int32_t TR::RegDepCopyRemoval::perform() { if (!cg()->supportsPassThroughCopyToNewVirtualRegister()) return 0; discardAllNodeChoices(); TR::TreeTop *tt; for (tt = comp()->getStartTree(); tt != NULL; tt = tt->getNextTreeTop()) { TR::Node *node = tt->getNode(); switch (node->getOpCodeValue()) { case TR::BBStart: if (!node->getBlock()->isExtensionOfPreviousBlock()) { if (trace()) traceMsg(comp(), "clearing remembered node choices at start of extended block at block_%d\n", node->getBlock()->getNumber()); discardAllNodeChoices(); } if (node->getNumChildren() > 0) processRegDeps(node->getFirstChild(), tt); break; case TR::BBEnd: if (node->getNumChildren() > 0) processRegDeps(node->getFirstChild(), tt); break; default: if (node->getOpCode().isSwitch()) { TR::Node *defaultDest = node->getSecondChild(); if (defaultDest->getNumChildren() > 0) processRegDeps(defaultDest->getFirstChild(), tt); } else if (node->getOpCode().isBranch()) { int nChildren = node->getNumChildren(); // only the last child may be GlRegDeps for (int i = 0; i < nChildren - 1; i++) TR_ASSERT(node->getChild(i)->getOpCodeValue() != TR::GlRegDeps, "GlRegDeps for branch is not the last child\n"); if (nChildren > 0) { TR::Node *lastChild = node->getChild(nChildren - 1); if (lastChild->getOpCodeValue() == TR::GlRegDeps) processRegDeps(lastChild, tt); } } break; } } return 1; // a bit arbitrary... }
static void removeGlRegDep(TR::Node * parent, TR_GlobalRegisterNumber registerNum, TR::Block *containingBlock, TR::Optimization *opt) { if (parent->getNumChildren() == 0) return; TR_ASSERT(parent->getNumChildren() > 0, "expected TR::GlRegDeps %p", parent); TR::Node * predGlRegDeps = parent->getLastChild(); if (predGlRegDeps->getOpCodeValue() != TR::GlRegDeps) // could be already removed return; TR_ASSERT(predGlRegDeps->getOpCodeValue() == TR::GlRegDeps, "expected TR::GlRegDeps"); for (int32_t i = predGlRegDeps->getNumChildren() - 1; i >= 0; --i) if (predGlRegDeps->getChild(i)->getGlobalRegisterNumber() == registerNum) { dumpOptDetails(opt->comp(), "%sRemove GlRegDep : %p\n", opt->optDetailString(), predGlRegDeps->getChild(i)); TR::Node *removedChild = predGlRegDeps->removeChild(i); if (removedChild->getReferenceCount() <= 1) { // The only remaining parent is the RegStore. Another pass of // deadTrees may be able to eliminate that. // opt->requestOpt(OMR::deadTreesElimination, true, containingBlock); } break; } if (predGlRegDeps->getNumChildren() == 0) parent->removeLastChild(); }
void OMR::CodeGenerator::evaluateChildrenWithMultipleRefCount(TR::Node * node) { for (int i=0; i < node->getNumChildren(); i++) { TR::Node *child = node->getChild(i); if (child->getRegister() == NULL) // not already evaluated { // Note: we assume things without a symbol reference don't // necessarily need to be evaluated here, and can wait // until they are actually needed. // // vft pointers are speical - we need to evaluate the object in all cases // but for nopable virtual guards we can wait to load and mask the pointer // until we actually need to use it // if (child->getReferenceCount() > 1 && (child->getOpCode().hasSymbolReference() || (child->getOpCodeValue() == TR::l2a && child->getChild(0)->containsCompressionSequence()))) { TR::SymbolReference *vftPointerSymRef = TR::comp()->getSymRefTab()->element(TR::SymbolReferenceTable::vftSymbol); if (node->isNopableInlineGuard() && self()->getSupportsVirtualGuardNOPing() && child->getOpCodeValue() == TR::aloadi && child->getChild(0)->getOpCode().hasSymbolReference() && child->getChild(0)->getSymbolReference() == vftPointerSymRef && child->getChild(0)->getOpCodeValue() == TR::aloadi) { if (!child->getChild(0)->getChild(0)->getRegister() && child->getChild(0)->getChild(0)->getReferenceCount() > 1) self()->evaluate(child->getChild(0)->getChild(0)); else self()->evaluateChildrenWithMultipleRefCount(child->getChild(0)->getChild(0)); } else { self()->evaluate(child); } } else { self()->evaluateChildrenWithMultipleRefCount(child); } } } }
