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/Support/CommandLine.h"
215 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
216 #include "llvm/Transforms/Utils/SSAUpdater.h"
217 
218 using namespace llvm;
219 
220 #define DEBUG_TYPE "wasm-lower-em-ehsjlj"
221 
222 static cl::list<std::string>
223     EHWhitelist("emscripten-cxx-exceptions-whitelist",
224                 cl::desc("The list of function names in which Emscripten-style "
225                          "exception handling is enabled (see emscripten "
226                          "EMSCRIPTEN_CATCHING_WHITELIST options)"),
227                 cl::CommaSeparated);
228 
229 namespace {
230 class WebAssemblyLowerEmscriptenEHSjLj final : public ModulePass {
231   bool EnableEH;   // Enable exception handling
232   bool EnableSjLj; // Enable setjmp/longjmp handling
233 
234   GlobalVariable *ThrewGV = nullptr;
235   GlobalVariable *ThrewValueGV = nullptr;
236   Function *GetTempRet0Func = nullptr;
237   Function *SetTempRet0Func = nullptr;
238   Function *ResumeF = nullptr;
239   Function *EHTypeIDF = nullptr;
240   Function *EmLongjmpF = nullptr;
241   Function *EmLongjmpJmpbufF = nullptr;
242   Function *SaveSetjmpF = nullptr;
243   Function *TestSetjmpF = nullptr;
244 
245   // __cxa_find_matching_catch_N functions.
246   // Indexed by the number of clauses in an original landingpad instruction.
247   DenseMap<int, Function *> FindMatchingCatches;
248   // Map of <function signature string, invoke_ wrappers>
249   StringMap<Function *> InvokeWrappers;
250   // Set of whitelisted function names for exception handling
251   std::set<std::string> EHWhitelistSet;
252 
253   StringRef getPassName() const override {
254     return "WebAssembly Lower Emscripten Exceptions";
255   }
256 
257   bool runEHOnFunction(Function &F);
258   bool runSjLjOnFunction(Function &F);
259   Function *getFindMatchingCatch(Module &M, unsigned NumClauses);
260 
261   template <typename CallOrInvoke> Value *wrapInvoke(CallOrInvoke *CI);
262   void wrapTestSetjmp(BasicBlock *BB, DebugLoc DL, Value *Threw,
263                       Value *SetjmpTable, Value *SetjmpTableSize, Value *&Label,
264                       Value *&LongjmpResult, BasicBlock *&EndBB);
265   template <typename CallOrInvoke> Function *getInvokeWrapper(CallOrInvoke *CI);
266 
267   bool areAllExceptionsAllowed() const { return EHWhitelistSet.empty(); }
268   bool canLongjmp(Module &M, const Value *Callee) const;
269   bool isEmAsmCall(Module &M, const Value *Callee) const;
270 
271   void rebuildSSA(Function &F);
272 
273 public:
274   static char ID;
275 
276   WebAssemblyLowerEmscriptenEHSjLj(bool EnableEH = true, bool EnableSjLj = true)
277       : ModulePass(ID), EnableEH(EnableEH), EnableSjLj(EnableSjLj) {
278     EHWhitelistSet.insert(EHWhitelist.begin(), EHWhitelist.end());
279   }
280   bool runOnModule(Module &M) override;
281 
282   void getAnalysisUsage(AnalysisUsage &AU) const override {
283     AU.addRequired<DominatorTreeWrapperPass>();
284   }
285 };
286 } // End anonymous namespace
287 
288 char WebAssemblyLowerEmscriptenEHSjLj::ID = 0;
289 INITIALIZE_PASS(WebAssemblyLowerEmscriptenEHSjLj, DEBUG_TYPE,
290                 "WebAssembly Lower Emscripten Exceptions / Setjmp / Longjmp",
291                 false, false)
292 
293 ModulePass *llvm::createWebAssemblyLowerEmscriptenEHSjLj(bool EnableEH,
294                                                          bool EnableSjLj) {
295   return new WebAssemblyLowerEmscriptenEHSjLj(EnableEH, EnableSjLj);
296 }
297 
298 static bool canThrow(const Value *V) {
299   if (const auto *F = dyn_cast<const Function>(V)) {
300     // Intrinsics cannot throw
301     if (F->isIntrinsic())
302       return false;
303     StringRef Name = F->getName();
304     // leave setjmp and longjmp (mostly) alone, we process them properly later
305     if (Name == "setjmp" || Name == "longjmp")
306       return false;
307     return !F->doesNotThrow();
308   }
309   // not a function, so an indirect call - can throw, we can't tell
310   return true;
311 }
312 
313 // Get a global variable with the given name.  If it doesn't exist declare it,
314 // which will generate an import and asssumes that it will exist at link time.
315 static GlobalVariable *getGlobalVariableI32(Module &M, IRBuilder<> &IRB,
316                                             const char *Name) {
317 
318   auto *GV =
319       dyn_cast<GlobalVariable>(M.getOrInsertGlobal(Name, IRB.getInt32Ty()));
320   if (!GV)
321     report_fatal_error(Twine("unable to create global: ") + Name);
322 
323   return GV;
324 }
325 
326 // Simple function name mangler.
327 // This function simply takes LLVM's string representation of parameter types
328 // and concatenate them with '_'. There are non-alphanumeric characters but llc
329 // is ok with it, and we need to postprocess these names after the lowering
330 // phase anyway.
331 static std::string getSignature(FunctionType *FTy) {
332   std::string Sig;
333   raw_string_ostream OS(Sig);
334   OS << *FTy->getReturnType();
335   for (Type *ParamTy : FTy->params())
336     OS << "_" << *ParamTy;
337   if (FTy->isVarArg())
338     OS << "_...";
339   Sig = OS.str();
340   Sig.erase(remove_if(Sig, isspace), Sig.end());
341   // When s2wasm parses .s file, a comma means the end of an argument. So a
342   // mangled function name can contain any character but a comma.
