Re: [PATCH v2] lib/memweight.c: open codes bitmap_weight()

[Matthew Wilcox <willy@infradead.org>]

On Sun, Aug 25, 2019 at 02:39:47PM +0300, Denis Efremov wrote:
> On 25.08.2019 09:11, Matthew Wilcox wrote:
> > On Sat, Aug 24, 2019 at 01:01:02PM +0300, Denis Efremov wrote:
> >> This patch open codes the bitmap_weight() call. The direct
> >> invocation of hweight_long() allows to remove the BUG_ON and
> >> excessive "longs to bits, bits to longs" conversion.
> > 
> > Honestly, that's not the problem with this function.  Take a look
> > at https://danluu.com/assembly-intrinsics/ for a _benchmarked_
> > set of problems with popcnt.
> > 
> >> BUG_ON was required to check that bitmap_weight() will return
> >> a correct value, i.e. the computed weight will fit the int type
> >> of the return value.
> > 
> > What?  No.  Look at the _arguments_ of bitmap_weight():
> > 
> > static __always_inline int bitmap_weight(const unsigned long *src, unsigned int nbits)
> 
> I'm not sure why it is INT_MAX then? I would expect in case we care only about arguments
> something like:
>  
> BUG_ON(longs >= UINT_MAX / BITS_PER_LONG);

People aren't always terribly consistent with INT_MAX vs UINT_MAX.
Also, bitmap_weight() should arguably return an unisnged int (it can't
legitimately return a negative value).

> >> With this patch memweight() controls the
> >> computation directly with size_t type everywhere. Thus, the BUG_ON
> >> becomes unnecessary.
> > 
> > Why are you bothering?  How are you allocating half a gigabyte of memory?
> > Why are you calling memweight() on half a gigabyte of memory?
> > 
> 
> No, we don't use such big arrays. However, it's possible to remove BUG_ON and make
> the code more "straight". Why do we need to "artificially" limit this function
> to arrays of a particular size if we can relatively simple omit this restriction?

You're not making a great case for changing the implementation of
memweight() here ...

> I don't know how the implementation of this optimization will look like in it's
> final shape, because of different hardware/compiler issues. It looks there are
> a number of different ways to do it https://arxiv.org/pdf/1611.07612.pdf, 
> http://0x80.pl/articles/sse-popcount.html.

The problem with using XMM registers is that they have to be saved/restored.
Not to mention the thermal issues caused by heavy usage of AVX instructions.

> However, if it will be based on popcnt instruction I would expect that
> hweight_long will also contain this intrinsics. Since version 4.9.2
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62011#c13 GCC knows of the
> false-dependency in popcnt and generates code to handle it

Ah!  Glad to see GCC knows about this problem and has worked around it.

> (e.g. xor https://godbolt.org/z/Q7AW_d) Thus, I would expect that it's
> possible to use popcnt intrinsics in hweight_long that would be natively
> optimized in all loops like "for (...) { res += hweight_long() }" without
> requiring manual unrolling like in builtin_popcnt_unrolled_errata_manual
> example of Dan Luu's optimization.

That might be expecting rather more from our compiler than is reasonable ...

> > 
> > Also, why does the trailer do this:
> > 
> >         for (; bytes > 0; bytes--, bitmap++)
> >                 ret += hweight8(*bitmap);
> > 
> > instead of calling hweight_long on *bitmap & mask?
> > 
> 
> Do you mean something like this?
> 
>         longs = bytes;
>         bytes = do_div(longs, sizeof(long));
>         bitmap_long = (const unsigned long *)bitmap;
>         if (longs) {
>                 for (; longs > 0; longs--, bitmap_long++)
>                         ret += hweight_long(*bitmap_long);
>         }
>         if (bytes) {
>                 ret += hweight_long(*bitmap_long &
>                                    ((0x1 << bytes * BITS_PER_BYTE) - 1));
>         }
> 
> The *bitmap_long will lead to buffer overflow here.

No it won't.  The CPU will access more bytes than the `bytes' argument
would seem to imply -- but it's going to have fetched that entire
cacheline anyway.  It might confuse a very strict bounds checking library,
but usually those just check you're not accessing outside your object,
which is going to be a multiple of 'sizeof(long)' anyway.

If we do something like this, we'll need to use an 'inverse' of that mask
on big-endian machines.  ie something more like:

	if (bytes) {
		unsigned long mask;
		if (_BIG_ENDIAN)
			mask = ~0UL >> (bytes * 8);
		else
			mask = ~0UL << (bytes * 8);
		ret += hweight_long(*bitmap_long & ~mask);
	}

Also we need a memweight() test to be sure we didn't get that wrong.