//===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
-//
+//
// The LLVM Compiler Infrastructure
//
// 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.
-//
+//
//===----------------------------------------------------------------------===//
//
-// Guarantees that all loops with identifiable, linear, induction variables will
-// be transformed to have a single, canonical, induction variable. After this
-// pass runs, it guarantees the the first PHI node of the header block in the
-// loop is the canonical induction variable if there is one.
+// This transformation analyzes and transforms the induction variables (and
+// computations derived from them) into simpler forms suitable for subsequent
+// analysis and transformation.
+//
+// This transformation makes 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/BasicBlock.h"
#include "llvm/Constants.h"
-#include "llvm/Type.h"
#include "llvm/Instructions.h"
-#include "llvm/Analysis/InductionVariable.h"
+#include "llvm/Type.h"
+#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Support/CFG.h"
-#include "llvm/Target/TargetData.h"
+#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Transforms/Utils/Local.h"
-#include "Support/Debug.h"
-#include "Support/Statistic.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/ADT/Statistic.h"
using namespace llvm;
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 *Loops;
- TargetData *TD;
+ LoopInfo *LI;
+ ScalarEvolution *SE;
+ bool Changed;
public:
virtual bool runOnFunction(Function &) {
- Loops = &getAnalysis<LoopInfo>();
- TD = &getAnalysis<TargetData>();
-
+ LI = &getAnalysis<LoopInfo>();
+ SE = &getAnalysis<ScalarEvolution>();
+ Changed = false;
+
// Induction Variables live in the header nodes of loops
- bool Changed = false;
- for (unsigned i = 0, e = Loops->getTopLevelLoops().size(); i != e; ++i)
- Changed |= runOnLoop(Loops->getTopLevelLoops()[i]);
+ for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
+ runOnLoop(*I);
return Changed;
}
- unsigned getTypeSize(const Type *Ty) {
- if (unsigned Size = Ty->getPrimitiveSize())
- return Size;
- return TD->getTypeSize(Ty); // Must be a pointer
- }
-
- Value *ComputeAuxIndVarValue(InductionVariable &IV, Value *CIV);
- void ReplaceIndVar(InductionVariable &IV, Value *Counter);
-
- bool runOnLoop(Loop *L);
-
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<TargetData>(); // Need pointer size
- AU.addRequired<LoopInfo>();
AU.addRequiredID(LoopSimplifyID);
+ AU.addRequired<ScalarEvolution>();
+ AU.addRequired<LoopInfo>();
AU.addPreservedID(LoopSimplifyID);
+ AU.addPreservedID(LCSSAID);
AU.setPreservesCFG();
}
+ private:
+ void runOnLoop(Loop *L);
+ void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader,
+ std::set<Instruction*> &DeadInsts);
+ Instruction *LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
+ SCEVExpander &RW);
+ void RewriteLoopExitValues(Loop *L);
+
+ void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
};
- RegisterOpt<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
+ RegisterPass<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
}
-Pass *llvm::createIndVarSimplifyPass() {
+FunctionPass *llvm::createIndVarSimplifyPass() {
return new IndVarSimplify();
}
+/// 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<Instruction*> &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<Instruction>(I->getOperand(i)))
+ Insts.insert(U);
+ SE->deleteInstructionFromRecords(I);
+ I->eraseFromParent();
+ Changed = true;
+ }
+ }
+}
-bool IndVarSimplify::runOnLoop(Loop *Loop) {
- // Transform all subloops before this loop...
- bool Changed = false;
- for (unsigned i = 0, e = Loop->getSubLoops().size(); i != e; ++i)
- Changed |= runOnLoop(Loop->getSubLoops()[i]);
-
- // 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->getHeader();
-
- // 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<InductionVariable> IndVars; // Induction variables for block
- BasicBlock::iterator AfterPHIIt = Header->begin();
- for (; PHINode *PN = dyn_cast<PHINode>(AfterPHIIt); ++AfterPHIIt)
- IndVars.push_back(InductionVariable(PN, Loops));
- // AfterPHIIt now points to first non-phi 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 canonical induction
- // variable, and checking to make sure they are not all unknown induction
- // variables. Keep track of the largest integer size of the induction
- // variable.
