This fixes a regression where projects that use the C++ stdlib overlay stop building because of a linker error:
```
ld: Shared cache eligible dylib cannot link to ineligible dylib '@rpath/libswiftCompatibilitySpan.dylib'.
```
Usages of `Span<T>` would generally cause a type metadata accessor to be emitted for `Swift.Span`. This becomes a problem with backdeployment, since Span is partially defined in a compatibility binary.
This change adds `@_alwaysEmitIntoClient` to generic usages of `Span` to prevent the type metadata accessor from being emitted into `libswiftCxx.a`.
rdar://152192080
This teaches ClangImporter to respect the `_Nonnull`/`_Nullable` arguments on templated function parameters.
Previously Swift would only import a non-annotated function overload. Using an overload that has either `_Nonnull` or `_Nullable` would result in a compiler error. The non-annotated overload would get imported with incorrect nullability: Swift would always assume non-null pointers, which was inconsistent with non-templated function parameters, which are mapped to implicitly unwrapped optionals.
With this change all three possible overloads are imported, and all of them get the correct nullability in Swift.
rdar://151939344
Update availability for CxxSpan<->Span, fix lifetimebound on parameters
with reference type
Because swift-ide-test doesn't care about typechecking,
std-span-interface.swift passed despite containing 2 separate errors.
This updates the test file to properly exercise the entire compilation
pipeline for the macro expansions, by running swift-frontend
-emit-module and calling each macro expansion.
The first issue was that CxxSpan initializers taking [Mutable]Span still
had their availability set to Swift 6.2+, even after back-deploying
caused [Mutable]Span to have availability back to Swift 5.0. Since
_SwiftifyImport expansions copy the availbility of Span, this resulted
in the macro expansions calling unavailable initializers. Interestingly
enough, this manifested itself in the form of a tripped assert in SIL
verification, because although we do now typecheck the expansions from
_SwiftifyImport, the compilation can still keep going after
`shouldEmitFunctionBody` returns false: the macro expansion declaration
is still there, but is now missing its definition, despite not being
external.
The second issue was when parameters with C++ reference types were
annotated with `[[clang::lifetimebound]]`. For parameters with a type
that is `Escapable`, this is normally done using `@lifetime(borrow
foo)`. However C++ reference parameters are imported as `inout`, which
requires the `@lifetime(&foo)` syntax.
rdar://151493400
rdar://151678415
* [Swiftify] Emit Mutable[Raw]Span when possible
Previously wrappers would use UnsafeMutable[Raw]Pointer for mutable
pointers, and Span for non-const std::span, to prevent the compiler from
complaining that MutableSpan didn't exist.
Now that MutableSpan has landed we can finally emit MutableSpan without
causing compilation errors. While we had (disabled) support for MutableSpan
syntax already, some unexpected semantic errors required additional
changes:
- Mutable[Raw]Span parameters need to be inout (for mutation)
- inout ~Escapable paramters need explicit lifetime annotations
- MutableSpan cannot be directly bitcast to std::span, because it is
~Copyable, so they need unwrapping to UnsafeMutableBufferPointer
rdar://147883022
* [Swiftify] Wrap if-expressions in Immediately Called Closures
When parameters in swiftified wrapper functions are nullable, we use
separate branches for the nil and nonnil cases, because
`withUnsafeBufferPointer` (and similar) cannot be called on nil.
If-expressions have some limitations on where they are allowed in the
grammar, and cannot be passed as arguments to a function. As such, when
the return value is also swiftified, we get an error when trying to
pass the if-expression to the UnsafeBufferPointer/Span constructor.
While it isn't pretty, the best way forward seems to be by wrapping the
if-expressions in Immediately Called Closures.
The closures have the side-effect of acting as a barrier for 'unsafe':
unsafe keywords outside the closure do not "reach" unsafe expressions
inside the closure. We therefore have to emit "unsafe" where unsafe
expressions are used, rather than just when returning.
rdar://148153063
Do not rely on the @_unsafeNonescapableResult attribute. That attribute is only
for temporarily working around bugs! And it only affects lifetime diagnostics within
the function. It has no affect on the caller's diagnostics, so it won't solve
this problem:
func macroGeneratedThunk() -> CxxSpan<Int> {
return _unsafeRemoveLifetime(Span...)
}
We cannot simply add @_unsafeRemoveLifetime to the thunk, because SwiftSyntax
does not natively support the attribute. We don't want to add SwiftSyntax
support because this attribute will never be supported syntax!
Instead, use `_overrideLifetime` copying the `Void` type to remove a dependency:
func macroGeneratedThunk() -> CxxSpan<Int> {
return _cxxOverrideLifetime(Span..., copying: ())
}
Unfortunately, this was not discovered earlier as swift-ide-test is not
invoking the SIL passes that produce this diagnostic. When creating
Swift spans from C++ spans we have no lifetime dependency information to
propagate as C++ spans are modeled as escapable types. Hence, this PR
introduces a helper function to bypass the lifetime checks triggered by
this discepancy. Hopefully, the new utility will go away as the lifetime
analysis matures on the Swift side and we get standardized way to deal
with unsafe lifetimes.
CxxSpan is trivial, but not immortal.
This initializer is diagnosed with an error after enabling trivial dependence
enforcement. Correct this with an _overrideLifetime call. This could be avoided
if we had a another way to tell the compiler that CxxSpan.__dataUnsafe()
produced a pointer with the same effective lifetime as the CxxSpan.
A first step towards creating safe overloads for C++ APIs using span
(rdar://139074571).
Note that we need to mark span as owned because it the libc++
implementation was mistakenly recognized as owned and might now rely on
span methods like `data` being renamed as `__dataUnsafe`. We will change
it under a new interop version. But for the time being, we want
consistent behavior across stdlib versions.
