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 };
45 static bool alu_instr_is_bool(nir_alu_instr
*instr
);
48 src_is_bool(nir_src src
)
52 if (src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
54 return alu_instr_is_bool(nir_instr_as_alu(src
.ssa
->parent_instr
));
58 alu_instr_is_bool(nir_alu_instr
*instr
)
64 return src_is_bool(instr
->src
[0].src
) && src_is_bool(instr
->src
[1].src
);
66 return src_is_bool(instr
->src
[0].src
);
68 return (nir_alu_type_get_base_type(nir_op_infos
[instr
->op
].output_type
)
74 match_value(const nir_search_value
*value
, nir_alu_instr
*instr
, unsigned src
,
75 unsigned num_components
, const uint8_t *swizzle
,
76 struct match_state
*state
)
78 uint8_t new_swizzle
[4];
80 /* If the source is an explicitly sized source, then we need to reset
81 * both the number of components and the swizzle.
83 if (nir_op_infos
[instr
->op
].input_sizes
[src
] != 0) {
84 num_components
= nir_op_infos
[instr
->op
].input_sizes
[src
];
85 swizzle
= identity_swizzle
;
88 for (unsigned i
= 0; i
< num_components
; ++i
)
89 new_swizzle
[i
] = instr
->src
[src
].swizzle
[swizzle
[i
]];
91 switch (value
->type
) {
92 case nir_search_value_expression
:
93 if (!instr
->src
[src
].src
.is_ssa
)
96 if (instr
->src
[src
].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
99 return match_expression(nir_search_value_as_expression(value
),
100 nir_instr_as_alu(instr
->src
[src
].src
.ssa
->parent_instr
),
101 num_components
, new_swizzle
, state
);
103 case nir_search_value_variable
: {
104 nir_search_variable
*var
= nir_search_value_as_variable(value
);
105 assert(var
->variable
< NIR_SEARCH_MAX_VARIABLES
);
107 if (state
->variables_seen
& (1 << var
->variable
)) {
108 if (!nir_srcs_equal(state
->variables
[var
->variable
].src
,
109 instr
->src
[src
].src
))
112 assert(!instr
->src
[src
].abs
&& !instr
->src
[src
].negate
);
114 for (unsigned i
= 0; i
< num_components
; ++i
) {
115 if (state
->variables
[var
->variable
].swizzle
[i
] != new_swizzle
[i
])
121 if (var
->is_constant
&&
122 instr
->src
[src
].src
.ssa
->parent_instr
->type
!= nir_instr_type_load_const
)
125 if (var
->type
!= nir_type_invalid
) {
126 if (instr
->src
[src
].src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
129 nir_alu_instr
*src_alu
=
130 nir_instr_as_alu(instr
->src
[src
].src
.ssa
->parent_instr
);
132 if (nir_alu_type_get_base_type(nir_op_infos
[src_alu
->op
].output_type
) !=
134 !(nir_alu_type_get_base_type(var
->type
) == nir_type_bool
&&
135 alu_instr_is_bool(src_alu
)))
139 state
->variables_seen
|= (1 << var
->variable
);
140 state
->variables
[var
->variable
].src
= instr
->src
[src
].src
;
141 state
->variables
[var
->variable
].abs
= false;
142 state
->variables
[var
->variable
].negate
= false;
144 for (unsigned i
= 0; i
< 4; ++i
) {
145 if (i
< num_components
)
146 state
->variables
[var
->variable
].swizzle
[i
] = new_swizzle
[i
];
148 state
->variables
[var
->variable
].swizzle
[i
] = 0;
155 case nir_search_value_constant
: {
156 nir_search_constant
*const_val
= nir_search_value_as_constant(value
);
158 if (!instr
->src
[src
].src
.is_ssa
)
161 if (instr
->src
[src
].src
.ssa
->parent_instr
->type
!= nir_instr_type_load_const
)
164 nir_load_const_instr
*load
=
165 nir_instr_as_load_const(instr
->src
[src
].src
.ssa
->parent_instr
);
167 switch (const_val
->type
) {
