Re: [PATCH] Fix PR47691: always run scev_const_prop before graphite

[Sebastian Pop <sebpop at gmail dot com>]

On Tue, Jul 26, 2011 at 10:43, Richard Guenther <rguenther@suse.de> wrote:
> Required? ?Do you mean you generate wrong code in graphite if it is
> not run?

Graphite won't generate wrong code if scev_cprop is not run: it will
fail on an assert in the code generation part.  We used to know the
scev for the induction variable (in the scop detection), and after the
detection of reductions we just decided that the IV was a reduction
(its value is used after the loop in the loop close phi), so we decide
to rewrite the IV using arrays to expose this "reduction" to the
graphite representation.  Finally when we arrive in the code gen,
and asking again for the scev of the IV, we get a "dont_know", and
so we are puzzled and abort.

> What if it just fails to handle things or does not handle
> them because things are too expensive?
>
> Please make graphite more robust instead.

Ok, in this case, what about setting gloog_error and stopping the code
generation instead of failing on this gcc_assert.

Thanks for your insistence ;-)

Sebastian