1//===- CoroFrame.cpp - Builds and manipulates coroutine frame -------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8// This file contains classes used to discover if for a particular value
9// its definition precedes and its uses follow a suspend block. This is
10// referred to as a suspend crossing value.
11//
12// Using the information discovered we form a Coroutine Frame structure to
13// contain those values. All uses of those values are replaced with appropriate
14// GEP + load from the coroutine frame. At the point of the definition we spill
15// the value into the coroutine frame.
16//===----------------------------------------------------------------------===//
17
18#include "CoroInternal.h"
19#include "llvm/ADT/ScopeExit.h"
20#include "llvm/ADT/SmallString.h"
21#include "llvm/Analysis/StackLifetime.h"
22#include "llvm/IR/DIBuilder.h"
23#include "llvm/IR/DebugInfo.h"
24#include "llvm/IR/Dominators.h"
25#include "llvm/IR/IRBuilder.h"
26#include "llvm/IR/InstIterator.h"
27#include "llvm/IR/IntrinsicInst.h"
28#include "llvm/IR/MDBuilder.h"
29#include "llvm/IR/Module.h"
30#include "llvm/Support/Compiler.h"
31#include "llvm/Support/Debug.h"
32#include "llvm/Support/OptimizedStructLayout.h"
33#include "llvm/Transforms/Coroutines/ABI.h"
34#include "llvm/Transforms/Coroutines/CoroInstr.h"
35#include "llvm/Transforms/Coroutines/MaterializationUtils.h"
36#include "llvm/Transforms/Coroutines/SpillUtils.h"
37#include "llvm/Transforms/Coroutines/SuspendCrossingInfo.h"
38#include "llvm/Transforms/Utils/BasicBlockUtils.h"
39#include "llvm/Transforms/Utils/Local.h"
40#include "llvm/Transforms/Utils/PromoteMemToReg.h"
41#include <algorithm>
42#include <optional>
43
44using namespace llvm;
45
46#define DEBUG_TYPE "coro-frame"
47
48namespace {
49class FrameTypeBuilder;
50// Mapping from the to-be-spilled value to all the users that need reload.
51struct FrameDataInfo {
52 // All the values (that are not allocas) that needs to be spilled to the
53 // frame.
54 coro::SpillInfo &Spills;
55 // Allocas contains all values defined as allocas that need to live in the
56 // frame.
57 SmallVectorImpl<coro::AllocaInfo> &Allocas;
58
59 FrameDataInfo(coro::SpillInfo &Spills,
60 SmallVectorImpl<coro::AllocaInfo> &Allocas)
61 : Spills(Spills), Allocas(Allocas) {}
62
63 SmallVector<Value *, 8> getAllDefs() const {
64 SmallVector<Value *, 8> Defs;
65 for (const auto &P : Spills)
66 Defs.push_back(Elt: P.first);
67 for (const auto &A : Allocas)
68 Defs.push_back(Elt: A.Alloca);
69 return Defs;
70 }
71
72 uint32_t getFieldIndex(Value *V) const {
73 auto Itr = FieldIndexMap.find(Val: V);
74 assert(Itr != FieldIndexMap.end() &&
75 "Value does not have a frame field index");
76 return Itr->second;
77 }
78
79 void setFieldIndex(Value *V, uint32_t Index) {
80 assert(FieldIndexMap.count(V) == 0 &&
81 "Cannot set the index for the same field twice.");
82 FieldIndexMap[V] = Index;
83 }
84
85 Align getAlign(Value *V) const {
86 auto Iter = FieldAlignMap.find(Val: V);
87 assert(Iter != FieldAlignMap.end());
88 return Iter->second;
89 }
90
91 void setAlign(Value *V, Align AL) {
92 assert(FieldAlignMap.count(V) == 0);
93 FieldAlignMap.insert(KV: {V, AL});
94 }
95
96 uint64_t getDynamicAlign(Value *V) const {
97 auto Iter = FieldDynamicAlignMap.find(Val: V);
98 assert(Iter != FieldDynamicAlignMap.end());
99 return Iter->second;
100 }
101
102 void setDynamicAlign(Value *V, uint64_t Align) {
103 assert(FieldDynamicAlignMap.count(V) == 0);
104 FieldDynamicAlignMap.insert(KV: {V, Align});
105 }
106
107 uint64_t getOffset(Value *V) const {
108 auto Iter = FieldOffsetMap.find(Val: V);
109 assert(Iter != FieldOffsetMap.end());
110 return Iter->second;
111 }
112
113 void setOffset(Value *V, uint64_t Offset) {
114 assert(FieldOffsetMap.count(V) == 0);
115 FieldOffsetMap.insert(KV: {V, Offset});
116 }
117
118 // Update field offset and alignment information from FrameTypeBuilder.
119 void updateLayoutInfo(FrameTypeBuilder &B);
120
121private:
122 // Map from values to their slot indexes on the frame (insertion order).
123 DenseMap<Value *, uint32_t> FieldIndexMap;
124 // Map from values to their alignment on the frame. They would be set after
125 // the frame is built.
126 DenseMap<Value *, Align> FieldAlignMap;
127 DenseMap<Value *, uint64_t> FieldDynamicAlignMap;
128 // Map from values to their offset on the frame. They would be set after
129 // the frame is built.
130 DenseMap<Value *, uint64_t> FieldOffsetMap;
131};
132} // namespace
133
134#ifndef NDEBUG
135static void dumpSpills(StringRef Title, const coro::SpillInfo &Spills) {
136 dbgs() << "------------- " << Title << " --------------\n";
137 for (const auto &E : Spills) {
138 E.first->dump();
139 dbgs() << " user: ";
140 for (auto *I : E.second)
141 I->dump();
142 }
143}
144
145static void dumpAllocas(const SmallVectorImpl<coro::AllocaInfo> &Allocas) {
146 dbgs() << "------------- Allocas --------------\n";
147 for (const auto &A : Allocas) {
148 A.Alloca->dump();
149 }
150}
151#endif
152
153namespace {
154using FieldIDType = size_t;
155// We cannot rely solely on natural alignment of a type when building a
156// coroutine frame and if the alignment specified on the Alloca instruction
157// differs from the natural alignment of the alloca type we will need to insert
158// padding.
159class FrameTypeBuilder {
160private:
161 struct Field {
162 uint64_t Size;
163 uint64_t Offset;
164 Align Alignment;
165 uint64_t DynamicAlignBuffer;
166 };
167
168 const DataLayout &DL;
169 uint64_t StructSize = 0;
170 Align StructAlign;
171 bool IsFinished = false;
172
173 std::optional<Align> MaxFrameAlignment;
174
175 SmallVector<Field, 8> Fields;
176 DenseMap<Value*, unsigned> FieldIndexByKey;
177
178public:
179 FrameTypeBuilder(const DataLayout &DL, std::optional<Align> MaxFrameAlignment)
180 : DL(DL), MaxFrameAlignment(MaxFrameAlignment) {}
181
182 /// Add a field to this structure for the storage of an `alloca`
183 /// instruction.
184 [[nodiscard]] FieldIDType addFieldForAlloca(AllocaInst *AI,
185 bool IsHeader = false) {
186 auto Size = AI->getAllocationSize(DL: AI->getDataLayout());
187 if (!Size || !Size->isFixed())
188 report_fatal_error(
189 reason: "Coroutines cannot handle non static or vscale allocas yet");
190 return addField(FieldSize: Size->getFixedValue(), FieldAlignment: AI->getAlign(), IsHeader);
191 }
192
193 /// We want to put the allocas whose lifetime-ranges are not overlapped
194 /// into one slot of coroutine frame.
195 /// Consider the example at:https://bugs.llvm.org/show_bug.cgi?id=45566
196 ///
197 /// cppcoro::task<void> alternative_paths(bool cond) {
198 /// if (cond) {
199 /// big_structure a;
200 /// process(a);
201 /// co_await something();
202 /// } else {
203 /// big_structure b;
204 /// process2(b);
205 /// co_await something();
206 /// }
207 /// }
208 ///
209 /// We want to put variable a and variable b in the same slot to
210 /// reduce the size of coroutine frame.
211 ///
212 /// This function use StackLifetime algorithm to partition the AllocaInsts in
213 /// Spills to non-overlapped sets in order to put Alloca in the same
214 /// non-overlapped set into the same slot in the Coroutine Frame. Then add
215 /// field for the allocas in the same non-overlapped set by using the largest
216 /// type as the field type.
217 ///
218 /// Side Effects: Because We sort the allocas, the order of allocas in the
219 /// frame may be different with the order in the source code.
220 void addFieldForAllocas(const Function &F, FrameDataInfo &FrameData,
221 coro::Shape &Shape, bool OptimizeFrame);
222
223 /// Add a field to this structure for a spill.
224 [[nodiscard]] FieldIDType addField(Type *Ty, MaybeAlign MaybeFieldAlignment,
225 bool IsHeader = false,
226 bool IsSpillOfValue = false) {
227 assert(Ty && "must provide a type for a field");
228 // The field size is the alloc size of the type.
229 uint64_t FieldSize = DL.getTypeAllocSize(Ty);
230 // The field alignment is usually the type alignment.
231 // But if we are spilling values we don't need to worry about ABI alignment
232 // concerns.
233 Align ABIAlign = DL.getABITypeAlign(Ty);
234 Align TyAlignment = ABIAlign;
235 if (IsSpillOfValue && MaxFrameAlignment && *MaxFrameAlignment < ABIAlign)
236 TyAlignment = *MaxFrameAlignment;
237 Align FieldAlignment = MaybeFieldAlignment.value_or(u&: TyAlignment);
238 return addField(FieldSize, FieldAlignment, IsHeader);
239 }
240
241 /// Add a field to this structure.
242 [[nodiscard]] FieldIDType addField(uint64_t FieldSize, Align FieldAlignment,
243 bool IsHeader = false) {
244 assert(!IsFinished && "adding fields to a finished builder");
245
246 // For an alloca with size=0, we don't need to add a field and they
247 // can just point to any index in the frame. Use index 0.
248 if (FieldSize == 0)
249 return 0;
250
251 // The field alignment could be bigger than the max frame case, in that case
252 // we request additional storage to be able to dynamically align the
253 // pointer.
254 uint64_t DynamicAlignBuffer = 0;
255 if (MaxFrameAlignment && (FieldAlignment > *MaxFrameAlignment)) {
256 DynamicAlignBuffer =
257 offsetToAlignment(Value: MaxFrameAlignment->value(), Alignment: FieldAlignment);
258 FieldAlignment = *MaxFrameAlignment;
259 FieldSize = FieldSize + DynamicAlignBuffer;
260 }
261
262 // Lay out header fields immediately.
263 uint64_t Offset;
264 if (IsHeader) {
265 Offset = alignTo(Size: StructSize, A: FieldAlignment);
266 StructSize = Offset + FieldSize;
267
268 // Everything else has a flexible offset.
269 } else {
270 Offset = OptimizedStructLayoutField::FlexibleOffset;
271 }
272
273 Fields.push_back(Elt: {.Size: FieldSize, .Offset: Offset, .Alignment: FieldAlignment, .DynamicAlignBuffer: DynamicAlignBuffer});
274 return Fields.size() - 1;
275 }
276
277 /// Finish the layout and compute final size and alignment.
278 void finish();
279
280 uint64_t getStructSize() const {
281 assert(IsFinished && "not yet finished!");
282 return StructSize;
283 }
284
285 Align getStructAlign() const {
286 assert(IsFinished && "not yet finished!");
287 return StructAlign;
288 }
289
290 Field getLayoutField(FieldIDType Id) const {
291 assert(IsFinished && "not yet finished!");
292 return Fields[Id];
293 }
294};
295} // namespace
296
297void FrameDataInfo::updateLayoutInfo(FrameTypeBuilder &B) {
298 auto Updater = [&](Value *I) {
299 uint32_t FieldIndex = getFieldIndex(V: I);
300 auto Field = B.getLayoutField(Id: FieldIndex);
301 setAlign(V: I, AL: Field.Alignment);
302 uint64_t dynamicAlign =
303 Field.DynamicAlignBuffer
304 ? Field.DynamicAlignBuffer + Field.Alignment.value()
305 : 0;
306 setDynamicAlign(V: I, Align: dynamicAlign);
307 setOffset(V: I, Offset: Field.Offset);
308 };
309 for (auto &S : Spills)
310 Updater(S.first);
311 for (const auto &A : Allocas)
312 Updater(A.Alloca);
313}
314
315void FrameTypeBuilder::addFieldForAllocas(const Function &F,
316 FrameDataInfo &FrameData,
317 coro::Shape &Shape,
318 bool OptimizeFrame) {
319 using AllocaSetType = SmallVector<AllocaInst *, 4>;
320 SmallVector<AllocaSetType, 4> NonOverlapedAllocas;
321
322 // We need to add field for allocas at the end of this function.
