/
githubmirror
/
galahad
Обзор
Документация
Войти
/
githubmirror
/
galahad
Код
Запросы
0
Пакеты
0
Релизы
0
Аналитика
Безопасность
master
src/java.base/share/classes/jdk/internal/foreign/SegmentBulkOperations.java
388 строк
19 KB
Igor Rudenko
8370822: Improve performance for MemorySegment::fill
03 ноя 2025, 17:17
03 ноя 2025, 17:17
8dbefc5
Код
Авторство
О чём код?
/* * Copyright (c) 2024, 2025, Oracle and/or its affiliates. All rights reserved. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * * This code is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License version 2 only, as * published by the Free Software Foundation. Oracle designates this * particular file as subject to the "Classpath" exception as provided * by Oracle in the LICENSE file that accompanied this code. * * This code is distributed in the hope that it will be useful, but WITHOUT * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License * version 2 for more details (a copy is included in the LICENSE file that * accompanied this code). * * You should have received a copy of the GNU General Public License version * 2 along with this work; if not, write to the Free Software Foundation, * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. * * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA * or visit www.oracle.com if you need additional information or have any * questions. */ package jdk.internal.foreign; import jdk.internal.misc.ScopedMemoryAccess; import jdk.internal.util.Architecture; import jdk.internal.util.ArraysSupport; import jdk.internal.vm.annotation.ForceInline; import jdk.internal.vm.annotation.Stable; import java.lang.foreign.MemorySegment; /** * This class contains optimized bulk operation methods that operate on one or several * memory segments. * <p> * Generally, the methods attempt to work with as-large-as-possible units of memory at * a time. * <p> * It should be noted that when invoking scoped memory access get/set operations, it * is imperative from a performance perspective to convey the sharp types from the * call site in order for the compiler to pick the correct Unsafe access variant. */ public final class SegmentBulkOperations { private SegmentBulkOperations() {} private static final ScopedMemoryAccess SCOPED_MEMORY_ACCESS = ScopedMemoryAccess.getScopedMemoryAccess(); private static final long LONG_MASK = ~7L; // The last three bits are zero private static final long BYTE_REPLICATOR = 0x0101010101010101L; // All the threshold values below MUST be a power of two and should preferably be // greater or equal to 2^3. private static final int NATIVE_THRESHOLD_FILL = powerOfPropertyOr("fill", 5); private static final int NATIVE_THRESHOLD_MISMATCH = powerOfPropertyOr("mismatch", 6); private static final int NATIVE_THRESHOLD_COPY = powerOfPropertyOr("copy", 6); @ForceInline public static MemorySegment fill(AbstractMemorySegmentImpl dst, byte value) { dst.checkReadOnly(false); if (dst.length == 0) { // Implicit state check dst.sessionImpl().checkValidState(); } else if (dst.length < NATIVE_THRESHOLD_FILL) { // 0 <= length < FILL_NATIVE_LIMIT : 0...0X...XXXX // Handle smaller segments directly without transitioning to native code final long u = Byte.toUnsignedLong(value); final long longValue = BYTE_REPLICATOR * u; int offset = 0; // 0...0X...X000 final int limit = (int) (dst.length & (NATIVE_THRESHOLD_FILL - 8)); for (; offset < limit; offset += Long.BYTES) { SCOPED_MEMORY_ACCESS.putLongUnaligned(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + offset, longValue, !Architecture.isLittleEndian()); } int remaining = (int) dst.length - limit; // 0...0X00 if (remaining >= Integer.BYTES) { SCOPED_MEMORY_ACCESS.putIntUnaligned(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + offset, (int) longValue, !Architecture.isLittleEndian()); offset += Integer.BYTES; remaining -= Integer.BYTES; } // 0...00X0 if (remaining >= Short.BYTES) { SCOPED_MEMORY_ACCESS.putShortUnaligned(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + offset, (short) longValue, !Architecture.isLittleEndian()); offset += Short.BYTES; remaining -= Short.BYTES; } // 0...000X if (remaining == 1) { SCOPED_MEMORY_ACCESS.putByte(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + offset, value); } // We have now fully handled 0...0X...XXXX } else { // Handle larger segments via native calls SCOPED_MEMORY_ACCESS.setMemory(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset(), dst.length, value); } return dst; } @ForceInline public static void copy(AbstractMemorySegmentImpl src, long srcOffset, AbstractMemorySegmentImpl dst, long dstOffset, long size) { Utils.checkNonNegativeIndex(size, "size"); // Implicit null check for src and dst src.checkAccess(srcOffset, size, true); dst.checkAccess(dstOffset, size, false); if (size <= 0) { // Do nothing } else if (size < NATIVE_THRESHOLD_COPY && !src.overlaps(dst)) { // 0 < size < FILL_NATIVE_LIMIT : 0...0X...XXXX // // Strictly, we could check for !src.asSlice(srcOffset, size).overlaps(dst.asSlice(dstOffset, size) but // this is a bit slower and it likely very unusual there is any difference in the outcome. Also, if there // is an overlap, we could tolerate one particular direction of overlap (but not the other). // 0...0X...X000 final int limit = (int) (size & (NATIVE_THRESHOLD_COPY - Long.BYTES)); int offset = 0; for (; offset < limit; offset += Long.BYTES) { final long v = SCOPED_MEMORY_ACCESS.getLongUnaligned(src.sessionImpl(), src.unsafeGetBase(), src.unsafeGetOffset() + srcOffset + offset, !Architecture.isLittleEndian()); SCOPED_MEMORY_ACCESS.putLongUnaligned(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + dstOffset + offset, v, !Architecture.isLittleEndian()); } int remaining = (int) size - offset; // 0...0X00 if (remaining >= Integer.BYTES) { final int v = SCOPED_MEMORY_ACCESS.getIntUnaligned(src.sessionImpl(), src.unsafeGetBase(),src.unsafeGetOffset() + srcOffset + offset, !Architecture.isLittleEndian()); SCOPED_MEMORY_ACCESS.putIntUnaligned(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + dstOffset + offset, v, !Architecture.isLittleEndian()); offset += Integer.BYTES; remaining -= Integer.BYTES; } // 0...00X0 if (remaining >= Short.BYTES) { final short v = SCOPED_MEMORY_ACCESS.getShortUnaligned(src.sessionImpl(), src.unsafeGetBase(), src.unsafeGetOffset() + srcOffset + offset, !Architecture.isLittleEndian()); SCOPED_MEMORY_ACCESS.putShortUnaligned(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + dstOffset + offset, v, !Architecture.isLittleEndian()); offset += Short.BYTES; remaining -= Short.BYTES; } // 0...000X if (remaining == 1) { final byte v = SCOPED_MEMORY_ACCESS.getByte(src.sessionImpl(), src.unsafeGetBase(), src.unsafeGetOffset() + srcOffset + offset); SCOPED_MEMORY_ACCESS.putByte(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + dstOffset + offset, v); } // We have now fully handled 0...0X...XXXX } else { // For larger sizes, the transition to native code pays off SCOPED_MEMORY_ACCESS.copyMemory(src.sessionImpl(), dst.sessionImpl(), src.unsafeGetBase(), src.unsafeGetOffset() + srcOffset, dst.unsafeGetBase(), dst.unsafeGetOffset() + dstOffset, size); } } private static final @Stable int[] POWERS_OF_31 = new int[]{ 0x0000001f, 0x000003c1, 0x0000745f, 0x000e1781, 0x01b4d89f, 0x34e63b41, 0x67e12cdf, 0x94446f01}; /** * {@return a 32-bit hash value calculated from the content in the provided * {@code segment} between the provided offsets} * <p> * The method is implemented as a 32-bit polynomial hash function equivalent to: * {@snippet lang=java : * final long length = toOffset - fromOffset; * segment.checkBounds(fromOffset, length); * int result = 1; * for (long i = fromOffset; i < toOffset; i++) { * result = 31 * result + segment.get(JAVA_BYTE, i); * } * return result; * } * but is potentially more performant. * * @param