25% performance regression on `main`

[jacobtylerwalls]

#1442 adds about 10-25% time to a pylint run, depending on how much inference would have taken place otherwise, because it infers every single name.

On my machine, with cProfile:
astroid lints in 46.2s on main
astroid lints in 36.6s with reverting #1442

Are we okay with that? If so, should we short-circuit before inferring in the pathlib brain if the interpreter is below 3.10? (#1442 only has to do with 3.10 behavior.)

[Pierre-Sassoulas]

I think we need to find a compromise here, inferring pathlib correctly is important, but not "25% slower does not matter" kind of important. I don't have a clear understanding of where we can "cut corner" or make the inference faster though.

[jacobtylerwalls]

Something like this:

diff --git a/astroid/brain/brain_pathlib.py b/astroid/brain/brain_pathlib.py
index 8ff3310b..c79f5b84 100644
--- a/astroid/brain/brain_pathlib.py
+++ b/astroid/brain/brain_pathlib.py
@@ -8,6 +8,7 @@ from collections.abc import Iterator
 
 from astroid import bases, context, inference_tip, nodes
 from astroid.builder import _extract_single_node
+from astroid.const import PY310_PLUS
 from astroid.exceptions import InferenceError, UseInferenceDefault
 from astroid.manager import AstroidManager
 
@@ -18,6 +19,8 @@ Path
 
 
 def _looks_like_parents_subscript(node: nodes.Subscript) -> bool:
+    if not PY310_PLUS:
+        return False
     if not (
         isinstance(node.value, nodes.Name)
         or isinstance(node.value, nodes.Attribute)
diff --git a/tests/unittest_brain_pathlib.py b/tests/unittest_brain_pathlib.py
index 4c9eb1b3..55c0c6a8 100644
--- a/tests/unittest_brain_pathlib.py
+++ b/tests/unittest_brain_pathlib.py
@@ -2,6 +2,7 @@
 # For details: https://github.com/PyCQA/astroid/blob/main/LICENSE
 # Copyright (c) https://github.com/PyCQA/astroid/blob/main/CONTRIBUTORS.txt
 
+import pytest
 
 import astroid
 from astroid import bases
@@ -26,6 +27,7 @@ def test_inference_parents() -> None:
     assert inferred[0].qname() == "pathlib._PathParents"
 
 
+@pytest.mark.skipif(not PY310_PLUS, reason="Brain only acts on Py310+")
 def test_inference_parents_subscript_index() -> None:
     """Test inference of ``pathlib.Path.parents``, accessed by index."""
     parents, path = astroid.extract_node(

[Pierre-Sassoulas]

Reading your earlier comment I had the impression that this would speed up old interpreters but that python 3.10 would still be 25% slower ?

[DanielNoord]

The issue is probably that we don't have any brains that use Subscripts with inference.