2 * Copyright © 2018 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"
27 #include "util/hash_table.h"
30 is_trivial_deref_cast(nir_deref_instr
*cast
)
32 nir_deref_instr
*parent
= nir_src_as_deref(cast
->parent
);
36 return cast
->mode
== parent
->mode
&&
37 cast
->type
== parent
->type
&&
38 cast
->dest
.ssa
.num_components
== parent
->dest
.ssa
.num_components
&&
39 cast
->dest
.ssa
.bit_size
== parent
->dest
.ssa
.bit_size
;
43 nir_deref_path_init(nir_deref_path
*path
,
44 nir_deref_instr
*deref
, void *mem_ctx
)
46 assert(deref
!= NULL
);
48 /* The length of the short path is at most ARRAY_SIZE - 1 because we need
49 * room for the NULL terminator.
51 static const int max_short_path_len
= ARRAY_SIZE(path
->_short_path
) - 1;
55 nir_deref_instr
**tail
= &path
->_short_path
[max_short_path_len
];
56 nir_deref_instr
**head
= tail
;
59 for (nir_deref_instr
*d
= deref
; d
; d
= nir_deref_instr_parent(d
)) {
60 if (d
->deref_type
== nir_deref_type_cast
&& is_trivial_deref_cast(d
))
63 if (count
<= max_short_path_len
)
67 if (count
<= max_short_path_len
) {
68 /* If we're under max_short_path_len, just use the short path. */
74 /* Just in case someone uses short_path by accident */
75 for (unsigned i
= 0; i
< ARRAY_SIZE(path
->_short_path
); i
++)
76 path
->_short_path
[i
] = (void *)(uintptr_t)0xdeadbeef;
79 path
->path
= ralloc_array(mem_ctx
, nir_deref_instr
*, count
+ 1);
80 head
= tail
= path
->path
+ count
;
82 for (nir_deref_instr
*d
= deref
; d
; d
= nir_deref_instr_parent(d
)) {
83 if (d
->deref_type
== nir_deref_type_cast
&& is_trivial_deref_cast(d
))
89 assert(head
== path
->path
);
90 assert(tail
== head
+ count
);
91 assert(*tail
== NULL
);
95 nir_deref_path_finish(nir_deref_path
*path
)
97 if (path
->path
< &path
->_short_path
[0] ||
98 path
->path
> &path
->_short_path
[ARRAY_SIZE(path
->_short_path
) - 1])
99 ralloc_free(path
->path
);
103 * Recursively removes unused deref instructions
106 nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
)
108 bool progress
= false;
110 for (nir_deref_instr
*d
= instr
; d
; d
= nir_deref_instr_parent(d
)) {
111 /* If anyone is using this deref, leave it alone */
112 assert(d
->dest
.is_ssa
);
113 if (!list_is_empty(&d
->dest
.ssa
.uses
))
116 nir_instr_remove(&d
->instr
);
124 nir_deref_instr_has_indirect(nir_deref_instr
*instr
)
126 while (instr
->deref_type
!= nir_deref_type_var
) {
127 /* Consider casts to be indirects */
128 if (instr
->deref_type
== nir_deref_type_cast
)
131 if ((instr
->deref_type
== nir_deref_type_array
||
132 instr
->deref_type
== nir_deref_type_ptr_as_array
) &&
133 !nir_src_is_const(instr
->arr
.index
))
136 instr
= nir_deref_instr_parent(instr
);
143 nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
)
145 for (; instr
; instr
= nir_deref_instr_parent(instr
)) {
146 if (instr
->deref_type
== nir_deref_type_array
&&
147 nir_src_is_const(instr
->arr
.index
) &&
148 nir_src_as_uint(instr
->arr
.index
) >=
149 glsl_get_length(nir_deref_instr_parent(instr
)->type
))
157 nir_deref_instr_has_complex_use(nir_deref_instr
*deref
)
159 nir_foreach_use(use_src
, &deref
->dest
.ssa
) {
160 nir_instr
*use_instr
= use_src
->parent_instr
;
162 switch (use_instr
->type
) {
163 case nir_instr_type_deref
: {
164 nir_deref_instr
*use_deref
= nir_instr_as_deref(use_instr
);
166 /* A var deref has no sources */
167 assert(use_deref
->deref_type
!= nir_deref_type_var
);
169 /* If a deref shows up in an array index or something like that, it's
172 if (use_src
!= &use_deref
