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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
|
//===- EmulateWideInt.cpp - Wide integer operation emulation ----*- C++ -*-===//
//
// 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
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Arith/Transforms/Passes.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Arith/Transforms/WideIntEmulationConverter.h"
#include "mlir/Dialect/Arith/Utils/Utils.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"
#include "mlir/Dialect/Func/Transforms/FuncConversions.h"
#include "mlir/Dialect/Vector/IR/VectorOps.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/TypeUtilities.h"
#include "mlir/Support/LogicalResult.h"
#include "mlir/Transforms/DialectConversion.h"
#include "llvm/ADT/APInt.h"
#include "llvm/Support/FormatVariadic.h"
#include "llvm/Support/MathExtras.h"
#include <cassert>
namespace mlir::arith {
#define GEN_PASS_DEF_ARITHEMULATEWIDEINT
#include "mlir/Dialect/Arith/Transforms/Passes.h.inc"
} // namespace mlir::arith
using namespace mlir;
//===----------------------------------------------------------------------===//
// Common Helper Functions
//===----------------------------------------------------------------------===//
/// Returns N bottom and N top bits from `value`, where N = `newBitWidth`.
/// Treats `value` as a 2*N bits-wide integer.
/// The bottom bits are returned in the first pair element, while the top bits
/// in the second one.
static std::pair<APInt, APInt> getHalves(const APInt &value,
unsigned newBitWidth) {
APInt low = value.extractBits(newBitWidth, 0);
APInt high = value.extractBits(newBitWidth, newBitWidth);
return {std::move(low), std::move(high)};
}
/// Returns the type with the last (innermost) dimension reduced to x1.
/// Scalarizes 1D vector inputs to match how we extract/insert vector values,
/// e.g.:
/// - vector<3x2xi16> --> vector<3x1xi16>
/// - vector<2xi16> --> i16
static Type reduceInnermostDim(VectorType type) {
if (type.getShape().size() == 1)
return type.getElementType();
auto newShape = to_vector(type.getShape());
newShape.back() = 1;
return VectorType::get(newShape, type.getElementType());
}
/// Extracts the `input` vector slice with elements at the last dimension offset
/// by `lastOffset`. Returns a value of vector type with the last dimension
/// reduced to x1 or fully scalarized, e.g.:
/// - vector<3x2xi16> --> vector<3x1xi16>
/// - vector<2xi16> --> i16
static Value extractLastDimSlice(ConversionPatternRewriter &rewriter,
Location loc, Value input,
int64_t lastOffset) {
ArrayRef<int64_t> shape = cast<VectorType>(input.getType()).getShape();
assert(lastOffset < shape.back() && "Offset out of bounds");
// Scalarize the result in case of 1D vectors.
if (shape.size() == 1)
return rewriter.create<vector::ExtractOp>(loc, input, lastOffset);
SmallVector<int64_t> offsets(shape.size(), 0);
offsets.back() = lastOffset;
auto sizes = llvm::to_vector(shape);
sizes.back() = 1;
SmallVector<int64_t> strides(shape.size(), 1);
return rewriter.create<vector::ExtractStridedSliceOp>(loc, input, offsets,
sizes, strides);
}
/// Extracts two vector slices from the `input` whose type is `vector<...x2T>`,
/// with the first element at offset 0 and the second element at offset 1.
static std::pair<Value, Value>
extractLastDimHalves(ConversionPatternRewriter &rewriter, Location loc,
Value input) {
return {extractLastDimSlice(rewriter, loc, input, 0),
extractLastDimSlice(rewriter, loc, input, 1)};
}
// Performs a vector shape cast to drop the trailing x1 dimension. If the
// `input` is a scalar, this is a noop.
static Value dropTrailingX1Dim(ConversionPatternRewriter &rewriter,
Location loc, Value input) {
auto vecTy = dyn_cast<VectorType>(input.getType());
if (!vecTy)
return input;
// Shape cast to drop the last x1 dimension.
ArrayRef<int64_t> shape = vecTy.getShape();
assert(shape.size() >= 2 && "Expected vector with at list two dims");
assert(shape.back() == 1 && "Expected the last vector dim to be x1");
auto newVecTy = VectorType::get(shape.drop_back(), vecTy.getElementType());
return rewriter.create<vector::ShapeCastOp>(loc, newVecTy, input);
}
/// Performs a vector shape cast to append an x1 dimension. If the
/// `input` is a scalar, this is a noop.
static Value appendX1Dim(ConversionPatternRewriter &rewriter, Location loc,
Value input) {
auto vecTy = dyn_cast<VectorType>(input.getType());
if (!vecTy)
return input;
// Add a trailing x1 dim.
auto newShape = llvm::to_vector(vecTy.getShape());
newShape.push_back(1);
auto newTy = VectorType::get(newShape, vecTy.getElementType());
return rewriter.create<vector::ShapeCastOp>(loc, newTy, input);
}
/// Inserts the `source` vector slice into the `dest` vector at offset
/// `lastOffset` in the last dimension. `source` can be a scalar when `dest` is
/// a 1D vector.
