NFC: BridgeJS: Descriptor-driven codegen, decoupling type knowledge from code generation

[krodak]

Context

Following the discussion in #496 - @kateinoigakukun raised a concern about adding more BridgeType cases before paying down complexity debt in the code generator. Specifically, the ad-hoc per-type handling should be moved out of codegen, and extending supported types shouldn't require invasive changes.

In my reply, I suggested we could define pairs of Swift/TS type bridging declaratively and generalize codegen logic - similar in spirit to how stack-based types already work.

This PR is an attempt at that approach. It's experimental - I'm completely open to feedback on the direction. We can rework it, split parts into separate PRs (e.g. the JS glue changes), or close this entirely if this isn't the right path. The goal was to explore what it takes to centralize type-specific ABI knowledge so codegen operates generically rather than switching on every type.

The idea

Today, each codegen function (lowerParameter, liftReturn, optionalLowerReturn, etc.) has its own switch over BridgeType encoding the same structural facts - "this type uses one i32", "this type returns via the stack", "this type's optional uses a side-channel function" - in slightly different ways. Adding a new simple type means touching many of these switches.

This PR introduces four declarative abstractions:

BridgeTypeDescriptor

Defined once per BridgeType, this captures the Wasm ABI shape as a struct:

• wasmParams - core Wasm parameter types for export direction (e.g. string = [i32, i32], bool = [i32], struct = [])
• importParams - parameter types for import direction (defaults to wasmParams; string overrides to [i32] since imports use object IDs)
• wasmReturnType / importReturnType - return type in export vs import direction
• optionalConvention - how Optional<T> is handled (see below), explicitly declared on every descriptor
• nilSentinel - bit pattern that represents nil for types with "extra inhabitants" (see below)
• usesStackLifting - whether multi-param stack lifting needs LIFO variable reordering
• accessorTransform - how to access the bridgeable value from a Swift accessor (identity, .jsObject member access, protocol downcast)
• lowerMethod - which bridge protocol method to call (stackReturn, fullReturn, pushParameter, none)

Codegen functions read descriptor fields instead of switching per type. For example, lowerStatements uses descriptor.lowerMethod and descriptor.accessorTransform rather than matching on .jsObject, .swiftProtocol, .swiftStruct, etc. individually. Return-type switches in renderCallStatement, callStaticProperty, and callPropertyGetter are unified via accessorTransform.applyToReturnBinding().

OptionalConvention

Captures how Optional<T> is lowered/lifted for a given wrapped type T:

public enum OptionalConvention: Sendable, Equatable {
    case stackABI                            // struct, array, dict - everything via stack
    case inlineFlag                          // bool, jsValue, closure, etc. - isSome + T's params
    case sideChannelReturn(OptionalSideChannel) // int, string, jsObject, etc. - side-channel for returns
}

Every descriptor explicitly declares its convention. There's no implicit derivation from wasmParams count - each type states exactly how its optionals behave.

NilSentinel

NilSentinel enum captures whether a type has a value that can represent nil without an extra isSome flag:

public enum NilSentinel: Sendable, Equatable {
    case none          // all bit patterns valid (int, float, etc.)
    case i32(Int32)    // specific i32 sentinel (0 for object IDs, -1 for enum tags)
    case pointer       // null pointer (0) for heap objects
}

Types with sentinels:

• jsObject, swiftProtocol - sentinel 0 (object IDs start at 2)
• swiftHeapObject - sentinel null pointer
• caseEnum, associatedValueEnum - sentinel -1 (never a valid case index)

The sentinel is used in optionalLowerReturn for the JS import direction - types with sentinels use a generic sentinel-based return path instead of per-type switch cases. The inner lowerReturn fragment is composed into an isSome ? <lowered> : <sentinel> pattern automatically.

JSCoercion

JS-side coercion info for simple scalar types - lift/lower transforms, variable hints, and optional scalar kind:

public struct JSCoercion: Sendable {
    let liftCoerce: String?           // e.g. "$0 !== 0" for bool
    let lowerCoerce: String?          // e.g. "$0 ? 1 : 0" for bool
    let stackLowerCoerce: String?     // override for stack-context lowering
    let varHint: String               // variable naming hint ("bool", "int", "f32", etc.)
    let optionalScalarKind: OptionalScalarKind?  // which side-channel storage slot to use
}

OptionalScalarKind is an enum (.bool, .int, .float, .double) that derives the side-channel storage name (tmpRetOptionalBool, etc.) and intrinsic function name (swift_js_return_optional_bool, etc.) from the case. This replaces the previous string-based optionalReturnStorage / optionalReturnFunc fields.

