X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FIndVarSimplify.cpp;h=0622b170c8c6643d13468853345f6b2a91db1efe;hb=40bf8b48cdb9961898dba1bc67320be1e49e3da1;hp=35fe697f0f9c63feccf6c3732c636343e0e671a3;hpb=f629309f74cf1a64aa7fd1cd5784fd7db9a8f59e;p=oota-llvm.git diff --git a/lib/Transforms/Scalar/IndVarSimplify.cpp b/lib/Transforms/Scalar/IndVarSimplify.cpp index 35fe697f0f9..0622b170c8c 100644 --- a/lib/Transforms/Scalar/IndVarSimplify.cpp +++ b/lib/Transforms/Scalar/IndVarSimplify.cpp @@ -1,207 +1,420 @@ //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===// +// +// The LLVM Compiler Infrastructure // -// InductionVariableSimplify - Transform induction variables in a program -// to all use a single cannonical induction variable per loop. +// This file was developed by the LLVM research group and is distributed under +// the University of Illinois Open Source License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This transformation analyzes and transforms the induction variables (and +// computations derived from them) into simpler forms suitable for subsequent +// analysis and transformation. +// +// This transformation make the following changes to each loop with an +// identifiable induction variable: +// 1. All loops are transformed to have a SINGLE canonical induction variable +// which starts at zero and steps by one. +// 2. The canonical induction variable is guaranteed to be the first PHI node +// in the loop header block. +// 3. Any pointer arithmetic recurrences are raised to use array subscripts. +// +// If the trip count of a loop is computable, this pass also makes the following +// changes: +// 1. The exit condition for the loop is canonicalized to compare the +// induction value against the exit value. This turns loops like: +// 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)' +// 2. Any use outside of the loop of an expression derived from the indvar +// is changed to compute the derived value outside of the loop, eliminating +// the dependence on the exit value of the induction variable. If the only +// purpose of the loop is to compute the exit value of some derived +// expression, this transformation will make the loop dead. +// +// This transformation should be followed by strength reduction after all of the +// desired loop transformations have been performed. Additionally, on targets +// where it is profitable, the loop could be transformed to count down to zero +// (the "do loop" optimization). // //===----------------------------------------------------------------------===// #include "llvm/Transforms/Scalar.h" -#include "llvm/Analysis/InductionVariable.h" -#include "llvm/Analysis/LoopInfo.h" -#include "llvm/Analysis/Writer.h" -#include "llvm/iPHINode.h" -#include "llvm/iOther.h" +#include "llvm/BasicBlock.h" +#include "llvm/Constant.h" +#include "llvm/Instructions.h" #include "llvm/Type.h" -#include "llvm/Constants.h" +#include "llvm/Analysis/ScalarEvolution.h" +#include "llvm/Analysis/LoopInfo.h" #include "llvm/Support/CFG.h" -#include "Support/STLExtras.h" -#include "Support/StatisticReporter.h" +#include "llvm/Transforms/Utils/Local.h" +#include "Support/CommandLine.h" +#include "Support/Statistic.h" +using namespace llvm; -static Statistic<> NumRemoved ("indvars\t\t- Number of aux indvars removed"); -static Statistic<> NumInserted("indvars\t\t- Number of cannonical indvars added"); +namespace { + Statistic<> NumRemoved ("indvars", "Number of aux indvars removed"); + Statistic<> NumPointer ("indvars", "Number of pointer indvars promoted"); + Statistic<> NumInserted("indvars", "Number of canonical indvars added"); + Statistic<> NumReplaced("indvars", "Number of exit values replaced"); + Statistic<> NumLFTR ("indvars", "Number of loop exit tests replaced"); + class IndVarSimplify : public FunctionPass { + LoopInfo *LI; + ScalarEvolution *SE; + bool