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
, void *mem_ctx
)
117 size_t size
= sizeof(struct deref_node
) +
118 type_get_length(type
) * sizeof(struct deref_node
*);
120 struct deref_node
*node
= rzalloc_size(mem_ctx
, 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 assert(arr
->deref_array_type
== nir_deref_array_type_direct
);
322 if (node
->children
[arr
->base_offset
] &&
323 deref_may_be_aliased_node(node
->children
[arr
->base_offset
],
324 deref
->child
, state
))
327 if (node
->wildcard
&&
328 deref_may_be_aliased_node(node
->wildcard
, deref
->child
, state
))
334 case nir_deref_type_struct
: {
335 nir_deref_struct
*str
= nir_deref_as_struct(deref
->child
);
336 if (node
->children
[str
->index
]) {
337 return deref_may_be_aliased_node(node
->children
[str
->index
],
338 deref
->child
, state
);
345 unreachable("Invalid nir_deref child type");
350 /* Returns true if there are no indirects that can ever touch this deref.
352 * For example, if the given deref is a[6].foo, then any uses of a[i].foo
353 * would cause this to return false, but a[i].bar would not affect it
354 * because it's a different structure member. A var_copy involving of
355 * a[*].bar also doesn't affect it because that can be lowered to entirely
356 * direct load/stores.
358 * We only support asking this question about fully-qualified derefs.
359 * Obviously, it's pointless to ask this about indirects, but we also
360 * rule-out wildcards. Handling Wildcard dereferences would involve
361 * checking each array index to make sure that there aren't any indirect
365 deref_may_be_aliased(nir_deref_var
*deref
,
366 struct lower_variables_state
*state
)
368 return deref_may_be_aliased_node(get_deref_node_for_var(deref
->var
, state
),
369 &deref
->deref
, state
);
373 register_load_instr(nir_intrinsic_instr
*load_instr
,
374 struct lower_variables_state
*state
)
376 struct deref_node
*node
= get_deref_node(load_instr
->variables
[0], state
);
380 if (node
->loads
== NULL
)
381 node
->loads
= _mesa_set_create(state
->dead_ctx
, _mesa_hash_pointer
,
382 _mesa_key_pointer_equal
);
384 _mesa_set_add(node
->loads
, load_instr
);
388 register_store_instr(nir_intrinsic_instr
*store_instr
,
389 struct lower_variables_state
*state
)
391 struct deref_node
*node
= get_deref_node(store_instr
->variables
[0], state
);
395 if (node
->stores
== NULL
)
396 node
->stores
= _mesa_set_create(state
->dead_ctx
, _mesa_hash_pointer
,
397 _mesa_key_pointer_equal
);
399 _mesa_set_add(node
->stores
, store_instr
);
403 register_copy_instr(nir_intrinsic_instr
*copy_instr
,
404 struct lower_variables_state
*state
)
406 for (unsigned idx
= 0; idx
< 2; idx
++) {
407 struct deref_node
*node
=
408 get_deref_node(copy_instr
->variables
[idx
], state
);
413 if (node
->copies
== NULL
)
414 node
->copies
= _mesa_set_create(state
->dead_ctx
, _mesa_hash_pointer
,
415 _mesa_key_pointer_equal
);
417 _mesa_set_add(node
->copies
, copy_instr
);
421 /* Registers all variable uses in the given block. */
423 register_variable_uses_block(nir_block
*block
, void *void_state
)
425 struct lower_variables_state
*state
= void_state
;
427 nir_foreach_instr_safe(block
, instr
) {
428 if (instr
->type
!= nir_instr_type_intrinsic
)
431 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
433 switch (intrin
->intrinsic
) {
434 case nir_intrinsic_load_var
:
435 register_load_instr(intrin
, state
);
438 case nir_intrinsic_store_var
:
439 register_store_instr(intrin
, state
);
442 case nir_intrinsic_copy_var
:
443 register_copy_instr(intrin
, state
);
454 /* Walks over all of the copy instructions to or from the given deref_node
455 * and lowers them to load/store intrinsics.
458 lower_copies_to_load_store(struct deref_node
*node
,
459 struct lower_variables_state
*state
)
464 struct set_entry
*copy_entry
;
465 set_foreach(node
->copies
, copy_entry
) {
466 nir_intrinsic_instr
*copy
= (void *)copy_entry
->key
;
468 nir_lower_var_copy_instr(copy
, state
->mem_ctx
);
470 for (unsigned i
= 0; i
< 2; ++i
) {
471 struct deref_node
*arg_node
=
472 get_deref_node(copy
->variables
[i
], state
);
474 if (arg_node
== NULL
)
477 struct set_entry
*arg_entry
= _mesa_set_search(arg_node
->copies
, copy
);
479 _mesa_set_remove(node
->copies
, arg_entry
);
482 nir_instr_remove(©
->instr
);
488 /* Returns a load_const instruction that represents the constant
489 * initializer for the given deref chain. The caller is responsible for
490 * ensuring that there actually is a constant initializer.
