| File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Transforms/Coroutines/CoroSplit.cpp |
| Warning: | line 463, column 19 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 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/SmallPtrSet.h" | |||
| 26 | #include "llvm/ADT/SmallVector.h" | |||
| 27 | #include "llvm/ADT/StringRef.h" | |||
| 28 | #include "llvm/ADT/Twine.h" | |||
| 29 | #include "llvm/Analysis/CFG.h" | |||
| 30 | #include "llvm/Analysis/CallGraph.h" | |||
| 31 | #include "llvm/Analysis/CallGraphSCCPass.h" | |||
| 32 | #include "llvm/Analysis/LazyCallGraph.h" | |||
| 33 | #include "llvm/IR/Argument.h" | |||
| 34 | #include "llvm/IR/Attributes.h" | |||
| 35 | #include "llvm/IR/BasicBlock.h" | |||
| 36 | #include "llvm/IR/CFG.h" | |||
| 37 | #include "llvm/IR/CallingConv.h" | |||
| 38 | #include "llvm/IR/Constants.h" | |||
| 39 | #include "llvm/IR/DataLayout.h" | |||
| 40 | #include "llvm/IR/DerivedTypes.h" | |||
| 41 | #include "llvm/IR/Dominators.h" | |||
| 42 | #include "llvm/IR/Function.h" | |||
| 43 | #include "llvm/IR/GlobalValue.h" | |||
| 44 | #include "llvm/IR/GlobalVariable.h" | |||
| 45 | #include "llvm/IR/IRBuilder.h" | |||
| 46 | #include "llvm/IR/InstIterator.h" | |||
| 47 | #include "llvm/IR/InstrTypes.h" | |||
| 48 | #include "llvm/IR/Instruction.h" | |||
| 49 | #include "llvm/IR/Instructions.h" | |||
| 50 | #include "llvm/IR/IntrinsicInst.h" | |||
| 51 | #include "llvm/IR/LLVMContext.h" | |||
| 52 | #include "llvm/IR/LegacyPassManager.h" | |||
| 53 | #include "llvm/IR/Module.h" | |||
| 54 | #include "llvm/IR/Type.h" | |||
| 55 | #include "llvm/IR/Value.h" | |||
| 56 | #include "llvm/IR/Verifier.h" | |||
| 57 | #include "llvm/InitializePasses.h" | |||
| 58 | #include "llvm/Pass.h" | |||
| 59 | #include "llvm/Support/Casting.h" | |||
| 60 | #include "llvm/Support/Debug.h" | |||
| 61 | #include "llvm/Support/PrettyStackTrace.h" | |||
| 62 | #include "llvm/Support/raw_ostream.h" | |||
| 63 | #include "llvm/Transforms/Scalar.h" | |||
| 64 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
| 65 | #include "llvm/Transforms/Utils/CallGraphUpdater.h" | |||
| 66 | #include "llvm/Transforms/Utils/Cloning.h" | |||
| 67 | #include "llvm/Transforms/Utils/Local.h" | |||
| 68 | #include "llvm/Transforms/Utils/ValueMapper.h" | |||
| 69 | #include <cassert> | |||
| 70 | #include <cstddef> | |||
| 71 | #include <cstdint> | |||
| 72 | #include <initializer_list> | |||
| 73 | #include <iterator> | |||
| 74 | ||||
| 75 | using namespace llvm; | |||
| 76 | ||||
| 77 | #define DEBUG_TYPE"coro-split" "coro-split" | |||
| 78 | ||||
| 79 | namespace { | |||
| 80 | ||||
| 81 | /// A little helper class for building | |||
| 82 | class CoroCloner { | |||
| 83 | public: | |||
| 84 | enum class Kind { | |||
| 85 | /// The shared resume function for a switch lowering. | |||
| 86 | SwitchResume, | |||
| 87 | ||||
| 88 | /// The shared unwind function for a switch lowering. | |||
| 89 | SwitchUnwind, | |||
| 90 | ||||
| 91 | /// The shared cleanup function for a switch lowering. | |||
| 92 | SwitchCleanup, | |||
| 93 | ||||
| 94 | /// An individual continuation function. | |||
| 95 | Continuation, | |||
| 96 | ||||
| 97 | /// An async resume function. | |||
| 98 | Async, | |||
| 99 | }; | |||
| 100 | ||||
| 101 | private: | |||
| 102 | Function &OrigF; | |||
| 103 | Function *NewF; | |||
| 104 | const Twine &Suffix; | |||
| 105 | coro::Shape &Shape; | |||
| 106 | Kind FKind; | |||
| 107 | ValueToValueMapTy VMap; | |||
| 108 | IRBuilder<> Builder; | |||
| 109 | Value *NewFramePtr = nullptr; | |||
| 110 | ||||
| 111 | /// The active suspend instruction; meaningful only for continuation and async | |||
| 112 | /// ABIs. | |||
| 113 | AnyCoroSuspendInst *ActiveSuspend = nullptr; | |||
| 114 | ||||
| 115 | public: | |||
| 116 | /// Create a cloner for a switch lowering. | |||
| 117 | CoroCloner(Function &OrigF, const Twine &Suffix, coro::Shape &Shape, | |||
| 118 | Kind FKind) | |||
| 119 | : OrigF(OrigF), NewF(nullptr), Suffix(Suffix), Shape(Shape), | |||
| 120 | FKind(FKind), Builder(OrigF.getContext()) { | |||
| 121 | assert(Shape.ABI == coro::ABI::Switch)((void)0); | |||
| 122 | } | |||
| 123 | ||||
| 124 | /// Create a cloner for a continuation lowering. | |||
| 125 | CoroCloner(Function &OrigF, const Twine &Suffix, coro::Shape &Shape, | |||
| 126 | Function *NewF, AnyCoroSuspendInst *ActiveSuspend) | |||
| 127 | : OrigF(OrigF), NewF(NewF), Suffix(Suffix), Shape(Shape), | |||
| 128 | FKind(Shape.ABI == coro::ABI::Async ? Kind::Async : Kind::Continuation), | |||
| 129 | Builder(OrigF.getContext()), ActiveSuspend(ActiveSuspend) { | |||
| 130 | assert(Shape.ABI == coro::ABI::Retcon ||((void)0) | |||
| 131 | Shape.ABI == coro::ABI::RetconOnce || Shape.ABI == coro::ABI::Async)((void)0); | |||
| 132 | assert(NewF && "need existing function for continuation")((void)0); | |||
| 133 | assert(ActiveSuspend && "need active suspend point for continuation")((void)0); | |||
| 134 | } | |||
| 135 | ||||
| 136 | Function *getFunction() const { | |||
| 137 | assert(NewF != nullptr && "declaration not yet set")((void)0); | |||
| 138 | return NewF; | |||
| 139 | } | |||
| 140 | ||||
| 141 | void create(); | |||
| 142 | ||||
| 143 | private: | |||
| 144 | bool isSwitchDestroyFunction() { | |||
| 145 | switch (FKind) { | |||
| 146 | case Kind::Async: | |||
| 147 | case Kind::Continuation: | |||
| 148 | case Kind::SwitchResume: | |||
| 149 | return false; | |||
| 150 | case Kind::SwitchUnwind: | |||
| 151 | case Kind::SwitchCleanup: | |||
| 152 | return true; | |||
| 153 | } | |||
| 154 | llvm_unreachable("Unknown CoroCloner::Kind enum")__builtin_unreachable(); | |||
| 155 | } | |||
| 156 | ||||
| 157 | void replaceEntryBlock(); | |||
| 158 | Value *deriveNewFramePointer(); | |||
| 159 | void replaceRetconOrAsyncSuspendUses(); | |||
| 160 | void replaceCoroSuspends(); | |||
| 161 | void replaceCoroEnds(); | |||
| 162 | void replaceSwiftErrorOps(); | |||
| 163 | void salvageDebugInfo(); | |||
| 164 | void handleFinalSuspend(); | |||
| 165 | }; | |||
| 166 | ||||
| 167 | } // end anonymous namespace | |||
| 168 | ||||
| 169 | static void maybeFreeRetconStorage(IRBuilder<> &Builder, | |||
| 170 | const coro::Shape &Shape, Value *FramePtr, | |||
| 171 | CallGraph *CG) { | |||
| 172 | assert(Shape.ABI == coro::ABI::Retcon ||((void)0) | |||
| 173 | Shape.ABI == coro::ABI::RetconOnce)((void)0); | |||
| 174 | if (Shape.RetconLowering.IsFrameInlineInStorage) | |||
| 175 | return; | |||
| 176 | ||||
| 177 | Shape.emitDealloc(Builder, FramePtr, CG); | |||
| 178 | } | |||
| 179 | ||||
| 180 | /// Replace an llvm.coro.end.async. | |||
| 181 | /// Will inline the must tail call function call if there is one. | |||
| 182 | /// \returns true if cleanup of the coro.end block is needed, false otherwise. | |||
| 183 | static bool replaceCoroEndAsync(AnyCoroEndInst *End) { | |||
| 184 | IRBuilder<> Builder(End); | |||
| 185 | ||||
| 186 | auto *EndAsync = dyn_cast<CoroAsyncEndInst>(End); | |||
| 187 | if (!EndAsync) { | |||
| 188 | Builder.CreateRetVoid(); | |||
| 189 | return true /*needs cleanup of coro.end block*/; | |||
| 190 | } | |||
| 191 | ||||
| 192 | auto *MustTailCallFunc = EndAsync->getMustTailCallFunction(); | |||
| 193 | if (!MustTailCallFunc) { | |||
| 194 | Builder.CreateRetVoid(); | |||
| 195 | return true /*needs cleanup of coro.end block*/; | |||
| 196 | } | |||
| 197 | ||||
| 198 | // Move the must tail call from the predecessor block into the end block. | |||
| 199 | auto *CoroEndBlock = End->getParent(); | |||
| 200 | auto *MustTailCallFuncBlock = CoroEndBlock->getSinglePredecessor(); | |||
| 201 | assert(MustTailCallFuncBlock && "Must have a single predecessor block")((void)0); | |||
| 202 | auto It = MustTailCallFuncBlock->getTerminator()->getIterator(); | |||
| 203 | auto *MustTailCall = cast<CallInst>(&*std::prev(It)); | |||
| 204 | CoroEndBlock->getInstList().splice( | |||
| 205 | End->getIterator(), MustTailCallFuncBlock->getInstList(), MustTailCall); | |||
| 206 | ||||
| 207 | // Insert the return instruction. | |||
| 208 | Builder.SetInsertPoint(End); | |||
| 209 | Builder.CreateRetVoid(); | |||
| 210 | InlineFunctionInfo FnInfo; | |||
| 211 | ||||
| 212 | // Remove the rest of the block, by splitting it into an unreachable block. | |||
| 213 | auto *BB = End->getParent(); | |||
| 214 | BB->splitBasicBlock(End); | |||
| 215 | BB->getTerminator()->eraseFromParent(); | |||
| 216 | ||||
| 217 | auto InlineRes = InlineFunction(*MustTailCall, FnInfo); | |||
| 218 | assert(InlineRes.isSuccess() && "Expected inlining to succeed")((void)0); | |||
| 219 | (void)InlineRes; | |||
| 220 | ||||
| 221 | // We have cleaned up the coro.end block above. | |||
| 222 | return false; | |||
| 223 | } | |||
| 224 | ||||
| 225 | /// Replace a non-unwind call to llvm.coro.end. | |||
| 226 | static void replaceFallthroughCoroEnd(AnyCoroEndInst *End, | |||
| 227 | const coro::Shape &Shape, Value *FramePtr, | |||
| 228 | bool InResume, CallGraph *CG) { | |||
| 229 | // Start inserting right before the coro.end. | |||
| 230 | IRBuilder<> Builder(End); | |||
| 231 | ||||
| 232 | // Create the return instruction. | |||
| 233 | switch (Shape.ABI) { | |||
| 234 | // The cloned functions in switch-lowering always return void. | |||
| 235 | case coro::ABI::Switch: | |||
| 236 | // coro.end doesn't immediately end the coroutine in the main function | |||
| 237 | // in this lowering, because we need to deallocate the coroutine. | |||
| 238 | if (!InResume) | |||
| 239 | return; | |||
| 240 | Builder.CreateRetVoid(); | |||
| 241 | break; | |||
| 242 | ||||
| 243 | // In async lowering this returns. | |||
| 244 | case coro::ABI::Async: { | |||
| 245 | bool CoroEndBlockNeedsCleanup = replaceCoroEndAsync(End); | |||
| 246 | if (!CoroEndBlockNeedsCleanup) | |||
| 247 | return; | |||
| 248 | break; | |||
| 249 | } | |||
| 250 | ||||
| 251 | // In unique continuation lowering, the continuations always return void. | |||
| 252 | // But we may have implicitly allocated storage. | |||
| 253 | case coro::ABI::RetconOnce: | |||
| 254 | maybeFreeRetconStorage(Builder, Shape, FramePtr, CG); | |||
| 255 | Builder.CreateRetVoid(); | |||
| 256 | break; | |||
| 257 | ||||
| 258 | // In non-unique continuation lowering, we signal completion by returning | |||
| 259 | // a null continuation. | |||
| 260 | case coro::ABI::Retcon: { | |||
| 261 | maybeFreeRetconStorage(Builder, Shape, FramePtr, CG); | |||
| 262 | auto RetTy = Shape.getResumeFunctionType()->getReturnType(); | |||
| 263 | auto RetStructTy = dyn_cast<StructType>(RetTy); | |||
| 264 | PointerType *ContinuationTy = | |||
| 265 | cast<PointerType>(RetStructTy ? RetStructTy->getElementType(0) : RetTy); | |||
| 266 | ||||
| 267 | Value *ReturnValue = ConstantPointerNull::get(ContinuationTy); | |||
| 268 | if (RetStructTy) { | |||
| 269 | ReturnValue = Builder.CreateInsertValue(UndefValue::get(RetStructTy), | |||
| 270 | ReturnValue, 0); | |||
| 271 | } | |||
| 272 | Builder.CreateRet(ReturnValue); | |||
| 273 | break; | |||
| 274 | } | |||
| 275 | } | |||
| 276 | ||||
| 277 | // Remove the rest of the block, by splitting it into an unreachable block. | |||
| 278 | auto *BB = End->getParent(); | |||
| 279 | BB->splitBasicBlock(End); | |||
| 280 | BB->getTerminator()->eraseFromParent(); | |||
| 281 | } | |||
| 282 | ||||
| 283 | /// Replace an unwind call to llvm.coro.end. | |||
| 284 | static void replaceUnwindCoroEnd(AnyCoroEndInst *End, const coro::Shape &Shape, | |||
| 285 | Value *FramePtr, bool InResume, | |||
| 286 | CallGraph *CG) { | |||
| 287 | IRBuilder<> Builder(End); | |||
| 288 | ||||
| 289 | switch (Shape.ABI) { | |||
| 290 | // In switch-lowering, this does nothing in the main function. | |||
| 291 | case coro::ABI::Switch: | |||
| 292 | if (!InResume) | |||
| 293 | return; | |||
| 294 | break; | |||
| 295 | // In async lowering this does nothing. | |||
| 296 | case coro::ABI::Async: | |||
| 297 | break; | |||
| 298 | // In continuation-lowering, this frees the continuation storage. | |||
| 299 | case coro::ABI::Retcon: | |||
| 300 | case coro::ABI::RetconOnce: | |||
| 301 | maybeFreeRetconStorage(Builder, Shape, FramePtr, CG); | |||
| 302 | break; | |||
| 303 | } | |||
| 304 | ||||
| 305 | // If coro.end has an associated bundle, add cleanupret instruction. | |||
| 306 | if (auto Bundle = End->getOperandBundle(LLVMContext::OB_funclet)) { | |||
| 307 | auto *FromPad = cast<CleanupPadInst>(Bundle->Inputs[0]); | |||
| 308 | auto *CleanupRet = Builder.CreateCleanupRet(FromPad, nullptr); | |||
| 309 | End->getParent()->splitBasicBlock(End); | |||
| 310 | CleanupRet->getParent()->getTerminator()->eraseFromParent(); | |||
| 311 | } | |||
| 312 | } | |||
| 313 | ||||
| 314 | static void replaceCoroEnd(AnyCoroEndInst *End, const coro::Shape &Shape, | |||
| 315 | Value *FramePtr, bool InResume, CallGraph *CG) { | |||
| 316 | if (End->isUnwind()) | |||
| 317 | replaceUnwindCoroEnd(End, Shape, FramePtr, InResume, CG); | |||
| 318 | else | |||
| 319 | replaceFallthroughCoroEnd(End, Shape, FramePtr, InResume, CG); | |||
| 320 | ||||
| 321 | auto &Context = End->getContext(); | |||
| 322 | End->replaceAllUsesWith(InResume ? ConstantInt::getTrue(Context) | |||
| 323 | : ConstantInt::getFalse(Context)); | |||
| 324 | End->eraseFromParent(); | |||
| 325 | } | |||
| 326 | ||||
| 327 | // Create an entry block for a resume function with a switch that will jump to | |||
| 328 | // suspend points. | |||
| 329 | static void createResumeEntryBlock(Function &F, coro::Shape &Shape) { | |||
| 330 | assert(Shape.ABI == coro::ABI::Switch)((void)0); | |||
| 331 | LLVMContext &C = F.getContext(); | |||
| 332 | ||||
| 333 | // resume.entry: | |||
| 334 | // %index.addr = getelementptr inbounds %f.Frame, %f.Frame* %FramePtr, i32 0, | |||
| 335 | // i32 2 | |||
