`indexWord8ArrayAsWord64#` not working on ARM?

During this bytestring PR we found that the following code causes a segfault on ARM Linux (32bit) (tested with GHC-8.10.1 and 8.10.7):

reverse :: ShortByteString -> ShortByteString
reverse = \sbs ->
    let l = length sbs
        ba = asBA sbs
    in create l (\mba -> go ba mba l)
  where
    go :: BA -> MBA s -> Int -> ST s ()
    go !ba !mba !l = case l `quotRem` 8 of
      (0, r) -> void $ goWord8Chunk 0 r
      (q, 0) -> goWord64Chunk 0 0 q
      (q, r) -> do
        i' <- goWord8Chunk 0 r
        goWord64Chunk i' 0 q
     where
      goWord64Chunk !off !i !cl
        | i >= cl = return ()
        | otherwise = do
            let w = indexWord64Array ba (off + (i * 8))
            writeWord64Array mba (cl - 1 - i) (byteSwap64 w)
            goWord64Chunk off (i+1) cl

      goWord8Chunk !i !cl
        | i >= cl = return i
        | otherwise = do
            let w = indexWord8Array ba i
            writeWord8Array mba (l - 1 - i) w
            goWord8Chunk (i+1) cl

indexWord64Array :: BA -> Int -> Word64
indexWord64Array (BA# ba#) (I# i#) = W64# (indexWord8ArrayAsWord64# ba# i#)

writeWord64Array :: MBA s -> Int -> Word64 -> ST s ()
writeWord64Array (MBA# mba#) (I# i#) (W64# w#) =
  ST $ \s -> case writeWord64Array# mba# i# w# s of
               s -> (# s, () #)

When replaced with a variant that only uses indexWord8Array/writeWord8Array the issue was gone.

We've been speculating about issues due to unaligned access, for which a CPP like this is used in other places. However @Bodigrim suggested this should maybe be handled by indexWord8ArrayAsWord64#?

Is this expected?