uint8_t *TR::PPCArrayCopyCallSnippet::emitSnippetBody() { TR::Node *node = getNode(); TR_ASSERT(node->getOpCodeValue() == TR::arraycopy && node->getChild(2)->getOpCode().isLoadConst(), "only valid for arraycopies with a constant length\n"); uint8_t *buffer = cg()->getBinaryBufferCursor(); getSnippetLabel()->setCodeLocation(buffer); TR::RealRegister *lengthReg = cg()->machine()->getRealRegister(_lengthRegNum); TR::Node *lengthNode = node->getChild(2); int64_t byteLen = (lengthNode->getType().isInt32() ? lengthNode->getInt() : lengthNode->getLongInt()); TR::InstOpCode opcode; // li lengthReg, #byteLen opcode.setOpCodeValue(TR::InstOpCode::li); buffer = opcode.copyBinaryToBuffer(buffer); lengthReg->setRegisterFieldRT((uint32_t *)buffer); TR_ASSERT(byteLen <= UPPER_IMMED,"byteLen too big to encode\n"); *(int32_t *)buffer |= byteLen; buffer += 4; return TR::PPCHelperCallSnippet::genHelperCall(buffer); }
bool TR::ILValidator::treesAreValid(TR::TreeTop *start, TR::TreeTop *stop) { checkSoundness(start, stop); for (PostorderNodeOccurrenceIterator iter(start, _comp, "VALIDATOR"); iter != stop; ++iter) { updateNodeState(iter); // General node validation // validateNode(iter); // // Additional specific kinds of validation // TR::Node *node = iter.currentNode(); if (node->getOpCodeValue() == TR::BBEnd) { // Determine whether this is the end of an extended block // bool isEndOfExtendedBlock = false; TR::TreeTop *nextTree = iter.currentTree()->getNextTreeTop(); if (nextTree) { validityRule(iter, nextTree->getNode()->getOpCodeValue() == TR::BBStart, "Expected BBStart after BBEnd"); isEndOfExtendedBlock = ! nextTree->getNode()->getBlock()->isExtensionOfPreviousBlock(); } else { isEndOfExtendedBlock = true; } if (isEndOfExtendedBlock) validateEndOfExtendedBlock(iter); } auto opcode = node->getOpCode(); if (opcode.expectedChildCount() != ILChildProp::UnspecifiedChildCount) { // Validate child expectations // const auto expChildCount = opcode.expectedChildCount(); const auto actChildCount = node->getNumChildren(); // validate child count if (!opcode.canHaveGlRegDeps()) { // in the common case, no GlRegDeps child is expect nor present validityRule(iter, actChildCount == expChildCount, "Child count %d does not match expected value of %d", actChildCount, expChildCount); } else if (actChildCount == (expChildCount + 1)) { // adjust expected child number to account for a possible extra GlRegDeps // child and make sure the last child is actually a GlRegDeps validityRule(iter, node->getChild(actChildCount - 1)->getOpCodeValue() == TR::GlRegDeps, "Child count %d does not match expected value of %d (%d without GlRegDeps) and last child is not a GlRegDeps", actChildCount, expChildCount + 1, expChildCount); } else { // if expected and actual child counts don't match, then the child // count is just wrong, even with an expected GlRegDeps validityRule(iter, actChildCount == expChildCount, "Child count %d matches neither expected values of %d (without GlRegDeps) nor %d (with GlRegDeps)", actChildCount, expChildCount, expChildCount + 1); } // validate child types for (auto i = 0; i < actChildCount; ++i) { auto childOpcode = node->getChild(i)->getOpCode(); if (childOpcode.getOpCodeValue() != TR::GlRegDeps) { const auto expChildType = opcode.expectedChildType(i); const auto actChildType = childOpcode.getDataType().getDataType(); const