[mypyc] Precompute set literals for "in" ops against / iteration over set literals

[ichard26]

Towards mypyc/mypyc#726. (There's a Python compatibility bug that needs to be fixed before the issue can be closed.)

For example, the set literals here are now precomputed as frozensets at module initialization.

x in {1, 2.0, "3"}
x not in {1, 2.0, "3"}

for _ in {1, 2.0, "3"}:
    ...

Set literal items supported:

• Anything supported by irbuild.constant_fold.constant_fold_expr()
  • String and integer literals
  • Final references to int/str values
  • Certain int and str unary/binary ops that evaluate to a constant value
• None, True, and False
• Float, byte, and complex literals
• Tuple literals with only items listed above

Results (using gcc-9 on 64-bit Ubuntu)

Master @ 98cc165

running in_set
..........
interpreted: 0.495790s (avg of 5 iterations; stdev 6.8%)
compiled:    0.810029s (avg of 5 iterations; stdev 1.5%)

compiled is 0.612x faster

running set_literal_iteration
.........................................................................................
interpreted: 0.020255s (avg of 45 iterations; stdev 2.5%)
compiled:    0.016336s (avg of 45 iterations; stdev 1.8%)

compiled is 1.240x faster

This PR

running in_set
..........
interpreted: 0.502020s (avg of 5 iterations; stdev 1.1%)
compiled:    0.390281s (avg of 5 iterations; stdev 6.2%)

compiled is 1.286x faster

running set_literal_iteration
..............................................................................................
interpreted: 0.019917s (avg of 47 iterations; stdev 2.2%)
compiled:    0.007134s (avg of 47 iterations; stdev 2.6%)

compiled is 2.792x faster

Benchmarks can be found here: mypyc/mypyc-benchmarks#32

[ichard26]

FYI, I wrote the PR description to be used as the commit message when this is squash merged into main. Please copy and paste it and save yourself some work! :)

[ichard26]

I'll push a functional work around to the mypyc bug that's breaking the build sometime later, but for those wondering "do people really loop over set literals often?" here's a GitHub search query for you (there's even an example in Black's source code 🙂).

[JukkaL]

Looks great, just a few minor comments. I like that this fixes some important cases where mypyc can generate code slower than CPython.

[JukkaL]

Style nit: Instead of having the assertions in driver.py, could you add them into the main [case ...] section inside test_* functions?

I'd like to move towards writing tests this way, since it makes larger test cases easier to read, and there's less risk that contributors forget that code in driver.py doesn't get compiled.

[ichard26]

Makes sense, PTAL 9cdfc98.