diff options
Diffstat (limited to 'compiler/Language/Haskell/Syntax/Expr.hs')
| -rw-r--r-- | compiler/Language/Haskell/Syntax/Expr.hs | 155 |
1 files changed, 4 insertions, 151 deletions
diff --git a/compiler/Language/Haskell/Syntax/Expr.hs b/compiler/Language/Haskell/Syntax/Expr.hs index 0e54adb8f4..3d6500d342 100644 --- a/compiler/Language/Haskell/Syntax/Expr.hs +++ b/compiler/Language/Haskell/Syntax/Expr.hs @@ -178,12 +178,8 @@ data HsExpr p -- The renamer renames record-field selectors to HsRecFld -- The typechecker preserves HsRecFld - | HsOverLabel (XOverLabel p) - (Maybe (IdP p)) FastString + | HsOverLabel (XOverLabel p) FastString -- ^ Overloaded label (Note [Overloaded labels] in GHC.OverloadedLabels) - -- @Just id@ means @RebindableSyntax@ is in use, and gives the id of the - -- in-scope 'fromLabel'. - -- NB: Not in use after typechecking | HsIPVar (XIPVar p) HsIPName -- ^ Implicit parameter (not in use after typechecking) @@ -224,8 +220,8 @@ data HsExpr p -- | Operator applications: -- NB Bracketed ops such as (+) come out as Vars. - -- NB We need an expr for the operator in an OpApp/Section since - -- the typechecker may need to apply the operator to a few types. + -- NB Sadly, we need an expr for the operator in an OpApp/Section since + -- the renamer may turn a HsVar into HsRecFld or HsUnboundVar | OpApp (XOpApp p) (LHsExpr p) -- left operand @@ -343,8 +339,6 @@ data HsExpr p -- See Note [Empty lists] | ExplicitList (XExplicitList p) -- Gives type of components of list - (Maybe (SyntaxExpr p)) - -- For OverloadedLists, the fromListN witness [LHsExpr p] -- | Record construction @@ -472,7 +466,7 @@ data HsExpr p | XExpr !(XXExpr p) -- Note [Trees that Grow] extension constructor for the - -- general idea, and Note [Rebindable syntax and HsExpansion] + -- general idea, and Note [Rebindable syntax and HsExpansion] in GHC.Hs.Expr -- for an example of how we use it. -- | The AST used to hard-refer to GhcPass, which was a layer violation. For now, @@ -484,147 +478,6 @@ type family PendingTcSplice' p -- --------------------------------------------------------------------- -{- -Note [Rebindable syntax and HsExpansion] -~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ - -We implement rebindable syntax (RS) support by performing a desugaring -in the renamer. We transform GhcPs expressions affected by RS into the -appropriate desugared form, but **annotated with the original expression**. - -Let us consider a piece of code like: - - {-# LANGUAGE RebindableSyntax #-} - ifThenElse :: Char -> () -> () -> () - ifThenElse _ _ _ = () - x = if 'a' then () else True - -The parsed AST for the RHS of x would look something like (slightly simplified): - - L locif (HsIf (L loca 'a') (L loctrue ()) (L locfalse True)) - -Upon seeing such an AST with RS on, we could transform it into a -mere function call, as per the RS rules, equivalent to the -following function application: - - ifThenElse 'a' () True - -which doesn't typecheck. But GHC would report an error about -not being able to match the third argument's type (Bool) with the -expected type: (), in the expression _as desugared_, i.e in -the aforementioned function application. But the user never -wrote a function application! This would be pretty bad. - -To remedy this, instead of transforming the original HsIf -node into mere applications of 'ifThenElse', we keep the -original 'if' expression around too, using the TTG -XExpr extension point to allow GHC to construct an -'HsExpansion' value that will keep track of the original -expression in its first field, and the desugared one in the -second field. The resulting renamed AST would look like: - - L locif (XExpr - (HsExpanded - (HsIf (L loca 'a') - (L loctrue ()) - (L locfalse True) - ) - (App (L generatedSrcSpan - (App (L generatedSrcSpan - (App (L generatedSrcSpan (Var ifThenElse)) - (L loca 'a') - ) - ) - (L loctrue ()) - ) - ) - (L locfalse True) - ) - ) - ) - -When comes the time to typecheck the program, we end up calling -tcMonoExpr on the AST above. If this expression gives rise to -a type error, then it will appear in a context line and GHC -will pretty-print it using the 'Outputable (HsExpansion a b)' -instance defined below, which *only prints the original -expression*. This is the gist of the idea, but is not quite -enough to recover the error messages that we had with the -SyntaxExpr-based, typechecking/desugaring-to-core time -implementation of rebindable syntax. The key idea is to decorate -some elements of the desugared expression so as to be able to -give them a special treatment when typechecking the desugared -expression, to print a different context line or skip one -altogether. - -Whenever we 'setSrcSpan' a 'generatedSrcSpan', we update a field in -TcLclEnv called 'tcl_in_gen_code', setting it to True, which indicates that we -entered generated code, i.e code fabricated by the compiler when rebinding some -syntax. If someone tries to push some error context line while that field is set -to True, the pushing won't actually happen and the context line is just dropped. -Once we 'setSrcSpan' a real span (for an expression that was in the original -source code), we set 'tcl_in_gen_code' back to False, indicating that we -"emerged from the generated code tunnel", and that the expressions we will be -processing are relevant to report in context lines again. - -You might wonder why we store a RealSrcSpan in addition to a Bool in -the TcLclEnv: could we not store a Maybe RealSrcSpan? The problem is -that we still generate constraints when processing generated code, -and a CtLoc must contain a RealSrcSpan -- otherwise, error messages -might appear without source locations. So we keep the RealSrcSpan of -the last location spotted that wasn't generated; it's as good as -we're going to get in generated code. Once we get to sub-trees that -are not generated, then we update the RealSrcSpan appropriately, and -set the tcl_in_gen_code Bool to False. - ---- - -A general recipe to follow this approach for new constructs could go as follows: - -- Remove any GhcRn-time SyntaxExpr extensions to the relevant constructor for your - construct, in HsExpr or related syntax data types. -- At renaming-time: - - take your original node of interest (HsIf above) - - rename its subexpressions (condition, true branch, false branch above) - - construct the suitable "rebound"-and-renamed result (ifThenElse call - above), where the 'SrcSpan' attached to any _fabricated node_ (the - HsVar/HsApp nodes, above) is set to 'generatedSrcSpan' - - take both the original node and that rebound-and-renamed result and wrap - them in an XExpr: XExpr (HsExpanded <original node> <desugared>) - - At typechecking-time: - - remove any logic that was previously dealing with your rebindable - construct, typically involving [tc]SyntaxOp, SyntaxExpr and friends. - - the XExpr (HsExpanded ... ...) case in tcExpr already makes sure that we - typecheck the desugared expression while reporting the original one in - errors - --} - --- See Note [Rebindable syntax and HsExpansion] just above. -data HsExpansion a b - = HsExpanded a b - deriving Data - --- | Build a "wrapped" 'HsExpansion' out of an extension constructor, --- and the two components of the expansion: original and desugared --- expressions. --- --- See Note [Rebindable Syntax and HsExpansion] above for more details. -mkExpanded - :: (HsExpansion a b -> b) -- ^ XExpr, XCmd, ... - -> a -- ^ source expression ('GhcPs') - -> b -- ^ "desugared" expression - -- ('GhcRn') - -> b -- ^ suitably wrapped - -- 'HsExpansion' -mkExpanded xwrap a b = xwrap (HsExpanded a b) - --- | Just print the original expression (the @a@). -instance (Outputable a, Outputable b) => Outputable (HsExpansion a b) where - ppr (HsExpanded a b) = ifPprDebug (vcat [ppr a, ppr b]) (ppr a) - --- --------------------------------------------------------------------- - -- | A pragma, written as {-# ... #-}, that may appear within an expression. data HsPragE p = HsPragSCC (XSCC p) |