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/ADT/SmallVector.h"
23 #include "llvm/Support/CallSite.h"
26 bool llvm::InlineFunction(CallInst *CI, CallGraph *CG, const TargetData *TD) {
27 return InlineFunction(CallSite(CI), CG, TD);
29 bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG, const TargetData *TD) {
30 return InlineFunction(CallSite(II), CG, TD);
33 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
34 /// in the body of the inlined function into invokes and turn unwind
35 /// instructions into branches to the invoke unwind dest.
37 /// II is the invoke instruction begin inlined. FirstNewBlock is the first
38 /// block of the inlined code (the last block is the end of the function),
39 /// and InlineCodeInfo is information about the code that got inlined.
40 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
41 ClonedCodeInfo &InlinedCodeInfo) {
42 BasicBlock *InvokeDest = II->getUnwindDest();
43 std::vector<Value*> InvokeDestPHIValues;
45 // If there are PHI nodes in the unwind destination block, we need to
46 // keep track of which values came into them from this invoke, then remove
47 // the entry for this block.
48 BasicBlock *InvokeBlock = II->getParent();
49 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) {
50 PHINode *PN = cast<PHINode>(I);
51 // Save the value to use for this edge.
52 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock));
55 Function *Caller = FirstNewBlock->getParent();
57 // The inlined code is currently at the end of the function, scan from the
58 // start of the inlined code to its end, checking for stuff we need to
60 if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) {
61 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
63 if (InlinedCodeInfo.ContainsCalls) {
64 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){
65 Instruction *I = BBI++;
67 // We only need to check for function calls: inlined invoke
68 // instructions require no special handling.
69 if (!isa<CallInst>(I)) continue;
70 CallInst *CI = cast<CallInst>(I);
72 // If this is an intrinsic function call or an inline asm, don't
73 // convert it to an invoke.
74 if ((CI->getCalledFunction() &&
75 CI->getCalledFunction()->getIntrinsicID()) ||
76 isa<InlineAsm>(CI->getCalledValue()))
79 // Convert this function call into an invoke instruction.
80 // First, split the basic block.
81 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc");
83 // Next, create the new invoke instruction, inserting it at the end
84 // of the old basic block.
85 SmallVector<Value*, 8> InvokeArgs(CI->op_begin()+1, CI->op_end());
87 new InvokeInst(CI->getCalledValue(), Split, InvokeDest,
88 &InvokeArgs[0], InvokeArgs.size(),
89 CI->getName(), BB->getTerminator());
90 II->setCallingConv(CI->getCallingConv());
92 // Make sure that anything using the call now uses the invoke!
93 CI->replaceAllUsesWith(II);
95 // Delete the unconditional branch inserted by splitBasicBlock
96 BB->getInstList().pop_back();
97 Split->getInstList().pop_front(); // Delete the original call
99 // Update any PHI nodes in the exceptional block to indicate that
100 // there is now a new entry in them.
102 for (BasicBlock::iterator I = InvokeDest->begin();
103 isa<PHINode>(I); ++I, ++i) {
104 PHINode *PN = cast<PHINode>(I);
105 PN->addIncoming(InvokeDestPHIValues[i], BB);
108 // This basic block is now complete, start scanning the next one.
113 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
114 // An UnwindInst requires special handling when it gets inlined into an
115 // invoke site. Once this happens, we know that the unwind would cause
116 // a control transfer to the invoke exception destination, so we can
117 // transform it into a direct branch to the exception destination.
118 new BranchInst(InvokeDest, UI);
120 // Delete the unwind instruction!
121 UI->getParent()->getInstList().pop_back();
123 // Update any PHI nodes in the exceptional block to indicate that
124 // there is now a new entry in them.
