java编译器对finally的优化

<Inside JVM>是市面上少数几本系统介绍JVM的书籍，(作者Bill Venners的网站artima也是非常有趣的Java技术社区)，这本书最有意思的地方可能就是作者为了演示JVM内部工作原理编写的Applet, 非常值得向大家推荐。<Inside JVM>最新的是第二版，也有些年头了，尽管JVM规范本身的变化并不大，但是java编译器的有些细节处理还是跟书里面不一样了，比如说现在的java编译器一般都会对finally作inline处理。

这个Applet演示的是下面这段程序的bytecode。
代码1:

/*
*copiedfromhttp://www.artima.com/insidejvm/applets/HopAround.html
*/

classClown...{

staticinthopAround()...{
inti=0;
while(true)...{
try...{
try...{
i=1;
}
finally...{//Thefirstfinallyclause
i=2;
}
i=3;
//Thisreturnnevercompletes,becauseof
//thecontinueinthesecondfinallyclause
returni;
}
finally...{//Thesecondfinallyclause
if(i==3)...{
//Thiscontinueoverridesthereturnstatement
continue;
}
}
}
}
}

其主要目的演示的是finally子句的处理， 在Inside JVM和JVM Spec的7.13节当中，finally都是通过jsr和ret指令实现的，按照书中的说法，hopAround()方法将被编译成47个jvm指令，如下所示:

代码2：

/**//*
*copiedfromhttp://www.artima.com/insidejvm/applets/HopAround.html
*/

0iconst_0//Pushconstant0
1istore_0//Popintolocalvar0:inti=0;

//Bothtryblocksstarthere(seeexceptiontable,below):
2iconst_1//Pushconstant1
3istore_0//Popintolocalvar0:i=1;
4jsr18//Jumptomini-subroutineatoffset18(the
//firstfinallyclause)
7goto24//Jumptooffset24(tojustbelowfirst
//finallyclause)

//Catchclauseforthefirstfinallyclause:
10astore4//Popthereferencetothrownexception,store
//inlocalvariable4
12jsr18//Jumptomini-subroutineatoffset18(the
//firstfinallyclause)
15aload4//Pushthereference(tothrownexception)
//fromlocalvariable4
17athrow//Rethrowthesameexception

//Thefirstfinallyclause:
18astore5//Storethereturnaddressinlocalvariable5
20iconst_2//Pushconstant2
21istore_0//Popintolocalvar0:i=2;
22ret5//Jumptoreturnaddressstoredinlocalvariable5

//Bytecodesforthecodejustafterthefirstfinallyclause:
24iconst_3//Pushconstant3
25istore_0//Popintolocalvar0:inti=3;

//Bytecodesforthereturnstatment:
26iload_0//Pushtheintfromlocal
//variable0(i,whichis3)
27istore_1//Popandstoretheintintolocal
//variable1(thereturnvalue,i)
28jsr39//Jumptomini-subroutineatoffset39(the
//secondfinallyclause)
31iload_1//Pushtheintfromlocalvariable1(the
//returnvalue)
32ireturn//Returntheintonthetopofthestack

//Catchclauseforthesecondfinallyclause:
33astore_2//Popthereferencetothrownexception,store
//inlocalvariable2
34jsr39//Jumptomini-subroutineatoffset39(the
//secondfinallyclause)
37aload_2//Pushthereference(tothrownexception)
//fromlocalvariable2
38athrow//Rethrowthesameexception

//Thesecondfinallyclause:
39astore_3//Storethereturnaddressinlocalvariable3
40iload_0//Pushtheintfromlocalvariable0(i)
41iconst_3//Pushconstant3
42if_icmpeq47//Ifthetoptwointsonthestackareequal,jump
//tooffset47:if(i==3){
45ret3//Jumptoreturnaddressstoredinlocalvariable3
47goto2//Jumptooffset2(thetopofthewhile
//block):continue;

但是如果用Sun JDK 1.4之后的编译器编译，这段程序实际上被编译成截然不同的38个jvm指令(如下所示)，其主要区别是finally 子句被inline到了正常代码和exception handling代码，避免了jsr/ret, 从而减少了JSR跳转/pop出栈/ret返回三条指令，在本例中，不但提高了运行效率，还减少了class文件的大小，其代价是finally block之中的内容被复制到了两个地方，如果finally block的指令很多，class文件可能会变大。

