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 function
11	/// calls to use Emscripten's library functions. The pass uses JavaScript's try
12	/// and catch mechanism in case of Emscripten EH/SjLj and Wasm EH intrinsics in
13	/// case of Emscripten SjLJ.
14	///
15	/// * Emscripten exception handling
16	/// This pass lowers invokes and landingpads into library functions in JS glue
17	/// code. Invokes are lowered into function wrappers called invoke wrappers that
18	/// exist in JS side, which wraps the original function call with JS try-catch.
19	/// If an exception occurred, cxa_throw() function in JS side sets some
20	/// variables (see below) so we can check whether an exception occurred from
21	/// wasm code and handle it appropriately.
22	///
23	/// * Emscripten setjmp-longjmp handling
24	/// This pass lowers setjmp to a reasonably-performant approach for emscripten.
25	/// The idea is that each block with a setjmp is broken up into two parts: the
26	/// part containing setjmp and the part right after the setjmp. The latter part
27	/// is either reached from the setjmp, or later from a longjmp. To handle the
28	/// longjmp, all calls that might longjmp are also called using invoke wrappers
29	/// and thus JS / try-catch. JS longjmp() function also sets some variables so
30	/// we can check / whether a longjmp occurred from wasm code. Each block with a
31	/// function call that might longjmp is also split up after the longjmp call.
32	/// After the longjmp call, we check whether a longjmp occurred, and if it did,
33	/// which setjmp it corresponds to, and jump to the right post-setjmp block.
34	/// We assume setjmp-longjmp handling always run after EH handling, which means
35	/// we don't expect any exception-related instructions when SjLj runs.
36	/// FIXME Currently this scheme does not support indirect call of setjmp,
37	/// because of the limitation of the scheme itself. fastcomp does not support it
38	/// either.
39	///
40	/// In detail, this pass does following things:
41	///
42	/// 1) Assumes the existence of global variables: __THREW__, __threwValue
43	/// __THREW__ and __threwValue are defined in compiler-rt in Emscripten.
44	/// These variables are used for both exceptions and setjmp/longjmps.
45	/// __THREW__ indicates whether an exception or a longjmp occurred or not. 0
46	/// means nothing occurred, 1 means an exception occurred, and other numbers
47	/// mean a longjmp occurred. In the case of longjmp, __THREW__ variable
48	/// indicates the corresponding setjmp buffer the longjmp corresponds to.
49	/// __threwValue is 0 for exceptions, and the argument to longjmp in case of
50	/// longjmp.
51	///
52	/// * Emscripten exception handling
53	///
54	/// 2) We assume the existence of setThrew and setTempRet0/getTempRet0 functions
55	/// at link time. setThrew exists in Emscripten's compiler-rt:
56	///
57	/// void setThrew(uintptr_t threw, int value) {
58	/// if (__THREW__ == 0) {
59	/// __THREW__ = threw;
60	/// __threwValue = value;
61	/// }
62	/// }
63	//
64	/// setTempRet0 is called from __cxa_find_matching_catch() in JS glue code.
65	/// In exception handling, getTempRet0 indicates the type of an exception
66	/// caught, and in setjmp/longjmp, it means the second argument to longjmp
67	/// function.
68	///
69	/// 3) Lower
70	/// invoke @func(arg1, arg2) to label %invoke.cont unwind label %lpad
71	/// into
72	/// __THREW__ = 0;
73	/// call @__invoke_SIG(func, arg1, arg2)
74	/// %__THREW__.val = __THREW__;
75	/// __THREW__ = 0;
76	/// if (%__THREW__.val == 1)
77	/// goto %lpad
78	/// else
79	/// goto %invoke.cont
80	/// SIG is a mangled string generated based on the LLVM IR-level function
81	/// signature. After LLVM IR types are lowered to the target wasm types,
82	/// the names for these wrappers will change based on wasm types as well,
83	/// as in invoke_vi (function takes an int and returns void). The bodies of
84	/// these wrappers will be generated in JS glue code, and inside those
85	/// wrappers we use JS try-catch to generate actual exception effects. It
86	/// also calls the original callee function. An example wrapper in JS code
87	/// would look like this:
88	/// function invoke_vi(index,a1) {
89	/// try {
90	/// Module["dynCall_vi"](index,a1); // This calls original callee
91	/// } catch(e) {
92	/// if (typeof e !== 'number' && e !== 'longjmp') throw e;
93	/// _setThrew(1, 0); // setThrew is called here
94	/// }
95	/// }
96	/// If an exception is thrown, __THREW__ will be set to true in a wrapper,
97	/// so we can jump to the right BB based on this value.
98	///
99	/// 4) Lower
100	/// %val = landingpad catch c1 catch c2 catch c3 ...
101	/// ... use %val ...
102	/// into
103	/// %fmc = call @__cxa_find_matching_catch_N(c1, c2, c3, ...)
104	/// %val = {%fmc, getTempRet0()}
105	/// ... use %val ...
106	/// Here N is a number calculated based on the number of clauses.
107	/// setTempRet0 is called from __cxa_find_matching_catch() in JS glue code.
108	///
109	/// 5) Lower
110	/// resume {%a, %b}
111	/// into
112	/// call @__resumeException(%a)
113	/// where __resumeException() is a function in JS glue code.
114	///
115	/// 6) Lower
116	/// call @llvm.eh.typeid.for(type) (intrinsic)
117	/// into
118	/// call @llvm_eh_typeid_for(type)
119	/// llvm_eh_typeid_for function will be generated in JS glue code.
120	///
121	/// * Emscripten setjmp / longjmp handling
122	///
123	/// If there are calls to longjmp()
124	///
125	/// 1) Lower
126	/// longjmp(env, val)
127	/// into
128	/// emscripten_longjmp(env, val)
129	///
130	/// If there are calls to setjmp()
131	///
132	/// 2) In the function entry that calls setjmp, initialize
133	/// functionInvocationId as follows:
134	///
135	/// functionInvocationId = alloca(4)
136	///
137	/// Note: the alloca size is not important as this pointer is
138	/// merely used for pointer comparisions.
139	///
140	/// 3) Lower
141	/// setjmp(env)
142	/// into
143	/// __wasm_setjmp(env, label, functionInvocationId)
144	///
145	/// __wasm_setjmp records the necessary info (the label and
146	/// functionInvocationId) to the "env".
147	/// A BB with setjmp is split into two after setjmp call in order to
148	/// make the post-setjmp BB the possible destination of longjmp BB.
149	///
150	/// 4) Lower every call that might longjmp into
151	/// __THREW__ = 0;
152	/// call @__invoke_SIG(func, arg1, arg2)
153	/// %__THREW__.val = __THREW__;
154	/// __THREW__ = 0;
155	/// %__threwValue.val = __threwValue;
156	/// if (%__THREW__.val != 0 & %__threwValue.val != 0) {
157	/// %label = __wasm_setjmp_test(%__THREW__.val, functionInvocationId);
158	/// if (%label == 0)
159	/// emscripten_longjmp(%__THREW__.val, %__threwValue.val);
160	/// setTempRet0(%__threwValue.val);
161	/// } else {
162	/// %label = -1;
163	/// }
164	/// longjmp_result = getTempRet0();
165	/// switch %label {
166	/// label 1: goto post-setjmp BB 1
167	/// label 2: goto post-setjmp BB 2
168	/// ...
169	/// default: goto splitted next BB
170	/// }
171	///
172	/// __wasm_setjmp_test examines the jmp buf to see if it was for a matching
173	/// setjmp call. After calling an invoke wrapper, if a longjmp occurred,
174	/// __THREW__ will be the address of matching jmp_buf buffer and
175	/// __threwValue be the second argument to longjmp.
176	/// __wasm_setjmp_test returns a setjmp label, a unique ID to each setjmp
177	/// callsite. Label 0 means this longjmp buffer does not correspond to one
178	/// of the setjmp callsites in this function, so in this case we just chain
179	/// the longjmp to the caller. Label -1 means no longjmp occurred.
180	/// Otherwise we jump to the right post-setjmp BB based on the label.
181	///
182	/// * Wasm setjmp / longjmp handling
183	/// This mode still uses some Emscripten library functions but not JavaScript's
184	/// try-catch mechanism. It instead uses Wasm exception handling intrinsics,
185	/// which will be lowered to exception handling instructions.
186	///
187	/// If there are calls to longjmp()
188	///
189	/// 1) Lower
190	/// longjmp(env, val)
191	/// into
192	/// __wasm_longjmp(env, val)
193	///
194	/// If there are calls to setjmp()
195	///
196	/// 2) and 3): The same as 2) and 3) in Emscripten SjLj.
197	/// (functionInvocationId initialization + setjmp callsite transformation)
198	///
199	/// 4) Create a catchpad with a wasm.catch() intrinsic, which returns the value
200	/// thrown by __wasm_longjmp function. In the runtime library, we have an
201	/// equivalent of the following struct:
202	///
203	/// struct __WasmLongjmpArgs {
204	/// void *env;
205	/// int val;
206	/// };
207	///
208	/// The thrown value here is a pointer to the struct. We use this struct to
209	/// transfer two values by throwing a single value. Wasm throw and catch
210	/// instructions are capable of throwing and catching multiple values, but
211	/// it also requires multivalue support that is currently not very reliable.
212	/// TODO Switch to throwing and catching two values without using the struct
213	///
214	/// All longjmpable function calls will be converted to an invoke that will
215	/// unwind to this catchpad in case a longjmp occurs. Within the catchpad, we
216	/// test the thrown values using __wasm_setjmp_test function as we do for
217	/// Emscripten SjLj. The main difference is, in Emscripten SjLj, we need to
218	/// transform every longjmpable callsite into a sequence of code including
219	/// __wasm_setjmp_test() call; in Wasm SjLj we do the testing in only one
220	/// place, in this catchpad.
221	///
222	/// After testing calling __wasm_setjmp_test(), if the longjmp does not
223	/// correspond to one of the setjmps within the current function, it rethrows
224	/// the longjmp by calling __wasm_longjmp(). If it corresponds to one of
225	/// setjmps in the function, we jump to the beginning of the function, which
226	/// contains a switch to each post-setjmp BB. Again, in Emscripten SjLj, this
227	/// switch is added for every longjmpable callsite; in Wasm SjLj we do this
228	/// only once at the top of the function. (after functionInvocationId
229	/// initialization)
230	///
231	/// The below is the pseudocode for what we have described
232	///
233	/// entry:
234	/// Initialize functionInvocationId
235	///
236	/// setjmp.dispatch:
237	/// switch %label {
238	/// label 1: goto post-setjmp BB 1
239	/// label 2: goto post-setjmp BB 2
240	/// ...
241	/// default: goto splitted next BB
242	/// }
243	/// ...
244	///
245	/// bb:
246	/// invoke void @foo() ;; foo is a longjmpable function
247	/// to label %next unwind label %catch.dispatch.longjmp
248	/// ...
