Re: [PATCH v2 06/10] softfloat: Implement float128_muladd

[Richard Henderson <richard.henderson@linaro.org>]

On 10/16/20 9:31 AM, Alex Bennée wrote:
>> +static void float128_unpack(FloatParts128 *p, float128 a, float_status *status)
>> +{
>> +    p->sign = extractFloat128Sign(a);
>> +    p->exp = extractFloat128Exp(a);
>> +    p->frac0 = extractFloat128Frac0(a);
>> +    p->frac1 = extractFloat128Frac1(a);
> 
> Here we are deviating from the exiting style, it would be nice if we
> could separate the raw unpack and have something like:
> 
> static const FloatFmt float128_params = {
>     FLOAT_PARAMS(15, 112)
> }
> 
> static inline FloatParts128 unpack128_raw(FloatFmt fmt, uint128_t raw)
> {
>     const int sign_pos = fmt.frac_size + fmt.exp_size;
> 
>     return (FloatParts128) {
>         .cls = float_class_unclassified,
>         .sign = extract128(raw, sign_pos, 1),
>         .exp = extract128(raw, fmt.frac_size, fmt.exp_size),
>         .frac1 = extract128_lo(raw, 0, fmt.frac_size),
>         .frac2 = extract128_hi(raw, 0, fmt.frac_size),
>     };
> }
> 
> So even if we end up duplicating a chunk of the code the form will be
> similar so when we side-by-side the logic we can see it works the same
> way.

I suppose, but unlike the smaller fp formats, we won't be able to reuse this.
Even if we pull in the x86 80-bit format and the m68k 96-bit format, there are
a number of fundamental differences.  E.g. the implicit bit

>> +        /* Add the implicit bit. */
>> +        p->frac0 |= UINT64_C(0x0001000000000000);

is not present in the x86 and m68k formats.

Finally, I'm continuing to use the existing Berkeley packing logic.  Which a
bit persnickety with where that implicit bit goes.  Our smaller formats put the
implicit bit at bit 62.


r~