https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117645
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |ra
--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to John David Anglin from comment #6)
> I believe this is a LRA bug.
>
> The problem is the following insn generated by the LRA pass
>
> (insn 181 178 59 9 (clobber (reg:TI 3 %r3 [111])) -1
> (nil))
>
> in the reload RTL dump. This causes the initialization of reg:TI 3
> to be deleted by DCE.
yes that definitely looks like a bug, I wonder how LRA decided to add that
clobber.
>
> The current implementation of addti3 on hppa only allows register
> operands. The problem is avoided if we improve addti3 to support
> the addition of small integers. The code needed to do this is
> similar to that used for adddi3 on 32-bit targets. Have a patch
> which I will install soon.
I assume this works around the issue right? And reverting the backend patch
will expose the issue again, right? I might take a few minutes to debug where
inside LRA, the clobber is coming from.
