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
, ")");
98 match_function_by_name(exec_list
*instructions
, const char *name
,
99 YYLTYPE
*loc
, exec_list
*actual_parameters
,
100 struct _mesa_glsl_parse_state
*state
)
103 ir_function
*f
= state
->symbols
->get_function(name
);
104 ir_function_signature
*sig
;
106 sig
= f
? f
->matching_signature(actual_parameters
) : NULL
;
108 /* FINISHME: This doesn't handle the case where shader X contains a
109 * FINISHME: matching signature but shader X + N contains an _exact_
110 * FINISHME: matching signature.
113 && (f
== NULL
|| state
->es_shader
|| !f
->has_user_signature())
114 && state
->symbols
->get_type(name
) == NULL
115 && (state
->language_version
== 110
116 || state
->symbols
->get_variable(name
) == NULL
)) {
117 /* The current shader doesn't contain a matching function or signature.
118 * Before giving up, look for the prototype in the built-in functions.
120 for (unsigned i
= 0; i
< state
->num_builtins_to_link
; i
++) {
121 ir_function
*builtin
;
122 builtin
= state
->builtins_to_link
[i
]->symbols
->get_function(name
);
123 sig
= builtin
? builtin
->matching_signature(actual_parameters
) : NULL
;
126 f
= new(ctx
) ir_function(name
);
127 state
->symbols
->add_global_function(f
);
128 emit_function(state
, f
);
131 f
->add_signature(sig
->clone_prototype(f
, NULL
));
137 exec_list post_call_conversions
;
140 /* Verify that 'out' and 'inout' actual parameters are lvalues. This
141 * isn't done in ir_function::matching_signature because that function
142 * cannot generate the necessary diagnostics.
144 * Also, validate that 'const_in' formal parameters (an extension of our
145 * IR) correspond to ir_constant actual parameters.
147 * Also, perform implicit conversion of arguments. Note: to implicitly
148 * convert out parameters, we need to place them in a temporary
149 * variable, and do the conversion after the call takes place. Since we
150 * haven't emitted the call yet, we'll place the post-call conversions
151 * in a temporary exec_list, and emit them later.
153 exec_list_iterator actual_iter
= actual_parameters
->iterator();
154 exec_list_iterator formal_iter
= sig
->parameters
.iterator();
156 while (actual_iter
.has_next()) {
157 ir_rvalue
*actual
= (ir_rvalue
*) actual_iter
.get();
158 ir_variable
*formal
= (ir_variable
*) formal_iter
.get();
160 assert(actual
!= NULL
);
161 assert(formal
!= NULL
);
163 if (formal
->mode
== ir_var_const_in
&& !actual
->as_constant()) {
164 _mesa_glsl_error(loc
, state
,
165 "parameter `%s' must be a constant expression",
169 if ((formal
->mode
== ir_var_out
)
170 || (formal
->mode
== ir_var_inout
)) {
171 const char *mode
= NULL
;
172 switch (formal
->mode
) {
173 case ir_var_out
: mode
= "out"; break;
174 case ir_var_inout
: mode
= "inout"; break;
175 default: assert(false); break;
177 /* FIXME: 'loc' is incorrect (as of 2011-01-21). It is always
180 if (actual
->variable_referenced()
181 && actual
->variable_referenced()->read_only
) {
182 _mesa_glsl_error(loc
, state
,
183 "function parameter '%s %s' references the "
184 "read-only variable '%s'",
186 actual
->variable_referenced()->name
);
188 } else if (!actual
->is_lvalue()) {
189 _mesa_glsl_error(loc
, state
,
190 "function parameter '%s %s' is not an lvalue",
195 if (formal
->type
->is_numeric() || formal
->type
->is_boolean()) {
196 switch (formal
->mode
) {
199 = convert_component(actual
, formal
->type
);
200 actual
->replace_with(converted
);
204 if (actual
->type
!= formal
->type
) {
205 /* To convert an out parameter, we need to create a
206 * temporary variable to hold the value before conversion,
207 * and then perform the conversion after the function call
210 * This has the effect of transforming code like this:
216 * Into IR that's equivalent to this:
220 * int out_parameter_conversion;
221 * f(out_parameter_conversion);
222 * value = float(out_parameter_conversion);
225 new(ctx
) ir_variable(formal
->type
,
226 "out_parameter_conversion",
228 instructions
->push_tail(tmp
);
229 ir_dereference_variable
*deref_tmp_1
230 = new(ctx
) ir_dereference_variable(tmp
);
231 ir_dereference_variable
*deref_tmp_2
232 = new(ctx
) ir_dereference_variable(tmp
);
233 ir_rvalue
*converted_tmp
234 = convert_component(deref_tmp_1
, actual
->type
);
235 ir_assignment
*assignment
236 = new(ctx
) ir_assignment(actual
, converted_tmp
);
237 post_call_conversions
.push_tail(assignment
);
238 actual
->replace_with(deref_tmp_2
);
242 /* Inout parameters should never require conversion, since that
243 * would require an implicit conversion to exist both to and
244 * from the formal parameter type, and there are no
245 * bidirectional implicit conversions.
247 assert (actual
->type
== formal
->type
);
250 assert (!"Illegal formal parameter mode");
259 /* Always insert the call in the instruction stream, and return a deref
260 * of its return val if it returns a value, since we don't know if
261 * the rvalue is going to be assigned to anything or not.
263 * Also insert any out parameter conversions after the call.
