Consolidate CC #21863
Consolidate CC #21863
Commits on Nov 22, 2024
-
-
-
Previously, we violated that assumption is we too the deep capture set of a capture reference wiht singleton type.
-
Align deep capture sets with reach capabilities
Count in dcs exactly those locations where a cap gets replaced by a reach capability.
-
Refactor handling of applications
Simplify code that handles applications, avoiding adding pieces of mutable state.
-
Handle reach capabilities correctly in markFree
The correct point to address charging reach capabilities is in markFree itself: When a reach capability goes out of scope, and that capability is not a parameter, we need to continue with the underlying capture set.
-
-
Drop the scheme where we only charge the last arrow of a curried lambda with elements used in the body. On the one hand, this is unsound without compensation measures (like, restricting to reach capabilities, or taking all capture sets of a named curried function as the underlying reference). On the other hand, this should be generalized to all closures and anonymous functions forming the right hand sides of methods.
-
Stop markFree at nested methods
Use sets of nested methods are anyway charged on call.
-
-
Re-enable handleEtaExpansionsSpecially setting
This gives better error messages. Previously we thought this would make reach capabilities unsound, but I don't see an issue with the latest design.
-
Disable reach capabilities in nested closures appearing without @use
With our current @use scheme, this is unsound. We leave the possibility to re-enable as a Config option which is disabled by default and comes with a warning that enabling it would be unsound.
-
Check that type parameters of methods and parent traits don't get instantiated with types containing a `cap` anywhere in covariant or invariant position.
-
-
-
-
Fix dcs for invariant type parameters
Also: add test that reach capabilities are contained inside boxes
-
-
Be more careful computing underlying types of reach capabilities
We can use the dcs only if there are no type variables.
-
-
The additional purity in the asInstanceOf target is not needed
-
-
Drop legacy caps.unsafe operations
Retain only caps.unsafe.unsafeAssumePure
-
Don't show redudundant existental wrappers
Don't show an `(ex$n: Exists) ->` if the bound variable does not appear in the result. The full type will be shown under -Ycc-debug. Also, avoid spurious ineffective mappings in widenReach.
-
-
Treat by-name closures specially in recheck
A by-name Closure node, which is produced by phase ElimByName gets a target type to indicate it's a contextual zero parameter closure. But for the purposes of rechecking and capture checking, it needs to be treated like a function. In particular the type of the closure needs to be derived from the result type of the anonymous function. Fixes scala#21920
-
-
-
-
-
