xref: /freebsd/contrib/llvm-project/llvm/lib/Target/WebAssembly/WebAssemblyLowerEmscriptenEHSjLj.cpp (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
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/ADT/StringExtras.h"
212 #include "llvm/IR/DebugInfoMetadata.h"
213 #include "llvm/IR/Dominators.h"
214 #include "llvm/IR/IRBuilder.h"
215 #include "llvm/Support/CommandLine.h"
216 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
217 #include "llvm/Transforms/Utils/SSAUpdater.h"
218 
219 using namespace llvm;
220 
221 #define DEBUG_TYPE "wasm-lower-em-ehsjlj"
222 
223 static cl::list<std::string>
224     EHAllowlist("emscripten-cxx-exceptions-allowed",
225                 cl::desc("The list of function names in which Emscripten-style "
226                          "exception handling is enabled (see emscripten "
227                          "EMSCRIPTEN_CATCHING_ALLOWED options)"),
228                 cl::CommaSeparated);
229 
230 namespace {
231 class WebAssemblyLowerEmscriptenEHSjLj final : public ModulePass {
232   bool EnableEH;   // Enable exception handling
233   bool EnableSjLj; // Enable setjmp/longjmp handling
234 
235   GlobalVariable *ThrewGV = nullptr;
236   GlobalVariable *ThrewValueGV = nullptr;
237   Function *GetTempRet0Func = nullptr;
238   Function *SetTempRet0Func = nullptr;
239   Function *ResumeF = nullptr;
240   Function *EHTypeIDF = nullptr;
241   Function *EmLongjmpF = nullptr;
242   Function *EmLongjmpJmpbufF = nullptr;
243   Function *SaveSetjmpF = nullptr;
244   Function *TestSetjmpF = nullptr;
245 
246   // __cxa_find_matching_catch_N functions.
247   // Indexed by the number of clauses in an original landingpad instruction.
248   DenseMap<int, Function *> FindMatchingCatches;
249   // Map of <function signature string, invoke_ wrappers>
250   StringMap<Function *> InvokeWrappers;
251   // Set of allowed function names for exception handling
252   std::set<std::string> EHAllowlistSet;
253 
254   StringRef getPassName() const override {
255     return "WebAssembly Lower Emscripten Exceptions";
256   }
257 
258   bool runEHOnFunction(Function &F);
259   bool runSjLjOnFunction(Function &F);
260   Function *getFindMatchingCatch(Module &M, unsigned NumClauses);
261 
262   Value *wrapInvoke(CallBase *CI);
263   void wrapTestSetjmp(BasicBlock *BB, DebugLoc DL, Value *Threw,
264                       Value *SetjmpTable, Value *SetjmpTableSize, Value *&Label,
265                       Value *&LongjmpResult, BasicBlock *&EndBB);
266   Function *getInvokeWrapper(CallBase *CI);
267 
268   bool areAllExceptionsAllowed() const { return EHAllowlistSet.empty(); }
269   bool canLongjmp(Module &M, const Value *Callee) const;
270   bool isEmAsmCall(Module &M, const Value *Callee) const;
271 
272   void rebuildSSA(Function &F);
273 
274 public:
275   static char ID;
276 
277   WebAssemblyLowerEmscriptenEHSjLj(bool EnableEH = true, bool EnableSjLj = true)
278       : ModulePass(ID), EnableEH(EnableEH), EnableSjLj(EnableSjLj) {
279     EHAllowlistSet.insert(EHAllowlist.begin(), EHAllowlist.end());
280   }
281   bool runOnModule(Module &M) override;
282 
283   void getAnalysisUsage(AnalysisUsage &AU) const override {
284     AU.addRequired<DominatorTreeWrapperPass>();
285   }
286 };
287 } // End anonymous namespace
288 
289 char WebAssemblyLowerEmscriptenEHSjLj::ID = 0;
290 INITIALIZE_PASS(WebAssemblyLowerEmscriptenEHSjLj, DEBUG_TYPE,
291                 "WebAssembly Lower Emscripten Exceptions / Setjmp / Longjmp",
292                 false, false)
293 
294 ModulePass *llvm::createWebAssemblyLowerEmscriptenEHSjLj(bool EnableEH,
295                                                          bool EnableSjLj) {
296   return new WebAssemblyLowerEmscriptenEHSjLj(EnableEH, EnableSjLj);
297 }
298 
299 static bool canThrow(const Value *V) {
300   if (const auto *F = dyn_cast<const Function>(V)) {
301     // Intrinsics cannot throw
302     if (F->isIntrinsic())
303       return false;
304     StringRef Name = F->getName();
305     // leave setjmp and longjmp (mostly) alone, we process them properly later
306     if (Name == "setjmp" || Name == "longjmp")
307       return false;
308     return !F->doesNotThrow();
309   }
310   // not a function, so an indirect call - can throw, we can't tell
311   return true;
312 }
313 
314 // Get a global variable with the given name.  If it doesn't exist declare it,
315 // which will generate an import and asssumes that it will exist at link time.
316 static GlobalVariable *getGlobalVariableI32(Module &M, IRBuilder<> &IRB,
317                                             const char *Name) {
318 
319   auto *GV =
320       dyn_cast<GlobalVariable>(M.getOrInsertGlobal(Name, IRB.getInt32Ty()));
321   if (!GV)
322     report_fatal_error(Twine("unable to create global: ") + Name);
323 
324   return GV;
325 }
326 
327 // Simple function name mangler.
328 // This function simply takes LLVM's string representation of parameter types
329 // and concatenate them with '_'. There are non-alphanumeric characters but llc
330 // is ok with it, and we need to postprocess these names after the lowering
331 // phase anyway.
