glsl2: Write vector constructor constants in a single assignment
[mesa.git] / src / glsl / ast_function.cpp
1 /*
2 * Copyright © 2010 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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.
22 */
23
24 #include "glsl_symbol_table.h"
25 #include "ast.h"
26 #include "glsl_types.h"
27 #include "ir.h"
28 #include "main/core.h" /* for MIN2 */
29
30 static ir_rvalue *
31 convert_component(ir_rvalue *src, const glsl_type *desired_type);
32
33 static unsigned
34 process_parameters(exec_list *instructions, exec_list *actual_parameters,
35 exec_list *parameters,
36 struct _mesa_glsl_parse_state *state)
37 {
38 unsigned count = 0;
39
40 foreach_list (n, parameters) {
41 ast_node *const ast = exec_node_data(ast_node, n, link);
42 ir_rvalue *result = ast->hir(instructions, state);
43
44 ir_constant *const constant = result->constant_expression_value();
45 if (constant != NULL)
46 result = constant;
47
48 actual_parameters->push_tail(result);
49 count++;
50 }
51
52 return count;
53 }
54
55
56 /**
57 * Generate a source prototype for a function signature
58 *
59 * \param return_type Return type of the function. May be \c NULL.
60 * \param name Name of the function.
61 * \param parameters Parameter list for the function. This may be either a
62 * formal or actual parameter list. Only the type is used.
63 *
64 * \return
65 * A talloced string representing the prototype of the function.
66 */
67 char *
68 prototype_string(const glsl_type *return_type, const char *name,
69 exec_list *parameters)
70 {
71 char *str = NULL;
72
73 if (return_type != NULL)
74 str = talloc_asprintf(str, "%s ", return_type->name);
75
76 str = talloc_asprintf_append(str, "%s(", name);
77
78 const char *comma = "";
79 foreach_list(node, parameters) {
80 const ir_instruction *const param = (ir_instruction *) node;
81
82 str = talloc_asprintf_append(str, "%s%s", comma, param->type->name);
83 comma = ", ";
84 }
85
86 str = talloc_strdup_append(str, ")");
87 return str;
88 }
89
90
91 static ir_rvalue *
92 process_call(exec_list *instructions, ir_function *f,
93 YYLTYPE *loc, exec_list *actual_parameters,
94 struct _mesa_glsl_parse_state *state)
95 {
96 void *ctx = state;
97
98 ir_function_signature *sig = f->matching_signature(actual_parameters);
99
100 /* The instructions param will be used when the FINISHMEs below are done */
101 (void) instructions;
102
103 if (sig != NULL) {
104 /* Verify that 'out' and 'inout' actual parameters are lvalues. This
105 * isn't done in ir_function::matching_signature because that function
106 * cannot generate the necessary diagnostics.
107 */
108 exec_list_iterator actual_iter = actual_parameters->iterator();
109 exec_list_iterator formal_iter = sig->parameters.iterator();
110
111 while (actual_iter.has_next()) {
112 ir_rvalue *actual = (ir_rvalue *) actual_iter.get();
113 ir_variable *formal = (ir_variable *) formal_iter.get();
114
115 assert(actual != NULL);
116 assert(formal != NULL);
117
118 if ((formal->mode == ir_var_out)
119 || (formal->mode == ir_var_inout)) {
120 if (! actual->is_lvalue()) {
121 /* FINISHME: Log a better diagnostic here. There is no way
122 * FINISHME: to tell the user which parameter is invalid.
123 */
124 _mesa_glsl_error(loc, state, "`%s' parameter is not lvalue",
125 (formal->mode == ir_var_out) ? "out" : "inout");
126 }
127 }
128
129 if (formal->type->is_numeric() || formal->type->is_boolean()) {
130 ir_rvalue *converted = convert_component(actual, formal->type);
131 actual->replace_with(converted);
132 }
133
134 actual_iter.next();
135 formal_iter.next();
136 }
137
138 /* Always insert the call in the instruction stream, and return a deref
139 * of its return val if it returns a value, since we don't know if
140 * the rvalue is going to be assigned to anything or not.
141 */
142 ir_call *call = new(ctx) ir_call(sig, actual_parameters);
143 if (!sig->return_type->is_void()) {
144 ir_variable *var;
145 ir_dereference_variable *deref;
146
147 var = new(ctx) ir_variable(sig->return_type,
148 talloc_asprintf(ctx, "%s_retval",
149 sig->function_name()),
150 ir_var_temporary);
151 instructions->push_tail(var);
152
153 deref = new(ctx) ir_dereference_variable(var);
154 ir_assignment *assign = new(ctx) ir_assignment(deref, call, NULL);
155 instructions->push_tail(assign);
156 if (state->language_version >= 120)
157 var->constant_value = call->constant_expression_value();
158
159 deref = new(ctx) ir_dereference_variable(var);
160 return deref;
161 } else {
162 instructions->push_tail(call);
163 return NULL;
164 }
165 } else {
166 char *str = prototype_string(NULL, f->name, actual_parameters);
167
168 _mesa_glsl_error(loc, state, "no matching function for call to `%s'",
169 str);
170 talloc_free(str);
171
172 const char *prefix = "candidates are: ";
173 foreach_list (node, &f->signatures) {
174 ir_function_signature *sig = (ir_function_signature *) node;
175
176 str = prototype_string(sig->return_type, f->name, &sig->parameters);
177 _mesa_glsl_error(loc, state, "%s%s\n", prefix, str);
178 talloc_free(str);
179
180 prefix = " ";
181 }
182
183 return ir_call::get_error_instruction(ctx);
184 }
185 }
186
187
188 static ir_rvalue *
189 match_function_by_name(exec_list *instructions, const char *name,
190 YYLTYPE *loc, exec_list *actual_parameters,
191 struct _mesa_glsl_parse_state *state)
192 {
193 void *ctx = state;
194 ir_function *f = state->symbols->get_function(name);
195
196 if (f == NULL) {
197 _mesa_glsl_error(loc, state, "function `%s' undeclared", name);
198 return ir_call::get_error_instruction(ctx);
199 }
200
201 /* Once we've determined that the function being called might exist, try
202 * to find an overload of the function that matches the parameters.
