Target/WebAssembly/WebAssemblyLowerEmscriptenEHSjLj.cpp

//=== WebAssemblyLowerEmscriptenEHSjLj.cpp - Lower exceptions for Emscripten =//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
///
/// \file
/// This file lowers exception-related instructions and setjmp/longjmp
/// function calls in order to use Emscripten's JavaScript try and catch
/// mechanism.
///
/// To handle exceptions and setjmp/longjmps, this scheme relies on JavaScript's
/// try and catch syntax and relevant exception-related libraries implemented
/// in JavaScript glue code that will be produced by Emscripten.
///
/// * Exception handling
/// This pass lowers invokes and landingpads into library functions in JS glue
/// code. Invokes are lowered into function wrappers called invoke wrappers that
/// exist in JS side, which wraps the original function call with JS try-catch.
/// If an exception occurred, cxa_throw() function in JS side sets some
/// variables (see below) so we can check whether an exception occurred from
/// wasm code and handle it appropriately.
///
/// * Setjmp-longjmp handling
/// This pass lowers setjmp to a reasonably-performant approach for emscripten.
/// The idea is that each block with a setjmp is broken up into two parts: the
/// part containing setjmp and the part right after the setjmp. The latter part
/// is either reached from the setjmp, or later from a longjmp. To handle the
/// longjmp, all calls that might longjmp are also called using invoke wrappers
/// and thus JS / try-catch. JS longjmp() function also sets some variables so
/// we can check / whether a longjmp occurred from wasm code. Each block with a
/// function call that might longjmp is also split up after the longjmp call.
/// After the longjmp call, we check whether a longjmp occurred, and if it did,
/// which setjmp it corresponds to, and jump to the right post-setjmp block.
/// We assume setjmp-longjmp handling always run after EH handling, which means
/// we don't expect any exception-related instructions when SjLj runs.
/// FIXME Currently this scheme does not support indirect call of setjmp,
/// because of the limitation of the scheme itself. fastcomp does not support it
/// either.
///
/// In detail, this pass does following things:
///
/// 1) Assumes the existence of global variables: __THREW__, __threwValue
///    __THREW__ and __threwValue are defined in compiler-rt in Emscripten.
///    These variables are used for both exceptions and setjmp/longjmps.
///    __THREW__ indicates whether an exception or a longjmp occurred or not. 0
///    means nothing occurred, 1 means an exception occurred, and other numbers
///    mean a longjmp occurred. In the case of longjmp, __THREW__ variable
///    indicates the corresponding setjmp buffer the longjmp corresponds to.
///    __threwValue is 0 for exceptions, and the argument to longjmp in case of
///    longjmp.
///
/// * Exception handling
///
/// 2) We assume the existence of setThrew and setTempRet0/getTempRet0 functions
///    at link time. setThrew exists in Emscripten's compiler-rt:
///
///    void setThrew(uintptr_t threw, int value) {
///      if (__THREW__ == 0) {
///        __THREW__ = threw;
///        __threwValue = value;
///      }
///    }
//
///    setTempRet0 is called from __cxa_find_matching_catch() in JS glue code.
///    In exception handling, getTempRet0 indicates the type of an exception
///    caught, and in setjmp/longjmp, it means the second argument to longjmp
///    function.
///
/// 3) Lower
///      invoke @func(arg1, arg2) to label %invoke.cont unwind label %lpad
///    into
///      __THREW__ = 0;
///      call @__invoke_SIG(func, arg1, arg2)
///      %__THREW__.val = __THREW__;
///      __THREW__ = 0;
///      if (%__THREW__.val == 1)
///        goto %lpad
///      else
///         goto %invoke.cont
///    SIG is a mangled string generated based on the LLVM IR-level function
///    signature. After LLVM IR types are lowered to the target wasm types,
///    the names for these wrappers will change based on wasm types as well,
///    as in invoke_vi (function takes an int and returns void). The bodies of
///    these wrappers will be generated in JS glue code, and inside those
///    wrappers we use JS try-catch to generate actual exception effects. It
///    also calls the original callee function. An example wrapper in JS code
///    would look like this:
///      function invoke_vi(index,a1) {
///        try {
///          Module["dynCall_vi"](index,a1); // This calls original callee
///        } catch(e) {
///          if (typeof e !== 'number' && e !== 'longjmp') throw e;
///          _setThrew(1, 0); // setThrew is called here
///        }
///      }
///    If an exception is thrown, __THREW__ will be set to true in a wrapper,
///    so we can jump to the right BB based on this value.
///
/// 4) Lower
///      %val = landingpad catch c1 catch c2 catch c3 ...
///      ... use %val ...
///    into
///      %fmc = call @__cxa_find_matching_catch_N(c1, c2, c3, ...)
///      %val = {%fmc, getTempRet0()}
///      ... use %val ...
///    Here N is a number calculated based on the number of clauses.
///    setTempRet0 is called from __cxa_find_matching_catch() in JS glue code.
///
/// 5) Lower
///      resume {%a, %b}
///    into
///      call @__resumeException(%a)
///    where __resumeException() is a function in JS glue code.
///
/// 6) Lower
///      call @llvm.eh.typeid.for(type) (intrinsic)
///    into
///      call @llvm_eh_typeid_for(type)
///    llvm_eh_typeid_for function will be generated in JS glue code.
///
/// * Setjmp / Longjmp handling
///
/// In case calls to longjmp() exists
///
/// 1) Lower
///      longjmp(buf, value)
///    into
///      emscripten_longjmp(buf, value)
///
/// In case calls to setjmp() exists
///
/// 2) In the function entry that calls setjmp, initialize setjmpTable and
///    sejmpTableSize as follows:
///      setjmpTableSize = 4;
///      setjmpTable = (int *) malloc(40);
///      setjmpTable[0] = 0;
///    setjmpTable and setjmpTableSize are used to call saveSetjmp() function in
///    Emscripten compiler-rt.
///
/// 3) Lower
///      setjmp(buf)
///    into
///      setjmpTable = saveSetjmp(buf, label, setjmpTable, setjmpTableSize);
///      setjmpTableSize = getTempRet0();
///    For each dynamic setjmp call, setjmpTable stores its ID (a number which
///    is incrementally assigned from 0) and its label (a unique number that
///    represents each callsite of setjmp). When we need more entries in
///    setjmpTable, it is reallocated in saveSetjmp() in Emscripten's
///    compiler-rt and it will return the new table address, and assign the new
///    table size in setTempRet0(). saveSetjmp also stores the setjmp's ID into
///    the buffer buf. A BB with setjmp is split into two after setjmp call in
///    order to make the post-setjmp BB the possible destination of longjmp BB.
///
///
/// 4) Lower every call that might longjmp into
///      __THREW__ = 0;
///      call @__invoke_SIG(func, arg1, arg2)
///      %__THREW__.val = __THREW__;
///      __THREW__ = 0;
///      %__threwValue.val = __threwValue;
///      if (%__THREW__.val != 0 & %__threwValue.val != 0) {
///        %label = testSetjmp(mem[%__THREW__.val], setjmpTable,
///                            setjmpTableSize);
///        if (%label == 0)
///          emscripten_longjmp(%__THREW__.val, %__threwValue.val);
///        setTempRet0(%__threwValue.val);
///      } else {
///        %label = -1;
///      }
///      longjmp_result = getTempRet0();
///      switch label {
///        label 1: goto post-setjmp BB 1
///        label 2: goto post-setjmp BB 2
///        ...
///        default: goto splitted next BB
///      }
///    testSetjmp examines setjmpTable to see if there is a matching setjmp
///    call. After calling an invoke wrapper, if a longjmp occurred, __THREW__
///    will be the address of matching jmp_buf buffer and __threwValue be the
///    second argument to longjmp. mem[%__THREW__.val] is a setjmp ID that is
///    stored in saveSetjmp. testSetjmp returns a setjmp label, a unique ID to
///    each setjmp callsite. Label 0 means this longjmp buffer does not
///    correspond to one of the setjmp callsites in this function, so in this
///    case we just chain the longjmp to the caller. Label -1 means no longjmp
///    occurred. Otherwise we jump to the right post-setjmp BB based on the
///    label.
///
///===----------------------------------------------------------------------===//

