bug#22689: [PATCH] Add logcount

[Eli Zaretskii]

> Date: Sat, 30 Sep 2017 09:16:36 -0400
> From: Mark Oteiza <mvoteiza@udel.edu>
> Cc: 22689@debbugs.gnu.org
> 
> On 30/09/17 at 02:42pm, Eli Zaretskii wrote:
> > > +#if defined (__POPCNT__) && defined (__GNUC__) && (__GNUC__> 4 || 
> > > (__GNUC__== 4 && __GNUC_MINOR__> 1))
> > > +#define HAVE_BUILTIN_POPCOUNTLL
> > > +#endif
> > 
> > Where does __POPCNT__ definition come from?  I don't see it in my
> > GCC's "gcc -dM" output.
> > 
> > As for the rest of the condition, I think you should use GNUC_PREREQ,
> > like this:
> > 
> >  #if GNUC_PREREQ (4, 1, 0)
> 
> I guess it comes from nowhere, that condition and the two helper
> functions come from here
> https://rosettacode.org/wiki/Population_count#C

I only see that symbol in the GCC sources, which I don't think are
relevant here.

If you drop the __POPCNT__ part, does the code still work for you?

> > > +static uint64_t
> > > +bitcount64 (uint64_t b)
> > > +{
> > > +  b -= (b >> 1) & 0x5555555555555555;
> > > +  b = (b & 0x3333333333333333) + ((b >> 2) & 0x3333333333333333);
> > > +  b = (b + (b >> 4)) & 0x0f0f0f0f0f0f0f0f;
> > > +  return (b * 0x0101010101010101) >> 56;
> > 
> > Shouldn't these constants have a ULL suffix, to make sure they are not
> > truncated to a 32-bit int?
> 
> Probably, I don't know--I'm out of my comfort zone here.

I'd use the suffixes.

> > > +#else  /* HAVE_BUILTIN_POPCOUNTLL */
> > > +  if (XINT (value) <= UINT_MAX)
> > > +    XSETINT (res, bitcount32 (XUINT (value)));
> > > +  else if (XINT (value) <= ULONG_MAX)
> > > +    XSETINT (res, bitcount64 (XUINT (value)));
> > 
> > The comparisons against Uxxx_MAX seem to assume that VALUE is
> > unsigned, but that's not guaranteed, right?
> 
> True.  Should I instead be doing
> 
>   XINT (value) <= xxx_MAX &&
>   XINT (value) >= xxx_MIN
> 
> or might there be a better check?  For negative values popcount
> typically counts ones of the two's complement

I'd just assign to an unsigned temporary, IIUC the semantics.

Btw, on Windows, a long is a 32-bit type, so I think we need
to check against ULONG_LONG_MAX as well.