Re: [PATCH v4 3/4] lib/find_bit: optimize find_next_bit() functions

[Yury Norov]

On Mon, Sep 19, 2022 at 04:45:54PM +0300, Andy Shevchenko wrote:
> On Wed, Sep 14, 2022 at 07:07:29PM -0700, Yury Norov wrote:
> > Over the past couple years, the function _find_next_bit() was extended
> > with parameters that modify its behavior to implement and- zero- and le-
> > flavors. The parameters are passed at compile time, but current design
> > prevents a compiler from optimizing out the conditionals.
> > 
> > As find_next_bit() API grows, I expect that more parameters will be added.
> > Current design would require more conditional code in _find_next_bit(),
> > which would bloat the helper even more and make it barely readable.
> > 
> > This patch replaces _find_next_bit() with a macro FIND_NEXT_BIT, and adds
> > a set of wrappers, so that the compile-time optimizations become possible.
> > 
> > The common logic is moved to the new macro, and all flavors may be
> > generated by providing a FETCH macro parameter, like in this example:
> > 
> >   #define FIND_NEXT_BIT(FETCH, MUNGE, size, start) ...
> > 
> >   find_next_xornot_and_bit(addr1, addr2, addr3, size, start)
> >   {
> > 	return FIND_NEXT_BIT(addr1[idx] ^ ~addr2[idx] & addr3[idx],
> > 				/* nop */, size, start);
> >   }
> > 
> > The FETCH may be of any complexity, as soon as it only refers the bitmap(s)
> > and an iterator idx.
> > 
> > MUNGE is here to support _le code generation for BE builds. May be
> > empty.
> > 
> > I ran find_bit_benchmark 16 times on top of 6.0-rc2 and 16 times on top
> > of 6.0-rc2 + this series. The results for kvm/x86_64 are:
> > 
> >                       v6.0-rc2  Optimized       Difference  Z-score
> > Random dense bitmap         ns         ns        ns      %
> > find_next_bit:          787735     670546    117189   14.9     3.97
> > find_next_zero_bit:     777492     664208    113284   14.6    10.51
> > find_last_bit:          830925     687573    143352   17.3     2.35
> > find_first_bit:        3874366    3306635    567731   14.7     1.84
> > find_first_and_bit:   40677125   37739887   2937238    7.2     1.36
> > find_next_and_bit:      347865     304456     43409   12.5     1.35
> > 
> > Random sparse bitmap
> > find_next_bit:           19816      14021      5795   29.2     6.10
> > find_next_zero_bit:    1318901    1223794     95107    7.2     1.41
> > find_last_bit:           14573      13514      1059    7.3     6.92
> > find_first_bit:        1313321    1249024     64297    4.9     1.53
> > find_first_and_bit:       8921       8098       823    9.2     4.56
> > find_next_and_bit:        9796       7176      2620   26.7     5.39
> > 
> > Where the statistics is significant (z-score > 3), the improvement
> > is ~15%.
> > 
> > According to the bloat-o-meter, the Image size is 10-11K less:
> > 
> > x86_64/defconfig:
> > add/remove: 32/14 grow/shrink: 61/782 up/down: 6344/-16521 (-10177)
> > 
> > arm64/defconfig:
> > add/remove: 3/2 grow/shrink: 50/714 up/down: 608/-11556 (-10948)
> 
> ...
> 
> >  /*
> 
> Seems like you wanted this to be a kernel doc, but it isn't right now.

No, I didn't. I can remove '@' below, if that concerns you.
 
> > - * This is a common helper function for find_next_bit, find_next_zero_bit, and
> > - * find_next_and_bit. The differences are:
> > - *  - The "invert" argument, which is XORed with each fetched word before
> > - *    searching it for one bits.
> > - *  - The optional "addr2", which is anded with "addr1" if present.
> > + * Common helper for find_next_bit() function family
> 
> In such case this should start with a name of the macro
> 
>  * FIND_NEXT_BIT - ...
> 
> > + * @FETCH: The expression that fetches and pre-processes each word of bitmap(s)
> > + * @MUNGE: The expression that post-processes a word containing found bit (may be empty)
> > + * @size: The bitmap size in bits
> > + * @start: The bitnumber to start searching at
> >   */
> 
> ...
> 
> > +#define FIND_NEXT_BIT(FETCH, MUNGE, size, start)				\
> > +({										\
> > +	unsigned long mask, idx, tmp, sz = (size), __start = (start);		\
> > +										\
> > +	if (unlikely(__start >= sz))						\
> > +		goto out;							\
> > +										\
> > +	mask = MUNGE(BITMAP_FIRST_WORD_MASK(__start));				\
> > +	idx = __start / BITS_PER_LONG;						\
> > +										\
> > +	for (tmp = (FETCH) & mask; !tmp; tmp = (FETCH)) {			\
> > +		if ((idx + 1) * BITS_PER_LONG >= sz)				\
> > +			goto out;						\
> > +		idx++;								\
> > +	}									\
> > +										\
> > +	sz = min(idx * BITS_PER_LONG + __ffs(MUNGE(tmp)), sz);			\
> > +out:										\
> 
> I dunno if GCC expression limits the scope of goto labels, but on the safe side
> you can add a prefix to it, so it becomes:
> 
> FIND_NEXT_BIT_out:
> 
> (or alike).

As Linus already said, the 'out' is function-scope. We can make it a
block-scope with __label__, but this would make an impression that we
are OK with stacking many FIND macros in a single function.

I spend some time trying to figure out a legitimate usecase for it, but
nothing came in mind. There are many real cases when we need 2 or more
find functions at once but all that cases would work with regular
wrappers around FIND_BIT(). Check this, for example:

https://lore.kernel.org/lkml/20220919210559.1509179-6-yury.norov@xxxxxxxxx/

I don't know how FIND_BIT() machinery will evolve with time. For now
it's a clean and neat local helper with a very straightforward usage.
Lets keep it simple now? If someone will decide to call FIND_BIT()
twice and fail, it would be a good hint that he's doing something
wrong.

> > +	sz;									\
> > +})
> 
> ...
> 
> > +unsigned long _find_next_zero_bit_le(const unsigned long *addr, unsigned
> > +		long size, unsigned long offset)
> 
> Usually we don't split parameters between lines.

Ok

Thanks,
Yury