FYI @dotnet/jit-contrib

Still a work in progress, but figured I'd ask for some feedback before getting too much further. Things I know of that are still to do:

• We may need to handle case where an allocatable local is not live across a funclet boundary and ends up in different registers in different regions. Currently for STORE_LCL_VAR we always take the reg from the lcl var.
• ABI classifier type support for funclet params?
• Check if typeless catch won't actually get an ex obj from runtime (I suppose we can pass nullptr if necessary)
• Codegen is still off -- there is a bogus block before funclet start, and no return
• This won't validate (I'm pretty sure) until the jit host extracts the funclets into separate functions.

@SingleAccretion this may be more complex than what you had envisioned.

@SingleAccretion 90066eb should cover most of your feedback except internal regs.

Tagging subscribers to 'arch-wasm': @lewing, @pavelsavara
See info in area-owners.md if you want to be subscribed.

Hopefully closer now...

@SingleAccretion in the future maybe just point out one instance of a recurring pattern? Otherwise stuff may get overlooked.

There's still an issue where a node can be collected twice, eg if it both needs a register and also has internal registers, and double processing a collected node is not good.

We could have the collector mark nodes perhaps, use this to avoid a double collection, and unmark them when we process them later.

Also RequestInternalRegister can be called multiple times for the same node.

Ok, one more batch of updates:

• rely on allocator's m_currentFunclet to track what assignments the lcl var dscs hold. If we think they're already correct, verify that they are, in debug.
• remove need for LIR flag by remembering the last collected node
• add node on why we don't need a debug live range update like LSRA has

@dotnet/jit-contrib PTAL

LGTM but I don't understand all of it.

/ba-g infrastructure issues