Re: [Qemu-devel] [PATCH v2 23/32] arm/helper.c: re-factor recpe and add recepe_f16

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

On 02/08/2018 09:31 AM, Alex Bennée wrote:
> +float16 HELPER(recpe_f16)(float16 input, void *fpstp)
> +{
> +    float_status *fpst = fpstp;
> +    float16 f16 = float16_squash_input_denormal(input, fpst);
> +    uint32_t f16_val = float16_val(f16);
> +    uint32_t f16_sign = float16_is_neg(f16);
> +    int f16_exp = extract32(f16_val, 10, 5);
> +    uint32_t f16_frac = extract32(f16_val, 0, 10);
> +    uint64_t f64_frac;
> +
> +    if (float16_is_any_nan(f16)) {
> +        float16 nan = f16;
> +        if (float16_is_signaling_nan(f16, fpst)) {
> +            float_raise(float_flag_invalid, fpst);
> +            nan = float16_maybe_silence_nan(f16, fpst);
> +        }
> +        if (fpst->default_nan_mode) {
> +            nan =  float16_default_nan(fpst);
> +        }
> +        return nan;
> +    } else if (float16_is_infinity(f16)) {
> +        return float16_set_sign(float16_zero, float16_is_neg(f16));
> +    } else if (float16_is_zero(f16)) {
> +        float_raise(float_flag_divbyzero, fpst);
> +        return float16_set_sign(float16_infinity, float16_is_neg(f16));
> +    } else if (float16_abs(f16) < (1 << 8)) {
> +        /* Abs(value) < 2.0^-14 */

The pseudocode I'm looking at says 2.0^-16.  But I think the code is right --
this is checking for two zero bits at the top of a denormal, so that is
2.0^(-14-2).


> +        float_raise(float_flag_overflow | float_flag_inexact, fpst);
> +        if (round_to_inf(fpst, f16_sign)) {
> +            return float16_set_sign(float16_infinity, f16_sign);
> +        } else {
> +            return float16_set_sign(float16_maxnorm, f16_sign);
> +        }
> +        /* FP16 has it's own flag FZ16 flag which is in a separate fpst*/
> +    } else if (f16_exp >= 14 && fpst->flush_to_zero) {

(1) The comment is confusing.
    (a) It's placement within the previous IF is not helpful,
    (b) Why mention the separate fpst, begging the question of
        where it is?  We're using it, of course, so...

(2) The exponent is still biased, so this isn't 2.0^14 you're testing.

>      } else if (f32_exp >= 253 && fpst->flush_to_zero) {

E.g. the single-precision version tests (2^)126 + (bias)127.


r~