IEEE-754 recommended functions

[simonbyrne]

As I was enjoying the entertaining read that is the IEEE-754 standard for floating point arithmetic, I noticed they recommend some functions (section 9.2), of which we currently lack:

• [ ] exp2m1: 2^x-1
• [ ] exp10m1: 10^x-1
• [ ] log2p1: log2(1+x) (might want to call this log21p, to match log1p?)
• [ ] log10p1: log10(1+x) (similarly, log101p?)
• [ ] rsqrt: 1/sqrt(x)
• [ ] compound: (1+x)^n (pow1p?)
• [ ] rootn: x^(1/n)
• [ ] various inverse trig-pi functions, e.g. atanpi: atan(x)/pi (somewhat confusingly, not all of these are mentioned in the table, although they are referred to in the text below).

There's also three different "pow" functions: pown for integer exponents, two for handling floating-point exponents pow and powr which seem to differ mostly in how signals are raised, so we could probably wait until #2976 is completed.

[andrioni]

I was thinking just today about how to deal with the rest #2976, namely the flags, but I haven't yet found a nice enough solution.

[lobingera]

Just as a comment: not everyone has IEEE library access, you are referencing a document that has a price tag.

[JeffBezanson]

I was able to find a free copy somewhere.

[lobingera]

I have f.e. (due to my job) IEEE library access and i guess a lot of universities also have, but still...

[johnmyleswhite]

To obey US laws, Julia issues should probably not post links to free copies of things that are under copyright.

[lobingera]

... but this is only a side discussion. I remember this operations to be available on FP HW and the question is rather, does LLVM support them directly or can we make julia recognize them?

[simonbyrne]

@lobingera I agree it's ridiculous that such a widely-used standard comes with a price tag, but if it's any consolation, it's even more expensive to buy it from ISO

[oscardssmith]

It's probably worth pointing out that as of Julia 1.8 we do have special paths for Float^Int which is the pown part of this.