1 //=== WebAssemblyLowerEmscriptenEHSjLj.cpp - Lower exceptions for Emscripten =// 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 /// This file lowers exception-related instructions and setjmp/longjmp 11 /// function calls in order to use Emscripten's JavaScript try and catch 12 /// mechanism. 13 /// 14 /// To handle exceptions and setjmp/longjmps, this scheme relies on JavaScript's 15 /// try and catch syntax and relevant exception-related libraries implemented 16 /// in JavaScript glue code that will be produced by Emscripten. This is similar 17 /// to the current Emscripten asm.js exception handling in fastcomp. For 18 /// fastcomp's EH / SjLj scheme, see these files in fastcomp LLVM branch: 19 /// (Location: https://github.com/kripken/emscripten-fastcomp) 20 /// lib/Target/JSBackend/NaCl/LowerEmExceptionsPass.cpp 21 /// lib/Target/JSBackend/NaCl/LowerEmSetjmp.cpp 22 /// lib/Target/JSBackend/JSBackend.cpp 23 /// lib/Target/JSBackend/CallHandlers.h 24 /// 25 /// * Exception handling 26 /// This pass lowers invokes and landingpads into library functions in JS glue 27 /// code. Invokes are lowered into function wrappers called invoke wrappers that 28 /// exist in JS side, which wraps the original function call with JS try-catch. 29 /// If an exception occurred, cxa_throw() function in JS side sets some 30 /// variables (see below) so we can check whether an exception occurred from 31 /// wasm code and handle it appropriately. 32 /// 33 /// * Setjmp-longjmp handling 34 /// This pass lowers setjmp to a reasonably-performant approach for emscripten. 35 /// The idea is that each block with a setjmp is broken up into two parts: the 36 /// part containing setjmp and the part right after the setjmp. The latter part 37 /// is either reached from the setjmp, or later from a longjmp. To handle the 38 /// longjmp, all calls that might longjmp are also called using invoke wrappers 39 /// and thus JS / try-catch. JS longjmp() function also sets some variables so 40 /// we can check / whether a longjmp occurred from wasm code. Each block with a 41 /// function call that might longjmp is also split up after the longjmp call. 42 /// After the longjmp call, we check whether a longjmp occurred, and if it did, 43 /// which setjmp it corresponds to, and jump to the right post-setjmp block. 44 /// We assume setjmp-longjmp handling always run after EH handling, which means 45 /// we don't expect any exception-related instructions when SjLj runs. 46 /// FIXME Currently this scheme does not support indirect call of setjmp, 47 /// because of the limitation of the scheme itself. fastcomp does not support it 48 /// either. 49 /// 50 /// In detail, this pass does following things: 51 /// 52 /// 1) Assumes the existence of global variables: __THREW__, __threwValue 53 /// __THREW__ and __threwValue will be set in invoke wrappers 54 /// in JS glue code. For what invoke wrappers are, refer to 3). These 55 /// variables are used for both exceptions and setjmp/longjmps. 56 /// __THREW__ indicates whether an exception or a longjmp occurred or not. 0 57 /// means nothing occurred, 1 means an exception occurred, and other numbers 58 /// mean a longjmp occurred. In the case of longjmp, __threwValue variable 59 /// indicates the corresponding setjmp buffer the longjmp corresponds to. 60 /// 61 /// * Exception handling 62 /// 63 /// 2) We assume the existence of setThrew and setTempRet0/getTempRet0 functions 64 /// at link time. 65 /// The global variables in 1) will exist in wasm address space, 66 /// but their values should be set in JS code, so these functions 67 /// as interfaces to JS glue code. These functions are equivalent to the 68 /// following JS functions, which actually exist in asm.js version of JS 69 /// library. 70 /// 71 /// function setThrew(threw, value) { 72 /// if (__THREW__ == 0) { 73 /// __THREW__ = threw; 74 /// __threwValue = value; 75 /// } 76 /// } 77 // 78 /// setTempRet0 is called from __cxa_find_matching_catch() in JS glue code. 79 /// 80 /// In exception handling, getTempRet0 indicates the type of an exception 81 /// caught, and in setjmp/longjmp, it means the second argument to longjmp 82 /// function. 83 /// 84 /// 3) Lower 85 /// invoke @func(arg1, arg2) to label %invoke.cont unwind label %lpad 86 /// into 87 /// __THREW__ = 0; 88 /// call @__invoke_SIG(func, arg1, arg2) 89 /// %__THREW__.val = __THREW__; 90 /// __THREW__ = 0; 91 /// if (%__THREW__.val == 1) 92 /// goto %lpad 93 /// else 94 /// goto %invoke.cont 95 /// SIG is a mangled string generated based on the LLVM IR-level function 96 /// signature. After LLVM IR types are lowered to the target wasm types, 97 /// the names for these wrappers will change based on wasm types as well, 98 /// as in invoke_vi (function takes an int and returns void). The bodies of 99 /// these wrappers will be generated in JS glue code, and inside those 100 /// wrappers we use JS try-catch to generate actual exception effects. It 101 /// also calls the original callee function. An example wrapper in JS code 102 /// would look like this: 103 /// function invoke_vi(index,a1) { 104 /// try { 105 /// Module["dynCall_vi"](index,a1); // This calls original callee 106 /// } catch(e) { 107 /// if (typeof e !== 'number' && e !== 'longjmp') throw e; 108 /// asm["setThrew"](1, 0); // setThrew is called here 109 /// } 110 /// } 111 /// If an exception is thrown, __THREW__ will be set to true in a wrapper, 112 /// so we can jump to the right BB based on this value. 113 /// 114 /// 4) Lower 115 /// %val = landingpad catch c1 catch c2 catch c3 ... 116 /// ... use %val ... 117 /// into 118 /// %fmc = call @__cxa_find_matching_catch_N(c1, c2, c3, ...) 119 /// %val = {%fmc, getTempRet0()} 120 /// ... use %val ... 