Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line | 
|---|---|---|---|
| 14 | pmbaty | 1 | //===- Cloning.h - Clone various parts of LLVM programs ---------*- C++ -*-===// | 
| 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 defines various functions that are used to clone chunks of LLVM | ||
| 10 | // code for various purposes.  This varies from copying whole modules into new | ||
| 11 | // modules, to cloning functions with different arguments, to inlining | ||
| 12 | // functions, to copying basic blocks to support loop unrolling or superblock | ||
| 13 | // formation, etc. | ||
| 14 | // | ||
| 15 | //===----------------------------------------------------------------------===// | ||
| 16 | |||
| 17 | #ifndef LLVM_TRANSFORMS_UTILS_CLONING_H | ||
| 18 | #define LLVM_TRANSFORMS_UTILS_CLONING_H | ||
| 19 | |||
| 20 | #include "llvm/ADT/SmallVector.h" | ||
| 21 | #include "llvm/ADT/Twine.h" | ||
| 22 | #include "llvm/Analysis/AssumptionCache.h" | ||
| 23 | #include "llvm/Analysis/InlineCost.h" | ||
| 24 | #include "llvm/IR/ValueHandle.h" | ||
| 25 | #include "llvm/Transforms/Utils/ValueMapper.h" | ||
| 26 | #include <functional> | ||
| 27 | #include <memory> | ||
| 28 | #include <vector> | ||
| 29 | |||
| 30 | namespace llvm { | ||
| 31 | |||
| 32 | class AAResults; | ||
| 33 | class AllocaInst; | ||
| 34 | class BasicBlock; | ||
| 35 | class BlockFrequencyInfo; | ||
| 36 | class CallGraph; | ||
| 37 | class DebugInfoFinder; | ||
| 38 | class DominatorTree; | ||
| 39 | class Function; | ||
| 40 | class Instruction; | ||
| 41 | class Loop; | ||
| 42 | class LoopInfo; | ||
| 43 | class Module; | ||
| 44 | class ProfileSummaryInfo; | ||
| 45 | class ReturnInst; | ||
| 46 | class DomTreeUpdater; | ||
| 47 | |||
| 48 | /// Return an exact copy of the specified module | ||
| 49 | std::unique_ptr<Module> CloneModule(const Module &M); | ||
| 50 | std::unique_ptr<Module> CloneModule(const Module &M, ValueToValueMapTy &VMap); | ||
| 51 | |||
| 52 | /// Return a copy of the specified module. The ShouldCloneDefinition function | ||
| 53 | /// controls whether a specific GlobalValue's definition is cloned. If the | ||
| 54 | /// function returns false, the module copy will contain an external reference | ||
| 55 | /// in place of the global definition. | ||
| 56 | std::unique_ptr<Module> | ||
| 57 | CloneModule(const Module &M, ValueToValueMapTy &VMap, | ||
| 58 | function_ref<bool(const GlobalValue *)> ShouldCloneDefinition); | ||
| 59 | |||
| 60 | /// This struct can be used to capture information about code | ||
| 61 | /// being cloned, while it is being cloned. | ||
| 62 | struct ClonedCodeInfo { | ||
| 63 |   /// This is set to true if the cloned code contains a normal call instruction. | ||
| 64 | bool ContainsCalls = false; | ||
| 65 | |||
| 66 |   /// This is set to true if there is memprof related metadata (memprof or | ||
| 67 |   /// callsite metadata) in the cloned code. | ||
| 68 | bool ContainsMemProfMetadata = false; | ||
| 69 | |||
| 70 |   /// This is set to true if the cloned code contains a 'dynamic' alloca. | ||
| 71 |   /// Dynamic allocas are allocas that are either not in the entry block or they | ||
| 72 |   /// are in the entry block but are not a constant size. | ||
| 73 | bool ContainsDynamicAllocas = false; | ||
| 74 | |||
| 75 |   /// All cloned call sites that have operand bundles attached are appended to | ||
| 76 |   /// this vector.  This vector may contain nulls or undefs if some of the | ||
| 77 |   /// originally inserted callsites were DCE'ed after they were cloned. | ||
| 78 | std::vector<WeakTrackingVH> OperandBundleCallSites; | ||
| 79 | |||
| 80 |   /// Like VMap, but maps only unsimplified instructions. Values in the map | ||
| 81 |   /// may be dangling, it is only intended to be used via isSimplified(), to | ||
| 82 |   /// check whether the main VMap mapping involves simplification or not. | ||
| 83 | DenseMap<const Value *, const Value *> OrigVMap; | ||
