Re: [RFC PATCH] Enabling ld -r ELF .stab/.stabstr optimization

[Ian Lance Taylor <ian at airs dot com>]

Alan Modra <amodra@bigpond.net.au> writes:

> On Wed, Sep 25, 2002 at 10:03:29PM -0700, Earl Chew wrote:
> > Here is a patch to enable .stab/.stabstr optimization in the presence
> > of ld -r. This patch was generated against binutils-2.10.1 --- sorry
> > the baseline is a little dated, but that's what I'm running.
> > 
> >          * bfd/elflink.h (elf_link_add_object_symbols): optimize .stabstr
> >          even if relocatable
> >          * ld/ldlang.c (wild_doit): discard debugging sections that have
> >          been marked SEC_EXCLUDE
> 
> I'm a little nervous about this change, although I haven't thought of
> any particular problem.  Suggestion:  Modify ld so that -r sets
> link_info.traditional_format, and implement a -fno-traditional-format
> option.  That way "ld -r" behaviour is unchanged unless you write
> "ld -r -fno-traditional-format".  Oh, and cofflink.c should have a
> similar change to your elflink.h change.

I think that having -r imply -ftraditional-format would have other
unexpected effects.  --traditional-format implies more things than
just optimizing .stabs sections.

Having -r optimize stabs sections will work poorly if the resulting
object is indeed linked with other objects, as opportunities for
optimizing stabs in the resulting executable between different -r
objects will be lost.  This will hurt a certain style of building
large programs, in which each directory contributes a single .o file
built using -r.

Therefore I believe we do need a new option here, but I don't think it
should be related to --traditional-format.  Instead, we need something
horrible like --optimize-stabs (the default for a final link) and
--no-optimize-stabs (the default for a relocateable link).  Then
--traditional-format should imply --no-optimize-stabs.

Ian