LoopInfo *LI;
ScalarEvolution *SE;
DominatorTree *DT;
+ SmallVector<WeakVH, 16> DeadInsts;
bool Changed;
public:
static char ID; // Pass identification, replacement for typeid
- IndVarSimplify() : LoopPass(&ID) {}
+ IndVarSimplify() : LoopPass(ID) {
+ initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
+ }
virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
}
private:
+ bool isValidRewrite(Value *FromVal, Value *ToVal);
+ void EliminateIVComparisons();
+ void EliminateIVRemainders();
void RewriteNonIntegerIVs(Loop *L);
ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
- Value *IndVar,
+ PHINode *IndVar,
BasicBlock *ExitingBlock,
BranchInst *BI,
SCEVExpander &Rewriter);
}
char IndVarSimplify::ID = 0;
-static RegisterPass<IndVarSimplify>
-X("indvars", "Canonicalize Induction Variables");
+INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
+ "Canonicalize Induction Variables", false, false)
+INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_PASS_DEPENDENCY(LoopInfo)
+INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
+INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
+INITIALIZE_PASS_DEPENDENCY(LCSSA)
+INITIALIZE_PASS_DEPENDENCY(IVUsers)
+INITIALIZE_PASS_END(IndVarSimplify, "indvars",
+ "Canonicalize Induction Variables", false, false)
Pass *llvm::createIndVarSimplifyPass() {
return new IndVarSimplify();
}
+/// isValidRewrite - Return true if the SCEV expansion generated by the
+/// rewriter can replace the original value. SCEV guarantees that it
+/// produces the same value, but the way it is produced may be illegal IR.
+/// Ideally, this function will only be called for verification.
+bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
+ // If an SCEV expression subsumed multiple pointers, its expansion could
+ // reassociate the GEP changing the base pointer. This is illegal because the
+ // final address produced by a GEP chain must be inbounds relative to its
+ // underlying object. Otherwise basic alias analysis, among other things,
+ // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
+ // producing an expression involving multiple pointers. Until then, we must
+ // bail out here.
+ //
+ // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
+ // because it understands lcssa phis while SCEV does not.
+ Value *FromPtr = FromVal;
+ Value *ToPtr = ToVal;
+ if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
+ FromPtr = GEP->getPointerOperand();
+ }
+ if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
+ ToPtr = GEP->getPointerOperand();
+ }
+ if (FromPtr != FromVal || ToPtr != ToVal) {
+ // Quickly check the common case
+ if (FromPtr == ToPtr)
+ return true;
+
+ // SCEV may have rewritten an expression that produces the GEP's pointer
+ // operand. That's ok as long as the pointer operand has the same base
+ // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
+ // base of a recurrence. This handles the case in which SCEV expansion
+ // converts a pointer type recurrence into a nonrecurrent pointer base
+ // indexed by an integer recurrence.
+ const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
+ const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
+ if (FromBase == ToBase)
+ return true;
+
+ DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
+ << *FromBase << " != " << *ToBase << "\n");
+
+ return false;
+ }
+ return true;
+}
+
/// 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
/// is actually a much broader range than just linear tests.
ICmpInst *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
const SCEV *BackedgeTakenCount,
- Value *IndVar,
+ PHINode *IndVar,
BasicBlock *ExitingBlock,
BranchInst *BI,
SCEVExpander &Rewriter) {
+ // Special case: If the backedge-taken count is a UDiv, it's very likely a
+ // UDiv that ScalarEvolution produced in order to compute a precise
+ // expression, rather than a UDiv from the user's code. If we can't find a
+ // UDiv in the code with some simple searching, assume the former and forego
+ // rewriting the loop.
+ if (isa<SCEVUDivExpr>(BackedgeTakenCount)) {
+ ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
+ if (!OrigCond) return 0;
+ const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
+ R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
+ if (R != BackedgeTakenCount) {
+ const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
+ L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
+ if (L != BackedgeTakenCount)
+ return 0;
+ }
+ }
+
// 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.
// Add one to the "backedge-taken" count to get the trip count.
// If this addition may overflow, we have to be more pessimistic and
// cast the induction variable before doing the add.
- const SCEV *Zero = SE->getIntegerSCEV(0, BackedgeTakenCount->getType());
+ const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0);
const SCEV *N =
SE->getAddExpr(BackedgeTakenCount,
- SE->getIntegerSCEV(1, BackedgeTakenCount->getType()));
+ SE->getConstant(BackedgeTakenCount->getType(), 1));
if ((isa<SCEVConstant>(N) && !N->isZero()) ||
- SE->isLoopGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
+ SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
// No overflow. Cast the sum.
RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
} else {
RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
IndVar->getType());
RHS = SE->getAddExpr(RHS,
- SE->getIntegerSCEV(1, IndVar->getType()));
+ SE->getConstant(IndVar->getType(), 1));
}
// The BackedgeTaken expression contains the number of times that the
// backedge branches to the loop header. This is one less than the
// number of times the loop executes, so use the incremented indvar.
- CmpIndVar = L->getCanonicalInductionVariableIncrement();
+ CmpIndVar = IndVar->getIncomingValueForBlock(ExitingBlock);
} else {
// We have to use the preincremented value...
RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
}
// Expand the code for the iteration count.
- assert(RHS->isLoopInvariant(L) &&
+ assert(SE->isLoopInvariant(RHS, L) &&
"Computed iteration count is not loop invariant!");
Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
/// happen later, except that it's more powerful in some cases, because it's
/// able to brute-force evaluate arbitrary instructions as long as they have
/// constant operands at the beginning of the loop.
-void IndVarSimplify::RewriteLoopExitValues(Loop *L,
- SCEVExpander &Rewriter) {
+void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
// Verify the input to the pass in already in LCSSA form.
assert(L->isLCSSAForm(*DT));
// and varies predictably *inside* the loop. Evaluate the value it
// contains when the loop exits, if possible.
const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
- if (!ExitValue->isLoopInvariant(L))
+ if (!SE->isLoopInvariant(ExitValue, L))
continue;
- Changed = true;
- ++NumReplaced;
-
Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
<< " LoopVal = " << *Inst << "\n");
+ if (!isValidRewrite(Inst, ExitVal)) {
+ DeadInsts.push_back(ExitVal);
+ continue;
+ }
+ Changed = true;
+ ++NumReplaced;
+
PN->setIncomingValue(i, ExitVal);
// If this instruction is dead now, delete it.
// If there are, change them into integer recurrences, permitting analysis by
// the SCEV routines.
//
- BasicBlock *Header = L->getHeader();
+ BasicBlock *Header = L->getHeader();
SmallVector<WeakVH, 8> PHIs;
for (BasicBlock::iterator I = Header->begin();
PHIs.push_back(PN);
for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
- if (PHINode *PN = dyn_cast_or_null<PHINode>(PHIs[i]))
+ if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
HandleFloatingPointIV(L, PN);
// If the loop previously had floating-point IV, ScalarEvolution
SE->forgetLoop(L);
}
+void IndVarSimplify::EliminateIVComparisons() {
+ // Look for ICmp users.
+ for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
+ IVStrideUse &UI = *I;
+ ICmpInst *ICmp = dyn_cast<ICmpInst>(UI.getUser());
+ if (!ICmp) continue;
+
+ bool Swapped = UI.getOperandValToReplace() == ICmp->getOperand(1);
+ ICmpInst::Predicate Pred = ICmp->getPredicate();
+ if (Swapped) Pred = ICmpInst::getSwappedPredicate(Pred);
+
+ // Get the SCEVs for the ICmp operands.
+ const SCEV *S = IU->getReplacementExpr(UI);
+ const SCEV *X = SE->getSCEV(ICmp->getOperand(!Swapped));
+
+ // Simplify unnecessary loops away.
+ const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
+ S = SE->getSCEVAtScope(S, ICmpLoop);
+ X = SE->getSCEVAtScope(X, ICmpLoop);
+
+ // If the condition is always true or always false, replace it with
+ // a constant value.
+ if (SE->isKnownPredicate(Pred, S, X))
+ ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
+ else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
+ ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
+ else
+ continue;
+
+ DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
+ DeadInsts.push_back(ICmp);
+ }
+}
+
+void IndVarSimplify::EliminateIVRemainders() {
+ // Look for SRem and URem users.
+ for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E; ++I) {
+ IVStrideUse &UI = *I;
+ BinaryOperator *Rem = dyn_cast<BinaryOperator>(UI.getUser());
+ if (!Rem) continue;
+
+ bool isSigned = Rem->getOpcode() == Instruction::SRem;
+ if (!isSigned && Rem->getOpcode() != Instruction::URem)
+ continue;
+
+ // We're only interested in the case where we know something about
+ // the numerator.
+ if (UI.getOperandValToReplace() != Rem->getOperand(0))
+ continue;
+
+ // Get the SCEVs for the ICmp operands.
+ const SCEV *S = SE->getSCEV(Rem->getOperand(0));
+ const SCEV *X = SE->getSCEV(Rem->getOperand(1));
+
+ // Simplify unnecessary loops away.
