1 /* 2 * Copyright (c) 2013, 2019, 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.code.ValueUtil.asRegister; 28 import static org.graalvm.compiler.asm.amd64.AMD64Assembler.AMD64BinaryArithmetic.XOR; 29 import static org.graalvm.compiler.lir.LIRInstruction.OperandFlag.CONST; 30 import static org.graalvm.compiler.lir.LIRInstruction.OperandFlag.ILLEGAL; 31 import static org.graalvm.compiler.lir.LIRInstruction.OperandFlag.REG; 32 33 import java.util.Objects; 34 35 import org.graalvm.compiler.asm.Label; 36 import org.graalvm.compiler.asm.amd64.AMD64Address; 37 import org.graalvm.compiler.asm.amd64.AMD64Address.Scale; 38 import org.graalvm.compiler.asm.amd64.AMD64Assembler; 39 import org.graalvm.compiler.asm.amd64.AMD64Assembler.ConditionFlag; 40 import org.graalvm.compiler.asm.amd64.AMD64Assembler.SSEOp; 41 import org.graalvm.compiler.asm.amd64.AMD64BaseAssembler.OperandSize; 42 import org.graalvm.compiler.asm.amd64.AMD64MacroAssembler; 43 import org.graalvm.compiler.asm.amd64.AVXKind; 44 import org.graalvm.compiler.core.common.LIRKind; 45 import org.graalvm.compiler.debug.GraalError; 46 import org.graalvm.compiler.lir.LIRInstructionClass; 47 import org.graalvm.compiler.lir.LIRValueUtil; 48 import org.graalvm.compiler.lir.Opcode; 49 import org.graalvm.compiler.lir.asm.CompilationResultBuilder; 50 import org.graalvm.compiler.lir.gen.LIRGeneratorTool; 51 52 import jdk.vm.ci.amd64.AMD64; 53 import jdk.vm.ci.amd64.AMD64.CPUFeature; 54 import jdk.vm.ci.amd64.AMD64Kind; 55 import jdk.vm.ci.code.Register; 56 import jdk.vm.ci.code.TargetDescription; 57 import jdk.vm.ci.meta.JavaKind; 58 import jdk.vm.ci.meta.Value; 59 60 /** 61 * Emits code which compares two arrays of the same length. If the CPU supports any vector 62 * instructions specialized code is emitted to leverage these instructions. 63 * 64 * This op can also compare arrays of different integer types (e.g. {@code byte[]} and 65 * {@code char[]}) with on-the-fly sign- or zero-extension. If one of the given arrays is a 66 * {@code char[]} array, the smaller elements are zero-extended, otherwise they are sign-extended. 67 */ 68 @Opcode("ARRAY_EQUALS") 69 public final class AMD64ArrayEqualsOp extends AMD64LIRInstruction { 70 public static final LIRInstructionClass<AMD64ArrayEqualsOp> TYPE = LIRInstructionClass.create(AMD64ArrayEqualsOp.class); 71 72 private final JavaKind kind1; 73 private final JavaKind kind2; 74 private final int arrayBaseOffset1; 75 private final int arrayBaseOffset2; 76 private final Scale arrayIndexScale1; 77 private final Scale arrayIndexScale2; 78 private final AVXKind.AVXSize vectorSize; 79 private final boolean signExtend; 80 81 @Def({REG}) private Value resultValue; 82 @Alive({REG}) private Value array1Value; 83 @Alive({REG}) private Value array2Value; 84 @Alive({REG, CONST}) private Value lengthValue; 85 @Temp({REG, ILLEGAL}) private Value temp1; 86 @Temp({REG, ILLEGAL}) private Value temp2; 87 @Temp({REG}) private Value temp3; 88 @Temp({REG, ILLEGAL}) private Value temp4; 89 90 @Temp({REG, ILLEGAL}) private Value temp5; 91 @Temp({REG, ILLEGAL}) private Value tempXMM; 92 93 @Temp({REG, ILLEGAL}) private Value vectorTemp1; 94 @Temp({REG, ILLEGAL}) private Value vectorTemp2; 95 @Temp({REG, ILLEGAL}) private Value vectorTemp3; 96 @Temp({REG, ILLEGAL}) private Value vectorTemp4; 97 98 public AMD64ArrayEqualsOp(LIRGeneratorTool tool, JavaKind kind1, JavaKind kind2, Value result, Value array1, Value array2, Value length, 99 boolean directPointers, int maxVectorSize) { 100 super(TYPE); 101 this.kind1 = kind1; 102 this.kind2 = kind2; 103 this.signExtend = kind1 != JavaKind.Char && kind2 != JavaKind.Char; 104 105 assert kind1.isNumericInteger() && kind2.isNumericInteger() || kind1 == kind2; 106 107 this.arrayBaseOffset1 = directPointers ? 0 : tool.getProviders().getMetaAccess().getArrayBaseOffset(kind1); 108 this.arrayBaseOffset2 = directPointers ? 