| 336 | // % index = load i32, i32* %index.addr | |||
| 337 | // switch i32 %index, label %unreachable [ | |||
| 338 | // i32 0, label %resume.0 | |||
| 339 | // i32 1, label %resume.1 | |||
| 340 | // ... | |||
| 341 | // ] | |||
| 342 | ||||
| 343 | auto *NewEntry = BasicBlock::Create(C, "resume.entry", &F); | |||
| 344 | auto *UnreachBB = BasicBlock::Create(C, "unreachable", &F); | |||
| 345 | ||||
| 346 | IRBuilder<> Builder(NewEntry); | |||
| 347 | auto *FramePtr = Shape.FramePtr; | |||
| 348 | auto *FrameTy = Shape.FrameTy; | |||
| 349 | auto *GepIndex = Builder.CreateStructGEP( | |||
| 350 | FrameTy, FramePtr, Shape.getSwitchIndexField(), "index.addr"); | |||
| 351 | auto *Index = Builder.CreateLoad(Shape.getIndexType(), GepIndex, "index"); | |||
| 352 | auto *Switch = | |||
| 353 | Builder.CreateSwitch(Index, UnreachBB, Shape.CoroSuspends.size()); | |||
| 354 | Shape.SwitchLowering.ResumeSwitch = Switch; | |||
| 355 | ||||
| 356 | size_t SuspendIndex = 0; | |||
| 357 | for (auto *AnyS : Shape.CoroSuspends) { | |||
| 358 | auto *S = cast<CoroSuspendInst>(AnyS); | |||
| 359 | ConstantInt *IndexVal = Shape.getIndex(SuspendIndex); | |||
| 360 | ||||
| 361 | // Replace CoroSave with a store to Index: | |||
| 362 | // %index.addr = getelementptr %f.frame... (index field number) | |||
| 363 | // store i32 0, i32* %index.addr1 | |||
| 364 | auto *Save = S->getCoroSave(); | |||
| 365 | Builder.SetInsertPoint(Save); | |||
| 366 | if (S->isFinal()) { | |||
| 367 | // Final suspend point is represented by storing zero in ResumeFnAddr. | |||
| 368 | auto *GepIndex = Builder.CreateStructGEP(FrameTy, FramePtr, | |||
| 369 | coro::Shape::SwitchFieldIndex::Resume, | |||
| 370 | "ResumeFn.addr"); | |||
| 371 | auto *NullPtr = ConstantPointerNull::get(cast<PointerType>( | |||
| 372 | FrameTy->getTypeAtIndex(coro::Shape::SwitchFieldIndex::Resume))); | |||
| 373 | Builder.CreateStore(NullPtr, GepIndex); | |||
| 374 | } else { | |||
| 375 | auto *GepIndex = Builder.CreateStructGEP( | |||
| 376 | FrameTy, FramePtr, Shape.getSwitchIndexField(), "index.addr"); | |||
| 377 | Builder.CreateStore(IndexVal, GepIndex); | |||
| 378 | } | |||
| 379 | Save->replaceAllUsesWith(ConstantTokenNone::get(C)); | |||
| 380 | Save->eraseFromParent(); | |||
| 381 | ||||
| 382 | // Split block before and after coro.suspend and add a jump from an entry | |||
| 383 | // switch: | |||
| 384 | // | |||
| 385 | // whateverBB: | |||
| 386 | // whatever | |||
| 387 | // %0 = call i8 @llvm.coro.suspend(token none, i1 false) | |||
| 388 | // switch i8 %0, label %suspend[i8 0, label %resume | |||
| 389 | // i8 1, label %cleanup] | |||
| 390 | // becomes: | |||
| 391 | // | |||
| 392 | // whateverBB: | |||
| 393 | // whatever | |||
| 394 | // br label %resume.0.landing | |||
| 395 | // | |||
| 396 | // resume.0: ; <--- jump from the switch in the resume.entry | |||
| 397 | // %0 = tail call i8 @llvm.coro.suspend(token none, i1 false) | |||
| 398 | // br label %resume.0.landing | |||
| 399 | // | |||
| 400 | // resume.0.landing: | |||
| 401 | // %1 = phi i8[-1, %whateverBB], [%0, %resume.0] | |||
| 402 | // switch i8 % 1, label %suspend [i8 0, label %resume | |||
| 403 | // i8 1, label %cleanup] | |||
| 404 | ||||
| 405 | auto *SuspendBB = S->getParent(); | |||
| 406 | auto *ResumeBB = | |||
| 407 | SuspendBB->splitBasicBlock(S, "resume." + Twine(SuspendIndex)); | |||
| 408 | auto *LandingBB = ResumeBB->splitBasicBlock( | |||
| 409 | S->getNextNode(), ResumeBB->getName() + Twine(".landing")); | |||
| 410 | Switch->addCase(IndexVal, ResumeBB); | |||
| 411 | ||||
| 412 | cast<BranchInst>(SuspendBB->getTerminator())->setSuccessor(0, LandingBB); | |||
| 413 | auto *PN = PHINode::Create(Builder.getInt8Ty(), 2, "", &LandingBB->front()); | |||
| 414 | S->replaceAllUsesWith(PN); | |||
| 415 | PN->addIncoming(Builder.getInt8(-1), SuspendBB); | |||
| 416 | PN->addIncoming(S, ResumeBB); | |||
| 417 | ||||
| 418 | ++SuspendIndex; | |||
| 419 | } | |||
| 420 | ||||
| 421 | Builder.SetInsertPoint(UnreachBB); | |||
| 422 | Builder.CreateUnreachable(); | |||
| 423 | ||||
| 424 | Shape.SwitchLowering.ResumeEntryBlock = NewEntry; | |||
| 425 | } | |||
| 426 | ||||
| 427 | ||||
| 428 | // Rewrite final suspend point handling. We do not use suspend index to | |||
| 429 | // represent the final suspend point. Instead we zero-out ResumeFnAddr in the | |||
| 430 | // coroutine frame, since it is undefined behavior to resume a coroutine | |||
| 431 | // suspended at the final suspend point. Thus, in the resume function, we can | |||
| 432 | // simply remove the last case (when coro::Shape is built, the final suspend | |||
| 433 | // point (if present) is always the last element of CoroSuspends array). | |||
| 434 | // In the destroy function, we add a code sequence to check if ResumeFnAddress | |||
| 435 | // is Null, and if so, jump to the appropriate label to handle cleanup from the | |||
| 436 | // final suspend point. | |||
| 437 | void CoroCloner::handleFinalSuspend() { | |||
| 438 | assert(Shape.ABI == coro::ABI::Switch &&((void)0) | |||
| 439 | Shape.SwitchLowering.HasFinalSuspend)((void)0); | |||
| 440 | auto *Switch = cast<SwitchInst>(VMap[Shape.SwitchLowering.ResumeSwitch]); | |||
| 441 | auto FinalCaseIt = std::prev(Switch->case_end()); | |||
| 442 | BasicBlock *ResumeBB = FinalCaseIt->getCaseSuccessor(); | |||
| 443 | Switch->removeCase(FinalCaseIt); | |||
| 444 | if (isSwitchDestroyFunction()) { | |||
| 445 | BasicBlock *OldSwitchBB = Switch->getParent(); | |||
| 446 | auto *NewSwitchBB = OldSwitchBB->splitBasicBlock(Switch, "Switch"); | |||
| 447 | Builder.SetInsertPoint(OldSwitchBB->getTerminator()); | |||
| 448 | auto *GepIndex = Builder.CreateStructGEP(Shape.FrameTy, NewFramePtr, | |||
| 449 | coro::Shape::SwitchFieldIndex::Resume, | |||
| 450 | "ResumeFn.addr"); | |||
| 451 | auto *Load = Builder.CreateLoad(Shape.getSwitchResumePointerType(), | |||
| 452 | GepIndex); | |||
| 453 | auto *Cond = Builder.CreateIsNull(Load); | |||
| 454 | Builder.CreateCondBr(Cond, ResumeBB, NewSwitchBB); | |||
| 455 | OldSwitchBB->getTerminator()->eraseFromParent(); | |||
| 456 | } | |||
| 457 | } | |||
| 458 | ||||
| 459 | static FunctionType * | |||
| 460 | getFunctionTypeFromAsyncSuspend(AnyCoroSuspendInst *Suspend) { | |||
| 461 | auto *AsyncSuspend = cast<CoroSuspendAsyncInst>(Suspend); | |||
| 462 | auto *StructTy = cast<StructType>(AsyncSuspend->getType()); | |||
| 463 | auto &Context = Suspend->getParent()->getParent()->getContext(); | |||
| ||||
| 464 | auto *VoidTy = Type::getVoidTy(Context); | |||
| 465 | return FunctionType::get(VoidTy, StructTy->elements(), false); | |||
| 466 | } | |||
| 467 | ||||
| 468 | static Function *createCloneDeclaration(Function &OrigF, coro::Shape &Shape, | |||
| 469 | const Twine &Suffix, | |||
| 470 | Module::iterator InsertBefore, | |||
| 471 | AnyCoroSuspendInst *ActiveSuspend) { | |||
| 472 | Module *M = OrigF.getParent(); | |||
| 473 | auto *FnTy = (Shape.ABI != coro::ABI::Async) | |||
| 474 | ? Shape.getResumeFunctionType() | |||
| 475 | : getFunctionTypeFromAsyncSuspend(ActiveSuspend); | |||
| 476 | ||||
| 477 | Function *NewF = | |||
| 478 | Function::Create(FnTy, GlobalValue::LinkageTypes::InternalLinkage, | |||
| 479 | OrigF.getName() + Suffix); | |||
| 480 | if (Shape.ABI != coro::ABI::Async) | |||
| 481 | NewF->addParamAttr(0, Attribute::NonNull); | |||
| 482 | ||||
| 483 | // For the async lowering ABI we can't guarantee that the context argument is | |||
| 484 | // not access via a different pointer not based on the argument. | |||
| 485 | if (Shape.ABI != coro::ABI::Async) | |||
| 486 | NewF->addParamAttr(0, Attribute::NoAlias); | |||
| 487 | ||||
| 488 | M->getFunctionList().insert(InsertBefore, NewF); | |||
| 489 | ||||
| 490 | return NewF; | |||
| 491 | } | |||
| 492 | ||||
| 493 | /// Replace uses of the active llvm.coro.suspend.retcon/async call with the | |||
| 494 | /// arguments to the continuation function. | |||
| 495 | /// | |||
| 496 | /// This assumes that the builder has a meaningful insertion point. | |||
| 497 | void CoroCloner::replaceRetconOrAsyncSuspendUses() { | |||
| 498 | assert(Shape.ABI == coro::ABI::Retcon || Shape.ABI == coro::ABI::RetconOnce ||((void)0) | |||
| 499 | Shape.ABI == coro::ABI::Async)((void)0); | |||
| 500 | ||||
| 501 | auto NewS = VMap[ActiveSuspend]; | |||
| 502 | if (NewS->use_empty()) return; | |||
| 503 | ||||
| 504 | // Copy out all the continuation arguments after the buffer pointer into | |||
| 505 | // an easily-indexed data structure for convenience. | |||
| 506 | SmallVector<Value*, 8> Args; | |||
| 507 | // The async ABI includes all arguments -- including the first argument. | |||
| 508 | bool IsAsyncABI = Shape.ABI == coro::ABI::Async; | |||
| 509 | for (auto I = IsAsyncABI ? NewF->arg_begin() : std::next(NewF->arg_begin()), | |||
| 510 | E = NewF->arg_end(); | |||
| 511 | I != E; ++I) | |||
| 512 | Args.push_back(&*I); | |||
| 513 | ||||
| 514 | // If the suspend returns a single scalar value, we can just do a simple | |||
| 515 | // replacement. | |||
| 516 | if (!isa<StructType>(NewS->getType())) { | |||
| 517 | assert(Args.size() == 1)((void)0); | |||
| 518 | NewS->replaceAllUsesWith(Args.front()); | |||
| 519 | return; | |||
| 520 | } | |||
| 521 | ||||
| 522 | // Try to peephole extracts of an aggregate return. | |||
| 523 | for (auto UI = NewS->use_begin(), UE = NewS->use_end(); UI != UE; ) { | |||
| 524 | auto EVI = dyn_cast<ExtractValueInst>((UI++)->getUser()); | |||
| 525 | if (!EVI || EVI->getNumIndices() != 1) | |||
| 526 | continue; | |||
| 527 | ||||
| 528 | EVI->replaceAllUsesWith(Args[EVI->getIndices().front()]); | |||
| 529 | EVI->eraseFromParent(); | |||
| 530 | } | |||
| 531 | ||||
| 532 | // If we have no remaining uses, we're done. | |||
| 533 | if (NewS->use_empty()) return; | |||
| 534 | ||||
| 535 | // Otherwise, we need to create an aggregate. | |||
| 536 | Value *Agg = UndefValue::get(NewS->getType()); | |||
| 537 | for (size_t I = 0, E = Args.size(); I != E; ++I) | |||
| 538 | Agg = Builder.CreateInsertValue(Agg, Args[I], I); | |||
| 539 | ||||
| 540 | NewS->replaceAllUsesWith(Agg); | |||
| 541 | } | |||
| 542 | ||||
| 543 | void CoroCloner::replaceCoroSuspends() { | |||
| 544 | Value *SuspendResult; | |||
| 545 | ||||
| 546 | switch (Shape.ABI) { | |||
| 547 | // In switch lowering, replace coro.suspend with the appropriate value | |||
| 548 | // for the type of function we're extracting. | |||
| 549 | // Replacing coro.suspend with (0) will result in control flow proceeding to | |||
| 550 | // a resume label associated with a suspend point, replacing it with (1) will | |||
| 551 | // result in control flow proceeding to a cleanup label associated with this | |||
| 552 | // suspend point. | |||
| 553 | case coro::ABI::Switch: | |||
| 554 | SuspendResult = Builder.getInt8(isSwitchDestroyFunction() ? 1 : 0); | |||
| 555 | break; | |||
| 556 | ||||
| 557 | // In async lowering there are no uses of the result. | |||
| 558 | case coro::ABI::Async: | |||
| 559 | return; | |||
| 560 | ||||
| 561 | // In returned-continuation lowering, the arguments from earlier | |||
| 562 | // continuations are theoretically arbitrary, and they should have been | |||
| 563 | // spilled. | |||
| 564 | case coro::ABI::RetconOnce: | |||
| 565 | case coro::ABI::Retcon: | |||
| 566 | return; | |||
| 567 | } | |||
| 568 | ||||
| 569 | for (AnyCoroSuspendInst *CS : Shape.CoroSuspends) { | |||
| 570 | // The active suspend was handled earlier. | |||
| 571 | if (CS == ActiveSuspend) continue; | |||
| 572 | ||||
| 573 | auto *MappedCS = cast<AnyCoroSuspendInst>(VMap[CS]); | |||
| 574 | MappedCS->replaceAllUsesWith(SuspendResult); | |||
| 575 | MappedCS->eraseFromParent(); | |||
| 576 | } | |||
| 577 | } | |||
| 578 | ||||
| 579 | void CoroCloner::replaceCoroEnds() { | |||
| 580 | for (AnyCoroEndInst *CE : Shape.CoroEnds) { | |||
| 581 | // We use a null call graph because there's no call graph node for | |||
| 582 | // the cloned function yet. We'll just be rebuilding that later. | |||
| 583 | auto *NewCE = cast<AnyCoroEndInst>(VMap[CE]); | |||
| 584 | replaceCoroEnd(NewCE, Shape, NewFramePtr, /*in resume*/ true, nullptr); | |||
| 585 | } | |||
| 586 | } | |||
| 587 | ||||
| 588 | static void replaceSwiftErrorOps(Function &F, coro::Shape &Shape, | |||
| 589 | ValueToValueMapTy *VMap) { | |||
| 590 | if (Shape.ABI == coro::ABI::Async && Shape.CoroSuspends.empty()) | |||
| 591 | return; | |||
| 592 | Value *CachedSlot = nullptr; | |||
| 593 | auto getSwiftErrorSlot = [&](Type *ValueTy) -> Value * { | |||
| 594 | if (CachedSlot) { | |||
| 595 | assert(CachedSlot->getType()->getPointerElementType() == ValueTy &&((void)0) | |||
| 596 | "multiple swifterror slots in function with different types")((void)0); | |||
| 597 | return CachedSlot; | |||
| 598 | } | |||
| 599 | ||||
| 600 | // Check if the function has a swifterror argument. | |||
| 601 | for (auto &Arg : F.args()) { | |||
| 602 | if (Arg.isSwiftError()) { | |||
| 603 | CachedSlot = &Arg; | |||
| 604 | assert(Arg.getType()->getPointerElementType() == ValueTy &&((void)0) | |||
| 605 | "swifterror argument does not have expected type")((void)0); | |||
| 606 | return &Arg; | |||
| 607 | } | |||
| 608 | } | |||
| 609 | ||||
| 610 | // Create a swifterror alloca. | |||
| 611 | IRBuilder<> Builder(F.getEntryBlock().getFirstNonPHIOrDbg()); | |||
| 612 | auto Alloca = Builder.CreateAlloca(ValueTy); | |||
| 613 | Alloca->setSwiftError(true); | |||
| 614 | ||||
| 615 | CachedSlot = Alloca; | |||
| 616 | return Alloca; | |||
| 617 | }; | |||
| 618 | ||||
| 619 | for (CallInst *Op : Shape.SwiftErrorOps) { | |||
| 620 | auto MappedOp = VMap ? cast<CallInst>((*VMap)[Op]) : Op; | |||
| 621 | IRBuilder<> Builder(MappedOp); | |||
| 622 | ||||
