2 * Copyright © 2014 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
24 * Jason Ekstrand (jason@jlekstrand.net)
33 struct deref_node
*parent
;
34 const struct glsl_type
*type
;
38 /* Only valid for things that end up in the direct list.
39 * Note that multiple nir_deref_vars may correspond to this node, but they
40 * will all be equivalent, so any is as good as the other.
43 struct exec_node direct_derefs_link
;
49 nir_ssa_def
**def_stack
;
50 nir_ssa_def
**def_stack_tail
;
52 struct deref_node
*wildcard
;
53 struct deref_node
*indirect
;
54 struct deref_node
*children
[0];
57 struct lower_variables_state
{
60 nir_function_impl
*impl
;
62 /* A hash table mapping variables to deref_node data */
63 struct hash_table
*deref_var_nodes
;
65 /* A hash table mapping fully-qualified direct dereferences, i.e.
66 * dereferences with no indirect or wildcard array dereferences, to
69 * At the moment, we only lower loads, stores, and copies that can be
70 * trivially lowered to loads and stores, i.e. copies with no indirects
71 * and no wildcards. If a part of a variable that is being loaded from
72 * and/or stored into is also involved in a copy operation with
73 * wildcards, then we lower that copy operation to loads and stores, but
74 * otherwise we leave copies with wildcards alone. Since the only derefs
75 * used in these loads, stores, and trivial copies are ones with no
76 * wildcards and no indirects, these are precisely the derefs that we
77 * can actually consider lowering.
79 struct exec_list direct_deref_nodes
;
81 /* Controls whether get_deref_node will add variables to the
82 * direct_deref_nodes table. This is turned on when we are initially
83 * scanning for load/store instructions. It is then turned off so we
84 * don't accidentally change the direct_deref_nodes table while we're
85 * iterating throug it.
87 bool add_to_direct_deref_nodes
;
89 /* A hash table mapping phi nodes to deref_state data */
90 struct hash_table
*phi_table
;
94 type_get_length(const struct glsl_type
*type
)
96 switch (glsl_get_base_type(type
)) {
97 case GLSL_TYPE_STRUCT
:
99 return glsl_get_length(type
);
100 case GLSL_TYPE_FLOAT
:
104 if (glsl_type_is_matrix(type
))
105 return glsl_get_matrix_columns(type
);
107 return glsl_get_vector_elements(type
);
109 unreachable("Invalid deref base type");
113 static struct deref_node
*
114 deref_node_create(struct deref_node
*parent
,
115 const struct glsl_type
*type
, nir_shader
*shader
)
117 size_t size
= sizeof(struct deref_node
) +
118 type_get_length(type
) * sizeof(struct deref_node
*);
120 struct deref_node
*node
= rzalloc_size(shader
, size
);
122 node
->parent
= parent
;
124 exec_node_init(&node
->direct_derefs_link
);
129 /* Returns the deref node associated with the given variable. This will be
130 * the root of the tree representing all of the derefs of the given variable.
132 static struct deref_node
*
133 get_deref_node_for_var(nir_variable
*var
, struct lower_variables_state
*state
)
135 struct deref_node
*node
;
137 struct hash_entry
*var_entry
=
138 _mesa_hash_table_search(state
->deref_var_nodes
, var
);
141 return var_entry
->data
;
143 node
= deref_node_create(NULL
, var
->type
, state
->dead_ctx
);
144 _mesa_hash_table_insert(state
->deref_var_nodes
, var
, node
);
149 /* Gets the deref_node for the given deref chain and creates it if it
150 * doesn't yet exist. If the deref is fully-qualified and direct and
151 * state->add_to_direct_deref_nodes is true, it will be added to the hash
152 * table of of fully-qualified direct derefs.
