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 opt_algebraic.cpp
27 * Takes advantage of association, commutivity, and other algebraic
28 * properties to simplify expressions.
32 #include "ir_visitor.h"
33 #include "ir_rvalue_visitor.h"
34 #include "ir_optimization.h"
35 #include "ir_builder.h"
36 #include "compiler/glsl_types.h"
37 #include "main/mtypes.h"
39 using namespace ir_builder
;
44 * Visitor class for replacing expressions with ir_constant values.
47 class ir_algebraic_visitor
: public ir_rvalue_visitor
{
49 ir_algebraic_visitor(bool native_integers
,
50 const struct gl_shader_compiler_options
*options
)
53 this->progress
= false;
55 this->native_integers
= native_integers
;
58 virtual ~ir_algebraic_visitor()
62 virtual ir_visitor_status
visit_enter(ir_assignment
*ir
);
64 ir_rvalue
*handle_expression(ir_expression
*ir
);
65 void handle_rvalue(ir_rvalue
**rvalue
);
66 bool reassociate_constant(ir_expression
*ir1
,
68 ir_constant
*constant
,
70 void reassociate_operands(ir_expression
*ir1
,
74 ir_rvalue
*swizzle_if_required(ir_expression
*expr
,
77 const struct gl_shader_compiler_options
*options
;
84 } /* unnamed namespace */
87 ir_algebraic_visitor::visit_enter(ir_assignment
*ir
)
89 ir_variable
*var
= ir
->lhs
->variable_referenced();
90 if (var
->data
.invariant
|| var
->data
.precise
) {
91 /* If we're assigning to an invariant or precise variable, just bail.
92 * Most of the algebraic optimizations aren't precision-safe.
94 * FINISHME: Find out which optimizations are precision-safe and enable
95 * then only for invariant or precise trees.
97 return visit_continue_with_parent
;
99 return visit_continue
;
104 is_vec_zero(ir_constant
*ir
)
106 return (ir
== NULL
) ? false : ir
->is_zero();
110 is_vec_one(ir_constant
*ir
)
112 return (ir
== NULL
) ? false : ir
->is_one();
116 is_vec_two(ir_constant
*ir
)
118 return (ir
== NULL
) ? false : ir
->is_value(2.0, 2);
122 is_vec_four(ir_constant
*ir
)
124 return (ir
== NULL
) ? false : ir
->is_value(4.0, 4);
128 is_vec_negative_one(ir_constant
*ir
)
130 return (ir
== NULL
) ? false : ir
->is_negative_one();
134 is_valid_vec_const(ir_constant
*ir
)
139 if (!ir
->type
->is_scalar() && !ir
->type
->is_vector())
146 is_less_than_one(ir_constant
*ir
)
148 assert(ir
->type
->is_float());
150 if (!is_valid_vec_const(ir
))
153 unsigned component
= 0;
154 for (int c
= 0; c
< ir
->type
->vector_elements
; c
++) {
155 if (ir
->get_float_component(c
) < 1.0f
)
159 return (component
== ir
->type
->vector_elements
);
163 is_greater_than_zero(ir_constant
*ir
)
165 assert(ir
->type
->is_float());
167 if (!is_valid_vec_const(ir
))
170 unsigned component
= 0;
171 for (int c
= 0; c
< ir
->type
->vector_elements
; c
++) {
172 if (ir
->get_float_component(c
) > 0.0f
)
176 return (component
== ir
->type
->vector_elements
);
180 update_type(ir_expression
*ir
)
182 if (ir
->operands
[0]->type
->is_vector())
183 ir
->type
= ir
->operands
[0]->type
;
185 ir
->type
= ir
->operands
[1]->type
;
188 /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
189 static ir_expression
