Line data Source code
1 : //===-- WebAssemblyFixFunctionBitcasts.cpp - Fix function bitcasts --------===//
2 : //
3 : // The LLVM Compiler Infrastructure
4 : //
5 : // This file is distributed under the University of Illinois Open Source
6 : // License. See LICENSE.TXT for details.
7 : //
8 : //===----------------------------------------------------------------------===//
9 : ///
10 : /// \file
11 : /// Fix bitcasted functions.
12 : ///
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.
16 : ///
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.
19 : ///
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
22 : /// common cases.
23 : ///
24 : //===----------------------------------------------------------------------===//
25 :
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"
35 : using namespace llvm;
36 :
37 : #define DEBUG_TYPE "wasm-fix-function-bitcasts"
38 :
39 : static cl::opt<bool>
40 : TemporaryWorkarounds("wasm-temporary-workarounds",
41 : cl::desc("Apply certain temporary workarounds"),
42 : cl::init(true), cl::Hidden);
43 :
44 : namespace {
45 : class FixFunctionBitcasts final : public ModulePass {
46 0 : StringRef getPassName() const override {
47 0 : return "WebAssembly Fix Function Bitcasts";
48 : }
49 :
50 305 : void getAnalysisUsage(AnalysisUsage &AU) const override {
51 305 : AU.setPreservesCFG();
52 305 : ModulePass::getAnalysisUsage(AU);
53 305 : }
54 :
55 : bool runOnModule(Module &M) override;
56 :
57 : public:
58 : static char ID;
59 305 : FixFunctionBitcasts() : ModulePass(ID) {}
60 : };
61 : } // End anonymous namespace
62 :
63 : char FixFunctionBitcasts::ID = 0;
64 199030 : INITIALIZE_PASS(FixFunctionBitcasts, DEBUG_TYPE,
65 : "Fix mismatching bitcasts for WebAssembly", false, false)
66 :
67 305 : ModulePass *llvm::createWebAssemblyFixFunctionBitcasts() {
68 305 : return new FixFunctionBitcasts();
69 : }
70 :
71 : // Recursively descend the def-use lists from V to find non-bitcast users of
72 : // bitcasts of V.
73 3801 : static void FindUses(Value *V, Function &F,
74 : SmallVectorImpl<std::pair<Use *, Function *>> &Uses,
75 : SmallPtrSetImpl<Constant *> &ConstantBCs) {
76 4771 : for (Use &U : V->uses()) {
77 970 : if (BitCastOperator *BC = dyn_cast<BitCastOperator>(U.getUser()))
78 64 : FindUses(BC, F, Uses, ConstantBCs);
79 1812 : else if (U.get()->getType() != F.getType()) {
80 : CallSite CS(U.getUser());
81 87 : if (!CS)
82 : // Skip uses that aren't immediately called
83 : continue;
84 : Value *Callee = CS.getCalledValue();
85 30 : if (Callee != V)
86 : // Skip calls where the function isn't the callee
87 : continue;
88 28 : if (isa<Constant>(U.get())) {
89 : // Only add constant bitcasts to the list once; they get RAUW'd
90 26 : auto c = ConstantBCs.insert(cast<Constant>(U.get()));
91 26 : if (!c.second)
92 4 : continue;
93 : }
94 24 : Uses.push_back(std::make_pair(&U, &F));
95 : }
96 : }
97 3801 : }
98 :
99 : // Create a wrapper function with type Ty that calls F (which may have a
100 : // different type). Attempt to support common bitcasted function idioms:
101 : // - Call with more arguments than needed: arguments are dropped
102 : // - Call with fewer arguments than needed: arguments are filled in with undef
103 : // - Return value is not needed: drop it
104 : // - Return value needed but not present: supply an undef
105 : //
106 : // If the all the argument types of trivially castable to one another (i.e.
107 : // I32 vs pointer type) then we don't create a wrapper at all (return nullptr
108 : // instead).
109 : //
110 : // If there is a type mismatch that we know would result in an invalid wasm
111 : // module then generate wrapper that contains unreachable (i.e. abort at
112 : // runtime). Such programs are deep into undefined behaviour territory,
113 : // but we choose to fail at runtime rather than generate and invalid module
114 : // or fail at compiler time. The reason we delay the error is that we want
115 : // to support the CMake which expects to be able to compile and link programs
116 : // that refer to functions with entirely incorrect signatures (this is how
117 : // CMake detects the existence of a function in a toolchain).
118 : //
119 : // For bitcasts that involve struct types we don't know at this stage if they
120 : // would be equivalent at the wasm level and so we can't know if we need to
121 : // generate a wrapper.