203 */
204 return process_call(instructions, f, loc, actual_parameters, state);
205 }
206
207
208 /**
209 * Perform automatic type conversion of constructor parameters
210 *
211 * This implements the rules in the "Conversion and Scalar Constructors"
212 * section (GLSL 1.10 section 5.4.1), not the "Implicit Conversions" rules.
213 */
214 static ir_rvalue *
215 convert_component(ir_rvalue *src, const glsl_type *desired_type)
216 {
217 void *ctx = talloc_parent(src);
218 const unsigned a = desired_type->base_type;
219 const unsigned b = src->type->base_type;
220 ir_expression *result = NULL;
221
222 if (src->type->is_error())
223 return src;
224
225 assert(a <= GLSL_TYPE_BOOL);
226 assert(b <= GLSL_TYPE_BOOL);
227
228 if ((a == b) || (src->type->is_integer() && desired_type->is_integer()))
229 return src;
230
231 switch (a) {
232 case GLSL_TYPE_UINT:
233 case GLSL_TYPE_INT:
234 if (b == GLSL_TYPE_FLOAT)
235 result = new(ctx) ir_expression(ir_unop_f2i, desired_type, src, NULL);
236 else {
237 assert(b == GLSL_TYPE_BOOL);
238 result = new(ctx) ir_expression(ir_unop_b2i, desired_type, src, NULL);
239 }
240 break;
241 case GLSL_TYPE_FLOAT:
242 switch (b) {
243 case GLSL_TYPE_UINT:
244 result = new(ctx) ir_expression(ir_unop_u2f, desired_type, src, NULL);
245 break;
246 case GLSL_TYPE_INT:
247 result = new(ctx) ir_expression(ir_unop_i2f, desired_type, src, NULL);
248 break;
249 case GLSL_TYPE_BOOL:
250 result = new(ctx) ir_expression(ir_unop_b2f, desired_type, src, NULL);
251 break;
252 }
253 break;
254 case GLSL_TYPE_BOOL:
255 switch (b) {
256 case GLSL_TYPE_UINT:
257 case GLSL_TYPE_INT:
258 result = new(ctx) ir_expression(ir_unop_i2b, desired_type, src, NULL);
259 break;
260 case GLSL_TYPE_FLOAT:
261 result = new(ctx) ir_expression(ir_unop_f2b, desired_type, src, NULL);
262 break;
263 }
264 break;
265 }
266
267 assert(result != NULL);
268
269 /* Try constant folding; it may fold in the conversion we just added. */
270 ir_constant *const constant = result->constant_expression_value();
271 return (constant != NULL) ? (ir_rvalue *) constant : (ir_rvalue *) result;
272 }
273
274 /**
275 * Dereference a specific component from a scalar, vector, or matrix
276 */
277 static ir_rvalue *
278 dereference_component(ir_rvalue *src, unsigned component)
279 {
280 void *ctx = talloc_parent(src);
281 assert(component < src->type->components());
282
283 /* If the source is a constant, just create a new constant instead of a
284 * dereference of the existing constant.
285 */
286 ir_constant *constant = src->as_constant();
287 if (constant)
288 return new(ctx) ir_constant(constant, component);
289
290 if (src->type->is_scalar()) {
291 return src;
292 } else if (src->type->is_vector()) {
293 return new(ctx) ir_swizzle(src, component, 0, 0, 0, 1);
294 } else {
295 assert(src->type->is_matrix());
296
297 /* Dereference a row of the matrix, then call this function again to get
298 * a specific element from that row.
299 */
300 const int c = component / src->type->column_type()->vector_elements;
301 const int r = component % src->type->column_type()->vector_elements;
302 ir_constant *const col_index = new(ctx) ir_constant(c);
303 ir_dereference *const col = new(ctx) ir_dereference_array(src, col_index);
304
305 col->type = src->type->column_type();
306
307 return dereference_component(col, r);
308 }
309
310 assert(!"Should not get here.");
311 return NULL;
312 }
313
314
315 static ir_rvalue *
316 process_array_constructor(exec_list *instructions,
317 const glsl_type *constructor_type,
318 YYLTYPE *loc, exec_list *parameters,
319 struct _mesa_glsl_parse_state *state)
320 {
321 void *ctx = state;
322 /* Array constructors come in two forms: sized and unsized. Sized array
323 * constructors look like 'vec4[2](a, b)', where 'a' and 'b' are vec4
324 * variables. In this case the number of parameters must exactly match the
325 * specified size of the array.
326 *
327 * Unsized array constructors look like 'vec4[](a, b)', where 'a' and 'b'
328 * are vec4 variables. In this case the size of the array being constructed
329 * is determined by the number of parameters.
330 *
331 * From page 52 (page 58 of the PDF) of the GLSL 1.50 spec:
332 *
333 * "There must be exactly the same number of arguments as the size of
334 * the array being constructed. If no size is present in the
335 * constructor, then the array is explicitly sized to the number of
336 * arguments provided. The arguments are assigned in order, starting at
337 * element 0, to the elements of the constructed array. Each argument
338 * must be the same type as the element type of the array, or be a type
339 * that can be converted to the element type of the array according to
340 * Section 4.1.10 "Implicit Conversions.""
