[project @ 2002-11-07 11:42:48 by simonpj]

------------------
	Fix an obscure bug in implicit parameters,
	interacting with lazy pattern matching
		------------------

		MERGE TO STABLE BRANCH

The problem was this:

  data UniqueSupply = US Integer

  newUnique :: (?uniqueSupply :: UniqueSupply) => Integer
  newUnique = r
            where US r = ?uniqueSupply

The lazy pattern match in the where clause killed GHC 5.04 because the
SourceType {?uniqueSupply::UniqueSupply} of the RHS of the 'where' didn't
look like a UniqueSupply.

The fix is simple: in DsUtils.mkSelectorBinds, use the pattern, not
the rhs, to get the type reqd.  More efficient too.

Test is typecheck/should_compile/tc164.hs

diff --git a/ghc/compiler/deSugar/DsUtils.lhs b/ghc/compiler/deSugar/DsUtils.lhs
index 790097c..3411ebf 100644 (file)
--- a/ghc/compiler/deSugar/DsUtils.lhs
+++ b/ghc/compiler/deSugar/DsUtils.lhs
@@ -498,7 +498,25 @@ mkSelectorBinds (VarPat v) val_expr
 
 mkSelectorBinds pat val_expr
   | isSingleton binders || is_simple_pat pat
-  = newSysLocalDs (exprType val_expr)  `thenDs` \ val_var ->
+  =    -- Given   p = e, where p binds x,y
+       -- we are going to make
+       --      v = p   (where v is fresh)
+       --      x = case v of p -> x
+       --      y = case v of p -> x
+
+       -- Make up 'v'
+       -- NB: give it the type of *pattern* p, not the type of the *rhs* e.
+       -- This does not matter after desugaring, but there's a subtle 
+       -- issue with implicit parameters. Consider
+       --      (x,y) = ?i
+       -- Then, ?i is given type {?i :: Int}, a SourceType, which is opaque
+       -- to the desugarer.  (Why opaque?  Because newtypes have to be.  Why
+       -- does it get that type?  So that when we abstract over it we get the
+       -- right top-level type  (?i::Int) => ...)
+       --
+       -- So to get the type of 'v', use the pattern not the rhs.  Often more
+       -- efficient too.
+    newSysLocalDs (hsPatType pat)      `thenDs` \ val_var ->
 
        -- For the error message we make one error-app, to avoid duplication.
        -- But we need it at different types... so we use coerce for that