323 llvm::scope_exit AddFieldForAllocasAtExit([&]() {
324 for (auto AllocaList : NonOverlapedAllocas) {
325 auto *LargestAI = *AllocaList.begin();
326 FieldIDType Id = addFieldForAlloca(AI: LargestAI);
327 for (auto *Alloca : AllocaList)
328 FrameData.setFieldIndex(V: Alloca, Index: Id);
329 }
330 });
331
332 if (!OptimizeFrame) {
333 for (const auto &A : FrameData.Allocas) {
334 AllocaInst *Alloca = A.Alloca;
335 NonOverlapedAllocas.emplace_back(Args: AllocaSetType(1, Alloca));
336 }
337 return;
338 }
339
340 // Because there are paths from the lifetime.start to coro.end
341 // for each alloca, the liferanges for every alloca is overlaped
342 // in the blocks who contain coro.end and the successor blocks.
343 // So we choose to skip there blocks when we calculate the liferange
344 // for each alloca. It should be reasonable since there shouldn't be uses
345 // in these blocks and the coroutine frame shouldn't be used outside the
346 // coroutine body.
347 //
348 // Note that the user of coro.suspend may not be SwitchInst. However, this
349 // case seems too complex to handle. And it is harmless to skip these
350 // patterns since it just prevend putting the allocas to live in the same
351 // slot.
352 DenseMap<SwitchInst *, BasicBlock *> DefaultSuspendDest;
353 for (auto *CoroSuspendInst : Shape.CoroSuspends) {
354 for (auto *U : CoroSuspendInst->users()) {
355 if (auto *ConstSWI = dyn_cast<SwitchInst>(Val: U)) {
356 auto *SWI = const_cast<SwitchInst *>(ConstSWI);
357 DefaultSuspendDest[SWI] = SWI->getDefaultDest();
358 SWI->setDefaultDest(SWI->getSuccessor(idx: 1));
359 }
360 }
361 }
362
363 auto ExtractAllocas = [&]() {
364 AllocaSetType Allocas;
365 Allocas.reserve(N: FrameData.Allocas.size());
366 for (const auto &A : FrameData.Allocas)
367 Allocas.push_back(Elt: A.Alloca);
368 return Allocas;
369 };
370 StackLifetime StackLifetimeAnalyzer(F, ExtractAllocas(),
371 StackLifetime::LivenessType::May);
372 StackLifetimeAnalyzer.run();
373 auto DoAllocasInterfere = [&](const AllocaInst *AI1, const AllocaInst *AI2) {
374 return StackLifetimeAnalyzer.getLiveRange(AI: AI1).overlaps(
375 Other: StackLifetimeAnalyzer.getLiveRange(AI: AI2));
376 };
377 auto GetAllocaSize = [&](const coro::AllocaInfo &A) {
378 std::optional<TypeSize> RetSize = A.Alloca->getAllocationSize(DL);
379 assert(RetSize && "Variable Length Arrays (VLA) are not supported.\n");
380 assert(!RetSize->isScalable() && "Scalable vectors are not yet supported");
381 return RetSize->getFixedValue();
382 };
383 // Put larger allocas in the front. So the larger allocas have higher
384 // priority to merge, which can save more space potentially. Also each
385 // AllocaSet would be ordered. So we can get the largest Alloca in one
386 // AllocaSet easily.
387 sort(C&: FrameData.Allocas, Comp: [&](const auto &Iter1, const auto &Iter2) {
388 return GetAllocaSize(Iter1) > GetAllocaSize(Iter2);
389 });
390 for (const auto &A : FrameData.Allocas) {
391 AllocaInst *Alloca = A.Alloca;
392 bool Merged = false;
393 // Try to find if the Alloca does not interfere with any existing
394 // NonOverlappedAllocaSet. If it is true, insert the alloca to that
395 // NonOverlappedAllocaSet.
396 for (auto &AllocaSet : NonOverlapedAllocas) {
397 assert(!AllocaSet.empty() && "Processing Alloca Set is not empty.\n");
398 bool NoInterference = none_of(Range&: AllocaSet, P: [&](auto Iter) {
399 return DoAllocasInterfere(Alloca, Iter);
400 });
401 // If the alignment of A is multiple of the alignment of B, the address
402 // of A should satisfy the requirement for aligning for B.
403 //
404 // There may be other more fine-grained strategies to handle the alignment
405 // infomation during the merging process. But it seems hard to handle
406 // these strategies and benefit little.
407 bool Alignable = [&]() -> bool {
408 auto *LargestAlloca = *AllocaSet.begin();
409 return LargestAlloca->getAlign().value() % Alloca->getAlign().value() ==
410 0;
411 }();
412 bool CouldMerge = NoInterference && Alignable;
413 if (!CouldMerge)
414 continue;
415 AllocaSet.push_back(Elt: Alloca);
416 Merged = true;
417 break;
418 }
419 if (!Merged) {
420 NonOverlapedAllocas.emplace_back(Args: AllocaSetType(1, Alloca));
421 }
422 }
423 // Recover the default target destination for each Switch statement
424 // reserved.
425 for (auto SwitchAndDefaultDest : DefaultSuspendDest) {
426 SwitchInst *SWI = SwitchAndDefaultDest.first;
427 BasicBlock *DestBB = SwitchAndDefaultDest.second;
428 SWI->setDefaultDest(DestBB);
429 }
430 // This Debug Info could tell us which allocas are merged into one slot.
431 LLVM_DEBUG(for (auto &AllocaSet
432 : NonOverlapedAllocas) {
433 if (AllocaSet.size() > 1) {
434 dbgs() << "In Function:" << F.getName() << "\n";
435 dbgs() << "Find Union Set "
436 << "\n";
437 dbgs() << "\tAllocas are \n";
438 for (auto Alloca : AllocaSet)
439 dbgs() << "\t\t" << *Alloca << "\n";
440 }
441 });
442}
443
444void FrameTypeBuilder::finish() {
445 assert(!IsFinished && "already finished!");
446
447 // Prepare the optimal-layout field array.
448 // The Id in the layout field is a pointer to our Field for it.
449 SmallVector<OptimizedStructLayoutField, 8> LayoutFields;
450 LayoutFields.reserve(N: Fields.size());
451 for (auto &Field : Fields) {
452 LayoutFields.emplace_back(Args: &Field, Args&: Field.Size, Args&: Field.Alignment,
453 Args&: Field.Offset);
454 }
455
456 // Perform layout to compute size, alignment, and field offsets.
457 auto SizeAndAlign = performOptimizedStructLayout(Fields: LayoutFields);
458 StructSize = SizeAndAlign.first;
459 StructAlign = SizeAndAlign.second;
460
461 auto getField = [](const OptimizedStructLayoutField &LayoutField) -> Field & {
462 return *static_cast<Field *>(const_cast<void*>(LayoutField.Id));
463 };
464
465 // Update field offsets from the computed layout.
466 for (auto &LayoutField : LayoutFields) {
467 auto &F = getField(LayoutField);
468 F.Offset = LayoutField.Offset;
469 }
470
471 IsFinished = true;
472}
473
474static void cacheDIVar(FrameDataInfo &FrameData,
475 DenseMap<Value *, DILocalVariable *> &DIVarCache) {
476 for (auto *V : FrameData.getAllDefs()) {
477 if (DIVarCache.contains(Val: V))
478 continue;
479
480 auto CacheIt = [&DIVarCache, V](const auto &Container) {
481 auto *I = llvm::find_if(Container, [](auto *DDI) {
482 return DDI->getExpression()->getNumElements() == 0;
483 });
484 if (I != Container.end())
485 DIVarCache.insert({V, (*I)->getVariable()});
486 };
487 CacheIt(findDVRDeclares(V));
488 CacheIt(findDVRDeclareValues(V));
489 }
490}
491
492/// Create name for Type. It uses MDString to store new created string to
493/// avoid memory leak.
494static StringRef solveTypeName(Type *Ty) {
495 if (Ty->isIntegerTy()) {
496 // The longest name in common may be '__int_128', which has 9 bits.
497 SmallString<16> Buffer;
498 raw_svector_ostream OS(Buffer);
499 OS << "__int_" << cast<IntegerType>(Val: Ty)->getBitWidth();
500 auto *MDName = MDString::get(Context&: Ty->getContext(), Str: OS.str());
501 return MDName->getString();
502 }
503
504 if (Ty->isFloatingPointTy()) {
505 if (Ty->isFloatTy())
506 return "__float_";
507 if (Ty->isDoubleTy())
508 return "__double_";
509 return "__floating_type_";
510 }
511
512 if (Ty->isPointerTy())
513 return "PointerType";
514
515 if (Ty->isStructTy()) {
516 if (!cast<StructType>(Val: Ty)->hasName())
517 return "__LiteralStructType_";
518
519 auto Name = Ty->getStructName();
520
521 SmallString<16> Buffer(Name);
522 for (auto &Iter : Buffer)
523 if (Iter == '.' || Iter == ':')
524 Iter = '_';
525 auto *MDName = MDString::get(Context&: Ty->getContext(), Str: Buffer.str());
526 return MDName->getString();
527 }
528
529 return "UnknownType";
530}
531
532static DIType *solveDIType(DIBuilder &Builder, Type *Ty,
533 const DataLayout &Layout, DIScope *Scope,
534 unsigned LineNum,
535 DenseMap<Type *, DIType *> &DITypeCache) {
536 if (DIType *DT = DITypeCache.lookup(Val: Ty))
537 return DT;
538
539 StringRef Name = solveTypeName(Ty);
540
541 DIType *RetType = nullptr;
542
543 if (Ty->isIntegerTy()) {
544 auto BitWidth = cast<IntegerType>(Val: Ty)->getBitWidth();
545 RetType = Builder.createBasicType(Name, SizeInBits: BitWidth, Encoding: dwarf::DW_ATE_signed,
546 Flags: llvm::DINode::FlagArtificial);
547 } else if (Ty->isFloatingPointTy()) {
548 RetType = Builder.createBasicType(Name, SizeInBits: Layout.getTypeSizeInBits(Ty),
549 Encoding: dwarf::DW_ATE_float,
550 Flags: llvm::DINode::FlagArtificial);
551 } else if (Ty->isPointerTy()) {
552 // Construct PointerType points to null (aka void *) instead of exploring
553 // pointee type to avoid infinite search problem. For example, we would be
554 // in trouble if we traverse recursively:
555 //
556 // struct Node {
557 // Node* ptr;
558 // };
559 RetType =
560 Builder.createPointerType(PointeeTy: nullptr, SizeInBits: Layout.getTypeSizeInBits(Ty),
561 AlignInBits: Layout.getABITypeAlign(Ty).value() * CHAR_BIT,
562 /*DWARFAddressSpace=*/std::nullopt, Name);
563 } else if (Ty->isStructTy()) {
564 auto *DIStruct = Builder.createStructType(
565 Scope, Name, File: Scope->getFile(), LineNumber: LineNum, SizeInBits: Layout.getTypeSizeInBits(Ty),
566 AlignInBits: Layout.getPrefTypeAlign(Ty).value() * CHAR_BIT,
567 Flags: llvm::DINode::FlagArtificial, DerivedFrom: nullptr, Elements: llvm::DINodeArray());
568
569 auto *StructTy = cast<StructType>(Val: Ty);
570 SmallVector<Metadata *, 16> Elements;
571 for (unsigned I = 0; I < StructTy->getNumElements(); I++) {
572 DIType *DITy = solveDIType(Builder, Ty: StructTy->getElementType(N: I), Layout,
573 Scope: DIStruct, LineNum, DITypeCache);
574 assert(DITy);
575 Elements.push_back(Elt: Builder.createMemberType(
576 Scope: DIStruct, Name: DITy->getName(), File: DIStruct->getFile(), LineNo: LineNum,
577 SizeInBits: DITy->getSizeInBits(), AlignInBits: DITy->getAlignInBits(),
578 OffsetInBits: Layout.getStructLayout(Ty: StructTy)->getElementOffsetInBits(Idx: I),
579 Flags: llvm::DINode::FlagArtificial, Ty: DITy));
580 }
581
582 Builder.replaceArrays(T&: DIStruct, Elements: Builder.getOrCreateArray(Elements));
583
584 RetType = DIStruct;
585 } else {
586 LLVM_DEBUG(dbgs() << "Unresolved Type: " << *Ty << "\n");
587 TypeSize Size = Layout.getTypeSizeInBits(Ty);
588 auto *CharSizeType = Builder.createBasicType(
589 Name, SizeInBits: 8, Encoding: dwarf::DW_ATE_unsigned_char, Flags: llvm::DINode::FlagArtificial);
590
591 if (Size <= 8)
592 RetType = CharSizeType;
593 else {
594 if (Size % 8 != 0)
595 Size = TypeSize::getFixed(ExactSize: Size + 8 - (Size % 8));
596
597 RetType = Builder.createArrayType(
598 Size, AlignInBits: Layout.getPrefTypeAlign(Ty).value(), Ty: CharSizeType,
599 Subscripts: Builder.getOrCreateArray(Elements: Builder.getOrCreateSubrange(Lo: 0, Count: Size / 8)));
600 }
601 }
602
603 DITypeCache.insert(KV: {Ty, RetType});
604 return RetType;
605}
606
607/// Build artificial debug info for C++ coroutine frames to allow users to
608/// inspect the contents of the frame directly
609///
610/// Create Debug information for coroutine frame with debug name "__coro_frame".
