Performance regression in base64-bytestring caused by dmd-anal regression.

Summary

base64-bytestring is about twice as slow on ghc_head than on 8.8

See also the discussion here: https://github.com/emilypi/base64/issues/3

This seems like it could be a result of the NCG changes I made recently.

Steps to reproduce

Run the benchmark suite of https://github.com/andrewthad/base64-bytes

Expected behavior

Performance should be equal or better.

Environment

Various. Tested on Linux and Windows