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_phi_builder.h"
35 struct deref_node *parent;
36 const struct glsl_type *type;
40 /* Only valid for things that end up in the direct list.
41 * Note that multiple nir_deref_vars may correspond to this node, but they
42 * will all be equivalent, so any is as good as the other.
45 struct exec_node direct_derefs_link;
51 struct nir_phi_builder_value *pb_value;
53 struct deref_node *wildcard;
54 struct deref_node *indirect;
55 struct deref_node *children[0];
58 struct lower_variables_state {
61 nir_function_impl *impl;
63 /* A hash table mapping variables to deref_node data */
64 struct hash_table *deref_var_nodes;
66 /* A hash table mapping fully-qualified direct dereferences, i.e.
67 * dereferences with no indirect or wildcard array dereferences, to
70 * At the moment, we only lower loads, stores, and copies that can be
71 * trivially lowered to loads and stores, i.e. copies with no indirects
72 * and no wildcards. If a part of a variable that is being loaded from
73 * and/or stored into is also involved in a copy operation with
74 * wildcards, then we lower that copy operation to loads and stores, but
75 * otherwise we leave copies with wildcards alone. Since the only derefs
76 * used in these loads, stores, and trivial copies are ones with no
77 * wildcards and no indirects, these are precisely the derefs that we
78 * can actually consider lowering.
80 struct exec_list direct_deref_nodes;
82 /* Controls whether get_deref_node will add variables to the
83 * direct_deref_nodes table. This is turned on when we are initially
84 * scanning for load/store instructions. It is then turned off so we
85 * don't accidentally change the direct_deref_nodes table while we're
86 * iterating throug it.
88 bool add_to_direct_deref_nodes;
90 struct nir_phi_builder *phi_builder;
93 static struct deref_node *
94 deref_node_create(struct deref_node *parent,
95 const struct glsl_type *type, nir_shader *shader)
97 size_t size = sizeof(struct deref_node) +
98 glsl_get_length(type) * sizeof(struct deref_node *);
100 struct deref_node *node = rzalloc_size(shader, size);
102 node->parent = parent;
104 exec_node_init(&node->direct_derefs_link);
109 /* Returns the deref node associated with the given variable. This will be
110 * the root of the tree representing all of the derefs of the given variable.
112 static struct deref_node *
113 get_deref_node_for_var(nir_variable *var, struct lower_variables_state *state)
115 struct deref_node *node;
117 struct hash_entry *var_entry =
118 _mesa_hash_table_search(state->deref_var_nodes, var);
121 return var_entry->data;
123 node = deref_node_create(NULL, var->type, state->dead_ctx);
124 _mesa_hash_table_insert(state->deref_var_nodes, var, node);
129 /* Gets the deref_node for the given deref chain and creates it if it
130 * doesn't yet exist. If the deref is fully-qualified and direct and
131 * state->add_to_direct_deref_nodes is true, it will be added to the hash
132 * table of of fully-qualified direct derefs.
134 static struct deref_node *
135 get_deref_node(nir_deref_var *deref, struct lower_variables_state *state)
137 bool is_direct = true;
139 /* Start at the base of the chain. */
140 struct deref_node *node = get_deref_node_for_var(deref->var, state);
141 assert(deref->deref.type == node->type);
143 for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) {
144 switch (tail->deref_type) {
145 case nir_deref_type_struct: {
146 nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
148 assert(deref_struct->index < glsl_get_length(node->type));
150 if (node->children[deref_struct->index] == NULL)
151 node->children[deref_struct->index] =
152 deref_node_create(node, tail->type, state->dead_ctx);
154 node = node->children[deref_struct->index];
158 case nir_deref_type_array: {
159 nir_deref_array *arr = nir_deref_as_array(tail);
161 switch (arr->deref_array_type) {
162 case nir_deref_array_type_direct:
163 /* This is possible if a loop unrolls and generates an
164 * out-of-bounds offset. We need to handle this at least
165 * somewhat gracefully.
