Back to home page

EIC code displayed by LXR

 
 

    


File indexing completed on 2026-05-10 08:44:42

0001 //===- Cloning.h - Clone various parts of LLVM programs ---------*- C++ -*-===//
0002 //
0003 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
0004 // See https://llvm.org/LICENSE.txt for license information.
0005 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
0006 //
0007 //===----------------------------------------------------------------------===//
0008 //
0009 // This file defines various functions that are used to clone chunks of LLVM
0010 // code for various purposes.  This varies from copying whole modules into new
0011 // modules, to cloning functions with different arguments, to inlining
0012 // functions, to copying basic blocks to support loop unrolling or superblock
0013 // formation, etc.
0014 //
0015 //===----------------------------------------------------------------------===//
0016 
0017 #ifndef LLVM_TRANSFORMS_UTILS_CLONING_H
0018 #define LLVM_TRANSFORMS_UTILS_CLONING_H
0019 
0020 #include "llvm/ADT/SmallVector.h"
0021 #include "llvm/ADT/Twine.h"
0022 #include "llvm/Analysis/AssumptionCache.h"
0023 #include "llvm/Analysis/InlineCost.h"
0024 #include "llvm/IR/BasicBlock.h"
0025 #include "llvm/IR/ValueHandle.h"
0026 #include "llvm/Transforms/Utils/ValueMapper.h"
0027 #include <functional>
0028 #include <memory>
0029 #include <vector>
0030 
0031 namespace llvm {
0032 
0033 class AAResults;
0034 class AllocaInst;
0035 class BasicBlock;
0036 class BlockFrequencyInfo;
0037 class DebugInfoFinder;
0038 class DominatorTree;
0039 class Function;
0040 class Instruction;
0041 class Loop;
0042 class LoopInfo;
0043 class Module;
0044 class PGOContextualProfile;
0045 class ProfileSummaryInfo;
0046 class ReturnInst;
0047 class DomTreeUpdater;
0048 
0049 /// Return an exact copy of the specified module
0050 std::unique_ptr<Module> CloneModule(const Module &M);
0051 std::unique_ptr<Module> CloneModule(const Module &M, ValueToValueMapTy &VMap);
0052 
0053 /// Return a copy of the specified module. The ShouldCloneDefinition function
0054 /// controls whether a specific GlobalValue's definition is cloned. If the
0055 /// function returns false, the module copy will contain an external reference
0056 /// in place of the global definition.
0057 std::unique_ptr<Module>
0058 CloneModule(const Module &M, ValueToValueMapTy &VMap,
0059             function_ref<bool(const GlobalValue *)> ShouldCloneDefinition);
0060 
0061 /// This struct can be used to capture information about code
0062 /// being cloned, while it is being cloned.
0063 struct ClonedCodeInfo {
0064   /// This is set to true if the cloned code contains a normal call instruction.
0065   bool ContainsCalls = false;
0066 
0067   /// This is set to true if there is memprof related metadata (memprof or
0068   /// callsite metadata) in the cloned code.
0069   bool ContainsMemProfMetadata = false;
0070 
0071   /// This is set to true if the cloned code contains a 'dynamic' alloca.
0072   /// Dynamic allocas are allocas that are either not in the entry block or they
0073   /// are in the entry block but are not a constant size.
0074   bool ContainsDynamicAllocas = false;
0075 
0076   /// All cloned call sites that have operand bundles attached are appended to
0077   /// this vector.  This vector may contain nulls or undefs if some of the
0078   /// originally inserted callsites were DCE'ed after they were cloned.
0079   std::vector<WeakTrackingVH> OperandBundleCallSites;
0080 
0081   /// Like VMap, but maps only unsimplified instructions. Values in the map
0082   /// may be dangling, it is only intended to be used via isSimplified(), to
0083   /// check whether the main VMap mapping involves simplification or not.
0084   DenseMap<const Value *, const Value *> OrigVMap;
0085 
0086   ClonedCodeInfo() = default;
0087 
0088   bool isSimplified(const Value *From, const Value *To) const {
0089     return OrigVMap.lookup(From) != To;
0090   }
0091 };
0092 
0093 /// Return a copy of the specified basic block, but without
0094 /// embedding the block into a particular function.  The block returned is an
0095 /// exact copy of the specified basic block, without any remapping having been
0096 /// performed.  Because of this, this is only suitable for applications where
0097 /// the basic block will be inserted into the same function that it was cloned
0098 /// from (loop unrolling would use this, for example).
