/* * Copyright (c) 2009, 2011, 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 org.graalvm.compiler.core.common.alloc; import java.util.ArrayList; import java.util.BitSet; import java.util.Comparator; import java.util.List; import java.util.PriorityQueue; import org.graalvm.compiler.core.common.cfg.AbstractBlockBase; import org.graalvm.compiler.core.common.cfg.Loop; /** * Computes an ordering of the block that can be used by the linear scan register allocator and the * machine code generator. The machine code generation order will start with the first block and * produce a straight sequence always following the most likely successor. Then it will continue * with the most likely path that was left out during this process. The process iteratively * continues until all blocks are scheduled. Additionally, it is guaranteed that all blocks of a * loop are scheduled before any block following the loop is scheduled. * * The machine code generator order includes reordering of loop headers such that the backward jump * is a conditional jump if there is only one loop end block. Additionally, the target of loop * backward jumps are always marked as aligned. Aligning the target of conditional jumps does not * bring a measurable benefit and is therefore avoided to keep the code size small. * * The linear scan register allocator order has an additional mechanism that prevents merge nodes * from being scheduled if there is at least one highly likely predecessor still unscheduled. This * increases the probability that the merge node and the corresponding predecessor are more closely * together in the schedule thus decreasing the probability for inserted phi moves. Also, the * algorithm sets the linear scan order number of the block that corresponds to its index in the * linear scan order. */ public final class ComputeBlockOrder { /** * The initial capacities of the worklists used for iteratively finding the block order. */ private static final int INITIAL_WORKLIST_CAPACITY = 10; /** * Divisor used for degrading the probability of the current path versus unscheduled paths at a * merge node when calculating the linear scan order. A high value means that predecessors of * merge nodes are more likely to be scheduled before the merge node. */ private static final int PENALTY_VERSUS_UNSCHEDULED = 10; /** * Computes the block order used for the linear scan register allocator. * * @return sorted list of blocks */ public static > AbstractBlockBase[] computeLinearScanOrder(int blockCount, T startBlock) { List order = new ArrayList<>(); BitSet visitedBlocks = new BitSet(blockCount); PriorityQueue worklist = initializeWorklist(startBlock, visitedBlocks); computeLinearScanOrder(order, worklist, visitedBlocks); assert checkOrder(order, blockCount); return order.toArray(new AbstractBlockBase[0]); } /** * Computes the block order used for code emission. * * @return sorted list of blocks */ public static > AbstractBlockBase[] computeCodeEmittingOrder(int blockCount, T startBlock) { List order = new ArrayList<>(); BitSet visitedBlocks = new BitSet(blockCount); PriorityQueue worklist = initializeWorklist(startBlock, visitedBlocks); computeCodeEmittingOrder(order, worklist, visitedBlocks); assert checkOrder(order, blockCount); return order.toArray(new AbstractBlockBase[0]); } /** * Iteratively adds paths to the code emission block order. */ private static > void computeCodeEmittingOrder(List order, PriorityQueue worklist, BitSet visitedBlocks) { while (!worklist.isEmpty()) { T nextImportantPath = worklist.poll(); addPathToCodeEmittingOrder(nextImportantPath, order, worklist, visitedBlocks); } } /** * Iteratively adds paths to the linear scan block order. */ private static > void computeLinearScanOrder(List order, PriorityQueue worklist, BitSet visitedBlocks) { while (!worklist.isEmpty()) { T nextImportantPath = worklist.poll(); do { nextImportantPath = addPathToLinearScanOrder(nextImportantPath, order, worklist, visitedBlocks); } while (nextImportantPath != null); } } /** * Initializes the priority queue used for the work list of blocks and adds the start block. */ private static > PriorityQueue initializeWorklist(T startBlock, BitSet visitedBlocks) { PriorityQueue result = new PriorityQueue<>(INITIAL_WORKLIST_CAPACITY, new BlockOrderComparator<>()); result.add(startBlock); visitedBlocks.set(startBlock.getId()); return result; } /** * Add a linear path to the linear scan order greedily following the most likely successor. */ private static > T addPathToLinearScanOrder(T block, List