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"
38 using namespace ir_builder
;
43 * Visitor class for replacing expressions with ir_constant values.
46 class ir_algebraic_visitor
: public ir_rvalue_visitor
{
48 ir_algebraic_visitor(bool native_integers
,
49 const struct gl_shader_compiler_options
*options
)
52 this->progress
= false;
54 this->native_integers
= native_integers
;
57 virtual ~ir_algebraic_visitor()
61 virtual ir_visitor_status
visit_enter(ir_assignment
*ir
);
63 ir_rvalue
*handle_expression(ir_expression
*ir
);
64 void handle_rvalue(ir_rvalue
**rvalue
);
65 bool reassociate_constant(ir_expression
*ir1
,
67 ir_constant
*constant
,
69 void reassociate_operands(ir_expression
*ir1
,
73 ir_rvalue
*swizzle_if_required(ir_expression
*expr
,
76 const struct gl_shader_compiler_options
*options
;
83 } /* unnamed namespace */
86 ir_algebraic_visitor::visit_enter(ir_assignment
*ir
)
88 ir_variable
*var
= ir
->lhs
->variable_referenced();
89 if (var
->data
.invariant
|| var
->data
.precise
) {
90 /* If we're assigning to an invariant or precise variable, just bail.
91 * Most of the algebraic optimizations aren't precision-safe.
93 * FINISHME: Find out which optimizations are precision-safe and enable
94 * then only for invariant or precise trees.
96 return visit_continue_with_parent
;
98 return visit_continue
;
103 is_vec_zero(ir_constant
*ir
)
105 return (ir
== NULL
) ? false : ir
->is_zero();
109 is_vec_one(ir_constant
*ir
)
111 return (ir
== NULL
) ? false : ir
->is_one();
115 is_vec_two(ir_constant
*ir
)
117 return (ir
== NULL
) ? false : ir
->is_value(2.0, 2);
121 is_vec_four(ir_constant
*ir
)
123 return (ir
== NULL
) ? false : ir
->is_value(4.0, 4);
127 is_vec_negative_one(ir_constant
*ir
)
129 return (ir
== NULL
) ? false : ir
->is_negative_one();
133 is_valid_vec_const(ir_constant
*ir
)
138 if (!ir
->type
->is_scalar() && !ir
->type
->is_vector())
145 is_less_than_one(ir_constant
*ir
)
147 assert(ir
->type
->is_float());
149 if (!is_valid_vec_const(ir
))
152 unsigned component
= 0;
153 for (int c
= 0; c
< ir
->type
->vector_elements
; c
++) {
154 if (ir
->get_float_component(c
) < 1.0f
)
158 return (component
== ir
->type
->vector_elements
);
162 is_greater_than_zero(ir_constant
*ir
)
164 assert(ir
->type
->is_float());
166 if (!is_valid_vec_const(ir
))
169 unsigned component
= 0;
170 for (int c
= 0; c
< ir
->type
->vector_elements
; c
++) {
171 if (ir
->get_float_component(c
) > 0.0f
)
175 return (component
== ir
->type
->vector_elements
);
179 update_type(ir_expression
*ir
)
181 if (ir
->operands
[0]->type
->is_vector())
182 ir
->type
= ir
->operands
[0]->type
;
184 ir
->type
= ir
->operands
[1]->type
;
187 /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
188 static ir_expression
*
189 try_replace_with_dot(ir_expression
*expr0
, ir_expression
*expr1
, void *mem_ctx
)
191 if (expr0
&& expr0
->operation
== ir_binop_add
&&
192 expr0
->type
->is_float() &&
193 expr1
&& expr1