332 static std::string getSignature(FunctionType *FTy) {
333   std::string Sig;
334   raw_string_ostream OS(Sig);
335   OS << *FTy->getReturnType();
336   for (Type *ParamTy : FTy->params())
337     OS << "_" << *ParamTy;
338   if (FTy->isVarArg())
339     OS << "_...";
340   Sig = OS.str();
341   Sig.erase(remove_if(Sig, isSpace), Sig.end());
342   // When s2wasm parses .s file, a comma means the end of an argument. So a
343   // mangled function name can contain any character but a comma.
344   std::replace(Sig.begin(), Sig.end(), ',', '.');
345   return Sig;
346 }
347 
348 static Function *getEmscriptenFunction(FunctionType *Ty, const Twine &Name,
349                                        Module *M) {
350   Function* F = Function::Create(Ty, GlobalValue::ExternalLinkage, Name, M);
351   // Tell the linker that this function is expected to be imported from the
352   // 'env' module.
353   if (!F->hasFnAttribute("wasm-import-module")) {
354     llvm::AttrBuilder B;
355     B.addAttribute("wasm-import-module", "env");
356     F->addAttributes(llvm::AttributeList::FunctionIndex, B);
357   }
358   if (!F->hasFnAttribute("wasm-import-name")) {
359     llvm::AttrBuilder B;
360     B.addAttribute("wasm-import-name", F->getName());
361     F->addAttributes(llvm::AttributeList::FunctionIndex, B);
362   }
363   return F;
364 }
365 
366 // Returns __cxa_find_matching_catch_N function, where N = NumClauses + 2.
367 // This is because a landingpad instruction contains two more arguments, a
368 // personality function and a cleanup bit, and __cxa_find_matching_catch_N
369 // functions are named after the number of arguments in the original landingpad
370 // instruction.
371 Function *
372 WebAssemblyLowerEmscriptenEHSjLj::getFindMatchingCatch(Module &M,
373                                                        unsigned NumClauses) {
374   if (FindMatchingCatches.count(NumClauses))
375     return FindMatchingCatches[NumClauses];
376   PointerType *Int8PtrTy = Type::getInt8PtrTy(M.getContext());
377   SmallVector<Type *, 16> Args(NumClauses, Int8PtrTy);
378   FunctionType *FTy = FunctionType::get(Int8PtrTy, Args, false);
379   Function *F = getEmscriptenFunction(
380       FTy, "__cxa_find_matching_catch_" + Twine(NumClauses + 2), &M);
381   FindMatchingCatches[NumClauses] = F;
382   return F;
383 }
384 
385 // Generate invoke wrapper seqence with preamble and postamble
386 // Preamble:
387 // __THREW__ = 0;
388 // Postamble:
389 // %__THREW__.val = __THREW__; __THREW__ = 0;
390 // Returns %__THREW__.val, which indicates whether an exception is thrown (or
391 // whether longjmp occurred), for future use.
392 Value *WebAssemblyLowerEmscriptenEHSjLj::wrapInvoke(CallBase *CI) {
393   LLVMContext &C = CI->getModule()->getContext();
394 
395   // If we are calling a function that is noreturn, we must remove that
396   // attribute. The code we insert here does expect it to return, after we
397   // catch the exception.
398   if (CI->doesNotReturn()) {
399     if (auto *F = CI->getCalledFunction())
400       F->removeFnAttr(Attribute::NoReturn);
401     CI->removeAttribute(AttributeList::FunctionIndex, Attribute::NoReturn);
402   }
403 
404   IRBuilder<> IRB(C);
405   IRB.SetInsertPoint(CI);
406 
407   // Pre-invoke
408   // __THREW__ = 0;
409   IRB.CreateStore(IRB.getInt32(0), ThrewGV);
410 
411   // Invoke function wrapper in JavaScript
412   SmallVector<Value *, 16> Args;
413   // Put the pointer to the callee as first argument, so it can be called
414   // within the invoke wrapper later
415   Args.push_back(CI->getCalledOperand());
416   Args.append(CI->arg_begin(), CI->arg_end());
417   CallInst *NewCall = IRB.CreateCall(getInvokeWrapper(CI), Args);
418   NewCall->takeName(CI);
419   NewCall->setCallingConv(CallingConv::WASM_EmscriptenInvoke);
420   NewCall->setDebugLoc(CI->getDebugLoc());
421 
422   // Because we added the pointer to the callee as first argument, all
423   // argument attribute indices have to be incremented by one.
424   SmallVector<AttributeSet, 8> ArgAttributes;
425   const AttributeList &InvokeAL = CI->getAttributes();
426 
427   // No attributes for the callee pointer.
428   ArgAttributes.push_back(AttributeSet());
429   // Copy the argument attributes from the original
430   for (unsigned I = 0, E = CI->getNumArgOperands(); I < E; ++I)
431     ArgAttributes.push_back(InvokeAL.getParamAttributes(I));
432 
433   AttrBuilder FnAttrs(InvokeAL.getFnAttributes());
434   if (FnAttrs.contains(Attribute::AllocSize)) {
435     // The allocsize attribute (if any) referes to parameters by index and needs
436     // to be adjusted.
437     unsigned SizeArg;
438     Optional<unsigned> NEltArg;
439     std::tie(SizeArg, NEltArg) = FnAttrs.getAllocSizeArgs();
440     SizeArg += 1;
441     if (NEltArg.hasValue())
442       NEltArg = NEltArg.getValue() + 1;
443     FnAttrs.addAllocSizeAttr(SizeArg, NEltArg);
444   }
445 
446   // Reconstruct the AttributesList based on the vector we constructed.
