Repository navigation
Expand file tree
/
Copy pathlift.go
More file actions
1817 lines (1659 loc) · 47.9 KB
/
Copy pathlift.go
File metadata and controls
1817 lines (1659 loc) · 47.9 KB
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
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
// Copyright 2013 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package ir
// This file defines the lifting pass which tries to "lift" Alloc
// cells (new/local variables) into SSA registers, replacing loads
// with the dominating stored value, eliminating loads and stores, and
// inserting φ- and σ-nodes as needed.
// Cited papers and resources:
//
// Ron Cytron et al. 1991. Efficiently computing SSA form...
// https://doi.acm.org/10.1145/115372.115320
//
// Cooper, Harvey, Kennedy. 2001. A Simple, Fast Dominance Algorithm.
// Software Practice and Experience 2001, 4:1-10.
// https://www.hipersoft.rice.edu/grads/publications/dom14.pdf
//
// Daniel Berlin, llvmdev mailing list, 2012.
// https://lists.cs.uiuc.edu/pipermail/llvmdev/2012-January/046638.html
// (Be sure to expand the whole thread.)
//
// C. Scott Ananian. 1997. The static single information form.
//
// Jeremy Singer. 2006. Static program analysis based on virtual register renaming.
// TODO(adonovan): opt: there are many optimizations worth evaluating, and
// the conventional wisdom for SSA construction is that a simple
// algorithm well engineered often beats those of better asymptotic
// complexity on all but the most egregious inputs.
//
// Danny Berlin suggests that the Cooper et al. algorithm for
// computing the dominance frontier is superior to Cytron et al.
// Furthermore he recommends that rather than computing the DF for the
// whole function then renaming all alloc cells, it may be cheaper to
// compute the DF for each alloc cell separately and throw it away.
//
// Consider exploiting liveness information to avoid creating dead
// φ-nodes which we then immediately remove.
//
// Also see many other "TODO: opt" suggestions in the code.
import (
"encoding/binary"
"fmt"
"os"
"slices"
)
// If true, show diagnostic information at each step of lifting.
// Very verbose.
const debugLifting = false
// domFrontier maps each block to the set of blocks in its dominance
// frontier. The outer slice is conceptually a map keyed by
// Block.Index. The inner slice is conceptually a set, possibly
// containing duplicates.
//
// TODO(adonovan): opt: measure impact of dups; consider a packed bit
// representation, e.g. big.Int, and bitwise parallel operations for
// the union step in the Children loop.
//
// domFrontier's methods mutate the slice's elements but not its
// length, so their receivers needn't be pointers.
type domFrontier BlockMap[[]*BasicBlock]
func (df domFrontier) add(u, v *BasicBlock) {
df[u.Index] = append(df[u.Index], v)
}
// build builds the dominance frontier df for the dominator tree of
// fn, using the algorithm found in A Simple, Fast Dominance
// Algorithm, Figure 5.
//
// TODO(adonovan): opt: consider Berlin approach, computing pruned SSA
// by pruning the entire IDF computation, rather than merely pruning
// the DF -> IDF step.
