Several fixes in order to make handling of types more consistent.
In particular:
- Expression types used within the constraint system itself always use
the type map.
- Prior to calling out to any TypeChecker functions, the types are
written back into the expression tree. After the call, the types are
read back into the constraint system type map.
- Some calls to directly set types on the expressions are reintroduced
in places they make sense (e.g. when building up new expressions that
are then passed to typeCheckExpressionShallow).
ConstraintSystem::setType() is still setting the type on the expression
nodes at the moment, because there is still some incorrect handling of
types that I need to track down (in particular some issues related to
closures do not appear to be handled correctly). Similarly, when we
cache the types in the constraint system map, we are not clearing them
from the expression tree yet.
The diagnostics have not been updated at all to use the types from
the constraint system where appropriate, and that will need to happen as
well before we can remove the write from ConstraintSystem::setType()
into the expression and enable the clearing of the expression tree type
when we cache it in the constraint system.
Previously this would cause strange diagnostics to be emitted when
values of type Bool! were the subject of if-expression conditions.
var isBoolean: Bool! = false
let conditional = isBoolean ? "Broken" : "Heart"
// ^ type 'Bool' is broken
Instead, look through IOUs before performing the member access to cover
our bases here.
These are used from within constraint system code, and for those uses we
need to be reading from the constraint system type map.
Add the parallel constraint system interfaces that call into the
Expr interfaces with the appropriate accessors.
- TypeMatchVisitor doesn't like seeing ErrorTypes -- stop if we have one.
- Common logic for replacing type parameters and variables with
UnresolvedType.
- Fix crash in coerceToType() if one of the two types has an UnresolvedType
in it, but not at the top level.
Fixes one compiler_crasher and some regressions with upcoming patches.
Protocol methods returning optionals, metatypes and functions
returning 'Self' are now handled correctly in Sema when
accessed with a base of existential type, eg:
protocol Clonable {
func maybeClone() -> Self?
func cloneMetatype() -> Self.Type
func getClonerFunc() -> () -> Self
}
let c: Clonable = ...
let _: Clonable = c.maybeClone()
let _: Clonable.Type = c.cloneMetatype()
Previously, we were unable to model this properly, for
several reasons:
- When opening the function type, we replaced the return
value in openedFullType _unconditionally_ with the
existential type. This was broken if the return type
was not the 'Self' parameter, but rather an optional,
metatype or function type involving the 'Self' parameter.
- It was also broken because we lost information; if the
return type originally contained an existential, we had
no way to draw the distinction. The 'hasArchetypeSelf()'
method was a hint that the original type contained a
type parameter, but it was inaccurate in some subtle cases.
- Since we performed this replacement on both the
'openedFullType' and 'openedType', CSApply had to "undo"
it to replace the existential type with the opened
existential archetype. This is because while the formal
type of the access returns the existential type, in
reality it returns the opened type, and CSApply has to
insert the correct ErasureExpr.
This was also done unconditionally, but even if it were
done more carefully, it would do the wrong thing because
of the 'loss of information' above.
- There was something fishy going on when we were dealing
with a constructor that was declared on the Optional type
itself, as seen in the fixed type_checker_crasher.
- TypeBase::eraseOpenedExistential() would transform a
MetatypeType of an opened existential into a MetatypeType
of an existential, rather than an ExistentialMetatypeType
of the existential type.
The new logic has the following improvements:
- When opening a function type, we replace the 'Self' type
parameter with the existential type in 'openedType', but
*not* 'openedFullType'. This simplifies CSApply, since it
doesn't have to "undo" this transformation when figuring
out what coercions to perform.
- We now only use replaceCovariantResultType() when working
with Self-returning methods on *classes*, which have more
severe restrictions than protocols. For protocols, we walk
the type using Type::transform() and handle all the cases
properly.
- Not really related to the above, but there was a whole
pile of dead code here concerning associated type references.
Note that SILGen still needs support for function conversions
involving an existential erasure; in the above Clonable example,
Sema now models this properly but SILGen still crashes:
let _: () -> Clonable = c.getClonerFunc()
Fixes <rdar://problem/28048391>, progress on <rdar://problem/21391055>.
When we have a value 'p' of type 'P.Type' for some protocol 'P',
we require the user to write 'p.init()' rather than 'p()' to
construct an instance of a concrete type conforming to 'P'.
If they write 'p()', we suggest a fixit to insert '.init',
however if there is also a subsequent member access on the
result, for example 'p().foo', we would crash.
Another invalid case that used to crash is when you construct
a protocol metatype directly and then access a member of it,
eg. 'P().foo'.