265 ir_call
*call
= new(ctx
) ir_call(sig
, actual_parameters
);
266 ir_dereference_variable
*deref
;
267 if (!sig
->return_type
->is_void()) {
268 /* If the function call is a constant expression, don't
269 * generate the instructions to call it; just generate an
270 * ir_constant representing the constant value.
272 * Function calls can only be constant expressions starting
275 if (state
->language_version
>= 120) {
276 ir_constant
*const_val
= call
->constant_expression_value();
284 var
= new(ctx
) ir_variable(sig
->return_type
,
285 ralloc_asprintf(ctx
, "%s_retval",
286 sig
->function_name()),
288 instructions
->push_tail(var
);
290 deref
= new(ctx
) ir_dereference_variable(var
);
291 ir_assignment
*assign
= new(ctx
) ir_assignment(deref
, call
, NULL
);
292 instructions
->push_tail(assign
);
294 deref
= new(ctx
) ir_dereference_variable(var
);
296 instructions
->push_tail(call
);
299 instructions
->append_list(&post_call_conversions
);
302 char *str
= prototype_string(NULL
, name
, actual_parameters
);
304 _mesa_glsl_error(loc
, state
, "no matching function for call to `%s'",
308 const char *prefix
= "candidates are: ";
310 for (int i
= -1; i
< (int) state
->num_builtins_to_link
; i
++) {
311 glsl_symbol_table
*syms
= i
>= 0 ? state
->builtins_to_link
[i
]->symbols
313 f
= syms
->get_function(name
);
317 foreach_list (node
, &f
->signatures
) {
318 ir_function_signature
*sig
= (ir_function_signature
*) node
;
320 str
= prototype_string(sig
->return_type
, f
->name
, &sig
->parameters
);
321 _mesa_glsl_error(loc
, state
, "%s%s", prefix
, str
);
329 return ir_call::get_error_instruction(ctx
);
335 * Perform automatic type conversion of constructor parameters
337 * This implements the rules in the "Conversion and Scalar Constructors"
338 * section (GLSL 1.10 section 5.4.1), not the "Implicit Conversions" rules.
341 convert_component(ir_rvalue
*src
, const glsl_type
*desired_type
)
343 void *ctx
= ralloc_parent(src
);
344 const unsigned a
= desired_type
->base_type
;
345 const unsigned b
= src
->type
->base_type
;
346 ir_expression
*result
= NULL
;
348 if (src
->type
->is_error())
351 assert(a
<= GLSL_TYPE_BOOL
);
352 assert(b
<= GLSL_TYPE_BOOL
);
361 result
= new(ctx
) ir_expression(ir_unop_i2u
, src
);
363 case GLSL_TYPE_FLOAT
:
364 result
= new(ctx
) ir_expression(ir_unop_i2u
,
365 new(ctx
) ir_expression(ir_unop_f2i
, src
));
368 result
= new(ctx
) ir_expression(ir_unop_i2u
,
369 new(ctx
) ir_expression(ir_unop_b2i
, src
));
376 result
= new(ctx
) ir_expression(ir_unop_u2i
, src
);
378 case GLSL_TYPE_FLOAT
:
379 result
= new(ctx
) ir_expression(ir_unop_f2i
, src
);
382 result
= new(ctx
) ir_expression(ir_unop_b2i
, src
);
386 case GLSL_TYPE_FLOAT
:
389 result
= new(ctx
) ir_expression(ir_unop_u2f
, desired_type
, src
, NULL
);
392 result
= new(ctx
) ir_expression(ir_unop_i2f
, desired_type
, src
, NULL
);
395 result
= new(ctx
) ir_expression(ir_unop_b2f
, desired_type
, src
, NULL
);
402 result
= new(ctx
) ir_expression(ir_unop_i2b
,
403 new(ctx
) ir_expression(ir_unop_u2i
, src
));
406 result
= new(ctx
) ir_expression(ir_unop_i2b
, desired_type
, src
, NULL
);
408 case GLSL_TYPE_FLOAT
:
409 result
= new(ctx
) ir_expression(ir_unop_f2b
, desired_type
, src
, NULL
);
415 assert(result
!= NULL
);
416 assert(result
->type
== desired_type
);
418 /* Try constant folding; it may fold in the conversion we just added. */
419 ir_constant
*const constant
= result
->constant_expression_value();
420 return (constant
!= NULL
) ? (ir_rvalue
*) constant
: (ir_rvalue
*) result
;
424 * Dereference a specific component from a scalar, vector, or matrix
427 dereference_component(ir_rvalue
*src
, unsigned component
)
429 void *ctx
= ralloc_parent(src
);
430 assert(component
< src
->type
->components());
432 /* If the source is a constant, just create a new constant instead of a
433 * dereference of the existing constant.
435 ir_constant
*constant
= src
->as_constant();
437 return new(ctx
) ir_constant(constant
, component
);
439 if (src
->type
->is_scalar()) {
441 } else if (src
->type
->is_vector()) {
442 return new(ctx
) ir_swizzle(src
, component
, 0, 0, 0, 1);
444 assert(src
->type
->is_matrix());
446 /* Dereference a row of the matrix, then call this function again to get
447 * a specific element from that row.
