Allow inline matches to bind type variables #5657
Conversation
abgruszecki
force-pushed
the
inline-type-binds
branch
from
89a164b to
8c32798
Compare
| @@ -1004,7 +1005,7 @@ trait Implicits { self: Typer => | |||
| val generated2 = | |||
| if (cand.isExtension) Applications.ExtMethodApply(generated1).withType(generated1.tpe) | |||
| else generated1 | |||
| SearchSuccess(generated2, ref, cand.level)(ctx.typerState) | |||
| SearchSuccess(generated2, ref, cand.level)(ctx.typerState, ctx.gadt) | |||
There was a problem hiding this comment.
Would it make sense to combine typerState and GADT map?
There was a problem hiding this comment.
I'm keeping my eye on it. I tried to do that, but couldn't find a clean approach. The problem is we don't want to set fresh typer state when we set fresh GADT state, and vice versa - it's really easiest if the two are separate objects. If current approach turns out to be too clumsy, I'll find a way to merge the two.
| ctx.gadt.addBound(sym, hi, isUpper = true) | ||
| } | ||
|
|
||
| def addTypeBindings(list: List[(TypeSymbol, Boolean)])(implicit ctx: Context): Unit = |
There was a problem hiding this comment.
Something looks suspicious here: list is a list of type variable occurrences. A symbol may appear several times in that list. Hence a symbol can appear several times in fromBuf with different copies.
There was a problem hiding this comment.
Maybe that's still OK since every symbol is referred to only once, at its binding location (not sure about that!). But then it would be good to make the accumulator findBoundVars more specific and robust. Why not combine findBoundVars and getBind?
There was a problem hiding this comment.
I've taken another look at my approach - it looks like currently it's really not possible to get multiple bindings in a list, but relying on that could be fragile and it's relatively obvious that we should do the same as IsFullyDefinedAccumulator, so I've reworked the List into a type TypeBindsMap = SimpleIdentityMap[TypeSymbol, java.lang.Boolean] and added appropriate value-merging logic to findBoundVars.
Unless I'm missing something, we cannot merge findBoundVars and getBinds, since the latter locates symbols bound by the current pattern (as opposed to enclosing patterns) and the former find the variance of the symbol by traversing the type. I've renamed findBoundVars to extractBindVariance and added a comment to make this clearer.
abgruszecki
force-pushed
the
inline-type-binds
branch
2 times, most recently
from
c1fe272 to
aa014d4
Compare
|
I've addresed the comments. The rebase was strictly to make the branch mergeable and didn't modify the commits that were there, modulo including 01078c4. |
Use SmartGADTMap and Constraint#approximation to infer and extract type values. Widen pattern-bound types in ProtoTypes so implicit search can find relevant candidates. When dropping unused bindings in Inliner, insert type bindings with a single pre-pass to correctly handle updating singleton types. Previously Select trees retained old type values, which prevented pickler roundtripping.
Re-typing the pattern in GADT context is unnecessary - reducing the pattern will on its own perform the same subtyping checks. Previously if a pattern binding a HK type was inlined, a warning saying that its kind is different from the kind of its parameter was emitted _after_ preparing for reduction. This replaced the type of the Bind node with an error, preventing the bind from being located and narrowed when reducing the inline match.
abgruszecki
force-pushed
the
inline-type-binds
branch
from
62a8499 to
e772dbb
Compare
Fixes #5405, #5574.
Github diff is unusually horrible for
dropUnusedBindings, so let me quote part of commit message: