X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FIPO%2FDeadArgumentElimination.cpp;h=dbeb365ec16e80b0349070163648e4fceba4fe45;hb=a3ec5d6eccb0fa3746050a3d3c739bd9718b6bd0;hp=f1599c361ff569c4763b858ecdc63730aaf47347;hpb=08227e4f17f5c7686bc6f0fe5d2c030e2fa05806;p=oota-llvm.git diff --git a/lib/Transforms/IPO/DeadArgumentElimination.cpp b/lib/Transforms/IPO/DeadArgumentElimination.cpp index f1599c361ff..dbeb365ec16 100644 --- a/lib/Transforms/IPO/DeadArgumentElimination.cpp +++ b/lib/Transforms/IPO/DeadArgumentElimination.cpp @@ -1,184 +1,841 @@ //===-- DeadArgumentElimination.cpp - Eliminate dead arguments ------------===// // +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// // This pass deletes dead arguments from internal functions. Dead argument // elimination removes arguments which are directly dead, as well as arguments -// only passed into function calls as dead arguments of other functions. +// only passed into function calls as dead arguments of other functions. This +// pass also deletes dead return values in a similar way. // // This pass is often useful as a cleanup pass to run after aggressive -// interprocedural passes, which add possibly-dead arguments. +// interprocedural passes, which add possibly-dead arguments or return values. // //===----------------------------------------------------------------------===// +#define DEBUG_TYPE "deadargelim" #include "llvm/Transforms/IPO.h" +#include "llvm/CallingConv.h" +#include "llvm/Constant.h" +#include "llvm/DerivedTypes.h" +#include "llvm/Instructions.h" +#include "llvm/IntrinsicInst.h" #include "llvm/Module.h" #include "llvm/Pass.h" -#include "llvm/DerivedTypes.h" -#include "llvm/Constant.h" -#include "llvm/iOther.h" -#include "llvm/iTerminators.h" #include "llvm/Support/CallSite.h" -#include "Support/Statistic.h" -#include "Support/iterator" +#include "llvm/Support/Debug.h" +#include "llvm/ADT/SmallVector.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/ADT/StringExtras.h" +#include "llvm/Support/Compiler.h" +#include #include +using namespace llvm; + +STATISTIC(NumArgumentsEliminated, "Number of unread args removed"); +STATISTIC(NumRetValsEliminated , "Number of unused return values removed"); namespace { - Statistic<> NumArgumentsEliminated("deadargelim", "Number of args removed"); + /// DAE - The dead argument elimination pass. + /// + class VISIBILITY_HIDDEN DAE : public ModulePass { + public: + + /// Struct that represents (part of) either a return value or a function + /// argument. Used so that arguments and return values can be used + /// interchangably. + struct RetOrArg { + RetOrArg(const Function* F, unsigned Idx, bool IsArg) : F(F), Idx(Idx), + IsArg(IsArg) {} + const Function *F; + unsigned Idx; + bool IsArg; + + /// Make RetOrArg comparable, so we can put it into a map. + bool operator<(const RetOrArg &O) const { + if (F != O.F) + return F < O.F; + else if (Idx != O.Idx) + return Idx < O.Idx; + else + return IsArg < O.IsArg; + } - struct DAE : public Pass { - bool run(Module &M); + /// Make RetOrArg comparable, so we can easily iterate the multimap. + bool operator==(const RetOrArg &O) const { + return F == O.F && Idx == O.Idx && IsArg == O.IsArg; + } + + std::string getDescription() { + return std::string((IsArg ? "Argument #" : "Return value #")) + + utostr(Idx) + " of function " + F->getName(); + } + }; + + /// Liveness enum - During our initial pass over the program, we determine + /// that things are either alive or maybe alive. We don't mark anything + /// explicitly dead (even if we know they are), since anything not alive + /// with no registered uses (in Uses) will never be marked alive and will + /// thus become dead in the end. + enum Liveness { Live, MaybeLive }; + + /// Convenience wrapper + RetOrArg CreateRet(const Function *F, unsigned Idx) { + return RetOrArg(F, Idx, false); + } + /// Convenience wrapper + RetOrArg CreateArg(const Function *F, unsigned Idx) { + return RetOrArg(F, Idx, true); + } + + typedef std::multimap UseMap; + /// This maps a return value or argument to any MaybeLive return values or + /// arguments it uses. This allows the MaybeLive values to be marked live + /// when any of its users is marked live. + /// For example (indices are left out for clarity): + /// - Uses[ret F] = ret G + /// This means that F calls G, and F returns the value returned by G. + /// - Uses[arg F] = ret G + /// This means that some function calls G and passes its result as an + /// argument to F. + /// - Uses[ret F] = arg F + /// This means that F returns one of its own arguments. + /// - Uses[arg F] = arg G + /// This means that G calls F and passes one of its own (G's) arguments + /// directly to F. + UseMap Uses; + + typedef std::set LiveSet; + + /// This set contains all values that have been determined to be live. + LiveSet LiveValues; + + typedef SmallVector UseVector; + + public: + static char ID; // Pass identification, replacement for typeid + DAE() : ModulePass((intptr_t)&ID) {} + bool runOnModule(Module &M); + + virtual bool ShouldHackArguments() const { return false; } + + private: + Liveness MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses); + Liveness SurveyUse(Value::use_iterator U, UseVector &MaybeLiveUses, + unsigned RetValNum = 0); + Liveness SurveyUses(Value *V, UseVector &MaybeLiveUses); + + void SurveyFunction(Function &F); + void MarkValue(const RetOrArg &RA, Liveness L, + const UseVector &MaybeLiveUses); + void MarkLive(RetOrArg RA); + void MarkLive(const Function &F); + bool RemoveDeadStuffFromFunction(Function *F); + bool DeleteDeadVarargs(Function &Fn); }; - RegisterOpt X("deadargelim", "Dead Argument Elimination"); } -// createDeadArgEliminationPass - This pass removes arguments from functions -// which are not used by the body of the function. -// -Pass *createDeadArgEliminationPass() { return new DAE(); } +char DAE::ID = 0; +static RegisterPass +X("deadargelim", "Dead Argument Elimination"); -// FunctionArgumentsIntrinsicallyAlive - Return true if the arguments of the -// specified function are intrinsically alive. -// -// We consider arguments of non-internal functions to be intrinsically alive as -// well as arguments to functions which have their "address taken". -// -static bool FunctionArgumentsIntrinsicallyAlive(const Function &F) { - if (!F.hasInternalLinkage()) return true; - - for (Value::use_const_iterator I = F.use_begin(), E = F.use_end(); I!=E; ++I){ - // If this use is anything other than a call site, the function is alive. - CallSite CS = CallSite::get(const_cast(*I)); - if (!CS.getInstruction()) return true; // Not a valid call site? - - // If the function is PASSED IN as an argument, its address has been taken - for (CallSite::arg_iterator AI = CS.arg_begin(), E = CS.arg_end(); AI != E; - ++AI) - if (AI->get() == &F) return true; +namespace { + /// DAH - DeadArgumentHacking pass - Same as dead argument elimination, but + /// deletes arguments to functions which are external. This is only for use + /// by bugpoint. + struct DAH : public DAE { + static char ID; + virtual bool ShouldHackArguments() const { return true; } + }; +} + +char DAH::ID = 0; +static RegisterPass +Y("deadarghaX0r", "Dead Argument Hacking (BUGPOINT USE ONLY; DO NOT USE)"); + +/// createDeadArgEliminationPass - This pass removes arguments from functions +/// which are not used by the body of the function. +/// +ModulePass *llvm::createDeadArgEliminationPass() { return new DAE(); } +ModulePass *llvm::createDeadArgHackingPass() { return new DAH(); } + +/// DeleteDeadVarargs - If this is an function that takes a ... list, and if +/// llvm.vastart is never called, the varargs list is dead for the function. +bool DAE::DeleteDeadVarargs(Function &Fn) { + assert(Fn.getFunctionType()->isVarArg() && "Function isn't varargs!"); + if (Fn.isDeclaration() || !Fn.hasInternalLinkage()) return false; + + // Ensure that the function is only directly called. + for (Value::use_iterator I = Fn.use_begin(), E = Fn.use_end(); I != E; ++I) { + // If this use is anything other than a call site, give up. + CallSite CS = CallSite::get(*I); + Instruction *TheCall = CS.getInstruction(); + if (!TheCall) return false; // Not a direct call site? + + // The addr of this function is passed to the call. + if (I.getOperandNo() != 0) return false; + } + + // Okay, we know we can transform this function if safe. Scan its body + // looking for calls to llvm.vastart. + for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { + for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { + if (IntrinsicInst *II = dyn_cast(I)) { + if (II->getIntrinsicID() == Intrinsic::vastart) + return false; + } + } + } + + // If we get here, there are no calls to llvm.vastart in the function body, + // remove the "..." and adjust all the calls. + + // Start by computing a new prototype for the function, which is the same as + // the old function, but doesn't have isVarArg set. + const FunctionType *FTy = Fn.getFunctionType(); + std::vector Params(FTy->param_begin(), FTy->param_end()); + FunctionType *NFTy = FunctionType::get(FTy->getReturnType(), Params, false); + unsigned NumArgs = Params.size(); + + // Create the new function body and insert it into the module... + Function *NF = Function::Create(NFTy, Fn.getLinkage()); + NF->copyAttributesFrom(&Fn); + Fn.getParent()->getFunctionList().insert(&Fn, NF); + NF->takeName(&Fn); + + // Loop over all of the callers of the function, transforming the call sites + // to pass in a smaller number of arguments into the new function. + // + std::vector Args; + while (!Fn.use_empty()) { + CallSite CS = CallSite::get(Fn.use_back()); + Instruction *Call = CS.getInstruction(); + + // Pass all the same arguments. + Args.assign(CS.arg_begin(), CS.arg_begin()+NumArgs); + + // Drop any attributes that were on the vararg arguments. + PAListPtr PAL = CS.getParamAttrs(); + if (!PAL.isEmpty() && PAL.getSlot(PAL.getNumSlots() - 1).Index > NumArgs) { + SmallVector ParamAttrsVec; + for (unsigned i = 0; PAL.getSlot(i).Index <= NumArgs; ++i) + ParamAttrsVec.push_back(PAL.getSlot(i)); + PAL = PAListPtr::get(ParamAttrsVec.begin(), ParamAttrsVec.end()); + } + + Instruction *New; + if (InvokeInst *II = dyn_cast(Call)) { + New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), + Args.begin(), Args.end(), "", Call); + cast(New)->setCallingConv(CS.getCallingConv()); + cast(New)->setParamAttrs(PAL); + } else { + New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); + cast(New)->setCallingConv(CS.getCallingConv()); + cast(New)->setParamAttrs(PAL); + if (cast(Call)->isTailCall()) + cast(New)->setTailCall(); + } + Args.clear(); + + if (!Call->use_empty()) + Call->replaceAllUsesWith(New); + + New->takeName(Call); + + // Finally, remove the old call from the program, reducing the use-count of + // F. + Call->eraseFromParent(); } - return false; + + // Since we have now created the new function, splice the body of the old + // function right into the new function, leaving the old rotting hulk of the + // function empty. + NF->getBasicBlockList().splice(NF->begin(), Fn.getBasicBlockList()); + + // Loop over the argument list, transfering uses of the old arguments over to + // the new arguments, also transfering over the names as well. While we're at + // it, remove the dead arguments from the DeadArguments list. + // + for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(), + I2 = NF->arg_begin(); I != E; ++I, ++I2) { + // Move the name and users over to the new version. + I->replaceAllUsesWith(I2); + I2->takeName(I); + } + + // Finally, nuke the old function. + Fn.eraseFromParent(); + return true; } -namespace { - enum ArgumentLiveness { Alive, MaybeLive, Dead }; +/// Convenience function that returns the number of return values. It returns 0 +/// for void functions and 1 for functions not returning a struct. It returns +/// the number of struct elements for functions returning a struct. +static unsigned NumRetVals(const Function *F) { + if (F->getReturnType() == Type::VoidTy) + return 0; + else if (const StructType *STy = dyn_cast(F->getReturnType())) + return STy->getNumElements(); + else + return 1; } -// getArgumentLiveness - Inspect an argument, determining if is known Alive -// (used in a computation), MaybeLive (only passed as an argument to a call), or -// Dead (not used). -static ArgumentLiveness getArgumentLiveness(const Argument &A) { - if (A.use_empty()) return Dead; // First check, directly dead? - - // Scan through all of the uses, looking for non-argument passing uses. - for (Value::use_const_iterator I = A.use_begin(), E = A.use_end(); I!=E;++I) { - CallSite CS = CallSite::get(const_cast(*I)); - if (!CS.getInstruction()) { - // If its used by something that is not a call or invoke, it's alive! - return Alive; +/// MarkIfNotLive - This checks Use for liveness in LiveValues. If Use is not +/// live, it adds Use to the MaybeLiveUses argument. Returns the determined +/// liveness of Use. +DAE::Liveness DAE::MarkIfNotLive(RetOrArg Use, UseVector &MaybeLiveUses) { + // We're live if our use is already marked as live. + if (LiveValues.count(Use)) + return Live; + + // We're maybe live otherwise, but remember that we must become live if + // Use becomes live. + MaybeLiveUses.push_back(Use); + return MaybeLive; +} + + +/// SurveyUse - This looks at a single use of an argument or return value +/// and determines if it should be alive or not. Adds this use to MaybeLiveUses +/// if it causes the used value to become MaybeAlive. +/// +/// RetValNum is the return value number to use when this use is used in a +/// return instruction. This is used in the recursion, you should always leave +/// it at 0. +DAE::Liveness DAE::SurveyUse(Value::use_iterator U, UseVector &MaybeLiveUses, + unsigned RetValNum) { + Value *V = *U; + if (ReturnInst *RI = dyn_cast(V)) { + // The value is returned from a function. It's only live when the + // function's return value is live. We use RetValNum here, for the case + // that U is really a use of an insertvalue instruction that uses the + // orginal Use. + RetOrArg Use = CreateRet(RI->getParent()->getParent(), RetValNum); + // We might be live, depending on the liveness of Use. + return MarkIfNotLive(Use, MaybeLiveUses); } - // If it's an indirect call, mark it alive... - Function *Callee = CS.getCalledFunction(); - if (!Callee) return Alive; + if (InsertValueInst *IV = dyn_cast(V)) { + if (U.getOperandNo() != InsertValueInst::getAggregateOperandIndex() + && IV->hasIndices()) + // The use we are examining is inserted into an aggregate. Our liveness + // depends on all uses of that aggregate, but if it is used as a return + // value, only index at which we were inserted counts. + RetValNum = *IV->idx_begin(); + + // Note that if we are used as the aggregate operand to the insertvalue, + // we don't change RetValNum, but do survey all our uses. + + Liveness Result = MaybeLive; + for (Value::use_iterator I = IV->use_begin(), + E = V->use_end(); I != E; ++I) { + Result = SurveyUse(I, MaybeLiveUses, RetValNum); + if (Result == Live) + break; + } + return Result; + } + CallSite CS = CallSite::get(V); + if (CS.getInstruction()) { + Function *F = CS.getCalledFunction(); + if (F) { + // Used in a direct call. + + // Find the argument number. We know for sure that this use is an + // argument, since if it was the function argument this would be an + // indirect call and the we know can't be looking at a value of the + // label type (for the invoke instruction). + unsigned ArgNo = CS.getArgumentNo(U.getOperandNo()); + + if (ArgNo >= F->getFunctionType()->getNumParams()) + // The value is passed in through a vararg! Must be live. + return Live; + + assert(CS.getArgument(ArgNo) + == CS.getInstruction()->getOperand(U.getOperandNo()) + && "Argument is not where we expected it"); + + // Value passed to a normal call. It's only live when the corresponding + // argument to the called function turns out live. + RetOrArg Use = CreateArg(F, ArgNo); + return MarkIfNotLive(Use, MaybeLiveUses); + } + } + // Used in any other way? Value must be live. + return Live; +} - // FIXME: check to see if it's passed through a va_arg area +/// SurveyUses - This looks at all the uses of the given value +/// Returns the Liveness deduced from the uses of this value. +/// +/// Adds all uses that cause the result to be MaybeLive to MaybeLiveRetUses. If +/// the result is Live, MaybeLiveUses might be modified but its content should +/// be ignored (since it might not be complete). +DAE::Liveness