167 if (arr->base_offset >= glsl_get_length(node->type))
170 if (node->children[arr->base_offset] == NULL)
171 node->children[arr->base_offset] =
172 deref_node_create(node, tail->type, state->dead_ctx);
174 node = node->children[arr->base_offset];
177 case nir_deref_array_type_indirect:
178 if (node->indirect == NULL)
179 node->indirect = deref_node_create(node, tail->type,
182 node = node->indirect;
186 case nir_deref_array_type_wildcard:
187 if (node->wildcard == NULL)
188 node->wildcard = deref_node_create(node, tail->type,
191 node = node->wildcard;
196 unreachable("Invalid array deref type");
201 unreachable("Invalid deref type");
207 /* Only insert if it isn't already in the list. */
208 if (is_direct && state->add_to_direct_deref_nodes &&
209 node->direct_derefs_link.next == NULL) {
211 assert(deref->var != NULL);
212 exec_list_push_tail(&state->direct_deref_nodes,
213 &node->direct_derefs_link);
219 /* \sa foreach_deref_node_match */
221 foreach_deref_node_worker(struct deref_node *node, nir_deref *deref,
222 bool (* cb)(struct deref_node *node,
223 struct lower_variables_state *state),
224 struct lower_variables_state *state)
226 if (deref->child == NULL) {
227 return cb(node, state);
229 switch (deref->child->deref_type) {
230 case nir_deref_type_array: {
231 nir_deref_array *arr = nir_deref_as_array(deref->child);
232 assert(arr->deref_array_type == nir_deref_array_type_direct);
233 if (node->children[arr->base_offset] &&
234 !foreach_deref_node_worker(node->children[arr->base_offset],
235 deref->child, cb, state))
238 if (node->wildcard &&
239 !foreach_deref_node_worker(node->wildcard,
240 deref->child, cb, state))
246 case nir_deref_type_struct: {
247 nir_deref_struct *str = nir_deref_as_struct(deref->child);
248 return foreach_deref_node_worker(node->children[str->index],
249 deref->child, cb, state);
253 unreachable("Invalid deref child type");
258 /* Walks over every "matching" deref_node and calls the callback. A node
259 * is considered to "match" if either refers to that deref or matches up t
260 * a wildcard. In other words, the following would match a[6].foo[3].bar:
267 * The given deref must be a full-length and fully qualified (no wildcards
268 * or indirects) deref chain.
271 foreach_deref_node_match(nir_deref_var *deref,
272 bool (* cb)(struct deref_node *node,
273 struct lower_variables_state *state),
274 struct lower_variables_state *state)
276 nir_deref_var var_deref = *deref;
277 var_deref.deref.child = NULL;
278 struct deref_node *node = get_deref_node(&var_deref, state);
283 return foreach_deref_node_worker(node, &deref->deref, cb, state);
286 /* \sa deref_may_be_aliased */
288 deref_may_be_aliased_node(struct deref_node *node, nir_deref *deref,
289 struct lower_variables_state *state)
291 if (deref->child == NULL) {
294 switch (deref->child->deref_type) {
295 case nir_deref_type_array: {
296 nir_deref_array *arr = nir_deref_as_array(deref->child);
297 if (arr->deref_array_type == nir_deref_array_type_indirect)
300 /* If there is an indirect at this level, we're aliased. */
304 assert(arr->deref_array_type == nir_deref_array_type_direct);
306 if (node->children[arr->base_offset] &&
307 deref_may_be_aliased_node(node->children[arr->base_offset],
308 deref->child, state))
311 if (node->wildcard &&
312 deref_may_be_aliased_node(node->wildcard, deref->child, state))
318 case nir_deref_type_struct: {
319 nir_deref_struct *str = nir_deref_as_struct(deref->child);
320 if (node->children[str->index]) {
321 return deref_may_be_aliased_node(node->children[str->index],
322 deref->child, state);
329 unreachable("Invalid nir_deref child type");
334 /* Returns true if there are no indirects that can ever touch this deref.
336 * For example, if the given deref is a[6].foo, then any uses of a[i].foo
337 * would cause this to return false, but a[i].bar would not affect it
338 * because it's a different structure member. A var_copy involving of
339 * a[*].bar also doesn't affect it because that can be lowered to entirely
340 * direct load/stores.