Changed; + public: + virtual bool runOnFunction(Function &) { + LI = &getAnalysis(); + SE = &getAnalysis(); + Changed = false; -// InsertCast - Cast Val to Ty, setting a useful name on the cast if Val has a -// name... -// -static Instruction *InsertCast(Instruction *Val, const Type *Ty, - BasicBlock::iterator It) { - Instruction *Cast = new CastInst(Val, Ty); - if (Val->hasName()) Cast->setName(Val->getName()+"-casted"); - Val->getParent()->getInstList().insert(It, Cast); - return Cast; + // Induction Variables live in the header nodes of loops + for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) + runOnLoop(*I); + return Changed; + } + + virtual void getAnalysisUsage(AnalysisUsage &AU) const { + AU.addRequiredID(LoopSimplifyID); + AU.addRequired(); + AU.addRequired(); + AU.addPreservedID(LoopSimplifyID); + AU.setPreservesCFG(); + } + private: + void runOnLoop(Loop *L); + void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader, + std::set &DeadInsts); + void LinearFunctionTestReplace(Loop *L, SCEV *IterationCount, + Value *IndVar, ScalarEvolutionRewriter &RW); + void RewriteLoopExitValues(Loop *L); + + void DeleteTriviallyDeadInstructions(std::set &Insts); + }; + RegisterOpt X("indvars", "Canonicalize Induction Variables"); } -static bool TransformLoop(LoopInfo *Loops, Loop *Loop) { - // Transform all subloops before this loop... - bool Changed = reduce_apply_bool(Loop->getSubLoops().begin(), - Loop->getSubLoops().end(), - std::bind1st(std::ptr_fun(TransformLoop), Loops)); - // Get the header node for this loop. All of the phi nodes that could be - // induction variables must live in this basic block. - // - BasicBlock *Header = Loop->getBlocks().front(); - - // Loop over all of the PHI nodes in the basic block, calculating the - // induction variables that they represent... stuffing the induction variable - // info into a vector... - // - std::vector IndVars; // Induction variables for block - BasicBlock::iterator AfterPHIIt = Header->begin(); - for (; PHINode *PN = dyn_cast(&*AfterPHIIt); ++AfterPHIIt) - IndVars.push_back(InductionVariable(PN, Loops)); - // AfterPHIIt now points to first nonphi instruction... - - // If there are no phi nodes in this basic block, there can't be indvars... - if (IndVars.empty()) return Changed; - - // Loop over the induction variables, looking for a cannonical induction - // variable, and checking to make sure they are not all unknown induction - // variables. - // - bool FoundIndVars = false; - InductionVariable *Cannonical = 0; - for (unsigned i = 0; i < IndVars.size(); ++i) { - if (IndVars[i].InductionType == InductionVariable::Cannonical) - Cannonical = &IndVars[i]; - if (IndVars[i].InductionType != InductionVariable::Unknown) - FoundIndVars = true; - } +Pass *llvm::createIndVarSimplifyPass() { + return new IndVarSimplify(); +} - // No induction variables, bail early... don't add a cannonnical indvar - if (!FoundIndVars) return Changed; - - // Okay, we want to convert other induction variables to use a cannonical - // indvar. If we don't have one, add one now... - if (!Cannonical) { - // Create the PHI node for the new induction variable - PHINode *PN = new PHINode(Type::UIntTy, "cann-indvar"); - - // Insert the phi node at the end of the other phi nodes... - AfterPHIIt = ++Header->getInstList().insert(AfterPHIIt, PN); - - // Create the increment instruction to add one to the counter... - Instruction *Add = BinaryOperator::create(Instruction::Add, PN, - ConstantUInt::get(Type::UIntTy,1), - "add1-indvar"); - - // Insert the add instruction after all of the PHI nodes... - Header->getInstList().insert(AfterPHIIt, Add); - - // Figure out which block is incoming and which is the backedge for the loop - BasicBlock *Incoming, *BackEdgeBlock; - pred_iterator PI = pred_begin(Header); - assert(PI != pred_end(Header) && "Loop headers should have 2 preds!"); - if (Loop->contains(*PI)) { // First pred is back edge... - BackEdgeBlock = *PI++; - Incoming = *PI++; - } else { - Incoming = *PI++; - BackEdgeBlock = *PI++; + +/// DeleteTriviallyDeadInstructions - If any of the instructions is the +/// specified set are trivially dead, delete them and see if this makes any of +/// their operands subsequently dead. +void IndVarSimplify:: +DeleteTriviallyDeadInstructions(std::set &Insts) { + while (!Insts.empty()) { + Instruction *I = *Insts.begin(); + Insts.erase(Insts.begin()); + if (isInstructionTriviallyDead(I)) { + for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) + if (Instruction *U = dyn_cast(I->getOperand(i))) + Insts.insert(U); + SE->deleteInstructionFromRecords(I); + I->getParent()->getInstList().erase(I); + Changed = true; } - assert(PI == pred_end(Header) && "Loop headers should have 2 preds!"); - - // Add incoming values for the PHI node... - PN->addIncoming(Constant::getNullValue(Type::UIntTy), Incoming); - PN->addIncoming(Add, BackEdgeBlock); - - // Analyze the new induction variable... - IndVars.push_back(InductionVariable(PN, Loops)); - assert(IndVars.back().InductionType == InductionVariable::Cannonical && - "Just inserted cannonical indvar that is not cannonical!"); - Cannonical = &IndVars.back(); - ++NumInserted; - Changed = true; } +} - DEBUG(std::cerr << "Induction variables:\n"); - // Get the current loop iteration count, which is always the value of the - // cannonical phi node... - // - PHINode *IterCount = Cannonical->Phi; +/// EliminatePointerRecurrence - Check to see if this is a trivial GEP pointer +/// recurrence. If so, change it into an integer recurrence, permitting +/// analysis by the SCEV routines. +void IndVarSimplify::EliminatePointerRecurrence(PHINode *PN, + BasicBlock *Preheader, + std::set &DeadInsts) { + assert(PN->getNumIncomingValues() == 2 && "Noncanonicalized loop!"); + unsigned PreheaderIdx = PN->getBasicBlockIndex(Preheader); + unsigned BackedgeIdx = PreheaderIdx^1; + if (GetElementPtrInst *GEPI = + dyn_cast(PN->getIncomingValue(BackedgeIdx))) + if (GEPI->getOperand(0) == PN) { + assert(GEPI->getNumOperands() == 2 && "GEP types must mismatch!"); + + // Okay, we found a pointer recurrence. Transform this pointer + // recurrence into an integer recurrence. Compute the value that gets + // added to the pointer at every iteration. + Value *AddedVal = GEPI->getOperand(1); - // Loop through and replace all of the auxillary induction variables with - // references to the primary induction variable... - // - for (unsigned i = 0; i < IndVars.size(); ++i) { - InductionVariable *IV = &IndVars[i]; - - DEBUG(std::cerr << IV); - - // Don't modify the cannonical indvar or unrecognized indvars... - if (IV != Cannonical && IV->InductionType != InductionVariable::Unknown) { - Instruction *Val = IterCount; - if (!isa(IV->Step) || // If the step != 1 - !cast(IV->Step)->equalsInt(1)) { - std::string Name; // Create a scale by the step value... - if (IV->Phi->hasName()) Name = IV->Phi->getName()+"-scale"; - - // If the types are not compatible, insert a cast now... - if (Val->getType() != IV->Step->getType()) - Val = InsertCast(Val, IV->Step->getType(), AfterPHIIt); - - Val = BinaryOperator::create(Instruction::Mul, Val, IV->Step, Name); - // Insert the phi node at the end of the other phi nodes... - Header->getInstList().insert(AfterPHIIt, Val); + // Insert a new integer PHI node into the top of the block. + PHINode *NewPhi = new PHINode(AddedVal->getType(), + PN->getName()+".rec", PN); + NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()), + Preheader); + // Create the new add instruction. + Value *NewAdd = BinaryOperator::create(Instruction::Add, NewPhi, + AddedVal, + GEPI->getName()+".rec", GEPI); + NewPhi->addIncoming(NewAdd, PN->getIncomingBlock(BackedgeIdx)); + + // Update the existing GEP to use the