2 * Copyright © 2014 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 DEALINGS
24 * Jason Ekstrand (jason@jlekstrand.net)
29 #include "nir_search.h"
34 unsigned variables_seen
;
35 nir_alu_src variables
[NIR_SEARCH_MAX_VARIABLES
];
39 match_expression(const nir_search_expression
*expr
, nir_alu_instr
*instr
,
40 unsigned num_components
, const uint8_t *swizzle
,
41 struct match_state
*state
);
43 static const uint8_t identity_swizzle
[] = { 0, 1, 2, 3 };
46 * Check if a source produces a value of the given type.
48 * Used for satisfying 'a@type' constraints.
51 src_is_type(nir_src src
, nir_alu_type type
)
53 assert(type
!= nir_type_invalid
);
58 /* Turn nir_type_bool32 into nir_type_bool...they're the same thing. */
59 if (nir_alu_type_get_base_type(type
) == nir_type_bool
)
62 if (src
.ssa
->parent_instr
->type
== nir_instr_type_alu
) {
63 nir_alu_instr
*src_alu
= nir_instr_as_alu(src
.ssa
->parent_instr
);
64 nir_alu_type output_type
= nir_op_infos
[src_alu
->op
].output_type
;
66 if (type
== nir_type_bool
) {
67 switch (src_alu
->op
) {
71 return src_is_type(src_alu
->src
[0].src
, nir_type_bool
) &&
72 src_is_type(src_alu
->src
[1].src
, nir_type_bool
);
74 return src_is_type(src_alu
->src
[0].src
, nir_type_bool
);
80 return nir_alu_type_get_base_type(output_type
) == type
;
81 } else if (src
.ssa
->parent_instr
->type
== nir_instr_type_intrinsic
) {
82 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(src
.ssa
->parent_instr
);
84 if (type
== nir_type_bool
) {
85 return intr
->intrinsic
== nir_intrinsic_load_front_face
||
86 intr
->intrinsic
== nir_intrinsic_load_helper_invocation
;
95 match_value(const nir_search_value
*value
, nir_alu_instr
*instr
, unsigned src
,
96 unsigned num_components
, const uint8_t *swizzle
,
97 struct match_state
*state
)
99 uint8_t new_swizzle
[4];
101 /* Searching only works on SSA values because, if it's not SSA, we can't
102 * know if the value changed between one instance of that value in the
103 * expression and another. Also, the replace operation will place reads of
104 * that value right before the last instruction in the expression we're
105 * replacing so those reads will happen after the original reads and may
106 * not be valid if they're register reads.
108 if (!instr
->src
[src
].src
.is_ssa
)
111 /* If the source is an explicitly sized source, then we need to reset
112 * both the number of components and the swizzle.
114 if (nir_op_infos
[instr
->op
].input_sizes
[src
] != 0) {
115 num_components
= nir_op_infos
[instr
->op
].input_sizes
[src
];
116 swizzle
= identity_swizzle
;
119 for (unsigned i
= 0; i
< num_components
; ++i
)
120 new_swizzle
[i
] = instr
->src
[src
].swizzle
[swizzle
[i
]];
122 /* If the value has a specific bit size and it doesn't match, bail */
123 if (value
->bit_size
&&
124 nir_src_bit_size(instr
->src
[src
].src
) != value
->bit_size
)
127 switch (value
->type
) {
128 case nir_search_value_expression
:
129 if (instr
->src
[src
].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
132 return match_expression(nir_search_value_as_expression(value
),
133 nir_instr_as_alu(instr
->src
[src
].src
.ssa
->parent_instr
),
134 num_components
, new_swizzle
, state
);
136 case nir_search_value_variable
: {
137 nir_search_variable
*var
= nir_search_value_as_variable(value
);
138 assert(var
->variable
< NIR_SEARCH_MAX_VARIABLES
);
140 if (state
->variables_seen
& (1 << var
->variable
)) {
141 if (state
->variables
[var
->variable
].src
.ssa
!= instr
->src
[src
].src
.ssa
)
144 assert(!instr
->src
[src
].abs
&& !instr
->src
[src
].negate
);
146 for (unsigned i
= 0; i
< num_components
; ++i
) {
147 if (state
->variables
