Re: 2 x86-64 ABI bugs in gcc 3.3 and 3.4

[Gabriel Dos Reis <gdr at integrable-solutions dot net>]

Mark Mitchell <mark@codesourcery.com> writes:

| Gerald Pfeifer wrote:
| 
| >Still, I am concerned about ABI breakage on release branches, especially
| >as 3.3.x is already used as a system compiler in several GNU/Linux distros
| >and FreeBSD.
| >
| Actually, you were right the first time: Gaby approved this before I
| did.  I, too, would like to avoid seeing changes go into 3.3.x without
| prior approval for 3.4.x; the oscillation back and forth that you
| point worries me a lot.

There are cases where you don't say anything -- probably because
you're busy or forget about it -- and I have to make decisions because
time is pressing. The cases Gerald pointed in the past are

   (1) regressions from very early versions of GCC.

   (2) they are present in both 3.3.x, 3.4.x (and sometime in
       mainline). 

   (3) whether they are fixed in 3.3.x not, they are going to be
       regressions in 3.4.x.  So the only question is where we
       put the regression point.

   (4) Given the above, and given the submission of patches that cure
       the problem, I believe it would be a mistake not to accept the
       patch in the name that it would introduce a regression.
       That won't be a regression, because the regression is _ealredy_
       present _before_ the problem got fixed in 3.3.x.

I'd like Gerald consider the above points and trust that I did make
analysis before approved patches he objected to in the past.  I think
in those cases, "regression" is being applied too literally without
considering the actual situation (where the regression was already
present). 

Granted, this particular patch of HJL is different.  But I did look at
the situation.

| 
| As for the patch itself, the ABI changes involved here are for truly
| obscure cases.

As I explained in another message, I was going *not* to approve the
patch no matter what you decide for 3.4.x.  Then I *did look* at the
actual PRs. 

| I thought aobut the situation quite a bit before
| approving the patch, and decided that it was probably better to fix
| this problem soon, so that 3.4.1 would be compatible with 3.5.x, even

That is what I thought we're going to do, given the actual PRs, your
past experience with ABI testing, debates and late take for GCC.

| though there are cases where the existing ABI would change between
| point releases.
| 
| In general, however, I completely agree; we should not change the ABI
| between minor releases.

Yes, in general, I think we all agree on that.

| It's possible that I made the wrong decision here.  If the people most
| involved with AMD64 feel that it would be best to back these patches
| out, that's fine by me.

The same applies for gcc-3_3-branch.

-- Gaby