1 //=== WebAssemblyLowerEmscriptenEHSjLj.cpp - Lower exceptions for Emscripten =//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 ///
10 /// \file
11 /// \brief This file lowers exception-related instructions and setjmp/longjmp
12 /// function calls in order to use Emscripten's JavaScript try and catch
13 /// mechanism.
14 ///
15 /// To handle exceptions and setjmp/longjmps, this scheme relies on JavaScript's
16 /// try and catch syntax and relevant exception-related libraries implemented
17 /// in JavaScript glue code that will be produced by Emscripten. This is similar
18 /// to the current Emscripten asm.js exception handling in fastcomp. For
19 /// fastcomp's EH / SjLj scheme, see these files in fastcomp LLVM branch:
20 /// (Location: https://github.com/kripken/emscripten-fastcomp)
21 /// lib/Target/JSBackend/NaCl/LowerEmExceptionsPass.cpp
22 /// lib/Target/JSBackend/NaCl/LowerEmSetjmp.cpp
23 /// lib/Target/JSBackend/JSBackend.cpp
24 /// lib/Target/JSBackend/CallHandlers.h
25 ///
26 /// * Exception handling
27 /// This pass lowers invokes and landingpads into library functions in JS glue
28 /// code. Invokes are lowered into function wrappers called invoke wrappers that
29 /// exist in JS side, which wraps the original function call with JS try-catch.
30 /// If an exception occurred, cxa_throw() function in JS side sets some
31 /// variables (see below) so we can check whether an exception occurred from
32 /// wasm code and handle it appropriately.
33 ///
34 /// * Setjmp-longjmp handling
35 /// This pass lowers setjmp to a reasonably-performant approach for emscripten.
36 /// The idea is that each block with a setjmp is broken up into two parts: the
37 /// part containing setjmp and the part right after the setjmp. The latter part
38 /// is either reached from the setjmp, or later from a longjmp. To handle the
39 /// longjmp, all calls that might longjmp are also called using invoke wrappers
40 /// and thus JS / try-catch. JS longjmp() function also sets some variables so
41 /// we can check / whether a longjmp occurred from wasm code. Each block with a
42 /// function call that might longjmp is also split up after the longjmp call.
43 /// After the longjmp call, we check whether a longjmp occurred, and if it did,
44 /// which setjmp it corresponds to, and jump to the right post-setjmp block.
45 /// We assume setjmp-longjmp handling always run after EH handling, which means
46 /// we don't expect any exception-related instructions when SjLj runs.
47 /// FIXME Currently this scheme does not support indirect call of setjmp,
48 /// because of the limitation of the scheme itself. fastcomp does not support it
49 /// either.
50 ///
51 /// In detail, this pass does following things:
52 ///
53 /// 1) Create three global variables: __THREW__, __threwValue, and __tempRet0.
54 ///    __tempRet0 will be set within __cxa_find_matching_catch() function in
55 ///    JS library, and __THREW__ and __threwValue will be set in invoke wrappers
56 ///    in JS glue code. For what invoke wrappers are, refer to 3). These
57 ///    variables are used for both exceptions and setjmp/longjmps.
58 ///    __THREW__ indicates whether an exception or a longjmp occurred or not. 0
59 ///    means nothing occurred, 1 means an exception occurred, and other numbers
60 ///    mean a longjmp occurred. In the case of longjmp, __threwValue variable
61 ///    indicates the corresponding setjmp buffer the longjmp corresponds to.
62 ///    In exception handling, __tempRet0 indicates the type of an exception
63 ///    caught, and in setjmp/longjmp, it means the second argument to longjmp
64 ///    function.
65 ///
66 /// * Exception handling
67 ///
68 /// 2) Create setThrew and setTempRet0 functions.
69 ///    The global variables created in 1) will exist in wasm address space,
70 ///    but their values should be set in JS code, so we provide these functions
71 ///    as interfaces to JS glue code. These functions are equivalent to the
72 ///    following JS functions, which actually exist in asm.js version of JS
73 ///    library.
74 ///
75 ///    function setThrew(threw, value) {
76 ///      if (__THREW__ == 0) {
77 ///        __THREW__ = threw;
78 ///        __threwValue = value;
79 ///      }
80 ///    }
81 ///
82 ///    function setTempRet0(value) {
83 ///      __tempRet0 = value;
84 ///    }
85 ///
86 /// 3) Lower
87 ///      invoke @func(arg1, arg2) to label %invoke.cont unwind label %lpad
88 ///    into
89 ///      __THREW__ = 0;
90 ///      call @__invoke_SIG(func, arg1, arg2)
91 ///      %__THREW__.val = __THREW__;
92 ///      __THREW__ = 0;
93 ///      if (%__THREW__.val == 1)
94 ///        goto %lpad
95 ///      else
96 ///         goto %invoke.cont
97 ///    SIG is a mangled string generated based on the LLVM IR-level function
98 ///    signature. After LLVM IR types are lowered to the target wasm types,
99 ///    the names for these wrappers will change based on wasm types as well,
100 ///    as in invoke_vi (function takes an int and returns void). The bodies of
101 ///    these wrappers will be generated in JS glue code, and inside those
102 ///    wrappers we use JS try-catch to generate actual exception effects. It
103 ///    also calls the original callee function. An example wrapper in JS code
104 ///    would look like this:
105 ///      function invoke_vi(index,a1) {
106 ///        try {
107 ///          Module["dynCall_vi"](index,a1); // This calls original callee
108 ///        } catch(e) {
109 ///          if (typeof e !== 'number' && e !== 'longjmp') throw e;
110 ///          asm["setThrew"](1, 0); // setThrew is called here
111 ///        }
112 ///      }
113 ///    If an exception is thrown, __THREW__ will be set to true in a wrapper,
114 ///    so we can jump to the right BB based on this value.
115 ///
116 /// 4) Lower
117 ///      %val = landingpad catch c1 catch c2 catch c3 ...
118 ///      ... use %val ...
119 ///    into
120 ///      %fmc = call @__cxa_find_matching_catch_N(c1, c2, c3, ...)
121 ///      %val = {%fmc, __tempRet0}
122 ///      ... use %val ...
123 ///    Here N is a number calculated based on the number of clauses.
124 ///    Global variable __tempRet0 is set within __cxa_find_matching_catch() in
125 ///    JS glue code.
126 ///
127 /// 5) Lower
128 ///      resume {%a, %b}
129 ///    into
130 ///      call @__resumeException(%a)
131 ///    where __resumeException() is a function in JS glue code.
132 ///
133 /// 6) Lower
134 ///      call @llvm.eh.typeid.for(type) (intrinsic)
135 ///    into
136 ///      call @llvm_eh_typeid_for(type)
137 ///    llvm_eh_typeid_for function will be generated in JS glue code.
138 ///
139 /// * Setjmp / Longjmp handling
140 ///
141 /// 7) In the function entry that calls setjmp, initialize setjmpTable and
142 ///    sejmpTableSize as follows:
143 ///      setjmpTableSize = 4;
144 ///      setjmpTable = (int *) malloc(40);
145 ///      setjmpTable[0] = 0;
146 ///    setjmpTable and setjmpTableSize are used in saveSetjmp() function in JS
147 ///    code.