126 for (BasicBlock::iterator I = InvokeDest->begin();
127 isa<PHINode>(I); ++I, ++i) {
128 PHINode *PN = cast<PHINode>(I);
129 PN->addIncoming(InvokeDestPHIValues[i], BB);
135 // Now that everything is happy, we have one final detail. The PHI nodes in
136 // the exception destination block still have entries due to the original
137 // invoke instruction. Eliminate these entries (which might even delete the
139 InvokeDest->removePredecessor(II->getParent());
142 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
143 /// into the caller, update the specified callgraph to reflect the changes we
144 /// made. Note that it's possible that not all code was copied over, so only
145 /// some edges of the callgraph will be remain.
146 static void UpdateCallGraphAfterInlining(const Function *Caller,
147 const Function *Callee,
148 Function::iterator FirstNewBlock,
149 DenseMap<const Value*, Value*> &ValueMap,
151 // Update the call graph by deleting the edge from Callee to Caller
152 CallGraphNode *CalleeNode = CG[Callee];
153 CallGraphNode *CallerNode = CG[Caller];
154 CallerNode->removeCallEdgeTo(CalleeNode);
156 // Since we inlined some uninlined call sites in the callee into the caller,
157 // add edges from the caller to all of the callees of the callee.
158 for (CallGraphNode::iterator I = CalleeNode->begin(),
159 E = CalleeNode->end(); I != E; ++I) {
160 const Instruction *OrigCall = I->first.getInstruction();
162 DenseMap<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall);
163 // Only copy the edge if the call was inlined!
164 if (VMI != ValueMap.end() && VMI->second) {
165 // If the call was inlined, but then constant folded, there is no edge to
166 // add. Check for this case.
167 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second))
168 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second);
174 // InlineFunction - This function inlines the called function into the basic
175 // block of the caller. This returns false if it is not possible to inline this
176 // call. The program is still in a well defined state if this occurs though.
178 // Note that this only does one level of inlining. For example, if the
179 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
180 // exists in the instruction stream. Similiarly this will inline a recursive
181 // function by one level.
183 bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) {
184 Instruction *TheCall = CS.getInstruction();
185 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
186 "Instruction not in function!");
188 const Function *CalledFunc = CS.getCalledFunction();
189 if (CalledFunc == 0 || // Can't inline external function or indirect
190 CalledFunc->isDeclaration() || // call, or call to a vararg function!
191 CalledFunc->getFunctionType()->isVarArg()) return false;
194 // If the call to the callee is a non-tail call, we must clear the 'tail'
195 // flags on any calls that we inline.
196 bool MustClearTailCallFlags =
197 isa<CallInst>(TheCall) && !cast<CallInst>(TheCall)->isTailCall();
199 BasicBlock *OrigBB = TheCall->getParent();
200 Function *Caller = OrigBB->getParent();
202 // Get an iterator to the last basic block in the function, which will have
203 // the new function inlined after it.
205 Function::iterator LastBlock = &Caller->back();
207 // Make sure to capture all of the return instructions from the cloned
209 std::vector<ReturnInst*> Returns;
210 ClonedCodeInfo InlinedFunctionInfo;
211 Function::iterator FirstNewBlock;
213 { // Scope to destroy ValueMap after cloning.
214 DenseMap<const Value*, Value*> ValueMap;
216 // Calculate the vector of arguments to pass into the function cloner, which
217 // matches up the formal to the actual argument values.
218 assert(std::distance(CalledFunc->arg_begin(), CalledFunc->arg_end()) ==
219 std::distance(CS.arg_begin(), CS.arg_end()) &&
220 "No varargs calls can be inlined!");
221 CallSite::arg_iterator AI = CS.arg_begin();
222 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
223 E = CalledFunc->arg_end(); I != E; ++I, ++AI)
226 // We want the inliner to prune the code as it copies. We would LOVE to
227 // have no dead or constant instructions leftover after inlining occurs
228 // (which can happen, e.g., because an argument was constant), but we'll be
229 // happy with whatever the cloner can do.
230 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i",
231 &InlinedFunctionInfo, TD);
233 // Remember the first block that is newly cloned over.
234 FirstNewBlock = LastBlock; ++FirstNewBlock;
236 // Update the callgraph if requested.