449 const int c
= component
/ src
->type
->column_type()->vector_elements
;
450 const int r
= component
% src
->type
->column_type()->vector_elements
;
451 ir_constant
*const col_index
= new(ctx
) ir_constant(c
);
452 ir_dereference
*const col
= new(ctx
) ir_dereference_array(src
, col_index
);
454 col
->type
= src
->type
->column_type();
456 return dereference_component(col
, r
);
459 assert(!"Should not get here.");
465 process_array_constructor(exec_list
*instructions
,
466 const glsl_type
*constructor_type
,
467 YYLTYPE
*loc
, exec_list
*parameters
,
468 struct _mesa_glsl_parse_state
*state
)
471 /* Array constructors come in two forms: sized and unsized. Sized array
472 * constructors look like 'vec4[2](a, b)', where 'a' and 'b' are vec4
473 * variables. In this case the number of parameters must exactly match the
474 * specified size of the array.
476 * Unsized array constructors look like 'vec4[](a, b)', where 'a' and 'b'
477 * are vec4 variables. In this case the size of the array being constructed
478 * is determined by the number of parameters.
480 * From page 52 (page 58 of the PDF) of the GLSL 1.50 spec:
482 * "There must be exactly the same number of arguments as the size of
483 * the array being constructed. If no size is present in the
484 * constructor, then the array is explicitly sized to the number of
485 * arguments provided. The arguments are assigned in order, starting at
486 * element 0, to the elements of the constructed array. Each argument
487 * must be the same type as the element type of the array, or be a type
488 * that can be converted to the element type of the array according to
489 * Section 4.1.10 "Implicit Conversions.""
491 exec_list actual_parameters
;
492 const unsigned parameter_count
=
493 process_parameters(instructions
, &actual_parameters
, parameters
, state
);
495 if ((parameter_count
== 0)
496 || ((constructor_type
->length
!= 0)
497 && (constructor_type
->length
!= parameter_count
))) {
498 const unsigned min_param
= (constructor_type
->length
== 0)
499 ? 1 : constructor_type
->length
;
501 _mesa_glsl_error(loc
, state
, "array constructor must have %s %u "
503 (constructor_type
->length
!= 0) ? "at least" : "exactly",
504 min_param
, (min_param
<= 1) ? "" : "s");
505 return ir_call::get_error_instruction(ctx
);
508 if (constructor_type
->length
== 0) {
510 glsl_type::get_array_instance(constructor_type
->element_type(),
512 assert(constructor_type
!= NULL
);
513 assert(constructor_type
->length
== parameter_count
);
516 bool all_parameters_are_constant
= true;
518 /* Type cast each parameter and, if possible, fold constants. */
519 foreach_list_safe(n
, &actual_parameters
) {
520 ir_rvalue
*ir
= (ir_rvalue
*) n
;
521 ir_rvalue
*result
= ir
;
523 /* Apply implicit conversions (not the scalar constructor rules!). See
524 * the spec quote above. */
525 if (constructor_type
->element_type()->is_float()) {
526 const glsl_type
*desired_type
=
527 glsl_type::get_instance(GLSL_TYPE_FLOAT
,
528 ir
->type
->vector_elements
,
529 ir
->type
->matrix_columns
);
530 if (result
->type
->can_implicitly_convert_to(desired_type
)) {
531 /* Even though convert_component() implements the constructor
532 * conversion rules (not the implicit conversion rules), its safe
533 * to use it here because we already checked that the implicit
534 * conversion is legal.
536 result
= convert_component(ir
, desired_type
);
540 if (result
->type
!= constructor_type
->element_type()) {
541 _mesa_glsl_error(loc
, state
, "type error in array constructor: "
542 "expected: %s, found %s",
543 constructor_type
->element_type()->name
,
547 /* Attempt to convert the parameter to a constant valued expression.
548 * After doing so, track whether or not all the parameters to the
549 * constructor are trivially constant valued expressions.
551 ir_rvalue
*const constant
= result
->constant_expression_value();
553 if (constant
!= NULL
)
556 all_parameters_are_constant
= false;
558 ir
->replace_with(result
);
561 if (all_parameters_are_constant
)
562 return new(ctx
) ir_constant(constructor_type
, &actual_parameters
);
564 ir_variable
*var
= new(ctx
) ir_variable(constructor_type
, "array_ctor",
566 instructions
->push_tail(var
);
569 foreach_list(node
, &actual_parameters
) {
570 ir_rvalue
*rhs
= (ir_rvalue
*) node
;
571 ir_rvalue
*lhs
= new(ctx
) ir_dereference_array(var
,
572 new(ctx
) ir_constant(i
));
574 ir_instruction
*assignment
= new(ctx
) ir_assignment(lhs
, rhs
, NULL
);
575 instructions
->push_tail(assignment
);
580 return new(ctx
) ir_dereference_variable(var
);
585 * Try to convert a record constructor to a constant expression
588 constant_record_constructor(const glsl_type
*constructor_type
,
589 exec_list
*parameters
, void *mem_ctx
)
591 foreach_list(node
, parameters
) {
592 ir_constant
*constant
= ((ir_instruction
*) node
)->as_constant();
593 if (constant
== NULL
)
595 node
->replace_with(constant
);
598 return new(mem_ctx
) ir_constant(constructor_type
, parameters
);
603 * Determine if a list consists of a single scalar r-value
606 single_scalar_parameter(exec_list
*parameters
)
608 const ir_rvalue
*const p
= (ir_rvalue
*) parameters
->head
;
609 assert(((ir_rvalue
*)p
)->as_rvalue() != NULL
);
611 return (p
->type
->is_scalar() && p
->next
->is_tail_sentinel());
616 * Generate inline code for a vector constructor
618 * The generated constructor code will consist of a temporary variable
619 * declaration of the same type as the constructor. A sequence of assignments
620 * from constructor parameters to the temporary will follow.