447   AttributeList NewCallAL =
448       AttributeList::get(C, AttributeSet::get(C, FnAttrs),
449                          InvokeAL.getRetAttributes(), ArgAttributes);
450   NewCall->setAttributes(NewCallAL);
451 
452   CI->replaceAllUsesWith(NewCall);
453 
454   // Post-invoke
455   // %__THREW__.val = __THREW__; __THREW__ = 0;
456   Value *Threw =
457       IRB.CreateLoad(IRB.getInt32Ty(), ThrewGV, ThrewGV->getName() + ".val");
458   IRB.CreateStore(IRB.getInt32(0), ThrewGV);
459   return Threw;
460 }
461 
462 // Get matching invoke wrapper based on callee signature
463 Function *WebAssemblyLowerEmscriptenEHSjLj::getInvokeWrapper(CallBase *CI) {
464   Module *M = CI->getModule();
465   SmallVector<Type *, 16> ArgTys;
466   FunctionType *CalleeFTy = CI->getFunctionType();
467 
468   std::string Sig = getSignature(CalleeFTy);
469   if (InvokeWrappers.find(Sig) != InvokeWrappers.end())
470     return InvokeWrappers[Sig];
471 
472   // Put the pointer to the callee as first argument
473   ArgTys.push_back(PointerType::getUnqual(CalleeFTy));
474   // Add argument types
475   ArgTys.append(CalleeFTy->param_begin(), CalleeFTy->param_end());
476 
477   FunctionType *FTy = FunctionType::get(CalleeFTy->getReturnType(), ArgTys,
478                                         CalleeFTy->isVarArg());
479   Function *F = getEmscriptenFunction(FTy, "__invoke_" + Sig, M);
480   InvokeWrappers[Sig] = F;
481   return F;
482 }
483 
484 bool WebAssemblyLowerEmscriptenEHSjLj::canLongjmp(Module &M,
485                                                   const Value *Callee) const {
486   if (auto *CalleeF = dyn_cast<Function>(Callee))
487     if (CalleeF->isIntrinsic())
488       return false;
489 
490   // Attempting to transform inline assembly will result in something like:
491   //     call void @__invoke_void(void ()* asm ...)
492   // which is invalid because inline assembly blocks do not have addresses
493   // and can't be passed by pointer. The result is a crash with illegal IR.
494   if (isa<InlineAsm>(Callee))
495     return false;
496   StringRef CalleeName = Callee->getName();
497 
498   // The reason we include malloc/free here is to exclude the malloc/free
499   // calls generated in setjmp prep / cleanup routines.
500   if (CalleeName == "setjmp" || CalleeName == "malloc" || CalleeName == "free")
501     return false;
502 
503   // There are functions in JS glue code
504   if (CalleeName == "__resumeException" || CalleeName == "llvm_eh_typeid_for" ||
505       CalleeName == "saveSetjmp" || CalleeName == "testSetjmp" ||
506       CalleeName == "getTempRet0" || CalleeName == "setTempRet0")
507     return false;
508 
509   // __cxa_find_matching_catch_N functions cannot longjmp
510   if (Callee->getName().startswith("__cxa_find_matching_catch_"))
511     return false;
512 
513   // Exception-catching related functions
514   if (CalleeName == "__cxa_begin_catch" || CalleeName == "__cxa_end_catch" ||
515       CalleeName == "__cxa_allocate_exception" || CalleeName == "__cxa_throw" ||
516       CalleeName == "__clang_call_terminate")
517     return false;
518 
519   // Otherwise we don't know
520   return true;
521 }
522 
523 bool WebAssemblyLowerEmscriptenEHSjLj::isEmAsmCall(Module &M,
524                                                    const Value *Callee) const {
525   StringRef CalleeName = Callee->getName();
526   // This is an exhaustive list from Emscripten's <emscripten/em_asm.h>.
527   return CalleeName == "emscripten_asm_const_int" ||
528          CalleeName == "emscripten_asm_const_double" ||
529          CalleeName == "emscripten_asm_const_int_sync_on_main_thread" ||
530          CalleeName == "emscripten_asm_const_double_sync_on_main_thread" ||
531          CalleeName == "emscripten_asm_const_async_on_main_thread";
532 }
533 
534 // Generate testSetjmp function call seqence with preamble and postamble.
535 // The code this generates is equivalent to the following JavaScript code:
536 // if (%__THREW__.val != 0 & threwValue != 0) {
537 //   %label = _testSetjmp(mem[%__THREW__.val], setjmpTable, setjmpTableSize);
538 //   if (%label == 0)
539 //     emscripten_longjmp(%__THREW__.val, threwValue);
540 //   setTempRet0(threwValue);
541 // } else {
542 //   %label = -1;
543 // }
544 // %longjmp_result = getTempRet0();
545 //
546 // As output parameters. returns %label, %longjmp_result, and the BB the last
547 // instruction (%longjmp_result = ...) is in.