*
190 try_replace_with_dot(ir_expression
*expr0
, ir_expression
*expr1
, void *mem_ctx
)
192 if (expr0
&& expr0
->operation
== ir_binop_add
&&
193 expr0
->type
->is_float() &&
194 expr1
&& expr1
->operation
== ir_binop_add
&&
195 expr1
->type
->is_float()) {
196 ir_swizzle
*x
= expr0
->operands
[0]->as_swizzle();
197 ir_swizzle
*y
= expr0
->operands
[1]->as_swizzle();
198 ir_swizzle
*z
= expr1
->operands
[0]->as_swizzle();
199 ir_swizzle
*w
= expr1
->operands
[1]->as_swizzle();
201 if (!x
|| x
->mask
.num_components
!= 1 ||
202 !y
|| y
->mask
.num_components
!= 1 ||
203 !z
|| z
->mask
.num_components
!= 1 ||
204 !w
|| w
->mask
.num_components
!= 1) {
208 bool swiz_seen
[4] = {false, false, false, false};
209 swiz_seen
[x
->mask
.x
] = true;
210 swiz_seen
[y
->mask
.x
] = true;
211 swiz_seen
[z
->mask
.x
] = true;
212 swiz_seen
[w
->mask
.x
] = true;
214 if (!swiz_seen
[0] || !swiz_seen
[1] ||
215 !swiz_seen
[2] || !swiz_seen
[3]) {
219 if (x
->val
->equals(y
->val
) &&
220 x
->val
->equals(z
->val
) &&
221 x
->val
->equals(w
->val
)) {
222 return dot(x
->val
, new(mem_ctx
) ir_constant(1.0f
, 4));
229 ir_algebraic_visitor::reassociate_operands(ir_expression
*ir1
,
234 ir_rvalue
*temp
= ir2
->operands
[op2
];
235 ir2
->operands
[op2
] = ir1
->operands
[op1
];
236 ir1
->operands
[op1
] = temp
;
238 /* Update the type of ir2. The type of ir1 won't have changed --
239 * base types matched, and at least one of the operands of the 2
240 * binops is still a vector if any of them were.
244 this->progress
= true;
248 * Reassociates a constant down a tree of adds or multiplies.
250 * Consider (2 * (a * (b * 0.5))). We want to end up with a * b.
253 ir_algebraic_visitor::reassociate_constant(ir_expression
*ir1
, int const_index
,
254 ir_constant
*constant
,
257 if (!ir2
|| ir1
->operation
!= ir2
->operation
)
260 /* Don't want to even think about matrices. */
261 if (ir1
->operands
[0]->type
->is_matrix() ||
262 ir1
->operands
[1]->type
->is_matrix() ||
263 ir2
->operands
[0]->type
->is_matrix() ||
264 ir2
->operands
[1]->type
->is_matrix())
267 void *mem_ctx
= ralloc_parent(ir2
);
269 ir_constant
*ir2_const
[2];
270 ir2_const
[0] = ir2
->operands
[0]->constant_expression_value(mem_ctx
);
271 ir2_const
[1] = ir2
->operands
[1]->constant_expression_value(mem_ctx
);
273 if (ir2_const
[0] && ir2_const
[1])
277 reassociate_operands(ir1
, const_index
, ir2
, 1);
279 } else if (ir2_const
[1]) {
280 reassociate_operands(ir1
, const_index
, ir2
, 0);
284 if (reassociate_constant(ir1
, const_index
, constant
,
285 ir2
->operands
[0]->as_expression())) {
290 if (reassociate_constant(ir1
, const_index
, constant
,
291 ir2
->operands
[1]->as_expression())) {
299 /* When eliminating an expression and just returning one of its operands,
300 * we may need to swizzle that operand out to a vector if the expression was
304 ir_algebraic_visitor::swizzle_if_required(ir_expression
*expr
,
307 if (expr
->type
->is_vector() && operand
->type
->is_scalar()) {
308 return new(mem_ctx
) ir_swizzle(operand
, 0, 0, 0, 0,
309 expr
->type
