一位面试官最近问了我这样一个问题:给定三个布尔变量a、b和c，如果三个变量中至少有两个为真，则返回true。
我的解决方案如下:
boolean atLeastTwo(boolean a, boolean b, boolean c) {
if ((a && b) || (b && c) || (a && c)) {
return true;
}
else{
return false;
}
}
他说这还可以进一步改进，但如何改进呢?
当前回答
在C:
return !!a + !!b + !!c >= 2;
其他回答
One thing I haven't seen others point out is that a standard thing to do in the "please write me some code" section of the job interview is to say "Could you improve that?" or "Are you completely happy with that" or "is that as optimized as possible?" when you say you are done. It's possible you heard "how would you improve that" as "this might be improved; how?". In this case changing the if(x) return true; else return false; idiom to just return x is an improvement - but be aware that there are times they just want to see how you react to the question. I have heard that some interviewers will insist there is a flaw in perfect code just to see how you cope with it.
这个怎么样:
(a - b) ? c : a
为什么不逐字执行呢?：）
(a?1:0)+(b?1:0)+(c?1:0) >= 2
在C语言中，你可以写a+b+ C >= 2(或者!!a+!! !b+!! !C >= 2，非常安全)。
为了回应TofuBeer对java字节码的比较，这里有一个简单的性能测试:
class Main
{
static boolean majorityDEAD(boolean a,boolean b,boolean c)
{
return a;
}
static boolean majority1(boolean a,boolean b,boolean c)
{
return a&&b || b&&c || a&&c;
}
static boolean majority2(boolean a,boolean b,boolean c)
{
return a ? b||c : b&&c;
}
static boolean majority3(boolean a,boolean b,boolean c)
{
return a&b | b&c | c&a;
}
static boolean majority4(boolean a,boolean b,boolean c)
{
return (a?1:0)+(b?1:0)+(c?1:0) >= 2;
}
static int loop1(boolean[] data, int i, int sz1, int sz2)
{
int sum = 0;
for(int j=i;j<i+sz1;j++)
{
for(int k=j;k<j+sz2;k++)
{
sum += majority1(data[i], data[j], data[k])?1:0;
sum += majority1(data[i], data[k], data[j])?1:0;
sum += majority1(data[j], data[k], data[i])?1:0;
sum += majority1(data[j], data[i], data[k])?1:0;
sum += majority1(data[k], data[i], data[j])?1:0;
sum += majority1(data[k], data[j], data[i])?1:0;
}
}
return sum;
}
static int loop2(boolean[] data, int i, int sz1, int sz2)
{
int sum = 0;
for(int j=i;j<i+sz1;j++)
{
for(int k=j;k<j+sz2;k++)
{
sum += majority2(data[i], data[j], data[k])?1:0;
sum += majority2(data[i], data[k], data[j])?1:0;
sum += majority2(data[j], data[k], data[i])?1:0;
sum += majority2(data[j], data[i], data[k])?1:0;
sum += majority2(data[k], data[i], data[j])?1:0;
sum += majority2(data[k], data[j], data[i])?1:0;
}
}
return sum;
}
static int loop3(boolean[] data, int i, int sz1, int sz2)
{
int sum = 0;
for(int j=i;j<i+sz1;j++)
{
for(int k=j;k<j+sz2;k++)
{
sum += majority3(data[i], data[j], data[k])?1:0;
sum += majority3(data[i], data[k], data[j])?1:0;
sum += majority3(data[j], data[k], data[i])?1:0;
