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"
29 #include "util/u_math.h"
32 struct split_var_state
{
36 nir_function_impl
*impl
;
38 nir_variable
*base_var
;
44 const struct glsl_type
*type
;
52 static const struct glsl_type
*
53 wrap_type_in_array(const struct glsl_type
*type
,
54 const struct glsl_type
*array_type
)
56 if (!glsl_type_is_array(array_type
))
59 const struct glsl_type
*elem_type
=
60 wrap_type_in_array(type
, glsl_get_array_element(array_type
));
61 assert(glsl_get_explicit_stride(array_type
) == 0);
62 return glsl_array_type(elem_type
, glsl_get_length(array_type
), 0);
66 num_array_levels_in_array_of_vector_type(const struct glsl_type
*type
)
70 if (glsl_type_is_array_or_matrix(type
)) {
72 type
= glsl_get_array_element(type
);
73 } else if (glsl_type_is_vector_or_scalar(type
)) {
76 /* Not an array of vectors */
83 init_field_for_type(struct field
*field
, struct field
*parent
,
84 const struct glsl_type
*type
,
86 struct split_var_state
*state
)
88 *field
= (struct field
) {
93 const struct glsl_type
*struct_type
= glsl_without_array(type
);
94 if (glsl_type_is_struct(struct_type
)) {
95 field
->num_fields
= glsl_get_length(struct_type
),
96 field
->fields
= ralloc_array(state
->mem_ctx
, struct field
,
98 for (unsigned i
= 0; i
< field
->num_fields
; i
++) {
99 char *field_name
= NULL
;
101 field_name
= ralloc_asprintf(state
->mem_ctx
, "%s_%s", name
,
102 glsl_get_struct_elem_name(struct_type
, i
));
104 field_name
= ralloc_asprintf(state
->mem_ctx
, "{unnamed %s}_%s",
105 glsl_get_type_name(struct_type
),
106 glsl_get_struct_elem_name(struct_type
, i
));
108 init_field_for_type(&field
->fields
[i
], field
,
109 glsl_get_struct_field(struct_type
, i
),
113 const struct glsl_type
*var_type
= type
;
114 for (struct field
*f
= field
->parent
; f
; f
= f
->parent
)
115 var_type
= wrap_type_in_array(var_type
, f
->type
);
117 nir_variable_mode mode
= state
->base_var
->data
.mode
;
118 if (mode
== nir_var_function
) {
119 field
->var
= nir_local_variable_create(state
->impl
, var_type
, name
);
121 field
->var
= nir_variable_create(state
->shader
, mode
, var_type
, name
);
127 split_var_list_structs(nir_shader
*shader
,
128 nir_function_impl
*impl
,
129 struct exec_list
*vars
,
130 struct hash_table
*var_field_map
,
133 struct split_var_state state
= {
139 struct exec_list split_vars
;
140 exec_list_make_empty(&split_vars
);
142 /* To avoid list confusion (we'll be adding things as we split variables),
143 * pull all of the variables we plan to split off of the list
145 nir_foreach_variable_safe(var
, vars
) {
146 if (!glsl_type_is_struct(glsl_without_array(var
->type
)))
149 exec_node_remove(&var
->node
);
150 exec_list_push_tail(&split_vars
, &var
->node
);
153 nir_foreach_variable(var
, &split_vars
) {
154 state
.base_var
= var
;
156 struct field
*root_field
= ralloc(mem_ctx
, struct field
);
157 init_field_for_type(root_field
, NULL
, var
->type
, var
->name
, &state
);
158 _mesa_hash_table_insert(var_field_map
, var
, root_field
);
161 return !exec_list_is_empty(&split_vars
);
165 split_struct_derefs_impl(nir_function_impl
*impl
,
166 struct hash_table
*var_field_map
,
167 nir_variable_mode modes
,
171 nir_builder_init(&b
, impl
);
173 nir_foreach_block(block
, impl
) {
174 nir_foreach_instr_safe(instr
, block
) {
175 if (instr
->type
!= nir_instr_type_deref
)
178 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
179 if (!(deref
->mode
& modes
))
182 /* Clean up any dead derefs we find lying around. They may refer to
183 * variables we're planning to split.
185 if (nir_deref_instr_remove_if_unused(deref
))
188 if (!glsl_type_is_vector_or_scalar(deref
->type
))
191 nir_variable
*base_var
= nir_deref_instr_get_variable(deref
);
192 struct hash_entry
*entry
=
193 _mesa_hash_table_search(var_field_map
, base_var
);
197 struct field
*root_field
= entry
->data
;
200 nir_deref_path_init(&path
, deref
, mem_ctx
);
202 struct field
*tail_field
= root_field
;
203 for (unsigned i
= 0; path
.path
[i
]; i
++) {
204 if (path
.path
[i
]->deref_type
!= nir_deref_type_struct
)
208 assert(glsl_type_is_struct(path
.path
[i
- 1]->type
));
209 assert(path
.path
[i
- 1]->type
==
210 glsl_without_array(tail_field
->type
));
212 tail_field
= &tail_field
->fields
[path
.path
[i
]->strct
.index
];
214 nir_variable
*split_var
= tail_field
->var
;
216 nir_deref_instr
*new_deref
= NULL
;
217 for (unsigned i
= 0; path
.path
[i
]; i
++) {
218 nir_deref_instr
*p
= path
.path
[i
];
219 b
.cursor
= nir_after_instr(&p
->instr
);
221 switch (p
->deref_type
) {
222 case nir_deref_type_var
:
223 assert(new_deref
== NULL
);
224 new_deref
= nir_build_deref_var(&b
, split_var
);
227 case nir_deref_type_array
:
228 case nir_deref_type_array_wildcard
:
229 new_deref
= nir_build_deref_follower(&b
, new_deref
, p
);
232 case nir_deref_type_struct
:
233 /* Nothing to do; we're splitting structs */
237 unreachable("Invalid deref type in path");
241 assert(new_deref
->type
== deref
->type
);
242 nir_ssa_def_rewrite_uses(&deref
->dest
.ssa
,
243 nir_src_for_ssa(&new_deref
->dest
.ssa
));
244 nir_deref_instr_remove_if_unused(deref
);
249 /** A pass for splitting structs into multiple variables
251 * This pass splits arrays of structs into multiple variables, one for each
252 * (possibly nested) structure member. After this pass completes, no
253 * variables of the given mode will contain a struct type.
