1 /*
   2  * Copyright (c) 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  */
  23 
  24 
  25 package org.graalvm.compiler.lir.aarch64;
  26 
  27 import static jdk.vm.ci.aarch64.AArch64.zr;
  28 import static jdk.vm.ci.code.ValueUtil.asRegister;
  29 import static org.graalvm.compiler.lir.LIRInstruction.OperandFlag.REG;
  30 
  31 import org.graalvm.compiler.asm.Label;
  32 import org.graalvm.compiler.asm.aarch64.AArch64Address;
  33 import org.graalvm.compiler.asm.aarch64.AArch64Assembler.ConditionFlag;
  34 import org.graalvm.compiler.asm.aarch64.AArch64MacroAssembler;
  35 import org.graalvm.compiler.core.common.LIRKind;
  36 import org.graalvm.compiler.lir.LIRInstructionClass;
  37 import org.graalvm.compiler.lir.Opcode;
  38 import org.graalvm.compiler.lir.asm.CompilationResultBuilder;
  39 import org.graalvm.compiler.lir.gen.LIRGeneratorTool;
  40 
  41 import jdk.vm.ci.code.Register;
  42 import jdk.vm.ci.meta.JavaKind;
  43 import jdk.vm.ci.meta.Value;
  44 
  45 /**
  46  * Emits code which compares two arrays lexicographically. If the CPU supports any vector
  47  * instructions specialized code is emitted to leverage these instructions.
  48  */
  49 @Opcode("ARRAY_COMPARE_TO")
  50 public final class AArch64ArrayCompareToOp extends AArch64LIRInstruction {
  51     public static final LIRInstructionClass<AArch64ArrayCompareToOp> TYPE = LIRInstructionClass.create(AArch64ArrayCompareToOp.class);
  52 
  53     private final JavaKind kind1;
  54     private final JavaKind kind2;
  55 
  56     private final int array1BaseOffset;
  57     private final int array2BaseOffset;
  58 
  59     @Def({REG}) protected Value resultValue;
  60 
  61     @Alive({REG}) protected Value array1Value;
  62     @Alive({REG}) protected Value array2Value;
  63     @Use({REG}) protected Value length1Value;
  64     @Use({REG}) protected Value length2Value;
  65     @Temp({REG}) protected Value length1ValueTemp;
  66     @Temp({REG}) protected Value length2ValueTemp;
  67 
  68     @Temp({REG}) protected Value temp1;
  69     @Temp({REG}) protected Value temp2;
  70     @Temp({REG}) protected Value temp3;
  71     @Temp({REG}) protected Value temp4;
  72     @Temp({REG}) protected Value temp5;
  73     @Temp({REG}) protected Value temp6;
  74 
  75     public AArch64ArrayCompareToOp(LIRGeneratorTool tool, JavaKind kind1, JavaKind kind2, Value result, Value array1, Value array2, Value length1, Value length2) {
  76         super(TYPE);
  77         this.kind1 = kind1;
  78         this.kind2 = kind2;
  79 
  80         // Both offsets should be the same but better be safe than sorry.
  81         this.array1BaseOffset = tool.getProviders().getMetaAccess().getArrayBaseOffset(kind1);
  82         this.array2BaseOffset = tool.getProviders().getMetaAccess().getArrayBaseOffset(kind2);
  83 
  84         this.resultValue = result;
  85 
  86         this.array1Value = array1;
  87         this.array2Value = array2;
  88 
  89         /*
  90          * The length values are inputs but are also killed like temporaries so need both Use and
  91          * Temp annotations, which will only work with fixed registers.
  92          */
  93 
  94         this.length1Value = length1;
  95         this.length2Value = length2;
  96         this.length1ValueTemp = length1;
  97         this.length2ValueTemp = length2;
  98 
  99         // Allocate some temporaries.