548 void WebAssemblyLowerEmscriptenEHSjLj::wrapTestSetjmp(
549     BasicBlock *BB, DebugLoc DL, Value *Threw, Value *SetjmpTable,
550     Value *SetjmpTableSize, Value *&Label, Value *&LongjmpResult,
551     BasicBlock *&EndBB) {
552   Function *F = BB->getParent();
553   LLVMContext &C = BB->getModule()->getContext();
554   IRBuilder<> IRB(C);
555   IRB.SetCurrentDebugLocation(DL);
556 
557   // if (%__THREW__.val != 0 & threwValue != 0)
558   IRB.SetInsertPoint(BB);
559   BasicBlock *ThenBB1 = BasicBlock::Create(C, "if.then1", F);
560   BasicBlock *ElseBB1 = BasicBlock::Create(C, "if.else1", F);
561   BasicBlock *EndBB1 = BasicBlock::Create(C, "if.end", F);
562   Value *ThrewCmp = IRB.CreateICmpNE(Threw, IRB.getInt32(0));
563   Value *ThrewValue = IRB.CreateLoad(IRB.getInt32Ty(), ThrewValueGV,
564                                      ThrewValueGV->getName() + ".val");
565   Value *ThrewValueCmp = IRB.CreateICmpNE(ThrewValue, IRB.getInt32(0));
566   Value *Cmp1 = IRB.CreateAnd(ThrewCmp, ThrewValueCmp, "cmp1");
567   IRB.CreateCondBr(Cmp1, ThenBB1, ElseBB1);
568 
569   // %label = _testSetjmp(mem[%__THREW__.val], _setjmpTable, _setjmpTableSize);
570   // if (%label == 0)
571   IRB.SetInsertPoint(ThenBB1);
572   BasicBlock *ThenBB2 = BasicBlock::Create(C, "if.then2", F);
573   BasicBlock *EndBB2 = BasicBlock::Create(C, "if.end2", F);
574   Value *ThrewInt = IRB.CreateIntToPtr(Threw, Type::getInt32PtrTy(C),
575                                        Threw->getName() + ".i32p");
576   Value *LoadedThrew = IRB.CreateLoad(IRB.getInt32Ty(), ThrewInt,
577                                       ThrewInt->getName() + ".loaded");
578   Value *ThenLabel = IRB.CreateCall(
579       TestSetjmpF, {LoadedThrew, SetjmpTable, SetjmpTableSize}, "label");
580   Value *Cmp2 = IRB.CreateICmpEQ(ThenLabel, IRB.getInt32(0));
581   IRB.CreateCondBr(Cmp2, ThenBB2, EndBB2);
582 
583   // emscripten_longjmp(%__THREW__.val, threwValue);
584   IRB.SetInsertPoint(ThenBB2);
585   IRB.CreateCall(EmLongjmpF, {Threw, ThrewValue});
586   IRB.CreateUnreachable();
587 
588   // setTempRet0(threwValue);
589   IRB.SetInsertPoint(EndBB2);
590   IRB.CreateCall(SetTempRet0Func, ThrewValue);
591   IRB.CreateBr(EndBB1);
592 
593   IRB.SetInsertPoint(ElseBB1);
594   IRB.CreateBr(EndBB1);
595 
596   // longjmp_result = getTempRet0();
597   IRB.SetInsertPoint(EndBB1);
598   PHINode *LabelPHI = IRB.CreatePHI(IRB.getInt32Ty(), 2, "label");
599   LabelPHI->addIncoming(ThenLabel, EndBB2);
600 
601   LabelPHI->addIncoming(IRB.getInt32(-1), ElseBB1);
602 
603   // Output parameter assignment
604   Label = LabelPHI;
605   EndBB = EndBB1;
606   LongjmpResult = IRB.CreateCall(GetTempRet0Func, None, "longjmp_result");
607 }
608 
609 void WebAssemblyLowerEmscriptenEHSjLj::rebuildSSA(Function &F) {
610   DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
611   DT.recalculate(F); // CFG has been changed
612   SSAUpdater SSA;
613   for (BasicBlock &BB : F) {
614     for (Instruction &I : BB) {
615       SSA.Initialize(I.getType(), I.getName());
616       SSA.AddAvailableValue(&BB, &I);
617       for (auto UI = I.use_begin(), UE = I.use_end(); UI != UE;) {
618         Use &U = *UI;
619         ++UI;
620         auto *User = cast<Instruction>(U.getUser());
621         if (auto *UserPN = dyn_cast<PHINode>(User))
622           if (UserPN->getIncomingBlock(U) == &BB)
623             continue;
624 
625         if (DT.dominates(&I, User))
626           continue;
627         SSA.RewriteUseAfterInsertions(U);
628       }
629     }
630   }
631 }
632 
633 bool WebAssemblyLowerEmscriptenEHSjLj::runOnModule(Module &M) {
634   LLVM_DEBUG(dbgs() << "********** Lower Emscripten EH & SjLj **********\n");
635 
636   LLVMContext &C = M.getContext();
637   IRBuilder<> IRB(C);
638 
639   Function *SetjmpF = M.getFunction("setjmp");
640   Function *LongjmpF = M.getFunction("longjmp");
641   bool SetjmpUsed = SetjmpF && !SetjmpF->use_empty();
642   bool LongjmpUsed = LongjmpF && !LongjmpF->use_empty();
643   bool DoSjLj = EnableSjLj && (SetjmpUsed || LongjmpUsed);
644 
645   // Declare (or get) global variables __THREW__, __threwValue, and
646   // getTempRet0/setTempRet0 function which are used in common for both
647   // exception handling and setjmp/longjmp handling
648   ThrewGV = getGlobalVariableI32(M, IRB, "__THREW__");
649   ThrewValueGV = getGlobalVariableI32(M, IRB, "__threwValue");
650   GetTempRet0Func = getEmscriptenFunction(
651       FunctionType::get(IRB.getInt32Ty(), false), "getTempRet0", &M);
652   SetTempRet0Func = getEmscriptenFunction(
653       FunctionType::get(IRB.getVoidTy(), IRB.getInt32Ty(), false),
654       "setTempRet0", &M);
655   GetTempRet0Func->setDoesNotThrow();
656   SetTempRet0Func->setDoesNotThrow();
657 
658   bool Changed = false;
659 
660   // Exception handling
661   if (EnableEH) {
662     // Register __resumeException function
663     FunctionType *ResumeFTy =
664         FunctionType::get(IRB.getVoidTy(), IRB.getInt8PtrTy(), false);
665     ResumeF = getEmscriptenFunction(ResumeFTy, "__resumeException", &M);
666 
667     // Register llvm_eh_typeid_for function
668     FunctionType *EHTypeIDTy =
669         FunctionType::get(IRB.getInt32Ty(), IRB.getInt8PtrTy(), false);
670     EHTypeIDF = getEmscriptenFunction(EHTypeIDTy, "llvm_eh_typeid_for", &M);
671 
672     for (Function &F : M) {
673       if (F.isDeclaration())
674         continue;
675       Changed |= runEHOnFunction(F);
676     }
677   }
678 
679   // Setjmp/longjmp handling
680   if (DoSjLj) {
681     Changed = true; // We have setjmp or longjmp somewhere
682 
683     if (LongjmpF) {
684       // Replace all uses of longjmp with emscripten_longjmp_jmpbuf, which is
685       // defined in JS code
686       EmLongjmpJmpbufF = getEmscriptenFunction(LongjmpF->getFunctionType(),
687                                                "emscripten_longjmp_jmpbuf", &M);
688       LongjmpF->replaceAllUsesWith(EmLongjmpJmpbufF);
689     }
690 
691     if (SetjmpF) {
