nir: Add a ptr_as_array deref type
[mesa.git] / src / compiler / nir / nir_instr_set.c
1 /*
2 * Copyright © 2014 Connor Abbott
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 */
23
24 #include "nir_instr_set.h"
25 #include "nir_vla.h"
26
27 #define HASH(hash, data) _mesa_fnv32_1a_accumulate((hash), (data))
28
29 static uint32_t
30 hash_src(uint32_t hash, const nir_src *src)
31 {
32 assert(src->is_ssa);
33 hash = HASH(hash, src->ssa);
34 return hash;
35 }
36
37 static uint32_t
38 hash_alu_src(uint32_t hash, const nir_alu_src *src, unsigned num_components)
39 {
40 hash = HASH(hash, src->abs);
41 hash = HASH(hash, src->negate);
42
43 for (unsigned i = 0; i < num_components; i++)
44 hash = HASH(hash, src->swizzle[i]);
45
46 hash = hash_src(hash, &src->src);
47 return hash;
48 }
49
50 static uint32_t
51 hash_alu(uint32_t hash, const nir_alu_instr *instr)
52 {
53 hash = HASH(hash, instr->op);
54 hash = HASH(hash, instr->dest.dest.ssa.num_components);
55 hash = HASH(hash, instr->dest.dest.ssa.bit_size);
56 /* We explicitly don't hash instr->dest.dest.exact */
57
58 if (nir_op_infos[instr->op].algebraic_properties & NIR_OP_IS_COMMUTATIVE) {
59 assert(nir_op_infos[instr->op].num_inputs == 2);
60 uint32_t hash0 = hash_alu_src(hash, &instr->src[0],
61 nir_ssa_alu_instr_src_components(instr, 0));
62 uint32_t hash1 = hash_alu_src(hash, &instr->src[1],
63 nir_ssa_alu_instr_src_components(instr, 1));
64 /* For commutative operations, we need some commutative way of
65 * combining the hashes. One option would be to XOR them but that
66 * means that anything with two identical sources will hash to 0 and
67 * that's common enough we probably don't want the guaranteed
68 * collision. Either addition or multiplication will also work.
69 */
70 hash = hash0 * hash1;
71 } else {
72 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
73 hash = hash_alu_src(hash, &instr->src[i],
74 nir_ssa_alu_instr_src_components(instr, i));
75 }
76 }
77
78 return hash;
79 }
80
81 static uint32_t
82 hash_deref(uint32_t hash, const nir_deref_instr *instr)
83 {
84 hash = HASH(hash, instr->deref_type);
85 hash = HASH(hash, instr->mode);
86 hash = HASH(hash, instr->type);
87
88 if (instr->deref_type == nir_deref_type_var)
89 return HASH(hash, instr->var);
90
91 hash = hash_src(hash, &instr->parent);
92
93 switch (instr->deref_type) {
94 case nir_deref_type_struct:
95 hash = HASH(hash, instr->strct.index);
96 break;
97
98 case nir_deref_type_array:
99 case nir_deref_type_ptr_as_array:
100 hash = hash_src(hash, &instr->arr.index);
101 break;
102
103 case nir_deref_type_cast:
104 hash = HASH(hash, instr->cast.ptr_stride);
105 break;
106
107 case nir_deref_type_var:
108 case nir_deref_type_array_wildcard:
109 /* Nothing to do */
110 break;
111
112 default:
113 unreachable("Invalid instruction deref type");
114 }
115
116 return hash;
117 }
118
119 static uint32_t
120 hash_load_const(uint32_t hash, const nir_load_const_instr *instr)
121 {
122 hash = HASH(hash, instr->def.num_components);
123
124 if (instr->def.bit_size == 1) {
125 for (unsigned i = 0; i < instr->def.num_components; i++) {
126 uint8_t b = instr->value.b[i];
127 hash = HASH(hash, b);
128 }
129 } else {
130 unsigned size = instr->def.num_components * (instr->def.bit_size / 8);
131 hash = _mesa_fnv32_1a_accumulate_block(hash, instr->value.f32, size);
132 }
133
134 return hash;
135 }
136
137 static int
138 cmp_phi_src(const void *data1, const void *data2)
139 {
140 nir_phi_src *src1 = *(nir_phi_src **)data1;
141 nir_phi_src *src2 = *(nir_phi_src **)data2;
142 return src1->pred - src2->pred;
143 }
144
145 static uint32_t
146 hash_phi(uint32_t hash, const nir_phi_instr *instr)
147 {
148 hash = HASH(hash, instr->instr.block);
149
150 /* sort sources by predecessor, since the order shouldn't matter */
151 unsigned num_preds = instr->instr.block->predecessors->entries;
152 NIR_VLA(nir_phi_src *, srcs, num_preds);
153 unsigned i = 0;
154 nir_foreach_phi_src(src, instr) {
155 srcs[i++] = src;
156 }
157
