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 ir_constant_expression.cpp
26 * Evaluate and process constant valued expressions
28 * In GLSL, constant valued expressions are used in several places. These
29 * must be processed and evaluated very early in the compilation process.
32 * * Initializers for uniforms
33 * * Initializers for \c const variables
37 #include "main/core.h" /* for MAX2, MIN2, CLAMP */
39 #include "ir_visitor.h"
40 #include "glsl_types.h"
41 #include "program/hash_table.h"
43 /* Using C99 rounding functions for roundToEven() implementation is
44 * difficult, because round(), rint, and nearbyint() are affected by
45 * fesetenv(), which the application may have done for its own
46 * purposes. Mesa's IROUND macro is close to what we want, but it
47 * rounds away from 0 on n + 0.5.
50 round_to_even(float val)
52 int rounded = IROUND(val);
54 if (val - floor(val) == 0.5) {
56 rounded += val > 0 ? -1 : 1;
63 dot(ir_constant *op0, ir_constant *op1)
65 assert(op0->type->is_float() && op1->type->is_float());
68 for (unsigned c = 0; c < op0->type->components(); c++)
69 result += op0->value.f[c] * op1->value.f[c];
74 /* This method is the only one supported by gcc. Unions in particular
75 * are iffy, and read-through-converted-pointer is killed by strict
76 * aliasing. OTOH, the compiler sees through the memcpy, so the
77 * resulting asm is reasonable.
80 bitcast_u2f(unsigned int u)
82 assert(sizeof(float) == sizeof(unsigned int));
84 memcpy(&f, &u, sizeof(f));
91 assert(sizeof(float) == sizeof(unsigned int));
93 memcpy(&u, &f, sizeof(f));
98 ir_rvalue::constant_expression_value(struct hash_table *variable_context)
100 assert(this->type->is_error());
105 ir_expression::constant_expression_value(struct hash_table *variable_context)
107 if (this->type->is_error())
110 ir_constant *op[Elements(this->operands)] = { NULL, };
111 ir_constant_data data;
113 memset(&data, 0, sizeof(data));
115 for (unsigned operand = 0; operand < this->get_num_operands(); operand++) {
116 op[operand] = this->operands[operand]->constant_expression_value(variable_context);
122 assert(op[0]->type->base_type == op[1]->type->base_type ||
123 this->operation == ir_binop_lshift ||
124 this->operation == ir_binop_rshift);
126 bool op0_scalar = op[0]->type->is_scalar();
127 bool op1_scalar = op[1] != NULL && op[1]->type->is_scalar();
129 /* When iterating over a vector or matrix's components, we want to increase
130 * the loop counter. However, for scalars, we want to stay at 0.
132 unsigned c0_inc = op0_scalar ? 0 : 1;
133 unsigned c1_inc = op1_scalar ? 0 : 1;
135 if (op1_scalar || !op[1]) {
136 components = op[0]->type->components();
138 components = op[1]->type->components();
141 void *ctx = ralloc_parent(this);
143 /* Handle array operations here, rather than below. */
144 if (op[0]->type->is_array()) {
145 assert(op[1] != NULL && op[1]->type->is_array());
146 switch (this->operation) {
147 case ir_binop_all_equal:
148 return new(ctx) ir_constant(op[0]->has_value(op[1]));
149 case ir_binop_any_nequal:
150 return new(ctx) ir_constant(!op[0]->has_value(op[1]));
157 switch (this->operation) {
158 case ir_unop_bit_not:
159 switch (op[0]->type->base_type) {
161 for (unsigned c = 0; c < components; c++)
162 data.i[c] = ~ op[0]->value.i[c];
165 for (unsigned c = 0; c < components; c++)
166 data.u[c] = ~ op[0]->value.u[c];
173 case ir_unop_logic_not:
174 assert(op[0]->type->base_type == GLSL_TYPE_BOOL);
175 for (unsigned c = 0; c < op[0]->type->components(); c++)
