1 //===-- WebAssemblyFixFunctionBitcasts.cpp - Fix function bitcasts --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// Fix bitcasted functions.
13 /// WebAssembly requires caller and callee signatures to match, however in LLVM,
14 /// some amount of slop is vaguely permitted. Detect mismatch by looking for
15 /// bitcasts of functions and rewrite them to use wrapper functions instead.
17 /// This doesn't catch all cases, such as when a function's address is taken in
18 /// one place and casted in another, but it works for many common cases.
20 /// Note that LLVM already optimizes away function bitcasts in common cases by
21 /// dropping arguments as needed, so this pass only ends up getting used in less
24 //===----------------------------------------------------------------------===//
26 #include "WebAssembly.h"
27 #include "llvm/IR/CallSite.h"
28 #include "llvm/IR/Constants.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/Module.h"
31 #include "llvm/IR/Operator.h"
32 #include "llvm/Pass.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/raw_ostream.h"
37 #define DEBUG_TYPE "wasm-fix-function-bitcasts"
40 class FixFunctionBitcasts final : public ModulePass {
41 StringRef getPassName() const override {
42 return "WebAssembly Fix Function Bitcasts";
45 void getAnalysisUsage(AnalysisUsage &AU) const override {
47 ModulePass::getAnalysisUsage(AU);
50 bool runOnModule(Module &M) override;
54 FixFunctionBitcasts() : ModulePass(ID) {}
56 } // End anonymous namespace
58 char FixFunctionBitcasts::ID = 0;
59 INITIALIZE_PASS(FixFunctionBitcasts, DEBUG_TYPE,
60 "Fix mismatching bitcasts for WebAssembly", false, false)
62 ModulePass *llvm::createWebAssemblyFixFunctionBitcasts() {
63 return new FixFunctionBitcasts();
66 // Recursively descend the def-use lists from V to find non-bitcast users of
68 static void FindUses(Value *V, Function &F,
69 SmallVectorImpl<std::pair<Use *, Function *>> &Uses,
70 SmallPtrSetImpl<Constant *> &ConstantBCs) {
71 for (Use &U : V->uses()) {
72 if (BitCastOperator *BC = dyn_cast<BitCastOperator>(U.getUser()))
73 FindUses(BC, F, Uses, ConstantBCs);
74 else if (U.get()->getType() != F.getType()) {
75 CallSite CS(U.getUser());
77 // Skip uses that aren't immediately called
79 Value *Callee = CS.getCalledValue();
81 // Skip calls where the function isn't the callee
83 if (isa<Constant>(U.get())) {
84 // Only add constant bitcasts to the list once; they get RAUW'd
85 auto c = ConstantBCs.insert(cast<Constant>(U.get()));
89 Uses.push_back(std::make_pair(&U, &F));
94 // Create a wrapper function with type Ty that calls F (which may have a
95 // different type). Attempt to support common bitcasted function idioms:
96 // - Call with more arguments than needed: arguments are dropped
97 // - Call with fewer arguments than needed: arguments are filled in with undef
98 // - Return value is not needed: drop it
99 // - Return value needed but not present: supply an undef
101 // If the all the argument types of trivially castable to one another (i.e.
102 // I32 vs pointer type) then we don't create a wrapper at all (return nullptr
105 // If there is a type mismatch that we know would result in an invalid wasm
106 // module then generate wrapper that contains unreachable (i.e. abort at
107 // runtime). Such programs are deep into undefined behaviour territory,
108 // but we choose to fail at runtime rather than generate and invalid module
109 // or fail at compiler time. The reason we delay the error is that we want
110 // to support the CMake which expects to be able to compile and link programs
111 // that refer to functions with entirely incorrect signatures (this is how
112 // CMake detects the existence of a function in a toolchain).
114 // For bitcasts that involve struct types we don't know at this stage if they
115 // would be equivalent at the wasm level and so we can't know if we need to
116 // generate a wrapper.
117 static Function *CreateWrapper(Function *F, FunctionType *Ty) {
118 Module *M = F->getParent();
120 Function *Wrapper = Function::Create(Ty, Function::PrivateLinkage,
121 F->getName() + "_bitcast", M);
122 BasicBlock *BB = BasicBlock::Create(M->getContext(), "body", Wrapper);
123 const DataLayout &DL = BB->getModule()->getDataLayout();
125 // Determine what arguments to pass.
