1 //===- LowerSetJmp.cpp - Code pertaining to lowering set/long jumps -------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the lowering of setjmp and longjmp to use the
11 // LLVM invoke and unwind instructions as necessary.
13 // Lowering of longjmp is fairly trivial. We replace the call with a
14 // call to the LLVM library function "__llvm_sjljeh_throw_longjmp()".
15 // This unwinds the stack for us calling all of the destructors for
16 // objects allocated on the stack.
18 // At a setjmp call, the basic block is split and the setjmp removed.
19 // The calls in a function that have a setjmp are converted to invoke
20 // where the except part checks to see if it's a longjmp exception and,
21 // if so, if it's handled in the function. If it is, then it gets the
22 // value returned by the longjmp and goes to where the basic block was
23 // split. Invoke instructions are handled in a similar fashion with the
24 // original except block being executed if it isn't a longjmp except
25 // that is handled by that function.
27 //===----------------------------------------------------------------------===//
29 //===----------------------------------------------------------------------===//
30 // FIXME: This pass doesn't deal with PHI statements just yet. That is,
31 // we expect this to occur before SSAification is done. This would seem
32 // to make sense, but in general, it might be a good idea to make this
33 // pass invokable via the "opt" command at will.
34 //===----------------------------------------------------------------------===//
36 #include "llvm/Transforms/IPO.h"
37 #include "llvm/Constants.h"
38 #include "llvm/DerivedTypes.h"
39 #include "llvm/Instructions.h"
40 #include "llvm/Intrinsics.h"
41 #include "llvm/Module.h"
42 #include "llvm/Pass.h"
43 #include "llvm/Support/CFG.h"
44 #include "llvm/Support/InstVisitor.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/ADT/DepthFirstIterator.h"
47 #include "llvm/ADT/Statistic.h"
48 #include "llvm/ADT/StringExtras.h"
49 #include "llvm/ADT/VectorExtras.h"
53 Statistic<> LongJmpsTransformed("lowersetjmp",
54 "Number of longjmps transformed");
55 Statistic<> SetJmpsTransformed("lowersetjmp",
56 "Number of setjmps transformed");
57 Statistic<> CallsTransformed("lowersetjmp",
58 "Number of calls invokified");
59 Statistic<> InvokesTransformed("lowersetjmp",
60 "Number of invokes modified");
62 //===--------------------------------------------------------------------===//
63 // LowerSetJmp pass implementation.
64 class LowerSetJmp : public ModulePass,
65 public InstVisitor<LowerSetJmp> {
66 // LLVM library functions...
67 Function* InitSJMap; // __llvm_sjljeh_init_setjmpmap
68 Function* DestroySJMap; // __llvm_sjljeh_destroy_setjmpmap
69 Function* AddSJToMap; // __llvm_sjljeh_add_setjmp_to_map
70 Function* ThrowLongJmp; // __llvm_sjljeh_throw_longjmp
71 Function* TryCatchLJ; // __llvm_sjljeh_try_catching_longjmp_exception
72 Function* IsLJException; // __llvm_sjljeh_is_longjmp_exception
73 Function* GetLJValue; // __llvm_sjljeh_get_longjmp_value
75 typedef std::pair<SwitchInst*, CallInst*> SwitchValuePair;
77 // Keep track of those basic blocks reachable via a depth-first search of
78 // the CFG from a setjmp call. We only need to transform those "call" and
79 // "invoke" instructions that are reachable from the setjmp call site.
80 std::set<BasicBlock*> DFSBlocks;
82 // The setjmp map is going to hold information about which setjmps
83 // were called (each setjmp gets its own number) and with which
84 // buffer it was called.
85 std::map<Function*, AllocaInst*> SJMap;
87 // The rethrow basic block map holds the basic block to branch to if
88 // the exception isn't handled in the current function and needs to
90 std::map<const Function*, BasicBlock*> RethrowBBMap;
92 // The preliminary basic block map holds a basic block that grabs the
93 // exception and determines if it's handled by the current function.
94 std::map<const Function*, BasicBlock*> PrelimBBMap;
96 // The switch/value map holds a switch inst/call inst pair. The
97 // switch inst controls which handler (if any) gets called and the
98 // value is the value returned to that handler by the call to
99 // __llvm_sjljeh_get_longjmp_value.
100 std::map<const Function*, SwitchValuePair> SwitchValMap;
102 // A map of which setjmps we've seen so far in a function.
