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 "glsl_types.h"
50 static bool debug
= false;
54 class acp_entry
: public exec_node
57 acp_entry(ir_variable
*lhs
, ir_variable
*rhs
, int write_mask
, int swizzle
[4])
61 this->write_mask
= write_mask
;
62 memcpy(this->swizzle
, swizzle
, sizeof(this->swizzle
));
65 acp_entry(acp_entry
*a
)
69 this->write_mask
= a
->write_mask
;
70 memcpy(this->swizzle
, a
->swizzle
, sizeof(this->swizzle
));
75 unsigned int write_mask
;
80 class kill_entry
: public exec_node
83 kill_entry(ir_variable
*var
, int write_mask
)
86 this->write_mask
= write_mask
;
90 unsigned int write_mask
;
93 class ir_copy_propagation_elements_visitor
: public ir_rvalue_visitor
{
95 ir_copy_propagation_elements_visitor()
97 this->progress
= false;
98 this->killed_all
= false;
99 this->mem_ctx
= ralloc_context(NULL
);
100 this->shader_mem_ctx
= NULL
;
101 this->acp
= new(mem_ctx
) exec_list
;
102 this->kills
= new(mem_ctx
) exec_list
;
104 ~ir_copy_propagation_elements_visitor()
106 ralloc_free(mem_ctx
);
109 virtual ir_visitor_status
visit_enter(class ir_loop
*);
110 virtual ir_visitor_status
visit_enter(class ir_function_signature
*);
111 virtual ir_visitor_status
visit_leave(class ir_assignment
*);
112 virtual ir_visitor_status
visit_enter(class ir_call
*);
113 virtual ir_visitor_status
visit_enter(class ir_if
*);
114 virtual ir_visitor_status
visit_leave(class ir_swizzle
*);
116 void handle_rvalue(ir_rvalue
**rvalue
);
118 void add_copy(ir_assignment
*ir
);
119 void kill(kill_entry
*k
);
120 void handle_if_block(exec_list
*instructions
);
122 /** List of acp_entry: The available copies to propagate */
125 * List of kill_entry: The variables whose values were killed in this
134 /* Context for our local data structures. */
136 /* Context for allocating new shader nodes. */
137 void *shader_mem_ctx
;
140 } /* unnamed namespace */
143 ir_copy_propagation_elements_visitor::visit_enter(ir_function_signature
*ir
)
145 /* Treat entry into a function signature as a completely separate
146 * block. Any instructions at global scope will be shuffled into
147 * main() at link time, so they're irrelevant to us.
149 exec_list
*orig_acp
= this->acp
;
150 exec_list
*orig_kills
= this->kills
;
151 bool orig_killed_all
= this->killed_all
;
153 this->acp
= new(mem_ctx
) exec_list
;
154 this->kills
= new(mem_ctx
) exec_list
;
155 this->killed_all
= false;
157 visit_list_elements(this, &ir
->body
);
159 ralloc_free(this->acp
);
160 ralloc_free(this->kills
);
162 this->kills
= orig_kills
;
163 this->acp
= orig_acp
;
164 this->killed_all
= orig_killed_all
;
166 return visit_continue_with_parent
;
170 ir_copy_propagation_elements_visitor::visit_leave(ir_assignment
*ir
)
172 ir_dereference_variable
*lhs
= ir
->lhs
->as_dereference_variable();
173 ir_variable
*var
= ir
->lhs
->variable_referenced();
175 if (var
->type
->is_scalar() || var
->type
->is_vector()) {
179 k
= new(this->kills
) kill_entry(var
, ir
->write_mask
);
181 k
= new(this->kills
) kill_entry(var
, ~0);
188 return visit_continue
;
192 ir_copy_propagation_elements_visitor::visit_leave(ir_swizzle
*)
194 /* Don't visit the values of swizzles since they are handled while
195 * visiting the swizzle itself.
197 return visit_continue
;
201 * Replaces dereferences of ACP RHS variables with ACP LHS variables.
203 * This is where the actual copy propagation occurs. Note that the
204 * rewriting of ir_dereference means that the ir_dereference instance
205 * must not be shared by multiple IR operations!