343   std::replace(Sig.begin(), Sig.end(), ',', '.');
344   return Sig;
345 }
346 
347 static void markAsImported(Function *F) {
348   // Tell the linker that this function is expected to be imported from the
349   // 'env' module.
350   if (!F->hasFnAttribute("wasm-import-module")) {
351     llvm::AttrBuilder B;
352     B.addAttribute("wasm-import-module", "env");
353     F->addAttributes(llvm::AttributeList::FunctionIndex, B);
354   }
355   if (!F->hasFnAttribute("wasm-import-name")) {
356     llvm::AttrBuilder B;
357     B.addAttribute("wasm-import-name", F->getName());
358     F->addAttributes(llvm::AttributeList::FunctionIndex, B);
359   }
360 }
361 
362 // Returns __cxa_find_matching_catch_N function, where N = NumClauses + 2.
363 // This is because a landingpad instruction contains two more arguments, a
364 // personality function and a cleanup bit, and __cxa_find_matching_catch_N
365 // functions are named after the number of arguments in the original landingpad
366 // instruction.
367 Function *
368 WebAssemblyLowerEmscriptenEHSjLj::getFindMatchingCatch(Module &M,
369                                                        unsigned NumClauses) {
370   if (FindMatchingCatches.count(NumClauses))
371     return FindMatchingCatches[NumClauses];
372   PointerType *Int8PtrTy = Type::getInt8PtrTy(M.getContext());
373   SmallVector<Type *, 16> Args(NumClauses, Int8PtrTy);
374   FunctionType *FTy = FunctionType::get(Int8PtrTy, Args, false);
375   Function *F = Function::Create(
376       FTy, GlobalValue::ExternalLinkage,
377       "__cxa_find_matching_catch_" + Twine(NumClauses + 2), &M);
378   markAsImported(F);
379   FindMatchingCatches[NumClauses] = F;
380   return F;
381 }
382 
383 // Generate invoke wrapper seqence with preamble and postamble
384 // Preamble:
385 // __THREW__ = 0;
386 // Postamble:
387 // %__THREW__.val = __THREW__; __THREW__ = 0;
388 // Returns %__THREW__.val, which indicates whether an exception is thrown (or
389 // whether longjmp occurred), for future use.
390 template <typename CallOrInvoke>
391 Value *WebAssemblyLowerEmscriptenEHSjLj::wrapInvoke(CallOrInvoke *CI) {
392   LLVMContext &C = CI->getModule()->getContext();
393 
394   // If we are calling a function that is noreturn, we must remove that
395   // attribute. The code we insert here does expect it to return, after we
396   // catch the exception.
397   if (CI->doesNotReturn()) {
398     if (auto *F = dyn_cast<Function>(CI->getCalledValue()))
399       F->removeFnAttr(Attribute::NoReturn);
400     CI->removeAttribute(AttributeList::FunctionIndex, Attribute::NoReturn);
401   }
402 
403   IRBuilder<> IRB(C);
404   IRB.SetInsertPoint(CI);
405 
406   // Pre-invoke
407   // __THREW__ = 0;
408   IRB.CreateStore(IRB.getInt32(0), ThrewGV);
409 
410   // Invoke function wrapper in JavaScript
411   SmallVector<Value *, 16> Args;
412   // Put the pointer to the callee as first argument, so it can be called
413   // within the invoke wrapper later
414   Args.push_back(CI->getCalledValue());
415   Args.append(CI->arg_begin(), CI->arg_end());
416   CallInst *NewCall = IRB.CreateCall(getInvokeWrapper(CI), Args);
417   NewCall->takeName(CI);
418   NewCall->setCallingConv(CallingConv::WASM_EmscriptenInvoke);
419   NewCall->setDebugLoc(CI->getDebugLoc());
420 
421   // Because we added the pointer to the callee as first argument, all
422   // argument attribute indices have to be incremented by one.
423   SmallVector<AttributeSet, 8> ArgAttributes;
424   const AttributeList &InvokeAL = CI->getAttributes();
425 
426   // No attributes for the callee pointer.
427   ArgAttributes.push_back(AttributeSet());
428   // Copy the argument attributes from the original
429   for (unsigned I = 0, E = CI->getNumArgOperands(); I < E; ++I)
430     ArgAttributes.push_back(InvokeAL.getParamAttributes(I));
431 
432   AttrBuilder FnAttrs(InvokeAL.getFnAttributes());
433   if (FnAttrs.contains(Attribute::AllocSize)) {
434     // The allocsize attribute (if any) referes to parameters by index and needs
435     // to be adjusted.
436     unsigned SizeArg;
437     Optional<unsigned> NEltArg;
438     std::tie(SizeArg, NEltArg) = FnAttrs.getAllocSizeArgs();
439     SizeArg += 1;
440     if (NEltArg.hasValue())
441       NEltArg = NEltArg.getValue() + 1;
442     FnAttrs.addAllocSizeAttr(SizeArg, NEltArg);
443   }
444 
445   // Reconstruct the AttributesList based on the vector we constructed.