static Value insertLastDimSlice(ConversionPatternRewriter &rewriter,
Location loc, Value source, Value dest,
int64_t lastOffset) {
ArrayRef<int64_t> shape = cast<VectorType>(dest.getType()).getShape();
assert(lastOffset < shape.back() && "Offset out of bounds");
// Handle scalar source.
if (isa<IntegerType>(source.getType()))
return rewriter.create<vector::InsertOp>(loc, source, dest, lastOffset);
SmallVector<int64_t> offsets(shape.size(), 0);
offsets.back() = lastOffset;
SmallVector<int64_t> strides(shape.size(), 1);
return rewriter.create<vector::InsertStridedSliceOp>(loc, source, dest,
offsets, strides);
}
/// Constructs a new vector of type `resultType` by creating a series of
/// insertions of `resultComponents`, each at the next offset of the last vector
/// dimension.
/// When all `resultComponents` are scalars, the result type is `vector<NxT>`;
/// when `resultComponents` are `vector<...x1xT>`s, the result type is
/// `vector<...xNxT>`, where `N` is the number of `resultComponents`.
static Value constructResultVector(ConversionPatternRewriter &rewriter,
Location loc, VectorType resultType,
ValueRange resultComponents) {
llvm::ArrayRef<int64_t> resultShape = resultType.getShape();
(void)resultShape;
assert(!resultShape.empty() && "Result expected to have dimensions");
assert(resultShape.back() == static_cast<int64_t>(resultComponents.size()) &&
"Wrong number of result components");
Value resultVec = createScalarOrSplatConstant(rewriter, loc, resultType, 0);
for (auto [i, component] : llvm::enumerate(resultComponents))
resultVec = insertLastDimSlice(rewriter, loc, component, resultVec, i);
return resultVec;
}
namespace {
//===----------------------------------------------------------------------===//
// ConvertConstant
//===----------------------------------------------------------------------===//
struct ConvertConstant final : OpConversionPattern<arith::ConstantOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::ConstantOp op, OpAdaptor,
ConversionPatternRewriter &rewriter) const override {
Type oldType = op.getType();
auto newType = getTypeConverter()->convertType<VectorType>(oldType);
if (!newType)
return rewriter.notifyMatchFailure(
op, llvm::formatv("unsupported type: {0}", op.getType()));
unsigned newBitWidth = newType.getElementTypeBitWidth();
Attribute oldValue = op.getValueAttr();
if (auto intAttr = dyn_cast<IntegerAttr>(oldValue)) {
auto [low, high] = getHalves(intAttr.getValue(), newBitWidth);
auto newAttr = DenseElementsAttr::get(newType, {low, high});
rewriter.replaceOpWithNewOp<arith::ConstantOp>(op, newAttr);
return success();
}
if (auto splatAttr = dyn_cast<SplatElementsAttr>(oldValue)) {
auto [low, high] =
getHalves(splatAttr.getSplatValue<APInt>(), newBitWidth);
int64_t numSplatElems = splatAttr.getNumElements();
SmallVector<APInt> values;
values.reserve(numSplatElems * 2);
for (int64_t i = 0; i < numSplatElems; ++i) {
values.push_back(low);
values.push_back(high);
}
auto attr = DenseElementsAttr::get(newType, values);
rewriter.replaceOpWithNewOp<arith::ConstantOp>(op, attr);
return success();
}
if (auto elemsAttr = dyn_cast<DenseElementsAttr>(oldValue)) {
int64_t numElems = elemsAttr.getNumElements();
SmallVector<APInt> values;
values.reserve(numElems * 2);
for (const APInt &origVal : elemsAttr.getValues<APInt>()) {
auto [low, high] = getHalves(origVal, newBitWidth);
values.push_back(std::move(low));
values.push_back(std::move(high));
}
auto attr = DenseElementsAttr::get(newType, values);
rewriter.replaceOpWithNewOp<arith::ConstantOp>(op, attr);
return success();
}
return rewriter.notifyMatchFailure(op.getLoc(),
"unhandled constant attribute");
}
};
//===----------------------------------------------------------------------===//
// ConvertAddI
//===----------------------------------------------------------------------===//
struct ConvertAddI final : OpConversionPattern<arith::AddIOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::AddIOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
auto newTy = getTypeConverter()->convertType<VectorType>(op.getType());
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
Type newElemTy = reduceInnermostDim(newTy);
auto [lhsElem0, lhsElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
auto [rhsElem0, rhsElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getRhs());
auto lowSum =
rewriter.create<arith::AddUIExtendedOp>(loc, lhsElem0, rhsElem0);
Value overflowVal =
rewriter.create<arith::ExtUIOp>(loc, newElemTy, lowSum.getOverflow());
Value high0 = rewriter.create<arith::AddIOp>(loc, overflowVal, lhsElem1);
Value high = rewriter.create<arith::AddIOp>(loc, high0, rhsElem1);
Value resultVec =
constructResultVector(rewriter, loc, newTy, {lowSum.getSum(), high});
rewriter.replaceOp(op, resultVec);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertBitwiseBinary
//===----------------------------------------------------------------------===//
/// Conversion pattern template for bitwise binary ops, e.g., `arith.andi`.
