/
ModuleImport.cpp
1732 lines (1539 loc) · 66.3 KB
/
ModuleImport.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
//===- ModuleImport.cpp - LLVM to MLIR conversion ---------------*- 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
//
//===----------------------------------------------------------------------===//
//
// This file implements the import of an LLVM IR module into an LLVM dialect
// module.
//
//===----------------------------------------------------------------------===//
#include "mlir/Target/LLVMIR/ModuleImport.h"
#include "mlir/Target/LLVMIR/Import.h"
#include "AttrKindDetail.h"
#include "DebugImporter.h"
#include "LoopAnnotationImporter.h"
#include "mlir/Dialect/DLTI/DLTI.h"
#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
#include "mlir/IR/Matchers.h"
#include "mlir/Interfaces/DataLayoutInterfaces.h"
#include "mlir/Tools/mlir-translate/Translation.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/ADT/ScopeExit.h"
#include "llvm/ADT/StringSet.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/InlineAsm.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Metadata.h"
#include "llvm/IR/Operator.h"
#include "llvm/Support/ModRef.h"
using namespace mlir;
using namespace mlir::LLVM;
using namespace mlir::LLVM::detail;
#include "mlir/Dialect/LLVMIR/LLVMConversionEnumsFromLLVM.inc"
// Utility to print an LLVM value as a string for passing to emitError().
// FIXME: Diagnostic should be able to natively handle types that have
// operator << (raw_ostream&) defined.
static std::string diag(const llvm::Value &value) {
std::string str;
llvm::raw_string_ostream os(str);
os << value;
return os.str();
}
// Utility to print an LLVM metadata node as a string for passing
// to emitError(). The module argument is needed to print the nodes
// canonically numbered.
static std::string diagMD(const llvm::Metadata *node,
const llvm::Module *module) {
std::string str;
llvm::raw_string_ostream os(str);
node->print(os, module, /*IsForDebug=*/true);
return os.str();
}
/// Returns the name of the global_ctors global variables.
static constexpr StringRef getGlobalCtorsVarName() {
return "llvm.global_ctors";
}
/// Returns the name of the global_dtors global variables.
static constexpr StringRef getGlobalDtorsVarName() {
return "llvm.global_dtors";
}
/// Returns the symbol name for the module-level metadata operation. It must not
/// conflict with the user namespace.
static constexpr StringRef getGlobalMetadataOpName() {
return "__llvm_global_metadata";
}
/// Returns a supported MLIR floating point type of the given bit width or null
/// if the bit width is not supported.
static FloatType getDLFloatType(MLIRContext &ctx, int32_t bitwidth) {
switch (bitwidth) {
case 16:
return FloatType::getF16(&ctx);
case 32:
return FloatType::getF32(&ctx);
case 64:
return FloatType::getF64(&ctx);
case 80:
return FloatType::getF80(&ctx);
case 128:
return FloatType::getF128(&ctx);
default:
return nullptr;
}
}
/// Converts the sync scope identifier of `inst` to the string representation
/// necessary to build an atomic LLVM dialect operation. Returns the empty
/// string if the operation has either no sync scope or the default system-level
/// sync scope attached. The atomic operations only set their sync scope
/// attribute if they have a non-default sync scope attached.
static StringRef getLLVMSyncScope(llvm::Instruction *inst) {
std::optional<llvm::SyncScope::ID> syncScopeID =
llvm::getAtomicSyncScopeID(inst);
if (!syncScopeID)
return "";
// Search the sync scope name for the given identifier. The default
// system-level sync scope thereby maps to the empty string.
SmallVector<StringRef> syncScopeName;
llvm::LLVMContext &llvmContext = inst->getContext();
llvmContext.getSyncScopeNames(syncScopeName);
auto *it = llvm::find_if(syncScopeName, [&](StringRef name) {
return *syncScopeID == llvmContext.getOrInsertSyncScopeID(name);
});
if (it != syncScopeName.end())
return *it;
llvm_unreachable("incorrect sync scope identifier");
}
/// Converts an array of unsigned indices to a signed integer position array.
static SmallVector<int64_t> getPositionFromIndices(ArrayRef<unsigned> indices) {
SmallVector<int64_t> position;
llvm::append_range(position, indices);
return position;
}
/// Converts the LLVM instructions that have a generated MLIR builder. Using a
/// static implementation method called from the module import ensures the
/// builders have to use the `moduleImport` argument and cannot directly call
/// import methods. As a result, both the intrinsic and the instruction MLIR
/// builders have to use the `moduleImport` argument and none of them has direct
/// access to the private module import methods.
static LogicalResult convertInstructionImpl(OpBuilder &odsBuilder,
llvm::Instruction *inst,
ModuleImport &moduleImport) {
// Copy the operands to an LLVM operands array reference for conversion.
SmallVector<llvm::Value *> operands(inst->operands());
ArrayRef<llvm::Value *> llvmOperands(operands);
// Convert all instructions that provide an MLIR builder.
