TypeRefiningGUFA errors on visitStructCmpxchg etc.

[kripken]

binaryen/src/ir/possible-contents.cpp

Lines 1076 to 1078 in 66c7166

	// TODO: Model the modification part of the RMW in addition to the read and
	// the write.
	addRoot(curr);

@tlively you added 4 of these TODOs and I think we need to fix them. Fuzzer testcase:

(module
 (type $struct (struct (field (mut eqref))))

 (func $atomic (param $0 (ref $struct)) (param $1 eqref) (result eqref)
  (struct.atomic.rmw.cmpxchg $struct 0
   (local.get $0)
   (local.get $1)
   (local.get $1)
  )
 )

 (func $call
  (drop
   (call $atomic
    (struct.new_default $struct)
    (ref.i31
     (i32.const 0)
    )
   )
  )
 )
)

$ bin/wasm-opt a.wat -all --type-refining-gufa --closed-world 
[wasm-validator error in function atomic] struct.atomic.rmw.cmpxchg replacement value must have the proper type, on 
(struct.atomic.rmw.cmpxchg $struct 0
 (local.get $0)
 (local.get $1)
 (local.get $1)
)
Fatal: error after opts

The problem is that this is not counted as a write, so GUFA thinks nothing is written to the field, and makes it nullref. And it then fails to validate to write an eqref there.

[kripken]

(actually the todos say we model the read&write already - but we don't)

[tlively]

I guess I've been forgetting to add new GC RMW test cases to the fuzzer ignore list. I'll do that as a short-term fix, but yeah, we'll need to fix this for real. Would you be able to take a look? I think it would take me a while to figure out what to do here, but I can do it if you'd prefer.