| File: | llvm/lib/Transforms/Utils/CodeExtractor.cpp |
| Warning: | line 944, column 33 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===- CodeExtractor.cpp - Pull code region into a new function -----------===// | |||
| 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 | // | |||
| 9 | // This file implements the interface to tear out a code region, such as an | |||
| 10 | // individual loop or a parallel section, into a new function, replacing it with | |||
| 11 | // a call to the new function. | |||
| 12 | // | |||
| 13 | //===----------------------------------------------------------------------===// | |||
| 14 | ||||
| 15 | #include "llvm/Transforms/Utils/CodeExtractor.h" | |||
| 16 | #include "llvm/ADT/ArrayRef.h" | |||
| 17 | #include "llvm/ADT/DenseMap.h" | |||
| 18 | #include "llvm/ADT/Optional.h" | |||
| 19 | #include "llvm/ADT/STLExtras.h" | |||
| 20 | #include "llvm/ADT/SetVector.h" | |||
| 21 | #include "llvm/ADT/SmallPtrSet.h" | |||
| 22 | #include "llvm/ADT/SmallVector.h" | |||
| 23 | #include "llvm/Analysis/AssumptionCache.h" | |||
| 24 | #include "llvm/Analysis/BlockFrequencyInfo.h" | |||
| 25 | #include "llvm/Analysis/BlockFrequencyInfoImpl.h" | |||
| 26 | #include "llvm/Analysis/BranchProbabilityInfo.h" | |||
| 27 | #include "llvm/Analysis/LoopInfo.h" | |||
| 28 | #include "llvm/IR/Argument.h" | |||
| 29 | #include "llvm/IR/Attributes.h" | |||
| 30 | #include "llvm/IR/BasicBlock.h" | |||
| 31 | #include "llvm/IR/CFG.h" | |||
| 32 | #include "llvm/IR/Constant.h" | |||
| 33 | #include "llvm/IR/Constants.h" | |||
| 34 | #include "llvm/IR/DIBuilder.h" | |||
| 35 | #include "llvm/IR/DataLayout.h" | |||
| 36 | #include "llvm/IR/DebugInfoMetadata.h" | |||
| 37 | #include "llvm/IR/DerivedTypes.h" | |||
| 38 | #include "llvm/IR/Dominators.h" | |||
| 39 | #include "llvm/IR/Function.h" | |||
| 40 | #include "llvm/IR/GlobalValue.h" | |||
| 41 | #include "llvm/IR/InstIterator.h" | |||
| 42 | #include "llvm/IR/InstrTypes.h" | |||
| 43 | #include "llvm/IR/Instruction.h" | |||
| 44 | #include "llvm/IR/Instructions.h" | |||
| 45 | #include "llvm/IR/IntrinsicInst.h" | |||
| 46 | #include "llvm/IR/Intrinsics.h" | |||
| 47 | #include "llvm/IR/LLVMContext.h" | |||
| 48 | #include "llvm/IR/MDBuilder.h" | |||
| 49 | #include "llvm/IR/Module.h" | |||
| 50 | #include "llvm/IR/PatternMatch.h" | |||
| 51 | #include "llvm/IR/Type.h" | |||
| 52 | #include "llvm/IR/User.h" | |||
| 53 | #include "llvm/IR/Value.h" | |||
| 54 | #include "llvm/IR/Verifier.h" | |||
| 55 | #include "llvm/Pass.h" | |||
| 56 | #include "llvm/Support/BlockFrequency.h" | |||
| 57 | #include "llvm/Support/BranchProbability.h" | |||
| 58 | #include "llvm/Support/Casting.h" | |||
| 59 | #include "llvm/Support/CommandLine.h" | |||
| 60 | #include "llvm/Support/Debug.h" | |||
| 61 | #include "llvm/Support/ErrorHandling.h" | |||
| 62 | #include "llvm/Support/raw_ostream.h" | |||
| 63 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | |||
| 64 | #include "llvm/Transforms/Utils/Local.h" | |||
| 65 | #include <cassert> | |||
| 66 | #include <cstdint> | |||
| 67 | #include <iterator> | |||
| 68 | #include <map> | |||
| 69 | #include <set> | |||
| 70 | #include <utility> | |||
| 71 | #include <vector> | |||
| 72 | ||||
| 73 | using namespace llvm; | |||
| 74 | using namespace llvm::PatternMatch; | |||
| 75 | using ProfileCount = Function::ProfileCount; | |||
| 76 | ||||
| 77 | #define DEBUG_TYPE"code-extractor" "code-extractor" | |||
| 78 | ||||
| 79 | // Provide a command-line option to aggregate function arguments into a struct | |||
| 80 | // for functions produced by the code extractor. This is useful when converting | |||
| 81 | // extracted functions to pthread-based code, as only one argument (void*) can | |||
| 82 | // be passed in to pthread_create(). | |||
| 83 | static cl::opt<bool> | |||
| 84 | AggregateArgsOpt("aggregate-extracted-args", cl::Hidden, | |||
| 85 | cl::desc("Aggregate arguments to code-extracted functions")); | |||
| 86 | ||||
| 87 | /// Test whether a block is valid for extraction. | |||
| 88 | static bool isBlockValidForExtraction(const BasicBlock &BB, | |||
| 89 | const SetVector<BasicBlock *> &Result, | |||
| 90 | bool AllowVarArgs, bool AllowAlloca) { | |||
| 91 | // taking the address of a basic block moved to another function is illegal | |||
| 92 | if (BB.hasAddressTaken()) | |||
| 93 | return false; | |||
| 94 | ||||
| 95 | // don't hoist code that uses another basicblock address, as it's likely to | |||
| 96 | // lead to unexpected behavior, like cross-function jumps | |||
| 97 | SmallPtrSet<User const *, 16> Visited; | |||
| 98 | SmallVector<User const *, 16> ToVisit; | |||
| 99 | ||||
| 100 | for (Instruction const &Inst : BB) | |||
| 101 | ToVisit.push_back(&Inst); | |||
| 102 | ||||
| 103 | while (!ToVisit.empty()) { | |||
| 104 | User const *Curr = ToVisit.pop_back_val(); | |||
| 105 | if (!Visited.insert(Curr).second) | |||
| 106 | continue; | |||
| 107 | if (isa<BlockAddress const>(Curr)) | |||
| 108 | return false; // even a reference to self is likely to be not compatible | |||
| 109 | ||||
| 110 | if (isa<Instruction>(Curr) && cast<Instruction>(Curr)->getParent() != &BB) | |||
| 111 | continue; | |||
| 112 | ||||
| 113 | for (auto const &U : Curr->operands()) { | |||
| 114 | if (auto *UU = dyn_cast<User>(U)) | |||
| 115 | ToVisit.push_back(UU); | |||
| 116 | } | |||
| 117 | } | |||
| 118 | ||||
| 119 | // If explicitly requested, allow vastart and alloca. For invoke instructions | |||
| 120 | // verify that extraction is valid. | |||
| 121 | for (BasicBlock::const_iterator I = BB.begin(), E = BB.end(); I != E; ++I) { | |||
| 122 | if (isa<AllocaInst>(I)) { | |||
| 123 | if (!AllowAlloca) | |||
| 124 | return false; | |||
| 125 | continue; | |||
| 126 | } | |||
| 127 | ||||
| 128 | if (const auto *II = dyn_cast<InvokeInst>(I)) { | |||
| 129 | // Unwind destination (either a landingpad, catchswitch, or cleanuppad) | |||
| 130 | // must be a part of the subgraph which is being extracted. | |||
| 131 | if (auto *UBB = II->getUnwindDest()) | |||
| 132 | if (!Result.count(UBB)) | |||
| 133 | return false; | |||
| 134 | continue; | |||
| 135 | } | |||
| 136 | ||||
| 137 | // All catch handlers of a catchswitch instruction as well as the unwind | |||
| 138 | // destination must be in the subgraph. | |||
| 139 | if (const auto *CSI = dyn_cast<CatchSwitchInst>(I)) { | |||
| 140 | if (auto *UBB = CSI->getUnwindDest()) | |||
| 141 | if (!Result.count(UBB)) | |||
| 142 | return false; | |||
| 143 | for (auto *HBB : CSI->handlers()) | |||
| 144 | if (!Result.count(const_cast<BasicBlock*>(HBB))) | |||
| 145 | return false; | |||
| 146 | continue; | |||
| 147 | } | |||
| 148 | ||||
| 149 | // Make sure that entire catch handler is within subgraph. It is sufficient | |||
| 150 | // to check that catch return's block is in the list. | |||
| 151 | if (const auto *CPI = dyn_cast<CatchPadInst>(I)) { | |||
| 152 | for (const auto *U : CPI->users()) | |||
| 153 | if (const auto *CRI = dyn_cast<CatchReturnInst>(U)) | |||
| 154 | if (!Result.count(const_cast<BasicBlock*>(CRI->getParent()))) | |||
| 155 | return false; | |||
| 156 | continue; | |||
| 157 | } | |||
| 158 | ||||
| 159 | // And do similar checks for cleanup handler - the entire handler must be | |||
| 160 | // in subgraph which is going to be extracted. For cleanup return should | |||
| 161 | // additionally check that the unwind destination is also in the subgraph. | |||
| 162 | if (const auto *CPI = dyn_cast<CleanupPadInst>(I)) { | |||
| 163 | for (const auto *U : CPI->users()) | |||
| 164 | if (const auto *CRI = dyn_cast<CleanupReturnInst>(U)) | |||
| 165 | if (!Result.count(const_cast<BasicBlock*>(CRI->getParent()))) | |||
| 166 | return false; | |||
| 167 | continue; | |||
| 168 | } | |||
| 169 | if (const auto *CRI = dyn_cast<CleanupReturnInst>(I)) { | |||
| 170 | if (auto *UBB = CRI->getUnwindDest()) | |||
| 171 | if (!Result.count(UBB)) | |||
| 172 | return false; | |||
| 173 | continue; | |||
| 174 | } | |||
| 175 | ||||
| 176 | if (const CallInst *CI = dyn_cast<CallInst>(I)) { | |||
| 177 | if (const Function *F = CI->getCalledFunction()) { | |||
| 178 | auto IID = F->getIntrinsicID(); | |||
| 179 | if (IID == Intrinsic::vastart) { | |||
| 180 | if (AllowVarArgs) | |||
| 181 | continue; | |||
| 182 | else | |||
| 183 | return false; | |||
| 184 | } | |||
| 185 | ||||
| 186 | // Currently, we miscompile outlined copies of eh_typid_for. There are | |||
| 187 | // proposals for fixing this in llvm.org/PR39545. | |||
| 188 | if (IID == Intrinsic::eh_typeid_for) | |||
| 189 | return false; | |||
| 190 | } | |||
| 191 | } | |||
| 192 | } | |||
| 193 | ||||
| 194 | return true; | |||
| 195 | } | |||
| 196 | ||||
| 197 | /// Build a set of blocks to extract if the input blocks are viable. | |||
| 198 | static SetVector<BasicBlock *> | |||
| 199 | buildExtractionBlockSet(ArrayRef<BasicBlock *> BBs, DominatorTree *DT, | |||
| 200 | bool AllowVarArgs, bool AllowAlloca) { | |||
| 201 | assert(!BBs.empty() && "The set of blocks to extract must be non-empty")((!BBs.empty() && "The set of blocks to extract must be non-empty" ) ? static_cast<void> (0) : __assert_fail ("!BBs.empty() && \"The set of blocks to extract must be non-empty\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 201, __PRETTY_FUNCTION__)); | |||
| 202 | SetVector<BasicBlock *> Result; | |||
| 203 | ||||
| 204 | // Loop over the blocks, adding them to our set-vector, and aborting with an | |||
| 205 | // empty set if we encounter invalid blocks. | |||
| 206 | for (BasicBlock *BB : BBs) { | |||
| 207 | // If this block is dead, don't process it. | |||
| 208 | if (DT && !DT->isReachableFromEntry(BB)) | |||
| 209 | continue; | |||
| 210 | ||||
| 211 | if (!Result.insert(BB)) | |||
| 212 | llvm_unreachable("Repeated basic blocks in extraction input")::llvm::llvm_unreachable_internal("Repeated basic blocks in extraction input" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 212); | |||
| 213 | } | |||
| 214 | ||||
| 215 | LLVM_DEBUG(dbgs() << "Region front block: " << Result.front()->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "Region front block: " << Result.front()->getName() << '\n'; } } while (false ) | |||
| 216 | << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "Region front block: " << Result.front()->getName() << '\n'; } } while (false ); | |||
| 217 | ||||
| 218 | for (auto *BB : Result) { | |||
| 219 | if (!isBlockValidForExtraction(*BB, Result, AllowVarArgs, AllowAlloca)) | |||
| 220 | return {}; | |||
| 221 | ||||
| 222 | // Make sure that the first block is not a landing pad. | |||
| 223 | if (BB == Result.front()) { | |||
| 224 | if (BB->isEHPad()) { | |||
| 225 | LLVM_DEBUG(dbgs() << "The first block cannot be an unwind block\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "The first block cannot be an unwind block\n" ; } } while (false); | |||
| 226 | return {}; | |||
| 227 | } | |||
| 228 | continue; | |||
| 229 | } | |||
| 230 | ||||
| 231 | // All blocks other than the first must not have predecessors outside of | |||
| 232 | // the subgraph which is being extracted. | |||
| 233 | for (auto *PBB : predecessors(BB)) | |||
| 234 | if (!Result.count(PBB)) { | |||
| 235 | LLVM_DEBUG(dbgs() << "No blocks in this region may have entries from "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "No blocks in this region may have entries from " "outside the region except for the first block!\n" << "Problematic source BB: " << BB->getName() << "\n" << "Problematic destination BB: " << PBB->getName() << "\n"; } } while (false) | |||
| 236 | "outside the region except for the first block!\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "No blocks in this region may have entries from " "outside the region except for the first block!\n" << "Problematic source BB: " << BB->getName() << "\n" << "Problematic destination BB: " << PBB->getName() << "\n"; } } while (false) | |||
