WLED_MM_Infinity/.github/cpp.instructions.md

---
applyTo: "**/*.cpp,**/*.h,**/*.hpp, **/*.ino"
---
# C++ Coding Conventions

See also: [CONTRIBUTING.md](../CONTRIBUTING.md) for general style guidelines that apply to all contributors.

## Formatting

- Indent with **2 spaces** (no tabs in C++ files)
- Opening braces on the same line is preferred (K&R style). Brace on a separate line (Allman style) is acceptable
- Single-statement `if` bodies may omit braces: `if (a == b) doStuff(a);`
- Space between keyword and parenthesis: `if (...)`, `for (...)`. No space between function name and parenthesis: `doStuff(a)`
- No enforced line-length limit; wrap when a line exceeds your editor width

## Naming

- **camelCase** for functions and variables: `setValuesFromMainSeg()`, `effectCurrent`
- **PascalCase** for classes and structs: `PinManagerClass`, `BusConfig`
- **UPPER_CASE** for macros and constants: `WLED_MAX_USERMODS`, `DEFAULT_CLIENT_SSID`

## Header Guards

Most headers use `#ifndef` / `#define` guards. Some newer headers add `#pragma once` before the guard:

```cpp
#ifndef WLED_EXAMPLE_H
#define WLED_EXAMPLE_H
// ...
#endif // WLED_EXAMPLE_H
```

## Comments

- `//` for inline comments, `/* ... */` for block comments. Always put a space after `//`
- Mark WLED-MM-specific changes with `// WLEDMM` or `// WLEDMM: description`:

```cpp
// WLEDMM: increased max bus count for larger installs
#ifndef WLED_MAX_BUSSES
  #define WLED_MAX_BUSSES 20  // WLEDMM default (upstream: 10)
#endif
```

- **AI attribution:** When a larger block of code is generated by an AI tool, mark it with an `// AI:` comment so reviewers know to scrutinize it:

```cpp
// AI: below section was generated by an AI
void calculateCRC(const uint8_t* data, size_t len) {
  ...
}
// AI: end of AI-generated section
```

  Single-line AI-assisted edits do not need the marker — use it when the AI produced a contiguous block that a human did not write line-by-line.

- **Function & feature comments:** Every non-trivial function should have a brief comment above it describing what it does. Include a note about each parameter when the names alone are not self-explanatory:

```cpp
/* *****
 * Apply gamma correction to a single color channel.
 * @param value  raw 8-bit channel value (0–255)
 * @param gamma  gamma exponent (typically 2.8)
 * @return       corrected 8-bit value
 ***** */
uint8_t gammaCorrect(uint8_t value, float gamma);
```

  Short accessor-style functions (getters/setters, one-liners) may skip this if their purpose is obvious from the name.

## Preprocessor & Feature Flags

- Prefer compile-time feature flags (`#ifdef` / `#ifndef`) over runtime checks where possible
- Platform differentiation: `ARDUINO_ARCH_ESP32` vs `ESP8266`
- WLED-MM fork detection: `_MoonModules_WLED_` (defined in `wled.h`)
- PSRAM availability: `BOARD_HAS_PSRAM`
- WLEDMM_FASTPATH is the default path; code under `#ifndef WLEDMM_FASTPATH` is deprecated and will be phased out.
- Flash-saving mode: `WLEDMM_SAVE_FLASH` (disables aggressive inlining)

## Error Handling

- `DEBUG_PRINTF()` / `DEBUG_PRINTLN()` for developer diagnostics (compiled out unless `-D WLED_DEBUG`)
- `USER_PRINTF()` / `USER_PRINTLN()` for user-visible messages (always compiled in)
- Don't rely on C++ exceptions — use return codes (`-1` / `false` for errors) and global flags (e.g. `errorFlag = ERR_LOW_MEM`). Some builds don't support C++ exceptions.