611/// The debug information for the fields of coroutine frame is constructed from
612/// the following way:
613/// 1. For all the value in the Frame, we search the use of dbg.declare to find
614/// the corresponding debug variables for the value. If we can find the
615/// debug variable, we can get full and accurate debug information.
616/// 2. If we can't get debug information in step 1 and 2, we could only try to
617/// build the DIType by Type. We did this in solveDIType. We only handle
618/// integer, float, double, integer type and struct type for now.
619static void buildFrameDebugInfo(Function &F, coro::Shape &Shape,
620 FrameDataInfo &FrameData) {
621 DISubprogram *DIS = F.getSubprogram();
622 // If there is no DISubprogram for F, it implies the function is compiled
623 // without debug info. So we also don't generate debug info for the frame.
624
625 if (!DIS || !DIS->getUnit())
626 return;
627
628 if (!dwarf::isCPlusPlus(S: static_cast<llvm::dwarf::SourceLanguage>(
629 DIS->getUnit()->getSourceLanguage().getUnversionedName())) ||
630 DIS->getUnit()->getEmissionKind() !=
631 DICompileUnit::DebugEmissionKind::FullDebug)
632 return;
633
634 assert(Shape.ABI == coro::ABI::Switch &&
635 "We could only build debug infomation for C++ coroutine now.\n");
636
637 DIBuilder DBuilder(*F.getParent(), /*AllowUnresolved*/ false);
638
639 DIFile *DFile = DIS->getFile();
640 unsigned LineNum = DIS->getLine();
641
642 DICompositeType *FrameDITy = DBuilder.createStructType(
643 Scope: DIS->getUnit(), Name: Twine(F.getName() + ".coro_frame_ty").str(), File: DFile,
644 LineNumber: LineNum, SizeInBits: Shape.FrameSize * 8, AlignInBits: Shape.FrameAlign.value() * 8,
645 Flags: llvm::DINode::FlagArtificial, DerivedFrom: nullptr, Elements: llvm::DINodeArray());
646 SmallVector<Metadata *, 16> Elements;
647 DataLayout Layout = F.getDataLayout();
648
649 DenseMap<Value *, DILocalVariable *> DIVarCache;
650 cacheDIVar(FrameData, DIVarCache);
651
652 // This counter is used to avoid same type names. e.g., there would be
653 // many i32 and i64 types in one coroutine. And we would use i32_0 and
654 // i32_1 to avoid the same type. Since it makes no sense the name of the
655 // fields confilicts with each other.
656 unsigned UnknownTypeNum = 0;
657 DenseMap<Type *, DIType *> DITypeCache;
658
659 auto addElement = [&](StringRef Name, uint64_t SizeInBits, uint64_t Alignment,
660 uint64_t Offset, DIType *DITy) {
661 Elements.push_back(Elt: DBuilder.createMemberType(
662 Scope: FrameDITy, Name, File: DFile, LineNo: LineNum, SizeInBits, AlignInBits: Alignment, OffsetInBits: Offset * 8,
663 Flags: llvm::DINode::FlagArtificial, Ty: DITy));
664 };
665
666 auto addDIDef = [&](Value *V) {
667 // Get the offset and alignment for this value.
668 uint64_t Offset = FrameData.getOffset(V);
669 Align Alignment = FrameData.getAlign(V);
670
671 std::string Name;
672 uint64_t SizeInBits;
673 DIType *DITy = nullptr;
674
675 auto It = DIVarCache.find(Val: V);
676 if (It != DIVarCache.end()) {
677 // Get the type from the debug variable.
678 Name = It->second->getName().str();
679 DITy = It->second->getType();
680 } else {
681 if (auto AI = dyn_cast<AllocaInst>(Val: V)) {
682 // Frame alloca
683 DITy = solveDIType(Builder&: DBuilder, Ty: AI->getAllocatedType(), Layout, Scope: FrameDITy,
684 LineNum, DITypeCache);
685 } else {
686 // Spill
687 DITy = solveDIType(Builder&: DBuilder, Ty: V->getType(), Layout, Scope: FrameDITy, LineNum,
688 DITypeCache);
689 }
690 assert(DITy && "SolveDIType shouldn't return nullptr.\n");
691 Name = DITy->getName().str();
692 Name += "_" + std::to_string(val: UnknownTypeNum);
693 UnknownTypeNum++;
694 }
695
696 if (auto AI = dyn_cast<AllocaInst>(Val: V)) {
697 // Lookup the total size of this alloca originally
698 auto Size = AI->getAllocationSize(DL: Layout);
699 assert(Size && Size->isFixed() &&
700 "unreachable due to addFieldForAlloca checks");
701 SizeInBits = Size->getFixedValue() * 8;
702 } else {
703 // Compute the size of the active data of this member for this spill
704 SizeInBits = Layout.getTypeSizeInBits(Ty: V->getType());
705 }
706
707 addElement(Name, SizeInBits, Alignment.value() * 8, Offset, DITy);
708 };
709
710 // For Switch ABI, add debug info for the added fields (resume, destroy).
711 if (Shape.ABI == coro::ABI::Switch) {
712 auto *FnPtrTy = Shape.getSwitchResumePointerType();
713 uint64_t PtrSize = Layout.getPointerSizeInBits(AS: FnPtrTy->getAddressSpace());
714 uint64_t PtrAlign =
715 Layout.getPointerABIAlignment(AS: FnPtrTy->getAddressSpace()).value() * 8;
716 auto *DIPtr = DBuilder.createPointerType(PointeeTy: nullptr, SizeInBits: PtrSize,
717 AlignInBits: FnPtrTy->getAddressSpace());
718 addElement("__resume_fn", PtrSize, PtrAlign, 0, DIPtr);
719 addElement("__destroy_fn", PtrSize, PtrAlign,
720 Shape.SwitchLowering.DestroyOffset, DIPtr);
721 uint64_t IndexSize =
722 Layout.getTypeSizeInBits(Ty: Shape.getIndexType()).getFixedValue();
723 addElement("__coro_index", IndexSize, Shape.SwitchLowering.IndexAlign * 8,
724 Shape.SwitchLowering.IndexOffset,
725 DBuilder.createBasicType(Name: "__coro_index",
726 SizeInBits: (IndexSize < 8) ? 8 : IndexSize,
727 Encoding: dwarf::DW_ATE_unsigned_char));
728 }
729 auto Defs = FrameData.getAllDefs();
730 for (auto *V : Defs)
731 addDIDef(V);
732
733 DBuilder.replaceArrays(T&: FrameDITy, Elements: DBuilder.getOrCreateArray(Elements));
734
735 auto *FrameDIVar =
736 DBuilder.createAutoVariable(Scope: DIS, Name: "__coro_frame", File: DFile, LineNo: LineNum,
737 Ty: FrameDITy, AlwaysPreserve: true, Flags: DINode::FlagArtificial);
738
739 // Subprogram would have ContainedNodes field which records the debug
740 // variables it contained. So we need to add __coro_frame to the
741 // ContainedNodes of it.
742 //
743 // If we don't add __coro_frame to the RetainedNodes, user may get
744 // `no symbol __coro_frame in context` rather than `__coro_frame`
745 // is optimized out, which is more precise.
746 DIS->retainNodes(NodesBegin: &FrameDIVar, NodesEnd: &FrameDIVar + 1);
747
748 // Construct the location for the frame debug variable. The column number
749 // is fake but it should be fine.
750 DILocation *DILoc =
751 DILocation::get(Context&: DIS->getContext(), Line: LineNum, /*Column=*/1, Scope: DIS);
752 assert(FrameDIVar->isValidLocationForIntrinsic(DILoc));
753
754 DbgVariableRecord *NewDVR =
755 new DbgVariableRecord(ValueAsMetadata::get(V: Shape.FramePtr), FrameDIVar,
756 DBuilder.createExpression(), DILoc,
757 DbgVariableRecord::LocationType::Declare);
758 BasicBlock::iterator It = Shape.getInsertPtAfterFramePtr();
759 It->getParent()->insertDbgRecordBefore(DR: NewDVR, Here: It);
760}
761
762// If there is memory accessing to promise alloca before CoroBegin
763static bool hasAccessingPromiseBeforeCB(const DominatorTree &DT,
764 coro::Shape &Shape) {
765 auto *PA = Shape.SwitchLowering.PromiseAlloca;
766 return llvm::any_of(Range: PA->uses(), P: [&](Use &U) {
767 auto *Inst = dyn_cast<Instruction>(Val: U.getUser());
768 if (!Inst || DT.dominates(Def: Shape.CoroBegin, User: Inst))
769 return false;
770
771 if (auto *CI = dyn_cast<CallInst>(Val: Inst)) {
772 // It is fine if the call wouldn't write to the Promise.
773 // This is possible for @llvm.coro.id intrinsics, which
774 // would take the promise as the second argument as a
775 // marker.
776 if (CI->onlyReadsMemory() || CI->onlyReadsMemory(OpNo: CI->getArgOperandNo(U: &U)))
777 return false;
778 return true;
779 }
780
781 return isa<StoreInst>(Val: Inst) ||
782 // It may take too much time to track the uses.
783 // Be conservative about the case the use may escape.
784 isa<GetElementPtrInst>(Val: Inst) ||
785 // There would always be a bitcast for the promise alloca
786 // before we enabled Opaque pointers. And now given
787 // opaque pointers are enabled by default. This should be
788 // fine.
789 isa<BitCastInst>(Val: Inst);
790 });
791}
792// Build the coroutine frame type as a byte array.
793// The frame layout includes:
794// - Resume function pointer at offset 0 (Switch ABI only)
795// - Destroy function pointer at offset ptrsize (Switch ABI only)
796// - Promise alloca (Switch ABI only, only if present)
797// - Suspend/Resume index
798// - Spilled values and allocas
799static void buildFrameLayout(Function &F, const DominatorTree &DT,
800 coro::Shape &Shape, FrameDataInfo &FrameData,
801 bool OptimizeFrame) {
802 const DataLayout &DL = F.getDataLayout();
803
804 // We will use this value to cap the alignment of spilled values.
805 std::optional<Align> MaxFrameAlignment;
806 if (Shape.ABI == coro::ABI::Async)
807 MaxFrameAlignment = Shape.AsyncLowering.getContextAlignment();
808 FrameTypeBuilder B(DL, MaxFrameAlignment);
809
810 AllocaInst *PromiseAlloca = Shape.getPromiseAlloca();
811 std::optional<FieldIDType> SwitchIndexFieldId;
812 IntegerType *SwitchIndexType = nullptr;
813
814 if (Shape.ABI == coro::ABI::Switch) {
815 auto *FnPtrTy = Shape.getSwitchResumePointerType();
816
817 // Add header fields for the resume and destroy functions.
818 // We can rely on these being perfectly packed.
