gh-118761: Improve import time of tomllib

[hukkin]

This is a port of two commits in Tomli repository (the diff of these commits being hukkin/tomli@42a570d...1da01ef).

This removes runtime overhead of string, typing, and tomllib._types imports from import tomllib.

main branch

$ hyperfine --warmup 8 "./python -c 'import tomllib'" "./python -c 'pass'"
Benchmark 1: ./python -c 'import tomllib'
  Time (mean ± σ):      70.3 ms ±   1.8 ms    [User: 66.0 ms, System: 4.3 ms]
  Range (min … max):    68.6 ms …  76.8 ms    41 runs
 
Benchmark 2: ./python -c 'pass'
  Time (mean ± σ):      26.2 ms ±   0.7 ms    [User: 23.3 ms, System: 3.0 ms]
  Range (min … max):    25.1 ms …  29.7 ms    110 runs
 
Summary
  './python -c 'pass'' ran
    2.69 ± 0.10 times faster than './python -c 'import tomllib''

PR branch

$ hyperfine --warmup 8 "./python -c 'import tomllib'" "./python -c 'pass'"
Benchmark 1: ./python -c 'import tomllib'
  Time (mean ± σ):      48.0 ms ±   0.7 ms    [User: 43.4 ms, System: 4.7 ms]
  Range (min … max):    46.8 ms …  50.2 ms    60 runs
 
Benchmark 2: ./python -c 'pass'
  Time (mean ± σ):      26.1 ms ±   0.4 ms    [User: 23.5 ms, System: 2.7 ms]
  Range (min … max):    25.3 ms …  27.4 ms    110 runs
 
Summary
  './python -c 'pass'' ran
    1.84 ± 0.04 times faster than './python -c 'import tomllib''

• Issue: Improve import time of various stdlib modules #118761

[AA-Turner]

I think that this should technically be a protocol similar to typeshed's SupportsRead. But for a slightly more radical proposal, would you support removing all the type annotations from tomllib? The ones in typeshed are used by all of the type checkers, and there's a risk of going out of sync, or unhelpful 'improvements' to adjust the (unused) type hints in Lib/tomllib.

[hukkin]

I think that this should technically be a protocol similar to typeshed's SupportsRead.

This is true. I changed this annotation because the the diff I was porting included an updated from typing import line that uses IO not BinaryIO. A trade-off was made choosing this annotation, and so far nobody complained about it.

But for a slightly more radical proposal, would you support removing all the type annotations from tomllib?

I wouldn't support it, at least not before Python 3.10 EOL date. Type annotations were inherited from Tomli, and having them here makes diffs smaller and easier to work with -> syncing changes such as this PR is a bit easier for me.

That said, no big deal for me. If folks want to remove the annotations, I'd be fine with it.

[AA-Turner]

Fair enough -- I'd remove them personally, but I'm not the maintainer of tomllib and being able to easily synchronise updates is a good reason -- we might re-visit in a couple of years!

IO[bytes] is probably fine for now the given this.

[vstinner]

cc @encukou who added the module in Python 3.11 (PEP 680).