{-# LANGUAGE LambdaCase #-} ----------------------------------------------------------------------------- -- | -- Module : GHC.JS.Optimizer -- Copyright : (c) The University of Glasgow 2001 -- License : BSD-style (see the file LICENSE) -- -- Maintainer : Jeffrey Young -- Luite Stegeman -- Sylvain Henry -- Josh Meredith -- Stability : experimental -- -- -- * Domain and Purpose -- -- GHC.JS.Optimizer is a shallow embedding of a peephole optimizer. That is, -- this module defines transformations over the JavaScript IR in -- 'GHC.JS.Syntax', transforming the IR forms from inefficient, or -- non-idiomatic, JavaScript to more efficient and idiomatic JavaScript. The -- optimizer is written in continuation passing style so optimizations -- compose. -- -- * Architecture of the optimizer -- -- The design is that each optimization pattern matches on the head of a -- block by pattern matching onto the head of the stream of nodes in the -- JavaScript IR. If an optimization gets a successful match then it performs -- whatever rewrite is necessary and then calls the 'loop' continuation. This -- ensures that the result of the optimization is subject to the same -- optimization, /and/ the rest of the optimizations. If there is no match -- then the optimization should call the 'next' continuation to pass the -- stream to the next optimization in the optimization chain. We then define -- the last "optimization" to be @tailLoop@ which selects the next block of -- code to optimize and begin the optimization pipeline again. ----------------------------------------------------------------------------- module GHC.JS.Optimizer ( jsOptimize ) where import Prelude import GHC.JS.Syntax import Control.Arrow {- Note [Unsafe JavaScript optimizations] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ There are a number of optimizations that the JavaScript Backend performs that are not sound with respect to arbritrary JavaScript. We still perform these optimizations because we are not optimizing arbritrary javascript and under the assumption that the JavaScript backend will not generate code that violates the soundness of the optimizer. For example, the @deadCodeElim@ optimization removes all statements that occur after a 'return' in JavaScript, however this is not always sound because of hoisting, consider this program: function foo() { var x = 2; bar(); return x; function bar() { x = 10; }} which is transformed to: function foo() { var x = 2; bar(); return x; }} The optimized form is clearly a program that goes wrong because `bar()` is no longer defined. But the JavaScript backend will never generate this code, so as long as that assumption holds we are safe to perform optimizations that would normally be unsafe. -} -------------------------------------------------------------------------------- -- Top level Driver -------------------------------------------------------------------------------- jsOptimize :: JStat -> JStat jsOptimize = go where p_opt = jsOptimize opt = jsOptimize' e_opt = jExprOptimize -- base case go (BlockStat xs) = BlockStat (opt xs) -- recursive cases go (ForStat i p s body) = ForStat (go i) (e_opt p) (go s) (p_opt body) go (ForInStat b i p body) = ForInStat b i p (p_opt body) go (WhileStat b c body) = WhileStat b (e_opt c) (p_opt body) go (SwitchStat s ps body) = SwitchStat s (fmap (second go) ps) (p_opt body) go (FuncStat i args body) = FuncStat i args (p_opt body) go (IfStat c t e) = IfStat (e_opt c) (p_opt t) (p_opt e) go (TryStat ths i c f) = TryStat (p_opt ths) i (p_opt c) (p_opt f) go (LabelStat lbl s) = LabelStat lbl (p_opt s) -- special case: drive the optimizer into expressions go (AssignStat id op rhs) = AssignStat (e_opt id) op (e_opt rhs) go (DeclStat i (Just e)) = DeclStat i (Just $ e_opt e) go (ReturnStat e) = ReturnStat (e_opt e) go (UOpStat op e) = UOpStat op (e_opt e) go (ApplStat f args) = ApplStat (e_opt f) (e_opt <$> args) -- all else is terminal, we match on these to force a warning in the event -- another constructor is added go x@BreakStat{} = x go x@ContinueStat{} = x go x@DeclStat{} = x -- match on the nothing case jsOptimize' :: [JStat] -> [JStat] jsOptimize' = runBlockOpt opts . single_pass_opts where opts :: BlockOpt opts = safe_opts <> unsafe_opts <> tailLoop -- tailloop must be last, see module description unsafe_opts :: BlockOpt unsafe_opts = mconcat [ deadCodeElim ] safe_opts :: BlockOpt safe_opts = mconcat [ declareAssign, combineOps ] single_pass_opts :: BlockTrans single_pass_opts = runBlockTrans sp_opts sp_opts = [flattenBlocks] -- | recur over a @JExpr@ and optimize the @JVal@s jExprOptimize :: JExpr -> JExpr -- the base case jExprOptimize (ValExpr val) = ValExpr (jValOptimize val) -- recursive cases jExprOptimize (SelExpr