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_instruction
*const param
= (ir_instruction
*) 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 if (actual
->variable_referenced()
156 && actual
->variable_referenced()->read_only
) {
157 _mesa_glsl_error(&loc
, state
,
158 "function parameter '%s %s' references the "
159 "read-only variable '%s'",
161 actual
->variable_referenced()->name
);
163 } else if (!actual
->is_lvalue()) {
164 _mesa_glsl_error(&loc
, state
,
165 "function parameter '%s %s' is not an lvalue",
171 actual_ir_node
= actual_ir_node
->next
;
172 actual_ast_node
= actual_ast_node
->next
;
178 * If a function call is generated, \c call_ir will point to it on exit.
179 * Otherwise \c call_ir will be set to \c NULL.
182 generate_call(exec_list
*instructions
, ir_function_signature
*sig
,
183 YYLTYPE
*loc
, exec_list
*actual_parameters
,
185 struct _mesa_glsl_parse_state
*state
)
188 exec_list post_call_conversions
;
192 /* Perform implicit conversion of arguments. For out parameters, we need
193 * to place them in a temporary variable and do the conversion after the
194 * call takes place. Since we haven't emitted the call yet, we'll place
195 * the post-call conversions in a temporary exec_list, and emit them later.
197 exec_list_iterator actual_iter
= actual_parameters
->iterator();
198 exec_list_iterator formal_iter
= sig
->parameters
.iterator();
200 while (actual_iter
.has_next()) {
201 ir_rvalue
*actual
= (ir_rvalue
*) actual_iter
.get();
202 ir_variable
*formal
= (ir_variable
*) formal_iter
.get();
204 assert(actual
!= NULL
);
205 assert(formal
!= NULL
);
207 if (formal
->type
->is_numeric() || formal
->type
->is_boolean()) {
208 switch (formal
->mode
) {
209 case ir_var_const_in
:
212 = convert_component(actual
, formal
->type
);
213 actual
->replace_with(converted
);
217 if (actual
->type
!= formal
->type
) {
218 /* To convert an out parameter, we need to create a
219 * temporary variable to hold the value before conversion,
220 * and then perform the conversion after the function call
223 * This has the effect of transforming code like this:
229 * Into IR that's equivalent to this:
233 * int out_parameter_conversion;
234 * f(out_parameter_conversion);
235 * value = float(out_parameter_conversion);
238 new(ctx
) ir_variable(formal
->type
,
239 "out_parameter_conversion",
241 instructions
->push_tail(tmp
);
242 ir_dereference_variable
*deref_tmp_1
243 = new(ctx
) ir_dereference_variable(tmp
);
244 ir_dereference_variable
*deref_tmp_2
245 = new(ctx
) ir_dereference_variable(tmp
);
246 ir_rvalue
*converted_tmp
247 = convert_component(deref_tmp_1
, actual
->type
);
248 ir_assignment
*assignment
249 = new(ctx
) ir_assignment(actual
, converted_tmp
);
250 post_call_conversions
.push_tail(assignment
);
251 actual
->replace_with(deref_tmp_2
);
255 /* Inout parameters should never require conversion, since that
256 * would require an implicit conversion to exist both to and
257 * from the formal parameter type, and there are no
258 * bidirectional implicit conversions.
260 assert (actual
->type
== formal
->type
);
263 assert (!"Illegal formal parameter mode");
272 /* If the function call is a constant expression, don't generate any
273 * instructions; just generate an ir_constant.
275 * Function calls were first allowed to be constant expressions in GLSL 1.20.
277 if (state
->language_version
>= 120) {
278 ir_constant
*value
= sig
->constant_expression_value(actual_parameters
);
284 ir_dereference_variable
*deref
= NULL
;
285 if (!sig
->return_type
->is_void()) {
286 /* Create a new temporary to hold the return value. */
289 var
= new(ctx
) ir_variable(sig
->return_type
,
290 ralloc_asprintf(ctx
, "%s_retval",
291 sig
->function_name()),
293 instructions
->push_tail(var
);
295 deref
= new(ctx
) ir_dereference_variable(var
);
297 ir_call
*call
= new(ctx
) ir_call(sig
, deref
, actual_parameters
);
298 instructions
->push_tail(call
);
300 /* Also emit any necessary out-parameter conversions. */
301 instructions
->append_list(&post_call_conversions
);
303 return deref
? deref
->clone(ctx
, NULL
) : NULL
;
307 * Given a function name and parameter list, find the matching signature.
309 static ir_function_signature
*
310 match_function_by_name(const char *name
,
311 exec_list
*actual_parameters
,
312 struct _mesa_glsl_parse_state
*state
)
315 ir_function
*f
= state
->symbols
->get_function(name
);
316 ir_function_signature
*local_sig
= NULL
;
317 ir_function_signature
*sig
= NULL
;
319 /* Is the function hidden by a record type constructor? */
320 if (state
->symbols
->get_type(name
))
321 goto done
; /* no match */
323 /* Is the function hidden by a variable (impossible in 1.10)? */
324 if (state
->language_version
!= 110 && state
->symbols
->get_variable(name
))
325 goto done
; /* no match */
328 /* Look for a match in the local shader. If exact, we're done. */
329 bool is_exact
= false;
330 sig
= local_sig
= f
->matching_signature(actual_parameters
, &is_exact
);
334 if (!state
->es_shader
&& f
->has_user_signature()) {
335 /* In desktop GL, the presence of a user-defined signature hides any
336 * built-in signatures, so we must ignore them. In contrast, in ES2
337 * user-defined signatures add new overloads, so we must proceed.
