1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
|
/*
* Copyright © 2016 Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#include "nir.h"
#include "nir/nir_builder.h"
#include "nir_constant_expressions.h"
#include "nir_control_flow.h"
#include "nir_loop_analyze.h"
/**
* Gets the single block that jumps back to the loop header. Already assumes
* there is exactly one such block.
*/
static nir_block*
find_continue_block(nir_loop *loop)
{
nir_block *header_block = nir_loop_first_block(loop);
nir_block *prev_block =
nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
assert(header_block->predecessors->entries == 2);
set_foreach(header_block->predecessors, pred_entry) {
if (pred_entry->key != prev_block)
return (nir_block*)pred_entry->key;
}
unreachable("Continue block not found!");
}
/**
* This optimization detects if statements at the tops of loops where the
* condition is a phi node of two constants and moves half of the if to above
* the loop and the other half of the if to the end of the loop. A simple for
* loop "for (int i = 0; i < 4; i++)", when run through the SPIR-V front-end,
* ends up looking something like this:
*
* vec1 32 ssa_0 = load_const (0x00000000)
* vec1 32 ssa_1 = load_const (0xffffffff)
* loop {
* block block_1:
* vec1 32 ssa_2 = phi block_0: ssa_0, block_7: ssa_5
* vec1 32 ssa_3 = phi block_0: ssa_0, block_7: ssa_1
* if ssa_2 {
* block block_2:
* vec1 32 ssa_4 = load_const (0x00000001)
* vec1 32 ssa_5 = iadd ssa_2, ssa_4
* } else {
* block block_3:
* }
* block block_4:
* vec1 32 ssa_6 = load_const (0x00000004)
* vec1 32 ssa_7 = ilt ssa_5, ssa_6
* if ssa_7 {
* block block_5:
* } else {
* block block_6:
* break
* }
* block block_7:
* }
*
* This turns it into something like this:
*
* // Stuff from block 1
* // Stuff from block 3
* loop {
* block block_1:
* vec1 32 ssa_3 = phi block_0: ssa_0, block_7: ssa_1
* vec1 32 ssa_6 = load_const (0x00000004)
* vec1 32 ssa_7 = ilt ssa_5, ssa_6
* if ssa_7 {
* block block_5:
* } else {
* block block_6:
* break
* }
* block block_7:
* // Stuff from block 1
* // Stuff from block 2
* vec1 32 ssa_4 = load_const (0x00000001)
* vec1 32 ssa_5 = iadd ssa_2, ssa_4
* }
*/
static bool
opt_peel_loop_initial_if(nir_loop *loop)
{
nir_block *header_block = nir_loop_first_block(loop);
MAYBE_UNUSED nir_block *prev_block =
nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
/* It would be insane if this were not true */
assert(_mesa_set_search(header_block->predecessors, prev_block));
/* The loop must have exactly one continue block which could be a block
* ending in a continue instruction or the "natural" continue from the
* last block in the loop back to the top.
*/
if (header_block->predecessors->entries != 2)
return false;
nir_block *continue_block = find_continue_block(loop);
nir_cf_node *if_node = nir_cf_node_next(&header_block->cf_node);
if (!if_node || if_node->type != nir_cf_node_if)
return false;
nir_if *nif = nir_cf_node_as_if(if_node);
assert(nif->condition.is_ssa);
nir_ssa_def *cond = nif->condition.ssa;
if (cond->parent_instr->type != nir_instr_type_phi)
return false;
nir_phi_instr *cond_phi = nir_instr_as_phi(cond->parent_instr);
if (cond->parent_instr->block != header_block)
return false;
/* We already know we have exactly one continue */
assert(exec_list_length(&cond_phi->srcs) == 2);
uint32_t entry_val = 0, continue_val = 0;
nir_foreach_phi_src(src, cond_phi) {
assert(src->src.is_ssa);
nir_const_value *const_src = nir_src_as_const_value(src->src);
if (!const_src)
return false;
if (src->pred == continue_block) {
continue_val = const_src->u32[0];
} else {
assert(src->pred == prev_block);
entry_val = const_src->u32[0];
}
}
/* If they both execute or both don't execute, this is a job for
* nir_dead_cf, not this pass.
