Re: bitfield and gcc

I wonder if somewhere in Postgres source "we" are relying on the GCC
"correct behaviour" regarding the read-modify-write of bitfield in
structures.

Take a read at this https://lwn.net/Articles/478657/

sorry if this was already mentioned.

Regards
Gaetano Mendola

On Sat, Feb 11, 2012 at 01:54, Gaetano Mendola <mendola(at)gmail(dot)com> wrote:
> I wonder if somewhere in Postgres source "we" are relying on the GCC
> "correct behaviour" regarding the read-modify-write of bitfield in
> structures.

Probably not. I'm pretty sure that we don't have any bitfields, since
not all compilers are happy with them. And it looks like this behavior
doesn't affect other kinds of struct fields.

It sounds like the GCC guys are saying that it's theoretically
possible that the compiler will generate 64-bit read-modify-writes
regardless of the struct member types. In this light, PostgreSQL code
is not "correct" -- our slock_t uses a char type on i386/AMD64/SPARC
and 32-bit int on IA-64/PPC64. There are plenty of places where it's
adjacent to other small fields.

However, I don't think the latter is a problem with any compilers in
practice, as that would break a lot more code than just btrfs and
Postgres.

Regards,
Marti

On Mon, Feb 13, 2012 at 5:38 AM, Marti Raudsepp <marti(at)juffo(dot)org> wrote:
> On Sat, Feb 11, 2012 at 01:54, Gaetano Mendola <mendola(at)gmail(dot)com> wrote:
>> I wonder if somewhere in Postgres source "we" are relying on the GCC
>> "correct behaviour" regarding the read-modify-write of bitfield in
>> structures.
>
> Probably not. I'm pretty sure that we don't have any bitfields, since
> not all compilers are happy with them. And it looks like this behavior
> doesn't affect other kinds of struct fields.

We do, actually: see spgist_private.h, itemid.h, regis.h, spell.h, and
ts_type.h. And maybe some others.

I'm not aware, however, of any cases where we put a lock in the same
structure as a bitfield, so I think we might be OK in that regard.
But the bit about <64-bit spinlocks next to other stuff is a bit
alarming. I continue to be astonished at the degree to which the gcc
developers seem not to care about the POLA. Padding out all of our
spinlocks to 64 bits would not be free: it would cost us significantly
in memory usage, if nothing else. I understand that it's not possible
to modify individual bits in a bitfield atomically, but generating a
64-bit-wide read-modify-write when the underlying base type is 4 bytes
or less is almost pure evil, IMHO.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company