2022-05-12 11:22:49

by Nick Desaulniers

[permalink] [raw]
Subject: Re: [PATCH v3 1/2] x86/asm/bitops: ffs: use __builtin_ffs to evaluate constant expressions

On Wed, May 11, 2022 at 5:04 PM Vincent Mailhol
<[email protected]> wrote:
>
> And finally, Nick Desaulniers pointed out in [2] that this also fixes
> a constant propagation missed-optimization in clang.
>
> [2] https://lore.kernel.org/all/CAKwvOdnH_gYv4qRN9pKY7jNTQK95xNeH1w1KZJJmvCkh8xJLBg@mail.gmail.com/

Regarding
https://github.com/llvm/llvm-project/issues/55394
it seems that functions with static linkage cannot be considered
library functions, so libcall optimization will not run on calls to
them. So the compiler might be able to do a better job for constants
if ffs() and friends indeed were not defined in a header as static
inline. But that relies on the compiler knowing these tricks; I think
the kernel's approach is just fine (better in fact, because we should
inline these tiny functions, regardless of LTO), but like this series
shows, there may be room for improvement for other functions within
the kernel that are defined as static inline in headers that are
normally found in a libc.

So I no longer think there's a missed optimization here, but at this
point, it's not worth a respin of the series IMO to just let sleeping
dogs lie.

Unless the x86 maintainers wouldn't mind dropping that line and link
when applying?
--
Thanks,
~Nick Desaulniers


2022-05-13 19:17:27

by Vincent MAILHOL

[permalink] [raw]
Subject: Re: [PATCH v3 1/2] x86/asm/bitops: ffs: use __builtin_ffs to evaluate constant expressions

On Thu. 12 May 2022 at 09:28, Nick Desaulniers <[email protected]> wrote:
> On Wed, May 11, 2022 at 5:04 PM Vincent Mailhol
> <[email protected]> wrote:
> >
> > And finally, Nick Desaulniers pointed out in [2] that this also fixes
> > a constant propagation missed-optimization in clang.
> >
> > [2] https://lore.kernel.org/all/CAKwvOdnH_gYv4qRN9pKY7jNTQK95xNeH1w1KZJJmvCkh8xJLBg@mail.gmail.com/
>
> Regarding
> https://github.com/llvm/llvm-project/issues/55394
> it seems that functions with static linkage cannot be considered
> library functions, so libcall optimization will not run on calls to
> them. So the compiler might be able to do a better job for constants
> if ffs() and friends indeed were not defined in a header as static
> inline. But that relies on the compiler knowing these tricks; I think
> the kernel's approach is just fine (better in fact, because we should
> inline these tiny functions, regardless of LTO), but like this series
> shows, there may be room for improvement for other functions within
> the kernel that are defined as static inline in headers that are
> normally found in a libc.
>
> So I no longer think there's a missed optimization here, but at this
> point, it's not worth a respin of the series IMO to just let sleeping
> dogs lie.
>
> Unless the x86 maintainers wouldn't mind dropping that line and link
> when applying?

Let me send the v4, this will save the x86 maintainers some manual
editing effort (add will add your "Review-by" tag in patch 2 while
doing so).