169 for (unsigned i
= 0; i
< num_components
; ++i
) {
171 switch (load
->def
.bit_size
) {
173 val
= load
->value
.f32
[new_swizzle
[i
]];
176 val
= load
->value
.f64
[new_swizzle
[i
]];
179 unreachable("unknown bit size");
182 if (val
!= const_val
->data
.d
)
188 for (unsigned i
= 0; i
< num_components
; ++i
) {
190 switch (load
->def
.bit_size
) {
192 val
= load
->value
.i32
[new_swizzle
[i
]];
195 val
= load
->value
.i64
[new_swizzle
[i
]];
198 unreachable("unknown bit size");
201 if (val
!= const_val
->data
.i
)
207 case nir_type_bool32
:
208 for (unsigned i
= 0; i
< num_components
; ++i
) {
210 switch (load
->def
.bit_size
) {
212 val
= load
->value
.u32
[new_swizzle
[i
]];
215 val
= load
->value
.u64
[new_swizzle
[i
]];
218 unreachable("unknown bit size");
221 if (val
!= const_val
->data
.u
)
227 unreachable("Invalid alu source type");
232 unreachable("Invalid search value type");
237 match_expression(const nir_search_expression
*expr
, nir_alu_instr
*instr
,
238 unsigned num_components
, const uint8_t *swizzle
,
239 struct match_state
*state
)
241 if (instr
->op
!= expr
->opcode
)
244 assert(instr
->dest
.dest
.is_ssa
);
246 state
->inexact_match
= expr
->inexact
|| state
->inexact_match
;
247 state
->has_exact_alu
= instr
->exact
|| state
->has_exact_alu
;
248 if (state
->inexact_match
&& state
->has_exact_alu
)
251 assert(!instr
->dest
.saturate
);
252 assert(nir_op_infos
[instr
->op
].num_inputs
> 0);
254 /* If we have an explicitly sized destination, we can only handle the
255 * identity swizzle. While dot(vec3(a, b, c).zxy) is a valid
256 * expression, we don't have the information right now to propagate that
257 * swizzle through. We can only properly propagate swizzles if the
258 * instruction is vectorized.
260 if (nir_op_infos
[instr
->op
].output_size
!= 0) {
261 for (unsigned i
= 0; i
< num_components
; i
++) {
267 /* Stash off the current variables_seen bitmask. This way we can
268 * restore it prior to matching in the commutative case below.
270 unsigned variables_seen_stash
= state
->variables_seen
;
273 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
274 if (!match_value(expr
->srcs
[i
], instr
, i
, num_components
,
284 if (nir_op_infos
[instr
->op
].algebraic_properties
& NIR_OP_IS_COMMUTATIVE
) {
285 assert(nir_op_infos
[instr
->op
].num_inputs
== 2);
287 /* Restore the variables_seen bitmask. If we don't do this, then we
288 * could end up with an erroneous failure due to variables found in the
289 * first match attempt above not matching those in the second.
291 state
->variables_seen
= variables_seen_stash
;
293 if (!match_value(expr
->srcs
[0], instr
, 1, num_components
,
297 return match_value(expr
->srcs
[1], instr
, 0, num_components
,
304 typedef struct bitsize_tree
{
306 struct bitsize_tree
*srcs
[4];
308 unsigned common_size
;
309 bool is_src_sized
[4];
313 unsigned src_size
[4];
316 static bitsize_tree
*
317 build_bitsize_tree(void *mem_ctx
, struct match_state
*state
,
318 const nir_search_value
*value
)
320 bitsize_tree
*tree
= ralloc(mem_ctx
, bitsize_tree
);
322 switch (value
->type
) {
323 case nir_search_value_expression
: {
324 nir_search_expression
*expr
= nir_search_value_as_expression(value
);
325 nir_op_info info
= nir_op_infos
[expr
->opcode
];
326 tree
->num_srcs
= info
.num_inputs
;
327 tree
->common_size
= 0;
328 for (unsigned i