249	///
250	/// catch.dispatch.longjmp:
251	/// %0 = catchswitch within none [label %catch.longjmp] unwind to caller
252	///
253	/// catch.longjmp:
254	/// %longjmp.args = wasm.catch() ;; struct __WasmLongjmpArgs
255	/// %env = load 'env' field from __WasmLongjmpArgs
256	/// %val = load 'val' field from __WasmLongjmpArgs
257	/// %label = __wasm_setjmp_test(%env, functionInvocationId);
258	/// if (%label == 0)
259	/// __wasm_longjmp(%env, %val)
260	/// catchret to %setjmp.dispatch
261	///
262	///===----------------------------------------------------------------------===//
263
264	#include "MCTargetDesc/WebAssemblyMCTargetDesc.h"
265	#include "WebAssembly.h"
266	#include "WebAssemblyTargetMachine.h"
267	#include "llvm/ADT/StringExtras.h"
268	#include "llvm/CodeGen/TargetPassConfig.h"
269	#include "llvm/CodeGen/WasmEHFuncInfo.h"
270	#include "llvm/IR/DebugInfoMetadata.h"
271	#include "llvm/IR/Dominators.h"
272	#include "llvm/IR/IRBuilder.h"
273	#include "llvm/IR/IntrinsicsWebAssembly.h"
274	#include "llvm/Support/CommandLine.h"
275	#include "llvm/Transforms/Utils/BasicBlockUtils.h"
276	#include "llvm/Transforms/Utils/Local.h"
277	#include "llvm/Transforms/Utils/SSAUpdater.h"
278	#include "llvm/Transforms/Utils/SSAUpdaterBulk.h"
279	#include <set>
280
281	using namespace llvm;
282
283	#define DEBUG_TYPE "wasm-lower-em-ehsjlj"
284
285	static cl::list<std::string>
286	EHAllowlist("emscripten-cxx-exceptions-allowed",
287	cl::desc("The list of function names in which Emscripten-style "
288	"exception handling is enabled (see emscripten "
289	"EMSCRIPTEN_CATCHING_ALLOWED options)"),
290	cl::CommaSeparated);
291
292	namespace {
293	class WebAssemblyLowerEmscriptenEHSjLj final : public ModulePass {
294	bool EnableEmEH; // Enable Emscripten exception handling
295	bool EnableEmSjLj; // Enable Emscripten setjmp/longjmp handling
296	bool EnableWasmSjLj; // Enable Wasm setjmp/longjmp handling
297	bool DoSjLj; // Whether we actually perform setjmp/longjmp handling
298
299	GlobalVariable *ThrewGV = nullptr; // __THREW__ (Emscripten)
300	GlobalVariable *ThrewValueGV = nullptr; // __threwValue (Emscripten)
301	Function *GetTempRet0F = nullptr; // getTempRet0() (Emscripten)
302	Function *SetTempRet0F = nullptr; // setTempRet0() (Emscripten)
303	Function *ResumeF = nullptr; // __resumeException() (Emscripten)
304	Function *EHTypeIDF = nullptr; // llvm.eh.typeid.for() (intrinsic)
305	Function *EmLongjmpF = nullptr; // emscripten_longjmp() (Emscripten)
306	Function *WasmSetjmpF = nullptr; // __wasm_setjmp() (Emscripten)
307	Function *WasmSetjmpTestF = nullptr; // __wasm_setjmp_test() (Emscripten)
308	Function *WasmLongjmpF = nullptr; // __wasm_longjmp() (Emscripten)
309	Function *CatchF = nullptr; // wasm.catch() (intrinsic)
310
311	// type of 'struct __WasmLongjmpArgs' defined in emscripten
312	Type *LongjmpArgsTy = nullptr;
313
314	// __cxa_find_matching_catch_N functions.
315	// Indexed by the number of clauses in an original landingpad instruction.
316	DenseMap<int, Function *> FindMatchingCatches;
317	// Map of <function signature string, invoke_ wrappers>
318	StringMap<Function *> InvokeWrappers;
319	// Set of allowed function names for exception handling
320	std::set<std::string> EHAllowlistSet;
321	// Functions that contains calls to setjmp
322	SmallPtrSet<Function *, 8> SetjmpUsers;
323
324	StringRef getPassName() const override {
325	return "WebAssembly Lower Emscripten Exceptions";
326	}
327
328	using InstVector = SmallVectorImpl<Instruction *>;
329	bool runEHOnFunction(Function &F);
330	bool runSjLjOnFunction(Function &F);
331	void handleLongjmpableCallsForEmscriptenSjLj(
332	Function &F, Instruction *FunctionInvocationId,
333	SmallVectorImpl<PHINode *> &SetjmpRetPHIs);
334	void
335	handleLongjmpableCallsForWasmSjLj(Function &F,
336	Instruction *FunctionInvocationId,
337	SmallVectorImpl<PHINode *> &SetjmpRetPHIs);
338	Function *getFindMatchingCatch(Module &M, unsigned NumClauses);
339
340	Value *wrapInvoke(CallBase *CI);
341	void wrapTestSetjmp(BasicBlock *BB, DebugLoc DL, Value *Threw,
342	Value *FunctionInvocationId, Value *&Label,
343	Value *&LongjmpResult, BasicBlock *&CallEmLongjmpBB,
344	PHINode *&CallEmLongjmpBBThrewPHI,
345	PHINode *&CallEmLongjmpBBThrewValuePHI,
346	BasicBlock *&EndBB);
347	Function *getInvokeWrapper(CallBase *CI);
348
349	bool areAllExceptionsAllowed() const { return EHAllowlistSet.empty(); }
350	bool supportsException(const Function *F) const {
351	return EnableEmEH && (areAllExceptionsAllowed() ||
352	EHAllowlistSet.count(x: std::string(F->getName())));
353	}
354	void replaceLongjmpWith(Function *LongjmpF, Function *NewF);
355
356	void rebuildSSA(Function &F);
357
358	public:
359	static char ID;
360
361	WebAssemblyLowerEmscriptenEHSjLj()
362	: ModulePass(ID), EnableEmEH(WebAssembly::WasmEnableEmEH),
363	EnableEmSjLj(WebAssembly::WasmEnableEmSjLj),
364	EnableWasmSjLj(WebAssembly::WasmEnableSjLj) {
365	assert(!(EnableEmSjLj && EnableWasmSjLj) &&
366	"Two SjLj modes cannot be turned on at the same time");
367	assert(!(EnableEmEH && EnableWasmSjLj) &&
368	"Wasm SjLj should be only used with Wasm EH");
369	EHAllowlistSet.insert(first: EHAllowlist.begin(), last: EHAllowlist.end());
370	}
371	bool runOnModule(Module &M) override;
372
373	void getAnalysisUsage(AnalysisUsage &AU) const override {
374	AU.addRequired<DominatorTreeWrapperPass>();
375	}
376	};
377	} // End anonymous namespace
378
379	char WebAssemblyLowerEmscriptenEHSjLj::ID = 0;
380	INITIALIZE_PASS(WebAssemblyLowerEmscriptenEHSjLj, DEBUG_TYPE,
381	"WebAssembly Lower Emscripten Exceptions / Setjmp / Longjmp",
382	false, false)
383
384	ModulePass *llvm::createWebAssemblyLowerEmscriptenEHSjLj() {
385	return new WebAssemblyLowerEmscriptenEHSjLj();
386	}
387
388	static bool canThrow(const Value *V) {
389	if (const auto *F = dyn_cast<const Function>(Val: V)) {
390	// Intrinsics cannot throw
391	if (F->isIntrinsic())
392	return false;
393	StringRef Name = F->getName();
394	// leave setjmp and longjmp (mostly) alone, we process them properly later
395	if (Name == "setjmp" || Name == "longjmp" || Name == "emscripten_longjmp")
396	return false;
397	return !F->doesNotThrow();
398	}
399	// not a function, so an indirect call - can throw, we can't tell
400	return true;
401	}
402
403	// Get a thread-local global variable with the given name. If it doesn't exist
404	// declare it, which will generate an import and assume that it will exist at
405	// link time.
406	static GlobalVariable *getGlobalVariable(Module &M, Type *Ty,
407	WebAssemblyTargetMachine &TM,
408	const char *Name) {
409	auto *GV = dyn_cast<GlobalVariable>(Val: M.getOrInsertGlobal(Name, Ty));
410	if (!GV)
411	report_fatal_error(reason: Twine("unable to create global: ") + Name);
412
413	// Variables created by this function are thread local. If the target does not
414	// support TLS, we depend on CoalesceFeaturesAndStripAtomics to downgrade it
415	// to non-thread-local ones, in which case we don't allow this object to be
416	// linked with other objects using shared memory.
417	GV->setThreadLocalMode(GlobalValue::GeneralDynamicTLSModel);
418	return GV;
419	}
420
421	// Simple function name mangler.
422	// This function simply takes LLVM's string representation of parameter types
423	// and concatenate them with '_'. There are non-alphanumeric characters but llc
424	// is ok with it, and we need to postprocess these names after the lowering
425	// phase anyway.
426	static std::string getSignature(FunctionType *FTy) {
427	std::string Sig;
428	raw_string_ostream OS(Sig);
429	OS << *FTy->getReturnType();
430	for (Type *ParamTy : FTy->params())
431	OS << "_" << *ParamTy;
432	if (FTy->isVarArg())
433	OS << "_...";
434	Sig = OS.str();
435	erase_if(C&: Sig, P: isSpace);
436	// When s2wasm parses .s file, a comma means the end of an argument. So a
437	// mangled function name can contain any character but a comma.
438	std::replace(first: Sig.begin(), last: Sig.end(), old_value: ',', new_value: '.');
439	return Sig;
440	}
441
442	static Function *getEmscriptenFunction(FunctionType *Ty, const Twine &Name,
443	Module *M) {
444	Function* F = Function::Create(Ty, Linkage: GlobalValue::ExternalLinkage, N: Name, M);
445	// Tell the linker that this function is expected to be imported from the
446	// 'env' module.
447	if (!F->hasFnAttribute(Kind: "wasm-import-module")) {
448	llvm::AttrBuilder B(M->getContext());
449	B.addAttribute(A: "wasm-import-module", V: "env");
450	F->addFnAttrs(Attrs: B);
451	}
452	if (!F->hasFnAttribute(Kind: "wasm-import-name")) {
453	llvm::AttrBuilder B(M->getContext());
454	B.addAttribute(A: "wasm-import-name", V: F->getName());
455	F->addFnAttrs(Attrs: B);
456	}
457	return F;
458	}
459
460	// Returns an integer type for the target architecture's address space.
461	// i32 for wasm32 and i64 for wasm64.
462	static Type *getAddrIntType(Module *M) {
463	IRBuilder<> IRB(M->getContext());
464	return IRB.getIntNTy(N: M->getDataLayout().getPointerSizeInBits());
465	}
466
467	// Returns an integer pointer type for the target architecture's address space.
468	// i32* for wasm32 and i64* for wasm64. With opaque pointers this is just a ptr
469	// in address space zero.
470	static Type *getAddrPtrType(Module *M) {
471	return PointerType::getUnqual(C&: M->getContext());
472	}
473
474	// Returns an integer whose type is the integer type for the target's address
475	// space. Returns (i32 C) for wasm32 and (i64 C) for wasm64, when C is the
476	// integer.
477	static Value *getAddrSizeInt(Module *M, uint64_t C) {
478	IRBuilder<> IRB(M->getContext());
479	return IRB.getIntN(N: M->getDataLayout().getPointerSizeInBits(), C);
480	}
481
482	// Returns __cxa_find_matching_catch_N function, where N = NumClauses + 2.
483	// This is because a landingpad instruction contains two more arguments, a
484	// personality function and a cleanup bit, and __cxa_find_matching_catch_N
485	// functions are named after the number of arguments in the original landingpad
486	// instruction.