#include "mlir/Dialect/LLVMIR/LLVMOpFromLLVMIRConversions.inc"
return failure();
}
/// Creates an attribute containing ABI and preferred alignment numbers parsed
/// a string. The string may be either "abi:preferred" or just "abi". In the
/// latter case, the preferred alignment is considered equal to ABI alignment.
static DenseIntElementsAttr parseDataLayoutAlignment(MLIRContext &ctx,
StringRef spec) {
auto i32 = IntegerType::get(&ctx, 32);
StringRef abiString, preferredString;
std::tie(abiString, preferredString) = spec.split(':');
int abi, preferred;
if (abiString.getAsInteger(/*Radix=*/10, abi))
return nullptr;
if (preferredString.empty())
preferred = abi;
else if (preferredString.getAsInteger(/*Radix=*/10, preferred))
return nullptr;
return DenseIntElementsAttr::get(VectorType::get({2}, i32), {abi, preferred});
}
/// Translate the given LLVM data layout into an MLIR equivalent using the DLTI
/// dialect.
DataLayoutSpecInterface
mlir::translateDataLayout(const llvm::DataLayout &dataLayout,
MLIRContext *context) {
assert(context && "expected MLIR context");
std::string layoutstr = dataLayout.getStringRepresentation();
// Remaining unhandled default layout defaults
// e (little endian if not set)
// p[n]:64:64:64 (non zero address spaces have 64-bit properties)
std::string append =
"p:64:64:64-S0-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f16:16:16-f64:"
"64:64-f128:128:128-v64:64:64-v128:128:128-a:0:64";
if (layoutstr.empty())
layoutstr = append;
else
layoutstr = layoutstr + "-" + append;
StringRef layout(layoutstr);
SmallVector<DataLayoutEntryInterface> entries;
StringSet<> seen;
while (!layout.empty()) {
// Split at '-'.
std::pair<StringRef, StringRef> split = layout.split('-');
StringRef current;
std::tie(current, layout) = split;
// Split at ':'.
StringRef kind, spec;
std::tie(kind, spec) = current.split(':');
if (seen.contains(kind))
continue;
seen.insert(kind);
char symbol = kind.front();
StringRef parameter = kind.substr(1);
if (symbol == 'i' || symbol == 'f') {
unsigned bitwidth;
if (parameter.getAsInteger(/*Radix=*/10, bitwidth))
return nullptr;
DenseIntElementsAttr params = parseDataLayoutAlignment(*context, spec);
if (!params)
return nullptr;
auto entry = DataLayoutEntryAttr::get(
symbol == 'i' ? static_cast<Type>(IntegerType::get(context, bitwidth))
: getDLFloatType(*context, bitwidth),
params);
entries.emplace_back(entry);
} else if (symbol == 'e' || symbol == 'E') {
auto value = StringAttr::get(
context, symbol == 'e' ? DLTIDialect::kDataLayoutEndiannessLittle
: DLTIDialect::kDataLayoutEndiannessBig);
auto entry = DataLayoutEntryAttr::get(
StringAttr::get(context, DLTIDialect::kDataLayoutEndiannessKey),
value);
entries.emplace_back(entry);
}
}
return DataLayoutSpecAttr::get(context, entries);
}
/// Get a topologically sorted list of blocks for the given function.
static SetVector<llvm::BasicBlock *>
getTopologicallySortedBlocks(llvm::Function *func) {
SetVector<llvm::BasicBlock *> blocks;
for (llvm::BasicBlock &bb : *func) {
if (blocks.count(&bb) == 0) {
llvm::ReversePostOrderTraversal<llvm::BasicBlock *> traversal(&bb);
blocks.insert(traversal.begin(), traversal.end());
}
}
assert(blocks.size() == func->size() && "some blocks are not sorted");
return blocks;
}
ModuleImport::ModuleImport(ModuleOp mlirModule,
std::unique_ptr<llvm::Module> llvmModule)
: builder(mlirModule->getContext()), context(mlirModule->getContext()),
mlirModule(mlirModule), llvmModule(std::move(llvmModule)),
iface(mlirModule->getContext()),
typeTranslator(*mlirModule->getContext()),
debugImporter(std::make_unique<DebugImporter>(mlirModule)),
loopAnnotationImporter(
std::make_unique<LoopAnnotationImporter>(builder)) {
builder.setInsertionPointToStart(mlirModule.getBody());
}
MetadataOp ModuleImport::getGlobalMetadataOp() {
if (globalMetadataOp)
return globalMetadataOp;
OpBuilder::InsertionGuard guard(builder);
builder.setInsertionPointToEnd(mlirModule.getBody());
return globalMetadataOp = builder.create<MetadataOp>(
mlirModule.getLoc(), getGlobalMetadataOpName());
}
LogicalResult ModuleImport::processTBAAMetadata(const llvm::MDNode *node) {
Location loc = mlirModule.getLoc();
SmallVector<const llvm::MDNode *> workList;
SetVector<const llvm::MDNode *> nodesToConvert;
workList.push_back(node);
while (!workList.empty()) {
const llvm::MDNode *current = workList.pop_back_val();
if (tbaaMapping.count(current))
continue;
// Allow cycles in TBAA metadata. Just import it as-is,
// and diagnose the problem during LLVMIR dialect verification.
