> On Fri, Oct 12, 2012 at 9:43 PM, Jan Hubicka <[email protected]> wrote:
> > Hi,
> > I finally tracked down twolf misoptimization triggered by my loop-unroll.c
> > changes. It has turned out to be semi-latent wrong code issue in webizer.
> > What happens is:
> >
> > 1) gcse.c drop REG_EQUAL note on the induction variable
> > 2) loop optimizer unrolls the loop enabling webizer to cleanup
> > 3) webizer do not track reaching refs into REG_EQUAL note because
> > the variable is dead and renames it to unused pseudo
> > Program is stil valid but the REG_EQUAL is bogus.
>
> What do you mean, "not track"? web.c does track EQ notes via DF this:
>
> web.c:315: df_set_flags (DF_NO_HARD_REGS + DF_EQ_NOTES);
>
> This is bogus. You're papering over the underlying bug of an invalid
> REG_EQUAL note. This patch is not OK!
The REG_EQUAL note is valid until webizer breaks it.
DF_EQ_NOTES makes UD chain problem to look into uses of REG_EQUAL and add all
reaching definitions into the list. To do the reaching definitions it uses RD
problem solution. This solution do not really care about REG_EQUAL notes
because it does only over definitions. It however uses results of LIVENESS
problem via REG_DEAD notes and the pruning trick. Now th REG_EQUAL note uses
dead register (that is aloved), because the liveness is computed ignoring
REG_EQUAL notes and thus the definition gets prunned from the list and
consequentely webizer misses the link.
>
> I did not add DF_RD_PRUNE_DEAD_DEFS for no reason. You're
> re-introducing a major compile time hog. Did you do timings on some
> significant body of code with -fweb enabled?
Well, I can not think how to fix it without
1) computing liveness with REG_EQUAL included prior RD that means a lot
of shuffling of REG_DEAD notes
2) making webizer to drop all REG_EQUAL notes that use dead register
(for that it will need to track liveness too)
Do you have better ideas?
Honza
>
> Ciao!
> Steven