446   AttributeList NewCallAL =
447       AttributeList::get(C, AttributeSet::get(C, FnAttrs),
448                          InvokeAL.getRetAttributes(), ArgAttributes);
449   NewCall->setAttributes(NewCallAL);
450 
451   CI->replaceAllUsesWith(NewCall);
452 
453   // Post-invoke
454   // %__THREW__.val = __THREW__; __THREW__ = 0;
455   Value *Threw =
456       IRB.CreateLoad(IRB.getInt32Ty(), ThrewGV, ThrewGV->getName() + ".val");
457   IRB.CreateStore(IRB.getInt32(0), ThrewGV);
458   return Threw;
459 }
460 
461 // Get matching invoke wrapper based on callee signature
462 template <typename CallOrInvoke>
463 Function *WebAssemblyLowerEmscriptenEHSjLj::getInvokeWrapper(CallOrInvoke *CI) {
464   Module *M = CI->getModule();
465   SmallVector<Type *, 16> ArgTys;
466   Value *Callee = CI->getCalledValue();
467   FunctionType *CalleeFTy;
468   if (auto *F = dyn_cast<Function>(Callee))
469     CalleeFTy = F->getFunctionType();
470   else {
471     auto *CalleeTy = cast<PointerType>(Callee->getType())->getElementType();
472     CalleeFTy = cast<FunctionType>(CalleeTy);
473   }
474 
475   std::string Sig = getSignature(CalleeFTy);
476   if (InvokeWrappers.find(Sig) != InvokeWrappers.end())
477     return InvokeWrappers[Sig];
478 
479   // Put the pointer to the callee as first argument
480   ArgTys.push_back(PointerType::getUnqual(CalleeFTy));
481   // Add argument types
482   ArgTys.append(CalleeFTy->param_begin(), CalleeFTy->param_end());
483 
484   FunctionType *FTy = FunctionType::get(CalleeFTy->getReturnType(), ArgTys,
485                                         CalleeFTy->isVarArg());
486   Function *F =
487       Function::Create(FTy, GlobalValue::ExternalLinkage, "__invoke_" + Sig, M);
488   markAsImported(F);
489   InvokeWrappers[Sig] = F;
490   return F;
491 }
492 
493 bool WebAssemblyLowerEmscriptenEHSjLj::canLongjmp(Module &M,
494                                                   const Value *Callee) const {
495   if (auto *CalleeF = dyn_cast<Function>(Callee))
496     if (CalleeF->isIntrinsic())
497       return false;
498 
499   // Attempting to transform inline assembly will result in something like:
500   //     call void @__invoke_void(void ()* asm ...)
501   // which is invalid because inline assembly blocks do not have addresses
502   // and can't be passed by pointer. The result is a crash with illegal IR.
503   if (isa<InlineAsm>(Callee))
504     return false;
505   StringRef CalleeName = Callee->getName();
506 
507   // The reason we include malloc/free here is to exclude the malloc/free
508   // calls generated in setjmp prep / cleanup routines.
509   if (CalleeName == "setjmp" || CalleeName == "malloc" || CalleeName == "free")
510     return false;
511 
512   // There are functions in JS glue code
513   if (CalleeName == "__resumeException" || CalleeName == "llvm_eh_typeid_for" ||
514       CalleeName == "saveSetjmp" || CalleeName == "testSetjmp" ||
515       CalleeName == "getTempRet0" || CalleeName == "setTempRet0")
516     return false;
517 
518   // __cxa_find_matching_catch_N functions cannot longjmp
519   if (Callee->getName().startswith("__cxa_find_matching_catch_"))
520     return false;
521 
522   // Exception-catching related functions
523   if (CalleeName == "__cxa_begin_catch" || CalleeName == "__cxa_end_catch" ||
524       CalleeName == "__cxa_allocate_exception" || CalleeName == "__cxa_throw" ||
525       CalleeName == "__clang_call_terminate")
526     return false;
527 
528   // Otherwise we don't know
529   return true;
530 }
531 
532 bool WebAssemblyLowerEmscriptenEHSjLj::isEmAsmCall(Module &M,
533                                                    const Value *Callee) const {
534   StringRef CalleeName = Callee->getName();
535   // This is an exhaustive list from Emscripten's <emscripten/em_asm.h>.
536   return CalleeName == "emscripten_asm_const_int" ||
537          CalleeName == "emscripten_asm_const_double" ||
538          CalleeName == "emscripten_asm_const_int_sync_on_main_thread" ||
539          CalleeName == "emscripten_asm_const_double_sync_on_main_thread" ||
540          CalleeName == "emscripten_asm_const_async_on_main_thread";
541 }
542 
543 // Generate testSetjmp function call seqence with preamble and postamble.
544 // The code this generates is equivalent to the following JavaScript code:
545 // if (%__THREW__.val != 0 & threwValue != 0) {
546 //   %label = _testSetjmp(mem[%__THREW__.val], setjmpTable, setjmpTableSize);
547 //   if (%label == 0)
548 //     emscripten_longjmp(%__THREW__.val, threwValue);
549 //   setTempRet0(threwValue);
550 // } else {
551 //   %label = -1;
552 // }
553 // %longjmp_result = getTempRet0();
554 //
555 // As output parameters. returns %label, %longjmp_result, and the BB the last
556 // instruction (%longjmp_result = ...) is in.
