1//===- CoroSplit.cpp - Converts a coroutine into a state machine ----------===//
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 pass builds the coroutine frame and outlines resume and destroy parts
9// of the coroutine into separate functions.
10//
11// We present a coroutine to an LLVM as an ordinary function with suspension
12// points marked up with intrinsics. We let the optimizer party on the coroutine
13// as a single function for as long as possible. Shortly before the coroutine is
14// eligible to be inlined into its callers, we split up the coroutine into parts
15// corresponding to an initial, resume and destroy invocations of the coroutine,
16// add them to the current SCC and restart the IPO pipeline to optimize the
17// coroutine subfunctions we extracted before proceeding to the caller of the
18// coroutine.
19//===----------------------------------------------------------------------===//
20
21#include "llvm/Transforms/Coroutines/CoroSplit.h"
22#include "CoroCloner.h"
23#include "CoroInternal.h"
24#include "llvm/ADT/DenseMap.h"
25#include "llvm/ADT/PriorityWorklist.h"
26#include "llvm/ADT/STLExtras.h"
27#include "llvm/ADT/SmallPtrSet.h"
28#include "llvm/ADT/SmallVector.h"
29#include "llvm/ADT/StringExtras.h"
30#include "llvm/ADT/StringRef.h"
31#include "llvm/ADT/Twine.h"
32#include "llvm/Analysis/BlockFrequencyInfo.h"
33#include "llvm/Analysis/CFG.h"
34#include "llvm/Analysis/CallGraph.h"
35#include "llvm/Analysis/ConstantFolding.h"
36#include "llvm/Analysis/LazyCallGraph.h"
37#include "llvm/Analysis/OptimizationRemarkEmitter.h"
38#include "llvm/Analysis/TargetTransformInfo.h"
39#include "llvm/BinaryFormat/Dwarf.h"
40#include "llvm/IR/Argument.h"
41#include "llvm/IR/Attributes.h"
42#include "llvm/IR/BasicBlock.h"
43#include "llvm/IR/CFG.h"
44#include "llvm/IR/Constants.h"
45#include "llvm/IR/DIBuilder.h"
46#include "llvm/IR/DataLayout.h"
47#include "llvm/IR/DebugInfo.h"
48#include "llvm/IR/DerivedTypes.h"
49#include "llvm/IR/Dominators.h"
50#include "llvm/IR/GlobalValue.h"
51#include "llvm/IR/GlobalVariable.h"
52#include "llvm/IR/InstIterator.h"
53#include "llvm/IR/InstrTypes.h"
54#include "llvm/IR/Instruction.h"
55#include "llvm/IR/Instructions.h"
56#include "llvm/IR/IntrinsicInst.h"
57#include "llvm/IR/LLVMContext.h"
58#include "llvm/IR/MDBuilder.h"
59#include "llvm/IR/Module.h"
60#include "llvm/IR/ProfDataUtils.h"
61#include "llvm/IR/Type.h"
62#include "llvm/IR/Value.h"
63#include "llvm/IR/Verifier.h"
64#include "llvm/Support/Casting.h"
65#include "llvm/Support/Debug.h"
66#include "llvm/Support/PrettyStackTrace.h"
67#include "llvm/Support/raw_ostream.h"
68#include "llvm/Transforms/Coroutines/MaterializationUtils.h"
69#include "llvm/Transforms/Scalar.h"
70#include "llvm/Transforms/Utils/BasicBlockUtils.h"
71#include "llvm/Transforms/Utils/CallGraphUpdater.h"
72#include "llvm/Transforms/Utils/Cloning.h"
73#include "llvm/Transforms/Utils/Local.h"
74#include <cassert>
75#include <cstddef>
76#include <cstdint>
77#include <initializer_list>
78#include <iterator>
79
80using namespace llvm;
81
82#define DEBUG_TYPE "coro-split"
83
84// FIXME:
85// Lower the intrinisc in CoroEarly phase if coroutine frame doesn't escape
86// and it is known that other transformations, for example, sanitizers
87// won't lead to incorrect code.
88static void lowerAwaitSuspend(IRBuilder<> &Builder, CoroAwaitSuspendInst *CB,
89 coro::Shape &Shape) {
90 auto Wrapper = CB->getWrapperFunction();
91 auto Awaiter = CB->getAwaiter();
92 auto FramePtr = CB->getFrame();
93
94 Builder.SetInsertPoint(CB);
95
96 CallBase *NewCall = nullptr;
97 // await_suspend has only 2 parameters, awaiter and handle.
98 // Copy parameter attributes from the intrinsic call, but remove the last,
99 // because the last parameter now becomes the function that is being called.
100 AttributeList NewAttributes =
101 CB->getAttributes().removeParamAttributes(C&: CB->getContext(), ArgNo: 2);
102
103 if (auto Invoke = dyn_cast<InvokeInst>(Val: CB)) {
104 auto WrapperInvoke =
105 Builder.CreateInvoke(Callee: Wrapper, NormalDest: Invoke->getNormalDest(),
106 UnwindDest: Invoke->getUnwindDest(), Args: {Awaiter, FramePtr});
107
108 WrapperInvoke->setCallingConv(Invoke->getCallingConv());
109 std::copy(first: Invoke->bundle_op_info_begin(), last: Invoke->bundle_op_info_end(),
110 result: WrapperInvoke->bundle_op_info_begin());
111 WrapperInvoke->setAttributes(NewAttributes);
112 WrapperInvoke->setDebugLoc(Invoke->getDebugLoc());
113 NewCall = WrapperInvoke;
114 } else if (auto Call = dyn_cast<CallInst>(Val: CB)) {
115 auto WrapperCall = Builder.CreateCall(Callee: Wrapper, Args: {Awaiter, FramePtr});
116
117 WrapperCall->setAttributes(NewAttributes);
118 WrapperCall->setDebugLoc(Call->getDebugLoc());
119 NewCall = WrapperCall;
120 } else {
121 llvm_unreachable("Unexpected coro_await_suspend invocation method");
122 }
123
124 if (CB->getCalledFunction()->getIntrinsicID() ==
125 Intrinsic::coro_await_suspend_handle) {
126 // Follow the lowered await_suspend call above with a lowered resume call
127 // to the returned coroutine.
128 if (auto *Invoke = dyn_cast<InvokeInst>(Val: CB)) {
129 // If the await_suspend call is an invoke, we continue in the next block.
130 Builder.SetInsertPoint(Invoke->getNormalDest()->getFirstInsertionPt());
131 }
132
133 coro::LowererBase LB(*Wrapper->getParent());
134 auto *ResumeAddr = LB.makeSubFnCall(Arg: NewCall, Index: CoroSubFnInst::ResumeIndex,
135 InsertPt: &*Builder.GetInsertPoint());
136
137 LLVMContext &Ctx = Builder.getContext();
138 FunctionType *ResumeTy = FunctionType::get(
139 Result: Type::getVoidTy(C&: Ctx), Params: PointerType::getUnqual(C&: Ctx), isVarArg: false);
140 auto *ResumeCall = Builder.CreateCall(FTy: ResumeTy, Callee: ResumeAddr, Args: {NewCall});
141
142 // We can't insert the 'ret' instruction and adjust the cc until the
143 // function has been split, so remember this for later.
144 Shape.SymmetricTransfers.push_back(Elt: ResumeCall);
145
146 NewCall = ResumeCall;
147 }
148
149 CB->replaceAllUsesWith(V: NewCall);
150 CB->eraseFromParent();
151}
152
153static void lowerAwaitSuspends(Function &F, coro::Shape &Shape) {
154 IRBuilder<> Builder(F.getContext());
155 for (auto *AWS : Shape.CoroAwaitSuspends)
156 lowerAwaitSuspend(Builder, CB: AWS, Shape);
157}
158
159static void maybeFreeRetconStorage(IRBuilder<> &Builder,
160 const coro::Shape &Shape, Value *FramePtr,
161 CallGraph *CG) {
162 assert(Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce);
163 if (Shape.RetconLowering.IsFrameInlineInStorage)
164 return;
165
166 Shape.emitDealloc(Builder, Ptr: FramePtr, CG);
167}
168
169/// Create a pointer to the switch destroy function field in the coroutine
170/// frame.
171static Value *createSwitchDestroyPtr(const coro::Shape &Shape,
172 IRBuilder<> &Builder, Value *FramePtr) {
173 auto *Offset = ConstantInt::get(Ty: Type::getInt64Ty(C&: FramePtr->getContext()),
174 V: Shape.SwitchLowering.DestroyOffset);
175 return Builder.CreateInBoundsPtrAdd(Ptr: FramePtr, Offset, Name: "destroy.addr");
176}
177
178/// Make resume-clone coro.free conditional on whether the frame is elided.
179///
180/// The destroy slot holds the cleanup clone for an elided frame and the destroy
181/// clone for a heap frame. Load it before user code can reentrantly destroy the
182/// enclosing caller frame, then use the cached comparison to suppress only the
183/// deallocation. The resume clone has already performed the shared coroutine
184/// cleanup, so calling either clone here would run that cleanup twice.
185static void replaceSwitchResumeCoroFree(const coro::Shape &Shape,
186 Function &Resume, Function &Cleanup) {
187 Value *FramePtr = Resume.getArg(i: 0);
188 IRBuilder<> EntryBuilder(Resume.getEntryBlock().getTerminator());
189 Value *DestroyAddr = createSwitchDestroyPtr(Shape, Builder&: EntryBuilder, FramePtr);
190 Value *DestroyFn = EntryBuilder.CreateLoad(Ty: Shape.getSwitchResumePointerType(),
191 Ptr: DestroyAddr, Name: "destroy");
192 Value *CleanupFn =
193 EntryBuilder.CreatePointerCast(V: &Cleanup, DestTy: DestroyFn->getType());
194 Value *IsElided =
195 EntryBuilder.CreateICmpEQ(LHS: DestroyFn, RHS: CleanupFn, Name: "is.elided");
196
197 SmallVector<CoroFreeInst *, 4> CoroFrees;
198 for (User *U : FramePtr->users()) {
199 if (auto *CF = dyn_cast<CoroFreeInst>(Val: U))
200 CoroFrees.push_back(Elt: CF);
201 }
202
203 for (CoroFreeInst *CF : CoroFrees) {
204 IRBuilder<> Builder(CF);
205 auto *Null = ConstantPointerNull::get(T: cast<PointerType>(Val: CF->getType()));
206 Value *Replacement =
207 Builder.CreateSelect(C: IsElided, True: Null, False: FramePtr, Name: "coro.free");
208 // Add unknown branch weights to the select since whether the frame is
209 // heap-allocated or elided cannot be determined.
210 applyProfMetadataIfEnabled(V: Replacement, setMetadataCallback: [&](Instruction *Inst) {
211 setExplicitlyUnknownBranchWeightsIfProfiled(I&: *Inst, DEBUG_TYPE,
212 F: Inst->getFunction());
213 });
214 CF->replaceAllUsesWith(V: Replacement);
215 CF->eraseFromParent();
216 }
217}
218
219/// Replace an llvm.coro.end.async.
220/// Will inline the must tail call function call if there is one.
221/// \returns true if cleanup of the coro.end block is needed, false otherwise.
222static bool replaceCoroEndAsync(AnyCoroEndInst *End) {
223 IRBuilder<> Builder(End);
224
225 auto *EndAsync = dyn_cast<CoroAsyncEndInst>(Val: End);
226 if (!EndAsync) {
227 Builder.CreateRetVoid();
228 return true /*needs cleanup of coro.end block*/;
229 }
230
231 auto *MustTailCallFunc = EndAsync->getMustTailCallFunction();
232 if (!MustTailCallFunc) {
233 Builder.CreateRetVoid();
234 return true /*needs cleanup of coro.end block*/;
235 }
236
237 // Move the must tail call from the predecessor block into the end block.
238 auto *CoroEndBlock = End->getParent();
239 auto *MustTailCallFuncBlock = CoroEndBlock->getSinglePredecessor();
240 assert(MustTailCallFuncBlock && "Must have a single predecessor block");
241 auto It = MustTailCallFuncBlock->getTerminator()->getIterator();
242 auto *MustTailCall = cast<CallInst>(Val: &*std::prev(x: It));
243 CoroEndBlock->splice(ToIt: End->getIterator(), FromBB: MustTailCallFuncBlock,
244 FromIt: MustTailCall->getIterator());
245
246 // Insert the return instruction.
247 Builder.SetInsertPoint(End);
248 Builder.CreateRetVoid();
249 InlineFunctionInfo FnInfo;
250
251 // Remove the rest of the block, by splitting it into an unreachable block.
252 auto *BB = End->getParent();
253 BB->splitBasicBlock(I: End);
254 BB->getTerminator()->eraseFromParent();
255
256 auto InlineRes = InlineFunction(CB&: *MustTailCall, IFI&: FnInfo);
257 assert(InlineRes.isSuccess() && "Expected inlining to succeed");
258 (void)InlineRes;
259
260 // We have cleaned up the coro.end block above.
261 return false;
262}
263
264/// Replace a non-unwind call to llvm.coro.end.
265static void replaceFallthroughCoroEnd(AnyCoroEndInst *End,
266 const coro::Shape &Shape, Value *FramePtr,
267 bool InRamp, CallGraph *CG) {
268 // Start inserting right before the coro.end.
269 IRBuilder<> Builder(End);
270
271 // Create the return instruction.
272 switch (Shape.ABI) {
273 // The cloned functions in switch-lowering always return void.
274 case coro::ABI::Switch:
275 assert(!cast<CoroEndInst>(End)->hasResults() &&
276 "switch coroutine should not return any values");
277 // coro.end doesn't immediately end the coroutine in the main function
278 // in this lowering, because we need to deallocate the coroutine.
279 if (InRamp)
280 return;
281 Builder.CreateRetVoid();
282 break;
283
284 // In async lowering this returns.
285 case coro::ABI::Async: {
286 bool CoroEndBlockNeedsCleanup = replaceCoroEndAsync(End);
287 if (!CoroEndBlockNeedsCleanup)
288 return;
289 break;
290 }
291
292 // In unique continuation lowering, the continuations always return void.
293 // But we may have implicitly allocated storage.