256 nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
)
258 void *mem_ctx
= ralloc_context(NULL
);
259 struct hash_table
*var_field_map
=
260 _mesa_hash_table_create(mem_ctx
, _mesa_hash_pointer
,
261 _mesa_key_pointer_equal
);
263 assert((modes
& (nir_var_private
| nir_var_function
)) == modes
);
265 bool has_global_splits
= false;
266 if (modes
& nir_var_private
) {
267 has_global_splits
= split_var_list_structs(shader
, NULL
,
269 var_field_map
, mem_ctx
);
272 bool progress
= false;
273 nir_foreach_function(function
, shader
) {
277 bool has_local_splits
= false;
278 if (modes
& nir_var_function
) {
279 has_local_splits
= split_var_list_structs(shader
, function
->impl
,
280 &function
->impl
->locals
,
281 var_field_map
, mem_ctx
);
284 if (has_global_splits
|| has_local_splits
) {
285 split_struct_derefs_impl(function
->impl
, var_field_map
,
288 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
289 nir_metadata_dominance
);
294 ralloc_free(mem_ctx
);
299 struct array_level_info
{
305 /* Only set if this is the tail end of the splitting */
309 struct array_split
*splits
;
312 struct array_var_info
{
313 nir_variable
*base_var
;
315 const struct glsl_type
*split_var_type
;
318 struct array_split root_split
;
321 struct array_level_info levels
[0];
325 init_var_list_array_infos(struct exec_list
*vars
,
326 struct hash_table
*var_info_map
,
329 bool has_array
= false;
331 nir_foreach_variable(var
, vars
) {
332 int num_levels
= num_array_levels_in_array_of_vector_type(var
->type
);
336 struct array_var_info
*info
=
337 rzalloc_size(mem_ctx
, sizeof(*info
) +
338 num_levels
* sizeof(info
->levels
[0]));
340 info
->base_var
= var
;
341 info
->num_levels
= num_levels
;
343 const struct glsl_type
*type
= var
->type
;
344 for (int i
= 0; i
< num_levels
; i
++) {
345 assert(glsl_get_explicit_stride(type
) == 0);
346 info
->levels
[i
].array_len
= glsl_get_length(type
);
347 type
= glsl_get_array_element(type
);
349 /* All levels start out initially as split */
350 info
->levels
[i
].split
= true;
353 _mesa_hash_table_insert(var_info_map
, var
, info
);
360 static struct array_var_info
*
361 get_array_var_info(nir_variable
*var
,
362 struct hash_table
*var_info_map
)
364 struct hash_entry
*entry
=
365 _mesa_hash_table_search(var_info_map
, var
);
366 return entry
? entry
->data
: NULL
;
369 static struct array_var_info
*
370 get_array_deref_info(nir_deref_instr
*deref
,
371 struct hash_table
*var_info_map
,
372 nir_variable_mode modes
)
374 if (!(deref
->mode
& modes
))
377 return get_array_var_info(nir_deref_instr_get_variable(deref
),
382 mark_array_deref_used(nir_deref_instr
*deref
,
383 struct hash_table
*var_info_map
,
384 nir_variable_mode modes
,
387 struct array_var_info
*info
=
388 get_array_deref_info(deref
, var_info_map
, modes
);
393 nir_deref_path_init(&path
, deref
, mem_ctx
);
395 /* Walk the path and look for indirects. If we have an array deref with an
396 * indirect, mark the given level as not being split.
398 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
399 nir_deref_instr
*p
= path
.path
[i
+ 1];
400 if (p
->deref_type
== nir_deref_type_array
&&
401 !nir_src_is_const(p
->arr
.index
))
402 info
->levels
[i
].split
= false;
407 mark_array_usage_impl(nir_function_impl
*impl
,
408 struct hash_table
*var_info_map
,
409 nir_variable_mode modes
,
412 nir_foreach_block(block
, impl
) {
413 nir_foreach_instr(instr
, block
) {
414 if (instr
->type
!= nir_instr_type_intrinsic
)
417 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
418 switch (intrin
->intrinsic
) {
419 case nir_intrinsic_copy_deref
:
420 mark_array_deref_used(nir_src_as_deref(intrin
->src
[1]),
421 var_info_map
, modes
, mem_ctx
);
424 case nir_intrinsic_load_deref
:
425 case nir_intrinsic_store_deref
:
426 mark_array_deref_used(nir_src_as_deref(intrin
->src
[0]),
427 var_info_map
, modes
, mem_ctx
);
438 create_split_array_vars(struct array_var_info
*var_info
,
440 struct array_split
*split
,
443 nir_function_impl
*impl
,
446 while (level
< var_info
->num_levels
&& !var_info
->levels
[level
].split
) {
447 name
= ralloc_asprintf(mem_ctx
, "%s[*]", name
);
451 if (level
== var_info
->num_levels
) {
452 /* We add parens to the variable name so it looks like "(foo[2][*])" so
453 * that further derefs will look like "(foo[2][*])[ssa_6]"
455 name
= ralloc_asprintf(mem_ctx
, "(%s)", name
);
457 nir_variable_mode mode
= var_info
->base_var
->data
.mode
;
458 if (mode
== nir_var_function
) {
459 split
->var
= nir_local_variable_create(impl
,
460 var_info
->split_var_type
, name
);
462 split
->var
= nir_variable_create(shader
, mode
,
463 var_info
->split_var_type
, name
);
466 assert(var_info
->levels
[level
].split
);
467 split
->num_splits
= var_info
->levels
[level
].array_len
;
468 split
->splits
= rzalloc_array(mem_ctx
, struct array_split
,
470 for (unsigned i
= 0; i
< split
->num_splits
; i
++) {
471 create_split_array_vars(var_info
, level
+ 1, &split
->splits
[i
],
472 ralloc_asprintf(mem_ctx
, "%s[%d]", name
, i
),
473 shader
, impl
, mem_ctx
);
479 split_var_list_arrays(nir_shader
*shader
,
480 nir_function_impl
*impl
,
481 struct exec_list
*vars
,
482 struct hash_table
*var_info_map
,
485 struct exec_list split_vars
;
486 exec_list_make_empty(&split_vars
);
488 nir_foreach_variable_safe(var
, vars
) {
489 struct array_var_info
*info
= get_array_var_info(var
, var_info_map
);
493 bool has_split
= false;
494 const struct glsl_type
*split_type
=
495 glsl_without_array_or_matrix(var
->type
);
496 for (int i
= info
->num_levels
- 1; i
>= 0; i
--) {
497 if (info
->levels
[i
].split
) {
502 /* If the original type was a matrix type, we'd like to keep that so
503 * we don't convert matrices into arrays.
