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 "CoroInstr.h" |
23 | #include "CoroInternal.h" |
24 | #include "llvm/ADT/DenseMap.h" |
25 | #include "llvm/ADT/PriorityWorklist.h" |
26 | #include "llvm/ADT/SmallPtrSet.h" |
27 | #include "llvm/ADT/SmallVector.h" |
28 | #include "llvm/ADT/StringRef.h" |
29 | #include "llvm/ADT/Twine.h" |
30 | #include "llvm/Analysis/CFG.h" |
31 | #include "llvm/Analysis/CallGraph.h" |
32 | #include "llvm/Analysis/ConstantFolding.h" |
33 | #include "llvm/Analysis/LazyCallGraph.h" |
34 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
35 | #include "llvm/Analysis/TargetTransformInfo.h" |
36 | #include "llvm/BinaryFormat/Dwarf.h" |
37 | #include "llvm/IR/Argument.h" |
38 | #include "llvm/IR/Attributes.h" |
39 | #include "llvm/IR/BasicBlock.h" |
40 | #include "llvm/IR/CFG.h" |
41 | #include "llvm/IR/CallingConv.h" |
42 | #include "llvm/IR/Constants.h" |
43 | #include "llvm/IR/DataLayout.h" |
44 | #include "llvm/IR/DerivedTypes.h" |
45 | #include "llvm/IR/Dominators.h" |
46 | #include "llvm/IR/Function.h" |
47 | #include "llvm/IR/GlobalValue.h" |
48 | #include "llvm/IR/GlobalVariable.h" |
49 | #include "llvm/IR/IRBuilder.h" |
50 | #include "llvm/IR/InstIterator.h" |
51 | #include "llvm/IR/InstrTypes.h" |
52 | #include "llvm/IR/Instruction.h" |
53 | #include "llvm/IR/Instructions.h" |
54 | #include "llvm/IR/IntrinsicInst.h" |
55 | #include "llvm/IR/LLVMContext.h" |
56 | #include "llvm/IR/Module.h" |
57 | #include "llvm/IR/Type.h" |
58 | #include "llvm/IR/Value.h" |
59 | #include "llvm/IR/Verifier.h" |
60 | #include "llvm/Support/Casting.h" |
61 | #include "llvm/Support/Debug.h" |
62 | #include "llvm/Support/PrettyStackTrace.h" |
63 | #include "llvm/Support/raw_ostream.h" |
64 | #include "llvm/Transforms/Scalar.h" |
65 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
66 | #include "llvm/Transforms/Utils/CallGraphUpdater.h" |
67 | #include "llvm/Transforms/Utils/Cloning.h" |
68 | #include "llvm/Transforms/Utils/Local.h" |
69 | #include "llvm/Transforms/Utils/ValueMapper.h" |
70 | #include <cassert> |
71 | #include <cstddef> |
72 | #include <cstdint> |
73 | #include <initializer_list> |
74 | #include <iterator> |
75 | |
76 | using namespace llvm; |
77 | |
78 | #define DEBUG_TYPE "coro-split" |
79 | |
80 | namespace { |
81 | |
82 | /// A little helper class for building |
83 | class CoroCloner { |
84 | public: |
85 | enum class Kind { |
86 | /// The shared resume function for a switch lowering. |
87 | SwitchResume, |
88 | |
89 | /// The shared unwind function for a switch lowering. |
90 | SwitchUnwind, |
91 | |
92 | /// The shared cleanup function for a switch lowering. |
93 | SwitchCleanup, |
94 | |
95 | /// An individual continuation function. |
96 | Continuation, |
97 | |
98 | /// An async resume function. |
99 | Async, |
100 | }; |
101 | |
102 | private: |
103 | Function &OrigF; |
104 | Function *NewF; |
105 | const Twine &Suffix; |
106 | coro::Shape &Shape; |
107 | Kind FKind; |
108 | ValueToValueMapTy VMap; |
109 | IRBuilder<> Builder; |
110 | Value *NewFramePtr = nullptr; |
111 | |
112 | /// The active suspend instruction; meaningful only for continuation and async |
113 | /// ABIs. |
114 | AnyCoroSuspendInst *ActiveSuspend = nullptr; |
115 | |
116 | public: |
117 | /// Create a cloner for a switch lowering. |
118 | CoroCloner(Function &OrigF, const Twine &Suffix, coro::Shape &Shape, |
119 | Kind FKind) |
120 | : OrigF(OrigF), NewF(nullptr), Suffix(Suffix), Shape(Shape), FKind(FKind), |
121 | Builder(OrigF.getContext()) { |
122 | assert(Shape.ABI == coro::ABI::Switch); |
123 | } |
124 | |
125 | /// Create a cloner for a continuation lowering. |
126 | CoroCloner(Function &OrigF, const Twine &Suffix, coro::Shape &Shape, |
127 | Function *NewF, AnyCoroSuspendInst *ActiveSuspend) |
128 | : OrigF(OrigF), NewF(NewF), Suffix(Suffix), Shape(Shape), |
129 | FKind(Shape.ABI == coro::ABI::Async ? Kind::Async : Kind::Continuation), |
130 | Builder(OrigF.getContext()), ActiveSuspend(ActiveSuspend) { |
131 | assert(Shape.ABI == coro::ABI::Retcon || |
132 | Shape.ABI == coro::ABI::RetconOnce || Shape.ABI == coro::ABI::Async); |
133 | assert(NewF && "need existing function for continuation" ); |
134 | assert(ActiveSuspend && "need active suspend point for continuation" ); |
135 | } |
136 | |
137 | Function *getFunction() const { |
138 | assert(NewF != nullptr && "declaration not yet set" ); |
139 | return NewF; |
140 | } |
141 | |
142 | void create(); |
143 | |
144 | private: |
145 | bool isSwitchDestroyFunction() { |
146 | switch (FKind) { |
147 | case Kind::Async: |
148 | case Kind::Continuation: |
149 | case Kind::SwitchResume: |
150 | return false; |
151 | case Kind::SwitchUnwind: |
152 | case Kind::SwitchCleanup: |
153 | return true; |
154 | } |
155 | llvm_unreachable("Unknown CoroCloner::Kind enum" ); |
156 | } |
157 | |
158 | void replaceEntryBlock(); |
159 | Value *deriveNewFramePointer(); |
160 | void replaceRetconOrAsyncSuspendUses(); |
161 | void replaceCoroSuspends(); |
162 | void replaceCoroEnds(); |
163 | void replaceSwiftErrorOps(); |
164 | void salvageDebugInfo(); |
165 | void handleFinalSuspend(); |
166 | }; |
167 | |
168 | } // end anonymous namespace |
169 | |
170 | // FIXME: |
171 | // Lower the intrinisc in CoroEarly phase if coroutine frame doesn't escape |
172 | // and it is known that other transformations, for example, sanitizers |
173 | // won't lead to incorrect code. |
174 | static void lowerAwaitSuspend(IRBuilder<> &Builder, CoroAwaitSuspendInst *CB) { |
175 | auto Wrapper = CB->getWrapperFunction(); |
176 | auto Awaiter = CB->getAwaiter(); |
177 | auto FramePtr = CB->getFrame(); |
178 | |
179 | Builder.SetInsertPoint(CB); |
180 | |
181 | CallBase *NewCall = nullptr; |
182 | // await_suspend has only 2 parameters, awaiter and handle. |
183 | // Copy parameter attributes from the intrinsic call, but remove the last, |
184 | // because the last parameter now becomes the function that is being called. |
185 | AttributeList NewAttributes = |
186 | CB->getAttributes().removeParamAttributes(C&: CB->getContext(), ArgNo: 2); |
187 | |
188 | if (auto Invoke = dyn_cast<InvokeInst>(Val: CB)) { |
189 | auto WrapperInvoke = |
190 | Builder.CreateInvoke(Callee: Wrapper, NormalDest: Invoke->getNormalDest(), |
191 | UnwindDest: Invoke->getUnwindDest(), Args: {Awaiter, FramePtr}); |
192 | |
193 | WrapperInvoke->setCallingConv(Invoke->getCallingConv()); |
194 | std::copy(first: Invoke->bundle_op_info_begin(), last: Invoke->bundle_op_info_end(), |
195 | result: WrapperInvoke->bundle_op_info_begin()); |
196 | WrapperInvoke->setAttributes(NewAttributes); |
197 | WrapperInvoke->setDebugLoc(Invoke->getDebugLoc()); |
198 | NewCall = WrapperInvoke; |
199 | } else if (auto Call = dyn_cast<CallInst>(Val: CB)) { |
200 | auto WrapperCall = Builder.CreateCall(Callee: Wrapper, Args: {Awaiter, FramePtr}); |
201 | |
202 | WrapperCall->setAttributes(NewAttributes); |
203 | WrapperCall->setDebugLoc(Call->getDebugLoc()); |
204 | NewCall = WrapperCall; |
205 | } else { |
206 | llvm_unreachable("Unexpected coro_await_suspend invocation method" ); |
207 | } |
208 | |
209 | CB->replaceAllUsesWith(V: NewCall); |
210 | CB->eraseFromParent(); |
211 | } |
212 | |
213 | static void lowerAwaitSuspends(Function &F, coro::Shape &Shape) { |
214 | IRBuilder<> Builder(F.getContext()); |
215 | for (auto *AWS : Shape.CoroAwaitSuspends) |
216 | lowerAwaitSuspend(Builder, CB: AWS); |
217 | } |
218 | |
219 | static void maybeFreeRetconStorage(IRBuilder<> &Builder, |
220 | const coro::Shape &Shape, Value *FramePtr, |
221 | CallGraph *CG) { |
222 | assert(Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce); |
223 | if (Shape.RetconLowering.IsFrameInlineInStorage) |
224 | return; |
225 | |
226 | Shape.emitDealloc(Builder, Ptr: FramePtr, CG); |
227 | } |
228 | |
229 | /// Replace an llvm.coro.end.async. |
230 | /// Will inline the must tail call function call if there is one. |
231 | /// \returns true if cleanup of the coro.end block is needed, false otherwise. |
232 | static bool replaceCoroEndAsync(AnyCoroEndInst *End) { |
233 | IRBuilder<> Builder(End); |
234 | |
235 | auto *EndAsync = dyn_cast<CoroAsyncEndInst>(Val: End); |
236 | if (!EndAsync) { |
237 | Builder.CreateRetVoid(); |
238 | return true /*needs cleanup of coro.end block*/; |
239 | } |
240 | |
241 | auto *MustTailCallFunc = EndAsync->getMustTailCallFunction(); |
242 | if (!MustTailCallFunc) { |
243 | Builder.CreateRetVoid(); |
244 | return true /*needs cleanup of coro.end block*/; |
245 | } |
246 | |
247 | // Move the must tail call from the predecessor block into the end block. |
248 | auto *CoroEndBlock = End->getParent(); |
249 | auto *MustTailCallFuncBlock = CoroEndBlock->getSinglePredecessor(); |
250 | assert(MustTailCallFuncBlock && "Must have a single predecessor block" ); |
251 | auto It = MustTailCallFuncBlock->getTerminator()->getIterator(); |
252 | auto *MustTailCall = cast<CallInst>(Val: &*std::prev(x: It)); |
253 | CoroEndBlock->splice(ToIt: End->getIterator(), FromBB: MustTailCallFuncBlock, |
254 | FromIt: MustTailCall->getIterator()); |
255 | |
256 | // Insert the return instruction. |
257 | Builder.SetInsertPoint(End); |
258 | Builder.CreateRetVoid(); |
259 | InlineFunctionInfo FnInfo; |
260 | |
261 | // Remove the rest of the block, by splitting it into an unreachable block. |
262 | auto *BB = End->getParent(); |
263 | BB->splitBasicBlock(I: End); |
264 | BB->getTerminator()->eraseFromParent(); |
265 | |
266 | auto InlineRes = InlineFunction(CB&: *MustTailCall, IFI&: FnInfo); |
267 | assert(InlineRes.isSuccess() && "Expected inlining to succeed" ); |
268 | (void)InlineRes; |
269 | |
270 | // We have cleaned up the coro.end block above. |
271 | return false; |
272 | } |
273 | |
274 | /// Replace a non-unwind call to llvm.coro.end. |
275 | static void replaceFallthroughCoroEnd(AnyCoroEndInst *End, |
276 | const coro::Shape &Shape, Value *FramePtr, |
277 | bool InResume, CallGraph *CG) { |
278 | // Start inserting right before the coro.end. |
279 | IRBuilder<> Builder(End); |
280 | |
281 | // Create the return instruction. |
282 | switch (Shape.ABI) { |
283 | // The cloned functions in switch-lowering always return void. |
284 | case coro::ABI::Switch: |
285 | assert(!cast<CoroEndInst>(End)->hasResults() && |
286 | "switch coroutine should not return any values" ); |
287 | // coro.end doesn't immediately end the coroutine in the main function |
288 | // in this lowering, because we need to deallocate the coroutine. |
289 | if (!InResume) |
290 | return; |
291 | Builder.CreateRetVoid(); |
292 | break; |
293 | |
294 | // In async lowering this returns. |
295 | case coro::ABI::Async: { |
296 | bool CoroEndBlockNeedsCleanup = replaceCoroEndAsync(End); |
297 | if (!CoroEndBlockNeedsCleanup) |
298 | return; |
299 | break; |
300 | } |
301 | |
302 | // In unique continuation lowering, the continuations always return void. |
303 | // But we may have implicitly allocated storage. |
304 | case coro::ABI::RetconOnce: { |
305 | maybeFreeRetconStorage(Builder, Shape, FramePtr, CG); |
306 | auto *CoroEnd = cast<CoroEndInst>(Val: End); |
307 | auto *RetTy = Shape.getResumeFunctionType()->getReturnType(); |
308 | |
309 | if (!CoroEnd->hasResults()) { |
310 | assert(RetTy->isVoidTy()); |
311 | Builder.CreateRetVoid(); |
312 | break; |
313 | } |
314 | |
315 | auto *CoroResults = CoroEnd->getResults(); |
316 | unsigned NumReturns = CoroResults->numReturns(); |
317 | |
318 | if (auto *RetStructTy = dyn_cast<StructType>(Val: RetTy)) { |
319 | assert(RetStructTy->getNumElements() == NumReturns && |
320 | "numbers of returns should match resume function singature" ); |
321 | Value *ReturnValue = UndefValue::get(T: RetStructTy); |
322 | unsigned Idx = 0; |
323 | for (Value *RetValEl : CoroResults->return_values()) |
324 | ReturnValue = Builder.CreateInsertValue(Agg: ReturnValue, Val: RetValEl, Idxs: Idx++); |
325 | Builder.CreateRet(V: ReturnValue); |
326 | } else if (NumReturns == 0) { |
327 | assert(RetTy->isVoidTy()); |
328 | Builder.CreateRetVoid(); |
329 | } else { |
330 | assert(NumReturns == 1); |
331 | Builder.CreateRet(V: *CoroResults->retval_begin()); |
