gcc 3.2.3 config/os/bits/i386/atomicity.h broken

[Ralf Corsepius]

Am Die, 2003-04-29 um 01.34 schrieb Loren James Rittle:
> In article <1051331346.28271.9123.camel@mccallum.corsepiu.faw.uni-ulm.de>,
> Ralf Corsepius<corsepiu@faw.uni-ulm.de> writes:
> 
> >> > But I have no opposition to having the lock;xaddl implementation
> >> > selected if the appropriate CPU model cpp predefine is defined.
> 
> >> We can't to this given the current architecture.
> > Why? I think we already do.
> 
> > Given the gcc-3.2.2/3 implementation of atomicity.h works sufficiently,
> > all we need to do is to detect such cases which do not support the xadd
> > instruction. 
> 
> > Then, AFAIS, #ifdef'ing on __tune_i386__ is sufficient. This is what I
> > try to exploit in my rtems-gcc-3.2.2-patch below.
> 
> What you have proposed is known as an ODR violation in C++.  ODR
> covers the internal implementation in addition to the ABI itself.  Had
> this header file only been used in library implementation and not been
> seen directly by application code (and random command line switches
> provided by the user), then we could make a patch as you suggest.

Bummer, bits/atomicity.h is publically used? How that?

> This is the aspect of the current architecture I refer to.
>
> >> No, that is the exact solution that doesn't work.  ;-)  There may be no
> >> #ifdef regions in the header except the entire multiple inclusion
> >> guard.  The reason is that this header is pulled into headers exposed
> >> to users for templating purposes.
> > I don't agree on this (cf. to my patch below). You just must not change
> > the ABI and API if ifdefs get encountered.
> 
> Since there is no locking on your __tune_i386__ path, it will not
> function as required (I think you might be looking at very old
> generic/atomicity.h code). 
I do not understand. The patch addresses gcc-3.2.x and contains parts
from current gcc-3.2-branch, only.

I.e. the #ifdef __tune_i386__ part is an exact copy from
libstdc++-v3/config/cpu/generic/bits/atomicity.h
while the #else part is an exact copy from
libstdc++-v3/config/cpu/i386/bits/atomicity.h

=> If what you say holds and if I understand correctly, the generic
implementation in gcc-3.2.x is broken.

> I much rather prefer the patch that Joel
> posted (this is where we find a spin lock implementation that works on
> i386 and all architectures which claim compatibility with i386/IA32).
How about using an external global c-function (or extern static
C++-function) or pointer to such function instead?

Then such #ifdefs could be used inside of the library could be
transparently hidden from user-code.

Ralf