819 (void)B.addField(Ty: FnPtrTy, MaybeFieldAlignment: MaybeAlign(), /*header*/ IsHeader: true);
820 (void)B.addField(Ty: FnPtrTy, MaybeFieldAlignment: MaybeAlign(), /*header*/ IsHeader: true);
821
822 // PromiseAlloca field needs to be explicitly added here because it's
823 // a header field with a fixed offset based on its alignment. Hence it
824 // needs special handling.
825 if (PromiseAlloca)
826 FrameData.setFieldIndex(
827 V: PromiseAlloca, Index: B.addFieldForAlloca(AI: PromiseAlloca, /*header*/ IsHeader: true));
828
829 // Add a field to store the suspend index. This doesn't need to
830 // be in the header.
831 unsigned IndexBits = std::max(a: 1U, b: Log2_64_Ceil(Value: Shape.CoroSuspends.size()));
832 SwitchIndexType = Type::getIntNTy(C&: F.getContext(), N: IndexBits);
833
834 SwitchIndexFieldId = B.addField(Ty: SwitchIndexType, MaybeFieldAlignment: MaybeAlign());
835 } else {
836 assert(PromiseAlloca == nullptr && "lowering doesn't support promises");
837 }
838
839 // Because multiple allocas may own the same field slot,
840 // we add allocas to field here.
841 B.addFieldForAllocas(F, FrameData, Shape, OptimizeFrame);
842 // Add PromiseAlloca to Allocas list so that
843 // 1. updateLayoutIndex could update its index after
844 // `performOptimizedStructLayout`
845 // 2. it is processed in insertSpills.
846 if (Shape.ABI == coro::ABI::Switch && PromiseAlloca) {
847 // We assume that no alias will be create before CoroBegin.
848 FrameData.Allocas.emplace_back(
849 Args&: PromiseAlloca, Args: DenseMap<Instruction *, std::optional<APInt>>{},
850 Args: hasAccessingPromiseBeforeCB(DT, Shape));
851 }
852 // Create an entry for every spilled value.
853 for (auto &S : FrameData.Spills) {
854 Type *FieldType = S.first->getType();
855 MaybeAlign MA;
856 // For byval arguments, we need to store the pointed value in the frame,
857 // instead of the pointer itself.
858 if (const Argument *A = dyn_cast<Argument>(Val: S.first)) {
859 if (A->hasByValAttr()) {
860 FieldType = A->getParamByValType();
861 MA = A->getParamAlign();
862 }
863 }
864 FieldIDType Id =
865 B.addField(Ty: FieldType, MaybeFieldAlignment: MA, IsHeader: false /*header*/, IsSpillOfValue: true /*IsSpillOfValue*/);
866 FrameData.setFieldIndex(V: S.first, Index: Id);
867 }
868
869 B.finish();
870
871 FrameData.updateLayoutInfo(B);
872 Shape.FrameAlign = B.getStructAlign();
873 Shape.FrameSize = B.getStructSize();
874
875 switch (Shape.ABI) {
876 case coro::ABI::Switch: {
877 // In the switch ABI, remember the function pointer and index field info.
878 // Resume and Destroy function pointers are in the frame header.
879 const DataLayout &DL = F.getDataLayout();
880 Shape.SwitchLowering.DestroyOffset = DL.getPointerSize();
881
882 auto IndexField = B.getLayoutField(Id: *SwitchIndexFieldId);
883 Shape.SwitchLowering.IndexType = SwitchIndexType;
884 Shape.SwitchLowering.IndexAlign = IndexField.Alignment.value();
885 Shape.SwitchLowering.IndexOffset = IndexField.Offset;
886
887 // Also round the frame size up to a multiple of its alignment, as is
888 // generally expected in C/C++.
889 Shape.FrameSize = alignTo(Size: Shape.FrameSize, A: Shape.FrameAlign);
890 break;
891 }
892
893 // In the retcon ABI, remember whether the frame is inline in the storage.
894 case coro::ABI::Retcon:
895 case coro::ABI::RetconOnce: {
896 auto Id = Shape.getRetconCoroId();
897 Shape.RetconLowering.IsFrameInlineInStorage
898 = (B.getStructSize() <= Id->getStorageSize() &&
899 B.getStructAlign() <= Id->getStorageAlignment());
900 break;
901 }
902 case coro::ABI::Async: {
903 Shape.AsyncLowering.FrameOffset =
904 alignTo(Size: Shape.AsyncLowering.ContextHeaderSize, A: Shape.FrameAlign);
905 // Also make the final context size a multiple of the context alignment to
906 // make allocation easier for allocators.
907 Shape.AsyncLowering.ContextSize =
908 alignTo(Size: Shape.AsyncLowering.FrameOffset + Shape.FrameSize,
909 A: Shape.AsyncLowering.getContextAlignment());
910 if (Shape.AsyncLowering.getContextAlignment() < Shape.FrameAlign) {
911 report_fatal_error(
912 reason: "The alignment requirment of frame variables cannot be higher than "
913 "the alignment of the async function context");
914 }
915 break;
916 }
917 }
918}
919
920/// If MaybeArgument is a byval Argument, return its byval type. Also removes
921/// the captures attribute, so that the argument *value* may be stored directly
922/// on the coroutine frame.
923static Type *extractByvalIfArgument(Value *MaybeArgument) {
924 if (auto *Arg = dyn_cast<Argument>(Val: MaybeArgument)) {
925 Arg->getParent()->removeParamAttr(ArgNo: Arg->getArgNo(), Kind: Attribute::Captures);
926
927 if (Arg->hasByValAttr())
928 return Arg->getParamByValType();
929 }
930 return nullptr;
931}
932
933/// Store Def into the coroutine frame.
934static void createStoreIntoFrame(IRBuilder<> &Builder, Value *Def,
935 Type *ByValTy, const coro::Shape &Shape,
936 const FrameDataInfo &FrameData) {
937 LLVMContext &Ctx = Shape.CoroBegin->getContext();
938 uint64_t Offset = FrameData.getOffset(V: Def);
939
940 Value *G = Shape.FramePtr;
941 if (Offset != 0) {
942 auto *OffsetVal = ConstantInt::get(Ty: Type::getInt64Ty(C&: Ctx), V: Offset);
943 G = Builder.CreateInBoundsPtrAdd(Ptr: G, Offset: OffsetVal,
944 Name: Def->getName() + Twine(".spill.addr"));
945 }
946 auto SpillAlignment = Align(FrameData.getAlign(V: Def));
947
948 // For byval arguments, copy the pointed-to value to the frame.
949 if (ByValTy) {
950 auto &DL = Builder.GetInsertBlock()->getDataLayout();
951 auto Size = DL.getTypeStoreSize(Ty: ByValTy);
952 // Def is a pointer to the byval argument
953 Builder.CreateMemCpy(Dst: G, DstAlign: SpillAlignment, Src: Def, SrcAlign: SpillAlignment, Size);
954 } else {
955 Builder.CreateAlignedStore(Val: Def, Ptr: G, Align: SpillAlignment);
956 }
957}
958
959/// Returns a pointer into the coroutine frame at the offset where Orig is
960/// located.
961static Value *createGEPToFramePointer(const FrameDataInfo &FrameData,
962 IRBuilder<> &Builder, coro::Shape &Shape,
963 Value *Orig) {
964 LLVMContext &Ctx = Shape.CoroBegin->getContext();
965 uint64_t Offset = FrameData.getOffset(V: Orig);
966 auto *OffsetVal = ConstantInt::get(Ty: Type::getInt64Ty(C&: Ctx), V: Offset);
967 Value *Ptr = Builder.CreateInBoundsPtrAdd(Ptr: Shape.FramePtr, Offset: OffsetVal);
968
969 if (auto *AI = dyn_cast<AllocaInst>(Val: Orig)) {
970 if (FrameData.getDynamicAlign(V: Orig) != 0) {
971 assert(FrameData.getDynamicAlign(Orig) == AI->getAlign().value());
972 auto *M = AI->getModule();
973 auto *IntPtrTy = M->getDataLayout().getIntPtrType(AI->getType());
974 auto *PtrValue = Builder.CreatePtrToInt(V: Ptr, DestTy: IntPtrTy);
975 auto *AlignMask = ConstantInt::get(Ty: IntPtrTy, V: AI->getAlign().value() - 1);
976 PtrValue = Builder.CreateAdd(LHS: PtrValue, RHS: AlignMask);
977 PtrValue = Builder.CreateAnd(LHS: PtrValue, RHS: Builder.CreateNot(V: AlignMask));
978 return Builder.CreateIntToPtr(V: PtrValue, DestTy: AI->getType());
979 }
980 // If the type of Ptr is not equal to the type of AllocaInst, it implies
981 // that the AllocaInst may be reused in the Frame slot of other AllocaInst.
982 // Note: If the strategy dealing with alignment changes, this cast must be
983 // refined
984 if (Ptr->getType() != Orig->getType())
985 Ptr = Builder.CreateAddrSpaceCast(V: Ptr, DestTy: Orig->getType(),
986 Name: Orig->getName() + Twine(".cast"));
987 }
988 return Ptr;
989}
990
991/// Find dbg.declare or dbg.declare_value records referencing `Def`. If none are
992/// found, walk up the load chain to find one.
993template <DbgVariableRecord::LocationType record_type>
994static TinyPtrVector<DbgVariableRecord *>
995findDbgRecordsThroughLoads(Function &F, Value *Def) {
996 static_assert(record_type == DbgVariableRecord::LocationType::Declare ||
997 record_type == DbgVariableRecord::LocationType::DeclareValue);
998 constexpr auto FindFunc =
999 record_type == DbgVariableRecord::LocationType::Declare
1000 ? findDVRDeclares
1001 : findDVRDeclareValues;
1002
1003 TinyPtrVector<DbgVariableRecord *> Records = FindFunc(Def);
1004
1005 if (!F.getSubprogram())
1006 return Records;
1007
1008 Value *CurDef = Def;
1009 while (Records.empty() && isa<LoadInst>(Val: CurDef)) {
1010 auto *LdInst = cast<LoadInst>(Val: CurDef);
1011 if (!LdInst->getType()->isPointerTy())
1012 break;
1013 CurDef = LdInst->getPointerOperand();
1014 if (!isa<AllocaInst, LoadInst>(Val: CurDef))
1015 break;
1016 Records = FindFunc(CurDef);
1017 }
1018
1019 return Records;
1020}
1021
1022// Helper function to handle allocas that may be accessed before CoroBegin.
1023// This creates a memcpy from the original alloca to the coroutine frame after
1024// CoroBegin, ensuring the frame has the correct initial values.
1025static void handleAccessBeforeCoroBegin(const FrameDataInfo &FrameData,
1026 coro::Shape &Shape,
1027 IRBuilder<> &Builder,
1028 AllocaInst *Alloca) {
1029 Value *Size = Builder.CreateAllocationSize(DestTy: Builder.getInt64Ty(), AI: Alloca);
1030 auto *G = createGEPToFramePointer(FrameData, Builder, Shape, Orig: Alloca);
1031 Builder.CreateMemCpy(Dst: G, DstAlign: FrameData.getAlign(V: Alloca), Src: Alloca,
1032 SrcAlign: Alloca->getAlign(), Size);
1033}
1034
1035// Replace all alloca and SSA values that are accessed across suspend points
1036// with GetElementPointer from coroutine frame + loads and stores. Create an
1037// AllocaSpillBB that will become the new entry block for the resume parts of
1038// the coroutine:
1039//
1040// %hdl = coro.begin(...)
1041// whatever
1042//
1043// becomes:
1044//
1045// %hdl = coro.begin(...)
1046// br label %AllocaSpillBB
1047//
1048// AllocaSpillBB:
1049// ; geps corresponding to allocas that were moved to coroutine frame
1050// br label PostSpill
1051//
1052// PostSpill:
1053// whatever
1054//
1055//
1056static void insertSpills(const FrameDataInfo &FrameData, coro::Shape &Shape) {
1057 LLVMContext &C = Shape.CoroBegin->getContext();
1058 Function *F = Shape.CoroBegin->getFunction();
1059 IRBuilder<> Builder(C);
1060 DominatorTree DT(*F);
1061 SmallDenseMap<Argument *, AllocaInst *, 4> ArgToAllocaMap;
1062
1063 MDBuilder MDB(C);
1064 // Create a TBAA tag for accesses to certain coroutine frame slots, so that
1065 // subsequent alias analysis will understand they do not intersect with
1066 // user memory.
