[PATCH] D40558: [ELF] - Trigger error when -R <filename> is given.

[Rui Ueyama via llvm-commits]

On Wed, Nov 29, 2017 at 8:59 AM, Rafael Avila de Espindola <
rafael.espindola at gmail.com> wrote:

> George Rimar <grimar at accesssoftek.com> writes:
>
> >>I think we shouldn't even emit an error for the second case. -rpath is
> the information for the dynamic linker, and the static linker should
> blindly store given -rpath >paths to a file header, because the static
> linker doesn't know the actual runtime environment. It is a valid use case
> that a given -rpath is not a directory in a build >environment but a
> directory in a runtime environment.
> >>
> >>Given that, I'm inclined to not to emit an error or a warning at all for
> -R and just make it an alias to -rpath. If you are so familiar with linker
> features that you want to >use --just-symbols, you are likely to know -R,
> -rpath and -just-symbols. -just-symbols is used very rarely, so in reality,
> I think this is not an issue.
> >
> > Two points from me:
> > "Given that, I'm inclined to not to emit an error or a warning at all
> for -R and just make it an alias to -rpath. If you are so familiar with
> linker features that you want to use --just-symbols, you are likely to know
> -R"
> >
> > Last part means that we should implement -R both for -rpath and
> --just-symbols, like original GNU linkers do. As it is what is expected by
> user, who familar with both.
> > Currently our behavior always behaves as -rpath without warnings or
> errors and produce broken output. And that nehavior is definetely not
> expected (and comments for PR35067 shows that).
> >
> > And then making -R an alias for one of these options is just not
> correct. If it's intention is to share 2 different behaviors it should be
> separate option with corresponding description, no ?
> >
> > ?Currently no our tests use -R for -rpath. What about if we drop
> supporting -R completely ? We can show a error suggesting to use explicitly
> use -rpath/--just-symbols instead.
>
> I *think* that using -R for --rpath is fairly common.
>
> I think I am back to favoring -R being an alias for rpath and producing
> an error for "--rpath file" or "-R file". Yes, the build machine might
> have a file and the intended run environment a directory, but that seems
> unlikely and easy to avoid.
>

I'm not in favor of that. Reasons being that (1) assuming that a runtime
environment is the same as a build environment is generally a bad idea even
though it can be easily avoided, (2) --just-symbols is used very rarely,
and (3) the -R option's behavior is undeniably a bit too hacky. Now seems
like a good time to kill that tricky behavior.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20171129/fe0dfb11/attachment.html>