if (!nodesToConvert.insert(current))
continue;
for (const llvm::MDOperand &operand : current->operands())
if (auto *opNode = dyn_cast_or_null<const llvm::MDNode>(operand.get()))
workList.push_back(opNode);
}
// If `node` is a valid TBAA root node, then return its identity
// string, otherwise return std::nullopt.
auto getIdentityIfRootNode =
[&](const llvm::MDNode *node) -> std::optional<StringRef> {
// Root node, e.g.:
// !0 = !{!"Simple C/C++ TBAA"}
if (node->getNumOperands() != 1)
return std::nullopt;
// If the operand is MDString, then assume that this is a root node.
if (const auto *op0 = dyn_cast<const llvm::MDString>(node->getOperand(0)))
return op0->getString();
return std::nullopt;
};
// If `node` looks like a TBAA type descriptor metadata,
// then return true, if it is a valid node, and false otherwise.
// If it does not look like a TBAA type descriptor metadata, then
// return std::nullopt.
// If `identity` and `memberTypes/Offsets` are non-null, then they will
// contain the converted metadata operands for a valid TBAA node (i.e. when
// true is returned).
auto isTypeDescriptorNode =
[&](const llvm::MDNode *node, StringRef *identity = nullptr,
SmallVectorImpl<Attribute> *memberTypes = nullptr,
SmallVectorImpl<int64_t> *memberOffsets =
nullptr) -> std::optional<bool> {
unsigned numOperands = node->getNumOperands();
// Type descriptor, e.g.:
// !1 = !{!"int", !0, /*optional*/i64 0} /* scalar int type */
// !2 = !{!"agg_t", !1, i64 0} /* struct agg_t { int x; } */
if (numOperands < 2)
return std::nullopt;
// TODO: support "new" format (D41501) for type descriptors,
// where the first operand is an MDNode.
const auto *identityNode =
dyn_cast<const llvm::MDString>(node->getOperand(0));
if (!identityNode)
return std::nullopt;
// This should be a type descriptor node.
if (identity)
*identity = identityNode->getString();
for (unsigned pairNum = 0, e = numOperands / 2; pairNum < e; ++pairNum) {
const auto *memberNode =
dyn_cast<const llvm::MDNode>(node->getOperand(2 * pairNum + 1));
if (!memberNode) {
emitError(loc) << "operand '" << 2 * pairNum + 1 << "' must be MDNode: "
<< diagMD(node, llvmModule.get());
return false;
}
int64_t offset = 0;
if (2 * pairNum + 2 >= numOperands) {
// Allow for optional 0 offset in 2-operand nodes.
if (numOperands != 2) {
emitError(loc) << "missing member offset: "
<< diagMD(node, llvmModule.get());
return false;
}
} else {
auto *offsetCI = llvm::mdconst::dyn_extract<llvm::ConstantInt>(
node->getOperand(2 * pairNum + 2));
if (!offsetCI) {
emitError(loc) << "operand '" << 2 * pairNum + 2
<< "' must be ConstantInt: "
<< diagMD(node, llvmModule.get());
return false;
}
offset = offsetCI->getZExtValue();
}
if (memberTypes)
memberTypes->push_back(tbaaMapping.lookup(memberNode));
if (memberOffsets)
memberOffsets->push_back(offset);
}
return true;
};
// If `node` looks like a TBAA access tag metadata,
// then return true, if it is a valid node, and false otherwise.
// If it does not look like a TBAA access tag metadata, then
// return std::nullopt.
// If the other arguments are non-null, then they will contain
// the converted metadata operands for a valid TBAA node (i.e. when true is
// returned).
auto isTagNode =
[&](const llvm::MDNode *node, SymbolRefAttr *baseSymRef = nullptr,
SymbolRefAttr *accessSymRef = nullptr, int64_t *offset = nullptr,
bool *isConstant = nullptr) -> std::optional<bool> {
// Access tag, e.g.:
// !3 = !{!1, !1, i64 0} /* scalar int access */
// !4 = !{!2, !1, i64 0} /* agg_t::x access */
//
// Optional 4th argument is ConstantInt 0/1 identifying whether
// the location being accessed is "constant" (see for details:
// https://llvm.org/docs/LangRef.html#representation).
unsigned numOperands = node->getNumOperands();
if (numOperands != 3 && numOperands != 4)
return std::nullopt;
const auto *baseMD = dyn_cast<const llvm::MDNode>(node->getOperand(0));
const auto *accessMD = dyn_cast<const llvm::MDNode>(node->getOperand(1));
auto *offsetCI =
llvm::mdconst::dyn_extract<llvm::ConstantInt>(node->getOperand(2));
if (!baseMD || !accessMD || !offsetCI)
return std::nullopt;
// TODO: support "new" TBAA format, if needed (see D41501).