341 */
342 exec_list actual_parameters;
343 const unsigned parameter_count =
344 process_parameters(instructions, &actual_parameters, parameters, state);
345
346 if ((parameter_count == 0)
347 || ((constructor_type->length != 0)
348 && (constructor_type->length != parameter_count))) {
349 const unsigned min_param = (constructor_type->length == 0)
350 ? 1 : constructor_type->length;
351
352 _mesa_glsl_error(loc, state, "array constructor must have %s %u "
353 "parameter%s",
354 (constructor_type->length != 0) ? "at least" : "exactly",
355 min_param, (min_param <= 1) ? "" : "s");
356 return ir_call::get_error_instruction(ctx);
357 }
358
359 if (constructor_type->length == 0) {
360 constructor_type =
361 glsl_type::get_array_instance(constructor_type->element_type(),
362 parameter_count);
363 assert(constructor_type != NULL);
364 assert(constructor_type->length == parameter_count);
365 }
366
367 bool all_parameters_are_constant = true;
368
369 /* Type cast each parameter and, if possible, fold constants. */
370 foreach_list_safe(n, &actual_parameters) {
371 ir_rvalue *ir = (ir_rvalue *) n;
372 ir_rvalue *result = ir;
373
374 /* Apply implicit conversions (not the scalar constructor rules!) */
375 if (constructor_type->element_type()->is_float()) {
376 const glsl_type *desired_type =
377 glsl_type::get_instance(GLSL_TYPE_FLOAT,
378 ir->type->vector_elements,
379 ir->type->matrix_columns);
380 result = convert_component(ir, desired_type);
381 }
382
383 if (result->type != constructor_type->element_type()) {
384 _mesa_glsl_error(loc, state, "type error in array constructor: "
385 "expected: %s, found %s",
386 constructor_type->element_type()->name,
387 result->type->name);
388 }
389
390 /* Attempt to convert the parameter to a constant valued expression.
391 * After doing so, track whether or not all the parameters to the
392 * constructor are trivially constant valued expressions.
393 */
394 ir_rvalue *const constant = result->constant_expression_value();
395
396 if (constant != NULL)
397 result = constant;
398 else
399 all_parameters_are_constant = false;
400
401 ir->replace_with(result);
402 }
403
404 if (all_parameters_are_constant)
405 return new(ctx) ir_constant(constructor_type, &actual_parameters);
406
407 ir_variable *var = new(ctx) ir_variable(constructor_type, "array_ctor",
408 ir_var_temporary);
409 instructions->push_tail(var);
410
411 int i = 0;
412 foreach_list(node, &actual_parameters) {
413 ir_rvalue *rhs = (ir_rvalue *) node;
414 ir_rvalue *lhs = new(ctx) ir_dereference_array(var,
415 new(ctx) ir_constant(i));
416
417 ir_instruction *assignment = new(ctx) ir_assignment(lhs, rhs, NULL);
418 instructions->push_tail(assignment);
419
420 i++;
421 }
422
423 return new(ctx) ir_dereference_variable(var);
424 }
425
426
427 /**
428 * Try to convert a record constructor to a constant expression
429 */
430 static ir_constant *
431 constant_record_constructor(const glsl_type *constructor_type,
432 YYLTYPE *loc, exec_list *parameters,
433 struct _mesa_glsl_parse_state *state)
434 {
435 void *ctx = state;
436 bool all_parameters_are_constant = true;
437
438 exec_node *node = parameters->head;
439 for (unsigned i = 0; i < constructor_type->length; i++) {
440 ir_instruction *ir = (ir_instruction *) node;
441
442 if (node->is_tail_sentinel()) {
443 _mesa_glsl_error(loc, state,
444 "insufficient parameters to constructor for `%s'",
445 constructor_type->name);
446 return NULL;
447 }
448
449 if (ir->type != constructor_type->fields.structure[i].type) {
450 _mesa_glsl_error(loc, state,
451 "parameter type mismatch in constructor for `%s' "
452 " (%s vs %s)",
453 constructor_type->name,
454 ir->type->name,
455 constructor_type->fields.structure[i].type->name);
456 return NULL;
457 }
458
459 if (ir->as_constant() == NULL)
460 all_parameters_are_constant = false;
461
462 node = node->next;
463 }
464
465 if (!all_parameters_are_constant)
466 return NULL;
467
468 return new(ctx) ir_constant(constructor_type, parameters);
469 }
470
471
472 /**
473 * Generate data for a constant matrix constructor w/a single scalar parameter
474 *
475 * Matrix constructors in GLSL can be passed a single scalar of the
476 * approriate type. In these cases, the resulting matrix is the identity
477 * matrix multipled by the specified scalar. This function generates data for
478 * that matrix.
479 *
480 * \param type Type of the desired matrix.
481 * \param initializer Scalar value used to initialize the matrix diagonal.
482 * \param data Location to store the resulting matrix.
483 */
484 void
485 generate_constructor_matrix(const glsl_type *type, ir_constant *initializer,
486 ir_constant_data *data)
487 {
488 switch (type->base_type) {
489 case GLSL_TYPE_UINT:
490 case GLSL_TYPE_INT:
491 for (unsigned i = 0; i < type->components(); i++)
492 data->u[i] = 0;
493
494 for (unsigned i = 0; i < type->matrix_columns; i++) {
495 /* The array offset of the ith row and column of the matrix.
496 */
497 const unsigned idx = (i * type->vector_elements) + i;
498
499 data->u[idx] = initializer->value.u[0];
500 }
501 break;
502
503 case GLSL_TYPE_FLOAT:
504 for (unsigned i = 0; i < type->components(); i++)
505 data->f[i] = 0;
506
507 for (unsigned i = 0; i < type->matrix_columns; i++) {
508 /* The array offset of the ith row and column of the matrix.
509 */
510 const unsigned idx = (i * type->vector_elements) + i;
511
512 data->f[idx] = initializer->value.f[0];
513 }
514
515 break;
516
517 default:
518 assert(!"Should not get here.");
519 break;
520 }
521 }
522
523
524 /**
525 * Generate data for a constant vector constructor w/a single scalar parameter
526 *
527 * Vector constructors in GLSL can be passed a single scalar of the
528 * approriate type. In these cases, the resulting vector contains the specified
529 * value in all components. This function generates data for that vector.
