Mercurial > hg > graal-jvmci-8
view graal/com.oracle.graal.compiler/src/com/oracle/graal/compiler/alloc/LinearScan.java @ 16369:f8ba57019a5d
LSRA spill optimization: move spill position to the dominator if at spill interval.
author | Josef Eisl <josef.eisl@jku.at> |
---|---|
date | Wed, 11 Jun 2014 17:02:27 +0200 |
parents | ef641ba1fb69 |
children | 3324ab9fe71a |
line wrap: on
line source
/* * Copyright (c) 2009, 2014, Oracle and/or its affiliates. All rights reserved. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * * This code is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License version 2 only, as * published by the Free Software Foundation. * * This code is distributed in the hope that it will be useful, but WITHOUT * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License * version 2 for more details (a copy is included in the LICENSE file that * accompanied this code). * * You should have received a copy of the GNU General Public License version * 2 along with this work; if not, write to the Free Software Foundation, * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. * * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA * or visit www.oracle.com if you need additional information or have any * questions. */ package com.oracle.graal.compiler.alloc; import static com.oracle.graal.api.code.CodeUtil.*; import static com.oracle.graal.api.code.ValueUtil.*; import static com.oracle.graal.compiler.GraalDebugConfig.*; import static com.oracle.graal.compiler.common.cfg.AbstractBlock.*; import static com.oracle.graal.lir.LIRValueUtil.*; import java.util.*; import com.oracle.graal.alloc.*; import com.oracle.graal.api.code.*; import com.oracle.graal.api.meta.*; import com.oracle.graal.compiler.alloc.Interval.RegisterBinding; import com.oracle.graal.compiler.alloc.Interval.RegisterPriority; import com.oracle.graal.compiler.alloc.Interval.SpillState; import com.oracle.graal.compiler.common.*; import com.oracle.graal.compiler.common.cfg.*; import com.oracle.graal.compiler.gen.*; import com.oracle.graal.debug.*; import com.oracle.graal.debug.Debug.Scope; import com.oracle.graal.lir.*; import com.oracle.graal.lir.LIRInstruction.InstructionStateProcedure; import com.oracle.graal.lir.LIRInstruction.InstructionValueProcedure; import com.oracle.graal.lir.LIRInstruction.OperandFlag; import com.oracle.graal.lir.LIRInstruction.OperandMode; import com.oracle.graal.lir.LIRInstruction.ValueProcedure; import com.oracle.graal.lir.StandardOp.MoveOp; import com.oracle.graal.nodes.*; import com.oracle.graal.options.*; import com.oracle.graal.phases.util.*; /** * An implementation of the linear scan register allocator algorithm described in <a * href="http://doi.acm.org/10.1145/1064979.1064998" * >"Optimized Interval Splitting in a Linear Scan Register Allocator"</a> by Christian Wimmer and * Hanspeter Moessenboeck. */ public final class LinearScan { final TargetDescription target; final LIR ir; final FrameMap frameMap; final RegisterAttributes[] registerAttributes; final Register[] registers; boolean callKillsRegisters; private static final int SPLIT_INTERVALS_CAPACITY_RIGHT_SHIFT = 1; public static class Options { // @formatter:off @Option(help = "Enable spill position optimization") public static final OptionValue<Boolean> LSRAOptimizeSpillPosition = new OptionValue<>(true); // @formatter:on } public static class BlockData { /** * Bit map specifying which operands are live upon entry to this block. These are values * used in this block or any of its successors where such value are not defined in this * block. The bit index of an operand is its {@linkplain LinearScan#operandNumber(Value) * operand number}. */ public BitSet liveIn; /** * Bit map specifying which operands are live upon exit from this block. These are values * used in a successor block that are either defined in this block or were live upon entry * to this block. The bit index of an operand is its * {@linkplain LinearScan#operandNumber(Value) operand number}. */ public BitSet liveOut; /** * Bit map specifying which operands are used (before being defined) in this block. That is, * these are the values that are live upon entry to the block. The bit index of an operand * is its {@linkplain LinearScan#operandNumber(Value) operand number}. */ public BitSet liveGen; /** * Bit map specifying which operands are defined/overwritten in this block. The bit index of * an operand is its {@linkplain LinearScan#operandNumber(Value) operand number}. */ public BitSet liveKill; } public final BlockMap<BlockData> blockData; /** * List of blocks in linear-scan order. This is only correct as long as the CFG does not change. */ final List<? extends AbstractBlock<?>> sortedBlocks; /** * Map from {@linkplain #operandNumber(Value) operand numbers} to intervals. */ Interval[] intervals; /** * The number of valid entries in {@link #intervals}. */ int intervalsSize; /** * The index of the first entry in {@link #intervals} for a * {@linkplain #createDerivedInterval(Interval) derived interval}. */ int firstDerivedIntervalIndex = -1; /** * Intervals sorted by {@link Interval#from()}. */ Interval[] sortedIntervals; /** * Map from an instruction {@linkplain LIRInstruction#id id} to the instruction. Entries should * be retrieved with {@link #instructionForId(int)} as the id is not simply an index into this * array. */ LIRInstruction[] opIdToInstructionMap; /** * Map from an instruction {@linkplain LIRInstruction#id id} to the {@linkplain AbstractBlock * block} containing the instruction. Entries should be retrieved with {@link #blockForId(int)} * as the id is not simply an index into this array. */ AbstractBlock<?>[] opIdToBlockMap; /** * Bit set for each variable that is contained in each loop. */ BitMap2D intervalInLoop; /** * The {@linkplain #operandNumber(Value) number} of the first variable operand allocated. */ private final int firstVariableNumber; public LinearScan(TargetDescription target, LIR ir, FrameMap frameMap) { this.target = target; this.ir = ir; this.frameMap = frameMap; this.sortedBlocks = ir.linearScanOrder(); this.registerAttributes = frameMap.registerConfig.getAttributesMap(); this.registers = target.arch.getRegisters(); this.firstVariableNumber = registers.length; this.blockData = new BlockMap<>(ir.getControlFlowGraph()); } public int getFirstLirInstructionId(AbstractBlock<?> block) { int result = ir.getLIRforBlock(block).get(0).id(); assert result >= 0; return result; } public int getLastLirInstructionId(AbstractBlock<?> block) { List<LIRInstruction> instructions = ir.getLIRforBlock(block); int result = instructions.get(instructions.size() - 1).id(); assert result >= 0; return result; } public static boolean isVariableOrRegister(Value value) { return isVariable(value) || isRegister(value); } /** * Converts an operand (variable or register) to an index in a flat address space covering all * the {@linkplain Variable variables} and {@linkplain RegisterValue registers} being processed * by this allocator. */ private int operandNumber(Value operand) { if (isRegister(operand)) { int number = asRegister(operand).number; assert number < firstVariableNumber; return number; } assert isVariable(operand) : operand; return firstVariableNumber + ((Variable) operand).index; } /** * Gets the number of operands. This value will increase by 1 for new variable. */ private int operandSize() { return firstVariableNumber + ir.numVariables(); } /** * Gets the highest operand number for a register operand. This value will never change. */ public int maxRegisterNumber() { return firstVariableNumber - 1; } static final IntervalPredicate IS_PRECOLORED_INTERVAL = new IntervalPredicate() { @Override public boolean apply(Interval i) { return isRegister(i.operand); } }; static final IntervalPredicate IS_VARIABLE_INTERVAL = new IntervalPredicate() { @Override public boolean apply(Interval i) { return isVariable(i.operand); } }; static final IntervalPredicate IS_STACK_INTERVAL = new IntervalPredicate() { @Override public boolean apply(Interval i) { return !isRegister(i.operand); } }; /** * Gets an object describing the attributes of a given register according to this register * configuration. */ RegisterAttributes attributes(Register reg) { return registerAttributes[reg.number]; } void assignSpillSlot(Interval interval) { // assign the canonical spill slot of the parent (if a part of the interval // is already spilled) or allocate a new spill slot if (interval.canMaterialize()) { interval.assignLocation(Value.ILLEGAL); } else if (interval.spillSlot() != null) { interval.assignLocation(interval.spillSlot()); } else { StackSlot slot = frameMap.allocateSpillSlot(interval.kind()); interval.setSpillSlot(slot); interval.assignLocation(slot); } } /** * Creates a new interval. * * @param operand the operand for the interval * @return the created interval */ Interval createInterval(AllocatableValue operand) { assert isLegal(operand); int operandNumber = operandNumber(operand); Interval interval = new Interval(operand, operandNumber); assert operandNumber < intervalsSize; assert intervals[operandNumber] == null; intervals[operandNumber] = interval; return interval; } /** * Creates an interval as a result of splitting or spilling another interval. * * @param source an interval being split of spilled * @return a new interval derived from {@code source} */ Interval createDerivedInterval(Interval source) { if (firstDerivedIntervalIndex == -1) { firstDerivedIntervalIndex = intervalsSize; } if (intervalsSize == intervals.length) { intervals = Arrays.copyOf(intervals, intervals.length + (intervals.length >> SPLIT_INTERVALS_CAPACITY_RIGHT_SHIFT)); } intervalsSize++; Variable variable = new Variable(source.kind(), ir.nextVariable()); Interval interval = createInterval(variable); assert intervals[intervalsSize - 1] == interval; return interval; } // access to block list (sorted in linear scan order) int blockCount() { return sortedBlocks.size(); } AbstractBlock<?