->parent
)
175 /* Anything that isn't a basic struct or array deref is considered to
176 * be a "complex" use. In particular, we don't allow ptr_as_array
177 * because we assume that opt_deref will turn any non-complex
178 * ptr_as_array derefs into regular array derefs eventually so passes
179 * which only want to handle simple derefs will pick them up in a
182 if (use_deref
->deref_type
!= nir_deref_type_struct
&&
183 use_deref
->deref_type
!= nir_deref_type_array_wildcard
&&
184 use_deref
->deref_type
!= nir_deref_type_array
)
187 if (nir_deref_instr_has_complex_use(use_deref
))
193 case nir_instr_type_intrinsic
: {
194 nir_intrinsic_instr
*use_intrin
= nir_instr_as_intrinsic(use_instr
);
195 switch (use_intrin
->intrinsic
) {
196 case nir_intrinsic_load_deref
:
197 assert(use_src
== &use_intrin
->src
[0]);
200 case nir_intrinsic_copy_deref
:
201 assert(use_src
== &use_intrin
->src
[0] ||
202 use_src
== &use_intrin
->src
[1]);
205 case nir_intrinsic_store_deref
:
206 /* A use in src[1] of a store means we're taking that pointer and
207 * writing it to a variable. Because we have no idea who will
208 * read that variable and what they will do with the pointer, it's
209 * considered a "complex" use. A use in src[0], on the other
210 * hand, is a simple use because we're just going to dereference
211 * it and write a value there.
213 if (use_src
== &use_intrin
->src
[0])
220 unreachable("Switch default failed");
228 nir_foreach_if_use(use
, &deref
->dest
.ssa
)
235 nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*deref
)
237 switch (deref
->deref_type
) {
238 case nir_deref_type_array
:
239 return glsl_get_explicit_stride(nir_deref_instr_parent(deref
)->type
);
240 case nir_deref_type_ptr_as_array
:
241 return nir_deref_instr_ptr_as_array_stride(nir_deref_instr_parent(deref
));
242 case nir_deref_type_cast
:
243 return deref
->cast
.ptr_stride
;
250 type_get_array_stride(const struct glsl_type
*elem_type
,
251 glsl_type_size_align_func size_align
)
253 unsigned elem_size
, elem_align
;
254 size_align(elem_type
, &elem_size
, &elem_align
);
255 return ALIGN_POT(elem_size
, elem_align
);
259 struct_type_get_field_offset(const struct glsl_type
*struct_type
,
260 glsl_type_size_align_func size_align
,
263 assert(glsl_type_is_struct_or_ifc(struct_type
));
265 for (unsigned i
= 0; i
<= field_idx
; i
++) {
266 unsigned elem_size
, elem_align
;
267 size_align(glsl_get_struct_field(struct_type
, i
), &elem_size
, &elem_align
);
268 offset
= ALIGN_POT(offset
, elem_align
);
276 nir_deref_instr_get_const_offset(nir_deref_instr
*deref
,
277 glsl_type_size_align_func size_align
)
280 nir_deref_path_init(&path
, deref
, NULL
);
282 assert(path
.path
[0]->deref_type
== nir_deref_type_var
);
285 for (nir_deref_instr
**p
= &path
.path
[1]; *p
; p
++) {
286 switch ((*p
)->deref_type
) {
287 case nir_deref_type_array
:
288 offset
+= nir_src_as_uint((*p
)->arr
.index
) *
289 type_get_array_stride((*p
)->type
, size_align
);
291 case nir_deref_type_struct
: {
292 /* p starts at path[1], so this is safe */
293 nir_deref_instr
*parent
= *(p
- 1);
294 offset
+= struct_type_get_field_offset(parent
->type
, size_align
,
299 unreachable("Unsupported deref type");
303 nir_deref_path_finish(&path
);
309 nir_build_deref_offset(nir_builder
*b
, nir_deref_instr
*deref
,
310 glsl_type_size_align_func size_align
)
313 nir_deref_path_init(&path
, deref
, NULL
);
315 assert(path
.path
[0]->deref_type
== nir_deref_type_var
);
317 nir_ssa_def
*offset
= nir_imm_intN_t(b