343 /* Local shader has no exact candidates; check the built-ins. */
344 _mesa_glsl_initialize_functions(state
);
345 for (unsigned i
= 0; i
< state
->num_builtins_to_link
; i
++) {
346 ir_function
*builtin
=
347 state
->builtins_to_link
[i
]->symbols
->get_function(name
);
351 bool is_exact
= false;
352 ir_function_signature
*builtin_sig
=
353 builtin
->matching_signature(actual_parameters
, &is_exact
);
355 if (builtin_sig
== NULL
)
358 /* If the built-in signature is exact, we can stop. */
365 /* We found an inexact match, which is better than nothing. However,
366 * we should keep searching for an exact match.
374 /* If the match is from a linked built-in shader, import the prototype. */
375 if (sig
!= local_sig
) {
377 f
= new(ctx
) ir_function(name
);
378 state
->symbols
->add_global_function(f
);
379 emit_function(state
, f
);
381 f
->add_signature(sig
->clone_prototype(f
, NULL
));
388 * Raise a "no matching function" error, listing all possible overloads the
389 * compiler considered so developers can figure out what went wrong.
392 no_matching_function_error(const char *name
,
394 exec_list
*actual_parameters
,
395 _mesa_glsl_parse_state
*state
)
397 char *str
= prototype_string(NULL
, name
, actual_parameters
);
398 _mesa_glsl_error(loc
, state
, "no matching function for call to `%s'", str
);
401 const char *prefix
= "candidates are: ";
403 for (int i
= -1; i
< (int) state
->num_builtins_to_link
; i
++) {
404 glsl_symbol_table
*syms
= i
>= 0 ? state
->builtins_to_link
[i
]->symbols
406 ir_function
*f
= syms
->get_function(name
);
410 foreach_list (node
, &f
->signatures
) {
411 ir_function_signature
*sig
= (ir_function_signature
*) node
;
413 str
= prototype_string(sig
->return_type
, f
->name
, &sig
->parameters
);
414 _mesa_glsl_error(loc
, state
, "%s%s", prefix
, str
);
423 * Perform automatic type conversion of constructor parameters
425 * This implements the rules in the "Conversion and Scalar Constructors"
426 * section (GLSL 1.10 section 5.4.1), not the "Implicit Conversions" rules.
429 convert_component(ir_rvalue
*src
, const glsl_type
*desired_type
)
431 void *ctx
= ralloc_parent(src
);
432 const unsigned a
= desired_type
->base_type
;
433 const unsigned b
= src
->type
->base_type
;
434 ir_expression
*result
= NULL
;
436 if (src
->type
->is_error())
439 assert(a
<= GLSL_TYPE_BOOL
);
440 assert(b
<= GLSL_TYPE_BOOL
);
449 result
= new(ctx
) ir_expression(ir_unop_i2u
, src
);
451 case GLSL_TYPE_FLOAT
:
452 result
= new(ctx
) ir_expression(ir_unop_i2u
,
453 new(ctx
) ir_expression(ir_unop_f2i
, src
));
456 result
= new(ctx
) ir_expression(ir_unop_i2u
,
457 new(ctx
) ir_expression(ir_unop_b2i
, src
));
464 result
= new(ctx
) ir_expression(ir_unop_u2i
, src
);
466 case GLSL_TYPE_FLOAT
:
467 result
= new(ctx
) ir_expression(ir_unop_f2i
, src
);
470 result
= new(ctx
) ir_expression(ir_unop_b2i
, src
);
474 case GLSL_TYPE_FLOAT
:
477 result
= new(ctx
) ir_expression(ir_unop_u2f
, desired_type
, src
, NULL
);
480 result
= new(ctx
) ir_expression(ir_unop_i2f
, desired_type
, src
, NULL
);
483 result
= new(ctx
) ir_expression(ir_unop_b2f
, desired_type
, src
, NULL
);
490 result
= new(ctx
) ir_expression(ir_unop_i2b
,
491 new(ctx
) ir_expression(ir_unop_u2i
, src
));
494 result
= new(ctx
) ir_expression(ir_unop_i2b
, desired_type
, src
, NULL
);
496 case GLSL_TYPE_FLOAT
:
497 result
= new(ctx
) ir_expression(ir_unop_f2b
, desired_type
, src
, NULL
);
503 assert(result
!= NULL
);
504 assert(result
->type
== desired_type
);
506 /* Try constant folding; it may fold in the conversion we just added. */
507 ir_constant
*const constant
= result
->constant_expression_value();
508 return (constant
!= NULL
) ? (ir_rvalue
*) constant
: (ir_rvalue
*) result
;
512 * Dereference a specific component from a scalar, vector, or matrix
515 dereference_component(ir_rvalue
*src
, unsigned component
)
517 void *ctx
= ralloc_parent(src
);
518 assert(component
< src
->type
->components());
520 /* If the source is a constant, just create a new constant instead of a
521 * dereference of the existing constant.
523 ir_constant
*constant
= src
->as_constant();
525 return new(ctx
) ir_constant(constant
, component
);
527 if (src
->type
->is_scalar()) {
529 } else if (src
->type
->is_vector()) {
530 return new(ctx
) ir_swizzle(src
, component
, 0, 0, 0, 1);
532 assert(src
->type
->is_matrix());
534 /* Dereference a row of the matrix, then call this function again to get
535 * a specific element from that row.