154 static struct deref_node
*
155 get_deref_node(nir_deref_var
*deref
, struct lower_variables_state
*state
)
157 bool is_direct
= true;
159 /* Start at the base of the chain. */
160 struct deref_node
*node
= get_deref_node_for_var(deref
->var
, state
);
161 assert(deref
->deref
.type
== node
->type
);
163 for (nir_deref
*tail
= deref
->deref
.child
; tail
; tail
= tail
->child
) {
164 switch (tail
->deref_type
) {
165 case nir_deref_type_struct
: {
166 nir_deref_struct
*deref_struct
= nir_deref_as_struct(tail
);
168 assert(deref_struct
->index
< type_get_length(node
->type
));
170 if (node
->children
[deref_struct
->index
] == NULL
)
171 node
->children
[deref_struct
->index
] =
172 deref_node_create(node
, tail
->type
, state
->dead_ctx
);
174 node
= node
->children
[deref_struct
->index
];
178 case nir_deref_type_array
: {
179 nir_deref_array
*arr
= nir_deref_as_array(tail
);
181 switch (arr
->deref_array_type
) {
182 case nir_deref_array_type_direct
:
183 /* This is possible if a loop unrolls and generates an
184 * out-of-bounds offset. We need to handle this at least
185 * somewhat gracefully.
187 if (arr
->base_offset
>= type_get_length(node
->type
))
190 if (node
->children
[arr
->base_offset
] == NULL
)
191 node
->children
[arr
->base_offset
] =
192 deref_node_create(node
, tail
->type
, state
->dead_ctx
);
194 node
= node
->children
[arr
->base_offset
];
197 case nir_deref_array_type_indirect
:
198 if (node
->indirect
== NULL
)
199 node
->indirect
= deref_node_create(node
, tail
->type
,
202 node
= node
->indirect
;
206 case nir_deref_array_type_wildcard
:
207 if (node
->wildcard
== NULL
)
208 node
->wildcard
= deref_node_create(node
, tail
->type
,
211 node
= node
->wildcard
;
216 unreachable("Invalid array deref type");
221 unreachable("Invalid deref type");
227 /* Only insert if it isn't already in the list. */
228 if (is_direct
&& state
->add_to_direct_deref_nodes
&&
229 node
->direct_derefs_link
.next
== NULL
) {
231 assert(deref
->var
!= NULL
);
232 exec_list_push_tail(&state
->direct_deref_nodes
,
233 &node
->direct_derefs_link
);
239 /* \sa foreach_deref_node_match */
241 foreach_deref_node_worker(struct deref_node
*node
, nir_deref
*deref
,
242 bool (* cb
)(struct deref_node
*node
,
243 struct lower_variables_state
*state
),
244 struct lower_variables_state
*state
)
246 if (deref
->child
== NULL
) {
247 return cb(node
, state
);
249 switch (deref
->child
->deref_type
) {
250 case nir_deref_type_array
: {
251 nir_deref_array
*arr
= nir_deref_as_array(deref
->child
);
252 assert(arr
->deref_array_type
== nir_deref_array_type_direct
);
253 if (node
->children
[arr
->base_offset
] &&
254 !foreach_deref_node_worker(node
->children
[arr
->base_offset
],
255 deref
->child
, cb
, state
))
258 if (node
->wildcard
&&
259 !foreach_deref_node_worker(node
->wildcard
,
260 deref
->child
, cb
, state
))
266 case nir_deref_type_struct
: {
267 nir_deref_struct
*str
= nir_deref_as_struct(deref
->child
);
268 return foreach_deref_node_worker(node
->children
[str
->index
],
269 deref
->child
, cb
, state
);
273 unreachable("Invalid deref child type");
278 /* Walks over every "matching" deref_node and calls the callback. A node
279 * is considered to "match" if either refers to that deref or matches up t
280 * a wildcard. In other words, the following would match a[6].foo[3].bar:
287 * The given deref must be a full-length and fully qualified (no wildcards
288 * or indirects) deref chain.