332 | } |
333 | CoroResults->replaceAllUsesWith( |
334 | V: ConstantTokenNone::get(Context&: CoroResults->getContext())); |
335 | CoroResults->eraseFromParent(); |
336 | break; |
337 | } |
338 | |
339 | // In non-unique continuation lowering, we signal completion by returning |
340 | // a null continuation. |
341 | case coro::ABI::Retcon: { |
342 | assert(!cast<CoroEndInst>(End)->hasResults() && |
343 | "retcon coroutine should not return any values" ); |
344 | maybeFreeRetconStorage(Builder, Shape, FramePtr, CG); |
345 | auto RetTy = Shape.getResumeFunctionType()->getReturnType(); |
346 | auto RetStructTy = dyn_cast<StructType>(Val: RetTy); |
347 | PointerType *ContinuationTy = |
348 | cast<PointerType>(Val: RetStructTy ? RetStructTy->getElementType(N: 0) : RetTy); |
349 | |
350 | Value *ReturnValue = ConstantPointerNull::get(T: ContinuationTy); |
351 | if (RetStructTy) { |
352 | ReturnValue = Builder.CreateInsertValue(Agg: UndefValue::get(T: RetStructTy), |
353 | Val: ReturnValue, Idxs: 0); |
354 | } |
355 | Builder.CreateRet(V: ReturnValue); |
356 | break; |
357 | } |
358 | } |
359 | |
360 | // Remove the rest of the block, by splitting it into an unreachable block. |
361 | auto *BB = End->getParent(); |
362 | BB->splitBasicBlock(I: End); |
363 | BB->getTerminator()->eraseFromParent(); |
364 | } |
365 | |
366 | // Mark a coroutine as done, which implies that the coroutine is finished and |
367 | // never get resumed. |
368 | // |
369 | // In resume-switched ABI, the done state is represented by storing zero in |
370 | // ResumeFnAddr. |
371 | // |
372 | // NOTE: We couldn't omit the argument `FramePtr`. It is necessary because the |
373 | // pointer to the frame in splitted function is not stored in `Shape`. |
374 | static void markCoroutineAsDone(IRBuilder<> &Builder, const coro::Shape &Shape, |
375 | Value *FramePtr) { |
376 | assert( |
377 | Shape.ABI == coro::ABI::Switch && |
378 | "markCoroutineAsDone is only supported for Switch-Resumed ABI for now." ); |
379 | auto *GepIndex = Builder.CreateStructGEP( |
380 | Ty: Shape.FrameTy, Ptr: FramePtr, Idx: coro::Shape::SwitchFieldIndex::Resume, |
381 | Name: "ResumeFn.addr" ); |
382 | auto *NullPtr = ConstantPointerNull::get(T: cast<PointerType>( |
383 | Val: Shape.FrameTy->getTypeAtIndex(N: coro::Shape::SwitchFieldIndex::Resume))); |
384 | Builder.CreateStore(Val: NullPtr, Ptr: GepIndex); |
385 | |
386 | // If the coroutine don't have unwind coro end, we could omit the store to |
387 | // the final suspend point since we could infer the coroutine is suspended |
388 | // at the final suspend point by the nullness of ResumeFnAddr. |
389 | // However, we can't skip it if the coroutine have unwind coro end. Since |
390 | // the coroutine reaches unwind coro end is considered suspended at the |
391 | // final suspend point (the ResumeFnAddr is null) but in fact the coroutine |
392 | // didn't complete yet. We need the IndexVal for the final suspend point |
393 | // to make the states clear. |
394 | if (Shape.SwitchLowering.HasUnwindCoroEnd && |
395 | Shape.SwitchLowering.HasFinalSuspend) { |
396 | assert(cast<CoroSuspendInst>(Shape.CoroSuspends.back())->isFinal() && |
397 | "The final suspend should only live in the last position of " |
398 | "CoroSuspends." ); |
399 | ConstantInt *IndexVal = Shape.getIndex(Value: Shape.CoroSuspends.size() - 1); |
400 | auto *FinalIndex = Builder.CreateStructGEP( |
401 | Ty: Shape.FrameTy, Ptr: FramePtr, Idx: Shape.getSwitchIndexField(), Name: "index.addr" ); |
402 | |
403 | Builder.CreateStore(Val: IndexVal, Ptr: FinalIndex); |
404 | } |
405 | } |
406 | |
407 | /// Replace an unwind call to llvm.coro.end. |
408 | static void replaceUnwindCoroEnd(AnyCoroEndInst *End, const coro::Shape &Shape, |
409 | Value *FramePtr, bool InResume, |
410 | CallGraph *CG) { |
411 | IRBuilder<> Builder(End); |
412 | |
413 | switch (Shape.ABI) { |
414 | // In switch-lowering, this does nothing in the main function. |
415 | case coro::ABI::Switch: { |
416 | // In C++'s specification, the coroutine should be marked as done |
417 | // if promise.unhandled_exception() throws. The frontend will |
418 | // call coro.end(true) along this path. |
419 | // |
420 | // FIXME: We should refactor this once there is other language |
421 | // which uses Switch-Resumed style other than C++. |
422 | markCoroutineAsDone(Builder, Shape, FramePtr); |
423 | if (!InResume) |
424 | return; |
425 | break; |
426 | } |
427 | // In async lowering this does nothing. |
428 | case coro::ABI::Async: |
429 | break; |
430 | // In continuation-lowering, this frees the continuation storage. |
431 | case coro::ABI::Retcon: |
432 | case coro::ABI::RetconOnce: |
433 | maybeFreeRetconStorage(Builder, Shape, FramePtr, CG); |
434 | break; |
435 | } |
436 | |
437 | // If coro.end has an associated bundle, add cleanupret instruction. |
438 | if (auto Bundle = End->getOperandBundle(ID: LLVMContext::OB_funclet)) { |
439 | auto *FromPad = cast<CleanupPadInst>(Val: Bundle->Inputs[0]); |
440 | auto *CleanupRet = Builder.CreateCleanupRet(CleanupPad: FromPad, UnwindBB: nullptr); |
441 | End->getParent()->splitBasicBlock(I: End); |
442 | CleanupRet->getParent()->getTerminator()->eraseFromParent(); |
443 | } |
444 | } |
445 | |
446 | static void replaceCoroEnd(AnyCoroEndInst *End, const coro::Shape &Shape, |
447 | Value *FramePtr, bool InResume, CallGraph *CG) { |
448 | if (End->isUnwind()) |
449 | replaceUnwindCoroEnd(End, Shape, FramePtr, InResume, CG); |
450 | else |
451 | replaceFallthroughCoroEnd(End, Shape, FramePtr, InResume, CG); |
452 | |
453 | auto &Context = End->getContext(); |
454 | End->replaceAllUsesWith(V: InResume ? ConstantInt::getTrue(Context) |
455 | : ConstantInt::getFalse(Context)); |
456 | End->eraseFromParent(); |
457 | } |
458 | |
459 | // In the resume function, we remove the last case (when coro::Shape is built, |
460 | // the final suspend point (if present) is always the last element of |
461 | // CoroSuspends array) since it is an undefined behavior to resume a coroutine |
462 | // suspended at the final suspend point. |
463 | // In the destroy function, if it isn't possible that the ResumeFnAddr is NULL |
464 | // and the coroutine doesn't suspend at the final suspend point actually (this |
465 | // is possible since the coroutine is considered suspended at the final suspend |
466 | // point if promise.unhandled_exception() exits via an exception), we can |
467 | // remove the last case. |
468 | void CoroCloner::handleFinalSuspend() { |
469 | assert(Shape.ABI == coro::ABI::Switch && |
470 | Shape.SwitchLowering.HasFinalSuspend); |
471 | |
472 | if (isSwitchDestroyFunction() && Shape.SwitchLowering.HasUnwindCoroEnd) |
473 | return; |
474 | |
475 | auto *Switch = cast<SwitchInst>(Val&: VMap[Shape.SwitchLowering.ResumeSwitch]); |
476 | auto FinalCaseIt = std::prev(x: Switch->case_end()); |
477 | BasicBlock *ResumeBB = FinalCaseIt->getCaseSuccessor(); |
478 | Switch->removeCase(I: FinalCaseIt); |
479 | if (isSwitchDestroyFunction()) { |
480 | BasicBlock *OldSwitchBB = Switch->getParent(); |
481 | auto *NewSwitchBB = OldSwitchBB->splitBasicBlock(I: Switch, BBName: "Switch" ); |
482 | Builder.SetInsertPoint(OldSwitchBB->getTerminator()); |
483 | |
484 | if (NewF->isCoroOnlyDestroyWhenComplete()) { |
485 | // When the coroutine can only be destroyed when complete, we don't need |
486 | // to generate code for other cases. |
487 | Builder.CreateBr(Dest: ResumeBB); |
488 | } else { |
489 | auto *GepIndex = Builder.CreateStructGEP( |
490 | Ty: Shape.FrameTy, Ptr: NewFramePtr, Idx: coro::Shape::SwitchFieldIndex::Resume, |
491 | Name: "ResumeFn.addr" ); |
492 | auto *Load = |
493 | Builder.CreateLoad(Ty: Shape.getSwitchResumePointerType(), Ptr: GepIndex); |
494 | auto *Cond = Builder.CreateIsNull(Arg: Load); |
495 | Builder.CreateCondBr(Cond, True: ResumeBB, False: NewSwitchBB); |
496 | } |
497 | OldSwitchBB->getTerminator()->eraseFromParent(); |
498 | } |
499 | } |
500 | |
501 | static FunctionType * |
502 | getFunctionTypeFromAsyncSuspend(AnyCoroSuspendInst *Suspend) { |
503 | auto *AsyncSuspend = cast<CoroSuspendAsyncInst>(Val: Suspend); |
504 | auto *StructTy = cast<StructType>(Val: AsyncSuspend->getType()); |
505 | auto &Context = Suspend->getParent()->getParent()->getContext(); |
506 | auto *VoidTy = Type::getVoidTy(C&: Context); |
507 | return FunctionType::get(Result: VoidTy, Params: StructTy->elements(), isVarArg: false); |
508 | } |
509 | |
510 | static Function *createCloneDeclaration(Function &OrigF, coro::Shape &Shape, |
511 | const Twine &Suffix, |
512 | Module::iterator InsertBefore, |
513 | AnyCoroSuspendInst *ActiveSuspend) { |
514 | Module *M = OrigF.getParent(); |
515 | auto *FnTy = (Shape.ABI != coro::ABI::Async) |
516 | ? Shape.getResumeFunctionType() |
517 | : getFunctionTypeFromAsyncSuspend(Suspend: ActiveSuspend); |
518 | |
519 | Function *NewF = |
520 | Function::Create(Ty: FnTy, Linkage: GlobalValue::LinkageTypes::InternalLinkage, |
521 | N: OrigF.getName() + Suffix); |
522 | |
523 | M->getFunctionList().insert(where: InsertBefore, New: NewF); |
524 | |
525 | return NewF; |
526 | } |
527 | |
528 | /// Replace uses of the active llvm.coro.suspend.retcon/async call with the |
529 | /// arguments to the continuation function. |
530 | /// |
531 | /// This assumes that the builder has a meaningful insertion point. |
532 | void CoroCloner::replaceRetconOrAsyncSuspendUses() { |
533 | assert(Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce || |
534 | Shape.ABI == coro::ABI::Async); |
535 | |
536 | auto NewS = VMap[ActiveSuspend]; |
537 | if (NewS->use_empty()) |
538 | return; |
539 | |
540 | // Copy out all the continuation arguments after the buffer pointer into |
541 | // an easily-indexed data structure for convenience. |
542 | SmallVector<Value *, 8> Args; |
543 | // The async ABI includes all arguments -- including the first argument. |
544 | bool IsAsyncABI = Shape.ABI == coro::ABI::Async; |
545 | for (auto I = IsAsyncABI ? NewF->arg_begin() : std::next(x: NewF->arg_begin()), |
546 | E = NewF->arg_end(); |
547 | I != E; ++I) |
548 | Args.push_back(Elt: &*I); |
549 | |
550 | // If the suspend returns a single scalar value, we can just do a simple |
551 | // replacement. |
552 | if (!isa<StructType>(Val: NewS->getType())) { |
553 | assert(Args.size() == 1); |
554 | NewS->replaceAllUsesWith(V: Args.front()); |
555 | return; |
556 | } |
557 | |
558 | // Try to peephole extracts of an aggregate return. |
559 | for (Use &U : llvm::make_early_inc_range(Range: NewS->uses())) { |
560 | auto *EVI = dyn_cast<ExtractValueInst>(Val: U.getUser()); |
561 | if (!EVI || EVI->getNumIndices() != 1) |
562 | continue; |
563 | |
564 | EVI->replaceAllUsesWith(V: Args[EVI->getIndices().front()]); |
565 | EVI->eraseFromParent(); |
566 | } |
567 | |
568 | // If we have no remaining uses, we're done. |
569 | if (NewS->use_empty()) |
570 | return; |
571 | |
572 | // Otherwise, we need to create an aggregate. |
573 | Value *Agg = PoisonValue::get(T: NewS->getType()); |
574 | for (size_t I = 0, E = Args.size(); I != E; ++I) |
575 | Agg = Builder.CreateInsertValue(Agg, Val: Args[I], Idxs: I); |
576 | |
577 | NewS->replaceAllUsesWith(V: Agg); |
578 | } |
579 | |
580 | void CoroCloner::replaceCoroSuspends() { |
581 | Value *SuspendResult; |
582 | |
583 | switch (Shape.ABI) { |
584 | // In switch lowering, replace coro.suspend with the appropriate value |
585 | // for the type of function we're extracting. |
586 | // Replacing coro.suspend with (0) will result in control flow proceeding to |
587 | // a resume label associated with a suspend point, replacing it with (1) will |
588 | // result in control flow proceeding to a cleanup label associated with this |
589 | // suspend point. |
590 | case coro::ABI::Switch: |
591 | SuspendResult = Builder.getInt8(C: isSwitchDestroyFunction() ? 1 : 0); |
592 | break; |
593 | |
594 | // In async lowering there are no uses of the result. |
595 | case coro::ABI::Async: |
596 | return; |
597 | |
598 | // In returned-continuation lowering, the arguments from earlier |
