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/SmallString.h"
23#include "llvm/ADT/Statistic.h"
24#include "llvm/ADT/StringSet.h"
25#include "llvm/Analysis/CallGraph.h"
26#include "llvm/IR/Module.h"
27#include "llvm/Support/CommandLine.h"
28#include "llvm/Support/Debug.h"
29#include "llvm/Support/GlobPattern.h"
30#include "llvm/Support/LineIterator.h"
31#include "llvm/Support/MemoryBuffer.h"
32#include "llvm/Support/raw_ostream.h"
33#include "llvm/TargetParser/Triple.h"
34#include "llvm/Transforms/IPO.h"
35using namespace llvm;
36
37#define DEBUG_TYPE "internalize"
38
39STATISTIC(NumAliases, "Number of aliases internalized");
40STATISTIC(NumFunctions, "Number of functions internalized");
41STATISTIC(NumGlobals, "Number of global vars internalized");
42
43// APIFile - A file which contains a list of symbol glob patterns that should
44// not be marked external.
45static cl::opt<std::string>
46 APIFile("internalize-public-api-file", cl::value_desc("filename"),
47 cl::desc("A file containing list of symbol names to preserve"));
48
49// APIList - A list of symbol glob patterns that should not be marked internal.
50static cl::list<std::string>
51 APIList("internalize-public-api-list", cl::value_desc("list"),
52 cl::desc("A list of symbol names to preserve"), cl::CommaSeparated);
53
54namespace {
55// Helper to load an API list to preserve from file and expose it as a functor
56// for internalization.
57class PreserveAPIList {
58public:
59 PreserveAPIList() {
60 if (!APIFile.empty())
61 LoadFile(Filename: APIFile);
62 for (StringRef Pattern : APIList)
63 addGlob(Pattern);
64 }
65
66 bool operator()(const GlobalValue &GV) {
67 return llvm::any_of(
68 Range&: ExternalNames, P: [&](GlobPattern &GP) { return GP.match(S: GV.getName()); });
69 }
70
71private:
72 // Contains the set of symbols loaded from file
73 SmallVector<GlobPattern> ExternalNames;
74
75 void addGlob(StringRef Pattern) {
76 auto GlobOrErr = GlobPattern::create(Pat: Pattern);
77 if (!GlobOrErr) {
78 errs() << "WARNING: when loading pattern: '"
79 << toString(E: GlobOrErr.takeError()) << "' ignoring";
80 return;
81 }
82 ExternalNames.emplace_back(Args: std::move(*GlobOrErr));
83 }
84
85 void LoadFile(StringRef Filename) {
86 // Load the APIFile...
87 ErrorOr<std::unique_ptr<MemoryBuffer>> BufOrErr =
88 MemoryBuffer::getFile(Filename);
89 if (!BufOrErr) {
90 errs() << "WARNING: Internalize couldn't load file '" << Filename
91 << "'! Continuing as if it's empty.\n";
92 return; // Just continue as if the file were empty
93 }
94 Buf = std::move(*BufOrErr);
95 for (line_iterator I(*Buf, true), E; I != E; ++I)
96 addGlob(Pattern: *I);
97 }
98
99 std::shared_ptr<MemoryBuffer> Buf;
100};
101} // end anonymous namespace
102
103bool InternalizePass::shouldPreserveGV(const GlobalValue &GV) {
104 // Function must be defined here
105 if (GV.isDeclaration())
106 return true;
107
108 // Available externally is really just a "declaration with a body".
109 if (GV.hasAvailableExternallyLinkage())
110 return true;
111
112 // Assume that dllexported symbols are referenced elsewhere
113 if (GV.hasDLLExportStorageClass())
114 return true;
115
116 // As the name suggests, externally initialized variables need preserving as
117 // they would be initialized elsewhere externally.
118 if (const auto *G = dyn_cast<GlobalVariable>(Val: &GV))
119 if (G->isExternallyInitialized())
120 return true;
121
122 // Already local, has nothing to do.
123 if (GV.hasLocalLinkage())
124 return false;
125
126 // Check some special cases
127 if (AlwaysPreserved.count(Key: GV.getName()))
128 return true;
129
130 return MustPreserveGV(GV);
131}
132
133bool InternalizePass::maybeInternalize(
134 GlobalValue &GV, DenseMap<const Comdat *, ComdatInfo> &ComdatMap) {
135 SmallString<0> ComdatName;
136 if (Comdat *C = GV.getComdat()) {
137 // For GlobalAlias, C is the aliasee object's comdat which may have been
138 // redirected. So ComdatMap may not contain C.
139 if (ComdatMap.lookup(Val: C).External)
140 return false;
141
142 if (auto *GO = dyn_cast<GlobalObject>(Val: &GV)) {
143 // If a comdat with one member is not externally visible, we can drop it.
144 // Otherwise, the comdat can be used to establish dependencies among the
145 // group of sections. Thus we have to keep the comdat but switch it to
146 // nodeduplicate.
147 // Note: nodeduplicate is not necessary for COFF. wasm doesn't support
148 // nodeduplicate.
