//===- 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.
-//
+// This file 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
+// 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.
//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "indvars"
#include "llvm/Transforms/Scalar.h"
#include "llvm/BasicBlock.h"
#include "llvm/Constants.h"
#include "llvm/Instructions.h"
#include "llvm/Type.h"
-#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopPass.h"
#include "llvm/Support/CFG.h"
+#include "llvm/Support/Compiler.h"
+#include "llvm/Support/Debug.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Support/CommandLine.h"
+#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
using namespace llvm;
-namespace {
- /// SCEVExpander - This class uses information about analyze scalars to
- /// rewrite expressions in canonical form.
- ///
- /// Clients should create an instance of this class when rewriting is needed,
- /// and destroying it when finished to allow the release of the associated
- /// memory.
- struct SCEVExpander : public SCEVVisitor<SCEVExpander, Value*> {
- ScalarEvolution &SE;
- LoopInfo &LI;
- std::map<SCEVHandle, Value*> InsertedExpressions;
- std::set<Instruction*> InsertedInstructions;
-
- Instruction *InsertPt;
-
- friend class SCEVVisitor<SCEVExpander, Value*>;
- public:
- SCEVExpander(ScalarEvolution &se, LoopInfo &li) : SE(se), LI(li) {}
-
- /// isInsertedInstruction - Return true if the specified instruction was
- /// inserted by the code rewriter. If so, the client should not modify the
- /// instruction.
- bool isInsertedInstruction(Instruction *I) const {
- return InsertedInstructions.count(I);
- }
-
- /// getOrInsertCanonicalInductionVariable - This method returns the
- /// canonical induction variable of the specified type for the specified
- /// loop (inserting one if there is none). A canonical induction variable
- /// starts at zero and steps by one on each iteration.
- Value *getOrInsertCanonicalInductionVariable(const Loop *L, const Type *Ty){
- assert((Ty->isInteger() || Ty->isFloatingPoint()) &&
- "Can only insert integer or floating point induction variables!");
- SCEVHandle H = SCEVAddRecExpr::get(SCEVUnknown::getIntegerSCEV(0, Ty),
- SCEVUnknown::getIntegerSCEV(1, Ty), L);
- return expand(H);
- }
-
- /// addInsertedValue - Remember the specified instruction as being the
- /// canonical form for the specified SCEV.
- void addInsertedValue(Instruction *I, SCEV *S) {
- InsertedExpressions[S] = (Value*)I;
- InsertedInstructions.insert(I);
- }
-
- /// expandCodeFor - Insert code to directly compute the specified SCEV
- /// expression into the program. The inserted code is inserted into the
- /// specified block.
- ///
- /// If a particular value sign is required, a type may be specified for the
- /// result.
- Value *expandCodeFor(SCEVHandle SH, Instruction *IP, const Type *Ty = 0) {
- // Expand the code for this SCEV.
- this->InsertPt = IP;
- return expandInTy(SH, Ty);
- }
-
- protected:
- Value *expand(SCEV *S) {
- // Check to see if we already expanded this.
- std::map<SCEVHandle, Value*>::iterator I = InsertedExpressions.find(S);
- if (I != InsertedExpressions.end())
- return I->second;
-
- Value *V = visit(S);
- InsertedExpressions[S] = V;
- return V;
- }
-
- Value *expandInTy(SCEV *S, const Type *Ty) {
- Value *V = expand(S);
- if (Ty && V->getType() != Ty) {
- // FIXME: keep track of the cast instruction.
- if (Constant *C = dyn_cast<Constant>(V))
- return ConstantExpr::getCast(C, Ty);
- else if (Instruction *I = dyn_cast<Instruction>(V)) {
- // Check to see if there is already a cast. If there is, use it.
- for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
- UI != E; ++UI) {
- if ((*UI)->getType() == Ty)
- if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI))) {
- BasicBlock::iterator It = I; ++It;
- while (isa<PHINode>(It)) ++It;
- if (It != BasicBlock::iterator(CI)) {
- // Splice the cast immediately after the operand in question.
