--- Comment #18 from amacleod at redhat dot com 2008-09-18 14:00 ---
Subject: Bug 37102
Author: amacleod
Date: Thu Sep 18 13:58:55 2008
New Revision: 140455
URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=140455
Log:
fix PR 37102 by having out of ssa remove dead PHI nodes.
Adde
--- Comment #17 from amacleod at redhat dot com 2008-09-17 21:55 ---
Created an attachment (id=16348)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16348&action=view)
Patch submitted
Doh. The last patch for the dead code removal had a disabling check in it that
I was using for som
--- Comment #16 from amacleod at redhat dot com 2008-09-17 14:48 ---
Created an attachment (id=16345)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16345&action=view)
Potential patch #2
This is the other option, eliminate dead PHIs. compiling all of GCC .c files
at -O3, it finds
--- Comment #15 from amacleod at redhat dot com 2008-09-17 14:39 ---
Created an attachment (id=16343)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16343&action=view)
potential patch 1
This is the first option which simply doesn't add the result of a PHI with no
uses to the partit
--- Comment #14 from rguenth at gcc dot gnu dot org 2008-09-17 14:38
---
Hm, it doesn't work - bootstrap fails as gengtype segfaults.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37102
--- Comment #13 from amacleod at redhat dot com 2008-09-17 14:34 ---
I was in the middle of updating this PR and taking possesion :-)
Upon further reflection, I don't think it is acceptable for out-of-ssa to
generate incorrect code simply because an optimization wasn't run before it.
S
--- Comment #12 from rguenth at gcc dot gnu dot org 2008-09-17 14:25
---
Ok, so I think we should be fine if we ignore PHI nodes with zero-use results
during building the elimination graph - chained unused PHIs will have lifeness
computed for all but the PHI node with the zero-use resul
--- Comment #11 from lthode at mail dot unomaha dot edu 2008-09-15 01:03
---
(In reply to comment #9)
> Thanks for the explanation, for the branch I would recommend an extra DCE
> pass right before pass_del_ssa. On the trunk we need to make sure to run
> this at -O0 as well. Note that
--- Comment #10 from amacleod at redhat dot com 2008-09-04 16:51 ---
As long as it removes any dead PHIs, it would be sufficient. Other types of
dead statements don't have 'unexpected' side effects across basic block
boundaries, and should be handled fine. Other than being a waste of eff
--- Comment #9 from rguenth at gcc dot gnu dot org 2008-09-04 16:17 ---
Thanks for the explanation, for the branch I would recommend an extra DCE
pass right before pass_del_ssa. On the trunk we need to make sure to run
this at -O0 as well. Note that the simple DCE can leave dead statem
--- Comment #8 from amacleod at redhat dot com 2008-09-04 16:09 ---
out of ssa generally expects that there is no dead code.
I think the original logic was that you never want to generate code for dead
statements, so DCE would be run just before out of ssa.
It assumes any conflicts ca
--- Comment #7 from rguenth at gcc dot gnu dot org 2008-09-04 14:53 ---
As on the trunk we are now feeding out-of-SSA with random dead statements
at -O0 we should look into this. Or schedule a DCE pass before it if it
cannot cope with dead statements.
--
rguenth at gcc dot gnu dot o
12 matches
Mail list logo