> blockAt(int index) { return sortedBlocks.get(index); } /** * Gets the size of the {@link BlockData#liveIn} and {@link BlockData#liveOut} sets for a basic * block. These sets do not include any operands allocated as a result of creating * {@linkplain #createDerivedInterval(Interval) derived intervals}. */ int liveSetSize() { return firstDerivedIntervalIndex == -1 ? operandSize() : firstDerivedIntervalIndex; } int numLoops() { return ir.getControlFlowGraph().getLoops().size(); } boolean isIntervalInLoop(int interval, int loop) { return intervalInLoop.at(interval, loop); } Interval intervalFor(int operandNumber) { return intervals[operandNumber]; } Interval intervalFor(Value operand) { int operandNumber = operandNumber(operand); assert operandNumber < intervalsSize; return intervals[operandNumber]; } Interval getOrCreateInterval(AllocatableValue operand) { Interval ret = intervalFor(operand); if (ret == null) { return createInterval(operand); } else { return ret; } } /** * Gets the highest instruction id allocated by this object. */ int maxOpId() { assert opIdToInstructionMap.length > 0 : "no operations"; return (opIdToInstructionMap.length - 1) << 1; } /** * Converts an {@linkplain LIRInstruction#id instruction id} to an instruction index. All LIR * instructions in a method have an index one greater than their linear-scan order predecesor * with the first instruction having an index of 0. */ static int opIdToIndex(int opId) { return opId >> 1; } /** * Retrieves the {@link LIRInstruction} based on its {@linkplain LIRInstruction#id id}. * * @param opId an instruction {@linkplain LIRInstruction#id id} * @return the instruction whose {@linkplain LIRInstruction#id} {@code == id} */ LIRInstruction instructionForId(int opId) { assert isEven(opId) : "opId not even"; LIRInstruction instr = opIdToInstructionMap[opIdToIndex(opId)]; assert instr.id() == opId; return instr; } /** * Gets the block containing a given instruction. * * @param opId an instruction {@linkplain LIRInstruction#id id} * @return the block containing the instruction denoted by {@code opId} */ AbstractBlock<?> blockForId(int opId) { assert opIdToBlockMap.length > 0 && opId >= 0 && opId <= maxOpId() + 1 : "opId out of range"; return opIdToBlockMap[opIdToIndex(opId)]; } boolean isBlockBegin(int opId) { return opId == 0 || blockForId(opId) != blockForId(opId - 1); } boolean coversBlockBegin(int opId1, int opId2) { return blockForId(opId1) != blockForId(opId2); } /** * Determines if an {@link LIRInstruction} destroys all caller saved registers. * * @param opId an instruction {@linkplain LIRInstruction#id id} * @return {@code true} if the instruction denoted by {@code id} destroys all caller saved * registers. */ boolean hasCall(int opId) { assert isEven(opId) : "opId not even"; return instructionForId(opId).destroysCallerSavedRegisters(); } /** * Eliminates moves from register to stack if the stack slot is known to be correct. */ void changeSpillDefinitionPos(Interval interval, int defPos) { assert interval.isSplitParent() : "can only be called for split parents"; switch (interval.spillState()) { case NoDefinitionFound: assert interval.spillDefinitionPos() == -1 : "must no be set before"; interval.setSpillDefinitionPos(defPos); interval.setSpillState(SpillState.NoSpillStore); break; case NoSpillStore: assert defPos <= interval.spillDefinitionPos() : "positions are processed in reverse order when intervals are created"; if (defPos < interval.spillDefinitionPos() - 2) { // second definition found, so no spill optimization possible for this interval interval.setSpillState(SpillState.NoOptimization); } else { // two consecutive definitions (because of two-operand LIR form) assert blockForId(defPos) == blockForId(interval.spillDefinitionPos()) : "block must be equal"; } break; case NoOptimization: // nothing to do break; default: throw new BailoutException("other states not allowed at this time"); } } // called during register allocation void changeSpillState(Interval interval, int spillPos) { switch (interval.spillState()) { case NoSpillStore: { int defLoopDepth = blockForId(interval.spillDefinitionPos()).getLoopDepth(); int spillLoopDepth = blockForId(spillPos).getLoopDepth(); if (defLoopDepth < spillLoopDepth) { // the loop depth of the spilling position is higher then the loop depth // at the definition of the interval . move write to memory out of loop. if (Options.LSRAOptimizeSpillPosition.getValue()) { // find best spill position in dominator the tree interval.setSpillState(SpillState.SpillInDominator); } else { // store at definition of the interval interval.setSpillState(SpillState.StoreAtDefinition); } } else { // the interval is currently spilled only once, so for now there is no // reason to store the interval at the definition interval.setSpillState(SpillState.OneSpillStore); } break; } case OneSpillStore: { if (Options.LSRAOptimizeSpillPosition.getValue()) { // the interval is spilled more then once interval.setSpillState(SpillState.SpillInDominator); } else { // it is better to store it to // memory at the definition interval.setSpillState(SpillState.StoreAtDefinition); } break; } case SpillInDominator: case StoreAtDefinition: case StartInMemory: case NoOptimization: case NoDefinitionFound: // nothing to do break; default: throw new BailoutException("other states not allowed at this time"); } } abstract static class IntervalPredicate { abstract boolean apply(Interval i); } private static Interval addToListSortedByDefinition(Interval first, Interval interval) { assert first != null; if (first == Interval.EndMarker) { interval.next = Interval.EndMarker; return interval; } int spillPos = interval.spillDefinitionPos(); if (first.spillDefinitionPos() > spillPos) { interval.next = first; return interval; } Interval current = first; while (current.next != Interval.EndMarker && current.next.spillDefinitionPos() < spillPos) { current = current.next; } interval.next = current.next; current.next = interval; return first; } private Interval.Pair createSortedByDefinitionLists(IntervalPredicate isList1, IntervalPredicate isList2) { assert isSorted(sortedIntervals) : "interval list is not sorted"; Interval list1 = Interval.EndMarker; Interval list2 = Interval.EndMarker; Interval v; int n = sortedIntervals.length; for (int i = 0; i < n; i++) { v = sortedIntervals[i]; if (v == null) { continue; } if (isList1.apply(v)) { list1 = addToListSortedByDefinition(list1, v); } else if (isList2 == null || isList2.apply(v)) { list2 = addToListSortedByDefinition(list2, v); } } return new Interval.Pair(list1, list2); } private static final IntervalPredicate mustStoreAtDominatorORDefinition = new IntervalPredicate() { @Override public boolean apply(Interval i) { return i.isSplitParent() && (i.spillState() == SpillState.StoreAtDefinition || i.spillState() == SpillState.SpillInDominator); } }; // called once before assignment of register numbers void eliminateSpillMoves() { try (Indent indent = Debug.logAndIndent("Eliminating unnecessary spill moves")) { // collect all intervals that must be stored after their definition. // the list is sorted by Interval.spillDefinitionPos Interval interval; interval = createSortedByDefinitionLists(mustStoreAtDominatorORDefinition, null).first; if (DetailedAsserts.getValue()) { checkIntervals(interval); } LIRInsertionBuffer insertionBuffer = new LIRInsertionBuffer(); for (AbstractBlock<?