| 237 | << "Problematic source BB: " << BB->getName() << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "No blocks in this region may have entries from " "outside the region except for the first block!\n" << "Problematic source BB: " << BB->getName() << "\n" << "Problematic destination BB: " << PBB->getName() << "\n"; } } while (false) | |||
| 238 | << "Problematic destination BB: " << PBB->getName()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "No blocks in this region may have entries from " "outside the region except for the first block!\n" << "Problematic source BB: " << BB->getName() << "\n" << "Problematic destination BB: " << PBB->getName() << "\n"; } } while (false) | |||
| 239 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "No blocks in this region may have entries from " "outside the region except for the first block!\n" << "Problematic source BB: " << BB->getName() << "\n" << "Problematic destination BB: " << PBB->getName() << "\n"; } } while (false); | |||
| 240 | return {}; | |||
| 241 | } | |||
| 242 | } | |||
| 243 | ||||
| 244 | return Result; | |||
| 245 | } | |||
| 246 | ||||
| 247 | CodeExtractor::CodeExtractor(ArrayRef<BasicBlock *> BBs, DominatorTree *DT, | |||
| 248 | bool AggregateArgs, BlockFrequencyInfo *BFI, | |||
| 249 | BranchProbabilityInfo *BPI, AssumptionCache *AC, | |||
| 250 | bool AllowVarArgs, bool AllowAlloca, | |||
| 251 | std::string Suffix) | |||
| 252 | : DT(DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI), | |||
| 253 | BPI(BPI), AC(AC), AllowVarArgs(AllowVarArgs), | |||
| 254 | Blocks(buildExtractionBlockSet(BBs, DT, AllowVarArgs, AllowAlloca)), | |||
| 255 | Suffix(Suffix) {} | |||
| 256 | ||||
| 257 | CodeExtractor::CodeExtractor(DominatorTree &DT, Loop &L, bool AggregateArgs, | |||
| 258 | BlockFrequencyInfo *BFI, | |||
| 259 | BranchProbabilityInfo *BPI, AssumptionCache *AC, | |||
| 260 | std::string Suffix) | |||
| 261 | : DT(&DT), AggregateArgs(AggregateArgs || AggregateArgsOpt), BFI(BFI), | |||
| 262 | BPI(BPI), AC(AC), AllowVarArgs(false), | |||
| 263 | Blocks(buildExtractionBlockSet(L.getBlocks(), &DT, | |||
| 264 | /* AllowVarArgs */ false, | |||
| 265 | /* AllowAlloca */ false)), | |||
| 266 | Suffix(Suffix) {} | |||
| 267 | ||||
| 268 | /// definedInRegion - Return true if the specified value is defined in the | |||
| 269 | /// extracted region. | |||
| 270 | static bool definedInRegion(const SetVector<BasicBlock *> &Blocks, Value *V) { | |||
| 271 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
| 272 | if (Blocks.count(I->getParent())) | |||
| 273 | return true; | |||
| 274 | return false; | |||
| 275 | } | |||
| 276 | ||||
| 277 | /// definedInCaller - Return true if the specified value is defined in the | |||
| 278 | /// function being code extracted, but not in the region being extracted. | |||
| 279 | /// These values must be passed in as live-ins to the function. | |||
| 280 | static bool definedInCaller(const SetVector<BasicBlock *> &Blocks, Value *V) { | |||
| 281 | if (isa<Argument>(V)) return true; | |||
| 282 | if (Instruction *I = dyn_cast<Instruction>(V)) | |||
| 283 | if (!Blocks.count(I->getParent())) | |||
| 284 | return true; | |||
| 285 | return false; | |||
| 286 | } | |||
| 287 | ||||
| 288 | static BasicBlock *getCommonExitBlock(const SetVector<BasicBlock *> &Blocks) { | |||
| 289 | BasicBlock *CommonExitBlock = nullptr; | |||
| 290 | auto hasNonCommonExitSucc = [&](BasicBlock *Block) { | |||
| 291 | for (auto *Succ : successors(Block)) { | |||
| 292 | // Internal edges, ok. | |||
| 293 | if (Blocks.count(Succ)) | |||
| 294 | continue; | |||
| 295 | if (!CommonExitBlock) { | |||
| 296 | CommonExitBlock = Succ; | |||
| 297 | continue; | |||
| 298 | } | |||
| 299 | if (CommonExitBlock != Succ) | |||
| 300 | return true; | |||
| 301 | } | |||
| 302 | return false; | |||
| 303 | }; | |||
| 304 | ||||
| 305 | if (any_of(Blocks, hasNonCommonExitSucc)) | |||
| 306 | return nullptr; | |||
| 307 | ||||
| 308 | return CommonExitBlock; | |||
| 309 | } | |||
| 310 | ||||
| 311 | CodeExtractorAnalysisCache::CodeExtractorAnalysisCache(Function &F) { | |||
| 312 | for (BasicBlock &BB : F) { | |||
| 313 | for (Instruction &II : BB.instructionsWithoutDebug()) | |||
| 314 | if (auto *AI = dyn_cast<AllocaInst>(&II)) | |||
| 315 | Allocas.push_back(AI); | |||
| 316 | ||||
| 317 | findSideEffectInfoForBlock(BB); | |||
| 318 | } | |||
| 319 | } | |||
| 320 | ||||
| 321 | void CodeExtractorAnalysisCache::findSideEffectInfoForBlock(BasicBlock &BB) { | |||
| 322 | for (Instruction &II : BB.instructionsWithoutDebug()) { | |||
| 323 | unsigned Opcode = II.getOpcode(); | |||
| 324 | Value *MemAddr = nullptr; | |||
| 325 | switch (Opcode) { | |||
| 326 | case Instruction::Store: | |||
| 327 | case Instruction::Load: { | |||
| 328 | if (Opcode == Instruction::Store) { | |||
| 329 | StoreInst *SI = cast<StoreInst>(&II); | |||
| 330 | MemAddr = SI->getPointerOperand(); | |||
| 331 | } else { | |||
| 332 | LoadInst *LI = cast<LoadInst>(&II); | |||
| 333 | MemAddr = LI->getPointerOperand(); | |||
| 334 | } | |||
| 335 | // Global variable can not be aliased with locals. | |||
| 336 | if (dyn_cast<Constant>(MemAddr)) | |||
| 337 | break; | |||
| 338 | Value *Base = MemAddr->stripInBoundsConstantOffsets(); | |||
| 339 | if (!isa<AllocaInst>(Base)) { | |||
| 340 | SideEffectingBlocks.insert(&BB); | |||
| 341 | return; | |||
| 342 | } | |||
| 343 | BaseMemAddrs[&BB].insert(Base); | |||
| 344 | break; | |||
| 345 | } | |||
| 346 | default: { | |||
| 347 | IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(&II); | |||
| 348 | if (IntrInst) { | |||
| 349 | if (IntrInst->isLifetimeStartOrEnd()) | |||
| 350 | break; | |||
| 351 | SideEffectingBlocks.insert(&BB); | |||
| 352 | return; | |||
| 353 | } | |||
| 354 | // Treat all the other cases conservatively if it has side effects. | |||
| 355 | if (II.mayHaveSideEffects()) { | |||
| 356 | SideEffectingBlocks.insert(&BB); | |||
| 357 | return; | |||
| 358 | } | |||
| 359 | } | |||
| 360 | } | |||
| 361 | } | |||
| 362 | } | |||
| 363 | ||||
| 364 | bool CodeExtractorAnalysisCache::doesBlockContainClobberOfAddr( | |||
| 365 | BasicBlock &BB, AllocaInst *Addr) const { | |||
| 366 | if (SideEffectingBlocks.count(&BB)) | |||
| 367 | return true; | |||
| 368 | auto It = BaseMemAddrs.find(&BB); | |||
| 369 | if (It != BaseMemAddrs.end()) | |||
| 370 | return It->second.count(Addr); | |||
| 371 | return false; | |||
| 372 | } | |||
| 373 | ||||
| 374 | bool CodeExtractor::isLegalToShrinkwrapLifetimeMarkers( | |||
| 375 | const CodeExtractorAnalysisCache &CEAC, Instruction *Addr) const { | |||
| 376 | AllocaInst *AI = cast<AllocaInst>(Addr->stripInBoundsConstantOffsets()); | |||
| 377 | Function *Func = (*Blocks.begin())->getParent(); | |||
| 378 | for (BasicBlock &BB : *Func) { | |||
| 379 | if (Blocks.count(&BB)) | |||
| 380 | continue; | |||
| 381 | if (CEAC.doesBlockContainClobberOfAddr(BB, AI)) | |||
| 382 | return false; | |||
| 383 | } | |||
| 384 | return true; | |||
| 385 | } | |||
| 386 | ||||
| 387 | BasicBlock * | |||
| 388 | CodeExtractor::findOrCreateBlockForHoisting(BasicBlock *CommonExitBlock) { | |||
| 389 | BasicBlock *SinglePredFromOutlineRegion = nullptr; | |||
| 390 | assert(!Blocks.count(CommonExitBlock) &&((!Blocks.count(CommonExitBlock) && "Expect a block outside the region!" ) ? static_cast<void> (0) : __assert_fail ("!Blocks.count(CommonExitBlock) && \"Expect a block outside the region!\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 391, __PRETTY_FUNCTION__)) | |||
| 391 | "Expect a block outside the region!")((!Blocks.count(CommonExitBlock) && "Expect a block outside the region!" ) ? static_cast<void> (0) : __assert_fail ("!Blocks.count(CommonExitBlock) && \"Expect a block outside the region!\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 391, __PRETTY_FUNCTION__)); | |||
| 392 | for (auto *Pred : predecessors(CommonExitBlock)) { | |||
| 393 | if (!Blocks.count(Pred)) | |||
| 394 | continue; | |||
| 395 | if (!SinglePredFromOutlineRegion) { | |||
| 396 | SinglePredFromOutlineRegion = Pred; | |||
| 397 | } else if (SinglePredFromOutlineRegion != Pred) { | |||
| 398 | SinglePredFromOutlineRegion = nullptr; | |||
| 399 | break; | |||
| 400 | } | |||
| 401 | } | |||
| 402 | ||||
| 403 | if (SinglePredFromOutlineRegion) | |||
| 404 | return SinglePredFromOutlineRegion; | |||
| 405 | ||||
| 406 | #ifndef NDEBUG | |||
| 407 | auto getFirstPHI = [](BasicBlock *BB) { | |||
| 408 | BasicBlock::iterator I = BB->begin(); | |||
| 409 | PHINode *FirstPhi = nullptr; | |||
| 410 | while (I != BB->end()) { | |||
| 411 | PHINode *Phi = dyn_cast<PHINode>(I); | |||
| 412 | if (!Phi) | |||
| 413 | break; | |||
| 414 | if (!FirstPhi) { | |||
| 415 | FirstPhi = Phi; | |||
| 416 | break; | |||
| 417 | } | |||
| 418 | } | |||
| 419 | return FirstPhi; | |||
| 420 | }; | |||
| 421 | // If there are any phi nodes, the single pred either exists or has already | |||
| 422 | // be created before code extraction. | |||
| 423 | assert(!getFirstPHI(CommonExitBlock) && "Phi not expected")((!getFirstPHI(CommonExitBlock) && "Phi not expected" ) ? static_cast<void> (0) : __assert_fail ("!getFirstPHI(CommonExitBlock) && \"Phi not expected\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 423, __PRETTY_FUNCTION__)); | |||
| 424 | #endif | |||
| 425 | ||||
| 426 | BasicBlock *NewExitBlock = CommonExitBlock->splitBasicBlock( | |||
| 427 | CommonExitBlock->getFirstNonPHI()->getIterator()); | |||
| 428 | ||||
| 429 | for (auto PI = pred_begin(CommonExitBlock), PE = pred_end(CommonExitBlock); | |||
| 430 | PI != PE;) { | |||
| 431 | BasicBlock *Pred = *PI++; | |||
| 432 | if (Blocks.count(Pred)) | |||
| 433 | continue; | |||
| 434 | Pred->getTerminator()->replaceUsesOfWith(CommonExitBlock, NewExitBlock); | |||
| 435 | } | |||
| 436 | // Now add the old exit block to the outline region. | |||
| 437 | Blocks.insert(CommonExitBlock); | |||
| 438 | return CommonExitBlock; | |||
| 439 | } | |||
| 440 | ||||
| 441 | // Find the pair of life time markers for address 'Addr' that are either | |||
| 442 | // defined inside the outline region or can legally be shrinkwrapped into the | |||
| 443 | // outline region. If there are not other untracked uses of the address, return | |||
| 444 | // the pair of markers if found; otherwise return a pair of nullptr. | |||
| 445 | CodeExtractor::LifetimeMarkerInfo | |||
| 446 | CodeExtractor::getLifetimeMarkers(const CodeExtractorAnalysisCache &CEAC, | |||
| 447 | Instruction *Addr, | |||
| 448 | BasicBlock *ExitBlock) const { | |||
| 449 | LifetimeMarkerInfo Info; | |||
| 450 | ||||
| 451 | for (User *U : Addr->users()) { | |||
| 452 | IntrinsicInst *IntrInst = dyn_cast<IntrinsicInst>(U); | |||
| 453 | if (IntrInst) { | |||
| 454 | if (IntrInst->getIntrinsicID() == Intrinsic::lifetime_start) { | |||
| 455 | // Do not handle the case where Addr has multiple start markers. | |||
| 456 | if (Info.LifeStart) | |||
| 457 | return {}; | |||
| 458 | Info.LifeStart = IntrInst; | |||
| 459 | } | |||
| 460 | if (IntrInst->getIntrinsicID() == Intrinsic::lifetime_end) { | |||
| 461 | if (Info.LifeEnd) | |||
| 462 | return {}; | |||
| 463 | Info.LifeEnd = IntrInst; | |||
| 464 | } | |||
| 465 | continue; | |||
| 466 | } | |||
| 467 | // Find untracked uses of the address, bail. | |||
| 468 | if (!definedInRegion(Blocks, U)) | |||
| 469 | return {}; | |||
| 470 | } | |||
| 471 | ||||
| 472 | if (!Info.LifeStart || !Info.LifeEnd) | |||
| 473 | return {}; | |||
| 474 | ||||
| 475 | Info.SinkLifeStart = !definedInRegion(Blocks, Info.LifeStart); | |||
| 476 | Info.HoistLifeEnd = !definedInRegion(Blocks, Info.LifeEnd); | |||
| 477 | // Do legality check. | |||
| 478 | if ((Info.SinkLifeStart || Info.HoistLifeEnd) && | |||
| 479 | !isLegalToShrinkwrapLifetimeMarkers(CEAC, Addr)) | |||
| 480 | return {}; | |||
| 481 | ||||
