[Bug tree-optimization/114151] [14 Regression] weird and inefficient codegen and addressing modes since g:a0b1798042d033fd2cc2c806afbb77875dd2909b

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114151

--- Comment #4 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 28 Feb 2024, tnfchris at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114151
> 
> --- Comment #3 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
> > 
> > This was a correctness fix btw, so I'm not sure we can easily recover - we
> > could try using niter information for CHREC_VARIABLE but then there's
> > variable niter here so I don't see a chance.
> > 
> 
> It looks like it's mostly SVE suffering here. Adv. SIMD an scalar codegen seems
> to have vastly improved with it. we see ~10% improvements due to better
> addressing for scalar.
> 
> It also only happens at -O3 but -O2 is fine, which is weird as I was expected
> IVopts to be enabled at -O2 too.

Note the underlying issue, analyzing { a, +, b } * c differently and
thus eventually dependent CHRECs failing to analyze should be independent
on the architecture.

What was definitely missing is consideration of POLY_INT_CSTs (and
variable polys, as I think there's no range info for those).

We do eventually want to improve how ranger behaves here.  I'm not sure
why when we do not provide a context 'stmt' it can't see to compute
a range valid at the SSA names point of definition?  (so basically
compute the global range)

Maybe there's another API to do that.  But I thought it was convenient
to use range_of_expr as that also handles GENERIC expression trees
to some extent and those are common within SCEV.