1 //===-- ModuleUtils.cpp - Functions to manipulate Modules -----------------===//
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 family of functions perform manipulations on Modules.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/Transforms/Utils/ModuleUtils.h"
14 #include "llvm/IR/DerivedTypes.h"
15 #include "llvm/IR/Function.h"
16 #include "llvm/IR/IRBuilder.h"
17 #include "llvm/IR/Module.h"
18 #include "llvm/Support/raw_ostream.h"
19 
20 using namespace llvm;
21 
22 static void appendToGlobalArray(const char *Array, Module &M, Function *F,
23                                 int Priority, Constant *Data) {
24   IRBuilder<> IRB(M.getContext());
25   FunctionType *FnTy = FunctionType::get(IRB.getVoidTy(), false);
26 
27   // Get the current set of static global constructors and add the new ctor
28   // to the list.
29   SmallVector<Constant *, 16> CurrentCtors;
30   StructType *EltTy;
31   if (GlobalVariable *GVCtor = M.getNamedGlobal(Array)) {
32     ArrayType *ATy = cast<ArrayType>(GVCtor->getValueType());
33     StructType *OldEltTy = cast<StructType>(ATy->getElementType());
34     // Upgrade a 2-field global array type to the new 3-field format if needed.
35     if (Data && OldEltTy->getNumElements() < 3)
36       EltTy = StructType::get(IRB.getInt32Ty(), PointerType::getUnqual(FnTy),
37                               IRB.getInt8PtrTy());
38     else
39       EltTy = OldEltTy;
40     if (Constant *Init = GVCtor->getInitializer()) {
41       unsigned n = Init->getNumOperands();
42       CurrentCtors.reserve(n + 1);
43       for (unsigned i = 0; i != n; ++i) {
44         auto Ctor = cast<Constant>(Init->getOperand(i));
45         if (EltTy != OldEltTy)
46           Ctor =
47               ConstantStruct::get(EltTy, Ctor->getAggregateElement((unsigned)0),
48                                   Ctor->getAggregateElement(1),
49                                   Constant::getNullValue(IRB.getInt8PtrTy()));
50         CurrentCtors.push_back(Ctor);
51       }
52     }
53     GVCtor->eraseFromParent();
54   } else {
55     // Use the new three-field struct if there isn't one already.
56     EltTy = StructType::get(IRB.getInt32Ty(), PointerType::getUnqual(FnTy),
57                             IRB.getInt8PtrTy());
58   }
59 
60   // Build a 2 or 3 field global_ctor entry.  We don't take a comdat key.
61   Constant *CSVals[3];
62   CSVals[0] = IRB.getInt32(Priority);
63   CSVals[1] = F;
64   // FIXME: Drop support for the two element form in LLVM 4.0.
65   if (EltTy->getNumElements() >= 3)
66     CSVals[2] = Data ? ConstantExpr::getPointerCast(Data, IRB.getInt8PtrTy())
67                      : Constant::getNullValue(IRB.getInt8PtrTy());
68   Constant *RuntimeCtorInit =
69       ConstantStruct::get(EltTy, makeArrayRef(CSVals, EltTy->getNumElements()));
70 
71   CurrentCtors.push_back(RuntimeCtorInit);
72 
73   // Create a new initializer.
74   ArrayType *AT = ArrayType::get(EltTy, CurrentCtors.size());
75   Constant *NewInit = ConstantArray::get(AT, CurrentCtors);
76 
77   // Create the new global variable and replace all uses of
78   // the old global variable with the new one.
