Re: [PATCH 2/3] Have different definitions of __pthread_tsd_main agree in size

[Rich Felker <dalias@aerifal.cx>]

On Mon, Feb 11, 2013 at 03:07:28PM +0100, Jens Gustedt wrote:
> Am Montag, den 11.02.2013, 08:44 -0500 schrieb Rich Felker:
> > The current intended usage is that __environ is used in ISO C
> > functions and startup code that must not reference the name environ
> > (since it's in the namespace reserved for the application), and POSIX
> > and extension functions use environ.
> 
> If this was the intended use, the effective use was not in line with
> it. execv and execvp had it differently.
> 
> > There's no reason the latter
> > _need_ to use environ though; it was just more convenient and less
> > implementation-specific. If you think there's a good reason to change
> > it after reading this, let's discuss it.
> 
> With what I have added in my other mail (and you have snipped :) in
> the current model there is even the danger that environ and __environ
> split into two different objects.
> 
> char **environ;
> 
> int main(void) {
>   return (long)&environ;
> }

This is a conforming C program, but not a conforming POSIX program.
The current code in musl should allow it to work as long as it does
not call functions in the standard library which are not defined by
ISO C. If any POSIX functions are used, the program is non-conforming
and invokes undefined behavior.

With that said, I don't see any reason this behavior is _desirable_ or
beneficial, so I'm fine with changing it and always using __environ
internally. We could even add a visibility attribute on it so that
GOT/PC-relative accesses get used in compilers which support
visibility, instead of the more expensive accesses through the GOT.

Rich