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_algebraic.cpp
27 * Takes advantage of association, commutivity, and other algebraic
28 * properties to simplify expressions.
32 #include "ir_visitor.h"
33 #include "ir_rvalue_visitor.h"
34 #include "ir_optimization.h"
35 #include "glsl_types.h"
40 * Visitor class for replacing expressions with ir_constant values.
43 class ir_algebraic_visitor
: public ir_rvalue_visitor
{
45 ir_algebraic_visitor()
47 this->progress
= false;
51 virtual ~ir_algebraic_visitor()
55 ir_rvalue
*handle_expression(ir_expression
*ir
);
56 void handle_rvalue(ir_rvalue
**rvalue
);
57 bool reassociate_constant(ir_expression
*ir1
,
59 ir_constant
*constant
,
61 void reassociate_operands(ir_expression
*ir1
,
65 ir_rvalue
*swizzle_if_required(ir_expression
*expr
,
73 } /* unnamed namespace */
76 is_vec_zero(ir_constant
*ir
)
78 return (ir
== NULL
) ? false : ir
->is_zero();
82 is_vec_one(ir_constant
*ir
)
84 return (ir
== NULL
) ? false : ir
->is_one();
88 is_vec_basis(ir_constant
*ir
)
90 return (ir
== NULL
) ? false : ir
->is_basis();
94 update_type(ir_expression
*ir
)
96 if (ir
->operands
[0]->type
->is_vector())
97 ir
->type
= ir
->operands
[0]->type
;
99 ir
->type
= ir
->operands
[1]->type
;
103 ir_algebraic_visitor::reassociate_operands(ir_expression
*ir1
,
108 ir_rvalue
*temp
= ir2
->operands
[op2
];
109 ir2
->operands
[op2
] = ir1
->operands
[op1
];
110 ir1
->operands
[op1
] = temp
;
112 /* Update the type of ir2. The type of ir1 won't have changed --
113 * base types matched, and at least one of the operands of the 2
114 * binops is still a vector if any of them were.
118 this->progress
= true;
122 * Reassociates a constant down a tree of adds or multiplies.
124 * Consider (2 * (a * (b * 0.5))). We want to send up with a * b.
127 ir_algebraic_visitor::reassociate_constant(ir_expression
*ir1
, int const_index
,
128 ir_constant
*constant
,
131 if (!ir2
|| ir1
->operation
!= ir2
->operation
)
134 /* Don't want to even think about matrices. */
135 if (ir1
->operands
[0]->type
->is_matrix() ||
136 ir1
->operands
[1]->type
->is_matrix() ||
137 ir2
->operands
[0]->type
->is_matrix() ||
138 ir2
->operands
[1]->type
->is_matrix())
141 ir_constant
*ir2_const
[2];
142 ir2_const
[0] = ir2
->operands
[0]->constant_expression_value();
143 ir2_const
[1] = ir2
->operands
[1]->constant_expression_value();
145 if (ir2_const
[0] && ir2_const
[1])
149 reassociate_operands(ir1
, const_index
, ir2
, 1);
151 } else if (ir2_const
[1]) {
152 reassociate_operands(ir1
, const_index
, ir2
, 0);
156 if (reassociate_constant(ir1
, const_index
, constant
,
157 ir2
->operands
[0]->as_expression())) {
162 if (reassociate_constant(ir1
, const_index
, constant
,
163 ir2
->operands
[1]->as_expression())) {
171 /* When eliminating an expression and just returning one of its operands,
172 * we may need to swizzle that operand out to a vector if the expression was
176 ir_algebraic_visitor::swizzle_if_required(ir_expression
*expr
,
179 if (expr
->type
->is_vector() && operand
->type
->is_scalar()) {
180 return new(mem_ctx
) ir_swizzle(operand
, 0, 0, 0, 0,
181 expr
->type
->vector_elements
);
187 ir_algebraic_visitor::handle_expression(ir_expression
*ir
)
189 ir_constant
*op_const
[2] = {NULL
, NULL
