\begin{code}
module SimplUtils (
-- Rebuilding
- mkLam, mkCase, prepareAlts, bindCaseBndr,
+ mkLam, mkCase, prepareAlts, tryEtaExpand,
-- Inlining,
preInlineUnconditionally, postInlineUnconditionally,
- activeInline, activeRule, inlineMode,
+ activeUnfolding, activeRule,
+ getUnfoldingInRuleMatch,
+ simplEnvForGHCi, updModeForInlineRules,
-- The continuation type
- SimplCont(..), DupFlag(..), LetRhsFlag(..),
+ SimplCont(..), DupFlag(..), ArgInfo(..),
+ isSimplified,
contIsDupable, contResultType, contIsTrivial, contArgs, dropArgs,
- countValArgs, countArgs,
- mkBoringStop, mkLazyArgStop, mkRhsStop, contIsRhsOrArg,
- interestingCallContext, interestingArgContext,
+ pushSimplifiedArgs, countValArgs, countArgs, addArgTo,
+ mkBoringStop, mkRhsStop, mkLazyArgStop, contIsRhsOrArg,
+ interestingCallContext,
interestingArg, mkArgInfo,
#include "HsVersions.h"
import SimplEnv
+import CoreMonad ( SimplifierMode(..), Tick(..) )
import DynFlags
import StaticFlags
import CoreSyn
import qualified CoreSubst
import PprCore
+import DataCon ( dataConCannotMatch )
import CoreFVs
import CoreUtils
-import Literal
+import CoreArity
import CoreUnfold
-import MkId
import Name
import Id
-import NewDemand
+import Var
+import Demand
import SimplMonad
-import Type
+import Type hiding( substTy )
+import Coercion hiding( substCo )
import TyCon
-import DataCon
-import Unify ( dataConCannotMatch )
import VarSet
import BasicTypes
import Util
+import MonadUtils
import Outputable
-import List( nub )
+import FastString
+import Pair
+
+import Data.List
\end{code}
\begin{code}
data SimplCont
= Stop -- An empty context, or hole, []
- OutType -- Type of the result
- LetRhsFlag
- Bool -- True <=> There is something interesting about
+ CallCtxt -- True <=> There is something interesting about
-- the context, and hence the inliner
-- should be a bit keener (see interestingCallContext)
- -- Two cases:
- -- (a) This is the RHS of a thunk whose type suggests
- -- that update-in-place would be possible
- -- (b) This is an argument of a function that has RULES
+ -- Specifically:
+ -- This is an argument of a function that has RULES
-- Inlining the call might allow the rule to fire
| CoerceIt -- C `cast` co
OutCoercion -- The coercion simplified
+ -- Invariant: never an identity coercion
SimplCont
| ApplyTo -- C arg
- DupFlag
- InExpr SimplEnv -- The argument and its static env
+ DupFlag -- See Note [DupFlag invariants]
+ InExpr StaticEnv -- The argument and its static env
SimplCont
| Select -- case C of alts
- DupFlag
- InId [InAlt] SimplEnv -- The case binder, alts, and subst-env
+ DupFlag -- See Note [DupFlag invariants]
+ InId [InAlt] StaticEnv -- The case binder, alts, and subst-env
SimplCont
-- The two strict forms have no DupFlag, because we never duplicate them
| StrictBind -- (\x* \xs. e) C
InId [InBndr] -- let x* = [] in e
- InExpr SimplEnv -- is a special case
+ InExpr StaticEnv -- is a special case
SimplCont
- | StrictArg -- e C
- OutExpr OutType -- e and its type
- (Bool,[Bool]) -- Whether the function at the head of e has rules,
- SimplCont -- plus strictness flags for further args
-
-data LetRhsFlag = AnArg -- It's just an argument not a let RHS
