X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTransforms%2FIPO%2FArgumentPromotion.cpp;h=94b44654d92279cb7b7dc911ec0a52d4c753f26c;hp=d66e2c4620230577df8987a021a463f2d12063dc;hb=7962dbdc6531cb44003dc53323e18c8ee9a20e19;hpb=19c874638d9478a5d5028854817a5ee72293bb2b diff --git a/lib/Transforms/IPO/ArgumentPromotion.cpp b/lib/Transforms/IPO/ArgumentPromotion.cpp index d66e2c46202..94b44654d92 100644 --- a/lib/Transforms/IPO/ArgumentPromotion.cpp +++ b/lib/Transforms/IPO/ArgumentPromotion.cpp @@ -29,26 +29,27 @@ // //===----------------------------------------------------------------------===// -#define DEBUG_TYPE "argpromotion" #include "llvm/Transforms/IPO.h" -#include "llvm/Constants.h" -#include "llvm/DerivedTypes.h" -#include "llvm/Module.h" -#include "llvm/CallGraphSCCPass.h" -#include "llvm/Instructions.h" -#include "llvm/Analysis/AliasAnalysis.h" -#include "llvm/Analysis/CallGraph.h" -#include "llvm/Target/TargetData.h" -#include "llvm/Support/CallSite.h" -#include "llvm/Support/CFG.h" -#include "llvm/Support/Debug.h" #include "llvm/ADT/DepthFirstIterator.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/StringExtras.h" -#include "llvm/Support/Compiler.h" +#include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/Analysis/CallGraph.h" +#include "llvm/Analysis/CallGraphSCCPass.h" +#include "llvm/IR/CFG.h" +#include "llvm/IR/CallSite.h" +#include "llvm/IR/Constants.h" +#include "llvm/IR/DerivedTypes.h" +#include "llvm/IR/Instructions.h" +#include "llvm/IR/LLVMContext.h" +#include "llvm/IR/Module.h" +#include "llvm/Support/Debug.h" +#include "llvm/Support/raw_ostream.h" #include using namespace llvm; +#define DEBUG_TYPE "argpromotion" + STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted"); STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted"); STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted"); @@ -57,51 +58,60 @@ STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated"); namespace { /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass. /// - struct VISIBILITY_HIDDEN ArgPromotion : public CallGraphSCCPass { - virtual void getAnalysisUsage(AnalysisUsage &AU) const { + struct ArgPromotion : public CallGraphSCCPass { + void getAnalysisUsage(AnalysisUsage &AU) const override { AU.addRequired(); - AU.addRequired(); CallGraphSCCPass::getAnalysisUsage(AU); } - virtual bool runOnSCC(const std::vector &SCC); + bool runOnSCC(CallGraphSCC &SCC) override; static char ID; // Pass identification, replacement for typeid - ArgPromotion(unsigned maxElements = 3) : CallGraphSCCPass(&ID), - maxElements(maxElements) {} + explicit ArgPromotion(unsigned maxElements = 3) + : CallGraphSCCPass(ID), maxElements(maxElements) { + initializeArgPromotionPass(*PassRegistry::getPassRegistry()); + } /// A vector used to hold the indices of a single GEP instruction typedef std::vector IndicesVector; private: - bool PromoteArguments(CallGraphNode *CGN); + CallGraphNode *PromoteArguments(CallGraphNode *CGN); bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const; - Function *DoPromotion(Function *F, - SmallPtrSet &ArgsToPromote, - SmallPtrSet &ByValArgsToTransform); + CallGraphNode *DoPromotion(Function *F, + SmallPtrSet &ArgsToPromote, + SmallPtrSet &ByValArgsToTransform); /// The maximum number of elements to expand, or 0 for unlimited. unsigned maxElements; }; } char ArgPromotion::ID = 0; -static RegisterPass -X("argpromotion", "Promote 'by reference' arguments to scalars"); +INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion", + "Promote 'by reference' arguments to scalars", false, false) +INITIALIZE_AG_DEPENDENCY(AliasAnalysis) +INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass) +INITIALIZE_PASS_END(ArgPromotion, "argpromotion", + "Promote 'by reference' arguments to scalars", false, false) Pass *llvm::createArgumentPromotionPass(unsigned maxElements) { return new ArgPromotion(maxElements); } -bool ArgPromotion::runOnSCC(const std::vector &SCC) { +bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) { bool Changed = false, LocalChange; do { // Iterate until we stop promoting from this SCC. LocalChange = false; // Attempt to promote arguments from all functions in this SCC. - for (unsigned i = 0, e = SCC.size(); i != e; ++i) - LocalChange |= PromoteArguments(SCC[i]); + for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { + if (CallGraphNode *CGN = PromoteArguments(*I)) { + LocalChange = true; + SCC.ReplaceNode(*I, CGN); + } + } Changed |= LocalChange; // Remember that we changed something. } while (LocalChange); - + return Changed; } @@ -110,116 +120,113 @@ bool ArgPromotion::runOnSCC(const std::vector &SCC) { /// example, all callers are direct). If safe to promote some arguments, it /// calls the DoPromotion method. /// -bool ArgPromotion::PromoteArguments(CallGraphNode *CGN) { +CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) { Function *F = CGN->getFunction(); // Make sure that it is local to this module. - if (!F || !F->hasInternalLinkage()) return false; + if (!F || !F->hasLocalLinkage()) return 0; // First check: see if there are any pointer arguments! If not, quick exit. - SmallVector, 16> PointerArgs; - unsigned ArgNo = 0; - for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); - I != E; ++I, ++ArgNo) - if (isa(I->getType())) - PointerArgs.push_back(std::pair(I, ArgNo)); - if (PointerArgs.empty()) return false; + SmallVector PointerArgs; + for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I) + if (I->getType()->isPointerTy()) + PointerArgs.push_back(I); + if (PointerArgs.empty()) return 0; // Second check: make sure that all callers are direct callers. We can't - // transform functions that have indirect callers. - for (Value::use_iterator UI = F->use_begin(), E = F->use_end(); - UI != E; ++UI) { - CallSite CS = CallSite::get(*UI); - if (!CS.getInstruction()) // "Taking the address" of the function - return false; - - // Ensure that this call site is CALLING the function, not passing it as - // an argument. - if (UI.getOperandNo() != 0) - return false; + // transform functions that have indirect callers. Also see if the function + // is self-recursive. + bool isSelfRecursive = false; + for (Use &U : F->uses()) { + CallSite CS(U.getUser()); + // Must be a direct call. + if (CS.getInstruction() == 0 || !CS.isCallee(&U)) return 0; + + if (CS.getInstruction()->getParent()->getParent() == F) + isSelfRecursive = true; } - + // Check to see which arguments are promotable. If an argument is promotable, // add it to ArgsToPromote. SmallPtrSet ArgsToPromote; SmallPtrSet ByValArgsToTransform; - for (unsigned i = 0; i != PointerArgs.size(); ++i) { - bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal); + for (unsigned i = 0, e = PointerArgs.size(); i != e; ++i) { + Argument *PtrArg = PointerArgs[i]; + Type *AgTy = cast(PtrArg->getType())->getElementType(); // If this is a byval argument, and if the aggregate type is small, just - // pass the elements, which is always safe. - Argument *PtrArg = PointerArgs[i].first; - if (isByVal) { - const Type *AgTy = cast(PtrArg->getType())->getElementType(); - if (const StructType *STy = dyn_cast(AgTy)) { + // pass the elements, which is always safe. This does not apply to + // inalloca. + if (PtrArg->hasByValAttr()) { + if (StructType *STy = dyn_cast(AgTy)) { if (maxElements > 0 && STy->getNumElements() > maxElements) { - DOUT << "argpromotion disable promoting argument '" - << PtrArg->getName() << "' because it would require adding more " - << "than " << maxElements << " arguments to the function.\n"; - } else { - // If all the elements are single-value types, we can promote it. - bool AllSimple = true; - for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) - if (!STy->getElementType(i)->isSingleValueType()) { - AllSimple = false; - break; - } - - // Safe to transform, don't even bother trying to "promote" it. - // Passing the elements as a scalar will allow scalarrepl to hack on - // the new alloca we introduce. - if (AllSimple) { - ByValArgsToTransform.insert(PtrArg); - continue; + DEBUG(dbgs() << "argpromotion disable promoting argument '" + << PtrArg->getName() << "' because it would require adding more" + << " than " << maxElements << " arguments to the function.