1067 // We do this only if a suitable TBAA root already exists in the module.
1068 MDNode *TBAATag = nullptr;
1069 if (auto *CppTBAAStr = MDString::getIfExists(Context&: C, Str: "Simple C++ TBAA")) {
1070 auto *TBAARoot = MDNode::getIfExists(Context&: C, MDs: CppTBAAStr);
1071 // Create a "fake" scalar type; all other types defined in the source
1072 // language will be assumed non-aliasing with this type.
1073 MDNode *Scalar = MDB.createTBAAScalarTypeNode(
1074 Name: (F->getName() + ".Frame Slot").str(), Parent: TBAARoot);
1075 TBAATag = MDB.createTBAAStructTagNode(BaseType: Scalar, AccessType: Scalar, Offset: 0);
1076 }
1077 for (auto const &E : FrameData.Spills) {
1078 Value *Def = E.first;
1079 Type *ByValTy = extractByvalIfArgument(MaybeArgument: Def);
1080
1081 Builder.SetInsertPoint(coro::getSpillInsertionPt(Shape, Def, DT));
1082 createStoreIntoFrame(Builder, Def, ByValTy, Shape, FrameData);
1083
1084 BasicBlock *CurrentBlock = nullptr;
1085 Value *CurrentReload = nullptr;
1086 for (auto *U : E.second) {
1087 // If we have not seen the use block, create a load instruction to reload
1088 // the spilled value from the coroutine frame. Populates the Value pointer
1089 // reference provided with the frame GEP.
1090 if (CurrentBlock != U->getParent()) {
1091 CurrentBlock = U->getParent();
1092 Builder.SetInsertPoint(TheBB: CurrentBlock,
1093 IP: CurrentBlock->getFirstInsertionPt());
1094
1095 auto *GEP = createGEPToFramePointer(FrameData, Builder, Shape, Orig: E.first);
1096 GEP->setName(E.first->getName() + Twine(".reload.addr"));
1097 if (ByValTy) {
1098 CurrentReload = GEP;
1099 } else {
1100 auto SpillAlignment = Align(FrameData.getAlign(V: Def));
1101 auto *LI =
1102 Builder.CreateAlignedLoad(Ty: E.first->getType(), Ptr: GEP, Align: SpillAlignment,
1103 Name: E.first->getName() + Twine(".reload"));
1104 if (TBAATag)
1105 LI->setMetadata(KindID: LLVMContext::MD_tbaa, Node: TBAATag);
1106 CurrentReload = LI;
1107 }
1108
1109 TinyPtrVector<DbgVariableRecord *> DVRs = findDbgRecordsThroughLoads<
1110 DbgVariableRecord::LocationType::Declare>(F&: *F, Def);
1111
1112 auto SalvageOne = [&](DbgVariableRecord *DDI) {
1113 // This dbg.declare is preserved for all coro-split function
1114 // fragments. It will be unreachable in the main function, and
1115 // processed by coro::salvageDebugInfo() by the Cloner.
1116 DbgVariableRecord *NewDVR = new DbgVariableRecord(
1117 ValueAsMetadata::get(V: CurrentReload), DDI->getVariable(),
1118 DDI->getExpression(), DDI->getDebugLoc(),
1119 DbgVariableRecord::LocationType::Declare);
1120 Builder.GetInsertPoint()->getParent()->insertDbgRecordBefore(
1121 DR: NewDVR, Here: Builder.GetInsertPoint());
1122 // This dbg.declare is for the main function entry point. It
1123 // will be deleted in all coro-split functions.
1124 coro::salvageDebugInfo(ArgToAllocaMap, DVR&: *DDI, UseEntryValue: false /*UseEntryValue*/);
1125 };
1126 for_each(Range&: DVRs, F: SalvageOne);
1127 }
1128
1129 TinyPtrVector<DbgVariableRecord *> DVRDeclareValues =
1130 findDbgRecordsThroughLoads<
1131 DbgVariableRecord::LocationType::DeclareValue>(F&: *F, Def);
1132
1133 auto SalvageOneCoro = [&](auto *DDI) {
1134 // This dbg.declare_value is preserved for all coro-split function
1135 // fragments. It will be unreachable in the main function, and
1136 // processed by coro::salvageDebugInfo() by the Cloner. However, convert
1137 // it to a dbg.declare to make sure future passes don't have to deal
1138 // with a dbg.declare_value.
1139 auto *VAM = ValueAsMetadata::get(V: CurrentReload);
1140 Type *Ty = VAM->getValue()->getType();
1141 // If the metadata type is not a pointer, emit a dbg.value instead.
1142 DbgVariableRecord *NewDVR = new DbgVariableRecord(
1143 ValueAsMetadata::get(V: CurrentReload), DDI->getVariable(),
1144 DDI->getExpression(), DDI->getDebugLoc(),
1145 Ty->isPointerTy() ? DbgVariableRecord::LocationType::Declare
1146 : DbgVariableRecord::LocationType::Value);
1147 Builder.GetInsertPoint()->getParent()->insertDbgRecordBefore(
1148 DR: NewDVR, Here: Builder.GetInsertPoint());
1149 // This dbg.declare_value is for the main function entry point. It
1150 // will be deleted in all coro-split functions.
1151 coro::salvageDebugInfo(ArgToAllocaMap, DVR&: *DDI, UseEntryValue: false /*UseEntryValue*/);
1152 };
1153 for_each(Range&: DVRDeclareValues, F: SalvageOneCoro);
1154
1155 // If we have a single edge PHINode, remove it and replace it with a
1156 // reload from the coroutine frame. (We already took care of multi edge
1157 // PHINodes by normalizing them in the rewritePHIs function).
1158 if (auto *PN = dyn_cast<PHINode>(Val: U)) {
1159 assert(PN->getNumIncomingValues() == 1 &&
1160 "unexpected number of incoming "
1161 "values in the PHINode");
1162 PN->replaceAllUsesWith(V: CurrentReload);
1163 PN->eraseFromParent();
1164 continue;
1165 }
1166
1167 // Replace all uses of CurrentValue in the current instruction with
1168 // reload.
1169 U->replaceUsesOfWith(From: Def, To: CurrentReload);
1170 // Instructions are added to Def's user list if the attached
1171 // debug records use Def. Update those now.
1172 for (DbgVariableRecord &DVR : filterDbgVars(R: U->getDbgRecordRange()))
1173 DVR.replaceVariableLocationOp(OldValue: Def, NewValue: CurrentReload, AllowEmpty: true);
1174 }
1175 }
1176
1177 BasicBlock *FramePtrBB = Shape.getInsertPtAfterFramePtr()->getParent();
1178
1179 auto SpillBlock = FramePtrBB->splitBasicBlock(
1180 I: Shape.getInsertPtAfterFramePtr(), BBName: "AllocaSpillBB");
1181 SpillBlock->splitBasicBlock(I: &SpillBlock->front(), BBName: "PostSpill");
1182 Shape.AllocaSpillBlock = SpillBlock;
1183
1184 // retcon and retcon.once lowering assumes all uses have been sunk.
1185 if (Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce ||
1186 Shape.ABI == coro::ABI::Async) {
1187 // If we found any allocas, replace all of their remaining uses with Geps.
1188 Builder.SetInsertPoint(TheBB: SpillBlock, IP: SpillBlock->begin());
1189 for (const auto &P : FrameData.Allocas) {
1190 AllocaInst *Alloca = P.Alloca;
1191 auto *G = createGEPToFramePointer(FrameData, Builder, Shape, Orig: Alloca);
1192
1193 // Remove any lifetime intrinsics, now that these are no longer allocas.
1194 for (User *U : make_early_inc_range(Range: Alloca->users())) {
1195 auto *I = cast<Instruction>(Val: U);
1196 if (I->isLifetimeStartOrEnd())
1197 I->eraseFromParent();
1198 }
1199
1200 // We are not using ReplaceInstWithInst(P.first, cast<Instruction>(G))
1201 // here, as we are changing location of the instruction.
1202 G->takeName(V: Alloca);
1203 Alloca->replaceAllUsesWith(V: G);
1204 Alloca->eraseFromParent();
1205 }
1206 return;
1207 }
1208
1209 // If we found any alloca, replace all of their remaining uses with GEP
1210 // instructions. To remain debugbility, we replace the uses of allocas for
1211 // dbg.declares and dbg.values with the reload from the frame.
1212 // Note: We cannot replace the alloca with GEP instructions indiscriminately,
1213 // as some of the uses may not be dominated by CoroBegin.
1214 Builder.SetInsertPoint(TheBB: Shape.AllocaSpillBlock,
1215 IP: Shape.AllocaSpillBlock->begin());
1216 SmallVector<Instruction *, 4> UsersToUpdate;
1217 for (const auto &A : FrameData.Allocas) {
1218 AllocaInst *Alloca = A.Alloca;
1219 UsersToUpdate.clear();
1220 for (User *U : make_early_inc_range(Range: Alloca->users())) {
1221 auto *I = cast<Instruction>(Val: U);
1222 // It is meaningless to retain the lifetime intrinsics refer for the
1223 // member of coroutine frames and the meaningless lifetime intrinsics
1224 // are possible to block further optimizations.
1225 if (I->isLifetimeStartOrEnd())
1226 I->eraseFromParent();
1227 else if (DT.dominates(Def: Shape.CoroBegin, User: I))
1228 UsersToUpdate.push_back(Elt: I);
1229 }
1230
1231 if (UsersToUpdate.empty())
1232 continue;
1233 auto *G = createGEPToFramePointer(FrameData, Builder, Shape, Orig: Alloca);
1234 G->setName(Alloca->getName() + Twine(".reload.addr"));
1235
1236 SmallVector<DbgVariableRecord *> DbgVariableRecords;
1237 findDbgUsers(V: Alloca, DbgVariableRecords);
1238 for (auto *DVR : DbgVariableRecords)
1239 DVR->replaceVariableLocationOp(OldValue: Alloca, NewValue: G);
1240
1241 for (Instruction *I : UsersToUpdate)
1242 I->replaceUsesOfWith(From: Alloca, To: G);
1243
1244 if (Alloca->user_empty())
1245 Alloca->eraseFromParent();
1246 }
1247 Builder.SetInsertPoint(&*Shape.getInsertPtAfterFramePtr());
1248 for (const auto &A : FrameData.Allocas) {
1249 AllocaInst *Alloca = A.Alloca;
1250 if (A.MayWriteBeforeCoroBegin) {
1251 // isEscaped really means potentially modified before CoroBegin.
1252 handleAccessBeforeCoroBegin(FrameData, Shape, Builder, Alloca);
1253 }
1254 // For each alias to Alloca created before CoroBegin but used after
1255 // CoroBegin, we recreate them after CoroBegin by applying the offset
1256 // to the pointer in the frame.
1257 for (const auto &Alias : A.Aliases) {
1258 auto *FramePtr =
1259 createGEPToFramePointer(FrameData, Builder, Shape, Orig: Alloca);
1260 auto &Value = *Alias.second;
1261 auto ITy = IntegerType::get(C, NumBits: Value.getBitWidth());
1262 auto *AliasPtr =
1263 Builder.CreateInBoundsPtrAdd(Ptr: FramePtr, Offset: ConstantInt::get(Ty: ITy, V: Value));
1264 Alias.first->replaceUsesWithIf(
1265 New: AliasPtr, ShouldReplace: [&](Use &U) { return DT.dominates(Def: Shape.CoroBegin, U); });
1266 }
1267 }
1268}
1269
1270// Moves the values in the PHIs in SuccBB that correspong to PredBB into a new
1271// PHI in InsertedBB.
1272static void movePHIValuesToInsertedBlock(BasicBlock *SuccBB,
1273 BasicBlock *InsertedBB,
1274 BasicBlock *PredBB,
1275 PHINode *UntilPHI = nullptr) {
1276 auto *PN = cast<PHINode>(Val: &SuccBB->front());
1277 do {
1278 int Index = PN->getBasicBlockIndex(BB: InsertedBB);
1279 Value *V = PN->getIncomingValue(i: Index);
1280 PHINode *InputV = PHINode::Create(
1281 Ty: V->getType(), NumReservedValues: 1, NameStr: V->getName() + Twine(".") + SuccBB->getName());
1282 InputV->insertBefore(InsertPos: InsertedBB->begin());
1283 InputV->addIncoming(V, BB: PredBB);
1284 PN->setIncomingValue(i: Index, V: InputV);
1285 PN = dyn_cast<PHINode>(Val: PN->getNextNode());
1286 } while (PN != UntilPHI);
1287}
1288
1289// Rewrites the PHI Nodes in a cleanuppad.