692       // Register saveSetjmp function
693       FunctionType *SetjmpFTy = SetjmpF->getFunctionType();
694       FunctionType *FTy =
695           FunctionType::get(Type::getInt32PtrTy(C),
696                             {SetjmpFTy->getParamType(0), IRB.getInt32Ty(),
697                              Type::getInt32PtrTy(C), IRB.getInt32Ty()},
698                             false);
699       SaveSetjmpF = getEmscriptenFunction(FTy, "saveSetjmp", &M);
700 
701       // Register testSetjmp function
702       FTy = FunctionType::get(
703           IRB.getInt32Ty(),
704           {IRB.getInt32Ty(), Type::getInt32PtrTy(C), IRB.getInt32Ty()}, false);
705       TestSetjmpF = getEmscriptenFunction(FTy, "testSetjmp", &M);
706 
707       FTy = FunctionType::get(IRB.getVoidTy(),
708                               {IRB.getInt32Ty(), IRB.getInt32Ty()}, false);
709       EmLongjmpF = getEmscriptenFunction(FTy, "emscripten_longjmp", &M);
710 
711       // Only traverse functions that uses setjmp in order not to insert
712       // unnecessary prep / cleanup code in every function
713       SmallPtrSet<Function *, 8> SetjmpUsers;
714       for (User *U : SetjmpF->users()) {
715         auto *UI = cast<Instruction>(U);
716         SetjmpUsers.insert(UI->getFunction());
717       }
718       for (Function *F : SetjmpUsers)
719         runSjLjOnFunction(*F);
720     }
721   }
722 
723   if (!Changed) {
724     // Delete unused global variables and functions
725     if (ResumeF)
726       ResumeF->eraseFromParent();
727     if (EHTypeIDF)
728       EHTypeIDF->eraseFromParent();
729     if (EmLongjmpF)
730       EmLongjmpF->eraseFromParent();
731     if (SaveSetjmpF)
732       SaveSetjmpF->eraseFromParent();
733     if (TestSetjmpF)
734       TestSetjmpF->eraseFromParent();
735     return false;
736   }
737 
738   return true;
739 }
740 
741 bool WebAssemblyLowerEmscriptenEHSjLj::runEHOnFunction(Function &F) {
742   Module &M = *F.getParent();
743   LLVMContext &C = F.getContext();
744   IRBuilder<> IRB(C);
745   bool Changed = false;
746   SmallVector<Instruction *, 64> ToErase;
747   SmallPtrSet<LandingPadInst *, 32> LandingPads;
748   bool AllowExceptions = areAllExceptionsAllowed() ||
749                          EHAllowlistSet.count(std::string(F.getName()));
750 
751   for (BasicBlock &BB : F) {
752     auto *II = dyn_cast<InvokeInst>(BB.getTerminator());
753     if (!II)
754       continue;
755     Changed = true;
756     LandingPads.insert(II->getLandingPadInst());
757     IRB.SetInsertPoint(II);
758 
759     bool NeedInvoke = AllowExceptions && canThrow(II->getCalledOperand());
760     if (NeedInvoke) {
761       // Wrap invoke with invoke wrapper and generate preamble/postamble
762       Value *Threw = wrapInvoke(II);
763       ToErase.push_back(II);
764 
765       // Insert a branch based on __THREW__ variable
766       Value *Cmp = IRB.CreateICmpEQ(Threw, IRB.getInt32(1), "cmp");
767       IRB.CreateCondBr(Cmp, II->getUnwindDest(), II->getNormalDest());
768 
769     } else {
770       // This can't throw, and we don't need this invoke, just replace it with a
771       // call+branch
772       SmallVector<Value *, 16> Args(II->arg_begin(), II->arg_end());
773       CallInst *NewCall =
774           IRB.CreateCall(II->getFunctionType(), II->getCalledOperand(), Args);
775       NewCall->takeName(II);
776       NewCall->setCallingConv(II->getCallingConv());
777       NewCall->setDebugLoc(II->getDebugLoc());
778       NewCall->setAttributes(II->getAttributes());
779       II->replaceAllUsesWith(NewCall);
780       ToErase.push_back(II);
781 
782       IRB.CreateBr(II->getNormalDest());
783 
784       // Remove any PHI node entries from the exception destination
785       II->getUnwindDest()->removePredecessor(&BB);
786     }
787   }
788 
789   // Process resume instructions
790   for (BasicBlock &BB : F) {
791     // Scan the body of the basic block for resumes
792     for (Instruction &I : BB) {
793       auto *RI = dyn_cast<ResumeInst>(&I);
794       if (!RI)
795         continue;
796       Changed = true;
797 
798       // Split the input into legal values
799       Value *Input = RI->getValue();
800       IRB.SetInsertPoint(RI);
801       Value *Low = IRB.CreateExtractValue(Input, 0, "low");
802       // Create a call to __resumeException function
803       IRB.CreateCall(ResumeF, {Low});
804       // Add a terminator to the block
805       IRB.CreateUnreachable();
806       ToErase.push_back(RI);
807     }
808   }
809 
810   // Process llvm.eh.typeid.for intrinsics
811   for (BasicBlock &BB : F) {
812     for (Instruction &I : BB) {
813       auto *CI = dyn_cast<CallInst>(&I);
814       if (!CI)
815         continue;
816       const Function *Callee = CI->getCalledFunction();
817       if (!Callee)
818         continue;
819       if (Callee->getIntrinsicID() != Intrinsic::eh_typeid_for)
820         continue;
821       Changed = true;
822 
823       IRB.SetInsertPoint(CI);
824       CallInst *NewCI =
825           IRB.CreateCall(EHTypeIDF, CI->getArgOperand(0), "typeid");
826       CI->replaceAllUsesWith(NewCI);
827       ToErase.push_back(CI);
828     }
829   }
830 
831   // Look for orphan landingpads, can occur in blocks with no predecessors
832   for (BasicBlock &BB : F) {
833     Instruction *I = BB.getFirstNonPHI();
834     if (auto *LPI = dyn_cast<LandingPadInst>(I))
835       LandingPads.insert(LPI);
836   }
837   Changed |= !LandingPads.empty();
838 
839   // Handle all the landingpad for this function together, as multiple invokes
840   // may share a single lp
841   for (LandingPadInst *LPI : LandingPads) {
842     IRB.SetInsertPoint(LPI);
843     SmallVector<Value *, 16> FMCArgs;
844     for (unsigned I = 0, E = LPI->getNumClauses(); I < E; ++I) {
845       Constant *Clause = LPI->getClause(I);
846       // As a temporary workaround for the lack of aggregate varargs support
847       // in the interface between JS and wasm, break out filter operands into
848       // their component elements.