#include "WebAssembly.h"
#include "WebAssemblyTargetMachine.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/SSAUpdater.h"

using namespace llvm;

#define DEBUG_TYPE "wasm-lower-em-ehsjlj"

static cl::list<std::string>
    EHAllowlist("emscripten-cxx-exceptions-allowed",
                cl::desc("The list of function names in which Emscripten-style "
                         "exception handling is enabled (see emscripten "
                         "EMSCRIPTEN_CATCHING_ALLOWED options)"),
                cl::CommaSeparated);

namespace {
class WebAssemblyLowerEmscriptenEHSjLj final : public ModulePass {
  bool EnableEH;   // Enable exception handling
  bool EnableSjLj; // Enable setjmp/longjmp handling
  bool DoSjLj;     // Whether we actually perform setjmp/longjmp handling

  GlobalVariable *ThrewGV = nullptr;
  GlobalVariable *ThrewValueGV = nullptr;
  Function *GetTempRet0Func = nullptr;
  Function *SetTempRet0Func = nullptr;
  Function *ResumeF = nullptr;
  Function *EHTypeIDF = nullptr;
  Function *EmLongjmpF = nullptr;
  Function *SaveSetjmpF = nullptr;
  Function *TestSetjmpF = nullptr;

  // __cxa_find_matching_catch_N functions.
  // Indexed by the number of clauses in an original landingpad instruction.
  DenseMap<int, Function *> FindMatchingCatches;
  // Map of <function signature string, invoke_ wrappers>
  StringMap<Function *> InvokeWrappers;
  // Set of allowed function names for exception handling
  std::set<std::string> EHAllowlistSet;
  // Functions that contains calls to setjmp
  SmallPtrSet<Function *, 8> SetjmpUsers;

  StringRef getPassName() const override {
    return "WebAssembly Lower Emscripten Exceptions";
  }

  bool runEHOnFunction(Function &F);
  bool runSjLjOnFunction(Function &F);
  Function *getFindMatchingCatch(Module &M, unsigned NumClauses);

  Value *wrapInvoke(CallBase *CI);
  void wrapTestSetjmp(BasicBlock *BB, DebugLoc DL, Value *Threw,
                      Value *SetjmpTable, Value *SetjmpTableSize, Value *&Label,
                      Value *&LongjmpResult, BasicBlock *&EndBB);
  Function *getInvokeWrapper(CallBase *CI);

  bool areAllExceptionsAllowed() const { return EHAllowlistSet.empty(); }
  bool canLongjmp(Module &M, const Value *Callee) const;
  bool isEmAsmCall(Module &M, const Value *Callee) const;
  bool supportsException(const Function *F) const {
    return EnableEH && (areAllExceptionsAllowed() ||
                        EHAllowlistSet.count(std::string(F->getName())));
  }

  void rebuildSSA(Function &F);

public:
  static char ID;

  WebAssemblyLowerEmscriptenEHSjLj(bool EnableEH = true, bool EnableSjLj = true)
      : ModulePass(ID), EnableEH(EnableEH), EnableSjLj(EnableSjLj) {
    EHAllowlistSet.insert(EHAllowlist.begin(), EHAllowlist.end());
  }
  bool runOnModule(Module &M) override;

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addRequired<DominatorTreeWrapperPass>();
  }
};
} // End anonymous namespace

char WebAssemblyLowerEmscriptenEHSjLj::ID = 0;
INITIALIZE_PASS(WebAssemblyLowerEmscriptenEHSjLj, DEBUG_TYPE,
                "WebAssembly Lower Emscripten Exceptions / Setjmp / Longjmp",
                false, false)

ModulePass *llvm::createWebAssemblyLowerEmscriptenEHSjLj(bool EnableEH,
                                                         bool EnableSjLj) {
  return new WebAssemblyLowerEmscriptenEHSjLj(EnableEH, EnableSjLj);
}

static bool canThrow(const Value *V) {
  if (const auto *F = dyn_cast<const Function>(V)) {
    // Intrinsics cannot throw
    if (F->isIntrinsic())
      return false;
    StringRef Name = F->getName();
    // leave setjmp and longjmp (mostly) alone, we process them properly later
    if (Name == "setjmp" || Name == "longjmp" || Name == "emscripten_longjmp")
      return false;
    return !F->doesNotThrow();
  }
  // not a function, so an indirect call - can throw, we can't tell
  return true;
}

// Get a global variable with the given name. If it doesn't exist declare it,
// which will generate an import and assume that it will exist at link time.
static GlobalVariable *getGlobalVariable(Module &M, Type *Ty,
                                         WebAssemblyTargetMachine &TM,
                                         const char *Name) {
  auto *GV = dyn_cast<GlobalVariable>(M.getOrInsertGlobal(Name, Ty));
  if (!GV)
    report_fatal_error(Twine("unable to create global: ") + Name);

  // If the target supports TLS, make this variable thread-local. We can't just
  // unconditionally make it thread-local and depend on
  // CoalesceFeaturesAndStripAtomics to downgrade it, because stripping TLS has
  // the side effect of disallowing the object from being linked into a
  // shared-memory module, which we don't want to be responsible for.
  auto *Subtarget = TM.getSubtargetImpl();
  auto TLS = Subtarget->hasAtomics() && Subtarget->hasBulkMemory()
                 ? GlobalValue::LocalExecTLSModel
                 : GlobalValue::NotThreadLocal;
  GV->setThreadLocalMode(TLS);
  return GV;
}

// Simple function name mangler.
// This function simply takes LLVM's string representation of parameter types
// and concatenate them with '_'. There are non-alphanumeric characters but llc
// is ok with it, and we need to postprocess these names after the lowering
// phase anyway.
static std::string getSignature(FunctionType *FTy) {
  std::string Sig;
  raw_string_ostream OS(Sig);
  OS << *FTy->getReturnType();
  for (Type *ParamTy : FTy->params())
    OS << "_" << *ParamTy;
  if (FTy->isVarArg())
    OS << "_...";
  Sig = OS.str();
  erase_if(Sig, isSpace);
  // When s2wasm parses .s file, a comma means the end of an argument. So a
  // mangled function name can contain any character but a comma.
  std::replace(Sig.begin(), Sig.end(), ',', '.');
  return Sig;
}

static Function *getEmscriptenFunction(FunctionType *Ty, const Twine &Name,
                                       Module *M) {
  Function* F = Function::Create(Ty, GlobalValue::ExternalLinkage, Name, M);
  // Tell the linker that this function is expected to be imported from the
  // 'env' module.
  if (!F->hasFnAttribute("wasm-import-module")) {
    llvm::AttrBuilder B;
    B.addAttribute("wasm-import-module", "env");
    F->addAttributes(llvm::AttributeList::FunctionIndex, B);
  }
  if (!F->hasFnAttribute("wasm-import-name")) {
    llvm::AttrBuilder B;
    B.addAttribute("wasm-import-name", F->getName());
    F->addAttributes(llvm::AttributeList::FunctionIndex, B);
  }
  return F;
}

