Unstable if/whiles are errored

[calebmkim]

We decided on the following requirement for while loops/ if statements:

Either: a) the condition port for if/while loops is stable or b) they have a comb group (in which case the Calyx compiler will compile it such that the loop will be stable).

One test case that I wasn't quite sure how to change: tests/backends/mlir/with-guards.futil, found here. I see why there's an error, but wasn't quite sure how to change it, since I'm not completely sure what the test is doing ?

Some of the test cases were relying on ThisComponent's input ports. I changed these by making a 1 bit register that holds the value of ThisComponent's input port and writing to that register at the very beginning. It adds an extra cycle.
However, for our method of compiling static while loops with static bodies in #1431, I think it's important that the condition is not ThisComponent's port.

[rachitnigam]

The backend tests aren’t great but they just check to see if we can generate backend code

[rachitnigam]

LGTM!

[calebmkim]

Ok, merging! (@paili0628 I think this can help unblock #1431).

[sampsyo]

Looks good overall! I do also want to point out that this doesn't entirely address the semantics question from #1438 and there is still room to hammer out a rigorous definition there. Namely, as @calebmkim notes, the policy is:

Either: a) the condition port for if/while loops is stable or b) they have a comb group (in which case the Calyx compiler will compile it such that the loop will be stable).

This leads to two weird cases:

• while cond is illegal, but while cond with cg is legal, even though cond doesn't come from anything defined in cg and in fact has nothing to do at all with cg. This is a little too tied to the current way remove-comb-groups works. It seems like it should be legal, for example, for an optimization pass to prune dead comb groups, which would transform a legal loop into an illegal one under this policy.
• while cond with cg is classified as legal but cond is actually affected by other assignments, outside of cg, that render it unreliable as a loop condition.

So the "unfinished business" from #1438 remains: clearly defining what conds are legal, in English. This well-formedness check is necessary but insufficient; the rule we are gravitating toward requires dynamic reasoning.

[calebmkim]

Yeah, I see what your'e saying.

As for the two cases:

while cond is illegal, but while cond with cg is legal, even though cond doesn't come from anything defined in cg and in fact has nothing to do at all with cg. This is a little too tied to the current way remove-comb-groups works. It seems like it should be legal, for example, for an optimization pass to prune dead comb groups, which would transform a legal loop into an illegal one under this policy.

Interesting point. Maybe we should make the condition for a valid while loop even more restrictive. We could change criterion b) to be something like "has a comb group, and that comb group writes to the component that cond reads from"? I don't think this would be too difficult to check, and it (at least to me) sort of makes sense why we would want to enforce this.

while cond with cg is classified as legal but cond is actually affected by other assignments, outside of cg, that render it unreliable as a loop condition.

I think I'd have to think about a specific case of how this would happen. I'll def keep it in mind tho.

Edit: this is also just my quick thoughts. I think it may require careful thinking to lay out exactly what we want to require from while ports.

[sampsyo]

I do think this is on the right track! I honestly don't think a sound static check is the most urgent thing here, but it's something to think about eventually.

It seems like there is room to generalize the concept of @stable to include:

• combinational logic driven continuously and stably from @stable signals
• stuff that is locally stable by the above definition when a certain comb group is active

Basically, it seems like the desired policy involves some kind of "propagation" of stability, but exactly the rules for that propagation remain unclear.

[rachitnigam]

Ah, that's a good point @sampsyo @calebmkim! It better be the case that the cond in while cond with cg is defined by the combinational group. It would be definitely weird if we allowed anything else!

[rachitnigam]

@calebmkim can we open an issue about this?