0099 ///
0100 /// Also, note that this function makes a direct copy of the basic block, and
0101 /// can thus produce illegal LLVM code.  In particular, it will copy any PHI
0102 /// nodes from the original block, even though there are no predecessors for the
0103 /// newly cloned block (thus, phi nodes will have to be updated).  Also, this
0104 /// block will branch to the old successors of the original block: these
0105 /// successors will have to have any PHI nodes updated to account for the new
0106 /// incoming edges.
0107 ///
0108 /// The correlation between instructions in the source and result basic blocks
0109 /// is recorded in the VMap map.
0110 ///
0111 /// If you have a particular suffix you'd like to use to add to any cloned
0112 /// names, specify it as the optional third parameter.
0113 ///
0114 /// If you would like the basic block to be auto-inserted into the end of a
0115 /// function, you can specify it as the optional fourth parameter.
0116 ///
0117 /// If you would like to collect additional information about the cloned
0118 /// function, you can specify a ClonedCodeInfo object with the optional fifth
0119 /// parameter.
0120 BasicBlock *CloneBasicBlock(const BasicBlock *BB, ValueToValueMapTy &VMap,
0121                             const Twine &NameSuffix = "", Function *F = nullptr,
0122                             ClonedCodeInfo *CodeInfo = nullptr);
0123 
0124 /// Return a copy of the specified function and add it to that
0125 /// function's module.  Also, any references specified in the VMap are changed
0126 /// to refer to their mapped value instead of the original one.  If any of the
0127 /// arguments to the function are in the VMap, the arguments are deleted from
0128 /// the resultant function.  The VMap is updated to include mappings from all of
0129 /// the instructions and basicblocks in the function from their old to new
0130 /// values.  The final argument captures information about the cloned code if
0131 /// non-null.
0132 ///
0133 /// \pre VMap contains no non-identity GlobalValue mappings.
0134 ///
0135 Function *CloneFunction(Function *F, ValueToValueMapTy &VMap,
0136                         ClonedCodeInfo *CodeInfo = nullptr);
0137 
0138 enum class CloneFunctionChangeType {
0139   LocalChangesOnly,
0140   GlobalChanges,
0141   DifferentModule,
0142   ClonedModule,
0143 };
0144 
0145 /// Clone OldFunc into NewFunc, transforming the old arguments into references
0146 /// to VMap values.  Note that if NewFunc already has basic blocks, the ones
0147 /// cloned into it will be added to the end of the function.  This function
0148 /// fills in a list of return instructions, and can optionally remap types
0149 /// and/or append the specified suffix to all values cloned.
0150 ///
0151 /// If \p Changes is \a CloneFunctionChangeType::LocalChangesOnly, VMap is
0152 /// required to contain no non-identity GlobalValue mappings. Otherwise,
0153 /// referenced metadata will be cloned.
0154 ///
0155 /// If \p Changes is less than \a CloneFunctionChangeType::DifferentModule
0156 /// indicating cloning into the same module (even if it's LocalChangesOnly), if
0157 /// debug info metadata transitively references a \a DISubprogram, it will be
0158 /// cloned, effectively upgrading \p Changes to GlobalChanges while suppressing
0159 /// cloning of types and compile units.
0160 ///
0161 /// If \p Changes is \a CloneFunctionChangeType::DifferentModule, the new
0162 /// module's \c !llvm.dbg.cu will get updated with any newly created compile
0163 /// units. (\a CloneFunctionChangeType::ClonedModule leaves that work for the
0164 /// caller.)
0165 ///
0166 /// FIXME: Consider simplifying this function by splitting out \a
0167 /// CloneFunctionMetadataInto() and expecting / updating callers to call it
0168 /// first when / how it's needed.
0169 void CloneFunctionInto(Function *NewFunc, const Function *OldFunc,
0170                        ValueToValueMapTy &VMap, CloneFunctionChangeType Changes,
0171                        SmallVectorImpl<ReturnInst *> &Returns,
0172                        const char *NameSuffix = "",
0173                        ClonedCodeInfo *CodeInfo = nullptr,
0174                        ValueMapTypeRemapper *TypeMapper = nullptr,
0175                        ValueMaterializer *Materializer = nullptr);
0176 
0177 /// Clone OldFunc's attributes into NewFunc, transforming values based on the
0178 /// mappings in VMap.
0179 void CloneFunctionAttributesInto(Function *NewFunc, const Function *OldFunc,
0180                                  ValueToValueMapTy &VMap,
0181                                  bool ModuleLevelChanges,
0182                                  ValueMapTypeRemapper *TypeMapper = nullptr,
0183                                  ValueMaterializer *Materializer = nullptr);
0184 
0185 /// Clone OldFunc's metadata into NewFunc.
