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)
29 #include "nir_builder.h"
30 #include "nir_deref.h"
31 #include "nir_phi_builder.h"
36 struct deref_node
*parent
;
37 const struct glsl_type
*type
;
41 /* Only valid for things that end up in the direct list.
42 * Note that multiple nir_deref_instrs may correspond to this node, but
43 * they will all be equivalent, so any is as good as the other.
46 struct exec_node direct_derefs_link
;
52 struct nir_phi_builder_value
*pb_value
;
54 /* True if this node is fully direct. If set, it must be in the children
55 * array of its parent.
59 struct deref_node
*wildcard
;
60 struct deref_node
*indirect
;
61 struct deref_node
*children
[0];
64 struct lower_variables_state
{
67 nir_function_impl
*impl
;
69 /* A hash table mapping variables to deref_node data */
70 struct hash_table
*deref_var_nodes
;
72 /* A hash table mapping fully-qualified direct dereferences, i.e.
73 * dereferences with no indirect or wildcard array dereferences, to
76 * At the moment, we only lower loads, stores, and copies that can be
77 * trivially lowered to loads and stores, i.e. copies with no indirects
78 * and no wildcards. If a part of a variable that is being loaded from
79 * and/or stored into is also involved in a copy operation with
80 * wildcards, then we lower that copy operation to loads and stores, but
81 * otherwise we leave copies with wildcards alone. Since the only derefs
82 * used in these loads, stores, and trivial copies are ones with no
83 * wildcards and no indirects, these are precisely the derefs that we
84 * can actually consider lowering.
86 struct exec_list direct_deref_nodes
;
88 /* Controls whether get_deref_node will add variables to the
89 * direct_deref_nodes table. This is turned on when we are initially
90 * scanning for load/store instructions. It is then turned off so we
91 * don't accidentally change the direct_deref_nodes table while we're
92 * iterating throug it.
94 bool add_to_direct_deref_nodes
;
96 struct nir_phi_builder
*phi_builder
;
99 static struct deref_node
*
100 deref_node_create(struct deref_node
*parent
,
101 const struct glsl_type
*type
,
102 bool is_direct
, void *mem_ctx
)
104 size_t size
= sizeof(struct deref_node
) +
105 glsl_get_length(type
) * sizeof(struct deref_node
*);
107 struct deref_node
*node
= rzalloc_size(mem_ctx
, size
);
109 node
->parent
= parent
;
110 exec_node_init(&node
->direct_derefs_link
);
111 node
->is_direct
= is_direct
;
116 /* Returns the deref node associated with the given variable. This will be
117 * the root of the tree representing all of the derefs of the given variable.
119 static struct deref_node
*
120 get_deref_node_for_var(nir_variable
*var
, struct lower_variables_state
*state
)
122 struct deref_node
*node
;
124 struct hash_entry
*var_entry
=
125 _mesa_hash_table_search(state
->deref_var_nodes
, var
);
128 return var_entry
->data
;
130 node
= deref_node_create(NULL
, var
->type
, true, state
->dead_ctx
);
131 _mesa_hash_table_insert(state
->deref_var_nodes
, var
, node
);
136 /* Gets the deref_node for the given deref chain and creates it if it
137 * doesn't yet exist. If the deref is fully-qualified and direct and
138 * state->add_to_direct_deref_nodes is true, it will be added to the hash
139 * table of of fully-qualified direct derefs.
141 static struct deref_node
*
142 get_deref_node_recur(nir_deref_instr
*deref
,
143 struct lower_variables_state
*state
)
145 if (deref
->deref_type
== nir_deref_type_var
)
146 return get_deref_node_for_var(deref
->var
, state
);
148 struct deref_node
*parent
=
149 get_deref_node_recur(nir_deref_instr_parent(deref
), state
);
151 switch (deref
->deref_type
) {
152 case nir_deref_type_struct
:
153 assert(glsl_type_is_struct(parent
->type
));
154 assert(deref
->strct
.index
< glsl_get_length(parent
->type
));
156 if (parent
->children
[deref
->strct
.index
] == NULL
) {
157 parent
->children
[deref
->strct
.index
] =
158 deref_node_create(parent
, deref
->type
, parent
->is_direct
,
162 return parent
->children
[deref
->strct
.index
];
164 case nir_deref_type_array
: {
165 if (nir_src_is_const(deref
->arr
.index
)) {
166 uint32_t index
= nir_src_as_uint(deref
->arr
.index
);
167 /* This is possible if a loop unrolls and generates an
168 * out-of-bounds offset. We need to handle this at least
169 * somewhat gracefully.
