2 * Copyright © 2014 Connor Abbott
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 #include "nir_instr_set.h"
26 #include "util/half_float.h"
29 src_is_ssa(nir_src
*src
, void *data
)
36 dest_is_ssa(nir_dest
*dest
, void *data
)
43 instr_each_src_and_dest_is_ssa(const nir_instr
*instr
)
45 if (!nir_foreach_dest((nir_instr
*)instr
, dest_is_ssa
, NULL
) ||
46 !nir_foreach_src((nir_instr
*)instr
, src_is_ssa
, NULL
))
52 /* This function determines if uses of an instruction can safely be rewritten
53 * to use another identical instruction instead. Note that this function must
54 * be kept in sync with hash_instr() and nir_instrs_equal() -- only
55 * instructions that pass this test will be handed on to those functions, and
56 * conversely they must handle everything that this function returns true for.
59 instr_can_rewrite(const nir_instr
*instr
)
61 /* We only handle SSA. */
62 assert(instr_each_src_and_dest_is_ssa(instr
));
64 switch (instr
->type
) {
65 case nir_instr_type_alu
:
66 case nir_instr_type_deref
:
67 case nir_instr_type_tex
:
68 case nir_instr_type_load_const
:
69 case nir_instr_type_phi
:
71 case nir_instr_type_intrinsic
:
72 return nir_intrinsic_can_reorder(nir_instr_as_intrinsic(instr
));
73 case nir_instr_type_call
:
74 case nir_instr_type_jump
:
75 case nir_instr_type_ssa_undef
:
77 case nir_instr_type_parallel_copy
:
79 unreachable("Invalid instruction type");
86 #define HASH(hash, data) _mesa_fnv32_1a_accumulate((hash), (data))
89 hash_src(uint32_t hash
, const nir_src
*src
)
92 hash
= HASH(hash
, src
->ssa
);
97 hash_alu_src(uint32_t hash
, const nir_alu_src
*src
, unsigned num_components
)
99 hash
= HASH(hash
, src
->abs
);
100 hash
= HASH(hash
, src
->negate
);
102 for (unsigned i
= 0; i
< num_components
; i
++)
103 hash
= HASH(hash
, src
->swizzle
[i
]);
105 hash
= hash_src(hash
, &src
->src
);
110 hash_alu(uint32_t hash
, const nir_alu_instr
*instr
)
112 hash
= HASH(hash
, instr
->op
);
114 /* We explicitly don't hash instr->exact. */
115 uint8_t flags
= instr
->no_signed_wrap
|
116 instr
->no_unsigned_wrap
<< 1;
117 hash
= HASH(hash
, flags
);
119 hash
= HASH(hash
, instr
->dest
.dest
.ssa
.num_components
);
120 hash
= HASH(hash
, instr
->dest
.dest
.ssa
.bit_size
);
122 if (nir_op_infos
[instr
->op
].algebraic_properties
& NIR_OP_IS_2SRC_COMMUTATIVE
) {
123 assert(nir_op_infos
[instr
->op
].num_inputs
>= 2);
125 uint32_t hash0
= hash_alu_src(hash
, &instr
->src
[0],
126 nir_ssa_alu_instr_src_components(instr
, 0));
127 uint32_t hash1
= hash_alu_src(hash
, &instr
->src
[1],
128 nir_ssa_alu_instr_src_components(instr
, 1));
129 /* For commutative operations, we need some commutative way of
130 * combining the hashes. One option would be to XOR them but that
131 * means that anything with two identical sources will hash to 0 and
132 * that's common enough we probably don't want the guaranteed
133 * collision. Either addition or multiplication will also work.