- | AnRhs -- It's the RHS of a let (so please float lets out of big lambdas)
+ | StrictArg -- f e1 ..en C
+ ArgInfo -- Specifies f, e1..en, Whether f has rules, etc
+ -- plus strictness flags for *further* args
+ CallCtxt -- Whether *this* argument position is interesting
+ SimplCont
+
+data ArgInfo
+ = ArgInfo {
+ ai_fun :: Id, -- The function
+ ai_args :: [OutExpr], -- ...applied to these args (which are in *reverse* order)
+ ai_rules :: [CoreRule], -- Rules for this function
+
+ ai_encl :: Bool, -- Flag saying whether this function
+ -- or an enclosing one has rules (recursively)
+ -- True => be keener to inline in all args
+
+ ai_strs :: [Bool], -- Strictness of remaining arguments
+ -- Usually infinite, but if it is finite it guarantees
+ -- that the function diverges after being given
+ -- that number of args
+ ai_discs :: [Int] -- Discounts for remaining arguments; non-zero => be keener to inline
+ -- Always infinite
+ }
-instance Outputable LetRhsFlag where
- ppr AnArg = ptext SLIT("arg")
- ppr AnRhs = ptext SLIT("rhs")
+addArgTo :: ArgInfo -> OutExpr -> ArgInfo
+addArgTo ai arg = ai { ai_args = arg : ai_args ai }
instance Outputable SimplCont where
- ppr (Stop ty is_rhs _) = ptext SLIT("Stop") <> brackets (ppr is_rhs) <+> ppr ty
- ppr (ApplyTo dup arg se cont) = ((ptext SLIT("ApplyTo") <+> ppr dup <+> pprParendExpr arg)
+ ppr (Stop interesting) = ptext (sLit "Stop") <> brackets (ppr interesting)
+ ppr (ApplyTo dup arg _ cont) = ((ptext (sLit "ApplyTo") <+> ppr dup <+> pprParendExpr arg)
{- $$ nest 2 (pprSimplEnv se) -}) $$ ppr cont
- ppr (StrictBind b _ _ _ cont) = (ptext SLIT("StrictBind") <+> ppr b) $$ ppr cont
- ppr (StrictArg f _ _ cont) = (ptext SLIT("StrictArg") <+> ppr f) $$ ppr cont
- ppr (Select dup bndr alts se cont) = (ptext SLIT("Select") <+> ppr dup <+> ppr bndr) $$
- (nest 4 (ppr alts)) $$ ppr cont
- ppr (CoerceIt co cont) = (ptext SLIT("CoerceIt") <+> ppr co) $$ ppr cont
+ ppr (StrictBind b _ _ _ cont) = (ptext (sLit "StrictBind") <+> ppr b) $$ ppr cont
+ ppr (StrictArg ai _ cont) = (ptext (sLit "StrictArg") <+> ppr (ai_fun ai)) $$ ppr cont
+ ppr (Select dup bndr alts se cont) = (ptext (sLit "Select") <+> ppr dup <+> ppr bndr) $$
+ (nest 2 $ vcat [ppr (seTvSubst se), ppr alts]) $$ ppr cont
+ ppr (CoerceIt co cont) = (ptext (sLit "CoerceIt") <+> ppr co) $$ ppr cont
-data DupFlag = OkToDup | NoDup
+data DupFlag = NoDup -- Unsimplified, might be big
+ | Simplified -- Simplified
+ | OkToDup -- Simplified and small
+
+isSimplified :: DupFlag -> Bool
+isSimplified NoDup = False
+isSimplified _ = True -- Invariant: the subst-env is empty
instance Outputable DupFlag where
- ppr OkToDup = ptext SLIT("ok")
- ppr NoDup = ptext SLIT("nodup")
+ ppr OkToDup = ptext (sLit "ok")
+ ppr NoDup = ptext (sLit "nodup")
+ ppr Simplified = ptext (sLit "simpl")
+\end{code}
+Note [DupFlag invariants]
+~~~~~~~~~~~~~~~~~~~~~~~~~
+In both (ApplyTo dup _ env k)
+ and (Select dup _ _ env k)
+the following invariants hold
+ (a) if dup = OkToDup, then continuation k is also ok-to-dup
+ (b) if dup = OkToDup or Simplified, the subst-env is empty
+ (and and hence no need to re-simplify)
+\begin{code}
-------------------
-mkBoringStop :: OutType -> SimplCont
-mkBoringStop ty = Stop ty AnArg False