121 /// Here N is a number calculated based on the number of clauses. 122 /// setTempRet0 is called from __cxa_find_matching_catch() in JS glue code. 123 /// 124 /// 5) Lower 125 /// resume {%a, %b} 126 /// into 127 /// call @__resumeException(%a) 128 /// where __resumeException() is a function in JS glue code. 129 /// 130 /// 6) Lower 131 /// call @llvm.eh.typeid.for(type) (intrinsic) 132 /// into 133 /// call @llvm_eh_typeid_for(type) 134 /// llvm_eh_typeid_for function will be generated in JS glue code. 135 /// 136 /// * Setjmp / Longjmp handling 137 /// 138 /// In case calls to longjmp() exists 139 /// 140 /// 1) Lower 141 /// longjmp(buf, value) 142 /// into 143 /// emscripten_longjmp_jmpbuf(buf, value) 144 /// emscripten_longjmp_jmpbuf will be lowered to emscripten_longjmp later. 145 /// 146 /// In case calls to setjmp() exists 147 /// 148 /// 2) In the function entry that calls setjmp, initialize setjmpTable and 149 /// sejmpTableSize as follows: 150 /// setjmpTableSize = 4; 151 /// setjmpTable = (int *) malloc(40); 152 /// setjmpTable[0] = 0; 153 /// setjmpTable and setjmpTableSize are used in saveSetjmp() function in JS 154 /// code. 155 /// 156 /// 3) Lower 157 /// setjmp(buf) 158 /// into 159 /// setjmpTable = saveSetjmp(buf, label, setjmpTable, setjmpTableSize); 160 /// setjmpTableSize = getTempRet0(); 161 /// For each dynamic setjmp call, setjmpTable stores its ID (a number which 162 /// is incrementally assigned from 0) and its label (a unique number that 163 /// represents each callsite of setjmp). When we need more entries in 164 /// setjmpTable, it is reallocated in saveSetjmp() in JS code and it will 165 /// return the new table address, and assign the new table size in 166 /// setTempRet0(). saveSetjmp also stores the setjmp's ID into the buffer 167 /// buf. A BB with setjmp is split into two after setjmp call in order to 168 /// make the post-setjmp BB the possible destination of longjmp BB. 169 /// 170 /// 171 /// 4) Lower every call that might longjmp into 172 /// __THREW__ = 0; 173 /// call @__invoke_SIG(func, arg1, arg2) 174 /// %__THREW__.val = __THREW__; 175 /// __THREW__ = 0; 176 /// if (%__THREW__.val != 0 & __threwValue != 0) { 177 /// %label = testSetjmp(mem[%__THREW__.val], setjmpTable, 178 /// setjmpTableSize); 179 /// if (%label == 0) 180 /// emscripten_longjmp(%__THREW__.val, __threwValue); 181 /// setTempRet0(__threwValue); 182 /// } else { 183 /// %label = -1; 184 /// } 185 /// longjmp_result = getTempRet0(); 186 /// switch label { 187 /// label 1: goto post-setjmp BB 1 188 /// label 2: goto post-setjmp BB 2 189 /// ... 190 /// default: goto splitted next BB 191 /// } 192 /// testSetjmp examines setjmpTable to see if there is a matching setjmp 193 /// call. After calling an invoke wrapper, if a longjmp occurred, __THREW__ 194 /// will be the address of matching jmp_buf buffer and __threwValue be the 195 /// second argument to longjmp. mem[__THREW__.val] is a setjmp ID that is 196 /// stored in saveSetjmp. testSetjmp returns a setjmp label, a unique ID to 197 /// each setjmp callsite. Label 0 means this longjmp buffer does not 198 /// correspond to one of the setjmp callsites in this function, so in this 199 /// case we just chain the longjmp to the caller. (Here we call 200 /// emscripten_longjmp, which is different from emscripten_longjmp_jmpbuf. 201 /// emscripten_longjmp_jmpbuf takes jmp_buf as its first argument, while 202 /// emscripten_longjmp takes an int. Both of them will eventually be lowered 203 /// to emscripten_longjmp in s2wasm, but here we need two signatures - we 204 /// can't translate an int value to a jmp_buf.) 205 /// Label -1 means no longjmp occurred. Otherwise we jump to the right 206 /// post-setjmp BB based on the label. 207 /// 208 ///===----------------------------------------------------------------------===// 209 210 #include "WebAssembly.h" 211 #include "llvm/IR/CallSite.h" 212 #include "llvm/IR/Dominators.h" 213 #include "llvm/IR/IRBuilder.h" 214 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 215 #include "llvm/Transforms/Utils/SSAUpdater.h" 216 217 using namespace llvm; 218 219 #define DEBUG_TYPE "wasm-lower-em-ehsjlj" 220 221 static cl::list<std::string> 222 EHWhitelist("emscripten-cxx-exceptions-whitelist", 223 cl::desc("The list of function names in which Emscripten-style " 224 "exception handling is enabled (see emscripten " 225 "EMSCRIPTEN_CATCHING_WHITELIST options)"), 226 cl::CommaSeparated); 227 228 namespace { 229 class WebAssemblyLowerEmscriptenEHSjLj final : public ModulePass { 230 static const char *ResumeFName; 231 static const char *EHTypeIDFName; 232 static const char *EmLongjmpFName; 233 static const char *EmLongjmpJmpbufFName; 234 static const char *SaveSetjmpFName; 235 static const char *TestSetjmpFName; 236 static const char *FindMatchingCatchPrefix; 237 static const char *InvokePrefix; 238 239 bool EnableEH; // Enable exception handling 240 bool EnableSjLj; // Enable setjmp/longjmp handling 241 242 GlobalVariable *ThrewGV = nullptr; 243 GlobalVariable *ThrewValueGV = nullptr; 244 Function *GetTempRet0Func = nullptr; 245 Function *SetTempRet0Func = nullptr; 246 Function *ResumeF = nullptr; 247 Function *EHTypeIDF = nullptr; 248 Function *EmLongjmpF = nullptr; 249 Function *EmLongjmpJmpbufF = nullptr; 250 Function *SaveSetjmpF = nullptr; 251 Function *TestSetjmpF = nullptr; 252 253 // __cxa_find_matching_catch_N functions. 254 // Indexed by the number of clauses in an original landingpad instruction. 