79   (void)new GlobalVariable(M, NewInit->getType(), false,
80                            GlobalValue::AppendingLinkage, NewInit, Array);
81 }
82 
83 void llvm::appendToGlobalCtors(Module &M, Function *F, int Priority, Constant *Data) {
84   appendToGlobalArray("llvm.global_ctors", M, F, Priority, Data);
85 }
86 
87 void llvm::appendToGlobalDtors(Module &M, Function *F, int Priority, Constant *Data) {
88   appendToGlobalArray("llvm.global_dtors", M, F, Priority, Data);
89 }
90 
91 static void appendToUsedList(Module &M, StringRef Name, ArrayRef<GlobalValue *> Values) {
92   GlobalVariable *GV = M.getGlobalVariable(Name);
93   SmallPtrSet<Constant *, 16> InitAsSet;
94   SmallVector<Constant *, 16> Init;
95   if (GV) {
96     ConstantArray *CA = dyn_cast<ConstantArray>(GV->getInitializer());
97     for (auto &Op : CA->operands()) {
98       Constant *C = cast_or_null<Constant>(Op);
99       if (InitAsSet.insert(C).second)
100         Init.push_back(C);
101     }
102     GV->eraseFromParent();
103   }
104 
105   Type *Int8PtrTy = llvm::Type::getInt8PtrTy(M.getContext());
106   for (auto *V : Values) {
107     Constant *C = ConstantExpr::getBitCast(V, Int8PtrTy);
108     if (InitAsSet.insert(C).second)
109       Init.push_back(C);
110   }
111 
112   if (Init.empty())
113     return;
114 
115   ArrayType *ATy = ArrayType::get(Int8PtrTy, Init.size());
116   GV = new llvm::GlobalVariable(M, ATy, false, GlobalValue::AppendingLinkage,
117                                 ConstantArray::get(ATy, Init), Name);
118   GV->setSection("llvm.metadata");
119 }
120 
121 void llvm::appendToUsed(Module &M, ArrayRef<GlobalValue *> Values) {
122   appendToUsedList(M, "llvm.used", Values);
123 }
124 
125 void llvm::appendToCompilerUsed(Module &M, ArrayRef<GlobalValue *> Values) {
126   appendToUsedList(M, "llvm.compiler.used", Values);
127 }
128 
129 FunctionCallee
130 llvm::declareSanitizerInitFunction(Module &M, StringRef InitName,
131                                    ArrayRef<Type *> InitArgTypes) {
132   assert(!InitName.empty() && "Expected init function name");
133   return M.getOrInsertFunction(
134       InitName,
135       FunctionType::get(Type::getVoidTy(M.getContext()), InitArgTypes, false),
136       AttributeList());
137 }
138 
139 std::pair<Function *, FunctionCallee> llvm::createSanitizerCtorAndInitFunctions(
140     Module &M, StringRef CtorName, StringRef InitName,
141     ArrayRef<Type *> InitArgTypes, ArrayRef<Value *> InitArgs,
142     StringRef VersionCheckName) {
143   assert(!InitName.empty() && "Expected init function name");
144   assert(InitArgs.size() == InitArgTypes.size() &&
145          "Sanitizer's init function expects different number of arguments");
146   FunctionCallee InitFunction =
147       declareSanitizerInitFunction(M, InitName, InitArgTypes);
148   Function *Ctor = Function::Create(
149       FunctionType::get(Type::getVoidTy(M.getContext()), false),
150       GlobalValue::InternalLinkage, CtorName, &M);
151   BasicBlock *CtorBB = BasicBlock::Create(M.getContext(), "", Ctor);
152   IRBuilder<> IRB(ReturnInst::Create(M.getContext(), CtorBB));
153   IRB.CreateCall(InitFunction, InitArgs);
154   if (!VersionCheckName.empty()) {
155     FunctionCallee VersionCheckFunction = M.getOrInsertFunction(
156         VersionCheckName, FunctionType::get(IRB.getVoidTy(), {}, false),
157         AttributeList());
158     IRB.CreateCall(VersionCheckFunction, {});
159   }
160   return std::make_pair(Ctor, InitFunction);
161 }
162 
163 std::pair<Function *, FunctionCallee>
164 llvm::getOrCreateSanitizerCtorAndInitFunctions(
165     Module &M, StringRef CtorName, StringRef InitName,
166     ArrayRef<Type *> InitArgTypes, ArrayRef<Value *> InitArgs,
167     function_ref<void(Function *, FunctionCallee)> FunctionsCreatedCallback,
168     StringRef VersionCheckName) {
169   assert(!CtorName.empty() && "Expected ctor function name");
170 
171   if (Function *Ctor = M.getFunction(CtorName))
172     // FIXME: Sink this logic into the module, similar to the handling of
173     // globals. This will make moving to a concurrent model much easier.