135 hash
= hash0
* hash1
;
137 for (unsigned i
= 2; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
138 hash
= hash_alu_src(hash
, &instr
->src
[i
],
139 nir_ssa_alu_instr_src_components(instr
, i
));
142 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
143 hash
= hash_alu_src(hash
, &instr
->src
[i
],
144 nir_ssa_alu_instr_src_components(instr
, i
));
152 hash_deref(uint32_t hash
, const nir_deref_instr
*instr
)
154 hash
= HASH(hash
, instr
->deref_type
);
155 hash
= HASH(hash
, instr
->mode
);
156 hash
= HASH(hash
, instr
->type
);
158 if (instr
->deref_type
== nir_deref_type_var
)
159 return HASH(hash
, instr
->var
);
161 hash
= hash_src(hash
, &instr
->parent
);
163 switch (instr
->deref_type
) {
164 case nir_deref_type_struct
:
165 hash
= HASH(hash
, instr
->strct
.index
);
168 case nir_deref_type_array
:
169 case nir_deref_type_ptr_as_array
:
170 hash
= hash_src(hash
, &instr
->arr
.index
);
173 case nir_deref_type_cast
:
174 hash
= HASH(hash
, instr
->cast
.ptr_stride
);
177 case nir_deref_type_var
:
178 case nir_deref_type_array_wildcard
:
183 unreachable("Invalid instruction deref type");
190 hash_load_const(uint32_t hash
, const nir_load_const_instr
*instr
)
192 hash
= HASH(hash
, instr
->def
.num_components
);
194 if (instr
->def
.bit_size
== 1) {
195 for (unsigned i
= 0; i
< instr
->def
.num_components
; i
++) {
196 uint8_t b
= instr
->value
[i
].b
;
197 hash
= HASH(hash
, b
);
200 unsigned size
= instr
->def
.num_components
* sizeof(*instr
->value
);
201 hash
= _mesa_fnv32_1a_accumulate_block(hash
, instr
->value
, size
);
208 cmp_phi_src(const void *data1
, const void *data2
)
210 nir_phi_src
*src1
= *(nir_phi_src
**)data1
;
211 nir_phi_src
*src2
= *(nir_phi_src
**)data2
;
212 return src1
->pred
- src2
->pred
;
216 hash_phi(uint32_t hash
, const nir_phi_instr
*instr
)
218 hash
= HASH(hash
, instr
->instr
.block
);
220 /* sort sources by predecessor, since the order shouldn't matter */
221 unsigned num_preds
= instr
->instr
.block
->predecessors
->entries
;
222 NIR_VLA(nir_phi_src
*, srcs
, num_preds
);
224 nir_foreach_phi_src(src
, instr
) {
228 qsort(srcs
, num_preds
, sizeof(nir_phi_src
*), cmp_phi_src
);
230 for (i
= 0; i
< num_preds
; i
++) {
231 hash
= hash_src(hash
, &srcs
[i
]->src
);
232 hash
= HASH(hash
, srcs
[i
]->pred
);
239 hash_intrinsic(uint32_t hash
, const nir_intrinsic_instr
*instr
)
241 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[instr
->intrinsic
];
242 hash
= HASH(hash
, instr
->intrinsic
);
244 if (info
->has_dest
) {
245 hash
= HASH(hash
, instr
->dest
.ssa
.num_components
);
246 hash
= HASH(hash
, instr
->dest
.ssa
.bit_size
);
249 hash
= _mesa_fnv32_1a_accumulate_block(hash
, instr
->const_index
,
251 * sizeof(instr
->const_index
[0]));
256 hash_tex(uint32_t hash
, const nir_tex_instr
*instr
)
258 hash
= HASH(hash
, instr
->op
);
259 hash
= HASH(hash
, instr
->num_srcs
);
261 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
262 hash
= HASH(hash
, instr
->src
[i
].src_type
);
263 hash
= hash_src(hash
, &instr
->src
[i
].src
);
266 hash
= HASH(hash
, instr
->coord_components
);
267 hash
= HASH(hash
, instr
->sampler_dim
);
268 hash
= HASH(hash
, instr
->is_array
);
269 hash
= HASH(hash
, instr
->is_shadow
);
270 hash
= HASH(hash
, instr
->is_new_style_shadow
);
271 unsigned component
= instr
->component
;
272 hash
= HASH(hash
, component
);
273 for (unsigned i
= 0; i
< 4; ++i
)
274 for (unsigned j
= 0; j
< 2; ++j
)
275 hash
= HASH(hash
, instr
->tg4_offsets
[i
][j
]);
276 hash
= HASH(hash
, instr
->texture_index
);
277 hash
= HASH(hash
, instr
->texture_array_size
);
278 hash
= HASH(hash
, instr
->sampler_index
);
279 hash
= HASH(hash
, instr
->texture_non_uniform
);
280 hash
= HASH(hash
, instr
->sampler_non_uniform
);
285 /* Computes a hash of an instruction for use in a hash table. Note that this
286 * will only work for instructions where instr_can_rewrite() returns true, and
287 * it should return identical hashes for two instructions that are the same
288 * according nir_instrs_equal().