- //
- InductionVariable *Canonical = 0;
- unsigned MaxSize = 0;
-
- for (unsigned i = 0; i != IndVars.size(); ++i) {
- InductionVariable &IV = IndVars[i];
- if (IV.InductionType != InductionVariable::Unknown) {
- unsigned IVSize = getTypeSize(IV.Phi->getType());
+/// 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<Instruction*> &DeadInsts) {
+ assert(PN->getNumIncomingValues() == 2 && "Noncanonicalized loop!");
+ unsigned PreheaderIdx = PN->getBasicBlockIndex(Preheader);
+ unsigned BackedgeIdx = PreheaderIdx^1;
+ if (GetElementPtrInst *GEPI =
+ dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
+ if (GEPI->getOperand(0) == PN) {
+ assert(GEPI->getNumOperands() == 2 && "GEP types must match!");
+
+ // 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);
+
+ // 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::createAdd(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);
+
+ // If the incoming value is a constant expr GEP, try peeling out the array
+ // 0 index if possible to make things simpler.
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEPI->getOperand(0)))
+ if (CE->getOpcode() == Instruction::GetElementPtr) {
+ unsigned NumOps = CE->getNumOperands();
+ assert(NumOps > 1 && "CE folding didn't work!");
+ if (CE->getOperand(NumOps-1)->isNullValue()) {
+ // Check to make sure the last index really is an array index.
+ gep_type_iterator GTI = gep_type_begin(CE);
+ for (unsigned i = 1, e = CE->getNumOperands()-1;
+ i != e; ++i, ++GTI)
+ /*empty*/;
+ if (isa<SequentialType>(*GTI)) {
+ // Pull the last index out of the constant expr GEP.
+ std::vector<Value*> CEIdxs(CE->op_begin()+1, CE->op_end()-1);
+ Constant *NCE = ConstantExpr::getGetElementPtr(CE->getOperand(0),
+ CEIdxs);
+ GetElementPtrInst *NGEPI =
+ new GetElementPtrInst(NCE, Constant::getNullValue(Type::IntTy),
+ NewAdd, GEPI->getName(), GEPI);
+ GEPI->replaceAllUsesWith(NGEPI);
+ GEPI->eraseFromParent();
+ GEPI = NGEPI;
+ }
+ }
+ }
+
+
+ // 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<PHINode>(InsertPos)) ++InsertPos;
+ std::string Name = PN->getName(); PN->setName("");
+ Value *PreInc =
+ new GetElementPtrInst(PN->getIncomingValue(PreheaderIdx),
+ std::vector<Value*>(1, NewPhi), Name,
+ InsertPos);
+ PN->replaceAllUsesWith(PreInc);
+ }
- if (IV.InductionType == InductionVariable::Canonical &&
- !isa<PointerType>(IV.Phi->getType()) && IVSize >= MaxSize)
- Canonical = &IV;
-
- if (IVSize > MaxSize) MaxSize = IVSize;
+ // Delete the old PHI for sure, and the GEP if its otherwise unused.
+ DeadInsts.insert(PN);
- // If this variable is larger than the currently identified canonical
- // indvar, the canonical indvar is not usable.
- if (Canonical && IVSize > getTypeSize(Canonical->Phi->getType()))
- Canonical = 0;
+ ++NumPointer;
+ Changed = true;
}
- }
+}
- // No induction variables, bail early... don't add a canonical indvar
- if (MaxSize == 0) return Changed;
-
- // Okay, we want to convert other induction variables to use a canonical
- // indvar. If we don't have one, add one now...