158 qsort(srcs, num_preds, sizeof(nir_phi_src *), cmp_phi_src);
159
160 for (i = 0; i < num_preds; i++) {
161 hash = hash_src(hash, &srcs[i]->src);
162 hash = HASH(hash, srcs[i]->pred);
163 }
164
165 return hash;
166 }
167
168 static uint32_t
169 hash_intrinsic(uint32_t hash, const nir_intrinsic_instr *instr)
170 {
171 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic];
172 hash = HASH(hash, instr->intrinsic);
173
174 if (info->has_dest) {
175 hash = HASH(hash, instr->dest.ssa.num_components);
176 hash = HASH(hash, instr->dest.ssa.bit_size);
177 }
178
179 hash = _mesa_fnv32_1a_accumulate_block(hash, instr->const_index,
180 info->num_indices
181 * sizeof(instr->const_index[0]));
182 return hash;
183 }
184
185 static uint32_t
186 hash_tex(uint32_t hash, const nir_tex_instr *instr)
187 {
188 hash = HASH(hash, instr->op);
189 hash = HASH(hash, instr->num_srcs);
190
191 for (unsigned i = 0; i < instr->num_srcs; i++) {
192 hash = HASH(hash, instr->src[i].src_type);
193 hash = hash_src(hash, &instr->src[i].src);
194 }
195
196 hash = HASH(hash, instr->coord_components);
197 hash = HASH(hash, instr->sampler_dim);
198 hash = HASH(hash, instr->is_array);
199 hash = HASH(hash, instr->is_shadow);
200 hash = HASH(hash, instr->is_new_style_shadow);
201 unsigned component = instr->component;
202 hash = HASH(hash, component);
203 hash = HASH(hash, instr->texture_index);
204 hash = HASH(hash, instr->texture_array_size);
205 hash = HASH(hash, instr->sampler_index);
206
207 return hash;
208 }
209
210 /* Computes a hash of an instruction for use in a hash table. Note that this
211 * will only work for instructions where instr_can_rewrite() returns true, and
212 * it should return identical hashes for two instructions that are the same
213 * according nir_instrs_equal().
214 */
215
216 static uint32_t
217 hash_instr(const void *data)
218 {
219 const nir_instr *instr = data;
220 uint32_t hash = _mesa_fnv32_1a_offset_bias;
221
222 switch (instr->type) {
223 case nir_instr_type_alu:
224 hash = hash_alu(hash, nir_instr_as_alu(instr));
225 break;
226 case nir_instr_type_deref:
227 hash = hash_deref(hash, nir_instr_as_deref(instr));
228 break;
229 case nir_instr_type_load_const:
230 hash = hash_load_const(hash, nir_instr_as_load_const(instr));
231 break;
232 case nir_instr_type_phi:
233 hash = hash_phi(hash, nir_instr_as_phi(instr));
234 break;
235 case nir_instr_type_intrinsic:
236 hash = hash_intrinsic(hash, nir_instr_as_intrinsic(instr));
237 break;
238 case nir_instr_type_tex:
239 hash = hash_tex(hash, nir_instr_as_tex(instr));
240 break;
241 default:
242 unreachable("Invalid instruction type");
243 }
244
245 return hash;
246 }
247
248 bool
249 nir_srcs_equal(nir_src src1, nir_src src2)
250 {
251 if (src1.is_ssa) {
252 if (src2.is_ssa) {
253 return src1.ssa == src2.ssa;
254 } else {
255 return false;
256 }
257 } else {
258 if (src2.is_ssa) {
259 return false;
260 } else {
261 if ((src1.reg.indirect == NULL) != (src2.reg.indirect == NULL))
262 return false;
263
264 if (src1.reg.indirect) {
265 if (!nir_srcs_equal(*src1.reg.indirect, *src2.reg.indirect))
266 return false;
267 }
268
269 return src1.reg.reg == src2.reg.reg &&
270 src1.reg.base_offset == src2.reg.base_offset;
271 }
272 }
273 }
274
275 bool
276 nir_alu_srcs_equal(const nir_alu_instr *alu1, const nir_alu_instr *alu2,
277 unsigned src1, unsigned src2)
278 {
279 if (alu1->src[src1].abs != alu2->src[src2].abs ||
280 alu1->src[src1].negate != alu2->src[src2].negate)
281 return false;
282
283 for (unsigned i = 0; i < nir_ssa_alu_instr_src_components(alu1, src1); i++) {
284 if (alu1->src[src1].swizzle[i] != alu2->src[src2].swizzle[i])
285 return false;
286 }
287
288 return nir_srcs_equal(alu1->src[src1].src, alu2->src[src2].src);
289 }
290
291 /* Returns "true" if two instructions are equal. Note that this will only
292 * work for the subset of instructions defined by instr_can_rewrite(). Also,
293 * it should only return "true" for instructions that hash_instr() will return
294 * the same hash for (ignoring collisions, of course).
