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_out
|| formal
->mode
== ir_var_inout
) {
136 const char *mode
= NULL
;
137 switch (formal
->mode
) {
138 case ir_var_out
: mode
= "out"; break;
139 case ir_var_inout
: mode
= "inout"; break;
140 default: assert(false); break;
143 /* This AST-based check catches errors like f(i++). The IR-based
144 * is_lvalue() is insufficient because the actual parameter at the
145 * IR-level is just a temporary value, which is an l-value.
147 if (actual_ast
->non_lvalue_description
!= NULL
) {
148 _mesa_glsl_error(&loc
, state
,
149 "function parameter '%s %s' references a %s",
151 actual_ast
->non_lvalue_description
);
155 ir_variable
*var
= actual
->variable_referenced();
157 var
->assigned
= true;
159 if (var
&& var
->read_only
) {
160 _mesa_glsl_error(&loc
, state
,
161 "function parameter '%s %s' references the "
162 "read-only variable '%s'",
164 actual
->variable_referenced()->name
);
166 } else if (!actual
->is_lvalue()) {
167 _mesa_glsl_error(&loc
, state
,
168 "function parameter '%s %s' is not an lvalue",
174 actual_ir_node
= actual_ir_node
->next
;
175 actual_ast_node
= actual_ast_node
->next
;
181 * If a function call is generated, \c call_ir will point to it on exit.
182 * Otherwise \c call_ir will be set to \c NULL.
185 generate_call(exec_list
*instructions
, ir_function_signature
*sig
,
186 YYLTYPE
*loc
, exec_list
*actual_parameters
,
188 struct _mesa_glsl_parse_state
*state
)
191 exec_list post_call_conversions
;
195 /* Perform implicit conversion of arguments. For out parameters, we need
196 * to place them in a temporary variable and do the conversion after the
197 * call takes place. Since we haven't emitted the call yet, we'll place
198 * the post-call conversions in a temporary exec_list, and emit them later.
200 exec_list_iterator actual_iter
= actual_parameters
->iterator();
201 exec_list_iterator formal_iter
= sig
->parameters
.iterator();
203 while (actual_iter
.has_next()) {
204 ir_rvalue
*actual
= (ir_rvalue
*) actual_iter
.get();
205 ir_variable
*formal
= (ir_variable
*) formal_iter
.get();
207 assert(actual
!= NULL
);
208 assert(formal
!= NULL
);
210 if (formal
->type
->is_numeric() || formal
->type
->is_boolean()) {
211 switch (formal
->mode
) {
212 case ir_var_const_in
:
215 = convert_component(actual
, formal
->type
);
216 actual
->replace_with(converted
);
220 if (actual
->type
!= formal
->type
) {
221 /* To convert an out parameter, we need to create a
222 * temporary variable to hold the value before conversion,
223 * and then perform the conversion after the function call
226 * This has the effect of transforming code like this:
232 * Into IR that's equivalent to this:
236 * int out_parameter_conversion;
237 * f(out_parameter_conversion);
238 * value = float(out_parameter_conversion);
241 new(ctx
) ir_variable(formal
->type
,
242 "out_parameter_conversion",
244 instructions
->push_tail(tmp
);
245 ir_dereference_variable
*deref_tmp_1
246 = new(ctx
) ir_dereference_variable(tmp
);
247 ir_dereference_variable
*deref_tmp_2
248 = new(ctx
) ir_dereference_variable(tmp
);
249 ir_rvalue
*converted_tmp
250 = convert_component(deref_tmp_1
, actual
->type
);
251 ir_assignment
*assignment
252 = new(ctx
) ir_assignment(actual
, converted_tmp
);
253 post_call_conversions
.push_tail(assignment
);
254 actual
->replace_with(deref_tmp_2
);
258 /* Inout parameters should never require conversion, since that
259 * would require an implicit conversion to exist both to and
260 * from the formal parameter type, and there are no
261 * bidirectional implicit conversions.
263 assert (actual
->type
== formal
->type
);
266 assert (!"Illegal formal parameter mode");
275 /* If the function call is a constant expression, don't generate any
276 * instructions; just generate an ir_constant.
278 * Function calls were first allowed to be constant expressions in GLSL 1.20.
280 if (state
->language_version
>= 120) {
281 ir_constant
*value
= sig
->constant_expression_value(actual_parameters
, NULL
);
287 ir_dereference_variable
*deref
= NULL
;
288 if (!sig
->return_type
->is_void()) {
289 /* Create a new temporary to hold the return value. */
292 var
= new(ctx
) ir_variable(sig
->return_type
,
293 ralloc_asprintf(ctx
, "%s_retval",
294 sig
->function_name()),
296 instructions
->push_tail(var
);
298 deref
= new(ctx
) ir_dereference_variable(var
);
300 ir_call
*call
= new(ctx
) ir_call(sig
, deref
, actual_parameters
);
301 instructions
->push_tail(call
);
303 /* Also emit any necessary out-parameter conversions. */
304 instructions
->append_list(&post_call_conversions
);
306 return deref
? deref
->clone(ctx
, NULL
) : NULL
;
310 * Given a function name and parameter list, find the matching signature.
