The new Swift parser uses a revamped set of syntax nodes for key paths
that better match the language. However, the C++ parser does not
produce anything like these syntax nodes, so switch to an "old"
keypath node style for now.
Remove the preallocated closure discriminator from KeyPathExpr and go back
to expanding them using an AutoClosureExpr inside of a CaptureListExpr now
that that's supported. This allows the discriminator to be assigned during
type checking without disturbing the indexing of explicit closure literals.
Previously, we would turn a key path literal like `\.foo` in function type
context into a double-wrapped closure like this:
```
foo(\.x) // before type checking
foo({ $kp$ in { $0[$kp$] } }(\.x)) // after type checking
```
in order to preserve the evaluation semantics of the key path literal. This
works but leads to some awkward raw SIL generated out of SILGen which misses
out on various SILGen peepholes and requires a fair number of passes to clean
up. The semantics can still be preserved with a single layer of closure, by
using a capture list:
```
foo({[$kp$ = \.x] in $0[$kp$] }) // after type checking
```
which generates better natural code out of SILGen, and is also (IMO) easier
to understand on human inspection.
Changing the AST representation did lead to a change in code generation that
interfered with the efficacy of CapturePropagation of key path literals; for
key path literals used as nonescaping closures, a mark_dependence of the
nonescaping function value on the key path was left behind, leaving the key
path object alive. The dependence is severed by the specialization done in
the pass, so update the pass to eliminate the dependence.
Compared to the previous patch, this version removes the attempt to have
the type-checked function expression carry the noescape-ness of its context,
and allows for coerceToType to introduce a function conversion instead, since
that FunctionConversionExpr is apparently load-bearing for default argument
generators.
Previously, we would turn a key path literal like `\.foo` in function type
context into a double-wrapped closure like this:
foo(\.x) // before type checking
foo({ $kp$ in { $0[$kp$] } }(\.x)) // after type checking
in order to preserve the evaluation semantics of the key path literal. This
works but leads to some awkward raw SIL generated out of SILGen which misses
out on various SILGen peepholes and requires a fair number of passes to clean
up. The semantics can still be preserved with a single layer of closure, by
using a capture list:
foo({[$kp$ = \.x] in $0[$kp$] }) // after type checking
which generates better natural code out of SILGen, and is also (IMO) easier
to understand on human inspection.
Changing the AST representation did lead to a change in code generation that
interfered with the efficacy of CapturePropagation of key path literals; for
key path literals used as nonescaping closures, a mark_dependence of the
nonescaping function value on the key path was left behind, leaving the key
path object alive. The dependence is severed by the specialization done in
the pass, so update the pass to eliminate the dependence.
Introduce the compiler directive `#_hasSymbol` which will be used to detect whether weakly linked symbols are present at runtime. It is intended for use in combination with `@_weakLinked import` or `-weak-link-at-target`.
```
if #_hasSymbol(foo(_:)) {
foo(42)
}
```
Parsing only; SILGen is coming in a later commit.
Resolves rdar://99342017
SequenceExprSyntax should have odd number elements. Previously 'a as b'
was parsed like:
```
(sequence_expr
(identifier_expr "a"),
(as_expr
'as'
(typeidentifier "b")))
```
So it had even number elements. Now it's parsed
```
(sequence_expr
(identifier_expr "a"),
(unresolved_as_expr 'as')
(type_expr
(typeidentifier "b")))
```
When the source code is invalid, this allows us to represent tokens that could not be used to form a valid syntax tree with more fidelity.
This commit does not start using GarbageNodes yet, it just sets everything up for them.
While skipping, if we encounter a token that looks
like it could be the start of a `/.../` regex
literal, fall back to parsing the function or type
body normally, as such a token could become a
regex literal. As such, it could treat `{` and
`}` as literal, or otherwise have contents that
would be lexically invalid Swift.
To avoid falling back in too many cases, we apply
the existing regex literal heuristics. Cases that
pass the heuristic fall back to regular parsing.
Cases that fail the heuristic are further checked
to make sure they wouldn't contain an unbalanced
`{` or `}`, but otherwise are allowed to be
skipped. This allows us to continue skipping for
most occurrences of infix and prefix `/`.
This is meant as a lower risk workaround to fix the
the issue, we ought to go back to handling regex
literals in the lexer.
Resolves rdar://95354010
Using the same feature set logic as experimental features, provide
feature names for "future" features, which are changes that will
become available with Swift 6. Use the feature check when determining
whether to implementation the feature instead of a language version
check, and map existing flags for these features (when available) over
to the feature set.
