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 update_type(ir_expression *ir)
90 if (ir->operands[0]->type->is_vector())
91 ir->type = ir->operands[0]->type;
93 ir->type = ir->operands[1]->type;
97 ir_algebraic_visitor::reassociate_operands(ir_expression *ir1,
102 ir_rvalue *temp = ir2->operands[op2];
103 ir2->operands[op2] = ir1->operands[op1];
104 ir1->operands[op1] = temp;
106 /* Update the type of ir2. The type of ir1 won't have changed --
107 * base types matched, and at least one of the operands of the 2
108 * binops is still a vector if any of them were.
112 this->progress = true;
116 * Reassociates a constant down a tree of adds or multiplies.
118 * Consider (2 * (a * (b * 0.5))). We want to send up with a * b.
121 ir_algebraic_visitor::reassociate_constant(ir_expression *ir1, int const_index,
122 ir_constant *constant,
125 if (!ir2 || ir1->operation != ir2->operation)
128 /* Don't want to even think about matrices. */
129 if (ir1->operands[0]->type->is_matrix() ||
130 ir1->operands[1]->type->is_matrix() ||
131 ir2->operands[0]->type->is_matrix() ||
132 ir2->operands[1]->type->is_matrix())
135 ir_constant *ir2_const[2];
136 ir2_const[0] = ir2->operands[0]->constant_expression_value();
137 ir2_const[1] = ir2->operands[1]->constant_expression_value();
139 if (ir2_const[0] && ir2_const[1])
143 reassociate_operands(ir1, const_index, ir2, 1);
145 } else if (ir2_const[1]) {
146 reassociate_operands(ir1, const_index, ir2, 0);
150 if (reassociate_constant(ir1, const_index, constant,
151 ir2->operands[0]->as_expression())) {
156 if (reassociate_constant(ir1, const_index, constant,
157 ir2->operands[1]->as_expression())) {
165 /* When eliminating an expression and just returning one of its operands,
166 * we may need to swizzle that operand out to a vector if the expression was
170 ir_algebraic_visitor::swizzle_if_required(ir_expression *expr,
173 if (expr->type->is_vector() && operand->type->is_scalar()) {
174 return new(mem_ctx) ir_swizzle(operand, 0, 0, 0, 0,
175 expr->type->vector_elements);
181 ir_algebraic_visitor::handle_expression(ir_expression *ir)
183 ir_constant *op_const[2] = {NULL, NULL};
184 ir_expression *op_expr[2] = {NULL, NULL};
188 assert(ir->get_num_operands() <= 2);
189 for (i = 0; i < ir->get_num_operands(); i++) {
190 if (ir->operands[i]->type->is_matrix())
193 op_const[i] = ir->operands[i]->constant_expression_value();
194 op_expr[i] = ir->operands[i]->as_expression();
197 if (this->mem_ctx == NULL)
198 this->mem_ctx = ralloc_parent(ir);
200 switch (ir->operation) {
201 case ir_unop_logic_not: {
202 enum ir_expression_operation new_op = ir_unop_logic_not;
204 if (op_expr[0] == NULL)
207 switch (op_expr[0]->operation) {
208 case ir_binop_less: new_op = ir_binop_gequal; break;
209 case ir_binop_greater: new_op = ir_binop_lequal; break;
210 case ir_binop_lequal: new_op = ir_binop_greater; break;
211 case ir_binop_gequal: new_op = ir_binop_less; break;
212 case ir_binop_equal: new_op = ir_binop_nequal; break;
213 case ir_binop_nequal: new_op = ir_binop_equal; break;
214 case ir_binop_all_equal: new_op = ir_binop_any_nequal; break;
215 case ir_binop_any_nequal: new_op = ir_binop_all_equal; break;
218 /* The default case handler is here to silence a warning from GCC.
