X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FUtils%2FInlineFunction.cpp;h=348c3e49ab6a1285cb8b679ae67767abe7f617c4;hb=2280ebd61416b73d0b6137f275b25af82e268d1f;hp=cee224aae52400ff9b6727ff40068ed053e9c66a;hpb=125329891f97baedef21e4b464ba70182c3fb45e;p=oota-llvm.git diff --git a/lib/Transforms/Utils/InlineFunction.cpp b/lib/Transforms/Utils/InlineFunction.cpp index cee224aae52..348c3e49ab6 100644 --- a/lib/Transforms/Utils/InlineFunction.cpp +++ b/lib/Transforms/Utils/InlineFunction.cpp @@ -10,6 +10,13 @@ // This file implements inlining of a function into a call site, resolving // parameters and the return value as appropriate. // +// The code in this file for handling inlines through invoke +// instructions preserves semantics only under some assumptions about +// the behavior of unwinders which correspond to gcc-style libUnwind +// exception personality functions. Eventually the IR will be +// improved to make this unnecessary, but until then, this code is +// marked [LIBUNWIND]. +// //===----------------------------------------------------------------------===// #include "llvm/Transforms/Utils/Cloning.h" @@ -17,119 +24,541 @@ #include "llvm/DerivedTypes.h" #include "llvm/Module.h" #include "llvm/Instructions.h" +#include "llvm/IntrinsicInst.h" #include "llvm/Intrinsics.h" #include "llvm/Attributes.h" #include "llvm/Analysis/CallGraph.h" +#include "llvm/Analysis/DebugInfo.h" +#include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Target/TargetData.h" +#include "llvm/Transforms/Utils/Local.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/StringExtras.h" #include "llvm/Support/CallSite.h" +#include "llvm/Support/IRBuilder.h" using namespace llvm; -bool llvm::InlineFunction(CallInst *CI, CallGraph *CG, const TargetData *TD) { - return InlineFunction(CallSite(CI), CG, TD); +bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI) { + return InlineFunction(CallSite(CI), IFI); +} +bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI) { + return InlineFunction(CallSite(II), IFI); +} + +/// [LIBUNWIND] Look for an llvm.eh.exception call in the given block. +static EHExceptionInst *findExceptionInBlock(BasicBlock *bb) { + for (BasicBlock::iterator i = bb->begin(), e = bb->end(); i != e; i++) { + EHExceptionInst *exn = dyn_cast(i); + if (exn) return exn; + } + + return 0; +} + +/// [LIBUNWIND] Look for the 'best' llvm.eh.selector instruction for +/// the given llvm.eh.exception call. +static EHSelectorInst *findSelectorForException(EHExceptionInst *exn) { + BasicBlock *exnBlock = exn->getParent(); + + EHSelectorInst *outOfBlockSelector = 0; + for (Instruction::use_iterator + ui = exn->use_begin(), ue = exn->use_end(); ui != ue; ++ui) { + EHSelectorInst *sel = dyn_cast(*ui); + if (!sel) continue; + + // Immediately accept an eh.selector in the same block as the + // excepton call. + if (sel->getParent() == exnBlock) return sel; + + // Otherwise, use the first selector we see. + if (!outOfBlockSelector) outOfBlockSelector = sel; + } + + return outOfBlockSelector; } -bool llvm::InlineFunction(InvokeInst *II, CallGraph *CG, const TargetData *TD) { - return InlineFunction(CallSite(II), CG, TD); + +/// [LIBUNWIND] Find the (possibly absent) call to @llvm.eh.selector +/// in the given landing pad. In principle, llvm.eh.exception is +/// required to be in the landing pad; in practice, SplitCriticalEdge +/// can break that invariant, and then inlining can break it further. +/// There's a real need for a reliable solution here, but until that +/// happens, we have some fragile workarounds here. +static EHSelectorInst *findSelectorForLandingPad(BasicBlock *lpad) { + // Look for an exception call in the actual landing pad. + EHExceptionInst *exn = findExceptionInBlock(lpad); + if (exn) return findSelectorForException(exn); + + // Okay, if that failed, look for one in an obvious successor. If + // we find one, we'll fix the IR by moving things back to the + // landing pad. + + bool dominates = true; // does the lpad dominate the exn call + BasicBlock *nonDominated = 0; // if not, the first non-dominated block + BasicBlock *lastDominated = 0; // and the block which branched to it + + BasicBlock *exnBlock = lpad; + + // We need to protect against lpads that lead into infinite loops. + SmallPtrSet visited; + visited.insert(exnBlock); + + do { + // We're not going to apply this hack to anything more complicated + // than a series of unconditional branches, so if the block + // doesn't terminate in an unconditional branch, just fail. More + // complicated cases can arise when, say, sinking a call into a + // split unwind edge and then inlining it; but that can do almost + // *anything* to the CFG, including leaving the selector + // completely unreachable. The only way to fix that properly is + // to (1) prohibit transforms which move the exception or selector + // values away from the landing pad, e.g. by producing them with + // instructions that are pinned to an edge like a phi, or + // producing them with not-really-instructions, and (2) making + // transforms which split edges deal with that. + BranchInst *branch = dyn_cast(&exnBlock->back()); + if (!branch || branch->isConditional()) return 0; + + BasicBlock *successor = branch->getSuccessor(0); + + // Fail if we found an infinite loop. + if (!visited.insert(successor)) return 0; + + // If the successor isn't dominated by exnBlock: + if (!successor->getSinglePredecessor()) { + // We don't want to have to deal with threading the exception + // through multiple levels of phi, so give up if we've already + // followed a non-dominating edge. + if (!dominates) return 0; + + // Otherwise, remember this as a non-dominating edge. + dominates = false; + nonDominated = successor; + lastDominated = exnBlock; + } + + exnBlock = successor; + + // Can we stop here? + exn = findExceptionInBlock(exnBlock); + } while (!exn); + + // Look for a selector call for the exception we found. + EHSelectorInst *selector = findSelectorForException(exn); + if (!selector) return 0; + + // The easy case is when the landing pad still dominates the + // exception call, in which case we can just move both calls back to + // the landing pad. + if (dominates) { + selector->moveBefore(lpad->getFirstNonPHI()); + exn->moveBefore(selector); + return selector; + } + + // Otherwise, we have to split at the first non-dominating block. + // The CFG looks basically like this: + // lpad: + // phis_0 + // insnsAndBranches_1 + // br label %nonDominated + // nonDominated: + // phis_2 + // insns_3 + // %exn = call i8* @llvm.eh.exception() + // insnsAndBranches_4 + // %selector = call @llvm.eh.selector(i8* %exn, ... + // We need to turn this into: + // lpad: + // phis_0 + // %exn0 = call i8* @llvm.eh.exception() + // %selector0 = call @llvm.eh.selector(i8* %exn0, ... + // insnsAndBranches_1 + // br label %split // from lastDominated + // nonDominated: + // phis_2 (without edge from lastDominated) + // %exn1 = call i8* @llvm.eh.exception() + // %selector1 = call i8* @llvm.eh.selector(i8* %exn1, ... + // br label %split + // split: + // phis_2 (edge from lastDominated, edge from split) + // %exn = phi ... + // %selector = phi ... + // insns_3 + // insnsAndBranches_4 + + assert(nonDominated); + assert(lastDominated); + + // First, make clones of the intrinsics to go in lpad. + EHExceptionInst *lpadExn = cast(exn->clone()); + EHSelectorInst *lpadSelector = cast(selector->clone()); + lpadSelector->setArgOperand(0, lpadExn); + lpadSelector->insertBefore(lpad->getFirstNonPHI()); + lpadExn->insertBefore(lpadSelector); + + // Split the non-dominated block. + BasicBlock *split = + nonDominated->splitBasicBlock(nonDominated->getFirstNonPHI(), + nonDominated->getName() + ".lpad-fix"); + + // Redirect the last dominated branch there. + cast(lastDominated->back()).setSuccessor(0, split); + + // Move the existing intrinsics to the end of the old block. + selector->moveBefore(&nonDominated->back()); + exn->moveBefore(selector); + + Instruction *splitIP = &split->front(); + + // For all the phis in nonDominated, make a new phi in split to join + // that phi with the edge from lastDominated. + for (BasicBlock::iterator + i = nonDominated->begin(), e = nonDominated->end(); i != e; ++i) { + PHINode *phi = dyn_cast(i); + if (!phi) break; + + PHINode *splitPhi = PHINode::Create(phi->getType(), 2, phi->getName(), + splitIP); + phi->replaceAllUsesWith(splitPhi); + splitPhi->addIncoming(phi, nonDominated); + splitPhi->addIncoming(phi->removeIncomingValue(lastDominated), + lastDominated); + } + + // Make new phis for the exception and selector. + PHINode *exnPhi = PHINode::Create(exn->getType(), 2, "", splitIP); + exn->replaceAllUsesWith(exnPhi); + selector->setArgOperand(0, exn); // except for this use + exnPhi->addIncoming(exn, nonDominated); + exnPhi->addIncoming(lpadExn, lastDominated); + + PHINode *selectorPhi = PHINode::Create(selector->getType(), 2, "", splitIP); + selector->replaceAllUsesWith(selectorPhi); + selectorPhi->addIncoming(selector, nonDominated); + selectorPhi->addIncoming(lpadSelector, lastDominated); + + return lpadSelector; +} + +namespace { + /// A class for recording information about inlining through an invoke. + class InvokeInliningInfo { + BasicBlock *OuterUnwindDest; + EHSelectorInst *OuterSelector; + BasicBlock *InnerUnwindDest; + PHINode *InnerExceptionPHI; + PHINode *InnerSelectorPHI; + SmallVector UnwindDestPHIValues; + + public: + InvokeInliningInfo(InvokeInst *II) : + OuterUnwindDest(II->getUnwindDest()), OuterSelector(0), + InnerUnwindDest(0), InnerExceptionPHI(0), InnerSelectorPHI(0) { + + // If there are PHI nodes in the unwind destination block, we + // need to keep track of which values came into them from the + // invoke before removing the edge from this block. + llvm::BasicBlock *invokeBB = II->getParent(); + for (BasicBlock::iterator I = OuterUnwindDest->begin(); + isa(I); ++I) { + // Save the value to use for this edge. + PHINode *phi = cast(I); + UnwindDestPHIValues.push_back(phi->getIncomingValueForBlock(invokeBB)); + } + } + + /// The outer unwind destination is the target of unwind edges + /// introduced for calls within the inlined function. + BasicBlock *getOuterUnwindDest() const { + return OuterUnwindDest; + } + + EHSelectorInst *getOuterSelector() { + if (!OuterSelector) + OuterSelector = findSelectorForLandingPad(OuterUnwindDest); + return OuterSelector; + } + + BasicBlock *getInnerUnwindDest(); + + bool forwardEHResume(CallInst *call, BasicBlock *src); + + /// Add incoming-PHI values to the unwind destination block for + /// the given basic block, using the values for the original + /// invoke's source block. + void addIncomingPHIValuesFor(BasicBlock *BB) const { + addIncomingPHIValuesForInto(BB, OuterUnwindDest); + } + + void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { + BasicBlock::iterator I = dest->begin(); + for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { + PHINode *phi = cast(I); + phi->addIncoming(UnwindDestPHIValues[i], src); + } + } + }; +} + +/// Get or create a target for the branch out of rewritten calls to +/// llvm.eh.resume. +BasicBlock *InvokeInliningInfo::getInnerUnwindDest() { + if (InnerUnwindDest) return InnerUnwindDest; + + // Find and hoist the llvm.eh.exception and llvm.eh.selector calls + // in the outer landing pad to immediately following the phis. + EHSelectorInst *selector = getOuterSelector(); + if (!selector) return 0; + + // The call to llvm.eh.exception *must* be in the landing pad. + Instruction *exn = cast(selector->getArgOperand(0)); + assert(exn->getParent() == OuterUnwindDest); + + // TODO: recognize when we've already done this, so that we don't + // get a linear number of these when inlining calls into lots of + // invokes with the same landing pad. + + // Do the hoisting. + Instruction *splitPoint = exn->getParent()->getFirstNonPHI(); + assert(splitPoint != selector && "selector-on-exception dominance broken!"); + if (splitPoint == exn) { + selector->removeFromParent(); + selector->insertAfter(exn); + splitPoint = selector->getNextNode(); + } else { + exn->moveBefore(splitPoint); + selector->moveBefore(splitPoint); + } + + // Split the landing pad. + InnerUnwindDest = OuterUnwindDest->splitBasicBlock(splitPoint, + OuterUnwindDest->getName() + ".body"); + + // The number of incoming edges we expect to the inner landing pad. + const unsigned phiCapacity = 2; + + // Create corresponding new phis for all the phis in the outer landing pad. + BasicBlock::iterator insertPoint = InnerUnwindDest->begin(); + BasicBlock::iterator I = OuterUnwindDest->begin(); + for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { + PHINode *outerPhi = cast(I); + PHINode *innerPhi = PHINode::Create(outerPhi->getType(), phiCapacity, + outerPhi->getName() + ".lpad-body", + insertPoint); + outerPhi->replaceAllUsesWith(innerPhi); + innerPhi->addIncoming(outerPhi, OuterUnwindDest); + } + + // Create a phi for the exception value... + InnerExceptionPHI = PHINode::Create(exn->getType(), phiCapacity, + "exn.lpad-body", insertPoint); + exn->replaceAllUsesWith(InnerExceptionPHI); + selector->setArgOperand(0, exn); // restore this use + InnerExceptionPHI->addIncoming(exn, OuterUnwindDest); + + // ...and the selector. + InnerSelectorPHI = PHINode::Create(selector->getType(), phiCapacity, + "selector.lpad-body", insertPoint); + selector->replaceAllUsesWith(InnerSelectorPHI); + InnerSelectorPHI->addIncoming(selector, OuterUnwindDest); + + // All done. + return InnerUnwindDest; +} + +/// [LIBUNWIND] Try to forward the given call, which logically occurs +/// at the end of the given block, as a branch to the inner unwind +/// block. Returns true if the call was forwarded. +bool InvokeInliningInfo::forwardEHResume(CallInst *call, BasicBlock *src) { + // First, check whether this is a call to the intrinsic. + Function *fn = dyn_cast(call->getCalledValue()); + if (!fn || fn->getName() != "llvm.eh.resume") + return false; + + // At this point, we need to return true on all paths, because + // otherwise we'll construct an invoke of the intrinsic, which is + // not well-formed. + + // Try to find or make an inner unwind dest, which will fail if we + // can't find a selector call for the outer unwind dest. + BasicBlock *dest = getInnerUnwindDest(); + bool hasSelector = (dest != 0); + + // If we failed, just use the outer unwind dest, dropping the + // exception and selector on the floor. + if (!hasSelector) + dest = OuterUnwindDest; + + // Make a branch. + BranchInst::Create(dest, src); + + // Update the phis in the destination. They were inserted in an + // order which makes this work. + addIncomingPHIValuesForInto(src, dest); + + if (hasSelector) { + InnerExceptionPHI->addIncoming(call->getArgOperand(0), src); + InnerSelectorPHI->addIncoming(call->getArgOperand(1), src); + } + + return true; +} + +/// [LIBUNWIND] Check whether this selector is "only cleanups": +/// call i32 @llvm.eh.selector(blah, blah, i32 0) +static bool isCleanupOnlySelector(EHSelectorInst *selector) { + if (selector->getNumArgOperands() != 3) return false; + ConstantInt *val = dyn_cast(selector->getArgOperand(2)); + return (val && val->isZero()); +} + +/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into +/// an invoke, we have to turn all of the calls that can throw into +/// invokes. This function analyze BB to see if there are any calls, and if so, +/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI +/// nodes in that block with the values specified in InvokeDestPHIValues. +/// +/// Returns true to indicate that the next block should be skipped. +static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, + InvokeInliningInfo &Invoke) { + for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { + Instruction *I = BBI++; + + // We only need to check for function calls: inlined invoke + // instructions require no special handling. + CallInst *CI = dyn_cast(I); + if (CI == 0) continue; + + // LIBUNWIND: merge selector instructions. + if (EHSelectorInst *Inner = dyn_cast(CI)) { + EHSelectorInst *Outer = Invoke.getOuterSelector(); + if (!Outer) continue; + + bool innerIsOnlyCleanup = isCleanupOnlySelector(Inner); + bool outerIsOnlyCleanup = isCleanupOnlySelector(Outer); + + // If both selectors contain only cleanups, we don't need to do + // anything. TODO: this is really just a very specific instance + // of a much more general optimization. + if (innerIsOnlyCleanup && outerIsOnlyCleanup) continue; + + // Otherwise, we just append the outer selector to the inner selector. + SmallVector NewSelector; + for (unsigned i = 0, e = Inner->getNumArgOperands(); i != e; ++i) + NewSelector.push_back(Inner->getArgOperand(i)); + for (unsigned i = 2, e = Outer->getNumArgOperands(); i != e; ++i) + NewSelector.push_back(Outer->getArgOperand(i)); + + CallInst *NewInner = + IRBuilder<>(Inner).CreateCall(Inner->getCalledValue(), + NewSelector.begin(), + NewSelector.end()); + // No need to copy attributes, calling convention, etc. + NewInner->takeName(Inner); + Inner->replaceAllUsesWith(NewInner); + Inner->eraseFromParent(); + continue; + } + + // If this call cannot unwind, don't convert it to an invoke. + if (CI->doesNotThrow()) + continue; + + // Convert this function call into an invoke instruction. + // First, split the basic block. + BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); + + // Delete the unconditional branch inserted by splitBasicBlock + BB->getInstList().pop_back(); + + // LIBUNWIND: If this is a call to @llvm.eh.resume, just branch + // directly to the new landing pad. + if (Invoke.forwardEHResume(CI, BB)) { + // TODO: 'Split' is now unreachable; clean it up. + + // We want to leave the original call intact so that the call + // graph and other structures won't get misled. We also have to + // avoid processing the next block, or we'll iterate here forever. + return true; + } + + // Otherwise, create the new invoke instruction. + ImmutableCallSite CS(CI); + SmallVector InvokeArgs(CS.arg_begin(), CS.arg_end()); + InvokeInst *II = + InvokeInst::Create(CI->getCalledValue(), Split, + Invoke.getOuterUnwindDest(), + InvokeArgs.begin(), InvokeArgs.end(), + CI->getName(), BB); + II->setCallingConv(CI->getCallingConv()); + II->setAttributes(CI->getAttributes()); + + // Make sure that anything using the call now uses the invoke! This also + // updates the CallGraph if present, because it uses a WeakVH. + CI->replaceAllUsesWith(II); + + Split->getInstList().pop_front(); // Delete the original call + + // Update any PHI nodes in the exceptional block to indicate that + // there is now a new entry in them. + Invoke.addIncomingPHIValuesFor(BB); + return false; + } + + return false; } + /// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls /// in the body of the inlined function into invokes and turn unwind /// instructions into branches to the invoke unwind dest. /// -/// II is the invoke instruction begin inlined. FirstNewBlock is the first +/// II is the invoke instruction being inlined. FirstNewBlock is the first /// block of the inlined code (the last block is the end of the function), /// and InlineCodeInfo is information about the code that got inlined. static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, ClonedCodeInfo &InlinedCodeInfo) { BasicBlock *InvokeDest = II->getUnwindDest(); - std::vector InvokeDestPHIValues; - - // If there are PHI nodes in the unwind destination block, we need to - // keep track of which values came into them from this invoke, then remove - // the entry for this block. - BasicBlock *InvokeBlock = II->getParent(); - for (BasicBlock::iterator I = InvokeDest->begin(); isa(I); ++I) { - PHINode *PN = cast(I); - // Save the value to use for this edge. - InvokeDestPHIValues.push_back(PN->getIncomingValueForBlock(InvokeBlock)); - } Function *Caller = FirstNewBlock->getParent(); // The inlined code is currently at the end of the function, scan from the // start of the inlined code to its end, checking for stuff we need to - // rewrite. - if (InlinedCodeInfo.ContainsCalls || InlinedCodeInfo.ContainsUnwinds) { - for (Function::iterator BB = FirstNewBlock, E = Caller->end(); - BB != E; ++BB) { - if (InlinedCodeInfo.ContainsCalls) { - for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ){ - Instruction *I = BBI++; - - // We only need to check for function calls: inlined invoke - // instructions require no special handling. - if (!isa(I)) continue; - CallInst *CI = cast(I); - - // If this call cannot unwind, don't convert it to an invoke. - if (CI->doesNotThrow()) - continue; - - // Convert this function call into an invoke instruction. - // First, split the basic block. - BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); - - // Next, create the new invoke instruction, inserting it at the end - // of the old basic block. - SmallVector InvokeArgs(CI->op_begin()+1, CI->op_end()); - InvokeInst *II = - InvokeInst::Create(CI->getCalledValue(), Split, InvokeDest, - InvokeArgs.begin(), InvokeArgs.end(), - CI->getName(), BB->getTerminator()); - II->setCallingConv(CI->getCallingConv()); - II->setAttributes(CI->getAttributes()); - - // Make sure that anything using the call now uses the invoke! - CI->replaceAllUsesWith(II); - - // Delete the unconditional branch inserted by splitBasicBlock - BB->getInstList().pop_back(); - Split->getInstList().pop_front(); // Delete the original call - - // Update any PHI nodes in the exceptional block to indicate that - // there is now a new entry in them. - unsigned i = 0; - for (BasicBlock::iterator I = InvokeDest->begin(); - isa(I); ++I, ++i) { - PHINode *PN = cast(I); - PN->addIncoming(InvokeDestPHIValues[i], BB); - } - - // This basic block is now complete, start scanning the next one. - break; - } - } + // rewrite. If the code doesn't have calls or unwinds, we know there is + // nothing to rewrite. + if (!InlinedCodeInfo.ContainsCalls && !InlinedCodeInfo.ContainsUnwinds) { + // Now that everything is happy, we have one final detail. The PHI nodes in + // the exception destination block still have entries due to the original + // invoke instruction. Eliminate these entries (which might even delete the + // PHI node) now. + InvokeDest->removePredecessor(II->getParent()); + return; + } - if (UnwindInst *UI = dyn_cast(BB->getTerminator())) { - // An