auto expChildTypeName = expChildType == ILChildProp::UnspecifiedChildType ? "UnspecifiedChildType" : TR::DataType::getName(expChildType); const auto actChildTypeName = TR::DataType::getName(actChildType); validityRule(iter, expChildType == ILChildProp::UnspecifiedChildType || actChildType == expChildType, "Child %d has unexpected type %s (expected %s)" , i, actChildTypeName, expChildTypeName); } else { // make sure the node is allowed to have a GlRegDeps child // and make sure that it is the last child validityRule(iter, opcode.canHaveGlRegDeps() && (i == actChildCount - 1), "Unexpected GlRegDeps child %d", i); } } } } return _isValidSoFar; }
int32_t TR::DeadTreesElimination::process(TR::TreeTop *startTree, TR::TreeTop *endTree) { TR::StackMemoryRegion stackRegion(*comp()->trMemory()); LongestPathMap longestPaths(std::less<TR::Node*>(), stackRegion); typedef TR::typed_allocator<CRAnchor, TR::Region&> CRAnchorAlloc; typedef TR::forward_list<CRAnchor, CRAnchorAlloc> CRAnchorList; CRAnchorList anchors(stackRegion); vcount_t visitCount = comp()->incOrResetVisitCount(); TR::TreeTop *treeTop; for (treeTop = startTree; (treeTop != endTree); treeTop = treeTop->getNextTreeTop()) treeTop->getNode()->initializeFutureUseCounts(visitCount); TR::Block *block = NULL; bool delayedRegStoresBeforeThisPass = _delayedRegStores; // Update visitCount as they are used in this optimization and need to be visitCount = comp()->incOrResetVisitCount(); for (TR::TreeTopIterator iter(startTree, comp()); iter != endTree; ++iter) { TR::Node *node = iter.currentTree()->getNode(); if (node->getOpCodeValue() == TR::BBStart) { block = node->getBlock(); if (!block->isExtensionOfPreviousBlock()) longestPaths.clear(); } int vcountLimit = MAX_VCOUNT - 3; if (comp()->getVisitCount() > vcountLimit) { dumpOptDetails(comp(), "%sVisit count %d exceeds limit %d; stopping\n", optDetailString(), comp()->getVisitCount(), vcountLimit); return 0; } // correct at all intermediate stages // if ((node->getOpCodeValue() != TR::treetop) && (!node->getOpCode().isAnchor() || (node->getFirstChild()->getReferenceCount() != 1)) && (!node->getOpCode().isStoreReg() || (node->getFirstChild()->getReferenceCount() != 1)) && (delayedRegStoresBeforeThisPass || (iter.currentTree() == block->getLastRealTreeTop()) || !node->getOpCode().isStoreReg() || (node->getVisitCount() == visitCount))) { if (node->getOpCode().isAnchor() && node->getFirstChild()->getOpCode().isLoadIndirect()) anchors.push_front(CRAnchor(iter.currentTree(), block)); TR::TransformUtil::recursivelySetNodeVisitCount(node, visitCount); continue; } if (node->getOpCode().isStoreReg()) _delayedRegStores = true; TR::Node *child = node->getFirstChild(); if (child->getOpCodeValue() == TR::PassThrough) { TR::Node *newChild = child->getFirstChild(); node->setAndIncChild(0, newChild); newChild->incFutureUseCount(); if (child->getReferenceCount() <= 1) optimizer()->prepareForNodeRemoval(child); child->recursivelyDecReferenceCount(); recursivelyDecFutureUseCount(child); child = newChild; } bool treeTopCanBeEliminated = false; // If the treetop child has been seen before then it must be anchored // somewhere above already; so we don't need the treetop to be anchoring // this node (as the computation is already done at the first reference to // the node). // if (visitCount == child->getVisitCount()) { treeTopCanBeEliminated = true; } else { TR::ILOpCode &childOpCode = child->getOpCode(); TR::ILOpCodes opCodeValue = childOpCode.getOpCodeValue(); bool seenConditionalBranch = false; bool callWithNoSideEffects = child->getOpCode().isCall() && child->getSymbolReference()->getSymbol()->isResolvedMethod() && child->getSymbolReference()->getSymbol()->castToResolvedMethodSymbol()->isSideEffectFree(); if (callWithNoSideEffects) { treeTopCanBeEliminated = true; } else if (!