local-cse in O3 introduces side-effecting across structurally isolated sub-trees thus prevent const prop & DCE

Given the following code:

(module
  (import "External" "external_function" (func $external_function))
  (func $_start (param $0 i64) (param $1 i32) (param $2 i32) (param $3 i32)
    (local $4 i32) (local $5 i32)
    i32.const 0
    call $foo
    drop
  )
  (func $foo (param $0 i32) (result i32)
    (local $1 i32) (local $2 i32) (local $4 i32) (local $5 i32) (local $scratch i32) (local $3 i64) (local $6 i64)
      i32.const 77986
      i32.load16_u
      i32.const 16
      local.tee $2
      i32.shl
      local.get $2
      i32.shr_s
      i32.const 65533
      i32.eq
      i32.const 77986
      i32.load16_u
      i32.const 16
      i32.shl
      i32.and
      if (result i32)  ;; label = @4
        call $external_function
        i32.const 0
      else
        i32.const 1
      end
    )
  (memory $0 258 258)
  (export "_start" (func $_start)))

wasm-opt (d0d970cb5) eliminates the dead br_if body by -all -O2 but can not do that by -all -O3.

Analysis

Similar but different to #7440, this time wasm-opt in O3 introduces side-effect in more complex structure by local-cse:

The figure below depicts how the local.cse transforms the input:

It introduces the local.tee for less code size, however, it also affects the constant propagation and further blocks the dead code elimination.

Different from #7440, whose solution is simply enhancing the optimization of the binaryOp regardless that the side effects of its children, this issues is more complex because the local-cse does CSE in different depth and sub-tree, is there any code logic which can check or deal with this issue?

Start by reproducing the provided WebAssembly module with wasm-opt d0d970cb5 using -all -O2 and -all -O3, then inspect the local-cse transformation and its interaction with constant propagation and dead-code elimination. Compare the resulting optimization trees, focusing on the introduced local.tee across structurally isolated sub-trees. Done means O3 can eliminate the dead branch without introducing an invalid side-effect dependency.