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.
25 * \file ir_constant_expression.cpp
26 * Evaluate and process constant valued expressions
28 * In GLSL, constant valued expressions are used in several places. These
29 * must be processed and evaluated very early in the compilation process.
32 * * Initializers for uniforms
33 * * Initializers for \c const variables
37 #include "main/core.h" /* for MAX2, MIN2, CLAMP */
38 #include "util/rounding.h" /* for _mesa_roundeven */
39 #include "util/half_float.h"
41 #include "compiler/glsl_types.h"
42 #include "util/hash_table.h"
45 dot_f(ir_constant
*op0
, ir_constant
*op1
)
47 assert(op0
->type
->is_float() && op1
->type
->is_float());
50 for (unsigned c
= 0; c
< op0
->type
->components(); c
++)
51 result
+= op0
->value
.f
[c
] * op1
->value
.f
[c
];
57 dot_d(ir_constant
*op0
, ir_constant
*op1
)
59 assert(op0
->type
->is_double() && op1
->type
->is_double());
62 for (unsigned c
= 0; c
< op0
->type
->components(); c
++)
63 result
+= op0
->value
.d
[c
] * op1
->value
.d
[c
];
68 /* This method is the only one supported by gcc. Unions in particular
69 * are iffy, and read-through-converted-pointer is killed by strict
70 * aliasing. OTOH, the compiler sees through the memcpy, so the
71 * resulting asm is reasonable.
74 bitcast_u2f(unsigned int u
)
76 assert(sizeof(float) == sizeof(unsigned int));
78 memcpy(&f
, &u
, sizeof(f
));
85 assert(sizeof(float) == sizeof(unsigned int));
87 memcpy(&u
, &f
, sizeof(f
));
92 * Evaluate one component of a floating-point 4x8 unpacking function.
95 (*pack_1x8_func_t
)(float);
98 * Evaluate one component of a floating-point 2x16 unpacking function.
101 (*pack_1x16_func_t
)(float);
104 * Evaluate one component of a floating-point 4x8 unpacking function.
107 (*unpack_1x8_func_t
)(uint8_t);
110 * Evaluate one component of a floating-point 2x16 unpacking function.
113 (*unpack_1x16_func_t
)(uint16_t);
116 * Evaluate a 2x16 floating-point packing function.
119 pack_2x16(pack_1x16_func_t pack_1x16
,
122 /* From section 8.4 of the GLSL ES 3.00 spec:
126 * The first component of the vector will be written to the least
127 * significant bits of the output; the last component will be written to
128 * the most significant bits.
130 * The specifications for the other packing functions contain similar
134 u
|= ((uint32_t) pack_1x16(x
) << 0);
135 u
|= ((uint32_t) pack_1x16(y
) << 16);
140 * Evaluate a 4x8 floating-point packing function.
143 pack_4x8(pack_1x8_func_t pack_1x8
,
144 float x
, float y
, float z
, float w
)
146 /* From section 8.4 of the GLSL 4.30 spec:
150 * The first component of the vector will be written to the least
151 * significant bits of the output; the last component will be written to
152 * the most significant bits.
154 * The specifications for the other packing functions contain similar
158 u
|= ((uint32_t) pack_1x8(x
) << 0);
159 u
|= ((uint32_t) pack_1x8(y
) << 8);
160 u
|= ((uint32_t) pack_1x8(z
) << 16);
161 u
|= ((uint32_t) pack_1x8(w
) << 24);
166 * Evaluate a 2x16 floating-point unpacking function.
169 unpack_2x16(unpack_1x16_func_t unpack_1x16
,
173 /* From section 8.4 of the GLSL ES 3.00 spec:
177 * The first component of the returned vector will be extracted from
178 * the least significant bits of the input; the last component will be
179 * extracted from the most significant bits.
181 * The specifications for the other unpacking functions contain similar
184 *x
= unpack_1x16((uint16_t) (u
& 0xffff));
185 *y
= unpack_1x16((uint16_t) (u
>> 16));
189 * Evaluate a 4x8 floating-point unpacking function.
