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.
24 #include "glsl_symbol_table.h"
26 #include "glsl_types.h"
28 #include "main/core.h" /* for MIN2 */
31 convert_component(ir_rvalue *src, const glsl_type *desired_type);
34 apply_implicit_conversion(const glsl_type *to, ir_rvalue * &from,
35 struct _mesa_glsl_parse_state *state);
38 process_parameters(exec_list *instructions, exec_list *actual_parameters,
39 exec_list *parameters,
40 struct _mesa_glsl_parse_state *state)
44 foreach_list (n, parameters) {
45 ast_node *const ast = exec_node_data(ast_node, n, link);
46 ir_rvalue *result = ast->hir(instructions, state);
48 ir_constant *const constant = result->constant_expression_value();
52 actual_parameters->push_tail(result);
61 * Generate a source prototype for a function signature
63 * \param return_type Return type of the function. May be \c NULL.
64 * \param name Name of the function.
65 * \param parameters List of \c ir_instruction nodes representing the
66 * parameter list for the function. This may be either a
67 * formal (\c ir_variable) or actual (\c ir_rvalue)
68 * parameter list. Only the type is used.
71 * A ralloced string representing the prototype of the function.
74 prototype_string(const glsl_type *return_type, const char *name,
75 exec_list *parameters)
79 if (return_type != NULL)
80 str = ralloc_asprintf(NULL, "%s ", return_type->name);
82 ralloc_asprintf_append(&str, "%s(", name);
84 const char *comma = "";
85 foreach_list(node, parameters) {
86 const ir_variable *const param = (ir_variable *) node;
88 ralloc_asprintf_append(&str, "%s%s", comma, param->type->name);
92 ralloc_strcat(&str, ")");
97 * Verify that 'out' and 'inout' actual parameters are lvalues. Also, verify
98 * that 'const_in' formal parameters (an extension in our IR) correspond to
99 * ir_constant actual parameters.
102 verify_parameter_modes(_mesa_glsl_parse_state *state,
103 ir_function_signature *sig,
104 exec_list &actual_ir_parameters,
105 exec_list &actual_ast_parameters)
107 exec_node *actual_ir_node = actual_ir_parameters.head;
108 exec_node *actual_ast_node = actual_ast_parameters.head;
110 foreach_list(formal_node, &sig->parameters) {
111 /* The lists must be the same length. */
112 assert(!actual_ir_node->is_tail_sentinel());
113 assert(!actual_ast_node->is_tail_sentinel());
115 const ir_variable *const formal = (ir_variable *) formal_node;
116 const ir_rvalue *const actual = (ir_rvalue *) actual_ir_node;
117 const ast_expression *const actual_ast =
118 exec_node_data(ast_expression, actual_ast_node, link);
120 /* FIXME: 'loc' is incorrect (as of 2011-01-21). It is always
123 YYLTYPE loc = actual_ast->get_location();
125 /* Verify that 'const_in' parameters are ir_constants. */
126 if (formal->mode == ir_var_const_in &&
127 actual->ir_type != ir_type_constant) {
128 _mesa_glsl_error(&loc, state,
129 "parameter `in %s' must be a constant expression",
134 /* Verify that 'out' and 'inout' actual parameters are lvalues. */
135 if (formal->mode == ir_var_function_out
136 || formal->mode == ir_var_function_inout) {
137 const char *mode = NULL;
138 switch (formal->mode) {
139 case ir_var_function_out: mode = "out"; break;
140 case ir_var_function_inout: mode = "inout"; break;
141 default: assert(false); break;
144 /* This AST-based check catches errors like f(i++). The IR-based
145 * is_lvalue() is insufficient because the actual parameter at the
146 * IR-level is just a temporary value, which is an l-value.
148 if (actual_ast->non_lvalue_description != NULL) {
149 _mesa_glsl_error(&loc, state,
150 "function parameter '%s %s' references a %s",
152 actual_ast->non_lvalue_description);
156 ir_variable *var = actual->variable_referenced();
158 var->assigned = true;
160 if (var && var->read_only) {
161 _mesa_glsl_error(&loc, state,
162 "function parameter '%s %s' references the "
163 "read-only variable '%s'",
165 actual->variable_referenced()->name);
167 } else if (!actual->is_lvalue()) {
168 /* Even though ir_binop_vector_extract is not an l-value, let it
169 * slop through. generate_call will handle it correctly.
171 ir_expression *const expr = ((ir_rvalue *) actual)->as_expression();
173 || expr->operation != ir_binop_vector_extract
174 || !expr->operands[0]->is_lvalue()) {
175 _mesa_glsl_error(&loc, state,
176 "function parameter '%s %s' is not an lvalue",
183 actual_ir_node = actual_ir_node->next;
184 actual_ast_node = actual_ast_node->next;
190 fix_parameter(void *mem_ctx, ir_rvalue *actual, const glsl_type *formal_type,
191 exec_list *before_instructions, exec_list *after_instructions,
192 bool parameter_is_inout)
194 ir_expression *const expr = actual->as_expression();
196 /* If the types match exactly and the parameter is not a vector-extract,
197 * nothing needs to be done to fix the parameter.
199 if (formal_type == actual->type
200 && (expr == NULL || expr->operation != ir_binop_vector_extract))
203 /* To convert an out parameter, we need to create a temporary variable to
204 * hold the value before conversion, and then perform the conversion after
205 * the function call returns.
207 * This has the effect of transforming code like this:
213 * Into IR that's equivalent to this:
217 * int out_parameter_conversion;
218 * f(out_parameter_conversion);
219 * value = float(out_parameter_conversion);
221 * If the parameter is an ir_expression of ir_binop_vector_extract,
222 * additional conversion is needed in the post-call re-write.
225 new(mem_ctx) ir_variable(formal_type, "inout_tmp", ir_var_temporary);
227 before_instructions->push_tail(tmp);
229 /* If the parameter is an inout parameter, copy the value of the actual
230 * parameter to the new temporary. Note that no type conversion is allowed
231 * here because inout parameters must match types exactly.