599 | // continuations are theoretically arbitrary, and they should have been |
600 | // spilled. |
601 | case coro::ABI::RetconOnce: |
602 | case coro::ABI::Retcon: |
603 | return; |
604 | } |
605 | |
606 | for (AnyCoroSuspendInst *CS : Shape.CoroSuspends) { |
607 | // The active suspend was handled earlier. |
608 | if (CS == ActiveSuspend) |
609 | continue; |
610 | |
611 | auto *MappedCS = cast<AnyCoroSuspendInst>(Val&: VMap[CS]); |
612 | MappedCS->replaceAllUsesWith(V: SuspendResult); |
613 | MappedCS->eraseFromParent(); |
614 | } |
615 | } |
616 | |
617 | void CoroCloner::replaceCoroEnds() { |
618 | for (AnyCoroEndInst *CE : Shape.CoroEnds) { |
619 | // We use a null call graph because there's no call graph node for |
620 | // the cloned function yet. We'll just be rebuilding that later. |
621 | auto *NewCE = cast<AnyCoroEndInst>(Val&: VMap[CE]); |
622 | replaceCoroEnd(End: NewCE, Shape, FramePtr: NewFramePtr, /*in resume*/ InResume: true, CG: nullptr); |
623 | } |
624 | } |
625 | |
626 | static void replaceSwiftErrorOps(Function &F, coro::Shape &Shape, |
627 | ValueToValueMapTy *VMap) { |
628 | if (Shape.ABI == coro::ABI::Async && Shape.CoroSuspends.empty()) |
629 | return; |
630 | Value *CachedSlot = nullptr; |
631 | auto getSwiftErrorSlot = [&](Type *ValueTy) -> Value * { |
632 | if (CachedSlot) |
633 | return CachedSlot; |
634 | |
635 | // Check if the function has a swifterror argument. |
636 | for (auto &Arg : F.args()) { |
637 | if (Arg.isSwiftError()) { |
638 | CachedSlot = &Arg; |
639 | return &Arg; |
640 | } |
641 | } |
642 | |
643 | // Create a swifterror alloca. |
644 | IRBuilder<> Builder(F.getEntryBlock().getFirstNonPHIOrDbg()); |
645 | auto Alloca = Builder.CreateAlloca(Ty: ValueTy); |
646 | Alloca->setSwiftError(true); |
647 | |
648 | CachedSlot = Alloca; |
649 | return Alloca; |
650 | }; |
651 | |
652 | for (CallInst *Op : Shape.SwiftErrorOps) { |
653 | auto MappedOp = VMap ? cast<CallInst>(Val&: (*VMap)[Op]) : Op; |
654 | IRBuilder<> Builder(MappedOp); |
655 | |
656 | // If there are no arguments, this is a 'get' operation. |
657 | Value *MappedResult; |
658 | if (Op->arg_empty()) { |
659 | auto ValueTy = Op->getType(); |
660 | auto Slot = getSwiftErrorSlot(ValueTy); |
661 | MappedResult = Builder.CreateLoad(Ty: ValueTy, Ptr: Slot); |
662 | } else { |
663 | assert(Op->arg_size() == 1); |
664 | auto Value = MappedOp->getArgOperand(i: 0); |
665 | auto ValueTy = Value->getType(); |
666 | auto Slot = getSwiftErrorSlot(ValueTy); |
667 | Builder.CreateStore(Val: Value, Ptr: Slot); |
668 | MappedResult = Slot; |
669 | } |
670 | |
671 | MappedOp->replaceAllUsesWith(V: MappedResult); |
672 | MappedOp->eraseFromParent(); |
673 | } |
674 | |
675 | // If we're updating the original function, we've invalidated SwiftErrorOps. |
676 | if (VMap == nullptr) { |
677 | Shape.SwiftErrorOps.clear(); |
678 | } |
679 | } |
680 | |
681 | /// Returns all DbgVariableIntrinsic in F. |
682 | static std::pair<SmallVector<DbgVariableIntrinsic *, 8>, |
683 | SmallVector<DbgVariableRecord *>> |
684 | collectDbgVariableIntrinsics(Function &F) { |
685 | SmallVector<DbgVariableIntrinsic *, 8> Intrinsics; |
686 | SmallVector<DbgVariableRecord *> DbgVariableRecords; |
687 | for (auto &I : instructions(F)) { |
688 | for (DbgVariableRecord &DVR : filterDbgVars(R: I.getDbgRecordRange())) |
689 | DbgVariableRecords.push_back(Elt: &DVR); |
690 | if (auto *DVI = dyn_cast<DbgVariableIntrinsic>(Val: &I)) |
691 | Intrinsics.push_back(Elt: DVI); |
692 | } |
693 | return {Intrinsics, DbgVariableRecords}; |
694 | } |
695 | |
696 | void CoroCloner::replaceSwiftErrorOps() { |
697 | ::replaceSwiftErrorOps(F&: *NewF, Shape, VMap: &VMap); |
698 | } |
699 | |
700 | void CoroCloner::salvageDebugInfo() { |
701 | auto [Worklist, DbgVariableRecords] = collectDbgVariableIntrinsics(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 = |
706 | llvm::Triple(OrigF.getParent()->getTargetTriple()).isArch64Bit(); |
707 | for (DbgVariableIntrinsic *DVI : Worklist) |
708 | coro::salvageDebugInfo(ArgToAllocaMap, DVI&: *DVI, OptimizeFrame: Shape.OptimizeFrame, |
709 | IsEntryPoint: UseEntryValue); |
710 | for (DbgVariableRecord *DVR : DbgVariableRecords) |
711 | coro::salvageDebugInfo(ArgToAllocaMap, DVR&: *DVR, OptimizeFrame: Shape.OptimizeFrame, |
712 | UseEntryValue); |
713 | |
714 | // Remove all salvaged dbg.declare intrinsics that became |
715 | // either unreachable or stale due to the CoroSplit transformation. |
716 | DominatorTree DomTree(*NewF); |
717 | auto IsUnreachableBlock = [&](BasicBlock *BB) { |
718 | return !isPotentiallyReachable(From: &NewF->getEntryBlock(), To: BB, ExclusionSet: nullptr, |
719 | DT: &DomTree); |
720 | }; |
721 | auto RemoveOne = [&](auto *DVI) { |
722 | if (IsUnreachableBlock(DVI->getParent())) |
723 | DVI->eraseFromParent(); |
724 | else if (isa_and_nonnull<AllocaInst>(DVI->getVariableLocationOp(0))) { |
725 | // Count all non-debuginfo uses in reachable blocks. |
726 | unsigned Uses = 0; |
727 | for (auto *User : DVI->getVariableLocationOp(0)->users()) |
728 | if (auto *I = dyn_cast<Instruction>(User)) |
729 | if (!isa<AllocaInst>(I) && !IsUnreachableBlock(I->getParent())) |
730 | ++Uses; |
731 | if (!Uses) |
732 | DVI->eraseFromParent(); |
733 | } |
734 | }; |
735 | for_each(Range&: Worklist, F: RemoveOne); |
736 | for_each(Range&: DbgVariableRecords, F: RemoveOne); |
737 | } |
738 | |
739 | void CoroCloner::replaceEntryBlock() { |
740 | // In the original function, the AllocaSpillBlock is a block immediately |
741 | // following the allocation of the frame object which defines GEPs for |
742 | // all the allocas that have been moved into the frame, and it ends by |
743 | // branching to the original beginning of the coroutine. Make this |
744 | // the entry block of the cloned function. |
745 | auto *Entry = cast<BasicBlock>(Val&: VMap[Shape.AllocaSpillBlock]); |
746 | auto *OldEntry = &NewF->getEntryBlock(); |
747 | Entry->setName("entry" + Suffix); |
748 | Entry->moveBefore(MovePos: OldEntry); |
749 | Entry->getTerminator()->eraseFromParent(); |
750 | |
751 | // Clear all predecessors of the new entry block. There should be |
752 | // exactly one predecessor, which we created when splitting out |
753 | // AllocaSpillBlock to begin with. |
754 | assert(Entry->hasOneUse()); |
755 | auto BranchToEntry = cast<BranchInst>(Val: Entry->user_back()); |
756 | assert(BranchToEntry->isUnconditional()); |
757 | Builder.SetInsertPoint(BranchToEntry); |
758 | Builder.CreateUnreachable(); |
759 | BranchToEntry->eraseFromParent(); |
760 | |
761 | // Branch from the entry to the appropriate place. |
762 | Builder.SetInsertPoint(Entry); |
763 | switch (Shape.ABI) { |
764 | case coro::ABI::Switch: { |
765 | // In switch-lowering, we built a resume-entry block in the original |
766 | // function. Make the entry block branch to this. |
767 | auto *SwitchBB = |
768 | cast<BasicBlock>(Val&: VMap[Shape.SwitchLowering.ResumeEntryBlock]); |
769 | Builder.CreateBr(Dest: SwitchBB); |
770 | break; |
771 | } |
772 | case coro::ABI::Async: |
773 | case coro::ABI::Retcon: |
774 | case coro::ABI::RetconOnce: { |
775 | // In continuation ABIs, we want to branch to immediately after the |
776 | // active suspend point. Earlier phases will have put the suspend in its |
777 | // own basic block, so just thread our jump directly to its successor. |
778 | assert((Shape.ABI == coro::ABI::Async && |
779 | isa<CoroSuspendAsyncInst>(ActiveSuspend)) || |
780 | ((Shape.ABI == coro::ABI::Retcon || |
781 | Shape.ABI == coro::ABI::RetconOnce) && |
782 | isa<CoroSuspendRetconInst>(ActiveSuspend))); |
783 | auto *MappedCS = cast<AnyCoroSuspendInst>(Val&: VMap[ActiveSuspend]); |
784 | auto Branch = cast<BranchInst>(Val: MappedCS->getNextNode()); |
785 | assert(Branch->isUnconditional()); |
786 | Builder.CreateBr(Dest: Branch->getSuccessor(i: 0)); |
787 | break; |
788 | } |
789 | } |
790 | |
791 | // Any static alloca that's still being used but not reachable from the new |
792 | // entry needs to be moved to the new entry. |
793 | Function *F = OldEntry->getParent(); |
794 | DominatorTree DT{*F}; |
795 | for (Instruction &I : llvm::make_early_inc_range(Range: instructions(F))) { |
796 | auto *Alloca = dyn_cast<AllocaInst>(Val: &I); |
797 | if (!Alloca || I.use_empty()) |
798 | continue; |
799 | if (DT.isReachableFromEntry(A: I.getParent()) || |
800 | !isa<ConstantInt>(Val: Alloca->getArraySize())) |
801 | continue; |
802 | I.moveBefore(BB&: *Entry, I: Entry->getFirstInsertionPt()); |
803 | } |
804 | } |
805 | |
806 | /// Derive the value of the new frame pointer. |
807 | Value *CoroCloner::deriveNewFramePointer() { |
808 | // Builder should be inserting to the front of the new entry block. |
809 | |
810 | switch (Shape.ABI) { |
811 | // In switch-lowering, the argument is the frame pointer. |
812 | case coro::ABI::Switch: |
813 | return &*NewF->arg_begin(); |
814 | // In async-lowering, one of the arguments is an async context as determined |
815 | // by the `llvm.coro.id.async` intrinsic. We can retrieve the async context of |
816 | // the resume function from the async context projection function associated |
817 | // with the active suspend. The frame is located as a tail to the async |
818 | // context header. |
819 | case coro::ABI::Async: { |
820 | auto *ActiveAsyncSuspend = cast<CoroSuspendAsyncInst>(Val: ActiveSuspend); |
821 | auto ContextIdx = ActiveAsyncSuspend->getStorageArgumentIndex() & 0xff; |
822 | auto *CalleeContext = NewF->getArg(i: ContextIdx); |
823 | auto *ProjectionFunc = |
824 | ActiveAsyncSuspend->getAsyncContextProjectionFunction(); |
825 | auto DbgLoc = |
826 | cast<CoroSuspendAsyncInst>(Val&: VMap[ActiveSuspend])->getDebugLoc(); |
827 | // Calling i8* (i8*) |
828 | auto *CallerContext = Builder.CreateCall(FTy: ProjectionFunc->getFunctionType(), |
829 | Callee: ProjectionFunc, Args: CalleeContext); |
830 | CallerContext->setCallingConv(ProjectionFunc->getCallingConv()); |
831 | CallerContext->setDebugLoc(DbgLoc); |
832 | // The frame is located after the async_context header. |
833 | auto &Context = Builder.getContext(); |
834 | auto *FramePtrAddr = Builder.CreateConstInBoundsGEP1_32( |
835 | Ty: Type::getInt8Ty(C&: Context), Ptr: CallerContext, |
836 | Idx0: Shape.AsyncLowering.FrameOffset, Name: "async.ctx.frameptr" ); |
837 | // Inline the projection function. |
838 | InlineFunctionInfo InlineInfo; |
839 | auto InlineRes = InlineFunction(CB&: *CallerContext, IFI&: InlineInfo); |
840 | assert(InlineRes.isSuccess()); |
841 | (void)InlineRes; |
842 | return FramePtrAddr; |
843 | } |
844 | // In continuation-lowering, the argument is the opaque storage. |
845 | case coro::ABI::Retcon: |
846 | case coro::ABI::RetconOnce: { |
847 | Argument *NewStorage = &*NewF->arg_begin(); |
848 | auto FramePtrTy = PointerType::getUnqual(C&: Shape.FrameTy->getContext()); |
849 | |
850 | // If the storage is inline, just bitcast to the storage to the frame type. |
851 | if (Shape.RetconLowering.IsFrameInlineInStorage) |
852 | return NewStorage; |
853 | |
854 | // Otherwise, load the real frame from the opaque storage. |
855 | return Builder.CreateLoad(Ty: FramePtrTy, Ptr: NewStorage); |
856 | } |
857 | } |
858 | llvm_unreachable("bad ABI" ); |
859 | } |
860 | |
861 | static void addFramePointerAttrs(AttributeList &Attrs, LLVMContext &Context, |
862 | unsigned ParamIndex, uint64_t Size, |
863 | Align Alignment, bool NoAlias) { |
864 | AttrBuilder ParamAttrs(Context); |
865 | ParamAttrs.addAttribute(Attribute::NonNull); |
866 | ParamAttrs.addAttribute(Attribute::NoUndef); |
867 | |
868 | if (NoAlias) |
869 | ParamAttrs.addAttribute(Attribute::NoAlias); |
870 | |
871 | ParamAttrs.addAlignmentAttr(Align: Alignment); |
872 | ParamAttrs.addDereferenceableAttr(Bytes: Size); |
873 | Attrs = Attrs.addParamAttributes(C&: Context, ArgNo: ParamIndex, B: ParamAttrs); |
874 | } |
875 | |
876 | static void addAsyncContextAttrs(AttributeList &Attrs, LLVMContext &Context, |
877 | unsigned ParamIndex) { |
878 | AttrBuilder ParamAttrs(Context); |
879 | ParamAttrs.addAttribute(Attribute::SwiftAsync); |
880 | Attrs = Attrs.addParamAttributes(C&: Context, ArgNo: ParamIndex, B: ParamAttrs); |