func (df domFrontier) build(fn *Function) {
for _, b := range fn.Blocks {
preds := b.Preds[0:len(b.Preds):len(b.Preds)]
if b == fn.Exit {
for i, v := range fn.fakeExits.values {
if v {
preds = append(preds, fn.Blocks[i])
}
}
}
if len(preds) >= 2 {
for _, p := range preds {
runner := p
for runner != b.dom.idom {
df.add(runner, b)
runner = runner.dom.idom
}
}
}
}
}
func buildDomFrontier(fn *Function) domFrontier {
df := make(domFrontier, len(fn.Blocks))
df.build(fn)
return df
}
type postDomFrontier BlockMap[[]*BasicBlock]
func (rdf postDomFrontier) add(u, v *BasicBlock) {
rdf[u.Index] = append(rdf[u.Index], v)
}
func (rdf postDomFrontier) build(fn *Function) {
for _, b := range fn.Blocks {
succs := b.Succs[0:len(b.Succs):len(b.Succs)]
if fn.fakeExits.Has(b) {
succs = append(succs, fn.Exit)
}
if len(succs) >= 2 {
for _, s := range succs {
runner := s
for runner != b.pdom.idom {
rdf.add(runner, b)
runner = runner.pdom.idom
}
}
}
}
}
func buildPostDomFrontier(fn *Function) postDomFrontier {
rdf := make(postDomFrontier, len(fn.Blocks))
rdf.build(fn)
return rdf
}
func removeInstr(refs []Instruction, instr Instruction) []Instruction {
return removeInstrsIf(refs, func(i Instruction) bool { return i == instr })
}
func removeInstrsIf(refs []Instruction, p func(Instruction) bool) []Instruction {
return slices.DeleteFunc(refs, p)
}
func clearInstrs(instrs []Instruction) {
for i := range instrs {
instrs[i] = nil
}
}
func numberNodesPerBlock(f *Function) {
for _, b := range f.Blocks {
var base ID
for _, instr := range b.Instrs {
if instr == nil {
continue
}
instr.setID(base)
base++
}
}
}
// lift replaces local and new Allocs accessed only with
// load/store by IR registers, inserting φ- and σ-nodes where necessary.
// The result is a program in pruned SSI form.
//
// Preconditions:
// - fn has no dead blocks (blockopt has run).
// - Def/use info (Operands and Referrers) is up-to-date.
// - The dominator tree is up-to-date.
func lift(fn *Function) bool {
// TODO(adonovan): opt: lots of little optimizations may be
// worthwhile here, especially if they cause us to avoid
// buildDomFrontier. For example:
//
// - Alloc never loaded? Eliminate.
// - Alloc never stored? Replace all loads with a zero constant.
// - Alloc stored once? Replace loads with dominating store;
// don't forget that an Alloc is itself an effective store
// of zero.
// - Alloc used only within a single block?
// Use degenerate algorithm avoiding φ-nodes.
// - Consider synergy with scalar replacement of aggregates (SRA).
// e.g. *(&x.f) where x is an Alloc.
// Perhaps we'd get better results if we generated this as x.f
// i.e. Field(x, .f) instead of Load(FieldIndex(x, .f)).
// Unclear.
//
// But we will start with the simplest correct code.
var df domFrontier
var rdf postDomFrontier
var closure *closure
var newPhis BlockMap[[]newPhi]
var newSigmas BlockMap[[]newSigma]
// During this pass we will replace some BasicBlock.Instrs
// (allocs, loads and stores) with nil, keeping a count in
// BasicBlock.gaps. At the end we will reset Instrs to the
// concatenation of all non-dead newPhis and non-nil Instrs
// for the block, reusing the original array if space permits.
// While we're here, we also eliminate 'rundefers'
// instructions and ssa:deferstack() in functions that contain no
// 'defer' instructions. Eliminate ssa:deferstack() if it does not
// escape.
usesDefer := false
deferstackAlloc, deferstackCall := deferstackPreamble(fn)
eliminateDeferStack := deferstackAlloc != nil && !deferstackAlloc.Heap
// Determine which allocs we can lift and number them densely.
// The renaming phase uses this numbering for compact maps.
