[swift-evolution] Pitch: Cross-module inlining and specialization

[Slava Pestov]

Hi all,

Here is a draft proposal that makes public a feature we’ve had for a while. Let me know what you think!

Cross-module inlining and specialization ("@inlinable")
Proposal: SE-NNNN <file:///Users/slava/NNNN-filename.md>
Authors: Slava Pestov <https://github.com/slavapestov>, Jordan Rose <https://github.com/jrose-apple>
Review Manager: TBD
Status: Initial pitch
Implementation: Already implemented as an underscored attribute @_inlineable
Introduction
We propose introducing an @inlinable attribute which exports the body of a function as part of a module's interface, making it available to the optimizer when referenced from other modules.

Motivation
One of the top priorities of the Swift 5 release is a design and implementation of the Swift ABI. This effort consists of three major tasks:

Finalizing the low-level function calling convention, layout of data types, and various runtime data structures. The goal here is to maintain compatibility across compiler versions, ensuring that we can continue to make improvements to the Swift compiler without breaking binaries built with an older version of the compiler.

Implementing support for library evolution, or the ability to make certain source-compatible changes, without breaking binary compatibility. Examples of source-compatible changes we are considering include adding new stored properties to structs and classes, removing private stored properties from structs and classes, adding new public methods to a class, or adding new protocol requirements that have a default implementation. The goal here is to maintain compatibility across framework versions, ensuring that framework authors can evolve their API without breaking binaries built against an older version of the framework. For more information about the resilience model, see the library evolution document <https://github.com/apple/swift/blob/master/docs/LibraryEvolution.rst> in the Swift repository.

Stabilizing the API of the standard library. The goal here is to ensure that the standard library can be deployed separately from client binaries and frameworks, without forcing recompilation of existing code.

All existing language features of Swift were designed with these goals in mind. In particular, the implementation of generic types and functions relies on runtime reified types to allow separate compilation and type checking of generic code.

Within the scope of a single module, the Swift compiler performs very aggressive optimization, including full and partial specialization of generic functions, inlining, and various forms of interprocedural analysis.

On the other hand, across module boundaries, runtime generics introduce unavoidable overhead, as reified type metadata must be passed between functions, and various indirect access patterns must be used to manipulate values of generic type. We believe that for most applications, this overhead is negligible compared to the actual work performed by the code itself.

However, for some advanced use cases, and in particular for the standard library, the overhead of runtime generics can dominate any useful work performed by the library. Examples include the various algorithms defined in protocol extensions of Sequence and Collection, for instance the mapmethod of the Sequence protocol. Here the algorithm is very simple and spends most of its time manipulating generic values and calling to a user-supplied closure; specialization and inlining can completely eliminate the algorithm of the higher-order function call and generate equivalent code to a hand-written loop manipulating concrete types.

We would like to annotate such functions with the @inlinable attribute. This will make their bodies available to the optimizer when building client code; on the other hand, calling such a function will cause it to be emitted into the client binary, meaning that if a library were to change the definition of such a function, only binaries built against the newer version of library will use the new definition.

Proposed solution
The @inlinable attribute causes the body of a function to be emitted as part of the module interface. For example, a framework can define a rather impractical implementation of an algorithm which returns true if all elements of a sequence are equal or if the sequence is empty, and falseotherwise:

@inlinable public func allEqual<T>(_ seq: T) -> Bool
    where T : Sequence, T.Element : Equatable {
  var iter = seq.makeIterator()
  guard let first = iter.next() else { return true }

  func rec(_ iter: inout T.Iterator) -> Bool {
    guard let next = iter.next() else { return true }
    return next == first && rec(&iter)
  }

  return rec(&iter)
}
A client binary built against this framework can call allEqual() and enjoy a possible performance improvement when built with optimizations enabled, due to the elimination of abstraction overhead.

On the other hand, once the framework author comes to their senses and implements an iterative solution to replace the recursive algorithm defined above, the client binary cannot make use of the more efficient implementation until recompiled.

Detailed design
The new @inlinable attribute can only be applied to the following kinds of declarations:

Functions and methods
Subscripts
Computed properties
Initializers
Deinitializers
The attribute can only be applied to public declarations. This is because the attribute only has an effect when the declaration is used from outside of the module. Within a module, the optimizer can always rely on the function body being available.

