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

Ivan Godard
Post count: 629

Remember, this is hardware, not compiler internals; there is no such thing as taints :-)

In hardware terms, your suggestion would turn into predication on the store, where the value of the predicate results from the exit test. There have been predicated machines, most recently Itanium and older ARMs; I did a compiler for the Data General Nova1200 once, which had predicated operations implemented as skip predicates; it was an interesting machine.

The problems arise in the implementation. Predication adds another operand to all the operations (or at least the non-speculable ones, on a Mill) which adds another data path between the Belt and the functional units, which is an expensive proposition. One also needs a way to encode the extra argument, which would completely change the encoding; store in particular fully saturates the argument space available. Then there’s the problem of keeping the predicate value live on the belt; it increases belt and scratchpad pressure.

This is not to say that it couldn’t be done; the Mill ISA already has conditional branches and conditional returns, which can be thought of as a form of predication. When the tool chain gets further and we start implementing pipelining in the specializer and try it on real code then we may rethink the design. For now, we don’t think we need anything beyond what we have now.