487	Function *
488	WebAssemblyLowerEmscriptenEHSjLj::getFindMatchingCatch(Module &M,
489	unsigned NumClauses) {
490	if (FindMatchingCatches.count(Val: NumClauses))
491	return FindMatchingCatches[NumClauses];
492	PointerType *Int8PtrTy = PointerType::getUnqual(C&: M.getContext());
493	SmallVector<Type *, 16> Args(NumClauses, Int8PtrTy);
494	FunctionType *FTy = FunctionType::get(Result: Int8PtrTy, Params: Args, isVarArg: false);
495	Function *F = getEmscriptenFunction(
496	Ty: FTy, Name: "__cxa_find_matching_catch_" + Twine(NumClauses + 2), M: &M);
497	FindMatchingCatches[NumClauses] = F;
498	return F;
499	}
500
501	// Generate invoke wrapper seqence with preamble and postamble
502	// Preamble:
503	// __THREW__ = 0;
504	// Postamble:
505	// %__THREW__.val = __THREW__; __THREW__ = 0;
506	// Returns %__THREW__.val, which indicates whether an exception is thrown (or
507	// whether longjmp occurred), for future use.
508	Value *WebAssemblyLowerEmscriptenEHSjLj::wrapInvoke(CallBase *CI) {
509	Module *M = CI->getModule();
510	LLVMContext &C = M->getContext();
511
512	IRBuilder<> IRB(C);
513	IRB.SetInsertPoint(CI);
514
515	// Pre-invoke
516	// __THREW__ = 0;
517	IRB.CreateStore(Val: getAddrSizeInt(M, C: 0), Ptr: ThrewGV);
518
519	// Invoke function wrapper in JavaScript
520	SmallVector<Value *, 16> Args;
521	// Put the pointer to the callee as first argument, so it can be called
522	// within the invoke wrapper later
523	Args.push_back(Elt: CI->getCalledOperand());
524	Args.append(in_start: CI->arg_begin(), in_end: CI->arg_end());
525	CallInst *NewCall = IRB.CreateCall(Callee: getInvokeWrapper(CI), Args);
526	NewCall->takeName(V: CI);
527	NewCall->setCallingConv(CallingConv::WASM_EmscriptenInvoke);
528	NewCall->setDebugLoc(CI->getDebugLoc());
529
530	// Because we added the pointer to the callee as first argument, all
531	// argument attribute indices have to be incremented by one.
532	SmallVector<AttributeSet, 8> ArgAttributes;
533	const AttributeList &InvokeAL = CI->getAttributes();
534
535	// No attributes for the callee pointer.
536	ArgAttributes.push_back(Elt: AttributeSet());
537	// Copy the argument attributes from the original
538	for (unsigned I = 0, E = CI->arg_size(); I < E; ++I)
539	ArgAttributes.push_back(Elt: InvokeAL.getParamAttrs(ArgNo: I));
540
541	AttrBuilder FnAttrs(CI->getContext(), InvokeAL.getFnAttrs());
542	if (auto Args = FnAttrs.getAllocSizeArgs()) {
543	// The allocsize attribute (if any) referes to parameters by index and needs
544	// to be adjusted.
545	auto [SizeArg, NEltArg] = *Args;
546	SizeArg += 1;
547	if (NEltArg)
548	NEltArg = *NEltArg + 1;
549	FnAttrs.addAllocSizeAttr(ElemSizeArg: SizeArg, NumElemsArg: NEltArg);
550	}
551	// In case the callee has 'noreturn' attribute, We need to remove it, because
552	// we expect invoke wrappers to return.
553	FnAttrs.removeAttribute(Attribute::NoReturn);
554
555	// Reconstruct the AttributesList based on the vector we constructed.
556	AttributeList NewCallAL = AttributeList::get(
557	C, FnAttrs: AttributeSet::get(C, B: FnAttrs), RetAttrs: InvokeAL.getRetAttrs(), ArgAttrs: ArgAttributes);
558	NewCall->setAttributes(NewCallAL);
559
560	CI->replaceAllUsesWith(V: NewCall);
561
562	// Post-invoke
563	// %__THREW__.val = __THREW__; __THREW__ = 0;
564	Value *Threw =
565	IRB.CreateLoad(Ty: getAddrIntType(M), Ptr: ThrewGV, Name: ThrewGV->getName() + ".val");
566	IRB.CreateStore(Val: getAddrSizeInt(M, C: 0), Ptr: ThrewGV);
567	return Threw;
568	}
569
570	// Get matching invoke wrapper based on callee signature
571	Function *WebAssemblyLowerEmscriptenEHSjLj::getInvokeWrapper(CallBase *CI) {
572	Module *M = CI->getModule();
573	SmallVector<Type *, 16> ArgTys;
574	FunctionType *CalleeFTy = CI->getFunctionType();
575
576	std::string Sig = getSignature(FTy: CalleeFTy);
577	if (InvokeWrappers.contains(Key: Sig))
578	return InvokeWrappers[Sig];
579
580	// Put the pointer to the callee as first argument
581	ArgTys.push_back(Elt: PointerType::getUnqual(ElementType: CalleeFTy));
582	// Add argument types
583	ArgTys.append(in_start: CalleeFTy->param_begin(), in_end: CalleeFTy->param_end());
584
585	FunctionType *FTy = FunctionType::get(Result: CalleeFTy->getReturnType(), Params: ArgTys,
586	isVarArg: CalleeFTy->isVarArg());
587	Function *F = getEmscriptenFunction(Ty: FTy, Name: "__invoke_" + Sig, M);
588	InvokeWrappers[Sig] = F;
589	return F;
590	}
591
592	static bool canLongjmp(const Value *Callee) {
593	if (auto *CalleeF = dyn_cast<Function>(Val: Callee))
594	if (CalleeF->isIntrinsic())
595	return false;
596
597	// Attempting to transform inline assembly will result in something like:
598	// call void @__invoke_void(void ()* asm ...)
599	// which is invalid because inline assembly blocks do not have addresses
600	// and can't be passed by pointer. The result is a crash with illegal IR.
601	if (isa<InlineAsm>(Val: Callee))
602	return false;
603	StringRef CalleeName = Callee->getName();
604
605	// TODO Include more functions or consider checking with mangled prefixes
606
607	// The reason we include malloc/free here is to exclude the malloc/free
608	// calls generated in setjmp prep / cleanup routines.
609	if (CalleeName == "setjmp" || CalleeName == "malloc" || CalleeName == "free")
610	return false;
611
612	// There are functions in Emscripten's JS glue code or compiler-rt
613	if (CalleeName == "__resumeException" || CalleeName == "llvm_eh_typeid_for" ||
614	CalleeName == "__wasm_setjmp" || CalleeName == "__wasm_setjmp_test" ||
615	CalleeName == "getTempRet0" || CalleeName == "setTempRet0")
616	return false;
617
618	// __cxa_find_matching_catch_N functions cannot longjmp
619	if (Callee->getName().starts_with(Prefix: "__cxa_find_matching_catch_"))
620	return false;
621
622	// Exception-catching related functions
623	//
624	// We intentionally treat __cxa_end_catch longjmpable in Wasm SjLj even though
625	// it surely cannot longjmp, in order to maintain the unwind relationship from
626	// all existing catchpads (and calls within them) to catch.dispatch.longjmp.
627	//
628	// In Wasm EH + Wasm SjLj, we
629	// 1. Make all catchswitch and cleanuppad that unwind to caller unwind to
630	// catch.dispatch.longjmp instead
631	// 2. Convert all longjmpable calls to invokes that unwind to
632	// catch.dispatch.longjmp
633	// But catchswitch BBs are removed in isel, so if an EH catchswitch (generated
634	// from an exception)'s catchpad does not contain any calls that are converted
635	// into invokes unwinding to catch.dispatch.longjmp, this unwind relationship
636	// (EH catchswitch BB -> catch.dispatch.longjmp BB) is lost and
637	// catch.dispatch.longjmp BB can be placed before the EH catchswitch BB in
638	// CFGSort.
639	// int ret = setjmp(buf);
640	// try {
641	// foo(); // longjmps
642	// } catch (...) {
643	// }
644	// Then in this code, if 'foo' longjmps, it first unwinds to 'catch (...)'
645	// catchswitch, and is not caught by that catchswitch because it is a longjmp,
646	// then it should next unwind to catch.dispatch.longjmp BB. But if this 'catch
647	// (...)' catchswitch -> catch.dispatch.longjmp unwind relationship is lost,
648	// it will not unwind to catch.dispatch.longjmp, producing an incorrect
649	// result.
650	//
651	// Every catchpad generated by Wasm C++ contains __cxa_end_catch, so we
652	// intentionally treat it as longjmpable to work around this problem. This is
653	// a hacky fix but an easy one.
654	//
655	// The comment block in findWasmUnwindDestinations() in
656	// SelectionDAGBuilder.cpp is addressing a similar problem.
657	if (CalleeName == "__cxa_end_catch")
658	return WebAssembly::WasmEnableSjLj;
659	if (CalleeName == "__cxa_begin_catch" ||
660	CalleeName == "__cxa_allocate_exception" || CalleeName == "__cxa_throw" ||
661	CalleeName == "__clang_call_terminate")
662	return false;
663
664	// std::terminate, which is generated when another exception occurs while
665	// handling an exception, cannot longjmp.
666	if (CalleeName == "_ZSt9terminatev")
667	return false;
668
669	// Otherwise we don't know
670	return true;
671	}
672
673	static bool isEmAsmCall(const Value *Callee) {
674	StringRef CalleeName = Callee->getName();
675	// This is an exhaustive list from Emscripten's <emscripten/em_asm.h>.
676	return CalleeName == "emscripten_asm_const_int" ||
677	CalleeName == "emscripten_asm_const_double" ||
678	CalleeName == "emscripten_asm_const_int_sync_on_main_thread" ||
679	CalleeName == "emscripten_asm_const_double_sync_on_main_thread" ||
680	CalleeName == "emscripten_asm_const_async_on_main_thread";
681	}
682
683	// Generate __wasm_setjmp_test function call seqence with preamble and
684	// postamble. The code this generates is equivalent to the following
685	// JavaScript code:
686	// %__threwValue.val = __threwValue;
687	// if (%__THREW__.val != 0 & %__threwValue.val != 0) {
688	// %label = __wasm_setjmp_test(%__THREW__.val, functionInvocationId);
689	// if (%label == 0)
690	// emscripten_longjmp(%__THREW__.val, %__threwValue.val);
691	// setTempRet0(%__threwValue.val);
692	// } else {
693	// %label = -1;
694	// }
695	// %longjmp_result = getTempRet0();
696	//
697	// As output parameters. returns %label, %longjmp_result, and the BB the last
698	// instruction (%longjmp_result = ...) is in.