- if (!Canonical) {
- // Create the PHI node for the new induction variable, and insert the phi
- // node at the start of the PHI nodes...
- const Type *IVType;
- switch (MaxSize) {
- default: assert(0 && "Unknown integer type size!");
- case 1: IVType = Type::UByteTy; break;
- case 2: IVType = Type::UShortTy; break;
- case 4: IVType = Type::UIntTy; break;
- case 8: IVType = Type::ULongTy; break;
- }
-
- PHINode *PN = new PHINode(IVType, "cann-indvar", Header->begin());
-
- // Create the increment instruction to add one to the counter...
- Instruction *Add = BinaryOperator::create(Instruction::Add, PN,
- ConstantUInt::get(IVType, 1),
- "next-indvar", AfterPHIIt);
-
- // 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++;
+/// LinearFunctionTestReplace - This method rewrites the exit condition of the
+/// loop to be a canonical != comparison against the incremented loop induction
+/// variable. This pass is able to rewrite the exit tests of any loop where the
+/// SCEV analysis can determine a loop-invariant trip count of the loop, which
+/// is actually a much broader range than just linear tests.
+///
+/// This method returns a "potentially dead" instruction whose computation chain
+/// should be deleted when convenient.
+Instruction *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
+ SCEV *IterationCount,
+ SCEVExpander &RW) {
+ // Find the exit block for the loop. We can currently only handle loops with
+ // a single exit.
+ std::vector<BasicBlock*> ExitBlocks;
+ L->getExitBlocks(ExitBlocks);
+ if (ExitBlocks.size() != 1) return 0;
+ BasicBlock *ExitBlock = ExitBlocks[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 0; // Multiple exits from loop to this block.
}
- assert(PI == pred_end(Header) && "Loop headers should have 2 preds!");
-
- // Add incoming values for the PHI node...
- PN->addIncoming(Constant::getNullValue(IVType), Incoming);
- PN->addIncoming(Add, BackEdgeBlock);
-
- // Analyze the new induction variable...
- IndVars.push_back(InductionVariable(PN, Loops));
- assert(IndVars.back().InductionType == InductionVariable::Canonical &&
- "Just inserted canonical indvar that is not canonical!");
- Canonical = &IndVars.back();
- ++NumInserted;
- Changed = true;
- } else {
- // If we have a canonical induction variable, make sure that it is the first
- // one in the basic block.
- if (&Header->front() != Canonical->Phi)
- Header->getInstList().splice(Header->begin(), Header->getInstList(),
- Canonical->Phi);
- }
+ assert(ExitingBlock && "Loop info is broken");
- DEBUG(std::cerr << "Induction variables:\n");
+ if (!isa<BranchInst>(ExitingBlock->getTerminator()))
+ return 0; // Can't rewrite non-branch yet
+ BranchInst *BI = cast<BranchInst>(ExitingBlock->getTerminator());
+ assert(BI->isConditional() && "Must be conditional to be part of loop!");
- // Get the current loop iteration count, which is always the value of the
- // canonical phi node...
- //
- PHINode *IterCount = Canonical->Phi;
+ Instruction *PotentiallyDeadInst = dyn_cast<Instruction>(BI->getCondition());
+
+ // If the exiting block is not the same as the backedge block, we must compare
+ // against the preincremented value, otherwise we prefer to compare against
+ // the post-incremented value.
+ BasicBlock *Header = L->getHeader();
+ pred_iterator HPI = pred_begin(Header);
+ assert(HPI != pred_end(Header) && "Loop with zero preds???");
+ if (!L->contains(*HPI)) ++HPI;
+ assert(HPI != pred_end(Header) && L->contains(*HPI) &&
+ "No backedge in loop?");
+
+ SCEVHandle TripCount = IterationCount;
+ Value *IndVar;
+ if (*HPI == ExitingBlock) {
+ // The IterationCount expression contains the number of times that the
+ // backedge actually branches to the loop header. This is one less than the
+ // number of times the loop executes, so add one to it.
