1 //===- InlineFunction.cpp - Code to perform function inlining -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // 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/Attributes.h"
22 #include "llvm/Analysis/CallGraph.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/Support/CallSite.h"
29 bool llvm::InlineFunction(CallInst *CI, CallGraph *CG, const TargetData *TD) {
30 return InlineFunction(CallSite(CI), CG, TD);
32 bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG, const TargetData *TD) {
33 return InlineFunction(CallSite(II), CG, TD);
36 /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls
37 /// in the body of the inlined function into invokes and turn unwind
38 /// instructions into branches to the invoke unwind dest.
40 /// II is the invoke instruction begin inlined. FirstNewBlock is the first
41 /// block of the inlined code (the last block is the end of the function),
42 /// and InlineCodeInfo is information about the code that got inlined.
43 static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock,
44 ClonedCodeInfo &InlinedCodeInfo) {
45 BasicBlock *InvokeDest = II->getUnwindDest();
46 std::vector<Value*> InvokeDestPHIValues;
48 // If there are PHI nodes in the unwind destination block, we need to
49 // keep track of which values came into them from this invoke, then remove
50 // the entry for this block.
51 BasicBlock *InvokeBlock = II->getParent();
52 for (BasicBlock::iterator I = InvokeDest->begin(); isa<PHINode>(I); ++I) {
53 PHINode *PN = cast<PHINode>(I);
54 // Save the value to use for this edge.
55 InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock));
58 Function *Caller = FirstNewBlock->getParent();
60 // The inlined code is currently at the end of the function, scan from the
61 // start of the inlined code to its end, checking for stuff we need to
63 if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) {
64 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
66 if (InlinedCodeInfo.ContainsCalls) {
67 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){
68 Instruction *I = BBI++;
70 // We only need to check for function calls: inlined invoke
71 // instructions require no special handling.
72 if (!isa<CallInst>(I)) continue;
73 CallInst *CI = cast<CallInst>(I);
75 // If this call cannot unwind, don't convert it to an invoke.
76 if (CI->doesNotThrow())
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 InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest,
88 InvokeArgs.begin(), InvokeArgs.end(),
89 CI->getName(), BB->getTerminator());
90 II->setCallingConv(CI->getCallingConv());
91 II->setAttributes(CI->getAttributes());
93 // Make sure that anything using the call now uses the invoke!
94 CI->replaceAllUsesWith(II);
96 // Delete the unconditional branch inserted by splitBasicBlock
97 BB->getInstList().pop_back();
98 Split->getInstList().pop_front(); // Delete the original call
100 // Update any PHI nodes in the exceptional block to indicate that
101 // there is now a new entry in them.
103 for (BasicBlock::iterator I = InvokeDest->begin();
104 isa<PHINode>(I); ++I, ++i) {
105 PHINode *PN = cast<PHINode>(I);
106 PN->addIncoming(InvokeDestPHIValues[i], BB);
109 // This basic block is now complete, start scanning the next one.
114 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
115 // An UnwindInst requires special handling when it gets inlined into an
116 // invoke site. Once this happens, we know that the unwind would cause
117 // a control transfer to the invoke exception destination, so we can
118 // transform it into a direct branch to the exception destination.
119 BranchInst::Create(InvokeDest, UI);
121 // Delete the unwind instruction!
122 UI->eraseFromParent();
124 // Update any PHI nodes in the exceptional block to indicate that
125 // there is now a new entry in them.
127 for (BasicBlock::iterator I = InvokeDest->begin();
128 isa<PHINode>(I); ++I, ++i) {
129 PHINode *PN = cast<PHINode>(I);
130 PN->addIncoming(InvokeDestPHIValues[i], BB);
136 // Now that everything is happy, we have one final detail. The PHI nodes in
137 // the exception destination block still have entries due to the original
138 // invoke instruction. Eliminate these entries (which might even delete the
140 InvokeDest->removePredecessor(II->getParent());
143 /// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee
144 /// into the caller, update the specified callgraph to reflect the changes we
145 /// made. Note that it's possible that not all code was copied over, so only
146 /// some edges of the callgraph may remain.
147 static void UpdateCallGraphAfterInlining(CallSite CS,
148 Function::iterator FirstNewBlock,
149 DenseMap<const Value*, Value*> &ValueMap,
151 const Function *Caller = CS.getInstruction()->getParent()->getParent();
152 const Function *Callee = CS.getCalledFunction();
153 CallGraphNode *CalleeNode = CG[Callee];
154 CallGraphNode *CallerNode = CG[Caller];
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 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end();
160 // Consider the case where CalleeNode == CallerNode.
