private boolean sortByCost(Cost[] loopCostList, int[] loopIndex) {
int maxlooptemp = loopCostList.length - 1;
boolean exchanged = true;
boolean permuted = false;
while (exchanged) { // Bubble sort the LoopCosts.
exchanged = false;
for (int i = 0; i < maxlooptemp; i++) {
int j = i + 1;
Cost lci = loopCostList[i];
Cost lci1 = loopCostList[j];
if (lci1 == null) continue;
if (lci.lessThan(lci1)) {
int temp = loopIndex[i];
loopIndex[i] = loopIndex[j];
loopIndex[j] = temp;
loopCostList[i] = lci1;
loopCostList[j] = lci;
exchanged = true;
}
}
permuted |= exchanged;
maxlooptemp--;
}
return permuted;
}
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;
}
private Cost computeRefCostSum(LoopHeaderChord loop, Vector<RefGroup> refGroups) {
Cost lc = new Cost();
int l = refGroups.size();
for (int i = 0; i < l; i++) {
RefGroup rg = refGroups.elementAt(i);
SubscriptExpr se = rg.getRepresentative();
Cost cost = computeRefCost(loop, se);
if (cost == null) continue;
lc.add(cost);
}
return lc;
}
private void tryPermute(LoopHeaderChord topLoop) {
Vector<LoopHeaderChord> loopNest = topLoop.getTightlyNestedLoops();
if (loopNest == null) return;
int loopDepth = loopNest.size();
LoopHeaderChord bottom = loopNest.get(loopDepth - 1);
if (!unsafe && !legalLoop(bottom)) return;
// Set the loop costs for the loops in a loop nest. The cost for a loop is
// the cost of executing the nest with that loop in the innermost nesting.
Table<Declaration, SubscriptExpr> arrayRefs = new Table<Declaration, SubscriptExpr>();
if (!topLoop.getSubscriptsRecursive(arrayRefs)) return;
graph = topLoop.getDDGraph(false);
if (graph == null) return;
if (trace) System.out.println(" " + graph);
int[] loopIndex = new int[loopDepth];
Cost[] loopCostList = new Cost[loopDepth];
Vector<RefGroup> refGroups = new Vector<RefGroup>(20);
for (int i = 0; i < loopDepth; i++) {
LoopHeaderChord loop = loopNest.elementAt(i);
if (trace) System.out.println(" " + i + " " + loop);
computeRefGroups(loop.getNestedLevel(), 2, 2, arrayRefs, refGroups);
Cost lc = computeRefCostSum(loop, refGroups);
if (trace) System.out.println(" " + i + " " + lc);
loopCostList[i] = lc;
loopIndex[i] = i; // the outtermost loop is at position 0
Cost tp = tripProduct(loopNest, loop);
if (trace) System.out.println(" " + i + " " + tp);
lc.multiply(tp);
if (trace) System.out.println(" " + i + " " + lc);
}
boolean permuted = sortByCost(loopCostList, loopIndex);
if (!permuted) return;
if (trace) {
System.out.print(" permute " + loopDepth);
System.out.print(":");
for (int i = 0; i < loopIndex.length; i++) System.out.print(" " + loopIndex[i]);
System.out.println("");
}
int[][] ddVec = getDDVec(arrayRefs, loopDepth);
if (trace) printDDInfo(ddVec, loopDepth);
if (!isLegal(loopIndex, ddVec)) return;
if (trace) System.out.println(" permute " + loopDepth);
int[] rank = new int[loopDepth];
// We will do sorting on the rank vector, which corresponds to the interchange we need.
for (int i = 0; i < loopDepth; i++) {
int loopNum = loopIndex[i];
rank[loopNum] = i;
}
if (trace) printOrder(rank);
boolean changed = true;
while (changed) {
changed = false;
for (int i = 0; i < loopDepth - 1; i++) {
int j = i + 1;
int outerRank = rank[i];
int innerRank = rank[j];
if (innerRank >= outerRank) continue;
LoopHeaderChord innerLoop = loopNest.elementAt(j);
LoopHeaderChord outerLoop = loopNest.elementAt(i);
if (!outerLoop.isDDComplete() || outerLoop.inhibitLoopPermute()) continue;
if (!innerLoop.isDDComplete() || innerLoop.inhibitLoopPermute()) continue;
changed = true;
rank[i] = innerRank;
rank[j] = outerRank;
loopNest.setElementAt(innerLoop, i);
loopNest.setElementAt(outerLoop, j);
performLoopInterchange(innerLoop, outerLoop);
}
}
}