+mkBoringStop :: SimplCont
+mkBoringStop = Stop BoringCtxt
-mkLazyArgStop :: OutType -> Bool -> SimplCont
-mkLazyArgStop ty has_rules = Stop ty AnArg (canUpdateInPlace ty || has_rules)
+mkRhsStop :: SimplCont -- See Note [RHS of lets] in CoreUnfold
+mkRhsStop = Stop (ArgCtxt False)
-mkRhsStop :: OutType -> SimplCont
-mkRhsStop ty = Stop ty AnRhs (canUpdateInPlace ty)
+mkLazyArgStop :: CallCtxt -> SimplCont
+mkLazyArgStop cci = Stop cci
+-------------------
+contIsRhsOrArg :: SimplCont -> Bool
contIsRhsOrArg (Stop {}) = True
contIsRhsOrArg (StrictBind {}) = True
contIsRhsOrArg (StrictArg {}) = True
-contIsRhsOrArg other = False
+contIsRhsOrArg _ = False
-------------------
contIsDupable :: SimplCont -> Bool
-contIsDupable (Stop {}) = True
-contIsDupable (ApplyTo OkToDup _ _ _) = True
-contIsDupable (Select OkToDup _ _ _ _) = True
+contIsDupable (Stop {}) = True
+contIsDupable (ApplyTo OkToDup _ _ _) = True -- See Note [DupFlag invariants]
+contIsDupable (Select OkToDup _ _ _ _) = True -- ...ditto...
contIsDupable (CoerceIt _ cont) = contIsDupable cont
-contIsDupable other = False
+contIsDupable _ = False
-------------------
contIsTrivial :: SimplCont -> Bool
-contIsTrivial (Stop {}) = True
+contIsTrivial (Stop {}) = True
contIsTrivial (ApplyTo _ (Type _) _ cont) = contIsTrivial cont
-contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
-contIsTrivial other = False
+contIsTrivial (ApplyTo _ (Coercion _) _ cont) = contIsTrivial cont
+contIsTrivial (CoerceIt _ cont) = contIsTrivial cont
+contIsTrivial _ = False
-------------------
-contResultType :: SimplCont -> OutType
-contResultType (Stop to_ty _ _) = to_ty
-contResultType (StrictArg _ _ _ cont) = contResultType cont
-contResultType (StrictBind _ _ _ _ cont) = contResultType cont
-contResultType (ApplyTo _ _ _ cont) = contResultType cont
-contResultType (CoerceIt _ cont) = contResultType cont
-contResultType (Select _ _ _ _ cont) = contResultType cont
+contResultType :: SimplEnv -> OutType -> SimplCont -> OutType
+contResultType env ty cont
+ = go cont ty
+ where
+ subst_ty se ty = SimplEnv.substTy (se `setInScope` env) ty
+ subst_co se co = SimplEnv.substCo (se `setInScope` env) co
+
+ go (Stop {}) ty = ty
+ go (CoerceIt co cont) _ = go cont (pSnd (coercionKind co))
+ go (StrictBind _ bs body se cont) _ = go cont (subst_ty se (exprType (mkLams bs body)))
+ go (StrictArg ai _ cont) _ = go cont (funResultTy (argInfoResultTy ai))
+ go (Select _ _ alts se cont) _ = go cont (subst_ty se (coreAltsType alts))
+ go (ApplyTo _ arg se cont) ty = go cont (apply_to_arg ty arg se)
+
+ apply_to_arg ty (Type ty_arg) se = applyTy ty (subst_ty se ty_arg)
+ apply_to_arg ty (Coercion co_arg) se = applyCo ty (subst_co se co_arg)
+ apply_to_arg ty _ _ = funResultTy ty
+
+argInfoResultTy :: ArgInfo -> OutType
+argInfoResultTy (ArgInfo { ai_fun = fun, ai_args = args })
+ = foldr (\arg fn_ty -> applyTypeToArg fn_ty arg) (idType fun) args
-------------------
countValArgs :: SimplCont -> Int
-countValArgs (ApplyTo _ (Type ty) se cont) = countValArgs cont
-countValArgs (ApplyTo _ val_arg se cont) = 1 + countValArgs cont
-countValArgs other = 0
+countValArgs (ApplyTo _ (Type _) _ cont) = countValArgs cont
