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