255 DenseMap<int, Function *> FindMatchingCatches; 256 // Map of <function signature string, invoke_ wrappers> 257 StringMap<Function *> InvokeWrappers; 258 // Set of whitelisted function names for exception handling 259 std::set<std::string> EHWhitelistSet; 260 261 StringRef getPassName() const override { 262 return "WebAssembly Lower Emscripten Exceptions"; 263 } 264 265 bool runEHOnFunction(Function &F); 266 bool runSjLjOnFunction(Function &F); 267 Function *getFindMatchingCatch(Module &M, unsigned NumClauses); 268 269 template <typename CallOrInvoke> Value *wrapInvoke(CallOrInvoke *CI); 270 void wrapTestSetjmp(BasicBlock *BB, Instruction *InsertPt, Value *Threw, 271 Value *SetjmpTable, Value *SetjmpTableSize, Value *&Label, 272 Value *&LongjmpResult, BasicBlock *&EndBB); 273 template <typename CallOrInvoke> Function *getInvokeWrapper(CallOrInvoke *CI); 274 275 bool areAllExceptionsAllowed() const { return EHWhitelistSet.empty(); } 276 bool canLongjmp(Module &M, const Value *Callee) const; 277 bool isEmAsmCall(Module &M, const Value *Callee) const; 278 279 void rebuildSSA(Function &F); 280 281 public: 282 static char ID; 283 284 WebAssemblyLowerEmscriptenEHSjLj(bool EnableEH = true, bool EnableSjLj = true) 285 : ModulePass(ID), EnableEH(EnableEH), EnableSjLj(EnableSjLj) { 286 EHWhitelistSet.insert(EHWhitelist.begin(), EHWhitelist.end()); 287 } 288 bool runOnModule(Module &M) override; 289 290 void getAnalysisUsage(AnalysisUsage &AU) const override { 291 AU.addRequired<DominatorTreeWrapperPass>(); 292 } 293 }; 294 } // End anonymous namespace 295 296 const char *WebAssemblyLowerEmscriptenEHSjLj::ResumeFName = "__resumeException"; 297 const char *WebAssemblyLowerEmscriptenEHSjLj::EHTypeIDFName = 298 "llvm_eh_typeid_for"; 299 const char *WebAssemblyLowerEmscriptenEHSjLj::EmLongjmpFName = 300 "emscripten_longjmp"; 301 const char *WebAssemblyLowerEmscriptenEHSjLj::EmLongjmpJmpbufFName = 302 "emscripten_longjmp_jmpbuf"; 303 const char *WebAssemblyLowerEmscriptenEHSjLj::SaveSetjmpFName = "saveSetjmp"; 304 const char *WebAssemblyLowerEmscriptenEHSjLj::TestSetjmpFName = "testSetjmp"; 305 const char *WebAssemblyLowerEmscriptenEHSjLj::FindMatchingCatchPrefix = 306 "__cxa_find_matching_catch_"; 307 const char *WebAssemblyLowerEmscriptenEHSjLj::InvokePrefix = "__invoke_"; 308 309 char WebAssemblyLowerEmscriptenEHSjLj::ID = 0; 310 INITIALIZE_PASS(WebAssemblyLowerEmscriptenEHSjLj, DEBUG_TYPE, 311 "WebAssembly Lower Emscripten Exceptions / Setjmp / Longjmp", 312 false, false) 313 314 ModulePass *llvm::createWebAssemblyLowerEmscriptenEHSjLj(bool EnableEH, 315 bool EnableSjLj) { 316 return new WebAssemblyLowerEmscriptenEHSjLj(EnableEH, EnableSjLj); 317 } 318 319 static bool canThrow(const Value *V) { 320 if (const auto *F = dyn_cast<const Function>(V)) { 321 // Intrinsics cannot throw 322 if (F->isIntrinsic()) 323 return false; 324 StringRef Name = F->getName(); 325 // leave setjmp and longjmp (mostly) alone, we process them properly later 326 if (Name == "setjmp" || Name == "longjmp") 327 return false; 328 return !F->doesNotThrow(); 329 } 330 // not a function, so an indirect call - can throw, we can't tell 331 return true; 332 } 333 334 // Get a global variable with the given name. If it doesn't exist declare it, 335 // which will generate an import and asssumes that it will exist at link time. 336 static GlobalVariable *getGlobalVariableI32(Module &M, IRBuilder<> &IRB, 337 const char *Name) { 338 339 auto* GV = dyn_cast<GlobalVariable>(M.getOrInsertGlobal(Name, IRB.getInt32Ty())); 340 if (!GV) 341 report_fatal_error(Twine("unable to create global: ") + Name); 342 343 return GV; 344 } 345 346 // Simple function name mangler. 347 // This function simply takes LLVM's string representation of parameter types 348 // and concatenate them with '_'. There are non-alphanumeric characters but llc 349 // is ok with it, and we need to postprocess these names after the lowering 350 // phase anyway. 351 static std::string getSignature(FunctionType *FTy) { 352 std::string Sig; 353 raw_string_ostream OS(Sig); 354 OS << *FTy->getReturnType(); 355 for (Type *ParamTy : FTy->params()) 356 OS << "_" << *ParamTy; 357 if (FTy->isVarArg()) 358 OS << "_..."; 359 Sig = OS.str(); 360 Sig.erase(remove_if(Sig, isspace), Sig.end()); 361 // When s2wasm parses .s file, a comma means the end of an argument. So a 362 // mangled function name can contain any character but a comma. 363 std::replace(Sig.begin(), Sig.end(), ',', '.'); 364 return Sig; 365 } 366 367 // Returns __cxa_find_matching_catch_N function, where N = NumClauses + 2. 368 // This is because a landingpad instruction contains two more arguments, a 369 // personality function and a cleanup bit, and __cxa_find_matching_catch_N 370 // functions are named after the number of arguments in the original landingpad 371 // instruction. 372 Function * 373 WebAssemblyLowerEmscriptenEHSjLj::getFindMatchingCatch(Module &M, 374 unsigned NumClauses) { 375 if (FindMatchingCatches.count(NumClauses)) 376 return FindMatchingCatches[NumClauses]; 377 PointerType *Int8PtrTy = Type::getInt8PtrTy(M.getContext()); 378 SmallVector<Type *, 16> Args(NumClauses, Int8PtrTy); 379 FunctionType *FTy = FunctionType::get(Int8PtrTy, Args, false); 380 Function *F = 381 Function::Create(FTy, GlobalValue::ExternalLinkage, 382 FindMatchingCatchPrefix + Twine(NumClauses + 2), &M); 383 FindMatchingCatches[NumClauses] = F; 384 return F; 385 } 386 387 // Generate invoke wrapper seqence with preamble and postamble 388 // Preamble: 389 // __THREW__ = 0; 390 // Postamble: 391 // %__THREW__.val = __THREW__; __THREW__ = 0; 392 // Returns %__THREW__.val, which indicates whether an exception is thrown (or 393 // whether longjmp occurred), for future use. 394 template <typename CallOrInvoke> 395 Value *WebAssemblyLowerEmscriptenEHSjLj::wrapInvoke(CallOrInvoke *CI) { 396 LLVMContext &C = CI->getModule()->getContext(); 397 398 // If we are calling a function that is noreturn, we must remove that 399 // attribute. The code we insert here does expect it to return, after we 400 // catch the exception. 