148 ///
149 /// 8) Lower
150 ///      setjmp(buf)
151 ///    into
152 ///      setjmpTable = saveSetjmp(buf, label, setjmpTable, setjmpTableSize);
153 ///      setjmpTableSize = __tempRet0;
154 ///    For each dynamic setjmp call, setjmpTable stores its ID (a number which
155 ///    is incrementally assigned from 0) and its label (a unique number that
156 ///    represents each callsite of setjmp). When we need more entries in
157 ///    setjmpTable, it is reallocated in saveSetjmp() in JS code and it will
158 ///    return the new table address, and assign the new table size in
159 ///    __tempRet0. saveSetjmp also stores the setjmp's ID into the buffer buf.
160 ///    A BB with setjmp is split into two after setjmp call in order to make the
161 ///    post-setjmp BB the possible destination of longjmp BB.
162 ///
163 /// 9) Lower
164 ///      longjmp(buf, value)
165 ///    into
166 ///      emscripten_longjmp_jmpbuf(buf, value)
167 ///    emscripten_longjmp_jmpbuf will be lowered to emscripten_longjmp later.
168 ///
169 /// 10) Lower every call that might longjmp into
170 ///      __THREW__ = 0;
171 ///      call @__invoke_SIG(func, arg1, arg2)
172 ///      %__THREW__.val = __THREW__;
173 ///      __THREW__ = 0;
174 ///      if (%__THREW__.val != 0 & __threwValue != 0) {
175 ///        %label = testSetjmp(mem[%__THREW__.val], setjmpTable,
176 ///                            setjmpTableSize);
177 ///        if (%label == 0)
178 ///          emscripten_longjmp(%__THREW__.val, __threwValue);
179 ///        __tempRet0 = __threwValue;
180 ///      } else {
181 ///        %label = -1;
182 ///      }
183 ///      longjmp_result = __tempRet0;
184 ///      switch label {
185 ///        label 1: goto post-setjmp BB 1
186 ///        label 2: goto post-setjmp BB 2
187 ///        ...
188 ///        default: goto splitted next BB
189 ///      }
190 ///     testSetjmp examines setjmpTable to see if there is a matching setjmp
191 ///     call. After calling an invoke wrapper, if a longjmp occurred, __THREW__
192 ///     will be the address of matching jmp_buf buffer and __threwValue be the
193 ///     second argument to longjmp. mem[__THREW__.val] is a setjmp ID that is
194 ///     stored in saveSetjmp. testSetjmp returns a setjmp label, a unique ID to
195 ///     each setjmp callsite. Label 0 means this longjmp buffer does not
196 ///     correspond to one of the setjmp callsites in this function, so in this
197 ///     case we just chain the longjmp to the caller. (Here we call
198 ///     emscripten_longjmp, which is different from emscripten_longjmp_jmpbuf.
199 ///     emscripten_longjmp_jmpbuf takes jmp_buf as its first argument, while
200 ///     emscripten_longjmp takes an int. Both of them will eventually be lowered
201 ///     to emscripten_longjmp in s2wasm, but here we need two signatures - we
202 ///     can't translate an int value to a jmp_buf.)
203 ///     Label -1 means no longjmp occurred. Otherwise we jump to the right
204 ///     post-setjmp BB based on the label.
205 ///
206 ///===----------------------------------------------------------------------===//
207 
208 #include "WebAssembly.h"
209 #include "llvm/IR/CallSite.h"
210 #include "llvm/IR/Dominators.h"
211 #include "llvm/IR/IRBuilder.h"
212 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
213 #include "llvm/Transforms/Utils/SSAUpdater.h"
214 
215 using namespace llvm;
216 
217 #define DEBUG_TYPE "wasm-lower-em-ehsjlj"
218 
219 static cl::list<std::string>
220     EHWhitelist("emscripten-cxx-exceptions-whitelist",
221                 cl::desc("The list of function names in which Emscripten-style "
222                          "exception handling is enabled (see emscripten "
223                          "EMSCRIPTEN_CATCHING_WHITELIST options)"),
224                 cl::CommaSeparated);
225 
226 namespace {
227 class WebAssemblyLowerEmscriptenEHSjLj final : public ModulePass {
228   static const char *ThrewGVName;
229   static const char *ThrewValueGVName;
230   static const char *TempRet0GVName;
231   static const char *ResumeFName;
232   static const char *EHTypeIDFName;
233   static const char *SetThrewFName;
234   static const char *SetTempRet0FName;
235   static const char *EmLongjmpFName;
236   static const char *EmLongjmpJmpbufFName;
237   static const char *SaveSetjmpFName;
238   static const char *TestSetjmpFName;
239   static const char *FindMatchingCatchPrefix;
240   static const char *InvokePrefix;
241 
242   bool EnableEH;   // Enable exception handling
243   bool EnableSjLj; // Enable setjmp/longjmp handling
244 
245   GlobalVariable *ThrewGV;
246   GlobalVariable *ThrewValueGV;
247   GlobalVariable *TempRet0GV;
248   Function *ResumeF;
249   Function *EHTypeIDF;
250   Function *EmLongjmpF;
251   Function *EmLongjmpJmpbufF;
252   Function *SaveSetjmpF;
253   Function *TestSetjmpF;
254 
255   // __cxa_find_matching_catch_N functions.
256   // Indexed by the number of clauses in an original landingpad instruction.