obj field) = SelExpr (jExprOptimize obj) field jExprOptimize (IdxExpr obj ix) = IdxExpr (jExprOptimize obj) (jExprOptimize ix) jExprOptimize (UOpExpr op exp) = UOpExpr op (jExprOptimize exp) jExprOptimize (IfExpr c t e) = IfExpr c (jExprOptimize t) (jExprOptimize e) jExprOptimize (ApplExpr f args ) = ApplExpr (jExprOptimize f) (jExprOptimize <$> args) jExprOptimize (InfixExpr op l r) = InfixExpr op (jExprOptimize l) (jExprOptimize r) -- | drive optimizations to anonymous functions and over expressions jValOptimize :: JVal -> JVal -- base case jValOptimize (JFunc args body) = JFunc args (jsOptimize body) -- recursive cases jValOptimize (JList exprs) = JList (jExprOptimize <$> exprs) jValOptimize (JHash hash) = JHash (jExprOptimize <$> hash) -- all else is terminal jValOptimize x@JVar{} = x jValOptimize x@JDouble{} = x jValOptimize x@JInt{} = x jValOptimize x@JStr{} = x jValOptimize x@JRegEx{} = x -- | A block transformation is a function from a stream of syntax to another -- stream type BlockTrans = [JStat] -> [JStat] -- | A BlockOpt is a function that alters the stream, and a continuation that -- represents the rest of the stream. The first @BlockTrans@ represents -- restarting the optimizer after a change has happened. The second @BlockTrans@ -- represents the rest of the continuation stream. newtype BlockOpt = BlockOpt (BlockTrans -> BlockTrans -> BlockTrans) -- | To merge two BlockOpt we first run the left-hand side optimization and -- capture the right-hand side in the continuation instance Semigroup BlockOpt where BlockOpt opt0 <> BlockOpt opt1 = BlockOpt $ \loop next -> opt0 loop (opt1 loop next) instance Monoid BlockOpt where -- don't loop, just finalize mempty = BlockOpt $ \_loop next -> next -- | loop until a fixpoint is reached runBlockOpt :: BlockOpt -> [JStat] -> [JStat] runBlockOpt (BlockOpt opt) xs = recur xs where recur = opt recur id runBlockTrans :: [BlockTrans] -> [JStat] -> [JStat] runBlockTrans opts = foldl (.) id opts -- | Perform all the optimizations on the tail of a block. tailLoop :: BlockOpt tailLoop = BlockOpt $ \loop next -> \case [] -> next [] -- this call to jsOptimize is required or else the optimizer will not -- properly recur down JStat. See the 'deadCodeElim' test for examples which -- were failing before this change (x:xs) -> next (jsOptimize x : loop xs) -------------------------------------------------------------------------------- -- Single Slot Optimizations -------------------------------------------------------------------------------- {- | Catch modify and assign operators: case 1: i = i + 1; ==> ++i; case 2: i = i - 1; ==> --i; case 3: i = i + n; ==> i += n; case 4: i = i - n; ==> i -= n; -} combineOps :: BlockOpt combineOps = BlockOpt $ \loop next -> \case -- find a op pattern, and rerun the optimizer on its result unless there is -- nothing to optimize, in which case call the next optimization (unchanged@(AssignStat ident@(ValExpr (JVar i)) AssignOp (InfixExpr op (ValExpr (JVar i')) e)) : xs) | i == i' -> case (op, e) of (AddOp, (ValExpr (JInt 1))) -> loop $ UOpStat PreIncOp ident : xs (SubOp, (ValExpr (JInt 1))) -> loop $ UOpStat PreDecOp ident : xs (AddOp, e') -> loop $ AssignStat ident AddAssignOp e' : xs (SubOp, e') -> loop $ AssignStat ident SubAssignOp e' : xs _ -> next $ unchanged : xs -- commutative cases (unchanged@(AssignStat ident@(ValExpr (JVar i)) AssignOp (InfixExpr op e (ValExpr (JVar i')))) : xs) | i == i' -> case (op, e) of (AddOp, (ValExpr (JInt 1))) -> loop $ UOpStat PreIncOp ident : xs (SubOp, (ValExpr (JInt 1))) -> loop $ UOpStat PreDecOp ident : xs (AddOp, e') -> loop $ AssignStat ident AddAssignOp e' : xs (SubOp, e') -> loop $ AssignStat ident SubAssignOp e' : xs _ -> next $ unchanged : xs -- general case, we had nothing to optimize in this case so call the next -- optimization xs -> next xs -------------------------------------------------------------------------------- -- Dual Slot Optimizations -------------------------------------------------------------------------------- -- | Catch 'var i; i = q;' ==> 'var i = q;' declareAssign :: BlockOpt declareAssign = BlockOpt $ \loop next -> \case ( (DeclStat i Nothing) : (AssignStat (ValExpr (JVar i')) AssignOp v) : xs ) | i == i' -> loop (DeclStat i (Just v) : xs) xs -> next xs -- | Eliminate all code after a return statement. This is a special case -- optimization that doesn't need to loop. See Note [Unsafe JavaScript -- optimizations] deadCodeElim :: BlockOpt deadCodeElim = BlockOpt $ \_loop next -> \case (x@ReturnStat{}:_) -> next [x] xs -> next xs -- | remove nested blocks flattenBlocks :: BlockTrans flattenBlocks (BlockStat y : ys) = flattenBlocks y ++ flattenBlocks ys flattenBlocks (x:xs) = x : flattenBlocks xs flattenBlocks [] = []