342 * We only support asking this question about fully-qualified derefs.
343 * Obviously, it's pointless to ask this about indirects, but we also
344 * rule-out wildcards. Handling Wildcard dereferences would involve
345 * checking each array index to make sure that there aren't any indirect
349 deref_may_be_aliased(nir_deref_var *deref,
350 struct lower_variables_state *state)
352 return deref_may_be_aliased_node(get_deref_node_for_var(deref->var, state),
353 &deref->deref, state);
357 register_load_instr(nir_intrinsic_instr *load_instr,
358 struct lower_variables_state *state)
360 struct deref_node *node = get_deref_node(load_instr->variables[0], state);
364 if (node->loads == NULL)
365 node->loads = _mesa_set_create(state->dead_ctx, _mesa_hash_pointer,
366 _mesa_key_pointer_equal);
368 _mesa_set_add(node->loads, load_instr);
372 register_store_instr(nir_intrinsic_instr *store_instr,
373 struct lower_variables_state *state)
375 struct deref_node *node = get_deref_node(store_instr->variables[0], state);
379 if (node->stores == NULL)
380 node->stores = _mesa_set_create(state->dead_ctx, _mesa_hash_pointer,
381 _mesa_key_pointer_equal);
383 _mesa_set_add(node->stores, store_instr);
387 register_copy_instr(nir_intrinsic_instr *copy_instr,
388 struct lower_variables_state *state)
390 for (unsigned idx = 0; idx < 2; idx++) {
391 struct deref_node *node =
392 get_deref_node(copy_instr->variables[idx], state);
397 if (node->copies == NULL)
398 node->copies = _mesa_set_create(state->dead_ctx, _mesa_hash_pointer,
399 _mesa_key_pointer_equal);
401 _mesa_set_add(node->copies, copy_instr);
405 /* Registers all variable uses in the given block. */
407 register_variable_uses_block(nir_block *block,
408 struct lower_variables_state *state)
410 nir_foreach_instr_safe(instr, block) {
411 if (instr->type != nir_instr_type_intrinsic)
414 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
416 switch (intrin->intrinsic) {
417 case nir_intrinsic_load_var:
418 register_load_instr(intrin, state);
421 case nir_intrinsic_store_var:
422 register_store_instr(intrin, state);
425 case nir_intrinsic_copy_var:
426 register_copy_instr(intrin, state);
437 /* Walks over all of the copy instructions to or from the given deref_node
438 * and lowers them to load/store intrinsics.
441 lower_copies_to_load_store(struct deref_node *node,
442 struct lower_variables_state *state)
447 struct set_entry *copy_entry;
448 set_foreach(node->copies, copy_entry) {
449 nir_intrinsic_instr *copy = (void *)copy_entry->key;
451 nir_lower_var_copy_instr(copy, state->shader);
453 for (unsigned i = 0; i < 2; ++i) {
454 struct deref_node *arg_node =
455 get_deref_node(copy->variables[i], state);
457 /* Only bother removing copy entries for other nodes */
458 if (arg_node == NULL || arg_node == node)
461 struct set_entry *arg_entry = _mesa_set_search(arg_node->copies, copy);
463 _mesa_set_remove(node->copies, arg_entry);
466 nir_instr_remove(©->instr);
474 /* Performs variable renaming by doing a DFS of the dominance tree
476 * This algorithm is very similar to the one outlined in "Efficiently
477 * Computing Static Single Assignment Form and the Control Dependence
478 * Graph" by Cytron et. al. The primary difference is that we only put one
479 * SSA def on the stack per block.
482 rename_variables_block(nir_block *block, struct lower_variables_state *state)
485 nir_builder_init(&b, state->impl);
487 nir_foreach_instr_safe(instr, block) {
488 if (instr->type != nir_instr_type_intrinsic)
491 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
493 switch (intrin->intrinsic) {
494 case nir_intrinsic_load_var: {
495 struct deref_node *node =
496 get_deref_node(intrin->variables[0], state);
499 /* If we hit this path then we are referencing an invalid
500 * value. Most likely, we unrolled something and are
501 * reading past the end of some array. In any case, this
502 * should result in an undefined value.