623 * An \c ir_dereference_variable of the temprorary generated in the constructor
627 emit_inline_vector_constructor(const glsl_type
*type
,
628 exec_list
*instructions
,
629 exec_list
*parameters
,
632 assert(!parameters
->is_empty());
634 ir_variable
*var
= new(ctx
) ir_variable(type
, "vec_ctor", ir_var_temporary
);
635 instructions
->push_tail(var
);
637 /* There are two kinds of vector constructors.
639 * - Construct a vector from a single scalar by replicating that scalar to
640 * all components of the vector.
642 * - Construct a vector from an arbirary combination of vectors and
643 * scalars. The components of the constructor parameters are assigned
644 * to the vector in order until the vector is full.
646 const unsigned lhs_components
= type
->components();
647 if (single_scalar_parameter(parameters
)) {
648 ir_rvalue
*first_param
= (ir_rvalue
*)parameters
->head
;
649 ir_rvalue
*rhs
= new(ctx
) ir_swizzle(first_param
, 0, 0, 0, 0,
651 ir_dereference_variable
*lhs
= new(ctx
) ir_dereference_variable(var
);
652 const unsigned mask
= (1U << lhs_components
) - 1;
654 assert(rhs
->type
== lhs
->type
);
656 ir_instruction
*inst
= new(ctx
) ir_assignment(lhs
, rhs
, NULL
, mask
);
657 instructions
->push_tail(inst
);
659 unsigned base_component
= 0;
660 unsigned base_lhs_component
= 0;
661 ir_constant_data data
;
662 unsigned constant_mask
= 0, constant_components
= 0;
664 memset(&data
, 0, sizeof(data
));
666 foreach_list(node
, parameters
) {
667 ir_rvalue
*param
= (ir_rvalue
*) node
;
668 unsigned rhs_components
= param
->type
->components();
670 /* Do not try to assign more components to the vector than it has!
672 if ((rhs_components
+ base_lhs_component
) > lhs_components
) {
673 rhs_components
= lhs_components
- base_lhs_component
;
676 const ir_constant
*const c
= param
->as_constant();
678 for (unsigned i
= 0; i
< rhs_components
; i
++) {
679 switch (c
->type
->base_type
) {
681 data
.u
[i
+ base_component
] = c
->get_uint_component(i
);
684 data
.i
[i
+ base_component
] = c
->get_int_component(i
);
686 case GLSL_TYPE_FLOAT
:
687 data
.f
[i
+ base_component
] = c
->get_float_component(i
);
690 data
.b
[i
+ base_component
] = c
->get_bool_component(i
);
693 assert(!"Should not get here.");
698 /* Mask of fields to be written in the assignment.
700 constant_mask
|= ((1U << rhs_components
) - 1) << base_lhs_component
;
701 constant_components
+= rhs_components
;
703 base_component
+= rhs_components
;
705 /* Advance the component index by the number of components
706 * that were just assigned.
708 base_lhs_component
+= rhs_components
;
711 if (constant_mask
!= 0) {
712 ir_dereference
*lhs
= new(ctx
) ir_dereference_variable(var
);
713 const glsl_type
*rhs_type
= glsl_type::get_instance(var
->type
->base_type
,
716 ir_rvalue
*rhs
= new(ctx
) ir_constant(rhs_type
, &data
);
718 ir_instruction
*inst
=
719 new(ctx
) ir_assignment(lhs
, rhs
, NULL
, constant_mask
);
720 instructions
->push_tail(inst
);
724 foreach_list(node
, parameters
) {
725 ir_rvalue
*param
= (ir_rvalue
*) node
;
726 unsigned rhs_components
= param
->type
->components();
728 /* Do not try to assign more components to the vector than it has!
730 if ((rhs_components
+ base_component
) > lhs_components
) {
731 rhs_components
= lhs_components
- base_component
;
734 const ir_constant
*const c
= param
->as_constant();
736 /* Mask of fields to be written in the assignment.
738 const unsigned write_mask
= ((1U << rhs_components
) - 1)
741 ir_dereference
*lhs
= new(ctx
) ir_dereference_variable(var
);
743 /* Generate a swizzle so that LHS and RHS sizes match.
746 new(ctx
) ir_swizzle(param
, 0, 1, 2, 3, rhs_components
);
748 ir_instruction
*inst
=
749 new(ctx
) ir_assignment(lhs
, rhs
, NULL
, write_mask
);
750 instructions
->push_tail(inst
);
753 /* Advance the component index by the number of components that were
756 base_component
+= rhs_components
;
759 return new(ctx
) ir_dereference_variable(var
);
764 * Generate assignment of a portion of a vector to a portion of a matrix column
766 * \param src_base First component of the source to be used in assignment
767 * \param column Column of destination to be assiged
768 * \param row_base First component of the destination column to be assigned
769 * \param count Number of components to be assigned
772 * \c src_base + \c count must be less than or equal to the number of components
773 * in the source vector.
776 assign_to_matrix_column(ir_variable
*var
, unsigned column
, unsigned row_base
,
777 ir_rvalue
*src
, unsigned src_base
, unsigned count
,
780 ir_constant
*col_idx
= new(mem_ctx
) ir_constant(column
);
781 ir_dereference
*column_ref
= new(mem_ctx
) ir_dereference_array(var
, col_idx
);
783 assert(column_ref
->type
->components() >= (row_base
+ count
));
784 assert(src
->type
->components() >= (src_base
+ count
));
786 /* Generate a swizzle that extracts the number of components from the source
787 * that are to be assigned to the column of the matrix.