| 623 | // If there are no arguments, this is a 'get' operation. | |||
| 624 | Value *MappedResult; | |||
| 625 | if (Op->getNumArgOperands() == 0) { | |||
| 626 | auto ValueTy = Op->getType(); | |||
| 627 | auto Slot = getSwiftErrorSlot(ValueTy); | |||
| 628 | MappedResult = Builder.CreateLoad(ValueTy, Slot); | |||
| 629 | } else { | |||
| 630 | assert(Op->getNumArgOperands() == 1)((void)0); | |||
| 631 | auto Value = MappedOp->getArgOperand(0); | |||
| 632 | auto ValueTy = Value->getType(); | |||
| 633 | auto Slot = getSwiftErrorSlot(ValueTy); | |||
| 634 | Builder.CreateStore(Value, Slot); | |||
| 635 | MappedResult = Slot; | |||
| 636 | } | |||
| 637 | ||||
| 638 | MappedOp->replaceAllUsesWith(MappedResult); | |||
| 639 | MappedOp->eraseFromParent(); | |||
| 640 | } | |||
| 641 | ||||
| 642 | // If we're updating the original function, we've invalidated SwiftErrorOps. | |||
| 643 | if (VMap == nullptr) { | |||
| 644 | Shape.SwiftErrorOps.clear(); | |||
| 645 | } | |||
| 646 | } | |||
| 647 | ||||
| 648 | void CoroCloner::replaceSwiftErrorOps() { | |||
| 649 | ::replaceSwiftErrorOps(*NewF, Shape, &VMap); | |||
| 650 | } | |||
| 651 | ||||
| 652 | void CoroCloner::salvageDebugInfo() { | |||
| 653 | SmallVector<DbgVariableIntrinsic *, 8> Worklist; | |||
| 654 | SmallDenseMap<llvm::Value *, llvm::AllocaInst *, 4> DbgPtrAllocaCache; | |||
| 655 | for (auto &BB : *NewF) | |||
| 656 | for (auto &I : BB) | |||
| 657 | if (auto *DVI = dyn_cast<DbgVariableIntrinsic>(&I)) | |||
| 658 | Worklist.push_back(DVI); | |||
| 659 | for (DbgVariableIntrinsic *DVI : Worklist) | |||
| 660 | coro::salvageDebugInfo(DbgPtrAllocaCache, DVI, Shape.ReuseFrameSlot); | |||
| 661 | ||||
| 662 | // Remove all salvaged dbg.declare intrinsics that became | |||
| 663 | // either unreachable or stale due to the CoroSplit transformation. | |||
| 664 | DominatorTree DomTree(*NewF); | |||
| 665 | auto IsUnreachableBlock = [&](BasicBlock *BB) { | |||
| 666 | return !isPotentiallyReachable(&NewF->getEntryBlock(), BB, nullptr, | |||
| 667 | &DomTree); | |||
| 668 | }; | |||
| 669 | for (DbgVariableIntrinsic *DVI : Worklist) { | |||
| 670 | if (IsUnreachableBlock(DVI->getParent())) | |||
| 671 | DVI->eraseFromParent(); | |||
| 672 | else if (dyn_cast_or_null<AllocaInst>(DVI->getVariableLocationOp(0))) { | |||
| 673 | // Count all non-debuginfo uses in reachable blocks. | |||
| 674 | unsigned Uses = 0; | |||
| 675 | for (auto *User : DVI->getVariableLocationOp(0)->users()) | |||
| 676 | if (auto *I = dyn_cast<Instruction>(User)) | |||
| 677 | if (!isa<AllocaInst>(I) && !IsUnreachableBlock(I->getParent())) | |||
| 678 | ++Uses; | |||
| 679 | if (!Uses) | |||
| 680 | DVI->eraseFromParent(); | |||
| 681 | } | |||
| 682 | } | |||
| 683 | } | |||
| 684 | ||||
| 685 | void CoroCloner::replaceEntryBlock() { | |||
| 686 | // In the original function, the AllocaSpillBlock is a block immediately | |||
| 687 | // following the allocation of the frame object which defines GEPs for | |||
| 688 | // all the allocas that have been moved into the frame, and it ends by | |||
| 689 | // branching to the original beginning of the coroutine. Make this | |||
| 690 | // the entry block of the cloned function. | |||
| 691 | auto *Entry = cast<BasicBlock>(VMap[Shape.AllocaSpillBlock]); | |||
| 692 | auto *OldEntry = &NewF->getEntryBlock(); | |||
| 693 | Entry->setName("entry" + Suffix); | |||
| 694 | Entry->moveBefore(OldEntry); | |||
| 695 | Entry->getTerminator()->eraseFromParent(); | |||
| 696 | ||||
| 697 | // Clear all predecessors of the new entry block. There should be | |||
| 698 | // exactly one predecessor, which we created when splitting out | |||
| 699 | // AllocaSpillBlock to begin with. | |||
| 700 | assert(Entry->hasOneUse())((void)0); | |||
| 701 | auto BranchToEntry = cast<BranchInst>(Entry->user_back()); | |||
| 702 | assert(BranchToEntry->isUnconditional())((void)0); | |||
| 703 | Builder.SetInsertPoint(BranchToEntry); | |||
| 704 | Builder.CreateUnreachable(); | |||
| 705 | BranchToEntry->eraseFromParent(); | |||
| 706 | ||||
| 707 | // Branch from the entry to the appropriate place. | |||
| 708 | Builder.SetInsertPoint(Entry); | |||
| 709 | switch (Shape.ABI) { | |||
| 710 | case coro::ABI::Switch: { | |||
| 711 | // In switch-lowering, we built a resume-entry block in the original | |||
| 712 | // function. Make the entry block branch to this. | |||
| 713 | auto *SwitchBB = | |||
| 714 | cast<BasicBlock>(VMap[Shape.SwitchLowering.ResumeEntryBlock]); | |||
| 715 | Builder.CreateBr(SwitchBB); | |||
| 716 | break; | |||
| 717 | } | |||
| 718 | case coro::ABI::Async: | |||
| 719 | case coro::ABI::Retcon: | |||
| 720 | case coro::ABI::RetconOnce: { | |||
| 721 | // In continuation ABIs, we want to branch to immediately after the | |||
| 722 | // active suspend point. Earlier phases will have put the suspend in its | |||
| 723 | // own basic block, so just thread our jump directly to its successor. | |||
| 724 | assert((Shape.ABI == coro::ABI::Async &&((void)0) | |||
| 725 | isa<CoroSuspendAsyncInst>(ActiveSuspend)) ||((void)0) | |||
| 726 | ((Shape.ABI == coro::ABI::Retcon ||((void)0) | |||
| 727 | Shape.ABI == coro::ABI::RetconOnce) &&((void)0) | |||
| 728 | isa<CoroSuspendRetconInst>(ActiveSuspend)))((void)0); | |||
| 729 | auto *MappedCS = cast<AnyCoroSuspendInst>(VMap[ActiveSuspend]); | |||
| 730 | auto Branch = cast<BranchInst>(MappedCS->getNextNode()); | |||
| 731 | assert(Branch->isUnconditional())((void)0); | |||
| 732 | Builder.CreateBr(Branch->getSuccessor(0)); | |||
| 733 | break; | |||
| 734 | } | |||
| 735 | } | |||
| 736 | ||||
| 737 | // Any static alloca that's still being used but not reachable from the new | |||
| 738 | // entry needs to be moved to the new entry. | |||
| 739 | Function *F = OldEntry->getParent(); | |||
| 740 | DominatorTree DT{*F}; | |||
| 741 | for (auto IT = inst_begin(F), End = inst_end(F); IT != End;) { | |||
| 742 | Instruction &I = *IT++; | |||
| 743 | auto *Alloca = dyn_cast<AllocaInst>(&I); | |||
| 744 | if (!Alloca || I.use_empty()) | |||
| 745 | continue; | |||
| 746 | if (DT.isReachableFromEntry(I.getParent()) || | |||
| 747 | !isa<ConstantInt>(Alloca->getArraySize())) | |||
| 748 | continue; | |||
| 749 | I.moveBefore(*Entry, Entry->getFirstInsertionPt()); | |||
| 750 | } | |||
| 751 | } | |||
| 752 | ||||
| 753 | /// Derive the value of the new frame pointer. | |||
| 754 | Value *CoroCloner::deriveNewFramePointer() { | |||
| 755 | // Builder should be inserting to the front of the new entry block. | |||
| 756 | ||||
| 757 | switch (Shape.ABI) { | |||
| 758 | // In switch-lowering, the argument is the frame pointer. | |||
| 759 | case coro::ABI::Switch: | |||
| 760 | return &*NewF->arg_begin(); | |||
| 761 | // In async-lowering, one of the arguments is an async context as determined | |||
| 762 | // by the `llvm.coro.id.async` intrinsic. We can retrieve the async context of | |||
| 763 | // the resume function from the async context projection function associated | |||
| 764 | // with the active suspend. The frame is located as a tail to the async | |||
| 765 | // context header. | |||
| 766 | case coro::ABI::Async: { | |||
| 767 | auto *ActiveAsyncSuspend = cast<CoroSuspendAsyncInst>(ActiveSuspend); | |||
| 768 | auto ContextIdx = ActiveAsyncSuspend->getStorageArgumentIndex() & 0xff; | |||
| 769 | auto *CalleeContext = NewF->getArg(ContextIdx); | |||
| 770 | auto *FramePtrTy = Shape.FrameTy->getPointerTo(); | |||
| 771 | auto *ProjectionFunc = | |||
| 772 | ActiveAsyncSuspend->getAsyncContextProjectionFunction(); | |||
| 773 | auto DbgLoc = | |||
| 774 | cast<CoroSuspendAsyncInst>(VMap[ActiveSuspend])->getDebugLoc(); | |||
| 775 | // Calling i8* (i8*) | |||
| 776 | auto *CallerContext = Builder.CreateCall( | |||
| 777 | cast<FunctionType>(ProjectionFunc->getType()->getPointerElementType()), | |||
| 778 | ProjectionFunc, CalleeContext); | |||
| 779 | CallerContext->setCallingConv(ProjectionFunc->getCallingConv()); | |||
| 780 | CallerContext->setDebugLoc(DbgLoc); | |||
| 781 | // The frame is located after the async_context header. | |||
| 782 | auto &Context = Builder.getContext(); | |||
| 783 | auto *FramePtrAddr = Builder.CreateConstInBoundsGEP1_32( | |||
| 784 | Type::getInt8Ty(Context), CallerContext, | |||
| 785 | Shape.AsyncLowering.FrameOffset, "async.ctx.frameptr"); | |||
| 786 | // Inline the projection function. | |||
| 787 | InlineFunctionInfo InlineInfo; | |||
| 788 | auto InlineRes = InlineFunction(*CallerContext, InlineInfo); | |||
| 789 | assert(InlineRes.isSuccess())((void)0); | |||
| 790 | (void)InlineRes; | |||
| 791 | return Builder.CreateBitCast(FramePtrAddr, FramePtrTy); | |||
| 792 | } | |||
| 793 | // In continuation-lowering, the argument is the opaque storage. | |||
| 794 | case coro::ABI::Retcon: | |||
| 795 | case coro::ABI::RetconOnce: { | |||
| 796 | Argument *NewStorage = &*NewF->arg_begin(); | |||
| 797 | auto FramePtrTy = Shape.FrameTy->getPointerTo(); | |||
| 798 | ||||
| 799 | // If the storage is inline, just bitcast to the storage to the frame type. | |||
| 800 | if (Shape.RetconLowering.IsFrameInlineInStorage) | |||
| 801 | return Builder.CreateBitCast(NewStorage, FramePtrTy); | |||
| 802 | ||||
| 803 | // Otherwise, load the real frame from the opaque storage. | |||
| 804 | auto FramePtrPtr = | |||
| 805 | Builder.CreateBitCast(NewStorage, FramePtrTy->getPointerTo()); | |||
| 806 | return Builder.CreateLoad(FramePtrTy, FramePtrPtr); | |||
| 807 | } | |||
| 808 | } | |||
| 809 | llvm_unreachable("bad ABI")__builtin_unreachable(); | |||
| 810 | } | |||
| 811 | ||||
| 812 | static void addFramePointerAttrs(AttributeList &Attrs, LLVMContext &Context, | |||
| 813 | unsigned ParamIndex, | |||
| 814 | uint64_t Size, Align Alignment) { | |||
| 815 | AttrBuilder ParamAttrs; | |||
| 816 | ParamAttrs.addAttribute(Attribute::NonNull); | |||
| 817 | ParamAttrs.addAttribute(Attribute::NoAlias); | |||
| 818 | ParamAttrs.addAlignmentAttr(Alignment); | |||
| 819 | ParamAttrs.addDereferenceableAttr(Size); | |||
| 820 | Attrs = Attrs.addParamAttributes(Context, ParamIndex, ParamAttrs); | |||
| 821 | } | |||
| 822 | ||||
| 823 | static void addAsyncContextAttrs(AttributeList &Attrs, LLVMContext &Context, | |||
| 824 | unsigned ParamIndex) { | |||
| 825 | AttrBuilder ParamAttrs; | |||
| 826 | ParamAttrs.addAttribute(Attribute::SwiftAsync); | |||
| 827 | Attrs = Attrs.addParamAttributes(Context, ParamIndex, ParamAttrs); | |||
| 828 | } | |||
| 829 | ||||
| 830 | static void addSwiftSelfAttrs(AttributeList &Attrs, LLVMContext &Context, | |||
| 831 | unsigned ParamIndex) { | |||
| 832 | AttrBuilder ParamAttrs; | |||
| 833 | ParamAttrs.addAttribute(Attribute::SwiftSelf); | |||
| 834 | Attrs = Attrs.addParamAttributes(Context, ParamIndex, ParamAttrs); | |||
| 835 | } | |||
| 836 | ||||
| 837 | /// Clone the body of the original function into a resume function of | |||
| 838 | /// some sort. | |||
| 839 | void CoroCloner::create() { | |||
| 840 | // Create the new function if we don't already have one. | |||
| 841 | if (!NewF) { | |||
| 842 | NewF = createCloneDeclaration(OrigF, Shape, Suffix, | |||
| 843 | OrigF.getParent()->end(), ActiveSuspend); | |||
| 844 | } | |||
| 845 | ||||
| 846 | // Replace all args with undefs. The buildCoroutineFrame algorithm already | |||
| 847 | // rewritten access to the args that occurs after suspend points with loads | |||
| 848 | // and stores to/from the coroutine frame. | |||
| 849 | for (Argument &A : OrigF.args()) | |||
| 850 | VMap[&A] = UndefValue::get(A.getType()); | |||
| 851 | ||||
| 852 | SmallVector<ReturnInst *, 4> Returns; | |||
| 853 | ||||
| 854 | // Ignore attempts to change certain attributes of the function. | |||
| 855 | // TODO: maybe there should be a way to suppress this during cloning? | |||
| 856 | auto savedVisibility = NewF->getVisibility(); | |||
| 857 | auto savedUnnamedAddr = NewF->getUnnamedAddr(); | |||
| 858 | auto savedDLLStorageClass = NewF->getDLLStorageClass(); | |||
| 859 | ||||
| 860 | // NewF's linkage (which CloneFunctionInto does *not* change) might not | |||
| 861 | // be compatible with the visibility of OrigF (which it *does* change), | |||
| 862 | // so protect against that. | |||
| 863 | auto savedLinkage = NewF->getLinkage(); | |||
| 864 | NewF->setLinkage(llvm::GlobalValue::ExternalLinkage); | |||
| 865 | ||||
| 866 | CloneFunctionInto(NewF, &OrigF, VMap, | |||
| 867 | CloneFunctionChangeType::LocalChangesOnly, Returns); | |||
| 868 | ||||
| 869 | auto &Context = NewF->getContext(); | |||
| 870 | ||||
| 871 | // For async functions / continuations, adjust the scope line of the | |||
| 872 | // clone to the line number of the suspend point. However, only | |||
| 873 | // adjust the scope line when the files are the same. This ensures | |||
| 874 | // line number and file name belong together. The scope line is | |||
| 875 | // associated with all pre-prologue instructions. This avoids a jump | |||
| 876 | // in the linetable from the function declaration to the suspend point. | |||
| 877 | if (DISubprogram *SP = NewF->getSubprogram()) { | |||
| 878 | assert(SP != OrigF.getSubprogram() && SP->isDistinct())((void)0); | |||
| 879 | if (ActiveSuspend) | |||
| 880 | if (auto DL = ActiveSuspend->getDebugLoc()) | |||
| 881 | if (SP->getFile() == DL->getFile()) | |||
| 882 | SP->setScopeLine(DL->getLine()); | |||
| 883 | // Update the linkage name to reflect the modified symbol name. It | |||
| 884 | // is necessary to update the linkage name in Swift, since the | |||
| 885 | // mangling changes for resume functions. It might also be the | |||
| 886 | // right thing to do in C++, but due to a limitation in LLVM's | |||
| 887 | // AsmPrinter we can only do this if the function doesn't have an | |||
| 888 | // abstract specification, since the DWARF backend expects the | |||