557 void WebAssemblyLowerEmscriptenEHSjLj::wrapTestSetjmp(
558     BasicBlock *BB, DebugLoc DL, Value *Threw, Value *SetjmpTable,
559     Value *SetjmpTableSize, Value *&Label, Value *&LongjmpResult,
560     BasicBlock *&EndBB) {
561   Function *F = BB->getParent();
562   LLVMContext &C = BB->getModule()->getContext();
563   IRBuilder<> IRB(C);
564   IRB.SetCurrentDebugLocation(DL);
565 
566   // if (%__THREW__.val != 0 & threwValue != 0)
567   IRB.SetInsertPoint(BB);
568   BasicBlock *ThenBB1 = BasicBlock::Create(C, "if.then1", F);
569   BasicBlock *ElseBB1 = BasicBlock::Create(C, "if.else1", F);
570   BasicBlock *EndBB1 = BasicBlock::Create(C, "if.end", F);
571   Value *ThrewCmp = IRB.CreateICmpNE(Threw, IRB.getInt32(0));
572   Value *ThrewValue = IRB.CreateLoad(IRB.getInt32Ty(), ThrewValueGV,
573                                      ThrewValueGV->getName() + ".val");
574   Value *ThrewValueCmp = IRB.CreateICmpNE(ThrewValue, IRB.getInt32(0));
575   Value *Cmp1 = IRB.CreateAnd(ThrewCmp, ThrewValueCmp, "cmp1");
576   IRB.CreateCondBr(Cmp1, ThenBB1, ElseBB1);
577 
578   // %label = _testSetjmp(mem[%__THREW__.val], _setjmpTable, _setjmpTableSize);
579   // if (%label == 0)
580   IRB.SetInsertPoint(ThenBB1);
581   BasicBlock *ThenBB2 = BasicBlock::Create(C, "if.then2", F);
582   BasicBlock *EndBB2 = BasicBlock::Create(C, "if.end2", F);
583   Value *ThrewInt = IRB.CreateIntToPtr(Threw, Type::getInt32PtrTy(C),
584                                        Threw->getName() + ".i32p");
585   Value *LoadedThrew = IRB.CreateLoad(IRB.getInt32Ty(), ThrewInt,
586                                       ThrewInt->getName() + ".loaded");
587   Value *ThenLabel = IRB.CreateCall(
588       TestSetjmpF, {LoadedThrew, SetjmpTable, SetjmpTableSize}, "label");
589   Value *Cmp2 = IRB.CreateICmpEQ(ThenLabel, IRB.getInt32(0));
590   IRB.CreateCondBr(Cmp2, ThenBB2, EndBB2);
591 
592   // emscripten_longjmp(%__THREW__.val, threwValue);
593   IRB.SetInsertPoint(ThenBB2);
594   IRB.CreateCall(EmLongjmpF, {Threw, ThrewValue});
595   IRB.CreateUnreachable();
596 
597   // setTempRet0(threwValue);
598   IRB.SetInsertPoint(EndBB2);
599   IRB.CreateCall(SetTempRet0Func, ThrewValue);
600   IRB.CreateBr(EndBB1);
601 
602   IRB.SetInsertPoint(ElseBB1);
603   IRB.CreateBr(EndBB1);
604 
605   // longjmp_result = getTempRet0();
606   IRB.SetInsertPoint(EndBB1);
607   PHINode *LabelPHI = IRB.CreatePHI(IRB.getInt32Ty(), 2, "label");
608   LabelPHI->addIncoming(ThenLabel, EndBB2);
609 
610   LabelPHI->addIncoming(IRB.getInt32(-1), ElseBB1);
611 
612   // Output parameter assignment
613   Label = LabelPHI;
614   EndBB = EndBB1;
615   LongjmpResult = IRB.CreateCall(GetTempRet0Func, None, "longjmp_result");
616 }
617 
618 void WebAssemblyLowerEmscriptenEHSjLj::rebuildSSA(Function &F) {
619   DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
620   DT.recalculate(F); // CFG has been changed
621   SSAUpdater SSA;
622   for (BasicBlock &BB : F) {
623     for (Instruction &I : BB) {
624       SSA.Initialize(I.getType(), I.getName());
625       SSA.AddAvailableValue(&BB, &I);
626       for (auto UI = I.use_begin(), UE = I.use_end(); UI != UE;) {
627         Use &U = *UI;
628         ++UI;
629         auto *User = cast<Instruction>(U.getUser());
630         if (auto *UserPN = dyn_cast<PHINode>(User))
631           if (UserPN->getIncomingBlock(U) == &BB)
632             continue;
633 
634         if (DT.dominates(&I, User))
635           continue;
636         SSA.RewriteUseAfterInsertions(U);
637       }
638     }
639   }
640 }
641 
642 bool WebAssemblyLowerEmscriptenEHSjLj::runOnModule(Module &M) {
643   LLVM_DEBUG(dbgs() << "********** Lower Emscripten EH & SjLj **********\n");
644 
645   LLVMContext &C = M.getContext();
646   IRBuilder<> IRB(C);
647 
648   Function *SetjmpF = M.getFunction("setjmp");
649   Function *LongjmpF = M.getFunction("longjmp");
650   bool SetjmpUsed = SetjmpF && !SetjmpF->use_empty();
651   bool LongjmpUsed = LongjmpF && !LongjmpF->use_empty();
652   bool DoSjLj = EnableSjLj && (SetjmpUsed || LongjmpUsed);
653 
654   // Declare (or get) global variables __THREW__, __threwValue, and
655   // getTempRet0/setTempRet0 function which are used in common for both
656   // exception handling and setjmp/longjmp handling
657   ThrewGV = getGlobalVariableI32(M, IRB, "__THREW__");
658   ThrewValueGV = getGlobalVariableI32(M, IRB, "__threwValue");
659   GetTempRet0Func =
660       Function::Create(FunctionType::get(IRB.getInt32Ty(), false),
661                        GlobalValue::ExternalLinkage, "getTempRet0", &M);
662   SetTempRet0Func = Function::Create(
663       FunctionType::get(IRB.getVoidTy(), IRB.getInt32Ty(), false),
664       GlobalValue::ExternalLinkage, "setTempRet0", &M);
665   GetTempRet0Func->setDoesNotThrow();
666   SetTempRet0Func->setDoesNotThrow();
667 
668   bool Changed = false;
669 
670   // Exception handling
671   if (EnableEH) {
672     // Register __resumeException function
673     FunctionType *ResumeFTy =