126 SmallVector<Value *, 4> Args;
127 Function::arg_iterator AI = Wrapper->arg_begin();
128 Function::arg_iterator AE = Wrapper->arg_end();
129 FunctionType::param_iterator PI = F->getFunctionType()->param_begin();
130 FunctionType::param_iterator PE = F->getFunctionType()->param_end();
131 bool TypeMismatch = false;
132 bool WrapperNeeded = false;
134 Type *ExpectedRtnType = F->getFunctionType()->getReturnType();
135 Type *RtnType = Ty->getReturnType();
137 if ((F->getFunctionType()->getNumParams() != Ty->getNumParams()) ||
138 (F->getFunctionType()->isVarArg() != Ty->isVarArg()) ||
139 (ExpectedRtnType != RtnType))
140 WrapperNeeded = true;
142 for (; AI != AE && PI != PE; ++AI, ++PI) {
143 Type *ArgType = AI->getType();
144 Type *ParamType = *PI;
146 if (ArgType == ParamType) {
147 Args.push_back(&*AI);
149 if (CastInst::isBitOrNoopPointerCastable(ArgType, ParamType, DL)) {
150 Instruction *PtrCast =
151 CastInst::CreateBitOrPointerCast(AI, ParamType, "cast");
152 BB->getInstList().push_back(PtrCast);
153 Args.push_back(PtrCast);
154 } else if (ArgType->isStructTy() || ParamType->isStructTy()) {
155 LLVM_DEBUG(dbgs() << "CreateWrapper: struct param type in bitcast: "
156 << F->getName() << "\n");
157 WrapperNeeded = false;
159 LLVM_DEBUG(dbgs() << "CreateWrapper: arg type mismatch calling: "
160 << F->getName() << "\n");
161 LLVM_DEBUG(dbgs() << "Arg[" << Args.size() << "] Expected: "
162 << *ParamType << " Got: " << *ArgType << "\n");
169 if (WrapperNeeded && !TypeMismatch) {
170 for (; PI != PE; ++PI)
171 Args.push_back(UndefValue::get(*PI));
173 for (; AI != AE; ++AI)
174 Args.push_back(&*AI);
176 CallInst *Call = CallInst::Create(F, Args, "", BB);
178 Type *ExpectedRtnType = F->getFunctionType()->getReturnType();
179 Type *RtnType = Ty->getReturnType();
180 // Determine what value to return.
181 if (RtnType->isVoidTy()) {
182 ReturnInst::Create(M->getContext(), BB);
183 } else if (ExpectedRtnType->isVoidTy()) {
184 LLVM_DEBUG(dbgs() << "Creating dummy return: " << *RtnType << "\n");
185 ReturnInst::Create(M->getContext(), UndefValue::get(RtnType), BB);
186 } else if (RtnType == ExpectedRtnType) {
187 ReturnInst::Create(M->getContext(), Call, BB);
188 } else if (CastInst::isBitOrNoopPointerCastable(ExpectedRtnType, RtnType,
191 CastInst::CreateBitOrPointerCast(Call, RtnType, "cast");
192 BB->getInstList().push_back(Cast);
193 ReturnInst::Create(M->getContext(), Cast, BB);
194 } else if (RtnType->isStructTy() || ExpectedRtnType->isStructTy()) {
195 LLVM_DEBUG(dbgs() << "CreateWrapper: struct return type in bitcast: "
196 << F->getName() << "\n");
197 WrapperNeeded = false;
199 LLVM_DEBUG(dbgs() << "CreateWrapper: return type mismatch calling: "
200 << F->getName() << "\n");
201 LLVM_DEBUG(dbgs() << "Expected: " << *ExpectedRtnType
202 << " Got: " << *RtnType << "\n");
208 // Create a new wrapper that simply contains `unreachable`.
