------- Comment #3 from law at redhat dot com 2007-03-12 20:06 -------
Subject: Re: [4.1/4.2/4.3 Regression] ICE with
computed goto and constants
On Mon, 2007-03-12 at 19:45 +0000, janis at gcc dot gnu dot org wrote:
>
> ------- Comment #2 from janis at gcc dot gnu dot org 2007-03-12 19:45 -------
> A regression hunt on powerpc-linux using the submitter's test case identified
> this patch:
>
> http://gcc.gnu.org/viewcvs?view=rev&rev=96084
>
> r96084 | law | 2005-03-08 03:39:19 +0000 (Tue, 08 Mar 2005)
Thanks for alerting me to this problem. I think the right fix is going
to simply be to enforce a rule that we can only optimize a computed goto
if the argument collapses down to a local LABEL_REF rather than a
generic invariant.
In the case where we have a constant or non-local LABEL_REF, the CFG
(before optimizing) ought to be conservatively correct(*). Optimizing
is impossible because we don't know which outgoing edge to keep and
which ones to throw away.
This ought to be a 1-2 line fix.
Jeff
(*) If the argument referred to a constant address in the current
function which does not correspond to any known addressable LABEL_REF
then we're hosed as the original unoptimized CFG is probably bogus.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30984