Force unwrapping the expression and propagating that type down to the
rest of the tree causes crashes when we go to request a different set
of protocols than we were expecting from it later. Make this
transformation local to the apply instead.
withoutActuallyEscaping has a signature like `<T..., U, V, W> (@nonescaping (T...) throws<U> -> V, (@escaping (T...) throws<U> -> V) -> W) -> W, but our type system for functions unfortunately isn't quite that expressive yet, so we need to special-case it. Set up the necessary type system when resolving an overload set to reference withoutActuallyEscaping, and if a type check succeeds, build a MakeTemporarilyEscapableExpr to represent it in the type-checked AST.
`type(of:)` has behavior whose type isn't directly representable in Swift's type system, since it produces both concrete and existential metatypes. In Swift 3 we put in a parser hack to turn `type(of: <expr>)` into a DynamicTypeExpr, but this effectively made `type(of:)` a reserved name. It's a bit more principled to put `Swift.type(of:)` on the same level as other declarations, even with its special-case type system behavior, and we can do this by special-casing the type system we produce during overload resolution if `Swift.type(of:)` shows up in an overload set. This also lays groundwork for handling other declarations we want to ostensibly behave like normal declarations but with otherwise inexpressible types, viz. `withoutActuallyEscaping` from SE-0110.
We were dropping the substitutions required to call the default
argument generator on the floor, when the correct solution is to
remap them to substitutions in terms of the base class signature.
Fixes <rdar://problem/29721571>.
The prior formulation of bridging conversions allowed conversion to
more-optional types, e.g., converting an "NSDate" to "Date?", which
was broken by my recent refactoring in this area. Allow bridging
conversions to more-optional types by introducing extra optional
injections at the end.
Fixes rdar://problem/29780527.
Specialize and improve the "downcast only unwraps optionals"
diagnostic to provide specific diagnostics + Fix-Its for the various
casts of forced cast, conditional cast, and "isa" check. Specifically:
* With a forced cast, customize the diagnostic. We still insert the
appropriate number of !'s, but now we remove the 'as! T' (if an
implicit conversion would suffice) or replace the 'as!' with 'as'
(if we still need a bridge)
* With a conditional cast, only emit a diagnostic if we're removing
just one level of optional. In such cases, we either have a no-op
(an implicit conversion would do) or we could just use 'as' to the
optional type, so emit a customized warning to do that. If we are
removing more than one level of optional, don't complain:
conditional casts can remove optionals. Add the appropriate Fix-Its
here.
* With an 'is' expression, only emit a diagnostic if we're removing
just one level of optional. In this case, the 'is' check is
equivalent to '!= nil'. Add a Fix-It for that.
Across the board, reduce the error to a warning. These are
semantically-well-formed casts, it's just that they could be written
better.
Fixes rdar://problem/28856049 and rdar://problem/22275685.
The type checker implements logic for handling checked casts in two
places: the constraint solver (for type-checking expressions
containing "as!" or "as?") and as a top-level entrypoint for
type-checking as?/as! for diagnostics and is/as patterns. Needless to
say, the two implementations were inconsistent, and in fact both were
wrong, leading to various problems---rejecting perfectly-valid "as!"
and "as?" casts outright, bogus warnings that particular as!/as? casts
always-succeed or always-fail when they wouldn't, and so on.
Start detangling the mess in two ways. First, drastically simplify the
handling of checked casts in the constraint solver, eliminating the
unprincipled "subtype" constraint checks that (among other things)
broke the handling of checked casts that involved bridging or optional
unwrapping. The simpler code is more permissive and more correct; it
essentially accepts that the user knows what she is doing with the
cast.
Second, make the type checker's checking of casts far more thorough,
which includes:
* When we're performing a collection cast, actually check that the
element types (and key types, for a dictionary) are castable, rather
than assuming all collection casts are legitimate. This means we'll
get more useful "always fails" and "always succeeds" diagnostics for
array/set/dictionary.
* Handle casts from a bridged value type to a subclass of the
corresponding bridged class type. Previously, these would be
incorrectly classified as "always fails".
While I'm here, eliminate a spurious diagnostic that occurs when using
a conditional cast ("as?") that could have been a coercion/bridging
conversion ("as"). The optional injection we synthesize to get the
resulting type correct was getting diagnosed as an implicit coercion,
but shouldn't have been.
Previously, bridging conversions were handled as a form of "explicit
conversion" that was treated along the same path as normal
conversions in matchTypes(). Historically, this made some
sense---bridging was just another form of conversion---however, Swift
now separates out bridging into a different kind of conversion that is
available only via an explicit "as". This change accomplishes a few
things:
* Improves type inference around "as" coercions. We were incorrectly
inferring type variables of the "x" in "x as T" in cases where a
bridging conversion was expected, which cause some type inference
failures (e.g., the SR-3319 regression).
