Inlining exported functions, -O vs -O3

Having the following unoptimized module with an exported and an internal function suitable for inlining:

(module
 (type $iii (func (param i32 i32) (result i32)))
 (export "add" (func $exported_add))
 (export "test" (func $test))
 (func $exported_add (; 0 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
  (i32.add
   (get_local $0)
   (get_local $1)
  )
 )
 (func $internal_sub (; 1 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
  (i32.sub
   (get_local $0)
   (get_local $1)
  )
 )
 (func $test (; 2 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
  (i32.add
   (call $exported_add
    (get_local $0)
    (get_local $1)
   )
   (call $internal_sub
    (get_local $0)
    (get_local $1)
   )
  )
 )
)

When optimizing with default optimization levels, this produces:

(module
 (type $iii (func (param i32 i32) (result i32)))
 (export "add" (func $exported_add))
 (export "test" (func $test))
 (func $exported_add (; 0 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
  (i32.add
   (get_local $0)
   (get_local $1)
  )
 )
 (func $test (; 1 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
  (i32.add
   (call $exported_add ;; exported call is not inlined
    (get_local $0)
    (get_local $1)
   )
   (i32.sub ;; internal call is inlined
    (get_local $0)
    (get_local $1)
   )
  )
 )
)

Which appears to keep the call to the exported function for size reasons (it inlines it in -O3, see below), but it probably should also inline in -O because an i32.add should always be smaller than a call with an additional function index immediate - is this correct?

For comparison, this is what's produced with -O3:

(module
 (type $iii (func (param i32 i32) (result i32)))
 (export "add" (func $exported_add))
 (export "test" (func $test))
 (func $exported_add (; 0 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
  (i32.add
   (get_local $0)
   (get_local $1)
  )
 )
 (func $test (; 1 ;) (type $iii) (param $0 i32) (param $1 i32) (result i32)
  (i32.add
   (i32.add ;; inlined as well
    (get_local $0)
    (get_local $1)
   )
   (i32.sub
    (get_local $0)
    (get_local $1)
   )
  )
 )
)

Fwiw, it also doesn't inline this call in -O2 without a shrinklevel, which seems odd, considering that -O3s also doesn't inline it (maybe) for size reasons.

Use the WAT module in the issue as the reproduction and compare output at -O, -O2, -O3, and -O3s, focusing on calls to exported versus internal functions. Determine whether the differing inlining decisions are intentional and whether the size behavior is consistent; done means the behavior is explained or corrected with an appropriate regression test.