699	void WebAssemblyLowerEmscriptenEHSjLj::wrapTestSetjmp(
700	BasicBlock *BB, DebugLoc DL, Value *Threw, Value *FunctionInvocationId,
701	Value *&Label, Value *&LongjmpResult, BasicBlock *&CallEmLongjmpBB,
702	PHINode *&CallEmLongjmpBBThrewPHI, PHINode *&CallEmLongjmpBBThrewValuePHI,
703	BasicBlock *&EndBB) {
704	Function *F = BB->getParent();
705	Module *M = F->getParent();
706	LLVMContext &C = M->getContext();
707	IRBuilder<> IRB(C);
708	IRB.SetCurrentDebugLocation(DL);
709
710	// if (%__THREW__.val != 0 & %__threwValue.val != 0)
711	IRB.SetInsertPoint(BB);
712	BasicBlock *ThenBB1 = BasicBlock::Create(Context&: C, Name: "if.then1", Parent: F);
713	BasicBlock *ElseBB1 = BasicBlock::Create(Context&: C, Name: "if.else1", Parent: F);
714	BasicBlock *EndBB1 = BasicBlock::Create(Context&: C, Name: "if.end", Parent: F);
715	Value *ThrewCmp = IRB.CreateICmpNE(LHS: Threw, RHS: getAddrSizeInt(M, C: 0));
716	Value *ThrewValue = IRB.CreateLoad(Ty: IRB.getInt32Ty(), Ptr: ThrewValueGV,
717	Name: ThrewValueGV->getName() + ".val");
718	Value *ThrewValueCmp = IRB.CreateICmpNE(LHS: ThrewValue, RHS: IRB.getInt32(C: 0));
719	Value *Cmp1 = IRB.CreateAnd(LHS: ThrewCmp, RHS: ThrewValueCmp, Name: "cmp1");
720	IRB.CreateCondBr(Cond: Cmp1, True: ThenBB1, False: ElseBB1);
721
722	// Generate call.em.longjmp BB once and share it within the function
723	if (!CallEmLongjmpBB) {
724	// emscripten_longjmp(%__THREW__.val, %__threwValue.val);
725	CallEmLongjmpBB = BasicBlock::Create(Context&: C, Name: "call.em.longjmp", Parent: F);
726	IRB.SetInsertPoint(CallEmLongjmpBB);
727	CallEmLongjmpBBThrewPHI = IRB.CreatePHI(Ty: getAddrIntType(M), NumReservedValues: 4, Name: "threw.phi");
728	CallEmLongjmpBBThrewValuePHI =
729	IRB.CreatePHI(Ty: IRB.getInt32Ty(), NumReservedValues: 4, Name: "threwvalue.phi");
730	CallEmLongjmpBBThrewPHI->addIncoming(V: Threw, BB: ThenBB1);
731	CallEmLongjmpBBThrewValuePHI->addIncoming(V: ThrewValue, BB: ThenBB1);
732	IRB.CreateCall(Callee: EmLongjmpF,
733	Args: {CallEmLongjmpBBThrewPHI, CallEmLongjmpBBThrewValuePHI});
734	IRB.CreateUnreachable();
735	} else {
736	CallEmLongjmpBBThrewPHI->addIncoming(V: Threw, BB: ThenBB1);
737	CallEmLongjmpBBThrewValuePHI->addIncoming(V: ThrewValue, BB: ThenBB1);
738	}
739
740	// %label = __wasm_setjmp_test(%__THREW__.val, functionInvocationId);
741	// if (%label == 0)
742	IRB.SetInsertPoint(ThenBB1);
743	BasicBlock *EndBB2 = BasicBlock::Create(Context&: C, Name: "if.end2", Parent: F);
744	Value *ThrewPtr =
745	IRB.CreateIntToPtr(V: Threw, DestTy: getAddrPtrType(M), Name: Threw->getName() + ".p");
746	Value *ThenLabel = IRB.CreateCall(Callee: WasmSetjmpTestF,
747	Args: {ThrewPtr, FunctionInvocationId}, Name: "label");
748	Value *Cmp2 = IRB.CreateICmpEQ(LHS: ThenLabel, RHS: IRB.getInt32(C: 0));
749	IRB.CreateCondBr(Cond: Cmp2, True: CallEmLongjmpBB, False: EndBB2);
750
751	// setTempRet0(%__threwValue.val);
752	IRB.SetInsertPoint(EndBB2);
753	IRB.CreateCall(Callee: SetTempRet0F, Args: ThrewValue);
754	IRB.CreateBr(Dest: EndBB1);
755
756	IRB.SetInsertPoint(ElseBB1);
757	IRB.CreateBr(Dest: EndBB1);
758
759	// longjmp_result = getTempRet0();
760	IRB.SetInsertPoint(EndBB1);
761	PHINode *LabelPHI = IRB.CreatePHI(Ty: IRB.getInt32Ty(), NumReservedValues: 2, Name: "label");
762	LabelPHI->addIncoming(V: ThenLabel, BB: EndBB2);
763
764	LabelPHI->addIncoming(V: IRB.getInt32(C: -1), BB: ElseBB1);
765
766	// Output parameter assignment
767	Label = LabelPHI;
768	EndBB = EndBB1;
769	LongjmpResult = IRB.CreateCall(Callee: GetTempRet0F, Args: std::nullopt, Name: "longjmp_result");
770	}
771
772	void WebAssemblyLowerEmscriptenEHSjLj::rebuildSSA(Function &F) {
773	DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
774	DT.recalculate(Func&: F); // CFG has been changed
775
776	SSAUpdaterBulk SSA;
777	for (BasicBlock &BB : F) {
778	for (Instruction &I : BB) {
779	unsigned VarID = SSA.AddVariable(Name: I.getName(), Ty: I.getType());
780	// If a value is defined by an invoke instruction, it is only available in
781	// its normal destination and not in its unwind destination.
782	if (auto *II = dyn_cast<InvokeInst>(Val: &I))
783	SSA.AddAvailableValue(Var: VarID, BB: II->getNormalDest(), V: II);
784	else
785	SSA.AddAvailableValue(Var: VarID, BB: &BB, V: &I);
786	for (auto &U : I.uses()) {
787	auto *User = cast<Instruction>(Val: U.getUser());
788	if (auto *UserPN = dyn_cast<PHINode>(Val: User))
789	if (UserPN->getIncomingBlock(U) == &BB)
790	continue;
791	if (DT.dominates(Def: &I, User))
792	continue;
793	SSA.AddUse(Var: VarID, U: &U);
794	}
795	}
796	}
797	SSA.RewriteAllUses(DT: &DT);
798	}
799
800	// Replace uses of longjmp with a new longjmp function in Emscripten library.
801	// In Emscripten SjLj, the new function is
802	// void emscripten_longjmp(uintptr_t, i32)
803	// In Wasm SjLj, the new function is
804	// void __wasm_longjmp(i8*, i32)
805	// Because the original libc longjmp function takes (jmp_buf*, i32), we need a
806	// ptrtoint/bitcast instruction here to make the type match. jmp_buf* will
807	// eventually be lowered to i32/i64 in the wasm backend.
808	void WebAssemblyLowerEmscriptenEHSjLj::replaceLongjmpWith(Function *LongjmpF,
809	Function *NewF) {
810	assert(NewF == EmLongjmpF || NewF == WasmLongjmpF);
811	Module *M = LongjmpF->getParent();
812	SmallVector<CallInst *, 8> ToErase;
813	LLVMContext &C = LongjmpF->getParent()->getContext();
814	IRBuilder<> IRB(C);
815
816	// For calls to longjmp, replace it with emscripten_longjmp/__wasm_longjmp and
817	// cast its first argument (jmp_buf*) appropriately
818	for (User *U : LongjmpF->users()) {
819	auto *CI = dyn_cast<CallInst>(Val: U);
820	if (CI && CI->getCalledFunction() == LongjmpF) {
821	IRB.SetInsertPoint(CI);
822	Value *Env = nullptr;
823	if (NewF == EmLongjmpF)
824	Env =
825	IRB.CreatePtrToInt(V: CI->getArgOperand(i: 0), DestTy: getAddrIntType(M), Name: "env");
826	else // WasmLongjmpF
827	Env = IRB.CreateBitCast(V: CI->getArgOperand(i: 0), DestTy: IRB.getPtrTy(), Name: "env");
828	IRB.CreateCall(Callee: NewF, Args: {Env, CI->getArgOperand(i: 1)});
829	ToErase.push_back(Elt: CI);
830	}
831	}
832	for (auto *I : ToErase)
833	I->eraseFromParent();
834
835	// If we have any remaining uses of longjmp's function pointer, replace it
836	// with (void(*)(jmp_buf*, int))emscripten_longjmp / __wasm_longjmp.
837	if (!LongjmpF->uses().empty()) {
838	Value *NewLongjmp =
839	IRB.CreateBitCast(V: NewF, DestTy: LongjmpF->getType(), Name: "longjmp.cast");
840	LongjmpF->replaceAllUsesWith(V: NewLongjmp);
841	}
842	}
843
844	static bool containsLongjmpableCalls(const Function *F) {
845	for (const auto &BB : *F)
846	for (const auto &I : BB)
847	if (const auto *CB = dyn_cast<CallBase>(Val: &I))
848	if (canLongjmp(Callee: CB->getCalledOperand()))
849	return true;
850	return false;
851	}
852
853	// When a function contains a setjmp call but not other calls that can longjmp,
854	// we don't do setjmp transformation for that setjmp. But we need to convert the
855	// setjmp calls into "i32 0" so they don't cause link time errors. setjmp always
856	// returns 0 when called directly.
857	static void nullifySetjmp(Function *F) {
858	Module &M = *F->getParent();
859	IRBuilder<> IRB(M.getContext());
860	Function *SetjmpF = M.getFunction(Name: "setjmp");
861	SmallVector<Instruction *, 1> ToErase;
862
863	for (User *U : make_early_inc_range(Range: SetjmpF->users())) {
864	auto *CB = cast<CallBase>(Val: U);
865	BasicBlock *BB = CB->getParent();
866	if (BB->getParent() != F) // in other function
867	continue;
868	CallInst *CI = nullptr;
869	// setjmp cannot throw. So if it is an invoke, lower it to a call
870	if (auto *II = dyn_cast<InvokeInst>(Val: CB))
871	CI = llvm::changeToCall(II);
872	else
873	CI = cast<CallInst>(Val: CB);
874	ToErase.push_back(Elt: CI);
875	CI->replaceAllUsesWith(V: IRB.getInt32(C: 0));
876	}
877	for (auto *I : ToErase)
878	I->eraseFromParent();
879	}
880
881	bool WebAssemblyLowerEmscriptenEHSjLj::runOnModule(Module &M) {
882	LLVM_DEBUG(dbgs() << "********** Lower Emscripten EH & SjLj **********\n");
883
884	LLVMContext &C = M.getContext();
885	IRBuilder<> IRB(C);
886
887	Function *SetjmpF = M.getFunction(Name: "setjmp");
888	Function *LongjmpF = M.getFunction(Name: "longjmp");
889
890	// In some platforms _setjmp and _longjmp are used instead. Change these to
891	// use setjmp/longjmp instead, because we later detect these functions by
892	// their names.