order, PriorityQueue worklist, BitSet visitedBlocks) { block.setLinearScanNumber(order.size()); order.add(block); T mostLikelySuccessor = findAndMarkMostLikelySuccessor(block, visitedBlocks); enqueueSuccessors(block, worklist, visitedBlocks); if (mostLikelySuccessor != null) { if (!mostLikelySuccessor.isLoopHeader() && mostLikelySuccessor.getPredecessorCount() > 1) { // We are at a merge. Check probabilities of predecessors that are not yet // scheduled. double unscheduledSum = 0.0; for (T pred : mostLikelySuccessor.getPredecessors()) { if (pred.getLinearScanNumber() == -1) { unscheduledSum += pred.getRelativeFrequency(); } } if (unscheduledSum > block.getRelativeFrequency() / PENALTY_VERSUS_UNSCHEDULED) { // Add this merge only after at least one additional predecessor gets scheduled. visitedBlocks.clear(mostLikelySuccessor.getId()); return null; } } return mostLikelySuccessor; } return null; } /** * Add a linear path to the code emission order greedily following the most likely successor. */ private static > void addPathToCodeEmittingOrder(T initialBlock, List order, PriorityQueue worklist, BitSet visitedBlocks) { T block = initialBlock; while (block != null) { // Skip loop headers if there is only a single loop end block to // make the backward jump be a conditional jump. if (!skipLoopHeader(block)) { // Align unskipped loop headers as they are the target of the backward jump. if (block.isLoopHeader()) { block.setAlign(true); } addBlock(block, order); } Loop loop = block.getLoop(); if (block.isLoopEnd() && skipLoopHeader(loop.getHeader())) { // This is the only loop end of a skipped loop header. // Add the header immediately afterwards. addBlock(loop.getHeader(), order); // Make sure the loop successors of the loop header are aligned // as they are the target // of the backward jump. for (T successor : loop.getHeader().getSuccessors()) { if (successor.getLoopDepth() == block.getLoopDepth()) { successor.setAlign(true); } } } T mostLikelySuccessor = findAndMarkMostLikelySuccessor(block, visitedBlocks); enqueueSuccessors(block, worklist, visitedBlocks); block = mostLikelySuccessor; } } /** * Adds a block to the ordering. */ private static > void addBlock(T header, List order) { assert !order.contains(header) : "Cannot insert block twice"; order.add(header); } /** * Find the highest likely unvisited successor block of a given block. */ private static > T findAndMarkMostLikelySuccessor(T block, BitSet visitedBlocks) { T result = null; for (T successor : block.getSuccessors()) { assert successor.getRelativeFrequency() >= 0.0 : "Relative frequencies must be positive"; if (!visitedBlocks.get(successor.getId()) && successor.getLoopDepth() >= block.getLoopDepth() && (result == null || successor.getRelativeFrequency() >= result.getRelativeFrequency())) { result = successor; } } if (result != null) { visitedBlocks.set(result.getId()); } return result; } /** * Add successor blocks into the given work list if they are not already marked as visited. */ private static > void enqueueSuccessors(T block, PriorityQueue worklist, BitSet visitedBlocks) { for (T successor : block.getSuccessors()) { if (!visitedBlocks.get(successor.getId())) { visitedBlocks.set(successor.getId()); worklist.add(successor); } } } /** * Skip the loop header block if the loop consists of more than one block and it has only a * single loop end block. */ private static > boolean skipLoopHeader(AbstractBlockBase block) { return (block.isLoopHeader() && !block.isLoopEnd() && block.getLoop().numBackedges() == 1); } /** * Checks that the ordering contains the expected number of blocks. */ private static boolean checkOrder(List> order, int expectedBlockCount) { assert order.size() == expectedBlockCount : String.format("Number of blocks in ordering (%d) does not match expected block count (%d)", order.size(), expectedBlockCount); return true; } /** * Comparator for sorting blocks based on loop depth and probability. */ private static class BlockOrderComparator> implements Comparator { private static final double EPSILON = 1E-6; @Override public int compare(T a, T b) { // Loop blocks before any loop exit block. The only exception are blocks that are // (almost) impossible to reach. if (a.getRelativeFrequency() > EPSILON && b.getRelativeFrequency() > EPSILON) { int diff = b.getLoopDepth() - a.getLoopDepth(); if (diff != 0) { return diff; } } // Blocks with high probability before blocks with low probability. if (a.getRelativeFrequency() > b.getRelativeFrequency()) { return -1; } else { return 1; } } } }