2 * Copyright © 2010 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * constant 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, constant, 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 constantright 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 CONSTANTRIGHT 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_constant_propagation.cpp
27 * Tracks assignments of constants to channels of variables, and
28 * usage of those constant channels with direct usage of the constants.
30 * This can lead to constant folding and algebraic optimizations in
31 * those later expressions, while causing no increase in instruction
32 * count (due to constants being generally free to load from a
33 * constant push buffer or as instruction immediate values) and
34 * possibly reducing register pressure.
38 #include "ir_visitor.h"
39 #include "ir_rvalue_visitor.h"
40 #include "ir_basic_block.h"
41 #include "ir_optimization.h"
42 #include "compiler/glsl_types.h"
43 #include "util/hash_table.h"
47 class acp_entry
: public exec_node
50 /* override operator new from exec_node */
51 DECLARE_LINEAR_ZALLOC_CXX_OPERATORS(acp_entry
)
53 acp_entry(ir_variable
*var
, unsigned write_mask
, ir_constant
*constant
)
58 this->write_mask
= write_mask
;
59 this->constant
= constant
;
60 this->initial_values
= write_mask
;
63 acp_entry(const acp_entry
*src
)
66 this->write_mask
= src
->write_mask
;
67 this->constant
= src
->constant
;
68 this->initial_values
= src
->initial_values
;
72 ir_constant
*constant
;
75 /** Mask of values initially available in the constant. */
76 unsigned initial_values
;
80 class kill_entry
: public exec_node
83 /* override operator new from exec_node */
84 DECLARE_LINEAR_ZALLOC_CXX_OPERATORS(kill_entry
)
86 kill_entry(ir_variable
*var
, unsigned write_mask
)
90 this->write_mask
= write_mask
;
97 class ir_constant_propagation_visitor
: public ir_rvalue_visitor
{
99 ir_constant_propagation_visitor()
103 mem_ctx
= ralloc_context(0);
104 this->lin_ctx
= linear_alloc_parent(this->mem_ctx
, 0);
105 this->acp
= new(mem_ctx
) exec_list
;
106 this->kills
= _mesa_hash_table_create(mem_ctx
, _mesa_hash_pointer
,
107 _mesa_key_pointer_equal
);
109 ~ir_constant_propagation_visitor()
111 ralloc_free(mem_ctx
);
114 virtual ir_visitor_status
visit_enter(class ir_loop
*);
115 virtual ir_visitor_status
visit_enter(class ir_function_signature
*);
116 virtual ir_visitor_status
visit_enter(class ir_function
*);
117 virtual ir_visitor_status
visit_leave(class ir_assignment
*);
118 virtual ir_visitor_status
visit_enter(class ir_call
*);
119 virtual ir_visitor_status
visit_enter(class ir_if
*);
121 void add_constant(ir_assignment
*ir
);
122 void constant_folding(ir_rvalue
**rvalue
);
123 void constant_propagation(ir_rvalue
**rvalue
);
124 void kill(ir_variable
*ir
, unsigned write_mask
);
125 void handle_if_block(exec_list
*instructions
);
126 void handle_rvalue(ir_rvalue
**rvalue
);
128 /** List of acp_entry: The available constants to propagate */
132 * Hash table of kill_entry: The masks of variables whose values were
133 * killed in this block.
147 ir_constant_propagation_visitor::constant_folding(ir_rvalue
**rvalue
)
149 if (this->in_assignee
|| *rvalue
== NULL
)
152 if (ir_constant_fold(rvalue
))
153 this->progress
= true;
155 ir_dereference_variable
*var_ref
= (*rvalue
)->as_dereference_variable();
156 if (var_ref
&& !var_ref
->type
->is_array()) {
157 ir_constant
*constant
= var_ref
->constant_expression_value();
160 this->progress
= true;
166 ir_constant_propagation_visitor::constant_propagation(ir_rvalue
**rvalue
) {
168 if (this->in_assignee
|| !*rvalue
)
171 const glsl_type
*type
= (*rvalue
)->type
;
172 if (!type
->is_scalar() && !type
->is_vector())
175 ir_swizzle
*swiz
= NULL
;
176 ir_dereference_variable
*deref
= (*rvalue
)->as_dereference_variable();
178 swiz
= (*rvalue
)->as_swizzle();
182 deref
= swiz
->val
->as_dereference_variable();
187 ir_constant_data data
;
188 memset(&data