209 Wrapper->eraseFromParent();
210 Wrapper = Function::Create(Ty, Function::PrivateLinkage,
211 F->getName() + "_bitcast_invalid", M);
212 BasicBlock *BB = BasicBlock::Create(M->getContext(), "body", Wrapper);
213 new UnreachableInst(M->getContext(), BB);
214 Wrapper->setName(F->getName() + "_bitcast_invalid");
215 } else if (!WrapperNeeded) {
216 LLVM_DEBUG(dbgs() << "CreateWrapper: no wrapper needed: " << F->getName()
218 Wrapper->eraseFromParent();
221 LLVM_DEBUG(dbgs() << "CreateWrapper: " << F->getName() << "\n");
225 // Test whether a main function with type FuncTy should be rewritten to have
227 bool shouldFixMainFunction(FunctionType *FuncTy, FunctionType *MainTy) {
228 // Only fix the main function if it's the standard zero-arg form. That way,
229 // the standard cases will work as expected, and users will see signature
230 // mismatches from the linker for non-standard cases.
231 return FuncTy->getReturnType() == MainTy->getReturnType() &&
232 FuncTy->getNumParams() == 0 &&
236 bool FixFunctionBitcasts::runOnModule(Module &M) {
237 LLVM_DEBUG(dbgs() << "********** Fix Function Bitcasts **********\n");
239 Function *Main = nullptr;
240 CallInst *CallMain = nullptr;
241 SmallVector<std::pair<Use *, Function *>, 0> Uses;
242 SmallPtrSet<Constant *, 2> ConstantBCs;
244 // Collect all the places that need wrappers.
245 for (Function &F : M) {
246 FindUses(&F, F, Uses, ConstantBCs);
248 // If we have a "main" function, and its type isn't
249 // "int main(int argc, char *argv[])", create an artificial call with it
250 // bitcasted to that type so that we generate a wrapper for it, so that
251 // the C runtime can call it.
252 if (F.getName() == "main") {
254 LLVMContext &C = M.getContext();
255 Type *MainArgTys[] = {Type::getInt32Ty(C),
256 PointerType::get(Type::getInt8PtrTy(C), 0)};
257 FunctionType *MainTy = FunctionType::get(Type::getInt32Ty(C), MainArgTys,
259 if (shouldFixMainFunction(F.getFunctionType(), MainTy)) {
260 LLVM_DEBUG(dbgs() << "Found `main` function with incorrect type: "
261 << *F.getFunctionType() << "\n");
262 Value *Args[] = {UndefValue::get(MainArgTys[0]),
263 UndefValue::get(MainArgTys[1])};
265 ConstantExpr::getBitCast(Main, PointerType::get(MainTy, 0));
266 CallMain = CallInst::Create(Casted, Args, "call_main");
267 Use *UseMain = &CallMain->getOperandUse(2);
268 Uses.push_back(std::make_pair(UseMain, &F));
273 DenseMap<std::pair<Function *, FunctionType *>, Function *> Wrappers;
275 for (auto &UseFunc : Uses) {
276 Use *U = UseFunc.first;
277 Function *F = UseFunc.second;
278 PointerType *PTy = cast<PointerType>(U->get()->getType());
279 FunctionType *Ty = dyn_cast<FunctionType>(PTy->getElementType());
281 // If the function is casted to something like i8* as a "generic pointer"
282 // to be later casted to something else, we can't generate a wrapper for it.
283 // Just ignore such casts for now.
287 auto Pair = Wrappers.insert(std::make_pair(std::make_pair(F, Ty), nullptr));
289 Pair.first->second = CreateWrapper(F, Ty);
291 Function *Wrapper = Pair.first->second;
295 if (isa<Constant>(U->get()))
296 U->get()->replaceAllUsesWith(Wrapper);
301 // If we created a wrapper for main, rename the wrapper so that it's the
302 // one that gets called from startup.
304 Main->setName("__original_main");
305 Function *MainWrapper =
306 cast<Function>(CallMain->getCalledValue()->stripPointerCasts());
308 if (Main->isDeclaration()) {
309 // The wrapper is not needed in this case as we don't need to export
310 // it to anyone else.
311 MainWrapper->eraseFromParent();
313 // Otherwise give the wrapper the same linkage as the original main
314 // function, so that it can be called from the same places.
315 MainWrapper->setName("main");
316 MainWrapper->setLinkage(Main->getLinkage());
317 MainWrapper->setVisibility(Main->getVisibility());