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
;
92 class copy_propagation_state
{
94 DECLARE_RZALLOC_CXX_OPERATORS(copy_propagation_state
);
96 copy_propagation_state(void *mem_ctx
, void *lin_ctx
)
98 /* Use 'this' as context for the tables, no explicit destruction
101 lhs_ht
= _mesa_hash_table_create(this, _mesa_hash_pointer
,
102 _mesa_key_pointer_equal
);
103 rhs_ht
= _mesa_hash_table_create(this, _mesa_hash_pointer
,
104 _mesa_key_pointer_equal
);
105 this->mem_ctx
= mem_ctx
;
106 this->lin_ctx
= lin_ctx
;
109 copy_propagation_state
* clone()
111 return new (ralloc_parent(this)) copy_propagation_state(this);
116 _mesa_hash_table_clear(lhs_ht
, NULL
);
117 _mesa_hash_table_clear(rhs_ht
, NULL
);
120 void erase(ir_variable
*var
, unsigned write_mask
)
122 /* removal of lhs entries */
123 hash_entry
*ht_entry
= _mesa_hash_table_search(lhs_ht
, var
);
125 exec_list
*lhs_list
= (exec_list
*) ht_entry
->data
;
126 foreach_in_list_safe(acp_entry
, entry
, lhs_list
) {
127 entry
->write_mask
= entry
->write_mask
& ~write_mask
;
128 if (entry
->write_mask
== 0) {
135 /* removal of rhs entries */
136 ht_entry
= _mesa_hash_table_search(rhs_ht
, var
);
138 exec_list
*rhs_list
= (exec_list
*) ht_entry
->data
;
141 while ((ref
= (acp_ref
*) rhs_list
->pop_head()) != NULL
) {
142 acp_entry
*entry
= ref
->entry
;
144 /* If entry is still in a list (not already removed by lhs entry
145 * removal above), remove it.
147 if (entry
->prev
|| entry
->next
)
153 exec_list
*read(ir_variable
*var
)
155 hash_entry
*ht_entry
= _mesa_hash_table_search(lhs_ht
, var
);
157 return (exec_list
*) ht_entry
->data
;
161 void write(ir_variable
*lhs
, ir_variable
*rhs
, unsigned write_mask
, int swizzle
[4])
163 acp_entry
*entry
= new(this->lin_ctx
) acp_entry(lhs
, rhs
, write_mask
, swizzle
);
165 /* lhs hash, hash of lhs -> acp_entry lists */
166 hash_entry
*ht_entry
= _mesa_hash_table_search(lhs_ht
, lhs
);
168 exec_list
*lhs_list
= (exec_list
*) ht_entry
->data
;
169 lhs_list
->push_tail(entry
);
171 exec_list
*lhs_list
= new(mem_ctx
) exec_list
;
172 lhs_list
->push_tail(entry
);
173 _mesa_hash_table_insert(lhs_ht
, lhs
, lhs_list
);
176 /* rhs hash, hash of rhs -> acp_entry pointers to lhs lists */
177 ht_entry
= _mesa_hash_table_search(rhs_ht
, rhs
);
179 exec_list
*rhs_list
= (exec_list
*) ht_entry
->data
;
180 rhs_list
->push_tail(&entry
->rhs_node
);
182 exec_list
*rhs_list
= new(mem_ctx
) exec_list
;
183 rhs_list
->push_tail(&entry
->rhs_node
);
184 _mesa_hash_table_insert(rhs_ht
, rhs
, rhs_list
);
189 explicit copy_propagation_state(copy_propagation_state
*original
)
191 lhs_ht
= _mesa_hash_table_clone(original
->lhs_ht
, this);
192 rhs_ht
= _mesa_hash_table_clone(original
->rhs_ht
, this);
193 lin_ctx
= original
->lin_ctx
;
194 mem_ctx
= original
->mem_ctx
;
197 /** Hash of acp_entry: The available copies to propagate */
200 /** Reverse index of the lhs_ht, to optimize finding uses of a certain variable. */
207 class kill_entry
: public exec_node
210 /* override operator new from exec_node */
211 DECLARE_LINEAR_ZALLOC_CXX_OPERATORS(kill_entry
)
213 kill_entry(ir_variable