103 std::map<const Function*, unsigned> SetJmpIDMap;
105 AllocaInst* GetSetJmpMap(Function* Func);
106 BasicBlock* GetRethrowBB(Function* Func);
107 SwitchValuePair GetSJSwitch(Function* Func, BasicBlock* Rethrow);
109 void TransformLongJmpCall(CallInst* Inst);
110 void TransformSetJmpCall(CallInst* Inst);
112 bool IsTransformableFunction(const std::string& Name);
114 void visitCallInst(CallInst& CI);
115 void visitInvokeInst(InvokeInst& II);
116 void visitReturnInst(ReturnInst& RI);
117 void visitUnwindInst(UnwindInst& UI);
119 bool runOnModule(Module& M);
120 bool doInitialization(Module& M);
123 RegisterOpt<LowerSetJmp> X("lowersetjmp", "Lower Set Jump");
124 } // end anonymous namespace
126 // run - Run the transformation on the program. We grab the function
127 // prototypes for longjmp and setjmp. If they are used in the program,
128 // then we can go directly to the places they're at and transform them.
129 bool LowerSetJmp::runOnModule(Module& M) {
130 bool Changed = false;
132 // These are what the functions are called.
133 Function* SetJmp = M.getNamedFunction("llvm.setjmp");
134 Function* LongJmp = M.getNamedFunction("llvm.longjmp");
136 // This program doesn't have longjmp and setjmp calls.
137 if ((!LongJmp || LongJmp->use_empty()) &&
138 (!SetJmp || SetJmp->use_empty())) return false;
140 // Initialize some values and functions we'll need to transform the
141 // setjmp/longjmp functions.
145 for (Value::use_iterator B = SetJmp->use_begin(), E = SetJmp->use_end();
147 BasicBlock* BB = cast<Instruction>(*B)->getParent();
148 for (df_ext_iterator<BasicBlock*> I = df_ext_begin(BB, DFSBlocks),
149 E = df_ext_end(BB, DFSBlocks); I != E; ++I)
153 while (!SetJmp->use_empty()) {
154 assert(isa<CallInst>(SetJmp->use_back()) &&
155 "User of setjmp intrinsic not a call?");
156 TransformSetJmpCall(cast<CallInst>(SetJmp->use_back()));
162 while (!LongJmp->use_empty()) {
163 assert(isa<CallInst>(LongJmp->use_back()) &&
164 "User of longjmp intrinsic not a call?");
165 TransformLongJmpCall(cast<CallInst>(LongJmp->use_back()));
169 // Now go through the affected functions and convert calls and invokes
171 for (std::map<Function*, AllocaInst*>::iterator
172 B = SJMap.begin(), E = SJMap.end(); B != E; ++B) {
173 Function* F = B->first;
174 for (Function::iterator BB = F->begin(), BE = F->end(); BB != BE; ++BB)
175 for (BasicBlock::iterator IB = BB->begin(), IE = BB->end(); IB != IE; ) {
177 if (IB != BB->end() && IB->getParent() != BB)
178 break; // The next instruction got moved to a different block!
184 RethrowBBMap.clear();
186 SwitchValMap.clear();
192 // doInitialization - For the lower long/setjmp pass, this ensures that a
193 // module contains a declaration for the intrisic functions we are going
194 // to call to convert longjmp and setjmp calls.
196 // This function is always successful, unless it isn't.
197 bool LowerSetJmp::doInitialization(Module& M)
199 const Type *SBPTy = PointerType::get(Type::SByteTy);
200 const Type *SBPPTy = PointerType::get(SBPTy);
202 // N.B. See llvm/runtime/GCCLibraries/libexception/SJLJ-Exception.h for
203 // a description of the following library functions.