789 if (count
< src
->type
->vector_elements
) {
790 src
= new(mem_ctx
) ir_swizzle(src
,
791 src_base
+ 0, src_base
+ 1,
792 src_base
+ 2, src_base
+ 3,
796 /* Mask of fields to be written in the assignment.
798 const unsigned write_mask
= ((1U << count
) - 1) << row_base
;
800 return new(mem_ctx
) ir_assignment(column_ref
, src
, NULL
, write_mask
);
805 * Generate inline code for a matrix constructor
807 * The generated constructor code will consist of a temporary variable
808 * declaration of the same type as the constructor. A sequence of assignments
809 * from constructor parameters to the temporary will follow.
812 * An \c ir_dereference_variable of the temprorary generated in the constructor
816 emit_inline_matrix_constructor(const glsl_type
*type
,
817 exec_list
*instructions
,
818 exec_list
*parameters
,
821 assert(!parameters
->is_empty());
823 ir_variable
*var
= new(ctx
) ir_variable(type
, "mat_ctor", ir_var_temporary
);
824 instructions
->push_tail(var
);
826 /* There are three kinds of matrix constructors.
828 * - Construct a matrix from a single scalar by replicating that scalar to
829 * along the diagonal of the matrix and setting all other components to
832 * - Construct a matrix from an arbirary combination of vectors and
833 * scalars. The components of the constructor parameters are assigned
834 * to the matrix in colum-major order until the matrix is full.
836 * - Construct a matrix from a single matrix. The source matrix is copied
837 * to the upper left portion of the constructed matrix, and the remaining
838 * elements take values from the identity matrix.
840 ir_rvalue
*const first_param
= (ir_rvalue
*) parameters
->head
;
841 if (single_scalar_parameter(parameters
)) {
842 /* Assign the scalar to the X component of a vec4, and fill the remaining
843 * components with zero.
845 ir_variable
*rhs_var
=
846 new(ctx
) ir_variable(glsl_type::vec4_type
, "mat_ctor_vec",
848 instructions
->push_tail(rhs_var
);
850 ir_constant_data zero
;
856 ir_instruction
*inst
=
857 new(ctx
) ir_assignment(new(ctx
) ir_dereference_variable(rhs_var
),
858 new(ctx
) ir_constant(rhs_var
->type
, &zero
),
860 instructions
->push_tail(inst
);
862 ir_dereference
*const rhs_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
864 inst
= new(ctx
) ir_assignment(rhs_ref
, first_param
, NULL
, 0x01);
865 instructions
->push_tail(inst
);
867 /* Assign the temporary vector to each column of the destination matrix
868 * with a swizzle that puts the X component on the diagonal of the
869 * matrix. In some cases this may mean that the X component does not
870 * get assigned into the column at all (i.e., when the matrix has more
871 * columns than rows).
873 static const unsigned rhs_swiz
[4][4] = {
880 const unsigned cols_to_init
= MIN2(type
->matrix_columns
,
881 type
->vector_elements
);
882 for (unsigned i
= 0; i
< cols_to_init
; i
++) {
883 ir_constant
*const col_idx
= new(ctx
) ir_constant(i
);
884 ir_rvalue
*const col_ref
= new(ctx
) ir_dereference_array(var
, col_idx
);
886 ir_rvalue
*const rhs_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
887 ir_rvalue
*const rhs
= new(ctx
) ir_swizzle(rhs_ref
, rhs_swiz
[i
],
888 type
->vector_elements
);
890 inst
= new(ctx
) ir_assignment(col_ref
, rhs
, NULL
);
891 instructions
->push_tail(inst
);
894 for (unsigned i
= cols_to_init
; i
< type
->matrix_columns
; i
++) {
895 ir_constant
*const col_idx
= new(ctx
) ir_constant(i
);
896 ir_rvalue
*const col_ref
= new(ctx
) ir_dereference_array(var
, col_idx
);
898 ir_rvalue
*const rhs_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
899 ir_rvalue
*const rhs
= new(ctx
) ir_swizzle(rhs_ref
, 1, 1, 1, 1,
900 type
->vector_elements
);
902 inst
= new(ctx
) ir_assignment(col_ref
, rhs
, NULL
);
903 instructions
->push_tail(inst
);
905 } else if (first_param
->type
->is_matrix()) {
906 /* From page 50 (56 of the PDF) of the GLSL 1.50 spec:
908 * "If a matrix is constructed from a matrix, then each component
909 * (column i, row j) in the result that has a corresponding
910 * component (column i, row j) in the argument will be initialized
911 * from there. All other components will be initialized to the
912 * identity matrix. If a matrix argument is given to a matrix
913 * constructor, it is an error to have any other arguments."
915 assert(first_param
->next
->is_tail_sentinel());
916 ir_rvalue
*const src_matrix
= first_param
;
918 /* If the source matrix is smaller, pre-initialize the relavent parts of
919 * the destination matrix to the identity matrix.
921 if ((src_matrix
->type
->matrix_columns
< var
->type
->matrix_columns
)
922 || (src_matrix
->type
->vector_elements
< var
->type
->vector_elements
)) {
924 /* If the source matrix has fewer rows, every column of the destination
925 * must be initialized. Otherwise only the columns in the destination
926 * that do not exist in the source must be initialized.
