[Bug tree-optimization/98542] Redundant loads in vectorised loop

["rsandifo at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98542

--- Comment #4 from rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> ---
(In reply to rguenther@suse.de from comment #3)
> On Wed, 6 Jan 2021, rsandifo at gcc dot gnu.org wrote:
> 
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98542
> > 
> > --- Comment #2 from rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> ---
> > With the follow-on mentioned (“i” being 32-bits rather than 64) we'd
> > still want to treat the structure access as a single group, even though
> > the fields are different sizes.
> 
> Supporting that is going to be tricky though.  Presumably this only
> works for { int, double }, not for { int, int, double } and accessing
> the 2nd int (or it all depends on endianess?).
It should work for both in the SVE example.  The idea is that the
int field will be loaded into an unpacked vector, with each 32-bit
element being stored in a 64-bit container.  We then need to extend
the int (if the int starts off in the low half) or right-shift the
container (if the int starts off in the high half).  The nice thing
about the gather example is that the extend case can be folded into
the gather itself.

Like you say, the choice between extending and shifting will depend
on which int we need and on the endianness.  All four combinations
should be doable though.