->vector_elements
);
315 ir_algebraic_visitor::handle_expression(ir_expression
*ir
)
317 ir_constant
*op_const
[4] = {NULL
, NULL
, NULL
, NULL
};
318 ir_expression
*op_expr
[4] = {NULL
, NULL
, NULL
, NULL
};
320 if (ir
->operation
== ir_binop_mul
&&
321 ir
->operands
[0]->type
->is_matrix() &&
322 ir
->operands
[1]->type
->is_vector()) {
323 ir_expression
*matrix_mul
= ir
->operands
[0]->as_expression();
325 if (matrix_mul
&& matrix_mul
->operation
== ir_binop_mul
&&
326 matrix_mul
->operands
[0]->type
->is_matrix() &&
327 matrix_mul
->operands
[1]->type
->is_matrix()) {
329 return mul(matrix_mul
->operands
[0],
330 mul(matrix_mul
->operands
[1], ir
->operands
[1]));
334 assert(ir
->num_operands
<= 4);
335 for (unsigned i
= 0; i
< ir
->num_operands
; i
++) {
336 if (ir
->operands
[i
]->type
->is_matrix())
340 ir
->operands
[i
]->constant_expression_value(ralloc_parent(ir
));
341 op_expr
[i
] = ir
->operands
[i
]->as_expression();
344 if (this->mem_ctx
== NULL
)
345 this->mem_ctx
= ralloc_parent(ir
);
347 switch (ir
->operation
) {
348 case ir_unop_bit_not
:
349 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_bit_not
)
350 return op_expr
[0]->operands
[0];
354 if (op_expr
[0] == NULL
)
357 switch (op_expr
[0]->operation
) {
360 return abs(op_expr
[0]->operands
[0]);
367 if (op_expr
[0] == NULL
)
370 if (op_expr
[0]->operation
== ir_unop_neg
) {
371 return op_expr
[0]->operands
[0];
376 if (op_expr
[0] == NULL
)
379 if (op_expr
[0]->operation
== ir_unop_log
) {
380 return op_expr
[0]->operands
[0];
385 if (op_expr
[0] == NULL
)
388 if (op_expr
[0]->operation
== ir_unop_exp
) {
389 return op_expr
[0]->operands
[0];
394 if (op_expr
[0] == NULL
)
397 if (op_expr
[0]->operation
== ir_unop_log2
) {
398 return op_expr
[0]->operands
[0];
401 if (!options
->EmitNoPow
&& op_expr
[0]->operation
== ir_binop_mul
) {
402 for (int log2_pos
= 0; log2_pos
< 2; log2_pos
++) {
403 ir_expression
*log2_expr
=
404 op_expr
[0]->operands
[log2_pos
]->as_expression();
406 if (log2_expr
&& log2_expr
->operation
== ir_unop_log2
) {
407 return new(mem_ctx
) ir_expression(ir_binop_pow
,
409 log2_expr
->operands
[0],
410 op_expr
[0]->operands
[1 - log2_pos
]);
417 if (op_expr
[0] == NULL
)
420 if (op_expr
[0]->operation
== ir_unop_exp2
) {
421 return op_expr
[0]->operands
[0];
427 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_trunc
) {
428 return new(mem_ctx
) ir_expression(ir
->operation
,
430 op_expr
[0]->operands
[0]);
434 case ir_unop_logic_not
: {
435 enum ir_expression_operation new_op
= ir_unop_logic_not
;
437 if (op_expr
[0] == NULL
)
440 switch (op_expr
[0]->operation
) {
441 case ir_binop_less
: new_op
= ir_binop_gequal
; break;
442 case ir_binop_gequal
: new_op
= ir_binop_less
; break;
443 case ir_binop_equal
: new_op
= ir_binop_nequal
; break;
444 case ir_binop_nequal
: new_op
= ir_binop_equal
; break;
445 case ir_binop_all_equal
: new_op
= ir_binop_any_nequal
; break;
446 case ir_binop_any_nequal
: new_op
= ir_binop_all_equal
; break;
449 /* The default case handler is here to silence a warning from GCC.
