This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.

Index Nav:	[Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav:	[Date Prev] [Date Next]	[Thread Prev] [Thread Next]
Other format:	[Raw text]

Re: [lno] Avoid use of rewrite_ssa_into_ssa in loop header copying

From: Daniel Berlin <dberlin at dberlin dot org>
To: Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>
Cc: gcc-patches at gcc dot gnu dot org,amacleod at redhat dot com,Andrew Pinski <pinskia at physics dot uc dot edu>,David Edelsohn <dje at watson dot ibm dot com>
Date: Sat, 17 Jul 2004 21:58:55 -0400
Subject: Re: [lno] Avoid use of rewrite_ssa_into_ssa in loop header copying
References: <20040718004412.GA31495@atrey.karlin.mff.cuni.cz> <5F44356C-D856-11D8-AE5E-000393A6D2F2@physics.uc.edu> <200407180136.i6I1a6D33200@makai.watson.ibm.com> <20040718013847.GA1686@atrey.karlin.mff.cuni.cz>

On Jul 17, 2004, at 9:38 PM, Zdenek Dvorak wrote:

Hello,

I almost think this is a bug in your code rather than PRE.
This is a bug in Zdenek's code.
an you know that how?

Because your change shouldn't change anything. But regardless, it could be a bug in GVN-PRE. Have you compared code before and after to see where "PRE is breaking"? Can you provide me a small testcase? Have you tried it on mainline with a similar code change? There hasn't been a PRE merge in quite a while, since before the VN changes, so it's possible it's a bug that has already been fixed.

Zdenek, please do not commit your patch to the LNO branch.


I already did.  If someone says me where the bug is, I will be happy to
fix it (and since the patch does not cause any problems except with
PRE which nobody on works on on LNO branch, I am not going to revert
it).

This of course, doesn't mean anything at all. Most "bugs in PRE" tend to be bugs elsewhere. At least, that is what history has shown. Not that GVN-PRE doesn't have bugs, it's just they are a well-defined set of failures. For example, the most "common" symptom is an ICE in find_or_generate_expression. This almost always means something has been marked as anticipated that isn't actually available where we think it is anticipated. This is either due to incorrect value numbering, or incorrect ssa given to PRE (the actual phi translation and anticipability computation is rather easy to prove correct, so bugs generally don't turn up there). The value numbering is rather stable now too.

However, regardless of all of this, the fact remains that merging this problem into the lno branch means that we can't rely on lno branch testing before submitting things to the mainline, since they won't include testing with PRE, which is enabled on the mainline!

This is a very bad idea to do when we are in the middle of trying to merge patches from that branch.

Zdenek

Follow-Ups:
- Re: [lno] Avoid use of rewrite_ssa_into_ssa in loop header copying
  - From: Daniel Berlin
- Re: [lno] Avoid use of rewrite_ssa_into_ssa in loop header copying
  - From: Zdenek Dvorak

References:
- [lno] Avoid use of rewrite_ssa_into_ssa in loop header copying
  - From: Zdenek Dvorak
- Re: [lno] Avoid use of rewrite_ssa_into_ssa in loop header copying
  - From: Andrew Pinski
- Re: [lno] Avoid use of rewrite_ssa_into_ssa in loop header copying
  - From: David Edelsohn
- Re: [lno] Avoid use of rewrite_ssa_into_ssa in loop header copying
  - From: Zdenek Dvorak

Index Nav:	[Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav:	[Date Prev] [Date Next]	[Thread Prev] [Thread Next]