, 0, deref
->dest
.ssa
.bit_size
);
318 for (nir_deref_instr
**p
= &path
.path
[1]; *p
; p
++) {
319 switch ((*p
)->deref_type
) {
320 case nir_deref_type_array
: {
321 nir_ssa_def
*index
= nir_ssa_for_src(b
, (*p
)->arr
.index
, 1);
322 int stride
= type_get_array_stride((*p
)->type
, size_align
);
323 offset
= nir_iadd(b
, offset
, nir_amul_imm(b
, index
, stride
));
326 case nir_deref_type_struct
: {
327 /* p starts at path[1], so this is safe */
328 nir_deref_instr
*parent
= *(p
- 1);
329 unsigned field_offset
=
330 struct_type_get_field_offset(parent
->type
, size_align
,
332 offset
= nir_iadd_imm(b
, offset
, field_offset
);
336 unreachable("Unsupported deref type");
340 nir_deref_path_finish(&path
);
346 nir_remove_dead_derefs_impl(nir_function_impl
*impl
)
348 bool progress
= false;
350 nir_foreach_block(block
, impl
) {
351 nir_foreach_instr_safe(instr
, block
) {
352 if (instr
->type
== nir_instr_type_deref
&&
353 nir_deref_instr_remove_if_unused(nir_instr_as_deref(instr
)))
359 nir_metadata_preserve(impl
, nir_metadata_block_index
|
360 nir_metadata_dominance
);
366 nir_remove_dead_derefs(nir_shader
*shader
)
368 bool progress
= false;
369 nir_foreach_function(function
, shader
) {
370 if (function
->impl
&& nir_remove_dead_derefs_impl(function
->impl
))
378 nir_fixup_deref_modes(nir_shader
*shader
)
380 nir_foreach_function(function
, shader
) {
384 nir_foreach_block(block
, function
->impl
) {
385 nir_foreach_instr(instr
, block
) {
386 if (instr
->type
!= nir_instr_type_deref
)
389 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
390 if (deref
->deref_type
== nir_deref_type_cast
)
393 nir_variable_mode parent_mode
;
394 if (deref
->deref_type
== nir_deref_type_var
) {
395 parent_mode
= deref
->var
->data
.mode
;
397 assert(deref
->parent
.is_ssa
);
398 nir_deref_instr
*parent
=
399 nir_instr_as_deref(deref
->parent
.ssa
->parent_instr
);
400 parent_mode
= parent
->mode
;
403 deref
->mode
= parent_mode
;
410 modes_may_alias(nir_variable_mode a
, nir_variable_mode b
)
412 /* Generic pointers can alias with SSBOs */
413 if ((a
== nir_var_mem_ssbo
|| a
== nir_var_mem_global
) &&
414 (b
== nir_var_mem_ssbo
|| b
== nir_var_mem_global
))
417 /* In the general case, pointers can only alias if they have the same mode.
419 * NOTE: In future, with things like OpenCL generic pointers, this may not
420 * be true and will have to be re-evaluated. However, with graphics only,
427 deref_path_contains_coherent_decoration(nir_deref_path
*path
)
429 assert(path
->path
[0]->deref_type
== nir_deref_type_var
);
431 if (path
->path
[0]->var
->data
.access
& ACCESS_COHERENT
)
434 for (nir_deref_instr
**p
= &path
->path
[1]; *p
; p
++) {
435 if ((*p
)->deref_type
!= nir_deref_type_struct
)
438 const struct glsl_type
*struct_type
= (*(p
- 1))->type
;
439 const struct glsl_struct_field
*field
=
440 glsl_get_struct_field_data(struct_type
, (*p
)->strct
.index
);
441 if (field
->memory_coherent
)
448 nir_deref_compare_result
449 nir_compare_deref_paths(nir_deref_path
*a_path
,
450 nir_deref_path
*b_path
)
452 if (!modes_may_alias(b_path
->path
[0]->mode
, a_path
->path
[0]->mode
))
453 return nir_derefs_do_not_alias
;
455 if (a_path
->path
[0]->deref_type
!= b_path
->path
[0]->deref_type
)
456 return nir_derefs_may_alias_bit
;
458 if (a_path
->path
[0]->deref_type
== nir_deref_type_var
) {
459 if (a_path
->path
[0]->var
!= b_path
->path
[0]->var
) {
460 /* Shader and function temporaries aren't backed by memory so two
461 * distinct variables never alias.
