#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/ConstantFolding.h"
-#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/InstructionSimplify.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/Analysis/ValueTracking.h"
-#include "llvm/Assembly/Writer.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Dominators.h"
#include "llvm/IR/GlobalAlias.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/Instructions.h"
INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution",
"Scalar Evolution Analysis", false, true)
INITIALIZE_PASS_DEPENDENCY(LoopInfo)
-INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution",
"Scalar Evolution Analysis", false, true)
void SCEV::print(raw_ostream &OS) const {
switch (getSCEVType()) {
case scConstant:
- WriteAsOperand(OS, cast<SCEVConstant>(this)->getValue(), false);
+ cast<SCEVConstant>(this)->getValue()->printAsOperand(OS, false);
return;
case scTruncate: {
const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(this);
if (AR->getNoWrapFlags(FlagNW) &&
!AR->getNoWrapFlags((NoWrapFlags)(FlagNUW | FlagNSW)))
OS << "nw><";
- WriteAsOperand(OS, AR->getLoop()->getHeader(), /*PrintType=*/false);
+ AR->getLoop()->getHeader()->printAsOperand(OS, /*PrintType=*/false);
OS << ">";
return;
}
Constant *FieldNo;
if (U->isOffsetOf(CTy, FieldNo)) {
OS << "offsetof(" << *CTy << ", ";
- WriteAsOperand(OS, FieldNo, false);
+ FieldNo->printAsOperand(OS, false);
OS << ")";
return;
}
// Otherwise just print it normally.
- WriteAsOperand(OS, U->getValue(), false);
+ U->getValue()->printAsOperand(OS, false);
return;
}
case scCouldNotCompute:
// Examine all exits and pick the most conservative values.
const SCEV *MaxBECount = getCouldNotCompute();
bool CouldComputeBECount = true;
+ BasicBlock *Latch = L->getLoopLatch(); // may be NULL.
+ const SCEV *LatchMaxCount = 0;
SmallVector<std::pair<BasicBlock *, const SCEV *>, 4> ExitCounts;
for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
ExitLimit EL = ComputeExitLimit(L, ExitingBlocks[i]);
// We cannot take the "min" MaxBECount, because non-unit stride loops may
// skip some loop tests. Taking the max over the exits is sufficiently
// conservative. TODO: We could do better taking into consideration
- // that (1) the loop has unit stride (2) the last loop test is
- // less-than/greater-than (3) any loop test is less-than/greater-than AND
- // falls-through some constant times less then the other tests.
- MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max);
+ // non-latch exits that dominate the latch.
+ if (EL.MustExit && ExitingBlocks[i] == Latch)
+ LatchMaxCount = EL.Max;
+ else
+ MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max);
}
}
-
+ // Be more precise in the easy case of a loop latch that must exit.
+ if (LatchMaxCount) {
+ MaxBECount = getUMinFromMismatchedTypes(MaxBECount, LatchMaxCount);
+ }
return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount);
}
IsSubExpr || EitherMayExit);
const SCEV *BECount = getCouldNotCompute();
const SCEV *MaxBECount = getCouldNotCompute();
+ bool MustExit = false;
if (EitherMayExit) {
// Both conditions must be true for the loop to continue executing.
// Choose the less conservative count.
MaxBECount = EL0.Max;
else
MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
+ MustExit = EL0.MustExit || EL1.MustExit;
} else {
// Both conditions must be true at the same time for the loop to exit.
// For now, be conservative.
MaxBECount = EL0.Max;
if (EL0.Exact == EL1.Exact)
BECount = EL0.Exact;
+ MustExit = EL0.MustExit && EL1.MustExit;
}
- return ExitLimit(BECount, MaxBECount);
+ return ExitLimit(BECount, MaxBECount, MustExit);
}
if (BO->getOpcode() == Instruction::Or) {
// Recurse on the operands of the or.
IsSubExpr || EitherMayExit);
const SCEV *BECount = getCouldNotCompute();
const SCEV *MaxBECount = getCouldNotCompute();
+ bool MustExit = false;
if (EitherMayExit) {
// Both conditions must be false for the loop to continue executing.
