Mercurial > hg > truffle
diff graal/GraalCompiler/src/com/sun/c1x/graph/package-info.java @ 2509:16b9a8b5ad39
Renamings Runtime=>GraalRuntime and Compiler=>GraalCompiler
author | Thomas Wuerthinger <thomas@wuerthinger.net> |
---|---|
date | Wed, 27 Apr 2011 11:50:44 +0200 |
parents | graal/Compiler/src/com/sun/c1x/graph/package-info.java@9ec15d6914ca |
children |
line wrap: on
line diff
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/graal/GraalCompiler/src/com/sun/c1x/graph/package-info.java Wed Apr 27 11:50:44 2011 +0200 @@ -0,0 +1,153 @@ +/* + * Copyright (c) 2010, 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. + */ + +/** + * <H2>IR Graph building</H2> + * + * The {@link com.sun.c1x.graph.IR} class drives the generation of the HIR graph for a method, making use of other + * utility classes in this package. + * + * The graph building is separated into a basic build phase ({@link com.sun.c1x.graph.IR#buildGraph} method)and + * (currently) two optimization phases ({@link com.sun.c1x.graph.IR#optimize1} and + * {@link com.sun.c1x.graph.IR#optimize2}) although the basic phase also does some (basic) optimizations. + * + * <H2>Basic Graph Build Phase</H2> + * + * {@code IR.buildGraph} creates an {@link com.sun.c1x.ir.IRScope topScope} object, + * that represents a context for inlining, and then invokes the constructor for the + * {@link com.sun.c1x.graph.GraphBuilder} class, passing the {@link com.sun.c1x.C1XCompilation}, and {@code IR} + * instances, which are cached. The following support objects are created in the constructor: + * + * <ul> + * <li>{@code memoryMap}: an instance of {@link com.sun.c1x.graph.MemoryMap} + * <li>{@code localValueMap}: an instance of {@link com.sun.c1x.opt.ValueMap} + * <li>{@code canonicalizer}: an instance of {@link com.sun.c1x.opt.Canonicalizer} + * </ul> + * + * Now the {@link com.sun.c1x.graph.GraphBuilder#build} is invoked with {@code topScope} as argument. + * + * <H3>{@code GraphBuilder.build}</H3> + * + * <ol> + * <li>The {@link com.sun.c1x.graph.IR#startBlock} field of the cached {@link com.sun.c1x.graph.IR} instance is set to a newly created + * {@link com.sun.c1x.ir.BlockBegin} node, with bytecode index 0 and then the {@link com.sun.c1x.graph.BlockMap} is + * constructed by calling {@link com.sun.c1x.C1XCompilation#getBlockMap}. This behaves slightly differently depending on + * whether this is an OSR compilation. If so, a new {@link com.sun.c1x.ir.BlockBegin} node is added to the map at the OSR bytecode + * index. The map is then built by the{@link com.sun.c1x.graph.BlockMap#build}, which takes a boolean argument that + * controls whether a second pass is made over the bytecodes to compute stores in loops. This always false for an OSR + * compilation (why?). Otherwise, it is only true if enabled by the {@link com.sun.c1x.C1XOptions#PhiLoopStores} + * compilation option. FInally some unneeded state from the map is removed by the {@link com.sun.c1x.graph.BlockMap#cleanup} method, and + * the statistics are updated. + * </li> + * + * <li>Next the {@link com.sun.c1x.graph.GraphBuilder#pushRootScope} method is called, with the passed-in {@link com.sun.c1x.ir.IRScope} + * object, the {@link com.sun.c1x.graph.BlockMap} returned by build and the {@code startBlock}. (Note: Unlike + * {@link com.sun.c1x.graph.GraphBuilder#pushScope}, this method does not propagate the + * {@link com.sun.c1x.graph.BlockMap#storesInLoops} field to the {@link com.sun.c1x.ir.IRScope} object, which means that + * {@link com.sun.c1x.ir.BlockBegin#insertLoopPhis} will always get null for this value. Is this a bug?). + * {@link com.sun.c1x.graph.GraphBuilder#pushRootScope} initializes the {@link com.sun.c1x.graph.GraphBuilder#scopeData} field with a + * {@link com.sun.c1x.graph.ScopeData} instance, with null parent. The + * {@link com.sun.c1x.graph.GraphBuilder#compilation} instance is called to get an {@link com.sun.cri.ri.RiConstantPool} + * , which is C1X's interface to constant pool information. The {@link com.sun.c1x.graph.GraphBuilder#curBlock} field is + * set to the {@code startBlock}. + * <p> + * + * Now a {@link com.sun.c1x.value.FrameState initialState} object is created by + * {@link com.sun.c1x.graph.GraphBuilder#stateAtEntry}. If the method is not static, then a {@link com.sun.c1x.ir.Local} + * instance is created at index 0. Since the receiver cannot be {@code null}, the + * {@link com.sun.c1x.ir.Value.Flag#NonNull} flag is set. Additional {@link com.sun.c1x.ir.Local} instances are created for the + * arguments to the method. The index is incremented by the number of slots occupied