- BasicBlock::InstListType &InstList =
- I->getParent()->getInstList();
- InstList.splice(It, CI->getParent()->getInstList(), CI);
- }
- return CI;
- }
- }
- BasicBlock::iterator IP = I; ++IP;
- if (InvokeInst *II = dyn_cast<InvokeInst>(I))
- IP = II->getNormalDest()->begin();
- while (isa<PHINode>(IP)) ++IP;
- return new CastInst(V, Ty, V->getName(), IP);
- } else {
- // FIXME: check to see if there is already a cast!
- return new CastInst(V, Ty, V->getName(), InsertPt);
- }
- }
- return V;
- }
-
- Value *visitConstant(SCEVConstant *S) {
- return S->getValue();
- }
-
- Value *visitTruncateExpr(SCEVTruncateExpr *S) {
- Value *V = expand(S->getOperand());
- return new CastInst(V, S->getType(), "tmp.", InsertPt);
- }
-
- Value *visitZeroExtendExpr(SCEVZeroExtendExpr *S) {
- Value *V = expandInTy(S->getOperand(),S->getType()->getUnsignedVersion());
- return new CastInst(V, S->getType(), "tmp.", InsertPt);
- }
-
- Value *visitAddExpr(SCEVAddExpr *S) {
- const Type *Ty = S->getType();
- Value *V = expandInTy(S->getOperand(S->getNumOperands()-1), Ty);
-
- // Emit a bunch of add instructions
- for (int i = S->getNumOperands()-2; i >= 0; --i)
- V = BinaryOperator::createAdd(V, expandInTy(S->getOperand(i), Ty),
- "tmp.", InsertPt);
- return V;
- }
-
- Value *visitMulExpr(SCEVMulExpr *S);
-
- Value *visitUDivExpr(SCEVUDivExpr *S) {
- const Type *Ty = S->getType();
- Value *LHS = expandInTy(S->getLHS(), Ty);
- Value *RHS = expandInTy(S->getRHS(), Ty);
- return BinaryOperator::createDiv(LHS, RHS, "tmp.", InsertPt);
- }
-
- Value *visitAddRecExpr(SCEVAddRecExpr *S);
-
- Value *visitUnknown(SCEVUnknown *S) {
- return S->getValue();
- }
- };
-}
-
-Value *SCEVExpander::visitMulExpr(SCEVMulExpr *S) {
- const Type *Ty = S->getType();
- int FirstOp = 0; // Set if we should emit a subtract.
- if (SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
- if (SC->getValue()->isAllOnesValue())
- FirstOp = 1;
-
- int i = S->getNumOperands()-2;
- Value *V = expandInTy(S->getOperand(i+1), Ty);
-
- // Emit a bunch of multiply instructions
- for (; i >= FirstOp; --i)
- V = BinaryOperator::createMul(V, expandInTy(S->getOperand(i), Ty),
- "tmp.", InsertPt);
- // -1 * ... ---> 0 - ...
- if (FirstOp == 1)
- V = BinaryOperator::createNeg(V, "tmp.", InsertPt);
- return V;
-}
-
-Value *SCEVExpander::visitAddRecExpr(SCEVAddRecExpr *S) {
- const Type *Ty = S->getType();
- const Loop *L = S->getLoop();
- // We cannot yet do fp recurrences, e.g. the xform of {X,+,F} --> X+{0,+,F}
- assert(Ty->isIntegral() && "Cannot expand fp recurrences yet!");
-
- // {X,+,F} --> X + {0,+,F}
- if (!isa<SCEVConstant>(S->getStart()) ||
- !cast<SCEVConstant>(S->getStart())->getValue()->isNullValue()) {
- Value *Start = expandInTy(S->getStart(), Ty);
- std::vector<SCEVHandle> NewOps(S->op_begin(), S->op_end());
- NewOps[0] = SCEVUnknown::getIntegerSCEV(0, Ty);
- Value *Rest = expandInTy(SCEVAddRecExpr::get(NewOps, L), Ty);
-
- // FIXME: look for an existing add to use.
