--- old/src/jdk.internal.vm.compiler/share/classes/org.graalvm.compiler.lir.amd64/src/org/graalvm/compiler/lir/amd64/AMD64ArrayEqualsOp.java 2019-03-09 03:56:44.572936422 +0100 +++ new/src/jdk.internal.vm.compiler/share/classes/org.graalvm.compiler.lir.amd64/src/org/graalvm/compiler/lir/amd64/AMD64ArrayEqualsOp.java 2019-03-09 03:56:44.204933819 +0100 @@ -41,7 +41,7 @@ import org.graalvm.compiler.asm.amd64.AMD64MacroAssembler; import org.graalvm.compiler.asm.amd64.AVXKind; import org.graalvm.compiler.core.common.LIRKind; -import org.graalvm.compiler.core.common.NumUtil; +import org.graalvm.compiler.debug.GraalError; import org.graalvm.compiler.lir.LIRInstructionClass; import org.graalvm.compiler.lir.Opcode; import org.graalvm.compiler.lir.asm.CompilationResultBuilder; @@ -51,18 +51,29 @@ import static org.graalvm.compiler.lir.LIRInstruction.OperandFlag.ILLEGAL; import static org.graalvm.compiler.lir.LIRInstruction.OperandFlag.REG; +import java.util.Objects; + /** * Emits code which compares two arrays of the same length. If the CPU supports any vector * instructions specialized code is emitted to leverage these instructions. + * + * This op can also compare arrays of different integer types (e.g. {@code byte[]} and + * {@code char[]}) with on-the-fly sign- or zero-extension. If one of the given arrays is a + * {@code char[]} array, the smaller elements are zero-extended, otherwise they are sign-extended. */ @Opcode("ARRAY_EQUALS") public final class AMD64ArrayEqualsOp extends AMD64LIRInstruction { public static final LIRInstructionClass TYPE = LIRInstructionClass.create(AMD64ArrayEqualsOp.class); - private final JavaKind kind; - private final int arrayBaseOffset; - private final int arrayIndexScale; - private final int constantByteLength; + private final JavaKind kind1; + private final JavaKind kind2; + private final int arrayBaseOffset1; + private final int arrayBaseOffset2; + private final Scale arrayIndexScale1; + private final Scale arrayIndexScale2; + private final AVXKind.AVXSize vectorSize; + private final int constantLength; + private final boolean signExtend; @Def({REG}) private Value resultValue; @Alive({REG}) private Value array1Value; @@ -81,20 +92,21 @@ @Temp({REG, ILLEGAL}) private Value vectorTemp3; @Temp({REG, ILLEGAL}) private Value vectorTemp4; - public AMD64ArrayEqualsOp(LIRGeneratorTool tool, JavaKind kind, Value result, Value array1, Value array2, Value length, + public AMD64ArrayEqualsOp(LIRGeneratorTool tool, JavaKind kind1, JavaKind kind2, Value result, Value array1, Value array2, Value length, int constantLength, boolean directPointers, int maxVectorSize) { super(TYPE); - this.kind = kind; - - this.arrayBaseOffset = directPointers ? 0 : tool.getProviders().getMetaAccess().getArrayBaseOffset(kind); - this.arrayIndexScale = tool.getProviders().getMetaAccess().getArrayIndexScale(kind); - - if (constantLength >= 0 && arrayIndexScale > 1) { - // scale length - this.constantByteLength = constantLength << NumUtil.log2Ceil(arrayIndexScale); - } else { - this.constantByteLength = constantLength; - } + this.kind1 = kind1; + this.kind2 = kind2; + this.signExtend = kind1 != JavaKind.Char && kind2 != JavaKind.Char; + + assert kind1.isNumericInteger() && kind2.isNumericInteger() || kind1 == kind2; + + this.arrayBaseOffset1 = directPointers ? 0 : tool.getProviders().getMetaAccess().getArrayBaseOffset(kind1); + this.arrayBaseOffset2 = directPointers ? 0 : tool.getProviders().getMetaAccess().getArrayBaseOffset(kind2); + this.arrayIndexScale1 = Objects.requireNonNull(Scale.fromInt(tool.getProviders().getMetaAccess().getArrayIndexScale(kind1))); + this.arrayIndexScale2 = Objects.requireNonNull(Scale.fromInt(tool.getProviders().getMetaAccess().getArrayIndexScale(kind2))); + this.vectorSize = ((AMD64) tool.target().arch).getFeatures().contains(CPUFeature.AVX2) && (maxVectorSize < 0 || maxVectorSize >= 32) ? AVXKind.AVXSize.YMM : AVXKind.AVXSize.XMM; + this.constantLength = constantLength; this.resultValue = result; this.array1Value = array1; @@ -107,10 +119,10 @@ this.temp3 = tool.newVariable(LIRKind.value(tool.target().arch.getWordKind())); this.temp4 = tool.newVariable(LIRKind.value(tool.target().arch.getWordKind())); - this.temp5 = kind.isNumericFloat() ? tool.newVariable(LIRKind.value(tool.target().arch.getWordKind())) : Value.ILLEGAL; - if (kind == JavaKind.Float) { + this.temp5 = kind1.isNumericFloat() || kind1 != kind2 ? tool.newVariable(LIRKind.value(tool.target().arch.getWordKind())) : Value.ILLEGAL; + if (kind1 == JavaKind.Float) { this.tempXMM = tool.newVariable(LIRKind.value(AMD64Kind.SINGLE)); - } else if (kind == JavaKind.Double) { + } else if (kind1 == JavaKind.Double) { this.tempXMM = tool.newVariable(LIRKind.value(AMD64Kind.DOUBLE)); } else { this.tempXMM = Value.ILLEGAL; @@ -119,7 +131,7 @@ // We only need the vector temporaries if we generate SSE code. if (supportsSSE41(tool.target())) { if (canGenerateConstantLengthCompare(tool.target())) { - LIRKind lirKind = LIRKind.value(supportsAVX2(tool.target()) && (maxVectorSize < 0 || maxVectorSize >= 32) ? AMD64Kind.V256_BYTE : AMD64Kind.V128_BYTE); + LIRKind lirKind = LIRKind.value(vectorSize == AVXKind.AVXSize.YMM ? AMD64Kind.V256_BYTE : AMD64Kind.V128_BYTE); this.vectorTemp1 = tool.newVariable(lirKind); this.vectorTemp2 = tool.newVariable(lirKind); this.vectorTemp3 = tool.newVariable(lirKind); @@ -139,7 +151,7 @@ } private boolean