->operation
== ir_binop_add
&&
194 expr1
->type
->is_float()) {
195 ir_swizzle
*x
= expr0
->operands
[0]->as_swizzle();
196 ir_swizzle
*y
= expr0
->operands
[1]->as_swizzle();
197 ir_swizzle
*z
= expr1
->operands
[0]->as_swizzle();
198 ir_swizzle
*w
= expr1
->operands
[1]->as_swizzle();
200 if (!x
|| x
->mask
.num_components
!= 1 ||
201 !y
|| y
->mask
.num_components
!= 1 ||
202 !z
|| z
->mask
.num_components
!= 1 ||
203 !w
|| w
->mask
.num_components
!= 1) {
207 bool swiz_seen
[4] = {false, false, false, false};
208 swiz_seen
[x
->mask
.x
] = true;
209 swiz_seen
[y
->mask
.x
] = true;
210 swiz_seen
[z
->mask
.x
] = true;
211 swiz_seen
[w
->mask
.x
] = true;
213 if (!swiz_seen
[0] || !swiz_seen
[1] ||
214 !swiz_seen
[2] || !swiz_seen
[3]) {
218 if (x
->val
->equals(y
->val
) &&
219 x
->val
->equals(z
->val
) &&
220 x
->val
->equals(w
->val
)) {
221 return dot(x
->val
, new(mem_ctx
) ir_constant(1.0f
, 4));
228 ir_algebraic_visitor::reassociate_operands(ir_expression
*ir1
,
233 ir_rvalue
*temp
= ir2
->operands
[op2
];
234 ir2
->operands
[op2
] = ir1
->operands
[op1
];
235 ir1
->operands
[op1
] = temp
;
237 /* Update the type of ir2. The type of ir1 won't have changed --
238 * base types matched, and at least one of the operands of the 2
239 * binops is still a vector if any of them were.
243 this->progress
= true;
247 * Reassociates a constant down a tree of adds or multiplies.
249 * Consider (2 * (a * (b * 0.5))). We want to end up with a * b.
252 ir_algebraic_visitor::reassociate_constant(ir_expression
*ir1
, int const_index
,
253 ir_constant
*constant
,
256 if (!ir2
|| ir1
->operation
!= ir2
->operation
)
259 /* Don't want to even think about matrices. */
260 if (ir1
->operands
[0]->type
->is_matrix() ||
261 ir1
->operands
[1]->type
->is_matrix() ||
262 ir2
->operands
[0]->type
->is_matrix() ||
263 ir2
->operands
[1]->type
->is_matrix())
266 void *mem_ctx
= ralloc_parent(ir2
);
268 ir_constant
*ir2_const
[2];
269 ir2_const
[0] = ir2
->operands
[0]->constant_expression_value(mem_ctx
);
270 ir2_const
[1] = ir2
->operands
[1]->constant_expression_value(mem_ctx
);
272 if (ir2_const
[0] && ir2_const
[1])
276 reassociate_operands(ir1
, const_index
, ir2
, 1);
278 } else if (ir2_const
[1]) {
279 reassociate_operands(ir1
, const_index
, ir2
, 0);
283 if (reassociate_constant(ir1
, const_index
, constant
,
284 ir2
->operands
[0]->as_expression())) {
289 if (reassociate_constant(ir1
, const_index
, constant
,
290 ir2
->operands
[1]->as_expression())) {
298 /* When eliminating an expression and just returning one of its operands,
299 * we may need to swizzle that operand out to a vector if the expression was
303 ir_algebraic_visitor::swizzle_if_required(ir_expression
*expr
,
306 if (expr
->type
->is_vector() && operand
->type
->is_scalar()) {
307 return new(mem_ctx
) ir_swizzle(operand
, 0, 0, 0, 0,
308 expr
->type
->vector_elements
);
314 ir_algebraic_visitor::handle_expression(ir_expression
*ir
)
316 ir_constant