291 foreach_deref_node_match(nir_deref_var
*deref
,
292 bool (* cb
)(struct deref_node
*node
,
293 struct lower_variables_state
*state
),
294 struct lower_variables_state
*state
)
296 nir_deref_var var_deref
= *deref
;
297 var_deref
.deref
.child
= NULL
;
298 struct deref_node
*node
= get_deref_node(&var_deref
, state
);
303 return foreach_deref_node_worker(node
, &deref
->deref
, cb
, state
);
306 /* \sa deref_may_be_aliased */
308 deref_may_be_aliased_node(struct deref_node
*node
, nir_deref
*deref
,
309 struct lower_variables_state
*state
)
311 if (deref
->child
== NULL
) {
314 switch (deref
->child
->deref_type
) {
315 case nir_deref_type_array
: {
316 nir_deref_array
*arr
= nir_deref_as_array(deref
->child
);
317 if (arr
->deref_array_type
== nir_deref_array_type_indirect
)
320 /* If there is an indirect at this level, we're aliased. */
324 assert(arr
->deref_array_type
== nir_deref_array_type_direct
);
326 if (node
->children
[arr
->base_offset
] &&
327 deref_may_be_aliased_node(node
->children
[arr
->base_offset
],
328 deref
->child
, state
))
331 if (node
->wildcard
&&
332 deref_may_be_aliased_node(node
->wildcard
, deref
->child
, state
))
338 case nir_deref_type_struct
: {
339 nir_deref_struct
*str
= nir_deref_as_struct(deref
->child
);
340 if (node
->children
[str
->index
]) {
341 return deref_may_be_aliased_node(node
->children
[str
->index
],
342 deref
->child
, state
);
349 unreachable("Invalid nir_deref child type");
354 /* Returns true if there are no indirects that can ever touch this deref.
356 * For example, if the given deref is a[6].foo, then any uses of a[i].foo
357 * would cause this to return false, but a[i].bar would not affect it
358 * because it's a different structure member. A var_copy involving of
359 * a[*].bar also doesn't affect it because that can be lowered to entirely
360 * direct load/stores.
362 * We only support asking this question about fully-qualified derefs.
363 * Obviously, it's pointless to ask this about indirects, but we also
364 * rule-out wildcards. Handling Wildcard dereferences would involve
365 * checking each array index to make sure that there aren't any indirect
369 deref_may_be_aliased(nir_deref_var
*deref
,
370 struct lower_variables_state
*state
)
372 return deref_may_be_aliased_node(get_deref_node_for_var(deref
->var
, state
),
373 &deref
->deref
, state
);
377 register_load_instr(nir_intrinsic_instr
*load_instr
,
378 struct lower_variables_state
*state
)
380 struct deref_node
*node
= get_deref_node(load_instr
->variables
[0], state
);
384 if (node
->loads
== NULL
)
385 node
->loads
= _mesa_set_create(state
->dead_ctx
, _mesa_hash_pointer
,
386 _mesa_key_pointer_equal
);
388 _mesa_set_add(node
->loads
, load_instr
);
392 register_store_instr(nir_intrinsic_instr
*store_instr
,
393 struct lower_variables_state
*state
)
395 struct deref_node
*node
= get_deref_node(store_instr
->variables
[0], state
);
399 if (node
->stores
== NULL
)
400 node
->stores
= _mesa_set_create(state
->dead_ctx
, _mesa_hash_pointer
,
401 _mesa_key_pointer_equal
);
403 _mesa_set_add(node
->stores
, store_instr
);
407 register_copy_instr(nir_intrinsic_instr
*copy_instr
,
408 struct lower_variables_state
*state
)
410 for (unsigned idx
= 0; idx
< 2; idx
++) {
411 struct deref_node
*node
=
412 get_deref_node(copy_instr
->variables
[idx
], state
);
417 if (node
->copies
== NULL
)
418 node
->copies
= _mesa_set_create(state
->dead_ctx
, _mesa_hash_pointer
,
419 _mesa_key_pointer_equal
);
421 _mesa_set_add(node
->copies
, copy_instr
);
425 /* Registers all variable uses in the given block. */
427 register_variable_uses_block(nir_block
*block
, void *void_state
)
429 struct lower_variables_state
*state
= void_state
;
431 nir_foreach_instr_safe(block
, instr
) {
432 if (instr
->type
!= nir_instr_type_intrinsic
)
435 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
437 switch (intrin
->intrinsic
) {
438 case nir_intrinsic_load_var
:
439 register_load_instr(intrin
, state
);
442 case nir_intrinsic_store_var
:
443 register_store_instr(intrin
, state
);
446 case nir_intrinsic_copy_var
:
447 register_copy_instr(intrin
, state
);
458 /* Walks over all of the copy instructions to or from the given deref_node
459 * and lowers them to load/store intrinsics.