numAllocs := 0
instructions := make(BlockMap[liftInstructions], len(fn.Blocks))
for i := range instructions {
instructions[i].insertInstructions = map[Instruction][]Instruction{}
}
// Number nodes, for liftable
numberNodesPerBlock(fn)
for _, b := range fn.Blocks {
b.gaps = 0
b.rundefers = 0
for _, instr := range b.Instrs {
switch instr := instr.(type) {
case *Alloc:
if !liftable(instr, instructions) {
instr.index = -1
continue
}
if numAllocs == 0 {
df = buildDomFrontier(fn)
rdf = buildPostDomFrontier(fn)
if len(fn.Blocks) > 2 {
closure = transitiveClosure(fn)
}
newPhis = make(BlockMap[[]newPhi], len(fn.Blocks))
newSigmas = make(BlockMap[[]newSigma], len(fn.Blocks))
if debugLifting {
title := false
for i, blocks := range df {
if blocks != nil {
if !title {
fmt.Fprintf(os.Stderr, "Dominance frontier of %s:\n", fn)
title = true
}
fmt.Fprintf(os.Stderr, "\t%s: %s\n", fn.Blocks[i], blocks)
}
}
}
}
instr.index = numAllocs
numAllocs++
case *Defer:
usesDefer = true
if eliminateDeferStack {
// Clear _DeferStack and remove references to loads
if instr._DeferStack != nil {
if refs := instr._DeferStack.Referrers(); refs != nil {
*refs = removeInstr(*refs, instr)
}
instr._DeferStack = nil
}
}
case *RunDefers:
b.rundefers++
}
}
}
if numAllocs > 0 {
for _, b := range fn.Blocks {
work := instructions[b.Index]
for _, rename := range work.renameAllocs {
for _, instr_ := range b.Instrs[rename.startingAt:] {
replace(instr_, rename.from, rename.to)
}
}
}
for _, b := range fn.Blocks {
work := instructions[b.Index]
if len(work.insertInstructions) != 0 {
newInstrs := make([]Instruction, 0, len(fn.Blocks)+len(work.insertInstructions)*3)
for _, instr := range b.Instrs {
if add, ok := work.insertInstructions[instr]; ok {
newInstrs = append(newInstrs, add...)
}
newInstrs = append(newInstrs, instr)
}
b.Instrs = newInstrs
}
}
// TODO(dh): remove inserted allocs that end up unused after lifting.
for _, b := range fn.Blocks {
for _, instr := range b.Instrs {
if instr, ok := instr.(*Alloc); ok && instr.index >= 0 {
liftAlloc(closure, df, rdf, instr, newPhis, newSigmas)
}
}
}
// renaming maps an alloc (keyed by index) to its replacement
// value. Initially the renaming contains nil, signifying the
// zero constant of the appropriate type; we construct the
// Const lazily at most once on each path through the domtree.
// TODO(adonovan): opt: cache per-function not per subtree.
renaming := make([]Value, numAllocs)
// Renaming.
rename(fn.Blocks[0], renaming, newPhis, newSigmas)
simplifyPhisAndSigmas(newPhis, newSigmas)
// Eliminate dead φ- and σ-nodes.
markLiveNodes(fn.Blocks, newPhis, newSigmas)
// Eliminate ssa:deferstack() call.
if eliminateDeferStack {
b := deferstackCall.block
for i, instr := range b.Instrs {
if instr == deferstackCall {
b.Instrs[i] = nil
b.gaps++
break
}
}
}
}
// Prepend remaining live φ-nodes to each block and possibly kill rundefers.
for _, b := range fn.Blocks {
var head []Instruction
if numAllocs > 0 {
nps := newPhis[b.Index]
head = make([]Instruction, 0, len(nps))
for _, pred := range b.Preds {
nss := newSigmas[pred.Index]
idx := pred.succIndex(b)
for _, newSigma := range nss {
if sigma := newSigma.sigmas[idx]; sigma != nil && sigma.live {
head = append(head, sigma)
// we didn't populate referrers before, as most
// sigma nodes will be killed
if refs := sigma.X.Referrers(); refs != nil {
*refs = append(*refs, sigma)
}
} else if sigma != nil {
sigma.block = nil
}
}
}
for _, np := range nps {
if np.phi.live {
head = append(head, np.phi)
} else {
for _, edge := range np.phi.Edges {
if refs := edge.Referrers(); refs != nil {
*refs = removeInstr(*refs, np.phi)
}
}
np.phi.block = nil
}
}
}
rundefersToKill := b.rundefers
if usesDefer {
rundefersToKill = 0
}
j := len(head)
if j+b.gaps+rundefersToKill == 0 {
continue // fast path: no new phis or gaps
}
// We could do straight copies instead of element-wise copies
// when both b.gaps and rundefersToKill are zero. However,
// that seems to only be the case ~1% of the time, which
// doesn't seem worth the extra branch.
