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/Analysis/CallGraph.h"
22 #include "llvm/Support/CallSite.h"
25 bool llvm::InlineFunction(CallInst *CI, CallGraph *CG) {
26 return InlineFunction(CallSite(CI), CG);
28 bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG) {
29 return InlineFunction(CallSite(II), CG);
32 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
33 /// in the body of the inlined function into invokes and turn unwind
34 /// instructions into branches to the invoke unwind dest.
36 /// II is the invoke instruction begin inlined. FirstNewBlock is the first
37 /// block of the inlined code (the last block is the end of the function),
38 /// and InlineCodeInfo is information about the code that got inlined.
39 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
40 ClonedCodeInfo &InlinedCodeInfo) {
41 BasicBlock *InvokeDest = II->getUnwindDest();
42 std::vector<Value*> InvokeDestPHIValues;
44 // If there are PHI nodes in the unwind destination block, we need to
45 // keep track of which values came into them from this invoke, then remove
46 // the entry for this block.
47 BasicBlock *InvokeBlock = II->getParent();
48 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) {
49 PHINode *PN = cast<PHINode>(I);
50 // Save the value to use for this edge.
51 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock));
54 Function *Caller = FirstNewBlock->getParent();
56 // The inlined code is currently at the end of the function, scan from the
57 // start of the inlined code to its end, checking for stuff we need to
59 if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) {
60 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
62 if (InlinedCodeInfo.ContainsCalls) {
63 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){
64 Instruction *I = BBI++;
66 // We only need to check for function calls: inlined invoke
67 // instructions require no special handling.
68 if (!isa<CallInst>(I)) continue;
69 CallInst *CI = cast<CallInst>(I);
71 // If this is an intrinsic function call, don't convert it to an
73 if (CI->getCalledFunction() &&
74 CI->getCalledFunction()->getIntrinsicID())
77 // Convert this function call into an invoke instruction.
78 // First, split the basic block.
79 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
81 // Next, create the new invoke instruction, inserting it at the end
82 // of the old basic block.
84 new InvokeInst(CI->getCalledValue(), Split, InvokeDest,
85 std::vector<Value*>(CI->op_begin()+1, CI->op_end()),
86 CI->getName(), BB->getTerminator());
87 II->setCallingConv(CI->getCallingConv());
89 // Make sure that anything using the call now uses the invoke!
90 CI->replaceAllUsesWith(II);
92 // Delete the unconditional branch inserted by splitBasicBlock
93 BB->getInstList().pop_back();
94 Split->getInstList().pop_front(); // Delete the original call
96 // Update any PHI nodes in the exceptional block to indicate that
97 // there is now a new entry in them.
99 for (BasicBlock::iterator I = InvokeDest->begin();
100 isa<PHINode>(I); ++I, ++i) {
101 PHINode *PN = cast<PHINode>(I);
102 PN->addIncoming(InvokeDestPHIValues[i], BB);
105 // This basic block is now complete, start scanning the next one.
110 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
111 // An UnwindInst requires special handling when it gets inlined into an
112 // invoke site. Once this happens, we know that the unwind would cause
113 // a control transfer to the invoke exception destination, so we can
114 // transform it into a direct branch to the exception destination.
115 new BranchInst(InvokeDest, UI);
117 // Delete the unwind instruction!
118 UI->getParent()->getInstList().pop_back();
120 // Update any PHI nodes in the exceptional block to indicate that
121 // there is now a new entry in them.
123 for (BasicBlock::iterator I = InvokeDest->begin();
124 isa<PHINode>(I); ++I, ++i) {
125 PHINode *PN = cast<PHINode>(I);
126 PN->addIncoming(InvokeDestPHIValues[i], BB);
132 // Now that everything is happy, we have one final detail. The PHI nodes in
133 // the exception destination block still have entries due to the original
134 // invoke instruction. Eliminate these entries (which might even delete the
136 InvokeDest->removePredecessor(II->getParent());
139 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
140 /// into the caller, update the specified callgraph to reflect the changes we
141 /// made. Note that it's possible that not all code was copied over, so only
142 /// some edges of the callgraph will be remain.