505 if (i
== info
->num_levels
- 1 &&
506 glsl_type_is_matrix(glsl_without_array(var
->type
))) {
507 split_type
= glsl_matrix_type(glsl_get_base_type(split_type
),
508 glsl_get_components(split_type
),
509 info
->levels
[i
].array_len
);
511 split_type
= glsl_array_type(split_type
, info
->levels
[i
].array_len
, 0);
516 info
->split_var_type
= split_type
;
517 /* To avoid list confusion (we'll be adding things as we split
518 * variables), pull all of the variables we plan to split off of the
519 * main variable list.
521 exec_node_remove(&var
->node
);
522 exec_list_push_tail(&split_vars
, &var
->node
);
524 assert(split_type
== var
->type
);
525 /* If we're not modifying this variable, delete the info so we skip
526 * it faster in later passes.
528 _mesa_hash_table_remove_key(var_info_map
, var
);
532 nir_foreach_variable(var
, &split_vars
) {
533 struct array_var_info
*info
= get_array_var_info(var
, var_info_map
);
534 create_split_array_vars(info
, 0, &info
->root_split
, var
->name
,
535 shader
, impl
, mem_ctx
);
538 return !exec_list_is_empty(&split_vars
);
542 deref_has_split_wildcard(nir_deref_path
*path
,
543 struct array_var_info
*info
)
548 assert(path
->path
[0]->var
== info
->base_var
);
549 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
550 if (path
->path
[i
+ 1]->deref_type
== nir_deref_type_array_wildcard
&&
551 info
->levels
[i
].split
)
559 array_path_is_out_of_bounds(nir_deref_path
*path
,
560 struct array_var_info
*info
)
565 assert(path
->path
[0]->var
== info
->base_var
);
566 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
567 nir_deref_instr
*p
= path
->path
[i
+ 1];
568 if (p
->deref_type
== nir_deref_type_array_wildcard
)
571 if (nir_src_is_const(p
->arr
.index
) &&
572 nir_src_as_uint(p
->arr
.index
) >= info
->levels
[i
].array_len
)
580 emit_split_copies(nir_builder
*b
,
581 struct array_var_info
*dst_info
, nir_deref_path
*dst_path
,
582 unsigned dst_level
, nir_deref_instr
*dst
,
583 struct array_var_info
*src_info
, nir_deref_path
*src_path
,
584 unsigned src_level
, nir_deref_instr
*src
)
586 nir_deref_instr
*dst_p
, *src_p
;
588 while ((dst_p
= dst_path
->path
[dst_level
+ 1])) {
589 if (dst_p
->deref_type
== nir_deref_type_array_wildcard
)
592 dst
= nir_build_deref_follower(b
, dst
, dst_p
);
596 while ((src_p
= src_path
->path
[src_level
+ 1])) {
597 if (src_p
->deref_type
== nir_deref_type_array_wildcard
)
600 src
= nir_build_deref_follower(b
, src
, src_p
);
604 if (src_p
== NULL
|| dst_p
== NULL
) {
605 assert(src_p
== NULL
&& dst_p
== NULL
);
606 nir_copy_deref(b
, dst
, src
);
608 assert(dst_p
->deref_type
== nir_deref_type_array_wildcard
&&
609 src_p
->deref_type
== nir_deref_type_array_wildcard
);
611 if ((dst_info
&& dst_info
->levels
[dst_level
].split
) ||
612 (src_info
&& src_info
->levels
[src_level
].split
)) {
613 /* There are no indirects at this level on one of the source or the
614 * destination so we are lowering it.
616 assert(glsl_get_length(dst_path
->path
[dst_level
]->type
) ==
617 glsl_get_length(src_path
->path
[src_level
]->type
));
618 unsigned len
= glsl_get_length(dst_path
->path
[dst_level
]->type
);
619 for (unsigned i
= 0; i
< len
; i
++) {
620 nir_ssa_def
*idx
= nir_imm_int(b
, i
);
621 emit_split_copies(b
, dst_info
, dst_path
, dst_level
+ 1,
622 nir_build_deref_array(b
, dst
, idx
),
623 src_info
, src_path
, src_level
+ 1,
624 nir_build_deref_array(b
, src
, idx
));
627 /* Neither side is being split so we just keep going */
628 emit_split_copies(b
, dst_info
, dst_path
, dst_level
+ 1,
629 nir_build_deref_array_wildcard(b
, dst
),
630 src_info
, src_path
, src_level
+ 1,
631 nir_build_deref_array_wildcard(b
, src
));
637 split_array_copies_impl(nir_function_impl
*impl
,
638 struct hash_table
*var_info_map
,
639 nir_variable_mode modes
,
643 nir_builder_init(&b
, impl
);
645 nir_foreach_block(block
, impl
) {
646 nir_foreach_instr_safe(instr
, block
) {
647 if (instr
->type
!= nir_instr_type_intrinsic
)
650 nir_intrinsic_instr
*copy
= nir_instr_as_intrinsic(instr
);
651 if (copy
->intrinsic
!= nir_intrinsic_copy_deref
)
654 nir_deref_instr
*dst_deref
= nir_src_as_deref(copy
->src
[0]);
655 nir_deref_instr
*src_deref
= nir_src_as_deref(copy
->src
[1]);
657 struct array_var_info
*dst_info
=
658 get_array_deref_info(dst_deref
, var_info_map
, modes
);
659 struct array_var_info
*src_info
=
660 get_array_deref_info(src_deref
, var_info_map
, modes
);
662 if (!src_info
&& !dst_info
)
665 nir_deref_path dst_path
, src_path
;
666 nir_deref_path_init(&dst_path
, dst_deref
, mem_ctx
);
667 nir_deref_path_init(&src_path
, src_deref
, mem_ctx
);
669 if (!deref_has_split_wildcard(&dst_path
, dst_info
) &&
670 !deref_has_split_wildcard(&src_path
, src_info
))
673 b
.cursor
= nir_instr_remove(©
->instr
);
675 emit_split_copies(&b
, dst_info
, &dst_path
, 0, dst_path
.path
[0],
676 src_info
, &src_path
, 0, src_path
.path
[0]);
682 split_array_access_impl(nir_function_impl
*impl
,
683 struct hash_table
*var_info_map
,
684 nir_variable_mode modes
,
688 nir_builder_init(&b
, impl
);
690 nir_foreach_block(block
, impl
) {
691 nir_foreach_instr_safe(instr
, block
) {
692 if (instr
->type
== nir_instr_type_deref
) {
693 /* Clean up any dead derefs we find lying around. They may refer
694 * to variables we're planning to split.
