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/AliasAnalysis.h" 23 #include "llvm/Analysis/AssumptionCache.h" 24 #include "llvm/Analysis/InlineCost.h" 25 #include "llvm/IR/CallSite.h" 26 #include "llvm/IR/ValueHandle.h" 27 #include "llvm/Transforms/Utils/ValueMapper.h" 28 #include <functional> 29 #include <memory> 30 #include <vector> 31 32 namespace llvm { 33 34 class AllocaInst; 35 class BasicBlock; 36 class BlockFrequencyInfo; 37 class CallInst; 38 class CallGraph; 39 class DebugInfoFinder; 40 class DominatorTree; 41 class Function; 42 class Instruction; 43 class InvokeInst; 44 class Loop; 45 class LoopInfo; 46 class Module; 47 class ProfileSummaryInfo; 48 class ReturnInst; 49 class DomTreeUpdater; 50 51 /// Return an exact copy of the specified module 52 std::unique_ptr<Module> CloneModule(const Module &M); 53 std::unique_ptr<Module> CloneModule(const Module &M, ValueToValueMapTy &VMap); 54 55 /// Return a copy of the specified module. The ShouldCloneDefinition function 56 /// controls whether a specific GlobalValue's definition is cloned. If the 57 /// function returns false, the module copy will contain an external reference 58 /// in place of the global definition. 59 std::unique_ptr<Module> 60 CloneModule(const Module &M, ValueToValueMapTy &VMap, 61 function_ref<bool(const GlobalValue *)> ShouldCloneDefinition); 62 63 /// This struct can be used to capture information about code 64 /// being cloned, while it is being cloned. 65 struct ClonedCodeInfo { 66 /// This is set to true if the cloned code contains a normal call instruction. 67 bool ContainsCalls = false; 68 69 /// This is set to true if the cloned code contains a 'dynamic' alloca. 70 /// Dynamic allocas are allocas that are either not in the entry block or they 71 /// are in the entry block but are not a constant size. 72 bool ContainsDynamicAllocas = false; 73 74 /// All cloned call sites that have operand bundles attached are appended to 75 /// this vector. This vector may contain nulls or undefs if some of the 76 /// originally inserted callsites were DCE'ed after they were cloned. 77 std::vector<WeakTrackingVH> OperandBundleCallSites; 78 79 ClonedCodeInfo() = default; 80 }; 81 82 /// Return a copy of the specified basic block, but without 83 /// embedding the block into a particular function. The block returned is an 84 /// exact copy of the specified basic block, without any remapping having been 85 /// performed. Because of this, this is only suitable for applications where 86 /// the basic block will be inserted into the same function that it was cloned 87 /// from (loop unrolling would use this, for example). 88 /// 89 /// Also, note that this function makes a direct copy of the basic block, and 90 /// can thus produce illegal LLVM code. In particular, it will copy any PHI 91 /// nodes from the original block, even though there are no predecessors for the 92 /// newly cloned block (thus, phi nodes will have to be updated). Also, this 93 /// block will branch to the old successors of the original block: these 94 /// successors will have to have any PHI nodes updated to account for the new 95 /// incoming edges. 96 /// 97 /// The correlation between instructions in the source and result basic blocks 98 /// is recorded in the VMap map. 99 /// 100 /// If you have a particular suffix you'd like to use to add to any cloned 101 /// names, specify it as the optional third parameter. 102 /// 103 /// If you would like the basic block to be auto-inserted into the end of a 104 /// function, you can specify it as the optional fourth parameter. 105 /// 106 /// If you would like to collect additional information about the cloned 107 /// function, you can specify a ClonedCodeInfo object with the optional fifth 108 /// parameter. 