forked from llvm/llvm-project
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathX86InstCombineIntrinsic.cpp
2016 lines (1788 loc) · 73.9 KB
/
X86InstCombineIntrinsic.cpp
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
//===-- X86InstCombineIntrinsic.cpp - X86 specific InstCombine pass -------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
/// \file
/// This file implements a TargetTransformInfo analysis pass specific to the
/// X86 target machine. It uses the target's detailed information to provide
/// more precise answers to certain TTI queries, while letting the target
/// independent and default TTI implementations handle the rest.
///
//===----------------------------------------------------------------------===//
#include "X86TargetTransformInfo.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/IntrinsicsX86.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Transforms/InstCombine/InstCombiner.h"
using namespace llvm;
#define DEBUG_TYPE "x86tti"
/// Return a constant boolean vector that has true elements in all positions
/// where the input constant data vector has an element with the sign bit set.
static Constant *getNegativeIsTrueBoolVec(Constant *V) {
VectorType *IntTy = VectorType::getInteger(cast<VectorType>(V->getType()));
V = ConstantExpr::getBitCast(V, IntTy);
V = ConstantExpr::getICmp(CmpInst::ICMP_SGT, Constant::getNullValue(IntTy),
V);
return V;
}
/// Convert the x86 XMM integer vector mask to a vector of bools based on
/// each element's most significant bit (the sign bit).
static Value *getBoolVecFromMask(Value *Mask) {
// Fold Constant Mask.
if (auto *ConstantMask = dyn_cast<ConstantDataVector>(Mask))
return getNegativeIsTrueBoolVec(ConstantMask);
// Mask was extended from a boolean vector.
Value *ExtMask;
if (PatternMatch::match(
Mask, PatternMatch::m_SExt(PatternMatch::m_Value(ExtMask))) &&
ExtMask->getType()->isIntOrIntVectorTy(1))
return ExtMask;
return nullptr;
}
// TODO: If the x86 backend knew how to convert a bool vector mask back to an
// XMM register mask efficiently, we could transform all x86 masked intrinsics
// to LLVM masked intrinsics and remove the x86 masked intrinsic defs.
static Instruction *simplifyX86MaskedLoad(IntrinsicInst &II, InstCombiner &IC) {
Value *Ptr = II.getOperand(0);
Value *Mask = II.getOperand(1);
Constant *ZeroVec = Constant::getNullValue(II.getType());
// Zero Mask - masked load instruction creates a zero vector.
if (isa<ConstantAggregateZero>(Mask))
return IC.replaceInstUsesWith(II, ZeroVec);
// The mask is constant or extended from a bool vector. Convert this x86
// intrinsic to the LLVM intrinsic to allow target-independent optimizations.
if (Value *BoolMask = getBoolVecFromMask(Mask)) {
// First, cast the x86 intrinsic scalar pointer to a vector pointer to match
// the LLVM intrinsic definition for the pointer argument.
unsigned AddrSpace = cast<PointerType>(Ptr->getType())->getAddressSpace();
PointerType *VecPtrTy = PointerType::get(II.getType(), AddrSpace);
Value *PtrCast = IC.Builder.CreateBitCast(Ptr, VecPtrTy, "castvec");
// The pass-through vector for an x86 masked load is a zero vector.
CallInst *NewMaskedLoad = IC.Builder.CreateMaskedLoad(
II.getType(), PtrCast, Align(1), BoolMask, ZeroVec);
return IC.replaceInstUsesWith(II, NewMaskedLoad);
}
return nullptr;
}
// TODO: If the x86 backend knew how to convert a bool vector mask back to an
// XMM register mask efficiently, we could transform all x86 masked intrinsics
// to LLVM masked intrinsics and remove the x86 masked intrinsic defs.
static bool simplifyX86MaskedStore(IntrinsicInst &II, InstCombiner &IC) {
Value *Ptr = II.getOperand(0);
Value *Mask = II.getOperand(1);
Value *Vec = II.getOperand(2);
// Zero Mask - this masked store instruction does nothing.
if (isa<ConstantAggregateZero>(Mask)) {
IC.eraseInstFromFunction(II);
return true;
}
// The SSE2 version is too weird (eg, unaligned but non-temporal) to do
// anything else at this level.
if (II.getIntrinsicID() == Intrinsic::x86_sse2_maskmov_dqu)
return false;
// The mask is constant or extended from a bool vector. Convert this x86
// intrinsic to the LLVM intrinsic to allow target-independent optimizations.
if (Value *BoolMask = getBoolVecFromMask(Mask)) {
unsigned AddrSpace = cast<PointerType>(Ptr->getType())->getAddressSpace();
PointerType *VecPtrTy = PointerType::get(Vec->getType(), AddrSpace);
Value *PtrCast = IC.Builder.CreateBitCast(Ptr, VecPtrTy, "castvec");
IC.Builder.CreateMaskedStore(Vec, PtrCast, Align(1), BoolMask);
// 'Replace uses' doesn't work for stores. Erase the original masked store.