// Returns an integer type for the target architecture's address space.
// i32 for wasm32 and i64 for wasm64.
static Type *getAddrIntType(Module *M) {
  IRBuilder<> IRB(M->getContext());
  return IRB.getIntNTy(M->getDataLayout().getPointerSizeInBits());
}

// Returns an integer pointer type for the target architecture's address space.
// i32* for wasm32 and i64* for wasm64.
static Type *getAddrPtrType(Module *M) {
  return Type::getIntNPtrTy(M->getContext(),
                            M->getDataLayout().getPointerSizeInBits());
}

// Returns an integer whose type is the integer type for the target's address
// space. Returns (i32 C) for wasm32 and (i64 C) for wasm64, when C is the
// integer.
static Value *getAddrSizeInt(Module *M, uint64_t C) {
  IRBuilder<> IRB(M->getContext());
  return IRB.getIntN(M->getDataLayout().getPointerSizeInBits(), C);
}

// Returns __cxa_find_matching_catch_N function, where N = NumClauses + 2.
// This is because a landingpad instruction contains two more arguments, a
// personality function and a cleanup bit, and __cxa_find_matching_catch_N
// functions are named after the number of arguments in the original landingpad
// instruction.
Function *
WebAssemblyLowerEmscriptenEHSjLj::getFindMatchingCatch(Module &M,
                                                       unsigned NumClauses) {
  if (FindMatchingCatches.count(NumClauses))
    return FindMatchingCatches[NumClauses];
  PointerType *Int8PtrTy = Type::getInt8PtrTy(M.getContext());
  SmallVector<Type *, 16> Args(NumClauses, Int8PtrTy);
  FunctionType *FTy = FunctionType::get(Int8PtrTy, Args, false);
  Function *F = getEmscriptenFunction(
      FTy, "__cxa_find_matching_catch_" + Twine(NumClauses + 2), &M);
  FindMatchingCatches[NumClauses] = F;
  return F;
}

// Generate invoke wrapper seqence with preamble and postamble
// Preamble:
// __THREW__ = 0;
// Postamble:
// %__THREW__.val = __THREW__; __THREW__ = 0;
// Returns %__THREW__.val, which indicates whether an exception is thrown (or
// whether longjmp occurred), for future use.
Value *WebAssemblyLowerEmscriptenEHSjLj::wrapInvoke(CallBase *CI) {
  Module *M = CI->getModule();
  LLVMContext &C = M->getContext();

  // If we are calling a function that is noreturn, we must remove that
  // attribute. The code we insert here does expect it to return, after we
  // catch the exception.
  if (CI->doesNotReturn()) {
    if (auto *F = CI->getCalledFunction())
      F->removeFnAttr(Attribute::NoReturn);
    CI->removeAttribute(AttributeList::FunctionIndex, Attribute::NoReturn);
  }

  IRBuilder<> IRB(C);
  IRB.SetInsertPoint(CI);

  // Pre-invoke
  // __THREW__ = 0;
  IRB.CreateStore(getAddrSizeInt(M, 0), ThrewGV);

  // Invoke function wrapper in JavaScript
  SmallVector<Value *, 16> Args;
  // Put the pointer to the callee as first argument, so it can be called
  // within the invoke wrapper later
  Args.push_back(CI->getCalledOperand());
  Args.append(CI->arg_begin(), CI->arg_end());
  CallInst *NewCall = IRB.CreateCall(getInvokeWrapper(CI), Args);
  NewCall->takeName(CI);
  NewCall->setCallingConv(CallingConv::WASM_EmscriptenInvoke);
  NewCall->setDebugLoc(CI->getDebugLoc());

  // Because we added the pointer to the callee as first argument, all
  // argument attribute indices have to be incremented by one.
  SmallVector<AttributeSet, 8> ArgAttributes;
  const AttributeList &InvokeAL = CI->getAttributes();

  // No attributes for the callee pointer.
  ArgAttributes.push_back(AttributeSet());
  // Copy the argument attributes from the original
  for (unsigned I = 0, E = CI->getNumArgOperands(); I < E; ++I)
    ArgAttributes.push_back(InvokeAL.getParamAttributes(I));

  AttrBuilder FnAttrs(InvokeAL.getFnAttributes());
  if (FnAttrs.contains(Attribute::AllocSize)) {
    // The allocsize attribute (if any) referes to parameters by index and needs
    // to be adjusted.
    unsigned SizeArg;
    Optional<unsigned> NEltArg;
    std::tie(SizeArg, NEltArg) = FnAttrs.getAllocSizeArgs();
    SizeArg += 1;
    if (NEltArg.hasValue())
      NEltArg = NEltArg.getValue() + 1;
    FnAttrs.addAllocSizeAttr(SizeArg, NEltArg);
  }

  // Reconstruct the AttributesList based on the vector we constructed.
  AttributeList NewCallAL =
      AttributeList::get(C, AttributeSet::get(C, FnAttrs),
                         InvokeAL.getRetAttributes(), ArgAttributes);
  NewCall->setAttributes(NewCallAL);

  CI->replaceAllUsesWith(NewCall);

  // Post-invoke
  // %__THREW__.val = __THREW__; __THREW__ = 0;
  Value *Threw =
      IRB.CreateLoad(getAddrIntType(M), ThrewGV, ThrewGV->getName() + ".val");
  IRB.CreateStore(getAddrSizeInt(M, 0), ThrewGV);
  return Threw;
}

// Get matching invoke wrapper based on callee signature
Function *WebAssemblyLowerEmscriptenEHSjLj::getInvokeWrapper(CallBase *CI) {
  Module *M = CI->getModule();
  SmallVector<Type *, 16> ArgTys;
  FunctionType *CalleeFTy = CI->getFunctionType();

  std::string Sig = getSignature(CalleeFTy);
  if (InvokeWrappers.find(Sig) != InvokeWrappers.end())
    return InvokeWrappers[Sig];

  // Put the pointer to the callee as first argument
  ArgTys.push_back(PointerType::getUnqual(CalleeFTy));
  // Add argument types
  ArgTys.append(CalleeFTy->param_begin(), CalleeFTy->param_end());

  FunctionType *FTy = FunctionType::get(CalleeFTy->getReturnType(), ArgTys,
                                        CalleeFTy->isVarArg());
  Function *F = getEmscriptenFunction(FTy, "__invoke_" + Sig, M);
  InvokeWrappers[Sig] = F;
  return F;
}

bool WebAssemblyLowerEmscriptenEHSjLj::canLongjmp(Module &M,
                                                  const Value *Callee) const {
  if (auto *CalleeF = dyn_cast<Function>(Callee))
    if (CalleeF->isIntrinsic())
      return false;