233 if (parameter_is_inout) {
234 /* Inout parameters should never require conversion, since that would
235 * require an implicit conversion to exist both to and from the formal
236 * parameter type, and there are no bidirectional implicit conversions.
238 assert (actual->type == formal_type);
240 ir_dereference_variable *const deref_tmp_1 =
241 new(mem_ctx) ir_dereference_variable(tmp);
242 ir_assignment *const assignment =
243 new(mem_ctx) ir_assignment(deref_tmp_1, actual);
244 before_instructions->push_tail(assignment);
247 /* Replace the parameter in the call with a dereference of the new
250 ir_dereference_variable *const deref_tmp_2 =
251 new(mem_ctx) ir_dereference_variable(tmp);
252 actual->replace_with(deref_tmp_2);
255 /* Copy the temporary variable to the actual parameter with optional
256 * type conversion applied.
258 ir_rvalue *rhs = new(mem_ctx) ir_dereference_variable(tmp);
259 if (actual->type != formal_type)
260 rhs = convert_component(rhs, actual->type);
262 ir_rvalue *lhs = actual;
263 if (expr != NULL && expr->operation == ir_binop_vector_extract) {
264 rhs = new(mem_ctx) ir_expression(ir_triop_vector_insert,
265 expr->operands[0]->type,
266 expr->operands[0]->clone(mem_ctx, NULL),
268 expr->operands[1]->clone(mem_ctx, NULL));
269 lhs = expr->operands[0]->clone(mem_ctx, NULL);
272 ir_assignment *const assignment_2 = new(mem_ctx) ir_assignment(lhs, rhs);
273 after_instructions->push_tail(assignment_2);
277 * If a function call is generated, \c call_ir will point to it on exit.
278 * Otherwise \c call_ir will be set to \c NULL.
281 generate_call(exec_list *instructions, ir_function_signature *sig,
282 exec_list *actual_parameters,
284 struct _mesa_glsl_parse_state *state)
287 exec_list post_call_conversions;
291 /* Perform implicit conversion of arguments. For out parameters, we need
292 * to place them in a temporary variable and do the conversion after the
293 * call takes place. Since we haven't emitted the call yet, we'll place
294 * the post-call conversions in a temporary exec_list, and emit them later.
296 exec_list_iterator actual_iter = actual_parameters->iterator();
297 exec_list_iterator formal_iter = sig->parameters.iterator();
299 while (actual_iter.has_next()) {
300 ir_rvalue *actual = (ir_rvalue *) actual_iter.get();
301 ir_variable *formal = (ir_variable *) formal_iter.get();
303 assert(actual != NULL);
304 assert(formal != NULL);
306 if (formal->type->is_numeric() || formal->type->is_boolean()) {
307 switch (formal->mode) {
308 case ir_var_const_in:
309 case ir_var_function_in: {
311 = convert_component(actual, formal->type);
312 actual->replace_with(converted);
315 case ir_var_function_out:
316 case ir_var_function_inout:
317 fix_parameter(ctx, actual, formal->type,
318 instructions, &post_call_conversions,
319 formal->mode == ir_var_function_inout);
322 assert (!"Illegal formal parameter mode");
331 /* If the function call is a constant expression, don't generate any
332 * instructions; just generate an ir_constant.
334 * Function calls were first allowed to be constant expressions in GLSL
335 * 1.20 and GLSL ES 3.00.
337 if (state->is_version(120, 300)) {
338 ir_constant *value = sig->constant_expression_value(actual_parameters, NULL);
344 ir_dereference_variable *deref = NULL;
345 if (!sig->return_type->is_void()) {
346 /* Create a new temporary to hold the return value. */
349 var = new(ctx) ir_variable(sig->return_type,
350 ralloc_asprintf(ctx, "%s_retval",
351 sig->function_name()),
353 instructions->push_tail(var);
355 deref = new(ctx) ir_dereference_variable(var);
357 ir_call *call = new(ctx) ir_call(sig, deref, actual_parameters);
358 instructions->push_tail(call);
360 /* Also emit any necessary out-parameter conversions. */
361 instructions->append_list(&post_call_conversions);
363 return deref ? deref->clone(ctx, NULL) : NULL;
367 * Given a function name and parameter list, find the matching signature.
369 static ir_function_signature *
370 match_function_by_name(const char *name,
371 exec_list *actual_parameters,
372 struct _mesa_glsl_parse_state *state)
375 ir_function *f = state->symbols->get_function(name);
376 ir_function_signature *local_sig = NULL;
377 ir_function_signature *sig = NULL;
379 /* Is the function hidden by a record type constructor? */
380 if (state->symbols->get_type(name))
381 goto done; /* no match */
383 /* Is the function hidden by a variable (impossible in 1.10)? */
384 if (!state->symbols->separate_function_namespace
385 && state->symbols->get_variable(name))
386 goto done; /* no match */
389 /* Look for a match in the local shader. If exact, we're done. */
390 bool is_exact = false;
391 sig = local_sig = f->matching_signature(actual_parameters, &is_exact);
395 if (!state->es_shader && f->has_user_signature()) {
396 /* In desktop GL, the presence of a user-defined signature hides any
397 * built-in signatures, so we must ignore them. In contrast, in ES2
398 * user-defined signatures add new overloads, so we must proceed.
404 /* Local shader has no exact candidates; check the built-ins. */
405 _mesa_glsl_initialize_functions(state);
406 for (unsigned i = 0; i < state->num_builtins_to_link; i++) {
407 ir_function *builtin =
408 state->builtins_to_link[i]->symbols->get_function(name);
412 bool is_exact = false;
413 ir_function_signature *builtin_sig =
414 builtin->matching_signature(actual_parameters, &is_exact);
416 if (builtin_sig == NULL)
419 /* If the built-in signature is exact, we can stop. */
426 /* We found an inexact match, which is better than nothing. However,
427 * we should keep searching for an exact match.
