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/IR/Module.h"
275#include "llvm/Support/CommandLine.h"
276#include "llvm/Transforms/Utils/BasicBlockUtils.h"
277#include "llvm/Transforms/Utils/Local.h"
278#include "llvm/Transforms/Utils/SSAUpdater.h"
279#include "llvm/Transforms/Utils/SSAUpdaterBulk.h"
280#include <set>
281
282using namespace llvm;
283
284#define DEBUG_TYPE "wasm-lower-em-ehsjlj"
285
286static cl::list<std::string>
287 EHAllowlist("emscripten-cxx-exceptions-allowed",
288 cl::desc("The list of function names in which Emscripten-style "
289 "exception handling is enabled (see emscripten "
290 "EMSCRIPTEN_CATCHING_ALLOWED options)"),
291 cl::CommaSeparated);
292
293namespace {
294class WebAssemblyLowerEmscriptenEHSjLj final : public ModulePass {
295 bool EnableEmEH; // Enable Emscripten exception handling
296 bool EnableEmSjLj; // Enable Emscripten setjmp/longjmp handling
297 bool EnableWasmSjLj; // Enable Wasm setjmp/longjmp handling
298 bool DoSjLj; // Whether we actually perform setjmp/longjmp handling
299
300 GlobalVariable *ThrewGV = nullptr; // __THREW__ (Emscripten)
301 GlobalVariable *ThrewValueGV = nullptr; // __threwValue (Emscripten)
302 Function *GetTempRet0F = nullptr; // getTempRet0() (Emscripten)
303 Function *SetTempRet0F = nullptr; // setTempRet0() (Emscripten)
304 Function *ResumeF = nullptr; // __resumeException() (Emscripten)
305 Function *EHTypeIDF = nullptr; // llvm.eh.typeid.for() (intrinsic)
306 Function *EmLongjmpF = nullptr; // emscripten_longjmp() (Emscripten)
307 Function *WasmSetjmpF = nullptr; // __wasm_setjmp() (Emscripten)
308 Function *WasmSetjmpTestF = nullptr; // __wasm_setjmp_test() (Emscripten)
309 Function *WasmLongjmpF = nullptr; // __wasm_longjmp() (Emscripten)
310 Function *CatchF = nullptr; // wasm.catch() (intrinsic)
311
312 // type of 'struct __WasmLongjmpArgs' defined in emscripten
313 Type *LongjmpArgsTy = nullptr;
314
315 // __cxa_find_matching_catch_N functions.
316 // Indexed by the number of clauses in an original landingpad instruction.
317 DenseMap<int, Function *> FindMatchingCatches;
318 // Map of <function signature string, invoke_ wrappers>
319 StringMap<Function *> InvokeWrappers;
320 // Set of allowed function names for exception handling
321 std::set<std::string, std::less<>> EHAllowlistSet;
322 // Functions that contains calls to setjmp
323 SmallPtrSet<Function *, 8> SetjmpUsers;
324
325 StringRef getPassName() const override {
326 return "WebAssembly Lower Emscripten Exceptions";
327 }
328
329 using InstVector = SmallVectorImpl<Instruction *>;
330 bool runEHOnFunction(Function &F);
331 bool runSjLjOnFunction(Function &F);
332 void handleLongjmpableCallsForEmscriptenSjLj(
333 Function &F, Instruction *FunctionInvocationId,
334 SmallVectorImpl<PHINode *> &SetjmpRetPHIs);
335 void
336 handleLongjmpableCallsForWasmSjLj(Function &F,
337 Instruction *FunctionInvocationId,
338 SmallVectorImpl<PHINode *> &SetjmpRetPHIs);
339 Function *getFindMatchingCatch(Module &M, unsigned NumClauses);
340
341 Value *wrapInvoke(CallBase *CI);
342 void wrapTestSetjmp(BasicBlock *BB, DebugLoc DL, Value *Threw,
343 Value *FunctionInvocationId, Value *&Label,
344 Value *&LongjmpResult, BasicBlock *&CallEmLongjmpBB,
345 PHINode *&CallEmLongjmpBBThrewPHI,
346 PHINode *&CallEmLongjmpBBThrewValuePHI,
347 BasicBlock *&EndBB);
348 Function *getInvokeWrapper(CallBase *CI);
349
350 bool areAllExceptionsAllowed() const { return EHAllowlistSet.empty(); }
351 bool supportsException(const Function *F) const {
352 return EnableEmEH &&
353 (areAllExceptionsAllowed() || EHAllowlistSet.count(x: F->getName()));
354 }
355 void replaceLongjmpWith(Function *LongjmpF, Function *NewF);
356
357 void rebuildSSA(Function &F);
358
359public:
360 static char ID;
361
362 WebAssemblyLowerEmscriptenEHSjLj()
363 : ModulePass(ID), EnableEmEH(WebAssembly::WasmEnableEmEH),
364 EnableEmSjLj(WebAssembly::WasmEnableEmSjLj),
365 EnableWasmSjLj(WebAssembly::WasmEnableSjLj) {
366 assert(!(EnableEmSjLj && EnableWasmSjLj) &&
367 "Two SjLj modes cannot be turned on at the same time");
368 assert(!(EnableEmEH && EnableWasmSjLj) &&
369 "Wasm SjLj should be only used with Wasm EH");
370 EHAllowlistSet.insert(first: EHAllowlist.begin(), last: EHAllowlist.end());
371 }
372 bool runOnModule(Module &M) override;
373
374 void getAnalysisUsage(AnalysisUsage &AU) const override {
375 AU.addRequired<DominatorTreeWrapperPass>();
376 }
377};
378} // End anonymous namespace
379
380char WebAssemblyLowerEmscriptenEHSjLj::ID = 0;
381INITIALIZE_PASS(WebAssemblyLowerEmscriptenEHSjLj, DEBUG_TYPE,
382 "WebAssembly Lower Emscripten Exceptions / Setjmp / Longjmp",
383 false, false)
384
385ModulePass *llvm::createWebAssemblyLowerEmscriptenEHSjLj() {
386 return new WebAssemblyLowerEmscriptenEHSjLj();
387}
388
389static bool canThrow(const Value *V) {
390 if (const auto *F = dyn_cast<const Function>(Val: V)) {
391 // Intrinsics cannot throw
392 if (F->isIntrinsic())
393 return false;
394 StringRef Name = F->getName();
395 // leave setjmp and longjmp (mostly) alone, we process them properly later
396 if (Name == "setjmp" || Name == "longjmp" || Name == "emscripten_longjmp")
397 return false;
398 return !F->doesNotThrow();
399 }
400 // not a function, so an indirect call - can throw, we can't tell
401 return true;
402}
403
404// Get a thread-local global variable with the given name. If it doesn't exist
405// declare it, which will generate an import and assume that it will exist at
406// link time.
407static GlobalVariable *getGlobalVariable(Module &M, Type *Ty,
408 WebAssemblyTargetMachine &TM,
409 const char *Name) {
410 auto *GV = dyn_cast<GlobalVariable>(Val: M.getOrInsertGlobal(Name, Ty));
411 if (!GV)
412 report_fatal_error(reason: Twine("unable to create global: ") + Name);
413
414 // Variables created by this function are thread local. If the target does not
415 // support TLS, we depend on CoalesceFeaturesAndStripAtomics to downgrade it
416 // to non-thread-local ones, in which case we don't allow this object to be
417 // linked with other objects using shared memory.
418 GV->setThreadLocalMode(GlobalValue::GeneralDynamicTLSModel);
419 return GV;
420}
421
422// Simple function name mangler.
423// This function simply takes LLVM's string representation of parameter types
424// and concatenate them with '_'. There are non-alphanumeric characters but llc
425// is ok with it, and we need to postprocess these names after the lowering
426// phase anyway.
427static std::string getSignature(FunctionType *FTy) {
428 std::string Sig;
429 raw_string_ostream OS(Sig);
430 OS << *FTy->getReturnType();
431 for (Type *ParamTy : FTy->params())
432 OS << "_" << *ParamTy;
433 if (FTy->isVarArg())
434 OS << "_...";
435 Sig = OS.str();
436 erase_if(C&: Sig, P: isSpace);
437 // When s2wasm parses .s file, a comma means the end of an argument. So a
438 // mangled function name can contain any character but a comma.
439 llvm::replace(Range&: Sig, OldValue: ',', NewValue: '.');
440 return Sig;
441}
442
443static Function *getFunction(FunctionType *Ty, const Twine &Name, Module *M) {
444 return Function::Create(Ty, Linkage: GlobalValue::ExternalLinkage, N: Name, M);
445}
446
447static void markAsImported(Function *F) {
448 // Tell the linker that this function is expected to be imported from the
449 // 'env' module. This is necessary for functions that do not have fixed names
450 // (e.g. __import_xyz). These names cannot be provided by any kind of shared
451 // or static library as instead we mark them explictly as imported.
452 if (!F->hasFnAttribute(Kind: "wasm-import-module")) {
453 llvm::AttrBuilder B(F->getParent()->getContext());
454 B.addAttribute(A: "wasm-import-module", V: "env");
455 F->addFnAttrs(Attrs: B);
456 }
457 if (!F->hasFnAttribute(Kind: "wasm-import-name")) {
458 llvm::AttrBuilder B(F->getParent()->getContext());
459 B.addAttribute(A: "wasm-import-name", V: F->getName());
460 F->addFnAttrs(Attrs: B);
461 }
462}
463
464// Returns an integer type for the target architecture's address space.
465// i32 for wasm32 and i64 for wasm64.
466static Type *getAddrIntType(Module *M) {
467 IRBuilder<> IRB(M->getContext());
468 return IRB.getIntNTy(N: M->getDataLayout().getPointerSizeInBits());
469}
470
471// Returns an integer pointer type for the target architecture's address space.
472// i32* for wasm32 and i64* for wasm64. With opaque pointers this is just a ptr
473// in address space zero.
474static Type *getAddrPtrType(Module *M) {
475 return PointerType::getUnqual(C&: M->getContext());
476}
477
478// Returns an integer whose type is the integer type for the target's address
479// space. Returns (i32 C) for wasm32 and (i64 C) for wasm64, when C is the
480// integer.
