Re: [RFC PATCH 2/6] softfloat: add int128_to_float* conversion methods

[Richard Henderson]

On 3/28/22 14:14, matheus.ferst@eldorado.org.br wrote:
> From: Matheus Ferst <matheus.ferst@eldorado.org.br>
>
> Based on parts_sint_to_float, implements parts_sint_to_float2 that
> receives a 128-bit signed integer via int64_t and uint64_t arguments.
>
> Signed-off-by: Matheus Ferst <matheus.ferst@eldorado.org.br>
> ---
>   fpu/softfloat-parts.c.inc | 37 +++++++++++++++++++++++++++++++++++++
>   fpu/softfloat.c           | 30 ++++++++++++++++++++++++++++++
>   include/fpu/softfloat.h   |  3 +++
>   3 files changed, 70 insertions(+)
>
> diff --git a/fpu/softfloat-parts.c.inc b/fpu/softfloat-parts.c.inc
> index 0bbecf835f..5f7f107a0d 100644
> --- a/fpu/softfloat-parts.c.inc
> +++ b/fpu/softfloat-parts.c.inc
> @@ -1196,6 +1196,43 @@ static void partsN(sint_to_float)(FloatPartsN *p, 
> int64_t a,
>       p->frac_hi = f << shift;
>   }
>   
> +static void partsN(sint_to_float2)(FloatPartsN *p, int64_t hi, uint64_t lo,
> +                                   int scale, float_status *status)
> +{
> +    uint64_t f = hi;
> +    int shift;
> +
> +    if (hi == 0) {
> +        parts_uint_to_float(p, lo, scale, status);
> +    } else {

We should also defer "small" negative numbers.

       if (hi == -1) {
           parts_uint_to_float(p, -lo, scale, status);
           p->sign = true;
           return;
       }

That should ensure...

> +        memset(p, 0, sizeof(*p));
> +        p->cls = float_class_normal;
> +        if (hi < 0) {
> +            lo = -lo;
> +            f = ~f + !lo;
> +            p->sign = true;
> +        }
> +        if (f != 0) {
> +            shift = clz64(f);
> +        } else {
> +            shift = 64 + clz64(lo);
> +        }

this case

> +        scale = MIN(MAX(scale, -0x10000), 0x10000);
> +
> +        p->exp = 127 - shift + scale;
> +
> +        if (shift >= 64) {
> +            f = lo;
> +            lo = 0;
> +            shift -= 64;
> +        }

and this case don't happen.

> +        p->frac_hi = shl_double(f, lo, shift);
> +        if (N > 64) {
> +            p->frac_lo = shl_double(lo, 0, shift);

Same comment about shl_double w/ 0.

Alternately, rewrite the whole thing in terms of uint_to_float2:

    if (hi >= 0) {
        uint_to_float2(p, hi, lo, scale, status);
    } else {
        lo = -lo;
        hi = ~hi + !lo;
        uint_to_float2(p, hi, lo, scale, status);
        p->sign = true;
    }


r~