// Choose the less conservative count.
MaxBECount = EL0.Max;
else
MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max);
+ MustExit = EL0.MustExit || EL1.MustExit;
} else {
// Both conditions must be false at the same time for the loop to exit.
// For now, be conservative.
MaxBECount = EL0.Max;
if (EL0.Exact == EL1.Exact)
BECount = EL0.Exact;
+ MustExit = EL0.MustExit && EL1.MustExit;
}
- return ExitLimit(BECount, MaxBECount);
+ return ExitLimit(BECount, MaxBECount, MustExit);
}
}
else
MaxBECount = getConstant(CountDown ? CR.getUnsignedMax()
: -CR.getUnsignedMin());
- return ExitLimit(Distance, MaxBECount);
+ return ExitLimit(Distance, MaxBECount, /*MustExit=*/true);
}
// If the recurrence is known not to wraparound, unsigned divide computes the
// that the value will either become zero (and thus the loop terminates), that
// the loop will terminate through some other exit condition first, or that
// the loop has undefined behavior. This means we can't "miss" the exit
- // value, even with nonunit stride.
+ // value, even with nonunit stride, and exit later via the same branch. Note
+ // that we can skip this exit if loop later exits via a different
+ // branch. Hence MustExit=false.
//
// This is only valid for expressions that directly compute the loop exit. It
// is invalid for subexpressions in which the loop may exit through this
// branch even if this subexpression is false. In that case, the trip count
// computed by this udiv could be smaller than the number of well-defined
// iterations.
- if (!IsSubExpr && AddRec->getNoWrapFlags(SCEV::FlagNW))
- return getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
-
+ if (!IsSubExpr && AddRec->getNoWrapFlags(SCEV::FlagNW)) {
+ const SCEV *Exact =
+ getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
+ return ExitLimit(Exact, Exact, /*MustExit=*/false);
+ }
// Then, try to solve the above equation provided that Start is constant.
if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),
// LHS' type is checked for above.
if (getTypeSizeInBits(LHS->getType()) >
getTypeSizeInBits(FoundLHS->getType())) {
- if (CmpInst::isSigned(Pred)) {
+ if (CmpInst::isSigned(FoundPred)) {
FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType());
FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType());
} else {
if (isa<SCEVCouldNotCompute>(MaxBECount))
MaxBECount = BECount;
- return ExitLimit(BECount, MaxBECount);
+ return ExitLimit(BECount, MaxBECount, /*MustExit=*/true);
}
ScalarEvolution::ExitLimit
if (isa<SCEVCouldNotCompute>(MaxBECount))
MaxBECount = BECount;
- return ExitLimit(BECount, MaxBECount);
+ return ExitLimit(BECount, MaxBECount, /*MustExit=*/true);
}
/// getNumIterationsInRange - Return the number of iterations of this loop that
return SE.getCouldNotCompute();
}
+static const APInt gcd(const SCEVConstant *C1, const SCEVConstant *C2) {
+ APInt A = C1->getValue()->getValue().abs();
+ APInt B = C2->getValue()->getValue().abs();
+ uint32_t ABW = A.getBitWidth();
+ uint32_t BBW = B.getBitWidth();
+
+ if (ABW > BBW)
+ B = B.zext(ABW);
+ else if (ABW < BBW)
+ A = A.zext(BBW);
+
+ return APIntOps::GreatestCommonDivisor(A, B);
+}
+
+static const APInt srem(const SCEVConstant *C1, const SCEVConstant *C2) {
+ APInt A = C1->getValue()->getValue();
+ APInt B = C2->getValue()->getValue();
+ uint32_t ABW = A.getBitWidth();
+ uint32_t BBW = B.getBitWidth();
+
+ if (ABW > BBW)
+ B = B.sext(ABW);
+ else if (ABW < BBW)
+ A = A.sext(BBW);
+
+ return APIntOps::srem(A, B);
+}
+
+static const APInt sdiv(const SCEVConstant *C1, const SCEVConstant *C2) {
+ APInt A = C1->getValue()->getValue();
+ APInt B = C2->getValue()->getValue();
+ uint32_t ABW = A.getBitWidth();
+ uint32_t BBW = B.getBitWidth();
+
+ if (ABW > BBW)
+ B = B.sext(ABW);
+ else if (ABW < BBW)
+ A = A.sext(BBW);
+
+ return APIntOps::sdiv(A, B);
+}
+
+namespace {
+struct SCEVGCD : public SCEVVisitor<SCEVGCD, const SCEV *> {
+public:
+ // Pattern match Step into Start. When Step is a multiply expression, find
+ // the largest subexpression of Step that appears in Start. When Start is an
+ // add expression, try to match Step in the subexpressions of Start, non
+ // matching subexpressions are returned under Remainder.