435 /* If the match is from a linked built-in shader, import the prototype. */
436 if (sig != local_sig) {
438 f = new(ctx) ir_function(name);
439 state->symbols->add_global_function(f);
440 emit_function(state, f);
442 f->add_signature(sig->clone_prototype(f, NULL));
449 * Raise a "no matching function" error, listing all possible overloads the
450 * compiler considered so developers can figure out what went wrong.
453 no_matching_function_error(const char *name,
455 exec_list *actual_parameters,
456 _mesa_glsl_parse_state *state)
458 char *str = prototype_string(NULL, name, actual_parameters);
459 _mesa_glsl_error(loc, state, "no matching function for call to `%s'", str);
462 const char *prefix = "candidates are: ";
464 for (int i = -1; i < (int) state->num_builtins_to_link; i++) {
465 glsl_symbol_table *syms = i >= 0 ? state->builtins_to_link[i]->symbols
467 ir_function *f = syms->get_function(name);
471 foreach_list (node, &f->signatures) {
472 ir_function_signature *sig = (ir_function_signature *) node;
474 str = prototype_string(sig->return_type, f->name, &sig->parameters);
475 _mesa_glsl_error(loc, state, "%s%s", prefix, str);
484 * Perform automatic type conversion of constructor parameters
486 * This implements the rules in the "Conversion and Scalar Constructors"
487 * section (GLSL 1.10 section 5.4.1), not the "Implicit Conversions" rules.
490 convert_component(ir_rvalue *src, const glsl_type *desired_type)
492 void *ctx = ralloc_parent(src);
493 const unsigned a = desired_type->base_type;
494 const unsigned b = src->type->base_type;
495 ir_expression *result = NULL;
497 if (src->type->is_error())
500 assert(a <= GLSL_TYPE_BOOL);
501 assert(b <= GLSL_TYPE_BOOL);
510 result = new(ctx) ir_expression(ir_unop_i2u, src);
512 case GLSL_TYPE_FLOAT:
513 result = new(ctx) ir_expression(ir_unop_f2u, src);
516 result = new(ctx) ir_expression(ir_unop_i2u,
517 new(ctx) ir_expression(ir_unop_b2i, src));
524 result = new(ctx) ir_expression(ir_unop_u2i, src);
526 case GLSL_TYPE_FLOAT:
527 result = new(ctx) ir_expression(ir_unop_f2i, src);
530 result = new(ctx) ir_expression(ir_unop_b2i, src);
534 case GLSL_TYPE_FLOAT:
537 result = new(ctx) ir_expression(ir_unop_u2f, desired_type, src, NULL);
540 result = new(ctx) ir_expression(ir_unop_i2f, desired_type, src, NULL);
543 result = new(ctx) ir_expression(ir_unop_b2f, desired_type, src, NULL);
550 result = new(ctx) ir_expression(ir_unop_i2b,
551 new(ctx) ir_expression(ir_unop_u2i, src));
554 result = new(ctx) ir_expression(ir_unop_i2b, desired_type, src, NULL);
556 case GLSL_TYPE_FLOAT:
557 result = new(ctx) ir_expression(ir_unop_f2b, desired_type, src, NULL);
563 assert(result != NULL);
564 assert(result->type == desired_type);
566 /* Try constant folding; it may fold in the conversion we just added. */
567 ir_constant *const constant = result->constant_expression_value();
568 return (constant != NULL) ? (ir_rvalue *) constant : (ir_rvalue *) result;
572 * Dereference a specific component from a scalar, vector, or matrix
575 dereference_component(ir_rvalue *src, unsigned component)
577 void *ctx = ralloc_parent(src);
578 assert(component < src->type->components());
580 /* If the source is a constant, just create a new constant instead of a
581 * dereference of the existing constant.
583 ir_constant *constant = src->as_constant();
585 return new(ctx) ir_constant(constant, component);
587 if (src->type->is_scalar()) {
589 } else if (src->type->is_vector()) {
590 return new(ctx) ir_swizzle(src, component, 0, 0, 0, 1);
592 assert(src->type->is_matrix());
594 /* Dereference a row of the matrix, then call this function again to get
595 * a specific element from that row.
597 const int c = component / src->type->column_type()->vector_elements;
598 const int r = component % src->type->column_type()->vector_elements;
599 ir_constant *const col_index = new(ctx) ir_constant(c);
600 ir_dereference *const col = new(ctx) ir_dereference_array(src, col_index);
602 col->type = src->type->column_type();
604 return dereference_component(col, r);
607 assert(!"Should not get here.");
613 process_array_constructor(exec_list *instructions,
614 const glsl_type *constructor_type,
615 YYLTYPE *loc, exec_list *parameters,
616 struct _mesa_glsl_parse_state *state)
619 /* Array constructors come in two forms: sized and unsized. Sized array
620 * constructors look like 'vec4[2](a, b)', where 'a' and 'b' are vec4
621 * variables. In this case the number of parameters must exactly match the
622 * specified size of the array.
624 * Unsized array constructors look like 'vec4[](a, b)', where 'a' and 'b'
625 * are vec4 variables. In this case the size of the array being constructed
626 * is determined by the number of parameters.
628 * From page 52 (page 58 of the PDF) of the GLSL 1.50 spec:
630 * "There must be exactly the same number of arguments as the size of
631 * the array being constructed. If no size is present in the
632 * constructor, then the array is explicitly sized to the number of
633 * arguments provided. The arguments are assigned in order, starting at
634 * element 0, to the elements of the constructed array. Each argument
635 * must be the same type as the element type of the array, or be a type
636 * that can be converted to the element type of the array according to
637 * Section 4.1.10 "Implicit Conversions.""