122 23 : static Function *CreateWrapper(Function *F, FunctionType *Ty) {
123 23 : Module *M = F->getParent();
124 :
125 : Function *Wrapper = Function::Create(Ty, Function::PrivateLinkage,
126 46 : F->getName() + "_bitcast", M);
127 46 : BasicBlock *BB = BasicBlock::Create(M->getContext(), "body", Wrapper);
128 23 : const DataLayout &DL = BB->getModule()->getDataLayout();
129 :
130 : // Determine what arguments to pass.
131 : SmallVector<Value *, 4> Args;
132 : Function::arg_iterator AI = Wrapper->arg_begin();
133 : Function::arg_iterator AE = Wrapper->arg_end();
134 23 : FunctionType::param_iterator PI = F->getFunctionType()->param_begin();
135 23 : FunctionType::param_iterator PE = F->getFunctionType()->param_end();
136 : bool TypeMismatch = false;
137 : bool WrapperNeeded = false;
138 :
139 : Type *ExpectedRtnType = F->getFunctionType()->getReturnType();
140 23 : Type *RtnType = Ty->getReturnType();
141 :
142 12 : if ((F->getFunctionType()->getNumParams() != Ty->getNumParams()) ||
143 35 : (F->getFunctionType()->isVarArg() != Ty->isVarArg()) ||
144 : (ExpectedRtnType != RtnType))
145 : WrapperNeeded = true;
146 :
147 32 : for (; AI != AE && PI != PE; ++AI, ++PI) {
148 12 : Type *ArgType = AI->getType();
149 12 : Type *ParamType = *PI;
150 :
151 12 : if (ArgType == ParamType) {
152 4 : Args.push_back(&*AI);
153 : } else {
154 8 : if (CastInst::isBitOrNoopPointerCastable(ArgType, ParamType, DL)) {
155 : Instruction *PtrCast =
156 4 : CastInst::CreateBitOrPointerCast(AI, ParamType, "cast");
157 4 : BB->getInstList().push_back(PtrCast);
158 4 : Args.push_back(PtrCast);
159 4 : } else if (ArgType->isStructTy() || ParamType->isStructTy()) {
160 : LLVM_DEBUG(dbgs() << "CreateWrapper: struct param type in bitcast: "
161 : << F->getName() << "\n");
162 : WrapperNeeded = false;
163 : } else {
164 : LLVM_DEBUG(dbgs() << "CreateWrapper: arg type mismatch calling: "
165 : << F->getName() << "\n");
166 : LLVM_DEBUG(dbgs() << "Arg[" << Args.size() << "] Expected: "
167 : << *ParamType << " Got: " << *ArgType << "\n");
168 : TypeMismatch = true;
169 : break;
170 : }
171 : }
172 : }
173 :
174 23 : if (WrapperNeeded && !TypeMismatch) {
175 22 : for (; PI != PE; ++PI)
176 5 : Args.push_back(UndefValue::get(*PI));
177 17 : if (F->isVarArg())
178 11 : for (; AI != AE; ++AI)
179 7 : Args.push_back(&*AI);
180 :
181 17 : CallInst *Call = CallInst::Create(F, Args, "", BB);
182 :
183 17 : Type *ExpectedRtnType = F->getFunctionType()->getReturnType();
184 17 : Type *RtnType = Ty->getReturnType();
185 : // Determine what value to return.
186 17 : if (RtnType->isVoidTy()) {
187 11 : ReturnInst::Create(M->getContext(), BB);
188 6 : } else if (ExpectedRtnType->isVoidTy()) {
189 : LLVM_DEBUG(dbgs() << "Creating dummy return: " << *RtnType << "\n");
190 3 : ReturnInst::Create(M->getContext(), UndefValue::get(RtnType), BB);
191 3 : } else if (RtnType == ExpectedRtnType) {
192 0 : ReturnInst::Create(M->getContext(), Call, BB);
193 3 : } else if (CastInst::isBitOrNoopPointerCastable(ExpectedRtnType, RtnType,
194 : DL)) {
195 : Instruction *Cast =
196 0 : CastInst::CreateBitOrPointerCast(Call, RtnType, "cast");
197 0 : BB->getInstList().push_back(Cast);
198 0 : ReturnInst::Create(M->getContext(), Cast, BB);
199 3 : } else if (RtnType->isStructTy() || ExpectedRtnType->isStructTy()) {
200 : LLVM_DEBUG(dbgs() << "CreateWrapper: struct return type in bitcast: "
201 : << F->getName() << "\n");
202 : WrapperNeeded = false;
203 : } else {
204 : LLVM_DEBUG(dbgs() << "CreateWrapper: return type mismatch calling: "
205 : << F->getName() << "\n");
206 : LLVM_DEBUG(dbgs() << "Expected: " << *ExpectedRtnType
207 : << " Got: " << *RtnType << "\n");
208 : TypeMismatch = true;
209 : }
210 : }
211 :
212 22 : if (TypeMismatch) {
213 : // Create a new wrapper that simply contains `unreachable`.