| 482 | // Check to see if we have a place to do hoisting, if not, bail. | |||
| 483 | if (Info.HoistLifeEnd && !ExitBlock) | |||
| 484 | return {}; | |||
| 485 | ||||
| 486 | return Info; | |||
| 487 | } | |||
| 488 | ||||
| 489 | void CodeExtractor::findAllocas(const CodeExtractorAnalysisCache &CEAC, | |||
| 490 | ValueSet &SinkCands, ValueSet &HoistCands, | |||
| 491 | BasicBlock *&ExitBlock) const { | |||
| 492 | Function *Func = (*Blocks.begin())->getParent(); | |||
| 493 | ExitBlock = getCommonExitBlock(Blocks); | |||
| 494 | ||||
| 495 | auto moveOrIgnoreLifetimeMarkers = | |||
| 496 | [&](const LifetimeMarkerInfo &LMI) -> bool { | |||
| 497 | if (!LMI.LifeStart) | |||
| 498 | return false; | |||
| 499 | if (LMI.SinkLifeStart) { | |||
| 500 | LLVM_DEBUG(dbgs() << "Sinking lifetime.start: " << *LMI.LifeStartdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "Sinking lifetime.start: " << *LMI.LifeStart << "\n"; } } while (false) | |||
| 501 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "Sinking lifetime.start: " << *LMI.LifeStart << "\n"; } } while (false); | |||
| 502 | SinkCands.insert(LMI.LifeStart); | |||
| 503 | } | |||
| 504 | if (LMI.HoistLifeEnd) { | |||
| 505 | LLVM_DEBUG(dbgs() << "Hoisting lifetime.end: " << *LMI.LifeEnd << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "Hoisting lifetime.end: " << *LMI.LifeEnd << "\n"; } } while (false); | |||
| 506 | HoistCands.insert(LMI.LifeEnd); | |||
| 507 | } | |||
| 508 | return true; | |||
| 509 | }; | |||
| 510 | ||||
| 511 | // Look up allocas in the original function in CodeExtractorAnalysisCache, as | |||
| 512 | // this is much faster than walking all the instructions. | |||
| 513 | for (AllocaInst *AI : CEAC.getAllocas()) { | |||
| 514 | BasicBlock *BB = AI->getParent(); | |||
| 515 | if (Blocks.count(BB)) | |||
| 516 | continue; | |||
| 517 | ||||
| 518 | // As a prior call to extractCodeRegion() may have shrinkwrapped the alloca, | |||
| 519 | // check whether it is actually still in the original function. | |||
| 520 | Function *AIFunc = BB->getParent(); | |||
| 521 | if (AIFunc != Func) | |||
| 522 | continue; | |||
| 523 | ||||
| 524 | LifetimeMarkerInfo MarkerInfo = getLifetimeMarkers(CEAC, AI, ExitBlock); | |||
| 525 | bool Moved = moveOrIgnoreLifetimeMarkers(MarkerInfo); | |||
| 526 | if (Moved) { | |||
| 527 | LLVM_DEBUG(dbgs() << "Sinking alloca: " << *AI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "Sinking alloca: " << *AI << "\n"; } } while (false); | |||
| 528 | SinkCands.insert(AI); | |||
| 529 | continue; | |||
| 530 | } | |||
| 531 | ||||
| 532 | // Follow any bitcasts. | |||
| 533 | SmallVector<Instruction *, 2> Bitcasts; | |||
| 534 | SmallVector<LifetimeMarkerInfo, 2> BitcastLifetimeInfo; | |||
| 535 | for (User *U : AI->users()) { | |||
| 536 | if (U->stripInBoundsConstantOffsets() == AI) { | |||
| 537 | Instruction *Bitcast = cast<Instruction>(U); | |||
| 538 | LifetimeMarkerInfo LMI = getLifetimeMarkers(CEAC, Bitcast, ExitBlock); | |||
| 539 | if (LMI.LifeStart) { | |||
| 540 | Bitcasts.push_back(Bitcast); | |||
| 541 | BitcastLifetimeInfo.push_back(LMI); | |||
| 542 | continue; | |||
| 543 | } | |||
| 544 | } | |||
| 545 | ||||
| 546 | // Found unknown use of AI. | |||
| 547 | if (!definedInRegion(Blocks, U)) { | |||
| 548 | Bitcasts.clear(); | |||
| 549 | break; | |||
| 550 | } | |||
| 551 | } | |||
| 552 | ||||
| 553 | // Either no bitcasts reference the alloca or there are unknown uses. | |||
| 554 | if (Bitcasts.empty()) | |||
| 555 | continue; | |||
| 556 | ||||
| 557 | LLVM_DEBUG(dbgs() << "Sinking alloca (via bitcast): " << *AI << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "Sinking alloca (via bitcast): " << *AI << "\n"; } } while (false); | |||
| 558 | SinkCands.insert(AI); | |||
| 559 | for (unsigned I = 0, E = Bitcasts.size(); I != E; ++I) { | |||
| 560 | Instruction *BitcastAddr = Bitcasts[I]; | |||
| 561 | const LifetimeMarkerInfo &LMI = BitcastLifetimeInfo[I]; | |||
| 562 | assert(LMI.LifeStart &&((LMI.LifeStart && "Unsafe to sink bitcast without lifetime markers" ) ? static_cast<void> (0) : __assert_fail ("LMI.LifeStart && \"Unsafe to sink bitcast without lifetime markers\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 563, __PRETTY_FUNCTION__)) | |||
| 563 | "Unsafe to sink bitcast without lifetime markers")((LMI.LifeStart && "Unsafe to sink bitcast without lifetime markers" ) ? static_cast<void> (0) : __assert_fail ("LMI.LifeStart && \"Unsafe to sink bitcast without lifetime markers\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 563, __PRETTY_FUNCTION__)); | |||
| 564 | moveOrIgnoreLifetimeMarkers(LMI); | |||
| 565 | if (!definedInRegion(Blocks, BitcastAddr)) { | |||
| 566 | LLVM_DEBUG(dbgs() << "Sinking bitcast-of-alloca: " << *BitcastAddrdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "Sinking bitcast-of-alloca: " << *BitcastAddr << "\n"; } } while (false) | |||
| 567 | << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "Sinking bitcast-of-alloca: " << *BitcastAddr << "\n"; } } while (false); | |||
| 568 | SinkCands.insert(BitcastAddr); | |||
| 569 | } | |||
| 570 | } | |||
| 571 | } | |||
| 572 | } | |||
| 573 | ||||
| 574 | bool CodeExtractor::isEligible() const { | |||
| 575 | if (Blocks.empty()) | |||
| 576 | return false; | |||
| 577 | BasicBlock *Header = *Blocks.begin(); | |||
| 578 | Function *F = Header->getParent(); | |||
| 579 | ||||
| 580 | // For functions with varargs, check that varargs handling is only done in the | |||
| 581 | // outlined function, i.e vastart and vaend are only used in outlined blocks. | |||
| 582 | if (AllowVarArgs && F->getFunctionType()->isVarArg()) { | |||
| 583 | auto containsVarArgIntrinsic = [](const Instruction &I) { | |||
| 584 | if (const CallInst *CI = dyn_cast<CallInst>(&I)) | |||
| 585 | if (const Function *Callee = CI->getCalledFunction()) | |||
| 586 | return Callee->getIntrinsicID() == Intrinsic::vastart || | |||
| 587 | Callee->getIntrinsicID() == Intrinsic::vaend; | |||
| 588 | return false; | |||
| 589 | }; | |||
| 590 | ||||
| 591 | for (auto &BB : *F) { | |||
| 592 | if (Blocks.count(&BB)) | |||
| 593 | continue; | |||
| 594 | if (llvm::any_of(BB, containsVarArgIntrinsic)) | |||
| 595 | return false; | |||
| 596 | } | |||
| 597 | } | |||
| 598 | return true; | |||
| 599 | } | |||
| 600 | ||||
| 601 | void CodeExtractor::findInputsOutputs(ValueSet &Inputs, ValueSet &Outputs, | |||
| 602 | const ValueSet &SinkCands) const { | |||
| 603 | for (BasicBlock *BB : Blocks) { | |||
| 604 | // If a used value is defined outside the region, it's an input. If an | |||
| 605 | // instruction is used outside the region, it's an output. | |||
| 606 | for (Instruction &II : *BB) { | |||
| 607 | for (auto &OI : II.operands()) { | |||
| 608 | Value *V = OI; | |||
| 609 | if (!SinkCands.count(V) && definedInCaller(Blocks, V)) | |||
| 610 | Inputs.insert(V); | |||
| 611 | } | |||
| 612 | ||||
| 613 | for (User *U : II.users()) | |||
| 614 | if (!definedInRegion(Blocks, U)) { | |||
| 615 | Outputs.insert(&II); | |||
| 616 | break; | |||
| 617 | } | |||
| 618 | } | |||
| 619 | } | |||
| 620 | } | |||
| 621 | ||||
| 622 | /// severSplitPHINodesOfEntry - If a PHI node has multiple inputs from outside | |||
| 623 | /// of the region, we need to split the entry block of the region so that the | |||
| 624 | /// PHI node is easier to deal with. | |||
| 625 | void CodeExtractor::severSplitPHINodesOfEntry(BasicBlock *&Header) { | |||
| 626 | unsigned NumPredsFromRegion = 0; | |||
| 627 | unsigned NumPredsOutsideRegion = 0; | |||
| 628 | ||||
| 629 | if (Header != &Header->getParent()->getEntryBlock()) { | |||
| 630 | PHINode *PN = dyn_cast<PHINode>(Header->begin()); | |||
| 631 | if (!PN) return; // No PHI nodes. | |||
| 632 | ||||
| 633 | // If the header node contains any PHI nodes, check to see if there is more | |||
| 634 | // than one entry from outside the region. If so, we need to sever the | |||
| 635 | // header block into two. | |||
| 636 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) | |||
| 637 | if (Blocks.count(PN->getIncomingBlock(i))) | |||
| 638 | ++NumPredsFromRegion; | |||
| 639 | else | |||
| 640 | ++NumPredsOutsideRegion; | |||
| 641 | ||||
| 642 | // If there is one (or fewer) predecessor from outside the region, we don't | |||
| 643 | // need to do anything special. | |||
| 644 | if (NumPredsOutsideRegion <= 1) return; | |||
| 645 | } | |||
| 646 | ||||
| 647 | // Otherwise, we need to split the header block into two pieces: one | |||
| 648 | // containing PHI nodes merging values from outside of the region, and a | |||
| 649 | // second that contains all of the code for the block and merges back any | |||
| 650 | // incoming values from inside of the region. | |||
| 651 | BasicBlock *NewBB = SplitBlock(Header, Header->getFirstNonPHI(), DT); | |||
| 652 | ||||
| 653 | // We only want to code extract the second block now, and it becomes the new | |||
| 654 | // header of the region. | |||
| 655 | BasicBlock *OldPred = Header; | |||
| 656 | Blocks.remove(OldPred); | |||
| 657 | Blocks.insert(NewBB); | |||
| 658 | Header = NewBB; | |||
| 659 | ||||
| 660 | // Okay, now we need to adjust the PHI nodes and any branches from within the | |||
| 661 | // region to go to the new header block instead of the old header block. | |||
| 662 | if (NumPredsFromRegion) { | |||
| 663 | PHINode *PN = cast<PHINode>(OldPred->begin()); | |||
| 664 | // Loop over all of the predecessors of OldPred that are in the region, | |||
| 665 | // changing them to branch to NewBB instead. | |||
| 666 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) | |||
| 667 | if (Blocks.count(PN->getIncomingBlock(i))) { | |||
| 668 | Instruction *TI = PN->getIncomingBlock(i)->getTerminator(); | |||
| 669 | TI->replaceUsesOfWith(OldPred, NewBB); | |||
| 670 | } | |||
| 671 | ||||
| 672 | // Okay, everything within the region is now branching to the right block, we | |||
| 673 | // just have to update the PHI nodes now, inserting PHI nodes into NewBB. | |||
| 674 | BasicBlock::iterator AfterPHIs; | |||
| 675 | for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) { | |||
| 676 | PHINode *PN = cast<PHINode>(AfterPHIs); | |||
| 677 | // Create a new PHI node in the new region, which has an incoming value | |||
| 678 | // from OldPred of PN. | |||
| 679 | PHINode *NewPN = PHINode::Create(PN->getType(), 1 + NumPredsFromRegion, | |||
| 680 | PN->getName() + ".ce", &NewBB->front()); | |||
| 681 | PN->replaceAllUsesWith(NewPN); | |||
| 682 | NewPN->addIncoming(PN, OldPred); | |||
| 683 | ||||
| 684 | // Loop over all of the incoming value in PN, moving them to NewPN if they | |||
| 685 | // are from the extracted region. | |||
| 686 | for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) { | |||
| 687 | if (Blocks.count(PN->getIncomingBlock(i))) { | |||
| 688 | NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i)); | |||
| 689 | PN->removeIncomingValue(i); | |||
| 690 | --i; | |||
| 691 | } | |||
| 692 | } | |||
| 693 | } | |||
| 694 | } | |||
| 695 | } | |||
| 696 | ||||
| 697 | /// severSplitPHINodesOfExits - if PHI nodes in exit blocks have inputs from | |||
| 698 | /// outlined region, we split these PHIs on two: one with inputs from region | |||
| 699 | /// and other with remaining incoming blocks; then first PHIs are placed in | |||
| 700 | /// outlined region. | |||
| 701 | void CodeExtractor::severSplitPHINodesOfExits( | |||
| 702 | const SmallPtrSetImpl<BasicBlock *> &Exits) { | |||
| 703 | for (BasicBlock *ExitBB : Exits) { | |||
| 704 | BasicBlock *NewBB = nullptr; | |||
| 705 | ||||
| 706 | for (PHINode &PN : ExitBB->phis()) { | |||
| 707 | // Find all incoming values from the outlining region. | |||
| 708 | SmallVector<unsigned, 2> IncomingVals; | |||
| 709 | for (unsigned i = 0; i < PN.getNumIncomingValues(); ++i) | |||
| 710 | if (Blocks.count(PN.getIncomingBlock(i))) | |||
