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/ParameterAttributes.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->setParamAttrs(CI->getParamAttrs());
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 will be remain.
147 static void UpdateCallGraphAfterInlining(const Function *Caller,
148 const Function *Callee,
149 Function::iterator FirstNewBlock,
150 DenseMap<const Value*, Value*> &ValueMap,
152 // Update the call graph by deleting the edge from Callee to Caller
153 CallGraphNode *CalleeNode = CG[Callee];
154 CallGraphNode *CallerNode = CG[Caller];
155 CallerNode->removeCallEdgeTo(CalleeNode);
157 // Since we inlined some uninlined call sites in the callee into the caller,
158 // add edges from the caller to all of the callees of the callee.
159 for (CallGraphNode::iterator I = CalleeNode->begin(),
160 E = CalleeNode->end(); I != E; ++I) {
161 const Instruction *OrigCall = I->first.getInstruction();
163 DenseMap<const Value*, Value*>::iterator VMI = ValueMap.find(OrigCall);
164 // Only copy the edge if the call was inlined!
165 if (VMI != ValueMap.end() && VMI->second) {
166 // If the call was inlined, but then constant folded, there is no edge to
167 // add. Check for this case.
168 if (Instruction *NewCall = dyn_cast<Instruction>(VMI->second))
169 CallerNode->addCalledFunction(CallSite::get(NewCall), I->second);
175 // InlineFunction - This function inlines the called function into the basic
176 // block of the caller. This returns false if it is not possible to inline this
177 // call. The program is still in a well defined state if this occurs though.
179 // Note that this only does one level of inlining. For example, if the
180 // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
181 // exists in the instruction stream. Similiarly this will inline a recursive
182 // function by one level.
184 bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) {
185 Instruction *TheCall = CS.getInstruction();
186 assert(TheCall->getParent() && TheCall->getParent()->getParent() &&
187 "Instruction not in function!");
189 const Function *CalledFunc = CS.getCalledFunction();
190 if (CalledFunc == 0 || // Can't inline external function or indirect
191 CalledFunc->isDeclaration() || // call, or call to a vararg function!
192 CalledFunc->getFunctionType()->isVarArg()) return false;
195 // If the call to the callee is a non-tail call, we must clear the 'tail'
196 // flags on any calls that we inline.
197 bool MustClearTailCallFlags =
198 isa<CallInst>(TheCall) && !cast<CallInst>(TheCall)->isTailCall();
200 // If the call to the callee cannot throw, set the 'nounwind' flag on any
201 // calls that we inline.
202 bool MarkNoUnwind = CS.doesNotThrow();
204 BasicBlock *OrigBB = TheCall->getParent();
205 Function *Caller = OrigBB->getParent();
207 // GC poses two hazards to inlining, which only occur when the callee has GC:
208 // 1. If the caller has no GC, then the callee's GC must be propagated to the
210 // 2. If the caller has a differing GC, it is invalid to inline.
211 if (CalledFunc->hasCollector()) {
212 if (!Caller->hasCollector())
213 Caller->setCollector(CalledFunc->getCollector());
214 else if (CalledFunc->getCollector() != Caller->getCollector())
218 // Get an iterator to the last basic block in the function, which will have
219 // the new function inlined after it.
221 Function::iterator LastBlock = &Caller->back();
223 // Make sure to capture all of the return instructions from the cloned
225 std::vector<ReturnInst*> Returns;
226 ClonedCodeInfo InlinedFunctionInfo;
227 Function::iterator FirstNewBlock;
229 { // Scope to destroy ValueMap after cloning.
230 DenseMap<const Value*, Value*> ValueMap;
232 assert(CalledFunc->arg_size() == CS.arg_size() &&
233 "No varargs calls can be inlined!");
235 // Calculate the vector of arguments to pass into the function cloner, which
236 // matches up the formal to the actual argument values.
