@marshallward wrote in #1 (comment):

I would like to see greater support for bit-reproducible numerical operations. This is a very high priority for us since our models are used in weather and climate forecasting, and much of our time is devoted to bit reproducibility of our numerical calculations.

A common problem is intrinsic Fortran reduction operations like sum(), where the order is ambiguous (deliberately, one might say), and therefore not reproducible. A more serious problem for us is transcendental functions, like exp() or cos(), which will give different results for different optimizations, and we typically cannot say where it was invoked (libm? Vendor library? etc.).

A standard library may be a place to provide bit-reproducible implementations.