893	Function *SetjmpF2 = M.getFunction(Name: "_setjmp");
894	Function *LongjmpF2 = M.getFunction(Name: "_longjmp");
895	if (SetjmpF2) {
896	if (SetjmpF) {
897	if (SetjmpF->getFunctionType() != SetjmpF2->getFunctionType())
898	report_fatal_error(reason: "setjmp and _setjmp have different function types");
899	} else {
900	SetjmpF = Function::Create(Ty: SetjmpF2->getFunctionType(),
901	Linkage: GlobalValue::ExternalLinkage, N: "setjmp", M);
902	}
903	SetjmpF2->replaceAllUsesWith(V: SetjmpF);
904	}
905	if (LongjmpF2) {
906	if (LongjmpF) {
907	if (LongjmpF->getFunctionType() != LongjmpF2->getFunctionType())
908	report_fatal_error(
909	reason: "longjmp and _longjmp have different function types");
910	} else {
911	LongjmpF = Function::Create(Ty: LongjmpF2->getFunctionType(),
912	Linkage: GlobalValue::ExternalLinkage, N: "setjmp", M);
913	}
914	LongjmpF2->replaceAllUsesWith(V: LongjmpF);
915	}
916
917	auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
918	assert(TPC && "Expected a TargetPassConfig");
919	auto &TM = TPC->getTM<WebAssemblyTargetMachine>();
920
921	// Declare (or get) global variables __THREW__, __threwValue, and
922	// getTempRet0/setTempRet0 function which are used in common for both
923	// exception handling and setjmp/longjmp handling
924	ThrewGV = getGlobalVariable(M, Ty: getAddrIntType(M: &M), TM, Name: "__THREW__");
925	ThrewValueGV = getGlobalVariable(M, Ty: IRB.getInt32Ty(), TM, Name: "__threwValue");
926	GetTempRet0F = getEmscriptenFunction(
927	Ty: FunctionType::get(Result: IRB.getInt32Ty(), isVarArg: false), Name: "getTempRet0", M: &M);
928	SetTempRet0F = getEmscriptenFunction(
929	Ty: FunctionType::get(Result: IRB.getVoidTy(), Params: IRB.getInt32Ty(), isVarArg: false),
930	Name: "setTempRet0", M: &M);
931	GetTempRet0F->setDoesNotThrow();
932	SetTempRet0F->setDoesNotThrow();
933
934	bool Changed = false;
935
936	// Function registration for exception handling
937	if (EnableEmEH) {
938	// Register __resumeException function
939	FunctionType *ResumeFTy =
940	FunctionType::get(Result: IRB.getVoidTy(), Params: IRB.getPtrTy(), isVarArg: false);
941	ResumeF = getEmscriptenFunction(Ty: ResumeFTy, Name: "__resumeException", M: &M);
942	ResumeF->addFnAttr(Attribute::NoReturn);
943
944	// Register llvm_eh_typeid_for function
945	FunctionType *EHTypeIDTy =
946	FunctionType::get(Result: IRB.getInt32Ty(), Params: IRB.getPtrTy(), isVarArg: false);
947	EHTypeIDF = getEmscriptenFunction(Ty: EHTypeIDTy, Name: "llvm_eh_typeid_for", M: &M);
948	}
949
950	// Functions that contains calls to setjmp but don't have other longjmpable
951	// calls within them.
952	SmallPtrSet<Function *, 4> SetjmpUsersToNullify;
953
954	if ((EnableEmSjLj || EnableWasmSjLj) && SetjmpF) {
955	// Precompute setjmp users
956	for (User *U : SetjmpF->users()) {
957	if (auto *CB = dyn_cast<CallBase>(Val: U)) {
958	auto *UserF = CB->getFunction();
959	// If a function that calls setjmp does not contain any other calls that
960	// can longjmp, we don't need to do any transformation on that function,
961	// so can ignore it
962	if (containsLongjmpableCalls(F: UserF))
963	SetjmpUsers.insert(Ptr: UserF);
964	else
965	SetjmpUsersToNullify.insert(Ptr: UserF);
966	} else {
967	std::string S;
968	raw_string_ostream SS(S);
969	SS << *U;
970	report_fatal_error(reason: Twine("Indirect use of setjmp is not supported: ") +
971	SS.str());
972	}
973	}
974	}
975
976	bool SetjmpUsed = SetjmpF && !SetjmpUsers.empty();
977	bool LongjmpUsed = LongjmpF && !LongjmpF->use_empty();
978	DoSjLj = (EnableEmSjLj | EnableWasmSjLj) && (SetjmpUsed || LongjmpUsed);
979
980	// Function registration and data pre-gathering for setjmp/longjmp handling
981	if (DoSjLj) {
982	assert(EnableEmSjLj || EnableWasmSjLj);
983	if (EnableEmSjLj) {
984	// Register emscripten_longjmp function
985	FunctionType *FTy = FunctionType::get(
986	Result: IRB.getVoidTy(), Params: {getAddrIntType(M: &M), IRB.getInt32Ty()}, isVarArg: false);
987	EmLongjmpF = getEmscriptenFunction(Ty: FTy, Name: "emscripten_longjmp", M: &M);
988	EmLongjmpF->addFnAttr(Attribute::NoReturn);
989	} else { // EnableWasmSjLj
990	Type *Int8PtrTy = IRB.getPtrTy();
991	// Register __wasm_longjmp function, which calls __builtin_wasm_longjmp.
992	FunctionType *FTy = FunctionType::get(
993	Result: IRB.getVoidTy(), Params: {Int8PtrTy, IRB.getInt32Ty()}, isVarArg: false);
994	WasmLongjmpF = getEmscriptenFunction(Ty: FTy, Name: "__wasm_longjmp", M: &M);
995	WasmLongjmpF->addFnAttr(Attribute::NoReturn);
996	}
997
998	if (SetjmpF) {
999	Type *Int8PtrTy = IRB.getPtrTy();
1000	Type *Int32PtrTy = IRB.getPtrTy();
1001	Type *Int32Ty = IRB.getInt32Ty();
1002
1003	// Register __wasm_setjmp function
1004	FunctionType *SetjmpFTy = SetjmpF->getFunctionType();
1005	FunctionType *FTy = FunctionType::get(
1006	Result: IRB.getVoidTy(), Params: {SetjmpFTy->getParamType(i: 0), Int32Ty, Int32PtrTy},
1007	isVarArg: false);
1008	WasmSetjmpF = getEmscriptenFunction(Ty: FTy, Name: "__wasm_setjmp", M: &M);
1009
1010	// Register __wasm_setjmp_test function
1011	FTy = FunctionType::get(Result: Int32Ty, Params: {Int32PtrTy, Int32PtrTy}, isVarArg: false);
1012	WasmSetjmpTestF = getEmscriptenFunction(Ty: FTy, Name: "__wasm_setjmp_test", M: &M);
1013
1014	// wasm.catch() will be lowered down to wasm 'catch' instruction in
1015	// instruction selection.
1016	CatchF = Intrinsic::getDeclaration(M: &M, Intrinsic::id: wasm_catch);
1017	// Type for struct __WasmLongjmpArgs
1018	LongjmpArgsTy = StructType::get(elt1: Int8PtrTy, // env
1019	elts: Int32Ty // val
1020	);
1021	}
1022	}
1023
1024	// Exception handling transformation
1025	if (EnableEmEH) {
1026	for (Function &F : M) {
1027	if (F.isDeclaration())
1028	continue;
1029	Changed |= runEHOnFunction(F);
1030	}
1031	}
1032
1033	// Setjmp/longjmp handling transformation
1034	if (DoSjLj) {
1035	Changed = true; // We have setjmp or longjmp somewhere
1036	if (LongjmpF)
1037	replaceLongjmpWith(LongjmpF, NewF: EnableEmSjLj ? EmLongjmpF : WasmLongjmpF);
1038	// Only traverse functions that uses setjmp in order not to insert
1039	// unnecessary prep / cleanup code in every function
1040	if (SetjmpF)
1041	for (Function *F : SetjmpUsers)
1042	runSjLjOnFunction(F&: *F);
1043	}
1044
1045	// Replace unnecessary setjmp calls with 0
1046	if ((EnableEmSjLj || EnableWasmSjLj) && !SetjmpUsersToNullify.empty()) {
1047	Changed = true;
1048	assert(SetjmpF);
1049	for (Function *F : SetjmpUsersToNullify)
1050	nullifySetjmp(F);
1051	}
1052
1053	// Delete unused global variables and functions
1054	for (auto *V : {ThrewGV, ThrewValueGV})
1055	if (V && V->use_empty())
1056	V->eraseFromParent();
1057	for (auto *V : {GetTempRet0F, SetTempRet0F, ResumeF, EHTypeIDF, EmLongjmpF,
1058	WasmSetjmpF, WasmSetjmpTestF, WasmLongjmpF, CatchF})
1059	if (V && V->use_empty())
1060	V->eraseFromParent();
1061
1062	return Changed;
1063	}
1064
1065	bool WebAssemblyLowerEmscriptenEHSjLj::runEHOnFunction(Function &F) {
1066	Module &M = *F.getParent();
1067	LLVMContext &C = F.getContext();
1068	IRBuilder<> IRB(C);
1069	bool Changed = false;
1070	SmallVector<Instruction *, 64> ToErase;
1071	SmallPtrSet<LandingPadInst *, 32> LandingPads;
1072
1073	// rethrow.longjmp BB that will be shared within the function.
1074	BasicBlock *RethrowLongjmpBB = nullptr;
1075	// PHI node for the loaded value of __THREW__ global variable in
1076	// rethrow.longjmp BB
1077	PHINode *RethrowLongjmpBBThrewPHI = nullptr;
1078
1079	for (BasicBlock &BB : F) {
1080	auto *II = dyn_cast<InvokeInst>(Val: BB.getTerminator());
1081	if (!II)
1082	continue;
1083	Changed = true;
1084	LandingPads.insert(Ptr: II->getLandingPadInst());
1085	IRB.SetInsertPoint(II);
1086
1087	const Value *Callee = II->getCalledOperand();
1088	bool NeedInvoke = supportsException(F: &F) && canThrow(V: Callee);
1089	if (NeedInvoke) {
1090	// Wrap invoke with invoke wrapper and generate preamble/postamble
1091	Value *Threw = wrapInvoke(CI: II);
1092	ToErase.push_back(Elt: II);
1093
1094	// If setjmp/longjmp handling is enabled, the thrown value can be not an
1095	// exception but a longjmp. If the current function contains calls to
1096	// setjmp, it will be appropriately handled in runSjLjOnFunction. But even
1097	// if the function does not contain setjmp calls, we shouldn't silently
1098	// ignore longjmps; we should rethrow them so they can be correctly
1099	// handled in somewhere up the call chain where setjmp is. __THREW__'s
1100	// value is 0 when nothing happened, 1 when an exception is thrown, and
1101	// other values when longjmp is thrown.
1102	//
1103	// if (%__THREW__.val == 0 || %__THREW__.val == 1)
1104	// goto %tail
1105	// else
1106	// goto %longjmp.rethrow
1107	//
1108	// rethrow.longjmp: ;; This is longjmp. Rethrow it
1109	// %__threwValue.val = __threwValue
1110	// emscripten_longjmp(%__THREW__.val, %__threwValue.val);
1111	//
1112	// tail: ;; Nothing happened or an exception is thrown
1113	// ... Continue exception handling ...