> block : sortedBlocks) { List<LIRInstruction> instructions = ir.getLIRforBlock(block); int numInst = instructions.size(); // iterate all instructions of the block. for (int j = 0; j < numInst; j++) { LIRInstruction op = instructions.get(j); int opId = op.id(); if (opId == -1) { MoveOp move = (MoveOp) op; // remove move from register to stack if the stack slot is guaranteed to be // correct. // only moves that have been inserted by LinearScan can be removed. assert isVariable(move.getResult()) : "LinearScan inserts only moves to variables"; Interval curInterval = intervalFor(move.getResult()); if (!isRegister(curInterval.location()) && curInterval.alwaysInMemory()) { // move target is a stack slot that is always correct, so eliminate // instruction if (Debug.isLogEnabled()) { Debug.log("eliminating move from interval %d to %d", operandNumber(move.getInput()), operandNumber(move.getResult())); } // null-instructions are deleted by assignRegNum instructions.set(j, null); } } else { // insert move from register to stack just after // the beginning of the interval assert interval == Interval.EndMarker || interval.spillDefinitionPos() >= opId : "invalid order"; assert interval == Interval.EndMarker || (interval.isSplitParent() && (interval.spillState() == SpillState.StoreAtDefinition || interval.spillState() == SpillState.SpillInDominator)) : "invalid interval"; while (interval != Interval.EndMarker && interval.spillDefinitionPos() == opId) { if (!interval.canMaterialize()) { if (!insertionBuffer.initialized()) { // prepare insertion buffer (appended when all instructions in // the block are processed) insertionBuffer.init(instructions); } // if we spill in a dominator we need to find the right location AllocatableValue fromLocation = interval.spillState() == SpillState.SpillInDominator ? interval.getSplitChildAtOpId(opId, OperandMode.DEF, this).location() : interval.location(); AllocatableValue toLocation = canonicalSpillOpr(interval); assert isRegister(fromLocation) : "from operand must be a register but is: " + fromLocation + " toLocation=" + toLocation + " spillState=" + interval.spillState(); assert isStackSlot(toLocation) : "to operand must be a stack slot"; insertionBuffer.append(j + 1, ir.getSpillMoveFactory().createMove(toLocation, fromLocation)); Debug.log("inserting move after definition of interval %d to stack slot %s at opId %d", interval.operandNumber, interval.spillSlot(), opId); } interval = interval.next; } } } // end of instruction iteration if (insertionBuffer.initialized()) { insertionBuffer.finish(); } } // end of block iteration assert interval == Interval.EndMarker : "missed an interval"; } } private static void checkIntervals(Interval interval) { Interval prev = null; Interval temp = interval; while (temp != Interval.EndMarker) { assert temp.spillDefinitionPos() > 0 : "invalid spill definition pos"; if (prev != null) { assert temp.spillDefinitionPos() >= prev.spillDefinitionPos() : "when intervals are sorted by from : then they must also be sorted by spillDefinitionPos"; } assert temp.spillSlot() != null || temp.canMaterialize() : "interval has no spill slot assigned"; assert temp.spillDefinitionPos() >= temp.from() : "invalid order"; Debug.log("interval %d (from %d to %d) must be stored at %d", temp.operandNumber, temp.from(), temp.to(), temp.spillDefinitionPos()); prev = temp; temp = temp.next; } } /** * Numbers all instructions in all blocks. The numbering follows the * {@linkplain ComputeBlockOrder linear scan order}. */ void numberInstructions() { intervalsSize = operandSize(); intervals = new Interval[intervalsSize + (intervalsSize >> SPLIT_INTERVALS_CAPACITY_RIGHT_SHIFT)]; ValueProcedure setVariableProc = new ValueProcedure() { @Override public Value doValue(Value value) { if (isVariable(value)) { getOrCreateInterval(asVariable(value)); } return value; } }; // Assign IDs to LIR nodes and build a mapping, lirOps, from ID to LIRInstruction node. int numInstructions = 0; for (AbstractBlock<?> block : sortedBlocks) { numInstructions += ir.getLIRforBlock(block).size(); } // initialize with correct length opIdToInstructionMap = new LIRInstruction[numInstructions]; opIdToBlockMap = new AbstractBlock<?>[numInstructions]; int opId = 0; int index = 0; for (AbstractBlock<?> block : sortedBlocks) { blockData.put(block, new BlockData()); List<LIRInstruction> instructions = ir.getLIRforBlock(block); int numInst = instructions.size(); for (int j = 0; j < numInst; j++) { LIRInstruction op = instructions.get(j); op.setId(opId); opIdToInstructionMap[index] = op; opIdToBlockMap[index] = block; assert instructionForId(opId) == op : "must match"; op.forEachTemp(setVariableProc); op.forEachOutput(setVariableProc); index++; opId += 2; // numbering of lirOps by two } } assert index == numInstructions : "must match"; assert (index << 1) == opId : "must match: " + (index << 1); } /** * Computes local live sets (i.e. {@link BlockData#liveGen} and {@link BlockData#liveKill}) * separately for each block. */ void computeLocalLiveSets() { int liveSize = liveSetSize(); intervalInLoop = new BitMap2D(operandSize(), numLoops()); // iterate all blocks for (final AbstractBlock<?> block : sortedBlocks) { try (Indent indent = Debug.logAndIndent("compute local live sets for block %d", block.getId())) { final BitSet liveGen = new BitSet(liveSize); final BitSet liveKill = new BitSet(liveSize); List<LIRInstruction> instructions = ir.getLIRforBlock(block); int numInst = instructions.size(); ValueProcedure useProc = new ValueProcedure() { @Override protected Value doValue(Value operand) { if (isVariable(operand)) { int operandNum = operandNumber(operand); if (!liveKill.get(operandNum)) { liveGen.set(operandNum); Debug.log("liveGen for operand %d", operandNum); } if (block.getLoop() != null) { intervalInLoop.setBit(operandNum, block.getLoop().getIndex()); } } if (DetailedAsserts.getValue()) { verifyInput(block, liveKill, operand); } return operand; } }; ValueProcedure stateProc = new ValueProcedure() { @Override public Value doValue(Value operand) { int operandNum = operandNumber(operand); if (!liveKill.get(operandNum)) { liveGen.set(operandNum); Debug.log("liveGen in state for operand %d", operandNum); } return operand; } }; ValueProcedure defProc = new ValueProcedure() { @Override public Value doValue(Value operand) { if (isVariable(operand)) { int varNum = operandNumber(operand); liveKill.set(varNum); Debug.log("liveKill for operand %d", varNum); if (block.getLoop() != null) { intervalInLoop.setBit(varNum, block.getLoop().getIndex()); } } if (DetailedAsserts.getValue()) { // fixed intervals are never live at block boundaries, so // they need not be processed in live sets // process them only in debug mode so that this can be checked verifyTemp(liveKill, operand); } return operand; } }; // iterate all instructions of the block for (int j = 0; j < numInst; j++) { final LIRInstruction op = instructions.get(j); try (Indent indent2 = Debug.logAndIndent("handle op %d", op.id())) { op.forEachInput(useProc); op.forEachAlive(useProc); // Add uses of live locals from interpreter's point of view for proper debug // information generation op.forEachState(stateProc); op.forEachTemp(defProc); op.forEachOutput(defProc); } } // end of instruction iteration BlockData blockSets = blockData.get(block); blockSets.liveGen = liveGen; blockSets.liveKill = liveKill; blockSets.liveIn = new BitSet(liveSize); blockSets.liveOut = new BitSet(liveSize); Debug.log("liveGen B%d %s", block.getId(), blockSets.liveGen); Debug.log("liveKill B%d %s", block.getId(), blockSets.liveKill); } } // end of block iteration } private void verifyTemp(BitSet liveKill, Value operand) { // fixed intervals are never live at block boundaries, so // they need not be processed in live sets // process them only in debug mode so that this can be checked if (isRegister(operand)) { if (isProcessed(operand)) { liveKill.set(operandNumber(operand)); } } } private void verifyInput(AbstractBlock<?> block, BitSet liveKill, Value operand) { // fixed intervals are never live at block boundaries, so // they need not be processed in live sets. // this is checked by these assertions to be sure about it. // the entry block may have incoming // values in registers, which is ok. if (isRegister(operand) && block != ir.getControlFlowGraph().getStartBlock()) { if (isProcessed(operand)) { assert liveKill.get(operandNumber(operand)) : "using fixed register that is not defined in this block"; } } } /** * Performs a backward dataflow analysis to compute global live sets (i.e. * {@link BlockData#liveIn} and {@link BlockData#liveOut}) for each block. */ void computeGlobalLiveSets() { try (Indent indent = Debug.logAndIndent("compute global live sets")) { int numBlocks = blockCount(); boolean changeOccurred; boolean changeOccurredInBlock; int iterationCount = 0; BitSet liveOut = new BitSet(liveSetSize()); // scratch set for calculations // Perform a backward dataflow analysis to compute liveOut and liveIn for each block. // The loop is executed until a fixpoint is reached (no changes in an iteration) do { changeOccurred = false; try (Indent indent2 = Debug.logAndIndent("new iteration %d", iterationCount)) { // iterate all blocks in reverse order for (int i = numBlocks - 1; i >= 0; i--) { AbstractBlock<?> block = blockAt(i); BlockData blockSets = blockData.get(block); changeOccurredInBlock = false; // liveOut(block) is the union of liveIn(sux), for successors sux of block int n = block.getSuccessorCount(); if (n > 0) { liveOut.clear(); // block has successors if (n > 0) { for (AbstractBlock<?> successor : block.getSuccessors()) { liveOut.or(blockData.get(successor).liveIn); } } if (!blockSets.liveOut.equals(liveOut)) { // A change occurred. Swap the old and new live out // sets to avoid copying. BitSet temp = blockSets.liveOut; blockSets.liveOut = liveOut; liveOut = temp; changeOccurred = true; changeOccurredInBlock = true; } } if (iterationCount == 0 || changeOccurredInBlock) { // liveIn(block) is the union of liveGen(block) with (liveOut(block) & // !liveKill(block)) // note: liveIn has to be computed only in first iteration // or if liveOut has changed! BitSet liveIn = blockSets.liveIn; liveIn.clear(); liveIn.or(blockSets.liveOut); liveIn.andNot(blockSets.liveKill); liveIn.or(blockSets.liveGen); Debug.log("block %d: livein = %s, liveout = %s", block.getId(), liveIn, blockSets.liveOut); } } iterationCount++; if (changeOccurred && iterationCount > 50) { throw new BailoutException("too many iterations in computeGlobalLiveSets"); } } } while (changeOccurred); if (DetailedAsserts.getValue()) { verifyLiveness(); } // check that the liveIn set of the first block is empty AbstractBlock<?