As an internal implementation detail, this makes it easier to reason
about when specific features are enabled (or not). If we decide to go
with piecemeal adoption support for features, it can provide an
alternative path to enabling features that feeds this mechanism.
Treat a prefix operator containing `/` the same as
the unapplied infix operator case, where we
tentatively lex. This means that we bail if there
is no closing `/` or the starting character is
invalid. This leaves binary operator containing
`/` in expression position as the last place where
we know that we definitely have a regex literal.
Teach the lexer not to consider `/` an operator
character when attempting to re-lex a regex
literal. This allows us to split off a prefix
operator.
Previously this was done after-the-fact in the
parser, but that didn't cover the unapplied infix
operator case, and didn't form a `tok::amp_prefix`
for `foo(&/.../)`, which led to a suboptimal
diagnostic.
This also now means we'll split an operator for
cases such as `foo(!/^/)` rather than treating it
as an unapplied infix operator.
rdar://92469917
This patch improves the error message emitted when the capture list
contains an item that is a sub-field of a struct/class/etc....
If the closure capture did not include `weak` at the beginning, the
presence of a period would cause the if-chain to fall through the
identifier checking, resulting in an error message about expecting a
`weak` keyword. Instead, I've opted to accept the period at that stage
of parsing so that we can fall through to a better error message.
For the following code
```
{ [self.field] in ... }
```
instead of emitting
`expected 'weak', 'unowned', or no specifier in capture list`,
we now emit
`fields may only be captured by assigning to a specific name`
with a fix-it that changes the code to
```
{ [ field = self.field ] in ... }
```
With `-enable-experimental-string-processing`,
start lexing `'` delimiters as regex literals (this
is just a placeholder delimiter for now). The
contents of which gets passed to the libswift
library, which can return an error string to be
emitted, or null for success.
The libswift side isn't yet hooked up to the Swift
regex parser, so for now just emit a dummy
diagnostic for regexes starting with quantifiers.
If successful, build an AST node which will be
emitted as an implicit call to an
`init(_regexString:)` initializer of an in-scope
`Regex` decl (which will eventually be a known
stdlib decl).
Now that the CSApply just uses components, we can
better split up the key path constructors to either
accept a set of resolved components, or a parsed
root or path.
Previously, when we reached the maximum nesting level, we changed the current token’s kind to an EOF token. A lot of places in the parser are not set up to expect this token change. The intended workaround was to check whether pushing a structure marker failed (which would change the token kind) and bail out parsing if this happened. This was fragile and caused assertion failures in assert builds.
Instead of changing the current token’s kind, and failing to push the structure marker, let the lexer know that it should cut off lexing, essentially making the input buffer stop at the current position. The parser will continue to consume its current token (`Parser.Tok`) and the next token that’s already lexed in the lexer (`Lexer.NextToken`) before reaching the emulated EOF token. Thus two more tokens are parsed than before, but that shouldn’t make much of a difference.
We used to avoid parsing labels for yield stmts
by passing in `SyntaxKind::YieldStmt` and relying
on this condition. However we no longer do so, we
pass `SyntaxKind::ExprList` instead.
Split up the expr list parsing members such that
there are separate entry points for tuple and
argument list parsing, and start using the argument
list parsing member for call and subscripts.
Completing a generic parameter to a nested type `Outer.Nested<#^COMPLETE^#>` crashed when used to initialize a local variable and returned no results when initializing a global variable.
This is because the generic argument parsing logic for nested types inside `parseExprPostfixSuffix` didn’t match that of global types in `parseExprIdentifier`. Make sure they do the same thing semantically.
Fixes rdar://77909679 [SR-14627]
This commit essentially consistes of the following steps:
- Add a new code completion key path component that represents the code completion token inside a key path. Previously, the key path would have an invalid component at the end if it contained a code completion token.
- When type checking the key path, model the code completion token’s result type by a new type variable that is unrelated to the previous components (because the code completion token might resolve to anything).
- Since the code completion token is now properly modelled in the constraint system, we can use the solver based code completion implementation and inspect any solution determined by the constraint solver. The base type for code completion is now the result type of the key path component that preceeds the code completion component.
This resolves bugs where code completion was not working correctly if the key path’s type had a generic base or result type. It’s also nice to have moved another completion type over to the solver-based implementation.
Resolves rdar://78779234 [SR-14685] and rdar://78779335 [SR-14703]