237 CallSite::arg_iterator AI = CS.arg_begin();
239 for (Function::const_arg_iterator I = CalledFunc->arg_begin(),
240 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) {
241 Value *ActualArg = *AI;
243 // When byval arguments actually inlined, we need to make the copy implied
244 // by them explicit. However, we don't do this if the callee is readonly
245 // or readnone, because the copy would be unneeded: the callee doesn't
246 // modify the struct.
247 if (CalledFunc->paramHasAttr(ArgNo+1, ParamAttr::ByVal) &&
248 !CalledFunc->onlyReadsMemory()) {
249 const Type *AggTy = cast<PointerType>(I->getType())->getElementType();
250 const Type *VoidPtrTy = PointerType::getUnqual(Type::Int8Ty);
252 // Create the alloca. If we have TargetData, use nice alignment.
254 if (TD) Align = TD->getPrefTypeAlignment(AggTy);
255 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, I->getName(),
256 Caller->begin()->begin());
258 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(),
259 Intrinsic::memcpy_i64);
260 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall);
261 Value *SrcCast = new BitCastInst(*AI, VoidPtrTy, "tmp", TheCall);
265 Size = ConstantExpr::getSizeOf(AggTy);
267 Size = ConstantInt::get(Type::Int64Ty, TD->getTypeStoreSize(AggTy));
269 // Always generate a memcpy of alignment 1 here because we don't know
270 // the alignment of the src pointer. Other optimizations can infer
272 Value *CallArgs[] = {
273 DestCast, SrcCast, Size, ConstantInt::get(Type::Int32Ty, 1)
275 CallInst *TheMemCpy =
276 CallInst::Create(MemCpyFn, CallArgs, CallArgs+4, "", TheCall);
278 // If we have a call graph, update it.
280 CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn);
281 CallGraphNode *CallerNode = (*CG)[Caller];
282 CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN);
285 // Uses of the argument in the function should use our new alloca
287 ActualArg = NewAlloca;
290 ValueMap[I] = ActualArg;
293 // We want the inliner to prune the code as it copies. We would LOVE to
294 // have no dead or constant instructions leftover after inlining occurs
295 // (which can happen, e.g., because an argument was constant), but we'll be
296 // happy with whatever the cloner can do.
297 CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i",
298 &InlinedFunctionInfo, TD);
300 // Remember the first block that is newly cloned over.
301 FirstNewBlock = LastBlock; ++FirstNewBlock;
303 // Update the callgraph if requested.
305 UpdateCallGraphAfterInlining(Caller, CalledFunc, FirstNewBlock, ValueMap,
309 // If there are any alloca instructions in the block that used to be the entry
310 // block for the callee, move them to the entry block of the caller. First
311 // calculate which instruction they should be inserted before. We insert the
312 // instructions at the end of the current alloca list.
315 BasicBlock::iterator InsertPoint = Caller->begin()->begin();
316 for (BasicBlock::iterator I = FirstNewBlock->begin(),
317 E = FirstNewBlock->end(); I != E; )
318 if (AllocaInst *AI = dyn_cast<AllocaInst>(I++)) {
319 // If the alloca is now dead, remove it. This often occurs due to code
321 if (AI->use_empty()) {
322 AI->eraseFromParent();
326 if (isa<Constant>(AI->getArraySize())) {
327 // Scan for the block of allocas that we can move over, and move them
329 while (isa<AllocaInst>(I) &&
330 isa<Constant>(cast<AllocaInst>(I)->getArraySize()))
333 // Transfer all of the allocas over in a block. Using splice means
334 // that the instructions aren't removed from the symbol table, then
336 Caller->getEntryBlock().getInstList().splice(
338 FirstNewBlock->getInstList(),
344 // If the inlined code contained dynamic alloca instructions, wrap the inlined
345 // code with llvm.stacksave/llvm.stackrestore intrinsics.
346 if (InlinedFunctionInfo.ContainsDynamicAllocas) {
347 Module *M = Caller->getParent();
348 // Get the two intrinsics we care about.