696 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
697 if (deref
->mode
& modes
)
698 nir_deref_instr_remove_if_unused(deref
);
702 if (instr
->type
!= nir_instr_type_intrinsic
)
705 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
706 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
&&
707 intrin
->intrinsic
!= nir_intrinsic_store_deref
&&
708 intrin
->intrinsic
!= nir_intrinsic_copy_deref
)
711 const unsigned num_derefs
=
712 intrin
->intrinsic
== nir_intrinsic_copy_deref
? 2 : 1;
714 for (unsigned d
= 0; d
< num_derefs
; d
++) {
715 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[d
]);
717 struct array_var_info
*info
=
718 get_array_deref_info(deref
, var_info_map
, modes
);
723 nir_deref_path_init(&path
, deref
, mem_ctx
);
725 b
.cursor
= nir_before_instr(&intrin
->instr
);
727 if (array_path_is_out_of_bounds(&path
, info
)) {
728 /* If one of the derefs is out-of-bounds, we just delete the
729 * instruction. If a destination is out of bounds, then it may
730 * have been in-bounds prior to shrinking so we don't want to
731 * accidentally stomp something. However, we've already proven
732 * that it will never be read so it's safe to delete. If a
733 * source is out of bounds then it is loading random garbage.
734 * For loads, we replace their uses with an undef instruction
735 * and for copies we just delete the copy since it was writing
736 * undefined garbage anyway and we may as well leave the random
737 * garbage in the destination alone.
739 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
741 nir_ssa_undef(&b
, intrin
->dest
.ssa
.num_components
,
742 intrin
->dest
.ssa
.bit_size
);
743 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
746 nir_instr_remove(&intrin
->instr
);
747 for (unsigned i
= 0; i
< num_derefs
; i
++)
748 nir_deref_instr_remove_if_unused(nir_src_as_deref(intrin
->src
[i
]));
752 struct array_split
*split
= &info
->root_split
;
753 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
754 if (info
->levels
[i
].split
) {
755 nir_deref_instr
*p
= path
.path
[i
+ 1];
756 unsigned index
= nir_src_as_uint(p
->arr
.index
);
757 assert(index
< info
->levels
[i
].array_len
);
758 split
= &split
->splits
[index
];
761 assert(!split
->splits
&& split
->var
);
763 nir_deref_instr
*new_deref
= nir_build_deref_var(&b
, split
->var
);
764 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
765 if (!info
->levels
[i
].split
) {
766 new_deref
= nir_build_deref_follower(&b
, new_deref
,
770 assert(new_deref
->type
== deref
->type
);
772 /* Rewrite the deref source to point to the split one */
773 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[d
],
774 nir_src_for_ssa(&new_deref
->dest
.ssa
));
775 nir_deref_instr_remove_if_unused(deref
);
781 /** A pass for splitting arrays of vectors into multiple variables
783 * This pass looks at arrays (possibly multiple levels) of vectors (not
784 * structures or other types) and tries to split them into piles of variables,
785 * one for each array element. The heuristic used is simple: If a given array
786 * level is never used with an indirect, that array level will get split.
788 * This pass probably could handles structures easily enough but making a pass
789 * that could see through an array of structures of arrays would be difficult
790 * so it's best to just run nir_split_struct_vars first.