> startBlock = ir.getControlFlowGraph().getStartBlock(); if (blockData.get(startBlock).liveIn.cardinality() != 0) { if (DetailedAsserts.getValue()) { reportFailure(numBlocks); } // bailout if this occurs in product mode. throw new GraalInternalError("liveIn set of first block must be empty: " + blockData.get(startBlock).liveIn); } } } private static NodeLIRBuilder getNodeLIRGeneratorFromDebugContext() { if (Debug.isEnabled()) { NodeLIRBuilder lirGen = Debug.contextLookup(NodeLIRBuilder.class); assert lirGen != null; return lirGen; } return null; } private static ValueNode getValueForOperandFromDebugContext(Value value) { NodeLIRBuilder gen = getNodeLIRGeneratorFromDebugContext(); if (gen != null) { return gen.valueForOperand(value); } return null; } private void reportFailure(int numBlocks) { try (Scope s = Debug.forceLog()) { try (Indent indent = Debug.logAndIndent("report failure")) { BitSet startBlockLiveIn = blockData.get(ir.getControlFlowGraph().getStartBlock()).liveIn; try (Indent indent2 = Debug.logAndIndent("Error: liveIn set of first block must be empty (when this fails, variables are used before they are defined):")) { for (int operandNum = startBlockLiveIn.nextSetBit(0); operandNum >= 0; operandNum = startBlockLiveIn.nextSetBit(operandNum + 1)) { Value operand = intervalFor(operandNum).operand; Debug.log("var %d; operand=%s; node=%s", operandNum, operand, getValueForOperandFromDebugContext(operand)); } } // print some additional information to simplify debugging for (int operandNum = startBlockLiveIn.nextSetBit(0); operandNum >= 0; operandNum = startBlockLiveIn.nextSetBit(operandNum + 1)) { Value operand = intervalFor(operandNum).operand; try (Indent indent2 = Debug.logAndIndent("---- Detailed information for var %d; operand=%s; node=%s ----", operandNum, operand, getValueForOperandFromDebugContext(operand))) { Deque<AbstractBlock<?>> definedIn = new ArrayDeque<>(); HashSet<AbstractBlock<?>> usedIn = new HashSet<>(); for (AbstractBlock<?> block : sortedBlocks) { if (blockData.get(block).liveGen.get(operandNum)) { usedIn.add(block); try (Indent indent3 = Debug.logAndIndent("used in block B%d", block.getId())) { for (LIRInstruction ins : ir.getLIRforBlock(block)) { try (Indent indent4 = Debug.logAndIndent("%d: %s", ins.id(), ins)) { ins.forEachState(new ValueProcedure() { @Override public Value doValue(Value liveStateOperand) { Debug.log("operand=%s", liveStateOperand); return liveStateOperand; } }); } } } } if (blockData.get(block).liveKill.get(operandNum)) { definedIn.add(block); try (Indent indent3 = Debug.logAndIndent("defined in block B%d", block.getId())) { for (LIRInstruction ins : ir.getLIRforBlock(block)) { Debug.log("%d: %s", ins.id(), ins); } } } } int[] hitCount = new int[numBlocks]; while (!definedIn.isEmpty()) { AbstractBlock<?> block = definedIn.removeFirst(); usedIn.remove(block); for (AbstractBlock<?> successor : block.getSuccessors()) { if (successor.isLoopHeader()) { if (!block.isLoopEnd()) { definedIn.add(successor); } } else { if (++hitCount[successor.getId()] == successor.getPredecessorCount()) { definedIn.add(successor); } } } } try (Indent indent3 = Debug.logAndIndent("**** offending usages are in: ")) { for (AbstractBlock<?> block : usedIn) { Debug.log("B%d", block.getId()); } } } } } } } private void verifyLiveness() { // check that fixed intervals are not live at block boundaries // (live set must be empty at fixed intervals) for (AbstractBlock<?> block : sortedBlocks) { for (int j = 0; j <= maxRegisterNumber(); j++) { assert !blockData.get(block).liveIn.get(j) : "liveIn set of fixed register must be empty"; assert !blockData.get(block).liveOut.get(j) : "liveOut set of fixed register must be empty"; assert !blockData.get(block).liveGen.get(j) : "liveGen set of fixed register must be empty"; } } } void addUse(AllocatableValue operand, int from, int to, RegisterPriority registerPriority, LIRKind kind) { if (!isProcessed(operand)) { return; } Interval interval = getOrCreateInterval(operand); if (!kind.equals(LIRKind.Illegal)) { interval.setKind(kind); } interval.addRange(from, to); // Register use position at even instruction id. interval.addUsePos(to & ~1, registerPriority); Debug.log("add use: %s, from %d to %d (%s)", interval, from, to, registerPriority.name()); } void addTemp(AllocatableValue operand, int tempPos, RegisterPriority registerPriority, LIRKind kind) { if (!isProcessed(operand)) { return; } Interval interval = getOrCreateInterval(operand); if (!kind.equals(LIRKind.Illegal)) { interval.setKind(kind); } interval.addRange(tempPos, tempPos + 1); interval.addUsePos(tempPos, registerPriority); interval.addMaterializationValue(null); Debug.log("add temp: %s tempPos %d (%s)", interval, tempPos, RegisterPriority.MustHaveRegister.name()); } boolean isProcessed(Value operand) { return !isRegister(operand) || attributes(asRegister(operand)).isAllocatable(); } void addDef(AllocatableValue operand, LIRInstruction op, RegisterPriority registerPriority, LIRKind kind) { if (!isProcessed(operand)) { return; } int defPos = op.id(); Interval interval = getOrCreateInterval(operand); if (!kind.equals(LIRKind.Illegal)) { interval.setKind(kind); } Range r = interval.first(); if (r.from <= defPos) { // Update the starting point (when a range is first created for a use, its // start is the beginning of the current block until a def is encountered.) r.from = defPos; interval.addUsePos(defPos, registerPriority); } else { // Dead value - make vacuous interval // also add register priority for dead intervals interval.addRange(defPos, defPos + 1); interval.addUsePos(defPos, registerPriority); Debug.log("Warning: def of operand %s at %d occurs without use", operand, defPos); } changeSpillDefinitionPos(interval, defPos); if (registerPriority == RegisterPriority.None && interval.spillState().ordinal() <= SpillState.StartInMemory.ordinal()) { // detection of method-parameters and roundfp-results interval.setSpillState(SpillState.StartInMemory); } interval.addMaterializationValue(LinearScan.getMaterializedValue(op, operand, interval)); Debug.log("add def: %s defPos %d (%s)", interval, defPos, registerPriority.name()); } /** * Determines the register priority for an instruction's output/result operand. */ static RegisterPriority registerPriorityOfOutputOperand(LIRInstruction op) { if (op instanceof MoveOp) { MoveOp move = (MoveOp) op; if (optimizeMethodArgument(move.getInput())) { return RegisterPriority.None; } } // all other operands require a register return RegisterPriority.MustHaveRegister; } /** * Determines the priority which with an instruction's input operand will be allocated a * register. */ static RegisterPriority registerPriorityOfInputOperand(EnumSet<OperandFlag> flags) { if (flags.contains(OperandFlag.STACK)) { return RegisterPriority.ShouldHaveRegister; } // all other operands require a register return RegisterPriority.MustHaveRegister; } private static boolean optimizeMethodArgument(Value value) { /* * Object method arguments that are passed on the stack are currently not optimized because * this requires that the runtime visits method arguments during stack walking. */ return isStackSlot(value) && asStackSlot(value).isInCallerFrame() && value.getKind() != Kind.Object; } /** * Optimizes moves related to incoming stack based arguments. The interval for the destination * of such moves is assigned the stack slot (which is in the caller's frame) as its spill slot. */ void handleMethodArguments(LIRInstruction op) { if (op instanceof MoveOp) { MoveOp move = (MoveOp) op; if (optimizeMethodArgument(move.getInput())) { StackSlot slot = asStackSlot(move.getInput()); if (DetailedAsserts.getValue()) { assert op.id() > 0 : "invalid id"; assert blockForId(op.id()).getPredecessorCount() == 0 : "move from stack must be in first block"; assert isVariable(move.getResult()) : "result of move must be a variable"; Debug.log("found move from stack slot %s to %s", slot, move.getResult()); } Interval interval = intervalFor(move.getResult()); interval.setSpillSlot(slot); interval.assignLocation(slot); } } } void addRegisterHint(final LIRInstruction op, final Value targetValue, OperandMode mode, EnumSet<OperandFlag> flags, final boolean hintAtDef) { if (flags.contains(OperandFlag.HINT) && isVariableOrRegister(targetValue)) { op.forEachRegisterHint(targetValue, mode, new ValueProcedure() { @Override protected Value doValue(Value registerHint) { if (isVariableOrRegister(registerHint)) { Interval from = getOrCreateInterval((AllocatableValue) registerHint); Interval to = getOrCreateInterval((AllocatableValue) targetValue); // hints always point from def to use if (hintAtDef) { to.setLocationHint(from); } else { from.setLocationHint(to); } Debug.log("operation at opId %d: added hint from interval %d to %d", op.id(), from.operandNumber, to.operandNumber); return registerHint; } return null; } }); } } void buildIntervals() { try (Indent indent = Debug.logAndIndent("build intervals")) { InstructionValueProcedure outputProc = new InstructionValueProcedure() { @Override public Value doValue(LIRInstruction op, Value operand, OperandMode mode, EnumSet<OperandFlag> flags) { if (isVariableOrRegister(operand)) { addDef((AllocatableValue) operand, op, registerPriorityOfOutputOperand(op), operand.getLIRKind()); addRegisterHint(op, operand, mode, flags, true); } return operand; } }; InstructionValueProcedure tempProc = new InstructionValueProcedure() { @Override public Value doValue(LIRInstruction op, Value operand, OperandMode mode, EnumSet<OperandFlag> flags) { if (isVariableOrRegister(operand)) { addTemp((AllocatableValue) operand, op.id(), RegisterPriority.MustHaveRegister, operand.getLIRKind()); addRegisterHint(op, operand, mode, flags, false); } return operand; } }; InstructionValueProcedure aliveProc = new InstructionValueProcedure() { @Override public Value doValue(LIRInstruction op, Value operand, OperandMode mode, EnumSet<OperandFlag> flags) { if (isVariableOrRegister(operand)) { RegisterPriority p = registerPriorityOfInputOperand(flags); final int opId = op.id(); final int blockFrom = getFirstLirInstructionId((blockForId(opId))); addUse((AllocatableValue) operand, blockFrom, opId + 1, p, operand.getLIRKind()); addRegisterHint(op, operand, mode, flags, false); } return operand; } }; InstructionValueProcedure inputProc = new InstructionValueProcedure() { @Override public Value doValue(LIRInstruction op, Value operand, OperandMode mode, EnumSet<OperandFlag> flags) { if (isVariableOrRegister(operand)) { final int opId = op.id(); final int blockFrom = getFirstLirInstructionId((blockForId(opId))); RegisterPriority p = registerPriorityOfInputOperand(flags); addUse((AllocatableValue) operand, blockFrom, opId, p, operand.getLIRKind()); addRegisterHint(op, operand, mode, flags, false); } return operand; } }; InstructionValueProcedure stateProc = new InstructionValueProcedure() { @Override public Value doValue(LIRInstruction op, Value operand) { final int opId = op.id(); final int blockFrom = getFirstLirInstructionId((blockForId(opId))); addUse((AllocatableValue) operand, blockFrom, opId + 1, RegisterPriority.None, operand.getLIRKind()); return operand; } }; // create a list with all caller-save registers (cpu, fpu, xmm) Register[] callerSaveRegs = frameMap.registerConfig.getCallerSaveRegisters(); // iterate all blocks in reverse order for (int i = blockCount() - 1; i >= 0; i--) { AbstractBlock<?> block = blockAt(i); try (Indent indent2 = Debug.logAndIndent("handle block %d", block.getId())) { List<LIRInstruction> instructions = ir.getLIRforBlock(block); final int blockFrom = getFirstLirInstructionId(block); int blockTo = getLastLirInstructionId(block); assert blockFrom == instructions.get(0).id(); assert blockTo == instructions.get(instructions.size() - 1).id(); // Update intervals for operands live at the end of this block; BitSet live = blockData.get(block).liveOut; for (int operandNum = live.nextSetBit(0); operandNum >= 0; operandNum = live.nextSetBit(operandNum + 1)) { assert live.get(operandNum) : "should not stop here otherwise"; AllocatableValue operand = intervalFor(operandNum).operand; Debug.log("live in %d: %s", operandNum, operand); addUse(operand, blockFrom, blockTo + 2, RegisterPriority.None, LIRKind.Illegal); // add special use positions for loop-end blocks when the // interval is used anywhere inside this loop. It's possible // that the block was part of a non-natural loop, so it might // have an invalid loop index. if (block.isLoopEnd() && block.getLoop() != null && isIntervalInLoop(operandNum, block.getLoop().getIndex())) { intervalFor(operandNum).addUsePos(blockTo + 1, RegisterPriority.LiveAtLoopEnd); } } // iterate all instructions of the block in reverse order. // definitions of intervals are processed before uses for (int j = instructions.size() - 1; j >= 0; j--) { final LIRInstruction op = instructions.get(j); final int opId = op.id(); try (Indent indent3 = Debug.logAndIndent("handle inst %d: %s", opId, op)) { // add a temp range for each register if operation destroys // caller-save registers if (op.destroysCallerSavedRegisters()) { for (Register r : callerSaveRegs) { if (attributes(r).isAllocatable()) { addTemp(r.asValue(), opId, RegisterPriority.None, LIRKind.Illegal); } } Debug.log("operation destroys all caller-save registers"); } op.forEachOutput(outputProc); op.forEachTemp(tempProc); op.forEachAlive(aliveProc); op.forEachInput(inputProc); // Add uses of live locals from interpreter's point of view for proper // debug information generation // Treat these operands as temp values (if the live range is extended // to a call site, the value would be in a register at // the call otherwise) op.forEachState(stateProc); // special steps for some instructions (especially moves) handleMethodArguments(op); } } // end of instruction iteration } } // end of block iteration // add the range [0, 1] to all fixed intervals. // the register allocator need not handle unhandled fixed intervals for (Interval interval : intervals) { if (interval != null && isRegister(interval.operand)) { interval.addRange(0, 1); } } } } // * Phase 5: actual register allocation private static boolean isSorted(Interval[] intervals) { int from = -1; for (Interval interval : intervals) { assert interval != null; assert from <= interval.from(); from = interval.from(); } return true; } static Interval addToList(Interval first, Interval prev, Interval interval) { Interval newFirst = first; if (prev != null) { prev.next = interval; } else { newFirst = interval; } return newFirst; } Interval.Pair createUnhandledLists(IntervalPredicate isList1, IntervalPredicate isList2) { assert isSorted(sortedIntervals) : "interval list is not sorted"; Interval list1 = Interval.EndMarker; Interval list2 = Interval.EndMarker; Interval list1Prev = null; Interval list2Prev = null; Interval v; int n = sortedIntervals.length; for (int i = 0; i < n; i++) { v = sortedIntervals[i]; if (v == null) { continue; } if (isList1.apply(v)) { list1 = addToList(list1, list1Prev, v); list1Prev = v; } else if (isList2 == null || isList2.apply(v)) { list2 = addToList(list2, list2Prev, v); list2Prev = v; } } if (list1Prev != null) { list1Prev.next = Interval.EndMarker; } if (list2Prev != null) { list2Prev.next = Interval.EndMarker; } assert list1Prev == null || list1Prev.next == Interval.EndMarker : "linear list ends not with sentinel"; assert list2Prev == null || list2Prev.next == Interval.EndMarker : "linear list ends not with sentinel"; return new Interval.Pair(list1, list2); } void sortIntervalsBeforeAllocation() { int sortedLen = 0; for (Interval interval : intervals) { if (interval != null) { sortedLen++; } } Interval[] sortedList = new Interval[sortedLen]; int sortedIdx = 0; int sortedFromMax = -1; // special sorting algorithm: the original interval-list is almost sorted, // only some intervals are swapped. So this is much faster than a complete QuickSort for (Interval interval : intervals) { if (interval != null) { int from = interval.from(); if (sortedFromMax <= from) { sortedList[sortedIdx++] = interval; sortedFromMax = interval.from(); } else { // the assumption that the intervals are already sorted failed, // so this interval must be sorted in manually int j; for (j = sortedIdx - 1; j >= 0 && from < sortedList[j].from(); j--) { sortedList[j + 1] = sortedList[j]; } sortedList[j + 1] = interval; sortedIdx++; } } } sortedIntervals = sortedList; } void sortIntervalsAfterAllocation() { if (firstDerivedIntervalIndex == -1) { // no intervals have been added during allocation, so sorted list is already up to date return; } Interval[] oldList = sortedIntervals; Interval[] newList = Arrays.copyOfRange(intervals, firstDerivedIntervalIndex, intervalsSize); int oldLen = oldList.length; int newLen = newList.length; // conventional sort-algorithm for new intervals Arrays.sort(newList, (Interval a, Interval b) -> a.from() - b.from()); // merge old and new list (both already sorted) into one combined list Interval[] combinedList = new Interval[oldLen + newLen]; int oldIdx = 0; int newIdx = 0; while (oldIdx + newIdx < combinedList.length) { if (newIdx >= newLen || (oldIdx < oldLen && oldList[oldIdx].from() <= newList[newIdx].from())) { combinedList[oldIdx + newIdx] = oldList[oldIdx]; oldIdx++; } else { combinedList[oldIdx + newIdx] = newList[newIdx]; newIdx++; } } sortedIntervals = combinedList; } public void allocateRegisters() { try (Indent indent = Debug.logAndIndent("allocate registers")) { Interval precoloredIntervals; Interval notPrecoloredIntervals; Interval.Pair result = createUnhandledLists(IS_PRECOLORED_INTERVAL, IS_VARIABLE_INTERVAL); precoloredIntervals = result.first; notPrecoloredIntervals = result.second; // allocate cpu registers LinearScanWalker lsw; if (OptimizingLinearScanWalker.Options.LSRAOptimization.getValue()) { lsw = new OptimizingLinearScanWalker(this, precoloredIntervals, notPrecoloredIntervals); } else { lsw = new LinearScanWalker(this, precoloredIntervals, notPrecoloredIntervals); } lsw.walk(); lsw.finishAllocation(); } } // * Phase 6: resolve data flow // (insert moves at edges between blocks if intervals have been split) // wrapper for Interval.splitChildAtOpId that performs a bailout in product mode // instead of returning null Interval splitChildAtOpId(Interval interval, int opId, LIRInstruction.OperandMode mode) { Interval result = interval.getSplitChildAtOpId(opId, mode, this); if (result != null) { Debug.log("Split child at pos %d of interval %s is %s", opId, interval, result); return result; } throw new BailoutException("LinearScan: interval is null"); } Interval intervalAtBlockBegin(AbstractBlock<?