504 nir_ssa_undef_instr *undef =
505 nir_ssa_undef_instr_create(state->shader,
506 intrin->num_components,
507 intrin->dest.ssa.bit_size);
509 nir_instr_insert_before(&intrin->instr, &undef->instr);
510 nir_instr_remove(&intrin->instr);
512 nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
513 nir_src_for_ssa(&undef->def));
517 if (!node->lower_to_ssa)
520 nir_alu_instr *mov = nir_alu_instr_create(state->shader,
522 mov->src[0].src = nir_src_for_ssa(
523 nir_phi_builder_value_get_block_def(node->pb_value, block));
524 for (unsigned i = intrin->num_components; i < 4; i++)
525 mov->src[0].swizzle[i] = 0;
527 assert(intrin->dest.is_ssa);
529 mov->dest.write_mask = (1 << intrin->num_components) - 1;
530 nir_ssa_dest_init(&mov->instr, &mov->dest.dest,
531 intrin->num_components,
532 intrin->dest.ssa.bit_size, NULL);
534 nir_instr_insert_before(&intrin->instr, &mov->instr);
535 nir_instr_remove(&intrin->instr);
537 nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
538 nir_src_for_ssa(&mov->dest.dest.ssa));
542 case nir_intrinsic_store_var: {
543 struct deref_node *node =
544 get_deref_node(intrin->variables[0], state);
547 /* Probably an out-of-bounds array store. That should be a
549 nir_instr_remove(&intrin->instr);
553 if (!node->lower_to_ssa)
556 assert(intrin->num_components ==
557 glsl_get_vector_elements(node->type));
559 assert(intrin->src[0].is_ssa);
561 nir_ssa_def *new_def;
562 b.cursor = nir_before_instr(&intrin->instr);
564 unsigned wrmask = nir_intrinsic_write_mask(intrin);
565 if (wrmask == (1 << intrin->num_components) - 1) {
566 /* Whole variable store - just copy the source. Note that
567 * intrin->num_components and intrin->src[0].ssa->num_components
571 for (unsigned i = 0; i < 4; i++)
572 swiz[i] = i < intrin->num_components ? i : 0;
574 new_def = nir_swizzle(&b, intrin->src[0].ssa, swiz,
575 intrin->num_components, false);
577 nir_ssa_def *old_def =
578 nir_phi_builder_value_get_block_def(node->pb_value, block);
579 /* For writemasked store_var intrinsics, we combine the newly
580 * written values with the existing contents of unwritten
581 * channels, creating a new SSA value for the whole vector.
583 nir_ssa_def *srcs[4];
584 for (unsigned i = 0; i < intrin->num_components; i++) {
585 if (wrmask & (1 << i)) {
586 srcs[i] = nir_channel(&b, intrin->src[0].ssa, i);
588 srcs[i] = nir_channel(&b, old_def, i);
591 new_def = nir_vec(&b, srcs, intrin->num_components);
594 assert(new_def->num_components == intrin->num_components);
596 nir_phi_builder_value_set_block_def(node->pb_value, block, new_def);
597 nir_instr_remove(&intrin->instr);
606 for (unsigned i = 0; i < block->num_dom_children; ++i)
607 rename_variables_block(block->dom_children[i], state);
612 /** Implements a pass to lower variable uses to SSA values
614 * This path walks the list of instructions and tries to lower as many
615 * local variable load/store operations to SSA defs and uses as it can.
616 * The process involves four passes:
618 * 1) Iterate over all of the instructions and mark where each local
619 * variable deref is used in a load, store, or copy. While we're at
620 * it, we keep track of all of the fully-qualified (no wildcards) and
621 * fully-direct references we see and store them in the
622 * direct_deref_nodes hash table.
