Re: [PATCH 1/2] LoongArch: bfd: Remove elf_seg_map condition in loongarch_elf_relax_section

[mengqinggang <mengqinggang@loongson.cn>]

Hi,

I will send a patch in the next few days to solve this issue and other 
issues
related to relaxation.


在 2023/10/5 下午7:19, Xi Ruoyao 写道:
> On Thu, 2023-10-05 at 18:09 +0800, Xi Ruoyao via Binutils wrote:
>> Hi Jinyang,
>>
>> Any progress on this?  During my attempt trying to balance relaxation
>> and scheduling better in GCC I found the elf_seg_map condition is
>> preventing relaxation on *every* shared library (when we are linking a
>> shared library the elf_seg_map of .text is NULL).
>>
>> On a modern system most code paths are in shared libraries, so it's
>> really bad not to perform the relaxation on them.  Esp. now we are
>> disabling explicit relocs for relaxation, so if we don't relax shared
>> libraries we are likely regressing the overall performance of the
>> system.
> Phew.  It's not only a performance issue.  It's actually a *correctness*
> issue because the condition also causes R_LARCH_ALIGN skipped, so some
> programs depending on code alignment (for e.g. duff-device-like code
> using pcaddi for eg) will be broken.
>
> Generally skipping the entire relaxation pass for any section containing
> R_LARCH_ALIGN is wrong (as we've discussed in
> https://github.com/llvm/llvm-project/pull/67424).  Even if we don't
> really relax a thing we still *must* process R_LARCH_ALIGN.
>
>> Sorry for disturbing you during the national holiday, feel free to defer
>> the reply until the holiday ends!