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 DebugInfoFinder; |
37 | class DominatorTree; |
38 | class Function; |
39 | class Instruction; |
40 | class Loop; |
41 | class LoopInfo; |
42 | class Module; |
43 | class ProfileSummaryInfo; |
44 | class ReturnInst; |
45 | class DomTreeUpdater; |
46 | |
47 | /// Return an exact copy of the specified module |
48 | std::unique_ptr<Module> CloneModule(const Module &M); |
49 | std::unique_ptr<Module> CloneModule(const Module &M, ValueToValueMapTy &VMap); |
50 | |
51 | /// Return a copy of the specified module. The ShouldCloneDefinition function |
52 | /// controls whether a specific GlobalValue's definition is cloned. If the |
53 | /// function returns false, the module copy will contain an external reference |
54 | /// in place of the global definition. |
55 | std::unique_ptr<Module> |
56 | CloneModule(const Module &M, ValueToValueMapTy &VMap, |
57 | function_ref<bool(const GlobalValue *)> ShouldCloneDefinition); |
58 | |
59 | /// This struct can be used to capture information about code |
60 | /// being cloned, while it is being cloned. |
61 | struct ClonedCodeInfo { |
62 | /// This is set to true if the cloned code contains a normal call instruction. |
63 | bool ContainsCalls = false; |
64 | |
65 | /// This is set to true if there is memprof related metadata (memprof or |
66 | /// callsite metadata) in the cloned code. |
67 | bool ContainsMemProfMetadata = 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 | /// Like VMap, but maps only unsimplified instructions. Values in the map |
80 | /// may be dangling, it is only intended to be used via isSimplified(), to |
81 | /// check whether the main VMap mapping involves simplification or not. |
82 | DenseMap<const Value *, const Value *> OrigVMap; |
83 | |
84 | ClonedCodeInfo() = default; |
85 | |
86 | bool isSimplified(const Value *From, const Value *To) const { |
87 | return OrigVMap.lookup(Val: From) != To; |
88 | } |
89 | }; |
90 | |
91 | /// Return a copy of the specified basic block, but without |
92 | /// embedding the block into a particular function. The block returned is an |
93 | /// exact copy of the specified basic block, without any remapping having been |
94 | /// performed. Because of this, this is only suitable for applications where |
95 | /// the basic block will be inserted into the same function that it was cloned |
96 | /// from (loop unrolling would use this, for example). |
97 | /// |
98 | /// Also, note that this function makes a direct copy of the basic block, and |
99 | /// can thus produce illegal LLVM code. In particular, it will copy any PHI |
100 | /// nodes from the original block, even though there are no predecessors for the |
101 | /// newly cloned block (thus, phi nodes will have to be updated). Also, this |
102 | /// block will branch to the old successors of the original block: these |
103 | /// successors will have to have any PHI nodes updated to account for the new |
104 | /// incoming edges. |
105 | /// |
106 | /// The correlation between instructions in the source and result basic blocks |
107 | /// is recorded in the VMap map. |
108 | /// |
109 | /// If you have a particular suffix you'd like to use to add to any cloned |
110 | /// names, specify it as the optional third parameter. |
111 | /// |
112 | /// If you would like the basic block to be auto-inserted into the end of a |
113 | /// function, you can specify it as the optional fourth parameter. |
114 | /// |
115 | /// If you would like to collect additional information about the cloned |
116 | /// function, you can specify a ClonedCodeInfo object with the optional fifth |
117 | /// parameter. |
118 | BasicBlock *CloneBasicBlock(const BasicBlock *BB, ValueToValueMapTy &VMap, |
119 | const Twine &NameSuffix = "" , Function *F = nullptr, |
120 | ClonedCodeInfo *CodeInfo = nullptr, |
121 | DebugInfoFinder *DIFinder = nullptr); |
122 | |
123 | /// Return a copy of the specified function and add it to that |
124 | /// function's module. Also, any references specified in the VMap are changed |
125 | /// to refer to their mapped value instead of the original one. If any of the |
126 | /// arguments to the function are in the VMap, the arguments are deleted from |
127 | /// the resultant function. The VMap is updated to include mappings from all of |
128 | /// the instructions and basicblocks in the function from their old to new |
129 | /// values. The final argument captures information about the cloned code if |
130 | /// non-null. |
131 | /// |
