The two GatherKinds no longer share any implementation, so there's
no point keeping the logic together. Doing this also allows removing
the acceptConstraintFn from gatherAllConstraints(), which further
simplifies depthFirstSearch().
Fixes a regression from commit 0c128e5db7.
The old depthFirstSearch() walked all adjacencies via the constraint graph,
and thus it would visit type variables that are currently inactive because
we're solving a conjunction element.
This was inconsistent with the new union-find which only formed the
connected components from the currently active type variables; adjacencies
involving inactive type variables are no longer considered.
Fix the inconsistency by changing gatherConstraints(), which used from
addTypeVariableConstraintsToWorkList(), to also skip inactive type
variables.
Fixes rdar://problem/143340082.
Don't attempt this optimization if call has number literals.
This is intended to narrowly fix situations like:
```swift
func test<T: FloatingPoint>(_: T) { ... }
func test<T: Numeric>(_: T) { ... }
test(42)
```
The call should use `<T: Numeric>` overload even though the
`<T: FloatingPoint>` is a more specialized version because
selecting `<T: Numeric>` doesn't introduce non-default literal
types.
This algorithm attempts to ensure that the solver always picks a disjunction
it knows the most about given the previously deduced type information.
For example in chains of operators like: `let _: (Double) -> Void = { 1 * 2 + $0 - 5 }`
The solver is going to start from `2 + $0` because `$0` is known to be `Double` and
then proceed to `1 * ...` and only after that to `... - 5`.
The algorithm is pretty simple:
- Collect "candidate" types for each argument
- If argument is bound then the set going to be represented by just one type
- Otherwise:
- Collect all the possible bindings
- Add default literal type (if any)
- Collect "candidate" types for result
- For each disjunction in the current scope:
- Compute a favoring score for each viable* overload choice:
- Compute score for each parameter:
- Match parameter flags to argument flags
- Match parameter types to a set of candidate argument types
- If it's an exact match
- Concrete type: score = 1.0
- Literal default: score = 0.3
- Highest scored candidate type wins.
- If none of the candidates match and they are all non-literal
remove overload choice from consideration.
- Average the score by dividing it by the number of parameters
to avoid disfavoring disjunctions with fewer arguments.
- Match result type to a set of candidates; add 1 to the score
if one of the candidate types matches exactly.
- The best choice score becomes a disjunction score
- Compute disjunction scores for all of the disjunctions in scope.
- Pick disjunction with the best overall score and favor choices with
the best local candidate scores (if some candidates have equal scores).
- Viable overloads include:
- non-disfavored
- non-disabled
- available
- non-generic (with current exception to SIMD)
Also introduce two new frontend flags:
The -solver-scope-threshold flag sets the maximum number of scopes, which was
previously hardcoded to 1 million.
The -solver-trail-threshold flag sets the maximum number of trail steps,
which defaults to 64 million.
When we replay a solution, we must record changes in the trail, so fix the
logic to do that. This fixes the first assertion failure with this test case.
The test case also exposed a second issue. We synthesize a CustomAttr in
applySolutionToClosurePropertyWrappers() with a type returned by simplifyType().
Eventually, CustomAttrNominalRequest::evaluate() looks at this type, and passes
it to directReferencesForType(). Unfortunately, this entry point does not
understand type aliases whose underlying type is a type parameter.
However, directReferencesForType() is the wrong thing to use here, and we
can just call getAnyNominal() instead.
Fixes rdar://139237781.
`filterDisjunction` should ignore the choices that are already
disabled while attempting to optimize disjunctions related to
dynamic member lookup.
Resolves: rdar://139314763
When we are using diagnostic transactions to disable immediate emission
of diagnostics, `DiagnosticEngine::hadAnyError()` no longer accurately
reports whether an error occurred. Thread the DiagnosticTransaction
into the ConstraintSystem so we can also check whether it contains an
error before emitting the fallback diagnostic.
Fixes rdar://128272346.
