Re: [PATCH v2] x86-64: Add -z mark-plt and -z nomark-plt

[Fangrui Song <i@maskray.me>]

On Wed, Sep 27, 2023 at 11:07 PM Jan Beulich via Binutils
<binutils@sourceware.org> wrote:
>
> On 27.09.2023 19:19, H.J. Lu via Binutils wrote:
> > --- a/ld/ld.texi
> > +++ b/ld/ld.texi
> > @@ -1475,6 +1475,11 @@ Specify that the object's filters be processed immediately at runtime.
> >  @item max-page-size=@var{value}
> >  Set the maximum memory page size supported to @var{value}.
> >
> > +@item mark-plt
> > +@itemx nomark-plt
> > +Mark PLT entries with dynamic tags, DT_X86_64_PLT, DT_X86_64_PLTSZ and
> > +DT_X86_64_PLTENT.  Supported for x86_64.
>
> Hmm, you use backwards compatibility issues as justification for not having
> a configure time way to default the new behavior to enabled, and then you
> don't mention the backwards compatibility aspect here?
>
> Jan

For linker defaults, they can be performed either on the linker side
or on the compiler driver side.
If GCC specifies an extra option to ld (like -z relro -z now), I think
it is fine for ld not to have a configure-time option.
(I understand that someone may argue that patching ld has a greater
coverage... I typically refute such arguments. This mark-plt is for a
rtld optimization that unlikely applies to the programs that directly
use ld anyway.)

At least for LLD that I maintain, we prefer not to have
target-specific differences and prefer that the compiler drivers pass
relevant options to the linker.
Linker not having to deal with too many defaults allows testing more
address-related properties in the output, which I find valuable to
catch many bugs without doing distribution testing.
(https://sourceware.org/bugzilla/show_bug.cgi?id=28602)