| 889 | // abstract specification to contain the linkage name and asserts | |||
| 890 | // that they are identical. | |||
| 891 | if (!SP->getDeclaration() && SP->getUnit() && | |||
| 892 | SP->getUnit()->getSourceLanguage() == dwarf::DW_LANG_Swift) | |||
| 893 | SP->replaceLinkageName(MDString::get(Context, NewF->getName())); | |||
| 894 | } | |||
| 895 | ||||
| 896 | NewF->setLinkage(savedLinkage); | |||
| 897 | NewF->setVisibility(savedVisibility); | |||
| 898 | NewF->setUnnamedAddr(savedUnnamedAddr); | |||
| 899 | NewF->setDLLStorageClass(savedDLLStorageClass); | |||
| 900 | ||||
| 901 | // Replace the attributes of the new function: | |||
| 902 | auto OrigAttrs = NewF->getAttributes(); | |||
| 903 | auto NewAttrs = AttributeList(); | |||
| 904 | ||||
| 905 | switch (Shape.ABI) { | |||
| 906 | case coro::ABI::Switch: | |||
| 907 | // Bootstrap attributes by copying function attributes from the | |||
| 908 | // original function. This should include optimization settings and so on. | |||
| 909 | NewAttrs = NewAttrs.addAttributes(Context, AttributeList::FunctionIndex, | |||
| 910 | OrigAttrs.getFnAttributes()); | |||
| 911 | ||||
| 912 | addFramePointerAttrs(NewAttrs, Context, 0, | |||
| 913 | Shape.FrameSize, Shape.FrameAlign); | |||
| 914 | break; | |||
| 915 | case coro::ABI::Async: { | |||
| 916 | auto *ActiveAsyncSuspend = cast<CoroSuspendAsyncInst>(ActiveSuspend); | |||
| 917 | if (OrigF.hasParamAttribute(Shape.AsyncLowering.ContextArgNo, | |||
| 918 | Attribute::SwiftAsync)) { | |||
| 919 | uint32_t ArgAttributeIndices = | |||
| 920 | ActiveAsyncSuspend->getStorageArgumentIndex(); | |||
| 921 | auto ContextArgIndex = ArgAttributeIndices & 0xff; | |||
| 922 | addAsyncContextAttrs(NewAttrs, Context, ContextArgIndex); | |||
| 923 | ||||
| 924 | // `swiftasync` must preceed `swiftself` so 0 is not a valid index for | |||
| 925 | // `swiftself`. | |||
| 926 | auto SwiftSelfIndex = ArgAttributeIndices >> 8; | |||
| 927 | if (SwiftSelfIndex) | |||
| 928 | addSwiftSelfAttrs(NewAttrs, Context, SwiftSelfIndex); | |||
| 929 | } | |||
| 930 | ||||
| 931 | // Transfer the original function's attributes. | |||
| 932 | auto FnAttrs = OrigF.getAttributes().getFnAttributes(); | |||
| 933 | NewAttrs = | |||
| 934 | NewAttrs.addAttributes(Context, AttributeList::FunctionIndex, FnAttrs); | |||
| 935 | break; | |||
| 936 | } | |||
| 937 | case coro::ABI::Retcon: | |||
| 938 | case coro::ABI::RetconOnce: | |||
| 939 | // If we have a continuation prototype, just use its attributes, | |||
| 940 | // full-stop. | |||
| 941 | NewAttrs = Shape.RetconLowering.ResumePrototype->getAttributes(); | |||
| 942 | ||||
| 943 | addFramePointerAttrs(NewAttrs, Context, 0, | |||
| 944 | Shape.getRetconCoroId()->getStorageSize(), | |||
| 945 | Shape.getRetconCoroId()->getStorageAlignment()); | |||
| 946 | break; | |||
| 947 | } | |||
| 948 | ||||
| 949 | switch (Shape.ABI) { | |||
| 950 | // In these ABIs, the cloned functions always return 'void', and the | |||
| 951 | // existing return sites are meaningless. Note that for unique | |||
| 952 | // continuations, this includes the returns associated with suspends; | |||
| 953 | // this is fine because we can't suspend twice. | |||
| 954 | case coro::ABI::Switch: | |||
| 955 | case coro::ABI::RetconOnce: | |||
| 956 | // Remove old returns. | |||
| 957 | for (ReturnInst *Return : Returns) | |||
| 958 | changeToUnreachable(Return); | |||
| 959 | break; | |||
| 960 | ||||
| 961 | // With multi-suspend continuations, we'll already have eliminated the | |||
| 962 | // original returns and inserted returns before all the suspend points, | |||
| 963 | // so we want to leave any returns in place. | |||
| 964 | case coro::ABI::Retcon: | |||
| 965 | break; | |||
| 966 | // Async lowering will insert musttail call functions at all suspend points | |||
| 967 | // followed by a return. | |||
| 968 | // Don't change returns to unreachable because that will trip up the verifier. | |||
| 969 | // These returns should be unreachable from the clone. | |||
| 970 | case coro::ABI::Async: | |||
| 971 | break; | |||
| 972 | } | |||
| 973 | ||||
| 974 | NewF->setAttributes(NewAttrs); | |||
| 975 | NewF->setCallingConv(Shape.getResumeFunctionCC()); | |||
| 976 | ||||
| 977 | // Set up the new entry block. | |||
| 978 | replaceEntryBlock(); | |||
| 979 | ||||
| 980 | Builder.SetInsertPoint(&NewF->getEntryBlock().front()); | |||
| 981 | NewFramePtr = deriveNewFramePointer(); | |||
| 982 | ||||
| 983 | // Remap frame pointer. | |||
| 984 | Value *OldFramePtr = VMap[Shape.FramePtr]; | |||
| 985 | NewFramePtr->takeName(OldFramePtr); | |||
| 986 | OldFramePtr->replaceAllUsesWith(NewFramePtr); | |||
| 987 | ||||
| 988 | // Remap vFrame pointer. | |||
| 989 | auto *NewVFrame = Builder.CreateBitCast( | |||
| 990 | NewFramePtr, Type::getInt8PtrTy(Builder.getContext()), "vFrame"); | |||
| 991 | Value *OldVFrame = cast<Value>(VMap[Shape.CoroBegin]); | |||
| 992 | OldVFrame->replaceAllUsesWith(NewVFrame); | |||
| 993 | ||||
| 994 | switch (Shape.ABI) { | |||
| 995 | case coro::ABI::Switch: | |||
| 996 | // Rewrite final suspend handling as it is not done via switch (allows to | |||
| 997 | // remove final case from the switch, since it is undefined behavior to | |||
| 998 | // resume the coroutine suspended at the final suspend point. | |||
| 999 | if (Shape.SwitchLowering.HasFinalSuspend) | |||
| 1000 | handleFinalSuspend(); | |||
| 1001 | break; | |||
| 1002 | case coro::ABI::Async: | |||
| 1003 | case coro::ABI::Retcon: | |||
| 1004 | case coro::ABI::RetconOnce: | |||
| 1005 | // Replace uses of the active suspend with the corresponding | |||
| 1006 | // continuation-function arguments. | |||
| 1007 | assert(ActiveSuspend != nullptr &&((void)0) | |||
| 1008 | "no active suspend when lowering a continuation-style coroutine")((void)0); | |||
| 1009 | replaceRetconOrAsyncSuspendUses(); | |||
| 1010 | break; | |||
| 1011 | } | |||
| 1012 | ||||
| 1013 | // Handle suspends. | |||
| 1014 | replaceCoroSuspends(); | |||
| 1015 | ||||
| 1016 | // Handle swifterror. | |||
| 1017 | replaceSwiftErrorOps(); | |||
| 1018 | ||||
| 1019 | // Remove coro.end intrinsics. | |||
| 1020 | replaceCoroEnds(); | |||
| 1021 | ||||
| 1022 | // Salvage debug info that points into the coroutine frame. | |||
| 1023 | salvageDebugInfo(); | |||
| 1024 | ||||
| 1025 | // Eliminate coro.free from the clones, replacing it with 'null' in cleanup, | |||
| 1026 | // to suppress deallocation code. | |||
| 1027 | if (Shape.ABI == coro::ABI::Switch) | |||
| 1028 | coro::replaceCoroFree(cast<CoroIdInst>(VMap[Shape.CoroBegin->getId()]), | |||
| 1029 | /*Elide=*/ FKind == CoroCloner::Kind::SwitchCleanup); | |||
| 1030 | } | |||
| 1031 | ||||
| 1032 | // Create a resume clone by cloning the body of the original function, setting | |||
| 1033 | // new entry block and replacing coro.suspend an appropriate value to force | |||
| 1034 | // resume or cleanup pass for every suspend point. | |||
| 1035 | static Function *createClone(Function &F, const Twine &Suffix, | |||
| 1036 | coro::Shape &Shape, CoroCloner::Kind FKind) { | |||
| 1037 | CoroCloner Cloner(F, Suffix, Shape, FKind); | |||
| 1038 | Cloner.create(); | |||
| 1039 | return Cloner.getFunction(); | |||
| 1040 | } | |||
| 1041 | ||||
| 1042 | /// Remove calls to llvm.coro.end in the original function. | |||
| 1043 | static void removeCoroEnds(const coro::Shape &Shape, CallGraph *CG) { | |||
| 1044 | for (auto End : Shape.CoroEnds) { | |||
| 1045 | replaceCoroEnd(End, Shape, Shape.FramePtr, /*in resume*/ false, CG); | |||
| 1046 | } | |||
| 1047 | } | |||
| 1048 | ||||
| 1049 | static void updateAsyncFuncPointerContextSize(coro::Shape &Shape) { | |||
| 1050 | assert(Shape.ABI == coro::ABI::Async)((void)0); | |||
| 1051 | ||||
| 1052 | auto *FuncPtrStruct = cast<ConstantStruct>( | |||
| 1053 | Shape.AsyncLowering.AsyncFuncPointer->getInitializer()); | |||
| 1054 | auto *OrigRelativeFunOffset = FuncPtrStruct->getOperand(0); | |||
| 1055 | auto *OrigContextSize = FuncPtrStruct->getOperand(1); | |||
| 1056 | auto *NewContextSize = ConstantInt::get(OrigContextSize->getType(), | |||
| 1057 | Shape.AsyncLowering.ContextSize); | |||
| 1058 | auto *NewFuncPtrStruct = ConstantStruct::get( | |||
| 1059 | FuncPtrStruct->getType(), OrigRelativeFunOffset, NewContextSize); | |||
| 1060 | ||||
| 1061 | Shape.AsyncLowering.AsyncFuncPointer->setInitializer(NewFuncPtrStruct); | |||
| 1062 | } | |||
| 1063 | ||||
| 1064 | static void replaceFrameSize(coro::Shape &Shape) { | |||
| 1065 | if (Shape.ABI == coro::ABI::Async) | |||
| 1066 | updateAsyncFuncPointerContextSize(Shape); | |||
| 1067 | ||||
| 1068 | if (Shape.CoroSizes.empty()) | |||
| 1069 | return; | |||
| 1070 | ||||
| 1071 | // In the same function all coro.sizes should have the same result type. | |||
| 1072 | auto *SizeIntrin = Shape.CoroSizes.back(); | |||
| 1073 | Module *M = SizeIntrin->getModule(); | |||
| 1074 | const DataLayout &DL = M->getDataLayout(); | |||
| 1075 | auto Size = DL.getTypeAllocSize(Shape.FrameTy); | |||
| 1076 | auto *SizeConstant = ConstantInt::get(SizeIntrin->getType(), Size); | |||
| 1077 | ||||
| 1078 | for (CoroSizeInst *CS : Shape.CoroSizes) { | |||
| 1079 | CS->replaceAllUsesWith(SizeConstant); | |||
| 1080 | CS->eraseFromParent(); | |||
| 1081 | } | |||
| 1082 | } | |||
| 1083 | ||||
| 1084 | // Create a global constant array containing pointers to functions provided and | |||
| 1085 | // set Info parameter of CoroBegin to point at this constant. Example: | |||
| 1086 | // | |||
| 1087 | // @f.resumers = internal constant [2 x void(%f.frame*)*] | |||
| 1088 | // [void(%f.frame*)* @f.resume, void(%f.frame*)* @f.destroy] | |||
| 1089 | // define void @f() { | |||
| 1090 | // ... | |||
| 1091 | // call i8* @llvm.coro.begin(i8* null, i32 0, i8* null, | |||
| 1092 | // i8* bitcast([2 x void(%f.frame*)*] * @f.resumers to i8*)) | |||
| 1093 | // | |||
| 1094 | // Assumes that all the functions have the same signature. | |||
| 1095 | static void setCoroInfo(Function &F, coro::Shape &Shape, | |||
| 1096 | ArrayRef<Function *> Fns) { | |||
| 1097 | // This only works under the switch-lowering ABI because coro elision | |||
| 1098 | // only works on the switch-lowering ABI. | |||
| 1099 | assert(Shape.ABI == coro::ABI::Switch)((void)0); | |||
| 1100 | ||||
| 1101 | SmallVector<Constant *, 4> Args(Fns.begin(), Fns.end()); | |||
| 1102 | assert(!Args.empty())((void)0); | |||
| 1103 | Function *Part = *Fns.begin(); | |||
| 1104 | Module *M = Part->getParent(); | |||
| 1105 | auto *ArrTy = ArrayType::get(Part->getType(), Args.size()); | |||
| 1106 | ||||
| 1107 | auto *ConstVal = ConstantArray::get(ArrTy, Args); | |||
| 1108 | auto *GV = new GlobalVariable(*M, ConstVal->getType(), /*isConstant=*/true, | |||
| 1109 | GlobalVariable::PrivateLinkage, ConstVal, | |||
| 1110 | F.getName() + Twine(".resumers")); | |||
| 1111 | ||||
| 1112 | // Update coro.begin instruction to refer to this constant. | |||
| 1113 | LLVMContext &C = F.getContext(); | |||
| 1114 | auto *BC = ConstantExpr::getPointerCast(GV, Type::getInt8PtrTy(C)); | |||
| 1115 | Shape.getSwitchCoroId()->setInfo(BC); | |||
| 1116 | } | |||
| 1117 | ||||
| 1118 | // Store addresses of Resume/Destroy/Cleanup functions in the coroutine frame. | |||
| 1119 | static void updateCoroFrame(coro::Shape &Shape, Function *ResumeFn, | |||
| 1120 | Function *DestroyFn, Function *CleanupFn) { | |||
| 1121 | assert(Shape.ABI == coro::ABI::Switch)((void)0); | |||
| 1122 | ||||
| 1123 | IRBuilder<> Builder(Shape.FramePtr->getNextNode()); | |||
| 1124 | auto *ResumeAddr = Builder.CreateStructGEP( | |||
| 1125 | Shape.FrameTy, Shape.FramePtr, coro::Shape::SwitchFieldIndex::Resume, | |||
| 1126 | "resume.addr"); | |||
| 1127 | Builder.CreateStore(ResumeFn, ResumeAddr); | |||
| 1128 | ||||
| 1129 | Value *DestroyOrCleanupFn = DestroyFn; | |||
| 1130 | ||||
| 1131 | CoroIdInst *CoroId = Shape.getSwitchCoroId(); | |||
| 1132 | if (CoroAllocInst *CA = CoroId->getCoroAlloc()) { | |||
| 1133 | // If there is a CoroAlloc and it returns false (meaning we elide the | |||
| 1134 | // allocation, use CleanupFn instead of DestroyFn). | |||
| 1135 | DestroyOrCleanupFn = Builder.CreateSelect(CA, DestroyFn, CleanupFn); | |||
| 1136 | } | |||
| 1137 | ||||
| 1138 | auto *DestroyAddr = Builder.CreateStructGEP( | |||
| 1139 | Shape.FrameTy, Shape.FramePtr, coro::Shape::SwitchFieldIndex::Destroy, | |||
| 1140 | "destroy.addr"); | |||
| 1141 | Builder.CreateStore(DestroyOrCleanupFn, DestroyAddr); | |||
| 1142 | } | |||
| 1143 | ||||
| 1144 | static void postSplitCleanup(Function &F) { | |||
| 1145 | removeUnreachableBlocks(F); | |||
| 1146 | ||||
| 1147 | // For now, we do a mandatory verification step because we don't | |||
| 1148 | // entirely trust this pass. Note that we don't want to add a verifier | |||
| 1149 | // pass to FPM below because it will also verify all the global data. | |||
| 1150 | if (verifyFunction(F, &errs())) | |||
| 1151 | report_fatal_error("Broken function"); | |||
| 1152 | } | |||
| 1153 | ||||
| 1154 | // Assuming we arrived at the block NewBlock from Prev instruction, store | |||
| 1155 | // PHI's incoming values in the ResolvedValues map. | |||
| 1156 | static void | |||
| 1157 | scanPHIsAndUpdateValueMap(Instruction *Prev, BasicBlock *NewBlock, | |||
| 1158 | DenseMap<Value *, Value *> &ResolvedValues) { | |||
| 1159 | auto *PrevBB = Prev->getParent(); | |||
| 1160 | for (PHINode &PN : NewBlock->phis()) { | |||
| 1161 | auto V = PN.getIncomingValueForBlock(PrevBB); | |||
| 1162 | // See if we already resolved it. | |||
| 1163 | auto VI = ResolvedValues.find(V); | |||
| 1164 | if (VI != ResolvedValues.end()) | |||
| 1165 | V = VI->second; | |||
| 1166 | // Remember the value. | |||
| 1167 | ResolvedValues[&PN] = V; | |||
| 1168 | } | |||
| 1169 | } | |||
