public String toString() {
StringBuffer buf = new StringBuffer("{RG ");
int len = refs.size();
if (len > 3) len = 3;
for (int i = 0; i < len; i++) {
if (i > 0) buf.append(',');
buf.append(refs.elementAt(i));
}
if (len < refs.size()) buf.append(", ...");
buf.append('}');
return buf.toString();
}
/**
* This routine is called to split blocks during register allocation. At this point we know that
* the block was legal before the spill stores and loads were inserted. For every hyperblock with
* spills that is no longer legal, we pick its largest predicate block and split that in half. One
* half is then placed into a new hyperblock.
*
* <p>The case that was causing trouble results from the classic diamond-shaped predicate flow
* graph:
*
* <pre>
* a
* / \
* b c
* \ /
* d
* </pre>
*
* where predicate block <code>a</code> contains most of the instructions, blocks <code>b</code>
* and <code>c</code> each contain a predicated branch, and block <code>d</code> is empty. Because
* of spilling, the hyperblock was too big but all of the predicate blocks were legal. As a result
* we attempted to split block <code>d</code> which did no good.
*
* @param blocks is a list of hyperblocks that have had spills inserted
* @return true if a block was split
*/
public final boolean splitBlocksWithSpills(Hyperblock hbStart, Vector<Hyperblock> blocks) {
// If all the blocks with spills are legal we are done.
boolean illegal = false;
int bs = blocks.size();
for (int i = 0; i < bs; i++) {
Hyperblock hb = blocks.get(i);
if (!hb.isLegalBlock(true)) {
illegal = true;
}
}
if (!illegal) {
return false;
}
// Remove any spill code because the register allocator will run again.
// And remove any blocks which are legal from the set to split.
for (int i = bs - 1; i > -1; i--) {
Hyperblock hb = blocks.get(i);
if (hb.isLegalBlock(true)) {
blocks.remove(i);
}
hb.removeSpillCode();
workingSet.add(hb);
}
// Split hyperblocks that have violations.
int bl = blocks.size();
for (int i = bl - 1; i > -1; i--) {
Hyperblock hb = blocks.remove(i);
splitHyperblock(hb);
}
// The register allocator and block splitter are going to run again.
// Analyze all the hyperblocks that were created or changed.
DataflowAnalysis df = new DataflowAnalysis(hbStart, regs);
df.computeLiveness3();
int sl = workingSet.size();
for (int i = sl - 1; i > -1; i--) {
Hyperblock hb = workingSet.remove(i);
hb.enterSSA();
hb.analyzeLeaveSSA();
}
return true;
}
public void getDeclList(AbstractCollection<Declaration> varList) {
int l = argList.size();
for (int i = 0; i < l; i++) {
Expression exp = argList.get(i);
exp.getDeclList(varList);
}
}
/**
* Compute the entry points for a list of hyperblocks to be inserted into the HFG. This method
* also removes the link between hbStart and all its successors.
*/
public static HashMap<Instruction, Hyperblock> computeEntries(
Hyperblock hbStart, Vector<Hyperblock> hbs) {
HashMap<Instruction, Hyperblock> entries = new HashMap<Instruction, Hyperblock>(203);
// Find the entry points in the new hyperblocks.
int l = hbs.size();
for (int i = 0; i < l; i++) {
Hyperblock hb = hbs.elementAt(i);
PredicateBlock start = hb.getFirstBlock();
Instruction first = start.getFirstInstruction();
entries.put(first, hb);
}
// The successors of the original hyperblock are also entry points.
int sl = hbStart.numOutEdges();
for (int i = 0; i < sl; i++) {
Hyperblock succ = (Hyperblock) hbStart.getOutEdge(0);
PredicateBlock start = succ.getFirstBlock();
Instruction first = start.getFirstInstruction();
hbStart.deleteOutEdge(succ);
succ.deleteInEdge(hbStart);
entries.put(first, succ);
}
return entries;
}
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;
}
/**
* Find a split point for a block with spills. <br>
* We know the hyperblock was legal before the insertion of spill code. Return the "deepest" split
* point in the predicate flow graph, or if there are no split points, the block with the most
* instructions.