template <typename BinaryOp>
struct ConvertBitwiseBinary final : OpConversionPattern<BinaryOp> {
using OpConversionPattern<BinaryOp>::OpConversionPattern;
using OpAdaptor = typename OpConversionPattern<BinaryOp>::OpAdaptor;
LogicalResult
matchAndRewrite(BinaryOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
auto newTy = this->getTypeConverter()->template convertType<VectorType>(
op.getType());
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
auto [lhsElem0, lhsElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
auto [rhsElem0, rhsElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getRhs());
Value resElem0 = rewriter.create<BinaryOp>(loc, lhsElem0, rhsElem0);
Value resElem1 = rewriter.create<BinaryOp>(loc, lhsElem1, rhsElem1);
Value resultVec =
constructResultVector(rewriter, loc, newTy, {resElem0, resElem1});
rewriter.replaceOp(op, resultVec);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertCmpI
//===----------------------------------------------------------------------===//
/// Returns the matching unsigned version of the given predicate `pred`, or the
/// same predicate if `pred` is not a signed.
static arith::CmpIPredicate toUnsignedPredicate(arith::CmpIPredicate pred) {
using P = arith::CmpIPredicate;
switch (pred) {
case P::sge:
return P::uge;
case P::sgt:
return P::ugt;
case P::sle:
return P::ule;
case P::slt:
return P::ult;
default:
return pred;
}
}
struct ConvertCmpI final : OpConversionPattern<arith::CmpIOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::CmpIOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
auto inputTy =
getTypeConverter()->convertType<VectorType>(op.getLhs().getType());
if (!inputTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
arith::CmpIPredicate highPred = adaptor.getPredicate();
arith::CmpIPredicate lowPred = toUnsignedPredicate(highPred);
auto [lhsElem0, lhsElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
auto [rhsElem0, rhsElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getRhs());
Value lowCmp =
rewriter.create<arith::CmpIOp>(loc, lowPred, lhsElem0, rhsElem0);
Value highCmp =
rewriter.create<arith::CmpIOp>(loc, highPred, lhsElem1, rhsElem1);
Value cmpResult{};
switch (highPred) {
case arith::CmpIPredicate::eq: {
cmpResult = rewriter.create<arith::AndIOp>(loc, lowCmp, highCmp);
break;
}
case arith::CmpIPredicate::ne: {
cmpResult = rewriter.create<arith::OrIOp>(loc, lowCmp, highCmp);
break;
}
default: {
// Handle inequality checks.
Value highEq = rewriter.create<arith::CmpIOp>(
loc, arith::CmpIPredicate::eq, lhsElem1, rhsElem1);
cmpResult =
rewriter.create<arith::SelectOp>(loc, highEq, lowCmp, highCmp);
break;
}
}
assert(cmpResult && "Unhandled case");
rewriter.replaceOp(op, dropTrailingX1Dim(rewriter, loc, cmpResult));
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertMulI
//===----------------------------------------------------------------------===//
struct ConvertMulI final : OpConversionPattern<arith::MulIOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::MulIOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
auto newTy = getTypeConverter()->convertType<VectorType>(op.getType());
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
auto [lhsElem0, lhsElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
auto [rhsElem0, rhsElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getRhs());
// The multiplication algorithm used is the standard (long) multiplication.
// Multiplying two i2N integers produces (at most) an i4N result, but
// because the calculation of top i2N is not necessary, we omit it.
auto mulLowLow =
rewriter.create<arith::MulUIExtendedOp>(loc, lhsElem0, rhsElem0);
Value mulLowHi = rewriter.create<arith::MulIOp>(loc, lhsElem0, rhsElem1);
Value mulHiLow = rewriter.create<arith::MulIOp>(loc, lhsElem1, rhsElem0);
Value resLow = mulLowLow.getLow();
Value resHi =
rewriter.create<arith::AddIOp>(loc, mulLowLow.getHigh(), mulLowHi);
resHi = rewriter.create<arith::AddIOp>(loc, resHi, mulHiLow);
Value resultVec =
constructResultVector(rewriter, loc, newTy, {resLow, resHi});
rewriter.replaceOp(op, resultVec);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertExtSI
//===----------------------------------------------------------------------===//
struct ConvertExtSI final : OpConversionPattern<arith::ExtSIOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::ExtSIOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
auto newTy = getTypeConverter()->convertType<VectorType>(op.getType());
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
Type newResultComponentTy = reduceInnermostDim(newTy);
// Sign-extend the input value to determine the low half of the result.
// Then, check if the low half is negative, and sign-extend the comparison
// result to get the high half.