171 if (index
>= glsl_get_length(parent
->type
))
174 if (parent
->children
[index
] == NULL
) {
175 parent
->children
[index
] =
176 deref_node_create(parent
, deref
->type
, parent
->is_direct
,
180 return parent
->children
[index
];
182 if (parent
->indirect
== NULL
) {
184 deref_node_create(parent
, deref
->type
, false, state
->dead_ctx
);
187 return parent
->indirect
;
192 case nir_deref_type_array_wildcard
:
193 if (parent
->wildcard
== NULL
) {
195 deref_node_create(parent
, deref
->type
, false, state
->dead_ctx
);
198 return parent
->wildcard
;
201 unreachable("Invalid deref type");
205 static struct deref_node
*
206 get_deref_node(nir_deref_instr
*deref
, struct lower_variables_state
*state
)
208 /* This pass only works on local variables. Just ignore any derefs with
211 if (deref
->mode
!= nir_var_local
)
214 struct deref_node
*node
= get_deref_node_recur(deref
, state
);
218 /* Insert the node in the direct derefs list. We only do this if it's not
219 * already in the list and we only bother for deref nodes which are used
220 * directly in a load or store.
222 if (node
->is_direct
&& state
->add_to_direct_deref_nodes
&&
223 node
->direct_derefs_link
.next
== NULL
) {
224 nir_deref_path_init(&node
->path
, deref
, state
->dead_ctx
);
225 assert(deref
->var
!= NULL
);
226 exec_list_push_tail(&state
->direct_deref_nodes
,
227 &node
->direct_derefs_link
);
233 /* \sa foreach_deref_node_match */
235 foreach_deref_node_worker(struct deref_node
*node
, nir_deref_instr
**path
,
236 void (* cb
)(struct deref_node
*node
,
237 struct lower_variables_state
*state
),
238 struct lower_variables_state
*state
)
245 switch ((*path
)->deref_type
) {
246 case nir_deref_type_struct
:
247 if (node
->children
[(*path
)->strct
.index
]) {
248 foreach_deref_node_worker(node
->children
[(*path
)->strct
.index
],
249 path
+ 1, cb
, state
);
253 case nir_deref_type_array
: {
254 uint32_t index
= nir_src_as_uint((*path
)->arr
.index
);
256 if (node
->children
[index
]) {
257 foreach_deref_node_worker(node
->children
[index
],
258 path
+ 1, cb
, state
);
261 if (node
->wildcard
) {
262 foreach_deref_node_worker(node
->wildcard
,
263 path
+ 1, cb
, state
);
269 unreachable("Unsupported deref type");
273 /* Walks over every "matching" deref_node and calls the callback. A node
274 * is considered to "match" if either refers to that deref or matches up t
275 * a wildcard. In other words, the following would match a[6].foo[3].bar:
282 * The given deref must be a full-length and fully qualified (no wildcards
283 * or indirects) deref chain.
286 foreach_deref_node_match(nir_deref_path
*path
,
287 void (* cb
)(struct deref_node
*node
,
288 struct lower_variables_state
*state
),
289 struct lower_variables_state
*state
)
291 assert(path
->path
[0]->deref_type
== nir_deref_type_var
);
292 struct deref_node
*node
= get_deref_node_for_var(path
->path
[0]->var
, state
);
297 foreach_deref_node_worker(node
, &path
->path
[1], cb
, state
);
300 /* \sa deref_may_be_aliased */
302 path_may_be_aliased_node(struct deref_node
*node
, nir_deref_instr
**path
,
303 struct lower_variables_state
*state
)
308 switch ((*path
)->deref_type
) {
309 case nir_deref_type_struct
:
310 if (node
->children
[(*path
)->strct
.index
]) {
311 return path_may_be_aliased_node(node
->children
[(*path
)->strct
.index
],
317 case nir_deref_type_array
: {
318 if (!nir_src_is_const((*path
)->arr
.index
))
321 uint32_t index
= nir_src_as_uint((*path
)->arr
.index
);
323 /* If there is an indirect at this level, we're aliased. */
327 if (node
->children
[index
] &&
328 path_may_be_aliased_node(node
->children
[index
],
332 if (node
->wildcard
&&
333 path_may_be_aliased_node(node
->wildcard
, path
+ 1, state
))
340 unreachable("Unsupported deref type");
344 /* Returns true if there are no indirects that can ever touch this deref.
