Optimizing the C-API

[ericsnowcurrently]

The general idea of optimizing the C-API is relevant to the overall goal of making CPython faster. Furthermore, though I don't have good data on this, my gut tells me the impact of optimizing the C-API would be meaningful.

This is a follow-up to #406. While that case might not be worth pursuing, it prompted some broader questions for me:

• do we have usage statistics for all the C-API functions? (via static analysis, e.g. of top 5000 PyPI; via profiling running apps)
• do we know how much (relative) time Python apps spend in the C-API? (see #411)
• do we have a good sense of what possible optimizations are possible relative to the C-API as a whole?
• what about for specific functions?
• do we have a clear picture of the high-level optimization categories?

For that last item, here are some examples of what I mean:

• optimize the implementation of an existing API function
• add a new version of an existing API function (modified signature enables new optimizations)
• add a replacement for an existing API function (or set of functions) that takes a fundamentally different, but more efficient, approach to providing useful capability
• redesign a struct (e.g. re-arrange its fields)
• hide/eliminate API functions/types that prevent optimizing the runtime internals