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 /* override operator new from exec_node */
74 DECLARE_LINEAR_ZALLOC_CXX_OPERATORS(acp_entry
)
76 acp_entry(ir_variable
*lhs
, ir_variable
*rhs
, int write_mask
, int swizzle
[4])
81 this->write_mask
= write_mask
;
82 memcpy(this->swizzle
, swizzle
, sizeof(this->swizzle
));
87 unsigned int write_mask
;
93 class kill_entry
: public exec_node
96 /* override operator new from exec_node */
97 DECLARE_LINEAR_ZALLOC_CXX_OPERATORS(kill_entry
)
99 kill_entry(ir_variable
*var
, int write_mask
)
102 this->write_mask
= write_mask
;
106 unsigned int write_mask
;
109 class ir_copy_propagation_elements_visitor
: public ir_rvalue_visitor
{
111 ir_copy_propagation_elements_visitor()
113 this->progress
= false;
114 this->killed_all
= false;
115 this->mem_ctx
= ralloc_context(NULL
);
116 this->lin_ctx
= linear_alloc_parent(this->mem_ctx
, 0);
117 this->shader_mem_ctx
= NULL
;
118 this->kills
= new(mem_ctx
) exec_list
;
122 ~ir_copy_propagation_elements_visitor()
124 ralloc_free(mem_ctx
);
129 lhs_ht
= _mesa_hash_table_create(mem_ctx
, _mesa_hash_pointer
,
130 _mesa_key_pointer_equal
);
131 rhs_ht
= _mesa_hash_table_create(mem_ctx
, _mesa_hash_pointer
,
132 _mesa_key_pointer_equal
);
137 _mesa_hash_table_destroy(lhs_ht
, NULL
);
138 _mesa_hash_table_destroy(rhs_ht
, NULL
);
141 void populate_acp(hash_table
*lhs
, hash_table
*rhs
)
143 struct hash_entry
*entry
;
145 hash_table_foreach(lhs
, entry
) {
146 _mesa_hash_table_insert(lhs_ht
, entry
->key
, entry
->data
);
149 hash_table_foreach(rhs
, entry
) {
150 _mesa_hash_table_insert(rhs_ht
, entry
->key
, entry
->data
);
154 void handle_loop(ir_loop
*, bool keep_acp
);
155 virtual ir_visitor_status
visit_enter(class ir_loop
*);
156 virtual ir_visitor_status
visit_enter(class ir_function_signature
*);
157 virtual ir_visitor_status
visit_leave(class ir_assignment
*);
158 virtual ir_visitor_status
visit_enter(class ir_call
*);
159 virtual ir_visitor_status
visit_enter(class ir_if
*);
160 virtual ir_visitor_status
visit_leave(class ir_swizzle
*);
162 void handle_rvalue(ir_rvalue
**rvalue
);
164 void add_copy(ir_assignment
*ir
);
165 void kill(kill_entry
*k
);
166 void handle_if_block(exec_list
*instructions
);
168 /** Hash of acp_entry: The available copies to propagate */
173 * List of kill_entry: The variables whose values were killed in this
182 /* Context for our local data structures. */
185 /* Context for allocating new shader nodes. */
186 void *shader_mem_ctx
;
189 } /* unnamed namespace */
192 ir_copy_propagation_elements_visitor::visit_enter(ir_function_signature
*ir
)
194 /* Treat entry into a function signature as a completely separate
195 * block. Any instructions at global scope will be shuffled into
196 * main() at link time, so they're irrelevant to us.