149 ComdatInfo &Info = ComdatMap.find(Val: C)->second;
150 if (Info.Size == 1)
151 GO->setComdat(nullptr);
152 else if (!IsWasm)
153 C->setSelectionKind(Comdat::NoDeduplicate);
154 }
155
156 if (GV.hasLocalLinkage())
157 return false;
158 } else {
159 if (GV.hasLocalLinkage())
160 return false;
161
162 if (shouldPreserveGV(GV))
163 return false;
164 }
165
166 GV.setVisibility(GlobalValue::DefaultVisibility);
167 GV.setLinkage(GlobalValue::InternalLinkage);
168 return true;
169}
170
171// If GV is part of a comdat and is externally visible, update the comdat size
172// and keep track of its comdat so that we don't internalize any of its members.
173void InternalizePass::checkComdat(
174 GlobalValue &GV, DenseMap<const Comdat *, ComdatInfo> &ComdatMap) {
175 Comdat *C = GV.getComdat();
176 if (!C)
177 return;
178
179 ComdatInfo &Info = ComdatMap.try_emplace(Key: C).first->second;
180 ++Info.Size;
181 if (shouldPreserveGV(GV))
182 Info.External = true;
183}
184
185bool InternalizePass::internalizeModule(Module &M) {
186 bool Changed = false;
187
188 SmallVector<GlobalValue *, 4> Used;
189 collectUsedGlobalVariables(M, Vec&: Used, CompilerUsed: false);
190
191 // Collect comdat size and visiblity information for the module.
192 DenseMap<const Comdat *, ComdatInfo> ComdatMap;
193 if (!M.getComdatSymbolTable().empty()) {
194 for (Function &F : M)
195 checkComdat(GV&: F, ComdatMap);
196 for (GlobalVariable &GV : M.globals())
197 checkComdat(GV, ComdatMap);
198 for (GlobalAlias &GA : M.aliases())
199 checkComdat(GV&: GA, ComdatMap);
200 }
201
202 // We must assume that globals in llvm.used have a reference that not even
203 // the linker can see, so we don't internalize them.
204 // For llvm.compiler.used the situation is a bit fuzzy. The assembler and
205 // linker can drop those symbols. If this pass is running as part of LTO,
206 // one might think that it could just drop llvm.compiler.used. The problem
207 // is that even in LTO llvm doesn't see every reference. For example,
208 // we don't see references from function local inline assembly. To be
209 // conservative, we internalize symbols in llvm.compiler.used, but we
210 // keep llvm.compiler.used so that the symbol is not deleted by llvm.
211 for (GlobalValue *V : Used) {
212 AlwaysPreserved.insert(key: V->getName());
213 }
214
215 // Never internalize the llvm.used symbol. It is used to implement
216 // attribute((used)).
217 // FIXME: Shouldn't this just filter on llvm.metadata section??
218 AlwaysPreserved.insert(key: "llvm.used");
219 AlwaysPreserved.insert(key: "llvm.compiler.used");
220
221 // Never internalize anchors used by the machine module info, else the info
222 // won't find them. (see MachineModuleInfo.)
223 AlwaysPreserved.insert(key: "llvm.global_ctors");
224 AlwaysPreserved.insert(key: "llvm.global_dtors");
225 AlwaysPreserved.insert(key: "llvm.global.annotations");
226
227 // Never internalize symbols code-gen inserts.
228 // FIXME: We should probably add this (and the __stack_chk_guard) via some
229 // type of call-back in CodeGen.
230 AlwaysPreserved.insert(key: "__stack_chk_fail");
231 if (Triple(M.getTargetTriple()).isOSAIX())
232 AlwaysPreserved.insert(key: "__ssp_canary_word");
233 else
234 AlwaysPreserved.insert(key: "__stack_chk_guard");
235
236 // Mark all functions not in the api as internal.
237 IsWasm = Triple(M.getTargetTriple()).isOSBinFormatWasm();
238 for (Function &I : M) {
239 if (!maybeInternalize(GV&: I, ComdatMap))
240 continue;
241 Changed = true;
242
243 ++NumFunctions;
244 LLVM_DEBUG(dbgs() << "Internalizing func " << I.getName() << "\n");
245 }
246
247 // Mark all global variables with initializers that are not in the api as
248 // internal as well.
249 for (auto &GV : M.globals()) {
250 if (!maybeInternalize(GV, ComdatMap))
251 continue;
252 Changed = true;
253
254 ++NumGlobals;
255 LLVM_DEBUG(dbgs() << "Internalized gvar " << GV.getName() << "\n");
256 }
257
258 // Mark all aliases that are not in the api as internal as well.
259 for (auto &GA : M.aliases()) {
260 if (!maybeInternalize(GV&: GA, ComdatMap))
261 continue;
262 Changed = true;
263
264 ++NumAliases;
265 LLVM_DEBUG(dbgs() << "Internalized alias " << GA.getName() << "\n");
266 }
267
268 return Changed;
269}
270
271InternalizePass::InternalizePass() : MustPreserveGV(PreserveAPIList()) {}
272
273PreservedAnalyses InternalizePass::run(Module &M, ModuleAnalysisManager &AM) {
274 if (!internalizeModule(M))
275 return PreservedAnalyses::all();
276
277 return PreservedAnalyses::none();
278}
279