+countValArgs (ApplyTo _ (Coercion _) _ cont) = countValArgs cont
+countValArgs (ApplyTo _ _ _ cont) = 1 + countValArgs cont
+countValArgs _ = 0
countArgs :: SimplCont -> Int
-countArgs (ApplyTo _ arg se cont) = 1 + countArgs cont
-countArgs other = 0
+countArgs (ApplyTo _ _ _ cont) = 1 + countArgs cont
+countArgs _ = 0
-contArgs :: SimplCont -> ([OutExpr], SimplCont)
+contArgs :: SimplCont -> (Bool, [ArgSummary], SimplCont)
-- Uses substitution to turn each arg into an OutExpr
-contArgs cont = go [] cont
+contArgs cont@(ApplyTo {})
+ = case go [] cont of { (args, cont') -> (False, args, cont') }
where
- go args (ApplyTo _ arg se cont) = go (substExpr se arg : args) cont
- go args cont = (reverse args, cont)
+ go args (ApplyTo _ arg se cont)
+ | isTypeArg arg = go args cont
+ | otherwise = go (is_interesting arg se : args) cont
+ go args cont = (reverse args, cont)
+
+ is_interesting arg se = interestingArg (substExpr (text "contArgs") se arg)
+ -- Do *not* use short-cutting substitution here
+ -- because we want to get as much IdInfo as possible
+
+contArgs cont = (True, [], cont)
+
+pushSimplifiedArgs :: SimplEnv -> [CoreExpr] -> SimplCont -> SimplCont
+pushSimplifiedArgs _env [] cont = cont
+pushSimplifiedArgs env (arg:args) cont = ApplyTo Simplified arg env (pushSimplifiedArgs env args cont)
+ -- The env has an empty SubstEnv
dropArgs :: Int -> SimplCont -> SimplCont
dropArgs 0 cont = cont
\end{code}
-\begin{code}
-interestingArg :: OutExpr -> Bool
- -- An argument is interesting if it has *some* structure
- -- We are here trying to avoid unfolding a function that
- -- is applied only to variables that have no unfolding
- -- (i.e. they are probably lambda bound): f x y z
- -- There is little point in inlining f here.
-interestingArg (Var v) = hasSomeUnfolding (idUnfolding v)
- -- Was: isValueUnfolding (idUnfolding v')
- -- But that seems over-pessimistic
- || isDataConWorkId v
- -- This accounts for an argument like
- -- () or [], which is definitely interesting
-interestingArg (Type _) = False
-interestingArg (App fn (Type _)) = interestingArg fn
-interestingArg (Note _ a) = interestingArg a
-
--- Idea (from Sam B); I'm not sure if it's a good idea, so commented out for now
--- interestingArg expr | isUnLiftedType (exprType expr)
--- -- Unlifted args are only ever interesting if we know what they are
--- = case expr of
--- Lit lit -> True
--- _ -> False
-
-interestingArg other = True
- -- Consider let x = 3 in f x
- -- The substitution will contain (x -> ContEx 3), and we want to
- -- to say that x is an interesting argument.
- -- But consider also (\x. f x y) y
- -- The substitution will contain (x -> ContEx y), and we want to say
- -- that x is not interesting (assuming y has no unfolding)
-\end{code}
-
-
-Comment about interestingCallContext
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Note [Interesting call context]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We want to avoid inlining an expression where there can't possibly be
any gain, such as in an argument position. Hence, if the continuation
is interesting (eg. a case scrutinee, application etc.) then we
contIsInteresting looks for case expressions with just a single
default case.
