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"
30 #include "util/u_math.h"
33 get_complex_used_vars(nir_shader
*shader
, void *mem_ctx
)
35 struct set
*complex_vars
= _mesa_pointer_set_create(mem_ctx
);
37 nir_foreach_function(function
, shader
) {
41 nir_foreach_block(block
, function
->impl
) {
42 nir_foreach_instr(instr
, block
) {
43 if (instr
->type
!= nir_instr_type_deref
)
46 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
48 /* We only need to consider var derefs because
49 * nir_deref_instr_has_complex_use is recursive.
51 if (deref
->deref_type
== nir_deref_type_var
&&
52 nir_deref_instr_has_complex_use(deref
))
53 _mesa_set_add(complex_vars
, deref
->var
);
61 struct split_var_state
{
65 nir_function_impl
*impl
;
67 nir_variable
*base_var
;
73 const struct glsl_type
*type
;
81 static const struct glsl_type
*
82 wrap_type_in_array(const struct glsl_type
*type
,
83 const struct glsl_type
*array_type
)
85 if (!glsl_type_is_array(array_type
))
88 const struct glsl_type
*elem_type
=
89 wrap_type_in_array(type
, glsl_get_array_element(array_type
));
90 assert(glsl_get_explicit_stride(array_type
) == 0);
91 return glsl_array_type(elem_type
, glsl_get_length(array_type
), 0);
95 num_array_levels_in_array_of_vector_type(const struct glsl_type
*type
)
99 if (glsl_type_is_array_or_matrix(type
)) {
101 type
= glsl_get_array_element(type
);
102 } else if (glsl_type_is_vector_or_scalar(type
)) {
105 /* Not an array of vectors */
112 init_field_for_type(struct field
*field
, struct field
*parent
,
113 const struct glsl_type
*type
,
115 struct split_var_state
*state
)
117 *field
= (struct field
) {
122 const struct glsl_type
*struct_type
= glsl_without_array(type
);
123 if (glsl_type_is_struct_or_ifc(struct_type
)) {
124 field
->num_fields
= glsl_get_length(struct_type
),
125 field
->fields
= ralloc_array(state
->mem_ctx
, struct field
,
127 for (unsigned i
= 0; i
< field
->num_fields
; i
++) {
128 char *field_name
= NULL
;
130 field_name
= ralloc_asprintf(state
->mem_ctx
, "%s_%s", name
,
131 glsl_get_struct_elem_name(struct_type
, i
));
133 field_name
= ralloc_asprintf(state
->mem_ctx
, "{unnamed %s}_%s",
134 glsl_get_type_name(struct_type
),
135 glsl_get_struct_elem_name(struct_type
, i
));
137 init_field_for_type(&field
->fields
[i
], field
,
138 glsl_get_struct_field(struct_type
, i
),
142 const struct glsl_type
*var_type
= type
;
143 for (struct field
*f
= field
->parent
; f
; f
= f
->parent
)
144 var_type
= wrap_type_in_array(var_type
, f
->type
);
146 nir_variable_mode mode
= state
->base_var
->data
.mode
;
147 if (mode
== nir_var_function_temp
) {
148 field
->var
= nir_local_variable_create(state
->impl
, var_type
, name
);
150 field
->var
= nir_variable_create(state
->shader
, mode
, var_type
, name
);
156 split_var_list_structs(nir_shader
*shader
,
157 nir_function_impl
*impl
,
158 struct exec_list
*vars
,
159 struct hash_table
*var_field_map
,
160 struct set
**complex_vars
,
163 struct split_var_state state
= {
169 struct exec_list split_vars
;
170 exec_list_make_empty(&split_vars
);
172 /* To avoid list confusion (we'll be adding things as we split variables),
173 * pull all of the variables we plan to split off of the list
175 nir_foreach_variable_safe(var
, vars
) {
176 if (!glsl_type_is_struct_or_ifc(glsl_without_array(var
->type
)))
179 if (*complex_vars
== NULL
)
180 *complex_vars
= get_complex_used_vars(shader
, mem_ctx
);
182 /* We can't split a variable that's referenced with deref that has any
183 * sort of complex usage.
185 if (_mesa_set_search(*complex_vars
, var
))
188 exec_node_remove(&var
->node
);
189 exec_list_push_tail(&split_vars
, &var
->node
);
192 nir_foreach_variable(var
, &split_vars
) {
193 state
.base_var
= var
;
195 struct field
*root_field
= ralloc(mem_ctx
, struct field
);
196 init_field_for_type(root_field
, NULL
, var
->type
, var
->name
, &state
);
197 _mesa_hash_table_insert(var_field_map
, var
, root_field
);
200 return !exec_list_is_empty(&split_vars
);
204 split_struct_derefs_impl(nir_function_impl
*impl
,
205 struct hash_table
*var_field_map
,
206 nir_variable_mode modes
,
210 nir_builder_init(&b
, impl
);
212 nir_foreach_block(block
, impl
) {
213 nir_foreach_instr_safe(instr
, block
) {
214 if (instr
->type
!= nir_instr_type_deref
)
217 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
218 if (!(deref
->mode
& modes
))
221 /* Clean up any dead derefs we find lying around. They may refer to
222 * variables we're planning to split.
224 if (nir_deref_instr_remove_if_unused(deref
))
227 if (!glsl_type_is_vector_or_scalar(deref
->type
))
230 nir_variable
*base_var
= nir_deref_instr_get_variable(deref
);
231 struct hash_entry
*entry
=
232 _mesa_hash_table_search(var_field_map
, base_var
);
236 struct field
*root_field
= entry
->data
;
239 nir_deref_path_init(&path
, deref
, mem_ctx
);
241 struct field
*tail_field
= root_field
;
242 for (unsigned i
= 0; path
.path
[i
]; i
++) {
243 if (path
.path
[i
]->deref_type
!= nir_deref_type_struct
)
247 assert(glsl_type_is_struct_or_ifc(path
.path
[i
- 1]->type
));
248 assert(path
.path
[i
- 1]->type
==
249 glsl_without_array(tail_field
->type
));
251 tail_field
= &tail_field
->fields
[path
.path
[i
]->strct
.index
];
253 nir_variable
*split_var
= tail_field
->var
;
255 nir_deref_instr
*new_deref
= NULL
;
256 for (unsigned i
= 0; path
.path
[i
]; i
++) {
257 nir_deref_instr
*p
= path
.path
[i
];
258 b
.cursor
= nir_after_instr(&p
->instr
);
260 switch (p
->deref_type
) {
261 case nir_deref_type_var
:
262 assert(new_deref
== NULL
);
263 new_deref
= nir_build_deref_var(&b
, split_var
);
266 case nir_deref_type_array
:
267 case nir_deref_type_array_wildcard
:
268 new_deref
= nir_build_deref_follower(&b
, new_deref
, p
);
271 case nir_deref_type_struct
:
272 /* Nothing to do; we're splitting structs */
276 unreachable("Invalid deref type in path");
280 assert(new_deref
->type
== deref
->type
);
281 nir_ssa_def_rewrite_uses(&deref
->dest
.ssa
,
282 nir_src_for_ssa(&new_deref
->dest
.ssa
));
283 nir_deref_instr_remove_if_unused(deref
);
288 /** A pass for splitting structs into multiple variables
290 * This pass splits arrays of structs into multiple variables, one for each
291 * (possibly nested) structure member. After this pass completes, no
292 * variables of the given mode will contain a struct type.
