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 assert(glsl_get_explicit_stride(type
) == 0);
345 info
->levels
[i
].array_len
= glsl_get_length(type
);
346 type
= glsl_get_array_element(type
);
348 /* All levels start out initially as split */
349 info
->levels
[i
].split
= true;
352 _mesa_hash_table_insert(var_info_map
, var
, info
);
359 static struct array_var_info
*
360 get_array_var_info(nir_variable
*var
,
361 struct hash_table
*var_info_map
)
363 struct hash_entry
*entry
=
364 _mesa_hash_table_search(var_info_map
, var
);
365 return entry
? entry
->data
: NULL
;
368 static struct array_var_info
*
369 get_array_deref_info(nir_deref_instr
*deref
,
370 struct hash_table
*var_info_map
,
371 nir_variable_mode modes
)
373 if (!(deref
->mode
& modes
))
376 return get_array_var_info(nir_deref_instr_get_variable(deref
),
381 mark_array_deref_used(nir_deref_instr
*deref
,
382 struct hash_table
*var_info_map
,
383 nir_variable_mode modes
,
386 struct array_var_info
*info
=
387 get_array_deref_info(deref
, var_info_map
, modes
);
392 nir_deref_path_init(&path
, deref
, mem_ctx
);
394 /* Walk the path and look for indirects. If we have an array deref with an
395 * indirect, mark the given level as not being split.
397 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
398 nir_deref_instr
*p
= path
.path
[i
+ 1];
399 if (p
->deref_type
== nir_deref_type_array
&&
400 !nir_src_is_const(p
->arr
.index
))
401 info
->levels
[i
].split
= false;
406 mark_array_usage_impl(nir_function_impl
*impl
,
407 struct hash_table
*var_info_map
,
408 nir_variable_mode modes
,
411 nir_foreach_block(block
, impl
) {
412 nir_foreach_instr(instr
, block
) {
413 if (instr
->type
!= nir_instr_type_intrinsic
)
416 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
417 switch (intrin
->intrinsic
) {
418 case nir_intrinsic_copy_deref
:
419 mark_array_deref_used(nir_src_as_deref(intrin
->src
[1]),
420 var_info_map
, modes
, mem_ctx
);
423 case nir_intrinsic_load_deref
:
424 case nir_intrinsic_store_deref
:
425 mark_array_deref_used(nir_src_as_deref(intrin
->src
[0]),
426 var_info_map
, modes
, mem_ctx
);
437 create_split_array_vars(struct array_var_info
*var_info
,
439 struct array_split
*split
,
442 nir_function_impl
*impl
,
445 while (level
< var_info
->num_levels
&& !var_info
->levels
[level
].split
) {
446 name
= ralloc_asprintf(mem_ctx
, "%s[*]", name
);
450 if (level
== var_info
->num_levels
) {
451 /* We add parens to the variable name so it looks like "(foo[2][*])" so
452 * that further derefs will look like "(foo[2][*])[ssa_6]"
454 name
= ralloc_asprintf(mem_ctx
, "(%s)", name
);
456 nir_variable_mode mode
= var_info
->base_var
->data
.mode
;
457 if (mode
== nir_var_function_temp
) {
458 split
->var
= nir_local_variable_create(impl
,
459 var_info
->split_var_type
, name
);
461 split
->var
= nir_variable_create(shader
, mode
,
462 var_info
->split_var_type
, name
);
465 assert(var_info
->levels
[level
].split
);
466 split
->num_splits
= var_info
->levels
[level
].array_len
;
467 split
->splits
= rzalloc_array(mem_ctx
, struct array_split
,
469 for (unsigned i
= 0; i
< split
->num_splits
; i
++) {
470 create_split_array_vars(var_info
, level
+ 1, &split
->splits
[i
],
471 ralloc_asprintf(mem_ctx
, "%s[%d]", name
, i
),
472 shader
, impl
, mem_ctx
);
478 split_var_list_arrays(nir_shader
*shader
,
479 nir_function_impl
*impl
,
480 struct exec_list
*vars
,
481 struct hash_table
*var_info_map
,
484 struct exec_list split_vars
;
485 exec_list_make_empty(&split_vars
);
487 nir_foreach_variable_safe(var
, vars
) {
488 struct array_var_info
*info
= get_array_var_info(var
, var_info_map
);
492 bool has_split
= false;
493 const struct glsl_type
*split_type
=
494 glsl_without_array_or_matrix(var
->type
);
495 for (int i
= info
->num_levels
- 1; i
>= 0; i
--) {
496 if (info
->levels
[i
].split
) {
501 /* If the original type was a matrix type, we'd like to keep that so
502 * we don't convert matrices into arrays.
504 if (i
== info
->num_levels
- 1 &&
505 glsl_type_is_matrix(glsl_without_array(var
->type
))) {
506 split_type
= glsl_matrix_type(glsl_get_base_type(split_type
),
507 glsl_get_components(split_type
),
508 info
->levels
[i
].array_len
);
510 split_type
= glsl_array_type(split_type
, info
->levels
[i
].array_len
, 0);
515 info
->split_var_type
= split_type
;
516 /* To avoid list confusion (we'll be adding things as we split
517 * variables), pull all of the variables we plan to split off of the
518 * main variable list.