639 exec_list actual_parameters;
640 const unsigned parameter_count =
641 process_parameters(instructions, &actual_parameters, parameters, state);
643 if ((parameter_count == 0)
644 || ((constructor_type->length != 0)
645 && (constructor_type->length != parameter_count))) {
646 const unsigned min_param = (constructor_type->length == 0)
647 ? 1 : constructor_type->length;
649 _mesa_glsl_error(loc, state, "array constructor must have %s %u "
651 (constructor_type->length != 0) ? "at least" : "exactly",
652 min_param, (min_param <= 1) ? "" : "s");
653 return ir_rvalue::error_value(ctx);
656 if (constructor_type->length == 0) {
658 glsl_type::get_array_instance(constructor_type->element_type(),
660 assert(constructor_type != NULL);
661 assert(constructor_type->length == parameter_count);
664 bool all_parameters_are_constant = true;
666 /* Type cast each parameter and, if possible, fold constants. */
667 foreach_list_safe(n, &actual_parameters) {
668 ir_rvalue *ir = (ir_rvalue *) n;
669 ir_rvalue *result = ir;
671 /* Apply implicit conversions (not the scalar constructor rules!). See
672 * the spec quote above. */
673 if (constructor_type->element_type()->is_float()) {
674 const glsl_type *desired_type =
675 glsl_type::get_instance(GLSL_TYPE_FLOAT,
676 ir->type->vector_elements,
677 ir->type->matrix_columns);
678 if (result->type->can_implicitly_convert_to(desired_type)) {
679 /* Even though convert_component() implements the constructor
680 * conversion rules (not the implicit conversion rules), its safe
681 * to use it here because we already checked that the implicit
682 * conversion is legal.
684 result = convert_component(ir, desired_type);
688 if (result->type != constructor_type->element_type()) {
689 _mesa_glsl_error(loc, state, "type error in array constructor: "
690 "expected: %s, found %s",
691 constructor_type->element_type()->name,
695 /* Attempt to convert the parameter to a constant valued expression.
696 * After doing so, track whether or not all the parameters to the
697 * constructor are trivially constant valued expressions.
699 ir_rvalue *const constant = result->constant_expression_value();
701 if (constant != NULL)
704 all_parameters_are_constant = false;
706 ir->replace_with(result);
709 if (all_parameters_are_constant)
710 return new(ctx) ir_constant(constructor_type, &actual_parameters);
712 ir_variable *var = new(ctx) ir_variable(constructor_type, "array_ctor",
714 instructions->push_tail(var);
717 foreach_list(node, &actual_parameters) {
718 ir_rvalue *rhs = (ir_rvalue *) node;
719 ir_rvalue *lhs = new(ctx) ir_dereference_array(var,
720 new(ctx) ir_constant(i));
722 ir_instruction *assignment = new(ctx) ir_assignment(lhs, rhs, NULL);
723 instructions->push_tail(assignment);
728 return new(ctx) ir_dereference_variable(var);
733 * Try to convert a record constructor to a constant expression
736 constant_record_constructor(const glsl_type *constructor_type,
737 exec_list *parameters, void *mem_ctx)
739 foreach_list(node, parameters) {
740 ir_constant *constant = ((ir_instruction *) node)->as_constant();
741 if (constant == NULL)
743 node->replace_with(constant);
746 return new(mem_ctx) ir_constant(constructor_type, parameters);
751 * Determine if a list consists of a single scalar r-value
754 single_scalar_parameter(exec_list *parameters)
756 const ir_rvalue *const p = (ir_rvalue *) parameters->head;
757 assert(((ir_rvalue *)p)->as_rvalue() != NULL);
759 return (p->type->is_scalar() && p->next->is_tail_sentinel());
764 * Generate inline code for a vector constructor
766 * The generated constructor code will consist of a temporary variable
767 * declaration of the same type as the constructor. A sequence of assignments
768 * from constructor parameters to the temporary will follow.
771 * An \c ir_dereference_variable of the temprorary generated in the constructor
775 emit_inline_vector_constructor(const glsl_type *type,
776 exec_list *instructions,
777 exec_list *parameters,
780 assert(!parameters->is_empty());
782 ir_variable *var = new(ctx) ir_variable(type, "vec_ctor", ir_var_temporary);
783 instructions->push_tail(var);
785 /* There are two kinds of vector constructors.
787 * - Construct a vector from a single scalar by replicating that scalar to
788 * all components of the vector.
790 * - Construct a vector from an arbirary combination of vectors and
791 * scalars. The components of the constructor parameters are assigned
792 * to the vector in order until the vector is full.
794 const unsigned lhs_components = type->components();
795 if (single_scalar_parameter(parameters)) {
796 ir_rvalue *first_param = (ir_rvalue *)parameters->head;
797 ir_rvalue *rhs = new(ctx) ir_swizzle(first_param, 0, 0, 0, 0,
799 ir_dereference_variable *lhs = new(ctx) ir_dereference_variable(var);
800 const unsigned mask = (1U << lhs_components) - 1;
802 assert(rhs->type == lhs->type);
804 ir_instruction *inst = new(ctx) ir_assignment(lhs, rhs, NULL, mask);
805 instructions->push_tail(inst);
807 unsigned base_component = 0;
808 unsigned base_lhs_component = 0;
809 ir_constant_data data;
810 unsigned constant_mask = 0, constant_components = 0;
812 memset(&data, 0, sizeof(data));
814 foreach_list(node, parameters) {
815 ir_rvalue *param = (ir_rvalue *) node;
816 unsigned rhs_components = param->type->components();
818 /* Do not try to assign more components to the vector than it has!
820 if ((rhs_components + base_lhs_component) > lhs_components) {
821 rhs_components = lhs_components - base_lhs_component;
824 const ir_constant *const c = param->as_constant();
826 for (unsigned i = 0; i < rhs_components; i++) {
827 switch (c->type->base_type) {
829 data.u[i + base_component] = c->get_uint_component(i);
832 data.i[i + base_component] = c->get_int_component(i);
834 case GLSL_TYPE_FLOAT:
835 data.f[i + base_component] = c->get_float_component(i);
838 data.b[i + base_component] = c->get_bool_component(i);
841 assert(!"Should not get here.");
846 /* Mask of fields to be written in the assignment.