295 */
296
297 static bool
298 nir_instrs_equal(const nir_instr *instr1, const nir_instr *instr2)
299 {
300 if (instr1->type != instr2->type)
301 return false;
302
303 switch (instr1->type) {
304 case nir_instr_type_alu: {
305 nir_alu_instr *alu1 = nir_instr_as_alu(instr1);
306 nir_alu_instr *alu2 = nir_instr_as_alu(instr2);
307
308 if (alu1->op != alu2->op)
309 return false;
310
311 /* TODO: We can probably acutally do something more inteligent such
312 * as allowing different numbers and taking a maximum or something
313 * here */
314 if (alu1->dest.dest.ssa.num_components != alu2->dest.dest.ssa.num_components)
315 return false;
316
317 if (alu1->dest.dest.ssa.bit_size != alu2->dest.dest.ssa.bit_size)
318 return false;
319
320 /* We explicitly don't hash instr->dest.dest.exact */
321
322 if (nir_op_infos[alu1->op].algebraic_properties & NIR_OP_IS_COMMUTATIVE) {
323 assert(nir_op_infos[alu1->op].num_inputs == 2);
324 return (nir_alu_srcs_equal(alu1, alu2, 0, 0) &&
325 nir_alu_srcs_equal(alu1, alu2, 1, 1)) ||
326 (nir_alu_srcs_equal(alu1, alu2, 0, 1) &&
327 nir_alu_srcs_equal(alu1, alu2, 1, 0));
328 } else {
329 for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) {
330 if (!nir_alu_srcs_equal(alu1, alu2, i, i))
331 return false;
332 }
333 }
334 return true;
335 }
336 case nir_instr_type_deref: {
337 nir_deref_instr *deref1 = nir_instr_as_deref(instr1);
338 nir_deref_instr *deref2 = nir_instr_as_deref(instr2);
339
340 if (deref1->deref_type != deref2->deref_type ||
341 deref1->mode != deref2->mode ||
342 deref1->type != deref2->type)
343 return false;
344
345 if (deref1->deref_type == nir_deref_type_var)
346 return deref1->var == deref2->var;
347
348 if (!nir_srcs_equal(deref1->parent, deref2->parent))
349 return false;
350
351 switch (deref1->deref_type) {
352 case nir_deref_type_struct:
353 if (deref1->strct.index != deref2->strct.index)
354 return false;
355 break;
356
357 case nir_deref_type_array:
358 case nir_deref_type_ptr_as_array:
359 if (!nir_srcs_equal(deref1->arr.index, deref2->arr.index))
360 return false;
361 break;
362
363 case nir_deref_type_cast:
364 if (deref1->cast.ptr_stride != deref2->cast.ptr_stride)
365 return false;
366 break;
367
368 case nir_deref_type_var:
369 case nir_deref_type_array_wildcard:
370 /* Nothing to do */
371 break;
372
373 default:
374 unreachable("Invalid instruction deref type");
375 }
376 return true;
377 }
378 case nir_instr_type_tex: {
379 nir_tex_instr *tex1 = nir_instr_as_tex(instr1);
380 nir_tex_instr *tex2 = nir_instr_as_tex(instr2);
381
382 if (tex1->op != tex2->op)
383 return false;
384
385 if (tex1->num_srcs != tex2->num_srcs)
386 return false;
387 for (unsigned i = 0; i < tex1->num_srcs; i++) {
388 if (tex1->src[i].src_type != tex2->src[i].src_type ||
389 !nir_srcs_equal(tex1->src[i].src, tex2->src[i].src)) {
390 return false;
391 }
392 }
393
394 if (tex1->coord_components != tex2->coord_components ||
395 tex1->sampler_dim != tex2->sampler_dim ||
396 tex1->is_array != tex2->is_array ||
397 tex1->is_shadow != tex2->is_shadow ||
398 tex1->is_new_style_shadow != tex2->is_new_style_shadow ||
399 tex1->component != tex2->component ||