674         FunctionType::get(IRB.getVoidTy(), IRB.getInt8PtrTy(), false);
675     ResumeF = Function::Create(ResumeFTy, GlobalValue::ExternalLinkage,
676                                "__resumeException", &M);
677 
678     // Register llvm_eh_typeid_for function
679     FunctionType *EHTypeIDTy =
680         FunctionType::get(IRB.getInt32Ty(), IRB.getInt8PtrTy(), false);
681     EHTypeIDF = Function::Create(EHTypeIDTy, GlobalValue::ExternalLinkage,
682                                  "llvm_eh_typeid_for", &M);
683 
684     for (Function &F : M) {
685       if (F.isDeclaration())
686         continue;
687       Changed |= runEHOnFunction(F);
688     }
689   }
690 
691   // Setjmp/longjmp handling
692   if (DoSjLj) {
693     Changed = true; // We have setjmp or longjmp somewhere
694 
695     if (LongjmpF) {
696       // Replace all uses of longjmp with emscripten_longjmp_jmpbuf, which is
697       // defined in JS code
698       EmLongjmpJmpbufF = Function::Create(LongjmpF->getFunctionType(),
699                                           GlobalValue::ExternalLinkage,
700                                           "emscripten_longjmp_jmpbuf", &M);
701 
702       LongjmpF->replaceAllUsesWith(EmLongjmpJmpbufF);
703     }
704 
705     if (SetjmpF) {
706       // Register saveSetjmp function
707       FunctionType *SetjmpFTy = SetjmpF->getFunctionType();
708       FunctionType *FTy =
709           FunctionType::get(Type::getInt32PtrTy(C),
710                             {SetjmpFTy->getParamType(0), IRB.getInt32Ty(),
711                              Type::getInt32PtrTy(C), IRB.getInt32Ty()},
712                             false);
713       SaveSetjmpF =
714           Function::Create(FTy, GlobalValue::ExternalLinkage, "saveSetjmp", &M);
715 
716       // Register testSetjmp function
717       FTy = FunctionType::get(
718           IRB.getInt32Ty(),
719           {IRB.getInt32Ty(), Type::getInt32PtrTy(C), IRB.getInt32Ty()}, false);
720       TestSetjmpF =
721           Function::Create(FTy, GlobalValue::ExternalLinkage, "testSetjmp", &M);
722 
723       FTy = FunctionType::get(IRB.getVoidTy(),
724                               {IRB.getInt32Ty(), IRB.getInt32Ty()}, false);
725       EmLongjmpF = Function::Create(FTy, GlobalValue::ExternalLinkage,
726                                     "emscripten_longjmp", &M);
727 
728       // Only traverse functions that uses setjmp in order not to insert
729       // unnecessary prep / cleanup code in every function
730       SmallPtrSet<Function *, 8> SetjmpUsers;
731       for (User *U : SetjmpF->users()) {
732         auto *UI = cast<Instruction>(U);
733         SetjmpUsers.insert(UI->getFunction());
734       }
735       for (Function *F : SetjmpUsers)
736         runSjLjOnFunction(*F);
737     }
738   }
739 
740   if (!Changed) {
741     // Delete unused global variables and functions
742     if (ResumeF)
743       ResumeF->eraseFromParent();
744     if (EHTypeIDF)
745       EHTypeIDF->eraseFromParent();
746     if (EmLongjmpF)
747       EmLongjmpF->eraseFromParent();
748     if (SaveSetjmpF)
749       SaveSetjmpF->eraseFromParent();
750     if (TestSetjmpF)
751       TestSetjmpF->eraseFromParent();
752     return false;
753   }
754 
755   return true;
756 }
757 
758 bool WebAssemblyLowerEmscriptenEHSjLj::runEHOnFunction(Function &F) {
759   Module &M = *F.getParent();
760   LLVMContext &C = F.getContext();
761   IRBuilder<> IRB(C);
762   bool Changed = false;
763   SmallVector<Instruction *, 64> ToErase;
764   SmallPtrSet<LandingPadInst *, 32> LandingPads;
765   bool AllowExceptions = areAllExceptionsAllowed() ||
766                          EHWhitelistSet.count(std::string(F.getName()));
767 
768   for (BasicBlock &BB : F) {
769     auto *II = dyn_cast<InvokeInst>(BB.getTerminator());
770     if (!II)
771       continue;
772     Changed = true;
773     LandingPads.insert(II->getLandingPadInst());
774     IRB.SetInsertPoint(II);
775 
776     bool NeedInvoke = AllowExceptions && canThrow(II->getCalledValue());
777     if (NeedInvoke) {
778       // Wrap invoke with invoke wrapper and generate preamble/postamble
779       Value *Threw = wrapInvoke(II);
780       ToErase.push_back(II);
781 
782       // Insert a branch based on __THREW__ variable
783       Value *Cmp = IRB.CreateICmpEQ(Threw, IRB.getInt32(1), "cmp");
784       IRB.CreateCondBr(Cmp, II->getUnwindDest(), II->getNormalDest());
785 
786     } else {
787       // This can't throw, and we don't need this invoke, just replace it with a
788       // call+branch
789       SmallVector<Value *, 16> Args(II->arg_begin(), II->arg_end());
790       CallInst *NewCall =
791           IRB.CreateCall(II->getFunctionType(), II->getCalledValue(), Args);
792       NewCall->takeName(II);
793       NewCall->setCallingConv(II->getCallingConv());
794       NewCall->setDebugLoc(II->getDebugLoc());
795       NewCall->setAttributes(II->getAttributes());
796       II->replaceAllUsesWith(NewCall);
797       ToErase.push_back(II);
798 
