On Fri, Mar 24, 2023 at 12:54:01PM +0200, Kalle Valo wrote: > Nathan Chancellor <nat...@kernel.org> writes: > > > On Wed, Mar 22, 2023 at 02:44:47PM +0200, Kalle Valo wrote: > >> Nathan Chancellor <nat...@kernel.org> writes: > >> > >> > Perhaps these would make doing allmodconfig builds with clang more > >> > frequently less painful for you? > >> > > >> > https://lore.kernel.org/llvm/20230319235619.GA18547@dev-arch.thelio-3990X/ > >> > >> Thank you, at least for me this is really helpful. > > > > Really glad to hear! I hope this helps make testing and verifying > > changes with clang and LLVM easier for developers and maintainers. > > It really does. And I hope you are able to update these packages in > future as well so that it would be easy to get the latest clang.
That is the current plan (I will push 16.0.1, 16.0.2, etc. as they are released), I have a relatively automated process for this going forward. > >> I tried now clang for the first time but seeing a strange problem. > >> > >> I prefer to define the compiler in GNUmakefile so it's easy to change > >> compilers and I don't need to remember the exact command line. So I have > >> this in the top level GNUmakefile (all the rest commented out): > >> > >> LLVM=/opt/clang/llvm-16.0.0/bin/ > >> > >> If I run 'make oldconfig' it seems to use clang but after I run just > >> 'make' it seems to switch back to the host GCC compiler and ask for GCC > >> specific config questions again. Workaround for this seems to be adding > >> 'export LLVM' to GNUmakefile, after that also 'make' uses clang as > >> expected. > > > > Interesting... I just tested with a basic GNUmakefile and everything > > seems to work fine without an export. At the same time, the export > > should not hurt anything, so as long as it works, that is what matters. > > Sure, once I figured out the quirks I can workaround them. I was just > hoping that other users would not have to go through the same hassle as > I did :) > > > If you have any further issues, please do not hesitate to reach out! > > This is nitpicking but it would be nice if the tarball contents wouldn't > conflict with each other. Now both llvm-16.0.0-aarch64.tar.gz and > llvm-16.0.0-x86_64.tar extract to the same directory llvm-16.0.0 with > same binary names. It would be much better if they would extract to > llvm-16.0.0-aarch64 and llvm-16.0.0-x86_64, respectively. > > For example, Arnd's crosstool packages don't conflict with each other: > > https://mirrors.edge.kernel.org/pub/tools/crosstool/ I could certainly do that but what is the use case for extracting both? You cannot run the aarch64 version on an x86_64 host and vice versa, so why bother extracting them? I had figured the architecture would be irrelevant once installed on the host, so I opted only to include it in the tarball name. Perhaps I should make it clearer that these are the host architectures, not the target architectures (because clang is multi-targeted, unlike GCC)? > And maybe request a similar llvm directory under pub/tools to make it > more official? :) Yes, I was talking that over with Nick recently, as having it under a group on kernel.org would make taking over maintainership easier should something happen to me :) Thanks for all the feedback so far, it is much appreciated! Cheers, Nathan