[var
->variable
].swizzle
[i
] != new_swizzle
[i
])
153 if (var
->is_constant
&&
154 instr
->src
[src
].src
.ssa
->parent_instr
->type
!= nir_instr_type_load_const
)
157 if (var
->cond
&& !var
->cond(instr
, src
, num_components
, new_swizzle
))
160 if (var
->type
!= nir_type_invalid
&&
161 !src_is_type(instr
->src
[src
].src
, var
->type
))
164 state
->variables_seen
|= (1 << var
->variable
);
165 state
->variables
[var
->variable
].src
= instr
->src
[src
].src
;
166 state
->variables
[var
->variable
].abs
= false;
167 state
->variables
[var
->variable
].negate
= false;
169 for (unsigned i
= 0; i
< 4; ++i
) {
170 if (i
< num_components
)
171 state
->variables
[var
->variable
].swizzle
[i
] = new_swizzle
[i
];
173 state
->variables
[var
->variable
].swizzle
[i
] = 0;
180 case nir_search_value_constant
: {
181 nir_search_constant
*const_val
= nir_search_value_as_constant(value
);
183 if (!instr
->src
[src
].src
.is_ssa
)
186 if (instr
->src
[src
].src
.ssa
->parent_instr
->type
!= nir_instr_type_load_const
)
189 nir_load_const_instr
*load
=
190 nir_instr_as_load_const(instr
->src
[src
].src
.ssa
->parent_instr
);
192 switch (const_val
->type
) {
194 for (unsigned i
= 0; i
< num_components
; ++i
) {
196 switch (load
->def
.bit_size
) {
198 val
= load
->value
.f32
[new_swizzle
[i
]];
201 val
= load
->value
.f64
[new_swizzle
[i
]];
204 unreachable("unknown bit size");
207 if (val
!= const_val
->data
.d
)
214 case nir_type_bool32
:
215 switch (load
->def
.bit_size
) {
217 for (unsigned i
= 0; i
< num_components
; ++i
) {
218 if (load
->value
.u32
[new_swizzle
[i
]] !=
219 (uint32_t)const_val
->data
.u
)
225 for (unsigned i
= 0; i
< num_components
; ++i
) {
226 if (load
->value
.u64
[new_swizzle
[i
]] != const_val
->data
.u
)
232 unreachable("unknown bit size");
236 unreachable("Invalid alu source type");
241 unreachable("Invalid search value type");
246 match_expression(const nir_search_expression
*expr
, nir_alu_instr
*instr
,
247 unsigned num_components
, const uint8_t *swizzle
,
248 struct match_state
*state
)
250 if (expr
->cond
&& !expr
->cond(instr
))
253 if (instr
->op
!= expr
->opcode
)
256 assert(instr
->dest
.dest
.is_ssa
);
258 if (expr
->value
.bit_size
&&
259 instr
->dest
.dest
.ssa
.bit_size
!= expr
->value
.bit_size
)
262 state
->inexact_match
= expr
->inexact
|| state
->inexact_match
;
263 state
->has_exact_alu
= instr
->exact
|| state
->has_exact_alu
;
264 if (state
->inexact_match
&& state
->has_exact_alu
)
267 assert(!instr
->dest
.saturate
);
268 assert(nir_op_infos
[instr
->op
].num_inputs
> 0);
270 /* If we have an explicitly sized destination, we can only handle the
271 * identity swizzle. While dot(vec3(a, b, c).zxy) is a valid
272 * expression, we don't have the information right now to propagate that
273 * swizzle through. We can only properly propagate swizzles if the
274 * instruction is vectorized.
276 if (nir_op_infos
[instr
->op
].output_size
!= 0) {
277 for (unsigned i
= 0; i
< num_components
; i
++) {
283 /* Stash off the current variables_seen bitmask. This way we can
284 * restore it prior to matching in the commutative case below.
286 unsigned variables_seen_stash
= state
->variables_seen
;
289 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
290 if (!match_value(expr
->srcs
[i
], instr
, i
, num_components
,
300 if (nir_op_infos
[instr
->op
].algebraic_properties
& NIR_OP_IS_COMMUTATIVE
) {
301 assert(nir_op_infos
[instr
->op
].num_inputs
== 2);
303 /* Restore the variables_seen bitmask. If we don't do this, then we
304 * could end up with an erroneous failure due to variables found in the
305 * first match attempt above not matching those in the second.