IC.eraseInstFromFunction(II);
return true;
}
return false;
}
static Value *simplifyX86immShift(const IntrinsicInst &II,
InstCombiner::BuilderTy &Builder) {
bool LogicalShift = false;
bool ShiftLeft = false;
bool IsImm = false;
switch (II.getIntrinsicID()) {
default:
llvm_unreachable("Unexpected intrinsic!");
case Intrinsic::x86_sse2_psrai_d:
case Intrinsic::x86_sse2_psrai_w:
case Intrinsic::x86_avx2_psrai_d:
case Intrinsic::x86_avx2_psrai_w:
case Intrinsic::x86_avx512_psrai_q_128:
case Intrinsic::x86_avx512_psrai_q_256:
case Intrinsic::x86_avx512_psrai_d_512:
case Intrinsic::x86_avx512_psrai_q_512:
case Intrinsic::x86_avx512_psrai_w_512:
IsImm = true;
LLVM_FALLTHROUGH;
case Intrinsic::x86_sse2_psra_d:
case Intrinsic::x86_sse2_psra_w:
case Intrinsic::x86_avx2_psra_d:
case Intrinsic::x86_avx2_psra_w:
case Intrinsic::x86_avx512_psra_q_128:
case Intrinsic::x86_avx512_psra_q_256:
case Intrinsic::x86_avx512_psra_d_512:
case Intrinsic::x86_avx512_psra_q_512:
case Intrinsic::x86_avx512_psra_w_512:
LogicalShift = false;
ShiftLeft = false;
break;
case Intrinsic::x86_sse2_psrli_d:
case Intrinsic::x86_sse2_psrli_q:
case Intrinsic::x86_sse2_psrli_w:
case Intrinsic::x86_avx2_psrli_d:
case Intrinsic::x86_avx2_psrli_q:
case Intrinsic::x86_avx2_psrli_w:
case Intrinsic::x86_avx512_psrli_d_512:
case Intrinsic::x86_avx512_psrli_q_512:
case Intrinsic::x86_avx512_psrli_w_512:
IsImm = true;
LLVM_FALLTHROUGH;
case Intrinsic::x86_sse2_psrl_d:
case Intrinsic::x86_sse2_psrl_q:
case Intrinsic::x86_sse2_psrl_w:
case Intrinsic::x86_avx2_psrl_d:
case Intrinsic::x86_avx2_psrl_q:
case Intrinsic::x86_avx2_psrl_w:
case Intrinsic::x86_avx512_psrl_d_512:
case Intrinsic::x86_avx512_psrl_q_512:
case Intrinsic::x86_avx512_psrl_w_512:
LogicalShift = true;
ShiftLeft = false;
break;
case Intrinsic::x86_sse2_pslli_d:
case Intrinsic::x86_sse2_pslli_q:
case Intrinsic::x86_sse2_pslli_w:
case Intrinsic::x86_avx2_pslli_d:
case Intrinsic::x86_avx2_pslli_q:
case Intrinsic::x86_avx2_pslli_w:
case Intrinsic::x86_avx512_pslli_d_512:
case Intrinsic::x86_avx512_pslli_q_512:
case Intrinsic::x86_avx512_pslli_w_512:
IsImm = true;
LLVM_FALLTHROUGH;
case Intrinsic::x86_sse2_psll_d:
case Intrinsic::x86_sse2_psll_q:
case Intrinsic::x86_sse2_psll_w:
case Intrinsic::x86_avx2_psll_d:
case Intrinsic::x86_avx2_psll_q:
case Intrinsic::x86_avx2_psll_w:
case Intrinsic::x86_avx512_psll_d_512:
case Intrinsic::x86_avx512_psll_q_512:
case Intrinsic::x86_avx512_psll_w_512:
LogicalShift = true;
ShiftLeft = true;
break;
}
assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
Value *Vec = II.getArgOperand(0);
Value *Amt = II.getArgOperand(1);
auto *VT = cast<FixedVectorType>(Vec->getType());
Type *SVT = VT->getElementType();
Type *AmtVT = Amt->getType();
unsigned VWidth = VT->getNumElements();
unsigned BitWidth = SVT->getPrimitiveSizeInBits();
// If the shift amount is guaranteed to be in-range we can replace it with a
// generic shift. If its guaranteed to be out of range, logical shifts combine
// to zero and arithmetic shifts are clamped to (BitWidth - 1).
if (IsImm) {
assert(AmtVT->isIntegerTy(32) && "Unexpected shift-by-immediate type");
KnownBits KnownAmtBits =
llvm::computeKnownBits(Amt, II.getModule()->getDataLayout());
if (KnownAmtBits.getMaxValue().ult(BitWidth)) {
Amt = Builder.CreateZExtOrTrunc(Amt, SVT);
Amt = Builder.CreateVectorSplat(VWidth, Amt);
return (LogicalShift ? (ShiftLeft ? Builder.CreateShl(Vec, Amt)
: Builder.CreateLShr(Vec, Amt))
: Builder.CreateAShr(Vec, Amt));
}
if (KnownAmtBits.getMinValue().uge(BitWidth)) {
if (LogicalShift)
return ConstantAggregateZero::get(VT);
Amt = ConstantInt::get(SVT, BitWidth - 1);
return Builder.CreateAShr(Vec, Builder.CreateVectorSplat(VWidth, Amt));
}
} else {
// Ensure the first element has an in-range value and the rest of the
// elements in the bottom 64 bits are zero.