292 hash_instr(const void *data
)
294 const nir_instr
*instr
= data
;
295 uint32_t hash
= _mesa_fnv32_1a_offset_bias
;
297 switch (instr
->type
) {
298 case nir_instr_type_alu
:
299 hash
= hash_alu(hash
, nir_instr_as_alu(instr
));
301 case nir_instr_type_deref
:
302 hash
= hash_deref(hash
, nir_instr_as_deref(instr
));
304 case nir_instr_type_load_const
:
305 hash
= hash_load_const(hash
, nir_instr_as_load_const(instr
));
307 case nir_instr_type_phi
:
308 hash
= hash_phi(hash
, nir_instr_as_phi(instr
));
310 case nir_instr_type_intrinsic
:
311 hash
= hash_intrinsic(hash
, nir_instr_as_intrinsic(instr
));
313 case nir_instr_type_tex
:
314 hash
= hash_tex(hash
, nir_instr_as_tex(instr
));
317 unreachable("Invalid instruction type");
324 nir_srcs_equal(nir_src src1
, nir_src src2
)
328 return src1
.ssa
== src2
.ssa
;
336 if ((src1
.reg
.indirect
== NULL
) != (src2
.reg
.indirect
== NULL
))
339 if (src1
.reg
.indirect
) {
340 if (!nir_srcs_equal(*src1
.reg
.indirect
, *src2
.reg
.indirect
))
344 return src1
.reg
.reg
== src2
.reg
.reg
&&
345 src1
.reg
.base_offset
== src2
.reg
.base_offset
;
351 * If the \p s is an SSA value that was generated by a negation instruction,
352 * that instruction is returned as a \c nir_alu_instr. Otherwise \c NULL is
355 static nir_alu_instr
*
356 get_neg_instr(nir_src s
)
358 nir_alu_instr
*alu
= nir_src_as_alu_instr(s
);
360 return alu
!= NULL
&& (alu
->op
== nir_op_fneg
|| alu
->op
== nir_op_ineg
)
365 nir_const_value_negative_equal(nir_const_value c1
,
367 nir_alu_type full_type
)
369 assert(nir_alu_type_get_base_type(full_type
) != nir_type_invalid
);
370 assert(nir_alu_type_get_type_size(full_type
) != 0);
373 case nir_type_float16
:
374 return _mesa_half_to_float(c1
.u16
) == -_mesa_half_to_float(c2
.u16
);
376 case nir_type_float32
:
377 return c1
.f32
== -c2
.f32
;
379 case nir_type_float64
:
380 return c1
.f64
== -c2
.f64
;
384 return c1
.i8
== -c2
.i8
;
387 case nir_type_uint16
:
388 return c1
.i16
== -c2
.i16
;
391 case nir_type_uint32
:
392 return c1
.i32
== -c2
.i32
;
395 case nir_type_uint64
:
396 return c1
.i64
== -c2
.i64
;
406 * Shallow compare of ALU srcs to determine if one is the negation of the other
408 * This function detects cases where \p alu1 is a constant and \p alu2 is a
409 * constant that is its negation. It will also detect cases where \p alu2 is
410 * an SSA value that is a \c nir_op_fneg applied to \p alu1 (and vice versa).
