Ensure word atomicity in System.arraycopy
The VM needs to ensure that updates to 16-bit and 32-bit fields
and array elements happen atomically. System.arraycopy was calling
memmmove(), which in bionic happens to copy a byte at a time.
The new plan is to have bionic provide a platform-optimized memmove
variant that makes guarantees about word sizes. If it is not present
(either because it hasn't been written for the current platform,
or we're not linking against bionic), we will use a trivial copy loop.
We assume that the memmove-by-word implementation does whatever
it needs to for best performance. For example, if the function
determines that memcpy() is word-safe and will be faster, it can
call that. The VM no longer makes assumptions about the internal
workings of libc functions.
I also rearranged the code to avoid making indirect calls, reading
function addresses from memory, and using integer multiplication
instructions when a simple shift will do. (gcc now generates the
whole thing as one function.)
Also, added some primitive array copy tests to 011-array-copy.
Bug 3398352
Change-Id: I4677ee1c87998697a93e61f39a48b3f391e8e11b
4 files changed