*/
if ((entry_val && continue_val) || (!entry_val && !continue_val))
return false;
struct exec_list *continue_list, *entry_list;
if (continue_val) {
continue_list = &nif->then_list;
entry_list = &nif->else_list;
} else {
continue_list = &nif->else_list;
entry_list = &nif->then_list;
}
/* We want to be moving the contents of entry_list to above the loop so it
* can't contain any break or continue instructions.
*/
foreach_list_typed(nir_cf_node, cf_node, node, entry_list) {
nir_foreach_block_in_cf_node(block, cf_node) {
nir_instr *last_instr = nir_block_last_instr(block);
if (last_instr && last_instr->type == nir_instr_type_jump)
return false;
}
}
/* We're about to re-arrange a bunch of blocks so make sure that we don't
* have deref uses which cross block boundaries. We don't want a deref
* accidentally ending up in a phi.
*/
nir_rematerialize_derefs_in_use_blocks_impl(
nir_cf_node_get_function(&loop->cf_node));
/* Before we do anything, convert the loop to LCSSA. We're about to
* replace a bunch of SSA defs with registers and this will prevent any of
* it from leaking outside the loop.
*/
nir_convert_loop_to_lcssa(loop);
nir_block *after_if_block =
nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
/* Get rid of phis in the header block since we will be duplicating it */
nir_lower_phis_to_regs_block(header_block);
/* Get rid of phis after the if since dominance will change */
nir_lower_phis_to_regs_block(after_if_block);
/* Get rid of SSA defs in the pieces we're about to move around */
nir_lower_ssa_defs_to_regs_block(header_block);
nir_foreach_block_in_cf_node(block, &nif->cf_node)
nir_lower_ssa_defs_to_regs_block(block);
nir_cf_list header, tmp;
nir_cf_extract(&header, nir_before_block(header_block),
nir_after_block(header_block));
nir_cf_list_clone(&tmp, &header, &loop->cf_node, NULL);
nir_cf_reinsert(&tmp, nir_before_cf_node(&loop->cf_node));
nir_cf_extract(&tmp, nir_before_cf_list(entry_list),
nir_after_cf_list(entry_list));
nir_cf_reinsert(&tmp, nir_before_cf_node(&loop->cf_node));
nir_cf_reinsert(&header, nir_after_block_before_jump(continue_block));
/* Get continue block again as the previous reinsert might have removed the block. */
continue_block = find_continue_block(loop);
nir_cf_extract(&tmp, nir_before_cf_list(continue_list),
nir_after_cf_list(continue_list));
nir_cf_reinsert(&tmp, nir_after_block_before_jump(continue_block));
nir_cf_node_remove(&nif->cf_node);
return true;
}
static bool
is_block_empty(nir_block *block)
{
return nir_cf_node_is_last(&block->cf_node) &&
exec_list_is_empty(&block->instr_list);
}
static bool
nir_block_ends_in_continue(nir_block *block)
{
if (exec_list_is_empty(&block->instr_list))
return false;
nir_instr *instr = nir_block_last_instr(block);
return instr->type == nir_instr_type_jump &&
nir_instr_as_jump(instr)->type == nir_jump_continue;
}
/**
* This optimization turns:
*
* loop {
* ...
* if (cond) {
* do_work_1();
* continue;
* } else {
* }
* do_work_2();
* }
*
* into:
*
* loop {
* ...
* if (cond) {
* do_work_1();
* continue;
* } else {
* do_work_2();
* }
* }
*
* The continue should then be removed by nir_opt_trivial_continues() and the
* loop can potentially be unrolled.
*
* Note: do_work_2() is only ever blocks and nested loops. We could also nest
* other if-statments in the branch which would allow further continues to
* be removed. However in practice this can result in increased register
* pressure.
*/
static bool
opt_if_loop_last_continue(nir_loop *loop)
{
/* Get the last if-stament in the loop */
nir_block *last_block = nir_loop_last_block(loop);
nir_cf_node *if_node = nir_cf_node_prev(&last_block->cf_node);
while (if_node) {
if (if_node->type == nir_cf_node_if)
break;
if_node = nir_cf_node_prev(if_node);
}
if (!if_node || if_node->type != nir_cf_node_if)
return false;
nir_if *nif = nir_cf_node_as_if(if_node);
nir_block *then_block = nir_if_last_then_block(nif);
nir_block *else_block = nir_if_last_else_block(nif);
bool then_ends_in_continue = nir_block_ends_in_continue(then_block);
bool else_ends_in_continue = nir_block_ends_in_continue(else_block);
/* If both branches end in a continue do nothing, this should be handled
* by nir_opt_dead_cf().
