Re: Minimum gcc and clang supported to generate FreeBSD binaries

On Thu, Jun 20, 2024, 11:43 AM Bakul Shah <bakul@iitbombay.org> wrote:

> On Jun 19, 2024, at 11:49 PM, Warner Losh <imp@bsdimp.com> wrote:
> >
> > Yea. We shouldn't. But it's kinda necessary to have the compilers tested
> all the time to spot regressions. This stuff is fiddly enough with 2 main
> compiles and 2 that kinda emulate these two...  comes a point that you need
> to say enough unless somebody is really, actively using it, our kinda
> support becomes the worst of both worlds: a random drag on development that
> isn't actually useful to anybody.
>
> You should'n't have to test with every compiler if the libraries and
> headers are standard compliant (by default).
>

Except things don't work that way. They are standard compliant (which
standard do you want?). The standard is the BSD API. In that API, we have
two versions of qsort_r: The old, pre-standard and the new standardized
version (at least for the moment). You can say you want a different
standard, (see _POSIX_C_SOURCE or _XOPEN_SOURCE), but that's on you.

And there's many things that are standards conforment in a specific
environment that need extensions. There's little way around that.

FreeBSD's headers go to extreme lengths to try to be compliant with all of
that. It's a tough act. Made tougher by some compilers not implementing
everything that the project needs to meet all these competing needs.

Is there no way of avoiding putting ELF specific bits in such a commonly
> used header file if it is only used for qsort (or a handful of such version
> specific functions)? Also note that somehow linux avoids this (not that
> FreeBSD should do what Linux does but just as a data point).
>

This is not an ELF specific bit, but a toolchain used by the compiler
specific bit. And it's needed for backwards compatibility with old releases.
Linux's system header files are not standards compliant, and often do not
avoid non-standard constructs. The toolchains in that environment know how
to swallow everything that's there, that's all.

Any way, one suggestion is to not use #Ifdef TCC but to use some generic
> name. Alternately, if tcc is the only outlier,


It is the only outlier, but I did it in a generic way.

tl;dr: This stuff is a lot more complex than you'd think going into it. :(

Warner

>