> On Nov 9, 2017, at 3:02 AM, Andi Kleen <[email protected]> wrote:
>
>> There's the series from Andi Kleen that enables LTO for Linux on x86:
>> https://lwn.net/Articles/512548/
>> https://github.com/andikleen/linux-misc/tree/lto-411-1
>>
>> It has solved many problems you also try to solve, and some patches
>> are looking very similar.
>>
>> At now we have different patchsets for gcc and clang, and it would be
>> better to have them together. One thing I'm worried is that you introduce
>> CONFIG_CLANG_LTO and use it for all cases, including that where more
>> generic CONFIG_LTO should be used.
>
> Yes would be good to merge the two. I've been looking at updating
> my old one.
>
> I don't cover any ARM code, but lots of generic code. My patches
> also worked on MIPS at least.
>
> There's also older patches to enable single-pass-linking for kallsyms,
> which is extremly useful for LTO build performance.
[Yury, thanks for the CC:]
Chiming in from the toolchain side, Linaro's Toolchain team will try to help with any GCC or Clang issues that are exposed by building kernel with LTO on arm64 / arm.
Regarding CONFIG_* options, I would expect most of the configuration changes to be equally valid for both GCC's and Clang's LTO support. Sami, I don't think it's fair to ask you to support both Clang and GCC in your patchset, but, where changes are obviously toolchain-agnostic, could you use CONFIG_LTO? And use CONFIG_LTO_CLANG for Clang-specific parts?
This way we will be able to avoid most of the refactoring when adding support for GCC's LTO.
Thank you,
--
Maxim Kuvyrkov
http://www.linaro.org
From 1583544633170484133@xxx Thu Nov 09 00:04:05 +0000 2017
X-GM-THRID: 1583065954961523588
X-Gmail-Labels: Inbox,Category Forums,HistoricalUnread