176 data.b[c] = !op[0]->value.b[c];
180 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
181 for (unsigned c = 0; c < op[0]->type->components(); c++) {
182 data.i[c] = (int) op[0]->value.f[c];
186 assert(op[0]->type->base_type == GLSL_TYPE_INT);
187 for (unsigned c = 0; c < op[0]->type->components(); c++) {
188 data.f[c] = (float) op[0]->value.i[c];
192 assert(op[0]->type->base_type == GLSL_TYPE_UINT);
193 for (unsigned c = 0; c < op[0]->type->components(); c++) {
194 data.f[c] = (float) op[0]->value.u[c];
198 assert(op[0]->type->base_type == GLSL_TYPE_BOOL);
199 for (unsigned c = 0; c < op[0]->type->components(); c++) {
200 data.f[c] = op[0]->value.b[c] ? 1.0F : 0.0F;
204 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
205 for (unsigned c = 0; c < op[0]->type->components(); c++) {
206 data.b[c] = op[0]->value.f[c] != 0.0F ? true : false;
210 assert(op[0]->type->base_type == GLSL_TYPE_BOOL);
211 for (unsigned c = 0; c < op[0]->type->components(); c++) {
212 data.u[c] = op[0]->value.b[c] ? 1 : 0;
216 assert(op[0]->type->is_integer());
217 for (unsigned c = 0; c < op[0]->type->components(); c++) {
218 data.b[c] = op[0]->value.u[c] ? true : false;
222 assert(op[0]->type->base_type == GLSL_TYPE_UINT);
223 for (unsigned c = 0; c < op[0]->type->components(); c++) {
224 data.i[c] = op[0]->value.u[c];
228 assert(op[0]->type->base_type == GLSL_TYPE_INT);
229 for (unsigned c = 0; c < op[0]->type->components(); c++) {
230 data.u[c] = op[0]->value.i[c];
233 case ir_unop_bitcast_i2f:
234 assert(op[0]->type->base_type == GLSL_TYPE_INT);
235 for (unsigned c = 0; c < op[0]->type->components(); c++) {
236 data.f[c] = bitcast_u2f(op[0]->value.i[c]);
239 case ir_unop_bitcast_f2i:
240 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
241 for (unsigned c = 0; c < op[0]->type->components(); c++) {
242 data.i[c] = bitcast_f2u(op[0]->value.f[c]);
245 case ir_unop_bitcast_u2f:
246 assert(op[0]->type->base_type == GLSL_TYPE_UINT);
247 for (unsigned c = 0; c < op[0]->type->components(); c++) {
248 data.f[c] = bitcast_u2f(op[0]->value.u[c]);
251 case ir_unop_bitcast_f2u:
252 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
253 for (unsigned c = 0; c < op[0]->type->components(); c++) {
254 data.u[c] = bitcast_f2u(op[0]->value.f[c]);
258 assert(op[0]->type->is_boolean());
260 for (unsigned c = 0; c < op[0]->type->components(); c++) {
261 if (op[0]->value.b[c])
267 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
268 for (unsigned c = 0; c < op[0]->type->components(); c++) {
269 data.f[c] = truncf(op[0]->value.f[c]);
273 case ir_unop_round_even:
274 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
275 for (unsigned c = 0; c < op[0]->type->components(); c++) {
276 data.f[c] = round_to_even(op[0]->value.f[c]);
281 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
282 for (unsigned c = 0; c < op[0]->type->components(); c++) {
283 data.f[c] = ceilf(op[0]->value.f[c]);
288 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
289 for (unsigned c = 0; c < op[0]->type->components(); c++) {
290 data.f[c] = floorf(op[0]->value.f[c]);
295 for (unsigned c = 0; c < op[0]->type->components(); c++) {
296 switch (this->type->base_type) {
303 case GLSL_TYPE_FLOAT:
304 data.f[c] = op[0]->value.f[c] - floor(op[0]->value.f[c]);
313 case ir_unop_sin_reduced:
314 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
315 for (unsigned c = 0; c < op[0]->type->components(); c++) {
316 data.f[c] = sinf(op[0]->value.f[c]);