454 if (new_op
!= ir_unop_logic_not
) {
455 return new(mem_ctx
) ir_expression(new_op
,
457 op_expr
[0]->operands
[0],
458 op_expr
[0]->operands
[1]);
464 case ir_unop_saturate
:
465 if (op_expr
[0] && op_expr
[0]->operation
== ir_binop_add
) {
466 ir_expression
*b2f_0
= op_expr
[0]->operands
[0]->as_expression();
467 ir_expression
*b2f_1
= op_expr
[0]->operands
[1]->as_expression();
469 if (b2f_0
&& b2f_0
->operation
== ir_unop_b2f
&&
470 b2f_1
&& b2f_1
->operation
== ir_unop_b2f
) {
471 return b2f(logic_or(b2f_0
->operands
[0], b2f_1
->operands
[0]));
476 /* This macro CANNOT use the do { } while(true) mechanism because
477 * then the breaks apply to the loop instead of the switch!
479 #define HANDLE_PACK_UNPACK_INVERSE(inverse_operation) \
481 ir_expression *const op = ir->operands[0]->as_expression(); \
484 if (op->operation == (inverse_operation)) \
485 return op->operands[0]; \
489 case ir_unop_unpack_uint_2x32
:
490 HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_uint_2x32
);
491 case ir_unop_pack_uint_2x32
:
492 HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_uint_2x32
);
493 case ir_unop_unpack_int_2x32
:
494 HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_int_2x32
);
495 case ir_unop_pack_int_2x32
:
496 HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_int_2x32
);
497 case ir_unop_unpack_double_2x32
:
498 HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_double_2x32
);
499 case ir_unop_pack_double_2x32
:
500 HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_double_2x32
);
502 #undef HANDLE_PACK_UNPACK_INVERSE
505 if (is_vec_zero(op_const
[0]))
506 return ir
->operands
[1];
507 if (is_vec_zero(op_const
[1]))
508 return ir
->operands
[0];
510 /* Replace (x + (-x)) with constant 0 */
511 for (int i
= 0; i
< 2; i
++) {
513 if (op_expr
[i
]->operation
== ir_unop_neg
) {
514 ir_rvalue
*other
= ir
->operands
[(i
+ 1) % 2];
515 if (other
&& op_expr
[i
]->operands
[0]->equals(other
)) {
516 return ir_constant::zero(ir
, ir
->type
);
522 /* Reassociate addition of constants so that we can do constant
525 if (op_const
[0] && !op_const
[1])
526 reassociate_constant(ir
, 0, op_const
[0], op_expr
[1]);
527 if (op_const
[1] && !op_const
[0])
528 reassociate_constant(ir
, 1, op_const
[1], op_expr
[0]);
530 /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
531 if (options
->OptimizeForAOS
) {
532 ir_expression
*expr
= try_replace_with_dot(op_expr
[0], op_expr
[1],
538 /* Replace (-x + y) * a + x and commutative variations with lrp(x, y, a).
541 * (x * -a) + (y * a) + x
542 * x + (x * -a) + (y * a)
543 * x * (1 - a) + y * a
546 for (int mul_pos
= 0; mul_pos
< 2; mul_pos
++) {
547 ir_expression
*mul
= op_expr
[mul_pos
];
549 if (!mul
|| mul
->operation
!= ir_binop_mul
)
552 /* Multiply found on one of the operands. Now check for an
553 * inner addition operation.