294 case coro::ABI::RetconOnce: {
295 maybeFreeRetconStorage(Builder, Shape, FramePtr, CG);
296 auto *CoroEnd = cast<CoroEndInst>(Val: End);
297 auto *RetTy = Shape.getResumeFunctionType()->getReturnType();
298
299 if (!CoroEnd->hasResults()) {
300 assert(RetTy->isVoidTy());
301 Builder.CreateRetVoid();
302 break;
303 }
304
305 auto *CoroResults = CoroEnd->getResults();
306 unsigned NumReturns = CoroResults->numReturns();
307
308 if (auto *RetStructTy = dyn_cast<StructType>(Val: RetTy)) {
309 assert(RetStructTy->getNumElements() == NumReturns &&
310 "numbers of returns should match resume function singature");
311 Value *ReturnValue = PoisonValue::get(T: RetStructTy);
312 unsigned Idx = 0;
313 for (Value *RetValEl : CoroResults->return_values())
314 ReturnValue = Builder.CreateInsertValue(Agg: ReturnValue, Val: RetValEl, Idxs: Idx++);
315 Builder.CreateRet(V: ReturnValue);
316 } else if (NumReturns == 0) {
317 assert(RetTy->isVoidTy());
318 Builder.CreateRetVoid();
319 } else {
320 assert(NumReturns == 1);
321 Builder.CreateRet(V: *CoroResults->retval_begin());
322 }
323 CoroResults->replaceAllUsesWith(
324 V: ConstantTokenNone::get(Context&: CoroResults->getContext()));
325 CoroResults->eraseFromParent();
326 break;
327 }
328
329 // In non-unique continuation lowering, we signal completion by returning
330 // a null continuation.
331 case coro::ABI::Retcon: {
332 assert(!cast<CoroEndInst>(End)->hasResults() &&
333 "retcon coroutine should not return any values");
334 maybeFreeRetconStorage(Builder, Shape, FramePtr, CG);
335 auto RetTy = Shape.getResumeFunctionType()->getReturnType();
336 auto RetStructTy = dyn_cast<StructType>(Val: RetTy);
337 PointerType *ContinuationTy =
338 cast<PointerType>(Val: RetStructTy ? RetStructTy->getElementType(N: 0) : RetTy);
339
340 Value *ReturnValue = ConstantPointerNull::get(T: ContinuationTy);
341 if (RetStructTy) {
342 ReturnValue = Builder.CreateInsertValue(Agg: PoisonValue::get(T: RetStructTy),
343 Val: ReturnValue, Idxs: 0);
344 }
345 Builder.CreateRet(V: ReturnValue);
346 break;
347 }
348 }
349
350 // Remove the rest of the block, by splitting it into an unreachable block.
351 auto *BB = End->getParent();
352 BB->splitBasicBlock(I: End);
353 BB->getTerminator()->eraseFromParent();
354}
355
356/// Create a pointer to the switch index field in the coroutine frame.
357static Value *createSwitchIndexPtr(const coro::Shape &Shape,
358 IRBuilder<> &Builder, Value *FramePtr) {
359 auto *Offset = ConstantInt::get(Ty: Type::getInt64Ty(C&: FramePtr->getContext()),
360 V: Shape.SwitchLowering.IndexOffset);
361 return Builder.CreateInBoundsPtrAdd(Ptr: FramePtr, Offset, Name: "index.addr");
362}
363
364// Mark a coroutine as done, which implies that the coroutine is finished and
365// never gets resumed.
366//
367// In resume-switched ABI, the done state is represented by storing zero in
368// ResumeFnAddr.
369//
370// NOTE: We couldn't omit the argument `FramePtr`. It is necessary because the
371// pointer to the frame in splitted function is not stored in `Shape`.
372static void markCoroutineAsDone(IRBuilder<> &Builder, const coro::Shape &Shape,
373 Value *FramePtr) {
374 assert(
375 Shape.ABI == coro::ABI::Switch &&
376 "markCoroutineAsDone is only supported for Switch-Resumed ABI for now.");
377 // Resume function pointer is always first
378 auto *NullPtr = ConstantPointerNull::get(T: Shape.getSwitchResumePointerType());
379 Builder.CreateStore(Val: NullPtr, Ptr: FramePtr);
380
381 // If the coroutine don't have unwind coro end, we could omit the store to
382 // the final suspend point since we could infer the coroutine is suspended
383 // at the final suspend point by the nullness of ResumeFnAddr.
384 // However, we can't skip it if the coroutine have unwind coro end. Since
385 // the coroutine reaches unwind coro end is considered suspended at the
386 // final suspend point (the ResumeFnAddr is null) but in fact the coroutine
387 // didn't complete yet. We need the IndexVal for the final suspend point
388 // to make the states clear.
389 if (Shape.SwitchLowering.HasUnwindCoroEnd &&
390 Shape.SwitchLowering.HasFinalSuspend) {
391 assert(cast<CoroSuspendInst>(Shape.CoroSuspends.back())->isFinal() &&
392 "The final suspend should only live in the last position of "
393 "CoroSuspends.");
394 ConstantInt *IndexVal = Shape.getIndex(Value: Shape.CoroSuspends.size() - 1);
395 Value *FinalIndex = createSwitchIndexPtr(Shape, Builder, FramePtr);
396 Builder.CreateStore(Val: IndexVal, Ptr: FinalIndex);
397 }
398}
399
400/// Replace an unwind call to llvm.coro.end.
401static void replaceUnwindCoroEnd(AnyCoroEndInst *End, const coro::Shape &Shape,
402 Value *FramePtr, bool InRamp, CallGraph *CG) {
403 IRBuilder<> Builder(End);
404
405 switch (Shape.ABI) {
406 // In switch-lowering, this does nothing in the main function.
407 case coro::ABI::Switch: {
408 // In C++'s specification, the coroutine should be marked as done
409 // if promise.unhandled_exception() throws. The frontend will
410 // call coro.end(true) along this path.
411 //
412 // FIXME: We should refactor this once there is other language
413 // which uses Switch-Resumed style other than C++.
414 markCoroutineAsDone(Builder, Shape, FramePtr);
415 if (InRamp)
416 return;
417 break;
418 }
419 // In async lowering this does nothing.
420 case coro::ABI::Async:
421 break;
422 // In continuation-lowering, this frees the continuation storage.
423 case coro::ABI::Retcon:
424 case coro::ABI::RetconOnce:
425 maybeFreeRetconStorage(Builder, Shape, FramePtr, CG);
426 break;
427 }
428
429 // If coro.end has an associated bundle, add cleanupret instruction.
430 if (auto Bundle = End->getOperandBundle(ID: LLVMContext::OB_funclet)) {
431 auto *FromPad = cast<CleanupPadInst>(Val: Bundle->Inputs[0]);
432 auto *CleanupRet = Builder.CreateCleanupRet(CleanupPad: FromPad, UnwindBB: nullptr);
433 End->getParent()->splitBasicBlock(I: End);
434 CleanupRet->getParent()->getTerminator()->eraseFromParent();
435 }
436}
437
438static void replaceCoroEnd(AnyCoroEndInst *End, const coro::Shape &Shape,
439 Value *FramePtr, bool InRamp, CallGraph *CG) {
440 if (End->isUnwind())
441 replaceUnwindCoroEnd(End, Shape, FramePtr, InRamp, CG);
442 else
443 replaceFallthroughCoroEnd(End, Shape, FramePtr, InRamp, CG);
444 End->eraseFromParent();
445}
446
447// In the resume function, we remove the last case (when coro::Shape is built,
448// the final suspend point (if present) is always the last element of
449// CoroSuspends array) since it is an undefined behavior to resume a coroutine
450// suspended at the final suspend point.
451// In the destroy function, if it isn't possible that the ResumeFnAddr is NULL
452// and the coroutine doesn't suspend at the final suspend point actually (this
453// is possible since the coroutine is considered suspended at the final suspend
454// point if promise.unhandled_exception() exits via an exception), we can
455// remove the last case.
456void coro::BaseCloner::handleFinalSuspend() {
457 assert(Shape.ABI == coro::ABI::Switch &&
458 Shape.SwitchLowering.HasFinalSuspend);
459
460 if (isSwitchDestroyFunction() && Shape.SwitchLowering.HasUnwindCoroEnd)
461 return;
462
463 auto *Switch = cast<SwitchInst>(Val&: VMap[Shape.SwitchLowering.ResumeSwitch]);
464 auto FinalCaseIt = std::prev(x: Switch->case_end());
465 BasicBlock *ResumeBB = FinalCaseIt->getCaseSuccessor();
466
467 // Use SwitchInstProfUpdateWrapper to remove the case, keeping the profile
468 // branch weights in sync with the switch successors.
469 SwitchInstProfUpdateWrapper SwitchWrapper(*Switch);
470 SwitchWrapper.removeCase(I: FinalCaseIt);
471 if (isSwitchDestroyFunction()) {
472 BasicBlock *OldSwitchBB = Switch->getParent();
473 auto *NewSwitchBB = OldSwitchBB->splitBasicBlock(I: Switch, BBName: "Switch");
474 Builder.SetInsertPoint(OldSwitchBB->getTerminator());
475
476 if (NewF->isCoroOnlyDestroyWhenComplete()) {
477 // When the coroutine can only be destroyed when complete, we don't need
478 // to generate code for other cases.
479 Builder.CreateBr(Dest: ResumeBB);
480 } else {
481 // Resume function pointer is always first
482 auto *Load =
483 Builder.CreateLoad(Ty: Shape.getSwitchResumePointerType(), Ptr: NewFramePtr);
484 auto *Cond = Builder.CreateIsNull(Arg: Load);
485 auto *Br = Builder.CreateCondBr(Cond, True: ResumeBB, False: NewSwitchBB);
486 applyProfMetadataIfEnabled(V: Br, setMetadataCallback: [&](Instruction *Inst) {
487 setExplicitlyUnknownBranchWeightsIfProfiled(I&: *Inst, DEBUG_TYPE,
488 F: Inst->getFunction());
489 });
490 }
491 OldSwitchBB->getTerminator()->eraseFromParent();
492 }
493}
494
495static FunctionType *
496getFunctionTypeFromAsyncSuspend(AnyCoroSuspendInst *Suspend) {
497 auto *AsyncSuspend = cast<CoroSuspendAsyncInst>(Val: Suspend);
498 auto *StructTy = cast<StructType>(Val: AsyncSuspend->getType());
499 auto &Context = Suspend->getParent()->getParent()->getContext();
500 auto *VoidTy = Type::getVoidTy(C&: Context);
501 return FunctionType::get(Result: VoidTy, Params: StructTy->elements(), isVarArg: false);
502}
503
504static Function *createCloneDeclaration(Function &OrigF, coro::Shape &Shape,
505 const Twine &Suffix,
506 Module::iterator InsertBefore,
507 AnyCoroSuspendInst *ActiveSuspend) {
508 Module *M = OrigF.getParent();
509 auto *FnTy = (Shape.ABI != coro::ABI::Async)
510 ? Shape.getResumeFunctionType()
511 : getFunctionTypeFromAsyncSuspend(Suspend: ActiveSuspend);
512
513 Function *NewF =
514 Function::Create(Ty: FnTy, Linkage: GlobalValue::LinkageTypes::InternalLinkage,
515 AddrSpace: OrigF.getAddressSpace(), N: OrigF.getName() + Suffix);
516
517 M->getFunctionList().insert(where: InsertBefore, New: NewF);
518
519 return NewF;
520}
521
522/// Replace uses of the active llvm.coro.suspend.retcon/async call with the
523/// arguments to the continuation function.
524///
525/// This assumes that the builder has a meaningful insertion point.
526void coro::BaseCloner::replaceRetconOrAsyncSuspendUses() {
527 assert(Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce ||
528 Shape.ABI == coro::ABI::Async);
529
530 auto NewS = VMap[ActiveSuspend];
531 if (NewS->use_empty())
532 return;
533
534 // Copy out all the continuation arguments after the buffer pointer into
535 // an easily-indexed data structure for convenience.
536 SmallVector<Value *, 8> Args;
537 // The async ABI includes all arguments -- including the first argument.
538 bool IsAsyncABI = Shape.ABI == coro::ABI::Async;
539 for (auto I = IsAsyncABI ? NewF->arg_begin() : std::next(x: NewF->arg_begin()),
540 E = NewF->arg_end();
541 I != E; ++I)
542 Args.push_back(Elt: &*I);
543
544 // If the suspend returns a single scalar value, we can just do a simple
545 // replacement.
546 if (!isa<StructType>(Val: NewS->getType())) {
547 assert(Args.size() == 1);
548 NewS->replaceAllUsesWith(V: Args.front());
549 return;
550 }
551
552 // Try to peephole extracts of an aggregate return.
553 for (Use &U : llvm::make_early_inc_range(Range: NewS->uses())) {
554 auto *EVI = dyn_cast<ExtractValueInst>(Val: U.getUser());
555 if (!EVI || EVI->getNumIndices() != 1)
556 continue;
557
558 EVI->replaceAllUsesWith(V: Args[EVI->getIndices().front()]);
559 EVI->eraseFromParent();
560 }
561
562 // If we have no remaining uses, we're done.
563 if (NewS->use_empty())
564 return;
565
566 // Otherwise, we need to create an aggregate.
567 Value *Aggr = PoisonValue::get(T: NewS->getType());
568 for (auto [Idx, Arg] : llvm::enumerate(First&: Args))
569 Aggr = Builder.CreateInsertValue(Agg: Aggr, Val: Arg, Idxs: Idx);
570
571 NewS->replaceAllUsesWith(V: Aggr);
572}
573
574void coro::BaseCloner::replaceCoroSuspends() {
575 Value *SuspendResult;
576
577 switch (Shape.ABI) {
578 // In switch lowering, replace coro.suspend with the appropriate value
579 // for the type of function we're extracting.
580 // Replacing coro.suspend with (0) will result in control flow proceeding to
581 // a resume label associated with a suspend point, replacing it with (1) will
582 // result in control flow proceeding to a cleanup label associated with this
583 // suspend point.
584 case coro::ABI::Switch:
585 SuspendResult = Builder.getInt8(C: isSwitchDestroyFunction() ? 1 : 0);
586 break;
587
588 // In async lowering there are no uses of the result.
589 case coro::ABI::Async:
590 return;
591
592 // In returned-continuation lowering, the arguments from earlier
593 // continuations are theoretically arbitrary, and they should have been
594 // spilled.
595 case coro::ABI::RetconOnce:
596 case coro::ABI::Retcon:
597 return;
598 }
599
600 for (AnyCoroSuspendInst *CS : Shape.CoroSuspends) {
601 // The active suspend was handled earlier.
602 if (CS == ActiveSuspend)
603 continue;
604
605 auto *MappedCS = cast<AnyCoroSuspendInst>(Val&: VMap[CS]);
606 MappedCS->replaceAllUsesWith(V: SuspendResult);
607 MappedCS->eraseFromParent();
608 }
609}
610
611void coro::BaseCloner::replaceCoroEnds() {
612 for (AnyCoroEndInst *CE : Shape.CoroEnds) {
613 // We use a null call graph because there's no call graph node for
614 // the cloned function yet. We'll just be rebuilding that later.