*op_const
[4] = {NULL
, NULL
, NULL
, NULL
};
317 ir_expression
*op_expr
[4] = {NULL
, NULL
, NULL
, NULL
};
319 if (ir
->operation
== ir_binop_mul
&&
320 ir
->operands
[0]->type
->is_matrix() &&
321 ir
->operands
[1]->type
->is_vector()) {
322 ir_expression
*matrix_mul
= ir
->operands
[0]->as_expression();
324 if (matrix_mul
&& matrix_mul
->operation
== ir_binop_mul
&&
325 matrix_mul
->operands
[0]->type
->is_matrix() &&
326 matrix_mul
->operands
[1]->type
->is_matrix()) {
328 return mul(matrix_mul
->operands
[0],
329 mul(matrix_mul
->operands
[1], ir
->operands
[1]));
333 assert(ir
->num_operands
<= 4);
334 for (unsigned i
= 0; i
< ir
->num_operands
; i
++) {
335 if (ir
->operands
[i
]->type
->is_matrix())
339 ir
->operands
[i
]->constant_expression_value(ralloc_parent(ir
));
340 op_expr
[i
] = ir
->operands
[i
]->as_expression();
343 if (this->mem_ctx
== NULL
)
344 this->mem_ctx
= ralloc_parent(ir
);
346 switch (ir
->operation
) {
347 case ir_unop_bit_not
:
348 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_bit_not
)
349 return op_expr
[0]->operands
[0];
353 if (op_expr
[0] == NULL
)
356 switch (op_expr
[0]->operation
) {
359 return abs(op_expr
[0]->operands
[0]);
366 if (op_expr
[0] == NULL
)
369 if (op_expr
[0]->operation
== ir_unop_neg
) {
370 return op_expr
[0]->operands
[0];
375 if (op_expr
[0] == NULL
)
378 if (op_expr
[0]->operation
== ir_unop_log
) {
379 return op_expr
[0]->operands
[0];
384 if (op_expr
[0] == NULL
)
387 if (op_expr
[0]->operation
== ir_unop_exp
) {
388 return op_expr
[0]->operands
[0];
393 if (op_expr
[0] == NULL
)
396 if (op_expr
[0]->operation
== ir_unop_log2
) {
397 return op_expr
[0]->operands
[0];
400 if (!options
->EmitNoPow
&& op_expr
[0]->operation
== ir_binop_mul
) {
401 for (int log2_pos
= 0; log2_pos
< 2; log2_pos
++) {
402 ir_expression
*log2_expr
=
403 op_expr
[0]->operands
[log2_pos
]->as_expression();
405 if (log2_expr
&& log2_expr
->operation
== ir_unop_log2
) {
406 return new(mem_ctx
) ir_expression(ir_binop_pow
,
408 log2_expr
->operands
[0],
409 op_expr
[0]->operands
[1 - log2_pos
]);
416 if (op_expr
[0] == NULL
)
419 if (op_expr
[0]->operation
== ir_unop_exp2
) {
420 return op_expr
[0]->operands
[0];
426 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_trunc
) {
427 return new(mem_ctx
) ir_expression(ir
->operation
,
429 op_expr
[0]->operands
[0]);
433 case ir_unop_logic_not
: {
434 enum ir_expression_operation new_op
= ir_unop_logic_not
;
436 if (op_expr
[0] == NULL
)
439 switch (op_expr
[0]->operation
) {
440 case ir_binop_less
: new_op
= ir_binop_gequal
; break;
441 case ir_binop_gequal
: new_op
= ir_binop_less
; break;
442 case ir_binop_equal
: new_op
= ir_binop_nequal
; break;
443 case ir_binop_nequal
: new_op
= ir_binop_equal
; break;
444 case ir_binop_all_equal
: new_op
= ir_binop_any_nequal
; break;
445 case ir_binop_any_nequal
: new_op
= ir_binop_all_equal
; break;
448 /* The default case handler is here to silence a warning from GCC.
