Re: [PATCH v1 1/2] LoongArch: Make LTO case independent in Makefile

[Tiezhu Yang <yangtiezhu@loongson.cn>]

On 2025/9/23 上午9:10, Nathan Chancellor wrote:
>>>> ifdef CONFIG_LTO_CLANG
>>>> ifdef CONFIG_LLD_HAS_ANNOTATE_TABLEJUMP
>>>> KBUILD_LDFLAGS                  += -mllvm --loongarch-annotate-tablejump
>>>> else
>>>> KBUILD_LDFLAGS                  += -mllvm --no-jump-tables
>>>> endif
> 
> There is no '--no-jump-tables' LLVM option so this will not work.
> Shouldn't -fno-jump-tables and -Zno-jump-tables take care of generating
> jump tables?
> 
>> I do not know whether this is valid, you can test it with llvm 18
>> and llvm 20 if you think it is a proper way.
> 
> For what it's worth, Huacai's original suggestion of
> 
>      KBUILD_LDFLAGS += $(call ld-option,-mllvm --loongarch-annotate-tablejump)
> 
> appears to work for me and I do not see any objtool warnings with LTO
> enabled but I might be missing something.

As discussed offline, Huacai will test this change and submit a patch
later.

>> But IIRC, there is objtool warning with llvm 18, I reported to llvm
>> developer Wang Lei and he fixed it as the following commit:
>>
>> [LoongArch] Avoid indirect branch jumps using the ra register
>> https://github.com/llvm/llvm-project/commit/21ef17c62645
>>
>> Actually, the above commit solved a performance issue of llvm compiler,
>> so I prefer to update the minimal llvm version to 20 for LoongArch.
> 
> I tend to let architecture maintainers make the call around minimum
> supported versions of compilers, so if that is how you would like to
> proceed, I am fine with that. I will say LLVM 20 is pretty new (released
> on March 4th, 2025) but I expect most of your users to probably be using
> bleeding edge tools for all the changes you make in the compiler and
> lower level libraries?

Usually, I like to use the latest LLVM upstream version to test the
kernel, but maybe somebody only uses the lower version, so keep the
minimal version as is.

Thanks,
Tiezhu