segment from which a content hash should be computed * @param fromOffset starting offset (inclusive) in the segment * @param toOffset ending offset (non-inclusive) in the segment * @throws WrongThreadException if this method is called from a thread {@code T}, * such that {@code srcSegment.isAccessibleBy(T) == false} * @throws IllegalStateException if the {@linkplain MemorySegment#scope() scope} * associated with {@code segment} is not * {@linkplain MemorySegment.Scope#isAlive() alive} * @throws IndexOutOfBoundsException if either {@code fromOffset} or {@code toOffset} * are {@code > segment.byteSize} * @throws IndexOutOfBoundsException if either {@code fromOffset} or {@code toOffset} * are {@code < 0} * @throws IndexOutOfBoundsException if {@code toOffset - fromOffset} is {@code < 0} */ @ForceInline public static int contentHash(AbstractMemorySegmentImpl segment, long fromOffset, long toOffset) { final long length = toOffset - fromOffset; segment.checkBounds(fromOffset, length); if (length == 0) { // The state has to be checked explicitly for zero-length segments segment.scope.checkValidState(); return 1; } int result = 1; final long longBytes = length & LONG_MASK; final long limit = fromOffset + longBytes; for (; fromOffset < limit; fromOffset += Long.BYTES) { long val = SCOPED_MEMORY_ACCESS.getLongUnaligned(segment.sessionImpl(), segment.unsafeGetBase(), segment.unsafeGetOffset() + fromOffset, !Architecture.isLittleEndian()); result = result * POWERS_OF_31[7] + ((byte) (val >>> 56)) * POWERS_OF_31[6] + ((byte) (val >>> 48)) * POWERS_OF_31[5] + ((byte) (val >>> 40)) * POWERS_OF_31[4] + ((byte) (val >>> 32)) * POWERS_OF_31[3] + ((byte) (val >>> 24)) * POWERS_OF_31[2] + ((byte) (val >>> 16)) * POWERS_OF_31[1] + ((byte) (val >>> 8)) * POWERS_OF_31[0] + ((byte) val); } int remaining = (int) (length - longBytes); // 0...0X00 if (remaining >= Integer.BYTES) { int val = SCOPED_MEMORY_ACCESS.getIntUnaligned(segment.sessionImpl(), segment.unsafeGetBase(), segment.unsafeGetOffset() + fromOffset, !Architecture.isLittleEndian()); result = result * POWERS_OF_31[3] + ((byte) (val >>> 24)) * POWERS_OF_31[2] + ((byte) (val >>> 16)) * POWERS_OF_31[1] + ((byte) (val >>> 8)) * POWERS_OF_31[0] + ((byte) val); fromOffset += Integer.BYTES; remaining -= Integer.BYTES; } // 0...00X0 if (remaining >= Short.BYTES) { short val = SCOPED_MEMORY_ACCESS.getShortUnaligned(segment.sessionImpl(), segment.unsafeGetBase(), segment.unsafeGetOffset() + fromOffset, !Architecture.isLittleEndian()); result = result * POWERS_OF_31[1] + ((byte) (val >>> 8)) * POWERS_OF_31[0] + ((byte) val); fromOffset += Short.BYTES; remaining -= Short.BYTES; } // 0...000X if (remaining == 1) { byte val = SCOPED_MEMORY_ACCESS.getByte(segment.sessionImpl(), segment.unsafeGetBase(), segment.unsafeGetOffset() + fromOffset); result = result * POWERS_OF_31[0] + val; } return result; } @ForceInline public static long mismatch(AbstractMemorySegmentImpl src, long srcFromOffset, long srcToOffset, AbstractMemorySegmentImpl dst, long dstFromOffset, long dstToOffset) { final long srcBytes = srcToOffset - srcFromOffset; final long dstBytes = dstToOffset - dstFromOffset; src.checkAccess(srcFromOffset, srcBytes, true); dst.checkAccess(dstFromOffset, dstBytes, true); final long length = Math.min(srcBytes, dstBytes); final boolean srcAndDstBytesDiffer = srcBytes != dstBytes; if (length == 0) { return srcAndDstBytesDiffer ? 