Why tableswitch requires more than constant-time?

Question

http://jroller.com/page/allex?entry=why_tableswitch_requires_more_than
Everyone knows that there is effective implementation of switch like this:
public void switching( int l )
switch( l )
case 0: m0(); break;
case 1: m1(); break;
case 2: m2(); break;
case 3: m3(); break;
case 4: m4(); break;
case 5: m5(); break;
}using indexing in array of offsets. So, switch is better that such ifs:
if( l == 0 ) m0(); else
if( l == 1 ) m1(); else
if( l == 2 ) m2(); else
if( l == 3 ) m3(); else
if( l == 4 ) m4(); else
if( l == 5 ) m5(); elseI'm testing JVM performance and found that such switch requires linear instead of constant time!
Server JVM executes switch exactly the same time as ifs. Client JVM executes switch x1.5 faster,
but execution time depends of a number of cases.
If you're interested, look at my test - I'm wrong or it's really JVM issue?
http://jroller.com/resources/a/allex/SwitchTest.java
http://jroller.com/resources/a/allex/SwitchTest.class - precompiled
http://jroller.com/resources/a/allex/SwitchTest.jad - decompiled with Jasper to be sure tableswitch is used. It is.
I've tried Sun JDK 1.4.2_10 and 1.5.0_06 32-bit on Windows XP x64 AMD Athlon 3800+ 64 x2.

Answer 1

Yes, this is a JVM issue, but it wasn't addressed until recently.
In the Mustang release, the server compiler will generate code for
constant-time execution of switch statements that are of sufficient
size and density. (Offhand, I think at least 16 cases are needed.)
If the switch is not large or dense enough, decision-tree style
code is generated.
You should try a recent Mustang release; I think you will be pleased.
In 5.0, a prototype of this contstant-time code was released on Sparc
only.
On all other 5.0 platforms, and in 1.4.2 and earlier, a decision-tree was
always produced for switch statements.
Mustang client compiler code for switches has not changed.