132 | /// \pre VMap contains no non-identity GlobalValue mappings. |
133 | /// |
134 | Function *CloneFunction(Function *F, ValueToValueMapTy &VMap, |
135 | ClonedCodeInfo *CodeInfo = nullptr); |
136 | |
137 | enum class CloneFunctionChangeType { |
138 | LocalChangesOnly, |
139 | GlobalChanges, |
140 | DifferentModule, |
141 | ClonedModule, |
142 | }; |
143 | |
144 | /// Clone OldFunc into NewFunc, transforming the old arguments into references |
145 | /// to VMap values. Note that if NewFunc already has basic blocks, the ones |
146 | /// cloned into it will be added to the end of the function. This function |
147 | /// fills in a list of return instructions, and can optionally remap types |
148 | /// and/or append the specified suffix to all values cloned. |
149 | /// |
150 | /// If \p Changes is \a CloneFunctionChangeType::LocalChangesOnly, VMap is |
151 | /// required to contain no non-identity GlobalValue mappings. Otherwise, |
152 | /// referenced metadata will be cloned. |
153 | /// |
154 | /// If \p Changes is less than \a CloneFunctionChangeType::DifferentModule |
155 | /// indicating cloning into the same module (even if it's LocalChangesOnly), if |
156 | /// debug info metadata transitively references a \a DISubprogram, it will be |
157 | /// cloned, effectively upgrading \p Changes to GlobalChanges while suppressing |
158 | /// cloning of types and compile units. |
159 | /// |
160 | /// If \p Changes is \a CloneFunctionChangeType::DifferentModule, the new |
161 | /// module's \c !llvm.dbg.cu will get updated with any newly created compile |
162 | /// units. (\a CloneFunctionChangeType::ClonedModule leaves that work for the |
163 | /// caller.) |
164 | /// |
165 | /// FIXME: Consider simplifying this function by splitting out \a |
166 | /// CloneFunctionMetadataInto() and expecting / updating callers to call it |
167 | /// first when / how it's needed. |
168 | void CloneFunctionInto(Function *NewFunc, const Function *OldFunc, |
169 | ValueToValueMapTy &VMap, CloneFunctionChangeType Changes, |
170 | SmallVectorImpl<ReturnInst *> &Returns, |
171 | const char *NameSuffix = "" , |
172 | ClonedCodeInfo *CodeInfo = nullptr, |
173 | ValueMapTypeRemapper *TypeMapper = nullptr, |
174 | ValueMaterializer *Materializer = nullptr); |
175 | |
176 | void CloneAndPruneIntoFromInst(Function *NewFunc, const Function *OldFunc, |
177 | const Instruction *StartingInst, |
178 | ValueToValueMapTy &VMap, bool ModuleLevelChanges, |
179 | SmallVectorImpl<ReturnInst *> &Returns, |
180 | const char *NameSuffix = "" , |
181 | ClonedCodeInfo *CodeInfo = nullptr); |
182 | |
183 | /// This works exactly like CloneFunctionInto, |
184 | /// except that it does some simple constant prop and DCE on the fly. The |
185 | /// effect of this is to copy significantly less code in cases where (for |
186 | /// example) a function call with constant arguments is inlined, and those |
187 | /// constant arguments cause a significant amount of code in the callee to be |
188 | /// dead. Since this doesn't produce an exactly copy of the input, it can't be |
189 | /// used for things like CloneFunction or CloneModule. |
190 | /// |
191 | /// If ModuleLevelChanges is false, VMap contains no non-identity GlobalValue |
192 | /// mappings. |
193 | /// |
194 | void CloneAndPruneFunctionInto(Function *NewFunc, const Function *OldFunc, |
195 | ValueToValueMapTy &VMap, bool ModuleLevelChanges, |
196 | SmallVectorImpl<ReturnInst*> &Returns, |
197 | const char *NameSuffix = "" , |
198 | ClonedCodeInfo *CodeInfo = nullptr); |
199 | |
200 | /// This class captures the data input to the InlineFunction call, and records |
201 | /// the auxiliary results produced by it. |
202 | class InlineFunctionInfo { |
203 | public: |
204 | explicit InlineFunctionInfo( |
205 | function_ref<AssumptionCache &(Function &)> GetAssumptionCache = nullptr, |
206 | ProfileSummaryInfo *PSI = nullptr, |
207 | BlockFrequencyInfo *CallerBFI = nullptr, |
208 | BlockFrequencyInfo *CalleeBFI = nullptr, bool UpdateProfile = true) |
209 | : GetAssumptionCache(GetAssumptionCache), PSI(PSI), CallerBFI(CallerBFI), |
210 | CalleeBFI(CalleeBFI), UpdateProfile(UpdateProfile) {} |
211 | |
212 | /// If non-null, InlineFunction will update the callgraph to reflect the |
213 | /// changes it makes. |
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. |
263 | /// |
264 | /// The callee's function attributes are merged into the callers' if |
265 | /// MergeAttributes is set to true. |
266 | InlineResult InlineFunction(CallBase &CB, InlineFunctionInfo &IFI, |