520 exec_node_remove(&var
->node
);
521 exec_list_push_tail(&split_vars
, &var
->node
);
523 assert(split_type
== var
->type
);
524 /* If we're not modifying this variable, delete the info so we skip
525 * it faster in later passes.
527 _mesa_hash_table_remove_key(var_info_map
, var
);
531 nir_foreach_variable(var
, &split_vars
) {
532 struct array_var_info
*info
= get_array_var_info(var
, var_info_map
);
533 create_split_array_vars(info
, 0, &info
->root_split
, var
->name
,
534 shader
, impl
, mem_ctx
);
537 return !exec_list_is_empty(&split_vars
);
541 deref_has_split_wildcard(nir_deref_path
*path
,
542 struct array_var_info
*info
)
547 assert(path
->path
[0]->var
== info
->base_var
);
548 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
549 if (path
->path
[i
+ 1]->deref_type
== nir_deref_type_array_wildcard
&&
550 info
->levels
[i
].split
)
558 array_path_is_out_of_bounds(nir_deref_path
*path
,
559 struct array_var_info
*info
)
564 assert(path
->path
[0]->var
== info
->base_var
);
565 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
566 nir_deref_instr
*p
= path
->path
[i
+ 1];
567 if (p
->deref_type
== nir_deref_type_array_wildcard
)
570 if (nir_src_is_const(p
->arr
.index
) &&
571 nir_src_as_uint(p
->arr
.index
) >= info
->levels
[i
].array_len
)
579 emit_split_copies(nir_builder
*b
,
580 struct array_var_info
*dst_info
, nir_deref_path
*dst_path
,
581 unsigned dst_level
, nir_deref_instr
*dst
,
582 struct array_var_info
*src_info
, nir_deref_path
*src_path
,
583 unsigned src_level
, nir_deref_instr
*src
)
585 nir_deref_instr
*dst_p
, *src_p
;
587 while ((dst_p
= dst_path
->path
[dst_level
+ 1])) {
588 if (dst_p
->deref_type
== nir_deref_type_array_wildcard
)
591 dst
= nir_build_deref_follower(b
, dst
, dst_p
);
595 while ((src_p
= src_path
->path
[src_level
+ 1])) {
596 if (src_p
->deref_type
== nir_deref_type_array_wildcard
)
599 src
= nir_build_deref_follower(b
, src
, src_p
);
603 if (src_p
== NULL
|| dst_p
== NULL
) {
604 assert(src_p
== NULL
&& dst_p
== NULL
);
605 nir_copy_deref(b
, dst
, src
);
607 assert(dst_p
->deref_type
== nir_deref_type_array_wildcard
&&
608 src_p
->deref_type
== nir_deref_type_array_wildcard
);
610 if ((dst_info
&& dst_info
->levels
[dst_level
].split
) ||
611 (src_info
&& src_info
->levels
[src_level
].split
)) {
612 /* There are no indirects at this level on one of the source or the
613 * destination so we are lowering it.
615 assert(glsl_get_length(dst_path
->path
[dst_level
]->type
) ==
616 glsl_get_length(src_path
->path
[src_level
]->type
));
617 unsigned len
= glsl_get_length(dst_path
->path
[dst_level
]->type
);
618 for (unsigned i
= 0; i
< len
; i
++) {
619 emit_split_copies(b
, dst_info
, dst_path
, dst_level
+ 1,
620 nir_build_deref_array_imm(b
, dst
, i
),
621 src_info
, src_path
, src_level
+ 1,
622 nir_build_deref_array_imm(b
, src
, i
));
625 /* Neither side is being split so we just keep going */
626 emit_split_copies(b
, dst_info
, dst_path
, dst_level
+ 1,
627 nir_build_deref_array_wildcard(b
, dst
),
628 src_info
, src_path
, src_level
+ 1,
629 nir_build_deref_array_wildcard(b
, src
));
635 split_array_copies_impl(nir_function_impl
*impl
,
636 struct hash_table
*var_info_map
,
637 nir_variable_mode modes
,
641 nir_builder_init(&b
, impl
);
643 nir_foreach_block(block
, impl
) {
644 nir_foreach_instr_safe(instr
, block
) {
645 if (instr
->type
!= nir_instr_type_intrinsic
)
648 nir_intrinsic_instr
*copy
= nir_instr_as_intrinsic(instr
);
649 if (copy
->intrinsic
!= nir_intrinsic_copy_deref
)
652 nir_deref_instr
*dst_deref
= nir_src_as_deref(copy
->src
[0]);
653 nir_deref_instr
*src_deref
= nir_src_as_deref(copy
->src
[1]);
655 struct array_var_info
*dst_info
=
656 get_array_deref_info(dst_deref
, var_info_map
, modes
);
657 struct array_var_info
*src_info
=
658 get_array_deref_info(src_deref
, var_info_map
, modes
);
660 if (!src_info
&& !dst_info
)
663 nir_deref_path dst_path
, src_path
;
664 nir_deref_path_init(&dst_path
, dst_deref
, mem_ctx
);
665 nir_deref_path_init(&src_path
, src_deref
, mem_ctx
);
667 if (!deref_has_split_wildcard(&dst_path
, dst_info
) &&
668 !deref_has_split_wildcard(&src_path
, src_info
))
671 b
.cursor
= nir_instr_remove(©
->instr
);
673 emit_split_copies(&b
, dst_info
, &dst_path
, 0, dst_path
.path
[0],
674 src_info
, &src_path
, 0, src_path
.path
[0]);
680 split_array_access_impl(nir_function_impl
*impl
,
681 struct hash_table
*var_info_map
,
682 nir_variable_mode modes
,
686 nir_builder_init(&b
, impl
);
688 nir_foreach_block(block
, impl
) {
689 nir_foreach_instr_safe(instr
, block
) {
690 if (instr
->type
== nir_instr_type_deref
) {
691 /* Clean up any dead derefs we find lying around. They may refer
692 * to variables we're planning to split.