530 *
531 * \param type Type of the desired vector.
532 * \param initializer Scalar value used to initialize the vector.
533 * \param data Location to store the resulting vector data.
534 */
535 void
536 generate_constructor_vector(const glsl_type *type, ir_constant *initializer,
537 ir_constant_data *data)
538 {
539 switch (type->base_type) {
540 case GLSL_TYPE_UINT:
541 case GLSL_TYPE_INT:
542 for (unsigned i = 0; i < type->components(); i++)
543 data->u[i] = initializer->value.u[0];
544
545 break;
546
547 case GLSL_TYPE_FLOAT:
548 for (unsigned i = 0; i < type->components(); i++)
549 data->f[i] = initializer->value.f[0];
550
551 break;
552
553 case GLSL_TYPE_BOOL:
554 for (unsigned i = 0; i < type->components(); i++)
555 data->b[i] = initializer->value.b[0];
556
557 break;
558
559 default:
560 assert(!"Should not get here.");
561 break;
562 }
563 }
564
565
566 /**
567 * Determine if a list consists of a single scalar r-value
568 */
569 bool
570 single_scalar_parameter(exec_list *parameters)
571 {
572 const ir_rvalue *const p = (ir_rvalue *) parameters->head;
573 assert(((ir_rvalue *)p)->as_rvalue() != NULL);
574
575 return (p->type->is_scalar() && p->next->is_tail_sentinel());
576 }
577
578
579 /**
580 * Generate inline code for a vector constructor
581 *
582 * The generated constructor code will consist of a temporary variable
583 * declaration of the same type as the constructor. A sequence of assignments
584 * from constructor parameters to the temporary will follow.
585 *
586 * \return
587 * An \c ir_dereference_variable of the temprorary generated in the constructor
588 * body.
589 */
590 ir_rvalue *
591 emit_inline_vector_constructor(const glsl_type *type,
592 exec_list *instructions,
593 exec_list *parameters,
594 void *ctx)
595 {
596 assert(!parameters->is_empty());
597
598 ir_variable *var = new(ctx) ir_variable(type, "vec_ctor", ir_var_temporary);
599 instructions->push_tail(var);
600
601 /* There are two kinds of vector constructors.
602 *
603 * - Construct a vector from a single scalar by replicating that scalar to
604 * all components of the vector.
605 *
606 * - Construct a vector from an arbirary combination of vectors and
607 * scalars. The components of the constructor parameters are assigned
608 * to the vector in order until the vector is full.
609 */
610 const unsigned lhs_components = type->components();
611 if (single_scalar_parameter(parameters)) {
612 ir_rvalue *first_param = (ir_rvalue *)parameters->head;
613 ir_rvalue *rhs = new(ctx) ir_swizzle(first_param, 0, 0, 0, 0,
614 lhs_components);
615 ir_dereference_variable *lhs = new(ctx) ir_dereference_variable(var);
616 const unsigned mask = (1U << lhs_components) - 1;
617
618 assert(rhs->type == lhs->type);
619
620 ir_instruction *inst = new(ctx) ir_assignment(lhs, rhs, NULL, mask);
621 instructions->push_tail(inst);
622 } else {
623 unsigned base_component = 0;
624 ir_constant_data data;
625 unsigned constant_mask = 0;
626
627 memset(&data, 0, sizeof(data));
628
629 foreach_list(node, parameters) {
630 ir_rvalue *param = (ir_rvalue *) node;
631 unsigned rhs_components = param->type->components();
632
633 /* Do not try to assign more components to the vector than it has!
634 */
635 if ((rhs_components + base_component) > lhs_components) {
636 rhs_components = lhs_components - base_component;
637 }
638
639 const ir_constant *const c = param->as_constant();
640 if (c != NULL) {
641 for (unsigned i = 0; i < rhs_components; i++) {
642 switch (c->type->base_type) {
643 case GLSL_TYPE_UINT:
644 data.u[i + base_component] = c->get_uint_component(i);
645 break;
646 case GLSL_TYPE_INT:
647 data.i[i + base_component] = c->get_int_component(i);
648 break;
649 case GLSL_TYPE_FLOAT:
650 data.f[i + base_component] = c->get_float_component(i);
651 break;
652 case GLSL_TYPE_BOOL:
653 data.b[i + base_component] = c->get_bool_component(i);
654 break;
655 default:
656 assert(!"Should not get here.");
657 break;
658 }
659 }
660
661 /* Mask of fields to be written in the assignment.
662 */
663 constant_mask |= ((1U << rhs_components) - 1) << base_component;
664 }
665
666 /* Advance the component index by the number of components that were
667 * just assigned.
668 */
669 base_component += rhs_components;
670 }
671
672 if (constant_mask != 0) {
673 ir_dereference *lhs = new(ctx) ir_dereference_variable(var);
674 ir_rvalue *rhs = new(ctx) ir_constant(var->type, &data);
675
676 ir_instruction *inst =
677 new(ctx) ir_assignment(lhs, rhs, NULL, constant_mask);
678 instructions->push_tail(inst);
679 }
680
681 base_component = 0;
682 foreach_list(node, parameters) {
683 ir_rvalue *param = (ir_rvalue *) node;
684 unsigned rhs_components = param->type->components();
685
686 /* Do not try to assign more components to the vector than it has!
687 */
688 if ((rhs_components + base_component) > lhs_components) {
689 rhs_components = lhs_components - base_component;
690 }
691
692 const ir_constant *const c = param->as_constant();
693 if (c == NULL) {
694 /* Generate a swizzle that puts the first element of the source at
695 * the location of the first element of the destination.
696 */
697 unsigned swiz[4] = { 0, 0, 0, 0 };
698 for (unsigned i = 0; i < rhs_components; i++)
699 swiz[i + base_component] = i;
700
701 /* Mask of fields to be written in the assignment.