481static Value *getAddrSizeInt(Module *M, uint64_t C) {
482 IRBuilder<> IRB(M->getContext());
483 return IRB.getIntN(N: M->getDataLayout().getPointerSizeInBits(), C);
484}
485
486// Returns __cxa_find_matching_catch_N function, where N = NumClauses + 2.
487// This is because a landingpad instruction contains two more arguments, a
488// personality function and a cleanup bit, and __cxa_find_matching_catch_N
489// functions are named after the number of arguments in the original landingpad
490// instruction.
491Function *
492WebAssemblyLowerEmscriptenEHSjLj::getFindMatchingCatch(Module &M,
493 unsigned NumClauses) {
494 auto [It, Inserted] = FindMatchingCatches.try_emplace(Key: NumClauses);
495 if (!Inserted)
496 return It->second;
497 PointerType *Int8PtrTy = PointerType::getUnqual(C&: M.getContext());
498 SmallVector<Type *, 16> Args(NumClauses, Int8PtrTy);
499 FunctionType *FTy = FunctionType::get(Result: Int8PtrTy, Params: Args, isVarArg: false);
500 Function *F = getFunction(
501 Ty: FTy, Name: "__cxa_find_matching_catch_" + Twine(NumClauses + 2), M: &M);
502 markAsImported(F);
503 It->second = F;
504 return F;
505}
506
507// Generate invoke wrapper seqence with preamble and postamble
508// Preamble:
509// __THREW__ = 0;
510// Postamble:
511// %__THREW__.val = __THREW__; __THREW__ = 0;
512// Returns %__THREW__.val, which indicates whether an exception is thrown (or
513// whether longjmp occurred), for future use.
514Value *WebAssemblyLowerEmscriptenEHSjLj::wrapInvoke(CallBase *CI) {
515 Module *M = CI->getModule();
516 LLVMContext &C = M->getContext();
517
518 IRBuilder<> IRB(C);
519 IRB.SetInsertPoint(CI);
520
521 // Pre-invoke
522 // __THREW__ = 0;
523 IRB.CreateStore(Val: getAddrSizeInt(M, C: 0), Ptr: ThrewGV);
524
525 // Invoke function wrapper in JavaScript
526 SmallVector<Value *, 16> Args;
527 // Put the pointer to the callee as first argument, so it can be called
528 // within the invoke wrapper later
529 Args.push_back(Elt: CI->getCalledOperand());
530 Args.append(in_start: CI->arg_begin(), in_end: CI->arg_end());
531 CallInst *NewCall = IRB.CreateCall(Callee: getInvokeWrapper(CI), Args);
532 NewCall->takeName(V: CI);
533 NewCall->setCallingConv(CallingConv::WASM_EmscriptenInvoke);
534 NewCall->setDebugLoc(CI->getDebugLoc());
535
536 // Because we added the pointer to the callee as first argument, all
537 // argument attribute indices have to be incremented by one.
538 SmallVector<AttributeSet, 8> ArgAttributes;
539 const AttributeList &InvokeAL = CI->getAttributes();
540
541 // No attributes for the callee pointer.
542 ArgAttributes.push_back(Elt: AttributeSet());
543 // Copy the argument attributes from the original
544 for (unsigned I = 0, E = CI->arg_size(); I < E; ++I)
545 ArgAttributes.push_back(Elt: InvokeAL.getParamAttrs(ArgNo: I));
546
547 AttrBuilder FnAttrs(CI->getContext(), InvokeAL.getFnAttrs());
548 if (auto Args = FnAttrs.getAllocSizeArgs()) {
549 // The allocsize attribute (if any) referes to parameters by index and needs
550 // to be adjusted.
551 auto [SizeArg, NEltArg] = *Args;
552 SizeArg += 1;
553 if (NEltArg)
554 NEltArg = *NEltArg + 1;
555 FnAttrs.addAllocSizeAttr(ElemSizeArg: SizeArg, NumElemsArg: NEltArg);
556 }
557 // In case the callee has 'noreturn' attribute, We need to remove it, because
558 // we expect invoke wrappers to return.
559 FnAttrs.removeAttribute(Val: Attribute::NoReturn);
560
561 // Reconstruct the AttributesList based on the vector we constructed.
562 AttributeList NewCallAL = AttributeList::get(
563 C, FnAttrs: AttributeSet::get(C, B: FnAttrs), RetAttrs: InvokeAL.getRetAttrs(), ArgAttrs: ArgAttributes);
564 NewCall->setAttributes(NewCallAL);
565
566 CI->replaceAllUsesWith(V: NewCall);
567
568 // Post-invoke
569 // %__THREW__.val = __THREW__; __THREW__ = 0;
570 Value *Threw =
571 IRB.CreateLoad(Ty: getAddrIntType(M), Ptr: ThrewGV, Name: ThrewGV->getName() + ".val");
572 IRB.CreateStore(Val: getAddrSizeInt(M, C: 0), Ptr: ThrewGV);
573 return Threw;
574}
575
576// Get matching invoke wrapper based on callee signature
577Function *WebAssemblyLowerEmscriptenEHSjLj::getInvokeWrapper(CallBase *CI) {
578 Module *M = CI->getModule();
579 SmallVector<Type *, 16> ArgTys;
580 FunctionType *CalleeFTy = CI->getFunctionType();
581
582 std::string Sig = getSignature(FTy: CalleeFTy);
583 auto It = InvokeWrappers.find(Key: Sig);
584 if (It != InvokeWrappers.end())
585 return It->second;
586
587 // Put the pointer to the callee as first argument
588 ArgTys.push_back(Elt: PointerType::getUnqual(C&: CI->getContext()));
589 // Add argument types
590 ArgTys.append(in_start: CalleeFTy->param_begin(), in_end: CalleeFTy->param_end());
591
592 FunctionType *FTy = FunctionType::get(Result: CalleeFTy->getReturnType(), Params: ArgTys,
593 isVarArg: CalleeFTy->isVarArg());
594 Function *F = getFunction(Ty: FTy, Name: "__invoke_" + Sig, M);
595 markAsImported(F);
596 InvokeWrappers[Sig] = F;
597 return F;
598}
599
600static bool canLongjmp(const Value *Callee) {
601 if (auto *CalleeF = dyn_cast<Function>(Val: Callee))
602 if (CalleeF->isIntrinsic())
603 return false;
604
605 // Attempting to transform inline assembly will result in something like:
606 // call void @__invoke_void(void ()* asm ...)
607 // which is invalid because inline assembly blocks do not have addresses
608 // and can't be passed by pointer. The result is a crash with illegal IR.
609 if (isa<InlineAsm>(Val: Callee))
610 return false;
611 StringRef CalleeName = Callee->getName();
612
613 // TODO Include more functions or consider checking with mangled prefixes
614
615 // The reason we include malloc/free here is to exclude the malloc/free
616 // calls generated in setjmp prep / cleanup routines.
617 if (CalleeName == "setjmp" || CalleeName == "malloc" || CalleeName == "free")
618 return false;
619
620 // There are functions in Emscripten's JS glue code or compiler-rt
621 if (CalleeName == "__resumeException" || CalleeName == "llvm_eh_typeid_for" ||
622 CalleeName == "__wasm_setjmp" || CalleeName == "__wasm_setjmp_test" ||
623 CalleeName == "getTempRet0" || CalleeName == "setTempRet0")
624 return false;
625
626 // __cxa_find_matching_catch_N functions cannot longjmp
627 if (Callee->getName().starts_with(Prefix: "__cxa_find_matching_catch_"))
628 return false;
629
630 // Exception-catching related functions
631 //
632 // We intentionally treat __cxa_end_catch longjmpable in Wasm SjLj even though
633 // it surely cannot longjmp, in order to maintain the unwind relationship from
634 // all existing catchpads (and calls within them) to catch.dispatch.longjmp.
635 //
636 // In Wasm EH + Wasm SjLj, we
637 // 1. Make all catchswitch and cleanuppad that unwind to caller unwind to
638 // catch.dispatch.longjmp instead
639 // 2. Convert all longjmpable calls to invokes that unwind to
640 // catch.dispatch.longjmp
641 // But catchswitch BBs are removed in isel, so if an EH catchswitch (generated
642 // from an exception)'s catchpad does not contain any calls that are converted
643 // into invokes unwinding to catch.dispatch.longjmp, this unwind relationship
644 // (EH catchswitch BB -> catch.dispatch.longjmp BB) is lost and
645 // catch.dispatch.longjmp BB can be placed before the EH catchswitch BB in
646 // CFGSort.
647 // int ret = setjmp(buf);
648 // try {
649 // foo(); // longjmps
650 // } catch (...) {
651 // }
652 // Then in this code, if 'foo' longjmps, it first unwinds to 'catch (...)'
653 // catchswitch, and is not caught by that catchswitch because it is a longjmp,
654 // then it should next unwind to catch.dispatch.longjmp BB. But if this 'catch
655 // (...)' catchswitch -> catch.dispatch.longjmp unwind relationship is lost,
656 // it will not unwind to catch.dispatch.longjmp, producing an incorrect
657 // result.
658 //
659 // Every catchpad generated by Wasm C++ contains __cxa_end_catch, so we
660 // intentionally treat it as longjmpable to work around this problem. This is
661 // a hacky fix but an easy one.
662 if (CalleeName == "__cxa_end_catch")
663 return WebAssembly::WasmEnableSjLj;
664 if (CalleeName == "__cxa_begin_catch" ||
665 CalleeName == "__cxa_allocate_exception" || CalleeName == "__cxa_throw" ||
666 CalleeName == "__clang_call_terminate")
667 return false;
668
669 // std::terminate, which is generated when another exception occurs while
670 // handling an exception, cannot longjmp.
671 if (CalleeName == "_ZSt9terminatev")
672 return false;
673
674 // Otherwise we don't know
675 return true;
676}
677
678static bool isEmAsmCall(const Value *Callee) {
679 StringRef CalleeName = Callee->getName();
680 // This is an exhaustive list from Emscripten's <emscripten/em_asm.h>.