  // Attempting to transform inline assembly will result in something like:
  //     call void @__invoke_void(void ()* asm ...)
  // which is invalid because inline assembly blocks do not have addresses
  // and can't be passed by pointer. The result is a crash with illegal IR.
  if (isa<InlineAsm>(Callee))
    return false;
  StringRef CalleeName = Callee->getName();

  // The reason we include malloc/free here is to exclude the malloc/free
  // calls generated in setjmp prep / cleanup routines.
  if (CalleeName == "setjmp" || CalleeName == "malloc" || CalleeName == "free")
    return false;

  // There are functions in Emscripten's JS glue code or compiler-rt
  if (CalleeName == "__resumeException" || CalleeName == "llvm_eh_typeid_for" ||
      CalleeName == "saveSetjmp" || CalleeName == "testSetjmp" ||
      CalleeName == "getTempRet0" || CalleeName == "setTempRet0")
    return false;

  // __cxa_find_matching_catch_N functions cannot longjmp
  if (Callee->getName().startswith("__cxa_find_matching_catch_"))
    return false;

  // Exception-catching related functions
  if (CalleeName == "__cxa_begin_catch" || CalleeName == "__cxa_end_catch" ||
      CalleeName == "__cxa_allocate_exception" || CalleeName == "__cxa_throw" ||
      CalleeName == "__clang_call_terminate")
    return false;

  // Otherwise we don't know
  return true;
}

bool WebAssemblyLowerEmscriptenEHSjLj::isEmAsmCall(Module &M,
                                                   const Value *Callee) const {
  StringRef CalleeName = Callee->getName();
  // This is an exhaustive list from Emscripten's <emscripten/em_asm.h>.
  return CalleeName == "emscripten_asm_const_int" ||
         CalleeName == "emscripten_asm_const_double" ||
         CalleeName == "emscripten_asm_const_int_sync_on_main_thread" ||
         CalleeName == "emscripten_asm_const_double_sync_on_main_thread" ||
         CalleeName == "emscripten_asm_const_async_on_main_thread";
}

// Generate testSetjmp function call seqence with preamble and postamble.
// The code this generates is equivalent to the following JavaScript code:
// %__threwValue.val = __threwValue;
// if (%__THREW__.val != 0 & %__threwValue.val != 0) {
//   %label = _testSetjmp(mem[%__THREW__.val], setjmpTable, setjmpTableSize);
//   if (%label == 0)
//     emscripten_longjmp(%__THREW__.val, %__threwValue.val);
//   setTempRet0(%__threwValue.val);
// } else {
//   %label = -1;
// }
// %longjmp_result = getTempRet0();
//
// As output parameters. returns %label, %longjmp_result, and the BB the last
// instruction (%longjmp_result = ...) is in.
void WebAssemblyLowerEmscriptenEHSjLj::wrapTestSetjmp(
    BasicBlock *BB, DebugLoc DL, Value *Threw, Value *SetjmpTable,
    Value *SetjmpTableSize, Value *&Label, Value *&LongjmpResult,
    BasicBlock *&EndBB) {
  Function *F = BB->getParent();
  Module *M = F->getParent();
  LLVMContext &C = M->getContext();
  IRBuilder<> IRB(C);
  IRB.SetCurrentDebugLocation(DL);

  // if (%__THREW__.val != 0 & %__threwValue.val != 0)
  IRB.SetInsertPoint(BB);
  BasicBlock *ThenBB1 = BasicBlock::Create(C, "if.then1", F);
  BasicBlock *ElseBB1 = BasicBlock::Create(C, "if.else1", F);
  BasicBlock *EndBB1 = BasicBlock::Create(C, "if.end", F);
  Value *ThrewCmp = IRB.CreateICmpNE(Threw, getAddrSizeInt(M, 0));
  Value *ThrewValue = IRB.CreateLoad(IRB.getInt32Ty(), ThrewValueGV,
                                     ThrewValueGV->getName() + ".val");
  Value *ThrewValueCmp = IRB.CreateICmpNE(ThrewValue, IRB.getInt32(0));
  Value *Cmp1 = IRB.CreateAnd(ThrewCmp, ThrewValueCmp, "cmp1");
  IRB.CreateCondBr(Cmp1, ThenBB1, ElseBB1);

  // %label = _testSetjmp(mem[%__THREW__.val], _setjmpTable, _setjmpTableSize);
  // if (%label == 0)
  IRB.SetInsertPoint(ThenBB1);
  BasicBlock *ThenBB2 = BasicBlock::Create(C, "if.then2", F);
  BasicBlock *EndBB2 = BasicBlock::Create(C, "if.end2", F);
  Value *ThrewPtr =
      IRB.CreateIntToPtr(Threw, getAddrPtrType(M), Threw->getName() + ".p");
  Value *LoadedThrew = IRB.CreateLoad(getAddrIntType(M), ThrewPtr,
                                      ThrewPtr->getName() + ".loaded");
  Value *ThenLabel = IRB.CreateCall(
      TestSetjmpF, {LoadedThrew, SetjmpTable, SetjmpTableSize}, "label");
  Value *Cmp2 = IRB.CreateICmpEQ(ThenLabel, IRB.getInt32(0));
  IRB.CreateCondBr(Cmp2, ThenBB2, EndBB2);

  // emscripten_longjmp(%__THREW__.val, %__threwValue.val);
  IRB.SetInsertPoint(ThenBB2);
  IRB.CreateCall(EmLongjmpF, {Threw, ThrewValue});
  IRB.CreateUnreachable();

  // setTempRet0(%__threwValue.val);
  IRB.SetInsertPoint(EndBB2);
  IRB.CreateCall(SetTempRet0Func, ThrewValue);
  IRB.CreateBr(EndBB1);

  IRB.SetInsertPoint(ElseBB1);
  IRB.CreateBr(EndBB1);

  // longjmp_result = getTempRet0();
  IRB.SetInsertPoint(EndBB1);
  PHINode *LabelPHI = IRB.CreatePHI(IRB.getInt32Ty(), 2, "label");
  LabelPHI->addIncoming(ThenLabel, EndBB2);

  LabelPHI->addIncoming(IRB.getInt32(-1), ElseBB1);

  // Output parameter assignment
  Label = LabelPHI;
  EndBB = EndBB1;
  LongjmpResult = IRB.CreateCall(GetTempRet0Func, None, "longjmp_result");
}

void WebAssemblyLowerEmscriptenEHSjLj::rebuildSSA(Function &F) {
  DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
  DT.recalculate(F); // CFG has been changed
  SSAUpdater SSA;
  for (BasicBlock &BB : F) {
    for (Instruction &I : BB) {
      SSA.Initialize(I.getType(), I.getName());
      SSA.AddAvailableValue(&BB, &I);
      for (auto UI = I.use_begin(), UE = I.use_end(); UI != UE;) {
        Use &U = *UI;
        ++UI;
        auto *User = cast<Instruction>(U.getUser());
        if (auto *UserPN = dyn_cast<PHINode>(User))
          if (UserPN->getIncomingBlock(U) == &BB)
            continue;

        if (DT.dominates(&I, User))
          continue;
        SSA.RewriteUseAfterInsertions(U);
      }
    }
  }
}