- return BinaryOperator::createAdd(Rest, Start, "tmp.", InsertPt);
- }
-
- // {0,+,1} --> Insert a canonical induction variable into the loop!
- if (S->getNumOperands() == 2 &&
- S->getOperand(1) == SCEVUnknown::getIntegerSCEV(1, Ty)) {
- // Create and insert the PHI node for the induction variable in the
- // specified loop.
- BasicBlock *Header = L->getHeader();
- PHINode *PN = new PHINode(Ty, "indvar", Header->begin());
- PN->addIncoming(Constant::getNullValue(Ty), L->getLoopPreheader());
-
- 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?");
-
- // Insert a unit add instruction right before the terminator corresponding
- // to the back-edge.
- Constant *One = Ty->isFloatingPoint() ? (Constant*)ConstantFP::get(Ty, 1.0)
- : ConstantInt::get(Ty, 1);
- Instruction *Add = BinaryOperator::createAdd(PN, One, "indvar.next",
- (*HPI)->getTerminator());
-
- pred_iterator PI = pred_begin(Header);
- if (*PI == L->getLoopPreheader())
- ++PI;
- PN->addIncoming(Add, *PI);
- return PN;
- }
-
- // Get the canonical induction variable I for this loop.
- Value *I = getOrInsertCanonicalInductionVariable(L, Ty);
-
- if (S->getNumOperands() == 2) { // {0,+,F} --> i*F
- Value *F = expandInTy(S->getOperand(1), Ty);
- return BinaryOperator::createMul(I, F, "tmp.", InsertPt);
- }
-
- // If this is a chain of recurrences, turn it into a closed form, using the
- // folders, then expandCodeFor the closed form. This allows the folders to
- // simplify the expression without having to build a bunch of special code
- // into this folder.
- SCEVHandle IH = SCEVUnknown::get(I); // Get I as a "symbolic" SCEV.
-
- SCEVHandle V = S->evaluateAtIteration(IH);
- //std::cerr << "Evaluated: " << *this << "\n to: " << *V << "\n";
-
- return expandInTy(V, Ty);
-}
-
+STATISTIC(NumRemoved , "Number of aux indvars removed");
+STATISTIC(NumPointer , "Number of pointer indvars promoted");
+STATISTIC(NumInserted, "Number of canonical indvars added");
+STATISTIC(NumReplaced, "Number of exit values replaced");
+STATISTIC(NumLFTR , "Number of loop exit tests replaced");
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 {
+ class VISIBILITY_HIDDEN IndVarSimplify : public LoopPass {
LoopInfo *LI;
ScalarEvolution *SE;
bool Changed;
public:
- virtual bool runOnFunction(Function &) {
- LI = &getAnalysis<LoopInfo>();
- SE = &getAnalysis<ScalarEvolution>();
- Changed = false;
-
- // 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<ScalarEvolution>();
- AU.addRequired<LoopInfo>();
- AU.addPreservedID(LoopSimplifyID);
- AU.setPreservesCFG();
- }
+ static char ID; // Pass identification, replacement for typeid
+ IndVarSimplify() : LoopPass(&ID) {}
+
+ bool runOnLoop(Loop *L, LPPassManager &LPM);
+ bool doInitialization(Loop *L, LPPassManager &LPM);
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<ScalarEvolution>();
+ AU.addRequiredID(LCSSAID);
+ AU.addRequiredID(LoopSimplifyID);
+ 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);
- void LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
- SCEVExpander &RW);
- void RewriteLoopExitValues(Loop *L);
+ Instruction *LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
+ SCEVExpander &RW);
+ void RewriteLoopExitValues(Loop *L, SCEV *IterationCount);
void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
+
+ void OptimizeCanonicalIVType(Loop *L);
};
- RegisterOpt<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
}
-FunctionPass *llvm::createIndVarSimplifyPass() {
+char IndVarSimplify::ID = 0;
+static RegisterPass<IndVarSimplify>
+X("indvars", "Canonicalize Induction Variables");
+
+LoopPass *llvm::createIndVarSimplifyPass() {
return new IndVarSimplify();
}
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);
+ SE->deleteValueFromRecords(I);
+ DOUT << "INDVARS: Deleting: " << *I;
I->eraseFromParent();
Changed = true;
}
/// 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,
+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)))
+ dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
if (GEPI->getOperand(0) == PN) {
- assert(GEPI->getNumOperands() == 2 && "GEP types must mismatch!");
-
+ assert(GEPI->getNumOperands() == 2 && "GEP types must match!");
+ DOUT << "INDVARS: Eliminating pointer recurrence: " << *GEPI;
+
// 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);
+ PHINode *NewPhi = PHINode::Create(AddedVal->getType(),
+ PN->getName()+".rec", PN);
NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()), Preheader);
// Create the new add instruction.
