Make sure to test everything with scalar as well to catch any weird edge
cases.
Not all opcodes are covered here, just the arithmetic ones. FP64 packing
is also ignored.
This is quite complex since we cannot flush Phi inside the case labels,
we have to do it outside by emitting a lot of manual branches ourselves.
This should be extremely rare, but we need to handle this case.
We made the mistake of registering a dependency on the atomic variable
even if the atomic result was forced to a temporary. There is no need to
register reads from atomic variables like this as we always force atomic
results to a temporary and argument read/writes do not need to be
tracked.
-1 (0xffffffff) literal means the component should be undefined.
Since we cannot express undefined directly, just use a 0 literal in the
appropriate type.
MSL does not support value semantics for arrays (sigh), so we need to
force constant references and deal with copies if we have a different
address space than what we end up guessing.