| 711 | IncomingVals.push_back(i); | |||
| 712 | ||||
| 713 | // Do not process PHI if there is one (or fewer) predecessor from region. | |||
| 714 | // If PHI has exactly one predecessor from region, only this one incoming | |||
| 715 | // will be replaced on codeRepl block, so it should be safe to skip PHI. | |||
| 716 | if (IncomingVals.size() <= 1) | |||
| 717 | continue; | |||
| 718 | ||||
| 719 | // Create block for new PHIs and add it to the list of outlined if it | |||
| 720 | // wasn't done before. | |||
| 721 | if (!NewBB) { | |||
| 722 | NewBB = BasicBlock::Create(ExitBB->getContext(), | |||
| 723 | ExitBB->getName() + ".split", | |||
| 724 | ExitBB->getParent(), ExitBB); | |||
| 725 | SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBB), | |||
| 726 | pred_end(ExitBB)); | |||
| 727 | for (BasicBlock *PredBB : Preds) | |||
| 728 | if (Blocks.count(PredBB)) | |||
| 729 | PredBB->getTerminator()->replaceUsesOfWith(ExitBB, NewBB); | |||
| 730 | BranchInst::Create(ExitBB, NewBB); | |||
| 731 | Blocks.insert(NewBB); | |||
| 732 | } | |||
| 733 | ||||
| 734 | // Split this PHI. | |||
| 735 | PHINode *NewPN = | |||
| 736 | PHINode::Create(PN.getType(), IncomingVals.size(), | |||
| 737 | PN.getName() + ".ce", NewBB->getFirstNonPHI()); | |||
| 738 | for (unsigned i : IncomingVals) | |||
| 739 | NewPN->addIncoming(PN.getIncomingValue(i), PN.getIncomingBlock(i)); | |||
| 740 | for (unsigned i : reverse(IncomingVals)) | |||
| 741 | PN.removeIncomingValue(i, false); | |||
| 742 | PN.addIncoming(NewPN, NewBB); | |||
| 743 | } | |||
| 744 | } | |||
| 745 | } | |||
| 746 | ||||
| 747 | void CodeExtractor::splitReturnBlocks() { | |||
| 748 | for (BasicBlock *Block : Blocks) | |||
| 749 | if (ReturnInst *RI = dyn_cast<ReturnInst>(Block->getTerminator())) { | |||
| 750 | BasicBlock *New = | |||
| 751 | Block->splitBasicBlock(RI->getIterator(), Block->getName() + ".ret"); | |||
| 752 | if (DT) { | |||
| 753 | // Old dominates New. New node dominates all other nodes dominated | |||
| 754 | // by Old. | |||
| 755 | DomTreeNode *OldNode = DT->getNode(Block); | |||
| 756 | SmallVector<DomTreeNode *, 8> Children(OldNode->begin(), | |||
| 757 | OldNode->end()); | |||
| 758 | ||||
| 759 | DomTreeNode *NewNode = DT->addNewBlock(New, Block); | |||
| 760 | ||||
| 761 | for (DomTreeNode *I : Children) | |||
| 762 | DT->changeImmediateDominator(I, NewNode); | |||
| 763 | } | |||
| 764 | } | |||
| 765 | } | |||
| 766 | ||||
| 767 | /// constructFunction - make a function based on inputs and outputs, as follows: | |||
| 768 | /// f(in0, ..., inN, out0, ..., outN) | |||
| 769 | Function *CodeExtractor::constructFunction(const ValueSet &inputs, | |||
| 770 | const ValueSet &outputs, | |||
| 771 | BasicBlock *header, | |||
| 772 | BasicBlock *newRootNode, | |||
| 773 | BasicBlock *newHeader, | |||
| 774 | Function *oldFunction, | |||
| 775 | Module *M) { | |||
| 776 | LLVM_DEBUG(dbgs() << "inputs: " << inputs.size() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "inputs: " << inputs .size() << "\n"; } } while (false); | |||
| ||||
| 777 | LLVM_DEBUG(dbgs() << "outputs: " << outputs.size() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "outputs: " << outputs .size() << "\n"; } } while (false); | |||
| 778 | ||||
| 779 | // This function returns unsigned, outputs will go back by reference. | |||
| 780 | switch (NumExitBlocks) { | |||
| 781 | case 0: | |||
| 782 | case 1: RetTy = Type::getVoidTy(header->getContext()); break; | |||
| 783 | case 2: RetTy = Type::getInt1Ty(header->getContext()); break; | |||
| 784 | default: RetTy = Type::getInt16Ty(header->getContext()); break; | |||
| 785 | } | |||
| 786 | ||||
| 787 | std::vector<Type *> paramTy; | |||
| 788 | ||||
| 789 | // Add the types of the input values to the function's argument list | |||
| 790 | for (Value *value : inputs) { | |||
| 791 | LLVM_DEBUG(dbgs() << "value used in func: " << *value << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "value used in func: " << *value << "\n"; } } while (false); | |||
| 792 | paramTy.push_back(value->getType()); | |||
| 793 | } | |||
| 794 | ||||
| 795 | // Add the types of the output values to the function's argument list. | |||
| 796 | for (Value *output : outputs) { | |||
| 797 | LLVM_DEBUG(dbgs() << "instr used in func: " << *output << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { dbgs() << "instr used in func: " << *output << "\n"; } } while (false); | |||
| 798 | if (AggregateArgs) | |||
| 799 | paramTy.push_back(output->getType()); | |||
| 800 | else | |||
| 801 | paramTy.push_back(PointerType::getUnqual(output->getType())); | |||
| 802 | } | |||
| 803 | ||||
| 804 | LLVM_DEBUG({do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { { dbgs() << "Function type: " << *RetTy << " f("; for (Type *i : paramTy) dbgs() << *i << ", "; dbgs() << ")\n"; }; } } while (false ) | |||
| 805 | dbgs() << "Function type: " << *RetTy << " f(";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { { dbgs() << "Function type: " << *RetTy << " f("; for (Type *i : paramTy) dbgs() << *i << ", "; dbgs() << ")\n"; }; } } while (false ) | |||
| 806 | for (Type *i : paramTy)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { { dbgs() << "Function type: " << *RetTy << " f("; for (Type *i : paramTy) dbgs() << *i << ", "; dbgs() << ")\n"; }; } } while (false ) | |||
| 807 | dbgs() << *i << ", ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { { dbgs() << "Function type: " << *RetTy << " f("; for (Type *i : paramTy) dbgs() << *i << ", "; dbgs() << ")\n"; }; } } while (false ) | |||
| 808 | dbgs() << ")\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { { dbgs() << "Function type: " << *RetTy << " f("; for (Type *i : paramTy) dbgs() << *i << ", "; dbgs() << ")\n"; }; } } while (false ) | |||
| 809 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { { dbgs() << "Function type: " << *RetTy << " f("; for (Type *i : paramTy) dbgs() << *i << ", "; dbgs() << ")\n"; }; } } while (false ); | |||
| 810 | ||||
| 811 | StructType *StructTy = nullptr; | |||
| 812 | if (AggregateArgs && (inputs.size() + outputs.size() > 0)) { | |||
| 813 | StructTy = StructType::get(M->getContext(), paramTy); | |||
| 814 | paramTy.clear(); | |||
| 815 | paramTy.push_back(PointerType::getUnqual(StructTy)); | |||
| 816 | } | |||
| 817 | FunctionType *funcType = | |||
| 818 | FunctionType::get(RetTy, paramTy, | |||
| 819 | AllowVarArgs && oldFunction->isVarArg()); | |||
| 820 | ||||
| 821 | std::string SuffixToUse = | |||
| 822 | Suffix.empty() | |||
| 823 | ? (header->getName().empty() ? "extracted" : header->getName().str()) | |||
| 824 | : Suffix; | |||
| 825 | // Create the new function | |||
| 826 | Function *newFunction = Function::Create( | |||
| 827 | funcType, GlobalValue::InternalLinkage, oldFunction->getAddressSpace(), | |||
| 828 | oldFunction->getName() + "." + SuffixToUse, M); | |||
| 829 | // If the old function is no-throw, so is the new one. | |||
| 830 | if (oldFunction->doesNotThrow()) | |||
| 831 | newFunction->setDoesNotThrow(); | |||
| 832 | ||||
| 833 | // Inherit the uwtable attribute if we need to. | |||
| 834 | if (oldFunction->hasUWTable()) | |||
| 835 | newFunction->setHasUWTable(); | |||
| 836 | ||||
| 837 | // Inherit all of the target dependent attributes and white-listed | |||
| 838 | // target independent attributes. | |||
| 839 | // (e.g. If the extracted region contains a call to an x86.sse | |||
| 840 | // instruction we need to make sure that the extracted region has the | |||
| 841 | // "target-features" attribute allowing it to be lowered. | |||
| 842 | // FIXME: This should be changed to check to see if a specific | |||
| 843 | // attribute can not be inherited. | |||
| 844 | for (const auto &Attr : oldFunction->getAttributes().getFnAttributes()) { | |||
| 845 | if (Attr.isStringAttribute()) { | |||
| 846 | if (Attr.getKindAsString() == "thunk") | |||
| 847 | continue; | |||
| 848 | } else | |||
| 849 | switch (Attr.getKindAsEnum()) { | |||
| 850 | // Those attributes cannot be propagated safely. Explicitly list them | |||
| 851 | // here so we get a warning if new attributes are added. This list also | |||
| 852 | // includes non-function attributes. | |||
| 853 | case Attribute::Alignment: | |||
| 854 | case Attribute::AllocSize: | |||
| 855 | case Attribute::ArgMemOnly: | |||
| 856 | case Attribute::Builtin: | |||
| 857 | case Attribute::ByVal: | |||
| 858 | case Attribute::Convergent: | |||
| 859 | case Attribute::Dereferenceable: | |||
| 860 | case Attribute::DereferenceableOrNull: | |||
| 861 | case Attribute::InAlloca: | |||
| 862 | case Attribute::InReg: | |||
| 863 | case Attribute::InaccessibleMemOnly: | |||
| 864 | case Attribute::InaccessibleMemOrArgMemOnly: | |||
| 865 | case Attribute::JumpTable: | |||
| 866 | case Attribute::Naked: | |||
| 867 | case Attribute::Nest: | |||
| 868 | case Attribute::NoAlias: | |||
| 869 | case Attribute::NoBuiltin: | |||
| 870 | case Attribute::NoCapture: | |||
| 871 | case Attribute::NoReturn: | |||
| 872 | case Attribute::NoSync: | |||
| 873 | case Attribute::None: | |||
| 874 | case Attribute::NonNull: | |||
| 875 | case Attribute::ReadNone: | |||
| 876 | case Attribute::ReadOnly: | |||
| 877 | case Attribute::Returned: | |||
| 878 | case Attribute::ReturnsTwice: | |||
| 879 | case Attribute::SExt: | |||
| 880 | case Attribute::Speculatable: | |||
| 881 | case Attribute::StackAlignment: | |||
| 882 | case Attribute::StructRet: | |||
| 883 | case Attribute::SwiftError: | |||
| 884 | case Attribute::SwiftSelf: | |||
| 885 | case Attribute::WillReturn: | |||
| 886 | case Attribute::WriteOnly: | |||
| 887 | case Attribute::ZExt: | |||
| 888 | case Attribute::ImmArg: | |||
| 889 | case Attribute::EndAttrKinds: | |||
| 890 | case Attribute::EmptyKey: | |||
| 891 | case Attribute::TombstoneKey: | |||
| 892 | continue; | |||
| 893 | // Those attributes should be safe to propagate to the extracted function. | |||
| 894 | case Attribute::AlwaysInline: | |||
| 895 | case Attribute::Cold: | |||
| 896 | case Attribute::NoRecurse: | |||
| 897 | case Attribute::InlineHint: | |||
| 898 | case Attribute::MinSize: | |||
| 899 | case Attribute::NoDuplicate: | |||
| 900 | case Attribute::NoFree: | |||
| 901 | case Attribute::NoImplicitFloat: | |||
| 902 | case Attribute::NoInline: | |||
| 903 | case Attribute::NonLazyBind: | |||
| 904 | case Attribute::NoRedZone: | |||
| 905 | case Attribute::NoUnwind: | |||
| 906 | case Attribute::OptForFuzzing: | |||
| 907 | case Attribute::OptimizeNone: | |||
| 908 | case Attribute::OptimizeForSize: | |||
| 909 | case Attribute::SafeStack: | |||
| 910 | case Attribute::ShadowCallStack: | |||
| 911 | case Attribute::SanitizeAddress: | |||
| 912 | case Attribute::SanitizeMemory: | |||
| 913 | case Attribute::SanitizeThread: | |||
| 914 | case Attribute::SanitizeHWAddress: | |||
| 915 | case Attribute::SanitizeMemTag: | |||
| 916 | case Attribute::SpeculativeLoadHardening: | |||
| 917 | case Attribute::StackProtect: | |||
| 918 | case Attribute::StackProtectReq: | |||
| 919 | case Attribute::StackProtectStrong: | |||
| 920 | case Attribute::StrictFP: | |||
| 921 | case Attribute::UWTable: | |||
| 922 | case Attribute::NoCfCheck: | |||
| 923 | break; | |||
| 924 | } | |||
| 925 | ||||
| 926 | newFunction->addFnAttr(Attr); | |||
| 927 | } | |||
| 928 | newFunction->getBasicBlockList().push_back(newRootNode); | |||
| 929 | ||||
| 930 | // Create an iterator to name all of the arguments we inserted. | |||
| 931 | Function::arg_iterator AI = newFunction->arg_begin(); | |||
| 932 | ||||
| 933 | // Rewrite all users of the inputs in the extracted region to use the | |||
| 934 | // arguments (or appropriate addressing into struct) instead. | |||
| 935 | for (unsigned i = 0, e = inputs.size(); i != e; ++i) { | |||
| 936 | Value *RewriteVal; | |||
| 937 | if (AggregateArgs
| |||
| 938 | Value *Idx[2]; | |||
