Permalink
332 lines (252 sloc) 13.3 KB

Make unsafe pointer nullability explicit using Optional

Introduction

In Objective-C, pointers (whether to objects or to a non-object type) can be marked as nullable or nonnull, depending on whether the pointer value can ever be null. In Swift, however, there is no such way to make this distinction for pointers to non-object types: an UnsafePointer<Int> might be null, or it might never be.

We already have a way to describe this: Optionals. This proposal makes UnsafePointer<Int> represent a non-nullable pointer, and UnsafePointer<Int>? a nullable pointer. This also allows us to preserve information about pointer nullability available in header files for imported C and Objective-C APIs.

swift-evolution thread: https://lists.swift.org/pipermail/swift-evolution/Week-of-Mon-20160314/012918.html

Motivation

Today, UnsafePointer and friends suffer from a problem inherited from C: every pointer value could potentially be null, and code that works with pointers may or may not expect this. Failing to take the null pointer case into account can lead to assertion failures or crashes. For example, pretty much every operation on UnsafePointer itself requires a valid pointer (reading, writing, and initializing the pointee or performing arithmetic operations).

Fortunately, when a type has a single invalid value for which no operations are valid, Swift already has a solution: Optionals. Applying this to pointer types makes things very clear: if the type is non-optional, the pointer will never be null, and if it is optional, the developer must take the "null pointer" case into account. This clarity has already been appreciated in Apple's Objective-C headers, which include nullability annotations for all pointer types (not just object pointers).

This change also allows developers working with pointers to take advantage of the many syntactic conveniences already built around optionals. For example, the standard library currently has a helper method on UnsafeMutablePointer called _setIfNonNil; with "optional pointers" this can be written simply and clearly:

ptr?.pointee = newValue

Finally, this change also reduces the number of types that conform to NilLiteralConvertible, a source of confusion for newcomers who (reasonably) associate nil directly with optionals. Currently the standard library includes the following NilLiteralConvertible types:

  • Optional
  • ImplicitlyUnwrappedOptional (subject of a separate proposal by Chris Willmore)
  • _OptionalNilComparisonType (used for optionalValue == nil)
  • UnsafePointer
  • UnsafeMutablePointer
  • AutoreleasingUnsafeMutablePointer
  • OpaquePointer

plus these Objective-C-specific types:

  • Selector
  • NSZone (only used to pass nil in Swift)

All of the italicized types would drop their conformance to NilLiteralConvertible; the "null pointer" would be represented by a nil optional of a particular type.

Proposed solution

  1. Have the compiler assume that all values with pointer type (the italicized types listed above) are non-null. This allows the representation of Optional.none for a pointer type to be a null pointer value.

  2. Drop NilLiteralConvertible conformance for all pointer types.

  3. Teach the Clang importer to treat _Nullable pointers as Optional (and _Null_unspecified pointers as ImplicitlyUnwrappedOptional).

  4. Deal with the fallout, i.e. adjust the compiler and the standard library to handle this new behavior.

  5. Test migration and improve the migrator as necessary.

This proposal does not include the removal of the NilLiteralConvertible protocol altogether; besides still having two distinct optional types, we've seen people wanting to use nil for their own types (e.g. JSON values). (Changing this in the future is not out of the question; it's just out of scope for this proposal.)

Detailed design

API Changes

  • Conformance to NilLiteralConvertible is removed from all types except Optional, ImplicitlyUnwrappedOptional, and _OptionalNilComparisonType, along with the implementation of init(nilLiteral:).

  • init(bitPattern: Int) and init(bitPattern: UInt) on all pointer types become failable; if the bit pattern represents a null pointer, nil is returned.

  • Should SE-0016 be accepted, the init(bitPattern:) initializers on Int and UInt will be changed to take optional pointers.

  • New initializers will be added to all pointer types to convert between optional pointer types (see below).

  • UnsafeBufferPointer's baseAddress property becomes nullable, along with its initializer parameter (see below).

  • Process.unsafeArgv is a pointer to a null-terminated C array of C strings, so its type changes from UnsafeMutablePointer<UnsafeMutablePointer<Int8>> to UnsafeMutablePointer<UnsafeMutablePointer<Int8>?>, i.e. the inner pointer type becomes optional.

  • NSErrorPointer becomes optional:

-public typealias NSErrorPointer = AutoreleasingUnsafeMutablePointer<NSError?>
+public typealias NSErrorPointer = AutoreleasingUnsafeMutablePointer<NSError?>?
  • A number of methods on String that came from NSString now have optional parameters:
   public func completePathIntoString(
-    outputName: UnsafeMutablePointer<String> = nil,
+    outputName: UnsafeMutablePointer<String>? = nil,
     caseSensitive: Bool,
-    matchesIntoArray: UnsafeMutablePointer<[String]> = nil,
+    matchesIntoArray: UnsafeMutablePointer<[String]>? = nil,
     filterTypes: [String]? = nil
   ) -> Int {
   public init(
     contentsOfFile path: String,
-    usedEncoding: UnsafeMutablePointer<NSStringEncoding> = nil
+    usedEncoding: UnsafeMutablePointer<NSStringEncoding>? = nil
   ) throws {

   public init(
     contentsOfURL url: NSURL,
-    usedEncoding enc: UnsafeMutablePointer<NSStringEncoding> = nil
+    usedEncoding enc: UnsafeMutablePointer<NSStringEncoding>? = nil
   ) throws {
   public func linguisticTags(
     in range: Range<Index>,
     scheme tagScheme: String,
     options opts: NSLinguisticTaggerOptions = [],
     orthography: NSOrthography? = nil,
-    tokenRanges: UnsafeMutablePointer<[Range<Index>]> = nil
+    tokenRanges: UnsafeMutablePointer<[Range<Index>]>? = nil
   ) -> [String] {
  • NSZone's no-argument initializer is gone. (It probably should have been removed already as part of the Swift 3 naming cleanup.)

