir_if
*if_stmt
= nir_cf_node_as_if(cf_node
);
808 /* Clone the copies for each branch of the if statement. The idea is
809 * that they both see the same state of available copies, but do not
810 * interfere to each other.
813 struct util_dynarray then_copies
;
814 util_dynarray_clone(&then_copies
, state
->mem_ctx
, copies
);
816 struct util_dynarray else_copies
;
817 util_dynarray_clone(&else_copies
, state
->mem_ctx
, copies
);
819 foreach_list_typed_safe(nir_cf_node
, cf_node
, node
, &if_stmt
->then_list
)
820 copy_prop_vars_cf_node(state
, &then_copies
, cf_node
);
822 foreach_list_typed_safe(nir_cf_node
, cf_node
, node
, &if_stmt
->else_list
)
823 copy_prop_vars_cf_node(state
, &else_copies
, cf_node
);
825 /* Both branches copies can be ignored, since the effect of running both
826 * branches was captured in the first pass that collects vars_written.
829 invalidate_copies_for_cf_node(state
, copies
, cf_node
);
834 case nir_cf_node_loop
: {
835 nir_loop
*loop
= nir_cf_node_as_loop(cf_node
);
837 /* Invalidate before cloning the copies for the loop, since the loop
838 * body can be executed more than once.
841 invalidate_copies_for_cf_node(state
, copies
, cf_node
);
843 struct util_dynarray loop_copies
;
844 util_dynarray_clone(&loop_copies
, state
->mem_ctx
, copies
);
846 foreach_list_typed_safe(nir_cf_node
, cf_node
, node
, &loop
->body
)
847 copy_prop_vars_cf_node(state
, &loop_copies
, cf_node
);
853 unreachable("Invalid CF node type");
858 nir_copy_prop_vars_impl(nir_function_impl
*impl
)
860 void *mem_ctx
= ralloc_context(NULL
);
862 struct copy_prop_var_state state
= {
865 .lin_ctx
= linear_zalloc_parent(mem_ctx
, 0),
867 .vars_written_map
= _mesa_pointer_hash_table_create(mem_ctx
),
870 gather_vars_written(&state
, NULL
, &impl
->cf_node
);
872 copy_prop_vars_cf_node(&state
, NULL
, &impl
->cf_node
);
874 if (state
.progress
) {
875 nir_metadata_preserve(impl
, nir_metadata_block_index
|
876 nir_metadata_dominance
);
879 impl
->valid_metadata
&= ~nir_metadata_not_properly_reset
;
883 ralloc_free(mem_ctx
);
884 return state
.progress
;
888 nir_opt_copy_prop_vars(nir_shader
*shader
)
890 bool progress
= false;
892 nir_foreach_function(function
, shader
) {
895 progress
|= nir_copy_prop_vars_impl(function
->impl
);