*var
, int write_mask
)
216 this->write_mask
= write_mask
;
220 unsigned int write_mask
;
223 class ir_copy_propagation_elements_visitor
: public ir_rvalue_visitor
{
225 ir_copy_propagation_elements_visitor()
227 this->progress
= false;
228 this->killed_all
= false;
229 this->mem_ctx
= ralloc_context(NULL
);
230 this->lin_ctx
= linear_alloc_parent(this->mem_ctx
, 0);
231 this->shader_mem_ctx
= NULL
;
232 this->kills
= new(mem_ctx
) exec_list
;
233 this->state
= new(mem_ctx
) copy_propagation_state(mem_ctx
, lin_ctx
);
235 ~ir_copy_propagation_elements_visitor()
237 ralloc_free(mem_ctx
);
240 void handle_loop(ir_loop
*, bool keep_acp
);
241 virtual ir_visitor_status
visit_enter(class ir_loop
*);
242 virtual ir_visitor_status
visit_enter(class ir_function_signature
*);
243 virtual ir_visitor_status
visit_leave(class ir_assignment
*);
244 virtual ir_visitor_status
visit_enter(class ir_call
*);
245 virtual ir_visitor_status
visit_enter(class ir_if
*);
246 virtual ir_visitor_status
visit_leave(class ir_swizzle
*);
248 void handle_rvalue(ir_rvalue
**rvalue
);
250 void add_copy(ir_assignment
*ir
);
251 void kill(kill_entry
*k
);
252 void handle_if_block(exec_list
*instructions
);
254 copy_propagation_state
*state
;
257 * List of kill_entry: The variables whose values were killed in this
266 /* Context for our local data structures. */
269 /* Context for allocating new shader nodes. */
270 void *shader_mem_ctx
;
273 } /* unnamed namespace */
276 ir_copy_propagation_elements_visitor::visit_enter(ir_function_signature
*ir
)
278 /* Treat entry into a function signature as a completely separate
279 * block. Any instructions at global scope will be shuffled into
280 * main() at link time, so they're irrelevant to us.
282 exec_list
*orig_kills
= this->kills
;
283 bool orig_killed_all
= this->killed_all
;
285 this->kills
= new(mem_ctx
) exec_list
;
286 this->killed_all
= false;
288 copy_propagation_state
*orig_state
= state
;
289 this->state
= new(mem_ctx
) copy_propagation_state(mem_ctx
, lin_ctx
);
291 visit_list_elements(this, &ir
->body
);
294 this->state
= orig_state
;
296 ralloc_free(this->kills
);
297 this->kills
= orig_kills
;
298 this->killed_all
= orig_killed_all
;
300 return visit_continue_with_parent
;
304 ir_copy_propagation_elements_visitor::visit_leave(ir_assignment
*ir
)
306 ir_dereference_variable
*lhs
= ir
->lhs
->as_dereference_variable();
307 ir_variable
*var
= ir
->lhs
->variable_referenced();
309 if (var
->type
->is_scalar() || var
->type
->is_vector()) {
313 k
= new(this->lin_ctx
) kill_entry(var
, ir
->write_mask
);
315 k
= new(this->lin_ctx
) kill_entry(var
, ~0);
322 return visit_continue
;
326 ir_copy_propagation_elements_visitor::visit_leave(ir_swizzle
*)
328 /* Don't visit the values of swizzles since they are handled while
329 * visiting the swizzle itself.
331 return visit_continue
;
335 * Replaces dereferences of ACP RHS variables with ACP LHS variables.
337 * This is where the actual copy propagation occurs. Note that the
338 * rewriting of ir_dereference means that the ir_dereference instance
339 * must not be shared by multiple IR operations!
