public Cost computeRefCost(LoopHeaderChord loop, SubscriptExpr se) {
Cost tripL = loop.getTripCount();
InductionVar ivar = loop.getPrimaryInductionVar();
if (ivar == null) return tripL;
// Compute
// tripL = (ubL - lbL + stepL) / stepL
int nsubs = se.numSubscripts(); // number of subscripts in the subscript
LoopHeaderChord tloop = loop.getTopLoop();
VariableDecl iv = ivar.getVar();
for (int i = 0; i < nsubs - 1; i++) { // for c code
// If the loop index variable of this loop appears in any of the
// subscripts other then the last, then return tripL.
Expr sI = se.getSubscript(i);
AffineExpr ae = tloop.isAffine(sI);
if (ae == null) return null;
if (ae.hasTerm(iv)) return tripL;
}
int fs = nsubs - 1;
Expr s0 = se.getSubscript(fs);
AffineExpr ae = tloop.isAffine(s0);
if (ae == null) return null;
int at = ae.getTermIndexOrig(iv);
long coeff = 0;
if (at >= 0) coeff = ae.getCoefficient(at);
if (coeff == 0) // Invariant Reuse.
return new Cost(1.0, 0);
long stepL = loop.getStepValue();
long stride = stepL * coeff;
if (stride < 0) stride = -stride;
// Unit Reuse.
Type et = se.getCoreType().getPointedTo();
int bs = et.memorySizeAsInt(Machine.currentMachine);
int cs = Machine.currentMachine.getCacheSize(bs);
if (stride <= cs) { // cache line or block size
tripL.multiply(stride);
tripL.divide(cs);
return tripL;
}
// No Reuse.
return tripL;
}
public void perform() {
if (trace) System.out.println("** LP " + scribble.getRoutineDecl().getName());
LoopHeaderChord lt = scribble.getLoopTree();
Vector<LoopHeaderChord> innerLoops = lt.getInnerLoops();
for (int li = 0; li < innerLoops.size(); li++) {
LoopHeaderChord loop = innerLoops.elementAt(li);
InductionVar ivar = loop.getPrimaryInductionVar();
if (trace) System.out.println(" lp " + loop.nestedLevel() + " " + loop);
if ((ivar == null) || !loop.isPerfectlyNested()) {
innerLoops.addVectors(loop.getInnerLoops());
continue;
}
if (loop.nestedLevel() < 2) // no need to check permutation for a simple nest
continue;
tryPermute(loop);
}
}