492 static nir_load_const_instr
*
493 get_const_initializer_load(const nir_deref_var
*deref
,
494 struct lower_variables_state
*state
)
496 nir_constant
*constant
= deref
->var
->constant_initializer
;
497 const nir_deref
*tail
= &deref
->deref
;
498 unsigned matrix_offset
= 0;
499 while (tail
->child
) {
500 switch (tail
->child
->deref_type
) {
501 case nir_deref_type_array
: {
502 nir_deref_array
*arr
= nir_deref_as_array(tail
->child
);
503 assert(arr
->deref_array_type
== nir_deref_array_type_direct
);
504 if (glsl_type_is_matrix(tail
->type
)) {
505 assert(arr
->deref
.child
== NULL
);
506 matrix_offset
= arr
->base_offset
;
508 constant
= constant
->elements
[arr
->base_offset
];
513 case nir_deref_type_struct
: {
514 constant
= constant
->elements
[nir_deref_as_struct(tail
->child
)->index
];
519 unreachable("Invalid deref child type");
525 nir_load_const_instr
*load
=
526 nir_load_const_instr_create(state
->mem_ctx
,
527 glsl_get_vector_elements(tail
->type
));
529 matrix_offset
*= load
->def
.num_components
;
530 for (unsigned i
= 0; i
< load
->def
.num_components
; i
++) {
531 switch (glsl_get_base_type(tail
->type
)) {
532 case GLSL_TYPE_FLOAT
:
535 load
->value
.u
[i
] = constant
->value
.u
[matrix_offset
+ i
];
538 load
->value
.u
[i
] = constant
->value
.b
[matrix_offset
+ i
] ?
539 NIR_TRUE
: NIR_FALSE
;
542 unreachable("Invalid immediate type");
549 /** Pushes an SSA def onto the def stack for the given node
551 * Each node is potentially associated with a stack of SSA definitions.
552 * This stack is used for determining what SSA definition reaches a given
553 * point in the program for variable renaming. The stack is always kept in
554 * dominance-order with at most one SSA def per block. If the SSA
555 * definition on the top of the stack is in the same block as the one being
556 * pushed, the top element is replaced.
559 def_stack_push(struct deref_node
*node
, nir_ssa_def
*def
,
560 struct lower_variables_state
*state
)
562 if (node
->def_stack
== NULL
) {
563 node
->def_stack
= ralloc_array(state
->dead_ctx
, nir_ssa_def
*,
564 state
->impl
->num_blocks
);
565 node
->def_stack_tail
= node
->def_stack
- 1;
568 if (node
->def_stack_tail
>= node
->def_stack
) {
569 nir_ssa_def
*top_def
= *node
->def_stack_tail
;
571 if (def
->parent_instr
->block
== top_def
->parent_instr
->block
) {
572 /* They're in the same block, just replace the top */
573 *node
->def_stack_tail
= def
;
578 *(++node
->def_stack_tail
) = def
;
581 /* Pop the top of the def stack if it's in the given block */
583 def_stack_pop_if_in_block(struct deref_node
*node
, nir_block
*block
)
585 /* If we're popping, then we have presumably pushed at some time in the
586 * past so this should exist.
588 assert(node
->def_stack
!= NULL
);
590 /* The stack is already empty. Do nothing. */
591 if (node
->def_stack_tail
< node
->def_stack
)
594 nir_ssa_def
*def
= *node
->def_stack_tail
;
595 if (def
->parent_instr
->block
== block
)
596 node
->def_stack_tail
--;
599 /** Retrieves the SSA definition on the top of the stack for the given
600 * node, if one exists. If the stack is empty, then we return the constant
601 * initializer (if it exists) or an SSA undef.
604 get_ssa_def_for_block(struct deref_node
*node
, nir_block
*block
,
605 struct lower_variables_state
*state
)
607 /* If we have something on the stack, go ahead and return it. We're
608 * assuming that the top of the stack dominates the given block.
610 if (node
->def_stack
&& node
->def_stack_tail
>= node
->def_stack
)
611 return *node
->def_stack_tail
;
613 /* If we got here then we don't have a definition that dominates the
614 * given block. This means that we need to add an undef and use that.
