John Maddock wrote:
Well, I'm biased :-) But I'd characterise it like this: the version that will be in Boost-1.35 uses only portable C++ code, and will forward to the platforms native version if there is one. It's also been tested on all the platforms we regression test on, and the only failure I know of turned out to be a bug in numeric_limits (confirmed by IBM). On the other hand, Johan's implementation peeks at the bits of the value to determine it's type, which is *much* faster for native floating point types, but *maybe* less portable for some platforms, and doesn't scale to user defined floating point types. Hopefully we'll end up with a combination of the best bits of the two versions in the end - and Johan's code review is coming up at the end of Feb if you're interested BTW.
HTH, John.
I fully agree with John's description. --Johan