1290static void rewritePHIsForCleanupPad(BasicBlock *CleanupPadBB,
1291 CleanupPadInst *CleanupPad) {
1292 // For every incoming edge to a CleanupPad we will create a new block holding
1293 // all incoming values in single-value PHI nodes. We will then create another
1294 // block to act as a dispather (as all unwind edges for related EH blocks
1295 // must be the same).
1296 //
1297 // cleanuppad:
1298 // %2 = phi i32[%0, %catchswitch], [%1, %catch.1]
1299 // %3 = cleanuppad within none []
1300 //
1301 // It will create:
1302 //
1303 // cleanuppad.corodispatch
1304 // %2 = phi i8[0, %catchswitch], [1, %catch.1]
1305 // %3 = cleanuppad within none []
1306 // switch i8 % 2, label %unreachable
1307 // [i8 0, label %cleanuppad.from.catchswitch
1308 // i8 1, label %cleanuppad.from.catch.1]
1309 // cleanuppad.from.catchswitch:
1310 // %4 = phi i32 [%0, %catchswitch]
1311 // br %label cleanuppad
1312 // cleanuppad.from.catch.1:
1313 // %6 = phi i32 [%1, %catch.1]
1314 // br %label cleanuppad
1315 // cleanuppad:
1316 // %8 = phi i32 [%4, %cleanuppad.from.catchswitch],
1317 // [%6, %cleanuppad.from.catch.1]
1318
1319 // Unreachable BB, in case switching on an invalid value in the dispatcher.
1320 auto *UnreachBB = BasicBlock::Create(
1321 Context&: CleanupPadBB->getContext(), Name: "unreachable", Parent: CleanupPadBB->getParent());
1322 IRBuilder<> Builder(UnreachBB);
1323 Builder.CreateUnreachable();
1324
1325 // Create a new cleanuppad which will be the dispatcher.
1326 auto *NewCleanupPadBB =
1327 BasicBlock::Create(Context&: CleanupPadBB->getContext(),
1328 Name: CleanupPadBB->getName() + Twine(".corodispatch"),
1329 Parent: CleanupPadBB->getParent(), InsertBefore: CleanupPadBB);
1330 Builder.SetInsertPoint(NewCleanupPadBB);
1331 auto *SwitchType = Builder.getInt8Ty();
1332 auto *SetDispatchValuePN =
1333 Builder.CreatePHI(Ty: SwitchType, NumReservedValues: pred_size(BB: CleanupPadBB));
1334 CleanupPad->removeFromParent();
1335 CleanupPad->insertAfter(InsertPos: SetDispatchValuePN->getIterator());
1336 auto *SwitchOnDispatch = Builder.CreateSwitch(V: SetDispatchValuePN, Dest: UnreachBB,
1337 NumCases: pred_size(BB: CleanupPadBB));
1338
1339 int SwitchIndex = 0;
1340 SmallVector<BasicBlock *, 8> Preds(predecessors(BB: CleanupPadBB));
1341 for (BasicBlock *Pred : Preds) {
1342 // Create a new cleanuppad and move the PHI values to there.
1343 auto *CaseBB = BasicBlock::Create(Context&: CleanupPadBB->getContext(),
1344 Name: CleanupPadBB->getName() +
1345 Twine(".from.") + Pred->getName(),
1346 Parent: CleanupPadBB->getParent(), InsertBefore: CleanupPadBB);
1347 updatePhiNodes(DestBB: CleanupPadBB, OldPred: Pred, NewPred: CaseBB);
1348 CaseBB->setName(CleanupPadBB->getName() + Twine(".from.") +
1349 Pred->getName());
1350 Builder.SetInsertPoint(CaseBB);
1351 Builder.CreateBr(Dest: CleanupPadBB);
1352 movePHIValuesToInsertedBlock(SuccBB: CleanupPadBB, InsertedBB: CaseBB, PredBB: NewCleanupPadBB);
1353
1354 // Update this Pred to the new unwind point.
1355 setUnwindEdgeTo(TI: Pred->getTerminator(), Succ: NewCleanupPadBB);
1356
1357 // Setup the switch in the dispatcher.
1358 auto *SwitchConstant = ConstantInt::get(Ty: SwitchType, V: SwitchIndex);
1359 SetDispatchValuePN->addIncoming(V: SwitchConstant, BB: Pred);
1360 SwitchOnDispatch->addCase(OnVal: SwitchConstant, Dest: CaseBB);
1361 SwitchIndex++;
1362 }
1363}
1364
1365static void cleanupSinglePredPHIs(Function &F) {
1366 SmallVector<PHINode *, 32> Worklist;
1367 for (auto &BB : F) {
1368 for (auto &Phi : BB.phis()) {
1369 if (Phi.getNumIncomingValues() == 1) {
1370 Worklist.push_back(Elt: &Phi);
1371 } else
1372 break;
1373 }
1374 }
1375 while (!Worklist.empty()) {
1376 auto *Phi = Worklist.pop_back_val();
1377 auto *OriginalValue = Phi->getIncomingValue(i: 0);
1378 Phi->replaceAllUsesWith(V: OriginalValue);
1379 }
1380}
1381
1382static void rewritePHIs(BasicBlock &BB) {
1383 // For every incoming edge we will create a block holding all
1384 // incoming values in a single PHI nodes.
1385 //
1386 // loop:
1387 // %n.val = phi i32[%n, %entry], [%inc, %loop]
1388 //
1389 // It will create:
1390 //
1391 // loop.from.entry:
1392 // %n.loop.pre = phi i32 [%n, %entry]
1393 // br %label loop
1394 // loop.from.loop:
1395 // %inc.loop.pre = phi i32 [%inc, %loop]
1396 // br %label loop
1397 //
1398 // After this rewrite, further analysis will ignore any phi nodes with more
1399 // than one incoming edge.
1400
1401 // TODO: Simplify PHINodes in the basic block to remove duplicate
1402 // predecessors.
1403
1404 // Special case for CleanupPad: all EH blocks must have the same unwind edge
1405 // so we need to create an additional "dispatcher" block.
1406 if (!BB.empty()) {
1407 if (auto *CleanupPad =
1408 dyn_cast_or_null<CleanupPadInst>(Val: BB.getFirstNonPHIIt())) {
1409 SmallVector<BasicBlock *, 8> Preds(predecessors(BB: &BB));
1410 for (BasicBlock *Pred : Preds) {
1411 if (CatchSwitchInst *CS =
1412 dyn_cast<CatchSwitchInst>(Val: Pred->getTerminator())) {
1413 // CleanupPad with a CatchSwitch predecessor: therefore this is an
1414 // unwind destination that needs to be handle specially.
1415 assert(CS->getUnwindDest() == &BB);
1416 (void)CS;
1417 rewritePHIsForCleanupPad(CleanupPadBB: &BB, CleanupPad);
1418 return;
1419 }
1420 }
1421 }
1422 }
1423
1424 LandingPadInst *LandingPad = nullptr;
1425 PHINode *ReplPHI = nullptr;
1426 if (!BB.empty()) {
1427 if ((LandingPad =
1428 dyn_cast_or_null<LandingPadInst>(Val: BB.getFirstNonPHIIt()))) {
1429 // ehAwareSplitEdge will clone the LandingPad in all the edge blocks.
1430 // We replace the original landing pad with a PHINode that will collect the
1431 // results from all of them.
1432 ReplPHI = PHINode::Create(Ty: LandingPad->getType(), NumReservedValues: 1, NameStr: "");
1433 ReplPHI->insertBefore(InsertPos: LandingPad->getIterator());
1434 ReplPHI->takeName(V: LandingPad);
1435 LandingPad->replaceAllUsesWith(V: ReplPHI);
1436 // We will erase the original landing pad at the end of this function after
1437 // ehAwareSplitEdge cloned it in the transition blocks.
1438 }
1439 }
1440
1441 SmallVector<BasicBlock *, 8> Preds(predecessors(BB: &BB));
1442 for (BasicBlock *Pred : Preds) {
1443 auto *IncomingBB = ehAwareSplitEdge(BB: Pred, Succ: &BB, OriginalPad: LandingPad, LandingPadReplacement: ReplPHI);
1444 IncomingBB->setName(BB.getName() + Twine(".from.") + Pred->getName());
1445
1446 // Stop the moving of values at ReplPHI, as this is either null or the PHI
1447 // that replaced the landing pad.
1448 movePHIValuesToInsertedBlock(SuccBB: &BB, InsertedBB: IncomingBB, PredBB: Pred, UntilPHI: ReplPHI);
1449 }
1450
1451 if (LandingPad) {
1452 // Calls to ehAwareSplitEdge function cloned the original lading pad.
1453 // No longer need it.
1454 LandingPad->eraseFromParent();
1455 }
1456}
1457
1458static void rewritePHIs(Function &F) {
1459 SmallVector<BasicBlock *, 8> WorkList;
1460
1461 for (BasicBlock &BB : F)
1462 if (auto *PN = dyn_cast<PHINode>(Val: &BB.front()))
1463 if (PN->getNumIncomingValues() > 1)
1464 WorkList.push_back(Elt: &BB);
1465
1466 for (BasicBlock *BB : WorkList)
1467 rewritePHIs(BB&: *BB);
1468}
1469
1470// Splits the block at a particular instruction unless it is the first
1471// instruction in the block with a single predecessor.
1472static BasicBlock *splitBlockIfNotFirst(Instruction *I, const Twine &Name) {
1473 auto *BB = I->getParent();
1474 if (&BB->front() == I) {
1475 if (BB->getSinglePredecessor()) {
1476 BB->setName(Name);
1477 return BB;
1478 }
1479 }
1480 return BB->splitBasicBlock(I, BBName: Name);
1481}
1482
1483// Split above and below a particular instruction so that it
1484// will be all alone by itself in a block.
1485static void splitAround(Instruction *I, const Twine &Name) {
1486 splitBlockIfNotFirst(I, Name);
1487 splitBlockIfNotFirst(I: I->getNextNode(), Name: "After" + Name);
1488}
1489
1490/// After we split the coroutine, will the given basic block be along
1491/// an obvious exit path for the resumption function?
1492static bool willLeaveFunctionImmediatelyAfter(BasicBlock *BB,
1493 unsigned depth = 3) {
1494 // If we've bottomed out our depth count, stop searching and assume
1495 // that the path might loop back.
1496 if (depth == 0) return false;
1497
1498 // If this is a suspend block, we're about to exit the resumption function.
1499 if (coro::isSuspendBlock(BB))
1500 return true;
1501
1502 // Recurse into the successors.
1503 for (auto *Succ : successors(BB)) {
1504 if (!willLeaveFunctionImmediatelyAfter(BB: Succ, depth: depth - 1))
1505 return false;
1506 }
1507
1508 // If none of the successors leads back in a loop, we're on an exit/abort.
1509 return true;
1510}
1511
1512static bool localAllocaNeedsStackSave(CoroAllocaAllocInst *AI) {
1513 // Look for a free that isn't sufficiently obviously followed by
1514 // either a suspend or a termination, i.e. something that will leave
1515 // the coro resumption frame.
1516 for (auto *U : AI->users()) {
1517 auto FI = dyn_cast<CoroAllocaFreeInst>(Val: U);
1518 if (!FI) continue;
1519
1520 if (!willLeaveFunctionImmediatelyAfter(BB: FI->getParent()))
1521 return true;
1522 }
1523
1524 // If we never found one, we don't need a stack save.
1525 return false;
1526}
1527
1528/// Turn each of the given local allocas into a normal (dynamic) alloca
1529/// instruction.
1530static void lowerLocalAllocas(ArrayRef<CoroAllocaAllocInst*> LocalAllocas,
1531 SmallVectorImpl<Instruction*> &DeadInsts) {
1532 for (auto *AI : LocalAllocas) {
1533 IRBuilder<> Builder(AI);
1534
1535 // Save the stack depth. Try to avoid doing this if the stackrestore
1536 // is going to immediately precede a return or something.
1537 Value *StackSave = nullptr;
1538 if (localAllocaNeedsStackSave(AI))
1539 StackSave = Builder.CreateStackSave();
1540
1541 // Allocate memory.