  • A small regression: optional pointers can no longer be passed using withVaList because it would require a conditional conformance to the CVarArg protocol. For now, using unsafeBitCast to reinterpret the optional pointer as an Int is the best alternative; Int has the same C variadic calling conventions as a pointer on all supported platforms.

Conversion between pointers

Currently each pointer type has initializers of this form:

init<OtherPointee>(_ otherPointer: UnsafePointer<OtherPointee>)

This simply makes a pointer with a different type but the same address as otherPointer. However, in making pointer nullability explicit, this now only converts non-nil pointers to non-nil pointers. In my experiments, this has led to this idiom becoming very common:

// Before:
let untypedPointer = UnsafePointer<Void>(ptr)

// After:
let untypedPointer = ptr.map(UnsafePointer<Void>.init)

// Usually the pointee type is actually inferred:
foo(ptr.map(UnsafePointer.init))

I consider this a bit more difficult to understand than the original code, at least at a glance. We should therefore add new initializers of the following form:

init?<OtherPointee>(_ otherPointer: UnsafePointer<OtherPointee>?) {
  guard let nonnullPointer = otherPointer else {
    return nil
  }
  self.init(nonnullPointer)
}

The body is for explanation purposes only; we'll make sure the actual implementation does not require an extra comparison.

(This would need to be an overload rather than replacing the previous initializer because the "non-null-ness" should be preserved through the type conversion.)

Note: It is very likely the existing initializers described here will be renamed (perhaps to init(bitPattern:)). In this case, the new initializers should adopt the same argument labels.

UnsafeBufferPointer

The type UnsafeBufferPointer represents a bounded typed memory region with no ownership or lifetime semantics; it is made up of a bare typed pointer (its baseAddress) and a length (count) and conforms to Collection. There is also a variant with mutable contents named UnsafeMutableBufferPointer.

For a buffer with 0 elements, there's no need to provide the address of allocated memory, since it can't be read from. This case is represented as a nil base address and a count of 0.

With this proposal, the baseAddress property becomes optional:

   /// Construct an Unsafe${Mutable}Pointer over the `count` contiguous
   /// `Element` instances beginning at `start`.
-  public init(start: Unsafe${Mutable}Pointer<Element>, count: Int) {
+  ///
+  /// If `start` is nil, `count` must be 0. However, `count` may be 0 even for
+  /// a nonzero `start`.
+  public init(start: Unsafe${Mutable}Pointer<Element>?, count: Int) {
-  public var baseAddress: Unsafe${Mutable}Pointer<Element> {
+  public var baseAddress: Unsafe${Mutable}Pointer<Element>? {

This does force clients using baseAddress to consider the possibility that the buffer does not represent allocated memory. However, we believe that most clients are either using the Collection conformance and ignoring the baseAddress property, or are immediately passing the pointer (and perhaps also the count) to a C API, most of which accept null pointers with a 0 count. In either case a null pointer should be treated no differently from any other address with a count of 0. This API also allows converting to and from a pair of (UnsafePointer?, Int) without losing information and without needing to explicitly handle the nil case.

Here is some data on standard library uses of UnsafeBuffer:

  • Used as Collection: 4 (mostly String operations)
  • Passed to C-style APIs (pointer and length): 1
  • Explicitly extracting the base address: 3 (all related to C strings)

A "use" here is roughly "mentioned at least once in a function body".

Impact on existing code

Any code that uses a pointer type (including Selector or NSZone) may be affected by this change. For the most part our existing logic to handle last year's nullability audit should cover this, but the implementer should test migration of several projects to see what issues might arise.

Anecdotally, in migrating the standard library to use this new logic I've been quite happy with nullability being made explicit. There are many places where a pointer really can't be nil.

Alternatives considered

The primary alternative here would be to leave everything as it is today, with UnsafePointer and friends including the null pointer as one of their normal values. This has obviously worked just fine for nearly two years of Swift, but it is leaving information on the table that can help avoid bugs, and is strange in a language that makes fluent use of Optional. As a fairly major source-breaking change, it is also something that we probably should do sooner rather than later in the language's evolution.

Félix Cloutier also noted that this may prove problematic for porting Swift to a platform where there are no invalid pointer values (usually an embedded platform). However, Chris Lattner thinks this potential future issue should not limit the improvements that can be made to Swift today (especially given the lack of several other features necessary for low-level system programming, such as volatile), and Doug Gregor pointed out that Clang and LLVM have the assumption that "0 is an invalid pointer" hardcoded in many places already. This is not an entirely satisfactory answer, but I agree that we should go ahead with the language change regardless.

Alternatives for UnsafeBufferPointer

The chosen API change for UnsafeBufferPointer does impose a cost on clients that want to access and use the base address themselves: they need to consider the nil case explicitly, where previously they wouldn't have had to. We considered several alternatives, including:

  • Using an arbitrary address with proper alignment whenever we would have used nil as a base address.

  • Eliminating nil from the withUnsafeBufferPointer APIs and then making the baseAddress property non-optional; clients that need to deal with nil could use an Optional UnsafeBufferPointer.

The ultimate consensus (both on the list and in off-list discussion with the Swift core team) was that neither of these behave well when using UnsafeBufferPointer to interoperate with C APIs, even if we could make withUnsafeBufferPointer. We do eliminate the possibility of using the type system to distinguish between "buffers that may have a null base address" and "buffers known to have a non-null base address", but we're expecting that distinction to not be a useful one anyway.