462 lower_copies_to_load_store(struct deref_node
*node
,
463 struct lower_variables_state
*state
)
468 struct set_entry
*copy_entry
;
469 set_foreach(node
->copies
, copy_entry
) {
470 nir_intrinsic_instr
*copy
= (void *)copy_entry
->key
;
472 nir_lower_var_copy_instr(copy
, state
->shader
);
474 for (unsigned i
= 0; i
< 2; ++i
) {
475 struct deref_node
*arg_node
=
476 get_deref_node(copy
->variables
[i
], state
);
478 if (arg_node
== NULL
)
481 struct set_entry
*arg_entry
= _mesa_set_search(arg_node
->copies
, copy
);
483 _mesa_set_remove(node
->copies
, arg_entry
);
486 nir_instr_remove(©
->instr
);
492 /** Pushes an SSA def onto the def stack for the given node
494 * Each node is potentially associated with a stack of SSA definitions.
495 * This stack is used for determining what SSA definition reaches a given
496 * point in the program for variable renaming. The stack is always kept in
497 * dominance-order with at most one SSA def per block. If the SSA
498 * definition on the top of the stack is in the same block as the one being
499 * pushed, the top element is replaced.
502 def_stack_push(struct deref_node
*node
, nir_ssa_def
*def
,
503 struct lower_variables_state
*state
)
505 if (node
->def_stack
== NULL
) {
506 node
->def_stack
= ralloc_array(state
->dead_ctx
, nir_ssa_def
*,
507 state
->impl
->num_blocks
);
508 node
->def_stack_tail
= node
->def_stack
- 1;
511 if (node
->def_stack_tail
>= node
->def_stack
) {
512 nir_ssa_def
*top_def
= *node
->def_stack_tail
;
514 if (def
->parent_instr
->block
== top_def
->parent_instr
->block
) {
515 /* They're in the same block, just replace the top */
516 *node
->def_stack_tail
= def
;
521 *(++node
->def_stack_tail
) = def
;
524 /* Pop the top of the def stack if it's in the given block */
526 def_stack_pop_if_in_block(struct deref_node
*node
, nir_block
*block
)
528 /* If we're popping, then we have presumably pushed at some time in the
529 * past so this should exist.
531 assert(node
->def_stack
!= NULL
);
533 /* The stack is already empty. Do nothing. */
534 if (node
->def_stack_tail
< node
->def_stack
)
537 nir_ssa_def
*def
= *node
->def_stack_tail
;
538 if (def
->parent_instr
->block
== block
)
539 node
->def_stack_tail
--;
542 /** Retrieves the SSA definition on the top of the stack for the given
543 * node, if one exists. If the stack is empty, then we return the constant
544 * initializer (if it exists) or an SSA undef.
547 get_ssa_def_for_block(struct deref_node
*node
, nir_block
*block
,
548 struct lower_variables_state
*state
)
550 /* If we have something on the stack, go ahead and return it. We're
551 * assuming that the top of the stack dominates the given block.
553 if (node
->def_stack
&& node
->def_stack_tail
>= node
->def_stack
)
554 return *node
->def_stack_tail
;
556 /* If we got here then we don't have a definition that dominates the
557 * given block. This means that we need to add an undef and use that.
559 nir_ssa_undef_instr
*undef
=
560 nir_ssa_undef_instr_create(state
->shader
,
561 glsl_get_vector_elements(node
->type
));
562 nir_instr_insert_before_cf_list(&state
->impl
->body
, &undef
->instr
);
563 def_stack_push(node
, &undef
->def
, state
);
567 /* Given a block and one of its predecessors, this function fills in the
568 * souces of the phi nodes to take SSA defs from the given predecessor.
569 * This function must be called exactly once per block/predecessor pair.
572 add_phi_sources(nir_block
*block
, nir_block
*pred
,
573 struct lower_variables_state
*state
)
575 nir_foreach_instr(block
, instr
) {
576 if (instr
->type
!= nir_instr_type_phi
)
579 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
581 struct hash_entry
*entry
=
582 _mesa_hash_table_search(state
->phi_table
, phi
);
586 struct deref_node
*node
= entry
->data
;
588 nir_phi_src
*src
= ralloc(phi
, nir_phi_src
);
590 src
->src
.is_ssa
= true;
591 src
->src
.ssa
= get_ssa_def_for_block(node
, pred
, state
);
593 _mesa_set_add(src
->src
.ssa
->uses
, instr
);
595 exec_list_push_tail(&phi
->srcs
, &src
->node
);
599 /* Performs variable renaming by doing a DFS of the dominance tree
601 * This algorithm is very similar to the one outlined in "Efficiently
602 * Computing Static Single Assignment Form and the Control Dependence
603 * Graph" by Cytron et. al. The primary difference is that we only put one
604 * SSA def on the stack per block.