recurrence. + GEPI->setOperand(0, PN->getIncomingValue(PreheaderIdx)); + + // Update the GEP to use the new recurrence we just inserted. + GEPI->setOperand(1, NewAdd); + + // Finally, if there are any other users of the PHI node, we must + // insert a new GEP instruction that uses the pre-incremented version + // of the induction amount. + if (!PN->use_empty()) { + BasicBlock::iterator InsertPos = PN; ++InsertPos; + while (isa(InsertPos)) ++InsertPos; + std::string Name = PN->getName(); PN->setName(""); + Value *PreInc = + new GetElementPtrInst(PN->getIncomingValue(PreheaderIdx), + std::vector(1, NewPhi), Name, + InsertPos); + PN->replaceAllUsesWith(PreInc); } - if (!isa(IV->Start) || // If the start != 0 - !cast(IV->Start)->isNullValue()) { - std::string Name; // Create a offset by the start value... - if (IV->Phi->hasName()) Name = IV->Phi->getName()+"-offset"; + // Delete the old PHI for sure, and the GEP if its otherwise unused. + DeadInsts.insert(PN); - // If the types are not compatible, insert a cast now... - if (Val->getType() != IV->Start->getType()) - Val = InsertCast(Val, IV->Start->getType(), AfterPHIIt); + ++NumPointer; + Changed = true; + } +} - Val = BinaryOperator::create(Instruction::Add, Val, IV->Start, Name); - // Insert the phi node at the end of the other phi nodes... - Header->getInstList().insert(AfterPHIIt, Val); - } +/// LinearFunctionTestReplace - This method rewrites the exit condition of the +/// loop to be a canonical != comparison against the loop induction variable. +/// This pass is able to rewrite the exit tests of any loop where the SCEV +/// analysis can determine the trip count of the loop, which is actually a much +/// broader range than just linear tests. +void IndVarSimplify::LinearFunctionTestReplace(Loop *L, SCEV *IterationCount, + Value *IndVar, + ScalarEvolutionRewriter &RW) { + // Find the exit block for the loop. We can currently only handle loops with + // a single exit. + if (L->getExitBlocks().size() != 1) return; + BasicBlock *ExitBlock = L->getExitBlocks()[0]; + + // Make sure there is only one predecessor block in the loop. + BasicBlock *ExitingBlock = 0; + for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock); + PI != PE; ++PI) + if (L->contains(*PI)) { + if (ExitingBlock == 0) + ExitingBlock = *PI; + else + return; // Multiple exits from loop to this block. + } + assert(ExitingBlock && "Loop info is broken"); - // If the PHI node has a different type than val is, insert a cast now... - if (Val->getType() != IV->Phi->getType()) - Val = InsertCast(Val, IV->Phi->getType(), AfterPHIIt); - - // Replace all uses of the old PHI node with the new computed value... - IV->Phi->replaceAllUsesWith(Val); + if (!isa(ExitingBlock->getTerminator())) + return; // Can't rewrite non-branch yet + BranchInst *BI = cast(ExitingBlock->getTerminator()); + assert(BI->isConditional() && "Must be conditional to be part of loop!"); - // Move the PHI name to it's new equivalent value... - std::string OldName = IV->Phi->getName(); - IV->Phi->setName(""); - Val->setName(OldName); + std::set InstructionsToDelete; + if (Instruction *Cond = dyn_cast(BI->getCondition())) + InstructionsToDelete.insert(Cond); - // Delete the old, now unused, phi node... - Header->getInstList().erase(IV->Phi); - Changed = true; - ++NumRemoved; - } - } + // Expand the code for the iteration count into the preheader of the loop. + BasicBlock *Preheader = L->getLoopPreheader(); + Value *ExitCnt = RW.ExpandCodeFor(IterationCount, Preheader->getTerminator(), + IndVar->getType()); + + // Insert a new setne or seteq instruction before the branch. + Instruction::BinaryOps Opcode; + if (L->contains(BI->getSuccessor(0))) + Opcode = Instruction::SetNE; + else + Opcode = Instruction::SetEQ; - return Changed; + Value *Cond = new SetCondInst(Opcode, IndVar, ExitCnt, "exitcond", BI); + BI->setCondition(Cond); + ++NumLFTR; + Changed = true; + + DeleteTriviallyDeadInstructions(InstructionsToDelete); } -namespace { - struct InductionVariableSimplify : public FunctionPass { - virtual bool runOnFunction(Function &) { - LoopInfo &LI = getAnalysis(); - // Induction Variables live in the header nodes of loops - return reduce_apply_bool(LI.getTopLevelLoops().begin(), - LI.getTopLevelLoops().end(), - std::bind1st(std::ptr_fun(TransformLoop), &LI)); - } - - virtual void getAnalysisUsage(AnalysisUsage &AU) const { - AU.addRequired(LoopInfo::ID); - AU.preservesCFG(); +/// RewriteLoopExitValues - Check to see if this loop has a computable +/// loop-invariant execution count. If so, this means that we can compute the +/// final value of any expressions that are recurrent in the loop, and +/// substitute the exit values from the loop into any instructions outside of +/// the loop that use the final values of the current expressions. +void IndVarSimplify::RewriteLoopExitValues(Loop *L) { + BasicBlock *Preheader = L->getLoopPreheader(); + + // Scan all of the instructions in the loop, looking at those that have + // extra-loop users and which are recurrences. + ScalarEvolutionRewriter Rewriter(*SE, *LI); + + // We insert the code into the preheader of the loop if the loop contains + // multiple exit blocks, or in the exit block if there is exactly one. + BasicBlock *BlockToInsertInto; + if (L->getExitBlocks().size() == 1) + BlockToInsertInto = L->getExitBlocks()[0]; + else + BlockToInsertInto = Preheader; + BasicBlock::iterator InsertPt = BlockToInsertInto->begin(); + while (isa(InsertPt)) ++InsertPt; + + std::set InstructionsToDelete; + + for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) + if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop... + BasicBlock *BB = L->getBlocks()[i]; + for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) + if (I->getType()->isInteger()) { // Is an integer instruction + SCEVHandle SH = SE->getSCEV(I); + if (SH->hasComputableLoopEvolution(L)) { // Varies predictably + // Find out if this predictably varying value is actually used + // outside of the loop. "extra" as opposed to "intra". + std::vector ExtraLoopUsers; + for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); + UI != E; ++UI) + if (!L->contains(cast(*UI)->getParent())) + ExtraLoopUsers.push_back(*UI); + if (!ExtraLoopUsers.empty()) { + // Okay, this instruction has a user outside of the current loop + // and varies predictably in this loop. Evaluate the value it + // contains when the loop exits, and insert code for it. + SCEVHandle ExitValue = SE->getSCEVAtScope(I,L->getParentLoop()); + if (!isa(ExitValue)) { + Changed = true; + ++NumReplaced; + Value *NewVal = Rewriter.ExpandCodeFor(ExitValue, InsertPt, + I->getType()); + + // Rewrite any users of the computed value outside of the loop + // with the newly computed value. + for (unsigned i = 0, e = ExtraLoopUsers.size(); i != e; ++i) + ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal); + + // If this instruction is dead now, schedule it to be removed. + if (I->use_empty()) + InstructionsToDelete.insert(I); + } + } + } + } } - }; - RegisterPass X("indvars", - "Cannonicalize Induction Variables"); + + DeleteTriviallyDeadInstructions(InstructionsToDelete); } -Pass *createIndVarSimplifyPass() { - return new InductionVariableSimplify(); + +void IndVarSimplify::runOnLoop(Loop *L) { + // First step. Check to see if there are any trivial GEP pointer recurrences. + // If there are, change them into integer recurrences, permitting analysis by + // the SCEV routines. + // + BasicBlock *Header = L->getHeader(); + BasicBlock *Preheader = L->getLoopPreheader(); + + std::set DeadInsts; + for (BasicBlock::iterator I = Header->begin(); + PHINode *PN = dyn_cast(I); ++I) + if (isa(PN->getType())) + EliminatePointerRecurrence(PN, Preheader, DeadInsts); + + if (!DeadInsts.empty()) + DeleteTriviallyDeadInstructions(DeadInsts); + + + // Next, transform