[fedora-java] Profiling ecj (was: Using the natively-built Eclipse compiler)

[Andrew Haley]

Ziga Mahkovec writes:
 > I upgraded to libgcj-4.0.0-0.32 and eclipse-ecj-3.1.0_fc-0.M5.12.
 > The results for compiling classpath:
 > 
 > ecj                |  ecj-native         |  jikes
 > -------------------------------------------------------------
 > real    1m31.786s  |  real    0m16.207s  |  real    0m8.423s
 > user    1m22.796s  |  user    0m14.925s  |  user    0m7.052s
 > sys     0m1.369s   |  sys     0m1.020s   |  sys     0m0.755s
 > 
 > (ecj-native went from 24.5s to 16.2s as a result of -O2)
 > 
 > I also uploaded oprofiles for ecj[1], ecj-native[2] and ecj-native-with-
 > the-old-gcjlib-tools-jar[3].
 > 
 > Note that I'm running a Centrino so the oprofile driver falls back to
 > timer interrupt mode.
 > 
 > [1] http://www.bootchart.org/misc/ecj/M5.12/opreport.ecj.log
 > [2] http://www.bootchart.org/misc/ecj/M5.12/opreport.ecj-native.log
 > [3] http://www.bootchart.org/misc/ecj/M5.12/opreport.ecj-native-gcjlib.log

Thank you very much for doing this.  It's very useful.

Jikes is famously fast, so I'm fairly pleased to see we're close.

It says here: "[Jikes] is astoundingly fast -- hundreds of thousands
of lines of code per minute (about ten times faster than other
compilers)"

There are some surprises here.  Is ecj fully interpreted code?
_Jv_InterpMethod::run seems only to take 0.0036% of the time.  I'm
going to investigate some more.

I'm sure that ecj-native isn't purely native code, because of this line:

399       2.2024  libgcj.so.6.0.0          _Jv_BytecodeVerifier::verify_instructions_0()

which is weird.

Thanks again,
Andrew.