694 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
695 if (deref
->mode
& modes
)
696 nir_deref_instr_remove_if_unused(deref
);
700 if (instr
->type
!= nir_instr_type_intrinsic
)
703 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
704 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
&&
705 intrin
->intrinsic
!= nir_intrinsic_store_deref
&&
706 intrin
->intrinsic
!= nir_intrinsic_copy_deref
)
709 const unsigned num_derefs
=
710 intrin
->intrinsic
== nir_intrinsic_copy_deref
? 2 : 1;
712 for (unsigned d
= 0; d
< num_derefs
; d
++) {
713 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[d
]);
715 struct array_var_info
*info
=
716 get_array_deref_info(deref
, var_info_map
, modes
);
721 nir_deref_path_init(&path
, deref
, mem_ctx
);
723 b
.cursor
= nir_before_instr(&intrin
->instr
);
725 if (array_path_is_out_of_bounds(&path
, info
)) {
726 /* If one of the derefs is out-of-bounds, we just delete the
727 * instruction. If a destination is out of bounds, then it may
728 * have been in-bounds prior to shrinking so we don't want to
729 * accidentally stomp something. However, we've already proven
730 * that it will never be read so it's safe to delete. If a
731 * source is out of bounds then it is loading random garbage.
732 * For loads, we replace their uses with an undef instruction
733 * and for copies we just delete the copy since it was writing
734 * undefined garbage anyway and we may as well leave the random
735 * garbage in the destination alone.
737 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
739 nir_ssa_undef(&b
, intrin
->dest
.ssa
.num_components
,
740 intrin
->dest
.ssa
.bit_size
);
741 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
744 nir_instr_remove(&intrin
->instr
);
745 for (unsigned i
= 0; i
< num_derefs
; i
++)
746 nir_deref_instr_remove_if_unused(nir_src_as_deref(intrin
->src
[i
]));
750 struct array_split
*split
= &info
->root_split
;
751 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
752 if (info
->levels
[i
].split
) {
753 nir_deref_instr
*p
= path
.path
[i
+ 1];
754 unsigned index
= nir_src_as_uint(p
->arr
.index
);
755 assert(index
< info
->levels
[i
].array_len
);
756 split
= &split
->splits
[index
];
759 assert(!split
->splits
&& split
->var
);
761 nir_deref_instr
*new_deref
= nir_build_deref_var(&b
, split
->var
);
762 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
763 if (!info
->levels
[i
].split
) {
764 new_deref
= nir_build_deref_follower(&b
, new_deref
,
768 assert(new_deref
->type
== deref
->type
);
770 /* Rewrite the deref source to point to the split one */
771 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[d
],
772 nir_src_for_ssa(&new_deref
->dest
.ssa
));
773 nir_deref_instr_remove_if_unused(deref
);
779 /** A pass for splitting arrays of vectors into multiple variables
781 * This pass looks at arrays (possibly multiple levels) of vectors (not
782 * structures or other types) and tries to split them into piles of variables,
783 * one for each array element. The heuristic used is simple: If a given array
784 * level is never used with an indirect, that array level will get split.
786 * This pass probably could handles structures easily enough but making a pass
787 * that could see through an array of structures of arrays would be difficult
788 * so it's best to just run nir_split_struct_vars first.
