Replacing CRlibm with CORE-MATH

[OlivierHnt]

In a recent triage call, we discussed replacing CRlibm with CORE-MATH, which provides a more complete set of correctly rounded functions.

On particular appeal is that they give a correctly rounded version of pow, which should fix long standing performance issues for powers in IntervalArithmetic.jl.

@giordano I noticed that you created the CRlibm_jll binaries with Yggdrasil.jl. I tried to reproduce this on my computer today for core-math, but I am out of my depth here and could not make it work. The repository on GitLab is not a library per se, but it contains all the C functions. Do you think we can compile CORE-MATH into some CoreMath_jll binaries as easily?

[lucaferranti]

havent forgotten, working hard hardly working

[dpsanders]

I'm not sure that Core-Math can be used for our purporse, since it apparently requires using fesetround to specify the rounding mode required, which has caused issues with LLVM in the past

[dpsanders]

Also changing the processor's rounding mode is not thread-safe.

I suggest reaching out to Paul Zimmermann to discuss this. (W e had an email conversation about it, but I dropped the ball.)

[dpsanders]

Hmm, apparently the state of fesetround is supposed to be per thread. But this will need careful implementation and testing!

[dpsanders]

The other approach with this is performance. Changing the rounding mode clears out the processor pipeline (or at least used to?) so will be much slower (?)

[dpsanders]

If I remember correctly he suggested using the prevfloat / next float approach for performance...!

[OlivierHnt]

Interesting, thanks for pointing this out.
Using Core-Math does not solve all the performance issues, but it does (or seems to) avoid using MPFR.
I will have to take a closer look at fesetround, I do not know much on the inner workings of these things. I do hope that a modern and maintained library for correctly rounded functions can be incorporated into Julia 🙂.