xref: /freebsd/contrib/llvm-project/llvm/lib/Target/WebAssembly/WebAssemblyFixFunctionBitcasts.cpp (revision 25ecdc7d52770caf1c9b44b5ec11f468f6b636f3)
1 //===-- WebAssemblyFixFunctionBitcasts.cpp - Fix function bitcasts --------===//
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 /// \file
10 /// Fix bitcasted functions.
11 ///
12 /// WebAssembly requires caller and callee signatures to match, however in LLVM,
13 /// some amount of slop is vaguely permitted. Detect mismatch by looking for
14 /// bitcasts of functions and rewrite them to use wrapper functions instead.
15 ///
16 /// This doesn't catch all cases, such as when a function's address is taken in
17 /// one place and casted in another, but it works for many common cases.
18 ///
19 /// Note that LLVM already optimizes away function bitcasts in common cases by
20 /// dropping arguments as needed, so this pass only ends up getting used in less
21 /// common cases.
22 ///
23 //===----------------------------------------------------------------------===//
24 
25 #include "WebAssembly.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/Instructions.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/IR/Operator.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
33 using namespace llvm;
34 
35 #define DEBUG_TYPE "wasm-fix-function-bitcasts"
36 
37 namespace {
38 class FixFunctionBitcasts final : public ModulePass {
39   StringRef getPassName() const override {
40     return "WebAssembly Fix Function Bitcasts";
41   }
42 
43   void getAnalysisUsage(AnalysisUsage &AU) const override {
44     AU.setPreservesCFG();
45     ModulePass::getAnalysisUsage(AU);
46   }
47 
48   bool runOnModule(Module &M) override;
49 
50 public:
51   static char ID;
52   FixFunctionBitcasts() : ModulePass(ID) {}
53 };
54 } // End anonymous namespace
55 
56 char FixFunctionBitcasts::ID = 0;
57 INITIALIZE_PASS(FixFunctionBitcasts, DEBUG_TYPE,
58                 "Fix mismatching bitcasts for WebAssembly", false, false)
59 
60 ModulePass *llvm::createWebAssemblyFixFunctionBitcasts() {
61   return new FixFunctionBitcasts();
62 }
63 
64 // Recursively descend the def-use lists from V to find non-bitcast users of
65 // bitcasts of V.
66 static void findUses(Value *V, Function &F,
67                      SmallVectorImpl<std::pair<Use *, Function *>> &Uses,
68                      SmallPtrSetImpl<Constant *> &ConstantBCs) {
69   for (Use &U : V->uses()) {
70     if (auto *BC = dyn_cast<BitCastOperator>(U.getUser()))
71       findUses(BC, F, Uses, ConstantBCs);
72     else if (auto *A = dyn_cast<GlobalAlias>(U.getUser()))
73       findUses(A, F, Uses, ConstantBCs);
74     else if (U.get()->getType() != F.getType()) {
75       CallBase *CB = dyn_cast<CallBase>(U.getUser());
76       if (!CB)
77         // Skip uses that aren't immediately called
78         continue;
79       Value *Callee = CB->getCalledOperand();
80       if (Callee != V)
81         // Skip calls where the function isn't the callee
82         continue;
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()));
86         if (!C.second)
87           continue;
88       }
89       Uses.push_back(std::make_pair(&U, &F));
90     }
91   }
92 }
93 
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
100 //
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
103 // instead).
104 //
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).
113 //
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();
119 
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();
124 
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;
133 
134   Type *ExpectedRtnType = F->getFunctionType()->getReturnType();
135   Type *RtnType = Ty->getReturnType();
136 
137   if ((F->getFunctionType()->getNumParams() != Ty->getNumParams()) ||
138       (F->getFunctionType()->isVarArg() != Ty->isVarArg()) ||
139       (ExpectedRtnType != RtnType))
140     WrapperNeeded = true;
141 
142   for (; AI != AE && PI != PE; ++AI, ++PI) {
143     Type *ArgType = AI->getType();
144     Type *ParamType = *PI;
145 
146     if (ArgType == ParamType) {
147       Args.push_back(&*AI);
148     } else {
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;
158       } else {
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");
163         TypeMismatch = true;
164         break;
165       }
166     }
167   }
168 
169   if (WrapperNeeded && !TypeMismatch) {
170     for (; PI != PE; ++PI)
171       Args.push_back(UndefValue::get(*PI));
172     if (F->isVarArg())
173       for (; AI != AE; ++AI)
174         Args.push_back(&*AI);
175 
176     CallInst *Call = CallInst::Create(F, Args, "", BB);
177 
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,
189                                                     DL)) {
190       Instruction *Cast =
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;
198     } else {
199       LLVM_DEBUG(dbgs() << "createWrapper: return type mismatch calling: "
200                         << F->getName() << "\n");
201       LLVM_DEBUG(dbgs() << "Expected: " << *ExpectedRtnType
202                         << " Got: " << *RtnType << "\n");
203       TypeMismatch = true;
204     }
205   }
206 
207   if (TypeMismatch) {
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()
217                       << "\n");
218     Wrapper->eraseFromParent();
219     return nullptr;
220   }
221   LLVM_DEBUG(dbgs() << "createWrapper: " << F->getName() << "\n");
222   return Wrapper;
223 }
224 
225 // Test whether a main function with type FuncTy should be rewritten to have
226 // type MainTy.