555 for (int inner_add_pos
= 0; inner_add_pos
< 2; inner_add_pos
++) {
556 ir_expression
*inner_add
=
557 mul
->operands
[inner_add_pos
]->as_expression();
559 if (!inner_add
|| inner_add
->operation
!= ir_binop_add
)
562 /* Inner addition found on one of the operands. Now check for
563 * one of the operands of the inner addition to be the negative
566 for (int neg_pos
= 0; neg_pos
< 2; neg_pos
++) {
568 inner_add
->operands
[neg_pos
]->as_expression();
570 if (!neg
|| neg
->operation
!= ir_unop_neg
)
573 ir_rvalue
*x_operand
= ir
->operands
[1 - mul_pos
];
575 if (!neg
->operands
[0]->equals(x_operand
))
578 ir_rvalue
*y_operand
= inner_add
->operands
[1 - neg_pos
];
579 ir_rvalue
*a_operand
= mul
->operands
[1 - inner_add_pos
];
581 if (x_operand
->type
!= y_operand
->type
||
582 x_operand
->type
!= a_operand
->type
)
585 return lrp(x_operand
, y_operand
, a_operand
);
593 if (is_vec_zero(op_const
[0]))
594 return neg(ir
->operands
[1]);
595 if (is_vec_zero(op_const
[1]))
596 return ir
->operands
[0];
600 if (is_vec_one(op_const
[0]))
601 return ir
->operands
[1];
602 if (is_vec_one(op_const
[1]))
603 return ir
->operands
[0];
605 if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1]))
606 return ir_constant::zero(ir
, ir
->type
);
608 if (is_vec_negative_one(op_const
[0]))
609 return neg(ir
->operands
[1]);
610 if (is_vec_negative_one(op_const
[1]))
611 return neg(ir
->operands
[0]);
613 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_b2f
&&
614 op_expr
[1] && op_expr
[1]->operation
== ir_unop_b2f
) {
615 return b2f(logic_and(op_expr
[0]->operands
[0], op_expr
[1]->operands
[0]));
618 /* Reassociate multiplication of constants so that we can do
621 if (op_const
[0] && !op_const
[1])
622 reassociate_constant(ir
, 0, op_const
[0], op_expr
[1]);
623 if (op_const
[1] && !op_const
[0])
624 reassociate_constant(ir
, 1, op_const
[1], op_expr
[0]);
628 * (mul (floor (add (abs x) 0.5) (sign x)))
632 * (trunc (add x (mul (sign x) 0.5)))
634 for (int i
= 0; i
< 2; i
++) {
635 ir_expression
*sign_expr
= ir
->operands
[i
]->as_expression();
636 ir_expression
*floor_expr
= ir
->operands
[1 - i
]->as_expression();
638 if (!sign_expr
|| sign_expr
->operation
!= ir_unop_sign
||
639 !floor_expr
|| floor_expr
->operation
!= ir_unop_floor
)
642 ir_expression
*add_expr
= floor_expr
->operands
[0]->as_expression();
643 if (!add_expr
|| add_expr
->operation
!= ir_binop_add
)
646 for (int j
= 0; j
< 2; j
++) {
647 ir_expression
*abs_expr
= add_expr
->operands
[j
]->as_expression();
648 if (!abs_expr
|| abs_expr
->operation
!= ir_unop_abs
)
651 ir_constant
*point_five
= add_expr
->operands
[1 - j
]->as_constant();
652 if (!point_five
|| !point_five
->is_value(0.5, 0))
655 if (abs_expr
->operands
[0]->equals(sign_expr
->operands
[0])) {
656 return trunc(add(abs_expr
->operands
[0],
657 mul(sign_expr
, point_five
)));
664 if (is_vec_one(op_const
[0]) && (
665 ir
->type
->is_float() || ir
->type
->is_double())) {
666 return new(mem_ctx
) ir_expression(ir_unop_rcp
,
667 ir
->operands
[1]->type
,
671 if (is_vec_one(op_const
[1]))
672 return ir
->operands
[0];
676 if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1]))
677 return ir_constant::zero(mem_ctx
, ir
->type
);
679 for (int i
= 0; i
< 2; i
++) {
683 unsigned components
[4] = { 0 }, count
= 0;
685 for (unsigned c
= 0; c
< op_const
[i
]->type
->vector_elements
; c
++) {
686 if (op_const
[i
]->is_zero())
689 components
[count
] = c
;
693 /* No channels had zero values; bail. */
694 if (count
>= op_const
[i
]->type
->vector_elements
)
697 ir_expression_operation op
= count
== 1 ?