174     if (Ctor->arg_size() == 0 ||
175         Ctor->getReturnType() == Type::getVoidTy(M.getContext()))
176       return {Ctor, declareSanitizerInitFunction(M, InitName, InitArgTypes)};
177 
178   Function *Ctor;
179   FunctionCallee InitFunction;
180   std::tie(Ctor, InitFunction) = llvm::createSanitizerCtorAndInitFunctions(
181       M, CtorName, InitName, InitArgTypes, InitArgs, VersionCheckName);
182   FunctionsCreatedCallback(Ctor, InitFunction);
183   return std::make_pair(Ctor, InitFunction);
184 }
185 
186 Function *llvm::getOrCreateInitFunction(Module &M, StringRef Name) {
187   assert(!Name.empty() && "Expected init function name");
188   if (Function *F = M.getFunction(Name)) {
189     if (F->arg_size() != 0 ||
190         F->getReturnType() != Type::getVoidTy(M.getContext())) {
191       std::string Err;
192       raw_string_ostream Stream(Err);
193       Stream << "Sanitizer interface function defined with wrong type: " << *F;
194       report_fatal_error(Err);
195     }
196     return F;
197   }
198   Function *F =
199       cast<Function>(M.getOrInsertFunction(Name, AttributeList(),
200                                            Type::getVoidTy(M.getContext()))
201                          .getCallee());
202 
203   appendToGlobalCtors(M, F, 0);
204 
205   return F;
206 }
207 
208 void llvm::filterDeadComdatFunctions(
209     Module &M, SmallVectorImpl<Function *> &DeadComdatFunctions) {
210   // Build a map from the comdat to the number of entries in that comdat we
211   // think are dead. If this fully covers the comdat group, then the entire
212   // group is dead. If we find another entry in the comdat group though, we'll
213   // have to preserve the whole group.
214   SmallDenseMap<Comdat *, int, 16> ComdatEntriesCovered;
215   for (Function *F : DeadComdatFunctions) {
216     Comdat *C = F->getComdat();
217     assert(C && "Expected all input GVs to be in a comdat!");
218     ComdatEntriesCovered[C] += 1;
219   }
220 
221   auto CheckComdat = [&](Comdat &C) {
222     auto CI = ComdatEntriesCovered.find(&C);
223     if (CI == ComdatEntriesCovered.end())
224       return;
225 
226     // If this could have been covered by a dead entry, just subtract one to
227     // account for it.
228     if (CI->second > 0) {
229       CI->second -= 1;
230       return;
231     }
232 
233     // If we've already accounted for all the entries that were dead, the
234     // entire comdat is alive so remove it from the map.
235     ComdatEntriesCovered.erase(CI);
236   };
237 
238   auto CheckAllComdats = [&] {
239     for (Function &F : M.functions())
240       if (Comdat *C = F.getComdat()) {
241         CheckComdat(*C);
242         if (ComdatEntriesCovered.empty())
243           return;
244       }
245     for (GlobalVariable &GV : M.globals())
246       if (Comdat *C = GV.getComdat()) {
247         CheckComdat(*C);
248         if (ComdatEntriesCovered.empty())
249           return;
250       }
251     for (GlobalAlias &GA : M.aliases())
252       if (Comdat *C = GA.getComdat()) {
253         CheckComdat(*C);
254         if (ComdatEntriesCovered.empty())
255           return;
256       }
257   };
258   CheckAllComdats();
259 
260   if (ComdatEntriesCovered.empty()) {
261     DeadComdatFunctions.clear();
262     return;
263   }
264 
265   // Remove the entries that were not covering.
266   erase_if(DeadComdatFunctions, [&](GlobalValue *GV) {
267     return ComdatEntriesCovered.find(GV->getComdat()) ==
268            ComdatEntriesCovered.end();
269   });
270 }
271 
272 std::string llvm::getUniqueModuleId(Module *M) {
273   MD5 Md5;
274   bool ExportsSymbols = false;
275   auto AddGlobal = [&](GlobalValue &GV) {
276     if (GV.isDeclaration() || GV.getName().startswith("llvm.") ||
277         !GV.hasExternalLinkage() || GV.hasComdat())
278       return;
279     ExportsSymbols = true;
280     Md5.update(GV.getName());
281     Md5.update(ArrayRef<uint8_t>{0});
282   };
283 
284   for (auto &F : *M)
285     AddGlobal(F);
286   for (auto &GV : M->globals())
287     AddGlobal(GV);
288   for (auto &GA : M->aliases())
289     AddGlobal(GA);
290   for (auto &IF : M->ifuncs())
291     AddGlobal(IF);
292 
293   if (!ExportsSymbols)
294     return "";
295 
296   MD5::MD5Result R;
297   Md5.final(R);
298 
299   SmallString<32> Str;
300   MD5::stringifyResult(R, Str);
301   return ("$" + Str).str();
302 }
303