Use unsafeWrite for speed

src/Data/Text/Internal/Reverse.hs

@@ -13,6 +13,7 @@ import GHC.Exts as Exts
 import Control.Monad.ST.Unsafe (unsafeIOToST)
 import Foreign.C.Types (CSize(..))
 #else
+import Control.Monad (when)
 import Control.Monad.ST (ST)
 import Data.Text.Internal.Encoding.Utf8 (utf8LengthByLeader)
 #endif
@@ -49,8 +50,18 @@ reversePoints
 reversePoints _ _ _ pOut | pOut < 0 = pure ()
 reversePoints src pIn dest pOut =
     let pointLength = utf8LengthByLeader (A.unsafeIndex src pIn)
+        pOut' = pOut - pointLength + 1
+        -- Repeated unsafeWrite is faster than copyI
+        copyByte i = A.unsafeWrite dest (pOut' + i) (A.unsafeIndex src (pIn + i))
     in do
-        A.copyI pointLength dest (pOut - pointLength + 1) src pIn
+        -- Unrolled for speed
+        copyByte 0
+        when (pointLength > 1) $ do
+            copyByte 1
+            when (pointLength > 2) $ do
+                copyByte 2
+                when (pointLength > 3) $
+                    copyByte 3
         reversePoints src (pIn + pointLength) dest (pOut - pointLength)
 #else
 reverse (Text (A.ByteArray ba) off len) = runST $ do