205 // void __llvm_sjljeh_init_setjmpmap(void**)
206 InitSJMap = M.getOrInsertFunction("__llvm_sjljeh_init_setjmpmap",
207 Type::VoidTy, SBPPTy, (Type *)0);
208 // void __llvm_sjljeh_destroy_setjmpmap(void**)
209 DestroySJMap = M.getOrInsertFunction("__llvm_sjljeh_destroy_setjmpmap",
210 Type::VoidTy, SBPPTy, (Type *)0);
212 // void __llvm_sjljeh_add_setjmp_to_map(void**, void*, unsigned)
213 AddSJToMap = M.getOrInsertFunction("__llvm_sjljeh_add_setjmp_to_map",
214 Type::VoidTy, SBPPTy, SBPTy,
215 Type::UIntTy, (Type *)0);
217 // void __llvm_sjljeh_throw_longjmp(int*, int)
218 ThrowLongJmp = M.getOrInsertFunction("__llvm_sjljeh_throw_longjmp",
219 Type::VoidTy, SBPTy, Type::IntTy,
222 // unsigned __llvm_sjljeh_try_catching_longjmp_exception(void **)
224 M.getOrInsertFunction("__llvm_sjljeh_try_catching_longjmp_exception",
225 Type::UIntTy, SBPPTy, (Type *)0);
227 // bool __llvm_sjljeh_is_longjmp_exception()
228 IsLJException = M.getOrInsertFunction("__llvm_sjljeh_is_longjmp_exception",
229 Type::BoolTy, (Type *)0);
231 // int __llvm_sjljeh_get_longjmp_value()
232 GetLJValue = M.getOrInsertFunction("__llvm_sjljeh_get_longjmp_value",
233 Type::IntTy, (Type *)0);
237 // IsTransformableFunction - Return true if the function name isn't one
238 // of the ones we don't want transformed. Currently, don't transform any
239 // "llvm.{setjmp,longjmp}" functions and none of the setjmp/longjmp error
240 // handling functions (beginning with __llvm_sjljeh_...they don't throw
242 bool LowerSetJmp::IsTransformableFunction(const std::string& Name) {
243 std::string SJLJEh("__llvm_sjljeh");
245 if (Name.size() > SJLJEh.size())
246 return std::string(Name.begin(), Name.begin() + SJLJEh.size()) != SJLJEh;
251 // TransformLongJmpCall - Transform a longjmp call into a call to the
252 // internal __llvm_sjljeh_throw_longjmp function. It then takes care of
253 // throwing the exception for us.
254 void LowerSetJmp::TransformLongJmpCall(CallInst* Inst)
256 const Type* SBPTy = PointerType::get(Type::SByteTy);
258 // Create the call to "__llvm_sjljeh_throw_longjmp". This takes the
259 // same parameters as "longjmp", except that the buffer is cast to a
260 // char*. It returns "void", so it doesn't need to replace any of
261 // Inst's uses and doesn't get a name.
262 CastInst* CI = new CastInst(Inst->getOperand(1), SBPTy, "LJBuf", Inst);
263 new CallInst(ThrowLongJmp, make_vector<Value*>(CI, Inst->getOperand(2), 0),
266 SwitchValuePair& SVP = SwitchValMap[Inst->getParent()->getParent()];
268 // If the function has a setjmp call in it (they are transformed first)
269 // we should branch to the basic block that determines if this longjmp
270 // is applicable here. Otherwise, issue an unwind.
272 new BranchInst(SVP.first->getParent(), Inst);
274 new UnwindInst(Inst);
276 // Remove all insts after the branch/unwind inst. Go from back to front to
277 // avoid replaceAllUsesWith if possible.
278 BasicBlock *BB = Inst->getParent();
279 Instruction *Removed;
281 Removed = &BB->back();
282 // If the removed instructions have any users, replace them now.
283 if (!Removed->use_empty())
284 Removed->replaceAllUsesWith(UndefValue::get(Removed->getType()));
285 Removed->eraseFromParent();
286 } while (Removed != Inst);
288 ++LongJmpsTransformed;
291 // GetSetJmpMap - Retrieve (create and initialize, if necessary) the
292 // setjmp map. This map is going to hold information about which setjmps
293 // were called (each setjmp gets its own number) and with which buffer it
294 // was called. There can be only one!
295 AllocaInst* LowerSetJmp::GetSetJmpMap(Function* Func)
297 if (SJMap[Func]) return SJMap[Func];
299 // Insert the setjmp map initialization before the first instruction in
301 Instruction* Inst = Func->getEntryBlock().begin();
302 assert(Inst && "Couldn't find even ONE instruction in entry block!");
304 // Fill in the alloca and call to initialize the SJ map.
305 const Type *SBPTy = PointerType::get(Type::SByteTy);
306 AllocaInst* Map = new AllocaInst(SBPTy, 0, "SJMap", Inst);
307 new CallInst(InitSJMap, make_vector<Value*>(Map, 0), "", Inst);
308 return SJMap[Func] = Map;
311 // GetRethrowBB - Only one rethrow basic block is needed per function.
312 // If this is a longjmp exception but not handled in this block, this BB
313 // performs the rethrow.
314 BasicBlock* LowerSetJmp::GetRethrowBB(Function* Func)
316 if (RethrowBBMap[Func]) return RethrowBBMap[Func];
318 // The basic block we're going to jump to if we need to rethrow the
320 BasicBlock* Rethrow = new BasicBlock("RethrowExcept", Func);
322 // Fill in the "Rethrow" BB with a call to rethrow the exception. This
323 // is the last instruction in the BB since at this point the runtime
324 // should exit this function and go to the next function.