assert(AmtVT->isVectorTy() && AmtVT->getPrimitiveSizeInBits() == 128 &&
cast<VectorType>(AmtVT)->getElementType() == SVT &&
"Unexpected shift-by-scalar type");
unsigned NumAmtElts = cast<FixedVectorType>(AmtVT)->getNumElements();
APInt DemandedLower = APInt::getOneBitSet(NumAmtElts, 0);
APInt DemandedUpper = APInt::getBitsSet(NumAmtElts, 1, NumAmtElts / 2);
KnownBits KnownLowerBits = llvm::computeKnownBits(
Amt, DemandedLower, II.getModule()->getDataLayout());
KnownBits KnownUpperBits = llvm::computeKnownBits(
Amt, DemandedUpper, II.getModule()->getDataLayout());
if (KnownLowerBits.getMaxValue().ult(BitWidth) &&
(DemandedUpper.isZero() || KnownUpperBits.isZero())) {
SmallVector<int, 16> ZeroSplat(VWidth, 0);
Amt = Builder.CreateShuffleVector(Amt, ZeroSplat);
return (LogicalShift ? (ShiftLeft ? Builder.CreateShl(Vec, Amt)
: Builder.CreateLShr(Vec, Amt))
: Builder.CreateAShr(Vec, Amt));
}
}
// Simplify if count is constant vector.
auto *CDV = dyn_cast<ConstantDataVector>(Amt);
if (!CDV)
return nullptr;
// SSE2/AVX2 uses all the first 64-bits of the 128-bit vector
// operand to compute the shift amount.
assert(AmtVT->isVectorTy() && AmtVT->getPrimitiveSizeInBits() == 128 &&
cast<VectorType>(AmtVT)->getElementType() == SVT &&
"Unexpected shift-by-scalar type");
// Concatenate the sub-elements to create the 64-bit value.
APInt Count(64, 0);
for (unsigned i = 0, NumSubElts = 64 / BitWidth; i != NumSubElts; ++i) {
unsigned SubEltIdx = (NumSubElts - 1) - i;
auto *SubElt = cast<ConstantInt>(CDV->getElementAsConstant(SubEltIdx));
Count <<= BitWidth;
Count |= SubElt->getValue().zextOrTrunc(64);
}
// If shift-by-zero then just return the original value.
if (Count.isZero())
return Vec;
// Handle cases when Shift >= BitWidth.
if (Count.uge(BitWidth)) {
// If LogicalShift - just return zero.
if (LogicalShift)
return ConstantAggregateZero::get(VT);
// If ArithmeticShift - clamp Shift to (BitWidth - 1).
Count = APInt(64, BitWidth - 1);
}
// Get a constant vector of the same type as the first operand.
auto ShiftAmt = ConstantInt::get(SVT, Count.zextOrTrunc(BitWidth));
auto ShiftVec = Builder.CreateVectorSplat(VWidth, ShiftAmt);
if (ShiftLeft)
return Builder.CreateShl(Vec, ShiftVec);
if (LogicalShift)
return Builder.CreateLShr(Vec, ShiftVec);
return Builder.CreateAShr(Vec, ShiftVec);
}
// Attempt to simplify AVX2 per-element shift intrinsics to a generic IR shift.
// Unlike the generic IR shifts, the intrinsics have defined behaviour for out
// of range shift amounts (logical - set to zero, arithmetic - splat sign bit).
static Value *simplifyX86varShift(const IntrinsicInst &II,
InstCombiner::BuilderTy &Builder) {
bool LogicalShift = false;
bool ShiftLeft = false;
switch (II.getIntrinsicID()) {
default:
llvm_unreachable("Unexpected intrinsic!");
case Intrinsic::x86_avx2_psrav_d:
case Intrinsic::x86_avx2_psrav_d_256:
case Intrinsic::x86_avx512_psrav_q_128:
case Intrinsic::x86_avx512_psrav_q_256:
case Intrinsic::x86_avx512_psrav_d_512:
case Intrinsic::x86_avx512_psrav_q_512:
case Intrinsic::x86_avx512_psrav_w_128:
case Intrinsic::x86_avx512_psrav_w_256:
case Intrinsic::x86_avx512_psrav_w_512:
LogicalShift = false;
ShiftLeft = false;
break;
case Intrinsic::x86_avx2_psrlv_d:
case Intrinsic::x86_avx2_psrlv_d_256:
case Intrinsic::x86_avx2_psrlv_q:
case Intrinsic::x86_avx2_psrlv_q_256:
case Intrinsic::x86_avx512_psrlv_d_512:
case Intrinsic::x86_avx512_psrlv_q_512:
case Intrinsic::x86_avx512_psrlv_w_128:
case Intrinsic::x86_avx512_psrlv_w_256:
case Intrinsic::x86_avx512_psrlv_w_512:
LogicalShift = true;
ShiftLeft = false;
break;
case Intrinsic::x86_avx2_psllv_d:
case Intrinsic::x86_avx2_psllv_d_256:
case Intrinsic::x86_avx2_psllv_q:
case Intrinsic::x86_avx2_psllv_q_256:
case Intrinsic::x86_avx512_psllv_d_512:
case Intrinsic::x86_avx512_psllv_q_512:
case Intrinsic::x86_avx512_psllv_w_128:
case Intrinsic::x86_avx512_psllv_w_256:
case Intrinsic::x86_avx512_psllv_w_512:
LogicalShift = true;
ShiftLeft = true;
break;
}
assert((LogicalShift || !ShiftLeft) && "Only logical shifts can shift left");
Value *Vec = II.getArgOperand(0);
Value *Amt = II.getArgOperand(1);
auto *VT = cast<FixedVectorType>(II.getType());
Type *SVT = VT->getElementType();
int NumElts = VT->getNumElements();
int BitWidth = SVT->getIntegerBitWidth();
// If the shift amount is guaranteed to be in-range we can replace it with a
// generic shift.