537 const int c
= component
/ src
->type
->column_type()->vector_elements
;
538 const int r
= component
% src
->type
->column_type()->vector_elements
;
539 ir_constant
*const col_index
= new(ctx
) ir_constant(c
);
540 ir_dereference
*const col
= new(ctx
) ir_dereference_array(src
, col_index
);
542 col
->type
= src
->type
->column_type();
544 return dereference_component(col
, r
);
547 assert(!"Should not get here.");
553 process_array_constructor(exec_list
*instructions
,
554 const glsl_type
*constructor_type
,
555 YYLTYPE
*loc
, exec_list
*parameters
,
556 struct _mesa_glsl_parse_state
*state
)
559 /* Array constructors come in two forms: sized and unsized. Sized array
560 * constructors look like 'vec4[2](a, b)', where 'a' and 'b' are vec4
561 * variables. In this case the number of parameters must exactly match the
562 * specified size of the array.
564 * Unsized array constructors look like 'vec4[](a, b)', where 'a' and 'b'
565 * are vec4 variables. In this case the size of the array being constructed
566 * is determined by the number of parameters.
568 * From page 52 (page 58 of the PDF) of the GLSL 1.50 spec:
570 * "There must be exactly the same number of arguments as the size of
571 * the array being constructed. If no size is present in the
572 * constructor, then the array is explicitly sized to the number of
573 * arguments provided. The arguments are assigned in order, starting at
574 * element 0, to the elements of the constructed array. Each argument
575 * must be the same type as the element type of the array, or be a type
576 * that can be converted to the element type of the array according to
577 * Section 4.1.10 "Implicit Conversions.""
579 exec_list actual_parameters
;
580 const unsigned parameter_count
=
581 process_parameters(instructions
, &actual_parameters
, parameters
, state
);
583 if ((parameter_count
== 0)
584 || ((constructor_type
->length
!= 0)
585 && (constructor_type
->length
!= parameter_count
))) {
586 const unsigned min_param
= (constructor_type
->length
== 0)
587 ? 1 : constructor_type
->length
;
589 _mesa_glsl_error(loc
, state
, "array constructor must have %s %u "
591 (constructor_type
->length
!= 0) ? "at least" : "exactly",
592 min_param
, (min_param
<= 1) ? "" : "s");
593 return ir_rvalue::error_value(ctx
);
596 if (constructor_type
->length
== 0) {
598 glsl_type::get_array_instance(constructor_type
->element_type(),
600 assert(constructor_type
!= NULL
);
601 assert(constructor_type
->length
== parameter_count
);
604 bool all_parameters_are_constant
= true;
606 /* Type cast each parameter and, if possible, fold constants. */
607 foreach_list_safe(n
, &actual_parameters
) {
608 ir_rvalue
*ir
= (ir_rvalue
*) n
;
609 ir_rvalue
*result
= ir
;
611 /* Apply implicit conversions (not the scalar constructor rules!). See
612 * the spec quote above. */
613 if (constructor_type
->element_type()->is_float()) {
614 const glsl_type
*desired_type
=
615 glsl_type::get_instance(GLSL_TYPE_FLOAT
,
616 ir
->type
->vector_elements
,
617 ir
->type
->matrix_columns
);
618 if (result
->type
->can_implicitly_convert_to(desired_type
)) {
619 /* Even though convert_component() implements the constructor
620 * conversion rules (not the implicit conversion rules), its safe
621 * to use it here because we already checked that the implicit
622 * conversion is legal.
624 result
= convert_component(ir
, desired_type
);
628 if (result
->type
!= constructor_type
->element_type()) {
629 _mesa_glsl_error(loc
, state
, "type error in array constructor: "
630 "expected: %s, found %s",
631 constructor_type
->element_type()->name
,
635 /* Attempt to convert the parameter to a constant valued expression.
636 * After doing so, track whether or not all the parameters to the
637 * constructor are trivially constant valued expressions.
639 ir_rvalue
*const constant
= result
->constant_expression_value();
641 if (constant
!= NULL
)
644 all_parameters_are_constant
= false;
646 ir
->replace_with(result
);
649 if (all_parameters_are_constant
)
650 return new(ctx
) ir_constant(constructor_type
, &actual_parameters
);
652 ir_variable
*var
= new(ctx
) ir_variable(constructor_type
, "array_ctor",
654 instructions
->push_tail(var
);
657 foreach_list(node
, &actual_parameters
) {
658 ir_rvalue
*rhs
= (ir_rvalue
*) node
;
659 ir_rvalue
*lhs
= new(ctx
) ir_dereference_array(var
,
660 new(ctx
) ir_constant(i
));
662 ir_instruction
*assignment
= new(ctx
) ir_assignment(lhs
, rhs
, NULL
);
663 instructions
->push_tail(assignment
);
668 return new(ctx
) ir_dereference_variable(var
);
673 * Try to convert a record constructor to a constant expression
676 constant_record_constructor(const glsl_type
*constructor_type
,
677 exec_list
*parameters
, void *mem_ctx
)
679 foreach_list(node
, parameters
) {
680 ir_constant
*constant
= ((ir_instruction
*) node
)->as_constant();
681 if (constant
== NULL
)
683 node
->replace_with(constant
);
686 return new(mem_ctx
) ir_constant(constructor_type
, parameters
);
691 * Determine if a list consists of a single scalar r-value
694 single_scalar_parameter(exec_list
*parameters
)
696 const ir_rvalue
*const p
= (ir_rvalue
*) parameters
->head
;
697 assert(((ir_rvalue
*)p
)->as_rvalue() != NULL
);
699 return (p
->type
->is_scalar() && p
->next
->is_tail_sentinel());
704 * Generate inline code for a vector constructor
706 * The generated constructor code will consist of a temporary variable
707 * declaration of the same type as the constructor. A sequence of assignments
708 * from constructor parameters to the temporary will follow.
