There was an import cycle in the Python modules:
- `qmk.build_targets` imported `qmk.cli.generate.compilation_database`;
- importing `qmk.cli.generate.compilation_database` requires
initializing `qmk.cli` first;
- the initialization of `qmk.cli` imported the modules for all CLI
commands;
- `qmk.cli.compile` imported `qmk.build_targets`.
This cycle did not matter in most cases, because `qmk.cli` was imported
first, and in that case importing `qmk.cli.generate.compilation_database`
did not trigger the initialization of `qmk.cli` again. However, there was
one corner case when `qmk.bulld_targets` was getting imported first:
- The `qmk find` command uses the `multiprocessing` module.
- The `multiprocessing` module uses the `spawn` start method on macOS
and Windows.
- When the `spawn` method is used, the child processes initialize
without any Python modules loaded, and the required modules are loaded
on demand by the `pickle` module when receiving the serialized objects
from the main process.
The result was that the `qmk find` command did not work properly on macOS
(and probably Windows too); it reported exceptions like this:
ImportError: cannot import name 'KeyboardKeymapBuildTarget' from partially initialized module 'qmk.build_targets' (most likely due to a circular import)
Moving the offending `qmk.cli.generate.compilation_database` import into
the method which actually uses it fixes the problem.
When multiple `-f FILTER` options were specified, `qmk find` did not
return anything at all instead of printing the list of entries that
matched all of the specified filters.
The problem was that the statement in `_filter_keymap_targets()` that
filled `targets` had a wrong indent and therefore was executed for every
filter instead of only once after applying all filters, and
`valid_keymaps` was actually an iterator and therefore could be used
only once. Moving the statement outside of the loop fixes the problem.
Functions in filters did not work properly except when used in the last
(or only) filter. The problem was caused by the peculiarity of the
`lambda` behavior in Python — any variables from the outer scope are
captured only by reference, therefore any subsequent reassignment of
those variables is propagated to all lambdas created earlier in the same
scope. Together with the laziness of `filter()` (it returns an iterator
which performs filtering on demand) this resulted in all function
filters using the values of the `key` and `value` variables which
correspond to the last filter in the sequence, therefore the result of
filtering was wrong if some filter with a function was not the last one
in the sequence.
Apparently the shortest way to make a Python lambda capture some
variables by value is to add arguments with default values for such
variables (default values are evaluated when the lambda is created, and
any subsequent reassignments in the outer scope no longer changes them).
This makes filters with functions work properly even when such filters
are not at the last position in the sequence.
Nick Brassel
Sergey Vlasov
Joel Challis
QMK Bot
Andrew Kannan
Ryan
Less/Rikki
Pablo Martínez