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_or_ifc(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_temp
) {
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_or_ifc(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_or_ifc(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_pointer_hash_table_create(mem_ctx
);
262 assert((modes
& (nir_var_shader_temp
| nir_var_function_temp
)) == modes
);
264 bool has_global_splits
= false;
265 if (modes
& nir_var_shader_temp
) {
266 has_global_splits
= split_var_list_structs(shader
, NULL
,
268 var_field_map
, mem_ctx
);
271 bool progress
= false;
272 nir_foreach_function(function
, shader
) {
276 bool has_local_splits
= false;
277 if (modes
& nir_var_function_temp
) {
278 has_local_splits
= split_var_list_structs(shader
, function
->impl
,
279 &function
->impl
->locals
,
280 var_field_map
, mem_ctx
);
283 if (has_global_splits
|| has_local_splits
) {
284 split_struct_derefs_impl(function
->impl
, var_field_map
,
287 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
288 nir_metadata_dominance
);
293 ralloc_free(mem_ctx
);
298 struct array_level_info
{
304 /* Only set if this is the tail end of the splitting */
308 struct array_split
*splits
;
311 struct array_var_info
{
312 nir_variable
*base_var
;
314 const struct glsl_type
*split_var_type
;
317 struct array_split root_split
;
320 struct array_level_info levels
[0];
324 init_var_list_array_infos(struct exec_list
*vars
,
325 struct hash_table
*var_info_map
,
328 bool has_array
= false;
330 nir_foreach_variable(var
, vars
) {
331 int num_levels
= num_array_levels_in_array_of_vector_type(var
->type
);
335 struct array_var_info
*info
=
336 rzalloc_size(mem_ctx
, sizeof(*info
) +
337 num_levels
* sizeof(info
->levels
[0]));
339 info
->base_var
= var
;
340 info
->num_levels
= num_levels
;
342 const struct glsl_type
*type
= var
->type
;
343 for (int i
= 0; i
< num_levels
; i
++) {
344 info
->levels
[i
].array_len
= glsl_get_length(type
);
345 type
= glsl_get_array_element(type
);
347 /* All levels start out initially as split */
348 info
->levels
[i
].split
= true;
351 _mesa_hash_table_insert(var_info_map
, var
, info
);
358 static struct array_var_info
*
359 get_array_var_info(nir_variable
*var
,
360 struct hash_table
*var_info_map
)
362 struct hash_entry
*entry
=
363 _mesa_hash_table_search(var_info_map
, var
);
364 return entry
? entry
->data
: NULL
;
367 static struct array_var_info
*
368 get_array_deref_info(nir_deref_instr
*deref
,
369 struct hash_table
*var_info_map
,
370 nir_variable_mode modes
)
372 if (!(deref
->mode
& modes
))
375 return get_array_var_info(nir_deref_instr_get_variable(deref
),
380 mark_array_deref_used(nir_deref_instr
*deref
,
381 struct hash_table
*var_info_map
,
382 nir_variable_mode modes
,
385 struct array_var_info
*info
=
386 get_array_deref_info(deref
, var_info_map
, modes
);
391 nir_deref_path_init(&path
, deref
, mem_ctx
);
393 /* Walk the path and look for indirects. If we have an array deref with an
394 * indirect, mark the given level as not being split.
396 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
397 nir_deref_instr
*p
= path
.path
[i
+ 1];
398 if (p
->deref_type
== nir_deref_type_array
&&
399 !nir_src_is_const(p
->arr
.index
))
400 info
->levels
[i
].split
= false;
405 mark_array_usage_impl(nir_function_impl
*impl
,
406 struct hash_table
*var_info_map
,
407 nir_variable_mode modes
,
410 nir_foreach_block(block
, impl
) {
411 nir_foreach_instr(instr
, block
) {
412 if (instr
->type
!= nir_instr_type_intrinsic
)
415 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
416 switch (intrin
->intrinsic
) {
417 case nir_intrinsic_copy_deref
:
418 mark_array_deref_used(nir_src_as_deref(intrin
->src
[1]),
419 var_info_map
, modes
, mem_ctx
);
422 case nir_intrinsic_load_deref
:
423 case nir_intrinsic_store_deref
:
424 mark_array_deref_used(nir_src_as_deref(intrin
->src
[0]),
425 var_info_map
, modes
, mem_ctx
);
436 create_split_array_vars(struct array_var_info
*var_info
,
438 struct array_split
*split
,
441 nir_function_impl
*impl
,
444 while (level
< var_info
->num_levels
&& !var_info
->levels
[level
].split
) {
445 name
= ralloc_asprintf(mem_ctx
, "%s[*]", name
);
449 if (level
== var_info
->num_levels
) {
450 /* We add parens to the variable name so it looks like "(foo[2][*])" so
451 * that further derefs will look like "(foo[2][*])[ssa_6]"
453 name
= ralloc_asprintf(mem_ctx
, "(%s)", name
);
455 nir_variable_mode mode
= var_info
->base_var
->data
.mode
;
456 if (mode
== nir_var_function_temp
) {
457 split
->var
= nir_local_variable_create(impl
,
458 var_info
->split_var_type
, name
);
460 split
->var
= nir_variable_create(shader
, mode
,
461 var_info
->split_var_type
, name
);
464 assert(var_info
->levels
[level
].split
);
465 split
->num_splits
= var_info
->levels
[level
].array_len
;
466 split
->splits
= rzalloc_array(mem_ctx
, struct array_split
,
468 for (unsigned i
= 0; i
< split
->num_splits
; i
++) {
469 create_split_array_vars(var_info
, level
+ 1, &split
->splits
[i
],
470 ralloc_asprintf(mem_ctx
, "%s[%d]", name
, i
),
471 shader
, impl
, mem_ctx
);
477 split_var_list_arrays(nir_shader
*shader
,
478 nir_function_impl
*impl
,
479 struct exec_list
*vars
,
480 struct hash_table
*var_info_map
,
483 struct exec_list split_vars
;
484 exec_list_make_empty(&split_vars
);
486 nir_foreach_variable_safe(var
, vars
) {
487 struct array_var_info
*info
= get_array_var_info(var
, var_info_map
);
491 bool has_split
= false;
492 const struct glsl_type
*split_type
=
493 glsl_without_array_or_matrix(var
->type
);
494 for (int i
= info
->num_levels
- 1; i
>= 0; i
--) {
495 if (info
->levels
[i
].split
) {
500 /* If the original type was a matrix type, we'd like to keep that so
501 * we don't convert matrices into arrays.
