1 //===-- LowerGC.cpp - Provide GC support for targets that don't -----------===//
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 lowering for the llvm.gc* intrinsics for targets that do
11 // not natively support them (which includes the C backend). Note that the code
12 // generated is not as efficient as it would be for targets that natively
13 // support the GC intrinsics, but it is useful for getting new targets
14 // up-and-running quickly.
16 // This pass implements the code transformation described in this paper:
17 // "Accurate Garbage Collection in an Uncooperative Environment"
18 // Fergus Henderson, ISMM, 2002
20 //===----------------------------------------------------------------------===//
22 #define DEBUG_TYPE "lowergc"
23 #include "llvm/Transforms/Scalar.h"
24 #include "llvm/Constants.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/Module.h"
28 #include "llvm/Pass.h"
32 class LowerGC : public FunctionPass {
33 /// GCRootInt, GCReadInt, GCWriteInt - The function prototypes for the
34 /// llvm.gcread/llvm.gcwrite/llvm.gcroot intrinsics.
35 Function *GCRootInt, *GCReadInt, *GCWriteInt;
37 /// GCRead/GCWrite - These are the functions provided by the garbage
38 /// collector for read/write barriers.
39 Function *GCRead, *GCWrite;
41 /// RootChain - This is the global linked-list that contains the chain of GC
43 GlobalVariable *RootChain;
45 /// MainRootRecordType - This is the type for a function root entry if it
47 const Type *MainRootRecordType;
49 LowerGC() : GCRootInt(0), GCReadInt(0), GCWriteInt(0),
50 GCRead(0), GCWrite(0), RootChain(0), MainRootRecordType(0) {}
51 virtual bool doInitialization(Module &M);
52 virtual bool runOnFunction(Function &F);
55 const StructType *getRootRecordType(unsigned NumRoots);
59 X("lowergc", "Lower GC intrinsics, for GCless code generators");
62 /// createLowerGCPass - This function returns an instance of the "lowergc"
63 /// pass, which lowers garbage collection intrinsics to normal LLVM code.
64 FunctionPass *llvm::createLowerGCPass() {
68 /// getRootRecordType - This function creates and returns the type for a root
69 /// record containing 'NumRoots' roots.
70 const StructType *LowerGC::getRootRecordType(unsigned NumRoots) {
71 // Build a struct that is a type used for meta-data/root pairs.
72 std::vector<const Type *> ST;
73 ST.push_back(GCRootInt->getFunctionType()->getParamType(0));
74 ST.push_back(GCRootInt->getFunctionType()->getParamType(1));
75 StructType *PairTy = StructType::get(ST);
77 // Build the array of pairs.
78 ArrayType *PairArrTy = ArrayType::get(PairTy, NumRoots);
80 // Now build the recursive list type.
81 PATypeHolder RootListH =
82 MainRootRecordType ? (Type*)MainRootRecordType : (Type*)OpaqueType::get();
84 ST.push_back(PointerType::get(RootListH)); // Prev pointer
85 ST.push_back(Type::UIntTy); // NumElements in array
86 ST.push_back(PairArrTy); // The pairs
87 StructType *RootList = StructType::get(ST);
88 if (MainRootRecordType)
91 assert(NumRoots == 0 && "The main struct type should have zero entries!");
92 cast<OpaqueType>((Type*)RootListH.get())->refineAbstractTypeTo(RootList);
93 MainRootRecordType = RootListH;
94 return cast<StructType>(RootListH.get());
97 /// doInitialization - If this module uses the GC intrinsics, find them now. If
98 /// not, this pass does not do anything.
99 bool LowerGC::doInitialization(Module &M) {
100 GCRootInt = M.getNamedFunction("llvm.gcroot");
101 GCReadInt = M.getNamedFunction("llvm.gcread");
102 GCWriteInt = M.getNamedFunction("llvm.gcwrite");
103 if (!GCRootInt && !GCReadInt && !GCWriteInt) return false;
105 PointerType *VoidPtr = PointerType::get(Type::SByteTy);
106 PointerType *VoidPtrPtr = PointerType::get(VoidPtr);
108 // If the program is using read/write barriers, find the implementations of
109 // them from the GC runtime library.
110 if (GCReadInt) // Make: sbyte* %llvm_gc_read(sbyte**)
111 GCRead = M.getOrInsertFunction("llvm_gc_read", VoidPtr, VoidPtr, VoidPtrPtr,
113 if (GCWriteInt) // Make: void %llvm_gc_write(sbyte*, sbyte**)
114 GCWrite = M.getOrInsertFunction("llvm_gc_write", Type::VoidTy,
115 VoidPtr, VoidPtr, VoidPtrPtr, (Type *)0);
117 // If the program has GC roots, get or create the global root list.
