#include <cmath>
using namespace llvm;
-STATISTIC(NumBruteForceEvaluations,
- "Number of brute force evaluations needed to "
- "calculate high-order polynomial exit values");
STATISTIC(NumArrayLenItCounts,
"Number of trip counts computed with array length");
STATISTIC(NumTripCountsComputed,
SCEV::~SCEV() {}
void SCEV::dump() const {
print(cerr);
+ cerr << '\n';
}
uint32_t SCEV::getBitWidth() const {
return LHS->getType();
}
+
+// SCEVSDivs - Only allow the creation of one SCEVSDivExpr for any particular
+// input. Don't use a SCEVHandle here, or else the object will never be
+// deleted!
+static ManagedStatic<std::map<std::pair<SCEV*, SCEV*>,
+ SCEVSDivExpr*> > SCEVSDivs;
+
+SCEVSDivExpr::~SCEVSDivExpr() {
+ SCEVSDivs->erase(std::make_pair(LHS, RHS));
+}
+
+void SCEVSDivExpr::print(std::ostream &OS) const {
+ OS << "(" << *LHS << " /s " << *RHS << ")";
+}
+
+const Type *SCEVSDivExpr::getType() const {
+ return LHS->getType();
+}
+
+
// SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any
// particular input. Don't use a SCEVHandle here, or else the object will never
// be deleted!
// The computation is correct in the face of overflow provided that the
// multiplication is performed _after_ the evaluation of the binomial
// coefficient.
- SCEVHandle Val =
- SE.getMulExpr(getOperand(i),
- BinomialCoefficient(It, i, SE,
- cast<IntegerType>(getType())));
- Result = SE.getAddExpr(Result, Val);
+ SCEVHandle Coeff = BinomialCoefficient(It, i, SE,
+ cast<IntegerType>(getType()));
+ if (isa<SCEVCouldNotCompute>(Coeff))
+ return Coeff;
+
+ Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff));
}
return Result;
}
// If we found some loop invariants, fold them into the recurrence.
if (!LIOps.empty()) {
- // NLI + LI + { Start,+,Step} --> NLI + { LI+Start,+,Step }
+ // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step}
LIOps.push_back(AddRec->getStart());
std::vector<SCEVHandle> AddRecOps(AddRec->op_begin(), AddRec->op_end());
// If we found some loop invariants, fold them into the recurrence.
if (!LIOps.empty()) {
- // NLI * LI * { Start,+,Step} --> NLI * { LI*Start,+,LI*Step }
+ // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step}
std::vector<SCEVHandle> NewOps;
NewOps.reserve(AddRec->getNumOperands());
if (LIOps.size() == 1) {
}
SCEVHandle ScalarEvolution::getUDivExpr(const SCEVHandle &LHS, const SCEVHandle &RHS) {
+ if (LHS == RHS)
+ return getIntegerSCEV(1, LHS->getType()); // X udiv X --> 1
+
if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
if (RHSC->getValue()->equalsInt(1))
- return LHS; // X udiv 1 --> x
+ return LHS; // X udiv 1 --> X
if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
Constant *LHSCV = LHSC->getValue();
}
}
- // FIXME: implement folding of (X*4)/4 when we know X*4 doesn't overflow.
-
SCEVUDivExpr *&Result = (*SCEVUDivs)[std::make_pair(LHS, RHS)];
if (Result == 0) Result = new SCEVUDivExpr(LHS, RHS);
return Result;
}
+SCEVHandle ScalarEvolution::getSDivExpr(const SCEVHandle &LHS, const SCEVHandle &RHS) {
+ if (LHS == RHS)
+ return getIntegerSCEV(1, LHS->getType()); // X sdiv X --> 1
+
+ if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) {
+ if (RHSC->getValue()->equalsInt(1))
+ return LHS; // X sdiv 1 --> X
+
+ if (RHSC->getValue()->isAllOnesValue())
+ return getNegativeSCEV(LHS); // X sdiv -1 --> -X
+
+ if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) {
+ Constant *LHSCV = LHSC->getValue();
+ Constant *RHSCV = RHSC->getValue();
+ return getUnknown(ConstantExpr::getSDiv(LHSCV, RHSCV));
+ }
+ }
+
+ SCEVSDivExpr *&Result = (*SCEVSDivs)[std::make_pair(LHS, RHS)];
+ if (Result == 0) Result = new SCEVSDivExpr(LHS, RHS);
+ return Result;
+}
+
/// SCEVAddRecExpr::get - Get a add recurrence expression for the
/// specified loop. Simplify the expression as much as possible.