109 BasicBlock *CloneBasicBlock(const BasicBlock *BB, ValueToValueMapTy &VMap, 110 const Twine &NameSuffix = "", Function *F = nullptr, 111 ClonedCodeInfo *CodeInfo = nullptr, 112 DebugInfoFinder *DIFinder = nullptr); 113 114 /// Return a copy of the specified function and add it to that 115 /// function's module. Also, any references specified in the VMap are changed 116 /// to refer to their mapped value instead of the original one. If any of the 117 /// arguments to the function are in the VMap, the arguments are deleted from 118 /// the resultant function. The VMap is updated to include mappings from all of 119 /// the instructions and basicblocks in the function from their old to new 120 /// values. The final argument captures information about the cloned code if 121 /// non-null. 122 /// 123 /// VMap contains no non-identity GlobalValue mappings and debug info metadata 124 /// will not be cloned. 125 /// 126 Function *CloneFunction(Function *F, ValueToValueMapTy &VMap, 127 ClonedCodeInfo *CodeInfo = nullptr); 128 129 /// Clone OldFunc into NewFunc, transforming the old arguments into references 130 /// to VMap values. Note that if NewFunc already has basic blocks, the ones 131 /// cloned into it will be added to the end of the function. This function 132 /// fills in a list of return instructions, and can optionally remap types 133 /// and/or append the specified suffix to all values cloned. 134 /// 135 /// If ModuleLevelChanges is false, VMap contains no non-identity GlobalValue 136 /// mappings. 137 /// 138 void CloneFunctionInto(Function *NewFunc, const Function *OldFunc, 139 ValueToValueMapTy &VMap, bool ModuleLevelChanges, 140 SmallVectorImpl<ReturnInst*> &Returns, 141 const char *NameSuffix = "", 142 ClonedCodeInfo *CodeInfo = nullptr, 143 ValueMapTypeRemapper *TypeMapper = nullptr, 144 ValueMaterializer *Materializer = nullptr); 145 146 void CloneAndPruneIntoFromInst(Function *NewFunc, const Function *OldFunc, 147 const Instruction *StartingInst, 148 ValueToValueMapTy &VMap, bool ModuleLevelChanges, 149 SmallVectorImpl<ReturnInst *> &Returns, 150 const char *NameSuffix = "", 151 ClonedCodeInfo *CodeInfo = nullptr); 152 153 /// This works exactly like CloneFunctionInto, 154 /// except that it does some simple constant prop and DCE on the fly. The 155 /// effect of this is to copy significantly less code in cases where (for 156 /// example) a function call with constant arguments is inlined, and those 157 /// constant arguments cause a significant amount of code in the callee to be 158 /// dead. Since this doesn't produce an exactly copy of the input, it can't be 159 /// used for things like CloneFunction or CloneModule. 160 /// 161 /// If ModuleLevelChanges is false, VMap contains no non-identity GlobalValue 162 /// mappings. 163 /// 164 void CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc, 165 ValueToValueMapTy &VMap, bool ModuleLevelChanges, 166 SmallVectorImpl<ReturnInst*> &Returns, 167 const char *NameSuffix = "", 168 ClonedCodeInfo *CodeInfo = nullptr, 169 Instruction *TheCall = nullptr); 170 171 /// This class captures the data input to the InlineFunction call, and records 172 /// the auxiliary results produced by it. 173 class InlineFunctionInfo { 174 public: 175 explicit InlineFunctionInfo(CallGraph *cg = nullptr, 176 std::function<AssumptionCache &(Function &)> 177 *GetAssumptionCache = nullptr, 178 ProfileSummaryInfo *PSI = nullptr, 179 BlockFrequencyInfo *CallerBFI = nullptr, 180 BlockFrequencyInfo *CalleeBFI = nullptr) CG(cg)181 : CG(cg), GetAssumptionCache(GetAssumptionCache), PSI(PSI), 182 CallerBFI(CallerBFI), CalleeBFI(CalleeBFI) {} 183 184 /// If non-null, InlineFunction will update the callgraph to reflect the 185 /// changes it makes. 186 CallGraph *CG; 187 std::function<AssumptionCache &(Function &)> *GetAssumptionCache; 188 ProfileSummaryInfo *PSI; 189 BlockFrequencyInfo *CallerBFI, *CalleeBFI; 190 191 /// InlineFunction fills this in with all static allocas that get copied into 192 /// the caller. 