793 nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
)
795 void *mem_ctx
= ralloc_context(NULL
);
796 struct hash_table
*var_info_map
=
797 _mesa_hash_table_create(mem_ctx
, _mesa_hash_pointer
,
798 _mesa_key_pointer_equal
);
800 assert((modes
& (nir_var_private
| nir_var_function
)) == modes
);
802 bool has_global_array
= false;
803 if (modes
& nir_var_private
) {
804 has_global_array
= init_var_list_array_infos(&shader
->globals
,
805 var_info_map
, mem_ctx
);
808 bool has_any_array
= false;
809 nir_foreach_function(function
, shader
) {
813 bool has_local_array
= false;
814 if (modes
& nir_var_function
) {
815 has_local_array
= init_var_list_array_infos(&function
->impl
->locals
,
816 var_info_map
, mem_ctx
);
819 if (has_global_array
|| has_local_array
) {
820 has_any_array
= true;
821 mark_array_usage_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
825 /* If we failed to find any arrays of arrays, bail early. */
826 if (!has_any_array
) {
827 ralloc_free(mem_ctx
);
831 bool has_global_splits
= false;
832 if (modes
& nir_var_private
) {
833 has_global_splits
= split_var_list_arrays(shader
, NULL
,
835 var_info_map
, mem_ctx
);
838 bool progress
= false;
839 nir_foreach_function(function
, shader
) {
843 bool has_local_splits
= false;
844 if (modes
& nir_var_function
) {
845 has_local_splits
= split_var_list_arrays(shader
, function
->impl
,
846 &function
->impl
->locals
,
847 var_info_map
, mem_ctx
);
850 if (has_global_splits
|| has_local_splits
) {
851 split_array_copies_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
852 split_array_access_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
854 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
855 nir_metadata_dominance
);
860 ralloc_free(mem_ctx
);
865 struct array_level_usage
{
868 /* The value UINT_MAX will be used to indicate an indirect */
870 unsigned max_written
;
872 /* True if there is a copy that isn't to/from a shrinkable array */
873 bool has_external_copy
;
874 struct set
*levels_copied
;
877 struct vec_var_usage
{
878 /* Convenience set of all components this variable has */
879 nir_component_mask_t all_comps
;
881 nir_component_mask_t comps_read
;
882 nir_component_mask_t comps_written
;
884 nir_component_mask_t comps_kept
;
886 /* True if there is a copy that isn't to/from a shrinkable vector */
887 bool has_external_copy
;
888 struct set
*vars_copied
;
891 struct array_level_usage levels
[0];
894 static struct vec_var_usage
*
895 get_vec_var_usage(nir_variable
*var
,
896 struct hash_table
*var_usage_map
,
897 bool add_usage_entry
, void *mem_ctx
)
899 struct hash_entry
*entry
= _mesa_hash_table_search(var_usage_map
, var
);
903 if (!add_usage_entry
)
906 /* Check to make sure that we are working with an array of vectors. We
907 * don't bother to shrink single vectors because we figure that we can
908 * clean it up better with SSA than by inserting piles of vecN instructions
909 * to compact results.
911 int num_levels
= num_array_levels_in_array_of_vector_type(var
->type
);
913 return NULL
; /* Not an array of vectors */
915 struct vec_var_usage
*usage
=
916 rzalloc_size(mem_ctx
, sizeof(*usage
) +
917 num_levels
* sizeof(usage
->levels
[0]));
919 usage
->num_levels
= num_levels
;
920 const struct glsl_type
*type
= var
->type
;
921 for (unsigned i
= 0; i
< num_levels
; i
++) {
922 usage
->levels
[i
].array_len
= glsl_get_length(type
);
923 assert(glsl_get_explicit_stride(type
) == 0);
924 type
= glsl_get_array_element(type
);
926 assert(glsl_type_is_vector_or_scalar(type
));
928 usage
->all_comps
= (1 << glsl_get_components(type
)) - 1;
930 _mesa_hash_table_insert(var_usage_map
, var
, usage
);
935 static struct vec_var_usage
*
936 get_vec_deref_usage(nir_deref_instr
*deref
,
937 struct hash_table
*var_usage_map
,
938 nir_variable_mode modes
,
939 bool add_usage_entry
, void *mem_ctx
)
941 if (!(deref
->mode
& modes
))
944 return get_vec_var_usage(nir_deref_instr_get_variable(deref
),
945 var_usage_map
, add_usage_entry
, mem_ctx
);
949 mark_deref_used(nir_deref_instr
*deref
,
950 nir_component_mask_t comps_read
,
951 nir_component_mask_t comps_written
,
952 nir_deref_instr
*copy_deref
,
953 struct hash_table
*var_usage_map
,
954 nir_variable_mode modes
,
957 if (!(deref
->mode
& modes
))
960 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
962 struct vec_var_usage
*usage
=
963 get_vec_var_usage(var
, var_usage_map
, true, mem_ctx
);
967 usage
->comps_read
|= comps_read
& usage
->all_comps
;
968 usage
->comps_written
|= comps_written
& usage
->all_comps
;
970 struct vec_var_usage
*copy_usage
= NULL
;
972 copy_usage
= get_vec_deref_usage(copy_deref
, var_usage_map
, modes
,
975 if (usage
->vars_copied
== NULL
) {
976 usage
->vars_copied
= _mesa_set_create(mem_ctx
, _mesa_hash_pointer
,
977 _mesa_key_pointer_equal
);
979 _mesa_set_add(usage
->vars_copied
, copy_usage
);
981 usage
->has_external_copy
= true;
986 nir_deref_path_init(&path
, deref
, mem_ctx
);
988 nir_deref_path copy_path
;
990 nir_deref_path_init(©_path
, copy_deref
, mem_ctx
);
993 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
994 struct array_level_usage
*level
= &usage
->levels
[i
];
995 nir_deref_instr
*deref
= path
.path
[i
+ 1];
996 assert(deref
->deref_type
== nir_deref_type_array
||
997 deref
->deref_type
== nir_deref_type_array_wildcard
);
1000 if (deref
->deref_type
== nir_deref_type_array
) {
1001 max_used
= nir_src_is_const(deref
->arr
.index
) ?
1002 nir_src_as_uint(deref
->arr
.index
) : UINT_MAX
;
1004 /* For wildcards, we read or wrote the whole thing. */
1005 assert(deref
->deref_type
== nir_deref_type_array_wildcard
);
1006 max_used
= level
->array_len
- 1;
1009 /* Match each wildcard level with the level on copy_usage */
1010 for (; copy_path
.path
[copy_i
+ 1]; copy_i
++) {
1011 if (copy_path
.path
[copy_i
+ 1]->deref_type
==
1012 nir_deref_type_array_wildcard
)
1015 struct array_level_usage
*copy_level
=
1016 ©_usage
->levels
[copy_i
++];
1018 if (level
->levels_copied
== NULL
) {
1019 level
->levels_copied
=
1020 _mesa_set_create(mem_ctx
, _mesa_hash_pointer
,
1021 _mesa_key_pointer_equal
);
1023 _mesa_set_add(level
->levels_copied
, copy_level
);
1025 /* We have a wildcard and it comes from a variable we aren't
1026 * tracking; flag it and we'll know to not shorten this array.