 100         this.temp1 = tool.newVariable(LIRKind.unknownReference(tool.target().arch.getWordKind()));
 101         this.temp2 = tool.newVariable(LIRKind.unknownReference(tool.target().arch.getWordKind()));
 102         this.temp3 = tool.newVariable(LIRKind.unknownReference(tool.target().arch.getWordKind()));
 103         this.temp4 = tool.newVariable(LIRKind.unknownReference(tool.target().arch.getWordKind()));
 104         this.temp5 = tool.newVariable(LIRKind.unknownReference(tool.target().arch.getWordKind()));
 105         this.temp6 = tool.newVariable(LIRKind.unknownReference(tool.target().arch.getWordKind()));
 106     }
 107 
 108     @Override
 109     protected void emitCode(CompilationResultBuilder crb, AArch64MacroAssembler masm) {
 110 
 111         Register result = asRegister(resultValue);
 112         Register length1 = asRegister(length1Value);
 113         Register length2 = asRegister(length2Value);
 114 
 115         Register array1 = asRegister(temp1);
 116         Register array2 = asRegister(temp2);
 117         Register length = asRegister(temp3);
 118         Register temp = asRegister(temp4);
 119         Register tailCount = asRegister(temp5);
 120         Register vecCount = asRegister(temp6);
 121 
 122         // Checkstyle: stop
 123         final Label BREAK_LABEL = new Label();
 124         final Label STRING_DIFFER_LABEL = new Label();
 125         final Label LENGTH_DIFFER_LABEL = new Label();
 126         final Label MAIN_LOOP_LABEL = new Label();
 127         final Label COMPARE_SHORT_LABEL = new Label();
 128         // Checkstyle: resume
 129 
 130         // Checkstyle: stop
 131         int CHAR_SIZE_BYTES = 1;
 132         int VECTOR_SIZE_BYTES = 8;
 133         int VECTOR_COUNT_BYTES = 8;
 134         // Checkstyle: resume
 135 
 136         // Byte is expanded to short if we compare strings with different encoding
 137         if (kind1 != kind2 || kind1 == JavaKind.Char) {
 138             CHAR_SIZE_BYTES = 2;
 139         }
 140 
 141         if (kind1 != kind2) {
 142             VECTOR_COUNT_BYTES = 4;
 143         }
 144 
 145         // Load array base addresses.
 146         masm.lea(array1, AArch64Address.createUnscaledImmediateAddress(asRegister(array1Value), array1BaseOffset));
 147         masm.lea(array2, AArch64Address.createUnscaledImmediateAddress(asRegister(array2Value), array2BaseOffset));
 148 
 149         // Calculate minimal length in chars for different kind case
 150         // Conditions could be squashed but lets keep it readable
 151         if (kind1 != kind2) {
 152             masm.lshr(64, length2, length2, 1);
 153         }
 154 
 155         if (kind1 == kind2 && kind1 == JavaKind.Char) {
 156             masm.lshr(64, length1, length1, 1);
 157             masm.lshr(64, length2, length2, 1);
 158         }
 159 
 160         masm.cmp(64, length1, length2);
 161         masm.cmov(64, length, length1, length2, ConditionFlag.LT);
 162 
 163         // One of strings is empty
 164         masm.cbz(64, length, LENGTH_DIFFER_LABEL);
 165 
 166         // Go back to bytes if necessary
 167         if (kind1 != kind2 || kind1 == JavaKind.Char) {
 168             masm.shl(64, length, length, 1);
 169         }
 170 
 171         masm.mov(64, vecCount, zr);
 172         masm.and(64, tailCount, length, VECTOR_SIZE_BYTES - 1); // tail count (in bytes)
 173         masm.ands(64, length, length, ~(VECTOR_SIZE_BYTES - 1));  // vector count (in bytes)
 174 
 175         // Length of string is less than VECTOR_SIZE, go to simple compare
 176         masm.branchConditionally(ConditionFlag.EQ, COMPARE_SHORT_LABEL);
 177 
 178         // MAIN_LOOP - read strings by 8 byte.
