Allow new just-symbols test to run on XCOFF and PE

[Maciej W. Rozycki]

On Tue, 28 Jul 2020, Alan Modra wrote:

> > >  Would you be able to update your change to keep that purpose?  What's 
> > > there with the affected targets that causes extra data to be present 
> > > ahead?
> > 
> > It's not extra data but rather that some aout formats don't even have
> > a .data section.
> > 
> > ~/build/gas/i386-msdos/ld$ ../binutils/objdump -s tmpdir/just-symbols-1obj 
> > 
> > tmpdir/just-symbols-1obj:     file format msdos
> > 
> > Contents of section .text:
> >  0000 12000000 00000000 00000000 00000000  ................

 Good point, yeah!  I reckon the run-time-relocatable .exe (MZ) format 
this target uses doesn't have a notion of sections at all and all the 
memory is RWX, so presenting the contents of a program as `.text' is as 
good a way as there could be (and then the more limited .com format is an 
entirely raw byte stream, so I guess we only support it with the `binary' 
BFD).

 NB somehow I missed `i386-msdos' when looking through bfd/config.bfd 
previously, so I have now added it to my list of targets to verify.

> > I'll reinstate the "file format" and "Contents of section" lines but
> > match any named section.
> 
> That failed on another aout target, ns32k, where the linker always
> gives a a non-empty .text section.  So I'm reverting the

 Indeed I'd expect at least some a.out targets to be inflexible enough to 
hardwire the presence of all the three of `.text'/`.data'/`.bss' even.

> just-symbols-1.dd change and applying the following.  The lang_check
> change fixes msp430-elf bfd_merge_private_bfd_data errors.

 Great, thank you for taking care of it!

> diff --git a/ld/testsuite/ld-misc/just-symbols.exp b/ld/testsuite/ld-misc/just-symbols.exp
> index 59c725d0d2..e3c012f0f6 100644
> --- a/ld/testsuite/ld-misc/just-symbols.exp
> +++ b/ld/testsuite/ld-misc/just-symbols.exp
> @@ -20,7 +20,13 @@
>  #
>  
>  # SH/PE targets complain about zero VMA.
> -if { [istarget sh-*-pe] } {
> +# i386-msdos doesn't have .data
> +# tic30-coff has a weird 32-bit reloc that divides sym value by four
> +# alpha-dec-vms wants a bunch of libs to link
> +if { [istarget sh-*-pe]
> +     || [istarget *-*-msdos]
> +     || [istarget *c30-*-*]
> +     || [istarget alpha-*-vms] } {

 I think there's no need to exclude `alpha-dec-vms' specifically for this 
test case, as the target wants to link in those libraries throughout and 
all the LD tests fail, e.g.:

./ld-new: cannot find -limagelib
./ld-new: cannot find -lstarlet
./ld-new: cannot find -lsys$public_vectors
FAIL: script

(which I discovered many years ago but never bothered to look into).

 It looks like those are hardcoded in the emulation itself rather than in 
a linker script:

static void
gld${EMULATION_NAME}_create_output_section_statements (void)
{
  lang_add_input_file ("imagelib", lang_input_file_is_l_enum, NULL);
  lang_add_input_file ("starlet", lang_input_file_is_l_enum, NULL);
  lang_add_input_file ("sys\$public_vectors", lang_input_file_is_l_enum, NULL);
}

so I think the right way to sort it might be by having the test framework 
prebuild stub versions of those, except for the native configuration; 
perhaps by making them empty-ish linker scripts so as to avoid a 
chicken-and-egg problem here.

  Maciej