192 unpack_4x8(unpack_1x8_func_t unpack_1x8
, uint32_t u
,
193 float *x
, float *y
, float *z
, float *w
)
195 /* From section 8.4 of the GLSL 4.30 spec:
199 * The first component of the returned vector will be extracted from
200 * the least significant bits of the input; the last component will be
201 * extracted from the most significant bits.
203 * The specifications for the other unpacking functions contain similar
206 *x
= unpack_1x8((uint8_t) (u
& 0xff));
207 *y
= unpack_1x8((uint8_t) (u
>> 8));
208 *z
= unpack_1x8((uint8_t) (u
>> 16));
209 *w
= unpack_1x8((uint8_t) (u
>> 24));
213 * Evaluate one component of packSnorm4x8.
216 pack_snorm_1x8(float x
)
218 /* From section 8.4 of the GLSL 4.30 spec:
222 * The conversion for component c of v to fixed point is done as
225 * packSnorm4x8: round(clamp(c, -1, +1) * 127.0)
228 _mesa_lroundevenf(CLAMP(x
, -1.0f
, +1.0f
) * 127.0f
);
232 * Evaluate one component of packSnorm2x16.
235 pack_snorm_1x16(float x
)
237 /* From section 8.4 of the GLSL ES 3.00 spec:
241 * The conversion for component c of v to fixed point is done as
244 * packSnorm2x16: round(clamp(c, -1, +1) * 32767.0)
247 _mesa_lroundevenf(CLAMP(x
, -1.0f
, +1.0f
) * 32767.0f
);
251 * Evaluate one component of unpackSnorm4x8.
254 unpack_snorm_1x8(uint8_t u
)
256 /* From section 8.4 of the GLSL 4.30 spec:
260 * The conversion for unpacked fixed-point value f to floating point is
263 * unpackSnorm4x8: clamp(f / 127.0, -1, +1)
265 return CLAMP((int8_t) u
/ 127.0f
, -1.0f
, +1.0f
);
269 * Evaluate one component of unpackSnorm2x16.
272 unpack_snorm_1x16(uint16_t u
)
274 /* From section 8.4 of the GLSL ES 3.00 spec:
278 * The conversion for unpacked fixed-point value f to floating point is
281 * unpackSnorm2x16: clamp(f / 32767.0, -1, +1)
283 return CLAMP((int16_t) u
/ 32767.0f
, -1.0f
, +1.0f
);
287 * Evaluate one component packUnorm4x8.
290 pack_unorm_1x8(float x
)
292 /* From section 8.4 of the GLSL 4.30 spec:
296 * The conversion for component c of v to fixed point is done as
299 * packUnorm4x8: round(clamp(c, 0, +1) * 255.0)
301 return (uint8_t) (int) _mesa_roundevenf(CLAMP(x
, 0.0f
, 1.0f
) * 255.0f
);
305 * Evaluate one component packUnorm2x16.
308 pack_unorm_1x16(float x
)
310 /* From section 8.4 of the GLSL ES 3.00 spec:
314 * The conversion for component c of v to fixed point is done as
317 * packUnorm2x16: round(clamp(c, 0, +1) * 65535.0)
319 return (uint16_t) (int)
320 _mesa_roundevenf(CLAMP(x
, 0.0f
, 1.0f
) * 65535.0f
);
324 * Evaluate one component of unpackUnorm4x8.
327 unpack_unorm_1x8(uint8_t u
)
329 /* From section 8.4 of the GLSL 4.30 spec:
333 * The conversion for unpacked fixed-point value f to floating point is
336 * unpackUnorm4x8: f / 255.0
338 return (float) u
/ 255.0f
;
342 * Evaluate one component of unpackUnorm2x16.
345 unpack_unorm_1x16(uint16_t u
)
347 /* From section 8.4 of the GLSL ES 3.00 spec:
351 * The conversion for unpacked fixed-point value f to floating point is
354 * unpackUnorm2x16: f / 65535.0
356 return (float) u
/ 65535.0f
;
360 * Evaluate one component of packHalf2x16.