257   DenseMap<int, Function *> FindMatchingCatches;
258   // Map of <function signature string, invoke_ wrappers>
259   StringMap<Function *> InvokeWrappers;
260   // Set of whitelisted function names for exception handling
261   std::set<std::string> EHWhitelistSet;
262 
263   StringRef getPassName() const override {
264     return "WebAssembly Lower Emscripten Exceptions";
265   }
266 
267   bool runEHOnFunction(Function &F);
268   bool runSjLjOnFunction(Function &F);
269   Function *getFindMatchingCatch(Module &M, unsigned NumClauses);
270 
271   template <typename CallOrInvoke> Value *wrapInvoke(CallOrInvoke *CI);
272   void wrapTestSetjmp(BasicBlock *BB, Instruction *InsertPt, Value *Threw,
273                       Value *SetjmpTable, Value *SetjmpTableSize, Value *&Label,
274                       Value *&LongjmpResult, BasicBlock *&EndBB);
275   template <typename CallOrInvoke> Function *getInvokeWrapper(CallOrInvoke *CI);
276 
277   bool areAllExceptionsAllowed() const { return EHWhitelistSet.empty(); }
278   bool canLongjmp(Module &M, const Value *Callee) const;
279 
280   void createSetThrewFunction(Module &M);
281   void createSetTempRet0Function(Module &M);
282 
283   void rebuildSSA(Function &F);
284 
285 public:
286   static char ID;
287 
288   WebAssemblyLowerEmscriptenEHSjLj(bool EnableEH = true, bool EnableSjLj = true)
289       : ModulePass(ID), EnableEH(EnableEH), EnableSjLj(EnableSjLj),
290         ThrewGV(nullptr), ThrewValueGV(nullptr), TempRet0GV(nullptr),
291         ResumeF(nullptr), EHTypeIDF(nullptr), EmLongjmpF(nullptr),
292         EmLongjmpJmpbufF(nullptr), SaveSetjmpF(nullptr), TestSetjmpF(nullptr) {
293     EHWhitelistSet.insert(EHWhitelist.begin(), EHWhitelist.end());
294   }
295   bool runOnModule(Module &M) override;
296 
297   void getAnalysisUsage(AnalysisUsage &AU) const override {
298     AU.addRequired<DominatorTreeWrapperPass>();
299   }
300 };
301 } // End anonymous namespace
302 
303 const char *WebAssemblyLowerEmscriptenEHSjLj::ThrewGVName = "__THREW__";
304 const char *WebAssemblyLowerEmscriptenEHSjLj::ThrewValueGVName = "__threwValue";
305 const char *WebAssemblyLowerEmscriptenEHSjLj::TempRet0GVName = "__tempRet0";
306 const char *WebAssemblyLowerEmscriptenEHSjLj::ResumeFName = "__resumeException";
307 const char *WebAssemblyLowerEmscriptenEHSjLj::EHTypeIDFName =
308     "llvm_eh_typeid_for";
309 const char *WebAssemblyLowerEmscriptenEHSjLj::SetThrewFName = "setThrew";
310 const char *WebAssemblyLowerEmscriptenEHSjLj::SetTempRet0FName = "setTempRet0";
311 const char *WebAssemblyLowerEmscriptenEHSjLj::EmLongjmpFName =
312     "emscripten_longjmp";
313 const char *WebAssemblyLowerEmscriptenEHSjLj::EmLongjmpJmpbufFName =
314     "emscripten_longjmp_jmpbuf";
315 const char *WebAssemblyLowerEmscriptenEHSjLj::SaveSetjmpFName = "saveSetjmp";
316 const char *WebAssemblyLowerEmscriptenEHSjLj::TestSetjmpFName = "testSetjmp";
317 const char *WebAssemblyLowerEmscriptenEHSjLj::FindMatchingCatchPrefix =
318     "__cxa_find_matching_catch_";
319 const char *WebAssemblyLowerEmscriptenEHSjLj::InvokePrefix = "__invoke_";
320 
321 char WebAssemblyLowerEmscriptenEHSjLj::ID = 0;
322 INITIALIZE_PASS(WebAssemblyLowerEmscriptenEHSjLj, DEBUG_TYPE,
323                 "WebAssembly Lower Emscripten Exceptions / Setjmp / Longjmp",
324                 false, false)
325 
326 ModulePass *llvm::createWebAssemblyLowerEmscriptenEHSjLj(bool EnableEH,
327                                                          bool EnableSjLj) {
328   return new WebAssemblyLowerEmscriptenEHSjLj(EnableEH, EnableSjLj);
329 }
330 
331 static bool canThrow(const Value *V) {
332   if (const auto *F = dyn_cast<const Function>(V)) {
333     // Intrinsics cannot throw
334     if (F->isIntrinsic())
335       return false;
336     StringRef Name = F->getName();
337     // leave setjmp and longjmp (mostly) alone, we process them properly later
338     if (Name == "setjmp" || Name == "longjmp")
339       return false;
340     return !F->doesNotThrow();
341   }
342   // not a function, so an indirect call - can throw, we can't tell
343   return true;
344 }
345 
346 // Returns an available name for a global value.
347 // If the proposed name already exists in the module, adds '_' at the end of
348 // the name until the name is available.
349 static inline std::string createGlobalValueName(const Module &M,
350                                                 const std::string &Propose) {
351   std::string Name = Propose;
352   while (M.getNamedGlobal(Name))
353     Name += "_";
354   return Name;
355 }
356 
357 // Simple function name mangler.
358 // This function simply takes LLVM's string representation of parameter types
359 // and concatenate them with '_'. There are non-alphanumeric characters but llc
360 // is ok with it, and we need to postprocess these names after the lowering
361 // phase anyway.
362 static std::string getSignature(FunctionType *FTy) {
363   std::string Sig;
364   raw_string_ostream OS(Sig);
365   OS << *FTy->getReturnType();
366   for (Type *ParamTy : FTy->params())
367     OS << "_" << *ParamTy;
368   if (FTy->isVarArg())
369     OS << "_...";
370   Sig = OS.str();
371   Sig.erase(remove_if(Sig, isspace), Sig.end());
372   // When s2wasm parses .s file, a comma means the end of an argument. So a
373   // mangled function name can contain any character but a comma.
374   std::replace(Sig.begin(), Sig.end(), ',', '.');
375   return Sig;
376 }
377 
378 // Returns __cxa_find_matching_catch_N function, where N = NumClauses + 2.
379 // This is because a landingpad instruction contains two more arguments, a
380 // personality function and a cleanup bit, and __cxa_find_matching_catch_N
381 // functions are named after the number of arguments in the original landingpad
382 // instruction.
383 Function *
384 WebAssemblyLowerEmscriptenEHSjLj::getFindMatchingCatch(Module &M,
385                                                        unsigned NumClauses) {
386   if (FindMatchingCatches.count(NumClauses))
387     return FindMatchingCatches[NumClauses];
388   PointerType *Int8PtrTy = Type::getInt8PtrTy(M.getContext());
389   SmallVector<Type *, 16> Args(NumClauses, Int8PtrTy);
390   FunctionType *FTy = FunctionType::get(Int8PtrTy, Args, false);
391   Function *F =
392       Function::Create(FTy, GlobalValue::ExternalLinkage,
393                        FindMatchingCatchPrefix + Twine(NumClauses + 2), &M);
394   FindMatchingCatches[NumClauses] = F;
395   return F;
396 }
397 
398 // Generate invoke wrapper seqence with preamble and postamble
399 // Preamble:
400 // __THREW__ = 0;
401 // Postamble:
402 // %__THREW__.val = __THREW__; __THREW__ = 0;
403 // Returns %__THREW__.val, which indicates whether an exception is thrown (or
404 // whether longjmp occurred), for future use.
405 template <typename CallOrInvoke>
406 Value *WebAssemblyLowerEmscriptenEHSjLj::wrapInvoke(CallOrInvoke *CI) {
407   LLVMContext &C = CI->getModule()->getContext();
408 
409   // If we are calling a function that is noreturn, we must remove that
410   // attribute. The code we insert here does expect it to return, after we
411   // catch the exception.
412   if (CI->doesNotReturn()) {
413     if (auto *F = dyn_cast<Function>(CI->getCalledValue()))
414       F->removeFnAttr(Attribute::NoReturn);
415     CI->removeAttribute(AttributeList::FunctionIndex, Attribute::NoReturn);
416   }
417 
418   IRBuilder<> IRB(C);
419   IRB.SetInsertPoint(CI);
420 
421   // Pre-invoke
422   // __THREW__ = 0;
423   IRB.CreateStore(IRB.getInt32(0), ThrewGV);
424 
425   // Invoke function wrapper in JavaScript
426   SmallVector<Value *, 16> Args;
427   // Put the pointer to the callee as first argument, so it can be called
428   // within the invoke wrapper later
429   Args.push_back(CI->getCalledValue());
430   Args.append(CI->arg_begin(), CI->arg_end());
431   CallInst *NewCall = IRB.CreateCall(getInvokeWrapper(CI), Args);
432   NewCall->takeName(CI);
433   NewCall->setCallingConv(CI->getCallingConv());
434   NewCall->setDebugLoc(CI->getDebugLoc());
435 
436   // Because we added the pointer to the callee as first argument, all
437   // argument attribute indices have to be incremented by one.