1542 auto Alloca = Builder.CreateAlloca(Ty: Builder.getInt8Ty(), ArraySize: AI->getSize());
1543 Alloca->setAlignment(AI->getAlignment());
1544
1545 for (auto *U : AI->users()) {
1546 // Replace gets with the allocation.
1547 if (isa<CoroAllocaGetInst>(Val: U)) {
1548 U->replaceAllUsesWith(V: Alloca);
1549
1550 // Replace frees with stackrestores. This is safe because
1551 // alloca.alloc is required to obey a stack discipline, although we
1552 // don't enforce that structurally.
1553 } else {
1554 auto FI = cast<CoroAllocaFreeInst>(Val: U);
1555 if (StackSave) {
1556 Builder.SetInsertPoint(FI);
1557 Builder.CreateStackRestore(Ptr: StackSave);
1558 }
1559 }
1560 DeadInsts.push_back(Elt: cast<Instruction>(Val: U));
1561 }
1562
1563 DeadInsts.push_back(Elt: AI);
1564 }
1565}
1566
1567/// Get the current swifterror value.
1568static Value *emitGetSwiftErrorValue(IRBuilder<> &Builder, Type *ValueTy,
1569 coro::Shape &Shape) {
1570 // Make a fake function pointer as a sort of intrinsic.
1571 auto FnTy = FunctionType::get(Result: ValueTy, Params: {}, isVarArg: false);
1572 auto Fn = ConstantPointerNull::get(T: Builder.getPtrTy());
1573
1574 auto Call = Builder.CreateCall(FTy: FnTy, Callee: Fn, Args: {});
1575 Shape.SwiftErrorOps.push_back(Elt: Call);
1576
1577 return Call;
1578}
1579
1580/// Set the given value as the current swifterror value.
1581///
1582/// Returns a slot that can be used as a swifterror slot.
1583static Value *emitSetSwiftErrorValue(IRBuilder<> &Builder, Value *V,
1584 coro::Shape &Shape) {
1585 // Make a fake function pointer as a sort of intrinsic.
1586 auto FnTy = FunctionType::get(Result: Builder.getPtrTy(),
1587 Params: {V->getType()}, isVarArg: false);
1588 auto Fn = ConstantPointerNull::get(T: Builder.getPtrTy());
1589
1590 auto Call = Builder.CreateCall(FTy: FnTy, Callee: Fn, Args: { V });
1591 Shape.SwiftErrorOps.push_back(Elt: Call);
1592
1593 return Call;
1594}
1595
1596/// Set the swifterror value from the given alloca before a call,
1597/// then put in back in the alloca afterwards.
1598///
1599/// Returns an address that will stand in for the swifterror slot
1600/// until splitting.
1601static Value *emitSetAndGetSwiftErrorValueAround(Instruction *Call,
1602 AllocaInst *Alloca,
1603 coro::Shape &Shape) {
1604 auto ValueTy = Alloca->getAllocatedType();
1605 IRBuilder<> Builder(Call);
1606
1607 // Load the current value from the alloca and set it as the
1608 // swifterror value.
1609 auto ValueBeforeCall = Builder.CreateLoad(Ty: ValueTy, Ptr: Alloca);
1610 auto Addr = emitSetSwiftErrorValue(Builder, V: ValueBeforeCall, Shape);
1611
1612 // Move to after the call. Since swifterror only has a guaranteed
1613 // value on normal exits, we can ignore implicit and explicit unwind
1614 // edges.
1615 if (isa<CallInst>(Val: Call)) {
1616 Builder.SetInsertPoint(Call->getNextNode());
1617 } else {
1618 auto Invoke = cast<InvokeInst>(Val: Call);
1619 Builder.SetInsertPoint(Invoke->getNormalDest()->getFirstNonPHIOrDbg());
1620 }
1621
1622 // Get the current swifterror value and store it to the alloca.
1623 auto ValueAfterCall = emitGetSwiftErrorValue(Builder, ValueTy, Shape);
1624 Builder.CreateStore(Val: ValueAfterCall, Ptr: Alloca);
1625
1626 return Addr;
1627}
1628
1629/// Eliminate a formerly-swifterror alloca by inserting the get/set
1630/// intrinsics and attempting to MemToReg the alloca away.
1631static void eliminateSwiftErrorAlloca(Function &F, AllocaInst *Alloca,
1632 coro::Shape &Shape) {
1633 for (Use &Use : llvm::make_early_inc_range(Range: Alloca->uses())) {
1634 // swifterror values can only be used in very specific ways.
1635 // We take advantage of that here.
1636 auto User = Use.getUser();
1637 if (isa<LoadInst>(Val: User) || isa<StoreInst>(Val: User))
1638 continue;
1639
1640 assert(isa<CallInst>(User) || isa<InvokeInst>(User));
1641 auto Call = cast<Instruction>(Val: User);
1642
1643 auto Addr = emitSetAndGetSwiftErrorValueAround(Call, Alloca, Shape);
1644
1645 // Use the returned slot address as the call argument.
1646 Use.set(Addr);
1647 }
1648
1649 // All the uses should be loads and stores now.
1650 assert(isAllocaPromotable(Alloca));
1651}
1652
1653/// "Eliminate" a swifterror argument by reducing it to the alloca case
1654/// and then loading and storing in the prologue and epilog.
1655///
1656/// The argument keeps the swifterror flag.
1657static void eliminateSwiftErrorArgument(Function &F, Argument &Arg,
1658 coro::Shape &Shape,
1659 SmallVectorImpl<AllocaInst*> &AllocasToPromote) {
1660 IRBuilder<> Builder(&F.getEntryBlock(),
1661 F.getEntryBlock().getFirstNonPHIOrDbg());
1662
1663 auto ArgTy = cast<PointerType>(Val: Arg.getType());
1664 auto ValueTy = PointerType::getUnqual(C&: F.getContext());
1665
1666 // Reduce to the alloca case:
1667
1668 // Create an alloca and replace all uses of the arg with it.
1669 auto Alloca = Builder.CreateAlloca(Ty: ValueTy, AddrSpace: ArgTy->getAddressSpace());
1670 Arg.replaceAllUsesWith(V: Alloca);
1671
1672 // Set an initial value in the alloca. swifterror is always null on entry.
1673 auto InitialValue = Constant::getNullValue(Ty: ValueTy);
1674 Builder.CreateStore(Val: InitialValue, Ptr: Alloca);
1675
1676 // Find all the suspends in the function and save and restore around them.
1677 for (auto *Suspend : Shape.CoroSuspends) {
1678 (void) emitSetAndGetSwiftErrorValueAround(Call: Suspend, Alloca, Shape);
1679 }
1680
1681 // Find all the coro.ends in the function and restore the error value.
1682 for (auto *End : Shape.CoroEnds) {
1683 Builder.SetInsertPoint(End);
1684 auto FinalValue = Builder.CreateLoad(Ty: ValueTy, Ptr: Alloca);
1685 (void) emitSetSwiftErrorValue(Builder, V: FinalValue, Shape);
1686 }
1687
1688 // Now we can use the alloca logic.
1689 AllocasToPromote.push_back(Elt: Alloca);
1690 eliminateSwiftErrorAlloca(F, Alloca, Shape);
1691}
1692
1693/// Eliminate all problematic uses of swifterror arguments and allocas
1694/// from the function. We'll fix them up later when splitting the function.
1695static void eliminateSwiftError(Function &F, coro::Shape &Shape) {
1696 SmallVector<AllocaInst*, 4> AllocasToPromote;
1697
1698 // Look for a swifterror argument.
1699 for (auto &Arg : F.args()) {
1700 if (!Arg.hasSwiftErrorAttr()) continue;
1701
1702 eliminateSwiftErrorArgument(F, Arg, Shape, AllocasToPromote);
1703 break;
1704 }
1705
1706 // Look for swifterror allocas.
1707 for (auto &Inst : F.getEntryBlock()) {
1708 auto Alloca = dyn_cast<AllocaInst>(Val: &Inst);
1709 if (!Alloca || !Alloca->isSwiftError()) continue;
1710
1711 // Clear the swifterror flag.
1712 Alloca->setSwiftError(false);
1713
1714 AllocasToPromote.push_back(Elt: Alloca);
1715 eliminateSwiftErrorAlloca(F, Alloca, Shape);
1716 }
1717
1718 // If we have any allocas to promote, compute a dominator tree and
1719 // promote them en masse.
1720 if (!AllocasToPromote.empty()) {
1721 DominatorTree DT(F);
1722 PromoteMemToReg(Allocas: AllocasToPromote, DT);
1723 }
1724}
1725
1726/// For each local variable that all of its user are only used inside one of
1727/// suspended region, we sink their lifetime.start markers to the place where
1728/// after the suspend block. Doing so minimizes the lifetime of each variable,
1729/// hence minimizing the amount of data we end up putting on the frame.
1730static void sinkLifetimeStartMarkers(Function &F, coro::Shape &Shape,
1731 SuspendCrossingInfo &Checker,
1732 const DominatorTree &DT) {
1733 if (F.hasOptNone())
1734 return;
1735
1736 // Collect all possible basic blocks which may dominate all uses of allocas.
1737 SmallPtrSet<BasicBlock *, 4> DomSet;
1738 DomSet.insert(Ptr: &F.getEntryBlock());
1739 for (auto *CSI : Shape.CoroSuspends) {
1740 BasicBlock *SuspendBlock = CSI->getParent();
1741 assert(coro::isSuspendBlock(SuspendBlock) &&
1742 SuspendBlock->getSingleSuccessor() &&
1743 "should have split coro.suspend into its own block");
1744 DomSet.insert(Ptr: SuspendBlock->getSingleSuccessor());
1745 }
1746
1747 for (Instruction &I : instructions(F)) {
1748 AllocaInst* AI = dyn_cast<AllocaInst>(Val: &I);
1749 if (!AI)
1750 continue;
1751
1752 for (BasicBlock *DomBB : DomSet) {
1753 bool Valid = true;
1754 SmallVector<Instruction *, 1> Lifetimes;
1755
1756 auto isLifetimeStart = [](Instruction* I) {
1757 if (auto* II = dyn_cast<IntrinsicInst>(Val: I))
1758 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1759 return false;
1760 };
1761
1762 auto collectLifetimeStart = [&](Instruction *U, AllocaInst *AI) {
1763 if (isLifetimeStart(U)) {
1764 Lifetimes.push_back(Elt: U);
1765 return true;
1766 }
1767 if (!U->hasOneUse() || U->stripPointerCasts() != AI)
1768 return false;
1769 if (isLifetimeStart(U->user_back())) {
1770 Lifetimes.push_back(Elt: U->user_back());
1771 return true;
1772 }
1773 return false;
1774 };
1775
1776 for (User *U : AI->users()) {
1777 Instruction *UI = cast<Instruction>(Val: U);
1778 // For all users except lifetime.start markers, if they are all
1779 // dominated by one of the basic blocks and do not cross
1780 // suspend points as well, then there is no need to spill the
1781 // instruction.
1782 if (!DT.dominates(A: DomBB, B: UI->getParent()) ||
1783 Checker.isDefinitionAcrossSuspend(DefBB: DomBB, U: UI)) {
1784 // Skip lifetime.start, GEP and bitcast used by lifetime.start
1785 // markers.
1786 if (collectLifetimeStart(UI, AI))
1787 continue;
1788 Valid = false;
1789 break;
1790 }
1791 }
1792 // Sink lifetime.start markers to dominate block when they are
1793 // only used outside the region.
1794 if (Valid && Lifetimes.size() != 0) {
1795 auto *NewLifetime = Lifetimes[0]->clone();
1796 NewLifetime->replaceUsesOfWith(From: NewLifetime->getOperand(i: 0), To: AI);
1797 NewLifetime->insertBefore(InsertPos: DomBB->getTerminator()->getIterator());
1798
1799 // All the outsided lifetime.start markers are no longer necessary.
