Mill Computing, Inc. Forums The Mill Tools Simulators Simulation Reply To: Simulation

Ivan Godard
Post count: 689

1) You list only the arguments to the label; there’s no way to list the belt contents anyway, because only the specializer knows that. At a join point anything else on the belt is dead anyway. Note that falling into the label is equivalent to an (elided) go to the label from the falling code.

2) Sorry my explanation wasn’t clear. Labels, like functions, can have an arbitrary argument list in the source, including the genAsm source, and functions (but not labels) can be VARARGS. The specializer maps these arguments in a member-dependent way to a set that are passed on the belt, and a remainder that are passed in memory (functions) or scratchpad (labels). To a first approximation you may assume that leading arguments that are less than or equal to the size of the largest member-supported scalar will be on the belt in argument order, while larger arguments (in particular quad data on a non-quad member) may be split into a tuple and passed piecemeal.

I grant that “must fit on the belt” was misleading; I was thinking in terms of the specializer implementation. Internally the specializer turns label declarations into fewer-argument labels (for which all arguments can go on the belt) plus a set of scratchpad locations that are reserved in the scratchpad map so they won’t be used for spill/fill. The go operations are also transformed into a set of explicit spill ops plus a conform for what gets passed on the belt, plus the transfer op itself (br and friends).

At the join, the current argument set comprises the incoming stuff on the belt and the set of scratchpad locations. About the only thing you could do at a multi-argument label in genAsm is to rebind the incoming arguments, and the rebindings name whatever belt position or scratchpad location they name, and can be used normally thereafter. Loops are not yet done, but I expect that the back-transfer will use the same mechanism to get all the loop-carried data to the next iteration.

VARARGS handling is not yet finalized. Currently we expect to pass the entire “…” list in memory (arguments before the “…” pass normally) but we may come up with better ideas at some point.