438   SmallVector<AttributeSet, 8> AttributesVec;
439   const AttributeList &InvokeAL = CI->getAttributes();
440 
441   // Add any return attributes.
442   AttributesVec.push_back(InvokeAL.getRetAttributes());
443   // No attributes for the callee pointer.
444   AttributesVec.push_back(AttributeSet());
445   // Copy the argument attributes from the original
446   for (unsigned i = 1, e = CI->getNumArgOperands(); i <= e; ++i) {
447     AttributesVec.push_back(InvokeAL.getParamAttributes(i));
448   }
449 
450   // Add any function attributes.
451   AttributesVec.push_back(InvokeAL.getFnAttributes());
452   // Reconstruct the AttributesList based on the vector we constructed.
453   AttributeList NewCallAL = AttributeList::get(C, AttributesVec);
454   NewCall->setAttributes(NewCallAL);
455 
456   CI->replaceAllUsesWith(NewCall);
457 
458   // Post-invoke
459   // %__THREW__.val = __THREW__; __THREW__ = 0;
460   Value *Threw = IRB.CreateLoad(ThrewGV, ThrewGV->getName() + ".val");
461   IRB.CreateStore(IRB.getInt32(0), ThrewGV);
462   return Threw;
463 }
464 
465 // Get matching invoke wrapper based on callee signature
466 template <typename CallOrInvoke>
467 Function *WebAssemblyLowerEmscriptenEHSjLj::getInvokeWrapper(CallOrInvoke *CI) {
468   Module *M = CI->getModule();
469   SmallVector<Type *, 16> ArgTys;
470   Value *Callee = CI->getCalledValue();
471   FunctionType *CalleeFTy;
472   if (auto *F = dyn_cast<Function>(Callee))
473     CalleeFTy = F->getFunctionType();
474   else {
475     auto *CalleeTy = cast<PointerType>(Callee->getType())->getElementType();
476     CalleeFTy = dyn_cast<FunctionType>(CalleeTy);
477   }
478 
479   std::string Sig = getSignature(CalleeFTy);
480   if (InvokeWrappers.find(Sig) != InvokeWrappers.end())
481     return InvokeWrappers[Sig];
482 
483   // Put the pointer to the callee as first argument
484   ArgTys.push_back(PointerType::getUnqual(CalleeFTy));
485   // Add argument types
486   ArgTys.append(CalleeFTy->param_begin(), CalleeFTy->param_end());
487 
488   FunctionType *FTy = FunctionType::get(CalleeFTy->getReturnType(), ArgTys,
489                                         CalleeFTy->isVarArg());
490   Function *F = Function::Create(FTy, GlobalValue::ExternalLinkage,
491                                  InvokePrefix + Sig, M);
492   InvokeWrappers[Sig] = F;
493   return F;
494 }
495 
496 bool WebAssemblyLowerEmscriptenEHSjLj::canLongjmp(Module &M,
497                                                   const Value *Callee) const {
498   if (auto *CalleeF = dyn_cast<Function>(Callee))
499     if (CalleeF->isIntrinsic())
500       return false;
501 
502   // The reason we include malloc/free here is to exclude the malloc/free
503   // calls generated in setjmp prep / cleanup routines.
504   Function *SetjmpF = M.getFunction("setjmp");
505   Function *MallocF = M.getFunction("malloc");
506   Function *FreeF = M.getFunction("free");
507   if (Callee == SetjmpF || Callee == MallocF || Callee == FreeF)
508     return false;
509 
510   // There are functions in JS glue code
511   if (Callee == ResumeF || Callee == EHTypeIDF || Callee == SaveSetjmpF ||
512       Callee == TestSetjmpF)
513     return false;
514 
515   // __cxa_find_matching_catch_N functions cannot longjmp
516   if (Callee->getName().startswith(FindMatchingCatchPrefix))
517     return false;
518 
519   // Exception-catching related functions
520   Function *BeginCatchF = M.getFunction("__cxa_begin_catch");
521   Function *EndCatchF = M.getFunction("__cxa_end_catch");
522   Function *AllocExceptionF = M.getFunction("__cxa_allocate_exception");
523   Function *ThrowF = M.getFunction("__cxa_throw");
524   Function *TerminateF = M.getFunction("__clang_call_terminate");
525   if (Callee == BeginCatchF || Callee == EndCatchF ||
526       Callee == AllocExceptionF || Callee == ThrowF || Callee == TerminateF)
527     return false;
528 
529   // Otherwise we don't know
530   return true;
531 }
532 
533 // Generate testSetjmp function call seqence with preamble and postamble.
534 // The code this generates is equivalent to the following JavaScript code:
535 // if (%__THREW__.val != 0 & threwValue != 0) {
536 //   %label = _testSetjmp(mem[%__THREW__.val], setjmpTable, setjmpTableSize);
537 //   if (%label == 0)
538 //     emscripten_longjmp(%__THREW__.val, threwValue);
539 //   __tempRet0 = threwValue;
540 // } else {
541 //   %label = -1;
542 // }
543 // %longjmp_result = __tempRet0;
544 //
545 // As output parameters. returns %label, %longjmp_result, and the BB the last
546 // instruction (%longjmp_result = ...) is in.