161 CallGraphNode::CalledFunctionsVector CallCache;
162 if (CalleeNode == CallerNode) {
163 CallCache.assign(I, E);
164 I = CallCache.begin();
168 for (; I != E; ++I) {
169 const Instruction *OrigCall = I->first.getInstruction();
171 DenseMap<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall);
172 // Only copy the edge if the call was inlined!
173 if (VMI != ValueMap.end() && VMI->second) {
174 // If the call was inlined, but then constant folded, there is no edge to
175 // add. Check for this case.
176 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second))
177 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second);
180 // Update the call graph by deleting the edge from Callee to Caller. We must
181 // do this after the loop above in case Caller and Callee are the same.
182 CallerNode->removeCallEdgeFor(CS);
186 // InlineFunction - This function inlines the called function into the basic
187 // block of the caller. This returns false if it is not possible to inline this
188 // call. The program is still in a well defined state if this occurs though.
190 // Note that this only does one level of inlining. For example, if the
191 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
192 // exists in the instruction stream. Similiarly this will inline a recursive
193 // function by one level.
195 bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) {
196 Instruction *TheCall = CS.getInstruction();
197 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
198 "Instruction not in function!");
200 const Function *CalledFunc = CS.getCalledFunction();
201 if (CalledFunc == 0 || // Can't inline external function or indirect
202 CalledFunc->isDeclaration() || // call, or call to a vararg function!
203 CalledFunc->getFunctionType()->isVarArg()) return false;
206 // If the call to the callee is a non-tail call, we must clear the 'tail'
207 // flags on any calls that we inline.
208 bool MustClearTailCallFlags =
209 isa<CallInst>(TheCall) && !cast<CallInst>(TheCall)->isTailCall();
211 // If the call to the callee cannot throw, set the 'nounwind' flag on any
212 // calls that we inline.
213 bool MarkNoUnwind = CS.doesNotThrow();
215 BasicBlock *OrigBB = TheCall->getParent();
216 Function *Caller = OrigBB->getParent();
218 // GC poses two hazards to inlining, which only occur when the callee has GC:
219 // 1. If the caller has no GC, then the callee's GC must be propagated to the
221 // 2. If the caller has a differing GC, it is invalid to inline.
222 if (CalledFunc->hasGC()) {
223 if (!Caller->hasGC())
224 Caller->setGC(CalledFunc->getGC());
225 else if (CalledFunc->getGC() != Caller->getGC())
229 // Get an iterator to the last basic block in the function, which will have
230 // the new function inlined after it.
232 Function::iterator LastBlock = &Caller->back();
234 // Make sure to capture all of the return instructions from the cloned
236 std::vector<ReturnInst*> Returns;
237 ClonedCodeInfo InlinedFunctionInfo;
238 Function::iterator FirstNewBlock;
240 { // Scope to destroy ValueMap after cloning.
241 DenseMap<const Value*, Value*> ValueMap;
243 assert(CalledFunc->arg_size() == CS.arg_size() &&
244 "No varargs calls can be inlined!");
246 // Calculate the vector of arguments to pass into the function cloner, which
247 // matches up the formal to the actual argument values.
248 CallSite::arg_iterator AI = CS.arg_begin();
250 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
251 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
252 Value *ActualArg = *AI;
254 // When byval arguments actually inlined, we need to make the copy implied
255 // by them explicit. However, we don't do this if the callee is readonly
256 // or readnone, because the copy would be unneeded: the callee doesn't
257 // modify the struct.
258 if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal) &&
259 !CalledFunc->onlyReadsMemory()) {
260 const Type *AggTy = cast<PointerType>(I->getType())->getElementType();
261 const Type *VoidPtrTy = PointerType::getUnqual(Type::Int8Ty);
263 // Create the alloca. If we have TargetData, use nice alignment.
265 if (TD) Align = TD->getPrefTypeAlignment(AggTy);
266 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, I->getName(),
267 Caller->begin()->begin());
269 const Type *Tys[] = { Type::Int64Ty };
270 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(),
273 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall);
274 Value *SrcCast = new BitCastInst(*AI, VoidPtrTy, "tmp", TheCall);
278 Size = ConstantExpr::getSizeOf(AggTy);
280 Size = ConstantInt::get(Type::Int64Ty, TD->getTypeStoreSize(AggTy));
282 // Always generate a memcpy of alignment 1 here because we don't know
283 // the alignment of the src pointer. Other optimizations can infer
285 Value *CallArgs[] = {
286 DestCast, SrcCast, Size, ConstantInt::get(Type::Int32Ty, 1)
288 CallInst *TheMemCpy =
289 CallInst::Create(MemCpyFn, CallArgs, CallArgs+4, "", TheCall);
291 // If we have a call graph, update it.