canGenerateConstantLengthCompare(TargetDescription target) { - return constantByteLength >= 0 && kind.isNumericInteger() && supportsSSE41(target); + return constantLength >= 0 && kind1.isNumericInteger() && (kind1 == kind2 || getElementsPerVector(AVXKind.AVXSize.XMM) <= constantLength) && supportsSSE41(target); } @Override @@ -153,26 +165,19 @@ Label done = new Label(); // Load array base addresses. - masm.leaq(array1, new AMD64Address(asRegister(array1Value), arrayBaseOffset)); - masm.leaq(array2, new AMD64Address(asRegister(array2Value), arrayBaseOffset)); + masm.leaq(array1, new AMD64Address(asRegister(array1Value), arrayBaseOffset1)); + masm.leaq(array2, new AMD64Address(asRegister(array2Value), arrayBaseOffset2)); if (canGenerateConstantLengthCompare(crb.target)) { - emitConstantLengthArrayCompareBytes(masm, array1, array2, asRegister(temp3), asRegister(temp4), - new Register[]{asRegister(vectorTemp1), asRegister(vectorTemp2), asRegister(vectorTemp3), asRegister(vectorTemp4)}, - falseLabel, constantByteLength, AVXKind.getRegisterSize(vectorTemp1).getBytes()); + emitConstantLengthArrayCompareBytes(crb, masm, array1, array2, asRegister(temp3), asRegister(temp4), + new Register[]{asRegister(vectorTemp1), asRegister(vectorTemp2), asRegister(vectorTemp3), asRegister(vectorTemp4)}, falseLabel); } else { Register length = asRegister(temp3); - - // Get array length in bytes. + // Get array length. masm.movl(length, asRegister(lengthValue)); - - if (arrayIndexScale > 1) { - masm.shll(length, NumUtil.log2Ceil(arrayIndexScale)); // scale length - } - - masm.movl(result, length); // copy - - emitArrayCompare(crb, masm, kind, result, array1, array2, length, temp4, temp5, tempXMM, vectorTemp1, vectorTemp2, trueLabel, falseLabel); + // copy + masm.movl(result, length); + emitArrayCompare(crb, masm, result, array1, array2, length, trueLabel, falseLabel); } // Return true @@ -188,19 +193,18 @@ masm.bind(done); } - private static void emitArrayCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, JavaKind kind, + private void emitArrayCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, Register result, Register array1, Register array2, Register length, - Value temp4, Value temp5, Value tempXMM, Value vectorTemp1, Value vectorTemp2, Label trueLabel, Label falseLabel) { - if (supportsAVX2(crb.target)) { - emitAVXCompare(crb, masm, kind, result, array1, array2, length, temp4, temp5, tempXMM, vectorTemp1, vectorTemp2, trueLabel, falseLabel); - } else if (supportsSSE41(crb.target)) { - // this code is used for AVX as well because our backend correctly ensures that - // VEX-prefixed instructions are emitted if AVX is supported - emitSSE41Compare(crb, masm, kind, result, array1, array2, length, temp4, temp5, tempXMM, vectorTemp1, vectorTemp2, trueLabel, falseLabel); + if (supportsSSE41(crb.target)) { + emitVectorCompare(crb, masm, result, array1, array2, length, trueLabel, falseLabel); + } + if (kind1 == kind2) { + emit8ByteCompare(crb, masm, result, array1, array2, length, trueLabel, falseLabel); + emitTailCompares(masm, result, array1, array2, length, trueLabel, falseLabel); + } else { + emitDifferentKindsElementWiseCompare(crb, masm, result, array1, array2, length, trueLabel, falseLabel); } - emit8ByteCompare(crb, masm, kind, result, array1, array2, length, temp4, tempXMM, trueLabel, falseLabel); - emitTailCompares(masm, kind, result, array1, array2, length, temp4, tempXMM, trueLabel, falseLabel); } /** @@ -215,49 +219,44 @@ } /** - * Vector size used in {@link #emitSSE41Compare}. + * Emits code that uses SSE4.1/AVX1 128-bit (16-byte) or AVX2 256-bit (32-byte) vector compares. */ - private static final int SSE4_1_VECTOR_SIZE = 16; - - /** - * Emits code that uses SSE4.1 128-bit (16-byte) vector compares. - */ - private static void emitSSE41Compare(CompilationResultBuilder crb, AMD64MacroAssembler masm, JavaKind kind, + private void emitVectorCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, Register result, Register array1, Register array2, Register length, - Value temp4, Value temp5, Value tempXMM, Value vectorTemp1, Value vectorTemp2, Label trueLabel, Label falseLabel) { assert supportsSSE41(crb.target); Register vector1 = asRegister(vectorTemp1); Register vector2 = asRegister(vectorTemp2); + int elementsPerVector = getElementsPerVector(vectorSize); + Label loop = new Label(); Label compareTail = new Label(); - boolean requiresNaNCheck = kind.isNumericFloat(); + boolean requiresNaNCheck = kind1.isNumericFloat(); Label loopCheck = new Label(); Label nanCheck = new Label(); // Compare 16-byte vectors - masm.andl(result, SSE4_1_VECTOR_SIZE - 1); // tail count (in bytes) - masm.andl(length, ~(SSE4_1_VECTOR_SIZE - 1)); // vector count (in bytes) + masm.andl(result, elementsPerVector - 1); // tail count + masm.andl(length, ~(elementsPerVector - 1)); // vector count masm.jcc(ConditionFlag.Zero, compareTail); - masm.leaq(array1, new AMD64Address(array1, length, Scale.Times1, 0)); - masm.leaq(array2, new AMD64Address(array2, length, Scale.Times1, 0)); + masm.leaq(array1, new AMD64Address(array1, length, arrayIndexScale1, 0)); + masm.leaq(array2, new AMD64Address(array2, length, arrayIndexScale2, 0)); masm.negq(length); // Align the main loop masm.align(crb.target.wordSize * 2); masm.bind(loop); - masm.movdqu(vector1, new AMD64Address(array1, length, Scale.Times1, 0)); - masm.movdqu(vector2, new AMD64Address(array2, length, Scale.Times1, 