198 exec_list
*orig_kills
= this->kills
;
199 bool orig_killed_all
= this->killed_all
;
201 hash_table
*orig_lhs_ht
= lhs_ht
;
202 hash_table
*orig_rhs_ht
= rhs_ht
;
204 this->kills
= new(mem_ctx
) exec_list
;
205 this->killed_all
= false;
209 visit_list_elements(this, &ir
->body
);
211 ralloc_free(this->kills
);
215 this->kills
= orig_kills
;
216 this->killed_all
= orig_killed_all
;
218 lhs_ht
= orig_lhs_ht
;
219 rhs_ht
= orig_rhs_ht
;
221 return visit_continue_with_parent
;
225 ir_copy_propagation_elements_visitor::visit_leave(ir_assignment
*ir
)
227 ir_dereference_variable
*lhs
= ir
->lhs
->as_dereference_variable();
228 ir_variable
*var
= ir
->lhs
->variable_referenced();
230 if (var
->type
->is_scalar() || var
->type
->is_vector()) {
234 k
= new(this->lin_ctx
) kill_entry(var
, ir
->write_mask
);
236 k
= new(this->lin_ctx
) kill_entry(var
, ~0);
243 return visit_continue
;
247 ir_copy_propagation_elements_visitor::visit_leave(ir_swizzle
*)
249 /* Don't visit the values of swizzles since they are handled while
250 * visiting the swizzle itself.
252 return visit_continue
;
256 * Replaces dereferences of ACP RHS variables with ACP LHS variables.
258 * This is where the actual copy propagation occurs. Note that the
259 * rewriting of ir_dereference means that the ir_dereference instance
260 * must not be shared by multiple IR operations!
263 ir_copy_propagation_elements_visitor::handle_rvalue(ir_rvalue
**ir
)
266 ir_dereference_variable
*deref_var
;
267 ir_variable
*source
[4] = {NULL
, NULL
, NULL
, NULL
};
268 int source_chan
[4] = {0, 0, 0, 0};
270 bool noop_swizzle
= true;
275 ir_swizzle
*swizzle
= (*ir
)->as_swizzle();
277 deref_var
= swizzle
->val
->as_dereference_variable();
281 swizzle_chan
[0] = swizzle
->mask
.x
;
282 swizzle_chan
[1] = swizzle
->mask
.y
;
283 swizzle_chan
[2] = swizzle
->mask
.z
;
284 swizzle_chan
[3] = swizzle
->mask
.w
;
285 chans
= swizzle
->type
->vector_elements
;
287 deref_var
= (*ir
)->as_dereference_variable();
295 chans
= deref_var
->type
->vector_elements
;
298 if (this->in_assignee
)
301 ir_variable
*var
= deref_var
->var
;
303 /* Try to find ACP entries covering swizzle_chan[], hoping they're
304 * the same source variable.
306 hash_entry
*ht_entry
= _mesa_hash_table_search(lhs_ht
, var
);
308 exec_list
*ht_list
= (exec_list
*) ht_entry
->data
;
309 foreach_in_list(acp_entry
, entry
, ht_list
) {
310 for (int c
= 0; c
< chans
; c
++) {
311 if (entry
->write_mask
& (1 << swizzle_chan
[c
])) {
312 source
[c
] = entry
->rhs
;
313 source_chan
[c
] = entry
->swizzle
[swizzle_chan
[c
]];
315 if (source_chan
[c
] != swizzle_chan
[c
])
316 noop_swizzle
= false;
322 /* Make sure all channels are copying from the same source variable. */
325 for (int c
= 1; c
< chans
; c
++) {
326 if (source
[c
] != source
[0])
331 shader_mem_ctx
= ralloc_parent(deref_var
);
333 /* Don't pointlessly replace the rvalue with itself (or a noop swizzle
334 * of itself, which would just be deleted by opt_noop_swizzle).
336 if (source
[0] == var
&& noop_swizzle
)
340 printf("Copy propagation from:\n");
344 deref_var
= new(shader_mem_ctx
) ir_dereference_variable(source
[0]);
345 *ir
= new(shader_mem_ctx
) ir_swizzle(deref_var
,
362 ir_copy_propagation_elements_visitor::visit_enter(ir_call
*ir
)
364 /* Do copy propagation on call parameters, but skip any out params */
365 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
366 actual_node
, &ir
->actual_parameters
) {
367 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
368 ir_rvalue
*ir
= (ir_rvalue
*) actual_node
;
369 if (sig_param
->data
.mode
!= ir_var_function_out
370 && sig_param
->data
.mode
!= ir_var_function_inout
) {
375 /* Since we're unlinked, we don't (necessarily) know the side effects of
376 * this call. So kill all copies.
