I guess this would turn what's currently an arcane crash message due to unused output to a silent transient register pressure. It's an improvement, but I'd want some feedback mechanism, maybe a debug!()?

But also, test failures.

We used to control every LIR instruction in YJIT's design, so it was useful to have a notification to remove unused LIR instructions because it was easy to identify why it happens and remove it, which also helps us simplify the implementation.

Now that we have an extra abstraction layer of HIR with multiple possible behaviors (HIR opt or not), I guess it's harder to remove unused instructions in ZJIT? I'm open to taking this direction, but I'd like to understand what's happening and why we need this, which I don't think are explained in Shopify#614 (comment) or Shopify#614 (comment).

Alternatively, if this happens only when you disable the HIR optimizer, we could take this path when it's disabled, and make it panic (probably with a better message than what we see today) when it's enabled. At least for the production mode, unused instructions are something to be fixed, somewhere in HIR, LIR, or the LIR backend.

In Shopify#614 it happens because the HIR ends up looking like this in the block:

Optimized HIR:
fn block in [email protected]:4:
bb0(v0:BasicObject, v1:BasicObject):
  v2:NilClass = Const Value(nil)
  v3:Any = Snapshot FrameState { pc: 0x0, stack: [], locals: [v1, v2] }
  v4:BasicObject = GetLocal l1, EP@3
  v5:Fixnum[1] = Const Value(1)
  Return v5

The GetLocal generates a load but it's never used. It's not cleaned up because we haven't run the optimizer.

Thanks for the explanation. It makes sense to allow such LIR to be compiled then. I still want the debug! log (at least when the HIR optimizer is enabled) in case the HIR optimizer misses it though.

I think you can call rb_zjit_prepare_options() in the called Option::unwrap() on a None value-ing tests to make debug!() work.