| 1170 | ||||
| 1171 | // Replace a sequence of branches leading to a ret, with a clone of a ret | |||
| 1172 | // instruction. Suspend instruction represented by a switch, track the PHI | |||
| 1173 | // values and select the correct case successor when possible. | |||
| 1174 | static bool simplifyTerminatorLeadingToRet(Instruction *InitialInst) { | |||
| 1175 | DenseMap<Value *, Value *> ResolvedValues; | |||
| 1176 | BasicBlock *UnconditionalSucc = nullptr; | |||
| 1177 | ||||
| 1178 | Instruction *I = InitialInst; | |||
| 1179 | while (I->isTerminator() || | |||
| 1180 | (isa<CmpInst>(I) && I->getNextNode()->isTerminator())) { | |||
| 1181 | if (isa<ReturnInst>(I)) { | |||
| 1182 | if (I != InitialInst) { | |||
| 1183 | // If InitialInst is an unconditional branch, | |||
| 1184 | // remove PHI values that come from basic block of InitialInst | |||
| 1185 | if (UnconditionalSucc) | |||
| 1186 | UnconditionalSucc->removePredecessor(InitialInst->getParent(), true); | |||
| 1187 | ReplaceInstWithInst(InitialInst, I->clone()); | |||
| 1188 | } | |||
| 1189 | return true; | |||
| 1190 | } | |||
| 1191 | if (auto *BR = dyn_cast<BranchInst>(I)) { | |||
| 1192 | if (BR->isUnconditional()) { | |||
| 1193 | BasicBlock *BB = BR->getSuccessor(0); | |||
| 1194 | if (I == InitialInst) | |||
| 1195 | UnconditionalSucc = BB; | |||
| 1196 | scanPHIsAndUpdateValueMap(I, BB, ResolvedValues); | |||
| 1197 | I = BB->getFirstNonPHIOrDbgOrLifetime(); | |||
| 1198 | continue; | |||
| 1199 | } | |||
| 1200 | } else if (auto *CondCmp = dyn_cast<CmpInst>(I)) { | |||
| 1201 | auto *BR = dyn_cast<BranchInst>(I->getNextNode()); | |||
| 1202 | if (BR && BR->isConditional() && CondCmp == BR->getCondition()) { | |||
| 1203 | // If the case number of suspended switch instruction is reduced to | |||
| 1204 | // 1, then it is simplified to CmpInst in llvm::ConstantFoldTerminator. | |||
| 1205 | // And the comparsion looks like : %cond = icmp eq i8 %V, constant. | |||
| 1206 | ConstantInt *CondConst = dyn_cast<ConstantInt>(CondCmp->getOperand(1)); | |||
| 1207 | if (CondConst && CondCmp->getPredicate() == CmpInst::ICMP_EQ) { | |||
| 1208 | Value *V = CondCmp->getOperand(0); | |||
| 1209 | auto it = ResolvedValues.find(V); | |||
| 1210 | if (it != ResolvedValues.end()) | |||
| 1211 | V = it->second; | |||
| 1212 | ||||
| 1213 | if (ConstantInt *Cond0 = dyn_cast<ConstantInt>(V)) { | |||
| 1214 | BasicBlock *BB = Cond0->equalsInt(CondConst->getZExtValue()) | |||
| 1215 | ? BR->getSuccessor(0) | |||
| 1216 | : BR->getSuccessor(1); | |||
| 1217 | scanPHIsAndUpdateValueMap(I, BB, ResolvedValues); | |||
| 1218 | I = BB->getFirstNonPHIOrDbgOrLifetime(); | |||
| 1219 | continue; | |||
| 1220 | } | |||
| 1221 | } | |||
| 1222 | } | |||
| 1223 | } else if (auto *SI = dyn_cast<SwitchInst>(I)) { | |||
| 1224 | Value *V = SI->getCondition(); | |||
| 1225 | auto it = ResolvedValues.find(V); | |||
| 1226 | if (it != ResolvedValues.end()) | |||
| 1227 | V = it->second; | |||
| 1228 | if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) { | |||
| 1229 | BasicBlock *BB = SI->findCaseValue(Cond)->getCaseSuccessor(); | |||
| 1230 | scanPHIsAndUpdateValueMap(I, BB, ResolvedValues); | |||
| 1231 | I = BB->getFirstNonPHIOrDbgOrLifetime(); | |||
| 1232 | continue; | |||
| 1233 | } | |||
| 1234 | } | |||
| 1235 | return false; | |||
| 1236 | } | |||
| 1237 | return false; | |||
| 1238 | } | |||
| 1239 | ||||
| 1240 | // Check whether CI obeys the rules of musttail attribute. | |||
| 1241 | static bool shouldBeMustTail(const CallInst &CI, const Function &F) { | |||
| 1242 | if (CI.isInlineAsm()) | |||
| 1243 | return false; | |||
| 1244 | ||||
| 1245 | // Match prototypes and calling conventions of resume function. | |||
| 1246 | FunctionType *CalleeTy = CI.getFunctionType(); | |||
| 1247 | if (!CalleeTy->getReturnType()->isVoidTy() || (CalleeTy->getNumParams() != 1)) | |||
| 1248 | return false; | |||
| 1249 | ||||
| 1250 | Type *CalleeParmTy = CalleeTy->getParamType(0); | |||
| 1251 | if (!CalleeParmTy->isPointerTy() || | |||
| 1252 | (CalleeParmTy->getPointerAddressSpace() != 0)) | |||
| 1253 | return false; | |||
| 1254 | ||||
| 1255 | if (CI.getCallingConv() != F.getCallingConv()) | |||
| 1256 | return false; | |||
| 1257 | ||||
| 1258 | // CI should not has any ABI-impacting function attributes. | |||
| 1259 | static const Attribute::AttrKind ABIAttrs[] = { | |||
| 1260 | Attribute::StructRet, Attribute::ByVal, Attribute::InAlloca, | |||
| 1261 | Attribute::Preallocated, Attribute::InReg, Attribute::Returned, | |||
| 1262 | Attribute::SwiftSelf, Attribute::SwiftError}; | |||
| 1263 | AttributeList Attrs = CI.getAttributes(); | |||
| 1264 | for (auto AK : ABIAttrs) | |||
| 1265 | if (Attrs.hasParamAttribute(0, AK)) | |||
| 1266 | return false; | |||
| 1267 | ||||
| 1268 | return true; | |||
| 1269 | } | |||
| 1270 | ||||
| 1271 | // Add musttail to any resume instructions that is immediately followed by a | |||
| 1272 | // suspend (i.e. ret). We do this even in -O0 to support guaranteed tail call | |||
| 1273 | // for symmetrical coroutine control transfer (C++ Coroutines TS extension). | |||
| 1274 | // This transformation is done only in the resume part of the coroutine that has | |||
| 1275 | // identical signature and calling convention as the coro.resume call. | |||
| 1276 | static void addMustTailToCoroResumes(Function &F) { | |||
| 1277 | bool changed = false; | |||
| 1278 | ||||
| 1279 | // Collect potential resume instructions. | |||
| 1280 | SmallVector<CallInst *, 4> Resumes; | |||
| 1281 | for (auto &I : instructions(F)) | |||
| 1282 | if (auto *Call = dyn_cast<CallInst>(&I)) | |||
| 1283 | if (shouldBeMustTail(*Call, F)) | |||
| 1284 | Resumes.push_back(Call); | |||
| 1285 | ||||
| 1286 | // Set musttail on those that are followed by a ret instruction. | |||
| 1287 | for (CallInst *Call : Resumes) | |||
| 1288 | if (simplifyTerminatorLeadingToRet(Call->getNextNode())) { | |||
| 1289 | Call->setTailCallKind(CallInst::TCK_MustTail); | |||
| 1290 | changed = true; | |||
| 1291 | } | |||
| 1292 | ||||
| 1293 | if (changed) | |||
| 1294 | removeUnreachableBlocks(F); | |||
| 1295 | } | |||
| 1296 | ||||
| 1297 | // Coroutine has no suspend points. Remove heap allocation for the coroutine | |||
| 1298 | // frame if possible. | |||
| 1299 | static void handleNoSuspendCoroutine(coro::Shape &Shape) { | |||
| 1300 | auto *CoroBegin = Shape.CoroBegin; | |||
| 1301 | auto *CoroId = CoroBegin->getId(); | |||
| 1302 | auto *AllocInst = CoroId->getCoroAlloc(); | |||
| 1303 | switch (Shape.ABI) { | |||
| 1304 | case coro::ABI::Switch: { | |||
| 1305 | auto SwitchId = cast<CoroIdInst>(CoroId); | |||
| 1306 | coro::replaceCoroFree(SwitchId, /*Elide=*/AllocInst != nullptr); | |||
| 1307 | if (AllocInst) { | |||
| 1308 | IRBuilder<> Builder(AllocInst); | |||
| 1309 | auto *Frame = Builder.CreateAlloca(Shape.FrameTy); | |||
| 1310 | Frame->setAlignment(Shape.FrameAlign); | |||
| 1311 | auto *VFrame = Builder.CreateBitCast(Frame, Builder.getInt8PtrTy()); | |||
| 1312 | AllocInst->replaceAllUsesWith(Builder.getFalse()); | |||
| 1313 | AllocInst->eraseFromParent(); | |||
| 1314 | CoroBegin->replaceAllUsesWith(VFrame); | |||
| 1315 | } else { | |||
| 1316 | CoroBegin->replaceAllUsesWith(CoroBegin->getMem()); | |||
| 1317 | } | |||
| 1318 | ||||
| 1319 | break; | |||
| 1320 | } | |||
| 1321 | case coro::ABI::Async: | |||
| 1322 | case coro::ABI::Retcon: | |||
| 1323 | case coro::ABI::RetconOnce: | |||
| 1324 | CoroBegin->replaceAllUsesWith(UndefValue::get(CoroBegin->getType())); | |||
| 1325 | break; | |||
| 1326 | } | |||
| 1327 | ||||
| 1328 | CoroBegin->eraseFromParent(); | |||
| 1329 | } | |||
| 1330 | ||||
| 1331 | // SimplifySuspendPoint needs to check that there is no calls between | |||
| 1332 | // coro_save and coro_suspend, since any of the calls may potentially resume | |||
| 1333 | // the coroutine and if that is the case we cannot eliminate the suspend point. | |||
| 1334 | static bool hasCallsInBlockBetween(Instruction *From, Instruction *To) { | |||
| 1335 | for (Instruction *I = From; I != To; I = I->getNextNode()) { | |||
| 1336 | // Assume that no intrinsic can resume the coroutine. | |||
| 1337 | if (isa<IntrinsicInst>(I)) | |||
| 1338 | continue; | |||
| 1339 | ||||
| 1340 | if (isa<CallBase>(I)) | |||
| 1341 | return true; | |||
| 1342 | } | |||
| 1343 | return false; | |||
| 1344 | } | |||
| 1345 | ||||
| 1346 | static bool hasCallsInBlocksBetween(BasicBlock *SaveBB, BasicBlock *ResDesBB) { | |||
| 1347 | SmallPtrSet<BasicBlock *, 8> Set; | |||
| 1348 | SmallVector<BasicBlock *, 8> Worklist; | |||
| 1349 | ||||
| 1350 | Set.insert(SaveBB); | |||
| 1351 | Worklist.push_back(ResDesBB); | |||
| 1352 | ||||
| 1353 | // Accumulate all blocks between SaveBB and ResDesBB. Because CoroSaveIntr | |||
| 1354 | // returns a token consumed by suspend instruction, all blocks in between | |||
| 1355 | // will have to eventually hit SaveBB when going backwards from ResDesBB. | |||
| 1356 | while (!Worklist.empty()) { | |||
| 1357 | auto *BB = Worklist.pop_back_val(); | |||
| 1358 | Set.insert(BB); | |||
| 1359 | for (auto *Pred : predecessors(BB)) | |||
| 1360 | if (Set.count(Pred) == 0) | |||
| 1361 | Worklist.push_back(Pred); | |||
| 1362 | } | |||
| 1363 | ||||
| 1364 | // SaveBB and ResDesBB are checked separately in hasCallsBetween. | |||
| 1365 | Set.erase(SaveBB); | |||
| 1366 | Set.erase(ResDesBB); | |||
| 1367 | ||||
| 1368 | for (auto *BB : Set) | |||
| 1369 | if (hasCallsInBlockBetween(BB->getFirstNonPHI(), nullptr)) | |||
| 1370 | return true; | |||
| 1371 | ||||
| 1372 | return false; | |||
| 1373 | } | |||
| 1374 | ||||
| 1375 | static bool hasCallsBetween(Instruction *Save, Instruction *ResumeOrDestroy) { | |||
| 1376 | auto *SaveBB = Save->getParent(); | |||
| 1377 | auto *ResumeOrDestroyBB = ResumeOrDestroy->getParent(); | |||
| 1378 | ||||
| 1379 | if (SaveBB == ResumeOrDestroyBB) | |||
| 1380 | return hasCallsInBlockBetween(Save->getNextNode(), ResumeOrDestroy); | |||
| 1381 | ||||
| 1382 | // Any calls from Save to the end of the block? | |||
| 1383 | if (hasCallsInBlockBetween(Save->getNextNode(), nullptr)) | |||
| 1384 | return true; | |||
| 1385 | ||||
| 1386 | // Any calls from begging of the block up to ResumeOrDestroy? | |||
| 1387 | if (hasCallsInBlockBetween(ResumeOrDestroyBB->getFirstNonPHI(), | |||
| 1388 | ResumeOrDestroy)) | |||
| 1389 | return true; | |||
| 1390 | ||||
| 1391 | // Any calls in all of the blocks between SaveBB and ResumeOrDestroyBB? | |||
| 1392 | if (hasCallsInBlocksBetween(SaveBB, ResumeOrDestroyBB)) | |||
| 1393 | return true; | |||
| 1394 | ||||
| 1395 | return false; | |||
| 1396 | } | |||
| 1397 | ||||
| 1398 | // If a SuspendIntrin is preceded by Resume or Destroy, we can eliminate the | |||
| 1399 | // suspend point and replace it with nornal control flow. | |||
| 1400 | static bool simplifySuspendPoint(CoroSuspendInst *Suspend, | |||
| 1401 | CoroBeginInst *CoroBegin) { | |||
| 1402 | Instruction *Prev = Suspend->getPrevNode(); | |||
| 1403 | if (!Prev) { | |||
| 1404 | auto *Pred = Suspend->getParent()->getSinglePredecessor(); | |||
| 1405 | if (!Pred) | |||
| 1406 | return false; | |||
| 1407 | Prev = Pred->getTerminator(); | |||
| 1408 | } | |||
| 1409 | ||||
| 1410 | CallBase *CB = dyn_cast<CallBase>(Prev); | |||
| 1411 | if (!CB) | |||
| 1412 | return false; | |||
| 1413 | ||||
| 1414 | auto *Callee = CB->getCalledOperand()->stripPointerCasts(); | |||
| 1415 | ||||
| 1416 | // See if the callsite is for resumption or destruction of the coroutine. | |||
| 1417 | auto *SubFn = dyn_cast<CoroSubFnInst>(Callee); | |||
| 1418 | if (!SubFn) | |||
| 1419 | return false; | |||
| 1420 | ||||
| 1421 | // Does not refer to the current coroutine, we cannot do anything with it. | |||
| 1422 | if (SubFn->getFrame() != CoroBegin) | |||
| 1423 | return false; | |||
| 1424 | ||||
| 1425 | // See if the transformation is safe. Specifically, see if there are any | |||
| 1426 | // calls in between Save and CallInstr. They can potenitally resume the | |||
| 1427 | // coroutine rendering this optimization unsafe. | |||
| 1428 | auto *Save = Suspend->getCoroSave(); | |||
| 1429 | if (hasCallsBetween(Save, CB)) | |||
| 1430 | return false; | |||
| 1431 | ||||
| 1432 | // Replace llvm.coro.suspend with the value that results in resumption over | |||
| 1433 | // the resume or cleanup path. | |||
| 1434 | Suspend->replaceAllUsesWith(SubFn->getRawIndex()); | |||
| 1435 | Suspend->eraseFromParent(); | |||
| 1436 | Save->eraseFromParent(); | |||
| 1437 | ||||
| 1438 | // No longer need a call to coro.resume or coro.destroy. | |||
| 1439 | if (auto *Invoke = dyn_cast<InvokeInst>(CB)) { | |||
| 1440 | BranchInst::Create(Invoke->getNormalDest(), Invoke); | |||
| 1441 | } | |||
| 1442 | ||||
| 1443 | // Grab the CalledValue from CB before erasing the CallInstr. | |||
| 1444 | auto *CalledValue = CB->getCalledOperand(); | |||
| 1445 | CB->eraseFromParent(); | |||
| 1446 | ||||
| 1447 | // If no more users remove it. Usually it is a bitcast of SubFn. | |||
| 1448 | if (CalledValue != SubFn && CalledValue->user_empty()) | |||
| 1449 | if (auto *I = dyn_cast<Instruction>(CalledValue)) | |||
| 1450 | I->eraseFromParent(); | |||
| 1451 | ||||
| 1452 | // Now we are good to remove SubFn. | |||
| 1453 | if (SubFn->user_empty()) | |||
| 1454 | SubFn->eraseFromParent(); | |||
| 1455 | ||||
| 1456 | return true; | |||
| 1457 | } | |||
| 1458 | ||||
| 1459 | // Remove suspend points that are simplified. | |||
| 1460 | static void simplifySuspendPoints(coro::Shape &Shape) { | |||
| 1461 | // Currently, the only simplification we do is switch-lowering-specific. | |||
| 1462 | if (Shape.ABI != coro::ABI::Switch) | |||
| 1463 | return; | |||
| 1464 | ||||
| 1465 | auto &S = Shape.CoroSuspends; | |||
