On 9/21/13 12:16 AM, Eli Zaretskii wrote:
>> Date: Fri, 20 Sep 2013 19:49:00 -0700
>> From: Daniel Colascione <dancol@dancol.org>
>> Cc: Emacs development discussions <emacs-devel@gnu.org>
>>
>>>> +static inline
>>>> +EMACS_INT
>>>> +popcount_size_t(size_t val)
>>>> +{
>>>> +  EMACS_INT count;
>>>> +
>>>> +#if defined __GNUC__ && BITS_PER_SIZE_T == 64
>>>> +  count = __builtin_popcountll (val);
>>>> +#elif defined __GNUC__ && BITS_PER_SIZE_T == 32
>>>> +  count = __builtin_popcount (val);
>>>> +#elif defined __MSC_VER && BITS_PER_SIZE_T == 64
>>>> +# pragma intrinsic __popcnt64
>>>> +  count = __popcnt64 (val);
>>>> +#elif defined __MSC_VER && BITS_PER_SIZE_T == 32
>>>> +# pragma intrinsic __popcnt
>>>> +  count = __popcnt (val);
>>>> +#else
>>>> +  {
>>>> +    EMACS_INT j;
>>>> +    count = 0;
>>>> +    for (j = 0; j < BITS_PER_SIZE_T; ++j)
>>>> +      count += !!((((size_t) 1) << j) & val);
>>>> +  }
>>>> +#endif
>>>
>>> Why loop? See http://en.wikipedia.org/wiki/Hamming_weight.
>>
>> I didn't want to put a lot of effort into a code path we'll probably
>> never use.  Recall that if we're using icc or gcc or Visual C++ or
>> Clang, we'll be using a compiler intrinsic, which will probably compile
>> down to a single machine instruction.
>>
>> By the way: can someone test that the Visual C++ alternate actually
>> works? I don't have access to a Windows machine at the moment.
> 
> I don't see why it won't work, per documentation on this page:
> 
>   http://msdn.microsoft.com/en-us/library/bb385231%28v=vs.90%29.aspx
> 
> However, I think you will need to make usage of these intrinsics
> compiler version dependent.  GCC supports them starting from 3.4,
> whereas MSVC seems to support them since Studio 2008, i.e. _MSC_VER =
> 1500 or higher.

Fair enough.

> It is also not clear to me what will the MSVC intrinsic do if the
> binary ever runs on a CPU that doesn't support SSE4, the MSDN
> documentation seems to say that the results are unpredictable,
> i.e. that there's no fallback, like GCC has in libgcc.  So perhaps we
> should also guard that with a Windows version (assuming that old
> machines will only ever run Windows 9x).

Good point. SSE4 is much too recent to require. Making the cpuid check
shouldn't be too hard. I have no way to actually test the fallback,
though. (It's easy to test the fallback code, but not that easy to test
falling back to it.)