848 constant_mask |= ((1U << rhs_components) - 1) << base_lhs_component;
849 constant_components += rhs_components;
851 base_component += rhs_components;
853 /* Advance the component index by the number of components
854 * that were just assigned.
856 base_lhs_component += rhs_components;
859 if (constant_mask != 0) {
860 ir_dereference *lhs = new(ctx) ir_dereference_variable(var);
861 const glsl_type *rhs_type = glsl_type::get_instance(var->type->base_type,
864 ir_rvalue *rhs = new(ctx) ir_constant(rhs_type, &data);
866 ir_instruction *inst =
867 new(ctx) ir_assignment(lhs, rhs, NULL, constant_mask);
868 instructions->push_tail(inst);
872 foreach_list(node, parameters) {
873 ir_rvalue *param = (ir_rvalue *) node;
874 unsigned rhs_components = param->type->components();
876 /* Do not try to assign more components to the vector than it has!
878 if ((rhs_components + base_component) > lhs_components) {
879 rhs_components = lhs_components - base_component;
882 const ir_constant *const c = param->as_constant();
884 /* Mask of fields to be written in the assignment.
886 const unsigned write_mask = ((1U << rhs_components) - 1)
889 ir_dereference *lhs = new(ctx) ir_dereference_variable(var);
891 /* Generate a swizzle so that LHS and RHS sizes match.
894 new(ctx) ir_swizzle(param, 0, 1, 2, 3, rhs_components);
896 ir_instruction *inst =
897 new(ctx) ir_assignment(lhs, rhs, NULL, write_mask);
898 instructions->push_tail(inst);
901 /* Advance the component index by the number of components that were
904 base_component += rhs_components;
907 return new(ctx) ir_dereference_variable(var);
912 * Generate assignment of a portion of a vector to a portion of a matrix column
914 * \param src_base First component of the source to be used in assignment
915 * \param column Column of destination to be assiged
916 * \param row_base First component of the destination column to be assigned
917 * \param count Number of components to be assigned
920 * \c src_base + \c count must be less than or equal to the number of components
921 * in the source vector.
924 assign_to_matrix_column(ir_variable *var, unsigned column, unsigned row_base,
925 ir_rvalue *src, unsigned src_base, unsigned count,
928 ir_constant *col_idx = new(mem_ctx) ir_constant(column);
929 ir_dereference *column_ref = new(mem_ctx) ir_dereference_array(var, col_idx);
931 assert(column_ref->type->components() >= (row_base + count));
932 assert(src->type->components() >= (src_base + count));
934 /* Generate a swizzle that extracts the number of components from the source
935 * that are to be assigned to the column of the matrix.
937 if (count < src->type->vector_elements) {
938 src = new(mem_ctx) ir_swizzle(src,
939 src_base + 0, src_base + 1,
940 src_base + 2, src_base + 3,
944 /* Mask of fields to be written in the assignment.
946 const unsigned write_mask = ((1U << count) - 1) << row_base;
948 return new(mem_ctx) ir_assignment(column_ref, src, NULL, write_mask);
953 * Generate inline code for a matrix constructor
955 * The generated constructor code will consist of a temporary variable
956 * declaration of the same type as the constructor. A sequence of assignments
957 * from constructor parameters to the temporary will follow.
960 * An \c ir_dereference_variable of the temprorary generated in the constructor
964 emit_inline_matrix_constructor(const glsl_type *type,
965 exec_list *instructions,
966 exec_list *parameters,
969 assert(!parameters->is_empty());
971 ir_variable *var = new(ctx) ir_variable(type, "mat_ctor", ir_var_temporary);
972 instructions->push_tail(var);
974 /* There are three kinds of matrix constructors.
976 * - Construct a matrix from a single scalar by replicating that scalar to
977 * along the diagonal of the matrix and setting all other components to
980 * - Construct a matrix from an arbirary combination of vectors and
981 * scalars. The components of the constructor parameters are assigned
982 * to the matrix in colum-major order until the matrix is full.
984 * - Construct a matrix from a single matrix. The source matrix is copied
985 * to the upper left portion of the constructed matrix, and the remaining
986 * elements take values from the identity matrix.
988 ir_rvalue *const first_param = (ir_rvalue *) parameters->head;
989 if (single_scalar_parameter(parameters)) {
990 /* Assign the scalar to the X component of a vec4, and fill the remaining
991 * components with zero.
993 ir_variable *rhs_var =
994 new(ctx) ir_variable(glsl_type::vec4_type, "mat_ctor_vec",
996 instructions->push_tail(rhs_var);
998 ir_constant_data zero;
1004 ir_instruction *inst =
1005 new(ctx) ir_assignment(new(ctx) ir_dereference_variable(rhs_var),
1006 new(ctx) ir_constant(rhs_var->type, &zero),
1008 instructions->push_tail(inst);
1010 ir_dereference *const rhs_ref = new(ctx) ir_dereference_variable(rhs_var);
1012 inst = new(ctx) ir_assignment(rhs_ref, first_param, NULL, 0x01);
1013 instructions->push_tail(inst);
1015 /* Assign the temporary vector to each column of the destination matrix
1016 * with a swizzle that puts the X component on the diagonal of the
1017 * matrix. In some cases this may mean that the X component does not
1018 * get assigned into the column at all (i.e., when the matrix has more
1019 * columns than rows).