363 pack_half_1x16(float x
)
365 return _mesa_float_to_half(x
);
369 * Evaluate one component of unpackHalf2x16.
372 unpack_half_1x16(uint16_t u
)
374 return _mesa_half_to_float(u
);
378 * Get the constant that is ultimately referenced by an r-value, in a constant
379 * expression evaluation context.
381 * The offset is used when the reference is to a specific column of a matrix.
384 constant_referenced(const ir_dereference
*deref
,
385 struct hash_table
*variable_context
,
386 ir_constant
*&store
, int &offset
)
391 if (variable_context
== NULL
)
394 switch (deref
->ir_type
) {
395 case ir_type_dereference_array
: {
396 const ir_dereference_array
*const da
=
397 (const ir_dereference_array
*) deref
;
399 ir_constant
*const index_c
=
400 da
->array_index
->constant_expression_value(variable_context
);
402 if (!index_c
|| !index_c
->type
->is_scalar() || !index_c
->type
->is_integer())
405 const int index
= index_c
->type
->base_type
== GLSL_TYPE_INT
?
406 index_c
->get_int_component(0) :
407 index_c
->get_uint_component(0);
409 ir_constant
*substore
;
412 const ir_dereference
*const deref
= da
->array
->as_dereference();
416 if (!constant_referenced(deref
, variable_context
, substore
, suboffset
))
419 const glsl_type
*const vt
= da
->array
->type
;
420 if (vt
->is_array()) {
421 store
= substore
->get_array_element(index
);
423 } else if (vt
->is_matrix()) {
425 offset
= index
* vt
->vector_elements
;
426 } else if (vt
->is_vector()) {
428 offset
= suboffset
+ index
;
434 case ir_type_dereference_record
: {
435 const ir_dereference_record
*const dr
=
436 (const ir_dereference_record
*) deref
;
438 const ir_dereference
*const deref
= dr
->record
->as_dereference();
442 ir_constant
*substore
;
445 if (!constant_referenced(deref
, variable_context
, substore
, suboffset
))
448 /* Since we're dropping it on the floor...
450 assert(suboffset
== 0);
452 store
= substore
->get_record_field(dr
->field
);
456 case ir_type_dereference_variable
: {
457 const ir_dereference_variable
*const dv
=
458 (const ir_dereference_variable
*) deref
;
460 hash_entry
*entry
= _mesa_hash_table_search(variable_context
, dv
->var
);
462 store
= (ir_constant
*) entry
->data
;
467 assert(!"Should not get here.");
471 return store
!= NULL
;
476 ir_rvalue::constant_expression_value(struct hash_table
*)
478 assert(this->type
->is_error());
483 bitfield_reverse(uint32_t v
)
485 /* http://graphics.stanford.edu/~seander/bithacks.html#BitReverseObvious */
486 uint32_t r
= v
; // r will be reversed bits of v; first get LSB of v
487 int s
= sizeof(v
) * CHAR_BIT
- 1; // extra shift needed at end
489 for (v
>>= 1; v
; v
>>= 1) {
494 r
<<= s
; // shift when v's highest bits are zero
500 find_msb_uint(uint32_t v
)
504 /* If v == 0, then the loop will terminate when count == 32. In that case
505 * 31-count will produce the -1 result required by GLSL findMSB().