sum += majority3(data[j], data[i], data[k])?1:0;
sum += majority3(data[k], data[i], data[j])?1:0;
sum += majority3(data[k], data[j], data[i])?1:0;
}
}
return sum;
}
static int loop4(boolean[] data, int i, int sz1, int sz2)
{
int sum = 0;
for(int j=i;j<i+sz1;j++)
{
for(int k=j;k<j+sz2;k++)
{
sum += majority4(data[i], data[j], data[k])?1:0;
sum += majority4(data[i], data[k], data[j])?1:0;
sum += majority4(data[j], data[k], data[i])?1:0;
sum += majority4(data[j], data[i], data[k])?1:0;
sum += majority4(data[k], data[i], data[j])?1:0;
sum += majority4(data[k], data[j], data[i])?1:0;
}
}
return sum;
}
static int loopDEAD(boolean[] data, int i, int sz1, int sz2)
{
int sum = 0;
for(int j=i;j<i+sz1;j++)
{
for(int k=j;k<j+sz2;k++)
{
sum += majorityDEAD(data[i], data[j], data[k])?1:0;
sum += majorityDEAD(data[i], data[k], data[j])?1:0;
sum += majorityDEAD(data[j], data[k], data[i])?1:0;
sum += majorityDEAD(data[j], data[i], data[k])?1:0;
sum += majorityDEAD(data[k], data[i], data[j])?1:0;
sum += majorityDEAD(data[k], data[j], data[i])?1:0;
}
}
return sum;
}
static void work()
{
boolean [] data = new boolean [10000];
java.util.Random r = new java.util.Random(0);
for(int i=0;i<data.length;i++)
data[i] = r.nextInt(2) > 0;
long t0,t1,t2,t3,t4,tDEAD;
int sz1 = 100;
int sz2 = 100;
int sum = 0;
t0 = System.currentTimeMillis();
for(int i=0;i<data.length-sz1-sz2;i++)
sum += loop1(data, i, sz1, sz2);
t1 = System.currentTimeMillis();
for(int i=0;i<data.length-sz1-sz2;i++)
sum += loop2(data, i, sz1, sz2);
t2 = System.currentTimeMillis();
for(int i=0;i<data.length-sz1-sz2;i++)
sum += loop3(data, i, sz1, sz2);
t3 = System.currentTimeMillis();
for(int i=0;i<data.length-sz1-sz2;i++)
sum += loop4(data, i, sz1, sz2);
t4 = System.currentTimeMillis();
for(int i=0;i<data.length-sz1-sz2;i++)
sum += loopDEAD(data, i, sz1, sz2);
tDEAD = System.currentTimeMillis();
System.out.println("a&&b || b&&c || a&&c : " + (t1-t0) + " ms");
System.out.println(" a ? b||c : b&&c : " + (t2-t1) + " ms");
System.out.println(" a&b | b&c | c&a : " + (t3-t2) + " ms");
System.out.println(" a + b + c >= 2 : " + (t4-t3) + " ms");
System.out.println(" DEAD : " + (tDEAD-t4) + " ms");
System.out.println("sum: "+sum);
}
public static void main(String[] args) throws InterruptedException
{
while(true)
{
work();
Thread.sleep(1000);
}
}
}
这将在我的机器上打印以下内容(在Intel Core 2 + sun java 1.6.0_15-b03上运行Ubuntu，带有HotSpot Server VM (14.1-b02，混合模式):
第一次和第二次迭代:
a&&b || b&&c || a&&c : 1740 ms
a ? b||c : b&&c : 1690 ms
a&b | b&c | c&a : 835 ms
a + b + c >= 2 : 348 ms
DEAD : 169 ms
sum: 1472612418
后来迭代:
a&&b || b&&c || a&&c : 1638 ms
a ? b||c : b&&c : 1612 ms
a&b | b&c | c&a : 779 ms
