Re: [PATCH 09/14] bitmap: extend bitmap_{get,set}_value8() to bitmap_{get,set}_bits()

[Alexander Lobakin]

From: Yury Norov <yury.norov@xxxxxxxxx>
Date: Mon, 9 Oct 2023 09:31:15 -0700

> + Alexander Potapenko <glider@xxxxxxxxxx>
> 
> On Mon, Oct 09, 2023 at 05:10:21PM +0200, Alexander Lobakin wrote:
>> Sometimes there's need to get a 8/16/...-bit piece of a bitmap at a
>> particular offset. Currently, there are only bitmap_{get,set}_value8()
>> to do that for 8 bits and that's it.
> 
> And also a series from Alexander Potapenko, which I really hope will
> get into the -next really soon. It introduces bitmap_read/write which
> can set up to BITS_PER_LONG at once, with no limitations on alignment
> of position and length:
> 
> https://lore.kernel.org/linux-arm-kernel/ZRXbOoKHHafCWQCW@yury-ThinkPad/T/#mc311037494229647088b3a84b9f0d9b50bf227cb
> 
> Can you consider building your series on top of it?

Yeah, I mentioned in the cover letter that I'm aware of it and in fact
it doesn't conflict much, as the functions I'm adding here get optimized
as much as the original bitmap_{get,set}_value8(), while Alexander's
generic helpers are heavier.
I realize lots of calls will be optimized as well due to the offset and
the width being compile-time constants, but not all of them. The idea of
keeping two pairs of helpers initially came from Andy if I understood
him correctly.
What do you think? I can provide some bloat-o-meter stats after
rebasing. And either way, I see no issue in basing this series on top of
Alex' one.

> 
>> Instead of introducing a separate pair for u16 and so on, which doesn't
>> scale well, extend the existing functions to be able to pass the wanted
>> value width. Make both offset and width arbitrary, but in order to not
>> over complicate the current logic and keep the helpers as optimized as
>> the current ones, require the width to be a pow-2 value and the offset
>> to be a multiple of the width, while the target piece should not cross
>> a %BITS_PER_LONG boundary and stay within one long.
>> Avoid adjusting all the already existing callsites by defining oneliner
>> wrapper macros named after the former functions. bloat-o-meter shows
>> almost no difference (+1-2 bytes in a couple of places), meaning the

[...]

Thanks,
Olek