| 84 | |||
| 85 | ClonedCodeInfo() = default; | ||
| 86 | |||
| 87 | bool isSimplified(const Value *From, const Value *To) const { | ||
| 88 | return OrigVMap.lookup(From) != To; | ||
| 89 |   } | ||
| 90 | }; | ||
| 91 | |||
| 92 | /// Return a copy of the specified basic block, but without | ||
| 93 | /// embedding the block into a particular function.  The block returned is an | ||
| 94 | /// exact copy of the specified basic block, without any remapping having been | ||
| 95 | /// performed.  Because of this, this is only suitable for applications where | ||
| 96 | /// the basic block will be inserted into the same function that it was cloned | ||
| 97 | /// from (loop unrolling would use this, for example). | ||
| 98 | /// | ||
| 99 | /// Also, note that this function makes a direct copy of the basic block, and | ||
| 100 | /// can thus produce illegal LLVM code.  In particular, it will copy any PHI | ||
| 101 | /// nodes from the original block, even though there are no predecessors for the | ||
| 102 | /// newly cloned block (thus, phi nodes will have to be updated).  Also, this | ||
| 103 | /// block will branch to the old successors of the original block: these | ||
| 104 | /// successors will have to have any PHI nodes updated to account for the new | ||
| 105 | /// incoming edges. | ||
| 106 | /// | ||
| 107 | /// The correlation between instructions in the source and result basic blocks | ||
| 108 | /// is recorded in the VMap map. | ||
| 109 | /// | ||
| 110 | /// If you have a particular suffix you'd like to use to add to any cloned | ||
| 111 | /// names, specify it as the optional third parameter. | ||
| 112 | /// | ||
| 113 | /// If you would like the basic block to be auto-inserted into the end of a | ||
| 114 | /// function, you can specify it as the optional fourth parameter. | ||
| 115 | /// | ||
| 116 | /// If you would like to collect additional information about the cloned | ||
| 117 | /// function, you can specify a ClonedCodeInfo object with the optional fifth | ||
| 118 | /// parameter. | ||
| 119 | BasicBlock *CloneBasicBlock(const BasicBlock *BB, ValueToValueMapTy &VMap, | ||
| 120 | const Twine &NameSuffix = "", Function *F = nullptr, | ||
| 121 | ClonedCodeInfo *CodeInfo = nullptr, | ||
| 122 | DebugInfoFinder *DIFinder = nullptr); | ||
| 123 | |||
| 124 | /// Return a copy of the specified function and add it to that | ||
| 125 | /// function's module.  Also, any references specified in the VMap are changed | ||
| 126 | /// to refer to their mapped value instead of the original one.  If any of the | ||
| 127 | /// arguments to the function are in the VMap, the arguments are deleted from | ||
| 128 | /// the resultant function.  The VMap is updated to include mappings from all of | ||
| 129 | /// the instructions and basicblocks in the function from their old to new | ||
| 130 | /// values.  The final argument captures information about the cloned code if | ||
| 131 | /// non-null. | ||
| 132 | /// | ||
| 133 | /// \pre VMap contains no non-identity GlobalValue mappings. | ||
| 134 | /// | ||
| 135 | Function *CloneFunction(Function *F, ValueToValueMapTy &VMap, | ||
| 136 | ClonedCodeInfo *CodeInfo = nullptr); | ||
| 137 | |||
| 138 | enum class CloneFunctionChangeType { | ||
| 139 | LocalChangesOnly, | ||
| 140 | GlobalChanges, | ||
| 141 | DifferentModule, | ||
| 142 | ClonedModule, | ||
| 143 | }; | ||
| 144 | |||
| 145 | /// Clone OldFunc into NewFunc, transforming the old arguments into references | ||
| 146 | /// to VMap values.  Note that if NewFunc already has basic blocks, the ones | ||
| 147 | /// cloned into it will be added to the end of the function.  This function | ||
| 148 | /// fills in a list of return instructions, and can optionally remap types | ||
| 149 | /// and/or append the specified suffix to all values cloned. | ||
| 150 | /// | ||
| 151 | /// If \p Changes is \a CloneFunctionChangeType::LocalChangesOnly, VMap is | ||
