Re: [PATCH 04/13] lib: introduce BITS_{FIRST,LAST} macro

[Yury Norov]

On Tue, Mar 16, 2021 at 01:42:45PM +0200, Andy Shevchenko wrote:
> On Tue, Mar 16, 2021 at 09:35:35AM +0100, Rasmus Villemoes wrote:
> > On 16/03/2021 02.54, Yury Norov wrote:
> > > BITMAP_{LAST,FIRST}_WORD_MASK() in linux/bitmap.h duplicates the
> > > functionality of GENMASK(). The scope of BITMAP* macros is wider
> > > than just bitmaps. This patch defines 4 new macros: BITS_FIRST(),
> > > BITS_LAST(), BITS_FIRST_MASK() and BITS_LAST_MASK() in linux/bits.h
> > > on top of GENMASK() and replaces BITMAP_{LAST,FIRST}_WORD_MASK()
> > > to avoid duplication and increase the scope of the macros.
> > > 
> > > This change doesn't affect code generation. On ARM64:
> > > scripts/bloat-o-meter vmlinux.before vmlinux
> > > add/remove: 1/2 grow/shrink: 2/0 up/down: 17/-16 (1)
> > > Function                                     old     new   delta
> > > ethtool_get_drvinfo                          900     908      +8
> > > e843419@0cf2_0001309d_7f0                      -       8      +8
> > > vermagic                                      48      49      +1
> > > e843419@0d45_000138bb_f68                      8       -      -8
> > > e843419@0cc9_00012bce_198c                     8       -      -8
> > 
> > [what on earth are those weird symbols?]
> > 
> > 
> > > diff --git a/include/linux/bits.h b/include/linux/bits.h
> > > index 7f475d59a097..8c191c29506e 100644
> > > --- a/include/linux/bits.h
> > > +++ b/include/linux/bits.h
> > > @@ -37,6 +37,12 @@
> > >  #define GENMASK(h, l) \
> > >  	(GENMASK_INPUT_CHECK(h, l) + __GENMASK(h, l))
> > >  
> > > +#define BITS_FIRST(nr)		GENMASK((nr), 0)
> > > +#define BITS_LAST(nr)		GENMASK(BITS_PER_LONG - 1, (nr))
> > > +
> > > +#define BITS_FIRST_MASK(nr)	BITS_FIRST((nr) % BITS_PER_LONG)
> > > +#define BITS_LAST_MASK(nr)	BITS_LAST((nr) % BITS_PER_LONG)
> > 
> > I don't think it's a good idea to propagate the unusual closed-range
> > semantics of GENMASK to those wrappers. Almost all C and kernel code
> > uses the 'inclusive lower bound, exclusive upper bound', and I'd expect
> > BITS_FIRST(5) to result in a word with five bits set, not six. So I
> > think these changes as-is make the code much harder to read and understand.
> > 
> > Regardless, please add some comments on the valid input ranges to the
> > macros, whether that ends up being 0 <= nr < BITS_PER_LONG or 0 < nr <=
> > BITS_PER_LONG or whatnot.
> > 
> > It would also be much easier to review if you just redefined the
> > BITMAP_LAST_WORD_MASK macros etc. in terms of these new things, so you
> > wouldn't have to do a lot of mechanical changes at the same time as
> > introducing the new ones - especially when those mechanical changes
> > involve adding a "minus 1" everywhere.
> 
> I tend to agree with Rasmus here.

OK. All this plus terrible GENMASK(high, low) design, when high goes
first, makes me feel like we need to deprecate GENMASK and propose a
new interface.

What do you think about this:
BITS_FIRST(bitnum)      -> [0, bitnum)
BITS_LAST(bitnum)       -> [bitnum, BITS_PER_LONG)
BITS_RANGE(begin, end)  -> [begin, end)

We can pick BITS_{LAST,FIRST} implementation from existing BITMAP_*_WORD_MASK
analogues, and make the BITS_RANGE like:
        #define BITS_RANGE(begin, end) BITS_FIRST(end) & BITS_LAST(begin)

Regarding BITMAP_*_WORD_MASK, I can save them in bitmap.h as aliases
to BITS_{LAST,FIRST} to avoid massive renaming. (Should I?)

Would this all work for you?