On Thu, 10 Aug 2017, Jackson Woodruff wrote:
> > We also change:
> > x / (- y) -> (-x) / y
> > Which requires -fno-trapping-math.
I don't see why that requires -fno-trapping-math. The exceptions should
be identical from both variants, as should the result, as far as defined
by C or IEEE 754 (it's possible it might affect the sign of a NaN result
on some architectures, but only where that sign is unspecified in C and
IEEE 754 terms, so such a change is OK).
Joseph S. Myers