APInt UpperBits =
APInt::getHighBitsSet(BitWidth, BitWidth - Log2_32(BitWidth));
if (llvm::MaskedValueIsZero(Amt, UpperBits,
II.getModule()->getDataLayout())) {
return (LogicalShift ? (ShiftLeft ? Builder.CreateShl(Vec, Amt)
: Builder.CreateLShr(Vec, Amt))
: Builder.CreateAShr(Vec, Amt));
}
// Simplify if all shift amounts are constant/undef.
auto *CShift = dyn_cast<Constant>(Amt);
if (!CShift)
return nullptr;
// Collect each element's shift amount.
// We also collect special cases: UNDEF = -1, OUT-OF-RANGE = BitWidth.
bool AnyOutOfRange = false;
SmallVector<int, 8> ShiftAmts;
for (int I = 0; I < NumElts; ++I) {
auto *CElt = CShift->getAggregateElement(I);
if (isa_and_nonnull<UndefValue>(CElt)) {
ShiftAmts.push_back(-1);
continue;
}
auto *COp = dyn_cast_or_null<ConstantInt>(CElt);
if (!COp)
return nullptr;
// Handle out of range shifts.
// If LogicalShift - set to BitWidth (special case).
// If ArithmeticShift - set to (BitWidth - 1) (sign splat).
APInt ShiftVal = COp->getValue();
if (ShiftVal.uge(BitWidth)) {
AnyOutOfRange = LogicalShift;
ShiftAmts.push_back(LogicalShift ? BitWidth : BitWidth - 1);
continue;
}
ShiftAmts.push_back((int)ShiftVal.getZExtValue());
}
// If all elements out of range or UNDEF, return vector of zeros/undefs.
// ArithmeticShift should only hit this if they are all UNDEF.
auto OutOfRange = [&](int Idx) { return (Idx < 0) || (BitWidth <= Idx); };
if (llvm::all_of(ShiftAmts, OutOfRange)) {
SmallVector<Constant *, 8> ConstantVec;
for (int Idx : ShiftAmts) {
if (Idx < 0) {
ConstantVec.push_back(UndefValue::get(SVT));
} else {
assert(LogicalShift && "Logical shift expected");
ConstantVec.push_back(ConstantInt::getNullValue(SVT));
}
}
return ConstantVector::get(ConstantVec);
}
// We can't handle only some out of range values with generic logical shifts.
if (AnyOutOfRange)
return nullptr;
// Build the shift amount constant vector.
SmallVector<Constant *, 8> ShiftVecAmts;
for (int Idx : ShiftAmts) {
if (Idx < 0)
ShiftVecAmts.push_back(UndefValue::get(SVT));
else
ShiftVecAmts.push_back(ConstantInt::get(SVT, Idx));
}
auto ShiftVec = ConstantVector::get(ShiftVecAmts);
if (ShiftLeft)
return Builder.CreateShl(Vec, ShiftVec);
if (LogicalShift)
return Builder.CreateLShr(Vec, ShiftVec);
return Builder.CreateAShr(Vec, ShiftVec);
}
static Value *simplifyX86pack(IntrinsicInst &II,
InstCombiner::BuilderTy &Builder, bool IsSigned) {
Value *Arg0 = II.getArgOperand(0);
Value *Arg1 = II.getArgOperand(1);
Type *ResTy = II.getType();
// Fast all undef handling.
if (isa<UndefValue>(Arg0) && isa<UndefValue>(Arg1))
return UndefValue::get(ResTy);
auto *ArgTy = cast<FixedVectorType>(Arg0->getType());
unsigned NumLanes = ResTy->getPrimitiveSizeInBits() / 128;
unsigned NumSrcElts = ArgTy->getNumElements();
assert(cast<FixedVectorType>(ResTy)->getNumElements() == (2 * NumSrcElts) &&
"Unexpected packing types");
unsigned NumSrcEltsPerLane = NumSrcElts / NumLanes;
unsigned DstScalarSizeInBits = ResTy->getScalarSizeInBits();
unsigned SrcScalarSizeInBits = ArgTy->getScalarSizeInBits();
assert(SrcScalarSizeInBits == (2 * DstScalarSizeInBits) &&
"Unexpected packing types");
// Constant folding.
if (!isa<Constant>(Arg0) || !isa<Constant>(Arg1))
return nullptr;
// Clamp Values - signed/unsigned both use signed clamp values, but they
// differ on the min/max values.
APInt MinValue, MaxValue;
if (IsSigned) {
// PACKSS: Truncate signed value with signed saturation.
// Source values less than dst minint are saturated to minint.
// Source values greater than dst maxint are saturated to maxint.
MinValue =
APInt::getSignedMinValue(DstScalarSizeInBits).sext(SrcScalarSizeInBits);
MaxValue =
APInt::getSignedMaxValue(DstScalarSizeInBits).sext(SrcScalarSizeInBits);
} else {
// PACKUS: Truncate signed value with unsigned saturation.
// Source values less than zero are saturated to zero.
// Source values greater than dst maxuint are saturated to maxuint.