702 */
703 const unsigned write_mask = ((1U << rhs_components) - 1)
704 << base_component;
705
706 ir_dereference *lhs = new(ctx) ir_dereference_variable(var);
707 ir_rvalue *rhs = new(ctx) ir_swizzle(param, swiz, lhs_components);
708
709 ir_instruction *inst =
710 new(ctx) ir_assignment(lhs, rhs, NULL, write_mask);
711 instructions->push_tail(inst);
712 }
713
714 /* Advance the component index by the number of components that were
715 * just assigned.
716 */
717 base_component += rhs_components;
718 }
719 }
720 return new(ctx) ir_dereference_variable(var);
721 }
722
723
724 /**
725 * Generate assignment of a portion of a vector to a portion of a matrix column
726 *
727 * \param src_base First component of the source to be used in assignment
728 * \param column Column of destination to be assiged
729 * \param row_base First component of the destination column to be assigned
730 * \param count Number of components to be assigned
731 *
732 * \note
733 * \c src_base + \c count must be less than or equal to the number of components
734 * in the source vector.
735 */
736 ir_instruction *
737 assign_to_matrix_column(ir_variable *var, unsigned column, unsigned row_base,
738 ir_rvalue *src, unsigned src_base, unsigned count,
739 void *mem_ctx)
740 {
741 ir_constant *col_idx = new(mem_ctx) ir_constant(column);
742 ir_dereference *column_ref = new(mem_ctx) ir_dereference_array(var, col_idx);
743
744 assert(column_ref->type->components() >= (row_base + count));
745 assert(src->type->components() >= (src_base + count));
746
747 /* Generate a swizzle that puts the first element of the source at the
748 * location of the first element of the destination.
749 */
750 unsigned swiz[4] = { src_base, src_base, src_base, src_base };
751 for (unsigned i = 0; i < count; i++)
752 swiz[i + row_base] = src_base + i;
753
754 ir_rvalue *const rhs =
755 new(mem_ctx) ir_swizzle(src, swiz, column_ref->type->components());
756
757 /* Mask of fields to be written in the assignment.
758 */
759 const unsigned write_mask = ((1U << count) - 1) << row_base;
760
761 return new(mem_ctx) ir_assignment(column_ref, rhs, NULL, write_mask);
762 }
763
764
765 /**
766 * Generate inline code for a matrix constructor
767 *
768 * The generated constructor code will consist of a temporary variable
769 * declaration of the same type as the constructor. A sequence of assignments
770 * from constructor parameters to the temporary will follow.
771 *
772 * \return
773 * An \c ir_dereference_variable of the temprorary generated in the constructor
774 * body.
775 */
776 ir_rvalue *
777 emit_inline_matrix_constructor(const glsl_type *type,
778 exec_list *instructions,
779 exec_list *parameters,
780 void *ctx)
781 {
782 assert(!parameters->is_empty());
783
784 ir_variable *var = new(ctx) ir_variable(type, "mat_ctor", ir_var_temporary);
785 instructions->push_tail(var);
786
787 /* There are three kinds of matrix constructors.
788 *
789 * - Construct a matrix from a single scalar by replicating that scalar to
790 * along the diagonal of the matrix and setting all other components to
791 * zero.
792 *
793 * - Construct a matrix from an arbirary combination of vectors and
794 * scalars. The components of the constructor parameters are assigned
795 * to the matrix in colum-major order until the matrix is full.
796 *
797 * - Construct a matrix from a single matrix. The source matrix is copied
798 * to the upper left portion of the constructed matrix, and the remaining
799 * elements take values from the identity matrix.
800 */
801 ir_rvalue *const first_param = (ir_rvalue *) parameters->head;
802 if (single_scalar_parameter(parameters)) {
803 /* Assign the scalar to the X component of a vec4, and fill the remaining
804 * components with zero.
805 */
806 ir_variable *rhs_var =
807 new(ctx) ir_variable(glsl_type::vec4_type, "mat_ctor_vec",
808 ir_var_temporary);
809 instructions->push_tail(rhs_var);
810
811 ir_constant_data zero;
812 zero.f[0] = 0.0;
813 zero.f[1] = 0.0;
814 zero.f[2] = 0.0;
815 zero.f[3] = 0.0;
816
817 ir_instruction *inst =
818 new(ctx) ir_assignment(new(ctx) ir_dereference_variable(rhs_var),
819 new(ctx) ir_constant(rhs_var->type, &zero),
820 NULL);
821 instructions->push_tail(inst);
822
823 ir_dereference *const rhs_ref = new(ctx) ir_dereference_variable(rhs_var);
824
825 inst = new(ctx) ir_assignment(rhs_ref, first_param, NULL, 0x01);
826 instructions->push_tail(inst);
827
828 /* Assign the temporary vector to each column of the destination matrix
829 * with a swizzle that puts the X component on the diagonal of the
830 * matrix. In some cases this may mean that the X component does not
831 * get assigned into the column at all (i.e., when the matrix has more
832 * columns than rows).