Value newOperand = appendX1Dim(rewriter, loc, adaptor.getIn());
Value extended = rewriter.createOrFold<arith::ExtSIOp>(
loc, newResultComponentTy, newOperand);
Value operandZeroCst =
createScalarOrSplatConstant(rewriter, loc, newResultComponentTy, 0);
Value signBit = rewriter.create<arith::CmpIOp>(
loc, arith::CmpIPredicate::slt, extended, operandZeroCst);
Value signValue =
rewriter.create<arith::ExtSIOp>(loc, newResultComponentTy, signBit);
Value resultVec =
constructResultVector(rewriter, loc, newTy, {extended, signValue});
rewriter.replaceOp(op, resultVec);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertExtUI
//===----------------------------------------------------------------------===//
struct ConvertExtUI final : OpConversionPattern<arith::ExtUIOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::ExtUIOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
auto newTy = getTypeConverter()->convertType<VectorType>(op.getType());
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
Type newResultComponentTy = reduceInnermostDim(newTy);
// Zero-extend the input value to determine the low half of the result.
// The high half is always zero.
Value newOperand = appendX1Dim(rewriter, loc, adaptor.getIn());
Value extended = rewriter.createOrFold<arith::ExtUIOp>(
loc, newResultComponentTy, newOperand);
Value zeroCst = createScalarOrSplatConstant(rewriter, loc, newTy, 0);
Value newRes = insertLastDimSlice(rewriter, loc, extended, zeroCst, 0);
rewriter.replaceOp(op, newRes);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertMaxMin
//===----------------------------------------------------------------------===//
template <typename SourceOp, arith::CmpIPredicate CmpPred>
struct ConvertMaxMin final : OpConversionPattern<SourceOp> {
using OpConversionPattern<SourceOp>::OpConversionPattern;
LogicalResult
matchAndRewrite(SourceOp op, typename SourceOp::Adaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
Type oldTy = op.getType();
auto newTy = dyn_cast_or_null<VectorType>(
this->getTypeConverter()->convertType(oldTy));
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
// Rewrite Max*I/Min*I as compare and select over original operands. Let
// the CmpI and Select emulation patterns handle the final legalization.
Value cmp =
rewriter.create<arith::CmpIOp>(loc, CmpPred, op.getLhs(), op.getRhs());
rewriter.replaceOpWithNewOp<arith::SelectOp>(op, cmp, op.getLhs(),
op.getRhs());
return success();
}
};
// Convert IndexCast ops
//===----------------------------------------------------------------------===//
/// Returns true iff the type is `index` or `vector<...index>`.
static bool isIndexOrIndexVector(Type type) {
if (isa<IndexType>(type))
return true;
if (auto vectorTy = dyn_cast<VectorType>(type))
if (isa<IndexType>(vectorTy.getElementType()))
return true;
return false;
}
template <typename CastOp>
struct ConvertIndexCastIntToIndex final : OpConversionPattern<CastOp> {
using OpConversionPattern<CastOp>::OpConversionPattern;
LogicalResult
matchAndRewrite(CastOp op, typename CastOp::Adaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Type resultType = op.getType();
if (!isIndexOrIndexVector(resultType))
return failure();
Location loc = op.getLoc();
Type inType = op.getIn().getType();
auto newInTy =
this->getTypeConverter()->template convertType<VectorType>(inType);
if (!newInTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", inType));
// Discard the high half of the input truncating the original value.
Value extracted = extractLastDimSlice(rewriter, loc, adaptor.getIn(), 0);
extracted = dropTrailingX1Dim(rewriter, loc, extracted);
rewriter.replaceOpWithNewOp<CastOp>(op, resultType, extracted);
return success();
}
};
template <typename CastOp, typename ExtensionOp>
struct ConvertIndexCastIndexToInt final : OpConversionPattern<CastOp> {
using OpConversionPattern<CastOp>::OpConversionPattern;
LogicalResult
matchAndRewrite(CastOp op, typename CastOp::Adaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Type inType = op.getIn().getType();
if (!isIndexOrIndexVector(inType))
return failure();
Location loc = op.getLoc();
auto *typeConverter =
this->template getTypeConverter<arith::WideIntEmulationConverter>();
Type resultType = op.getType();
auto newTy = typeConverter->template convertType<VectorType>(resultType);
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", resultType));
// Emit an index cast over the matching narrow type.
Type narrowTy =
rewriter.getIntegerType(typeConverter->getMaxTargetIntBitWidth());
if (auto vecTy = dyn_cast<VectorType>(resultType))
narrowTy = VectorType::get(vecTy.getShape(), narrowTy);
// Sign or zero-extend the result. Let the matching conversion pattern
// legalize the extension op.