0 : tool.getProviders().getMetaAccess().getArrayBaseOffset(kind2); 109 this.arrayIndexScale1 = Objects.requireNonNull(Scale.fromInt(tool.getProviders().getMetaAccess().getArrayIndexScale(kind1))); 110 this.arrayIndexScale2 = Objects.requireNonNull(Scale.fromInt(tool.getProviders().getMetaAccess().getArrayIndexScale(kind2))); 111 this.vectorSize = ((AMD64) tool.target().arch).getFeatures().contains(CPUFeature.AVX2) && (maxVectorSize < 0 || maxVectorSize >= 32) ? AVXKind.AVXSize.YMM : AVXKind.AVXSize.XMM; 112 113 this.resultValue = result; 114 this.array1Value = array1; 115 this.array2Value = array2; 116 this.lengthValue = length; 117 118 // Allocate some temporaries. 119 if (supportsSSE41(tool.target()) && canGenerateConstantLengthCompare(tool.target()) && !constantLengthCompareNeedsTmpArrayPointers()) { 120 this.temp1 = Value.ILLEGAL; 121 this.temp2 = Value.ILLEGAL; 122 } else { 123 this.temp1 = tool.newVariable(LIRKind.unknownReference(tool.target().arch.getWordKind())); 124 this.temp2 = tool.newVariable(LIRKind.unknownReference(tool.target().arch.getWordKind())); 125 } 126 this.temp3 = tool.newVariable(LIRKind.value(tool.target().arch.getWordKind())); 127 if (supportsSSE41(tool.target()) && canGenerateConstantLengthCompare(tool.target())) { 128 this.temp4 = Value.ILLEGAL; 129 this.temp5 = Value.ILLEGAL; 130 } else { 131 this.temp4 = tool.newVariable(LIRKind.value(tool.target().arch.getWordKind())); 132 this.temp5 = kind1.isNumericFloat() || kind1 != kind2 ? tool.newVariable(LIRKind.value(tool.target().arch.getWordKind())) : Value.ILLEGAL; 133 } 134 135 if (kind1 == JavaKind.Float) { 136 this.tempXMM = tool.newVariable(LIRKind.value(AMD64Kind.SINGLE)); 137 } else if (kind1 == JavaKind.Double) { 138 this.tempXMM = tool.newVariable(LIRKind.value(AMD64Kind.DOUBLE)); 139 } else { 140 this.tempXMM = Value.ILLEGAL; 141 } 142 143 // We only need the vector temporaries if we generate SSE code. 144 if (supportsSSE41(tool.target())) { 145 if (canGenerateConstantLengthCompare(tool.target())) { 146 LIRKind lirKind = LIRKind.value(vectorSize == AVXKind.AVXSize.YMM ? AMD64Kind.V256_BYTE : AMD64Kind.V128_BYTE); 147 this.vectorTemp1 = tool.newVariable(lirKind); 148 this.vectorTemp2 = tool.newVariable(lirKind); 149 this.vectorTemp3 = tool.newVariable(lirKind); 150 this.vectorTemp4 = tool.newVariable(lirKind); 151 } else { 152 this.vectorTemp1 = tool.newVariable(LIRKind.value(AMD64Kind.DOUBLE)); 153 this.vectorTemp2 = tool.newVariable(LIRKind.value(AMD64Kind.DOUBLE)); 154 this.vectorTemp3 = Value.ILLEGAL; 155 this.vectorTemp4 = Value.ILLEGAL; 156 } 157 } else { 158 this.vectorTemp1 = Value.ILLEGAL; 159 this.vectorTemp2 = Value.ILLEGAL; 160 this.vectorTemp3 = Value.ILLEGAL; 161 this.vectorTemp4 = Value.ILLEGAL; 162 } 163 } 164 165 private boolean canGenerateConstantLengthCompare(TargetDescription target) { 166 return LIRValueUtil.isJavaConstant(lengthValue) && kind1.isNumericInteger() && (kind1 == kind2 || getElementsPerVector(AVXKind.AVXSize.XMM) <= constantLength()) && supportsSSE41(target); 167 } 168 169 private int constantLength() { 170 return LIRValueUtil.asJavaConstant(lengthValue).asInt(); 171 } 172 173 @Override 174 public void emitCode(CompilationResultBuilder crb, AMD64MacroAssembler masm) { 175 Register result = asRegister(resultValue); 176 177 Label trueLabel = new Label(); 178 Label falseLabel = new Label(); 179 Label done = new Label(); 180 181 if (canGenerateConstantLengthCompare(crb.target)) { 182 emitConstantLengthArrayCompareBytes(crb, masm, new Register[]{asRegister(vectorTemp1), asRegister(vectorTemp2), asRegister(vectorTemp3), asRegister(vectorTemp4)}, falseLabel); 183 } else { 184 Register array1 = asRegister(temp1); 185 Register array2 = asRegister(temp2); 186 // Load array base addresses. 187 masm.leaq(array1, new AMD64Address(asRegister(array1Value), arrayBaseOffset1)); 188 masm.leaq(array2, new AMD64Address(asRegister(array2Value), arrayBaseOffset2)); 189 Register length = asRegister(temp3); 190 // Get array length. 