238 UpdateCallGraphAfterInlining(Caller, CalledFunc, FirstNewBlock, ValueMap,
242 // If there are any alloca instructions in the block that used to be the entry
243 // block for the callee, move them to the entry block of the caller. First
244 // calculate which instruction they should be inserted before. We insert the
245 // instructions at the end of the current alloca list.
248 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
249 for (BasicBlock::iterator I = FirstNewBlock->begin(),
250 E = FirstNewBlock->end(); I != E; )
251 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++)) {
252 // If the alloca is now dead, remove it. This often occurs due to code
254 if (AI->use_empty()) {
255 AI->eraseFromParent();
259 if (isa<Constant>(AI->getArraySize())) {
260 // Scan for the block of allocas that we can move over, and move them
262 while (isa<AllocaInst>(I) &&
263 isa<Constant>(cast<AllocaInst>(I)->getArraySize()))
266 // Transfer all of the allocas over in a block. Using splice means
267 // that the instructions aren't removed from the symbol table, then
269 Caller->getEntryBlock().getInstList().splice(
271 FirstNewBlock->getInstList(),
277 // If the inlined code contained dynamic alloca instructions, wrap the inlined
278 // code with llvm.stacksave/llvm.stackrestore intrinsics.
279 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
280 Module *M = Caller->getParent();
281 const Type *BytePtr = PointerType::get(Type::Int8Ty);
282 // Get the two intrinsics we care about.
283 Constant *StackSave, *StackRestore;
284 StackSave = M->getOrInsertFunction("llvm.stacksave", BytePtr, NULL);
285 StackRestore = M->getOrInsertFunction("llvm.stackrestore", Type::VoidTy,
288 // If we are preserving the callgraph, add edges to the stacksave/restore
289 // functions for the calls we insert.
290 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0;
292 // We know that StackSave/StackRestore are Function*'s, because they are
293 // intrinsics which must have the right types.
294 StackSaveCGN = CG->getOrInsertFunction(cast<Function>(StackSave));
295 StackRestoreCGN = CG->getOrInsertFunction(cast<Function>(StackRestore));
296 CallerNode = (*CG)[Caller];
299 // Insert the llvm.stacksave.
300 CallInst *SavedPtr = new CallInst(StackSave, "savedstack",
301 FirstNewBlock->begin());
302 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN);
304 // Insert a call to llvm.stackrestore before any return instructions in the
306 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
307 CallInst *CI = new CallInst(StackRestore, SavedPtr, "", Returns[i]);
308 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN);
311 // Count the number of StackRestore calls we insert.
312 unsigned NumStackRestores = Returns.size();
314 // If we are inlining an invoke instruction, insert restores before each
315 // unwind. These unwinds will be rewritten into branches later.
316 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) {
317 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
319 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
320 new CallInst(StackRestore, SavedPtr, "", UI);
326 // If we are inlining tail call instruction through a call site that isn't
327 // marked 'tail', we must remove the tail marker for any calls in the inlined
329 if (MustClearTailCallFlags && InlinedFunctionInfo.ContainsCalls) {
330 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
332 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
333 if (CallInst *CI = dyn_cast<CallInst>(I))
334 CI->setTailCall(false);
337 // If we are inlining for an invoke instruction, we must make sure to rewrite
338 // any inlined 'unwind' instructions into branches to the invoke exception
339 // destination, and call instructions into invoke instructions.
340 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
341 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
343 // If we cloned in _exactly one_ basic block, and if that block ends in a
344 // return instruction, we splice the body of the inlined callee directly into
345 // the calling basic block.
346 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
347 // Move all of the instructions right before the call.
348 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
349 FirstNewBlock->begin(), FirstNewBlock->end());
350 // Remove the cloned basic block.
351 Caller->getBasicBlockList().pop_back();
353 // If the call site was an invoke instruction, add a branch to the normal
355 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
356 new BranchInst(II->getNormalDest(), TheCall);
358 // If the return instruction returned a value, replace uses of the call with
359 // uses of the returned value.
360 if (!TheCall->use_empty())
361 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
363 // Since we are now done with the Call/Invoke, we can delete it.