| 152 | /// required to contain no non-identity GlobalValue mappings. Otherwise, | ||
| 153 | /// referenced metadata will be cloned. | ||
| 154 | /// | ||
| 155 | /// If \p Changes is less than \a CloneFunctionChangeType::DifferentModule | ||
| 156 | /// indicating cloning into the same module (even if it's LocalChangesOnly), if | ||
| 157 | /// debug info metadata transitively references a \a DISubprogram, it will be | ||
| 158 | /// cloned, effectively upgrading \p Changes to GlobalChanges while suppressing | ||
| 159 | /// cloning of types and compile units. | ||
| 160 | /// | ||
| 161 | /// If \p Changes is \a CloneFunctionChangeType::DifferentModule, the new | ||
| 162 | /// module's \c !llvm.dbg.cu will get updated with any newly created compile | ||
| 163 | /// units. (\a CloneFunctionChangeType::ClonedModule leaves that work for the | ||
| 164 | /// caller.) | ||
| 165 | /// | ||
| 166 | /// FIXME: Consider simplifying this function by splitting out \a | ||
| 167 | /// CloneFunctionMetadataInto() and expecting / updating callers to call it | ||
| 168 | /// first when / how it's needed. | ||
| 169 | void CloneFunctionInto(Function *NewFunc, const Function *OldFunc, | ||
| 170 |                        ValueToValueMapTy &VMap, CloneFunctionChangeType Changes, | ||
| 171 | SmallVectorImpl<ReturnInst *> &Returns, | ||
| 172 | const char *NameSuffix = "", | ||
| 173 | ClonedCodeInfo *CodeInfo = nullptr, | ||
| 174 | ValueMapTypeRemapper *TypeMapper = nullptr, | ||
| 175 | ValueMaterializer *Materializer = nullptr); | ||
| 176 | |||
| 177 | void CloneAndPruneIntoFromInst(Function *NewFunc, const Function *OldFunc, | ||
| 178 | const Instruction *StartingInst, | ||
| 179 | ValueToValueMapTy &VMap, bool ModuleLevelChanges, | ||
| 180 | SmallVectorImpl<ReturnInst *> &Returns, | ||
| 181 | const char *NameSuffix = "", | ||
| 182 | ClonedCodeInfo *CodeInfo = nullptr); | ||
| 183 | |||
| 184 | /// This works exactly like CloneFunctionInto, | ||
| 185 | /// except that it does some simple constant prop and DCE on the fly.  The | ||
| 186 | /// effect of this is to copy significantly less code in cases where (for | ||
| 187 | /// example) a function call with constant arguments is inlined, and those | ||
| 188 | /// constant arguments cause a significant amount of code in the callee to be | ||
| 189 | /// dead.  Since this doesn't produce an exactly copy of the input, it can't be | ||
| 190 | /// used for things like CloneFunction or CloneModule. | ||
| 191 | /// | ||
| 192 | /// If ModuleLevelChanges is false, VMap contains no non-identity GlobalValue | ||
| 193 | /// mappings. | ||
| 194 | /// | ||
| 195 | void CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc, | ||
| 196 | ValueToValueMapTy &VMap, bool ModuleLevelChanges, | ||
| 197 | SmallVectorImpl<ReturnInst*> &Returns, | ||
| 198 | const char *NameSuffix = "", | ||
| 199 | ClonedCodeInfo *CodeInfo = nullptr); | ||
| 200 | |||
| 201 | /// This class captures the data input to the InlineFunction call, and records | ||
| 202 | /// the auxiliary results produced by it. | ||
| 203 | class InlineFunctionInfo { | ||
| 204 | public: | ||
| 205 | explicit InlineFunctionInfo( | ||
| 206 | CallGraph *cg = nullptr, | ||
| 207 | function_ref<AssumptionCache &(Function &)> GetAssumptionCache = nullptr, | ||
| 208 | ProfileSummaryInfo *PSI = nullptr, | ||
| 209 | BlockFrequencyInfo *CallerBFI = nullptr, | ||
| 210 | BlockFrequencyInfo *CalleeBFI = nullptr, bool UpdateProfile = true) | ||
| 211 | : CG(cg), GetAssumptionCache(GetAssumptionCache), PSI(PSI), | ||
| 212 | CallerBFI(CallerBFI), CalleeBFI(CalleeBFI), | ||
| 213 | UpdateProfile(UpdateProfile) {} | ||
| 214 | |||
| 215 |   /// If non-null, InlineFunction will update the callgraph to reflect the | ||
| 216 |   /// changes it makes. | ||
| 217 | CallGraph *CG; | ||
| 218 | function_ref<AssumptionCache &(Function &)> GetAssumptionCache; | ||
| 219 | ProfileSummaryInfo *PSI; | ||