0)); - masm.pxor(vector1, vector2); - masm.ptest(vector1, vector1); + emitVectorLoad1(masm, vector1, array1, length, 0, vectorSize); + emitVectorLoad2(masm, vector2, array2, length, 0, vectorSize); + emitVectorCmp(masm, vector1, vector2, vectorSize); masm.jcc(ConditionFlag.NotZero, requiresNaNCheck ? nanCheck : falseLabel); masm.bind(loopCheck); - masm.addq(length, SSE4_1_VECTOR_SIZE); + masm.addq(length, elementsPerVector); masm.jcc(ConditionFlag.NotZero, loop); masm.testl(result, result); @@ -267,7 +266,7 @@ Label unalignedCheck = new Label(); masm.jmpb(unalignedCheck); masm.bind(nanCheck); - emitFloatCompareWithinRange(crb, masm, kind, array1, array2, length, temp4, temp5, tempXMM, 0, falseLabel, SSE4_1_VECTOR_SIZE); + emitFloatCompareWithinRange(crb, masm, array1, array2, length, 0, falseLabel, elementsPerVector); masm.jmpb(loopCheck); masm.bind(unalignedCheck); } @@ -276,13 +275,12 @@ * Compare the remaining bytes with an unaligned memory load aligned to the end of the * array. */ - masm.movdqu(vector1, new AMD64Address(array1, result, Scale.Times1, -SSE4_1_VECTOR_SIZE)); - masm.movdqu(vector2, new AMD64Address(array2, result, Scale.Times1, -SSE4_1_VECTOR_SIZE)); - masm.pxor(vector1, vector2); - masm.ptest(vector1, vector1); + emitVectorLoad1(masm, vector1, array1, result, scaleDisplacement1(-vectorSize.getBytes()), vectorSize); + emitVectorLoad2(masm, vector2, array2, result, scaleDisplacement2(-vectorSize.getBytes()), vectorSize); + emitVectorCmp(masm, vector1, vector2, vectorSize); if (requiresNaNCheck) { masm.jcc(ConditionFlag.Zero, trueLabel); - emitFloatCompareWithinRange(crb, masm, kind, array1, array2, result, temp4, temp5, tempXMM, -SSE4_1_VECTOR_SIZE, falseLabel, SSE4_1_VECTOR_SIZE); + emitFloatCompareWithinRange(crb, masm, array1, array2, result, -vectorSize.getBytes(), falseLabel, elementsPerVector); } else { masm.jcc(ConditionFlag.NotZero, falseLabel); } @@ -292,90 +290,159 @@ masm.movl(length, result); } - /** - * Returns if the underlying AMD64 architecture supports AVX instructions. - * - * @param target target description of the underlying architecture - * @return true if the underlying architecture supports AVX - */ - private static boolean supportsAVX2(TargetDescription target) { - AMD64 arch = (AMD64) target.arch; - return arch.getFeatures().contains(CPUFeature.AVX2); + private int getElementsPerVector(AVXKind.AVXSize vSize) { + return vSize.getBytes() >> Math.max(arrayIndexScale1.log2, arrayIndexScale2.log2); } - /** - * Vector size used in {@link #emitAVXCompare}. - */ - private static final int AVX_VECTOR_SIZE = 32; + private void emitVectorLoad1(AMD64MacroAssembler asm, Register dst, Register src, int displacement, AVXKind.AVXSize size) { + emitVectorLoad1(asm, dst, src, Register.None, displacement, size); + } - private static void emitAVXCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, JavaKind kind, Register result, - Register array1, Register array2, Register length, - Value temp4, Value temp5, Value tempXMM, Value vectorTemp1, Value vectorTemp2, - Label trueLabel, Label falseLabel) { - assert supportsAVX2(crb.target); + private void emitVectorLoad2(AMD64MacroAssembler asm, Register dst, Register src, int displacement, AVXKind.AVXSize size) { + emitVectorLoad2(asm, dst, src, Register.None, displacement, size); + } - Register vector1 = asRegister(vectorTemp1); - Register vector2 = asRegister(vectorTemp2); + private void emitVectorLoad1(AMD64MacroAssembler asm, Register dst, Register src, Register index, int displacement, AVXKind.AVXSize size) { + emitVectorLoad(asm, dst, src, index, displacement, arrayIndexScale1, arrayIndexScale2, size); + } - Label loop = new Label(); - Label compareTail = new Label(); + private void emitVectorLoad2(AMD64MacroAssembler asm, Register dst, Register src, Register index, int displacement, AVXKind.AVXSize size) { + emitVectorLoad(asm, dst, src, index, displacement, arrayIndexScale2, arrayIndexScale1, size); + } - boolean requiresNaNCheck = kind.isNumericFloat(); - Label loopCheck = new Label(); - Label nanCheck = new Label(); + private void emitVectorLoad(AMD64MacroAssembler asm, Register dst, Register src, Register index, int displacement, Scale ownScale, Scale otherScale, AVXKind.AVXSize size) { + AMD64Address address = new AMD64Address(src, index, ownScale, displacement); + if (ownScale.value < otherScale.value) { + if (size == AVXKind.AVXSize.YMM) { + getAVX2LoadAndExtendOp(ownScale, otherScale, signExtend).emit(asm, size, dst, address); + } else { + loadAndExtendSSE(asm, dst, address, ownScale, otherScale, signExtend); + } + } else { + if (size == AVXKind.AVXSize.YMM) { + asm.vmovdqu(dst, address); + } else { + asm.movdqu(dst, address); + } + } + } - // Compare 32-byte vectors - masm.andl(result, AVX_VECTOR_SIZE - 1); // tail count (in bytes) - masm.andl(length, ~(AVX_VECTOR_SIZE - 1)); // vector count (in bytes) - masm.jcc(ConditionFlag.Zero, compareTail); + private int scaleDisplacement1(int displacement) { + return scaleDisplacement(displacement, arrayIndexScale1, arrayIndexScale2); + } - masm.leaq(array1, new AMD64Address(array1, length, Scale.Times1, 0)); - masm.leaq(array2, new AMD64Address(array2, length, Scale.Times1, 0)); - masm.negq(length); + private int scaleDisplacement2(int displacement) { + return scaleDisplacement(displacement, arrayIndexScale2, arrayIndexScale1); + } - // Align the main loop - masm.align(crb.target.wordSize * 2); - masm.bind(loop); - masm.vmovdqu(vector1, new AMD64Address(array1, length, Scale.Times1, 0)); - masm.vmovdqu(vector2, new AMD64Address(array2, length, Scale.Times1, 0)); - masm.vpxor(vector1, vector1, vector2); - masm.vptest(vector1, vector1); - masm.jcc(ConditionFlag.NotZero, requiresNaNCheck ? nanCheck : falseLabel); + private static int scaleDisplacement(int displacement, Scale ownScale, Scale otherScale) { + if (ownScale.value < otherScale.value) { + return displacement >> (otherScale.log2 - ownScale.log2); + } + return displacement; + } - masm.bind(loopCheck); - masm.addq(length, AVX_VECTOR_SIZE); - masm.jcc(ConditionFlag.NotZero, loop); + private static AMD64Assembler.VexRMOp getAVX2LoadAndExtendOp(Scale ownScale, Scale otherScale, boolean signExtend) { + switch (ownScale) { + case Times1: + switch (otherScale) { + case Times2: + return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXBW : AMD64Assembler.VexRMOp.VPMOVZXBW; + case Times4: + return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXBD : AMD64Assembler.VexRMOp.VPMOVZXBD; + case Times8: + return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXBQ : AMD64Assembler.VexRMOp.VPMOVZXBQ; + } + throw GraalError.shouldNotReachHere(); + case Times2: + switch (otherScale) { + case Times4: + return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXWD : AMD64Assembler.VexRMOp.VPMOVZXWD; + case Times8: + return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXWQ : AMD64Assembler.VexRMOp.VPMOVZXWQ; + } + throw GraalError.shouldNotReachHere(); + case Times4: + return signExtend ? AMD64Assembler.VexRMOp.VPMOVSXDQ : AMD64Assembler.VexRMOp.VPMOVZXDQ; + } + throw GraalError.shouldNotReachHere(); + } + + private static void loadAndExtendSSE(AMD64MacroAssembler asm, Register dst, AMD64Address src, Scale ownScale, Scale otherScale, boolean signExtend) { + switch (ownScale) { + case Times1: + switch (otherScale) { + case Times2: + if (signExtend) { + asm.pmovsxbw(dst, src); + } else { + asm.pmovzxbw(dst, src); + } + return; + case Times4: + if (signExtend) { + asm.pmovsxbd(dst, src); + } else { + asm.pmovzxbd(dst, src); + } + return; + case Times8: + if (signExtend) { + asm.pmovsxbq(dst, src); + } else { + asm.pmovzxbq(dst, src); + } + return; + } + throw GraalError.shouldNotReachHere(); + case Times2: + switch (otherScale) { + case Times4: + if (signExtend) { + asm.pmovsxwd(dst, src); + } else { + asm.pmovzxwd(dst, src); + } + return; + case Times8: + if (signExtend) { + asm.pmovsxwq(dst, src); + } else { + asm.pmovzxwq(dst, src); + } + return; + } + throw GraalError.shouldNotReachHere(); + case Times4: + if (signExtend) { + asm.pmovsxdq(dst, src); + } else { + asm.pmovzxdq(dst, src); + } + return; + } + throw GraalError.shouldNotReachHere(); + } - masm.testl(result, result); - masm.jcc(ConditionFlag.Zero, trueLabel); + private static void emitVectorCmp(AMD64MacroAssembler masm, Register vector1, Register vector2, AVXKind.AVXSize size) { + emitVectorXor(masm, vector1, vector2, size); + emitVectorTest(masm, vector1, size); + } - if (requiresNaNCheck) { - Label unalignedCheck = new Label(); - masm.jmpb(unalignedCheck); - masm.bind(nanCheck); - emitFloatCompareWithinRange(crb, masm, kind, array1, array2, length, temp4, temp5, tempXMM, 0, falseLabel, AVX_VECTOR_SIZE); - masm.jmpb(loopCheck); - masm.bind(unalignedCheck); + private static void emitVectorXor(AMD64MacroAssembler masm, Register vector1, Register vector2, AVXKind.AVXSize size) { + if (size == AVXKind.AVXSize.YMM) { + masm.vpxor(vector1, vector1, vector2); + } else { + masm.pxor(vector1, vector2); } + } - /* - * Compare the remaining bytes with an unaligned memory load aligned to the end of the - * array. - */ - masm.vmovdqu(vector1, new AMD64Address(array1, result, Scale.Times1, -AVX_VECTOR_SIZE)); - masm.vmovdqu(vector2, new AMD64Address(array2, result, Scale.Times1, -AVX_VECTOR_SIZE)); - masm.vpxor(vector1, vector1, vector2); - masm.vptest(vector1, vector1); - if (requiresNaNCheck) { - masm.jcc(ConditionFlag.Zero, trueLabel); - emitFloatCompareWithinRange(crb, masm, kind, array1, array2, result, temp4, temp5, tempXMM, -AVX_VECTOR_SIZE, falseLabel, AVX_VECTOR_SIZE); + private static void emitVectorTest(AMD64MacroAssembler masm, Register vector1, AVXKind.AVXSize size) { + if (size == AVXKind.AVXSize.YMM) { + masm.vptest(vector1, vector1); } else { - masm.jcc(ConditionFlag.NotZero, falseLabel); + masm.ptest(vector1, vector1); } - masm.jmp(trueLabel); - - masm.bind(compareTail); - masm.movl(length, result); } /** @@ -386,34 +453,37 @@ /** * Emits code that uses 8-byte vector compares. */ - private static void emit8ByteCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, JavaKind kind, Register result, Register array1, Register array2, Register length, Value temp4, - Value tempXMM, Label trueLabel, Label falseLabel) { + private void emit8ByteCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, + Register result, Register array1, Register array2, Register length, Label trueLabel, Label falseLabel) { + assert kind1 == kind2; Label loop = new Label(); Label compareTail = new Label(); - boolean requiresNaNCheck = kind.isNumericFloat(); + int elementsPerVector = 8 >> arrayIndexScale1.log2; + + boolean requiresNaNCheck = kind1.isNumericFloat(); Label loopCheck = new Label(); Label nanCheck = new Label(); Register temp = asRegister(temp4); - masm.andl(result, VECTOR_SIZE - 1); // tail