208 ir_copy_propagation_elements_visitor::handle_rvalue(ir_rvalue
**ir
)
211 ir_dereference_variable
*deref_var
;
212 ir_variable
*source
[4] = {NULL
, NULL
, NULL
, NULL
};
213 int source_chan
[4] = {0, 0, 0, 0};
215 bool noop_swizzle
= true;
220 ir_swizzle
*swizzle
= (*ir
)->as_swizzle();
222 deref_var
= swizzle
->val
->as_dereference_variable();
226 swizzle_chan
[0] = swizzle
->mask
.x
;
227 swizzle_chan
[1] = swizzle
->mask
.y
;
228 swizzle_chan
[2] = swizzle
->mask
.z
;
229 swizzle_chan
[3] = swizzle
->mask
.w
;
230 chans
= swizzle
->type
->vector_elements
;
232 deref_var
= (*ir
)->as_dereference_variable();
240 chans
= deref_var
->type
->vector_elements
;
243 if (this->in_assignee
)
246 ir_variable
*var
= deref_var
->var
;
248 /* Try to find ACP entries covering swizzle_chan[], hoping they're
249 * the same source variable.
251 foreach_in_list(acp_entry
, entry
, this->acp
) {
252 if (var
== entry
->lhs
) {
253 for (int c
= 0; c
< chans
; c
++) {
254 if (entry
->write_mask
& (1 << swizzle_chan
[c
])) {
255 source
[c
] = entry
->rhs
;
256 source_chan
[c
] = entry
->swizzle
[swizzle_chan
[c
]];
258 if (source_chan
[c
] != swizzle_chan
[c
])
259 noop_swizzle
= false;
265 /* Make sure all channels are copying from the same source variable. */
268 for (int c
= 1; c
< chans
; c
++) {
269 if (source
[c
] != source
[0])
274 shader_mem_ctx
= ralloc_parent(deref_var
);
276 /* Don't pointlessly replace the rvalue with itself (or a noop swizzle
277 * of itself, which would just be deleted by opt_noop_swizzle).
279 if (source
[0] == var
&& noop_swizzle
)
283 printf("Copy propagation from:\n");
287 deref_var
= new(shader_mem_ctx
) ir_dereference_variable(source
[0]);
288 *ir
= new(shader_mem_ctx
) ir_swizzle(deref_var
,
305 ir_copy_propagation_elements_visitor::visit_enter(ir_call
*ir
)
307 /* Do copy propagation on call parameters, but skip any out params */
308 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
309 actual_node
, &ir
->actual_parameters
) {
310 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
311 ir_rvalue
*ir
= (ir_rvalue
*) actual_node
;
312 if (sig_param
->data
.mode
!= ir_var_function_out
313 && sig_param
->data
.mode
!= ir_var_function_inout
) {
318 /* Since we're unlinked, we don't (necessarily) know the side effects of
319 * this call. So kill all copies.
322 this->killed_all
= true;
324 return visit_continue_with_parent
;
328 ir_copy_propagation_elements_visitor::handle_if_block(exec_list
*instructions
)
330 exec_list
*orig_acp
= this->acp
;
331 exec_list
*orig_kills
= this->kills
;
332 bool orig_killed_all
= this->killed_all
;
334 this->acp
= new(mem_ctx
) exec_list
;
335 this->kills
= new(mem_ctx
) exec_list
;
336 this->killed_all
= false;
338 /* Populate the initial acp with a copy of the original */
339 foreach_in_list(acp_entry
, a
, orig_acp
) {
340 this->acp
->push_tail(new(this->acp
) acp_entry(a
));
343 visit_list_elements(this, instructions
);
345 if (this->killed_all
) {
346 orig_acp
->make_empty();
349 exec_list
*new_kills
= this->kills
;
350 this->kills
= orig_kills
;
351 ralloc_free(this->acp
);
352 this->acp
= orig_acp
;
353 this->killed_all
= this->killed_all
|| orig_killed_all
;
355 /* Move the new kills into the parent block's list, removing them
356 * from the parent's ACP list in the process.