\n"); + continue; + } + + // If all the elements are single-value types, we can promote it. + bool AllSimple = true; + for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { + if (!STy->getElementType(i)->isSingleValueType()) { + AllSimple = false; + break; } } + + // Safe to transform, don't even bother trying to "promote" it. + // Passing the elements as a scalar will allow scalarrepl to hack on + // the new alloca we introduce. + if (AllSimple) { + ByValArgsToTransform.insert(PtrArg); + continue; + } } } + // If the argument is a recursive type and we're in a recursive + // function, we could end up infinitely peeling the function argument. + if (isSelfRecursive) { + if (StructType *STy = dyn_cast(AgTy)) { + bool RecursiveType = false; + for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { + if (STy->getElementType(i) == PtrArg->getType()) { + RecursiveType = true; + break; + } + } + if (RecursiveType) + continue; + } + } + // Otherwise, see if we can promote the pointer to its value. - if (isSafeToPromoteArgument(PtrArg, isByVal)) + if (isSafeToPromoteArgument(PtrArg, PtrArg->hasByValOrInAllocaAttr())) ArgsToPromote.insert(PtrArg); } // No promotable pointer arguments. - if (ArgsToPromote.empty() && ByValArgsToTransform.empty()) return false; - - Function *NewF = DoPromotion(F, ArgsToPromote, ByValArgsToTransform); - - // Update the call graph to know that the function has been transformed. - getAnalysis().changeFunction(F, NewF); - return true; -} + if (ArgsToPromote.empty() && ByValArgsToTransform.empty()) + return 0; -/// IsAlwaysValidPointer - Return true if the specified pointer is always legal -/// to load. -static bool IsAlwaysValidPointer(Value *V) { - if (isa(V) || isa(V)) return true; - if (GetElementPtrInst *GEP = dyn_cast(V)) - return IsAlwaysValidPointer(GEP->getOperand(0)); - if (ConstantExpr *CE = dyn_cast(V)) - if (CE->getOpcode() == Instruction::GetElementPtr) - return IsAlwaysValidPointer(CE->getOperand(0)); - - return false; + return DoPromotion(F, ArgsToPromote, ByValArgsToTransform); } -/// AllCalleesPassInValidPointerForArgument - Return true if we can prove that +/// AllCallersPassInValidPointerForArgument - Return true if we can prove that /// all callees pass in a valid pointer for the specified function argument. -static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) { +static bool AllCallersPassInValidPointerForArgument(Argument *Arg) { Function *Callee = Arg->getParent(); - unsigned ArgNo = std::distance(Callee->arg_begin(), - Function::arg_iterator(Arg)); + unsigned ArgNo = Arg->getArgNo(); // Look at all call sites of the function. At this pointer we know we only // have direct callees. - for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end(); - UI != E; ++UI) { - CallSite CS = CallSite::get(*UI); - assert(CS.getInstruction() && "Should only have direct calls!"); + for (User *U : Callee->users()) { + CallSite CS(U); + assert(CS && "Should only have direct calls!"); - if (!IsAlwaysValidPointer(CS.getArgument(ArgNo))) + if (!CS.getArgument(ArgNo)->isDereferenceablePointer()) return false; } return true; @@ -234,10 +241,7 @@ static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix, const ArgPromotion::IndicesVector &Longer) { if (Prefix.size() > Longer.size()) return false; - for (unsigned i = 0, e = Prefix.size(); i != e; ++i) - if (Prefix[i] != Longer[i]) - return false; - return true; + return std::equal(Prefix.begin(), Prefix.end(), Longer.begin()); } @@ -256,7 +260,7 @@ static bool PrefixIn(const ArgPromotion::IndicesVector &Indices, return Low != Set.end() && IsPrefix(*Low, Indices); } -/// Mark the given indices (ToMark) as safe in the the given set of indices +/// Mark the given indices (ToMark) as safe in the given set of indices /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there /// is already a prefix of Indices in Safe, Indices are implicitely marked safe /// already. Furthermore, any indices that Indices is itself a prefix of, are @@ -297,7 +301,8 @@ static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark, /// This method limits promotion of aggregates to only promote up to three /// elements of the aggregate in order to avoid exploding the number of /// arguments passed in. -bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { +bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, + bool isByValOrInAlloca) const { typedef std::set GEPIndicesSet; // Quick exit for unused arguments @@ -319,6 +324,9 @@ bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { // // This set will contain all sets of indices that are loaded in the entry // block, and thus are safe to unconditionally load in the caller. + // + // This optimization is also safe for InAlloca parameters, because it verifies + // that the address isn't captured. GEPIndicesSet SafeToUnconditionallyLoad; // This