+ const Loop *ICmpLoop = LI->getLoopFor(Rem->getParent());
+ S = SE->getSCEVAtScope(S, ICmpLoop);
+ X = SE->getSCEVAtScope(X, ICmpLoop);
+
+ // i % n --> i if i is in [0,n).
+ if ((!isSigned || SE->isKnownNonNegative(S)) &&
+ SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
+ S, X))
+ Rem->replaceAllUsesWith(Rem->getOperand(0));
+ else {
+ // (i+1) % n --> (i+1)==n?0:(i+1) if i is in [0,n).
+ const SCEV *LessOne =
+ SE->getMinusSCEV(S, SE->getConstant(S->getType(), 1));
+ if ((!isSigned || SE->isKnownNonNegative(LessOne)) &&
+ SE->isKnownPredicate(isSigned ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT,
+ LessOne, X)) {
+ ICmpInst *ICmp = new ICmpInst(Rem, ICmpInst::ICMP_EQ,
+ Rem->getOperand(0), Rem->getOperand(1),
+ "tmp");
+ SelectInst *Sel =
+ SelectInst::Create(ICmp,
+ ConstantInt::get(Rem->getType(), 0),
+ Rem->getOperand(0), "tmp", Rem);
+ Rem->replaceAllUsesWith(Sel);
+ } else
+ continue;
+ }
+
+ // Inform IVUsers about the new users.
+ if (Instruction *I = dyn_cast<Instruction>(Rem->getOperand(0)))
+ IU->AddUsersIfInteresting(I);
+
+ DEBUG(dbgs() << "INDVARS: Simplified rem: " << *Rem << '\n');
+ DeadInsts.push_back(Rem);
+ }
+}
+
bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
+ // If LoopSimplify form is not available, stay out of trouble. Some notes:
+ // - LSR currently only supports LoopSimplify-form loops. Indvars'
+ // canonicalization can be a pessimization without LSR to "clean up"
+ // afterwards.
+ // - We depend on having a preheader; in particular,
+ // Loop::getCanonicalInductionVariable only supports loops with preheaders,
+ // and we're in trouble if we can't find the induction variable even when
+ // we've manually inserted one.
+ if (!L->isLoopSimplifyForm())
+ return false;
+
IU = &getAnalysis<IVUsers>();
LI = &getAnalysis<LoopInfo>();
SE = &getAnalysis<ScalarEvolution>();
DT = &getAnalysis<DominatorTree>();
+ DeadInsts.clear();
Changed = false;
// If there are any floating-point recurrences, attempt to
if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
RewriteLoopExitValues(L, Rewriter);
+ // Simplify ICmp IV users.
+ EliminateIVComparisons();
+
+ // Simplify SRem and URem IV users.
+ EliminateIVRemainders();
+
// Compute the type of the largest recurrence expression, and decide whether
// a canonical induction variable should be inserted.
const Type *LargestType = 0;
// Now that we know the largest of the induction variable expressions
// in this loop, insert a canonical induction variable of the largest size.
- Value *IndVar = 0;
+ PHINode *IndVar = 0;
if (NeedCannIV) {
// Check to see if the loop already has any canonical-looking induction
// variables. If any are present and wider than the planned canonical
ExitingBlock, BI, Rewriter);
}
- // Rewrite IV-derived expressions. Clears the rewriter cache.
+ // Rewrite IV-derived expressions.
RewriteIVExpressions(L, Rewriter);
+ // Clear the rewriter cache, because values that are in the rewriter's cache
+ // can be deleted in the loop below, causing the AssertingVH in the cache to
+ // trigger.
+ Rewriter.clear();
+
+ // Now that we're done iterating through lists, clean up any instructions
+ // which are now dead.
+ while (!DeadInsts.empty())
+ if (Instruction *Inst =
+ dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
+ RecursivelyDeleteTriviallyDeadInstructions(Inst);
+
// The Rewriter may not be used from this point on.
// Loop-invariant instructions in the preheader that aren't used in the
return Changed;
}
-void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
- SmallVector<WeakVH, 16> DeadInsts;
+// 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, unless
+// it can be expanded to a trivial value.
+static bool isSafe(const SCEV *S, const Loop *L, ScalarEvolution *SE) {
+ // Loop-invariant values are safe.
+ if (SE->isLoopInvariant(S, L)) return true;
+
+ // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
+ // to transform them into efficient code.
+ if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
+ return AR->isAffine();
+
+ // An add is safe it all its operands are safe.
+ if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
+ for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
+ E = Commutative->op_end(); I != E; ++I)
+ if (!isSafe(*I, L, SE)) return false;
+ return true;
+ }
+
+ // A cast is safe if its operand is.