count (in bytes) - masm.andl(length, ~(VECTOR_SIZE - 1)); // vector count (in bytes) + masm.andl(result, elementsPerVector - 1); // tail count + masm.andl(length, ~(elementsPerVector - 1)); // vector count masm.jcc(ConditionFlag.Zero, compareTail); - masm.leaq(array1, new AMD64Address(array1, length, Scale.Times1, 0)); - masm.leaq(array2, new AMD64Address(array2, length, Scale.Times1, 0)); + masm.leaq(array1, new AMD64Address(array1, length, arrayIndexScale1, 0)); + masm.leaq(array2, new AMD64Address(array2, length, arrayIndexScale2, 0)); masm.negq(length); // Align the main loop masm.align(crb.target.wordSize * 2); masm.bind(loop); - masm.movq(temp, new AMD64Address(array1, length, Scale.Times1, 0)); - masm.cmpq(temp, new AMD64Address(array2, length, Scale.Times1, 0)); + masm.movq(temp, new AMD64Address(array1, length, arrayIndexScale1, 0)); + masm.cmpq(temp, new AMD64Address(array2, length, arrayIndexScale2, 0)); masm.jcc(ConditionFlag.NotEqual, requiresNaNCheck ? nanCheck : falseLabel); masm.bind(loopCheck); - masm.addq(length, VECTOR_SIZE); + masm.addq(length, elementsPerVector); masm.jccb(ConditionFlag.NotZero, loop); masm.testl(result, result); @@ -425,8 +495,8 @@ masm.jmpb(unalignedCheck); masm.bind(nanCheck); // At most two iterations, unroll in the emitted code. - for (int offset = 0; offset < VECTOR_SIZE; offset += kind.getByteCount()) { - emitFloatCompare(masm, kind, array1, array2, length, temp4, tempXMM, offset, falseLabel, kind.getByteCount() == VECTOR_SIZE); + for (int offset = 0; offset < VECTOR_SIZE; offset += kind1.getByteCount()) { + emitFloatCompare(masm, array1, array2, length, offset, falseLabel, kind1.getByteCount() == VECTOR_SIZE); } masm.jmpb(loopCheck); masm.bind(unalignedCheck); @@ -436,13 +506,13 @@ * Compare the remaining bytes with an unaligned memory load aligned to the end of the * array. */ - masm.movq(temp, new AMD64Address(array1, result, Scale.Times1, -VECTOR_SIZE)); - masm.cmpq(temp, new AMD64Address(array2, result, Scale.Times1, -VECTOR_SIZE)); + masm.movq(temp, new AMD64Address(array1, result, arrayIndexScale1, -VECTOR_SIZE)); + masm.cmpq(temp, new AMD64Address(array2, result, arrayIndexScale2, -VECTOR_SIZE)); if (requiresNaNCheck) { masm.jcc(ConditionFlag.Equal, trueLabel); // At most two iterations, unroll in the emitted code. - for (int offset = 0; offset < VECTOR_SIZE; offset += kind.getByteCount()) { - emitFloatCompare(masm, kind, array1, array2, result, temp4, tempXMM, -VECTOR_SIZE + offset, falseLabel, kind.getByteCount() == VECTOR_SIZE); + for (int offset = 0; offset < VECTOR_SIZE; offset += kind1.getByteCount()) { + emitFloatCompare(masm, array1, array2, result, -VECTOR_SIZE + offset, falseLabel, kind1.getByteCount() == VECTOR_SIZE); } } else { masm.jccb(ConditionFlag.NotEqual, falseLabel); @@ -456,34 +526,35 @@ /** * Emits code to compare the remaining 1 to 4 bytes. */ - private static void emitTailCompares(AMD64MacroAssembler masm, JavaKind kind, Register result, Register array1, Register array2, Register length, Value temp4, Value tempXMM, - Label trueLabel, Label falseLabel) { + private void emitTailCompares(AMD64MacroAssembler masm, + Register result, Register array1, Register array2, Register length, Label trueLabel, Label falseLabel) { + assert kind1 == kind2; Label compare2Bytes = new Label(); Label compare1Byte = new Label(); Register temp = asRegister(temp4); - if (kind.getByteCount() <= 4) { + if (kind1.getByteCount() <= 4) { // Compare trailing 4 bytes, if any. - masm.testl(result, 4); + masm.testl(result, arrayIndexScale1.log2 == 0 ? 4 : 4 >> arrayIndexScale1.log2); masm.jccb(ConditionFlag.Zero, compare2Bytes); masm.movl(temp, new AMD64Address(array1, 0)); masm.cmpl(temp, new AMD64Address(array2, 0)); - if (kind == JavaKind.Float) { + if (kind1 == JavaKind.Float) { masm.jccb(ConditionFlag.Equal, trueLabel); - emitFloatCompare(masm, kind, array1, array2, Register.None, temp4, tempXMM, 0, falseLabel, true); + emitFloatCompare(masm, array1, array2, Register.None, 0, falseLabel, true); masm.jmpb(trueLabel); } else { masm.jccb(ConditionFlag.NotEqual, falseLabel); } - if (kind.getByteCount() <= 2) { + if (kind1.getByteCount() <= 2) { // Move array pointers forward. masm.leaq(array1, new AMD64Address(array1, 4)); masm.leaq(array2, new AMD64Address(array2, 4)); // Compare trailing 2 bytes, if any. masm.bind(compare2Bytes); - masm.testl(result, 2); + masm.testl(result, arrayIndexScale1.log2 == 0 ? 2 : 2 >> arrayIndexScale1.log2); masm.jccb(ConditionFlag.Zero, compare1Byte); masm.movzwl(temp, new AMD64Address(array1, 0)); masm.movzwl(length, new AMD64Address(array2, 0)); @@ -491,7 +562,7 @@ masm.jccb(ConditionFlag.NotEqual, falseLabel); // The one-byte tail compare is only required for boolean and byte arrays. - if (kind.getByteCount() <= 1) { + if (kind1.getByteCount() <= 1) { // Move array pointers forward before we compare the last trailing byte. masm.leaq(array1, new AMD64Address(array1, 2)); masm.leaq(array2, new AMD64Address(array2, 2)); @@ -513,28 +584,82 @@ } } + private void emitDifferentKindsElementWiseCompare(CompilationResultBuilder crb, AMD64MacroAssembler masm, + Register result, Register array1, Register array2, Register length, Label trueLabel, Label falseLabel) { + assert kind1 != kind2; + assert kind1.isNumericInteger() && kind2.isNumericInteger(); + Label loop = new Label(); + Label compareTail = new Label(); + + int elementsPerLoopIteration = 4; + + Register tmp1 = asRegister(temp4); + Register tmp2 = asRegister(temp5); + + masm.andl(result, elementsPerLoopIteration - 1); // tail count + masm.andl(length, ~(elementsPerLoopIteration - 1)); // bulk loop count + masm.jcc(ConditionFlag.Zero, compareTail); + + masm.leaq(array1, new AMD64Address(array1, length, arrayIndexScale1, 0)); + masm.leaq(array2, new AMD64Address(array2, length, arrayIndexScale2, 0)); + masm.negq(length); + + // clear comparison registers because of the missing movzlq instruction + masm.xorq(tmp1, tmp1); + masm.xorq(tmp2, tmp2); + + // Align the main loop + masm.align(crb.target.wordSize * 2); + masm.bind(loop); + for (int i = 0; i < elementsPerLoopIteration; i++) { + emitMovBytes(masm, tmp1, new AMD64Address(array1, length, arrayIndexScale1, i << arrayIndexScale1.log2), kind1.getByteCount()); + emitMovBytes(masm, tmp2, new AMD64Address(array2, length, arrayIndexScale2, i << arrayIndexScale2.log2), kind2.getByteCount()); + masm.cmpq(tmp1, tmp2); + masm.jcc(ConditionFlag.NotEqual, falseLabel); + } + masm.addq(length, elementsPerLoopIteration); + masm.jccb(ConditionFlag.NotZero, loop); + + masm.bind(compareTail); + masm.testl(result, result); + masm.jcc(ConditionFlag.Zero, trueLabel); + for (int i = 0; i < elementsPerLoopIteration - 1; i++) { + emitMovBytes(masm, tmp1, new AMD64Address(array1, length, arrayIndexScale1, 0), kind1.getByteCount()); + emitMovBytes(masm, tmp2, new AMD64Address(array2, length, arrayIndexScale2, 0), kind2.getByteCount()); + masm.cmpq(tmp1, tmp2); + masm.jcc(ConditionFlag.NotEqual, falseLabel); + if (i < elementsPerLoopIteration - 2) { + masm.incrementq(length, 1); + masm.decrementq(result, 1); + masm.jcc(ConditionFlag.Zero, trueLabel); + } else { + masm.jmpb(trueLabel); + } + } + } + /** * Emits code to fall through if {@code src} is NaN, otherwise jump to {@code branchOrdered}. */ - private static void emitNaNCheck(AMD64MacroAssembler masm, JavaKind kind, Value tempXMM, AMD64Address src, Label branchIfNonNaN) { - assert kind.isNumericFloat(); + private void emitNaNCheck(AMD64MacroAssembler masm, AMD64Address src, Label branchIfNonNaN) { + assert kind1.isNumericFloat(); Register tempXMMReg = asRegister(tempXMM); - if (kind == JavaKind.Float) { + if (kind1 == JavaKind.Float) { masm.movflt(tempXMMReg, src); } else { masm.movdbl(tempXMMReg, src); } - SSEOp.UCOMIS.emit(masm, kind == JavaKind.Float ? OperandSize.PS : OperandSize.PD, tempXMMReg, tempXMMReg); + SSEOp.UCOMIS.emit(masm, kind1 == JavaKind.Float ? OperandSize.PS : OperandSize.PD, tempXMMReg, tempXMMReg); masm.jcc(ConditionFlag.NoParity, branchIfNonNaN); } /** * Emits code to compare if two floats are bitwise equal or both NaN. */ - private static void emitFloatCompare(AMD64MacroAssembler masm, JavaKind kind, Register base1, Register base2, Register index, Value temp4, Value tempXMM, int offset, Label falseLabel, + private void emitFloatCompare(AMD64MacroAssembler masm, Register base1, Register base2, Register index, int offset, Label falseLabel, boolean skipBitwiseCompare) { - AMD64Address address1 = new AMD64Address(base1, index, Scale.Times1, offset); - AMD64Address address2 = new AMD64Address(base2, index, Scale.Times1, offset); + AMD64Address address1 = new AMD64Address(base1, index, arrayIndexScale1, offset); + AMD64Address address2 = new AMD64Address(base2, index, arrayIndexScale2, offset); Label bitwiseEqual = new Label(); @@ -542,7 +667,7 @@ // Bitwise compare Register temp = asRegister(temp4); - if (kind == JavaKind.Float) { + if (kind1 == JavaKind.Float) { masm.movl(temp, address1); masm.cmpl(temp, address2); } else { @@ -552,8 +677,8 @@ masm.jccb(ConditionFlag.Equal, bitwiseEqual); } - emitNaNCheck(masm, kind, tempXMM, address1, falseLabel); - emitNaNCheck(masm, kind, tempXMM, address2, falseLabel); + emitNaNCheck(masm, address1, falseLabel); + emitNaNCheck(masm, address2, falseLabel); masm.bind(bitwiseEqual); } @@ -561,9 +686,9 @@ /** * Emits code to compare float equality within a range. */ - private static void emitFloatCompareWithinRange(CompilationResultBuilder crb, AMD64MacroAssembler masm, JavaKind kind, Register base1, Register base2, Register index, Value temp4, Value temp5, - Value tempXMM, int offset, Label falseLabel, int range) { - assert kind.isNumericFloat(); + private void emitFloatCompareWithinRange(CompilationResultBuilder crb, AMD64MacroAssembler masm, + Register base1, Register base2, Register index, int offset, Label falseLabel, int range) { + assert kind1.isNumericFloat(); Label loop = new Label(); Register i = asRegister(temp5); @@ -572,9 +697,9 @@ // Align the main loop masm.align(crb.target.wordSize * 2); masm.bind(loop); - emitFloatCompare(masm, kind, base1, base2, index, temp4, tempXMM, offset, falseLabel, kind.getByteCount() == range); - masm.addq(index, kind.getByteCount()); - masm.addq(i, kind.getByteCount()); + emitFloatCompare(masm, base1, base2, index, offset, falseLabel, range == 1); + masm.incrementq(index, 1); + masm.incrementq(i, 1); masm.jccb(ConditionFlag.NotZero, loop); // Floats within the range are equal, revert change to the register index masm.subq(index, range); @@ -585,189 +710,104 @@ * {@code arrayPtr1[0..nBytes]} and {@code arrayPtr2[0..nBytes]}. If they match, execution * continues directly after the emitted code block, otherwise we jump to {@code noMatch}. */ - private static void emitConstantLengthArrayCompareBytes( + private void emitConstantLengthArrayCompareBytes( + CompilationResultBuilder crb, AMD64MacroAssembler asm, Register arrayPtr1, Register arrayPtr2, Register tmp1, Register tmp2, Register[] tmpVectors, - Label noMatch, - int nBytes, - int bytesPerVector) { - assert bytesPerVector >= 16; - if (nBytes == 0) { + Label noMatch) { + if (constantLength == 0) { // do nothing return; } - if (nBytes < 16) { + AVXKind.AVXSize vSize = vectorSize; + if (constantLength < getElementsPerVector(vectorSize)) { + vSize = AVXKind.AVXSize.XMM; + } + int elementsPerVector = getElementsPerVector(vSize); + if (elementsPerVector > constantLength) { + assert kind1 == kind2; + int byteLength = constantLength << arrayIndexScale1.log2; // array is shorter than any vector register, use regular CMP instructions - int movSize = (nBytes < 2) ? 