346 * For example, if the given deref is a[6].foo, then any uses of a[i].foo
347 * would cause this to return false, but a[i].bar would not affect it
348 * because it's a different structure member. A var_copy involving of
349 * a[*].bar also doesn't affect it because that can be lowered to entirely
350 * direct load/stores.
352 * We only support asking this question about fully-qualified derefs.
353 * Obviously, it's pointless to ask this about indirects, but we also
354 * rule-out wildcards. Handling Wildcard dereferences would involve
355 * checking each array index to make sure that there aren't any indirect
359 path_may_be_aliased(nir_deref_path
*path
,
360 struct lower_variables_state
*state
)
362 assert(path
->path
[0]->deref_type
== nir_deref_type_var
);
363 nir_variable
*var
= path
->path
[0]->var
;
365 return path_may_be_aliased_node(get_deref_node_for_var(var
, state
),
366 &path
->path
[1], state
);
370 register_load_instr(nir_intrinsic_instr
*load_instr
,
371 struct lower_variables_state
*state
)
373 nir_deref_instr
*deref
= nir_src_as_deref(load_instr
->src
[0]);
374 struct deref_node
*node
= get_deref_node(deref
, state
);
378 if (node
->loads
== NULL
)
379 node
->loads
= _mesa_set_create(state
->dead_ctx
, _mesa_hash_pointer
,
380 _mesa_key_pointer_equal
);
382 _mesa_set_add(node
->loads
, load_instr
);
386 register_store_instr(nir_intrinsic_instr
*store_instr
,
387 struct lower_variables_state
*state
)
389 nir_deref_instr
*deref
= nir_src_as_deref(store_instr
->src
[0]);
390 struct deref_node
*node
= get_deref_node(deref
, state
);
394 if (node
->stores
== NULL
)
395 node
->stores
= _mesa_set_create(state
->dead_ctx
, _mesa_hash_pointer
,
396 _mesa_key_pointer_equal
);
398 _mesa_set_add(node
->stores
, store_instr
);
402 register_copy_instr(nir_intrinsic_instr
*copy_instr
,
403 struct lower_variables_state
*state
)
405 for (unsigned idx
= 0; idx
< 2; idx
++) {
406 nir_deref_instr
*deref
= nir_src_as_deref(copy_instr
->src
[idx
]);
407 struct deref_node
*node
= get_deref_node(deref
, state
);
411 if (node
->copies
== NULL
)
412 node
->copies
= _mesa_set_create(state
->dead_ctx
, _mesa_hash_pointer
,
413 _mesa_key_pointer_equal
);
415 _mesa_set_add(node
->copies
, copy_instr
);
420 register_variable_uses(nir_function_impl
*impl
,
421 struct lower_variables_state
*state
)
423 nir_foreach_block(block
, impl
) {
424 nir_foreach_instr_safe(instr
, block
) {
425 if (instr
->type
!= nir_instr_type_intrinsic
)
428 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
430 switch (intrin
->intrinsic
) {
431 case nir_intrinsic_load_deref
:
432 register_load_instr(intrin
, state
);
435 case nir_intrinsic_store_deref
:
436 register_store_instr(intrin
, state
);
439 case nir_intrinsic_copy_deref
:
440 register_copy_instr(intrin
, state
);
450 /* Walks over all of the copy instructions to or from the given deref_node
451 * and lowers them to load/store intrinsics.
454 lower_copies_to_load_store(struct deref_node
*node
,
455 struct lower_variables_state
*state
)
461 nir_builder_init(&b
, state
->impl
);
463 set_foreach(node
->copies
, copy_entry
) {
464 nir_intrinsic_instr
*copy
= (void *)copy_entry
->key
;
466 nir_lower_deref_copy_instr(&b
, copy
);
468 for (unsigned i
= 0; i
< 2; ++i
) {
469 nir_deref_instr
*arg_deref
= nir_src_as_deref(copy
->src
[i
]);
470 struct deref_node
*arg_node
= get_deref_node(arg_deref
, state
);
472 /* Only bother removing copy entries for other nodes */
473 if (arg_node
== NULL
|| arg_node
== node
)
476 struct set_entry
*arg_entry
= _mesa_set_search(arg_node
->copies
, copy
);
478 _mesa_set_remove(arg_node
->copies
, arg_entry
);
481 nir_instr_remove(©
->instr
);
487 /* Performs variable renaming
489 * This algorithm is very similar to the one outlined in "Efficiently
490 * Computing Static Single Assignment Form and the Control Dependence
491 * Graph" by Cytron et al. The primary difference is that we only put one
492 * SSA def on the stack per block.
