[Bug rtl-optimization/103006] [9/10/11/12 Regression] wrong code at -O1 or -O2 on x86_64-linux-gnu by r7-7101

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103006

--- Comment #10 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 2 Nov 2021, jakub at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103006
> 
> Jakub Jelinek <jakub at gcc dot gnu.org> changed:
> 
>            What    |Removed                     |Added
> ----------------------------------------------------------------------------
>            Keywords|ra                          |
> 
> --- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> We don't have that many unrolling passes though, and perhaps we could use
> compute_live_vars and decide based on that.  Though I guess the addresses of
> the vars could be even then hoisted before such loops, though it is unclear why
> it would be done, it can't be VN because the vars don't exist outside of the
> loop.
> Say LIM can do that though...

Yes, LIM can do it as can loop header copying or jump threading that
happens to peel an iteration.  Also there's no dataflow barrier that
prevents addresses from being moved so that "no pass does this" isn't
a good excuse.

That said, I think the fix is to the code computing stack slot reuse.