代码3：

staticinthopAround();
Code:
0:iconst_0//pushconst0
1:istore_0//popintolocalvar0:i=0

//startbothtryblock
2:iconst_1//pushconst1
3:istore_0//popintoi

//startfirstfinallyblock(itisinlined)
4:iconst_2//pushconst2
5:istore_0//popintolocalvar0(i)
6:goto14//skipexceptionhandling

//startexceptionhandlingforsecondtryblock
9:astore_1//popthrownexceptionreferenceintolocalvar1
//startfirstfinallyblock(inlineagain)
10:iconst_2//pushconst2
11:istore_0//popinto0
12:aload_1//pushthereference(theexception)fromlocalvar1
13:athrow//throwexeption

//goonthefirsttryblock
14:iconst_3//pushconst3
15:istore_0//popintolocalvar0(i)
16:iload_0//pushvaluefromlocalvar0(i)
17:istore_1//popintolocalvar1(returnvalue)

//startsecondfinallyblock(inlined)
18:iload_0//pushvaluefromlocalvar0(i)
19:iconst_3//pushconst3
20:if_icmpne26//ifthetoptwointsonthestackarenotequal,jumpto26
23:goto2//loop

//goonthefirsttryblock
26:iload_1//pushvaluefromlocalvar1(returnvalue)
27:ireturn//returnmethod

//startexceptionhandlingforfirsttryblock
28:astore_2//popthrownexceptionreferenceintolocalvar2
//startsecondfinallyblock(inlinedagain)
29:iload_0//pushvaluefromlocalvar0(i)
30:iconst_3//pushconst3
31:if_icmpne37//ifthetoptwointsonthestackarenotequal,jumpto37
34:goto2//loop
37:aload_2//pushthereference(theexception)fromlocalvar2
38:athrow//throwexception

Exceptiontable:
fromtotargettype
249any
9109any
21828any
282928any

Eclipse编译器显然也将finally inline了，Eclipse 3.2甚至disable了编译器不inline finally的选项，ECJ编译出来的hopAround()由39条指令组成:

代码4：

staticinthopAround();
Code:
0:iconst_0
1:istore_0

//startbothtry
2:iconst_1
3:istore_0

//skipexceptionhandling
4:goto12

//exceptionhandlingforfirsttryblock
7:astore_1
//firstfinallyblockinlined
8:iconst_2
9:istore_0
10:aload_1
11:athrow

//firstfinallyblockinlinedagain
12:iconst_2
13:istore_0

//goonsecondtryblock
14:iconst_3
15:istore_0
16:iload_0
17:istore_3

//secondfinallyblockinlined
18:iload_0
19:iconst_3
20:if_icmpne26
23:goto2

//return
26:iload_3
27:ireturn

//exceptionhandlingforsecondtry
28:astore_2
//secondefinallyinlined
29:iload_0
30:iconst_3
31:if_icmpne37
34:goto2
37:aload_2
38:athrow

//whatdoesthismean?
39:goto2
Exceptiontable:
fromtotargettype
277any
21828any

区别不大，Eclipse编译器将exception handling的代码放在了正常路径的finally block之前，多出来的一句是第39句goto 2, 比较匪夷所思， 因为这一句显然是无法被执行到的。还有一个区别是exception table缩减为两项，少掉的两条是针对exception handling代码自身的，Sun javac加上的这两条意欲何为也有点让人迷惑...也许因为这个原因，ECJ编译出来的class要稍微小一些(对于这个class, 222bytes vs. 235bytes)。

这两种主流Java编译器都对finally做了inline处理，所以我们在平常写java代码的时候应该注意在try/catch/finally里面不要做太多的分支，并且finally block中包含的内容尽量不要太多，否则class文件可能会多占用不少空间。

还有一个tips: javac和Eclipse编译器缺省都带有debug信息，加上-g:none选项之后，class文件要小一半，对于Eclipse编译器尤其明显。

最后一个tips送给有耐心读到这里的人: 请代码3中的14-17行，在进入finally block之前，JVM将把变量i当前的值存入另外一个本地变量(istore_1)，并且在26-27行将此值返回，所以如果在finally block中修改了i的值，对返回值是没有影响的，如果要将修改过的i返回，请在finally中直接return i。更多finally实现的细节请参考Inside JVM:).