1028 level
->has_external_copy
= true;
1033 level
->max_written
= MAX2(level
->max_written
, max_used
);
1035 level
->max_read
= MAX2(level
->max_read
, max_used
);
1040 src_is_load_deref(nir_src src
, nir_src deref_src
)
1043 assert(deref_src
.is_ssa
);
1045 if (src
.ssa
->parent_instr
->type
!= nir_instr_type_intrinsic
)
1048 nir_intrinsic_instr
*load
= nir_instr_as_intrinsic(src
.ssa
->parent_instr
);
1049 if (load
->intrinsic
!= nir_intrinsic_load_deref
)
1052 assert(load
->src
[0].is_ssa
);
1054 return load
->src
[0].ssa
== deref_src
.ssa
;
1057 /* Returns all non-self-referential components of a store instruction. A
1058 * component is self-referential if it comes from the same component of a load
1059 * instruction on the same deref. If the only data in a particular component
1060 * of a variable came directly from that component then it's undefined. The
1061 * only way to get defined data into a component of a variable is for it to
1062 * get written there by something outside or from a different component.
1064 * This is a fairly common pattern in shaders that come from either GLSL IR or
1065 * GLSLang because both glsl_to_nir and GLSLang implement write-masking with
1068 static nir_component_mask_t
1069 get_non_self_referential_store_comps(nir_intrinsic_instr
*store
)
1071 nir_component_mask_t comps
= nir_intrinsic_write_mask(store
);
1073 assert(store
->src
[1].is_ssa
);
1074 nir_instr
*src_instr
= store
->src
[1].ssa
->parent_instr
;
1075 if (src_instr
->type
!= nir_instr_type_alu
)
1078 nir_alu_instr
*src_alu
= nir_instr_as_alu(src_instr
);
1080 if (src_alu
->op
== nir_op_imov
||
1081 src_alu
->op
== nir_op_fmov
) {
1082 /* If it's just a swizzle of a load from the same deref, discount any
1083 * channels that don't move in the swizzle.
1085 if (src_is_load_deref(src_alu
->src
[0].src
, store
->src
[0])) {
1086 for (unsigned i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; i
++) {
1087 if (src_alu
->src
[0].swizzle
[i
] == i
)
1088 comps
&= ~(1u << i
);
1091 } else if (src_alu
->op
== nir_op_vec2
||
1092 src_alu
->op
== nir_op_vec3
||
1093 src_alu
->op
== nir_op_vec4
) {
1094 /* If it's a vec, discount any channels that are just loads from the
1095 * same deref put in the same spot.
1097 for (unsigned i
= 0; i
< nir_op_infos
[src_alu
->op
].num_inputs
; i
++) {
1098 if (src_is_load_deref(src_alu
->src
[i
].src
, store
->src
[0]) &&
1099 src_alu
->src
[i
].swizzle
[0] == i
)
1100 comps
&= ~(1u << i
);
1108 find_used_components_impl(nir_function_impl
*impl
,
1109 struct hash_table
*var_usage_map
,
1110 nir_variable_mode modes
,
1113 nir_foreach_block(block
, impl
) {
1114 nir_foreach_instr(instr
, block
) {
1115 if (instr
->type
!= nir_instr_type_intrinsic
)
1118 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1119 switch (intrin
->intrinsic
) {
1120 case nir_intrinsic_load_deref
:
1121 mark_deref_used(nir_src_as_deref(intrin
->src
[0]),
1122 nir_ssa_def_components_read(&intrin
->dest
.ssa
), 0,
1123 NULL
, var_usage_map
, modes
, mem_ctx
);
1126 case nir_intrinsic_store_deref
:
1127 mark_deref_used(nir_src_as_deref(intrin
->src
[0]),
1128 0, get_non_self_referential_store_comps(intrin
),
1129 NULL
, var_usage_map
, modes
, mem_ctx
);
1132 case nir_intrinsic_copy_deref
: {
1133 /* Just mark everything used for copies. */
1134 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
1135 nir_deref_instr
*src
= nir_src_as_deref(intrin
->src
[1]);
1136 mark_deref_used(dst
, 0, ~0, src
, var_usage_map
, modes
, mem_ctx
);
1137 mark_deref_used(src
, ~0, 0, dst
, var_usage_map
, modes
, mem_ctx
);
1149 shrink_vec_var_list(struct exec_list
*vars
,
1150 struct hash_table
*var_usage_map
)
1152 /* Initialize the components kept field of each variable. This is the
1153 * AND of the components written and components read. If a component is
1154 * written but never read, it's dead. If it is read but never written,
1155 * then all values read are undefined garbage and we may as well not read
1158 * The same logic applies to the array length. We make the array length
1159 * the minimum needed required length between read and write and plan to
1160 * discard any OOB access. The one exception here is indirect writes
1161 * because we don't know where they will land and we can't shrink an array
1162 * with indirect writes because previously in-bounds writes may become
1163 * out-of-bounds and have undefined behavior.
1165 * Also, if we have a copy that to/from something we can't shrink, we need
1166 * to leave components and array_len of any wildcards alone.
1168 nir_foreach_variable(var
, vars
) {
1169 struct vec_var_usage
*usage
=
1170 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1174 assert(usage
->comps_kept
== 0);
1175 if (usage
->has_external_copy
)
1176 usage
->comps_kept
= usage
->all_comps
;
1178 usage
->comps_kept
= usage
->comps_read
& usage
->comps_written
;
1180 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1181 struct array_level_usage
*level
= &usage
->levels
[i
];
1182 assert(level
->array_len
> 0);
1184 if (level
->max_written
== UINT_MAX
|| level
->has_external_copy
)
1185 continue; /* Can't shrink */
1187 unsigned max_used
= MIN2(level
->max_read
, level
->max_written
);
1188 level
->array_len
= MIN2(max_used
, level
->array_len
- 1) + 1;
1192 /* In order for variable copies to work, we have to have the same data type
1193 * on the source and the destination. In order to satisfy this, we run a
1194 * little fixed-point algorithm to transitively ensure that we get enough
1195 * components and array elements for this to hold for all copies.