Value underlyingVal =
rewriter.create<CastOp>(loc, narrowTy, adaptor.getIn());
rewriter.replaceOpWithNewOp<ExtensionOp>(op, resultType, underlyingVal);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertSelect
//===----------------------------------------------------------------------===//
struct ConvertSelect final : OpConversionPattern<arith::SelectOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::SelectOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
auto newTy = getTypeConverter()->convertType<VectorType>(op.getType());
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
auto [trueElem0, trueElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getTrueValue());
auto [falseElem0, falseElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getFalseValue());
Value cond = appendX1Dim(rewriter, loc, adaptor.getCondition());
Value resElem0 =
rewriter.create<arith::SelectOp>(loc, cond, trueElem0, falseElem0);
Value resElem1 =
rewriter.create<arith::SelectOp>(loc, cond, trueElem1, falseElem1);
Value resultVec =
constructResultVector(rewriter, loc, newTy, {resElem0, resElem1});
rewriter.replaceOp(op, resultVec);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertShLI
//===----------------------------------------------------------------------===//
struct ConvertShLI final : OpConversionPattern<arith::ShLIOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::ShLIOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
Type oldTy = op.getType();
auto newTy = getTypeConverter()->convertType<VectorType>(oldTy);
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
Type newOperandTy = reduceInnermostDim(newTy);
// `oldBitWidth` == `2 * newBitWidth`
unsigned newBitWidth = newTy.getElementTypeBitWidth();
auto [lhsElem0, lhsElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
Value rhsElem0 = extractLastDimSlice(rewriter, loc, adaptor.getRhs(), 0);
// Assume that the shift amount is < 2 * newBitWidth. Calculate the low and
// high halves of the results separately:
// 1. low := LHS.low shli RHS
//
// 2. high := a or b or c, where:
// a) Bits from LHS.high, shifted by the RHS.
// b) Bits from LHS.low, shifted right. These come into play when
// RHS < newBitWidth, e.g.:
// [0000][llll] shli 3 --> [0lll][l000]
// ^
// |
// [llll] shrui (4 - 3)
// c) Bits from LHS.low, shifted left. These matter when
// RHS > newBitWidth, e.g.:
// [0000][llll] shli 7 --> [l000][0000]
// ^
// |
// [llll] shli (7 - 4)
//
// Because shifts by values >= newBitWidth are undefined, we ignore the high
// half of RHS, and introduce 'bounds checks' to account for
// RHS.low > newBitWidth.
//
// TODO: Explore possible optimizations.
Value zeroCst = createScalarOrSplatConstant(rewriter, loc, newOperandTy, 0);
Value elemBitWidth =
createScalarOrSplatConstant(rewriter, loc, newOperandTy, newBitWidth);
Value illegalElemShift = rewriter.create<arith::CmpIOp>(
loc, arith::CmpIPredicate::uge, rhsElem0, elemBitWidth);
Value shiftedElem0 =
rewriter.create<arith::ShLIOp>(loc, lhsElem0, rhsElem0);
Value resElem0 = rewriter.create<arith::SelectOp>(loc, illegalElemShift,
zeroCst, shiftedElem0);
Value cappedShiftAmount = rewriter.create<arith::SelectOp>(
loc, illegalElemShift, elemBitWidth, rhsElem0);
Value rightShiftAmount =
rewriter.create<arith::SubIOp>(loc, elemBitWidth, cappedShiftAmount);
Value shiftedRight =
rewriter.create<arith::ShRUIOp>(loc, lhsElem0, rightShiftAmount);
Value overshotShiftAmount =
rewriter.create<arith::SubIOp>(loc, rhsElem0, elemBitWidth);
Value shiftedLeft =
rewriter.create<arith::ShLIOp>(loc, lhsElem0, overshotShiftAmount);
Value shiftedElem1 =
rewriter.create<arith::ShLIOp>(loc, lhsElem1, rhsElem0);
Value resElem1High = rewriter.create<arith::SelectOp>(
loc, illegalElemShift, zeroCst, shiftedElem1);
Value resElem1Low = rewriter.create<arith::SelectOp>(
loc, illegalElemShift, shiftedLeft, shiftedRight);
Value resElem1 =
rewriter.create<arith::OrIOp>(loc, resElem1Low, resElem1High);
Value resultVec =
constructResultVector(rewriter, loc, newTy, {resElem0, resElem1});
rewriter.replaceOp(op, resultVec);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertShRUI
//===----------------------------------------------------------------------===//
struct ConvertShRUI final : OpConversionPattern<arith::ShRUIOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::ShRUIOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
Type oldTy = op.getType();
auto newTy = getTypeConverter()->convertType<VectorType>(oldTy);
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
Type newOperandTy = reduceInnermostDim(newTy);
// `oldBitWidth` == `2 * newBitWidth`
unsigned newBitWidth = newTy.getElementTypeBitWidth();
auto [lhsElem0, lhsElem1] =
extractLastDimHalves(rewriter, loc, adaptor.getLhs());
Value rhsElem0 = extractLastDimSlice(rewriter, loc, adaptor.getRhs(), 0);
// Assume that the shift amount is < 2 * newBitWidth. Calculate the low and
// high halves of the results separately:
// 1. low := a or b or c, where:
// a) Bits from LHS.low, shifted by the RHS.
// b) Bits from LHS.high, shifted left. These matter when
// RHS < newBitWidth, e.g.:
// [hhhh][0000] shrui 3 --> [000h][hhh0]
// ^
// |
// [hhhh] shli (4 - 1)
// c) Bits from LHS.high, shifted right. These come into play when
// RHS > newBitWidth, e.g.:
// [hhhh][0000] shrui 7 --> [0000][000h]
// ^
// |
// [hhhh] shrui (7 - 4)
//
// 2. high := LHS.high shrui RHS
//
// Because shifts by values >= newBitWidth are undefined, we ignore the high
// half of RHS, and introduce 'bounds checks' to account for
// RHS.low > newBitWidth.