+ static const SCEV *findGCD(ScalarEvolution &SE, const SCEV *Start,
+ const SCEV *Step, const SCEV **Remainder) {
+ assert(Remainder && "Remainder should not be NULL");
+ SCEVGCD R(SE, Step, SE.getConstant(Step->getType(), 0));
+ const SCEV *Res = R.visit(Start);
+ *Remainder = R.Remainder;
+ return Res;
+ }
+
+ SCEVGCD(ScalarEvolution &S, const SCEV *G, const SCEV *R)
+ : SE(S), GCD(G), Remainder(R) {
+ Zero = SE.getConstant(GCD->getType(), 0);
+ One = SE.getConstant(GCD->getType(), 1);
+ }
+
+ const SCEV *visitConstant(const SCEVConstant *Constant) {
+ if (GCD == Constant || Constant == Zero)
+ return GCD;
+
+ if (const SCEVConstant *CGCD = dyn_cast<SCEVConstant>(GCD)) {
+ const SCEV *Res = SE.getConstant(gcd(Constant, CGCD));
+ if (Res != One)
+ return Res;
+
+ Remainder = SE.getConstant(srem(Constant, CGCD));
+ Constant = cast<SCEVConstant>(SE.getMinusSCEV(Constant, Remainder));
+ Res = SE.getConstant(gcd(Constant, CGCD));
+ return Res;
+ }
+
+ // When GCD is not a constant, it could be that the GCD is an Add, Mul,
+ // AddRec, etc., in which case we want to find out how many times the
+ // Constant divides the GCD: we then return that as the new GCD.
+ const SCEV *Rem = Zero;
+ const SCEV *Res = findGCD(SE, GCD, Constant, &Rem);
+
+ if (Res == One || Rem != Zero) {
+ Remainder = Constant;
+ return One;
+ }
+
+ assert(isa<SCEVConstant>(Res) && "Res should be a constant");
+ Remainder = SE.getConstant(srem(Constant, cast<SCEVConstant>(Res)));
+ return Res;
+ }
+
+ const SCEV *visitTruncateExpr(const SCEVTruncateExpr *Expr) {
+ if (GCD != Expr)
+ Remainder = Expr;
+ return GCD;
+ }
+
+ const SCEV *visitZeroExtendExpr(const SCEVZeroExtendExpr *Expr) {
+ if (GCD != Expr)
+ Remainder = Expr;
+ return GCD;
+ }
+
+ const SCEV *visitSignExtendExpr(const SCEVSignExtendExpr *Expr) {
+ if (GCD != Expr)
+ Remainder = Expr;
+ return GCD;
+ }
+
+ const SCEV *visitAddExpr(const SCEVAddExpr *Expr) {
+ if (GCD == Expr)
+ return GCD;
+
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
+ const SCEV *Rem = Zero;
+ const SCEV *Res = findGCD(SE, Expr->getOperand(e - 1 - i), GCD, &Rem);
+
+ // FIXME: There may be ambiguous situations: for instance,
+ // GCD(-4 + (3 * %m), 2 * %m) where 2 divides -4 and %m divides (3 * %m).
+ // The order in which the AddExpr is traversed computes a different GCD
+ // and Remainder.
+ if (Res != One)
+ GCD = Res;
+ if (Rem != Zero)
+ Remainder = SE.getAddExpr(Remainder, Rem);
+ }
+
+ return GCD;
+ }
+
+ const SCEV *visitMulExpr(const SCEVMulExpr *Expr) {
+ if (GCD == Expr)
+ return GCD;
+
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
+ if (Expr->getOperand(i) == GCD)
+ return GCD;
+ }
+
+ // If we have not returned yet, it means that GCD is not part of Expr.
