If packing hits a blackhole, the packing thread should block
and retry, to get consistent behaviour at the Haskell level.
To enable this from Haskell, the core routine now checks and
evaluates the blackhole before retrying (i.e. blocks on the
known-to-be-blackholed data).
As the blackhole is not necessarily (and most likely not) the
graphroot, the C code writes its address into the buffer at
position 0. This is OK since packing fails anyway, so the buffer
won't hold any useful data.
Pack.c should now be thread-safe for unpacking and packing,
as its functions pass around an explicitly allocated state.
changes should eventually go back into the RTS code (but we
need to see the performance first).
The PAP case has not been rewritten yet, it should be revised
to pack the bitmap instead of tagging the packed stack elements.