791 nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
)
793 void *mem_ctx
= ralloc_context(NULL
);
794 struct hash_table
*var_info_map
= _mesa_pointer_hash_table_create(mem_ctx
);
796 assert((modes
& (nir_var_shader_temp
| nir_var_function_temp
)) == modes
);
798 bool has_global_array
= false;
799 if (modes
& nir_var_shader_temp
) {
800 has_global_array
= init_var_list_array_infos(&shader
->globals
,
801 var_info_map
, mem_ctx
);
804 bool has_any_array
= false;
805 nir_foreach_function(function
, shader
) {
809 bool has_local_array
= false;
810 if (modes
& nir_var_function_temp
) {
811 has_local_array
= init_var_list_array_infos(&function
->impl
->locals
,
812 var_info_map
, mem_ctx
);
815 if (has_global_array
|| has_local_array
) {
816 has_any_array
= true;
817 mark_array_usage_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
821 /* If we failed to find any arrays of arrays, bail early. */
822 if (!has_any_array
) {
823 ralloc_free(mem_ctx
);
827 bool has_global_splits
= false;
828 if (modes
& nir_var_shader_temp
) {
829 has_global_splits
= split_var_list_arrays(shader
, NULL
,
831 var_info_map
, mem_ctx
);
834 bool progress
= false;
835 nir_foreach_function(function
, shader
) {
839 bool has_local_splits
= false;
840 if (modes
& nir_var_function_temp
) {
841 has_local_splits
= split_var_list_arrays(shader
, function
->impl
,
842 &function
->impl
->locals
,
843 var_info_map
, mem_ctx
);
846 if (has_global_splits
|| has_local_splits
) {
847 split_array_copies_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
848 split_array_access_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
850 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
851 nir_metadata_dominance
);
856 ralloc_free(mem_ctx
);
861 struct array_level_usage
{
864 /* The value UINT_MAX will be used to indicate an indirect */
866 unsigned max_written
;
868 /* True if there is a copy that isn't to/from a shrinkable array */
869 bool has_external_copy
;
870 struct set
*levels_copied
;
873 struct vec_var_usage
{
874 /* Convenience set of all components this variable has */
875 nir_component_mask_t all_comps
;
877 nir_component_mask_t comps_read
;
878 nir_component_mask_t comps_written
;
880 nir_component_mask_t comps_kept
;
882 /* True if there is a copy that isn't to/from a shrinkable vector */
883 bool has_external_copy
;
884 struct set
*vars_copied
;
887 struct array_level_usage levels
[0];
890 static struct vec_var_usage
*
891 get_vec_var_usage(nir_variable
*var
,
892 struct hash_table
*var_usage_map
,
893 bool add_usage_entry
, void *mem_ctx
)
895 struct hash_entry
*entry
= _mesa_hash_table_search(var_usage_map
, var
);
899 if (!add_usage_entry
)
902 /* Check to make sure that we are working with an array of vectors. We
903 * don't bother to shrink single vectors because we figure that we can
904 * clean it up better with SSA than by inserting piles of vecN instructions
905 * to compact results.
907 int num_levels
= num_array_levels_in_array_of_vector_type(var
->type
);
909 return NULL
; /* Not an array of vectors */
911 struct vec_var_usage
*usage
=
912 rzalloc_size(mem_ctx
, sizeof(*usage
) +
913 num_levels
* sizeof(usage
->levels
[0]));
915 usage
->num_levels
= num_levels
;
916 const struct glsl_type
*type
= var
->type
;
917 for (unsigned i
= 0; i
< num_levels
; i
++) {
918 usage
->levels
[i
].array_len
= glsl_get_length(type
);
919 assert(glsl_get_explicit_stride(type
) == 0);
920 type
= glsl_get_array_element(type
);
922 assert(glsl_type_is_vector_or_scalar(type
));
924 usage
->all_comps
= (1 << glsl_get_components(type
)) - 1;
926 _mesa_hash_table_insert(var_usage_map
, var
, usage
);
931 static struct vec_var_usage
*
932 get_vec_deref_usage(nir_deref_instr
*deref
,
933 struct hash_table
*var_usage_map
,
934 nir_variable_mode modes
,
935 bool add_usage_entry
, void *mem_ctx
)
937 if (!(deref
->mode
& modes
))
940 return get_vec_var_usage(nir_deref_instr_get_variable(deref
),
941 var_usage_map
, add_usage_entry
, mem_ctx
);
945 mark_deref_used(nir_deref_instr
*deref
,
946 nir_component_mask_t comps_read
,
947 nir_component_mask_t comps_written
,
948 nir_deref_instr
*copy_deref
,
949 struct hash_table
*var_usage_map
,
950 nir_variable_mode modes
,
953 if (!(deref
->mode
& modes
))
956 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
958 struct vec_var_usage
*usage
=
959 get_vec_var_usage(var
, var_usage_map
, true, mem_ctx
);
963 usage
->comps_read
|= comps_read
& usage
->all_comps
;
964 usage
->comps_written
|= comps_written
& usage
->all_comps
;
966 struct vec_var_usage
*copy_usage
= NULL
;
968 copy_usage
= get_vec_deref_usage(copy_deref
, var_usage_map
, modes
,
971 if (usage
->vars_copied
== NULL
) {
972 usage
->vars_copied
= _mesa_pointer_set_create(mem_ctx
);
974 _mesa_set_add(usage
->vars_copied
, copy_usage
);
976 usage
->has_external_copy
= true;
981 nir_deref_path_init(&path
, deref
, mem_ctx
);
983 nir_deref_path copy_path
;
985 nir_deref_path_init(©_path
, copy_deref
, mem_ctx
);
988 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
989 struct array_level_usage
*level
= &usage
->levels
[i
];
990 nir_deref_instr
*deref
= path
.path
[i
+ 1];
991 assert(deref
->deref_type
== nir_deref_type_array
||
992 deref
->deref_type
== nir_deref_type_array_wildcard
);
995 if (deref
->deref_type
== nir_deref_type_array
) {
996 max_used
= nir_src_is_const(deref
->arr
.index
) ?
997 nir_src_as_uint(deref
->arr
.index
) : UINT_MAX
;
999 /* For wildcards, we read or wrote the whole thing. */
1000 assert(deref
->deref_type
== nir_deref_type_array_wildcard
);
1001 max_used
= level
->array_len
- 1;
1004 /* Match each wildcard level with the level on copy_usage */
1005 for (; copy_path
.path
[copy_i
+ 1]; copy_i
++) {
1006 if (copy_path
.path
[copy_i
+ 1]->deref_type
==
1007 nir_deref_type_array_wildcard
)
1010 struct array_level_usage
*copy_level
=
1011 ©_usage
->levels
[copy_i
++];
1013 if (level
->levels_copied
== NULL
) {
1014 level
->levels_copied
= _mesa_pointer_set_create(mem_ctx
);
1016 _mesa_set_add(level
->levels_copied
, copy_level
);
1018 /* We have a wildcard and it comes from a variable we aren't
1019 * tracking; flag it and we'll know to not shorten this array.
1021 level
->has_external_copy
= true;
1026 level
->max_written
= MAX2(level
->max_written
, max_used
);
1028 level
->max_read
= MAX2(level
->max_read
, max_used
);
1033 src_is_load_deref(nir_src src
, nir_src deref_src
)
1036 assert(deref_src
.is_ssa
);
1038 if (src
.ssa
->parent_instr
->type
!= nir_instr_type_intrinsic
)
1041 nir_intrinsic_instr
*load
= nir_instr_as_intrinsic(src
.ssa
->parent_instr
);
1042 if (load
->intrinsic
!= nir_intrinsic_load_deref
)
1045 assert(load
->src
[0].is_ssa
);
1047 return load
->src
[0].ssa
== deref_src
.ssa
;
1050 /* Returns all non-self-referential components of a store instruction. A
1051 * component is self-referential if it comes from the same component of a load
1052 * instruction on the same deref. If the only data in a particular component
1053 * of a variable came directly from that component then it's undefined. The
1054 * only way to get defined data into a component of a variable is for it to
1055 * get written there by something outside or from a different component.
1057 * This is a fairly common pattern in shaders that come from either GLSL IR or
1058 * GLSLang because both glsl_to_nir and GLSLang implement write-masking with
1061 static nir_component_mask_t
1062 get_non_self_referential_store_comps(nir_intrinsic_instr
*store
)
1064 nir_component_mask_t comps
= nir_intrinsic_write_mask(store
);
1066 assert(store
->src
[1].is_ssa
);
1067 nir_instr
*src_instr
= store
->src
[1].ssa
->parent_instr
;
1068 if (src_instr
->type
!= nir_instr_type_alu
)
1071 nir_alu_instr
*src_alu
= nir_instr_as_alu(src_instr
);
1073 if (src_alu
->op
== nir_op_imov
||
1074 src_alu
->op
== nir_op_fmov
) {
1075 /* If it's just a swizzle of a load from the same deref, discount any
1076 * channels that don't move in the swizzle.
1078 if (src_is_load_deref(src_alu
->src
[0].src
, store
->src
[0])) {
1079 for (unsigned i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; i
++) {
1080 if (src_alu
->src
[0].swizzle
[i
] == i
)
1081 comps
&= ~(1u << i
);
1084 } else if (src_alu
->op
== nir_op_vec2
||
1085 src_alu
->op
== nir_op_vec3
||
1086 src_alu
->op
== nir_op_vec4
) {
1087 /* If it's a vec, discount any channels that are just loads from the
1088 * same deref put in the same spot.
1090 for (unsigned i
= 0; i
< nir_op_infos
[src_alu
->op
].num_inputs
; i
++) {
1091 if (src_is_load_deref(src_alu
->src
[i
].src
, store
->src
[0]) &&
1092 src_alu
->src
[i
].swizzle
[0] == i
)
1093 comps
&= ~(1u << i
);
1101 find_used_components_impl(nir_function_impl
*impl
,
1102 struct hash_table
*var_usage_map
,
1103 nir_variable_mode modes
,
1106 nir_foreach_block(block
, impl
) {
1107 nir_foreach_instr(instr
, block
) {
1108 if (instr
->type
!= nir_instr_type_intrinsic
)
1111 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1112 switch (intrin
->intrinsic
) {
1113 case nir_intrinsic_load_deref
:
1114 mark_deref_used(nir_src_as_deref(intrin
->src
[0]),
1115 nir_ssa_def_components_read(&intrin
->dest
.ssa
), 0,
1116 NULL
, var_usage_map
, modes
, mem_ctx
);
1119 case nir_intrinsic_store_deref
:
1120 mark_deref_used(nir_src_as_deref(intrin
->src
[0]),
1121 0, get_non_self_referential_store_comps(intrin
),
1122 NULL
, var_usage_map
, modes
, mem_ctx
);
1125 case nir_intrinsic_copy_deref
: {
1126 /* Just mark everything used for copies. */
1127 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
1128 nir_deref_instr
*src
= nir_src_as_deref(intrin
->src
[1]);
1129 mark_deref_used(dst
, 0, ~0, src
, var_usage_map
, modes
, mem_ctx
);
1130 mark_deref_used(src
, ~0, 0, dst
, var_usage_map
, modes
, mem_ctx
);
1142 shrink_vec_var_list(struct exec_list
*vars
,
1143 struct hash_table
*var_usage_map
)
1145 /* Initialize the components kept field of each variable. This is the
1146 * AND of the components written and components read. If a component is
1147 * written but never read, it's dead. If it is read but never written,
1148 * then all values read are undefined garbage and we may as well not read
1151 * The same logic applies to the array length. We make the array length
1152 * the minimum needed required length between read and write and plan to
1153 * discard any OOB access. The one exception here is indirect writes
1154 * because we don't know where they will land and we can't shrink an array
1155 * with indirect writes because previously in-bounds writes may become
1156 * out-of-bounds and have undefined behavior.
1158 * Also, if we have a copy that to/from something we can't shrink, we need
1159 * to leave components and array_len of any wildcards alone.
1161 nir_foreach_variable(var
, vars
) {
1162 struct vec_var_usage
*usage
=
1163 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1167 assert(usage
->comps_kept
== 0);
1168 if (usage
->has_external_copy
)
1169 usage
->comps_kept
= usage
->all_comps
;
1171 usage
->comps_kept
= usage
->comps_read
& usage
->comps_written
;
1173 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1174 struct array_level_usage
*level
= &usage
->levels
[i
];
1175 assert(level
->array_len
> 0);
1177 if (level
->max_written
== UINT_MAX
|| level
->has_external_copy
)
1178 continue; /* Can't shrink */
1180 unsigned max_used
= MIN2(level
->max_read
, level
->max_written
);
1181 level
->array_len
= MIN2(max_used
, level
->array_len
- 1) + 1;
1185 /* In order for variable copies to work, we have to have the same data type
1186 * on the source and the destination. In order to satisfy this, we run a
1187 * little fixed-point algorithm to transitively ensure that we get enough
1188 * components and array elements for this to hold for all copies.
1192 fp_progress
= false;
1193 nir_foreach_variable(var
, vars
) {
1194 struct vec_var_usage
*var_usage
=
1195 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1196 if (!var_usage
|| !var_usage
->vars_copied
)
1199 set_foreach(var_usage
->vars_copied
, copy_entry
) {
1200 struct vec_var_usage
*copy_usage
= (void *)copy_entry
->key
;
1201 if (copy_usage
->comps_kept
!= var_usage
->comps_kept
) {
1202 nir_component_mask_t comps_kept
=
1203 (var_usage
->comps_kept
| copy_usage
->comps_kept
);
1204 var_usage
->comps_kept
= comps_kept
;
1205 copy_usage
->comps_kept
= comps_kept
;
1210 for (unsigned i
= 0; i
< var_usage
->num_levels
; i
++) {
1211 struct array_level_usage
*var_level
= &var_usage
->levels
[i
];
1212 if (!var_level
->levels_copied
)
1215 set_foreach(var_level
->levels_copied
, copy_entry
) {
1216 struct array_level_usage
*copy_level
= (void *)copy_entry
->key
;
1217 if (var_level
->array_len
!= copy_level
->array_len
) {
1218 unsigned array_len
=
1219 MAX2(var_level
->array_len
, copy_level
->array_len
);
1220 var_level
->array_len
= array_len
;
1221 copy_level
->array_len
= array_len
;
1227 } while (fp_progress
);
1229 bool vars_shrunk
= false;
1230 nir_foreach_variable_safe(var
, vars
) {
1231 struct vec_var_usage
*usage
=
1232 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1236 bool shrunk
= false;
1237 const struct glsl_type
*vec_type
= var
->type
;
1238 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1239 /* If we've reduced the array to zero elements at some level, just
1240 * set comps_kept to 0 and delete the variable.
1242 if (usage
->levels
[i
].array_len
== 0) {
1243 usage
->comps_kept
= 0;
1247 assert(usage
->levels
[i
].array_len
<= glsl_get_length(vec_type
));
1248 if (usage
->levels
[i
].array_len
< glsl_get_length(vec_type
))
1250 vec_type
= glsl_get_array_element(vec_type
);
1252 assert(glsl_type_is_vector_or_scalar(vec_type
));
1254 assert(usage
->comps_kept
== (usage
->comps_kept
& usage
->all_comps
));
1255 if (usage
->comps_kept
!= usage
->all_comps
)
1258 if (usage
->comps_kept
== 0) {
1259 /* This variable is dead, remove it */
1261 exec_node_remove(&var
->node
);
1266 /* This variable doesn't need to be shrunk. Remove it from the
1267 * hash table so later steps will ignore it.
1269 _mesa_hash_table_remove_key(var_usage_map
, var
);
1273 /* Build the new var type */
1274 unsigned new_num_comps
= util_bitcount(usage
->comps_kept
);
1275 const struct glsl_type
*new_type
=
1276 glsl_vector_type(glsl_get_base_type(vec_type
), new_num_comps
);
1277 for (int i
= usage
->num_levels
- 1; i
>= 0; i
--) {
1278 assert(usage
->levels
[i
].array_len
> 0);
1279 /* If the original type was a matrix type, we'd like to keep that so
1280 * we don't convert matrices into arrays.
1282 if (i
== usage
->num_levels
- 1 &&
1283 glsl_type_is_matrix(glsl_without_array(var
->type
)) &&
1284 new_num_comps
> 1 && usage
->levels
[i
].array_len
> 1) {
1285 new_type
= glsl_matrix_type(glsl_get_base_type(new_type
),
1287 usage
->levels
[i
].array_len
);
1289 new_type
= glsl_array_type(new_type
, usage
->levels
[i
].array_len
, 0);
1292 var
->type
= new_type
;
1301 vec_deref_is_oob(nir_deref_instr
*deref
,
1302 struct vec_var_usage
*usage
)
1304 nir_deref_path path
;
1305 nir_deref_path_init(&path
, deref
, NULL
);
1308 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1309 nir_deref_instr
*p
= path
.path
[i
+ 1];
1310 if (p
->deref_type
== nir_deref_type_array_wildcard
)
1313 if (nir_src_is_const(p
->arr
.index
) &&
1314 nir_src_as_uint(p
->arr
.index
) >= usage
->levels
[i
].array_len
) {
1320 nir_deref_path_finish(&path
);
1326 vec_deref_is_dead_or_oob(nir_deref_instr
*deref
,
1327 struct hash_table
*var_usage_map
,
1328 nir_variable_mode modes
)
1330 struct vec_var_usage
*usage
=
1331 get_vec_deref_usage(deref
, var_usage_map
, modes
, false, NULL
);
1335 return usage
->comps_kept
== 0 || vec_deref_is_oob(deref
, usage
);
1339 shrink_vec_var_access_impl(nir_function_impl
*impl
,
1340 struct hash_table
*var_usage_map
,
1341 nir_variable_mode modes
)
1344 nir_builder_init(&b
, impl
);
1346 nir_foreach_block(block
, impl
) {
1347 nir_foreach_instr_safe(instr
, block
) {
1348 switch (instr
->type
) {
1349 case nir_instr_type_deref
: {
1350 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
1351 if (!(deref
->mode
& modes
))
1354 /* Clean up any dead derefs we find lying around. They may refer
1355 * to variables we've deleted.
1357 if (nir_deref_instr_remove_if_unused(deref
))
1360 /* Update the type in the deref to keep the types consistent as
1361 * you walk down the chain. We don't need to check if this is one
1362 * of the derefs we're shrinking because this is a no-op if it
1363 * isn't. The worst that could happen is that we accidentally fix
1366 if (deref
->deref_type
== nir_deref_type_var
) {
1367 deref
->type
= deref
->var
->type
;
1368 } else if (deref
->deref_type
== nir_deref_type_array
||
1369 deref
->deref_type
== nir_deref_type_array_wildcard
) {
1370 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
1371 assert(glsl_type_is_array(parent
->type
) ||
1372 glsl_type_is_matrix(parent
->type
));
1373 deref
->type
= glsl_get_array_element(parent
->type
);
1378 case nir_instr_type_intrinsic
: {
1379 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1381 /* If we have a copy whose source or destination has been deleted
1382 * because we determined the variable was dead, then we just
1383 * delete the copy instruction. If the source variable was dead
1384 * then it was writing undefined garbage anyway and if it's the
1385 * destination variable that's dead then the write isn't needed.