547 void WebAssemblyLowerEmscriptenEHSjLj::wrapTestSetjmp(
548     BasicBlock *BB, Instruction *InsertPt, Value *Threw, Value *SetjmpTable,
549     Value *SetjmpTableSize, Value *&Label, Value *&LongjmpResult,
550     BasicBlock *&EndBB) {
551   Function *F = BB->getParent();
552   LLVMContext &C = BB->getModule()->getContext();
553   IRBuilder<> IRB(C);
554   IRB.SetInsertPoint(InsertPt);
555 
556   // if (%__THREW__.val != 0 & threwValue != 0)
557   IRB.SetInsertPoint(BB);
558   BasicBlock *ThenBB1 = BasicBlock::Create(C, "if.then1", F);
559   BasicBlock *ElseBB1 = BasicBlock::Create(C, "if.else1", F);
560   BasicBlock *EndBB1 = BasicBlock::Create(C, "if.end", F);
561   Value *ThrewCmp = IRB.CreateICmpNE(Threw, IRB.getInt32(0));
562   Value *ThrewValue =
563       IRB.CreateLoad(ThrewValueGV, ThrewValueGV->getName() + ".val");
564   Value *ThrewValueCmp = IRB.CreateICmpNE(ThrewValue, IRB.getInt32(0));
565   Value *Cmp1 = IRB.CreateAnd(ThrewCmp, ThrewValueCmp, "cmp1");
566   IRB.CreateCondBr(Cmp1, ThenBB1, ElseBB1);
567 
568   // %label = _testSetjmp(mem[%__THREW__.val], _setjmpTable, _setjmpTableSize);
569   // if (%label == 0)
570   IRB.SetInsertPoint(ThenBB1);
571   BasicBlock *ThenBB2 = BasicBlock::Create(C, "if.then2", F);
572   BasicBlock *EndBB2 = BasicBlock::Create(C, "if.end2", F);
573   Value *ThrewInt = IRB.CreateIntToPtr(Threw, Type::getInt32PtrTy(C),
574                                        Threw->getName() + ".i32p");
575   Value *LoadedThrew =
576       IRB.CreateLoad(ThrewInt, ThrewInt->getName() + ".loaded");
577   Value *ThenLabel = IRB.CreateCall(
578       TestSetjmpF, {LoadedThrew, SetjmpTable, SetjmpTableSize}, "label");
579   Value *Cmp2 = IRB.CreateICmpEQ(ThenLabel, IRB.getInt32(0));
580   IRB.CreateCondBr(Cmp2, ThenBB2, EndBB2);
581 
582   // emscripten_longjmp(%__THREW__.val, threwValue);
583   IRB.SetInsertPoint(ThenBB2);
584   IRB.CreateCall(EmLongjmpF, {Threw, ThrewValue});
585   IRB.CreateUnreachable();
586 
587   // __tempRet0 = threwValue;
588   IRB.SetInsertPoint(EndBB2);
589   IRB.CreateStore(ThrewValue, TempRet0GV);
590   IRB.CreateBr(EndBB1);
591 
592   IRB.SetInsertPoint(ElseBB1);
593   IRB.CreateBr(EndBB1);
594 
595   // longjmp_result = __tempRet0;
596   IRB.SetInsertPoint(EndBB1);
597   PHINode *LabelPHI = IRB.CreatePHI(IRB.getInt32Ty(), 2, "label");
598   LabelPHI->addIncoming(ThenLabel, EndBB2);
599 
600   LabelPHI->addIncoming(IRB.getInt32(-1), ElseBB1);
601 
602   // Output parameter assignment
603   Label = LabelPHI;
604   EndBB = EndBB1;
605   LongjmpResult = IRB.CreateLoad(TempRet0GV, "longjmp_result");
606 }
607 
608 // Create setThrew function
609 // function setThrew(threw, value) {
610 //   if (__THREW__ == 0) {
611 //     __THREW__ = threw;
612 //     __threwValue = value;
613 //   }
614 // }
615 void WebAssemblyLowerEmscriptenEHSjLj::createSetThrewFunction(Module &M) {
616   LLVMContext &C = M.getContext();
617   IRBuilder<> IRB(C);
618 
619   assert(!M.getNamedGlobal(SetThrewFName) && "setThrew already exists");
620   Type *Params[] = {IRB.getInt32Ty(), IRB.getInt32Ty()};
621   FunctionType *FTy = FunctionType::get(IRB.getVoidTy(), Params, false);
622   Function *F =
623       Function::Create(FTy, GlobalValue::ExternalLinkage, SetThrewFName, &M);
624   Argument *Arg1 = &*(F->arg_begin());
625   Argument *Arg2 = &*std::next(F->arg_begin());
626   Arg1->setName("threw");
627   Arg2->setName("value");
628   BasicBlock *EntryBB = BasicBlock::Create(C, "entry", F);
629   BasicBlock *ThenBB = BasicBlock::Create(C, "if.then", F);
630   BasicBlock *EndBB = BasicBlock::Create(C, "if.end", F);
631 
632   IRB.SetInsertPoint(EntryBB);
633   Value *Threw = IRB.CreateLoad(ThrewGV, ThrewGV->getName() + ".val");
634   Value *Cmp = IRB.CreateICmpEQ(Threw, IRB.getInt32(0), "cmp");
635   IRB.CreateCondBr(Cmp, ThenBB, EndBB);
636 
637   IRB.SetInsertPoint(ThenBB);
638   IRB.CreateStore(Arg1, ThrewGV);
639   IRB.CreateStore(Arg2, ThrewValueGV);
640   IRB.CreateBr(EndBB);
641 
642   IRB.SetInsertPoint(EndBB);
643   IRB.CreateRetVoid();
644 }
645 
646 // Create setTempRet0 function
647 // function setTempRet0(value) {
648 //   __tempRet0 = value;
649 // }
650 void WebAssemblyLowerEmscriptenEHSjLj::createSetTempRet0Function(Module &M) {
651   LLVMContext &C = M.getContext();
652   IRBuilder<> IRB(C);
653 
654   assert(!M.getNamedGlobal(SetTempRet0FName) && "setTempRet0 already exists");
655   Type *Params[] = {IRB.getInt32Ty()};
656   FunctionType *FTy = FunctionType::get(IRB.getVoidTy(), Params, false);
657   Function *F =
658       Function::Create(FTy, GlobalValue::ExternalLinkage, SetTempRet0FName, &M);
659   F->arg_begin()->setName("value");
660   BasicBlock *EntryBB = BasicBlock::Create(C, "entry", F);
661   IRB.SetInsertPoint(EntryBB);
662   IRB.CreateStore(&*F->arg_begin(), TempRet0GV);
663   IRB.CreateRetVoid();
664 }
665 
666 void WebAssemblyLowerEmscriptenEHSjLj::rebuildSSA(Function &F) {
667   DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
668   DT.recalculate(F); // CFG has been changed
669   SSAUpdater SSA;
670   for (BasicBlock &BB : F) {
671     for (Instruction &I : BB) {
672       for (auto UI = I.use_begin(), UE = I.use_end(); UI != UE;) {
673         Use &U = *UI;
674         ++UI;
675         SSA.Initialize(I.getType(), I.getName());
676         SSA.AddAvailableValue(&BB, &I);
677         Instruction *User = cast<Instruction>(U.getUser());
678         if (User->getParent() == &BB)
679           continue;
680 
681         if (PHINode *UserPN = dyn_cast<PHINode>(User))
682           if (UserPN->getIncomingBlock(U) == &BB)
683             continue;
684 
685         if (DT.dominates(&I, User))
686           continue;
687         SSA.RewriteUseAfterInsertions(U);
688       }
689     }
690   }
691 }
692 
693 bool WebAssemblyLowerEmscriptenEHSjLj::runOnModule(Module &M) {
694   LLVMContext &C = M.getContext();
695   IRBuilder<> IRB(C);
696 
697   Function *SetjmpF = M.getFunction("setjmp");
698   Function *LongjmpF = M.getFunction("longjmp");
699   bool SetjmpUsed = SetjmpF && !SetjmpF->use_empty();
700   bool LongjmpUsed = LongjmpF && !LongjmpF->use_empty();
701   bool DoSjLj = EnableSjLj && (SetjmpUsed || LongjmpUsed);
702 
703   // Create global variables __THREW__, threwValue, and __tempRet0, which are
704   // used in common for both exception handling and setjmp/longjmp handling
705   ThrewGV = new GlobalVariable(M, IRB.getInt32Ty(), false,
706                                GlobalValue::ExternalLinkage, IRB.getInt32(0),
707                                createGlobalValueName(M, ThrewGVName));
708   ThrewValueGV = new GlobalVariable(
709       M, IRB.getInt32Ty(), false, GlobalValue::ExternalLinkage, IRB.getInt32(0),
710       createGlobalValueName(M, ThrewValueGVName));
711   TempRet0GV = new GlobalVariable(M, IRB.getInt32Ty(), false,
712                                   GlobalValue::ExternalLinkage, IRB.getInt32(0),
713                                   createGlobalValueName(M, TempRet0GVName));
714 
715   bool Changed = false;
716 
717   // Exception handling
718   if (EnableEH) {
719     // Register __resumeException function
720     FunctionType *ResumeFTy =
721         FunctionType::get(IRB.getVoidTy(), IRB.getInt8PtrTy(), false);
722     ResumeF = Function::Create(ResumeFTy, GlobalValue::ExternalLinkage,
723                                ResumeFName, &M);
724 
725     // Register llvm_eh_typeid_for function
726     FunctionType *EHTypeIDTy =