615 auto *NewCE = cast<AnyCoroEndInst>(Val&: VMap[CE]);
616 replaceCoroEnd(End: NewCE, Shape, FramePtr: NewFramePtr, /*in ramp*/ InRamp: false, CG: nullptr);
617 }
618}
619
620void coro::BaseCloner::replaceCoroIsInRamp() {
621 auto &Ctx = OrigF.getContext();
622 for (auto *II : Shape.CoroIsInRampInsts) {
623 auto *NewII = cast<CoroIsInRampInst>(Val&: VMap[II]);
624 NewII->replaceAllUsesWith(V: ConstantInt::getFalse(Context&: Ctx));
625 NewII->eraseFromParent();
626 }
627}
628
629static void replaceSwiftErrorOps(Function &F, coro::Shape &Shape,
630 ValueToValueMapTy *VMap) {
631 if (Shape.ABI == coro::ABI::Async && Shape.CoroSuspends.empty())
632 return;
633 Value *CachedSlot = nullptr;
634 auto getSwiftErrorSlot = [&](Type *ValueTy) -> Value * {
635 if (CachedSlot)
636 return CachedSlot;
637
638 // Check if the function has a swifterror argument.
639 for (auto &Arg : F.args()) {
640 if (Arg.isSwiftError()) {
641 CachedSlot = &Arg;
642 return &Arg;
643 }
644 }
645
646 // Create a swifterror alloca.
647 IRBuilder<> Builder(&F.getEntryBlock(),
648 F.getEntryBlock().getFirstNonPHIOrDbg());
649 auto Alloca = Builder.CreateAlloca(Ty: ValueTy);
650 Alloca->setSwiftError(true);
651
652 CachedSlot = Alloca;
653 return Alloca;
654 };
655
656 for (CallInst *Op : Shape.SwiftErrorOps) {
657 auto MappedOp = VMap ? cast<CallInst>(Val&: (*VMap)[Op]) : Op;
658 IRBuilder<> Builder(MappedOp);
659
660 // If there are no arguments, this is a 'get' operation.
661 Value *MappedResult;
662 if (Op->arg_empty()) {
663 auto ValueTy = Op->getType();
664 auto Slot = getSwiftErrorSlot(ValueTy);
665 MappedResult = Builder.CreateLoad(Ty: ValueTy, Ptr: Slot);
666 } else {
667 assert(Op->arg_size() == 1);
668 auto Value = MappedOp->getArgOperand(i: 0);
669 auto ValueTy = Value->getType();
670 auto Slot = getSwiftErrorSlot(ValueTy);
671 Builder.CreateStore(Val: Value, Ptr: Slot);
672 MappedResult = Slot;
673 }
674
675 MappedOp->replaceAllUsesWith(V: MappedResult);
676 MappedOp->eraseFromParent();
677 }
678
679 // If we're updating the original function, we've invalidated SwiftErrorOps.
680 if (VMap == nullptr) {
681 Shape.SwiftErrorOps.clear();
682 }
683}
684
685/// Returns all debug records in F.
686static SmallVector<DbgVariableRecord *>
687collectDbgVariableRecords(Function &F) {
688 SmallVector<DbgVariableRecord *> DbgVariableRecords;
689 for (auto &I : instructions(F)) {
690 for (DbgVariableRecord &DVR : filterDbgVars(R: I.getDbgRecordRange()))
691 DbgVariableRecords.push_back(Elt: &DVR);
692 }
693 return DbgVariableRecords;
694}
695
696void coro::BaseCloner::replaceSwiftErrorOps() {
697 ::replaceSwiftErrorOps(F&: *NewF, Shape, VMap: &VMap);
698}
699
700void coro::BaseCloner::salvageDebugInfo() {
701 auto DbgVariableRecords = collectDbgVariableRecords(F&: *NewF);
702 SmallDenseMap<Argument *, AllocaInst *, 4> ArgToAllocaMap;
703
704 // Only 64-bit ABIs have a register we can refer to with the entry value.
705 bool UseEntryValue = OrigF.getParent()->getTargetTriple().isArch64Bit();
706 for (DbgVariableRecord *DVR : DbgVariableRecords)
707 coro::salvageDebugInfo(ArgToAllocaMap, DVR&: *DVR, UseEntryValue);
708
709 // Remove all salvaged dbg.declare intrinsics that became
710 // either unreachable or stale due to the CoroSplit transformation.
711 DominatorTree DomTree(*NewF);
712 auto IsUnreachableBlock = [&](BasicBlock *BB) {
713 return !isPotentiallyReachable(From: &NewF->getEntryBlock(), To: BB, ExclusionSet: nullptr,
714 DT: &DomTree);
715 };
716 auto RemoveOne = [&](DbgVariableRecord *DVI) {
717 if (IsUnreachableBlock(DVI->getParent()))
718 DVI->eraseFromParent();
719 else if (isa_and_nonnull<AllocaInst>(Val: DVI->getVariableLocationOp(OpIdx: 0))) {
720 // Count all non-debuginfo uses in reachable blocks.
721 unsigned Uses = 0;
722 for (auto *User : DVI->getVariableLocationOp(OpIdx: 0)->users())
723 if (auto *I = dyn_cast<Instruction>(Val: User))
724 if (!isa<AllocaInst>(Val: I) && !IsUnreachableBlock(I->getParent()))
725 ++Uses;
726 if (!Uses)
727 DVI->eraseFromParent();
728 }
729 };
730 for_each(Range&: DbgVariableRecords, F: RemoveOne);
731}
732
733void coro::BaseCloner::replaceEntryBlock() {
734 // In the original function, the AllocaSpillBlock is a block immediately
735 // following the allocation of the frame object which defines GEPs for
736 // all the allocas that have been moved into the frame, and it ends by
737 // branching to the original beginning of the coroutine. Make this
738 // the entry block of the cloned function.
739 auto *Entry = cast<BasicBlock>(Val&: VMap[Shape.AllocaSpillBlock]);
740 auto *OldEntry = &NewF->getEntryBlock();
741 Entry->setName("entry" + Suffix);
742 Entry->moveBefore(MovePos: OldEntry);
743 Entry->getTerminator()->eraseFromParent();
744
745 // Clear all predecessors of the new entry block. There should be
746 // exactly one predecessor, which we created when splitting out
747 // AllocaSpillBlock to begin with.
748 assert(Entry->hasOneUse());
749 auto BranchToEntry = cast<UncondBrInst>(Val: Entry->user_back());
750 Builder.SetInsertPoint(BranchToEntry);
751 Builder.CreateUnreachable();
752 BranchToEntry->eraseFromParent();
753
754 // Branch from the entry to the appropriate place.
755 Builder.SetInsertPoint(Entry);
756 switch (Shape.ABI) {
757 case coro::ABI::Switch: {
758 // In switch-lowering, we built a resume-entry block in the original
759 // function. Make the entry block branch to this.
760 auto *SwitchBB =
761 cast<BasicBlock>(Val&: VMap[Shape.SwitchLowering.ResumeEntryBlock]);
762 Builder.CreateBr(Dest: SwitchBB);
763 SwitchBB->moveAfter(MovePos: Entry);
764 break;
765 }
766 case coro::ABI::Async:
767 case coro::ABI::Retcon:
768 case coro::ABI::RetconOnce: {
769 // In continuation ABIs, we want to branch to immediately after the
770 // active suspend point. Earlier phases will have put the suspend in its
771 // own basic block, so just thread our jump directly to its successor.
772 assert((Shape.ABI == coro::ABI::Async &&
773 isa<CoroSuspendAsyncInst>(ActiveSuspend)) ||
774 ((Shape.ABI == coro::ABI::Retcon ||
775 Shape.ABI == coro::ABI::RetconOnce) &&
776 isa<CoroSuspendRetconInst>(ActiveSuspend)));
777 auto *MappedCS = cast<AnyCoroSuspendInst>(Val&: VMap[ActiveSuspend]);
778 auto Branch = cast<UncondBrInst>(Val: MappedCS->getNextNode());
779 Builder.CreateBr(Dest: Branch->getSuccessor(i: 0));
780 break;
781 }
782 }
783
784 // Any static alloca that's still being used but not reachable from the new
785 // entry needs to be moved to the new entry.
786 Function *F = OldEntry->getParent();
787 DominatorTree DT{*F};
788 for (Instruction &I : llvm::make_early_inc_range(Range: instructions(F))) {
789 auto *Alloca = dyn_cast<AllocaInst>(Val: &I);
790 if (!Alloca || I.use_empty())
791 continue;
792 if (DT.isReachableFromEntry(A: I.getParent()) ||
793 !isa<ConstantInt>(Val: Alloca->getArraySize()))
794 continue;
795 I.moveBefore(BB&: *Entry, I: Entry->getFirstInsertionPt());
796 }
797}
798
799/// Derive the value of the new frame pointer.
800Value *coro::BaseCloner::deriveNewFramePointer() {
801 // Builder should be inserting to the front of the new entry block.
802
803 switch (Shape.ABI) {
804 // In switch-lowering, the argument is the frame pointer.
805 case coro::ABI::Switch:
806 return &*NewF->arg_begin();
807 // In async-lowering, one of the arguments is an async context as determined
808 // by the `llvm.coro.id.async` intrinsic. We can retrieve the async context of
809 // the resume function from the async context projection function associated
810 // with the active suspend. The frame is located as a tail to the async
811 // context header.
812 case coro::ABI::Async: {
813 auto *ActiveAsyncSuspend = cast<CoroSuspendAsyncInst>(Val: ActiveSuspend);
814 auto ContextIdx = ActiveAsyncSuspend->getStorageArgumentIndex() & 0xff;
815 auto *CalleeContext = NewF->getArg(i: ContextIdx);
816 auto *ProjectionFunc =
817 ActiveAsyncSuspend->getAsyncContextProjectionFunction();
818 auto DbgLoc =
819 cast<CoroSuspendAsyncInst>(Val&: VMap[ActiveSuspend])->getDebugLoc();
820 // Calling i8* (i8*)
821 auto *CallerContext = Builder.CreateCall(FTy: ProjectionFunc->getFunctionType(),
822 Callee: ProjectionFunc, Args: CalleeContext);
823 CallerContext->setCallingConv(ProjectionFunc->getCallingConv());
824 CallerContext->setDebugLoc(DbgLoc);
825 // The frame is located after the async_context header.
826 auto &Context = Builder.getContext();
827 auto *FramePtrAddr = Builder.CreateInBoundsPtrAdd(
828 Ptr: CallerContext,
829 Offset: ConstantInt::get(Ty: Type::getInt64Ty(C&: Context),
830 V: Shape.AsyncLowering.FrameOffset),
831 Name: "async.ctx.frameptr");
832 // Inline the projection function.
833 InlineFunctionInfo InlineInfo;
834 auto InlineRes = InlineFunction(CB&: *CallerContext, IFI&: InlineInfo);
835 assert(InlineRes.isSuccess());
836 (void)InlineRes;
837 return FramePtrAddr;
838 }
839 // In continuation-lowering, the argument is the opaque storage.
840 case coro::ABI::Retcon:
841 case coro::ABI::RetconOnce: {
842 Argument *NewStorage = &*NewF->arg_begin();
843 auto FramePtrTy = PointerType::getUnqual(C&: Shape.FramePtr->getContext());
844
845 // If the storage is inline, just bitcast to the storage to the frame type.
846 if (Shape.RetconLowering.IsFrameInlineInStorage)
847 return NewStorage;
848
849 // Otherwise, load the real frame from the opaque storage.
850 return Builder.CreateLoad(Ty: FramePtrTy, Ptr: NewStorage);
851 }
852 }
853 llvm_unreachable("bad ABI");
854}
855
856/// Adjust the scope line of the funclet to the first line number after the
857/// suspend point. This avoids a jump in the line table from the function
858/// declaration (where prologue instructions are attributed to) to the suspend
859/// point.
860/// Only adjust the scope line when the files are the same.
861/// If no candidate line number is found, fallback to the line of ActiveSuspend.
862static void updateScopeLine(Instruction *ActiveSuspend,
863 DISubprogram &SPToUpdate) {
864 if (!ActiveSuspend)
865 return;
866
867 // No subsequent instruction -> fallback to the location of ActiveSuspend.
868 if (!ActiveSuspend->getNextNode()) {
869 if (auto DL = ActiveSuspend->getDebugLoc())
870 if (SPToUpdate.getFile() == DL->getFile())
871 SPToUpdate.setScopeLine(DL->getLine());
872 return;
873 }
874
875 BasicBlock::iterator Successor = ActiveSuspend->getNextNode()->getIterator();
876 // Corosplit splits the BB around ActiveSuspend, so the meaningful
877 // instructions are not in the same BB.
878 // FIXME: remove this hardcoded number of tries.
879 for (unsigned Repeat = 0; Repeat < 2; Repeat++) {
880 auto *Branch = dyn_cast_or_null<UncondBrInst>(Val&: Successor);
881 if (!Branch)
882 break;
883 Successor = Branch->getSuccessor()->getFirstNonPHIOrDbg();
884 }
885
886 // Find the first successor of ActiveSuspend with a non-zero line location.
887 // If that matches the file of ActiveSuspend, use it.
888 BasicBlock *PBB = Successor->getParent();
889 for (; Successor != PBB->end(); Successor = std::next(x: Successor)) {
890 Successor = skipDebugIntrinsics(It: Successor);
891 auto DL = Successor->getDebugLoc();
892 if (!DL || DL.getLine() == 0)
893 continue;
894
895 if (SPToUpdate.getFile() == DL->getFile()) {
896 SPToUpdate.setScopeLine(DL.getLine());
897 return;
898 }
899
900 break;
901 }
902
903 // If the search above failed, fallback to the location of ActiveSuspend.
904 if (auto DL = ActiveSuspend->getDebugLoc())
905 if (SPToUpdate.getFile() == DL->getFile())
906 SPToUpdate.setScopeLine(DL->getLine());
907}
908
909static void addFramePointerAttrs(AttributeList &Attrs, LLVMContext &Context,
910 unsigned ParamIndex, uint64_t Size,
911 Align Alignment, bool NoAlias) {
912 AttrBuilder ParamAttrs(Context);
913 ParamAttrs.addAttribute(Val: Attribute::NonNull);
914 ParamAttrs.addAttribute(Val: Attribute::NoUndef);
915
916 if (NoAlias)
917 ParamAttrs.addAttribute(Val: Attribute::NoAlias);
918
919 ParamAttrs.addAlignmentAttr(Align: Alignment);
920 ParamAttrs.addDereferenceableAttr(Bytes: Size);
921 Attrs = Attrs.addParamAttributes(C&: Context, ArgNo: ParamIndex, B: ParamAttrs);
922}
923
924static void addAsyncContextAttrs(AttributeList &Attrs, LLVMContext &Context,
925 unsigned ParamIndex) {
926 AttrBuilder ParamAttrs(Context);
927 ParamAttrs.addAttribute(Val: Attribute::SwiftAsync);
928 Attrs = Attrs.addParamAttributes(C&: Context, ArgNo: ParamIndex, B: ParamAttrs);
929}
930
931static void addSwiftSelfAttrs(AttributeList &Attrs, LLVMContext &Context,
932 unsigned ParamIndex) {
933 AttrBuilder ParamAttrs(Context);
934 ParamAttrs.addAttribute(Val: Attribute::SwiftSelf);
935 Attrs = Attrs.addParamAttributes(C&: Context, ArgNo: ParamIndex, B: ParamAttrs);
936}
937
938/// Clone the body of the original function into a resume function of
939/// some sort.