> block, int operandNumber) { return splitChildAtOpId(intervalFor(operandNumber), getFirstLirInstructionId(block), LIRInstruction.OperandMode.DEF); } Interval intervalAtBlockEnd(AbstractBlock<?> block, int operandNumber) { return splitChildAtOpId(intervalFor(operandNumber), getLastLirInstructionId(block) + 1, LIRInstruction.OperandMode.DEF); } void resolveCollectMappings(AbstractBlock<?> fromBlock, AbstractBlock<?> toBlock, MoveResolver moveResolver) { assert moveResolver.checkEmpty(); int numOperands = operandSize(); BitSet liveAtEdge = blockData.get(toBlock).liveIn; // visit all variables for which the liveAtEdge bit is set for (int operandNum = liveAtEdge.nextSetBit(0); operandNum >= 0; operandNum = liveAtEdge.nextSetBit(operandNum + 1)) { assert operandNum < numOperands : "live information set for not exisiting interval"; assert blockData.get(fromBlock).liveOut.get(operandNum) && blockData.get(toBlock).liveIn.get(operandNum) : "interval not live at this edge"; Interval fromInterval = intervalAtBlockEnd(fromBlock, operandNum); Interval toInterval = intervalAtBlockBegin(toBlock, operandNum); if (fromInterval != toInterval && !fromInterval.location().equals(toInterval.location())) { // need to insert move instruction moveResolver.addMapping(fromInterval, toInterval); } } } void resolveFindInsertPos(AbstractBlock<?> fromBlock, AbstractBlock<?> toBlock, MoveResolver moveResolver) { if (fromBlock.getSuccessorCount() <= 1) { Debug.log("inserting moves at end of fromBlock B%d", fromBlock.getId()); List<LIRInstruction> instructions = ir.getLIRforBlock(fromBlock); LIRInstruction instr = instructions.get(instructions.size() - 1); if (instr instanceof StandardOp.JumpOp) { // insert moves before branch moveResolver.setInsertPosition(instructions, instructions.size() - 1); } else { moveResolver.setInsertPosition(instructions, instructions.size()); } } else { Debug.log("inserting moves at beginning of toBlock B%d", toBlock.getId()); if (DetailedAsserts.getValue()) { assert ir.getLIRforBlock(fromBlock).get(0) instanceof StandardOp.LabelOp : "block does not start with a label"; // because the number of predecessor edges matches the number of // successor edges, blocks which are reached by switch statements // may have be more than one predecessor but it will be guaranteed // that all predecessors will be the same. for (AbstractBlock<?> predecessor : toBlock.getPredecessors()) { assert fromBlock == predecessor : "all critical edges must be broken"; } } moveResolver.setInsertPosition(ir.getLIRforBlock(toBlock), 1); } } /** * Inserts necessary moves (spilling or reloading) at edges between blocks for intervals that * have been split. */ void resolveDataFlow() { try (Indent indent = Debug.logAndIndent("resolve data flow")) { int numBlocks = blockCount(); MoveResolver moveResolver = new MoveResolver(this); BitSet blockCompleted = new BitSet(numBlocks); BitSet alreadyResolved = new BitSet(numBlocks); for (AbstractBlock<?> block : sortedBlocks) { // check if block has only one predecessor and only one successor if (block.getPredecessorCount() == 1 && block.getSuccessorCount() == 1) { List<LIRInstruction> instructions = ir.getLIRforBlock(block); assert instructions.get(0) instanceof StandardOp.LabelOp : "block must start with label"; assert instructions.get(instructions.size() - 1) instanceof StandardOp.JumpOp : "block with successor must end with unconditional jump"; // check if block is empty (only label and branch) if (instructions.size() == 2) { AbstractBlock<?> pred = block.getPredecessors().iterator().next(); AbstractBlock<?> sux = block.getSuccessors().iterator().next(); // prevent optimization of two consecutive blocks if (!blockCompleted.get(pred.getLinearScanNumber()) && !blockCompleted.get(sux.getLinearScanNumber())) { Debug.log(" optimizing empty block B%d (pred: B%d, sux: B%d)", block.getId(), pred.getId(), sux.getId()); blockCompleted.set(block.getLinearScanNumber()); // directly resolve between pred and sux (without looking // at the empty block // between) resolveCollectMappings(pred, sux, moveResolver); if (moveResolver.hasMappings()) { moveResolver.setInsertPosition(instructions, 1); moveResolver.resolveAndAppendMoves(); } } } } } for (AbstractBlock<?> fromBlock : sortedBlocks) { if (!blockCompleted.get(fromBlock.getLinearScanNumber())) { alreadyResolved.clear(); alreadyResolved.or(blockCompleted); for (AbstractBlock<?> toBlock : fromBlock.getSuccessors()) { // check for duplicate edges between the same blocks (can happen with switch // blocks) if (!alreadyResolved.get(toBlock.getLinearScanNumber())) { Debug.log("processing edge between B%d and B%d", fromBlock.getId(), toBlock.getId()); alreadyResolved.set(toBlock.getLinearScanNumber()); // collect all intervals that have been split between // fromBlock and toBlock resolveCollectMappings(fromBlock, toBlock, moveResolver); if (moveResolver.hasMappings()) { resolveFindInsertPos(fromBlock, toBlock, moveResolver); moveResolver.resolveAndAppendMoves(); } } } } } } } // * Phase 7: assign register numbers back to LIR // (includes computation of debug information and oop maps) static StackSlot canonicalSpillOpr(Interval interval) { assert interval.spillSlot() != null : "canonical spill slot not set"; return interval.spillSlot(); } /** * Assigns the allocated location for an LIR instruction operand back into the instruction. * * @param operand an LIR instruction operand * @param opId the id of the LIR instruction using {@code operand} * @param mode the usage mode for {@code operand} by the instruction * @return the location assigned for the operand */ private Value colorLirOperand(Variable operand, int opId, OperandMode mode) { Interval interval = intervalFor(operand); assert interval != null : "interval must exist"; if (opId != -1) { if (DetailedAsserts.getValue()) { AbstractBlock<?> block = blockForId(opId); if (block.getSuccessorCount() <= 1 && opId == getLastLirInstructionId(block)) { // check if spill moves could have been appended at the end of this block, but // before the branch instruction. So the split child information for this branch // would // be incorrect. LIRInstruction instr = ir.getLIRforBlock(block).get(ir.getLIRforBlock(block).size() - 1); if (instr instanceof StandardOp.JumpOp) { if (blockData.get(block).liveOut.get(operandNumber(operand))) { assert false : "can't get split child for the last branch of a block because the information would be incorrect (moves are inserted before the branch in resolveDataFlow)"; } } } } // operands are not changed when an interval is split during allocation, // so search the right interval here interval = splitChildAtOpId(interval, opId, mode); } if (isIllegal(interval.location()) && interval.canMaterialize()) { assert mode != OperandMode.DEF; return interval.getMaterializedValue(); } return interval.location(); } private boolean isMaterialized(AllocatableValue operand, int opId, OperandMode mode) { Interval interval = intervalFor(operand); assert interval != null : "interval must exist"; if (opId != -1) { // operands are not changed when an interval is split during allocation, // so search the right interval here interval = splitChildAtOpId(interval, opId, mode); } return isIllegal(interval.location()) && interval.canMaterialize(); } protected IntervalWalker initIntervalWalker(IntervalPredicate predicate) { // setup lists of potential oops for walking Interval oopIntervals; Interval nonOopIntervals; oopIntervals = createUnhandledLists(predicate, null).first; // intervals that have no oops inside need not to be processed. // to ensure a walking until the last instruction id, add a dummy interval // with a high operation id nonOopIntervals = new Interval(Value.ILLEGAL, -1); nonOopIntervals.addRange(Integer.MAX_VALUE - 2, Integer.MAX_VALUE - 1); return new IntervalWalker(this, oopIntervals, nonOopIntervals); } /** * Visits all intervals for a frame state. The frame state use this information to build the OOP * maps. */ void markFrameLocations(IntervalWalker iw, LIRInstruction op, LIRFrameState info) { Debug.log("creating oop map at opId %d", op.id()); // walk before the current operation . intervals that start at // the operation (i.e. output operands of the operation) are not // included in the oop map iw.walkBefore(op.id()); // TODO(je) we could pass this as parameter AbstractBlock<?