463 static const nir_variable_mode temp_var_modes
=
464 nir_var_shader_temp
| nir_var_function_temp
;
465 if ((a_path
->path
[0]->mode
& temp_var_modes
) ||
466 (b_path
->path
[0]->mode
& temp_var_modes
))
467 return nir_derefs_do_not_alias
;
469 /* If they are both declared coherent or have coherent somewhere in
470 * their path (due to a member of an interface being declared
471 * coherent), we have to assume we that we could have any kind of
472 * aliasing. Otherwise, they could still alias but the client didn't
473 * tell us and that's their fault.
475 if (deref_path_contains_coherent_decoration(a_path
) &&
476 deref_path_contains_coherent_decoration(b_path
))
477 return nir_derefs_may_alias_bit
;
479 /* If we can chase the deref all the way back to the variable and
480 * they're not the same variable and at least one is not declared
481 * coherent, we know they can't possibly alias.
483 return nir_derefs_do_not_alias
;
486 assert(a_path
->path
[0]->deref_type
== nir_deref_type_cast
);
487 /* If they're not exactly the same cast, it's hard to compare them so we
488 * just assume they alias. Comparing casts is tricky as there are lots
489 * of things such as mode, type, etc. to make sure work out; for now, we
490 * just assume nit_opt_deref will combine them and compare the deref
493 * TODO: At some point in the future, we could be clever and understand
494 * that a float[] and int[] have the same layout and aliasing structure
495 * but double[] and vec3[] do not and we could potentially be a bit
498 if (a_path
->path
[0] != b_path
->path
[0])
499 return nir_derefs_may_alias_bit
;
502 /* Start off assuming they fully compare. We ignore equality for now. In
503 * the end, we'll determine that by containment.
505 nir_deref_compare_result result
= nir_derefs_may_alias_bit
|
506 nir_derefs_a_contains_b_bit
|
507 nir_derefs_b_contains_a_bit
;
509 nir_deref_instr
**a_p
= &a_path
->path
[1];
510 nir_deref_instr
**b_p
= &b_path
->path
[1];
511 while (*a_p
!= NULL
&& *a_p
== *b_p
) {
516 /* We're at either the tail or the divergence point between the two deref
517 * paths. Look to see if either contains a ptr_as_array deref. It it
518 * does we don't know how to safely make any inferences. Hopefully,
519 * nir_opt_deref will clean most of these up and we can start inferring
522 * In theory, we could do a bit better. For instance, we could detect the
523 * case where we have exactly one ptr_as_array deref in the chain after the
524 * divergence point and it's matched in both chains and the two chains have
525 * different constant indices.