698 ir_binop_mul
: ir_binop_dot
;
700 /* Swizzle both operands to remove the channels that were zero. */
702 ir_expression(op
, ir
->type
,
703 new(mem_ctx
) ir_swizzle(ir
->operands
[0],
705 new(mem_ctx
) ir_swizzle(ir
->operands
[1],
711 case ir_binop_gequal
:
713 case ir_binop_nequal
:
714 for (int add_pos
= 0; add_pos
< 2; add_pos
++) {
715 ir_expression
*add
= op_expr
[add_pos
];
717 if (!add
|| add
->operation
!= ir_binop_add
)
720 ir_constant
*zero
= op_const
[1 - add_pos
];
721 if (!is_vec_zero(zero
))
724 /* We are allowed to add scalars with a vector or matrix. In that
725 * case lets just exit early.
727 if (add
->operands
[0]->type
!= add
->operands
[1]->type
)
730 /* Depending of the zero position we want to optimize
731 * (0 cmp x+y) into (-x cmp y) or (x+y cmp 0) into (x cmp -y)
734 return new(mem_ctx
) ir_expression(ir
->operation
,
735 neg(add
->operands
[0]),
738 return new(mem_ctx
) ir_expression(ir
->operation
,
740 neg(add
->operands
[1]));
745 case ir_binop_all_equal
:
746 case ir_binop_any_nequal
:
747 if (ir
->operands
[0]->type
->is_scalar() &&
748 ir
->operands
[1]->type
->is_scalar())
749 return new(mem_ctx
) ir_expression(ir
->operation
== ir_binop_all_equal
750 ? ir_binop_equal
: ir_binop_nequal
,
755 case ir_binop_rshift
:
756 case ir_binop_lshift
:
758 if (is_vec_zero(op_const
[0]))
759 return ir
->operands
[0];
761 if (is_vec_zero(op_const
[1]))
762 return ir
->operands
[0];
765 case ir_binop_logic_and
:
766 if (is_vec_one(op_const
[0])) {
767 return ir
->operands
[1];
768 } else if (is_vec_one(op_const
[1])) {
769 return ir
->operands
[0];
770 } else if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1])) {
771 return ir_constant::zero(mem_ctx
, ir
->type
);
772 } else if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_logic_not
&&
773 op_expr
[1] && op_expr
[1]->operation
== ir_unop_logic_not
) {
775 * (not A) and (not B) === not (A or B)
777 return logic_not(logic_or(op_expr
[0]->operands
[0],
778 op_expr
[1]->operands
[0]));
779 } else if (ir
->operands
[0]->equals(ir
->operands
[1])) {
781 return ir
->operands
[0];
785 case ir_binop_logic_xor
:
786 if (is_vec_zero(op_const
[0])) {
787 return ir
->operands
[1];
788 } else if (is_vec_zero(op_const
[1])) {
789 return ir
->operands
[0];
790 } else if (is_vec_one(op_const
[0])) {
791 return logic_not(ir
->operands
[1]);
792 } else if (is_vec_one(op_const
[1])) {
793 return logic_not(ir
->operands
[0]);
794 } else if (ir
->operands
[0]->equals(ir
->operands
[1])) {
795 /* (a ^^ a) == false */
796 return ir_constant::zero(mem_ctx
, ir
->type
);
800 case ir_binop_logic_or
:
801 if (is_vec_zero(op_const
[0])) {
802 return ir
->operands
[1];
803 } else if (is_vec_zero(op_const
[1])) {
804 return ir
->operands
[0];
805 } else if (is_vec_one(op_const
[0]) || is_vec_one(op_const
[1])) {
806 ir_constant_data data
;
808 for (unsigned i
= 0; i
< 16; i
++)
811 return new(mem_ctx
) ir_constant(ir
->type
, &data
);
812 } else if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_logic_not
&&
813 op_expr
[1] && op_expr
[1]->operation
== ir_unop_logic_not
) {
815 * (not A) or (not B) === not (A and B)
817 return logic_not(logic_and(op_expr
[0]->operands
[0],
818 op_expr
[1]->operands
[0]));
819 } else if (ir
->operands
[0]->equals(ir
->operands
[1])) {
821 return ir
->operands
[0];
827 if (is_vec_one(op_const
[0]))
831 if (is_vec_one(op_const
[1]))
832 return ir
->operands
