//
// 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.
//
//===----------------------------------------------------------------------===//
//
//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "indvars"
#include "llvm/Transforms/Scalar.h"
#include "llvm/BasicBlock.h"
#include "llvm/Constants.h"
#include "llvm/Type.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 {
- 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");
+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");
- class IndVarSimplify : public FunctionPass {
+namespace {
+ 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.addPreservedID(LCSSAID);
- AU.setPreservesCFG();
- }
+ static char ID; // Pass identification, replacement for typeid
+ IndVarSimplify() : LoopPass((intptr_t)&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);
Instruction *LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
};
- RegisterPass<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;
}
dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
if (GEPI->getOperand(0) == PN) {
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.
/*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);
}
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 0;
BasicBlock *ExitBlock = ExitBlocks[0];
// 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;
// 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
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
- }
- }
- }
+ 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;
}
-
- // Next instruction. Continue instruction skips this.
- ++I;
}
}
-
+ }
+
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) {
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
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: It is an extremely bad idea to indvar substitute anything more
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;
-
- if (!isa<SCEVCouldNotCompute>(IterationCount))
+ DOUT << "INDVARS: New CanIV: " << *IndVar;
+
+ if (!isa<SCEVCouldNotCompute>(IterationCount)) {
+ if (IterationCount->getType()->getPrimitiveSizeInBits() <
+ LargestType->getPrimitiveSizeInBits())
+ IterationCount = SE->getZeroExtendExpr(IterationCount, LargestType);
+ else if (IterationCount->getType() != LargestType)
+ IterationCount = SE->getTruncateExpr(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 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 = CastInst::create(Instruction::Trunc, 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.
+ // Rewrite all induction variables in terms of the canonical induction
+ // variable.
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);
+ 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);
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;
DeleteTriviallyDeadInstructions(DeadInsts);
- if (mustPreserveAnalysisID(LCSSAID)) assert(L->isLCSSAForm());
+ assert(L->isLCSSAForm());
+ return Changed;
}
projects / oota-llvm.git / blobdiff