decode() is never called polymorphically - all call sites in
read_message_() use concrete types. The only indirect call site was
decode_to_message(), which also always knows the concrete type.
Convert decode_to_message() to a template so the concrete type is
preserved, allowing decode() to be non-virtual. The two classes that
override decode() (ExecuteServiceArgument, ExecuteServiceRequest) now
hide the base method, which works since all calls use concrete types.
This removes one vtable slot (4 bytes) from each decodable message
class vtable, saving ~148 bytes of flash.
Since init_log_buffer_ is now called from the constructor (before
pre_setup sets global_logger), calling disable_loop() would trigger
ESP_LOGVV which dereferences the null global_logger pointer.
The loop self-disables on its first iteration when no messages are
found, so the explicit disable in init_log_buffer_ was unnecessary.
Move TaskLogBuffer allocation from init_log_buffer() (called at
DIAGNOSTICS priority) into the Logger constructor (called at
EARLY_INIT priority). This ensures the buffer exists before
global_logger is set, eliminating a window where another FreeRTOS
task could dereference a null log_buffer_ pointer.
Fixes crash: Guru Meditation Error: Core 0 panic'ed (Load access fault)
in TaskLogBuffer::send_message_thread_safe when a task logs before
init_log_buffer() is called.
Move trivial null-check getter from component.cpp to component.h
so the compiler can inline it at call sites, eliminating function
call overhead in hot logging paths.
Extend the precomputed-tag approach from fixed32 key fields to all forced
fields with single-byte tags (field IDs 1-15). The code generator now
emits write_raw_byte(tag) followed by the raw encode primitive instead
of calling the full encode_* method.
For varint types (uint32, uint64, sint32, sint64, int64, bool, enum),
this eliminates the zero-check branch and encode_field_raw indirection.
For length-delimited types (bytes, string), it additionally skips the
encode_string wrapper.
Benchmarked on real hardware with BluetoothLERawAdvertisementsResponse
(12 advertisements per message, 10000 iterations):
ESP32 (Xtensa dual-core 240MHz):
encode: 38498 -> 30460 ns/op (-20.9%)
calc+encode: 48479 -> 40458 ns/op (-16.6%)
ESP32-C3 (RISC-V single-core 160MHz):
encode: 54199 -> 40342 ns/op (-25.6%)
calc+encode: 57800 -> 51365 ns/op (-11.1%)
The single-message case (via write_protobuf_packet) is the most common
path. Peeling the first loop iteration and outlining the multi-message
batch path avoids the ~300-byte StaticVector<iovec> stack allocation
on the hot path.
Plaintext write_protobuf_messages:
- Stack frame: 352 → 64 bytes
- Code size: 246 → 127 bytes
Noise write_protobuf_messages:
- Extracted encrypt_noise_message_ helper for reuse
- Same peeling pattern with outlined batch path
Use real TCP sockets instead of AF_UNIX socketpair so TCP_NODELAY
succeeds during init() and the benchmark exercises the full write
path. Replace hardcoded message type 38 with SensorStateResponse::MESSAGE_TYPE.
Avoid benchmarking heap allocation by pre-reserving the buffer
to typical TCP MSS size and reusing it across iterations, matching
real-world usage where the buffer persists across writes.
J. Nick Koston