681 return CalleeName == "emscripten_asm_const_int" ||
682 CalleeName == "emscripten_asm_const_double" ||
683 CalleeName == "emscripten_asm_const_int_sync_on_main_thread" ||
684 CalleeName == "emscripten_asm_const_double_sync_on_main_thread" ||
685 CalleeName == "emscripten_asm_const_async_on_main_thread";
686}
687
688// Generate __wasm_setjmp_test function call seqence with preamble and
689// postamble. The code this generates is equivalent to the following
690// JavaScript code:
691// %__threwValue.val = __threwValue;
692// if (%__THREW__.val != 0 & %__threwValue.val != 0) {
693// %label = __wasm_setjmp_test(%__THREW__.val, functionInvocationId);
694// if (%label == 0)
695// emscripten_longjmp(%__THREW__.val, %__threwValue.val);
696// setTempRet0(%__threwValue.val);
697// } else {
698// %label = -1;
699// }
700// %longjmp_result = getTempRet0();
701//
702// As output parameters. returns %label, %longjmp_result, and the BB the last
703// instruction (%longjmp_result = ...) is in.
704void WebAssemblyLowerEmscriptenEHSjLj::wrapTestSetjmp(
705 BasicBlock *BB, DebugLoc DL, Value *Threw, Value *FunctionInvocationId,
706 Value *&Label, Value *&LongjmpResult, BasicBlock *&CallEmLongjmpBB,
707 PHINode *&CallEmLongjmpBBThrewPHI, PHINode *&CallEmLongjmpBBThrewValuePHI,
708 BasicBlock *&EndBB) {
709 Function *F = BB->getParent();
710 Module *M = F->getParent();
711 LLVMContext &C = M->getContext();
712 IRBuilder<> IRB(C);
713 IRB.SetCurrentDebugLocation(DL);
714
715 // if (%__THREW__.val != 0 & %__threwValue.val != 0)
716 IRB.SetInsertPoint(BB);
717 BasicBlock *ThenBB1 = BasicBlock::Create(Context&: C, Name: "if.then1", Parent: F);
718 BasicBlock *ElseBB1 = BasicBlock::Create(Context&: C, Name: "if.else1", Parent: F);
719 BasicBlock *EndBB1 = BasicBlock::Create(Context&: C, Name: "if.end", Parent: F);
720 Value *ThrewCmp = IRB.CreateICmpNE(LHS: Threw, RHS: getAddrSizeInt(M, C: 0));
721 Value *ThrewValue = IRB.CreateLoad(Ty: IRB.getInt32Ty(), Ptr: ThrewValueGV,
722 Name: ThrewValueGV->getName() + ".val");
723 Value *ThrewValueCmp = IRB.CreateICmpNE(LHS: ThrewValue, RHS: IRB.getInt32(C: 0));
724 Value *Cmp1 = IRB.CreateAnd(LHS: ThrewCmp, RHS: ThrewValueCmp, Name: "cmp1");
725 IRB.CreateCondBr(Cond: Cmp1, True: ThenBB1, False: ElseBB1);
726
727 // Generate call.em.longjmp BB once and share it within the function
728 if (!CallEmLongjmpBB) {
729 // emscripten_longjmp(%__THREW__.val, %__threwValue.val);
730 CallEmLongjmpBB = BasicBlock::Create(Context&: C, Name: "call.em.longjmp", Parent: F);
731 IRB.SetInsertPoint(CallEmLongjmpBB);
732 CallEmLongjmpBBThrewPHI = IRB.CreatePHI(Ty: getAddrIntType(M), NumReservedValues: 4, Name: "threw.phi");
733 CallEmLongjmpBBThrewValuePHI =
734 IRB.CreatePHI(Ty: IRB.getInt32Ty(), NumReservedValues: 4, Name: "threwvalue.phi");
735 CallEmLongjmpBBThrewPHI->addIncoming(V: Threw, BB: ThenBB1);
736 CallEmLongjmpBBThrewValuePHI->addIncoming(V: ThrewValue, BB: ThenBB1);
737 IRB.CreateCall(Callee: EmLongjmpF,
738 Args: {CallEmLongjmpBBThrewPHI, CallEmLongjmpBBThrewValuePHI});
739 IRB.CreateUnreachable();
740 } else {
741 CallEmLongjmpBBThrewPHI->addIncoming(V: Threw, BB: ThenBB1);
742 CallEmLongjmpBBThrewValuePHI->addIncoming(V: ThrewValue, BB: ThenBB1);
743 }
744
745 // %label = __wasm_setjmp_test(%__THREW__.val, functionInvocationId);
746 // if (%label == 0)
747 IRB.SetInsertPoint(ThenBB1);
748 BasicBlock *EndBB2 = BasicBlock::Create(Context&: C, Name: "if.end2", Parent: F);
749 Value *ThrewPtr =
750 IRB.CreateIntToPtr(V: Threw, DestTy: getAddrPtrType(M), Name: Threw->getName() + ".p");
751 Value *ThenLabel = IRB.CreateCall(Callee: WasmSetjmpTestF,
752 Args: {ThrewPtr, FunctionInvocationId}, Name: "label");
753 Value *Cmp2 = IRB.CreateICmpEQ(LHS: ThenLabel, RHS: IRB.getInt32(C: 0));
754 IRB.CreateCondBr(Cond: Cmp2, True: CallEmLongjmpBB, False: EndBB2);
755
756 // setTempRet0(%__threwValue.val);
757 IRB.SetInsertPoint(EndBB2);
758 IRB.CreateCall(Callee: SetTempRet0F, Args: ThrewValue);
759 IRB.CreateBr(Dest: EndBB1);
760
761 IRB.SetInsertPoint(ElseBB1);
762 IRB.CreateBr(Dest: EndBB1);
763
764 // longjmp_result = getTempRet0();
765 IRB.SetInsertPoint(EndBB1);
766 PHINode *LabelPHI = IRB.CreatePHI(Ty: IRB.getInt32Ty(), NumReservedValues: 2, Name: "label");
767 LabelPHI->addIncoming(V: ThenLabel, BB: EndBB2);
768
769 LabelPHI->addIncoming(V: IRB.getInt32(C: -1), BB: ElseBB1);
770
771 // Output parameter assignment
772 Label = LabelPHI;
773 EndBB = EndBB1;
774 LongjmpResult = IRB.CreateCall(Callee: GetTempRet0F, Args: {}, Name: "longjmp_result");
775}
776
777void WebAssemblyLowerEmscriptenEHSjLj::rebuildSSA(Function &F) {
778 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
779 DT.recalculate(Func&: F); // CFG has been changed
780
781 SSAUpdaterBulk SSA;
782 for (BasicBlock &BB : F) {
783 for (Instruction &I : BB) {
784 if (I.getType()->isVoidTy())
785 continue;
786 unsigned VarID = SSA.AddVariable(Name: I.getName(), Ty: I.getType());
787 // If a value is defined by an invoke instruction, it is only available in
788 // its normal destination and not in its unwind destination.
789 if (auto *II = dyn_cast<InvokeInst>(Val: &I))
790 SSA.AddAvailableValue(Var: VarID, BB: II->getNormalDest(), V: II);
791 else
792 SSA.AddAvailableValue(Var: VarID, BB: &BB, V: &I);
793 for (auto &U : I.uses()) {
794 auto *User = cast<Instruction>(Val: U.getUser());
795 if (auto *UserPN = dyn_cast<PHINode>(Val: User))
796 if (UserPN->getIncomingBlock(U) == &BB)
797 continue;
798 if (DT.dominates(Def: &I, User))
799 continue;
800 SSA.AddUse(Var: VarID, U: &U);
801 }
802 }
803 }
804 SSA.RewriteAllUses(DT: &DT);
805}
806
807// Replace uses of longjmp with a new longjmp function in Emscripten library.
808// In Emscripten SjLj, the new function is
809// void emscripten_longjmp(uintptr_t, i32)
810// In Wasm SjLj, the new function is
811// void __wasm_longjmp(i8*, i32)
812// Because the original libc longjmp function takes (jmp_buf*, i32), we need a
813// ptrtoint/bitcast instruction here to make the type match. jmp_buf* will
814// eventually be lowered to i32/i64 in the wasm backend.
815void WebAssemblyLowerEmscriptenEHSjLj::replaceLongjmpWith(Function *LongjmpF,
816 Function *NewF) {
817 assert(NewF == EmLongjmpF || NewF == WasmLongjmpF);
818 Module *M = LongjmpF->getParent();
819 SmallVector<CallInst *, 8> ToErase;
820 LLVMContext &C = LongjmpF->getParent()->getContext();
821 IRBuilder<> IRB(C);
822
823 // For calls to longjmp, replace it with emscripten_longjmp/__wasm_longjmp and
824 // cast its first argument (jmp_buf*) appropriately
825 for (User *U : LongjmpF->users()) {
826 auto *CI = dyn_cast<CallInst>(Val: U);
827 if (CI && CI->getCalledFunction() == LongjmpF) {
828 IRB.SetInsertPoint(CI);
829 Value *Env = nullptr;
830 if (NewF == EmLongjmpF)
831 Env =
832 IRB.CreatePtrToInt(V: CI->getArgOperand(i: 0), DestTy: getAddrIntType(M), Name: "env");
833 else // WasmLongjmpF
834 Env = IRB.CreateBitCast(V: CI->getArgOperand(i: 0), DestTy: IRB.getPtrTy(), Name: "env");
835 IRB.CreateCall(Callee: NewF, Args: {Env, CI->getArgOperand(i: 1)});
836 ToErase.push_back(Elt: CI);
837 }
838 }
839 for (auto *I : ToErase)
840 I->eraseFromParent();
841
842 // If we have any remaining uses of longjmp's function pointer, replace it
843 // with (void(*)(jmp_buf*, int))emscripten_longjmp / __wasm_longjmp.
844 if (!LongjmpF->uses().empty()) {
845 Value *NewLongjmp =
846 IRB.CreateBitCast(V: NewF, DestTy: LongjmpF->getType(), Name: "longjmp.cast");
847 LongjmpF->replaceAllUsesWith(V: NewLongjmp);
848 }
849}
850
851static bool containsLongjmpableCalls(const Function *F) {
852 for (const auto &BB : *F)
853 for (const auto &I : BB)
854 if (const auto *CB = dyn_cast<CallBase>(Val: &I))
855 if (canLongjmp(Callee: CB->getCalledOperand()))
856 return true;
857 return false;
858}
859
860// When a function contains a setjmp call but not other calls that can longjmp,
861// we don't do setjmp transformation for that setjmp. But we need to convert the
862// setjmp calls into "i32 0" so they don't cause link time errors. setjmp always
863// returns 0 when called directly.