849       if (LPI->isFilter(I)) {
850         auto *ATy = cast<ArrayType>(Clause->getType());
851         for (unsigned J = 0, E = ATy->getNumElements(); J < E; ++J) {
852           Value *EV = IRB.CreateExtractValue(Clause, makeArrayRef(J), "filter");
853           FMCArgs.push_back(EV);
854         }
855       } else
856         FMCArgs.push_back(Clause);
857     }
858 
859     // Create a call to __cxa_find_matching_catch_N function
860     Function *FMCF = getFindMatchingCatch(M, FMCArgs.size());
861     CallInst *FMCI = IRB.CreateCall(FMCF, FMCArgs, "fmc");
862     Value *Undef = UndefValue::get(LPI->getType());
863     Value *Pair0 = IRB.CreateInsertValue(Undef, FMCI, 0, "pair0");
864     Value *TempRet0 = IRB.CreateCall(GetTempRet0Func, None, "tempret0");
865     Value *Pair1 = IRB.CreateInsertValue(Pair0, TempRet0, 1, "pair1");
866 
867     LPI->replaceAllUsesWith(Pair1);
868     ToErase.push_back(LPI);
869   }
870 
871   // Erase everything we no longer need in this function
872   for (Instruction *I : ToErase)
873     I->eraseFromParent();
874 
875   return Changed;
876 }
877 
878 // This tries to get debug info from the instruction before which a new
879 // instruction will be inserted, and if there's no debug info in that
880 // instruction, tries to get the info instead from the previous instruction (if
881 // any). If none of these has debug info and a DISubprogram is provided, it
882 // creates a dummy debug info with the first line of the function, because IR
883 // verifier requires all inlinable callsites should have debug info when both a
884 // caller and callee have DISubprogram. If none of these conditions are met,
885 // returns empty info.
886 static DebugLoc getOrCreateDebugLoc(const Instruction *InsertBefore,
887                                     DISubprogram *SP) {
888   assert(InsertBefore);
889   if (InsertBefore->getDebugLoc())
890     return InsertBefore->getDebugLoc();
891   const Instruction *Prev = InsertBefore->getPrevNode();
892   if (Prev && Prev->getDebugLoc())
893     return Prev->getDebugLoc();
894   if (SP)
895     return DILocation::get(SP->getContext(), SP->getLine(), 1, SP);
896   return DebugLoc();
897 }
898 
899 bool WebAssemblyLowerEmscriptenEHSjLj::runSjLjOnFunction(Function &F) {
900   Module &M = *F.getParent();
901   LLVMContext &C = F.getContext();
902   IRBuilder<> IRB(C);
903   SmallVector<Instruction *, 64> ToErase;
904   // Vector of %setjmpTable values
905   std::vector<Instruction *> SetjmpTableInsts;
906   // Vector of %setjmpTableSize values
907   std::vector<Instruction *> SetjmpTableSizeInsts;
908 
909   // Setjmp preparation
910 
911   // This instruction effectively means %setjmpTableSize = 4.
912   // We create this as an instruction intentionally, and we don't want to fold
913   // this instruction to a constant 4, because this value will be used in
914   // SSAUpdater.AddAvailableValue(...) later.
915   BasicBlock &EntryBB = F.getEntryBlock();
916   DebugLoc FirstDL = getOrCreateDebugLoc(&*EntryBB.begin(), F.getSubprogram());
917   BinaryOperator *SetjmpTableSize = BinaryOperator::Create(
918       Instruction::Add, IRB.getInt32(4), IRB.getInt32(0), "setjmpTableSize",
919       &*EntryBB.getFirstInsertionPt());
920   SetjmpTableSize->setDebugLoc(FirstDL);
921   // setjmpTable = (int *) malloc(40);
922   Instruction *SetjmpTable = CallInst::CreateMalloc(
923       SetjmpTableSize, IRB.getInt32Ty(), IRB.getInt32Ty(), IRB.getInt32(40),
924       nullptr, nullptr, "setjmpTable");
925   SetjmpTable->setDebugLoc(FirstDL);
926   // CallInst::CreateMalloc may return a bitcast instruction if the result types
927   // mismatch. We need to set the debug loc for the original call too.