412 * This function does not detect the general case when \p alu1 and \p alu2 are
413 * SSA values that are the negations of each other (e.g., \p alu1 represents
414 * (a * b) and \p alu2 represents (-a * b)).
417 * It is the responsibility of the caller to ensure that the component counts,
418 * write masks, and base types of the sources being compared are compatible.
421 nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
422 const nir_alu_instr
*alu2
,
423 unsigned src1
, unsigned src2
)
426 for (unsigned i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; i
++) {
427 assert(nir_alu_instr_channel_used(alu1
, src1
, i
) ==
428 nir_alu_instr_channel_used(alu2
, src2
, i
));
431 if (nir_op_infos
[alu1
->op
].input_types
[src1
] == nir_type_float
) {
432 assert(nir_op_infos
[alu1
->op
].input_types
[src1
] ==
433 nir_op_infos
[alu2
->op
].input_types
[src2
]);
435 assert(nir_op_infos
[alu1
->op
].input_types
[src1
] == nir_type_int
);
436 assert(nir_op_infos
[alu2
->op
].input_types
[src2
] == nir_type_int
);
440 if (alu1
->src
[src1
].abs
!= alu2
->src
[src2
].abs
)
443 bool parity
= alu1
->src
[src1
].negate
!= alu2
->src
[src2
].negate
;
445 /* Handling load_const instructions is tricky. */
447 const nir_const_value
*const const1
=
448 nir_src_as_const_value(alu1
->src
[src1
].src
);
450 if (const1
!= NULL
) {
451 /* Assume that constant folding will eliminate source mods and unary
457 const nir_const_value
*const const2
=
458 nir_src_as_const_value(alu2
->src
[src2
].src
);
463 if (nir_src_bit_size(alu1
->src
[src1
].src
) !=
464 nir_src_bit_size(alu2
->src
[src2
].src
))
467 const nir_alu_type full_type
= nir_op_infos
[alu1
->op
].input_types
[src1
] |
468 nir_src_bit_size(alu1
->src
[src1
].src
);
469 for (unsigned i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; i
++) {
470 if (nir_alu_instr_channel_used(alu1
, src1
, i
) &&
471 !nir_const_value_negative_equal(const1
[alu1
->src
[src1
].swizzle
[i
]],
472 const2
[alu2
->src
[src2
].swizzle
[i
]],
480 uint8_t alu1_swizzle
[4] = {0};
481 nir_src alu1_actual_src
;
482 nir_alu_instr
*neg1
= get_neg_instr(alu1
->src
[src1
].src
);
486 alu1_actual_src
= neg1
->src
[0].src
;
488 for (unsigned i
= 0; i
< nir_ssa_alu_instr_src_components(neg1
, 0); i
++)
489 alu1_swizzle
[i
] = neg1
->src
[0].swizzle
[i
];
491 alu1_actual_src
= alu1
->src
[src1
].src
;
493 for (unsigned i
= 0; i
< nir_ssa_alu_instr_src_components(alu1
, src1
); i
++)
497 uint8_t alu2_swizzle
[4] = {0};
498 nir_src alu2_actual_src
;
499 nir_alu_instr
*neg2
= get_neg_instr(alu2
->src
[src2
].src
);
503 alu2_actual_src
= neg2
->src
[0].src
;
505 for (unsigned i
= 0; i
< nir_ssa_alu_instr_src_components(neg2
, 0); i
++)
506 alu2_swizzle
[i
] = neg2
->src
[0].swizzle
[i
];
508 alu2_actual_src
= alu2
->src
[src2
].src
;
510 for (unsigned i
= 0; i
< nir_ssa_alu_instr_src_components(alu2
, src2
); i