864static void nullifySetjmp(Function *F) {
865 Module &M = *F->getParent();
866 IRBuilder<> IRB(M.getContext());
867 Function *SetjmpF = M.getFunction(Name: "setjmp");
868 SmallVector<Instruction *, 1> ToErase;
869
870 for (User *U : make_early_inc_range(Range: SetjmpF->users())) {
871 auto *CB = cast<CallBase>(Val: U);
872 BasicBlock *BB = CB->getParent();
873 if (BB->getParent() != F) // in other function
874 continue;
875 CallInst *CI = nullptr;
876 // setjmp cannot throw. So if it is an invoke, lower it to a call
877 if (auto *II = dyn_cast<InvokeInst>(Val: CB))
878 CI = llvm::changeToCall(II);
879 else
880 CI = cast<CallInst>(Val: CB);
881 ToErase.push_back(Elt: CI);
882 CI->replaceAllUsesWith(V: IRB.getInt32(C: 0));
883 }
884 for (auto *I : ToErase)
885 I->eraseFromParent();
886}
887
888bool WebAssemblyLowerEmscriptenEHSjLj::runOnModule(Module &M) {
889 LLVM_DEBUG(dbgs() << "********** Lower Emscripten EH & SjLj **********\n");
890
891 LLVMContext &C = M.getContext();
892 IRBuilder<> IRB(C);
893
894 Function *SetjmpF = M.getFunction(Name: "setjmp");
895 Function *LongjmpF = M.getFunction(Name: "longjmp");
896
897 // In some platforms _setjmp and _longjmp are used instead. Change these to
898 // use setjmp/longjmp instead, because we later detect these functions by
899 // their names.
900 Function *SetjmpF2 = M.getFunction(Name: "_setjmp");
901 Function *LongjmpF2 = M.getFunction(Name: "_longjmp");
902 if (SetjmpF2) {
903 if (SetjmpF) {
904 if (SetjmpF->getFunctionType() != SetjmpF2->getFunctionType())
905 report_fatal_error(reason: "setjmp and _setjmp have different function types");
906 } else {
907 SetjmpF = Function::Create(Ty: SetjmpF2->getFunctionType(),
908 Linkage: GlobalValue::ExternalLinkage, N: "setjmp", M);
909 }
910 SetjmpF2->replaceAllUsesWith(V: SetjmpF);
911 }
912 if (LongjmpF2) {
913 if (LongjmpF) {
914 if (LongjmpF->getFunctionType() != LongjmpF2->getFunctionType())
915 report_fatal_error(
916 reason: "longjmp and _longjmp have different function types");
917 } else {
918 LongjmpF = Function::Create(Ty: LongjmpF2->getFunctionType(),
919 Linkage: GlobalValue::ExternalLinkage, N: "setjmp", M);
920 }
921 LongjmpF2->replaceAllUsesWith(V: LongjmpF);
922 }
923
924 auto *TPC = getAnalysisIfAvailable<TargetPassConfig>();
925 assert(TPC && "Expected a TargetPassConfig");
926 auto &TM = TPC->getTM<WebAssemblyTargetMachine>();
927
928 // Declare (or get) global variables __THREW__, __threwValue, and
929 // getTempRet0/setTempRet0 function which are used in common for both
930 // exception handling and setjmp/longjmp handling
931 ThrewGV = getGlobalVariable(M, Ty: getAddrIntType(M: &M), TM, Name: "__THREW__");
932 ThrewValueGV = getGlobalVariable(M, Ty: IRB.getInt32Ty(), TM, Name: "__threwValue");
933 GetTempRet0F = getFunction(Ty: FunctionType::get(Result: IRB.getInt32Ty(), isVarArg: false),
934 Name: "getTempRet0", M: &M);
935 SetTempRet0F =
936 getFunction(Ty: FunctionType::get(Result: IRB.getVoidTy(), Params: IRB.getInt32Ty(), isVarArg: false),
937 Name: "setTempRet0", M: &M);
938 GetTempRet0F->setDoesNotThrow();
939 SetTempRet0F->setDoesNotThrow();
940
941 bool Changed = false;
942
943 // Function registration for exception handling
944 if (EnableEmEH) {
945 // Register __resumeException function
946 FunctionType *ResumeFTy =
947 FunctionType::get(Result: IRB.getVoidTy(), Params: IRB.getPtrTy(), isVarArg: false);
948 ResumeF = getFunction(Ty: ResumeFTy, Name: "__resumeException", M: &M);
949 ResumeF->addFnAttr(Kind: Attribute::NoReturn);
950
951 // Register llvm_eh_typeid_for function
952 FunctionType *EHTypeIDTy =
953 FunctionType::get(Result: IRB.getInt32Ty(), Params: IRB.getPtrTy(), isVarArg: false);
954 EHTypeIDF = getFunction(Ty: EHTypeIDTy, Name: "llvm_eh_typeid_for", M: &M);
955 }
956
957 // Functions that contains calls to setjmp but don't have other longjmpable
958 // calls within them.
959 SmallPtrSet<Function *, 4> SetjmpUsersToNullify;
960
961 if ((EnableEmSjLj || EnableWasmSjLj) && SetjmpF) {
962 // Precompute setjmp users
963 for (User *U : SetjmpF->users()) {
964 if (auto *CB = dyn_cast<CallBase>(Val: U)) {
965 auto *UserF = CB->getFunction();
966 // If a function that calls setjmp does not contain any other calls that
967 // can longjmp, we don't need to do any transformation on that function,
968 // so can ignore it
969 if (containsLongjmpableCalls(F: UserF))
970 SetjmpUsers.insert(Ptr: UserF);
971 else
972 SetjmpUsersToNullify.insert(Ptr: UserF);
973 } else {
974 std::string S;
975 raw_string_ostream SS(S);
976 SS << *U;
977 report_fatal_error(reason: Twine("Indirect use of setjmp is not supported: ") +
978 SS.str());
979 }
980 }
981 }
982
983 bool SetjmpUsed = SetjmpF && !SetjmpUsers.empty();
984 bool LongjmpUsed = LongjmpF && !LongjmpF->use_empty();
985 DoSjLj = (EnableEmSjLj | EnableWasmSjLj) && (SetjmpUsed || LongjmpUsed);
986
987 // Function registration and data pre-gathering for setjmp/longjmp handling
988 if (DoSjLj) {
989 assert(EnableEmSjLj || EnableWasmSjLj);
990 if (EnableEmSjLj) {
991 // Register emscripten_longjmp function
992 FunctionType *FTy = FunctionType::get(
993 Result: IRB.getVoidTy(), Params: {getAddrIntType(M: &M), IRB.getInt32Ty()}, isVarArg: false);
994 EmLongjmpF = getFunction(Ty: FTy, Name: "emscripten_longjmp", M: &M);
995 EmLongjmpF->addFnAttr(Kind: Attribute::NoReturn);
996 } else { // EnableWasmSjLj
997 Type *Int8PtrTy = IRB.getPtrTy();
998 // Register __wasm_longjmp function, which calls __builtin_wasm_longjmp.
999 FunctionType *FTy = FunctionType::get(
1000 Result: IRB.getVoidTy(), Params: {Int8PtrTy, IRB.getInt32Ty()}, isVarArg: false);
1001 WasmLongjmpF = getFunction(Ty: FTy, Name: "__wasm_longjmp", M: &M);
1002 WasmLongjmpF->addFnAttr(Kind: Attribute::NoReturn);
1003 }
1004
1005 if (SetjmpF) {
1006 Type *Int8PtrTy = IRB.getPtrTy();
1007 Type *Int32PtrTy = IRB.getPtrTy();
1008 Type *Int32Ty = IRB.getInt32Ty();
1009
1010 // Register __wasm_setjmp function
1011 FunctionType *SetjmpFTy = SetjmpF->getFunctionType();
1012 FunctionType *FTy = FunctionType::get(
1013 Result: IRB.getVoidTy(), Params: {SetjmpFTy->getParamType(i: 0), Int32Ty, Int32PtrTy},
1014 isVarArg: false);
1015 WasmSetjmpF = getFunction(Ty: FTy, Name: "__wasm_setjmp", M: &M);
1016
1017 // Register __wasm_setjmp_test function
1018 FTy = FunctionType::get(Result: Int32Ty, Params: {Int32PtrTy, Int32PtrTy}, isVarArg: false);
1019 WasmSetjmpTestF = getFunction(Ty: FTy, Name: "__wasm_setjmp_test", M: &M);
1020
1021 // wasm.catch() will be lowered down to wasm 'catch' instruction in
1022 // instruction selection.
1023 CatchF = Intrinsic::getOrInsertDeclaration(M: &M, id: Intrinsic::wasm_catch);
1024 // Type for struct __WasmLongjmpArgs
1025 LongjmpArgsTy = StructType::get(elt1: Int8PtrTy, // env
1026 elts: Int32Ty // val
1027 );
1028 }
1029 }
1030
1031 // Exception handling transformation
1032 if (EnableEmEH) {
1033 for (Function &F : M) {
1034 if (F.isDeclaration())
1035 continue;
1036 Changed |= runEHOnFunction(F);
1037 }
1038 }
1039
1040 // Setjmp/longjmp handling transformation
1041 if (DoSjLj) {
1042 Changed = true; // We have setjmp or longjmp somewhere
1043 if (LongjmpF)
1044 replaceLongjmpWith(LongjmpF, NewF: EnableEmSjLj ? EmLongjmpF : WasmLongjmpF);
1045 // Only traverse functions that uses setjmp in order not to insert
1046 // unnecessary prep / cleanup code in every function
1047 if (SetjmpF)
1048 for (Function *F : SetjmpUsers)
1049 runSjLjOnFunction(F&: *F);
1050 }
1051
1052 // Replace unnecessary setjmp calls with 0
1053 if ((EnableEmSjLj || EnableWasmSjLj) && !SetjmpUsersToNullify.empty()) {
1054 Changed = true;
1055 assert(SetjmpF);
1056 for (Function *F : SetjmpUsersToNullify)
1057 nullifySetjmp(F);
1058 }
1059
1060 // Delete unused global variables and functions
1061 for (auto *V : {ThrewGV, ThrewValueGV})
1062 if (V && V->use_empty())
1063 V->eraseFromParent();
1064 for (auto *V : {GetTempRet0F, SetTempRet0F, ResumeF, EHTypeIDF, EmLongjmpF,
1065 WasmSetjmpF, WasmSetjmpTestF, WasmLongjmpF, CatchF})
1066 if (V && V->use_empty())
1067 V->eraseFromParent();
1068
1069 return Changed;
1070}
1071
1072bool WebAssemblyLowerEmscriptenEHSjLj::runEHOnFunction(Function &F) {
1073 Module &M = *F.getParent();
1074 LLVMContext &C = F.getContext();
1075 IRBuilder<> IRB(C);
1076 bool Changed = false;
1077 SmallVector<Instruction *, 64> ToErase;
1078 SmallPtrSet<LandingPadInst *, 32> LandingPads;
1079
1080 // rethrow.longjmp BB that will be shared within the function.
