I wanted to run some perf checks on this, but don't remember any typealias-heavy codebase to use. Selfcheck almost definitely won't show anything - there are barely any type aliases in mypy code - and isolated benchmarks won't be representative in this case. Any suggestions?

@sterliakov IIRC altair from mypy primer is quite sensitive to type aliases.

Timed for _ in {1..8}; do time mypy altair tests >/dev/null; rm -rf .mypy_cache/; done using mypyc-compiled version at -O3 (gcc 13.3.0 on linux) and the same dependencies as installed by primer. First two warm-up runs were discarded. Results are given as $\text{mean} \pm \text{sd}$ in seconds (computed with stdlib statistics.fmean and statistics.stdev, resp.).

• PR: $31.2 \pm 0.4$
• is_recursive_pair: $30.9 \pm 0.3$
• master: $31.3 \pm 0.4$

There are 210 matches of \bTypeAlias\b in altair, and they do indeed seem to use them heavily, so this was a good benchmark, thanks!

So I'm switching to is_recursive_pair because it is not significantly slower (and even appears to be faster which is reasonable), and there's no significant perf hit from this PR compared to master - all numbers are equal within 1 SD.

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