0186 ///
0187 /// The caller is expected to populate \p VMap beforehand and set an appropriate
0188 /// \p RemapFlag. Subprograms/CUs/types that were already mapped to themselves
0189 /// won't be duplicated.
0190 ///
0191 /// NOTE: This function doesn't clone !llvm.dbg.cu when cloning into a different
0192 /// module. Use CloneFunctionInto for that behavior.
0193 void CloneFunctionMetadataInto(Function &NewFunc, const Function &OldFunc,
0194                                ValueToValueMapTy &VMap, RemapFlags RemapFlag,
0195                                ValueMapTypeRemapper *TypeMapper = nullptr,
0196                                ValueMaterializer *Materializer = nullptr,
0197                                const MetadataSetTy *IdentityMD = nullptr);
0198 
0199 /// Clone OldFunc's body into NewFunc.
0200 void CloneFunctionBodyInto(Function &NewFunc, const Function &OldFunc,
0201                            ValueToValueMapTy &VMap, RemapFlags RemapFlag,
0202                            SmallVectorImpl<ReturnInst *> &Returns,
0203                            const char *NameSuffix = "",
0204                            ClonedCodeInfo *CodeInfo = nullptr,
0205                            ValueMapTypeRemapper *TypeMapper = nullptr,
0206                            ValueMaterializer *Materializer = nullptr,
0207                            const MetadataSetTy *IdentityMD = nullptr);
0208 
0209 void CloneAndPruneIntoFromInst(Function *NewFunc, const Function *OldFunc,
0210                                const Instruction *StartingInst,
0211                                ValueToValueMapTy &VMap, bool ModuleLevelChanges,
0212                                SmallVectorImpl<ReturnInst *> &Returns,
0213                                const char *NameSuffix = "",
0214                                ClonedCodeInfo *CodeInfo = nullptr);
0215 
0216 /// This works exactly like CloneFunctionInto,
0217 /// except that it does some simple constant prop and DCE on the fly.  The
0218 /// effect of this is to copy significantly less code in cases where (for
0219 /// example) a function call with constant arguments is inlined, and those
0220 /// constant arguments cause a significant amount of code in the callee to be
0221 /// dead.  Since this doesn't produce an exactly copy of the input, it can't be
0222 /// used for things like CloneFunction or CloneModule.
0223 ///
0224 /// If ModuleLevelChanges is false, VMap contains no non-identity GlobalValue
0225 /// mappings.
0226 ///
0227 void CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc,
0228                                ValueToValueMapTy &VMap, bool ModuleLevelChanges,
0229                                SmallVectorImpl<ReturnInst*> &Returns,
0230                                const char *NameSuffix = "",
0231                                ClonedCodeInfo *CodeInfo = nullptr);
0232 
0233 /// Collect debug information such as types, compile units, and other
0234 /// subprograms that are reachable from \p F and can be considered global for
0235 /// the purposes of cloning (and hence not needing to be cloned).
0236 ///
0237 /// What debug information should be processed depends on \p Changes: when
0238 /// cloning into the same module we process \p F's subprogram and instructions;
0239 /// when into a cloned module, neither of those.
0240 ///
0241 /// Returns DISubprogram of the cloned function when cloning into the same
0242 /// module or nullptr otherwise.
0243 DISubprogram *CollectDebugInfoForCloning(const Function &F,
0244                                          CloneFunctionChangeType Changes,
0245                                          DebugInfoFinder &DIFinder);
0246 
0247 /// Based on \p Changes and \p DIFinder return debug info that needs to be
0248 /// identity mapped during Metadata cloning.
0249 ///
0250 /// NOTE: Such \a MetadataSetTy can be used by \a CloneFunction* to directly
0251 /// specify metadata that should be identity mapped (and hence not cloned). The
0252 /// metadata will be identity mapped in \a ValueToValueMapTy on first use. There
0253 /// are several reasons for doing it this way rather than eagerly identity
0254 /// mapping metadata nodes in a \a ValueMap:
0255 /// 1. Mapping metadata is not cheap, particularly because of tracking.
0256 /// 2. When cloning a Function we identity map lots of global module-level
0257 ///    metadata to avoid cloning it, while only a fraction of it is actually
0258 ///    used by the function. Mapping on first use is a lot faster for modules
0259 ///    with meaningful amount of debug info.
0260 /// 3. Eagerly identity mapping metadata makes it harder to cache module-level
0261 ///    data (e.g. a set of metadata nodes in a \a DICompileUnit).