 179         masm.bind(MAIN_LOOP_LABEL);
 180         if (kind1 != kind2) {
 181             // Load 32 bits ad unpack it to entire 64bit register
 182             masm.ldr(32, result, AArch64Address.createRegisterOffsetAddress(array1, vecCount, false));
 183             masm.ubfm(64, temp, result, 0, 7);
 184             masm.lshr(64, result, result, 8);
 185             masm.bfm(64, temp, result, 48, 7);
 186             masm.lshr(64, result, result, 8);
 187             masm.bfm(64, temp, result, 32, 7);
 188             masm.lshr(64, result, result, 8);
 189             masm.bfm(64, temp, result, 16, 7);
 190             // Unpacked value placed in temp now
 191 
 192             masm.shl(64, result, vecCount, 1);
 193             masm.ldr(64, result, AArch64Address.createRegisterOffsetAddress(array2, result, false));
 194         } else {
 195             masm.ldr(64, temp, AArch64Address.createRegisterOffsetAddress(array1, vecCount, false));
 196             masm.ldr(64, result, AArch64Address.createRegisterOffsetAddress(array2, vecCount, false));
 197         }
 198         masm.eor(64, result, temp, result);
 199         masm.cbnz(64, result, STRING_DIFFER_LABEL);
 200         masm.add(64, vecCount, vecCount, VECTOR_COUNT_BYTES);
 201         masm.cmp(64, vecCount, length);
 202         masm.branchConditionally(ConditionFlag.LT, MAIN_LOOP_LABEL);
 203         // End of MAIN_LOOP
 204 
 205         // Strings are equal and no TAIL go to END
 206         masm.cbz(64, tailCount, LENGTH_DIFFER_LABEL);
 207 
 208         // Compaire tail of long string ...
 209         masm.lea(array1, AArch64Address.createRegisterOffsetAddress(array1, length, false));
 210         masm.lea(array2, AArch64Address.createRegisterOffsetAddress(array2, length, false));
 211 
 212         // ... or string less than vector length
 213         masm.bind(COMPARE_SHORT_LABEL);
 214         for (int i = 0; i < VECTOR_COUNT_BYTES; i += CHAR_SIZE_BYTES) {
 215             if (kind1 != kind2) {
 216                 masm.ldr(8, temp, AArch64Address.createUnscaledImmediateAddress(array1, i / 2));
 217             } else {
 218                 masm.ldr(8 * CHAR_SIZE_BYTES, temp, AArch64Address.createUnscaledImmediateAddress(array1, i));
 219             }
 220 
 221             masm.ldr(8 * CHAR_SIZE_BYTES, result, AArch64Address.createUnscaledImmediateAddress(array2, i));
 222 
 223             if (kind1 != kind2 && kind1 == JavaKind.Char) {
 224                 // Weird swap of substraction order
 225                 masm.subs(64, result, result, temp);
 226             } else {
 227                 masm.subs(64, result, temp, result);
 228             }
 229 
 230             masm.branchConditionally(ConditionFlag.NE, BREAK_LABEL);
 231             masm.subs(64, tailCount, tailCount, CHAR_SIZE_BYTES);
 232             masm.branchConditionally(ConditionFlag.EQ, LENGTH_DIFFER_LABEL);
 233         }
 234 
 235         // STRING_DIFFER extract exact value of a difference
 236         masm.bind(STRING_DIFFER_LABEL);
 237         masm.rbit(64, tailCount, result);
 238         masm.clz(64, vecCount, tailCount);
 239         masm.and(64, vecCount, vecCount, ~((8 * CHAR_SIZE_BYTES) - 1)); // Round to byte or short
 240 
 241         masm.eor(64, result, temp, result);
 242         masm.ashr(64, result, result, vecCount);
 243         masm.ashr(64, temp, temp, vecCount);
 244 
 245         masm.and(64, result, result, 0xFFFF >>> (16 - (8 * CHAR_SIZE_BYTES))); // 0xFF or 0xFFFF
 246         masm.and(64, temp, temp, 0xFFFF >>> (16 - (8 * CHAR_SIZE_BYTES)));
 247 
 248         masm.sub(64, result, temp, result);
 249         masm.branchConditionally(ConditionFlag.AL, BREAK_LABEL);
 250         // End of STRING_DIFFER
 251 
 252         // Strings are equials up to length,
 253         // return length difference in chars
 254         masm.bind(LENGTH_DIFFER_LABEL);
 255         if (kind1 != kind2 && kind1 == JavaKind.Char) {
 256             // Weird swap of substraction order
 257             masm.sub(64, result, length2, length1);
 258         } else {
 259             masm.sub(64, result, length1, length2);
 260         }
 261 
 262         // We are done
 263         masm.bind(BREAK_LABEL);
 264     }
 265 
 266 } // class