+ Constant *OneC = ConstantInt::get(IterationCount->getType(), 1);
+ TripCount = SCEVAddExpr::get(IterationCount, SCEVUnknown::get(OneC));
+ IndVar = L->getCanonicalInductionVariableIncrement();
+ } else {
+ // We have to use the preincremented value...
+ IndVar = L->getCanonicalInductionVariable();
+ }
- // Loop through and replace all of the auxiliary induction variables with
- // references to the canonical induction variable...
- //
- for (unsigned i = 0; i != IndVars.size(); ++i) {
- InductionVariable *IV = &IndVars[i];
+ // Expand the code for the iteration count into the preheader of the loop.
+ BasicBlock *Preheader = L->getLoopPreheader();
+ Value *ExitCnt = RW.expandCodeFor(TripCount, 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;
+
+ Value *Cond = new SetCondInst(Opcode, IndVar, ExitCnt, "exitcond", BI);
+ BI->setCondition(Cond);
+ ++NumLFTR;
+ Changed = true;
+ return PotentiallyDeadInst;
+}
- DEBUG(IV->print(std::cerr));
- // Don't modify the canonical indvar or unrecognized indvars...
- if (IV != Canonical && IV->InductionType != InductionVariable::Unknown) {
- ReplaceIndVar(*IV, IterCount);
- Changed = true;
- ++NumRemoved;
+/// 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.
+ SCEVExpander 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;
+ std::vector<BasicBlock*> ExitBlocks;
+ L->getExitBlocks(ExitBlocks);
+ if (ExitBlocks.size() == 1)
+ BlockToInsertInto = ExitBlocks[0];
+ else
+ BlockToInsertInto = Preheader;
+ BasicBlock::iterator InsertPt = BlockToInsertInto->begin();
+ while (isa<PHINode>(InsertPt)) ++InsertPt;
+
+ bool HasConstantItCount = isa<SCEVConstant>(SE->getIterationCount(L));
+
+ std::set<Instruction*> 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;) {
+ if (I->getType()->isInteger()) { // Is an integer instruction
+ SCEVHandle SH = SE->getSCEV(I);
+ if (SH->hasComputableLoopEvolution(L) || // Varies predictably
+ HasConstantItCount) {
+ // Find out if this predictably varying value is actually used
+ // outside of the loop. "extra" as opposed to "intra".
+ std::vector<Instruction*> ExtraLoopUsers;
+ for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
+ UI != E; ++UI) {
+ Instruction *User = cast<Instruction>(*UI);
+ if (!L->contains(User->getParent())) {
+ // If this is a PHI node in the exit block and we're inserting,
+ // into the exit block, it must have a single entry. In this
+ // case, we can't insert the code after the PHI and have the PHI
+ // still use it. Instead, don't insert the the PHI.
+ if (PHINode *PN = dyn_cast<PHINode>(User)) {
+ // FIXME: This is a case where LCSSA pessimizes code, this
+ // should be fixed better.
+ if (PN->getNumOperands() == 2 &&
+ PN->getParent() == BlockToInsertInto)
+ continue;
+ }
+ ExtraLoopUsers.push_back(User);
+ }
+ }
+
+ 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<SCEVCouldNotCompute>(ExitValue)) {
+ Changed = true;
+ ++NumReplaced;
+ // Remember the next instruction. The rewriter can move code
+ // around in some cases.
+ BasicBlock::iterator NextI = I; ++NextI;
+
+ 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) {
+ PHINode* PN = dyn_cast<PHINode>(ExtraLoopUsers[i]);
+ if (PN && PN->getNumOperands() == 2 &&
+ !L->contains(PN->getParent())) {
+ // We're dealing with an LCSSA Phi. Handle it specially.