0262 MetadataSetTy FindDebugInfoToIdentityMap(CloneFunctionChangeType Changes,
0263                                          DebugInfoFinder &DIFinder,
0264                                          DISubprogram *SPClonedWithinModule);
0265 
0266 /// This class captures the data input to the InlineFunction call, and records
0267 /// the auxiliary results produced by it.
0268 class InlineFunctionInfo {
0269 public:
0270   explicit InlineFunctionInfo(
0271       function_ref<AssumptionCache &(Function &)> GetAssumptionCache = nullptr,
0272       ProfileSummaryInfo *PSI = nullptr,
0273       BlockFrequencyInfo *CallerBFI = nullptr,
0274       BlockFrequencyInfo *CalleeBFI = nullptr, bool UpdateProfile = true)
0275       : GetAssumptionCache(GetAssumptionCache), PSI(PSI), CallerBFI(CallerBFI),
0276         CalleeBFI(CalleeBFI), UpdateProfile(UpdateProfile) {}
0277 
0278   /// If non-null, InlineFunction will update the callgraph to reflect the
0279   /// changes it makes.
0280   function_ref<AssumptionCache &(Function &)> GetAssumptionCache;
0281   ProfileSummaryInfo *PSI;
0282   BlockFrequencyInfo *CallerBFI, *CalleeBFI;
0283 
0284   /// InlineFunction fills this in with all static allocas that get copied into
0285   /// the caller.
0286   SmallVector<AllocaInst *, 4> StaticAllocas;
0287 
0288   /// InlineFunction fills this in with callsites that were inlined from the
0289   /// callee. This is only filled in if CG is non-null.
0290   SmallVector<WeakTrackingVH, 8> InlinedCalls;
0291 
0292   /// All of the new call sites inlined into the caller.
0293   ///
0294   /// 'InlineFunction' fills this in by scanning the inlined instructions, and
0295   /// only if CG is null. If CG is non-null, instead the value handle
0296   /// `InlinedCalls` above is used.
0297   SmallVector<CallBase *, 8> InlinedCallSites;
0298 
0299   /// Update profile for callee as well as cloned version. We need to do this
0300   /// for regular inlining, but not for inlining from sample profile loader.
0301   bool UpdateProfile;
0302 
0303   void reset() {
0304     StaticAllocas.clear();
0305     InlinedCalls.clear();
0306     InlinedCallSites.clear();
0307   }
0308 };
0309 
0310 /// This function inlines the called function into the basic
0311 /// block of the caller.  This returns false if it is not possible to inline
0312 /// this call.  The program is still in a well defined state if this occurs
0313 /// though.
0314 ///
0315 /// Note that this only does one level of inlining.  For example, if the
0316 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
0317 /// exists in the instruction stream.  Similarly this will inline a recursive
0318 /// function by one level.
0319 ///
0320 /// Note that while this routine is allowed to cleanup and optimize the
0321 /// *inlined* code to minimize the actual inserted code, it must not delete
0322 /// code in the caller as users of this routine may have pointers to
0323 /// instructions in the caller that need to remain stable.
0324 ///
0325 /// If ForwardVarArgsTo is passed, inlining a function with varargs is allowed
0326 /// and all varargs at the callsite will be passed to any calls to
0327 /// ForwardVarArgsTo. The caller of InlineFunction has to make sure any varargs
0328 /// are only used by ForwardVarArgsTo.
0329 ///
0330 /// The callee's function attributes are merged into the callers' if
0331 /// MergeAttributes is set to true.
0332 InlineResult InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
0333                             bool MergeAttributes = false,
0334                             AAResults *CalleeAAR = nullptr,
0335                             bool InsertLifetime = true,
0336                             Function *ForwardVarArgsTo = nullptr);
0337 
0338 /// Same as above, but it will update the contextual profile. If the contextual
0339 /// profile is invalid (i.e. not loaded because it is not present), it defaults
0340 /// to the behavior of the non-contextual profile updating variant above. This
0341 /// makes it easy to drop-in replace uses of the non-contextual overload.
0342 InlineResult InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
0343                             PGOContextualProfile &CtxProf,
0344                             bool MergeAttributes = false,
0345                             AAResults *CalleeAAR = nullptr,
0346                             bool InsertLifetime = true,
0347                             Function *ForwardVarArgsTo = nullptr);
0348 
0349 /// Clones a loop \p OrigLoop.  Returns the loop and the blocks in \p
0350 /// Blocks.
0351 ///
0352 /// Updates LoopInfo and DominatorTree assuming the loop is dominated by block
0353 /// \p LoopDomBB.  Insert the new blocks before block specified in \p Before.