191 if (LIRValueUtil.isJavaConstant(lengthValue)) { 192 masm.movl(length, constantLength()); 193 } else { 194 masm.movl(length, asRegister(lengthValue)); 195 } 196 // copy 197 masm.movl(result, length); 198 emitArrayCompare(crb, masm, result, array1, array2, length, trueLabel, falseLabel); 199 } 200 201 // Return true 202 masm.bind(trueLabel); 203 masm.movl(result, 1); 204 masm.jmpb(done); 205 206 // Return false 207 masm.bind(falseLabel); 208 masm.xorl(result, result); 209 210 // That's it 211 masm.bind(done); 212 } 213 214 private void emitArrayCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, 215 Register result, Register array1, Register array2, Register length, 216 Label trueLabel, Label falseLabel) { 217 if (supportsSSE41(crb.target)) { 218 emitVectorCompare(crb, masm, result, array1, array2, length, trueLabel, falseLabel); 219 } 220 if (kind1 == kind2) { 221 emit8ByteCompare(crb, masm, result, array1, array2, length, trueLabel, falseLabel); 222 emitTailCompares(masm, result, array1, array2, length, trueLabel, falseLabel); 223 } else { 224 emitDifferentKindsElementWiseCompare(crb, masm, result, array1, array2, length, trueLabel, falseLabel); 225 } 226 } 227 228 /** 229 * Returns if the underlying AMD64 architecture supports SSE 4.1 instructions. 230 * 231 * @param target target description of the underlying architecture 232 * @return true if the underlying architecture supports SSE 4.1 233 */ 234 private static boolean supportsSSE41(TargetDescription target) { 235 AMD64 arch = (AMD64) target.arch; 236 return arch.getFeatures().contains(CPUFeature.SSE4_1); 237 } 238 239 /** 240 * Emits code that uses SSE4.1/AVX1 128-bit (16-byte) or AVX2 256-bit (32-byte) vector compares. 241 */ 242 private void emitVectorCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, 243 Register result, Register array1, Register array2, Register length, 244 Label trueLabel, Label falseLabel) { 245 assert supportsSSE41(crb.target); 246 247 Register vector1 = asRegister(vectorTemp1); 248 Register vector2 = asRegister(vectorTemp2); 249 250 int elementsPerVector = getElementsPerVector(vectorSize); 251 252 Label loop = new Label(); 253 Label compareTail = new Label(); 254 255 boolean requiresNaNCheck = kind1.isNumericFloat(); 256 Label loopCheck = new Label(); 257 Label nanCheck = new Label(); 258 259 // Compare 16-byte vectors 260 masm.andl(result, elementsPerVector - 1); // tail count 261 masm.andl(length, ~(elementsPerVector - 1)); // vector count 262 masm.jcc(ConditionFlag.Zero, compareTail); 263 264 masm.leaq(array1, new AMD64Address(array1, length, arrayIndexScale1, 0)); 265 masm.leaq(array2, new AMD64Address(array2, length, arrayIndexScale2, 0)); 266 masm.negq(length); 267 268 // Align the main loop 269 masm.align(crb.target.wordSize * 2); 270 masm.bind(loop); 271 emitVectorLoad1(masm, vector1, array1, length, 0, vectorSize); 272 emitVectorLoad2(masm, vector2, array2, length, 0, vectorSize); 273 emitVectorCmp(masm, vector1, vector2, vectorSize); 274 masm.jcc(ConditionFlag.NotZero, requiresNaNCheck ? nanCheck : falseLabel); 275 276 masm.bind(loopCheck); 277 masm.addq(length, elementsPerVector); 278 masm.jcc(ConditionFlag.NotZero, loop); 279 280 masm.testl(result, result); 281 masm.jcc(ConditionFlag.Zero, trueLabel); 282 283 if (requiresNaNCheck) { 284 Label unalignedCheck = new Label(); 285 masm.jmpb(unalignedCheck); 286 masm.bind(nanCheck); 287 emitFloatCompareWithinRange(crb, masm, array1, array2, length, 0, falseLabel, elementsPerVector); 288 masm.jmpb(loopCheck); 289 masm.bind(unalignedCheck); 290 } 291 292 /* 293 * Compare the remaining bytes with an unaligned memory load aligned to the end of the 294 * array. 295 */ 296 emitVectorLoad1(masm, vector1, array1, result, scaleDisplacement1(-vectorSize.getBytes()), vectorSize); 297 emitVectorLoad2(masm, vector2, array2, result, scaleDisplacement2(-vectorSize.getBytes()), vectorSize); 298 emitVectorCmp(masm, vector1, vector2, vectorSize); 299 if (requiresNaNCheck) { 300 masm.jcc(ConditionFlag.Zero, trueLabel); 301 emitFloatCompareWithinRange(crb, masm, array1, array2, result, -vectorSize.getBytes(), falseLabel, elementsPerVector); 302 } else { 303 masm.jcc(ConditionFlag.NotZero, falseLabel); 304 } 305 masm.jmp(trueLabel); 306 307 masm.bind(compareTail); 308 masm.movl(length, result); 309 } 310 311 private int getElementsPerVector(AVXKind.AVXSize vSize) { 312 return vSize.getBytes() >> Math.max(arrayIndexScale1.log2, arrayIndexScale2.log2); 313 } 314 315 