607 rename_variables_block(nir_block
*block
, struct lower_variables_state
*state
)
609 nir_foreach_instr_safe(block
, instr
) {
610 if (instr
->type
== nir_instr_type_phi
) {
611 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
613 struct hash_entry
*entry
=
614 _mesa_hash_table_search(state
->phi_table
, phi
);
616 /* This can happen if we already have phi nodes in the program
617 * that were not created in this pass.
622 struct deref_node
*node
= entry
->data
;
624 def_stack_push(node
, &phi
->dest
.ssa
, state
);
625 } else if (instr
->type
== nir_instr_type_intrinsic
) {
626 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
628 switch (intrin
->intrinsic
) {
629 case nir_intrinsic_load_var
: {
630 struct deref_node
*node
=
631 get_deref_node(intrin
->variables
[0], state
);
634 /* If we hit this path then we are referencing an invalid
635 * value. Most likely, we unrolled something and are
636 * reading past the end of some array. In any case, this
637 * should result in an undefined value.
639 nir_ssa_undef_instr
*undef
=
640 nir_ssa_undef_instr_create(state
->shader
,
641 intrin
->num_components
);
643 nir_instr_insert_before(&intrin
->instr
, &undef
->instr
);
644 nir_instr_remove(&intrin
->instr
);
646 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
647 nir_src_for_ssa(&undef
->def
),
652 if (!node
->lower_to_ssa
)
655 nir_alu_instr
*mov
= nir_alu_instr_create(state
->shader
,
657 mov
->src
[0].src
.is_ssa
= true;
658 mov
->src
[0].src
.ssa
= get_ssa_def_for_block(node
, block
, state
);
659 for (unsigned i
= intrin
->num_components
; i
< 4; i
++)
660 mov
->src
[0].swizzle
[i
] = 0;
662 assert(intrin
->dest
.is_ssa
);
664 mov
->dest
.write_mask
= (1 << intrin
->num_components
) - 1;
665 nir_ssa_dest_init(&mov
->instr
, &mov
->dest
.dest
,
666 intrin
->num_components
, NULL
);
668 nir_instr_insert_before(&intrin
->instr
, &mov
->instr
);
669 nir_instr_remove(&intrin
->instr
);
671 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
672 nir_src_for_ssa(&mov
->dest
.dest
.ssa
),
677 case nir_intrinsic_store_var
: {
678 struct deref_node
*node
=
679 get_deref_node(intrin
->variables
[0], state
);
682 /* Probably an out-of-bounds array store. That should be a
684 nir_instr_remove(&intrin
->instr
);
688 if (!node
->lower_to_ssa
)
691 assert(intrin
->num_components
==
692 glsl_get_vector_elements(node
->type
));
694 assert(intrin
->src
[0].is_ssa
);
696 nir_alu_instr
*mov
= nir_alu_instr_create(state
->shader
,
698 mov
->src
[0].src
.is_ssa
= true;
699 mov
->src
[0].src
.ssa
= intrin
->src
[0].ssa
;
700 for (unsigned i
= intrin
->num_components
; i
< 4; i
++)
701 mov
->src
[0].swizzle
[i
] = 0;
703 mov
->dest
.write_mask
= (1 << intrin
->num_components
) - 1;
704 nir_ssa_dest_init(&mov
->instr
, &mov
->dest
.dest
,
705 intrin
->num_components
, NULL
);
707 nir_instr_insert_before(&intrin
->instr
, &mov
->instr
);
709 def_stack_push(node
, &mov
->dest
.dest
.ssa
, state
);
711 /* We'll wait to remove the instruction until the next pass
712 * where we pop the node we just pushed back off the stack.