503 if (i
== info
->num_levels
- 1 &&
504 glsl_type_is_matrix(glsl_without_array(var
->type
))) {
505 split_type
= glsl_matrix_type(glsl_get_base_type(split_type
),
506 glsl_get_components(split_type
),
507 info
->levels
[i
].array_len
);
509 split_type
= glsl_array_type(split_type
, info
->levels
[i
].array_len
, 0);
514 info
->split_var_type
= split_type
;
515 /* To avoid list confusion (we'll be adding things as we split
516 * variables), pull all of the variables we plan to split off of the
517 * main variable list.
519 exec_node_remove(&var
->node
);
520 exec_list_push_tail(&split_vars
, &var
->node
);
522 assert(split_type
== glsl_get_bare_type(var
->type
));
523 /* If we're not modifying this variable, delete the info so we skip
524 * it faster in later passes.
526 _mesa_hash_table_remove_key(var_info_map
, var
);
530 nir_foreach_variable(var
, &split_vars
) {
531 struct array_var_info
*info
= get_array_var_info(var
, var_info_map
);
532 create_split_array_vars(info
, 0, &info
->root_split
, var
->name
,
533 shader
, impl
, mem_ctx
);
536 return !exec_list_is_empty(&split_vars
);
540 deref_has_split_wildcard(nir_deref_path
*path
,
541 struct array_var_info
*info
)
546 assert(path
->path
[0]->var
== info
->base_var
);
547 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
548 if (path
->path
[i
+ 1]->deref_type
== nir_deref_type_array_wildcard
&&
549 info
->levels
[i
].split
)
557 array_path_is_out_of_bounds(nir_deref_path
*path
,
558 struct array_var_info
*info
)
563 assert(path
->path
[0]->var
== info
->base_var
);
564 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
565 nir_deref_instr
*p
= path
->path
[i
+ 1];
566 if (p
->deref_type
== nir_deref_type_array_wildcard
)
569 if (nir_src_is_const(p
->arr
.index
) &&
570 nir_src_as_uint(p
->arr
.index
) >= info
->levels
[i
].array_len
)
578 emit_split_copies(nir_builder
*b
,
579 struct array_var_info
*dst_info
, nir_deref_path
*dst_path
,
580 unsigned dst_level
, nir_deref_instr
*dst
,
581 struct array_var_info
*src_info
, nir_deref_path
*src_path
,
582 unsigned src_level
, nir_deref_instr
*src
)
584 nir_deref_instr
*dst_p
, *src_p
;
586 while ((dst_p
= dst_path
->path
[dst_level
+ 1])) {
587 if (dst_p
->deref_type
== nir_deref_type_array_wildcard
)
590 dst
= nir_build_deref_follower(b
, dst
, dst_p
);
594 while ((src_p
= src_path
->path
[src_level
+ 1])) {
595 if (src_p
->deref_type
== nir_deref_type_array_wildcard
)
598 src
= nir_build_deref_follower(b
, src
, src_p
);
602 if (src_p
== NULL
|| dst_p
== NULL
) {
603 assert(src_p
== NULL
&& dst_p
== NULL
);
604 nir_copy_deref(b
, dst
, src
);
606 assert(dst_p
->deref_type
== nir_deref_type_array_wildcard
&&
607 src_p
->deref_type
== nir_deref_type_array_wildcard
);
609 if ((dst_info
&& dst_info
->levels
[dst_level
].split
) ||
610 (src_info
&& src_info
->levels
[src_level
].split
)) {
611 /* There are no indirects at this level on one of the source or the
612 * destination so we are lowering it.