400 tex1->texture_index != tex2->texture_index ||
401 tex1->texture_array_size != tex2->texture_array_size ||
402 tex1->sampler_index != tex2->sampler_index) {
403 return false;
404 }
405
406 return true;
407 }
408 case nir_instr_type_load_const: {
409 nir_load_const_instr *load1 = nir_instr_as_load_const(instr1);
410 nir_load_const_instr *load2 = nir_instr_as_load_const(instr2);
411
412 if (load1->def.num_components != load2->def.num_components)
413 return false;
414
415 if (load1->def.bit_size != load2->def.bit_size)
416 return false;
417
418 if (load1->def.bit_size == 1) {
419 unsigned size = load1->def.num_components * sizeof(bool);
420 return memcmp(load1->value.b, load2->value.b, size) == 0;
421 } else {
422 unsigned size = load1->def.num_components * (load1->def.bit_size / 8);
423 return memcmp(load1->value.f32, load2->value.f32, size) == 0;
424 }
425 }
426 case nir_instr_type_phi: {
427 nir_phi_instr *phi1 = nir_instr_as_phi(instr1);
428 nir_phi_instr *phi2 = nir_instr_as_phi(instr2);
429
430 if (phi1->instr.block != phi2->instr.block)
431 return false;
432
433 nir_foreach_phi_src(src1, phi1) {
434 nir_foreach_phi_src(src2, phi2) {
435 if (src1->pred == src2->pred) {
436 if (!nir_srcs_equal(src1->src, src2->src))
437 return false;
438
439 break;
440 }
441 }
442 }
443
444 return true;
445 }
446 case nir_instr_type_intrinsic: {
447 nir_intrinsic_instr *intrinsic1 = nir_instr_as_intrinsic(instr1);
448 nir_intrinsic_instr *intrinsic2 = nir_instr_as_intrinsic(instr2);
449 const nir_intrinsic_info *info =
450 &nir_intrinsic_infos[intrinsic1->intrinsic];
451
452 if (intrinsic1->intrinsic != intrinsic2->intrinsic ||
453 intrinsic1->num_components != intrinsic2->num_components)
454 return false;
455
456 if (info->has_dest && intrinsic1->dest.ssa.num_components !=
457 intrinsic2->dest.ssa.num_components)
458 return false;
459
460 if (info->has_dest && intrinsic1->dest.ssa.bit_size !=
461 intrinsic2->dest.ssa.bit_size)
462 return false;
463
464 for (unsigned i = 0; i < info->num_srcs; i++) {
465 if (!nir_srcs_equal(intrinsic1->src[i], intrinsic2->src[i]))
466 return false;
467 }
468
469 for (unsigned i = 0; i < info->num_indices; i++) {
470 if (intrinsic1->const_index[i] != intrinsic2->const_index[i])
471 return false;
472 }
473
474 return true;
475 }
476 case nir_instr_type_call:
477 case nir_instr_type_jump:
478 case nir_instr_type_ssa_undef:
479 case nir_instr_type_parallel_copy:
480 default:
481 unreachable("Invalid instruction type");
482 }
483
484 unreachable("All cases in the above switch should return");
485 }
486
487 static bool
488 src_is_ssa(nir_src *src, void *data)
489 {
490 (void) data;
491 return src->is_ssa;
492 }
493
494 static bool
495 dest_is_ssa(nir_dest *dest, void *data)
496 {
497 (void) data;
498 return dest->is_ssa;
499 }
500
501 /* This function determines if uses of an instruction can safely be rewritten
502 * to use another identical instruction instead. Note that this function must
503 * be kept in sync with hash_instr() and nir_instrs_equal() -- only
504 * instructions that pass this test will be handed on to those functions, and
505 * conversely they must handle everything that this function returns true for.