// Remove dead instructions, add phis and sigmas
ns := len(b.Instrs) + j - b.gaps - rundefersToKill
if ns <= cap(b.Instrs) {
// b.Instrs has enough capacity to store all instructions
// OPT(dh): check cap vs the actually required space; if
// there is a big enough difference, it may be worth
// allocating a new slice, to avoid pinning memory.
dst := b.Instrs[:cap(b.Instrs)]
i := len(dst) - 1
for n := len(b.Instrs) - 1; n >= 0; n-- {
instr := dst[n]
if instr == nil {
continue
}
if !usesDefer {
if _, ok := instr.(*RunDefers); ok {
continue
}
}
dst[i] = instr
i--
}
off := i + 1 - len(head)
// aid GC
clearInstrs(dst[:off])
dst = dst[off:]
copy(dst, head)
b.Instrs = dst
} else {
// not enough space, so allocate a new slice and copy
// over.
dst := make([]Instruction, ns)
copy(dst, head)
for _, instr := range b.Instrs {
if instr == nil {
continue
}
if !usesDefer {
if _, ok := instr.(*RunDefers); ok {
continue
}
}
dst[j] = instr
j++
}
b.Instrs = dst
}
}
// Remove any fn.Locals that were lifted.
j := 0
for _, l := range fn.Locals {
if l.index < 0 {
fn.Locals[j] = l
j++
}
}
// Nil out fn.Locals[j:] to aid GC.
for i := j; i < len(fn.Locals); i++ {
fn.Locals[i] = nil
}
fn.Locals = fn.Locals[:j]
return numAllocs > 0
}
func hasDirectReferrer(instr Instruction) bool {
for _, instr := range *instr.Referrers() {
switch instr.(type) {
case *Phi, *Sigma:
// ignore
default:
return true
}
}
return false
}
func markLiveNodes(blocks []*BasicBlock, newPhis BlockMap[[]newPhi], newSigmas BlockMap[[]newSigma]) {
// Phis and sigmas may become dead due to optimization passes. We may also insert more nodes than strictly
// necessary, e.g. sigma nodes for constants, which will never be used.
// Phi and sigma nodes are considered live if a non-phi, non-sigma
// node uses them. Once we find a node that is live, we mark all
// of its operands as used, too.
for _, npList := range newPhis {
for _, np := range npList {
phi := np.phi
if !phi.live && hasDirectReferrer(phi) {
markLivePhi(phi)
}
}
}
for _, npList := range newSigmas {
for _, np := range npList {
for _, sigma := range np.sigmas {
if sigma != nil && !sigma.live && hasDirectReferrer(sigma) {
markLiveSigma(sigma)
}
}
}
}
// Existing φ-nodes due to && and || operators
// are all considered live (see Go issue 19622).
for _, b := range blocks {
for _, phi := range b.phis() {
markLivePhi(phi.(*Phi))
}
}
}
func markLivePhi(phi *Phi) {
phi.live = true
for _, rand := range phi.Edges {
switch rand := rand.(type) {
case *Phi:
if !rand.live {
markLivePhi(rand)
}
case *Sigma:
if !rand.live {
markLiveSigma(rand)
}
}
}
}
func markLiveSigma(sigma *Sigma) {
sigma.live = true
switch rand := sigma.X.(type) {
case *Phi:
if !rand.live {
markLivePhi(rand)
}
case *Sigma:
if !rand.live {
markLiveSigma(rand)
}
}
}
// simplifyPhisAndSigmas removes duplicate phi and sigma nodes,
// and replaces trivial phis with non-phi alternatives. Phi
// nodes where all edges are identical, or consist of only the phi
// itself and one other value, may be replaced with the value.