| 939 | Idx[0] = Constant::getNullValue(Type::getInt32Ty(header->getContext())); | |||
| 940 | Idx[1] = ConstantInt::get(Type::getInt32Ty(header->getContext()), i); | |||
| 941 | Instruction *TI = newFunction->begin()->getTerminator(); | |||
| 942 | GetElementPtrInst *GEP = GetElementPtrInst::Create( | |||
| 943 | StructTy, &*AI, Idx, "gep_" + inputs[i]->getName(), TI); | |||
| 944 | RewriteVal = new LoadInst(StructTy->getElementType(i), GEP, | |||
| ||||
| 945 | "loadgep_" + inputs[i]->getName(), TI); | |||
| 946 | } else | |||
| 947 | RewriteVal = &*AI++; | |||
| 948 | ||||
| 949 | std::vector<User *> Users(inputs[i]->user_begin(), inputs[i]->user_end()); | |||
| 950 | for (User *use : Users) | |||
| 951 | if (Instruction *inst = dyn_cast<Instruction>(use)) | |||
| 952 | if (Blocks.count(inst->getParent())) | |||
| 953 | inst->replaceUsesOfWith(inputs[i], RewriteVal); | |||
| 954 | } | |||
| 955 | ||||
| 956 | // Set names for input and output arguments. | |||
| 957 | if (!AggregateArgs) { | |||
| 958 | AI = newFunction->arg_begin(); | |||
| 959 | for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI) | |||
| 960 | AI->setName(inputs[i]->getName()); | |||
| 961 | for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI) | |||
| 962 | AI->setName(outputs[i]->getName()+".out"); | |||
| 963 | } | |||
| 964 | ||||
| 965 | // Rewrite branches to basic blocks outside of the loop to new dummy blocks | |||
| 966 | // within the new function. This must be done before we lose track of which | |||
| 967 | // blocks were originally in the code region. | |||
| 968 | std::vector<User *> Users(header->user_begin(), header->user_end()); | |||
| 969 | for (auto &U : Users) | |||
| 970 | // The BasicBlock which contains the branch is not in the region | |||
| 971 | // modify the branch target to a new block | |||
| 972 | if (Instruction *I = dyn_cast<Instruction>(U)) | |||
| 973 | if (I->isTerminator() && I->getFunction() == oldFunction && | |||
| 974 | !Blocks.count(I->getParent())) | |||
| 975 | I->replaceUsesOfWith(header, newHeader); | |||
| 976 | ||||
| 977 | return newFunction; | |||
| 978 | } | |||
| 979 | ||||
| 980 | /// Erase lifetime.start markers which reference inputs to the extraction | |||
| 981 | /// region, and insert the referenced memory into \p LifetimesStart. | |||
| 982 | /// | |||
| 983 | /// The extraction region is defined by a set of blocks (\p Blocks), and a set | |||
| 984 | /// of allocas which will be moved from the caller function into the extracted | |||
| 985 | /// function (\p SunkAllocas). | |||
| 986 | static void eraseLifetimeMarkersOnInputs(const SetVector<BasicBlock *> &Blocks, | |||
| 987 | const SetVector<Value *> &SunkAllocas, | |||
| 988 | SetVector<Value *> &LifetimesStart) { | |||
| 989 | for (BasicBlock *BB : Blocks) { | |||
| 990 | for (auto It = BB->begin(), End = BB->end(); It != End;) { | |||
| 991 | auto *II = dyn_cast<IntrinsicInst>(&*It); | |||
| 992 | ++It; | |||
| 993 | if (!II || !II->isLifetimeStartOrEnd()) | |||
| 994 | continue; | |||
| 995 | ||||
| 996 | // Get the memory operand of the lifetime marker. If the underlying | |||
| 997 | // object is a sunk alloca, or is otherwise defined in the extraction | |||
| 998 | // region, the lifetime marker must not be erased. | |||
| 999 | Value *Mem = II->getOperand(1)->stripInBoundsOffsets(); | |||
| 1000 | if (SunkAllocas.count(Mem) || definedInRegion(Blocks, Mem)) | |||
| 1001 | continue; | |||
| 1002 | ||||
| 1003 | if (II->getIntrinsicID() == Intrinsic::lifetime_start) | |||
| 1004 | LifetimesStart.insert(Mem); | |||
| 1005 | II->eraseFromParent(); | |||
| 1006 | } | |||
| 1007 | } | |||
| 1008 | } | |||
| 1009 | ||||
| 1010 | /// Insert lifetime start/end markers surrounding the call to the new function | |||
| 1011 | /// for objects defined in the caller. | |||
| 1012 | static void insertLifetimeMarkersSurroundingCall( | |||
| 1013 | Module *M, ArrayRef<Value *> LifetimesStart, ArrayRef<Value *> LifetimesEnd, | |||
| 1014 | CallInst *TheCall) { | |||
| 1015 | LLVMContext &Ctx = M->getContext(); | |||
| 1016 | auto Int8PtrTy = Type::getInt8PtrTy(Ctx); | |||
| 1017 | auto NegativeOne = ConstantInt::getSigned(Type::getInt64Ty(Ctx), -1); | |||
| 1018 | Instruction *Term = TheCall->getParent()->getTerminator(); | |||
| 1019 | ||||
| 1020 | // The memory argument to a lifetime marker must be a i8*. Cache any bitcasts | |||
| 1021 | // needed to satisfy this requirement so they may be reused. | |||
| 1022 | DenseMap<Value *, Value *> Bitcasts; | |||
| 1023 | ||||
| 1024 | // Emit lifetime markers for the pointers given in \p Objects. Insert the | |||
| 1025 | // markers before the call if \p InsertBefore, and after the call otherwise. | |||
| 1026 | auto insertMarkers = [&](Function *MarkerFunc, ArrayRef<Value *> Objects, | |||
| 1027 | bool InsertBefore) { | |||
| 1028 | for (Value *Mem : Objects) { | |||
| 1029 | assert((!isa<Instruction>(Mem) || cast<Instruction>(Mem)->getFunction() ==(((!isa<Instruction>(Mem) || cast<Instruction>(Mem )->getFunction() == TheCall->getFunction()) && "Input memory not defined in original function" ) ? static_cast<void> (0) : __assert_fail ("(!isa<Instruction>(Mem) || cast<Instruction>(Mem)->getFunction() == TheCall->getFunction()) && \"Input memory not defined in original function\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1031, __PRETTY_FUNCTION__)) | |||
| 1030 | TheCall->getFunction()) &&(((!isa<Instruction>(Mem) || cast<Instruction>(Mem )->getFunction() == TheCall->getFunction()) && "Input memory not defined in original function" ) ? static_cast<void> (0) : __assert_fail ("(!isa<Instruction>(Mem) || cast<Instruction>(Mem)->getFunction() == TheCall->getFunction()) && \"Input memory not defined in original function\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1031, __PRETTY_FUNCTION__)) | |||
| 1031 | "Input memory not defined in original function")(((!isa<Instruction>(Mem) || cast<Instruction>(Mem )->getFunction() == TheCall->getFunction()) && "Input memory not defined in original function" ) ? static_cast<void> (0) : __assert_fail ("(!isa<Instruction>(Mem) || cast<Instruction>(Mem)->getFunction() == TheCall->getFunction()) && \"Input memory not defined in original function\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1031, __PRETTY_FUNCTION__)); | |||
| 1032 | Value *&MemAsI8Ptr = Bitcasts[Mem]; | |||
| 1033 | if (!MemAsI8Ptr) { | |||
| 1034 | if (Mem->getType() == Int8PtrTy) | |||
| 1035 | MemAsI8Ptr = Mem; | |||
| 1036 | else | |||
| 1037 | MemAsI8Ptr = | |||
| 1038 | CastInst::CreatePointerCast(Mem, Int8PtrTy, "lt.cast", TheCall); | |||
| 1039 | } | |||
| 1040 | ||||
| 1041 | auto Marker = CallInst::Create(MarkerFunc, {NegativeOne, MemAsI8Ptr}); | |||
| 1042 | if (InsertBefore) | |||
| 1043 | Marker->insertBefore(TheCall); | |||
| 1044 | else | |||
| 1045 | Marker->insertBefore(Term); | |||
| 1046 | } | |||
| 1047 | }; | |||
| 1048 | ||||
| 1049 | if (!LifetimesStart.empty()) { | |||
| 1050 | auto StartFn = llvm::Intrinsic::getDeclaration( | |||
| 1051 | M, llvm::Intrinsic::lifetime_start, Int8PtrTy); | |||
| 1052 | insertMarkers(StartFn, LifetimesStart, /*InsertBefore=*/true); | |||
| 1053 | } | |||
| 1054 | ||||
| 1055 | if (!LifetimesEnd.empty()) { | |||
| 1056 | auto EndFn = llvm::Intrinsic::getDeclaration( | |||
| 1057 | M, llvm::Intrinsic::lifetime_end, Int8PtrTy); | |||
| 1058 | insertMarkers(EndFn, LifetimesEnd, /*InsertBefore=*/false); | |||
| 1059 | } | |||
| 1060 | } | |||
| 1061 | ||||
| 1062 | /// emitCallAndSwitchStatement - This method sets up the caller side by adding | |||
| 1063 | /// the call instruction, splitting any PHI nodes in the header block as | |||
| 1064 | /// necessary. | |||
| 1065 | CallInst *CodeExtractor::emitCallAndSwitchStatement(Function *newFunction, | |||
| 1066 | BasicBlock *codeReplacer, | |||
| 1067 | ValueSet &inputs, | |||
| 1068 | ValueSet &outputs) { | |||
| 1069 | // Emit a call to the new function, passing in: *pointer to struct (if | |||
| 1070 | // aggregating parameters), or plan inputs and allocated memory for outputs | |||
| 1071 | std::vector<Value *> params, StructValues, ReloadOutputs, Reloads; | |||
| 1072 | ||||
| 1073 | Module *M = newFunction->getParent(); | |||
| 1074 | LLVMContext &Context = M->getContext(); | |||
| 1075 | const DataLayout &DL = M->getDataLayout(); | |||
| 1076 | CallInst *call = nullptr; | |||
| 1077 | ||||
| 1078 | // Add inputs as params, or to be filled into the struct | |||
| 1079 | unsigned ArgNo = 0; | |||
| 1080 | SmallVector<unsigned, 1> SwiftErrorArgs; | |||
| 1081 | for (Value *input : inputs) { | |||
| 1082 | if (AggregateArgs) | |||
| 1083 | StructValues.push_back(input); | |||
| 1084 | else { | |||
| 1085 | params.push_back(input); | |||
| 1086 | if (input->isSwiftError()) | |||
| 1087 | SwiftErrorArgs.push_back(ArgNo); | |||
| 1088 | } | |||
| 1089 | ++ArgNo; | |||
| 1090 | } | |||
| 1091 | ||||
| 1092 | // Create allocas for the outputs | |||
| 1093 | for (Value *output : outputs) { | |||
| 1094 | if (AggregateArgs) { | |||
| 1095 | StructValues.push_back(output); | |||
| 1096 | } else { | |||
| 1097 | AllocaInst *alloca = | |||
| 1098 | new AllocaInst(output->getType(), DL.getAllocaAddrSpace(), | |||
| 1099 | nullptr, output->getName() + ".loc", | |||
| 1100 | &codeReplacer->getParent()->front().front()); | |||
| 1101 | ReloadOutputs.push_back(alloca); | |||
| 1102 | params.push_back(alloca); | |||
| 1103 | } | |||
| 1104 | } | |||
| 1105 | ||||
| 1106 | StructType *StructArgTy = nullptr; | |||
| 1107 | AllocaInst *Struct = nullptr; | |||
| 1108 | if (AggregateArgs && (inputs.size() + outputs.size() > 0)) { | |||
| 1109 | std::vector<Type *> ArgTypes; | |||
| 1110 | for (ValueSet::iterator v = StructValues.begin(), | |||
| 1111 | ve = StructValues.end(); v != ve; ++v) | |||
| 1112 | ArgTypes.push_back((*v)->getType()); | |||
| 1113 | ||||
| 1114 | // Allocate a struct at the beginning of this function | |||
| 1115 | StructArgTy = StructType::get(newFunction->getContext(), ArgTypes); | |||
| 1116 | Struct = new AllocaInst(StructArgTy, DL.getAllocaAddrSpace(), nullptr, | |||
| 1117 | "structArg", | |||
| 1118 | &codeReplacer->getParent()->front().front()); | |||
| 1119 | params.push_back(Struct); | |||
| 1120 | ||||
| 1121 | for (unsigned i = 0, e = inputs.size(); i != e; ++i) { | |||
| 1122 | Value *Idx[2]; | |||
| 1123 | Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); | |||
| 1124 | Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), i); | |||
| 1125 | GetElementPtrInst *GEP = GetElementPtrInst::Create( | |||
| 1126 | StructArgTy, Struct, Idx, "gep_" + StructValues[i]->getName()); | |||
| 1127 | codeReplacer->getInstList().push_back(GEP); | |||
| 1128 | StoreInst *SI = new StoreInst(StructValues[i], GEP); | |||
| 1129 | codeReplacer->getInstList().push_back(SI); | |||
| 1130 | } | |||
| 1131 | } | |||
| 1132 | ||||
| 1133 | // Emit the call to the function | |||
| 1134 | call = CallInst::Create(newFunction, params, | |||
| 1135 | NumExitBlocks > 1 ? "targetBlock" : ""); | |||
| 1136 | // Add debug location to the new call, if the original function has debug | |||
| 1137 | // info. In that case, the terminator of the entry block of the extracted | |||
| 1138 | // function contains the first debug location of the extracted function, | |||
| 1139 | // set in extractCodeRegion. | |||
| 1140 | if (codeReplacer->getParent()->getSubprogram()) { | |||
| 1141 | if (auto DL = newFunction->getEntryBlock().getTerminator()->getDebugLoc()) | |||
| 1142 | call->setDebugLoc(DL); | |||
| 1143 | } | |||
| 1144 | codeReplacer->getInstList().push_back(call); | |||
| 1145 | ||||
| 1146 | // Set swifterror parameter attributes. | |||
| 1147 | for (unsigned SwiftErrArgNo : SwiftErrorArgs) { | |||
| 1148 | call->addParamAttr(SwiftErrArgNo, Attribute::SwiftError); | |||
| 1149 | newFunction->addParamAttr(SwiftErrArgNo, Attribute::SwiftError); | |||
| 1150 | } | |||
| 1151 | ||||
