implement significand in native Julia

[simonbyrne]

significand utilised non-standard libm functions, which will be removed from the next release of Openlibm (JuliaMath/openlibm#85). This implements them in native Julia.

(I also changed some comparisons from x==zero(x) to x==0: LLVM now optimises these away (see #9030); this change is purely for notational convenience)

[kmsquire]

Any reason not to also add a version for Float16?

[simonbyrne]

@kmsquire No, only because we didn't have them before. Now done.

[quinnj]

This is great to have for the grisu code as well. We currently promote Float16s up because there wasn't a significand defined. I'll have to dig back into that code soon and do another pass.

[jiahao]

bump

+1

[simonbyrne]

Updated.

[ViralBShah]

Should we be adding more tests to ensure full coverage - especially since we have our own implementation (even though it is pretty simple)?

[simonbyrne]

Probably, I'll see what I can do.

[simonbyrne]

I got a bit carried away, and implemented general methods for significand, exponent and frexp across the different types. I also added some more tests.

[StefanKarpinski]

wow! very cool. feel free to merge whenever you feel confident that this is correct and passing tests.

[quinnj]

Awesome work as usual @simonbyrne. With all three defined for our float types, I can probably cut about 250 lines out of the grisu code. I'll have to do another pass to work that out once this gets merged.

[simonbyrne]

I can't get the osx build to pass, but this doesn't seem to be related.