Types that return non-nil jsCoercion are handled through a generic scalarFragments() builder that returns a (lift, lower) pair. Scalar lowerParameter is always .identity since JS auto-coerces. This removes all per-type scalar fragment functions (boolLowerParameter, uintLiftReturn, etc.).

Compositional optional handling

optionalLowerParameter and optionalLiftParameter no longer contain per-type switches. They compose T's existing lowerParameter/liftParameter fragment inside an isSome conditional:

optionalLowerParameter: Runs T's lowerParameter fragment into a buffer printer, captures results into outer variables, and wraps any cleanup in a scoped closure. The buffer approach lets us detect whether the inner fragment actually produces cleanup code - the innerCleanup variable and its associated emission are only generated when the inner fragment has cleanup lines, avoiding dead code in the generated JS.

optionalLiftParameter: Runs T's liftParameter fragment into a buffer. If the buffer is empty (pure expression), uses a ternary isSome ? expr : null. If it has side effects, wraps in an if/else block.

New types added to lowerParameter or liftParameter automatically get correct optional handling for free.

The same compositional approach extends to struct fields - structFieldLowerFragment for nullable wrapped types delegates to the inner type's non-optional lowering fragment, wrapping it in an isSome conditional with placeholder pushes in the else branch. The same conditional cleanup emission applies here.

Import/export param unification via importParams

Types like string, rawValueEnum(.string), and swiftStruct have different parameter shapes depending on direction:

• Export: string uses (bytes, length) pair
• Import: string uses a single object ID (value)

The importParams field on the descriptor captures this difference. loweringParameterInfo (used when Swift calls JS) always reads importParams, while liftParameterInfo (used when JS calls Swift) reads wasmParams. This eliminated the per-type parameter info switches - the default path now reads directly from the descriptor.

liftExpression collapse

In StackCodegen.liftExpression, 15 types all generated the same TypeName.bridgeJSLiftParameter() pattern. These are collapsed to a default case, keeping only the types that need genuinely different codegen: .jsObject(className?) (wrapping constructor), .nullable/.array/.dictionary (delegation), .closure (uses JSObject), and .void/.namespaceEnum (literal ()).

Similarly, liftNullableExpression is collapsed from a 15-type explicit case list to a 2-case switch: named jsObject needing a .map wrapper vs everything else using direct lift.

What adding a new type (e.g. UUID) would look like after this PR

1. Add the BridgeType case - e.g. .uuid in BridgeType enum

2. Define its descriptor - UUID maps to string on the wire, so:

   case .uuid:
       return BridgeTypeDescriptor(
           wasmParams: [("bytes", .i32), ("length", .i32)],
           importParams: [("value", .i32)],
           wasmReturnType: nil,
           importReturnType: .i32,
           optionalConvention: .sideChannelReturn(.storage),
           accessorTransform: .identity,
           lowerMethod: .stackReturn
       )

3. JS glue - jsCoercion stays nil (same as string). UUID uses string's wire format, so the existing bespoke string fragments handle the JS-side encoding/decoding. scalarFragments() only applies to types like int/bool/float that need simple JS coercion expressions.

4. Add Swift runtime conformances - BridgeJSLowerStackReturn / BridgeJSLiftParameter that convert between UUID and its wire representation (UUID(uuidString:)! / .uuidString)

5. Add skeleton recognition - teach the type parser to recognize Foundation.UUID and emit .uuid

Codegen functions (lowerStatements, liftReturn, optionalLower*, optionalLift*, JS glue generation) pick up the new type automatically through the descriptor and coercion info - no need to add cases to each of them individually.

For a type like URL, the same pattern applies (URL also maps to string on the wire via .absoluteString / URL(string:)!). The Foundation-gating (#if canImport(Foundation)) would go in the runtime conformances for Embedded Swift compatibility.

[kateinoigakukun]

I think the schema-driven approach makes sense to me.

On the other hand, I think what we really need to think more about is how to composite the type descriptors. e.g. Optional should have a different descriptor depending on wrapped type T, so we need to define a general rule for that. Or define a fallback convention that works without knowing the T's descriptor by using Stack ABI and define specialized descriptors for known cases.

I still haven't checked the entire changes yet so I might be missing something 🙇

[krodak]

@kateinoigakukun thanks for looking at this, I'll think on your feedback and try to progress in this direction when I can; in parallel I'll look for more simplifications around intrinsics like we are both currently doing 👌🏻

[krodak]

@kateinoigakukun updated PR description and made some more changes and fixes, no rush on this, but PR should be in a good to have a look for further discussion; let me know if some of the changes would satisfy your earlier remarks partially