+ if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
+ return isSafe(C->getOperand(), L, SE);
+ // A udiv is safe if its operands are.
+ if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
+ return isSafe(UD->getLHS(), L, SE) &&
+ isSafe(UD->getRHS(), L, SE);
+
+ // SCEVUnknown is always safe.
+ if (isa<SCEVUnknown>(S))
+ return true;
+
+ // Nothing else is safe.
+ return false;
+}
+
+void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
// Rewrite all induction variable expressions in terms of the canonical
// induction variable.
//
// the need for the code evaluation methods to insert induction variables
// of different sizes.
for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
- const SCEV *Stride = UI->getStride();
Value *Op = UI->getOperandValToReplace();
const Type *UseTy = Op->getType();
Instruction *User = UI->getUser();
// Evaluate the expression out of the loop, if possible.
if (!L->contains(UI->getUser())) {
const SCEV *ExitVal = SE->getSCEVAtScope(AR, L->getParentLoop());
- if (ExitVal->isLoopInvariant(L))
+ if (SE->isLoopInvariant(ExitVal, L))
AR = ExitVal;
}
// currently can only reduce affine polynomials. For now just disable
// indvar subst on anything more complex than an affine addrec, unless
// it can be expanded to a trivial value.
- if (!AR->isLoopInvariant(L) && !Stride->isLoopInvariant(L))
+ if (!isSafe(AR, L, SE))
continue;
// Determine the insertion point for this user. By default, insert
// Now expand it into actual Instructions and patch it into place.
Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
+ DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
+ << " into = " << *NewVal << "\n");
+
+ if (!isValidRewrite(Op, NewVal)) {
+ DeadInsts.push_back(NewVal);
+ continue;
+ }
// Inform ScalarEvolution that this value is changing. The change doesn't
// affect its value, but it does potentially affect which use lists the
// value will be on after the replacement, which affects ScalarEvolution's
NewVal->takeName(Op);
User->replaceUsesOfWith(Op, NewVal);
UI->setOperandValToReplace(NewVal);
- DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
- << " into = " << *NewVal << "\n");
+
++NumRemoved;
Changed = true;
// The old value may be dead now.
DeadInsts.push_back(Op);
}
-
- // Clear the rewriter cache, because values that are in the rewriter's cache
- // can be deleted in the loop below, causing the AssertingVH in the cache to
- // trigger.
- Rewriter.clear();
- // Now that we're done iterating through lists, clean up any instructions
- // which are now dead.
- while (!DeadInsts.empty())
- if (Instruction *Inst =
- dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val()))
- RecursivelyDeleteTriviallyDeadInstructions(Inst);
}
/// If there's a single exit block, sink any loop-invariant values that
bool UsedInLoop = false;
for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
UI != UE; ++UI) {
- BasicBlock *UseBB = cast<Instruction>(UI)->getParent();
- if (PHINode *P = dyn_cast<PHINode>(UI)) {
+ User *U = *UI;
+ BasicBlock *UseBB = cast<Instruction>(U)->getParent();
+ if (PHINode *P = dyn_cast<PHINode>(U)) {
unsigned i =
PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
UseBB = P->getIncomingBlock(i);
}
}
-/// convertToInt - Convert APF to an integer, if possible.
-static bool convertToInt(const APFloat &APF, uint64_t &intVal) {
+/// ConvertToSInt - Convert APF to an integer, if possible.
+static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
bool isExact = false;
if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
return false;
- if (APF.convertToInteger(&intVal, 32, APF.isNegative(),
- APFloat::rmTowardZero, &isExact) != APFloat::opOK)
- return false;
- if (!isExact)
+ // See if we can convert this to an int64_t
+ uint64_t UIntVal;
+ if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
+ &isExact) != APFloat::opOK || !isExact)
return false;
+ IntVal = UIntVal;
return true;
}
// Check incoming value.
ConstantFP *InitValueVal =
dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
- if (!InitValueVal) return;
-
- uint64_t InitValue;
- if (!convertToInt(InitValueVal->getValueAPF(), InitValue))
+
+ int64_t InitValue;
+ if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
return;
// Check IV increment. Reject this PN if increment operation is not
BinaryOperator *Incr =
dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
-
+
// If this is not an add of the PHI with a constantfp, or if the constant fp
// is not an integer, bail out.
ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
- uint64_t IntValue;
+ int64_t IncValue;
if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
- !convertToInt(IncValueVal->getValueAPF(), IntValue))
+ !ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
return;
// Check Incr uses. One user is PN and the other user is an exit condition
// used by the conditional terminator.
