Nathan Meyers wrote:
> Nick Lawson wrote:
> >
> > I'm pretty sure try{}catch{} catch blocks add NO overhead to code,
> > unless the exception actually gets thrown. But exceptions are
> > supposed to be
> > exceptional, so who cares how slow it is ?
> > ....
>
> Times for attached program (JDK1.2pre2, glibc2.1, RH6, 200MHZ PPro,
> 500000000 iterations):
>
> No JIT sunwjit tya
> ------ ------- ---
> With try/catch 278793 17925 33894
> Without 237905 15364 35803
>
> %slower 17.2% 16.7% -5.3%
>
> I'd be surprised to see a zero-overhead try/catch, although tya seems to
> be on to something :-).
>
> Results are unaffected by -O compilation.
>
> Nathan
Ok, NO overhead for try..catch is an exaggeration. But only a small one
- typically there is exactly one extra bytecode instruction per try,
to branch round the catch{} clauses. Most C++ compilers produce similar
code,
and it would be easy for an optimizing java compiler to produce code
without even this overhead (although I haven't seen one that does so yet).
The byte-code produced by jdk1.2pre-v2 for the Hello program is reproduced
below.
Suns JDK 1.2 produces exactly the same code (of course!).
The extra instructions for the try..catch are marked with a !
Theres just 3 of them, and only one gets executed if the exception is not
thrown.
I guess the question was whether try..catch is faster or slower
than the alternatives. I think the program below is more
like the sort of situation that occurs in real life.
The times for this (Pentium 166, Sun Windows 95 JDK 1.2, Blackdown 1.2pre-v2
+ glibc 2.1 + OpenLinux 2.2,
500,000,000 iterations):
VM: kaffe Sun Sun Blackdown Blackdown
JIT: symcjit none sunwjit none
----- ------- ---- --------- ----------
With try/catch 388,212 12,360 545,410 136,207 466,300
Without 396,401 15,490 601,000 143,784 508,770
%faster 2% 20% 9.2% 5.2% 8.3%
- the last column is estimated from 50,000,000 iterations.
I didn't believe the Sun VM figures either, but I tried both tests
several times, and the figures are consistent, scale linearly with
the number of iterations, and match wall-clock time. So it looks
like symcjit is pretty smart !
Nick
-----------------------------------------
import java.util.*;
public class Hello2 {
public static void main(String[] argv) throws Exception {
int count = Integer.parseInt(argv[0]);
Boolean foobar = new Boolean(false);
Date date1 = new Date();
for (int i = 0; i < count;) {
try {
if (!foobar.booleanValue()) i++;
}
catch (Exception e) {
System.err.println( "Oops!" );
}
}
Date date2 = new Date();
Date date3 = new Date();
for (int i = 0; i < count;) {
if(foobar != null) {
if (!foobar.booleanValue()) i++;
}
else {
System.err.println( "Oops!" );
}
}
Date date4 = new Date();
System.out.println("First loop (try): " +
(date2.getTime() - date1.getTime()));
System.out.println("Second loop (no try): " +
(date4.getTime() - date3.getTime()));
}
}
-----------------------------------------
Bytecode from Nathans Hello.java:
Code Disassembly:
Addr Op Operands Code
---- -- -------- ----
...............
...............
0011 03 iconst_0 // for( int i = 0; .....
0012 36 04 istore L4 // (store i)
0014 a7 001b goto +1b = 002f
try #1
0017 84 0401 iinc L4,1 // { i++;
001a 1d iload_3
001b 99 000b ifeq +b = 0026 // if( foobar ) ...
001e bb 0005 new #5 <Class Exception>
0021 59 dup
0022 b7 000b invokespecial #11 <Method void Exception.<init>()>
0025 bf athrow // throw ...
0026 1d iload_3 // foobar = foobar &&
true;
0027 3e istore_3
end try #1
! 0028 a7 0007 goto +7 = 002f // branch around catch
catch #1 (Exception)
! 002b 57 pop // start catch clause
! 002c a7 0003 goto +3 = 002f // end catch clause
002f 15 04 iload L4 // }
0031 1b iload_1
0032 a1 ffe5 if_icmplt -1b = 0017 // end for loop 1
...............
...............
0047 03 iconst_0 // for( int i = 0; .....
0048 36 07 istore L7 // (store i)
004a a7 0014 goto +14 = 005e
004d 84 0701 iinc L7,1 // { i++;
0050 1d iload_3
0051 99 000b ifeq +b = 005c // if( foobar ) ...
0054 bb 0005 new #5 <Class Exception>
0057 59 dup
0058 b7 000b invokespecial #11 <Method void Exception.<init>()>
005b bf athrow // throw ...
005c 1d iload_3
005d 3e istore_3 // foobar = foobar &&
true;
005e 15 07 iload L7 // }
0060 1b iload_1
0061 a1 ffec if_icmplt -14 = 004d // end for loop 2
...............
-----------------------------------------
----------------------------------------------------------------------
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