833 */
834 static const unsigned rhs_swiz[4][4] = {
835 { 0, 1, 1, 1 },
836 { 1, 0, 1, 1 },
837 { 1, 1, 0, 1 },
838 { 1, 1, 1, 0 }
839 };
840
841 const unsigned cols_to_init = MIN2(type->matrix_columns,
842 type->vector_elements);
843 for (unsigned i = 0; i < cols_to_init; i++) {
844 ir_constant *const col_idx = new(ctx) ir_constant(i);
845 ir_rvalue *const col_ref = new(ctx) ir_dereference_array(var, col_idx);
846
847 ir_rvalue *const rhs_ref = new(ctx) ir_dereference_variable(rhs_var);
848 ir_rvalue *const rhs = new(ctx) ir_swizzle(rhs_ref, rhs_swiz[i],
849 type->vector_elements);
850
851 inst = new(ctx) ir_assignment(col_ref, rhs, NULL);
852 instructions->push_tail(inst);
853 }
854
855 for (unsigned i = cols_to_init; i < type->matrix_columns; i++) {
856 ir_constant *const col_idx = new(ctx) ir_constant(i);
857 ir_rvalue *const col_ref = new(ctx) ir_dereference_array(var, col_idx);
858
859 ir_rvalue *const rhs_ref = new(ctx) ir_dereference_variable(rhs_var);
860 ir_rvalue *const rhs = new(ctx) ir_swizzle(rhs_ref, 1, 1, 1, 1,
861 type->vector_elements);
862
863 inst = new(ctx) ir_assignment(col_ref, rhs, NULL);
864 instructions->push_tail(inst);
865 }
866 } else if (first_param->type->is_matrix()) {
867 /* From page 50 (56 of the PDF) of the GLSL 1.50 spec:
868 *
869 * "If a matrix is constructed from a matrix, then each component
870 * (column i, row j) in the result that has a corresponding
871 * component (column i, row j) in the argument will be initialized
872 * from there. All other components will be initialized to the
873 * identity matrix. If a matrix argument is given to a matrix
874 * constructor, it is an error to have any other arguments."
875 */
876 assert(first_param->next->is_tail_sentinel());
877 ir_rvalue *const src_matrix = first_param;
878
879 /* If the source matrix is smaller, pre-initialize the relavent parts of
880 * the destination matrix to the identity matrix.
881 */
882 if ((src_matrix->type->matrix_columns < var->type->matrix_columns)
883 || (src_matrix->type->vector_elements < var->type->vector_elements)) {
884
885 /* If the source matrix has fewer rows, every column of the destination
886 * must be initialized. Otherwise only the columns in the destination
887 * that do not exist in the source must be initialized.
888 */
889 unsigned col =
890 (src_matrix->type->vector_elements < var->type->vector_elements)
891 ? 0 : src_matrix->type->matrix_columns;
892
893 const glsl_type *const col_type = var->type->column_type();
894 for (/* empty */; col < var->type->matrix_columns; col++) {
895 ir_constant_data ident;
896
897 ident.f[0] = 0.0;
898 ident.f[1] = 0.0;
899 ident.f[2] = 0.0;
900 ident.f[3] = 0.0;
901
902 ident.f[col] = 1.0;
903
904 ir_rvalue *const rhs = new(ctx) ir_constant(col_type, &ident);
905
906 ir_rvalue *const lhs =
907 new(ctx) ir_dereference_array(var, new(ctx) ir_constant(col));
908
909 ir_instruction *inst = new(ctx) ir_assignment(lhs, rhs, NULL);
910 instructions->push_tail(inst);
911 }
912 }
913
914 /* Assign columns from the source matrix to the destination matrix.
915 *
916 * Since the parameter will be used in the RHS of multiple assignments,
917 * generate a temporary and copy the paramter there.
918 */
919 ir_variable *const rhs_var =
920 new(ctx) ir_variable(first_param->type, "mat_ctor_mat",
921 ir_var_temporary);
922 instructions->push_tail(rhs_var);
923
924 ir_dereference *const rhs_var_ref =
925 new(ctx) ir_dereference_variable(rhs_var);
926 ir_instruction *const inst =
927 new(ctx) ir_assignment(rhs_var_ref, first_param, NULL);
928 instructions->push_tail(inst);
929
930
931 unsigned swiz[4] = { 0, 0, 0, 0 };
932 for (unsigned i = 1; i < src_matrix->type->vector_elements; i++)
933 swiz[i] = i;
934
935 const unsigned last_col = MIN2(src_matrix->type->matrix_columns,
936 var->type->matrix_columns);
937 const unsigned write_mask = (1U << var->type->vector_elements) - 1;
938
939 for (unsigned i = 0; i < last_col; i++) {
940 ir_dereference *const lhs =
941 new(ctx) ir_dereference_array(var, new(ctx) ir_constant(i));
942 ir_rvalue *const rhs_col =
943 new(ctx) ir_dereference_array(rhs_var, new(ctx) ir_constant(i));
944
945 /* If one matrix has columns that are smaller than the columns of the
946 * other matrix, wrap the column access of the larger with a swizzle
947 * so that the LHS and RHS of the assignment have the same size (and
948 * therefore have the same type).
949 *
950 * It would be perfectly valid to unconditionally generate the
951 * swizzles, this this will typically result in a more compact IR tree.
952 */
953 ir_rvalue *rhs;
954 if (lhs->type->vector_elements != rhs_col->type->vector_elements) {
955 rhs = new(ctx) ir_swizzle(rhs_col, swiz,
956 lhs->type->vector_elements);
957 } else {
958 rhs = rhs_col;
959 }
960
961 assert(lhs->type == rhs->type);
962
963 ir_instruction *inst =
964 new(ctx) ir_assignment(lhs, rhs, NULL, write_mask);
965 instructions->push_tail(inst);
966 }
967 } else {
968 const unsigned cols = type->matrix_columns;
969 const unsigned rows = type->vector_elements;
970 unsigned col_idx = 0;
971 unsigned row_idx = 0;
972
973 foreach_list (node, parameters) {
974 ir_rvalue *const rhs = (ir_rvalue *) node;
975 const unsigned components_remaining_this_column = rows - row_idx;
976 unsigned rhs_components = rhs->type->components();
977 unsigned rhs_base = 0;
978
979 /* Since the parameter might be used in the RHS of two assignments,
980 * generate a temporary and copy the paramter there.
981 */
982 ir_variable *rhs_var =
983 new(ctx) ir_variable(rhs->type, "mat_ctor_vec", ir_var_temporary);
984 instructions->push_tail(rhs_var);
985
986 ir_dereference *rhs_var_ref =
987 new(ctx) ir_dereference_variable(rhs_var);
988 ir_instruction *inst = new(ctx) ir_assignment(rhs_var_ref, rhs, NULL);
989 instructions->push_tail(inst);
990
991 /* Assign the current parameter to as many components of the matrix
992 * as it will fill.