## Strings

- (Optional) Use `F("string")` for string constants (stores in PROGMEM, saves RAM)
- Use `const char*` for temporary/parsed strings
- Avoid `String` (Arduino heap-allocated string) in hot paths; acceptable in config/setup code

## Memory

- **PSRAM-aware allocation**: use `d_malloc()` (prefer DRAM), `p_malloc()` (prefer PSRAM) from `util.h`
- **Avoid Variable Length Arrays (VLAs)**: GCC/Clang support VLAs as an extension (they are not part of the C++ standard), so they look like a legitimate feature — but they are allocated on the stack at runtime. On ESP32/ESP8266, FreeRTOS task stacks are typically only 2–8 KB; a VLA whose size depends on a runtime parameter (segment dimensions, pixel counts, etc.) can silently exhaust the stack and cause the program to behave in unexpected ways or crash. Prefer a fixed-size array with a compile-time bound, or heap allocation (`d_malloc` / `p_malloc`) for dynamically sized buffers. **Any VLA must be explicitly justified in the source code or PR.**
- **Larger buffers** (LED data, JSON documents) should use PSRAM when available and technically feasible
- **Hot-path**: some data should stay in DRAM or IRAM for performance reasons
- Memory efficiency matters, but is less critical on boards with PSRAM

## `const` and `constexpr`

`const` is a promise to the compiler that a value (or object) will not change - a function declared with a `const char* message` parameter is not allowed to modify the content of `message`.
This pattern enables optimizations and makes intent clear to reviewers.

### `const` locals

Adding `const` to a local variable that is only assigned once is not necessary — but it **is** required when the variable is passed to a function that takes a `const` parameter (pointer or reference). In hot-path code, `const` on cached locals helps the compiler keep values in registers:

```cpp
const uint_fast16_t cols = virtualWidth();
const uint_fast16_t rows = virtualHeight();
```

### `const` references to avoid copies

Pass and store objects by `const &` (or `&`) instead of copying them implicitly. This avoids constructing temporary objects on every access — especially important in loops:

```cpp
const auto &m = _mappings[i];          // reference, not a copy (bus_manager.cpp)
const CRGB& c = ledBuffer[pix];        // alias — avoids creating a temporary CRGB instance
```

For function parameters that are read-only, prefer `const &`:

```cpp
BusDigital(BusConfig &bc, uint8_t nr, const ColorOrderMap &com);
```

### `constexpr` over `#define`

Prefer `constexpr` for compile-time constants. Unlike `#define`, `constexpr` respects scope and type safety, keeping the global namespace clean:

```cpp
// Prefer:
constexpr uint32_t TWO_CHANNEL_MASK = 0x00FF00FF;
constexpr int WLED_MAX_BUSSES = WLED_MAX_DIGITAL_CHANNELS + WLED_MAX_ANALOG_CHANNELS;

// Avoid (when possible):
#define TWO_CHANNEL_MASK 0x00FF00FF
```

Note: `#define` is still needed for conditional compilation guards (`#ifdef`), platform macros, and values that must be overridable from build flags.

### `static_assert` over `#error`

Use `static_assert` instead of the C-style `#if … #error … #endif` pattern when validating compile-time constants. It provides a clear message and works with `constexpr` values:

```cpp
// Prefer:
constexpr int WLED_MAX_BUSSES = WLED_MAX_DIGITAL_CHANNELS + WLED_MAX_ANALOG_CHANNELS;
static_assert(WLED_MAX_BUSSES <= 32, "WLED_MAX_BUSSES exceeds hard limit");

// Avoid:
#if (WLED_MAX_BUSSES > 32)
  #error "WLED_MAX_BUSSES exceeds hard limit"
#endif
```

### `static` and `const` class methods

#### `const` member functions

Marking a member function `const` tells the compiler that it does not modify the object's state:

```cpp
uint16_t length() const { return _len; }
bool     isActive() const { return _active; }
```

Benefits for GCC/Xtensa/RISC-V:
- The compiler knows the method cannot write to `this`, so it is free to **keep member values in registers** across the call and avoid reload barriers.
- `const` methods can be called on `const` objects and `const` references — essential when passing large objects as `const &` to avoid copying.
- `const` allows the compiler to **eliminate redundant loads**: if a caller already has a member value cached, the compiler can prove the `const` call cannot invalidate it.

Declare every getter, query, or inspection method `const`. If you need to mark a member `mutable` to work around this (e.g. for a cache or counter), document the reason.

#### `static` member functions

A `static` member function has no implicit `this` pointer. This has two distinct advantages:

1. **Smaller code, faster calls**: no `this` is passed in a register. On Xtensa and RISC-V, this removes one register argument from every call site and prevents the compiler from emitting `this`-preservation code around inlined blocks.
2. **Better inlining**: GCC can inline a `static` method with more certainty because it cannot be overridden by a derived class (no virtual dispatch ambiguity) and has no aliasing concern through `this`.