881 | } |
882 | |
883 | static void addSwiftSelfAttrs(AttributeList &Attrs, LLVMContext &Context, |
884 | unsigned ParamIndex) { |
885 | AttrBuilder ParamAttrs(Context); |
886 | ParamAttrs.addAttribute(Attribute::SwiftSelf); |
887 | Attrs = Attrs.addParamAttributes(C&: Context, ArgNo: ParamIndex, B: ParamAttrs); |
888 | } |
889 | |
890 | /// Clone the body of the original function into a resume function of |
891 | /// some sort. |
892 | void CoroCloner::create() { |
893 | // Create the new function if we don't already have one. |
894 | if (!NewF) { |
895 | NewF = createCloneDeclaration(OrigF, Shape, Suffix, |
896 | InsertBefore: OrigF.getParent()->end(), ActiveSuspend); |
897 | } |
898 | |
899 | // Replace all args with dummy instructions. If an argument is the old frame |
900 | // pointer, the dummy will be replaced by the new frame pointer once it is |
901 | // computed below. Uses of all other arguments should have already been |
902 | // rewritten by buildCoroutineFrame() to use loads/stores on the coroutine |
903 | // frame. |
904 | SmallVector<Instruction *> DummyArgs; |
905 | for (Argument &A : OrigF.args()) { |
906 | DummyArgs.push_back(Elt: new FreezeInst(PoisonValue::get(T: A.getType()))); |
907 | VMap[&A] = DummyArgs.back(); |
908 | } |
909 | |
910 | SmallVector<ReturnInst *, 4> Returns; |
911 | |
912 | // Ignore attempts to change certain attributes of the function. |
913 | // TODO: maybe there should be a way to suppress this during cloning? |
914 | auto savedVisibility = NewF->getVisibility(); |
915 | auto savedUnnamedAddr = NewF->getUnnamedAddr(); |
916 | auto savedDLLStorageClass = NewF->getDLLStorageClass(); |
917 | |
918 | // NewF's linkage (which CloneFunctionInto does *not* change) might not |
919 | // be compatible with the visibility of OrigF (which it *does* change), |
920 | // so protect against that. |
921 | auto savedLinkage = NewF->getLinkage(); |
922 | NewF->setLinkage(llvm::GlobalValue::ExternalLinkage); |
923 | |
924 | CloneFunctionInto(NewFunc: NewF, OldFunc: &OrigF, VMap, |
925 | Changes: CloneFunctionChangeType::LocalChangesOnly, Returns); |
926 | |
927 | auto &Context = NewF->getContext(); |
928 | |
929 | // For async functions / continuations, adjust the scope line of the |
930 | // clone to the line number of the suspend point. However, only |
931 | // adjust the scope line when the files are the same. This ensures |
932 | // line number and file name belong together. The scope line is |
933 | // associated with all pre-prologue instructions. This avoids a jump |
934 | // in the linetable from the function declaration to the suspend point. |
935 | if (DISubprogram *SP = NewF->getSubprogram()) { |
936 | assert(SP != OrigF.getSubprogram() && SP->isDistinct()); |
937 | if (ActiveSuspend) |
938 | if (auto DL = ActiveSuspend->getDebugLoc()) |
939 | if (SP->getFile() == DL->getFile()) |
940 | SP->setScopeLine(DL->getLine()); |
941 | // Update the linkage name to reflect the modified symbol name. It |
942 | // is necessary to update the linkage name in Swift, since the |
943 | // mangling changes for resume functions. It might also be the |
944 | // right thing to do in C++, but due to a limitation in LLVM's |
945 | // AsmPrinter we can only do this if the function doesn't have an |
946 | // abstract specification, since the DWARF backend expects the |
947 | // abstract specification to contain the linkage name and asserts |
948 | // that they are identical. |
949 | if (SP->getUnit() && |
950 | SP->getUnit()->getSourceLanguage() == dwarf::DW_LANG_Swift) { |
951 | SP->replaceLinkageName(LN: MDString::get(Context, Str: NewF->getName())); |
952 | if (auto *Decl = SP->getDeclaration()) { |
953 | auto *NewDecl = DISubprogram::get( |
954 | Context&: Decl->getContext(), Scope: Decl->getScope(), Name: Decl->getName(), |
955 | LinkageName: NewF->getName(), File: Decl->getFile(), Line: Decl->getLine(), Type: Decl->getType(), |
956 | ScopeLine: Decl->getScopeLine(), ContainingType: Decl->getContainingType(), |
957 | VirtualIndex: Decl->getVirtualIndex(), ThisAdjustment: Decl->getThisAdjustment(), |
958 | Flags: Decl->getFlags(), SPFlags: Decl->getSPFlags(), Unit: Decl->getUnit(), |
959 | TemplateParams: Decl->getTemplateParams(), Declaration: nullptr, RetainedNodes: Decl->getRetainedNodes(), |
960 | ThrownTypes: Decl->getThrownTypes(), Annotations: Decl->getAnnotations(), |
961 | TargetFuncName: Decl->getTargetFuncName()); |
962 | SP->replaceDeclaration(Decl: NewDecl); |
963 | } |
964 | } |
965 | } |
966 | |
967 | NewF->setLinkage(savedLinkage); |
968 | NewF->setVisibility(savedVisibility); |
969 | NewF->setUnnamedAddr(savedUnnamedAddr); |
970 | NewF->setDLLStorageClass(savedDLLStorageClass); |
971 | // The function sanitizer metadata needs to match the signature of the |
972 | // function it is being attached to. However this does not hold for split |
973 | // functions here. Thus remove the metadata for split functions. |
974 | if (Shape.ABI == coro::ABI::Switch && |
975 | NewF->hasMetadata(KindID: LLVMContext::MD_func_sanitize)) |
976 | NewF->eraseMetadata(KindID: LLVMContext::MD_func_sanitize); |
977 | |
978 | // Replace the attributes of the new function: |
979 | auto OrigAttrs = NewF->getAttributes(); |
980 | auto NewAttrs = AttributeList(); |
981 | |
982 | switch (Shape.ABI) { |
983 | case coro::ABI::Switch: |
984 | // Bootstrap attributes by copying function attributes from the |
985 | // original function. This should include optimization settings and so on. |
986 | NewAttrs = NewAttrs.addFnAttributes( |
987 | C&: Context, B: AttrBuilder(Context, OrigAttrs.getFnAttrs())); |
988 | |
989 | addFramePointerAttrs(Attrs&: NewAttrs, Context, ParamIndex: 0, Size: Shape.FrameSize, |
990 | Alignment: Shape.FrameAlign, /*NoAlias=*/false); |
991 | break; |
992 | case coro::ABI::Async: { |
993 | auto *ActiveAsyncSuspend = cast<CoroSuspendAsyncInst>(Val: ActiveSuspend); |
994 | if (OrigF.hasParamAttribute(ArgNo: Shape.AsyncLowering.ContextArgNo, |
995 | Attribute::Kind: SwiftAsync)) { |
996 | uint32_t ArgAttributeIndices = |
997 | ActiveAsyncSuspend->getStorageArgumentIndex(); |
998 | auto ContextArgIndex = ArgAttributeIndices & 0xff; |
999 | addAsyncContextAttrs(Attrs&: NewAttrs, Context, ParamIndex: ContextArgIndex); |
1000 | |
1001 | // `swiftasync` must preceed `swiftself` so 0 is not a valid index for |
1002 | // `swiftself`. |
1003 | auto SwiftSelfIndex = ArgAttributeIndices >> 8; |
1004 | if (SwiftSelfIndex) |
1005 | addSwiftSelfAttrs(Attrs&: NewAttrs, Context, ParamIndex: SwiftSelfIndex); |
1006 | } |
1007 | |
1008 | // Transfer the original function's attributes. |
1009 | auto FnAttrs = OrigF.getAttributes().getFnAttrs(); |
1010 | NewAttrs = NewAttrs.addFnAttributes(C&: Context, B: AttrBuilder(Context, FnAttrs)); |
1011 | break; |
1012 | } |
1013 | case coro::ABI::Retcon: |
1014 | case coro::ABI::RetconOnce: |
1015 | // If we have a continuation prototype, just use its attributes, |
1016 | // full-stop. |
1017 | NewAttrs = Shape.RetconLowering.ResumePrototype->getAttributes(); |
1018 | |
1019 | /// FIXME: Is it really good to add the NoAlias attribute? |
1020 | addFramePointerAttrs(Attrs&: NewAttrs, Context, ParamIndex: 0, |
1021 | Size: Shape.getRetconCoroId()->getStorageSize(), |
1022 | Alignment: Shape.getRetconCoroId()->getStorageAlignment(), |
1023 | /*NoAlias=*/true); |
1024 | |
1025 | break; |
1026 | } |
1027 | |
1028 | switch (Shape.ABI) { |
1029 | // In these ABIs, the cloned functions always return 'void', and the |
1030 | // existing return sites are meaningless. Note that for unique |
1031 | // continuations, this includes the returns associated with suspends; |
1032 | // this is fine because we can't suspend twice. |
1033 | case coro::ABI::Switch: |
1034 | case coro::ABI::RetconOnce: |
1035 | // Remove old returns. |
1036 | for (ReturnInst *Return : Returns) |
1037 | changeToUnreachable(I: Return); |
1038 | break; |
1039 | |
1040 | // With multi-suspend continuations, we'll already have eliminated the |
1041 | // original returns and inserted returns before all the suspend points, |
1042 | // so we want to leave any returns in place. |
1043 | case coro::ABI::Retcon: |
1044 | break; |
1045 | // Async lowering will insert musttail call functions at all suspend points |
1046 | // followed by a return. |
1047 | // Don't change returns to unreachable because that will trip up the verifier. |
1048 | // These returns should be unreachable from the clone. |
1049 | case coro::ABI::Async: |
1050 | break; |
1051 | } |
1052 | |
1053 | NewF->setAttributes(NewAttrs); |
1054 | NewF->setCallingConv(Shape.getResumeFunctionCC()); |
1055 | |
1056 | // Set up the new entry block. |
1057 | replaceEntryBlock(); |
1058 | |
1059 | Builder.SetInsertPoint(&NewF->getEntryBlock().front()); |
1060 | NewFramePtr = deriveNewFramePointer(); |
1061 | |
1062 | // Remap frame pointer. |
1063 | Value *OldFramePtr = VMap[Shape.FramePtr]; |
1064 | NewFramePtr->takeName(V: OldFramePtr); |
1065 | OldFramePtr->replaceAllUsesWith(V: NewFramePtr); |
1066 | |
1067 | // Remap vFrame pointer. |
1068 | auto *NewVFrame = Builder.CreateBitCast( |
1069 | V: NewFramePtr, DestTy: PointerType::getUnqual(C&: Builder.getContext()), Name: "vFrame" ); |
1070 | Value *OldVFrame = cast<Value>(Val&: VMap[Shape.CoroBegin]); |
1071 | if (OldVFrame != NewVFrame) |
1072 | OldVFrame->replaceAllUsesWith(V: NewVFrame); |
1073 | |
1074 | // All uses of the arguments should have been resolved by this point, |
1075 | // so we can safely remove the dummy values. |
1076 | for (Instruction *DummyArg : DummyArgs) { |
1077 | DummyArg->replaceAllUsesWith(V: PoisonValue::get(T: DummyArg->getType())); |
1078 | DummyArg->deleteValue(); |
1079 | } |
1080 | |
1081 | switch (Shape.ABI) { |
1082 | case coro::ABI::Switch: |
1083 | // Rewrite final suspend handling as it is not done via switch (allows to |
1084 | // remove final case from the switch, since it is undefined behavior to |
1085 | // resume the coroutine suspended at the final suspend point. |
1086 | if (Shape.SwitchLowering.HasFinalSuspend) |
1087 | handleFinalSuspend(); |
1088 | break; |
1089 | case coro::ABI::Async: |
1090 | case coro::ABI::Retcon: |
1091 | case coro::ABI::RetconOnce: |
1092 | // Replace uses of the active suspend with the corresponding |
1093 | // continuation-function arguments. |
1094 | assert(ActiveSuspend != nullptr && |
1095 | "no active suspend when lowering a continuation-style coroutine" ); |
1096 | replaceRetconOrAsyncSuspendUses(); |
1097 | break; |
1098 | } |
1099 | |
1100 | // Handle suspends. |
1101 | replaceCoroSuspends(); |
1102 | |
1103 | // Handle swifterror. |
1104 | replaceSwiftErrorOps(); |
1105 | |
1106 | // Remove coro.end intrinsics. |
1107 | replaceCoroEnds(); |
1108 | |
1109 | // Salvage debug info that points into the coroutine frame. |
1110 | salvageDebugInfo(); |
1111 | |
1112 | // Eliminate coro.free from the clones, replacing it with 'null' in cleanup, |
1113 | // to suppress deallocation code. |
1114 | if (Shape.ABI == coro::ABI::Switch) |
1115 | coro::replaceCoroFree(CoroId: cast<CoroIdInst>(Val&: VMap[Shape.CoroBegin->getId()]), |
1116 | /*Elide=*/FKind == CoroCloner::Kind::SwitchCleanup); |
1117 | } |
1118 | |
1119 | static void updateAsyncFuncPointerContextSize(coro::Shape &Shape) { |
1120 | assert(Shape.ABI == coro::ABI::Async); |
1121 | |
1122 | auto *FuncPtrStruct = cast<ConstantStruct>( |
1123 | Val: Shape.AsyncLowering.AsyncFuncPointer->getInitializer()); |
1124 | auto *OrigRelativeFunOffset = FuncPtrStruct->getOperand(i_nocapture: 0); |
1125 | auto *OrigContextSize = FuncPtrStruct->getOperand(i_nocapture: 1); |
1126 | auto *NewContextSize = ConstantInt::get(Ty: OrigContextSize->getType(), |
1127 | V: Shape.AsyncLowering.ContextSize); |
1128 | auto *NewFuncPtrStruct = ConstantStruct::get( |
1129 | T: FuncPtrStruct->getType(), Vs: OrigRelativeFunOffset, Vs: NewContextSize); |
1130 | |
1131 | Shape.AsyncLowering.AsyncFuncPointer->setInitializer(NewFuncPtrStruct); |
1132 | } |
1133 | |
1134 | static void replaceFrameSizeAndAlignment(coro::Shape &Shape) { |
1135 | if (Shape.ABI == coro::ABI::Async) |
