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.amd64;
  26 
  27 import static jdk.vm.ci.amd64.AMD64.k1;
  28 import static jdk.vm.ci.amd64.AMD64.k2;
  29 import static jdk.vm.ci.amd64.AMD64.rdi;
  30 import static jdk.vm.ci.amd64.AMD64.rdx;
  31 import static jdk.vm.ci.amd64.AMD64.rsi;
  32 
  33 import static jdk.vm.ci.code.ValueUtil.asRegister;
  34 import static org.graalvm.compiler.lir.LIRInstruction.OperandFlag.REG;
  35 
  36 import jdk.vm.ci.amd64.AMD64;
  37 import org.graalvm.compiler.asm.Label;
  38 import org.graalvm.compiler.asm.amd64.AMD64Address;
  39 import org.graalvm.compiler.asm.amd64.AMD64Assembler;
  40 import org.graalvm.compiler.asm.amd64.AMD64MacroAssembler;
  41 import org.graalvm.compiler.core.common.LIRKind;
  42 import org.graalvm.compiler.lir.LIRInstructionClass;
  43 import org.graalvm.compiler.lir.Opcode;
  44 import org.graalvm.compiler.lir.asm.CompilationResultBuilder;
  45 import org.graalvm.compiler.lir.gen.LIRGeneratorTool;
  46 
  47 import jdk.vm.ci.amd64.AMD64Kind;
  48 import jdk.vm.ci.code.Register;
  49 import jdk.vm.ci.meta.Value;
  50 
  51 @Opcode("AMD64_STRING_INFLATE")
  52 public final class AMD64StringLatin1InflateOp extends AMD64LIRInstruction {
  53     public static final LIRInstructionClass<AMD64StringLatin1InflateOp> TYPE = LIRInstructionClass.create(AMD64StringLatin1InflateOp.class);
  54 
  55     @Alive({REG}) private Value rsrc;
  56     @Alive({REG}) private Value rdst;
  57     @Alive({REG}) private Value rlen;
  58 
  59     @Temp({REG}) private Value vtmp1;
  60     @Temp({REG}) private Value rtmp2;
  61 
  62     public AMD64StringLatin1InflateOp(LIRGeneratorTool tool, Value src, Value dst, Value len) {
  63         super(TYPE);
  64 
  65         assert asRegister(src).equals(rsi);
  66         assert asRegister(dst).equals(rdi);
  67         assert asRegister(len).equals(rdx);
  68 
  69         rsrc = src;
  70         rdst = dst;
  71         rlen = len;
  72 
  73         vtmp1 = tool.newVariable(LIRKind.value(AMD64Kind.V512_BYTE));
  74         rtmp2 = tool.newVariable(LIRKind.value(AMD64Kind.DWORD));
  75     }
  76 
  77     @Override
  78     public void emitCode(CompilationResultBuilder crb, AMD64MacroAssembler masm) {
  79         Register src = asRegister(rsrc);
  80         Register dst = asRegister(rdst);
  81         Register len = asRegister(rlen);
  82 
  83         Register tmp1 = asRegister(vtmp1);
  84         Register tmp2 = asRegister(rtmp2);
  85 
  86         byteArrayInflate(masm, src, dst, len, tmp1, tmp2);
  87     }
  88 
  89     /**
  90      * Inflate a Latin1 string using a byte[] array representation into a UTF16 string using a
  91      * char[] array representation.
  92      *
  93      * @param masm the assembler
  94      * @param src (rsi) the start address of source byte[] to be inflated
  95      * @param dst (rdi) the start address of destination char[] array
  96      * @param len (rdx) the length
  97      * @param vtmp (xmm) temporary xmm register
  98      * @param tmp (gpr) temporary gpr register
  99      */
 100     private static void byteArrayInflate(AMD64MacroAssembler masm, Register src, Register dst, Register len, Register vtmp, Register tmp) {
 101         assert vtmp.getRegisterCategory().equals(AMD64.XMM);
 102 
 103         Label labelDone = new Label();
 104         Label labelBelowThreshold = new Label();
 105 
 106         assert src.number != dst.number && src.number != len.number && src.number != tmp.number;
 107         assert dst.number != len.number && dst.number != tmp.number;
 108         assert len.number != tmp.number;
 109 
 110         if (masm.supports(AMD64.CPUFeature.AVX512BW) &&
 111                         masm.supports(AMD64.CPUFeature.AVX512VL) &&
 112                         masm.supports(AMD64.CPUFeature.BMI2)) {
 113 
 114             // If the length of the string is less than 16, we chose not to use the
 115             // AVX512 instructions.