378 _mesa_hash_table_clear(lhs_ht
, NULL
);
379 _mesa_hash_table_clear(rhs_ht
, NULL
);
381 this->killed_all
= true;
383 return visit_continue_with_parent
;
387 ir_copy_propagation_elements_visitor::handle_if_block(exec_list
*instructions
)
389 exec_list
*orig_kills
= this->kills
;
390 bool orig_killed_all
= this->killed_all
;
392 hash_table
*orig_lhs_ht
= lhs_ht
;
393 hash_table
*orig_rhs_ht
= rhs_ht
;
395 this->kills
= new(mem_ctx
) exec_list
;
396 this->killed_all
= false;
400 /* Populate the initial acp with a copy of the original */
401 populate_acp(orig_lhs_ht
, orig_rhs_ht
);
403 visit_list_elements(this, instructions
);
405 if (this->killed_all
) {
406 _mesa_hash_table_clear(orig_lhs_ht
, NULL
);
407 _mesa_hash_table_clear(orig_rhs_ht
, NULL
);
410 exec_list
*new_kills
= this->kills
;
411 this->kills
= orig_kills
;
412 this->killed_all
= this->killed_all
|| orig_killed_all
;
416 lhs_ht
= orig_lhs_ht
;
417 rhs_ht
= orig_rhs_ht
;
419 /* Move the new kills into the parent block's list, removing them
420 * from the parent's ACP list in the process.
422 foreach_in_list_safe(kill_entry
, k
, new_kills
) {
426 ralloc_free(new_kills
);
430 ir_copy_propagation_elements_visitor::visit_enter(ir_if
*ir
)
432 ir
->condition
->accept(this);
434 handle_if_block(&ir
->then_instructions
);
435 handle_if_block(&ir
->else_instructions
);
437 /* handle_if_block() already descended into the children. */
438 return visit_continue_with_parent
;
442 ir_copy_propagation_elements_visitor::handle_loop(ir_loop
*ir
, bool keep_acp
)
444 exec_list
*orig_kills
= this->kills
;
445 bool orig_killed_all
= this->killed_all
;
447 hash_table
*orig_lhs_ht
= lhs_ht
;
448 hash_table
*orig_rhs_ht
= rhs_ht
;
450 /* FINISHME: For now, the initial acp for loops is totally empty.
451 * We could go through once, then go through again with the acp
452 * cloned minus the killed entries after the first run through.
454 this->kills
= new(mem_ctx
) exec_list
;
455 this->killed_all
= false;
460 /* Populate the initial acp with a copy of the original */
461 populate_acp(orig_lhs_ht
, orig_rhs_ht
);
464 visit_list_elements(this, &ir
->body_instructions
);
466 if (this->killed_all
) {
467 _mesa_hash_table_clear(orig_lhs_ht
, NULL
);
468 _mesa_hash_table_clear(orig_rhs_ht
, NULL
);
471 exec_list
*new_kills
= this->kills
;
472 this->kills
= orig_kills
;
473 this->killed_all
= this->killed_all
|| orig_killed_all
;
477 lhs_ht
= orig_lhs_ht
;
478 rhs_ht
= orig_rhs_ht
;
480 foreach_in_list_safe(kill_entry
, k
, new_kills
) {
484 ralloc_free(new_kills
);
488 ir_copy_propagation_elements_visitor::visit_enter(ir_loop
*ir
)
490 handle_loop(ir
, false);
491 handle_loop(ir
, true);
493 /* already descended into the children. */
494 return visit_continue_with_parent
;
497 /* Remove any entries currently in the ACP for this kill. */
499 ir_copy_propagation_elements_visitor::kill(kill_entry
*k
)
501 /* removal of lhs entries */
502 hash_entry
*ht_entry
= _mesa_hash_table_search(lhs_ht
, k
->var
);
504 exec_list
*lhs_list
= (exec_list
*) ht_entry
->data
;
505 foreach_in_list_safe(acp_entry
, entry
, lhs_list
) {
506 entry
->write_mask
= entry
->write_mask
& ~k
->write_mask
;
507 if (entry
->write_mask
== 0) {
514 /* removal of rhs entries */
515 ht_entry
= _mesa_hash_table_search(rhs_ht
, k
->var
);
517 exec_list
*rhs_list
= (exec_list
*) ht_entry
->data
;
520 while ((ref
= (acp_ref
*) rhs_list
->pop_head()) != NULL
) {
521 acp_entry
*entry
= ref
->entry
;
523 /* If entry is still in a list (not already removed by lhs entry
524 * removal above), remove it.
