1 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
3 // This file implements inlining of a function into a call site, resolving
4 // parameters and the return value as appropriate.
6 // FIXME: This pass should transform alloca instructions in the called function
7 // into malloc/free pairs! Or perhaps it should refuse to inline them!
9 //===----------------------------------------------------------------------===//
11 #include "llvm/Transforms/Utils/Cloning.h"
12 #include "llvm/Constant.h"
13 #include "llvm/DerivedTypes.h"
14 #include "llvm/Module.h"
15 #include "llvm/Instructions.h"
16 #include "llvm/Intrinsics.h"
17 #include "llvm/Support/CallSite.h"
18 #include "llvm/Transforms/Utils/Local.h"
20 bool InlineFunction(CallInst *CI) { return InlineFunction(CallSite(CI)); }
21 bool InlineFunction(InvokeInst *II) { return InlineFunction(CallSite(II)); }
23 // InlineFunction - This function inlines the called function into the basic
24 // block of the caller. This returns false if it is not possible to inline this
25 // call. The program is still in a well defined state if this occurs though.
27 // Note that this only does one level of inlining. For example, if the
28 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
29 // exists in the instruction stream. Similiarly this will inline a recursive
30 // function by one level.
32 bool InlineFunction(CallSite CS) {
33 Instruction *TheCall = CS.getInstruction();
34 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
35 "Instruction not in function!");
37 const Function *CalledFunc = CS.getCalledFunction();
38 if (CalledFunc == 0 || // Can't inline external function or indirect
39 CalledFunc->isExternal() || // call, or call to a vararg function!
40 CalledFunc->getFunctionType()->isVarArg()) return false;
42 BasicBlock *OrigBB = TheCall->getParent();
43 Function *Caller = OrigBB->getParent();
45 // We want to clone the entire callee function into the whole between the
46 // "starter" and "ender" blocks. How we accomplish this depends on whether
47 // this is an invoke instruction or a call instruction.
49 BasicBlock *InvokeDest = 0; // Exception handling destination
50 BasicBlock *AfterCallBB;
51 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
52 AfterCallBB = II->getNormalDest();
53 InvokeDest = II->getExceptionalDest();
55 // Add an unconditional branch to make this look like the CallInst case...
56 new BranchInst(AfterCallBB, TheCall);
58 // Remove (unlink) the InvokeInst from the function...
59 OrigBB->getInstList().remove(TheCall);
60 } else { // It's a call
61 // If this is a call instruction, we need to split the basic block that the
64 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
65 CalledFunc->getName()+".entry");
66 // Remove (unlink) the CallInst from the function...
67 AfterCallBB->getInstList().remove(TheCall);
70 // If we have a return value generated by this call, convert it into a PHI
71 // node that gets values from each of the old RET instructions in the original
75 if (!TheCall->use_empty()) {
76 // The PHI node should go at the front of the new basic block to merge all
77 // possible incoming values.
79 PHI = new PHINode(CalledFunc->getReturnType(), TheCall->getName(),
80 AfterCallBB->begin());
82 // Anything that used the result of the function call should now use the PHI
83 // node as their operand.
85 TheCall->replaceAllUsesWith(PHI);
88 // Get an iterator to the last basic block in the function, which will have
89 // the new function inlined after it.
91 Function::iterator LastBlock = &Caller->back();
93 // Calculate the vector of arguments to pass into the function cloner...
94 std::map<const Value*, Value*> ValueMap;
95 assert(std::distance(CalledFunc->abegin(), CalledFunc->aend()) ==
96 std::distance(CS.arg_begin(), CS.arg_end()) &&
97 "No varargs calls can be inlined!");
99 CallSite::arg_iterator AI = CS.arg_begin();
100 for (Function::const_aiterator I = CalledFunc->abegin(), E=CalledFunc->aend();
104 // Since we are now done with the Call/Invoke, we can delete it.
107 // Make a vector to capture the return instructions in the cloned function...
108 std::vector<ReturnInst*> Returns;
110 // Populate the value map with all of the globals in the program.
111 // FIXME: This should be the default for CloneFunctionInto!
112 Module &M = *Caller->getParent();
113 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
115 for (Module::giterator I = M.gbegin(), E = M.gend(); I != E; ++I)
118 // Do all of the hard part of cloning the callee into the caller...
119 CloneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i");
121 // Loop over all of the return instructions, turning them into unconditional
122 // branches to the merge point now...
123 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
124 ReturnInst *RI = Returns[i];
125 BasicBlock *BB = RI->getParent();
127 // Add a branch to the merge point where the PHI node lives if it exists.
128 new BranchInst(AfterCallBB, RI);
130 if (PHI) { // The PHI node should include this value!