func simplifyPhisAndSigmas(newPhis BlockMap[[]newPhi], newSigmas BlockMap[[]newSigma]) {
// temporary numbering of values used in phis so that we can build map keys
var id ID
for _, npList := range newPhis {
for _, np := range npList {
for _, edge := range np.phi.Edges {
edge.setID(id)
id++
}
}
}
// find all phis that are trivial and can be replaced with a
// non-phi value. run until we reach a fixpoint, because replacing
// a phi may make other phis trivial.
for changed := true; changed; {
changed = false
for _, npList := range newPhis {
for _, np := range npList {
if np.phi.live {
// we're reusing 'live' to mean 'dead' in the context of simplifyPhisAndSigmas
continue
}
if r, ok := isUselessPhi(np.phi); ok {
// useless phi, replace its uses with the
// replacement value. the dead phi pass will clean
// up the phi afterwards.
replaceAll(np.phi, r)
np.phi.live = true
changed = true
}
}
}
// Replace duplicate sigma nodes with a single node. These nodes exist when multiple allocs get replaced with the
// same dominating store.
for _, sigmaList := range newSigmas {
primarySigmas := map[struct {
succ int
v Value
}]*Sigma{}
for _, sigmas := range sigmaList {
for succ, sigma := range sigmas.sigmas {
if sigma == nil {
continue
}
if sigma.live {
// we're reusing 'live' to mean 'dead' in the context of simplifyPhisAndSigmas
continue
}
key := struct {
succ int
v Value
}{succ, sigma.X}
if alt, ok := primarySigmas[key]; ok {
replaceAll(sigma, alt)
sigma.live = true
changed = true
} else {
primarySigmas[key] = sigma
}
}
}
}
// Replace duplicate phi nodes with a single node. As far as we know, these duplicate nodes only ever exist
// because of the previous sigma deduplication.
keyb := make([]byte, 0, 4*8)
for _, npList := range newPhis {
primaryPhis := map[string]*Phi{}
for _, np := range npList {
if np.phi.live {
continue
}
if n := len(np.phi.Edges) * 8; cap(keyb) >= n {
keyb = keyb[:n]
} else {
keyb = make([]byte, n, n*2)
}
for i, e := range np.phi.Edges {
binary.LittleEndian.PutUint64(keyb[i*8:i*8+8], uint64(e.ID()))
}
if alt, ok := primaryPhis[string(keyb)]; ok {
replaceAll(np.phi, alt)
np.phi.live = true
changed = true
} else {
primaryPhis[string(keyb)] = np.phi
}
}
}
}
for _, npList := range newPhis {
for _, np := range npList {
np.phi.live = false
for _, edge := range np.phi.Edges {
edge.setID(0)
}
}
}
for _, sigmaList := range newSigmas {
for _, sigmas := range sigmaList {
for _, sigma := range sigmas.sigmas {
if sigma != nil {
sigma.live = false
}
}
}
}
}
type BlockSet struct {
idx int
values []bool
count int
}
func NewBlockSet(size int) *BlockSet {
return &BlockSet{values: make([]bool, size)}
}
func (s *BlockSet) Set(s2 *BlockSet) {
copy(s.values, s2.values)
s.count = 0
for _, v := range s.values {
if v {
s.count++
}
}
}
func (s *BlockSet) Num() int {
return s.count
}
func (s *BlockSet) Has(b *BasicBlock) bool {
if b.Index >= len(s.values) {
return false
}
return s.values[b.Index]
}
// add adds b to the set and returns true if the set changed.
func (s *BlockSet) Add(b *BasicBlock) bool {
if s.values[b.Index] {
return false
}
s.count++
s.values[b.Index] = true
s.idx = b.Index
return true
}
func (s *BlockSet) Clear() {
for j := range s.values {
s.values[j] = false
}
s.count = 0
}
// take removes an arbitrary element from a set s and
// returns its index, or returns -1 if empty.