614 assert(glsl_get_length(dst_path
->path
[dst_level
]->type
) ==
615 glsl_get_length(src_path
->path
[src_level
]->type
));
616 unsigned len
= glsl_get_length(dst_path
->path
[dst_level
]->type
);
617 for (unsigned i
= 0; i
< len
; i
++) {
618 emit_split_copies(b
, dst_info
, dst_path
, dst_level
+ 1,
619 nir_build_deref_array_imm(b
, dst
, i
),
620 src_info
, src_path
, src_level
+ 1,
621 nir_build_deref_array_imm(b
, src
, i
));
624 /* Neither side is being split so we just keep going */
625 emit_split_copies(b
, dst_info
, dst_path
, dst_level
+ 1,
626 nir_build_deref_array_wildcard(b
, dst
),
627 src_info
, src_path
, src_level
+ 1,
628 nir_build_deref_array_wildcard(b
, src
));
634 split_array_copies_impl(nir_function_impl
*impl
,
635 struct hash_table
*var_info_map
,
636 nir_variable_mode modes
,
640 nir_builder_init(&b
, impl
);
642 nir_foreach_block(block
, impl
) {
643 nir_foreach_instr_safe(instr
, block
) {
644 if (instr
->type
!= nir_instr_type_intrinsic
)
647 nir_intrinsic_instr
*copy
= nir_instr_as_intrinsic(instr
);
648 if (copy
->intrinsic
!= nir_intrinsic_copy_deref
)
651 nir_deref_instr
*dst_deref
= nir_src_as_deref(copy
->src
[0]);
652 nir_deref_instr
*src_deref
= nir_src_as_deref(copy
->src
[1]);
654 struct array_var_info
*dst_info
=
655 get_array_deref_info(dst_deref
, var_info_map
, modes
);
656 struct array_var_info
*src_info
=
657 get_array_deref_info(src_deref
, var_info_map
, modes
);
659 if (!src_info
&& !dst_info
)
662 nir_deref_path dst_path
, src_path
;
663 nir_deref_path_init(&dst_path
, dst_deref
, mem_ctx
);
664 nir_deref_path_init(&src_path
, src_deref
, mem_ctx
);
666 if (!deref_has_split_wildcard(&dst_path
, dst_info
) &&
667 !deref_has_split_wildcard(&src_path
, src_info
))
670 b
.cursor
= nir_instr_remove(©
->instr
);
672 emit_split_copies(&b
, dst_info
, &dst_path
, 0, dst_path
.path
[0],
673 src_info
, &src_path
, 0, src_path
.path
[0]);
679 split_array_access_impl(nir_function_impl
*impl
,
680 struct hash_table
*var_info_map
,
681 nir_variable_mode modes
,
685 nir_builder_init(&b
, impl
);
687 nir_foreach_block(block
, impl
) {
688 nir_foreach_instr_safe(instr
, block
) {
689 if (instr
->type
== nir_instr_type_deref
) {
690 /* Clean up any dead derefs we find lying around. They may refer
691 * to variables we're planning to split.
693 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
694 if (deref
->mode
& modes
)
695 nir_deref_instr_remove_if_unused(deref
);
699 if (instr
->type
!= nir_instr_type_intrinsic
)
702 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
703 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
&&
704 intrin
->intrinsic
!= nir_intrinsic_store_deref
&&
705 intrin
->intrinsic
!= nir_intrinsic_copy_deref
)
708 const unsigned num_derefs
=
709 intrin
->intrinsic
== nir_intrinsic_copy_deref
? 2 : 1;
711 for (unsigned d
= 0; d
< num_derefs
; d
++) {
712 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[d
]);
714 struct array_var_info
*info
=
715 get_array_deref_info(deref
, var_info_map
, modes
);
720 nir_deref_path_init(&path
, deref
, mem_ctx
);
722 b
.cursor
= nir_before_instr(&intrin
->instr
);
724 if (array_path_is_out_of_bounds(&path
, info
)) {
725 /* If one of the derefs is out-of-bounds, we just delete the
726 * instruction. If a destination is out of bounds, then it may
727 * have been in-bounds prior to shrinking so we don't want to
728 * accidentally stomp something. However, we've already proven
729 * that it will never be read so it's safe to delete. If a
730 * source is out of bounds then it is loading random garbage.
731 * For loads, we replace their uses with an undef instruction
732 * and for copies we just delete the copy since it was writing
733 * undefined garbage anyway and we may as well leave the random
734 * garbage in the destination alone.
736 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
738 nir_ssa_undef(&b
, intrin
->dest
.ssa
.num_components
,
739 intrin
->dest
.ssa
.bit_size
);
740 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
743 nir_instr_remove(&intrin
->instr
);
744 for (unsigned i
= 0; i
< num_derefs
; i
++)
745 nir_deref_instr_remove_if_unused(nir_src_as_deref(intrin
->src
[i
]));
749 struct array_split
*split
= &info
->root_split
;
750 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
751 if (info
->levels
[i
].split
) {
752 nir_deref_instr
*p
= path
.path
[i
+ 1];
753 unsigned index
= nir_src_as_uint(p
->arr
.index
);
754 assert(index
< info
->levels
[i
].array_len
);
755 split
= &split
->splits
[index
];
758 assert(!split
->splits
&& split
->var
);
760 nir_deref_instr
*new_deref
= nir_build_deref_var(&b
, split
->var
);
761 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
762 if (!info
->levels
[i
].split
) {
763 new_deref
= nir_build_deref_follower(&b
, new_deref
,
767 assert(new_deref
->type
== deref
->type
);
769 /* Rewrite the deref source to point to the split one */
770 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[d
],
771 nir_src_for_ssa(&new_deref
->dest
.ssa
));
772 nir_deref_instr_remove_if_unused(deref
);
778 /** A pass for splitting arrays of vectors into multiple variables
780 * This pass looks at arrays (possibly multiple levels) of vectors (not
781 * structures or other types) and tries to split them into piles of variables,
782 * one for each array element. The heuristic used is simple: If a given array
783 * level is never used with an indirect, that array level will get split.
