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 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/Cloning.h"
16 #include "llvm/Constants.h"
17 #include "llvm/DerivedTypes.h"
18 #include "llvm/Module.h"
19 #include "llvm/Instructions.h"
20 #include "llvm/Intrinsics.h"
21 #include "llvm/Support/CallSite.h"
24 bool llvm::InlineFunction(CallInst *CI) { return InlineFunction(CallSite(CI)); }
25 bool llvm::InlineFunction(InvokeInst *II) {return InlineFunction(CallSite(II));}
27 // InlineFunction - This function inlines the called function into the basic
28 // block of the caller. This returns false if it is not possible to inline this
29 // call. The program is still in a well defined state if this occurs though.
31 // Note that this only does one level of inlining. For example, if the
32 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
33 // exists in the instruction stream. Similiarly this will inline a recursive
34 // function by one level.
36 bool llvm::InlineFunction(CallSite CS) {
37 Instruction *TheCall = CS.getInstruction();
38 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
39 "Instruction not in function!");
41 const Function *CalledFunc = CS.getCalledFunction();
42 if (CalledFunc == 0 || // Can't inline external function or indirect
43 CalledFunc->isExternal() || // call, or call to a vararg function!
44 CalledFunc->getFunctionType()->isVarArg()) return false;
47 // If the call to the callee is a non-tail call, we must clear the 'tail'
48 // flags on any calls that we inline.
49 bool MustClearTailCallFlags =
50 isa<CallInst>(TheCall) && !cast<CallInst>(TheCall)->isTailCall();
52 BasicBlock *OrigBB = TheCall->getParent();
53 Function *Caller = OrigBB->getParent();
55 // Get an iterator to the last basic block in the function, which will have
56 // the new function inlined after it.
58 Function::iterator LastBlock = &Caller->back();
60 // Make sure to capture all of the return instructions from the cloned
62 std::vector<ReturnInst*> Returns;
63 { // Scope to destroy ValueMap after cloning.
64 // Calculate the vector of arguments to pass into the function cloner...
65 std::map<const Value*, Value*> ValueMap;
66 assert(std::distance(CalledFunc->arg_begin(), CalledFunc->arg_end()) ==
67 std::distance(CS.arg_begin(), CS.arg_end()) &&
68 "No varargs calls can be inlined!");
70 CallSite::arg_iterator AI = CS.arg_begin();
71 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
72 E = CalledFunc->arg_end(); I != E; ++I, ++AI)
75 // Clone the entire body of the callee into the caller.
76 CloneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i");
79 // Remember the first block that is newly cloned over.
80 Function::iterator FirstNewBlock = LastBlock; ++FirstNewBlock;
82 // If there are any alloca instructions in the block that used to be the entry
83 // block for the callee, move them to the entry block of the caller. First
84 // calculate which instruction they should be inserted before. We insert the
85 // instructions at the end of the current alloca list.
88 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
89 for (BasicBlock::iterator I = FirstNewBlock->begin(),
90 E = FirstNewBlock->end(); I != E; )
91 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++))
92 if (isa<Constant>(AI->getArraySize())) {
93 // Scan for the block of allocas that we can move over, and move them
95 while (isa<AllocaInst>(I) &&
96 isa<Constant>(cast<AllocaInst>(I)->getArraySize()))
99 // Transfer all of the allocas over in a block. Using splice means
100 // that they instructions aren't removed from the symbol table, then
102 Caller->front().getInstList().splice(InsertPoint,
103 FirstNewBlock->getInstList(),
108 // If we are inlining tail call instruction through an invoke or
109 if (MustClearTailCallFlags) {
110 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
112 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
113 if (CallInst *CI = dyn_cast<CallInst>(I))
114 CI->setTailCall(false);
117 // If we are inlining for an invoke instruction, we must make sure to rewrite
118 // any inlined 'unwind' instructions into branches to the invoke exception
119 // destination, and call instructions into invoke instructions.
120 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
121 BasicBlock *InvokeDest = II->getUnwindDest();
122 std::vector<Value*> InvokeDestPHIValues;
124 // If there are PHI nodes in the exceptional destination block, we need to
125 // keep track of which values came into them from this invoke, then remove
126 // the entry for this block.
127 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) {
128 PHINode *PN = cast<PHINode>(I);
129 // Save the value to use for this edge...
130 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(OrigBB));
133 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
135 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
136 // We only need to check for function calls: inlined invoke instructions
137 // require no special handling...
138 if (CallInst *CI = dyn_cast<CallInst>(I)) {
139 // Convert this function call into an invoke instruction... if it's
140 // not an intrinsic function call (which are known to not unwind).
141 if (CI->getCalledFunction() &&
142 CI->getCalledFunction()->getIntrinsicID()) {
145 // First, split the basic block...
146 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
148 // Next, create the new invoke instruction, inserting it at the end
149 // of the old basic block.
151 new InvokeInst(CI->getCalledValue(), Split, InvokeDest,
152 std::vector<Value*>(CI->op_begin()+1, CI->op_end()),
153 CI->getName(), BB->getTerminator());
154 II->setCallingConv(CI->getCallingConv());
156 // Make sure that anything using the call now uses the invoke!
157 CI->replaceAllUsesWith(II);
159 // Delete the unconditional branch inserted by splitBasicBlock
160 BB->getInstList().pop_back();
161 Split->getInstList().pop_front(); // Delete the original call
163 // Update any PHI nodes in the exceptional block to indicate that
164 // there is now a new entry in them.
166 for (BasicBlock::iterator I = InvokeDest->begin();
167 isa<PHINode>(I); ++I, ++i) {
168 PHINode *PN = cast<PHINode>(I);
169 PN->addIncoming(InvokeDestPHIValues[i], BB);
172 // This basic block is now complete, start scanning the next one.
