1 //===-- Internalize.cpp - Mark functions internal -------------------------===//
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 pass loops over all of the functions and variables in the input module.
10 // If the function or variable does not need to be preserved according to the
11 // client supplied callback, it is marked as internal.
12 //
13 // This transformation would not be legal in a regular compilation, but it gets
14 // extra information from the linker about what is safe.
15 //
16 // For example: Internalizing a function with external linkage. Only if we are
17 // told it is only used from within this module, it is safe to do it.
18 //
19 //===----------------------------------------------------------------------===//
20 
21 #include "llvm/Transforms/IPO/Internalize.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/ADT/StringSet.h"
25 #include "llvm/ADT/Triple.h"
26 #include "llvm/Analysis/CallGraph.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/InitializePasses.h"
29 #include "llvm/Pass.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/LineIterator.h"
33 #include "llvm/Support/MemoryBuffer.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Transforms/IPO.h"
36 #include "llvm/Transforms/Utils/GlobalStatus.h"
37 #include "llvm/Transforms/Utils/ModuleUtils.h"
38 using namespace llvm;
39 
40 #define DEBUG_TYPE "internalize"
41 
42 STATISTIC(NumAliases, "Number of aliases internalized");
43 STATISTIC(NumFunctions, "Number of functions internalized");
44 STATISTIC(NumGlobals, "Number of global vars internalized");
45 
46 // APIFile - A file which contains a list of symbols that should not be marked
47 // external.
48 static cl::opt<std::string>
49     APIFile("internalize-public-api-file", cl::value_desc("filename"),
50             cl::desc("A file containing list of symbol names to preserve"));
51 
52 // APIList - A list of symbols that should not be marked internal.
53 static cl::list<std::string>
54     APIList("internalize-public-api-list", cl::value_desc("list"),
55             cl::desc("A list of symbol names to preserve"), cl::CommaSeparated);
56 
57 namespace {
58 // Helper to load an API list to preserve from file and expose it as a functor
59 // for internalization.
60 class PreserveAPIList {
61 public:
62   PreserveAPIList() {
63     if (!APIFile.empty())
64       LoadFile(APIFile);
65     ExternalNames.insert(APIList.begin(), APIList.end());
66   }
67 
68   bool operator()(const GlobalValue &GV) {
69     return ExternalNames.count(GV.getName());
70   }
71 
72 private:
73   // Contains the set of symbols loaded from file
74   StringSet<> ExternalNames;
75 
76   void LoadFile(StringRef Filename) {
77     // Load the APIFile...
78     ErrorOr<std::unique_ptr<MemoryBuffer>> Buf =
79         MemoryBuffer::getFile(Filename);
80     if (!Buf) {
81       errs() << "WARNING: Internalize couldn't load file '" << Filename
82              << "'! Continuing as if it's empty.\n";
83       return; // Just continue as if the file were empty
84     }
85     for (line_iterator I(*Buf->get(), true), E; I != E; ++I)
86       ExternalNames.insert(*I);
87   }
88 };
89 } // end anonymous namespace
90 
91 bool InternalizePass::shouldPreserveGV(const GlobalValue &GV) {
92   // Function must be defined here
93   if (GV.isDeclaration())
94     return true;
95 
96   // Available externally is really just a "declaration with a body".
97   if (GV.hasAvailableExternallyLinkage())
98     return true;
99 
100   // Assume that dllexported symbols are referenced elsewhere
101   if (GV.hasDLLExportStorageClass())
102     return true;
103 
104   // Already local, has nothing to do.
105   if (GV.hasLocalLinkage())
106     return false;
107 
108   // Check some special cases
109   if (AlwaysPreserved.count(GV.getName()))
110     return true;
111 
112   return MustPreserveGV(GV);
113 }
114 
115 bool InternalizePass::maybeInternalize(
116     GlobalValue &GV, DenseMap<const Comdat *, ComdatInfo> &ComdatMap) {
117   SmallString<0> ComdatName;
118   if (Comdat *C = GV.getComdat()) {
119     // For GlobalAlias, C is the aliasee object's comdat which may have been
120     // redirected. So ComdatMap may not contain C.