325 new UnwindInst(Rethrow);
326 return RethrowBBMap[Func] = Rethrow;
329 // GetSJSwitch - Return the switch statement that controls which handler
330 // (if any) gets called and the value returned to that handler.
331 LowerSetJmp::SwitchValuePair LowerSetJmp::GetSJSwitch(Function* Func,
334 if (SwitchValMap[Func].first) return SwitchValMap[Func];
336 BasicBlock* LongJmpPre = new BasicBlock("LongJmpBlkPre", Func);
337 BasicBlock::InstListType& LongJmpPreIL = LongJmpPre->getInstList();
339 // Keep track of the preliminary basic block for some of the other
341 PrelimBBMap[Func] = LongJmpPre;
343 // Grab the exception.
345 CallInst(IsLJException, std::vector<Value*>(), "IsLJExcept");
346 LongJmpPreIL.push_back(Cond);
348 // The "decision basic block" gets the number associated with the
349 // setjmp call returning to switch on and the value returned by
351 BasicBlock* DecisionBB = new BasicBlock("LJDecisionBB", Func);
352 BasicBlock::InstListType& DecisionBBIL = DecisionBB->getInstList();
354 new BranchInst(DecisionBB, Rethrow, Cond, LongJmpPre);
356 // Fill in the "decision" basic block.
357 CallInst* LJVal = new CallInst(GetLJValue, std::vector<Value*>(), "LJVal");
358 DecisionBBIL.push_back(LJVal);
359 CallInst* SJNum = new
360 CallInst(TryCatchLJ, make_vector<Value*>(GetSetJmpMap(Func), 0), "SJNum");
361 DecisionBBIL.push_back(SJNum);
363 SwitchInst* SI = new SwitchInst(SJNum, Rethrow, 0, DecisionBB);
364 return SwitchValMap[Func] = SwitchValuePair(SI, LJVal);
367 // TransformSetJmpCall - The setjmp call is a bit trickier to transform.
368 // We're going to convert all setjmp calls to nops. Then all "call" and
369 // "invoke" instructions in the function are converted to "invoke" where
370 // the "except" branch is used when returning from a longjmp call.
371 void LowerSetJmp::TransformSetJmpCall(CallInst* Inst)
373 BasicBlock* ABlock = Inst->getParent();
374 Function* Func = ABlock->getParent();
376 // Add this setjmp to the setjmp map.
377 const Type* SBPTy = PointerType::get(Type::SByteTy);
378 CastInst* BufPtr = new CastInst(Inst->getOperand(1), SBPTy, "SBJmpBuf", Inst);
379 new CallInst(AddSJToMap,
380 make_vector<Value*>(GetSetJmpMap(Func), BufPtr,
381 ConstantUInt::get(Type::UIntTy,
382 SetJmpIDMap[Func]++), 0),
385 // We are guaranteed that there are no values live across basic blocks
386 // (because we are "not in SSA form" yet), but there can still be values live
387 // in basic blocks. Because of this, splitting the setjmp block can cause
388 // values above the setjmp to not dominate uses which are after the setjmp
389 // call. For all of these occasions, we must spill the value to the stack.
391 std::set<Instruction*> InstrsAfterCall;
393 // The call is probably very close to the end of the basic block, for the
394 // common usage pattern of: 'if (setjmp(...))', so keep track of the
395 // instructions after the call.
396 for (BasicBlock::iterator I = ++BasicBlock::iterator(Inst), E = ABlock->end();
398 InstrsAfterCall.insert(I);
400 for (BasicBlock::iterator II = ABlock->begin();
401 II != BasicBlock::iterator(Inst); ++II)
402 // Loop over all of the uses of instruction. If any of them are after the
403 // call, "spill" the value to the stack.
404 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
406 if (cast<Instruction>(*UI)->getParent() != ABlock ||
407 InstrsAfterCall.count(cast<Instruction>(*UI))) {
408 DemoteRegToStack(*II);
411 InstrsAfterCall.clear();
413 // Change the setjmp call into a branch statement. We'll remove the
414 // setjmp call in a little bit. No worries.
415 BasicBlock* SetJmpContBlock = ABlock->splitBasicBlock(Inst);
416 assert(SetJmpContBlock && "Couldn't split setjmp BB!!");
418 SetJmpContBlock->setName(ABlock->getName()+"SetJmpCont");
420 // Add the SetJmpContBlock to the set of blocks reachable from a setjmp.
421 DFSBlocks.insert(SetJmpContBlock);
423 // This PHI node will be in the new block created from the
424 // splitBasicBlock call.
425 PHINode* PHI = new PHINode(Type::IntTy, "SetJmpReturn", Inst);
427 // Coming from a call to setjmp, the return is 0.