1081 BasicBlock *RethrowLongjmpBB = nullptr;
1082 // PHI node for the loaded value of __THREW__ global variable in
1083 // rethrow.longjmp BB
1084 PHINode *RethrowLongjmpBBThrewPHI = nullptr;
1085
1086 for (BasicBlock &BB : F) {
1087 auto *II = dyn_cast<InvokeInst>(Val: BB.getTerminator());
1088 if (!II)
1089 continue;
1090 Changed = true;
1091 LandingPads.insert(Ptr: II->getLandingPadInst());
1092 IRB.SetInsertPoint(II);
1093
1094 const Value *Callee = II->getCalledOperand();
1095 bool NeedInvoke = supportsException(F: &F) && canThrow(V: Callee);
1096 if (NeedInvoke) {
1097 // Wrap invoke with invoke wrapper and generate preamble/postamble
1098 Value *Threw = wrapInvoke(CI: II);
1099 ToErase.push_back(Elt: II);
1100
1101 // If setjmp/longjmp handling is enabled, the thrown value can be not an
1102 // exception but a longjmp. If the current function contains calls to
1103 // setjmp, it will be appropriately handled in runSjLjOnFunction. But even
1104 // if the function does not contain setjmp calls, we shouldn't silently
1105 // ignore longjmps; we should rethrow them so they can be correctly
1106 // handled in somewhere up the call chain where setjmp is. __THREW__'s
1107 // value is 0 when nothing happened, 1 when an exception is thrown, and
1108 // other values when longjmp is thrown.
1109 //
1110 // if (%__THREW__.val == 0 || %__THREW__.val == 1)
1111 // goto %tail
1112 // else
1113 // goto %longjmp.rethrow
1114 //
1115 // rethrow.longjmp: ;; This is longjmp. Rethrow it
1116 // %__threwValue.val = __threwValue
1117 // emscripten_longjmp(%__THREW__.val, %__threwValue.val);
1118 //
1119 // tail: ;; Nothing happened or an exception is thrown
1120 // ... Continue exception handling ...
1121 if (DoSjLj && EnableEmSjLj && !SetjmpUsers.count(Ptr: &F) &&
1122 canLongjmp(Callee)) {
1123 // Create longjmp.rethrow BB once and share it within the function
1124 if (!RethrowLongjmpBB) {
1125 RethrowLongjmpBB = BasicBlock::Create(Context&: C, Name: "rethrow.longjmp", Parent: &F);
1126 IRB.SetInsertPoint(RethrowLongjmpBB);
1127 RethrowLongjmpBBThrewPHI =
1128 IRB.CreatePHI(Ty: getAddrIntType(M: &M), NumReservedValues: 4, Name: "threw.phi");
1129 RethrowLongjmpBBThrewPHI->addIncoming(V: Threw, BB: &BB);
1130 Value *ThrewValue = IRB.CreateLoad(Ty: IRB.getInt32Ty(), Ptr: ThrewValueGV,
1131 Name: ThrewValueGV->getName() + ".val");
1132 IRB.CreateCall(Callee: EmLongjmpF, Args: {RethrowLongjmpBBThrewPHI, ThrewValue});
1133 IRB.CreateUnreachable();
1134 } else {
1135 RethrowLongjmpBBThrewPHI->addIncoming(V: Threw, BB: &BB);
1136 }
1137
1138 IRB.SetInsertPoint(II); // Restore the insert point back
1139 BasicBlock *Tail = BasicBlock::Create(Context&: C, Name: "tail", Parent: &F);
1140 Value *CmpEqOne =
1141 IRB.CreateICmpEQ(LHS: Threw, RHS: getAddrSizeInt(M: &M, C: 1), Name: "cmp.eq.one");
1142 Value *CmpEqZero =
1143 IRB.CreateICmpEQ(LHS: Threw, RHS: getAddrSizeInt(M: &M, C: 0), Name: "cmp.eq.zero");
1144 Value *Or = IRB.CreateOr(LHS: CmpEqZero, RHS: CmpEqOne, Name: "or");
1145 IRB.CreateCondBr(Cond: Or, True: Tail, False: RethrowLongjmpBB);
1146 IRB.SetInsertPoint(Tail);
1147 BB.replaceSuccessorsPhiUsesWith(Old: &BB, New: Tail);
1148 }
1149
1150 // Insert a branch based on __THREW__ variable
1151 Value *Cmp = IRB.CreateICmpEQ(LHS: Threw, RHS: getAddrSizeInt(M: &M, C: 1), Name: "cmp");
1152 IRB.CreateCondBr(Cond: Cmp, True: II->getUnwindDest(), False: II->getNormalDest());
1153
1154 } else {
1155 // This can't throw, and we don't need this invoke, just replace it with a
1156 // call+branch
1157 changeToCall(II);
1158 }
1159 }
1160
1161 // Process resume instructions
1162 for (BasicBlock &BB : F) {
1163 // Scan the body of the basic block for resumes
1164 for (Instruction &I : BB) {
1165 auto *RI = dyn_cast<ResumeInst>(Val: &I);
1166 if (!RI)
1167 continue;
1168 Changed = true;
1169
1170 // Split the input into legal values
1171 Value *Input = RI->getValue();
1172 IRB.SetInsertPoint(RI);
1173 Value *Low = IRB.CreateExtractValue(Agg: Input, Idxs: 0, Name: "low");
1174 // Create a call to __resumeException function
1175 IRB.CreateCall(Callee: ResumeF, Args: {Low});
1176 // Add a terminator to the block
1177 IRB.CreateUnreachable();
1178 ToErase.push_back(Elt: RI);
1179 }
1180 }
1181
1182 // Process llvm.eh.typeid.for intrinsics
1183 for (BasicBlock &BB : F) {
1184 for (Instruction &I : BB) {
1185 auto *CI = dyn_cast<CallInst>(Val: &I);
1186 if (!CI)
1187 continue;
1188 const Function *Callee = CI->getCalledFunction();
1189 if (!Callee)
1190 continue;
1191 if (Callee->getIntrinsicID() != Intrinsic::eh_typeid_for)
1192 continue;
1193 Changed = true;
1194
1195 IRB.SetInsertPoint(CI);
1196 CallInst *NewCI =
1197 IRB.CreateCall(Callee: EHTypeIDF, Args: CI->getArgOperand(i: 0), Name: "typeid");
1198 CI->replaceAllUsesWith(V: NewCI);
1199 ToErase.push_back(Elt: CI);
1200 }
1201 }
1202
1203 // Look for orphan landingpads, can occur in blocks with no predecessors
1204 for (BasicBlock &BB : F) {
1205 BasicBlock::iterator I = BB.getFirstNonPHIIt();
1206 if (auto *LPI = dyn_cast<LandingPadInst>(Val&: I))
1207 LandingPads.insert(Ptr: LPI);
1208 }
1209 Changed |= !LandingPads.empty();
1210
1211 // Handle all the landingpad for this function together, as multiple invokes
1212 // may share a single lp
1213 for (LandingPadInst *LPI : LandingPads) {
1214 IRB.SetInsertPoint(LPI);
1215 SmallVector<Value *, 16> FMCArgs;
1216 for (unsigned I = 0, E = LPI->getNumClauses(); I < E; ++I) {
1217 Constant *Clause = LPI->getClause(Idx: I);
1218 // TODO Handle filters (= exception specifications).
1219 // https://github.com/llvm/llvm-project/issues/49740
1220 if (LPI->isCatch(Idx: I))
1221 FMCArgs.push_back(Elt: Clause);
1222 }
1223
1224 // Create a call to __cxa_find_matching_catch_N function
1225 Function *FMCF = getFindMatchingCatch(M, NumClauses: FMCArgs.size());
1226 CallInst *FMCI = IRB.CreateCall(Callee: FMCF, Args: FMCArgs, Name: "fmc");
1227 Value *Poison = PoisonValue::get(T: LPI->getType());
1228 Value *Pair0 = IRB.CreateInsertValue(Agg: Poison, Val: FMCI, Idxs: 0, Name: "pair0");
1229 Value *TempRet0 = IRB.CreateCall(Callee: GetTempRet0F, Args: {}, Name: "tempret0");
1230 Value *Pair1 = IRB.CreateInsertValue(Agg: Pair0, Val: TempRet0, Idxs: 1, Name: "pair1");
1231
1232 LPI->replaceAllUsesWith(V: Pair1);
1233 ToErase.push_back(Elt: LPI);
1234 }
1235
1236 // Erase everything we no longer need in this function
1237 for (Instruction *I : ToErase)
1238 I->eraseFromParent();
1239
1240 return Changed;
1241}
1242
1243// This tries to get debug info from the instruction before which a new
1244// instruction will be inserted, and if there's no debug info in that
1245// instruction, tries to get the info instead from the previous instruction (if
1246// any). If none of these has debug info and a DISubprogram is provided, it
1247// creates a dummy debug info with the first line of the function, because IR
1248// verifier requires all inlinable callsites should have debug info when both a
1249// caller and callee have DISubprogram. If none of these conditions are met,
1250// returns empty info.