527 for (nir_deref_instr
**t_p
= a_p
; *t_p
; t_p
++) {
528 if ((*t_p
)->deref_type
== nir_deref_type_ptr_as_array
)
529 return nir_derefs_may_alias_bit
;
531 for (nir_deref_instr
**t_p
= b_p
; *t_p
; t_p
++) {
532 if ((*t_p
)->deref_type
== nir_deref_type_ptr_as_array
)
533 return nir_derefs_may_alias_bit
;
536 while (*a_p
!= NULL
&& *b_p
!= NULL
) {
537 nir_deref_instr
*a_tail
= *(a_p
++);
538 nir_deref_instr
*b_tail
= *(b_p
++);
540 switch (a_tail
->deref_type
) {
541 case nir_deref_type_array
:
542 case nir_deref_type_array_wildcard
: {
543 assert(b_tail
->deref_type
== nir_deref_type_array
||
544 b_tail
->deref_type
== nir_deref_type_array_wildcard
);
546 if (a_tail
->deref_type
== nir_deref_type_array_wildcard
) {
547 if (b_tail
->deref_type
!= nir_deref_type_array_wildcard
)
548 result
&= ~nir_derefs_b_contains_a_bit
;
549 } else if (b_tail
->deref_type
== nir_deref_type_array_wildcard
) {
550 if (a_tail
->deref_type
!= nir_deref_type_array_wildcard
)
551 result
&= ~nir_derefs_a_contains_b_bit
;
553 assert(a_tail
->deref_type
== nir_deref_type_array
&&
554 b_tail
->deref_type
== nir_deref_type_array
);
555 assert(a_tail
->arr
.index
.is_ssa
&& b_tail
->arr
.index
.is_ssa
);
557 if (nir_src_is_const(a_tail
->arr
.index
) &&
558 nir_src_is_const(b_tail
->arr
.index
)) {
559 /* If they're both direct and have different offsets, they
560 * don't even alias much less anything else.
562 if (nir_src_as_uint(a_tail
->arr
.index
) !=
563 nir_src_as_uint(b_tail
->arr
.index
))
564 return nir_derefs_do_not_alias
;
565 } else if (a_tail
->arr
.index
.ssa
== b_tail
->arr
.index
.ssa
) {
566 /* They're the same indirect, continue on */
568 /* They're not the same index so we can't prove anything about
571 result
&= ~(nir_derefs_a_contains_b_bit
| nir_derefs_b_contains_a_bit
);
577 case nir_deref_type_struct
: {
578 /* If they're different struct members, they don't even alias */
579 if (a_tail
->strct
.index
!= b_tail
->strct
.index
)
580 return nir_derefs_do_not_alias
;
585 unreachable("Invalid deref type");
589 /* If a is longer than b, then it can't contain b */
591 result
&= ~nir_derefs_a_contains_b_bit
;
593 result
&= ~nir_derefs_b_contains_a_bit
;
595 /* If a contains b and b contains a they must be equal. */
596 if ((result
& nir_derefs_a_contains_b_bit
) && (result
& nir_derefs_b_contains_a_bit
))
597 result
|= nir_derefs_equal_bit
;
602 nir_deref_compare_result
603 nir_compare_derefs(nir_deref_instr
*a
, nir_deref_instr
*b
)
606 return nir_derefs_equal_bit
| nir_derefs_may_alias_bit
|
607 nir_derefs_a_contains_b_bit
| nir_derefs_b_contains_a_bit
;
610 nir_deref_path a_path
, b_path
;
611 nir_deref_path_init(&a_path
, a
, NULL
);
612 nir_deref_path_init(&b_path
, b
, NULL
);
613 assert(a_path
.path
[0]->deref_type
== nir_deref_type_var
||
614 a_path
.path
[0]->deref_type
== nir_deref_type_cast
);
615 assert(b_path
.path
[0]->deref_type
== nir_deref_type_var
||
616 b_path
.path
[0]->deref_type
== nir_deref_type_cast
);
618 nir_deref_compare_result result
= nir_compare_deref_paths(&a_path
, &b_path
);
620 nir_deref_path_finish(&a_path
);
621 nir_deref_path_finish(&b_path
);
626 struct rematerialize_deref_state