+ Instruction* LCSSAInsertPt = BlockToInsertInto->begin();
+
+ Instruction* NewInstr = dyn_cast<Instruction>(NewVal);
+ if (NewInstr && !isa<PHINode>(NewInstr) &&
+ !L->contains(NewInstr->getParent()))
+ for (unsigned j = 0; j < NewInstr->getNumOperands(); ++j){
+ Instruction* PredI =
+ dyn_cast<Instruction>(NewInstr->getOperand(j));
+ if (PredI && L->contains(PredI->getParent())) {
+ PHINode* NewLCSSA = new PHINode(PredI->getType(),
+ PredI->getName() + ".lcssa",
+ LCSSAInsertPt);
+ NewLCSSA->addIncoming(PredI,
+ BlockToInsertInto->getSinglePredecessor());
+
+ NewInstr->replaceUsesOfWith(PredI, NewLCSSA);
+ }
+ }
+
+ PN->replaceAllUsesWith(NewVal);
+ PN->eraseFromParent();
+ } else {
+ ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal);
+ }
+ }
+
+ // If this instruction is dead now, schedule it to be removed.
+ if (I->use_empty())
+ InstructionsToDelete.insert(I);
+ I = NextI;
+ continue; // Skip the ++I
+ }
+ }
+ }
+ }
+
+ // Next instruction. Continue instruction skips this.
+ ++I;
+ }
}
- }
- return Changed;
+ DeleteTriviallyDeadInstructions(InstructionsToDelete);
}
-/// ComputeAuxIndVarValue - Given an auxillary induction variable, compute and
-/// return a value which will always be equal to the induction variable PHI, but
-/// is based off of the canonical induction variable CIV.
-///
-Value *IndVarSimplify::ComputeAuxIndVarValue(InductionVariable &IV, Value *CIV){
- Instruction *Phi = IV.Phi;
- const Type *IVTy = Phi->getType();
- if (isa<PointerType>(IVTy)) // If indexing into a pointer, make the
- IVTy = TD->getIntPtrType(); // index the appropriate type.
-
- BasicBlock::iterator AfterPHIIt = Phi;
- while (isa<PHINode>(AfterPHIIt)) ++AfterPHIIt;
-
- Value *Val = CIV;
- if (Val->getType() != IVTy)
- Val = new CastInst(Val, IVTy, Val->getName(), AfterPHIIt);
-
- if (!isa<ConstantInt>(IV.Step) || // If the step != 1
- !cast<ConstantInt>(IV.Step)->equalsInt(1)) {
-
- // If the types are not compatible, insert a cast now...
- if (IV.Step->getType() != IVTy)
- IV.Step = new CastInst(IV.Step, IVTy, IV.Step->getName(), AfterPHIIt);
-
- Val = BinaryOperator::create(Instruction::Mul, Val, IV.Step,
- Phi->getName()+"-scale", AfterPHIIt);
- }
-
- // If this is a pointer indvar...
- if (isa<PointerType>(Phi->getType())) {
- std::vector<Value*> Idx;
- // FIXME: this should not be needed when we fix PR82!
- if (Val->getType() != Type::LongTy)
- Val = new CastInst(Val, Type::LongTy, Val->getName(), AfterPHIIt);
- Idx.push_back(Val);
- Val = new GetElementPtrInst(IV.Start, Idx,
- Phi->getName()+"-offset",
- AfterPHIIt);
-
- } else if (!isa<Constant>(IV.Start) || // If Start != 0...
- !cast<Constant>(IV.Start)->isNullValue()) {
- // If the types are not compatible, insert a cast now...
- if (IV.Start->getType() != IVTy)
- IV.Start = new CastInst(IV.Start, IVTy, IV.Start->getName(),
- AfterPHIIt);
-
- // Insert the instruction after the phi nodes...