if (Operands.back()->isZero()) {
Operands.pop_back();
- return getAddRecExpr(Operands, L); // { X,+,0 } --> X
+ return getAddRecExpr(Operands, L); // {X,+,0} --> X
}
// Canonicalize nested AddRecs in by nesting them in order of loop depth.
void setSCEV(Value *V, const SCEVHandle &H) {
bool isNew = Scalars.insert(std::make_pair(V, H)).second;
assert(isNew && "This entry already existed!");
+ isNew = false;
}
/// specified less-than comparison will execute. If not computable, return
/// UnknownValue. isSigned specifies whether the less-than is signed.
SCEVHandle HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L,
- bool isSigned);
+ bool isSigned, bool trueWhenEqual);
+
+ /// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
+ /// (which may not be an immediate predecessor) which has exactly one
+ /// successor from which BB is reachable, or null if no such block is
+ /// found.
+ BasicBlock* getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB);
/// executesAtLeastOnce - Test whether entry to the loop is protected by
/// a conditional between LHS and RHS.
- bool executesAtLeastOnce(const Loop *L, bool isSigned, SCEV *LHS, SCEV *RHS);
+ bool executesAtLeastOnce(const Loop *L, bool isSigned, bool trueWhenEqual,
+ SCEV *LHS, SCEV *RHS);
+
+ /// potentialInfiniteLoop - Test whether the loop might jump over the exit value
+ /// due to wrapping.
+ bool potentialInfiniteLoop(SCEV *Stride, SCEV *RHS, bool isSigned,
+ bool trueWhenEqual);
/// getConstantEvolutionLoopExitValue - If we know that the specified Phi is
/// in the header of its containing loop, we know the loop executes a
return MinOpRes;
}
- // SCEVUDivExpr, SCEVUnknown
+ // SCEVUDivExpr, SCEVSDivExpr, SCEVUnknown
return 0;
}
case Instruction::UDiv:
return SE.getUDivExpr(getSCEV(U->getOperand(0)),
getSCEV(U->getOperand(1)));
+ case Instruction::SDiv:
+ return SE.getSDivExpr(getSCEV(U->getOperand(0)),
+ getSCEV(U->getOperand(1)));
case Instruction::Sub:
return SE.getMinusSCEV(getSCEV(U->getOperand(0)),
getSCEV(U->getOperand(1)));
break;
case Instruction::LShr:
- // Turn logical shift right of a constant into a unsigned divide.
+ // Turn logical shift right of a constant into an unsigned divide.
if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) {
uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
Constant *X = ConstantInt::get(
// At this point, we would like to compute how many iterations of the
// loop the predicate will return true for these inputs.
- if (isa<SCEVConstant>(LHS) && !isa<SCEVConstant>(RHS)) {
- // If there is a constant, force it into the RHS.
+ if (LHS->isLoopInvariant(L) && !RHS->isLoopInvariant(L)) {
+ // If there is a loop-invariant, force it into the RHS.