143 static void UpdateCallGraphAfterInlining(const Function *Caller,
144 const Function *Callee,
145 Function::iterator FirstNewBlock,
146 std::map<const Value*, Value*> &ValueMap,
148 // Update the call graph by deleting the edge from Callee to Caller
149 CallGraphNode *CalleeNode = CG[Callee];
150 CallGraphNode *CallerNode = CG[Caller];
151 CallerNode->removeCallEdgeTo(CalleeNode);
153 // Since we inlined some uninlined call sites in the callee into the caller,
154 // add edges from the caller to all of the callees of the callee.
155 for (CallGraphNode::iterator I = CalleeNode->begin(),
156 E = CalleeNode->end(); I != E; ++I) {
157 const Instruction *OrigCall = I->first.getInstruction();
159 std::map<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall);
160 // Only copy the edge if the call was inlined!
161 if (VMI != ValueMap.end() && VMI->second) {
162 // If the call was inlined, but then constant folded, there is no edge to
163 // add. Check for this case.
164 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second))
165 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second);
171 // InlineFunction - This function inlines the called function into the basic
172 // block of the caller. This returns false if it is not possible to inline this
173 // call. The program is still in a well defined state if this occurs though.
175 // Note that this only does one level of inlining. For example, if the
176 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
177 // exists in the instruction stream. Similiarly this will inline a recursive
178 // function by one level.
180 bool llvm::InlineFunction(CallSite CS, CallGraph *CG) {
181 Instruction *TheCall = CS.getInstruction();
182 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
183 "Instruction not in function!");
185 const Function *CalledFunc = CS.getCalledFunction();
186 if (CalledFunc == 0 || // Can't inline external function or indirect
187 CalledFunc->isExternal() || // call, or call to a vararg function!
188 CalledFunc->getFunctionType()->isVarArg()) return false;
191 // If the call to the callee is a non-tail call, we must clear the 'tail'
192 // flags on any calls that we inline.
193 bool MustClearTailCallFlags =
194 isa<CallInst>(TheCall) && !cast<CallInst>(TheCall)->isTailCall();
196 BasicBlock *OrigBB = TheCall->getParent();
197 Function *Caller = OrigBB->getParent();
199 // Get an iterator to the last basic block in the function, which will have
200 // the new function inlined after it.
202 Function::iterator LastBlock = &Caller->back();
204 // Make sure to capture all of the return instructions from the cloned
206 std::vector<ReturnInst*> Returns;
207 ClonedCodeInfo InlinedFunctionInfo;
208 Function::iterator FirstNewBlock;
210 { // Scope to destroy ValueMap after cloning.
211 std::map<const Value*, Value*> ValueMap;
213 // Calculate the vector of arguments to pass into the function cloner, which
214 // matches up the formal to the actual argument values.
215 assert(std::distance(CalledFunc->arg_begin(), CalledFunc->arg_end()) ==
216 std::distance(CS.arg_begin(), CS.arg_end()) &&
217 "No varargs calls can be inlined!");
218 CallSite::arg_iterator AI = CS.arg_begin();
219 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
220 E = CalledFunc->arg_end(); I != E; ++I, ++AI)
223 // We want the inliner to prune the code as it copies. We would LOVE to
224 // have no dead or constant instructions leftover after inlining occurs
225 // (which can happen, e.g., because an argument was constant), but we'll be
226 // happy with whatever the cloner can do.
227 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i",
228 &InlinedFunctionInfo);
230 // Remember the first block that is newly cloned over.
231 FirstNewBlock = LastBlock; ++FirstNewBlock;
233 // Update the callgraph if requested.
