When we're not serializing SIL for all function bodies, @_transparent
functions can only reference internal functions that are declared
@_versioned, otherwise there's no serialized body and no public entry
point, so any client that inlines the @_transparent function will
not be able to link.
This patch adds the minimum set of @_versioned declarations to allow
a non-optimized build of the standard library and overlays.
Recall that this attribute is just a temporary hack to make progress
on building the standard library with resilience enabled.
Once availability and resilience learn about each other, @_versioned
will be replaced by having an availability annotation on an internal
declaration. Invariants will be diagnosed by Sema instead of asserting
in the SIL verifier.
Finally, the set of "internal but available" declarations will
eventually be audited instead of determined by experimentation.
This almost closes out https://bugs.swift.org/browse/SR-267.
The remaining issue is an interaction between SIL optimizations and
serialization that will be fixed with some upcoming changes to the
optimizer.
The Generator interface is specialized for iteration and is often faster than indexing.
Using _initialize_to results in a ~10% speedup for Dictionary. For collections with more complex iteration state (such as search trees with O(log(n)) indexing), this leads to a complexity class improvement.
This requires a small change to the testsuite, because the generate() method of Defaulted*RangeReplaceableSlice called Array(self), which now leads to infinite recursion.
https://bugs.swift.org/browse/SR-610
...because it is apparently more efficient in some cases. Technically
we don't do this in ALL places, because it would be unfortunate if
the implementation of _successorInPlace() were self recursive :-)
An optimization should be added in order for the new one to be
efficient, i.e. if the `count` value is equal to `1`, the underlying
`Builtin.destroy` should be called, instead of
`Builtin.destroyArray`.
At DaveA's suggestion, I took a mostly mechanical approach to this:
pointers and numeric types start using += 1, and indexes use
i = i.successor(). The index model is likely to be revised in
Swift 3 anyway, so micro-optimizing this code syntactically isn't
super important.
There is some performance concern of this patch, since some
in-place succesor operations are more efficient than
i = i.successor(). The one that seems particularly at issue is the
instance in the implementation of partition(), which I changed to
use i._successorInPlace(). If other instances lead to a perf issue,
they can be changed to use that as well.