Use `static` for any method that does not need access to instance members:

```cpp
// Factory / utility — no instance needed:
static BusConfig fromJson(JsonObject obj);

// Pure computation helpers:
static uint8_t  gamma8(uint8_t val);
static uint32_t colorBalance(uint32_t color, uint8_t r, uint8_t g, uint8_t b);
```

`static` also communicates intent clearly: a reviewer immediately knows the method is stateless and safe to call without a fully constructed object.

> **Rule of thumb**: if a method does not read or write any member variable, make it `static`. If it only reads member variables, make it `const`. Note: `static` methods cannot also be `const`-qualified because there is no implicit `this` pointer to be const — just use `static`. Both qualifiers reduce coupling and improve generated code on all ESP32 targets.

---

## Hot-Path Optimization

The hot path is the per-frame pixel pipeline: **Segment → Strip → BusManager → Bus(Digital,HUB75,Network) or PolyBus → LED driver**. Speed is the top priority here. The patterns below are taken from existing hot-path code (`FX_fcn.cpp`, `FX_2Dfcn.cpp`, `bus_manager.cpp`, `colorTools.hpp`) and should be followed when modifying these files.

Note: `FX.cpp` (effect functions) is written by many contributors and has diverse styles — that is acceptable. The guidelines below apply starting from pixel set/get operations and below.

### Function Attributes

Stack the appropriate attributes on hot-path functions. Defined in `const.h`:

| Attribute | Meaning | When to use |
|---|---|---|
| `__attribute__((hot))` | Branch-prediction hint | hot-path functions with complex logic |
| `IRAM_ATTR` | Place in fast IRAM (ESP32) | Critical per-pixel functions (e.g. `BusDigital::setPixelColor`) |
| `IRAM_ATTR_YN` | IRAM on ESP32, no-op on ESP8266 | Hot functions that ESP8266 can't fit in IRAM |
| `WLED_O2_ATTR` | Force `-O2` optimization | Most hot-path functions |
| `WLED_O3_ATTR` | Force `-O3,fast-math` | Innermost color math (e.g. `color_blend`) |
| `[[gnu::hot]] inline` | Modern C++ attribute + inline | Header-defined accessors (e.g. `progress()`, `currentBri()`) |

Note: `WLED_O3_ATTR` sometimes causes performance loss compared to `WLED_O2_ATTR`. Choose optimization levels based on test results.

Example signature:

```cpp
void IRAM_ATTR_YN WLED_O2_ATTR __attribute__((hot)) Segment::setPixelColor(int i, uint32_t col)
```

### Use `uint_fast` Types for Locals

Use `uint_fast8_t` and `uint_fast16_t` for loop counters, indices, and temporary calculations in hot paths. These let the compiler pick the CPU's native word size (32-bit on ESP32), avoiding unnecessary narrow-type masking:

```cpp
uint_fast8_t count = numBusses;
for (uint_fast8_t i = 0; i < count; i++) { ... }
```

Keep `uint8_t` / `uint16_t` for struct fields and stored data where memory layout matters.

### Cache Members to Locals Before Loops

Copy class members and virtual-call results to local variables before entering a loop:

```cpp
uint_fast8_t count = numBusses;          // avoid repeated member access
for (uint_fast8_t i = 0; i < count; i++) {
  Bus* const b = busses[i];              // const pointer hints to compiler
  uint_fast16_t bstart = b->getStart();
  uint_fast16_t blen = b->getLength();
  ...
}
```

### Unsigned Range Check

Replace two-comparison range tests with a single unsigned subtraction:

```cpp
// Instead of: if (pix >= bstart && pix < bstart + blen)
if ((uint_fast16_t)(pix - bstart) < blen)  // also catches negative pix via unsigned underflow
```

### Early Returns

Guard every hot-path function with the cheapest necessary checks first:

```cpp
if (!isActive()) return;                    // inactive segment
if (unsigned(i) >= virtualLength()) return; // bounds check (catches negative i too)
```

