Re: [PR42897, VTA] fix vectorization of non-PHI DEF used outside loop

[Richard Guenther <richard dot guenther at gmail dot com>]

On Tue, Feb 2, 2010 at 10:12 AM, Alexandre Oliva <aoliva@redhat.com> wrote:
> I have come up with two patches for this PR. ?Each one by itself would
> be enough to fix this particular instance of the problem, but together
> they retain the most debug information and avoid similar problems that
> might arise in other circumstances.
>
> The first patch arranges for us to reset debug stmts outside a loop that
> refer to a vectorized DEF inside the loop, if there isn't any non-debug
> use outside the loop, even if the DEF is regarded as relevant for the
> vectorizer. ?We used to do this only for vectorizer-irrelevant DEFs,
> which proved to be insufficient.
>
> The second patch avoids the propagation of PHI nodes when they're
> replaced with a non-PHI DEF. ?The patch doesn't actually check that the
> replacement is taking place at the right spot, but I expect all
> non-permanent removals of PHI DEFs to do this.
>
> Regstrapped on x86_64-linux-gnu. ?Ok to install?

Ok.

I'm a little confused by the 2nd patch - we usually remove
PHI nodes permanently by calling remove_phi_node (gsi, true),
but that will end up calling gsi_remove (gsi, false) - and thus
will not insert debug temps (well, it will, but only through
the call of release_ssa_name from remove_phi_node).

So, why not do

  if (gimple_code (stmt) != GIMPLE_PHI)
    insert_debug_temps_for_defs (i);

in gsi_remove and instead do whatever makes sense in
remove_phi_node?

Thanks,
Richard.