1251static DebugLoc getOrCreateDebugLoc(const Instruction *InsertBefore,
1252 DISubprogram *SP) {
1253 assert(InsertBefore);
1254 if (InsertBefore->getDebugLoc())
1255 return InsertBefore->getDebugLoc();
1256 const Instruction *Prev = InsertBefore->getPrevNode();
1257 if (Prev && Prev->getDebugLoc())
1258 return Prev->getDebugLoc();
1259 if (SP)
1260 return DILocation::get(Context&: SP->getContext(), Line: SP->getLine(), Column: 1, Scope: SP);
1261 return DebugLoc();
1262}
1263
1264bool WebAssemblyLowerEmscriptenEHSjLj::runSjLjOnFunction(Function &F) {
1265 assert(EnableEmSjLj || EnableWasmSjLj);
1266 Module &M = *F.getParent();
1267 LLVMContext &C = F.getContext();
1268 IRBuilder<> IRB(C);
1269 SmallVector<Instruction *, 64> ToErase;
1270
1271 // Setjmp preparation
1272
1273 BasicBlock *Entry = &F.getEntryBlock();
1274 DebugLoc FirstDL = getOrCreateDebugLoc(InsertBefore: &*Entry->begin(), SP: F.getSubprogram());
1275 SplitBlock(Old: Entry, SplitPt: &*Entry->getFirstInsertionPt());
1276
1277 IRB.SetInsertPoint(Entry->getTerminator()->getIterator());
1278 // This alloca'ed pointer is used by the runtime to identify function
1279 // invocations. It's just for pointer comparisons. It will never be
1280 // dereferenced.
1281 Instruction *FunctionInvocationId =
1282 IRB.CreateAlloca(Ty: IRB.getInt32Ty(), ArraySize: nullptr, Name: "functionInvocationId");
1283 FunctionInvocationId->setDebugLoc(FirstDL);
1284
1285 // Setjmp transformation
1286 SmallVector<PHINode *, 4> SetjmpRetPHIs;
1287 Function *SetjmpF = M.getFunction(Name: "setjmp");
1288 for (auto *U : make_early_inc_range(Range: SetjmpF->users())) {
1289 auto *CB = cast<CallBase>(Val: U);
1290 BasicBlock *BB = CB->getParent();
1291 if (BB->getParent() != &F) // in other function
1292 continue;
1293 if (CB->getOperandBundle(ID: LLVMContext::OB_funclet)) {
1294 std::string S;
1295 raw_string_ostream SS(S);
1296 SS << "In function " + F.getName() +
1297 ": setjmp within a catch clause is not supported in Wasm EH:\n";
1298 SS << *CB;
1299 report_fatal_error(reason: StringRef(SS.str()));
1300 }
1301
1302 CallInst *CI = nullptr;
1303 // setjmp cannot throw. So if it is an invoke, lower it to a call
1304 if (auto *II = dyn_cast<InvokeInst>(Val: CB))
1305 CI = llvm::changeToCall(II);
1306 else
1307 CI = cast<CallInst>(Val: CB);
1308
1309 // The tail is everything right after the call, and will be reached once
1310 // when setjmp is called, and later when longjmp returns to the setjmp
1311 BasicBlock *Tail = SplitBlock(Old: BB, SplitPt: CI->getNextNode());
1312 // Add a phi to the tail, which will be the output of setjmp, which
1313 // indicates if this is the first call or a longjmp back. The phi directly
1314 // uses the right value based on where we arrive from
1315 IRB.SetInsertPoint(TheBB: Tail, IP: Tail->getFirstNonPHIIt());
1316 PHINode *SetjmpRet = IRB.CreatePHI(Ty: IRB.getInt32Ty(), NumReservedValues: 2, Name: "setjmp.ret");
1317
1318 // setjmp initial call returns 0
1319 SetjmpRet->addIncoming(V: IRB.getInt32(C: 0), BB);
1320 // The proper output is now this, not the setjmp call itself
1321 CI->replaceAllUsesWith(V: SetjmpRet);
1322 // longjmp returns to the setjmp will add themselves to this phi
1323 SetjmpRetPHIs.push_back(Elt: SetjmpRet);
1324
1325 // Fix call target
1326 // Our index in the function is our place in the array + 1 to avoid index
1327 // 0, because index 0 means the longjmp is not ours to handle.
1328 IRB.SetInsertPoint(CI);
1329 Value *Args[] = {CI->getArgOperand(i: 0), IRB.getInt32(C: SetjmpRetPHIs.size()),
1330 FunctionInvocationId};
1331 IRB.CreateCall(Callee: WasmSetjmpF, Args);
1332 ToErase.push_back(Elt: CI);
1333 }
1334
1335 // Handle longjmpable calls.
1336 if (EnableEmSjLj)
1337 handleLongjmpableCallsForEmscriptenSjLj(F, FunctionInvocationId,
1338 SetjmpRetPHIs);
1339 else // EnableWasmSjLj
1340 handleLongjmpableCallsForWasmSjLj(F, FunctionInvocationId, SetjmpRetPHIs);
1341
1342 // Erase everything we no longer need in this function
1343 for (Instruction *I : ToErase)
1344 I->eraseFromParent();
1345
1346 // Finally, our modifications to the cfg can break dominance of SSA variables.
1347 // For example, in this code,
1348 // if (x()) { .. setjmp() .. }
1349 // if (y()) { .. longjmp() .. }
1350 // We must split the longjmp block, and it can jump into the block splitted
1351 // from setjmp one. But that means that when we split the setjmp block, it's
1352 // first part no longer dominates its second part - there is a theoretically
1353 // possible control flow path where x() is false, then y() is true and we
1354 // reach the second part of the setjmp block, without ever reaching the first
1355 // part. So, we rebuild SSA form here.
1356 rebuildSSA(F);
1357 return true;
1358}
1359
1360// Update each call that can longjmp so it can return to the corresponding
1361// setjmp. Refer to 4) of "Emscripten setjmp/longjmp handling" section in the
1362// comments at top of the file for details.
1363void WebAssemblyLowerEmscriptenEHSjLj::handleLongjmpableCallsForEmscriptenSjLj(
1364 Function &F, Instruction *FunctionInvocationId,
1365 SmallVectorImpl<PHINode *> &SetjmpRetPHIs) {
1366 Module &M = *F.getParent();
1367 LLVMContext &C = F.getContext();
1368 IRBuilder<> IRB(C);
1369 SmallVector<Instruction *, 64> ToErase;
1370
1371 // call.em.longjmp BB that will be shared within the function.
1372 BasicBlock *CallEmLongjmpBB = nullptr;
1373 // PHI node for the loaded value of __THREW__ global variable in
1374 // call.em.longjmp BB
1375 PHINode *CallEmLongjmpBBThrewPHI = nullptr;
1376 // PHI node for the loaded value of __threwValue global variable in
1377 // call.em.longjmp BB
1378 PHINode *CallEmLongjmpBBThrewValuePHI = nullptr;
1379 // rethrow.exn BB that will be shared within the function.
1380 BasicBlock *RethrowExnBB = nullptr;
1381
1382 // Because we are creating new BBs while processing and don't want to make
1383 // all these newly created BBs candidates again for longjmp processing, we
1384 // first make the vector of candidate BBs.
1385 std::vector<BasicBlock *> BBs;
1386 for (BasicBlock &BB : F)
1387 BBs.push_back(x: &BB);
1388
1389 // BBs.size() will change within the loop, so we query it every time
1390 for (unsigned I = 0; I < BBs.size(); I++) {
1391 BasicBlock *BB = BBs[I];
1392 for (Instruction &I : *BB) {
1393 if (isa<InvokeInst>(Val: &I)) {
1394 std::string S;
1395 raw_string_ostream SS(S);
1396 SS << "In function " << F.getName()
1397 << ": When using Wasm EH with Emscripten SjLj, there is a "
1398 "restriction that `setjmp` function call and exception cannot be "
1399 "used within the same function:\n";
1400 SS << I;
1401 report_fatal_error(reason: StringRef(SS.str()));
1402 }
1403 auto *CI = dyn_cast<CallInst>(Val: &I);
1404 if (!CI)
1405 continue;
1406
1407 const Value *Callee = CI->getCalledOperand();
1408 if (!canLongjmp(Callee))
1409 continue;
1410 if (isEmAsmCall(Callee))
1411 report_fatal_error(reason: "Cannot use EM_ASM* alongside setjmp/longjmp in " +
1412 F.getName() +
1413 ". Please consider using EM_JS, or move the "
1414 "EM_ASM into another function.",
1415 gen_crash_diag: false);
1416
1417 Value *Threw = nullptr;
1418 BasicBlock *Tail;
1419 if (Callee->getName().starts_with(Prefix: "__invoke_")) {
1420 // If invoke wrapper has already been generated for this call in
1421 // previous EH phase, search for the load instruction
1422 // %__THREW__.val = __THREW__;
1423 // in postamble after the invoke wrapper call
1424 LoadInst *ThrewLI = nullptr;
1425 StoreInst *ThrewResetSI = nullptr;
1426 for (auto I = std::next(x: BasicBlock::iterator(CI)), IE = BB->end();
1427 I != IE; ++I) {
1428 if (auto *LI = dyn_cast<LoadInst>(Val&: I))
1429 if (auto *GV = dyn_cast<GlobalVariable>(Val: LI->getPointerOperand()))
1430 if (GV == ThrewGV) {
1431 Threw = ThrewLI = LI;
1432 break;
1433 }
1434 }
1435 // Search for the store instruction after the load above
1436 // __THREW__ = 0;
1437 for (auto I = std::next(x: BasicBlock::iterator(ThrewLI)), IE = BB->end();
1438 I != IE; ++I) {
1439 if (auto *SI = dyn_cast<StoreInst>(Val&: I)) {
1440 if (auto *GV = dyn_cast<GlobalVariable>(Val: SI->getPointerOperand())) {
1441 if (GV == ThrewGV &&
1442 SI->getValueOperand() == getAddrSizeInt(M: &M, C: 0)) {
1443 ThrewResetSI = SI;
1444 break;
1445 }
1446 }
1447 }
1448 }
1449 assert(Threw && ThrewLI && "Cannot find __THREW__ load after invoke");
1450 assert(ThrewResetSI && "Cannot find __THREW__ store after invoke");
1451 Tail = SplitBlock(Old: BB, SplitPt: ThrewResetSI->getNextNode());
1452
1453 } else {
1454 // Wrap call with invoke wrapper and generate preamble/postamble
1455 Threw = wrapInvoke(CI);
1456 ToErase.push_back(Elt: CI);
1457 Tail = SplitBlock(Old: BB, SplitPt: CI->getNextNode());
1458
1459 // If exception handling is enabled, the thrown value can be not a
1460 // longjmp but an exception, in which case we shouldn't silently ignore
1461 // exceptions; we should rethrow them.