785 * This pass probably could handles structures easily enough but making a pass
786 * that could see through an array of structures of arrays would be difficult
787 * so it's best to just run nir_split_struct_vars first.
790 nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
)
792 void *mem_ctx
= ralloc_context(NULL
);
793 struct hash_table
*var_info_map
= _mesa_pointer_hash_table_create(mem_ctx
);
795 assert((modes
& (nir_var_shader_temp
| nir_var_function_temp
)) == modes
);
797 bool has_global_array
= false;
798 if (modes
& nir_var_shader_temp
) {
799 has_global_array
= init_var_list_array_infos(&shader
->globals
,
800 var_info_map
, mem_ctx
);
803 bool has_any_array
= false;
804 nir_foreach_function(function
, shader
) {
808 bool has_local_array
= false;
809 if (modes
& nir_var_function_temp
) {
810 has_local_array
= init_var_list_array_infos(&function
->impl
->locals
,
811 var_info_map
, mem_ctx
);
814 if (has_global_array
|| has_local_array
) {
815 has_any_array
= true;
816 mark_array_usage_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
820 /* If we failed to find any arrays of arrays, bail early. */
821 if (!has_any_array
) {
822 ralloc_free(mem_ctx
);
826 bool has_global_splits
= false;
827 if (modes
& nir_var_shader_temp
) {
828 has_global_splits
= split_var_list_arrays(shader
, NULL
,
830 var_info_map
, mem_ctx
);
833 bool progress
= false;
834 nir_foreach_function(function
, shader
) {
838 bool has_local_splits
= false;
839 if (modes
& nir_var_function_temp
) {
840 has_local_splits
= split_var_list_arrays(shader
, function
->impl
,
841 &function
->impl
->locals
,
842 var_info_map
, mem_ctx
);
845 if (has_global_splits
|| has_local_splits
) {
846 split_array_copies_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
847 split_array_access_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
849 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
850 nir_metadata_dominance
);
855 ralloc_free(mem_ctx
);
860 struct array_level_usage
{
863 /* The value UINT_MAX will be used to indicate an indirect */
865 unsigned max_written
;
867 /* True if there is a copy that isn't to/from a shrinkable array */
868 bool has_external_copy
;
869 struct set
*levels_copied
;
872 struct vec_var_usage
{
873 /* Convenience set of all components this variable has */
874 nir_component_mask_t all_comps
;
876 nir_component_mask_t comps_read
;
877 nir_component_mask_t comps_written
;
879 nir_component_mask_t comps_kept
;
881 /* True if there is a copy that isn't to/from a shrinkable vector */
882 bool has_external_copy
;
883 struct set
*vars_copied
;
886 struct array_level_usage levels
[0];
889 static struct vec_var_usage
*
890 get_vec_var_usage(nir_variable
*var
,
891 struct hash_table
*var_usage_map
,
892 bool add_usage_entry
, void *mem_ctx
)
894 struct hash_entry
*entry
= _mesa_hash_table_search(var_usage_map
, var
);
898 if (!add_usage_entry
)
901 /* Check to make sure that we are working with an array of vectors. We
902 * don't bother to shrink single vectors because we figure that we can
903 * clean it up better with SSA than by inserting piles of vecN instructions
904 * to compact results.
906 int num_levels
= num_array_levels_in_array_of_vector_type(var
->type
);
908 return NULL
; /* Not an array of vectors */
910 struct vec_var_usage
*usage
=
911 rzalloc_size(mem_ctx
, sizeof(*usage
) +
912 num_levels
* sizeof(usage
->levels
[0]));
914 usage
->num_levels
= num_levels
;
915 const struct glsl_type
*type
= var
->type
;
916 for (unsigned i
= 0; i
< num_levels
; i
++) {
917 usage
->levels
[i
].array_len
= glsl_get_length(type
);
918 type
= glsl_get_array_element(type
);
920 assert(glsl_type_is_vector_or_scalar(type
));
922 usage
->all_comps
= (1 << glsl_get_components(type
)) - 1;
924 _mesa_hash_table_insert(var_usage_map
, var
, usage
);
929 static struct vec_var_usage
*
930 get_vec_deref_usage(nir_deref_instr
*deref
,
931 struct hash_table
*var_usage_map
,
932 nir_variable_mode modes
,
933 bool add_usage_entry
, void *mem_ctx
)
935 if (!(deref
->mode
& modes
))
938 return get_vec_var_usage(nir_deref_instr_get_variable(deref
),
939 var_usage_map
, add_usage_entry
, mem_ctx
);
943 mark_deref_used(nir_deref_instr
*deref
,
944 nir_component_mask_t comps_read
,
945 nir_component_mask_t comps_written
,
946 nir_deref_instr
*copy_deref
,
947 struct hash_table
*var_usage_map
,
948 nir_variable_mode modes
,
951 if (!(deref
->mode
& modes
))
954 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
956 struct vec_var_usage
*usage
=
957 get_vec_var_usage(var
, var_usage_map
, true, mem_ctx
);
961 usage
->comps_read
|= comps_read
& usage
->all_comps
;
962 usage
->comps_written
|= comps_written
& usage
->all_comps
;
964 struct vec_var_usage
*copy_usage
= NULL
;
966 copy_usage
= get_vec_deref_usage(copy_deref
, var_usage_map
, modes
,
969 if (usage
->vars_copied
== NULL
) {
970 usage
->vars_copied
= _mesa_pointer_set_create(mem_ctx
);
972 _mesa_set_add(usage
->vars_copied
, copy_usage
);
974 usage
->has_external_copy
= true;
979 nir_deref_path_init(&path
, deref
, mem_ctx
);
981 nir_deref_path copy_path
;
983 nir_deref_path_init(©_path
, copy_deref
, mem_ctx
);
986 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
987 struct array_level_usage
*level
= &usage
->levels
[i
];
988 nir_deref_instr
*deref
= path
.path
[i
+ 1];
989 assert(deref
->deref_type
== nir_deref_type_array
||
990 deref
->deref_type
== nir_deref_type_array_wildcard
);
993 if (deref
->deref_type
== nir_deref_type_array
) {
994 max_used
= nir_src_is_const(deref
->arr
.index
) ?