// In the "old" format the first operand of the access type
// metadata is MDString. We have to distinguish the formats,
// because access tags have the same structure, but different
// meaning for the operands.
if (accessMD->getNumOperands() < 1 ||
!isa<llvm::MDString>(accessMD->getOperand(0)))
return std::nullopt;
bool isConst = false;
if (numOperands == 4) {
auto *isConstantCI =
llvm::mdconst::dyn_extract<llvm::ConstantInt>(node->getOperand(3));
if (!isConstantCI) {
emitError(loc) << "operand '3' must be ConstantInt: "
<< diagMD(node, llvmModule.get());
return false;
}
isConst = isConstantCI->getValue()[0];
}
if (baseSymRef)
*baseSymRef = tbaaMapping.lookup(baseMD);
if (accessSymRef)
*accessSymRef = tbaaMapping.lookup(accessMD);
if (offset)
*offset = offsetCI->getZExtValue();
if (isConstant)
*isConstant = isConst;
return true;
};
// Helper to compute a unique symbol name that includes the given `baseName`.
// Uses the size of the mapping to unique the symbol name.
auto getUniqueSymbolName = [&](StringRef baseName) {
return (Twine("tbaa_") + Twine(baseName) + Twine('_') +
Twine(tbaaMapping.size()))
.str();
};
// Insert new operations at the end of the MetadataOp.
OpBuilder::InsertionGuard guard(builder);
builder.setInsertionPointToEnd(&getGlobalMetadataOp().getBody().back());
StringAttr metadataOpName = SymbolTable::getSymbolName(getGlobalMetadataOp());
// On the first walk, create SymbolRefAttr's and map them
// to nodes in `nodesToConvert`.
for (const auto *current : nodesToConvert) {
if (std::optional<StringRef> identity = getIdentityIfRootNode(current)) {
if (identity.value().empty())
return emitError(loc) << "TBAA root node must have non-empty identity: "
<< diagMD(current, llvmModule.get());
// The root nodes do not have operands, so we can create
// the TBAARootMetadataOp on the first walk.
auto rootNode = builder.create<TBAARootMetadataOp>(
loc, getUniqueSymbolName("root"), identity.value());
tbaaMapping.try_emplace(current, FlatSymbolRefAttr::get(rootNode));
continue;
}
if (std::optional<bool> isValid = isTypeDescriptorNode(current)) {
if (!isValid.value())
return failure();
tbaaMapping.try_emplace(
current, FlatSymbolRefAttr::get(builder.getContext(),
getUniqueSymbolName("type_desc")));
continue;
}
if (std::optional<bool> isValid = isTagNode(current)) {
if (!isValid.value())
return failure();
// TBAATagOp symbols must be referred by their fully qualified
// names, so create a path to TBAATagOp symbol.
tbaaMapping.try_emplace(
current, SymbolRefAttr::get(
builder.getContext(), metadataOpName,
FlatSymbolRefAttr::get(builder.getContext(),
getUniqueSymbolName("tag"))));
continue;
}
return emitError(loc) << "unsupported TBAA node format: "
<< diagMD(current, llvmModule.get());
}
// On the second walk, create TBAA operations using the symbol names from the
// map.
for (const auto *current : nodesToConvert) {
StringRef identity;
SmallVector<Attribute> memberTypes;
SmallVector<int64_t> memberOffsets;
if (std::optional<bool> isValid = isTypeDescriptorNode(
current, &identity, &memberTypes, &memberOffsets)) {
assert(isValid.value() && "type descriptor node must be valid");
builder.create<TBAATypeDescriptorOp>(
loc, tbaaMapping.lookup(current).getLeafReference(),
builder.getStringAttr(identity), builder.getArrayAttr(memberTypes),
memberOffsets);
continue;
}
SymbolRefAttr baseSymRef, accessSymRef;
int64_t offset;
bool isConstant;
if (std::optional<bool> isValid = isTagNode(
current, &baseSymRef, &accessSymRef, &offset, &isConstant)) {
assert(isValid.value() && "access tag node must be valid");
builder.create<TBAATagOp>(
loc, tbaaMapping.lookup(current).getLeafReference(),
baseSymRef.getLeafReference(), accessSymRef.getLeafReference(),
offset, isConstant);
continue;
}
}
return success();
}
LogicalResult
ModuleImport::processAccessGroupMetadata(const llvm::MDNode *node) {
Location loc = mlirModule.getLoc();
if (failed(loopAnnotationImporter->translateAccessGroup(
node, loc, getGlobalMetadataOp())))
return emitError(loc) << "unsupported access group node: "
<< diagMD(node, llvmModule.get());
return success();
}
LogicalResult
ModuleImport::processAliasScopeMetadata(const llvm::MDNode *node) {
Location loc = mlirModule.getLoc();
// Helper that verifies the node has a self reference operand.