312 static ir_function_signature
*
313 match_function_by_name(const char *name
,
314 exec_list
*actual_parameters
,
315 struct _mesa_glsl_parse_state
*state
)
318 ir_function
*f
= state
->symbols
->get_function(name
);
319 ir_function_signature
*local_sig
= NULL
;
320 ir_function_signature
*sig
= NULL
;
322 /* Is the function hidden by a record type constructor? */
323 if (state
->symbols
->get_type(name
))
324 goto done
; /* no match */
326 /* Is the function hidden by a variable (impossible in 1.10)? */
327 if (state
->language_version
!= 110 && state
->symbols
->get_variable(name
))
328 goto done
; /* no match */
331 /* Look for a match in the local shader. If exact, we're done. */
332 bool is_exact
= false;
333 sig
= local_sig
= f
->matching_signature(actual_parameters
, &is_exact
);
337 if (!state
->es_shader
&& f
->has_user_signature()) {
338 /* In desktop GL, the presence of a user-defined signature hides any
339 * built-in signatures, so we must ignore them. In contrast, in ES2
340 * user-defined signatures add new overloads, so we must proceed.
346 /* Local shader has no exact candidates; check the built-ins. */
347 _mesa_glsl_initialize_functions(state
);
348 for (unsigned i
= 0; i
< state
->num_builtins_to_link
; i
++) {
349 ir_function
*builtin
=
350 state
->builtins_to_link
[i
]->symbols
->get_function(name
);
354 bool is_exact
= false;
355 ir_function_signature
*builtin_sig
=
356 builtin
->matching_signature(actual_parameters
, &is_exact
);
358 if (builtin_sig
== NULL
)
361 /* If the built-in signature is exact, we can stop. */
368 /* We found an inexact match, which is better than nothing. However,
369 * we should keep searching for an exact match.
377 /* If the match is from a linked built-in shader, import the prototype. */
378 if (sig
!= local_sig
) {
380 f
= new(ctx
) ir_function(name
);
381 state
->symbols
->add_global_function(f
);
382 emit_function(state
, f
);
384 f
->add_signature(sig
->clone_prototype(f
, NULL
));
391 * Raise a "no matching function" error, listing all possible overloads the
392 * compiler considered so developers can figure out what went wrong.
395 no_matching_function_error(const char *name
,
397 exec_list
*actual_parameters
,
398 _mesa_glsl_parse_state
*state
)
400 char *str
= prototype_string(NULL
, name
, actual_parameters
);
401 _mesa_glsl_error(loc
, state
, "no matching function for call to `%s'", str
);
404 const char *prefix
= "candidates are: ";
406 for (int i
= -1; i
< (int) state
->num_builtins_to_link
; i
++) {
407 glsl_symbol_table
*syms
= i
>= 0 ? state
->builtins_to_link
[i
]->symbols
409 ir_function
*f
= syms
->get_function(name
);
413 foreach_list (node
, &f
->signatures
) {
414 ir_function_signature
*sig
= (ir_function_signature
*) node
;
416 str
= prototype_string(sig
->return_type
, f
->name
, &sig
->parameters
);
417 _mesa_glsl_error(loc
, state
, "%s%s", prefix
, str
);
426 * Perform automatic type conversion of constructor parameters
428 * This implements the rules in the "Conversion and Scalar Constructors"
429 * section (GLSL 1.10 section 5.4.1), not the "Implicit Conversions" rules.
432 convert_component(ir_rvalue
*src
, const glsl_type
*desired_type
)
434 void *ctx
= ralloc_parent(src
);
435 const unsigned a
= desired_type
->base_type
;
436 const unsigned b
= src
->type
->base_type
;
437 ir_expression
*result
= NULL
;
439 if (src
->type
->is_error())
442 assert(a
<= GLSL_TYPE_BOOL
);
443 assert(b
<= GLSL_TYPE_BOOL
);
452 result
= new(ctx
) ir_expression(ir_unop_i2u
, src
);
454 case GLSL_TYPE_FLOAT
:
455 result
= new(ctx
) ir_expression(ir_unop_i2u
,
456 new(ctx
) ir_expression(ir_unop_f2i
, src
));
459 result
= new(ctx
) ir_expression(ir_unop_i2u
,
460 new(ctx
) ir_expression(ir_unop_b2i
, src
));
467 result
= new(ctx
) ir_expression(ir_unop_u2i
, src
);
469 case GLSL_TYPE_FLOAT
:
470 result
= new(ctx
) ir_expression(ir_unop_f2i
, src
);
473 result
= new(ctx
) ir_expression(ir_unop_b2i
, src
);
477 case GLSL_TYPE_FLOAT
:
480 result
= new(ctx
) ir_expression(ir_unop_u2f
, desired_type
, src
, NULL
);
483 result
= new(ctx
) ir_expression(ir_unop_i2f
, desired_type
, src
, NULL
);
486 result
= new(ctx
) ir_expression(ir_unop_b2f
, desired_type
, src
, NULL
);
493 result
= new(ctx
) ir_expression(ir_unop_i2b
,
494 new(ctx
) ir_expression(ir_unop_u2i
, src
));
497 result
= new(ctx
) ir_expression(ir_unop_i2b
, desired_type
, src
, NULL
);
499 case GLSL_TYPE_FLOAT
:
500 result
= new(ctx
) ir_expression(ir_unop_f2b
, desired_type
, src
, NULL
);
506 assert(result
!= NULL
);
507 assert(result
->type
== desired_type
);
509 /* Try constant folding; it may fold in the conversion we just added. */
510 ir_constant
*const constant
= result
->constant_expression_value();
511 return (constant
!= NULL
) ? (ir_rvalue
*) constant
: (ir_rvalue
*) result
;
515 * Dereference a specific component from a scalar, vector, or matrix
518 dereference_component(ir_rvalue
*src
, unsigned component
)
520 void *ctx
= ralloc_parent(src
);
521 assert(component
< src
->type
->components());
523 /* If the source is a constant, just create a new constant instead of a
524 * dereference of the existing constant.