401 if (CI->doesNotReturn()) { 402 if (auto *F = dyn_cast<Function>(CI->getCalledValue())) 403 F->removeFnAttr(Attribute::NoReturn); 404 CI->removeAttribute(AttributeList::FunctionIndex, Attribute::NoReturn); 405 } 406 407 IRBuilder<> IRB(C); 408 IRB.SetInsertPoint(CI); 409 410 // Pre-invoke 411 // __THREW__ = 0; 412 IRB.CreateStore(IRB.getInt32(0), ThrewGV); 413 414 // Invoke function wrapper in JavaScript 415 SmallVector<Value *, 16> Args; 416 // Put the pointer to the callee as first argument, so it can be called 417 // within the invoke wrapper later 418 Args.push_back(CI->getCalledValue()); 419 Args.append(CI->arg_begin(), CI->arg_end()); 420 CallInst *NewCall = IRB.CreateCall(getInvokeWrapper(CI), Args); 421 NewCall->takeName(CI); 422 NewCall->setCallingConv(CallingConv::WASM_EmscriptenInvoke); 423 NewCall->setDebugLoc(CI->getDebugLoc()); 424 425 // Because we added the pointer to the callee as first argument, all 426 // argument attribute indices have to be incremented by one. 427 SmallVector<AttributeSet, 8> ArgAttributes; 428 const AttributeList &InvokeAL = CI->getAttributes(); 429 430 // No attributes for the callee pointer. 431 ArgAttributes.push_back(AttributeSet()); 432 // Copy the argument attributes from the original 433 for (unsigned I = 0, E = CI->getNumArgOperands(); I < E; ++I) 434 ArgAttributes.push_back(InvokeAL.getParamAttributes(I)); 435 436 AttrBuilder FnAttrs(InvokeAL.getFnAttributes()); 437 if (FnAttrs.contains(Attribute::AllocSize)) { 438 // The allocsize attribute (if any) referes to parameters by index and needs 439 // to be adjusted. 440 unsigned SizeArg; 441 Optional<unsigned> NEltArg; 442 std::tie(SizeArg, NEltArg) = FnAttrs.getAllocSizeArgs(); 443 SizeArg += 1; 444 if (NEltArg.hasValue()) 445 NEltArg = NEltArg.getValue() + 1; 446 FnAttrs.addAllocSizeAttr(SizeArg, NEltArg); 447 } 448 449 // Reconstruct the AttributesList based on the vector we constructed. 450 AttributeList NewCallAL = 451 AttributeList::get(C, AttributeSet::get(C, FnAttrs), 452 InvokeAL.getRetAttributes(), ArgAttributes); 453 NewCall->setAttributes(NewCallAL); 454 455 CI->replaceAllUsesWith(NewCall); 456 457 // Post-invoke 458 // %__THREW__.val = __THREW__; __THREW__ = 0; 459 Value *Threw = 460 IRB.CreateLoad(IRB.getInt32Ty(), ThrewGV, ThrewGV->getName() + ".val"); 461 IRB.CreateStore(IRB.getInt32(0), ThrewGV); 462 return Threw; 463 } 464 465 // Get matching invoke wrapper based on callee signature 466 template <typename CallOrInvoke> 467 Function *WebAssemblyLowerEmscriptenEHSjLj::getInvokeWrapper(CallOrInvoke *CI) { 468 Module *M = CI->getModule(); 469 SmallVector<Type *, 16> ArgTys; 470 Value *Callee = CI->getCalledValue(); 471 FunctionType *CalleeFTy; 472 if (auto *F = dyn_cast<Function>(Callee)) 473 CalleeFTy = F->getFunctionType(); 474 else { 475 auto *CalleeTy = cast<PointerType>(Callee->getType())->getElementType(); 476 CalleeFTy = dyn_cast<FunctionType>(CalleeTy); 477 } 478 479 std::string Sig = getSignature(CalleeFTy); 480 if (InvokeWrappers.find(Sig) != InvokeWrappers.end()) 481 return InvokeWrappers[Sig]; 482 483 // Put the pointer to the callee as first argument 484 ArgTys.push_back(PointerType::getUnqual(CalleeFTy)); 485 // Add argument types 486 ArgTys.append(CalleeFTy->param_begin(), CalleeFTy->param_end()); 487 488 FunctionType *FTy = FunctionType::get(CalleeFTy->getReturnType(), ArgTys, 489 CalleeFTy->isVarArg()); 490 Function *F = Function::Create(FTy, GlobalValue::ExternalLinkage, 491 InvokePrefix + Sig, M); 492 InvokeWrappers[Sig] = F; 493 return F; 494 } 495 496 bool WebAssemblyLowerEmscriptenEHSjLj::canLongjmp(Module &M, 497 const Value *Callee) const { 498 if (auto *CalleeF = dyn_cast<Function>(Callee)) 499 if (CalleeF->isIntrinsic()) 500 return false; 501 502 // Attempting to transform inline assembly will result in something like: 503 // call void @__invoke_void(void ()* asm ...) 504 // which is invalid because inline assembly blocks do not have addresses 505 // and can't be passed by pointer. The result is a crash with illegal IR. 506 if (isa<InlineAsm>(Callee)) 507 return false; 508 509 // The reason we include malloc/free here is to exclude the malloc/free 510 // calls generated in setjmp prep / cleanup routines. 511 Function *SetjmpF = M.getFunction("setjmp"); 512 Function *MallocF = M.getFunction("malloc"); 513 Function *FreeF = M.getFunction("free"); 514 if (Callee == SetjmpF || Callee == MallocF || Callee == FreeF) 515 return false; 516 517 // There are functions in JS glue code 518 if (Callee == ResumeF || Callee == EHTypeIDF || Callee == SaveSetjmpF || 519 Callee == TestSetjmpF) 520 return false; 521 522 // __cxa_find_matching_catch_N functions cannot longjmp 523 if (Callee->getName().startswith(FindMatchingCatchPrefix)) 524 return false; 525 526 // Exception-catching related functions 527 Function *BeginCatchF = M.getFunction("__cxa_begin_catch"); 528 Function *EndCatchF = M.getFunction("__cxa_end_catch"); 529 Function *AllocExceptionF = M.getFunction("__cxa_allocate_exception"); 530 Function *ThrowF = M.getFunction("__cxa_throw"); 531 Function *TerminateF = M.getFunction("__clang_call_terminate"); 532 if (Callee == BeginCatchF || Callee == EndCatchF || 533 Callee == AllocExceptionF || Callee == ThrowF || Callee == TerminateF || 534 Callee == GetTempRet0Func || Callee == SetTempRet0Func) 535 return false; 536 537 // Otherwise we don't know 538 return true; 539 } 540 541 bool WebAssemblyLowerEmscriptenEHSjLj::isEmAsmCall(Module &M, 542 const Value *Callee) const { 543 // This is an exhaustive list from Emscripten's <emscripten/em_asm.h>. 