711 * An \c ir_dereference_variable of the temprorary generated in the constructor
715 emit_inline_vector_constructor(const glsl_type
*type
,
716 exec_list
*instructions
,
717 exec_list
*parameters
,
720 assert(!parameters
->is_empty());
722 ir_variable
*var
= new(ctx
) ir_variable(type
, "vec_ctor", ir_var_temporary
);
723 instructions
->push_tail(var
);
725 /* There are two kinds of vector constructors.
727 * - Construct a vector from a single scalar by replicating that scalar to
728 * all components of the vector.
730 * - Construct a vector from an arbirary combination of vectors and
731 * scalars. The components of the constructor parameters are assigned
732 * to the vector in order until the vector is full.
734 const unsigned lhs_components
= type
->components();
735 if (single_scalar_parameter(parameters
)) {
736 ir_rvalue
*first_param
= (ir_rvalue
*)parameters
->head
;
737 ir_rvalue
*rhs
= new(ctx
) ir_swizzle(first_param
, 0, 0, 0, 0,
739 ir_dereference_variable
*lhs
= new(ctx
) ir_dereference_variable(var
);
740 const unsigned mask
= (1U << lhs_components
) - 1;
742 assert(rhs
->type
== lhs
->type
);
744 ir_instruction
*inst
= new(ctx
) ir_assignment(lhs
, rhs
, NULL
, mask
);
745 instructions
->push_tail(inst
);
747 unsigned base_component
= 0;
748 unsigned base_lhs_component
= 0;
749 ir_constant_data data
;
750 unsigned constant_mask
= 0, constant_components
= 0;
752 memset(&data
, 0, sizeof(data
));
754 foreach_list(node
, parameters
) {
755 ir_rvalue
*param
= (ir_rvalue
*) node
;
756 unsigned rhs_components
= param
->type
->components();
758 /* Do not try to assign more components to the vector than it has!
760 if ((rhs_components
+ base_lhs_component
) > lhs_components
) {
761 rhs_components
= lhs_components
- base_lhs_component
;
764 const ir_constant
*const c
= param
->as_constant();
766 for (unsigned i
= 0; i
< rhs_components
; i
++) {
767 switch (c
->type
->base_type
) {
769 data
.u
[i
+ base_component
] = c
->get_uint_component(i
);
772 data
.i
[i
+ base_component
] = c
->get_int_component(i
);
774 case GLSL_TYPE_FLOAT
:
775 data
.f
[i
+ base_component
] = c
->get_float_component(i
);
778 data
.b
[i
+ base_component
] = c
->get_bool_component(i
);
781 assert(!"Should not get here.");
786 /* Mask of fields to be written in the assignment.
788 constant_mask
|= ((1U << rhs_components
) - 1) << base_lhs_component
;
789 constant_components
+= rhs_components
;
791 base_component
+= rhs_components
;
793 /* Advance the component index by the number of components
794 * that were just assigned.
796 base_lhs_component
+= rhs_components
;
799 if (constant_mask
!= 0) {
800 ir_dereference
*lhs
= new(ctx
) ir_dereference_variable(var
);
801 const glsl_type
*rhs_type
= glsl_type::get_instance(var
->type
->base_type
,
804 ir_rvalue
*rhs
= new(ctx
) ir_constant(rhs_type
, &data
);
806 ir_instruction
*inst
=
807 new(ctx
) ir_assignment(lhs
, rhs
, NULL
, constant_mask
);
808 instructions
->push_tail(inst
);
812 foreach_list(node
, parameters
) {
813 ir_rvalue
*param
= (ir_rvalue
*) node
;
814 unsigned rhs_components
= param
->type
->components();
816 /* Do not try to assign more components to the vector than it has!
818 if ((rhs_components
+ base_component
) > lhs_components
) {
819 rhs_components
= lhs_components
- base_component
;
822 const ir_constant
*const c
= param
->as_constant();
824 /* Mask of fields to be written in the assignment.
826 const unsigned write_mask
= ((1U << rhs_components
) - 1)
829 ir_dereference
*lhs
= new(ctx
) ir_dereference_variable(var
);
831 /* Generate a swizzle so that LHS and RHS sizes match.
834 new(ctx
) ir_swizzle(param
, 0, 1, 2, 3, rhs_components
);
836 ir_instruction
*inst
=
837 new(ctx
) ir_assignment(lhs
, rhs
, NULL
, write_mask
);
838 instructions
->push_tail(inst
);
841 /* Advance the component index by the number of components that were
844 base_component
+= rhs_components
;
847 return new(ctx
) ir_dereference_variable(var
);
852 * Generate assignment of a portion of a vector to a portion of a matrix column
854 * \param src_base First component of the source to be used in assignment
855 * \param column Column of destination to be assiged
856 * \param row_base First component of the destination column to be assigned
857 * \param count Number of components to be assigned
860 * \c src_base + \c count must be less than or equal to the number of components
861 * in the source vector.
864 assign_to_matrix_column(ir_variable
*var
, unsigned column
, unsigned row_base
,
865 ir_rvalue
*src
, unsigned src_base
, unsigned count
,
868 ir_constant
*col_idx
= new(mem_ctx
) ir_constant(column
);
869 ir_dereference
*column_ref
= new(mem_ctx
) ir_dereference_array(var
, col_idx
);
871 assert(column_ref
->type
->components() >= (row_base
+ count
));
872 assert(src
->type
->components() >= (src_base
+ count
));
874 /* Generate a swizzle that extracts the number of components from the source
875 * that are to be assigned to the column of the matrix.
