Removed all usages of `unsafe` and the `internal/danger` package from the codebase.
1. **`unstable/ast.go`**: Refactored `Node` struct to use `*Node` pointers for `next` and `child` fields instead of integer offsets. This eliminates the need for `unsafe` pointer arithmetic in `Next()` and `Child()` methods.
2. **`unstable/builder.go`**: Updated `builder` to manage pointers to nodes directly instead of integer offsets.
3. **`unstable/parser.go`**:
* Replaced `danger.SubsliceOffset` with safe capacity-based calculation (`cap(p.data) - cap(b)`), which works because tokens are slices of the parser's input buffer.
4. **`strict.go`** & **`errors.go`**: Replaced `danger.BytesRange` and `danger.SubsliceOffset` with safe slice capacity arithmetic.
5. **`unmarshaler.go`**: Replaced `map[danger.TypeID]...` with `map[uintptr]...` for the field paths cache using `reflect.ValueOf(t).Pointer()`. This removes the need for `unsafe` access to `reflect.Type` internals.
6. **`internal/tracker/seen_test.go`**: Replaced `unsafe.Sizeof` with `reflect.TypeOf(...).Size()`.
7. **`internal/danger`**: Deleted the package entirely.
Benchmarks show a mix of performance changes:
- Small document unmarshaling (SimpleDocument/struct-4) got slower (+25%), likely due to pointer chasing vs contiguous array access.
- Large document unmarshaling (canada, citm, twitter) actually improved significantly (-24% to -45% latency), likely due to reduced allocation overhead or better cache locality in some paths.
- Memory usage for large datasets decreased significantly (-50% to -60% B/op).
- Overall geomean latency improved by ~6%.
No public interfaces were changed. All tests pass.
Remove internal/danger package and replace unsafe pointer arithmetic with direct pointer manipulation. Update AST node references to use pointers instead of integer offsets. This improves code safety and maintainability.
Co-authored-by: thomas.pelletier <thomas.pelletier@bedrockrobotics.com>
Cursor Agent