Re: [PATCH 04/13] Kbuild: Rust support

[Nick Desaulniers <ndesaulniers@google.com>]

On Wed, Apr 14, 2021 at 5:43 PM Miguel Ojeda
<miguel.ojeda.sandonis@gmail.com> wrote:
>
> On Thu, Apr 15, 2021 at 1:19 AM Nick Desaulniers
> <ndesaulniers@google.com> wrote:
> >
> > -Oz in clang typically generates larger kernel code than -Os; LLVM
> > seems to aggressively emit libcalls even when the setup for a call
> > would be larger than the inlined call itself.  Is z smaller than s for
> > the existing rust examples?
>
> I will check if the `s`/`z` flags have the exact same semantics as
> they do in Clang, but as a quick test (quite late here, sorry!), yes,
> it seems `z` is smaller:
>
>       text data bss    dec   hex filename
>
>     126568    8 104 126680 1eed8 drivers/android/rust_binder.o [s]
>     122923    8 104 123035 1e09b drivers/android/rust_binder.o [z]
>
>     212351    0   0 212351 33d7f rust/core.o [s]
>     207653    0   0 207653 32b25 rust/core.o [z]

cool, thanks for verifying. LGTM

> > This is a mess; who thought it would be a good idea to support
> > compiling the rust code at a different optimization level than the
> > rest of the C code in the kernel?  Do we really need that flexibility
> > for Rust kernel code, or can we drop this feature?
>
> I did :P
>
> The idea is that, since it seemed to work out of the box when I tried,
> it could be nice to keep for debugging and for having another degree
> of freedom when testing the compiler/nightlies etc.
>
> Also, it is not intended for users, which is why I put it in the
> "hacking" menu -- users should still only modify the usual global
> option.
>
> However, it is indeed strange for the kernel and I don't mind dropping
> it if people want to see it out (one could still do it manually if
> needed...).
>
> (Related: from what I have been told, the kernel does not support
> lower levels in C just due to old problems with compilers; but those
> may be gone now).

IIRC the kernel (or at least x86_64 defconfig) cannot be built at -O0,
which is too bad if developers were myopically focused on build times.
It would have been nice to have something like
CONFIG_CC_OPTIMIZE_FOR_COMPILE_TIME to join
CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE and CONFIG_CC_OPTIMIZE_FOR_SIZE,
but maybe it's still possible to support one day.  (¿Por qué no los
tres? Perhaps a false-trichotomy? Sorry, but those 3 are somewhat at
odds for compilation).

Until then, I don't see why we need to permit developers to express
such flexibility for just the Rust code, or have it differ from the
intent of the C code. Does it make sense to set RUST_OPT_LEVEL_3 and
CC_OPTIMIZE_FOR_SIZE? I doubt it. That doesn't seem like a development
feature, but a mistake.  YAGNI.  Instead developers should clarify
what they care about in terms of high level intent; if someone wants
to micromanage optimization level flags in their forks they don't need
a Kconfig to do it (they're either going to hack KBUILD_CFLAGS,
CFLAGS_*.o, or KCFLAGS), and there's probably better mechanisms for
fine-tooth precision of optimizing actually hot code or not via PGO
and AutoFDO.
https://lore.kernel.org/lkml/20210407211704.367039-1-morbo@google.com/
-- 
Thanks,
~Nick Desaulniers