| 220 | BlockFrequencyInfo *CallerBFI, *CalleeBFI; | ||
| 221 | |||
| 222 |   /// InlineFunction fills this in with all static allocas that get copied into | ||
| 223 |   /// the caller. | ||
| 224 | SmallVector<AllocaInst *, 4> StaticAllocas; | ||
| 225 | |||
| 226 |   /// InlineFunction fills this in with callsites that were inlined from the | ||
| 227 |   /// callee. This is only filled in if CG is non-null. | ||
| 228 | SmallVector<WeakTrackingVH, 8> InlinedCalls; | ||
| 229 | |||
| 230 |   /// All of the new call sites inlined into the caller. | ||
| 231 |   /// | ||
| 232 |   /// 'InlineFunction' fills this in by scanning the inlined instructions, and | ||
| 233 |   /// only if CG is null. If CG is non-null, instead the value handle | ||
| 234 |   /// `InlinedCalls` above is used. | ||
| 235 | SmallVector<CallBase *, 8> InlinedCallSites; | ||
| 236 | |||
| 237 |   /// Update profile for callee as well as cloned version. We need to do this | ||
| 238 |   /// for regular inlining, but not for inlining from sample profile loader. | ||
| 239 | bool UpdateProfile; | ||
| 240 | |||
| 241 | void reset() { | ||
| 242 | StaticAllocas.clear(); | ||
| 243 | InlinedCalls.clear(); | ||
| 244 | InlinedCallSites.clear(); | ||
| 245 |   } | ||
| 246 | }; | ||
| 247 | |||
| 248 | /// This function inlines the called function into the basic | ||
| 249 | /// block of the caller.  This returns false if it is not possible to inline | ||
| 250 | /// this call.  The program is still in a well defined state if this occurs | ||
| 251 | /// though. | ||
| 252 | /// | ||
| 253 | /// Note that this only does one level of inlining.  For example, if the | ||
| 254 | /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now | ||
| 255 | /// exists in the instruction stream.  Similarly this will inline a recursive | ||
| 256 | /// function by one level. | ||
| 257 | /// | ||
| 258 | /// Note that while this routine is allowed to cleanup and optimize the | ||
| 259 | /// *inlined* code to minimize the actual inserted code, it must not delete | ||
| 260 | /// code in the caller as users of this routine may have pointers to | ||
| 261 | /// instructions in the caller that need to remain stable. | ||
| 262 | /// | ||
| 263 | /// If ForwardVarArgsTo is passed, inlining a function with varargs is allowed | ||
| 264 | /// and all varargs at the callsite will be passed to any calls to | ||
| 265 | /// ForwardVarArgsTo. The caller of InlineFunction has to make sure any varargs | ||
| 266 | /// are only used by ForwardVarArgsTo. | ||
| 267 | /// | ||
| 268 | /// The callee's function attributes are merged into the callers' if | ||
| 269 | /// MergeAttributes is set to true. | ||
| 270 | InlineResult InlineFunction(CallBase &CB, InlineFunctionInfo &IFI, | ||
| 271 | bool MergeAttributes = false, | ||
| 272 | AAResults *CalleeAAR = nullptr, | ||
| 273 | bool InsertLifetime = true, | ||
| 274 | Function *ForwardVarArgsTo = nullptr); | ||
| 275 | |||
| 276 | /// Clones a loop \p OrigLoop.  Returns the loop and the blocks in \p | ||
| 277 | /// Blocks. | ||
| 278 | /// | ||
| 279 | /// Updates LoopInfo and DominatorTree assuming the loop is dominated by block | ||
| 280 | /// \p LoopDomBB.  Insert the new blocks before block specified in \p Before. | ||
| 281 | /// Note: Only innermost loops are supported. | ||
| 282 | Loop *cloneLoopWithPreheader(BasicBlock *Before, BasicBlock *LoopDomBB, | ||
| 283 | Loop *OrigLoop, ValueToValueMapTy &VMap, | ||
| 284 | const Twine &NameSuffix, LoopInfo *LI, | ||
| 285 |                              DominatorTree *DT, | ||
| 286 | SmallVectorImpl<BasicBlock *> &Blocks); | ||
| 287 | |||
| 288 | /// Remaps instructions in \p Blocks using the mapping in \p VMap. | ||
| 289 | void remapInstructionsInBlocks(const SmallVectorImpl<BasicBlock *> &Blocks, | ||
| 290 | ValueToValueMapTy &VMap); | ||
| 291 | |||
| 292 | /// Split edge between BB and PredBB and duplicate all non-Phi instructions | ||