*/
if (then_ends_in_continue && else_ends_in_continue)
return false;
if (!then_ends_in_continue && !else_ends_in_continue)
return false;
/* Move the last block of the loop inside the last if-statement */
nir_cf_list tmp;
nir_cf_extract(&tmp, nir_after_cf_node(if_node),
nir_after_block(last_block));
if (then_ends_in_continue) {
nir_cursor last_blk_cursor = nir_after_cf_list(&nif->else_list);
nir_cf_reinsert(&tmp,
nir_after_block_before_jump(last_blk_cursor.block));
} else {
nir_cursor last_blk_cursor = nir_after_cf_list(&nif->then_list);
nir_cf_reinsert(&tmp,
nir_after_block_before_jump(last_blk_cursor.block));
}
/* In order to avoid running nir_lower_regs_to_ssa_impl() every time an if
* opt makes progress we leave nir_opt_trivial_continues() to remove the
* continue now that the end of the loop has been simplified.
*/
return true;
}
/* Walk all the phis in the block immediately following the if statement and
* swap the blocks.
*/
static void
rewrite_phi_predecessor_blocks(nir_if *nif,
nir_block *old_then_block,
nir_block *old_else_block,
nir_block *new_then_block,
nir_block *new_else_block)
{
nir_block *after_if_block =
nir_cf_node_as_block(nir_cf_node_next(&nif->cf_node));
nir_foreach_instr(instr, after_if_block) {
if (instr->type != nir_instr_type_phi)
continue;
nir_phi_instr *phi = nir_instr_as_phi(instr);
foreach_list_typed(nir_phi_src, src, node, &phi->srcs) {
if (src->pred == old_then_block) {
src->pred = new_then_block;
} else if (src->pred == old_else_block) {
src->pred = new_else_block;
}
}
}
}
/**
* This optimization turns:
*
* if (cond) {
* } else {
* do_work();
* }
*
* into:
*
* if (!cond) {
* do_work();
* } else {
* }
*/
static bool
opt_if_simplification(nir_builder *b, nir_if *nif)
{
/* Only simplify if the then block is empty and the else block is not. */
if (!is_block_empty(nir_if_first_then_block(nif)) ||
is_block_empty(nir_if_first_else_block(nif)))
return false;
/* Make sure the condition is a comparison operation. */
nir_instr *src_instr = nif->condition.ssa->parent_instr;
if (src_instr->type != nir_instr_type_alu)
return false;
nir_alu_instr *alu_instr = nir_instr_as_alu(src_instr);
if (!nir_alu_instr_is_comparison(alu_instr))
return false;
/* Insert the inverted instruction and rewrite the condition. */
b->cursor = nir_after_instr(&alu_instr->instr);
nir_ssa_def *new_condition =
nir_inot(b, &alu_instr->dest.dest.ssa);
nir_if_rewrite_condition(nif, nir_src_for_ssa(new_condition));
/* Grab pointers to the last then/else blocks for fixing up the phis. */
nir_block *then_block = nir_if_last_then_block(nif);
nir_block *else_block = nir_if_last_else_block(nif);
rewrite_phi_predecessor_blocks(nif, then_block, else_block, else_block,
then_block);
/* Finally, move the else block to the then block. */
nir_cf_list tmp;
nir_cf_extract(&tmp, nir_before_cf_list(&nif->else_list),
nir_after_cf_list(&nif->else_list));
nir_cf_reinsert(&tmp, nir_before_cf_list(&nif->then_list));
return true;
}
/**
* This optimization simplifies potential loop terminators which then allows
* other passes such as opt_if_simplification() and loop unrolling to progress
* further:
*
* if (cond) {
* ... then block instructions ...
* } else {
* ...
* break;
* }
*
* into:
*
* if (cond) {
* } else {
* ...
* break;
* }
* ... then block instructions ...
