2 * Copyright © 2010 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
21 * DEALINGS IN THE SOFTWARE.
25 * \file opt_copy_propagation_elements.cpp
27 * Replaces usage of recently-copied components of variables with the
28 * previous copy of the variable.
30 * This pass can be compared with opt_copy_propagation, which operands
31 * on arbitrary whole-variable copies. However, in order to handle
32 * the copy propagation of swizzled variables or writemasked writes,
33 * we want to track things on a channel-wise basis. I found that
34 * trying to mix the swizzled/writemasked support here with the
35 * whole-variable stuff in opt_copy_propagation.cpp just made a mess,
36 * so this is separate despite the ACP handling being somewhat
39 * This should reduce the number of MOV instructions in the generated
40 * programs unless copy propagation is also done on the LIR, and may
41 * help anyway by triggering other optimizations that live in the HIR.
45 #include "ir_rvalue_visitor.h"
46 #include "ir_basic_block.h"
47 #include "ir_optimization.h"
48 #include "compiler/glsl_types.h"
49 #include "util/hash_table.h"
51 static bool debug = false;
57 /* Class that refers to acp_entry in another exec_list. Used
58 * when making removals based on rhs.
60 class acp_ref : public exec_node
70 class acp_entry : public exec_node
73 acp_entry(ir_variable *lhs, ir_variable *rhs, int write_mask, int swizzle[4])
78 this->write_mask = write_mask;
79 memcpy(this->swizzle, swizzle, sizeof(this->swizzle));
84 unsigned int write_mask;
90 class kill_entry : public exec_node
93 kill_entry(ir_variable *var, int write_mask)
96 this->write_mask = write_mask;
100 unsigned int write_mask;
103 class ir_copy_propagation_elements_visitor : public ir_rvalue_visitor {
105 ir_copy_propagation_elements_visitor()
107 this->progress = false;
108 this->killed_all = false;
109 this->mem_ctx = ralloc_context(NULL);
110 this->shader_mem_ctx = NULL;
111 this->kills = new(mem_ctx) exec_list;
115 ~ir_copy_propagation_elements_visitor()
117 ralloc_free(mem_ctx);
122 lhs_ht = _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
123 _mesa_key_pointer_equal);
124 rhs_ht = _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
125 _mesa_key_pointer_equal);
130 _mesa_hash_table_destroy(lhs_ht, NULL);
131 _mesa_hash_table_destroy(rhs_ht, NULL);
134 void populate_acp(hash_table *lhs, hash_table *rhs)
136 struct hash_entry *entry;
138 hash_table_foreach(lhs, entry) {
139 _mesa_hash_table_insert(lhs_ht, entry->key, entry->data);
142 hash_table_foreach(rhs, entry) {
143 _mesa_hash_table_insert(rhs_ht, entry->key, entry->data);
147 void handle_loop(ir_loop *, bool keep_acp);
148 virtual ir_visitor_status visit_enter(class ir_loop *);
149 virtual ir_visitor_status visit_enter(class ir_function_signature *);
150 virtual ir_visitor_status visit_leave(class ir_assignment *);
151 virtual ir_visitor_status visit_enter(class ir_call *);
152 virtual ir_visitor_status visit_enter(class ir_if *);
153 virtual ir_visitor_status visit_leave(class ir_swizzle *);
155 void handle_rvalue(ir_rvalue **rvalue);
157 void add_copy(ir_assignment *ir);
158 void kill(kill_entry *k);
159 void handle_if_block(exec_list *instructions);
161 /** Hash of acp_entry: The available copies to propagate */
166 * List of kill_entry: The variables whose values were killed in this
175 /* Context for our local data structures. */
177 /* Context for allocating new shader nodes. */
178 void *shader_mem_ctx;
181 } /* unnamed namespace */
184 ir_copy_propagation_elements_visitor::visit_enter(ir_function_signature *ir)
186 /* Treat entry into a function signature as a completely separate
187 * block. Any instructions at global scope will be shuffled into
188 * main() at link time, so they're irrelevant to us.