349 Constant *StackSave, *StackRestore;
350 StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave);
351 StackRestore = Intrinsic::getDeclaration(M, Intrinsic::stackrestore);
353 // If we are preserving the callgraph, add edges to the stacksave/restore
354 // functions for the calls we insert.
355 CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0;
357 // We know that StackSave/StackRestore are Function*'s, because they are
358 // intrinsics which must have the right types.
359 StackSaveCGN = CG->getOrInsertFunction(cast<Function>(StackSave));
360 StackRestoreCGN = CG->getOrInsertFunction(cast<Function>(StackRestore));
361 CallerNode = (*CG)[Caller];
364 // Insert the llvm.stacksave.
365 CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack",
366 FirstNewBlock->begin());
367 if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN);
369 // Insert a call to llvm.stackrestore before any return instructions in the
371 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
372 CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]);
373 if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN);
376 // Count the number of StackRestore calls we insert.
377 unsigned NumStackRestores = Returns.size();
379 // If we are inlining an invoke instruction, insert restores before each
380 // unwind. These unwinds will be rewritten into branches later.
381 if (InlinedFunctionInfo.ContainsUnwinds && isa<InvokeInst>(TheCall)) {
382 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
384 if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) {
385 CallInst::Create(StackRestore, SavedPtr, "", UI);
391 // If we are inlining tail call instruction through a call site that isn't
392 // marked 'tail', we must remove the tail marker for any calls in the inlined
393 // code. Also, calls inlined through a 'nounwind' call site should be marked
395 if (InlinedFunctionInfo.ContainsCalls &&
396 (MustClearTailCallFlags || MarkNoUnwind)) {
397 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
399 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
400 if (CallInst *CI = dyn_cast<CallInst>(I)) {
401 if (MustClearTailCallFlags)
402 CI->setTailCall(false);
404 CI->setDoesNotThrow();
408 // If we are inlining through a 'nounwind' call site then any inlined 'unwind'
409 // instructions are unreachable.
410 if (InlinedFunctionInfo.ContainsUnwinds && MarkNoUnwind)
411 for (Function::iterator BB = FirstNewBlock, E = Caller->end();
413 TerminatorInst *Term = BB->getTerminator();
414 if (isa<UnwindInst>(Term)) {
415 new UnreachableInst(Term);
416 BB->getInstList().erase(Term);
420 // If we are inlining for an invoke instruction, we must make sure to rewrite
421 // any inlined 'unwind' instructions into branches to the invoke exception
422 // destination, and call instructions into invoke instructions.
423 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
424 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo);
426 // If we cloned in _exactly one_ basic block, and if that block ends in a
427 // return instruction, we splice the body of the inlined callee directly into
428 // the calling basic block.
429 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) {
430 // Move all of the instructions right before the call.
431 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(),
432 FirstNewBlock->begin(), FirstNewBlock->end());
433 // Remove the cloned basic block.
434 Caller->getBasicBlockList().pop_back();
436 // If the call site was an invoke instruction, add a branch to the normal
438 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall))
439 BranchInst::Create(II->getNormalDest(), TheCall);
441 // If the return instruction returned a value, replace uses of the call with
442 // uses of the returned value.
443 if (!TheCall->use_empty()) {
444 ReturnInst *R = Returns[0];
445 TheCall->replaceAllUsesWith(R->getReturnValue());
447 // Since we are now done with the Call/Invoke, we can delete it.
448 TheCall->eraseFromParent();
450 // Since we are now done with the return instruction, delete it also.
451 Returns[0]->eraseFromParent();
453 // We are now done with the inlining.
457 // Otherwise, we have the normal case, of more than one block to inline or
458 // multiple return sites.
460 // We want to clone the entire callee function into the hole between the
461 // "starter" and "ender" blocks. How we accomplish this depends on whether
462 // this is an invoke instruction or a call instruction.