321 case ir_unop_cos_reduced:
322 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
323 for (unsigned c = 0; c < op[0]->type->components(); c++) {
324 data.f[c] = cosf(op[0]->value.f[c]);
329 for (unsigned c = 0; c < op[0]->type->components(); c++) {
330 switch (this->type->base_type) {
332 data.u[c] = -((int) op[0]->value.u[c]);
335 data.i[c] = -op[0]->value.i[c];
337 case GLSL_TYPE_FLOAT:
338 data.f[c] = -op[0]->value.f[c];
347 for (unsigned c = 0; c < op[0]->type->components(); c++) {
348 switch (this->type->base_type) {
350 data.u[c] = op[0]->value.u[c];
353 data.i[c] = op[0]->value.i[c];
355 data.i[c] = -data.i[c];
357 case GLSL_TYPE_FLOAT:
358 data.f[c] = fabs(op[0]->value.f[c]);
367 for (unsigned c = 0; c < op[0]->type->components(); c++) {
368 switch (this->type->base_type) {
370 data.u[c] = op[0]->value.i[c] > 0;
373 data.i[c] = (op[0]->value.i[c] > 0) - (op[0]->value.i[c] < 0);
375 case GLSL_TYPE_FLOAT:
376 data.f[c] = float((op[0]->value.f[c] > 0)-(op[0]->value.f[c] < 0));
385 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
386 for (unsigned c = 0; c < op[0]->type->components(); c++) {
387 switch (this->type->base_type) {
389 if (op[0]->value.u[c] != 0.0)
390 data.u[c] = 1 / op[0]->value.u[c];
393 if (op[0]->value.i[c] != 0.0)
394 data.i[c] = 1 / op[0]->value.i[c];
396 case GLSL_TYPE_FLOAT:
397 if (op[0]->value.f[c] != 0.0)
398 data.f[c] = 1.0F / op[0]->value.f[c];
407 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
408 for (unsigned c = 0; c < op[0]->type->components(); c++) {
409 data.f[c] = 1.0F / sqrtf(op[0]->value.f[c]);
414 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
415 for (unsigned c = 0; c < op[0]->type->components(); c++) {
416 data.f[c] = sqrtf(op[0]->value.f[c]);
421 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
422 for (unsigned c = 0; c < op[0]->type->components(); c++) {
423 data.f[c] = expf(op[0]->value.f[c]);
428 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
429 for (unsigned c = 0; c < op[0]->type->components(); c++) {
430 data.f[c] = exp2f(op[0]->value.f[c]);
435 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
436 for (unsigned c = 0; c < op[0]->type->components(); c++) {
437 data.f[c] = logf(op[0]->value.f[c]);
442 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
443 for (unsigned c = 0; c < op[0]->type->components(); c++) {
444 data.f[c] = log2f(op[0]->value.f[c]);
450 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
451 for (unsigned c = 0; c < op[0]->type->components(); c++) {
457 assert(op[0]->type->base_type == GLSL_TYPE_FLOAT);
458 for (unsigned c = 0; c < op[0]->type->components(); c++) {
459 data.f[c] = powf(op[0]->value.f[c], op[1]->value.f[c]);
464 data.f[0] = dot(op[0], op[1]);
468 assert(op[0]->type == op[1]->type || op0_scalar || op1_scalar);
469 for (unsigned c = 0, c0 = 0, c1 = 0;
471 c0 += c0_inc, c1 += c1_inc, c++) {
473 switch (op[0]->type->base_type) {
475 data.u[c] = MIN2(op[0]->value.u[c0], op[1]->value.u[c1]);
478 data.i[c] = MIN2(op[0]->value.i[c0], op[1]->value.i[c1]);
480 case GLSL_TYPE_FLOAT:
481 data.f[c] = MIN2(op[0]->value.f[c0], op[1]->value.f[c1]);
490 assert(op[0]->type == op[1]->type || op0_scalar || op1_scalar);
491 for (unsigned c = 0, c0 = 0, c1 = 0;
493 c0 += c0_inc, c1 += c1_inc, c++) {
495 switch (op[0]->type->base_type) {
497 data.u[c] = MAX2(op[0]->value.u[c0], op[1]->value.u[c1]);
500 data.i[c] = MAX2(op[0]->value.i[c0], op[1]->value.i[c1]);
502 case GLSL_TYPE_FLOAT:
503 data.f[c] = MAX2(op[0]->value.f[c0], op[1]->value.f[c1]);
512 assert(op[0]->type == op[1]->type || op0_scalar || op1_scalar);
513 for (unsigned c = 0, c0 = 0, c1 = 0;
515 c0 += c0_inc, c1 += c1_inc, c++) {
517 switch (op[0]->type->base_type) {
519 data.u[c] = op[0]->value.u[c0] + op[1]->value.u[c1];
522 data.i[c] = op[0]->value.i[c0] + op[1]->value.i[c1];
524 case GLSL_TYPE_FLOAT:
525 data.f[c] = op[0]->value.f[c0] + op[1]->value.f[c1];
534 assert(op[0]->type == op[1]->type || op0_scalar || op1_scalar);
535 for (unsigned c = 0, c0 = 0, c1 = 0;
537 c0 += c0_inc, c1 += c1_inc, c++) {
539 switch (op[0]->type->base_type) {
541 data.u[c] = op[0]->value.u[c0] - op[1]->value.u[c1];
544 data.i[c] = op[0]->value.i[c0] - op[1]->value.i[c1];
546 case GLSL_TYPE_FLOAT:
547 data.f[c] = op[0]->value.f[c0] - op[1]->value.f[c1];
556 /* Check for equal types, or unequal types involving scalars */
557 if ((op[0]->type == op[1]->type && !op[0]->type->is_matrix())
558 || op0_scalar || op1_scalar) {
559 for (unsigned c = 0, c0 = 0, c1 = 0;
561 c0 += c0_inc, c1 += c1_inc, c++) {
563 switch (op[0]->type->base_type) {
565 data.u[c] = op[0]->value.u[c0] * op[1]->value.u[c1];
568 data.i[c] = op[0]->value.i[c0] * op[1]->value.i[c1];
570 case GLSL_TYPE_FLOAT:
571 data.f[c] = op[0]->value.f[c0] * op[1]->value.f[c1];
578 assert(op[0]->type->is_matrix() || op[1]->type->is_matrix());
580 /* Multiply an N-by-M matrix with an M-by-P matrix. Since either
581 * matrix can be a GLSL vector, either N or P can be 1.
583 * For vec*mat, the vector is treated as a row vector. This
584 * means the vector is a 1-row x M-column matrix.
586 * For mat*vec, the vector is treated as a column vector. Since
587 * matrix_columns is 1 for vectors, this just works.
589 const unsigned n = op[0]->type->is_vector()
590 ? 1 : op[0]->type->vector_elements;
591 const unsigned m = op[1]->type->vector_elements;
592 const unsigned p = op[1]->type->matrix_columns;
593 for (unsigned j = 0; j < p; j++) {
594 for (unsigned i = 0; i < n; i++) {
595 for (unsigned k = 0; k < m; k++) {
596 data.f[i+n*j] += op[0]->value.f[i+n*k]*op[1]->value.f[k+m*j];
604 /* FINISHME: Emit warning when division-by-zero is detected. */
605 assert(op[0]->type == op[1]->type || op0_scalar || op1_scalar);
606 for (unsigned c = 0, c0 = 0, c1 = 0;
608 c0 += c0_inc, c1 += c1_inc, c++) {
610 switch (op[0]->type->base_type) {
612 if (op[1]->value.u[c1] == 0) {
615 data.u[c] = op[0]->value.u[c0] / op[1]->value.u[c1];
619 if (op[1]->value.i[c1] == 0) {
622 data.i[c] = op[0]->value.i[c0] / op[1]->value.i[c1];
625 case GLSL_TYPE_FLOAT:
626 data.f[c] = op[0]->value.f[c0] / op[1]->value.f[c1];
635 /* FINISHME: Emit warning when division-by-zero is detected. */
636 assert(op[0]->type == op[1]->type || op0_scalar || op1_scalar);
637 for (unsigned c = 0, c0 = 0, c1 = 0;
639 c0 += c0_inc, c1 += c1_inc, c++) {
641 switch (op[0]->type->base_type) {
643 if (op[1]->value.u[c1] == 0) {
646 data.u[c] = op[0]->value.u[c0] % op[1]->value.u[c1];
650 if (op[1]->value.i[c1] == 0) {
653 data.i[c] = op[0]->value.i[c0] % op[1]->value.i[c1];
656 case GLSL_TYPE_FLOAT:
657 /* We don't use fmod because it rounds toward zero; GLSL specifies
660 data.f[c] = op[0]->value.f[c0] - op[1]->value.f[c1]
661 * floorf(op[0]->value.f[c0] / op[1]->value.f[c1]);
670 case ir_binop_logic_and:
671 assert(op[0]->type->base_type == GLSL_TYPE_BOOL);
672 for (unsigned c = 0; c < op[0]->type->components(); c++)
673 data.b[c] = op[0]->value.b[c] && op[1]->value.b[c];
