Re: [PATCH RFC] target/arm: Implement SVE2 FLOGB

[Richard Henderson]

On 4/30/20 10:20 AM, Stephen Long wrote:
> +DO_ZPZ_FP(flogb_s, float32, H1_4, float32_log2)
> +DO_ZPZ_FP(flogb_d, float64,     , float64_log2)

Please read the instruction description more carefully.  The result is not the
full log2 of the input:


> This instruction returns the signed integer base 2 logarithm of each 
> floating-point input element | X | after normalization.  This is the unbiased 
> exponent of X used in the representation of the floating-point value, such 
> that, for positive X , X = significand × 2 exponent.

Please look at Library pseudocode for aarch64/functions/sve/FPLogB in the 
manual.

You then use the helpers from softfloat.h like so

    if (float32_is_normal(x)) {
        // extract exponent and remove bias
        return extract32(x, 23, 8) - 127;
    } else if (float32_is_infinity(x)) {
        return INT32_MAX;
    } else if (float32_is_any_nan(x) || float32_is_zero(x)) {
        return INT32_MIN;
    } else {
        // denormal
        // extract fraction, normalize vs 2**-127.
        int shift = 9 - clz32(extract32(0, 23));
        return -127 - shift + 1;
    }


r~