};
190 ir_expression
*op_expr
[2] = {NULL
, NULL
};
194 assert(ir
->get_num_operands() <= 2);
195 for (i
= 0; i
< ir
->get_num_operands(); i
++) {
196 if (ir
->operands
[i
]->type
->is_matrix())
199 op_const
[i
] = ir
->operands
[i
]->constant_expression_value();
200 op_expr
[i
] = ir
->operands
[i
]->as_expression();
203 if (this->mem_ctx
== NULL
)
204 this->mem_ctx
= ralloc_parent(ir
);
206 switch (ir
->operation
) {
207 case ir_unop_logic_not
: {
208 enum ir_expression_operation new_op
= ir_unop_logic_not
;
210 if (op_expr
[0] == NULL
)
213 switch (op_expr
[0]->operation
) {
214 case ir_binop_less
: new_op
= ir_binop_gequal
; break;
215 case ir_binop_greater
: new_op
= ir_binop_lequal
; break;
216 case ir_binop_lequal
: new_op
= ir_binop_greater
; break;
217 case ir_binop_gequal
: new_op
= ir_binop_less
; break;
218 case ir_binop_equal
: new_op
= ir_binop_nequal
; break;
219 case ir_binop_nequal
: new_op
= ir_binop_equal
; break;
220 case ir_binop_all_equal
: new_op
= ir_binop_any_nequal
; break;
221 case ir_binop_any_nequal
: new_op
= ir_binop_all_equal
; break;
224 /* The default case handler is here to silence a warning from GCC.
229 if (new_op
!= ir_unop_logic_not
) {
230 this->progress
= true;
231 return new(mem_ctx
) ir_expression(new_op
,
233 op_expr
[0]->operands
[0],
234 op_expr
[0]->operands
[1]);
241 if (is_vec_zero(op_const
[0])) {
242 this->progress
= true;
243 return swizzle_if_required(ir
, ir
->operands
[1]);
245 if (is_vec_zero(op_const
[1])) {
246 this->progress
= true;
247 return swizzle_if_required(ir
, ir
->operands
[0]);
250 /* Reassociate addition of constants so that we can do constant
253 if (op_const
[0] && !op_const
[1])
254 reassociate_constant(ir
, 0, op_const
[0],
255 ir
->operands
[1]->as_expression());
256 if (op_const
[1] && !op_const
[0])
257 reassociate_constant(ir
, 1, op_const
[1],
258 ir
->operands
[0]->as_expression());
262 if (is_vec_zero(op_const
[0])) {
263 this->progress
= true;
264 temp
= new(mem_ctx
) ir_expression(ir_unop_neg
,
265 ir
->operands
[1]->type
,
268 return swizzle_if_required(ir
, temp
);
270 if (is_vec_zero(op_const
[1])) {
271 this->progress
= true;
272 return swizzle_if_required(ir
, ir
->operands
[0]);
277 if (is_vec_one(op_const
[0])) {
278 this->progress
= true;
279 return swizzle_if_required(ir
, ir
->operands
[1]);
281 if (is_vec_one(op_const
[1])) {
282 this->progress
= true;
283 return swizzle_if_required(ir
, ir
->operands
[0]);
286 if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1])) {
287 this->progress
= true;
288 return ir_constant::zero(ir
, ir
->type
);
291 /* Reassociate multiplication of constants so that we can do
294 if (op_const
[0] && !op_const
[1])
295 reassociate_constant(ir
, 0, op_const
[0],
296 ir
->operands
[1]->as_expression());
297 if (op_const
[1] && !op_const
[0])
298 reassociate_constant(ir
, 1, op_const
[1],
299 ir
->operands
[0]->as_expression());
304 if (is_vec_one(op_const
[0]) && ir
->type
->base_type
== GLSL_TYPE_FLOAT
) {
305 this->progress
= true;
306 temp
= new(mem_ctx
) ir_expression(ir_unop_rcp
,
307 ir
->operands
[1]->type
,
310 return swizzle_if_required(ir
, temp
);
312 if (is_vec_one(op_const
[1])) {
313 this->progress
= true;
314 return swizzle_if_required(ir
, ir
->operands