- Val = BinaryOperator::create(Instruction::Add, Val, IV.Start,
- Phi->getName()+"-offset", AfterPHIIt);
+
+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<Instruction*> DeadInsts;
+ for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
+ PHINode *PN = cast<PHINode>(I);
+ if (isa<PointerType>(PN->getType()))
+ EliminatePointerRecurrence(PN, Preheader, DeadInsts);
}
-
- // If the PHI node has a different type than val is, insert a cast now...
- if (Val->getType() != Phi->getType())
- Val = new CastInst(Val, Phi->getType(), Val->getName(), AfterPHIIt);
- // Move the PHI name to it's new equivalent value...
- std::string OldName = Phi->getName();
- Phi->setName("");
- Val->setName(OldName);
+ if (!DeadInsts.empty())
+ DeleteTriviallyDeadInstructions(DeadInsts);
- return Val;
-}
+ // Next, transform all loops nesting inside of this loop.
+ for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
+ runOnLoop(*I);
-// ReplaceIndVar - Replace all uses of the specified induction variable with
-// expressions computed from the specified loop iteration counter variable.
-// Return true if instructions were deleted.
-void IndVarSimplify::ReplaceIndVar(InductionVariable &IV, Value *CIV) {
- Value *IndVarVal = 0;
- PHINode *Phi = IV.Phi;
-
- assert(Phi->getNumOperands() == 4 &&
- "Only expect induction variables in canonical loops!");
-
- // Remember the incoming values used by the PHI node
- std::vector<Value*> PHIOps;
- PHIOps.reserve(2);
- PHIOps.push_back(Phi->getIncomingValue(0));
- PHIOps.push_back(Phi->getIncomingValue(1));
-
- // Delete all of the operands of the PHI node... FIXME, this should be more
- // intelligent.
- Phi->dropAllReferences();
-
- // Now that we are rid of unneeded uses of the PHI node, replace any remaining
- // ones with the appropriate code using the canonical induction variable.
- while (!Phi->use_empty()) {
- Instruction *U = cast<Instruction>(Phi->use_back());
-
- // TODO: Perform LFTR here if possible
- if (0) {
-
- } else {
- // Replace all uses of the old PHI node with the new computed value...
- if (IndVarVal == 0)
- IndVarVal = ComputeAuxIndVarValue(IV, CIV);
- U->replaceUsesOfWith(Phi, IndVarVal);
+ // 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<SCEVCouldNotCompute>(IterationCount))
+ RewriteLoopExitValues(L);
+
+ // Next, analyze all of the induction variables in the loop, canonicalizing
+ // auxillary induction variables.
+ std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
+
+ for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
+ PHINode *PN = cast<PHINode>(I);
+ if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable!
+ SCEVHandle SCEV = SE->getSCEV(PN);
+ if (SCEV->hasComputableLoopEvolution(L))
+ // FIXME: It is an extremely bad idea to indvar substitute anything more
+ // complex than affine induction variables. Doing so will put expensive
+ // polynomial evaluations inside of the loop, and the str reduction pass
+ // currently can only reduce affine polynomials. For now just disable
+ // indvar subst on anything more complex than an affine addrec.
+ if (SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SCEV))
+ if (AR->isAffine())
+ IndVars.push_back(std::make_pair(PN, SCEV));
}
}
- // If the PHI is the last user of any instructions for computing PHI nodes
- // that are irrelevant now, delete those instructions.
- while (!PHIOps.empty()) {
- Instruction *MaybeDead = dyn_cast<Instruction>(PHIOps.back());
- PHIOps.pop_back();
-
- if (MaybeDead && isInstructionTriviallyDead(MaybeDead) &&
- (!isa<PHINode>(MaybeDead) ||
- MaybeDead->getParent() != Phi->getParent())) {
- PHIOps.insert(PHIOps.end(), MaybeDead->op_begin(),
- MaybeDead->op_end());
- MaybeDead->getParent()->getInstList().erase(MaybeDead);
-
- // Erase any duplicates entries in the PHIOps list.