727         FunctionType::get(IRB.getInt32Ty(), IRB.getInt8PtrTy(), false);
728     EHTypeIDF = Function::Create(EHTypeIDTy, GlobalValue::ExternalLinkage,
729                                  EHTypeIDFName, &M);
730 
731     for (Function &F : M) {
732       if (F.isDeclaration())
733         continue;
734       Changed |= runEHOnFunction(F);
735     }
736   }
737 
738   // Setjmp/longjmp handling
739   if (DoSjLj) {
740     Changed = true; // We have setjmp or longjmp somewhere
741 
742     Function *MallocF = M.getFunction("malloc");
743     Function *FreeF = M.getFunction("free");
744     if (!MallocF || !FreeF)
745       report_fatal_error(
746           "malloc and free must be linked into the module if setjmp is used");
747 
748     // Register saveSetjmp function
749     FunctionType *SetjmpFTy = SetjmpF->getFunctionType();
750     SmallVector<Type *, 4> Params = {SetjmpFTy->getParamType(0),
751                                      IRB.getInt32Ty(), Type::getInt32PtrTy(C),
752                                      IRB.getInt32Ty()};
753     FunctionType *FTy =
754         FunctionType::get(Type::getInt32PtrTy(C), Params, false);
755     SaveSetjmpF = Function::Create(FTy, GlobalValue::ExternalLinkage,
756                                    SaveSetjmpFName, &M);
757 
758     // Register testSetjmp function
759     Params = {IRB.getInt32Ty(), Type::getInt32PtrTy(C), IRB.getInt32Ty()};
760     FTy = FunctionType::get(IRB.getInt32Ty(), Params, false);
761     TestSetjmpF = Function::Create(FTy, GlobalValue::ExternalLinkage,
762                                    TestSetjmpFName, &M);
763 
764     if (LongjmpF) {
765       // Replace all uses of longjmp with emscripten_longjmp_jmpbuf, which is
766       // defined in JS code
767       EmLongjmpJmpbufF = Function::Create(LongjmpF->getFunctionType(),
768                                           GlobalValue::ExternalLinkage,
769                                           EmLongjmpJmpbufFName, &M);
770 
771       LongjmpF->replaceAllUsesWith(EmLongjmpJmpbufF);
772     }
773     FTy = FunctionType::get(IRB.getVoidTy(),
774                             {IRB.getInt32Ty(), IRB.getInt32Ty()}, false);
775     EmLongjmpF =
776         Function::Create(FTy, GlobalValue::ExternalLinkage, EmLongjmpFName, &M);
777 
778     // Only traverse functions that uses setjmp in order not to insert
779     // unnecessary prep / cleanup code in every function
780     SmallPtrSet<Function *, 8> SetjmpUsers;
781     for (User *U : SetjmpF->users()) {
782       auto *UI = cast<Instruction>(U);
783       SetjmpUsers.insert(UI->getFunction());
784     }
785     for (Function *F : SetjmpUsers)
786       runSjLjOnFunction(*F);
787   }
788 
789   if (!Changed) {
790     // Delete unused global variables and functions
791     ThrewGV->eraseFromParent();
792     ThrewValueGV->eraseFromParent();
793     TempRet0GV->eraseFromParent();
794     if (ResumeF)
795       ResumeF->eraseFromParent();
796     if (EHTypeIDF)
797       EHTypeIDF->eraseFromParent();
798     if (EmLongjmpF)
799       EmLongjmpF->eraseFromParent();
800     if (SaveSetjmpF)
801       SaveSetjmpF->eraseFromParent();
802     if (TestSetjmpF)
803       TestSetjmpF->eraseFromParent();
804     return false;
805   }
806 
807   // If we have made any changes while doing exception handling or
808   // setjmp/longjmp handling, we have to create these functions for JavaScript
809   // to call.
810   createSetThrewFunction(M);
811   createSetTempRet0Function(M);
812 
813   return true;
814 }
815 
816 bool WebAssemblyLowerEmscriptenEHSjLj::runEHOnFunction(Function &F) {
817   Module &M = *F.getParent();
818   LLVMContext &C = F.getContext();
819   IRBuilder<> IRB(C);
820   bool Changed = false;
821   SmallVector<Instruction *, 64> ToErase;
822   SmallPtrSet<LandingPadInst *, 32> LandingPads;
823   bool AllowExceptions =
824       areAllExceptionsAllowed() || EHWhitelistSet.count(F.getName());
825 
826   for (BasicBlock &BB : F) {
827     auto *II = dyn_cast<InvokeInst>(BB.getTerminator());
828     if (!II)
829       continue;
830     Changed = true;
831     LandingPads.insert(II->getLandingPadInst());
832     IRB.SetInsertPoint(II);
833 
834     bool NeedInvoke = AllowExceptions && canThrow(II->getCalledValue());
835     if (NeedInvoke) {
836       // Wrap invoke with invoke wrapper and generate preamble/postamble
837       Value *Threw = wrapInvoke(II);
838       ToErase.push_back(II);
839 
840       // Insert a branch based on __THREW__ variable
841       Value *Cmp = IRB.CreateICmpEQ(Threw, IRB.getInt32(1), "cmp");
842       IRB.CreateCondBr(Cmp, II->getUnwindDest(), II->getNormalDest());
843 
844     } else {
845       // This can't throw, and we don't need this invoke, just replace it with a
846       // call+branch
847       SmallVector<Value *, 16> Args(II->arg_begin(), II->arg_end());
848       CallInst *NewCall = IRB.CreateCall(II->getCalledValue(), Args);
849       NewCall->takeName(II);
850       NewCall->setCallingConv(II->getCallingConv());
851       NewCall->setDebugLoc(II->getDebugLoc());
852       NewCall->setAttributes(II->getAttributes());
853       II->replaceAllUsesWith(NewCall);
854       ToErase.push_back(II);
855 
856       IRB.CreateBr(II->getNormalDest());
857 
858       // Remove any PHI node entries from the exception destination
859       II->getUnwindDest()->removePredecessor(&BB);
860     }
861   }
862 
863   // Process resume instructions
864   for (BasicBlock &BB : F) {
865     // Scan the body of the basic block for resumes
866     for (Instruction &I : BB) {
867       auto *RI = dyn_cast<ResumeInst>(&I);
868       if (!RI)
869         continue;
870 
871       // Split the input into legal values
872       Value *Input = RI->getValue();
873       IRB.SetInsertPoint(RI);
874       Value *Low = IRB.CreateExtractValue(Input, 0, "low");
875       // Create a call to __resumeException function
876       IRB.CreateCall(ResumeF, {Low});
877       // Add a terminator to the block
878       IRB.CreateUnreachable();
879       ToErase.push_back(RI);
880     }
881   }
882 
883   // Process llvm.eh.typeid.for intrinsics
884   for (BasicBlock &BB : F) {
885     for (Instruction &I : BB) {
886       auto *CI = dyn_cast<CallInst>(&I);
887       if (!CI)
888         continue;
889       const Function *Callee = CI->getCalledFunction();
890       if (!Callee)
891         continue;
892       if (Callee->getIntrinsicID() != Intrinsic::eh_typeid_for)
893         continue;
894 
895       IRB.SetInsertPoint(CI);
896       CallInst *NewCI =
897           IRB.CreateCall(EHTypeIDF, CI->getArgOperand(0), "typeid");
898       CI->replaceAllUsesWith(NewCI);
899       ToErase.push_back(CI);
900     }
901   }
902 
903   // Look for orphan landingpads, can occur in blocks with no predecesors
904   for (BasicBlock &BB : F) {
905     Instruction *I = BB.getFirstNonPHI();
906     if (auto *LPI = dyn_cast<LandingPadInst>(I))
907       LandingPads.insert(LPI);
908   }
909 
910   // Handle all the landingpad for this function together, as multiple invokes
911   // may share a single lp
912   for (LandingPadInst *LPI : LandingPads) {
913     IRB.SetInsertPoint(LPI);
914     SmallVector<Value *, 16> FMCArgs;
915     for (unsigned i = 0, e = LPI->getNumClauses(); i < e; ++i) {
916       Constant *Clause = LPI->getClause(i);
917       // As a temporary workaround for the lack of aggregate varargs support
918       // in the interface between JS and wasm, break out filter operands into
919       // their component elements.