+
\begin{code}
-interestingCallContext :: Bool -- False <=> no args at all
- -> Bool -- False <=> no value args
- -> SimplCont -> Bool
- -- The "lone-variable" case is important. I spent ages
- -- messing about with unsatisfactory varaints, but this is nice.
- -- The idea is that if a variable appear all alone
- -- as an arg of lazy fn, or rhs Stop
- -- as scrutinee of a case Select
- -- as arg of a strict fn ArgOf
- -- then we should not inline it (unless there is some other reason,
- -- e.g. is is the sole occurrence). We achieve this by making
- -- interestingCallContext return False for a lone variable.
- --
- -- Why? At least in the case-scrutinee situation, turning
- -- let x = (a,b) in case x of y -> ...
- -- into
- -- let x = (a,b) in case (a,b) of y -> ...
- -- and thence to
- -- let x = (a,b) in let y = (a,b) in ...
- -- is bad if the binding for x will remain.
- --
- -- Another example: I discovered that strings
- -- were getting inlined straight back into applications of 'error'
- -- because the latter is strict.
- -- s = "foo"
- -- f = \x -> ...(error s)...
-
- -- Fundamentally such contexts should not ecourage inlining because
- -- the context can ``see'' the unfolding of the variable (e.g. case or a RULE)
- -- so there's no gain.
- --
- -- However, even a type application or coercion isn't a lone variable.
- -- Consider
- -- case $fMonadST @ RealWorld of { :DMonad a b c -> c }
- -- We had better inline that sucker! The case won't see through it.
- --
- -- For now, I'm treating treating a variable applied to types
- -- in a *lazy* context "lone". The motivating example was
- -- f = /\a. \x. BIG
- -- g = /\a. \y. h (f a)
- -- There's no advantage in inlining f here, and perhaps
- -- a significant disadvantage. Hence some_val_args in the Stop case
-
-interestingCallContext some_args some_val_args cont
+interestingCallContext :: SimplCont -> CallCtxt
+-- See Note [Interesting call context]
+interestingCallContext cont
= interesting cont
where
- interesting (Select {}) = some_args
- interesting (ApplyTo {}) = True -- Can happen if we have (coerce t (f x)) y
- -- Perhaps True is a bit over-keen, but I've
- -- seen (coerce f) x, where f has an INLINE prag,
- -- So we have to give some motivaiton for inlining it
- interesting (StrictArg {}) = some_val_args
- interesting (StrictBind {}) = some_val_args -- ??
- interesting (Stop ty _ interesting) = some_val_args && interesting
- interesting (CoerceIt _ cont) = interesting cont
+ interesting (Select _ bndr _ _ _)
+ | isDeadBinder bndr = CaseCtxt
+ | otherwise = ArgCtxt False -- If the binder is used, this
+ -- is like a strict let
+ -- See Note [RHS of lets] in CoreUnfold
+
+ interesting (ApplyTo _ arg _ cont)
+ | isTypeArg arg = interesting cont
+ | otherwise = ValAppCtxt -- Can happen if we have (f Int |> co) y
+ -- If f has an INLINE prag we need to give it some
+ -- motivation to inline. See Note [Cast then apply]
+ -- in CoreUnfold
+
+ interesting (StrictArg _ cci _) = cci
+ interesting (StrictBind {}) = BoringCtxt
+ interesting (Stop cci) = cci
+ interesting (CoerceIt _ cont) = interesting cont
-- If this call is the arg of a strict function, the context
-- is a bit interesting. If we inline here, we may get useful
-- evaluation information to avoid repeated evals: e.g.