675 case ir_binop_logic_xor:
676 assert(op[0]->type->base_type == GLSL_TYPE_BOOL);
677 for (unsigned c = 0; c < op[0]->type->components(); c++)
678 data.b[c] = op[0]->value.b[c] ^ op[1]->value.b[c];
680 case ir_binop_logic_or:
681 assert(op[0]->type->base_type == GLSL_TYPE_BOOL);
682 for (unsigned c = 0; c < op[0]->type->components(); c++)
683 data.b[c] = op[0]->value.b[c] || op[1]->value.b[c];
687 assert(op[0]->type == op[1]->type);
688 for (unsigned c = 0; c < op[0]->type->components(); c++) {
689 switch (op[0]->type->base_type) {
691 data.b[c] = op[0]->value.u[c] < op[1]->value.u[c];
694 data.b[c] = op[0]->value.i[c] < op[1]->value.i[c];
696 case GLSL_TYPE_FLOAT:
697 data.b[c] = op[0]->value.f[c] < op[1]->value.f[c];
704 case ir_binop_greater:
705 assert(op[0]->type == op[1]->type);
706 for (unsigned c = 0; c < op[0]->type->components(); c++) {
707 switch (op[0]->type->base_type) {
709 data.b[c] = op[0]->value.u[c] > op[1]->value.u[c];
712 data.b[c] = op[0]->value.i[c] > op[1]->value.i[c];
714 case GLSL_TYPE_FLOAT:
715 data.b[c] = op[0]->value.f[c] > op[1]->value.f[c];
722 case ir_binop_lequal:
723 assert(op[0]->type == op[1]->type);
724 for (unsigned c = 0; c < op[0]->type->components(); c++) {
725 switch (op[0]->type->base_type) {
727 data.b[c] = op[0]->value.u[c] <= op[1]->value.u[c];
730 data.b[c] = op[0]->value.i[c] <= op[1]->value.i[c];
732 case GLSL_TYPE_FLOAT:
733 data.b[c] = op[0]->value.f[c] <= op[1]->value.f[c];
740 case ir_binop_gequal:
741 assert(op[0]->type == op[1]->type);
742 for (unsigned c = 0; c < op[0]->type->components(); c++) {
743 switch (op[0]->type->base_type) {
745 data.b[c] = op[0]->value.u[c] >= op[1]->value.u[c];
748 data.b[c] = op[0]->value.i[c] >= op[1]->value.i[c];
750 case GLSL_TYPE_FLOAT:
751 data.b[c] = op[0]->value.f[c] >= op[1]->value.f[c];
759 assert(op[0]->type == op[1]->type);
760 for (unsigned c = 0; c < components; c++) {
761 switch (op[0]->type->base_type) {
763 data.b[c] = op[0]->value.u[c] == op[1]->value.u[c];
766 data.b[c] = op[0]->value.i[c] == op[1]->value.i[c];
768 case GLSL_TYPE_FLOAT:
769 data.b[c] = op[0]->value.f[c] == op[1]->value.f[c];
772 data.b[c] = op[0]->value.b[c] == op[1]->value.b[c];
779 case ir_binop_nequal:
780 assert(op[0]->type == op[1]->type);
781 for (unsigned c = 0; c < components; c++) {
782 switch (op[0]->type->base_type) {
784 data.b[c] = op[0]->value.u[c] != op[1]->value.u[c];
787 data.b[c] = op[0]->value.i[c] != op[1]->value.i[c];
789 case GLSL_TYPE_FLOAT:
790 data.b[c] = op[0]->value.f[c] != op[1]->value.f[c];
793 data.b[c] = op[0]->value.b[c] != op[1]->value.b[c];
800 case ir_binop_all_equal:
801 data.b[0] = op[0]->has_value(op[1]);
803 case ir_binop_any_nequal:
804 data.b[0] = !op[0]->has_value(op[1]);
807 case ir_binop_lshift:
808 for (unsigned c = 0, c0 = 0, c1 = 0;
810 c0 += c0_inc, c1 += c1_inc, c++) {
812 if (op[0]->type->base_type == GLSL_TYPE_INT &&
813 op[1]->type->base_type == GLSL_TYPE_INT) {
814 data.i[c] = op[0]->value.i[c0] << op[1]->value.i[c1];
816 } else if (op[0]->type->base_type == GLSL_TYPE_INT &&
817 op[1]->type->base_type == GLSL_TYPE_UINT) {
818 data.i[c] = op[0]->value.i[c0] << op[1]->value.u[c1];
820 } else if (op[0]->type->base_type == GLSL_TYPE_UINT &&
821 op[1]->type->base_type == GLSL_TYPE_INT) {
822 data.u[c] = op[0]->value.u[c0] << op[1]->value.i[c1];
824 } else if (op[0]->type->base_type == GLSL_TYPE_UINT &&
825 op[1]->type->base_type == GLSL_TYPE_UINT) {
826 data.u[c] = op[0]->value.u[c0] << op[1]->value.u[c1];