0354 /// Note: Only innermost loops are supported.
0355 Loop *cloneLoopWithPreheader(BasicBlock *Before, BasicBlock *LoopDomBB,
0356                              Loop *OrigLoop, ValueToValueMapTy &VMap,
0357                              const Twine &NameSuffix, LoopInfo *LI,
0358                              DominatorTree *DT,
0359                              SmallVectorImpl<BasicBlock *> &Blocks);
0360 
0361 /// Remaps instructions in \p Blocks using the mapping in \p VMap.
0362 void remapInstructionsInBlocks(ArrayRef<BasicBlock *> Blocks,
0363                                ValueToValueMapTy &VMap);
0364 
0365 /// Split edge between BB and PredBB and duplicate all non-Phi instructions
0366 /// from BB between its beginning and the StopAt instruction into the split
0367 /// block. Phi nodes are not duplicated, but their uses are handled correctly:
0368 /// we replace them with the uses of corresponding Phi inputs. ValueMapping
0369 /// is used to map the original instructions from BB to their newly-created
0370 /// copies. Returns the split block.
0371 BasicBlock *DuplicateInstructionsInSplitBetween(BasicBlock *BB,
0372                                                 BasicBlock *PredBB,
0373                                                 Instruction *StopAt,
0374                                                 ValueToValueMapTy &ValueMapping,
0375                                                 DomTreeUpdater &DTU);
0376 
0377 /// Updates profile information by adjusting the entry count by adding
0378 /// EntryDelta then scaling callsite information by the new count divided by the
0379 /// old count. VMap is used during inlinng to also update the new clone
0380 void updateProfileCallee(
0381     Function *Callee, int64_t EntryDelta,
0382     const ValueMap<const Value *, WeakTrackingVH> *VMap = nullptr);
0383 
0384 /// Find the 'llvm.experimental.noalias.scope.decl' intrinsics in the specified
0385 /// basic blocks and extract their scope. These are candidates for duplication
0386 /// when cloning.
0387 void identifyNoAliasScopesToClone(
0388     ArrayRef<BasicBlock *> BBs, SmallVectorImpl<MDNode *> &NoAliasDeclScopes);
0389 
0390 /// Find the 'llvm.experimental.noalias.scope.decl' intrinsics in the specified
0391 /// instruction range and extract their scope. These are candidates for
0392 /// duplication when cloning.
0393 void identifyNoAliasScopesToClone(
0394     BasicBlock::iterator Start, BasicBlock::iterator End,
0395     SmallVectorImpl<MDNode *> &NoAliasDeclScopes);
0396 
0397 /// Duplicate the specified list of noalias decl scopes.
0398 /// The 'Ext' string is added as an extension to the name.
0399 /// Afterwards, the ClonedScopes contains the mapping of the original scope
0400 /// MDNode onto the cloned scope.
0401 /// Be aware that the cloned scopes are still part of the original scope domain.
0402 void cloneNoAliasScopes(
0403     ArrayRef<MDNode *> NoAliasDeclScopes,
0404     DenseMap<MDNode *, MDNode *> &ClonedScopes,
0405     StringRef Ext, LLVMContext &Context);
0406 
0407 /// Adapt the metadata for the specified instruction according to the
0408 /// provided mapping. This is normally used after cloning an instruction, when
0409 /// some noalias scopes needed to be cloned.
0410 void adaptNoAliasScopes(
0411     llvm::Instruction *I, const DenseMap<MDNode *, MDNode *> &ClonedScopes,
0412     LLVMContext &Context);
0413 
0414 /// Clone the specified noalias decl scopes. Then adapt all instructions in the
0415 /// NewBlocks basicblocks to the cloned versions.
0416 /// 'Ext' will be added to the duplicate scope names.
0417 void cloneAndAdaptNoAliasScopes(ArrayRef<MDNode *> NoAliasDeclScopes,
0418                                 ArrayRef<BasicBlock *> NewBlocks,
0419                                 LLVMContext &Context, StringRef Ext);
0420 
0421 /// Clone the specified noalias decl scopes. Then adapt all instructions in the
0422 /// [IStart, IEnd] (IEnd included !) range to the cloned versions. 'Ext' will be
0423 /// added to the duplicate scope names.
0424 void cloneAndAdaptNoAliasScopes(ArrayRef<MDNode *> NoAliasDeclScopes,
0425                                 Instruction *IStart, Instruction *IEnd,
0426                                 LLVMContext &Context, StringRef Ext);
0427 } // end namespace llvm
0428 
0429 #endif // LLVM_TRANSFORMS_UTILS_CLONING_H