Jason, If the software or circuitry is IEEE 758-2008 compliant, all those quiet NaNs are usable. Most vendors select one or two of them (for quiet 64bit nans, usually 0x7ff8000000000000 and/or 0xfff8000000000000) and do everything NaN related with, say, those two. I hope vendors are not saying they have a Standards complying product when that is untrue. About compliance and NaNs, my impression is "you don`t have to use them, but they are expected to be present." They may play by different rules.
Stuart, The horse and arriage similie for a quiet nan makes sense to me. Travelling in unfamiliar places, occasionally noticing something of interest .. or getting a reminder .. and gathering some small, revealing information to place it in the carriage knowing that and it will arrive with me when the horse returns home. As I read it (his paper, not my redaction), William Kahan expressed an abiding regard for the efficacy and utillity of quiet NaNs as a numerical software engineers' participatory tool. That's what prompted me to to write the module. Thank you both for the thoughts. On Monday, August 3, 2015 at 12:56:57 PM UTC-4, Jason Riedy wrote: > > And Jeffrey Sarnoff writes: > > AFAIK Julia and most other languages use one or two of each in > > most circumstances. > > And many chips produce only one, the platform's "canonical" NaN. > Some pass one of the argument NaNs through but rarely will > specify which. > >