190 exec_list *orig_kills = this->kills;
191 bool orig_killed_all = this->killed_all;
193 hash_table *orig_lhs_ht = lhs_ht;
194 hash_table *orig_rhs_ht = rhs_ht;
196 this->kills = new(mem_ctx) exec_list;
197 this->killed_all = false;
201 visit_list_elements(this, &ir->body);
203 ralloc_free(this->kills);
207 this->kills = orig_kills;
208 this->killed_all = orig_killed_all;
210 lhs_ht = orig_lhs_ht;
211 rhs_ht = orig_rhs_ht;
213 return visit_continue_with_parent;
217 ir_copy_propagation_elements_visitor::visit_leave(ir_assignment *ir)
219 ir_dereference_variable *lhs = ir->lhs->as_dereference_variable();
220 ir_variable *var = ir->lhs->variable_referenced();
222 if (var->type->is_scalar() || var->type->is_vector()) {
226 k = new(this->kills) kill_entry(var, ir->write_mask);
228 k = new(this->kills) kill_entry(var, ~0);
235 return visit_continue;
239 ir_copy_propagation_elements_visitor::visit_leave(ir_swizzle *)
241 /* Don't visit the values of swizzles since they are handled while
242 * visiting the swizzle itself.
244 return visit_continue;
248 * Replaces dereferences of ACP RHS variables with ACP LHS variables.
250 * This is where the actual copy propagation occurs. Note that the
251 * rewriting of ir_dereference means that the ir_dereference instance
252 * must not be shared by multiple IR operations!
255 ir_copy_propagation_elements_visitor::handle_rvalue(ir_rvalue **ir)
258 ir_dereference_variable *deref_var;
259 ir_variable *source[4] = {NULL, NULL, NULL, NULL};
260 int source_chan[4] = {0, 0, 0, 0};
262 bool noop_swizzle = true;
267 ir_swizzle *swizzle = (*ir)->as_swizzle();
269 deref_var = swizzle->val->as_dereference_variable();
273 swizzle_chan[0] = swizzle->mask.x;
274 swizzle_chan[1] = swizzle->mask.y;
275 swizzle_chan[2] = swizzle->mask.z;
276 swizzle_chan[3] = swizzle->mask.w;
277 chans = swizzle->type->vector_elements;
279 deref_var = (*ir)->as_dereference_variable();
287 chans = deref_var->type->vector_elements;
290 if (this->in_assignee)
293 ir_variable *var = deref_var->var;
295 /* Try to find ACP entries covering swizzle_chan[], hoping they're
296 * the same source variable.
298 hash_entry *ht_entry = _mesa_hash_table_search(lhs_ht, var);
300 exec_list *ht_list = (exec_list *) ht_entry->data;
301 foreach_in_list(acp_entry, entry, ht_list) {
302 for (int c = 0; c < chans; c++) {
303 if (entry->write_mask & (1 << swizzle_chan[c])) {
304 source[c] = entry->rhs;
305 source_chan[c] = entry->swizzle[swizzle_chan[c]];
307 if (source_chan[c] != swizzle_chan[c])
308 noop_swizzle = false;
314 /* Make sure all channels are copying from the same source variable. */
317 for (int c = 1; c < chans; c++) {
318 if (source[c] != source[0])
323 shader_mem_ctx = ralloc_parent(deref_var);
325 /* Don't pointlessly replace the rvalue with itself (or a noop swizzle
326 * of itself, which would just be deleted by opt_noop_swizzle).