293 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn);
294 CallGraphNode *CallerNode = (*CG)[Caller];
295 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN);
298 // Uses of the argument in the function should use our new alloca
300 ActualArg = NewAlloca;
303 ValueMap[I] = ActualArg;
306 // We want the inliner to prune the code as it copies. We would LOVE to
307 // have no dead or constant instructions leftover after inlining occurs
308 // (which can happen, e.g., because an argument was constant), but we'll be
309 // happy with whatever the cloner can do.
310 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i",
311 &InlinedFunctionInfo, TD);
313 // Remember the first block that is newly cloned over.
314 FirstNewBlock = LastBlock; ++FirstNewBlock;
316 // Update the callgraph if requested.
318 UpdateCallGraphAfterInlining(CS, FirstNewBlock, ValueMap, *CG);
321 // If there are any alloca instructions in the block that used to be the entry
322 // block for the callee, move them to the entry block of the caller. First
323 // calculate which instruction they should be inserted before. We insert the
324 // instructions at the end of the current alloca list.
327 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
328 for (BasicBlock::iterator I = FirstNewBlock->begin(),
329 E = FirstNewBlock->end(); I != E; )
330 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++)) {
331 // If the alloca is now dead, remove it. This often occurs due to code
333 if (AI->use_empty()) {
334 AI->eraseFromParent();
338 if (isa<Constant>(AI->getArraySize())) {
339 // Scan for the block of allocas that we can move over, and move them
341 while (isa<AllocaInst>(I) &&
342 isa<Constant>(cast<AllocaInst>(I)->getArraySize()))
345 // Transfer all of the allocas over in a block. Using splice means
346 // that the instructions aren't removed from the symbol table, then
348 Caller->getEntryBlock().getInstList().splice(
350 FirstNewBlock->getInstList(),
356 // If the inlined code contained dynamic alloca instructions, wrap the inlined
357 // code with llvm.stacksave/llvm.stackrestore intrinsics.
358 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
359 Module *M = Caller->getParent();
360 // Get the two intrinsics we care about.
361 Constant *StackSave, *StackRestore;
362 StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
363 StackRestore = Intrinsic::getDeclaration(M, Intrinsic::stackrestore);
365 // If we are preserving the callgraph, add edges to the stacksave/restore
366 // functions for the calls we insert.
367 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0;
369 // We know that StackSave/StackRestore are Function*'s, because they are
370 // intrinsics which must have the right types.
371 StackSaveCGN = CG->getOrInsertFunction(cast<Function>(StackSave));
372 StackRestoreCGN = CG->getOrInsertFunction(cast<Function>(StackRestore));
373 CallerNode = (*CG)[Caller];
376 // Insert the llvm.stacksave.
377 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack",
378 FirstNewBlock->begin());
379 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN);
381 // Insert a call to llvm.stackrestore before any return instructions in the
383 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
384 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]);
385 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN);
388 // Count the number of StackRestore calls we insert.
389 unsigned NumStackRestores = Returns.size();
391 // If we are inlining an invoke instruction, insert restores before each
392 // unwind. These unwinds will be rewritten into branches later.
393 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) {
394 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
396 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
397 CallInst::Create(StackRestore, SavedPtr, "", UI);
403 // If we are inlining tail call instruction through a call site that isn't
404 // marked 'tail', we must remove the tail marker for any calls in the inlined
405 // code. Also, calls inlined through a 'nounwind' call site should be marked
407 if (InlinedFunctionInfo.ContainsCalls &&
408 (MustClearTailCallFlags || MarkNoUnwind)) {
409 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
411 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
412 if (CallInst *CI = dyn_cast<CallInst>(I)) {
413 if (MustClearTailCallFlags)
414 CI->setTailCall(false);
416 CI->setDoesNotThrow();
420 // If we are inlining through a 'nounwind' call site then any inlined 'unwind'
421 // instructions are unreachable.
422 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind)
423 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
425 TerminatorInst *Term = BB->getTerminator();
426 if (isa<UnwindInst>(Term)) {
427 new UnreachableInst(Term);
428 BB->getInstList().erase(Term);
432 // If we are inlining for an invoke instruction, we must make sure to rewrite
433 // any inlined 'unwind' instructions into branches to the invoke exception
434 // destination, and call instructions into invoke instructions.
435 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
436 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
438 // If we cloned in _exactly one_ basic block, and if that block ends in a
439 // return instruction, we splice the body of the inlined callee directly into
440 // the calling basic block.