131 assert(RI->getReturnValue() && "Ret should have value!");
132 assert(RI->getReturnValue()->getType() == PHI->getType() &&
133 "Ret value not consistent in function!");
134 PHI->addIncoming(RI->getReturnValue(), BB);
137 // Delete the return instruction now
138 BB->getInstList().erase(RI);
141 // Check to see if the PHI node only has one argument. This is a common
142 // case resulting from there only being a single return instruction in the
143 // function call. Because this is so common, eliminate the PHI node.
145 if (PHI && PHI->getNumIncomingValues() == 1) {
146 PHI->replaceAllUsesWith(PHI->getIncomingValue(0));
147 PHI->getParent()->getInstList().erase(PHI);
150 // Change the branch that used to go to AfterCallBB to branch to the first
151 // basic block of the inlined function.
153 TerminatorInst *Br = OrigBB->getTerminator();
154 assert(Br && Br->getOpcode() == Instruction::Br &&
155 "splitBasicBlock broken!");
156 Br->setOperand(0, ++LastBlock);
158 // If there are any alloca instructions in the block that used to be the entry
159 // block for the callee, move them to the entry block of the caller. First
160 // calculate which instruction they should be inserted before. We insert the
161 // instructions at the end of the current alloca list.
163 if (isa<AllocaInst>(LastBlock->begin())) {
164 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
165 while (isa<AllocaInst>(InsertPoint)) ++InsertPoint;
167 for (BasicBlock::iterator I = LastBlock->begin(), E = LastBlock->end();
169 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) {
170 ++I; // Move to the next instruction
171 LastBlock->getInstList().remove(AI);
172 Caller->front().getInstList().insert(InsertPoint, AI);
178 // If we just inlined a call due to an invoke instruction, scan the inlined
179 // function checking for function calls that should now be made into invoke
180 // instructions, and for llvm.exc.rethrow()'s which should be turned into
183 for (Function::iterator BB = LastBlock, E = Caller->end(); BB != E; ++BB)
184 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
185 // We only need to check for function calls: inlined invoke instructions
186 // require no special handling...
187 if (CallInst *CI = dyn_cast<CallInst>(I)) {
188 // FIXME: this should use annotations of the LLVM functions themselves
189 // to determine whether or not the function can throw.
190 bool ShouldInvokify = true;
192 if (Function *F = CI->getCalledFunction())
193 if (unsigned ID = F->getIntrinsicID())
194 if (ID == LLVMIntrinsic::unwind) {
195 // llvm.unwind requires special handling when it gets inlined
196 // into an invoke site. Once this happens, we know that the
197 // unwind would cause a control transfer to the invoke exception
198 // destination, so we can transform it into a direct branch to
199 // the exception destination.
200 BranchInst *BI = new BranchInst(InvokeDest, CI);
202 // Note that since any instructions after the rethrow/branch are
203 // dead, we must delete them now (otherwise the terminator we
204 // just inserted wouldn't be at the end of the basic block!)
205 BasicBlock *CurBB = BB;
206 while (&CurBB->back() != BI) {
207 Instruction *I = &CurBB->back();
209 I->replaceAllUsesWith(Constant::getNullValue(I->getType()));
210 CurBB->getInstList().pop_back();
213 break; // Done with this basic block!
214 } else if (ID == LLVMIntrinsic::exc_setcurrent ||
215 ID == LLVMIntrinsic::exc_getcurrent) {
216 ShouldInvokify = false; // Not correct to invokify exc.throw!
219 // If we should convert this function into an invoke instruction, do
221 if (ShouldInvokify) {
222 // First, split the basic block...
223 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
225 // Next, create the new invoke instruction, inserting it at the end
226 // of the old basic block.
227 new InvokeInst(CI->getCalledValue(), Split, InvokeDest,
228 std::vector<Value*>(CI->op_begin()+1, CI->op_end()),
229 CI->getName(), BB->getTerminator());
231 // Delete the unconditional branch inserted by splitBasicBlock
232 BB->getInstList().pop_back();
233 Split->getInstList().pop_front(); // Delete the original call
235 // This basic block is now complete, start scanning the next one.
245 // Now that the function is correct, make it a little bit nicer. In
246 // particular, move the basic blocks inserted from the end of the function
247 // into the space made by splitting the source basic block.
249 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
250 LastBlock, Caller->end());
252 // We should always be able to fold the entry block of the function into the
253 // single predecessor of the block...
254 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
255 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
256 SimplifyCFG(CalleeEntry);
258 // Okay, continue the CFG cleanup. It's often the case that there is only a
259 // single return instruction in the callee function. If this is the case,
260 // then we have an unconditional branch from the return block to the
261 // 'AfterCallBB'. Check for this case, and eliminate the branch is possible.
262 SimplifyCFG(AfterCallBB);