1021 static const unsigned rhs_swiz[4][4] = {
1028 const unsigned cols_to_init = MIN2(type->matrix_columns,
1029 type->vector_elements);
1030 for (unsigned i = 0; i < cols_to_init; i++) {
1031 ir_constant *const col_idx = new(ctx) ir_constant(i);
1032 ir_rvalue *const col_ref = new(ctx) ir_dereference_array(var, col_idx);
1034 ir_rvalue *const rhs_ref = new(ctx) ir_dereference_variable(rhs_var);
1035 ir_rvalue *const rhs = new(ctx) ir_swizzle(rhs_ref, rhs_swiz[i],
1036 type->vector_elements);
1038 inst = new(ctx) ir_assignment(col_ref, rhs, NULL);
1039 instructions->push_tail(inst);
1042 for (unsigned i = cols_to_init; i < type->matrix_columns; i++) {
1043 ir_constant *const col_idx = new(ctx) ir_constant(i);
1044 ir_rvalue *const col_ref = new(ctx) ir_dereference_array(var, col_idx);
1046 ir_rvalue *const rhs_ref = new(ctx) ir_dereference_variable(rhs_var);
1047 ir_rvalue *const rhs = new(ctx) ir_swizzle(rhs_ref, 1, 1, 1, 1,
1048 type->vector_elements);
1050 inst = new(ctx) ir_assignment(col_ref, rhs, NULL);
1051 instructions->push_tail(inst);
1053 } else if (first_param->type->is_matrix()) {
1054 /* From page 50 (56 of the PDF) of the GLSL 1.50 spec:
1056 * "If a matrix is constructed from a matrix, then each component
1057 * (column i, row j) in the result that has a corresponding
1058 * component (column i, row j) in the argument will be initialized
1059 * from there. All other components will be initialized to the
1060 * identity matrix. If a matrix argument is given to a matrix
1061 * constructor, it is an error to have any other arguments."
1063 assert(first_param->next->is_tail_sentinel());
1064 ir_rvalue *const src_matrix = first_param;
1066 /* If the source matrix is smaller, pre-initialize the relavent parts of
1067 * the destination matrix to the identity matrix.
1069 if ((src_matrix->type->matrix_columns < var->type->matrix_columns)
1070 || (src_matrix->type->vector_elements < var->type->vector_elements)) {
1072 /* If the source matrix has fewer rows, every column of the destination
1073 * must be initialized. Otherwise only the columns in the destination
1074 * that do not exist in the source must be initialized.
1077 (src_matrix->type->vector_elements < var->type->vector_elements)
1078 ? 0 : src_matrix->type->matrix_columns;
1080 const glsl_type *const col_type = var->type->column_type();
1081 for (/* empty */; col < var->type->matrix_columns; col++) {
1082 ir_constant_data ident;
1091 ir_rvalue *const rhs = new(ctx) ir_constant(col_type, &ident);
1093 ir_rvalue *const lhs =
1094 new(ctx) ir_dereference_array(var, new(ctx) ir_constant(col));
1096 ir_instruction *inst = new(ctx) ir_assignment(lhs, rhs, NULL);
1097 instructions->push_tail(inst);
1101 /* Assign columns from the source matrix to the destination matrix.
1103 * Since the parameter will be used in the RHS of multiple assignments,
1104 * generate a temporary and copy the paramter there.
1106 ir_variable *const rhs_var =
1107 new(ctx) ir_variable(first_param->type, "mat_ctor_mat",
1109 instructions->push_tail(rhs_var);
1111 ir_dereference *const rhs_var_ref =
1112 new(ctx) ir_dereference_variable(rhs_var);
1113 ir_instruction *const inst =
1114 new(ctx) ir_assignment(rhs_var_ref, first_param, NULL);
1115 instructions->push_tail(inst);
1117 const unsigned last_row = MIN2(src_matrix->type->vector_elements,
1118 var->type->vector_elements);
1119 const unsigned last_col = MIN2(src_matrix->type->matrix_columns,
1120 var->type->matrix_columns);
1122 unsigned swiz[4] = { 0, 0, 0, 0 };
1123 for (unsigned i = 1; i < last_row; i++)
1126 const unsigned write_mask = (1U << last_row) - 1;
1128 for (unsigned i = 0; i < last_col; i++) {
1129 ir_dereference *const lhs =
1130 new(ctx) ir_dereference_array(var, new(ctx) ir_constant(i));
1131 ir_rvalue *const rhs_col =
1132 new(ctx) ir_dereference_array(rhs_var, new(ctx) ir_constant(i));
1134 /* If one matrix has columns that are smaller than the columns of the
1135 * other matrix, wrap the column access of the larger with a swizzle
1136 * so that the LHS and RHS of the assignment have the same size (and
1137 * therefore have the same type).
1139 * It would be perfectly valid to unconditionally generate the
1140 * swizzles, this this will typically result in a more compact IR tree.
1143 if (lhs->type->vector_elements != rhs_col->type->vector_elements) {
1144 rhs = new(ctx) ir_swizzle(rhs_col, swiz, last_row);
1149 ir_instruction *inst =
1150 new(ctx) ir_assignment(lhs, rhs, NULL, write_mask);
1151 instructions->push_tail(inst);
1154 const unsigned cols = type->matrix_columns;
1155 const unsigned rows = type->vector_elements;
1156 unsigned col_idx = 0;
1157 unsigned row_idx = 0;
1159 foreach_list (node, parameters) {
1160 ir_rvalue *const rhs = (ir_rvalue *) node;
1161 const unsigned components_remaining_this_column = rows - row_idx;
1162 unsigned rhs_components = rhs->type->components();
1163 unsigned rhs_base = 0;
1165 /* Since the parameter might be used in the RHS of two assignments,
1166 * generate a temporary and copy the paramter there.
1168 ir_variable *rhs_var =
1169 new(ctx) ir_variable(rhs->type, "mat_ctor_vec", ir_var_temporary);
1170 instructions->push_tail(rhs_var);
1172 ir_dereference *rhs_var_ref =
1173 new(ctx) ir_dereference_variable(rhs_var);
1174 ir_instruction *inst = new(ctx) ir_assignment(rhs_var_ref, rhs, NULL);
1175 instructions->push_tail(inst);
1177 /* Assign the current parameter to as many components of the matrix
1180 * NOTE: A single vector parameter can span two matrix columns. A
1181 * single vec4, for example, can completely fill a mat2.