507 while (((v
& (1u << 31)) == 0) && count
!= 32) {
516 find_msb_int(int32_t v
)
518 /* If v is signed, findMSB() returns the position of the most significant
521 return find_msb_uint(v
< 0 ? ~v
: v
);
525 ldexpf_flush_subnormal(float x
, int exp
)
527 const float result
= ldexpf(x
, exp
);
529 /* Flush subnormal values to zero. */
530 return !isnormal(result
) ? copysignf(0.0f
, x
) : result
;
534 ldexp_flush_subnormal(double x
, int exp
)
536 const double result
= ldexp(x
, exp
);
538 /* Flush subnormal values to zero. */
539 return !isnormal(result
) ? copysign(0.0, x
) : result
;
543 bitfield_extract_uint(uint32_t value
, int offset
, int bits
)
547 else if (offset
< 0 || bits
< 0)
548 return 0; /* Undefined, per spec. */
549 else if (offset
+ bits
> 32)
550 return 0; /* Undefined, per spec. */
552 value
<<= 32 - bits
- offset
;
559 bitfield_extract_int(int32_t value
, int offset
, int bits
)
563 else if (offset
< 0 || bits
< 0)
564 return 0; /* Undefined, per spec. */
565 else if (offset
+ bits
> 32)
566 return 0; /* Undefined, per spec. */
568 value
<<= 32 - bits
- offset
;
575 bitfield_insert(uint32_t base
, uint32_t insert
, int offset
, int bits
)
579 else if (offset
< 0 || bits
< 0)
580 return 0; /* Undefined, per spec. */
581 else if (offset
+ bits
> 32)
582 return 0; /* Undefined, per spec. */
584 unsigned insert_mask
= ((1ull << bits
) - 1) << offset
;
587 insert
&= insert_mask
;
588 base
&= ~insert_mask
;
590 return base
| insert
;
595 ir_expression::constant_expression_value(struct hash_table
*variable_context
)
597 if (this->type
->is_error())
600 ir_constant
*op
[ARRAY_SIZE(this->operands
)] = { NULL
, };
601 ir_constant_data data
;
603 memset(&data
, 0, sizeof(data
));
605 for (unsigned operand
= 0; operand
< this->get_num_operands(); operand
++) {
606 op
[operand
] = this->operands
[operand
]->constant_expression_value(variable_context
);
612 switch (this->operation
) {
613 case ir_binop_lshift
:
614 case ir_binop_rshift
:
616 case ir_binop_interpolate_at_offset
:
617 case ir_binop_interpolate_at_sample
:
618 case ir_binop_vector_extract
:
620 case ir_triop_bitfield_extract
:
624 assert(op
[0]->type
->base_type
== op
[1]->type
->base_type
);
628 bool op0_scalar
= op
[0]->type
->is_scalar();
629 bool op1_scalar
= op
[1] != NULL
&& op
[1]->type
->is_scalar();
631 /* When iterating over a vector or matrix's components, we want to increase
632 * the loop counter. However, for scalars, we want to stay at 0.
634 unsigned c0_inc
= op0_scalar
? 0 : 1;
635 unsigned c1_inc
= op1_scalar
? 0 : 1;
637 if (op1_scalar
|| !op
[1]) {
638 components
= op
[0]->type
->components();
640 components
= op
[1]->type
->components();
643 void *ctx
= ralloc_parent(this);
645 /* Handle array operations here, rather than below. */
646 if (op
[0]->type
->is_array()) {
647 assert(op
[1] != NULL
&& op
[1]->type
->is_array());
648 switch (this->operation
) {
649 case ir_binop_all_equal
:
650 return new(ctx
) ir_constant(op
[0]->has_value(op
[1]));
651 case ir_binop_any_nequal
:
652 return new(ctx
) ir_constant(!op
[0]->has_value(op
[1]));
659 #include "ir_expression_operation_constant.h"
661 return new(ctx
) ir_constant(this->type
, &data
);
666 ir_texture::constant_expression_value(struct hash_table
*)
668 /* texture lookups aren't constant expressions */
674 ir_swizzle::constant_expression_value(struct hash_table
*variable_context
)
676 ir_constant
*v
= this->val
->constant_expression_value(variable_context
);
679 ir_constant_data data
= { { 0 } };
681 const unsigned swiz_idx
[4] = {
682 this->mask
.x
, this->mask
.y
, this->mask
.z
, this->mask
.w
685 for (unsigned i
= 0; i
< this->mask
.num_components
; i
++) {
686 switch (v
->type
->base_type
) {
688 case GLSL_TYPE_INT
: data
.u
[i
] = v
->value
.u
[swiz_idx
[i
]]; break;
689 case GLSL_TYPE_FLOAT
: data
.f
[i
] = v
->value
.f
[swiz_idx
[i
]]; break;
690 case GLSL_TYPE_BOOL
: data
.b
[i
] = v
->value
.b
[swiz_idx
[i
]]; break;
691 case GLSL_TYPE_DOUBLE
:data
.d
[i
] = v
->value
.d
[swiz_idx
[i
]]; break;
692 default: assert(!"Should not get here."); break;
696 void *ctx
= ralloc_parent(this);
697 return new(ctx
) ir_constant(this->type
, &data
);
704 ir_dereference_variable::constant_expression_value(struct hash_table
*variable_context
)
708 /* Give priority to the context hashtable, if it exists */
709 if (variable_context
) {
710 hash_entry
*entry
= _mesa_hash_table_search(variable_context
, var
);
713 return (ir_constant
*) entry
->data
;
716 /* The constant_value of a uniform variable is its initializer,
717 * not the lifetime constant value of the uniform.
