/**
* 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;
}
/**
* 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;
}
/** The main routine for block splitting. */
public final void split(Hyperblock hbStart) {
Stack<Node> wl = WorkArea.<Node>getStack("split");
DataflowAnalysis df = new DataflowAnalysis(hbStart, regs);
int trips = 0;
// Add all the hyperblocks to the working set.
hbStart.nextVisit();
hbStart.setVisited();
wl.push(hbStart);
while (!wl.isEmpty()) {
Hyperblock hb = (Hyperblock) wl.pop();
hb.pushOutEdges(wl);
workingSet.add(hb);
}
// Split the hyperblocks.
while (!workingSet.isEmpty()) {
df.computeLiveness3();
wl.addAll(workingSet);
while (!wl.isEmpty()) {
Hyperblock hb = (Hyperblock) wl.pop();
hb.enterSSA();
hb.analyzeLeaveSSA();
if (!hb.isLegalBlock(true)) {
splitHyperblock(hb);
} else {
workingSet.remove(hb);
}
}
if ((++trips % WARN_SPLIT_ATTEMPTS) == 0) {
System.err.println(
"** Warning: the block splitter has run "
+ trips
+ " times for "
+ gen.getCurrentRoutine().getName()
+ "().");
}
}
WorkArea.<Node>returnStack(wl);
}
/** Split an unpredicated predicate block. */
private void splitBlock(Hyperblock hb, PredicateBlock block) {
int chunkSize = (block.getBlockSize() + block.getFanout()) / 2;
int maxChunk = (int) (Trips2Machine.maxBlockSize * MAXFILL);
if (chunkSize > maxChunk) {
chunkSize = maxChunk;
}
Instruction splitLocation = findSplitLocation(hb, block, chunkSize);
PredicateBlock start = block.cut(splitLocation, gen);
Hyperblock nhb = new Hyperblock(start, regs);
// Insert the new hyperblock into the HFG.
for (int i = hb.numOutEdges() - 1; i > -1; i--) {
Hyperblock out = (Hyperblock) hb.getOutEdge(i);
out.replaceInEdge(hb, nhb);
hb.deleteOutEdge(out);
nhb.addOutEdge(out);
}
hb.addOutEdge(nhb);
nhb.addInEdge(hb);
workingSet.add(nhb);
hb.invalidateDomination();
hb.findLastBlock();
hb.determinePredicatesBranches();
nhb.findLastBlock();
nhb.determinePredicatesBranches();
// Update the return block if it has changed.
if (gen.getReturnBlock() == hb) {
gen.setReturnBlock(nhb);
}
}
/** 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);
}
/** 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);
}