| 1152 | Function::arg_iterator OutputArgBegin = newFunction->arg_begin(); | |||
| 1153 | unsigned FirstOut = inputs.size(); | |||
| 1154 | if (!AggregateArgs) | |||
| 1155 | std::advance(OutputArgBegin, inputs.size()); | |||
| 1156 | ||||
| 1157 | // Reload the outputs passed in by reference. | |||
| 1158 | for (unsigned i = 0, e = outputs.size(); i != e; ++i) { | |||
| 1159 | Value *Output = nullptr; | |||
| 1160 | if (AggregateArgs) { | |||
| 1161 | Value *Idx[2]; | |||
| 1162 | Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); | |||
| 1163 | Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i); | |||
| 1164 | GetElementPtrInst *GEP = GetElementPtrInst::Create( | |||
| 1165 | StructArgTy, Struct, Idx, "gep_reload_" + outputs[i]->getName()); | |||
| 1166 | codeReplacer->getInstList().push_back(GEP); | |||
| 1167 | Output = GEP; | |||
| 1168 | } else { | |||
| 1169 | Output = ReloadOutputs[i]; | |||
| 1170 | } | |||
| 1171 | LoadInst *load = new LoadInst(outputs[i]->getType(), Output, | |||
| 1172 | outputs[i]->getName() + ".reload"); | |||
| 1173 | Reloads.push_back(load); | |||
| 1174 | codeReplacer->getInstList().push_back(load); | |||
| 1175 | std::vector<User *> Users(outputs[i]->user_begin(), outputs[i]->user_end()); | |||
| 1176 | for (unsigned u = 0, e = Users.size(); u != e; ++u) { | |||
| 1177 | Instruction *inst = cast<Instruction>(Users[u]); | |||
| 1178 | if (!Blocks.count(inst->getParent())) | |||
| 1179 | inst->replaceUsesOfWith(outputs[i], load); | |||
| 1180 | } | |||
| 1181 | } | |||
| 1182 | ||||
| 1183 | // Now we can emit a switch statement using the call as a value. | |||
| 1184 | SwitchInst *TheSwitch = | |||
| 1185 | SwitchInst::Create(Constant::getNullValue(Type::getInt16Ty(Context)), | |||
| 1186 | codeReplacer, 0, codeReplacer); | |||
| 1187 | ||||
| 1188 | // Since there may be multiple exits from the original region, make the new | |||
| 1189 | // function return an unsigned, switch on that number. This loop iterates | |||
| 1190 | // over all of the blocks in the extracted region, updating any terminator | |||
| 1191 | // instructions in the to-be-extracted region that branch to blocks that are | |||
| 1192 | // not in the region to be extracted. | |||
| 1193 | std::map<BasicBlock *, BasicBlock *> ExitBlockMap; | |||
| 1194 | ||||
| 1195 | unsigned switchVal = 0; | |||
| 1196 | for (BasicBlock *Block : Blocks) { | |||
| 1197 | Instruction *TI = Block->getTerminator(); | |||
| 1198 | for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) | |||
| 1199 | if (!Blocks.count(TI->getSuccessor(i))) { | |||
| 1200 | BasicBlock *OldTarget = TI->getSuccessor(i); | |||
| 1201 | // add a new basic block which returns the appropriate value | |||
| 1202 | BasicBlock *&NewTarget = ExitBlockMap[OldTarget]; | |||
| 1203 | if (!NewTarget) { | |||
| 1204 | // If we don't already have an exit stub for this non-extracted | |||
| 1205 | // destination, create one now! | |||
| 1206 | NewTarget = BasicBlock::Create(Context, | |||
| 1207 | OldTarget->getName() + ".exitStub", | |||
| 1208 | newFunction); | |||
| 1209 | unsigned SuccNum = switchVal++; | |||
| 1210 | ||||
| 1211 | Value *brVal = nullptr; | |||
| 1212 | switch (NumExitBlocks) { | |||
| 1213 | case 0: | |||
| 1214 | case 1: break; // No value needed. | |||
| 1215 | case 2: // Conditional branch, return a bool | |||
| 1216 | brVal = ConstantInt::get(Type::getInt1Ty(Context), !SuccNum); | |||
| 1217 | break; | |||
| 1218 | default: | |||
| 1219 | brVal = ConstantInt::get(Type::getInt16Ty(Context), SuccNum); | |||
| 1220 | break; | |||
| 1221 | } | |||
| 1222 | ||||
| 1223 | ReturnInst::Create(Context, brVal, NewTarget); | |||
| 1224 | ||||
| 1225 | // Update the switch instruction. | |||
| 1226 | TheSwitch->addCase(ConstantInt::get(Type::getInt16Ty(Context), | |||
| 1227 | SuccNum), | |||
| 1228 | OldTarget); | |||
| 1229 | } | |||
| 1230 | ||||
| 1231 | // rewrite the original branch instruction with this new target | |||
| 1232 | TI->setSuccessor(i, NewTarget); | |||
| 1233 | } | |||
| 1234 | } | |||
| 1235 | ||||
| 1236 | // Store the arguments right after the definition of output value. | |||
| 1237 | // This should be proceeded after creating exit stubs to be ensure that invoke | |||
| 1238 | // result restore will be placed in the outlined function. | |||
| 1239 | Function::arg_iterator OAI = OutputArgBegin; | |||
| 1240 | for (unsigned i = 0, e = outputs.size(); i != e; ++i) { | |||
| 1241 | auto *OutI = dyn_cast<Instruction>(outputs[i]); | |||
| 1242 | if (!OutI) | |||
| 1243 | continue; | |||
| 1244 | ||||
| 1245 | // Find proper insertion point. | |||
| 1246 | BasicBlock::iterator InsertPt; | |||
| 1247 | // In case OutI is an invoke, we insert the store at the beginning in the | |||
| 1248 | // 'normal destination' BB. Otherwise we insert the store right after OutI. | |||
| 1249 | if (auto *InvokeI = dyn_cast<InvokeInst>(OutI)) | |||
| 1250 | InsertPt = InvokeI->getNormalDest()->getFirstInsertionPt(); | |||
| 1251 | else if (auto *Phi = dyn_cast<PHINode>(OutI)) | |||
| 1252 | InsertPt = Phi->getParent()->getFirstInsertionPt(); | |||
| 1253 | else | |||
| 1254 | InsertPt = std::next(OutI->getIterator()); | |||
| 1255 | ||||
| 1256 | Instruction *InsertBefore = &*InsertPt; | |||
| 1257 | assert((InsertBefore->getFunction() == newFunction ||(((InsertBefore->getFunction() == newFunction || Blocks.count (InsertBefore->getParent())) && "InsertPt should be in new function" ) ? static_cast<void> (0) : __assert_fail ("(InsertBefore->getFunction() == newFunction || Blocks.count(InsertBefore->getParent())) && \"InsertPt should be in new function\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1259, __PRETTY_FUNCTION__)) | |||
| 1258 | Blocks.count(InsertBefore->getParent())) &&(((InsertBefore->getFunction() == newFunction || Blocks.count (InsertBefore->getParent())) && "InsertPt should be in new function" ) ? static_cast<void> (0) : __assert_fail ("(InsertBefore->getFunction() == newFunction || Blocks.count(InsertBefore->getParent())) && \"InsertPt should be in new function\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1259, __PRETTY_FUNCTION__)) | |||
| 1259 | "InsertPt should be in new function")(((InsertBefore->getFunction() == newFunction || Blocks.count (InsertBefore->getParent())) && "InsertPt should be in new function" ) ? static_cast<void> (0) : __assert_fail ("(InsertBefore->getFunction() == newFunction || Blocks.count(InsertBefore->getParent())) && \"InsertPt should be in new function\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1259, __PRETTY_FUNCTION__)); | |||
| 1260 | assert(OAI != newFunction->arg_end() &&((OAI != newFunction->arg_end() && "Number of output arguments should match " "the amount of defined values") ? static_cast<void> (0 ) : __assert_fail ("OAI != newFunction->arg_end() && \"Number of output arguments should match \" \"the amount of defined values\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1262, __PRETTY_FUNCTION__)) | |||
| 1261 | "Number of output arguments should match "((OAI != newFunction->arg_end() && "Number of output arguments should match " "the amount of defined values") ? static_cast<void> (0 ) : __assert_fail ("OAI != newFunction->arg_end() && \"Number of output arguments should match \" \"the amount of defined values\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1262, __PRETTY_FUNCTION__)) | |||
| 1262 | "the amount of defined values")((OAI != newFunction->arg_end() && "Number of output arguments should match " "the amount of defined values") ? static_cast<void> (0 ) : __assert_fail ("OAI != newFunction->arg_end() && \"Number of output arguments should match \" \"the amount of defined values\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1262, __PRETTY_FUNCTION__)); | |||
| 1263 | if (AggregateArgs) { | |||
| 1264 | Value *Idx[2]; | |||
| 1265 | Idx[0] = Constant::getNullValue(Type::getInt32Ty(Context)); | |||
| 1266 | Idx[1] = ConstantInt::get(Type::getInt32Ty(Context), FirstOut + i); | |||
| 1267 | GetElementPtrInst *GEP = GetElementPtrInst::Create( | |||
| 1268 | StructArgTy, &*OAI, Idx, "gep_" + outputs[i]->getName(), | |||
| 1269 | InsertBefore); | |||
| 1270 | new StoreInst(outputs[i], GEP, InsertBefore); | |||
| 1271 | // Since there should be only one struct argument aggregating | |||
| 1272 | // all the output values, we shouldn't increment OAI, which always | |||
| 1273 | // points to the struct argument, in this case. | |||
| 1274 | } else { | |||
| 1275 | new StoreInst(outputs[i], &*OAI, InsertBefore); | |||
| 1276 | ++OAI; | |||
| 1277 | } | |||
| 1278 | } | |||
| 1279 | ||||
| 1280 | // Now that we've done the deed, simplify the switch instruction. | |||
| 1281 | Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType(); | |||
| 1282 | switch (NumExitBlocks) { | |||
| 1283 | case 0: | |||
| 1284 | // There are no successors (the block containing the switch itself), which | |||
| 1285 | // means that previously this was the last part of the function, and hence | |||
| 1286 | // this should be rewritten as a `ret' | |||
| 1287 | ||||
| 1288 | // Check if the function should return a value | |||
| 1289 | if (OldFnRetTy->isVoidTy()) { | |||
| 1290 | ReturnInst::Create(Context, nullptr, TheSwitch); // Return void | |||
| 1291 | } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) { | |||
| 1292 | // return what we have | |||
| 1293 | ReturnInst::Create(Context, TheSwitch->getCondition(), TheSwitch); | |||
| 1294 | } else { | |||
| 1295 | // Otherwise we must have code extracted an unwind or something, just | |||
| 1296 | // return whatever we want. | |||
| 1297 | ReturnInst::Create(Context, | |||
| 1298 | Constant::getNullValue(OldFnRetTy), TheSwitch); | |||
| 1299 | } | |||
| 1300 | ||||
| 1301 | TheSwitch->eraseFromParent(); | |||
| 1302 | break; | |||
| 1303 | case 1: | |||
| 1304 | // Only a single destination, change the switch into an unconditional | |||
| 1305 | // branch. | |||
| 1306 | BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch); | |||
| 1307 | TheSwitch->eraseFromParent(); | |||
| 1308 | break; | |||
| 1309 | case 2: | |||
| 1310 | BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2), | |||
| 1311 | call, TheSwitch); | |||
| 1312 | TheSwitch->eraseFromParent(); | |||
| 1313 | break; | |||
| 1314 | default: | |||
| 1315 | // Otherwise, make the default destination of the switch instruction be one | |||
| 1316 | // of the other successors. | |||
| 1317 | TheSwitch->setCondition(call); | |||
| 1318 | TheSwitch->setDefaultDest(TheSwitch->getSuccessor(NumExitBlocks)); | |||
| 1319 | // Remove redundant case | |||
| 1320 | TheSwitch->removeCase(SwitchInst::CaseIt(TheSwitch, NumExitBlocks-1)); | |||
| 1321 | break; | |||
| 1322 | } | |||
| 1323 | ||||
| 1324 | // Insert lifetime markers around the reloads of any output values. The | |||
| 1325 | // allocas output values are stored in are only in-use in the codeRepl block. | |||
| 1326 | insertLifetimeMarkersSurroundingCall(M, ReloadOutputs, ReloadOutputs, call); | |||
| 1327 | ||||
| 1328 | return call; | |||
| 1329 | } | |||
| 1330 | ||||
| 1331 | void CodeExtractor::moveCodeToFunction(Function *newFunction) { | |||
| 1332 | Function *oldFunc = (*Blocks.begin())->getParent(); | |||
| 1333 | Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList(); | |||
| 1334 | Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList(); | |||
| 1335 | ||||
| 1336 | for (BasicBlock *Block : Blocks) { | |||
| 1337 | // Delete the basic block from the old function, and the list of blocks | |||
| 1338 | oldBlocks.remove(Block); | |||
| 1339 | ||||
| 1340 | // Insert this basic block into the new function | |||
| 1341 | newBlocks.push_back(Block); | |||
| 1342 | } | |||
| 1343 | } | |||
| 1344 | ||||
| 1345 | void CodeExtractor::calculateNewCallTerminatorWeights( | |||
| 1346 | BasicBlock *CodeReplacer, | |||
| 1347 | DenseMap<BasicBlock *, BlockFrequency> &ExitWeights, | |||