//
// TODO: Explore possible optimizations.
Value zeroCst = createScalarOrSplatConstant(rewriter, loc, newOperandTy, 0);
Value elemBitWidth =
createScalarOrSplatConstant(rewriter, loc, newOperandTy, newBitWidth);
Value illegalElemShift = rewriter.create<arith::CmpIOp>(
loc, arith::CmpIPredicate::uge, rhsElem0, elemBitWidth);
Value shiftedElem0 =
rewriter.create<arith::ShRUIOp>(loc, lhsElem0, rhsElem0);
Value resElem0Low = rewriter.create<arith::SelectOp>(loc, illegalElemShift,
zeroCst, shiftedElem0);
Value shiftedElem1 =
rewriter.create<arith::ShRUIOp>(loc, lhsElem1, rhsElem0);
Value resElem1 = rewriter.create<arith::SelectOp>(loc, illegalElemShift,
zeroCst, shiftedElem1);
Value cappedShiftAmount = rewriter.create<arith::SelectOp>(
loc, illegalElemShift, elemBitWidth, rhsElem0);
Value leftShiftAmount =
rewriter.create<arith::SubIOp>(loc, elemBitWidth, cappedShiftAmount);
Value shiftedLeft =
rewriter.create<arith::ShLIOp>(loc, lhsElem1, leftShiftAmount);
Value overshotShiftAmount =
rewriter.create<arith::SubIOp>(loc, rhsElem0, elemBitWidth);
Value shiftedRight =
rewriter.create<arith::ShRUIOp>(loc, lhsElem1, overshotShiftAmount);
Value resElem0High = rewriter.create<arith::SelectOp>(
loc, illegalElemShift, shiftedRight, shiftedLeft);
Value resElem0 =
rewriter.create<arith::OrIOp>(loc, resElem0Low, resElem0High);
Value resultVec =
constructResultVector(rewriter, loc, newTy, {resElem0, resElem1});
rewriter.replaceOp(op, resultVec);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertShRSI
//===----------------------------------------------------------------------===//
struct ConvertShRSI final : OpConversionPattern<arith::ShRSIOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::ShRSIOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
Type oldTy = op.getType();
auto newTy = getTypeConverter()->convertType<VectorType>(oldTy);
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", op.getType()));
Value lhsElem1 = extractLastDimSlice(rewriter, loc, adaptor.getLhs(), 1);
Value rhsElem0 = extractLastDimSlice(rewriter, loc, adaptor.getRhs(), 0);
Type narrowTy = rhsElem0.getType();
int64_t origBitwidth = newTy.getElementTypeBitWidth() * 2;
// Rewrite this as an bitwise or of `arith.shrui` and sign extension bits.
// Perform as many ops over the narrow integer type as possible and let the
// other emulation patterns convert the rest.
Value elemZero = createScalarOrSplatConstant(rewriter, loc, narrowTy, 0);
Value signBit = rewriter.create<arith::CmpIOp>(
loc, arith::CmpIPredicate::slt, lhsElem1, elemZero);
signBit = dropTrailingX1Dim(rewriter, loc, signBit);
// Create a bit pattern of either all ones or all zeros. Then shift it left
// to calculate the sign extension bits created by shifting the original
// sign bit right.
Value allSign = rewriter.create<arith::ExtSIOp>(loc, oldTy, signBit);
Value maxShift =
createScalarOrSplatConstant(rewriter, loc, narrowTy, origBitwidth);
Value numNonSignExtBits =
rewriter.create<arith::SubIOp>(loc, maxShift, rhsElem0);
numNonSignExtBits = dropTrailingX1Dim(rewriter, loc, numNonSignExtBits);
numNonSignExtBits =
rewriter.create<arith::ExtUIOp>(loc, oldTy, numNonSignExtBits);
Value signBits =
rewriter.create<arith::ShLIOp>(loc, allSign, numNonSignExtBits);
// Use original arguments to create the right shift.
Value shrui =
rewriter.create<arith::ShRUIOp>(loc, op.getLhs(), op.getRhs());
Value shrsi = rewriter.create<arith::OrIOp>(loc, shrui, signBits);
// Handle shifting by zero. This is necessary when the `signBits` shift is
// invalid.
Value isNoop = rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::eq,
rhsElem0, elemZero);
isNoop = dropTrailingX1Dim(rewriter, loc, isNoop);
rewriter.replaceOpWithNewOp<arith::SelectOp>(op, isNoop, op.getLhs(),
shrsi);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertSIToFP
//===----------------------------------------------------------------------===//
struct ConvertSIToFP final : OpConversionPattern<arith::SIToFPOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::SIToFPOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op.getLoc();
Value in = op.getIn();
Type oldTy = in.getType();
auto newTy = getTypeConverter()->convertType<VectorType>(oldTy);
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", oldTy));
unsigned oldBitWidth = getElementTypeOrSelf(oldTy).getIntOrFloatBitWidth();
Value zeroCst = createScalarOrSplatConstant(rewriter, loc, oldTy, 0);
Value oneCst = createScalarOrSplatConstant(rewriter, loc, oldTy, 1);
Value allOnesCst = createScalarOrSplatConstant(
rewriter, loc, oldTy, APInt::getAllOnes(oldBitWidth));
// To avoid operating on very large unsigned numbers, perform the
// conversion on the absolute value. Then, decide whether to negate the
// result or not based on that sign bit. We assume two's complement and
// implement negation by flipping all bits and adding 1.