928   auto *MallocCall = SetjmpTable->stripPointerCasts();
929   if (auto *MallocCallI = dyn_cast<Instruction>(MallocCall)) {
930     MallocCallI->setDebugLoc(FirstDL);
931   }
932   // setjmpTable[0] = 0;
933   IRB.SetInsertPoint(SetjmpTableSize);
934   IRB.CreateStore(IRB.getInt32(0), SetjmpTable);
935   SetjmpTableInsts.push_back(SetjmpTable);
936   SetjmpTableSizeInsts.push_back(SetjmpTableSize);
937 
938   // Setjmp transformation
939   std::vector<PHINode *> SetjmpRetPHIs;
940   Function *SetjmpF = M.getFunction("setjmp");
941   for (User *U : SetjmpF->users()) {
942     auto *CI = dyn_cast<CallInst>(U);
943     if (!CI)
944       report_fatal_error("Does not support indirect calls to setjmp");
945 
946     BasicBlock *BB = CI->getParent();
947     if (BB->getParent() != &F) // in other function
948       continue;
949 
950     // The tail is everything right after the call, and will be reached once
951     // when setjmp is called, and later when longjmp returns to the setjmp
952     BasicBlock *Tail = SplitBlock(BB, CI->getNextNode());
953     // Add a phi to the tail, which will be the output of setjmp, which
954     // indicates if this is the first call or a longjmp back. The phi directly
955     // uses the right value based on where we arrive from
956     IRB.SetInsertPoint(Tail->getFirstNonPHI());
957     PHINode *SetjmpRet = IRB.CreatePHI(IRB.getInt32Ty(), 2, "setjmp.ret");
958 
959     // setjmp initial call returns 0
960     SetjmpRet->addIncoming(IRB.getInt32(0), BB);
961     // The proper output is now this, not the setjmp call itself
962     CI->replaceAllUsesWith(SetjmpRet);
963     // longjmp returns to the setjmp will add themselves to this phi
964     SetjmpRetPHIs.push_back(SetjmpRet);
965 
966     // Fix call target
967     // Our index in the function is our place in the array + 1 to avoid index
968     // 0, because index 0 means the longjmp is not ours to handle.
969     IRB.SetInsertPoint(CI);
970     Value *Args[] = {CI->getArgOperand(0), IRB.getInt32(SetjmpRetPHIs.size()),
971                      SetjmpTable, SetjmpTableSize};
972     Instruction *NewSetjmpTable =
973         IRB.CreateCall(SaveSetjmpF, Args, "setjmpTable");
974     Instruction *NewSetjmpTableSize =
975         IRB.CreateCall(GetTempRet0Func, None, "setjmpTableSize");
976     SetjmpTableInsts.push_back(NewSetjmpTable);
977     SetjmpTableSizeInsts.push_back(NewSetjmpTableSize);
978     ToErase.push_back(CI);
979   }
980 
981   // Update each call that can longjmp so it can return to a setjmp where
982   // relevant.
983 
984   // Because we are creating new BBs while processing and don't want to make
985   // all these newly created BBs candidates again for longjmp processing, we
986   // first make the vector of candidate BBs.
987   std::vector<BasicBlock *> BBs;
988   for (BasicBlock &BB : F)
989     BBs.push_back(&BB);
990 
991   // BBs.size() will change within the loop, so we query it every time
992   for (unsigned I = 0; I < BBs.size(); I++) {
993     BasicBlock *BB = BBs[I];
994     for (Instruction &I : *BB) {
995       assert(!isa<InvokeInst>(&I));
996       auto *CI = dyn_cast<CallInst>(&I);
997       if (!CI)
998         continue;
999 
1000       const Value *Callee = CI->getCalledOperand();
1001       if (!canLongjmp(M, Callee))
1002         continue;
1003       if (isEmAsmCall(M, Callee))
1004         report_fatal_error("Cannot use EM_ASM* alongside setjmp/longjmp in " +
1005                                F.getName() +
1006                                ". Please consider using EM_JS, or move the "
1007                                "EM_ASM into another function.",
1008                            false);
1009 
1010       Value *Threw = nullptr;
1011       BasicBlock *Tail;
1012       if (Callee->getName().startswith("__invoke_")) {
1013         // If invoke wrapper has already been generated for this call in
1014         // previous EH phase, search for the load instruction
1015         // %__THREW__.val = __THREW__;
1016         // in postamble after the invoke wrapper call
1017         LoadInst *ThrewLI = nullptr;
1018         StoreInst *ThrewResetSI = nullptr;
1019         for (auto I = std::next(BasicBlock::iterator(CI)), IE = BB->end();
1020              I != IE; ++I) {
1021           if (auto *LI = dyn_cast<LoadInst>(I))
1022             if (auto *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand()))
1023               if (GV == ThrewGV) {
1024                 Threw = ThrewLI = LI;
1025                 break;
1026               }
1027         }
1028         // Search for the store instruction after the load above
1029         // __THREW__ = 0;
1030         for (auto I = std::next(BasicBlock::iterator(ThrewLI)), IE = BB->end();
1031              I != IE; ++I) {
1032           if (auto *SI = dyn_cast<StoreInst>(I))
1033             if (auto *GV = dyn_cast<GlobalVariable>(SI->getPointerOperand()))
1034               if (GV == ThrewGV && SI->getValueOperand() == IRB.getInt32(0)) {
1035                 ThrewResetSI = SI;
1036                 break;
1037               }
1038         }
1039         assert(Threw && ThrewLI && "Cannot find __THREW__ load after invoke");
1040         assert(ThrewResetSI && "Cannot find __THREW__ store after invoke");
1041         Tail = SplitBlock(BB, ThrewResetSI->getNextNode());
1042 
1043       } else {
1044         // Wrap call with invoke wrapper and generate preamble/postamble
1045         Threw = wrapInvoke(CI);
1046         ToErase.push_back(CI);
1047         Tail = SplitBlock(BB, CI->getNextNode());
1048       }
1049 
1050       // We need to replace the terminator in Tail - SplitBlock makes BB go
1051       // straight to Tail, we need to check if a longjmp occurred, and go to the
1052       // right setjmp-tail if so
1053       ToErase.push_back(BB->getTerminator());
1054 
1055       // Generate a function call to testSetjmp function and preamble/postamble
1056       // code to figure out (1) whether longjmp occurred (2) if longjmp
1057       // occurred, which setjmp it corresponds to
1058       Value *Label = nullptr;
1059       Value *LongjmpResult = nullptr;
1060       BasicBlock *EndBB = nullptr;
1061       wrapTestSetjmp(BB, CI->getDebugLoc(), Threw, SetjmpTable, SetjmpTableSize,
1062                      Label, LongjmpResult, EndBB);
1063       assert(Label && LongjmpResult && EndBB);
1064 
1065       // Create switch instruction
1066       IRB.SetInsertPoint(EndBB);
1067       IRB.SetCurrentDebugLocation(EndBB->getInstList().back().getDebugLoc());
1068       SwitchInst *SI = IRB.CreateSwitch(Label, Tail, SetjmpRetPHIs.size());
1069       // -1 means no longjmp happened, continue normally (will hit the default
1070       // switch case). 0 means a longjmp that is not ours to handle, needs a
1071       // rethrow. Otherwise the index is the same as the index in P+1 (to avoid
1072       // 0).
