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/Statistic.h"
23#include "llvm/ADT/StringSet.h"
24#include "llvm/Analysis/CallGraph.h"
25#include "llvm/IR/Module.h"
26#include "llvm/Support/CommandLine.h"
27#include "llvm/Support/Debug.h"
28#include "llvm/Support/GlobPattern.h"
29#include "llvm/Support/LineIterator.h"
30#include "llvm/Support/MemoryBuffer.h"
31#include "llvm/Support/raw_ostream.h"
32#include "llvm/TargetParser/Triple.h"
33#include "llvm/Transforms/IPO.h"
34using namespace llvm;
35
36#define DEBUG_TYPE "internalize"
37
38STATISTIC(NumAliases, "Number of aliases internalized");
39STATISTIC(NumFunctions, "Number of functions internalized");
40STATISTIC(NumGlobals, "Number of global vars internalized");
41
42// APIFile - A file which contains a list of symbol glob patterns that should
43// not be marked external.
44static cl::opt<std::string>
45 APIFile("internalize-public-api-file", cl::value_desc("filename"),
46 cl::desc("A file containing list of symbol names to preserve"));
47
48// APIList - A list of symbol glob patterns that should not be marked internal.
49static cl::list<std::string>
50 APIList("internalize-public-api-list", cl::value_desc("list"),
51 cl::desc("A list of symbol names to preserve"), cl::CommaSeparated);
52
53namespace {
54// Helper to load an API list to preserve from file and expose it as a functor
55// for internalization.
56class PreserveAPIList {
57public:
58 PreserveAPIList() {
59 if (!APIFile.empty())
60 LoadFile(Filename: APIFile);
61 for (StringRef Pattern : APIList)
62 addGlob(Pattern);
63 }
64
65 bool operator()(const GlobalValue &GV) {
66 return llvm::any_of(
67 Range&: ExternalNames, P: [&](GlobPattern &GP) { return GP.match(S: GV.getName()); });
68 }
69
70private:
71 // Contains the set of symbols loaded from file
72 SmallVector<GlobPattern> ExternalNames;
73
74 void addGlob(StringRef Pattern) {
75 auto GlobOrErr = GlobPattern::create(Pat: Pattern);
76 if (!GlobOrErr) {
77 errs() << "WARNING: when loading pattern: '"
78 << toString(E: GlobOrErr.takeError()) << "' ignoring";
79 return;
80 }
81 ExternalNames.emplace_back(Args: std::move(*GlobOrErr));
82 }
83
84 void LoadFile(StringRef Filename) {
85 // Load the APIFile...
86 ErrorOr<std::unique_ptr<MemoryBuffer>> BufOrErr =
87 MemoryBuffer::getFile(Filename);
88 if (!BufOrErr) {
89 errs() << "WARNING: Internalize couldn't load file '" << Filename
90 << "'! Continuing as if it's empty.\n";
91 return; // Just continue as if the file were empty
92 }
93 Buf = std::move(*BufOrErr);
94 for (line_iterator I(*Buf, true), E; I != E; ++I)
95 addGlob(Pattern: *I);
96 }
97
98 std::shared_ptr<MemoryBuffer> Buf;
99};
100} // end anonymous namespace
101
102bool InternalizePass::shouldPreserveGV(const GlobalValue &GV) {
103 // Function must be defined here
104 if (GV.isDeclaration())
105 return true;
106
107 // Available externally is really just a "declaration with a body".
108 if (GV.hasAvailableExternallyLinkage())
109 return true;
110
111 // Assume that dllexported symbols are referenced elsewhere
112 if (GV.hasDLLExportStorageClass())
113 return true;
114
115 // As the name suggests, externally initialized variables need preserving as
116 // they would be initialized elsewhere externally.
117 if (const auto *G = dyn_cast<GlobalVariable>(Val: &GV))
118 if (G->isExternallyInitialized())
119 return true;
120
121 // Already local, has nothing to do.
122 if (GV.hasLocalLinkage())
123 return false;
124
125 // Check some special cases
126 if (AlwaysPreserved.count(Key: GV.getName()))
127 return true;
128
129 return MustPreserveGV(GV);
130}
131
132bool InternalizePass::maybeInternalize(
133 GlobalValue &GV, DenseMap<const Comdat *, ComdatInfo> &ComdatMap) {
134 if (Comdat *C = GV.getComdat()) {
135 // For GlobalAlias, C is the aliasee object's comdat which may have been
136 // redirected. So ComdatMap may not contain C.
137 if (ComdatMap.lookup(Val: C).External)
138 return false;
139
140 if (auto *GO = dyn_cast<GlobalObject>(Val: &GV)) {
141 // If a comdat with one member is not externally visible, we can drop it.
142 // Otherwise, the comdat can be used to establish dependencies among the
143 // group of sections. Thus we have to keep the comdat but switch it to
144 // nodeduplicate.
