Re: [ARM-FDPIC 00/12] FDPIC ABI for ARM

[Christophe Lyon <christophe dot lyon at st dot com>]

Hi Nick,

On 23/03/2018 16:08, Nick Clifton wrote:
> Hi Christophe,
>
>    As a first pass over this patch set, I applied them to the sources
>    and ran my regression tests.  Unfortunately there were several new
>    failures:

Thanks for your interest in this patch series.

> Checking Binutils in: arm-none-eabi ... LD: 6  done
>    LD REGRESSION: FDPIC ARM shared library little
>    LD REGRESSION: FDPIC ARM dynamic executable little
>    LD REGRESSION: FDPIC ARM shared library big
>    LD REGRESSION: FDPIC ARM dynamic executable big
>    LD REGRESSION: FDPIC ARM shared library little endian M profile
>    LD REGRESSION: FDPIC ARM dynamic executable little endian M profile
> Checking Binutils in: armeb-eabi ... LD: 6  done
>    LD REGRESSION: FDPIC ARM shared library little
>    LD REGRESSION: FDPIC ARM dynamic executable little
>    LD REGRESSION: FDPIC ARM shared library big
>    LD REGRESSION: FDPIC ARM dynamic executable big
>    LD REGRESSION: FDPIC ARM shared library little endian M profile
>    LD REGRESSION: FDPIC ARM dynamic executable little endian M profile
>
> It seems that these tests are all failing for the same reason:
>
>    ld-new: unrecognised emulation mode: armelf_linux_fdpiceabi
>
>
> Checking Binutils in: arm-netbsdelf ... GAS: 1  LD: 2  done
>    GAS REGRESSION: FDPIC relocations
> Checking Binutils in: arm-nto ... GAS: 1  done
>    GAS REGRESSION: FDPIC relocations
>
> Meanwhile these tests fail for a similar, but not quite the same reason:
>
>    Fatal error: selected target format 'elf32-littlearm-fdpic' unknown
>
> Would you mind fixing these up please ?

I'm looking at it. All these failures are caused by the tests added by this series.

I did test them on arm-linux-gnueabihf, armeb-linux-gnueabihf and-linux-uclibceabi,
but overlooked testing on arm-eabi and others. As usual, not tested means it
does not work :(

My homework on de-hardcoding things is obviously not complete :)


The simple route would probably be just to skip the tests on those targets,
but I have been trying to do something better, but didn't succeed so far.

Currently, gas sets the OSABI flag to ELFOSABI_ARM_FDPIC if the user
supplied -fdpic. I think this flag should set the linker in the right
mode (fdpic/no-fdpic).

It seems I have to set arm_elf32_fdpic_le_vec as default (targ_defvec in config.bfd),
such that the linker uses that in priority (otherwise arm_elf32_le_vec also
accepts to read an FDPIC object file, but fails to flag it as "FDPIC" (because
elf-object_p is happy with ELFOSABI_NONE).

However, if I do that, the linker then defaults to generating FDPIC files,
and thus breaks arm-linux-gnueabihf, unless I force -m armelf_linux_eabi.

I have looked at other targets, for instance sh adds -m shelf_fd for its fdpic tests.
That's what the patches I have posted do. Is this the only way?

How can I have ld support both models transparently without needing to use -m ?

Thanks,

Christophe


> Cheers
>    Nick
> .