On Sun, 2011-11-06 at 16:37 +0000, Duncan Coutts wrote:
So I was preparing to commit this change in base and validating ghc when I discovered a more subtle issue in the pretty package:
Consider
a <> empty <+> b
So having tried to fix this by setting <> higher precedence than <+>, we simply run into the reverse problem a <+> empty <> b The concrete example is in ghc: ptext (sLit "In module") <+> quotes (ppr (is_mod decl_spec)) <+> source_import <> colon This produces strings like "In module `foo':" with the old/current fixities and produces "In module `foo' :" with my suggested fixities. The reason is that source_import is usually empty: source_import | mi_boot iface = ptext (sLit "(hi-boot interface)") | otherwise = empty so if we simplify, it's something like: test "foo" <+> empty <> colon which is "foo:" if it's all infixl, but becomes "foo :" if <> binds more tightly than <+>, since: a <+> empty <> b = a <+> (empty <> b) = a <+> b So maybe we should just conclude that pretty printing should be left associative, building from the left. Either that or we go and track down all the places that use empty in mixed <> <+> contexts. Or any other suggestions for fixes for the pretty package that'd let it use the proposed Data.Monoid.<> with infixr <> ? Could the behaviour of <+> be altered to work in a right associative world? Should we reconsider Data.Monoid.<> if we can't make it work with the pretty package? Duncan