On 3/13/21 3:48 PM, Richard Henderson wrote:
> On 3/11/21 2:07 PM, Eric Blake wrote:
>> + /*
>> + * Add in a fudge-factor (2^53 when double is IEEE format) for
>> + * all scales less than P (2^50), so that things like
>> + * 12.345M with unit 1000 produce 12345000 instead of
>> + * 12344999.
>> + */
>> + if (mul > 1e49) {
>
> The comment says less than, the code says greater than.
Shoot. A demonstration that I did not have an environment that actually
reproduced the bug (and my request for help in figuring out how to kick
off a gitlab CI run that would catch it).
>
>
>> An alternative patch might be writing (uint64_t)(fraction * mul + 0.5)
>> (that is, introduce the fudge factor after the multiplication instead
>> of before). Preferences?
>
> I think I would prefer this, or for further rounding error reduction,
> fma(fraction, mul, 0.5).
Indeed, fma() sounds a bit nicer at minimizing the chance for double
rounding errors.
--
Eric Blake, Principal Software Engineer
Red Hat, Inc. +1-919-301-3226
Virtualization: qemu.org | libvirt.org