235 UpdateCallGraphAfterInlining(Caller, CalledFunc, FirstNewBlock, ValueMap,
239 // If there are any alloca instructions in the block that used to be the entry
240 // block for the callee, move them to the entry block of the caller. First
241 // calculate which instruction they should be inserted before. We insert the
242 // instructions at the end of the current alloca list.
245 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
246 for (BasicBlock::iterator I = FirstNewBlock->begin(),
247 E = FirstNewBlock->end(); I != E; )
248 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++)) {
249 // If the alloca is now dead, remove it. This often occurs due to code
251 if (AI->use_empty()) {
252 AI->eraseFromParent();
256 if (isa<Constant>(AI->getArraySize())) {
257 // Scan for the block of allocas that we can move over, and move them
259 while (isa<AllocaInst>(I) &&
260 isa<Constant>(cast<AllocaInst>(I)->getArraySize()))
263 // Transfer all of the allocas over in a block. Using splice means
264 // that they instructions aren't removed from the symbol table, then
266 Caller->front().getInstList().splice(InsertPoint,
267 FirstNewBlock->getInstList(),
273 // If the inlined code contained dynamic alloca instructions, wrap the inlined
274 // code with llvm.stacksave/llvm.stackrestore intrinsics.
275 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
276 Module *M = Caller->getParent();
277 const Type *SBytePtr = PointerType::get(Type::SByteTy);
278 // Get the two intrinsics we care about.
279 Function *StackSave, *StackRestore;
280 StackSave = M->getOrInsertFunction("llvm.stacksave", SBytePtr, NULL);
281 StackRestore = M->getOrInsertFunction("llvm.stackrestore", Type::VoidTy,
284 // If we are preserving the callgraph, add edges to the stacksave/restore
285 // functions for the calls we insert.
286 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0;
288 StackSaveCGN = CG->getOrInsertFunction(StackSave);
289 StackRestoreCGN = CG->getOrInsertFunction(StackRestore);
290 CallerNode = (*CG)[Caller];
293 // Insert the llvm.stacksave.
294 CallInst *SavedPtr = new CallInst(StackSave, "savedstack",
295 FirstNewBlock->begin());
296 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN);
298 // Insert a call to llvm.stackrestore before any return instructions in the
300 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
301 CallInst *CI = new CallInst(StackRestore, SavedPtr, "", Returns[i]);
302 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN);
305 // Count the number of StackRestore calls we insert.
306 unsigned NumStackRestores = Returns.size();
308 // If we are inlining an invoke instruction, insert restores before each
309 // unwind. These unwinds will be rewritten into branches later.
310 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) {
311 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
313 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
314 new CallInst(StackRestore, SavedPtr, "", UI);
320 // If we are inlining tail call instruction through a call site that isn't
321 // marked 'tail', we must remove the tail marker for any calls in the inlined
323 if (MustClearTailCallFlags && InlinedFunctionInfo.ContainsCalls) {
324 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
326 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
327 if (CallInst *CI = dyn_cast<CallInst>(I))
328 CI->setTailCall(false);
331 // If we are inlining for an invoke instruction, we must make sure to rewrite
332 // any inlined 'unwind' instructions into branches to the invoke exception
333 // destination, and call instructions into invoke instructions.
334 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
335 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
337 // If we cloned in _exactly one_ basic block, and if that block ends in a
338 // return instruction, we splice the body of the inlined callee directly into
339 // the calling basic block.
340 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
341 // Move all of the instructions right before the call.
342 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
343 FirstNewBlock->begin(), FirstNewBlock->end());
344 // Remove the cloned basic block.
345 Caller->getBasicBlockList().pop_back();
347 // If the call site was an invoke instruction, add a branch to the normal
349 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
350 new BranchInst(II->getNormalDest(), TheCall);
352 // If the return instruction returned a value, replace uses of the call with
353 // uses of the returned value.
354 if (!TheCall->use_empty())
355 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
357 // Since we are now done with the Call/Invoke, we can delete it.
358 TheCall->getParent()->getInstList().erase(TheCall);
360 // Since we are now done with the return instruction, delete it also.