| 1348 | BranchProbabilityInfo *BPI) { | |||
| 1349 | using Distribution = BlockFrequencyInfoImplBase::Distribution; | |||
| 1350 | using BlockNode = BlockFrequencyInfoImplBase::BlockNode; | |||
| 1351 | ||||
| 1352 | // Update the branch weights for the exit block. | |||
| 1353 | Instruction *TI = CodeReplacer->getTerminator(); | |||
| 1354 | SmallVector<unsigned, 8> BranchWeights(TI->getNumSuccessors(), 0); | |||
| 1355 | ||||
| 1356 | // Block Frequency distribution with dummy node. | |||
| 1357 | Distribution BranchDist; | |||
| 1358 | ||||
| 1359 | // Add each of the frequencies of the successors. | |||
| 1360 | for (unsigned i = 0, e = TI->getNumSuccessors(); i < e; ++i) { | |||
| 1361 | BlockNode ExitNode(i); | |||
| 1362 | uint64_t ExitFreq = ExitWeights[TI->getSuccessor(i)].getFrequency(); | |||
| 1363 | if (ExitFreq != 0) | |||
| 1364 | BranchDist.addExit(ExitNode, ExitFreq); | |||
| 1365 | else | |||
| 1366 | BPI->setEdgeProbability(CodeReplacer, i, BranchProbability::getZero()); | |||
| 1367 | } | |||
| 1368 | ||||
| 1369 | // Check for no total weight. | |||
| 1370 | if (BranchDist.Total == 0) | |||
| 1371 | return; | |||
| 1372 | ||||
| 1373 | // Normalize the distribution so that they can fit in unsigned. | |||
| 1374 | BranchDist.normalize(); | |||
| 1375 | ||||
| 1376 | // Create normalized branch weights and set the metadata. | |||
| 1377 | for (unsigned I = 0, E = BranchDist.Weights.size(); I < E; ++I) { | |||
| 1378 | const auto &Weight = BranchDist.Weights[I]; | |||
| 1379 | ||||
| 1380 | // Get the weight and update the current BFI. | |||
| 1381 | BranchWeights[Weight.TargetNode.Index] = Weight.Amount; | |||
| 1382 | BranchProbability BP(Weight.Amount, BranchDist.Total); | |||
| 1383 | BPI->setEdgeProbability(CodeReplacer, Weight.TargetNode.Index, BP); | |||
| 1384 | } | |||
| 1385 | TI->setMetadata( | |||
| 1386 | LLVMContext::MD_prof, | |||
| 1387 | MDBuilder(TI->getContext()).createBranchWeights(BranchWeights)); | |||
| 1388 | } | |||
| 1389 | ||||
| 1390 | /// Erase debug info intrinsics which refer to values in \p F but aren't in | |||
| 1391 | /// \p F. | |||
| 1392 | static void eraseDebugIntrinsicsWithNonLocalRefs(Function &F) { | |||
| 1393 | for (Instruction &I : instructions(F)) { | |||
| 1394 | SmallVector<DbgVariableIntrinsic *, 4> DbgUsers; | |||
| 1395 | findDbgUsers(DbgUsers, &I); | |||
| 1396 | for (DbgVariableIntrinsic *DVI : DbgUsers) | |||
| 1397 | if (DVI->getFunction() != &F) | |||
| 1398 | DVI->eraseFromParent(); | |||
| 1399 | } | |||
| 1400 | } | |||
| 1401 | ||||
| 1402 | /// Fix up the debug info in the old and new functions by pointing line | |||
| 1403 | /// locations and debug intrinsics to the new subprogram scope, and by deleting | |||
| 1404 | /// intrinsics which point to values outside of the new function. | |||
| 1405 | static void fixupDebugInfoPostExtraction(Function &OldFunc, Function &NewFunc, | |||
| 1406 | CallInst &TheCall) { | |||
| 1407 | DISubprogram *OldSP = OldFunc.getSubprogram(); | |||
| 1408 | LLVMContext &Ctx = OldFunc.getContext(); | |||
| 1409 | ||||
| 1410 | if (!OldSP) { | |||
| 1411 | // Erase any debug info the new function contains. | |||
| 1412 | stripDebugInfo(NewFunc); | |||
| 1413 | // Make sure the old function doesn't contain any non-local metadata refs. | |||
| 1414 | eraseDebugIntrinsicsWithNonLocalRefs(NewFunc); | |||
| 1415 | return; | |||
| 1416 | } | |||
| 1417 | ||||
| 1418 | // Create a subprogram for the new function. Leave out a description of the | |||
| 1419 | // function arguments, as the parameters don't correspond to anything at the | |||
| 1420 | // source level. | |||
| 1421 | assert(OldSP->getUnit() && "Missing compile unit for subprogram")((OldSP->getUnit() && "Missing compile unit for subprogram" ) ? static_cast<void> (0) : __assert_fail ("OldSP->getUnit() && \"Missing compile unit for subprogram\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1421, __PRETTY_FUNCTION__)); | |||
| 1422 | DIBuilder DIB(*OldFunc.getParent(), /*AllowUnresolvedNodes=*/false, | |||
| 1423 | OldSP->getUnit()); | |||
| 1424 | auto SPType = DIB.createSubroutineType(DIB.getOrCreateTypeArray(None)); | |||
| 1425 | DISubprogram::DISPFlags SPFlags = DISubprogram::SPFlagDefinition | | |||
| 1426 | DISubprogram::SPFlagOptimized | | |||
| 1427 | DISubprogram::SPFlagLocalToUnit; | |||
| 1428 | auto NewSP = DIB.createFunction( | |||
| 1429 | OldSP->getUnit(), NewFunc.getName(), NewFunc.getName(), OldSP->getFile(), | |||
| 1430 | /*LineNo=*/0, SPType, /*ScopeLine=*/0, DINode::FlagZero, SPFlags); | |||
| 1431 | NewFunc.setSubprogram(NewSP); | |||
| 1432 | ||||
| 1433 | // Debug intrinsics in the new function need to be updated in one of two | |||
| 1434 | // ways: | |||
| 1435 | // 1) They need to be deleted, because they describe a value in the old | |||
| 1436 | // function. | |||
| 1437 | // 2) They need to point to fresh metadata, e.g. because they currently | |||
| 1438 | // point to a variable in the wrong scope. | |||
| 1439 | SmallDenseMap<DINode *, DINode *> RemappedMetadata; | |||
| 1440 | SmallVector<Instruction *, 4> DebugIntrinsicsToDelete; | |||
| 1441 | for (Instruction &I : instructions(NewFunc)) { | |||
| 1442 | auto *DII = dyn_cast<DbgInfoIntrinsic>(&I); | |||
| 1443 | if (!DII) | |||
| 1444 | continue; | |||
| 1445 | ||||
| 1446 | // Point the intrinsic to a fresh label within the new function. | |||
| 1447 | if (auto *DLI = dyn_cast<DbgLabelInst>(&I)) { | |||
| 1448 | DILabel *OldLabel = DLI->getLabel(); | |||
| 1449 | DINode *&NewLabel = RemappedMetadata[OldLabel]; | |||
| 1450 | if (!NewLabel) | |||
| 1451 | NewLabel = DILabel::get(Ctx, NewSP, OldLabel->getName(), | |||
| 1452 | OldLabel->getFile(), OldLabel->getLine()); | |||
| 1453 | DLI->setArgOperand(0, MetadataAsValue::get(Ctx, NewLabel)); | |||
| 1454 | continue; | |||
| 1455 | } | |||
| 1456 | ||||
| 1457 | // If the location isn't a constant or an instruction, delete the | |||
| 1458 | // intrinsic. | |||
| 1459 | auto *DVI = cast<DbgVariableIntrinsic>(DII); | |||
| 1460 | Value *Location = DVI->getVariableLocation(); | |||
| 1461 | if (!Location || | |||
| 1462 | (!isa<Constant>(Location) && !isa<Instruction>(Location))) { | |||
| 1463 | DebugIntrinsicsToDelete.push_back(DVI); | |||
| 1464 | continue; | |||
| 1465 | } | |||
| 1466 | ||||
| 1467 | // If the variable location is an instruction but isn't in the new | |||
| 1468 | // function, delete the intrinsic. | |||
| 1469 | Instruction *LocationInst = dyn_cast<Instruction>(Location); | |||
| 1470 | if (LocationInst && LocationInst->getFunction() != &NewFunc) { | |||
| 1471 | DebugIntrinsicsToDelete.push_back(DVI); | |||
| 1472 | continue; | |||
| 1473 | } | |||
| 1474 | ||||
| 1475 | // Point the intrinsic to a fresh variable within the new function. | |||
| 1476 | DILocalVariable *OldVar = DVI->getVariable(); | |||
| 1477 | DINode *&NewVar = RemappedMetadata[OldVar]; | |||
| 1478 | if (!NewVar) | |||
| 1479 | NewVar = DIB.createAutoVariable( | |||
| 1480 | NewSP, OldVar->getName(), OldVar->getFile(), OldVar->getLine(), | |||
| 1481 | OldVar->getType(), /*AlwaysPreserve=*/false, DINode::FlagZero, | |||
| 1482 | OldVar->getAlignInBits()); | |||
| 1483 | DVI->setArgOperand(1, MetadataAsValue::get(Ctx, NewVar)); | |||
| 1484 | } | |||
| 1485 | for (auto *DII : DebugIntrinsicsToDelete) | |||
| 1486 | DII->eraseFromParent(); | |||
| 1487 | DIB.finalizeSubprogram(NewSP); | |||
| 1488 | ||||
| 1489 | // Fix up the scope information attached to the line locations in the new | |||
| 1490 | // function. | |||
| 1491 | for (Instruction &I : instructions(NewFunc)) { | |||
| 1492 | if (const DebugLoc &DL = I.getDebugLoc()) | |||
| 1493 | I.setDebugLoc(DebugLoc::get(DL.getLine(), DL.getCol(), NewSP)); | |||
| 1494 | ||||
| 1495 | // Loop info metadata may contain line locations. Fix them up. | |||
| 1496 | auto updateLoopInfoLoc = [&Ctx, | |||
| 1497 | NewSP](const DILocation &Loc) -> DILocation * { | |||
| 1498 | return DILocation::get(Ctx, Loc.getLine(), Loc.getColumn(), NewSP, | |||
| 1499 | nullptr); | |||
| 1500 | }; | |||
| 1501 | updateLoopMetadataDebugLocations(I, updateLoopInfoLoc); | |||
| 1502 | } | |||
| 1503 | if (!TheCall.getDebugLoc()) | |||
| 1504 | TheCall.setDebugLoc(DebugLoc::get(0, 0, OldSP)); | |||
| 1505 | ||||
| 1506 | eraseDebugIntrinsicsWithNonLocalRefs(NewFunc); | |||
| 1507 | } | |||
| 1508 | ||||
| 1509 | Function * | |||
| 1510 | CodeExtractor::extractCodeRegion(const CodeExtractorAnalysisCache &CEAC) { | |||
| 1511 | if (!isEligible()) | |||
| 1512 | return nullptr; | |||
| 1513 | ||||
| 1514 | // Assumption: this is a single-entry code region, and the header is the first | |||
| 1515 | // block in the region. | |||
| 1516 | BasicBlock *header = *Blocks.begin(); | |||
| 1517 | Function *oldFunction = header->getParent(); | |||
| 1518 | ||||
| 1519 | // Calculate the entry frequency of the new function before we change the root | |||
| 1520 | // block. | |||
| 1521 | BlockFrequency EntryFreq; | |||
| 1522 | if (BFI) { | |||
| 1523 | assert(BPI && "Both BPI and BFI are required to preserve profile info")((BPI && "Both BPI and BFI are required to preserve profile info" ) ? static_cast<void> (0) : __assert_fail ("BPI && \"Both BPI and BFI are required to preserve profile info\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1523, __PRETTY_FUNCTION__)); | |||
| 1524 | for (BasicBlock *Pred : predecessors(header)) { | |||
| 1525 | if (Blocks.count(Pred)) | |||
| 1526 | continue; | |||
| 1527 | EntryFreq += | |||
| 1528 | BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, header); | |||
| 1529 | } | |||
| 1530 | } | |||
| 1531 | ||||
| 1532 | // Remove @llvm.assume calls that will be moved to the new function from the | |||
| 1533 | // old function's assumption cache. | |||
| 1534 | for (BasicBlock *Block : Blocks) { | |||
| 1535 | for (auto It = Block->begin(), End = Block->end(); It != End;) { | |||
| 1536 | Instruction *I = &*It; | |||
| 1537 | ++It; | |||
| 1538 | ||||
| 1539 | if (match(I, m_Intrinsic<Intrinsic::assume>())) { | |||
| 1540 | if (AC) | |||
| 1541 | AC->unregisterAssumption(cast<CallInst>(I)); | |||
| 1542 | I->eraseFromParent(); | |||
| 1543 | } | |||
| 1544 | } | |||
| 1545 | } | |||
| 1546 | ||||
| 1547 | // If we have any return instructions in the region, split those blocks so | |||
| 1548 | // that the return is not in the region. | |||
| 1549 | splitReturnBlocks(); | |||
| 1550 | ||||
| 1551 | // Calculate the exit blocks for the extracted region and the total exit | |||
| 1552 | // weights for each of those blocks. | |||
| 1553 | DenseMap<BasicBlock *, BlockFrequency> ExitWeights; | |||
| 1554 | SmallPtrSet<BasicBlock *, 1> ExitBlocks; | |||
| 1555 | for (BasicBlock *Block : Blocks) { | |||
| 1556 | for (succ_iterator SI = succ_begin(Block), SE = succ_end(Block); SI != SE; | |||
| 1557 | ++SI) { | |||
| 1558 | if (!Blocks.count(*SI)) { | |||
| 1559 | // Update the branch weight for this successor. | |||
| 1560 | if (BFI) { | |||
| 1561 | BlockFrequency &BF = ExitWeights[*SI]; | |||
| 1562 | BF += BFI->getBlockFreq(Block) * BPI->getEdgeProbability(Block, *SI); | |||
| 1563 | } | |||
| 1564 | ExitBlocks.insert(*SI); | |||
| 1565 | } | |||
| 1566 | } | |||
| 1567 | } | |||
| 1568 | NumExitBlocks = ExitBlocks.size(); | |||
| 1569 | ||||
| 1570 | // If we have to split PHI nodes of the entry or exit blocks, do so now. | |||
| 1571 | severSplitPHINodesOfEntry(header); | |||
| 1572 | severSplitPHINodesOfExits(ExitBlocks); | |||
| 1573 | ||||
| 1574 | // This takes place of the original loop | |||
| 1575 | BasicBlock *codeReplacer = BasicBlock::Create(header->getContext(), | |||
| 1576 | "codeRepl", oldFunction, | |||
| 1577 | header); | |||
| 1578 | ||||
| 1579 | // The new function needs a root node because other nodes can branch to the | |||
| 1580 | // head of the region, but the entry node of a function cannot have preds. | |||