995 nir_src_as_uint(deref
->arr
.index
) : UINT_MAX
;
997 /* For wildcards, we read or wrote the whole thing. */
998 assert(deref
->deref_type
== nir_deref_type_array_wildcard
);
999 max_used
= level
->array_len
- 1;
1002 /* Match each wildcard level with the level on copy_usage */
1003 for (; copy_path
.path
[copy_i
+ 1]; copy_i
++) {
1004 if (copy_path
.path
[copy_i
+ 1]->deref_type
==
1005 nir_deref_type_array_wildcard
)
1008 struct array_level_usage
*copy_level
=
1009 ©_usage
->levels
[copy_i
++];
1011 if (level
->levels_copied
== NULL
) {
1012 level
->levels_copied
= _mesa_pointer_set_create(mem_ctx
);
1014 _mesa_set_add(level
->levels_copied
, copy_level
);
1016 /* We have a wildcard and it comes from a variable we aren't
1017 * tracking; flag it and we'll know to not shorten this array.
1019 level
->has_external_copy
= true;
1024 level
->max_written
= MAX2(level
->max_written
, max_used
);
1026 level
->max_read
= MAX2(level
->max_read
, max_used
);
1031 src_is_load_deref(nir_src src
, nir_src deref_src
)
1034 assert(deref_src
.is_ssa
);
1036 if (src
.ssa
->parent_instr
->type
!= nir_instr_type_intrinsic
)
1039 nir_intrinsic_instr
*load
= nir_instr_as_intrinsic(src
.ssa
->parent_instr
);
1040 if (load
->intrinsic
!= nir_intrinsic_load_deref
)
1043 assert(load
->src
[0].is_ssa
);
1045 return load
->src
[0].ssa
== deref_src
.ssa
;
1048 /* Returns all non-self-referential components of a store instruction. A
1049 * component is self-referential if it comes from the same component of a load
1050 * instruction on the same deref. If the only data in a particular component
1051 * of a variable came directly from that component then it's undefined. The
1052 * only way to get defined data into a component of a variable is for it to
1053 * get written there by something outside or from a different component.
1055 * This is a fairly common pattern in shaders that come from either GLSL IR or
1056 * GLSLang because both glsl_to_nir and GLSLang implement write-masking with
1059 static nir_component_mask_t
1060 get_non_self_referential_store_comps(nir_intrinsic_instr
*store
)
1062 nir_component_mask_t comps
= nir_intrinsic_write_mask(store
);
1064 assert(store
->src
[1].is_ssa
);
1065 nir_instr
*src_instr
= store
->src
[1].ssa
->parent_instr
;
1066 if (src_instr
->type
!= nir_instr_type_alu
)
1069 nir_alu_instr
*src_alu
= nir_instr_as_alu(src_instr
);
1071 if (src_alu
->op
== nir_op_imov
||
1072 src_alu
->op
== nir_op_fmov
) {
1073 /* If it's just a swizzle of a load from the same deref, discount any
1074 * channels that don't move in the swizzle.
1076 if (src_is_load_deref(src_alu
->src
[0].src
, store
->src
[0])) {
1077 for (unsigned i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; i
++) {
1078 if (src_alu
->src
[0].swizzle
[i
] == i
)
1079 comps
&= ~(1u << i
);
1082 } else if (src_alu
->op
== nir_op_vec2
||
1083 src_alu
->op
== nir_op_vec3
||
1084 src_alu
->op
== nir_op_vec4
) {
1085 /* If it's a vec, discount any channels that are just loads from the
1086 * same deref put in the same spot.
