This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: PATCH: Use edge probabilities in interblock and ebb schedulers
- From: Maxim Kuvyrkov <mkuvyrkov at ispras dot ru>
- To: Vladimir Makarov <vmakarov at redhat dot com>
- Cc: gcc-patches at gcc dot gnu dot org, Peter Steinmetz <steinmtz at us dot ibm dot com>, Richard Henderson <rth at redhat dot com>, Andrey Belevantsev <abel at ispras dot ru>
- Date: Tue, 31 Jan 2006 15:59:46 -0500
- Subject: Re: PATCH: Use edge probabilities in interblock and ebb schedulers
- References: <43DFB970.5010708@ispras.ru> <43DFC1CE.5040406@redhat.com>
Vladimir Makarov wrote:
> Could you be more specific what is wrong with bb->frequency.
It is not consistant, i.e.
int f = 0;
FOR_EACH_EDGE (e, bb->preds) { f += EDGE_FREQUENCY (e); }
gcc_assert (bb->frequency == f);
will fail sometime.
In rare cases it can be greater than BB_FREQ_MAX. Some time ago I've
tried to investigate the cause, that led to some CFG optimization (I
think it is cfgcleanup.c: try_crossjump_to_edge (), but not sure).
>
> Even if you found wrong values of bb->frequency, it should be fixed in
> the first place not in just one place.
>
> So you should use bb->frequency in the patch even if their values are
> wrong and you can not fix them in the 1st place or you have no time for
> this.
The reason to keep frequencies in separate array is to calculate them
independently for each ebb, so we won't end up with many zero
frequencies in case of large CFGs.
BTW, it is not true that ammp shows +11% - it is 0%, actually. Sorry for
this measurement fault.
--
Maxim