// Replace uses of longjmp with emscripten_longjmp. emscripten_longjmp takes
// arguments of type {i32, i32} (wasm32) / {i64, i32} (wasm64) and longjmp takes
// {jmp_buf*, i32}, so we need a ptrtoint instruction here to make the type
// match. jmp_buf* will eventually be lowered to i32 in the wasm backend.
static void replaceLongjmpWithEmscriptenLongjmp(Function *LongjmpF,
                                                Function *EmLongjmpF) {
  Module *M = LongjmpF->getParent();
  SmallVector<CallInst *, 8> ToErase;
  LLVMContext &C = LongjmpF->getParent()->getContext();
  IRBuilder<> IRB(C);

  // For calls to longjmp, replace it with emscripten_longjmp and cast its first
  // argument (jmp_buf*) to int
  for (User *U : LongjmpF->users()) {
    auto *CI = dyn_cast<CallInst>(U);
    if (CI && CI->getCalledFunction() == LongjmpF) {
      IRB.SetInsertPoint(CI);
      Value *Jmpbuf =
          IRB.CreatePtrToInt(CI->getArgOperand(0), getAddrIntType(M), "jmpbuf");
      IRB.CreateCall(EmLongjmpF, {Jmpbuf, CI->getArgOperand(1)});
      ToErase.push_back(CI);
    }
  }
  for (auto *I : ToErase)
    I->eraseFromParent();

  // If we have any remaining uses of longjmp's function pointer, replace it
  // with (int(*)(jmp_buf*, int))emscripten_longjmp.
  if (!LongjmpF->uses().empty()) {
    Value *EmLongjmp =
        IRB.CreateBitCast(EmLongjmpF, LongjmpF->getType(), "em_longjmp");
    LongjmpF->replaceAllUsesWith(EmLongjmp);
  }
}

bool WebAssemblyLowerEmscriptenEHSjLj::runOnModule(Module &M) {
  LLVM_DEBUG(dbgs() << "********** Lower Emscripten EH & SjLj **********\n");

  LLVMContext &C = M.getContext();
  IRBuilder<> IRB(C);

  Function *SetjmpF = M.getFunction("setjmp");
  Function *LongjmpF = M.getFunction("longjmp");
  bool SetjmpUsed = SetjmpF && !SetjmpF->use_empty();
  bool LongjmpUsed = LongjmpF && !LongjmpF->use_empty();
  DoSjLj = EnableSjLj && (SetjmpUsed || LongjmpUsed);

  auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
  assert(TPC && "Expected a TargetPassConfig");
  auto &TM = TPC->getTM<WebAssemblyTargetMachine>();

  if (EnableEH && TM.Options.ExceptionModel == ExceptionHandling::Wasm)
    report_fatal_error("-exception-model=wasm not allowed with "
                       "-enable-emscripten-cxx-exceptions");

  // Declare (or get) global variables __THREW__, __threwValue, and
  // getTempRet0/setTempRet0 function which are used in common for both
  // exception handling and setjmp/longjmp handling
  ThrewGV = getGlobalVariable(M, getAddrIntType(&M), TM, "__THREW__");
  ThrewValueGV = getGlobalVariable(M, IRB.getInt32Ty(), TM, "__threwValue");
  GetTempRet0Func = getEmscriptenFunction(
      FunctionType::get(IRB.getInt32Ty(), false), "getTempRet0", &M);
  SetTempRet0Func = getEmscriptenFunction(
      FunctionType::get(IRB.getVoidTy(), IRB.getInt32Ty(), false),
      "setTempRet0", &M);
  GetTempRet0Func->setDoesNotThrow();
  SetTempRet0Func->setDoesNotThrow();

  bool Changed = false;

  // Function registration for exception handling
  if (EnableEH) {
    // Register __resumeException function
    FunctionType *ResumeFTy =
        FunctionType::get(IRB.getVoidTy(), IRB.getInt8PtrTy(), false);
    ResumeF = getEmscriptenFunction(ResumeFTy, "__resumeException", &M);

    // Register llvm_eh_typeid_for function
    FunctionType *EHTypeIDTy =
        FunctionType::get(IRB.getInt32Ty(), IRB.getInt8PtrTy(), false);
    EHTypeIDF = getEmscriptenFunction(EHTypeIDTy, "llvm_eh_typeid_for", &M);
  }

  // Function registration and data pre-gathering for setjmp/longjmp handling
  if (DoSjLj) {
    // Register emscripten_longjmp function
    FunctionType *FTy = FunctionType::get(
        IRB.getVoidTy(), {getAddrIntType(&M), IRB.getInt32Ty()}, false);
    EmLongjmpF = getEmscriptenFunction(FTy, "emscripten_longjmp", &M);

    if (SetjmpF) {
      // Register saveSetjmp function
      FunctionType *SetjmpFTy = SetjmpF->getFunctionType();
      FTy = FunctionType::get(Type::getInt32PtrTy(C),
                              {SetjmpFTy->getParamType(0), IRB.getInt32Ty(),
                               Type::getInt32PtrTy(C), IRB.getInt32Ty()},
                              false);
      SaveSetjmpF = getEmscriptenFunction(FTy, "saveSetjmp", &M);

      // Register testSetjmp function
      FTy = FunctionType::get(
          IRB.getInt32Ty(),
          {getAddrIntType(&M), Type::getInt32PtrTy(C), IRB.getInt32Ty()},
          false);
      TestSetjmpF = getEmscriptenFunction(FTy, "testSetjmp", &M);

      // Precompute setjmp users
      for (User *U : SetjmpF->users()) {
        auto *UI = cast<Instruction>(U);
        SetjmpUsers.insert(UI->getFunction());
      }
    }
  }

  // Exception handling transformation
  if (EnableEH) {
    for (Function &F : M) {
      if (F.isDeclaration())
        continue;
      Changed |= runEHOnFunction(F);
    }
  }

  // Setjmp/longjmp handling transformation
  if (DoSjLj) {
    Changed = true; // We have setjmp or longjmp somewhere
    if (LongjmpF)
      replaceLongjmpWithEmscriptenLongjmp(LongjmpF, EmLongjmpF);
    // Only traverse functions that uses setjmp in order not to insert
    // unnecessary prep / cleanup code in every function
    if (SetjmpF)
      for (Function *F : SetjmpUsers)
        runSjLjOnFunction(*F);
  }

  if (!Changed) {
    // Delete unused global variables and functions
    if (ResumeF)
      ResumeF->eraseFromParent();
    if (EHTypeIDF)
      EHTypeIDF->eraseFromParent();
    if (EmLongjmpF)
      EmLongjmpF->eraseFromParent();
    if (SaveSetjmpF)
      SaveSetjmpF->eraseFromParent();
    if (TestSetjmpF)
      TestSetjmpF->eraseFromParent();
    return false;
  }

  return true;
}

bool WebAssemblyLowerEmscriptenEHSjLj::runEHOnFunction(Function &F) {
  Module &M = *F.getParent();
  LLVMContext &C = F.getContext();
  IRBuilder<> IRB(C);
  bool Changed = false;
  SmallVector<Instruction *, 64> ToErase;
  SmallPtrSet<LandingPadInst *, 32> LandingPads;

  for (BasicBlock &BB : F) {
    auto *II = dyn_cast<InvokeInst>(BB.getTerminator());
    if (!II)
      continue;
    Changed = true;
    LandingPads.insert(II->getLandingPadInst());
    IRB.SetInsertPoint(II);

    const Value *Callee = II->getCalledOperand();
    bool NeedInvoke = supportsException(&F) && canThrow(Callee);
    if (NeedInvoke) {
      // Wrap invoke with invoke wrapper and generate preamble/postamble
      Value *Threw = wrapInvoke(II);
      ToErase.push_back(II);

