With a conservative analysis it's possible to make an underestimation on the names one can remove. This shouldn't be done on object files but at a level where program flow information is still preserved. A flow-independent analysis using a superset lattice over the callable names should do the trick for a first-order implementation that will look good enough compared to shipping 37mb for google authentication. Further you could make it flow-sensitive for a more accurate analysis with the caveat that the runtime will be significantly longer.