{
630 struct hash_table
*cache
;
633 static nir_deref_instr
*
634 rematerialize_deref_in_block(nir_deref_instr
*deref
,
635 struct rematerialize_deref_state
*state
)
637 if (deref
->instr
.block
== state
->block
)
641 state
->cache
= _mesa_pointer_hash_table_create(NULL
);
644 struct hash_entry
*cached
= _mesa_hash_table_search(state
->cache
, deref
);
648 nir_builder
*b
= &state
->builder
;
649 nir_deref_instr
*new_deref
=
650 nir_deref_instr_create(b
->shader
, deref
->deref_type
);
651 new_deref
->mode
= deref
->mode
;
652 new_deref
->type
= deref
->type
;
654 if (deref
->deref_type
== nir_deref_type_var
) {
655 new_deref
->var
= deref
->var
;
657 nir_deref_instr
*parent
= nir_src_as_deref(deref
->parent
);
659 parent
= rematerialize_deref_in_block(parent
, state
);
660 new_deref
->parent
= nir_src_for_ssa(&parent
->dest
.ssa
);
662 nir_src_copy(&new_deref
->parent
, &deref
->parent
, new_deref
);
666 switch (deref
->deref_type
) {
667 case nir_deref_type_var
:
668 case nir_deref_type_array_wildcard
:
669 /* Nothing more to do */
672 case nir_deref_type_cast
:
673 new_deref
->cast
.ptr_stride
= deref
->cast
.ptr_stride
;
676 case nir_deref_type_array
:
677 case nir_deref_type_ptr_as_array
:
678 assert(!nir_src_as_deref(deref
->arr
.index
));
679 nir_src_copy(&new_deref
->arr
.index
, &deref
->arr
.index
, new_deref
);
682 case nir_deref_type_struct
:
683 new_deref
->strct
.index
= deref
->strct
.index
;
687 unreachable("Invalid deref instruction type");
690 nir_ssa_dest_init(&new_deref
->instr
, &new_deref
->dest
,
691 deref
->dest
.ssa
.num_components
,
692 deref
->dest
.ssa
.bit_size
,
693 deref
->dest
.ssa
.name
);
694 nir_builder_instr_insert(b
, &new_deref
->instr
);
700 rematerialize_deref_src(nir_src
*src
, void *_state
)
702 struct rematerialize_deref_state
*state
= _state
;
704 nir_deref_instr
*deref
= nir_src_as_deref(*src
);
708 nir_deref_instr
*block_deref
= rematerialize_deref_in_block(deref
, state
);
709 if (block_deref
!= deref
) {
710 nir_instr_rewrite_src(src
->parent_instr
, src
,
711 nir_src_for_ssa(&block_deref
->dest
.ssa
));
712 nir_deref_instr_remove_if_unused(deref
);
713 state
->progress
= true;
719 /** Re-materialize derefs in every block
721 * This pass re-materializes deref instructions in every block in which it is
722 * used. After this pass has been run, every use of a deref will be of a
723 * deref in the same block as the use. Also, all unused derefs will be
724 * deleted as a side-effect.
726 * Derefs used as sources of phi instructions are not rematerialized.
729 nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
)
731 struct rematerialize_deref_state state
= { 0 };
732 nir_builder_init(&state
.builder
, impl
);
734 nir_foreach_block(block
, impl
) {
737 /* Start each block with a fresh cache */
739 _mesa_hash_table_clear(state
.cache
, NULL
);
741 nir_foreach_instr_safe(instr
, block
) {
742 if (instr
->type
== nir_instr_type_deref
&&
743 nir_deref_instr_remove_if_unused(nir_instr_as_deref(instr
)))
746 /* If a deref is used in a phi, we can't rematerialize it, as the new
747 * derefs would appear before the phi, which is not valid.