++)
514 for (unsigned i
= 0; i
< nir_ssa_alu_instr_src_components(alu1
, src1
); i
++) {
515 if (alu1_swizzle
[alu1
->src
[src1
].swizzle
[i
]] !=
516 alu2_swizzle
[alu2
->src
[src2
].swizzle
[i
]])
520 return parity
&& nir_srcs_equal(alu1_actual_src
, alu2_actual_src
);
524 nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
525 unsigned src1
, unsigned src2
)
527 if (alu1
->src
[src1
].abs
!= alu2
->src
[src2
].abs
||
528 alu1
->src
[src1
].negate
!= alu2
->src
[src2
].negate
)
531 for (unsigned i
= 0; i
< nir_ssa_alu_instr_src_components(alu1
, src1
); i
++) {
532 if (alu1
->src
[src1
].swizzle
[i
] != alu2
->src
[src2
].swizzle
[i
])
536 return nir_srcs_equal(alu1
->src
[src1
].src
, alu2
->src
[src2
].src
);
539 /* Returns "true" if two instructions are equal. Note that this will only
540 * work for the subset of instructions defined by instr_can_rewrite(). Also,
541 * it should only return "true" for instructions that hash_instr() will return
542 * the same hash for (ignoring collisions, of course).
546 nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
)
548 assert(instr_can_rewrite(instr1
) && instr_can_rewrite(instr2
));
550 if (instr1
->type
!= instr2
->type
)
553 switch (instr1
->type
) {
554 case nir_instr_type_alu
: {
555 nir_alu_instr
*alu1
= nir_instr_as_alu(instr1
);
556 nir_alu_instr
*alu2
= nir_instr_as_alu(instr2
);
558 if (alu1
->op
!= alu2
->op
)
561 /* We explicitly don't compare instr->exact. */
563 if (alu1
->no_signed_wrap
!= alu2
->no_signed_wrap
)
566 if (alu1
->no_unsigned_wrap
!= alu2
->no_unsigned_wrap
)
569 /* TODO: We can probably acutally do something more inteligent such
570 * as allowing different numbers and taking a maximum or something
572 if (alu1
->dest
.dest
.ssa
.num_components
!= alu2
->dest
.dest
.ssa
.num_components
)
575 if (alu1
->dest
.dest
.ssa
.bit_size
!= alu2
->dest
.dest
.ssa
.bit_size
)
578 if (nir_op_infos
[alu1
->op
].algebraic_properties
& NIR_OP_IS_2SRC_COMMUTATIVE
) {
579 if ((!nir_alu_srcs_equal(alu1
, alu2
, 0, 0) ||
580 !nir_alu_srcs_equal(alu1
, alu2
, 1, 1)) &&
581 (!nir_alu_srcs_equal(alu1
, alu2
, 0, 1) ||
582 !nir_alu_srcs_equal(alu1
, alu2
, 1, 0)))
585 for (unsigned i
= 2; i
< nir_op_infos
[alu1
->op
].num_inputs
; i
++) {
586 if (!nir_alu_srcs_equal(alu1
, alu2
, i
, i
))
590 for (unsigned i
= 0; i
< nir_op_infos
[alu1
->op
].num_inputs
; i
++) {
591 if (!nir_alu_srcs_equal(alu1
, alu2
, i
, i
))
597 case nir_instr_type_deref
: {
598 nir_deref_instr
*deref1
= nir_instr_as_deref(instr1
);
599 nir_deref_instr
*deref2
= nir_instr_as_deref(instr2
);
601 if (deref1
->deref_type
!= deref2
->deref_type
||
602 deref1
->mode
!= deref2
->mode
||
603 deref1
->type
!= deref2
->type
)
606 if (deref1
->deref_type
== nir_deref_type_var
)
607 return deref1
->var
== deref2
->var
;
609 if (!nir_srcs_equal(deref1
->parent
, deref2
->parent
))
612 switch (deref1
