Before, mbed_app.json files were explicitly ignored when building tests.
This was mostly because you could have multiple mbed_app.json files in the
tree (for instance, in test case folders) and the behavior would be
undefined. Now the tools explicitly ensure that there aren't multiple
mbed_app.json files in your source files. So auto discovery of
mbed_app.json for testing is being reintroduced.
This addresses the issue where building tests via test.py always triggered
a clean build. This is because the mbed_config.h file was being deleted from
the shared OS build to ensure that the correct config was always being
used. However, this contanstly triggered a rebuild of the OS since the
config file was not present.
Due to the shared build, having multiple app configurations that could
override the OS settings is not possible. For this reason, we now ignore
app config files unless explicitly set via the command line option
'--app-config'. Though there will now be two mbed_config.h files in the
include path of the build, it shouldn't matter since the contents will be
the same.
Custom targets were the origin of a number of issues with the mbed
tools, so it was decided that they need to be removed. This PR does just
that (and moves the "custom_targets" part of the config system tests
into a separate, per-test "targets.json" file to preserve the test
functionality).
The configuration system uses `mbed_config.h` to export configuration
data, but when this change was made, the documentation was not properly
updated. This commit updates the documentation with the correct
behaviour of the configuration system.
(long commit message ahead. Sorry about that, it can't be helped.)
This commit changs targets definition from Python to JSON format, as
part of the configuration mechanism implementation. There is a new file
under workspace_tools/ called "targets.json" which contains the target
definitions. "targets.py" remains, but becomes a wrapper on top of
"targets.json", with the same interface as before. This has the
advantage of not requiring code changes outside "targets.py".
Most of the JSON definitions of targets were automatically generated by a
script (available upon request since it doesn't make a lot of sense to
include it here), only those targets that had more than one parent in
the Python implementation were converted by hand. The target definitions
should be pretty self-explanatory. A number of things are different in
the JSON implementation (this is just a summary, check docs/mbed_targets.md
(also part of this PR) for a more complete description):
- "program_cycle_s" is now a value (as opposed to a function in the
Python implementation), since it only returned a number in all the
Python target implementations. The main definition that actually contains
some code (in class "Target") remains in target.py
- array values in "macros" and "extra_labels" can be modified
dynamically. Values can be added using "macros_add" and
"extra_labels_add" or removed using "macros_remove" and
"extra_labels_remove". This mechanism is available for all attributes
with a list type, but it's currently enabled only for "macros" and
"extra_labels" to keep things simple.
- "init_hooks"/"binary_hook" are now implemented in terms of a single
JSON key valled "post_binary_hook". The corresponding code is also in
"targets.py", under the various TargetCode classes (see for example
LPC4088Code in targets.py).
Just like in the Python implementation, a target can inherit from zero,
one or more targets. The resolution order for the target's attributes
follows the one used by the Python code (I used
http://makina-corpus.com/blog/metier/2014/python-tutorial-understanding-python-mro-class-search-path
as a reference for the implementation of resolution order).
This is obviously a very dangerous commit, since it affects all targets.
I tested compilation for a number of targets (K64F, LPC1768, NRF51822)
but there's definitely a lot more to be done in terms of testing.
I also tried to test in a different way: I wrote a script that imports the
old (Python) and the new (JSON) implementations and verifies that the
attributes in the old implementations exist and have the same values
in the new implementations (it also verifies that the attribute
resolution order is the same in the two implementations). If you're
interested, the script is here:
https://gist.github.com/bogdanm/c9d8cf34214109a4b9079befed6b3c0c
And the results of running the script are below (note that the script
outputs only the target names that were found to be problematic):
NRF51_MICROBIT_BOOT:
Resolution order is different in old and new
old: ['NRF51_MICROBIT_BOOT', 'MCU_NRF51_16K_BOOT_S110', 'MCU_NRF51_16K_BOOT_BASE', 'MCU_NRF51_16K_BASE', 'MCU_NRF51', 'Target', 'MCU_NRF51_S110']
new: ['NRF51_MICROBIT_BOOT', 'MCU_NRF51_16K_BOOT_S110', 'MCU_NRF51_S110', 'MCU_NRF51_16K_BOOT_BASE', 'MCU_NRF51_16K_BASE', 'MCU_NRF51', 'Target']
'extra_labels' has different values in old and new
old: ['NORDIC', 'MCU_NRF51', 'MCU_NRF51822', 'MCU_NORDIC_16K', 'MCU_NRF51_16K', 'MCU_NRF51_16K_BOOT', 'MCU_NRF51_16K_S110', 'NRF51_MICROBIT']
new: ['NORDIC', 'MCU_NRF51', 'MCU_NRF51822', 'MCU_NORDIC_16K', 'MCU_NRF51_16K', 'MCU_NRF51_16K_S110', 'MCU_NRF51_16K_BOOT', 'NRF51_MICROBIT']
'macros' has different values in old and new
old: ['NRF51', 'TARGET_NRF51822', 'TARGET_MCU_NORDIC_16K', 'TARGET_MCU_NRF51_16K', 'TARGET_MCU_NRF51_16K_BOOT', 'TARGET_OTA_ENABLED', 'TARGET_MCU_NRF51_16K_S110', 'TARGET_NRF51_MICROBIT', 'TARGET_NRF_LFCLK_RC']
new: ['NRF51', 'TARGET_NRF51822', 'TARGET_MCU_NORDIC_16K', 'TARGET_MCU_NRF51_16K', 'TARGET_MCU_NRF51_16K_S110', 'TARGET_MCU_NRF51_16K_BOOT', 'TARGET_OTA_ENABLED', 'TARGET_NRF51_MICROBIT', 'TARGET_NRF_LFCLK_RC']
NRF51_MICROBIT:
Resolution order is different in old and new
old: ['NRF51_MICROBIT', 'MCU_NRF51_16K_S110', 'MCU_NRF51_16K_BASE', 'MCU_NRF51', 'Target', 'MCU_NRF51_S110']
new: ['NRF51_MICROBIT', 'MCU_NRF51_16K_S110', 'MCU_NRF51_S110', 'MCU_NRF51_16K_BASE', 'MCU_NRF51', 'Target']
'extra_labels' has different values in old and new
old: ['NORDIC', 'MCU_NRF51', 'MCU_NRF51822', 'MCU_NORDIC_16K', 'MCU_NRF51_16K', 'MCU_NRF51_16K_S110']
new: ['NORDIC', 'MCU_NRF51', 'MCU_NRF51822', 'MCU_NRF51_16K_S110', 'MCU_NORDIC_16K', 'MCU_NRF51_16K']
'macros' has different values in old and new
old: ['NRF51', 'TARGET_NRF51822', 'TARGET_MCU_NORDIC_16K', 'TARGET_MCU_NRF51_16K', 'TARGET_MCU_NRF51_16K_S110', 'TARGET_NRF_LFCLK_RC']
new: ['NRF51', 'TARGET_NRF51822', 'TARGET_MCU_NRF51_16K_S110', 'TARGET_MCU_NORDIC_16K', 'TARGET_MCU_NRF51_16K', 'TARGET_NRF_LFCLK_RC']
NRF51_MICROBIT_OTA:
Resolution order is different in old and new
old: ['NRF51_MICROBIT_OTA', 'MCU_NRF51_16K_OTA_S110', 'MCU_NRF51_16K_OTA_BASE', 'MCU_NRF51_16K_BASE', 'MCU_NRF51', 'Target', 'MCU_NRF51_S110']
new: ['NRF51_MICROBIT_OTA', 'MCU_NRF51_16K_OTA_S110', 'MCU_NRF51_S110', 'MCU_NRF51_16K_OTA_BASE', 'MCU_NRF51_16K_BASE', 'MCU_NRF51', 'Target']
'extra_labels' has different values in old and new
old: ['NORDIC', 'MCU_NRF51', 'MCU_NRF51822', 'MCU_NORDIC_16K', 'MCU_NRF51_16K', 'MCU_NRF51_16K_OTA', 'MCU_NRF51_16K_S110', 'NRF51_MICROBIT']
new: ['NORDIC', 'MCU_NRF51', 'MCU_NRF51822', 'MCU_NORDIC_16K', 'MCU_NRF51_16K', 'MCU_NRF51_16K_S110', 'MCU_NRF51_16K_OTA', 'NRF51_MICROBIT']
'macros' has different values in old and new
old: ['NRF51', 'TARGET_NRF51822', 'TARGET_MCU_NORDIC_16K', 'TARGET_MCU_NRF51_16K', 'TARGET_MCU_NRF51_16K_OTA', 'TARGET_OTA_ENABLED', 'TARGET_MCU_NRF51_16K_S110', 'TARGET_NRF51_MICROBIT', 'TARGET_NRF_LFCLK_RC']
new: ['NRF51', 'TARGET_NRF51822', 'TARGET_MCU_NORDIC_16K', 'TARGET_MCU_NRF51_16K', 'TARGET_MCU_NRF51_16K_S110', 'TARGET_MCU_NRF51_16K_OTA', 'TARGET_OTA_ENABLED', 'TARGET_NRF51_MICROBIT', 'TARGET_NRF_LFCLK_RC']
NOT OK: ['NRF51_MICROBIT', 'NRF51_MICROBIT_BOOT', 'NRF51_MICROBIT_OTA']
The reasons for the above output are subtle and related to the
extremely weird way in which we defined target data in the Python
implementation: we used both class attributes and instance attributes.
This can complicate resolution order quite a bit and those two levels
don't exist in JSON: there's only one attribute type (equivalent to
Python's instance attributes). To make that work, I had to change the
inheritance order of the above targets (that use multiple inheritance)
which in turn changed the order of some macros and extra_labels (and of
course the resolution order). No harm done: the values are the same,
only their ordering is different. I don't believe this causes any
problems for 'extra_labels' and 'macros'.
This method of testing has its limitations though; in particular, it
can't test the hooks. I'm opened to ideas about how to test this better,
but I think that we need to remember that this commit might break some
targets and keep an eye out for "weird errors" in the future.
We want to use (apparently) requirements.txt to store mbed's Python package dependencies.
Users can use this file with pip to install those requirements.
sarahmarshy
Sarah Marsh
Sam Grove
Russ Butler
Brian Daniels
andreas.larsson
Bogdan Marinescu
Jimmy Brisson
kclin
Bogdan Marinescu
Christopher Haster
Seppo Takalo
Mihail Stoyanov
Marcelo Salazar
Mihail Stoyanov
Irit Arkin
Nabil Bendafi
Przemek Wirkus
Yoong Hor Meng