723 if (block
->successors
[0])
724 add_phi_sources(block
->successors
[0], block
, state
);
725 if (block
->successors
[1])
726 add_phi_sources(block
->successors
[1], block
, state
);
728 for (unsigned i
= 0; i
< block
->num_dom_children
; ++i
)
729 rename_variables_block(block
->dom_children
[i
], state
);
731 /* Now we iterate over the instructions and pop off any SSA defs that we
732 * pushed in the first loop.
734 nir_foreach_instr_safe(block
, instr
) {
735 if (instr
->type
== nir_instr_type_phi
) {
736 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
738 struct hash_entry
*entry
=
739 _mesa_hash_table_search(state
->phi_table
, phi
);
741 /* This can happen if we already have phi nodes in the program
742 * that were not created in this pass.
747 struct deref_node
*node
= entry
->data
;
749 def_stack_pop_if_in_block(node
, block
);
750 } else if (instr
->type
== nir_instr_type_intrinsic
) {
751 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
753 if (intrin
->intrinsic
!= nir_intrinsic_store_var
)
756 struct deref_node
*node
= get_deref_node(intrin
->variables
[0], state
);
760 if (!node
->lower_to_ssa
)
763 def_stack_pop_if_in_block(node
, block
);
764 nir_instr_remove(&intrin
->instr
);
771 /* Inserts phi nodes for all variables marked lower_to_ssa
773 * This is the same algorithm as presented in "Efficiently Computing Static
774 * Single Assignment Form and the Control Dependence Graph" by Cytron et.
778 insert_phi_nodes(struct lower_variables_state
*state
)
780 NIR_VLA_ZERO(unsigned, work
, state
->impl
->num_blocks
);
781 NIR_VLA_ZERO(unsigned, has_already
, state
->impl
->num_blocks
);
784 * Since the work flags already prevent us from inserting a node that has
785 * ever been inserted into W, we don't need to use a set to represent W.
786 * Also, since no block can ever be inserted into W more than once, we know
787 * that the maximum size of W is the number of basic blocks in the
788 * function. So all we need to handle W is an array and a pointer to the
789 * next element to be inserted and the next element to be removed.
791 NIR_VLA(nir_block
*, W
, state
->impl
->num_blocks
);
793 unsigned w_start
, w_end
;
794 unsigned iter_count
= 0;
796 foreach_list_typed(struct deref_node
, node
, direct_derefs_link
,
797 &state
->direct_deref_nodes
) {
798 if (node
->stores
== NULL
)
801 if (!node
->lower_to_ssa
)
807 struct set_entry
*store_entry
;
808 set_foreach(node
->stores
, store_entry
) {
809 nir_intrinsic_instr
*store
= (nir_intrinsic_instr
*)store_entry
->key
;
810 if (work
[store
->instr
.block
->index
] < iter_count
)
811 W
[w_end
++] = store
->instr
.block
;
812 work
[store
->instr
.block
->index
] = iter_count
;
815 while (w_start
!= w_end
) {
816 nir_block
*cur
= W
[w_start
++];
817 struct set_entry
*dom_entry
;
818 set_foreach(cur
->dom_frontier
, dom_entry
) {
819 nir_block
*next
= (nir_block
*) dom_entry
->key
;
822 * If there's more than one return statement, then the end block
823 * can be a join point for some definitions. However, there are
824 * no instructions in the end block, so nothing would use those
825 * phi nodes. Of course, we couldn't place those phi nodes
826 * anyways due to the restriction of having no instructions in the
829 if (next
== state
->impl
->end_block
)
832 if (has_already
[next
->index
] < iter_count
) {
833 nir_phi_instr
*phi
= nir_phi_instr_create(state
->shader
);
834 nir_ssa_dest_init(&phi
->instr
, &phi
->dest
,
835 glsl_get_vector_elements(node
->type
), NULL
);
836 nir_instr_insert_before_block(next
, &phi
->instr
);
838 _mesa_hash_table_insert(state
->phi_table
, phi
, node
);
840 has_already
[next
->index
] = iter_count
;
841 if (work
[next
->index
] < iter_count
) {
842 work
[next
->index
] = iter_count
;
852 /** Implements a pass to lower variable uses to SSA values
854 * This path walks the list of instructions and tries to lower as many
855 * local variable load/store operations to SSA defs and uses as it can.