267 | bool MergeAttributes = false, |
268 | AAResults *CalleeAAR = nullptr, |
269 | bool InsertLifetime = true, |
270 | Function *ForwardVarArgsTo = nullptr); |
271 | |
272 | /// Clones a loop \p OrigLoop. Returns the loop and the blocks in \p |
273 | /// Blocks. |
274 | /// |
275 | /// Updates LoopInfo and DominatorTree assuming the loop is dominated by block |
276 | /// \p LoopDomBB. Insert the new blocks before block specified in \p Before. |
277 | /// Note: Only innermost loops are supported. |
278 | Loop *(BasicBlock *Before, BasicBlock *LoopDomBB, |
279 | Loop *OrigLoop, ValueToValueMapTy &VMap, |
280 | const Twine &NameSuffix, LoopInfo *LI, |
281 | DominatorTree *DT, |
282 | SmallVectorImpl<BasicBlock *> &Blocks); |
283 | |
284 | /// Remaps instructions in \p Blocks using the mapping in \p VMap. |
285 | void remapInstructionsInBlocks(ArrayRef<BasicBlock *> Blocks, |
286 | ValueToValueMapTy &VMap); |
287 | |
288 | /// Split edge between BB and PredBB and duplicate all non-Phi instructions |
289 | /// from BB between its beginning and the StopAt instruction into the split |
290 | /// block. Phi nodes are not duplicated, but their uses are handled correctly: |
291 | /// we replace them with the uses of corresponding Phi inputs. ValueMapping |
292 | /// is used to map the original instructions from BB to their newly-created |
293 | /// copies. Returns the split block. |
294 | BasicBlock *DuplicateInstructionsInSplitBetween(BasicBlock *BB, |
295 | BasicBlock *PredBB, |
296 | Instruction *StopAt, |
297 | ValueToValueMapTy &ValueMapping, |
298 | DomTreeUpdater &DTU); |
299 | |
300 | /// Updates profile information by adjusting the entry count by adding |
301 | /// EntryDelta then scaling callsite information by the new count divided by the |
302 | /// old count. VMap is used during inlinng to also update the new clone |
303 | void updateProfileCallee( |
304 | Function *Callee, int64_t EntryDelta, |
305 | const ValueMap<const Value *, WeakTrackingVH> *VMap = nullptr); |
306 | |
307 | /// Find the 'llvm.experimental.noalias.scope.decl' intrinsics in the specified |
308 | /// basic blocks and extract their scope. These are candidates for duplication |
309 | /// when cloning. |
310 | void identifyNoAliasScopesToClone( |
311 | ArrayRef<BasicBlock *> BBs, SmallVectorImpl<MDNode *> &NoAliasDeclScopes); |
312 | |
313 | /// Find the 'llvm.experimental.noalias.scope.decl' intrinsics in the specified |
314 | /// instruction range and extract their scope. These are candidates for |
315 | /// duplication when cloning. |
316 | void identifyNoAliasScopesToClone( |
317 | BasicBlock::iterator Start, BasicBlock::iterator End, |
318 | SmallVectorImpl<MDNode *> &NoAliasDeclScopes); |
319 | |
320 | /// Duplicate the specified list of noalias decl scopes. |
321 | /// The 'Ext' string is added as an extension to the name. |
322 | /// Afterwards, the ClonedScopes contains the mapping of the original scope |
323 | /// MDNode onto the cloned scope. |
324 | /// Be aware that the cloned scopes are still part of the original scope domain. |
325 | void cloneNoAliasScopes( |
326 | ArrayRef<MDNode *> NoAliasDeclScopes, |
327 | DenseMap<MDNode *, MDNode *> &ClonedScopes, |
328 | StringRef Ext, LLVMContext &Context); |
329 | |
330 | /// Adapt the metadata for the specified instruction according to the |
331 | /// provided mapping. This is normally used after cloning an instruction, when |
332 | /// some noalias scopes needed to be cloned. |
333 | void adaptNoAliasScopes( |
334 | llvm::Instruction *I, const DenseMap<MDNode *, MDNode *> &ClonedScopes, |
335 | LLVMContext &Context); |
336 | |
337 | /// Clone the specified noalias decl scopes. Then adapt all instructions in the |
338 | /// NewBlocks basicblocks to the cloned versions. |
339 | /// 'Ext' will be added to the duplicate scope names. |
340 | void cloneAndAdaptNoAliasScopes(ArrayRef<MDNode *> NoAliasDeclScopes, |
341 | ArrayRef<BasicBlock *> NewBlocks, |
342 | LLVMContext &Context, StringRef Ext); |
343 | |
344 | /// Clone the specified noalias decl scopes. Then adapt all instructions in the |
345 | /// [IStart, IEnd] (IEnd included !) range to the cloned versions. 'Ext' will be |
346 | /// added to the duplicate scope names. |
347 | void cloneAndAdaptNoAliasScopes(ArrayRef<MDNode *> NoAliasDeclScopes, |
348 | Instruction *IStart, Instruction *IEnd, |
349 | LLVMContext &Context, StringRef Ext); |
350 | } // end namespace llvm |
351 | |
352 | #endif // LLVM_TRANSFORMS_UTILS_CLONING_H |
353 | |