### Avoid Nested Calls — Fast Path / Complex Path

Avoid calling non-inline functions or making complex decisions inside per-pixel hot-path code. When a function has both a common simple case and a rare complex case, split it into two variants and choose once per frame rather than per pixel:

```cpp
// Decision made once per frame in startFrame(), stored in a bool
bool simpleSegment = _isSuperSimpleSegment;

// Per-pixel loop — no complex branching inside
if (simpleSegment)
  setPixelColorXY_fast(x, y, col, scaled_col, cols, rows);  // inline, no bounds checks
else
  setPixelColorXY_slow(x, y, col);  // full validation, grouping, mirroring
```

The same principle applies to color utilities — `color_add()` accepts a `fast` flag so callers can choose saturating adds (no branches) vs. ratio-preserving adds (with division) without an inner-loop decision:

```cpp
uint32_t color_add(uint32_t c1, uint32_t c2, bool fast=false);
```

General rules:
- Keep the per-pixel fast path free of non-inline function calls, multi-way branches and complex switch-case decisions.
- Hoist the "which path?" decision out of the inner loop (once per frame or per segment)
- It is acceptable to duplicate some code between fast and complex variants to keep the fast path lean

### Function Pointers to Eliminate Repeated Decisions

When the same decision (e.g. "which drawing routine?") would be evaluated for every pixel, assign the chosen variant to a function pointer once and let the inner loop call through the pointer. This removes the branch entirely — the calling code (e.g. the GIF decoder loop) only ever invokes one function per frame, with no per-pixel decision.

`image_loader.cpp` demonstrates the pattern: `calculateScaling()` picks the best drawing callback once per frame based on segment dimensions and GIF size, then passes it to the decoder via `setDrawPixelCallback()`:

```cpp
// calculateScaling() — called once per frame
if ((perPixelX < 2) && (perPixelY < 2))
  decoder.setDrawPixelCallback(drawPixelCallbackDownScale2D);   // downscale-only variant
else
  decoder.setDrawPixelCallback(drawPixelCallback2D);            // full-scaling variant
```

Each callback is a small, single-purpose function with no internal branching — the decoder's per-pixel loop never re-evaluates which strategy to use.

### Template Specialization (Advanced)

Templates can eliminate runtime decisions by generating separate code paths at compile time. For example, a pixel setter could be templated on color order or channel count so the compiler removes dead branches and produces tight, specialized machine code:

```cpp
template<bool hasWhite>
void setChannel(uint8_t* out, uint32_t col) {
  out[0] = R(col); out[1] = G(col); out[2] = B(col);
  if constexpr (hasWhite) out[3] = W(col);  // compiled out when hasWhite==false
}
```

Use sparingly — each instantiation duplicates code in flash. On ESP8266 and small-flash ESP32 boards this can exhaust IRAM/flash. Prefer templates only when the hot path is measurably faster and the number of instantiations is small (2–4).

### RAII Lock-Free Synchronization (Advanced)

Where contention is rare and the critical section is short, consider replacing mutex-based locking with lock-free techniques using `std::atomic` and RAII scoped guards. A scoped guard sets a flag on construction and clears it on destruction, guaranteeing cleanup even on early return:

```cpp
struct ScopedBusyFlag {
  std::atomic<bool>& flag;
  bool acquired;
  ScopedBusyFlag(std::atomic<bool>& f) : flag(f), acquired(false) {
    bool expected = false;
    acquired = flag.compare_exchange_strong(expected, true);
  }
  ~ScopedBusyFlag() { if (acquired) flag.store(false); }
  explicit operator bool() const { return acquired; }
};

// Usage
static std::atomic<bool> busySending{false};
ScopedBusyFlag guard(busySending);
if (!guard) return;  // another task is already sending
// ... do work — flag auto-clears when guard goes out of scope
```

This avoids FreeRTOS semaphore overhead and the risk of forgetting `esp32SemGive`. There are no current examples of this pattern in the codebase — consult with maintainers before introducing it in new code, to ensure it aligns with the project's synchronization conventions.