For similar reasons, the attribute cannot be applied to local declarations, that is, declarations nested inside functions or statements. However, local functions and closure expressions defined inside public @inlinable functions are always implicitly @inlinable.

When applied to subscripts or computed properties, the attribute applies to the getter, setter, didSetand willSet, if present.

The compiler will enforce certain restrictions on bodies of inlinable declarations:

inlinable declarations cannot define local types. This is because all types have a unique identity in the Swift runtime, visible to the language in the form of the == operator on metatype values. It is not clear what it would mean if two different libraries inline the same local type from a third library, with all three libraries linked together into the same binary. This becomes even worse if two different versions of the same inlinable function appear inside the same binary.

inlinable declarations can only reference other public declarations. This is because they can be emitted into the client binary, and are therefore limited to referencing symbols that the client binary can reference.

Note: The restrictions enforced on the bodies of @inlinable declarations are exactly those that we have in place on default argument expressions of public functions in Swift 4.

Source compatibility
The introduction of the @inlinable attribute is an additive change to the language and has no impact on source compatibility.

Effect on ABI stability
The introduction of the @inlinable attribute does not change the ABI of existing declarations. However, adding @inlinable to an existing declaration changes ABI, because the declaration will no longer have a public entry point in the generated library. Removing @inlinable from an existing declaration does not change ABI, because it merely introduces a new public symbol in the generated library.

We have discussed adding a "versioned @inlinable" variant that preserves the public entry point for older clients, while making the declaration inlinable for newer clients. This will likely be a separate proposal and discussion.

Effect on API resilience
Because a declaration marked @inlinable is not part of the library ABI, removing such a declaration is a binary-compatible, but source-incompatible change.

Any changes to the body of a declaration marked @inlinable should be considered very carefully. As a general guideline, we feel that @inlinable makes the most sense with "obviously correct" algorithms which manipulate other data types abstractly through protocols, so that any future changes to an @inlinable declaration are optimizations that do not change observed behavior.

Comparison with other languages
The closest language feature to the @inlinable attribute is found in C and C++. In C and C++, the concept of a header file is similar to Swift's binary swiftmodule files, except they are written by hand and not generated by the compiler. Swift's public declarations are roughly analogous to declarations whose prototypes appear in a header file.

Header files mostly contain declarations without bodies, but can also declare static inlinefunctions with bodies. Such functions are not part of the binary interface of the library, and are instead emitted into client code when referenced. As with @inlinable declarations, static inlinefunctions can only reference other "public" declarations, that is, those that are defined in other header files.

Alternatives considered
One possible alterative would be to add a new compiler mode where all declarations become implicitly @inlinable.

However, such a compilation mode would not solve the problem of delivering a stable ABI and standard library which can be deployed separately from user code. We don't want all declaration bodies in the standard library to be available to the optimizer when building user code.

While such a feature might be useful for users who build private frameworks that are always shipped together their application without resilience concerns, we do not feel it aligns with our goals for ABI stability, and at best it should be a separate discussion.

For similar reasons, we do not feel that an "opt-out" attribute that can be applied to declarations to mark them non-inlinable makes sense.

We have also considered generalizing @inlinable to allow it to be applied to entire blocks of declarations, for example at the level of an extension. As we gain more experience with using this attribute in the standard library we might decide this would be a useful addition, but we feel that for now, it makes sense to focus on the case of a single inlinable declaration instead. Any future generalizations can be introduced as additive language features.

We originally used the spelling @inlineable for the attribute. However, we settled on @inlinable for consistency with the Decodable and Encodable protocols, which are named as they are and not Decodeable and Encodeable.

Finally, we have considered some alternate spellings for this attribute. The name @inlinable is somewhat of a misnomer, because nothing about it actually forces the compiler to inline the declaration; it might simply generate a concrete specialization of it, or look at the body as part of an interprocedural analysis, or completely ignore the body. We have considered @alwaysEmitIntoClient as a more accurate, but awkward, spelling of the attribute's behavior.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20171002/f2175ecb/attachment.html>