877 if (count
< src
->type
->vector_elements
) {
878 src
= new(mem_ctx
) ir_swizzle(src
,
879 src_base
+ 0, src_base
+ 1,
880 src_base
+ 2, src_base
+ 3,
884 /* Mask of fields to be written in the assignment.
886 const unsigned write_mask
= ((1U << count
) - 1) << row_base
;
888 return new(mem_ctx
) ir_assignment(column_ref
, src
, NULL
, write_mask
);
893 * Generate inline code for a matrix constructor
895 * The generated constructor code will consist of a temporary variable
896 * declaration of the same type as the constructor. A sequence of assignments
897 * from constructor parameters to the temporary will follow.
900 * An \c ir_dereference_variable of the temprorary generated in the constructor
904 emit_inline_matrix_constructor(const glsl_type
*type
,
905 exec_list
*instructions
,
906 exec_list
*parameters
,
909 assert(!parameters
->is_empty());
911 ir_variable
*var
= new(ctx
) ir_variable(type
, "mat_ctor", ir_var_temporary
);
912 instructions
->push_tail(var
);
914 /* There are three kinds of matrix constructors.
916 * - Construct a matrix from a single scalar by replicating that scalar to
917 * along the diagonal of the matrix and setting all other components to
920 * - Construct a matrix from an arbirary combination of vectors and
921 * scalars. The components of the constructor parameters are assigned
922 * to the matrix in colum-major order until the matrix is full.
924 * - Construct a matrix from a single matrix. The source matrix is copied
925 * to the upper left portion of the constructed matrix, and the remaining
926 * elements take values from the identity matrix.
928 ir_rvalue
*const first_param
= (ir_rvalue
*) parameters
->head
;
929 if (single_scalar_parameter(parameters
)) {
930 /* Assign the scalar to the X component of a vec4, and fill the remaining
931 * components with zero.
933 ir_variable
*rhs_var
=
934 new(ctx
) ir_variable(glsl_type::vec4_type
, "mat_ctor_vec",
936 instructions
->push_tail(rhs_var
);
938 ir_constant_data zero
;
944 ir_instruction
*inst
=
945 new(ctx
) ir_assignment(new(ctx
) ir_dereference_variable(rhs_var
),
946 new(ctx
) ir_constant(rhs_var
->type
, &zero
),
948 instructions
->push_tail(inst
);
950 ir_dereference
*const rhs_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
952 inst
= new(ctx
) ir_assignment(rhs_ref
, first_param
, NULL
, 0x01);
953 instructions
->push_tail(inst
);
955 /* Assign the temporary vector to each column of the destination matrix
956 * with a swizzle that puts the X component on the diagonal of the
957 * matrix. In some cases this may mean that the X component does not
958 * get assigned into the column at all (i.e., when the matrix has more
959 * columns than rows).
961 static const unsigned rhs_swiz
[4][4] = {
968 const unsigned cols_to_init
= MIN2(type
->matrix_columns
,
969 type
->vector_elements
);
970 for (unsigned i
= 0; i
< cols_to_init
; i
++) {
971 ir_constant
*const col_idx
= new(ctx
) ir_constant(i
);
972 ir_rvalue
*const col_ref
= new(ctx
) ir_dereference_array(var
, col_idx
);
974 ir_rvalue
*const rhs_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
975 ir_rvalue
*const rhs
= new(ctx
) ir_swizzle(rhs_ref
, rhs_swiz
[i
],
976 type
->vector_elements
);
978 inst
= new(ctx
) ir_assignment(col_ref
, rhs
, NULL
);
979 instructions
->push_tail(inst
);
982 for (unsigned i
= cols_to_init
; i
< type
->matrix_columns
; i
++) {
983 ir_constant
*const col_idx
= new(ctx
) ir_constant(i
);
984 ir_rvalue
*const col_ref
= new(ctx
) ir_dereference_array(var
, col_idx
);
986 ir_rvalue
*const rhs_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
987 ir_rvalue
*const rhs
= new(ctx
) ir_swizzle(rhs_ref
, 1, 1, 1, 1,
988 type
->vector_elements
);
990 inst
= new(ctx
) ir_assignment(col_ref
, rhs
, NULL
);
991 instructions
->push_tail(inst
);
993 } else if (first_param
->type
->is_matrix()) {
994 /* From page 50 (56 of the PDF) of the GLSL 1.50 spec:
996 * "If a matrix is constructed from a matrix, then each component
997 * (column i, row j) in the result that has a corresponding
998 * component (column i, row j) in the argument will be initialized
999 * from there. All other components will be initialized to the
1000 * identity matrix. If a matrix argument is given to a matrix
1001 * constructor, it is an error to have any other arguments."
1003 assert(first_param
->next
->is_tail_sentinel());
1004 ir_rvalue
*const src_matrix
= first_param
;
1006 /* If the source matrix is smaller, pre-initialize the relavent parts of
1007 * the destination matrix to the identity matrix.
1009 if ((src_matrix
->type
->matrix_columns
< var
->type
->matrix_columns
)
1010 || (src_matrix
->type
->vector_elements
< var
->type
->vector_elements
)) {
1012 /* If the source matrix has fewer rows, every column of the destination
1013 * must be initialized. Otherwise only the columns in the destination
1014 * that do not exist in the source must be initialized.
