Look in GHC.Unicode.toLower, and it is in terms of:

foreign import ccall unsafe “u_towlower”
towlower :: CInt -> CInt

So you’re calling out to C each time.

An ascii only version should be faster, e.g.,

toLower c@(C# c#)
| isAsciiUpper c = C# (chr# (ord# c# +# 32#))
| isAscii c = c
| isUpper c = unsafeChr (ord c `minusInt` ord ‘A’ `plusInt` ord ‘a’)
| otherwise = c

or your version.