/**
* Search for direct loads in the taken side of a guard
*
* @param firstBlock The guard's branch destination
* @param coldPathLoads BitVector of symbol reference numbers for any direct loads seen until the merge back to mainline
*/
static void collectColdPathLoads(TR::Block* firstBlock, TR_BitVector &coldPathLoads)
{
TR_Stack<TR::Block*> blocksToCheck(TR::comp()->trMemory(), 8, false, stackAlloc);
blocksToCheck.push(firstBlock);
TR::NodeChecklist checklist(TR::comp());
coldPathLoads.empty();
while (!blocksToCheck.isEmpty())
{
TR::Block *block = blocksToCheck.pop();
for (TR::TreeTop *tt = block->getFirstRealTreeTop(); tt->getNode()->getOpCodeValue() != TR::BBEnd; tt = tt->getNextTreeTop())
collectDirectLoads(tt->getNode(), coldPathLoads, checklist);
// Search for any successors that have not merged with the mainline
for (auto itr = block->getSuccessors().begin(), end = block->getSuccessors().end(); itr != end; ++itr)
{
TR::Block *dest = (*itr)->getTo()->asBlock();
if (dest != TR::comp()->getFlowGraph()->getEnd() && dest->getPredecessors().size() == 1)
blocksToCheck.push(dest);
}
}
}
// This opt tries to reduce merge backs from cold code that are the result of inliner
// gnerated nopable virtual guards
// It looks for one basic pattern
//
// guard1 -> cold1
// BBEND
// BBSTART
// guard2 -> cold2
// if guard1 is the guard for a method which calls the method guard2 protects or cold1 is
// a predecessor of cold2 (a situation commonly greated by virtual guard tail splitter) we
// can transform the guards as follows when guard1 and guard2 a
// guard1 -> cold1
// BBEND
// BBSTART
// guard2 -> cold1
// This is safe because there are no trees between the guards and calling the caller will
// result in the call to the callee if we need to patch guard2. cold2 and its mergebacks
// can then be eliminated
//
// In addition this opt will try to move guard2 up from the end of a block to the
// start of the block. We can do this if guard2 is an HCR guard and there is no GC point
// between BBSTART and guard2 since HCR is a stop-the-world event.
//
// Finally, there is a simple tail splitting step run before the analysis of a guard if we
// detect that the taken side of the guard merges back in the next block - this happens
// for some empty methods and is common for Object.<init> at the top of constructors.
int32_t TR_VirtualGuardHeadMerger::perform() {
static char *disableVGHeadMergerTailSplitting = feGetEnv("TR_DisableVGHeadMergerTailSplitting");
TR::CFG *cfg = comp()->getFlowGraph();
// Cache the loads for the outer guard's cold path
TR_BitVector coldPathLoads(comp()->trMemory()->currentStackRegion());
TR_BitVector privArgSymRefs(comp()->trMemory()->currentStackRegion());
bool evaluatedColdPathLoads = false;
for (TR::Block *block = optimizer()->getMethodSymbol()->getFirstTreeTop()->getNode()->getBlock();
block; block = block->getNextBlock())
{
TR::Node *guard1 = block->getLastRealTreeTop()->getNode();
if (isMergeableGuard(guard1))
{
if (trace())
traceMsg(comp(), "Found mergeable guard in block_%d\n", block->getNumber());
TR::Block *cold1 = guard1->getBranchDestination()->getEnclosingBlock();
// check for an immediate merge back from the cold block and
// tail split one block if we can - we only handle splitting a block
// ending in a fallthrough, a branch or a goto for now for simplicity
if (!disableVGHeadMergerTailSplitting &&
(cold1->getSuccessors().size() == 1) &&
cold1->hasSuccessor(block->getNextBlock()) &&
cold1->getLastRealTreeTop()->getNode()->getOpCode().isGoto())
{
// TODO handle moving code earlier in the block down below the guard
// tail split
if ((block->getNextBlock()->getSuccessors().size() == 1) ||
((block->getNextBlock()->getSuccessors().size() == 2) &&
block->getNextBlock()->getLastRealTreeTop()->getNode()->getOpCode().isBranch()) &&
performTransformation(comp(), "%sCloning block_%d and placing clone after block_%d to reduce HCR guard nops\n", OPT_DETAILS, block->getNextBlock()->getNumber(), cold1->getNumber()))