1183 if (rhs_components >= components_remaining_this_column) {
1184 const unsigned count = MIN2(rhs_components,
1185 components_remaining_this_column);
1187 rhs_var_ref = new(ctx) ir_dereference_variable(rhs_var);
1189 ir_instruction *inst = assign_to_matrix_column(var, col_idx,
1193 instructions->push_tail(inst);
1201 /* If there is data left in the parameter and components left to be
1202 * set in the destination, emit another assignment. It is possible
1203 * that the assignment could be of a vec4 to the last element of the
1204 * matrix. In this case col_idx==cols, but there is still data
1205 * left in the source parameter. Obviously, don't emit an assignment
1206 * to data outside the destination matrix.
1208 if ((col_idx < cols) && (rhs_base < rhs_components)) {
1209 const unsigned count = rhs_components - rhs_base;
1211 rhs_var_ref = new(ctx) ir_dereference_variable(rhs_var);
1213 ir_instruction *inst = assign_to_matrix_column(var, col_idx,
1218 instructions->push_tail(inst);
1225 return new(ctx) ir_dereference_variable(var);
1230 emit_inline_record_constructor(const glsl_type *type,
1231 exec_list *instructions,
1232 exec_list *parameters,
1235 ir_variable *const var =
1236 new(mem_ctx) ir_variable(type, "record_ctor", ir_var_temporary);
1237 ir_dereference_variable *const d = new(mem_ctx) ir_dereference_variable(var);
1239 instructions->push_tail(var);
1241 exec_node *node = parameters->head;
1242 for (unsigned i = 0; i < type->length; i++) {
1243 assert(!node->is_tail_sentinel());
1245 ir_dereference *const lhs =
1246 new(mem_ctx) ir_dereference_record(d->clone(mem_ctx, NULL),
1247 type->fields.structure[i].name);
1249 ir_rvalue *const rhs = ((ir_instruction *) node)->as_rvalue();
1250 assert(rhs != NULL);
1252 ir_instruction *const assign = new(mem_ctx) ir_assignment(lhs, rhs, NULL);
1254 instructions->push_tail(assign);
1263 ast_function_expression::hir(exec_list *instructions,
1264 struct _mesa_glsl_parse_state *state)
1267 /* There are three sorts of function calls.
1269 * 1. constructors - The first subexpression is an ast_type_specifier.
1270 * 2. methods - Only the .length() method of array types.
1271 * 3. functions - Calls to regular old functions.
1273 * Method calls are actually detected when the ast_field_selection
1274 * expression is handled.
1276 if (is_constructor()) {
1277 const ast_type_specifier *type = (ast_type_specifier *) subexpressions[0];
1278 YYLTYPE loc = type->get_location();
1281 const glsl_type *const constructor_type = type->glsl_type(& name, state);
1283 /* constructor_type can be NULL if a variable with the same name as the
1284 * structure has come into scope.
1286 if (constructor_type == NULL) {
1287 _mesa_glsl_error(& loc, state, "unknown type `%s' (structure name "
1288 "may be shadowed by a variable with the same name)",
1290 return ir_rvalue::error_value(ctx);
1294 /* Constructors for samplers are illegal.
1296 if (constructor_type->is_sampler()) {
1297 _mesa_glsl_error(& loc, state, "cannot construct sampler type `%s'",
1298 constructor_type->name);
1299 return ir_rvalue::error_value(ctx);
1302 if (constructor_type->is_array()) {
1303 if (!state->check_version(120, 300, &loc,
1304 "array constructors forbidden")) {
1305 return ir_rvalue::error_value(ctx);
1308 return process_array_constructor(instructions, constructor_type,
1309 & loc, &this->expressions, state);
1313 /* There are two kinds of constructor call. Constructors for built-in
1314 * language types, such as mat4 and vec2, are free form. The only
1315 * requirement is that the parameters must provide enough values of the
1316 * correct scalar type. Constructors for arrays and structures must
1317 * have the exact number of parameters with matching types in the
1318 * correct order. These constructors follow essentially the same type
1319 * matching rules as functions.
1321 if (constructor_type->is_record()) {
1322 exec_list actual_parameters;
1324 process_parameters(instructions, &actual_parameters,
1325 &this->expressions, state);
1327 exec_node *node = actual_parameters.head;
1328 for (unsigned i = 0; i < constructor_type->length; i++) {
1329 ir_rvalue *ir = (ir_rvalue *) node;
1331 if (node->is_tail_sentinel()) {
1332 _mesa_glsl_error(&loc, state,
1333 "insufficient parameters to constructor "
1335 constructor_type->name);
1336 return ir_rvalue::error_value(ctx);
1339 if (apply_implicit_conversion(constructor_type->fields.structure[i].type,
1341 node->replace_with(ir);
1343 _mesa_glsl_error(&loc, state,
1344 "parameter type mismatch in constructor "
1345 "for `%s.%s' (%s vs %s)",
1346 constructor_type->name,
1347 constructor_type->fields.structure[i].name,
1349 constructor_type->fields.structure[i].type->name);
1350 return ir_rvalue::error_value(ctx);;
1356 if (!node->is_tail_sentinel()) {
1357 _mesa_glsl_error(&loc, state, "too many parameters in constructor "
1358 "for `%s'", constructor_type->name);
1359 return ir_rvalue::error_value(ctx);
1362 ir_rvalue *const constant =
1363 constant_record_constructor(constructor_type, &actual_parameters,
1366 return (constant != NULL)
1368 : emit_inline_record_constructor(constructor_type, instructions,
1369 &actual_parameters, state);
1372 if (!constructor_type->is_numeric() && !constructor_type->is_boolean())
1373 return ir_rvalue::error_value(ctx);
1375 /* Total number of components of the type being constructed. */
1376 const unsigned type_components = constructor_type->components();
1378 /* Number of components from parameters that have actually been
1379 * consumed. This is used to perform several kinds of error checking.