std::swap(LHS, RHS);
Cond = ICmpInst::getSwappedPredicate(Cond);
}
break;
}
case ICmpInst::ICMP_SLT: {
- SCEVHandle TC = HowManyLessThans(LHS, RHS, L, true);
+ SCEVHandle TC = HowManyLessThans(LHS, RHS, L, true, false);
if (!isa<SCEVCouldNotCompute>(TC)) return TC;
break;
}
case ICmpInst::ICMP_SGT: {
SCEVHandle TC = HowManyLessThans(SE.getNotSCEV(LHS),
- SE.getNotSCEV(RHS), L, true);
+ SE.getNotSCEV(RHS), L, true, false);
if (!isa<SCEVCouldNotCompute>(TC)) return TC;
break;
}
case ICmpInst::ICMP_ULT: {
- SCEVHandle TC = HowManyLessThans(LHS, RHS, L, false);
+ SCEVHandle TC = HowManyLessThans(LHS, RHS, L, false, false);
if (!isa<SCEVCouldNotCompute>(TC)) return TC;
break;
}
case ICmpInst::ICMP_UGT: {
SCEVHandle TC = HowManyLessThans(SE.getNotSCEV(LHS),
- SE.getNotSCEV(RHS), L, false);
+ SE.getNotSCEV(RHS), L, false, false);
+ if (!isa<SCEVCouldNotCompute>(TC)) return TC;
+ break;
+ }
+ case ICmpInst::ICMP_SLE: {
+ SCEVHandle TC = HowManyLessThans(LHS, RHS, L, true, true);
+ if (!isa<SCEVCouldNotCompute>(TC)) return TC;
+ break;
+ }
+ case ICmpInst::ICMP_SGE: {
+ SCEVHandle TC = HowManyLessThans(SE.getNotSCEV(LHS),
+ SE.getNotSCEV(RHS), L, true, true);
+ if (!isa<SCEVCouldNotCompute>(TC)) return TC;
+ break;
+ }
+ case ICmpInst::ICMP_ULE: {
+ SCEVHandle TC = HowManyLessThans(LHS, RHS, L, false, true);
+ if (!isa<SCEVCouldNotCompute>(TC)) return TC;
+ break;
+ }
+ case ICmpInst::ICMP_UGE: {
+ SCEVHandle TC = HowManyLessThans(SE.getNotSCEV(LHS),
+ SE.getNotSCEV(RHS), L, false, true);
if (!isa<SCEVCouldNotCompute>(TC)) return TC;
break;
}
return Comm;
}
- if (SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) {
- SCEVHandle LHS = getSCEVAtScope(Div->getLHS(), L);
+ if (SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(V)) {
+ SCEVHandle LHS = getSCEVAtScope(UDiv->getLHS(), L);
if (LHS == UnknownValue) return LHS;
- SCEVHandle RHS = getSCEVAtScope(Div->getRHS(), L);
+ SCEVHandle RHS = getSCEVAtScope(UDiv->getRHS(), L);
if (RHS == UnknownValue) return RHS;
- if (LHS == Div->getLHS() && RHS == Div->getRHS())
- return Div; // must be loop invariant
+ if (LHS == UDiv->getLHS() && RHS == UDiv->getRHS())
+ return UDiv; // must be loop invariant
return SE.getUDivExpr(LHS, RHS);
}
+ if (SCEVSDivExpr *SDiv = dyn_cast<SCEVSDivExpr>(V)) {
+ SCEVHandle LHS = getSCEVAtScope(SDiv->getLHS(), L);
+ if (LHS == UnknownValue) return LHS;
+ SCEVHandle RHS = getSCEVAtScope(SDiv->getRHS(), L);
+ if (RHS == UnknownValue) return RHS;
+ if (LHS == SDiv->getLHS() && RHS == SDiv->getRHS())
+ return SDiv; // must be loop invariant
+ return SE.getSDivExpr(LHS, RHS);
+ }
+
// If this is a loop recurrence for a loop that does not contain L, then we
// are dealing with the final value computed by the loop.
if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) {
// The divisions must be performed as signed divisions.
APInt NegB(-B);
APInt TwoA( A << 1 );
+ if (TwoA.isMinValue()) {
+ SCEV *CNC = new SCEVCouldNotCompute();
+ return std::make_pair(CNC, CNC);
+ }
+
ConstantInt *Solution1 = ConstantInt::get((NegB + SqrtVal).sdiv(TwoA));
ConstantInt *Solution2 = ConstantInt::get((NegB - SqrtVal).sdiv(TwoA));
return UnknownValue;
}
+/// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB
+/// (which may not be an immediate predecessor) which has exactly one
+/// successor from which BB is reachable, or null if no such block is
+/// found.
+///
+BasicBlock *
+ScalarEvolutionsImpl::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) {
+ // If the block has a unique predecessor, the predecessor must have
+ // no other successors from which BB is reachable.
+ if (BasicBlock *Pred = BB->getSinglePredecessor())
+ return Pred;
+
+ // A loop's header is defined to be a block that dominates the loop.
+ // If the loop has a preheader, it must be a block that has exactly
+ // one successor that can reach BB. This is slightly more strict
+ // than necessary, but works if critical edges are split.