func (s *BlockSet) Take() int {
// [i, end]
for i := s.idx; i < len(s.values); i++ {
if s.values[i] {
s.values[i] = false
s.idx = i
s.count--
return i
}
}
// [start, i)
for i := 0; i < s.idx; i++ {
if s.values[i] {
s.values[i] = false
s.idx = i
s.count--
return i
}
}
return -1
}
type closure struct {
span []uint32
reachables BlockMap[interval]
}
type interval uint32
const (
flagMask = 1 << 31
numBits = 20
lengthBits = 32 - numBits - 1
lengthMask = (1<<lengthBits - 1) << numBits
numMask = 1<<numBits - 1
)
func (c closure) has(s, v *BasicBlock) bool {
idx := uint32(v.Index)
if idx == 1 || s.Dominates(v) {
return true
}
r := c.reachable(s.Index)
for i := 0; i < len(r); i++ {
inv := r[i]
var start, end uint32
if inv&flagMask == 0 {
// small interval
start = uint32(inv & numMask)
end = start + uint32(inv&lengthMask)>>numBits
} else {
// large interval
i++
start = uint32(inv & numMask)
end = uint32(r[i])
}
if idx >= start && idx <= end {
return true
}
}
return false
}
func (c closure) reachable(id int) []interval {
return c.reachables[c.span[id]:c.span[id+1]]
}
func (c closure) walk(current *BasicBlock, b *BasicBlock, visited []bool) {
// TODO(dh): the 'current' argument seems to be unused
// TODO(dh): there's no reason for this to be a method
visited[b.Index] = true
for _, succ := range b.Succs {
if visited[succ.Index] {
continue
}
visited[succ.Index] = true
c.walk(current, succ, visited)
}
}
func transitiveClosure(fn *Function) *closure {
reachable := make(BlockMap[bool], len(fn.Blocks))
c := &closure{}
c.span = make([]uint32, len(fn.Blocks)+1)
addInterval := func(start, end uint32) {
if l := end - start; l <= 1<<lengthBits-1 {
n := interval(l<<numBits | start)
c.reachables = append(c.reachables, n)
} else {
n1 := interval(1<<31 | start)
n2 := interval(end)
c.reachables = append(c.reachables, n1, n2)
}
}
for i, b := range fn.Blocks[1:] {
for i := range reachable {
reachable[i] = false
}
c.walk(b, b, reachable)
start := ^uint32(0)
for id, isReachable := range reachable {
if !isReachable {
if start != ^uint32(0) {
end := uint32(id) - 1
addInterval(start, end)
start = ^uint32(0)
}
continue
} else if start == ^uint32(0) {
start = uint32(id)
}
}
if start != ^uint32(0) {
addInterval(start, uint32(len(reachable))-1)
}
c.span[i+2] = uint32(len(c.reachables))
}
return c
}
// newPhi is a pair of a newly introduced φ-node and the lifted Alloc
// it replaces.
type newPhi struct {
phi *Phi
alloc *Alloc
}
type newSigma struct {
alloc *Alloc
sigmas []*Sigma
}
type liftInstructions struct {
insertInstructions map[Instruction][]Instruction
renameAllocs []struct {
from *Alloc
to *Alloc
startingAt int
}
}
// liftable determines if alloc can be lifted, and records instructions to split partially liftable allocs.
//
// In the trivial case, all uses of the alloc can be lifted. This is the case when it is only used for storing into and
// loading from. In that case, no instructions are recorded.
//
// In the more complex case, the alloc is used for storing into and loading from, but it is also used as a value, for
// example because it gets passed to a function, e.g. fn(&x). In this case, uses of the alloc fall into one of two
// categories: those that can be lifted and those that can't. A boundary forms between these two categories in the
// function's control flow: Once an unliftable use is encountered, the alloc is no longer liftable for the remainder of
// the basic block the use is in, nor in any blocks reachable from it.