| 1581 | BasicBlock *newFuncRoot = BasicBlock::Create(header->getContext(), | |||
| 1582 | "newFuncRoot"); | |||
| 1583 | auto *BranchI = BranchInst::Create(header); | |||
| 1584 | // If the original function has debug info, we have to add a debug location | |||
| 1585 | // to the new branch instruction from the artificial entry block. | |||
| 1586 | // We use the debug location of the first instruction in the extracted | |||
| 1587 | // blocks, as there is no other equivalent line in the source code. | |||
| 1588 | if (oldFunction->getSubprogram()) { | |||
| 1589 | any_of(Blocks, [&BranchI](const BasicBlock *BB) { | |||
| 1590 | return any_of(*BB, [&BranchI](const Instruction &I) { | |||
| 1591 | if (!I.getDebugLoc()) | |||
| 1592 | return false; | |||
| 1593 | BranchI->setDebugLoc(I.getDebugLoc()); | |||
| 1594 | return true; | |||
| 1595 | }); | |||
| 1596 | }); | |||
| 1597 | } | |||
| 1598 | newFuncRoot->getInstList().push_back(BranchI); | |||
| 1599 | ||||
| 1600 | ValueSet inputs, outputs, SinkingCands, HoistingCands; | |||
| 1601 | BasicBlock *CommonExit = nullptr; | |||
| 1602 | findAllocas(CEAC, SinkingCands, HoistingCands, CommonExit); | |||
| 1603 | assert(HoistingCands.empty() || CommonExit)((HoistingCands.empty() || CommonExit) ? static_cast<void> (0) : __assert_fail ("HoistingCands.empty() || CommonExit", "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1603, __PRETTY_FUNCTION__)); | |||
| 1604 | ||||
| 1605 | // Find inputs to, outputs from the code region. | |||
| 1606 | findInputsOutputs(inputs, outputs, SinkingCands); | |||
| 1607 | ||||
| 1608 | // Now sink all instructions which only have non-phi uses inside the region. | |||
| 1609 | // Group the allocas at the start of the block, so that any bitcast uses of | |||
| 1610 | // the allocas are well-defined. | |||
| 1611 | AllocaInst *FirstSunkAlloca = nullptr; | |||
| 1612 | for (auto *II : SinkingCands) { | |||
| 1613 | if (auto *AI = dyn_cast<AllocaInst>(II)) { | |||
| 1614 | AI->moveBefore(*newFuncRoot, newFuncRoot->getFirstInsertionPt()); | |||
| 1615 | if (!FirstSunkAlloca) | |||
| 1616 | FirstSunkAlloca = AI; | |||
| 1617 | } | |||
| 1618 | } | |||
| 1619 | assert((SinkingCands.empty() || FirstSunkAlloca) &&(((SinkingCands.empty() || FirstSunkAlloca) && "Did not expect a sink candidate without any allocas" ) ? static_cast<void> (0) : __assert_fail ("(SinkingCands.empty() || FirstSunkAlloca) && \"Did not expect a sink candidate without any allocas\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1620, __PRETTY_FUNCTION__)) | |||
| 1620 | "Did not expect a sink candidate without any allocas")(((SinkingCands.empty() || FirstSunkAlloca) && "Did not expect a sink candidate without any allocas" ) ? static_cast<void> (0) : __assert_fail ("(SinkingCands.empty() || FirstSunkAlloca) && \"Did not expect a sink candidate without any allocas\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1620, __PRETTY_FUNCTION__)); | |||
| 1621 | for (auto *II : SinkingCands) { | |||
| 1622 | if (!isa<AllocaInst>(II)) { | |||
| 1623 | cast<Instruction>(II)->moveAfter(FirstSunkAlloca); | |||
| 1624 | } | |||
| 1625 | } | |||
| 1626 | ||||
| 1627 | if (!HoistingCands.empty()) { | |||
| 1628 | auto *HoistToBlock = findOrCreateBlockForHoisting(CommonExit); | |||
| 1629 | Instruction *TI = HoistToBlock->getTerminator(); | |||
| 1630 | for (auto *II : HoistingCands) | |||
| 1631 | cast<Instruction>(II)->moveBefore(TI); | |||
| 1632 | } | |||
| 1633 | ||||
| 1634 | // Collect objects which are inputs to the extraction region and also | |||
| 1635 | // referenced by lifetime start markers within it. The effects of these | |||
| 1636 | // markers must be replicated in the calling function to prevent the stack | |||
| 1637 | // coloring pass from merging slots which store input objects. | |||
| 1638 | ValueSet LifetimesStart; | |||
| 1639 | eraseLifetimeMarkersOnInputs(Blocks, SinkingCands, LifetimesStart); | |||
| 1640 | ||||
| 1641 | // Construct new function based on inputs/outputs & add allocas for all defs. | |||
| 1642 | Function *newFunction = | |||
| 1643 | constructFunction(inputs, outputs, header, newFuncRoot, codeReplacer, | |||
| 1644 | oldFunction, oldFunction->getParent()); | |||
| 1645 | ||||
| 1646 | // Update the entry count of the function. | |||
| 1647 | if (BFI) { | |||
| 1648 | auto Count = BFI->getProfileCountFromFreq(EntryFreq.getFrequency()); | |||
| 1649 | if (Count.hasValue()) | |||
| 1650 | newFunction->setEntryCount( | |||
| 1651 | ProfileCount(Count.getValue(), Function::PCT_Real)); // FIXME | |||
| 1652 | BFI->setBlockFreq(codeReplacer, EntryFreq.getFrequency()); | |||
| 1653 | } | |||
| 1654 | ||||
| 1655 | CallInst *TheCall = | |||
| 1656 | emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs); | |||
| 1657 | ||||
| 1658 | moveCodeToFunction(newFunction); | |||
| 1659 | ||||
| 1660 | // Replicate the effects of any lifetime start/end markers which referenced | |||
| 1661 | // input objects in the extraction region by placing markers around the call. | |||
| 1662 | insertLifetimeMarkersSurroundingCall( | |||
| 1663 | oldFunction->getParent(), LifetimesStart.getArrayRef(), {}, TheCall); | |||
| 1664 | ||||
| 1665 | // Propagate personality info to the new function if there is one. | |||
| 1666 | if (oldFunction->hasPersonalityFn()) | |||
| 1667 | newFunction->setPersonalityFn(oldFunction->getPersonalityFn()); | |||
| 1668 | ||||
| 1669 | // Update the branch weights for the exit block. | |||
| 1670 | if (BFI && NumExitBlocks > 1) | |||
| 1671 | calculateNewCallTerminatorWeights(codeReplacer, ExitWeights, BPI); | |||
| 1672 | ||||
| 1673 | // Loop over all of the PHI nodes in the header and exit blocks, and change | |||
| 1674 | // any references to the old incoming edge to be the new incoming edge. | |||
| 1675 | for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) { | |||
| 1676 | PHINode *PN = cast<PHINode>(I); | |||
| 1677 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) | |||
| 1678 | if (!Blocks.count(PN->getIncomingBlock(i))) | |||
| 1679 | PN->setIncomingBlock(i, newFuncRoot); | |||
| 1680 | } | |||
| 1681 | ||||
| 1682 | for (BasicBlock *ExitBB : ExitBlocks) | |||
| 1683 | for (PHINode &PN : ExitBB->phis()) { | |||
| 1684 | Value *IncomingCodeReplacerVal = nullptr; | |||
| 1685 | for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { | |||
| 1686 | // Ignore incoming values from outside of the extracted region. | |||
| 1687 | if (!Blocks.count(PN.getIncomingBlock(i))) | |||
| 1688 | continue; | |||
| 1689 | ||||
| 1690 | // Ensure that there is only one incoming value from codeReplacer. | |||
| 1691 | if (!IncomingCodeReplacerVal) { | |||
| 1692 | PN.setIncomingBlock(i, codeReplacer); | |||
| 1693 | IncomingCodeReplacerVal = PN.getIncomingValue(i); | |||
| 1694 | } else | |||
| 1695 | assert(IncomingCodeReplacerVal == PN.getIncomingValue(i) &&((IncomingCodeReplacerVal == PN.getIncomingValue(i) && "PHI has two incompatbile incoming values from codeRepl") ? static_cast <void> (0) : __assert_fail ("IncomingCodeReplacerVal == PN.getIncomingValue(i) && \"PHI has two incompatbile incoming values from codeRepl\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1696, __PRETTY_FUNCTION__)) | |||
| 1696 | "PHI has two incompatbile incoming values from codeRepl")((IncomingCodeReplacerVal == PN.getIncomingValue(i) && "PHI has two incompatbile incoming values from codeRepl") ? static_cast <void> (0) : __assert_fail ("IncomingCodeReplacerVal == PN.getIncomingValue(i) && \"PHI has two incompatbile incoming values from codeRepl\"" , "/build/llvm-toolchain-snapshot-11~++20200309111110+2c36c23f347/llvm/lib/Transforms/Utils/CodeExtractor.cpp" , 1696, __PRETTY_FUNCTION__)); | |||
| 1697 | } | |||
| 1698 | } | |||
| 1699 | ||||
| 1700 | fixupDebugInfoPostExtraction(*oldFunction, *newFunction, *TheCall); | |||
| 1701 | ||||
| 1702 | // Mark the new function `noreturn` if applicable. Terminators which resume | |||
| 1703 | // exception propagation are treated as returning instructions. This is to | |||
| 1704 | // avoid inserting traps after calls to outlined functions which unwind. | |||
| 1705 | bool doesNotReturn = none_of(*newFunction, [](const BasicBlock &BB) { | |||
| 1706 | const Instruction *Term = BB.getTerminator(); | |||
| 1707 | return isa<ReturnInst>(Term) || isa<ResumeInst>(Term); | |||
| 1708 | }); | |||
| 1709 | if (doesNotReturn) | |||
| 1710 | newFunction->setDoesNotReturn(); | |||
| 1711 | ||||
| 1712 | LLVM_DEBUG(if (verifyFunction(*newFunction, &errs())) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { if (verifyFunction(*newFunction, &errs ())) { newFunction->dump(); report_fatal_error("verification of newFunction failed!" ); }; } } while (false) | |||
| 1713 | newFunction->dump();do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { if (verifyFunction(*newFunction, &errs ())) { newFunction->dump(); report_fatal_error("verification of newFunction failed!" ); }; } } while (false) | |||
| 1714 | report_fatal_error("verification of newFunction failed!");do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { if (verifyFunction(*newFunction, &errs ())) { newFunction->dump(); report_fatal_error("verification of newFunction failed!" ); }; } } while (false) | |||
| 1715 | })do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { if (verifyFunction(*newFunction, &errs ())) { newFunction->dump(); report_fatal_error("verification of newFunction failed!" ); }; } } while (false); | |||
| 1716 | LLVM_DEBUG(if (verifyFunction(*oldFunction))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { if (verifyFunction(*oldFunction)) report_fatal_error ("verification of oldFunction failed!"); } } while (false) | |||
| 1717 | report_fatal_error("verification of oldFunction failed!"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { if (verifyFunction(*oldFunction)) report_fatal_error ("verification of oldFunction failed!"); } } while (false); | |||
| 1718 | LLVM_DEBUG(if (AC && verifyAssumptionCache(*oldFunction, *newFunction, AC))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { if (AC && verifyAssumptionCache (*oldFunction, *newFunction, AC)) report_fatal_error("Stale Asumption cache for old Function!" ); } } while (false) | |||
| 1719 | report_fatal_error("Stale Asumption cache for old Function!"))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("code-extractor")) { if (AC && verifyAssumptionCache (*oldFunction, *newFunction, AC)) report_fatal_error("Stale Asumption cache for old Function!" ); } } while (false); | |||
| 1720 | return newFunction; | |||
| 1721 | } | |||
| 1722 | ||||
| 1723 | bool CodeExtractor::verifyAssumptionCache(const Function &OldFunc, | |||
| 1724 | const Function &NewFunc, | |||
| 1725 | AssumptionCache *AC) { | |||
| 1726 | for (auto AssumeVH : AC->assumptions()) { | |||
| 1727 | CallInst *I = dyn_cast_or_null<CallInst>(AssumeVH); | |||
| 1728 | if (!I) | |||
| 1729 | continue; | |||
| 1730 | ||||
| 1731 | // There shouldn't be any llvm.assume intrinsics in the new function. | |||
| 1732 | if (I->getFunction() != &OldFunc) | |||
| 1733 | return true; | |||
| 1734 | ||||
| 1735 | // There shouldn't be any stale affected values in the assumption cache | |||
| 1736 | // that were previously in the old function, but that have now been moved | |||
| 1737 | // to the new function. | |||
| 1738 | for (auto AffectedValVH : AC->assumptionsFor(I->getOperand(0))) { | |||
| 1739 | CallInst *AffectedCI = dyn_cast_or_null<CallInst>(AffectedValVH); | |||
| 1740 | if (!AffectedCI) | |||
| 1741 | continue; | |||
| 1742 | if (AffectedCI->getFunction() != &OldFunc) | |||
| 1743 | return true; | |||
| 1744 | auto *AssumedInst = dyn_cast<Instruction>(AffectedCI->getOperand(0)); | |||
| 1745 | if (AssumedInst->getFunction() != &OldFunc) | |||
| 1746 | return true; | |||
| 1747 | } | |||
| 1748 | } | |||
| 1749 | return false; | |||
| 1750 | } |