358 foreach_in_list_safe(kill_entry
, k
, new_kills
) {
362 ralloc_free(new_kills
);
366 ir_copy_propagation_elements_visitor::visit_enter(ir_if
*ir
)
368 ir
->condition
->accept(this);
370 handle_if_block(&ir
->then_instructions
);
371 handle_if_block(&ir
->else_instructions
);
373 /* handle_if_block() already descended into the children. */
374 return visit_continue_with_parent
;
378 ir_copy_propagation_elements_visitor::visit_enter(ir_loop
*ir
)
380 exec_list
*orig_acp
= this->acp
;
381 exec_list
*orig_kills
= this->kills
;
382 bool orig_killed_all
= this->killed_all
;
384 /* FINISHME: For now, the initial acp for loops is totally empty.
385 * We could go through once, then go through again with the acp
386 * cloned minus the killed entries after the first run through.
388 this->acp
= new(mem_ctx
) exec_list
;
389 this->kills
= new(mem_ctx
) exec_list
;
390 this->killed_all
= false;
392 visit_list_elements(this, &ir
->body_instructions
);
394 if (this->killed_all
) {
395 orig_acp
->make_empty();
398 exec_list
*new_kills
= this->kills
;
399 this->kills
= orig_kills
;
400 ralloc_free(this->acp
);
401 this->acp
= orig_acp
;
402 this->killed_all
= this->killed_all
|| orig_killed_all
;
404 foreach_in_list_safe(kill_entry
, k
, new_kills
) {
408 ralloc_free(new_kills
);
410 /* already descended into the children. */
411 return visit_continue_with_parent
;
414 /* Remove any entries currently in the ACP for this kill. */
416 ir_copy_propagation_elements_visitor::kill(kill_entry
*k
)
418 foreach_in_list_safe(acp_entry
, entry
, acp
) {
419 if (entry
->lhs
== k
->var
) {
420 entry
->write_mask
= entry
->write_mask
& ~k
->write_mask
;
421 if (entry
->write_mask
== 0) {
426 if (entry
->rhs
== k
->var
) {
431 /* If we were on a list, remove ourselves before inserting */
435 ralloc_steal(this->kills
, k
);
436 this->kills
->push_tail(k
);
440 * Adds directly-copied channels between vector variables to the available
441 * copy propagation list.
444 ir_copy_propagation_elements_visitor::add_copy(ir_assignment
*ir
)
447 int orig_swizzle
[4] = {0, 1, 2, 3};
453 ir_dereference_variable
*lhs
= ir
->lhs
->as_dereference_variable();
454 if (!lhs
|| !(lhs
->type
->is_scalar() || lhs
->type
->is_vector()))
457 ir_dereference_variable
*rhs
= ir
->rhs
->as_dereference_variable();
459 ir_swizzle
*swiz
= ir
->rhs
->as_swizzle();
463 rhs
= swiz
->val
->as_dereference_variable();
467 orig_swizzle
[0] = swiz
->mask
.x
;
468 orig_swizzle
[1] = swiz
->mask
.y
;
469 orig_swizzle
[2] = swiz
->mask
.z
;
470 orig_swizzle
[3] = swiz
->mask
.w
;
473 /* Move the swizzle channels out to the positions they match in the
474 * destination. We don't want to have to rewrite the swizzle[]
475 * array every time we clear a bit of the write_mask.
478 for (int i
= 0; i
< 4; i
++) {
479 if (ir
->write_mask
& (1 << i
))
480 swizzle
[i
] = orig_swizzle
[j
++];
483 int write_mask
= ir
->write_mask
;
484 if (lhs
->var
== rhs
->var
) {
485 /* If this is a copy from the variable to itself, then we need
486 * to be sure not to include the updated channels from this
487 * instruction in the set of new source channels to be
488 * copy-propagated from.
490 for (int i
= 0; i
< 4; i
++) {
491 if (ir
->write_mask
& (1 << orig_swizzle
[i
]))
492 write_mask
&= ~(1 << i
);
496 entry
= new(this->mem_ctx
) acp_entry(lhs
->var
, rhs
->var
, write_mask
,
498 this->acp
->push_tail(entry
);
502 do_copy_propagation_elements(exec_list
*instructions
)
504 ir_copy_propagation_elements_visitor v
;
506 visit_list_elements(&v
, instructions
);