Fix data races in type canonicalization (#3852)

Fix two potential sources of data races identified with the help of thread
sanitizer.

First, keep a lock on the global HeapType store as long as it can reach
temporary types to ensure that no other threads observe the temporary types, for
example if another thread concurrently constructs a new HeapType with the same
shape as one being canonicalized here. This cannot happen with Types because
they are hashed in the global store by pointer identity, which has not yet
escaped the builder, rather than shape.

Second, in the shape canonicalizer, do not replace children of the new, minimal
HeapTypeInfos if they are already canonical. Even though these writes are always
no-ops, they still race because they are visible to other threads via canonical
Types.

0 files changed, 0 insertions, 0 deletions