307 state
->variables_seen
= variables_seen_stash
;
309 if (!match_value(expr
->srcs
[0], instr
, 1, num_components
,
313 return match_value(expr
->srcs
[1], instr
, 0, num_components
,
320 typedef struct bitsize_tree
{
322 struct bitsize_tree
*srcs
[4];
324 unsigned common_size
;
325 bool is_src_sized
[4];
329 unsigned src_size
[4];
332 static bitsize_tree
*
333 build_bitsize_tree(void *mem_ctx
, struct match_state
*state
,
334 const nir_search_value
*value
)
336 bitsize_tree
*tree
= rzalloc(mem_ctx
, bitsize_tree
);
338 switch (value
->type
) {
339 case nir_search_value_expression
: {
340 nir_search_expression
*expr
= nir_search_value_as_expression(value
);
341 nir_op_info info
= nir_op_infos
[expr
->opcode
];
342 tree
->num_srcs
= info
.num_inputs
;
343 tree
->common_size
= 0;
344 for (unsigned i
= 0; i
< info
.num_inputs
; i
++) {
345 tree
->is_src_sized
[i
] = !!nir_alu_type_get_type_size(info
.input_types
[i
]);
346 if (tree
->is_src_sized
[i
])
347 tree
->src_size
[i
] = nir_alu_type_get_type_size(info
.input_types
[i
]);
348 tree
->srcs
[i
] = build_bitsize_tree(mem_ctx
, state
, expr
->srcs
[i
]);
350 tree
->is_dest_sized
= !!nir_alu_type_get_type_size(info
.output_type
);
351 if (tree
->is_dest_sized
)
352 tree
->dest_size
= nir_alu_type_get_type_size(info
.output_type
);
356 case nir_search_value_variable
: {
357 nir_search_variable
*var
= nir_search_value_as_variable(value
);
359 tree
->is_dest_sized
= true;
360 tree
->dest_size
= nir_src_bit_size(state
->variables
[var
->variable
].src
);
364 case nir_search_value_constant
: {
366 tree
->is_dest_sized
= false;
367 tree
->common_size
= 0;
372 if (value
->bit_size
) {
373 assert(!tree
->is_dest_sized
|| tree
->dest_size
== value
->bit_size
);
374 tree
->common_size
= value
->bit_size
;
381 bitsize_tree_filter_up(bitsize_tree
*tree
)
383 for (unsigned i
= 0; i
< tree
->num_srcs
; i
++) {
384 unsigned src_size
= bitsize_tree_filter_up(tree
->srcs
[i
]);
388 if (tree
->is_src_sized
[i
]) {
389 assert(src_size
== tree
->src_size
[i
]);
390 } else if (tree
->common_size
!= 0) {
391 assert(src_size
== tree
->common_size
);
392 tree
->src_size
[i
] = src_size
;
394 tree
->common_size
= src_size
;
395 tree
->src_size
[i
] = src_size
;
399 if (tree
->num_srcs
&& tree
->common_size
) {
400 if (tree
->dest_size
== 0)
401 tree
->dest_size
= tree
->common_size
;
402 else if (!tree
->is_dest_sized
)
403 assert(tree
->dest_size
== tree
->common_size
);
405 for (unsigned i
= 0; i
< tree
->num_srcs
; i
++) {
406 if (!tree
->src_size
[i
])
407 tree
->src_size
[i
] = tree
->common_size
;
411 return tree
->dest_size
;
415 bitsize_tree_filter_down(bitsize_tree
*tree
, unsigned size
)
418 assert(tree
->dest_size
== size
);
420 tree
->dest_size
= size
;
422 if (!tree
->is_dest_sized
) {
423 if (tree
->common_size
)
424 assert(tree
->common_size
== size
);
426 tree
->common_size
= size
;
429 for (unsigned i
= 0; i
< tree
->num_srcs
; i
++) {
430 if (!tree
->src_size
[i
]) {
431 assert(tree
->common_size
);
432 tree
->src_size
[i
] = tree
->common_size
;
434 bitsize_tree_filter_down(tree
->srcs
[i
], tree
->src_size
[i
]);
439 construct_value(const nir_search_value
*value
,
440 unsigned num_components
, bitsize_tree
*bitsize
,
441 struct match_state
*state
,
442 nir_instr
*instr
, void *mem_ctx
)
444 switch (value
->type
) {
445 case nir_search_value_expression
: {
446 const nir_search_expression
*expr
= nir_search_value_as_expression(value
);
448 if (nir_op_infos
[expr
->opcode
].output_size
!= 0)
449 num_components
= nir_op_infos
[expr
->opcode
].output_size
;
451 nir_alu_instr
*alu
= nir_alu_instr_create(mem_ctx
, expr
->opcode
);
452 nir_ssa_dest_init(&alu
->instr
, &alu
->dest
.dest
, num_components
,
453 bitsize
->dest_size
, NULL
);
454 alu
->dest
.write_mask
= (1 << num_components
) - 1;
455 alu
->dest
.saturate
= false;
457 /* We have no way of knowing what values in a given search expression
458 * map to a particular replacement value. Therefore, if the
459 * expression we are replacing has any exact values, the entire
460 * replacement should be exact.
462 alu
->exact
= state
->has_exact_alu
;
464 for (unsigned i
= 0; i
< nir_op_infos
[expr
->opcode
].num_inputs
; i
++) {
465 /* If the source is an explicitly sized source, then we need to reset
466 * the number of components to match.
