https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59833
--- Comment #16 from joseph at codesourcery dot com <joseph at codesourcery dot com> --- On Fri, 16 Feb 2018, egallager at gcc dot gnu.org wrote: > > powerpc failure of floating-point extensions to quiet signaling NaNs > > (because loads implicitly extend from float to double in a way that's > > defined as bit-manipulation rather than a convertFormat operation) is bug > > 56828. > > but bug 56828 was closed as INVALID... Maybe the test in that bug in fact only covered cases involving long double (which is what that closure was for). There is an issue with conversion from float to double on powerpc (FPRs always store values in double format, and a load of a float value is a bitwise operation that never raises exceptions but produces a double sNaN when loading a float sNaN - which can be stored back to float in memory as a float sNaN with no exceptions - meaning that converting a float value in registers to double is treated as a no-op). It's normally OK to do that, if the converted value will be used in double arithmetic - but if it's stored back to memory as a double without having gone through arithmetic other than abs / negate / copysign, it will incorrectly remain a double sNaN, and in that case, given -fsignaling-nans, something should be done to convert it explicitly to a qNaN (e.g. use an frsp instruction, which would do nothing to a value representable as float which is not a signaling NaN).
