[fedora-java] Using the natively-built Eclipse compiler

[Andrew Haley]

Bryce McKinlay writes:
 > Ziga Mahkovec wrote:
 > 
 > >Ah, much better, thanks.  Here are the revised times (BTW, that's a
 > >1.5GHz Pentium M):
 > >
 > >HelloWorld
 > >
 > >ecj                |  ecj-native         |  jikes
 > >-------------------------------------------------------------
 > >real    0m1.863s   |  real    0m1.614s   |  real    0m0.067s
 > >user    0m1.758s   |  user    0m1.536s   |  user    0m0.050s
 > >sys     0m0.103s   |  sys     0m0.076s   |  sys     0m0.012s
 > >
 > >
 > >GNU Classpath (cd lib; make)
 > >
 > >ecj                |  ecj-native         |  jikes
 > >-------------------------------------------------------------
 > >real    1m24.539s  |  real    0m24.552s  |  real    0m9.439s
 > >user    1m23.157s  |  user    0m23.047s  |  user    0m7.486s
 > >sys     0m1.142s   |  sys     0m1.139s   |  sys     0m0.771s
 > >
 > >
 > >Note that classpath sources need a tiny hack to keep ecj from crashing.
 > >But that looks like an upstream Eclipse bug, since I could reproduce it
 > >running ecj with java-1.5.0-sun.
 > >
 > >Andrew, I don't suppose you still need oprofile data (or are there other
 > >places where gcjlib:// loading might be a problem)?

It's not so very important now, but it is still interesting.

By the way, if you're not used to oprofile: the docs make it look
heinously difficult to use, but it's really very easy!  The only hard
part is reading the docs...

I do
  sudo sh -c 'opcontrol --reset ; opcontrol --start'; ./a.out trash flibber poo; sudo sh -c 'opcontrol --stop'

to measure a job, then

  sudo opreport -l

to get the report.  It's really that easy.

The trouble with very short runtimes like this one is that you might
not have enough samples.

 > The ecj-native times still look somewhat slow. Back in the RHUG days, a 
 > gcj-compiled ecj was faster than jikes at building classpath. Perhaps 
 > the BC-ABI stuff is slowing us down a bit.

Yeah.  Oprofile might still be interesting.

Andrew.