1136 | updateAsyncFuncPointerContextSize(Shape); |
1137 | |
1138 | for (CoroAlignInst *CA : Shape.CoroAligns) { |
1139 | CA->replaceAllUsesWith( |
1140 | V: ConstantInt::get(Ty: CA->getType(), V: Shape.FrameAlign.value())); |
1141 | CA->eraseFromParent(); |
1142 | } |
1143 | |
1144 | if (Shape.CoroSizes.empty()) |
1145 | return; |
1146 | |
1147 | // In the same function all coro.sizes should have the same result type. |
1148 | auto *SizeIntrin = Shape.CoroSizes.back(); |
1149 | Module *M = SizeIntrin->getModule(); |
1150 | const DataLayout &DL = M->getDataLayout(); |
1151 | auto Size = DL.getTypeAllocSize(Ty: Shape.FrameTy); |
1152 | auto *SizeConstant = ConstantInt::get(Ty: SizeIntrin->getType(), V: Size); |
1153 | |
1154 | for (CoroSizeInst *CS : Shape.CoroSizes) { |
1155 | CS->replaceAllUsesWith(V: SizeConstant); |
1156 | CS->eraseFromParent(); |
1157 | } |
1158 | } |
1159 | |
1160 | static void postSplitCleanup(Function &F) { |
1161 | removeUnreachableBlocks(F); |
1162 | |
1163 | #ifndef NDEBUG |
1164 | // For now, we do a mandatory verification step because we don't |
1165 | // entirely trust this pass. Note that we don't want to add a verifier |
1166 | // pass to FPM below because it will also verify all the global data. |
1167 | if (verifyFunction(F, OS: &errs())) |
1168 | report_fatal_error(reason: "Broken function" ); |
1169 | #endif |
1170 | } |
1171 | |
1172 | // Assuming we arrived at the block NewBlock from Prev instruction, store |
1173 | // PHI's incoming values in the ResolvedValues map. |
1174 | static void |
1175 | scanPHIsAndUpdateValueMap(Instruction *Prev, BasicBlock *NewBlock, |
1176 | DenseMap<Value *, Value *> &ResolvedValues) { |
1177 | auto *PrevBB = Prev->getParent(); |
1178 | for (PHINode &PN : NewBlock->phis()) { |
1179 | auto V = PN.getIncomingValueForBlock(BB: PrevBB); |
1180 | // See if we already resolved it. |
1181 | auto VI = ResolvedValues.find(Val: V); |
1182 | if (VI != ResolvedValues.end()) |
1183 | V = VI->second; |
1184 | // Remember the value. |
1185 | ResolvedValues[&PN] = V; |
1186 | } |
1187 | } |
1188 | |
1189 | // Replace a sequence of branches leading to a ret, with a clone of a ret |
1190 | // instruction. Suspend instruction represented by a switch, track the PHI |
1191 | // values and select the correct case successor when possible. |
1192 | static bool simplifyTerminatorLeadingToRet(Instruction *InitialInst) { |
1193 | // There is nothing to simplify. |
1194 | if (isa<ReturnInst>(Val: InitialInst)) |
1195 | return false; |
1196 | |
1197 | DenseMap<Value *, Value *> ResolvedValues; |
1198 | assert(InitialInst->getModule()); |
1199 | const DataLayout &DL = InitialInst->getModule()->getDataLayout(); |
1200 | |
1201 | auto TryResolveConstant = [&ResolvedValues](Value *V) { |
1202 | auto It = ResolvedValues.find(Val: V); |
1203 | if (It != ResolvedValues.end()) |
1204 | V = It->second; |
1205 | return dyn_cast<ConstantInt>(Val: V); |
1206 | }; |
1207 | |
1208 | Instruction *I = InitialInst; |
1209 | while (true) { |
1210 | if (isa<ReturnInst>(Val: I)) { |
1211 | assert(!cast<ReturnInst>(I)->getReturnValue()); |
1212 | ReplaceInstWithInst(From: InitialInst, To: I->clone()); |
1213 | return true; |
1214 | } |
1215 | |
1216 | if (auto *BR = dyn_cast<BranchInst>(Val: I)) { |
1217 | unsigned SuccIndex = 0; |
1218 | if (BR->isConditional()) { |
1219 | // Handle the case the condition of the conditional branch is constant. |
1220 | // e.g., |
1221 | // |
1222 | // br i1 false, label %cleanup, label %CoroEnd |
1223 | // |
1224 | // It is possible during the transformation. We could continue the |
1225 | // simplifying in this case. |
1226 | ConstantInt *Cond = TryResolveConstant(BR->getCondition()); |
1227 | if (!Cond) |
1228 | return false; |
1229 | |
1230 | SuccIndex = Cond->isOne() ? 0 : 1; |
1231 | } |
1232 | |
1233 | BasicBlock *Succ = BR->getSuccessor(i: SuccIndex); |
1234 | scanPHIsAndUpdateValueMap(Prev: I, NewBlock: Succ, ResolvedValues); |
1235 | I = Succ->getFirstNonPHIOrDbgOrLifetime(); |
1236 | continue; |
1237 | } |
1238 | |
1239 | if (auto *Cmp = dyn_cast<CmpInst>(Val: I)) { |
1240 | // If the case number of suspended switch instruction is reduced to |
1241 | // 1, then it is simplified to CmpInst in llvm::ConstantFoldTerminator. |
1242 | // Try to constant fold it. |
1243 | ConstantInt *Cond0 = TryResolveConstant(Cmp->getOperand(i_nocapture: 0)); |
1244 | ConstantInt *Cond1 = TryResolveConstant(Cmp->getOperand(i_nocapture: 1)); |
1245 | if (Cond0 && Cond1) { |
1246 | ConstantInt *Result = |
1247 | dyn_cast_or_null<ConstantInt>(Val: ConstantFoldCompareInstOperands( |
1248 | Predicate: Cmp->getPredicate(), LHS: Cond0, RHS: Cond1, DL)); |
1249 | if (Result) { |
1250 | ResolvedValues[Cmp] = Result; |
1251 | I = I->getNextNode(); |
1252 | continue; |
1253 | } |
1254 | } |
1255 | } |
1256 | |
1257 | if (auto *SI = dyn_cast<SwitchInst>(Val: I)) { |
1258 | ConstantInt *Cond = TryResolveConstant(SI->getCondition()); |
1259 | if (!Cond) |
1260 | return false; |
1261 | |
1262 | BasicBlock *Succ = SI->findCaseValue(C: Cond)->getCaseSuccessor(); |
1263 | scanPHIsAndUpdateValueMap(Prev: I, NewBlock: Succ, ResolvedValues); |
1264 | I = Succ->getFirstNonPHIOrDbgOrLifetime(); |
1265 | continue; |
1266 | } |
1267 | |
1268 | if (I->isDebugOrPseudoInst() || I->isLifetimeStartOrEnd() || |
1269 | wouldInstructionBeTriviallyDead(I)) { |
1270 | // We can skip instructions without side effects. If their values are |
1271 | // needed, we'll notice later, e.g. when hitting a conditional branch. |
1272 | I = I->getNextNode(); |
1273 | continue; |
1274 | } |
1275 | |
1276 | break; |
1277 | } |
1278 | |
1279 | return false; |
1280 | } |
1281 | |
1282 | // Check whether CI obeys the rules of musttail attribute. |
1283 | static bool shouldBeMustTail(const CallInst &CI, const Function &F) { |
1284 | if (CI.isInlineAsm()) |
1285 | return false; |
1286 | |
1287 | // Match prototypes and calling conventions of resume function. |
1288 | FunctionType *CalleeTy = CI.getFunctionType(); |
1289 | if (!CalleeTy->getReturnType()->isVoidTy() || (CalleeTy->getNumParams() != 1)) |
1290 | return false; |
1291 | |
1292 | Type *CalleeParmTy = CalleeTy->getParamType(i: 0); |
1293 | if (!CalleeParmTy->isPointerTy() || |
1294 | (CalleeParmTy->getPointerAddressSpace() != 0)) |
1295 | return false; |
1296 | |
1297 | if (CI.getCallingConv() != F.getCallingConv()) |
1298 | return false; |
1299 | |
1300 | // CI should not has any ABI-impacting function attributes. |
1301 | static const Attribute::AttrKind ABIAttrs[] = { |
1302 | Attribute::StructRet, Attribute::ByVal, Attribute::InAlloca, |
1303 | Attribute::Preallocated, Attribute::InReg, Attribute::Returned, |
1304 | Attribute::SwiftSelf, Attribute::SwiftError}; |
1305 | AttributeList Attrs = CI.getAttributes(); |
1306 | for (auto AK : ABIAttrs) |
1307 | if (Attrs.hasParamAttr(0, AK)) |
1308 | return false; |
1309 | |
1310 | return true; |
1311 | } |
1312 | |
1313 | // Coroutine has no suspend points. Remove heap allocation for the coroutine |
1314 | // frame if possible. |
1315 | static void handleNoSuspendCoroutine(coro::Shape &Shape) { |
1316 | auto *CoroBegin = Shape.CoroBegin; |
1317 | auto *CoroId = CoroBegin->getId(); |
1318 | auto *AllocInst = CoroId->getCoroAlloc(); |
1319 | switch (Shape.ABI) { |
1320 | case coro::ABI::Switch: { |
1321 | auto SwitchId = cast<CoroIdInst>(Val: CoroId); |
1322 | coro::replaceCoroFree(CoroId: SwitchId, /*Elide=*/AllocInst != nullptr); |
1323 | if (AllocInst) { |
1324 | IRBuilder<> Builder(AllocInst); |
1325 | auto *Frame = Builder.CreateAlloca(Ty: Shape.FrameTy); |
1326 | Frame->setAlignment(Shape.FrameAlign); |
1327 | AllocInst->replaceAllUsesWith(V: Builder.getFalse()); |
1328 | AllocInst->eraseFromParent(); |
1329 | CoroBegin->replaceAllUsesWith(V: Frame); |
1330 | } else { |
1331 | CoroBegin->replaceAllUsesWith(V: CoroBegin->getMem()); |
1332 | } |
1333 | |
1334 | break; |
1335 | } |
1336 | case coro::ABI::Async: |
1337 | case coro::ABI::Retcon: |
1338 | case coro::ABI::RetconOnce: |
1339 | CoroBegin->replaceAllUsesWith(V: UndefValue::get(T: CoroBegin->getType())); |
1340 | break; |
1341 | } |
1342 | |
1343 | CoroBegin->eraseFromParent(); |
1344 | } |
1345 | |
1346 | // SimplifySuspendPoint needs to check that there is no calls between |
1347 | // coro_save and coro_suspend, since any of the calls may potentially resume |
1348 | // the coroutine and if that is the case we cannot eliminate the suspend point. |
1349 | static bool hasCallsInBlockBetween(Instruction *From, Instruction *To) { |
1350 | for (Instruction *I = From; I != To; I = I->getNextNode()) { |
1351 | // Assume that no intrinsic can resume the coroutine. |
1352 | if (isa<IntrinsicInst>(Val: I)) |
1353 | continue; |
1354 | |
1355 | if (isa<CallBase>(Val: I)) |
1356 | return true; |
1357 | } |
1358 | return false; |
1359 | } |
1360 | |
1361 | static bool hasCallsInBlocksBetween(BasicBlock *SaveBB, BasicBlock *ResDesBB) { |
1362 | SmallPtrSet<BasicBlock *, 8> Set; |
1363 | SmallVector<BasicBlock *, 8> Worklist; |
1364 | |
1365 | Set.insert(Ptr: SaveBB); |
1366 | Worklist.push_back(Elt: ResDesBB); |
1367 | |
1368 | // Accumulate all blocks between SaveBB and ResDesBB. Because CoroSaveIntr |
1369 | // returns a token consumed by suspend instruction, all blocks in between |
1370 | // will have to eventually hit SaveBB when going backwards from ResDesBB. |
1371 | while (!Worklist.empty()) { |
1372 | auto *BB = Worklist.pop_back_val(); |
1373 | Set.insert(Ptr: BB); |
1374 | for (auto *Pred : predecessors(BB)) |
1375 | if (!Set.contains(Ptr: Pred)) |
1376 | Worklist.push_back(Elt: Pred); |
1377 | } |
1378 | |
1379 | // SaveBB and ResDesBB are checked separately in hasCallsBetween. |
1380 | Set.erase(Ptr: SaveBB); |
1381 | Set.erase(Ptr: ResDesBB); |
1382 | |
1383 | for (auto *BB : Set) |
1384 | if (hasCallsInBlockBetween(From: BB->getFirstNonPHI(), To: nullptr)) |
1385 | return true; |
1386 | |
1387 | return false; |
1388 | } |
1389 | |
1390 | static bool hasCallsBetween(Instruction *Save, Instruction *ResumeOrDestroy) { |
1391 | auto *SaveBB = Save->getParent(); |
1392 | auto *ResumeOrDestroyBB = ResumeOrDestroy->getParent(); |
1393 | |
1394 | if (SaveBB == ResumeOrDestroyBB) |
1395 | return hasCallsInBlockBetween(From: Save->getNextNode(), To: ResumeOrDestroy); |
1396 | |
1397 | // Any calls from Save to the end of the block? |
1398 | if (hasCallsInBlockBetween(From: Save->getNextNode(), To: nullptr)) |
1399 | return true; |
1400 | |
1401 | // Any calls from begging of the block up to ResumeOrDestroy? |
1402 | if (hasCallsInBlockBetween(From: ResumeOrDestroyBB->getFirstNonPHI(), |
1403 | To: ResumeOrDestroy)) |
1404 | return true; |
1405 | |
1406 | // Any calls in all of the blocks between SaveBB and ResumeOrDestroyBB? |
1407 | if (hasCallsInBlocksBetween(SaveBB, ResDesBB: ResumeOrDestroyBB)) |
1408 | return true; |
1409 | |
1410 | return false; |
1411 | } |
1412 | |
1413 | // If a SuspendIntrin is preceded by Resume or Destroy, we can eliminate the |
1414 | // suspend point and replace it with nornal control flow. |
1415 | static bool simplifySuspendPoint(CoroSuspendInst *Suspend, |
1416 | CoroBeginInst *CoroBegin) { |
1417 | Instruction *Prev = Suspend->getPrevNode(); |
1418 | if (!Prev) { |
1419 | auto *Pred = Suspend->getParent()->getSinglePredecessor(); |
1420 | if (!Pred) |
1421 | return false; |
1422 | Prev = Pred->getTerminator(); |
1423 | } |
1424 | |
1425 | CallBase *CB = dyn_cast<CallBase>(Val: Prev); |
1426 | if (!CB) |
1427 | return false; |
1428 | |
1429 | auto *Callee = CB->getCalledOperand()->stripPointerCasts(); |
1430 | |
1431 | // See if the callsite is for resumption or destruction of the coroutine. |
1432 | auto *SubFn = dyn_cast<CoroSubFnInst>(Val: Callee); |
1433 | if (!SubFn) |
1434 | return false; |
1435 | |
1436 | // Does not refer to the current coroutine, we cannot do anything with it. |
1437 | if (SubFn->getFrame() != CoroBegin) |
1438 | return false; |
1439 | |
1440 | // See if the transformation is safe. Specifically, see if there are any |
1441 | // calls in between Save and CallInstr. They can potenitally resume the |