- Value *NewAdd = BinaryOperator::createAdd(NewPhi, AddedVal,
+ 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);
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(GEPI);
- for (unsigned i = 1, e = GEPI->getNumOperands()-1;
+ 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);
+ SmallVector<Value*, 8> 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);
+ &CEIdxs[0],
+ CEIdxs.size());
+ Value *Idx[2];
+ Idx[0] = Constant::getNullValue(Type::Int32Ty);
+ Idx[1] = NewAdd;
+ GetElementPtrInst *NGEPI = GetElementPtrInst::Create(
+ NCE, Idx, Idx + 2,
+ GEPI->getName(), GEPI);
+ SE->deleteValueFromRecords(GEPI);
GEPI->replaceAllUsesWith(NGEPI);
GEPI->eraseFromParent();
GEPI = NGEPI;
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);
+ GetElementPtrInst::Create(PN->getIncomingValue(PreheaderIdx),
+ NewPhi, "", InsertPos);
+ PreInc->takeName(PN);
PN->replaceAllUsesWith(PreInc);
}
/// 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.
-void IndVarSimplify::LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
- SCEVExpander &RW) {
+///
+/// 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;
+ SmallVector<BasicBlock*, 8> ExitBlocks;
L->getExitBlocks(ExitBlocks);
- if (ExitBlocks.size() != 1) return;
+ if (ExitBlocks.size() != 1) return 0;
BasicBlock *ExitBlock = ExitBlocks[0];
// Make sure there is only one predecessor block in the loop.
if (ExitingBlock == 0)
ExitingBlock = *PI;
else
- return; // Multiple exits from loop to this block.
+ return 0; // Multiple exits from loop to this block.
}
assert(ExitingBlock && "Loop info is broken");
if (!isa<BranchInst>(ExitingBlock->getTerminator()))
- return; // Can't rewrite non-branch yet
+ 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!");
- std::set<Instruction*> InstructionsToDelete;
- if (Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()))
- InstructionsToDelete.insert(Cond);
-
+ 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.
// 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));
+ ConstantInt *OneC = ConstantInt::get(IterationCount->getType(), 1);
+ TripCount = SE->getAddExpr(IterationCount, SE->getConstant(OneC));
IndVar = L->getCanonicalInductionVariableIncrement();
} else {
// We have to use the preincremented value...
IndVar = L->getCanonicalInductionVariable();
}
+
+ DOUT << "INDVARS: LFTR: TripCount = " << *TripCount
+ << " IndVar = " << *IndVar << "\n";
// 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());
+ Value *ExitCnt = RW.expandCodeFor(TripCount, Preheader->getTerminator());
- // Insert a new setne or seteq instruction before the branch.
- Instruction::BinaryOps Opcode;
+ // Insert a new icmp_ne or icmp_eq instruction before the branch.