193 SmallVector<AllocaInst *, 4> StaticAllocas; 194 195 /// InlineFunction fills this in with callsites that were inlined from the 196 /// callee. This is only filled in if CG is non-null. 197 SmallVector<WeakTrackingVH, 8> InlinedCalls; 198 199 /// All of the new call sites inlined into the caller. 200 /// 201 /// 'InlineFunction' fills this in by scanning the inlined instructions, and 202 /// only if CG is null. If CG is non-null, instead the value handle 203 /// `InlinedCalls` above is used. 204 SmallVector<CallSite, 8> InlinedCallSites; 205 reset()206 void reset() { 207 StaticAllocas.clear(); 208 InlinedCalls.clear(); 209 InlinedCallSites.clear(); 210 } 211 }; 212 213 /// This function inlines the called function into the basic 214 /// block of the caller. This returns false if it is not possible to inline 215 /// this call. The program is still in a well defined state if this occurs 216 /// though. 217 /// 218 /// Note that this only does one level of inlining. For example, if the 219 /// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 220 /// exists in the instruction stream. Similarly this will inline a recursive 221 /// function by one level. 222 /// 223 /// Note that while this routine is allowed to cleanup and optimize the 224 /// *inlined* code to minimize the actual inserted code, it must not delete 225 /// code in the caller as users of this routine may have pointers to 226 /// instructions in the caller that need to remain stable. 227 /// 228 /// If ForwardVarArgsTo is passed, inlining a function with varargs is allowed 229 /// and all varargs at the callsite will be passed to any calls to 230 /// ForwardVarArgsTo. The caller of InlineFunction has to make sure any varargs 231 /// are only used by ForwardVarArgsTo. 232 InlineResult InlineFunction(CallBase *CB, InlineFunctionInfo &IFI, 233 AAResults *CalleeAAR = nullptr, 234 bool InsertLifetime = true); 235 InlineResult InlineFunction(CallSite CS, InlineFunctionInfo &IFI, 236 AAResults *CalleeAAR = nullptr, 237 bool InsertLifetime = true, 238 Function *ForwardVarArgsTo = nullptr); 239 240 /// Clones a loop \p OrigLoop. Returns the loop and the blocks in \p 241 /// Blocks. 242 /// 243 /// Updates LoopInfo and DominatorTree assuming the loop is dominated by block 244 /// \p LoopDomBB. Insert the new blocks before block specified in \p Before. 245 /// Note: Only innermost loops are supported. 246 Loop *cloneLoopWithPreheader(BasicBlock *Before, BasicBlock *LoopDomBB, 247 Loop *OrigLoop, ValueToValueMapTy &VMap, 248 const Twine &NameSuffix, LoopInfo *LI, 249 DominatorTree *DT, 250 SmallVectorImpl<BasicBlock *> &Blocks); 251 252 /// Remaps instructions in \p Blocks using the mapping in \p VMap. 253 void remapInstructionsInBlocks(const SmallVectorImpl<BasicBlock *> &Blocks, 254 ValueToValueMapTy &VMap); 255 256 /// Split edge between BB and PredBB and duplicate all non-Phi instructions 257 /// from BB between its beginning and the StopAt instruction into the split 258 /// block. Phi nodes are not duplicated, but their uses are handled correctly: 259 /// we replace them with the uses of corresponding Phi inputs. ValueMapping 260 /// is used to map the original instructions from BB to their newly-created 261 /// copies. Returns the split block. 262 BasicBlock *DuplicateInstructionsInSplitBetween(BasicBlock *BB, 263 BasicBlock *PredBB, 264 Instruction *StopAt, 265 ValueToValueMapTy &ValueMapping, 266 DomTreeUpdater &DTU); 267 268 /// Updates profile information by adjusting the entry count by adding 269 /// entryDelta then scaling callsite information by the new count divided by the 270 /// old count. VMap is used during inlinng to also update the new clone 271 void updateProfileCallee( 272 Function *Callee, int64_t entryDelta, 273 const ValueMap<const Value *, WeakTrackingVH> *VMap = nullptr); 274 275 } // end namespace llvm 276 277 #endif // LLVM_TRANSFORMS_UTILS_CLONING_H 278