((childOpCode.isCall() && !callWithNoSideEffects) || childOpCode.isStore() || ((opCodeValue == TR::New || opCodeValue == TR::anewarray || opCodeValue == TR::newarray) && child->getReferenceCount() > 1) || opCodeValue == TR::multianewarray || opCodeValue == TR::MergeNew || opCodeValue == TR::checkcast || opCodeValue == TR::Prefetch || opCodeValue == TR::iu2l || ((childOpCode.isDiv() || childOpCode.isRem()) && child->getNumChildren() == 3))) { // Perform the rather complex check to see whether its safe // to disconnect the child node from the treetop // bool safeToReplaceNode = false; if (child->getReferenceCount() == 1) { safeToReplaceNode = true; #ifdef J9_PROJECT_SPECIFIC if (child->getOpCode().isPackedExponentiation()) { // pdexp has a possible message side effect in truncating or no significant digits left cases safeToReplaceNode = false; } #endif if (opCodeValue == TR::loadaddr) treeTopCanBeEliminated = true; } else if (!_cannotBeEliminated) { safeToReplaceNode = isSafeToReplaceNode( child, iter.currentTree(), &seenConditionalBranch, visitCount, comp(), &_targetTrees, _cannotBeEliminated, longestPaths); } if (safeToReplaceNode) { if (childOpCode.hasSymbolReference()) { TR::SymbolReference *symRef = child->getSymbolReference(); if (symRef->getSymbol()->isAuto() || symRef->getSymbol()->isParm()) treeTopCanBeEliminated = true; else { if (childOpCode.isLoad() || (opCodeValue == TR::loadaddr) || (opCodeValue == TR::instanceof) || (((opCodeValue == TR::New) || (opCodeValue == TR::anewarray || opCodeValue == TR::newarray)) && ///child->getFirstChild()->isNonNegative())) child->markedAllocationCanBeRemoved())) // opCodeValue == TR::multianewarray || // opCodeValue == TR::MergeNew) treeTopCanBeEliminated = true; } } else treeTopCanBeEliminated = true; } } // Fix for the case when a float to non-float conversion node swings // down past a branch on IA32; this would cause a FP value to be commoned // across a branch where there was none originally; this causes pblms // as a value is left on the stack. // if (treeTopCanBeEliminated && seenConditionalBranch) { if (!cg()->getSupportsJavaFloatSemantics()) { if (child->getOpCode().isConversion() || child->getOpCode().isBooleanCompare()) { if (child->getFirstChild()->getOpCode().isFloatingPoint() && !child->getOpCode().isFloatingPoint()) treeTopCanBeEliminated = false; } } } if (treeTopCanBeEliminated) { TR::NodeChecklist visited(comp()); bool containsFloatingPoint = false; for (int32_t i = 0; i < child->getNumChildren(); ++i) { // Anchor nodes with reference count > 1 // bool highGlobalIndex = false; if (fixUpTree(child->getChild(i), iter.currentTree(), visited, highGlobalIndex, self(), visitCount)) containsFloatingPoint = true; if (highGlobalIndex) { dumpOptDetails(comp(), "%sGlobal index limit exceeded; stopping\n", optDetailString()); return 0; } } if (seenConditionalBranch && containsFloatingPoint) { if (!cg()->getSupportsJavaFloatSemantics()) treeTopCanBeEliminated = false; } } } // Update visitCount as they are used in this optimization and need to be // correct at all intermediate stages // if (!treeTopCanBeEliminated) TR::TransformUtil::recursivelySetNodeVisitCount(node, visitCount); if (treeTopCanBeEliminated) { TR::TreeTop *prevTree = iter.currentTree()->getPrevTreeTop(); TR::TreeTop *nextTree = iter.currentTree()->getNextTreeTop(); if (!node->getOpCode().isStoreReg() || (node->getFirstChild()->getReferenceCount() == 1)) { // Actually going to remove the treetop now // if (performTransformation(comp(), "%sRemove tree : [" POINTER_PRINTF_FORMAT "] ([" POINTER_PRINTF_FORMAT "] = %s)\n", optDetailString(), node, node->getFirstChild(), node->getFirstChild()->getOpCode().getName())) { prevTree->join(nextTree); optimizer()->prepareForNodeRemoval(node); ///child->recursivelyDecReferenceCount(); node->recursivelyDecReferenceCount(); recursivelyDecFutureUseCount(child); iter.jumpTo(prevTree); if (child->getReferenceCount() == 1) requestOpt(OMR::treeSimplification, true, block); if (nextTree->getNode()->getOpCodeValue() == TR::Goto && prevTree->getNode()->getOpCodeValue() == TR::BBStart && !prevTree->getNode()->getBlock()->isExtensionOfPreviousBlock()) { requestOpt( OMR::redundantGotoElimination, prevTree->getNode()->getBlock()); } } } else { if (performTransformation(comp(), "%sMove tree : [" POINTER_PRINTF_FORMAT "]([" POINTER_PRINTF_FORMAT "] = %s) to end of block\n", optDetailString(), node, node->getFirstChild(), node->getFirstChild()->getOpCode().getName())) { prevTree->join(nextTree); node->setVisitCount(visitCount); TR::TreeTop *lastTree = findLastTreetop(block, prevTree); TR::TreeTop *prevLastTree = lastTree->getPrevTreeTop(); TR::TreeTop *cursorTreeTop = nextTree; while (cursorTreeTop != lastTree) { if (cursorTreeTop->getNode()->getOpCode().isStoreReg() && (cursorTreeTop->getNode()->getGlobalRegisterNumber() == iter.currentTree()->getNode()->getGlobalRegisterNumber())) { lastTree = cursorTreeTop; prevLastTree = lastTree->getPrevTreeTop(); break; } cursorTreeTop = cursorTreeTop->getNextTreeTop(); } if (lastTree->getNode()->getOpCodeValue() == TR::BBStart) { prevLastTree = lastTree; lastTree = block->getExit(); } TR::Node *lastNode = lastTree->getNode(); TR::Node *prevLastNode = prevLastTree->getNode(); if (lastNode->getOpCode().isIf() && !lastNode->getOpCode().isCompBranchOnly() && prevLastNode->getOpCode().isStoreReg() && ((prevLastNode->getFirstChild() == lastNode->getFirstChild()) || (prevLastNode->getFirstChild() == lastNode->getSecondChild()))) { lastTree = prevLastTree; prevLastTree = lastTree->getPrevTreeTop(); } prevLastTree->join(iter.currentTree()); iter.currentTree()->join(lastTree); iter.jumpTo(prevTree); requestOpt(OMR::treeSimplification, true, block); } } } } for (auto it = anchors.begin(); it != anchors.end(); ++it) { TR::Node *anchor = it->tree->getNode(); TR::Node *load = anchor->getChild(0); if (load->getReferenceCount() > 1) continue; // We can eliminate the indirect load immediately, but for the moment the // subtree providing the base object has to be anchored. TR::Node *heapBase = anchor->getChild(1); TR::Node::recreate(anchor, TR::treetop); anchor->setAndIncChild(0, load->getChild(0)); anchor->setChild(1, NULL); anchor->setNumChildren(1); if (!heapBase->getOpCode().isLoadConst()) { it->tree->insertAfter( TR::TreeTop::create( comp(), TR::Node::create(heapBase, TR::treetop, 1, heapBase))); } load->recursivelyDecReferenceCount(); heapBase->recursivelyDecReferenceCount(); // A later pass of dead trees can likely move (or even remove) the base // object expression. requestOpt(OMR::deadTreesElimination, true, it->block); } return 1; // actual cost }