929 (src_matrix
->type
->vector_elements
< var
->type
->vector_elements
)
930 ? 0 : src_matrix
->type
->matrix_columns
;
932 const glsl_type
*const col_type
= var
->type
->column_type();
933 for (/* empty */; col
< var
->type
->matrix_columns
; col
++) {
934 ir_constant_data ident
;
943 ir_rvalue
*const rhs
= new(ctx
) ir_constant(col_type
, &ident
);
945 ir_rvalue
*const lhs
=
946 new(ctx
) ir_dereference_array(var
, new(ctx
) ir_constant(col
));
948 ir_instruction
*inst
= new(ctx
) ir_assignment(lhs
, rhs
, NULL
);
949 instructions
->push_tail(inst
);
953 /* Assign columns from the source matrix to the destination matrix.
955 * Since the parameter will be used in the RHS of multiple assignments,
956 * generate a temporary and copy the paramter there.
958 ir_variable
*const rhs_var
=
959 new(ctx
) ir_variable(first_param
->type
, "mat_ctor_mat",
961 instructions
->push_tail(rhs_var
);
963 ir_dereference
*const rhs_var_ref
=
964 new(ctx
) ir_dereference_variable(rhs_var
);
965 ir_instruction
*const inst
=
966 new(ctx
) ir_assignment(rhs_var_ref
, first_param
, NULL
);
967 instructions
->push_tail(inst
);
969 const unsigned last_row
= MIN2(src_matrix
->type
->vector_elements
,
970 var
->type
->vector_elements
);
971 const unsigned last_col
= MIN2(src_matrix
->type
->matrix_columns
,
972 var
->type
->matrix_columns
);
974 unsigned swiz
[4] = { 0, 0, 0, 0 };
975 for (unsigned i
= 1; i
< last_row
; i
++)
978 const unsigned write_mask
= (1U << last_row
) - 1;
980 for (unsigned i
= 0; i
< last_col
; i
++) {
981 ir_dereference
*const lhs
=
982 new(ctx
) ir_dereference_array(var
, new(ctx
) ir_constant(i
));
983 ir_rvalue
*const rhs_col
=
984 new(ctx
) ir_dereference_array(rhs_var
, new(ctx
) ir_constant(i
));
986 /* If one matrix has columns that are smaller than the columns of the
987 * other matrix, wrap the column access of the larger with a swizzle
988 * so that the LHS and RHS of the assignment have the same size (and
989 * therefore have the same type).
991 * It would be perfectly valid to unconditionally generate the
992 * swizzles, this this will typically result in a more compact IR tree.
995 if (lhs
->type
->vector_elements
!= rhs_col
->type
->vector_elements
) {
996 rhs
= new(ctx
) ir_swizzle(rhs_col
, swiz
, last_row
);
1001 ir_instruction
*inst
=
1002 new(ctx
) ir_assignment(lhs
, rhs
, NULL
, write_mask
);
1003 instructions
->push_tail(inst
);
1006 const unsigned cols
= type
->matrix_columns
;
1007 const unsigned rows
= type
->vector_elements
;
1008 unsigned col_idx
= 0;
1009 unsigned row_idx
= 0;
1011 foreach_list (node
, parameters
) {
1012 ir_rvalue
*const rhs
= (ir_rvalue
*) node
;
1013 const unsigned components_remaining_this_column
= rows
- row_idx
;
1014 unsigned rhs_components
= rhs
->type
->components();
1015 unsigned rhs_base
= 0;
1017 /* Since the parameter might be used in the RHS of two assignments,
1018 * generate a temporary and copy the paramter there.
1020 ir_variable
*rhs_var
=
1021 new(ctx
) ir_variable(rhs
->type
, "mat_ctor_vec", ir_var_temporary
);
1022 instructions
->push_tail(rhs_var
);
1024 ir_dereference
*rhs_var_ref
=
1025 new(ctx
) ir_dereference_variable(rhs_var
);
1026 ir_instruction
*inst
= new(ctx
) ir_assignment(rhs_var_ref
, rhs
, NULL
);
1027 instructions
->push_tail(inst
);
1029 /* Assign the current parameter to as many components of the matrix
1032 * NOTE: A single vector parameter can span two matrix columns. A
1033 * single vec4, for example, can completely fill a mat2.
1035 if (rhs_components
>= components_remaining_this_column
) {
1036 const unsigned count
= MIN2(rhs_components
,
1037 components_remaining_this_column
);
1039 rhs_var_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
1041 ir_instruction
*inst
= assign_to_matrix_column(var
, col_idx
,
1045 instructions
->push_tail(inst
);
1053 /* If there is data left in the parameter and components left to be
1054 * set in the destination, emit another assignment. It is possible
1055 * that the assignment could be of a vec4 to the last element of the
1056 * matrix. In this case col_idx==cols, but there is still data
1057 * left in the source parameter. Obviously, don't emit an assignment
1058 * to data outside the destination matrix.
