Re: Linux 6.3-rc3

[Masahiro Yamada <masahiroy@kernel.org>]

Hello Linus,


Thanks for giving me some more homeworks.


On Thu, Mar 23, 2023 at 1:56 AM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> On Wed, Mar 22, 2023 at 9:40 AM Sedat Dilek <sedat.dilek@gmail.com> wrote:
> >
> > You have to pass `make LLVM=1` in any case... to `oldconfig` or when
> > adding any MAKEFLAGS like -j${number-of-available-cpus}.
>
> I actually think we should look (again) at just making the compiler
> choice (and the prefix) be a Kconfig option.
>
> That would simplify *so* many use cases.
>
> It used to be that gcc was "THE compiler" and anything else was just
> an odd toy special case, but that's clearly not true any more.
>
> So it would be lovely to make the kernel choice a Kconfig choice - so
> you'd set it only at config time, and then after that a kernel build
> wouldn't need special flags any more, and you'd never need to play
> games with GNUmakefile or anything like that.


Presumably, this is the right direction.

To achieve it, Kconfig needs to have some mechanism to evaluate
shell commands dynamically.

If a user switches the toolchain set between GCC and LLVM
while running the Kconfig, $(cc-option) in Kconfig files must
be re-calculated.

Currently, Kconfig cannot do it. All macros are static - they are
expanded in the parse stage, and become constant strings.

Ulf Magnusson and I discussed the dynamic approach a few years back,
but I adopted the static way since it is much simpler.
We need to reconsider the dynamic approach to do this correctly.
I do not think it is too difficult technically.
We just need to come up with a decent syntax.



> Yes, you'd still use environment variables (or make arguments) for
> that initial Kconfig, but that's no different from the other
> environment variables we already have, like KCONFIG_SEED that kconfig
> uses internally, but also things like "$(ARCH)" that we already use
> *inside* the Kconfig files themselves.
>
> I really dislike how you have to set ARCH and CROSS_COMPILE etc
> externally, and can't just have them *in* the config file.
>
> So when you do cross-compiles, right now you have to do something like
>
>     make ARCH=i386 allmodconfig
>
> to build the .config file, but then you have to *repeat* that
> ARCH=i386 when you actually build things:
>
>     make ARCH=i386
>
> because the ARCH choice ends up being in the .config file, but the
> makefiles themselves always take it from the environment.
>
> There are good historical reasons for our behavior (and probably a
> number of extant practical reasons too), but it's a bit annoying, and
> it would be lovely if we could start moving away from this model.
>
>             Linus


Moving ARCH into the .config file needs careful thoughts, I think.

Not all targets include the .config file.
For example, "make clean", "make help", etc.

It is unclear which targets require explicit ARCH= option.

One solution is to move "archhelp", "CLEAN_FILES" etc.
from arch/*/Makefile to the top Makefile.
We will lose per-arch splitting in several places, though.


U-Boot adopts this model - 'ARCH' is determined in the Kconfig time,
so users do not need to give ARCH= option from the command line.

https://github.com/u-boot/u-boot/blob/v2023.01/arch/Kconfig#L44

You may get a quick idea of what it will look like.



I will take a look at this direction (the compiler choice in Kconfig first),
but it will not happen soonish due to the limited time for upstream work.


--
Best Regards

Masahiro Yamada