506 */
507
508 static bool
509 instr_can_rewrite(nir_instr *instr)
510 {
511 /* We only handle SSA. */
512 if (!nir_foreach_dest(instr, dest_is_ssa, NULL) ||
513 !nir_foreach_src(instr, src_is_ssa, NULL))
514 return false;
515
516 switch (instr->type) {
517 case nir_instr_type_alu:
518 case nir_instr_type_deref:
519 case nir_instr_type_tex:
520 case nir_instr_type_load_const:
521 case nir_instr_type_phi:
522 return true;
523 case nir_instr_type_intrinsic: {
524 const nir_intrinsic_info *info =
525 &nir_intrinsic_infos[nir_instr_as_intrinsic(instr)->intrinsic];
526 return (info->flags & NIR_INTRINSIC_CAN_ELIMINATE) &&
527 (info->flags & NIR_INTRINSIC_CAN_REORDER);
528 }
529 case nir_instr_type_call:
530 case nir_instr_type_jump:
531 case nir_instr_type_ssa_undef:
532 return false;
533 case nir_instr_type_parallel_copy:
534 default:
535 unreachable("Invalid instruction type");
536 }
537
538 return false;
539 }
540
541 static nir_ssa_def *
542 nir_instr_get_dest_ssa_def(nir_instr *instr)
543 {
544 switch (instr->type) {
545 case nir_instr_type_alu:
546 assert(nir_instr_as_alu(instr)->dest.dest.is_ssa);
547 return &nir_instr_as_alu(instr)->dest.dest.ssa;
548 case nir_instr_type_deref:
549 assert(nir_instr_as_deref(instr)->dest.is_ssa);
550 return &nir_instr_as_deref(instr)->dest.ssa;
551 case nir_instr_type_load_const:
552 return &nir_instr_as_load_const(instr)->def;
553 case nir_instr_type_phi:
554 assert(nir_instr_as_phi(instr)->dest.is_ssa);
555 return &nir_instr_as_phi(instr)->dest.ssa;
556 case nir_instr_type_intrinsic:
557 assert(nir_instr_as_intrinsic(instr)->dest.is_ssa);
558 return &nir_instr_as_intrinsic(instr)->dest.ssa;
559 case nir_instr_type_tex:
560 assert(nir_instr_as_tex(instr)->dest.is_ssa);
561 return &nir_instr_as_tex(instr)->dest.ssa;
562 default:
563 unreachable("We never ask for any of these");
564 }
565 }
566
567 static bool
568 cmp_func(const void *data1, const void *data2)
569 {
570 return nir_instrs_equal(data1, data2);
571 }
572
573 struct set *
574 nir_instr_set_create(void *mem_ctx)
575 {
576 return _mesa_set_create(mem_ctx, hash_instr, cmp_func);
577 }
578
579 void
580 nir_instr_set_destroy(struct set *instr_set)
581 {
582 _mesa_set_destroy(instr_set, NULL);
583 }
584
585 bool
586 nir_instr_set_add_or_rewrite(struct set *instr_set, nir_instr *instr)
587 {
588 if (!instr_can_rewrite(instr))
589 return false;
590
591 struct set_entry *entry = _mesa_set_search(instr_set, instr);
592 if (entry) {
593 nir_ssa_def *def = nir_instr_get_dest_ssa_def(instr);
594 nir_instr *match = (nir_instr *) entry->key;
595 nir_ssa_def *new_def = nir_instr_get_dest_ssa_def(match);
596
597 /* It's safe to replace an exact instruction with an inexact one as
598 * long as we make it exact. If we got here, the two instructions are
599 * exactly identical in every other way so, once we've set the exact
600 * bit, they are the same.
601 */
602 if (instr->type == nir_instr_type_alu && nir_instr_as_alu(instr)->exact)
603 nir_instr_as_alu(match)->exact = true;
604
605 nir_ssa_def_rewrite_uses(def, nir_src_for_ssa(new_def));
606 return true;
607 }
608
609 _mesa_set_add(instr_set, instr);
610 return false;
611 }
612
613 void
614 nir_instr_set_remove(struct set *instr_set, nir_instr *instr)
615 {
616 if (!instr_can_rewrite(instr))
617 return;
618
619 struct set_entry *entry = _mesa_set_search(instr_set, instr);
620 if (entry)
621 _mesa_set_remove(instr_set, entry);
622 }
623