Mill Computing, Inc. Forums The Mill Architecture Execution Reply To: Execution

Post count: 78

Please forgive me if these questions have been answered elsewhere. But since this topic mentions both logical/physical belt positions as well as frame IDs, I hope it’s OK to pose here.

However the Mill implements the belt and new-belt-per-function-call, there’s a finite number of belt IDs in any given implementation.
Q0: Correct?

I assume that a sufficiently deep series of function calls will need more belt IDs than the (particular family member’s) hardware supports.
Q1: In which case, what is the Mill’s behavior?

I’ll guess that some mechanism (the spiller or a close relative) has to offload a sizable chunk of internal state, so there will be available Belt IDs for the rest of those function calls. (And the spilled Belts/Belt-ID-state must eventually be refilled to complete the deeply-nested series of calls.)

Q2: If the Mill occasionally has to perform such Belt-ID spill/fill operations, does the compiler have any choice about the granularity of the spill/fill (e.g. a full spill, half, quarter, exactly Z beltIDs), or is the entire belt-ID internal state always spilled? The reason I ask is that infrequent, expensive digressions can cause problems for workloads with tight timing constraints. (And I’m hoping the Mill can be applied to such workloads, in a reasonably clean way.)

Thanks in advance for any explanation.