Re: [PATCH] arm64/io: add constant-argument check

[Mark Rutland <mark.rutland@arm.com>]

On Wed, May 29, 2024 at 06:15:57PM +0200, Arnd Bergmann wrote:
> On Wed, May 29, 2024, at 17:08, Mark Rutland wrote:
> > On Wed, May 29, 2024 at 02:29:37PM +0200, Arnd Bergmann wrote:
> >> On Wed, May 29, 2024, at 13:14, Mark Rutland wrote:
> 
> >> 
> >> Yes, your version addresses both failures I ran into, and
> >> I think all other theoretical cases.
> >> 
> >> I would prefer to combine both though, using __always_inline
> >> to force the compiler to pick the inline version over
> >> __iowrite32_copy_full() even when it is optimizing for size
> >> and it decides the inline version is larger, but removing
> >> the extra complexity from the macro.
> >
> > Sorry, I'm not sure what you mean here. I don't see anything handling
> > optimizing for size today so I'm not sure what change your suggesting to
> > force the use of the inline version; AFAICT that'd always be forced for
> > a suitable constant size.
> >
> > What change are you suggesting?
> 
> What I meant is that reason gcc chooses to not inline
> the macro is when we build with CONFIG_CC_OPTIMIZE_FOR_SIZE.
> 
> Since it doesn't know that __const_memcpy_toio_aligned64()
> is intended to be small after inlining, it sometimes
> decides against it, which (with just my patch) would
> fall back to the out-of-line __iowrite32_copy_full()
> while trying to generate smaller code.
> 
> The __always_inline annotation just overrides the
> calculation.

Ah, ok.

I think what you're suggesting is:

* Add the __always_inline annotations, as in my patch.

* Move the __builtin_constant_p check into __const_iowrite32_copy(), as in your
  patch.

* Remove the __iowrite32_copy() macro and rename __const_iowrite32_copy() to
  __iowrite32_copy(), removing the redundant logic.

Assuming so, that makes total sense to me.

Mark.