profiling report fixes between 6.8.3 and 6.10?
marlowsd at gmail.com
Tue Oct 21 10:40:10 EDT 2008
Claus Reinke wrote:
> Forgot to ask earlier, when I still had the offending profiles:-(
> Have there been any profiling report fixes after 6.8.3? I had a case
> where 6.8.3 -Pa profiling lied to me. An SCC kept being reported as
> doing substantial allocations, even after I reduced it to no more than
> readArray (in ST). On looking closer into the nested listing, it turned
> out that the SCC was given 0 entries, in spite of high time and
> alloc, and listed under a parent that never even called it directly.
> At which point I switched to a ghc head, and got more realistic
> profile reports. If that is due to a bug fix, it would be worthwhile
> mentioning it in the release notes; if there have been no bug fixes
> in the area, I should probably try to recreate the situation.
There haven't been any fixes in this area that I'm aware of, you've
probably just been lucky. There are known bugs with those symptoms, and
we've been holding off fixing them because nobody really understands how
cost-centre stacks are supposed to work. Fortunately profiling often gives
reasonable-looking results, but just occasionally odd things happen.
More information about the Cvs-ghc