+ ICmpInst::Predicate Opcode;
if (L->contains(BI->getSuccessor(0)))
- Opcode = Instruction::SetNE;
+ Opcode = ICmpInst::ICMP_NE;
else
- Opcode = Instruction::SetEQ;
+ Opcode = ICmpInst::ICMP_EQ;
- Value *Cond = new SetCondInst(Opcode, IndVar, ExitCnt, "exitcond", BI);
+ Value *Cond = new ICmpInst(Opcode, IndVar, ExitCnt, "exitcond", BI);
BI->setCondition(Cond);
++NumLFTR;
Changed = true;
-
- DeleteTriviallyDeadInstructions(InstructionsToDelete);
+ return PotentiallyDeadInst;
}
/// 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) {
+void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEV *IterationCount) {
BasicBlock *Preheader = L->getLoopPreheader();
// Scan all of the instructions in the loop, looking at those that have
// 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);
+ SmallVector<BasicBlock*, 8> ExitBlocks;
+ L->getUniqueExitBlocks(ExitBlocks);
if (ExitBlocks.size() == 1)
BlockToInsertInto = ExitBlocks[0];
else
BlockToInsertInto = Preheader;
- BasicBlock::iterator InsertPt = BlockToInsertInto->begin();
- while (isa<PHINode>(InsertPt)) ++InsertPt;
+ BasicBlock::iterator InsertPt = BlockToInsertInto->getFirstNonPHI();
- bool HasConstantItCount = isa<SCEVConstant>(SE->getIterationCount(L));
+ bool HasConstantItCount = isa<SCEVConstant>(IterationCount);
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; ++I)
- 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<User*> ExtraLoopUsers;
- for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
- UI != E; ++UI)
- if (!L->contains(cast<Instruction>(*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<SCEVCouldNotCompute>(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);
- }
- }
- }
+ std::map<Instruction*, Value*> ExitValues;
+
+ // Find all values that are computed inside the loop, but used outside of it.
+ // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
+ // the exit blocks of the loop to find them.
+ for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
+ BasicBlock *ExitBB = ExitBlocks[i];
+
+ // If there are no PHI nodes in this exit block, then no values defined
+ // inside the loop are used on this path, skip it.
+ PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
+ if (!PN) continue;
+
+ unsigned NumPreds = PN->getNumIncomingValues();
+
+ // Iterate over all of the PHI nodes.
+ BasicBlock::iterator BBI = ExitBB->begin();
+ while ((PN = dyn_cast<PHINode>(BBI++))) {
+
+ // Iterate over all of the values in all the PHI nodes.
+ for (unsigned i = 0; i != NumPreds; ++i) {
+ // If the value being merged in is not integer or is not defined
+ // in the loop, skip it.
+ Value *InVal = PN->getIncomingValue(i);
+ if (!isa<Instruction>(InVal) ||
+ // SCEV only supports integer expressions for now.
+ !isa<IntegerType>(InVal->getType()))
+ continue;
+
+ // If this pred is for a subloop, not L itself, skip it.
+ if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
+ continue; // The Block is in a subloop, skip it.
+
+ // Check that InVal is defined in the loop.
+ Instruction *Inst = cast<Instruction>(InVal);
+ if (!L->contains(Inst->getParent()))
+ continue;
+
+ // We require that this value either have a computable evolution or that
+ // the loop have a constant iteration count. In the case where the loop
+ // has a constant iteration count, we can sometimes force evaluation of
+ // the exit value through brute force.
+ SCEVHandle SH = SE->getSCEV(Inst);
+ if (!SH->hasComputableLoopEvolution(L) && !HasConstantItCount)
+ continue; // Cannot get exit evolution for the loop value.
+
+ // Okay, this instruction has a user outside of the current loop
+ // and varies predictably *inside* the loop. Evaluate the value it
+ // contains when the loop exits, if possible.
+ SCEVHandle ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
+ if (isa<SCEVCouldNotCompute>(ExitValue) ||
+ !ExitValue->isLoopInvariant(L))
+ continue;
+
+ Changed = true;
+ ++NumReplaced;
+
+ // See if we already computed the exit value for the instruction, if so,
+ // just reuse it.