*/
static bool
opt_if_loop_terminator(nir_if *nif)
{
nir_block *break_blk = NULL;
nir_block *continue_from_blk = NULL;
bool continue_from_then = true;
nir_block *last_then = nir_if_last_then_block(nif);
nir_block *last_else = nir_if_last_else_block(nif);
if (nir_block_ends_in_break(last_then)) {
break_blk = last_then;
continue_from_blk = last_else;
continue_from_then = false;
} else if (nir_block_ends_in_break(last_else)) {
break_blk = last_else;
continue_from_blk = last_then;
}
/* Continue if the if-statement contained no jumps at all */
if (!break_blk)
return false;
/* If the continue from block is empty then return as there is nothing to
* move.
*/
nir_block *first_continue_from_blk = continue_from_then ?
nir_if_first_then_block(nif) :
nir_if_first_else_block(nif);
if (is_block_empty(first_continue_from_blk))
return false;
if (!nir_is_trivial_loop_if(nif, break_blk))
return false;
/* Finally, move the continue from branch after the if-statement. */
nir_cf_list tmp;
nir_cf_extract(&tmp, nir_before_block(first_continue_from_blk),
nir_after_block(continue_from_blk));
nir_cf_reinsert(&tmp, nir_after_cf_node(&nif->cf_node));
return true;
}
static bool
evaluate_if_condition(nir_if *nif, nir_cursor cursor, bool *value)
{
nir_block *use_block = nir_cursor_current_block(cursor);
if (nir_block_dominates(nir_if_first_then_block(nif), use_block)) {
*value = true;
return true;
} else if (nir_block_dominates(nir_if_first_else_block(nif), use_block)) {
*value = false;
return true;
} else {
return false;
}
}
static nir_ssa_def *
clone_alu_and_replace_src_defs(nir_builder *b, const nir_alu_instr *alu,
nir_ssa_def **src_defs)
{
nir_alu_instr *nalu = nir_alu_instr_create(b->shader, alu->op);
nalu->exact = alu->exact;
nir_ssa_dest_init(&nalu->instr, &nalu->dest.dest,
alu->dest.dest.ssa.num_components,
alu->dest.dest.ssa.bit_size, alu->dest.dest.ssa.name);
nalu->dest.saturate = alu->dest.saturate;
nalu->dest.write_mask = alu->dest.write_mask;
for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
assert(alu->src[i].src.is_ssa);
nalu->src[i].src = nir_src_for_ssa(src_defs[i]);
nalu->src[i].negate = alu->src[i].negate;
nalu->src[i].abs = alu->src[i].abs;
memcpy(nalu->src[i].swizzle, alu->src[i].swizzle,
sizeof(nalu->src[i].swizzle));
}
nir_builder_instr_insert(b, &nalu->instr);
return &nalu->dest.dest.ssa;;
}
/*
* This propagates if condition evaluation down the chain of some alu
* instructions. For example by checking the use of some of the following alu
* instruction we can eventually replace ssa_107 with NIR_TRUE.
*
* loop {
* block block_1:
* vec1 32 ssa_85 = load_const (0x00000002)
* vec1 32 ssa_86 = ieq ssa_48, ssa_85
* vec1 32 ssa_87 = load_const (0x00000001)
* vec1 32 ssa_88 = ieq ssa_48, ssa_87
* vec1 32 ssa_89 = ior ssa_86, ssa_88
* vec1 32 ssa_90 = ieq ssa_48, ssa_0
* vec1 32 ssa_91 = ior ssa_89, ssa_90
* if ssa_86 {
* block block_2:
* ...
* break
* } else {
* block block_3:
* }
* block block_4:
* if ssa_88 {
* block block_5:
* ...
* break
* } else {
* block block_6:
* }
* block block_7:
* if ssa_90 {
* block block_8:
* ...