223 if (new_op != ir_unop_logic_not) {
224 this->progress = true;
225 return new(mem_ctx) ir_expression(new_op,
227 op_expr[0]->operands[0],
228 op_expr[0]->operands[1]);
235 if (is_vec_zero(op_const[0])) {
236 this->progress = true;
237 return swizzle_if_required(ir, ir->operands[1]);
239 if (is_vec_zero(op_const[1])) {
240 this->progress = true;
241 return swizzle_if_required(ir, ir->operands[0]);
244 /* Reassociate addition of constants so that we can do constant
247 if (op_const[0] && !op_const[1])
248 reassociate_constant(ir, 0, op_const[0],
249 ir->operands[1]->as_expression());
250 if (op_const[1] && !op_const[0])
251 reassociate_constant(ir, 1, op_const[1],
252 ir->operands[0]->as_expression());
256 if (is_vec_zero(op_const[0])) {
257 this->progress = true;
258 temp = new(mem_ctx) ir_expression(ir_unop_neg,
259 ir->operands[1]->type,
262 return swizzle_if_required(ir, temp);
264 if (is_vec_zero(op_const[1])) {
265 this->progress = true;
266 return swizzle_if_required(ir, ir->operands[0]);
271 if (is_vec_one(op_const[0])) {
272 this->progress = true;
273 return swizzle_if_required(ir, ir->operands[1]);
275 if (is_vec_one(op_const[1])) {
276 this->progress = true;
277 return swizzle_if_required(ir, ir->operands[0]);
280 if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
281 this->progress = true;
282 return ir_constant::zero(ir, ir->type);
285 /* Reassociate multiplication of constants so that we can do
288 if (op_const[0] && !op_const[1])
289 reassociate_constant(ir, 0, op_const[0],
290 ir->operands[1]->as_expression());
291 if (op_const[1] && !op_const[0])
292 reassociate_constant(ir, 1, op_const[1],
293 ir->operands[0]->as_expression());
298 if (is_vec_one(op_const[0]) && ir->type->base_type == GLSL_TYPE_FLOAT) {
299 this->progress = true;
300 temp = new(mem_ctx) ir_expression(ir_unop_rcp,
301 ir->operands[1]->type,
304 return swizzle_if_required(ir, temp);
306 if (is_vec_one(op_const[1])) {
307 this->progress = true;
308 return swizzle_if_required(ir, ir->operands[0]);
312 case ir_binop_logic_and:
313 /* FINISHME: Also simplify (a && a) to (a). */
314 if (is_vec_one(op_const[0])) {
315 this->progress = true;
316 return ir->operands[1];
317 } else if (is_vec_one(op_const[1])) {
318 this->progress = true;
319 return ir->operands[0];
320 } else if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
321 this->progress = true;
322 return ir_constant::zero(mem_ctx, ir->type);
326 case ir_binop_logic_xor:
327 /* FINISHME: Also simplify (a ^^ a) to (false). */
328 if (is_vec_zero(op_const[0])) {
329 this->progress = true;
330 return ir->operands[1];
331 } else if (is_vec_zero(op_const[1])) {
332 this->progress = true;
333 return ir->operands[0];
334 } else if (is_vec_one(op_const[0])) {
335 this->progress = true;
336 return new(mem_ctx) ir_expression(ir_unop_logic_not, ir->type,
337 ir->operands[1], NULL);
338 } else if (is_vec_one(op_const[1])) {
339 this->progress = true;
340 return new(mem_ctx) ir_expression(ir_unop_logic_not, ir->type,
341 ir->operands[0], NULL);
345 case ir_binop_logic_or:
346 /* FINISHME: Also simplify (a || a) to (a). */
347 if (is_vec_zero(op_const[0])) {
348 this->progress = true;
349 return ir->operands[1];
350 } else if (is_vec_zero(op_const[1])) {
351 this->progress = true;
352 return ir->operands[0];
353 } else if (is_vec_one(op_const[0]) || is_vec_one(op_const[1])) {
354 ir_constant_data data;
356 for (unsigned i = 0; i < 16; i++)
359 this->progress = true;
360 return new(mem_ctx) ir_constant(ir->type, &data);
365 if (op_expr[0] && op_expr[0]->operation == ir_unop_rcp) {
366 this->progress = true;
367 return op_expr[0]->operands[0];
370 /* FINISHME: We should do rcp(rsq(x)) -> sqrt(x) for some
371 * backends, except that some backends will have done sqrt ->
372 * rcp(rsq(x)) and we don't want to undo it for them.
375 /* As far as we know, all backends are OK with rsq. */
376 if (op_expr[0] && op_expr[0]->operation == ir_unop_sqrt) {
377 this->progress = true;
378 temp = new(mem_ctx) ir_expression(ir_unop_rsq,
379 op_expr[0]->operands[0]->type,
380 op_expr[0]->operands[0],
382 return swizzle_if_required(ir, temp);
395 ir_algebraic_visitor::handle_rvalue(ir_rvalue **rvalue)
400 ir_expression *expr = (*rvalue)->as_expression();
401 if (!expr || expr->operation == ir_quadop_vector)
404 *rvalue = handle_expression(expr);
408 do_algebraic(exec_list *instructions)
410 ir_algebraic_visitor v;
412 visit_list_elements(&v, instructions);