1088 for (unsigned i
= 0; i
< nir_op_infos
[src_alu
->op
].num_inputs
; i
++) {
1089 if (src_is_load_deref(src_alu
->src
[i
].src
, store
->src
[0]) &&
1090 src_alu
->src
[i
].swizzle
[0] == i
)
1091 comps
&= ~(1u << i
);
1099 find_used_components_impl(nir_function_impl
*impl
,
1100 struct hash_table
*var_usage_map
,
1101 nir_variable_mode modes
,
1104 nir_foreach_block(block
, impl
) {
1105 nir_foreach_instr(instr
, block
) {
1106 if (instr
->type
!= nir_instr_type_intrinsic
)
1109 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1110 switch (intrin
->intrinsic
) {
1111 case nir_intrinsic_load_deref
:
1112 mark_deref_used(nir_src_as_deref(intrin
->src
[0]),
1113 nir_ssa_def_components_read(&intrin
->dest
.ssa
), 0,
1114 NULL
, var_usage_map
, modes
, mem_ctx
);
1117 case nir_intrinsic_store_deref
:
1118 mark_deref_used(nir_src_as_deref(intrin
->src
[0]),
1119 0, get_non_self_referential_store_comps(intrin
),
1120 NULL
, var_usage_map
, modes
, mem_ctx
);
1123 case nir_intrinsic_copy_deref
: {
1124 /* Just mark everything used for copies. */
1125 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
1126 nir_deref_instr
*src
= nir_src_as_deref(intrin
->src
[1]);
1127 mark_deref_used(dst
, 0, ~0, src
, var_usage_map
, modes
, mem_ctx
);
1128 mark_deref_used(src
, ~0, 0, dst
, var_usage_map
, modes
, mem_ctx
);
1140 shrink_vec_var_list(struct exec_list
*vars
,
1141 struct hash_table
*var_usage_map
)
1143 /* Initialize the components kept field of each variable. This is the
1144 * AND of the components written and components read. If a component is
1145 * written but never read, it's dead. If it is read but never written,
1146 * then all values read are undefined garbage and we may as well not read
1149 * The same logic applies to the array length. We make the array length
1150 * the minimum needed required length between read and write and plan to
1151 * discard any OOB access. The one exception here is indirect writes
1152 * because we don't know where they will land and we can't shrink an array
1153 * with indirect writes because previously in-bounds writes may become
1154 * out-of-bounds and have undefined behavior.
1156 * Also, if we have a copy that to/from something we can't shrink, we need
1157 * to leave components and array_len of any wildcards alone.
1159 nir_foreach_variable(var
, vars
) {
1160 struct vec_var_usage
*usage
=
1161 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1165 assert(usage
->comps_kept
== 0);
1166 if (usage
->has_external_copy
)
1167 usage
->comps_kept
= usage
->all_comps
;
1169 usage
->comps_kept
= usage
->comps_read
& usage
->comps_written
;
1171 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1172 struct array_level_usage
*level
= &usage
->levels
[i
];
1173 assert(level
->array_len
> 0);
1175 if (level
->max_written
== UINT_MAX
|| level
->has_external_copy
)
1176 continue; /* Can't shrink */
1178 unsigned max_used
= MIN2(level
->max_read
, level
->max_written
);
1179 level
->array_len
= MIN2(max_used
, level
->array_len
- 1) + 1;
1183 /* In order for variable copies to work, we have to have the same data type
1184 * on the source and the destination. In order to satisfy this, we run a
1185 * little fixed-point algorithm to transitively ensure that we get enough
1186 * components and array elements for this to hold for all copies.
1190 fp_progress
= false;
1191 nir_foreach_variable(var
, vars
) {
1192 struct vec_var_usage
*var_usage
=
1193 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1194 if (!var_usage
|| !var_usage
->vars_copied
)
1197 set_foreach(var_usage
->vars_copied
, copy_entry
) {
1198 struct vec_var_usage
*copy_usage
= (void *)copy_entry
->key
;
1199 if (copy_usage
->comps_kept
!= var_usage
->comps_kept
) {
1200 nir_component_mask_t comps_kept
=
1201 (var_usage
->comps_kept
| copy_usage
->comps_kept
);
1202 var_usage
->comps_kept
= comps_kept
;
1203 copy_usage
->comps_kept
= comps_kept
;
1208 for (unsigned i
= 0; i
< var_usage
->num_levels
; i
++) {
1209 struct array_level_usage
*var_level
= &var_usage
->levels
[i
];
1210 if (!var_level
->levels_copied
)
1213 set_foreach(var_level
->levels_copied
, copy_entry
) {
1214 struct array_level_usage
*copy_level
= (void *)copy_entry
->key
;
1215 if (var_level
->array_len
!= copy_level
->array_len
) {
1216 unsigned array_len
=
1217 MAX2(var_level
->array_len
, copy_level
->array_len
);
1218 var_level
->array_len
= array_len
;
1219 copy_level
->array_len
= array_len
;
1225 } while (fp_progress
);
1227 bool vars_shrunk
= false;
1228 nir_foreach_variable_safe(var
, vars
) {
1229 struct vec_var_usage
*usage
=
1230 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1234 bool shrunk
= false;
1235 const struct glsl_type
*vec_type
= var
->type
;
1236 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1237 /* If we've reduced the array to zero elements at some level, just
1238 * set comps_kept to 0 and delete the variable.
