Re: [RFC PATCH 4/6] softfloat: add float*_to_int128 conversion methods

[Richard Henderson]

On 3/28/22 14:14, matheus.ferst@eldorado.org.br wrote:
> +static int64_t partsN(float_to_sint2)(FloatPartsN *p, FloatRoundMode rmode,
> +                                      int scale, float_status *s, uint64_t *lo)
> +{
> +    int flags = 0;
> +    uint64_t hi;

Similar comment about 'lo' vs '*lo'.

> +        if (p->exp <= DECOMPOSED_BINARY_POINT) {
> +            hi = 0;
> +            *lo = p->frac_hi >> (DECOMPOSED_BINARY_POINT - p->exp);
> +        } else if (p->exp <= 127) {
> +            int shift = 127 - p->exp;
> +            hi = shr_double(0, p->frac_hi, shift);

Same comment about double-shift w/ 0.

> +            if (N > 64) {
> +                *lo = shr_double(p->frac_hi, p->frac_lo, shift);
> +            } else {
> +                *lo = shr_double(p->frac_hi, 0, shift);
> +            }
> +        } else {
> +            hi = UINT64_MAX;
> +            *lo = UINT64_MAX;
> +        }
> +        if (p->sign) {
> +            if (hi < INT64_MIN || (hi == INT64_MIN && *lo == 0)) {
> +                *lo = -*lo;
> +                hi = ~hi + !*lo;
> +            } else {
> +                flags = float_flag_invalid | float_flag_invalid_cvti;
> +                hi = INT64_MIN;
> +                *lo = 0;
> +            }
> +        } else if (hi > INT64_MAX) {
> +            flags = float_flag_invalid | float_flag_invalid_cvti;
> +            hi = INT64_MAX;
> +            *lo = UINT64_MAX;
> +        }

Hmm.  It seemed easy to arrange the code this way with just a uint64_t, but here I think 
it might be worth detecting overflow earlier, via exp.

    if (p->exp < 127) {
        /* No overflow possible */
        int shift = 127 - p->exp;
        if (shift >= 64) {
            hi = 0;
            lo = p->frac_hi >> (shift - 64);
        } else {
            hi = p->frac_hi >> shift;
            lo = shr_double(p->frac_hi, N > 64 ? p->frac_lo : 0, shift);
        }
        if (p->sign) {
            lo = -lo;
            hi = ~hi + !lo;
        }
        break;
    }
    /* The only valid 127-bit number is UINT128_MIN. */
    if (p->exp == 127 &&
        p->sign &&
        p->frac_hi == DECOMPOSED_IMPLICIT_BIT &&
        (N <= 64 || p->frac_lo == 0)) {
        hi = INT64_MIN;
        lo = 0;
        break;
    }
    /* Overflow. */
    flags = ...;
    if (p->sign)
        ...


r~