1 : ((nBytes < 4) ? 2 : ((nBytes < 8) ? 4 : 8)); + int movSize = (byteLength < 2) ? 1 : ((byteLength < 4) ? 2 : ((byteLength < 8) ? 4 : 8)); emitMovBytes(asm, tmp1, new AMD64Address(arrayPtr1), movSize); emitMovBytes(asm, tmp2, new AMD64Address(arrayPtr2), movSize); emitCmpBytes(asm, tmp1, tmp2, movSize); asm.jcc(AMD64Assembler.ConditionFlag.NotEqual, noMatch); - if (nBytes > movSize) { - emitMovBytes(asm, tmp1, new AMD64Address(arrayPtr1, nBytes - movSize), movSize); - emitMovBytes(asm, tmp2, new AMD64Address(arrayPtr2, nBytes - movSize), movSize); + if (byteLength > movSize) { + emitMovBytes(asm, tmp1, new AMD64Address(arrayPtr1, byteLength - movSize), movSize); + emitMovBytes(asm, tmp2, new AMD64Address(arrayPtr2, byteLength - movSize), movSize); emitCmpBytes(asm, tmp1, tmp2, movSize); asm.jcc(AMD64Assembler.ConditionFlag.NotEqual, noMatch); } - } else if (nBytes < 32 && bytesPerVector >= 32) { - // we could use YMM registers, but the array is too short, force XMM registers - int bytesPerXMMVector = AVXKind.AVXSize.XMM.getBytes(); - AMD64Assembler.VexMoveOp.VMOVDQU.emit(asm, AVXKind.AVXSize.XMM, tmpVectors[0], new AMD64Address(arrayPtr1)); - AMD64Assembler.VexMoveOp.VMOVDQU.emit(asm, AVXKind.AVXSize.XMM, tmpVectors[1], new AMD64Address(arrayPtr2)); - AMD64Assembler.VexRVMOp.VPXOR.emit(asm, AVXKind.AVXSize.XMM, tmpVectors[0], tmpVectors[0], tmpVectors[1]); - if (nBytes > bytesPerXMMVector) { - AMD64Assembler.VexMoveOp.VMOVDQU.emit(asm, AVXKind.AVXSize.XMM, tmpVectors[2], new AMD64Address(arrayPtr1, nBytes - bytesPerXMMVector)); - AMD64Assembler.VexMoveOp.VMOVDQU.emit(asm, AVXKind.AVXSize.XMM, tmpVectors[3], new AMD64Address(arrayPtr2, nBytes - bytesPerXMMVector)); - AMD64Assembler.VexRVMOp.VPXOR.emit(asm, AVXKind.AVXSize.XMM, tmpVectors[2], tmpVectors[2], tmpVectors[3]); - AMD64Assembler.VexRMOp.VPTEST.emit(asm, AVXKind.AVXSize.XMM, tmpVectors[2], tmpVectors[2]); + } else { + int elementsPerVectorLoop = 2 * elementsPerVector; + int tailCount = constantLength & (elementsPerVectorLoop - 1); + int vectorCount = constantLength & ~(elementsPerVectorLoop - 1); + int bytesPerVector = vSize.getBytes(); + if (vectorCount > 0) { + Label loopBegin = new Label(); + asm.leaq(arrayPtr1, new AMD64Address(arrayPtr1, vectorCount << arrayIndexScale1.log2)); + asm.leaq(arrayPtr2, new AMD64Address(arrayPtr2, vectorCount << arrayIndexScale2.log2)); + asm.movq(tmp1, -vectorCount); + asm.align(crb.target.wordSize * 2); + asm.bind(loopBegin); + emitVectorLoad1(asm, tmpVectors[0], arrayPtr1, tmp1, 0, vSize); + emitVectorLoad2(asm, tmpVectors[1], arrayPtr2, tmp1, 0, vSize); + emitVectorLoad1(asm, tmpVectors[2], arrayPtr1, tmp1, scaleDisplacement1(bytesPerVector), vSize); + emitVectorLoad2(asm, tmpVectors[3], arrayPtr2, tmp1, scaleDisplacement2(bytesPerVector), vSize); + emitVectorXor(asm, tmpVectors[0], tmpVectors[1], vSize); + emitVectorXor(asm, tmpVectors[2], tmpVectors[3], vSize); + emitVectorTest(asm, tmpVectors[0], vSize); asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - } - AMD64Assembler.VexRMOp.VPTEST.emit(asm, AVXKind.AVXSize.XMM, tmpVectors[0], tmpVectors[0]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - } else if (bytesPerVector >= 32) { - // AVX2 supported, use YMM vectors - assert asm.supports(CPUFeature.AVX2); - int loopCount = nBytes / (bytesPerVector * 2); - int rest = nBytes % (bytesPerVector * 2); - if (loopCount > 0) { - if (0 < rest && rest < bytesPerVector) { - loopCount--; - } - if (loopCount > 0) { - if (loopCount > 1) { - asm.movl(tmp1, loopCount); - } - Label loopBegin = new Label(); - asm.bind(loopBegin); - asm.vmovdqu(tmpVectors[0], new AMD64Address(arrayPtr1)); - asm.vmovdqu(tmpVectors[1], new AMD64Address(arrayPtr2)); - asm.vmovdqu(tmpVectors[2], new AMD64Address(arrayPtr1, bytesPerVector)); - asm.vmovdqu(tmpVectors[3], new AMD64Address(arrayPtr2, bytesPerVector)); - asm.vpxor(tmpVectors[0], tmpVectors[0], tmpVectors[1]); - asm.vpxor(tmpVectors[2], tmpVectors[2], tmpVectors[3]); - asm.vptest(tmpVectors[0], tmpVectors[0]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - asm.vptest(tmpVectors[2], tmpVectors[2]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - asm.addq(arrayPtr1, bytesPerVector * 2); - asm.addq(arrayPtr2, bytesPerVector * 2); - if (loopCount > 1) { - asm.decrementl(tmp1); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, loopBegin); - } - } - if (0 < rest && rest < bytesPerVector) { - asm.vmovdqu(tmpVectors[0], new AMD64Address(arrayPtr1)); - asm.vmovdqu(tmpVectors[1], new AMD64Address(arrayPtr2)); - asm.vmovdqu(tmpVectors[2], new AMD64Address(arrayPtr1, bytesPerVector)); - asm.vmovdqu(tmpVectors[3], new