544 Function *EmAsmConstIntF = M.getFunction("emscripten_asm_const_int"); 545 Function *EmAsmConstDoubleF = M.getFunction("emscripten_asm_const_double"); 546 Function *EmAsmConstIntSyncMainF = 547 M.getFunction("emscripten_asm_const_int_sync_on_main_thread"); 548 Function *EmAsmConstDoubleSyncMainF = 549 M.getFunction("emscripten_asm_const_double_sync_on_main_thread"); 550 Function *EmAsmConstAsyncMainF = 551 M.getFunction("emscripten_asm_const_async_on_main_thread"); 552 553 return Callee == EmAsmConstIntF || Callee == EmAsmConstDoubleF || 554 Callee == EmAsmConstIntSyncMainF || 555 Callee == EmAsmConstDoubleSyncMainF || Callee == EmAsmConstAsyncMainF; 556 } 557 558 // Generate testSetjmp function call seqence with preamble and postamble. 559 // The code this generates is equivalent to the following JavaScript code: 560 // if (%__THREW__.val != 0 & threwValue != 0) { 561 // %label = _testSetjmp(mem[%__THREW__.val], setjmpTable, setjmpTableSize); 562 // if (%label == 0) 563 // emscripten_longjmp(%__THREW__.val, threwValue); 564 // setTempRet0(threwValue); 565 // } else { 566 // %label = -1; 567 // } 568 // %longjmp_result = getTempRet0(); 569 // 570 // As output parameters. returns %label, %longjmp_result, and the BB the last 571 // instruction (%longjmp_result = ...) is in. 572 void WebAssemblyLowerEmscriptenEHSjLj::wrapTestSetjmp( 573 BasicBlock *BB, Instruction *InsertPt, Value *Threw, Value *SetjmpTable, 574 Value *SetjmpTableSize, Value *&Label, Value *&LongjmpResult, 575 BasicBlock *&EndBB) { 576 Function *F = BB->getParent(); 577 LLVMContext &C = BB->getModule()->getContext(); 578 IRBuilder<> IRB(C); 579 IRB.SetInsertPoint(InsertPt); 580 581 // if (%__THREW__.val != 0 & threwValue != 0) 582 IRB.SetInsertPoint(BB); 583 BasicBlock *ThenBB1 = BasicBlock::Create(C, "if.then1", F); 584 BasicBlock *ElseBB1 = BasicBlock::Create(C, "if.else1", F); 585 BasicBlock *EndBB1 = BasicBlock::Create(C, "if.end", F); 586 Value *ThrewCmp = IRB.CreateICmpNE(Threw, IRB.getInt32(0)); 587 Value *ThrewValue = IRB.CreateLoad(IRB.getInt32Ty(), ThrewValueGV, 588 ThrewValueGV->getName() + ".val"); 589 Value *ThrewValueCmp = IRB.CreateICmpNE(ThrewValue, IRB.getInt32(0)); 590 Value *Cmp1 = IRB.CreateAnd(ThrewCmp, ThrewValueCmp, "cmp1"); 591 IRB.CreateCondBr(Cmp1, ThenBB1, ElseBB1); 592 593 // %label = _testSetjmp(mem[%__THREW__.val], _setjmpTable, _setjmpTableSize); 594 // if (%label == 0) 595 IRB.SetInsertPoint(ThenBB1); 596 BasicBlock *ThenBB2 = BasicBlock::Create(C, "if.then2", F); 597 BasicBlock *EndBB2 = BasicBlock::Create(C, "if.end2", F); 598 Value *ThrewInt = IRB.CreateIntToPtr(Threw, Type::getInt32PtrTy(C), 599 Threw->getName() + ".i32p"); 600 Value *LoadedThrew = IRB.CreateLoad(IRB.getInt32Ty(), ThrewInt, 601 ThrewInt->getName() + ".loaded"); 602 Value *ThenLabel = IRB.CreateCall( 603 TestSetjmpF, {LoadedThrew, SetjmpTable, SetjmpTableSize}, "label"); 604 Value *Cmp2 = IRB.CreateICmpEQ(ThenLabel, IRB.getInt32(0)); 605 IRB.CreateCondBr(Cmp2, ThenBB2, EndBB2); 606 607 // emscripten_longjmp(%__THREW__.val, threwValue); 608 IRB.SetInsertPoint(ThenBB2); 609 IRB.CreateCall(EmLongjmpF, {Threw, ThrewValue}); 610 IRB.CreateUnreachable(); 611 612 // setTempRet0(threwValue); 613 IRB.SetInsertPoint(EndBB2); 614 IRB.CreateCall(SetTempRet0Func, ThrewValue); 615 IRB.CreateBr(EndBB1); 616 617 IRB.SetInsertPoint(ElseBB1); 618 IRB.CreateBr(EndBB1); 619 620 // longjmp_result = getTempRet0(); 621 IRB.SetInsertPoint(EndBB1); 622 PHINode *LabelPHI = IRB.CreatePHI(IRB.getInt32Ty(), 2, "label"); 623 LabelPHI->addIncoming(ThenLabel, EndBB2); 624 625 LabelPHI->addIncoming(IRB.getInt32(-1), ElseBB1); 626 627 // Output parameter assignment 628 Label = LabelPHI; 629 EndBB = EndBB1; 630 LongjmpResult = IRB.CreateCall(GetTempRet0Func, None, "longjmp_result"); 631 } 632 633 void WebAssemblyLowerEmscriptenEHSjLj::rebuildSSA(Function &F) { 634 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree(); 635 DT.recalculate(F); // CFG has been changed 636 SSAUpdater SSA; 637 for (BasicBlock &BB : F) { 638 for (Instruction &I : BB) { 639 for (auto UI = I.use_begin(), UE = I.use_end(); UI != UE;) { 640 Use &U = *UI; 641 ++UI; 642 SSA.Initialize(I.getType(), I.getName()); 643 SSA.AddAvailableValue(&BB, &I); 644 auto *User = cast<Instruction>(U.getUser()); 645 if (User->getParent() == &BB) 646 continue; 647 648 if (auto *UserPN = dyn_cast<PHINode>(User)) 649 if (UserPN->getIncomingBlock(U) == &BB) 650 continue; 651 652 if (DT.dominates(&I, User)) 653 continue; 654 SSA.RewriteUseAfterInsertions(U); 655 } 656 } 657 } 658 } 659 660 bool WebAssemblyLowerEmscriptenEHSjLj::runOnModule(Module &M) { 661 LLVM_DEBUG(dbgs() << "********** Lower Emscripten EH & SjLj **********\n"); 662 663 LLVMContext &C = M.getContext(); 664 IRBuilder<> IRB(C); 665 666 Function *SetjmpF = M.getFunction("setjmp"); 667 Function *LongjmpF = M.getFunction("longjmp"); 668 bool SetjmpUsed = SetjmpF && !SetjmpF->use_empty(); 669 bool LongjmpUsed = LongjmpF && !LongjmpF->use_empty(); 670 bool DoSjLj = EnableSjLj && (SetjmpUsed || LongjmpUsed); 671 672 // Declare (or get) global variables __THREW__, __threwValue, and 673 // getTempRet0/setTempRet0 function which are used in common for both 674 // exception handling and setjmp/longjmp handling 675 ThrewGV = getGlobalVariableI32(M, IRB, "__THREW__"); 676 ThrewValueGV = getGlobalVariableI32(M, IRB, "__threwValue"); 677 GetTempRet0Func = 678 Function::Create(FunctionType::get(IRB.getInt32Ty(), false), 679 GlobalValue::ExternalLinkage, "getTempRet0", &M); 680 SetTempRet0Func = Function::Create( 681 FunctionType::get(IRB.getVoidTy(), IRB.getInt32Ty(), false), 682 GlobalValue::ExternalLinkage, "setTempRet0", &M); 683 GetTempRet0Func->setDoesNotThrow(); 684 SetTempRet0Func->setDoesNotThrow(); 685 686 bool Changed = false; 687 