Value::use_iterator IncrUse = Incr->use_begin();
- Instruction *U1 = cast<Instruction>(IncrUse++);
+ Instruction *U1 = cast<Instruction>(*IncrUse++);
if (IncrUse == Incr->use_end()) return;
- Instruction *U2 = cast<Instruction>(IncrUse++);
+ Instruction *U2 = cast<Instruction>(*IncrUse++);
if (IncrUse != Incr->use_end()) return;
// Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
if (Compare == 0 || !Compare->hasOneUse() ||
!isa<BranchInst>(Compare->use_back()))
return;
-
+
BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
+ // We need to verify that the branch actually controls the iteration count
+ // of the loop. If not, the new IV can overflow and no one will notice.
+ // The branch block must be in the loop and one of the successors must be out
+ // of the loop.
+ assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
+ if (!L->contains(TheBr->getParent()) ||
+ (L->contains(TheBr->getSuccessor(0)) &&
+ L->contains(TheBr->getSuccessor(1))))
+ return;
+
+
// If it isn't a comparison with an integer-as-fp (the exit value), we can't
// transform it.
ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
- uint64_t ExitValue;
- if (ExitValueVal == 0 || !convertToInt(ExitValueVal->getValueAPF(),ExitValue))
+ int64_t ExitValue;
+ if (ExitValueVal == 0 ||
+ !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
return;
// Find new predicate for integer comparison.
default: return; // Unknown comparison.
case CmpInst::FCMP_OEQ:
case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
+ case CmpInst::FCMP_ONE:
+ case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
case CmpInst::FCMP_OGT:
case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
case CmpInst::FCMP_OGE:
case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
}
+ // We convert the floating point induction variable to a signed i32 value if
+ // we can. This is only safe if the comparison will not overflow in a way
+ // that won't be trapped by the integer equivalent operations. Check for this
+ // now.
+ // TODO: We could use i64 if it is native and the range requires it.
+
+ // The start/stride/exit values must all fit in signed i32.
+ if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
+ return;
+
+ // If not actually striding (add x, 0.0), avoid touching the code.
+ if (IncValue == 0)
+ return;
+
+ // Positive and negative strides have different safety conditions.
+ if (IncValue > 0) {
+ // If we have a positive stride, we require the init to be less than the
+ // exit value and an equality or less than comparison.
+ if (InitValue >= ExitValue ||
+ NewPred == CmpInst::ICMP_SGT || NewPred == CmpInst::ICMP_SGE)
+ return;
+
+ uint32_t Range = uint32_t(ExitValue-InitValue);
+ if (NewPred == CmpInst::ICMP_SLE) {
+ // Normalize SLE -> SLT, check for infinite loop.
+ if (++Range == 0) return; // Range overflows.
+ }
+
+ unsigned Leftover = Range % uint32_t(IncValue);
+
+ // If this is an equality comparison, we require that the strided value
+ // exactly land on the exit value, otherwise the IV condition will wrap
+ // around and do things the fp IV wouldn't.
+ if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
+ Leftover != 0)
+ return;
+
+ // If the stride would wrap around the i32 before exiting, we can't
+ // transform the IV.
+ if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
+ return;
+
+ } else {
+ // If we have a negative stride, we require the init to be greater than the
+ // exit value and an equality or greater than comparison.
+ if (InitValue >= ExitValue ||
+ NewPred == CmpInst::ICMP_SLT || NewPred == CmpInst::ICMP_SLE)
+ return;
+
+ uint32_t Range = uint32_t(InitValue-ExitValue);
+ if (NewPred == CmpInst::ICMP_SGE) {
+ // Normalize SGE -> SGT, check for infinite loop.
+ if (++Range == 0) return; // Range overflows.
+ }
+
+ unsigned Leftover = Range % uint32_t(-IncValue);
+
+ // If this is an equality comparison, we require that the strided value
+ // exactly land on the exit value, otherwise the IV condition will wrap
+ // around and do things the fp IV wouldn't.
+ if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
+ Leftover != 0)
+ return;
+
+ // If the stride would wrap around the i32 before exiting, we can't
+ // transform the IV.
+ if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
+ return;
+ }
+
const IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
-
- // Insert new i32 integer induction variable.
- PHINode *NewPHI = PHINode::Create(Int32Ty, PN->getName()+".int", PN);
+
+ // Insert new integer induction variable.
+ PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
PN->getIncomingBlock(IncomingEdge));
Value *NewAdd =
- BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IntValue),
+ BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
Incr->getName()+".int", Incr);
NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
projects / oota-llvm.git / blobdiff