https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69132
Wang Xuancong <xuancong84 at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |RESOLVED
Resolution|--- |INVALID
--- Comment #4 from Wang Xuancong <xuancong84 at gmail dot com> ---
Thanks!
I have just learned one important thing today, that all *rcp*-based
instructions need at least one Newton-Rhapson iteration to restore accuracy,
otherwise, it loses 50% accuracy. Take note that the original FLOAT32 has
23-bits in the significand. Without any Newton-Rhapson, you are left with only
11-bits. Ouch!