719 if (var
->data
.mode
== ir_var_uniform
)
722 if (!var
->constant_value
)
725 return var
->constant_value
->clone(ralloc_parent(var
), NULL
);
730 ir_dereference_array::constant_expression_value(struct hash_table
*variable_context
)
732 ir_constant
*array
= this->array
->constant_expression_value(variable_context
);
733 ir_constant
*idx
= this->array_index
->constant_expression_value(variable_context
);
735 if ((array
!= NULL
) && (idx
!= NULL
)) {
736 void *ctx
= ralloc_parent(this);
737 if (array
->type
->is_matrix()) {
738 /* Array access of a matrix results in a vector.
740 const unsigned column
= idx
->value
.u
[0];
742 const glsl_type
*const column_type
= array
->type
->column_type();
744 /* Offset in the constant matrix to the first element of the column
747 const unsigned mat_idx
= column
* column_type
->vector_elements
;
749 ir_constant_data data
= { { 0 } };
751 switch (column_type
->base_type
) {
754 for (unsigned i
= 0; i
< column_type
->vector_elements
; i
++)
755 data
.u
[i
] = array
->value
.u
[mat_idx
+ i
];
759 case GLSL_TYPE_FLOAT
:
760 for (unsigned i
= 0; i
< column_type
->vector_elements
; i
++)
761 data
.f
[i
] = array
->value
.f
[mat_idx
+ i
];
765 case GLSL_TYPE_DOUBLE
:
766 for (unsigned i
= 0; i
< column_type
->vector_elements
; i
++)
767 data
.d
[i
] = array
->value
.d
[mat_idx
+ i
];
772 assert(!"Should not get here.");
776 return new(ctx
) ir_constant(column_type
, &data
);
777 } else if (array
->type
->is_vector()) {
778 const unsigned component
= idx
->value
.u
[0];
780 return new(ctx
) ir_constant(array
, component
);
782 const unsigned index
= idx
->value
.u
[0];
783 return array
->get_array_element(index
)->clone(ctx
, NULL
);
791 ir_dereference_record::constant_expression_value(struct hash_table
*)
793 ir_constant
*v
= this->record
->constant_expression_value();
795 return (v
!= NULL
) ? v
->get_record_field(this->field
) : NULL
;
800 ir_assignment::constant_expression_value(struct hash_table
*)
802 /* FINISHME: Handle CEs involving assignment (return RHS) */
808 ir_constant::constant_expression_value(struct hash_table
*)
815 ir_call::constant_expression_value(struct hash_table
*variable_context
)
817 return this->callee
->constant_expression_value(&this->actual_parameters
, variable_context
);
821 bool ir_function_signature::constant_expression_evaluate_expression_list(const struct exec_list
&body
,
822 struct hash_table
*variable_context
,
823 ir_constant
**result
)
825 foreach_in_list(ir_instruction
, inst
, &body
) {
826 switch(inst
->ir_type
) {
828 /* (declare () type symbol) */
829 case ir_type_variable
: {
830 ir_variable
*var
= inst
->as_variable();
831 _mesa_hash_table_insert(variable_context
, var
, ir_constant::zero(this, var
->type
));
835 /* (assign [condition] (write-mask) (ref) (value)) */
836 case ir_type_assignment
: {
837 ir_assignment
*asg
= inst
->as_assignment();
838 if (asg
->condition
) {
839 ir_constant
*cond
= asg
->condition
->constant_expression_value(variable_context
);
842 if (!cond
->get_bool_component(0))
846 ir_constant
*store
= NULL
;
849 if (!constant_referenced(asg
->lhs
, variable_context
, store
, offset
))
852 ir_constant
*value
= asg
->rhs
->constant_expression_value(variable_context
);
857 store
->copy_masked_offset(value
, offset
, asg
->write_mask
);
861 /* (return (expression)) */
864 *result
= inst
->as_return()->value
->constant_expression_value(variable_context
);
865 return *result
!= NULL
;
867 /* (call name (ref) (params))*/
869 ir_call
*call
= inst
->as_call();
871 /* Just say no to void functions in constant expressions. We
872 * don't need them at that point.