+ const SCEV *PartialGCD = One;
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
+ const SCEV *Rem = Zero;
+ const SCEV *Res = findGCD(SE, Expr->getOperand(i), GCD, &Rem);
+ if (Rem != Zero)
+ // GCD does not divide Expr->getOperand(i).
+ continue;
+
+ if (Res == GCD)
+ return GCD;
+ PartialGCD = SE.getMulExpr(PartialGCD, Res);
+ if (PartialGCD == GCD)
+ return GCD;
+ }
+
+ if (PartialGCD != One)
+ return PartialGCD;
+
+ Remainder = Expr;
+ const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(GCD);
+ if (!Mul)
+ return PartialGCD;
+
+ // When the GCD is a multiply expression, try to decompose it:
+ // this occurs when Step does not divide the Start expression
+ // as in: {(-4 + (3 * %m)),+,(2 * %m)}
+ for (int i = 0, e = Mul->getNumOperands(); i < e; ++i) {
+ const SCEV *Rem = Zero;
+ const SCEV *Res = findGCD(SE, Expr, Mul->getOperand(i), &Rem);
+ if (Rem == Zero) {
+ Remainder = Rem;
+ return Res;
+ }
+ }
+
+ return PartialGCD;
+ }
+
+ const SCEV *visitUDivExpr(const SCEVUDivExpr *Expr) {
+ if (GCD != Expr)
+ Remainder = Expr;
+ return GCD;
+ }
+
+ const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) {
+ if (GCD == Expr)
+ return GCD;
+
+ if (!Expr->isAffine()) {
+ Remainder = Expr;
+ return GCD;
+ }
+
+ const SCEV *Rem = Zero;
+ const SCEV *Res = findGCD(SE, Expr->getOperand(0), GCD, &Rem);
+ if (Rem != Zero)
+ Remainder = SE.getAddExpr(Remainder, Rem);
+
+ Rem = Zero;
+ Res = findGCD(SE, Expr->getOperand(1), Res, &Rem);
+ if (Rem != Zero) {
+ Remainder = Expr;
+ return GCD;
+ }
+
+ return Res;
+ }
+
+ const SCEV *visitSMaxExpr(const SCEVSMaxExpr *Expr) {
+ if (GCD != Expr)
+ Remainder = Expr;
+ return GCD;
+ }
+
+ const SCEV *visitUMaxExpr(const SCEVUMaxExpr *Expr) {
+ if (GCD != Expr)
+ Remainder = Expr;
+ return GCD;
+ }
+
+ const SCEV *visitUnknown(const SCEVUnknown *Expr) {
+ if (GCD != Expr)
+ Remainder = Expr;
+ return GCD;
+ }
+
+ const SCEV *visitCouldNotCompute(const SCEVCouldNotCompute *Expr) {
+ return One;
+ }
+
+private:
+ ScalarEvolution &SE;
+ const SCEV *GCD, *Remainder, *Zero, *One;
+};
+
+struct SCEVDivision : public SCEVVisitor<SCEVDivision, const SCEV *> {
+public:
+ // Remove from Start all multiples of Step.
+ static const SCEV *divide(ScalarEvolution &SE, const SCEV *Start,
+ const SCEV *Step) {
+ SCEVDivision D(SE, Step);
+ const SCEV *Rem = D.Zero;
+ (void)Rem;
+ // The division is guaranteed to succeed: Step should divide Start with no
+ // remainder.
