[gnugo-devel] oprofile

[Arend Bayer]

Paul wrote:

> Btw, why did you use some `oprofile' (what is that at all?)  I thought
> `-pg' was meant to create profiling information suitable for `gprof'.
> Or is it some general format?  In any case, `gprof' gives you call

oprofile is a potentially very useful set of userspace tools and a
kernel module for linux on x86. Other than -pg and gprof it can generate
system wide profiling (including all libriaries and kernel itself), and
does not need recompilation of the software. It also has much lower
overhead (they claim <= 5%). It can generate callgraphs, too, but for
some reason not on my hardware. Overall, it is still in beta stage, I
think. (They claim alpha, but seems to be overly modest.)

The more important benefit, available on Pentium 4 and later or roughly
equivalent Athlon processors, is that it can use the processor's
performance counters and, e.g., generate a listing of L2 cache misses
per source code line. Since bad performance of a function on a modern
processor is, so my understanding, usually caused by either cache misses
or bad branch prediction, it might be pretty useful for generating
a bunch of 2% speedup-patches.

See oprofile.sf.net.

Btw, while talking about the callgraph features of gprof, I assume you
know that the "% time" in the callgraph is pretty meaningless for gnugo?
See BUGS in man gprof.

Arend