[commit: ghc] wip/nested-cpr: Add Note [non-algebraic or open body type warning] (b1561d1)

[git at git.haskell.org]

Repository : ssh://git@git.haskell.org/ghc

On branch  : wip/nested-cpr
Link       : http://ghc.haskell.org/trac/ghc/changeset/b1561d1250f0006ba5a4f0f6e265cb7081451f3e/ghc

>---------------------------------------------------------------

commit b1561d1250f0006ba5a4f0f6e265cb7081451f3e
Author: Joachim Breitner <mail at joachim-breitner.de>
Date:   Wed Dec 4 17:12:07 2013 +0000

    Add Note [non-algebraic or open body type warning]


>---------------------------------------------------------------

b1561d1250f0006ba5a4f0f6e265cb7081451f3e
 compiler/stranal/WwLib.lhs |   20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/compiler/stranal/WwLib.lhs b/compiler/stranal/WwLib.lhs
index fc94c9b..4acf255 100644
--- a/compiler/stranal/WwLib.lhs
+++ b/compiler/stranal/WwLib.lhs
@@ -540,6 +540,7 @@ mkWWcpr body_ty res
        Just con_tag | Just stuff <- deepSplitCprType_maybe con_tag body_ty
                     -> mkWWcpr_help stuff
                     |  otherwise
+                       -- See Note [non-algebraic or open body type warning]
                     -> WARN( True, text "mkWWcpr: non-algebraic or open body type" <+> ppr body_ty )
                        return (id, id, body_ty)
 
@@ -590,6 +591,25 @@ mkUnpackCase scrut co uniq boxing_con unpk_args body
     bndr = mk_ww_local uniq (exprType casted_scrut)
 \end{code}
 
+Note [non-algebraic or open body type warning]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+There are a few cases where the W/W transformation is told that something
+returns a constructor, but the type at hand doesn't really match this. One
+real-world example involves unsafeCoerce:
+  foo = IO a
+  foo = unsafeCoere c_exit
+  foreign import ccall "c_exit" c_exit :: IO ()
+Here CPR will tell you that `foo` returns a () constructor for sure, but trying
+to create a worker/wrapper for type `a` obviously fails.
+(This was a real example until ee8e792  in libraries/base.)
+
+It does not seem feasilbe to avoid all such cases already in the analyser (and
+after all, the analysis is not really wrong), so we simply do nothing here in
+mkWWcpr. But we still want to emit warning with -DDEBUG, to hopefully catch
+other cases where something went avoidably wrong.
+
+
 Note [Profiling and unpacking]
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 If the original function looked like