GitLab

If you compile the following file:

module Main (main) where

import qualified Data.Char as C
import           Data.Char (isLetter, isLower)
import           Data.Maybe (mapMaybe)
import qualified Data.Text as T
import           Data.Text (Text)

main :: IO ()
main = print $ t_toTitle_1stNotLower
             $ T.pack "\4309" -- ვ

-- Taken from the text test suite
-- https://github.com/haskell/text/blob/44ec2cee65e5326ed943370e424f60d4ae6206d1/tests/Tests/Properties.hs#L388-L396
t_toTitle_1stNotLower :: Text -> Bool
t_toTitle_1stNotLower = and . notLow . T.toTitle . T.filter stable
    where notLow = mapMaybe (fmap (not . isLower) . (T.find isLetter)) . T.words
          -- Surprise! The Spanish/Portuguese ordinal indicators changed
          -- from category Ll (letter, lowercase) to Lo (letter, other)
          -- in Unicode 7.0
          -- Oh, and there exist lowercase-only letters (see previous test)
          stable c = if isLower c
                     then C.toUpper c /= c
                     else c /= '\170' && c /= '\186'

Then you'll get different runtime results depending on which version of GHC you use. Here is the answer with GHC 8.6.5:

$ /opt/ghc/8.6.5/bin/ghc Bug.hs -fforce-recomp -v0 && ./Bug                                        
True

And here is the answer with GHC 8.8.1, 8.10.1-alpha2, and HEAD:

$ /opt/ghc/8.8.1/bin/ghc Bug.hs -fforce-recomp -v0 && ./Bug                                        
False

This discrepancy manifests itself regardless of what GHC optimization level is used.