328 if (source[0] == var && noop_swizzle)
332 printf("Copy propagation from:\n");
336 deref_var = new(shader_mem_ctx) ir_dereference_variable(source[0]);
337 *ir = new(shader_mem_ctx) ir_swizzle(deref_var,
354 ir_copy_propagation_elements_visitor::visit_enter(ir_call *ir)
356 /* Do copy propagation on call parameters, but skip any out params */
357 foreach_two_lists(formal_node, &ir->callee->parameters,
358 actual_node, &ir->actual_parameters) {
359 ir_variable *sig_param = (ir_variable *) formal_node;
360 ir_rvalue *ir = (ir_rvalue *) actual_node;
361 if (sig_param->data.mode != ir_var_function_out
362 && sig_param->data.mode != ir_var_function_inout) {
367 /* Since we're unlinked, we don't (necessarily) know the side effects of
368 * this call. So kill all copies.
370 _mesa_hash_table_clear(lhs_ht, NULL);
371 _mesa_hash_table_clear(rhs_ht, NULL);
373 this->killed_all = true;
375 return visit_continue_with_parent;
379 ir_copy_propagation_elements_visitor::handle_if_block(exec_list *instructions)
381 exec_list *orig_kills = this->kills;
382 bool orig_killed_all = this->killed_all;
384 hash_table *orig_lhs_ht = lhs_ht;
385 hash_table *orig_rhs_ht = rhs_ht;
387 this->kills = new(mem_ctx) exec_list;
388 this->killed_all = false;
392 /* Populate the initial acp with a copy of the original */
393 populate_acp(orig_lhs_ht, orig_rhs_ht);
395 visit_list_elements(this, instructions);
397 if (this->killed_all) {
398 _mesa_hash_table_clear(orig_lhs_ht, NULL);
399 _mesa_hash_table_clear(orig_rhs_ht, NULL);
402 exec_list *new_kills = this->kills;
403 this->kills = orig_kills;
404 this->killed_all = this->killed_all || orig_killed_all;
408 lhs_ht = orig_lhs_ht;
409 rhs_ht = orig_rhs_ht;
411 /* Move the new kills into the parent block's list, removing them
412 * from the parent's ACP list in the process.
414 foreach_in_list_safe(kill_entry, k, new_kills) {
418 ralloc_free(new_kills);
422 ir_copy_propagation_elements_visitor::visit_enter(ir_if *ir)
424 ir->condition->accept(this);
426 handle_if_block(&ir->then_instructions);
427 handle_if_block(&ir->else_instructions);
429 /* handle_if_block() already descended into the children. */
430 return visit_continue_with_parent;
434 ir_copy_propagation_elements_visitor::handle_loop(ir_loop *ir, bool keep_acp)
436 exec_list *orig_kills = this->kills;
437 bool orig_killed_all = this->killed_all;
439 hash_table *orig_lhs_ht = lhs_ht;
440 hash_table *orig_rhs_ht = rhs_ht;
442 /* FINISHME: For now, the initial acp for loops is totally empty.
443 * We could go through once, then go through again with the acp
444 * cloned minus the killed entries after the first run through.
446 this->kills = new(mem_ctx) exec_list;
447 this->killed_all = false;
452 /* Populate the initial acp with a copy of the original */
453 populate_acp(orig_lhs_ht, orig_rhs_ht);
456 visit_list_elements(this, &ir->body_instructions);
458 if (this->killed_all) {
459 _mesa_hash_table_clear(orig_lhs_ht, NULL);
460 _mesa_hash_table_clear(orig_rhs_ht, NULL);
463 exec_list *new_kills = this->kills;
464 this->kills = orig_kills;
465 this->killed_all = this->killed_all || orig_killed_all;
469 lhs_ht = orig_lhs_ht;
470 rhs_ht = orig_rhs_ht;
472 foreach_in_list_safe(kill_entry, k, new_kills) {
476 ralloc_free(new_kills);
480 ir_copy_propagation_elements_visitor::visit_enter(ir_loop *ir)
482 handle_loop(ir, false);
483 handle_loop(ir, true);
485 /* already descended into the children. */
486 return visit_continue_with_parent;
489 /* Remove any entries currently in the ACP for this kill. */
491 ir_copy_propagation_elements_visitor::kill(kill_entry *k)
493 /* removal of lhs entries */
494 hash_entry *ht_entry = _mesa_hash_table_search(lhs_ht, k->var);
496 exec_list *lhs_list = (exec_list *) ht_entry->data;
497 foreach_in_list_safe(acp_entry, entry, lhs_list) {
498 entry->write_mask = entry->write_mask & ~k->write_mask;
499 if (entry->write_mask == 0) {
506 /* removal of rhs entries */
507 ht_entry = _mesa_hash_table_search(rhs_ht, k->var);
509 exec_list *rhs_list = (exec_list *) ht_entry->data;
512 while ((ref = (acp_ref *) rhs_list->pop_head()) != NULL) {
513 acp_entry *entry = ref->entry;
515 /* If entry is still in a list (not already removed by lhs entry
516 * removal above), remove it.