1114	if (DoSjLj && EnableEmSjLj && !SetjmpUsers.count(Ptr: &F) &&
1115	canLongjmp(Callee)) {
1116	// Create longjmp.rethrow BB once and share it within the function
1117	if (!RethrowLongjmpBB) {
1118	RethrowLongjmpBB = BasicBlock::Create(Context&: C, Name: "rethrow.longjmp", Parent: &F);
1119	IRB.SetInsertPoint(RethrowLongjmpBB);
1120	RethrowLongjmpBBThrewPHI =
1121	IRB.CreatePHI(Ty: getAddrIntType(M: &M), NumReservedValues: 4, Name: "threw.phi");
1122	RethrowLongjmpBBThrewPHI->addIncoming(V: Threw, BB: &BB);
1123	Value *ThrewValue = IRB.CreateLoad(Ty: IRB.getInt32Ty(), Ptr: ThrewValueGV,
1124	Name: ThrewValueGV->getName() + ".val");
1125	IRB.CreateCall(Callee: EmLongjmpF, Args: {RethrowLongjmpBBThrewPHI, ThrewValue});
1126	IRB.CreateUnreachable();
1127	} else {
1128	RethrowLongjmpBBThrewPHI->addIncoming(V: Threw, BB: &BB);
1129	}
1130
1131	IRB.SetInsertPoint(II); // Restore the insert point back
1132	BasicBlock *Tail = BasicBlock::Create(Context&: C, Name: "tail", Parent: &F);
1133	Value *CmpEqOne =
1134	IRB.CreateICmpEQ(LHS: Threw, RHS: getAddrSizeInt(M: &M, C: 1), Name: "cmp.eq.one");
1135	Value *CmpEqZero =
1136	IRB.CreateICmpEQ(LHS: Threw, RHS: getAddrSizeInt(M: &M, C: 0), Name: "cmp.eq.zero");
1137	Value *Or = IRB.CreateOr(LHS: CmpEqZero, RHS: CmpEqOne, Name: "or");
1138	IRB.CreateCondBr(Cond: Or, True: Tail, False: RethrowLongjmpBB);
1139	IRB.SetInsertPoint(Tail);
1140	BB.replaceSuccessorsPhiUsesWith(Old: &BB, New: Tail);
1141	}
1142
1143	// Insert a branch based on __THREW__ variable
1144	Value *Cmp = IRB.CreateICmpEQ(LHS: Threw, RHS: getAddrSizeInt(M: &M, C: 1), Name: "cmp");
1145	IRB.CreateCondBr(Cond: Cmp, True: II->getUnwindDest(), False: II->getNormalDest());
1146
1147	} else {
1148	// This can't throw, and we don't need this invoke, just replace it with a
1149	// call+branch
1150	changeToCall(II);
1151	}
1152	}
1153
1154	// Process resume instructions
1155	for (BasicBlock &BB : F) {
1156	// Scan the body of the basic block for resumes
1157	for (Instruction &I : BB) {
1158	auto *RI = dyn_cast<ResumeInst>(Val: &I);
1159	if (!RI)
1160	continue;
1161	Changed = true;
1162
1163	// Split the input into legal values
1164	Value *Input = RI->getValue();
1165	IRB.SetInsertPoint(RI);
1166	Value *Low = IRB.CreateExtractValue(Agg: Input, Idxs: 0, Name: "low");
1167	// Create a call to __resumeException function
1168	IRB.CreateCall(Callee: ResumeF, Args: {Low});
1169	// Add a terminator to the block
1170	IRB.CreateUnreachable();
1171	ToErase.push_back(Elt: RI);
1172	}
1173	}
1174
1175	// Process llvm.eh.typeid.for intrinsics
1176	for (BasicBlock &BB : F) {
1177	for (Instruction &I : BB) {
1178	auto *CI = dyn_cast<CallInst>(Val: &I);
1179	if (!CI)
1180	continue;
1181	const Function *Callee = CI->getCalledFunction();
1182	if (!Callee)
1183	continue;
1184	if (Callee->getIntrinsicID() != Intrinsic::eh_typeid_for)
1185	continue;
1186	Changed = true;
1187
1188	IRB.SetInsertPoint(CI);
1189	CallInst *NewCI =
1190	IRB.CreateCall(Callee: EHTypeIDF, Args: CI->getArgOperand(i: 0), Name: "typeid");
1191	CI->replaceAllUsesWith(V: NewCI);
1192	ToErase.push_back(Elt: CI);
1193	}
1194	}
1195
1196	// Look for orphan landingpads, can occur in blocks with no predecessors
1197	for (BasicBlock &BB : F) {
1198	Instruction *I = BB.getFirstNonPHI();
1199	if (auto *LPI = dyn_cast<LandingPadInst>(Val: I))
1200	LandingPads.insert(Ptr: LPI);
1201	}
1202	Changed |= !LandingPads.empty();
1203
1204	// Handle all the landingpad for this function together, as multiple invokes
1205	// may share a single lp
1206	for (LandingPadInst *LPI : LandingPads) {
1207	IRB.SetInsertPoint(LPI);
1208	SmallVector<Value *, 16> FMCArgs;
1209	for (unsigned I = 0, E = LPI->getNumClauses(); I < E; ++I) {
1210	Constant *Clause = LPI->getClause(Idx: I);
1211	// TODO Handle filters (= exception specifications).
1212	// https://github.com/llvm/llvm-project/issues/49740
1213	if (LPI->isCatch(Idx: I))
1214	FMCArgs.push_back(Elt: Clause);
1215	}
1216
1217	// Create a call to __cxa_find_matching_catch_N function
1218	Function *FMCF = getFindMatchingCatch(M, NumClauses: FMCArgs.size());
1219	CallInst *FMCI = IRB.CreateCall(Callee: FMCF, Args: FMCArgs, Name: "fmc");
1220	Value *Poison = PoisonValue::get(T: LPI->getType());
1221	Value *Pair0 = IRB.CreateInsertValue(Agg: Poison, Val: FMCI, Idxs: 0, Name: "pair0");
1222	Value *TempRet0 = IRB.CreateCall(Callee: GetTempRet0F, Args: std::nullopt, Name: "tempret0");
1223	Value *Pair1 = IRB.CreateInsertValue(Agg: Pair0, Val: TempRet0, Idxs: 1, Name: "pair1");
1224
1225	LPI->replaceAllUsesWith(V: Pair1);
1226	ToErase.push_back(Elt: LPI);
1227	}
1228
1229	// Erase everything we no longer need in this function
1230	for (Instruction *I : ToErase)
1231	I->eraseFromParent();
1232
1233	return Changed;
1234	}
1235
1236	// This tries to get debug info from the instruction before which a new
1237	// instruction will be inserted, and if there's no debug info in that
1238	// instruction, tries to get the info instead from the previous instruction (if
1239	// any). If none of these has debug info and a DISubprogram is provided, it
1240	// creates a dummy debug info with the first line of the function, because IR
1241	// verifier requires all inlinable callsites should have debug info when both a
1242	// caller and callee have DISubprogram. If none of these conditions are met,
1243	// returns empty info.
1244	static DebugLoc getOrCreateDebugLoc(const Instruction *InsertBefore,
1245	DISubprogram *SP) {
1246	assert(InsertBefore);
1247	if (InsertBefore->getDebugLoc())
1248	return InsertBefore->getDebugLoc();
1249	const Instruction *Prev = InsertBefore->getPrevNode();
1250	if (Prev && Prev->getDebugLoc())
1251	return Prev->getDebugLoc();
1252	if (SP)
1253	return DILocation::get(Context&: SP->getContext(), Line: SP->getLine(), Column: 1, Scope: SP);
1254	return DebugLoc();
1255	}
1256
1257	bool WebAssemblyLowerEmscriptenEHSjLj::runSjLjOnFunction(Function &F) {
1258	assert(EnableEmSjLj || EnableWasmSjLj);
1259	Module &M = *F.getParent();
1260	LLVMContext &C = F.getContext();
1261	IRBuilder<> IRB(C);
1262	SmallVector<Instruction *, 64> ToErase;
1263
1264	// Setjmp preparation
1265
1266	BasicBlock *Entry = &F.getEntryBlock();
1267	DebugLoc FirstDL = getOrCreateDebugLoc(InsertBefore: &*Entry->begin(), SP: F.getSubprogram());
1268	SplitBlock(Old: Entry, SplitPt: &*Entry->getFirstInsertionPt());
1269
1270	IRB.SetInsertPoint(Entry->getTerminator()->getIterator());
1271	// This alloca'ed pointer is used by the runtime to identify function
1272	// invocations. It's just for pointer comparisons. It will never be
1273	// dereferenced.
1274	Instruction *FunctionInvocationId =
1275	IRB.CreateAlloca(Ty: IRB.getInt32Ty(), ArraySize: nullptr, Name: "functionInvocationId");
1276	FunctionInvocationId->setDebugLoc(FirstDL);
1277
1278	// Setjmp transformation
1279	SmallVector<PHINode *, 4> SetjmpRetPHIs;
1280	Function *SetjmpF = M.getFunction(Name: "setjmp");
1281	for (auto *U : make_early_inc_range(Range: SetjmpF->users())) {
1282	auto *CB = cast<CallBase>(Val: U);
1283	BasicBlock *BB = CB->getParent();
1284	if (BB->getParent() != &F) // in other function
1285	continue;
1286	if (CB->getOperandBundle(ID: LLVMContext::OB_funclet)) {
1287	std::string S;
1288	raw_string_ostream SS(S);
1289	SS << "In function " + F.getName() +
1290	": setjmp within a catch clause is not supported in Wasm EH:\n";
1291	SS << *CB;
1292	report_fatal_error(reason: StringRef(SS.str()));
1293	}
1294
1295	CallInst *CI = nullptr;
1296	// setjmp cannot throw. So if it is an invoke, lower it to a call
1297	if (auto *II = dyn_cast<InvokeInst>(Val: CB))
1298	CI = llvm::changeToCall(II);
1299	else
1300	CI = cast<CallInst>(Val: CB);
1301
1302	// The tail is everything right after the call, and will be reached once
1303	// when setjmp is called, and later when longjmp returns to the setjmp
1304	BasicBlock *Tail = SplitBlock(Old: BB, SplitPt: CI->getNextNode());
1305	// Add a phi to the tail, which will be the output of setjmp, which
1306	// indicates if this is the first call or a longjmp back. The phi directly
1307	// uses the right value based on where we arrive from
1308	IRB.SetInsertPoint(TheBB: Tail, IP: Tail->getFirstNonPHIIt());
1309	PHINode *SetjmpRet = IRB.CreatePHI(Ty: IRB.getInt32Ty(), NumReservedValues: 2, Name: "setjmp.ret");
1310
1311	// setjmp initial call returns 0
1312	SetjmpRet->addIncoming(V: IRB.getInt32(C: 0), BB);
1313	// The proper output is now this, not the setjmp call itself
1314	CI->replaceAllUsesWith(V: SetjmpRet);
1315	// longjmp returns to the setjmp will add themselves to this phi
1316	SetjmpRetPHIs.push_back(Elt: SetjmpRet);
1317
1318	// Fix call target
1319	// Our index in the function is our place in the array + 1 to avoid index
1320	// 0, because index 0 means the longjmp is not ours to handle.
1321	IRB.SetInsertPoint(CI);
1322	Value *Args[] = {CI->getArgOperand(i: 0), IRB.getInt32(C: SetjmpRetPHIs.size()),
1323	FunctionInvocationId};
1324	IRB.CreateCall(Callee: WasmSetjmpF, Args);
1325	ToErase.push_back(Elt: CI);
1326	}
1327
1328	// Handle longjmpable calls.
1329	if (EnableEmSjLj)
1330	handleLongjmpableCallsForEmscriptenSjLj(F, FunctionInvocationId,
1331	SetjmpRetPHIs);
1332	else // EnableWasmSjLj
1333	handleLongjmpableCallsForWasmSjLj(F, FunctionInvocationId, SetjmpRetPHIs);
1334
1335	// Erase everything we no longer need in this function
1336	for (Instruction *I : ToErase)
1337	I->eraseFromParent();
1338
1339	// Finally, our modifications to the cfg can break dominance of SSA variables.
1340	// For example, in this code,
1341	// if (x()) { .. setjmp() .. }
1342	// if (y()) { .. longjmp() .. }
1343	// We must split the longjmp block, and it can jump into the block splitted
1344	// from setjmp one. But that means that when we split the setjmp block, it's
1345	// first part no longer dominates its second part - there is a theoretically
1346	// possible control flow path where x() is false, then y() is true and we
1347	// reach the second part of the setjmp block, without ever reaching the first
1348	// part. So, we rebuild SSA form here.