119 const StructType *RootListTy = getRootRecordType(0);
120 const Type *PRLTy = PointerType::get(RootListTy);
121 M.addTypeName("llvm_gc_root_ty", RootListTy);
123 // Get the root chain if it already exists.
124 RootChain = M.getGlobalVariable("llvm_gc_root_chain", PRLTy);
125 if (RootChain == 0) {
126 // If the root chain does not exist, insert a new one with linkonce
128 RootChain = new GlobalVariable(PRLTy, false,
129 GlobalValue::LinkOnceLinkage,
130 Constant::getNullValue(PRLTy),
131 "llvm_gc_root_chain", &M);
132 } else if (RootChain->hasExternalLinkage() && RootChain->isExternal()) {
133 RootChain->setInitializer(Constant::getNullValue(PRLTy));
134 RootChain->setLinkage(GlobalValue::LinkOnceLinkage);
140 /// Coerce - If the specified operand number of the specified instruction does
141 /// not have the specified type, insert a cast.
142 static void Coerce(Instruction *I, unsigned OpNum, Type *Ty) {
143 if (I->getOperand(OpNum)->getType() != Ty) {
144 if (Constant *C = dyn_cast<Constant>(I->getOperand(OpNum)))
145 I->setOperand(OpNum, ConstantExpr::getCast(C, Ty));
147 CastInst *CI = new CastInst(I->getOperand(OpNum), Ty, "", I);
148 I->setOperand(OpNum, CI);
153 /// runOnFunction - If the program is using GC intrinsics, replace any
154 /// read/write intrinsics with the appropriate read/write barrier calls, then
155 /// inline them. Finally, build the data structures for
156 bool LowerGC::runOnFunction(Function &F) {
157 // Quick exit for programs that are not using GC mechanisms.
158 if (!GCRootInt && !GCReadInt && !GCWriteInt) return false;
160 PointerType *VoidPtr = PointerType::get(Type::SByteTy);
161 PointerType *VoidPtrPtr = PointerType::get(VoidPtr);
163 // If there are read/write barriers in the program, perform a quick pass over
164 // the function eliminating them. While we are at it, remember where we see
165 // calls to llvm.gcroot.
166 std::vector<CallInst*> GCRoots;
167 std::vector<CallInst*> NormalCalls;
169 bool MadeChange = false;
170 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
171 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;)
172 if (CallInst *CI = dyn_cast<CallInst>(II++)) {
173 if (!CI->getCalledFunction() ||
174 !CI->getCalledFunction()->getIntrinsicID())
175 NormalCalls.push_back(CI); // Remember all normal function calls.
177 if (Function *F = CI->getCalledFunction())
179 GCRoots.push_back(CI);
180 else if (F == GCReadInt || F == GCWriteInt) {
181 if (F == GCWriteInt) {
182 // Change a llvm.gcwrite call to call llvm_gc_write instead.
183 CI->setOperand(0, GCWrite);
184 // Insert casts of the operands as needed.
185 Coerce(CI, 1, VoidPtr);
186 Coerce(CI, 2, VoidPtr);
187 Coerce(CI, 3, VoidPtrPtr);
189 Coerce(CI, 1, VoidPtr);
190 Coerce(CI, 2, VoidPtrPtr);
191 if (CI->getType() == VoidPtr) {
192 CI->setOperand(0, GCRead);
194 // Create a whole new call to replace the old one.
195 CallInst *NC = new CallInst(GCRead, CI->getOperand(1),
198 Value *NV = new CastInst(NC, CI->getType(), "", CI);
199 CI->replaceAllUsesWith(NV);
200 BB->getInstList().erase(CI);
209 // If there are no GC roots in this function, then there is no need to create
210 // a GC list record for it.
211 if (GCRoots.empty()) return MadeChange;
213 // Okay, there are GC roots in this function. On entry to the function, add a
214 // record to the llvm_gc_root_chain, and remove it on exit.
216 // Create the alloca, and zero it out.
217 const StructType *RootListTy = getRootRecordType(GCRoots.size());
218 AllocaInst *AI = new AllocaInst(RootListTy, 0, "gcroots", F.begin()->begin());
220 // Insert the memset call after all of the allocas in the function.
221 BasicBlock::iterator IP = AI;
222 while (isa<AllocaInst>(IP)) ++IP;
224 Constant *Zero = ConstantUInt::get(Type::UIntTy, 0);
225 Constant *One = ConstantUInt::get(Type::UIntTy, 1);
227 // Get a pointer to the prev pointer.
