Filter to possible package paths before trying to resolve a module.

[falsedrow]

With a long sys.path (it's got 300 entries), this removes 94% of stat syscalls from running mypy. With all the filesystem caching, that's only a small time savings, though it will depend on your filesystem. Local benchmarks showed a 20% time savings but they're pretty noisy from all the I/O.

[hauntsaninja]

Awesome stuff! I pushed two commits to try and get tests passing, I'll also run this through the benchmarking setup that I had going :-)

[hauntsaninja]

Okay, I benchmarked ac79825 as the following on mypy -c 'import torch' (see #17948 for description of the environments):

• 1.09x faster in long environment
• 1.06x faster in openai environment
• 1.42x faster in long environment (incremental)
• 1.22x faster in openai environment (incremental)

I'll check the extra normpath is slowing us down, if so should be very easy to save (update: looks like this might be a few 100ms, so a couple more percent)

Doing this also uncovered a minor regression from #15347 , which I shall try to fix... (update: I think best done in a new PR)

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