121     if (ComdatMap.lookup(C).External)
122       return false;
123 
124     if (auto *GO = dyn_cast<GlobalObject>(&GV)) {
125       // If a comdat with one member is not externally visible, we can drop it.
126       // Otherwise, the comdat can be used to establish dependencies among the
127       // group of sections. Thus we have to keep the comdat but switch it to
128       // noduplicates.
129       // Note: noduplicates is not necessary for COFF. wasm doesn't support
130       // noduplicates.
131       ComdatInfo &Info = ComdatMap.find(C)->second;
132       if (Info.Size == 1)
133         GO->setComdat(nullptr);
134       else if (!IsWasm)
135         C->setSelectionKind(Comdat::NoDuplicates);
136     }
137 
138     if (GV.hasLocalLinkage())
139       return false;
140   } else {
141     if (GV.hasLocalLinkage())
142       return false;
143 
144     if (shouldPreserveGV(GV))
145       return false;
146   }
147 
148   GV.setVisibility(GlobalValue::DefaultVisibility);
149   GV.setLinkage(GlobalValue::InternalLinkage);
150   return true;
151 }
152 
153 // If GV is part of a comdat and is externally visible, update the comdat size
154 // and keep track of its comdat so that we don't internalize any of its members.
155 void InternalizePass::checkComdat(
156     GlobalValue &GV, DenseMap<const Comdat *, ComdatInfo> &ComdatMap) {
157   Comdat *C = GV.getComdat();
158   if (!C)
159     return;
160 
161   ComdatInfo &Info = ComdatMap.try_emplace(C).first->second;
162   ++Info.Size;
163   if (shouldPreserveGV(GV))
164     Info.External = true;
165 }
166 
167 bool InternalizePass::internalizeModule(Module &M, CallGraph *CG) {
168   bool Changed = false;
169   CallGraphNode *ExternalNode = CG ? CG->getExternalCallingNode() : nullptr;
170 
171   SmallVector<GlobalValue *, 4> Used;
172   collectUsedGlobalVariables(M, Used, false);
173 
174   // Collect comdat size and visiblity information for the module.
175   DenseMap<const Comdat *, ComdatInfo> ComdatMap;
176   if (!M.getComdatSymbolTable().empty()) {
177     for (Function &F : M)
178       checkComdat(F, ComdatMap);
179     for (GlobalVariable &GV : M.globals())
180       checkComdat(GV, ComdatMap);
181     for (GlobalAlias &GA : M.aliases())
182       checkComdat(GA, ComdatMap);
183   }
184 
185   // We must assume that globals in llvm.used have a reference that not even
186   // the linker can see, so we don't internalize them.
187   // For llvm.compiler.used the situation is a bit fuzzy. The assembler and
188   // linker can drop those symbols. If this pass is running as part of LTO,
189   // one might think that it could just drop llvm.compiler.used. The problem
190   // is that even in LTO llvm doesn't see every reference. For example,
191   // we don't see references from function local inline assembly. To be
192   // conservative, we internalize symbols in llvm.compiler.used, but we
193   // keep llvm.compiler.used so that the symbol is not deleted by llvm.
194   for (GlobalValue *V : Used) {
195     AlwaysPreserved.insert(V->getName());
196   }
197 
198   // Mark all functions not in the api as internal.
199   IsWasm = Triple(M.getTargetTriple()).isOSBinFormatWasm();
200   for (Function &I : M) {
201     if (!maybeInternalize(I, ComdatMap))
202       continue;
203     Changed = true;
204 
205     if (ExternalNode)
206       // Remove a callgraph edge from the external node to this function.
207       ExternalNode->removeOneAbstractEdgeTo((*CG)[&I]);
208 
209     ++NumFunctions;
210     LLVM_DEBUG(dbgs() << "Internalizing func " << I.getName() << "\n");
211   }
212 
213   // Never internalize the llvm.used symbol.  It is used to implement
214   // attribute((used)).
215   // FIXME: Shouldn't this just filter on llvm.metadata section??