private void emitVectorLoad1(AMD64MacroAssembler asm, Register dst, Register src, int displacement, AVXKind.AVXSize size) { 316 emitVectorLoad1(asm, dst, src, Register.None, displacement, size); 317 } 318 319 private void emitVectorLoad2(AMD64MacroAssembler asm, Register dst, Register src, int displacement, AVXKind.AVXSize size) { 320 emitVectorLoad2(asm, dst, src, Register.None, displacement, size); 321 } 322 323 private void emitVectorLoad1(AMD64MacroAssembler asm, Register dst, Register src, Register index, int displacement, AVXKind.AVXSize size) { 324 emitVectorLoad(asm, dst, src, index, displacement, arrayIndexScale1, arrayIndexScale2, size); 325 } 326 327 private void emitVectorLoad2(AMD64MacroAssembler asm, Register dst, Register src, Register index, int displacement, AVXKind.AVXSize size) { 328 emitVectorLoad(asm, dst, src, index, displacement, arrayIndexScale2, arrayIndexScale1, size); 329 } 330 331 private void emitVectorLoad(AMD64MacroAssembler asm, Register dst, Register src, Register index, int displacement, Scale ownScale, Scale otherScale, AVXKind.AVXSize size) { 332 AMD64Address address = new AMD64Address(src, index, ownScale, displacement); 333 if (ownScale.value < otherScale.value) { 334 if (size == AVXKind.AVXSize.YMM) { 335 getAVX2LoadAndExtendOp(ownScale, otherScale, signExtend).emit(asm, size, dst, address); 336 } else { 337 loadAndExtendSSE(asm, dst, address, ownScale, otherScale, signExtend); 338 } 339 } else { 340 if (size == AVXKind.AVXSize.YMM) { 341 asm.vmovdqu(dst, address); 342 } else { 343 asm.movdqu(dst, address); 344 } 345 } 346 } 347 348 private int scaleDisplacement1(int displacement) { 349 return scaleDisplacement(displacement, arrayIndexScale1, arrayIndexScale2); 350 } 351 352 private int scaleDisplacement2(int displacement) { 353 return scaleDisplacement(displacement, arrayIndexScale2, arrayIndexScale1); 354 } 355 356 private static int scaleDisplacement(int displacement, Scale ownScale, Scale otherScale) { 357 if (ownScale.value < otherScale.value) { 358 return displacement >> (otherScale.log2 - ownScale.log2); 359 } 360 return displacement; 361 } 362 363 private static AMD64Assembler.VexRMOp getAVX2LoadAndExtendOp(Scale ownScale, Scale otherScale, boolean signExtend) { 364 switch (ownScale) { 365 case Times1: 366 switch (otherScale) { 367 case Times2: 368 return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXBW : AMD64Assembler.VexRMOp.VPMOVZXBW; 369 case Times4: 370 return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXBD : AMD64Assembler.VexRMOp.VPMOVZXBD; 371 case Times8: 372 return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXBQ : AMD64Assembler.VexRMOp.VPMOVZXBQ; 373 } 374 throw GraalError.shouldNotReachHere(); 375 case Times2: 376 switch (otherScale) { 377 case Times4: 378 return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXWD : AMD64Assembler.VexRMOp.VPMOVZXWD; 379 case Times8: 380 return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXWQ : AMD64Assembler.VexRMOp.VPMOVZXWQ; 381 } 382 throw GraalError.shouldNotReachHere(); 383 case Times4: 384 return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXDQ : AMD64Assembler.VexRMOp.VPMOVZXDQ; 385 } 386 throw GraalError.shouldNotReachHere(); 387 } 388 389 private static void loadAndExtendSSE(AMD64MacroAssembler asm, Register dst, AMD64Address src, Scale ownScale, Scale otherScale, boolean signExtend) { 390 switch (ownScale) { 391 case Times1: 392 switch (otherScale) { 393 case Times2: 394 if (signExtend) { 395 asm.pmovsxbw(dst, src); 396 } else { 397 asm.pmovzxbw(dst, src); 398 } 399 return; 400 case Times4: 401 if (signExtend) { 402 asm.pmovsxbd(dst, src); 403 } else { 404 asm.pmovzxbd(dst, src); 405 } 406 return; 407 case Times8: 408 if (signExtend) { 409 asm.pmovsxbq(dst, src); 410 } else { 411 asm.pmovzxbq(dst, src); 412 } 413 return; 414 } 415 throw GraalError.shouldNotReachHere(); 416 case Times2: 417 switch (otherScale) { 418 case Times4: 419 if (signExtend) { 420 asm.pmovsxwd(dst, src); 421 } else { 422 asm.pmovzxwd(dst, src); 423 } 424 return; 425 case Times8: 426 if (signExtend) { 427 asm.pmovsxwq(dst, src); 428 } else { 429 asm.pmovzxwq(dst, src); 430 } 431 return; 432 } 433 throw GraalError.shouldNotReachHere(); 434 case Times4: 435 if (signExtend) { 436 asm.pmovsxdq(dst, src); 437 } else { 438 asm.pmovzxdq(dst, src); 439 } 440 return; 441 } 442 throw GraalError.shouldNotReachHere(); 443 } 444 445 private static void emitVectorCmp(AMD64MacroAssembler masm, Register vector1, Register vector2, AVXKind.AVXSize size) { 446 emitVectorXor(masm, vector1, vector2, size); 447 emitVectorTest(masm, vector1, size); 448 } 449 450 private static void emitVectorXor(AMD64MacroAssembler masm, Register vector1, Register vector2, AVXKind.AVXSize size) { 451 if (size == AVXKind.AVXSize.YMM) { 452 masm.vpxor(vector1, vector1, vector2); 453 } else { 454 masm.pxor(vector1, vector2); 455 } 456 } 457 458 private static void emitVectorTest(AMD64MacroAssembler masm, Register vector1, AVXKind.AVXSize size) { 459 if (size == AVXKind.AVXSize.YMM) { 460 masm.vptest(vector1, vector1); 461 } else { 462 masm.ptest(vector1, vector1); 463 } 464 } 465 466 /** 467 * Vector size used in {@link #emit8ByteCompare}. 468 */ 469 private static final int VECTOR_SIZE = 8; 470 471 /** 472 * Emits code that uses 8-byte vector compares. 473 */ 474 private void emit8ByteCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, 475 Register result, Register array1, Register array2, Register length, Label trueLabel, Label falseLabel) { 476 assert kind1 == kind2; 477 Label loop = new Label(); 478 Label compareTail = new Label(); 479 480 int elementsPerVector = 8 >> arrayIndexScale1.log2; 481 482 boolean requiresNaNCheck = kind1.isNumericFloat(); 483 Label loopCheck = new Label(); 484 Label nanCheck = new Label(); 485 486 Register temp = asRegister(temp4); 487 488 masm.andl(result, elementsPerVector - 1); // tail count 489 masm.andl(length, ~(elementsPerVector - 1)); // vector count 490 masm.jcc(ConditionFlag.Zero, compareTail); 491 492 masm.leaq(array1, new AMD64Address(array1, length, arrayIndexScale1, 0)); 493 masm.leaq(array2, new AMD64Address(array2, length, arrayIndexScale2, 0)); 494 masm.negq(length); 495 496 // Align the main loop 497 masm.align(crb.target.wordSize * 2); 498 masm.bind(loop); 499 masm.movq(temp, new AMD64Address(array1, length, arrayIndexScale1, 0)); 500 masm.cmpq(temp, new AMD64Address(array2, length, arrayIndexScale2, 0)); 501 masm.jcc(ConditionFlag.NotEqual, requiresNaNCheck ? nanCheck : falseLabel); 502 503 masm.bind(loopCheck); 504 masm.addq(length, elementsPerVector); 505 masm.jccb(ConditionFlag.NotZero, loop); 506 507 masm.testl(result, result); 508 masm.jcc(ConditionFlag.Zero, trueLabel); 509 510 if (requiresNaNCheck) { 511 // NaN check is slow path and hence placed outside of the main loop. 512 Label unalignedCheck = new Label(); 513 masm.jmpb(unalignedCheck); 514 masm.bind(nanCheck); 515 // At most two iterations, unroll in the emitted code. 516 for (int offset = 0; offset < VECTOR_SIZE; offset += kind1.getByteCount()) { 517 emitFloatCompare(masm, array1, array2, length, offset, falseLabel, kind1.getByteCount() == VECTOR_SIZE); 518 } 519 masm.jmpb(loopCheck); 520 masm.bind(unalignedCheck); 521 } 522 523 /* 524 * Compare the remaining bytes with an unaligned memory load aligned to the end of the 525 * array. 526 */ 527 masm.movq(temp, new AMD64Address(array1, result, arrayIndexScale1, -VECTOR_SIZE)); 528 masm.cmpq(temp, new AMD64Address(array2, result, arrayIndexScale2, -VECTOR_SIZE)); 529 if (requiresNaNCheck) { 530 masm.jcc(ConditionFlag.Equal, trueLabel); 531 // At most two iterations, unroll in the emitted code. 532 for (int offset = 0; offset < VECTOR_SIZE; offset += kind1.getByteCount()) { 533 emitFloatCompare(masm, array1, array2, result, -VECTOR_SIZE + offset, falseLabel, kind1.getByteCount() == VECTOR_SIZE); 534 } 535 } else { 536 masm.jccb(ConditionFlag.NotEqual, falseLabel); 537 } 538 masm.jmpb(trueLabel); 539 540 masm.bind(compareTail); 541 masm.movl(length, result); 542 } 543 544 /** 545 * Emits code to compare the remaining 1 to 4 bytes. 