1462 // __THREW__'s value is 0 when nothing happened, 1 when an exception is
1463 // thrown, other values when longjmp is thrown.
1464 //
1465 // if (%__THREW__.val == 1)
1466 // goto %eh.rethrow
1467 // else
1468 // goto %normal
1469 //
1470 // eh.rethrow: ;; Rethrow exception
1471 // %exn = call @__cxa_find_matching_catch_2() ;; Retrieve thrown ptr
1472 // __resumeException(%exn)
1473 //
1474 // normal:
1475 // <-- Insertion point. Will insert sjlj handling code from here
1476 // goto %tail
1477 //
1478 // tail:
1479 // ...
1480 if (supportsException(F: &F) && canThrow(V: Callee)) {
1481 // We will add a new conditional branch. So remove the branch created
1482 // when we split the BB
1483 ToErase.push_back(Elt: BB->getTerminator());
1484
1485 // Generate rethrow.exn BB once and share it within the function
1486 if (!RethrowExnBB) {
1487 RethrowExnBB = BasicBlock::Create(Context&: C, Name: "rethrow.exn", Parent: &F);
1488 IRB.SetInsertPoint(RethrowExnBB);
1489 CallInst *Exn =
1490 IRB.CreateCall(Callee: getFindMatchingCatch(M, NumClauses: 0), Args: {}, Name: "exn");
1491 IRB.CreateCall(Callee: ResumeF, Args: {Exn});
1492 IRB.CreateUnreachable();
1493 }
1494
1495 IRB.SetInsertPoint(CI);
1496 BasicBlock *NormalBB = BasicBlock::Create(Context&: C, Name: "normal", Parent: &F);
1497 Value *CmpEqOne =
1498 IRB.CreateICmpEQ(LHS: Threw, RHS: getAddrSizeInt(M: &M, C: 1), Name: "cmp.eq.one");
1499 IRB.CreateCondBr(Cond: CmpEqOne, True: RethrowExnBB, False: NormalBB);
1500
1501 IRB.SetInsertPoint(NormalBB);
1502 IRB.CreateBr(Dest: Tail);
1503 BB = NormalBB; // New insertion point to insert __wasm_setjmp_test()
1504 }
1505 }
1506
1507 // We need to replace the terminator in Tail - SplitBlock makes BB go
1508 // straight to Tail, we need to check if a longjmp occurred, and go to the
1509 // right setjmp-tail if so
1510 ToErase.push_back(Elt: BB->getTerminator());
1511
1512 // Generate a function call to __wasm_setjmp_test function and
1513 // preamble/postamble code to figure out (1) whether longjmp
1514 // occurred (2) if longjmp occurred, which setjmp it corresponds to
1515 Value *Label = nullptr;
1516 Value *LongjmpResult = nullptr;
1517 BasicBlock *EndBB = nullptr;
1518 wrapTestSetjmp(BB, DL: CI->getDebugLoc(), Threw, FunctionInvocationId, Label,
1519 LongjmpResult, CallEmLongjmpBB, CallEmLongjmpBBThrewPHI,
1520 CallEmLongjmpBBThrewValuePHI, EndBB);
1521 assert(Label && LongjmpResult && EndBB);
1522
1523 // Create switch instruction
1524 IRB.SetInsertPoint(EndBB);
1525 IRB.SetCurrentDebugLocation(EndBB->back().getDebugLoc());
1526 SwitchInst *SI = IRB.CreateSwitch(V: Label, Dest: Tail, NumCases: SetjmpRetPHIs.size());
1527 // -1 means no longjmp happened, continue normally (will hit the default
1528 // switch case). 0 means a longjmp that is not ours to handle, needs a
1529 // rethrow. Otherwise the index is the same as the index in P+1 (to avoid
1530 // 0).
1531 for (unsigned I = 0; I < SetjmpRetPHIs.size(); I++) {
1532 SI->addCase(OnVal: IRB.getInt32(C: I + 1), Dest: SetjmpRetPHIs[I]->getParent());
1533 SetjmpRetPHIs[I]->addIncoming(V: LongjmpResult, BB: EndBB);
1534 }
1535
1536 // We are splitting the block here, and must continue to find other calls
1537 // in the block - which is now split. so continue to traverse in the Tail
1538 BBs.push_back(x: Tail);
1539 }
1540 }
1541
1542 for (Instruction *I : ToErase)
1543 I->eraseFromParent();
1544}
1545
1546static BasicBlock *getCleanupRetUnwindDest(const CleanupPadInst *CPI) {
1547 for (const User *U : CPI->users())
1548 if (const auto *CRI = dyn_cast<CleanupReturnInst>(Val: U))
1549 return CRI->getUnwindDest();
1550 return nullptr;
1551}
1552
1553// Create a catchpad in which we catch a longjmp's env and val arguments, test
1554// if the longjmp corresponds to one of setjmps in the current function, and if
1555// so, jump to the setjmp dispatch BB from which we go to one of post-setjmp
1556// BBs. Refer to 4) of "Wasm setjmp/longjmp handling" section in the comments at
1557// top of the file for details.
1558void WebAssemblyLowerEmscriptenEHSjLj::handleLongjmpableCallsForWasmSjLj(
1559 Function &F, Instruction *FunctionInvocationId,
1560 SmallVectorImpl<PHINode *> &SetjmpRetPHIs) {
1561 Module &M = *F.getParent();
1562 LLVMContext &C = F.getContext();
1563 IRBuilder<> IRB(C);
1564
1565 // A function with catchswitch/catchpad instruction should have a personality
1566 // function attached to it. Search for the wasm personality function, and if
1567 // it exists, use it, and if it doesn't, create a dummy personality function.
1568 // (SjLj is not going to call it anyway.)
1569 if (!F.hasPersonalityFn()) {
1570 StringRef PersName = getEHPersonalityName(Pers: EHPersonality::Wasm_CXX);
1571 FunctionType *PersType =
1572 FunctionType::get(Result: IRB.getInt32Ty(), /* isVarArg */ true);
1573 Value *PersF = M.getOrInsertFunction(Name: PersName, T: PersType).getCallee();
1574 F.setPersonalityFn(
1575 cast<Constant>(Val: IRB.CreateBitCast(V: PersF, DestTy: IRB.getPtrTy())));
1576 }
1577
1578 // Use the entry BB's debugloc as a fallback
1579 BasicBlock *Entry = &F.getEntryBlock();
1580 DebugLoc FirstDL = getOrCreateDebugLoc(InsertBefore: &*Entry->begin(), SP: F.getSubprogram());
1581 IRB.SetCurrentDebugLocation(FirstDL);
1582
1583 // Add setjmp.dispatch BB right after the entry block. Because we have
1584 // initialized functionInvocationId in the entry block and split the
1585 // rest into another BB, here 'OrigEntry' is the function's original entry
1586 // block before the transformation.
1587 //
1588 // entry:
1589 // functionInvocationId initialization
1590 // setjmp.dispatch:
1591 // switch will be inserted here later
1592 // entry.split: (OrigEntry)
1593 // the original function starts here
1594 BasicBlock *OrigEntry = Entry->getNextNode();
1595 BasicBlock *SetjmpDispatchBB =
1596 BasicBlock::Create(Context&: C, Name: "setjmp.dispatch", Parent: &F, InsertBefore: OrigEntry);
1597 cast<BranchInst>(Val: Entry->getTerminator())->setSuccessor(idx: 0, NewSucc: SetjmpDispatchBB);
1598
1599 // Create catch.dispatch.longjmp BB and a catchswitch instruction
1600 BasicBlock *CatchDispatchLongjmpBB =
1601 BasicBlock::Create(Context&: C, Name: "catch.dispatch.longjmp", Parent: &F);
1602 IRB.SetInsertPoint(CatchDispatchLongjmpBB);
1603 CatchSwitchInst *CatchSwitchLongjmp =
1604 IRB.CreateCatchSwitch(ParentPad: ConstantTokenNone::get(Context&: C), UnwindBB: nullptr, NumHandlers: 1);
1605
1606 // Create catch.longjmp BB and a catchpad instruction
1607 BasicBlock *CatchLongjmpBB = BasicBlock::Create(Context&: C, Name: "catch.longjmp", Parent: &F);
1608 CatchSwitchLongjmp->addHandler(Dest: CatchLongjmpBB);
1609 IRB.SetInsertPoint(CatchLongjmpBB);
1610 CatchPadInst *CatchPad = IRB.CreateCatchPad(ParentPad: CatchSwitchLongjmp, Args: {});
1611
1612 // Wasm throw and catch instructions can throw and catch multiple values, but
1613 // that requires multivalue support in the toolchain, which is currently not
1614 // very reliable. We instead throw and catch a pointer to a struct value of
1615 // type 'struct __WasmLongjmpArgs', which is defined in Emscripten.