- std::vector<Value*>::iterator It =
- std::find(PHIOps.begin(), PHIOps.end(), MaybeDead);
- while (It != PHIOps.end()) {
- PHIOps.erase(It);
- It = std::find(PHIOps.begin(), PHIOps.end(), MaybeDead);
+ // If there are no induction variables in the loop, there is nothing more to
+ // do.
+ if (IndVars.empty()) {
+ // Actually, if we know how many times the loop iterates, lets insert a
+ // canonical induction variable to help subsequent passes.
+ if (!isa<SCEVCouldNotCompute>(IterationCount)) {
+ SCEVExpander Rewriter(*SE, *LI);
+ Rewriter.getOrInsertCanonicalInductionVariable(L,
+ IterationCount->getType());
+ if (Instruction *I = LinearFunctionTestReplace(L, IterationCount,
+ Rewriter)) {
+ std::set<Instruction*> InstructionsToDelete;
+ InstructionsToDelete.insert(I);
+ DeleteTriviallyDeadInstructions(InstructionsToDelete);
}
}
+ return;
}
- // Delete the old, now unused, phi node...
- Phi->getParent()->getInstList().erase(Phi);
-}
+ // 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.
+ SCEVExpander 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.
+ LargestType = LargestType->getUnsignedVersion();
+ Value *IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
+ ++NumInserted;
+ Changed = true;
+
+ if (!isa<SCEVCouldNotCompute>(IterationCount))
+ if (Instruction *DI = LinearFunctionTestReplace(L, IterationCount,Rewriter))
+ DeadInsts.insert(DI);
+
+ // 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<PHINode>(InsertPt)) ++InsertPt;
+
+ // 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.
+ if (DifferingSizes) {
+ bool InsertedSizes[17] = { false };
+ InsertedSizes[LargestType->getPrimitiveSize()] = true;
+ for (unsigned i = 0, e = IndVars.size(); i != e; ++i)
+ if (!InsertedSizes[IndVars[i].first->getType()->getPrimitiveSize()]) {
+ PHINode *PN = IndVars[i].first;
+ InsertedSizes[PN->getType()->getPrimitiveSize()] = true;
+ Instruction *New = new CastInst(IndVar,
+ PN->getType()->getUnsignedVersion(),
+ "indvar", InsertPt);
+ Rewriter.addInsertedValue(New, SE->getSCEV(New));
+ }
+ }
+
+ // 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.
+ std::map<unsigned, Value*> InsertedSizes;
+ while (!IndVars.empty()) {
+ PHINode *PN = IndVars.back().first;
+ Value *NewVal = Rewriter.expandCodeFor(IndVars.back().second, InsertPt,
+ PN->getType());
+ std::string Name = PN->getName();
+ PN->setName("");
+ NewVal->setName(Name);
+
+ // Replace the old PHI Node with the inserted computation.
+ PN->replaceAllUsesWith(NewVal);
+ DeadInsts.insert(PN);
+ IndVars.pop_back();
+ ++NumRemoved;
+ Changed = true;
+ }
+#if 0
+ // Now replace all derived expressions in the loop body with simpler
+ // expressions.
+ 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
+ !I->use_empty() &&
+ !Rewriter.isInsertedInstruction(I)) {
+ SCEVHandle SH = SE->getSCEV(I);
+ Value *V = Rewriter.expandCodeFor(SH, I, I->getType());
+ if (V != I) {
+ if (isa<Instruction>(V)) {
+ std::string Name = I->getName();
+ I->setName("");
+ V->setName(Name);
+ }
+ I->replaceAllUsesWith(V);
+ DeadInsts.insert(I);
+ ++NumRemoved;
+ Changed = true;
+ }
+ }
+ }
+#endif
+
+ DeleteTriviallyDeadInstructions(DeadInsts);
+
+ if (mustPreserveAnalysisID(LCSSAID)) assert(L->isLCSSAForm());
+}