= 0; i
< info
.num_inputs
; i
++) {
329 tree
->is_src_sized
[i
] = !!nir_alu_type_get_type_size(info
.input_types
[i
]);
330 if (tree
->is_src_sized
[i
])
331 tree
->src_size
[i
] = nir_alu_type_get_type_size(info
.input_types
[i
]);
332 tree
->srcs
[i
] = build_bitsize_tree(mem_ctx
, state
, expr
->srcs
[i
]);
334 tree
->is_dest_sized
= !!nir_alu_type_get_type_size(info
.output_type
);
335 if (tree
->is_dest_sized
)
336 tree
->dest_size
= nir_alu_type_get_type_size(info
.output_type
);
340 case nir_search_value_variable
: {
341 nir_search_variable
*var
= nir_search_value_as_variable(value
);
343 tree
->is_dest_sized
= true;
344 tree
->dest_size
= nir_src_bit_size(state
->variables
[var
->variable
].src
);
348 case nir_search_value_constant
: {
350 tree
->is_dest_sized
= false;
351 tree
->common_size
= 0;
360 bitsize_tree_filter_up(bitsize_tree
*tree
)
362 for (unsigned i
= 0; i
< tree
->num_srcs
; i
++) {
363 unsigned src_size
= bitsize_tree_filter_up(tree
->srcs
[i
]);
367 if (tree
->is_src_sized
[i
]) {
368 assert(src_size
== tree
->src_size
[i
]);
369 } else if (tree
->common_size
!= 0) {
370 assert(src_size
== tree
->common_size
);
371 tree
->src_size
[i
] = src_size
;
373 tree
->common_size
= src_size
;
374 tree
->src_size
[i
] = src_size
;
378 if (tree
->num_srcs
&& tree
->common_size
) {
379 if (tree
->dest_size
== 0)
380 tree
->dest_size
= tree
->common_size
;
381 else if (!tree
->is_dest_sized
)
382 assert(tree
->dest_size
== tree
->common_size
);
384 for (unsigned i
= 0; i
< tree
->num_srcs
; i
++) {
385 if (!tree
->src_size
[i
])
386 tree
->src_size
[i
] = tree
->common_size
;
390 return tree
->dest_size
;
394 bitsize_tree_filter_down(bitsize_tree
*tree
, unsigned size
)
397 assert(tree
->dest_size
== size
);
399 tree
->dest_size
= size
;
401 if (!tree
->is_dest_sized
) {
402 if (tree
->common_size
)
403 assert(tree
->common_size
== size
);
405 tree
->common_size
= size
;
408 for (unsigned i
= 0; i
< tree
->num_srcs
; i
++) {
409 if (!tree
->src_size
[i
]) {
410 assert(tree
->common_size
);
411 tree
->src_size
[i
] = tree
->common_size
;
413 bitsize_tree_filter_down(tree
->srcs
[i
], tree
->src_size
[i
]);
418 construct_value(const nir_search_value
*value
,
419 unsigned num_components
, bitsize_tree
*bitsize
,
420 struct match_state
*state
,
421 nir_instr
*instr
, void *mem_ctx
)
423 switch (value
->type
) {
424 case nir_search_value_expression
: {
425 const nir_search_expression
*expr
= nir_search_value_as_expression(value
);
427 if (nir_op_infos
[expr
->opcode
].output_size
!= 0)
428 num_components
= nir_op_infos
[expr
->opcode
].output_size
;
430 nir_alu_instr
*alu
= nir_alu_instr_create(mem_ctx
, expr
->opcode
);
431 nir_ssa_dest_init(&alu
->instr
, &alu
->dest
.dest
, num_components
,
432 bitsize
->dest_size
, NULL
);
433 alu
->dest
.write_mask
= (1 << num_components
) - 1;
434 alu
->dest
.saturate
= false;
436 /* We have no way of knowing what values in a given search expression
437 * map to a particular replacement value. Therefore, if the
438 * expression we are replacing has any exact values, the entire
439 * replacement should be exact.
441 alu
->exact
= state
->has_exact_alu
;
443 for (unsigned i
= 0; i
< nir_op_infos
[expr
->opcode
].num_inputs
; i
++) {
444 /* If the source is an explicitly sized source, then we need to reset
445 * the number of components to match.