227 static 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 &&
233          !FuncTy->isVarArg();
234 }
235 
236 bool FixFunctionBitcasts::runOnModule(Module &M) {
237   LLVM_DEBUG(dbgs() << "********** Fix Function Bitcasts **********\n");
238 
239   Function *Main = nullptr;
240   CallInst *CallMain = nullptr;
241   SmallVector<std::pair<Use *, Function *>, 0> Uses;
242   SmallPtrSet<Constant *, 2> ConstantBCs;
243 
244   // Collect all the places that need wrappers.
245   for (Function &F : M) {
246     // Skip to fix when the function is swiftcc because swiftcc allows
247     // bitcast type difference for swiftself and swifterror.
248     if (F.getCallingConv() == CallingConv::Swift)
249       continue;
250     findUses(&F, F, Uses, ConstantBCs);
251 
252     // If we have a "main" function, and its type isn't
253     // "int main(int argc, char *argv[])", create an artificial call with it
254     // bitcasted to that type so that we generate a wrapper for it, so that
255     // the C runtime can call it.
256     if (F.getName() == "main") {
257       Main = &F;
258       LLVMContext &C = M.getContext();
259       Type *MainArgTys[] = {Type::getInt32Ty(C),
260                             PointerType::get(Type::getInt8PtrTy(C), 0)};
261       FunctionType *MainTy = FunctionType::get(Type::getInt32Ty(C), MainArgTys,
262                                                /*isVarArg=*/false);
263       if (shouldFixMainFunction(F.getFunctionType(), MainTy)) {
264         LLVM_DEBUG(dbgs() << "Found `main` function with incorrect type: "
265                           << *F.getFunctionType() << "\n");
266         Value *Args[] = {UndefValue::get(MainArgTys[0]),
267                          UndefValue::get(MainArgTys[1])};
268         Value *Casted =
269             ConstantExpr::getBitCast(Main, PointerType::get(MainTy, 0));
270         CallMain = CallInst::Create(MainTy, Casted, Args, "call_main");
271         Use *UseMain = &CallMain->getOperandUse(2);
272         Uses.push_back(std::make_pair(UseMain, &F));
273       }
274     }
275   }
276 
277   DenseMap<std::pair<Function *, FunctionType *>, Function *> Wrappers;
278 
279   for (auto &UseFunc : Uses) {
280     Use *U = UseFunc.first;
281     Function *F = UseFunc.second;
282     auto *PTy = cast<PointerType>(U->get()->getType());
283     auto *Ty = dyn_cast<FunctionType>(PTy->getElementType());
284 
285     // If the function is casted to something like i8* as a "generic pointer"
286     // to be later casted to something else, we can't generate a wrapper for it.
287     // Just ignore such casts for now.
288     if (!Ty)
289       continue;
290 
291     auto Pair = Wrappers.insert(std::make_pair(std::make_pair(F, Ty), nullptr));
292     if (Pair.second)
293       Pair.first->second = createWrapper(F, Ty);
294 
295     Function *Wrapper = Pair.first->second;
296     if (!Wrapper)
297       continue;
298 
299     if (isa<Constant>(U->get()))
300       U->get()->replaceAllUsesWith(Wrapper);
301     else
302       U->set(Wrapper);
303   }
304 
305   // If we created a wrapper for main, rename the wrapper so that it's the
306   // one that gets called from startup.
307   if (CallMain) {
308     Main->setName("__original_main");
309     auto *MainWrapper =
310         cast<Function>(CallMain->getCalledOperand()->stripPointerCasts());
311     delete CallMain;
312     if (Main->isDeclaration()) {
313       // The wrapper is not needed in this case as we don't need to export
314       // it to anyone else.
315       MainWrapper->eraseFromParent();
316     } else {
317       // Otherwise give the wrapper the same linkage as the original main
318       // function, so that it can be called from the same places.
319       MainWrapper->setName("main");
320       MainWrapper->setLinkage(Main->getLinkage());
321       MainWrapper->setVisibility(Main->getVisibility());
322     }
323   }
324 
325   return true;
326 }
327