1349	rebuildSSA(F);
1350	return true;
1351	}
1352
1353	// Update each call that can longjmp so it can return to the corresponding
1354	// setjmp. Refer to 4) of "Emscripten setjmp/longjmp handling" section in the
1355	// comments at top of the file for details.
1356	void WebAssemblyLowerEmscriptenEHSjLj::handleLongjmpableCallsForEmscriptenSjLj(
1357	Function &F, Instruction *FunctionInvocationId,
1358	SmallVectorImpl<PHINode *> &SetjmpRetPHIs) {
1359	Module &M = *F.getParent();
1360	LLVMContext &C = F.getContext();
1361	IRBuilder<> IRB(C);
1362	SmallVector<Instruction *, 64> ToErase;
1363
1364	// call.em.longjmp BB that will be shared within the function.
1365	BasicBlock *CallEmLongjmpBB = nullptr;
1366	// PHI node for the loaded value of __THREW__ global variable in
1367	// call.em.longjmp BB
1368	PHINode *CallEmLongjmpBBThrewPHI = nullptr;
1369	// PHI node for the loaded value of __threwValue global variable in
1370	// call.em.longjmp BB
1371	PHINode *CallEmLongjmpBBThrewValuePHI = nullptr;
1372	// rethrow.exn BB that will be shared within the function.
1373	BasicBlock *RethrowExnBB = nullptr;
1374
1375	// Because we are creating new BBs while processing and don't want to make
1376	// all these newly created BBs candidates again for longjmp processing, we
1377	// first make the vector of candidate BBs.
1378	std::vector<BasicBlock *> BBs;
1379	for (BasicBlock &BB : F)
1380	BBs.push_back(x: &BB);
1381
1382	// BBs.size() will change within the loop, so we query it every time
1383	for (unsigned I = 0; I < BBs.size(); I++) {
1384	BasicBlock *BB = BBs[I];
1385	for (Instruction &I : *BB) {
1386	if (isa<InvokeInst>(Val: &I)) {
1387	std::string S;
1388	raw_string_ostream SS(S);
1389	SS << "In function " << F.getName()
1390	<< ": When using Wasm EH with Emscripten SjLj, there is a "
1391	"restriction that `setjmp` function call and exception cannot be "
1392	"used within the same function:\n";
1393	SS << I;
1394	report_fatal_error(reason: StringRef(SS.str()));
1395	}
1396	auto *CI = dyn_cast<CallInst>(Val: &I);
1397	if (!CI)
1398	continue;
1399
1400	const Value *Callee = CI->getCalledOperand();
1401	if (!canLongjmp(Callee))
1402	continue;
1403	if (isEmAsmCall(Callee))
1404	report_fatal_error(reason: "Cannot use EM_ASM* alongside setjmp/longjmp in " +
1405	F.getName() +
1406	". Please consider using EM_JS, or move the "
1407	"EM_ASM into another function.",
1408	gen_crash_diag: false);
1409
1410	Value *Threw = nullptr;
1411	BasicBlock *Tail;
1412	if (Callee->getName().starts_with(Prefix: "__invoke_")) {
1413	// If invoke wrapper has already been generated for this call in
1414	// previous EH phase, search for the load instruction
1415	// %__THREW__.val = __THREW__;
1416	// in postamble after the invoke wrapper call
1417	LoadInst *ThrewLI = nullptr;
1418	StoreInst *ThrewResetSI = nullptr;
1419	for (auto I = std::next(x: BasicBlock::iterator(CI)), IE = BB->end();
1420	I != IE; ++I) {
1421	if (auto *LI = dyn_cast<LoadInst>(Val&: I))
1422	if (auto *GV = dyn_cast<GlobalVariable>(Val: LI->getPointerOperand()))
1423	if (GV == ThrewGV) {
1424	Threw = ThrewLI = LI;
1425	break;
1426	}
1427	}
1428	// Search for the store instruction after the load above
1429	// __THREW__ = 0;
1430	for (auto I = std::next(x: BasicBlock::iterator(ThrewLI)), IE = BB->end();
1431	I != IE; ++I) {
1432	if (auto *SI = dyn_cast<StoreInst>(Val&: I)) {
1433	if (auto *GV = dyn_cast<GlobalVariable>(Val: SI->getPointerOperand())) {
1434	if (GV == ThrewGV &&
1435	SI->getValueOperand() == getAddrSizeInt(M: &M, C: 0)) {
1436	ThrewResetSI = SI;
1437	break;
1438	}
1439	}
1440	}
1441	}
1442	assert(Threw && ThrewLI && "Cannot find __THREW__ load after invoke");
1443	assert(ThrewResetSI && "Cannot find __THREW__ store after invoke");
1444	Tail = SplitBlock(Old: BB, SplitPt: ThrewResetSI->getNextNode());
1445
1446	} else {
1447	// Wrap call with invoke wrapper and generate preamble/postamble
1448	Threw = wrapInvoke(CI);
1449	ToErase.push_back(Elt: CI);
1450	Tail = SplitBlock(Old: BB, SplitPt: CI->getNextNode());
1451
1452	// If exception handling is enabled, the thrown value can be not a
1453	// longjmp but an exception, in which case we shouldn't silently ignore
1454	// exceptions; we should rethrow them.
1455	// __THREW__'s value is 0 when nothing happened, 1 when an exception is
1456	// thrown, other values when longjmp is thrown.
1457	//
1458	// if (%__THREW__.val == 1)
1459	// goto %eh.rethrow
1460	// else
1461	// goto %normal
1462	//
1463	// eh.rethrow: ;; Rethrow exception
1464	// %exn = call @__cxa_find_matching_catch_2() ;; Retrieve thrown ptr
1465	// __resumeException(%exn)
1466	//
1467	// normal:
1468	// <-- Insertion point. Will insert sjlj handling code from here
1469	// goto %tail
1470	//
1471	// tail:
1472	// ...
1473	if (supportsException(F: &F) && canThrow(V: Callee)) {
1474	// We will add a new conditional branch. So remove the branch created
1475	// when we split the BB
1476	ToErase.push_back(Elt: BB->getTerminator());
1477
1478	// Generate rethrow.exn BB once and share it within the function
1479	if (!RethrowExnBB) {
1480	RethrowExnBB = BasicBlock::Create(Context&: C, Name: "rethrow.exn", Parent: &F);
1481	IRB.SetInsertPoint(RethrowExnBB);
1482	CallInst *Exn =
1483	IRB.CreateCall(Callee: getFindMatchingCatch(M, NumClauses: 0), Args: {}, Name: "exn");
1484	IRB.CreateCall(Callee: ResumeF, Args: {Exn});
1485	IRB.CreateUnreachable();
1486	}
1487
1488	IRB.SetInsertPoint(CI);
1489	BasicBlock *NormalBB = BasicBlock::Create(Context&: C, Name: "normal", Parent: &F);
1490	Value *CmpEqOne =
1491	IRB.CreateICmpEQ(LHS: Threw, RHS: getAddrSizeInt(M: &M, C: 1), Name: "cmp.eq.one");
1492	IRB.CreateCondBr(Cond: CmpEqOne, True: RethrowExnBB, False: NormalBB);
1493
1494	IRB.SetInsertPoint(NormalBB);
1495	IRB.CreateBr(Dest: Tail);
1496	BB = NormalBB; // New insertion point to insert __wasm_setjmp_test()
1497	}
1498	}
1499
1500	// We need to replace the terminator in Tail - SplitBlock makes BB go
1501	// straight to Tail, we need to check if a longjmp occurred, and go to the
1502	// right setjmp-tail if so
1503	ToErase.push_back(Elt: BB->getTerminator());
1504
1505	// Generate a function call to __wasm_setjmp_test function and
1506	// preamble/postamble code to figure out (1) whether longjmp
1507	// occurred (2) if longjmp occurred, which setjmp it corresponds to
1508	Value *Label = nullptr;
1509	Value *LongjmpResult = nullptr;
1510	BasicBlock *EndBB = nullptr;
1511	wrapTestSetjmp(BB, DL: CI->getDebugLoc(), Threw, FunctionInvocationId, Label,
1512	LongjmpResult, CallEmLongjmpBB, CallEmLongjmpBBThrewPHI,
1513	CallEmLongjmpBBThrewValuePHI, EndBB);
1514	assert(Label && LongjmpResult && EndBB);
1515
1516	// Create switch instruction
1517	IRB.SetInsertPoint(EndBB);
1518	IRB.SetCurrentDebugLocation(EndBB->back().getDebugLoc());
1519	SwitchInst *SI = IRB.CreateSwitch(V: Label, Dest: Tail, NumCases: SetjmpRetPHIs.size());
1520	// -1 means no longjmp happened, continue normally (will hit the default
1521	// switch case). 0 means a longjmp that is not ours to handle, needs a
1522	// rethrow. Otherwise the index is the same as the index in P+1 (to avoid
1523	// 0).
1524	for (unsigned I = 0; I < SetjmpRetPHIs.size(); I++) {
1525	SI->addCase(OnVal: IRB.getInt32(C: I + 1), Dest: SetjmpRetPHIs[I]->getParent());
1526	SetjmpRetPHIs[I]->addIncoming(V: LongjmpResult, BB: EndBB);
1527	}
1528
1529	// We are splitting the block here, and must continue to find other calls
1530	// in the block - which is now split. so continue to traverse in the Tail
1531	BBs.push_back(x: Tail);
1532	}
1533	}
1534
1535	for (Instruction *I : ToErase)
1536	I->eraseFromParent();
1537	}
1538
1539	static BasicBlock *getCleanupRetUnwindDest(const CleanupPadInst *CPI) {
1540	for (const User *U : CPI->users())
1541	if (const auto *CRI = dyn_cast<CleanupReturnInst>(Val: U))
1542	return CRI->getUnwindDest();
1543	return nullptr;
1544	}
1545
1546	// Create a catchpad in which we catch a longjmp's env and val arguments, test
1547	// if the longjmp corresponds to one of setjmps in the current function, and if
1548	// so, jump to the setjmp dispatch BB from which we go to one of post-setjmp
1549	// BBs. Refer to 4) of "Wasm setjmp/longjmp handling" section in the comments at
1550	// top of the file for details.
1551	void WebAssemblyLowerEmscriptenEHSjLj::handleLongjmpableCallsForWasmSjLj(
1552	Function &F, Instruction *FunctionInvocationId,
1553	SmallVectorImpl<PHINode *> &SetjmpRetPHIs) {
1554	Module &M = *F.getParent();
1555	LLVMContext &C = F.getContext();
1556	IRBuilder<> IRB(C);
1557
1558	// A function with catchswitch/catchpad instruction should have a personality
1559	// function attached to it. Search for the wasm personality function, and if
1560	// it exists, use it, and if it doesn't, create a dummy personality function.
1561	// (SjLj is not going to call it anyway.)
