On 3.10 and earlier with or without PYTHONHOME, an executable in a build directory has sys.path as below:

0
C:\cpython-3.10\PCbuild\amd64\python310.zip
C:\cpython-3.10\DLLs
C:\cpython-3.10\lib
C:\cpython-3.10\PCbuild\amd64
C:\cpython-3.10
C:\cpython-3.10\lib\site-packages

1
C:\cpython-3.10\PCbuild\amd64\python310.zip
C:\cpython-3.10\DLLs
C:\cpython-3.10\lib
C:\cpython-3.10\PCbuild\amd64
C:\cpython-3.10
C:\cpython-3.10\lib\site-packages

2
C:\cpython-3.10\PCbuild\amd64\python310.zip
C:\cpython-3.10\DLLs
C:\cpython-3.10\lib
C:\cpython-3.10\PCbuild\amd64
C:\cpython-3.10
C:\cpython-3.10\lib\site-packages

Okay, so the problem is that home is set on the second call, which means getpath skips checking for build directory landmarks.

IIRC, other tests were broken if we check against real_executable for the build landmarks and ignore PYTHONHOME (presumably because they used symlinks to a new directory for tests?), so this probably just needs new logic.

However, the fix in #92980 might also affect this, by not recalculating sys.path on subsequent calls. You might want to test again before changing too much.

IIRC, other tests were broken if we check against real_executable for the build landmarks and ignore PYTHONHOME (presumably because they used symlinks to a new directory for tests?), so this probably just needs new logic.

test_getpath passes the following test cases:

• test_symlink_buildtree_win32
• test_buildtree_pythonhome_win32

I can also start python in their file layouts, but test_embed were broken. Neither PYTHONHOME nor home worked around. New logic may be needed for this. Do I understand what you mean?

Had a quick look at your changes, but honestly am going to need a lot more time to figure out that the logic is sound. That also concerns me, because I really want to avoid putting any logic for this stuff back in C - moving to Python was to fix that ;-)

Another idea that might work: what if we just add a config field for "running in-tree build"? It can start as "unspecified" and then be updated if we detect landmarks. Then even if home is already set but we know we're in tree, we can go down the path that overrides platlibdir. What do you think?

Thanks for your feedback.

IIUC, we have no way to update the original config? getpath updates the config which was duplicated at pyinit_core(), and the copy disappears after system configuration with the update of global config (_Py_path_config: deprecated?).
For example, module_search_paths_set and _is_python_build fields never change in the original config when repeating Py_InitializeFromConfig().

I think the flag will work if you would be ok with saving it into the global config.

Oh I see, it's the path config state that's keeping the values around. If we copy _is_python_build into there and only ever set it ourselves, that would be fine, right?

My experiment (#92411) passed the repeating tests, keeping _is_python_build as a member of _Py_path_config. I now think it can be saved as a static variable in pathconfig.c.

All of _Py_path_config is basically just a static variable - it never gets referenced outside of this module. But we're trying to minimize static variables, so that one day we can get them to be per-runtime rather than per-process. So please, keep it in the struct.

@neonene Is this fixed now or there is still work to be done ?

I assume we delay 3.11 backporting just in case.