+ assert(Step == SCEVGCD::findGCD(SE, Start, Step, &Rem) && Rem == D.Zero &&
+ "Step should divide Start with no remainder.");
+ return D.visit(Start);
+ }
+
+ SCEVDivision(ScalarEvolution &S, const SCEV *G) : SE(S), GCD(G) {
+ Zero = SE.getConstant(GCD->getType(), 0);
+ One = SE.getConstant(GCD->getType(), 1);
+ }
+
+ const SCEV *visitConstant(const SCEVConstant *Constant) {
+ if (GCD == Constant)
+ return One;
+
+ if (const SCEVConstant *CGCD = dyn_cast<SCEVConstant>(GCD))
+ return SE.getConstant(sdiv(Constant, CGCD));
+ return Constant;
+ }
+
+ const SCEV *visitTruncateExpr(const SCEVTruncateExpr *Expr) {
+ if (GCD == Expr)
+ return One;
+ return Expr;
+ }
+
+ const SCEV *visitZeroExtendExpr(const SCEVZeroExtendExpr *Expr) {
+ if (GCD == Expr)
+ return One;
+ return Expr;
+ }
+
+ const SCEV *visitSignExtendExpr(const SCEVSignExtendExpr *Expr) {
+ if (GCD == Expr)
+ return One;
+ return Expr;
+ }
+
+ const SCEV *visitAddExpr(const SCEVAddExpr *Expr) {
+ if (GCD == Expr)
+ return One;
+
+ SmallVector<const SCEV *, 2> Operands;
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i)
+ Operands.push_back(divide(SE, Expr->getOperand(i), GCD));
+
+ if (Operands.size() == 1)
+ return Operands[0];
+ return SE.getAddExpr(Operands);
+ }
+
+ const SCEV *visitMulExpr(const SCEVMulExpr *Expr) {
+ if (GCD == Expr)
+ return One;
+
+ bool FoundGCDTerm = false;
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i)
+ if (Expr->getOperand(i) == GCD)
+ FoundGCDTerm = true;
+
+ SmallVector<const SCEV *, 2> Operands;
+ if (FoundGCDTerm) {
+ FoundGCDTerm = false;
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
+ if (FoundGCDTerm)
+ Operands.push_back(Expr->getOperand(i));
+ else if (Expr->getOperand(i) == GCD)
+ FoundGCDTerm = true;
+ else
+ Operands.push_back(Expr->getOperand(i));
+ }
+ } else {
+ FoundGCDTerm = false;
+ const SCEV *PartialGCD = One;
+ for (int i = 0, e = Expr->getNumOperands(); i < e; ++i) {
+ if (PartialGCD == GCD) {
+ Operands.push_back(Expr->getOperand(i));
+ continue;
+ }
+
+ const SCEV *Rem = Zero;
+ const SCEV *Res = SCEVGCD::findGCD(SE, Expr->getOperand(i), GCD, &Rem);
+ if (Rem == Zero) {
+ PartialGCD = SE.getMulExpr(PartialGCD, Res);
+ Operands.push_back(divide(SE, Expr->getOperand(i), GCD));
+ } else {
+ Operands.push_back(Expr->getOperand(i));
+ }
+ }
+ }
+
+ if (Operands.size() == 1)
+ return Operands[0];
+ return SE.getMulExpr(Operands);
+ }
+
+ const SCEV *visitUDivExpr(const SCEVUDivExpr *Expr) {
+ if (GCD == Expr)
+ return One;
+ return Expr;
+ }
+
+ const SCEV *visitAddRecExpr(const SCEVAddRecExpr *Expr) {
+ if (GCD == Expr)
+ return One;
+
+ assert(Expr->isAffine() && "Expr should be affine");
+
+ const SCEV *Start = divide(SE, Expr->getStart(), GCD);
+ const SCEV *Step = divide(SE, Expr->getStepRecurrence(SE), GCD);
+
+ return SE.getAddRecExpr(Start, Step, Expr->getLoop(),
+ Expr->getNoWrapFlags());
+ }
+
+ const SCEV *visitSMaxExpr(const SCEVSMaxExpr *Expr) {
+ if (GCD == Expr)
+ return One;
+ return Expr;
+ }
+
+ const SCEV *visitUMaxExpr(const SCEVUMaxExpr *Expr) {
+ if (GCD == Expr)
+ return One;
+ return Expr;
+ }
+
+ const SCEV *visitUnknown(const SCEVUnknown *Expr) {
+ if (GCD == Expr)
+ return One;
+ return Expr;
+ }
+
+ const SCEV *visitCouldNotCompute(const SCEVCouldNotCompute *Expr) {
+ return Expr;
+ }
+
+private:
+ ScalarEvolution &SE;
+ const SCEV *GCD, *Zero, *One;
+};
+}
+
+/// Splits the SCEV into two vectors of SCEVs representing the subscripts and
+/// sizes of an array access. Returns the remainder of the delinearization that
+/// is the offset start of the array. The SCEV->delinearize algorithm computes
+/// the multiples of SCEV coefficients: that is a pattern matching of sub
+/// expressions in the stride and base of a SCEV corresponding to the
+/// computation of a GCD (greatest common divisor) of base and stride. When
+/// SCEV->delinearize fails, it returns the SCEV unchanged.