1199 fp_progress
= false;
1200 nir_foreach_variable(var
, vars
) {
1201 struct vec_var_usage
*var_usage
=
1202 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1203 if (!var_usage
|| !var_usage
->vars_copied
)
1206 set_foreach(var_usage
->vars_copied
, copy_entry
) {
1207 struct vec_var_usage
*copy_usage
= (void *)copy_entry
->key
;
1208 if (copy_usage
->comps_kept
!= var_usage
->comps_kept
) {
1209 nir_component_mask_t comps_kept
=
1210 (var_usage
->comps_kept
| copy_usage
->comps_kept
);
1211 var_usage
->comps_kept
= comps_kept
;
1212 copy_usage
->comps_kept
= comps_kept
;
1217 for (unsigned i
= 0; i
< var_usage
->num_levels
; i
++) {
1218 struct array_level_usage
*var_level
= &var_usage
->levels
[i
];
1219 if (!var_level
->levels_copied
)
1222 set_foreach(var_level
->levels_copied
, copy_entry
) {
1223 struct array_level_usage
*copy_level
= (void *)copy_entry
->key
;
1224 if (var_level
->array_len
!= copy_level
->array_len
) {
1225 unsigned array_len
=
1226 MAX2(var_level
->array_len
, copy_level
->array_len
);
1227 var_level
->array_len
= array_len
;
1228 copy_level
->array_len
= array_len
;
1234 } while (fp_progress
);
1236 bool vars_shrunk
= false;
1237 nir_foreach_variable_safe(var
, vars
) {
1238 struct vec_var_usage
*usage
=
1239 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1243 bool shrunk
= false;
1244 const struct glsl_type
*vec_type
= var
->type
;
1245 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1246 /* If we've reduced the array to zero elements at some level, just
1247 * set comps_kept to 0 and delete the variable.
1249 if (usage
->levels
[i
].array_len
== 0) {
1250 usage
->comps_kept
= 0;
1254 assert(usage
->levels
[i
].array_len
<= glsl_get_length(vec_type
));
1255 if (usage
->levels
[i
].array_len
< glsl_get_length(vec_type
))
1257 vec_type
= glsl_get_array_element(vec_type
);
1259 assert(glsl_type_is_vector_or_scalar(vec_type
));
1261 assert(usage
->comps_kept
== (usage
->comps_kept
& usage
->all_comps
));
1262 if (usage
->comps_kept
!= usage
->all_comps
)
1265 if (usage
->comps_kept
== 0) {
1266 /* This variable is dead, remove it */
1268 exec_node_remove(&var
->node
);
1273 /* This variable doesn't need to be shrunk. Remove it from the
1274 * hash table so later steps will ignore it.
1276 _mesa_hash_table_remove_key(var_usage_map
, var
);
1280 /* Build the new var type */
1281 unsigned new_num_comps
= util_bitcount(usage
->comps_kept
);
1282 const struct glsl_type
*new_type
=
1283 glsl_vector_type(glsl_get_base_type(vec_type
), new_num_comps
);
1284 for (int i
= usage
->num_levels
- 1; i
>= 0; i
--) {
1285 assert(usage
->levels
[i
].array_len
> 0);
1286 /* If the original type was a matrix type, we'd like to keep that so
1287 * we don't convert matrices into arrays.
1289 if (i
== usage
->num_levels
- 1 &&
1290 glsl_type_is_matrix(glsl_without_array(var
->type
)) &&
1291 new_num_comps
> 1 && usage
->levels
[i
].array_len
> 1) {
1292 new_type
= glsl_matrix_type(glsl_get_base_type(new_type
),
1294 usage
->levels
[i
].array_len
);
1296 new_type
= glsl_array_type(new_type
, usage
->levels
[i
].array_len
, 0);
1299 var
->type
= new_type
;
1308 vec_deref_is_oob(nir_deref_instr
*deref
,
1309 struct vec_var_usage
*usage
)
1311 nir_deref_path path
;
1312 nir_deref_path_init(&path
, deref
, NULL
);
1315 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1316 nir_deref_instr
*p
= path
.path
[i
+ 1];
1317 if (p
->deref_type
== nir_deref_type_array_wildcard
)
1320 if (nir_src_is_const(p
->arr
.index
) &&
1321 nir_src_as_uint(p
->arr
.index
) >= usage
->levels
[i
].array_len
) {
1327 nir_deref_path_finish(&path
);
1333 vec_deref_is_dead_or_oob(nir_deref_instr
*deref
,
1334 struct hash_table
*var_usage_map
,
1335 nir_variable_mode modes
)
1337 struct vec_var_usage
*usage
=
1338 get_vec_deref_usage(deref
, var_usage_map
, modes
, false, NULL
);
1342 return usage
->comps_kept
== 0 || vec_deref_is_oob(deref
, usage
);
1346 shrink_vec_var_access_impl(nir_function_impl
*impl
,
1347 struct hash_table
*var_usage_map
,
1348 nir_variable_mode modes
)
1351 nir_builder_init(&b
, impl
);
1353 nir_foreach_block(block
, impl
) {
1354 nir_foreach_instr_safe(instr
, block
) {
1355 switch (instr
->type
) {
1356 case nir_instr_type_deref
: {
1357 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
1358 if (!(deref
->mode
& modes
))
1361 /* Clean up any dead derefs we find lying around. They may refer
1362 * to variables we've deleted.
1364 if (nir_deref_instr_remove_if_unused(deref
))
1367 /* Update the type in the deref to keep the types consistent as
1368 * you walk down the chain. We don't need to check if this is one
1369 * of the derefs we're shrinking because this is a no-op if it
1370 * isn't. The worst that could happen is that we accidentally fix
1373 if (deref
->deref_type
== nir_deref_type_var
) {
1374 deref
->type
= deref
->var
->type
;
1375 } else if (deref
->deref_type
== nir_deref_type_array
||
1376 deref
->deref_type
== nir_deref_type_array_wildcard
) {
1377 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
1378 assert(glsl_type_is_array(parent
->type
) ||
1379 glsl_type_is_matrix(parent
->type
));
1380 deref
->type
= glsl_get_array_element(parent
->type
);
1385 case nir_instr_type_intrinsic
: {
1386 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1388 /* If we have a copy whose source or destination has been deleted
1389 * because we determined the variable was dead, then we just
1390 * delete the copy instruction. If the source variable was dead
1391 * then it was writing undefined garbage anyway and if it's the
1392 * destination variable that's dead then the write isn't needed.
