--------------------------------
-- Top level function inlining
--------------------------------
--- This transformation inlines top level bindings that have been generated by
--- the compiler and are really simple. Really simple currently means that the
--- normalized form only contains a single binding, which catches most of the
+-- This transformation inlines simple top level bindings. Simple
+-- currently means that the body is only a single application (though
+-- the complexity of the arguments is not currently checked) or that the
+-- normalized form only contains a single binding. This should catch most of the
-- cases where a top level function is created that simply calls a type class
-- method with a type and dictionary argument, e.g.
-- fromInteger = GHC.Num.fromInteger (SizedWord D8) $dNum
-- which is later called using simply
-- fromInteger (smallInteger 10)
--- By inlining such calls to simple, compiler generated functions, we prevent
--- huge amounts of trivial components in the VHDL output, which the user never
--- wanted. We never inline user-defined functions, since we want to preserve
--- all structure defined by the user. Currently this includes all functions
--- that were created by funextract, since we would get loops otherwise.
+--
+-- These useless wrappers are created by GHC automatically. If we don't
+-- inline them, we get loads of useless components cluttering the
+-- generated VHDL.
+--
+-- Note that the inlining could also inline simple functions defined by
+-- the user, not just GHC generated functions. It turns out to be near
+-- impossible to reliably determine what functions are generated and
+-- what functions are user-defined. Instead of guessing (which will
+-- inline less than we want) we will just inline all simple functions.
--
-- Only functions that are actually completely applied and bound by a
-- variable in a let expression are inlined. These are the expressions
-- that will eventually generate instantiations of trivial components.
-- By not inlining any other reference, we also prevent looping problems
-- with funextract and inlinedict.
---
--- Note that "defined by the compiler" isn't completely watertight, since GHC
--- doesn't seem to set all those names as "system names", we apply some
--- guessing here.
inlinetoplevel, inlinetopleveltop :: Transform
inlinetoplevel (LetBinding:_) expr =
case collectArgs expr of
- -- Any system name is candidate for inlining. Never inline
- -- user-defined functions, to preserve structure.
- (Var f, args) | not $ isUserDefined f -> do
+ (Var f, args) -> do
body_maybe <- needsInline f
case body_maybe of
Just body -> do
--------------------------------
-- Dictionary inlining
--------------------------------
--- Inline all top level dictionaries, so we can use them to resolve
--- class methods based on the dictionary passed.
-inlinedict c expr@(Var f) | Id.isDictId f = do
- body_maybe <- Trans.lift $ getGlobalBind f
+-- Inline all top level dictionaries, that are in a position where
+-- classopresolution can actually resolve them. This makes this
+-- transformation look similar to classoperesolution below, but we'll
+-- keep them separated for clarity. By not inlining other dictionaries,
+-- we prevent expression sizes exploding when huge type level integer
+-- dictionaries are inlined which can never be expanded (in casts, for
+-- example).
+inlinedict c expr@(App (App (Var sel) ty) (Var dict)) | not is_builtin && is_classop = do
+ body_maybe <- Trans.lift $ getGlobalBind dict
case body_maybe of
+ -- No body available (no source available, or a local variable /
+ -- argument)
Nothing -> return expr
- Just body -> change body
+ Just body -> change (App (App (Var sel) ty) body)
+ where
+ -- Is this a builtin function / method?
+ is_builtin = elem (Name.getOccString sel) builtinIds
+ -- Are we dealing with a class operation selector?
+ is_classop = Maybe.isJust (Id.isClassOpId_maybe sel)
-- Leave all other expressions unchanged
inlinedict c expr = return expr
projects / matthijs / master-project / cλash.git / blobdiff