749 if (instr
->type
== nir_instr_type_phi
)
752 state
.builder
.cursor
= nir_before_instr(instr
);
753 nir_foreach_src(instr
, rematerialize_deref_src
, &state
);
757 nir_if
*following_if
= nir_block_get_following_if(block
);
759 assert(!nir_src_as_deref(following_if
->condition
));
763 _mesa_hash_table_destroy(state
.cache
, NULL
);
765 return state
.progress
;
769 nir_deref_instr_fixup_child_types(nir_deref_instr
*parent
)
771 nir_foreach_use(use
, &parent
->dest
.ssa
) {
772 if (use
->parent_instr
->type
!= nir_instr_type_deref
)
775 nir_deref_instr
*child
= nir_instr_as_deref(use
->parent_instr
);
776 switch (child
->deref_type
) {
777 case nir_deref_type_var
:
778 unreachable("nir_deref_type_var cannot be a child");
780 case nir_deref_type_array
:
781 case nir_deref_type_array_wildcard
:
782 child
->type
= glsl_get_array_element(parent
->type
);
785 case nir_deref_type_ptr_as_array
:
786 child
->type
= parent
->type
;
789 case nir_deref_type_struct
:
790 child
->type
= glsl_get_struct_field(parent
->type
,
794 case nir_deref_type_cast
:
795 /* We stop the recursion here */
799 /* Recurse into children */
800 nir_deref_instr_fixup_child_types(child
);
805 is_trivial_array_deref_cast(nir_deref_instr
*cast
)
807 assert(is_trivial_deref_cast(cast
));
809 nir_deref_instr
*parent
= nir_src_as_deref(cast
->parent
);
811 if (parent
->deref_type
== nir_deref_type_array
) {
812 return cast
->cast
.ptr_stride
==
813 glsl_get_explicit_stride(nir_deref_instr_parent(parent
)->type
);
814 } else if (parent
->deref_type
== nir_deref_type_ptr_as_array
) {
815 return cast
->cast
.ptr_stride
==
816 nir_deref_instr_ptr_as_array_stride(parent
);
823 is_deref_ptr_as_array(nir_instr
*instr
)
825 return instr
->type
== nir_instr_type_deref
&&
826 nir_instr_as_deref(instr
)->deref_type
== nir_deref_type_ptr_as_array
;
830 * Remove casts that just wrap other casts.
833 opt_remove_cast_cast(nir_deref_instr
*cast
)
835 nir_deref_instr
*first_cast
= cast
;
838 nir_deref_instr
*parent
= nir_deref_instr_parent(first_cast
);
839 if (parent
== NULL
|| parent
->deref_type
!= nir_deref_type_cast
)
843 if (cast
== first_cast
)
846 nir_instr_rewrite_src(&cast
->instr
, &cast
->parent
,
847 nir_src_for_ssa(first_cast
->parent
.ssa
));
852 opt_remove_sampler_cast(nir_deref_instr
*cast
)
854 assert(cast
->deref_type
== nir_deref_type_cast
);
855 nir_deref_instr
*parent
= nir_src_as_deref(cast
->parent
);
859 /* Strip both types down to their non-array type and bail if there are any
860 * discrepancies in array lengths.
862 const struct glsl_type
*parent_type
= parent
->type
;
863 const struct glsl_type
*cast_type
= cast
->type
;
864 while (glsl_type_is_array(parent_type
) && glsl_type_is_array(cast_type
)) {
865 if (glsl_get_length(parent_type
) != glsl_get_length(cast_type
))
867 parent_type
= glsl_get_array_element(parent_type
);
868 cast_type
= glsl_get_array_element(cast_type
);
871 if (glsl_type_is_array(parent_type
) || glsl_type_is_array(cast_type
))
874 if (!glsl_type_is_sampler(parent_type
) ||
875 cast_type
!= glsl_bare_sampler_type())
878 /* We're a cast from a more detailed sampler type to a bare sampler */
879 nir_ssa_def_rewrite_uses(&cast
->dest
.ssa
,
880 nir_src_for_ssa(&parent
->dest
.ssa
));
881 nir_instr_remove(&cast
->instr
);
883 /* Recursively crawl the deref tree and clean up types */
884 nir_deref_instr_fixup_child_types(parent
);
890 * Is this casting a struct to a contained struct.
891 * struct a { struct b field0 };
893 * deref_cast (structb *)ssa_5 (function_temp structb);
895 * deref_struct &ssa_5->field0 (function_temp structb);
896 * This allows subsequent copy propagation to work.
