The prefab'ed value witness tables for reference storage types are a
premature optimization. Not all scenarios are covered, and those that
are "look suspect" according to John McCall.
As part of this, rename TypeMetadataRecordKind to TypeReferenceKind
and consistently give it three bits of storage.
The better modelling of these type references appears to have been
sufficient to make dynamic conformance checks succeed, which is good
but unexpected.
One expression in the new hashing implementation is going exponential,
accounting for a huge amount of type-checking type. Add (admittedly ugly)
“as UInt64” annotations to greatly reduce the time to type-check this
expression.
*Ahem* type-checking time for the standard library goes from 24s->14s with
this change. Added a type-checker “slow” performance test and captured
the problem in rdar://problem/42672946.
NSObject defines its own interface for hashing and equality. Subclasses must implement custom hashing by overriding the hash property, or NSSet and NSDictionary will not work correctly.
Unfortunately, we currently define NSObject.hashValue and NSObject.hash(into:) as overridable, which leads people to accidentally override one of these instead of hash. hashValue is explicitly declared open, while hash(into:) is automatically synthesized as such.
Making hashValue non-open is a potentially source breaking change, but we can at least provide an explicit, non-overridable definition for hash(into:), and update the documentation to explain the contract.
ThreadLocalStorage is not accessible in an inlinable context, so drop
all such annotations. Also, drop a redundant one from Optional to
suppress a warning.
NFC
From what I see the only fields are DATA_VALUE_WITNESS which
all have type size_t. I converted them to use the target-dependent
`StoredSize`. While I was around I fixed also isValueInline()
to do the right thing (it was using ValueBuffer instead of
TargetValueBuffer) and all the getters for the data value witnesses.
<rdar://problem/41546568>
This is a bit easier than the fully general case where both the external descriptor and local pattern have captured arguments (because of generics or subscript indices) since we don't have to combine the two argument files in one component.
This was useful for bringup before compiler support for building key path patterns was implemented, but now it just hardcodes a bunch of implementation details that expose some internal details as ABI and interfere with refactorings. Remove it now that the same functionality is tested at a higher level.
Returning `Self` was probably a hack from when shifts were required to be homogenous and we didn't have good generic inits between integer types. Returning Int is more natural, because:
1. Dimensional analysis says that we shouldn't expect the logarithm to have the same type.
2. Int is guaranteed to be able to represent any object size, and BinaryInteger does not admit sparse representations, so the logarithm of the value cannot exceed allocated storage size.
3. Every single use case in the stdlib and tests is either unchanged or simplified by this change.
* Update std lib to Swift 5.0
* Disable Unicode.* warnings for now
* Slow path to resiliently handle the case where an unknown rounding rule is passed
* Remove resilience from Encoding/DecodingError (which should only happen on slow paths anyway)
* internal typealiases now need @usableFromInline
* Force inlining on Array._owner.get