295 nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
)
297 void *mem_ctx
= ralloc_context(NULL
);
298 struct hash_table
*var_field_map
=
299 _mesa_pointer_hash_table_create(mem_ctx
);
300 struct set
*complex_vars
= NULL
;
302 assert((modes
& (nir_var_shader_temp
| nir_var_function_temp
)) == modes
);
304 bool has_global_splits
= false;
305 if (modes
& nir_var_shader_temp
) {
306 has_global_splits
= split_var_list_structs(shader
, NULL
,
313 bool progress
= false;
314 nir_foreach_function(function
, shader
) {
318 bool has_local_splits
= false;
319 if (modes
& nir_var_function_temp
) {
320 has_local_splits
= split_var_list_structs(shader
, function
->impl
,
321 &function
->impl
->locals
,
327 if (has_global_splits
|| has_local_splits
) {
328 split_struct_derefs_impl(function
->impl
, var_field_map
,
331 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
332 nir_metadata_dominance
);
337 ralloc_free(mem_ctx
);
342 struct array_level_info
{
348 /* Only set if this is the tail end of the splitting */
352 struct array_split
*splits
;
355 struct array_var_info
{
356 nir_variable
*base_var
;
358 const struct glsl_type
*split_var_type
;
361 struct array_split root_split
;
364 struct array_level_info levels
[0];
368 init_var_list_array_infos(nir_shader
*shader
,
369 struct exec_list
*vars
,
370 struct hash_table
*var_info_map
,
371 struct set
**complex_vars
,
374 bool has_array
= false;
376 nir_foreach_variable(var
, vars
) {
377 int num_levels
= num_array_levels_in_array_of_vector_type(var
->type
);
381 if (*complex_vars
== NULL
)
382 *complex_vars
= get_complex_used_vars(shader
, mem_ctx
);
384 /* We can't split a variable that's referenced with deref that has any
385 * sort of complex usage.
387 if (_mesa_set_search(*complex_vars
, var
))
390 struct array_var_info
*info
=
391 rzalloc_size(mem_ctx
, sizeof(*info
) +
392 num_levels
* sizeof(info
->levels
[0]));
394 info
->base_var
= var
;
395 info
->num_levels
= num_levels
;
397 const struct glsl_type
*type
= var
->type
;
398 for (int i
= 0; i
< num_levels
; i
++) {
399 info
->levels
[i
].array_len
= glsl_get_length(type
);
400 type
= glsl_get_array_element(type
);
402 /* All levels start out initially as split */
403 info
->levels
[i
].split
= true;
406 _mesa_hash_table_insert(var_info_map
, var
, info
);
413 static struct array_var_info
*
414 get_array_var_info(nir_variable
*var
,
415 struct hash_table
*var_info_map
)
417 struct hash_entry
*entry
=
418 _mesa_hash_table_search(var_info_map
, var
);
419 return entry
? entry
->data
: NULL
;
422 static struct array_var_info
*
423 get_array_deref_info(nir_deref_instr
*deref
,
424 struct hash_table
*var_info_map
,
425 nir_variable_mode modes
)
427 if (!(deref
->mode
& modes
))
430 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
434 return get_array_var_info(var
, var_info_map
);
438 mark_array_deref_used(nir_deref_instr
*deref
,
439 struct hash_table
*var_info_map
,
440 nir_variable_mode modes
,
443 struct array_var_info
*info
=
444 get_array_deref_info(deref
, var_info_map
, modes
);
449 nir_deref_path_init(&path
, deref
, mem_ctx
);
451 /* Walk the path and look for indirects. If we have an array deref with an
452 * indirect, mark the given level as not being split.
454 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
455 nir_deref_instr
*p
= path
.path
[i
+ 1];
456 if (p
->deref_type
== nir_deref_type_array
&&
457 !nir_src_is_const(p
->arr
.index
))
458 info
->levels
[i
].split
= false;
463 mark_array_usage_impl(nir_function_impl
*impl
,
464 struct hash_table
*var_info_map
,
465 nir_variable_mode modes
,
468 nir_foreach_block(block
, impl
) {
469 nir_foreach_instr(instr
, block
) {
470 if (instr
->type
!= nir_instr_type_intrinsic
)
473 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
474 switch (intrin
->intrinsic
) {
475 case nir_intrinsic_copy_deref
:
476 mark_array_deref_used(nir_src_as_deref(intrin
->src
[1]),
477 var_info_map
, modes
, mem_ctx
);
480 case nir_intrinsic_load_deref
:
481 case nir_intrinsic_store_deref
:
482 mark_array_deref_used(nir_src_as_deref(intrin
->src
[0]),
483 var_info_map
, modes
, mem_ctx
);
494 create_split_array_vars(struct array_var_info
*var_info
,
496 struct array_split
*split
,
499 nir_function_impl
*impl
,
502 while (level
< var_info
->num_levels
&& !var_info
->levels
[level
].split
) {
503 name
= ralloc_asprintf(mem_ctx
, "%s[*]", name
);
507 if (level
== var_info
->num_levels
) {
508 /* We add parens to the variable name so it looks like "(foo[2][*])" so
509 * that further derefs will look like "(foo[2][*])[ssa_6]"
511 name
= ralloc_asprintf(mem_ctx
, "(%s)", name
);
513 nir_variable_mode mode
= var_info
->base_var
->data
.mode
;
514 if (mode
== nir_var_function_temp
) {
515 split
->var
= nir_local_variable_create(impl
,
516 var_info
->split_var_type
, name
);
518 split
->var
= nir_variable_create(shader
, mode
,
519 var_info
->split_var_type
, name
);
522 assert(var_info
->levels
[level
].split
);
523 split
->num_splits
= var_info
->levels
[level
].array_len
;
524 split
->splits
= rzalloc_array(mem_ctx
, struct array_split
,
526 for (unsigned i
= 0; i
< split
->num_splits
; i
++) {
527 create_split_array_vars(var_info
, level
+ 1, &split
->splits
[i
],
528 ralloc_asprintf(mem_ctx
, "%s[%d]", name
, i
),
529 shader
, impl
, mem_ctx
);
535 split_var_list_arrays(nir_shader
*shader
,
536 nir_function_impl
*impl
,
537 struct exec_list
*vars
,
538 struct hash_table
*var_info_map
,
541 struct exec_list split_vars
;
542 exec_list_make_empty(&split_vars
);
544 nir_foreach_variable_safe(var
, vars
) {
545 struct array_var_info
*info
= get_array_var_info(var
, var_info_map
);
549 bool has_split
= false;
550 const struct glsl_type
*split_type
=
551 glsl_without_array_or_matrix(var
->type
);
552 for (int i
= info
->num_levels
- 1; i
>= 0; i
--) {
553 if (info
->levels
[i
].split
) {
558 /* If the original type was a matrix type, we'd like to keep that so
559 * we don't convert matrices into arrays.
561 if (i
== info
->num_levels
- 1 &&
562 glsl_type_is_matrix(glsl_without_array(var
->type
))) {
563 split_type
= glsl_matrix_type(glsl_get_base_type(split_type
),
564 glsl_get_components(split_type
),
565 info
->levels
[i
].array_len
);
567 split_type
= glsl_array_type(split_type
, info
->levels
[i
].array_len
, 0);
572 info
->split_var_type
= split_type
;
573 /* To avoid list confusion (we'll be adding things as we split
574 * variables), pull all of the variables we plan to split off of the
575 * main variable list.
577 exec_node_remove(&var
->node
);
578 exec_list_push_tail(&split_vars
, &var
->node
);
580 assert(split_type
== glsl_get_bare_type(var
->type
));
581 /* If we're not modifying this variable, delete the info so we skip
582 * it faster in later passes.
584 _mesa_hash_table_remove_key(var_info_map
, var
);
588 nir_foreach_variable(var
, &split_vars
) {
589 struct array_var_info
*info
= get_array_var_info(var
, var_info_map
);
590 create_split_array_vars(info
, 0, &info
->root_split
, var
->name
,
591 shader
, impl
, mem_ctx
);
594 return !exec_list_is_empty(&split_vars
);
598 deref_has_split_wildcard(nir_deref_path
*path
,
599 struct array_var_info
*info
)
604 assert(path
->path
[0]->var
== info
->base_var
);
605 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
606 if (path
->path
[i
+ 1]->deref_type
== nir_deref_type_array_wildcard
&&
607 info
->levels
[i
].split
)
615 array_path_is_out_of_bounds(nir_deref_path
*path
,
616 struct array_var_info
*info
)
621 assert(path
->path
[0]->var
== info
->base_var
);
622 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
623 nir_deref_instr
*p
= path
->path
[i
+ 1];
624 if (p
->deref_type
== nir_deref_type_array_wildcard
)
627 if (nir_src_is_const(p
->arr
.index
) &&
628 nir_src_as_uint(p
->arr
.index
) >= info
->levels
[i
].array_len
)
636 emit_split_copies(nir_builder
*b
,
637 struct array_var_info
*dst_info
, nir_deref_path
*dst_path
,
638 unsigned dst_level
, nir_deref_instr
*dst
,
639 struct array_var_info
*src_info
, nir_deref_path
*src_path
,
640 unsigned src_level
, nir_deref_instr
*src
)
642 nir_deref_instr
*dst_p
, *src_p
;
644 while ((dst_p
= dst_path
->path
[dst_level
+ 1])) {
645 if (dst_p
->deref_type
== nir_deref_type_array_wildcard
)
648 dst
= nir_build_deref_follower(b
, dst
, dst_p
);
652 while ((src_p
= src_path
->path
[src_level
+ 1])) {
653 if (src_p
->deref_type
== nir_deref_type_array_wildcard
)
656 src
= nir_build_deref_follower(b
, src
, src_p
);
660 if (src_p
== NULL
|| dst_p
== NULL
) {
661 assert(src_p
== NULL
&& dst_p
== NULL
);
662 nir_copy_deref(b
, dst
, src
);
664 assert(dst_p
->deref_type
== nir_deref_type_array_wildcard
&&
665 src_p
->deref_type
== nir_deref_type_array_wildcard
);
667 if ((dst_info
&& dst_info
->levels
[dst_level
].split
) ||
668 (src_info
&& src_info
->levels
[src_level
].split
)) {
669 /* There are no indirects at this level on one of the source or the
670 * destination so we are lowering it.
672 assert(glsl_get_length(dst_path
->path
[dst_level
]->type
) ==
673 glsl_get_length(src_path
->path
[src_level
]->type
));
674 unsigned len
= glsl_get_length(dst_path
->path
[dst_level
]->type
);
675 for (unsigned i
= 0; i
< len
; i
++) {
676 emit_split_copies(b
, dst_info
, dst_path
, dst_level
+ 1,
677 nir_build_deref_array_imm(b
, dst
, i
),
678 src_info
, src_path
, src_level
+ 1,
679 nir_build_deref_array_imm(b
, src
, i
));
682 /* Neither side is being split so we just keep going */
683 emit_split_copies(b
, dst_info
, dst_path
, dst_level
+ 1,
684 nir_build_deref_array_wildcard(b
, dst
),
685 src_info
, src_path
, src_level
+ 1,
686 nir_build_deref_array_wildcard(b
, src
));
692 split_array_copies_impl(nir_function_impl
*impl
,
693 struct hash_table
*var_info_map
,
694 nir_variable_mode modes
,
698 nir_builder_init(&b
, impl
);
700 nir_foreach_block(block
, impl
) {
701 nir_foreach_instr_safe(instr
, block
) {
702 if (instr
->type
!= nir_instr_type_intrinsic
)
705 nir_intrinsic_instr
*copy
= nir_instr_as_intrinsic(instr
);
706 if (copy
->intrinsic
!= nir_intrinsic_copy_deref
)
709 nir_deref_instr
*dst_deref
= nir_src_as_deref(copy
->src
[0]);
710 nir_deref_instr
*src_deref
= nir_src_as_deref(copy
->src
[1]);
712 struct array_var_info
*dst_info
=
713 get_array_deref_info(dst_deref
, var_info_map
, modes
);
714 struct array_var_info
*src_info
=
715 get_array_deref_info(src_deref
, var_info_map
, modes
);
717 if (!src_info
&& !dst_info
)
720 nir_deref_path dst_path
, src_path
;
721 nir_deref_path_init(&dst_path
, dst_deref
, mem_ctx
);
722 nir_deref_path_init(&src_path
, src_deref
, mem_ctx
);
724 if (!deref_has_split_wildcard(&dst_path
, dst_info
) &&
725 !deref_has_split_wildcard(&src_path
, src_info
))
728 b
.cursor
= nir_instr_remove(©
->instr
);
730 emit_split_copies(&b
, dst_info
, &dst_path
, 0, dst_path
.path
[0],
731 src_info
, &src_path
, 0, src_path
.path
[0]);
737 split_array_access_impl(nir_function_impl
*impl
,
738 struct hash_table
*var_info_map
,
739 nir_variable_mode modes
,
743 nir_builder_init(&b
, impl
);
745 nir_foreach_block(block
, impl
) {
746 nir_foreach_instr_safe(instr
, block
) {
747 if (instr
->type
== nir_instr_type_deref
) {
748 /* Clean up any dead derefs we find lying around. They may refer
749 * to variables we're planning to split.
751 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
752 if (deref
->mode
& modes
)
753 nir_deref_instr_remove_if_unused(deref
);
757 if (instr
->type
!= nir_instr_type_intrinsic
)
760 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
761 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
&&
762 intrin
->intrinsic
!= nir_intrinsic_store_deref
&&
763 intrin
->intrinsic
!= nir_intrinsic_copy_deref
)
766 const unsigned num_derefs
=
767 intrin
->intrinsic
== nir_intrinsic_copy_deref
? 2 : 1;
769 for (unsigned d
= 0; d
< num_derefs
; d
++) {
770 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[d
]);
772 struct array_var_info
*info
=
773 get_array_deref_info(deref
, var_info_map
, modes
);
778 nir_deref_path_init(&path
, deref
, mem_ctx
);
780 b
.cursor
= nir_before_instr(&intrin
->instr
);
782 if (array_path_is_out_of_bounds(&path
, info
)) {
783 /* If one of the derefs is out-of-bounds, we just delete the
784 * instruction. If a destination is out of bounds, then it may
785 * have been in-bounds prior to shrinking so we don't want to
786 * accidentally stomp something. However, we've already proven
787 * that it will never be read so it's safe to delete. If a
788 * source is out of bounds then it is loading random garbage.
789 * For loads, we replace their uses with an undef instruction
790 * and for copies we just delete the copy since it was writing
791 * undefined garbage anyway and we may as well leave the random
792 * garbage in the destination alone.
794 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
796 nir_ssa_undef(&b
, intrin
->dest
.ssa
.num_components
,
797 intrin
->dest
.ssa
.bit_size
);
798 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
801 nir_instr_remove(&intrin
->instr
);
802 for (unsigned i
= 0; i
< num_derefs
; i
++)
803 nir_deref_instr_remove_if_unused(nir_src_as_deref(intrin
->src
[i
]));
807 struct array_split
*split
= &info
->root_split
;
808 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
809 if (info
->levels
[i
].split
) {
810 nir_deref_instr
*p
= path
.path
[i
+ 1];
811 unsigned index
= nir_src_as_uint(p
->arr
.index
);
812 assert(index
< info
->levels
[i
].array_len
);
813 split
= &split
->splits
[index
];
816 assert(!split
->splits
&& split
->var
);
818 nir_deref_instr
*new_deref
= nir_build_deref_var(&b
, split
->var
);
819 for (unsigned i
= 0; i
< info
->num_levels
; i
++) {
820 if (!info
->levels
[i
].split
) {
821 new_deref
= nir_build_deref_follower(&b
, new_deref
,
825 assert(new_deref
->type
== deref
->type
);
827 /* Rewrite the deref source to point to the split one */
828 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[d
],
829 nir_src_for_ssa(&new_deref
->dest
.ssa
));
830 nir_deref_instr_remove_if_unused(deref
);
836 /** A pass for splitting arrays of vectors into multiple variables
838 * This pass looks at arrays (possibly multiple levels) of vectors (not
839 * structures or other types) and tries to split them into piles of variables,
840 * one for each array element. The heuristic used is simple: If a given array
841 * level is never used with an indirect, that array level will get split.
843 * This pass probably could handles structures easily enough but making a pass
844 * that could see through an array of structures of arrays would be difficult
845 * so it's best to just run nir_split_struct_vars first.
848 nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
)
850 void *mem_ctx
= ralloc_context(NULL
);
851 struct hash_table
*var_info_map
= _mesa_pointer_hash_table_create(mem_ctx
);
852 struct set
*complex_vars
= NULL
;
854 assert((modes
& (nir_var_shader_temp
| nir_var_function_temp
)) == modes
);
856 bool has_global_array
= false;
857 if (modes
& nir_var_shader_temp
) {
858 has_global_array
= init_var_list_array_infos(shader
,
865 bool has_any_array
= false;
866 nir_foreach_function(function
, shader
) {
870 bool has_local_array
= false;
871 if (modes
& nir_var_function_temp
) {
872 has_local_array
= init_var_list_array_infos(shader
,
873 &function
->impl
->locals
,
879 if (has_global_array
|| has_local_array
) {
880 has_any_array
= true;
881 mark_array_usage_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
885 /* If we failed to find any arrays of arrays, bail early. */
886 if (!has_any_array
) {
887 ralloc_free(mem_ctx
);
891 bool has_global_splits
= false;
892 if (modes
& nir_var_shader_temp
) {
893 has_global_splits
= split_var_list_arrays(shader
, NULL
,
895 var_info_map
, mem_ctx
);
898 bool progress
= false;
899 nir_foreach_function(function
, shader
) {
903 bool has_local_splits
= false;
904 if (modes
& nir_var_function_temp
) {
905 has_local_splits
= split_var_list_arrays(shader
, function
->impl
,
906 &function
->impl
->locals
,
907 var_info_map
, mem_ctx
);
910 if (has_global_splits
|| has_local_splits
) {
911 split_array_copies_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
912 split_array_access_impl(function
->impl
, var_info_map
, modes
, mem_ctx
);
914 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
915 nir_metadata_dominance
);
920 ralloc_free(mem_ctx
);
925 struct array_level_usage
{
928 /* The value UINT_MAX will be used to indicate an indirect */
930 unsigned max_written
;
932 /* True if there is a copy that isn't to/from a shrinkable array */
933 bool has_external_copy
;
934 struct set
*levels_copied
;
937 struct vec_var_usage
{
938 /* Convenience set of all components this variable has */
939 nir_component_mask_t all_comps
;
941 nir_component_mask_t comps_read
;
942 nir_component_mask_t comps_written
;
944 nir_component_mask_t comps_kept
;
946 /* True if there is a copy that isn't to/from a shrinkable vector */
947 bool has_external_copy
;
948 bool has_complex_use
;
949 struct set
*vars_copied
;
952 struct array_level_usage levels
[0];
955 static struct vec_var_usage
*
956 get_vec_var_usage(nir_variable
*var
,
957 struct hash_table
*var_usage_map
,
958 bool add_usage_entry
, void *mem_ctx
)
960 struct hash_entry
*entry
= _mesa_hash_table_search(var_usage_map
, var
);
964 if (!add_usage_entry
)
967 /* Check to make sure that we are working with an array of vectors. We
968 * don't bother to shrink single vectors because we figure that we can
969 * clean it up better with SSA than by inserting piles of vecN instructions
970 * to compact results.
972 int num_levels
= num_array_levels_in_array_of_vector_type(var
->type
);
974 return NULL
; /* Not an array of vectors */
976 struct vec_var_usage
*usage
=
977 rzalloc_size(mem_ctx
, sizeof(*usage
) +
978 num_levels
* sizeof(usage
->levels
[0]));
980 usage
->num_levels
= num_levels
;
981 const struct glsl_type
*type
= var
->type
;
982 for (unsigned i
= 0; i
< num_levels
; i
++) {
983 usage
->levels
[i
].array_len
= glsl_get_length(type
);
984 type
= glsl_get_array_element(type
);
986 assert(glsl_type_is_vector_or_scalar(type
));
988 usage
->all_comps
= (1 << glsl_get_components(type
)) - 1;
990 _mesa_hash_table_insert(var_usage_map
, var
, usage
);
995 static struct vec_var_usage
*
996 get_vec_deref_usage(nir_deref_instr
*deref
,
997 struct hash_table
*var_usage_map
,
998 nir_variable_mode modes
,
999 bool add_usage_entry
, void *mem_ctx
)
1001 if (!(deref
->mode
& modes
))
1004 return get_vec_var_usage(nir_deref_instr_get_variable(deref
),
1005 var_usage_map
, add_usage_entry
, mem_ctx
);
1009 mark_deref_if_complex(nir_deref_instr
*deref
,
1010 struct hash_table
*var_usage_map
,
1011 nir_variable_mode modes
,
1014 if (!(deref
->mode
& modes
))
1017 /* Only bother with var derefs because nir_deref_instr_has_complex_use is
1020 if (deref
->deref_type
!= nir_deref_type_var
)
1023 if (!nir_deref_instr_has_complex_use(deref
))
1026 struct vec_var_usage
*usage
=
1027 get_vec_var_usage(deref
->var
, var_usage_map
, true, mem_ctx
);
1031 usage
->has_complex_use
= true;
1035 mark_deref_used(nir_deref_instr
*deref
,
1036 nir_component_mask_t comps_read
,
1037 nir_component_mask_t comps_written
,
1038 nir_deref_instr
*copy_deref
,
1039 struct hash_table
*var_usage_map
,
1040 nir_variable_mode modes
,
1043 if (!(deref
->mode
& modes
))
1046 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
1050 struct vec_var_usage
*usage
=
1051 get_vec_var_usage(var
, var_usage_map
, true, mem_ctx
);
1055 usage
->comps_read
|= comps_read
& usage
->all_comps
;
1056 usage
->comps_written
|= comps_written
& usage
->all_comps
;
1058 struct vec_var_usage
*copy_usage
= NULL
;
1060 copy_usage
= get_vec_deref_usage(copy_deref
, var_usage_map
, modes
,
1063 if (usage
->vars_copied
== NULL
) {
1064 usage
->vars_copied
= _mesa_pointer_set_create(mem_ctx
);
1066 _mesa_set_add(usage
->vars_copied
, copy_usage
);
1068 usage
->has_external_copy
= true;
1072 nir_deref_path path
;
1073 nir_deref_path_init(&path
, deref
, mem_ctx
);
1075 nir_deref_path copy_path
;
1077 nir_deref_path_init(©_path
, copy_deref
, mem_ctx
);
1079 unsigned copy_i
= 0;
1080 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1081 struct array_level_usage
*level
= &usage
->levels
[i
];
1082 nir_deref_instr
*deref
= path
.path
[i
+ 1];
1083 assert(deref
->deref_type
== nir_deref_type_array
||
1084 deref
->deref_type
== nir_deref_type_array_wildcard
);
1087 if (deref
->deref_type
== nir_deref_type_array
) {
1088 max_used
= nir_src_is_const(deref
->arr
.index
) ?
1089 nir_src_as_uint(deref
->arr
.index
) : UINT_MAX
;
1091 /* For wildcards, we read or wrote the whole thing. */
1092 assert(deref
->deref_type
== nir_deref_type_array_wildcard
);
1093 max_used
= level
->array_len
- 1;
1096 /* Match each wildcard level with the level on copy_usage */
1097 for (; copy_path
.path
[copy_i
+ 1]; copy_i
++) {
1098 if (copy_path
.path
[copy_i
+ 1]->deref_type
==
1099 nir_deref_type_array_wildcard
)
1102 struct array_level_usage
*copy_level
=
1103 ©_usage
->levels
[copy_i
++];
1105 if (level
->levels_copied
== NULL
) {
1106 level
->levels_copied
= _mesa_pointer_set_create(mem_ctx
);
1108 _mesa_set_add(level
->levels_copied
, copy_level
);
1110 /* We have a wildcard and it comes from a variable we aren't
1111 * tracking; flag it and we'll know to not shorten this array.
1113 level
->has_external_copy
= true;
1118 level
->max_written
= MAX2(level
->max_written
, max_used
);
1120 level
->max_read
= MAX2(level
->max_read
, max_used
);
1125 src_is_load_deref(nir_src src
, nir_src deref_src
)
1127 nir_intrinsic_instr
*load
= nir_src_as_intrinsic(src
);
1128 if (load
== NULL
|| load
->intrinsic
!= nir_intrinsic_load_deref
)
1131 assert(load
->src
[0].is_ssa
);
1133 return load
->src
[0].ssa
== deref_src
.ssa
;
1136 /* Returns all non-self-referential components of a store instruction. A
1137 * component is self-referential if it comes from the same component of a load
1138 * instruction on the same deref. If the only data in a particular component
1139 * of a variable came directly from that component then it's undefined. The
1140 * only way to get defined data into a component of a variable is for it to
1141 * get written there by something outside or from a different component.
1143 * This is a fairly common pattern in shaders that come from either GLSL IR or
1144 * GLSLang because both glsl_to_nir and GLSLang implement write-masking with
1147 static nir_component_mask_t
1148 get_non_self_referential_store_comps(nir_intrinsic_instr
*store
)
1150 nir_component_mask_t comps
= nir_intrinsic_write_mask(store
);
1152 assert(store
->src
[1].is_ssa
);
1153 nir_instr
*src_instr
= store
->src
[1].ssa
->parent_instr
;
1154 if (src_instr
->type
!= nir_instr_type_alu
)
1157 nir_alu_instr
*src_alu
= nir_instr_as_alu(src_instr
);
1159 if (src_alu
->op
== nir_op_mov
) {
1160 /* If it's just a swizzle of a load from the same deref, discount any
1161 * channels that don't move in the swizzle.
1163 if (src_is_load_deref(src_alu
->src
[0].src
, store
->src
[0])) {
1164 for (unsigned i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; i
++) {
1165 if (src_alu
->src
[0].swizzle
[i
] == i
)
1166 comps
&= ~(1u << i
);
1169 } else if (src_alu
->op
== nir_op_vec2
||
1170 src_alu
->op
== nir_op_vec3
||
1171 src_alu
->op
== nir_op_vec4
) {
1172 /* If it's a vec, discount any channels that are just loads from the
1173 * same deref put in the same spot.
1175 for (unsigned i
= 0; i
< nir_op_infos
[src_alu
->op
].num_inputs
; i
++) {
1176 if (src_is_load_deref(src_alu
->src
[i
].src
, store
->src
[0]) &&
1177 src_alu
->src
[i
].swizzle
[0] == i
)
1178 comps
&= ~(1u << i
);
1186 find_used_components_impl(nir_function_impl
*impl
,
1187 struct hash_table
*var_usage_map
,
1188 nir_variable_mode modes
,
1191 nir_foreach_block(block
, impl
) {
1192 nir_foreach_instr(instr
, block
) {
1193 if (instr
->type
== nir_instr_type_deref
) {
1194 mark_deref_if_complex(nir_instr_as_deref(instr
),
1195 var_usage_map
, modes
, mem_ctx
);
1198 if (instr
->type
!= nir_instr_type_intrinsic
)
1201 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1202 switch (intrin
->intrinsic
) {
1203 case nir_intrinsic_load_deref
:
1204 mark_deref_used(nir_src_as_deref(intrin
->src
[0]),
1205 nir_ssa_def_components_read(&intrin
->dest
.ssa
), 0,
1206 NULL
, var_usage_map
, modes
, mem_ctx
);
1209 case nir_intrinsic_store_deref
:
1210 mark_deref_used(nir_src_as_deref(intrin
->src
[0]),
1211 0, get_non_self_referential_store_comps(intrin
),
1212 NULL
, var_usage_map
, modes
, mem_ctx
);
1215 case nir_intrinsic_copy_deref
: {
1216 /* Just mark everything used for copies. */
1217 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
1218 nir_deref_instr
*src
= nir_src_as_deref(intrin
->src
[1]);
1219 mark_deref_used(dst
, 0, ~0, src
, var_usage_map
, modes
, mem_ctx
);
1220 mark_deref_used(src
, ~0, 0, dst
, var_usage_map
, modes
, mem_ctx
);
1232 shrink_vec_var_list(struct exec_list
*vars
,
1233 struct hash_table
*var_usage_map
)
1235 /* Initialize the components kept field of each variable. This is the
1236 * AND of the components written and components read. If a component is
1237 * written but never read, it's dead. If it is read but never written,
1238 * then all values read are undefined garbage and we may as well not read
1241 * The same logic applies to the array length. We make the array length
1242 * the minimum needed required length between read and write and plan to
1243 * discard any OOB access. The one exception here is indirect writes
1244 * because we don't know where they will land and we can't shrink an array
1245 * with indirect writes because previously in-bounds writes may become
1246 * out-of-bounds and have undefined behavior.
1248 * Also, if we have a copy that to/from something we can't shrink, we need
1249 * to leave components and array_len of any wildcards alone.
1251 nir_foreach_variable(var
, vars
) {
1252 struct vec_var_usage
*usage
=
1253 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1257 assert(usage
->comps_kept
== 0);
1258 if (usage
->has_external_copy
|| usage
->has_complex_use
)
1259 usage
->comps_kept
= usage
->all_comps
;
1261 usage
->comps_kept
= usage
->comps_read
& usage
->comps_written
;
1263 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1264 struct array_level_usage
*level
= &usage
->levels
[i
];
1265 assert(level
->array_len
> 0);
1267 if (level
->max_written
== UINT_MAX
|| level
->has_external_copy
||
1268 usage
->has_complex_use
)
1269 continue; /* Can't shrink */
1271 unsigned max_used
= MIN2(level
->max_read
, level
->max_written
);
1272 level
->array_len
= MIN2(max_used
, level
->array_len
- 1) + 1;
1276 /* In order for variable copies to work, we have to have the same data type
1277 * on the source and the destination. In order to satisfy this, we run a
1278 * little fixed-point algorithm to transitively ensure that we get enough
1279 * components and array elements for this to hold for all copies.
1283 fp_progress
= false;
1284 nir_foreach_variable(var
, vars
) {
1285 struct vec_var_usage
*var_usage
=
1286 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1287 if (!var_usage
|| !var_usage
->vars_copied
)
1290 set_foreach(var_usage
->vars_copied
, copy_entry
) {
1291 struct vec_var_usage
*copy_usage
= (void *)copy_entry
->key
;
1292 if (copy_usage
->comps_kept
!= var_usage
->comps_kept
) {
1293 nir_component_mask_t comps_kept
=
1294 (var_usage
->comps_kept
| copy_usage
->comps_kept
);
1295 var_usage
->comps_kept
= comps_kept
;
1296 copy_usage
->comps_kept
= comps_kept
;
1301 for (unsigned i
= 0; i
< var_usage
->num_levels
; i
++) {
1302 struct array_level_usage
*var_level
= &var_usage
->levels
[i
];
1303 if (!var_level
->levels_copied
)
1306 set_foreach(var_level
->levels_copied
, copy_entry
) {
1307 struct array_level_usage
*copy_level
= (void *)copy_entry
->key
;
1308 if (var_level
->array_len
!= copy_level
->array_len
) {
1309 unsigned array_len
=
1310 MAX2(var_level
->array_len
, copy_level
->array_len
);
1311 var_level
->array_len
= array_len
;
1312 copy_level
->array_len
= array_len
;
1318 } while (fp_progress
);
1320 bool vars_shrunk
= false;
1321 nir_foreach_variable_safe(var
, vars
) {
1322 struct vec_var_usage
*usage
=
1323 get_vec_var_usage(var
, var_usage_map
, false, NULL
);
1327 bool shrunk
= false;
1328 const struct glsl_type
*vec_type
= var
->type
;
1329 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1330 /* If we've reduced the array to zero elements at some level, just
1331 * set comps_kept to 0 and delete the variable.
1333 if (usage
->levels
[i
].array_len
== 0) {
1334 usage
->comps_kept
= 0;
1338 assert(usage
->levels
[i
].array_len
<= glsl_get_length(vec_type
));
1339 if (usage
->levels
[i
].array_len
< glsl_get_length(vec_type
))
1341 vec_type
= glsl_get_array_element(vec_type
);
1343 assert(glsl_type_is_vector_or_scalar(vec_type
));
1345 assert(usage
->comps_kept
== (usage
->comps_kept
& usage
->all_comps
));
1346 if (usage
->comps_kept
!= usage
->all_comps
)
1349 if (usage
->comps_kept
== 0) {
1350 /* This variable is dead, remove it */
1352 exec_node_remove(&var
->node
);
1357 /* This variable doesn't need to be shrunk. Remove it from the
1358 * hash table so later steps will ignore it.
1360 _mesa_hash_table_remove_key(var_usage_map
, var
);
1364 /* Build the new var type */
1365 unsigned new_num_comps
= util_bitcount(usage
->comps_kept
);
1366 const struct glsl_type
*new_type
=
1367 glsl_vector_type(glsl_get_base_type(vec_type
), new_num_comps
);
1368 for (int i
= usage
->num_levels
- 1; i
>= 0; i
--) {
1369 assert(usage
->levels
[i
].array_len
> 0);
1370 /* If the original type was a matrix type, we'd like to keep that so
1371 * we don't convert matrices into arrays.
1373 if (i
== usage
->num_levels
- 1 &&
1374 glsl_type_is_matrix(glsl_without_array(var
->type
)) &&
1375 new_num_comps
> 1 && usage
->levels
[i
].array_len
> 1) {
1376 new_type
= glsl_matrix_type(glsl_get_base_type(new_type
),
1378 usage
->levels
[i
].array_len
);
1380 new_type
= glsl_array_type(new_type
, usage
->levels
[i
].array_len
, 0);
1383 var
->type
= new_type
;
1392 vec_deref_is_oob(nir_deref_instr
*deref
,
1393 struct vec_var_usage
*usage
)
1395 nir_deref_path path
;
1396 nir_deref_path_init(&path
, deref
, NULL
);
1399 for (unsigned i
= 0; i
< usage
->num_levels
; i
++) {
1400 nir_deref_instr
*p
= path
.path
[i
+ 1];
1401 if (p
->deref_type
== nir_deref_type_array_wildcard
)
1404 if (nir_src_is_const(p
->arr
.index
) &&
1405 nir_src_as_uint(p
->arr
.index
) >= usage
->levels
[i
].array_len
) {
1411 nir_deref_path_finish(&path
);
1417 vec_deref_is_dead_or_oob(nir_deref_instr
*deref
,
1418 struct hash_table
*var_usage_map
,
1419 nir_variable_mode modes
)
1421 struct vec_var_usage
*usage
=
1422 get_vec_deref_usage(deref
, var_usage_map
, modes
, false, NULL
);
1426 return usage
->comps_kept
== 0 || vec_deref_is_oob(deref
, usage
);
1430 shrink_vec_var_access_impl(nir_function_impl
*impl
,
1431 struct hash_table
*var_usage_map
,
1432 nir_variable_mode modes
)
1435 nir_builder_init(&b
, impl
);
1437 nir_foreach_block(block
, impl
) {
1438 nir_foreach_instr_safe(instr
, block
) {
1439 switch (instr
->type
) {
1440 case nir_instr_type_deref
: {
1441 nir_deref_instr
*deref
= nir_instr_as_deref(instr
);
1442 if (!(deref
->mode
& modes
))
1445 /* Clean up any dead derefs we find lying around. They may refer
1446 * to variables we've deleted.
1448 if (nir_deref_instr_remove_if_unused(deref
))
1451 /* Update the type in the deref to keep the types consistent as
1452 * you walk down the chain. We don't need to check if this is one
1453 * of the derefs we're shrinking because this is a no-op if it
1454 * isn't. The worst that could happen is that we accidentally fix
1457 if (deref
->deref_type
== nir_deref_type_var
) {
1458 deref
->type
= deref
->var
->type
;
1459 } else if (deref
->deref_type
== nir_deref_type_array
||
1460 deref
->deref_type
== nir_deref_type_array_wildcard
) {
1461 nir_deref_instr
*parent
= nir_deref_instr_parent(deref
);
1462 assert(glsl_type_is_array(parent
->type
) ||
1463 glsl_type_is_matrix(parent
->type
));
1464 deref
->type
= glsl_get_array_element(parent
->type
);
1469 case nir_instr_type_intrinsic
: {
1470 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1472 /* If we have a copy whose source or destination has been deleted
1473 * because we determined the variable was dead, then we just
1474 * delete the copy instruction. If the source variable was dead
1475 * then it was writing undefined garbage anyway and if it's the
1476 * destination variable that's dead then the write isn't needed.
1478 if (intrin
->intrinsic
== nir_intrinsic_copy_deref
) {
1479 nir_deref_instr
*dst
= nir_src_as_deref(intrin
->src
[0]);
1480 nir_deref_instr
*src
= nir_src_as_deref(intrin
->src
[1]);
1481 if (vec_deref_is_dead_or_oob(dst
, var_usage_map
, modes
) ||
1482 vec_deref_is_dead_or_oob(src
, var_usage_map
, modes
)) {
1483 nir_instr_remove(&intrin
->instr
);
1484 nir_deref_instr_remove_if_unused(dst
);
1485 nir_deref_instr_remove_if_unused(src
);
1490 if (intrin
->intrinsic
!= nir_intrinsic_load_deref
&&
1491 intrin
->intrinsic
!= nir_intrinsic_store_deref
)
1494 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
1495 if (!(deref
->mode
& modes
))
1498 struct vec_var_usage
*usage
=
1499 get_vec_deref_usage(deref
, var_usage_map
, modes
, false, NULL
);
1503 if (usage
->comps_kept
== 0 || vec_deref_is_oob(deref
, usage
)) {
1504 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
1506 nir_ssa_undef(&b
, intrin
->dest
.ssa
.num_components
,
1507 intrin
->dest
.ssa
.bit_size
);
1508 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
1509 nir_src_for_ssa(u
));
1511 nir_instr_remove(&intrin
->instr
);
1512 nir_deref_instr_remove_if_unused(deref
);
1516 /* If we're not dropping any components, there's no need to
1519 if (usage
->comps_kept
== usage
->all_comps
)
1522 if (intrin
->intrinsic
== nir_intrinsic_load_deref
) {
1523 b
.cursor
= nir_after_instr(&intrin
->instr
);
1525 nir_ssa_def
*undef
=
1526 nir_ssa_undef(&b
, 1, intrin
->dest
.ssa
.bit_size
);
1527 nir_ssa_def
*vec_srcs
[NIR_MAX_VEC_COMPONENTS
];
1529 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1530 if (usage
->comps_kept
& (1u << i
))
1531 vec_srcs
[i
] = nir_channel(&b
, &intrin
->dest
.ssa
, c
++);
1533 vec_srcs
[i
] = undef
;
1535 nir_ssa_def
*vec
= nir_vec(&b
, vec_srcs
, intrin
->num_components
);
1537 nir_ssa_def_rewrite_uses_after(&intrin
->dest
.ssa
,
1538 nir_src_for_ssa(vec
),
1541 /* The SSA def is now only used by the swizzle. It's safe to
1542 * shrink the number of components.
1544 assert(list_length(&intrin
->dest
.ssa
.uses
) == c
);
1545 intrin
->num_components
= c
;
1546 intrin
->dest
.ssa
.num_components
= c
;
1548 nir_component_mask_t write_mask
=
1549 nir_intrinsic_write_mask(intrin
);
1551 unsigned swizzle
[NIR_MAX_VEC_COMPONENTS
];
1552 nir_component_mask_t new_write_mask
= 0;
1554 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
1555 if (usage
->comps_kept
& (1u << i
)) {
1557 if (write_mask
& (1u << i
))
1558 new_write_mask
|= 1u << c
;
1563 b
.cursor
= nir_before_instr(&intrin
->instr
);
1565 nir_ssa_def
*swizzled
=
1566 nir_swizzle(&b
, intrin
->src
[1].ssa
, swizzle
, c
);
1568 /* Rewrite to use the compacted source */
1569 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[1],
1570 nir_src_for_ssa(swizzled
));
1571 nir_intrinsic_set_write_mask(intrin
, new_write_mask
);
1572 intrin
->num_components
= c
;
1585 function_impl_has_vars_with_modes(nir_function_impl
*impl
,
1586 nir_variable_mode modes
)
1588 nir_shader
*shader
= impl
->function
->shader
;
1590 if ((modes
& nir_var_shader_temp
) && !exec_list_is_empty(&shader
->globals
))
1593 if ((modes
& nir_var_function_temp
) && !exec_list_is_empty(&impl
->locals
))
1599 /** Attempt to shrink arrays of vectors
1601 * This pass looks at variables which contain a vector or an array (possibly
1602 * multiple dimensions) of vectors and attempts to lower to a smaller vector
1603 * or array. If the pass can prove that a component of a vector (or array of
1604 * vectors) is never really used, then that component will be removed.
1605 * Similarly, the pass attempts to shorten arrays based on what elements it
1606 * can prove are never read or never contain valid data.
1609 nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
)
1611 assert((modes
& (nir_var_shader_temp
| nir_var_function_temp
)) == modes
);
1613 void *mem_ctx
= ralloc_context(NULL
);
1615 struct hash_table
*var_usage_map
=
1616 _mesa_pointer_hash_table_create(mem_ctx
);
1618 bool has_vars_to_shrink
= false;
1619 nir_foreach_function(function
, shader
) {
1620 if (!function
->impl
)
1623 /* Don't even bother crawling the IR if we don't have any variables.
1624 * Given that this pass deletes any unused variables, it's likely that
1625 * we will be in this scenario eventually.
1627 if (function_impl_has_vars_with_modes(function
->impl
, modes
)) {
1628 has_vars_to_shrink
= true;
1629 find_used_components_impl(function
->impl
, var_usage_map
,
1633 if (!has_vars_to_shrink
) {
1634 ralloc_free(mem_ctx
);
1638 bool globals_shrunk
= false;
1639 if (modes
& nir_var_shader_temp
)
1640 globals_shrunk
= shrink_vec_var_list(&shader
->globals
, var_usage_map
);
1642 bool progress
= false;
1643 nir_foreach_function(function
, shader
) {
1644 if (!function
->impl
)
1647 bool locals_shrunk
= false;
1648 if (modes
& nir_var_function_temp
) {
1649 locals_shrunk
= shrink_vec_var_list(&function
->impl
->locals
,
1653 if (globals_shrunk
|| locals_shrunk
) {
1654 shrink_vec_var_access_impl(function
->impl
, var_usage_map
, modes
);
1656 nir_metadata_preserve(function
->impl
, nir_metadata_block_index
|
1657 nir_metadata_dominance
);
1662 ralloc_free(mem_ctx
);