1 //===-- WasmEHPrepare - Prepare excepton handling for WebAssembly --------===//
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 // This transformation is designed for use by code generators which use
10 // WebAssembly exception handling scheme. This currently supports C++
11 // exceptions.
12 //
13 // WebAssembly exception handling uses Windows exception IR for the middle level
14 // representation. This pass does the following transformation for every
15 // catchpad block:
16 // (In C-style pseudocode)
17 //
18 // - Before:
19 // catchpad ...
20 // exn = wasm.get.exception();
21 // selector = wasm.get.selector();
22 // ...
23 //
24 // - After:
25 // catchpad ...
26 // exn = wasm.extract.exception();
27 // // Only add below in case it's not a single catch (...)
28 // wasm.landingpad.index(index);
29 // __wasm_lpad_context.lpad_index = index;
30 // __wasm_lpad_context.lsda = wasm.lsda();
31 // _Unwind_CallPersonality(exn);
32 // selector = __wasm.landingpad_context.selector;
33 // ...
34 //
35 //
36 // * Background: Direct personality function call
37 // In WebAssembly EH, the VM is responsible for unwinding the stack once an
38 // exception is thrown. After the stack is unwound, the control flow is
39 // transfered to WebAssembly 'catch' instruction.
40 //
41 // Unwinding the stack is not done by libunwind but the VM, so the personality
42 // function in libcxxabi cannot be called from libunwind during the unwinding
43 // process. So after a catch instruction, we insert a call to a wrapper function
44 // in libunwind that in turn calls the real personality function.
45 //
46 // In Itanium EH, if the personality function decides there is no matching catch
47 // clause in a call frame and no cleanup action to perform, the unwinder doesn't
48 // stop there and continues unwinding. But in Wasm EH, the unwinder stops at
49 // every call frame with a catch intruction, after which the personality
50 // function is called from the compiler-generated user code here.
51 //
52 // In libunwind, we have this struct that serves as a communincation channel
53 // between the compiler-generated user code and the personality function in
54 // libcxxabi.
55 //
56 // struct _Unwind_LandingPadContext {
57 // uintptr_t lpad_index;
58 // uintptr_t lsda;
59 // uintptr_t selector;
60 // };
61 // struct _Unwind_LandingPadContext __wasm_lpad_context = ...;
62 //
63 // And this wrapper in libunwind calls the personality function.
64 //
65 // _Unwind_Reason_Code _Unwind_CallPersonality(void *exception_ptr) {
66 // struct _Unwind_Exception *exception_obj =
67 // (struct _Unwind_Exception *)exception_ptr;
68 // _Unwind_Reason_Code ret = __gxx_personality_v0(
69 // 1, _UA_CLEANUP_PHASE, exception_obj->exception_class, exception_obj,
70 // (struct _Unwind_Context *)__wasm_lpad_context);
71 // return ret;
72 // }
73 //
74 // We pass a landing pad index, and the address of LSDA for the current function
75 // to the wrapper function _Unwind_CallPersonality in libunwind, and we retrieve
76 // the selector after it returns.
77 //
78 //===----------------------------------------------------------------------===//
79
80 #include "llvm/ADT/SetVector.h"
81 #include "llvm/ADT/Statistic.h"
82 #include "llvm/ADT/Triple.h"
83 #include "llvm/CodeGen/Passes.h"
84 #include "llvm/CodeGen/TargetLowering.h"
85 #include "llvm/CodeGen/TargetSubtargetInfo.h"
86 #include "llvm/CodeGen/WasmEHFuncInfo.h"
87 #include "llvm/IR/Dominators.h"
88 #include "llvm/IR/IRBuilder.h"
89 #include "llvm/IR/Intrinsics.h"
90 #include "llvm/IR/IntrinsicsWebAssembly.h"
91 #include "llvm/InitializePasses.h"
92 #include "llvm/Pass.h"