by the + * {@link com.sun.cri.ci.CiKind} corresponding to the argument type. All the {@link com.sun.c1x.ir.Local} instances are stored in the + * {@link com.sun.c1x.value.FrameState} using the {@link com.sun.c1x.value.FrameState#storeLocal} method. This {@link com.sun.c1x.value.FrameState} is then + * merged into the {@link com.sun.c1x.ir.BlockBegin#stateBefore} for the {@code startBlock}, which just results in a + * copy since {@code stateBefore} will be {@code null}. + * </li> + * <li> + * This step sets up three instance fields: {@link com.sun.c1x.graph.GraphBuilder#curBlock} and + * {@link com.sun.c1x.graph.GraphBuilder#lastInstr} to {@code startBlock} and + * {@link com.sun.c1x.graph.GraphBuilder#curState} to {@code initialState}. (N.B. the setting of {@code curBlock} is + * redundant as it is done in {@link com.sun.c1x.graph.GraphBuilder#pushRootScope}). + * </li> + * <li> + * Step 4 contains special handling for synchronized methods (TBD), otherwise it calls + * {@link com.sun.c1x.graph.GraphBuilder#finishStartBlock} which adds a {@link com.sun.c1x.ir.Base} block as the end of + * the {@code startBlock}. The {@link com.sun.c1x.ir.Base} block has one successor set to the (entry) block with flag + * {@link com.sun.c1x.ir.BlockBegin.BlockFlag#StandardEntry}, that was created by {@link com.sun.c1x.graph.BlockMap#build} (and possibly a + * successor to an OSREntry block). + * </li> + * <li> + * Then the {@link com.sun.c1x.ir.IRScope#lockStackSize} is computed. (TBD) + * </li> + * <li> + * Then the method is checked for being intrinsic, i.e., one that has a hard-wired implementation known to C1X. If so, + * and {@link com.sun.c1x.C1XOptions#OptIntrinsify} is set, an attempt is made to inline it (TBD). Otherwise, or if the + * intrinsification fails, normal processing continues by adding the entry block to the + * {@link com.sun.c1x.graph.ScopeData} work list (kept topologically sorted) and calling + * {@link com.sun.c1x.graph.GraphBuilder#iterateAllBlocks}. + * </li> + * <li> + * Finally there is some cleanup code for synchronized blocks and OSR compilations. + * </li> + * </ol> + * + * <H3>{@link com.sun.c1x.graph.GraphBuilder#iterateAllBlocks}</H3> + * {@link com.sun.c1x.graph#iterateAllBlocks} repeatedly removes a block from the work list and, if not already visited, marks it so, + * kills the current memory map, sets {@link com.sun.c1x.graph.GraphBuilder#curBlock}, {@link com.sun.c1x.graph.GraphBuilder#curState} and {@link com.sun.c1x.graph.GraphBuilder#lastInstr} and then calls + * {@link com.sun.c1x.graph.GraphBuilder#iterateBytecodesForBlock}. + * + * This process continues until all the blocks have been visited (processed) after which control returns to {@code + * build}. + * <p> + + * <H3>{@link com.sun.c1x.graph.GraphBuilder#iterateBytecodesForBlock}</H3> + * + * {@link com.sun.c1x.graph.GraphBuilder#iterateBytecodesForBlock} performs an abstract interpretation of the bytecodes in the block, appending new + * nodes as necessary, until the last added node is an instance of {@link com.sun.c1x.ir.BlockEnd}. (Note: It has an + * explicit check for finding a new {@link com.sun.c1x.ir.BlockBegin} before a {@link com.sun.c1x.ir.BlockEnd} but + * {@link com.sun.c1x.graph.BlockMap#moveSuccessorLists} has a similar check so this may be redundant). For example, + * consider the following bytecodes: + * + * <pre> + * <code> + * 0: iconst_0 + * 1: istore_2 + * 2: goto 22 + * </code> + * </pre> + * + * The {@code iconst_0} bytecode causes a {@link com.sun.c1x.ir.Constant} node representing zero to be pushed on the + * {@link com.sun.c1x.graph.GraphBuilder#curState} stack and the node to be appended to the {@link com.sun.c1x.ir.BlockBegin} (entry) node associated with index 0. + * The {@code istore_2} causes the node to be popped of the stack and stored in the local slot 2. No IR node is + * generated for the {@code istore_2}. The {@code goto} creates a {@link com.sun.c1x.ir.Goto} node which is a subclass + * of {@link com.sun.c1x.ir.BlockEnd}, so this terminates the iteration. As part of termination the {@link com.sun.c1x.ir.Goto} node is marked as the + * end node of the current block and the {@link com.sun.c1x.value.FrameState} is propagated to the successor node(s) by merging any + * existing {@link com.sun.c1x.value.FrameState} with the current state. If the target is a loop header node this involves inserting + * {@link com.sun.c1x.ir.Phi} nodes. Finally, the target node is added to the {@code scopeData} work list. + * <p> + * + * + * @author Ben Titzer + * @author Mick Jordan + * + */ +package com.sun.c1x.graph;