UnwindInst requires special handling when it gets inlined into an - // invoke site. Once this happens, we know that the unwind would cause - // a control transfer to the invoke exception destination, so we can - // transform it into a direct branch to the exception destination. - BranchInst::Create(InvokeDest, UI); - - // Delete the unwind instruction! - UI->eraseFromParent(); - - // Update any PHI nodes in the exceptional block to indicate that - // there is now a new entry in them. - unsigned i = 0; - for (BasicBlock::iterator I = InvokeDest->begin(); - isa(I); ++I, ++i) { - PHINode *PN = cast(I); - PN->addIncoming(InvokeDestPHIValues[i], BB); - } + InvokeInliningInfo Invoke(II); + + for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ + if (InlinedCodeInfo.ContainsCalls) + if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) { + // Honor a request to skip the next block. We don't need to + // consider UnwindInsts in this case either. + ++BB; + continue; } + + if (UnwindInst *UI = dyn_cast(BB->getTerminator())) { + // An UnwindInst requires special handling when it gets inlined into an + // invoke site. Once this happens, we know that the unwind would cause + // a control transfer to the invoke exception destination, so we can + // transform it into a direct branch to the exception destination. + BranchInst::Create(InvokeDest, UI); + + // Delete the unwind instruction! + UI->eraseFromParent(); + + // Update any PHI nodes in the exceptional block to indicate that + // there is now a new entry in them. + Invoke.addIncomingPHIValuesFor(BB); } } @@ -146,8 +575,9 @@ static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, /// some edges of the callgraph may remain. static void UpdateCallGraphAfterInlining(CallSite CS, Function::iterator FirstNewBlock, - DenseMap &ValueMap, - CallGraph &CG) { + ValueToValueMapTy &VMap, + InlineFunctionInfo &IFI) { + CallGraph &CG = *IFI.CG; const Function *Caller = CS.getInstruction()->getParent()->getParent(); const Function *Callee = CS.getCalledFunction(); CallGraphNode *CalleeNode = CG[Callee]; @@ -166,22 +596,188 @@ static void UpdateCallGraphAfterInlining(CallSite CS, } for (; I != E; ++I) { - const Instruction *OrigCall = I->first.getInstruction(); + const Value *OrigCall = I->first; - DenseMap::iterator VMI = ValueMap.find(OrigCall); + ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); // Only copy the edge if the call was inlined! - if (VMI != ValueMap.end() && VMI->second) { - // If the call was inlined, but then constant folded, there is no edge to - // add. Check for this case. - if (Instruction *NewCall = dyn_cast(VMI->second)) - CallerNode->addCalledFunction(CallSite::get(NewCall), I->second); - } + if (VMI == VMap.end() || VMI->second == 0) + continue; + + // If the call was inlined, but then constant folded, there is no edge to + // add. Check for this case. + Instruction *NewCall = dyn_cast(VMI->second); + if (NewCall == 0) continue; + + // Remember that this call site got inlined for the client of + // InlineFunction. + IFI.InlinedCalls.push_back(NewCall); + + // It's possible that inlining the callsite will cause it to go from an + // indirect to a direct call by resolving a function pointer. If this + // happens, set the callee of the new call site to a more precise + // destination. This can also happen if the call graph node of the caller + // was just unnecessarily imprecise. + if (I->second->getFunction() == 0) + if (Function *F = CallSite(NewCall).getCalledFunction()) { + // Indirect call site resolved to direct call. + CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); + + continue; + } + + CallerNode->addCalledFunction(CallSite(NewCall), I->second); } + // Update the call graph by deleting the edge from Callee to Caller. We must // do this after the loop above in case Caller and Callee are the same. CallerNode->removeCallEdgeFor(CS); } +/// HandleByValArgument - When inlining a call site that has a byval argument, +/// we have to make the implicit memcpy explicit by adding it. +static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, + const Function *CalledFunc, + InlineFunctionInfo &IFI, + unsigned ByValAlignment) { + const Type *AggTy = cast(Arg->getType())->getElementType(); + + // If the called function is readonly, then it could not mutate the caller's + // copy of the byval'd memory. In this case, it is safe to elide the copy and + // temporary. + if (CalledFunc->onlyReadsMemory()) { + // If the byval argument has a specified alignment that is greater than the + // passed in pointer, then we either have to round up the input pointer or + // give up on this transformation. + if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. + return Arg; + + // If the pointer is already known to be sufficiently aligned, or if we can + // round it up to a larger alignment, then we don't need a temporary. + if (getOrEnforceKnownAlignment(Arg, ByValAlignment, + IFI.TD) >= ByValAlignment) + return Arg; + + // Otherwise, we have to make a memcpy to get a safe alignment. This is bad + // for code quality, but rarely happens and is required for correctness. + } + + LLVMContext &Context = Arg->getContext(); + + const Type *VoidPtrTy = Type::getInt8PtrTy(Context); + + // Create the alloca. If we have TargetData, use nice alignment. + unsigned Align = 1; + if (IFI.TD) + Align = IFI.TD->getPrefTypeAlignment(AggTy); + + // If the byval had an alignment specified, we *must* use at least that + // alignment, as it is required by the byval argument (and uses of the + // pointer inside