->deref_type
) {
613 case nir_deref_type_struct
:
614 if (deref1
->strct
.index
!= deref2
->strct
.index
)
618 case nir_deref_type_array
:
619 case nir_deref_type_ptr_as_array
:
620 if (!nir_srcs_equal(deref1
->arr
.index
, deref2
->arr
.index
))
624 case nir_deref_type_cast
:
625 if (deref1
->cast
.ptr_stride
!= deref2
->cast
.ptr_stride
)
629 case nir_deref_type_var
:
630 case nir_deref_type_array_wildcard
:
635 unreachable("Invalid instruction deref type");
639 case nir_instr_type_tex
: {
640 nir_tex_instr
*tex1
= nir_instr_as_tex(instr1
);
641 nir_tex_instr
*tex2
= nir_instr_as_tex(instr2
);
643 if (tex1
->op
!= tex2
->op
)
646 if (tex1
->num_srcs
!= tex2
->num_srcs
)
648 for (unsigned i
= 0; i
< tex1
->num_srcs
; i
++) {
649 if (tex1
->src
[i
].src_type
!= tex2
->src
[i
].src_type
||
650 !nir_srcs_equal(tex1
->src
[i
].src
, tex2
->src
[i
].src
)) {
655 if (tex1
->coord_components
!= tex2
->coord_components
||
656 tex1
->sampler_dim
!= tex2
->sampler_dim
||
657 tex1
->is_array
!= tex2
->is_array
||
658 tex1
->is_shadow
!= tex2
->is_shadow
||
659 tex1
->is_new_style_shadow
!= tex2
->is_new_style_shadow
||
660 tex1
->component
!= tex2
->component
||
661 tex1
->texture_index
!= tex2
->texture_index
||
662 tex1
->texture_array_size
!= tex2
->texture_array_size
||
663 tex1
->sampler_index
!= tex2
->sampler_index
) {
667 if (memcmp(tex1
->tg4_offsets
, tex2
->tg4_offsets
,
668 sizeof(tex1
->tg4_offsets
)))
673 case nir_instr_type_load_const
: {
674 nir_load_const_instr
*load1
= nir_instr_as_load_const(instr1
);
675 nir_load_const_instr
*load2
= nir_instr_as_load_const(instr2
);
677 if (load1
->def
.num_components
!= load2
->def
.num_components
)
680 if (load1
->def
.bit_size
!= load2
->def
.bit_size
)
683 if (load1
->def
.bit_size
== 1) {
684 for (unsigned i
= 0; i
< load1
->def
.num_components
; ++i
) {
685 if (load1
->value
[i
].b
!= load2
->value
[i
].b
)
689 unsigned size
= load1
->def
.num_components
* sizeof(*load1
->value
);
690 if (memcmp(load1
->value
, load2
->value
, size
) != 0)
695 case nir_instr_type_phi
: {
696 nir_phi_instr
*phi1
= nir_instr_as_phi(instr1
);
697 nir_phi_instr
*phi2
= nir_instr_as_phi(instr2
);
699 if (phi1
->instr
.block
!= phi2
->instr
.block
)
702 nir_foreach_phi_src(src1
, phi1
) {
703 nir_foreach_phi_src(src2
, phi2
) {
704 if (src1
->pred
== src2
->pred
) {
705 if (!nir_srcs_equal(src1
->src
, src2
->src
))
715 case nir_instr_type_intrinsic
: {
716 nir_intrinsic_instr
*intrinsic1
= nir_instr_as_intrinsic(instr1
);
717 nir_intrinsic_instr
*intrinsic2
= nir_instr_as_intrinsic(instr2
);
718 const nir_intrinsic_info
*info
=
719 &nir_intrinsic_infos
[intrinsic1
->intrinsic
];
721 if (intrinsic1
->intrinsic
!= intrinsic2
->intrinsic
||
722 intrinsic1
->num_components
!= intrinsic2
->num_components
)
725 if (info
->has_dest
&& intrinsic1
->dest
.ssa
.num_components
!=
726 intrinsic2