361 Returns[0]->getParent()->getInstList().erase(Returns[0]);
363 // We are now done with the inlining.
367 // Otherwise, we have the normal case, of more than one block to inline or
368 // multiple return sites.
370 // We want to clone the entire callee function into the hole between the
371 // "starter" and "ender" blocks. How we accomplish this depends on whether
372 // this is an invoke instruction or a call instruction.
373 BasicBlock *AfterCallBB;
374 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
376 // Add an unconditional branch to make this look like the CallInst case...
377 BranchInst *NewBr = new BranchInst(II->getNormalDest(), TheCall);
379 // Split the basic block. This guarantees that no PHI nodes will have to be
380 // updated due to new incoming edges, and make the invoke case more
381 // symmetric to the call case.
382 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
383 CalledFunc->getName()+".exit");
385 } else { // It's a call
386 // If this is a call instruction, we need to split the basic block that
387 // the call lives in.
389 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
390 CalledFunc->getName()+".exit");
393 // Change the branch that used to go to AfterCallBB to branch to the first
394 // basic block of the inlined function.
396 TerminatorInst *Br = OrigBB->getTerminator();
397 assert(Br && Br->getOpcode() == Instruction::Br &&
398 "splitBasicBlock broken!");
399 Br->setOperand(0, FirstNewBlock);
402 // Now that the function is correct, make it a little bit nicer. In
403 // particular, move the basic blocks inserted from the end of the function
404 // into the space made by splitting the source basic block.
406 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
407 FirstNewBlock, Caller->end());
409 // Handle all of the return instructions that we just cloned in, and eliminate
410 // any users of the original call/invoke instruction.
411 if (Returns.size() > 1) {
412 // The PHI node should go at the front of the new basic block to merge all
413 // possible incoming values.
416 if (!TheCall->use_empty()) {
417 PHI = new PHINode(CalledFunc->getReturnType(),
418 TheCall->getName(), AfterCallBB->begin());
420 // Anything that used the result of the function call should now use the
421 // PHI node as their operand.
423 TheCall->replaceAllUsesWith(PHI);
426 // Loop over all of the return instructions, turning them into unconditional
427 // branches to the merge point now, and adding entries to the PHI node as
429 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
430 ReturnInst *RI = Returns[i];
433 assert(RI->getReturnValue() && "Ret should have value!");
434 assert(RI->getReturnValue()->getType() == PHI->getType() &&
435 "Ret value not consistent in function!");
436 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
439 // Add a branch to the merge point where the PHI node lives if it exists.
440 new BranchInst(AfterCallBB, RI);
442 // Delete the return instruction now
443 RI->getParent()->getInstList().erase(RI);
446 } else if (!Returns.empty()) {
447 // Otherwise, if there is exactly one return value, just replace anything
448 // using the return value of the call with the computed value.
449 if (!TheCall->use_empty())
450 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
452 // Splice the code from the return block into the block that it will return
453 // to, which contains the code that was after the call.
454 BasicBlock *ReturnBB = Returns[0]->getParent();
455 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
456 ReturnBB->getInstList());
458 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
459 ReturnBB->replaceAllUsesWith(AfterCallBB);
461 // Delete the return instruction now and empty ReturnBB now.
462 Returns[0]->eraseFromParent();
463 ReturnBB->eraseFromParent();
464 } else if (!TheCall->use_empty()) {
465 // No returns, but something is using the return value of the call. Just
467 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
470 // Since we are now done with the Call/Invoke, we can delete it.
471 TheCall->eraseFromParent();
473 // We should always be able to fold the entry block of the function into the
474 // single predecessor of the block...
475 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
476 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
478 // Splice the code entry block into calling block, right before the
479 // unconditional branch.
480 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
481 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
483 // Remove the unconditional branch.
484 OrigBB->getInstList().erase(Br);
486 // Now we can remove the CalleeEntry block, which is now empty.
487 Caller->getBasicBlockList().erase(CalleeEntry);