Re: [PATCH] Implement new hook for max_align_t_align

[Carlos O'Donell <carlos at redhat dot com>]

On 10/12/2016 10:17 AM, John David Anglin wrote:
> On 2016-10-12 9:48 AM, Jason Merrill wrote:
>> On Wed, Oct 12, 2016 at 4:02 AM, Jakub Jelinek <jakub@redhat.com> wrote:
>>> On Wed, Oct 12, 2016 at 09:52:04AM +0200, Florian Weimer wrote:
>>>> dropping the alignment means that the padding before the lock member
>>>> vanishes.  Consequently, we have just created a silent ABI change in
>>>> application code, which is a big no-no.
>>> Sure, it would be an ABI change, but how many users would it affect?
>>>
>>>> Since this is PA-RISC, which is essentially dead (neither HPE nor Debian
>>>> ship it anymore), I stand by my suggestion to bump the fundamental alignment
>>> Or just drop support for a dead arch?
>>>
>>>> instead.  Sure, it is a bit inefficient, but this will only affect PA-RISC
>>>> users.  It does not even cause work for PA-RISC porters. Conversely, if we
>>>> work on this to come up with a different fix, many more people will be
>>>> affected (because they don't get all the nice things we could work on
>>>> instead), and we may need to maintain a special GCC kludge for the
>>>> alternative solution, impacting GCC developers in particular.
>>> But sure, bumping malloc alignment is probably easiest.  And people who want
>>> performance have better options than to stay on 32-bit PA-RISC anyway.
>> Or we could do nothing and tell people to ignore the harmless warning.
> The warning is an issue because of -Werror.  However, it appears easy to suppress it in the PA
> backend.  I have a patch that I'm testing.
> 
> We are discussing offline regarding the glibc issue.  It easy to bump the alignment of malloc
> but I take Jakub's point and maybe we should break the ABI.  Debian unstable churns
> quickly, and I think we would be better off being consistent with the current max_align_t
> and 8-byte aligned malloc.

I am against breaking the ABI.

I would rather see us bump malloc alignment up to 16-bytes.

The last time I changed this alignment it _immediately_ broken libstdc++ boostrap
because it's using exactly the kind of embedded pthread_mutext_t we're talking about
breaking.

So in that case the debian builds in unstable broke right away, and I had to revert
the change. We'd have to BinNMU a bunch of things to get this working again.

Again I think our two options are, in my order of preference:

- Disable the warning via a PA backend change.
- Bump malloc alignment.

I am sensitive to the first change being something that carries with it extra
maintenance burden, so I'm happy to see the second solution chosen if that's what
everyone wants (Florian's suggestion).

-- 
Cheers,
Carlos.