856 * The process involves four passes:
858 * 1) Iterate over all of the instructions and mark where each local
859 * variable deref is used in a load, store, or copy. While we're at
860 * it, we keep track of all of the fully-qualified (no wildcards) and
861 * fully-direct references we see and store them in the
862 * direct_deref_nodes hash table.
864 * 2) Walk over the the list of fully-qualified direct derefs generated in
865 * the previous pass. For each deref, we determine if it can ever be
866 * aliased, i.e. if there is an indirect reference anywhere that may
867 * refer to it. If it cannot be aliased, we mark it for lowering to an
868 * SSA value. At this point, we lower any var_copy instructions that
869 * use the given deref to load/store operations and, if the deref has a
870 * constant initializer, we go ahead and add a load_const value at the
871 * beginning of the function with the initialized value.
873 * 3) Walk over the list of derefs we plan to lower to SSA values and
874 * insert phi nodes as needed.
876 * 4) Perform "variable renaming" by replacing the load/store instructions
877 * with SSA definitions and SSA uses.
880 nir_lower_vars_to_ssa_impl(nir_function_impl
*impl
)
882 struct lower_variables_state state
;
884 state
.shader
= impl
->overload
->function
->shader
;
885 state
.dead_ctx
= ralloc_context(state
.shader
);
888 state
.deref_var_nodes
= _mesa_hash_table_create(state
.dead_ctx
,
890 _mesa_key_pointer_equal
);
891 exec_list_make_empty(&state
.direct_deref_nodes
);
892 state
.phi_table
= _mesa_hash_table_create(state
.dead_ctx
,
894 _mesa_key_pointer_equal
);
896 /* Build the initial deref structures and direct_deref_nodes table */
897 state
.add_to_direct_deref_nodes
= true;
898 nir_foreach_block(impl
, register_variable_uses_block
, &state
);
900 struct set
*outputs
= _mesa_set_create(state
.dead_ctx
,
902 _mesa_key_pointer_equal
);
904 bool progress
= false;
906 nir_metadata_require(impl
, nir_metadata_block_index
);
908 /* We're about to iterate through direct_deref_nodes. Don't modify it. */
909 state
.add_to_direct_deref_nodes
= false;
911 foreach_list_typed_safe(struct deref_node
, node
, direct_derefs_link
,
912 &state
.direct_deref_nodes
) {
913 nir_deref_var
*deref
= node
->deref
;
915 if (deref
->var
->data
.mode
!= nir_var_local
) {
916 exec_node_remove(&node
->direct_derefs_link
);
920 if (deref_may_be_aliased(deref
, &state
)) {
921 exec_node_remove(&node
->direct_derefs_link
);
925 node
->lower_to_ssa
= true;
928 if (deref
->var
->constant_initializer
) {
929 nir_load_const_instr
*load
=
930 nir_deref_get_const_initializer_load(state
.shader
, deref
);
931 nir_ssa_def_init(&load
->instr
, &load
->def
,
932 glsl_get_vector_elements(node
->type
), NULL
);
933 nir_instr_insert_before_cf_list(&impl
->body
, &load
->instr
);
934 def_stack_push(node
, &load
->def
, &state
);
937 if (deref
->var
->data
.mode
== nir_var_shader_out
)
938 _mesa_set_add(outputs
, node
);
940 foreach_deref_node_match(deref
, lower_copies_to_load_store
, &state
);
946 nir_metadata_require(impl
, nir_metadata_dominance
);
948 /* We may have lowered some copy instructions to load/store
949 * instructions. The uses from the copy instructions hav already been
950 * removed but we need to rescan to ensure that the uses from the newly
951 * added load/store instructions are registered. We need this
952 * information for phi node insertion below.
954 nir_foreach_block(impl
, register_variable_uses_block
, &state
);
956 insert_phi_nodes(&state
);
957 rename_variables_block(impl
->start_block
, &state
);
959 nir_metadata_preserve(impl
, nir_metadata_block_index
|
960 nir_metadata_dominance
);
962 ralloc_free(state
.dead_ctx
);
968 nir_lower_vars_to_ssa(nir_shader
*shader
)
970 nir_foreach_overload(shader
, overload
) {
972 nir_lower_vars_to_ssa_impl(overload
->impl
);