On Tue, 11 Apr 2006, chefren wrote: > On 04/11/06 11:39, Otto Moerbeek wrote: > > On Thu, 6 Apr 2006, chefren wrote: > > > On 04/05/06 02:07, Andrew Pinski wrote: > > > > > > > Actually I bet ntohs16 is violating C aliasing rules. > > > > > > Interesting, how do you figure that? > > > > The following diff (backported from gcc PR > > <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10692>) seems to fix the > > problem. I hardly know i386 assembly, so please check if the produced > > code is correct. > > Cool! We patched & compiled gcc and inspected the new assembly output, it > indeed fixes the bug. > > The bug is an over-enthusiastic optimization in gcc, which removes a "store > temporary value to stack" without full justification. > > How did you manage to find that Bugzilla PR based on this problem description? > It does not seem trivial!
I knew debian gcc 3.3.5 doesn't contain the bug. So I diffed debian's gcc against our gcc, and went through the changelog. That gave me a list of candidate diffs; I ordered the list based on a gut feeling. Then I worked trough the diffs to test them. I was lucky, the second diff I tried was the right one. Only later I saw the diff was marked "m68k". > > > If I see things correctly, this fix is not in either 3.3.5 or 3.3.6. > > Debian -stable backported it into their 3.3.5 gcc. > > Yes. > > The Bugzilla PR indicates that someone applies the fix to gcc 3.4 and beyond, > the "Debian GCC maintainers" suggest applying it to gcc 3.3, but the story > ends there. > > The fix was never applied to the gcc 3.3 branch. > > (See line 6936 in > http://gcc.gnu.org/viewcvs/branches/gcc-3_3-branch/gcc/reload1.c?view=markup ) > > ---- > We notice that the upcoming OpenBSD 3.9 still uses gcc 3.3.5. > > Is the gcc 3.3 branch still a supported product for the gcc people, and this a > fix that "slipped through the cracks" in their usual maintenance process? I have no idea if the 3.3 branch is actively maintained by the gcc people. > Or does OpenBSD 3.9 use a "slightly long-in-the-tooth" gcc? In that case, how > is the process of back porting all bug fixes made to gcc 3.4 and onward > arranged for OpenBSD? We use 3.3.5, because it suits us. I believe that the 3.4 branch has some changes that make it difficult to include our local changes like propolice. Also, 3.4 is even slower and more memory hungry than 3.3. As for bug fixes, we tend to do that on a "as needed". > > Now the funny thing is that according to the PR, this is a mk68k > > specific bug, although the fix is not in target specific code. > > Yes, the originator hit it with mk68k, we hit it with i686, but we see nothing > obvious about the bug that makes clear that it can't be hit with i386 for > example. > > Thanks for enlightening us. I think the bug is misclassified by the gcc people. I'm currently building on several platforms to see if there are any regressions. -Otto