-------------------
mkArgInfo :: Id
+ -> [CoreRule] -- Rules for function
-> Int -- Number of value args
- -> SimplCont -- Context of the cal
- -> (Bool, [Bool]) -- Arg info
--- The arg info consists of
--- * A Bool indicating if the function has rules (recursively)
--- * A [Bool] indicating strictness for each arg
--- The [Bool] is usually infinite, but if it is finite it
--- guarantees that the function diverges after being given
--- that number of args
-
-mkArgInfo fun n_val_args call_cont
- = (interestingArgContext fun call_cont, fun_stricts)
+ -> SimplCont -- Context of the call
+ -> ArgInfo
+
+mkArgInfo fun rules n_val_args call_cont
+ | n_val_args < idArity fun -- Note [Unsaturated functions]
+ = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
+ , ai_encl = False
+ , ai_strs = vanilla_stricts
+ , ai_discs = vanilla_discounts }
+ | otherwise
+ = ArgInfo { ai_fun = fun, ai_args = [], ai_rules = rules
+ , ai_encl = interestingArgContext rules call_cont
+ , ai_strs = add_type_str (idType fun) arg_stricts
+ , ai_discs = arg_discounts }
where
- vanilla_stricts, fun_stricts :: [Bool]
+ vanilla_discounts, arg_discounts :: [Int]
+ vanilla_discounts = repeat 0
+ arg_discounts = case idUnfolding fun of
+ CoreUnfolding {uf_guidance = UnfIfGoodArgs {ug_args = discounts}}
+ -> discounts ++ vanilla_discounts
+ _ -> vanilla_discounts
+
+ vanilla_stricts, arg_stricts :: [Bool]
vanilla_stricts = repeat False
- fun_stricts
- = case splitStrictSig (idNewStrictness fun) of
+ arg_stricts
+ = case splitStrictSig (idStrictness fun) of
(demands, result_info)
| not (demands `lengthExceeds` n_val_args)
-> -- Enough args, use the strictness given.
map isStrictDmd demands -- Finite => result is bottom
else
map isStrictDmd demands ++ vanilla_stricts
+ | otherwise
+ -> WARN( True, text "More demands than arity" <+> ppr fun <+> ppr (idArity fun)
+ <+> ppr n_val_args <+> ppr demands )
+ vanilla_stricts -- Not enough args, or no strictness
+
+ add_type_str :: Type -> [Bool] -> [Bool]
+ -- If the function arg types are strict, record that in the 'strictness bits'
+ -- No need to instantiate because unboxed types (which dominate the strict
+ -- types) can't instantiate type variables.
+ -- add_type_str is done repeatedly (for each call); might be better
+ -- once-for-all in the function
+ -- But beware primops/datacons with no strictness
+ add_type_str _ [] = []
+ add_type_str fun_ty strs -- Look through foralls
+ | Just (_, fun_ty') <- splitForAllTy_maybe fun_ty -- Includes coercions
+ = add_type_str fun_ty' strs
+ add_type_str fun_ty (str:strs) -- Add strict-type info
+ | Just (arg_ty, fun_ty') <- splitFunTy_maybe fun_ty
+ = (str || isStrictType arg_ty) : add_type_str fun_ty' strs
+ add_type_str _ strs
+ = strs
+
+{- Note [Unsaturated functions]
+ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Consider (test eyeball/inline4)
+ x = a:as
+ y = f x
+where f has arity 2. Then we do not want to inline 'x', because
+it'll just be floated out again. Even if f has lots of discounts
+on its first argument -- it must be saturated for these to kick in
+-}
- other -> vanilla_stricts -- Not enough args, or no strictness
-
-interestingArgContext :: Id -> SimplCont -> Bool
+interestingArgContext :: [CoreRule] -> SimplCont -> Bool
-- If the argument has form (f x y), where x,y are boring,
-- and f is marked INLINE, then we don't want to inline f.
-- But if the context of the argument is
-- where g has rules, then we *do* want to inline f, in case it
-- exposes a rule that might fire. Similarly, if the context is
-- h (g (f x x))
--- where h has rules, then we do want to inline f.