993 *
994 * NOTE: A single vector parameter can span two matrix columns. A
995 * single vec4, for example, can completely fill a mat2.
996 */
997 if (rhs_components >= components_remaining_this_column) {
998 const unsigned count = MIN2(rhs_components,
999 components_remaining_this_column);
1000
1001 rhs_var_ref = new(ctx) ir_dereference_variable(rhs_var);
1002
1003 ir_instruction *inst = assign_to_matrix_column(var, col_idx,
1004 row_idx,
1005 rhs_var_ref, 0,
1006 count, ctx);
1007 instructions->push_tail(inst);
1008
1009 rhs_base = count;
1010
1011 col_idx++;
1012 row_idx = 0;
1013 }
1014
1015 /* If there is data left in the parameter and components left to be
1016 * set in the destination, emit another assignment. It is possible
1017 * that the assignment could be of a vec4 to the last element of the
1018 * matrix. In this case col_idx==cols, but there is still data
1019 * left in the source parameter. Obviously, don't emit an assignment
1020 * to data outside the destination matrix.
1021 */
1022 if ((col_idx < cols) && (rhs_base < rhs_components)) {
1023 const unsigned count = rhs_components - rhs_base;
1024
1025 rhs_var_ref = new(ctx) ir_dereference_variable(rhs_var);
1026
1027 ir_instruction *inst = assign_to_matrix_column(var, col_idx,
1028 row_idx,
1029 rhs_var_ref,
1030 rhs_base,
1031 count, ctx);
1032 instructions->push_tail(inst);
1033
1034 row_idx += count;
1035 }
1036 }
1037 }
1038
1039 return new(ctx) ir_dereference_variable(var);
1040 }
1041
1042
1043 ir_rvalue *
1044 ast_function_expression::hir(exec_list *instructions,
1045 struct _mesa_glsl_parse_state *state)
1046 {
1047 void *ctx = state;
1048 /* There are three sorts of function calls.
1049 *
1050 * 1. constructors - The first subexpression is an ast_type_specifier.
1051 * 2. methods - Only the .length() method of array types.
1052 * 3. functions - Calls to regular old functions.
1053 *
1054 * Method calls are actually detected when the ast_field_selection
1055 * expression is handled.
1056 */
1057 if (is_constructor()) {
1058 const ast_type_specifier *type = (ast_type_specifier *) subexpressions[0];
1059 YYLTYPE loc = type->get_location();
1060 const char *name;
1061
1062 const glsl_type *const constructor_type = type->glsl_type(& name, state);
1063
1064
1065 /* Constructors for samplers are illegal.
1066 */
1067 if (constructor_type->is_sampler()) {
1068 _mesa_glsl_error(& loc, state, "cannot construct sampler type `%s'",
1069 constructor_type->name);
1070 return ir_call::get_error_instruction(ctx);
1071 }
1072
1073 if (constructor_type->is_array()) {
1074 if (state->language_version <= 110) {
1075 _mesa_glsl_error(& loc, state,
1076 "array constructors forbidden in GLSL 1.10");
1077 return ir_call::get_error_instruction(ctx);
1078 }
1079
1080 return process_array_constructor(instructions, constructor_type,
1081 & loc, &this->expressions, state);
1082 }
1083
1084 /* There are two kinds of constructor call. Constructors for built-in
1085 * language types, such as mat4 and vec2, are free form. The only
1086 * requirement is that the parameters must provide enough values of the
1087 * correct scalar type. Constructors for arrays and structures must
1088 * have the exact number of parameters with matching types in the
1089 * correct order. These constructors follow essentially the same type
1090 * matching rules as functions.
1091 */
1092 if (!constructor_type->is_numeric() && !constructor_type->is_boolean())
1093 return ir_call::get_error_instruction(ctx);
1094
1095 /* Total number of components of the type being constructed. */
1096 const unsigned type_components = constructor_type->components();
1097
1098 /* Number of components from parameters that have actually been
1099 * consumed. This is used to perform several kinds of error checking.
1100 */
1101 unsigned components_used = 0;
1102
1103 unsigned matrix_parameters = 0;
1104 unsigned nonmatrix_parameters = 0;
1105 exec_list actual_parameters;
1106
1107 foreach_list (n, &this->expressions) {
1108 ast_node *ast = exec_node_data(ast_node, n, link);
1109 ir_rvalue *result = ast->hir(instructions, state)->as_rvalue();
1110
1111 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
1112 *
1113 * "It is an error to provide extra arguments beyond this
1114 * last used argument."
1115 */
1116 if (components_used >= type_components) {
1117 _mesa_glsl_error(& loc, state, "too many parameters to `%s' "
1118 "constructor",
1119 constructor_type->name);
1120 return ir_call::get_error_instruction(ctx);
1121 }
1122
1123 if (!result->type->is_numeric() && !result->type->is_boolean()) {
1124 _mesa_glsl_error(& loc, state, "cannot construct `%s' from a "
1125 "non-numeric data type",
1126 constructor_type->name);
1127 return ir_call::get_error_instruction(ctx);
1128 }
1129
1130 /* Count the number of matrix and nonmatrix parameters. This
1131 * is used below to enforce some of the constructor rules.
1132 */
1133 if (result->type->is_matrix())
1134 matrix_parameters++;
1135 else
1136 nonmatrix_parameters++;
1137
1138 actual_parameters.push_tail(result);
1139 components_used += result->type->components();
1140 }
1141
1142 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
1143 *
1144 * "It is an error to construct matrices from other matrices. This
1145 * is reserved for future use."