auto verifySelfRef = [](const llvm::MDNode *node) {
return node->getNumOperands() != 0 &&
node == dyn_cast<llvm::MDNode>(node->getOperand(0));
};
// Helper that verifies the given operand is a string or does not exist.
auto verifyDescription = [](const llvm::MDNode *node, unsigned idx) {
return idx >= node->getNumOperands() ||
isa<llvm::MDString>(node->getOperand(idx));
};
// Helper that creates an alias scope domain operation.
auto createAliasScopeDomainOp = [&](const llvm::MDNode *aliasDomain) {
StringAttr description = nullptr;
if (aliasDomain->getNumOperands() >= 2)
if (auto *operand = dyn_cast<llvm::MDString>(aliasDomain->getOperand(1)))
description = builder.getStringAttr(operand->getString());
std::string name = llvm::formatv("domain_{0}", aliasScopeMapping.size());
return builder.create<AliasScopeDomainMetadataOp>(loc, name, description);
};
// Collect the alias scopes and domains to translate them.
for (const llvm::MDOperand &operand : node->operands()) {
if (const auto *scope = dyn_cast<llvm::MDNode>(operand)) {
llvm::AliasScopeNode aliasScope(scope);
const llvm::MDNode *domain = aliasScope.getDomain();
// Verify the scope node points to valid scope metadata which includes
// verifying its domain. Perform the verification before looking it up in
// the alias scope mapping since it could have been inserted as a domain
// node before.
if (!verifySelfRef(scope) || !domain || !verifyDescription(scope, 2))
return emitError(loc) << "unsupported alias scope node: "
<< diagMD(scope, llvmModule.get());
if (!verifySelfRef(domain) || !verifyDescription(domain, 1))
return emitError(loc) << "unsupported alias domain node: "
<< diagMD(domain, llvmModule.get());
if (aliasScopeMapping.count(scope))
continue;
// Set the insertion point to the end of the global metadata operation.
MetadataOp metadataOp = getGlobalMetadataOp();
StringAttr metadataOpName =
SymbolTable::getSymbolName(getGlobalMetadataOp());
OpBuilder::InsertionGuard guard(builder);
builder.setInsertionPointToEnd(&metadataOp.getBody().back());
// Convert the domain metadata node if it has not been translated before.
auto it = aliasScopeMapping.find(aliasScope.getDomain());
if (it == aliasScopeMapping.end()) {
auto aliasScopeDomainOp = createAliasScopeDomainOp(domain);
auto symbolRef = SymbolRefAttr::get(
builder.getContext(), metadataOpName,
FlatSymbolRefAttr::get(builder.getContext(),
aliasScopeDomainOp.getSymName()));
it = aliasScopeMapping.try_emplace(domain, symbolRef).first;
}
// Convert the scope metadata node if it has not been converted before.
StringAttr description = nullptr;
if (!aliasScope.getName().empty())
description = builder.getStringAttr(aliasScope.getName());
std::string name = llvm::formatv("scope_{0}", aliasScopeMapping.size());
auto aliasScopeOp = builder.create<AliasScopeMetadataOp>(
loc, name, it->getSecond().getLeafReference().getValue(),
description);
auto symbolRef =
SymbolRefAttr::get(builder.getContext(), metadataOpName,
FlatSymbolRefAttr::get(builder.getContext(),
aliasScopeOp.getSymName()));
aliasScopeMapping.try_emplace(aliasScope.getNode(), symbolRef);
}
}
return success();
}
FailureOr<SmallVector<SymbolRefAttr>>
ModuleImport::lookupAliasScopeAttrs(const llvm::MDNode *node) const {
SmallVector<SymbolRefAttr> aliasScopes;
aliasScopes.reserve(node->getNumOperands());
for (const llvm::MDOperand &operand : node->operands()) {
auto *node = cast<llvm::MDNode>(operand.get());
aliasScopes.push_back(aliasScopeMapping.lookup(node));
}
// Return failure if one of the alias scope lookups failed.
if (llvm::is_contained(aliasScopes, nullptr))
return failure();
return aliasScopes;
}
LogicalResult ModuleImport::convertMetadata() {
OpBuilder::InsertionGuard guard(builder);
builder.setInsertionPointToEnd(mlirModule.getBody());
for (const llvm::Function &func : llvmModule->functions()) {
for (const llvm::Instruction &inst : llvm::instructions(func)) {
// Convert access group metadata nodes.
if (llvm::MDNode *node =
inst.getMetadata(llvm::LLVMContext::MD_access_group))
if (failed(processAccessGroupMetadata(node)))
return failure();
// Convert alias analysis metadata nodes.