*/
private PredicateBlock findSplitPointSpills(Hyperblock hb) {
Vector<PredicateBlock> wl = new Vector<PredicateBlock>();
PredicateBlock start = hb.getFirstBlock();
PredicateBlock last = hb.getLastBlock();
PredicateBlock biggest = null;
PredicateBlock splitPoint = null;
int biggestSize = 0;
if (start.numOutEdges() == 0) {
return start;
}
start.nextVisit();
start.setVisited();
wl.add(start);
// Find the block with the most "real" instructions.
// We do not include fanout, nulls, or spill code.
while (!wl.isEmpty()) {
int sl = wl.size();
for (int i = 0; i < sl; i++) {
PredicateBlock block = wl.get(i);
int size = 0;
for (Instruction inst = block.getFirstInstruction(); inst != null; inst = inst.getNext()) {
size++; // TODO? Should we use the real size of the instruction?
}
if (size >= biggestSize) {
if (block != last) {
biggestSize = size;
biggest = block;
}
}
if (block.isSplitPoint() && (block != start)) {
splitPoint = block;
}
}
wl = hb.getNextPFGLevel(wl);
}
return (splitPoint != null) ? splitPoint : biggest;
}
/**
* Create a representation of a range from min to max.
*
* @param min is the lower bound
* @param max is the upper bound
*/
public static Bound create(Expression min, Expression max) {
if (max == null) {
if (min == null) return noBound;
max = LiteralMap.put(0x7ffffff, Machine.currentMachine.getIntegerCalcType());
}
if (bounds != null) {
int n = bounds.size();
for (int i = 0; i < n; i++) {
Bound ta = bounds.elementAt(i);
if (!min.equivalent(ta.min)) continue;
if (!max.equivalent(ta.max)) continue;
return ta;
}
}
Bound a = new Bound(min, max);
return a;
}
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);
}
}
/** Find the predicate block in a hyperblock to split. */
private PredicateBlock findSplitPoint(Hyperblock hb) {
Vector<PredicateBlock> wl = new Vector<PredicateBlock>();
int totalSize = hb.getFanout() + hb.getBlockSize();
int splits = (totalSize / Trips2Machine.maxBlockSize) + 1;
int splitSize = totalSize / splits;
int hbSize = 0;
PredicateBlock start = hb.getFirstBlock();
PredicateBlock lastUnpredicated = null;
int lastUnpredicatedHBSize = 0;
assert (hb.numSpills() == 0) : "This method should not be called for blocks with spills.";
start.nextVisit();
start.setVisited();
wl.add(start);
while (!wl.isEmpty()) {
int l = wl.size();
int levelSize = 0;
int levelLSID = 0;
// Compute the statistics for this level of the PFG.
for (int i = 0; i < l; i++) {
PredicateBlock block = wl.get(i);
int blockSize = block.getBlockSize() + block.getFanout();
int id = block.maxLSID();
levelSize += blockSize;
if (id > levelLSID) {
levelLSID = id;
}
// Remember the block and the hyperblock size if this block is unpredicated
// and not the special exit block.
// TODO - Can we remove the restriction on being the last block now?
if (!block.isPredicated()) {
if (block.numOutEdges() > 0) {
if (lastUnpredicatedHBSize < (blockSize + hbSize)) {
lastUnpredicatedHBSize = blockSize + hbSize;
lastUnpredicated = block;
}
}
}
}
// Determine if all the blocks can be added to the hyperblock.
int size = hbSize + levelSize;
if ((size > Trips2Machine.maxBlockSize) || (levelLSID >= Trips2Machine.maxLSQEntries)) {
break;
}
hbSize = size;
wl = hb.getNextPFGLevel(wl);
}
assert (!wl.isEmpty()) : "This block does not need to be split?";
// If there is only one unpredicated block in the level and it is
// not the special exit block use it. Or if this is the only
// block in the PFG.
int l = wl.size();
if (l == 1) {
PredicateBlock block = wl.get(0);
if (!block.isPredicated()) {
if ((start == block) || (block.numOutEdges() > 0)) {
// System.out.println("block");
return block;
}
}
}
// Is there a last known unpredicated block of adequate size use it.
if (lastUnpredicated != null) {
if (lastUnpredicatedHBSize >= splitSize) {
// System.out.println("*** last unpred is greater than split sz " + splitSize);
return lastUnpredicated;
}
}
// Try to find a parent that's unpredicated unless the parent is
// the first block.