180 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
181 // An UnwindInst requires special handling when it gets inlined into an
182 // invoke site. Once this happens, we know that the unwind would cause
183 // a control transfer to the invoke exception destination, so we can
184 // transform it into a direct branch to the exception destination.
185 new BranchInst(InvokeDest, UI);
187 // Delete the unwind instruction!
188 UI->getParent()->getInstList().pop_back();
190 // Update any PHI nodes in the exceptional block to indicate that
191 // there is now a new entry in them.
193 for (BasicBlock::iterator I = InvokeDest->begin();
194 isa<PHINode>(I); ++I, ++i) {
195 PHINode *PN = cast<PHINode>(I);
196 PN->addIncoming(InvokeDestPHIValues[i], BB);
201 // Now that everything is happy, we have one final detail. The PHI nodes in
202 // the exception destination block still have entries due to the original
203 // invoke instruction. Eliminate these entries (which might even delete the
205 InvokeDest->removePredecessor(II->getParent());
208 // If we cloned in _exactly one_ basic block, and if that block ends in a
209 // return instruction, we splice the body of the inlined callee directly into
210 // the calling basic block.
211 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
212 // Move all of the instructions right before the call.
213 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
214 FirstNewBlock->begin(), FirstNewBlock->end());
215 // Remove the cloned basic block.
216 Caller->getBasicBlockList().pop_back();
218 // If the call site was an invoke instruction, add a branch to the normal
220 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
221 new BranchInst(II->getNormalDest(), TheCall);
223 // If the return instruction returned a value, replace uses of the call with
224 // uses of the returned value.
225 if (!TheCall->use_empty())
226 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
228 // Since we are now done with the Call/Invoke, we can delete it.
229 TheCall->getParent()->getInstList().erase(TheCall);
231 // Since we are now done with the return instruction, delete it also.
232 Returns[0]->getParent()->getInstList().erase(Returns[0]);
234 // We are now done with the inlining.
238 // Otherwise, we have the normal case, of more than one block to inline or
239 // multiple return sites.
241 // We want to clone the entire callee function into the hole between the
242 // "starter" and "ender" blocks. How we accomplish this depends on whether
243 // this is an invoke instruction or a call instruction.
244 BasicBlock *AfterCallBB;
245 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
247 // Add an unconditional branch to make this look like the CallInst case...
248 BranchInst *NewBr = new BranchInst(II->getNormalDest(), TheCall);
250 // Split the basic block. This guarantees that no PHI nodes will have to be
251 // updated due to new incoming edges, and make the invoke case more
252 // symmetric to the call case.
253 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
254 CalledFunc->getName()+".exit");
256 } else { // It's a call
257 // If this is a call instruction, we need to split the basic block that
258 // the call lives in.
260 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
261 CalledFunc->getName()+".exit");
264 // Change the branch that used to go to AfterCallBB to branch to the first
265 // basic block of the inlined function.
267 TerminatorInst *Br = OrigBB->getTerminator();
268 assert(Br && Br->getOpcode() == Instruction::Br &&
269 "splitBasicBlock broken!");
270 Br->setOperand(0, FirstNewBlock);
273 // Now that the function is correct, make it a little bit nicer. In
274 // particular, move the basic blocks inserted from the end of the function
275 // into the space made by splitting the source basic block.
277 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
278 FirstNewBlock, Caller->end());
280 // Handle all of the return instructions that we just cloned in, and eliminate
281 // any users of the original call/invoke instruction.
282 if (Returns.size() > 1) {
283 // The PHI node should go at the front of the new basic block to merge all
284 // possible incoming values.
287 if (!TheCall->use_empty()) {
288 PHI = new PHINode(CalledFunc->getReturnType(),
289 TheCall->getName(), AfterCallBB->begin());
291 // Anything that used the result of the function call should now use the
292 // PHI node as their operand.
294 TheCall->replaceAllUsesWith(PHI);
297 // Loop over all of the return instructions, turning them into unconditional
298 // branches to the merge point now, and adding entries to the PHI node as
300 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
301 ReturnInst *RI = Returns[i];
304 assert(RI->getReturnValue() && "Ret should have value!");
305 assert(RI->getReturnValue()->getType() == PHI->getType() &&
306 "Ret value not consistent in function!");
307 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
310 // Add a branch to the merge point where the PHI node lives if it exists.
311 new BranchInst(AfterCallBB, RI);
313 // Delete the return instruction now
314 RI->getParent()->getInstList().erase(RI);
317 } else if (!Returns.empty()) {
318 // Otherwise, if there is exactly one return value, just replace anything
319 // using the return value of the call with the computed value.
320 if (!TheCall->use_empty())
321 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
323 // Splice the code from the return block into the block that it will return
324 // to, which contains the code that was after the call.
325 BasicBlock *ReturnBB = Returns[0]->getParent();
326 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
327 ReturnBB->getInstList());
329 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
330 ReturnBB->replaceAllUsesWith(AfterCallBB);
332 // Delete the return instruction now and empty ReturnBB now.
333 Returns[0]->eraseFromParent();
334 ReturnBB->eraseFromParent();
335 } else if (!TheCall->use_empty()) {
336 // No returns, but something is using the return value of the call. Just
338 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
341 // Since we are now done with the Call/Invoke, we can delete it.
342 TheCall->eraseFromParent();
344 // We should always be able to fold the entry block of the function into the
345 // single predecessor of the block...
346 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
347 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
349 // Splice the code entry block into calling block, right before the
350 // unconditional branch.
351 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
352 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
354 // Remove the unconditional branch.
355 OrigBB->getInstList().erase(Br);
357 // Now we can remove the CalleeEntry block, which is now empty.
358 Caller->getBasicBlockList().erase(CalleeEntry);