* Detangles checking for bridging conversions from other conversions,
so it's easier to isolate when we're applying a bridging
conversion.
* Explicitly handle optionals when dealing with bridging conversions,
addressing a number of problems with incorrect diagnostics, e.g.,
complains about "unrelated type" cast failures that would succeed at
runtime.
Addresses rdar://problem/29496775 / SR-3319 / SR-2365.
- The DeclContext versions of these methods have equivalents
on the DeclContext class; use them instead.
- The GenericEnvironment versions of these methods are now
static methods on the GenericEnvironment class. Note that
these are not made redundant by the instance methods on
GenericEnvironment, since the static methods can also be
called with a null GenericEnvironment, in which case they
just assert that the type is fully concrete.
- Remove some unnecessary #includes of ArchetypeBuilder.h
and GenericEnvironment.h. Now changes to these files
result in a lot less recompilation.
Changes:
* Terminate all namespaces with the correct closing comment.
* Make sure argument names in comments match the corresponding parameter name.
* Remove redundant get() calls on smart pointers.
* Prefer using "override" or "final" instead of "virtual". Remove "virtual" where appropriate.
- TypeAliasDecl::getAliasType() is gone. Now, getDeclaredInterfaceType()
always returns the NameAliasType.
- NameAliasTypes now always desugar to the underlying type as an
interface type.
- The NameAliasType of a generic type alias no longer desugars to an
UnboundGenericType; call TypeAliasDecl::getUnboundGenericType() if you
want that.
- The "lazy mapTypeOutOfContext()" hack for deserialized TypeAliasDecls
is gone.
- The process of constructing a synthesized TypeAliasDecl is much simpler
now; instead of calling computeType(), setInterfaceType() and then
setting the recursive properties in the right order, just call
setUnderlyingType(), passing it either an interface type or a
contextual type.
In particular, many places weren't setting the recursive properties,
such as the ClangImporter and deserialization. This meant that queries
such as hasArchetype() or hasTypeParameter() would return incorrect
results on NameAliasTypes, which caused various subtle problems.
- Finally, add some more tests for generic typealiases, most of which
fail because they're still pretty broken.
In Swift 3.0.1, argument labels are ignored when calling a function
having a single parameter of 'Any' type. That is, if we have:
func foo(_: Any) {}
Both of the following were accepted in a no-assert build (an assert
build would crash, but the GM builds of Xcode ship with asserts off):
foo(123)
foo(data: 123)
This behavior was fixed by 578e36a7e1,
but unfortunately we have to revert to the old behavior *and* defeat
the assertion when in Swift 3 mode.
Swift 4 mode still has the correct behavior, where the second call
'foo(data: 123)' produces a diagnostic.
Now, I have to pour myself a strong drink to forget this ever happened.
Fixes <rdar://problem/28952837>.
There's a general problem where a SubscriptExpr has an argument
that's a LoadExpr loading a tuple from an lvalue. For some reason
we don't construct the ParenExpr in this case, which confused
CSDiag.
Also, in Swift 3 mode, add a total hack to fudge things in
matchCallArguments() in the case where we erroneously lost
ParenType sugar.
This parameter implements getType() for the given expression, making
it possible to use this from within the constraint system, which now
has it's own side map for types of expressions.
Update CSGen/CSApply/CSSolver to primarily use getType() from
ConstraintSystem.
Currently getType() just returns the type on the expression. As with
setType(), which continues to set the type on the expression, this
will be updated once all the other changes are in place.
This change also moves coerceToRValue from TypeChecker to
CosntraintSystem so that it can access the expression type map in the
constraint system.
I've been unable to reproduce the issue that hit on the builder, and
still expect this change to be NFC, so trying it out again. If it
fails, I'll revert again and try another shot at reproducing.
Original commit message:
We create new expressions that have the type on the expression
set. Make sure we capture these types in the constraint system type
map so that we can refer to types uniformally by consulting the map.
NFC.
(cherry picked from commit 57d5d974ff)
We create new expressions that have the type on the expression
set. Make sure we capture these types in the constraint system type
map so that we can refer to types uniformally by consulting the map.
NFC.
The setType() function in ConstraintSystem still sets the types on the
expressions themselves and that will continue until everything is
moved over to using the map.
Unfortunately this exposed cases where we are currently setting the
type multiple times to different values, so I've commented out the
assert that I previously added. I will circle back and start looking
into those issues once everything is moved over.
NFC for now.
