This fixes a source of non-determinism. The IDE/complete_constructor
test would sometimes fail depending on the order in which prior tests
ran, since those prior tests might populate the code completion cache.
C++ atomic's fetch_sub returns the previous value, where we want to
check the new value. This was causing massive memory leaks in SourceKit.
For ThreadSafeRefCountedBase, just switch to the one in LLVM that's
already correct. We should move the VPTR one to LLVM as well and then
we can get rid of this header.
rdar://problem/27358273
When we miss the in-memory (libcache-based) code completion cache, we
can now chain to an on-disk code completion cache. This drastically
improves the time and peak memory usage it takes to do the first code
completion (ie. before the in-memory cache is warm) if we've done the
same lookup before.
The on-disk cache, like the in-memory cache is tied to the specific
compiled swift and clang module files (.swiftmodule and .pcm), and will
consider itself out of date if they are modified. Responsibility for
deleting completely dead/unreachable cache files falls to the client.
Most of this commit is adding a simple serialization and deserialization
for CodeCompletionResults and CodeCompletionStrings. The format is very
simple, using an array of fixed size CodeCompletionResults, with offsets
into two blobs: one for CodeCompletionString::Chunks, and one for
strings. Currently that gives us about 5.8 MB for all the results in
Cocoa, but it's very compressible if we decide we want to reduce it
(gzip'd it is ~1.2 MB for the same data).
Swift SVN r28369