for (int i = 0; i < l; i++) {
PredicateBlock block = wl.get(i);
int pl = block.numInEdges();
for (int j = 0; j < pl; j++) {
PredicateBlock pred = (PredicateBlock) block.getInEdge(j);
if (!pred.isPredicated() && pred.numInEdges() > 1) {
// System.out.println("unpred parent not start");
return pred;
}
}
}
// Reverse if-convert the largest block in this level which is not
// an exit. Although this seems like a good idea, there is not
// always enough room in the hyperblock to fanout the live-outs to
// the write instructions. Don't do this for now. -- Aaron
PredicateBlock candidate = null;
int largest = 0;
// for (int i = 0; i < l; i++) {
// PredicateBlock block = (PredicateBlock) wl.get(i);
// int bsize = block.getBlockSize() + block.getFanout() + block.getSpillSize();
// if ((bsize > largest) && (block.numBranches() == 0)) {
// largest = bsize;
// candidate = block;
// }
// }
// if (candidate != null) {
// //System.out.println("level no exit");
// return candidate;
// }
// Reverse if-convert a parent which is not start.
// Prefer parents that are split points.
for (int i = 0; i < l; i++) {
PredicateBlock block = wl.get(i);
int pl = block.numInEdges();
for (int j = 0; j < pl; j++) {
PredicateBlock pred = (PredicateBlock) block.getInEdge(j);
if (pred != start) {
if (pred.isSplitPoint()) {
// System.out.println("pred out isSplit not start");
return pred;
}
candidate = pred;
}
}
}
if (candidate != null) {
// System.out.println("pred out not start");
return candidate;
}
// Reverse if-convert the largest successor of start without an exit.
largest = 0;
for (int i = 0; i < start.numOutEdges(); i++) {
PredicateBlock block = (PredicateBlock) start.getOutEdge(i);
int bsize = block.getBlockSize() + block.getFanout() + block.getSpillSize();
if ((bsize > largest) && !block.hasBranch()) {
largest = bsize;
candidate = block;
}
}
if (candidate != null) {
// System.out.println("start successor no exit");
return candidate;
}
// System.out.println("1st start successor ?");
return (PredicateBlock) start.getOutEdge(0);
}
/** Return the number of AST children of this node. */
public int numChildren() {
return 1 + labels.size();
}
public final int numLabels() {
return labels.size();
}
/**
* Compute the reference groups for this loop. Two array references are in the same group with
* respect to this loop if - there is a loop-independent dependence between them, or - the
* dependence distance(dependence vector entry dl) for this loop is less than some constant
* *dist*, and all other dependence vector entries are 0. - the two array refer to the same array
* and differ by at most *dist2* in the first subscript dimension, where d is less than or equal
* to the cache line size in terms of array elements. All other subscripts must be identical.
* Notes: Here we assume dist1 = dist2 = 2
*/
private void computeRefGroups(
int level,
int dist1,
int dist2,
Table<Declaration, SubscriptExpr> arrayRefs,
Vector<RefGroup> refGroups) {
if (arrayRefs == null) return;
Enumeration<Declaration> ek = arrayRefs.keys();
while (ek.hasMoreElements()) {
VariableDecl vd = (VariableDecl) ek.nextElement();
String s = vd.getName(); // array name
Object[] v = arrayRefs.getRowArray(vd); // vector of SubscriptExpr's
int vi = v.length;
for (int j = vi - 1; j >= 0; j--) {
SubscriptExpr sr = (SubscriptExpr) v[j];
Vector<LoopHeaderChord> allRelatedLoops =
sr.allRelatedLoops(); // ** Incorrect when something like a[(j+i][j]
int arls = allRelatedLoops.size();
int firstsub = arls - 1; // ** Making an invalid assumption here
// Process the list of references r' with which r has a data
// dependence, and r is the source(data flows from r to r').
