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 "glsl_types.h"
46 class acp_entry
: public exec_node
49 acp_entry(ir_variable
*var
, unsigned write_mask
, ir_constant
*constant
)
54 this->write_mask
= write_mask
;
55 this->constant
= constant
;
56 this->initial_values
= write_mask
;
59 acp_entry(const acp_entry
*src
)
62 this->write_mask
= src
->write_mask
;
63 this->constant
= src
->constant
;
64 this->initial_values
= src
->initial_values
;
68 ir_constant
*constant
;
71 /** Mask of values initially available in the constant. */
72 unsigned initial_values
;
76 class kill_entry
: public exec_node
79 kill_entry(ir_variable
*var
, unsigned write_mask
)
83 this->write_mask
= write_mask
;
90 class ir_constant_propagation_visitor
: public ir_rvalue_visitor
{
92 ir_constant_propagation_visitor()
96 mem_ctx
= ralloc_context(0);
97 this->acp
= new(mem_ctx
) exec_list
;
98 this->kills
= new(mem_ctx
) exec_list
;
100 ~ir_constant_propagation_visitor()
102 ralloc_free(mem_ctx
);
105 virtual ir_visitor_status
visit_enter(class ir_loop
*);
106 virtual ir_visitor_status
visit_enter(class ir_function_signature
*);
107 virtual ir_visitor_status
visit_enter(class ir_function
*);
108 virtual ir_visitor_status
visit_leave(class ir_assignment
*);
109 virtual ir_visitor_status
visit_enter(class ir_call
*);
110 virtual ir_visitor_status
visit_enter(class ir_if
*);
112 void add_constant(ir_assignment
*ir
);
113 void constant_folding(ir_rvalue
**rvalue
);
114 void constant_propagation(ir_rvalue
**rvalue
);
115 void kill(ir_variable
*ir
, unsigned write_mask
);
116 void handle_if_block(exec_list
*instructions
);
117 void handle_rvalue(ir_rvalue
**rvalue
);
119 /** List of acp_entry: The available constants to propagate */
123 * List of kill_entry: The masks of variables whose values were
124 * killed in this block.
137 ir_constant_propagation_visitor::constant_folding(ir_rvalue
**rvalue
) {
139 if (*rvalue
== NULL
|| (*rvalue
)->ir_type
== ir_type_constant
)
142 /* Note that we visit rvalues one leaving. So if an expression has a
143 * non-constant operand, no need to go looking down it to find if it's
144 * constant. This cuts the time of this pass down drastically.
146 ir_expression
*expr
= (*rvalue
)->as_expression();
148 for (unsigned int i
= 0; i
< expr
->get_num_operands(); i
++) {
149 if (!expr
->operands
[i
]->as_constant())
154 /* Ditto for swizzles. */
155 ir_swizzle
*swiz
= (*rvalue
)->as_swizzle();
156 if (swiz
&& !swiz
->val
->as_constant())
159 ir_constant
*constant
= (*rvalue
)->constant_expression_value();
162 this->progress
= true;
167 ir_constant_propagation_visitor::constant_propagation(ir_rvalue
**rvalue
) {
169 if (this->in_assignee
|| !*rvalue
)
172 const glsl_type
*type
= (*rvalue
)->type
;
173 if (!type
->is_scalar() && !type
->is_vector())
176 ir_swizzle
*swiz
= NULL
;
177 ir_dereference_variable
*deref
= (*rvalue
)->as_dereference_variable();
179 swiz
= (*rvalue
)->as_swizzle();
183 deref
= swiz
->val
->as_dereference_variable();
188 ir_constant_data data
;
189 memset(&data
, 0, sizeof(data
));
191 for (unsigned int i
= 0; i
< type
->components(); i
++) {
193 acp_entry
*found
= NULL
;
197 case 0: channel
= swiz
->mask
.x
; break;
198 case 1: channel
= swiz
->mask
.y
; break;
199 case 2: channel
= swiz
->mask
.z
; break;
200 case 3: channel
= swiz
->mask
.w
; break;
201 default: assert(!"shouldn't be reached"); channel
= 0; break;
207 foreach_in_list(acp_entry
, entry
, this->acp
) {
208 if (entry
->var
== deref
->var
&& entry
->write_mask
& (1 << channel
)) {
218 for (int j
= 0; j
< 4; j
++) {
221 if (found
->initial_values
& (1 << j
))
225 switch (type
->base_type
) {
226 case GLSL_TYPE_FLOAT
:
227 data
.f
[i
] = found
->constant
->value
.f
[rhs_channel
];
229 case GLSL_TYPE_DOUBLE
:
230 data
.d
[i
] = found
->constant
->value
.d
[rhs_channel
];
233 data
.i
[i
] = found
->constant
->value
.i
[rhs_channel
];
236 data
.u
[i
] = found
->constant
->value
.u
[rhs_channel
];
239 data
.b
[i
] = found
->constant
->value
.b
[rhs_channel
];
242 assert(!"not reached");
247 *rvalue
= new(ralloc_parent(deref
)) ir_constant(type
, &data
);
248 this->progress
= true;
252 ir_constant_propagation_visitor::handle_rvalue(ir_rvalue
**rvalue
)
254 constant_propagation(rvalue
);
255 constant_folding(rvalue
);
259 ir_constant_propagation_visitor::visit_enter(ir_function_signature
*ir
)
261 /* Treat entry into a function signature as a completely separate
262 * block. Any instructions at global scope will be shuffled into
263 * main() at link time, so they're irrelevant to us.