940void coro::BaseCloner::create() {
941 assert(NewF);
942
943 // Replace all args with dummy instructions. If an argument is the old frame
944 // pointer, the dummy will be replaced by the new frame pointer once it is
945 // computed below. Uses of all other arguments should have already been
946 // rewritten by buildCoroutineFrame() to use loads/stores on the coroutine
947 // frame.
948 SmallVector<Instruction *> DummyArgs;
949 for (Argument &A : OrigF.args()) {
950 DummyArgs.push_back(Elt: new FreezeInst(PoisonValue::get(T: A.getType())));
951 VMap[&A] = DummyArgs.back();
952 }
953
954 SmallVector<ReturnInst *, 4> Returns;
955
956 // Ignore attempts to change certain attributes of the function.
957 // TODO: maybe there should be a way to suppress this during cloning?
958 auto savedVisibility = NewF->getVisibility();
959 auto savedUnnamedAddr = NewF->getUnnamedAddr();
960 auto savedDLLStorageClass = NewF->getDLLStorageClass();
961
962 // NewF's linkage (which CloneFunctionInto does *not* change) might not
963 // be compatible with the visibility of OrigF (which it *does* change),
964 // so protect against that.
965 auto savedLinkage = NewF->getLinkage();
966 NewF->setLinkage(llvm::GlobalValue::ExternalLinkage);
967
968 CloneFunctionInto(NewFunc: NewF, OldFunc: &OrigF, VMap,
969 Changes: CloneFunctionChangeType::LocalChangesOnly, Returns);
970
971 auto &Context = NewF->getContext();
972
973 if (DISubprogram *SP = NewF->getSubprogram()) {
974 assert(SP != OrigF.getSubprogram() && SP->isDistinct());
975 updateScopeLine(ActiveSuspend, SPToUpdate&: *SP);
976
977 // Update the linkage name and the function name to reflect the modified
978 // name.
979 MDString *NewLinkageName = MDString::get(Context, Str: NewF->getName());
980 SP->replaceLinkageName(LN: NewLinkageName);
981 if (DISubprogram *Decl = SP->getDeclaration()) {
982 TempDISubprogram NewDecl = Decl->clone();
983 NewDecl->replaceLinkageName(LN: NewLinkageName);
984 SP->replaceDeclaration(Decl: MDNode::replaceWithUniqued(N: std::move(NewDecl)));
985 }
986 }
987
988 NewF->setLinkage(savedLinkage);
989 NewF->setVisibility(savedVisibility);
990 NewF->setUnnamedAddr(savedUnnamedAddr);
991 NewF->setDLLStorageClass(savedDLLStorageClass);
992 // The function sanitizer metadata needs to match the signature of the
993 // function it is being attached to. However this does not hold for split
994 // functions here. Thus remove the metadata for split functions.
995 if (Shape.ABI == coro::ABI::Switch &&
996 NewF->hasMetadata(KindID: LLVMContext::MD_func_sanitize))
997 NewF->eraseMetadata(KindID: LLVMContext::MD_func_sanitize);
998
999 // Replace the attributes of the new function:
1000 auto OrigAttrs = NewF->getAttributes();
1001 auto NewAttrs = AttributeList();
1002
1003 switch (Shape.ABI) {
1004 case coro::ABI::Switch:
1005 // Bootstrap attributes by copying function attributes from the
1006 // original function. This should include optimization settings and so on.
1007 NewAttrs = NewAttrs.addFnAttributes(
1008 C&: Context, B: AttrBuilder(Context, OrigAttrs.getFnAttrs()));
1009
1010 addFramePointerAttrs(Attrs&: NewAttrs, Context, ParamIndex: 0, Size: Shape.FrameSize,
1011 Alignment: Shape.FrameAlign, /*NoAlias=*/false);
1012 break;
1013 case coro::ABI::Async: {
1014 auto *ActiveAsyncSuspend = cast<CoroSuspendAsyncInst>(Val: ActiveSuspend);
1015 if (OrigF.hasParamAttribute(ArgNo: Shape.AsyncLowering.ContextArgNo,
1016 Kind: Attribute::SwiftAsync)) {
1017 uint32_t ArgAttributeIndices =
1018 ActiveAsyncSuspend->getStorageArgumentIndex();
1019 auto ContextArgIndex = ArgAttributeIndices & 0xff;
1020 addAsyncContextAttrs(Attrs&: NewAttrs, Context, ParamIndex: ContextArgIndex);
1021
1022 // `swiftasync` must preceed `swiftself` so 0 is not a valid index for
1023 // `swiftself`.
1024 auto SwiftSelfIndex = ArgAttributeIndices >> 8;
1025 if (SwiftSelfIndex)
1026 addSwiftSelfAttrs(Attrs&: NewAttrs, Context, ParamIndex: SwiftSelfIndex);
1027 }
1028
1029 // Transfer the original function's attributes.
1030 auto FnAttrs = OrigF.getAttributes().getFnAttrs();
1031 NewAttrs = NewAttrs.addFnAttributes(C&: Context, B: AttrBuilder(Context, FnAttrs));
1032 break;
1033 }
1034 case coro::ABI::Retcon:
1035 case coro::ABI::RetconOnce:
1036 // If we have a continuation prototype, just use its attributes,
1037 // full-stop.
1038 NewAttrs = Shape.RetconLowering.ResumePrototype->getAttributes();
1039
1040 /// FIXME: Is it really good to add the NoAlias attribute?
1041 addFramePointerAttrs(Attrs&: NewAttrs, Context, ParamIndex: 0,
1042 Size: Shape.getRetconCoroId()->getStorageSize(),
1043 Alignment: Shape.getRetconCoroId()->getStorageAlignment(),
1044 /*NoAlias=*/true);
1045
1046 break;
1047 }
1048
1049 switch (Shape.ABI) {
1050 // In these ABIs, the cloned functions always return 'void', and the
1051 // existing return sites are meaningless. Note that for unique
1052 // continuations, this includes the returns associated with suspends;
1053 // this is fine because we can't suspend twice.
1054 case coro::ABI::Switch:
1055 case coro::ABI::RetconOnce:
1056 // Remove old returns.
1057 for (ReturnInst *Return : Returns)
1058 changeToUnreachable(I: Return);
1059 break;
1060
1061 // With multi-suspend continuations, we'll already have eliminated the
1062 // original returns and inserted returns before all the suspend points,
1063 // so we want to leave any returns in place.
1064 case coro::ABI::Retcon:
1065 break;
1066 // Async lowering will insert musttail call functions at all suspend points
1067 // followed by a return.
1068 // Don't change returns to unreachable because that will trip up the verifier.
1069 // These returns should be unreachable from the clone.
1070 case coro::ABI::Async:
1071 break;
1072 }
1073
1074 NewF->setAttributes(NewAttrs);
1075 NewF->setCallingConv(Shape.getResumeFunctionCC());
1076
1077 // Set up the new entry block.
1078 replaceEntryBlock();
1079
1080 // Turn symmetric transfers into musttail calls.
1081 for (CallInst *ResumeCall : Shape.SymmetricTransfers) {
1082 ResumeCall = cast<CallInst>(Val&: VMap[ResumeCall]);
1083 if (TTI.supportsTailCallFor(CB: ResumeCall)) {
1084 // FIXME: Could we support symmetric transfer effectively without
1085 // musttail?
1086 ResumeCall->setTailCallKind(CallInst::TCK_MustTail);
1087 }
1088
1089 // Put a 'ret void' after the call, and split any remaining instructions to
1090 // an unreachable block.
1091 BasicBlock *BB = ResumeCall->getParent();
1092 BB->splitBasicBlock(I: ResumeCall->getNextNode());
1093 Builder.SetInsertPoint(BB->getTerminator());
1094 Builder.CreateRetVoid();
1095 BB->getTerminator()->eraseFromParent();
1096 }
1097
1098 Builder.SetInsertPoint(&NewF->getEntryBlock().front());
1099 NewFramePtr = deriveNewFramePointer();
1100
1101 // Remap frame pointer.
1102 Value *OldFramePtr = VMap[Shape.FramePtr];
1103 NewFramePtr->takeName(V: OldFramePtr);
1104 OldFramePtr->replaceAllUsesWith(V: NewFramePtr);
1105
1106 // Remap vFrame pointer.
1107 auto *NewVFrame = Builder.CreateBitCast(
1108 V: NewFramePtr, DestTy: PointerType::getUnqual(C&: Builder.getContext()), Name: "vFrame");
1109 Value *OldVFrame = cast<Value>(Val&: VMap[Shape.CoroBegin]);
1110 if (OldVFrame != NewVFrame)
1111 OldVFrame->replaceAllUsesWith(V: NewVFrame);
1112
1113 // All uses of the arguments should have been resolved by this point,
1114 // so we can safely remove the dummy values.
1115 for (Instruction *DummyArg : DummyArgs) {
1116 DummyArg->replaceAllUsesWith(V: PoisonValue::get(T: DummyArg->getType()));
1117 DummyArg->deleteValue();
1118 }
1119
1120 switch (Shape.ABI) {
1121 case coro::ABI::Switch:
1122 // Rewrite final suspend handling as it is not done via switch (allows to
1123 // remove final case from the switch, since it is undefined behavior to
1124 // resume the coroutine suspended at the final suspend point.
1125 if (Shape.SwitchLowering.HasFinalSuspend)
1126 handleFinalSuspend();
1127 break;
1128 case coro::ABI::Async:
1129 case coro::ABI::Retcon:
1130 case coro::ABI::RetconOnce:
1131 // Replace uses of the active suspend with the corresponding
1132 // continuation-function arguments.
1133 assert(ActiveSuspend != nullptr &&
1134 "no active suspend when lowering a continuation-style coroutine");
1135 replaceRetconOrAsyncSuspendUses();
1136 break;
1137 }
1138
1139 // Handle suspends.
1140 replaceCoroSuspends();
1141
1142 // Handle swifterror.
1143 replaceSwiftErrorOps();
1144
1145 // Remove coro.end intrinsics.
1146 replaceCoroEnds();
1147
1148 replaceCoroIsInRamp();
1149
1150 // Salvage debug info that points into the coroutine frame.
1151 salvageDebugInfo();
1152}
1153
1154void coro::SwitchCloner::create() {
1155 // Create a new function matching the original type
1156 NewF = createCloneDeclaration(OrigF, Shape, Suffix, InsertBefore: OrigF.getParent()->end(),
1157 ActiveSuspend);
1158
1159 // Clone the function
1160 coro::BaseCloner::create();
1161
1162 // Override EntryCount for the cloned resume function with the true sum of
1163 // all suspension points profile counts.
1164 if (FKind == coro::CloneKind::SwitchResume && OrigF.hasProfileData() &&
1165 Shape.ResumeEntryCount.has_value()) {
1166 NewF->setEntryCount(Count: Shape.ResumeEntryCount.value());
1167 }
1168
1169 // Replacing coro.free with 'null' in cleanup to suppress deallocation code.
1170 if (FKind == coro::CloneKind::SwitchCleanup)
1171 elideCoroFree(FramePtr: NewFramePtr);
1172}
1173
1174static void updateAsyncFuncPointerContextSize(coro::Shape &Shape) {
1175 assert(Shape.ABI == coro::ABI::Async);
1176
1177 auto *FuncPtrStruct = cast<ConstantStruct>(
1178 Val: Shape.AsyncLowering.AsyncFuncPointer->getInitializer());
1179 auto *OrigRelativeFunOffset = FuncPtrStruct->getOperand(i_nocapture: 0);
1180 auto *OrigContextSize = FuncPtrStruct->getOperand(i_nocapture: 1);
1181 auto *NewContextSize = ConstantInt::get(Ty: OrigContextSize->getType(),
1182 V: Shape.AsyncLowering.ContextSize);
1183 auto *NewFuncPtrStruct = ConstantStruct::get(
1184 T: FuncPtrStruct->getType(), Vs: OrigRelativeFunOffset, Vs: NewContextSize);
1185
1186 Shape.AsyncLowering.AsyncFuncPointer->setInitializer(NewFuncPtrStruct);
1187}
1188
1189static void replaceFrameSizeAndAlignment(coro::Shape &Shape) {
1190 if (Shape.ABI == coro::ABI::Async)
1191 updateAsyncFuncPointerContextSize(Shape);
1192
1193 for (CoroAlignInst *CA : Shape.CoroAligns) {
1194 CA->replaceAllUsesWith(
1195 V: ConstantInt::get(Ty: CA->getType(), V: Shape.FrameAlign.value()));
1196 CA->eraseFromParent();
1197 }
1198
1199 if (Shape.CoroSizes.empty())
1200 return;
1201
1202 // In the same function all coro.sizes should have the same result type.
1203 auto *SizeIntrin = Shape.CoroSizes.back();
1204 auto *SizeConstant = ConstantInt::get(Ty: SizeIntrin->getType(),
1205 V: TypeSize::getFixed(ExactSize: Shape.FrameSize));
1206
1207 for (CoroSizeInst *CS : Shape.CoroSizes) {
1208 CS->replaceAllUsesWith(V: SizeConstant);
1209 CS->eraseFromParent();
1210 }
1211}
1212
1213static void postSplitCleanup(Function &F) {
1214 removeUnreachableBlocks(F);
1215
1216#ifndef NDEBUG
1217 // For now, we do a mandatory verification step because we don't
1218 // entirely trust this pass. Note that we don't want to add a verifier
1219 // pass to FPM below because it will also verify all the global data.
1220 if (verifyFunction(F, &errs()))
1221 report_fatal_error("Broken function");
1222#endif
1223}
1224
1225// Coroutine has no suspend points. Remove heap allocation for the coroutine
1226// frame if possible.