llvm::AAMDNodes aliasAnalysisNodes = inst.getAAMetadata();
if (!aliasAnalysisNodes)
continue;
if (aliasAnalysisNodes.TBAA)
if (failed(processTBAAMetadata(aliasAnalysisNodes.TBAA)))
return failure();
if (aliasAnalysisNodes.Scope)
if (failed(processAliasScopeMetadata(aliasAnalysisNodes.Scope)))
return failure();
if (aliasAnalysisNodes.NoAlias)
if (failed(processAliasScopeMetadata(aliasAnalysisNodes.NoAlias)))
return failure();
}
}
return success();
}
LogicalResult ModuleImport::convertGlobals() {
for (llvm::GlobalVariable &globalVar : llvmModule->globals()) {
if (globalVar.getName() == getGlobalCtorsVarName() ||
globalVar.getName() == getGlobalDtorsVarName()) {
if (failed(convertGlobalCtorsAndDtors(&globalVar))) {
return emitError(mlirModule.getLoc())
<< "unhandled global variable: " << diag(globalVar);
}
continue;
}
if (failed(convertGlobal(&globalVar))) {
return emitError(mlirModule.getLoc())
<< "unhandled global variable: " << diag(globalVar);
}
}
return success();
}
LogicalResult ModuleImport::convertFunctions() {
for (llvm::Function &func : llvmModule->functions())
if (failed(processFunction(&func)))
return failure();
return success();
}
void ModuleImport::setNonDebugMetadataAttrs(llvm::Instruction *inst,
Operation *op) {
SmallVector<std::pair<unsigned, llvm::MDNode *>> allMetadata;
inst->getAllMetadataOtherThanDebugLoc(allMetadata);
for (auto &[kind, node] : allMetadata) {
if (!iface.isConvertibleMetadata(kind))
continue;
if (failed(iface.setMetadataAttrs(builder, kind, node, op, *this))) {
Location loc = debugImporter->translateLoc(inst->getDebugLoc());
emitWarning(loc) << "unhandled metadata: "
<< diagMD(node, llvmModule.get()) << " on "
<< diag(*inst);
}
}
}
void ModuleImport::setFastmathFlagsAttr(llvm::Instruction *inst,
Operation *op) const {
auto iface = cast<FastmathFlagsInterface>(op);
// Even if the imported operation implements the fastmath interface, the
// original instruction may not have fastmath flags set. Exit if an
// instruction, such as a non floating-point function call, does not have
// fastmath flags.
if (!isa<llvm::FPMathOperator>(inst))
return;
llvm::FastMathFlags flags = inst->getFastMathFlags();
// Set the fastmath bits flag-by-flag.
FastmathFlags value = {};
value = bitEnumSet(value, FastmathFlags::nnan, flags.noNaNs());
value = bitEnumSet(value, FastmathFlags::ninf, flags.noInfs());
value = bitEnumSet(value, FastmathFlags::nsz, flags.noSignedZeros());
value = bitEnumSet(value, FastmathFlags::arcp, flags.allowReciprocal());
value = bitEnumSet(value, FastmathFlags::contract, flags.allowContract());
value = bitEnumSet(value, FastmathFlags::afn, flags.approxFunc());
value = bitEnumSet(value, FastmathFlags::reassoc, flags.allowReassoc());
FastmathFlagsAttr attr = FastmathFlagsAttr::get(builder.getContext(), value);
iface->setAttr(iface.getFastmathAttrName(), attr);
}
// We only need integers, floats, doubles, and vectors and tensors thereof for
// attributes. Scalar and vector types are converted to the standard
// equivalents. Array types are converted to ranked tensors; nested array types
// are converted to multi-dimensional tensors or vectors, depending on the
// innermost type being a scalar or a vector.
Type ModuleImport::getStdTypeForAttr(Type type) {
if (!type)
return nullptr;
if (type.isa<IntegerType, FloatType>())
return type;
// LLVM vectors can only contain scalars.
if (LLVM::isCompatibleVectorType(type)) {
llvm::ElementCount numElements = LLVM::getVectorNumElements(type);
if (numElements.isScalable()) {
emitError(UnknownLoc::get(context)) << "scalable vectors not supported";
return nullptr;
}
Type elementType = getStdTypeForAttr(LLVM::getVectorElementType(type));
if (!elementType)
return nullptr;
return VectorType::get(numElements.getKnownMinValue(), elementType);
}
// LLVM arrays can contain other arrays or vectors.
if (auto arrayType = type.dyn_cast<LLVMArrayType>()) {
// Recover the nested array shape.
SmallVector<int64_t, 4> shape;
shape.push_back(arrayType.getNumElements());
while (arrayType.getElementType().isa<LLVMArrayType>()) {
arrayType = arrayType.getElementType().cast<LLVMArrayType>();
shape.push_back(arrayType.getNumElements());
}
// If the innermost type is a vector, use the multi-dimensional vector as
// attribute type.
if (LLVM::isCompatibleVectorType(arrayType.getElementType())) {
llvm::ElementCount numElements =
LLVM::getVectorNumElements(arrayType.getElementType());
if (numElements.isScalable()) {
emitError(UnknownLoc::get(context)) << "scalable vectors not supported";
return nullptr;
}
shape.push_back(numElements.getKnownMinValue());
Type elementType = getStdTypeForAttr(
LLVM::getVectorElementType(arrayType.getElementType()));
if (!elementType)
return nullptr;
return VectorType::get(shape, elementType);
}
// Otherwise use a tensor.
