Mill Computing, Inc. › Forums › The Mill › Architecture › Hard/soft realtime › Reply To: Hard/soft realtime
Mr. (Dr.?) Godard,
First off, thanks much for your reply to my question! Indeed, clarification helps; please continue with that, if you would.
In one of your talks describing how the Mill does function calls, you said words to the effect that, “The Mill provides each function call with the (illusion of) having a completely new belt, pre-populated with that function’s arguments.” I seem to recall that in the same talk, you mentioned that this behavior was accomplished by simply changing (incrementing?) the “BeltID,” and voila, new belt.
What I’m trying to learn is whether this behavior of the Mill (that all functions start with a new, pre-populated-with-args belt), however mechanized, implies some finite on-chip resource (e.g. counters of finite width, or a finite number of hardware-realized belt structures, etc.) that could be exhausted by, for instance, a deeply-nested sequence of short function calls, such that the Mill would eventually have to pause to spill some “state of multiple belts” to memory, in order to maintain the (illusion of) each new function starting with its own pre-populated belt. It is the HRT consequences of this sort of “save/restore” (if it exists in the Mill) that I’d like to better understand.
My thoughts on this may be contaminated by assumptions from other architectures (hardware stack machines, or SPARC’s large-but-finite number of windowing registers.) If there truly no resource limitation on the Mill’s “depth of stack-able belts”, other than the spiller’s memory bandwidth and memory size, that’s wonderful! In which case, I’ll be eagerly awaiting more detail on how the spiller works.
Since I haven’t yet seen anything in the Mill architecture that corresponds to a conventional stack pointer, may I request that you (or somebody else knowledgeable) add a bit more info to the glossary thread about what a Mill stack is and how Mill stacks, frames and BeltIDs work, as soon as possible, modulo all the patent filings to be done.
I’ll guess that there must be (more-persistent-than-belt-entry) special registers or their functional equivalents lurking under the hood, even if the programmer’s model cannot see (all of) them. The instruction/bundle pointer seems like one, and a stack pointer (or its functional equivalent) would seem to be another. My apologies if I’ve misinterpreted what I’ve read and watched.
Thanks much for your time, attention and reply!