//
// We record instructions that split the alloc into two allocs: one that is used in liftable uses, and one that is used
// in unliftable uses. Whenever we encounter a boundary between liftable and unliftable uses or blocks, we emit a pair
// of Load and Store that copy the value from the liftable alloc into the unliftable alloc. Taking these instructions
// into account, the normal lifting machinery will completely lift the liftable alloc, store the correct lifted values
// into the unliftable alloc, and will not at all lift the unliftable alloc.
//
// In Go syntax, the transformation looks somewhat like this:
//
// func foo() {
// x := 32
// if cond {
// println(x)
// escape(&x)
// println(x)
// } else {
// println(x)
// }
// println(x)
// }
//
// transforms into
//
// func fooSplitAlloc() {
// x := 32
// var x_ int
// if cond {
// println(x)
// x_ = x
// escape(&x_)
// println(x_)
// } else {
// println(x)
// x_ = x
// }
// println(x_)
// }
func liftable(alloc *Alloc, instructions BlockMap[liftInstructions]) bool {
fn := alloc.block.parent
// Don't lift result values in functions that defer
// calls that may recover from panic.
if fn.hasDefer {
if slices.Contains(fn.results, alloc) {
return false
}
}
type blockDesc struct {
// is the block (partially) unliftable, because it contains unliftable instructions or is reachable by an unliftable block
isUnliftable bool
hasLiftableLoad bool
hasLiftableOther bool
// we need to emit stores in predecessors because the unliftable use is in a phi
storeInPreds bool
lastLiftable int
firstUnliftable int
}
blocks := make(BlockMap[blockDesc], len(fn.Blocks))
for _, b := range fn.Blocks {
blocks[b.Index].lastLiftable = -1
blocks[b.Index].firstUnliftable = len(b.Instrs) + 1
}
// Look at all uses of the alloc and deduce which blocks have liftable or unliftable instructions.
for _, instr := range alloc.referrers {
// Find the first unliftable use
desc := &blocks[instr.Block().Index]
hasUnliftable := false
inHead := false
switch instr := instr.(type) {
case *Store:
if instr.Val == alloc {
hasUnliftable = true
}
case *Load:
case *DebugRef:
case *Phi, *Sigma:
inHead = true
hasUnliftable = true
default:
hasUnliftable = true
}
if hasUnliftable {
desc.isUnliftable = true
if int(instr.ID()) < desc.firstUnliftable {
desc.firstUnliftable = int(instr.ID())
}
if inHead {
desc.storeInPreds = true
desc.firstUnliftable = 0
}
}
}
for _, instr := range alloc.referrers {
// Find the last liftable use, taking the previously calculated firstUnliftable into consideration
desc := &blocks[instr.Block().Index]
if int(instr.ID()) >= desc.firstUnliftable {
continue
}
hasLiftable := false
switch instr := instr.(type) {
case *Store:
if instr.Val != alloc {
desc.hasLiftableOther = true
hasLiftable = true
}
case *Load:
desc.hasLiftableLoad = true
hasLiftable = true
case *DebugRef:
desc.hasLiftableOther = true
}
if hasLiftable {
if int(instr.ID()) > desc.lastLiftable {
desc.lastLiftable = int(instr.ID())
}
}
}
for i := range blocks {
// Update firstUnliftable to be one after lastLiftable. We do this to include the unliftable's preceding
// DebugRefs in the renaming.
if blocks[i].lastLiftable == -1 && !blocks[i].storeInPreds {
// There are no liftable instructions (for this alloc) in this block. Set firstUnliftable to the
// first non-head instruction to avoid inserting the store before phi instructions, which would
// fail validation.
first := -1
instrLoop:
for i, instr := range fn.Blocks[i].Instrs {
switch instr.(type) {
case *Phi, *Sigma:
default:
first = i
break instrLoop