1227static void handleNoSuspendCoroutine(coro::Shape &Shape) {
1228 auto *CoroBegin = Shape.CoroBegin;
1229 switch (Shape.ABI) {
1230 case coro::ABI::Switch: {
1231 if (auto *AllocInst = Shape.getSwitchCoroId()->getCoroAlloc()) {
1232 coro::elideCoroFree(FramePtr: CoroBegin);
1233
1234 IRBuilder<> Builder(AllocInst);
1235 // Create an alloca for a byte array of the frame size
1236 auto *FrameTy = ArrayType::get(ElementType: Type::getInt8Ty(C&: Builder.getContext()),
1237 NumElements: Shape.FrameSize);
1238 auto *Frame = Builder.CreateAlloca(
1239 Ty: FrameTy, ArraySize: nullptr, Name: AllocInst->getFunction()->getName() + ".Frame");
1240 Frame->setAlignment(Shape.FrameAlign);
1241 AllocInst->replaceAllUsesWith(V: Builder.getFalse());
1242 AllocInst->eraseFromParent();
1243 CoroBegin->replaceAllUsesWith(V: Frame);
1244 } else {
1245 CoroBegin->replaceAllUsesWith(V: CoroBegin->getMem());
1246 }
1247
1248 break;
1249 }
1250 case coro::ABI::Async:
1251 case coro::ABI::Retcon:
1252 case coro::ABI::RetconOnce:
1253 CoroBegin->replaceAllUsesWith(V: PoisonValue::get(T: CoroBegin->getType()));
1254 break;
1255 }
1256
1257 CoroBegin->eraseFromParent();
1258 Shape.CoroBegin = nullptr;
1259}
1260
1261// SimplifySuspendPoint needs to check that there is no calls between
1262// coro_save and coro_suspend, since any of the calls may potentially resume
1263// the coroutine and if that is the case we cannot eliminate the suspend point.
1264static bool hasCallsInBlockBetween(iterator_range<BasicBlock::iterator> R) {
1265 for (Instruction &I : R) {
1266 // Assume that no intrinsic can resume the coroutine.
1267 if (isa<IntrinsicInst>(Val: I))
1268 continue;
1269
1270 if (isa<CallBase>(Val: I))
1271 return true;
1272 }
1273 return false;
1274}
1275
1276static bool hasCallsInBlocksBetween(BasicBlock *SaveBB, BasicBlock *ResDesBB) {
1277 SmallPtrSet<BasicBlock *, 8> Set;
1278 SmallVector<BasicBlock *, 8> Worklist;
1279
1280 Set.insert(Ptr: SaveBB);
1281 Worklist.push_back(Elt: ResDesBB);
1282
1283 // Accumulate all blocks between SaveBB and ResDesBB. Because CoroSaveIntr
1284 // returns a token consumed by suspend instruction, all blocks in between
1285 // will have to eventually hit SaveBB when going backwards from ResDesBB.
1286 while (!Worklist.empty()) {
1287 auto *BB = Worklist.pop_back_val();
1288 Set.insert(Ptr: BB);
1289 for (auto *Pred : predecessors(BB))
1290 if (!Set.contains(Ptr: Pred))
1291 Worklist.push_back(Elt: Pred);
1292 }
1293
1294 // SaveBB and ResDesBB are checked separately in hasCallsBetween.
1295 Set.erase(Ptr: SaveBB);
1296 Set.erase(Ptr: ResDesBB);
1297
1298 for (auto *BB : Set)
1299 if (hasCallsInBlockBetween(R: {BB->getFirstNonPHIIt(), BB->end()}))
1300 return true;
1301
1302 return false;
1303}
1304
1305static bool hasCallsBetween(Instruction *Save, Instruction *ResumeOrDestroy) {
1306 auto *SaveBB = Save->getParent();
1307 auto *ResumeOrDestroyBB = ResumeOrDestroy->getParent();
1308 BasicBlock::iterator SaveIt = Save->getIterator();
1309 BasicBlock::iterator ResumeOrDestroyIt = ResumeOrDestroy->getIterator();
1310
1311 if (SaveBB == ResumeOrDestroyBB)
1312 return hasCallsInBlockBetween(R: {std::next(x: SaveIt), ResumeOrDestroyIt});
1313
1314 // Any calls from Save to the end of the block?
1315 if (hasCallsInBlockBetween(R: {std::next(x: SaveIt), SaveBB->end()}))
1316 return true;
1317
1318 // Any calls from begging of the block up to ResumeOrDestroy?
1319 if (hasCallsInBlockBetween(
1320 R: {ResumeOrDestroyBB->getFirstNonPHIIt(), ResumeOrDestroyIt}))
1321 return true;
1322
1323 // Any calls in all of the blocks between SaveBB and ResumeOrDestroyBB?
1324 if (hasCallsInBlocksBetween(SaveBB, ResDesBB: ResumeOrDestroyBB))
1325 return true;
1326
1327 return false;
1328}
1329
1330// If a SuspendIntrin is preceded by Resume or Destroy, we can eliminate the
1331// suspend point and replace it with nornal control flow.
1332static bool simplifySuspendPoint(CoroSuspendInst *Suspend,
1333 CoroBeginInst *CoroBegin) {
1334 Instruction *Prev = Suspend->getPrevNode();
1335 if (!Prev) {
1336 auto *Pred = Suspend->getParent()->getSinglePredecessor();
1337 if (!Pred)
1338 return false;
1339 Prev = Pred->getTerminator();
1340 }
1341
1342 CallBase *CB = dyn_cast<CallBase>(Val: Prev);
1343 if (!CB)
1344 return false;
1345
1346 auto *Callee = CB->getCalledOperand()->stripPointerCasts();
1347
1348 // See if the callsite is for resumption or destruction of the coroutine.
1349 auto *SubFn = dyn_cast<CoroSubFnInst>(Val: Callee);
1350 if (!SubFn)
1351 return false;
1352
1353 // Does not refer to the current coroutine, we cannot do anything with it.
1354 if (SubFn->getFrame() != CoroBegin)
1355 return false;
1356
1357 // See if the transformation is safe. Specifically, see if there are any
1358 // calls in between Save and CallInstr. They can potenitally resume the
1359 // coroutine rendering this optimization unsafe.
1360 auto *Save = Suspend->getCoroSave();
1361 if (hasCallsBetween(Save, ResumeOrDestroy: CB))
1362 return false;
1363
1364 // Replace llvm.coro.suspend with the value that results in resumption over
1365 // the resume or cleanup path.
1366 Suspend->replaceAllUsesWith(V: SubFn->getRawIndex());
1367 Suspend->eraseFromParent();
1368 Save->eraseFromParent();
1369
1370 // No longer need a call to coro.resume or coro.destroy.
1371 if (auto *Invoke = dyn_cast<InvokeInst>(Val: CB)) {
1372 UncondBrInst::Create(Target: Invoke->getNormalDest(), InsertBefore: Invoke->getIterator());
1373 }
1374
1375 // Grab the CalledValue from CB before erasing the CallInstr.
1376 auto *CalledValue = CB->getCalledOperand();
1377 CB->eraseFromParent();
1378
1379 // If no more users remove it. Usually it is a bitcast of SubFn.
1380 if (CalledValue != SubFn && CalledValue->user_empty())
1381 if (auto *I = dyn_cast<Instruction>(Val: CalledValue))
1382 I->eraseFromParent();
1383
1384 // Now we are good to remove SubFn.
1385 if (SubFn->user_empty())
1386 SubFn->eraseFromParent();
1387
1388 return true;
1389}
1390
1391// Remove suspend points that are simplified.
1392static void simplifySuspendPoints(coro::Shape &Shape) {
1393 // Currently, the only simplification we do is switch-lowering-specific.
1394 if (Shape.ABI != coro::ABI::Switch)
1395 return;
1396
1397 auto &S = Shape.CoroSuspends;
1398 size_t I = 0, N = S.size();
1399 if (N == 0)
1400 return;
1401
1402 size_t ChangedFinalIndex = std::numeric_limits<size_t>::max();
1403 while (true) {
1404 auto SI = cast<CoroSuspendInst>(Val: S[I]);
1405 // Leave final.suspend to handleFinalSuspend since it is undefined behavior
1406 // to resume a coroutine suspended at the final suspend point.
1407 if (!SI->isFinal() && simplifySuspendPoint(Suspend: SI, CoroBegin: Shape.CoroBegin)) {
1408 if (--N == I)
1409 break;
1410
1411 std::swap(a&: S[I], b&: S[N]);
1412
1413 if (cast<CoroSuspendInst>(Val: S[I])->isFinal()) {
1414 assert(Shape.SwitchLowering.HasFinalSuspend);
1415 ChangedFinalIndex = I;
1416 }
1417
1418 continue;
1419 }
1420 if (++I == N)
1421 break;
1422 }
1423 S.resize(N);
1424
1425 // Maintain final.suspend in case final suspend was swapped.
1426 // Due to we requrie the final suspend to be the last element of CoroSuspends.
1427 if (ChangedFinalIndex < N) {
1428 assert(cast<CoroSuspendInst>(S[ChangedFinalIndex])->isFinal());
1429 std::swap(a&: S[ChangedFinalIndex], b&: S.back());
1430 }
1431}
1432
1433namespace {
1434
1435struct SwitchCoroutineSplitter {
1436 static void split(Function &F, coro::Shape &Shape,
1437 SmallVectorImpl<Function *> &Clones,
1438 TargetTransformInfo &TTI) {
1439 assert(Shape.ABI == coro::ABI::Switch);
1440
1441 // Create a resume clone by cloning the body of the original function,
1442 // setting new entry block and replacing coro.suspend an appropriate value
1443 // to force resume or cleanup pass for every suspend point.
1444 createResumeEntryBlock(F, Shape);
1445 auto *ResumeClone = coro::SwitchCloner::createClone(
1446 OrigF&: F, Suffix: ".resume", Shape, FKind: coro::CloneKind::SwitchResume, TTI);
1447 auto *DestroyClone = coro::SwitchCloner::createClone(
1448 OrigF&: F, Suffix: ".destroy", Shape, FKind: coro::CloneKind::SwitchUnwind, TTI);
1449 auto *CleanupClone = coro::SwitchCloner::createClone(
1450 OrigF&: F, Suffix: ".cleanup", Shape, FKind: coro::CloneKind::SwitchCleanup, TTI);
1451
1452 if (Shape.SwitchLowering.HasCoroElideNoAllocVariant)
1453 replaceSwitchResumeCoroFree(Shape, Resume&: *ResumeClone, Cleanup&: *CleanupClone);
1454
1455 postSplitCleanup(F&: *ResumeClone);
1456 postSplitCleanup(F&: *DestroyClone);
1457 postSplitCleanup(F&: *CleanupClone);
1458
1459 // Store addresses resume/destroy/cleanup functions in the coroutine frame.
1460 updateCoroFrame(Shape, ResumeFn: ResumeClone, DestroyFn: DestroyClone, CleanupFn: CleanupClone);
1461
1462 assert(Clones.empty());
1463 Clones.push_back(Elt: ResumeClone);
1464 Clones.push_back(Elt: DestroyClone);
1465 Clones.push_back(Elt: CleanupClone);
1466
1467 // Create a constant array referring to resume/destroy/clone functions
1468 // pointed by the last argument of @llvm.coro.info, so that CoroElide pass
1469 // can determined correct function to call.
1470 setCoroInfo(F, Shape, Fns: Clones);
1471 }
1472
1473 // Create a variant of ramp function that does not perform heap allocation
1474 // for a switch ABI coroutine.
1475 //
1476 // The newly split `.noalloc` ramp function has the following differences:
1477 // - Has one additional frame pointer parameter in lieu of dynamic
1478 // allocation.
1479 // - Suppressed allocations by replacing coro.alloc and coro.free.
1480 static Function *createNoAllocVariant(Function &F, coro::Shape &Shape,
1481 SmallVectorImpl<Function *> &Clones) {
1482 assert(Shape.ABI == coro::ABI::Switch);
1483 auto *OrigFnTy = F.getFunctionType();
1484 auto OldParams = OrigFnTy->params();
1485
1486 SmallVector<Type *> NewParams;
1487 NewParams.reserve(N: OldParams.size() + 1);
1488 NewParams.append(in_start: OldParams.begin(), in_end: OldParams.end());
1489 NewParams.push_back(Elt: PointerType::getUnqual(C&: Shape.FramePtr->getContext()));
1490
1491 auto *NewFnTy = FunctionType::get(Result: OrigFnTy->getReturnType(), Params: NewParams,
1492 isVarArg: OrigFnTy->isVarArg());
1493 Function *NoAllocF = Function::Create(
1494 Ty: NewFnTy, Linkage: F.getLinkage(), AddrSpace: F.getAddressSpace(), N: F.getName() + ".noalloc");
1495
1496 ValueToValueMapTy VMap;
1497 unsigned int Idx = 0;
1498 for (const auto &I : F.args()) {
1499 VMap[&I] = NoAllocF->getArg(i: Idx++);
1500 }
1501 // We just appended the frame pointer as the last argument of the new
1502 // function.
1503 auto FrameIdx = NoAllocF->arg_size() - 1;
1504 SmallVector<ReturnInst *, 4> Returns;
1505 CloneFunctionInto(NewFunc: NoAllocF, OldFunc: &F, VMap,
1506 Changes: CloneFunctionChangeType::LocalChangesOnly, Returns);
1507
1508 if (Shape.CoroBegin) {
1509 auto *NewCoroBegin =
1510 cast_if_present<CoroBeginInst>(Val&: VMap[Shape.CoroBegin]);
1511 coro::elideCoroFree(FramePtr: NewCoroBegin);
1512 coro::suppressCoroAllocs(CoroId: cast<CoroIdInst>(Val: NewCoroBegin->getId()));
1513 NewCoroBegin->replaceAllUsesWith(V: NoAllocF->getArg(i: FrameIdx));
1514 NewCoroBegin->eraseFromParent();
1515 }
1516
1517 Module *M = F.getParent();
1518 M->getFunctionList().insert(where: M->end(), New: NoAllocF);
1519
1520 removeUnreachableBlocks(F&: *NoAllocF);
1521 auto NewAttrs = NoAllocF->getAttributes();
1522 // When we elide allocation, we read these attributes to determine the
1523 // frame size and alignment.
1524 addFramePointerAttrs(Attrs&: NewAttrs, Context&: NoAllocF->getContext(), ParamIndex: FrameIdx,
1525 Size: Shape.FrameSize, Alignment: Shape.FrameAlign,
1526 /*NoAlias=*/false);
1527
1528 NoAllocF->setAttributes(NewAttrs);
1529
1530 Clones.push_back(Elt: NoAllocF);
1531 // Reset the original function's coro info, make the new noalloc variant
1532 // connected to the original ramp function.
1533 setCoroInfo(F, Shape, Fns: Clones);
1534 // After copying, set the linkage to internal linkage. Original function
1535 // may have different linkage, but optimization dependent on this function
1536 // generally relies on LTO.
