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.cpp
27 * Moves usage of recently-copied variables to the previous copy of
30 * This should reduce the number of MOV instructions in the generated
31 * programs unless copy propagation is also done on the LIR, and may
32 * help anyway by triggering other optimizations that live in the HIR.
36 #include "ir_visitor.h"
37 #include "ir_basic_block.h"
38 #include "ir_optimization.h"
39 #include "compiler/glsl_types.h"
43 class acp_entry
: public exec_node
46 acp_entry(ir_variable
*lhs
, ir_variable
*rhs
)
59 class kill_entry
: public exec_node
62 kill_entry(ir_variable
*var
)
71 class ir_copy_propagation_visitor
: public ir_hierarchical_visitor
{
73 ir_copy_propagation_visitor()
76 mem_ctx
= ralloc_context(0);
77 this->acp
= new(mem_ctx
) exec_list
;
78 this->kills
= new(mem_ctx
) exec_list
;
81 ~ir_copy_propagation_visitor()
86 virtual ir_visitor_status
visit(class ir_dereference_variable
*);
87 void handle_loop(class ir_loop
*, bool keep_acp
);
88 virtual ir_visitor_status
visit_enter(class ir_loop
*);
89 virtual ir_visitor_status
visit_enter(class ir_function_signature
*);
90 virtual ir_visitor_status
visit_enter(class ir_function
*);
91 virtual ir_visitor_status
visit_leave(class ir_assignment
*);
92 virtual ir_visitor_status
visit_enter(class ir_call
*);
93 virtual ir_visitor_status
visit_enter(class ir_if
*);
95 void add_copy(ir_assignment
*ir
);
96 void kill(ir_variable
*ir
);
97 void handle_if_block(exec_list
*instructions
);
99 /** List of acp_entry: The available copies to propagate */
102 * List of kill_entry: The variables whose values were killed in this
114 } /* unnamed namespace */
117 ir_copy_propagation_visitor::visit_enter(ir_function_signature
*ir
)
119 /* Treat entry into a function signature as a completely separate
120 * block. Any instructions at global scope will be shuffled into
121 * main() at link time, so they're irrelevant to us.
123 exec_list
*orig_acp
= this->acp
;
124 exec_list
*orig_kills
= this->kills
;
125 bool orig_killed_all
= this->killed_all
;
127 this->acp
= new(mem_ctx
) exec_list
;
128 this->kills
= new(mem_ctx
) exec_list
;
129 this->killed_all
= false;
131 visit_list_elements(this, &ir
->body
);
133 ralloc_free(this->acp
);
134 ralloc_free(this->kills
);
136 this->kills
= orig_kills
;
137 this->acp
= orig_acp
;
138 this->killed_all
= orig_killed_all
;
140 return visit_continue_with_parent
;
144 ir_copy_propagation_visitor::visit_leave(ir_assignment
*ir
)
146 kill(ir
->lhs
->variable_referenced());
150 return visit_continue
;
154 ir_copy_propagation_visitor::visit_enter(ir_function
*ir
)
157 return visit_continue
;
161 * Replaces dereferences of ACP RHS variables with ACP LHS variables.
163 * This is where the actual copy propagation occurs. Note that the
164 * rewriting of ir_dereference means that the ir_dereference instance
165 * must not be shared by multiple IR operations!
168 ir_copy_propagation_visitor::visit(ir_dereference_variable
*ir
)
170 if (this->in_assignee
)
171 return visit_continue
;
173 ir_variable
*var
= ir
->var
;
175 foreach_in_list(acp_entry
, entry
, this->acp
) {
176 if (var
== entry
->lhs
) {
177 ir
->var
= entry
->rhs
;
178 this->progress
= true;
183 return visit_continue
;
188 ir_copy_propagation_visitor::visit_enter(ir_call
*ir
)
190 /* Do copy propagation on call parameters, but skip any out params */
191 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
192 actual_node
, &ir
->actual_parameters
) {
193 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
194 ir_rvalue
*ir
= (ir_rvalue
*) actual_node
;
195 if (sig_param
->data
.mode
!= ir_var_function_out
196 && sig_param
->data
.mode
!= ir_var_function_inout
) {
201 /* Since we're unlinked, we don't (necessarily) know the side effects of
202 * this call. So kill all copies.