// Note that this relies on the the other conversion patterns to legalize
// created ops and narrow the bit widths.
Value isNeg = rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::slt,
in, zeroCst);
Value bitwiseNeg = rewriter.create<arith::XOrIOp>(loc, in, allOnesCst);
Value neg = rewriter.create<arith::AddIOp>(loc, bitwiseNeg, oneCst);
Value abs = rewriter.create<arith::SelectOp>(loc, isNeg, neg, in);
Value absResult = rewriter.create<arith::UIToFPOp>(loc, op.getType(), abs);
Value negResult = rewriter.create<arith::NegFOp>(loc, absResult);
rewriter.replaceOpWithNewOp<arith::SelectOp>(op, isNeg, negResult,
absResult);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertUIToFP
//===----------------------------------------------------------------------===//
struct ConvertUIToFP final : OpConversionPattern<arith::UIToFPOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::UIToFPOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op.getLoc();
Type oldTy = op.getIn().getType();
auto newTy = getTypeConverter()->convertType<VectorType>(oldTy);
if (!newTy)
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported type: {0}", oldTy));
unsigned newBitWidth = newTy.getElementTypeBitWidth();
auto [low, hi] = extractLastDimHalves(rewriter, loc, adaptor.getIn());
Value lowInt = dropTrailingX1Dim(rewriter, loc, low);
Value hiInt = dropTrailingX1Dim(rewriter, loc, hi);
Value zeroCst =
createScalarOrSplatConstant(rewriter, loc, hiInt.getType(), 0);
// The final result has the following form:
// if (hi == 0) return uitofp(low)
// else return uitofp(low) + uitofp(hi) * 2^BW
//
// where `BW` is the bitwidth of the narrowed integer type. We emit a
// select to make it easier to fold-away the `hi` part calculation when it
// is known to be zero.
//
// Note 1: The emulation is precise only for input values that have exact
// integer representation in the result floating point type, and may lead
// loss of precision otherwise.
//
// Note 2: We do not strictly need the `hi == 0`, case, but it makes
// constant folding easier.
Value hiEqZero = rewriter.create<arith::CmpIOp>(
loc, arith::CmpIPredicate::eq, hiInt, zeroCst);
Type resultTy = op.getType();
Type resultElemTy = getElementTypeOrSelf(resultTy);
Value lowFp = rewriter.create<arith::UIToFPOp>(loc, resultTy, lowInt);
Value hiFp = rewriter.create<arith::UIToFPOp>(loc, resultTy, hiInt);
int64_t pow2Int = int64_t(1) << newBitWidth;
TypedAttr pow2Attr =
rewriter.getFloatAttr(resultElemTy, static_cast<double>(pow2Int));
if (auto vecTy = dyn_cast<VectorType>(resultTy))
pow2Attr = SplatElementsAttr::get(vecTy, pow2Attr);
Value pow2Val = rewriter.create<arith::ConstantOp>(loc, resultTy, pow2Attr);
Value hiVal = rewriter.create<arith::MulFOp>(loc, hiFp, pow2Val);
Value result = rewriter.create<arith::AddFOp>(loc, lowFp, hiVal);
rewriter.replaceOpWithNewOp<arith::SelectOp>(op, hiEqZero, lowFp, result);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertTruncI
//===----------------------------------------------------------------------===//
struct ConvertTruncI final : OpConversionPattern<arith::TruncIOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(arith::TruncIOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op.getLoc();
// Check if the result type is legal for this target. Currently, we do not
// support truncation to types wider than supported by the target.
if (!getTypeConverter()->isLegal(op.getType()))
return rewriter.notifyMatchFailure(
loc, llvm::formatv("unsupported truncation result type: {0}",
op.getType()));
// Discard the high half of the input. Truncate the low half, if
// necessary.
