Re: [PATCH] [RFC] vect: Fix infinite loop while determining peeling amount

[Richard Biener <richard.guenther@gmail.com>]

On Mon, Jul 27, 2020 at 4:20 PM Stefan Schulze Frielinghaus
<stefansf@linux.ibm.com> wrote:
>
> On Mon, Jul 27, 2020 at 12:29:11PM +0200, Richard Biener wrote:
> > On Mon, Jul 27, 2020 at 11:45 AM Richard Sandiford
> > <richard.sandiford@arm.com> wrote:
> > >
> > > Richard Biener <richard.guenther@gmail.com> writes:
> > > > On Mon, Jul 27, 2020 at 11:09 AM Richard Sandiford
> > > > <richard.sandiford@arm.com> wrote:
> > > >>
> > > >> Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> > > >> > On Wed, Jul 22, 2020 at 5:18 PM Stefan Schulze Frielinghaus via
> > > >> > Gcc-patches <gcc-patches@gcc.gnu.org> wrote:
> > > >> >>
> > > >> >> This is a follow up to commit 5c9669a0e6c respectively discussion
> > > >> >> https://gcc.gnu.org/pipermail/gcc-patches/2020-June/549132.html
> > > >> >>
> > > >> >> In case that an alignment constraint is less than the size of a
> > > >> >> corresponding scalar type, ensure that we advance at least by one
> > > >> >> iteration.  For example, on s390x we have for a long double an alignment
> > > >> >> constraint of 8 bytes whereas the size is 16 bytes.  Therefore,
> > > >> >> TARGET_ALIGN / DR_SIZE equals zero resulting in an infinite loop which
> > > >> >> can be reproduced by the following MWE:
> > > >> >
> > > >> > But we guard this case with vector_alignment_reachable_p, so we shouldn't
> > > >> > have ended up here and the patch looks bogus.
> > > >>
> > > >> The above sounds like it ought to count as reachable alignment though.
> > > >> If a type requires a lower alignment than its size, then that's even
> > > >> more easily reachable than a type that requires the same alignment as
> > > >> the size.  I guess at one extreme, a target alignment of 1 is always
> > > >> reachable.
> > > >
> > > > Well, if the element alignment is 8 but its size is 16 then when presumably
> > > > the desired vector alignment is a multiple of 16 we can never reach it.
> > > > Isn't this the case here?
> > >
> > > If the desired vector alignment (TARGET_ALIGN) is a multiple of 16 then
> > > TARGET_ALIGN / DR_SIZE will be nonzero and the problem the patch is
> > > fixing wouldn't occur.  I agree that we might never be able to reach
> > > that alignment if the pointer starts out misaligned by 8 bytes.
> > >
> > > But I think that's why it makes sense for the target to only ask
> > > for 8-byte alignment for vectors too, if it can cope with it.  8-byte
> > > alignment should always be achievable if the scalars are ABI-aligned.
> > > And if the target does ask for only 8-byte alignment, TARGET_ALIGN /
> > > DR_SIZE would be zero and the loop would never progress, which is the
> > > problem that the patch is fixing.
> > >
> > > It would even make sense for the target to ask for 1-byte alignment,
> > > if the target doesn't care about alignment at all.
> >
> > Hmm, OK.  Guess I still think we should detect this somewhere upward
> > and avoid this peeling compute at all.  Somehow.
>
> I've been playing around with another solution which works for me by
> changing vector_alignment_reachable_p to return also false if the
> alignment requirements are already satisfied, i.e., by adding:
>
> if (known_alignment_for_access_p (dr_info) && aligned_access_p (dr_info))
>   return false;

That sounds wrong, instead ...

> Though, I'm not entirely sure whether this makes it better or not.
> Strictly speaking if the alignment was reachable before peeling, then
> reaching alignment with peeling is also possible but probably not what
> was intended.  So I guess returning false in this case is sensible.  Any
> comments?

... why is the DR considered for peeling at all?  If it is already
aligned there's
no point to do that.  If we want to align another DR then the loop you fix
should run on that DRs align/size, no?

Richard.

> Thanks,
> Stefan
>
> >
> > Richard.
> >
> > > Thanks,
> > > Richard