688 // Exception handling 689 if (EnableEH) { 690 // Register __resumeException function 691 FunctionType *ResumeFTy = 692 FunctionType::get(IRB.getVoidTy(), IRB.getInt8PtrTy(), false); 693 ResumeF = Function::Create(ResumeFTy, GlobalValue::ExternalLinkage, 694 ResumeFName, &M); 695 696 // Register llvm_eh_typeid_for function 697 FunctionType *EHTypeIDTy = 698 FunctionType::get(IRB.getInt32Ty(), IRB.getInt8PtrTy(), false); 699 EHTypeIDF = Function::Create(EHTypeIDTy, GlobalValue::ExternalLinkage, 700 EHTypeIDFName, &M); 701 702 for (Function &F : M) { 703 if (F.isDeclaration()) 704 continue; 705 Changed |= runEHOnFunction(F); 706 } 707 } 708 709 // Setjmp/longjmp handling 710 if (DoSjLj) { 711 Changed = true; // We have setjmp or longjmp somewhere 712 713 if (LongjmpF) { 714 // Replace all uses of longjmp with emscripten_longjmp_jmpbuf, which is 715 // defined in JS code 716 EmLongjmpJmpbufF = Function::Create(LongjmpF->getFunctionType(), 717 GlobalValue::ExternalLinkage, 718 EmLongjmpJmpbufFName, &M); 719 720 LongjmpF->replaceAllUsesWith(EmLongjmpJmpbufF); 721 } 722 723 if (SetjmpF) { 724 // Register saveSetjmp function 725 FunctionType *SetjmpFTy = SetjmpF->getFunctionType(); 726 SmallVector<Type *, 4> Params = {SetjmpFTy->getParamType(0), 727 IRB.getInt32Ty(), Type::getInt32PtrTy(C), 728 IRB.getInt32Ty()}; 729 FunctionType *FTy = 730 FunctionType::get(Type::getInt32PtrTy(C), Params, false); 731 SaveSetjmpF = Function::Create(FTy, GlobalValue::ExternalLinkage, 732 SaveSetjmpFName, &M); 733 734 // Register testSetjmp function 735 Params = {IRB.getInt32Ty(), Type::getInt32PtrTy(C), IRB.getInt32Ty()}; 736 FTy = FunctionType::get(IRB.getInt32Ty(), Params, false); 737 TestSetjmpF = Function::Create(FTy, GlobalValue::ExternalLinkage, 738 TestSetjmpFName, &M); 739 740 FTy = FunctionType::get(IRB.getVoidTy(), 741 {IRB.getInt32Ty(), IRB.getInt32Ty()}, false); 742 EmLongjmpF = Function::Create(FTy, GlobalValue::ExternalLinkage, 743 EmLongjmpFName, &M); 744 745 // Only traverse functions that uses setjmp in order not to insert 746 // unnecessary prep / cleanup code in every function 747 SmallPtrSet<Function *, 8> SetjmpUsers; 748 for (User *U : SetjmpF->users()) { 749 auto *UI = cast<Instruction>(U); 750 SetjmpUsers.insert(UI->getFunction()); 751 } 752 for (Function *F : SetjmpUsers) 753 runSjLjOnFunction(*F); 754 } 755 } 756 757 if (!Changed) { 758 // Delete unused global variables and functions 759 if (ResumeF) 760 ResumeF->eraseFromParent(); 761 if (EHTypeIDF) 762 EHTypeIDF->eraseFromParent(); 763 if (EmLongjmpF) 764 EmLongjmpF->eraseFromParent(); 765 if (SaveSetjmpF) 766 SaveSetjmpF->eraseFromParent(); 767 if (TestSetjmpF) 768 TestSetjmpF->eraseFromParent(); 769 return false; 770 } 771 772 return true; 773 } 774 775 bool WebAssemblyLowerEmscriptenEHSjLj::runEHOnFunction(Function &F) { 776 Module &M = *F.getParent(); 777 LLVMContext &C = F.getContext(); 778 IRBuilder<> IRB(C); 779 bool Changed = false; 780 SmallVector<Instruction *, 64> ToErase; 781 SmallPtrSet<LandingPadInst *, 32> LandingPads; 782 bool AllowExceptions = 783 areAllExceptionsAllowed() || EHWhitelistSet.count(F.getName()); 784 785 for (BasicBlock &BB : F) { 786 auto *II = dyn_cast<InvokeInst>(BB.getTerminator()); 787 if (!II) 788 continue; 789 Changed = true; 790 LandingPads.insert(II->getLandingPadInst()); 791 IRB.SetInsertPoint(II); 792 793 bool NeedInvoke = AllowExceptions && canThrow(II->getCalledValue()); 794 if (NeedInvoke) { 795 // Wrap invoke with invoke wrapper and generate preamble/postamble 796 Value *Threw = wrapInvoke(II); 797 ToErase.push_back(II); 798 799 // Insert a branch based on __THREW__ variable 800 Value *Cmp = IRB.CreateICmpEQ(Threw, IRB.getInt32(1), "cmp"); 801 IRB.CreateCondBr(Cmp, II->getUnwindDest(), II->getNormalDest()); 802 803 } else { 804 // This can't throw, and we don't need this invoke, just replace it with a 805 // call+branch 806 SmallVector<Value *, 16> Args(II->arg_begin(), II->arg_end()); 807 CallInst *NewCall = 808 IRB.CreateCall(II->getFunctionType(), II->getCalledValue(), Args); 809 NewCall->takeName(II); 810 NewCall->setCallingConv(II->getCallingConv()); 811 NewCall->setDebugLoc(II->getDebugLoc()); 812 NewCall->setAttributes(II->getAttributes()); 813 II->replaceAllUsesWith(NewCall); 814 ToErase.push_back(II); 815 816 IRB.CreateBr(II->getNormalDest()); 817 818 // Remove any PHI node entries from the exception destination 819 II->getUnwindDest()->removePredecessor(&BB); 820 } 821 } 822 823 // Process resume instructions 824 for (BasicBlock &BB : F) { 825 // Scan the body of the basic block for resumes 826 for (Instruction &I : BB) { 827 auto *RI = dyn_cast<ResumeInst>(&I); 828 if (!RI) 829 continue; 830 831 // Split the input into legal values 832 Value *Input = RI->getValue(); 833 IRB.SetInsertPoint(RI); 834 Value *Low = IRB.CreateExtractValue(Input, 0, "low"); 835 // Create a call to __resumeException function 836 IRB.CreateCall(ResumeF, {Low}); 837 // Add a terminator to the block 838 IRB.CreateUnreachable(); 839 ToErase.push_back(RI); 840 } 841 } 842 843 // Process llvm.eh.typeid.for intrinsics 844 for (BasicBlock &BB : F) { 845 for (Instruction &I : BB) { 846 auto *CI = dyn_cast<CallInst>(&I); 847 if (!CI) 848 continue; 849 const Function *Callee = CI->getCalledFunction(); 850 if (!Callee) 851 continue; 852 if (Callee->getIntrinsicID() != Intrinsic::eh_typeid_for) 853 continue; 854 855 IRB.SetInsertPoint(CI); 856 CallInst *NewCI = 857 IRB.CreateCall(EHTypeIDF, CI->getArgOperand(0), "typeid"); 858 CI->replaceAllUsesWith(NewCI); 859 ToErase.push_back(CI); 860 } 861 } 862 863 // Look for orphan landingpads, can occur in blocks with no predecessors 864 for (BasicBlock &BB : F) { 865 Instruction *I = BB.getFirstNonPHI(); 866 if (auto *LPI = dyn_cast<LandingPadInst>(I)) 867 LandingPads.insert(LPI); 868 } 869 870 // Handle all the landingpad for this function together, as multiple invokes 871 // may share a single lp 872 for (LandingPadInst *LPI : LandingPads) { 873 IRB.SetInsertPoint(LPI); 874 SmallVector<Value *, 16> FMCArgs; 875 for (unsigned I = 0, E = LPI->getNumClauses(); I < E; ++I) { 876 Constant *Clause = LPI->getClause(I); 877 // As a temporary workaround for the lack of aggregate varargs support 878 // in the interface between JS and wasm, break out filter operands into 879 // their component elements. 