799       IRB.CreateBr(II->getNormalDest());
800 
801       // Remove any PHI node entries from the exception destination
802       II->getUnwindDest()->removePredecessor(&BB);
803     }
804   }
805 
806   // Process resume instructions
807   for (BasicBlock &BB : F) {
808     // Scan the body of the basic block for resumes
809     for (Instruction &I : BB) {
810       auto *RI = dyn_cast<ResumeInst>(&I);
811       if (!RI)
812         continue;
813       Changed = true;
814 
815       // Split the input into legal values
816       Value *Input = RI->getValue();
817       IRB.SetInsertPoint(RI);
818       Value *Low = IRB.CreateExtractValue(Input, 0, "low");
819       // Create a call to __resumeException function
820       IRB.CreateCall(ResumeF, {Low});
821       // Add a terminator to the block
822       IRB.CreateUnreachable();
823       ToErase.push_back(RI);
824     }
825   }
826 
827   // Process llvm.eh.typeid.for intrinsics
828   for (BasicBlock &BB : F) {
829     for (Instruction &I : BB) {
830       auto *CI = dyn_cast<CallInst>(&I);
831       if (!CI)
832         continue;
833       const Function *Callee = CI->getCalledFunction();
834       if (!Callee)
835         continue;
836       if (Callee->getIntrinsicID() != Intrinsic::eh_typeid_for)
837         continue;
838       Changed = true;
839 
840       IRB.SetInsertPoint(CI);
841       CallInst *NewCI =
842           IRB.CreateCall(EHTypeIDF, CI->getArgOperand(0), "typeid");
843       CI->replaceAllUsesWith(NewCI);
844       ToErase.push_back(CI);
845     }
846   }
847 
848   // Look for orphan landingpads, can occur in blocks with no predecessors
849   for (BasicBlock &BB : F) {
850     Instruction *I = BB.getFirstNonPHI();
851     if (auto *LPI = dyn_cast<LandingPadInst>(I))
852       LandingPads.insert(LPI);
853   }
854   Changed |= !LandingPads.empty();
855 
856   // Handle all the landingpad for this function together, as multiple invokes
857   // may share a single lp
858   for (LandingPadInst *LPI : LandingPads) {
859     IRB.SetInsertPoint(LPI);
860     SmallVector<Value *, 16> FMCArgs;
861     for (unsigned I = 0, E = LPI->getNumClauses(); I < E; ++I) {
862       Constant *Clause = LPI->getClause(I);
863       // As a temporary workaround for the lack of aggregate varargs support
864       // in the interface between JS and wasm, break out filter operands into
865       // their component elements.
866       if (LPI->isFilter(I)) {
867         auto *ATy = cast<ArrayType>(Clause->getType());
868         for (unsigned J = 0, E = ATy->getNumElements(); J < E; ++J) {
869           Value *EV = IRB.CreateExtractValue(Clause, makeArrayRef(J), "filter");
870           FMCArgs.push_back(EV);
871         }
872       } else
873         FMCArgs.push_back(Clause);
874     }
875 
876     // Create a call to __cxa_find_matching_catch_N function
877     Function *FMCF = getFindMatchingCatch(M, FMCArgs.size());
878     CallInst *FMCI = IRB.CreateCall(FMCF, FMCArgs, "fmc");
879     Value *Undef = UndefValue::get(LPI->getType());
880     Value *Pair0 = IRB.CreateInsertValue(Undef, FMCI, 0, "pair0");
881     Value *TempRet0 = IRB.CreateCall(GetTempRet0Func, None, "tempret0");
882     Value *Pair1 = IRB.CreateInsertValue(Pair0, TempRet0, 1, "pair1");
883 
884     LPI->replaceAllUsesWith(Pair1);
885     ToErase.push_back(LPI);
886   }
887 
888   // Erase everything we no longer need in this function
889   for (Instruction *I : ToErase)
890     I->eraseFromParent();
891 
892   return Changed;
893 }
894 
895 bool WebAssemblyLowerEmscriptenEHSjLj::runSjLjOnFunction(Function &F) {
896   Module &M = *F.getParent();
897   LLVMContext &C = F.getContext();
898   IRBuilder<> IRB(C);
899   SmallVector<Instruction *, 64> ToErase;
900   // Vector of %setjmpTable values
901   std::vector<Instruction *> SetjmpTableInsts;
902   // Vector of %setjmpTableSize values
903   std::vector<Instruction *> SetjmpTableSizeInsts;
904 
905   // Setjmp preparation
906 
907   // This instruction effectively means %setjmpTableSize = 4.
908   // We create this as an instruction intentionally, and we don't want to fold
909   // this instruction to a constant 4, because this value will be used in
910   // SSAUpdater.AddAvailableValue(...) later.
911   BasicBlock &EntryBB = F.getEntryBlock();
912   DebugLoc FirstDL = EntryBB.begin()->getDebugLoc();
913   BinaryOperator *SetjmpTableSize = BinaryOperator::Create(
914       Instruction::Add, IRB.getInt32(4), IRB.getInt32(0), "setjmpTableSize",
915       &*EntryBB.getFirstInsertionPt());
916   SetjmpTableSize->setDebugLoc(FirstDL);
917   // setjmpTable = (int *) malloc(40);
918   Instruction *SetjmpTable = CallInst::CreateMalloc(
919       SetjmpTableSize, IRB.getInt32Ty(), IRB.getInt32Ty(), IRB.getInt32(40),
920       nullptr, nullptr, "setjmpTable");
921   SetjmpTable->setDebugLoc(FirstDL);
922   // CallInst::CreateMalloc may return a bitcast instruction if the result types
923   // mismatch. We need to set the debug loc for the original call too.