1381 unsigned components_used = 0;
1383 unsigned matrix_parameters = 0;
1384 unsigned nonmatrix_parameters = 0;
1385 exec_list actual_parameters;
1387 foreach_list (n, &this->expressions) {
1388 ast_node *ast = exec_node_data(ast_node, n, link);
1389 ir_rvalue *result = ast->hir(instructions, state)->as_rvalue();
1391 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
1393 * "It is an error to provide extra arguments beyond this
1394 * last used argument."
1396 if (components_used >= type_components) {
1397 _mesa_glsl_error(& loc, state, "too many parameters to `%s' "
1399 constructor_type->name);
1400 return ir_rvalue::error_value(ctx);
1403 if (!result->type->is_numeric() && !result->type->is_boolean()) {
1404 _mesa_glsl_error(& loc, state, "cannot construct `%s' from a "
1405 "non-numeric data type",
1406 constructor_type->name);
1407 return ir_rvalue::error_value(ctx);
1410 /* Count the number of matrix and nonmatrix parameters. This
1411 * is used below to enforce some of the constructor rules.
1413 if (result->type->is_matrix())
1414 matrix_parameters++;
1416 nonmatrix_parameters++;
1418 actual_parameters.push_tail(result);
1419 components_used += result->type->components();
1422 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
1424 * "It is an error to construct matrices from other matrices. This
1425 * is reserved for future use."
1427 if (matrix_parameters > 0
1428 && constructor_type->is_matrix()
1429 && !state->check_version(120, 100, &loc,
1430 "cannot construct `%s' from a matrix",
1431 constructor_type->name)) {
1432 return ir_rvalue::error_value(ctx);
1435 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
1437 * "If a matrix argument is given to a matrix constructor, it is
1438 * an error to have any other arguments."
1440 if ((matrix_parameters > 0)
1441 && ((matrix_parameters + nonmatrix_parameters) > 1)
1442 && constructor_type->is_matrix()) {
1443 _mesa_glsl_error(& loc, state, "for matrix `%s' constructor, "
1444 "matrix must be only parameter",
1445 constructor_type->name);
1446 return ir_rvalue::error_value(ctx);
1449 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
1451 * "In these cases, there must be enough components provided in the
1452 * arguments to provide an initializer for every component in the
1453 * constructed value."
1455 if (components_used < type_components && components_used != 1
1456 && matrix_parameters == 0) {
1457 _mesa_glsl_error(& loc, state, "too few components to construct "
1459 constructor_type->name);
1460 return ir_rvalue::error_value(ctx);
1463 /* Later, we cast each parameter to the same base type as the
1464 * constructor. Since there are no non-floating point matrices, we
1465 * need to break them up into a series of column vectors.
1467 if (constructor_type->base_type != GLSL_TYPE_FLOAT) {
1468 foreach_list_safe(n, &actual_parameters) {
1469 ir_rvalue *matrix = (ir_rvalue *) n;
1471 if (!matrix->type->is_matrix())
1474 /* Create a temporary containing the matrix. */
1475 ir_variable *var = new(ctx) ir_variable(matrix->type, "matrix_tmp",
1477 instructions->push_tail(var);
1478 instructions->push_tail(new(ctx) ir_assignment(new(ctx)
1479 ir_dereference_variable(var), matrix, NULL));
1480 var->constant_value = matrix->constant_expression_value();
1482 /* Replace the matrix with dereferences of its columns. */
1483 for (int i = 0; i < matrix->type->matrix_columns; i++) {
1484 matrix->insert_before(new (ctx) ir_dereference_array(var,
1485 new(ctx) ir_constant(i)));
1491 bool all_parameters_are_constant = true;
1493 /* Type cast each parameter and, if possible, fold constants.*/
1494 foreach_list_safe(n, &actual_parameters) {
1495 ir_rvalue *ir = (ir_rvalue *) n;
1497 const glsl_type *desired_type =
1498 glsl_type::get_instance(constructor_type->base_type,
1499 ir->type->vector_elements,
1500 ir->type->matrix_columns);
1501 ir_rvalue *result = convert_component(ir, desired_type);
1503 /* Attempt to convert the parameter to a constant valued expression.
1504 * After doing so, track whether or not all the parameters to the
1505 * constructor are trivially constant valued expressions.
1507 ir_rvalue *const constant = result->constant_expression_value();
1509 if (constant != NULL)
1512 all_parameters_are_constant = false;
1515 ir->replace_with(result);
1519 /* If all of the parameters are trivially constant, create a
1520 * constant representing the complete collection of parameters.
1522 if (all_parameters_are_constant) {
1523 return new(ctx) ir_constant(constructor_type, &actual_parameters);
1524 } else if (constructor_type->is_scalar()) {
1525 return dereference_component((ir_rvalue *) actual_parameters.head,
1527 } else if (constructor_type->is_vector()) {
1528 return emit_inline_vector_constructor(constructor_type,
1533 assert(constructor_type->is_matrix());
1534 return emit_inline_matrix_constructor(constructor_type,
1540 const ast_expression *id = subexpressions[0];
1541 const char *func_name = id->primary_expression.identifier;
1542 YYLTYPE loc = id->get_location();
1543 exec_list actual_parameters;
1545 process_parameters(instructions, &actual_parameters, &this->expressions,
1548 ir_function_signature *sig =
1549 match_function_by_name(func_name, &actual_parameters, state);
1551 ir_call *call = NULL;
1552 ir_rvalue *value = NULL;
1554 no_matching_function_error(func_name, &loc, &actual_parameters, state);
1555 value = ir_rvalue::error_value(ctx);
1556 } else if (!verify_parameter_modes(state, sig, actual_parameters, this->expressions)) {
1557 /* an error has already been emitted */
1558 value = ir_rvalue::error_value(ctx);
1560 value = generate_call(instructions, sig, &actual_parameters,
1567 return ir_rvalue::error_value(ctx);