214 4 : Wrapper->eraseFromParent();
215 : Wrapper = Function::Create(Ty, Function::PrivateLinkage,
216 8 : F->getName() + "_bitcast_invalid", M);
217 4 : BasicBlock *BB = BasicBlock::Create(M->getContext(), "body", Wrapper);
218 4 : new UnreachableInst(M->getContext(), BB);
219 4 : Wrapper->setName(F->getName() + "_bitcast_invalid");
220 19 : } else if (!WrapperNeeded) {
221 : LLVM_DEBUG(dbgs() << "CreateWrapper: no wrapper needed: " << F->getName()
222 : << "\n");
223 5 : Wrapper->eraseFromParent();
224 5 : return nullptr;
225 : }
226 : LLVM_DEBUG(dbgs() << "CreateWrapper: " << F->getName() << "\n");
227 : return Wrapper;
228 : }
229 :
230 306 : bool FixFunctionBitcasts::runOnModule(Module &M) {
231 : Function *Main = nullptr;
232 : CallInst *CallMain = nullptr;
233 : SmallVector<std::pair<Use *, Function *>, 0> Uses;
234 : SmallPtrSet<Constant *, 2> ConstantBCs;
235 :
236 : // Collect all the places that need wrappers.
237 4043 : for (Function &F : M) {
238 3737 : FindUses(&F, F, Uses, ConstantBCs);
239 :
240 : // If we have a "main" function, and its type isn't
241 : // "int main(int argc, char *argv[])", create an artificial call with it
242 : // bitcasted to that type so that we generate a wrapper for it, so that
243 : // the C runtime can call it.
244 3737 : if (!TemporaryWorkarounds && !F.isDeclaration() && F.getName() == "main") {
245 : Main = &F;
246 2 : LLVMContext &C = M.getContext();
247 2 : Type *MainArgTys[] = {Type::getInt32Ty(C),
248 2 : PointerType::get(Type::getInt8PtrTy(C), 0)};
249 2 : FunctionType *MainTy = FunctionType::get(Type::getInt32Ty(C), MainArgTys,
250 : /*isVarArg=*/false);
251 2 : if (F.getFunctionType() != MainTy) {
252 : LLVM_DEBUG(dbgs() << "Found `main` function with incorrect type: "
253 : << *F.getFunctionType() << "\n");
254 1 : Value *Args[] = {UndefValue::get(MainArgTys[0]),
255 1 : UndefValue::get(MainArgTys[1])};
256 : Value *Casted =
257 1 : ConstantExpr::getBitCast(Main, PointerType::get(MainTy, 0));
258 1 : CallMain = CallInst::Create(Casted, Args, "call_main");
259 1 : Use *UseMain = &CallMain->getOperandUse(2);
260 1 : Uses.push_back(std::make_pair(UseMain, &F));
261 : }
262 : }
263 : }
264 :
265 : DenseMap<std::pair<Function *, FunctionType *>, Function *> Wrappers;
266 :
267 331 : for (auto &UseFunc : Uses) {
268 25 : Use *U = UseFunc.first;
269 25 : Function *F = UseFunc.second;
270 25 : PointerType *PTy = cast<PointerType>(U->get()->getType());
271 25 : FunctionType *Ty = dyn_cast<FunctionType>(PTy->getElementType());
272 :
273 : // If the function is casted to something like i8* as a "generic pointer"
274 : // to be later casted to something else, we can't generate a wrapper for it.
275 : // Just ignore such casts for now.
276 : if (!Ty)
277 5 : continue;
278 :
279 25 : auto Pair = Wrappers.insert(std::make_pair(std::make_pair(F, Ty), nullptr));
280 25 : if (Pair.second)
281 23 : Pair.first->second = CreateWrapper(F, Ty);
282 :
283 25 : Function *Wrapper = Pair.first->second;
284 25 : if (!Wrapper)
285 : continue;
286 :
287 40 : if (isa<Constant>(U->get()))
288 18 : U->get()->replaceAllUsesWith(Wrapper);
289 : else
290 2 : U->set(Wrapper);
291 : }
292 :
293 : // If we created a wrapper for main, rename the wrapper so that it's the
294 : // one that gets called from startup.
295 306 : if (CallMain) {
296 2 : Main->setName("__original_main");
297 : Function *MainWrapper =
298 : cast<Function>(CallMain->getCalledValue()->stripPointerCasts());
299 2 : MainWrapper->setName("main");
300 : MainWrapper->setLinkage(Main->getLinkage());
301 : MainWrapper->setVisibility(Main->getVisibility());
302 : Main->setLinkage(Function::PrivateLinkage);
303 : Main->setVisibility(Function::DefaultVisibility);
304 1 : delete CallMain;
305 : }
306 :
307 306 : return true;
308 : }
|