+ Value *&ExitVal = ExitValues[Inst];
+ if (!ExitVal)
+ ExitVal = Rewriter.expandCodeFor(ExitValue, InsertPt);
+
+ DOUT << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal
+ << " LoopVal = " << *Inst << "\n";
+
+ PN->setIncomingValue(i, ExitVal);
+
+ // If this instruction is dead now, schedule it to be removed.
+ if (Inst->use_empty())
+ InstructionsToDelete.insert(Inst);
+
+ // See if this is a single-entry LCSSA PHI node. If so, we can (and
+ // have to) remove
+ // the PHI entirely. This is safe, because the NewVal won't be variant
+ // in the loop, so we don't need an LCSSA phi node anymore.
+ if (NumPreds == 1) {
+ SE->deleteValueFromRecords(PN);
+ PN->replaceAllUsesWith(ExitVal);
+ PN->eraseFromParent();
+ break;
}
+ }
}
-
+ }
+
DeleteTriviallyDeadInstructions(InstructionsToDelete);
}
+bool IndVarSimplify::doInitialization(Loop *L, LPPassManager &LPM) {
-void IndVarSimplify::runOnLoop(Loop *L) {
+ Changed = false;
// 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();
-
+ SE = &LPM.getAnalysis<ScalarEvolution>();
+
std::set<Instruction*> DeadInsts;
for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
PHINode *PN = cast<PHINode>(I);
if (!DeadInsts.empty())
DeleteTriviallyDeadInstructions(DeadInsts);
+ return Changed;
+}
+
+bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
+
- // Next, transform all loops nesting inside of this loop.
- for (LoopInfo::iterator I = L->begin(), E = L->end(); I != E; ++I)
- runOnLoop(*I);
+ LI = &getAnalysis<LoopInfo>();
+ SE = &getAnalysis<ScalarEvolution>();
+
+ Changed = false;
+ BasicBlock *Header = L->getHeader();
+ std::set<Instruction*> DeadInsts;
+
+ // Verify the input to the pass in already in LCSSA form.
+ assert(L->isLCSSAForm());
// 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
//
SCEVHandle IterationCount = SE->getIterationCount(L);
if (!isa<SCEVCouldNotCompute>(IterationCount))
- RewriteLoopExitValues(L);
+ RewriteLoopExitValues(L, IterationCount);
// Next, analyze all of the induction variables in the loop, canonicalizing
// auxillary induction variables.
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!
+ if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable!
SCEVHandle SCEV = SE->getSCEV(PN);
if (SCEV->hasComputableLoopEvolution(L))
- // FIXME: Without a strength reduction pass, it is an extremely bad idea
- // to indvar substitute anything more complex than a linear induction
- // variable. Doing so will put expensive multiply instructions inside
- // of the loop. For now just disable indvar subst on anything more
- // complex than a linear addrec.
+ // 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->getNumOperands() == 2 && isa<SCEVConstant>(AR->getOperand(1)))
+ if (AR->isAffine())
IndVars.push_back(std::make_pair(PN, SCEV));
}
}
SCEVExpander Rewriter(*SE, *LI);
Rewriter.getOrInsertCanonicalInductionVariable(L,
IterationCount->getType());
- LinearFunctionTestReplace(L, IterationCount, Rewriter);
+ if (Instruction *I = LinearFunctionTestReplace(L, IterationCount,
+ Rewriter)) {
+ std::set<Instruction*> InstructionsToDelete;
+ InstructionsToDelete.insert(I);
+ DeleteTriviallyDeadInstructions(InstructionsToDelete);
+ }
}
- return;
+ return Changed;
}
// Compute the type of the largest recurrence expression.
