Re: [PATCH v15 5/6] rust: time: Add wrapper for fsleep() function

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

On Tue, Apr 29, 2025 at 12:14:04PM -0700, Boqun Feng wrote:
> On Tue, Apr 29, 2025 at 08:33:47PM +0200, Arnd Bergmann wrote:
> > On Tue, Apr 29, 2025, at 19:15, Boqun Feng wrote:
> > > On Tue, Apr 29, 2025 at 06:11:02PM +0200, Arnd Bergmann wrote:
> > >> On Tue, Apr 29, 2025, at 18:03, Boqun Feng wrote:
> > >
> > > Would it make sense if we rely on compiler optimization when it's
> > > avaiable (for x86_64, arm64, riscv, etc), and only call ktime_to_ms() if

In case I wasn't clear, nowadays Rust compiler supports optimizating
constant division into multi/shift on x86_64, arm64, riscv already, the
optimization is only not availabe for arm32. (It's actually an
optimization provided by LLVM I think)

Regards,
Boqun

> > > not? The downside of calling ktime_to_ms() are:
> > >
> > > * it's a call function, and cannot be inlined with LTO or INLINE_HELPER:
> > >
> > > 	https://lore.kernel.org/all/20250319205141.3528424-1-gary@garyguo.net/
> > >
> > > * it doesn't provide the overflow checking even if
> > >   CONFIG_RUST_OVERFLOW_CHECKS=y
> > >
> > > Thoughts?
> > 
> > The function call overhead is tiny compared to replacing a 64-bit
> > division with a constant mult/shift.
> > 
> 
> Just to be clear, are you essientially saying that even in C,
> ktime_to_ms() is not worth inlining? Because the call overhead is tiny
> compared to the function own cost?
> 
> My impression is that on x86 at least, function call is 10+ cycles, and
> multiply is 3 cycles, so I would think that ktime_to_ms() itself is at
> most 10 cycles. Maybe I'm out of date of the modern micro-architecture?
> 
> > What is the possible overflow that can happen here? For a constant
> > division at least there is no chance of divide-by-zero. Do you mean
> > truncating to 32 bit?
> > 
> 
> I was referring the last part of Miguel's email:
> 
> 	https://lore.kernel.org/rust-for-linux/CANiq72mMRpY4NC4_8v_wDpq6Z3qs99Y8gXd-7XL_3Bed58gkJg@mail.gmail.com/
> 
> Regards,
> Boqun
> 
> >      Arnd