+ if (Loop *L = LI.getLoopFor(BB))
+ return L->getLoopPreheader();
+
+ return 0;
+}
+
/// executesAtLeastOnce - Test whether entry to the loop is protected by
/// a conditional between LHS and RHS.
bool ScalarEvolutionsImpl::executesAtLeastOnce(const Loop *L, bool isSigned,
+ bool trueWhenEqual,
SCEV *LHS, SCEV *RHS) {
BasicBlock *Preheader = L->getLoopPreheader();
BasicBlock *PreheaderDest = L->getHeader();
// Starting at the preheader, climb up the predecessor chain, as long as
- // there are unique predecessors, looking for a conditional branch that
- // protects the loop.
- //
- // This is a conservative apporoximation of a climb of the
- // control-dependence predecessors.
-
- for (; Preheader; PreheaderDest = Preheader,
- Preheader = Preheader->getSinglePredecessor()) {
+ // there are predecessors that can be found that have unique successors
+ // leading to the original header.
+ for (; Preheader;
+ PreheaderDest = Preheader,
+ Preheader = getPredecessorWithUniqueSuccessorForBB(Preheader)) {
BranchInst *LoopEntryPredicate =
dyn_cast<BranchInst>(Preheader->getTerminator());
switch (Cond) {
case ICmpInst::ICMP_UGT:
- if (isSigned) continue;
+ if (isSigned || trueWhenEqual) continue;
std::swap(PreCondLHS, PreCondRHS);
Cond = ICmpInst::ICMP_ULT;
break;
case ICmpInst::ICMP_SGT:
- if (!isSigned) continue;
+ if (!isSigned || trueWhenEqual) continue;
std::swap(PreCondLHS, PreCondRHS);
Cond = ICmpInst::ICMP_SLT;
break;
case ICmpInst::ICMP_ULT:
- if (isSigned) continue;
+ if (isSigned || trueWhenEqual) continue;
break;
case ICmpInst::ICMP_SLT:
- if (!isSigned) continue;
+ if (!isSigned || trueWhenEqual) continue;
+ break;
+ case ICmpInst::ICMP_UGE:
+ if (isSigned || !trueWhenEqual) continue;
+ std::swap(PreCondLHS, PreCondRHS);
+ Cond = ICmpInst::ICMP_ULE;
+ break;
+ case ICmpInst::ICMP_SGE:
+ if (!isSigned || !trueWhenEqual) continue;
+ std::swap(PreCondLHS, PreCondRHS);
+ Cond = ICmpInst::ICMP_SLE;
+ break;
+ case ICmpInst::ICMP_ULE:
+ if (isSigned || !trueWhenEqual) continue;
+ break;
+ case ICmpInst::ICMP_SLE:
+ if (!isSigned || !trueWhenEqual) continue;
break;
default:
continue;
return false;
}
+/// potentialInfiniteLoop - Test whether the loop might jump over the exit value
+/// due to wrapping around 2^n.
+bool ScalarEvolutionsImpl::potentialInfiniteLoop(SCEV *Stride, SCEV *RHS,
+ bool isSigned, bool trueWhenEqual) {
+ // Return true when the distance from RHS to maxint > Stride.
+
+ SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride);
+ if (!SC)
+ return true;
+
+ if (SC->getValue()->isZero())
+ return true;
+ if (!trueWhenEqual && SC->getValue()->isOne())
+ return false;
+
+ SCEVConstant *R = dyn_cast<SCEVConstant>(RHS);
+ if (!R)
+ return true;
+
+ if (isSigned)
+ return true; // XXX: because we don't have an sdiv scev.
+
+ // If negative, it wraps around every iteration, but we don't care about that.
+ APInt S = SC->getValue()->getValue().abs();
+
+ APInt Dist = APInt::getMaxValue(R->getValue()->getBitWidth()) -
+ R->getValue()->getValue();
+
+ if (trueWhenEqual)
+ return !S.ult(Dist);
+ else
+ return !S.ule(Dist);
+}
+
/// HowManyLessThans - Return the number of times a backedge containing the
/// specified less-than comparison will execute. If not computable, return
/// UnknownValue.
SCEVHandle ScalarEvolutionsImpl::
-HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L, bool isSigned) {
+HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L,
+ bool isSigned, bool trueWhenEqual) {
// Only handle: "ADDREC < LoopInvariant".
if (!RHS->isLoopInvariant(L)) return UnknownValue;
return UnknownValue;
if (AddRec->isAffine()) {
- // FORNOW: We only support unit strides.