      // If setjmp/longjmp handling is enabled, the thrown value can be not an
      // exception but a longjmp. If the current function contains calls to
      // setjmp, it will be appropriately handled in runSjLjOnFunction. But even
      // if the function does not contain setjmp calls, we shouldn't silently
      // ignore longjmps; we should rethrow them so they can be correctly
      // handled in somewhere up the call chain where setjmp is.
      // __THREW__'s value is 0 when nothing happened, 1 when an exception is
      // thrown, other values when longjmp is thrown.
      //
      // if (%__THREW__.val == 0 || %__THREW__.val == 1)
      //   goto %tail
      // else
      //   goto %longjmp.rethrow
      //
      // longjmp.rethrow: ;; This is longjmp. Rethrow it
      //   %__threwValue.val = __threwValue
      //   emscripten_longjmp(%__THREW__.val, %__threwValue.val);
      //
      // tail: ;; Nothing happened or an exception is thrown
      //   ... Continue exception handling ...
      if (DoSjLj && !SetjmpUsers.count(&F) && canLongjmp(M, Callee)) {
        BasicBlock *Tail = BasicBlock::Create(C, "tail", &F);
        BasicBlock *RethrowBB = BasicBlock::Create(C, "longjmp.rethrow", &F);
        Value *CmpEqOne =
            IRB.CreateICmpEQ(Threw, getAddrSizeInt(&M, 1), "cmp.eq.one");
        Value *CmpEqZero =
            IRB.CreateICmpEQ(Threw, getAddrSizeInt(&M, 0), "cmp.eq.zero");
        Value *Or = IRB.CreateOr(CmpEqZero, CmpEqOne, "or");
        IRB.CreateCondBr(Or, Tail, RethrowBB);
        IRB.SetInsertPoint(RethrowBB);
        Value *ThrewValue = IRB.CreateLoad(IRB.getInt32Ty(), ThrewValueGV,
                                           ThrewValueGV->getName() + ".val");
        IRB.CreateCall(EmLongjmpF, {Threw, ThrewValue});

        IRB.CreateUnreachable();
        IRB.SetInsertPoint(Tail);
      }

      // Insert a branch based on __THREW__ variable
      Value *Cmp = IRB.CreateICmpEQ(Threw, getAddrSizeInt(&M, 1), "cmp");
      IRB.CreateCondBr(Cmp, II->getUnwindDest(), II->getNormalDest());

    } else {
      // This can't throw, and we don't need this invoke, just replace it with a
      // call+branch
      SmallVector<Value *, 16> Args(II->args());
      CallInst *NewCall =
          IRB.CreateCall(II->getFunctionType(), II->getCalledOperand(), Args);
      NewCall->takeName(II);
      NewCall->setCallingConv(II->getCallingConv());
      NewCall->setDebugLoc(II->getDebugLoc());
      NewCall->setAttributes(II->getAttributes());
      II->replaceAllUsesWith(NewCall);
      ToErase.push_back(II);

      IRB.CreateBr(II->getNormalDest());

      // Remove any PHI node entries from the exception destination
      II->getUnwindDest()->removePredecessor(&BB);
    }
  }

  // Process resume instructions
  for (BasicBlock &BB : F) {
    // Scan the body of the basic block for resumes
    for (Instruction &I : BB) {
      auto *RI = dyn_cast<ResumeInst>(&I);
      if (!RI)
        continue;
      Changed = true;

      // Split the input into legal values
      Value *Input = RI->getValue();
      IRB.SetInsertPoint(RI);
      Value *Low = IRB.CreateExtractValue(Input, 0, "low");
      // Create a call to __resumeException function
      IRB.CreateCall(ResumeF, {Low});
      // Add a terminator to the block
      IRB.CreateUnreachable();
      ToErase.push_back(RI);
    }
  }

  // Process llvm.eh.typeid.for intrinsics
  for (BasicBlock &BB : F) {
    for (Instruction &I : BB) {
      auto *CI = dyn_cast<CallInst>(&I);
      if (!CI)
        continue;
      const Function *Callee = CI->getCalledFunction();
      if (!Callee)
        continue;
      if (Callee->getIntrinsicID() != Intrinsic::eh_typeid_for)
        continue;
      Changed = true;

      IRB.SetInsertPoint(CI);
      CallInst *NewCI =
          IRB.CreateCall(EHTypeIDF, CI->getArgOperand(0), "typeid");
      CI->replaceAllUsesWith(NewCI);
      ToErase.push_back(CI);
    }
  }

  // Look for orphan landingpads, can occur in blocks with no predecessors
  for (BasicBlock &BB : F) {
    Instruction *I = BB.getFirstNonPHI();
    if (auto *LPI = dyn_cast<LandingPadInst>(I))
      LandingPads.insert(LPI);
  }
  Changed |= !LandingPads.empty();

  // Handle all the landingpad for this function together, as multiple invokes
  // may share a single lp
  for (LandingPadInst *LPI : LandingPads) {
    IRB.SetInsertPoint(LPI);
    SmallVector<Value *, 16> FMCArgs;
    for (unsigned I = 0, E = LPI->getNumClauses(); I < E; ++I) {
      Constant *Clause = LPI->getClause(I);
      // TODO Handle filters (= exception specifications).
      // https://bugs.llvm.org/show_bug.cgi?id=50396
      if (LPI->isCatch(I))
        FMCArgs.push_back(Clause);
    }

    // Create a call to __cxa_find_matching_catch_N function
    Function *FMCF = getFindMatchingCatch(M, FMCArgs.size());
    CallInst *FMCI = IRB.CreateCall(FMCF, FMCArgs, "fmc");
    Value *Undef = UndefValue::get(LPI->getType());
    Value *Pair0 = IRB.CreateInsertValue(Undef, FMCI, 0, "pair0");
    Value *TempRet0 = IRB.CreateCall(GetTempRet0Func, None, "tempret0");
    Value *Pair1 = IRB.CreateInsertValue(Pair0, TempRet0, 1, "pair1");