616 nir_ssa_undef_instr
*undef
=
617 nir_ssa_undef_instr_create(state
->mem_ctx
,
618 glsl_get_vector_elements(node
->type
));
619 nir_instr_insert_before_cf_list(&state
->impl
->body
, &undef
->instr
);
620 def_stack_push(node
, &undef
->def
, state
);
624 /* Given a block and one of its predecessors, this function fills in the
625 * souces of the phi nodes to take SSA defs from the given predecessor.
626 * This function must be called exactly once per block/predecessor pair.
629 add_phi_sources(nir_block
*block
, nir_block
*pred
,
630 struct lower_variables_state
*state
)
632 nir_foreach_instr(block
, instr
) {
633 if (instr
->type
!= nir_instr_type_phi
)
636 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
638 struct hash_entry
*entry
=
639 _mesa_hash_table_search(state
->phi_table
, phi
);
643 struct deref_node
*node
= entry
->data
;
645 nir_phi_src
*src
= ralloc(phi
, nir_phi_src
);
647 src
->src
.is_ssa
= true;
648 src
->src
.ssa
= get_ssa_def_for_block(node
, pred
, state
);
650 _mesa_set_add(src
->src
.ssa
->uses
, instr
);
652 exec_list_push_tail(&phi
->srcs
, &src
->node
);
656 /* Performs variable renaming by doing a DFS of the dominance tree
658 * This algorithm is very similar to the one outlined in "Efficiently
659 * Computing Static Single Assignment Form and the Control Dependence
660 * Graph" by Cytron et. al. The primary difference is that we only put one
661 * SSA def on the stack per block.
664 rename_variables_block(nir_block
*block
, struct lower_variables_state
*state
)
666 nir_foreach_instr_safe(block
, instr
) {
667 if (instr
->type
== nir_instr_type_phi
) {
668 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
670 struct hash_entry
*entry
=
671 _mesa_hash_table_search(state
->phi_table
, phi
);
673 /* This can happen if we already have phi nodes in the program
674 * that were not created in this pass.
679 struct deref_node
*node
= entry
->data
;
681 def_stack_push(node
, &phi
->dest
.ssa
, state
);
682 } else if (instr
->type
== nir_instr_type_intrinsic
) {
683 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
685 switch (intrin
->intrinsic
) {
686 case nir_intrinsic_load_var
: {
687 struct deref_node
*node
=
688 get_deref_node(intrin
->variables
[0], state
);
691 /* If we hit this path then we are referencing an invalid
692 * value. Most likely, we unrolled something and are
693 * reading past the end of some array. In any case, this
694 * should result in an undefined value.
696 nir_ssa_undef_instr
*undef
=
697 nir_ssa_undef_instr_create(state
->mem_ctx
,
698 intrin
->num_components
);
700 nir_instr_insert_before(&intrin
->instr
, &undef
->instr
);
701 nir_instr_remove(&intrin
->instr
);
703 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
704 nir_src_for_ssa(&undef
->def
),
709 if (!node
->lower_to_ssa
)
712 nir_alu_instr
*mov
= nir_alu_instr_create(state
->mem_ctx
,
714 mov
->src
[0].src
.is_ssa
= true;
715 mov
->src
[0].src
.ssa
= get_ssa_def_for_block(node
, block
, state
);
716 for (unsigned i
= intrin
->num_components
; i
< 4; i
++)
717 mov
->src
[0].swizzle
[i
] = 0;
719 assert(intrin
->dest
.is_ssa
);
721 mov
->dest
.write_mask
= (1 << intrin
->num_components
) - 1;
722 nir_ssa_dest_init(&mov
->instr
, &mov
->dest
.dest
,
723 intrin
->num_components
, NULL
);
725 nir_instr_insert_before(&intrin
->instr
, &mov
->instr
);
726 nir_instr_remove(&intrin
->instr
);
728 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
729 nir_src_for_ssa(&mov
->dest
.dest
.ssa
),
734 case nir_intrinsic_store_var
: {
735 struct deref_node
*node
=
736 get_deref_node(intrin
->variables
[0], state
);
739 /* Probably an out-of-bounds array store. That should be a
741 nir_instr_remove(&intrin
->instr
);
745 if (!node
->lower_to_ssa
)
748 assert(intrin
->num_components
==
749 glsl_get_vector_elements(node
->type
));
751 assert(intrin
->src
[0].is_ssa
);
753 nir_alu_instr
*mov
= nir_alu_instr_create(state
->mem_ctx
,
755 mov
->src
[0].src
.is_ssa
= true;
756 mov
->src
[0].src
.ssa
= intrin
->src
[0].ssa
;
757 for (unsigned i
= intrin
->num_components
; i
< 4; i
++)
758 mov
->src
[0].swizzle
[i
] = 0;
760 mov
->dest
.write_mask
= (1 << intrin
->num_components
) - 1;
761 nir_ssa_dest_init(&mov
->instr
, &mov
->dest
.dest
,
762 intrin
->num_components
, NULL
);
764 nir_instr_insert_before(&intrin
->instr
, &mov
->instr
);
766 def_stack_push(node
, &mov
->dest
.dest
.ssa
, state
);
768 /* We'll wait to remove the instruction until the next pass
769 * where we pop the node we just pushed back off the stack.
