1 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
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 inlining of a function into a call site, resolving
11 // parameters and the return value as appropriate.
13 // FIXME: This pass should transform alloca instructions in the called function
14 // into alloca/dealloca pairs! Or perhaps it should refuse to inline them!
16 //===----------------------------------------------------------------------===//
18 #include "llvm/Transforms/Utils/Cloning.h"
19 #include "llvm/Constant.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Module.h"
22 #include "llvm/Instructions.h"
23 #include "llvm/Intrinsics.h"
24 #include "llvm/Support/CallSite.h"
25 #include "llvm/Transforms/Utils/Local.h"
28 bool llvm::InlineFunction(CallInst *CI) { return InlineFunction(CallSite(CI)); }
29 bool llvm::InlineFunction(InvokeInst *II) {return InlineFunction(CallSite(II));}
31 // InlineFunction - This function inlines the called function into the basic
32 // block of the caller. This returns false if it is not possible to inline this
33 // call. The program is still in a well defined state if this occurs though.
35 // Note that this only does one level of inlining. For example, if the
36 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
37 // exists in the instruction stream. Similiarly this will inline a recursive
38 // function by one level.
40 bool llvm::InlineFunction(CallSite CS) {
41 Instruction *TheCall = CS.getInstruction();
42 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
43 "Instruction not in function!");
45 const Function *CalledFunc = CS.getCalledFunction();
46 if (CalledFunc == 0 || // Can't inline external function or indirect
47 CalledFunc->isExternal() || // call, or call to a vararg function!
48 CalledFunc->getFunctionType()->isVarArg()) return false;
50 BasicBlock *OrigBB = TheCall->getParent();
51 Function *Caller = OrigBB->getParent();
53 // Get an iterator to the last basic block in the function, which will have
54 // the new function inlined after it.
56 Function::iterator LastBlock = &Caller->back();
58 // Make sure to capture all of the return instructions from the cloned
60 std::vector<ReturnInst*> Returns;
61 { // Scope to destroy ValueMap after cloning.
62 // Calculate the vector of arguments to pass into the function cloner...
63 std::map<const Value*, Value*> ValueMap;
64 assert(std::distance(CalledFunc->abegin(), CalledFunc->aend()) ==
65 std::distance(CS.arg_begin(), CS.arg_end()) &&
66 "No varargs calls can be inlined!");
68 CallSite::arg_iterator AI = CS.arg_begin();
69 for (Function::const_aiterator I = CalledFunc->abegin(),
70 E = CalledFunc->aend(); I != E; ++I, ++AI)
73 // Clone the entire body of the callee into the caller.
74 CloneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i");
77 // Remember the first block that is newly cloned over.
78 Function::iterator FirstNewBlock = LastBlock; ++FirstNewBlock;
80 // If there are any alloca instructions in the block that used to be the entry
81 // block for the callee, move them to the entry block of the caller. First
82 // calculate which instruction they should be inserted before. We insert the
83 // instructions at the end of the current alloca list.
85 if (isa<AllocaInst>(FirstNewBlock->begin())) {
86 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
87 for (BasicBlock::iterator I = FirstNewBlock->begin(),
88 E = FirstNewBlock->end(); I != E; )
89 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++))
90 if (isa<Constant>(AI->getArraySize())) {
91 // Scan for the block of allocas that we can move over.
92 while (isa<AllocaInst>(I) &&
93 isa<Constant>(cast<AllocaInst>(I)->getArraySize()))
96 // Transfer all of the allocas over in a block. Using splice means
97 // that they instructions aren't removed from the symbol table, then
99 Caller->front().getInstList().splice(InsertPoint,
100 FirstNewBlock->getInstList(),
105 // If we are inlining for an invoke instruction, we must make sure to rewrite
106 // any inlined 'unwind' instructions into branches to the invoke exception
107 // destination, and call instructions into invoke instructions.
108 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
109 BasicBlock *InvokeDest = II->getUnwindDest();
110 std::vector<Value*> InvokeDestPHIValues;
112 // If there are PHI nodes in the exceptional destination block, we need to
113 // keep track of which values came into them from this invoke, then remove
114 // the entry for this block.
115 for (BasicBlock::iterator I = InvokeDest->begin();
116 PHINode *PN = dyn_cast<PHINode>(I); ++I)
117 // Save the value to use for this edge...
118 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(OrigBB));
120 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
122 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
123 // We only need to check for function calls: inlined invoke instructions
124 // require no special handling...
125 if (CallInst *CI = dyn_cast<CallInst>(I)) {
126 // Convert this function call into an invoke instruction...
128 // First, split the basic block...
129 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
131 // Next, create the new invoke instruction, inserting it at the end
132 // of the old basic block.
134 new InvokeInst(CI->getCalledValue(), Split, InvokeDest,
135 std::vector<Value*>(CI->op_begin()+1, CI->op_end()),
136 CI->getName(), BB->getTerminator());
138 // Make sure that anything using the call now uses the invoke!
139 CI->replaceAllUsesWith(II);
141 // Delete the unconditional branch inserted by splitBasicBlock
142 BB->getInstList().pop_back();
143 Split->getInstList().pop_front(); // Delete the original call
145 // Update any PHI nodes in the exceptional block to indicate that
146 // there is now a new entry in them.
148 for (BasicBlock::iterator I = InvokeDest->begin();
149 PHINode *PN = dyn_cast<PHINode>(I); ++I, ++i)
150 PN->addIncoming(InvokeDestPHIValues[i], BB);
152 // This basic block is now complete, start scanning the next one.