495 rename_variables(struct lower_variables_state
*state
)
498 nir_builder_init(&b
, state
->impl
);
500 nir_foreach_block(block
, state
->impl
) {
501 nir_foreach_instr_safe(instr
, block
) {
502 if (instr
->type
!= nir_instr_type_intrinsic
)
505 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
507 switch (intrin
->intrinsic
) {
508 case nir_intrinsic_load_deref
: {
509 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
510 if (deref
->mode
!= nir_var_local
)
513 struct deref_node
*node
= get_deref_node(deref
, state
);
515 /* If we hit this path then we are referencing an invalid
516 * value. Most likely, we unrolled something and are
517 * reading past the end of some array. In any case, this
518 * should result in an undefined value.
520 nir_ssa_undef_instr
*undef
=
521 nir_ssa_undef_instr_create(state
->shader
,
522 intrin
->num_components
,
523 intrin
->dest
.ssa
.bit_size
);
525 nir_instr_insert_before(&intrin
->instr
, &undef
->instr
);
526 nir_instr_remove(&intrin
->instr
);
528 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
529 nir_src_for_ssa(&undef
->def
));
533 if (!node
->lower_to_ssa
)
536 nir_alu_instr
*mov
= nir_alu_instr_create(state
->shader
,
538 mov
->src
[0].src
= nir_src_for_ssa(
539 nir_phi_builder_value_get_block_def(node
->pb_value
, block
));
540 for (unsigned i
= intrin
->num_components
; i
< 4; i
++)
541 mov
->src
[0].swizzle
[i
] = 0;
543 assert(intrin
->dest
.is_ssa
);
545 mov
->dest
.write_mask
= (1 << intrin
->num_components
) - 1;
546 nir_ssa_dest_init(&mov
->instr
, &mov
->dest
.dest
,
547 intrin
->num_components
,
548 intrin
->dest
.ssa
.bit_size
, NULL
);
550 nir_instr_insert_before(&intrin
->instr
, &mov
->instr
);
551 nir_instr_remove(&intrin
->instr
);
553 nir_ssa_def_rewrite_uses(&intrin
->dest
.ssa
,
554 nir_src_for_ssa(&mov
->dest
.dest
.ssa
));
558 case nir_intrinsic_store_deref
: {
559 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
560 if (deref
->mode
!= nir_var_local
)
563 struct deref_node
*node
= get_deref_node(deref
, state
);
565 assert(intrin
->src
[1].is_ssa
);
566 nir_ssa_def
*value
= intrin
->src
[1].ssa
;
569 /* Probably an out-of-bounds array store. That should be a
571 nir_instr_remove(&intrin
->instr
);
575 if (!node
->lower_to_ssa
)
578 assert(intrin
->num_components
==
579 glsl_get_vector_elements(node
->type
));
581 nir_ssa_def
*new_def
;
582 b
.cursor
= nir_before_instr(&intrin
->instr
);
584 unsigned wrmask
= nir_intrinsic_write_mask(intrin
);
585 if (wrmask
== (1 << intrin
->num_components
) - 1) {
586 /* Whole variable store - just copy the source. Note that
587 * intrin->num_components and value->num_components
591 for (unsigned i
= 0; i
< 4; i
++)
592 swiz
[i
] = i
< intrin
->num_components
? i
: 0;
594 new_def
= nir_swizzle(&b
, value
, swiz
,
595 intrin
->num_components
, false);
597 nir_ssa_def
*old_def
=
598 nir_phi_builder_value_get_block_def(node
->pb_value
, block
);
599 /* For writemasked store_var intrinsics, we combine the newly
600 * written values with the existing contents of unwritten
601 * channels, creating a new SSA value for the whole vector.
603 nir_ssa_def
*srcs
[4];
604 for (unsigned i
= 0; i
< intrin
->num_components
; i
++) {
605 if (wrmask
& (1 << i
)) {
606 srcs
[i
] = nir_channel(&b
, value
, i
);
608 srcs
[i
] = nir_channel(&b
, old_def
, i
);
611 new_def
= nir_vec(&b
, srcs
, intrin
->num_components
);
614 assert(new_def
->num_components
== intrin
->num_components
);
616 nir_phi_builder_value_set_block_def(node
->pb_value
, block
, new_def
);
617 nir_instr_remove(&intrin
->instr
);
630 /** Implements a pass to lower variable uses to SSA values
632 * This path walks the list of instructions and tries to lower as many
633 * local variable load/store operations to SSA defs and uses as it can.