468 if (nir_op_infos
[alu
->op
].input_sizes
[i
] != 0)
469 num_components
= nir_op_infos
[alu
->op
].input_sizes
[i
];
471 alu
->src
[i
] = construct_value(expr
->srcs
[i
],
472 num_components
, bitsize
->srcs
[i
],
473 state
, instr
, mem_ctx
);
476 nir_instr_insert_before(instr
, &alu
->instr
);
479 val
.src
= nir_src_for_ssa(&alu
->dest
.dest
.ssa
);
482 memcpy(val
.swizzle
, identity_swizzle
, sizeof val
.swizzle
);
487 case nir_search_value_variable
: {
488 const nir_search_variable
*var
= nir_search_value_as_variable(value
);
489 assert(state
->variables_seen
& (1 << var
->variable
));
491 nir_alu_src val
= { NIR_SRC_INIT
};
492 nir_alu_src_copy(&val
, &state
->variables
[var
->variable
], mem_ctx
);
494 assert(!var
->is_constant
);
499 case nir_search_value_constant
: {
500 const nir_search_constant
*c
= nir_search_value_as_constant(value
);
501 nir_load_const_instr
*load
=
502 nir_load_const_instr_create(mem_ctx
, 1, bitsize
->dest_size
);
506 load
->def
.name
= ralloc_asprintf(load
, "%f", c
->data
.d
);
507 switch (bitsize
->dest_size
) {
509 load
->value
.f32
[0] = c
->data
.d
;
512 load
->value
.f64
[0] = c
->data
.d
;
515 unreachable("unknown bit size");
520 load
->def
.name
= ralloc_asprintf(load
, "%" PRIi64
, c
->data
.i
);
521 switch (bitsize
->dest_size
) {
523 load
->value
.i32
[0] = c
->data
.i
;
526 load
->value
.i64
[0] = c
->data
.i
;
529 unreachable("unknown bit size");
534 load
->def
.name
= ralloc_asprintf(load
, "%" PRIu64
, c
->data
.u
);
535 switch (bitsize
->dest_size
) {
537 load
->value
.u32
[0] = c
->data
.u
;
540 load
->value
.u64
[0] = c
->data
.u
;
543 unreachable("unknown bit size");
547 case nir_type_bool32
:
548 load
->value
.u32
[0] = c
->data
.u
;
551 unreachable("Invalid alu source type");
554 nir_instr_insert_before(instr
, &load
->instr
);
557 val
.src
= nir_src_for_ssa(&load
->def
);
560 memset(val
.swizzle
, 0, sizeof val
.swizzle
);
566 unreachable("Invalid search value type");
571 nir_replace_instr(nir_alu_instr
*instr
, const nir_search_expression
*search
,
572 const nir_search_value
*replace
, void *mem_ctx
)
574 uint8_t swizzle
[4] = { 0, 0, 0, 0 };
576 for (unsigned i
= 0; i
< instr
->dest
.dest
.ssa
.num_components
; ++i
)
579 assert(instr
->dest
.dest
.is_ssa
);
581 struct match_state state
;
582 state
.inexact_match
= false;
583 state
.has_exact_alu
= false;
584 state
.variables_seen
= 0;
586 if (!match_expression(search
, instr
, instr
->dest
.dest
.ssa
.num_components
,
590 void *bitsize_ctx
= ralloc_context(NULL
);
591 bitsize_tree
*tree
= build_bitsize_tree(bitsize_ctx
, &state
, replace
);
592 bitsize_tree_filter_up(tree
);
593 bitsize_tree_filter_down(tree
, instr
->dest
.dest
.ssa
.bit_size
);
595 /* Inserting a mov may be unnecessary. However, it's much easier to
596 * simply let copy propagation clean this up than to try to go through
597 * and rewrite swizzles ourselves.
599 nir_alu_instr
*mov
= nir_alu_instr_create(mem_ctx
, nir_op_imov
);
600 mov
->dest
.write_mask
= instr
->dest
.write_mask
;
601 nir_ssa_dest_init(&mov
->instr
, &mov
->dest
.dest
,
602 instr
->dest
.dest
.ssa
.num_components
,
603 instr
->dest
.dest
.ssa
.bit_size
, NULL
);
605 mov
->src
[0] = construct_value(replace
,
606 instr
->dest
.dest
.ssa
.num_components
, tree
,
607 &state
, &instr
->instr
, mem_ctx
);
608 nir_instr_insert_before(&instr
->instr
, &mov
->instr
);
610 nir_ssa_def_rewrite_uses(&instr
->dest
.dest
.ssa
,
611 nir_src_for_ssa(&mov
->dest
.dest
.ssa
));
613 /* We know this one has no more uses because we just rewrote them all,
614 * so we can remove it. The rest of the matched expression, however, we
615 * don't know so much about. We'll just let dead code clean them up.
617 nir_instr_remove(&instr
->instr
);
619 ralloc_free(bitsize_ctx
);