+///
+/// For example: when analyzing the memory access A[i][j][k] in this loop nest
+///
+/// void foo(long n, long m, long o, double A[n][m][o]) {
+///
+/// for (long i = 0; i < n; i++)
+/// for (long j = 0; j < m; j++)
+/// for (long k = 0; k < o; k++)
+/// A[i][j][k] = 1.0;
+/// }
+///
+/// the delinearization input is the following AddRec SCEV:
+///
+/// AddRec: {{{%A,+,(8 * %m * %o)}<%for.i>,+,(8 * %o)}<%for.j>,+,8}<%for.k>
+///
+/// From this SCEV, we are able to say that the base offset of the access is %A
+/// because it appears as an offset that does not divide any of the strides in
+/// the loops:
+///
+/// CHECK: Base offset: %A
+///
+/// and then SCEV->delinearize determines the size of some of the dimensions of
+/// the array as these are the multiples by which the strides are happening:
+///
+/// CHECK: ArrayDecl[UnknownSize][%m][%o] with elements of sizeof(double) bytes.
+///
+/// Note that the outermost dimension remains of UnknownSize because there are
+/// no strides that would help identifying the size of the last dimension: when
+/// the array has been statically allocated, one could compute the size of that
+/// dimension by dividing the overall size of the array by the size of the known
+/// dimensions: %m * %o * 8.
+///
+/// Finally delinearize provides the access functions for the array reference
+/// that does correspond to A[i][j][k] of the above C testcase:
+///
+/// CHECK: ArrayRef[{0,+,1}<%for.i>][{0,+,1}<%for.j>][{0,+,1}<%for.k>]
+///
+/// The testcases are checking the output of a function pass:
+/// DelinearizationPass that walks through all loads and stores of a function
+/// asking for the SCEV of the memory access with respect to all enclosing
+/// loops, calling SCEV->delinearize on that and printing the results.
+const SCEV *
+SCEVAddRecExpr::delinearize(ScalarEvolution &SE,
+ SmallVectorImpl<const SCEV *> &Subscripts,
+ SmallVectorImpl<const SCEV *> &Sizes) const {
+ // Early exit in case this SCEV is not an affine multivariate function.
+ if (!this->isAffine())
+ return this;
+
+ const SCEV *Start = this->getStart();
+ const SCEV *Step = this->getStepRecurrence(SE);
+
+ // Build the SCEV representation of the cannonical induction variable in the
+ // loop of this SCEV.
+ const SCEV *Zero = SE.getConstant(this->getType(), 0);
+ const SCEV *One = SE.getConstant(this->getType(), 1);
+ const SCEV *IV =
+ SE.getAddRecExpr(Zero, One, this->getLoop(), this->getNoWrapFlags());
+
+ DEBUG(dbgs() << "(delinearize: " << *this << "\n");
+
+ // Currently we fail to delinearize when the stride of this SCEV is 1. We
+ // could decide to not fail in this case: we could just return 1 for the size
+ // of the subscript, and this same SCEV for the access function.
+ if (Step == One) {
+ DEBUG(dbgs() << "failed to delinearize " << *this << "\n)\n");
+ return this;
+ }
+
+ // Find the GCD and Remainder of the Start and Step coefficients of this SCEV.