518 if (entry->prev || entry->next)
523 /* If we were on a list, remove ourselves before inserting */
527 ralloc_steal(this->kills, k);
528 this->kills->push_tail(k);
532 * Adds directly-copied channels between vector variables to the available
533 * copy propagation list.
536 ir_copy_propagation_elements_visitor::add_copy(ir_assignment *ir)
539 int orig_swizzle[4] = {0, 1, 2, 3};
545 ir_dereference_variable *lhs = ir->lhs->as_dereference_variable();
546 if (!lhs || !(lhs->type->is_scalar() || lhs->type->is_vector()))
549 ir_dereference_variable *rhs = ir->rhs->as_dereference_variable();
551 ir_swizzle *swiz = ir->rhs->as_swizzle();
555 rhs = swiz->val->as_dereference_variable();
559 orig_swizzle[0] = swiz->mask.x;
560 orig_swizzle[1] = swiz->mask.y;
561 orig_swizzle[2] = swiz->mask.z;
562 orig_swizzle[3] = swiz->mask.w;
565 /* Move the swizzle channels out to the positions they match in the
566 * destination. We don't want to have to rewrite the swizzle[]
567 * array every time we clear a bit of the write_mask.
570 for (int i = 0; i < 4; i++) {
571 if (ir->write_mask & (1 << i))
572 swizzle[i] = orig_swizzle[j++];
575 int write_mask = ir->write_mask;
576 if (lhs->var == rhs->var) {
577 /* If this is a copy from the variable to itself, then we need
578 * to be sure not to include the updated channels from this
579 * instruction in the set of new source channels to be
580 * copy-propagated from.
582 for (int i = 0; i < 4; i++) {
583 if (ir->write_mask & (1 << orig_swizzle[i]))
584 write_mask &= ~(1 << i);
588 if (lhs->var->data.precise != rhs->var->data.precise)
591 entry = new(this->mem_ctx) acp_entry(lhs->var, rhs->var, write_mask,
594 /* lhs hash, hash of lhs -> acp_entry lists */
595 hash_entry *ht_entry = _mesa_hash_table_search(lhs_ht, lhs->var);
597 exec_list *lhs_list = (exec_list *) ht_entry->data;
598 lhs_list->push_tail(entry);
600 exec_list *lhs_list = new(mem_ctx) exec_list;
601 lhs_list->push_tail(entry);
602 _mesa_hash_table_insert(lhs_ht, lhs->var, lhs_list);
605 /* rhs hash, hash of rhs -> acp_entry pointers to lhs lists */
606 ht_entry = _mesa_hash_table_search(rhs_ht, rhs->var);
608 exec_list *rhs_list = (exec_list *) ht_entry->data;
609 rhs_list->push_tail(&entry->rhs_node);
611 exec_list *rhs_list = new(mem_ctx) exec_list;
612 rhs_list->push_tail(&entry->rhs_node);
613 _mesa_hash_table_insert(rhs_ht, rhs->var, rhs_list);
618 do_copy_propagation_elements(exec_list *instructions)
620 ir_copy_propagation_elements_visitor v;
622 visit_list_elements(&v, instructions);