0 : -1; } else if (length < NATIVE_THRESHOLD_MISMATCH) { return mismatch(src, srcFromOffset, dst, dstFromOffset, 0, (int) length, srcAndDstBytesDiffer); } else { long i; if (SCOPED_MEMORY_ACCESS.getByte(src.sessionImpl(), src.unsafeGetBase(), src.unsafeGetOffset() + srcFromOffset) != SCOPED_MEMORY_ACCESS.getByte(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + dstFromOffset)) { return 0; } i = vectorizedMismatchLargeForBytes(src.sessionImpl(), dst.sessionImpl(), src.unsafeGetBase(), src.unsafeGetOffset() + srcFromOffset, dst.unsafeGetBase(), dst.unsafeGetOffset() + dstFromOffset, length); if (i >= 0) { return i; } final long remaining = ~i; assert remaining < 8 : "remaining greater than 7: " + remaining; i = length - remaining; return mismatch(src, srcFromOffset + i, dst, dstFromOffset + i, i, (int) remaining, srcAndDstBytesDiffer); } } // Mismatch is handled in chunks of 64 (unroll of eight 8s), 8, 4, 2, and 1 byte(s). @ForceInline private static long mismatch(AbstractMemorySegmentImpl src, long srcFromOffset, AbstractMemorySegmentImpl dst, long dstFromOffset, long start, int length, boolean srcAndDstBytesDiffer) { int offset = 0; final int limit = length & (NATIVE_THRESHOLD_MISMATCH - 8); for (; offset < limit; offset += Long.BYTES) { final long s = SCOPED_MEMORY_ACCESS.getLongUnaligned(src.sessionImpl(), src.unsafeGetBase(), src.unsafeGetOffset() + srcFromOffset + offset, false); final long d = SCOPED_MEMORY_ACCESS.getLongUnaligned(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + dstFromOffset + offset, false); if (s != d) { return start + offset + mismatch(s, d); } } int remaining = length - offset; // 0...0X00 if (remaining >= Integer.BYTES) { final int s = SCOPED_MEMORY_ACCESS.getIntUnaligned(src.sessionImpl(), src.unsafeGetBase(), src.unsafeGetOffset() + srcFromOffset + offset, false); final int d = SCOPED_MEMORY_ACCESS.getIntUnaligned(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + dstFromOffset + offset, false); if (s != d) { return start + offset + mismatch(s, d); } offset += Integer.BYTES; remaining -= Integer.BYTES; } // 0...00X0 if (remaining >= Short.BYTES) { final short s = SCOPED_MEMORY_ACCESS.getShortUnaligned(src.sessionImpl(), src.unsafeGetBase(), src.unsafeGetOffset() + srcFromOffset + offset, false); final short d = SCOPED_MEMORY_ACCESS.getShortUnaligned(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + dstFromOffset + offset, false); if (s != d) { return start + offset + mismatch(s, d); } offset += Short.BYTES; remaining -= Short.BYTES; } // 0...000X if (remaining == 1) { final byte s = SCOPED_MEMORY_ACCESS.getByte(src.sessionImpl(), src.unsafeGetBase(), src.unsafeGetOffset() + srcFromOffset + offset); final byte d = SCOPED_MEMORY_ACCESS.getByte(dst.sessionImpl(), dst.unsafeGetBase(), dst.unsafeGetOffset() + dstFromOffset + offset); if (s != d) { return start + offset; } } return srcAndDstBytesDiffer ? (start + length) : -1; // We have now fully handled 0...0X...XXXX } @ForceInline private static int mismatch(long first, long second) { final long x = first ^ second; return Long.numberOfTrailingZeros(x) / 8; } @ForceInline private static int mismatch(int first, int second) { final int x = first ^ second; return Integer.numberOfTrailingZeros(x) / 8; } @ForceInline private static int mismatch(short first, short second) { return ((0xff & first) == (0xff & second)) ? 1 : 0; } /** * Mismatch over long lengths. */ private static long vectorizedMismatchLargeForBytes(MemorySessionImpl aSession, MemorySessionImpl bSession, Object a, long aOffset, Object b, long bOffset, long length) { long off = 0; long remaining = length; int i, size; boolean lastSubRange = false; while (remaining > 7 && !lastSubRange) { if (remaining > Integer.MAX_VALUE) { size = Integer.MAX_VALUE; } else { size = (int) remaining; lastSubRange = true; } i = SCOPED_MEMORY_ACCESS.vectorizedMismatch(aSession, bSession, a, aOffset + off, b, bOffset + off, size, ArraysSupport.LOG2_ARRAY_BYTE_INDEX_SCALE); if (i >= 0) return off + i; i = size - ~i; off += i; remaining -= i; } return ~remaining; } static final String PROPERTY_PATH = "java.lang.foreign.native.threshold.power."; // The returned value is in the interval [0, 2^30] static int powerOfPropertyOr(String name, int defaultPower) { final int power = Integer.getInteger(PROPERTY_PATH + name, defaultPower); return 1 << Math.clamp(power, 0, Integer.SIZE - 2); } }