MinValue = APInt::getZero(SrcScalarSizeInBits);
MaxValue = APInt::getLowBitsSet(SrcScalarSizeInBits, DstScalarSizeInBits);
}
auto *MinC = Constant::getIntegerValue(ArgTy, MinValue);
auto *MaxC = Constant::getIntegerValue(ArgTy, MaxValue);
Arg0 = Builder.CreateSelect(Builder.CreateICmpSLT(Arg0, MinC), MinC, Arg0);
Arg1 = Builder.CreateSelect(Builder.CreateICmpSLT(Arg1, MinC), MinC, Arg1);
Arg0 = Builder.CreateSelect(Builder.CreateICmpSGT(Arg0, MaxC), MaxC, Arg0);
Arg1 = Builder.CreateSelect(Builder.CreateICmpSGT(Arg1, MaxC), MaxC, Arg1);
// Shuffle clamped args together at the lane level.
SmallVector<int, 32> PackMask;
for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
for (unsigned Elt = 0; Elt != NumSrcEltsPerLane; ++Elt)
PackMask.push_back(Elt + (Lane * NumSrcEltsPerLane));
for (unsigned Elt = 0; Elt != NumSrcEltsPerLane; ++Elt)
PackMask.push_back(Elt + (Lane * NumSrcEltsPerLane) + NumSrcElts);
}
auto *Shuffle = Builder.CreateShuffleVector(Arg0, Arg1, PackMask);
// Truncate to dst size.
return Builder.CreateTrunc(Shuffle, ResTy);
}
static Value *simplifyX86movmsk(const IntrinsicInst &II,
InstCombiner::BuilderTy &Builder) {
Value *Arg = II.getArgOperand(0);
Type *ResTy = II.getType();
// movmsk(undef) -> zero as we must ensure the upper bits are zero.
if (isa<UndefValue>(Arg))
return Constant::getNullValue(ResTy);
auto *ArgTy = dyn_cast<FixedVectorType>(Arg->getType());
// We can't easily peek through x86_mmx types.
if (!ArgTy)
return nullptr;
// Expand MOVMSK to compare/bitcast/zext:
// e.g. PMOVMSKB(v16i8 x):
// %cmp = icmp slt <16 x i8> %x, zeroinitializer
// %int = bitcast <16 x i1> %cmp to i16
// %res = zext i16 %int to i32
unsigned NumElts = ArgTy->getNumElements();
Type *IntegerVecTy = VectorType::getInteger(ArgTy);
Type *IntegerTy = Builder.getIntNTy(NumElts);
Value *Res = Builder.CreateBitCast(Arg, IntegerVecTy);
Res = Builder.CreateICmpSLT(Res, Constant::getNullValue(IntegerVecTy));
Res = Builder.CreateBitCast(Res, IntegerTy);
Res = Builder.CreateZExtOrTrunc(Res, ResTy);
return Res;
}
static Value *simplifyX86addcarry(const IntrinsicInst &II,
InstCombiner::BuilderTy &Builder) {
Value *CarryIn = II.getArgOperand(0);
Value *Op1 = II.getArgOperand(1);
Value *Op2 = II.getArgOperand(2);
Type *RetTy = II.getType();
Type *OpTy = Op1->getType();
assert(RetTy->getStructElementType(0)->isIntegerTy(8) &&
RetTy->getStructElementType(1) == OpTy && OpTy == Op2->getType() &&
"Unexpected types for x86 addcarry");
// If carry-in is zero, this is just an unsigned add with overflow.
if (match(CarryIn, PatternMatch::m_ZeroInt())) {
Value *UAdd = Builder.CreateIntrinsic(Intrinsic::uadd_with_overflow, OpTy,
{Op1, Op2});
// The types have to be adjusted to match the x86 call types.
Value *UAddResult = Builder.CreateExtractValue(UAdd, 0);
Value *UAddOV = Builder.CreateZExt(Builder.CreateExtractValue(UAdd, 1),
Builder.getInt8Ty());
Value *Res = UndefValue::get(RetTy);
Res = Builder.CreateInsertValue(Res, UAddOV, 0);
return Builder.CreateInsertValue(Res, UAddResult, 1);
}
return nullptr;
}
static Value *simplifyX86insertps(const IntrinsicInst &II,
InstCombiner::BuilderTy &Builder) {
auto *CInt = dyn_cast<ConstantInt>(II.getArgOperand(2));
if (!CInt)
return nullptr;
auto *VecTy = cast<FixedVectorType>(II.getType());
assert(VecTy->getNumElements() == 4 && "insertps with wrong vector type");
// The immediate permute control byte looks like this:
// [3:0] - zero mask for each 32-bit lane
// [5:4] - select one 32-bit destination lane
// [7:6] - select one 32-bit source lane
uint8_t Imm = CInt->getZExtValue();
uint8_t ZMask = Imm & 0xf;
uint8_t DestLane = (Imm >> 4) & 0x3;
uint8_t SourceLane = (Imm >> 6) & 0x3;
ConstantAggregateZero *ZeroVector = ConstantAggregateZero::get(VecTy);
// If all zero mask bits are set, this was just a weird way to
// generate a zero vector.
if (ZMask == 0xf)
return ZeroVector;
// Initialize by passing all of the first source bits through.
int ShuffleMask[4] = {0, 1, 2, 3};
// We may replace the second operand with the zero vector.
Value *V1 = II.getArgOperand(1);
if (ZMask) {
// If the zero mask is being used with a single input or the zero mask
// overrides the destination lane, this is a shuffle with the zero vector.
if ((II.getArgOperand(0) == II.getArgOperand(1)) ||
(ZMask & (1 << DestLane))) {
V1 = ZeroVector;
// We may still move 32-bits of the first source vector from one lane
// to another.