| 1466 | size_t I = 0, N = S.size(); | |||
| 1467 | if (N == 0) | |||
| 1468 | return; | |||
| 1469 | while (true) { | |||
| 1470 | auto SI = cast<CoroSuspendInst>(S[I]); | |||
| 1471 | // Leave final.suspend to handleFinalSuspend since it is undefined behavior | |||
| 1472 | // to resume a coroutine suspended at the final suspend point. | |||
| 1473 | if (!SI->isFinal() && simplifySuspendPoint(SI, Shape.CoroBegin)) { | |||
| 1474 | if (--N == I) | |||
| 1475 | break; | |||
| 1476 | std::swap(S[I], S[N]); | |||
| 1477 | continue; | |||
| 1478 | } | |||
| 1479 | if (++I == N) | |||
| 1480 | break; | |||
| 1481 | } | |||
| 1482 | S.resize(N); | |||
| 1483 | } | |||
| 1484 | ||||
| 1485 | static void splitSwitchCoroutine(Function &F, coro::Shape &Shape, | |||
| 1486 | SmallVectorImpl<Function *> &Clones) { | |||
| 1487 | assert(Shape.ABI == coro::ABI::Switch)((void)0); | |||
| 1488 | ||||
| 1489 | createResumeEntryBlock(F, Shape); | |||
| 1490 | auto ResumeClone = createClone(F, ".resume", Shape, | |||
| 1491 | CoroCloner::Kind::SwitchResume); | |||
| 1492 | auto DestroyClone = createClone(F, ".destroy", Shape, | |||
| 1493 | CoroCloner::Kind::SwitchUnwind); | |||
| 1494 | auto CleanupClone = createClone(F, ".cleanup", Shape, | |||
| 1495 | CoroCloner::Kind::SwitchCleanup); | |||
| 1496 | ||||
| 1497 | postSplitCleanup(*ResumeClone); | |||
| 1498 | postSplitCleanup(*DestroyClone); | |||
| 1499 | postSplitCleanup(*CleanupClone); | |||
| 1500 | ||||
| 1501 | addMustTailToCoroResumes(*ResumeClone); | |||
| 1502 | ||||
| 1503 | // Store addresses resume/destroy/cleanup functions in the coroutine frame. | |||
| 1504 | updateCoroFrame(Shape, ResumeClone, DestroyClone, CleanupClone); | |||
| 1505 | ||||
| 1506 | assert(Clones.empty())((void)0); | |||
| 1507 | Clones.push_back(ResumeClone); | |||
| 1508 | Clones.push_back(DestroyClone); | |||
| 1509 | Clones.push_back(CleanupClone); | |||
| 1510 | ||||
| 1511 | // Create a constant array referring to resume/destroy/clone functions pointed | |||
| 1512 | // by the last argument of @llvm.coro.info, so that CoroElide pass can | |||
| 1513 | // determined correct function to call. | |||
| 1514 | setCoroInfo(F, Shape, Clones); | |||
| 1515 | } | |||
| 1516 | ||||
| 1517 | static void replaceAsyncResumeFunction(CoroSuspendAsyncInst *Suspend, | |||
| 1518 | Value *Continuation) { | |||
| 1519 | auto *ResumeIntrinsic = Suspend->getResumeFunction(); | |||
| 1520 | auto &Context = Suspend->getParent()->getParent()->getContext(); | |||
| 1521 | auto *Int8PtrTy = Type::getInt8PtrTy(Context); | |||
| 1522 | ||||
| 1523 | IRBuilder<> Builder(ResumeIntrinsic); | |||
| 1524 | auto *Val = Builder.CreateBitOrPointerCast(Continuation, Int8PtrTy); | |||
| 1525 | ResumeIntrinsic->replaceAllUsesWith(Val); | |||
| 1526 | ResumeIntrinsic->eraseFromParent(); | |||
| 1527 | Suspend->setOperand(CoroSuspendAsyncInst::ResumeFunctionArg, | |||
| 1528 | UndefValue::get(Int8PtrTy)); | |||
| 1529 | } | |||
| 1530 | ||||
| 1531 | /// Coerce the arguments in \p FnArgs according to \p FnTy in \p CallArgs. | |||
| 1532 | static void coerceArguments(IRBuilder<> &Builder, FunctionType *FnTy, | |||
| 1533 | ArrayRef<Value *> FnArgs, | |||
| 1534 | SmallVectorImpl<Value *> &CallArgs) { | |||
| 1535 | size_t ArgIdx = 0; | |||
| 1536 | for (auto paramTy : FnTy->params()) { | |||
| 1537 | assert(ArgIdx < FnArgs.size())((void)0); | |||
| 1538 | if (paramTy != FnArgs[ArgIdx]->getType()) | |||
| 1539 | CallArgs.push_back( | |||
| 1540 | Builder.CreateBitOrPointerCast(FnArgs[ArgIdx], paramTy)); | |||
| 1541 | else | |||
| 1542 | CallArgs.push_back(FnArgs[ArgIdx]); | |||
| 1543 | ++ArgIdx; | |||
| 1544 | } | |||
| 1545 | } | |||
| 1546 | ||||
| 1547 | CallInst *coro::createMustTailCall(DebugLoc Loc, Function *MustTailCallFn, | |||
| 1548 | ArrayRef<Value *> Arguments, | |||
| 1549 | IRBuilder<> &Builder) { | |||
| 1550 | auto *FnTy = | |||
| 1551 | cast<FunctionType>(MustTailCallFn->getType()->getPointerElementType()); | |||
| 1552 | // Coerce the arguments, llvm optimizations seem to ignore the types in | |||
| 1553 | // vaarg functions and throws away casts in optimized mode. | |||
| 1554 | SmallVector<Value *, 8> CallArgs; | |||
| 1555 | coerceArguments(Builder, FnTy, Arguments, CallArgs); | |||
| 1556 | ||||
| 1557 | auto *TailCall = Builder.CreateCall(FnTy, MustTailCallFn, CallArgs); | |||
| 1558 | TailCall->setTailCallKind(CallInst::TCK_MustTail); | |||
| 1559 | TailCall->setDebugLoc(Loc); | |||
| 1560 | TailCall->setCallingConv(MustTailCallFn->getCallingConv()); | |||
| 1561 | return TailCall; | |||
| 1562 | } | |||
| 1563 | ||||
| 1564 | static void splitAsyncCoroutine(Function &F, coro::Shape &Shape, | |||
| 1565 | SmallVectorImpl<Function *> &Clones) { | |||
| 1566 | assert(Shape.ABI == coro::ABI::Async)((void)0); | |||
| 1567 | assert(Clones.empty())((void)0); | |||
| 1568 | // Reset various things that the optimizer might have decided it | |||
| 1569 | // "knows" about the coroutine function due to not seeing a return. | |||
| 1570 | F.removeFnAttr(Attribute::NoReturn); | |||
| 1571 | F.removeAttribute(AttributeList::ReturnIndex, Attribute::NoAlias); | |||
| 1572 | F.removeAttribute(AttributeList::ReturnIndex, Attribute::NonNull); | |||
| 1573 | ||||
| 1574 | auto &Context = F.getContext(); | |||
| 1575 | auto *Int8PtrTy = Type::getInt8PtrTy(Context); | |||
| 1576 | ||||
| 1577 | auto *Id = cast<CoroIdAsyncInst>(Shape.CoroBegin->getId()); | |||
| 1578 | IRBuilder<> Builder(Id); | |||
| 1579 | ||||
| 1580 | auto *FramePtr = Id->getStorage(); | |||
| 1581 | FramePtr = Builder.CreateBitOrPointerCast(FramePtr, Int8PtrTy); | |||
| 1582 | FramePtr = Builder.CreateConstInBoundsGEP1_32( | |||
| 1583 | Type::getInt8Ty(Context), FramePtr, Shape.AsyncLowering.FrameOffset, | |||
| 1584 | "async.ctx.frameptr"); | |||
| 1585 | ||||
| 1586 | // Map all uses of llvm.coro.begin to the allocated frame pointer. | |||
| 1587 | { | |||
| 1588 | // Make sure we don't invalidate Shape.FramePtr. | |||
| 1589 | TrackingVH<Instruction> Handle(Shape.FramePtr); | |||
| 1590 | Shape.CoroBegin->replaceAllUsesWith(FramePtr); | |||
| 1591 | Shape.FramePtr = Handle.getValPtr(); | |||
| 1592 | } | |||
| 1593 | ||||
| 1594 | // Create all the functions in order after the main function. | |||
| 1595 | auto NextF = std::next(F.getIterator()); | |||
| 1596 | ||||
| 1597 | // Create a continuation function for each of the suspend points. | |||
| 1598 | Clones.reserve(Shape.CoroSuspends.size()); | |||
| 1599 | for (size_t Idx = 0, End = Shape.CoroSuspends.size(); Idx != End; ++Idx) { | |||
| 1600 | auto *Suspend = cast<CoroSuspendAsyncInst>(Shape.CoroSuspends[Idx]); | |||
| 1601 | ||||
| 1602 | // Create the clone declaration. | |||
| 1603 | auto ResumeNameSuffix = ".resume."; | |||
| 1604 | auto ProjectionFunctionName = | |||
| 1605 | Suspend->getAsyncContextProjectionFunction()->getName(); | |||
| 1606 | bool UseSwiftMangling = false; | |||
| 1607 | if (ProjectionFunctionName.equals("__swift_async_resume_project_context")) { | |||
| 1608 | ResumeNameSuffix = "TQ"; | |||
| 1609 | UseSwiftMangling = true; | |||
| 1610 | } else if (ProjectionFunctionName.equals( | |||
| 1611 | "__swift_async_resume_get_context")) { | |||
| 1612 | ResumeNameSuffix = "TY"; | |||
| 1613 | UseSwiftMangling = true; | |||
| 1614 | } | |||
| 1615 | auto *Continuation = createCloneDeclaration( | |||
| 1616 | F, Shape, | |||
| 1617 | UseSwiftMangling ? ResumeNameSuffix + Twine(Idx) + "_" | |||
| 1618 | : ResumeNameSuffix + Twine(Idx), | |||
| 1619 | NextF, Suspend); | |||
| 1620 | Clones.push_back(Continuation); | |||
| 1621 | ||||
| 1622 | // Insert a branch to a new return block immediately before the suspend | |||
| 1623 | // point. | |||
| 1624 | auto *SuspendBB = Suspend->getParent(); | |||
| 1625 | auto *NewSuspendBB = SuspendBB->splitBasicBlock(Suspend); | |||
| 1626 | auto *Branch = cast<BranchInst>(SuspendBB->getTerminator()); | |||
| 1627 | ||||
| 1628 | // Place it before the first suspend. | |||
| 1629 | auto *ReturnBB = | |||
| 1630 | BasicBlock::Create(F.getContext(), "coro.return", &F, NewSuspendBB); | |||
| 1631 | Branch->setSuccessor(0, ReturnBB); | |||
| 1632 | ||||
| 1633 | IRBuilder<> Builder(ReturnBB); | |||
| 1634 | ||||
| 1635 | // Insert the call to the tail call function and inline it. | |||
| 1636 | auto *Fn = Suspend->getMustTailCallFunction(); | |||
| 1637 | SmallVector<Value *, 8> Args(Suspend->args()); | |||
| 1638 | auto FnArgs = ArrayRef<Value *>(Args).drop_front( | |||
| 1639 | CoroSuspendAsyncInst::MustTailCallFuncArg + 1); | |||
| 1640 | auto *TailCall = | |||
| 1641 | coro::createMustTailCall(Suspend->getDebugLoc(), Fn, FnArgs, Builder); | |||
| 1642 | Builder.CreateRetVoid(); | |||
| 1643 | InlineFunctionInfo FnInfo; | |||
| 1644 | auto InlineRes = InlineFunction(*TailCall, FnInfo); | |||
| 1645 | assert(InlineRes.isSuccess() && "Expected inlining to succeed")((void)0); | |||
| 1646 | (void)InlineRes; | |||
| 1647 | ||||
| 1648 | // Replace the lvm.coro.async.resume intrisic call. | |||
| 1649 | replaceAsyncResumeFunction(Suspend, Continuation); | |||
| 1650 | } | |||
| 1651 | ||||
| 1652 | assert(Clones.size() == Shape.CoroSuspends.size())((void)0); | |||
| 1653 | for (size_t Idx = 0, End = Shape.CoroSuspends.size(); Idx != End; ++Idx) { | |||
| 1654 | auto *Suspend = Shape.CoroSuspends[Idx]; | |||
| 1655 | auto *Clone = Clones[Idx]; | |||
| 1656 | ||||
| 1657 | CoroCloner(F, "resume." + Twine(Idx), Shape, Clone, Suspend).create(); | |||
| 1658 | } | |||
| 1659 | } | |||
| 1660 | ||||
| 1661 | static void splitRetconCoroutine(Function &F, coro::Shape &Shape, | |||
| 1662 | SmallVectorImpl<Function *> &Clones) { | |||
| 1663 | assert(Shape.ABI == coro::ABI::Retcon ||((void)0) | |||
| 1664 | Shape.ABI == coro::ABI::RetconOnce)((void)0); | |||
| 1665 | assert(Clones.empty())((void)0); | |||
| 1666 | ||||
| 1667 | // Reset various things that the optimizer might have decided it | |||
| 1668 | // "knows" about the coroutine function due to not seeing a return. | |||
| 1669 | F.removeFnAttr(Attribute::NoReturn); | |||
| 1670 | F.removeAttribute(AttributeList::ReturnIndex, Attribute::NoAlias); | |||
| 1671 | F.removeAttribute(AttributeList::ReturnIndex, Attribute::NonNull); | |||
| 1672 | ||||
| 1673 | // Allocate the frame. | |||
| 1674 | auto *Id = cast<AnyCoroIdRetconInst>(Shape.CoroBegin->getId()); | |||
| 1675 | Value *RawFramePtr; | |||
| 1676 | if (Shape.RetconLowering.IsFrameInlineInStorage) { | |||
| 1677 | RawFramePtr = Id->getStorage(); | |||
| 1678 | } else { | |||
| 1679 | IRBuilder<> Builder(Id); | |||
| 1680 | ||||
| 1681 | // Determine the size of the frame. | |||
| 1682 | const DataLayout &DL = F.getParent()->getDataLayout(); | |||
| 1683 | auto Size = DL.getTypeAllocSize(Shape.FrameTy); | |||
| 1684 | ||||
| 1685 | // Allocate. We don't need to update the call graph node because we're | |||
| 1686 | // going to recompute it from scratch after splitting. | |||
| 1687 | // FIXME: pass the required alignment | |||
| 1688 | RawFramePtr = Shape.emitAlloc(Builder, Builder.getInt64(Size), nullptr); | |||
| 1689 | RawFramePtr = | |||
| 1690 | Builder.CreateBitCast(RawFramePtr, Shape.CoroBegin->getType()); | |||
| 1691 | ||||
| 1692 | // Stash the allocated frame pointer in the continuation storage. | |||
| 1693 | auto Dest = Builder.CreateBitCast(Id->getStorage(), | |||
| 1694 | RawFramePtr->getType()->getPointerTo()); | |||
| 1695 | Builder.CreateStore(RawFramePtr, Dest); | |||
| 1696 | } | |||
| 1697 | ||||
| 1698 | // Map all uses of llvm.coro.begin to the allocated frame pointer. | |||
| 1699 | { | |||
| 1700 | // Make sure we don't invalidate Shape.FramePtr. | |||
| 1701 | TrackingVH<Instruction> Handle(Shape.FramePtr); | |||
| 1702 | Shape.CoroBegin->replaceAllUsesWith(RawFramePtr); | |||
| 1703 | Shape.FramePtr = Handle.getValPtr(); | |||
| 1704 | } | |||
| 1705 | ||||
| 1706 | // Create a unique return block. | |||
| 1707 | BasicBlock *ReturnBB = nullptr; | |||
| 1708 | SmallVector<PHINode *, 4> ReturnPHIs; | |||
| 1709 | ||||
| 1710 | // Create all the functions in order after the main function. | |||
| 1711 | auto NextF = std::next(F.getIterator()); | |||
| 1712 | ||||
| 1713 | // Create a continuation function for each of the suspend points. | |||
| 1714 | Clones.reserve(Shape.CoroSuspends.size()); | |||
| 1715 | for (size_t i = 0, e = Shape.CoroSuspends.size(); i != e; ++i) { | |||
| 1716 | auto Suspend = cast<CoroSuspendRetconInst>(Shape.CoroSuspends[i]); | |||
| 1717 | ||||
| 1718 | // Create the clone declaration. | |||
| 1719 | auto Continuation = | |||
| 1720 | createCloneDeclaration(F, Shape, ".resume." + Twine(i), NextF, nullptr); | |||
| 1721 | Clones.push_back(Continuation); | |||
| 1722 | ||||
| 1723 | // Insert a branch to the unified return block immediately before | |||
| 1724 | // the suspend point. | |||
| 1725 | auto SuspendBB = Suspend->getParent(); | |||
| 1726 | auto NewSuspendBB = SuspendBB->splitBasicBlock(Suspend); | |||
| 1727 | auto Branch = cast<BranchInst>(SuspendBB->getTerminator()); | |||
| 1728 | ||||
| 1729 | // Create the unified return block. | |||
| 1730 | if (!ReturnBB) { | |||
| 1731 | // Place it before the first suspend. | |||
| 1732 | ReturnBB = BasicBlock::Create(F.getContext(), "coro.return", &F, | |||
| 1733 | NewSuspendBB); | |||
| 1734 | Shape.RetconLowering.ReturnBlock = ReturnBB; | |||
| 1735 | ||||
| 1736 | IRBuilder<> Builder(ReturnBB); | |||
| 1737 | ||||
| 1738 | // Create PHIs for all the return values. | |||
| 1739 | assert(ReturnPHIs.empty())((void)0); | |||
| 1740 | ||||
| 1741 | // First, the continuation. | |||
| 1742 | ReturnPHIs.push_back(Builder.CreatePHI(Continuation->getType(), | |||
| 1743 | Shape.CoroSuspends.size())); | |||
| 1744 | ||||
| 1745 | // Next, all the directly-yielded values. | |||