[0];
834 /* pow(2,x) == exp2(x) */
835 if (is_vec_two(op_const
[0]))
836 return expr(ir_unop_exp2
, ir
->operands
[1]);
838 if (is_vec_two(op_const
[1])) {
839 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[1]->type
, "x",
841 base_ir
->insert_before(x
);
842 base_ir
->insert_before(assign(x
, ir
->operands
[0]));
846 if (is_vec_four(op_const
[1])) {
847 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[1]->type
, "x",
849 base_ir
->insert_before(x
);
850 base_ir
->insert_before(assign(x
, ir
->operands
[0]));
852 ir_variable
*squared
= new(ir
) ir_variable(ir
->operands
[1]->type
,
855 base_ir
->insert_before(squared
);
856 base_ir
->insert_before(assign(squared
, mul(x
, x
)));
857 return mul(squared
, squared
);
864 if (!ir
->type
->is_float() || options
->EmitNoSat
)
867 /* Replace min(max) operations and its commutative combinations with
868 * a saturate operation
870 for (int op
= 0; op
< 2; op
++) {
871 ir_expression
*inner_expr
= op_expr
[op
];
872 ir_constant
*outer_const
= op_const
[1 - op
];
873 ir_expression_operation op_cond
= (ir
->operation
== ir_binop_max
) ?
874 ir_binop_min
: ir_binop_max
;
876 if (!inner_expr
|| !outer_const
|| (inner_expr
->operation
!= op_cond
))
879 /* One of these has to be a constant */
880 if (!inner_expr
->operands
[0]->as_constant() &&
881 !inner_expr
->operands
[1]->as_constant())
884 /* Found a min(max) combination. Now try to see if its operands
885 * meet our conditions that we can do just a single saturate operation
887 for (int minmax_op
= 0; minmax_op
< 2; minmax_op
++) {
888 ir_rvalue
*x
= inner_expr
->operands
[minmax_op
];
889 ir_rvalue
*y
= inner_expr
->operands
[1 - minmax_op
];
891 ir_constant
*inner_const
= y
->as_constant();
895 /* min(max(x, 0.0), 1.0) is sat(x) */
896 if (ir
->operation
== ir_binop_min
&&
897 inner_const
->is_zero() &&
898 outer_const
->is_one())
901 /* max(min(x, 1.0), 0.0) is sat(x) */
902 if (ir
->operation
== ir_binop_max
&&
903 inner_const
->is_one() &&
904 outer_const
->is_zero())
907 /* min(max(x, 0.0), b) where b < 1.0 is sat(min(x, b)) */
908 if (ir
->operation
== ir_binop_min
&&
909 inner_const
->is_zero() &&
910 is_less_than_one(outer_const
))
911 return saturate(expr(ir_binop_min
, x
, outer_const
));
913 /* max(min(x, b), 0.0) where b < 1.0 is sat(min(x, b)) */
914 if (ir
->operation
== ir_binop_max
&&
915 is_less_than_one(inner_const
) &&
916 outer_const
->is_zero())
917 return saturate(expr(ir_binop_min
, x
, inner_const
));
919 /* max(min(x, 1.0), b) where b > 0.0 is sat(max(x, b)) */
920 if (ir
->operation
== ir_binop_max
&&
921 inner_const
->is_one() &&
922 is_greater_than_zero(outer_const
))
923 return saturate(expr(ir_binop_max
, x
, outer_const
));
925 /* min(max(x, b), 1.0) where b > 0.0 is sat(max(x, b)) */
926 if (ir
->operation
== ir_binop_min
&&
927 is_greater_than_zero(inner_const
) &&
928 outer_const
->is_one())
929 return saturate(expr(ir_binop_max
, x
, inner_const
));
936 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_rcp
)
937 return op_expr
[0]->operands
[0];
939 if (op_expr
[0] && (op_expr
[0]->operation
== ir_unop_exp2
||
940 op_expr
[0]->operation
== ir_unop_exp
)) {
941 return new(mem_ctx
) ir_expression(op_expr
[0]->operation
, ir
->type
,
942 neg(op_expr
[0]->operands
[0]));
945 /* While ir_to_mesa.cpp will lower sqrt(x) to rcp(rsq(x)), it does so at
946 * its IR level, so we can always apply this transformation.