1240 if (usage
->levels
[i
].array_len
== 0) {
1241 usage
->comps_kept
= 0;
1245 assert(usage
->levels
[i
].array_len
<= glsl_get_length(vec_type
));
1246 if (usage
->levels
[i
].array_len
< glsl_get_length(vec_type
))
1248 vec_type
= glsl_get_array_element(vec_type
);
1250 assert(glsl_type_is_vector_or_scalar(vec_type
));
1252 assert(usage
->comps_kept
== (usage
->comps_kept
& usage
->all_comps
));
1253 if (usage
->comps_kept
!= usage
->all_comps
)
1256 if (usage
->comps_kept
== 0) {
1257 /* This variable is dead, remove it */
1259 exec_node_remove(&var
->node
);
1264 /* This variable doesn't need to be shrunk. Remove it from the
1265 * hash table so later steps will ignore it.
1267 _mesa_hash_table_remove_key(var_usage_map
, var
);
1271 /* Build the new var type */
1272 unsigned new_num_comps
= util_bitcount(usage
->comps_kept
);
1273 const struct glsl_type
*new_type
=
1274 glsl_vector_type(glsl_get_base_type(vec_type
), new_num_comps
);
1275 for (int i
= usage
->num_levels
- 1; i
>= 0; i
--) {
1276 assert(usage
->levels
[i
].array_len
> 0);
1277 /* If the original type was a matrix type, we'd like to keep that so
1278 * we don't convert matrices into arrays.
1280 if (i
== usage
->num_levels
- 1 &&
1281 glsl_type_is_matrix(glsl_without_array(var
->type
)) &&
1282 new_num_comps
> 1 && usage
->levels
[i
].array_len
> 1) {
1283 new_type
= glsl_matrix_type(glsl_get_base_type(new_type
),
1285 usage
->levels
[i
].array_len
);
1287 new_type
= glsl_array_type(new_type
, usage
->levels
[i
].array_len
, 0);
1290 var
->type
= new_type
;
1299 vec_deref_is_oob(nir_deref_instr
*deref
,
1300 struct vec_var_usage
*usage
)
1302 nir_deref_path path
;
1303 nir_deref_path_init(&path
, deref
, NULL
);
1306 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1307 nir_deref_instr
*p
= path
.path
[i
+ 1];
1308 if (p
->deref_type
== nir_deref_type_array_wildcard
)
1311 if (nir_src_is_const(p
->arr
.index
) &&
1312 nir_src_as_uint(p
->arr
.index
) >= usage
->levels
[i
].array_len
) {
1318 nir_deref_path_finish(&path
);
1324 vec_deref_is_dead_or_oob(nir_deref_instr
*deref
,
1325 struct hash_table
*var_usage_map
,
1326 nir_variable_mode modes
)
1328 struct vec_var_usage
*usage
=
1329 get_vec_deref_usage(deref
, var_usage_map
, modes
, false, NULL
);
1333 return usage
->comps_kept
== 0 || vec_deref_is_oob(deref
, usage
);
1337 shrink_vec_var_access_impl(nir_function_impl
*impl
,
1338 struct hash_table
*var_usage_map
,
1339 nir_variable_mode modes
)
1342 nir_builder_init(&b
, impl
);
1344 nir_foreach_block(block
, impl
) {
1345 nir_foreach_instr_safe(instr
, block
) {
1346 switch (instr
->type
) {
1347 case nir_instr_type_deref
: {
1348 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
1349 if (!(deref
->mode
& modes
))
1352 /* Clean up any dead derefs we find lying around. They may refer
1353 * to variables we've deleted.
1355 if (nir_deref_instr_remove_if_unused(deref
))
1358 /* Update the type in the deref to keep the types consistent as
1359 * you walk down the chain. We don't need to check if this is one
1360 * of the derefs we're shrinking because this is a no-op if it
1361 * isn't. The worst that could happen is that we accidentally fix
1364 if (deref
->deref_type
== nir_deref_type_var
) {
1365 deref
->type
= deref
->var
->type
;
1366 } else if (deref
->deref_type
== nir_deref_type_array
||
1367 deref
->deref_type
== nir_deref_type_array_wildcard
) {
1368 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
1369 assert(glsl_type_is_array(parent
->type
) ||
1370 glsl_type_is_matrix(parent
->type
));
1371 deref
->type
= glsl_get_array_element(parent
->type
);
1376 case nir_instr_type_intrinsic
: {
1377 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1379 /* If we have a copy whose source or destination has been deleted
1380 * because we determined the variable was dead, then we just
1381 * delete the copy instruction. If the source variable was dead
1382 * then it was writing undefined garbage anyway and if it's the
1383 * destination variable that's dead then the write isn't needed.