526 if (entry
->prev
|| entry
->next
)
531 /* If we were on a list, remove ourselves before inserting */
535 this->kills
->push_tail(k
);
539 * Adds directly-copied channels between vector variables to the available
540 * copy propagation list.
543 ir_copy_propagation_elements_visitor::add_copy(ir_assignment
*ir
)
546 int orig_swizzle
[4] = {0, 1, 2, 3};
552 ir_dereference_variable
*lhs
= ir
->lhs
->as_dereference_variable();
553 if (!lhs
|| !(lhs
->type
->is_scalar() || lhs
->type
->is_vector()))
556 ir_dereference_variable
*rhs
= ir
->rhs
->as_dereference_variable();
558 ir_swizzle
*swiz
= ir
->rhs
->as_swizzle();
562 rhs
= swiz
->val
->as_dereference_variable();
566 orig_swizzle
[0] = swiz
->mask
.x
;
567 orig_swizzle
[1] = swiz
->mask
.y
;
568 orig_swizzle
[2] = swiz
->mask
.z
;
569 orig_swizzle
[3] = swiz
->mask
.w
;
572 /* Move the swizzle channels out to the positions they match in the
573 * destination. We don't want to have to rewrite the swizzle[]
574 * array every time we clear a bit of the write_mask.
577 for (int i
= 0; i
< 4; i
++) {
578 if (ir
->write_mask
& (1 << i
))
579 swizzle
[i
] = orig_swizzle
[j
++];
582 int write_mask
= ir
->write_mask
;
583 if (lhs
->var
== rhs
->var
) {
584 /* If this is a copy from the variable to itself, then we need
585 * to be sure not to include the updated channels from this
586 * instruction in the set of new source channels to be
587 * copy-propagated from.
589 for (int i
= 0; i
< 4; i
++) {
590 if (ir
->write_mask
& (1 << orig_swizzle
[i
]))
591 write_mask
&= ~(1 << i
);
595 if (lhs
->var
->data
.precise
!= rhs
->var
->data
.precise
)
598 entry
= new(this->lin_ctx
) acp_entry(lhs
->var
, rhs
->var
, write_mask
,
601 /* lhs hash, hash of lhs -> acp_entry lists */
602 hash_entry
*ht_entry
= _mesa_hash_table_search(lhs_ht
, lhs
->var
);
604 exec_list
*lhs_list
= (exec_list
*) ht_entry
->data
;
605 lhs_list
->push_tail(entry
);
607 exec_list
*lhs_list
= new(mem_ctx
) exec_list
;
608 lhs_list
->push_tail(entry
);
609 _mesa_hash_table_insert(lhs_ht
, lhs
->var
, lhs_list
);
612 /* rhs hash, hash of rhs -> acp_entry pointers to lhs lists */
613 ht_entry
= _mesa_hash_table_search(rhs_ht
, rhs
->var
);
615 exec_list
*rhs_list
= (exec_list
*) ht_entry
->data
;
616 rhs_list
->push_tail(&entry
->rhs_node
);
618 exec_list
*rhs_list
= new(mem_ctx
) exec_list
;
619 rhs_list
->push_tail(&entry
->rhs_node
);
620 _mesa_hash_table_insert(rhs_ht
, rhs
->var
, rhs_list
);
625 do_copy_propagation_elements(exec_list
*instructions
)
627 ir_copy_propagation_elements_visitor v
;
629 visit_list_elements(&v
, instructions
);