a + b + c >= 2 : 905 ms
DEAD : 221 ms
我想知道，对于(a + b + c >= 2)情况，java虚拟机可以做什么来降低性能。
下面是如果我用-client VM开关运行java会发生什么:
a&&b || b&&c || a&&c : 4034 ms
a ? b||c : b&&c : 2215 ms
a&b | b&c | c&a : 1347 ms
a + b + c >= 2 : 6589 ms
DEAD : 1016 ms
神秘……
如果我在GNU Java解释器中运行它，它会变慢近100倍，但是a&&b || b&&c || a&&c版本胜出。
在运行OS X的最新代码中，豆腐啤酒的结果:
a&&b || b&&c || a&&c : 1358 ms
a ? b||c : b&&c : 1187 ms
a&b | b&c | c&a : 410 ms
a + b + c >= 2 : 602 ms
DEAD : 161 ms
Paul Wagland使用Mac Java 1.6.0_26-b03-383-11A511的结果
a&&b || b&&c || a&&c : 394 ms
a ? b||c : b&&c : 435 ms
a&b | b&c | c&a : 420 ms
a + b + c >= 2 : 640 ms
a ^ b ? c : a : 571 ms
a != b ? c : a : 487 ms
DEAD : 170 ms
我的第一个想法是
return (a||b)&&(b||c)
但为了便于阅读，我喜欢你们提出的a+b+c>=2的解决方案
以下是目前为止的答案:
public class X
{
static boolean a(final boolean a, final boolean b, final boolean c)
{
return ((a && b) || (b && c) || (a && c));
}
static boolean b(final boolean a, final boolean b, final boolean c)
{
return a ? (b || c) : (b && c);
}
static boolean c(final boolean a, final boolean b, final boolean c)
{
return ((a & b) | (b & c) | (c & a));
}
static boolean d(final boolean a, final boolean b, final boolean c)
{
return ((a?1:0)+(b?1:0)+(c?1:0) >= 2);
}
}
并通过反编译器运行它们(javap -c X > results.txt):
Compiled from "X.java"
public class X extends java.lang.Object{
public X();
Code:
0: aload_0
1: invokespecial #1; //Method java/lang/Object."<init>":()V
4: return
static boolean a(boolean, boolean, boolean);
Code:
0: iload_0
1: ifeq 8
4: iload_1
5: ifne 24
8: iload_1
9: ifeq 16
12: iload_2
13: ifne 24
16: iload_0
17: ifeq 28
20: iload_2
21: ifeq 28
24: iconst_1
25: goto 29
28: iconst_0
29: ireturn
static boolean b(boolean, boolean, boolean);
Code:
0: iload_0
1: ifeq 20
4: iload_1
5: ifne 12
8: iload_2
9: ifeq 16
12: iconst_1
13: goto 33
16: iconst_0
17: goto 33
20: iload_1
21: ifeq 32
24: iload_2
25: ifeq 32
28: iconst_1
29: goto 33
32: iconst_0
33: ireturn
static boolean c(boolean, boolean, boolean);
Code:
0: iload_0
1: iload_1
2: iand
3: iload_1
4: iload_2
5: iand
6: ior
7: iload_2
8: iload_0
9: iand
10: ior
11: ireturn
static boolean d(boolean, boolean, boolean);
Code:
0: iload_0
1: ifeq 8
4: iconst_1
5: goto 9
8: iconst_0
9: iload_1
10: ifeq 17
13: iconst_1
14: goto 18
17: iconst_0
18: iadd
19: iload_2
20: ifeq 27
23: iconst_1
24: goto 28
27: iconst_0
28: iadd
29: iconst_2
30: if_icmplt 37
33: iconst_1
34: goto 38
37: iconst_0
38: ireturn
}
你可以看到?:那些比你原来的修复版本稍好。最好的方法是完全避免分支的方法。从更少的指令(在大多数情况下)的角度来看，这很好，对于CPU的分支预测部分来说也更好，因为分支预测中的错误猜测可能会导致CPU失速。
我觉得最有效的是《月影》里的那个。它平均使用最少的指令，减少了CPU中管道停顿的机会。
为了100%确定，您需要找出每条指令的成本(以CPU周期为单位)，不幸的是，这是不容易获得的(您必须查看热点的源代码，然后查看CPU供应商的规格，以确定每条生成指令所花费的时间)。
请参阅Rotsor对代码的运行时分析的更新答案。