1060 if ((col_idx
< cols
) && (rhs_base
< rhs_components
)) {
1061 const unsigned count
= rhs_components
- rhs_base
;
1063 rhs_var_ref
= new(ctx
) ir_dereference_variable(rhs_var
);
1065 ir_instruction
*inst
= assign_to_matrix_column(var
, col_idx
,
1070 instructions
->push_tail(inst
);
1077 return new(ctx
) ir_dereference_variable(var
);
1082 emit_inline_record_constructor(const glsl_type
*type
,
1083 exec_list
*instructions
,
1084 exec_list
*parameters
,
1087 ir_variable
*const var
=
1088 new(mem_ctx
) ir_variable(type
, "record_ctor", ir_var_temporary
);
1089 ir_dereference_variable
*const d
= new(mem_ctx
) ir_dereference_variable(var
);
1091 instructions
->push_tail(var
);
1093 exec_node
*node
= parameters
->head
;
1094 for (unsigned i
= 0; i
< type
->length
; i
++) {
1095 assert(!node
->is_tail_sentinel());
1097 ir_dereference
*const lhs
=
1098 new(mem_ctx
) ir_dereference_record(d
->clone(mem_ctx
, NULL
),
1099 type
->fields
.structure
[i
].name
);
1101 ir_rvalue
*const rhs
= ((ir_instruction
*) node
)->as_rvalue();
1102 assert(rhs
!= NULL
);
1104 ir_instruction
*const assign
= new(mem_ctx
) ir_assignment(lhs
, rhs
, NULL
);
1106 instructions
->push_tail(assign
);
1115 ast_function_expression::hir(exec_list
*instructions
,
1116 struct _mesa_glsl_parse_state
*state
)
1119 /* There are three sorts of function calls.
1121 * 1. constructors - The first subexpression is an ast_type_specifier.
1122 * 2. methods - Only the .length() method of array types.
1123 * 3. functions - Calls to regular old functions.
1125 * Method calls are actually detected when the ast_field_selection
1126 * expression is handled.
1128 if (is_constructor()) {
1129 const ast_type_specifier
*type
= (ast_type_specifier
*) subexpressions
[0];
1130 YYLTYPE loc
= type
->get_location();
1133 const glsl_type
*const constructor_type
= type
->glsl_type(& name
, state
);
1135 /* constructor_type can be NULL if a variable with the same name as the
1136 * structure has come into scope.
1138 if (constructor_type
== NULL
) {
1139 _mesa_glsl_error(& loc
, state
, "unknown type `%s' (structure name "
1140 "may be shadowed by a variable with the same name)",
1142 return ir_call::get_error_instruction(ctx
);
1146 /* Constructors for samplers are illegal.
1148 if (constructor_type
->is_sampler()) {
1149 _mesa_glsl_error(& loc
, state
, "cannot construct sampler type `%s'",
1150 constructor_type
->name
);
1151 return ir_call::get_error_instruction(ctx
);
1154 if (constructor_type
->is_array()) {
1155 if (state
->language_version
<= 110) {
1156 _mesa_glsl_error(& loc
, state
,
1157 "array constructors forbidden in GLSL 1.10");
1158 return ir_call::get_error_instruction(ctx
);
1161 return process_array_constructor(instructions
, constructor_type
,
1162 & loc
, &this->expressions
, state
);
1166 /* There are two kinds of constructor call. Constructors for built-in
1167 * language types, such as mat4 and vec2, are free form. The only
1168 * requirement is that the parameters must provide enough values of the
1169 * correct scalar type. Constructors for arrays and structures must
1170 * have the exact number of parameters with matching types in the
1171 * correct order. These constructors follow essentially the same type
1172 * matching rules as functions.
1174 if (constructor_type
->is_record()) {
1175 exec_list actual_parameters
;
1177 process_parameters(instructions
, &actual_parameters
,
1178 &this->expressions
, state
);
1180 exec_node
*node
= actual_parameters
.head
;
1181 for (unsigned i
= 0; i
< constructor_type
->length
; i
++) {
1182 ir_rvalue
*ir
= (ir_rvalue
*) node
;
1184 if (node
->is_tail_sentinel()) {
1185 _mesa_glsl_error(&loc
, state
,
1186 "insufficient parameters to constructor "
1188 constructor_type
->name
);
1189 return ir_call::get_error_instruction(ctx
);
1192 if (apply_implicit_conversion(constructor_type
->fields
.structure
[i
].type
,
1194 node
->replace_with(ir
);
1196 _mesa_glsl_error(&loc
, state
,
1197 "parameter type mismatch in constructor "
1198 "for `%s.%s' (%s vs %s)",
1199 constructor_type
->name
,
1200 constructor_type
->fields
.structure
[i
].name
,
1202 constructor_type
->fields
.structure
[i
].type
->name
);
1203 return ir_call::get_error_instruction(ctx
);;
1209 if (!node
->is_tail_sentinel()) {
1210 _mesa_glsl_error(&loc
, state
, "too many parameters in constructor "
1211 "for `%s'", constructor_type
->name
);
1212 return ir_call::get_error_instruction(ctx
);
1215 ir_rvalue
*const constant
=
1216 constant_record_constructor(constructor_type
, &actual_parameters
,
1219 return (constant
!= NULL
)
1221 : emit_inline_record_constructor(constructor_type
, instructions
,
1222 &actual_parameters
, state
);
1225 if (!constructor_type
->is_numeric() && !constructor_type
->is_boolean())
1226 return ir_call::get_error_instruction(ctx
);
1228 /* Total number of components of the type being constructed. */
1229 const unsigned type_components
= constructor_type
->components();
1231 /* Number of components from parameters that have actually been
1232 * consumed. This is used to perform several kinds of error checking.
1234 unsigned components_used
= 0;
1236 unsigned matrix_parameters
= 0;
1237 unsigned nonmatrix_parameters
= 0;
1238 exec_list actual_parameters
;
1240 foreach_list (n
, &this->expressions
) {
1241 ast_node
*ast
= exec_node_data(ast_node
, n
, link
);
1242 ir_rvalue
*result
= ast
->hir(instructions
, state
)->as_rvalue();
1244 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
1246 * "It is an error to provide extra arguments beyond this
1247 * last used argument."