216   AlwaysPreserved.insert("llvm.used");
217   AlwaysPreserved.insert("llvm.compiler.used");
218 
219   // Never internalize anchors used by the machine module info, else the info
220   // won't find them.  (see MachineModuleInfo.)
221   AlwaysPreserved.insert("llvm.global_ctors");
222   AlwaysPreserved.insert("llvm.global_dtors");
223   AlwaysPreserved.insert("llvm.global.annotations");
224 
225   // Never internalize symbols code-gen inserts.
226   // FIXME: We should probably add this (and the __stack_chk_guard) via some
227   // type of call-back in CodeGen.
228   AlwaysPreserved.insert("__stack_chk_fail");
229   if (Triple(M.getTargetTriple()).isOSAIX())
230     AlwaysPreserved.insert("__ssp_canary_word");
231   else
232     AlwaysPreserved.insert("__stack_chk_guard");
233 
234   // Mark all global variables with initializers that are not in the api as
235   // internal as well.
236   for (auto &GV : M.globals()) {
237     if (!maybeInternalize(GV, ComdatMap))
238       continue;
239     Changed = true;
240 
241     ++NumGlobals;
242     LLVM_DEBUG(dbgs() << "Internalized gvar " << GV.getName() << "\n");
243   }
244 
245   // Mark all aliases that are not in the api as internal as well.
246   for (auto &GA : M.aliases()) {
247     if (!maybeInternalize(GA, ComdatMap))
248       continue;
249     Changed = true;
250 
251     ++NumAliases;
252     LLVM_DEBUG(dbgs() << "Internalized alias " << GA.getName() << "\n");
253   }
254 
255   return Changed;
256 }
257 
258 InternalizePass::InternalizePass() : MustPreserveGV(PreserveAPIList()) {}
259 
260 PreservedAnalyses InternalizePass::run(Module &M, ModuleAnalysisManager &AM) {
261   if (!internalizeModule(M, AM.getCachedResult<CallGraphAnalysis>(M)))
262     return PreservedAnalyses::all();
263 
264   PreservedAnalyses PA;
265   PA.preserve<CallGraphAnalysis>();
266   return PA;
267 }
268 
269 namespace {
270 class InternalizeLegacyPass : public ModulePass {
271   // Client supplied callback to control wheter a symbol must be preserved.
272   std::function<bool(const GlobalValue &)> MustPreserveGV;
273 
274 public:
275   static char ID; // Pass identification, replacement for typeid
276 
277   InternalizeLegacyPass() : ModulePass(ID), MustPreserveGV(PreserveAPIList()) {}
278 
279   InternalizeLegacyPass(std::function<bool(const GlobalValue &)> MustPreserveGV)
280       : ModulePass(ID), MustPreserveGV(std::move(MustPreserveGV)) {
281     initializeInternalizeLegacyPassPass(*PassRegistry::getPassRegistry());
282   }
283 
284   bool runOnModule(Module &M) override {
285     if (skipModule(M))
286       return false;
287 
288     CallGraphWrapperPass *CGPass =
289         getAnalysisIfAvailable<CallGraphWrapperPass>();
290     CallGraph *CG = CGPass ? &CGPass->getCallGraph() : nullptr;
291     return internalizeModule(M, MustPreserveGV, CG);
292   }
293 
294   void getAnalysisUsage(AnalysisUsage &AU) const override {
295     AU.setPreservesCFG();
296     AU.addPreserved<CallGraphWrapperPass>();
297   }
298 };
299 }
300 
301 char InternalizeLegacyPass::ID = 0;
302 INITIALIZE_PASS(InternalizeLegacyPass, "internalize",
303                 "Internalize Global Symbols", false, false)
304 
305 ModulePass *llvm::createInternalizePass() {
306   return new InternalizeLegacyPass();
307 }
308 
309 ModulePass *llvm::createInternalizePass(
310     std::function<bool(const GlobalValue &)> MustPreserveGV) {
311   return new InternalizeLegacyPass(std::move(MustPreserveGV));
312 }
313