342 ir_copy_propagation_elements_visitor::handle_rvalue(ir_rvalue
**ir
)
345 ir_dereference_variable
*deref_var
;
346 ir_variable
*source
[4] = {NULL
, NULL
, NULL
, NULL
};
347 int source_chan
[4] = {0, 0, 0, 0};
349 bool noop_swizzle
= true;
354 ir_swizzle
*swizzle
= (*ir
)->as_swizzle();
356 deref_var
= swizzle
->val
->as_dereference_variable();
360 swizzle_chan
[0] = swizzle
->mask
.x
;
361 swizzle_chan
[1] = swizzle
->mask
.y
;
362 swizzle_chan
[2] = swizzle
->mask
.z
;
363 swizzle_chan
[3] = swizzle
->mask
.w
;
364 chans
= swizzle
->type
->vector_elements
;
366 deref_var
= (*ir
)->as_dereference_variable();
374 chans
= deref_var
->type
->vector_elements
;
377 if (this->in_assignee
)
380 ir_variable
*var
= deref_var
->var
;
382 /* Try to find ACP entries covering swizzle_chan[], hoping they're
383 * the same source variable.
385 exec_list
*ht_list
= state
->read(var
);
387 foreach_in_list(acp_entry
, entry
, ht_list
) {
388 for (int c
= 0; c
< chans
; c
++) {
389 if (entry
->write_mask
& (1 << swizzle_chan
[c
])) {
390 source
[c
] = entry
->rhs
;
391 source_chan
[c
] = entry
->swizzle
[swizzle_chan
[c
]];
393 if (source_chan
[c
] != swizzle_chan
[c
])
394 noop_swizzle
= false;
400 /* Make sure all channels are copying from the same source variable. */
403 for (int c
= 1; c
< chans
; c
++) {
404 if (source
[c
] != source
[0])
409 shader_mem_ctx
= ralloc_parent(deref_var
);
411 /* Don't pointlessly replace the rvalue with itself (or a noop swizzle
412 * of itself, which would just be deleted by opt_noop_swizzle).
414 if (source
[0] == var
&& noop_swizzle
)
418 printf("Copy propagation from:\n");
422 deref_var
= new(shader_mem_ctx
) ir_dereference_variable(source
[0]);
423 *ir
= new(shader_mem_ctx
) ir_swizzle(deref_var
,
440 ir_copy_propagation_elements_visitor::visit_enter(ir_call
*ir
)
442 /* Do copy propagation on call parameters, but skip any out params */
443 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
444 actual_node
, &ir
->actual_parameters
) {
445 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
446 ir_rvalue
*ir
= (ir_rvalue
*) actual_node
;
447 if (sig_param
->data
.mode
!= ir_var_function_out
448 && sig_param
->data
.mode
!= ir_var_function_inout
) {
453 /* Since we're unlinked, we don't (necessarily) know the side effects of
454 * this call. So kill all copies.
456 this->state
->erase_all();
457 this->killed_all
= true;
459 return visit_continue_with_parent
;
463 ir_copy_propagation_elements_visitor::handle_if_block(exec_list
*instructions
)
465 exec_list
*orig_kills
= this->kills
;
466 bool orig_killed_all
= this->killed_all
;
468 this->kills
= new(mem_ctx
) exec_list
;
469 this->killed_all
= false;
471 /* Populate the initial acp with a copy of the original */
472 copy_propagation_state
*orig_state
= state
;
473 this->state
= orig_state
->clone();
475 visit_list_elements(this, instructions
);
478 this->state
= orig_state
;
480 if (this->killed_all
)
481 this->state
->erase_all();
483 exec_list
*new_kills
= this->kills
;
484 this->kills
= orig_kills
;
485 this->killed_all
= this->killed_all
|| orig_killed_all
;
487 /* Move the new kills into the parent block's list, removing them
488 * from the parent's ACP list in the process.