all loops nesting inside of this loop. + for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I) + runOnLoop(*I); + + // Check to see if this loop has a computable loop-invariant execution count. + // If so, this means that we can compute the final value of any expressions + // that are recurrent in the loop, and substitute the exit values from the + // loop into any instructions outside of the loop that use the final values of + // the current expressions. + // + SCEVHandle IterationCount = SE->getIterationCount(L); + if (!isa(IterationCount)) + RewriteLoopExitValues(L); + + // Next, analyze all of the induction variables in the loop, canonicalizing + // auxillary induction variables. + std::vector > IndVars; + + for (BasicBlock::iterator I = Header->begin(); + PHINode *PN = dyn_cast(I); ++I) + if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable! + SCEVHandle SCEV = SE->getSCEV(PN); + if (SCEV->hasComputableLoopEvolution(L)) + if (SE->shouldSubstituteIndVar(SCEV)) // HACK! + IndVars.push_back(std::make_pair(PN, SCEV)); + } + + // If there are no induction variables in the loop, there is nothing more to + // do. + if (IndVars.empty()) return; + + // Compute the type of the largest recurrence expression. + // + const Type *LargestType = IndVars[0].first->getType(); + bool DifferingSizes = false; + for (unsigned i = 1, e = IndVars.size(); i != e; ++i) { + const Type *Ty = IndVars[i].first->getType(); + DifferingSizes |= Ty->getPrimitiveSize() != LargestType->getPrimitiveSize(); + if (Ty->getPrimitiveSize() > LargestType->getPrimitiveSize()) + LargestType = Ty; + } + + // Create a rewriter object which we'll use to transform the code with. + ScalarEvolutionRewriter Rewriter(*SE, *LI); + + // Now that we know the largest of of the induction variables in this loop, + // insert a canonical induction variable of the largest size. + Value *IndVar = Rewriter.GetOrInsertCanonicalInductionVariable(L,LargestType); + ++NumInserted; + Changed = true; + + if (!isa(IterationCount)) + LinearFunctionTestReplace(L, IterationCount, IndVar, Rewriter); + +#if 0 + // If there were induction variables of other sizes, cast the primary + // induction variable to the right size for them, avoiding the need for the + // code evaluation methods to insert induction variables of different sizes. + // FIXME! + if (DifferingSizes) { + std::map InsertedSizes; + for (unsigned i = 0, e = IndVars.size(); i != e; ++i) { + } + } +#endif + + // Now that we have a canonical induction variable, we can rewrite any + // recurrences in terms of the induction variable. Start with the auxillary + // induction variables, and recursively rewrite any of their uses. + BasicBlock::iterator InsertPt = Header->begin(); + while (isa(InsertPt)) ++InsertPt; + + while (!IndVars.empty()) { + PHINode *PN = IndVars.back().first; + Value *NewVal = Rewriter.ExpandCodeFor(IndVars.back().second, InsertPt, + PN->getType()); + // Replace the old PHI Node with the inserted computation. + PN->replaceAllUsesWith(NewVal); + DeadInsts.insert(PN); + IndVars.pop_back(); + ++NumRemoved; + Changed = true; + } + + DeleteTriviallyDeadInstructions(DeadInsts); + + // TODO: In the future we could replace all instructions in the loop body with + // simpler expressions. It's not clear how useful this would be though or if + // the code expansion cost would be worth it! We probably shouldn't do this + // until we have a way to reuse expressions already in the code. +#if 0 + for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i) + if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop... + BasicBlock *BB = L->getBlocks()[i]; + for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) + if (I->getType()->isInteger() && // Is an integer instruction + !Rewriter.isInsertedInstruction(I)) { + SCEVHandle SH = SE->getSCEV(I); + } + } +#endif }