 116             masm.testl(len, -16);
 117             masm.jcc(AMD64Assembler.ConditionFlag.Zero, labelBelowThreshold);
 118 
 119             Label labelAvx512Tail = new Label();
 120             // Test for suitable number chunks with respect to the size of the vector
 121             // operation, mask off remaining number of chars (bytes) to inflate (such
 122             // that 'len' will always hold the number of bytes left to inflate) after
 123             // committing to the vector loop.
 124             // Adjust vector pointers to upper address bounds and inverse loop index.
 125             // This will keep the loop condition simple.
 126             //
 127             // NOTE: The above idiom/pattern is used in all the loops below.
 128 
 129             masm.movl(tmp, len);
 130             masm.andl(tmp, -32);     // The vector count (in chars).
 131             masm.jccb(AMD64Assembler.ConditionFlag.Zero, labelAvx512Tail);
 132             masm.andl(len, 32 - 1);  // The tail count (in chars).
 133 
 134             masm.leaq(src, new AMD64Address(src, tmp, AMD64Address.Scale.Times1));
 135             masm.leaq(dst, new AMD64Address(dst, tmp, AMD64Address.Scale.Times2));
 136             masm.negq(tmp);
 137 
 138             Label labelAvx512Loop = new Label();
 139             // Inflate 32 chars per iteration, reading 256-bit compact vectors
 140             // and writing 512-bit inflated ditto.
 141             masm.bind(labelAvx512Loop);
 142             masm.evpmovzxbw(vtmp, new AMD64Address(src, tmp, AMD64Address.Scale.Times1));
 143             masm.evmovdqu16(new AMD64Address(dst, tmp, AMD64Address.Scale.Times2), vtmp);
 144             masm.addq(tmp, 32);
 145             masm.jcc(AMD64Assembler.ConditionFlag.NotZero, labelAvx512Loop);
 146 
 147             masm.bind(labelAvx512Tail);
 148             // All done if the tail count is zero.
 149             masm.testl(len, len);
 150             masm.jcc(AMD64Assembler.ConditionFlag.Zero, labelDone);
 151 
 152             masm.kmovq(k2, k1);      // Save k1
 153 
 154             // Compute (1 << N) - 1 = ~(~0 << N), where N is the remaining number
 155             // of characters to process.
 156             masm.movl(tmp, -1);
 157             masm.shlxl(tmp, tmp, len);
 158             masm.notl(tmp);
 159 
 160             masm.kmovd(k1, tmp);
 161             masm.evpmovzxbw(vtmp, k1, new AMD64Address(src));
 162             masm.evmovdqu16(new AMD64Address(dst), k1, vtmp);
 163             masm.kmovq(k1, k2);      // Restore k1
 164             masm.jmp(labelDone);
 165         }
 166 
 167         if (masm.supports(AMD64.CPUFeature.SSE4_1)) {
 168 
 169             Label labelSSETail = new Label();
 170 
 171             if (masm.supports(AMD64.CPUFeature.AVX2)) {
 172 
 173                 Label labelAvx2Tail = new Label();
 174 
 175                 masm.movl(tmp, len);
 176                 masm.andl(tmp, -16);
 177                 masm.jccb(AMD64Assembler.ConditionFlag.Zero, labelAvx2Tail);
 178                 masm.andl(len, 16 - 1);
 179 
 180                 masm.leaq(src, new AMD64Address(src, tmp, AMD64Address.Scale.Times1));
 181                 masm.leaq(dst, new AMD64Address(dst, tmp, AMD64Address.Scale.Times2));
 182                 masm.negq(tmp);
 183 
 184                 Label labelAvx2Loop = new Label();
 185                 // Inflate 16 bytes (chars) per iteration, reading 128-bit compact vectors
 186                 // and writing 256-bit inflated ditto.
