Without tiering, PortableMath's Sin took 25.5 ns against the runtime's
15.0 ns in Release; marked AggressiveOptimization it takes 17.2 ns (Atan2
18.1 against 17.0). Results are unchanged: the cross-platform digest test
passes in Release and Debug on Linux and in Release on Windows.
System.Math's Sin, Cos and Atan2 call the C runtime, and Windows' UCRT and
Linux's glibc differ in the last bit at many angles that are not quarter
turns. Nesting turns those bits into different tie and tolerance decisions,
so the same job nests differently on the desktop and on Linux (Project
Memory 518f3b58).
PortableMath ports fdlibm's Sin, Cos, Atan and Atan2, which use only IEEE
exact operations, and builds Tan, Asin and Acos on them with Sqrt. Results
stay within one ulp of the runtime (a few for the derived functions) and
hash to the same digest on Linux and on Windows (DesktopPC). Nothing calls
it yet.
aj