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