[Bug tree-optimization/65206] Vectorized version of loop is removed.

["rguenther at suse dot de" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65206

--- Comment #9 from rguenther at suse dot de <rguenther at suse dot de> ---
On Mon, 2 Jan 2017, jakub at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65206
> 
> --- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> Adding some flag on the MASK_STORE or MASK_LOAD is not hard, it can be in
> another argument, or some GF_*, whatever.  But I don't understand here what is
> the difference between originally pointer based and array based access. 
> Doesn't for non-masked stores or loads forwprop propagate array accesses into
> pointer stores/loads anyway?

The issue is that dependence analysis can't handle mixed pointer / array
accesses (it would need to treat the array access as pointer-based,
thus comment #5 which talks about always providing both sets of
DR->indices variants).

Yes, forwprop propagates the loads/stores but it correctly conservatively
handles propagating p = &a[i] to *p (it doesn't propagate) or
p = &a[1] to *p (it creates MEM[a, 4], not an array access).  The "hack"
forwards it as array access to make dependence analysis happy.

The real fix is comment #5 computing (maybe lazily) a DR->indices
variant for array accesses using get_inner_reference on the
full reference and analyzing it as MEM[base + offset, bitoff],
thus if it were *&ref.

That makes more refs "compatible" for initialize_data_dependence_relation.

Implementation-wise it's not that easy to implement this though :/