DAE::SurveyUses(Value *V, UseVector &MaybeLiveUses) { + // Assume it's dead (which will only hold if there are no uses at all..). + Liveness Result = MaybeLive; + // Check each use. + for (Value::use_iterator I = V->use_begin(), + E = V->use_end(); I != E; ++I) { + Result = SurveyUse(I, MaybeLiveUses); + if (Result == Live) + break; } - - return MaybeLive; // It must be used, but only as argument to a function + return Result; } -// isMaybeLiveArgumentNowAlive - Check to see if Arg is alive. At this point, -// we know that the only uses of Arg are to be passed in as an argument to a -// function call. Check to see if the formal argument passed in is in the -// LiveArguments set. If so, return true. +// SurveyFunction - This performs the initial survey of the specified function, +// checking out whether or not it uses any of its incoming arguments or whether +// any callers use the return value. This fills in the LiveValues set and Uses +// map. // -static bool isMaybeLiveArgumentNowAlive(Argument *Arg, - const std::set &LiveArguments) { - for (Value::use_iterator I = Arg->use_begin(), E = Arg->use_end(); I!=E; ++I){ - CallSite CS = CallSite::get(*I); +// We consider arguments of non-internal functions to be intrinsically alive as +// well as arguments to functions which have their "address taken". +// +void DAE::SurveyFunction(Function &F) { + unsigned RetCount = NumRetVals(&F); + // Assume all return values are dead + typedef SmallVector RetVals; + RetVals RetValLiveness(RetCount, MaybeLive); + + typedef SmallVector RetUses; + // These vectors map each return value to the uses that make it MaybeLive, so + // we can add those to the Uses map if the return value really turns out to be + // MaybeLive. Initialized to a list of RetCount empty lists. + RetUses MaybeLiveRetUses(RetCount); + + for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) + if (ReturnInst *RI = dyn_cast(BB->getTerminator())) + if (RI->getNumOperands() != 0 && RI->getOperand(0)->getType() + != F.getFunctionType()->getReturnType()) { + // We don't support old style multiple return values. + MarkLive(F); + return; + } + + if (!F.hasInternalLinkage() && (!ShouldHackArguments() || F.isIntrinsic())) { + MarkLive(F); + return; + } + + DOUT << "DAE - Inspecting callers for fn: " << F.getName() << "\n"; + // Keep track of the number of live retvals, so we can skip checks once all + // of them turn out to be live. + unsigned NumLiveRetVals = 0; + const Type *STy = dyn_cast(F.getReturnType()); + // Loop all uses of the function. + for (Value::use_iterator I = F.use_begin(), E = F.use_end(); I != E; ++I) { + // If the function is PASSED IN as an argument, its address has been + // taken. + if (I.getOperandNo() != 0) { + MarkLive(F); + return; + } + + // If this use is anything other than a call site, the function is alive. + CallSite CS = CallSite::get(*I); + Instruction *TheCall = CS.getInstruction(); + if (!TheCall) { // Not a direct call site? + MarkLive(F); + return; + } + + // If we end up here, we are looking at a direct call to our function. + + // Now, check how our return value(s) is/are used in this caller. Don't + // bother checking return values if all of them are live already. + if (NumLiveRetVals != RetCount) { + if (STy) { + // Check all uses of the return value. + for (Value::use_iterator I = TheCall->use_begin(), + E = TheCall->use_end(); I != E; ++I) { + ExtractValueInst *Ext = dyn_cast(*I); + if (Ext && Ext->hasIndices()) { + // This use uses a part of our return value, survey the uses of + // that part and store the results for this index only. + unsigned Idx = *Ext->idx_begin(); + if (RetValLiveness[Idx] != Live) { + RetValLiveness[Idx] = SurveyUses(Ext, MaybeLiveRetUses[Idx]); + if (RetValLiveness[Idx] == Live) + NumLiveRetVals++; + } + } else { + // Used by something else than extractvalue. Mark all return + // values as live. + for (unsigned i = 0; i != RetCount; ++i ) + RetValLiveness[i] = Live; + NumLiveRetVals = RetCount; + break; + } + } + } else { + // Single return value + RetValLiveness[0] = SurveyUses(TheCall, MaybeLiveRetUses[0]); + if (RetValLiveness[0] == Live) + NumLiveRetVals = RetCount; + } + } + } - // We know that this can only be used for direct calls... - Function *Callee = cast(CS.getCalledValue()); - - // Loop over all of the arguments (because Arg may be passed into the call - // multiple times) and check to see if any are now alive... - CallSite::arg_iterator CSAI = CS.arg_begin(); - for (Function::aiterator AI = Callee->abegin(), E = Callee->aend(); - AI != E; ++AI, ++CSAI) - // If this is the argument we are looking for, check to see if it's alive - if (*CSAI == Arg && LiveArguments.count(AI)) - return true; + // Now we've inspected all callers, record the liveness of our return values. + for (unsigned i = 0; i != RetCount; ++i) + MarkValue(CreateRet(&F, i), RetValLiveness[i], MaybeLiveRetUses[i]); + + DOUT << "DAE - Inspecting args for fn: " << F.getName() << "\n"; + + // Now, check all of our arguments. + unsigned i = 0; + UseVector MaybeLiveArgUses; + for (Function::arg_iterator AI = F.arg_begin(), + E = F.arg_end(); AI != E; ++AI, ++i) { + // See what the effect of this use is (recording any uses that cause + // MaybeLive in MaybeLiveArgUses). + Liveness Result = SurveyUses(AI, MaybeLiveArgUses); + // Mark the result. + MarkValue(CreateArg(&F, i), Result, MaybeLiveArgUses); + // Clear the vector again for the next iteration. + MaybeLiveArgUses.clear(); } - return false; } -// MarkArgumentLive - The MaybeLive argument 'Arg' is now known to be alive. -// Mark it live in the specified sets and recursively mark arguments in callers -// live that are needed to pass in a value. -// -static void MarkArgumentLive(Argument *Arg, - std::set &MaybeLiveArguments, - std::set &LiveArguments, - const std::multimap &CallSites) { - DEBUG(std::cerr << " MaybeLive argument now live: " << Arg->getName()<<"\n"); - assert(MaybeLiveArguments.count(Arg) && !LiveArguments.count(Arg) && - "Arg not MaybeLive?"); - MaybeLiveArguments.erase(Arg); - LiveArguments.insert(Arg); - - // Loop over all of the call sites of the function, making any arguments - // passed in to provide a value for this argument live as necessary. - // - Function *Fn = Arg->getParent(); - unsigned ArgNo = std::distance(Fn->abegin(), Function::aiterator(Arg)); - - std::multimap::const_iterator I = - CallSites.lower_bound(Fn); - for (; I != CallSites.end() && I->first == Fn; ++I) { - const CallSite &CS = I->second; - if (Argument *ActualArg = dyn_cast(*(CS.arg_begin()+ArgNo))) - if (MaybeLiveArguments.count(ActualArg)) - MarkArgumentLive(ActualArg, MaybeLiveArguments, LiveArguments, - CallSites); +/// MarkValue - This function marks the liveness of RA depending on L. If L is +/// MaybeLive, it also takes all uses in MaybeLiveUses and records them in Uses, +/// such that RA will be marked live if any use in MaybeLiveUses gets marked +/// live later on. +void DAE::MarkValue(const RetOrArg &RA, Liveness L, + const UseVector &MaybeLiveUses) { + switch (L) { + case Live: MarkLive(RA); break; + case MaybeLive: + { + // Note any uses of this value, so this return value can be + // marked live whenever one of the uses becomes live. + for (UseVector::const_iterator UI = MaybeLiveUses.begin(), + UE = MaybeLiveUses.end(); UI != UE; ++UI) + Uses.insert(std::make_pair(*UI, RA)); + break; + } } } -// RemoveDeadArgumentsFromFunction - We know that F has dead arguments, as -// specified by the DeadArguments list. Transform the function and all of the -// callees of the function to not have these arguments. +/// MarkLive - Mark the given Function as alive, meaning that it cannot be +/// changed in any way. Additionally, +/// mark any values that are used as this function's parameters or by its return +/// values (according to Uses) live as well. +void DAE::MarkLive(const Function &F) { + DOUT << "DAE - Intrinsically live fn: " << F.getName() << "\n"; + // Mark all arguments as live. + for (unsigned i = 0, e = F.arg_size(); i != e; ++i) + MarkLive(CreateArg(&F, i)); + // Mark all return values as live. + for (unsigned i = 0, e = NumRetVals(&F); i != e; ++i) + MarkLive(CreateRet(&F, i)); +} + +/// MarkLive - Mark the given return value or argument as live. Additionally, +/// mark any values that are used by this value (according to Uses) live as +/// well. +void DAE::MarkLive(RetOrArg RA) { + if (!LiveValues.insert(RA).second) + return; // We were already marked Live. + + if (RA.IsArg) + DOUT << "DAE - Marking argument " << RA.Idx << " to function " + << RA.F->getNameStart() << " live\n"; + else + DOUT << "DAE - Marking return value " << RA.Idx << " of function " + << RA.F->getNameStart() << " live\n"; + + // We don't use upper_bound (or equal_range) here, because our recursive call + // to ourselves is likely to cause the upper_bound (which is the first value + // not belonging to RA) to become erased and the iterator invalidated. + UseMap::iterator Begin = Uses.lower_bound(RA); + UseMap::iterator E = Uses.end(); + UseMap::iterator I; + for (I = Begin; I != E && I->first == RA; ++I) + MarkLive(I->second); + + // Erase RA from the Uses map (from the lower bound to wherever we ended up + // after the loop). + Uses.erase(Begin, I); +} + +// RemoveDeadStuffFromFunction - Remove any arguments and return values from F +// that are not in LiveValues. Transform the function and all of the callees of +// the function to not have these arguments and return values. // -static void RemoveDeadArgumentsFromFunction(Function *F, - std::set &DeadArguments){ +bool DAE::RemoveDeadStuffFromFunction(Function *F) { + // Quick exit path for external functions + if (!F->hasInternalLinkage() && (!ShouldHackArguments() || F->isIntrinsic())) + return false; + // Start by computing a new prototype for the function, which is the same as - // the old function, but has fewer arguments. + // the old function, but has fewer arguments and a different return type. const FunctionType *FTy = F->getFunctionType(); std::vector Params; - for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I) - if (!DeadArguments.count(I)) + // Set up to build a new list of parameter attributes. + SmallVector ParamAttrsVec; + const PAListPtr &PAL = F->getParamAttrs(); + + // The existing function return attributes. + ParameterAttributes RAttrs = PAL.getParamAttrs(0); + + + // Find out the new return value. + + const Type *RetTy = FTy->getReturnType(); + const Type *NRetTy = NULL; + unsigned RetCount = NumRetVals(F); + // Explicitly track if anything changed, for debugging. + bool Changed = false; + // -1 means unused, other numbers are the new index + SmallVector NewRetIdxs(RetCount, -1); + std::vector RetTypes; + if (RetTy == Type::VoidTy) { + NRetTy = Type::VoidTy; + } else { + const StructType *STy = dyn_cast(RetTy); + if (STy) + // Look at each of the original return values individually. + for (unsigned i = 0; i != RetCount; ++i) { + RetOrArg Ret = CreateRet(F, i); + if (LiveValues.erase(Ret)) { + RetTypes.push_back(STy->getElementType(i)); + NewRetIdxs[i] = RetTypes.size() - 1; + } else { + ++NumRetValsEliminated; + DOUT << "DAE - Removing return value " << i << " from " + << F->getNameStart() << "\n"; + Changed = true; + } + } + else + // We used to return a single value. + if (LiveValues.erase(CreateRet(F, 0))) { + RetTypes.push_back(RetTy); + NewRetIdxs[0] = 0; + } else { + DOUT << "DAE - Removing return value from " << F->getNameStart() + << "\n"; + ++NumRetValsEliminated; + Changed = true; + } + if (RetTypes.size() > 1 || (STy && STy->getNumElements()==RetTypes.size())) + // More than one return type? Return a struct with them. Also, if we used + // to return a struct and didn't change the number of return values, + // return a struct again. This prevents changing {something} into + // something and {} into void. + // Make the new struct packed if we used to return a packed struct + // already. + NRetTy = StructType::get(RetTypes, STy->isPacked()); + else if (RetTypes.size() == 1) + // One return type? Just a simple value then, but only if we didn't use to + // return a struct with that simple value before. + NRetTy = RetTypes.front(); + else if (RetTypes.size() == 0) + // No return types? Make it void, but only if we didn't use to return {}. + NRetTy = Type::VoidTy; + } + + assert(NRetTy && "No new return type found?"); + + // Remove any incompatible attributes, but only if we removed all return + // values. Otherwise, ensure that we don't have any conflicting attributes + // here. Currently, this should not be possible, but special handling might be + // required when new return value attributes are added. + if (NRetTy == Type::VoidTy) + RAttrs &= ~ParamAttr::typeIncompatible(NRetTy); + else + assert((RAttrs & ParamAttr::typeIncompatible(NRetTy)) == 0 + && "Return attributes no longer compatible?"); + + if (RAttrs) + ParamAttrsVec.push_back(ParamAttrsWithIndex::get(0, RAttrs)); + + // Remember which arguments are still alive. + SmallVector ArgAlive(FTy->getNumParams(), false); + // Construct the new parameter list from non-dead arguments. Also construct + // a new set of parameter attributes to correspond. Skip the first parameter + // attribute, since that belongs to the return value. + unsigned i = 0; + for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); + I != E; ++I, ++i) { + RetOrArg Arg = CreateArg(F, i); + if (LiveValues.erase(Arg)) { Params.push_back(I->getType()); + ArgAlive[i] = true; + + // Get the original parameter attributes (skipping the first one, that is + // for the return value. + if (ParameterAttributes Attrs = PAL.getParamAttrs(i + 1)) + ParamAttrsVec.push_back(ParamAttrsWithIndex::get(Params.size(), Attrs)); + } else { + ++NumArgumentsEliminated; + DOUT << "DAE - Removing argument " << i << " (" << I->getNameStart() + << ") from " << F->getNameStart() << "\n"; + Changed = true; + } + } + + // Reconstruct the ParamAttrsList based on the vector we constructed. + PAListPtr NewPAL = PAListPtr::get(ParamAttrsVec.begin(), ParamAttrsVec.end()); + + // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which + // have zero fixed arguments. + // + // Note that we apply this hack for a vararg fuction that does not have any + // arguments anymore, but did have them before (so don't bother fixing + // functions that were already broken wrt CWriter). + bool ExtraArgHack = false; + if (Params.empty() && FTy->isVarArg() && FTy->getNumParams() != 0) { + ExtraArgHack = true; + Params.push_back(Type::Int32Ty); + } + + // Create the new function type based on the recomputed parameters. + FunctionType *NFTy = FunctionType::get(NRetTy, Params, FTy->isVarArg()); + + // No change? + if (NFTy == FTy) + return false; + + // The function type is only allowed to be different if we actually left out + // an argument or return value. + assert(Changed && "Function type changed while no arguments or return values" + "were removed!"); - FunctionType *NFTy = FunctionType::get(FTy->getReturnType(), Params, - FTy->isVarArg()); - // Create the new function body and insert it into the module... - Function *NF = new Function(NFTy, Function::InternalLinkage, F->getName()); + Function *NF = Function::Create(NFTy, F->getLinkage()); + NF->copyAttributesFrom(F); + NF->setParamAttrs(NewPAL); + // Insert the new function before the old function, so we won't be processing + // it again. F->getParent()->getFunctionList().insert(F, NF); + NF->takeName(F); // Loop over all of the callers of the function, transforming the call sites // to pass in a smaller number of arguments into the new function. // + std::vector Args; while (!F->use_empty()) { CallSite CS = CallSite::get(F->use_back()); Instruction *Call = CS.getInstruction(); - CS.setCalledFunction(NF); // Reduce the uses count of F - - // Loop over the operands, deleting dead ones... - CallSite::arg_iterator AI = CS.arg_begin(); - for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I) - if (DeadArguments.count(I)) { // Remove operands for dead arguments - AI = Call->op_erase(AI); - } else { - ++AI; // Leave live operands alone... + + ParamAttrsVec.clear(); + const PAListPtr &CallPAL = CS.getParamAttrs(); + + // The call return attributes. + ParameterAttributes RAttrs = CallPAL.getParamAttrs(0); + // Adjust in case the function was changed to return void. + RAttrs &= ~ParamAttr::typeIncompatible(NF->getReturnType()); + if (RAttrs) + ParamAttrsVec.push_back(ParamAttrsWithIndex::get(0, RAttrs)); + + // Declare these outside of the loops, so we can reuse them for