441 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
442 // Move all of the instructions right before the call.
443 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
444 FirstNewBlock->begin(), FirstNewBlock->end());
445 // Remove the cloned basic block.
446 Caller->getBasicBlockList().pop_back();
448 // If the call site was an invoke instruction, add a branch to the normal
450 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
451 BranchInst::Create(II->getNormalDest(), TheCall);
453 // If the return instruction returned a value, replace uses of the call with
454 // uses of the returned value.
455 if (!TheCall->use_empty()) {
456 ReturnInst *R = Returns[0];
457 TheCall->replaceAllUsesWith(R->getReturnValue());
459 // Since we are now done with the Call/Invoke, we can delete it.
460 TheCall->eraseFromParent();
462 // Since we are now done with the return instruction, delete it also.
463 Returns[0]->eraseFromParent();
465 // We are now done with the inlining.
469 // Otherwise, we have the normal case, of more than one block to inline or
470 // multiple return sites.
472 // We want to clone the entire callee function into the hole between the
473 // "starter" and "ender" blocks. How we accomplish this depends on whether
474 // this is an invoke instruction or a call instruction.
475 BasicBlock *AfterCallBB;
476 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
478 // Add an unconditional branch to make this look like the CallInst case...
479 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
481 // Split the basic block. This guarantees that no PHI nodes will have to be
482 // updated due to new incoming edges, and make the invoke case more
483 // symmetric to the call case.
484 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
485 CalledFunc->getName()+".exit");
487 } else { // It's a call
488 // If this is a call instruction, we need to split the basic block that
489 // the call lives in.
491 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
492 CalledFunc->getName()+".exit");
495 // Change the branch that used to go to AfterCallBB to branch to the first
496 // basic block of the inlined function.
498 TerminatorInst *Br = OrigBB->getTerminator();
499 assert(Br && Br->getOpcode() == Instruction::Br &&
500 "splitBasicBlock broken!");
501 Br->setOperand(0, FirstNewBlock);
504 // Now that the function is correct, make it a little bit nicer. In
505 // particular, move the basic blocks inserted from the end of the function
506 // into the space made by splitting the source basic block.
507 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
508 FirstNewBlock, Caller->end());
510 // Handle all of the return instructions that we just cloned in, and eliminate
511 // any users of the original call/invoke instruction.
512 const Type *RTy = CalledFunc->getReturnType();
514 if (Returns.size() > 1) {
515 // The PHI node should go at the front of the new basic block to merge all
516 // possible incoming values.
518 if (!TheCall->use_empty()) {
519 PHI = PHINode::Create(RTy, TheCall->getName(),
520 AfterCallBB->begin());
521 // Anything that used the result of the function call should now use the
522 // PHI node as their operand.
523 TheCall->replaceAllUsesWith(PHI);
526 // Loop over all of the return instructions adding entries to the PHI node
529 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
530 ReturnInst *RI = Returns[i];
531 assert(RI->getReturnValue()->getType() == PHI->getType() &&
532 "Ret value not consistent in function!");
533 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
537 // Add a branch to the merge points and remove return instructions.
538 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
539 ReturnInst *RI = Returns[i];
540 BranchInst::Create(AfterCallBB, RI);
541 RI->eraseFromParent();
543 } else if (!Returns.empty()) {
544 // Otherwise, if there is exactly one return value, just replace anything
545 // using the return value of the call with the computed value.
546 if (!TheCall->use_empty())
547 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
549 // Splice the code from the return block into the block that it will return
550 // to, which contains the code that was after the call.
551 BasicBlock *ReturnBB = Returns[0]->getParent();
552 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
553 ReturnBB->getInstList());
555 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
556 ReturnBB->replaceAllUsesWith(AfterCallBB);
558 // Delete the return instruction now and empty ReturnBB now.
559 Returns[0]->eraseFromParent();
560 ReturnBB->eraseFromParent();
561 } else if (!TheCall->use_empty()) {
562 // No returns, but something is using the return value of the call. Just
564 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
567 // Since we are now done with the Call/Invoke, we can delete it.
568 TheCall->eraseFromParent();
570 // We should always be able to fold the entry block of the function into the
571 // single predecessor of the block...
572 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
573 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
575 // Splice the code entry block into calling block, right before the
576 // unconditional branch.
577 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
578 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
580 // Remove the unconditional branch.
581 OrigBB->getInstList().erase(Br);
583 // Now we can remove the CalleeEntry block, which is now empty.
584 Caller->getBasicBlockList().erase(CalleeEntry);