634 * The process involves four passes:
636 * 1) Iterate over all of the instructions and mark where each local
637 * variable deref is used in a load, store, or copy. While we're at
638 * it, we keep track of all of the fully-qualified (no wildcards) and
639 * fully-direct references we see and store them in the
640 * direct_deref_nodes hash table.
642 * 2) Walk over the list of fully-qualified direct derefs generated in
643 * the previous pass. For each deref, we determine if it can ever be
644 * aliased, i.e. if there is an indirect reference anywhere that may
645 * refer to it. If it cannot be aliased, we mark it for lowering to an
646 * SSA value. At this point, we lower any var_copy instructions that
647 * use the given deref to load/store operations.
649 * 3) Walk over the list of derefs we plan to lower to SSA values and
650 * insert phi nodes as needed.
652 * 4) Perform "variable renaming" by replacing the load/store instructions
653 * with SSA definitions and SSA uses.
656 nir_lower_vars_to_ssa_impl(nir_function_impl
*impl
)
658 struct lower_variables_state state
;
660 state
.shader
= impl
->function
->shader
;
661 state
.dead_ctx
= ralloc_context(state
.shader
);
664 state
.deref_var_nodes
= _mesa_hash_table_create(state
.dead_ctx
,
666 _mesa_key_pointer_equal
);
667 exec_list_make_empty(&state
.direct_deref_nodes
);
669 /* Build the initial deref structures and direct_deref_nodes table */
670 state
.add_to_direct_deref_nodes
= true;
672 register_variable_uses(impl
, &state
);
674 bool progress
= false;
676 nir_metadata_require(impl
, nir_metadata_block_index
);
678 /* We're about to iterate through direct_deref_nodes. Don't modify it. */
679 state
.add_to_direct_deref_nodes
= false;
681 foreach_list_typed_safe(struct deref_node
, node
, direct_derefs_link
,
682 &state
.direct_deref_nodes
) {
683 nir_deref_path
*path
= &node
->path
;
685 assert(path
->path
[0]->deref_type
== nir_deref_type_var
);
687 /* We don't build deref nodes for non-local variables */
688 assert(path
->path
[0]->var
->data
.mode
== nir_var_local
);
690 if (path_may_be_aliased(path
, &state
)) {
691 exec_node_remove(&node
->direct_derefs_link
);
695 node
->lower_to_ssa
= true;
698 foreach_deref_node_match(path
, lower_copies_to_load_store
, &state
);
704 nir_metadata_require(impl
, nir_metadata_dominance
);
706 /* We may have lowered some copy instructions to load/store
707 * instructions. The uses from the copy instructions hav already been
708 * removed but we need to rescan to ensure that the uses from the newly
709 * added load/store instructions are registered. We need this
710 * information for phi node insertion below.
712 register_variable_uses(impl
, &state
);
714 state
.phi_builder
= nir_phi_builder_create(state
.impl
);
716 NIR_VLA(BITSET_WORD
, store_blocks
, BITSET_WORDS(state
.impl
->num_blocks
));
717 foreach_list_typed(struct deref_node
, node
, direct_derefs_link
,
718 &state
.direct_deref_nodes
) {
719 if (!node
->lower_to_ssa
)
722 memset(store_blocks
, 0,
723 BITSET_WORDS(state
.impl
->num_blocks
) * sizeof(*store_blocks
));
725 assert(node
->path
.path
[0]->var
->constant_initializer
== NULL
);
728 set_foreach(node
->stores
, store_entry
) {
729 nir_intrinsic_instr
*store
=
730 (nir_intrinsic_instr
*)store_entry
->key
;
731 BITSET_SET(store_blocks
, store
->instr
.block
->index
);
736 nir_phi_builder_add_value(state
.phi_builder
,
737 glsl_get_vector_elements(node
->type
),
738 glsl_get_bit_size(node
->type
),
742 rename_variables(&state
);
744 nir_phi_builder_finish(state
.phi_builder
);
746 nir_metadata_preserve(impl
, nir_metadata_block_index
|
747 nir_metadata_dominance
);
749 ralloc_free(state
.dead_ctx
);
755 nir_lower_vars_to_ssa(nir_shader
*shader
)
757 bool progress
= false;
759 nir_foreach_function(function
, shader
) {
761 progress
|= nir_lower_vars_to_ssa_impl(function
->impl
);