546 */ 547 private void emitTailCompares(AMD64MacroAssembler masm, 548 Register result, Register array1, Register array2, Register length, Label trueLabel, Label falseLabel) { 549 assert kind1 == kind2; 550 Label compare2Bytes = new Label(); 551 Label compare1Byte = new Label(); 552 553 Register temp = asRegister(temp4); 554 555 if (kind1.getByteCount() <= 4) { 556 // Compare trailing 4 bytes, if any. 557 masm.testl(result, arrayIndexScale1.log2 == 0 ? 4 : 4 >> arrayIndexScale1.log2); 558 masm.jccb(ConditionFlag.Zero, compare2Bytes); 559 masm.movl(temp, new AMD64Address(array1, 0)); 560 masm.cmpl(temp, new AMD64Address(array2, 0)); 561 if (kind1 == JavaKind.Float) { 562 masm.jccb(ConditionFlag.Equal, trueLabel); 563 emitFloatCompare(masm, array1, array2, Register.None, 0, falseLabel, true); 564 masm.jmpb(trueLabel); 565 } else { 566 masm.jccb(ConditionFlag.NotEqual, falseLabel); 567 } 568 if (kind1.getByteCount() <= 2) { 569 // Move array pointers forward. 570 masm.leaq(array1, new AMD64Address(array1, 4)); 571 masm.leaq(array2, new AMD64Address(array2, 4)); 572 573 // Compare trailing 2 bytes, if any. 574 masm.bind(compare2Bytes); 575 masm.testl(result, arrayIndexScale1.log2 == 0 ? 2 : 2 >> arrayIndexScale1.log2); 576 masm.jccb(ConditionFlag.Zero, compare1Byte); 577 masm.movzwl(temp, new AMD64Address(array1, 0)); 578 masm.movzwl(length, new AMD64Address(array2, 0)); 579 masm.cmpl(temp, length); 580 masm.jccb(ConditionFlag.NotEqual, falseLabel); 581 582 // The one-byte tail compare is only required for boolean and byte arrays. 583 if (kind1.getByteCount() <= 1) { 584 // Move array pointers forward before we compare the last trailing byte. 585 masm.leaq(array1, new AMD64Address(array1, 2)); 586 masm.leaq(array2, new AMD64Address(array2, 2)); 587 588 // Compare trailing byte, if any. 589 masm.bind(compare1Byte); 590 masm.testl(result, 1); 591 masm.jccb(ConditionFlag.Zero, trueLabel); 592 masm.movzbl(temp, new AMD64Address(array1, 0)); 593 masm.movzbl(length, new AMD64Address(array2, 0)); 594 masm.cmpl(temp, length); 595 masm.jccb(ConditionFlag.NotEqual, falseLabel); 596 } else { 597 masm.bind(compare1Byte); 598 } 599 } else { 600 masm.bind(compare2Bytes); 601 } 602 } 603 } 604 605 private void emitDifferentKindsElementWiseCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, 606 Register result, Register array1, Register array2, Register length, Label trueLabel, Label falseLabel) { 607 assert kind1 != kind2; 608 assert kind1.isNumericInteger() && kind2.isNumericInteger(); 609 Label loop = new Label(); 610 Label compareTail = new Label(); 611 612 int elementsPerLoopIteration = 4; 613 614 Register tmp1 = asRegister(temp4); 615 Register tmp2 = asRegister(temp5); 616 617 masm.andl(result, elementsPerLoopIteration - 1); // tail count 618 masm.andl(length, ~(elementsPerLoopIteration - 1)); // bulk loop count 619 masm.jcc(ConditionFlag.Zero, compareTail); 620 621 masm.leaq(array1, new AMD64Address(array1, length, arrayIndexScale1, 0)); 622 masm.leaq(array2, new AMD64Address(array2, length, arrayIndexScale2, 0)); 623 masm.negq(length); 624 625 // clear comparison registers because of the missing movzlq instruction 626 masm.xorq(tmp1, tmp1); 627 masm.xorq(tmp2, tmp2); 628 629 // Align the main loop 630 masm.align(crb.target.wordSize * 2); 631 masm.bind(loop); 632 for (int i = 0; i < elementsPerLoopIteration; i++) { 633 emitMovBytes(masm, tmp1, new AMD64Address(array1, length, arrayIndexScale1, i << arrayIndexScale1.log2), kind1.getByteCount()); 634 emitMovBytes(masm, tmp2, new AMD64Address(array2, length, arrayIndexScale2, i << arrayIndexScale2.log2), kind2.getByteCount()); 635 masm.cmpq(tmp1, tmp2); 636 masm.jcc(ConditionFlag.NotEqual, falseLabel); 637 } 638 masm.addq(length, elementsPerLoopIteration); 639 masm.jccb(ConditionFlag.NotZero, loop); 640 641 masm.bind(compareTail); 642 masm.testl(result, result); 643 masm.jcc(ConditionFlag.Zero, trueLabel); 644 for (int i = 0; i < elementsPerLoopIteration - 1; i++) { 645 emitMovBytes(masm, tmp1, new AMD64Address(array1, length, arrayIndexScale1, 0), kind1.getByteCount()); 646 emitMovBytes(masm, tmp2, new AMD64Address(array2, length, arrayIndexScale2, 0), kind2.getByteCount()); 647 masm.cmpq(tmp1, tmp2); 648 masm.jcc(ConditionFlag.NotEqual, falseLabel); 649 if (i < elementsPerLoopIteration - 2) { 650 masm.incrementq(length, 1); 651 masm.decrementq(result, 1); 652 masm.jcc(ConditionFlag.Zero, trueLabel); 653 } else { 654 masm.jmpb(trueLabel); 655 } 656 } 657 } 658 659 /** 660 * Emits code to fall through if {@code src} is NaN, otherwise jump to {@code branchOrdered}. 