924   auto *MallocCall = SetjmpTable->stripPointerCasts();
925   if (auto *MallocCallI = dyn_cast<Instruction>(MallocCall)) {
926     MallocCallI->setDebugLoc(FirstDL);
927   }
928   // setjmpTable[0] = 0;
929   IRB.SetInsertPoint(SetjmpTableSize);
930   IRB.CreateStore(IRB.getInt32(0), SetjmpTable);
931   SetjmpTableInsts.push_back(SetjmpTable);
932   SetjmpTableSizeInsts.push_back(SetjmpTableSize);
933 
934   // Setjmp transformation
935   std::vector<PHINode *> SetjmpRetPHIs;
936   Function *SetjmpF = M.getFunction("setjmp");
937   for (User *U : SetjmpF->users()) {
938     auto *CI = dyn_cast<CallInst>(U);
939     if (!CI)
940       report_fatal_error("Does not support indirect calls to setjmp");
941 
942     BasicBlock *BB = CI->getParent();
943     if (BB->getParent() != &F) // in other function
944       continue;
945 
946     // The tail is everything right after the call, and will be reached once
947     // when setjmp is called, and later when longjmp returns to the setjmp
948     BasicBlock *Tail = SplitBlock(BB, CI->getNextNode());
949     // Add a phi to the tail, which will be the output of setjmp, which
950     // indicates if this is the first call or a longjmp back. The phi directly
951     // uses the right value based on where we arrive from
952     IRB.SetInsertPoint(Tail->getFirstNonPHI());
953     PHINode *SetjmpRet = IRB.CreatePHI(IRB.getInt32Ty(), 2, "setjmp.ret");
954 
955     // setjmp initial call returns 0
956     SetjmpRet->addIncoming(IRB.getInt32(0), BB);
957     // The proper output is now this, not the setjmp call itself
958     CI->replaceAllUsesWith(SetjmpRet);
959     // longjmp returns to the setjmp will add themselves to this phi
960     SetjmpRetPHIs.push_back(SetjmpRet);
961 
962     // Fix call target
963     // Our index in the function is our place in the array + 1 to avoid index
964     // 0, because index 0 means the longjmp is not ours to handle.
965     IRB.SetInsertPoint(CI);
966     Value *Args[] = {CI->getArgOperand(0), IRB.getInt32(SetjmpRetPHIs.size()),
967                      SetjmpTable, SetjmpTableSize};
968     Instruction *NewSetjmpTable =
969         IRB.CreateCall(SaveSetjmpF, Args, "setjmpTable");
970     Instruction *NewSetjmpTableSize =
971         IRB.CreateCall(GetTempRet0Func, None, "setjmpTableSize");
972     SetjmpTableInsts.push_back(NewSetjmpTable);
973     SetjmpTableSizeInsts.push_back(NewSetjmpTableSize);
974     ToErase.push_back(CI);
975   }
976 
977   // Update each call that can longjmp so it can return to a setjmp where
978   // relevant.
979 
980   // Because we are creating new BBs while processing and don't want to make
981   // all these newly created BBs candidates again for longjmp processing, we
982   // first make the vector of candidate BBs.
983   std::vector<BasicBlock *> BBs;
984   for (BasicBlock &BB : F)
985     BBs.push_back(&BB);
986 
987   // BBs.size() will change within the loop, so we query it every time
988   for (unsigned I = 0; I < BBs.size(); I++) {
989     BasicBlock *BB = BBs[I];
990     for (Instruction &I : *BB) {
991       assert(!isa<InvokeInst>(&I));
992       auto *CI = dyn_cast<CallInst>(&I);
993       if (!CI)
994         continue;
995 
996       const Value *Callee = CI->getCalledValue();
997       if (!canLongjmp(M, Callee))
998         continue;
999       if (isEmAsmCall(M, Callee))
1000         report_fatal_error("Cannot use EM_ASM* alongside setjmp/longjmp in " +
1001                                F.getName() +
1002                                ". Please consider using EM_JS, or move the "
1003                                "EM_ASM into another function.",
1004                            false);
1005 
1006       Value *Threw = nullptr;
1007       BasicBlock *Tail;
1008       if (Callee->getName().startswith("__invoke_")) {
1009         // If invoke wrapper has already been generated for this call in
1010         // previous EH phase, search for the load instruction
1011         // %__THREW__.val = __THREW__;
1012         // in postamble after the invoke wrapper call
1013         LoadInst *ThrewLI = nullptr;
1014         StoreInst *ThrewResetSI = nullptr;
1015         for (auto I = std::next(BasicBlock::iterator(CI)), IE = BB->end();
1016              I != IE; ++I) {
1017           if (auto *LI = dyn_cast<LoadInst>(I))
1018             if (auto *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand()))
1019               if (GV == ThrewGV) {
1020                 Threw = ThrewLI = LI;
1021                 break;
1022               }
1023         }
1024         // Search for the store instruction after the load above
1025         // __THREW__ = 0;
1026         for (auto I = std::next(BasicBlock::iterator(ThrewLI)), IE = BB->end();
1027              I != IE; ++I) {
1028           if (auto *SI = dyn_cast<StoreInst>(I))
1029             if (auto *GV = dyn_cast<GlobalVariable>(SI->getPointerOperand()))
1030               if (GV == ThrewGV && SI->getValueOperand() == IRB.getInt32(0)) {
1031                 ThrewResetSI = SI;
1032                 break;
1033               }
1034         }
1035         assert(Threw && ThrewLI && "Cannot find __THREW__ load after invoke");
1036         assert(ThrewResetSI && "Cannot find __THREW__ store after invoke");
1037         Tail = SplitBlock(BB, ThrewResetSI->getNextNode());
1038 
1039       } else {
1040         // Wrap call with invoke wrapper and generate preamble/postamble
1041         Threw = wrapInvoke(CI);
1042         ToErase.push_back(CI);
1043         Tail = SplitBlock(BB, CI->getNextNode());
1044       }
1045 
1046       // We need to replace the terminator in Tail - SplitBlock makes BB go
1047       // straight to Tail, we need to check if a longjmp occurred, and go to the
1048       // right setjmp-tail if so
1049       ToErase.push_back(BB->getTerminator());
1050 
1051       // Generate a function call to testSetjmp function and preamble/postamble
1052       // code to figure out (1) whether longjmp occurred (2) if longjmp
1053       // occurred, which setjmp it corresponds to
1054       Value *Label = nullptr;
1055       Value *LongjmpResult = nullptr;
1056       BasicBlock *EndBB = nullptr;
1057       wrapTestSetjmp(BB, CI->getDebugLoc(), Threw, SetjmpTable, SetjmpTableSize,
1058                      Label, LongjmpResult, EndBB);
1059       assert(Label && LongjmpResult && EndBB);
1060 
1061       // Create switch instruction
1062       IRB.SetInsertPoint(EndBB);
1063       IRB.SetCurrentDebugLocation(EndBB->getInstList().back().getDebugLoc());
1064       SwitchInst *SI = IRB.CreateSwitch(Label, Tail, SetjmpRetPHIs.size());
1065       // -1 means no longjmp happened, continue normally (will hit the default
1066       // switch case). 0 means a longjmp that is not ours to handle, needs a
1067       // rethrow. Otherwise the index is the same as the index in P+1 (to avoid
1068       // 0).