447 if (nir_op_infos
[alu
->op
].input_sizes
[i
] != 0)
448 num_components
= nir_op_infos
[alu
->op
].input_sizes
[i
];
450 alu
->src
[i
] = construct_value(expr
->srcs
[i
],
451 num_components
, bitsize
->srcs
[i
],
452 state
, instr
, mem_ctx
);
455 nir_instr_insert_before(instr
, &alu
->instr
);
458 val
.src
= nir_src_for_ssa(&alu
->dest
.dest
.ssa
);
461 memcpy(val
.swizzle
, identity_swizzle
, sizeof val
.swizzle
);
466 case nir_search_value_variable
: {
467 const nir_search_variable
*var
= nir_search_value_as_variable(value
);
468 assert(state
->variables_seen
& (1 << var
->variable
));
470 nir_alu_src val
= { NIR_SRC_INIT
};
471 nir_alu_src_copy(&val
, &state
->variables
[var
->variable
], mem_ctx
);
473 assert(!var
->is_constant
);
478 case nir_search_value_constant
: {
479 const nir_search_constant
*c
= nir_search_value_as_constant(value
);
480 nir_load_const_instr
*load
=
481 nir_load_const_instr_create(mem_ctx
, 1, bitsize
->dest_size
);
485 load
->def
.name
= ralloc_asprintf(load
, "%f", c
->data
.d
);
486 switch (bitsize
->dest_size
) {
488 load
->value
.f32
[0] = c
->data
.d
;
491 load
->value
.f64
[0] = c
->data
.d
;
494 unreachable("unknown bit size");
499 load
->def
.name
= ralloc_asprintf(load
, "%" PRIi64
, c
->data
.i
);
500 switch (bitsize
->dest_size
) {
502 load
->value
.i32
[0] = c
->data
.i
;
505 load
->value
.i64
[0] = c
->data
.i
;
508 unreachable("unknown bit size");
513 load
->def
.name
= ralloc_asprintf(load
, "%" PRIu64
, c
->data
.u
);
514 switch (bitsize
->dest_size
) {
516 load
->value
.u32
[0] = c
->data
.u
;
519 load
->value
.u64
[0] = c
->data
.u
;
522 unreachable("unknown bit size");
525 case nir_type_bool32
:
526 load
->value
.u32
[0] = c
->data
.u
;
529 unreachable("Invalid alu source type");
532 nir_instr_insert_before(instr
, &load
->instr
);
535 val
.src
= nir_src_for_ssa(&load
->def
);
538 memset(val
.swizzle
, 0, sizeof val
.swizzle
);
544 unreachable("Invalid search value type");
549 nir_replace_instr(nir_alu_instr
*instr
, const nir_search_expression
*search
,
550 const nir_search_value
*replace
, void *mem_ctx
)
552 uint8_t swizzle
[4] = { 0, 0, 0, 0 };
554 for (unsigned i
= 0; i
< instr
->dest
.dest
.ssa
.num_components
; ++i
)
557 assert(instr
->dest
.dest
.is_ssa
);
559 struct match_state state
;
560 state
.inexact_match
= false;
561 state
.has_exact_alu
= false;
562 state
.variables_seen
= 0;
564 if (!match_expression(search
, instr
, instr
->dest
.dest
.ssa
.num_components
,
568 void *bitsize_ctx
= ralloc_context(NULL
);
569 bitsize_tree
*tree
= build_bitsize_tree(bitsize_ctx
, &state
, replace
);
570 bitsize_tree_filter_up(tree
);
571 bitsize_tree_filter_down(tree
, instr
->dest
.dest
.ssa
.bit_size
);
573 /* Inserting a mov may be unnecessary. However, it's much easier to
574 * simply let copy propagation clean this up than to try to go through
575 * and rewrite swizzles ourselves.
577 nir_alu_instr
*mov
= nir_alu_instr_create(mem_ctx
, nir_op_imov
);
578 mov
->dest
.write_mask
= instr
->dest
.write_mask
;
579 nir_ssa_dest_init(&mov
->instr
, &mov
->dest
.dest
,
580 instr
->dest
.dest
.ssa
.num_components
,
581 instr
->dest
.dest
.ssa
.bit_size
, NULL
);
583 mov
->src
[0] = construct_value(replace
,
584 instr
->dest
.dest
.ssa
.num_components
, tree
,
585 &state
, &instr
->instr
, mem_ctx
);
586 nir_instr_insert_before(&instr
->instr
, &mov
->instr
);
588 nir_ssa_def_rewrite_uses(&instr
->dest
.dest
.ssa
,
589 nir_src_for_ssa(&mov
->dest
.dest
.ssa
));
591 /* We know this one has no more uses because we just rewrote them all,
592 * so we can remove it. The rest of the matched expression, however, we
593 * don't know so much about. We'll just let dead code clean them up.
595 nir_instr_remove(&instr
->instr
);
597 ralloc_free(bitsize_ctx
);