//===- 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 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.
#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/Support/CFG.h"
#include "llvm/Support/GetElementPtrTypeIterator.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, InstList, 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);
-}
-
-
namespace {
Statistic<> NumRemoved ("indvars", "Number of aux indvars removed");
Statistic<> NumPointer ("indvars", "Number of pointer indvars promoted");
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);
- void LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
- SCEVExpander &RW);
+ 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");
}
FunctionPass *llvm::createIndVarSimplifyPass() {
/// 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!");
+
// 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 *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)) {
/// 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;
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.
BI->setCondition(Cond);
++NumLFTR;
Changed = true;
-
- DeleteTriviallyDeadInstructions(InstructionsToDelete);
+ return PotentiallyDeadInst;
}
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; ++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<User*> ExtraLoopUsers;
+ std::vector<Instruction*> 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);
+ 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
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)
- ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal);
+ 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;
+ }
}
DeleteTriviallyDeadInstructions(InstructionsToDelete);
//
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 (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;
}
Changed = true;
if (!isa<SCEVCouldNotCompute>(IterationCount))
- LinearFunctionTestReplace(L, IterationCount, Rewriter);
+ 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
DeadInsts.insert(I);
++NumRemoved;
Changed = true;
- }
+ }
}
}
#endif
DeleteTriviallyDeadInstructions(DeadInsts);
+
+ if (mustPreserveAnalysisID(LCSSAID)) assert(L->isLCSSAForm());
}