This means all cross-module references and all mangled names will
consistently use the Swift 4 name (the canonical type), no special
handling required.
The main thing we lose here is that the Swift 4 names of imported
types become usable in Swift 3 mode without any diagnostics, similar
to how most language features introduced in Swift 4 are available in
Swift 3 mode. It also implies that the Swift 4 name will show up in
demangled names.
rdar://problem/31616162
When a type is renamed, we leave behind a "compatibility typealias"
whose underlying type uses the new name. For generic types, though, we
were using the generic parameters and environment of the original
type, which is completely bogus. "Fix" this by just dropping the
generic part entirely and making a typealias that refers to the
/unbound/ generic type, as if written as `typealias OldName = NewName`
instead of `typealias OldName<Element> = NewName<Element>`. The rest
of the compiler can handle that fine.
That is, the stubs we generate when you rename a C global function
imported as a type member using the SwiftName API note. (See the
test case changes.) Previously we hit an assertion.
For good measure, also fix versioned stubs for types-as-members,
which were always added to their original context rather than the
new context.
rdar://problem/31435658
Generic Objective-C classes with this annotation will be imported as
non-generic in Swift. The Swift 3 behavior hardcoded a certain set of
class /hierarchies/ as permanently non-generic, and this is preserved
in Swift 3 mode.
Actually using this API note in a versioned way (as opposed to just
marking the class non-generic in all language versions) will cause
horrible source compatibility problems in the mix-and-match cases,
where Swift 3 code presents a non-generic type that Swift 4 expects to
be generic or vice versa. Fixes for this will come later; right now
it's more important to add support for the feature at all.
To avoid unwanted changes in Swift 4, this commit also adds API notes
to make any existing classes in the previously-hardcoded set continue
to import as non-generic even in Swift 4. The difference is that
/subclasses/ of these classes may come in as generic. (If we want to
make a change here, that can be a separate commit.)
rdar://problem/31226414 (Swift side of rdar://problem/28455962)
If a top-level declaration is imported as a member in Swift 4 but not
in Swift 3, it would still show up in the lookup table for the
containing type in Swift 3 mode. We would import it, and then try to
add the top-level declaration to the containing type.
I'm about to redo this anyway so that the versioned stub will show up
(the one for the Swift 4 name) but this is the narrow fix that avoids
the assertion failure we were seeing.
rdar://problem/31161489
Unlike values, we can't import multiple copies of types under
different names and get good results. Instead, we make a typealias
that points back to the original type. Make sure this typealias is
flagged with whatever version is appropriate, rather than always using
"Swift 2".
When a C declaration is marked unavailable with a replacement, we look
for the replacement to see how it would be imported into Swift. Make
sure we do that with respect to the active language version.
This doesn't actually have any effect yet, but if we start importing
both Swift 3 and Swift 4 versions of protocol requirements and the
non-active one is unavailable, we might mistakenly mark the protocol
un-implementable even when the requirements that are needed are all
there. (Hopefully we would never make a protocol /less/ available in a
newer release, of course.) The test case is designed to catch that.
...and Swift 4 versions in Swift 3, and Swift 2 and "raw" versions in
both. This allows the compiler to produce sensible errors and fix-its
when someone uses the "wrong" name for an API. The diagnostics
certainly have room to improve, but at least the essentials are there.
Note that this commit only addresses /top-level/ decls, i.e. those
found by lookup into a module. We're still limited to producing all
members of a nominal type up front, so that'll require a slightly
different approach.
Part of rdar://problem/29170671
When importing a property as accessor methods (rather than as a
property), we were still importing the type of the accessor methods as
if they were Swift getters and setters of a property, which
(necessarily) homogenizes the types. The homogenization is unnecessary
when importing as accessor methods, because the methods are
independent, so just import the accessor method types as if the
property did not exist. This is particularly useful for maintaining
Swift 3 source compatibility for cases where Swift 4 turns a
getter/setter pair into a null_resettable property.
Fixes rdar://problem/30075571.
