The acos function in the NumberDecimal library has a bug for special-case argument handling, where acos(-1.0) returns 0 rather than Pi. For
SERVER-32930 we are implementing the $acos expression and need this fixed. The new IntelRDFPMathLib20U2 June 2018 update of the library has this big fix, but seems to have either introduced a new issue in code changes for the power function, or exposes optimization bugs in GCC. So it looks like SERVER-37998 might take longer to do than expected.
So, we should for now just cherry-pick this change and associated test cases in readtest.in while we wait for upgrade of the library as a whole.