Re: [PATCH next 0/3] lib: Implement mul_u64_u64_div_u64_roundup()

[David Laight <david.laight.linux@gmail.com>]

On Fri, 16 May 2025 11:47:58 +0200
Uwe Kleine-König <u.kleine-koenig@baylibre.com> wrote:

> Hello David,
> 
> On Sat, Apr 05, 2025 at 09:45:27PM +0100, David Laight wrote:
> > The pwm-stm32.c code wants a 'rounding up' version of mul_u64_u64_div_u64().
> > This can be done simply by adding 'divisor - 1' to the 128bit product.
> > Implement mul_u64_add_u64_div_u64(a, b, c, d) = (a * b + c)/d based on the
> > existing code.
> > Define mul_u64_u64_div_u64(a, b, d) as mul_u64_add_u64_div_u64(a, b, 0, d) and
> > mul_u64_u64_div_u64_roundup(a, b, d) as mul_u64_add_u64_div_u64(a, b, d-1, d).
> > 
> > Only x86-64 has an optimsed (asm) version of the function.
> > That is optimised to avoid the 'add c' when c is known to be zero.
> > In all other cases the extra code will be noise compared to the software
> > divide code.
> > 
> > I've updated the test module to test mul_u64_u64_div_u64_roundup() and
> > also enhanced it to verify the C division code on x86-64.
> > 
> > Note that the code generated by gcc (eg for 32bit x86) just for the multiply
> > is rather more horrid than one would expect (clang does better).
> > I dread to think how long the divide loop takes.
> > And I'm not at all sure the call in kernel/sched/cputime.c isn't in a
> > relatively common path (rather than just hardware initialisation).
> > 
> > David Laight (3):
> >   lib: Add mul_u64_add_u64_div_u64() and mul_u64_u64_div_u64_roundup()
> >   lib: Add tests for mul_u64_u64_div_u64_roundup()
> >   lib: Update the muldiv64 tests to verify the C on x86-64  
> 
> I wonder what happend to this series. I'd like to make use of
> mul_u64_u64_div_u64_roundup() so I'd be interested to get this into the
> mainline.

I've a WIP rewrite of the divide code, speeds it up considerably for
'not amd-64'.

IIRC (the test machine is powered off) the test cases with random data
do down from over 900 clocks to below 150 for x86-32.
The 64bit code (on x64-x64 but skipping the divide instruction) are ~70
clocks rather than nearer 180.
Both those clock counts are almost data independent.

The code relies on the cpu having an 'unsigned long/unsigned long'
instruction.

I got reasonable code for 32bit x64 running gcc 12.2 from debian.
Then copied it to godbolt and found gcc before 15.0 making a pigs
breakfast of it (all clang versions are fine).
So some rework.

I still need to actually plumb it into the kernel sources.

I do want to test the divide loop on x86-64, at least for 64bit.
Probably by including the code directly into the test module.

	David