should we forward our complex math to libm?

[llvmbot]

Bugzilla Link	21504
Version	unspecified
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@hfinkel,@mclow

Extended Description

Is there any reason we're not delegating the math in to <complex.h>? I don't believe any of us look at this code too closely (or very often), so delegating to the lower library means we wouldn't have to worry about it any more.

It looks like we don't have any tests for accuracy of these functions, whereas the math libraries will be focused on that.

[llvmbot]

(in particular it looks like the implementations of the inverse trig are thousands of lines shorter than the FreeBSD libm ones, which makes me suspect that that they're not taking the intricacies of floating point math into account and are assuming they can get away with the math textbook definitions. even libm only claims 4 ulps!)

[hfinkel]

I can certainly support forwarding some of the complicated functions to libm. The simple things likely to inlined should stay in the header... and we should have tests for these if we don't already.

Realistically, we should have some tests for the other things too... at least in so far as we make sure they they're calling the correct function ;)