### Pre-Compute Outside Loops

Move invariant calculations before the loop. Pre-compute reciprocals to replace division with multiplication:

```cpp
const uint_fast16_t cols = virtualWidth();
const uint_fast16_t rows = virtualHeight();
uint_fast8_t fadeRate = (255 - rate) >> 1;
float mappedRate_r = 1.0f / (float(fadeRate) + 1.1f);  // reciprocal — avoid division inside loop
```

### Parallel Channel Processing

Process R+B and W+G channels simultaneously using the two-channel mask pattern:

```cpp
constexpr uint32_t TWO_CHANNEL_MASK = 0x00FF00FF;
uint32_t rb = (((c1 & TWO_CHANNEL_MASK) * amount) >> 8) & TWO_CHANNEL_MASK;
uint32_t wg = (((c1 >> 8) & TWO_CHANNEL_MASK) * amount) & ~TWO_CHANNEL_MASK;
return rb | wg;
```

### Bit Shifts Over Division (mainly for RISC-V boards)

ESP32 and ESP32-S3 (Xtensa core) have a fast "integer divide" instruction, so manual shifts rarely help. 
The compiler already converts power-of-two unsigned divisions to shifts at `-O2`.
On RISC-V-based boards (ESP32-C3, ESP32-C6, ESP32-C5) explicit shifts can be beneficial:

Prefer bit shifts for power-of-two operations:

```cpp
position >> 3     // instead of position / 8
(255U - rate) >> 1 // instead of (255 - rate) / 2
i & 0x0007        // instead of i % 8
```

**Important**: The bit-shifted expression should be unsigned. On some MCUs, "signed right-shift" is implemented by an "arithmetic shift right" that duplicates the sign bit: ``0b1010 >> 1 = 0b1101``.

### Static Caching for Expensive Computations

Cache results in static locals when the input rarely changes between calls:

```cpp
static uint16_t lastKelvin = 0;
static byte correctionRGB[4] = {255,255,255,0};
if (lastKelvin != kelvin) {
  colorKtoRGB(kelvin, correctionRGB);  // expensive — only recalculate when input changes
  lastKelvin = kelvin;
}
```

### Inlining Strategy

- Move frequently-called small functions to headers for inlining (e.g. `WS2812FX::setPixelColor` is in `FX.h`)
- Use `static inline` for file-local helpers
- On ESP32 with `WLEDMM_FASTPATH`, color utilities are inlined from `colorTools.hpp`; on ESP8266 or `WLEDMM_SAVE_FLASH`, they fall back to `colors.cpp`

### Colors

- Store and pass colors as `uint32_t` (0xWWRRGGBB)
- Extract channels with macros: `R(c)`, `G(c)`, `B(c)`, `W(c)`, compose with `RGBW32(r,g,b,w)`
- Use `CRGB` (FastLED type) mainly when interfacing with FastLED functions; convert at boundaries
- Use 16-bit intermediates for channel math to ensure 32-bit (not 64-bit) arithmetic:
  ```cpp
  uint16_t r1 = R(color1);  // 16-bit intermediate keeps the multiply result in 32 bits, avoiding 64-bit promotion
  ```

## Multi-Task Synchronization

ESP32 runs multiple FreeRTOS tasks concurrently (e.g. network handling, LED output, JSON parsing). Use the WLED-MM mutex macros for synchronization — they expand to FreeRTOS recursive semaphore calls on ESP32 and compile to no-ops on ESP8266:

| Macro | Signature | Description |
|---|---|---|
| `esp32SemTake(mux, timeout)` | `mux`: `SemaphoreHandle_t`, `timeout`: milliseconds | Acquire a recursive mutex. Returns `pdTRUE` on success, `pdFALSE` on timeout. |
| `esp32SemGive(mux)` | `mux`: `SemaphoreHandle_t` | Release a previously acquired recursive mutex. |

Pre-defined mutex handles (declared in `wled.h`):

| Mutex | Protects |
|---|---|
| `busDrawMux` | Concurrent `strip.show()` and `strip.service()` — acquire before writing pixels from background tasks (DDP, E1.31, Art-Net) |
| `segmentMux` | Segment array modifications — acquire before adding, removing, or iterating segments |
| `jsonBufferLockMutex` | Shared JSON document buffer |
| `presetFileMux` | `presets.json` file reads and writes |

Usage pattern:

```cpp
if (esp32SemTake(busDrawMux, 200) == pdTRUE) { // wait max 200 ms
  // ... critical section ...
  esp32SemGive(busDrawMux);
} else {
  // fallback code or error reporting
}
```

Always pair every `esp32SemTake` with a matching `esp32SemGive`. Choose a timeout appropriate for the operation — typically 200 ms for drawing, up to 2500 ms for file I/O.