453 if (new_op
!= ir_unop_logic_not
) {
454 return new(mem_ctx
) ir_expression(new_op
,
456 op_expr
[0]->operands
[0],
457 op_expr
[0]->operands
[1]);
463 case ir_unop_saturate
:
464 if (op_expr
[0] && op_expr
[0]->operation
== ir_binop_add
) {
465 ir_expression
*b2f_0
= op_expr
[0]->operands
[0]->as_expression();
466 ir_expression
*b2f_1
= op_expr
[0]->operands
[1]->as_expression();
468 if (b2f_0
&& b2f_0
->operation
== ir_unop_b2f
&&
469 b2f_1
&& b2f_1
->operation
== ir_unop_b2f
) {
470 return b2f(logic_or(b2f_0
->operands
[0], b2f_1
->operands
[0]));
475 /* This macro CANNOT use the do { } while(true) mechanism because
476 * then the breaks apply to the loop instead of the switch!
478 #define HANDLE_PACK_UNPACK_INVERSE(inverse_operation) \
480 ir_expression *const op = ir->operands[0]->as_expression(); \
483 if (op->operation == (inverse_operation)) \
484 return op->operands[0]; \
488 case ir_unop_unpack_uint_2x32
:
489 HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_uint_2x32
);
490 case ir_unop_pack_uint_2x32
:
491 HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_uint_2x32
);
492 case ir_unop_unpack_int_2x32
:
493 HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_int_2x32
);
494 case ir_unop_pack_int_2x32
:
495 HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_int_2x32
);
496 case ir_unop_unpack_double_2x32
:
497 HANDLE_PACK_UNPACK_INVERSE(ir_unop_pack_double_2x32
);
498 case ir_unop_pack_double_2x32
:
499 HANDLE_PACK_UNPACK_INVERSE(ir_unop_unpack_double_2x32
);
501 #undef HANDLE_PACK_UNPACK_INVERSE
504 if (is_vec_zero(op_const
[0]))
505 return ir
->operands
[1];
506 if (is_vec_zero(op_const
[1]))
507 return ir
->operands
[0];
509 /* Reassociate addition of constants so that we can do constant
512 if (op_const
[0] && !op_const
[1])
513 reassociate_constant(ir
, 0, op_const
[0], op_expr
[1]);
514 if (op_const
[1] && !op_const
[0])
515 reassociate_constant(ir
, 1, op_const
[1], op_expr
[0]);
517 /* Recognize (v.x + v.y) + (v.z + v.w) as dot(v, 1.0) */
518 if (options
->OptimizeForAOS
) {
519 ir_expression
*expr
= try_replace_with_dot(op_expr
[0], op_expr
[1],
525 /* Replace (-x + y) * a + x and commutative variations with lrp(x, y, a).
528 * (x * -a) + (y * a) + x
529 * x + (x * -a) + (y * a)
530 * x * (1 - a) + y * a
533 for (int mul_pos
= 0; mul_pos
< 2; mul_pos
++) {
534 ir_expression
*mul
= op_expr
[mul_pos
];
536 if (!mul
|| mul
->operation
!= ir_binop_mul
)
539 /* Multiply found on one of the operands. Now check for an
540 * inner addition operation.
