Re: [PATCH v1] rust: time: Avoid 64-bit integer division

[Boqun Feng <boqun.feng@gmail.com>]

On Thu, May 01, 2025 at 05:11:44PM +0200, Arnd Bergmann wrote:
> On Thu, May 1, 2025, at 15:20, Boqun Feng wrote:
> > On Thu, May 01, 2025 at 06:12:02AM -0700, Boqun Feng wrote:
> >> On Thu, May 01, 2025 at 10:07:17PM +0900, FUJITA Tomonori wrote:
> >> > On Thu, 1 May 2025 05:26:54 -0700
> >> > Boqun Feng <boqun.feng@gmail.com> wrote:
> >> > 
> >> > > On Thu, May 01, 2025 at 10:58:18AM +0900, FUJITA Tomonori wrote:
> >> > >> Avoid 64-bit integer division that 32-bit architectures don't
> >> > >> implement generally. This uses ktime_to_ms() and ktime_to_us()
> >> > >> instead.
> >> > >> 
> >> > >> The timer abstraction needs i64 / u32 division so C's div_s64() can be
> >> > >> used but ktime_to_ms() and ktime_to_us() provide a simpler solution
> >> > >> for this timer abstraction problem. On some architectures, there is
> >> > >> room to optimize the implementation of them, but such optimization can
> >> > >> be done if and when it becomes necessary.
> >> > >> 
> >> > > 
> >> > > Nacked-by: Boqun Feng <boqun.feng@gmail.com>
> >> > > 
> >> > > As I said a few times, we should rely on compiler's optimization when
> >> > > available, i.e. it's a problem that ARM compiler doesn't have this
> >> > > optimization, don't punish other architecture of no reason.
> 
> What is Arm specific here? I'm not aware of the compiler doing anything

Because Arm is the only 32bit architecture that selects CONFIG_HAVE_RUST
for non-UML cases, i.e. this is the only 32bit architecture that has
this problem. If your point is we should do this for all 32bit
architectures, then I won't disagree. Just s/CONFIG_ARM/CONFIG_32BIT
then.

Regards,
Boqun

> different from the other 32-bit architectures, though most are missing
> an optimized __arch_xprod_64() and fall back to slightly worse code
> from the asm-generic version.
> 
> > Copy-paste errors:
> >
> > 	    unsafe { bindings::ktime_to_ms(ns) }
> >
> >> 	}
> >> 
> >> 	#[cfg(not(CONFIG_ARM))]
> >> 	fn ns_to_ms(ns: i64) -> i64 {
> >> 	    self.as_nanos() / NSEC_PER_MSEC
> >
> > 	    ns / NSEC_PER_MSEC
> 
> I'm sure this is still broken on all 32-bit targets.
> 
>      Arnd