The machine verifier enforces the property that no block is listed twice in the successor or predecessor list. For a function with the optnone function this is not always true after instruction selection. Example:

> cat test.ll
define void @bar() #0 {
entry:
  br i1 undef, label %exit, label %exit

exit:
  ret void
}

attributes #0 = { noinline optnone }
> llc -verify-machineinstrs test.ll
...
*** Bad machine code: MBB has duplicate entries in its successor list. ***
...

I'm not sure why the machine verifier enforces this property, without the optnone attribute the codegenprepare appears to remove those pointless jumps though I am not sure yet whether this is on purpose to avoid the verification problems.