831 case ir_binop_rshift:
832 for (unsigned c = 0, c0 = 0, c1 = 0;
834 c0 += c0_inc, c1 += c1_inc, c++) {
836 if (op[0]->type->base_type == GLSL_TYPE_INT &&
837 op[1]->type->base_type == GLSL_TYPE_INT) {
838 data.i[c] = op[0]->value.i[c0] >> op[1]->value.i[c1];
840 } else if (op[0]->type->base_type == GLSL_TYPE_INT &&
841 op[1]->type->base_type == GLSL_TYPE_UINT) {
842 data.i[c] = op[0]->value.i[c0] >> op[1]->value.u[c1];
844 } else if (op[0]->type->base_type == GLSL_TYPE_UINT &&
845 op[1]->type->base_type == GLSL_TYPE_INT) {
846 data.u[c] = op[0]->value.u[c0] >> op[1]->value.i[c1];
848 } else if (op[0]->type->base_type == GLSL_TYPE_UINT &&
849 op[1]->type->base_type == GLSL_TYPE_UINT) {
850 data.u[c] = op[0]->value.u[c0] >> op[1]->value.u[c1];
855 case ir_binop_bit_and:
856 for (unsigned c = 0, c0 = 0, c1 = 0;
858 c0 += c0_inc, c1 += c1_inc, c++) {
860 switch (op[0]->type->base_type) {
862 data.i[c] = op[0]->value.i[c0] & op[1]->value.i[c1];
865 data.u[c] = op[0]->value.u[c0] & op[1]->value.u[c1];
873 case ir_binop_bit_or:
874 for (unsigned c = 0, c0 = 0, c1 = 0;
876 c0 += c0_inc, c1 += c1_inc, c++) {
878 switch (op[0]->type->base_type) {
880 data.i[c] = op[0]->value.i[c0] | op[1]->value.i[c1];
883 data.u[c] = op[0]->value.u[c0] | op[1]->value.u[c1];
891 case ir_binop_bit_xor:
892 for (unsigned c = 0, c0 = 0, c1 = 0;
894 c0 += c0_inc, c1 += c1_inc, c++) {
896 switch (op[0]->type->base_type) {
898 data.i[c] = op[0]->value.i[c0] ^ op[1]->value.i[c1];
901 data.u[c] = op[0]->value.u[c0] ^ op[1]->value.u[c1];
909 case ir_quadop_vector:
910 for (unsigned c = 0; c < this->type->vector_elements; c++) {
911 switch (this->type->base_type) {
913 data.i[c] = op[c]->value.i[0];
916 data.u[c] = op[c]->value.u[0];
918 case GLSL_TYPE_FLOAT:
919 data.f[c] = op[c]->value.f[0];
928 /* FINISHME: Should handle all expression types. */
932 return new(ctx) ir_constant(this->type, &data);
937 ir_texture::constant_expression_value(struct hash_table *variable_context)
939 /* texture lookups aren't constant expressions */
945 ir_swizzle::constant_expression_value(struct hash_table *variable_context)
947 ir_constant *v = this->val->constant_expression_value(variable_context);
950 ir_constant_data data = { { 0 } };
952 const unsigned swiz_idx[4] = {
953 this->mask.x, this->mask.y, this->mask.z, this->mask.w
956 for (unsigned i = 0; i < this->mask.num_components; i++) {
957 switch (v->type->base_type) {
959 case GLSL_TYPE_INT: data.u[i] = v->value.u[swiz_idx[i]]; break;
960 case GLSL_TYPE_FLOAT: data.f[i] = v->value.f[swiz_idx[i]]; break;
961 case GLSL_TYPE_BOOL: data.b[i] = v->value.b[swiz_idx[i]]; break;
962 default: assert(!"Should not get here."); break;
966 void *ctx = ralloc_parent(this);
967 return new(ctx) ir_constant(this->type, &data);
974 ir_dereference_variable::constant_referenced(struct hash_table *variable_context,
975 ir_constant *&store, int &offset) const
977 if (variable_context) {
978 store = (ir_constant *)hash_table_find(variable_context, var);
987 ir_dereference_variable::constant_expression_value(struct hash_table *variable_context)
989 /* This may occur during compile and var->type is glsl_type::error_type */
993 /* Give priority to the context hashtable, if it exists */
994 if (variable_context) {
995 ir_constant *value = (ir_constant *)hash_table_find(variable_context, var);
1000 /* The constant_value of a uniform variable is its initializer,
1001 * not the lifetime constant value of the uniform.