920       if (LPI->isFilter(i)) {
921         auto *ATy = cast<ArrayType>(Clause->getType());
922         for (unsigned j = 0, e = ATy->getNumElements(); j < e; ++j) {
923           Value *EV = IRB.CreateExtractValue(Clause, makeArrayRef(j), "filter");
924           FMCArgs.push_back(EV);
925         }
926       } else
927         FMCArgs.push_back(Clause);
928     }
929 
930     // Create a call to __cxa_find_matching_catch_N function
931     Function *FMCF = getFindMatchingCatch(M, FMCArgs.size());
932     CallInst *FMCI = IRB.CreateCall(FMCF, FMCArgs, "fmc");
933     Value *Undef = UndefValue::get(LPI->getType());
934     Value *Pair0 = IRB.CreateInsertValue(Undef, FMCI, 0, "pair0");
935     Value *TempRet0 =
936         IRB.CreateLoad(TempRet0GV, TempRet0GV->getName() + ".val");
937     Value *Pair1 = IRB.CreateInsertValue(Pair0, TempRet0, 1, "pair1");
938 
939     LPI->replaceAllUsesWith(Pair1);
940     ToErase.push_back(LPI);
941   }
942 
943   // Erase everything we no longer need in this function
944   for (Instruction *I : ToErase)
945     I->eraseFromParent();
946 
947   return Changed;
948 }
949 
950 bool WebAssemblyLowerEmscriptenEHSjLj::runSjLjOnFunction(Function &F) {
951   Module &M = *F.getParent();
952   LLVMContext &C = F.getContext();
953   IRBuilder<> IRB(C);
954   SmallVector<Instruction *, 64> ToErase;
955   // Vector of %setjmpTable values
956   std::vector<Instruction *> SetjmpTableInsts;
957   // Vector of %setjmpTableSize values
958   std::vector<Instruction *> SetjmpTableSizeInsts;
959 
960   // Setjmp preparation
961 
962   // This instruction effectively means %setjmpTableSize = 4.
963   // We create this as an instruction intentionally, and we don't want to fold
964   // this instruction to a constant 4, because this value will be used in
965   // SSAUpdater.AddAvailableValue(...) later.
966   BasicBlock &EntryBB = F.getEntryBlock();
967   BinaryOperator *SetjmpTableSize = BinaryOperator::Create(
968       Instruction::Add, IRB.getInt32(4), IRB.getInt32(0), "setjmpTableSize",
969       &*EntryBB.getFirstInsertionPt());
970   // setjmpTable = (int *) malloc(40);
971   Instruction *SetjmpTable = CallInst::CreateMalloc(
972       SetjmpTableSize, IRB.getInt32Ty(), IRB.getInt32Ty(), IRB.getInt32(40),
973       nullptr, nullptr, "setjmpTable");
974   // setjmpTable[0] = 0;
975   IRB.SetInsertPoint(SetjmpTableSize);
976   IRB.CreateStore(IRB.getInt32(0), SetjmpTable);
977   SetjmpTableInsts.push_back(SetjmpTable);
978   SetjmpTableSizeInsts.push_back(SetjmpTableSize);
979 
980   // Setjmp transformation
981   std::vector<PHINode *> SetjmpRetPHIs;
982   Function *SetjmpF = M.getFunction("setjmp");
983   for (User *U : SetjmpF->users()) {
984     auto *CI = dyn_cast<CallInst>(U);
985     if (!CI)
986       report_fatal_error("Does not support indirect calls to setjmp");
987 
988     BasicBlock *BB = CI->getParent();
989     if (BB->getParent() != &F) // in other function
990       continue;
991 
992     // The tail is everything right after the call, and will be reached once
993     // when setjmp is called, and later when longjmp returns to the setjmp
994     BasicBlock *Tail = SplitBlock(BB, CI->getNextNode());
995     // Add a phi to the tail, which will be the output of setjmp, which
996     // indicates if this is the first call or a longjmp back. The phi directly
997     // uses the right value based on where we arrive from
998     IRB.SetInsertPoint(Tail->getFirstNonPHI());
999     PHINode *SetjmpRet = IRB.CreatePHI(IRB.getInt32Ty(), 2, "setjmp.ret");
1000 
1001     // setjmp initial call returns 0
1002     SetjmpRet->addIncoming(IRB.getInt32(0), BB);
1003     // The proper output is now this, not the setjmp call itself
1004     CI->replaceAllUsesWith(SetjmpRet);
1005     // longjmp returns to the setjmp will add themselves to this phi
1006     SetjmpRetPHIs.push_back(SetjmpRet);
1007 
1008     // Fix call target
1009     // Our index in the function is our place in the array + 1 to avoid index
1010     // 0, because index 0 means the longjmp is not ours to handle.
1011     IRB.SetInsertPoint(CI);
1012     Value *Args[] = {CI->getArgOperand(0), IRB.getInt32(SetjmpRetPHIs.size()),
1013                      SetjmpTable, SetjmpTableSize};
1014     Instruction *NewSetjmpTable =
1015         IRB.CreateCall(SaveSetjmpF, Args, "setjmpTable");
1016     Instruction *NewSetjmpTableSize =
1017         IRB.CreateLoad(TempRet0GV, "setjmpTableSize");
1018     SetjmpTableInsts.push_back(NewSetjmpTable);
1019     SetjmpTableSizeInsts.push_back(NewSetjmpTableSize);
1020     ToErase.push_back(CI);
1021   }
1022 
1023   // Update each call that can longjmp so it can return to a setjmp where
1024   // relevant.
1025 
1026   // Because we are creating new BBs while processing and don't want to make
1027   // all these newly created BBs candidates again for longjmp processing, we
1028   // first make the vector of candidate BBs.