> block = blockForId(op.id()); // Iterate through active intervals for (Interval interval = iw.activeLists.get(RegisterBinding.Fixed); interval != Interval.EndMarker; interval = interval.next) { Value operand = interval.operand; assert interval.currentFrom() <= op.id() && op.id() <= interval.currentTo() : "interval should not be active otherwise"; assert isVariable(interval.operand) : "fixed interval found"; // Check if this range covers the instruction. Intervals that // start or end at the current operation are not included in the // oop map, except in the case of patching moves. For patching // moves, any intervals which end at this instruction are included // in the oop map since we may safepoint while doing the patch // before we've consumed the inputs. if (op.id() < interval.currentTo() && !isIllegal(interval.location())) { // caller-save registers must not be included into oop-maps at calls assert !op.destroysCallerSavedRegisters() || !isRegister(operand) || !isCallerSave(operand) : "interval is in a caller-save register at a call . register will be overwritten"; info.markLocation(interval.location(), frameMap); // Spill optimization: when the stack value is guaranteed to be always correct, // then it must be added to the oop map even if the interval is currently in a // register int spillPos = interval.spillDefinitionPos(); if (interval.spillState() != SpillState.SpillInDominator) { if (interval.alwaysInMemory() && op.id() > interval.spillDefinitionPos() && !interval.location().equals(interval.spillSlot())) { assert interval.spillDefinitionPos() > 0 : "position not set correctly"; assert spillPos > 0 : "position not set correctly"; assert interval.spillSlot() != null : "no spill slot assigned"; assert !isRegister(interval.operand) : "interval is on stack : so stack slot is registered twice"; info.markLocation(interval.spillSlot(), frameMap); } } else { AbstractBlock<?> spillBlock = blockForId(spillPos); if (interval.alwaysInMemory() && !interval.location().equals(interval.spillSlot())) { if ((spillBlock.equals(block) && op.id() > spillPos) || AbstractBlock.dominates(spillBlock, block)) { assert spillPos > 0 : "position not set correctly"; assert interval.spillSlot() != null : "no spill slot assigned"; assert !isRegister(interval.operand) : "interval is on stack : so stack slot is registered twice"; info.markLocation(interval.spillSlot(), frameMap); } } } } } } private boolean isCallerSave(Value operand) { return attributes(asRegister(operand)).isCallerSave(); } private InstructionValueProcedure debugInfoProc = new InstructionValueProcedure() { @Override public Value doValue(LIRInstruction op, Value operand) { int tempOpId = op.id(); OperandMode mode = OperandMode.USE; AbstractBlock<?> block = blockForId(tempOpId); if (block.getSuccessorCount() == 1 && tempOpId == getLastLirInstructionId(block)) { // generating debug information for the last instruction of a block. // if this instruction is a branch, spill moves are inserted before this branch // and so the wrong operand would be returned (spill moves at block boundaries // are not // considered in the live ranges of intervals) // Solution: use the first opId of the branch target block instead. final LIRInstruction instr = ir.getLIRforBlock(block).get(ir.getLIRforBlock(block).size() - 1); if (instr instanceof StandardOp.JumpOp) { if (blockData.get(block).liveOut.get(operandNumber(operand))) { tempOpId = getFirstLirInstructionId(block.getSuccessors().iterator().next()); mode = OperandMode.DEF; } } } // Get current location of operand // The operand must be live because debug information is considered when building // the intervals // if the interval is not live, colorLirOperand will cause an assert on failure Value result = colorLirOperand((Variable) operand, tempOpId, mode); assert !hasCall(tempOpId) || isStackSlot(result) || isConstant(result) || !isCallerSave(result) : "cannot have caller-save register operands at calls"; return result; } }; private void computeDebugInfo(IntervalWalker iw, final LIRInstruction op, LIRFrameState info) { info.initDebugInfo(frameMap, !op.destroysCallerSavedRegisters() || !callKillsRegisters); markFrameLocations(iw, op, info); info.forEachState(op, debugInfoProc); info.finish(op, frameMap); } private void assignLocations(List<LIRInstruction> instructions, final IntervalWalker iw) { int numInst = instructions.size(); boolean hasDead = false; InstructionValueProcedure assignProc = new InstructionValueProcedure() { @Override public Value doValue(LIRInstruction op, Value operand, OperandMode mode, EnumSet<OperandFlag> flags) { if (isVariable(operand)) { return colorLirOperand((Variable) operand, op.id(), mode); } return operand; } }; InstructionStateProcedure stateProc = new InstructionStateProcedure() { @Override protected void doState(LIRInstruction op, LIRFrameState state) { computeDebugInfo(iw, op, state); } }; for (int j = 0; j < numInst; j++) { final LIRInstruction op = instructions.get(j); if (op == null) { // this can happen when spill-moves are removed in eliminateSpillMoves hasDead = true; continue; } // remove useless moves MoveOp move = null; if (op instanceof MoveOp) { move = (MoveOp) op; AllocatableValue result = move.getResult(); if (isVariable(result) && isMaterialized(result, op.id(), OperandMode.DEF)) { /* * This happens if a materializable interval is originally not spilled but then * kicked out in LinearScanWalker.splitForSpilling(). When kicking out such an * interval this move operation was already generated. */ instructions.set(j, null); hasDead = true; continue; } } op.forEachInput(assignProc); op.forEachAlive(assignProc); op.forEachTemp(assignProc); op.forEachOutput(assignProc); // compute reference map and debug information op.forEachState(stateProc); // remove useless moves if (move != null) { if (move.getInput().equals(move.getResult())) { instructions.set(j, null); hasDead = true; } } } if (hasDead) { // Remove null values from the list. instructions.removeAll(Collections.singleton(null)); } } private void assignLocations() { IntervalWalker iw = initIntervalWalker(IS_STACK_INTERVAL); try (Indent indent = Debug.logAndIndent("assign locations")) { for (AbstractBlock<?> block : sortedBlocks) { try (Indent indent2 = Debug.logAndIndent("assign locations in block B%d", block.getId())) { assignLocations(ir.getLIRforBlock(block), iw); } } } } public void allocate() { /* * This is the point to enable debug logging for the whole register allocation. */ try (Indent indent = Debug.logAndIndent("LinearScan allocate")) { try (Scope s = Debug.scope("LifetimeAnalysis")) { numberInstructions(); printLir("Before register allocation", true); computeLocalLiveSets(); computeGlobalLiveSets(); buildIntervals(); sortIntervalsBeforeAllocation(); } catch (Throwable e) { throw Debug.handle(e); } try (Scope s = Debug.scope("RegisterAllocation")) { printIntervals("Before register allocation"); allocateRegisters(); } catch (Throwable e) { throw Debug.handle(e); } if (Options.LSRAOptimizeSpillPosition.getValue()) { try (Scope s = Debug.scope("OptimizeSpillPosition")) { optimizeSpillPosition(); } catch (Throwable e) { throw Debug.handle(e); } } try (Scope s = Debug.scope("ResolveDataFlow")) { resolveDataFlow(); } catch (Throwable e) { throw Debug.handle(e); } try (Scope s = Debug.scope("DebugInfo")) { frameMap.finish(); printIntervals("After register allocation"); printLir("After register allocation", true); sortIntervalsAfterAllocation(); if (DetailedAsserts.getValue()) { verify(); } eliminateSpillMoves(); assignLocations(); if (DetailedAsserts.getValue()) { verifyIntervals(); } } catch (Throwable e) { throw Debug.handle(e); } printLir("After register number assignment", true); } } private DebugMetric betterSpillPos = Debug.metric("BetterSpillPosition"); private DebugMetric betterSpillPosWithLowerProbability = Debug.metric("BetterSpillPositionWithLowerProbability"); private void optimizeSpillPosition() { for (Interval interval : intervals) { if (interval != null && interval.isSplitParent() && interval.spillState() == SpillState.SpillInDominator) { AbstractBlock<?> defBlock = blockForId(interval.spillDefinitionPos()); AbstractBlock<?> spillBlock = null; Interval firstSpillChild = null; try (Indent indent = Debug.logAndIndent("interval %s (%s)", interval, defBlock)) { for (Interval splitChild : interval.getSplitChildren()) { if (isStackSlot(splitChild.location())) { if (firstSpillChild == null || splitChild.from() < firstSpillChild.from()) { firstSpillChild = splitChild; } else { assert firstSpillChild.from() < splitChild.from(); } // iterate all blocks where the interval has use positions for (AbstractBlock<?> splitBlock : blocksForSplitChild(splitChild)) { assert dominates(defBlock, splitBlock) : String.format("Definition does not dominate the spill block %s !dom %s (interval %s)", defBlock, splitBlock, interval); Debug.log("Split interval %s, block %s", splitChild, splitBlock); if (spillBlock == null) { spillBlock = splitBlock; } else { spillBlock = nearestCommonDominator(spillBlock, splitBlock); assert spillBlock != null; } } } } if (spillBlock == null) { // no spill interval interval.setSpillState(SpillState.StoreAtDefinition); } else { // move out of loops if (defBlock.getLoopDepth() < spillBlock.getLoopDepth()) { spillBlock = moveSpillOutOfLoop(defBlock, spillBlock); } /* * If the spill block is the begin of the first split child (aka the value * is on the stack) spill in the dominator. */ assert firstSpillChild != null; if (!defBlock.equals(spillBlock) && spillBlock.equals(blockForId(firstSpillChild.from()))) { AbstractBlock<?