145 // Note: nodeduplicate is not necessary for COFF. wasm doesn't support
146 // nodeduplicate.
147 ComdatInfo &Info = ComdatMap.find(Val: C)->second;
148 if (Info.Size == 1)
149 GO->setComdat(nullptr);
150 else if (!IsWasm)
151 C->setSelectionKind(Comdat::NoDeduplicate);
152 }
153
154 if (GV.hasLocalLinkage())
155 return false;
156 } else {
157 if (GV.hasLocalLinkage())
158 return false;
159
160 if (shouldPreserveGV(GV))
161 return false;
162 }
163
164 GV.setVisibility(GlobalValue::DefaultVisibility);
165 GV.setLinkage(GlobalValue::InternalLinkage);
166 return true;
167}
168
169// If GV is part of a comdat and is externally visible, update the comdat size
170// and keep track of its comdat so that we don't internalize any of its members.
171void InternalizePass::checkComdat(
172 GlobalValue &GV, DenseMap<const Comdat *, ComdatInfo> &ComdatMap) {
173 Comdat *C = GV.getComdat();
174 if (!C)
175 return;
176
177 ComdatInfo &Info = ComdatMap[C];
178 ++Info.Size;
179 if (shouldPreserveGV(GV))
180 Info.External = true;
181}
182
183bool InternalizePass::internalizeModule(Module &M) {
184 bool Changed = false;
185
186 SmallVector<GlobalValue *, 4> Used;
187 collectUsedGlobalVariables(M, Vec&: Used, CompilerUsed: false);
188
189 // Collect comdat size and visiblity information for the module.
190 DenseMap<const Comdat *, ComdatInfo> ComdatMap;
191 if (!M.getComdatSymbolTable().empty()) {
192 for (Function &F : M)
193 checkComdat(GV&: F, ComdatMap);
194 for (GlobalVariable &GV : M.globals())
195 checkComdat(GV, ComdatMap);
196 for (GlobalAlias &GA : M.aliases())
197 checkComdat(GV&: GA, ComdatMap);
198 }
199
200 // We must assume that globals in llvm.used have a reference that not even
201 // the linker can see, so we don't internalize them.
202 // For llvm.compiler.used the situation is a bit fuzzy. The assembler and
203 // linker can drop those symbols. If this pass is running as part of LTO,
204 // one might think that it could just drop llvm.compiler.used. The problem
205 // is that even in LTO llvm doesn't see every reference. For example,
206 // we don't see references from function local inline assembly. To be
207 // conservative, we internalize symbols in llvm.compiler.used, but we
208 // keep llvm.compiler.used so that the symbol is not deleted by llvm.
209 for (GlobalValue *V : Used) {
210 AlwaysPreserved.insert(key: V->getName());
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(key: "llvm.used");
217 AlwaysPreserved.insert(key: "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(key: "llvm.global_ctors");
222 AlwaysPreserved.insert(key: "llvm.global_dtors");
223 AlwaysPreserved.insert(key: "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(key: "__stack_chk_fail");
229 if (M.getTargetTriple().isOSAIX())
230 AlwaysPreserved.insert(key: "__ssp_canary_word");
231 else
232 AlwaysPreserved.insert(key: "__stack_chk_guard");
233
234 // Preserve the RPC interface for GPU host callbacks when internalizing.
235 if (M.getTargetTriple().isNVPTX())
236 AlwaysPreserved.insert(key: "__llvm_rpc_client");
237
238 // Mark all functions not in the api as internal.
239 IsWasm = M.getTargetTriple().isOSBinFormatWasm();
240 for (Function &I : M) {
241 if (!maybeInternalize(GV&: I, ComdatMap))
242 continue;
243 Changed = true;
244
245 ++NumFunctions;
246 LLVM_DEBUG(dbgs() << "Internalizing func " << I.getName() << "\n");
247 }
248
249 // Mark all global variables with initializers that are not in the api as
250 // internal as well.
251 for (auto &GV : M.globals()) {
252 if (!maybeInternalize(GV, ComdatMap))
253 continue;
254 Changed = true;
255
256 ++NumGlobals;
257 LLVM_DEBUG(dbgs() << "Internalized gvar " << GV.getName() << "\n");
258 }
259
260 // Mark all aliases that are not in the api as internal as well.
261 for (auto &GA : M.aliases()) {
262 if (!maybeInternalize(GV&: GA, ComdatMap))
263 continue;
264 Changed = true;
265
266 ++NumAliases;
267 LLVM_DEBUG(dbgs() << "Internalized alias " << GA.getName() << "\n");
268 }
269
270 return Changed;
271}
272
273InternalizePass::InternalizePass() : MustPreserveGV(PreserveAPIList()) {}
274
275PreservedAnalyses InternalizePass::run(Module &M, ModuleAnalysisManager &AM) {
276 if (!internalizeModule(M))
277 return PreservedAnalyses::all();
278
279 return PreservedAnalyses::none();
280}
281