Swift side of this new flag. This allows Objective-C framework authors
to replace a pair of methods by properties without breaking source
compatibility. This is especially important for class properties,
which were only introduced last year.
Still to come: importing the accessors even when this flag isn't set,
in order to provide better QoI when migrating from a method interface
to a property interface.
Part of rdar://problem/28455962
Clang now accepts an option -fapinotes-swift-version=XX that describes
which Swift version to use when applying API notes to the Clang
AST. Pass this option down to Clang based on the Swift major version
number (e.g., 3 or 4), allowing API notes to introduce
Swift-version-specific behavior.
Fixes rdar://problem/28617631.
Enable newly-introduced Clang functionality to load API notes stored
alongside the headers for a module or, for a framework, in an APINotes
subdirectory within the framework. Fixes rdar://problem/28455644.
This can be used as QoI for things like 'NSASCIIStringEncoding', which
is going to canonically be `String.Encoding.ascii` if/when SE-0086 is
accepted.
We can't actually handle this in NS_SWIFT_NAME (that is, we can't
/import/ something onto a nested type), but that's okay: we already
have stricter limitations on NS_SWIFT_NAME enforced by Clang.
rdar://problem/26352374
Most cases fall out from swift_name validation on the Clang side dropping invalid API notes, though the validation on the Clang side is conservative and misses some cases. We still have Swift-side work to fall back to the original name here.
When attempting to compile Swift 2 code (or any Swift code using the
Swift 2 names) in Swift 3, the compiler diagnostics are often entirely
useless because the names have changed radically enough that one
generally gets "no member named 'foo'" errors rather than a helpful
"'foo' was renamed to 'bar'" error. This makes for a very poor user
experience when (e.g.) trying to move Swift 2 code forward to Swift 3.
To improve the experience, when the Swift 2 and Swift 3 names of an
API differ, the Clang importer will produce a "stub" declaration that
matches the Swift 2 API. That stub will be marked with a synthesized
attribute
@available(unavailable, renamed: "the-swift-3-name")
that enables better diagnostics (e.g., "'foo' is unavailable: renamed
to 'bar') along with Fix-Its (courtesy of @jrose-apple's recent work)
that fix the Swift 2 code to compile in Swift 3.
This change addresses much of rdar://problem/25309323 (concerning QoI
of Swift 2 code compiled with a Swift 3 compiler), but some cleanup
remains.
Also fixes a leak I introduced with the String -> NSString bridging;
we're always dealing with guaranteed +0, so borrow rather than forward
the "self" argument.
Using Clang to compile API notes introduces a dependency on the
underlying Clang, which can cause problems as Clang's support for API
notes evolves. Re-introduce the "-apinotes" driver flag into the Swift
driver to allow us to compile API notes, and use that (rather than
Clang) to precompile API notes for the overlays.
Most of this is in updating the standard library, SDK overlays, and
piles of test cases to use the new names. No surprises here, although
this shows us some potential heuristic tweaks.
There is one substantive compiler change that needs to be factored out
involving synthesizing calls to copyWithZone()/copy(zone:). Aside from
that, there are four failing tests:
Swift :: ClangModules/objc_parse.swift
Swift :: Interpreter/SDK/Foundation_test.swift
Swift :: Interpreter/SDK/archiving_generic_swift_class.swift
Swift :: Interpreter/SDK/objc_currying.swift
due to two independent remaining compiler bugs:
* We're not getting partial ordering between NSCoder's
encode(AnyObject, forKey: String) and NSKeyedArchiver's version of
that method, and
* Dynamic lookup (into AnyObject) doesn't know how to find the new
names. We need the Swift name lookup tables enabled to address this.
Doing so is safe even though we have mock SDK. The include paths for
modules with the same name in the real and mock SDKs are different, and
the module files will be distinct (because they will have a different
hash).
This reduces test runtime on OS X by 30% and brings it under a minute on
a 16-core machine.
This also uncovered some problems with some tests -- even when run for
iOS configurations, some tests would still run with macosx triple. I
fixed the tests where I noticed this issue.
rdar://problem/19125022
Swift SVN r23683