    LPI->replaceAllUsesWith(Pair1);
    ToErase.push_back(LPI);
  }

  // Erase everything we no longer need in this function
  for (Instruction *I : ToErase)
    I->eraseFromParent();

  return Changed;
}

// This tries to get debug info from the instruction before which a new
// instruction will be inserted, and if there's no debug info in that
// instruction, tries to get the info instead from the previous instruction (if
// any). If none of these has debug info and a DISubprogram is provided, it
// creates a dummy debug info with the first line of the function, because IR
// verifier requires all inlinable callsites should have debug info when both a
// caller and callee have DISubprogram. If none of these conditions are met,
// returns empty info.
static DebugLoc getOrCreateDebugLoc(const Instruction *InsertBefore,
                                    DISubprogram *SP) {
  assert(InsertBefore);
  if (InsertBefore->getDebugLoc())
    return InsertBefore->getDebugLoc();
  const Instruction *Prev = InsertBefore->getPrevNode();
  if (Prev && Prev->getDebugLoc())
    return Prev->getDebugLoc();
  if (SP)
    return DILocation::get(SP->getContext(), SP->getLine(), 1, SP);
  return DebugLoc();
}

bool WebAssemblyLowerEmscriptenEHSjLj::runSjLjOnFunction(Function &F) {
  Module &M = *F.getParent();
  LLVMContext &C = F.getContext();
  IRBuilder<> IRB(C);
  SmallVector<Instruction *, 64> ToErase;
  // Vector of %setjmpTable values
  std::vector<Instruction *> SetjmpTableInsts;
  // Vector of %setjmpTableSize values
  std::vector<Instruction *> SetjmpTableSizeInsts;

  // Setjmp preparation

  // This instruction effectively means %setjmpTableSize = 4.
  // We create this as an instruction intentionally, and we don't want to fold
  // this instruction to a constant 4, because this value will be used in
  // SSAUpdater.AddAvailableValue(...) later.
  BasicBlock &EntryBB = F.getEntryBlock();
  DebugLoc FirstDL = getOrCreateDebugLoc(&*EntryBB.begin(), F.getSubprogram());
  BinaryOperator *SetjmpTableSize = BinaryOperator::Create(
      Instruction::Add, IRB.getInt32(4), IRB.getInt32(0), "setjmpTableSize",
      &*EntryBB.getFirstInsertionPt());
  SetjmpTableSize->setDebugLoc(FirstDL);
  // setjmpTable = (int *) malloc(40);
  Instruction *SetjmpTable = CallInst::CreateMalloc(
      SetjmpTableSize, IRB.getInt32Ty(), IRB.getInt32Ty(), IRB.getInt32(40),
      nullptr, nullptr, "setjmpTable");
  SetjmpTable->setDebugLoc(FirstDL);
  // CallInst::CreateMalloc may return a bitcast instruction if the result types
  // mismatch. We need to set the debug loc for the original call too.
  auto *MallocCall = SetjmpTable->stripPointerCasts();
  if (auto *MallocCallI = dyn_cast<Instruction>(MallocCall)) {
    MallocCallI->setDebugLoc(FirstDL);
  }
  // setjmpTable[0] = 0;
  IRB.SetInsertPoint(SetjmpTableSize);
  IRB.CreateStore(IRB.getInt32(0), SetjmpTable);
  SetjmpTableInsts.push_back(SetjmpTable);
  SetjmpTableSizeInsts.push_back(SetjmpTableSize);

  // Setjmp transformation
  std::vector<PHINode *> SetjmpRetPHIs;
  Function *SetjmpF = M.getFunction("setjmp");
  for (User *U : SetjmpF->users()) {
    auto *CI = dyn_cast<CallInst>(U);
    if (!CI)
      report_fatal_error("Does not support indirect calls to setjmp");

    BasicBlock *BB = CI->getParent();
    if (BB->getParent() != &F) // in other function
      continue;

    // The tail is everything right after the call, and will be reached once
    // when setjmp is called, and later when longjmp returns to the setjmp
    BasicBlock *Tail = SplitBlock(BB, CI->getNextNode());
    // Add a phi to the tail, which will be the output of setjmp, which
    // indicates if this is the first call or a longjmp back. The phi directly
    // uses the right value based on where we arrive from
    IRB.SetInsertPoint(Tail->getFirstNonPHI());
    PHINode *SetjmpRet = IRB.CreatePHI(IRB.getInt32Ty(), 2, "setjmp.ret");

    // setjmp initial call returns 0
    SetjmpRet->addIncoming(IRB.getInt32(0), BB);
    // The proper output is now this, not the setjmp call itself
    CI->replaceAllUsesWith(SetjmpRet);
    // longjmp returns to the setjmp will add themselves to this phi
    SetjmpRetPHIs.push_back(SetjmpRet);

    // Fix call target
    // Our index in the function is our place in the array + 1 to avoid index
    // 0, because index 0 means the longjmp is not ours to handle.
    IRB.SetInsertPoint(CI);
    Value *Args[] = {CI->getArgOperand(0), IRB.getInt32(SetjmpRetPHIs.size()),
                     SetjmpTable, SetjmpTableSize};
    Instruction *NewSetjmpTable =
        IRB.CreateCall(SaveSetjmpF, Args, "setjmpTable");
    Instruction *NewSetjmpTableSize =
        IRB.CreateCall(GetTempRet0Func, None, "setjmpTableSize");
    SetjmpTableInsts.push_back(NewSetjmpTable);
    SetjmpTableSizeInsts.push_back(NewSetjmpTableSize);
    ToErase.push_back(CI);
  }

  // Update each call that can longjmp so it can return to a setjmp where
  // relevant.

  // Because we are creating new BBs while processing and don't want to make
  // all these newly created BBs candidates again for longjmp processing, we
  // first make the vector of candidate BBs.
  std::vector<BasicBlock *> BBs;
  for (BasicBlock &BB : F)
    BBs.push_back(&BB);

  // BBs.size() will change within the loop, so we query it every time
  for (unsigned I = 0; I < BBs.size(); I++) {
    BasicBlock *BB = BBs[I];
    for (Instruction &I : *BB) {
      assert(!isa<InvokeInst>(&I));
      auto *CI = dyn_cast<CallInst>(&I);
      if (!CI)
        continue;

      const Value *Callee = CI->getCalledOperand();
      if (!canLongjmp(M, Callee))
        continue;
      if (isEmAsmCall(M, Callee))
        report_fatal_error("Cannot use EM_ASM* alongside setjmp/longjmp in " +
                               F.getName() +
                               ". Please consider using EM_JS, or move the "
                               "EM_ASM into another function.",
                           false);

      Value *Threw = nullptr;
      BasicBlock *Tail;
      if (Callee->getName().startswith("__invoke_")) {
        // If invoke wrapper has already been generated for this call in
        // previous EH phase, search for the load instruction
        // %__THREW__.val = __THREW__;
        // in postamble after the invoke wrapper call
        LoadInst *ThrewLI = nullptr;
        StoreInst *ThrewResetSI = nullptr;
        for (auto I = std::next(BasicBlock::iterator(CI)), IE = BB->end();
             I != IE; ++I) {
          if (auto *LI = dyn_cast<LoadInst>(I))
            if (auto *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand()))
              if (GV == ThrewGV) {
                Threw = ThrewLI = LI;
                break;
              }
        }
        // Search for the store instruction after the load above
        // __THREW__ = 0;
        for (auto I = std::next(BasicBlock::iterator(ThrewLI)), IE = BB->end();
             I != IE; ++I) {
          if (auto *SI = dyn_cast<StoreInst>(I)) {
            if (auto *GV = dyn_cast<GlobalVariable>(SI->getPointerOperand())) {
              if (GV == ThrewGV &&
                  SI->getValueOperand() == getAddrSizeInt(&M, 0)) {
                ThrewResetSI = SI;
                break;
              }
            }
          }
        }
        assert(Threw && ThrewLI && "Cannot find __THREW__ load after invoke");
        assert(ThrewResetSI && "Cannot find __THREW__ store after invoke");
        Tail = SplitBlock(BB, ThrewResetSI->getNextNode());