1442 | // coroutine rendering this optimization unsafe. |
1443 | auto *Save = Suspend->getCoroSave(); |
1444 | if (hasCallsBetween(Save, ResumeOrDestroy: CB)) |
1445 | return false; |
1446 | |
1447 | // Replace llvm.coro.suspend with the value that results in resumption over |
1448 | // the resume or cleanup path. |
1449 | Suspend->replaceAllUsesWith(V: SubFn->getRawIndex()); |
1450 | Suspend->eraseFromParent(); |
1451 | Save->eraseFromParent(); |
1452 | |
1453 | // No longer need a call to coro.resume or coro.destroy. |
1454 | if (auto *Invoke = dyn_cast<InvokeInst>(Val: CB)) { |
1455 | BranchInst::Create(IfTrue: Invoke->getNormalDest(), InsertBefore: Invoke->getIterator()); |
1456 | } |
1457 | |
1458 | // Grab the CalledValue from CB before erasing the CallInstr. |
1459 | auto *CalledValue = CB->getCalledOperand(); |
1460 | CB->eraseFromParent(); |
1461 | |
1462 | // If no more users remove it. Usually it is a bitcast of SubFn. |
1463 | if (CalledValue != SubFn && CalledValue->user_empty()) |
1464 | if (auto *I = dyn_cast<Instruction>(Val: CalledValue)) |
1465 | I->eraseFromParent(); |
1466 | |
1467 | // Now we are good to remove SubFn. |
1468 | if (SubFn->user_empty()) |
1469 | SubFn->eraseFromParent(); |
1470 | |
1471 | return true; |
1472 | } |
1473 | |
1474 | // Remove suspend points that are simplified. |
1475 | static void simplifySuspendPoints(coro::Shape &Shape) { |
1476 | // Currently, the only simplification we do is switch-lowering-specific. |
1477 | if (Shape.ABI != coro::ABI::Switch) |
1478 | return; |
1479 | |
1480 | auto &S = Shape.CoroSuspends; |
1481 | size_t I = 0, N = S.size(); |
1482 | if (N == 0) |
1483 | return; |
1484 | |
1485 | size_t ChangedFinalIndex = std::numeric_limits<size_t>::max(); |
1486 | while (true) { |
1487 | auto SI = cast<CoroSuspendInst>(Val: S[I]); |
1488 | // Leave final.suspend to handleFinalSuspend since it is undefined behavior |
1489 | // to resume a coroutine suspended at the final suspend point. |
1490 | if (!SI->isFinal() && simplifySuspendPoint(Suspend: SI, CoroBegin: Shape.CoroBegin)) { |
1491 | if (--N == I) |
1492 | break; |
1493 | |
1494 | std::swap(a&: S[I], b&: S[N]); |
1495 | |
1496 | if (cast<CoroSuspendInst>(Val: S[I])->isFinal()) { |
1497 | assert(Shape.SwitchLowering.HasFinalSuspend); |
1498 | ChangedFinalIndex = I; |
1499 | } |
1500 | |
1501 | continue; |
1502 | } |
1503 | if (++I == N) |
1504 | break; |
1505 | } |
1506 | S.resize(N); |
1507 | |
1508 | // Maintain final.suspend in case final suspend was swapped. |
1509 | // Due to we requrie the final suspend to be the last element of CoroSuspends. |
1510 | if (ChangedFinalIndex < N) { |
1511 | assert(cast<CoroSuspendInst>(S[ChangedFinalIndex])->isFinal()); |
1512 | std::swap(a&: S[ChangedFinalIndex], b&: S.back()); |
1513 | } |
1514 | } |
1515 | |
1516 | namespace { |
1517 | |
1518 | struct SwitchCoroutineSplitter { |
1519 | static void split(Function &F, coro::Shape &Shape, |
1520 | SmallVectorImpl<Function *> &Clones, |
1521 | TargetTransformInfo &TTI) { |
1522 | assert(Shape.ABI == coro::ABI::Switch); |
1523 | |
1524 | createResumeEntryBlock(F, Shape); |
1525 | auto *ResumeClone = |
1526 | createClone(F, Suffix: ".resume" , Shape, FKind: CoroCloner::Kind::SwitchResume); |
1527 | auto *DestroyClone = |
1528 | createClone(F, Suffix: ".destroy" , Shape, FKind: CoroCloner::Kind::SwitchUnwind); |
1529 | auto *CleanupClone = |
1530 | createClone(F, Suffix: ".cleanup" , Shape, FKind: CoroCloner::Kind::SwitchCleanup); |
1531 | |
1532 | postSplitCleanup(F&: *ResumeClone); |
1533 | postSplitCleanup(F&: *DestroyClone); |
1534 | postSplitCleanup(F&: *CleanupClone); |
1535 | |
1536 | // Adding musttail call to support symmetric transfer. |
1537 | // Skip targets which don't support tail call. |
1538 | // |
1539 | // FIXME: Could we support symmetric transfer effectively without musttail |
1540 | // call? |
1541 | if (TTI.supportsTailCalls()) |
1542 | addMustTailToCoroResumes(F&: *ResumeClone, TTI); |
1543 | |
1544 | // Store addresses resume/destroy/cleanup functions in the coroutine frame. |
1545 | updateCoroFrame(Shape, ResumeFn: ResumeClone, DestroyFn: DestroyClone, CleanupFn: CleanupClone); |
1546 | |
1547 | assert(Clones.empty()); |
1548 | Clones.push_back(Elt: ResumeClone); |
1549 | Clones.push_back(Elt: DestroyClone); |
1550 | Clones.push_back(Elt: CleanupClone); |
1551 | |
1552 | // Create a constant array referring to resume/destroy/clone functions |
1553 | // pointed by the last argument of @llvm.coro.info, so that CoroElide pass |
1554 | // can determined correct function to call. |
1555 | setCoroInfo(F, Shape, Fns: Clones); |
1556 | } |
1557 | |
1558 | private: |
1559 | // Create a resume clone by cloning the body of the original function, setting |
1560 | // new entry block and replacing coro.suspend an appropriate value to force |
1561 | // resume or cleanup pass for every suspend point. |
1562 | static Function *createClone(Function &F, const Twine &Suffix, |
1563 | coro::Shape &Shape, CoroCloner::Kind FKind) { |
1564 | CoroCloner Cloner(F, Suffix, Shape, FKind); |
1565 | Cloner.create(); |
1566 | return Cloner.getFunction(); |
1567 | } |
1568 | |
1569 | // Create an entry block for a resume function with a switch that will jump to |
1570 | // suspend points. |
1571 | static void createResumeEntryBlock(Function &F, coro::Shape &Shape) { |
1572 | LLVMContext &C = F.getContext(); |
1573 | |
1574 | // resume.entry: |
1575 | // %index.addr = getelementptr inbounds %f.Frame, %f.Frame* %FramePtr, i32 |
1576 | // 0, i32 2 % index = load i32, i32* %index.addr switch i32 %index, label |
1577 | // %unreachable [ |
1578 | // i32 0, label %resume.0 |
1579 | // i32 1, label %resume.1 |
1580 | // ... |
1581 | // ] |
1582 | |
1583 | auto *NewEntry = BasicBlock::Create(Context&: C, Name: "resume.entry" , Parent: &F); |
1584 | auto *UnreachBB = BasicBlock::Create(Context&: C, Name: "unreachable" , Parent: &F); |
1585 | |
1586 | IRBuilder<> Builder(NewEntry); |
1587 | auto *FramePtr = Shape.FramePtr; |
1588 | auto *FrameTy = Shape.FrameTy; |
1589 | auto *GepIndex = Builder.CreateStructGEP( |
1590 | Ty: FrameTy, Ptr: FramePtr, Idx: Shape.getSwitchIndexField(), Name: "index.addr" ); |
1591 | auto *Index = Builder.CreateLoad(Ty: Shape.getIndexType(), Ptr: GepIndex, Name: "index" ); |
1592 | auto *Switch = |
1593 | Builder.CreateSwitch(V: Index, Dest: UnreachBB, NumCases: Shape.CoroSuspends.size()); |
1594 | Shape.SwitchLowering.ResumeSwitch = Switch; |
1595 | |
1596 | size_t SuspendIndex = 0; |
1597 | for (auto *AnyS : Shape.CoroSuspends) { |
1598 | auto *S = cast<CoroSuspendInst>(Val: AnyS); |
1599 | ConstantInt *IndexVal = Shape.getIndex(Value: SuspendIndex); |
1600 | |
1601 | // Replace CoroSave with a store to Index: |
1602 | // %index.addr = getelementptr %f.frame... (index field number) |
1603 | // store i32 %IndexVal, i32* %index.addr1 |
1604 | auto *Save = S->getCoroSave(); |
1605 | Builder.SetInsertPoint(Save); |
1606 | if (S->isFinal()) { |
1607 | // The coroutine should be marked done if it reaches the final suspend |
1608 | // point. |
1609 | markCoroutineAsDone(Builder, Shape, FramePtr); |
1610 | } else { |
1611 | auto *GepIndex = Builder.CreateStructGEP( |
1612 | Ty: FrameTy, Ptr: FramePtr, Idx: Shape.getSwitchIndexField(), Name: "index.addr" ); |
1613 | Builder.CreateStore(Val: IndexVal, Ptr: GepIndex); |
1614 | } |
1615 | |
1616 | Save->replaceAllUsesWith(V: ConstantTokenNone::get(Context&: C)); |
1617 | Save->eraseFromParent(); |
1618 | |
1619 | // Split block before and after coro.suspend and add a jump from an entry |
1620 | // switch: |
1621 | // |
1622 | // whateverBB: |
1623 | // whatever |
1624 | // %0 = call i8 @llvm.coro.suspend(token none, i1 false) |
1625 | // switch i8 %0, label %suspend[i8 0, label %resume |
1626 | // i8 1, label %cleanup] |
1627 | // becomes: |
1628 | // |
1629 | // whateverBB: |
1630 | // whatever |
1631 | // br label %resume.0.landing |
1632 | // |
1633 | // resume.0: ; <--- jump from the switch in the resume.entry |
1634 | // %0 = tail call i8 @llvm.coro.suspend(token none, i1 false) |
1635 | // br label %resume.0.landing |
1636 | // |
1637 | // resume.0.landing: |
1638 | // %1 = phi i8[-1, %whateverBB], [%0, %resume.0] |
1639 | // switch i8 % 1, label %suspend [i8 0, label %resume |
1640 | // i8 1, label %cleanup] |
1641 | |
1642 | auto *SuspendBB = S->getParent(); |
1643 | auto *ResumeBB = |
1644 | SuspendBB->splitBasicBlock(I: S, BBName: "resume." + Twine(SuspendIndex)); |
1645 | auto *LandingBB = ResumeBB->splitBasicBlock( |
1646 | I: S->getNextNode(), BBName: ResumeBB->getName() + Twine(".landing" )); |
1647 | Switch->addCase(OnVal: IndexVal, Dest: ResumeBB); |
1648 | |
1649 | cast<BranchInst>(Val: SuspendBB->getTerminator())->setSuccessor(idx: 0, NewSucc: LandingBB); |
1650 | auto *PN = PHINode::Create(Ty: Builder.getInt8Ty(), NumReservedValues: 2, NameStr: "" ); |
1651 | PN->insertBefore(InsertPos: LandingBB->begin()); |
1652 | S->replaceAllUsesWith(V: PN); |
1653 | PN->addIncoming(V: Builder.getInt8(C: -1), BB: SuspendBB); |
1654 | PN->addIncoming(V: S, BB: ResumeBB); |
1655 | |
1656 | ++SuspendIndex; |
1657 | } |
1658 | |
1659 | Builder.SetInsertPoint(UnreachBB); |
1660 | Builder.CreateUnreachable(); |
1661 | |
1662 | Shape.SwitchLowering.ResumeEntryBlock = NewEntry; |
1663 | } |
1664 | |
1665 | // Add musttail to any resume instructions that is immediately followed by a |
1666 | // suspend (i.e. ret). We do this even in -O0 to support guaranteed tail call |
1667 | // for symmetrical coroutine control transfer (C++ Coroutines TS extension). |
1668 | // This transformation is done only in the resume part of the coroutine that |
1669 | // has identical signature and calling convention as the coro.resume call. |
1670 | static void addMustTailToCoroResumes(Function &F, TargetTransformInfo &TTI) { |
1671 | bool Changed = false; |
1672 | |
1673 | // Collect potential resume instructions. |
1674 | SmallVector<CallInst *, 4> Resumes; |
1675 | for (auto &I : instructions(F)) |
1676 | if (auto *Call = dyn_cast<CallInst>(Val: &I)) |
1677 | if (shouldBeMustTail(CI: *Call, F)) |
1678 | Resumes.push_back(Elt: Call); |
1679 | |
1680 | // Set musttail on those that are followed by a ret instruction. |
1681 | for (CallInst *Call : Resumes) |
1682 | // Skip targets which don't support tail call on the specific case. |
1683 | if (TTI.supportsTailCallFor(CB: Call) && |
1684 | simplifyTerminatorLeadingToRet(InitialInst: Call->getNextNode())) { |
1685 | Call->setTailCallKind(CallInst::TCK_MustTail); |
1686 | Changed = true; |
1687 | } |
1688 | |
1689 | if (Changed) |
1690 | removeUnreachableBlocks(F); |
1691 | } |
1692 | |
1693 | // Store addresses of Resume/Destroy/Cleanup functions in the coroutine frame. |
1694 | static void updateCoroFrame(coro::Shape &Shape, Function *ResumeFn, |
1695 | Function *DestroyFn, Function *CleanupFn) { |
1696 | IRBuilder<> Builder(&*Shape.getInsertPtAfterFramePtr()); |
1697 | |
1698 | auto *ResumeAddr = Builder.CreateStructGEP( |
1699 | Ty: Shape.FrameTy, Ptr: Shape.FramePtr, Idx: coro::Shape::SwitchFieldIndex::Resume, |
1700 | Name: "resume.addr" ); |
1701 | Builder.CreateStore(Val: ResumeFn, Ptr: ResumeAddr); |
1702 | |
1703 | Value *DestroyOrCleanupFn = DestroyFn; |
1704 | |
1705 | CoroIdInst *CoroId = Shape.getSwitchCoroId(); |
1706 | if (CoroAllocInst *CA = CoroId->getCoroAlloc()) { |
1707 | // If there is a CoroAlloc and it returns false (meaning we elide the |
1708 | // allocation, use CleanupFn instead of DestroyFn). |
1709 | DestroyOrCleanupFn = Builder.CreateSelect(C: CA, True: DestroyFn, False: CleanupFn); |
1710 | } |
1711 | |
1712 | auto *DestroyAddr = Builder.CreateStructGEP( |
1713 | Ty: Shape.FrameTy, Ptr: Shape.FramePtr, Idx: coro::Shape::SwitchFieldIndex::Destroy, |
1714 | Name: "destroy.addr" ); |
1715 | Builder.CreateStore(Val: DestroyOrCleanupFn, Ptr: DestroyAddr); |
1716 | } |
1717 | |
1718 | // Create a global constant array containing pointers to functions provided |
1719 | // and set Info parameter of CoroBegin to point at this constant. Example: |
1720 | // |
1721 | // @f.resumers = internal constant [2 x void(%f.frame*)*] |