+ const SCEV *Remainder = NULL;
+ const SCEV *GCD = SCEVGCD::findGCD(SE, Start, Step, &Remainder);
+
+ DEBUG(dbgs() << "GCD: " << *GCD << "\n");
+ DEBUG(dbgs() << "Remainder: " << *Remainder << "\n");
+
+ // Same remark as above: we currently fail the delinearization, although we
+ // can very well handle this special case.
+ if (GCD == One) {
+ DEBUG(dbgs() << "failed to delinearize " << *this << "\n)\n");
+ return this;
+ }
+
+ // As findGCD computed Remainder, GCD divides "Start - Remainder." The
+ // Quotient is then this SCEV without Remainder, scaled down by the GCD. The
+ // Quotient is what will be used in the next subscript delinearization.
+ const SCEV *Quotient =
+ SCEVDivision::divide(SE, SE.getMinusSCEV(Start, Remainder), GCD);
+ DEBUG(dbgs() << "Quotient: " << *Quotient << "\n");
+
+ const SCEV *Rem;
+ if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Quotient))
+ // Recursively call delinearize on the Quotient until there are no more
+ // multiples that can be recognized.
+ Rem = AR->delinearize(SE, Subscripts, Sizes);
+ else
+ Rem = Quotient;
+
+ // Scale up the cannonical induction variable IV by whatever remains from the
+ // Step after division by the GCD: the GCD is the size of all the sub-array.
+ if (Step != GCD) {
+ Step = SCEVDivision::divide(SE, Step, GCD);
+ IV = SE.getMulExpr(IV, Step);
+ }
+ // The access function in the current subscript is computed as the cannonical
+ // induction variable IV (potentially scaled up by the step) and offset by
+ // Rem, the offset of delinearization in the sub-array.
+ const SCEV *Index = SE.getAddExpr(IV, Rem);
+
+ // Record the access function and the size of the current subscript.
+ Subscripts.push_back(Index);
+ Sizes.push_back(GCD);
+
+#ifndef NDEBUG
+ int Size = Sizes.size();
+ DEBUG(dbgs() << "succeeded to delinearize " << *this << "\n");
+ DEBUG(dbgs() << "ArrayDecl[UnknownSize]");
+ for (int i = 0; i < Size - 1; i++)
+ DEBUG(dbgs() << "[" << *Sizes[i] << "]");
+ DEBUG(dbgs() << " with elements of " << *Sizes[Size - 1] << " bytes.\n");
+
+ DEBUG(dbgs() << "ArrayRef");
+ for (int i = 0; i < Size; i++)
+ DEBUG(dbgs() << "[" << *Subscripts[i] << "]");
+ DEBUG(dbgs() << "\n)\n");
+#endif
+
+ return Remainder;
+}
//===----------------------------------------------------------------------===//
// SCEVCallbackVH Class Implementation
LI = &getAnalysis<LoopInfo>();
TD = getAnalysisIfAvailable<DataLayout>();
TLI = &getAnalysis<TargetLibraryInfo>();
- DT = &getAnalysis<DominatorTree>();
+ DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
return false;
}
void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequiredTransitive<LoopInfo>();
- AU.addRequiredTransitive<DominatorTree>();
+ AU.addRequiredTransitive<DominatorTreeWrapperPass>();
AU.addRequired<TargetLibraryInfo>();
}
PrintLoopInfo(OS, SE, *I);
OS << "Loop ";
- WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
+ L->getHeader()->printAsOperand(OS, /*PrintType=*/false);
OS << ": ";
SmallVector<BasicBlock *, 8> ExitBlocks;
OS << "\n"
"Loop ";
- WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false);
+ L->getHeader()->printAsOperand(OS, /*PrintType=*/false);
OS << ": ";
if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) {
ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this);
OS << "Classifying expressions for: ";
- WriteAsOperand(OS, F, /*PrintType=*/false);
+ F->printAsOperand(OS, /*PrintType=*/false);
OS << "\n";
for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) {
}
OS << "Determining loop execution counts for: ";
- WriteAsOperand(OS, F, /*PrintType=*/false);
+ F->printAsOperand(OS, /*PrintType=*/false);
OS << "\n";
for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
PrintLoopInfo(OS, &SE, *I);
projects / oota-llvm.git / blobdiff