542 for (int inner_add_pos
= 0; inner_add_pos
< 2; inner_add_pos
++) {
543 ir_expression
*inner_add
=
544 mul
->operands
[inner_add_pos
]->as_expression();
546 if (!inner_add
|| inner_add
->operation
!= ir_binop_add
)
549 /* Inner addition found on one of the operands. Now check for
550 * one of the operands of the inner addition to be the negative
553 for (int neg_pos
= 0; neg_pos
< 2; neg_pos
++) {
555 inner_add
->operands
[neg_pos
]->as_expression();
557 if (!neg
|| neg
->operation
!= ir_unop_neg
)
560 ir_rvalue
*x_operand
= ir
->operands
[1 - mul_pos
];
562 if (!neg
->operands
[0]->equals(x_operand
))
565 ir_rvalue
*y_operand
= inner_add
->operands
[1 - neg_pos
];
566 ir_rvalue
*a_operand
= mul
->operands
[1 - inner_add_pos
];
568 if (x_operand
->type
!= y_operand
->type
||
569 x_operand
->type
!= a_operand
->type
)
572 return lrp(x_operand
, y_operand
, a_operand
);
580 if (is_vec_zero(op_const
[0]))
581 return neg(ir
->operands
[1]);
582 if (is_vec_zero(op_const
[1]))
583 return ir
->operands
[0];
587 if (is_vec_one(op_const
[0]))
588 return ir
->operands
[1];
589 if (is_vec_one(op_const
[1]))
590 return ir
->operands
[0];
592 if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1]))
593 return ir_constant::zero(ir
, ir
->type
);
595 if (is_vec_negative_one(op_const
[0]))
596 return neg(ir
->operands
[1]);
597 if (is_vec_negative_one(op_const
[1]))
598 return neg(ir
->operands
[0]);
600 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_b2f
&&
601 op_expr
[1] && op_expr
[1]->operation
== ir_unop_b2f
) {
602 return b2f(logic_and(op_expr
[0]->operands
[0], op_expr
[1]->operands
[0]));
605 /* Reassociate multiplication of constants so that we can do
608 if (op_const
[0] && !op_const
[1])
609 reassociate_constant(ir
, 0, op_const
[0], op_expr
[1]);
610 if (op_const
[1] && !op_const
[0])
611 reassociate_constant(ir
, 1, op_const
[1], op_expr
[0]);
615 * (mul (floor (add (abs x) 0.5) (sign x)))
619 * (trunc (add x (mul (sign x) 0.5)))
621 for (int i
= 0; i
< 2; i
++) {
622 ir_expression
*sign_expr
= ir
->operands
[i
]->as_expression();
623 ir_expression
*floor_expr
= ir
->operands
[1 - i
]->as_expression();
625 if (!sign_expr
|| sign_expr
->operation
!= ir_unop_sign
||
626 !floor_expr
|| floor_expr
->operation
!= ir_unop_floor
)
629 ir_expression
*add_expr
= floor_expr
->operands
[0]->as_expression();
630 if (!add_expr
|| add_expr
->operation
!= ir_binop_add
)
633 for (int j
= 0; j
< 2; j
++) {
634 ir_expression
*abs_expr
= add_expr
->operands
[j
]->as_expression();
635 if (!abs_expr
|| abs_expr
->operation
!= ir_unop_abs
)
638 ir_constant
*point_five
= add_expr
->operands
[1 - j
]->as_constant();
639 if (!point_five
|| !point_five
->is_value(0.5, 0))
642 if (abs_expr
->operands
[0]->equals(sign_expr
->operands
[0])) {
643 return trunc(add(abs_expr
->operands
[0],
644 mul(sign_expr
, point_five
)));
651 if (is_vec_one(op_const
[0]) && (
652 ir
->type
->is_float() || ir
->type
->is_double())) {
653 return new(mem_ctx
) ir_expression(ir_unop_rcp
,
654 ir
->operands
[1]->type
,
658 if (is_vec_one(op_const
[1]))
659 return ir
->operands
[0];
663 if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1]))
664 return ir_constant::zero(mem_ctx
, ir
->type
);
666 for (int i
= 0; i
< 2; i
++) {
670 unsigned components
[4] = { 0 }, count
= 0;
672 for (unsigned c
= 0; c
< op_const
[i
]->type
->vector_elements
; c
++) {
673 if (op_const
[i
]->is_zero())
676 components
[count
] = c
;
680 /* No channels had zero values; bail. */
681 if (count
>= op_const
[i
]->type
->vector_elements
)
684 ir_expression_operation op
= count
== 1 ?