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())
+ DifferingSizes |=
+ Ty->getPrimitiveSizeInBits() != LargestType->getPrimitiveSizeInBits();
+ if (Ty->getPrimitiveSizeInBits() > LargestType->getPrimitiveSizeInBits())
LargestType = Ty;
}
// 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;
+ DOUT << "INDVARS: New CanIV: " << *IndVar;
- if (!isa<SCEVCouldNotCompute>(IterationCount))
- LinearFunctionTestReplace(L, IterationCount, Rewriter);
+ if (!isa<SCEVCouldNotCompute>(IterationCount)) {
+ IterationCount = SE->getTruncateOrZeroExtend(IterationCount, LargestType);
+ 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;
+ BasicBlock::iterator InsertPt = Header->getFirstNonPHI();
// 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()]) {
+ SmallVector<unsigned,4> InsertedSizes;
+ InsertedSizes.push_back(LargestType->getPrimitiveSizeInBits());
+ for (unsigned i = 0, e = IndVars.size(); i != e; ++i) {
+ unsigned ithSize = IndVars[i].first->getType()->getPrimitiveSizeInBits();
+ if (std::find(InsertedSizes.begin(), InsertedSizes.end(), ithSize)
+ == InsertedSizes.end()) {
PHINode *PN = IndVars[i].first;
- InsertedSizes[PN->getType()->getPrimitiveSize()] = true;
- Instruction *New = new CastInst(IndVar,
- PN->getType()->getUnsignedVersion(),
- "indvar", InsertPt);
+ InsertedSizes.push_back(ithSize);
+ Instruction *New = new TruncInst(IndVar, PN->getType(), "indvar",
+ InsertPt);
Rewriter.addInsertedValue(New, SE->getSCEV(New));
+ DOUT << "INDVARS: Made trunc IV for " << *PN
+ << " NewVal = " << *New << "\n";
}
+ }
}
- // 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;
+ // Rewrite all induction variables in terms of the canonical induction
+ // variable.
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);
+ Value *NewVal = Rewriter.expandCodeFor(IndVars.back().second, InsertPt);
+ DOUT << "INDVARS: Rewrote IV '" << *IndVars.back().second << "' " << *PN
+ << " into = " << *NewVal << "\n";
+ NewVal->takeName(PN);
// Replace the old PHI Node with the inserted computation.
PN->replaceAllUsesWith(NewVal);
#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 (LoopInfo::block_iterator I = L->block_begin(), E = L->block_end();
+ I != E; ++I) {
+ BasicBlock *BB = *I;
+ if (LI->getLoopFor(BB) == L) { // Not in a subloop...
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
if (I->getType()->isInteger() && // Is an integer instruction
!I->use_empty() &&
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);
- }
+ if (isa<Instruction>(V))
+ V->takeName(I);
I->replaceAllUsesWith(V);
DeadInsts.insert(I);
++NumRemoved;
Changed = true;
- }
+ }
}
}
+ }
#endif
DeleteTriviallyDeadInstructions(DeadInsts);
+ OptimizeCanonicalIVType(L);
+ assert(L->isLCSSAForm());
+ return Changed;
}
+
+/// OptimizeCanonicalIVType - If loop induction variable is always
+/// sign or zero extended then extend the type of the induction
+/// variable.
+void IndVarSimplify::OptimizeCanonicalIVType(Loop *L) {
+ PHINode *PH = L->getCanonicalInductionVariable();
+ if (!PH) return;
+
+ // Check loop iteration count.
+ SCEVHandle IC = SE->getIterationCount(L);
+ if (isa<SCEVCouldNotCompute>(IC)) return;
+ SCEVConstant *IterationCount = dyn_cast<SCEVConstant>(IC);
+ if (!IterationCount) return;
+
+ unsigned IncomingEdge = L->contains(PH->getIncomingBlock(0));
+ unsigned BackEdge = IncomingEdge^1;
+
+ // Check IV uses. If all IV uses are either SEXT or ZEXT (except
+ // IV increment instruction) then this IV is suitable for this
+ // transformation.