| 1746 | for (auto ResultTy : Shape.getRetconResultTypes()) | |||
| 1747 | ReturnPHIs.push_back(Builder.CreatePHI(ResultTy, | |||
| 1748 | Shape.CoroSuspends.size())); | |||
| 1749 | ||||
| 1750 | // Build the return value. | |||
| 1751 | auto RetTy = F.getReturnType(); | |||
| 1752 | ||||
| 1753 | // Cast the continuation value if necessary. | |||
| 1754 | // We can't rely on the types matching up because that type would | |||
| 1755 | // have to be infinite. | |||
| 1756 | auto CastedContinuationTy = | |||
| 1757 | (ReturnPHIs.size() == 1 ? RetTy : RetTy->getStructElementType(0)); | |||
| 1758 | auto *CastedContinuation = | |||
| 1759 | Builder.CreateBitCast(ReturnPHIs[0], CastedContinuationTy); | |||
| 1760 | ||||
| 1761 | Value *RetV; | |||
| 1762 | if (ReturnPHIs.size() == 1) { | |||
| 1763 | RetV = CastedContinuation; | |||
| 1764 | } else { | |||
| 1765 | RetV = UndefValue::get(RetTy); | |||
| 1766 | RetV = Builder.CreateInsertValue(RetV, CastedContinuation, 0); | |||
| 1767 | for (size_t I = 1, E = ReturnPHIs.size(); I != E; ++I) | |||
| 1768 | RetV = Builder.CreateInsertValue(RetV, ReturnPHIs[I], I); | |||
| 1769 | } | |||
| 1770 | ||||
| 1771 | Builder.CreateRet(RetV); | |||
| 1772 | } | |||
| 1773 | ||||
| 1774 | // Branch to the return block. | |||
| 1775 | Branch->setSuccessor(0, ReturnBB); | |||
| 1776 | ReturnPHIs[0]->addIncoming(Continuation, SuspendBB); | |||
| 1777 | size_t NextPHIIndex = 1; | |||
| 1778 | for (auto &VUse : Suspend->value_operands()) | |||
| 1779 | ReturnPHIs[NextPHIIndex++]->addIncoming(&*VUse, SuspendBB); | |||
| 1780 | assert(NextPHIIndex == ReturnPHIs.size())((void)0); | |||
| 1781 | } | |||
| 1782 | ||||
| 1783 | assert(Clones.size() == Shape.CoroSuspends.size())((void)0); | |||
| 1784 | for (size_t i = 0, e = Shape.CoroSuspends.size(); i != e; ++i) { | |||
| 1785 | auto Suspend = Shape.CoroSuspends[i]; | |||
| 1786 | auto Clone = Clones[i]; | |||
| 1787 | ||||
| 1788 | CoroCloner(F, "resume." + Twine(i), Shape, Clone, Suspend).create(); | |||
| 1789 | } | |||
| 1790 | } | |||
| 1791 | ||||
| 1792 | namespace { | |||
| 1793 | class PrettyStackTraceFunction : public PrettyStackTraceEntry { | |||
| 1794 | Function &F; | |||
| 1795 | public: | |||
| 1796 | PrettyStackTraceFunction(Function &F) : F(F) {} | |||
| 1797 | void print(raw_ostream &OS) const override { | |||
| 1798 | OS << "While splitting coroutine "; | |||
| 1799 | F.printAsOperand(OS, /*print type*/ false, F.getParent()); | |||
| 1800 | OS << "\n"; | |||
| 1801 | } | |||
| 1802 | }; | |||
| 1803 | } | |||
| 1804 | ||||
| 1805 | static coro::Shape splitCoroutine(Function &F, | |||
| 1806 | SmallVectorImpl<Function *> &Clones, | |||
| 1807 | bool ReuseFrameSlot) { | |||
| 1808 | PrettyStackTraceFunction prettyStackTrace(F); | |||
| 1809 | ||||
| 1810 | // The suspend-crossing algorithm in buildCoroutineFrame get tripped | |||
| 1811 | // up by uses in unreachable blocks, so remove them as a first pass. | |||
| 1812 | removeUnreachableBlocks(F); | |||
| 1813 | ||||
| 1814 | coro::Shape Shape(F, ReuseFrameSlot); | |||
| 1815 | if (!Shape.CoroBegin) | |||
| 1816 | return Shape; | |||
| 1817 | ||||
| 1818 | simplifySuspendPoints(Shape); | |||
| 1819 | buildCoroutineFrame(F, Shape); | |||
| 1820 | replaceFrameSize(Shape); | |||
| 1821 | ||||
| 1822 | // If there are no suspend points, no split required, just remove | |||
| 1823 | // the allocation and deallocation blocks, they are not needed. | |||
| 1824 | if (Shape.CoroSuspends.empty()) { | |||
| 1825 | handleNoSuspendCoroutine(Shape); | |||
| 1826 | } else { | |||
| 1827 | switch (Shape.ABI) { | |||
| 1828 | case coro::ABI::Switch: | |||
| 1829 | splitSwitchCoroutine(F, Shape, Clones); | |||
| 1830 | break; | |||
| 1831 | case coro::ABI::Async: | |||
| 1832 | splitAsyncCoroutine(F, Shape, Clones); | |||
| 1833 | break; | |||
| 1834 | case coro::ABI::Retcon: | |||
| 1835 | case coro::ABI::RetconOnce: | |||
| 1836 | splitRetconCoroutine(F, Shape, Clones); | |||
| 1837 | break; | |||
| 1838 | } | |||
| 1839 | } | |||
| 1840 | ||||
| 1841 | // Replace all the swifterror operations in the original function. | |||
| 1842 | // This invalidates SwiftErrorOps in the Shape. | |||
| 1843 | replaceSwiftErrorOps(F, Shape, nullptr); | |||
| 1844 | ||||
| 1845 | return Shape; | |||
| 1846 | } | |||
| 1847 | ||||
| 1848 | static void | |||
| 1849 | updateCallGraphAfterCoroutineSplit(Function &F, const coro::Shape &Shape, | |||
| 1850 | const SmallVectorImpl<Function *> &Clones, | |||
| 1851 | CallGraph &CG, CallGraphSCC &SCC) { | |||
| 1852 | if (!Shape.CoroBegin) | |||
| 1853 | return; | |||
| 1854 | ||||
| 1855 | removeCoroEnds(Shape, &CG); | |||
| 1856 | postSplitCleanup(F); | |||
| 1857 | ||||
| 1858 | // Update call graph and add the functions we created to the SCC. | |||
| 1859 | coro::updateCallGraph(F, Clones, CG, SCC); | |||
| 1860 | } | |||
| 1861 | ||||
| 1862 | static void updateCallGraphAfterCoroutineSplit( | |||
| 1863 | LazyCallGraph::Node &N, const coro::Shape &Shape, | |||
| 1864 | const SmallVectorImpl<Function *> &Clones, LazyCallGraph::SCC &C, | |||
| 1865 | LazyCallGraph &CG, CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR, | |||
| 1866 | FunctionAnalysisManager &FAM) { | |||
| 1867 | if (!Shape.CoroBegin) | |||
| 1868 | return; | |||
| 1869 | ||||
| 1870 | for (llvm::AnyCoroEndInst *End : Shape.CoroEnds) { | |||
| 1871 | auto &Context = End->getContext(); | |||
| 1872 | End->replaceAllUsesWith(ConstantInt::getFalse(Context)); | |||
| 1873 | End->eraseFromParent(); | |||
| 1874 | } | |||
| 1875 | ||||
| 1876 | if (!Clones.empty()) { | |||
| 1877 | switch (Shape.ABI) { | |||
| 1878 | case coro::ABI::Switch: | |||
| 1879 | // Each clone in the Switch lowering is independent of the other clones. | |||
| 1880 | // Let the LazyCallGraph know about each one separately. | |||
| 1881 | for (Function *Clone : Clones) | |||
| 1882 | CG.addSplitFunction(N.getFunction(), *Clone); | |||
| 1883 | break; | |||
| 1884 | case coro::ABI::Async: | |||
| 1885 | case coro::ABI::Retcon: | |||
| 1886 | case coro::ABI::RetconOnce: | |||
| 1887 | // Each clone in the Async/Retcon lowering references of the other clones. | |||
| 1888 | // Let the LazyCallGraph know about all of them at once. | |||
| 1889 | if (!Clones.empty()) | |||
| 1890 | CG.addSplitRefRecursiveFunctions(N.getFunction(), Clones); | |||
| 1891 | break; | |||
| 1892 | } | |||
| 1893 | ||||
| 1894 | // Let the CGSCC infra handle the changes to the original function. | |||
| 1895 | updateCGAndAnalysisManagerForCGSCCPass(CG, C, N, AM, UR, FAM); | |||
| 1896 | } | |||
| 1897 | ||||
| 1898 | // Do some cleanup and let the CGSCC infra see if we've cleaned up any edges | |||
| 1899 | // to the split functions. | |||
| 1900 | postSplitCleanup(N.getFunction()); | |||
| 1901 | updateCGAndAnalysisManagerForFunctionPass(CG, C, N, AM, UR, FAM); | |||
| 1902 | } | |||
| 1903 | ||||
| 1904 | // When we see the coroutine the first time, we insert an indirect call to a | |||
| 1905 | // devirt trigger function and mark the coroutine that it is now ready for | |||
| 1906 | // split. | |||
| 1907 | // Async lowering uses this after it has split the function to restart the | |||
| 1908 | // pipeline. | |||
| 1909 | static void prepareForSplit(Function &F, CallGraph &CG, | |||
| 1910 | bool MarkForAsyncRestart = false) { | |||
| 1911 | Module &M = *F.getParent(); | |||
| 1912 | LLVMContext &Context = F.getContext(); | |||
| 1913 | #ifndef NDEBUG1 | |||
| 1914 | Function *DevirtFn = M.getFunction(CORO_DEVIRT_TRIGGER_FN"coro.devirt.trigger"); | |||
| 1915 | assert(DevirtFn && "coro.devirt.trigger function not found")((void)0); | |||
| 1916 | #endif | |||
| 1917 | ||||
| 1918 | F.addFnAttr(CORO_PRESPLIT_ATTR"coroutine.presplit", MarkForAsyncRestart | |||
| 1919 | ? ASYNC_RESTART_AFTER_SPLIT"2" | |||
| 1920 | : PREPARED_FOR_SPLIT"1"); | |||
| 1921 | ||||
| 1922 | // Insert an indirect call sequence that will be devirtualized by CoroElide | |||
| 1923 | // pass: | |||
| 1924 | // %0 = call i8* @llvm.coro.subfn.addr(i8* null, i8 -1) | |||
| 1925 | // %1 = bitcast i8* %0 to void(i8*)* | |||
| 1926 | // call void %1(i8* null) | |||
| 1927 | coro::LowererBase Lowerer(M); | |||
| 1928 | Instruction *InsertPt = | |||
| 1929 | MarkForAsyncRestart ? F.getEntryBlock().getFirstNonPHIOrDbgOrLifetime() | |||
| 1930 | : F.getEntryBlock().getTerminator(); | |||
| 1931 | auto *Null = ConstantPointerNull::get(Type::getInt8PtrTy(Context)); | |||
| 1932 | auto *DevirtFnAddr = | |||
| 1933 | Lowerer.makeSubFnCall(Null, CoroSubFnInst::RestartTrigger, InsertPt); | |||
| 1934 | FunctionType *FnTy = FunctionType::get(Type::getVoidTy(Context), | |||
| 1935 | {Type::getInt8PtrTy(Context)}, false); | |||
| 1936 | auto *IndirectCall = CallInst::Create(FnTy, DevirtFnAddr, Null, "", InsertPt); | |||
| 1937 | ||||
| 1938 | // Update CG graph with an indirect call we just added. | |||
| 1939 | CG[&F]->addCalledFunction(IndirectCall, CG.getCallsExternalNode()); | |||
| 1940 | } | |||
| 1941 | ||||
| 1942 | // Make sure that there is a devirtualization trigger function that the | |||
| 1943 | // coro-split pass uses to force a restart of the CGSCC pipeline. If the devirt | |||
| 1944 | // trigger function is not found, we will create one and add it to the current | |||
| 1945 | // SCC. | |||
| 1946 | static void createDevirtTriggerFunc(CallGraph &CG, CallGraphSCC &SCC) { | |||
| 1947 | Module &M = CG.getModule(); | |||
| 1948 | if (M.getFunction(CORO_DEVIRT_TRIGGER_FN"coro.devirt.trigger")) | |||
| 1949 | return; | |||
| 1950 | ||||
| 1951 | LLVMContext &C = M.getContext(); | |||
| 1952 | auto *FnTy = FunctionType::get(Type::getVoidTy(C), Type::getInt8PtrTy(C), | |||
| 1953 | /*isVarArg=*/false); | |||
| 1954 | Function *DevirtFn = | |||
| 1955 | Function::Create(FnTy, GlobalValue::LinkageTypes::PrivateLinkage, | |||
| 1956 | CORO_DEVIRT_TRIGGER_FN"coro.devirt.trigger", &M); | |||
| 1957 | DevirtFn->addFnAttr(Attribute::AlwaysInline); | |||
| 1958 | auto *Entry = BasicBlock::Create(C, "entry", DevirtFn); | |||
| 1959 | ReturnInst::Create(C, Entry); | |||
| 1960 | ||||
| 1961 | auto *Node = CG.getOrInsertFunction(DevirtFn); | |||
| 1962 | ||||
| 1963 | SmallVector<CallGraphNode *, 8> Nodes(SCC.begin(), SCC.end()); | |||
| 1964 | Nodes.push_back(Node); | |||
| 1965 | SCC.initialize(Nodes); | |||
| 1966 | } | |||
| 1967 | ||||
| 1968 | /// Replace a call to llvm.coro.prepare.retcon. | |||
| 1969 | static void replacePrepare(CallInst *Prepare, LazyCallGraph &CG, | |||
| 1970 | LazyCallGraph::SCC &C) { | |||
| 1971 | auto CastFn = Prepare->getArgOperand(0); // as an i8* | |||
| 1972 | auto Fn = CastFn->stripPointerCasts(); // as its original type | |||
| 1973 | ||||
| 1974 | // Attempt to peephole this pattern: | |||
| 1975 | // %0 = bitcast [[TYPE]] @some_function to i8* | |||
| 1976 | // %1 = call @llvm.coro.prepare.retcon(i8* %0) | |||
| 1977 | // %2 = bitcast %1 to [[TYPE]] | |||
| 1978 | // ==> | |||
| 1979 | // %2 = @some_function | |||
| 1980 | for (auto UI = Prepare->use_begin(), UE = Prepare->use_end(); UI != UE;) { | |||
| 1981 | // Look for bitcasts back to the original function type. | |||
| 1982 | auto *Cast = dyn_cast<BitCastInst>((UI++)->getUser()); | |||
| 1983 | if (!Cast || Cast->getType() != Fn->getType()) | |||
| 1984 | continue; | |||
| 1985 | ||||
| 1986 | // Replace and remove the cast. | |||
| 1987 | Cast->replaceAllUsesWith(Fn); | |||
| 1988 | Cast->eraseFromParent(); | |||
| 1989 | } | |||
| 1990 | ||||
| 1991 | // Replace any remaining uses with the function as an i8*. | |||
| 1992 | // This can never directly be a callee, so we don't need to update CG. | |||
| 1993 | Prepare->replaceAllUsesWith(CastFn); | |||
| 1994 | Prepare->eraseFromParent(); | |||
| 1995 | ||||
| 1996 | // Kill dead bitcasts. | |||
| 1997 | while (auto *Cast = dyn_cast<BitCastInst>(CastFn)) { | |||
| 1998 | if (!Cast->use_empty()) | |||
| 1999 | break; | |||
| 2000 | CastFn = Cast->getOperand(0); | |||
| 2001 | Cast->eraseFromParent(); | |||
| 2002 | } | |||
| 2003 | } | |||
| 2004 | /// Replace a call to llvm.coro.prepare.retcon. | |||
| 2005 | static void replacePrepare(CallInst *Prepare, CallGraph &CG) { | |||
| 2006 | auto CastFn = Prepare->getArgOperand(0); // as an i8* | |||
| 2007 | auto Fn = CastFn->stripPointerCasts(); // as its original type | |||
| 2008 | ||||
| 2009 | // Find call graph nodes for the preparation. | |||
| 2010 | CallGraphNode *PrepareUserNode = nullptr, *FnNode = nullptr; | |||
| 2011 | if (auto ConcreteFn = dyn_cast<Function>(Fn)) { | |||
| 2012 | PrepareUserNode = CG[Prepare->getFunction()]; | |||
| 2013 | FnNode = CG[ConcreteFn]; | |||
| 2014 | } | |||
| 2015 | ||||
| 2016 | // Attempt to peephole this pattern: | |||
| 2017 | // %0 = bitcast [[TYPE]] @some_function to i8* | |||
| 2018 | // %1 = call @llvm.coro.prepare.retcon(i8* %0) | |||
| 2019 | // %2 = bitcast %1 to [[TYPE]] | |||
| 2020 | // ==> | |||
| 2021 | // %2 = @some_function | |||
| 2022 | for (auto UI = Prepare->use_begin(), UE = Prepare->use_end(); | |||
| 2023 | UI != UE; ) { | |||
| 2024 | // Look for bitcasts back to the original function type. | |||
| 2025 | auto *Cast = dyn_cast<BitCastInst>((UI++)->getUser()); | |||
| 2026 | if (!Cast || Cast->getType() != Fn->getType()) continue; | |||
| 2027 | ||||
| 2028 | // Check whether the replacement will introduce new direct calls. | |||
| 2029 | // If so, we'll need to update the call graph. | |||
| 2030 | if (PrepareUserNode) { | |||
| 2031 | for (auto &Use : Cast->uses()) { | |||
| 2032 | if (auto *CB = dyn_cast<CallBase>(Use.getUser())) { | |||
| 2033 | if (!CB->isCallee(&Use)) | |||
| 2034 | continue; | |||
| 2035 | PrepareUserNode->removeCallEdgeFor(*CB); | |||
| 2036 | PrepareUserNode->addCalledFunction(CB, FnNode); | |||
| 2037 | } | |||