AMD64Address(arrayPtr2, bytesPerVector)); - asm.vpxor(tmpVectors[0], tmpVectors[0], tmpVectors[1]); - asm.vpxor(tmpVectors[2], tmpVectors[2], tmpVectors[3]); - asm.vptest(tmpVectors[0], tmpVectors[0]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - asm.vptest(tmpVectors[2], tmpVectors[2]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - asm.vmovdqu(tmpVectors[0], new AMD64Address(arrayPtr1, bytesPerVector + rest)); - asm.vmovdqu(tmpVectors[1], new AMD64Address(arrayPtr2, bytesPerVector + rest)); - asm.vpxor(tmpVectors[0], tmpVectors[0], tmpVectors[1]); - asm.vptest(tmpVectors[0], tmpVectors[0]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - } - } - if (rest >= bytesPerVector) { - asm.vmovdqu(tmpVectors[0], new AMD64Address(arrayPtr1)); - asm.vmovdqu(tmpVectors[1], new AMD64Address(arrayPtr2)); - asm.vpxor(tmpVectors[0], tmpVectors[0], tmpVectors[1]); - if (rest > bytesPerVector) { - asm.vmovdqu(tmpVectors[2], new AMD64Address(arrayPtr1, rest - bytesPerVector)); - asm.vmovdqu(tmpVectors[3], new AMD64Address(arrayPtr2, rest - bytesPerVector)); - asm.vpxor(tmpVectors[2], tmpVectors[2], tmpVectors[3]); - asm.vptest(tmpVectors[2], tmpVectors[2]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - } - asm.vptest(tmpVectors[0], tmpVectors[0]); + emitVectorTest(asm, tmpVectors[2], vSize); asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); + asm.addq(tmp1, elementsPerVectorLoop); + asm.jcc(AMD64Assembler.ConditionFlag.NotZero, loopBegin); } - } else { - // on AVX or SSE, use XMM vectors - int loopCount = nBytes / (bytesPerVector * 2); - int rest = nBytes % (bytesPerVector * 2); - if (loopCount > 0) { - if (0 < rest && rest < bytesPerVector) { - loopCount--; - } - if (loopCount > 0) { - if (loopCount > 1) { - asm.movl(tmp1, loopCount); - } - Label loopBegin = new Label(); - asm.bind(loopBegin); - asm.movdqu(tmpVectors[0], new AMD64Address(arrayPtr1)); - asm.movdqu(tmpVectors[1], new AMD64Address(arrayPtr2)); - asm.movdqu(tmpVectors[2], new AMD64Address(arrayPtr1, bytesPerVector)); - asm.movdqu(tmpVectors[3], new AMD64Address(arrayPtr2, bytesPerVector)); - asm.pxor(tmpVectors[0], tmpVectors[1]); - asm.pxor(tmpVectors[2], tmpVectors[3]); - asm.ptest(tmpVectors[0], tmpVectors[0]); + if (tailCount > 0) { + emitVectorLoad1(asm, tmpVectors[0], arrayPtr1, (tailCount << arrayIndexScale1.log2) - scaleDisplacement1(bytesPerVector), vSize); + emitVectorLoad2(asm, tmpVectors[1], arrayPtr2, (tailCount << arrayIndexScale2.log2) - scaleDisplacement2(bytesPerVector), vSize); + emitVectorXor(asm, tmpVectors[0], tmpVectors[1], vSize); + if (tailCount > elementsPerVector) { + emitVectorLoad1(asm, tmpVectors[2], arrayPtr1, 0, vSize); + emitVectorLoad2(asm, tmpVectors[3], arrayPtr2, 0, vSize); + emitVectorXor(asm, tmpVectors[2], tmpVectors[3], vSize); + emitVectorTest(asm, tmpVectors[2], vSize); asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - asm.ptest(tmpVectors[2], tmpVectors[2]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - asm.addq(arrayPtr1, bytesPerVector * 2); - asm.addq(arrayPtr2, bytesPerVector * 2); - if (loopCount > 1) { - asm.decrementl(tmp1); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, loopBegin); - } } - if (0 < rest && rest < bytesPerVector) { - asm.movdqu(tmpVectors[0], new AMD64Address(arrayPtr1)); - asm.movdqu(tmpVectors[1], new AMD64Address(arrayPtr2)); - asm.movdqu(tmpVectors[2], new AMD64Address(arrayPtr1, bytesPerVector)); - asm.movdqu(tmpVectors[3], new AMD64Address(arrayPtr2, bytesPerVector)); - asm.pxor(tmpVectors[0], tmpVectors[1]); - asm.pxor(tmpVectors[2], tmpVectors[3]); - asm.ptest(tmpVectors[0], tmpVectors[0]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - asm.ptest(tmpVectors[2], tmpVectors[2]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - asm.movdqu(tmpVectors[0], new AMD64Address(arrayPtr1, bytesPerVector + rest)); - asm.movdqu(tmpVectors[1], new AMD64Address(arrayPtr2, bytesPerVector + rest)); - asm.pxor(tmpVectors[0], tmpVectors[1]); - asm.ptest(tmpVectors[0], tmpVectors[0]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - } - } - if (rest >= bytesPerVector) { - asm.movdqu(tmpVectors[0], new AMD64Address(arrayPtr1)); - asm.movdqu(tmpVectors[1], new AMD64Address(arrayPtr2)); - asm.pxor(tmpVectors[0], tmpVectors[1]); - if (rest > bytesPerVector) { - asm.movdqu(tmpVectors[2], new AMD64Address(arrayPtr1, rest - bytesPerVector)); - asm.movdqu(tmpVectors[3], new AMD64Address(arrayPtr2, rest - bytesPerVector)); - asm.pxor(tmpVectors[2], tmpVectors[3]); - asm.ptest(tmpVectors[2], tmpVectors[2]); - asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); - } - asm.ptest(tmpVectors[0], tmpVectors[0]); + emitVectorTest(asm, tmpVectors[0], vSize); asm.jcc(AMD64Assembler.ConditionFlag.NotZero, noMatch); } } } - private static void emitMovBytes(AMD64MacroAssembler asm, Register dst, AMD64Address src, int size) { + private void emitMovBytes(AMD64MacroAssembler asm, Register dst, AMD64Address src, int size) { switch (size) { case 1: - asm.movzbl(dst, src); + if (signExtend) { + asm.movsbq(dst, src); + } else { + asm.movzbq(dst, src); + } break; case 2: - asm.movzwl(dst, src); + if (signExtend) { + asm.movswq(dst, src); + } else { + asm.movzwq(dst, src); + } break; case 4: - asm.movl(dst, src); + if (signExtend) { + asm.movslq(dst, src); + } else { + // there is no movzlq + asm.movl(dst, src); + } break; case 8: asm.movq(dst, src);