948 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_rsq
)
949 return sqrt(op_expr
[0]->operands
[0]);
951 /* As far as we know, all backends are OK with rsq. */
952 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_sqrt
) {
953 return rsq(op_expr
[0]->operands
[0]);
959 /* Operands are op0 * op1 + op2. */
960 if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1])) {
961 return ir
->operands
[2];
962 } else if (is_vec_zero(op_const
[2])) {
963 return mul(ir
->operands
[0], ir
->operands
[1]);
964 } else if (is_vec_one(op_const
[0])) {
965 return add(ir
->operands
[1], ir
->operands
[2]);
966 } else if (is_vec_one(op_const
[1])) {
967 return add(ir
->operands
[0], ir
->operands
[2]);
972 /* Operands are (x, y, a). */
973 if (is_vec_zero(op_const
[2])) {
974 return ir
->operands
[0];
975 } else if (is_vec_one(op_const
[2])) {
976 return ir
->operands
[1];
977 } else if (ir
->operands
[0]->equals(ir
->operands
[1])) {
978 return ir
->operands
[0];
979 } else if (is_vec_zero(op_const
[0])) {
980 return mul(ir
->operands
[1], ir
->operands
[2]);
981 } else if (is_vec_zero(op_const
[1])) {
982 unsigned op2_components
= ir
->operands
[2]->type
->vector_elements
;
985 switch (ir
->type
->base_type
) {
986 case GLSL_TYPE_FLOAT16
:
987 one
= new(mem_ctx
) ir_constant(float16_t::one(), op2_components
);
989 case GLSL_TYPE_FLOAT
:
990 one
= new(mem_ctx
) ir_constant(1.0f
, op2_components
);
992 case GLSL_TYPE_DOUBLE
:
993 one
= new(mem_ctx
) ir_constant(1.0, op2_components
);
997 unreachable("unexpected type");
1000 return mul(ir
->operands
[0], add(one
, neg(ir
->operands
[2])));
1005 if (is_vec_one(op_const
[0]))
1006 return ir
->operands
[1];
1007 if (is_vec_zero(op_const
[0]))
1008 return ir
->operands
[2];
1011 /* Remove interpolateAt* instructions for demoted inputs. They are
1012 * assigned a constant expression to facilitate this.
1014 case ir_unop_interpolate_at_centroid
:
1015 case ir_binop_interpolate_at_offset
:
1016 case ir_binop_interpolate_at_sample
:
1018 return ir
->operands
[0];
1029 ir_algebraic_visitor::handle_rvalue(ir_rvalue
**rvalue
)
1034 ir_expression
*expr
= (*rvalue
)->as_expression();
1035 if (!expr
|| expr
->operation
== ir_quadop_vector
)
1038 ir_rvalue
*new_rvalue
= handle_expression(expr
);
1039 if (new_rvalue
== *rvalue
)
1042 /* If the expr used to be some vec OP scalar returning a vector, and the
1043 * optimization gave us back a scalar, we still need to turn it into a
1046 *rvalue
= swizzle_if_required(expr
, new_rvalue
);
1048 this->progress
= true;
1052 do_algebraic(exec_list
*instructions
, bool native_integers
,
1053 const struct gl_shader_compiler_options
*options
)
1055 ir_algebraic_visitor
v(native_integers
, options
);
1057 visit_list_elements(&v
, instructions
);