set contains all the sets of indices that we are planning to promote. @@ -326,7 +334,7 @@ bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { GEPIndicesSet ToPromote; // If the pointer is always valid, any load with first index 0 is valid. - if(isByVal || AllCalleesPassInValidPointerForArgument(Arg)) + if (isByValOrInAlloca || AllCallersPassInValidPointerForArgument(Arg)) SafeToUnconditionallyLoad.insert(IndicesVector(1, 0)); // First, iterate the entry block and mark loads of (geps of) arguments as @@ -366,24 +374,25 @@ bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { // not (GEP+)loads, or any (GEP+)loads that are not safe to promote. SmallVector Loads; IndicesVector Operands; - for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end(); - UI != E; ++UI) { + for (Use &U : Arg->uses()) { + User *UR = U.getUser(); Operands.clear(); - if (LoadInst *LI = dyn_cast(*UI)) { - if (LI->isVolatile()) return false; // Don't hack volatile loads + if (LoadInst *LI = dyn_cast(UR)) { + // Don't hack volatile/atomic loads + if (!LI->isSimple()) return false; Loads.push_back(LI); // Direct loads are equivalent to a GEP with a zero index and then a load. Operands.push_back(0); - } else if (GetElementPtrInst *GEP = dyn_cast(*UI)) { + } else if (GetElementPtrInst *GEP = dyn_cast(UR)) { if (GEP->use_empty()) { // Dead GEP's cause trouble later. Just remove them if we run into // them. getAnalysis().deleteValue(GEP); GEP->eraseFromParent(); - // TODO: This runs the above loop over and over again for dead GEPS + // TODO: This runs the above loop over and over again for dead GEPs // Couldn't we just do increment the UI iterator earlier and erase the // use? - return isSafeToPromoteArgument(Arg, isByVal); + return isSafeToPromoteArgument(Arg, isByValOrInAlloca); } // Ensure that all of the indices are constants. @@ -395,10 +404,10 @@ bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { return false; // Not a constant operand GEP! // Ensure that the only users of the GEP are load instructions. - for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end(); - UI != E; ++UI) - if (LoadInst *LI = dyn_cast(*UI)) { - if (LI->isVolatile()) return false; // Don't hack volatile loads + for (User *GEPU : GEP->users()) + if (LoadInst *LI = dyn_cast(GEPU)) { + // Don't hack volatile/atomic loads + if (!LI->isSimple()) return false; Loads.push_back(LI); } else { // Other uses than load? @@ -418,9 +427,9 @@ bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { // to do. if (ToPromote.find(Operands) == ToPromote.end()) { if (maxElements > 0 && ToPromote.size() == maxElements) { - DOUT << "argpromotion not promoting argument '" - << Arg->getName() << "' because it would require adding more " - << "than " << maxElements << " arguments to the function.\n"; + DEBUG(dbgs() << "argpromotion not promoting argument '" + << Arg->getName() << "' because it would require adding more " + << "than " << maxElements << " arguments to the function.\n"); // We limit aggregate promotion to only promoting up to a fixed number // of elements of the aggregate. return false; @@ -441,7 +450,6 @@ bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { SmallPtrSet TranspBlocks; AliasAnalysis &AA = getAnalysis(); - TargetData &TD = getAnalysis(); for (unsigned i = 0, e = Loads.size(); i != e; ++i) { // Check to see if the load is invalidated from the start of the block to @@ -449,22 +457,21 @@ bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { LoadInst *Load = Loads[i]; BasicBlock *BB = Load->getParent(); - const PointerType *LoadTy = - cast(Load->getPointerOperand()->getType()); - unsigned LoadSize = (unsigned)TD.getTypeStoreSize(LoadTy->getElementType()); - - if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize)) + AliasAnalysis::Location Loc = AA.getLocation(Load); + if (AA.canInstructionRangeModify(BB->front(), *Load, Loc)) return false; // Pointer is invalidated! // Now check every path from the entry block to the load for transparency. // To do this, we perform a depth first search on the inverse CFG from the // loading block. - for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) + for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { + BasicBlock *P = *PI; for (idf_ext_iterator > - I = idf_ext_begin(*PI, TranspBlocks), - E = idf_ext_end(*PI, TranspBlocks); I != E; ++I) - if (AA.canBasicBlockModify(**I, Arg, LoadSize)) + I = idf_ext_begin(P, TranspBlocks), + E = idf_ext_end(P, TranspBlocks); I != E; ++I) + if (AA.canBasicBlockModify(**I, Loc)) return false; + } } // If the path from the entry of the function to each load is free of @@ -476,14 +483,14 @@ bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const { /// DoPromotion - This method actually performs the promotion of the specified /// arguments, and returns the new function. At this point, we know that it's /// safe to do so. -Function *ArgPromotion::DoPromotion(Function *F, - SmallPtrSet &ArgsToPromote, +CallGraphNode *ArgPromotion::DoPromotion(Function *F, + SmallPtrSet &ArgsToPromote, SmallPtrSet &ByValArgsToTransform) { // Start by computing a new prototype for the function, which is the same as // the old function, but has modified arguments. - const FunctionType *FTy = F->getFunctionType(); - std::vector Params; + FunctionType *FTy = F->getFunctionType(); + std::vector Params; typedef std::set ScalarizeTable; @@ -499,17 +506,20 @@ Function *ArgPromotion::DoPromotion(Function *F, // OriginalLoads - Keep track of a representative load instruction from the // original function so that we can tell the alias analysis implementation // what the new GEP/Load instructions we are inserting look like. - std::map OriginalLoads; + // We need to keep the original loads for each argument and the elements + // of the argument that are accessed. + std::map, LoadInst*> OriginalLoads; - // Attributes - Keep track of the parameter attributes for the arguments + // Attribute - Keep track of the parameter attributes for the arguments // that we are *not* promoting. For the ones that we do promote, the parameter // attributes are lost - SmallVector AttributesVec; - const AttrListPtr &PAL = F->getAttributes(); + SmallVector AttributesVec; + const AttributeSet &PAL = F->getAttributes(); // Add any return attributes. - if (Attributes attrs = PAL.getRetAttributes()) - AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); + if (PAL.hasAttributes(AttributeSet::ReturnIndex)) + AttributesVec.push_back(AttributeSet::get(F->getContext(), + PAL.getRetAttributes())); // First, determine the new argument list unsigned ArgIndex = 1; @@ -517,16 +527,20 @@ Function *ArgPromotion::DoPromotion(Function *F, ++I, ++ArgIndex) { if (ByValArgsToTransform.count(I)) { // Simple byval argument? Just add all the struct element types. - const Type *AgTy = cast(I->getType())->getElementType(); - const StructType *STy = cast(AgTy); + Type *AgTy = cast(I->getType())->getElementType(); + StructType *STy = cast(AgTy); for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) Params.push_back(STy->getElementType(i)); ++NumByValArgsPromoted; } else if (!ArgsToPromote.count(I)) { // Unchanged argument Params.push_back(I->getType()); - if (Attributes attrs = PAL.getParamAttributes(ArgIndex)) - AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs)); + AttributeSet attrs = PAL.getParamAttributes(ArgIndex); + if (attrs.hasAttributes(ArgIndex)) { + AttrBuilder B(attrs, ArgIndex); + AttributesVec. + push_back(AttributeSet::get(F->getContext(), Params.size(), B)); + } } else if (I->use_empty()) { // Dead argument (which are always marked as promotable) ++NumArgumentsDead; @@ -537,16 +551,15 @@ Function *ArgPromotion::DoPromotion(Function *F, // In this table, we will track which indices are loaded from the argument // (where direct loads are tracked as no indices). ScalarizeTable &ArgIndices = ScalarizedElements[I]; - for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; - ++UI) { - Instruction *User = cast(*UI); - assert(isa(User) || isa(User)); + for (User *U : I->users()) { + Instruction *UI = cast(U); + assert(isa(UI) || isa(UI)); IndicesVector Indices; - Indices.reserve(User->getNumOperands() - 1); + Indices.reserve(UI->getNumOperands() - 1); // Since loads will only have a single operand, and GEPs only a single // non-index operand, this will record direct loads without any indices, // and gep+loads with the GEP indices. - for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end(); + for (User::op_iterator II = UI->op_begin() + 1, IE = UI->op_end(); II != IE; ++II) Indices.push_back(cast(*II)->getSExtValue()); // GEPs with a single 0 index can be merged with direct loads @@ -554,20 +567,19 @@ Function *ArgPromotion::DoPromotion(Function *F, Indices.clear(); ArgIndices.insert(Indices); LoadInst *OrigLoad; - if (LoadInst *L = dyn_cast(User)) + if (LoadInst *L = dyn_cast(UI)) OrigLoad = L; else // Take any load, we will use it only to update Alias Analysis - OrigLoad = cast(User->use_back()); - OriginalLoads[Indices] = OrigLoad; + OrigLoad = cast(UI->user_back()); + OriginalLoads[std::make_pair(I, Indices)] = OrigLoad; } // Add a parameter to the function for each element passed in. for (ScalarizeTable::iterator SI = ArgIndices.begin(), E = ArgIndices.end(); SI != E; ++SI) { - Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), - &*SI->begin(), - SI->size())); + // not allowed to dereference ->begin() if size() is 0 + Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), *SI)); assert(Params.back()); } @@ -579,29 +591,26 @@ Function *ArgPromotion::DoPromotion(Function *F, } // Add any function attributes. - if (Attributes attrs = PAL.getFnAttributes()) - AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); + if (PAL.hasAttributes(AttributeSet::FunctionIndex)) + AttributesVec.push_back(AttributeSet::get(FTy->getContext(), + PAL.getFnAttributes())); - const Type *RetTy = FTy->getReturnType(); - - // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which - // have zero fixed arguments. - bool ExtraArgHack = false; - if (Params.empty() && FTy->isVarArg()) { - ExtraArgHack = true; - Params.push_back(Type::Int32Ty); - } + Type *RetTy = FTy->getReturnType(); // Construct the new function type using the new arguments. FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg()); - // Create the new function body and insert it into the module... + // Create the new function body and insert it into the module. Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName()); NF->copyAttributesFrom(F); + + DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n" + << "From: " << *F); + // Recompute the parameter attributes list based on the new arguments for // the function. - NF->setAttributes(AttrListPtr::get(AttributesVec.begin(), AttributesVec.end())); + NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec)); AttributesVec.clear(); F->getParent()->getFunctionList().insert(F, NF); @@ -613,20 +622,25 @@ Function *ArgPromotion::DoPromotion(Function *F, // Get the callgraph information that we need to update to reflect our // changes. - CallGraph &CG = getAnalysis(); + CallGraph &CG = getAnalysis().getCallGraph(); + + // Get a new callgraph node for NF. + CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF); // Loop over all of the callers of the function, transforming the call sites // to pass in the loaded pointers. // SmallVector Args; while (!F->use_empty()) { - CallSite CS = CallSite::get(F->use_back()); + CallSite CS(F->user_back()); + assert(CS.getCalledFunction() == F); Instruction *Call = CS.getInstruction(); - const AttrListPtr &CallPAL = CS.getAttributes(); + const AttributeSet &CallPAL = CS.getAttributes(); // Add any return attributes. - if (Attributes attrs = CallPAL.getRetAttributes()) - AttributesVec.push_back(AttributeWithIndex::get(0, attrs)); + if (CallPAL.hasAttributes(AttributeSet::ReturnIndex)) + AttributesVec.push_back(AttributeSet::get(F->getContext(), + CallPAL.getRetAttributes())); // Loop over the operands, inserting GEP and loads in the caller as // appropriate. @@ -637,17 +651,20 @@ Function *ArgPromotion::DoPromotion(Function *F, if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) { Args.push_back(*AI); // Unmodified argument - if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) - AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); - + if (CallPAL.hasAttributes(ArgIndex)) { + AttrBuilder B(CallPAL, ArgIndex); + AttributesVec. + push_back(AttributeSet::get(F->getContext(), Args.size(), B)); + } } else if (ByValArgsToTransform.count(I)) { // Emit a GEP and load for each element of the struct. - const Type *AgTy = cast(I->getType())->getElementType(); - const StructType *STy = cast(AgTy); - Value *Idxs[2] = { ConstantInt::get(Type::Int32Ty, 0), 0 }; + Type *AgTy = cast(I->getType())->getElementType(); + StructType *STy = cast(AgTy); + Value *Idxs[2] = { + ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { - Idxs[1] = ConstantInt::get(Type::Int32Ty, i); - Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2, + Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); + Value *Idx = GetElementPtrInst::Create(*AI, Idxs, (*AI)->getName()+"."