On Thu, 24 May 2018, Segher Boessenkool wrote: > Which semantics doesn't it have? (Other than the lack of non-default > rounding modes, or is that what you meant?)
It has problems with rounding modes, exceptions and various operations involving large or small arguments or results. Some of these apply even in round-to-nearest and when the arguments are exactly representable as double (spurious overflow from (long double) DBL_MAX / 3.0L, for example), which makes it unambiguous that a widening from DFmode to IFmode is never appropriate. (Of course, code doing such a widening would be bad for other reasons anyway, since it should only be widening from a non-hardware mode to a hardware one, and IFmode is never a hardware mode.) -- Joseph S. Myers jos...@codesourcery.com