ShuffleMask[DestLane] = SourceLane;
// The zero mask may override the previous insert operation.
for (unsigned i = 0; i < 4; ++i)
if ((ZMask >> i) & 0x1)
ShuffleMask[i] = i + 4;
} else {
// TODO: Model this case as 2 shuffles or a 'logical and' plus shuffle?
return nullptr;
}
} else {
// Replace the selected destination lane with the selected source lane.
ShuffleMask[DestLane] = SourceLane + 4;
}
return Builder.CreateShuffleVector(II.getArgOperand(0), V1, ShuffleMask);
}
/// Attempt to simplify SSE4A EXTRQ/EXTRQI instructions using constant folding
/// or conversion to a shuffle vector.
static Value *simplifyX86extrq(IntrinsicInst &II, Value *Op0,
ConstantInt *CILength, ConstantInt *CIIndex,
InstCombiner::BuilderTy &Builder) {
auto LowConstantHighUndef = [&](uint64_t Val) {
Type *IntTy64 = Type::getInt64Ty(II.getContext());
Constant *Args[] = {ConstantInt::get(IntTy64, Val),
UndefValue::get(IntTy64)};
return ConstantVector::get(Args);
};
// See if we're dealing with constant values.
auto *C0 = dyn_cast<Constant>(Op0);
auto *CI0 =
C0 ? dyn_cast_or_null<ConstantInt>(C0->getAggregateElement((unsigned)0))
: nullptr;
// Attempt to constant fold.
if (CILength && CIIndex) {
// From AMD documentation: "The bit index and field length are each six
// bits in length other bits of the field are ignored."
APInt APIndex = CIIndex->getValue().zextOrTrunc(6);
APInt APLength = CILength->getValue().zextOrTrunc(6);
unsigned Index = APIndex.getZExtValue();
// From AMD documentation: "a value of zero in the field length is
// defined as length of 64".
unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue();
// From AMD documentation: "If the sum of the bit index + length field
// is greater than 64, the results are undefined".
unsigned End = Index + Length;
// Note that both field index and field length are 8-bit quantities.
// Since variables 'Index' and 'Length' are unsigned values
// obtained from zero-extending field index and field length
// respectively, their sum should never wrap around.
if (End > 64)
return UndefValue::get(II.getType());
// If we are inserting whole bytes, we can convert this to a shuffle.
// Lowering can recognize EXTRQI shuffle masks.
if ((Length % 8) == 0 && (Index % 8) == 0) {
// Convert bit indices to byte indices.
Length /= 8;
Index /= 8;
Type *IntTy8 = Type::getInt8Ty(II.getContext());
auto *ShufTy = FixedVectorType::get(IntTy8, 16);
SmallVector<int, 16> ShuffleMask;
for (int i = 0; i != (int)Length; ++i)
ShuffleMask.push_back(i + Index);
for (int i = Length; i != 8; ++i)
ShuffleMask.push_back(i + 16);
for (int i = 8; i != 16; ++i)
ShuffleMask.push_back(-1);
Value *SV = Builder.CreateShuffleVector(
Builder.CreateBitCast(Op0, ShufTy),
ConstantAggregateZero::get(ShufTy), ShuffleMask);
return Builder.CreateBitCast(SV, II.getType());
}
// Constant Fold - shift Index'th bit to lowest position and mask off
// Length bits.
if (CI0) {
APInt Elt = CI0->getValue();
Elt.lshrInPlace(Index);
Elt = Elt.zextOrTrunc(Length);
return LowConstantHighUndef(Elt.getZExtValue());
}
// If we were an EXTRQ call, we'll save registers if we convert to EXTRQI.
if (II.getIntrinsicID() == Intrinsic::x86_sse4a_extrq) {
Value *Args[] = {Op0, CILength, CIIndex};
Module *M = II.getModule();
Function *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_extrqi);
return Builder.CreateCall(F, Args);
}
}
// Constant Fold - extraction from zero is always {zero, undef}.
if (CI0 && CI0->isZero())
return LowConstantHighUndef(0);
return nullptr;
}
/// Attempt to simplify SSE4A INSERTQ/INSERTQI instructions using constant
/// folding or conversion to a shuffle vector.
static Value *simplifyX86insertq(IntrinsicInst &II, Value *Op0, Value *Op1,
APInt APLength, APInt APIndex,
InstCombiner::BuilderTy &Builder) {
// From AMD documentation: "The bit index and field length are each six bits
// in length other bits of the field are ignored."
APIndex = APIndex.zextOrTrunc(6);
APLength = APLength.zextOrTrunc(6);
// Attempt to constant fold.
unsigned Index = APIndex.getZExtValue();
// From AMD documentation: "a value of zero in the field length is
// defined as length of 64".
unsigned Length = APLength == 0 ? 64 : APLength.getZExtValue();
// From AMD documentation: "If the sum of the bit index + length field
// is greater than 64, the results are undefined".
unsigned End = Index + Length;
// Note that both field index and field length are 8-bit quantities.
// Since variables 'Index' and 'Length' are unsigned values
// obtained from zero-extending field index and field length
// respectively, their sum should never wrap around.
if (End > 64)
return UndefValue::get(II.getType());
// If we are inserting whole bytes, we can convert this to a shuffle.