1249 if (components_used
>= type_components
) {
1250 _mesa_glsl_error(& loc
, state
, "too many parameters to `%s' "
1252 constructor_type
->name
);
1253 return ir_call::get_error_instruction(ctx
);
1256 if (!result
->type
->is_numeric() && !result
->type
->is_boolean()) {
1257 _mesa_glsl_error(& loc
, state
, "cannot construct `%s' from a "
1258 "non-numeric data type",
1259 constructor_type
->name
);
1260 return ir_call::get_error_instruction(ctx
);
1263 /* Count the number of matrix and nonmatrix parameters. This
1264 * is used below to enforce some of the constructor rules.
1266 if (result
->type
->is_matrix())
1267 matrix_parameters
++;
1269 nonmatrix_parameters
++;
1271 actual_parameters
.push_tail(result
);
1272 components_used
+= result
->type
->components();
1275 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
1277 * "It is an error to construct matrices from other matrices. This
1278 * is reserved for future use."
1280 if (state
->language_version
== 110 && matrix_parameters
> 0
1281 && constructor_type
->is_matrix()) {
1282 _mesa_glsl_error(& loc
, state
, "cannot construct `%s' from a "
1283 "matrix in GLSL 1.10",
1284 constructor_type
->name
);
1285 return ir_call::get_error_instruction(ctx
);
1288 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
1290 * "If a matrix argument is given to a matrix constructor, it is
1291 * an error to have any other arguments."
1293 if ((matrix_parameters
> 0)
1294 && ((matrix_parameters
+ nonmatrix_parameters
) > 1)
1295 && constructor_type
->is_matrix()) {
1296 _mesa_glsl_error(& loc
, state
, "for matrix `%s' constructor, "
1297 "matrix must be only parameter",
1298 constructor_type
->name
);
1299 return ir_call::get_error_instruction(ctx
);
1302 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
1304 * "In these cases, there must be enough components provided in the
1305 * arguments to provide an initializer for every component in the
1306 * constructed value."
1308 if (components_used
< type_components
&& components_used
!= 1
1309 && matrix_parameters
== 0) {
1310 _mesa_glsl_error(& loc
, state
, "too few components to construct "
1312 constructor_type
->name
);
1313 return ir_call::get_error_instruction(ctx
);
1316 /* Later, we cast each parameter to the same base type as the
1317 * constructor. Since there are no non-floating point matrices, we
1318 * need to break them up into a series of column vectors.
1320 if (constructor_type
->base_type
!= GLSL_TYPE_FLOAT
) {
1321 foreach_list_safe(n
, &actual_parameters
) {
1322 ir_rvalue
*matrix
= (ir_rvalue
*) n
;
1324 if (!matrix
->type
->is_matrix())
1327 /* Create a temporary containing the matrix. */
1328 ir_variable
*var
= new(ctx
) ir_variable(matrix
->type
, "matrix_tmp",
1330 instructions
->push_tail(var
);
1331 instructions
->push_tail(new(ctx
) ir_assignment(new(ctx
)
1332 ir_dereference_variable(var
), matrix
, NULL
));
1333 var
->constant_value
= matrix
->constant_expression_value();
1335 /* Replace the matrix with dereferences of its columns. */
1336 for (int i
= 0; i
< matrix
->type
->matrix_columns
; i
++) {
1337 matrix
->insert_before(new (ctx
) ir_dereference_array(var
,
1338 new(ctx
) ir_constant(i
)));
1344 bool all_parameters_are_constant
= true;
1346 /* Type cast each parameter and, if possible, fold constants.*/
1347 foreach_list_safe(n
, &actual_parameters
) {
1348 ir_rvalue
*ir
= (ir_rvalue
*) n
;
1350 const glsl_type
*desired_type
=
1351 glsl_type::get_instance(constructor_type
->base_type
,
1352 ir
->type
->vector_elements
,
1353 ir
->type
->matrix_columns
);
1354 ir_rvalue
*result
= convert_component(ir
, desired_type
);
1356 /* Attempt to convert the parameter to a constant valued expression.
1357 * After doing so, track whether or not all the parameters to the
1358 * constructor are trivially constant valued expressions.
1360 ir_rvalue
*const constant
= result
->constant_expression_value();
1362 if (constant
!= NULL
)
1365 all_parameters_are_constant
= false;
1368 ir
->replace_with(result
);
1372 /* If all of the parameters are trivially constant, create a
1373 * constant representing the complete collection of parameters.
1375 if (all_parameters_are_constant
) {
1376 return new(ctx
) ir_constant(constructor_type
, &actual_parameters
);
1377 } else if (constructor_type
->is_scalar()) {
1378 return dereference_component((ir_rvalue
*) actual_parameters
.head
,
1380 } else if (constructor_type
->is_vector()) {
1381 return emit_inline_vector_constructor(constructor_type
,
1386 assert(constructor_type
->is_matrix());
1387 return emit_inline_matrix_constructor(constructor_type
,
1393 const ast_expression
*id
= subexpressions
[0];
1394 YYLTYPE loc
= id
->get_location();
1395 exec_list actual_parameters
;
1397 process_parameters(instructions
, &actual_parameters
, &this->expressions
,
1400 return match_function_by_name(instructions
,
1401 id
->primary_expression
.identifier
, & loc
,
1402 &actual_parameters
, state
);
1405 return ir_call::get_error_instruction(ctx
);