Also, existing strageness in insn_idx being incremented for non-insns?
It shouldn't be - it's just using the presence of the memory
instrumentation as a proxy for the start of a instruction and dealing
with the slightly different start of block boundary.
Should it be named something else? I haven't looked at how it's
really used in the end.
We need the insn idx to find the registered callbacks for a given
instruction later. We could maybe embed a metadata TCGOp that could
track this for us but that might make TCG a bit more confusing as it
doesn't really need that information?