// Lowering can recognize INSERTQI shuffle masks.
if ((Length % 8) == 0 && (Index % 8) == 0) {
// Convert bit indices to byte indices.
Length /= 8;
Index /= 8;
Type *IntTy8 = Type::getInt8Ty(II.getContext());
auto *ShufTy = FixedVectorType::get(IntTy8, 16);
SmallVector<int, 16> ShuffleMask;
for (int i = 0; i != (int)Index; ++i)
ShuffleMask.push_back(i);
for (int i = 0; i != (int)Length; ++i)
ShuffleMask.push_back(i + 16);
for (int i = Index + Length; i != 8; ++i)
ShuffleMask.push_back(i);
for (int i = 8; i != 16; ++i)
ShuffleMask.push_back(-1);
Value *SV = Builder.CreateShuffleVector(Builder.CreateBitCast(Op0, ShufTy),
Builder.CreateBitCast(Op1, ShufTy),
ShuffleMask);
return Builder.CreateBitCast(SV, II.getType());
}
// See if we're dealing with constant values.
auto *C0 = dyn_cast<Constant>(Op0);
auto *C1 = dyn_cast<Constant>(Op1);
auto *CI00 =
C0 ? dyn_cast_or_null<ConstantInt>(C0->getAggregateElement((unsigned)0))
: nullptr;
auto *CI10 =
C1 ? dyn_cast_or_null<ConstantInt>(C1->getAggregateElement((unsigned)0))
: nullptr;
// Constant Fold - insert bottom Length bits starting at the Index'th bit.
if (CI00 && CI10) {
APInt V00 = CI00->getValue();
APInt V10 = CI10->getValue();
APInt Mask = APInt::getLowBitsSet(64, Length).shl(Index);
V00 = V00 & ~Mask;
V10 = V10.zextOrTrunc(Length).zextOrTrunc(64).shl(Index);
APInt Val = V00 | V10;
Type *IntTy64 = Type::getInt64Ty(II.getContext());
Constant *Args[] = {ConstantInt::get(IntTy64, Val.getZExtValue()),
UndefValue::get(IntTy64)};
return ConstantVector::get(Args);
}
// If we were an INSERTQ call, we'll save demanded elements if we convert to
// INSERTQI.
if (II.getIntrinsicID() == Intrinsic::x86_sse4a_insertq) {
Type *IntTy8 = Type::getInt8Ty(II.getContext());
Constant *CILength = ConstantInt::get(IntTy8, Length, false);
Constant *CIIndex = ConstantInt::get(IntTy8, Index, false);
Value *Args[] = {Op0, Op1, CILength, CIIndex};
Module *M = II.getModule();
Function *F = Intrinsic::getDeclaration(M, Intrinsic::x86_sse4a_insertqi);
return Builder.CreateCall(F, Args);
}
return nullptr;
}
/// Attempt to convert pshufb* to shufflevector if the mask is constant.
static Value *simplifyX86pshufb(const IntrinsicInst &II,
InstCombiner::BuilderTy &Builder) {
auto *V = dyn_cast<Constant>(II.getArgOperand(1));
if (!V)
return nullptr;
auto *VecTy = cast<FixedVectorType>(II.getType());
unsigned NumElts = VecTy->getNumElements();
assert((NumElts == 16 || NumElts == 32 || NumElts == 64) &&
"Unexpected number of elements in shuffle mask!");
// Construct a shuffle mask from constant integers or UNDEFs.
int Indexes[64];
// Each byte in the shuffle control mask forms an index to permute the
// corresponding byte in the destination operand.
for (unsigned I = 0; I < NumElts; ++I) {
Constant *COp = V->getAggregateElement(I);
if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
return nullptr;
if (isa<UndefValue>(COp)) {
Indexes[I] = -1;
continue;
}
int8_t Index = cast<ConstantInt>(COp)->getValue().getZExtValue();
// If the most significant bit (bit[7]) of each byte of the shuffle
// control mask is set, then zero is written in the result byte.
// The zero vector is in the right-hand side of the resulting
// shufflevector.
// The value of each index for the high 128-bit lane is the least
// significant 4 bits of the respective shuffle control byte.
Index = ((Index < 0) ? NumElts : Index & 0x0F) + (I & 0xF0);
Indexes[I] = Index;
}
auto V1 = II.getArgOperand(0);
auto V2 = Constant::getNullValue(VecTy);
return Builder.CreateShuffleVector(V1, V2, makeArrayRef(Indexes, NumElts));
}
/// Attempt to convert vpermilvar* to shufflevector if the mask is constant.
static Value *simplifyX86vpermilvar(const IntrinsicInst &II,
InstCombiner::BuilderTy &Builder) {
auto *V = dyn_cast<Constant>(II.getArgOperand(1));
if (!V)
return nullptr;
auto *VecTy = cast<FixedVectorType>(II.getType());
unsigned NumElts = VecTy->getNumElements();
bool IsPD = VecTy->getScalarType()->isDoubleTy();
unsigned NumLaneElts = IsPD ? 2 : 4;
assert(NumElts == 16 || NumElts == 8 || NumElts == 4 || NumElts == 2);
// Construct a shuffle mask from constant integers or UNDEFs.
int Indexes[16];
// The intrinsics only read one or two bits, clear the rest.
for (unsigned I = 0; I < NumElts; ++I) {
Constant *COp = V->getAggregateElement(I);
if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
return nullptr;
if (isa<UndefValue>(COp)) {
Indexes[I] = -1;
continue;
}
APInt Index = cast<ConstantInt>(COp)->getValue();
Index = Index.zextOrTrunc(32).getLoBits(2);
// The PD variants uses bit 1 to select per-lane element index, so
// shift down to convert to generic shuffle mask index.
if (IsPD)
Index.lshrInPlace(1);
// The _256 variants are a bit trickier since the mask bits always index
// into the corresponding 128 half. In order to convert to a generic
// shuffle, we have to make that explicit.