, 0, sizeof(data
));
190 for (unsigned int i
= 0; i
< type
->components(); i
++) {
192 acp_entry
*found
= NULL
;
196 case 0: channel
= swiz
->mask
.x
; break;
197 case 1: channel
= swiz
->mask
.y
; break;
198 case 2: channel
= swiz
->mask
.z
; break;
199 case 3: channel
= swiz
->mask
.w
; break;
200 default: assert(!"shouldn't be reached"); channel
= 0; break;
206 foreach_in_list(acp_entry
, entry
, this->acp
) {
207 if (entry
->var
== deref
->var
&& entry
->write_mask
& (1 << channel
)) {
217 for (int j
= 0; j
< 4; j
++) {
220 if (found
->initial_values
& (1 << j
))
224 switch (type
->base_type
) {
225 case GLSL_TYPE_FLOAT
:
226 data
.f
[i
] = found
->constant
->value
.f
[rhs_channel
];
228 case GLSL_TYPE_DOUBLE
:
229 data
.d
[i
] = found
->constant
->value
.d
[rhs_channel
];
232 data
.i
[i
] = found
->constant
->value
.i
[rhs_channel
];
235 data
.u
[i
] = found
->constant
->value
.u
[rhs_channel
];
238 data
.b
[i
] = found
->constant
->value
.b
[rhs_channel
];
241 assert(!"not reached");
246 *rvalue
= new(ralloc_parent(deref
)) ir_constant(type
, &data
);
247 this->progress
= true;
251 ir_constant_propagation_visitor::handle_rvalue(ir_rvalue
**rvalue
)
253 constant_propagation(rvalue
);
254 constant_folding(rvalue
);
258 ir_constant_propagation_visitor::visit_enter(ir_function_signature
*ir
)
260 /* Treat entry into a function signature as a completely separate
261 * block. Any instructions at global scope will be shuffled into
262 * main() at link time, so they're irrelevant to us.
264 exec_list
*orig_acp
= this->acp
;
265 hash_table
*orig_kills
= this->kills
;
266 bool orig_killed_all
= this->killed_all
;
268 this->acp
= new(mem_ctx
) exec_list
;
269 this->kills
= _mesa_hash_table_create(mem_ctx
, _mesa_hash_pointer
,
270 _mesa_key_pointer_equal
);
271 this->killed_all
= false;
273 visit_list_elements(this, &ir
->body
);
275 this->kills
= orig_kills
;
276 this->acp
= orig_acp
;
277 this->killed_all
= orig_killed_all
;
279 return visit_continue_with_parent
;
283 ir_constant_propagation_visitor::visit_leave(ir_assignment
*ir
)
285 constant_folding(&ir
->rhs
);
287 if (this->in_assignee
)
288 return visit_continue
;
290 unsigned kill_mask
= ir
->write_mask
;
291 if (ir
->lhs
->as_dereference_array()) {
292 /* The LHS of the assignment uses an array indexing operator (e.g. v[i]
293 * = ...;). Since we only try to constant propagate vectors and
294 * scalars, this means that either (a) array indexing is being used to
295 * select a vector component, or (b) the variable in question is neither
296 * a scalar or a vector, so we don't care about it. In the former case,
297 * we want to kill the whole vector, since in general we can't predict
298 * which vector component will be selected by array indexing. In the
299 * latter case, it doesn't matter what we do, so go ahead and kill the
300 * whole variable anyway.
302 * Note that if the array index is constant (e.g. v[2] = ...;), we could
303 * in principle be smarter, but we don't need to, because a future
304 * optimization pass will convert it to a simple assignment with the
309 kill(ir
->lhs
->variable_referenced(), kill_mask
);
313 return visit_continue
;
317 ir_constant_propagation_visitor::visit_enter(ir_function
*ir
)
320 return visit_continue
;
324 ir_constant_propagation_visitor::visit_enter(ir_call
*ir
)
326 /* Do constant propagation on call parameters, but skip any out params */
327 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
328 actual_node
, &ir
->actual_parameters
) {
329 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
330 ir_rvalue
*param
= (ir_rvalue
*) actual_node
;
331 if (sig_param
->data
.mode
!= ir_var_function_out
332 && sig_param
->data
.mode
!= ir_var_function_inout
) {
333 ir_rvalue
*new_param
= param
;
334 handle_rvalue(&new_param
);
335 if (new_param
!= param
)
336 param
->replace_with(new_param
);
342 /* Since we're unlinked, we don't (necssarily) know the side effects of
343 * this call. So kill all copies.