364 TheCall->getParent()->getInstList().erase(TheCall);
366 // Since we are now done with the return instruction, delete it also.
367 Returns[0]->getParent()->getInstList().erase(Returns[0]);
369 // We are now done with the inlining.
373 // Otherwise, we have the normal case, of more than one block to inline or
374 // multiple return sites.
376 // We want to clone the entire callee function into the hole between the
377 // "starter" and "ender" blocks. How we accomplish this depends on whether
378 // this is an invoke instruction or a call instruction.
379 BasicBlock *AfterCallBB;
380 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
382 // Add an unconditional branch to make this look like the CallInst case...
383 BranchInst *NewBr = new BranchInst(II->getNormalDest(), TheCall);
385 // Split the basic block. This guarantees that no PHI nodes will have to be
386 // updated due to new incoming edges, and make the invoke case more
387 // symmetric to the call case.
388 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
389 CalledFunc->getName()+".exit");
391 } else { // It's a call
392 // If this is a call instruction, we need to split the basic block that
393 // the call lives in.
395 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
396 CalledFunc->getName()+".exit");
399 // Change the branch that used to go to AfterCallBB to branch to the first
400 // basic block of the inlined function.
402 TerminatorInst *Br = OrigBB->getTerminator();
403 assert(Br && Br->getOpcode() == Instruction::Br &&
404 "splitBasicBlock broken!");
405 Br->setOperand(0, FirstNewBlock);
408 // Now that the function is correct, make it a little bit nicer. In
409 // particular, move the basic blocks inserted from the end of the function
410 // into the space made by splitting the source basic block.
412 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
413 FirstNewBlock, Caller->end());
415 // Handle all of the return instructions that we just cloned in, and eliminate
416 // any users of the original call/invoke instruction.
417 if (Returns.size() > 1) {
418 // The PHI node should go at the front of the new basic block to merge all
419 // possible incoming values.
422 if (!TheCall->use_empty()) {
423 PHI = new PHINode(CalledFunc->getReturnType(),
424 TheCall->getName(), AfterCallBB->begin());
426 // Anything that used the result of the function call should now use the
427 // PHI node as their operand.
429 TheCall->replaceAllUsesWith(PHI);
432 // Loop over all of the return instructions, turning them into unconditional
433 // branches to the merge point now, and adding entries to the PHI node as
435 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
436 ReturnInst *RI = Returns[i];
439 assert(RI->getReturnValue() && "Ret should have value!");
440 assert(RI->getReturnValue()->getType() == PHI->getType() &&
441 "Ret value not consistent in function!");
442 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
445 // Add a branch to the merge point where the PHI node lives if it exists.
446 new BranchInst(AfterCallBB, RI);
448 // Delete the return instruction now
449 RI->getParent()->getInstList().erase(RI);
452 } else if (!Returns.empty()) {
453 // Otherwise, if there is exactly one return value, just replace anything
454 // using the return value of the call with the computed value.
455 if (!TheCall->use_empty())
456 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
458 // Splice the code from the return block into the block that it will return
459 // to, which contains the code that was after the call.
460 BasicBlock *ReturnBB = Returns[0]->getParent();
461 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
462 ReturnBB->getInstList());
464 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
465 ReturnBB->replaceAllUsesWith(AfterCallBB);
467 // Delete the return instruction now and empty ReturnBB now.
468 Returns[0]->eraseFromParent();
469 ReturnBB->eraseFromParent();
470 } else if (!TheCall->use_empty()) {
471 // No returns, but something is using the return value of the call. Just
473 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
476 // Since we are now done with the Call/Invoke, we can delete it.
477 TheCall->eraseFromParent();
479 // We should always be able to fold the entry block of the function into the
480 // single predecessor of the block...
481 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
482 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
484 // Splice the code entry block into calling block, right before the
485 // unconditional branch.
486 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
487 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
489 // Remove the unconditional branch.
490 OrigBB->getInstList().erase(Br);
492 // Now we can remove the CalleeEntry block, which is now empty.
493 Caller->getBasicBlockList().erase(CalleeEntry);