Value extracted = extractLastDimSlice(rewriter, loc, adaptor.getIn(), 0);
extracted = dropTrailingX1Dim(rewriter, loc, extracted);
Value truncated =
rewriter.createOrFold<arith::TruncIOp>(loc, op.getType(), extracted);
rewriter.replaceOp(op, truncated);
return success();
}
};
//===----------------------------------------------------------------------===//
// ConvertVectorPrint
//===----------------------------------------------------------------------===//
struct ConvertVectorPrint final : OpConversionPattern<vector::PrintOp> {
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(vector::PrintOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
rewriter.replaceOpWithNewOp<vector::PrintOp>(op, adaptor.getSource());
return success();
}
};
//===----------------------------------------------------------------------===//
// Pass Definition
//===----------------------------------------------------------------------===//
struct EmulateWideIntPass final
: arith::impl::ArithEmulateWideIntBase<EmulateWideIntPass> {
using ArithEmulateWideIntBase::ArithEmulateWideIntBase;
void runOnOperation() override {
if (!llvm::isPowerOf2_32(widestIntSupported) || widestIntSupported < 2) {
signalPassFailure();
return;
}
Operation *op = getOperation();
MLIRContext *ctx = op->getContext();
arith::WideIntEmulationConverter typeConverter(widestIntSupported);
ConversionTarget target(*ctx);
target.addDynamicallyLegalOp<func::FuncOp>([&typeConverter](Operation *op) {
return typeConverter.isLegal(cast<func::FuncOp>(op).getFunctionType());
});
auto opLegalCallback = [&typeConverter](Operation *op) {
return typeConverter.isLegal(op);
};
target.addDynamicallyLegalOp<func::CallOp, func::ReturnOp>(opLegalCallback);
target
.addDynamicallyLegalDialect<arith::ArithDialect, vector::VectorDialect>(
opLegalCallback);
RewritePatternSet patterns(ctx);
arith::populateArithWideIntEmulationPatterns(typeConverter, patterns);
if (failed(applyPartialConversion(op, target, std::move(patterns))))
signalPassFailure();
}
};
} // end anonymous namespace
//===----------------------------------------------------------------------===//
// Public Interface Definition
//===----------------------------------------------------------------------===//
arith::WideIntEmulationConverter::WideIntEmulationConverter(
unsigned widestIntSupportedByTarget)
: maxIntWidth(widestIntSupportedByTarget) {
assert(llvm::isPowerOf2_32(widestIntSupportedByTarget) &&
"Only power-of-two integers with are supported");
assert(widestIntSupportedByTarget >= 2 && "Integer type too narrow");
// Allow unknown types.
addConversion([](Type ty) -> std::optional<Type> { return ty; });
// Scalar case.
addConversion([this](IntegerType ty) -> std::optional<Type> {
unsigned width = ty.getWidth();
if (width <= maxIntWidth)
return ty;
// i2N --> vector<2xiN>
if (width == 2 * maxIntWidth)
return VectorType::get(2, IntegerType::get(ty.getContext(), maxIntWidth));
return std::nullopt;
});
// Vector case.
addConversion([this](VectorType ty) -> std::optional<Type> {
auto intTy = dyn_cast<IntegerType>(ty.getElementType());
if (!intTy)
return ty;
unsigned width = intTy.getWidth();
if (width <= maxIntWidth)
return ty;
// vector<...xi2N> --> vector<...x2xiN>
if (width == 2 * maxIntWidth) {
auto newShape = to_vector(ty.getShape());
newShape.push_back(2);
return VectorType::get(newShape,
IntegerType::get(ty.getContext(), maxIntWidth));
}
return std::nullopt;
});
// Function case.
addConversion([this](FunctionType ty) -> std::optional<Type> {
// Convert inputs and results, e.g.:
// (i2N, i2N) -> i2N --> (vector<2xiN>, vector<2xiN>) -> vector<2xiN>
SmallVector<Type> inputs;
if (failed(convertTypes(ty.getInputs(), inputs)))
return std::nullopt;
SmallVector<Type> results;
if (failed(convertTypes(ty.getResults(), results)))
return std::nullopt;
return FunctionType::get(ty.getContext(), inputs, results);
});
}
void arith::populateArithWideIntEmulationPatterns(
WideIntEmulationConverter &typeConverter, RewritePatternSet &patterns) {
// Populate `func.*` conversion patterns.
populateFunctionOpInterfaceTypeConversionPattern<func::FuncOp>(patterns,
typeConverter);
populateCallOpTypeConversionPattern(patterns, typeConverter);
populateReturnOpTypeConversionPattern(patterns, typeConverter);
// Populate `arith.*` conversion patterns.
patterns.add<
// Misc ops.
ConvertConstant, ConvertCmpI, ConvertSelect, ConvertVectorPrint,
// Binary ops.
ConvertAddI, ConvertMulI, ConvertShLI, ConvertShRSI, ConvertShRUI,
ConvertMaxMin<arith::MaxUIOp, arith::CmpIPredicate::ugt>,
ConvertMaxMin<arith::MaxSIOp, arith::CmpIPredicate::sgt>,
ConvertMaxMin<arith::MinUIOp, arith::CmpIPredicate::ult>,
ConvertMaxMin<arith::MinSIOp, arith::CmpIPredicate::slt>,
// Bitwise binary ops.
ConvertBitwiseBinary<arith::AndIOp>, ConvertBitwiseBinary<arith::OrIOp>,
ConvertBitwiseBinary<arith::XOrIOp>,
// Extension and truncation ops.
ConvertExtSI, ConvertExtUI, ConvertTruncI,
// Cast ops.
ConvertIndexCastIntToIndex<arith::IndexCastOp>,
ConvertIndexCastIntToIndex<arith::IndexCastUIOp>,
ConvertIndexCastIndexToInt<arith::IndexCastOp, arith::ExtSIOp>,
ConvertIndexCastIndexToInt<arith::IndexCastUIOp, arith::ExtUIOp>,
ConvertSIToFP, ConvertUIToFP>(typeConverter, patterns.getContext());
}
|