346 this->killed_all
= true;
348 return visit_continue_with_parent
;
352 ir_constant_propagation_visitor::handle_if_block(exec_list
*instructions
)
354 exec_list
*orig_acp
= this->acp
;
355 hash_table
*orig_kills
= this->kills
;
356 bool orig_killed_all
= this->killed_all
;
358 this->acp
= new(mem_ctx
) exec_list
;
359 this->kills
= _mesa_hash_table_create(mem_ctx
, _mesa_hash_pointer
,
360 _mesa_key_pointer_equal
);
361 this->killed_all
= false;
363 /* Populate the initial acp with a constant of the original */
364 foreach_in_list(acp_entry
, a
, orig_acp
) {
365 this->acp
->push_tail(new(this->lin_ctx
) acp_entry(a
));
368 visit_list_elements(this, instructions
);
370 if (this->killed_all
) {
371 orig_acp
->make_empty();
374 hash_table
*new_kills
= this->kills
;
375 this->kills
= orig_kills
;
376 this->acp
= orig_acp
;
377 this->killed_all
= this->killed_all
|| orig_killed_all
;
380 hash_table_foreach(new_kills
, htk
) {
381 kill_entry
*k
= (kill_entry
*) htk
->data
;
382 kill(k
->var
, k
->write_mask
);
387 ir_constant_propagation_visitor::visit_enter(ir_if
*ir
)
389 ir
->condition
->accept(this);
390 handle_rvalue(&ir
->condition
);
392 handle_if_block(&ir
->then_instructions
);
393 handle_if_block(&ir
->else_instructions
);
395 /* handle_if_block() already descended into the children. */
396 return visit_continue_with_parent
;
400 ir_constant_propagation_visitor::visit_enter(ir_loop
*ir
)
402 exec_list
*orig_acp
= this->acp
;
403 hash_table
*orig_kills
= this->kills
;
404 bool orig_killed_all
= this->killed_all
;
406 /* FINISHME: For now, the initial acp for loops is totally empty.
407 * We could go through once, then go through again with the acp
408 * cloned minus the killed entries after the first run through.
410 this->acp
= new(mem_ctx
) exec_list
;
411 this->kills
= _mesa_hash_table_create(mem_ctx
, _mesa_hash_pointer
,
412 _mesa_key_pointer_equal
);
413 this->killed_all
= false;
415 visit_list_elements(this, &ir
->body_instructions
);
417 if (this->killed_all
) {
418 orig_acp
->make_empty();
421 hash_table
*new_kills
= this->kills
;
422 this->kills
= orig_kills
;
423 this->acp
= orig_acp
;
424 this->killed_all
= this->killed_all
|| orig_killed_all
;
427 hash_table_foreach(new_kills
, htk
) {
428 kill_entry
*k
= (kill_entry
*) htk
->data
;
429 kill(k
->var
, k
->write_mask
);
432 /* already descended into the children. */
433 return visit_continue_with_parent
;
437 ir_constant_propagation_visitor::kill(ir_variable
*var
, unsigned write_mask
)
441 /* We don't track non-vectors. */
442 if (!var
->type
->is_vector() && !var
->type
->is_scalar())
445 /* Remove any entries currently in the ACP for this kill. */
446 foreach_in_list_safe(acp_entry
, entry
, this->acp
) {
447 if (entry
->var
== var
) {
448 entry
->write_mask
&= ~write_mask
;
449 if (entry
->write_mask
== 0)
454 /* Add this writemask of the variable to the hash table of killed
455 * variables in this block.
457 hash_entry
*kill_hash_entry
= _mesa_hash_table_search(this->kills
, var
);
458 if (kill_hash_entry
) {
459 kill_entry
*entry
= (kill_entry
*) kill_hash_entry
->data
;
460 entry
->write_mask
|= write_mask
;
463 /* Not already in the hash table. Make new entry. */
464 _mesa_hash_table_insert(this->kills
, var
,
465 new(this->lin_ctx
) kill_entry(var
, write_mask
));
469 * Adds an entry to the available constant list if it's a plain assignment
470 * of a variable to a variable.
473 ir_constant_propagation_visitor::add_constant(ir_assignment
*ir
)
483 ir_dereference_variable
*deref
= ir
->lhs
->as_dereference_variable();
484 ir_constant
*constant
= ir
->rhs
->as_constant();
486 if (!deref
|| !constant
)
489 /* Only do constant propagation on vectors. Constant matrices,
490 * arrays, or structures would require more work elsewhere.
492 if (!deref
->var
->type
->is_vector() && !deref
->var
->type
->is_scalar())
495 /* We can't do copy propagation on buffer variables, since the underlying
496 * memory storage is shared across multiple threads we can't be sure that
497 * the variable value isn't modified between this assignment and the next
498 * instruction where its value is read.
500 if (deref
->var
->data
.mode
== ir_var_shader_storage
||
501 deref
->var
->data
.mode
== ir_var_shader_shared
)
504 entry
= new(this->lin_ctx
) acp_entry(deref
->var
, ir
->write_mask
, constant
);
505 this->acp
->push_tail(entry
);
508 } /* unnamed namespace */
511 * Does a constant propagation pass on the code present in the instruction stream.
514 do_constant_propagation(exec_list
*instructions
)
516 ir_constant_propagation_visitor v
;
518 visit_list_elements(&v
, instructions
);