228 std::vector<Value*> Par;
231 Value *PrevPtrPtr = new GetElementPtrInst(AI, Par, "prevptrptr", IP);
233 // Load the previous pointer.
234 Value *PrevPtr = new LoadInst(RootChain, "prevptr", IP);
235 // Store the previous pointer into the prevptrptr
236 new StoreInst(PrevPtr, PrevPtrPtr, IP);
238 // Set the number of elements in this record.
239 Par[1] = ConstantUInt::get(Type::UIntTy, 1);
240 Value *NumEltsPtr = new GetElementPtrInst(AI, Par, "numeltsptr", IP);
241 new StoreInst(ConstantUInt::get(Type::UIntTy, GCRoots.size()), NumEltsPtr,IP);
243 Par[1] = ConstantUInt::get(Type::UIntTy, 2);
246 const PointerType *PtrLocTy =
247 cast<PointerType>(GCRootInt->getFunctionType()->getParamType(0));
248 Constant *Null = ConstantPointerNull::get(PtrLocTy);
250 // Initialize all of the gcroot records now, and eliminate them as we go.
251 for (unsigned i = 0, e = GCRoots.size(); i != e; ++i) {
252 // Initialize the meta-data pointer.
253 Par[2] = ConstantUInt::get(Type::UIntTy, i);
255 Value *MetaDataPtr = new GetElementPtrInst(AI, Par, "MetaDataPtr", IP);
256 assert(isa<Constant>(GCRoots[i]->getOperand(2)) && "Must be a constant");
257 new StoreInst(GCRoots[i]->getOperand(2), MetaDataPtr, IP);
259 // Initialize the root pointer to null on entry to the function.
261 Value *RootPtrPtr = new GetElementPtrInst(AI, Par, "RootEntPtr", IP);
262 new StoreInst(Null, RootPtrPtr, IP);
264 // Each occurrance of the llvm.gcroot intrinsic now turns into an
265 // initialization of the slot with the address and a zeroing out of the
266 // address specified.
267 new StoreInst(Constant::getNullValue(PtrLocTy->getElementType()),
268 GCRoots[i]->getOperand(1), GCRoots[i]);
269 new StoreInst(GCRoots[i]->getOperand(1), RootPtrPtr, GCRoots[i]);
270 GCRoots[i]->getParent()->getInstList().erase(GCRoots[i]);
273 // Now that the record is all initialized, store the pointer into the global
275 Value *C = new CastInst(AI, PointerType::get(MainRootRecordType), "", IP);
276 new StoreInst(C, RootChain, IP);
278 // On exit from the function we have to remove the entry from the GC root
279 // chain. Doing this is straight-forward for return and unwind instructions:
280 // just insert the appropriate copy.
281 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
282 if (isa<UnwindInst>(BB->getTerminator()) ||
283 isa<ReturnInst>(BB->getTerminator())) {
284 // We could reuse the PrevPtr loaded on entry to the function, but this
285 // would make the value live for the whole function, which is probably a
286 // bad idea. Just reload the value out of our stack entry.
287 PrevPtr = new LoadInst(PrevPtrPtr, "prevptr", BB->getTerminator());
288 new StoreInst(PrevPtr, RootChain, BB->getTerminator());
291 // If an exception is thrown from a callee we have to make sure to
292 // unconditionally take the record off the stack. For this reason, we turn
293 // all call instructions into invoke whose cleanup pops the entry off the
294 // stack. We only insert one cleanup block, which is shared by all invokes.
295 if (!NormalCalls.empty()) {
296 // Create the shared cleanup block.
297 BasicBlock *Cleanup = new BasicBlock("gc_cleanup", &F);
298 UnwindInst *UI = new UnwindInst(Cleanup);
299 PrevPtr = new LoadInst(PrevPtrPtr, "prevptr", UI);
300 new StoreInst(PrevPtr, RootChain, UI);
302 // Loop over all of the function calls, turning them into invokes.
303 while (!NormalCalls.empty()) {
304 CallInst *CI = NormalCalls.back();
305 BasicBlock *CBB = CI->getParent();
306 NormalCalls.pop_back();
308 // Split the basic block containing the function call.
309 BasicBlock *NewBB = CBB->splitBasicBlock(CI, CBB->getName()+".cont");
311 // Remove the unconditional branch inserted at the end of the CBB.
312 CBB->getInstList().pop_back();
313 NewBB->getInstList().remove(CI);
315 // Create a new invoke instruction.
316 Value *II = new InvokeInst(CI->getCalledValue(), NewBB, Cleanup,
317 std::vector<Value*>(CI->op_begin()+1,
320 CI->replaceAllUsesWith(II);