+utostr(i), Call); // TODO: Tell AA about the new values? @@ -657,61 +674,70 @@ Function *ArgPromotion::DoPromotion(Function *F, // Non-dead argument: insert GEPs and loads as appropriate. ScalarizeTable &ArgIndices = ScalarizedElements[I]; // Store the Value* version of the indices in here, but declare it now - // for reuse + // for reuse. std::vector Ops; for (ScalarizeTable::iterator SI = ArgIndices.begin(), E = ArgIndices.end(); SI != E; ++SI) { Value *V = *AI; - LoadInst *OrigLoad = OriginalLoads[*SI]; + LoadInst *OrigLoad = OriginalLoads[std::make_pair(I, *SI)]; if (!SI->empty()) { Ops.reserve(SI->size()); - const Type *ElTy = V->getType(); + Type *ElTy = V->getType(); for (IndicesVector::const_iterator II = SI->begin(), IE = SI->end(); II != IE; ++II) { // Use i32 to index structs, and i64 for others (pointers/arrays). // This satisfies GEP constraints. - const Type *IdxTy = (isa(ElTy) ? Type::Int32Ty : Type::Int64Ty); + Type *IdxTy = (ElTy->isStructTy() ? + Type::getInt32Ty(F->getContext()) : + Type::getInt64Ty(F->getContext())); Ops.push_back(ConstantInt::get(IdxTy, *II)); - // Keep track of the type we're currently indexing + // Keep track of the type we're currently indexing. ElTy = cast(ElTy)->getTypeAtIndex(*II); } - // And create a GEP to extract those indices - V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(), - V->getName()+".idx", Call); + // And create a GEP to extract those indices. + V = GetElementPtrInst::Create(V, Ops, V->getName()+".idx", Call); Ops.clear(); AA.copyValue(OrigLoad->getOperand(0), V); } - Args.push_back(new LoadInst(V, V->getName()+".val", Call)); + // Since we're replacing a load make sure we take the alignment + // of the previous load. + LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call); + newLoad->setAlignment(OrigLoad->getAlignment()); + // Transfer the TBAA info too. + newLoad->setMetadata(LLVMContext::MD_tbaa, + OrigLoad->getMetadata(LLVMContext::MD_tbaa)); + Args.push_back(newLoad); AA.copyValue(OrigLoad, Args.back()); } } - if (ExtraArgHack) - Args.push_back(Constant::getNullValue(Type::Int32Ty)); - - // Push any varargs arguments on the list + // Push any varargs arguments on the list. for (; AI != CS.arg_end(); ++AI, ++ArgIndex) { Args.push_back(*AI); - if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex)) - AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs)); + if (CallPAL.hasAttributes(ArgIndex)) { + AttrBuilder B(CallPAL, ArgIndex); + AttributesVec. + push_back(AttributeSet::get(F->getContext(), Args.size(), B)); + } } // Add any function attributes. - if (Attributes attrs = CallPAL.getFnAttributes()) - AttributesVec.push_back(AttributeWithIndex::get(~0, attrs)); + if (CallPAL.hasAttributes(AttributeSet::FunctionIndex)) + AttributesVec.push_back(AttributeSet::get(Call->getContext(), + CallPAL.getFnAttributes())); Instruction *New; if (InvokeInst *II = dyn_cast(Call)) { New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), - Args.begin(), Args.end(), "", Call); + Args, "", Call); cast(New)->setCallingConv(CS.getCallingConv()); - cast(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), - AttributesVec.end())); + cast(New)->setAttributes(AttributeSet::get(II->getContext(), + AttributesVec)); } else { - New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); + New = CallInst::Create(NF, Args, "", Call); cast(New)->setCallingConv(CS.getCallingConv()); - cast(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(), - AttributesVec.end())); + cast(New)->setAttributes(AttributeSet::get(New->getContext(), + AttributesVec)); if (cast(Call)->isTailCall()) cast(New)->setTailCall(); } @@ -723,7 +749,8 @@ Function *ArgPromotion::DoPromotion(Function *F, AA.replaceWithNewValue(Call, New); // Update the callgraph to know that the callsite has been transformed. - CG[Call->getParent()->getParent()]->replaceCallSite(Call, New); + CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()]; + CalleeNode->replaceCallEdge(Call, New, NF_CGN); if (!Call->use_empty()) { Call->replaceAllUsesWith(New); @@ -740,8 +767,8 @@ Function *ArgPromotion::DoPromotion(Function *F, // function empty. NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList()); - // Loop over the argument list, transfering uses of the old arguments over to - // the new arguments, also