1387 if (intrin
->intrinsic
== nir_intrinsic_copy_deref
) {
1388 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
1389 nir_deref_instr
*src
= nir_src_as_deref(intrin
->src
[1]);
1390 if (vec_deref_is_dead_or_oob(dst
, var_usage_map
, modes
) ||
1391 vec_deref_is_dead_or_oob(src
, var_usage_map
, modes
)) {
1392 nir_instr_remove(&intrin
->instr
);
1393 nir_deref_instr_remove_if_unused(dst
);
1394 nir_deref_instr_remove_if_unused(src
);
1399 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
&&
1400 intrin
->intrinsic
!= nir_intrinsic_store_deref
)
1403 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1404 if (!(deref
->mode
& modes
))
1407 struct vec_var_usage
*usage
=
1408 get_vec_deref_usage(deref
, var_usage_map
, modes
, false, NULL
);
1412 if (usage
->comps_kept
== 0 || vec_deref_is_oob(deref
, usage
)) {
1413 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
1415 nir_ssa_undef(&b
, intrin
->dest
.ssa
.num_components
,
1416 intrin
->dest
.ssa
.bit_size
);
1417 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
1418 nir_src_for_ssa(u
));
1420 nir_instr_remove(&intrin
->instr
);
1421 nir_deref_instr_remove_if_unused(deref
);
1425 /* If we're not dropping any components, there's no need to
1428 if (usage
->comps_kept
== usage
->all_comps
)
1431 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
1432 b
.cursor
= nir_after_instr(&intrin
->instr
);
1434 nir_ssa_def
*undef
=
1435 nir_ssa_undef(&b
, 1, intrin
->dest
.ssa
.bit_size
);
1436 nir_ssa_def
*vec_srcs
[NIR_MAX_VEC_COMPONENTS
];
1438 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1439 if (usage
->comps_kept
& (1u << i
))
1440 vec_srcs
[i
] = nir_channel(&b
, &intrin
->dest
.ssa
, c
++);
1442 vec_srcs
[i
] = undef
;
1444 nir_ssa_def
*vec
= nir_vec(&b
, vec_srcs
, intrin
->num_components
);
1446 nir_ssa_def_rewrite_uses_after(&intrin
->dest
.ssa
,
1447 nir_src_for_ssa(vec
),
1450 /* The SSA def is now only used by the swizzle. It's safe to
1451 * shrink the number of components.
1453 assert(list_length(&intrin
->dest
.ssa
.uses
) == c
);
1454 intrin
->num_components
= c
;
1455 intrin
->dest
.ssa
.num_components
= c
;
1457 nir_component_mask_t write_mask
=
1458 nir_intrinsic_write_mask(intrin
);
1460 unsigned swizzle
[NIR_MAX_VEC_COMPONENTS
];
1461 nir_component_mask_t new_write_mask
= 0;
1463 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1464 if (usage
->comps_kept
& (1u << i
)) {
1466 if (write_mask
& (1u << i
))
1467 new_write_mask
|= 1u << c
;
1472 b
.cursor
= nir_before_instr(&intrin
->instr
);
1474 nir_ssa_def
*swizzled
=
1475 nir_swizzle(&b
, intrin
->src
[1].ssa
, swizzle
, c
, false);
1477 /* Rewrite to use the compacted source */
1478 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[1],
1479 nir_src_for_ssa(swizzled
));
1480 nir_intrinsic_set_write_mask(intrin
, new_write_mask
);
1481 intrin
->num_components
= c
;
1494 function_impl_has_vars_with_modes(nir_function_impl
*impl
,
1495 nir_variable_mode modes
)
1497 nir_shader
*shader
= impl
->function
->shader
;
1499 if ((modes
& nir_var_shader_temp
) && !exec_list_is_empty(&shader
->globals
))
1502 if ((modes
& nir_var_function_temp
) && !exec_list_is_empty(&impl
->locals
))
1508 /** Attempt to shrink arrays of vectors
1510 * This pass looks at variables which contain a vector or an array (possibly
1511 * multiple dimensions) of vectors and attempts to lower to a smaller vector
1512 * or array. If the pass can prove that a component of a vector (or array of
1513 * vectors) is never really used, then that component will be removed.
1514 * Similarly, the pass attempts to shorten arrays based on what elements it
1515 * can prove are never read or never contain valid data.
1518 nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
)
1520 assert((modes
& (nir_var_shader_temp
| nir_var_function_temp
)) == modes
);
1522 void *mem_ctx
= ralloc_context(NULL
);
1524 struct hash_table
*var_usage_map
=
1525 _mesa_pointer_hash_table_create(mem_ctx
);
1527 bool has_vars_to_shrink
= false;
1528 nir_foreach_function(function
, shader
) {
1529 if (!function
->impl
)
1532 /* Don't even bother crawling the IR if we don't have any variables.
1533 * Given that this pass deletes any unused variables, it's likely that
1534 * we will be in this scenario eventually.
1536 if (function_impl_has_vars_with_modes(function
->impl
, modes
)) {
1537 has_vars_to_shrink
= true;
1538 find_used_components_impl(function
->impl
, var_usage_map
,
1542 if (!has_vars_to_shrink
) {
1543 ralloc_free(mem_ctx
);
1547 bool globals_shrunk
= false;
1548 if (modes
& nir_var_shader_temp
)
1549 globals_shrunk
= shrink_vec_var_list(&shader
->globals
, var_usage_map
);
1551 bool progress
= false;
1552 nir_foreach_function(function
, shader
) {
1553 if (!function
->impl
)
1556 bool locals_shrunk
= false;
1557 if (modes
& nir_var_function_temp
) {
1558 locals_shrunk
= shrink_vec_var_list(&function
->impl
->locals
,
1562 if (globals_shrunk
|| locals_shrunk
) {
1563 shrink_vec_var_access_impl(function
->impl
, var_usage_map
, modes
);
1565 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
1566 nir_metadata_dominance
);
1571 ralloc_free(mem_ctx
);