[Bug tree-optimization/62217] [4.9/5 Regression] DOM confuses complete unrolling which in turn causes VRP to warn

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62217

--- Comment #8 from rguenther at suse dot de <rguenther at suse dot de> ---
On Fri, 13 Feb 2015, law at redhat dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62217
> 
> --- Comment #7 from Jeffrey A. Law <law at redhat dot com> ---
> But replacement with the most dominating name (presumably a default def
> dominates everything) isn't going to help here.
> 
> In many ways we'd be better off if we didn't propagate from those equality
> comparisons -- unless they allowed some other later simplification.  But we
> don't have a good way to make that determination.

That's true.  We can also always choose the most immediate dominating
definition for the reason that it might carry more information.

But that might be a loss in some cases as well.

>  Which ultimately let to the
> uncprop pass which we run very late to try and put things back the way they
> were.
>
> I wonder if running uncprop between DOM1 & the late unroller would be worth the
> extra pass.  And if so, where does it make the most sense.

uncprop is mostly to help out-of-SSA PHI coalescing so it's sth different
(and should better be integrated with the out-of-SSA machinery)