1017 (src_matrix
->type
->vector_elements
< var
->type
->vector_elements
)
1018 ? 0 : src_matrix
->type
->matrix_columns
;
1020 const glsl_type
*const col_type
= var
->type
->column_type();
1021 for (/* empty */; col
< var
->type
->matrix_columns
; col
++) {
1022 ir_constant_data ident
;
1031 ir_rvalue
*const rhs
= new(ctx
) ir_constant(col_type
, &ident
);
1033 ir_rvalue
*const lhs
=
1034 new(ctx
) ir_dereference_array(var
, new(ctx
) ir_constant(col
));
1036 ir_instruction
*inst
= new(ctx
) ir_assignment(lhs
, rhs
, NULL
);
1037 instructions
->push_tail(inst
);
1041 /* Assign columns from the source matrix to the destination matrix.
1043 * Since the parameter will be used in the RHS of multiple assignments,
1044 * generate a temporary and copy the paramter there.
1046 ir_variable
*const rhs_var
=
1047 new(ctx
) ir_variable(first_param
->type
, "mat_ctor_mat",
1049 instructions
->push_tail(rhs_var
);
1051 ir_dereference
*const rhs_var_ref
=
1052 new(ctx
) ir_dereference_variable(rhs_var
);
1053 ir_instruction
*const inst
=
1054 new(ctx
) ir_assignment(rhs_var_ref
, first_param
, NULL
);
1055 instructions
->push_tail(inst
);
1057 const unsigned last_row
= MIN2(src_matrix
->type
->vector_elements
,
1058 var
->type
->vector_elements
);
1059 const unsigned last_col
= MIN2(src_matrix
->type
->matrix_columns
,
1060 var
->type
->matrix_columns
);
1062 unsigned swiz
[4] = { 0, 0, 0, 0 };
1063 for (unsigned i
= 1; i
< last_row
; i
++)
1066 const unsigned write_mask
= (1U << last_row
) - 1;
1068 for (unsigned i
= 0; i
< last_col
; i
++) {
1069 ir_dereference
*const lhs
=
1070 new(ctx
) ir_dereference_array(var
, new(ctx
) ir_constant(i
));
1071 ir_rvalue
*const rhs_col
=
1072 new(ctx
) ir_dereference_array(rhs_var
, new(ctx
) ir_constant(i
));
1074 /* If one matrix has columns that are smaller than the columns of the
1075 * other matrix, wrap the column access of the larger with a swizzle
1076 * so that the LHS and RHS of the assignment have the same size (and
1077 * therefore have the same type).
1079 * It would be perfectly valid to unconditionally generate the
1080 * swizzles, this this will typically result in a more compact IR tree.
1083 if (lhs
->type
->vector_elements
!= rhs_col
->type
->vector_elements
) {
1084 rhs
= new(ctx
) ir_swizzle(rhs_col
, swiz
, last_row
);
1089 ir_instruction
*inst
=
1090 new(ctx
) ir_assignment(lhs
, rhs
, NULL
, write_mask
);
1091 instructions
->push_tail(inst
);
1094 const unsigned cols
= type
->matrix_columns
;
1095 const unsigned rows
= type
->vector_elements
;
1096 unsigned col_idx
= 0;
1097 unsigned row_idx
= 0;
1099 foreach_list (node
, parameters
) {
1100 ir_rvalue
*const rhs
= (ir_rvalue
*) node
;
1101 const unsigned components_remaining_this_column
= rows
- row_idx
;
1102 unsigned rhs_components
= rhs
->type
->components();
1103 unsigned rhs_base
= 0;
1105 /* Since the parameter might be used in the RHS of two assignments,
1106 * generate a temporary and copy the paramter there.
1108 ir_variable
*rhs_var
=
1109 new(ctx
) ir_variable(rhs
->type
, "mat_ctor_vec", ir_var_temporary
);
1110 instructions
->push_tail(rhs_var
);
1112 ir_dereference
*rhs_var_ref
=
1113 new(ctx
) ir_dereference_variable(rhs_var
);
1114 ir_instruction
*inst
= new(ctx
) ir_assignment(rhs_var_ref
, rhs
, NULL
);
1115 instructions
->push_tail(inst
);
1117 /* Assign the current parameter to as many components of the matrix
1120 * NOTE: A single vector parameter can span two matrix columns. A
1121 * single vec4, for example, can completely fill a mat2.
1123 if (rhs_components
>= components_remaining_this_column
) {
1124 const unsigned count
= MIN2(rhs_components
,
1125 components_remaining_this_column
);
1127 rhs_var_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
1129 ir_instruction
*inst
= assign_to_matrix_column(var
, col_idx
,
1133 instructions
->push_tail(inst
);
1141 /* If there is data left in the parameter and components left to be
1142 * set in the destination, emit another assignment. It is possible
1143 * that the assignment could be of a vec4 to the last element of the
1144 * matrix. In this case col_idx==cols, but there is still data
1145 * left in the source parameter. Obviously, don't emit an assignment
1146 * to data outside the destination matrix.
1148 if ((col_idx
< cols
) && (rhs_base
< rhs_components
)) {
1149 const unsigned count
= rhs_components
- rhs_base
;
1151 rhs_var_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
1153 ir_instruction
*inst
= assign_to_matrix_column(var
, col_idx
,
1158 instructions
->push_tail(inst
);
1165 return new(ctx
) ir_dereference_variable(var
);
1170 emit_inline_record_constructor(const glsl_type
*type
,
1171 exec_list
*instructions
,
1172 exec_list
*parameters
,
1175 ir_variable
*const var
=
1176 new(mem_ctx
) ir_variable(type
, "record_ctor", ir_var_temporary
);
1177 ir_dereference_variable
*const d
= new(mem_ctx
) ir_dereference_variable(var
);
1179 instructions
->push_tail(var
);
1181 exec_node
*node
= parameters
->head
;
1182 for (unsigned i
= 0; i
< type
->length
; i
++) {
1183 assert(!node
->is_tail_sentinel());
1185 ir_dereference
*const lhs
=
1186 new(mem_ctx
) ir_dereference_record(d
->clone(mem_ctx
, NULL
),
1187 type
->fields
.structure
[i
].name
);
1189 ir_rvalue
*const rhs
= ((ir_instruction
*) node
)->as_rvalue();
1190 assert(rhs
!= NULL
);
1192 ir_instruction
*const assign
= new(mem_ctx
) ir_assignment(lhs
, rhs
, NULL
);
1194 instructions
->push_tail(assign
);
1203 ast_function_expression::hir(exec_list
*instructions
,
1204 struct _mesa_glsl_parse_state
*state
)
1207 /* There are three sorts of function calls.