1722 | // [void(%f.frame*)* @f.resume, void(%f.frame*)* |
1723 | // @f.destroy] |
1724 | // define void @f() { |
1725 | // ... |
1726 | // call i8* @llvm.coro.begin(i8* null, i32 0, i8* null, |
1727 | // i8* bitcast([2 x void(%f.frame*)*] * @f.resumers to |
1728 | // i8*)) |
1729 | // |
1730 | // Assumes that all the functions have the same signature. |
1731 | static void setCoroInfo(Function &F, coro::Shape &Shape, |
1732 | ArrayRef<Function *> Fns) { |
1733 | // This only works under the switch-lowering ABI because coro elision |
1734 | // only works on the switch-lowering ABI. |
1735 | SmallVector<Constant *, 4> Args(Fns.begin(), Fns.end()); |
1736 | assert(!Args.empty()); |
1737 | Function *Part = *Fns.begin(); |
1738 | Module *M = Part->getParent(); |
1739 | auto *ArrTy = ArrayType::get(ElementType: Part->getType(), NumElements: Args.size()); |
1740 | |
1741 | auto *ConstVal = ConstantArray::get(T: ArrTy, V: Args); |
1742 | auto *GV = new GlobalVariable(*M, ConstVal->getType(), /*isConstant=*/true, |
1743 | GlobalVariable::PrivateLinkage, ConstVal, |
1744 | F.getName() + Twine(".resumers" )); |
1745 | |
1746 | // Update coro.begin instruction to refer to this constant. |
1747 | LLVMContext &C = F.getContext(); |
1748 | auto *BC = ConstantExpr::getPointerCast(C: GV, Ty: PointerType::getUnqual(C)); |
1749 | Shape.getSwitchCoroId()->setInfo(BC); |
1750 | } |
1751 | }; |
1752 | |
1753 | } // namespace |
1754 | |
1755 | static void replaceAsyncResumeFunction(CoroSuspendAsyncInst *Suspend, |
1756 | Value *Continuation) { |
1757 | auto *ResumeIntrinsic = Suspend->getResumeFunction(); |
1758 | auto &Context = Suspend->getParent()->getParent()->getContext(); |
1759 | auto *Int8PtrTy = PointerType::getUnqual(C&: Context); |
1760 | |
1761 | IRBuilder<> Builder(ResumeIntrinsic); |
1762 | auto *Val = Builder.CreateBitOrPointerCast(V: Continuation, DestTy: Int8PtrTy); |
1763 | ResumeIntrinsic->replaceAllUsesWith(V: Val); |
1764 | ResumeIntrinsic->eraseFromParent(); |
1765 | Suspend->setOperand(i_nocapture: CoroSuspendAsyncInst::ResumeFunctionArg, |
1766 | Val_nocapture: UndefValue::get(T: Int8PtrTy)); |
1767 | } |
1768 | |
1769 | /// Coerce the arguments in \p FnArgs according to \p FnTy in \p CallArgs. |
1770 | static void coerceArguments(IRBuilder<> &Builder, FunctionType *FnTy, |
1771 | ArrayRef<Value *> FnArgs, |
1772 | SmallVectorImpl<Value *> &CallArgs) { |
1773 | size_t ArgIdx = 0; |
1774 | for (auto *paramTy : FnTy->params()) { |
1775 | assert(ArgIdx < FnArgs.size()); |
1776 | if (paramTy != FnArgs[ArgIdx]->getType()) |
1777 | CallArgs.push_back( |
1778 | Elt: Builder.CreateBitOrPointerCast(V: FnArgs[ArgIdx], DestTy: paramTy)); |
1779 | else |
1780 | CallArgs.push_back(Elt: FnArgs[ArgIdx]); |
1781 | ++ArgIdx; |
1782 | } |
1783 | } |
1784 | |
1785 | CallInst *coro::createMustTailCall(DebugLoc Loc, Function *MustTailCallFn, |
1786 | TargetTransformInfo &TTI, |
1787 | ArrayRef<Value *> Arguments, |
1788 | IRBuilder<> &Builder) { |
1789 | auto *FnTy = MustTailCallFn->getFunctionType(); |
1790 | // Coerce the arguments, llvm optimizations seem to ignore the types in |
1791 | // vaarg functions and throws away casts in optimized mode. |
1792 | SmallVector<Value *, 8> CallArgs; |
1793 | coerceArguments(Builder, FnTy, FnArgs: Arguments, CallArgs); |
1794 | |
1795 | auto *TailCall = Builder.CreateCall(FTy: FnTy, Callee: MustTailCallFn, Args: CallArgs); |
1796 | // Skip targets which don't support tail call. |
1797 | if (TTI.supportsTailCallFor(CB: TailCall)) { |
1798 | TailCall->setTailCallKind(CallInst::TCK_MustTail); |
1799 | } |
1800 | TailCall->setDebugLoc(Loc); |
1801 | TailCall->setCallingConv(MustTailCallFn->getCallingConv()); |
1802 | return TailCall; |
1803 | } |
1804 | |
1805 | static void splitAsyncCoroutine(Function &F, coro::Shape &Shape, |
1806 | SmallVectorImpl<Function *> &Clones, |
1807 | TargetTransformInfo &TTI) { |
1808 | assert(Shape.ABI == coro::ABI::Async); |
1809 | assert(Clones.empty()); |
1810 | // Reset various things that the optimizer might have decided it |
1811 | // "knows" about the coroutine function due to not seeing a return. |
1812 | F.removeFnAttr(Attribute::NoReturn); |
1813 | F.removeRetAttr(Attribute::NoAlias); |
1814 | F.removeRetAttr(Attribute::NonNull); |
1815 | |
1816 | auto &Context = F.getContext(); |
1817 | auto *Int8PtrTy = PointerType::getUnqual(C&: Context); |
1818 | |
1819 | auto *Id = cast<CoroIdAsyncInst>(Val: Shape.CoroBegin->getId()); |
1820 | IRBuilder<> Builder(Id); |
1821 | |
1822 | auto *FramePtr = Id->getStorage(); |
1823 | FramePtr = Builder.CreateBitOrPointerCast(V: FramePtr, DestTy: Int8PtrTy); |
1824 | FramePtr = Builder.CreateConstInBoundsGEP1_32( |
1825 | Ty: Type::getInt8Ty(C&: Context), Ptr: FramePtr, Idx0: Shape.AsyncLowering.FrameOffset, |
1826 | Name: "async.ctx.frameptr" ); |
1827 | |
1828 | // Map all uses of llvm.coro.begin to the allocated frame pointer. |
1829 | { |
1830 | // Make sure we don't invalidate Shape.FramePtr. |
1831 | TrackingVH<Value> Handle(Shape.FramePtr); |
1832 | Shape.CoroBegin->replaceAllUsesWith(V: FramePtr); |
1833 | Shape.FramePtr = Handle.getValPtr(); |
1834 | } |
1835 | |
1836 | // Create all the functions in order after the main function. |
1837 | auto NextF = std::next(x: F.getIterator()); |
1838 | |
1839 | // Create a continuation function for each of the suspend points. |
1840 | Clones.reserve(N: Shape.CoroSuspends.size()); |
1841 | for (size_t Idx = 0, End = Shape.CoroSuspends.size(); Idx != End; ++Idx) { |
1842 | auto *Suspend = cast<CoroSuspendAsyncInst>(Val: Shape.CoroSuspends[Idx]); |
1843 | |
1844 | // Create the clone declaration. |
1845 | auto ResumeNameSuffix = ".resume." ; |
1846 | auto ProjectionFunctionName = |
1847 | Suspend->getAsyncContextProjectionFunction()->getName(); |
1848 | bool UseSwiftMangling = false; |
1849 | if (ProjectionFunctionName.equals(RHS: "__swift_async_resume_project_context" )) { |
1850 | ResumeNameSuffix = "TQ" ; |
1851 | UseSwiftMangling = true; |
1852 | } else if (ProjectionFunctionName.equals( |
1853 | RHS: "__swift_async_resume_get_context" )) { |
1854 | ResumeNameSuffix = "TY" ; |
1855 | UseSwiftMangling = true; |
1856 | } |
1857 | auto *Continuation = createCloneDeclaration( |
1858 | OrigF&: F, Shape, |
1859 | Suffix: UseSwiftMangling ? ResumeNameSuffix + Twine(Idx) + "_" |
1860 | : ResumeNameSuffix + Twine(Idx), |
1861 | InsertBefore: NextF, ActiveSuspend: Suspend); |
1862 | Clones.push_back(Elt: Continuation); |
1863 | |
1864 | // Insert a branch to a new return block immediately before the suspend |
1865 | // point. |
1866 | auto *SuspendBB = Suspend->getParent(); |
1867 | auto *NewSuspendBB = SuspendBB->splitBasicBlock(I: Suspend); |
1868 | auto *Branch = cast<BranchInst>(Val: SuspendBB->getTerminator()); |
1869 | |
1870 | // Place it before the first suspend. |
1871 | auto *ReturnBB = |
1872 | BasicBlock::Create(Context&: F.getContext(), Name: "coro.return" , Parent: &F, InsertBefore: NewSuspendBB); |
1873 | Branch->setSuccessor(idx: 0, NewSucc: ReturnBB); |
1874 | |
1875 | IRBuilder<> Builder(ReturnBB); |
1876 | |
1877 | // Insert the call to the tail call function and inline it. |
1878 | auto *Fn = Suspend->getMustTailCallFunction(); |
1879 | SmallVector<Value *, 8> Args(Suspend->args()); |
1880 | auto FnArgs = ArrayRef<Value *>(Args).drop_front( |
1881 | N: CoroSuspendAsyncInst::MustTailCallFuncArg + 1); |
1882 | auto *TailCall = coro::createMustTailCall(Loc: Suspend->getDebugLoc(), MustTailCallFn: Fn, TTI, |
1883 | Arguments: FnArgs, Builder); |
1884 | Builder.CreateRetVoid(); |
1885 | InlineFunctionInfo FnInfo; |
1886 | (void)InlineFunction(CB&: *TailCall, IFI&: FnInfo); |
1887 | |
1888 | // Replace the lvm.coro.async.resume intrisic call. |
1889 | replaceAsyncResumeFunction(Suspend, Continuation); |
1890 | } |
1891 | |
1892 | assert(Clones.size() == Shape.CoroSuspends.size()); |
1893 | for (size_t Idx = 0, End = Shape.CoroSuspends.size(); Idx != End; ++Idx) { |
1894 | auto *Suspend = Shape.CoroSuspends[Idx]; |
1895 | auto *Clone = Clones[Idx]; |
1896 | |
1897 | CoroCloner(F, "resume." + Twine(Idx), Shape, Clone, Suspend).create(); |
1898 | } |
1899 | } |
1900 | |
1901 | static void splitRetconCoroutine(Function &F, coro::Shape &Shape, |
1902 | SmallVectorImpl<Function *> &Clones) { |
1903 | assert(Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce); |
1904 | assert(Clones.empty()); |
1905 | |
1906 | // Reset various things that the optimizer might have decided it |
1907 | // "knows" about the coroutine function due to not seeing a return. |
1908 | F.removeFnAttr(Attribute::NoReturn); |
1909 | F.removeRetAttr(Attribute::NoAlias); |
1910 | F.removeRetAttr(Attribute::NonNull); |
1911 | |
1912 | // Allocate the frame. |
1913 | auto *Id = cast<AnyCoroIdRetconInst>(Val: Shape.CoroBegin->getId()); |
1914 | Value *RawFramePtr; |
1915 | if (Shape.RetconLowering.IsFrameInlineInStorage) { |
1916 | RawFramePtr = Id->getStorage(); |
1917 | } else { |
1918 | IRBuilder<> Builder(Id); |
1919 | |
1920 | // Determine the size of the frame. |
1921 | const DataLayout &DL = F.getParent()->getDataLayout(); |
1922 | auto Size = DL.getTypeAllocSize(Ty: Shape.FrameTy); |
1923 | |
1924 | // Allocate. We don't need to update the call graph node because we're |
1925 | // going to recompute it from scratch after splitting. |
1926 | // FIXME: pass the required alignment |
1927 | RawFramePtr = Shape.emitAlloc(Builder, Size: Builder.getInt64(C: Size), CG: nullptr); |
1928 | RawFramePtr = |
1929 | Builder.CreateBitCast(V: RawFramePtr, DestTy: Shape.CoroBegin->getType()); |
1930 | |
1931 | // Stash the allocated frame pointer in the continuation storage. |
1932 | Builder.CreateStore(Val: RawFramePtr, Ptr: Id->getStorage()); |
1933 | } |
1934 | |
1935 | // Map all uses of llvm.coro.begin to the allocated frame pointer. |
1936 | { |
1937 | // Make sure we don't invalidate Shape.FramePtr. |
1938 | TrackingVH<Value> Handle(Shape.FramePtr); |
1939 | Shape.CoroBegin->replaceAllUsesWith(V: RawFramePtr); |
1940 | Shape.FramePtr = Handle.getValPtr(); |
1941 | } |
1942 | |
1943 | // Create a unique return block. |
1944 | BasicBlock *ReturnBB = nullptr; |
1945 | SmallVector<PHINode *, 4> ReturnPHIs; |
1946 | |
1947 | // Create all the functions in order after the main function. |
1948 | auto NextF = std::next(x: F.getIterator()); |
1949 | |
1950 | // Create a continuation function for each of the suspend points. |
1951 | Clones.reserve(N: Shape.CoroSuspends.size()); |
1952 | for (size_t i = 0, e = Shape.CoroSuspends.size(); i != e; ++i) { |
1953 | auto Suspend = cast<CoroSuspendRetconInst>(Val: Shape.CoroSuspends[i]); |
1954 | |
1955 | // Create the clone declaration. |
1956 | auto Continuation = |
1957 | createCloneDeclaration(OrigF&: F, Shape, Suffix: ".resume." + Twine(i), InsertBefore: NextF, ActiveSuspend: nullptr); |
1958 | Clones.push_back(Elt: Continuation); |
1959 | |
1960 | // Insert a branch to the unified return block immediately before |
1961 | // the suspend point. |
1962 | auto SuspendBB = Suspend->getParent(); |
1963 | auto NewSuspendBB = SuspendBB->splitBasicBlock(I: Suspend); |
1964 | auto Branch = cast<BranchInst>(Val: SuspendBB->getTerminator()); |
1965 | |
1966 | // Create the unified return block. |
1967 | if (!ReturnBB) { |
1968 | // Place it before the first suspend. |
1969 | ReturnBB = |
1970 | BasicBlock::Create(Context&: F.getContext(), Name: "coro.return" , Parent: &F, InsertBefore: NewSuspendBB); |
1971 | Shape.RetconLowering.ReturnBlock = ReturnBB; |
1972 | |
1973 | IRBuilder<> Builder(ReturnBB); |
1974 | |
1975 | // Create PHIs for all the return values. |
1976 | assert(ReturnPHIs.empty()); |
1977 | |
1978 | // First, the continuation. |
1979 | ReturnPHIs.push_back(Elt: Builder.CreatePHI(Ty: Continuation->getType(), |
1980 | NumReservedValues: Shape.CoroSuspends.size())); |
1981 | |
1982 | // Next, all the directly-yielded values. |
1983 | for (auto *ResultTy : Shape.getRetconResultTypes()) |
1984 | ReturnPHIs.push_back( |
1985 | Elt: Builder.CreatePHI(Ty: ResultTy, NumReservedValues: Shape.CoroSuspends.size())); |
1986 | |