780 if (block
->successors
[0])
781 add_phi_sources(block
->successors
[0], block
, state
);
782 if (block
->successors
[1])
783 add_phi_sources(block
->successors
[1], block
, state
);
785 for (unsigned i
= 0; i
< block
->num_dom_children
; ++i
)
786 rename_variables_block(block
->dom_children
[i
], state
);
788 /* Now we iterate over the instructions and pop off any SSA defs that we
789 * pushed in the first loop.
791 nir_foreach_instr_safe(block
, instr
) {
792 if (instr
->type
== nir_instr_type_phi
) {
793 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
795 struct hash_entry
*entry
=
796 _mesa_hash_table_search(state
->phi_table
, phi
);
798 /* This can happen if we already have phi nodes in the program
799 * that were not created in this pass.
804 struct deref_node
*node
= entry
->data
;
806 def_stack_pop_if_in_block(node
, block
);
807 } else if (instr
->type
== nir_instr_type_intrinsic
) {
808 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
810 if (intrin
->intrinsic
!= nir_intrinsic_store_var
)
813 struct deref_node
*node
= get_deref_node(intrin
->variables
[0], state
);
817 if (!node
->lower_to_ssa
)
820 def_stack_pop_if_in_block(node
, block
);
821 nir_instr_remove(&intrin
->instr
);
828 /* Inserts phi nodes for all variables marked lower_to_ssa
830 * This is the same algorithm as presented in "Efficiently Computing Static
831 * Single Assignment Form and the Control Dependence Graph" by Cytron et.
835 insert_phi_nodes(struct lower_variables_state
*state
)
837 NIR_VLA_ZERO(unsigned, work
, state
->impl
->num_blocks
);
838 NIR_VLA_ZERO(unsigned, has_already
, state
->impl
->num_blocks
);
841 * Since the work flags already prevent us from inserting a node that has
842 * ever been inserted into W, we don't need to use a set to represent W.
843 * Also, since no block can ever be inserted into W more than once, we know
844 * that the maximum size of W is the number of basic blocks in the
845 * function. So all we need to handle W is an array and a pointer to the
846 * next element to be inserted and the next element to be removed.
848 NIR_VLA(nir_block
*, W
, state
->impl
->num_blocks
);
850 unsigned w_start
, w_end
;
851 unsigned iter_count
= 0;
853 foreach_list_typed(struct deref_node
, node
, direct_derefs_link
,
854 &state
->direct_deref_nodes
) {
855 if (node
->stores
== NULL
)
858 if (!node
->lower_to_ssa
)
864 struct set_entry
*store_entry
;
865 set_foreach(node
->stores
, store_entry
) {
866 nir_intrinsic_instr
*store
= (nir_intrinsic_instr
*)store_entry
->key
;
867 if (work
[store
->instr
.block
->index
] < iter_count
)
868 W
[w_end
++] = store
->instr
.block
;
869 work
[store
->instr
.block
->index
] = iter_count
;
872 while (w_start
!= w_end
) {
873 nir_block
*cur
= W
[w_start
++];
874 struct set_entry
*dom_entry
;
875 set_foreach(cur
->dom_frontier
, dom_entry
) {
876 nir_block
*next
= (nir_block
*) dom_entry
->key
;
879 * If there's more than one return statement, then the end block
880 * can be a join point for some definitions. However, there are
881 * no instructions in the end block, so nothing would use those
882 * phi nodes. Of course, we couldn't place those phi nodes
883 * anyways due to the restriction of having no instructions in the
886 if (next
== state
->impl
->end_block
)
889 if (has_already
[next
->index
] < iter_count
) {
890 nir_phi_instr
*phi
= nir_phi_instr_create(state
->mem_ctx
);
891 nir_ssa_dest_init(&phi
->instr
, &phi
->dest
,
892 glsl_get_vector_elements(node
->type
), NULL
);
893 nir_instr_insert_before_block(next
, &phi
->instr
);
895 _mesa_hash_table_insert(state
->phi_table
, phi
, node
);
897 has_already
[next
->index
] = iter_count
;
898 if (work
[next
->index
] < iter_count
) {
899 work
[next
->index
] = iter_count
;
909 /** Implements a pass to lower variable uses to SSA values
911 * This path walks the list of instructions and tries to lower as many
912 * local variable load/store operations to SSA defs and uses as it can.