1209 * 1. constructors - The first subexpression is an ast_type_specifier.
1210 * 2. methods - Only the .length() method of array types.
1211 * 3. functions - Calls to regular old functions.
1213 * Method calls are actually detected when the ast_field_selection
1214 * expression is handled.
1216 if (is_constructor()) {
1217 const ast_type_specifier
*type
= (ast_type_specifier
*) subexpressions
[0];
1218 YYLTYPE loc
= type
->get_location();
1221 const glsl_type
*const constructor_type
= type
->glsl_type(& name
, state
);
1223 /* constructor_type can be NULL if a variable with the same name as the
1224 * structure has come into scope.
1226 if (constructor_type
== NULL
) {
1227 _mesa_glsl_error(& loc
, state
, "unknown type `%s' (structure name "
1228 "may be shadowed by a variable with the same name)",
1230 return ir_rvalue::error_value(ctx
);
1234 /* Constructors for samplers are illegal.
1236 if (constructor_type
->is_sampler()) {
1237 _mesa_glsl_error(& loc
, state
, "cannot construct sampler type `%s'",
1238 constructor_type
->name
);
1239 return ir_rvalue::error_value(ctx
);
1242 if (constructor_type
->is_array()) {
1243 if (state
->language_version
<= 110) {
1244 _mesa_glsl_error(& loc
, state
,
1245 "array constructors forbidden in GLSL 1.10");
1246 return ir_rvalue::error_value(ctx
);
1249 return process_array_constructor(instructions
, constructor_type
,
1250 & loc
, &this->expressions
, state
);
1254 /* There are two kinds of constructor call. Constructors for built-in
1255 * language types, such as mat4 and vec2, are free form. The only
1256 * requirement is that the parameters must provide enough values of the
1257 * correct scalar type. Constructors for arrays and structures must
1258 * have the exact number of parameters with matching types in the
1259 * correct order. These constructors follow essentially the same type
1260 * matching rules as functions.
1262 if (constructor_type
->is_record()) {
1263 exec_list actual_parameters
;
1265 process_parameters(instructions
, &actual_parameters
,
1266 &this->expressions
, state
);
1268 exec_node
*node
= actual_parameters
.head
;
1269 for (unsigned i
= 0; i
< constructor_type
->length
; i
++) {
1270 ir_rvalue
*ir
= (ir_rvalue
*) node
;
1272 if (node
->is_tail_sentinel()) {
1273 _mesa_glsl_error(&loc
, state
,
1274 "insufficient parameters to constructor "
1276 constructor_type
->name
);
1277 return ir_rvalue::error_value(ctx
);
1280 if (apply_implicit_conversion(constructor_type
->fields
.structure
[i
].type
,
1282 node
->replace_with(ir
);
1284 _mesa_glsl_error(&loc
, state
,
1285 "parameter type mismatch in constructor "
1286 "for `%s.%s' (%s vs %s)",
1287 constructor_type
->name
,
1288 constructor_type
->fields
.structure
[i
].name
,
1290 constructor_type
->fields
.structure
[i
].type
->name
);
1291 return ir_rvalue::error_value(ctx
);;
1297 if (!node
->is_tail_sentinel()) {
1298 _mesa_glsl_error(&loc
, state
, "too many parameters in constructor "
1299 "for `%s'", constructor_type
->name
);
1300 return ir_rvalue::error_value(ctx
);
1303 ir_rvalue
*const constant
=
1304 constant_record_constructor(constructor_type
, &actual_parameters
,
1307 return (constant
!= NULL
)
1309 : emit_inline_record_constructor(constructor_type
, instructions
,
1310 &actual_parameters
, state
);
1313 if (!constructor_type
->is_numeric() && !constructor_type
->is_boolean())
1314 return ir_rvalue::error_value(ctx
);
1316 /* Total number of components of the type being constructed. */
1317 const unsigned type_components
= constructor_type
->components();
1319 /* Number of components from parameters that have actually been
1320 * consumed. This is used to perform several kinds of error checking.
1322 unsigned components_used
= 0;
1324 unsigned matrix_parameters
= 0;
1325 unsigned nonmatrix_parameters
= 0;
1326 exec_list actual_parameters
;
1328 foreach_list (n
, &this->expressions
) {
1329 ast_node
*ast
= exec_node_data(ast_node
, n
, link
);
1330 ir_rvalue
*result
= ast
->hir(instructions
, state
)->as_rvalue();
1332 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
1334 * "It is an error to provide extra arguments beyond this
1335 * last used argument."
1337 if (components_used
>= type_components
) {
1338 _mesa_glsl_error(& loc
, state
, "too many parameters to `%s' "
1340 constructor_type
->name
);
1341 return ir_rvalue::error_value(ctx
);
1344 if (!result
->type
->is_numeric() && !result
->type
->is_boolean()) {
1345 _mesa_glsl_error(& loc
, state
, "cannot construct `%s' from a "
1346 "non-numeric data type",
1347 constructor_type
->name
);
1348 return ir_rvalue::error_value(ctx
);
1351 /* Count the number of matrix and nonmatrix parameters. This
1352 * is used below to enforce some of the constructor rules.