1537 NoAllocF->setLinkage(llvm::GlobalValue::InternalLinkage);
1538 return NoAllocF;
1539 }
1540
1541private:
1542 // Create an entry block for a resume function with a switch that will jump to
1543 // suspend points.
1544 static void createResumeEntryBlock(Function &F, coro::Shape &Shape) {
1545 LLVMContext &C = F.getContext();
1546
1547 DIBuilder DBuilder(*F.getParent(), /*AllowUnresolved*/ false);
1548 DISubprogram *DIS = F.getSubprogram();
1549 // If there is no DISubprogram for F, it implies the function is compiled
1550 // without debug info. So we also don't generate debug info for the
1551 // suspension points.
1552 bool AddDebugLabels = DIS && DIS->getUnit() &&
1553 (DIS->getUnit()->getEmissionKind() ==
1554 DICompileUnit::DebugEmissionKind::FullDebug);
1555
1556 // resume.entry:
1557 // %index.addr = getelementptr inbounds %f.Frame, %f.Frame* %FramePtr, i32
1558 // 0, i32 2 % index = load i32, i32* %index.addr switch i32 %index, label
1559 // %unreachable [
1560 // i32 0, label %resume.0
1561 // i32 1, label %resume.1
1562 // ...
1563 // ]
1564
1565 auto *NewEntry = BasicBlock::Create(Context&: C, Name: "resume.entry", Parent: &F);
1566 auto *UnreachBB = BasicBlock::Create(Context&: C, Name: "unreachable", Parent: &F);
1567
1568 IRBuilder<> Builder(NewEntry);
1569 auto *FramePtr = Shape.FramePtr;
1570 Value *GepIndex = createSwitchIndexPtr(Shape, Builder, FramePtr);
1571 auto *Index = Builder.CreateLoad(Ty: Shape.getIndexType(), Ptr: GepIndex, Name: "index");
1572 auto *Switch =
1573 Builder.CreateSwitch(V: Index, Dest: UnreachBB, NumCases: Shape.CoroSuspends.size());
1574 Shape.SwitchLowering.ResumeSwitch = Switch;
1575
1576 // Split all coro.suspend calls
1577 size_t SuspendIndex = 0;
1578 SmallVector<uint64_t, 8> SwitchWeights64;
1579 // Default destination (unreachable) has weight 0
1580 SwitchWeights64.push_back(Elt: 0);
1581
1582 for (auto *AnyS : Shape.CoroSuspends) {
1583 auto *S = cast<CoroSuspendInst>(Val: AnyS);
1584 ConstantInt *IndexVal = Shape.getIndex(Value: SuspendIndex);
1585
1586 // Replace CoroSave with a store to Index:
1587 // %index.addr = getelementptr %f.frame... (index field number)
1588 // store i32 %IndexVal, i32* %index.addr1
1589 auto *Save = S->getCoroSave();
1590 Builder.SetInsertPoint(Save);
1591 if (S->isFinal()) {
1592 // The coroutine should be marked done if it reaches the final suspend
1593 // point.
1594 markCoroutineAsDone(Builder, Shape, FramePtr);
1595 } else {
1596 Value *GepIndex = createSwitchIndexPtr(Shape, Builder, FramePtr);
1597 Builder.CreateStore(Val: IndexVal, Ptr: GepIndex);
1598 }
1599
1600 Save->replaceAllUsesWith(V: ConstantTokenNone::get(Context&: C));
1601 Save->eraseFromParent();
1602
1603 // Split block before and after coro.suspend and add a jump from an entry
1604 // switch:
1605 //
1606 // whateverBB:
1607 // whatever
1608 // %0 = call i8 @llvm.coro.suspend(token none, i1 false)
1609 // switch i8 %0, label %suspend[i8 0, label %resume
1610 // i8 1, label %cleanup]
1611 // becomes:
1612 //
1613 // whateverBB:
1614 // whatever
1615 // br label %resume.0.landing
1616 //
1617 // resume.0: ; <--- jump from the switch in the resume.entry
1618 // #dbg_label(...) ; <--- artificial label for debuggers
1619 // %0 = tail call i8 @llvm.coro.suspend(token none, i1 false)
1620 // br label %resume.0.landing
1621 //
1622 // resume.0.landing:
1623 // %1 = phi i8[-1, %whateverBB], [%0, %resume.0]
1624 // switch i8 % 1, label %suspend [i8 0, label %resume
1625 // i8 1, label %cleanup]
1626
1627 auto *SuspendBB = S->getParent();
1628 auto *ResumeBB =
1629 SuspendBB->splitBasicBlock(I: S, BBName: "resume." + Twine(SuspendIndex));
1630 auto *LandingBB = ResumeBB->splitBasicBlock(
1631 I: S->getNextNode(), BBName: ResumeBB->getName() + Twine(".landing"));
1632 Switch->addCase(OnVal: IndexVal, Dest: ResumeBB);
1633
1634 // Get pre-split frequency for this suspend point
1635 uint64_t Weight = 1; // Default fallback weight
1636 auto It = Shape.SuspendFreqs.find(Val: AnyS);
1637 if (It != Shape.SuspendFreqs.end()) {
1638 Weight = It->second;
1639 }
1640 SwitchWeights64.push_back(Elt: Weight);
1641
1642 cast<UncondBrInst>(Val: SuspendBB->getTerminator())->setSuccessor(LandingBB);
1643 auto *PN = PHINode::Create(Ty: Builder.getInt8Ty(), NumReservedValues: 2, NameStr: "");
1644 PN->insertBefore(InsertPos: LandingBB->begin());
1645 S->replaceAllUsesWith(V: PN);
1646 PN->addIncoming(V: Builder.getInt8(C: -1), BB: SuspendBB);
1647 PN->addIncoming(V: S, BB: ResumeBB);
1648
1649 if (AddDebugLabels) {
1650 if (DebugLoc SuspendLoc = S->getDebugLoc()) {
1651 std::string LabelName =
1652 ("__coro_resume_" + Twine(SuspendIndex)).str();
1653 // Take the "inlined at" location recursively, if present. This is
1654 // mandatory as the DILabel insertion checks that the scopes of label
1655 // and the attached location match. This is not the case when the
1656 // suspend location has been inlined due to pointing to the original
1657 // scope.
1658 DILocation *DILoc = SuspendLoc;
1659 while (DILocation *InlinedAt = DILoc->getInlinedAt())
1660 DILoc = InlinedAt;
1661
1662 DILabel *ResumeLabel =
1663 DBuilder.createLabel(Scope: DIS, Name: LabelName, File: DILoc->getFile(),
1664 LineNo: SuspendLoc.getLine(), Column: SuspendLoc.getCol(),
1665 /*IsArtificial=*/true,
1666 /*CoroSuspendIdx=*/SuspendIndex,
1667 /*AlwaysPreserve=*/false);
1668 DBuilder.insertLabel(LabelInfo: ResumeLabel, DL: DILoc, InsertPt: ResumeBB->begin());
1669 }
1670 }
1671
1672 ++SuspendIndex;
1673 }
1674
1675 if (!Shape.SuspendFreqs.empty()) {
1676 auto SwitchWeights32 = llvm::fitWeights(Weights: SwitchWeights64);
1677 MDBuilder MDB(C);
1678 Switch->setMetadata(KindID: LLVMContext::MD_prof,
1679 Node: MDB.createBranchWeights(Weights: SwitchWeights32));
1680 }
1681
1682 Builder.SetInsertPoint(UnreachBB);
1683 Builder.CreateUnreachable();
1684 DBuilder.finalize();
1685
1686 Shape.SwitchLowering.ResumeEntryBlock = NewEntry;
1687 }
1688
1689 // Store addresses of Resume/Destroy/Cleanup functions in the coroutine frame.
1690 static void updateCoroFrame(coro::Shape &Shape, Function *ResumeFn,
1691 Function *DestroyFn, Function *CleanupFn) {
1692 IRBuilder<> Builder(&*Shape.getInsertPtAfterFramePtr());
1693 LLVMContext &C = ResumeFn->getContext();
1694
1695 // Resume function pointer
1696 Value *ResumeAddr = Shape.FramePtr;
1697 Builder.CreateStore(Val: ResumeFn, Ptr: ResumeAddr);
1698
1699 Value *DestroyOrCleanupFn = DestroyFn;
1700
1701 CoroIdInst *CoroId = Shape.getSwitchCoroId();
1702 if (CoroAllocInst *CA = CoroId->getCoroAlloc()) {
1703 // If there is a CoroAlloc and it returns false (meaning we elide the
1704 // allocation, use CleanupFn instead of DestroyFn).
1705 DestroyOrCleanupFn = Builder.CreateSelect(C: CA, True: DestroyFn, False: CleanupFn);
1706 applyProfMetadataIfEnabled(V: DestroyOrCleanupFn, setMetadataCallback: [&](Instruction *Inst) {
1707 setExplicitlyUnknownBranchWeightsIfProfiled(I&: *Inst, DEBUG_TYPE,
1708 F: CoroId->getFunction());
1709 });
1710 }
1711
1712 // Destroy function pointer
1713 Value *DestroyAddr = Builder.CreateInBoundsPtrAdd(
1714 Ptr: Shape.FramePtr,
1715 Offset: ConstantInt::get(Ty: Type::getInt64Ty(C),
1716 V: Shape.SwitchLowering.DestroyOffset),
1717 Name: "destroy.addr");
1718 Builder.CreateStore(Val: DestroyOrCleanupFn, Ptr: DestroyAddr);
1719 }
1720
1721 // Create a global constant array containing pointers to functions provided
1722 // and set Info parameter of CoroBegin to point at this constant. Example:
1723 //
1724 // @f.resumers = internal constant [2 x void(%f.frame*)*]
1725 // [void(%f.frame*)* @f.resume, void(%f.frame*)*
1726 // @f.destroy]
1727 // define void @f() {
1728 // ...
1729 // call i8* @llvm.coro.begin(i8* null, i32 0, i8* null,
1730 // i8* bitcast([2 x void(%f.frame*)*] * @f.resumers to
1731 // i8*))
1732 //
1733 // Assumes that all the functions have the same signature.
1734 static void setCoroInfo(Function &F, coro::Shape &Shape,
1735 ArrayRef<Function *> Fns) {
1736 // This only works under the switch-lowering ABI because coro elision
1737 // only works on the switch-lowering ABI.
1738 SmallVector<Constant *, 4> Args(Fns);
1739 assert(!Args.empty());
1740 Function *Part = *Fns.begin();
1741 Module *M = Part->getParent();
1742 auto *ArrTy = ArrayType::get(ElementType: Part->getType(), NumElements: Args.size());
1743
1744 auto *ConstVal = ConstantArray::get(T: ArrTy, V: Args);
1745 auto *GV = new GlobalVariable(*M, ConstVal->getType(), /*isConstant=*/true,
1746 GlobalVariable::PrivateLinkage, ConstVal,
1747 F.getName() + Twine(".resumers"));
1748
1749 // Update coro.begin instruction to refer to this constant.
1750 LLVMContext &C = F.getContext();
1751 auto *BC = ConstantExpr::getPointerCast(C: GV, Ty: PointerType::getUnqual(C));
1752 Shape.getSwitchCoroId()->setInfo(BC);
1753 }
1754};
1755
1756} // namespace
1757
1758static void replaceAsyncResumeFunction(CoroSuspendAsyncInst *Suspend,
1759 Value *Continuation) {
1760 auto *ResumeIntrinsic = Suspend->getResumeFunction();
1761 auto &Context = Suspend->getParent()->getParent()->getContext();
1762 auto *Int8PtrTy = PointerType::getUnqual(C&: Context);
1763
1764 IRBuilder<> Builder(ResumeIntrinsic);
1765 auto *Val = Builder.CreateBitOrPointerCast(V: Continuation, DestTy: Int8PtrTy);
1766 ResumeIntrinsic->replaceAllUsesWith(V: Val);
1767 ResumeIntrinsic->eraseFromParent();
1768 Suspend->setOperand(i_nocapture: CoroSuspendAsyncInst::ResumeFunctionArg,
1769 Val_nocapture: PoisonValue::get(T: Int8PtrTy));
1770}
1771
1772/// Coerce the arguments in \p FnArgs according to \p FnTy in \p CallArgs.
1773static void coerceArguments(IRBuilder<> &Builder, FunctionType *FnTy,
1774 ArrayRef<Value *> FnArgs,
1775 SmallVectorImpl<Value *> &CallArgs) {
1776 size_t ArgIdx = 0;
1777 for (auto *paramTy : FnTy->params()) {
1778 assert(ArgIdx < FnArgs.size());
1779 if (paramTy != FnArgs[ArgIdx]->getType())
1780 CallArgs.push_back(
1781 Elt: Builder.CreateBitOrPointerCast(V: FnArgs[ArgIdx], DestTy: paramTy));
1782 else
1783 CallArgs.push_back(Elt: FnArgs[ArgIdx]);
1784 ++ArgIdx;
1785 }
1786}
1787
1788CallInst *coro::createMustTailCall(DebugLoc Loc, Function *MustTailCallFn,
1789 TargetTransformInfo &TTI,
1790 ArrayRef<Value *> Arguments,
1791 IRBuilder<> &Builder) {
1792 auto *FnTy = MustTailCallFn->getFunctionType();
1793 // Coerce the arguments, llvm optimizations seem to ignore the types in
1794 // vaarg functions and throws away casts in optimized mode.
1795 SmallVector<Value *, 8> CallArgs;
1796 coerceArguments(Builder, FnTy, FnArgs: Arguments, CallArgs);
1797
1798 auto *TailCall = Builder.CreateCall(FTy: FnTy, Callee: MustTailCallFn, Args: CallArgs);
1799 // Skip targets which don't support tail call.
1800 if (TTI.supportsTailCallFor(CB: TailCall)) {
1801 TailCall->setTailCallKind(CallInst::TCK_MustTail);
1802 }
1803 TailCall->setDebugLoc(Loc);
1804 TailCall->setCallingConv(MustTailCallFn->getCallingConv());
1805 return TailCall;
1806}
1807
1808void coro::AsyncABI::splitCoroutine(Function &F, coro::Shape &Shape,
1809 SmallVectorImpl<Function *> &Clones,
1810 TargetTransformInfo &TTI) {
1811 assert(Shape.ABI == coro::ABI::Async);
1812 assert(Clones.empty());
1813 // Reset various things that the optimizer might have decided it
1814 // "knows" about the coroutine function due to not seeing a return.
