https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67011
Jonathan Wakely <redi at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2018-05-14
Ever confirmed|0 |1
--- Comment #4 from Jonathan Wakely <redi at gcc dot gnu.org> ---
(In reply to Matthias Kretz from comment #3)
> The problem is that (result_type(1) - __aurng()) must never be zero (because
> std::log might raise FE_DIVBYZERO for std::log(0)). If __aurng() is a
> distribution guaranteed to exclude result_type(1) then this works fine.
> AFAIK std::generate_canoncial<float> sometimes generates a 1 (cf.
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63176).
PR 63176 is fixed now.
> Also, I'm wondering whether the 1 - rng wouldn't reduce the randomness.
> Consider:
> std::log(numeric_limits<float>::min()) -> -87.3365478515625
> std::log(1.f - nextafter(1.f, 0.f)) -> -16.635532379150390625
> IIUC, the effect of log(1 - rng) is that the distance between random values
> from std::exponential_distribution is considerable larger than for log(rng).
Hmm, let's confirm the bug for that, instead of the division-by-zero that
should be fixed.