265 exec_list
*orig_acp
= this->acp
;
266 exec_list
*orig_kills
= this->kills
;
267 bool orig_killed_all
= this->killed_all
;
269 this->acp
= new(mem_ctx
) exec_list
;
270 this->kills
= new(mem_ctx
) exec_list
;
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 exec_list
*orig_kills
= this->kills
;
356 bool orig_killed_all
= this->killed_all
;
358 this->acp
= new(mem_ctx
) exec_list
;
359 this->kills
= new(mem_ctx
) exec_list
;
360 this->killed_all
= false;
362 /* Populate the initial acp with a constant of the original */
363 foreach_in_list(acp_entry
, a
, orig_acp
) {
364 this->acp
->push_tail(new(this->mem_ctx
) acp_entry(a
));
367 visit_list_elements(this, instructions
);
369 if (this->killed_all
) {
370 orig_acp
->make_empty();
373 exec_list
*new_kills
= this->kills
;
374 this->kills
= orig_kills
;
375 this->acp
= orig_acp
;
376 this->killed_all
= this->killed_all
|| orig_killed_all
;
378 foreach_in_list(kill_entry
, k
, new_kills
) {
379 kill(k
->var
, k
->write_mask
);
384 ir_constant_propagation_visitor::visit_enter(ir_if
*ir
)
386 ir
->condition
->accept(this);
387 handle_rvalue(&ir
->condition
);
389 handle_if_block(&ir
->then_instructions
);
390 handle_if_block(&ir
->else_instructions
);
392 /* handle_if_block() already descended into the children. */
393 return visit_continue_with_parent
;
397 ir_constant_propagation_visitor::visit_enter(ir_loop
*ir
)
399 exec_list
*orig_acp
= this->acp
;
400 exec_list
*orig_kills
= this->kills
;
401 bool orig_killed_all
= this->killed_all
;
403 /* FINISHME: For now, the initial acp for loops is totally empty.
404 * We could go through once, then go through again with the acp
405 * cloned minus the killed entries after the first run through.
407 this->acp
= new(mem_ctx
) exec_list
;
408 this->kills
= new(mem_ctx
) exec_list
;
409 this->killed_all
= false;
411 visit_list_elements(this, &ir
->body_instructions
);
413 if (this->killed_all
) {
414 orig_acp
->make_empty();
417 exec_list
*new_kills
= this->kills
;
418 this->kills
= orig_kills
;
419 this->acp
= orig_acp
;
420 this->killed_all
= this->killed_all
|| orig_killed_all
;
422 foreach_in_list(kill_entry
, k
, new_kills
) {
423 kill(k
->var
, k
->write_mask
);
426 /* already descended into the children. */
427 return visit_continue_with_parent
;
431 ir_constant_propagation_visitor::kill(ir_variable
*var
, unsigned write_mask
)
435 /* We don't track non-vectors. */
436 if (!var
->type
->is_vector() && !var
->type
->is_scalar())
439 /* Remove any entries currently in the ACP for this kill. */
440 foreach_in_list_safe(acp_entry
, entry
, this->acp
) {
441 if (entry
->var
== var
) {
442 entry
->write_mask
&= ~write_mask
;
443 if (entry
->write_mask
== 0)
448 /* Add this writemask of the variable to the list of killed
449 * variables in this block.
451 foreach_in_list(kill_entry
, entry
, this->kills
) {
452 if (entry
->var
== var
) {
453 entry
->write_mask
|= write_mask
;
457 /* Not already in the list. Make new entry. */
458 this->kills
->push_tail(new(this->mem_ctx
) kill_entry(var
, write_mask
));
462 * Adds an entry to the available constant list if it's a plain assignment
463 * of a variable to a variable.
466 ir_constant_propagation_visitor::add_constant(ir_assignment
*ir
)
476 ir_dereference_variable
*deref
= ir
->lhs
->as_dereference_variable();
477 ir_constant
*constant
= ir
->rhs
->as_constant();
479 if (!deref
|| !constant
)
482 /* Only do constant propagation on vectors. Constant matrices,
483 * arrays, or structures would require more work elsewhere.
485 if (!deref
->var
->type
->is_vector() && !deref
->var
->type
->is_scalar())
488 /* We can't do copy propagation on buffer variables, since the underlying
489 * memory storage is shared across multiple threads we can't be sure that
490 * the variable value isn't modified between this assignment and the next
491 * instruction where its value is read.
493 if (deref
->var
->data
.mode
== ir_var_shader_storage
)
496 entry
= new(this->mem_ctx
) acp_entry(deref
->var
, ir
->write_mask
, constant
);
497 this->acp
->push_tail(entry
);
500 } /* unnamed namespace */
503 * Does a constant propagation pass on the code present in the instruction stream.
506 do_constant_propagation(exec_list
*instructions
)
508 ir_constant_propagation_visitor v
;
510 visit_list_elements(&v
, instructions
);