10% runtime regression in rnflow Polyhedron benchmark

[Dominique Dhumieres]

Hi Uros,

> Recent patch introduced 10% runtime regression on x86_64 targets in
> rnflow Polyhedron benchmark [1]. Did somebody alread bisected to the
> offending patch?

I see it since some time. It is on my TODO list to open a new PR.
You can suppress the slowdown with -fno-tree-loop-if-convert.

I think I have already done the bissection, but right now I am
chasing another problem and I'll need a couple hours to get to
it.

Cheers,

Dominique