* break
* } else {
* block block_9:
* }
* block block_10:
* vec1 32 ssa_107 = inot ssa_91
* if ssa_107 {
* block block_11:
* break
* } else {
* block block_12:
* }
* }
*/
static bool
propagate_condition_eval(nir_builder *b, nir_if *nif, nir_src *use_src,
nir_src *alu_use, nir_alu_instr *alu,
bool is_if_condition)
{
bool bool_value;
b->cursor = nir_before_src(alu_use, is_if_condition);
if (!evaluate_if_condition(nif, b->cursor, &bool_value))
return false;
nir_ssa_def *def[4] = {0};
for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
if (alu->src[i].src.ssa == use_src->ssa) {
def[i] = nir_imm_bool(b, bool_value);
} else {
def[i] = alu->src[i].src.ssa;
}
}
nir_ssa_def *nalu = clone_alu_and_replace_src_defs(b, alu, def);
/* Rewrite use to use new alu instruction */
nir_src new_src = nir_src_for_ssa(nalu);
if (is_if_condition)
nir_if_rewrite_condition(alu_use->parent_if, new_src);
else
nir_instr_rewrite_src(alu_use->parent_instr, alu_use, new_src);
return true;
}
static bool
can_propagate_through_alu(nir_src *src)
{
if (src->parent_instr->type != nir_instr_type_alu)
return false;
nir_alu_instr *alu = nir_instr_as_alu(src->parent_instr);
switch (alu->op) {
case nir_op_ior:
case nir_op_iand:
case nir_op_inot:
case nir_op_b2i32:
return true;
case nir_op_bcsel:
return src == &alu->src[0].src;
default:
return false;
}
}
static bool
evaluate_condition_use(nir_builder *b, nir_if *nif, nir_src *use_src,
bool is_if_condition)
{
bool progress = false;
b->cursor = nir_before_src(use_src, is_if_condition);
bool bool_value;
if (evaluate_if_condition(nif, b->cursor, &bool_value)) {
/* Rewrite use to use const */
nir_src imm_src = nir_src_for_ssa(nir_imm_bool(b, bool_value));
if (is_if_condition)
nir_if_rewrite_condition(use_src->parent_if, imm_src);
else
nir_instr_rewrite_src(use_src->parent_instr, use_src, imm_src);
progress = true;
}
if (!is_if_condition && can_propagate_through_alu(use_src)) {
nir_alu_instr *alu = nir_instr_as_alu(use_src->parent_instr);
nir_foreach_use_safe(alu_use, &alu->dest.dest.ssa) {
progress |= propagate_condition_eval(b, nif, use_src, alu_use, alu,
false);
}
nir_foreach_if_use_safe(alu_use, &alu->dest.dest.ssa) {
progress |= propagate_condition_eval(b, nif, use_src, alu_use, alu,
true);
}
}
return progress;
}
static bool
opt_if_evaluate_condition_use(nir_builder *b, nir_if *nif)
{
bool progress = false;
/* Evaluate any uses of the if condition inside the if branches */
assert(nif->condition.is_ssa);
nir_foreach_use_safe(use_src, nif->condition.ssa) {
progress |= evaluate_condition_use(b, nif, use_src, false);
}
nir_foreach_if_use_safe(use_src, nif->condition.ssa) {
if (use_src->parent_if != nif)
progress |= evaluate_condition_use(b, nif, use_src, true);
}
return progress;
}
static void
simple_merge_if(nir_if *dest_if, nir_if *src_if, bool dest_if_then,
bool src_if_then)
{
/* Now merge the if branch */
nir_block *dest_blk = dest_if_then ? nir_if_last_then_block(dest_if)
: nir_if_last_else_block(dest_if);
struct exec_list *list = src_if_then ? &src_if->then_list
: &src_if->else_list;
nir_cf_list if_cf_list;
nir_cf_extract(&if_cf_list, nir_before_cf_list(list),
nir_after_cf_list(list));
nir_cf_reinsert(&if_cf_list, nir_after_block(dest_blk));
}
static bool
opt_if_merge(nir_if *nif)
{
bool progress = false;
nir_block *next_blk = nir_cf_node_cf_tree_next(&nif->cf_node);
if (next_blk && nif->condition.is_ssa) {
nir_if *next_if = nir_block_get_following_if(next_blk);
if (next_if && next_if->condition.is_ssa) {
/* Here we merge two consecutive ifs that have the same
* condition e.g:
*
* if ssa_12 {
* ...
* } else {
* ...
* }
* if ssa_12 {
* ...
* } else {
* ...
* }
*
* Note: This only merges if-statements when the block between them
* is empty. The reason we don't try to merge ifs that just have phis
* between them is because this can results in increased register
* pressure. For example when merging if ladders created by indirect
* indexing.