--- The interesting_arg_ctxt flag makes this happen; if it's
+-- where h has rules, then we do want to inline f; hence the
+-- call_cont argument to interestingArgContext
+--
+-- The ai-rules flag makes this happen; if it's
-- set, the inliner gets just enough keener to inline f
-- regardless of how boring f's arguments are, if it's marked INLINE
--
-- The alternative would be to *always* inline an INLINE function,
-- regardless of how boring its context is; but that seems overkill
-- For example, it'd mean that wrapper functions were always inlined
-interestingArgContext fn cont
- = idHasRules fn || go cont
+interestingArgContext rules call_cont
+ = notNull rules || enclosing_fn_has_rules
where
- go (Select {}) = False
- go (ApplyTo {}) = False
- go (StrictArg {}) = True
- go (StrictBind {}) = False -- ??
- go (CoerceIt _ c) = go c
- go (Stop _ _ interesting) = interesting
+ enclosing_fn_has_rules = go call_cont
--------------------
-canUpdateInPlace :: Type -> Bool
--- Consider let x = <wurble> in ...
--- If <wurble> returns an explicit constructor, we might be able
--- to do update in place. So we treat even a thunk RHS context
--- as interesting if update in place is possible. We approximate
--- this by seeing if the type has a single constructor with a
--- small arity. But arity zero isn't good -- we share the single copy
--- for that case, so no point in sharing.
-
-canUpdateInPlace ty
- | not opt_UF_UpdateInPlace = False
- | otherwise
- = case splitTyConApp_maybe ty of
- Nothing -> False
- Just (tycon, _) -> case tyConDataCons_maybe tycon of
- Just [dc] -> arity == 1 || arity == 2
- where
- arity = dataConRepArity dc
- other -> False
-\end{code}
+ go (Select {}) = False
+ go (ApplyTo {}) = False
+ go (StrictArg _ cci _) = interesting cci
+ go (StrictBind {}) = False -- ??
+ go (CoerceIt _ c) = go c
+ go (Stop cci) = interesting cci
+ interesting (ArgCtxt rules) = rules
+ interesting _ = False
+\end{code}
%************************************************************************
%* *
-\subsection{Decisions about inlining}
+ SimplifierMode
%* *
%************************************************************************
-Inlining is controlled partly by the SimplifierMode switch. This has two
-settings:
-
- SimplGently (a) Simplifying before specialiser/full laziness
- (b) Simplifiying inside INLINE pragma
- (c) Simplifying the LHS of a rule
- (d) Simplifying a GHCi expression or Template
- Haskell splice
+The SimplifierMode controls several switches; see its definition in
+CoreMonad
+ sm_rules :: Bool -- Whether RULES are enabled
+ sm_inline :: Bool -- Whether inlining is enabled
+ sm_case_case :: Bool -- Whether case-of-case is enabled
+ sm_eta_expand :: Bool -- Whether eta-expansion is enabled
- SimplPhase n Used at all other times
+\begin{code}
+simplEnvForGHCi :: DynFlags -> SimplEnv
+simplEnvForGHCi dflags
+ = mkSimplEnv $ SimplMode { sm_names = ["GHCi"]
+ , sm_phase = InitialPhase
+ , sm_rules = rules_on
+ , sm_inline = False
+ , sm_eta_expand = eta_expand_on
+ , sm_case_case = True }
+ where
+ rules_on = dopt Opt_EnableRewriteRules dflags
+ eta_expand_on = dopt Opt_DoLambdaEtaExpansion dflags
+ -- Do not do any inlining, in case we expose some unboxed
+ -- tuple stuff that confuses the bytecode interpreter
+
+updModeForInlineRules :: Activation -> SimplifierMode -> SimplifierMode
+-- See Note [Simplifying inside InlineRules]
+updModeForInlineRules inline_rule_act current_mode
+ = current_mode { sm_phase = phaseFromActivation inline_rule_act
+ , sm_inline = True
+ , sm_eta_expand = False }
+ -- For sm_rules, just inherit; sm_rules might be "off"
+ -- becuase of -fno-enable-rewrite-rules
+ where
+ phaseFromActivation (ActiveAfter n) = Phase n
+ phaseFromActivation _ = InitialPhase
+\end{code}
-The key thing about SimplGently is that it does no call-site inlining.
