[PATCH 1/5] Fix size of external_reloc for pe-aarch64

[Tamar Christina]

> -----Original Message-----
> From: Jan Beulich <jbeulich@suse.com>
> Sent: Tuesday, December 20, 2022 1:10 PM
> To: Mark Harmstone <mark@harmstone.com>
> Cc: binutils@sourceware.org; wej22007@outlook.com;
> zac.walker@linaro.org; Tamar Christina <Tamar.Christina@arm.com>
> Subject: Re: [PATCH 1/5] Fix size of external_reloc for pe-aarch64
> 
> On 20.12.2022 13:59, Mark Harmstone wrote:
> > On 16/12/22 10:47, Tamar Christina wrote:
> >>
> >> The code was taken from the Arm port which weirdly had a different
> >> size for this whether it was a WINCE target or not.  For WINCE target it has
> 10, but for
> >> everything else 14 (See include/coff/arm.h).   Since I didn't have
> relocations
> >> support anyway (the original intention was only to support objcopy
> >> workflow) I didn't dig too deeply into why Arm has it as 14 and the extra
> field.
> >>
> >> That said 10 looks to be the correct value.
> >>
> >> Thanks,
> >> Tamar
> >
> > Thanks Tamar. Sorry, I should have copied you in in the first place.
> >
> > I can't remember the specifics, but I definitely needed this patch to get
> EXEs working on Windows.
> 
> But EXEs don't use this kind of relocation; the base relocations (fixups) used
> there are encoded in an entirely different way. The relocs here should
> matter for are COFF object files only.

Agreed, the relocs here are COFF only.  That said 10 still looks like the correct value.
Microsoft's documentation seems to have become a bit unclear after they reworded it for ASLR support.
Originally the Image files didn't have relocations as by then they would have been resolved already, but for ASLR
there is exposed a .reloc section now so the loader can do the fixups.  That's what the docs now describe:
https://learn.microsoft.com/en-us/windows/win32/debug/pe-format#the-reloc-section-image-only

They however no longer describe COFF relocations at all.  If we find an older spec (like Revision 6.0 - February 1999),
when it was still doc/pdf like https://courses.cs.washington.edu/courses/cse378/03wi/lectures/LinkerFiles/coff.pdf 
you'll see that the COFF relocations are described in section 5.2 are still 10 bytes long.

This also coincides with LLVM's description of the COFF relocations https://llvm.org/doxygen/structllvm_1_1COFF_1_1relocation.html

It's highly likely that there is a bug in the Arm port, because the most used PE target on Arm is WINCE which sets 10.

Cheers,
Tamar

> Jan