1562	if (!F.hasPersonalityFn()) {
1563	StringRef PersName = getEHPersonalityName(Pers: EHPersonality::Wasm_CXX);
1564	FunctionType *PersType =
1565	FunctionType::get(Result: IRB.getInt32Ty(), /* isVarArg */ true);
1566	Value *PersF = M.getOrInsertFunction(Name: PersName, T: PersType).getCallee();
1567	F.setPersonalityFn(
1568	cast<Constant>(Val: IRB.CreateBitCast(V: PersF, DestTy: IRB.getPtrTy())));
1569	}
1570
1571	// Use the entry BB's debugloc as a fallback
1572	BasicBlock *Entry = &F.getEntryBlock();
1573	DebugLoc FirstDL = getOrCreateDebugLoc(InsertBefore: &*Entry->begin(), SP: F.getSubprogram());
1574	IRB.SetCurrentDebugLocation(FirstDL);
1575
1576	// Add setjmp.dispatch BB right after the entry block. Because we have
1577	// initialized functionInvocationId in the entry block and split the
1578	// rest into another BB, here 'OrigEntry' is the function's original entry
1579	// block before the transformation.
1580	//
1581	// entry:
1582	// functionInvocationId initialization
1583	// setjmp.dispatch:
1584	// switch will be inserted here later
1585	// entry.split: (OrigEntry)
1586	// the original function starts here
1587	BasicBlock *OrigEntry = Entry->getNextNode();
1588	BasicBlock *SetjmpDispatchBB =
1589	BasicBlock::Create(Context&: C, Name: "setjmp.dispatch", Parent: &F, InsertBefore: OrigEntry);
1590	cast<BranchInst>(Val: Entry->getTerminator())->setSuccessor(idx: 0, NewSucc: SetjmpDispatchBB);
1591
1592	// Create catch.dispatch.longjmp BB and a catchswitch instruction
1593	BasicBlock *CatchDispatchLongjmpBB =
1594	BasicBlock::Create(Context&: C, Name: "catch.dispatch.longjmp", Parent: &F);
1595	IRB.SetInsertPoint(CatchDispatchLongjmpBB);
1596	CatchSwitchInst *CatchSwitchLongjmp =
1597	IRB.CreateCatchSwitch(ParentPad: ConstantTokenNone::get(Context&: C), UnwindBB: nullptr, NumHandlers: 1);
1598
1599	// Create catch.longjmp BB and a catchpad instruction
1600	BasicBlock *CatchLongjmpBB = BasicBlock::Create(Context&: C, Name: "catch.longjmp", Parent: &F);
1601	CatchSwitchLongjmp->addHandler(Dest: CatchLongjmpBB);
1602	IRB.SetInsertPoint(CatchLongjmpBB);
1603	CatchPadInst *CatchPad = IRB.CreateCatchPad(ParentPad: CatchSwitchLongjmp, Args: {});
1604
1605	// Wasm throw and catch instructions can throw and catch multiple values, but
1606	// that requires multivalue support in the toolchain, which is currently not
1607	// very reliable. We instead throw and catch a pointer to a struct value of
1608	// type 'struct __WasmLongjmpArgs', which is defined in Emscripten.
1609	Instruction *LongjmpArgs =
1610	IRB.CreateCall(Callee: CatchF, Args: {IRB.getInt32(C: WebAssembly::C_LONGJMP)}, Name: "thrown");
1611	Value *EnvField =
1612	IRB.CreateConstGEP2_32(Ty: LongjmpArgsTy, Ptr: LongjmpArgs, Idx0: 0, Idx1: 0, Name: "env_gep");
1613	Value *ValField =
1614	IRB.CreateConstGEP2_32(Ty: LongjmpArgsTy, Ptr: LongjmpArgs, Idx0: 0, Idx1: 1, Name: "val_gep");
1615	// void *env = __wasm_longjmp_args.env;
1616	Instruction *Env = IRB.CreateLoad(Ty: IRB.getPtrTy(), Ptr: EnvField, Name: "env");
1617	// int val = __wasm_longjmp_args.val;
1618	Instruction *Val = IRB.CreateLoad(Ty: IRB.getInt32Ty(), Ptr: ValField, Name: "val");
1619
1620	// %label = __wasm_setjmp_test(%env, functionInvocatinoId);
1621	// if (%label == 0)
1622	// __wasm_longjmp(%env, %val)
1623	// catchret to %setjmp.dispatch
1624	BasicBlock *ThenBB = BasicBlock::Create(Context&: C, Name: "if.then", Parent: &F);
1625	BasicBlock *EndBB = BasicBlock::Create(Context&: C, Name: "if.end", Parent: &F);
1626	Value *EnvP = IRB.CreateBitCast(V: Env, DestTy: getAddrPtrType(M: &M), Name: "env.p");
1627	Value *Label = IRB.CreateCall(Callee: WasmSetjmpTestF, Args: {EnvP, FunctionInvocationId},
1628	OpBundles: OperandBundleDef("funclet", CatchPad), Name: "label");
1629	Value *Cmp = IRB.CreateICmpEQ(LHS: Label, RHS: IRB.getInt32(C: 0));
1630	IRB.CreateCondBr(Cond: Cmp, True: ThenBB, False: EndBB);
1631
1632	IRB.SetInsertPoint(ThenBB);
1633	CallInst *WasmLongjmpCI = IRB.CreateCall(
1634	Callee: WasmLongjmpF, Args: {Env, Val}, OpBundles: OperandBundleDef("funclet", CatchPad));
1635	IRB.CreateUnreachable();
1636
1637	IRB.SetInsertPoint(EndBB);
1638	// Jump to setjmp.dispatch block
1639	IRB.CreateCatchRet(CatchPad, BB: SetjmpDispatchBB);
1640
1641	// Go back to setjmp.dispatch BB
1642	// setjmp.dispatch:
1643	// switch %label {
1644	// label 1: goto post-setjmp BB 1
1645	// label 2: goto post-setjmp BB 2
1646	// ...
1647	// default: goto splitted next BB
1648	// }
1649	IRB.SetInsertPoint(SetjmpDispatchBB);
1650	PHINode *LabelPHI = IRB.CreatePHI(Ty: IRB.getInt32Ty(), NumReservedValues: 2, Name: "label.phi");
1651	LabelPHI->addIncoming(V: Label, BB: EndBB);
1652	LabelPHI->addIncoming(V: IRB.getInt32(C: -1), BB: Entry);
1653	SwitchInst *SI = IRB.CreateSwitch(V: LabelPHI, Dest: OrigEntry, NumCases: SetjmpRetPHIs.size());
1654	// -1 means no longjmp happened, continue normally (will hit the default
1655	// switch case). 0 means a longjmp that is not ours to handle, needs a
1656	// rethrow. Otherwise the index is the same as the index in P+1 (to avoid
1657	// 0).
1658	for (unsigned I = 0; I < SetjmpRetPHIs.size(); I++) {
1659	SI->addCase(OnVal: IRB.getInt32(C: I + 1), Dest: SetjmpRetPHIs[I]->getParent());
1660	SetjmpRetPHIs[I]->addIncoming(V: Val, BB: SetjmpDispatchBB);
1661	}
1662
1663	// Convert all longjmpable call instructions to invokes that unwind to the
1664	// newly created catch.dispatch.longjmp BB.
1665	SmallVector<CallInst *, 64> LongjmpableCalls;
1666	for (auto *BB = &*F.begin(); BB; BB = BB->getNextNode()) {
1667	for (auto &I : *BB) {
1668	auto *CI = dyn_cast<CallInst>(Val: &I);
1669	if (!CI)
1670	continue;
1671	const Value *Callee = CI->getCalledOperand();
1672	if (!canLongjmp(Callee))
1673	continue;
1674	if (isEmAsmCall(Callee))
1675	report_fatal_error(reason: "Cannot use EM_ASM* alongside setjmp/longjmp in " +
1676	F.getName() +
1677	". Please consider using EM_JS, or move the "
1678	"EM_ASM into another function.",
1679	gen_crash_diag: false);
1680	// This is __wasm_longjmp() call we inserted in this function, which
1681	// rethrows the longjmp when the longjmp does not correspond to one of
1682	// setjmps in this function. We should not convert this call to an invoke.
1683	if (CI == WasmLongjmpCI)
1684	continue;
1685	LongjmpableCalls.push_back(Elt: CI);
1686	}
1687	}
1688
1689	for (auto *CI : LongjmpableCalls) {
1690	// Even if the callee function has attribute 'nounwind', which is true for
1691	// all C functions, it can longjmp, which means it can throw a Wasm
1692	// exception now.
1693	CI->removeFnAttr(Attribute::NoUnwind);
1694	if (Function *CalleeF = CI->getCalledFunction())
1695	CalleeF->removeFnAttr(Attribute::NoUnwind);
1696
1697	// Change it to an invoke and make it unwind to the catch.dispatch.longjmp
1698	// BB. If the call is enclosed in another catchpad/cleanuppad scope, unwind
1699	// to its parent pad's unwind destination instead to preserve the scope
1700	// structure. It will eventually unwind to the catch.dispatch.longjmp.
1701	SmallVector<OperandBundleDef, 1> Bundles;
1702	BasicBlock *UnwindDest = nullptr;
1703	if (auto Bundle = CI->getOperandBundle(ID: LLVMContext::OB_funclet)) {
1704	Instruction *FromPad = cast<Instruction>(Val: Bundle->Inputs[0]);
1705	while (!UnwindDest) {
1706	if (auto *CPI = dyn_cast<CatchPadInst>(Val: FromPad)) {
1707	UnwindDest = CPI->getCatchSwitch()->getUnwindDest();
1708	break;
1709	}
1710	if (auto *CPI = dyn_cast<CleanupPadInst>(Val: FromPad)) {
1711	// getCleanupRetUnwindDest() can return nullptr when
1712	// 1. This cleanuppad's matching cleanupret uwninds to caller
1713	// 2. There is no matching cleanupret because it ends with
1714	// unreachable.
1715	// In case of 2, we need to traverse the parent pad chain.
1716	UnwindDest = getCleanupRetUnwindDest(CPI);
1717	Value *ParentPad = CPI->getParentPad();
1718	if (isa<ConstantTokenNone>(Val: ParentPad))
1719	break;
1720	FromPad = cast<Instruction>(Val: ParentPad);
1721	}
1722	}
1723	}
1724	if (!UnwindDest)
1725	UnwindDest = CatchDispatchLongjmpBB;
1726	changeToInvokeAndSplitBasicBlock(CI, UnwindEdge: UnwindDest);
1727	}
1728
1729	SmallVector<Instruction *, 16> ToErase;
1730	for (auto &BB : F) {
1731	if (auto *CSI = dyn_cast<CatchSwitchInst>(Val: BB.getFirstNonPHI())) {
1732	if (CSI != CatchSwitchLongjmp && CSI->unwindsToCaller()) {
1733	IRB.SetInsertPoint(CSI);
1734	ToErase.push_back(Elt: CSI);
1735	auto *NewCSI = IRB.CreateCatchSwitch(ParentPad: CSI->getParentPad(),
1736	UnwindBB: CatchDispatchLongjmpBB, NumHandlers: 1);
1737	NewCSI->addHandler(Dest: *CSI->handler_begin());
1738	NewCSI->takeName(V: CSI);
1739	CSI->replaceAllUsesWith(V: NewCSI);
1740	}
1741	}
1742
1743	if (auto *CRI = dyn_cast<CleanupReturnInst>(Val: BB.getTerminator())) {
1744	if (CRI->unwindsToCaller()) {
1745	IRB.SetInsertPoint(CRI);
1746	ToErase.push_back(Elt: CRI);
1747	IRB.CreateCleanupRet(CleanupPad: CRI->getCleanupPad(), UnwindBB: CatchDispatchLongjmpBB);
1748	}
1749	}
1750	}
1751
1752	for (Instruction *I : ToErase)
1753	I->eraseFromParent();
1754	}
1755