**Important**: Not every shared resource needs a mutex. Some synchronization is guaranteed by the overall control flow. For example, `volatile bool` flags like `suspendStripService`, `doInitBusses`, `loadLedmap`, and `OTAisRunning` (declared in `wled.h`) are checked sequentially in the main loop (`wled.cpp`), so they serialize access without requiring a semaphore. Use mutexes when true concurrent access from multiple FreeRTOS tasks is possible and race-conditions can lead to unexpected behaviour. Rely on control-flow ordering when operations are sequenced within the same loop iteration.

### `delay()` vs `yield()` in FreeRTOS Tasks

On ESP32, `delay(ms)` calls `vTaskDelay(ms / portTICK_PERIOD_MS)`, which **suspends only the calling task**. The FreeRTOS scheduler immediately runs all other ready tasks. This differs from ESP8266, where `delay()` stalled the entire system unless `yield()` was called inside.

**`delay()` in `loopTask` is allowed.** The Arduino `loop()` function runs inside `loopTask`. Calling `delay()` there does not block the network stack, audio FFT, LED DMA, or any other FreeRTOS task.

**`yield()` is a no-op in WLED-MM on ESP32.** `WLEDMM_FASTPATH` redefines `yield()` to an empty macro:

```cpp
#define yield() {}  // WLEDMM: yield() is completely unnecessary on ESP32
```

Even in stock arduino-esp32, `yield()` calls `vTaskDelay(0)`, which only switches to tasks at equal or higher priority — the IDLE task (priority 0) is never reached. 
**Do not use `yield()` to pace ESP32 tasks or assume it feeds any watchdog**.

**Custom `xTaskCreate()` tasks must call `delay(1)` in their loop, not `yield()`.** Without a real blocking call, the IDLE task is starved. The IDLE watchdog panic is the first visible symptom — but the damage starts earlier: deleted task memory leaks, software timers stop firing, light sleep is disabled, and Wi-Fi/BT idle hooks don't run. See `esp-idf.instructions.md` for a full explanation of what IDLE does. Structure custom tasks like this:

```cpp
// WRONG — IDLE task is never scheduled; yield() does not feed the idle task watchdog.
void myTask(void*) {
  for (;;) {
    doWork();
    yield();
  }
}

// CORRECT — delay(1) suspends the task for ≥1 ms, IDLE task runs, IDLE watchdog is fed
void myTask(void*) {
  for (;;) {
    doWork();
    delay(1);  // DO NOT REMOVE — lets IDLE(0) run and feeds its watchdog
  }
}
```

Prefer blocking FreeRTOS primitives (`xQueueReceive`, `ulTaskNotifyTake`, `vTaskDelayUntil`) over `delay(1)` polling where precise timing or event-driven behaviour is needed.

**Watchdog note.** WLED-MM disables the Task Watchdog by default (`WLED_WATCHDOG_TIMEOUT 0` in `wled.h`). When enabled, `esp_task_wdt_reset()` is called at the end of each `loop()` iteration. Long blocking operations inside `loop()` — such as OTA downloads or slow file I/O — must call `esp_task_wdt_reset()` periodically, or be restructured so the main loop is not blocked for longer than the configured timeout.

## General

- Follow the existing style in the file you are editing
- If possible, use `static` for local (C-style) variables and functions (keeps the global namespace clean)
- Avoid unexplained "magic numbers". Prefer named constants (`constexpr`) or C-style `#define` constants for repeated numbers that have the same meaning
- Include `"wled.h"` as the primary project header where needed
- **Float-to-unsigned conversion is undefined behavior when the value is out of range.** Converting a negative `float` directly to an unsigned integer type (`uint8_t`, `uint16_t`, …) is UB per the C++ standard — the Xtensa (ESP32) toolchain may silently wrap, but RISC-V (ESP32-C3/C6) can produce different results due to clamping. Cast through a signed integer first:
  ```cpp
  // Undefined behavior — avoid:
  uint8_t angle = 40.74f * atan2f(dy, dx);   // negative float → uint8_t is UB
  
  // Correct — cast through int first:
  // atan2f returns [-π..+π], scaled ≈ [-128..+128] as int; uint8_t wraps negative ints via 2's complement (e.g. -1 → 255)
  uint8_t angle = int(40.74f * atan2f(dy, dx));  // float→int (defined), int→uint8_t (defined)
  ```