93 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
94
95 using namespace llvm;
96
97 #define DEBUG_TYPE "wasmehprepare"
98
99 namespace {
100 class WasmEHPrepare : public FunctionPass {
101 Type *LPadContextTy = nullptr; // type of 'struct _Unwind_LandingPadContext'
102 GlobalVariable *LPadContextGV = nullptr; // __wasm_lpad_context
103
104 // Field addresses of struct _Unwind_LandingPadContext
105 Value *LPadIndexField = nullptr; // lpad_index field
106 Value *LSDAField = nullptr; // lsda field
107 Value *SelectorField = nullptr; // selector
108
109 Function *ThrowF = nullptr; // wasm.throw() intrinsic
110 Function *LPadIndexF = nullptr; // wasm.landingpad.index() intrinsic
111 Function *LSDAF = nullptr; // wasm.lsda() intrinsic
112 Function *GetExnF = nullptr; // wasm.get.exception() intrinsic
113 Function *ExtractExnF = nullptr; // wasm.extract.exception() intrinsic
114 Function *GetSelectorF = nullptr; // wasm.get.ehselector() intrinsic
115 FunctionCallee CallPersonalityF =
116 nullptr; // _Unwind_CallPersonality() wrapper
117
118 bool prepareEHPads(Function &F);
119 bool prepareThrows(Function &F);
120
121 void prepareEHPad(BasicBlock *BB, bool NeedLSDA, unsigned Index = 0);
122 void prepareTerminateCleanupPad(BasicBlock *BB);
123
124 public:
125 static char ID; // Pass identification, replacement for typeid
126
WasmEHPrepare()127 WasmEHPrepare() : FunctionPass(ID) {}
128
129 bool doInitialization(Module &M) override;
130 bool runOnFunction(Function &F) override;
131
getPassName() const132 StringRef getPassName() const override {
133 return "WebAssembly Exception handling preparation";
134 }
135 };
136 } // end anonymous namespace
137
138 char WasmEHPrepare::ID = 0;
139 INITIALIZE_PASS(WasmEHPrepare, DEBUG_TYPE, "Prepare WebAssembly exceptions",
140 false, false)
141
createWasmEHPass()142 FunctionPass *llvm::createWasmEHPass() { return new WasmEHPrepare(); }
143
doInitialization(Module & M)144 bool WasmEHPrepare::doInitialization(Module &M) {
145 IRBuilder<> IRB(M.getContext());
146 LPadContextTy = StructType::get(IRB.getInt32Ty(), // lpad_index
147 IRB.getInt8PtrTy(), // lsda
148 IRB.getInt32Ty() // selector
149 );
150 return false;
151 }
152
153 // Erase the specified BBs if the BB does not have any remaining predecessors,
154 // and also all its dead children.
155 template <typename Container>
eraseDeadBBsAndChildren(const Container & BBs)156 static void eraseDeadBBsAndChildren(const Container &BBs) {
157 SmallVector<BasicBlock *, 8> WL(BBs.begin(), BBs.end());
158 while (!WL.empty()) {
159 auto *BB = WL.pop_back_val();
160 if (pred_begin(BB) != pred_end(BB))
161 continue;
162 WL.append(succ_begin(BB), succ_end(BB));
163 DeleteDeadBlock(BB);
164 }
165 }
166
runOnFunction(Function & F)167 bool WasmEHPrepare::runOnFunction(Function &F) {
168 bool Changed = false;
169 Changed |= prepareThrows(F);
170 Changed |= prepareEHPads(F);
171 return Changed;
172 }
173
prepareThrows(Function & F)174 bool WasmEHPrepare::prepareThrows(Function &F) {
175 Module &M = *F.getParent();
176 IRBuilder<> IRB(F.getContext());
177 bool Changed = false;
178
179 // wasm.throw() intinsic, which will be lowered to wasm 'throw' instruction.
180 ThrowF = Intrinsic::getDeclaration(&M, Intrinsic::wasm_throw);
181 // Insert an unreachable instruction after a call to @llvm.wasm.throw and
182 // delete all following instructions within the BB, and delete all the dead
183 // children of the BB as well.
184 for (User *U : ThrowF->users()) {
185 // A call to @llvm.wasm.throw() is only generated from __cxa_throw()
186 // builtin call within libcxxabi, and cannot be an InvokeInst.