875 if (!call
->return_deref
)
878 ir_constant
*store
= NULL
;
881 if (!constant_referenced(call
->return_deref
, variable_context
,
885 ir_constant
*value
= call
->constant_expression_value(variable_context
);
890 store
->copy_offset(value
, offset
);
894 /* (if condition (then-instructions) (else-instructions)) */
896 ir_if
*iif
= inst
->as_if();
898 ir_constant
*cond
= iif
->condition
->constant_expression_value(variable_context
);
899 if (!cond
|| !cond
->type
->is_boolean())
902 exec_list
&branch
= cond
->get_bool_component(0) ? iif
->then_instructions
: iif
->else_instructions
;
905 if (!constant_expression_evaluate_expression_list(branch
, variable_context
, result
))
908 /* If there was a return in the branch chosen, drop out now. */
915 /* Every other expression type, we drop out. */
921 /* Reaching the end of the block is not an error condition */
929 ir_function_signature::constant_expression_value(exec_list
*actual_parameters
, struct hash_table
*variable_context
)
931 const glsl_type
*type
= this->return_type
;
932 if (type
== glsl_type::void_type
)
935 /* From the GLSL 1.20 spec, page 23:
936 * "Function calls to user-defined functions (non-built-in functions)
937 * cannot be used to form constant expressions."
939 if (!this->is_builtin())
943 * Of the builtin functions, only the texture lookups and the noise
944 * ones must not be used in constant expressions. They all include
945 * specific opcodes so they don't need to be special-cased at this
949 /* Initialize the table of dereferencable names with the function
950 * parameters. Verify their const-ness on the way.
952 * We expect the correctness of the number of parameters to have
953 * been checked earlier.
955 hash_table
*deref_hash
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
956 _mesa_key_pointer_equal
);
958 /* If "origin" is non-NULL, then the function body is there. So we
959 * have to use the variable objects from the object with the body,
960 * but the parameter instanciation on the current object.
962 const exec_node
*parameter_info
= origin
? origin
->parameters
.get_head_raw() : parameters
.get_head_raw();
964 foreach_in_list(ir_rvalue
, n
, actual_parameters
) {
965 ir_constant
*constant
= n
->constant_expression_value(variable_context
);
966 if (constant
== NULL
) {
967 _mesa_hash_table_destroy(deref_hash
, NULL
);
972 ir_variable
*var
= (ir_variable
*)parameter_info
;
973 _mesa_hash_table_insert(deref_hash
, var
, constant
);
975 parameter_info
= parameter_info
->next
;
978 ir_constant
*result
= NULL
;
980 /* Now run the builtin function until something non-constant
981 * happens or we get the result.
983 if (constant_expression_evaluate_expression_list(origin
? origin
->body
: body
, deref_hash
, &result
) && result
)
984 result
= result
->clone(ralloc_parent(this), NULL
);
986 _mesa_hash_table_destroy(deref_hash
, NULL
);