1815 F.removeFnAttr(Kind: Attribute::NoReturn);
1816 F.removeRetAttr(Kind: Attribute::NoAlias);
1817 F.removeRetAttr(Kind: Attribute::NonNull);
1818
1819 auto &Context = F.getContext();
1820 auto *Int8PtrTy = PointerType::getUnqual(C&: Context);
1821
1822 auto *Id = Shape.getAsyncCoroId();
1823 IRBuilder<> Builder(Id);
1824
1825 auto *FramePtr = Id->getStorage();
1826 FramePtr = Builder.CreateBitOrPointerCast(V: FramePtr, DestTy: Int8PtrTy);
1827 FramePtr = Builder.CreateInBoundsPtrAdd(
1828 Ptr: FramePtr,
1829 Offset: ConstantInt::get(Ty: Type::getInt64Ty(C&: Context),
1830 V: Shape.AsyncLowering.FrameOffset),
1831 Name: "async.ctx.frameptr");
1832
1833 // Map all uses of llvm.coro.begin to the allocated frame pointer.
1834 {
1835 // Make sure we don't invalidate Shape.FramePtr.
1836 TrackingVH<Value> Handle(Shape.FramePtr);
1837 Shape.CoroBegin->replaceAllUsesWith(V: FramePtr);
1838 Shape.FramePtr = Handle.getValPtr();
1839 }
1840
1841 // Create all the functions in order after the main function.
1842 auto NextF = std::next(x: F.getIterator());
1843
1844 // Create a continuation function for each of the suspend points.
1845 Clones.reserve(N: Shape.CoroSuspends.size());
1846 for (auto [Idx, CS] : llvm::enumerate(First&: Shape.CoroSuspends)) {
1847 auto *Suspend = cast<CoroSuspendAsyncInst>(Val: CS);
1848
1849 // Create the clone declaration.
1850 auto ResumeNameSuffix = ".resume.";
1851 auto ProjectionFunctionName =
1852 Suspend->getAsyncContextProjectionFunction()->getName();
1853 bool UseSwiftMangling = false;
1854 if (ProjectionFunctionName == "__swift_async_resume_project_context") {
1855 ResumeNameSuffix = "TQ";
1856 UseSwiftMangling = true;
1857 } else if (ProjectionFunctionName == "__swift_async_resume_get_context") {
1858 ResumeNameSuffix = "TY";
1859 UseSwiftMangling = true;
1860 }
1861 auto *Continuation = createCloneDeclaration(
1862 OrigF&: F, Shape,
1863 Suffix: UseSwiftMangling ? ResumeNameSuffix + Twine(Idx) + "_"
1864 : ResumeNameSuffix + Twine(Idx),
1865 InsertBefore: NextF, ActiveSuspend: Suspend);
1866 Clones.push_back(Elt: Continuation);
1867
1868 // Insert a branch to a new return block immediately before the suspend
1869 // point.
1870 auto *SuspendBB = Suspend->getParent();
1871 auto *NewSuspendBB = SuspendBB->splitBasicBlock(I: Suspend);
1872 auto *Branch = cast<UncondBrInst>(Val: SuspendBB->getTerminator());
1873
1874 // Place it before the first suspend.
1875 auto *ReturnBB =
1876 BasicBlock::Create(Context&: F.getContext(), Name: "coro.return", Parent: &F, InsertBefore: NewSuspendBB);
1877 Branch->setSuccessor(idx: 0, NewSucc: ReturnBB);
1878
1879 IRBuilder<> Builder(ReturnBB);
1880
1881 // Insert the call to the tail call function and inline it.
1882 auto *Fn = Suspend->getMustTailCallFunction();
1883 SmallVector<Value *, 8> Args(Suspend->args());
1884 auto FnArgs = ArrayRef<Value *>(Args).drop_front(
1885 N: CoroSuspendAsyncInst::MustTailCallFuncArg + 1);
1886 auto *TailCall = coro::createMustTailCall(Loc: Suspend->getDebugLoc(), MustTailCallFn: Fn, TTI,
1887 Arguments: FnArgs, Builder);
1888 Builder.CreateRetVoid();
1889 InlineFunctionInfo FnInfo;
1890 (void)InlineFunction(CB&: *TailCall, IFI&: FnInfo);
1891
1892 // Replace the lvm.coro.async.resume intrisic call.
1893 replaceAsyncResumeFunction(Suspend, Continuation);
1894 }
1895
1896 assert(Clones.size() == Shape.CoroSuspends.size());
1897
1898 for (auto [Idx, CS] : llvm::enumerate(First&: Shape.CoroSuspends)) {
1899 auto *Suspend = CS;
1900 auto *Clone = Clones[Idx];
1901
1902 coro::BaseCloner::createClone(OrigF&: F, Suffix: "resume." + Twine(Idx), Shape, NewF: Clone,
1903 ActiveSuspend: Suspend, TTI);
1904 }
1905}
1906
1907void coro::AnyRetconABI::splitCoroutine(Function &F, coro::Shape &Shape,
1908 SmallVectorImpl<Function *> &Clones,
1909 TargetTransformInfo &TTI) {
1910 assert(Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce);
1911 assert(Clones.empty());
1912
1913 // Reset various things that the optimizer might have decided it
1914 // "knows" about the coroutine function due to not seeing a return.
1915 F.removeFnAttr(Kind: Attribute::NoReturn);
1916 F.removeRetAttr(Kind: Attribute::NoAlias);
1917 F.removeRetAttr(Kind: Attribute::NonNull);
1918
1919 // Allocate the frame.
1920 auto *Id = Shape.getRetconCoroId();
1921 Value *RawFramePtr;
1922 if (Shape.RetconLowering.IsFrameInlineInStorage) {
1923 RawFramePtr = Id->getStorage();
1924 } else {
1925 IRBuilder<> Builder(Id);
1926
1927 auto FrameSize = Builder.getInt64(C: Shape.FrameSize);
1928
1929 // Allocate. We don't need to update the call graph node because we're
1930 // going to recompute it from scratch after splitting.
1931 // FIXME: pass the required alignment
1932 RawFramePtr = Shape.emitAlloc(Builder, Size: FrameSize, CG: nullptr);
1933 RawFramePtr =
1934 Builder.CreateBitCast(V: RawFramePtr, DestTy: Shape.CoroBegin->getType());
1935
1936 // Stash the allocated frame pointer in the continuation storage.
1937 Builder.CreateStore(Val: RawFramePtr, Ptr: Id->getStorage());
1938 }
1939
1940 // Map all uses of llvm.coro.begin to the allocated frame pointer.
1941 {
1942 // Make sure we don't invalidate Shape.FramePtr.
1943 TrackingVH<Value> Handle(Shape.FramePtr);
1944 Shape.CoroBegin->replaceAllUsesWith(V: RawFramePtr);
1945 Shape.FramePtr = Handle.getValPtr();
1946 }
1947
1948 // Create a unique return block.
1949 BasicBlock *ReturnBB = nullptr;
1950 PHINode *ContinuationPhi = nullptr;
1951 SmallVector<PHINode *, 4> ReturnPHIs;
1952
1953 // Create all the functions in order after the main function.
1954 auto NextF = std::next(x: F.getIterator());
1955
1956 // Create a continuation function for each of the suspend points.
1957 Clones.reserve(N: Shape.CoroSuspends.size());
1958 for (auto [Idx, CS] : llvm::enumerate(First&: Shape.CoroSuspends)) {
1959 auto Suspend = cast<CoroSuspendRetconInst>(Val: CS);
1960
1961 // Create the clone declaration.
1962 auto Continuation = createCloneDeclaration(
1963 OrigF&: F, Shape, Suffix: ".resume." + Twine(Idx), InsertBefore: NextF, ActiveSuspend: nullptr);
1964 Clones.push_back(Elt: Continuation);
1965
1966 // Insert a branch to the unified return block immediately before
1967 // the suspend point.
1968 auto SuspendBB = Suspend->getParent();
1969 auto NewSuspendBB = SuspendBB->splitBasicBlock(I: Suspend);
1970 auto Branch = cast<UncondBrInst>(Val: SuspendBB->getTerminator());
1971
1972 // Create the unified return block.
1973 if (!ReturnBB) {
1974 // Place it before the first suspend.
1975 ReturnBB =
1976 BasicBlock::Create(Context&: F.getContext(), Name: "coro.return", Parent: &F, InsertBefore: NewSuspendBB);
1977 Shape.RetconLowering.ReturnBlock = ReturnBB;
1978
1979 IRBuilder<> Builder(ReturnBB);
1980
1981 // First, the continuation.
1982 ContinuationPhi =
1983 Builder.CreatePHI(Ty: Continuation->getType(), NumReservedValues: Shape.CoroSuspends.size());
1984
1985 // Create PHIs for all other return values.
1986 assert(ReturnPHIs.empty());
1987
1988 // Next, all the directly-yielded values.
1989 for (auto *ResultTy : Shape.getRetconResultTypes())
1990 ReturnPHIs.push_back(
1991 Elt: Builder.CreatePHI(Ty: ResultTy, NumReservedValues: Shape.CoroSuspends.size()));
1992
1993 // Build the return value.
1994 auto RetTy = F.getReturnType();
1995
1996 // Cast the continuation value if necessary.
1997 // We can't rely on the types matching up because that type would
1998 // have to be infinite.
1999 auto CastedContinuationTy =
2000 (ReturnPHIs.empty() ? RetTy : RetTy->getStructElementType(N: 0));
2001 auto *CastedContinuation =
2002 Builder.CreateBitCast(V: ContinuationPhi, DestTy: CastedContinuationTy);
2003
2004 Value *RetV = CastedContinuation;
2005 if (!ReturnPHIs.empty()) {
2006 auto ValueIdx = 0;
2007 RetV = PoisonValue::get(T: RetTy);
2008 RetV = Builder.CreateInsertValue(Agg: RetV, Val: CastedContinuation, Idxs: ValueIdx++);
2009
2010 for (auto Phi : ReturnPHIs)
2011 RetV = Builder.CreateInsertValue(Agg: RetV, Val: Phi, Idxs: ValueIdx++);
2012 }
2013
2014 Builder.CreateRet(V: RetV);
2015 }
2016
2017 // Branch to the return block.
2018 Branch->setSuccessor(idx: 0, NewSucc: ReturnBB);
2019 assert(ContinuationPhi);
2020 ContinuationPhi->addIncoming(V: Continuation, BB: SuspendBB);
2021 for (auto [Phi, VUse] :
2022 llvm::zip_equal(t&: ReturnPHIs, u: Suspend->value_operands()))
2023 Phi->addIncoming(V: VUse, BB: SuspendBB);
2024 }
2025
2026 assert(Clones.size() == Shape.CoroSuspends.size());
2027
2028 for (auto [Idx, CS] : llvm::enumerate(First&: Shape.CoroSuspends)) {
2029 auto Suspend = CS;
2030 auto Clone = Clones[Idx];
2031
2032 coro::BaseCloner::createClone(OrigF&: F, Suffix: "resume." + Twine(Idx), Shape, NewF: Clone,
2033 ActiveSuspend: Suspend, TTI);
2034 }
2035}
2036
2037namespace {
2038class PrettyStackTraceFunction : public PrettyStackTraceEntry {
2039 Function &F;
2040
2041public:
2042 PrettyStackTraceFunction(Function &F) : F(F) {}
2043 void print(raw_ostream &OS) const override {
2044 OS << "While splitting coroutine ";
2045 F.printAsOperand(O&: OS, /*print type*/ PrintType: false, M: F.getParent());
2046 OS << "\n";
2047 }
2048};
2049} // namespace
2050
2051/// Remove calls to llvm.coro.end in the original function.
2052static void removeCoroEndsFromRampFunction(const coro::Shape &Shape) {
2053 if (Shape.ABI != coro::ABI::Switch) {
2054 for (auto *End : Shape.CoroEnds) {
2055 replaceCoroEnd(End, Shape, FramePtr: Shape.FramePtr, /*in ramp*/ InRamp: true, CG: nullptr);
2056 }
2057 } else {
2058 for (llvm::AnyCoroEndInst *End : Shape.CoroEnds)
2059 End->eraseFromParent();
2060 }
2061}
2062
2063static void removeCoroIsInRampFromRampFunction(const coro::Shape &Shape) {
2064 for (auto *II : Shape.CoroIsInRampInsts) {
2065 auto &Ctx = II->getContext();
2066 II->replaceAllUsesWith(V: ConstantInt::getTrue(Context&: Ctx));
2067 II->eraseFromParent();
2068 }
2069}
2070
2071static bool hasSafeElideCaller(Function &F) {
2072 for (auto *U : F.users()) {
2073 if (auto *CB = dyn_cast<CallBase>(Val: U)) {
2074 auto *Caller = CB->getFunction();
2075 if (Caller && Caller->isPresplitCoroutine() &&
2076 CB->hasFnAttr(Kind: llvm::Attribute::CoroElideSafe))
2077 return true;
2078 }
2079 }
2080 return false;
2081}
2082
2083void coro::SwitchABI::splitCoroutine(Function &F, coro::Shape &Shape,
2084 SmallVectorImpl<Function *> &Clones,
2085 TargetTransformInfo &TTI) {
2086 SwitchCoroutineSplitter::split(F, Shape, Clones, TTI);
2087}
2088
2089static void doSplitCoroutine(Function &F, SmallVectorImpl<Function *> &Clones,
2090 coro::BaseABI &ABI, TargetTransformInfo &TTI,
2091 bool OptimizeFrame) {
2092 PrettyStackTraceFunction prettyStackTrace(F);
2093
2094 auto &Shape = ABI.Shape;
2095 assert(Shape.CoroBegin);
2096
2097 lowerAwaitSuspends(F, Shape);
2098
2099 simplifySuspendPoints(Shape);
2100
2101 normalizeCoroutine(F, Shape, TTI);
2102 ABI.buildCoroutineFrame(OptimizeFrame);
2103 replaceFrameSizeAndAlignment(Shape);
2104
2105 bool isNoSuspendCoroutine = Shape.CoroSuspends.empty();
2106
2107 bool shouldCreateNoAllocVariant =
2108 !isNoSuspendCoroutine && Shape.ABI == coro::ABI::Switch &&
2109 hasSafeElideCaller(F) && !F.hasFnAttribute(Kind: llvm::Attribute::NoInline);
2110 if (Shape.ABI == coro::ABI::Switch)
2111 Shape.SwitchLowering.HasCoroElideNoAllocVariant =
2112 shouldCreateNoAllocVariant;
2113
2114 // If there are no suspend points, no split required, just remove
2115 // the allocation and deallocation blocks, they are not needed.
2116 if (isNoSuspendCoroutine) {
2117 handleNoSuspendCoroutine(Shape);
2118 } else {
2119 ABI.splitCoroutine(F, Shape, Clones, TTI);
2120 }
2121
2122 // Replace all the swifterror operations in the original function.
2123 // This invalidates SwiftErrorOps in the Shape.
2124 replaceSwiftErrorOps(F, Shape, VMap: nullptr);
2125
2126 // Salvage debug intrinsics that point into the coroutine frame in the
2127 // original function. The Cloner has already salvaged debug info in the new
2128 // coroutine funclets.