526 ir_constant
*constant
= src
->as_constant();
528 return new(ctx
) ir_constant(constant
, component
);
530 if (src
->type
->is_scalar()) {
532 } else if (src
->type
->is_vector()) {
533 return new(ctx
) ir_swizzle(src
, component
, 0, 0, 0, 1);
535 assert(src
->type
->is_matrix());
537 /* Dereference a row of the matrix, then call this function again to get
538 * a specific element from that row.
540 const int c
= component
/ src
->type
->column_type()->vector_elements
;
541 const int r
= component
% src
->type
->column_type()->vector_elements
;
542 ir_constant
*const col_index
= new(ctx
) ir_constant(c
);
543 ir_dereference
*const col
= new(ctx
) ir_dereference_array(src
, col_index
);
545 col
->type
= src
->type
->column_type();
547 return dereference_component(col
, r
);
550 assert(!"Should not get here.");
556 process_array_constructor(exec_list
*instructions
,
557 const glsl_type
*constructor_type
,
558 YYLTYPE
*loc
, exec_list
*parameters
,
559 struct _mesa_glsl_parse_state
*state
)
562 /* Array constructors come in two forms: sized and unsized. Sized array
563 * constructors look like 'vec4[2](a, b)', where 'a' and 'b' are vec4
564 * variables. In this case the number of parameters must exactly match the
565 * specified size of the array.
567 * Unsized array constructors look like 'vec4[](a, b)', where 'a' and 'b'
568 * are vec4 variables. In this case the size of the array being constructed
569 * is determined by the number of parameters.
571 * From page 52 (page 58 of the PDF) of the GLSL 1.50 spec:
573 * "There must be exactly the same number of arguments as the size of
574 * the array being constructed. If no size is present in the
575 * constructor, then the array is explicitly sized to the number of
576 * arguments provided. The arguments are assigned in order, starting at
577 * element 0, to the elements of the constructed array. Each argument
578 * must be the same type as the element type of the array, or be a type
579 * that can be converted to the element type of the array according to
580 * Section 4.1.10 "Implicit Conversions.""
582 exec_list actual_parameters
;
583 const unsigned parameter_count
=
584 process_parameters(instructions
, &actual_parameters
, parameters
, state
);
586 if ((parameter_count
== 0)
587 || ((constructor_type
->length
!= 0)
588 && (constructor_type
->length
!= parameter_count
))) {
589 const unsigned min_param
= (constructor_type
->length
== 0)
590 ? 1 : constructor_type
->length
;
592 _mesa_glsl_error(loc
, state
, "array constructor must have %s %u "
594 (constructor_type
->length
!= 0) ? "at least" : "exactly",
595 min_param
, (min_param
<= 1) ? "" : "s");
596 return ir_rvalue::error_value(ctx
);
599 if (constructor_type
->length
== 0) {
601 glsl_type::get_array_instance(constructor_type
->element_type(),
603 assert(constructor_type
!= NULL
);
604 assert(constructor_type
->length
== parameter_count
);
607 bool all_parameters_are_constant
= true;
609 /* Type cast each parameter and, if possible, fold constants. */
610 foreach_list_safe(n
, &actual_parameters
) {
611 ir_rvalue
*ir
= (ir_rvalue
*) n
;
612 ir_rvalue
*result
= ir
;
614 /* Apply implicit conversions (not the scalar constructor rules!). See
615 * the spec quote above. */
616 if (constructor_type
->element_type()->is_float()) {
617 const glsl_type
*desired_type
=
618 glsl_type::get_instance(GLSL_TYPE_FLOAT
,
619 ir
->type
->vector_elements
,
620 ir
->type
->matrix_columns
);
621 if (result
->type
->can_implicitly_convert_to(desired_type
)) {
622 /* Even though convert_component() implements the constructor
623 * conversion rules (not the implicit conversion rules), its safe
624 * to use it here because we already checked that the implicit
625 * conversion is legal.
627 result
= convert_component(ir
, desired_type
);
631 if (result
->type
!= constructor_type
->element_type()) {
632 _mesa_glsl_error(loc
, state
, "type error in array constructor: "
633 "expected: %s, found %s",
634 constructor_type
->element_type()->name
,
638 /* Attempt to convert the parameter to a constant valued expression.
639 * After doing so, track whether or not all the parameters to the
640 * constructor are trivially constant valued expressions.
642 ir_rvalue
*const constant
= result
->constant_expression_value();
644 if (constant
!= NULL
)
647 all_parameters_are_constant
= false;
649 ir
->replace_with(result
);
652 if (all_parameters_are_constant
)
653 return new(ctx
) ir_constant(constructor_type
, &actual_parameters
);
655 ir_variable
*var
= new(ctx
) ir_variable(constructor_type
, "array_ctor",
657 instructions
->push_tail(var
);
660 foreach_list(node
, &actual_parameters
) {
661 ir_rvalue
*rhs
= (ir_rvalue
*) node
;
662 ir_rvalue
*lhs
= new(ctx
) ir_dereference_array(var
,
663 new(ctx
) ir_constant(i
));
665 ir_instruction
*assignment
= new(ctx
) ir_assignment(lhs
, rhs
, NULL
);
666 instructions
->push_tail(assignment
);
671 return new(ctx
) ir_dereference_variable(var
);
676 * Try to convert a record constructor to a constant expression
679 constant_record_constructor(const glsl_type
*constructor_type
,
680 exec_list
*parameters
, void *mem_ctx
)
682 foreach_list(node
, parameters
) {
683 ir_constant
*constant
= ((ir_instruction
*) node
)->as_constant();
684 if (constant
== NULL
)
686 node
->replace_with(constant
);
689 return new(mem_ctx
) ir_constant(constructor_type
, parameters
);
694 * Determine if a list consists of a single scalar r-value
697 single_scalar_parameter(exec_list
*parameters
)
699 const ir_rvalue
*const p
= (ir_rvalue
*) parameters
->head
;
700 assert(((ir_rvalue
*)p
)->as_rvalue() != NULL
);
702 return (p
->type
->is_scalar() && p
->next
->is_tail_sentinel());
707 * Generate inline code for a vector constructor
709 * The generated constructor code will consist of a temporary variable
710 * declaration of the same type as the constructor. A sequence of assignments
711 * from constructor parameters to the temporary will follow.