880 if (LPI->isFilter(I)) { 881 auto *ATy = cast<ArrayType>(Clause->getType()); 882 for (unsigned J = 0, E = ATy->getNumElements(); J < E; ++J) { 883 Value *EV = IRB.CreateExtractValue(Clause, makeArrayRef(J), "filter"); 884 FMCArgs.push_back(EV); 885 } 886 } else 887 FMCArgs.push_back(Clause); 888 } 889 890 // Create a call to __cxa_find_matching_catch_N function 891 Function *FMCF = getFindMatchingCatch(M, FMCArgs.size()); 892 CallInst *FMCI = IRB.CreateCall(FMCF, FMCArgs, "fmc"); 893 Value *Undef = UndefValue::get(LPI->getType()); 894 Value *Pair0 = IRB.CreateInsertValue(Undef, FMCI, 0, "pair0"); 895 Value *TempRet0 = IRB.CreateCall(GetTempRet0Func, None, "tempret0"); 896 Value *Pair1 = IRB.CreateInsertValue(Pair0, TempRet0, 1, "pair1"); 897 898 LPI->replaceAllUsesWith(Pair1); 899 ToErase.push_back(LPI); 900 } 901 902 // Erase everything we no longer need in this function 903 for (Instruction *I : ToErase) 904 I->eraseFromParent(); 905 906 return Changed; 907 } 908 909 bool WebAssemblyLowerEmscriptenEHSjLj::runSjLjOnFunction(Function &F) { 910 Module &M = *F.getParent(); 911 LLVMContext &C = F.getContext(); 912 IRBuilder<> IRB(C); 913 SmallVector<Instruction *, 64> ToErase; 914 // Vector of %setjmpTable values 915 std::vector<Instruction *> SetjmpTableInsts; 916 // Vector of %setjmpTableSize values 917 std::vector<Instruction *> SetjmpTableSizeInsts; 918 919 // Setjmp preparation 920 921 // This instruction effectively means %setjmpTableSize = 4. 922 // We create this as an instruction intentionally, and we don't want to fold 923 // this instruction to a constant 4, because this value will be used in 924 // SSAUpdater.AddAvailableValue(...) later. 925 BasicBlock &EntryBB = F.getEntryBlock(); 926 BinaryOperator *SetjmpTableSize = BinaryOperator::Create( 927 Instruction::Add, IRB.getInt32(4), IRB.getInt32(0), "setjmpTableSize", 928 &*EntryBB.getFirstInsertionPt()); 929 // setjmpTable = (int *) malloc(40); 930 Instruction *SetjmpTable = CallInst::CreateMalloc( 931 SetjmpTableSize, IRB.getInt32Ty(), IRB.getInt32Ty(), IRB.getInt32(40), 932 nullptr, nullptr, "setjmpTable"); 933 // setjmpTable[0] = 0; 934 IRB.SetInsertPoint(SetjmpTableSize); 935 IRB.CreateStore(IRB.getInt32(0), SetjmpTable); 936 SetjmpTableInsts.push_back(SetjmpTable); 937 SetjmpTableSizeInsts.push_back(SetjmpTableSize); 938 939 // Setjmp transformation 940 std::vector<PHINode *> SetjmpRetPHIs; 941 Function *SetjmpF = M.getFunction("setjmp"); 942 for (User *U : SetjmpF->users()) { 943 auto *CI = dyn_cast<CallInst>(U); 944 if (!CI) 945 report_fatal_error("Does not support indirect calls to setjmp"); 946 947 BasicBlock *BB = CI->getParent(); 948 if (BB->getParent() != &F) // in other function 949 continue; 950 951 // The tail is everything right after the call, and will be reached once 952 // when setjmp is called, and later when longjmp returns to the setjmp 953 BasicBlock *Tail = SplitBlock(BB, CI->getNextNode()); 954 // Add a phi to the tail, which will be the output of setjmp, which 955 // indicates if this is the first call or a longjmp back. The phi directly 956 // uses the right value based on where we arrive from 957 IRB.SetInsertPoint(Tail->getFirstNonPHI()); 958 PHINode *SetjmpRet = IRB.CreatePHI(IRB.getInt32Ty(), 2, "setjmp.ret"); 959 960 // setjmp initial call returns 0 961 SetjmpRet->addIncoming(IRB.getInt32(0), BB); 962 // The proper output is now this, not the setjmp call itself 963 CI->replaceAllUsesWith(SetjmpRet); 964 // longjmp returns to the setjmp will add themselves to this phi 965 SetjmpRetPHIs.push_back(SetjmpRet); 966 967 // Fix call target 968 // Our index in the function is our place in the array + 1 to avoid index 969 // 0, because index 0 means the longjmp is not ours to handle. 970 IRB.SetInsertPoint(CI); 971 Value *Args[] = {CI->getArgOperand(0), IRB.getInt32(SetjmpRetPHIs.size()), 972 SetjmpTable, SetjmpTableSize}; 973 Instruction *NewSetjmpTable = 974 IRB.CreateCall(SaveSetjmpF, Args, "setjmpTable"); 975 Instruction *NewSetjmpTableSize = 976 IRB.CreateCall(GetTempRet0Func, None, "setjmpTableSize"); 977 SetjmpTableInsts.push_back(NewSetjmpTable); 978 SetjmpTableSizeInsts.push_back(NewSetjmpTableSize); 979 ToErase.push_back(CI); 980 } 981 982 // Update each call that can longjmp so it can return to a setjmp where 983 // relevant. 