+ bool isSEXT = false;
+ BinaryOperator *Incr = NULL;
+ const Type *NewType = NULL;
+ for(Value::use_iterator UI = PH->use_begin(), UE = PH->use_end();
+ UI != UE; ++UI) {
+ const Type *CandidateType = NULL;
+ if (ZExtInst *ZI = dyn_cast<ZExtInst>(UI))
+ CandidateType = ZI->getDestTy();
+ else if (SExtInst *SI = dyn_cast<SExtInst>(UI)) {
+ CandidateType = SI->getDestTy();
+ isSEXT = true;
+ }
+ else if ((Incr = dyn_cast<BinaryOperator>(UI))) {
+ // Validate IV increment instruction.
+ if (PH->getIncomingValue(BackEdge) == Incr)
+ continue;
+ }
+ if (!CandidateType) {
+ NewType = NULL;
+ break;
+ }
+ if (!NewType)
+ NewType = CandidateType;
+ else if (NewType != CandidateType) {
+ NewType = NULL;
+ break;
+ }
+ }
+
+ // IV uses are not suitable then avoid this transformation.
+ if (!NewType || !Incr)
+ return;
+
+ // IV increment instruction has two uses, one is loop exit condition
+ // and second is the IV (phi node) itself.
+ ICmpInst *Exit = NULL;
+ for(Value::use_iterator II = Incr->use_begin(), IE = Incr->use_end();
+ II != IE; ++II) {
+ if (PH == *II) continue;
+ Exit = dyn_cast<ICmpInst>(*II);
+ break;
+ }
+ if (!Exit) return;
+ ConstantInt *EV = dyn_cast<ConstantInt>(Exit->getOperand(0));
+ if (!EV)
+ EV = dyn_cast<ConstantInt>(Exit->getOperand(1));
+ if (!EV) return;
+
+ // Check iteration count max value to avoid loops that wrap around IV.
+ APInt ICount = IterationCount->getValue()->getValue();
+ if (ICount.isNegative()) return;
+ uint32_t BW = PH->getType()->getPrimitiveSizeInBits();
+ APInt Max = (isSEXT ? APInt::getSignedMaxValue(BW) : APInt::getMaxValue(BW));
+ if (ICount.getZExtValue() > Max.getZExtValue()) return;
+
+ // Extend IV type.
+
+ SCEVExpander Rewriter(*SE, *LI);
+ Value *NewIV = Rewriter.getOrInsertCanonicalInductionVariable(L,NewType);
+ PHINode *NewPH = cast<PHINode>(NewIV);
+ Instruction *NewIncr = cast<Instruction>(NewPH->getIncomingValue(BackEdge));
+
+ // Replace all SEXT or ZEXT uses.
+ SmallVector<Instruction *, 4> PHUses;
+ for(Value::use_iterator UI = PH->use_begin(), UE = PH->use_end();
+ UI != UE; ++UI) {
+ Instruction *I = cast<Instruction>(UI);
+ PHUses.push_back(I);
+ }
+ while (!PHUses.empty()){
+ Instruction *Use = PHUses.back(); PHUses.pop_back();
+ if (Incr == Use) continue;
+
+ SE->deleteValueFromRecords(Use);
+ Use->replaceAllUsesWith(NewIV);
+ Use->eraseFromParent();
+ }
+
+ // Replace exit condition.
+ ConstantInt *NEV = ConstantInt::get(NewType, EV->getZExtValue());
+ Instruction *NE = new ICmpInst(Exit->getPredicate(),
+ NewIncr, NEV, "new.exit",
+ Exit->getParent()->getTerminator());
+ SE->deleteValueFromRecords(Exit);
+ Exit->replaceAllUsesWith(NE);
+ Exit->eraseFromParent();
+
+ // Remove old IV and increment instructions.
+ SE->deleteValueFromRecords(PH);
+ PH->removeIncomingValue((unsigned)0);
+ PH->removeIncomingValue((unsigned)0);
+ SE->deleteValueFromRecords(Incr);
+ Incr->eraseFromParent();
+}
+