      } else {
        // Wrap call with invoke wrapper and generate preamble/postamble
        Threw = wrapInvoke(CI);
        ToErase.push_back(CI);
        Tail = SplitBlock(BB, CI->getNextNode());

        // If exception handling is enabled, the thrown value can be not a
        // longjmp but an exception, in which case we shouldn't silently ignore
        // exceptions; we should rethrow them.
        // __THREW__'s value is 0 when nothing happened, 1 when an exception is
        // thrown, other values when longjmp is thrown.
        //
        // if (%__THREW__.val == 1)
        //   goto %eh.rethrow
        // else
        //   goto %normal
        //
        // eh.rethrow: ;; Rethrow exception
        //   %exn = call @__cxa_find_matching_catch_2() ;; Retrieve thrown ptr
        //   __resumeException(%exn)
        //
        // normal:
        //   <-- Insertion point. Will insert sjlj handling code from here
        //   goto %tail
        //
        // tail:
        //   ...
        if (supportsException(&F) && canThrow(Callee)) {
          IRB.SetInsertPoint(CI);
          // We will add a new conditional branch. So remove the branch created
          // when we split the BB
          ToErase.push_back(BB->getTerminator());
          BasicBlock *NormalBB = BasicBlock::Create(C, "normal", &F);
          BasicBlock *RethrowBB = BasicBlock::Create(C, "eh.rethrow", &F);
          Value *CmpEqOne =
              IRB.CreateICmpEQ(Threw, getAddrSizeInt(&M, 1), "cmp.eq.one");
          IRB.CreateCondBr(CmpEqOne, RethrowBB, NormalBB);
          IRB.SetInsertPoint(RethrowBB);
          CallInst *Exn = IRB.CreateCall(getFindMatchingCatch(M, 0), {}, "exn");
          IRB.CreateCall(ResumeF, {Exn});
          IRB.CreateUnreachable();
          IRB.SetInsertPoint(NormalBB);
          IRB.CreateBr(Tail);
          BB = NormalBB; // New insertion point to insert testSetjmp()
        }
      }

      // We need to replace the terminator in Tail - SplitBlock makes BB go
      // straight to Tail, we need to check if a longjmp occurred, and go to the
      // right setjmp-tail if so
      ToErase.push_back(BB->getTerminator());

      // Generate a function call to testSetjmp function and preamble/postamble
      // code to figure out (1) whether longjmp occurred (2) if longjmp
      // occurred, which setjmp it corresponds to
      Value *Label = nullptr;
      Value *LongjmpResult = nullptr;
      BasicBlock *EndBB = nullptr;
      wrapTestSetjmp(BB, CI->getDebugLoc(), Threw, SetjmpTable, SetjmpTableSize,
                     Label, LongjmpResult, EndBB);
      assert(Label && LongjmpResult && EndBB);

      // Create switch instruction
      IRB.SetInsertPoint(EndBB);
      IRB.SetCurrentDebugLocation(EndBB->getInstList().back().getDebugLoc());
      SwitchInst *SI = IRB.CreateSwitch(Label, Tail, SetjmpRetPHIs.size());
      // -1 means no longjmp happened, continue normally (will hit the default
      // switch case). 0 means a longjmp that is not ours to handle, needs a
      // rethrow. Otherwise the index is the same as the index in P+1 (to avoid
      // 0).
      for (unsigned I = 0; I < SetjmpRetPHIs.size(); I++) {
        SI->addCase(IRB.getInt32(I + 1), SetjmpRetPHIs[I]->getParent());
        SetjmpRetPHIs[I]->addIncoming(LongjmpResult, EndBB);
      }

      // We are splitting the block here, and must continue to find other calls
      // in the block - which is now split. so continue to traverse in the Tail
      BBs.push_back(Tail);
    }
  }

  // Erase everything we no longer need in this function
  for (Instruction *I : ToErase)
    I->eraseFromParent();

  // Free setjmpTable buffer before each return instruction
  for (BasicBlock &BB : F) {
    Instruction *TI = BB.getTerminator();
    if (isa<ReturnInst>(TI)) {
      DebugLoc DL = getOrCreateDebugLoc(TI, F.getSubprogram());
      auto *Free = CallInst::CreateFree(SetjmpTable, TI);
      Free->setDebugLoc(DL);
      // CallInst::CreateFree may create a bitcast instruction if its argument
      // types mismatch. We need to set the debug loc for the bitcast too.
      if (auto *FreeCallI = dyn_cast<CallInst>(Free)) {
        if (auto *BitCastI = dyn_cast<BitCastInst>(FreeCallI->getArgOperand(0)))
          BitCastI->setDebugLoc(DL);
      }
    }
  }

  // Every call to saveSetjmp can change setjmpTable and setjmpTableSize
  // (when buffer reallocation occurs)
  // entry:
  //   setjmpTableSize = 4;
  //   setjmpTable = (int *) malloc(40);
  //   setjmpTable[0] = 0;
  // ...
  // somebb:
  //   setjmpTable = saveSetjmp(buf, label, setjmpTable, setjmpTableSize);
  //   setjmpTableSize = getTempRet0();
  // So we need to make sure the SSA for these variables is valid so that every
  // saveSetjmp and testSetjmp calls have the correct arguments.
  SSAUpdater SetjmpTableSSA;
  SSAUpdater SetjmpTableSizeSSA;
  SetjmpTableSSA.Initialize(Type::getInt32PtrTy(C), "setjmpTable");
  SetjmpTableSizeSSA.Initialize(Type::getInt32Ty(C), "setjmpTableSize");
  for (Instruction *I : SetjmpTableInsts)
    SetjmpTableSSA.AddAvailableValue(I->getParent(), I);
  for (Instruction *I : SetjmpTableSizeInsts)
    SetjmpTableSizeSSA.AddAvailableValue(I->getParent(), I);

  for (auto UI = SetjmpTable->use_begin(), UE = SetjmpTable->use_end();
       UI != UE;) {
    // Grab the use before incrementing the iterator.
    Use &U = *UI;
    // Increment the iterator before removing the use from the list.
    ++UI;
    if (auto *I = dyn_cast<Instruction>(U.getUser()))
      if (I->getParent() != &EntryBB)
        SetjmpTableSSA.RewriteUse(U);
  }
  for (auto UI = SetjmpTableSize->use_begin(), UE = SetjmpTableSize->use_end();
       UI != UE;) {
    Use &U = *UI;
    ++UI;
    if (auto *I = dyn_cast<Instruction>(U.getUser()))
      if (I->getParent() != &EntryBB)
        SetjmpTableSizeSSA.RewriteUse(U);
  }

  // Finally, our modifications to the cfg can break dominance of SSA variables.
  // For example, in this code,
  // if (x()) { .. setjmp() .. }
  // if (y()) { .. longjmp() .. }
  // We must split the longjmp block, and it can jump into the block splitted
  // from setjmp one. But that means that when we split the setjmp block, it's
  // first part no longer dominates its second part - there is a theoretically
  // possible control flow path where x() is false, then y() is true and we
  // reach the second part of the setjmp block, without ever reaching the first
  // part. So, we rebuild SSA form here.
  rebuildSSA(F);
  return true;
}