- SCEVHandle One = SE.getIntegerSCEV(1, RHS->getType());
- if (AddRec->getOperand(1) != One)
+ SCEVHandle Stride = AddRec->getOperand(1);
+ if (potentialInfiniteLoop(Stride, RHS, isSigned, trueWhenEqual))
return UnknownValue;
- // We know the LHS is of the form {n,+,1} and the RHS is some loop-invariant
- // m. So, we count the number of iterations in which {n,+,1} < m is true.
- // Note that we cannot simply return max(m-n,0) because it's not safe to
+ // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant
+ // m. So, we count the number of iterations in which {n,+,s} < m is true.
+ // Note that we cannot simply return max(m-n,0)/s because it's not safe to
// treat m-n as signed nor unsigned due to overflow possibility.
// First, we get the value of the LHS in the first iteration: n
SCEVHandle Start = AddRec->getOperand(0);
- if (executesAtLeastOnce(L, isSigned,
- SE.getMinusSCEV(AddRec->getOperand(0), One), RHS)) {
- // Since we know that the condition is true in order to enter the loop,
- // we know that it will run exactly m-n times.
- return SE.getMinusSCEV(RHS, Start);
- } else {
- // Then, we get the value of the LHS in the first iteration in which the
- // above condition doesn't hold. This equals to max(m,n).
- SCEVHandle End = isSigned ? SE.getSMaxExpr(RHS, Start)
- : SE.getUMaxExpr(RHS, Start);
-
- // Finally, we subtract these two values to get the number of times the
- // backedge is executed: max(m,n)-n.
- return SE.getMinusSCEV(End, Start);
+ SCEVHandle One = SE.getIntegerSCEV(1, RHS->getType());
+
+ // Assuming that the loop will run at least once, we know that it will
+ // run (m-n)/s times.
+ SCEVHandle End = RHS;
+
+ if (!executesAtLeastOnce(L, isSigned, trueWhenEqual,
+ SE.getMinusSCEV(Start, One), RHS)) {
+ // If not, we get the value of the LHS in the first iteration in which
+ // the above condition doesn't hold. This equals to max(m,n).
+ End = isSigned ? SE.getSMaxExpr(RHS, Start)
+ : SE.getUMaxExpr(RHS, Start);
}
+
+ // If the expression is less-than-or-equal to, we need to extend the
+ // loop by one iteration.
+ //
+ // The loop won't actually run (m-n)/s times because the loop iterations
+ // might not divide cleanly. For example, if you have {2,+,5} u< 10 the
+ // division would equal one, but the loop runs twice putting the
+ // induction variable at 12.
+
+ if (!trueWhenEqual)
+ // (Stride - 1) is correct only because we know it's unsigned.
+ // What we really want is to decrease the magnitude of Stride by one.
+ Start = SE.getMinusSCEV(Start, SE.getMinusSCEV(Stride, One));
+ else
+ Start = SE.getMinusSCEV(Start, Stride);
+
+ // Finally, we subtract these two values to get the number of times the
+ // backedge is executed: max(m,n)-n.
+ return SE.getUDivExpr(SE.getMinusSCEV(End, Start), Stride);
}
return UnknownValue;
}
}
- // Fallback, if this is a general polynomial, figure out the progression
- // through brute force: evaluate until we find an iteration that fails the
- // test. This is likely to be slow, but getting an accurate trip count is
- // incredibly important, we will be able to simplify the exit test a lot, and
- // we are almost guaranteed to get a trip count in this case.
- ConstantInt *TestVal = ConstantInt::get(getType(), 0);
- ConstantInt *EndVal = TestVal; // Stop when we wrap around.
- do {
- ++NumBruteForceEvaluations;
- SCEVHandle Val = evaluateAtIteration(SE.getConstant(TestVal), SE);
- if (!isa<SCEVConstant>(Val)) // This shouldn't happen.
- return new SCEVCouldNotCompute();
-
- // Check to see if we found the value!
- if (!Range.contains(cast<SCEVConstant>(Val)->getValue()->getValue()))
- return SE.getConstant(TestVal);
-
- // Increment to test the next index.
- TestVal = ConstantInt::get(TestVal->getValue()+1);
- } while (TestVal != EndVal);
-
return new SCEVCouldNotCompute();
}
for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
if (I->getType()->isInteger()) {
OS << *I;
- OS << " --> ";
+ OS << " --> ";
SCEVHandle SV = getSCEV(&*I);
SV->print(OS);
OS << "\t\t";