1146 */
1147 if ((state->language_version <= 110) && (matrix_parameters > 0)
1148 && constructor_type->is_matrix()) {
1149 _mesa_glsl_error(& loc, state, "cannot construct `%s' from a "
1150 "matrix in GLSL 1.10",
1151 constructor_type->name);
1152 return ir_call::get_error_instruction(ctx);
1153 }
1154
1155 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
1156 *
1157 * "If a matrix argument is given to a matrix constructor, it is
1158 * an error to have any other arguments."
1159 */
1160 if ((matrix_parameters > 0)
1161 && ((matrix_parameters + nonmatrix_parameters) > 1)
1162 && constructor_type->is_matrix()) {
1163 _mesa_glsl_error(& loc, state, "for matrix `%s' constructor, "
1164 "matrix must be only parameter",
1165 constructor_type->name);
1166 return ir_call::get_error_instruction(ctx);
1167 }
1168
1169 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
1170 *
1171 * "In these cases, there must be enough components provided in the
1172 * arguments to provide an initializer for every component in the
1173 * constructed value."
1174 */
1175 if ((components_used < type_components) && (components_used != 1)) {
1176 _mesa_glsl_error(& loc, state, "too few components to construct "
1177 "`%s'",
1178 constructor_type->name);
1179 return ir_call::get_error_instruction(ctx);
1180 }
1181
1182 /* Later, we cast each parameter to the same base type as the
1183 * constructor. Since there are no non-floating point matrices, we
1184 * need to break them up into a series of column vectors.
1185 */
1186 if (constructor_type->base_type != GLSL_TYPE_FLOAT) {
1187 foreach_list_safe(n, &actual_parameters) {
1188 ir_rvalue *matrix = (ir_rvalue *) n;
1189
1190 if (!matrix->type->is_matrix())
1191 continue;
1192
1193 /* Create a temporary containing the matrix. */
1194 ir_variable *var = new(ctx) ir_variable(matrix->type, "matrix_tmp",
1195 ir_var_temporary);
1196 instructions->push_tail(var);
1197 instructions->push_tail(new(ctx) ir_assignment(new(ctx)
1198 ir_dereference_variable(var), matrix, NULL));
1199 var->constant_value = matrix->constant_expression_value();
1200
1201 /* Replace the matrix with dereferences of its columns. */
1202 for (int i = 0; i < matrix->type->matrix_columns; i++) {
1203 matrix->insert_before(new (ctx) ir_dereference_array(var,
1204 new(ctx) ir_constant(i)));
1205 }
1206 matrix->remove();
1207 }
1208 }
1209
1210 bool all_parameters_are_constant = true;
1211
1212 /* Type cast each parameter and, if possible, fold constants.*/
1213 foreach_list_safe(n, &actual_parameters) {
1214 ir_rvalue *ir = (ir_rvalue *) n;
1215
1216 const glsl_type *desired_type =
1217 glsl_type::get_instance(constructor_type->base_type,
1218 ir->type->vector_elements,
1219 ir->type->matrix_columns);
1220 ir_rvalue *result = convert_component(ir, desired_type);
1221
1222 /* Attempt to convert the parameter to a constant valued expression.
1223 * After doing so, track whether or not all the parameters to the
1224 * constructor are trivially constant valued expressions.
1225 */
1226 ir_rvalue *const constant = result->constant_expression_value();
1227
1228 if (constant != NULL)
1229 result = constant;
1230 else
1231 all_parameters_are_constant = false;
1232
1233 if (result != ir) {
1234 ir->replace_with(result);
1235 }
1236 }
1237
1238 /* If all of the parameters are trivially constant, create a
1239 * constant representing the complete collection of parameters.
1240 */
1241 if (all_parameters_are_constant) {
1242 if (components_used >= type_components)
1243 return new(ctx) ir_constant(constructor_type,
1244 & actual_parameters);
1245
1246 /* The above case must handle all scalar constructors.
1247 */
1248 assert(constructor_type->is_vector()
1249 || constructor_type->is_matrix());
1250
1251 /* Constructors with exactly one component are special for
1252 * vectors and matrices. For vectors it causes all elements of
1253 * the vector to be filled with the value. For matrices it
1254 * causes the matrix to be filled with 0 and the diagonal to be
1255 * filled with the value.
1256 */
1257 ir_constant_data data = { { 0 } };
1258 ir_constant *const initializer =
1259 (ir_constant *) actual_parameters.head;
1260 if (constructor_type->is_matrix())
1261 generate_constructor_matrix(constructor_type, initializer,
1262 &data);
1263 else
1264 generate_constructor_vector(constructor_type, initializer,
1265 &data);
1266
1267 return new(ctx) ir_constant(constructor_type, &data);
1268 } else if (constructor_type->is_scalar()) {
1269 return dereference_component((ir_rvalue *) actual_parameters.head,
1270 0);
1271 } else if (constructor_type->is_vector()) {
1272 return emit_inline_vector_constructor(constructor_type,
1273 instructions,
1274 &actual_parameters,
1275 ctx);
1276 } else {
1277 assert(constructor_type->is_matrix());
1278 return emit_inline_matrix_constructor(constructor_type,
1279 instructions,
1280 &actual_parameters,
1281 ctx);
1282 }
1283 } else {
1284 const ast_expression *id = subexpressions[0];
1285 YYLTYPE loc = id->get_location();
1286 exec_list actual_parameters;
1287
1288 process_parameters(instructions, &actual_parameters, &this->expressions,
1289 state);
1290
1291 const glsl_type *const type =
1292 state->symbols->get_type(id->primary_expression.identifier);
1293
1294 if ((type != NULL) && type->is_record()) {
1295 ir_constant *constant =
1296 constant_record_constructor(type, &loc, &actual_parameters, state);
1297
1298 if (constant != NULL)
1299 return constant;
1300 }
1301
1302 return match_function_by_name(instructions,
1303 id->primary_expression.identifier, & loc,
1304 &actual_parameters, state);
1305 }
1306
1307 return ir_call::get_error_instruction(ctx);
1308 }