*/
if (nif->condition.ssa == next_if->condition.ssa &&
exec_list_is_empty(&next_blk->instr_list)) {
simple_merge_if(nif, next_if, true, true);
simple_merge_if(nif, next_if, false, false);
nir_block *new_then_block = nir_if_last_then_block(nif);
nir_block *new_else_block = nir_if_last_else_block(nif);
nir_block *old_then_block = nir_if_last_then_block(next_if);
nir_block *old_else_block = nir_if_last_else_block(next_if);
/* Rewrite the predecessor block for any phis following the second
* if-statement.
*/
rewrite_phi_predecessor_blocks(next_if, old_then_block,
old_else_block,
new_then_block,
new_else_block);
/* Move phis after merged if to avoid them being deleted when we
* remove the merged if-statement.
*/
nir_block *after_next_if_block =
nir_cf_node_as_block(nir_cf_node_next(&next_if->cf_node));
nir_foreach_instr_safe(instr, after_next_if_block) {
if (instr->type != nir_instr_type_phi)
break;
exec_node_remove(&instr->node);
exec_list_push_tail(&next_blk->instr_list, &instr->node);
instr->block = next_blk;
}
nir_cf_node_remove(&next_if->cf_node);
progress = true;
}
}
}
return progress;
}
static bool
opt_if_cf_list(nir_builder *b, struct exec_list *cf_list)
{
bool progress = false;
foreach_list_typed(nir_cf_node, cf_node, node, cf_list) {
switch (cf_node->type) {
case nir_cf_node_block:
break;
case nir_cf_node_if: {
nir_if *nif = nir_cf_node_as_if(cf_node);
progress |= opt_if_cf_list(b, &nif->then_list);
progress |= opt_if_cf_list(b, &nif->else_list);
progress |= opt_if_loop_terminator(nif);
progress |= opt_if_merge(nif);
progress |= opt_if_simplification(b, nif);
break;
}
case nir_cf_node_loop: {
nir_loop *loop = nir_cf_node_as_loop(cf_node);
progress |= opt_if_cf_list(b, &loop->body);
progress |= opt_peel_loop_initial_if(loop);
progress |= opt_if_loop_last_continue(loop);
break;
}
case nir_cf_node_function:
unreachable("Invalid cf type");
}
}
return progress;
}
/**
* These optimisations depend on nir_metadata_block_index and therefore must
* not do anything to cause the metadata to become invalid.
*/
static bool
opt_if_safe_cf_list(nir_builder *b, struct exec_list *cf_list)
{
bool progress = false;
foreach_list_typed(nir_cf_node, cf_node, node, cf_list) {
switch (cf_node->type) {
case nir_cf_node_block:
break;
case nir_cf_node_if: {
nir_if *nif = nir_cf_node_as_if(cf_node);
progress |= opt_if_safe_cf_list(b, &nif->then_list);
progress |= opt_if_safe_cf_list(b, &nif->else_list);
progress |= opt_if_evaluate_condition_use(b, nif);
break;
}
case nir_cf_node_loop: {
nir_loop *loop = nir_cf_node_as_loop(cf_node);
progress |= opt_if_safe_cf_list(b, &loop->body);
break;
}
case nir_cf_node_function:
unreachable("Invalid cf type");
}
}
return progress;
}
bool
nir_opt_if(nir_shader *shader)
{
bool progress = false;
nir_foreach_function(function, shader) {
if (function->impl == NULL)
continue;
nir_builder b;
nir_builder_init(&b, function->impl);
nir_metadata_require(function->impl, nir_metadata_block_index |
nir_metadata_dominance);
progress = opt_if_safe_cf_list(&b, &function->impl->body);
nir_metadata_preserve(function->impl, nir_metadata_block_index |
nir_metadata_dominance);
if (opt_if_cf_list(&b, &function->impl->body)) {
nir_metadata_preserve(function->impl, nir_metadata_none);
/* If that made progress, we're no longer really in SSA form. We
* need to convert registers back into SSA defs and clean up SSA defs
* that don't dominate their uses.
*/
nir_lower_regs_to_ssa_impl(function->impl);
progress = true;
} else {
#ifndef NDEBUG
function->impl->valid_metadata &= ~nir_metadata_not_properly_reset;
#endif
}
}
return progress;
}
|