159 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
160 // An UnwindInst requires special handling when it gets inlined into an
161 // invoke site. Once this happens, we know that the unwind would cause
162 // a control transfer to the invoke exception destination, so we can
163 // transform it into a direct branch to the exception destination.
164 new BranchInst(InvokeDest, UI);
166 // Delete the unwind instruction!
167 UI->getParent()->getInstList().pop_back();
169 // Update any PHI nodes in the exceptional block to indicate that
170 // there is now a new entry in them.
172 for (BasicBlock::iterator I = InvokeDest->begin();
173 PHINode *PN = dyn_cast<PHINode>(I); ++I, ++i)
174 PN->addIncoming(InvokeDestPHIValues[i], BB);
178 // Now that everything is happy, we have one final detail. The PHI nodes in
179 // the exception destination block still have entries due to the original
180 // invoke instruction. Eliminate these entries (which might even delete the
182 InvokeDest->removePredecessor(II->getParent());
185 // If we cloned in _exactly one_ basic block, and if that block ends in a
186 // return instruction, we splice the body of the inlined callee directly into
187 // the calling basic block.
188 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
189 // Move all of the instructions right before the call.
190 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
191 FirstNewBlock->begin(), FirstNewBlock->end());
192 // Remove the cloned basic block.
193 Caller->getBasicBlockList().pop_back();
195 // If the call site was an invoke instruction, add a branch to the normal
197 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
198 new BranchInst(II->getNormalDest(), TheCall);
200 // If the return instruction returned a value, replace uses of the call with
201 // uses of the returned value.
202 if (!TheCall->use_empty())
203 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
205 // Since we are now done with the Call/Invoke, we can delete it.
206 TheCall->getParent()->getInstList().erase(TheCall);
208 // Since we are now done with the return instruction, delete it also.
209 Returns[0]->getParent()->getInstList().erase(Returns[0]);
211 // We are now done with the inlining.
215 // Otherwise, we have the normal case, of more than one block to inline or
216 // multiple return sites.
218 // We want to clone the entire callee function into the hole between the
219 // "starter" and "ender" blocks. How we accomplish this depends on whether
220 // this is an invoke instruction or a call instruction.
221 BasicBlock *AfterCallBB;
222 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
224 // Add an unconditional branch to make this look like the CallInst case...
225 BranchInst *NewBr = new BranchInst(II->getNormalDest(), TheCall);
227 // Split the basic block. This guarantees that no PHI nodes will have to be
228 // updated due to new incoming edges, and make the invoke case more
229 // symmetric to the call case.
230 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
231 CalledFunc->getName()+".entry");
233 } else { // It's a call
234 // If this is a call instruction, we need to split the basic block that
235 // the call lives in.
237 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
238 CalledFunc->getName()+".entry");
241 // Change the branch that used to go to AfterCallBB to branch to the first
242 // basic block of the inlined function.
244 TerminatorInst *Br = OrigBB->getTerminator();
245 assert(Br && Br->getOpcode() == Instruction::Br &&
246 "splitBasicBlock broken!");
247 Br->setOperand(0, FirstNewBlock);
250 // Now that the function is correct, make it a little bit nicer. In
251 // particular, move the basic blocks inserted from the end of the function
252 // into the space made by splitting the source basic block.
254 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
255 FirstNewBlock, Caller->end());
257 // Handle all of the return instructions that we just cloned in, and eliminate
258 // any users of the original call/invoke instruction.
259 if (Returns.size() > 1) {
260 // The PHI node should go at the front of the new basic block to merge all
261 // possible incoming values.
264 if (!TheCall->use_empty()) {
265 PHI = new PHINode(CalledFunc->getReturnType(),
266 TheCall->getName(), AfterCallBB->begin());
268 // Anything that used the result of the function call should now use the
269 // PHI node as their operand.
271 TheCall->replaceAllUsesWith(PHI);
274 // Loop over all of the return instructions, turning them into unconditional
275 // branches to the merge point now, and adding entries to the PHI node as
277 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
278 ReturnInst *RI = Returns[i];
281 assert(RI->getReturnValue() && "Ret should have value!");
282 assert(RI->getReturnValue()->getType() == PHI->getType() &&
283 "Ret value not consistent in function!");
284 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
287 // Add a branch to the merge point where the PHI node lives if it exists.
288 new BranchInst(AfterCallBB, RI);
290 // Delete the return instruction now
291 RI->getParent()->getInstList().erase(RI);
294 } else if (!Returns.empty()) {
295 // Otherwise, if there is exactly one return value, just replace anything
296 // using the return value of the call with the computed value.
297 if (!TheCall->use_empty())
298 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
300 // Add a branch to the merge point where the PHI node lives if it exists.
301 new BranchInst(AfterCallBB, Returns[0]);
303 // Delete the return instruction now
304 Returns[0]->getParent()->getInstList().erase(Returns[0]);
307 // Since we are now done with the Call/Invoke, we can delete it.
308 TheCall->getParent()->getInstList().erase(TheCall);
310 // We should always be able to fold the entry block of the function into the
311 // single predecessor of the block...
312 assert(cast<BranchInst>(Br)->isUnconditional() &&"splitBasicBlock broken!");
313 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
314 SimplifyCFG(CalleeEntry);
316 // Okay, continue the CFG cleanup. It's often the case that there is only a
317 // single return instruction in the callee function. If this is the case,
318 // then we have an unconditional branch from the return block to the
319 // 'AfterCallBB'. Check for this case, and eliminate the branch is possible.
320 SimplifyCFG(AfterCallBB);