1069       for (unsigned I = 0; I < SetjmpRetPHIs.size(); I++) {
1070         SI->addCase(IRB.getInt32(I + 1), SetjmpRetPHIs[I]->getParent());
1071         SetjmpRetPHIs[I]->addIncoming(LongjmpResult, EndBB);
1072       }
1073 
1074       // We are splitting the block here, and must continue to find other calls
1075       // in the block - which is now split. so continue to traverse in the Tail
1076       BBs.push_back(Tail);
1077     }
1078   }
1079 
1080   // Erase everything we no longer need in this function
1081   for (Instruction *I : ToErase)
1082     I->eraseFromParent();
1083 
1084   // Free setjmpTable buffer before each return instruction
1085   for (BasicBlock &BB : F) {
1086     Instruction *TI = BB.getTerminator();
1087     if (isa<ReturnInst>(TI)) {
1088       auto *Free = CallInst::CreateFree(SetjmpTable, TI);
1089       Free->setDebugLoc(TI->getDebugLoc());
1090       // CallInst::CreateFree may create a bitcast instruction if its argument
1091       // types mismatch. We need to set the debug loc for the bitcast too.
1092       if (auto *FreeCallI = dyn_cast<CallInst>(Free)) {
1093         if (auto *BitCastI = dyn_cast<BitCastInst>(FreeCallI->getArgOperand(0)))
1094           BitCastI->setDebugLoc(TI->getDebugLoc());
1095       }
1096     }
1097   }
1098 
1099   // Every call to saveSetjmp can change setjmpTable and setjmpTableSize
1100   // (when buffer reallocation occurs)
1101   // entry:
1102   //   setjmpTableSize = 4;
1103   //   setjmpTable = (int *) malloc(40);
1104   //   setjmpTable[0] = 0;
1105   // ...
1106   // somebb:
1107   //   setjmpTable = saveSetjmp(buf, label, setjmpTable, setjmpTableSize);
1108   //   setjmpTableSize = getTempRet0();
1109   // So we need to make sure the SSA for these variables is valid so that every
1110   // saveSetjmp and testSetjmp calls have the correct arguments.
1111   SSAUpdater SetjmpTableSSA;
1112   SSAUpdater SetjmpTableSizeSSA;
1113   SetjmpTableSSA.Initialize(Type::getInt32PtrTy(C), "setjmpTable");
1114   SetjmpTableSizeSSA.Initialize(Type::getInt32Ty(C), "setjmpTableSize");
1115   for (Instruction *I : SetjmpTableInsts)
1116     SetjmpTableSSA.AddAvailableValue(I->getParent(), I);
1117   for (Instruction *I : SetjmpTableSizeInsts)
1118     SetjmpTableSizeSSA.AddAvailableValue(I->getParent(), I);
1119 
1120   for (auto UI = SetjmpTable->use_begin(), UE = SetjmpTable->use_end();
1121        UI != UE;) {
1122     // Grab the use before incrementing the iterator.
1123     Use &U = *UI;
1124     // Increment the iterator before removing the use from the list.
1125     ++UI;
1126     if (auto *I = dyn_cast<Instruction>(U.getUser()))
1127       if (I->getParent() != &EntryBB)
1128         SetjmpTableSSA.RewriteUse(U);
1129   }
1130   for (auto UI = SetjmpTableSize->use_begin(), UE = SetjmpTableSize->use_end();
1131        UI != UE;) {
1132     Use &U = *UI;
1133     ++UI;
1134     if (auto *I = dyn_cast<Instruction>(U.getUser()))
1135       if (I->getParent() != &EntryBB)
1136         SetjmpTableSizeSSA.RewriteUse(U);
1137   }
1138 
1139   // Finally, our modifications to the cfg can break dominance of SSA variables.
1140   // For example, in this code,
1141   // if (x()) { .. setjmp() .. }
1142   // if (y()) { .. longjmp() .. }
1143   // We must split the longjmp block, and it can jump into the block splitted
1144   // from setjmp one. But that means that when we split the setjmp block, it's
1145   // first part no longer dominates its second part - there is a theoretically
1146   // possible control flow path where x() is false, then y() is true and we
1147   // reach the second part of the setjmp block, without ever reaching the first
1148   // part. So, we rebuild SSA form here.
1149   rebuildSSA(F);
1150   return true;
1151 }
1152