1003 if (var->mode == ir_var_uniform)
1006 if (!var->constant_value)
1009 return var->constant_value->clone(ralloc_parent(var), NULL);
1014 ir_dereference_array::constant_referenced(struct hash_table *variable_context,
1015 ir_constant *&store, int &offset) const
1017 ir_constant *index_c = array_index->constant_expression_value(variable_context);
1019 if (!index_c || !index_c->type->is_scalar() || !index_c->type->is_integer()) {
1025 int index = index_c->type->base_type == GLSL_TYPE_INT ?
1026 index_c->get_int_component(0) :
1027 index_c->get_uint_component(0);
1029 ir_constant *substore;
1031 const ir_dereference *deref = array->as_dereference();
1038 deref->constant_referenced(variable_context, substore, suboffset);
1046 const glsl_type *vt = substore->type;
1047 if (vt->is_array()) {
1048 store = substore->get_array_element(index);
1052 if (vt->is_matrix()) {
1054 offset = index * vt->vector_elements;
1057 if (vt->is_vector()) {
1059 offset = suboffset + index;
1068 ir_dereference_array::constant_expression_value(struct hash_table *variable_context)
1070 ir_constant *array = this->array->constant_expression_value(variable_context);
1071 ir_constant *idx = this->array_index->constant_expression_value(variable_context);
1073 if ((array != NULL) && (idx != NULL)) {
1074 void *ctx = ralloc_parent(this);
1075 if (array->type->is_matrix()) {
1076 /* Array access of a matrix results in a vector.
1078 const unsigned column = idx->value.u[0];
1080 const glsl_type *const column_type = array->type->column_type();
1082 /* Offset in the constant matrix to the first element of the column
1085 const unsigned mat_idx = column * column_type->vector_elements;
1087 ir_constant_data data = { { 0 } };
1089 switch (column_type->base_type) {
1090 case GLSL_TYPE_UINT:
1092 for (unsigned i = 0; i < column_type->vector_elements; i++)
1093 data.u[i] = array->value.u[mat_idx + i];
1097 case GLSL_TYPE_FLOAT:
1098 for (unsigned i = 0; i < column_type->vector_elements; i++)
1099 data.f[i] = array->value.f[mat_idx + i];
1104 assert(!"Should not get here.");
1108 return new(ctx) ir_constant(column_type, &data);
1109 } else if (array->type->is_vector()) {
1110 const unsigned component = idx->value.u[0];
1112 return new(ctx) ir_constant(array, component);
1114 const unsigned index = idx->value.u[0];
1115 return array->get_array_element(index)->clone(ctx, NULL);
1123 ir_dereference_record::constant_referenced(struct hash_table *variable_context,
1124 ir_constant *&store, int &offset) const
1126 ir_constant *substore;
1128 const ir_dereference *deref = record->as_dereference();
1135 deref->constant_referenced(variable_context, substore, suboffset);
1143 store = substore->get_record_field(field);
1148 ir_dereference_record::constant_expression_value(struct hash_table *variable_context)
1150 ir_constant *v = this->record->constant_expression_value();
1152 return (v != NULL) ? v->get_record_field(this->field) : NULL;
1157 ir_assignment::constant_expression_value(struct hash_table *variable_context)
1159 /* FINISHME: Handle CEs involving assignment (return RHS) */
1165 ir_constant::constant_expression_value(struct hash_table *variable_context)
1172 ir_call::constant_expression_value(struct hash_table *variable_context)
1174 return this->callee->constant_expression_value(&this->actual_parameters, variable_context);
1178 bool ir_function_signature::constant_expression_evaluate_expression_list(const struct exec_list &body,
1179 struct hash_table *variable_context,
1180 ir_constant **result)
1182 foreach_list(n, &body) {
1183 ir_instruction *inst = (ir_instruction *)n;
1184 switch(inst->ir_type) {
1186 /* (declare () type symbol) */
1187 case ir_type_variable: {
1188 ir_variable *var = inst->as_variable();