Type elementType = getStdTypeForAttr(arrayType.getElementType());
if (!elementType)
return nullptr;
return RankedTensorType::get(shape, elementType);
}
return nullptr;
}
// Get the given constant as an attribute. Not all constants can be represented
// as attributes.
Attribute ModuleImport::getConstantAsAttr(llvm::Constant *value) {
if (auto *ci = dyn_cast<llvm::ConstantInt>(value))
return builder.getIntegerAttr(
IntegerType::get(context, ci->getType()->getBitWidth()),
ci->getValue());
if (auto *c = dyn_cast<llvm::ConstantDataArray>(value))
if (c->isString())
return builder.getStringAttr(c->getAsString());
if (auto *c = dyn_cast<llvm::ConstantFP>(value)) {
llvm::Type *type = c->getType();
FloatType floatTy;
if (type->isBFloatTy())
floatTy = FloatType::getBF16(context);
else
floatTy = getDLFloatType(*context, type->getScalarSizeInBits());
assert(floatTy && "unsupported floating point type");
return builder.getFloatAttr(floatTy, c->getValueAPF());
}
if (auto *f = dyn_cast<llvm::Function>(value))
return SymbolRefAttr::get(builder.getContext(), f->getName());
// Convert constant data to a dense elements attribute.
if (auto *cd = dyn_cast<llvm::ConstantDataSequential>(value)) {
Type type = convertType(cd->getElementType());
auto attrType = getStdTypeForAttr(convertType(cd->getType()))
.dyn_cast_or_null<ShapedType>();
if (!attrType)
return nullptr;
if (type.isa<IntegerType>()) {
SmallVector<APInt, 8> values;
values.reserve(cd->getNumElements());
for (unsigned i = 0, e = cd->getNumElements(); i < e; ++i)
values.push_back(cd->getElementAsAPInt(i));
return DenseElementsAttr::get(attrType, values);
}
if (type.isa<Float32Type, Float64Type>()) {
SmallVector<APFloat, 8> values;
values.reserve(cd->getNumElements());
for (unsigned i = 0, e = cd->getNumElements(); i < e; ++i)
values.push_back(cd->getElementAsAPFloat(i));
return DenseElementsAttr::get(attrType, values);
}
return nullptr;
}
// Unpack constant aggregates to create dense elements attribute whenever
// possible. Return nullptr (failure) otherwise.
if (isa<llvm::ConstantAggregate>(value)) {
auto outerType = getStdTypeForAttr(convertType(value->getType()))
.dyn_cast_or_null<ShapedType>();
if (!outerType)
return nullptr;
SmallVector<Attribute, 8> values;
SmallVector<int64_t, 8> shape;
for (unsigned i = 0, e = value->getNumOperands(); i < e; ++i) {
auto nested = getConstantAsAttr(value->getAggregateElement(i))
.dyn_cast_or_null<DenseElementsAttr>();
if (!nested)
return nullptr;
values.append(nested.value_begin<Attribute>(),
nested.value_end<Attribute>());
}
return DenseElementsAttr::get(outerType, values);
}
return nullptr;
}
LogicalResult ModuleImport::convertGlobal(llvm::GlobalVariable *globalVar) {
// Insert the global after the last one or at the start of the module.