714 * An \c ir_dereference_variable of the temprorary generated in the constructor
718 emit_inline_vector_constructor(const glsl_type
*type
,
719 exec_list
*instructions
,
720 exec_list
*parameters
,
723 assert(!parameters
->is_empty());
725 ir_variable
*var
= new(ctx
) ir_variable(type
, "vec_ctor", ir_var_temporary
);
726 instructions
->push_tail(var
);
728 /* There are two kinds of vector constructors.
730 * - Construct a vector from a single scalar by replicating that scalar to
731 * all components of the vector.
733 * - Construct a vector from an arbirary combination of vectors and
734 * scalars. The components of the constructor parameters are assigned
735 * to the vector in order until the vector is full.
737 const unsigned lhs_components
= type
->components();
738 if (single_scalar_parameter(parameters
)) {
739 ir_rvalue
*first_param
= (ir_rvalue
*)parameters
->head
;
740 ir_rvalue
*rhs
= new(ctx
) ir_swizzle(first_param
, 0, 0, 0, 0,
742 ir_dereference_variable
*lhs
= new(ctx
) ir_dereference_variable(var
);
743 const unsigned mask
= (1U << lhs_components
) - 1;
745 assert(rhs
->type
== lhs
->type
);
747 ir_instruction
*inst
= new(ctx
) ir_assignment(lhs
, rhs
, NULL
, mask
);
748 instructions
->push_tail(inst
);
750 unsigned base_component
= 0;
751 unsigned base_lhs_component
= 0;
752 ir_constant_data data
;
753 unsigned constant_mask
= 0, constant_components
= 0;
755 memset(&data
, 0, sizeof(data
));
757 foreach_list(node
, parameters
) {
758 ir_rvalue
*param
= (ir_rvalue
*) node
;
759 unsigned rhs_components
= param
->type
->components();
761 /* Do not try to assign more components to the vector than it has!
763 if ((rhs_components
+ base_lhs_component
) > lhs_components
) {
764 rhs_components
= lhs_components
- base_lhs_component
;
767 const ir_constant
*const c
= param
->as_constant();
769 for (unsigned i
= 0; i
< rhs_components
; i
++) {
770 switch (c
->type
->base_type
) {
772 data
.u
[i
+ base_component
] = c
->get_uint_component(i
);
775 data
.i
[i
+ base_component
] = c
->get_int_component(i
);
777 case GLSL_TYPE_FLOAT
:
778 data
.f
[i
+ base_component
] = c
->get_float_component(i
);
781 data
.b
[i
+ base_component
] = c
->get_bool_component(i
);
784 assert(!"Should not get here.");
789 /* Mask of fields to be written in the assignment.
791 constant_mask
|= ((1U << rhs_components
) - 1) << base_lhs_component
;
792 constant_components
+= rhs_components
;
794 base_component
+= rhs_components
;
796 /* Advance the component index by the number of components
797 * that were just assigned.
799 base_lhs_component
+= rhs_components
;
802 if (constant_mask
!= 0) {
803 ir_dereference
*lhs
= new(ctx
) ir_dereference_variable(var
);
804 const glsl_type
*rhs_type
= glsl_type::get_instance(var
->type
->base_type
,
807 ir_rvalue
*rhs
= new(ctx
) ir_constant(rhs_type
, &data
);
809 ir_instruction
*inst
=
810 new(ctx
) ir_assignment(lhs
, rhs
, NULL
, constant_mask
);
811 instructions
->push_tail(inst
);
815 foreach_list(node
, parameters
) {
816 ir_rvalue
*param
= (ir_rvalue
*) node
;
817 unsigned rhs_components
= param
->type
->components();
819 /* Do not try to assign more components to the vector than it has!
821 if ((rhs_components
+ base_component
) > lhs_components
) {
822 rhs_components
= lhs_components
- base_component
;
825 const ir_constant
*const c
= param
->as_constant();
827 /* Mask of fields to be written in the assignment.
829 const unsigned write_mask
= ((1U << rhs_components
) - 1)
832 ir_dereference
*lhs
= new(ctx
) ir_dereference_variable(var
);
834 /* Generate a swizzle so that LHS and RHS sizes match.
837 new(ctx
) ir_swizzle(param
, 0, 1, 2, 3, rhs_components
);
839 ir_instruction
*inst
=
840 new(ctx
) ir_assignment(lhs
, rhs
, NULL
, write_mask
);
841 instructions
->push_tail(inst
);
844 /* Advance the component index by the number of components that were
847 base_component
+= rhs_components
;
850 return new(ctx
) ir_dereference_variable(var
);
855 * Generate assignment of a portion of a vector to a portion of a matrix column
857 * \param src_base First component of the source to be used in assignment
858 * \param column Column of destination to be assiged
859 * \param row_base First component of the destination column to be assigned
860 * \param count Number of components to be assigned
863 * \c src_base + \c count must be less than or equal to the number of components
864 * in the source vector.
