This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: calculate_global_regs_live
> > I tried looking at callers of update_life_info. It looks like most of
> > them already pass all blocks when extent != UPDATE_LIFE_LOCAL.
> > Exceptions:
> > cfgrtl.c purge_all_dead_edges
> > recog.c split_all_insns
> > sched-rgn.c schedule_insns
>
> Yes, I did the same search.
>
> > Since the majority already do this, I suspect the run time performance
> > won't be impacted too much if we assume this.
>
> I don't know (but of course I don't have your experience). There is perhaps
> a trade-off here, because infinite loops in c_g_r_l appear to be very rare.
> And we could make them even rarer by clearing the BBs in 'blocks'.
>
> > There is a particularly interesting bit in flow.c in the function
> > update_life_info_in_dirty_blocks. If called with a subset of blocks,
> > and extent is not UPDATE_LIFE_LOCAL, then it already adds in all
> > blocks. There is a comment that says this was needed to fix a pentium4
> > bootstrap failure. So again, it seems someone ran into a related
> > problem. And not surprisingly, it was Richard Henderson again. See
> > http://gcc.gnu.org/ml/gcc-patches/2002-05/msg02253.html
>
> Oh! yes, the very same problem. Could you lend me your glasses (or your eyes
> if you prefer)? :-)
>
> > This message refers to suggestions from Jan to solve the problem you are
> > looking into. Maybe you can find them on the mailing list archive? Or
> > maybe Jan remembers?
>
> Jan, do you have any recollection about suggestions to guarantee that
> calculate_global_regs_live from flow.c will always terminate?
>
> The problem is described here:
> http://gcc.gnu.org/ml/gcc/2003-12/msg00698.html
> It was also reported here:
> http://gcc.gnu.org/ml/gcc/2001-02/msg00476.html
> And Richard hinted at your suggestions here:
> http://gcc.gnu.org/ml/gcc-patches/2002-05/msg02253.html
It is kind of feature to limit work needed to update liveness, but it is
broken as it ineed don't have to converge. In the past we always
somehow papered around.
There is patch by Zdenek to ensure finitarity of the propagation that I
think is best compromise sollution for this problem unless we want to go
for DJ-graphs based liveness computation that is much more involved and
can't embedde dead code removal as we do right now.
http://gcc.gnu.org/ml/gcc-patches/2003-01/msg02059.html
Honza
>
> --
> Eric Botcazou