Re: [PATCH] random: get_random_u64_below()
From: David Laight
Date: Sat Mar 15 2025 - 16:55:46 EST
On Sat, 15 Mar 2025 14:20:46 -0400
Kent Overstreet <kent.overstreet@xxxxxxxxx> wrote:
> On Sat, Mar 15, 2025 at 01:52:34PM +0000, David Laight wrote:
> > On Thu, 13 Mar 2025 12:38:10 -0400
> > Kent Overstreet <kent.overstreet@xxxxxxxxx> wrote:
> >
> > > bcachefs needs this, for sampling devices to read from based on squared
> > > device latencies.
> > >
> > > this uses the same algorithm as get_random_u32_below: since the multiply
> > > uses the top and bottom halves separately, it works out fairly well.
> >
> > Adding two separate copies of much the same code is silly.
> > Given what the code is doing, does it ever make any sense to inline it.
> >
> > Inlining the original get_random_u32_below(ceil) that did
> > (random_u32() * ((1ull << 32) / ceil) >> 32
> > (for constant ceil) made sense.
> > While good enough for most purposes it was replaced by the much more
> > expensive function that guarantees that all the output values are
> > equally likely - rather than just evenly distributed.
>
> Expensive!? It adds a multiply.
I make it two multiplies and a loop.
Have you looked at what happens on 32bit systems?
>
> That % gets constant folded, in the inlined case, and in the non-inline
> case it's hit only a small fraction of the, time, for typical ceil.
If the % is only a small fraction on the cost for the non-inline case
(and it is over 100 clocks on many cpu that people still use) then
why inline anything?
A quick look shows divide being 'only moderately slow' on zen3 and coffee lake.
What you might want to do is pass -ceil % ceil through to a real function
(especially if constant).
Oh I guess you haven't actually tested the version you submitted.
Time to play 'spot the silly error'.
David