187 auto *ThrowI = cast<CallInst>(U);
188 if (ThrowI->getFunction() != &F)
189 continue;
190 Changed = true;
191 auto *BB = ThrowI->getParent();
192 SmallVector<BasicBlock *, 4> Succs(succ_begin(BB), succ_end(BB));
193 auto &InstList = BB->getInstList();
194 InstList.erase(std::next(BasicBlock::iterator(ThrowI)), InstList.end());
195 IRB.SetInsertPoint(BB);
196 IRB.CreateUnreachable();
197 eraseDeadBBsAndChildren(Succs);
198 }
199
200 return Changed;
201 }
202
prepareEHPads(Function & F)203 bool WasmEHPrepare::prepareEHPads(Function &F) {
204 Module &M = *F.getParent();
205 IRBuilder<> IRB(F.getContext());
206
207 SmallVector<BasicBlock *, 16> CatchPads;
208 SmallVector<BasicBlock *, 16> CleanupPads;
209 for (BasicBlock &BB : F) {
210 if (!BB.isEHPad())
211 continue;
212 auto *Pad = BB.getFirstNonPHI();
213 if (isa<CatchPadInst>(Pad))
214 CatchPads.push_back(&BB);
215 else if (isa<CleanupPadInst>(Pad))
216 CleanupPads.push_back(&BB);
217 }
218
219 if (CatchPads.empty() && CleanupPads.empty())
220 return false;
221 assert(F.hasPersonalityFn() && "Personality function not found");
222
223 // __wasm_lpad_context global variable
224 LPadContextGV = cast<GlobalVariable>(
225 M.getOrInsertGlobal("__wasm_lpad_context", LPadContextTy));
226 LPadIndexField = IRB.CreateConstGEP2_32(LPadContextTy, LPadContextGV, 0, 0,
227 "lpad_index_gep");
228 LSDAField =
229 IRB.CreateConstGEP2_32(LPadContextTy, LPadContextGV, 0, 1, "lsda_gep");
230 SelectorField = IRB.CreateConstGEP2_32(LPadContextTy, LPadContextGV, 0, 2,
231 "selector_gep");
232
233 // wasm.landingpad.index() intrinsic, which is to specify landingpad index
234 LPadIndexF = Intrinsic::getDeclaration(&M, Intrinsic::wasm_landingpad_index);
235 // wasm.lsda() intrinsic. Returns the address of LSDA table for the current
236 // function.
237 LSDAF = Intrinsic::getDeclaration(&M, Intrinsic::wasm_lsda);
238 // wasm.get.exception() and wasm.get.ehselector() intrinsics. Calls to these
239 // are generated in clang.
240 GetExnF = Intrinsic::getDeclaration(&M, Intrinsic::wasm_get_exception);
241 GetSelectorF = Intrinsic::getDeclaration(&M, Intrinsic::wasm_get_ehselector);
242
243 // wasm.extract.exception() is the same as wasm.get.exception() but it does
244 // not take a token argument. This will be lowered down to EXTRACT_EXCEPTION
245 // pseudo instruction in instruction selection, which will be expanded using
246 // 'br_on_exn' instruction later.
247 ExtractExnF =
248 Intrinsic::getDeclaration(&M, Intrinsic::wasm_extract_exception);
249
250 // _Unwind_CallPersonality() wrapper function, which calls the personality
251 CallPersonalityF = M.getOrInsertFunction(
252 "_Unwind_CallPersonality", IRB.getInt32Ty(), IRB.getInt8PtrTy());
253 if (Function *F = dyn_cast<Function>(CallPersonalityF.getCallee()))
254 F->setDoesNotThrow();
255
256 unsigned Index = 0;
257 for (auto *BB : CatchPads) {
258 auto *CPI = cast<CatchPadInst>(BB->getFirstNonPHI());
259 // In case of a single catch (...), we don't need to emit LSDA
260 if (CPI->getNumArgOperands() == 1 &&
261 cast<Constant>(CPI->getArgOperand(0))->isNullValue())
262 prepareEHPad(BB, false);
263 else
264 prepareEHPad(BB, true, Index++);
265 }
266
267 // Cleanup pads don't need LSDA.
268 for (auto *BB : CleanupPads)
269 prepareEHPad(BB, false);
270
271 return true;
272 }
273
274 // Prepare an EH pad for Wasm EH handling. If NeedLSDA is false, Index is
275 // ignored.
prepareEHPad(BasicBlock * BB,bool NeedLSDA,unsigned Index)276 void WasmEHPrepare::prepareEHPad(BasicBlock *BB, bool NeedLSDA,
277 unsigned Index) {
278 assert(BB->isEHPad() && "BB is not an EHPad!");
279 IRBuilder<> IRB(BB->getContext());
280 IRB.SetInsertPoint(&*BB->getFirstInsertionPt());
281
282 auto *FPI = cast<FuncletPadInst>(BB->getFirstNonPHI());
283 Instruction *GetExnCI = nullptr, *GetSelectorCI = nullptr;
284 for (auto &U : FPI->uses()) {
285 if (auto *CI = dyn_cast<CallInst>(U.getUser())) {
286 if (CI->getCalledValue() == GetExnF)
287 GetExnCI = CI;
288 if (CI->getCalledValue() == GetSelectorF)
289 GetSelectorCI = CI;
290 }
291 }
292
293 // Cleanup pads w/o __clang_call_terminate call do not have any of
294 // wasm.get.exception() or wasm.get.ehselector() calls. We need to do nothing.