1354 if (result
->type
->is_matrix())
1355 matrix_parameters
++;
1357 nonmatrix_parameters
++;
1359 actual_parameters
.push_tail(result
);
1360 components_used
+= result
->type
->components();
1363 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
1365 * "It is an error to construct matrices from other matrices. This
1366 * is reserved for future use."
1368 if (state
->language_version
== 110 && matrix_parameters
> 0
1369 && constructor_type
->is_matrix()) {
1370 _mesa_glsl_error(& loc
, state
, "cannot construct `%s' from a "
1371 "matrix in GLSL 1.10",
1372 constructor_type
->name
);
1373 return ir_rvalue::error_value(ctx
);
1376 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
1378 * "If a matrix argument is given to a matrix constructor, it is
1379 * an error to have any other arguments."
1381 if ((matrix_parameters
> 0)
1382 && ((matrix_parameters
+ nonmatrix_parameters
) > 1)
1383 && constructor_type
->is_matrix()) {
1384 _mesa_glsl_error(& loc
, state
, "for matrix `%s' constructor, "
1385 "matrix must be only parameter",
1386 constructor_type
->name
);
1387 return ir_rvalue::error_value(ctx
);
1390 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
1392 * "In these cases, there must be enough components provided in the
1393 * arguments to provide an initializer for every component in the
1394 * constructed value."
1396 if (components_used
< type_components
&& components_used
!= 1
1397 && matrix_parameters
== 0) {
1398 _mesa_glsl_error(& loc
, state
, "too few components to construct "
1400 constructor_type
->name
);
1401 return ir_rvalue::error_value(ctx
);
1404 /* Later, we cast each parameter to the same base type as the
1405 * constructor. Since there are no non-floating point matrices, we
1406 * need to break them up into a series of column vectors.
1408 if (constructor_type
->base_type
!= GLSL_TYPE_FLOAT
) {
1409 foreach_list_safe(n
, &actual_parameters
) {
1410 ir_rvalue
*matrix
= (ir_rvalue
*) n
;
1412 if (!matrix
->type
->is_matrix())
1415 /* Create a temporary containing the matrix. */
1416 ir_variable
*var
= new(ctx
) ir_variable(matrix
->type
, "matrix_tmp",
1418 instructions
->push_tail(var
);
1419 instructions
->push_tail(new(ctx
) ir_assignment(new(ctx
)
1420 ir_dereference_variable(var
), matrix
, NULL
));
1421 var
->constant_value
= matrix
->constant_expression_value();
1423 /* Replace the matrix with dereferences of its columns. */
1424 for (int i
= 0; i
< matrix
->type
->matrix_columns
; i
++) {
1425 matrix
->insert_before(new (ctx
) ir_dereference_array(var
,
1426 new(ctx
) ir_constant(i
)));
1432 bool all_parameters_are_constant
= true;
1434 /* Type cast each parameter and, if possible, fold constants.*/
1435 foreach_list_safe(n
, &actual_parameters
) {
1436 ir_rvalue
*ir
= (ir_rvalue
*) n
;
1438 const glsl_type
*desired_type
=
1439 glsl_type::get_instance(constructor_type
->base_type
,
1440 ir
->type
->vector_elements
,
1441 ir
->type
->matrix_columns
);
1442 ir_rvalue
*result
= convert_component(ir
, desired_type
);
1444 /* Attempt to convert the parameter to a constant valued expression.
1445 * After doing so, track whether or not all the parameters to the
1446 * constructor are trivially constant valued expressions.
1448 ir_rvalue
*const constant
= result
->constant_expression_value();
1450 if (constant
!= NULL
)
1453 all_parameters_are_constant
= false;
1456 ir
->replace_with(result
);
1460 /* If all of the parameters are trivially constant, create a
1461 * constant representing the complete collection of parameters.
1463 if (all_parameters_are_constant
) {
1464 return new(ctx
) ir_constant(constructor_type
, &actual_parameters
);
1465 } else if (constructor_type
->is_scalar()) {
1466 return dereference_component((ir_rvalue
*) actual_parameters
.head
,
1468 } else if (constructor_type
->is_vector()) {
1469 return emit_inline_vector_constructor(constructor_type
,
1474 assert(constructor_type
->is_matrix());
1475 return emit_inline_matrix_constructor(constructor_type
,
1481 const ast_expression
*id
= subexpressions
[0];
1482 const char *func_name
= id
->primary_expression
.identifier
;
1483 YYLTYPE loc
= id
->get_location();
1484 exec_list actual_parameters
;
1486 process_parameters(instructions
, &actual_parameters
, &this->expressions
,
1489 ir_function_signature
*sig
=
1490 match_function_by_name(func_name
, &actual_parameters
, state
);
1492 ir_call
*call
= NULL
;
1493 ir_rvalue
*value
= NULL
;
1495 no_matching_function_error(func_name
, &loc
, &actual_parameters
, state
);
1496 value
= ir_rvalue::error_value(ctx
);
1497 } else if (!verify_parameter_modes(state
, sig
, actual_parameters
, this->expressions
)) {
1498 /* an error has already been emitted */
1499 value
= ir_rvalue::error_value(ctx
);
1501 value
= generate_call(instructions
, sig
, &loc
, &actual_parameters
,
1508 return ir_rvalue::error_value(ctx
);