| 293 | /// from BB between its beginning and the StopAt instruction into the split | ||
| 294 | /// block. Phi nodes are not duplicated, but their uses are handled correctly: | ||
| 295 | /// we replace them with the uses of corresponding Phi inputs. ValueMapping | ||
| 296 | /// is used to map the original instructions from BB to their newly-created | ||
| 297 | /// copies. Returns the split block. | ||
| 298 | BasicBlock *DuplicateInstructionsInSplitBetween(BasicBlock *BB, | ||
| 299 |                                                 BasicBlock *PredBB, | ||
| 300 |                                                 Instruction *StopAt, | ||
| 301 |                                                 ValueToValueMapTy &ValueMapping, | ||
| 302 | DomTreeUpdater &DTU); | ||
| 303 | |||
| 304 | /// Updates profile information by adjusting the entry count by adding | ||
| 305 | /// EntryDelta then scaling callsite information by the new count divided by the | ||
| 306 | /// old count. VMap is used during inlinng to also update the new clone | ||
| 307 | void updateProfileCallee( | ||
| 308 | Function *Callee, int64_t EntryDelta, | ||
| 309 | const ValueMap<const Value *, WeakTrackingVH> *VMap = nullptr); | ||
| 310 | |||
| 311 | /// Find the 'llvm.experimental.noalias.scope.decl' intrinsics in the specified | ||
| 312 | /// basic blocks and extract their scope. These are candidates for duplication | ||
| 313 | /// when cloning. | ||
| 314 | void identifyNoAliasScopesToClone( | ||
| 315 | ArrayRef<BasicBlock *> BBs, SmallVectorImpl<MDNode *> &NoAliasDeclScopes); | ||
| 316 | |||
| 317 | /// Find the 'llvm.experimental.noalias.scope.decl' intrinsics in the specified | ||
| 318 | /// instruction range and extract their scope. These are candidates for | ||
| 319 | /// duplication when cloning. | ||
| 320 | void identifyNoAliasScopesToClone( | ||
| 321 | BasicBlock::iterator Start, BasicBlock::iterator End, | ||
| 322 | SmallVectorImpl<MDNode *> &NoAliasDeclScopes); | ||
| 323 | |||
| 324 | /// Duplicate the specified list of noalias decl scopes. | ||
| 325 | /// The 'Ext' string is added as an extension to the name. | ||
| 326 | /// Afterwards, the ClonedScopes contains the mapping of the original scope | ||
| 327 | /// MDNode onto the cloned scope. | ||
| 328 | /// Be aware that the cloned scopes are still part of the original scope domain. | ||
| 329 | void cloneNoAliasScopes( | ||
| 330 | ArrayRef<MDNode *> NoAliasDeclScopes, | ||
| 331 | DenseMap<MDNode *, MDNode *> &ClonedScopes, | ||
| 332 | StringRef Ext, LLVMContext &Context); | ||
| 333 | |||
| 334 | /// Adapt the metadata for the specified instruction according to the | ||
| 335 | /// provided mapping. This is normally used after cloning an instruction, when | ||
| 336 | /// some noalias scopes needed to be cloned. | ||
| 337 | void adaptNoAliasScopes( | ||
| 338 | llvm::Instruction *I, const DenseMap<MDNode *, MDNode *> &ClonedScopes, | ||
| 339 | LLVMContext &Context); | ||
| 340 | |||
| 341 | /// Clone the specified noalias decl scopes. Then adapt all instructions in the | ||
| 342 | /// NewBlocks basicblocks to the cloned versions. | ||
| 343 | /// 'Ext' will be added to the duplicate scope names. | ||
| 344 | void cloneAndAdaptNoAliasScopes(ArrayRef<MDNode *> NoAliasDeclScopes, | ||
| 345 | ArrayRef<BasicBlock *> NewBlocks, | ||
| 346 | LLVMContext &Context, StringRef Ext); | ||
| 347 | |||
| 348 | /// Clone the specified noalias decl scopes. Then adapt all instructions in the | ||
| 349 | /// [IStart, IEnd] (IEnd included !) range to the cloned versions. 'Ext' will be | ||
| 350 | /// added to the duplicate scope names. | ||
| 351 | void cloneAndAdaptNoAliasScopes(ArrayRef<MDNode *> NoAliasDeclScopes, | ||
| 352 | Instruction *IStart, Instruction *IEnd, | ||
| 353 | LLVMContext &Context, StringRef Ext); | ||
| 354 | } // end namespace llvm | ||
| 355 | |||
| 356 | #endif // LLVM_TRANSFORMS_UTILS_CLONING_H |