the callee). + Align = std::max(Align, ByValAlignment); + + Function *Caller = TheCall->getParent()->getParent(); + + Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), + &*Caller->begin()->begin()); + // Emit a memcpy. + const Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; + Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), + Intrinsic::memcpy, + Tys, 3); + Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); + Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); + + Value *Size; + if (IFI.TD == 0) + Size = ConstantExpr::getSizeOf(AggTy); + else + Size = ConstantInt::get(Type::getInt64Ty(Context), + IFI.TD->getTypeStoreSize(AggTy)); + + // Always generate a memcpy of alignment 1 here because we don't know + // the alignment of the src pointer. Other optimizations can infer + // better alignment. + Value *CallArgs[] = { + DestCast, SrcCast, Size, + ConstantInt::get(Type::getInt32Ty(Context), 1), + ConstantInt::getFalse(Context) // isVolatile + }; + IRBuilder<>(TheCall).CreateCall(MemCpyFn, CallArgs, CallArgs+5); + + // Uses of the argument in the function should use our new alloca + // instead. + return NewAlloca; +} + +// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime +// intrinsic. +static bool isUsedByLifetimeMarker(Value *V) { + for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; + ++UI) { + if (IntrinsicInst *II = dyn_cast(*UI)) { + switch (II->getIntrinsicID()) { + default: break; + case Intrinsic::lifetime_start: + case Intrinsic::lifetime_end: + return true; + } + } + } + return false; +} + +// hasLifetimeMarkers - Check whether the given alloca already has +// lifetime.start or lifetime.end intrinsics. +static bool hasLifetimeMarkers(AllocaInst *AI) { + const Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext()); + if (AI->getType() == Int8PtrTy) + return isUsedByLifetimeMarker(AI); + + // Do a scan to find all the casts to i8*. + for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E; + ++I) { + if (I->getType() != Int8PtrTy) continue; + if (I->stripPointerCasts() != AI) continue; + if (isUsedByLifetimeMarker(*I)) + return true; + } + return false; +} + +/// updateInlinedAtInfo - Helper function used by fixupLineNumbers to recursively +/// update InlinedAtEntry of a DebugLoc. +static DebugLoc updateInlinedAtInfo(const DebugLoc &DL, + const DebugLoc &InlinedAtDL, + LLVMContext &Ctx) { + if (MDNode *IA = DL.getInlinedAt(Ctx)) { + DebugLoc NewInlinedAtDL + = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx); + return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), + NewInlinedAtDL.getAsMDNode(Ctx)); + } + + return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), + InlinedAtDL.getAsMDNode(Ctx)); +} + + +/// fixupLineNumbers - Update inlined instructions' line numbers to +/// to encode location where these instructions are inlined. +static void fixupLineNumbers(Function *Fn, Function::iterator FI, + Instruction *TheCall) { + DebugLoc TheCallDL = TheCall->getDebugLoc(); + if (TheCallDL.isUnknown()) + return; + + for (; FI != Fn->end(); ++FI) { + for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); + BI != BE; ++BI) { + DebugLoc DL = BI->getDebugLoc(); + if (!DL.isUnknown()) + BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext())); + } + } +} // InlineFunction - This function inlines the called function into the basic // block of the caller. This returns false if it is not possible to inline this @@ -189,24 +785,27 @@ static void UpdateCallGraphAfterInlining(CallSite CS, // // Note that this only does one level of inlining. For example, if the // instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now -// exists in the instruction stream. Similiarly this will inline a recursive +// exists in the instruction stream. Similarly this will inline a recursive // function by one level. // -bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { +bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI) { Instruction *TheCall = CS.getInstruction(); + LLVMContext &Context = TheCall->getContext(); assert(TheCall->getParent() && TheCall->getParent()->getParent() && "Instruction not in function!"); + // If IFI has any state in it, zap it before we fill it in. + IFI.reset(); + const Function *CalledFunc = CS.getCalledFunction(); if (CalledFunc == 0 || // Can't inline external function or indirect CalledFunc->isDeclaration() || // call, or call to a vararg function! CalledFunc->getFunctionType()->isVarArg()) return false; - - // If the call to the callee is a non-tail call, we must clear the 'tail' + // If the call to the callee is not a tail call, we must clear the 'tail' // flags on any calls that we inline. bool MustClearTailCallFlags = - isa(TheCall) && !cast(TheCall)->isTailCall(); + !