tree 99c06337e4132476c0bd9a82140bbdd4e7605920
parent 65b3e80f46f80fca3e1fb857620b6b3574687aa9
author Hans Boehm <hboehm@google.com> 1424904793 -0800
committer Hans Boehm <hboehm@google.com> 1425669724 -0800

Fix cos() prescaling, clean up inverse, add tests

Cosine arguments close to pi were not correctly prescaled, resulting
in a significantly more expensive computation.

The choice of binary search vs interpolation in the inverse function
algorithm was somewhat random, in that it sometimes counted binary
search steps as "successful" and sometimes not.  This caused more binary
search steps than intended, which unfortunately sometimes improved matters.
This fixes the unpredictability and adjust the heuristics to get
performance on tan(atan(large)) back to roughly where we started.

Increase precision a bit more aggressively in the inverse function to
potentially allow superlinear convergence in more cases.  I couldn't
reconstruct any meaningful rationale for the old formula.

Add SlowCRTest.  Reorganize and fix tests to be consistent with Android
guidelines, more or less.  Update COPYRIGHT.txt to reflect the fact
that we own SlowCRTest.

Change-Id: I8e2d3d11672ddfe24fd8ad0c74d05ef11b12cb03