| 2038 | } | |||
| 2039 | } | |||
| 2040 | ||||
| 2041 | // Replace and remove the cast. | |||
| 2042 | Cast->replaceAllUsesWith(Fn); | |||
| 2043 | Cast->eraseFromParent(); | |||
| 2044 | } | |||
| 2045 | ||||
| 2046 | // Replace any remaining uses with the function as an i8*. | |||
| 2047 | // This can never directly be a callee, so we don't need to update CG. | |||
| 2048 | Prepare->replaceAllUsesWith(CastFn); | |||
| 2049 | Prepare->eraseFromParent(); | |||
| 2050 | ||||
| 2051 | // Kill dead bitcasts. | |||
| 2052 | while (auto *Cast = dyn_cast<BitCastInst>(CastFn)) { | |||
| 2053 | if (!Cast->use_empty()) break; | |||
| 2054 | CastFn = Cast->getOperand(0); | |||
| 2055 | Cast->eraseFromParent(); | |||
| 2056 | } | |||
| 2057 | } | |||
| 2058 | ||||
| 2059 | static bool replaceAllPrepares(Function *PrepareFn, LazyCallGraph &CG, | |||
| 2060 | LazyCallGraph::SCC &C) { | |||
| 2061 | bool Changed = false; | |||
| 2062 | for (auto PI = PrepareFn->use_begin(), PE = PrepareFn->use_end(); PI != PE;) { | |||
| 2063 | // Intrinsics can only be used in calls. | |||
| 2064 | auto *Prepare = cast<CallInst>((PI++)->getUser()); | |||
| 2065 | replacePrepare(Prepare, CG, C); | |||
| 2066 | Changed = true; | |||
| 2067 | } | |||
| 2068 | ||||
| 2069 | return Changed; | |||
| 2070 | } | |||
| 2071 | ||||
| 2072 | /// Remove calls to llvm.coro.prepare.retcon, a barrier meant to prevent | |||
| 2073 | /// IPO from operating on calls to a retcon coroutine before it's been | |||
| 2074 | /// split. This is only safe to do after we've split all retcon | |||
| 2075 | /// coroutines in the module. We can do that this in this pass because | |||
| 2076 | /// this pass does promise to split all retcon coroutines (as opposed to | |||
| 2077 | /// switch coroutines, which are lowered in multiple stages). | |||
| 2078 | static bool replaceAllPrepares(Function *PrepareFn, CallGraph &CG) { | |||
| 2079 | bool Changed = false; | |||
| 2080 | for (auto PI = PrepareFn->use_begin(), PE = PrepareFn->use_end(); | |||
| 2081 | PI != PE; ) { | |||
| 2082 | // Intrinsics can only be used in calls. | |||
| 2083 | auto *Prepare = cast<CallInst>((PI++)->getUser()); | |||
| 2084 | replacePrepare(Prepare, CG); | |||
| 2085 | Changed = true; | |||
| 2086 | } | |||
| 2087 | ||||
| 2088 | return Changed; | |||
| 2089 | } | |||
| 2090 | ||||
| 2091 | static bool declaresCoroSplitIntrinsics(const Module &M) { | |||
| 2092 | return coro::declaresIntrinsics(M, {"llvm.coro.begin", | |||
| 2093 | "llvm.coro.prepare.retcon", | |||
| 2094 | "llvm.coro.prepare.async"}); | |||
| 2095 | } | |||
| 2096 | ||||
| 2097 | static void addPrepareFunction(const Module &M, | |||
| 2098 | SmallVectorImpl<Function *> &Fns, | |||
| 2099 | StringRef Name) { | |||
| 2100 | auto *PrepareFn = M.getFunction(Name); | |||
| 2101 | if (PrepareFn && !PrepareFn->use_empty()) | |||
| 2102 | Fns.push_back(PrepareFn); | |||
| 2103 | } | |||
| 2104 | ||||
| 2105 | PreservedAnalyses CoroSplitPass::run(LazyCallGraph::SCC &C, | |||
| 2106 | CGSCCAnalysisManager &AM, | |||
| 2107 | LazyCallGraph &CG, CGSCCUpdateResult &UR) { | |||
| 2108 | // NB: One invariant of a valid LazyCallGraph::SCC is that it must contain a | |||
| 2109 | // non-zero number of nodes, so we assume that here and grab the first | |||
| 2110 | // node's function's module. | |||
| 2111 | Module &M = *C.begin()->getFunction().getParent(); | |||
| 2112 | auto &FAM = | |||
| 2113 | AM.getResult<FunctionAnalysisManagerCGSCCProxy>(C, CG).getManager(); | |||
| 2114 | ||||
| 2115 | if (!declaresCoroSplitIntrinsics(M)) | |||
| ||||
| 2116 | return PreservedAnalyses::all(); | |||
| 2117 | ||||
| 2118 | // Check for uses of llvm.coro.prepare.retcon/async. | |||
| 2119 | SmallVector<Function *, 2> PrepareFns; | |||
| 2120 | addPrepareFunction(M, PrepareFns, "llvm.coro.prepare.retcon"); | |||
| 2121 | addPrepareFunction(M, PrepareFns, "llvm.coro.prepare.async"); | |||
| 2122 | ||||
| 2123 | // Find coroutines for processing. | |||
| 2124 | SmallVector<LazyCallGraph::Node *, 4> Coroutines; | |||
| 2125 | for (LazyCallGraph::Node &N : C) | |||
| 2126 | if (N.getFunction().hasFnAttribute(CORO_PRESPLIT_ATTR"coroutine.presplit")) | |||
| 2127 | Coroutines.push_back(&N); | |||
| 2128 | ||||
| 2129 | if (Coroutines.empty() && PrepareFns.empty()) | |||
| 2130 | return PreservedAnalyses::all(); | |||
| 2131 | ||||
| 2132 | if (Coroutines.empty()) { | |||
| 2133 | for (auto *PrepareFn : PrepareFns) { | |||
| 2134 | replaceAllPrepares(PrepareFn, CG, C); | |||
| 2135 | } | |||
| 2136 | } | |||
| 2137 | ||||
| 2138 | // Split all the coroutines. | |||
| 2139 | for (LazyCallGraph::Node *N : Coroutines) { | |||
| 2140 | Function &F = N->getFunction(); | |||
| 2141 | LLVM_DEBUG(dbgs() << "CoroSplit: Processing coroutine '" << F.getName()do { } while (false) | |||
| 2142 | << "' state: "do { } while (false) | |||
| 2143 | << F.getFnAttribute(CORO_PRESPLIT_ATTR).getValueAsString()do { } while (false) | |||
| 2144 | << "\n")do { } while (false); | |||
| 2145 | F.removeFnAttr(CORO_PRESPLIT_ATTR"coroutine.presplit"); | |||
| 2146 | ||||
| 2147 | SmallVector<Function *, 4> Clones; | |||
| 2148 | const coro::Shape Shape = splitCoroutine(F, Clones, ReuseFrameSlot); | |||
| 2149 | updateCallGraphAfterCoroutineSplit(*N, Shape, Clones, C, CG, AM, UR, FAM); | |||
| 2150 | ||||
| 2151 | if (!Shape.CoroSuspends.empty()) { | |||
| 2152 | // Run the CGSCC pipeline on the original and newly split functions. | |||
| 2153 | UR.CWorklist.insert(&C); | |||
| 2154 | for (Function *Clone : Clones) | |||
| 2155 | UR.CWorklist.insert(CG.lookupSCC(CG.get(*Clone))); | |||
| 2156 | } | |||
| 2157 | } | |||
| 2158 | ||||
| 2159 | if (!PrepareFns.empty()) { | |||
| 2160 | for (auto *PrepareFn : PrepareFns) { | |||
| 2161 | replaceAllPrepares(PrepareFn, CG, C); | |||
| 2162 | } | |||
| 2163 | } | |||
| 2164 | ||||
| 2165 | return PreservedAnalyses::none(); | |||
| 2166 | } | |||
| 2167 | ||||
| 2168 | namespace { | |||
| 2169 | ||||
| 2170 | // We present a coroutine to LLVM as an ordinary function with suspension | |||
| 2171 | // points marked up with intrinsics. We let the optimizer party on the coroutine | |||
| 2172 | // as a single function for as long as possible. Shortly before the coroutine is | |||
| 2173 | // eligible to be inlined into its callers, we split up the coroutine into parts | |||
| 2174 | // corresponding to initial, resume and destroy invocations of the coroutine, | |||
| 2175 | // add them to the current SCC and restart the IPO pipeline to optimize the | |||
| 2176 | // coroutine subfunctions we extracted before proceeding to the caller of the | |||
| 2177 | // coroutine. | |||
| 2178 | struct CoroSplitLegacy : public CallGraphSCCPass { | |||
| 2179 | static char ID; // Pass identification, replacement for typeid | |||
| 2180 | ||||
| 2181 | CoroSplitLegacy(bool ReuseFrameSlot = false) | |||
| 2182 | : CallGraphSCCPass(ID), ReuseFrameSlot(ReuseFrameSlot) { | |||
| 2183 | initializeCoroSplitLegacyPass(*PassRegistry::getPassRegistry()); | |||
| 2184 | } | |||
| 2185 | ||||
| 2186 | bool Run = false; | |||
| 2187 | bool ReuseFrameSlot; | |||
| 2188 | ||||
| 2189 | // A coroutine is identified by the presence of coro.begin intrinsic, if | |||
| 2190 | // we don't have any, this pass has nothing to do. | |||
| 2191 | bool doInitialization(CallGraph &CG) override { | |||
| 2192 | Run = declaresCoroSplitIntrinsics(CG.getModule()); | |||
| 2193 | return CallGraphSCCPass::doInitialization(CG); | |||
| 2194 | } | |||
| 2195 | ||||
| 2196 | bool runOnSCC(CallGraphSCC &SCC) override { | |||
| 2197 | if (!Run) | |||
| 2198 | return false; | |||
| 2199 | ||||
| 2200 | // Check for uses of llvm.coro.prepare.retcon. | |||
| 2201 | SmallVector<Function *, 2> PrepareFns; | |||
| 2202 | auto &M = SCC.getCallGraph().getModule(); | |||
| 2203 | addPrepareFunction(M, PrepareFns, "llvm.coro.prepare.retcon"); | |||
| 2204 | addPrepareFunction(M, PrepareFns, "llvm.coro.prepare.async"); | |||
| 2205 | ||||
| 2206 | // Find coroutines for processing. | |||
| 2207 | SmallVector<Function *, 4> Coroutines; | |||
| 2208 | for (CallGraphNode *CGN : SCC) | |||
| 2209 | if (auto *F = CGN->getFunction()) | |||
| 2210 | if (F->hasFnAttribute(CORO_PRESPLIT_ATTR"coroutine.presplit")) | |||
| 2211 | Coroutines.push_back(F); | |||
| 2212 | ||||
| 2213 | if (Coroutines.empty() && PrepareFns.empty()) | |||
| 2214 | return false; | |||
| 2215 | ||||
| 2216 | CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph(); | |||
| 2217 | ||||
| 2218 | if (Coroutines.empty()) { | |||
| 2219 | bool Changed = false; | |||
| 2220 | for (auto *PrepareFn : PrepareFns) | |||
| 2221 | Changed |= replaceAllPrepares(PrepareFn, CG); | |||
| 2222 | return Changed; | |||
| 2223 | } | |||
| 2224 | ||||
| 2225 | createDevirtTriggerFunc(CG, SCC); | |||
| 2226 | ||||
| 2227 | // Split all the coroutines. | |||
| 2228 | for (Function *F : Coroutines) { | |||
| 2229 | Attribute Attr = F->getFnAttribute(CORO_PRESPLIT_ATTR"coroutine.presplit"); | |||
| 2230 | StringRef Value = Attr.getValueAsString(); | |||
| 2231 | LLVM_DEBUG(dbgs() << "CoroSplit: Processing coroutine '" << F->getName()do { } while (false) | |||
| 2232 | << "' state: " << Value << "\n")do { } while (false); | |||
| 2233 | // Async lowering marks coroutines to trigger a restart of the pipeline | |||
| 2234 | // after it has split them. | |||
| 2235 | if (Value == ASYNC_RESTART_AFTER_SPLIT"2") { | |||
| 2236 | F->removeFnAttr(CORO_PRESPLIT_ATTR"coroutine.presplit"); | |||
| 2237 | continue; | |||
| 2238 | } | |||
| 2239 | if (Value == UNPREPARED_FOR_SPLIT"0") { | |||
| 2240 | prepareForSplit(*F, CG); | |||
| 2241 | continue; | |||
| 2242 | } | |||
| 2243 | F->removeFnAttr(CORO_PRESPLIT_ATTR"coroutine.presplit"); | |||
| 2244 | ||||
| 2245 | SmallVector<Function *, 4> Clones; | |||
| 2246 | const coro::Shape Shape = splitCoroutine(*F, Clones, ReuseFrameSlot); | |||
| 2247 | updateCallGraphAfterCoroutineSplit(*F, Shape, Clones, CG, SCC); | |||
| 2248 | if (Shape.ABI == coro::ABI::Async) { | |||
| 2249 | // Restart SCC passes. | |||
| 2250 | // Mark function for CoroElide pass. It will devirtualize causing a | |||
| 2251 | // restart of the SCC pipeline. | |||
| 2252 | prepareForSplit(*F, CG, true /*MarkForAsyncRestart*/); | |||
| 2253 | } | |||
| 2254 | } | |||
| 2255 | ||||
| 2256 | for (auto *PrepareFn : PrepareFns) | |||
| 2257 | replaceAllPrepares(PrepareFn, CG); | |||
| 2258 | ||||
| 2259 | return true; | |||
| 2260 | } | |||
| 2261 | ||||
| 2262 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
| 2263 | CallGraphSCCPass::getAnalysisUsage(AU); | |||
| 2264 | } | |||
| 2265 | ||||
| 2266 | StringRef getPassName() const override { return "Coroutine Splitting"; } | |||
| 2267 | }; | |||
| 2268 | ||||
| 2269 | } // end anonymous namespace | |||
| 2270 | ||||
| 2271 | char CoroSplitLegacy::ID = 0; | |||
| 2272 | ||||
| 2273 | INITIALIZE_PASS_BEGIN(static void *initializeCoroSplitLegacyPassOnce(PassRegistry & Registry) { | |||
| 2274 | CoroSplitLegacy, "coro-split",static void *initializeCoroSplitLegacyPassOnce(PassRegistry & Registry) { | |||
| 2275 | "Split coroutine into a set of functions driving its state machine", false,static void *initializeCoroSplitLegacyPassOnce(PassRegistry & Registry) { | |||
| 2276 | false)static void *initializeCoroSplitLegacyPassOnce(PassRegistry & Registry) { | |||
| 2277 | INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)initializeCallGraphWrapperPassPass(Registry); | |||
| 2278 | INITIALIZE_PASS_END(PassInfo *PI = new PassInfo( "Split coroutine into a set of functions driving its state machine" , "coro-split", &CoroSplitLegacy::ID, PassInfo::NormalCtor_t (callDefaultCtor<CoroSplitLegacy>), false, false); Registry .registerPass(*PI, true); return PI; } static llvm::once_flag InitializeCoroSplitLegacyPassFlag; void llvm::initializeCoroSplitLegacyPass (PassRegistry &Registry) { llvm::call_once(InitializeCoroSplitLegacyPassFlag , initializeCoroSplitLegacyPassOnce, std::ref(Registry)); } | |||
| 2279 | CoroSplitLegacy, "coro-split",PassInfo *PI = new PassInfo( "Split coroutine into a set of functions driving its state machine" , "coro-split", &CoroSplitLegacy::ID, PassInfo::NormalCtor_t (callDefaultCtor<CoroSplitLegacy>), false, false); Registry .registerPass(*PI, true); return PI; } static llvm::once_flag InitializeCoroSplitLegacyPassFlag; void llvm::initializeCoroSplitLegacyPass (PassRegistry &Registry) { llvm::call_once(InitializeCoroSplitLegacyPassFlag , initializeCoroSplitLegacyPassOnce, std::ref(Registry)); } | |||
| 2280 | "Split coroutine into a set of functions driving its state machine", false,PassInfo *PI = new PassInfo( "Split coroutine into a set of functions driving its state machine" , "coro-split", &CoroSplitLegacy::ID, PassInfo::NormalCtor_t (callDefaultCtor<CoroSplitLegacy>), false, false); Registry .registerPass(*PI, true); return PI; } static llvm::once_flag InitializeCoroSplitLegacyPassFlag; void llvm::initializeCoroSplitLegacyPass (PassRegistry &Registry) { llvm::call_once(InitializeCoroSplitLegacyPassFlag , initializeCoroSplitLegacyPassOnce, std::ref(Registry)); } | |||
| 2281 | false)PassInfo *PI = new PassInfo( "Split coroutine into a set of functions driving its state machine" , "coro-split", &CoroSplitLegacy::ID, PassInfo::NormalCtor_t (callDefaultCtor<CoroSplitLegacy>), false, false); Registry .registerPass(*PI, true); return PI; } static llvm::once_flag InitializeCoroSplitLegacyPassFlag; void llvm::initializeCoroSplitLegacyPass (PassRegistry &Registry) { llvm::call_once(InitializeCoroSplitLegacyPassFlag , initializeCoroSplitLegacyPassOnce, std::ref(Registry)); } | |||
| 2282 | ||||
| 2283 | Pass *llvm::createCoroSplitLegacyPass(bool ReuseFrameSlot) { | |||
| 2284 | return new CoroSplitLegacy(ReuseFrameSlot); | |||
| 2285 | } |