Index += APInt(32, (I / NumLaneElts) * NumLaneElts);
Indexes[I] = Index.getZExtValue();
}
auto V1 = II.getArgOperand(0);
return Builder.CreateShuffleVector(V1, makeArrayRef(Indexes, NumElts));
}
/// Attempt to convert vpermd/vpermps to shufflevector if the mask is constant.
static Value *simplifyX86vpermv(const IntrinsicInst &II,
InstCombiner::BuilderTy &Builder) {
auto *V = dyn_cast<Constant>(II.getArgOperand(1));
if (!V)
return nullptr;
auto *VecTy = cast<FixedVectorType>(II.getType());
unsigned Size = VecTy->getNumElements();
assert((Size == 4 || Size == 8 || Size == 16 || Size == 32 || Size == 64) &&
"Unexpected shuffle mask size");
// Construct a shuffle mask from constant integers or UNDEFs.
int Indexes[64];
for (unsigned I = 0; I < Size; ++I) {
Constant *COp = V->getAggregateElement(I);
if (!COp || (!isa<UndefValue>(COp) && !isa<ConstantInt>(COp)))
return nullptr;
if (isa<UndefValue>(COp)) {
Indexes[I] = -1;
continue;
}
uint32_t Index = cast<ConstantInt>(COp)->getZExtValue();
Index &= Size - 1;
Indexes[I] = Index;
}
auto V1 = II.getArgOperand(0);
return Builder.CreateShuffleVector(V1, makeArrayRef(Indexes, Size));
}
Optional<Instruction *>
X86TTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
auto SimplifyDemandedVectorEltsLow = [&IC](Value *Op, unsigned Width,
unsigned DemandedWidth) {
APInt UndefElts(Width, 0);
APInt DemandedElts = APInt::getLowBitsSet(Width, DemandedWidth);
return IC.SimplifyDemandedVectorElts(Op, DemandedElts, UndefElts);
};
Intrinsic::ID IID = II.getIntrinsicID();
switch (IID) {
case Intrinsic::x86_bmi_bextr_32:
case Intrinsic::x86_bmi_bextr_64:
case Intrinsic::x86_tbm_bextri_u32:
case Intrinsic::x86_tbm_bextri_u64:
// If the RHS is a constant we can try some simplifications.
if (auto *C = dyn_cast<ConstantInt>(II.getArgOperand(1))) {
uint64_t Shift = C->getZExtValue();
uint64_t Length = (Shift >> 8) & 0xff;
Shift &= 0xff;
unsigned BitWidth = II.getType()->getIntegerBitWidth();
// If the length is 0 or the shift is out of range, replace with zero.
if (Length == 0 || Shift >= BitWidth) {
return IC.replaceInstUsesWith(II, ConstantInt::get(II.getType(), 0));
}
// If the LHS is also a constant, we can completely constant fold this.
if (auto *InC = dyn_cast<ConstantInt>(II.getArgOperand(0))) {
uint64_t Result = InC->getZExtValue() >> Shift;
if (Length > BitWidth)
Length = BitWidth;
Result &= maskTrailingOnes<uint64_t>(Length);
return IC.replaceInstUsesWith(II,
ConstantInt::get(II.getType(), Result));
}
// TODO should we turn this into 'and' if shift is 0? Or 'shl' if we
// are only masking bits that a shift already cleared?
}
break;
case Intrinsic::x86_bmi_bzhi_32:
case Intrinsic::x86_bmi_bzhi_64:
// If the RHS is a constant we can try some simplifications.
if (auto *C = dyn_cast<ConstantInt>(II.getArgOperand(1))) {
uint64_t Index = C->getZExtValue() & 0xff;
unsigned BitWidth = II.getType()->getIntegerBitWidth();
if (Index >= BitWidth) {
return IC.replaceInstUsesWith(II, II.getArgOperand(0));
}
if (Index == 0) {
return IC.replaceInstUsesWith(II, ConstantInt::get(II.getType(), 0));
}
// If the LHS is also a constant, we can completely constant fold this.
if (auto *InC = dyn_cast<ConstantInt>(II.getArgOperand(0))) {
uint64_t Result = InC->getZExtValue();
Result &= maskTrailingOnes<uint64_t>(Index);
return IC.replaceInstUsesWith(II,
ConstantInt::get(II.getType(), Result));
}
// TODO should we convert this to an AND if the RHS is constant?
}
break;
case Intrinsic::x86_bmi_pext_32:
case Intrinsic::x86_bmi_pext_64:
if (auto *MaskC = dyn_cast<ConstantInt>(II.getArgOperand(1))) {
if (MaskC->isNullValue()) {
return IC.replaceInstUsesWith(II, ConstantInt::get(II.getType(), 0));
}
if (MaskC->isAllOnesValue()) {
return IC.replaceInstUsesWith(II, II.getArgOperand(0));
}
if (MaskC->getValue().isShiftedMask()) {