(isa(TheCall) && cast(TheCall)->isTailCall()); // If the call to the callee cannot throw, set the 'nounwind' flag on any // calls that we inline. @@ -233,12 +832,12 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { // Make sure to capture all of the return instructions from the cloned // function. - std::vector Returns; + SmallVector Returns; ClonedCodeInfo InlinedFunctionInfo; Function::iterator FirstNewBlock; - { // Scope to destroy ValueMap after cloning. - DenseMap ValueMap; + { // Scope to destroy VMap after cloning. + ValueToValueMapTy VMap; assert(CalledFunc->arg_size() == CS.arg_size() && "No varargs calls can be inlined!"); @@ -255,67 +854,36 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { // by them explicit. However, we don't do this if the callee is readonly // or readnone, because the copy would be unneeded: the callee doesn't // modify the struct. - if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal) && - !CalledFunc->onlyReadsMemory()) { - const Type *AggTy = cast(I->getType())->getElementType(); - const Type *VoidPtrTy = PointerType::getUnqual(Type::Int8Ty); - - // Create the alloca. If we have TargetData, use nice alignment. - unsigned Align = 1; - if (TD) Align = TD->getPrefTypeAlignment(AggTy); - Value *NewAlloca = new AllocaInst(AggTy, 0, Align, I->getName(), - Caller->begin()->begin()); - // Emit a memcpy. - const Type *Tys[] = { Type::Int64Ty }; - Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), - Intrinsic::memcpy, - Tys, 1); - Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); - Value *SrcCast = new BitCastInst(*AI, VoidPtrTy, "tmp", TheCall); - - Value *Size; - if (TD == 0) - Size = ConstantExpr::getSizeOf(AggTy); - else - Size = ConstantInt::get(Type::Int64Ty, TD->getTypeStoreSize(AggTy)); - - // Always generate a memcpy of alignment 1 here because we don't know - // the alignment of the src pointer. Other optimizations can infer - // better alignment. - Value *CallArgs[] = { - DestCast, SrcCast, Size, ConstantInt::get(Type::Int32Ty, 1) - }; - CallInst *TheMemCpy = - CallInst::Create(MemCpyFn, CallArgs, CallArgs+4, "", TheCall); - - // If we have a call graph, update it. - if (CG) { - CallGraphNode *MemCpyCGN = CG->getOrInsertFunction(MemCpyFn); - CallGraphNode *CallerNode = (*CG)[Caller]; - CallerNode->addCalledFunction(TheMemCpy, MemCpyCGN); - } - - // Uses of the argument in the function should use our new alloca - // instead. - ActualArg = NewAlloca; + if (CalledFunc->paramHasAttr(ArgNo+1, Attribute::ByVal)) { + ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, + CalledFunc->getParamAlignment(ArgNo+1)); + + // Calls that we inline may use the new alloca, so we need to clear + // their 'tail' flags if HandleByValArgument introduced a new alloca and + // the callee has calls. + MustClearTailCallFlags |= ActualArg != *AI; } - ValueMap[I] = ActualArg; + VMap[I] = ActualArg; } // We want the inliner to prune the code as it copies. We would LOVE to // have no dead or constant instructions leftover after inlining occurs // (which can happen, e.g., because an argument was constant), but we'll be // happy with whatever the cloner can do. - CloneAndPruneFunctionInto(Caller, CalledFunc, ValueMap, Returns, ".i", - &InlinedFunctionInfo, TD); + CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, + /*ModuleLevelChanges=*/false, Returns, ".i", + &InlinedFunctionInfo, IFI.TD, TheCall); // Remember the first block that is newly cloned over. FirstNewBlock = LastBlock; ++FirstNewBlock; // Update the callgraph if requested. - if (CG) - UpdateCallGraphAfterInlining(CS, FirstNewBlock, ValueMap, *CG); + if (IFI.CG) + UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); + + // Update inlined instructions' line number information. + fixupLineNumbers(Caller, FirstNewBlock, TheCall); } // If there are any alloca instructions in the block that used to be the entry @@ -326,31 +894,58 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { { BasicBlock::iterator InsertPoint = Caller->begin()->begin(); for (BasicBlock::iterator I = FirstNewBlock->begin(), - E = FirstNewBlock->end(); I != E; ) - if (AllocaInst *AI = dyn_cast(I++)) { - // If the alloca is now dead, remove it. This often occurs due to code - // specialization. - if (AI->use_empty()) { - AI->eraseFromParent(); - continue; - } + E = FirstNewBlock->end(); I != E; ) { + AllocaInst *AI = dyn_cast(I++); + if (AI == 0) continue; + + // If the alloca is now dead, remove it. This often occurs due to code + // specialization. + if (AI->use_empty()) { + AI->eraseFromParent(); + continue; + } - if (isa(AI->getArraySize())) { - // Scan for the block of allocas that we can move over, and move them - // all at once. - while (isa(I) && - isa(cast(I)->getArraySize())) - ++I; - - // Transfer all of the allocas over in a block. Using splice means - // that the instructions aren't removed from the symbol table, then - // reinserted. - Caller->getEntryBlock().getInstList().splice( - InsertPoint, - FirstNewBlock->getInstList(), - AI, I); - } + if (!isa(AI->getArraySize())) + continue; + + // Keep track of the static allocas that we inline into the caller. + IFI.StaticAllocas.push_back(AI); + + // Scan for the block of allocas that we can move over, and move them + // all at once. + while (isa(I) && + isa(cast(I)->getArraySize())) { + IFI.StaticAllocas.push_back(cast(I)); + ++I; } + + // Transfer all of the allocas over in a block. Using splice means + // that the instructions aren't removed from the symbol table, then + // reinserted. + Caller->getEntryBlock().getInstList().splice(InsertPoint, + FirstNewBlock->getInstList(), + AI, I); + } + } + + // Leave lifetime markers for the static alloca's, scoping them to the + // function we just inlined. + if (!IFI.StaticAllocas.empty()) { + IRBuilder<> builder(FirstNewBlock->begin()); + for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { + AllocaInst *AI = IFI.StaticAllocas[ai]; + + // If the alloca is already scoped to something smaller than the whole + // function then there's no need to add redundant, less accurate markers. + if (hasLifetimeMarkers(AI)) + continue; + + builder.CreateLifetimeStart(AI); + for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) { + IRBuilder<> builder(Returns[ri]); + builder.CreateLifetimeEnd(AI); + } + } } // If the inlined code contained dynamic alloca instructions, wrap the inlined @@ -358,31 +953,17 