867 assign_to_matrix_column(ir_variable
*var
, unsigned column
, unsigned row_base
,
868 ir_rvalue
*src
, unsigned src_base
, unsigned count
,
871 ir_constant
*col_idx
= new(mem_ctx
) ir_constant(column
);
872 ir_dereference
*column_ref
= new(mem_ctx
) ir_dereference_array(var
, col_idx
);
874 assert(column_ref
->type
->components() >= (row_base
+ count
));
875 assert(src
->type
->components() >= (src_base
+ count
));
877 /* Generate a swizzle that extracts the number of components from the source
878 * that are to be assigned to the column of the matrix.
880 if (count
< src
->type
->vector_elements
) {
881 src
= new(mem_ctx
) ir_swizzle(src
,
882 src_base
+ 0, src_base
+ 1,
883 src_base
+ 2, src_base
+ 3,
887 /* Mask of fields to be written in the assignment.
889 const unsigned write_mask
= ((1U << count
) - 1) << row_base
;
891 return new(mem_ctx
) ir_assignment(column_ref
, src
, NULL
, write_mask
);
896 * Generate inline code for a matrix constructor
898 * The generated constructor code will consist of a temporary variable
899 * declaration of the same type as the constructor. A sequence of assignments
900 * from constructor parameters to the temporary will follow.
903 * An \c ir_dereference_variable of the temprorary generated in the constructor
907 emit_inline_matrix_constructor(const glsl_type
*type
,
908 exec_list
*instructions
,
909 exec_list
*parameters
,
912 assert(!parameters
->is_empty());
914 ir_variable
*var
= new(ctx
) ir_variable(type
, "mat_ctor", ir_var_temporary
);
915 instructions
->push_tail(var
);
917 /* There are three kinds of matrix constructors.
919 * - Construct a matrix from a single scalar by replicating that scalar to
920 * along the diagonal of the matrix and setting all other components to
923 * - Construct a matrix from an arbirary combination of vectors and
924 * scalars. The components of the constructor parameters are assigned
925 * to the matrix in colum-major order until the matrix is full.
927 * - Construct a matrix from a single matrix. The source matrix is copied
928 * to the upper left portion of the constructed matrix, and the remaining
929 * elements take values from the identity matrix.
931 ir_rvalue
*const first_param
= (ir_rvalue
*) parameters
->head
;
932 if (single_scalar_parameter(parameters
)) {
933 /* Assign the scalar to the X component of a vec4, and fill the remaining
934 * components with zero.
936 ir_variable
*rhs_var
=
937 new(ctx
) ir_variable(glsl_type::vec4_type
, "mat_ctor_vec",
939 instructions
->push_tail(rhs_var
);
941 ir_constant_data zero
;
947 ir_instruction
*inst
=
948 new(ctx
) ir_assignment(new(ctx
) ir_dereference_variable(rhs_var
),
949 new(ctx
) ir_constant(rhs_var
->type
, &zero
),
951 instructions
->push_tail(inst
);
953 ir_dereference
*const rhs_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
955 inst
= new(ctx
) ir_assignment(rhs_ref
, first_param
, NULL
, 0x01);
956 instructions
->push_tail(inst
);
958 /* Assign the temporary vector to each column of the destination matrix
959 * with a swizzle that puts the X component on the diagonal of the
960 * matrix. In some cases this may mean that the X component does not
961 * get assigned into the column at all (i.e., when the matrix has more
962 * columns than rows).
964 static const unsigned rhs_swiz
[4][4] = {
971 const unsigned cols_to_init
= MIN2(type
->matrix_columns
,
972 type
->vector_elements
);
973 for (unsigned i
= 0; i
< cols_to_init
; i
++) {
974 ir_constant
*const col_idx
= new(ctx
) ir_constant(i
);
975 ir_rvalue
*const col_ref
= new(ctx
) ir_dereference_array(var
, col_idx
);
977 ir_rvalue
*const rhs_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
978 ir_rvalue
*const rhs
= new(ctx
) ir_swizzle(rhs_ref
, rhs_swiz
[i
],
979 type
->vector_elements
);
981 inst
= new(ctx
) ir_assignment(col_ref
, rhs
, NULL
);
982 instructions
->push_tail(inst
);
985 for (unsigned i
= cols_to_init
; i
< type
->matrix_columns
; i
++) {
986 ir_constant
*const col_idx
= new(ctx
) ir_constant(i
);
987 ir_rvalue
*const col_ref
= new(ctx
) ir_dereference_array(var
, col_idx
);
989 ir_rvalue
*const rhs_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
990 ir_rvalue
*const rhs
= new(ctx
) ir_swizzle(rhs_ref
, 1, 1, 1, 1,
991 type
->vector_elements
);
993 inst
= new(ctx
) ir_assignment(col_ref
, rhs
, NULL
);
994 instructions
->push_tail(inst
);
996 } else if (first_param
->type
->is_matrix()) {
997 /* From page 50 (56 of the PDF) of the GLSL 1.50 spec:
999 * "If a matrix is constructed from a matrix, then each component
1000 * (column i, row j) in the result that has a corresponding
1001 * component (column i, row j) in the argument will be initialized
1002 * from there. All other components will be initialized to the
1003 * identity matrix. If a matrix argument is given to a matrix
1004 * constructor, it is an error to have any other arguments."
1006 assert(first_param
->next
->is_tail_sentinel());
1007 ir_rvalue
*const src_matrix
= first_param
;
1009 /* If the source matrix is smaller, pre-initialize the relavent parts of
1010 * the destination matrix to the identity matrix.
1012 if ((src_matrix
->type
->matrix_columns
< var
->type
->matrix_columns
)
1013 || (src_matrix
->type
->vector_elements
< var
->type
->vector_elements
)) {
1015 /* If the source matrix has fewer rows, every column of the destination
1016 * must be initialized. Otherwise only the columns in the destination
1017 * that do not exist in the source must be initialized.