685 ir_binop_mul
: ir_binop_dot
;
687 /* Swizzle both operands to remove the channels that were zero. */
689 ir_expression(op
, ir
->type
,
690 new(mem_ctx
) ir_swizzle(ir
->operands
[0],
692 new(mem_ctx
) ir_swizzle(ir
->operands
[1],
698 case ir_binop_gequal
:
700 case ir_binop_nequal
:
701 for (int add_pos
= 0; add_pos
< 2; add_pos
++) {
702 ir_expression
*add
= op_expr
[add_pos
];
704 if (!add
|| add
->operation
!= ir_binop_add
)
707 ir_constant
*zero
= op_const
[1 - add_pos
];
708 if (!is_vec_zero(zero
))
711 /* Depending of the zero position we want to optimize
712 * (0 cmp x+y) into (-x cmp y) or (x+y cmp 0) into (x cmp -y)
715 return new(mem_ctx
) ir_expression(ir
->operation
,
716 neg(add
->operands
[0]),
719 return new(mem_ctx
) ir_expression(ir
->operation
,
721 neg(add
->operands
[1]));
726 case ir_binop_all_equal
:
727 case ir_binop_any_nequal
:
728 if (ir
->operands
[0]->type
->is_scalar() &&
729 ir
->operands
[1]->type
->is_scalar())
730 return new(mem_ctx
) ir_expression(ir
->operation
== ir_binop_all_equal
731 ? ir_binop_equal
: ir_binop_nequal
,
736 case ir_binop_rshift
:
737 case ir_binop_lshift
:
739 if (is_vec_zero(op_const
[0]))
740 return ir
->operands
[0];
742 if (is_vec_zero(op_const
[1]))
743 return ir
->operands
[0];
746 case ir_binop_logic_and
:
747 if (is_vec_one(op_const
[0])) {
748 return ir
->operands
[1];
749 } else if (is_vec_one(op_const
[1])) {
750 return ir
->operands
[0];
751 } else if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1])) {
752 return ir_constant::zero(mem_ctx
, ir
->type
);
753 } else if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_logic_not
&&
754 op_expr
[1] && op_expr
[1]->operation
== ir_unop_logic_not
) {
756 * (not A) and (not B) === not (A or B)
758 return logic_not(logic_or(op_expr
[0]->operands
[0],
759 op_expr
[1]->operands
[0]));
760 } else if (ir
->operands
[0]->equals(ir
->operands
[1])) {
762 return ir
->operands
[0];
766 case ir_binop_logic_xor
:
767 if (is_vec_zero(op_const
[0])) {
768 return ir
->operands
[1];
769 } else if (is_vec_zero(op_const
[1])) {
770 return ir
->operands
[0];
771 } else if (is_vec_one(op_const
[0])) {
772 return logic_not(ir
->operands
[1]);
773 } else if (is_vec_one(op_const
[1])) {
774 return logic_not(ir
->operands
[0]);
775 } else if (ir
->operands
[0]->equals(ir
->operands
[1])) {
776 /* (a ^^ a) == false */
777 return ir_constant::zero(mem_ctx
, ir
->type
);
781 case ir_binop_logic_or
:
782 if (is_vec_zero(op_const
[0])) {
783 return ir
->operands
[1];
784 } else if (is_vec_zero(op_const
[1])) {
785 return ir
->operands
[0];
786 } else if (is_vec_one(op_const
[0]) || is_vec_one(op_const
[1])) {
787 ir_constant_data data
;
789 for (unsigned i
= 0; i
< 16; i
++)
792 return new(mem_ctx
) ir_constant(ir
->type
, &data
);
793 } else if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_logic_not
&&
794 op_expr
[1] && op_expr
[1]->operation
== ir_unop_logic_not
) {
796 * (not A) or (not B) === not (A and B)
798 return logic_not(logic_and(op_expr
[0]->operands
[0],
799 op_expr
[1]->operands
[0]));
800 } else if (ir
->operands
[0]->equals(ir
->operands
[1])) {
802 return ir
->operands
[0];
808 if (is_vec_one(op_const
[0]))
812 if (is_vec_one(op_const
[1]))
813 return ir
->operands
[0];
815 /* pow(2,x) == exp2(x) */
816 if (is_vec_two(op_const
[0]))
