Re: [PATCH 6/8] softfloat: Implement float128_muladd

[David Hildenbrand]

On 24.09.20 15:30, Richard Henderson wrote:
> On 9/24/20 12:56 AM, David Hildenbrand wrote:
>> I do wonder if a type for Int256 would make sense - instead of > manually 
>> passing these arrays.
> 
> I could do that.  It does name better, I suppose, in passing.  So long as
> you're happy having the guts of the type be public, and not wrapping 
> everything
> like we do for Int128...

We can do that once we have hardware+compiler support for 256bit ;)

> 
> Either is better than the softfloat style, which would pass 12 arguments to
> these helpers... ;-)

Indeed

> 
>>> +static void shortShift256Left(uint64_t p[4], unsigned count)
>>> +{
>>> +    int negcount = -count & 63;
>>
>> That's the same as "64 - count", right? (which I find easier to get)
> 
> In this case, yes.
> 
> Of course, more hosts have a "neg" instruction than they do a
> "subtract-from-immediate" instruction.  When the shift instruction only
> examines the low N bits, the "& 63" is optimized away, so this can result in 1
> fewer instruction in the end.
> 
> Which is why I almost always use this form, and it's already all over the code
> inherited from upstream.
> 
>> Wow, that's a beast :)
> 
> But not much worse than muladd_floats(), I'm pleased to say.

Definitely!


-- 
Thanks,

David / dhildenb