Re: ☠ Buildbot (Sourceware): binutils-gdb - failed test (failure) test (failure) (master)

[Nick Clifton <nickc@redhat.com>]

Hi Palmer,

> I can't think of any reason the RISC-V attributes would specifically trip up on an objcopy that doesn't change anything else.  We've got some string-based attributes, but 
> if nothing else changes they should stay the same too.  There's also some attribute merging code, but again that shouldn't change anything.
> 
> That said, I wouldn't be surprised if we have some bug floating around there somewhere.  The attributes aren't all that widely used (they sort of just store a bunch of 
> stuff that doesn't have much meaning, we've gotten burned by backwards compatability there a few times).  I think tooling mostly ignores them these days.
> 
> LMK if you want Nelson or I to look, but happpy to have the help ;)

Thanks for the offer, but I think that I have already found the problem.
The issue was that the .riscv_attributes section was appearing before
the .text section in BFD's list of sections to allocate to segments.
This was then tripping up on a bug in the fix for PR 31450 which made
the BFD library think that it needed to rework the section to segment
mapping which then meant that the RISCV_ATTRIBUTES segment was moved to
a different address which finally resulted in objcopy producing a changed
binary.  (phew!)

Anyway I am testing a patch locally which I think will fix the problem.

Cheers
   Nick