> dom = spillBlock.getDominator(); Debug.log("Spill block (%s) is the beginning of a spill child -> use dominator (%s)", spillBlock, dom); spillBlock = dom; } if (!defBlock.equals(spillBlock)) { assert dominates(defBlock, spillBlock); betterSpillPos.increment(); Debug.log("Better spill position found (Block %s)", spillBlock); if (defBlock.probability() <= spillBlock.probability()) { // better spill block has the same probability -> do nothing assert spillBlock.probability() - defBlock.probability() < 0.00000001 : "Check whether the probability difference is within epsilon"; interval.setSpillState(SpillState.StoreAtDefinition); } else { betterSpillPosWithLowerProbability.increment(); interval.setSpillDefinitionPos(getFirstLirInstructionId(spillBlock)); } } else { // definition is the best choice interval.setSpillState(SpillState.StoreAtDefinition); } } } } } } /** * Iterate over all {@link AbstractBlock blocks} of an interval with an use position. */ private class UseBlockIterator implements Iterator<AbstractBlock<?>> { int nextOpId; Interval interval; public UseBlockIterator(Interval interval) { // the first use position is the begin of the interval nextOpId = interval.from(); this.interval = interval; } public AbstractBlock<?> next() { AbstractBlock<?> block = blockForId(nextOpId); int nextBlockIndex = block.getLinearScanNumber() + 1; if (nextBlockIndex < sortedBlocks.size()) { int from = getFirstLirInstructionId(sortedBlocks.get(nextBlockIndex)); assert from > nextOpId : "cannot go backwards"; assert from > maxOpId() || isBlockBegin(from) : "nextOpId is not a block begin"; assert from > maxOpId() || blockForId(from).getLinearScanNumber() == block.getLinearScanNumber() + 1 : "nextOpId is not the beginning of the next block"; nextOpId = interval.nextUsage(RegisterPriority.None, from); } else { // already at last the last block nextOpId = Integer.MAX_VALUE; } return block; } public boolean hasNext() { return nextOpId < interval.to(); } } private Iterable<AbstractBlock<?>> blocksForSplitChild(Interval interval) { return new Iterable<AbstractBlock<?>>() { public Iterator<AbstractBlock<?>> iterator() { return new UseBlockIterator(interval); } }; } private static AbstractBlock<?> moveSpillOutOfLoop(AbstractBlock<?> defBlock, AbstractBlock<?> spillBlock) { int defLoopDepth = defBlock.getLoopDepth(); for (AbstractBlock<?> block = spillBlock.getDominator(); !defBlock.equals(block); block = block.getDominator()) { assert block != null : "spill block not dominated by definition block?"; if (block.getLoopDepth() <= defLoopDepth) { assert block.getLoopDepth() == defLoopDepth : "Cannot spill an interval outside of the loop where it is defined!"; return block; } } return defBlock; } private AbstractBlock<?> nearestCommonDominator(AbstractBlock<?> a, AbstractBlock<?> b) { assert a != null; assert b != null; try (Indent indent = Debug.logAndIndent("nearest common dominator of %s and %s", a, b)) { if (a.equals(b)) { return a; } // collect a's dominators BitSet aDom = new BitSet(sortedBlocks.size()); // a != b for (AbstractBlock<?> x = a; x != null; x = x.getDominator()) { aDom.set(x.getId()); } // walk b's dominator for (AbstractBlock<?> x = b; x != null; x = x.getDominator()) { if (aDom.get(x.getId())) { Debug.log("found %s", x); return x; } } } Debug.log("no common dominator found"); return null; } void printIntervals(String label) { if (Debug.isLogEnabled()) { try (Indent indent = Debug.logAndIndent("intervals %s", label)) { for (Interval interval : intervals) { if (interval != null) { Debug.log("%s", interval.logString(this)); } } try (Indent indent2 = Debug.logAndIndent("Basic Blocks")) { for (int i = 0; i < blockCount(); i++) { AbstractBlock<?> block = blockAt(i); Debug.log("B%d [%d, %d, %s] ", block.getId(), getFirstLirInstructionId(block), getLastLirInstructionId(block), block.getLoop()); } } } } Debug.dump(Arrays.copyOf(intervals, intervalsSize), label); } void printLir(String label, @SuppressWarnings("unused") boolean hirValid) { Debug.dump(ir, label); } boolean verify() { // (check that all intervals have a correct register and that no registers are overwritten) verifyIntervals(); verifyRegisters(); Debug.log("no errors found"); return true; } private void verifyRegisters() { // Enable this logging to get output for the verification process. try (Indent indent = Debug.logAndIndent("verifying register allocation")) { RegisterVerifier verifier = new RegisterVerifier(this); verifier.verify(blockAt(0)); } } void verifyIntervals() { try (Indent indent = Debug.logAndIndent("verifying intervals")) { int len = intervalsSize; for (int i = 0; i < len; i++) { Interval i1 = intervals[i]; if (i1 == null) { continue; } i1.checkSplitChildren(); if (i1.operandNumber != i) { Debug.log("Interval %d is on position %d in list", i1.operandNumber, i); Debug.log(i1.logString(this)); throw new GraalInternalError(""); } if (isVariable(i1.operand) && i1.kind().equals(LIRKind.Illegal)) { Debug.log("Interval %d has no type assigned", i1.operandNumber); Debug.log(i1.logString(this)); throw new GraalInternalError(""); } if (i1.location() == null) { Debug.log("Interval %d has no register assigned", i1.operandNumber); Debug.log(i1.logString(this)); throw new GraalInternalError(""); } if (i1.first() == Range.EndMarker) { Debug.log("Interval %d has no Range", i1.operandNumber); Debug.log(i1.logString(this)); throw new GraalInternalError(""); } for (Range r = i1.first(); r != Range.EndMarker; r = r.next) { if (r.from >= r.to) { Debug.log("Interval %d has zero length range", i1.operandNumber); Debug.log(i1.logString(this)); throw new GraalInternalError(""); } } for (int j = i + 1; j < len; j++) { Interval i2 = intervals[j]; if (i2 == null) { continue; } // special intervals that are created in MoveResolver // . ignore them because the range information has no meaning there if (i1.from() == 1 && i1.to() == 2) { continue; } if (i2.from() == 1 && i2.to() == 2) { continue; } Value l1 = i1.location(); Value l2 = i2.location(); if (i1.intersects(i2) && !isIllegal(l1) && (l1.equals(l2))) { if (DetailedAsserts.getValue()) { Debug.log("Intervals %d and %d overlap and have the same register assigned", i1.operandNumber, i2.operandNumber); Debug.log(i1.logString(this)); Debug.log(i2.logString(this)); } throw new BailoutException(""); } } } } } class CheckProcedure extends ValueProcedure { boolean ok; Interval curInterval; @Override protected Value doValue(Value operand) { if (isRegister(operand)) { if (intervalFor(operand) == curInterval) { ok = true; } } return operand; } } void verifyNoOopsInFixedIntervals() { try (Indent indent = Debug.logAndIndent("verifying that no oops are in fixed intervals *")) { CheckProcedure checkProc = new CheckProcedure(); Interval fixedIntervals; Interval otherIntervals; fixedIntervals = createUnhandledLists(IS_PRECOLORED_INTERVAL, null).first; // to ensure a walking until the last instruction id, add a dummy interval // with a high operation id otherIntervals = new Interval(Value.ILLEGAL, -1); otherIntervals.addRange(Integer.MAX_VALUE - 2, Integer.MAX_VALUE - 1); IntervalWalker iw = new IntervalWalker(this, fixedIntervals, otherIntervals); for (AbstractBlock<?> block : sortedBlocks) { List<LIRInstruction> instructions = ir.getLIRforBlock(block); for (int j = 0; j < instructions.size(); j++) { LIRInstruction op = instructions.get(j); if (op.hasState()) { iw.walkBefore(op.id()); boolean checkLive = true; // Make sure none of the fixed registers is live across an // oopmap since we can't handle that correctly. if (checkLive) { for (Interval interval = iw.activeLists.get(RegisterBinding.Fixed); interval != Interval.EndMarker; interval = interval.next) { if (interval.currentTo() > op.id() + 1) { // This interval is live out of this op so make sure // that this interval represents some value that's // referenced by this op either as an input or output. checkProc.curInterval = interval; checkProc.ok = false; op.forEachInput(checkProc); op.forEachAlive(checkProc); op.forEachTemp(checkProc); op.forEachOutput(checkProc); assert checkProc.ok : "fixed intervals should never be live across an oopmap point"; } } } } } } } } /** * Returns a value for a interval definition, which can be used for re-materialization. * * @param op An instruction which defines a value * @param operand The destination operand of the instruction * @param interval The interval for this defined value. * @return Returns the value which is moved to the instruction and which can be reused at all * reload-locations in case the interval of this instruction is spilled. Currently this * can only be a {@link Constant}. */ public static Constant getMaterializedValue(LIRInstruction op, Value operand, Interval interval) { if (op instanceof MoveOp) { MoveOp move = (MoveOp) op; if (move.getInput() instanceof Constant) { /* * Check if the interval has any uses which would accept an stack location (priority * == ShouldHaveRegister). Rematerialization of such intervals can result in a * degradation, because rematerialization always inserts a constant load, even if * the value is not needed in a register. */ Interval.UsePosList usePosList = interval.usePosList(); int numUsePos = usePosList.size(); for (int useIdx = 0; useIdx < numUsePos; useIdx++) { Interval.RegisterPriority priority = usePosList.registerPriority(useIdx); if (priority == Interval.RegisterPriority.ShouldHaveRegister) { return null; } } return (Constant) move.getInput(); } } return null; } }