1800 for (Instruction *S : Lifetimes)
1801 S->eraseFromParent();
1802
1803 break;
1804 }
1805 }
1806 }
1807}
1808
1809static std::optional<std::pair<Value &, DIExpression &>>
1810salvageDebugInfoImpl(SmallDenseMap<Argument *, AllocaInst *, 4> &ArgToAllocaMap,
1811 bool UseEntryValue, Function *F, Value *Storage,
1812 DIExpression *Expr, bool SkipOutermostLoad) {
1813 IRBuilder<> Builder(F->getContext());
1814 auto InsertPt = F->getEntryBlock().getFirstInsertionPt();
1815 while (isa<IntrinsicInst>(Val: InsertPt))
1816 ++InsertPt;
1817 Builder.SetInsertPoint(TheBB: &F->getEntryBlock(), IP: InsertPt);
1818
1819 while (auto *Inst = dyn_cast_or_null<Instruction>(Val: Storage)) {
1820 if (auto *LdInst = dyn_cast<LoadInst>(Val: Inst)) {
1821 Storage = LdInst->getPointerOperand();
1822 // FIXME: This is a heuristic that works around the fact that
1823 // LLVM IR debug intrinsics cannot yet distinguish between
1824 // memory and value locations: Because a dbg.declare(alloca) is
1825 // implicitly a memory location no DW_OP_deref operation for the
1826 // last direct load from an alloca is necessary. This condition
1827 // effectively drops the *last* DW_OP_deref in the expression.
1828 if (!SkipOutermostLoad)
1829 Expr = DIExpression::prepend(Expr, Flags: DIExpression::DerefBefore);
1830 } else if (auto *StInst = dyn_cast<StoreInst>(Val: Inst)) {
1831 Storage = StInst->getValueOperand();
1832 } else {
1833 SmallVector<uint64_t, 16> Ops;
1834 SmallVector<Value *, 0> AdditionalValues;
1835 Value *Op = llvm::salvageDebugInfoImpl(
1836 I&: *Inst, CurrentLocOps: Expr ? Expr->getNumLocationOperands() : 0, Ops,
1837 AdditionalValues);
1838 if (!Op || !AdditionalValues.empty()) {
1839 // If salvaging failed or salvaging produced more than one location
1840 // operand, give up.
1841 break;
1842 }
1843 Storage = Op;
1844 Expr = DIExpression::appendOpsToArg(Expr, Ops, ArgNo: 0, /*StackValue*/ false);
1845 }
1846 SkipOutermostLoad = false;
1847 }
1848 if (!Storage)
1849 return std::nullopt;
1850
1851 auto *StorageAsArg = dyn_cast<Argument>(Val: Storage);
1852
1853 const bool IsSingleLocationExpression = Expr->isSingleLocationExpression();
1854 // Use an EntryValue when requested (UseEntryValue) for swift async Arguments.
1855 // Entry values in variadic expressions are not supported.
1856 const bool WillUseEntryValue =
1857 UseEntryValue && StorageAsArg &&
1858 StorageAsArg->hasAttribute(Kind: Attribute::SwiftAsync) &&
1859 !Expr->isEntryValue() && IsSingleLocationExpression;
1860
1861 if (WillUseEntryValue)
1862 Expr = DIExpression::prepend(Expr, Flags: DIExpression::EntryValue);
1863
1864 // If the coroutine frame is an Argument, store it in an alloca to improve
1865 // its availability (e.g. registers may be clobbered).
1866 // Avoid this if the value is guaranteed to be available through other means
1867 // (e.g. swift ABI guarantees).
1868 // Avoid this if multiple location expressions are involved, as LLVM does not
1869 // know how to prepend a deref in this scenario.
1870 if (StorageAsArg && !WillUseEntryValue && IsSingleLocationExpression) {
1871 auto &Cached = ArgToAllocaMap[StorageAsArg];
1872 if (!Cached) {
1873 Cached = Builder.CreateAlloca(Ty: Storage->getType(), AddrSpace: 0, ArraySize: nullptr,
1874 Name: Storage->getName() + ".debug");
1875 Builder.CreateStore(Val: Storage, Ptr: Cached);
1876 }
1877 Storage = Cached;
1878 // FIXME: LLVM lacks nuanced semantics to differentiate between
1879 // memory and direct locations at the IR level. The backend will
1880 // turn a dbg.declare(alloca, ..., DIExpression()) into a memory
1881 // location. Thus, if there are deref and offset operations in the
1882 // expression, we need to add a DW_OP_deref at the *start* of the
1883 // expression to first load the contents of the alloca before
1884 // adjusting it with the expression.
1885 Expr = DIExpression::prepend(Expr, Flags: DIExpression::DerefBefore);
1886 }
1887
1888 Expr = Expr->foldConstantMath();
1889 return {{*Storage, *Expr}};
1890}
1891
1892void coro::salvageDebugInfo(
1893 SmallDenseMap<Argument *, AllocaInst *, 4> &ArgToAllocaMap,
1894 DbgVariableRecord &DVR, bool UseEntryValue) {
1895
1896 Function *F = DVR.getFunction();
1897 // Follow the pointer arithmetic all the way to the incoming
1898 // function argument and convert into a DIExpression.
1899 bool SkipOutermostLoad = DVR.isDbgDeclare() || DVR.isDbgDeclareValue();
1900 Value *OriginalStorage = DVR.getVariableLocationOp(OpIdx: 0);
1901
1902 auto SalvagedInfo =
1903 ::salvageDebugInfoImpl(ArgToAllocaMap, UseEntryValue, F, Storage: OriginalStorage,
1904 Expr: DVR.getExpression(), SkipOutermostLoad);
1905 if (!SalvagedInfo)
1906 return;
1907
1908 Value *Storage = &SalvagedInfo->first;
1909 DIExpression *Expr = &SalvagedInfo->second;
1910
1911 DVR.replaceVariableLocationOp(OldValue: OriginalStorage, NewValue: Storage);
1912 DVR.setExpression(Expr);
1913 // We only hoist dbg.declare and dbg.declare_value today since it doesn't make
1914 // sense to hoist dbg.value since it does not have the same function wide
1915 // guarantees that dbg.declare does.
1916 if (DVR.getType() == DbgVariableRecord::LocationType::Declare ||
1917 DVR.getType() == DbgVariableRecord::LocationType::DeclareValue) {
1918 std::optional<BasicBlock::iterator> InsertPt;
1919 if (auto *I = dyn_cast<Instruction>(Val: Storage)) {
1920 InsertPt = I->getInsertionPointAfterDef();
1921 // Update DILocation only if variable was not inlined.
1922 DebugLoc ILoc = I->getDebugLoc();
1923 DebugLoc DVRLoc = DVR.getDebugLoc();
1924 if (ILoc && DVRLoc &&
1925 DVRLoc->getScope()->getSubprogram() ==
1926 ILoc->getScope()->getSubprogram())
1927 DVR.setDebugLoc(ILoc);
1928 } else if (isa<Argument>(Val: Storage))
1929 InsertPt = F->getEntryBlock().begin();
1930 if (InsertPt) {
1931 DVR.removeFromParent();
1932 // If there is a dbg.declare_value being reinserted, insert it as a
1933 // dbg.declare instead, so that subsequent passes don't have to deal with
1934 // a dbg.declare_value.
1935 if (DVR.getType() == DbgVariableRecord::LocationType::DeclareValue) {
1936 auto *MD = DVR.getRawLocation();
1937 if (auto *VAM = dyn_cast<ValueAsMetadata>(Val: MD)) {
1938 Type *Ty = VAM->getValue()->getType();
1939 if (Ty->isPointerTy())
1940 DVR.Type = DbgVariableRecord::LocationType::Declare;
1941 else
1942 DVR.Type = DbgVariableRecord::LocationType::Value;
1943 }
1944 }
1945 (*InsertPt)->getParent()->insertDbgRecordBefore(DR: &DVR, Here: *InsertPt);
1946 }
1947 }
1948}
1949
1950void coro::normalizeCoroutine(Function &F, coro::Shape &Shape,
1951 TargetTransformInfo &TTI) {
1952 // Don't eliminate swifterror in async functions that won't be split.
1953 if (Shape.ABI != coro::ABI::Async || !Shape.CoroSuspends.empty())
1954 eliminateSwiftError(F, Shape);
1955
1956 if (Shape.ABI == coro::ABI::Switch &&
1957 Shape.SwitchLowering.PromiseAlloca) {
1958 Shape.getSwitchCoroId()->clearPromise();
1959 }
1960
1961 // Make sure that all coro.save, coro.suspend and the fallthrough coro.end
1962 // intrinsics are in their own blocks to simplify the logic of building up
1963 // SuspendCrossing data.
1964 for (auto *CSI : Shape.CoroSuspends) {
1965 if (auto *Save = CSI->getCoroSave())
1966 splitAround(I: Save, Name: "CoroSave");
1967 splitAround(I: CSI, Name: "CoroSuspend");
1968 }
1969
1970 // Put CoroEnds into their own blocks.
1971 for (AnyCoroEndInst *CE : Shape.CoroEnds) {
1972 splitAround(I: CE, Name: "CoroEnd");
1973
1974 // Emit the musttail call function in a new block before the CoroEnd.
1975 // We do this here so that the right suspend crossing info is computed for
1976 // the uses of the musttail call function call. (Arguments to the coro.end
1977 // instructions would be ignored)
1978 if (auto *AsyncEnd = dyn_cast<CoroAsyncEndInst>(Val: CE)) {
1979 auto *MustTailCallFn = AsyncEnd->getMustTailCallFunction();
1980 if (!MustTailCallFn)
1981 continue;
1982 IRBuilder<> Builder(AsyncEnd);
1983 SmallVector<Value *, 8> Args(AsyncEnd->args());
1984 auto Arguments = ArrayRef<Value *>(Args).drop_front(N: 3);
1985 auto *Call = coro::createMustTailCall(
1986 Loc: AsyncEnd->getDebugLoc(), MustTailCallFn, TTI, Arguments, Builder);
1987 splitAround(I: Call, Name: "MustTailCall.Before.CoroEnd");
1988 }
1989 }
1990
1991 // Later code makes structural assumptions about single predecessors phis e.g
1992 // that they are not live across a suspend point.
1993 cleanupSinglePredPHIs(F);
1994
1995 // Transforms multi-edge PHI Nodes, so that any value feeding into a PHI will
1996 // never have its definition separated from the PHI by the suspend point.
1997 rewritePHIs(F);
1998}
1999
2000void coro::BaseABI::buildCoroutineFrame(bool OptimizeFrame) {
2001 SuspendCrossingInfo Checker(F, Shape);
2002 doRematerializations(F, Checker, IsMaterializable);
2003
2004 const DominatorTree DT(F);
2005 if (Shape.ABI != coro::ABI::Async && Shape.ABI != coro::ABI::Retcon &&
2006 Shape.ABI != coro::ABI::RetconOnce)
2007 sinkLifetimeStartMarkers(F, Shape, Checker, DT);
2008
2009 // All values (that are not allocas) that needs to be spilled to the frame.
2010 coro::SpillInfo Spills;
2011 // All values defined as allocas that need to live in the frame.
2012 SmallVector<coro::AllocaInfo, 8> Allocas;
2013
2014 // Collect the spills for arguments and other not-materializable values.
2015 coro::collectSpillsFromArgs(Spills, F, Checker);
2016 SmallVector<Instruction *, 4> DeadInstructions;
2017 SmallVector<CoroAllocaAllocInst *, 4> LocalAllocas;
2018 coro::collectSpillsAndAllocasFromInsts(Spills, Allocas, DeadInstructions,
2019 LocalAllocas, F, Checker, DT, Shape);
2020 coro::collectSpillsFromDbgInfo(Spills, F, Checker);
2021
2022 LLVM_DEBUG(dumpAllocas(Allocas));
2023 LLVM_DEBUG(dumpSpills("Spills", Spills));
2024
2025 if (Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce ||
2026 Shape.ABI == coro::ABI::Async)
2027 sinkSpillUsesAfterCoroBegin(DT, CoroBegin: Shape.CoroBegin, Spills, Allocas);
2028
2029 // Build frame layout
2030 FrameDataInfo FrameData(Spills, Allocas);
2031 buildFrameLayout(F, DT, Shape, FrameData, OptimizeFrame);
2032 Shape.FramePtr = Shape.CoroBegin;
2033 // For now, this works for C++ programs only.
2034 buildFrameDebugInfo(F, Shape, FrameData);
2035 // Insert spills and reloads
2036 insertSpills(FrameData, Shape);
2037 lowerLocalAllocas(LocalAllocas, DeadInsts&: DeadInstructions);
2038
2039 for (auto *I : DeadInstructions)
2040 I->eraseFromParent();
2041}
2042