用于内存缩减的Java整数标志和按位操作的性能

Question

使用整数标志和按位操作是减少大容量对象内存占用的有效方法吗

• 内存占用

  我的理解是boolean通常在JVM实现中存储为int。这是正确的吗？在这种情况下，32个标志无疑代表了内存占用的大幅减少

  当然，尽管JVM实现各不相同，但情况并非总是如此

• 性能

  据我所知，CPU是非常数字驱动的，按位运算的效率与计算中的效率差不多

  与布尔运算相比，使用位运算是否会降低性能，甚至增加性能

• 备选方案

  有没有更好的方法来完成同样的事情？枚举是否允许组合标志，即FLAGX = FLAG1 | FLAG2

---

示例代码

注意，最后一个方法propogateMove()是递归的，每秒可能被调用数百次，并且对应用程序的响应性有直接影响，因此使用标志来避免逻辑位和调用其他方法

// FLAGS helper functions
private final void setclear(int mask, boolean set) { if (set) set(mask); else clear(mask); }
private final void set(int mask) { flags |= mask; }
private final void clear(int mask) { flags &= ~mask; }
private final boolean test(int mask) { return ((flags & mask) == mask); }



// Flags //////////////////////////////////////////////////////////////////////

private static final boolean HORIZONTAL          = true;
private static final boolean VERTICAL            = false;

private static final int ORIENT                  = 0x00000001;
private static final int DISPLAY                 = 0x00000002;
private static final int HSHRINK                 = 0x00000004;
private static final int VSHRINK                 = 0x00000008;
private static final int SHRINK = HSHRINK | VSHRINK;

private static final int TILE_IMAGE              = 0x00000010;
private static final int CURSOR                  = 0x00000020;
private static final int MOUSEINSIDE             = 0x00000040;
private static final int MOUSEINSIDE_BLOCKED     = 0x00000080;

private static final int CONSTRAIN               = 0x00000100;
private static final int CONSTRAIN_DESCENDENT    = 0x00000200;
private static final int PLACE                   = 0x00000400;
private static final int PLACE_DESCENDENT        = 0x00000800;
private static final int REFLOW = CONSTRAIN | CONSTRAIN_DESCENDENT | PLACE | PLACE_DESCENDENT;

private static final int PACK                    = 0x00001000;
private static final int CLIP                    = 0x00002000;
private static final int HAS_WIDTH_SLACK         = 0x00004000;
private static final int HAS_HEIGHT_SLACK        = 0x00008000;

private static final int ALIGN_TOP               = 0x00010000;
private static final int ALIGN_BOTTOM            = 0x00020000;
private static final int ALIGN_LEFT              = 0x00040000;
private static final int ALIGN_RIGHT             = 0x00080000;
private static final int ALIGNS = ALIGN_TOP | ALIGN_BOTTOM | ALIGN_LEFT | ALIGN_RIGHT;
private static final int ALIGN_TOPLEFT = ALIGN_TOP | ALIGN_LEFT;
private static final int ALIGN_TOPRIGHT = ALIGN_TOP | ALIGN_RIGHT;
private static final int ALIGN_BOTTOMLEFT = ALIGN_BOTTOM | ALIGN_LEFT;
private static final int ALIGN_BOTTOMRIGHT = ALIGN_BOTTOM | ALIGN_RIGHT;

private static final int ENTER_TRAP              = 0x00100000;
private static final int LEAVE_TRAP              = 0x00200000;
private static final int _MOVE_TRAP              = 0x00400000;
private static final int MOVE_TRAP               = 0x00800000;

private static final int CHILDREN_READ_TRAP      = 0x01000000;
private static final int CHILDREN_TRAP           = 0x02000000;
private static final int PLACE_CLEAN             = 0x03000000;

private static final int SHRINK_TRAP             = 0x04000000;
private static final int HSHRINK_TRAP            = 0x10000000;
private static final int VSHRINK_TRAP            = 0x20000000;

//private static final int UNUSED                = 0x40000000;
//private static final int UNUSED                = 0x80000000;




// Flags in switch ////////////////////////////////////////////////////////////

/** get align value as a string from align flags */
private JS alignToJS() {
    switch(flags & ALIGNS) {
        case (ALIGN_TOPLEFT):
            return SC_align_topleft;
        case (ALIGN_BOTTOMLEFT):
            return SC_align_bottomleft;
        case (ALIGN_TOPRIGHT):
            return SC_align_topright;
        case (ALIGN_BOTTOMRIGHT):
            return SC_align_bottomright;
        case ALIGN_TOP:
            return SC_align_top;
        case ALIGN_BOTTOM:
            return SC_align_bottom;
        case ALIGN_LEFT:
            return SC_align_left;
        case ALIGN_RIGHT:
            return SC_align_right;
        case 0: // CENTER
            return SC_align_center;
        default:
            throw new Error("This should never happen; invalid alignment flags: " + (flags & ALIGNS));
    }
}




// Flags in logic /////////////////////////////////////////////////////////////

private final boolean propagateMove(int mousex, int mousey) throws JSExn {
    // start with pre-event _Move which preceeds Enter/Leave
    if (test(_MOVE_TRAP)) {
        if (Interpreter.CASCADE_PREVENTED == justTriggerTraps(SC__Move, JSU.T)) {
            // _Move cascade prevention induces Leave
            propagateLeave();
            // propagate cascade prevention
            return true;
        }
    }

