The discovery for pytest is still quite a bit slower than running plain -m pytest --collect-only so I investigated a bit more and the next most significant (at least for our test suite) slowdown is caused by the pytest.Item.location call used to get line number of the test:

This is because the .location returns tuple of path and line number, of which only the line number is used by the discovery, but computing the path is rather expensive. This PR proposes an optimization which uses internal pytest API in exactly the same way as the location function itself, but without also computing the path.

In particular, the location call hierarchy is:

• Item.location
• PyobjMixin.reportinfo
• getfslineno
• Code.from_function
• Code.firstlineno + Code.path - this is the expensive call we need to avoid

This PR calls Code.from_function directly and only gets .firstlineno from it.

Both profiles are captured with this PR cherry-picked to main branch prior to the discovery result compaction introduced in 5389068 which causes significant performance hit that overshadows the optimization done here, for that I created a separate PR #26016 with proposed fix.

This change makes the discovery about 20% faster for our suite, now almost on par with the raw pytest call