 187                 masm.bind(labelAvx2Loop);
 188                 masm.vpmovzxbw(vtmp, new AMD64Address(src, tmp, AMD64Address.Scale.Times1));
 189                 masm.vmovdqu(new AMD64Address(dst, tmp, AMD64Address.Scale.Times2), vtmp);
 190                 masm.addq(tmp, 16);
 191                 masm.jcc(AMD64Assembler.ConditionFlag.NotZero, labelAvx2Loop);
 192 
 193                 masm.bind(labelBelowThreshold);
 194                 masm.bind(labelAvx2Tail);
 195 
 196                 masm.movl(tmp, len);
 197                 masm.andl(tmp, -8);
 198                 masm.jccb(AMD64Assembler.ConditionFlag.Zero, labelSSETail);
 199                 masm.andl(len, 8 - 1);
 200 
 201                 // Inflate another 8 bytes before final tail copy.
 202                 masm.pmovzxbw(vtmp, new AMD64Address(src));
 203                 masm.movdqu(new AMD64Address(dst), vtmp);
 204                 masm.addq(src, 8);
 205                 masm.addq(dst, 16);
 206 
 207                 // Fall-through to labelSSETail.
 208             } else {
 209                 // When there is no AVX2 support available, we use AVX/SSE support to
 210                 // inflate into maximum 128-bits per operation.
 211 
 212                 masm.movl(tmp, len);
 213                 masm.andl(tmp, -8);
 214                 masm.jccb(AMD64Assembler.ConditionFlag.Zero, labelSSETail);
 215                 masm.andl(len, 8 - 1);
 216 
 217                 masm.leaq(src, new AMD64Address(src, tmp, AMD64Address.Scale.Times1));
 218                 masm.leaq(dst, new AMD64Address(dst, tmp, AMD64Address.Scale.Times2));
 219                 masm.negq(tmp);
 220 
 221                 Label labelSSECopy8Loop = new Label();
 222                 // Inflate 8 bytes (chars) per iteration, reading 64-bit compact vectors
 223                 // and writing 128-bit inflated ditto.
 224                 masm.bind(labelSSECopy8Loop);
 225                 masm.pmovzxbw(vtmp, new AMD64Address(src, tmp, AMD64Address.Scale.Times1));
 226                 masm.movdqu(new AMD64Address(dst, tmp, AMD64Address.Scale.Times2), vtmp);
 227                 masm.addq(tmp, 8);
 228                 masm.jcc(AMD64Assembler.ConditionFlag.NotZero, labelSSECopy8Loop);
 229 
 230                 // Fall-through to labelSSETail.
 231             }
 232 
 233             Label labelCopyChars = new Label();
 234 
 235             masm.bind(labelSSETail);
 236             masm.cmpl(len, 4);
 237             masm.jccb(AMD64Assembler.ConditionFlag.Less, labelCopyChars);
 238 
 239             masm.movdl(vtmp, new AMD64Address(src));
 240             masm.pmovzxbw(vtmp, vtmp);
 241             masm.movq(new AMD64Address(dst), vtmp);
 242             masm.subq(len, 4);
 243             masm.addq(src, 4);
 244             masm.addq(dst, 8);
 245 
 246             masm.bind(labelCopyChars);
 247         }
 248 
 249         // Inflate any remaining characters (bytes) using a vanilla implementation.
 250         masm.testl(len, len);
 251         masm.jccb(AMD64Assembler.ConditionFlag.Zero, labelDone);
 252         masm.leaq(src, new AMD64Address(src, len, AMD64Address.Scale.Times1));
 253         masm.leaq(dst, new AMD64Address(dst, len, AMD64Address.Scale.Times2));
 254         masm.negq(len);
 255 
 256         Label labelCopyCharsLoop = new Label();
 257         // Inflate a single byte (char) per iteration.
 258         masm.bind(labelCopyCharsLoop);
 259         masm.movzbl(tmp, new AMD64Address(src, len, AMD64Address.Scale.Times1));
 260         masm.movw(new AMD64Address(dst, len, AMD64Address.Scale.Times2), tmp);
 261         masm.incrementq(len, 1);
 262         masm.jcc(AMD64Assembler.ConditionFlag.NotZero, labelCopyCharsLoop);
 263 
 264         masm.bind(labelDone);
 265     }
 266 
 267 }