661 */ 662 private void emitNaNCheck(AMD64MacroAssembler masm, AMD64Address src, Label branchIfNonNaN) { 663 assert kind1.isNumericFloat(); 664 Register tempXMMReg = asRegister(tempXMM); 665 if (kind1 == JavaKind.Float) { 666 masm.movflt(tempXMMReg, src); 667 } else { 668 masm.movdbl(tempXMMReg, src); 669 } 670 SSEOp.UCOMIS.emit(masm, kind1 == JavaKind.Float ? OperandSize.PS : OperandSize.PD, tempXMMReg, tempXMMReg); 671 masm.jcc(ConditionFlag.NoParity, branchIfNonNaN); 672 } 673 674 /** 675 * Emits code to compare if two floats are bitwise equal or both NaN. 676 */ 677 private void emitFloatCompare(AMD64MacroAssembler masm, Register base1, Register base2, Register index, int offset, Label falseLabel, 678 boolean skipBitwiseCompare) { 679 AMD64Address address1 = new AMD64Address(base1, index, arrayIndexScale1, offset); 680 AMD64Address address2 = new AMD64Address(base2, index, arrayIndexScale2, offset); 681 682 Label bitwiseEqual = new Label(); 683 684 if (!skipBitwiseCompare) { 685 // Bitwise compare 686 Register temp = asRegister(temp4); 687 688 if (kind1 == JavaKind.Float) { 689 masm.movl(temp, address1); 690 masm.cmpl(temp, address2); 691 } else { 692 masm.movq(temp, address1); 693 masm.cmpq(temp, address2); 694 } 695 masm.jccb(ConditionFlag.Equal, bitwiseEqual); 696 } 697 698 emitNaNCheck(masm, address1, falseLabel); 699 emitNaNCheck(masm, address2, falseLabel); 700 701 masm.bind(bitwiseEqual); 702 } 703 704 /** 705 * Emits code to compare float equality within a range. 706 */ 707 private void emitFloatCompareWithinRange(CompilationResultBuilder crb, AMD64MacroAssembler masm, 708 Register base1, Register base2, Register index, int offset, Label falseLabel, int range) { 709 assert kind1.isNumericFloat(); 710 Label loop = new Label(); 711 Register i = asRegister(temp5); 712 713 masm.movq(i, range); 714 masm.negq(i); 715 // Align the main loop 716 masm.align(crb.target.wordSize * 2); 717 masm.bind(loop); 718 emitFloatCompare(masm, base1, base2, index, offset, falseLabel, range == 1); 719 masm.incrementq(index, 1); 720 masm.incrementq(i, 1); 721 masm.jccb(ConditionFlag.NotZero, loop); 722 // Floats within the range are equal, revert change to the register index 723 masm.subq(index, range); 724 } 725 726 private boolean constantLengthCompareNeedsTmpArrayPointers() { 727 AVXKind.AVXSize vSize = vectorSize; 728 if (constantLength() < getElementsPerVector(vectorSize)) { 729 vSize = AVXKind.AVXSize.XMM; 730 } 731 int vectorCount = constantLength() & ~(2 * getElementsPerVector(vSize) - 1); 732 return vectorCount > 0; 733 } 734 735 /** 736 * Emits specialized assembly for checking equality of memory regions 737 * {@code arrayPtr1[0..nBytes]} and {@code arrayPtr2[0..nBytes]}. If they match, execution 738 * continues directly after the emitted code block, otherwise we jump to {@code noMatch}. 739 */ 740 private void emitConstantLengthArrayCompareBytes( 741 CompilationResultBuilder crb, 742 AMD64MacroAssembler asm, 743 Register[] tmpVectors, 744 Label noMatch) { 745 if (constantLength() == 0) { 746 // do nothing 747 return; 748 } 749 Register arrayPtr1 = asRegister(array1Value); 750 Register arrayPtr2 = asRegister(array2Value); 751 Register tmp = asRegister(temp3); 752 AVXKind.AVXSize vSize = vectorSize; 753 if (constantLength() < getElementsPerVector(vectorSize)) { 754 vSize = AVXKind.AVXSize.XMM; 755 } 756 int elementsPerVector = getElementsPerVector(vSize); 757 if (elementsPerVector > constantLength()) { 758 assert kind1 == kind2; 759 int byteLength = constantLength() << arrayIndexScale1.log2; 760 // array is shorter than any vector register, use regular XOR instructions 761 int movSize = (byteLength < 2) ? 1 : ((byteLength < 4) ? 2 : ((byteLength < 8) ? 