tailSplitBlock(block, cold1);
}
// guard motion is fairly complex but what we want to achieve around guard1 is a sequence
// of relocated privarg blocks, followed by a sequence of runtime patchable guards going to
// guard1's cold block, followed by a sequence of stop-the-world guards going to guard1's
// cold block
//
// The following code is to setup the various insert points based on the following diagrams
// of basic blocks:
//
// start: setup: end result after moving runtime guard'
// | | +-------+ <-- privargIns
// | | <-- privargIns |
// +-------+ <-- runtimeIns +-------+
// | | | | Guard'|
// | | V +-------+ <-- runtimeIns
// +-------+ +-------+ |
// | Guard | | Guard | V
// +-------+ +-------+ <-- HCRIns +-------+
// | ===> | ===> | Guard |
// V V +-------+ <-- HCRIns
// +-------+ +-------+ |
// | | | | V
// | | | | +-------+
//
// Note we always split the block - this may create an empty block but preserves the incoming
// control flow we leave the rest to block extension to fix later
block = block->split(block->getLastRealTreeTop(), cfg, true, false);
TR::Block *privargIns = block->getPrevBlock();
TR::Block *runtimeIns = block->getPrevBlock();
TR::Block *HCRIns = block;
// New outer guard so cold paths must be evaluated
evaluatedColdPathLoads = false;
// scan for candidate guards to merge with guard1 identified above
for (TR::Block *nextBlock = block->getNextBlock(); nextBlock; nextBlock = nextBlock->getNextBlock())
{
if (!(nextBlock->getPredecessors().size() == 1) ||
!nextBlock->hasPredecessor(block))
{
break;
}
TR::TreeTop *guard2Tree = NULL;
if (isMergeableGuard(nextBlock->getFirstRealTreeTop()->getNode()))
{
guard2Tree = nextBlock->getFirstRealTreeTop();
}
else if (isMergeableGuard(nextBlock->getLastRealTreeTop()->getNode()))
{
guard2Tree = nextBlock->getLastRealTreeTop();
}
else
break;
TR::Node *guard2 = guard2Tree->getNode();
TR::Block *guard2Block = nextBlock;
// It is not possible to shift an OSR guard unless the destination is already an OSR point
// as the necessary OSR state will not be available
if (guard2->isOSRGuard() && !guard1->isOSRGuard())
break;
TR::Block *insertPoint = isStopTheWorldGuard(guard2) ? HCRIns : runtimeIns;
if (!safeToMoveGuard(insertPoint, guard2Tree, guard1->getBranchDestination(), privArgSymRefs))
break;
// now we figure out if we can redirect guard2 to guard1's cold block
// ie can we do the head merge
TR::Block *cold2 = guard2->getBranchDestination()->getEnclosingBlock();
if (guard1->getInlinedSiteIndex() == guard2->getInlinedSiteIndex())
{
if (trace())
traceMsg(comp(), " Guard1 [%p] is guarding the same call as Guard2 [%p] - proceeding with guard merging\n", guard1, guard2);
}
else if (guard2->getInlinedSiteIndex() > -1 &&
guard1->getInlinedSiteIndex() == TR::comp()->getInlinedCallSite(guard2->getInlinedSiteIndex())._byteCodeInfo.getCallerIndex())
{
if (trace())
traceMsg(comp(), " Guard1 [%p] is the caller of Guard2 [%p] - proceeding with guard merging\n", guard1, guard2);
}
else if ((cold1->getSuccessors().size() == 1) &&
cold1->hasSuccessor(cold2))
{
if (trace())
traceMsg(comp(), " Guard1 cold destination block_%d has guard2 cold destination block_%d as its only successor - proceeding with guard merging\n", cold1->getNumber(), cold2->getNumber());
}
else
{
if (trace())
traceMsg(comp(), " Cold1 block_%d and cold2 block_%d of guard2 [%p] in unknown relationship - abandon the merge attempt\n", cold1->getNumber(), cold2->getNumber(), guard2);
break;
}
// Runtime guards will shift their privargs, so it is necessary to check such a move is safe
// This is possible if a privarg temp was recycled for the inner call site, with a prior use as an
// argument for the outer call site. As the privargs for the inner call site must be evaluated before
// both guards, this would result in the recycled temp holding the incorrect value if the guard is ever
// taken.