    // REMARK: anything from here on in is a partial interruption relative
    // to this box so we can not call propagateLeave() directly upon it
    int i;
    boolean interrupted = false;

    if (!test(PACK)) {
        // absolute layout - allows for interruption by overlaying siblings
        for (Box b = getChild(i=treeSize()-1); b != null; b = getChild(--i)) {
            if (!b.test(DISPLAY)) {
                continue;
            }
            if (interrupted) {
                b.propagateLeave();
                continue;
            }
            int b_mx = mousex-getXInParent(b);
            int b_my = mousey-getYInParent(b);
            if (b.inside(b_mx, b_my)) {
                if (b.propagateMove(b_mx, b_my)) {
                    interrupted = true;
                }
            } else {
                b.propagateLeave();
            }
        }
    } else {
        // packed layout - interrupted still applies, plus packedhit shortcut
        boolean packedhit = false;
        for (Box b = getChild(i=treeSize()-1); b != null; b = getChild(--i)) {
            if (!b.test(DISPLAY)) {
                continue;
            }
            if (packedhit) {
                b.propagateLeave();
                continue;
            }
            int b_mx = mousex-getXInParent(b);
            int b_my = mousey-getYInParent(b);
            if (b.inside(b_mx, b_my)) {
                packedhit = true;
                if (b.propagateMove(b_mx, b_my)) {
                    interrupted = true;
                }
            } else {
                b.propagateLeave();
            }
        }
    }

    // child prevented cascade during _Move/Move which blocks
    // Enter on this box - invoking Leave if necessary
    if (interrupted) {
        if (test(MOUSEINSIDE)) {
            if (!test(MOUSEINSIDE_BLOCKED)) {
                // mouse previously inside, now blocked so invoke Leave
                set(MOUSEINSIDE_BLOCKED);
                if (test(LEAVE_TRAP)) {
                    justTriggerTraps(SC_Leave, JSU.T);
                }
            }
        } else {
            // mouse not previously inside, Enter not yet triggered, so
            // do not invoke Leave
            set(MOUSEINSIDE);
            set(MOUSEINSIDE_BLOCKED);
        }
        // propagate cascade prevention 
        return true;
    }

    // set cursor if applicable to this box
    if (test(CURSOR)) {
        Surface s = getSurface();
        if (s!=null && !s.cursorset) {
            s.cursor = JSU.toString(getAndTriggerTraps(SC_cursor));
            s.cursorset = true;
        }
    }

    // fire Enter traps
    if (!test(MOUSEINSIDE)) {
        set(MOUSEINSIDE);
        if (test(ENTER_TRAP)) {
            justTriggerTraps(SC_Enter, JSU.T);
        }
    }

    // finish post-event Move which follows Enter/Leave
    if (test(MOVE_TRAP)) {
        if (Interpreter.CASCADE_PREVENTED == justTriggerTraps(SC_Move, JSU.T)) {
            // propagate cascade prevention
            return true;
        }
    }

    // propagation uninterrupted
    return false;
}

Answer 1

It is my understanding that commonly a boolean is stored as an int in a JVM implementation. Is this correct?

当然，这取决于JVM的实现，但对于主流CPU上的实现可能是如此

In which case surely the 32 flags represent a large memory footprint reduction.

如果一个类中实际上有32个标志，并且该类有大量实例，那么是的。如果您的实例从未超过几百个，那么就不值得担心了

It is my understanding that CPUs are very number driven and bitwise operations are about as efficient as things come in computing.

这是真的

Is there a performance penalty - or even gain - to using bitwise operations over boolean operations?

这还取决于内存使用情况。如果只对少数对象进行非常密集的操作，则按位操作可能会减慢速度。如果有很多对象，由于缓存性能更好，减少的内存可能会大大提高性能

Is there a better way of accomplishing the same thing? Does an Enum allow the combination of flags i.e. FLAGX = FLAG1 | FLAG2?

您可以（并且应该）使用^{}，而不是自己执行逐位操作。是的，如果您可以使用enum和^{}，它会更干净，但是如果您有许多enum，每个都有几个元素，那么由于多个EnumSet实例的开销，它可能不会产生所需的内存节省

Answer 2

Is using an integer flag and bitwise operations an effective way of reducing the memory footprint of high volume Objects?

如果一个对象有大量这样的标志，并且有大量这样的对象（或者减少堆使用是一个关键问题），那么这可能是一个值得进行的微优化。否则没有（海事组织）

这样做的问题是，它使代码更难阅读，也更脆弱。因此，只有当内存使用是一个关键问题时，您才应该考虑它

Answer 3

是的，布尔值存储为32位整数。方法签名区分布尔值和整数，但在其他情况下，它们的处理方式相同。这是JVM字节码规范的一部分

Java上的按位操作直接映射到CPU上执行相同按位操作的单个指令，因此它确实非常快。当然，将每个值保存在其自己的32位字中会更快（这样就根本不需要执行任何逐位操作）。使用您的方法将节省内存并花费更多的CPU周期

Java没有用于组合枚举的运算符，我不认为这有什么意义

Answer 4

因为你不能相信布尔值是以位或整数的形式存储的（我认为后者是实际的实现），是的，我认为位掩码和标志确实提高了性能

我在一些项目中使用过它们。它们的可读性较差，但我从不在意这一点，因为我的观点是每个程序员都应该理解二进制符号和算术

建议：对于java 7，我们可以这样定义数字文本：

   private static final int ENTER_TRAP = 0b00000000000100000000000000000000;

甚至像这样：

   private static final int ENTER_TRAP = 0b0000_0000_0001_0000_0000_0000_0000_0000;