1029   std::vector<BasicBlock *> BBs;
1030   for (BasicBlock &BB : F)
1031     BBs.push_back(&BB);
1032 
1033   // BBs.size() will change within the loop, so we query it every time
1034   for (unsigned i = 0; i < BBs.size(); i++) {
1035     BasicBlock *BB = BBs[i];
1036     for (Instruction &I : *BB) {
1037       assert(!isa<InvokeInst>(&I));
1038       auto *CI = dyn_cast<CallInst>(&I);
1039       if (!CI)
1040         continue;
1041 
1042       const Value *Callee = CI->getCalledValue();
1043       if (!canLongjmp(M, Callee))
1044         continue;
1045 
1046       Value *Threw = nullptr;
1047       BasicBlock *Tail;
1048       if (Callee->getName().startswith(InvokePrefix)) {
1049         // If invoke wrapper has already been generated for this call in
1050         // previous EH phase, search for the load instruction
1051         // %__THREW__.val = __THREW__;
1052         // in postamble after the invoke wrapper call
1053         LoadInst *ThrewLI = nullptr;
1054         StoreInst *ThrewResetSI = nullptr;
1055         for (auto I = std::next(BasicBlock::iterator(CI)), IE = BB->end();
1056              I != IE; ++I) {
1057           if (auto *LI = dyn_cast<LoadInst>(I))
1058             if (auto *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand()))
1059               if (GV == ThrewGV) {
1060                 Threw = ThrewLI = LI;
1061                 break;
1062               }
1063         }
1064         // Search for the store instruction after the load above
1065         // __THREW__ = 0;
1066         for (auto I = std::next(BasicBlock::iterator(ThrewLI)), IE = BB->end();
1067              I != IE; ++I) {
1068           if (auto *SI = dyn_cast<StoreInst>(I))
1069             if (auto *GV = dyn_cast<GlobalVariable>(SI->getPointerOperand()))
1070               if (GV == ThrewGV && SI->getValueOperand() == IRB.getInt32(0)) {
1071                 ThrewResetSI = SI;
1072                 break;
1073               }
1074         }
1075         assert(Threw && ThrewLI && "Cannot find __THREW__ load after invoke");
1076         assert(ThrewResetSI && "Cannot find __THREW__ store after invoke");
1077         Tail = SplitBlock(BB, ThrewResetSI->getNextNode());
1078 
1079       } else {
1080         // Wrap call with invoke wrapper and generate preamble/postamble
1081         Threw = wrapInvoke(CI);
1082         ToErase.push_back(CI);
1083         Tail = SplitBlock(BB, CI->getNextNode());
1084       }
1085 
1086       // We need to replace the terminator in Tail - SplitBlock makes BB go
1087       // straight to Tail, we need to check if a longjmp occurred, and go to the
1088       // right setjmp-tail if so
1089       ToErase.push_back(BB->getTerminator());
1090 
1091       // Generate a function call to testSetjmp function and preamble/postamble
1092       // code to figure out (1) whether longjmp occurred (2) if longjmp
1093       // occurred, which setjmp it corresponds to
1094       Value *Label = nullptr;
1095       Value *LongjmpResult = nullptr;
1096       BasicBlock *EndBB = nullptr;
1097       wrapTestSetjmp(BB, CI, Threw, SetjmpTable, SetjmpTableSize, Label,
1098                      LongjmpResult, EndBB);
1099       assert(Label && LongjmpResult && EndBB);
1100 
1101       // Create switch instruction
1102       IRB.SetInsertPoint(EndBB);
1103       SwitchInst *SI = IRB.CreateSwitch(Label, Tail, SetjmpRetPHIs.size());
1104       // -1 means no longjmp happened, continue normally (will hit the default
1105       // switch case). 0 means a longjmp that is not ours to handle, needs a
1106       // rethrow. Otherwise the index is the same as the index in P+1 (to avoid
1107       // 0).
1108       for (unsigned i = 0; i < SetjmpRetPHIs.size(); i++) {
1109         SI->addCase(IRB.getInt32(i + 1), SetjmpRetPHIs[i]->getParent());
1110         SetjmpRetPHIs[i]->addIncoming(LongjmpResult, EndBB);
1111       }
1112 
1113       // We are splitting the block here, and must continue to find other calls
1114       // in the block - which is now split. so continue to traverse in the Tail
1115       BBs.push_back(Tail);
1116     }
1117   }
1118 
1119   // Erase everything we no longer need in this function
1120   for (Instruction *I : ToErase)
1121     I->eraseFromParent();
1122 
1123   // Free setjmpTable buffer before each return instruction
1124   for (BasicBlock &BB : F) {
1125     TerminatorInst *TI = BB.getTerminator();
1126     if (isa<ReturnInst>(TI))
1127       CallInst::CreateFree(SetjmpTable, TI);
1128   }
1129 
1130   // Every call to saveSetjmp can change setjmpTable and setjmpTableSize
1131   // (when buffer reallocation occurs)
1132   // entry:
1133   //   setjmpTableSize = 4;
1134   //   setjmpTable = (int *) malloc(40);
1135   //   setjmpTable[0] = 0;
1136   // ...
1137   // somebb:
1138   //   setjmpTable = saveSetjmp(buf, label, setjmpTable, setjmpTableSize);
1139   //   setjmpTableSize = __tempRet0;
1140   // So we need to make sure the SSA for these variables is valid so that every
1141   // saveSetjmp and testSetjmp calls have the correct arguments.
1142   SSAUpdater SetjmpTableSSA;
1143   SSAUpdater SetjmpTableSizeSSA;
1144   SetjmpTableSSA.Initialize(Type::getInt32PtrTy(C), "setjmpTable");
1145   SetjmpTableSizeSSA.Initialize(Type::getInt32Ty(C), "setjmpTableSize");
1146   for (Instruction *I : SetjmpTableInsts)
1147     SetjmpTableSSA.AddAvailableValue(I->getParent(), I);
1148   for (Instruction *I : SetjmpTableSizeInsts)
1149     SetjmpTableSizeSSA.AddAvailableValue(I->getParent(), I);
1150 
1151   for (auto UI = SetjmpTable->use_begin(), UE = SetjmpTable->use_end();
1152        UI != UE;) {
1153     // Grab the use before incrementing the iterator.
1154     Use &U = *UI;
1155     // Increment the iterator before removing the use from the list.
1156     ++UI;
1157     if (Instruction *I = dyn_cast<Instruction>(U.getUser()))
1158       if (I->getParent() != &EntryBB)
1159         SetjmpTableSSA.RewriteUse(U);
1160   }
1161   for (auto UI = SetjmpTableSize->use_begin(), UE = SetjmpTableSize->use_end();
1162        UI != UE;) {
1163     Use &U = *UI;
1164     ++UI;
1165     if (Instruction *I = dyn_cast<Instruction>(U.getUser()))
1166       if (I->getParent() != &EntryBB)
1167         SetjmpTableSizeSSA.RewriteUse(U);
1168   }
1169 
1170   // Finally, our modifications to the cfg can break dominance of SSA variables.
1171   // For example, in this code,
1172   // if (x()) { .. setjmp() .. }
1173   // if (y()) { .. longjmp() .. }
1174   // We must split the longjmp block, and it can jump into the block splitted
1175   // from setjmp one. But that means that when we split the setjmp block, it's
1176   // first part no longer dominates its second part - there is a theoretically
1177   // possible control flow path where x() is false, then y() is true and we
1178   // reach the second part of the setjmp block, without ever reaching the first
1179   // part. So, we rebuild SSA form here.
1180   rebuildSSA(F);
1181   return true;
1182 }
1183