463 BasicBlock *AfterCallBB;
464 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
466 // Add an unconditional branch to make this look like the CallInst case...
467 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall);
469 // Split the basic block. This guarantees that no PHI nodes will have to be
470 // updated due to new incoming edges, and make the invoke case more
471 // symmetric to the call case.
472 AfterCallBB = OrigBB->splitBasicBlock(NewBr,
473 CalledFunc->getName()+".exit");
475 } else { // It's a call
476 // If this is a call instruction, we need to split the basic block that
477 // the call lives in.
479 AfterCallBB = OrigBB->splitBasicBlock(TheCall,
480 CalledFunc->getName()+".exit");
483 // Change the branch that used to go to AfterCallBB to branch to the first
484 // basic block of the inlined function.
486 TerminatorInst *Br = OrigBB->getTerminator();
487 assert(Br && Br->getOpcode() == Instruction::Br &&
488 "splitBasicBlock broken!");
489 Br->setOperand(0, FirstNewBlock);
492 // Now that the function is correct, make it a little bit nicer. In
493 // particular, move the basic blocks inserted from the end of the function
494 // into the space made by splitting the source basic block.
495 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(),
496 FirstNewBlock, Caller->end());
498 // Handle all of the return instructions that we just cloned in, and eliminate
499 // any users of the original call/invoke instruction.
500 const Type *RTy = CalledFunc->getReturnType();
502 if (Returns.size() > 1) {
503 // The PHI node should go at the front of the new basic block to merge all
504 // possible incoming values.
506 if (!TheCall->use_empty()) {
507 PHI = PHINode::Create(RTy, TheCall->getName(),
508 AfterCallBB->begin());
509 // Anything that used the result of the function call should now use the
510 // PHI node as their operand.
511 TheCall->replaceAllUsesWith(PHI);
514 // Loop over all of the return instructions adding entries to the PHI node as
517 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
518 ReturnInst *RI = Returns[i];
519 assert(RI->getReturnValue()->getType() == PHI->getType() &&
520 "Ret value not consistent in function!");
521 PHI->addIncoming(RI->getReturnValue(), RI->getParent());
525 // Add a branch to the merge points and remove retrun instructions.
526 for (unsigned i = 0, e = Returns.size(); i != e; ++i) {
527 ReturnInst *RI = Returns[i];
528 BranchInst::Create(AfterCallBB, RI);
529 RI->eraseFromParent();
531 } else if (!Returns.empty()) {
532 // Otherwise, if there is exactly one return value, just replace anything
533 // using the return value of the call with the computed value.
534 if (!TheCall->use_empty())
535 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue());
537 // Splice the code from the return block into the block that it will return
538 // to, which contains the code that was after the call.
539 BasicBlock *ReturnBB = Returns[0]->getParent();
540 AfterCallBB->getInstList().splice(AfterCallBB->begin(),
541 ReturnBB->getInstList());
543 // Update PHI nodes that use the ReturnBB to use the AfterCallBB.
544 ReturnBB->replaceAllUsesWith(AfterCallBB);
546 // Delete the return instruction now and empty ReturnBB now.
547 Returns[0]->eraseFromParent();
548 ReturnBB->eraseFromParent();
549 } else if (!TheCall->use_empty()) {
550 // No returns, but something is using the return value of the call. Just
552 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType()));
555 // Since we are now done with the Call/Invoke, we can delete it.
556 TheCall->eraseFromParent();
558 // We should always be able to fold the entry block of the function into the
559 // single predecessor of the block...
560 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!");
561 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0);
563 // Splice the code entry block into calling block, right before the
564 // unconditional branch.
565 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList());
566 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes
568 // Remove the unconditional branch.
569 OrigBB->getInstList().erase(Br);
571 // Now we can remove the CalleeEntry block, which is now empty.
572 Caller->getBasicBlockList().erase(CalleeEntry);