RefGroup rg = new RefGroup(s, sr);
Object[] edges = graph.getEdges(sr);
int len = edges.length;
for (int i = 0; i < len; i++) {
DDEdge edge = (DDEdge) edges[i];
if (edge.isSpatial()) continue;
// Condition(1)-(a) in McKinley's paper
if (edge.isLoopIndependentDependency()) { // add rP to the RefGroup of r:
rg.add(edge);
continue;
}
// Condition(1)-(b) in McKinley's paper
computeEdgeRefs(edge, rg, level, dist1);
if (arls <= 0) continue;
// Condition(2) in McKinley's paper
// rlevel is the level of the loop related to the first subscript.
int rlevel = allRelatedLoops.elementAt(firstsub).getNestedLevel();
computeEdgeRefs(edge, rg, rlevel, dist2);
}
boolean isInExistingRefGroups = false;
int rgl = refGroups.size();
for (int i = 0; i < rgl; i++) {
RefGroup rg2 = refGroups.elementAt(i);
if (!rg2.getName().equals(s)) continue;
isInExistingRefGroups = rg2.contains(rg);
if (isInExistingRefGroups) {
rg2.add(rg);
break;
}
}
if (!isInExistingRefGroups) refGroups.addElement(rg);
}
}
}
private void printRefGroups(Vector<RefGroup> refGroups) {
for (int i = 0; i < refGroups.size(); i++) {
RefGroup rg = refGroups.elementAt(i);
System.out.println(rg);
}
}
public SubscriptExpr getRepresentative() {
if (refs.size() > 0) return (SubscriptExpr) refs.elementAt(0);
return null;
}
/** Reverse if-convert the given predicate block from the hyperblock. */
private void reverseIfConvert(Hyperblock hb, PredicateBlock start) {
Stack<Node> wl = WorkArea.<Node>getStack("reverseIfConvert");
Stack<PredicateBlock> reverse = WorkArea.<PredicateBlock>getStack("reverseIfConvert");
Vector<PredicateBlock> blocks = new Vector<PredicateBlock>();
Vector<Hyperblock> hbs = new Vector<Hyperblock>();
// Find the blocks which need to be reverse if-converted.
start.nextVisit();
start.setVisited();
wl.add(start);
while (!wl.isEmpty()) {
PredicateBlock block = (PredicateBlock) wl.pop();
block.pushOutEdges(wl);
for (int i = 0; i < block.numInEdges(); i++) {
PredicateBlock pred = (PredicateBlock) block.getInEdge(i);
if (!pred.visited()) {
blocks.add(block);
break;
} else if (blocks.contains(pred) && block.numInEdges() > 1) {
blocks.add(block);
break;
}
}
}
// Order the blocks to reverse if-convert based on their depth from the root.
PredicateBlock head = hb.getFirstBlock();
Vector<PredicateBlock> wl2 = new Vector<PredicateBlock>();
head.nextVisit();
head.setVisited();
wl2.add(head);
while (!wl2.isEmpty()) {
int l = wl2.size();
for (int i = 0; i < l; i++) {
PredicateBlock block = wl2.get(i);
if (blocks.contains(block)) {
blocks.remove(block);
reverse.push(block);
}
}
wl2 = hb.getNextPFGLevel(wl2);
}
// Remove the special "dummy" last block from the PFG.
PredicateBlock last = hb.getLastBlock();
assert (last.numOutEdges() == 0 && !last.isPredicated());
if (last.getFirstInstruction() == null) {
for (int i = last.numInEdges() - 1; i > -1; i--) {
PredicateBlock pred = (PredicateBlock) last.getInEdge(i);
pred.deleteOutEdge(last);
last.deleteInEdge(pred);
}
reverse.remove(last);
}
// Reverse if-convert.
while (!reverse.isEmpty()) {
PredicateBlock block = reverse.pop();
Hyperblock hbn = reverseIfConvertBlock(block);
hbs.add(hbn);
workingSet.add(hbn);
}
// Update the PFG.
hb.updateLastBlock();
hb.invalidateDomination(); // The dominators are now invalid.
// Insert the new hyperblocks in the HFG.
HashMap<Instruction, Hyperblock> entries = computeEntries(hb, hbs);
hbs.add(hb);
Hyperblock.computeHyperblockFlowGraph(hbs, entries);
// Update the return block. Since 'hbs' is an ordered list, the
// first element in the list is the hyperblock with the return
// because this was the original tail of the PFG which was reverse
// if-converted.
if (hb == gen.getReturnBlock()) {
gen.setReturnBlock(hbs.firstElement());
}
WorkArea.<Node>returnStack(wl);
WorkArea.<PredicateBlock>returnStack(reverse);
}
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);
}
}
}
/** Return the number of AST children of this node. */
public int numChildren() {
return 1 + argList.size();
}