899 opt_replace_struct_wrapper_cast(nir_builder
*b
, nir_deref_instr
*cast
)
901 nir_deref_instr
*parent
= nir_src_as_deref(cast
->parent
);
905 if (!glsl_type_is_struct(parent
->type
))
908 if (glsl_get_struct_field_offset(parent
->type
, 0) != 0)
911 if (cast
->type
!= glsl_get_struct_field(parent
->type
, 0))
914 nir_deref_instr
*replace
= nir_build_deref_struct(b
, parent
, 0);
915 nir_ssa_def_rewrite_uses(&cast
->dest
.ssa
, nir_src_for_ssa(&replace
->dest
.ssa
));
916 nir_deref_instr_remove_if_unused(cast
);
921 opt_deref_cast(nir_builder
*b
, nir_deref_instr
*cast
)
925 if (opt_replace_struct_wrapper_cast(b
, cast
))
928 if (opt_remove_sampler_cast(cast
))
931 progress
= opt_remove_cast_cast(cast
);
932 if (!is_trivial_deref_cast(cast
))
935 bool trivial_array_cast
= is_trivial_array_deref_cast(cast
);
937 assert(cast
->dest
.is_ssa
);
938 assert(cast
->parent
.is_ssa
);
940 nir_foreach_use_safe(use_src
, &cast
->dest
.ssa
) {
941 /* If this isn't a trivial array cast, we can't propagate into
942 * ptr_as_array derefs.
944 if (is_deref_ptr_as_array(use_src
->parent_instr
) &&
948 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, cast
->parent
);
952 /* If uses would be a bit crazy */
953 assert(list_is_empty(&cast
->dest
.ssa
.if_uses
));
955 if (nir_deref_instr_remove_if_unused(cast
))
962 opt_deref_ptr_as_array(nir_builder
*b
, nir_deref_instr
*deref
)
964 assert(deref
->deref_type
== nir_deref_type_ptr_as_array
);
966 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
968 if (nir_src_is_const(deref
->arr
.index
) &&
969 nir_src_as_int(deref
->arr
.index
) == 0) {
970 /* If it's a ptr_as_array deref with an index of 0, it does nothing
971 * and we can just replace its uses with its parent.
973 * The source of a ptr_as_array deref always has a deref_type of
974 * nir_deref_type_array or nir_deref_type_cast. If it's a cast, it
975 * may be trivial and we may be able to get rid of that too. Any
976 * trivial cast of trivial cast cases should be handled already by
977 * opt_deref_cast() above.
979 if (parent
->deref_type
== nir_deref_type_cast
&&
980 is_trivial_deref_cast(parent
))
981 parent
= nir_deref_instr_parent(parent
);
982 nir_ssa_def_rewrite_uses(&deref
->dest
.ssa
,
983 nir_src_for_ssa(&parent
->dest
.ssa
));
984 nir_instr_remove(&deref
->instr
);
988 if (parent
->deref_type
!= nir_deref_type_array
&&
989 parent
->deref_type
!= nir_deref_type_ptr_as_array
)
992 assert(parent
->parent
.is_ssa
);
993 assert(parent
->arr
.index
.is_ssa
);
994 assert(deref
->arr
.index
.is_ssa
);
996 nir_ssa_def
*new_idx
= nir_iadd(b
, parent
->arr
.index
.ssa
,
997 deref
->arr
.index
.ssa
);
999 deref
->deref_type
= parent
->deref_type
;
1000 nir_instr_rewrite_src(&deref
->instr
, &deref
->parent
, parent
->parent
);
1001 nir_instr_rewrite_src(&deref
->instr
, &deref
->arr
.index
,
1002 nir_src_for_ssa(new_idx
));
1007 nir_opt_deref_impl(nir_function_impl
*impl
)
1009 bool progress
= false;
1012 nir_builder_init(&b
, impl
);
1014 nir_foreach_block(block
, impl
) {
1015 nir_foreach_instr_safe(instr
, block
) {
1016 if (instr
->type
!= nir_instr_type_deref
)
1019 b
.cursor
= nir_before_instr(instr
);
1021 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
1022 switch (deref
->deref_type
) {
1023 case nir_deref_type_ptr_as_array
:
1024 if (opt_deref_ptr_as_array(&b
, deref
))
1028 case nir_deref_type_cast
:
1029 if (opt_deref_cast(&b
, deref
))
1041 nir_metadata_preserve(impl
, nir_metadata_block_index
|
1042 nir_metadata_dominance
);
1044 nir_metadata_preserve(impl
, nir_metadata_all
);
1051 nir_opt_deref(nir_shader
*shader
)
1053 bool progress
= false;
1055 nir_foreach_function(func
, shader
) {
1056 if (func
->impl
&& nir_opt_deref_impl(func
->impl
))