1073       for (unsigned I = 0; I < SetjmpRetPHIs.size(); I++) {
1074         SI->addCase(IRB.getInt32(I + 1), SetjmpRetPHIs[I]->getParent());
1075         SetjmpRetPHIs[I]->addIncoming(LongjmpResult, EndBB);
1076       }
1077 
1078       // We are splitting the block here, and must continue to find other calls
1079       // in the block - which is now split. so continue to traverse in the Tail
1080       BBs.push_back(Tail);
1081     }
1082   }
1083 
1084   // Erase everything we no longer need in this function
1085   for (Instruction *I : ToErase)
1086     I->eraseFromParent();
1087 
1088   // Free setjmpTable buffer before each return instruction
1089   for (BasicBlock &BB : F) {
1090     Instruction *TI = BB.getTerminator();
1091     if (isa<ReturnInst>(TI)) {
1092       DebugLoc DL = getOrCreateDebugLoc(TI, F.getSubprogram());
1093       auto *Free = CallInst::CreateFree(SetjmpTable, TI);
1094       Free->setDebugLoc(DL);
1095       // CallInst::CreateFree may create a bitcast instruction if its argument
1096       // types mismatch. We need to set the debug loc for the bitcast too.
1097       if (auto *FreeCallI = dyn_cast<CallInst>(Free)) {
1098         if (auto *BitCastI = dyn_cast<BitCastInst>(FreeCallI->getArgOperand(0)))
1099           BitCastI->setDebugLoc(DL);
1100       }
1101     }
1102   }
1103 
1104   // Every call to saveSetjmp can change setjmpTable and setjmpTableSize
1105   // (when buffer reallocation occurs)
1106   // entry:
1107   //   setjmpTableSize = 4;
1108   //   setjmpTable = (int *) malloc(40);
1109   //   setjmpTable[0] = 0;
1110   // ...
1111   // somebb:
1112   //   setjmpTable = saveSetjmp(buf, label, setjmpTable, setjmpTableSize);
1113   //   setjmpTableSize = getTempRet0();
1114   // So we need to make sure the SSA for these variables is valid so that every
1115   // saveSetjmp and testSetjmp calls have the correct arguments.
1116   SSAUpdater SetjmpTableSSA;
1117   SSAUpdater SetjmpTableSizeSSA;
1118   SetjmpTableSSA.Initialize(Type::getInt32PtrTy(C), "setjmpTable");
1119   SetjmpTableSizeSSA.Initialize(Type::getInt32Ty(C), "setjmpTableSize");
1120   for (Instruction *I : SetjmpTableInsts)
1121     SetjmpTableSSA.AddAvailableValue(I->getParent(), I);
1122   for (Instruction *I : SetjmpTableSizeInsts)
1123     SetjmpTableSizeSSA.AddAvailableValue(I->getParent(), I);
1124 
1125   for (auto UI = SetjmpTable->use_begin(), UE = SetjmpTable->use_end();
1126        UI != UE;) {
1127     // Grab the use before incrementing the iterator.
1128     Use &U = *UI;
1129     // Increment the iterator before removing the use from the list.
1130     ++UI;
1131     if (auto *I = dyn_cast<Instruction>(U.getUser()))
1132       if (I->getParent() != &EntryBB)
1133         SetjmpTableSSA.RewriteUse(U);
1134   }
1135   for (auto UI = SetjmpTableSize->use_begin(), UE = SetjmpTableSize->use_end();
1136        UI != UE;) {
1137     Use &U = *UI;
1138     ++UI;
1139     if (auto *I = dyn_cast<Instruction>(U.getUser()))
1140       if (I->getParent() != &EntryBB)
1141         SetjmpTableSizeSSA.RewriteUse(U);
1142   }
1143 
1144   // Finally, our modifications to the cfg can break dominance of SSA variables.
1145   // For example, in this code,
1146   // if (x()) { .. setjmp() .. }
1147   // if (y()) { .. longjmp() .. }
1148   // We must split the longjmp block, and it can jump into the block splitted
1149   // from setjmp one. But that means that when we split the setjmp block, it's
1150   // first part no longer dominates its second part - there is a theoretically
1151   // possible control flow path where x() is false, then y() is true and we
1152   // reach the second part of the setjmp block, without ever reaching the first
1153   // part. So, we rebuild SSA form here.
1154   rebuildSSA(F);
1155   return true;
1156 }
1157