1189 hash_table_insert(variable_context, ir_constant::zero(this, var->type), var);
1193 /* (assign [condition] (write-mask) (ref) (value)) */
1194 case ir_type_assignment: {
1195 ir_assignment *asg = inst->as_assignment();
1196 if (asg->condition) {
1197 ir_constant *cond = asg->condition->constant_expression_value(variable_context);
1200 if (!cond->get_bool_component(0))
1204 ir_constant *store = NULL;
1206 asg->lhs->constant_referenced(variable_context, store, offset);
1211 ir_constant *value = asg->rhs->constant_expression_value(variable_context);
1216 store->copy_masked_offset(value, offset, asg->write_mask);
1220 /* (return (expression)) */
1221 case ir_type_return:
1223 *result = inst->as_return()->value->constant_expression_value(variable_context);
1224 return *result != NULL;
1226 /* (call name (ref) (params))*/
1227 case ir_type_call: {
1228 ir_call *call = inst->as_call();
1230 /* Just say no to void functions in constant expressions. We
1231 * don't need them at that point.
1234 if (!call->return_deref)
1237 ir_constant *store = NULL;
1239 call->return_deref->constant_referenced(variable_context, store, offset);
1244 ir_constant *value = call->constant_expression_value(variable_context);
1249 store->copy_offset(value, offset);
1253 /* (if condition (then-instructions) (else-instructions)) */
1255 ir_if *iif = inst->as_if();
1257 ir_constant *cond = iif->condition->constant_expression_value(variable_context);
1258 if (!cond || !cond->type->is_boolean())
1261 exec_list &branch = cond->get_bool_component(0) ? iif->then_instructions : iif->else_instructions;
1264 if (!constant_expression_evaluate_expression_list(branch, variable_context, result))
1267 /* If there was a return in the branch chosen, drop out now. */
1274 /* Every other expression type, we drop out. */
1280 /* Reaching the end of the block is not an error condition */
1288 ir_function_signature::constant_expression_value(exec_list *actual_parameters, struct hash_table *variable_context)
1290 const glsl_type *type = this->return_type;
1291 if (type == glsl_type::void_type)
1294 /* From the GLSL 1.20 spec, page 23:
1295 * "Function calls to user-defined functions (non-built-in functions)
1296 * cannot be used to form constant expressions."
1298 if (!this->is_builtin)
1302 * Of the builtin functions, only the texture lookups and the noise
1303 * ones must not be used in constant expressions. They all include
1304 * specific opcodes so they don't need to be special-cased at this
1308 /* Initialize the table of dereferencable names with the function
1309 * parameters. Verify their const-ness on the way.
1311 * We expect the correctness of the number of parameters to have
1312 * been checked earlier.
1314 hash_table *deref_hash = hash_table_ctor(8, hash_table_pointer_hash,
1315 hash_table_pointer_compare);
1317 /* If "origin" is non-NULL, then the function body is there. So we
1318 * have to use the variable objects from the object with the body,
1319 * but the parameter instanciation on the current object.
1321 const exec_node *parameter_info = origin ? origin->parameters.head : parameters.head;
1323 foreach_list(n, actual_parameters) {
1324 ir_constant *constant = ((ir_rvalue *) n)->constant_expression_value(variable_context);
1325 if (constant == NULL) {
1326 hash_table_dtor(deref_hash);
1331 ir_variable *var = (ir_variable *)parameter_info;
1332 hash_table_insert(deref_hash, constant, var);
1334 parameter_info = parameter_info->next;
1337 ir_constant *result = NULL;
1339 /* Now run the builtin function until something non-constant
1340 * happens or we get the result.
1342 if (constant_expression_evaluate_expression_list(origin ? origin->body : body, deref_hash, &result) && result)
1343 result = result->clone(ralloc_parent(this), NULL);
1345 hash_table_dtor(deref_hash);