Re: benchmark result

[Andrew Haley <aph at redhat dot com>]

Mathieu Lacage writes:
 > On Mon, 2004-12-06 at 13:56 +0000, Andrew Haley wrote:
 > > We could improve the performance of these collection classes with
 > > little work.  What is required is someone to study the code, do some
 > > profiling, and fix things.
 > 
 > Ok, I did a bit of that. I used the attached trivial benchmark with
 > 500000 iterations. On my pentium4, with 
 > "/opt/gcc-4.0.0/bin/gcj -gdwarf-2 -O3 -mtune=pentium4
 > MainArrayBench.java --main=MainArrayBench -o bench"

Thanks.

 > I reliably get:
 > 
 > [mlacage@chronos home]$ time -p ./bench 5000000
 > sum: 1642668640
 > real 2.62
 > user 2.30
 > sys 0.22
 > 
 > Then, I modified slightly ArrayList to include its own implementation of
 > Iterator which accesses the data member directly without doing through
 > the ArrayList getter but which does all the checking implemented in
 > AbstractList.Iterator (basically, copy/paste AbstractList.Iterator and
 > replace get(pos++) with data[pos++]. A sample of that is also attached
 > to this email). I reliably got:
 > 
 > [mlacage@chronos home]$ time -p ./bench 5000000
 > sum: 1642668640
 > real 2.53
 > user 2.16
 > sys 0.24
 > 
 > I doubt you could get much better results and keep the specified
 > semantics of the iterators. Well, it should be possible to inline the
 > iterator methods directly in the iterator caller but that would involve
 > knowledge of the iterator implementation in the iterator clients which
 > is something I cannot imagine possible in the current gcc codebase.
 > Maybe the linker could help... 

It's not exactly impossible, but it is tricky.  I wonder if the
iterator and all its methods can be final, but still conform to the
spec?  That would, at least, allow all the checkMod() calls to be
inlined.

Andrew.