[0]);
319 if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1])) {
320 this->progress
= true;
321 return ir_constant::zero(mem_ctx
, ir
->type
);
323 if (is_vec_basis(op_const
[0])) {
324 this->progress
= true;
325 unsigned component
= 0;
326 for (unsigned c
= 0; c
< op_const
[0]->type
->vector_elements
; c
++) {
327 if (op_const
[0]->value
.f
[c
] == 1.0)
330 return new(mem_ctx
) ir_swizzle(ir
->operands
[1], component
, 0, 0, 0, 1);
332 if (is_vec_basis(op_const
[1])) {
333 this->progress
= true;
334 unsigned component
= 0;
335 for (unsigned c
= 0; c
< op_const
[1]->type
->vector_elements
; c
++) {
336 if (op_const
[1]->value
.f
[c
] == 1.0)
339 return new(mem_ctx
) ir_swizzle(ir
->operands
[0], component
, 0, 0, 0, 1);
343 case ir_binop_logic_and
:
344 /* FINISHME: Also simplify (a && a) to (a). */
345 if (is_vec_one(op_const
[0])) {
346 this->progress
= true;
347 return ir
->operands
[1];
348 } else if (is_vec_one(op_const
[1])) {
349 this->progress
= true;
350 return ir
->operands
[0];
351 } else if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1])) {
352 this->progress
= true;
353 return ir_constant::zero(mem_ctx
, ir
->type
);
357 case ir_binop_logic_xor
:
358 /* FINISHME: Also simplify (a ^^ a) to (false). */
359 if (is_vec_zero(op_const
[0])) {
360 this->progress
= true;
361 return ir
->operands
[1];
362 } else if (is_vec_zero(op_const
[1])) {
363 this->progress
= true;
364 return ir
->operands
[0];
365 } else if (is_vec_one(op_const
[0])) {
366 this->progress
= true;
367 return new(mem_ctx
) ir_expression(ir_unop_logic_not
, ir
->type
,
368 ir
->operands
[1], NULL
);
369 } else if (is_vec_one(op_const
[1])) {
370 this->progress
= true;
371 return new(mem_ctx
) ir_expression(ir_unop_logic_not
, ir
->type
,
372 ir
->operands
[0], NULL
);
376 case ir_binop_logic_or
:
377 /* FINISHME: Also simplify (a || a) to (a). */
378 if (is_vec_zero(op_const
[0])) {
379 this->progress
= true;
380 return ir
->operands
[1];
381 } else if (is_vec_zero(op_const
[1])) {
382 this->progress
= true;
383 return ir
->operands
[0];
384 } else if (is_vec_one(op_const
[0]) || is_vec_one(op_const
[1])) {
385 ir_constant_data data
;
387 for (unsigned i
= 0; i
< 16; i
++)
390 this->progress
= true;
391 return new(mem_ctx
) ir_constant(ir
->type
, &data
);
396 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_rcp
) {
397 this->progress
= true;
398 return op_expr
[0]->operands
[0];
401 /* FINISHME: We should do rcp(rsq(x)) -> sqrt(x) for some
402 * backends, except that some backends will have done sqrt ->
403 * rcp(rsq(x)) and we don't want to undo it for them.
406 /* As far as we know, all backends are OK with rsq. */
407 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_sqrt
) {
408 this->progress
= true;
409 temp
= new(mem_ctx
) ir_expression(ir_unop_rsq
,
410 op_expr
[0]->operands
[0]->type
,
411 op_expr
[0]->operands
[0],
413 return swizzle_if_required(ir
, temp
);
426 ir_algebraic_visitor::handle_rvalue(ir_rvalue
**rvalue
)
431 ir_expression
*expr
= (*rvalue
)->as_expression();
432 if (!expr
|| expr
->operation
== ir_quadop_vector
)
435 *rvalue
= handle_expression(expr
);
439 do_algebraic(exec_list
*instructions
)
441 ir_algebraic_visitor v
;
443 visit_list_elements(&v
, instructions
);