913 * The process involves four passes:
915 * 1) Iterate over all of the instructions and mark where each local
916 * variable deref is used in a load, store, or copy. While we're at
917 * it, we keep track of all of the fully-qualified (no wildcards) and
918 * fully-direct references we see and store them in the
919 * direct_deref_nodes hash table.
921 * 2) Walk over the the list of fully-qualified direct derefs generated in
922 * the previous pass. For each deref, we determine if it can ever be
923 * aliased, i.e. if there is an indirect reference anywhere that may
924 * refer to it. If it cannot be aliased, we mark it for lowering to an
925 * SSA value. At this point, we lower any var_copy instructions that
926 * use the given deref to load/store operations and, if the deref has a
927 * constant initializer, we go ahead and add a load_const value at the
928 * beginning of the function with the initialized value.
930 * 3) Walk over the list of derefs we plan to lower to SSA values and
931 * insert phi nodes as needed.
933 * 4) Perform "variable renaming" by replacing the load/store instructions
934 * with SSA definitions and SSA uses.
937 nir_lower_vars_to_ssa_impl(nir_function_impl
*impl
)
939 struct lower_variables_state state
;
941 state
.mem_ctx
= ralloc_parent(impl
);
942 state
.dead_ctx
= ralloc_context(state
.mem_ctx
);
945 state
.deref_var_nodes
= _mesa_hash_table_create(state
.dead_ctx
,
947 _mesa_key_pointer_equal
);
948 exec_list_make_empty(&state
.direct_deref_nodes
);
949 state
.phi_table
= _mesa_hash_table_create(state
.dead_ctx
,
951 _mesa_key_pointer_equal
);
953 /* Build the initial deref structures and direct_deref_nodes table */
954 state
.add_to_direct_deref_nodes
= true;
955 nir_foreach_block(impl
, register_variable_uses_block
, &state
);
957 struct set
*outputs
= _mesa_set_create(state
.dead_ctx
,
959 _mesa_key_pointer_equal
);
961 bool progress
= false;
963 nir_metadata_require(impl
, nir_metadata_block_index
);
965 /* We're about to iterate through direct_deref_nodes. Don't modify it. */
966 state
.add_to_direct_deref_nodes
= false;
968 foreach_list_typed_safe(struct deref_node
, node
, direct_derefs_link
,
969 &state
.direct_deref_nodes
) {
970 nir_deref_var
*deref
= node
->deref
;
972 if (deref
->var
->data
.mode
!= nir_var_local
) {
973 exec_node_remove(&node
->direct_derefs_link
);
977 if (deref_may_be_aliased(deref
, &state
)) {
978 exec_node_remove(&node
->direct_derefs_link
);
982 node
->lower_to_ssa
= true;
985 if (deref
->var
->constant_initializer
) {
986 nir_load_const_instr
*load
= get_const_initializer_load(deref
, &state
);
987 nir_ssa_def_init(&load
->instr
, &load
->def
,
988 glsl_get_vector_elements(node
->type
), NULL
);
989 nir_instr_insert_before_cf_list(&impl
->body
, &load
->instr
);
990 def_stack_push(node
, &load
->def
, &state
);
993 if (deref
->var
->data
.mode
== nir_var_shader_out
)
994 _mesa_set_add(outputs
, node
);
996 foreach_deref_node_match(deref
, lower_copies_to_load_store
, &state
);
1002 nir_metadata_require(impl
, nir_metadata_dominance
);
1004 /* We may have lowered some copy instructions to load/store
1005 * instructions. The uses from the copy instructions hav already been
1006 * removed but we need to rescan to ensure that the uses from the newly
1007 * added load/store instructions are registered. We need this
1008 * information for phi node insertion below.
1010 nir_foreach_block(impl
, register_variable_uses_block
, &state
);
1012 insert_phi_nodes(&state
);
1013 rename_variables_block(impl
->start_block
, &state
);
1015 nir_metadata_preserve(impl
, nir_metadata_block_index
|
1016 nir_metadata_dominance
);
1018 ralloc_free(state
.dead_ctx
);
1024 nir_lower_vars_to_ssa(nir_shader
*shader
)
1026 nir_foreach_overload(shader
, overload
) {
1028 nir_lower_vars_to_ssa_impl(overload
->impl
);