1020 (src_matrix
->type
->vector_elements
< var
->type
->vector_elements
)
1021 ? 0 : src_matrix
->type
->matrix_columns
;
1023 const glsl_type
*const col_type
= var
->type
->column_type();
1024 for (/* empty */; col
< var
->type
->matrix_columns
; col
++) {
1025 ir_constant_data ident
;
1034 ir_rvalue
*const rhs
= new(ctx
) ir_constant(col_type
, &ident
);
1036 ir_rvalue
*const lhs
=
1037 new(ctx
) ir_dereference_array(var
, new(ctx
) ir_constant(col
));
1039 ir_instruction
*inst
= new(ctx
) ir_assignment(lhs
, rhs
, NULL
);
1040 instructions
->push_tail(inst
);
1044 /* Assign columns from the source matrix to the destination matrix.
1046 * Since the parameter will be used in the RHS of multiple assignments,
1047 * generate a temporary and copy the paramter there.
1049 ir_variable
*const rhs_var
=
1050 new(ctx
) ir_variable(first_param
->type
, "mat_ctor_mat",
1052 instructions
->push_tail(rhs_var
);
1054 ir_dereference
*const rhs_var_ref
=
1055 new(ctx
) ir_dereference_variable(rhs_var
);
1056 ir_instruction
*const inst
=
1057 new(ctx
) ir_assignment(rhs_var_ref
, first_param
, NULL
);
1058 instructions
->push_tail(inst
);
1060 const unsigned last_row
= MIN2(src_matrix
->type
->vector_elements
,
1061 var
->type
->vector_elements
);
1062 const unsigned last_col
= MIN2(src_matrix
->type
->matrix_columns
,
1063 var
->type
->matrix_columns
);
1065 unsigned swiz
[4] = { 0, 0, 0, 0 };
1066 for (unsigned i
= 1; i
< last_row
; i
++)
1069 const unsigned write_mask
= (1U << last_row
) - 1;
1071 for (unsigned i
= 0; i
< last_col
; i
++) {
1072 ir_dereference
*const lhs
=
1073 new(ctx
) ir_dereference_array(var
, new(ctx
) ir_constant(i
));
1074 ir_rvalue
*const rhs_col
=
1075 new(ctx
) ir_dereference_array(rhs_var
, new(ctx
) ir_constant(i
));
1077 /* If one matrix has columns that are smaller than the columns of the
1078 * other matrix, wrap the column access of the larger with a swizzle
1079 * so that the LHS and RHS of the assignment have the same size (and
1080 * therefore have the same type).
1082 * It would be perfectly valid to unconditionally generate the
1083 * swizzles, this this will typically result in a more compact IR tree.
1086 if (lhs
->type
->vector_elements
!= rhs_col
->type
->vector_elements
) {
1087 rhs
= new(ctx
) ir_swizzle(rhs_col
, swiz
, last_row
);
1092 ir_instruction
*inst
=
1093 new(ctx
) ir_assignment(lhs
, rhs
, NULL
, write_mask
);
1094 instructions
->push_tail(inst
);
1097 const unsigned cols
= type
->matrix_columns
;
1098 const unsigned rows
= type
->vector_elements
;
1099 unsigned col_idx
= 0;
1100 unsigned row_idx
= 0;
1102 foreach_list (node
, parameters
) {
1103 ir_rvalue
*const rhs
= (ir_rvalue
*) node
;
1104 const unsigned components_remaining_this_column
= rows
- row_idx
;
1105 unsigned rhs_components
= rhs
->type
->components();
1106 unsigned rhs_base
= 0;
1108 /* Since the parameter might be used in the RHS of two assignments,
1109 * generate a temporary and copy the paramter there.
1111 ir_variable
*rhs_var
=
1112 new(ctx
) ir_variable(rhs
->type
, "mat_ctor_vec", ir_var_temporary
);
1113 instructions
->push_tail(rhs_var
);
1115 ir_dereference
*rhs_var_ref
=
1116 new(ctx
) ir_dereference_variable(rhs_var
);
1117 ir_instruction
*inst
= new(ctx
) ir_assignment(rhs_var_ref
, rhs
, NULL
);
1118 instructions
->push_tail(inst
);
1120 /* Assign the current parameter to as many components of the matrix
1123 * NOTE: A single vector parameter can span two matrix columns. A
1124 * single vec4, for example, can completely fill a mat2.
1126 if (rhs_components
>= components_remaining_this_column
) {
1127 const unsigned count
= MIN2(rhs_components
,
1128 components_remaining_this_column
);
1130 rhs_var_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
1132 ir_instruction
*inst
= assign_to_matrix_column(var
, col_idx
,
1136 instructions
->push_tail(inst
);
1144 /* If there is data left in the parameter and components left to be
1145 * set in the destination, emit another assignment. It is possible
1146 * that the assignment could be of a vec4 to the last element of the
1147 * matrix. In this case col_idx==cols, but there is still data
1148 * left in the source parameter. Obviously, don't emit an assignment
1149 * to data outside the destination matrix.
