GHC 9.0 or later gives duplicate -Wincomplete-uni-patterns warnings

Consider this program:

{-# OPTIONS_GHC -Wincomplete-uni-patterns #-}
module Bug where

a, b, c :: Bool
a = False
b = False
c = False

mx :: Maybe ()
mx = Just ()

f :: ()
f | a = ()
  | otherwise
  = let Just x = mx
     in x

If you compile this, you'll get an -Wincomplete-uni-patterns warning, as expected:

$ ghc-9.2 Bug.hs -fforce-recomp
[1 of 1] Compiling Bug              ( Bug.hs, Bug.o )

Bug.hs:15:9: warning: [-Wincomplete-uni-patterns]
    Pattern match(es) are non-exhaustive
    In a pattern binding:
        Patterns of type ‘Maybe ()’ not matched: Nothing
   |
15 |   = let Just x = mx
   |         ^^^^^^^^^^^

Now change the definition of f so that its first guard is predicated on both a and b:

f :: ()
f | a, b = ()
  | otherwise
  = let Just x = mx
     in x

This time, the warning will look strange:

$ ghc-9.2 Bug.hs -fforce-recomp
[1 of 1] Compiling Bug              ( Bug.hs, Bug.o )

Bug.hs:15:9: warning: [-Wincomplete-uni-patterns]
    Pattern match(es) are non-exhaustive
    In a pattern binding:
        Patterns of type ‘Maybe ()’ not matched:
            Nothing
            Nothing
   |
15 |   = let Just x = mx
   |         ^^^^^^^^^^^

Notice that there are duplicate warnings about the Nothing pattern! In fact, the more conjuncts you add, the more duplicates there will be. For instance, making the first guard be | a, b, c will result in a total of three warnings about Nothing.

Ideally, we'd only have just one warning about Nothing. In fact, this was the behavior prior to GHC 9.0:

$ ghc-8.10 Bug.hs -fforce-recomp
[1 of 1] Compiling Bug              ( Bug.hs, Bug.o )

Bug.hs:15:9: warning: [-Wincomplete-uni-patterns]
    Pattern match(es) are non-exhaustive
    In a pattern binding: Patterns not matched: Nothing
   |
15 |   = let Just x = mx
   |         ^^^^^^^^^^^

This regression was introduced at some point between GHC 8.10 and 9.0, as 9.0 is the first version to exhibit the bug.