Re: [PATCH] x86: mark some mpspec inline functions as __init

[Arnd Bergmann]

On Thu, Feb 25, 2021 at 1:42 PM Borislav Petkov <bp@xxxxxxxxx> wrote:
>
> On Thu, Feb 25, 2021 at 01:18:21PM +0100, Arnd Bergmann wrote:
> > Either way works correctly, I don't care much, but picked the __init
> > annotation as it seemed more intuitive. If the compiler decides to
> > make it out-of-line for whatever reason,
>
> Well, frankly, I see no good reason for not inlining a function body
> which is a single call. And gcc does it just fine. And previous clangs
> did too, so why does clang-13 do it differently?
>
> IOW, could it be that you're fixing something that ain't broke?

Maybe Nick can give some more background here. He mentioned
elsewhere that inlining in clang-13 was completely rewritten and is generally
better than it was before. I'm not sure whether this particular instance
is a case where clang could in fact show an advantage in not inlining
a function, or if this is one case where it got worse and needs to be
tuned better.

In the end, inlining is a bunch of heuristics that are intended help
most of the time, but both (old) clang and gcc get it wrong in cases that
should have been decided the other way. Here we get a new method
that may go wrong in other ways.

        Arnd