1385 if (intrin
->intrinsic
== nir_intrinsic_copy_deref
) {
1386 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
1387 nir_deref_instr
*src
= nir_src_as_deref(intrin
->src
[1]);
1388 if (vec_deref_is_dead_or_oob(dst
, var_usage_map
, modes
) ||
1389 vec_deref_is_dead_or_oob(src
, var_usage_map
, modes
)) {
1390 nir_instr_remove(&intrin
->instr
);
1391 nir_deref_instr_remove_if_unused(dst
);
1392 nir_deref_instr_remove_if_unused(src
);
1397 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
&&
1398 intrin
->intrinsic
!= nir_intrinsic_store_deref
)
1401 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1402 if (!(deref
->mode
& modes
))
1405 struct vec_var_usage
*usage
=
1406 get_vec_deref_usage(deref
, var_usage_map
, modes
, false, NULL
);
1410 if (usage
->comps_kept
== 0 || vec_deref_is_oob(deref
, usage
)) {
1411 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
1413 nir_ssa_undef(&b
, intrin
->dest
.ssa
.num_components
,
1414 intrin
->dest
.ssa
.bit_size
);
1415 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
1416 nir_src_for_ssa(u
));
1418 nir_instr_remove(&intrin
->instr
);
1419 nir_deref_instr_remove_if_unused(deref
);
1423 /* If we're not dropping any components, there's no need to
1426 if (usage
->comps_kept
== usage
->all_comps
)
1429 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
1430 b
.cursor
= nir_after_instr(&intrin
->instr
);
1432 nir_ssa_def
*undef
=
1433 nir_ssa_undef(&b
, 1, intrin
->dest
.ssa
.bit_size
);
1434 nir_ssa_def
*vec_srcs
[NIR_MAX_VEC_COMPONENTS
];
1436 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1437 if (usage
->comps_kept
& (1u << i
))
1438 vec_srcs
[i
] = nir_channel(&b
, &intrin
->dest
.ssa
, c
++);
1440 vec_srcs
[i
] = undef
;
1442 nir_ssa_def
*vec
= nir_vec(&b
, vec_srcs
, intrin
->num_components
);
1444 nir_ssa_def_rewrite_uses_after(&intrin
->dest
.ssa
,
1445 nir_src_for_ssa(vec
),
1448 /* The SSA def is now only used by the swizzle. It's safe to
1449 * shrink the number of components.
1451 assert(list_length(&intrin
->dest
.ssa
.uses
) == c
);
1452 intrin
->num_components
= c
;
1453 intrin
->dest
.ssa
.num_components
= c
;
1455 nir_component_mask_t write_mask
=
1456 nir_intrinsic_write_mask(intrin
);
1458 unsigned swizzle
[NIR_MAX_VEC_COMPONENTS
];
1459 nir_component_mask_t new_write_mask
= 0;
1461 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1462 if (usage
->comps_kept
& (1u << i
)) {
1464 if (write_mask
& (1u << i
))
1465 new_write_mask
|= 1u << c
;
1470 b
.cursor
= nir_before_instr(&intrin
->instr
);
1472 nir_ssa_def
*swizzled
=
1473 nir_swizzle(&b
, intrin
->src
[1].ssa
, swizzle
, c
, false);
1475 /* Rewrite to use the compacted source */
1476 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[1],
1477 nir_src_for_ssa(swizzled
));
1478 nir_intrinsic_set_write_mask(intrin
, new_write_mask
);
1479 intrin
->num_components
= c
;
1492 function_impl_has_vars_with_modes(nir_function_impl
*impl
,
1493 nir_variable_mode modes
)
1495 nir_shader
*shader
= impl
->function
->shader
;
1497 if ((modes
& nir_var_shader_temp
) && !exec_list_is_empty(&shader
->globals
))
1500 if ((modes
& nir_var_function_temp
) && !exec_list_is_empty(&impl
->locals
))
1506 /** Attempt to shrink arrays of vectors
1508 * This pass looks at variables which contain a vector or an array (possibly
1509 * multiple dimensions) of vectors and attempts to lower to a smaller vector
1510 * or array. If the pass can prove that a component of a vector (or array of
1511 * vectors) is never really used, then that component will be removed.
1512 * Similarly, the pass attempts to shorten arrays based on what elements it
1513 * can prove are never read or never contain valid data.
1516 nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
)
1518 assert((modes
& (nir_var_shader_temp
| nir_var_function_temp
)) == modes
);
1520 void *mem_ctx
= ralloc_context(NULL
);
1522 struct hash_table
*var_usage_map
=
1523 _mesa_pointer_hash_table_create(mem_ctx
);
1525 bool has_vars_to_shrink
= false;
1526 nir_foreach_function(function
, shader
) {
1527 if (!function
->impl
)
1530 /* Don't even bother crawling the IR if we don't have any variables.
1531 * Given that this pass deletes any unused variables, it's likely that
1532 * we will be in this scenario eventually.
1534 if (function_impl_has_vars_with_modes(function
->impl
, modes
)) {
1535 has_vars_to_shrink
= true;
1536 find_used_components_impl(function
->impl
, var_usage_map
,
1540 if (!has_vars_to_shrink
) {
1541 ralloc_free(mem_ctx
);
1545 bool globals_shrunk
= false;
1546 if (modes
& nir_var_shader_temp
)
1547 globals_shrunk
= shrink_vec_var_list(&shader
->globals
, var_usage_map
);
1549 bool progress
= false;
1550 nir_foreach_function(function
, shader
) {
1551 if (!function
->impl
)
1554 bool locals_shrunk
= false;
1555 if (modes
& nir_var_function_temp
) {
1556 locals_shrunk
= shrink_vec_var_list(&function
->impl
->locals
,
1560 if (globals_shrunk
|| locals_shrunk
) {
1561 shrink_vec_var_access_impl(function
->impl
, var_usage_map
, modes
);
1563 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
1564 nir_metadata_dominance
);
1569 ralloc_free(mem_ctx
);