[Bug ld/28410] Prevent region check failures when relaxation is not final

[lewis.revill at embecosm dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=28410

lewis.revill at embecosm dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |lewis.revill at embecosm dot 
com

--- Comment #3 from lewis.revill at embecosm dot com ---
(In reply to Nelson Chu from comment #2)
> commit abd20cb637008da9d32018b4b03973e119388a0a
> Refs: users/ARM/embedded-gdb-master-2018q4-7811-gabd20cb
> Author:     Nelson Chu <nelson.chu@sifive.com>
> AuthorDate: Tue Nov 17 19:39:52 2020 -0800
> Commit:     Nelson Chu <nelson.chu@sifive.com>
> CommitDate: Sat Nov 21 09:41:58 2020 +0800
> 
>     RISC-V: Relax PCREL to GPREL while doing other relaxations is dangerous.
> 
> commit ebdcad3fddf6ec21f6d4dcc702379a12718cf0c4
> Refs: users/ARM/embedded-gdb-master-2018q4-9197-gebdcad3
> Author:     Nelson Chu <nelson.chu@sifive.com>
> AuthorDate: Fri Dec 18 10:59:41 2020 +0800
> Commit:     Nelson Chu <nelson.chu@sifive.com>
> CommitDate: Thu Mar 11 17:27:13 2021 +0800
> 
>     RISC-V: Improve multiple relax passes problem.
> 
> 
> The commit ebdcad causes troubles for the above testcase.  It is used to
> increase the relax chances of the pcrel relaxations, since I separate them
> to the different relax pass in commit abd20c, to resolve the dangerous
> relocation problem.
> 
> However, Lewis Revill's solution should work well, but the changes of the
> generic code (lang_relax_sections) is beyond my expectation.  For now we may
> have three options,
> 
> 1. Keep the commit ebdcad, and use Lewis Revill's solution to resolve the
> checking region problem.
> 
> 2. Rewrite the commit ebdcad, don't use an extra target `restart_relax',
> just find a way to enhance the `again' and lang_relax_sections can rerun the
> whole relaxations.
> 
> 3. Revert the commit ebdcad in the short-term, since maybe not so many pcgp
> relaxations will be affected.  For long-term, we should rewrite the commit
> abd20c.  Perhaps we shouldn't separate the pcgp relaxations to the different
> relax pass, we should find a way to update the pcgp_relocs table once we
> actually delete the code.
> 
> My personal opinion is the third option since I really don’t want to change
> the generic code only for riscv target.

I had thought about a possible approach to rewrite the commit, such that we can
restart the relaxation process again without modifying the `again` pointer.
Perhaps its worth me working on this if that's a preferable solution to this,
the only caveat is that the exact behaviour of what gets repeated wouldn't be
preserved - though this may be fine to achieve the same result.

-- 
You are receiving this mail because:
You are on the CC list for the bug.