Generic Specialization
Part of the Sun ABI draft. These are proposed requirements, not guarantees of current releases. Open decisions must be resolved before stabilization.
Generic specialization
Suppose a bundle retains a generic operation that selects the first array element:
library: generic first<Element>, referring to a particular helper declaration
app: requests first<Point>
compiler: generates code using Point's layout and the recorded helperAdding a new overload named helper must not silently redirect the old body's
call. Parsing its serialized fields successfully does not establish that its
names, ownership operations, or layout assumptions still mean the same thing.
The design keeps compile-time specialization. It does not require runtime type-erased generics. Retaining generic bodies is separate from optionally embedding complete source: complete source also contains ordinary implementations and inputs needed to rebuild the bundle. Specialization alone does not rebuild its existing compiled code.
| Minimal example | Level | Requirement |
|---|---|---|
An old generic first is specialized for a new Point type. | MUST | Specialize an exported body using its declared semantic revision and exact referenced declarations. |
An old body calls helper; a new overload must not change that binding. | MUST NOT | Silently apply changed name lookup, overload selection, ownership, intrinsic, or layout rules to an old body. |
The bundle already contains first specialized for i32. | MUST | Reuse it only when its identity and contracts match the requested specialization. |
A new element type has stricter alignment than i32. | MUST | Generate a compatible specialization using the new type's declared layout and operations, or reject the combination. |
| The old generic expects assumptions that the new element type cannot satisfy. | MUST NOT | Change a new type's layout to satisfy an old generic's ABI. |
Open decisions
A versioned serialized AST with explicit compatibility checks is one starting point. A representation with resolved declarations and types is a later option. Both need defined semantics; compatible protobuf fields alone are insufficient.