[tmori]

Makes the Hakoniwa sensor device of the RISC-V Erbium board a Devicetree node instead of a memory.txt DEV (EXDEV) device.

What changed

• New model hakoniwa,sensor-registers (board/common/src/hako_sensor_registers_model.c), ported from the EXDEV glue with its behaviour unchanged:
  • it latches each sample at the sample's own time, with a data-ready pulse;
  • it writes the committed motor outputs to hil_actuator;
  • the test vehicle and the trace format are the same.
• The same plugin library registers it next to hakoniwa,sensor-device. Both the Hakoniwa link build and the in-process aircraft build (HAKO_SENSOR_DEVICE_BUILD_INPROC) therefore carry it.
• boards/erbium/erbium-hakoniwa.dts places the node at 0x20000000 with PLIC source 5. erbium.dtb is unchanged.
• The test generator maps each Hakoniwa plugin compatible that a board uses.
• New unit test hako-sensor-registers (test vehicle). A deliberately broken interrupt pulse makes it fail.

Checks

Test vehicle on PX4 aifoundry_erbium_hakoniwa, 3e9 clocks, against the C path with the EXDEV device:

• the trace is identical: 3,899 sensor, GPS and actuator lines, with exact float bits;
• the firmware console is identical (39,216 bytes); only the start-up log lines differ.

Hakoniwa x500 take-off to 2.5 m, headless, 3 runs each:

Configuration	RTF
Before: C path + EXDEV, aircraft service	1.68 / 2.07 / 1.78
After: Devicetree plugin, aircraft service	1.93 / 1.64 / 1.89
After: Devicetree plugin, in-process aircraft	2.08 / 2.09 / 2.06

All runs PASS, holding 2.48-2.51 m. With the viewer at real time, 3 runs of the Devicetree plugin also passed.

ctest passes on macOS and on Ubuntu 24.04 (566/566).

Found on the way (not changed here): the core's EXDEV loader keeps a pointer into the MPU region table, and the Devicetree models reallocate that table. A DEV combined with a Devicetree can therefore read freed memory. athrill-target-riscv now refuses that combination.

🤖 Generated with Claude Code