1616 Instruction *LongjmpArgs =
1617 IRB.CreateCall(Callee: CatchF, Args: {IRB.getInt32(C: WebAssembly::C_LONGJMP)}, Name: "thrown");
1618 Value *EnvField =
1619 IRB.CreateConstGEP2_32(Ty: LongjmpArgsTy, Ptr: LongjmpArgs, Idx0: 0, Idx1: 0, Name: "env_gep");
1620 Value *ValField =
1621 IRB.CreateConstGEP2_32(Ty: LongjmpArgsTy, Ptr: LongjmpArgs, Idx0: 0, Idx1: 1, Name: "val_gep");
1622 // void *env = __wasm_longjmp_args.env;
1623 Instruction *Env = IRB.CreateLoad(Ty: IRB.getPtrTy(), Ptr: EnvField, Name: "env");
1624 // int val = __wasm_longjmp_args.val;
1625 Instruction *Val = IRB.CreateLoad(Ty: IRB.getInt32Ty(), Ptr: ValField, Name: "val");
1626
1627 // %label = __wasm_setjmp_test(%env, functionInvocatinoId);
1628 // if (%label == 0)
1629 // __wasm_longjmp(%env, %val)
1630 // catchret to %setjmp.dispatch
1631 BasicBlock *ThenBB = BasicBlock::Create(Context&: C, Name: "if.then", Parent: &F);
1632 BasicBlock *EndBB = BasicBlock::Create(Context&: C, Name: "if.end", Parent: &F);
1633 Value *EnvP = IRB.CreateBitCast(V: Env, DestTy: getAddrPtrType(M: &M), Name: "env.p");
1634 Value *Label = IRB.CreateCall(Callee: WasmSetjmpTestF, Args: {EnvP, FunctionInvocationId},
1635 OpBundles: OperandBundleDef("funclet", CatchPad), Name: "label");
1636 Value *Cmp = IRB.CreateICmpEQ(LHS: Label, RHS: IRB.getInt32(C: 0));
1637 IRB.CreateCondBr(Cond: Cmp, True: ThenBB, False: EndBB);
1638
1639 IRB.SetInsertPoint(ThenBB);
1640 CallInst *WasmLongjmpCI = IRB.CreateCall(
1641 Callee: WasmLongjmpF, Args: {Env, Val}, OpBundles: OperandBundleDef("funclet", CatchPad));
1642 IRB.CreateUnreachable();
1643
1644 IRB.SetInsertPoint(EndBB);
1645 // Jump to setjmp.dispatch block
1646 IRB.CreateCatchRet(CatchPad, BB: SetjmpDispatchBB);
1647
1648 // Go back to setjmp.dispatch BB
1649 // setjmp.dispatch:
1650 // switch %label {
1651 // label 1: goto post-setjmp BB 1
1652 // label 2: goto post-setjmp BB 2
1653 // ...
1654 // default: goto splitted next BB
1655 // }
1656 IRB.SetInsertPoint(SetjmpDispatchBB);
1657 PHINode *LabelPHI = IRB.CreatePHI(Ty: IRB.getInt32Ty(), NumReservedValues: 2, Name: "label.phi");
1658 LabelPHI->addIncoming(V: Label, BB: EndBB);
1659 LabelPHI->addIncoming(V: IRB.getInt32(C: -1), BB: Entry);
1660 SwitchInst *SI = IRB.CreateSwitch(V: LabelPHI, Dest: OrigEntry, NumCases: SetjmpRetPHIs.size());
1661 // -1 means no longjmp happened, continue normally (will hit the default
1662 // switch case). 0 means a longjmp that is not ours to handle, needs a
1663 // rethrow. Otherwise the index is the same as the index in P+1 (to avoid
1664 // 0).
1665 for (unsigned I = 0; I < SetjmpRetPHIs.size(); I++) {
1666 SI->addCase(OnVal: IRB.getInt32(C: I + 1), Dest: SetjmpRetPHIs[I]->getParent());
1667 SetjmpRetPHIs[I]->addIncoming(V: Val, BB: SetjmpDispatchBB);
1668 }
1669
1670 // Convert all longjmpable call instructions to invokes that unwind to the
1671 // newly created catch.dispatch.longjmp BB.
1672 SmallVector<CallInst *, 64> LongjmpableCalls;
1673 for (auto *BB = &*F.begin(); BB; BB = BB->getNextNode()) {
1674 for (auto &I : *BB) {
1675 auto *CI = dyn_cast<CallInst>(Val: &I);
1676 if (!CI)
1677 continue;
1678 const Value *Callee = CI->getCalledOperand();
1679 if (!canLongjmp(Callee))
1680 continue;
1681 if (isEmAsmCall(Callee))
1682 report_fatal_error(reason: "Cannot use EM_ASM* alongside setjmp/longjmp in " +
1683 F.getName() +
1684 ". Please consider using EM_JS, or move the "
1685 "EM_ASM into another function.",
1686 gen_crash_diag: false);
1687 // This is __wasm_longjmp() call we inserted in this function, which
1688 // rethrows the longjmp when the longjmp does not correspond to one of
1689 // setjmps in this function. We should not convert this call to an invoke.
1690 if (CI == WasmLongjmpCI)
1691 continue;
1692 LongjmpableCalls.push_back(Elt: CI);
1693 }
1694 }
1695
1696 SmallDenseMap<BasicBlock *, SmallSetVector<BasicBlock *, 4>, 4>
1697 UnwindDestToNewPreds;
1698 for (auto *CI : LongjmpableCalls) {
1699 // Even if the callee function has attribute 'nounwind', which is true for
1700 // all C functions, it can longjmp, which means it can throw a Wasm
1701 // exception now.
1702 CI->removeFnAttr(Kind: Attribute::NoUnwind);
1703 if (Function *CalleeF = CI->getCalledFunction())
1704 CalleeF->removeFnAttr(Kind: Attribute::NoUnwind);
1705
1706 // Change it to an invoke and make it unwind to the catch.dispatch.longjmp
1707 // BB. If the call is enclosed in another catchpad/cleanuppad scope, unwind
1708 // to its parent pad's unwind destination instead to preserve the scope
1709 // structure. It will eventually unwind to the catch.dispatch.longjmp.
1710 BasicBlock *UnwindDest = nullptr;
1711 if (auto Bundle = CI->getOperandBundle(ID: LLVMContext::OB_funclet)) {
1712 Instruction *FromPad = cast<Instruction>(Val: Bundle->Inputs[0]);
1713 while (!UnwindDest) {
1714 if (auto *CPI = dyn_cast<CatchPadInst>(Val: FromPad)) {
1715 UnwindDest = CPI->getCatchSwitch()->getUnwindDest();
1716 break;
1717 }
1718 if (auto *CPI = dyn_cast<CleanupPadInst>(Val: FromPad)) {
1719 // getCleanupRetUnwindDest() can return nullptr when
1720 // 1. This cleanuppad's matching cleanupret uwninds to caller
1721 // 2. There is no matching cleanupret because it ends with
1722 // unreachable.
1723 // In case of 2, we need to traverse the parent pad chain.
1724 UnwindDest = getCleanupRetUnwindDest(CPI);
1725 Value *ParentPad = CPI->getParentPad();
1726 if (isa<ConstantTokenNone>(Val: ParentPad))
1727 break;
1728 FromPad = cast<Instruction>(Val: ParentPad);
1729 }
1730 }
1731 }
1732 if (!UnwindDest)
1733 UnwindDest = CatchDispatchLongjmpBB;
1734 // Because we are changing a longjmpable call to an invoke, its unwind
1735 // destination can be an existing EH pad that already have phis, and the BB
1736 // with the newly created invoke will become a new predecessor of that EH
1737 // pad. In this case we need to add the new predecessor to those phis.
1738 UnwindDestToNewPreds[UnwindDest].insert(X: CI->getParent());
1739 changeToInvokeAndSplitBasicBlock(CI, UnwindEdge: UnwindDest);
1740 }
1741
1742 SmallVector<Instruction *, 16> ToErase;
1743 for (auto &BB : F) {
1744 if (auto *CSI = dyn_cast<CatchSwitchInst>(Val: BB.getFirstNonPHIIt())) {
1745 if (CSI != CatchSwitchLongjmp && CSI->unwindsToCaller()) {
1746 IRB.SetInsertPoint(CSI);
1747 ToErase.push_back(Elt: CSI);
1748 auto *NewCSI = IRB.CreateCatchSwitch(ParentPad: CSI->getParentPad(),
1749 UnwindBB: CatchDispatchLongjmpBB, NumHandlers: 1);
1750 NewCSI->addHandler(Dest: *CSI->handler_begin());
1751 NewCSI->takeName(V: CSI);
1752 CSI->replaceAllUsesWith(V: NewCSI);
1753 }
1754 }
1755
1756 if (auto *CRI = dyn_cast<CleanupReturnInst>(Val: BB.getTerminator())) {
1757 if (CRI->unwindsToCaller()) {
1758 IRB.SetInsertPoint(CRI);
1759 ToErase.push_back(Elt: CRI);
1760 IRB.CreateCleanupRet(CleanupPad: CRI->getCleanupPad(), UnwindBB: CatchDispatchLongjmpBB);
1761 }
1762 }
1763 }
1764
1765 for (Instruction *I : ToErase)
1766 I->eraseFromParent();
1767
1768 // Add entries for new predecessors to phis in unwind destinations. We use
1769 // 'poison' as a placeholder value. We should make sure the phis have a valid
1770 // set of predecessors before running SSAUpdater, because SSAUpdater
1771 // internally can use existing phis to gather predecessor info rather than
1772 // scanning the actual CFG (See FindPredecessorBlocks in SSAUpdater.cpp for
1773 // details).
1774 for (auto &[UnwindDest, NewPreds] : UnwindDestToNewPreds) {
1775 for (PHINode &PN : UnwindDest->phis()) {
1776 for (auto *NewPred : NewPreds) {
1777 assert(PN.getBasicBlockIndex(NewPred) == -1);
1778 PN.addIncoming(V: PoisonValue::get(T: PN.getType()), BB: NewPred);
1779 }
1780 }
1781 }
1782
1783 // For unwind destinations for newly added invokes to longjmpable functions,
1784 // calculate incoming values for the newly added predecessors using
1785 // SSAUpdater. We add existing values in the phis to SSAUpdater as available
1786 // values and let it calculate what the value should be at the end of new
1787 // incoming blocks.
1788 for (auto &[UnwindDest, NewPreds] : UnwindDestToNewPreds) {
1789 for (PHINode &PN : UnwindDest->phis()) {
1790 SSAUpdater SSA;
1791 SSA.Initialize(Ty: PN.getType(), Name: PN.getName());
1792 for (unsigned Idx = 0, E = PN.getNumIncomingValues(); Idx != E; ++Idx) {
1793 if (NewPreds.contains(key: PN.getIncomingBlock(i: Idx)))
1794 continue;
1795 Value *V = PN.getIncomingValue(i: Idx);
1796 if (auto *II = dyn_cast<InvokeInst>(Val: V))
1797 SSA.AddAvailableValue(BB: II->getNormalDest(), V: II);
1798 else if (auto *I = dyn_cast<Instruction>(Val: V))
1799 SSA.AddAvailableValue(BB: I->getParent(), V: I);
1800 else
1801 SSA.AddAvailableValue(BB: PN.getIncomingBlock(i: Idx), V);
1802 }
1803 for (auto *NewPred : NewPreds)
1804 PN.setIncomingValueForBlock(BB: NewPred, V: SSA.GetValueAtEndOfBlock(BB: NewPred));
1805 assert(PN.isComplete());
1806 }
1807 }
1808}
1809