817 return expr(ir_unop_exp2
, ir
->operands
[1]);
819 if (is_vec_two(op_const
[1])) {
820 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[1]->type
, "x",
822 base_ir
->insert_before(x
);
823 base_ir
->insert_before(assign(x
, ir
->operands
[0]));
827 if (is_vec_four(op_const
[1])) {
828 ir_variable
*x
= new(ir
) ir_variable(ir
->operands
[1]->type
, "x",
830 base_ir
->insert_before(x
);
831 base_ir
->insert_before(assign(x
, ir
->operands
[0]));
833 ir_variable
*squared
= new(ir
) ir_variable(ir
->operands
[1]->type
,
836 base_ir
->insert_before(squared
);
837 base_ir
->insert_before(assign(squared
, mul(x
, x
)));
838 return mul(squared
, squared
);
845 if (!ir
->type
->is_float() || options
->EmitNoSat
)
848 /* Replace min(max) operations and its commutative combinations with
849 * a saturate operation
851 for (int op
= 0; op
< 2; op
++) {
852 ir_expression
*inner_expr
= op_expr
[op
];
853 ir_constant
*outer_const
= op_const
[1 - op
];
854 ir_expression_operation op_cond
= (ir
->operation
== ir_binop_max
) ?
855 ir_binop_min
: ir_binop_max
;
857 if (!inner_expr
|| !outer_const
|| (inner_expr
->operation
!= op_cond
))
860 /* One of these has to be a constant */
861 if (!inner_expr
->operands
[0]->as_constant() &&
862 !inner_expr
->operands
[1]->as_constant())
865 /* Found a min(max) combination. Now try to see if its operands
866 * meet our conditions that we can do just a single saturate operation
868 for (int minmax_op
= 0; minmax_op
< 2; minmax_op
++) {
869 ir_rvalue
*x
= inner_expr
->operands
[minmax_op
];
870 ir_rvalue
*y
= inner_expr
->operands
[1 - minmax_op
];
872 ir_constant
*inner_const
= y
->as_constant();
876 /* min(max(x, 0.0), 1.0) is sat(x) */
877 if (ir
->operation
== ir_binop_min
&&
878 inner_const
->is_zero() &&
879 outer_const
->is_one())
882 /* max(min(x, 1.0), 0.0) is sat(x) */
883 if (ir
->operation
== ir_binop_max
&&
884 inner_const
->is_one() &&
885 outer_const
->is_zero())
888 /* min(max(x, 0.0), b) where b < 1.0 is sat(min(x, b)) */
889 if (ir
->operation
== ir_binop_min
&&
890 inner_const
->is_zero() &&
891 is_less_than_one(outer_const
))
892 return saturate(expr(ir_binop_min
, x
, outer_const
));
894 /* max(min(x, b), 0.0) where b < 1.0 is sat(min(x, b)) */
895 if (ir
->operation
== ir_binop_max
&&
896 is_less_than_one(inner_const
) &&
897 outer_const
->is_zero())
898 return saturate(expr(ir_binop_min
, x
, inner_const
));
900 /* max(min(x, 1.0), b) where b > 0.0 is sat(max(x, b)) */
901 if (ir
->operation
== ir_binop_max
&&
902 inner_const
->is_one() &&
903 is_greater_than_zero(outer_const
))
904 return saturate(expr(ir_binop_max
, x
, outer_const
));
906 /* min(max(x, b), 1.0) where b > 0.0 is sat(max(x, b)) */
907 if (ir
->operation
== ir_binop_min
&&
908 is_greater_than_zero(inner_const
) &&
909 outer_const
->is_one())
910 return saturate(expr(ir_binop_max
, x
, inner_const
));
917 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_rcp
)
918 return op_expr
[0]->operands
[0];
920 if (op_expr
[0] && (op_expr
[0]->operation
== ir_unop_exp2
||
921 op_expr
[0]->operation
== ir_unop_exp
)) {
922 return new(mem_ctx
) ir_expression(op_expr
[0]->operation
, ir
->type
,
923 neg(op_expr
[0]->operands
[0]));
926 /* While ir_to_mesa.cpp will lower sqrt(x) to rcp(rsq(x)), it does so at
927 * its IR level, so we can always apply this transformation.
