[Bug target/108583] [13 Regression] wrong code with vector division by uint16 at -O2

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108583

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Tamar Christina from comment #7)
> (In reply to rsandifo@gcc.gnu.org from comment #6)
> > (In reply to Tamar Christina from comment #3)
> > > The vectorizer has this context but since we didn't want a new IFN the
> > > context should instead be derivable in
> > > targetm.vectorize.can_special_div_by_const hook.
> > I probably got lost in the threading, but could you point to where
> > the idea of using an ifn was rejected?  I know there was pushback
> > against hard-coding a specific constant, but that doesn't prevent
> > the use of ifns in itself.  (E.g. the gather/scatter IFNs have
> > constant arguments that are checked against what the target allows.)
> 
> https://patchwork.sourceware.org/project/gcc/patch/patch-15779-tamar@arm.com/
> #124788 is where the original patch used an IFN.
> https://patchwork.sourceware.org/project/gcc/patch/patch-15779-tamar@arm.com/
> #124863 where Richi wouldn't ACK a new optab as it didn't have a physical
> instruction backing it on any ISA, and in fairness, no one else did either
> effectively stranding the change.

Yep.  The other option would have to expose whatever special instruction
is used in the end as IFN and rewrite the division during vectorization
"properly", either via some pattern or via doing the expansion when
emitting the vectorized division instruction.

In hindsight that might have been the better option, but then ...

> https://patchwork.sourceware.org/project/gcc/patch/patch-15779-tamar@arm.com/
> #125144 where I pinged and got no response to the ping. After which I went
> on IRC and asked Richi how he'd like me to proceed.  In reply to this I was
> instructed he would like to proceed the same way vector permutes are
> currently handled with `can_perm` etc.  and that's where the patch thread
> picks off back on the ML.

... this looked simpler and also reasonable.  Only that it breaks now.