1394 if (intrin
->intrinsic
== nir_intrinsic_copy_deref
) {
1395 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
1396 nir_deref_instr
*src
= nir_src_as_deref(intrin
->src
[1]);
1397 if (vec_deref_is_dead_or_oob(dst
, var_usage_map
, modes
) ||
1398 vec_deref_is_dead_or_oob(src
, var_usage_map
, modes
)) {
1399 nir_instr_remove(&intrin
->instr
);
1400 nir_deref_instr_remove_if_unused(dst
);
1401 nir_deref_instr_remove_if_unused(src
);
1406 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
&&
1407 intrin
->intrinsic
!= nir_intrinsic_store_deref
)
1410 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1411 if (!(deref
->mode
& modes
))
1414 struct vec_var_usage
*usage
=
1415 get_vec_deref_usage(deref
, var_usage_map
, modes
, false, NULL
);
1419 if (usage
->comps_kept
== 0 || vec_deref_is_oob(deref
, usage
)) {
1420 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
1422 nir_ssa_undef(&b
, intrin
->dest
.ssa
.num_components
,
1423 intrin
->dest
.ssa
.bit_size
);
1424 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
1425 nir_src_for_ssa(u
));
1427 nir_instr_remove(&intrin
->instr
);
1428 nir_deref_instr_remove_if_unused(deref
);
1432 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
1433 b
.cursor
= nir_after_instr(&intrin
->instr
);
1435 nir_ssa_def
*undef
=
1436 nir_ssa_undef(&b
, 1, intrin
->dest
.ssa
.bit_size
);
1437 nir_ssa_def
*vec_srcs
[NIR_MAX_VEC_COMPONENTS
];
1439 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1440 if (usage
->comps_kept
& (1u << i
))
1441 vec_srcs
[i
] = nir_channel(&b
, &intrin
->dest
.ssa
, c
++);
1443 vec_srcs
[i
] = undef
;
1445 nir_ssa_def
*vec
= nir_vec(&b
, vec_srcs
, intrin
->num_components
);
1447 nir_ssa_def_rewrite_uses_after(&intrin
->dest
.ssa
,
1448 nir_src_for_ssa(vec
),
1451 /* The SSA def is now only used by the swizzle. It's safe to
1452 * shrink the number of components.
1454 assert(list_length(&intrin
->dest
.ssa
.uses
) == c
);
1455 intrin
->num_components
= c
;
1456 intrin
->dest
.ssa
.num_components
= c
;
1458 nir_component_mask_t write_mask
=
1459 nir_intrinsic_write_mask(intrin
);
1461 unsigned swizzle
[NIR_MAX_VEC_COMPONENTS
];
1462 nir_component_mask_t new_write_mask
= 0;
1464 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1465 if (usage
->comps_kept
& (1u << i
)) {
1467 if (write_mask
& (1u << i
))
1468 new_write_mask
|= 1u << c
;
1473 b
.cursor
= nir_before_instr(&intrin
->instr
);
1475 nir_ssa_def
*swizzled
=
1476 nir_swizzle(&b
, intrin
->src
[1].ssa
, swizzle
, c
, false);
1478 /* Rewrite to use the compacted source */
1479 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[1],
1480 nir_src_for_ssa(swizzled
));
1481 nir_intrinsic_set_write_mask(intrin
, new_write_mask
);
1482 intrin
->num_components
= c
;
1495 function_impl_has_vars_with_modes(nir_function_impl
*impl
,
1496 nir_variable_mode modes
)
1498 nir_shader
*shader
= impl
->function
->shader
;
1500 if ((modes
& nir_var_private
) && !exec_list_is_empty(&shader
->globals
))
1503 if ((modes
& nir_var_function
) && !exec_list_is_empty(&impl
->locals
))
1509 /** Attempt to shrink arrays of vectors
1511 * This pass looks at variables which contain a vector or an array (possibly
1512 * multiple dimensions) of vectors and attempts to lower to a smaller vector
1513 * or array. If the pass can prove that a component of a vector (or array of
1514 * vectors) is never really used, then that component will be removed.
1515 * Similarly, the pass attempts to shorten arrays based on what elements it
1516 * can prove are never read or never contain valid data.
1519 nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
)
1521 assert((modes
& (nir_var_private
| nir_var_function
)) == modes
);
1523 void *mem_ctx
= ralloc_context(NULL
);
1525 struct hash_table
*var_usage_map
=
1526 _mesa_hash_table_create(mem_ctx
, _mesa_hash_pointer
,
1527 _mesa_key_pointer_equal
);
1529 bool has_vars_to_shrink
= false;
1530 nir_foreach_function(function
, shader
) {
1531 if (!function
->impl
)
1534 /* Don't even bother crawling the IR if we don't have any variables.
1535 * Given that this pass deletes any unused variables, it's likely that
1536 * we will be in this scenario eventually.
1538 if (function_impl_has_vars_with_modes(function
->impl
, modes
)) {
1539 has_vars_to_shrink
= true;
1540 find_used_components_impl(function
->impl
, var_usage_map
,
1544 if (!has_vars_to_shrink
) {
1545 ralloc_free(mem_ctx
);
1549 bool globals_shrunk
= false;
1550 if (modes
& nir_var_private
)
1551 globals_shrunk
= shrink_vec_var_list(&shader
->globals
, var_usage_map
);
1553 bool progress
= false;
1554 nir_foreach_function(function
, shader
) {
1555 if (!function
->impl
)
1558 bool locals_shrunk
= false;
1559 if (modes
& nir_var_function
) {
1560 locals_shrunk
= shrink_vec_var_list(&function
->impl
->locals
,
1564 if (globals_shrunk
|| locals_shrunk
) {
1565 shrink_vec_var_access_impl(function
->impl
, var_usage_map
, modes
);
1567 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
1568 nir_metadata_dominance
);
1573 ralloc_free(mem_ctx
);