https://bugs.kde.org/show_bug.cgi?id=491667
--- Comment #3 from Mark Wielaard <m...@klomp.org> ---
And the following comment at the top:
/* KNOWN LIMITATIONS 2014-Nov-16
* Correctness: FMAXNM, FMINNM are implemented the same as FMAX/FMIN.
Also FP comparison "unordered" .. is implemented as normal FP
comparison.
Both should be fixed. They behave incorrectly in the presence of
NaNs.
FMULX is treated the same as FMUL. That's also not correct.
* Floating multiply-add (etc) insns. Are split into a multiply and
an add, and so suffer double rounding and hence sometimes the
least significant mantissa bit is incorrect. Fix: use the IR
multiply-add IROps instead.
* FRINTX might be need updating to set the inexact computation FPSR flag
* Ditto FCVTXN. No idea what "round to odd" means. This implementation
just rounds to nearest.
*/
--
You are receiving this mail because:
You are watching all bug changes.