->dest
.ssa
.num_components
)
729 if (info
->has_dest
&& intrinsic1
->dest
.ssa
.bit_size
!=
730 intrinsic2
->dest
.ssa
.bit_size
)
733 for (unsigned i
= 0; i
< info
->num_srcs
; i
++) {
734 if (!nir_srcs_equal(intrinsic1
->src
[i
], intrinsic2
->src
[i
]))
738 for (unsigned i
= 0; i
< info
->num_indices
; i
++) {
739 if (intrinsic1
->const_index
[i
] != intrinsic2
->const_index
[i
])
745 case nir_instr_type_call
:
746 case nir_instr_type_jump
:
747 case nir_instr_type_ssa_undef
:
748 case nir_instr_type_parallel_copy
:
750 unreachable("Invalid instruction type");
753 unreachable("All cases in the above switch should return");
757 nir_instr_get_dest_ssa_def(nir_instr
*instr
)
759 switch (instr
->type
) {
760 case nir_instr_type_alu
:
761 assert(nir_instr_as_alu(instr
)->dest
.dest
.is_ssa
);
762 return &nir_instr_as_alu(instr
)->dest
.dest
.ssa
;
763 case nir_instr_type_deref
:
764 assert(nir_instr_as_deref(instr
)->dest
.is_ssa
);
765 return &nir_instr_as_deref(instr
)->dest
.ssa
;
766 case nir_instr_type_load_const
:
767 return &nir_instr_as_load_const(instr
)->def
;
768 case nir_instr_type_phi
:
769 assert(nir_instr_as_phi(instr
)->dest
.is_ssa
);
770 return &nir_instr_as_phi(instr
)->dest
.ssa
;
771 case nir_instr_type_intrinsic
:
772 assert(nir_instr_as_intrinsic(instr
)->dest
.is_ssa
);
773 return &nir_instr_as_intrinsic(instr
)->dest
.ssa
;
774 case nir_instr_type_tex
:
775 assert(nir_instr_as_tex(instr
)->dest
.is_ssa
);
776 return &nir_instr_as_tex(instr
)->dest
.ssa
;
778 unreachable("We never ask for any of these");
783 cmp_func(const void *data1
, const void *data2
)
785 return nir_instrs_equal(data1
, data2
);
789 nir_instr_set_create(void *mem_ctx
)
791 return _mesa_set_create(mem_ctx
, hash_instr
, cmp_func
);
795 nir_instr_set_destroy(struct set
*instr_set
)
797 _mesa_set_destroy(instr_set
, NULL
);
801 nir_instr_set_add_or_rewrite(struct set
*instr_set
, nir_instr
*instr
)
803 if (!instr_can_rewrite(instr
))
806 struct set_entry
*e
= _mesa_set_search_or_add(instr_set
, instr
);
807 nir_instr
*match
= (nir_instr
*) e
->key
;
808 if (match
!= instr
) {
809 nir_ssa_def
*def
= nir_instr_get_dest_ssa_def(instr
);
810 nir_ssa_def
*new_def
= nir_instr_get_dest_ssa_def(match
);
812 /* It's safe to replace an exact instruction with an inexact one as
813 * long as we make it exact. If we got here, the two instructions are
814 * exactly identical in every other way so, once we've set the exact
815 * bit, they are the same.
817 if (instr
->type
== nir_instr_type_alu
&& nir_instr_as_alu(instr
)->exact
)
818 nir_instr_as_alu(match
)->exact
= true;
820 nir_ssa_def_rewrite_uses(def
, nir_src_for_ssa(new_def
));
828 nir_instr_set_remove(struct set
*instr_set
, nir_instr
*instr
)
830 if (!instr_can_rewrite(instr
))
833 struct set_entry
*entry
= _mesa_set_search(instr_set
, instr
);
835 _mesa_set_remove(instr_set
, entry
);