void TR::DeadTreesElimination::prePerformOnBlocks() { _cannotBeEliminated = false; _delayedRegStores = false; _targetTrees.deleteAll(); // Walk through all the blocks to remove trivial dead trees of the form // treetop // => node // The problem with these trees is in the scenario where the earlier use // of 'node' is also dead. However, our analysis won't find that because // the reference count is > 1. vcount_t visitCount = comp()->incOrResetVisitCount(); for (TR::TreeTop *tt = comp()->getStartTree(); tt != 0; tt = tt->getNextTreeTop()) { bool removed = false; TR::Node *node = tt->getNode(); if (node->getOpCodeValue() == TR::treetop && node->getFirstChild()->getVisitCount() == visitCount && performTransformation(comp(), "%sRemove trivial dead tree: %p\n", optDetailString(), node)) { TR::TransformUtil::removeTree(comp(), tt); removed = true; } else { if (node->getOpCode().isCheck() && node->getFirstChild()->getOpCode().isCall() && node->getFirstChild()->getReferenceCount() == 1 && node->getFirstChild()->getSymbolReference()->getSymbol()->isResolvedMethod() && node->getFirstChild()->getSymbolReference()->getSymbol()->castToResolvedMethodSymbol()->isSideEffectFree() && performTransformation(comp(), "%sRemove dead check of side-effect free call: %p\n", optDetailString(), node)) { TR::TransformUtil::removeTree(comp(), tt); removed = true; } } if (removed && tt->getNextTreeTop()->getNode()->getOpCodeValue() == TR::Goto && tt->getPrevTreeTop()->getNode()->getOpCodeValue() == TR::BBStart && !tt->getPrevTreeTop()->getNode()->getBlock()->isExtensionOfPreviousBlock()) { requestOpt(OMR::redundantGotoElimination, tt->getEnclosingBlock()); } if (node->getVisitCount() >= visitCount) continue; TR::TransformUtil::recursivelySetNodeVisitCount(tt->getNode(), visitCount); } // If the last use of an iRegLoad has been removed, then remove the node from // the BBStart and remove the corresponding dependency node from each of the block's // predecessors. // while (1) { bool glRegDepRemoved = false; for (TR::Block * b = comp()->getStartBlock(); b; b = b->getNextBlock()) { TR::TreeTop * startTT = b->getEntry(); TR::Node * startNode = startTT->getNode(); if (startNode->getNumChildren() > 0 && !debug("disableEliminationOfGlRegDeps")) { TR::Node * glRegDeps = startNode->getFirstChild(); TR_ASSERT(glRegDeps->getOpCodeValue() == TR::GlRegDeps, "expected TR::GlRegDeps"); for (int32_t i = glRegDeps->getNumChildren() - 1; i >= 0; --i) { TR::Node * dep = glRegDeps->getChild(i); if (dep->getReferenceCount() == 1 && (!dep->getOpCode().isFloatingPoint() || cg()->getSupportsJavaFloatSemantics()) && performTransformation(comp(), "%sRemove GlRegDep : %p\n", optDetailString(), glRegDeps->getChild(i))) { glRegDeps->removeChild(i); glRegDepRemoved = true; TR_GlobalRegisterNumber registerNum = dep->getGlobalRegisterNumber(); for (auto e = b->getPredecessors().begin(); e != b->getPredecessors().end(); ++e) { TR::Block * pred = toBlock((*e)->getFrom()); if (pred == comp()->getFlowGraph()->getStart()) continue; TR::Node * parent = pred->getLastRealTreeTop()->getNode(); if ( parent->getOpCode().isJumpWithMultipleTargets() && parent->getOpCode().hasBranchChildren()) { for (int32_t j = parent->getCaseIndexUpperBound() - 1; j > 0; --j) { TR::Node * caseNode = parent->getChild(j); TR_ASSERT(caseNode->getOpCode().isCase() || caseNode->getOpCodeValue() == TR::branch, "having problems navigating a switch"); if (caseNode->getBranchDestination() == startTT && caseNode->getNumChildren() > 0 && 0) // can't do this now that all glRegDeps are hung off the default branch removeGlRegDep(caseNode, registerNum, pred, this); } } else if (!parent->getOpCode().isReturn() && parent->getOpCodeValue() != TR::igoto && !( parent->getOpCode().isJumpWithMultipleTargets() && parent->getOpCode().hasBranchChildren()) && !(parent->getOpCodeValue()==TR::treetop && parent->getFirstChild()->getOpCode().isCall() && parent->getFirstChild()->getOpCode().isIndirect())) { if (pred->getNextBlock() == b) parent = pred->getExit()->getNode(); removeGlRegDep(parent, registerNum, pred, this); } } } } if (glRegDeps->getNumChildren() == 0) startNode->removeChild(0); } } if (!glRegDepRemoved) break; } }