1151 if ((col_idx
< cols
) && (rhs_base
< rhs_components
)) {
1152 const unsigned count
= rhs_components
- rhs_base
;
1154 rhs_var_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
1156 ir_instruction
*inst
= assign_to_matrix_column(var
, col_idx
,
1161 instructions
->push_tail(inst
);
1168 return new(ctx
) ir_dereference_variable(var
);
1173 emit_inline_record_constructor(const glsl_type
*type
,
1174 exec_list
*instructions
,
1175 exec_list
*parameters
,
1178 ir_variable
*const var
=
1179 new(mem_ctx
) ir_variable(type
, "record_ctor", ir_var_temporary
);
1180 ir_dereference_variable
*const d
= new(mem_ctx
) ir_dereference_variable(var
);
1182 instructions
->push_tail(var
);
1184 exec_node
*node
= parameters
->head
;
1185 for (unsigned i
= 0; i
< type
->length
; i
++) {
1186 assert(!node
->is_tail_sentinel());
1188 ir_dereference
*const lhs
=
1189 new(mem_ctx
) ir_dereference_record(d
->clone(mem_ctx
, NULL
),
1190 type
->fields
.structure
[i
].name
);
1192 ir_rvalue
*const rhs
= ((ir_instruction
*) node
)->as_rvalue();
1193 assert(rhs
!= NULL
);
1195 ir_instruction
*const assign
= new(mem_ctx
) ir_assignment(lhs
, rhs
, NULL
);
1197 instructions
->push_tail(assign
);
1206 ast_function_expression::hir(exec_list
*instructions
,
1207 struct _mesa_glsl_parse_state
*state
)
1210 /* There are three sorts of function calls.
1212 * 1. constructors - The first subexpression is an ast_type_specifier.
1213 * 2. methods - Only the .length() method of array types.
1214 * 3. functions - Calls to regular old functions.
1216 * Method calls are actually detected when the ast_field_selection
1217 * expression is handled.
1219 if (is_constructor()) {
1220 const ast_type_specifier
*type
= (ast_type_specifier
*) subexpressions
[0];
1221 YYLTYPE loc
= type
->get_location();
1224 const glsl_type
*const constructor_type
= type
->glsl_type(& name
, state
);
1226 /* constructor_type can be NULL if a variable with the same name as the
1227 * structure has come into scope.
1229 if (constructor_type
== NULL
) {
1230 _mesa_glsl_error(& loc
, state
, "unknown type `%s' (structure name "
1231 "may be shadowed by a variable with the same name)",
1233 return ir_rvalue::error_value(ctx
);
1237 /* Constructors for samplers are illegal.
1239 if (constructor_type
->is_sampler()) {
1240 _mesa_glsl_error(& loc
, state
, "cannot construct sampler type `%s'",
1241 constructor_type
->name
);
1242 return ir_rvalue::error_value(ctx
);
1245 if (constructor_type
->is_array()) {
1246 if (state
->language_version
<= 110) {
1247 _mesa_glsl_error(& loc
, state
,
1248 "array constructors forbidden in GLSL 1.10");
1249 return ir_rvalue::error_value(ctx
);
1252 return process_array_constructor(instructions
, constructor_type
,
1253 & loc
, &this->expressions
, state
);
1257 /* There are two kinds of constructor call. Constructors for built-in
1258 * language types, such as mat4 and vec2, are free form. The only
1259 * requirement is that the parameters must provide enough values of the
1260 * correct scalar type. Constructors for arrays and structures must
1261 * have the exact number of parameters with matching types in the
1262 * correct order. These constructors follow essentially the same type
1263 * matching rules as functions.
1265 if (constructor_type
->is_record()) {
1266 exec_list actual_parameters
;
1268 process_parameters(instructions
, &actual_parameters
,
1269 &this->expressions
, state
);
1271 exec_node
*node
= actual_parameters
.head
;
1272 for (unsigned i
= 0; i
< constructor_type
->length
; i
++) {
1273 ir_rvalue
*ir
= (ir_rvalue
*) node
;
1275 if (node
->is_tail_sentinel()) {
1276 _mesa_glsl_error(&loc
, state
,
1277 "insufficient parameters to constructor "
1279 constructor_type
->name
);
1280 return ir_rvalue::error_value(ctx
);
1283 if (apply_implicit_conversion(constructor_type
->fields
.structure
[i
].type
,
1285 node
->replace_with(ir
);
1287 _mesa_glsl_error(&loc
, state
,
1288 "parameter type mismatch in constructor "
1289 "for `%s.%s' (%s vs %s)",
1290 constructor_type
->name
,
1291 constructor_type
->fields
.structure
[i
].name
,
1293 constructor_type
->fields
.structure
[i
].type
->name
);
1294 return ir_rvalue::error_value(ctx
);;
1300 if (!node
->is_tail_sentinel()) {
1301 _mesa_glsl_error(&loc
, state
, "too many parameters in constructor "
1302 "for `%s'", constructor_type
->name
);
1303 return ir_rvalue::error_value(ctx
);
1306 ir_rvalue
*const constant
=
1307 constant_record_constructor(constructor_type
, &actual_parameters
,
1310 return (constant
!= NULL
)
1312 : emit_inline_record_constructor(constructor_type
, instructions
,
1313 &actual_parameters
, state
);
1316 if (!constructor_type
->is_numeric() && !constructor_type
->is_boolean())
1317 return ir_rvalue::error_value(ctx
);
1319 /* Total number of components of the type being constructed. */
1320 const unsigned type_components
= constructor_type
->components();
1322 /* Number of components from parameters that have actually been
1323 * consumed. This is used to perform several kinds of error checking.
1325 unsigned components_used
= 0;
1327 unsigned matrix_parameters
= 0;
1328 unsigned nonmatrix_parameters
= 0;
1329 exec_list actual_parameters
;
1331 foreach_list (n
, &this->expressions
) {
1332 ast_node
*ast
= exec_node_data(ast_node
, n
, link
);
1333 ir_rvalue
*result
= ast
->hir(instructions
, state
)->as_rvalue();
1335 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
1337 * "It is an error to provide extra arguments beyond this
1338 * last used argument."