295 if (!GetExnCI) {
296 assert(!GetSelectorCI &&
297 "wasm.get.ehselector() cannot exist w/o wasm.get.exception()");
298 return;
299 }
300
301 Instruction *ExtractExnCI = IRB.CreateCall(ExtractExnF, {}, "exn");
302 GetExnCI->replaceAllUsesWith(ExtractExnCI);
303 GetExnCI->eraseFromParent();
304
305 // In case it is a catchpad with single catch (...) or a cleanuppad, we don't
306 // need to call personality function because we don't need a selector.
307 if (!NeedLSDA) {
308 if (GetSelectorCI) {
309 assert(GetSelectorCI->use_empty() &&
310 "wasm.get.ehselector() still has uses!");
311 GetSelectorCI->eraseFromParent();
312 }
313 return;
314 }
315 IRB.SetInsertPoint(ExtractExnCI->getNextNode());
316
317 // This is to create a map of <landingpad EH label, landingpad index> in
318 // SelectionDAGISel, which is to be used in EHStreamer to emit LSDA tables.
319 // Pseudocode: wasm.landingpad.index(Index);
320 IRB.CreateCall(LPadIndexF, {FPI, IRB.getInt32(Index)});
321
322 // Pseudocode: __wasm_lpad_context.lpad_index = index;
323 IRB.CreateStore(IRB.getInt32(Index), LPadIndexField);
324
325 // Store LSDA address only if this catchpad belongs to a top-level
326 // catchswitch. If there is another catchpad that dominates this pad, we don't
327 // need to store LSDA address again, because they are the same throughout the
328 // function and have been already stored before.
329 // TODO Can we not store LSDA address in user function but make libcxxabi
330 // compute it?
331 auto *CPI = cast<CatchPadInst>(FPI);
332 if (isa<ConstantTokenNone>(CPI->getCatchSwitch()->getParentPad()))
333 // Pseudocode: __wasm_lpad_context.lsda = wasm.lsda();
334 IRB.CreateStore(IRB.CreateCall(LSDAF), LSDAField);
335
336 // Pseudocode: _Unwind_CallPersonality(exn);
337 CallInst *PersCI = IRB.CreateCall(CallPersonalityF, ExtractExnCI,
338 OperandBundleDef("funclet", CPI));
339 PersCI->setDoesNotThrow();
340
341 // Pseudocode: int selector = __wasm.landingpad_context.selector;
342 Instruction *Selector =
343 IRB.CreateLoad(IRB.getInt32Ty(), SelectorField, "selector");
344
345 // Replace the return value from wasm.get.ehselector() with the selector value
346 // loaded from __wasm_lpad_context.selector.
347 assert(GetSelectorCI && "wasm.get.ehselector() call does not exist");
348 GetSelectorCI->replaceAllUsesWith(Selector);
349 GetSelectorCI->eraseFromParent();
350 }
351
calculateWasmEHInfo(const Function * F,WasmEHFuncInfo & EHInfo)352 void llvm::calculateWasmEHInfo(const Function *F, WasmEHFuncInfo &EHInfo) {
353 // If an exception is not caught by a catchpad (i.e., it is a foreign
354 // exception), it will unwind to its parent catchswitch's unwind destination.
355 // We don't record an unwind destination for cleanuppads because every
356 // exception should be caught by it.
357 for (const auto &BB : *F) {
358 if (!BB.isEHPad())
359 continue;
360 const Instruction *Pad = BB.getFirstNonPHI();
361
362 if (const auto *CatchPad = dyn_cast<CatchPadInst>(Pad)) {
363 const auto *UnwindBB = CatchPad->getCatchSwitch()->getUnwindDest();
364 if (!UnwindBB)
365 continue;
366 const Instruction *UnwindPad = UnwindBB->getFirstNonPHI();
367 if (const auto *CatchSwitch = dyn_cast<CatchSwitchInst>(UnwindPad))
368 // Currently there should be only one handler per a catchswitch.
369 EHInfo.setEHPadUnwindDest(&BB, *CatchSwitch->handlers().begin());
370 else // cleanuppad
371 EHInfo.setEHPadUnwindDest(&BB, UnwindBB);
372 }
373 }
374 }
375