the second + // loop, which loops the varargs. + CallSite::arg_iterator I = CS.arg_begin(); + unsigned i = 0; + // Loop over those operands, corresponding to the normal arguments to the + // original function, and add those that are still alive. + for (unsigned e = FTy->getNumParams(); i != e; ++I, ++i) + if (ArgAlive[i]) { + Args.push_back(*I); + // Get original parameter attributes, but skip return attributes. + if (ParameterAttributes Attrs = CallPAL.getParamAttrs(i + 1)) + ParamAttrsVec.push_back(ParamAttrsWithIndex::get(Args.size(), Attrs)); + } + + if (ExtraArgHack) + Args.push_back(UndefValue::get(Type::Int32Ty)); + + // Push any varargs arguments on the list. Don't forget their attributes. + for (CallSite::arg_iterator E = CS.arg_end(); I != E; ++I, ++i) { + Args.push_back(*I); + if (ParameterAttributes Attrs = CallPAL.getParamAttrs(i + 1)) + ParamAttrsVec.push_back(ParamAttrsWithIndex::get(Args.size(), Attrs)); + } + + // Reconstruct the ParamAttrsList based on the vector we constructed. + PAListPtr NewCallPAL = PAListPtr::get(ParamAttrsVec.begin(), + ParamAttrsVec.end()); + + Instruction *New; + if (InvokeInst *II = dyn_cast(Call)) { + New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(), + Args.begin(), Args.end(), "", Call); + cast(New)->setCallingConv(CS.getCallingConv()); + cast(New)->setParamAttrs(NewCallPAL); + } else { + New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call); + cast(New)->setCallingConv(CS.getCallingConv()); + cast(New)->setParamAttrs(NewCallPAL); + if (cast(Call)->isTailCall()) + cast(New)->setTailCall(); + } + Args.clear(); + + if (!Call->use_empty()) { + if (New->getType() == Call->getType()) { + // Return type not changed? Just replace users then. + Call->replaceAllUsesWith(New); + New->takeName(Call); + } else if (New->getType() == Type::VoidTy) { + // Our return value has uses, but they will get removed later on. + // Replace by null for now. + Call->replaceAllUsesWith(Constant::getNullValue(Call->getType())); + } else { + assert(isa(RetTy) && "Return type changed, but not into a" + "void. The old return type must have" + "been a struct!"); + // The original return value was a struct, update all uses (which are + // all extractvalue instructions). + for (Value::use_iterator I = Call->use_begin(), E = Call->use_end(); + I != E;) { + assert(isa(*I) && "Return value not only used by" + "extractvalue?"); + ExtractValueInst *EV = cast(*I); + // Increment now, since we're about to throw away this use. + ++I; + assert(EV->hasIndices() && "Return value used by extractvalue without" + "indices?"); + unsigned Idx = *EV->idx_begin(); + if (NewRetIdxs[Idx] != -1) { + if (RetTypes.size() > 1) { + // We're still returning a struct, create a new extractvalue + // instruction with the first index updated + std::vector NewIdxs(EV->idx_begin(), EV->idx_end()); + NewIdxs[0] = NewRetIdxs[Idx]; + Value *NEV = ExtractValueInst::Create(New, NewIdxs.begin(), + NewIdxs.end(), "retval", + EV); + EV->replaceAllUsesWith(NEV); + EV->eraseFromParent(); + } else { + // We are now only returning a simple value, remove the + // extractvalue. + EV->replaceAllUsesWith(New); + EV->eraseFromParent(); + } + } else { + // Value unused, replace uses by null for now, they will get removed + // later on. + EV->replaceAllUsesWith(Constant::getNullValue(EV->getType())); + EV->eraseFromParent(); + } + } + New->takeName(Call); } + } + + // Finally, remove the old call from the program, reducing the use-count of + // F. + Call->eraseFromParent(); } // Since we have now created the new function, splice the body of the old @@ -187,118 +844,99 @@ static void RemoveDeadArgumentsFromFunction(Function *F, 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. While we're at - // it, remove the dead arguments from the DeadArguments list. - // - for (Function::aiterator I = F->abegin(), E = F->aend(), I2 = NF->abegin(); - I != E; ++I) - if (!DeadArguments.count(I)) { + // the new arguments, also transfering over the names as well. + i = 0; + for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(), + I2 = NF->arg_begin(); I != E; ++I, ++i) + if (ArgAlive[i]) { // If this is a live argument, move the name and users over to the new // version. I->replaceAllUsesWith(I2); - I2->setName(I->getName()); + I2->takeName(I); ++I2; } else { // If this argument is dead, replace any uses of it with null constants - // (these are guaranteed to only be operands to call instructions which - // will later be simplified). + // (these are guaranteed to become unused later on). I->replaceAllUsesWith(Constant::getNullValue(I->getType())); - DeadArguments.erase(I); } + // If we change the return value of the function we must rewrite any return + // instructions. Check this now. + if (F->getReturnType() != NF->getReturnType()) + for (Function::iterator BB = NF->begin(), E = NF->end(); BB != E; ++BB) + if (ReturnInst *RI = dyn_cast(BB->getTerminator())) { + Value *RetVal; + + if (NFTy->getReturnType() == Type::VoidTy) { + RetVal = 0; + } else { + assert (isa(RetTy)); + // The original return value was a struct, insert + // extractvalue/insertvalue chains to extract only the values we need + // to return and insert them into our new result. + // This does generate messy code, but we'll let it to instcombine to + // clean that up. + Value *OldRet = RI->getOperand(0); + // Start out building up our return value from undef + RetVal = llvm::UndefValue::get(NRetTy); + for (unsigned i = 0; i != RetCount; ++i) + if (NewRetIdxs[i] != -1) { + ExtractValueInst *EV = ExtractValueInst::Create(OldRet, i, + "oldret", RI); + if (RetTypes.size() > 1) { + // We're still returning a struct, so reinsert the value into + // our new return value at the new index + + RetVal = InsertValueInst::Create(RetVal, EV, NewRetIdxs[i], + "newret", RI); + } else { + // We are now only returning a simple value, so just return the + // extracted value. + RetVal = EV; + } + } + } + // Replace the return instruction with one returning the new return + // value (possibly 0 if we became void). + ReturnInst::Create(RetVal, RI); + BB->getInstList().erase(RI); + } + // Now that the old function is dead, delete it. - F->getParent()->getFunctionList().erase(F); -} + F->eraseFromParent(); -bool DAE::run(Module &M) { - // First phase: loop through the module, determining which arguments are live. - // We assume all arguments are dead unless proven otherwise (allowing us to - // determing that dead arguments passed into recursive functions are dead). - // - std::set LiveArguments, MaybeLiveArguments, DeadArguments; - std::multimap CallSites; - - DEBUG(std::cerr << "DAE - Determining liveness\n"); - for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) { - Function &Fn = *I; - // If the function is intrinsically alive, just mark the arguments alive. - if (FunctionArgumentsIntrinsicallyAlive(Fn)) { - for (Function::aiterator AI = Fn.abegin(), E = Fn.aend(); AI != E; ++AI) - LiveArguments.insert(AI); - DEBUG(std::cerr << " Args intrinsically live for fn: " << Fn.getName() - << "\n"); - } else { - DEBUG(std::cerr << " Inspecting args for fn: " << Fn.getName() << "\n"); - - // If it is not intrinsically alive, we know that all users of the - // function are call sites. Mark all of the arguments live which are - // directly used, and keep track of all of the call sites of this function - // if there are any arguments we assume that are dead. - // - bool AnyMaybeLiveArgs = false; - for (Function::aiterator AI = Fn.abegin(), E = Fn.aend(); AI != E; ++AI) - switch (getArgumentLiveness(*AI)) { - case Alive: - DEBUG(std::cerr << " Arg live by use: " << AI->getName() << "\n"); - LiveArguments.insert(AI); - break; - case Dead: - DEBUG(std::cerr << " Arg definately dead: " <getName()<<"\n"); - DeadArguments.insert(AI); - break; - case MaybeLive: - DEBUG(std::cerr << " Arg only passed to calls: " - << AI->getName() << "\n"); - AnyMaybeLiveArgs = true; - MaybeLiveArguments.insert(AI); - break; - } + return true; +} - // If there are any "MaybeLive" arguments, we need to check callees of - // this function when/if they become alive. Record which functions are - // callees... - if (AnyMaybeLiveArgs) - for (Value::use_iterator I = Fn.use_begin(), E = Fn.use_end(); - I != E; ++I) - CallSites.insert(std::make_pair(&Fn, CallSite::get(*I))); - } +bool DAE::runOnModule(Module &M) { + bool Changed = false; + + // First pass: Do a simple check to see if any functions can have their "..." + // removed. We can do this if they never call va_start. This loop cannot be + // fused with the next loop, because deleting a function invalidates + // information computed while surveying other functions. + DOUT << "DAE - Deleting dead varargs\n"; + for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { + Function &F = *I++; + if (F.getFunctionType()->isVarArg()) + Changed |= DeleteDeadVarargs(F); } - // Now we loop over all of the MaybeLive arguments, promoting them to be live - // arguments if one of the calls that uses the arguments to the calls they are - // passed into requires them to be live. Of course this could make other - // arguments live, so process callers recursively. - // - // Because elements can be removed from the MaybeLiveArguments list, copy it - // to a temporary vector. + // Second phase:loop through the module, determining which arguments are live. + // We assume all arguments are dead unless proven otherwise (allowing us to + // determine that dead arguments passed into recursive functions are dead). // - std::vector TmpArgList(MaybeLiveArguments.begin(), - MaybeLiveArguments.end()); - for (unsigned i = 0, e = TmpArgList.size(); i != e; ++i) { - Argument *MLA = TmpArgList[i]; - if (MaybeLiveArguments.count(MLA) && - isMaybeLiveArgumentNowAlive(MLA, LiveArguments)) { - MarkArgumentLive(MLA, MaybeLiveArguments, LiveArguments, CallSites); - } - } - - // Recover memory early... - CallSites.clear(); - - // At this point, we know that all arguments in DeadArguments and - // MaybeLiveArguments are dead. If the two sets are empty, there is nothing - // to do. - if (MaybeLiveArguments.empty() && DeadArguments.empty()) - return false; + DOUT << "DAE - Determining liveness\n"; + for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) + SurveyFunction(*I); - // Otherwise, compact into one set, and start eliminating the arguments from - // the functions. - DeadArguments.insert(MaybeLiveArguments.begin(), MaybeLiveArguments.end()); - MaybeLiveArguments.clear(); - - NumArgumentsEliminated += DeadArguments.size(); - while (!DeadArguments.empty()) - RemoveDeadArgumentsFromFunction((*DeadArguments.begin())->getParent(), - DeadArguments); - return true; + // Now, remove all dead arguments and return values from each function in + // turn + for (Module::iterator I = M.begin(), E = M.end(); I != E; ) { + // Increment now, because the function will probably get removed (ie + // replaced by a new one). + Function *F = I++; + Changed |= RemoveDeadStuffFromFunction(F); + } + return Changed; }