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { if (InlinedFunctionInfo.ContainsDynamicAllocas) { Module *M = Caller->getParent(); // Get the two intrinsics we care about. - Constant *StackSave, *StackRestore; - StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); - StackRestore = Intrinsic::getDeclaration(M, Intrinsic::stackrestore); - - // If we are preserving the callgraph, add edges to the stacksave/restore - // functions for the calls we insert. - CallGraphNode *StackSaveCGN = 0, *StackRestoreCGN = 0, *CallerNode = 0; - if (CG) { - // We know that StackSave/StackRestore are Function*'s, because they are - // intrinsics which must have the right types. - StackSaveCGN = CG->getOrInsertFunction(cast(StackSave)); - StackRestoreCGN = CG->getOrInsertFunction(cast(StackRestore)); - CallerNode = (*CG)[Caller]; - } + Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); + Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); // Insert the llvm.stacksave. - CallInst *SavedPtr = CallInst::Create(StackSave, "savedstack", - FirstNewBlock->begin()); - if (CG) CallerNode->addCalledFunction(SavedPtr, StackSaveCGN); + CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin()) + .CreateCall(StackSave, "savedstack"); // Insert a call to llvm.stackrestore before any return instructions in the // inlined function. for (unsigned i = 0, e = Returns.size(); i != e; ++i) { - CallInst *CI = CallInst::Create(StackRestore, SavedPtr, "", Returns[i]); - if (CG) CallerNode->addCalledFunction(CI, StackRestoreCGN); + IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr); } // Count the number of StackRestore calls we insert. @@ -394,7 +975,7 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB) if (UnwindInst *UI = dyn_cast(BB->getTerminator())) { - CallInst::Create(StackRestore, SavedPtr, "", UI); + IRBuilder<>(UI).CreateCall(StackRestore, SavedPtr); ++NumStackRestores; } } @@ -424,7 +1005,7 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { BB != E; ++BB) { TerminatorInst *Term = BB->getTerminator(); if (isa(Term)) { - new UnreachableInst(Term); + new UnreachableInst(Context, Term); BB->getInstList().erase(Term); } } @@ -454,7 +1035,10 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { // uses of the returned value. if (!TheCall->use_empty()) { ReturnInst *R = Returns[0]; - TheCall->replaceAllUsesWith(R->getReturnValue()); + if (TheCall == R->getReturnValue()) + TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); + else + TheCall->replaceAllUsesWith(R->getReturnValue()); } // Since we are now done with the Call/Invoke, we can delete it. TheCall->eraseFromParent(); @@ -511,12 +1095,12 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { // any users of the original call/invoke instruction. const Type *RTy = CalledFunc->getReturnType(); + PHINode *PHI = 0; if (Returns.size() > 1) { // The PHI node should go at the front of the new basic block to merge all // possible incoming values. - PHINode *PHI = 0; if (!TheCall->use_empty()) { - PHI = PHINode::Create(RTy, TheCall->getName(), + PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), AfterCallBB->begin()); // Anything that used the result of the function call should now use the // PHI node as their operand. @@ -534,6 +1118,7 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { } } + // Add a branch to the merge points and remove return instructions. for (unsigned i = 0, e = Returns.size(); i != e; ++i) { ReturnInst *RI = Returns[i]; @@ -543,18 +1128,22 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { } else if (!Returns.empty()) { // Otherwise, if there is exactly one return value, just replace anything // using the return value of the call with the computed value. - if (!TheCall->use_empty()) - TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); + if (!TheCall->use_empty()) { + if (TheCall == Returns[0]->getReturnValue()) + TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); + else + TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); + } + + // Update PHI nodes that use the ReturnBB to use the AfterCallBB. + BasicBlock *ReturnBB = Returns[0]->getParent(); + ReturnBB->replaceAllUsesWith(AfterCallBB); // Splice the code from the return block into the block that it will return // to, which contains the code that was after the call. - BasicBlock *ReturnBB = Returns[0]->getParent(); AfterCallBB->getInstList().splice(AfterCallBB->begin(), ReturnBB->getInstList()); - // Update PHI nodes that use the ReturnBB to use the AfterCallBB. - ReturnBB->replaceAllUsesWith(AfterCallBB); - // Delete the return instruction now and empty ReturnBB now. Returns[0]->eraseFromParent(); ReturnBB->eraseFromParent(); @@ -574,8 +1163,8 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { // Splice the code entry block into calling block, right before the // unconditional branch. - OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes + OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); // Remove the unconditional branch. OrigBB->getInstList().erase(Br); @@ -583,5 +1172,14 @@ bool llvm::InlineFunction(CallSite CS, CallGraph *CG, const TargetData *TD) { // Now we can remove the CalleeEntry block, which is now empty. Caller->getBasicBlockList().erase(CalleeEntry); + // If we inserted a phi node, check to see if it has a single value (e.g. all + // the entries are the same or undef). If so, remove the PHI so it doesn't + // block other optimizations. + if (PHI) + if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { + PHI->replaceAllUsesWith(V); + PHI->eraseFromParent(); + } + return true; }