transfering over the names as well. + // Loop over the argument list, transferring uses of the old arguments over to + // the new arguments, also transferring over the names as well. // for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), I2 = NF->arg_begin(); I != E; ++I) { @@ -761,17 +788,19 @@ Function *ArgPromotion::DoPromotion(Function *F, Instruction *InsertPt = NF->begin()->begin(); // Just add all the struct element types. - const Type *AgTy = cast(I->getType())->getElementType(); + Type *AgTy = cast(I->getType())->getElementType(); Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt); - const StructType *STy = cast(AgTy); - Value *Idxs[2] = { ConstantInt::get(Type::Int32Ty, 0), 0 }; + StructType *STy = cast(AgTy); + Value *Idxs[2] = { + ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 }; for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { - Idxs[1] = ConstantInt::get(Type::Int32Ty, i); - std::string Name = TheAlloca->getName()+"."+utostr(i); - Value *Idx = GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2, - Name, InsertPt); - I2->setName(I->getName()+"."+utostr(i)); + Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i); + Value *Idx = + GetElementPtrInst::Create(TheAlloca, Idxs, + TheAlloca->getName()+"."+Twine(i), + InsertPt); + I2->setName(I->getName()+"."+Twine(i)); new StoreInst(I2++, Idx, InsertPt); } @@ -779,6 +808,15 @@ Function *ArgPromotion::DoPromotion(Function *F, I->replaceAllUsesWith(TheAlloca); TheAlloca->takeName(I); AA.replaceWithNewValue(I, TheAlloca); + + // If the alloca is used in a call, we must clear the tail flag since + // the callee now uses an alloca from the caller. + for (User *U : TheAlloca->users()) { + CallInst *Call = dyn_cast(U); + if (!Call) + continue; + Call->setTailCall(false); + } continue; } @@ -793,17 +831,17 @@ Function *ArgPromotion::DoPromotion(Function *F, ScalarizeTable &ArgIndices = ScalarizedElements[I]; while (!I->use_empty()) { - if (LoadInst *LI = dyn_cast(I->use_back())) { + if (LoadInst *LI = dyn_cast(I->user_back())) { assert(ArgIndices.begin()->empty() && "Load element should sort to front!"); I2->setName(I->getName()+".val"); LI->replaceAllUsesWith(I2); AA.replaceWithNewValue(LI, I2); LI->eraseFromParent(); - DOUT << "*** Promoted load of argument '" << I->getName() - << "' in function '" << F->getName() << "'\n"; + DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName() + << "' in function '" << F->getName() << "'\n"); } else { - GetElementPtrInst *GEP = cast(I->use_back()); + GetElementPtrInst *GEP = cast(I->user_back()); IndicesVector Operands; Operands.reserve(GEP->getNumIndices()); for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end(); @@ -827,13 +865,13 @@ Function *ArgPromotion::DoPromotion(Function *F, NewName += ".val"; TheArg->setName(NewName); - DOUT << "*** Promoted agg argument '" << TheArg->getName() - << "' of function '" << NF->getName() << "'\n"; + DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName() + << "' of function '" << NF->getName() << "'\n"); // All of the uses must be load instructions. Replace them all with // the argument specified by ArgNo. while (!GEP->use_empty()) { - LoadInst *L = cast(GEP->use_back()); + LoadInst *L = cast(GEP->user_back()); L->replaceAllUsesWith(TheArg); AA.replaceWithNewValue(L, TheArg); L->eraseFromParent(); @@ -844,19 +882,23 @@ Function *ArgPromotion::DoPromotion(Function *F, } // Increment I2 past all of the arguments added for this promoted pointer. - for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i) - ++I2; + std::advance(I2, ArgIndices.size()); } - // Notify the alias analysis implementation that we inserted a new argument. - if (ExtraArgHack) - AA.copyValue(Constant::getNullValue(Type::Int32Ty), NF->arg_begin()); - - // Tell the alias analysis that the old function is about to disappear. AA.replaceWithNewValue(F, NF); - // Now that the old function is dead, delete it. - F->eraseFromParent(); - return NF; + + NF_CGN->stealCalledFunctionsFrom(CG[F]); + + // Now that the old function is dead, delete it. If there is a dangling + // reference to the CallgraphNode, just leave the dead function around for + // someone else to nuke. + CallGraphNode *CGN = CG[F]; + if (CGN->getNumReferences() == 0) + delete CG.removeFunctionFromModule(CGN); + else + F->setLinkage(Function::ExternalLinkage); + + return NF_CGN; }