2129 SmallDenseMap<Argument *, AllocaInst *, 4> ArgToAllocaMap;
2130 auto DbgVariableRecords = collectDbgVariableRecords(F);
2131 for (DbgVariableRecord *DVR : DbgVariableRecords)
2132 coro::salvageDebugInfo(ArgToAllocaMap, DVR&: *DVR, UseEntryValue: false /*UseEntryValue*/);
2133
2134 removeCoroEndsFromRampFunction(Shape);
2135 removeCoroIsInRampFromRampFunction(Shape);
2136
2137 if (shouldCreateNoAllocVariant)
2138 SwitchCoroutineSplitter::createNoAllocVariant(F, Shape, Clones);
2139}
2140
2141static LazyCallGraph::SCC &updateCallGraphAfterCoroutineSplit(
2142 LazyCallGraph::Node &N, const coro::Shape &Shape,
2143 const SmallVectorImpl<Function *> &Clones, LazyCallGraph::SCC &C,
2144 LazyCallGraph &CG, CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR,
2145 FunctionAnalysisManager &FAM) {
2146
2147 auto *CurrentSCC = &C;
2148 if (!Clones.empty()) {
2149 switch (Shape.ABI) {
2150 case coro::ABI::Switch:
2151 // The resume clone's elided-frame check holds a reference to the cleanup
2152 // clone. Add the cleanup clone first, so populating the resume node does
2153 // not materialize an unregistered cleanup node.
2154 if (Shape.SwitchLowering.HasCoroElideNoAllocVariant) {
2155 assert(Clones.size() >= 3 && "expected switch coroutine clones");
2156 CG.addSplitFunction(OriginalFunction&: N.getFunction(), NewFunction&: *Clones[2]);
2157 CG.addSplitFunction(OriginalFunction&: N.getFunction(), NewFunction&: *Clones[1]);
2158 CG.addSplitFunction(OriginalFunction&: N.getFunction(), NewFunction&: *Clones[0]);
2159 for (Function *Clone : drop_begin(RangeOrContainer: Clones, N: 3))
2160 CG.addSplitFunction(OriginalFunction&: N.getFunction(), NewFunction&: *Clone);
2161 } else {
2162 // Each clone in the Switch lowering is independent of the other
2163 // clones. Let the LazyCallGraph know about each one separately.
2164 for (Function *Clone : Clones)
2165 CG.addSplitFunction(OriginalFunction&: N.getFunction(), NewFunction&: *Clone);
2166 }
2167 break;
2168 case coro::ABI::Async:
2169 case coro::ABI::Retcon:
2170 case coro::ABI::RetconOnce:
2171 // Each clone in the Async/Retcon lowering references of the other clones.
2172 // Let the LazyCallGraph know about all of them at once.
2173 if (!Clones.empty())
2174 CG.addSplitRefRecursiveFunctions(OriginalFunction&: N.getFunction(), NewFunctions: Clones);
2175 break;
2176 }
2177
2178 // Let the CGSCC infra handle the changes to the original function.
2179 CurrentSCC = &updateCGAndAnalysisManagerForCGSCCPass(G&: CG, C&: *CurrentSCC, N, AM,
2180 UR, FAM);
2181 }
2182
2183 // Do some cleanup and let the CGSCC infra see if we've cleaned up any edges
2184 // to the split functions.
2185 postSplitCleanup(F&: N.getFunction());
2186 CurrentSCC = &updateCGAndAnalysisManagerForFunctionPass(G&: CG, C&: *CurrentSCC, N,
2187 AM, UR, FAM);
2188 return *CurrentSCC;
2189}
2190
2191/// Replace a call to llvm.coro.prepare.retcon.
2192static void replacePrepare(CallInst *Prepare, LazyCallGraph &CG,
2193 LazyCallGraph::SCC &C) {
2194 auto CastFn = Prepare->getArgOperand(i: 0); // as an i8*
2195 auto Fn = CastFn->stripPointerCasts(); // as its original type
2196
2197 // Attempt to peephole this pattern:
2198 // %0 = bitcast [[TYPE]] @some_function to i8*
2199 // %1 = call @llvm.coro.prepare.retcon(i8* %0)
2200 // %2 = bitcast %1 to [[TYPE]]
2201 // ==>
2202 // %2 = @some_function
2203 for (Use &U : llvm::make_early_inc_range(Range: Prepare->uses())) {
2204 // Look for bitcasts back to the original function type.
2205 auto *Cast = dyn_cast<BitCastInst>(Val: U.getUser());
2206 if (!Cast || Cast->getType() != Fn->getType())
2207 continue;
2208
2209 // Replace and remove the cast.
2210 Cast->replaceAllUsesWith(V: Fn);
2211 Cast->eraseFromParent();
2212 }
2213
2214 // Replace any remaining uses with the function as an i8*.
2215 // This can never directly be a callee, so we don't need to update CG.
2216 Prepare->replaceAllUsesWith(V: CastFn);
2217 Prepare->eraseFromParent();
2218
2219 // Kill dead bitcasts.
2220 while (auto *Cast = dyn_cast<BitCastInst>(Val: CastFn)) {
2221 if (!Cast->use_empty())
2222 break;
2223 CastFn = Cast->getOperand(i_nocapture: 0);
2224 Cast->eraseFromParent();
2225 }
2226}
2227
2228static bool replaceAllPrepares(Function *PrepareFn, LazyCallGraph &CG,
2229 LazyCallGraph::SCC &C) {
2230 bool Changed = false;
2231 for (Use &P : llvm::make_early_inc_range(Range: PrepareFn->uses())) {
2232 // Intrinsics can only be used in calls.
2233 auto *Prepare = cast<CallInst>(Val: P.getUser());
2234 replacePrepare(Prepare, CG, C);
2235 Changed = true;
2236 }
2237
2238 return Changed;
2239}
2240
2241static void addPrepareFunction(const Module &M,
2242 SmallVectorImpl<Function *> &Fns,
2243 StringRef Name) {
2244 auto *PrepareFn = M.getFunction(Name);
2245 if (PrepareFn && !PrepareFn->use_empty())
2246 Fns.push_back(Elt: PrepareFn);
2247}
2248
2249static std::unique_ptr<coro::BaseABI>
2250CreateNewABI(Function &F, coro::Shape &S,
2251 std::function<bool(Instruction &)> IsMatCallback,
2252 const SmallVector<CoroSplitPass::BaseABITy> GenCustomABIs) {
2253 if (S.CoroBegin->hasCustomABI()) {
2254 unsigned CustomABI = S.CoroBegin->getCustomABI();
2255 if (CustomABI >= GenCustomABIs.size())
2256 llvm_unreachable("Custom ABI not found amoung those specified");
2257 return GenCustomABIs[CustomABI](F, S);
2258 }
2259
2260 switch (S.ABI) {
2261 case coro::ABI::Switch:
2262 return std::make_unique<coro::SwitchABI>(args&: F, args&: S, args&: IsMatCallback);
2263 case coro::ABI::Async:
2264 return std::make_unique<coro::AsyncABI>(args&: F, args&: S, args&: IsMatCallback);
2265 case coro::ABI::Retcon:
2266 return std::make_unique<coro::AnyRetconABI>(args&: F, args&: S, args&: IsMatCallback);
2267 case coro::ABI::RetconOnce:
2268 return std::make_unique<coro::AnyRetconABI>(args&: F, args&: S, args&: IsMatCallback);
2269 }
2270 llvm_unreachable("Unknown ABI");
2271}
2272
2273CoroSplitPass::CoroSplitPass(bool OptimizeFrame)
2274 : CreateAndInitABI([](Function &F, coro::Shape &S) {
2275 std::unique_ptr<coro::BaseABI> ABI =
2276 CreateNewABI(F, S, IsMatCallback: coro::isTriviallyMaterializable, GenCustomABIs: {});
2277 ABI->init();
2278 return ABI;
2279 }),
2280 OptimizeFrame(OptimizeFrame) {}
2281
2282CoroSplitPass::CoroSplitPass(
2283 SmallVector<CoroSplitPass::BaseABITy> GenCustomABIs, bool OptimizeFrame)
2284 : CreateAndInitABI([=](Function &F, coro::Shape &S) {
2285 std::unique_ptr<coro::BaseABI> ABI =
2286 CreateNewABI(F, S, IsMatCallback: coro::isTriviallyMaterializable, GenCustomABIs);
2287 ABI->init();
2288 return ABI;
2289 }),
2290 OptimizeFrame(OptimizeFrame) {}
2291
2292// For back compatibility, constructor takes a materializable callback and
2293// creates a generator for an ABI with a modified materializable callback.
2294CoroSplitPass::CoroSplitPass(std::function<bool(Instruction &)> IsMatCallback,
2295 bool OptimizeFrame)
2296 : CreateAndInitABI([=](Function &F, coro::Shape &S) {
2297 std::unique_ptr<coro::BaseABI> ABI =
2298 CreateNewABI(F, S, IsMatCallback, GenCustomABIs: {});
2299 ABI->init();
2300 return ABI;
2301 }),
2302 OptimizeFrame(OptimizeFrame) {}
2303
2304// For back compatibility, constructor takes a materializable callback and
2305// creates a generator for an ABI with a modified materializable callback.
2306CoroSplitPass::CoroSplitPass(
2307 std::function<bool(Instruction &)> IsMatCallback,
2308 SmallVector<CoroSplitPass::BaseABITy> GenCustomABIs, bool OptimizeFrame)
2309 : CreateAndInitABI([=](Function &F, coro::Shape &S) {
2310 std::unique_ptr<coro::BaseABI> ABI =
2311 CreateNewABI(F, S, IsMatCallback, GenCustomABIs);
2312 ABI->init();
2313 return ABI;
2314 }),
2315 OptimizeFrame(OptimizeFrame) {}
2316
2317PreservedAnalyses CoroSplitPass::run(LazyCallGraph::SCC &C,
2318 CGSCCAnalysisManager &AM,
2319 LazyCallGraph &CG, CGSCCUpdateResult &UR) {
2320 // NB: One invariant of a valid LazyCallGraph::SCC is that it must contain a
2321 // non-zero number of nodes, so we assume that here and grab the first
2322 // node's function's module.
2323 Module &M = *C.begin()->getFunction().getParent();
2324 auto &FAM =
2325 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(IR&: C, ExtraArgs&: CG).getManager();
2326
2327 // Check for uses of llvm.coro.prepare.retcon/async.
2328 SmallVector<Function *, 2> PrepareFns;
2329 addPrepareFunction(M, Fns&: PrepareFns, Name: "llvm.coro.prepare.retcon");
2330 addPrepareFunction(M, Fns&: PrepareFns, Name: "llvm.coro.prepare.async");
2331
2332 // Find coroutines for processing.
2333 SmallVector<LazyCallGraph::Node *> Coroutines;
2334 for (LazyCallGraph::Node &N : C)
2335 if (N.getFunction().isPresplitCoroutine())
2336 Coroutines.push_back(Elt: &N);
2337
2338 if (Coroutines.empty() && PrepareFns.empty())
2339 return PreservedAnalyses::all();
2340
2341 auto *CurrentSCC = &C;
2342 // Split all the coroutines.
2343 for (LazyCallGraph::Node *N : Coroutines) {
2344 Function &F = N->getFunction();
2345 LLVM_DEBUG(dbgs() << "CoroSplit: Processing coroutine '" << F.getName()
2346 << "\n");
2347
2348 // The suspend-crossing algorithm in buildCoroutineFrame gets tripped up
2349 // by unreachable blocks, so remove them as a first pass. Remove the
2350 // unreachable blocks before collecting intrinsics into Shape.
2351 removeUnreachableBlocks(F);
2352
2353 coro::Shape Shape(F);
2354 if (!Shape.CoroBegin)
2355 continue;
2356
2357 F.setSplittedCoroutine();
2358
2359 // Query BFI and populate SuspendFreqs right before splitting.
2360 auto &BFI = FAM.getResult<BlockFrequencyAnalysis>(IR&: F);
2361 for (auto *AnyS : Shape.CoroSuspends) {
2362 BasicBlock *BB = AnyS->getParent();
2363 uint64_t Freq = BFI.getBlockFreq(BB).getFrequency();
2364 Shape.SuspendFreqs[AnyS] = Freq;
2365
2366 // Query BFI to get the actual estimated execution profile count of the
2367 // basic block where this suspension point resides.
2368 std::optional<uint64_t> Count =
2369 BFI.getBlockProfileCount(BB, /*AllowSynthetic=*/true);
2370 if (Count.has_value()) {
2371 if (!Shape.ResumeEntryCount.has_value()) {
2372 // For the first suspend point visited, initialize the total sum.
2373 Shape.ResumeEntryCount = Count.value();
2374 } else {
2375 // Accumulate the absolute execution count of each subsequent suspend
2376 // point into the total sum.
2377 Shape.ResumeEntryCount.value() += Count.value();
2378 }
2379 }
2380 }
2381
2382 std::unique_ptr<coro::BaseABI> ABI = CreateAndInitABI(F, Shape);
2383
2384 SmallVector<Function *, 4> Clones;
2385 auto &TTI = FAM.getResult<TargetIRAnalysis>(IR&: F);
2386 doSplitCoroutine(F, Clones, ABI&: *ABI, TTI, OptimizeFrame);
2387 CurrentSCC = &updateCallGraphAfterCoroutineSplit(
2388 N&: *N, Shape, Clones, C&: *CurrentSCC, CG, AM, UR, FAM);
2389
2390 auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F);
2391 ORE.emit(RemarkBuilder: [&]() {
2392 return OptimizationRemark(DEBUG_TYPE, "CoroSplit", &F)
2393 << "Split '" << ore::NV("function", F.getName())
2394 << "' (frame_size=" << ore::NV("frame_size", Shape.FrameSize)
2395 << ", align=" << ore::NV("align", Shape.FrameAlign.value()) << ")";
2396 });
2397
2398 if (!Shape.CoroSuspends.empty()) {
2399 // Run the CGSCC pipeline on the original and newly split functions.
2400 UR.CWorklist.insert(X: CurrentSCC);
2401 for (Function *Clone : Clones)
2402 UR.CWorklist.insert(X: CG.lookupSCC(N&: CG.get(F&: *Clone)));
2403 } else if (Shape.ABI == coro::ABI::Async) {
2404 // Reprocess the function to inline the tail called return function of
2405 // coro.async.end.
2406 UR.CWorklist.insert(X: &C);
2407 }
2408 }
2409
2410 for (auto *PrepareFn : PrepareFns) {
2411 replaceAllPrepares(PrepareFn, CG, C&: *CurrentSCC);
2412 }
2413
2414 return PreservedAnalyses::none();
2415}
2416