984 985 // Because we are creating new BBs while processing and don't want to make 986 // all these newly created BBs candidates again for longjmp processing, we 987 // first make the vector of candidate BBs. 988 std::vector<BasicBlock *> BBs; 989 for (BasicBlock &BB : F) 990 BBs.push_back(&BB); 991 992 // BBs.size() will change within the loop, so we query it every time 993 for (unsigned I = 0; I < BBs.size(); I++) { 994 BasicBlock *BB = BBs[I]; 995 for (Instruction &I : *BB) { 996 assert(!isa<InvokeInst>(&I)); 997 auto *CI = dyn_cast<CallInst>(&I); 998 if (!CI) 999 continue; 1000 1001 const Value *Callee = CI->getCalledValue(); 1002 if (!canLongjmp(M, Callee)) 1003 continue; 1004 if (isEmAsmCall(M, Callee)) 1005 report_fatal_error("Cannot use EM_ASM* alongside setjmp/longjmp in " + 1006 F.getName() + 1007 ". Please consider using EM_JS, or move the " 1008 "EM_ASM into another function.", 1009 false); 1010 1011 Value *Threw = nullptr; 1012 BasicBlock *Tail; 1013 if (Callee->getName().startswith(InvokePrefix)) { 1014 // If invoke wrapper has already been generated for this call in 1015 // previous EH phase, search for the load instruction 1016 // %__THREW__.val = __THREW__; 1017 // in postamble after the invoke wrapper call 1018 LoadInst *ThrewLI = nullptr; 1019 StoreInst *ThrewResetSI = nullptr; 1020 for (auto I = std::next(BasicBlock::iterator(CI)), IE = BB->end(); 1021 I != IE; ++I) { 1022 if (auto *LI = dyn_cast<LoadInst>(I)) 1023 if (auto *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand())) 1024 if (GV == ThrewGV) { 1025 Threw = ThrewLI = LI; 1026 break; 1027 } 1028 } 1029 // Search for the store instruction after the load above 1030 // __THREW__ = 0; 1031 for (auto I = std::next(BasicBlock::iterator(ThrewLI)), IE = BB->end(); 1032 I != IE; ++I) { 1033 if (auto *SI = dyn_cast<StoreInst>(I)) 1034 if (auto *GV = dyn_cast<GlobalVariable>(SI->getPointerOperand())) 1035 if (GV == ThrewGV && SI->getValueOperand() == IRB.getInt32(0)) { 1036 ThrewResetSI = SI; 1037 break; 1038 } 1039 } 1040 assert(Threw && ThrewLI && "Cannot find __THREW__ load after invoke"); 1041 assert(ThrewResetSI && "Cannot find __THREW__ store after invoke"); 1042 Tail = SplitBlock(BB, ThrewResetSI->getNextNode()); 1043 1044 } else { 1045 // Wrap call with invoke wrapper and generate preamble/postamble 1046 Threw = wrapInvoke(CI); 1047 ToErase.push_back(CI); 1048 Tail = SplitBlock(BB, CI->getNextNode()); 1049 } 1050 1051 // We need to replace the terminator in Tail - SplitBlock makes BB go 1052 // straight to Tail, we need to check if a longjmp occurred, and go to the 1053 // right setjmp-tail if so 1054 ToErase.push_back(BB->getTerminator()); 1055 1056 // Generate a function call to testSetjmp function and preamble/postamble 1057 // code to figure out (1) whether longjmp occurred (2) if longjmp 1058 // occurred, which setjmp it corresponds to 1059 Value *Label = nullptr; 1060 Value *LongjmpResult = nullptr; 1061 BasicBlock *EndBB = nullptr; 1062 wrapTestSetjmp(BB, CI, Threw, SetjmpTable, SetjmpTableSize, Label, 1063 LongjmpResult, EndBB); 1064 assert(Label && LongjmpResult && EndBB); 1065 1066 // Create switch instruction 1067 IRB.SetInsertPoint(EndBB); 1068 SwitchInst *SI = IRB.CreateSwitch(Label, Tail, SetjmpRetPHIs.size()); 1069 // -1 means no longjmp happened, continue normally (will hit the default 1070 // switch case). 0 means a longjmp that is not ours to handle, needs a 1071 // rethrow. Otherwise the index is the same as the index in P+1 (to avoid 1072 // 0). 1073 for (unsigned I = 0; I < SetjmpRetPHIs.size(); I++) { 1074 SI->addCase(IRB.getInt32(I + 1), SetjmpRetPHIs[I]->getParent()); 1075 SetjmpRetPHIs[I]->addIncoming(LongjmpResult, EndBB); 1076 } 1077 1078 // We are splitting the block here, and must continue to find other calls 1079 // in the block - which is now split. so continue to traverse in the Tail 1080 BBs.push_back(Tail); 1081 } 1082 } 1083 1084 // Erase everything we no longer need in this function 1085 for (Instruction *I : ToErase) 1086 I->eraseFromParent(); 1087 1088 // Free setjmpTable buffer before each return instruction 1089 for (BasicBlock &BB : F) { 1090 Instruction *TI = BB.getTerminator(); 1091 if (isa<ReturnInst>(TI)) 1092 CallInst::CreateFree(SetjmpTable, TI); 1093 } 1094 1095 // Every call to saveSetjmp can change setjmpTable and setjmpTableSize 1096 // (when buffer reallocation occurs) 1097 // entry: 1098 // setjmpTableSize = 4; 1099 // setjmpTable = (int *) malloc(40); 1100 // setjmpTable[0] = 0; 1101 // ... 1102 // somebb: 1103 // setjmpTable = saveSetjmp(buf, label, setjmpTable, setjmpTableSize); 1104 // setjmpTableSize = getTempRet0(); 1105 // So we need to make sure the SSA for these variables is valid so that every 1106 // saveSetjmp and testSetjmp calls have the correct arguments. 1107 SSAUpdater SetjmpTableSSA; 1108 SSAUpdater SetjmpTableSizeSSA; 1109 SetjmpTableSSA.Initialize(Type::getInt32PtrTy(C), "setjmpTable"); 1110 SetjmpTableSizeSSA.Initialize(Type::getInt32Ty(C), "setjmpTableSize"); 1111 for (Instruction *I : SetjmpTableInsts) 1112 SetjmpTableSSA.AddAvailableValue(I->getParent(), I); 1113 for (Instruction *I : SetjmpTableSizeInsts) 1114 SetjmpTableSizeSSA.AddAvailableValue(I->getParent(), I); 1115 1116 for (auto UI = SetjmpTable->use_begin(), UE = SetjmpTable->use_end(); 1117 UI != UE;) { 1118 // Grab the use before incrementing the iterator. 1119 Use &U = *UI; 1120 // Increment the iterator before removing the use from the list. 1121 ++UI; 1122 if (auto *I = dyn_cast<Instruction>(U.getUser())) 1123 if (I->getParent() != &EntryBB) 1124 SetjmpTableSSA.RewriteUse(U); 1125 } 1126 for (auto UI = SetjmpTableSize->use_begin(), UE = SetjmpTableSize->use_end(); 1127 UI != UE;) { 1128 Use &U = *UI; 1129 ++UI; 1130 if (auto *I = dyn_cast<Instruction>(U.getUser())) 1131 if (I->getParent() != &EntryBB) 1132 SetjmpTableSizeSSA.RewriteUse(U); 1133 } 1134 1135 // Finally, our modifications to the cfg can break dominance of SSA variables. 1136 // For example, in this code, 1137 // if (x()) { .. setjmp() .. } 1138 // if (y()) { .. longjmp() .. } 1139 // We must split the longjmp block, and it can jump into the block splitted 1140 // from setjmp one. But that means that when we split the setjmp block, it's 1141 // first part no longer dominates its second part - there is a theoretically 1142 // possible control flow path where x() is false, then y() is true and we 1143 // reach the second part of the setjmp block, without ever reaching the first 1144 // part. So, we rebuild SSA form here. 1145 rebuildSSA(F); 1146 return true; 1147 } 1148