1340 if (components_used
>= type_components
) {
1341 _mesa_glsl_error(& loc
, state
, "too many parameters to `%s' "
1343 constructor_type
->name
);
1344 return ir_rvalue::error_value(ctx
);
1347 if (!result
->type
->is_numeric() && !result
->type
->is_boolean()) {
1348 _mesa_glsl_error(& loc
, state
, "cannot construct `%s' from a "
1349 "non-numeric data type",
1350 constructor_type
->name
);
1351 return ir_rvalue::error_value(ctx
);
1354 /* Count the number of matrix and nonmatrix parameters. This
1355 * is used below to enforce some of the constructor rules.
1357 if (result
->type
->is_matrix())
1358 matrix_parameters
++;
1360 nonmatrix_parameters
++;
1362 actual_parameters
.push_tail(result
);
1363 components_used
+= result
->type
->components();
1366 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
1368 * "It is an error to construct matrices from other matrices. This
1369 * is reserved for future use."
1371 if (state
->language_version
== 110 && matrix_parameters
> 0
1372 && constructor_type
->is_matrix()) {
1373 _mesa_glsl_error(& loc
, state
, "cannot construct `%s' from a "
1374 "matrix in GLSL 1.10",
1375 constructor_type
->name
);
1376 return ir_rvalue::error_value(ctx
);
1379 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
1381 * "If a matrix argument is given to a matrix constructor, it is
1382 * an error to have any other arguments."
1384 if ((matrix_parameters
> 0)
1385 && ((matrix_parameters
+ nonmatrix_parameters
) > 1)
1386 && constructor_type
->is_matrix()) {
1387 _mesa_glsl_error(& loc
, state
, "for matrix `%s' constructor, "
1388 "matrix must be only parameter",
1389 constructor_type
->name
);
1390 return ir_rvalue::error_value(ctx
);
1393 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
1395 * "In these cases, there must be enough components provided in the
1396 * arguments to provide an initializer for every component in the
1397 * constructed value."
1399 if (components_used
< type_components
&& components_used
!= 1
1400 && matrix_parameters
== 0) {
1401 _mesa_glsl_error(& loc
, state
, "too few components to construct "
1403 constructor_type
->name
);
1404 return ir_rvalue::error_value(ctx
);
1407 /* Later, we cast each parameter to the same base type as the
1408 * constructor. Since there are no non-floating point matrices, we
1409 * need to break them up into a series of column vectors.
1411 if (constructor_type
->base_type
!= GLSL_TYPE_FLOAT
) {
1412 foreach_list_safe(n
, &actual_parameters
) {
1413 ir_rvalue
*matrix
= (ir_rvalue
*) n
;
1415 if (!matrix
->type
->is_matrix())
1418 /* Create a temporary containing the matrix. */
1419 ir_variable
*var
= new(ctx
) ir_variable(matrix
->type
, "matrix_tmp",
1421 instructions
->push_tail(var
);
1422 instructions
->push_tail(new(ctx
) ir_assignment(new(ctx
)
1423 ir_dereference_variable(var
), matrix
, NULL
));
1424 var
->constant_value
= matrix
->constant_expression_value();
1426 /* Replace the matrix with dereferences of its columns. */
1427 for (int i
= 0; i
< matrix
->type
->matrix_columns
; i
++) {
1428 matrix
->insert_before(new (ctx
) ir_dereference_array(var
,
1429 new(ctx
) ir_constant(i
)));
1435 bool all_parameters_are_constant
= true;
1437 /* Type cast each parameter and, if possible, fold constants.*/
1438 foreach_list_safe(n
, &actual_parameters
) {
1439 ir_rvalue
*ir
= (ir_rvalue
*) n
;
1441 const glsl_type
*desired_type
=
1442 glsl_type::get_instance(constructor_type
->base_type
,
1443 ir
->type
->vector_elements
,
1444 ir
->type
->matrix_columns
);
1445 ir_rvalue
*result
= convert_component(ir
, desired_type
);
1447 /* Attempt to convert the parameter to a constant valued expression.
1448 * After doing so, track whether or not all the parameters to the
1449 * constructor are trivially constant valued expressions.
1451 ir_rvalue
*const constant
= result
->constant_expression_value();
1453 if (constant
!= NULL
)
1456 all_parameters_are_constant
= false;
1459 ir
->replace_with(result
);
1463 /* If all of the parameters are trivially constant, create a
1464 * constant representing the complete collection of parameters.
1466 if (all_parameters_are_constant
) {
1467 return new(ctx
) ir_constant(constructor_type
, &actual_parameters
);
1468 } else if (constructor_type
->is_scalar()) {
1469 return dereference_component((ir_rvalue
*) actual_parameters
.head
,
1471 } else if (constructor_type
->is_vector()) {
1472 return emit_inline_vector_constructor(constructor_type
,
1477 assert(constructor_type
->is_matrix());
1478 return emit_inline_matrix_constructor(constructor_type
,
1484 const ast_expression
*id
= subexpressions
[0];
1485 const char *func_name
= id
->primary_expression
.identifier
;
1486 YYLTYPE loc
= id
->get_location();
1487 exec_list actual_parameters
;
1489 process_parameters(instructions
, &actual_parameters
, &this->expressions
,
1492 ir_function_signature
*sig
=
1493 match_function_by_name(func_name
, &actual_parameters
, state
);
1495 ir_call
*call
= NULL
;
1496 ir_rvalue
*value
= NULL
;
1498 no_matching_function_error(func_name
, &loc
, &actual_parameters
, state
);
1499 value
= ir_rvalue::error_value(ctx
);
1500 } else if (!verify_parameter_modes(state
, sig
, actual_parameters
, this->expressions
)) {
1501 /* an error has already been emitted */
1502 value
= ir_rvalue::error_value(ctx
);
1504 value
= generate_call(instructions
, sig
, &loc
, &actual_parameters
,
1511 return ir_rvalue::error_value(ctx
);