428 PHI->addIncoming(ConstantInt::getNullValue(Type::IntTy), ABlock);
430 // Add the case for this setjmp's number...
431 SwitchValuePair SVP = GetSJSwitch(Func, GetRethrowBB(Func));
432 SVP.first->addCase(ConstantUInt::get(Type::UIntTy, SetJmpIDMap[Func] - 1),
435 // Value coming from the handling of the exception.
436 PHI->addIncoming(SVP.second, SVP.second->getParent());
438 // Replace all uses of this instruction with the PHI node created by
439 // the eradication of setjmp.
440 Inst->replaceAllUsesWith(PHI);
441 Inst->getParent()->getInstList().erase(Inst);
443 ++SetJmpsTransformed;
446 // visitCallInst - This converts all LLVM call instructions into invoke
447 // instructions. The except part of the invoke goes to the "LongJmpBlkPre"
448 // that grabs the exception and proceeds to determine if it's a longjmp
450 void LowerSetJmp::visitCallInst(CallInst& CI)
452 if (CI.getCalledFunction())
453 if (!IsTransformableFunction(CI.getCalledFunction()->getName()) ||
454 CI.getCalledFunction()->isIntrinsic()) return;
456 BasicBlock* OldBB = CI.getParent();
458 // If not reachable from a setjmp call, don't transform.
459 if (!DFSBlocks.count(OldBB)) return;
461 BasicBlock* NewBB = OldBB->splitBasicBlock(CI);
462 assert(NewBB && "Couldn't split BB of \"call\" instruction!!");
463 DFSBlocks.insert(NewBB);
464 NewBB->setName("Call2Invoke");
466 Function* Func = OldBB->getParent();
468 // Construct the new "invoke" instruction.
469 TerminatorInst* Term = OldBB->getTerminator();
470 std::vector<Value*> Params(CI.op_begin() + 1, CI.op_end());
472 InvokeInst(CI.getCalledValue(), NewBB, PrelimBBMap[Func],
473 Params, CI.getName(), Term);
475 // Replace the old call inst with the invoke inst and remove the call.
476 CI.replaceAllUsesWith(II);
477 CI.getParent()->getInstList().erase(&CI);
479 // The old terminator is useless now that we have the invoke inst.
480 Term->getParent()->getInstList().erase(Term);
484 // visitInvokeInst - Converting the "invoke" instruction is fairly
485 // straight-forward. The old exception part is replaced by a query asking
486 // if this is a longjmp exception. If it is, then it goes to the longjmp
487 // exception blocks. Otherwise, control is passed the old exception.
488 void LowerSetJmp::visitInvokeInst(InvokeInst& II)
490 if (II.getCalledFunction())
491 if (!IsTransformableFunction(II.getCalledFunction()->getName()) ||
492 II.getCalledFunction()->isIntrinsic()) return;
494 BasicBlock* BB = II.getParent();
496 // If not reachable from a setjmp call, don't transform.
497 if (!DFSBlocks.count(BB)) return;
499 BasicBlock* NormalBB = II.getNormalDest();
500 BasicBlock* ExceptBB = II.getUnwindDest();
502 Function* Func = BB->getParent();
503 BasicBlock* NewExceptBB = new BasicBlock("InvokeExcept", Func);
504 BasicBlock::InstListType& InstList = NewExceptBB->getInstList();
506 // If this is a longjmp exception, then branch to the preliminary BB of
507 // the longjmp exception handling. Otherwise, go to the old exception.
508 CallInst* IsLJExcept = new
509 CallInst(IsLJException, std::vector<Value*>(), "IsLJExcept");
510 InstList.push_back(IsLJExcept);
512 new BranchInst(PrelimBBMap[Func], ExceptBB, IsLJExcept, NewExceptBB);
514 II.setUnwindDest(NewExceptBB);
515 ++InvokesTransformed;
518 // visitReturnInst - We want to destroy the setjmp map upon exit from the
520 void LowerSetJmp::visitReturnInst(ReturnInst &RI) {
521 Function* Func = RI.getParent()->getParent();
522 new CallInst(DestroySJMap, make_vector<Value*>(GetSetJmpMap(Func), 0),
526 // visitUnwindInst - We want to destroy the setjmp map upon exit from the
528 void LowerSetJmp::visitUnwindInst(UnwindInst &UI) {
529 Function* Func = UI.getParent()->getParent();
530 new CallInst(DestroySJMap, make_vector<Value*>(GetSetJmpMap(Func), 0),
534 ModulePass *llvm::createLowerSetJmpPass() {
535 return new LowerSetJmp();