929 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_rsq
)
930 return sqrt(op_expr
[0]->operands
[0]);
932 /* As far as we know, all backends are OK with rsq. */
933 if (op_expr
[0] && op_expr
[0]->operation
== ir_unop_sqrt
) {
934 return rsq(op_expr
[0]->operands
[0]);
940 /* Operands are op0 * op1 + op2. */
941 if (is_vec_zero(op_const
[0]) || is_vec_zero(op_const
[1])) {
942 return ir
->operands
[2];
943 } else if (is_vec_zero(op_const
[2])) {
944 return mul(ir
->operands
[0], ir
->operands
[1]);
945 } else if (is_vec_one(op_const
[0])) {
946 return add(ir
->operands
[1], ir
->operands
[2]);
947 } else if (is_vec_one(op_const
[1])) {
948 return add(ir
->operands
[0], ir
->operands
[2]);
953 /* Operands are (x, y, a). */
954 if (is_vec_zero(op_const
[2])) {
955 return ir
->operands
[0];
956 } else if (is_vec_one(op_const
[2])) {
957 return ir
->operands
[1];
958 } else if (ir
->operands
[0]->equals(ir
->operands
[1])) {
959 return ir
->operands
[0];
960 } else if (is_vec_zero(op_const
[0])) {
961 return mul(ir
->operands
[1], ir
->operands
[2]);
962 } else if (is_vec_zero(op_const
[1])) {
963 unsigned op2_components
= ir
->operands
[2]->type
->vector_elements
;
966 switch (ir
->type
->base_type
) {
967 case GLSL_TYPE_FLOAT
:
968 one
= new(mem_ctx
) ir_constant(1.0f
, op2_components
);
970 case GLSL_TYPE_DOUBLE
:
971 one
= new(mem_ctx
) ir_constant(1.0, op2_components
);
975 unreachable("unexpected type");
978 return mul(ir
->operands
[0], add(one
, neg(ir
->operands
[2])));
983 if (is_vec_one(op_const
[0]))
984 return ir
->operands
[1];
985 if (is_vec_zero(op_const
[0]))
986 return ir
->operands
[2];
989 /* Remove interpolateAt* instructions for demoted inputs. They are
990 * assigned a constant expression to facilitate this.
992 case ir_unop_interpolate_at_centroid
:
993 case ir_binop_interpolate_at_offset
:
994 case ir_binop_interpolate_at_sample
:
996 return ir
->operands
[0];
1007 ir_algebraic_visitor::handle_rvalue(ir_rvalue
**rvalue
)
1012 ir_expression
*expr
= (*rvalue
)->as_expression();
1013 if (!expr
|| expr
->operation
== ir_quadop_vector
)
1016 ir_rvalue
*new_rvalue
= handle_expression(expr
);
1017 if (new_rvalue
== *rvalue
)
1020 /* If the expr used to be some vec OP scalar returning a vector, and the
1021 * optimization gave us back a scalar, we still need to turn it into a
1024 *rvalue
= swizzle_if_required(expr
, new_rvalue
);
1026 this->progress
= true;
1030 do_algebraic(exec_list
*instructions
, bool native_integers
,
1031 const struct gl_shader_compiler_options
*options
)
1033 ir_algebraic_visitor
v(native_integers
, options
);
1035 visit_list_elements(&v
, instructions
);