1987 | // Build the return value. |
1988 | auto RetTy = F.getReturnType(); |
1989 | |
1990 | // Cast the continuation value if necessary. |
1991 | // We can't rely on the types matching up because that type would |
1992 | // have to be infinite. |
1993 | auto CastedContinuationTy = |
1994 | (ReturnPHIs.size() == 1 ? RetTy : RetTy->getStructElementType(N: 0)); |
1995 | auto *CastedContinuation = |
1996 | Builder.CreateBitCast(V: ReturnPHIs[0], DestTy: CastedContinuationTy); |
1997 | |
1998 | Value *RetV; |
1999 | if (ReturnPHIs.size() == 1) { |
2000 | RetV = CastedContinuation; |
2001 | } else { |
2002 | RetV = PoisonValue::get(T: RetTy); |
2003 | RetV = Builder.CreateInsertValue(Agg: RetV, Val: CastedContinuation, Idxs: 0); |
2004 | for (size_t I = 1, E = ReturnPHIs.size(); I != E; ++I) |
2005 | RetV = Builder.CreateInsertValue(Agg: RetV, Val: ReturnPHIs[I], Idxs: I); |
2006 | } |
2007 | |
2008 | Builder.CreateRet(V: RetV); |
2009 | } |
2010 | |
2011 | // Branch to the return block. |
2012 | Branch->setSuccessor(idx: 0, NewSucc: ReturnBB); |
2013 | ReturnPHIs[0]->addIncoming(V: Continuation, BB: SuspendBB); |
2014 | size_t NextPHIIndex = 1; |
2015 | for (auto &VUse : Suspend->value_operands()) |
2016 | ReturnPHIs[NextPHIIndex++]->addIncoming(V: &*VUse, BB: SuspendBB); |
2017 | assert(NextPHIIndex == ReturnPHIs.size()); |
2018 | } |
2019 | |
2020 | assert(Clones.size() == Shape.CoroSuspends.size()); |
2021 | for (size_t i = 0, e = Shape.CoroSuspends.size(); i != e; ++i) { |
2022 | auto Suspend = Shape.CoroSuspends[i]; |
2023 | auto Clone = Clones[i]; |
2024 | |
2025 | CoroCloner(F, "resume." + Twine(i), Shape, Clone, Suspend).create(); |
2026 | } |
2027 | } |
2028 | |
2029 | namespace { |
2030 | class PrettyStackTraceFunction : public PrettyStackTraceEntry { |
2031 | Function &F; |
2032 | |
2033 | public: |
2034 | PrettyStackTraceFunction(Function &F) : F(F) {} |
2035 | void print(raw_ostream &OS) const override { |
2036 | OS << "While splitting coroutine " ; |
2037 | F.printAsOperand(O&: OS, /*print type*/ PrintType: false, M: F.getParent()); |
2038 | OS << "\n" ; |
2039 | } |
2040 | }; |
2041 | } // namespace |
2042 | |
2043 | static coro::Shape |
2044 | splitCoroutine(Function &F, SmallVectorImpl<Function *> &Clones, |
2045 | TargetTransformInfo &TTI, bool OptimizeFrame, |
2046 | std::function<bool(Instruction &)> MaterializableCallback) { |
2047 | PrettyStackTraceFunction prettyStackTrace(F); |
2048 | |
2049 | // The suspend-crossing algorithm in buildCoroutineFrame get tripped |
2050 | // up by uses in unreachable blocks, so remove them as a first pass. |
2051 | removeUnreachableBlocks(F); |
2052 | |
2053 | coro::Shape Shape(F, OptimizeFrame); |
2054 | if (!Shape.CoroBegin) |
2055 | return Shape; |
2056 | |
2057 | lowerAwaitSuspends(F, Shape); |
2058 | |
2059 | simplifySuspendPoints(Shape); |
2060 | buildCoroutineFrame(F, Shape, TTI, MaterializableCallback); |
2061 | replaceFrameSizeAndAlignment(Shape); |
2062 | |
2063 | // If there are no suspend points, no split required, just remove |
2064 | // the allocation and deallocation blocks, they are not needed. |
2065 | if (Shape.CoroSuspends.empty()) { |
2066 | handleNoSuspendCoroutine(Shape); |
2067 | } else { |
2068 | switch (Shape.ABI) { |
2069 | case coro::ABI::Switch: |
2070 | SwitchCoroutineSplitter::split(F, Shape, Clones, TTI); |
2071 | break; |
2072 | case coro::ABI::Async: |
2073 | splitAsyncCoroutine(F, Shape, Clones, TTI); |
2074 | break; |
2075 | case coro::ABI::Retcon: |
2076 | case coro::ABI::RetconOnce: |
2077 | splitRetconCoroutine(F, Shape, Clones); |
2078 | break; |
2079 | } |
2080 | } |
2081 | |
2082 | // Replace all the swifterror operations in the original function. |
2083 | // This invalidates SwiftErrorOps in the Shape. |
2084 | replaceSwiftErrorOps(F, Shape, VMap: nullptr); |
2085 | |
2086 | // Salvage debug intrinsics that point into the coroutine frame in the |
2087 | // original function. The Cloner has already salvaged debug info in the new |
2088 | // coroutine funclets. |
2089 | SmallDenseMap<Argument *, AllocaInst *, 4> ArgToAllocaMap; |
2090 | auto [DbgInsts, DbgVariableRecords] = collectDbgVariableIntrinsics(F); |
2091 | for (auto *DDI : DbgInsts) |
2092 | coro::salvageDebugInfo(ArgToAllocaMap, DVI&: *DDI, OptimizeFrame: Shape.OptimizeFrame, |
2093 | IsEntryPoint: false /*UseEntryValue*/); |
2094 | for (DbgVariableRecord *DVR : DbgVariableRecords) |
2095 | coro::salvageDebugInfo(ArgToAllocaMap, DVR&: *DVR, OptimizeFrame: Shape.OptimizeFrame, |
2096 | UseEntryValue: false /*UseEntryValue*/); |
2097 | return Shape; |
2098 | } |
2099 | |
2100 | /// Remove calls to llvm.coro.end in the original function. |
2101 | static void removeCoroEnds(const coro::Shape &Shape) { |
2102 | for (auto *End : Shape.CoroEnds) { |
2103 | replaceCoroEnd(End, Shape, FramePtr: Shape.FramePtr, /*in resume*/ InResume: false, CG: nullptr); |
2104 | } |
2105 | } |
2106 | |
2107 | static void updateCallGraphAfterCoroutineSplit( |
2108 | LazyCallGraph::Node &N, const coro::Shape &Shape, |
2109 | const SmallVectorImpl<Function *> &Clones, LazyCallGraph::SCC &C, |
2110 | LazyCallGraph &CG, CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR, |
2111 | FunctionAnalysisManager &FAM) { |
2112 | if (!Shape.CoroBegin) |
2113 | return; |
2114 | |
2115 | if (Shape.ABI != coro::ABI::Switch) |
2116 | removeCoroEnds(Shape); |
2117 | else { |
2118 | for (llvm::AnyCoroEndInst *End : Shape.CoroEnds) { |
2119 | auto &Context = End->getContext(); |
2120 | End->replaceAllUsesWith(V: ConstantInt::getFalse(Context)); |
2121 | End->eraseFromParent(); |
2122 | } |
2123 | } |
2124 | |
2125 | if (!Clones.empty()) { |
2126 | switch (Shape.ABI) { |
2127 | case coro::ABI::Switch: |
2128 | // Each clone in the Switch lowering is independent of the other clones. |
2129 | // Let the LazyCallGraph know about each one separately. |
2130 | for (Function *Clone : Clones) |
2131 | CG.addSplitFunction(OriginalFunction&: N.getFunction(), NewFunction&: *Clone); |
2132 | break; |
2133 | case coro::ABI::Async: |
2134 | case coro::ABI::Retcon: |
2135 | case coro::ABI::RetconOnce: |
2136 | // Each clone in the Async/Retcon lowering references of the other clones. |
2137 | // Let the LazyCallGraph know about all of them at once. |
2138 | if (!Clones.empty()) |
2139 | CG.addSplitRefRecursiveFunctions(OriginalFunction&: N.getFunction(), NewFunctions: Clones); |
2140 | break; |
2141 | } |
2142 | |
2143 | // Let the CGSCC infra handle the changes to the original function. |
2144 | updateCGAndAnalysisManagerForCGSCCPass(G&: CG, C, N, AM, UR, FAM); |
2145 | } |
2146 | |
2147 | // Do some cleanup and let the CGSCC infra see if we've cleaned up any edges |
2148 | // to the split functions. |
2149 | postSplitCleanup(F&: N.getFunction()); |
2150 | updateCGAndAnalysisManagerForFunctionPass(G&: CG, C, N, AM, UR, FAM); |
2151 | } |
2152 | |
2153 | /// Replace a call to llvm.coro.prepare.retcon. |
2154 | static void replacePrepare(CallInst *Prepare, LazyCallGraph &CG, |
2155 | LazyCallGraph::SCC &C) { |
2156 | auto CastFn = Prepare->getArgOperand(i: 0); // as an i8* |
2157 | auto Fn = CastFn->stripPointerCasts(); // as its original type |
2158 | |
2159 | // Attempt to peephole this pattern: |
2160 | // %0 = bitcast [[TYPE]] @some_function to i8* |
2161 | // %1 = call @llvm.coro.prepare.retcon(i8* %0) |
2162 | // %2 = bitcast %1 to [[TYPE]] |
2163 | // ==> |
2164 | // %2 = @some_function |
2165 | for (Use &U : llvm::make_early_inc_range(Range: Prepare->uses())) { |
2166 | // Look for bitcasts back to the original function type. |
2167 | auto *Cast = dyn_cast<BitCastInst>(Val: U.getUser()); |
2168 | if (!Cast || Cast->getType() != Fn->getType()) |
2169 | continue; |
2170 | |
2171 | // Replace and remove the cast. |
2172 | Cast->replaceAllUsesWith(V: Fn); |
2173 | Cast->eraseFromParent(); |
2174 | } |
2175 | |
2176 | // Replace any remaining uses with the function as an i8*. |
2177 | // This can never directly be a callee, so we don't need to update CG. |
2178 | Prepare->replaceAllUsesWith(V: CastFn); |
2179 | Prepare->eraseFromParent(); |
2180 | |
2181 | // Kill dead bitcasts. |
2182 | while (auto *Cast = dyn_cast<BitCastInst>(Val: CastFn)) { |
2183 | if (!Cast->use_empty()) |
2184 | break; |
2185 | CastFn = Cast->getOperand(i_nocapture: 0); |
2186 | Cast->eraseFromParent(); |
2187 | } |
2188 | } |
2189 | |
2190 | static bool replaceAllPrepares(Function *PrepareFn, LazyCallGraph &CG, |
2191 | LazyCallGraph::SCC &C) { |
2192 | bool Changed = false; |
2193 | for (Use &P : llvm::make_early_inc_range(Range: PrepareFn->uses())) { |
2194 | // Intrinsics can only be used in calls. |
2195 | auto *Prepare = cast<CallInst>(Val: P.getUser()); |
2196 | replacePrepare(Prepare, CG, C); |
2197 | Changed = true; |
2198 | } |
2199 | |
2200 | return Changed; |
2201 | } |
2202 | |
2203 | static void addPrepareFunction(const Module &M, |
2204 | SmallVectorImpl<Function *> &Fns, |
2205 | StringRef Name) { |
2206 | auto *PrepareFn = M.getFunction(Name); |
2207 | if (PrepareFn && !PrepareFn->use_empty()) |
2208 | Fns.push_back(Elt: PrepareFn); |
2209 | } |
2210 | |
2211 | CoroSplitPass::CoroSplitPass(bool OptimizeFrame) |
2212 | : MaterializableCallback(coro::defaultMaterializable), |
2213 | OptimizeFrame(OptimizeFrame) {} |
2214 | |
2215 | PreservedAnalyses CoroSplitPass::run(LazyCallGraph::SCC &C, |
2216 | CGSCCAnalysisManager &AM, |
2217 | LazyCallGraph &CG, CGSCCUpdateResult &UR) { |
2218 | // NB: One invariant of a valid LazyCallGraph::SCC is that it must contain a |
2219 | // non-zero number of nodes, so we assume that here and grab the first |
2220 | // node's function's module. |
2221 | Module &M = *C.begin()->getFunction().getParent(); |
2222 | auto &FAM = |
2223 | AM.getResult<FunctionAnalysisManagerCGSCCProxy>(IR&: C, ExtraArgs&: CG).getManager(); |
2224 | |
2225 | // Check for uses of llvm.coro.prepare.retcon/async. |
2226 | SmallVector<Function *, 2> PrepareFns; |
2227 | addPrepareFunction(M, Fns&: PrepareFns, Name: "llvm.coro.prepare.retcon" ); |
2228 | addPrepareFunction(M, Fns&: PrepareFns, Name: "llvm.coro.prepare.async" ); |
2229 | |
2230 | // Find coroutines for processing. |
2231 | SmallVector<LazyCallGraph::Node *> Coroutines; |
2232 | for (LazyCallGraph::Node &N : C) |
2233 | if (N.getFunction().isPresplitCoroutine()) |
2234 | Coroutines.push_back(Elt: &N); |
2235 | |
2236 | if (Coroutines.empty() && PrepareFns.empty()) |
2237 | return PreservedAnalyses::all(); |
2238 | |
2239 | if (Coroutines.empty()) { |
2240 | for (auto *PrepareFn : PrepareFns) { |
2241 | replaceAllPrepares(PrepareFn, CG, C); |
2242 | } |
2243 | } |
2244 | |
2245 | // Split all the coroutines. |
2246 | for (LazyCallGraph::Node *N : Coroutines) { |
2247 | Function &F = N->getFunction(); |
2248 | LLVM_DEBUG(dbgs() << "CoroSplit: Processing coroutine '" << F.getName() |
2249 | << "\n" ); |
2250 | F.setSplittedCoroutine(); |
2251 | |
2252 | SmallVector<Function *, 4> Clones; |
2253 | auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F); |
2254 | const coro::Shape Shape = |
2255 | splitCoroutine(F, Clones, TTI&: FAM.getResult<TargetIRAnalysis>(IR&: F), |
2256 | OptimizeFrame, MaterializableCallback); |
2257 | updateCallGraphAfterCoroutineSplit(N&: *N, Shape, Clones, C, CG, AM, UR, FAM); |
2258 | |
2259 | ORE.emit(RemarkBuilder: [&]() { |
2260 | return OptimizationRemark(DEBUG_TYPE, "CoroSplit" , &F) |
2261 | << "Split '" << ore::NV("function" , F.getName()) |
2262 | << "' (frame_size=" << ore::NV("frame_size" , Shape.FrameSize) |
2263 | << ", align=" << ore::NV("align" , Shape.FrameAlign.value()) << ")" ; |
2264 | }); |
2265 | |
2266 | if (!Shape.CoroSuspends.empty()) { |
2267 | // Run the CGSCC pipeline on the original and newly split functions. |
2268 | UR.CWorklist.insert(X: &C); |
2269 | for (Function *Clone : Clones) |
2270 | UR.CWorklist.insert(X: CG.lookupSCC(N&: CG.get(F&: *Clone))); |
2271 | } |
2272 | } |
2273 | |
2274 | if (!PrepareFns.empty()) { |
2275 | for (auto *PrepareFn : PrepareFns) { |
2276 | replaceAllPrepares(PrepareFn, CG, C); |
2277 | } |
2278 | } |
2279 | |
2280 | return PreservedAnalyses::none(); |
2281 | } |
2282 | |