624 * 2) Walk over the the list of fully-qualified direct derefs generated in
625 * the previous pass. For each deref, we determine if it can ever be
626 * aliased, i.e. if there is an indirect reference anywhere that may
627 * refer to it. If it cannot be aliased, we mark it for lowering to an
628 * SSA value. At this point, we lower any var_copy instructions that
629 * use the given deref to load/store operations and, if the deref has a
630 * constant initializer, we go ahead and add a load_const value at the
631 * beginning of the function with the initialized value.
633 * 3) Walk over the list of derefs we plan to lower to SSA values and
634 * insert phi nodes as needed.
636 * 4) Perform "variable renaming" by replacing the load/store instructions
637 * with SSA definitions and SSA uses.
640 nir_lower_vars_to_ssa_impl(nir_function_impl *impl)
642 struct lower_variables_state state;
644 state.shader = impl->function->shader;
645 state.dead_ctx = ralloc_context(state.shader);
648 state.deref_var_nodes = _mesa_hash_table_create(state.dead_ctx,
650 _mesa_key_pointer_equal);
651 exec_list_make_empty(&state.direct_deref_nodes);
653 /* Build the initial deref structures and direct_deref_nodes table */
654 state.add_to_direct_deref_nodes = true;
656 nir_foreach_block(block, impl) {
657 register_variable_uses_block(block, &state);
660 bool progress = false;
662 nir_metadata_require(impl, nir_metadata_block_index);
664 /* We're about to iterate through direct_deref_nodes. Don't modify it. */
665 state.add_to_direct_deref_nodes = false;
667 foreach_list_typed_safe(struct deref_node, node, direct_derefs_link,
668 &state.direct_deref_nodes) {
669 nir_deref_var *deref = node->deref;
671 if (deref->var->data.mode != nir_var_local) {
672 exec_node_remove(&node->direct_derefs_link);
676 if (deref_may_be_aliased(deref, &state)) {
677 exec_node_remove(&node->direct_derefs_link);
681 node->lower_to_ssa = true;
684 foreach_deref_node_match(deref, lower_copies_to_load_store, &state);
690 nir_metadata_require(impl, nir_metadata_dominance);
692 /* We may have lowered some copy instructions to load/store
693 * instructions. The uses from the copy instructions hav already been
694 * removed but we need to rescan to ensure that the uses from the newly
695 * added load/store instructions are registered. We need this
696 * information for phi node insertion below.
698 nir_foreach_block(block, impl) {
699 register_variable_uses_block(block, &state);
702 state.phi_builder = nir_phi_builder_create(state.impl);
704 NIR_VLA(BITSET_WORD, store_blocks, BITSET_WORDS(state.impl->num_blocks));
705 foreach_list_typed(struct deref_node, node, direct_derefs_link,
706 &state.direct_deref_nodes) {
707 if (!node->lower_to_ssa)
710 memset(store_blocks, 0,
711 BITSET_WORDS(state.impl->num_blocks) * sizeof(*store_blocks));
714 struct set_entry *store_entry;
715 set_foreach(node->stores, store_entry) {
716 nir_intrinsic_instr *store =
717 (nir_intrinsic_instr *)store_entry->key;
718 BITSET_SET(store_blocks, store->instr.block->index);
722 if (node->deref->var->constant_initializer)
723 BITSET_SET(store_blocks, 0);
726 nir_phi_builder_add_value(state.phi_builder,
727 glsl_get_vector_elements(node->type),
728 glsl_get_bit_size(node->type),
731 if (node->deref->var->constant_initializer) {
732 nir_load_const_instr *load =
733 nir_deref_get_const_initializer_load(state.shader, node->deref);
734 nir_instr_insert_before_cf_list(&impl->body, &load->instr);
735 nir_phi_builder_value_set_block_def(node->pb_value,
736 nir_start_block(impl), &load->def);
740 rename_variables_block(nir_start_block(impl), &state);
742 nir_phi_builder_finish(state.phi_builder);
744 nir_metadata_preserve(impl, nir_metadata_block_index |
745 nir_metadata_dominance);
747 ralloc_free(state.dead_ctx);
753 nir_lower_vars_to_ssa(nir_shader *shader)
755 nir_foreach_function(function, shader) {
757 nir_lower_vars_to_ssa_impl(function->impl);