4 : 8)); 762 emitMovBytes(asm, tmp, new AMD64Address(arrayPtr1, arrayBaseOffset1), movSize); 763 emitXorBytes(asm, tmp, new AMD64Address(arrayPtr2, arrayBaseOffset2), movSize); 764 asm.jccb(AMD64Assembler.ConditionFlag.NotZero, noMatch); 765 if (byteLength > movSize) { 766 emitMovBytes(asm, tmp, new AMD64Address(arrayPtr1, arrayBaseOffset1 + byteLength - movSize), movSize); 767 emitXorBytes(asm, tmp, new AMD64Address(arrayPtr2, arrayBaseOffset2 + byteLength - movSize), movSize); 768 asm.jccb(AMD64Assembler.ConditionFlag.NotZero, noMatch); 769 } 770 } else { 771 int elementsPerVectorLoop = 2 * elementsPerVector; 772 int tailCount = constantLength() & (elementsPerVectorLoop - 1); 773 int vectorCount = constantLength() & ~(elementsPerVectorLoop - 1); 774 int bytesPerVector = vSize.getBytes(); 775 if (vectorCount > 0) { 776 Label loopBegin = new Label(); 777 Register tmpArrayPtr1 = asRegister(temp1); 778 Register tmpArrayPtr2 = asRegister(temp2); 779 asm.leaq(tmpArrayPtr1, new AMD64Address(arrayPtr1, vectorCount << arrayIndexScale1.log2)); 780 asm.leaq(tmpArrayPtr2, new AMD64Address(arrayPtr2, vectorCount << arrayIndexScale2.log2)); 781 arrayPtr1 = tmpArrayPtr1; 782 arrayPtr2 = tmpArrayPtr2; 783 asm.movq(tmp, -vectorCount); 784 asm.align(crb.target.wordSize * 2); 785 asm.bind(loopBegin); 786 emitVectorLoad1(asm, tmpVectors[0], arrayPtr1, tmp, arrayBaseOffset1, vSize); 787 emitVectorLoad2(asm, tmpVectors[1], arrayPtr2, tmp, arrayBaseOffset2, vSize); 788 emitVectorLoad1(asm, tmpVectors[2], arrayPtr1, tmp, arrayBaseOffset1 + scaleDisplacement1(bytesPerVector), vSize); 789 emitVectorLoad2(asm, tmpVectors[3], arrayPtr2, tmp, arrayBaseOffset2 + scaleDisplacement2(bytesPerVector), vSize); 790 emitVectorXor(asm, tmpVectors[0], tmpVectors[1], vSize); 791 emitVectorXor(asm, tmpVectors[2], tmpVectors[3], vSize); 792 emitVectorTest(asm, tmpVectors[0], vSize); 793 asm.jccb(AMD64Assembler.ConditionFlag.NotZero, noMatch); 794 emitVectorTest(asm, tmpVectors[2], vSize); 795 asm.jccb(AMD64Assembler.ConditionFlag.NotZero, noMatch); 796 asm.addq(tmp, elementsPerVectorLoop); 797 asm.jccb(AMD64Assembler.ConditionFlag.NotZero, loopBegin); 798 } 799 if (tailCount > 0) { 800 emitVectorLoad1(asm, tmpVectors[0], arrayPtr1, arrayBaseOffset1 + (tailCount << arrayIndexScale1.log2) - scaleDisplacement1(bytesPerVector), vSize); 801 emitVectorLoad2(asm, tmpVectors[1], arrayPtr2, arrayBaseOffset2 + (tailCount << arrayIndexScale2.log2) - scaleDisplacement2(bytesPerVector), vSize); 802 emitVectorXor(asm, tmpVectors[0], tmpVectors[1], vSize); 803 if (tailCount > elementsPerVector) { 804 emitVectorLoad1(asm, tmpVectors[2], arrayPtr1, arrayBaseOffset1, vSize); 805 emitVectorLoad2(asm, tmpVectors[3], arrayPtr2, arrayBaseOffset2, vSize); 806 emitVectorXor(asm, tmpVectors[2], tmpVectors[3], vSize); 807 emitVectorTest(asm, tmpVectors[2], vSize); 808 asm.jccb(AMD64Assembler.ConditionFlag.NotZero, noMatch); 809 } 810 emitVectorTest(asm, tmpVectors[0], vSize); 811 asm.jccb(AMD64Assembler.ConditionFlag.NotZero, noMatch); 812 } 813 } 814 } 815 816 private void emitMovBytes(AMD64MacroAssembler asm, Register dst, AMD64Address src, int size) { 817 switch (size) { 818 case 1: 819 if (signExtend) { 820 asm.movsbq(dst, src); 821 } else { 822 asm.movzbq(dst, src); 823 } 824 break; 825 case 2: 826 if (signExtend) { 827 asm.movswq(dst, src); 828 } else { 829 asm.movzwq(dst, src); 830 } 831 break; 832 case 4: 833 if (signExtend) { 834 asm.movslq(dst, src); 835 } else { 836 // there is no movzlq 837 asm.movl(dst, src); 838 } 839 break; 840 case 8: 841 asm.movq(dst, src); 842 break; 843 default: 844 throw new IllegalStateException(); 845 } 846 } 847 848 private static void emitXorBytes(AMD64MacroAssembler asm, Register dst, AMD64Address src, int size) { 849 OperandSize opSize = getOperandSize(size); 850 XOR.getRMOpcode(opSize).emit(asm, opSize, dst, src); 851 } 852 853 private static OperandSize getOperandSize(int size) { 854 switch (size) { 855 case 1: 856 return OperandSize.BYTE; 857 case 2: 858 return OperandSize.WORD; 859 case 4: 860 return OperandSize.DWORD; 861 case 8: 862 return OperandSize.QWORD; 863 default: 864 throw new IllegalStateException(); 865 } 866 } 867 }