OpBuilder::InsertionGuard guard(builder);
if (!globalInsertionOp)
builder.setInsertionPointToStart(mlirModule.getBody());
else
builder.setInsertionPointAfter(globalInsertionOp);
Attribute valueAttr;
if (globalVar->hasInitializer())
valueAttr = getConstantAsAttr(globalVar->getInitializer());
Type type = convertType(globalVar->getValueType());
uint64_t alignment = 0;
llvm::MaybeAlign maybeAlign = globalVar->getAlign();
if (maybeAlign.has_value()) {
llvm::Align align = *maybeAlign;
alignment = align.value();
}
GlobalOp globalOp = builder.create<GlobalOp>(
mlirModule.getLoc(), type, globalVar->isConstant(),
convertLinkageFromLLVM(globalVar->getLinkage()), globalVar->getName(),
valueAttr, alignment, /*addr_space=*/globalVar->getAddressSpace(),
/*dso_local=*/globalVar->isDSOLocal(),
/*thread_local=*/globalVar->isThreadLocal());
globalInsertionOp = globalOp;
if (globalVar->hasInitializer() && !valueAttr) {
clearBlockAndValueMapping();
Block *block = builder.createBlock(&globalOp.getInitializerRegion());
setConstantInsertionPointToStart(block);
FailureOr<Value> initializer =
convertConstantExpr(globalVar->getInitializer());
if (failed(initializer))
return failure();
builder.create<ReturnOp>(globalOp.getLoc(), *initializer);
}
if (globalVar->hasAtLeastLocalUnnamedAddr()) {
globalOp.setUnnamedAddr(
convertUnnamedAddrFromLLVM(globalVar->getUnnamedAddr()));
}
if (globalVar->hasSection())
globalOp.setSection(globalVar->getSection());
globalOp.setVisibility_(
convertVisibilityFromLLVM(globalVar->getVisibility()));
return success();
}
LogicalResult
ModuleImport::convertGlobalCtorsAndDtors(llvm::GlobalVariable *globalVar) {
if (!globalVar->hasInitializer() || !globalVar->hasAppendingLinkage())
return failure();
auto *initializer =
dyn_cast<llvm::ConstantArray>(globalVar->getInitializer());
if (!initializer)
return failure();
SmallVector<Attribute> funcs;
SmallVector<int32_t> priorities;
for (llvm::Value *operand : initializer->operands()) {
auto *aggregate = dyn_cast<llvm::ConstantAggregate>(operand);
if (!aggregate || aggregate->getNumOperands() != 3)
return failure();
auto *priority = dyn_cast<llvm::ConstantInt>(aggregate->getOperand(0));
auto *func = dyn_cast<llvm::Function>(aggregate->getOperand(1));
auto *data = dyn_cast<llvm::Constant>(aggregate->getOperand(2));
if (!priority || !func || !data)
return failure();
// GlobalCtorsOps and GlobalDtorsOps do not support non-null data fields.
if (!data->isNullValue())
return failure();
funcs.push_back(FlatSymbolRefAttr::get(context, func->getName()));
priorities.push_back(priority->getValue().getZExtValue());
}
OpBuilder::InsertionGuard guard(builder);
if (!globalInsertionOp)
builder.setInsertionPointToStart(mlirModule.getBody());
else
builder.setInsertionPointAfter(globalInsertionOp);
if (globalVar->getName() == getGlobalCtorsVarName()) {
globalInsertionOp = builder.create<LLVM::GlobalCtorsOp>(
mlirModule.getLoc(), builder.getArrayAttr(funcs),
builder.getI32ArrayAttr(priorities));
return success();
}
globalInsertionOp = builder.create<LLVM::GlobalDtorsOp>(
mlirModule.getLoc(), builder.getArrayAttr(funcs),
builder.getI32ArrayAttr(priorities));
return success();
}
SetVector<llvm::Constant *>
ModuleImport::getConstantsToConvert(llvm::Constant *constant) {
// Return the empty set if the constant has been translated before.
if (valueMapping.count(constant))
return {};
// Traverse the constants in post-order and stop the traversal if a constant
// already has a `valueMapping` from an earlier constant translation or if the
// constant is traversed a second time.
SetVector<llvm::Constant *> orderedSet;
SetVector<llvm::Constant *> workList;
DenseMap<llvm::Constant *, SmallVector<llvm::Constant *>> adjacencyLists;
workList.insert(constant);
while (!workList.empty()) {
llvm::Constant *current = workList.back();
// Collect all dependencies of the current constant and add them to the
// adjacency list if none has been computed before.
auto adjacencyIt = adjacencyLists.find(current);
if (adjacencyIt == adjacencyLists.end()) {
adjacencyIt = adjacencyLists.try_emplace(current).first;
// Add all constant operands to the adjacency list and skip any other
// values such as basic block addresses.
for (llvm::Value *operand : current->operands())
if (auto *constDependency = dyn_cast<llvm::Constant>(operand))
adjacencyIt->getSecond().push_back(constDependency);
// Use the getElementValue method to add the dependencies of zero
// initialized aggregate constants since they do not take any operands.
if (auto *constAgg = dyn_cast<llvm::ConstantAggregateZero>(current)) {
unsigned numElements = constAgg->getElementCount().getFixedValue();
for (unsigned i = 0, e = numElements; i != e; ++i)
adjacencyIt->getSecond().push_back(constAgg->getElementValue(i));
}
}
// Add the current constant to the `orderedSet` of the traversed nodes if
// all its dependencies have been traversed before. Additionally, remove the
// constant from the `workList` and continue the traversal.
if (adjacencyIt->getSecond().empty()) {
orderedSet.insert(current);
workList.pop_back();
continue;
}
// Add the next dependency from the adjacency list to the `workList` and
// continue the traversal. Remove the dependency from the adjacency list to
// mark that it has been processed. Only enqueue the dependency if it has no
// `valueMapping` from an earlier translation and if it has not been
// enqueued before.
llvm::Constant *dependency = adjacencyIt->getSecond().pop_back_val();
if (valueMapping.count(dependency) || workList.count(dependency) ||
orderedSet.count(dependency))
continue;
workList.insert(dependency);