205 this->killed_all
= true;
207 return visit_continue_with_parent
;
211 ir_copy_propagation_visitor::handle_if_block(exec_list
*instructions
)
213 exec_list
*orig_acp
= this->acp
;
214 exec_list
*orig_kills
= this->kills
;
215 bool orig_killed_all
= this->killed_all
;
217 this->acp
= new(mem_ctx
) exec_list
;
218 this->kills
= new(mem_ctx
) exec_list
;
219 this->killed_all
= false;
221 /* Populate the initial acp with a copy of the original */
222 foreach_in_list(acp_entry
, a
, orig_acp
) {
223 this->acp
->push_tail(new(this->acp
) acp_entry(a
->lhs
, a
->rhs
));
226 visit_list_elements(this, instructions
);
228 if (this->killed_all
) {
229 orig_acp
->make_empty();
232 exec_list
*new_kills
= this->kills
;
233 this->kills
= orig_kills
;
234 ralloc_free(this->acp
);
235 this->acp
= orig_acp
;
236 this->killed_all
= this->killed_all
|| orig_killed_all
;
238 foreach_in_list(kill_entry
, k
, new_kills
) {
242 ralloc_free(new_kills
);
246 ir_copy_propagation_visitor::visit_enter(ir_if
*ir
)
248 ir
->condition
->accept(this);
250 handle_if_block(&ir
->then_instructions
);
251 handle_if_block(&ir
->else_instructions
);
253 /* handle_if_block() already descended into the children. */
254 return visit_continue_with_parent
;
258 ir_copy_propagation_visitor::handle_loop(ir_loop
*ir
, bool keep_acp
)
260 exec_list
*orig_acp
= this->acp
;
261 exec_list
*orig_kills
= this->kills
;
262 bool orig_killed_all
= this->killed_all
;
264 this->acp
= new(mem_ctx
) exec_list
;
265 this->kills
= new(mem_ctx
) exec_list
;
266 this->killed_all
= false;
269 /* Populate the initial acp with a copy of the original */
270 foreach_in_list(acp_entry
, a
, orig_acp
) {
271 this->acp
->push_tail(new(this->acp
) acp_entry(a
->lhs
, a
->rhs
));
275 visit_list_elements(this, &ir
->body_instructions
);
277 if (this->killed_all
) {
278 orig_acp
->make_empty();
281 exec_list
*new_kills
= this->kills
;
282 this->kills
= orig_kills
;
283 ralloc_free(this->acp
);
284 this->acp
= orig_acp
;
285 this->killed_all
= this->killed_all
|| orig_killed_all
;
287 foreach_in_list(kill_entry
, k
, new_kills
) {
291 ralloc_free(new_kills
);
295 ir_copy_propagation_visitor::visit_enter(ir_loop
*ir
)
297 /* Make a conservative first pass over the loop with an empty ACP set.
298 * This also removes any killed entries from the original ACP set.
300 handle_loop(ir
, false);
302 /* Then, run it again with the real ACP set, minus any killed entries.
303 * This takes care of propagating values from before the loop into it.
305 handle_loop(ir
, true);
307 /* already descended into the children. */
308 return visit_continue_with_parent
;
312 ir_copy_propagation_visitor::kill(ir_variable
*var
)
316 /* Remove any entries currently in the ACP for this kill. */
317 foreach_in_list_safe(acp_entry
, entry
, acp
) {
318 if (entry
->lhs
== var
|| entry
->rhs
== var
) {
323 /* Add the LHS variable to the list of killed variables in this block.
325 this->kills
->push_tail(new(this->kills
) kill_entry(var
));
329 * Adds an entry to the available copy list if it's a plain assignment
330 * of a variable to a variable.
333 ir_copy_propagation_visitor::add_copy(ir_assignment
*ir
)
340 ir_variable
*lhs_var
= ir
->whole_variable_written();
341 ir_variable
*rhs_var
= ir
->rhs
->whole_variable_referenced();
343 if ((lhs_var
!= NULL
) && (rhs_var
!= NULL
)) {
344 if (lhs_var
== rhs_var
) {
345 /* This is a dumb assignment, but we've conveniently noticed
346 * it here. Removing it now would mess up the loop iteration
347 * calling us. Just flag it to not execute, and someone else
348 * will clean up the mess.
350 ir
->condition
= new(ralloc_parent(ir
)) ir_constant(false);
351 this->progress
= true;
352 } else if (lhs_var
->data
.mode
!= ir_var_shader_storage
&&
353 lhs_var
->data
.mode
!= ir_var_shader_shared
&&
354 lhs_var
->data
.precise
== rhs_var
->data
.precise
) {
355 entry
= new(this->acp
) acp_entry(lhs_var
, rhs_var
);
356 this->acp
->push_tail(entry
);
362 * Does a copy propagation pass on the code present in the instruction stream.
365 do_copy_propagation(exec_list
*instructions
)
367 ir_copy_propagation_visitor v
;
369 visit_list_elements(&v
, instructions
);