490 foreach_in_list_safe(kill_entry
, k
, new_kills
) {
494 ralloc_free(new_kills
);
498 ir_copy_propagation_elements_visitor::visit_enter(ir_if
*ir
)
500 ir
->condition
->accept(this);
502 handle_if_block(&ir
->then_instructions
);
503 handle_if_block(&ir
->else_instructions
);
505 /* handle_if_block() already descended into the children. */
506 return visit_continue_with_parent
;
510 ir_copy_propagation_elements_visitor::handle_loop(ir_loop
*ir
, bool keep_acp
)
512 exec_list
*orig_kills
= this->kills
;
513 bool orig_killed_all
= this->killed_all
;
515 this->kills
= new(mem_ctx
) exec_list
;
516 this->killed_all
= false;
518 copy_propagation_state
*orig_state
= state
;
521 /* Populate the initial acp with a copy of the original */
522 this->state
= orig_state
->clone();
524 this->state
= new(mem_ctx
) copy_propagation_state(mem_ctx
, lin_ctx
);
527 visit_list_elements(this, &ir
->body_instructions
);
530 this->state
= orig_state
;
532 if (this->killed_all
)
533 this->state
->erase_all();
535 exec_list
*new_kills
= this->kills
;
536 this->kills
= orig_kills
;
537 this->killed_all
= this->killed_all
|| orig_killed_all
;
539 foreach_in_list_safe(kill_entry
, k
, new_kills
) {
543 ralloc_free(new_kills
);
547 ir_copy_propagation_elements_visitor::visit_enter(ir_loop
*ir
)
549 handle_loop(ir
, false);
550 handle_loop(ir
, true);
552 /* already descended into the children. */
553 return visit_continue_with_parent
;
556 /* Remove any entries currently in the ACP for this kill. */
558 ir_copy_propagation_elements_visitor::kill(kill_entry
*k
)
560 state
->erase(k
->var
, k
->write_mask
);
562 /* If we were on a list, remove ourselves before inserting */
566 this->kills
->push_tail(k
);
570 * Adds directly-copied channels between vector variables to the available
571 * copy propagation list.
574 ir_copy_propagation_elements_visitor::add_copy(ir_assignment
*ir
)
576 int orig_swizzle
[4] = {0, 1, 2, 3};
582 ir_dereference_variable
*lhs
= ir
->lhs
->as_dereference_variable();
583 if (!lhs
|| !(lhs
->type
->is_scalar() || lhs
->type
->is_vector()))
586 if (lhs
->var
->data
.mode
== ir_var_shader_storage
||
587 lhs
->var
->data
.mode
== ir_var_shader_shared
)
590 ir_dereference_variable
*rhs
= ir
->rhs
->as_dereference_variable();
592 ir_swizzle
*swiz
= ir
->rhs
->as_swizzle();
596 rhs
= swiz
->val
->as_dereference_variable();
600 orig_swizzle
[0] = swiz
->mask
.x
;
601 orig_swizzle
[1] = swiz
->mask
.y
;
602 orig_swizzle
[2] = swiz
->mask
.z
;
603 orig_swizzle
[3] = swiz
->mask
.w
;
606 if (rhs
->var
->data
.mode
== ir_var_shader_storage
||
607 rhs
->var
->data
.mode
== ir_var_shader_shared
)
610 /* Move the swizzle channels out to the positions they match in the
611 * destination. We don't want to have to rewrite the swizzle[]
612 * array every time we clear a bit of the write_mask.
615 for (int i
= 0; i
< 4; i
++) {
616 if (ir
->write_mask
& (1 << i
))
617 swizzle
[i
] = orig_swizzle
[j
++];
620 int write_mask
= ir
->write_mask
;
621 if (lhs
->var
== rhs
->var
) {
622 /* If this is a copy from the variable to itself, then we need
623 * to be sure not to include the updated channels from this
624 * instruction in the set of new source channels to be
625 * copy-propagated from.
627 for (int i
= 0; i
< 4; i
++) {
628 if (ir
->write_mask
& (1 << orig_swizzle
[i
]))
629 write_mask
&= ~(1 << i
);
633 if (lhs
->var
->data
.precise
!= rhs
->var
->data
.precise
)
636 state
->write(lhs
->var
, rhs
->var
, write_mask
, swizzle
);
640 do_copy_propagation_elements(exec_list
*instructions
)
642 ir_copy_propagation_elements_visitor v
;
644 visit_list_elements(&v
, instructions
);