if (!isStopTheWorldGuard(guard2))
{
if (!evaluatedColdPathLoads)
{
collectColdPathLoads(cold1, coldPathLoads);
evaluatedColdPathLoads = true;
}
if (coldPathLoads.intersects(privArgSymRefs))
{
if (trace())
traceMsg(comp(), " Recycled temp live in cold1 block_%d and used as privarg before guard2 [%p] - stop guard merging", cold1->getNumber(), guard2);
break;
}
}
if (!performTransformation(comp(), "%sRedirecting %s guard [%p] in block_%d to parent guard cold block_%d\n", OPT_DETAILS, isStopTheWorldGuard(guard2) ? "stop the world" : "runtime", guard2, guard2Block->getNumber(), cold1->getNumber()))
continue;
if (guard2->getBranchDestination() != guard1->getBranchDestination())
guard2Block->changeBranchDestination(guard1->getBranchDestination(), cfg);
if (guard2Tree != guard2Block->getFirstRealTreeTop())
{
cfg->setStructure(NULL);
// We should leave code ahead of an HCR guard in place because:
// 1, it might have side effect to runtime guards after it, moving it up might cause us to falsely merge
// the subsequent runtime guards
// 2, it might contain live monitor, moving it up above a guard can affect the monitor's live range
if (!isStopTheWorldGuard(guard2))
{
// the block created above guard2 contains only privarg treetops or monitor stores if
// guard2 is a runtime-patchable guard and is safe to merge. We need to move the priv
// args up to the runtime insert point and leave the monitor stores in place
// It's safe to do so because there is no data dependency between the monitor store and
// the priv arg store, because the priv arg store does not load the value from the temp
// holding the monitored object
// Split priv arg stores from monitor stores
// Monitor store is generated for the caller of the method guard2 protects, so should appear before
// priv arg stores for the method guard2 protects
TR::Block *privargBlock = guard2Block;
guard2Block = splitRuntimeGuardBlock(comp(), guard2Block, cfg);
if (privargBlock != guard2Block)
{
if (trace())
traceMsg(comp(), " Moving privarg block_%d after block_%d\n", privargBlock->getNumber(), privargIns->getNumber());
moveBlockAfterDest(cfg, privargBlock, privargIns);
if (HCRIns == privargIns)
HCRIns = privargBlock;
if (runtimeIns == privargIns)
runtimeIns = privargBlock;
privargIns = privargBlock;
// refresh the insertPoint since it could be stale after the above updates
insertPoint = runtimeIns;
}
}
guard2Block = guard2Block->split(guard2Tree, cfg, true, false);
if (trace())
traceMsg(comp(), " Created new block_%d to hold guard [%p] from block_%d\n", guard2Block->getNumber(), guard2, guard2Block->getNumber());
}
if (insertPoint != guard2Block->getPrevBlock())
{
TR::DebugCounter::incStaticDebugCounter(comp(), TR::DebugCounter::debugCounterName(comp(), "headMerger/%s_%s/(%s)", isStopTheWorldGuard(guard1) ? "stop the world" : "runtime", isStopTheWorldGuard(guard2) ? "stop the world" : "runtime", comp()->signature()));
cfg->setStructure(NULL);
block = nextBlock = guard2Block->getPrevBlock();
if (trace())
traceMsg(comp(), " Moving guard2 block block_%d after block_%d\n", guard2Block->getNumber(), insertPoint->getNumber());
moveBlockAfterDest(cfg, guard2Block, insertPoint);
if (HCRIns == insertPoint)
HCRIns = guard2Block;
if (runtimeIns == insertPoint)
runtimeIns = guard2Block;
}
else
{
block = guard2Block;
}
guard1 = guard2;
}
}
}
return 1;
}