-Before full laziness we must be careful not to inline wrappers,
-because doing so inhibits floating
+Note [Inlining in gentle mode]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Something is inlined if
+ (i) the sm_inline flag is on, AND
+ (ii) the thing has an INLINE pragma, AND
+ (iii) the thing is inlinable in the earliest phase.
+
+Example of why (iii) is important:
+ {-# INLINE [~1] g #-}
+ g = ...
+
+ {-# INLINE f #-}
+ f x = g (g x)
+
+If we were to inline g into f's inlining, then an importing module would
+never be able to do
+ f e --> g (g e) ---> RULE fires
+because the InlineRule for f has had g inlined into it.
+
+On the other hand, it is bad not to do ANY inlining into an
+InlineRule, because then recursive knots in instance declarations
+don't get unravelled.
+
+However, *sometimes* SimplGently must do no call-site inlining at all
+(hence sm_inline = False). Before full laziness we must be careful
+not to inline wrappers, because doing so inhibits floating
e.g. ...(case f x of ...)...
==> ...(case (case x of I# x# -> fw x#) of ...)...
==> ...(case x of I# x# -> case fw x# of ...)...
anything, because the byte-code interpreter might get confused about
unboxed tuples and suchlike.
-INLINE pragmas
-~~~~~~~~~~~~~~
-SimplGently is also used as the mode to simplify inside an InlineMe note.
-
-\begin{code}
-inlineMode :: SimplifierMode
-inlineMode = SimplGently
-\end{code}
-
-It really is important to switch off inlinings inside such
-expressions. Consider the following example
+Note [Simplifying inside InlineRules]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+We must take care with simplification inside InlineRules (which come from
+INLINE pragmas).
+First, consider the following example
let f = \pq -> BIG
in
let g = \y -> f y y
{-# INLINE g #-}
in ...g...g...g...g...g...
-
-Now, if that's the ONLY occurrence of f, it will be inlined inside g,
-and thence copied multiple times when g is inlined.
-
-
-This function may be inlinined in other modules, so we
-don't want to remove (by inlining) calls to functions that have
-specialisations, or that may have transformation rules in an importing
-scope.
-
-E.g. {-# INLINE f #-}
- f x = ...g...
-
-and suppose that g is strict *and* has specialisations. If we inline
-g's wrapper, we deny f the chance of getting the specialised version
-of g when f is inlined at some call site (perhaps in some other
-module).
-
+Now, if that's the ONLY occurrence of f, it might be inlined inside g,
+and thence copied multiple times when g is inlined. HENCE we treat
+any occurrence in an InlineRule as a multiple occurrence, not a single
+one; see OccurAnal.addRuleUsage.
+
+Second, we do want *do* to some modest rules/inlining stuff in InlineRules,
+partly to eliminate senseless crap, and partly to break the recursive knots
+generated by instance declarations.
+
+However, suppose we have
+ {-# INLINE <act> f #-}
+ f = <rhs>
+meaning "inline f in phases p where activation <act>(p) holds".
+Then what inlinings/rules can we apply to the copy of <rhs> captured in
+f's InlineRule? Our model is that literally <rhs> is substituted for
+f when it is inlined. So our conservative plan (implemented by
+updModeForInlineRules) is this:
+
+ -------------------------------------------------------------
+ When simplifying the RHS of an InlineRule, set the phase to the
+ phase in which the InlineRule first becomes active
+ -------------------------------------------------------------
+
+That ensures that
+
+ a) Rules/inlinings that *cease* being active before p will
+ not apply to the InlineRule rhs, consistent with it being
+ inlined in its *original* form in phase p.
+
+ b) Rules/inlinings that only become active *after* p will
+ not apply to the InlineRule rhs, again to be consistent with
+ inlining the *original* rhs in phase p.
+
+For example,
+ {-# INLINE f #-}
+ f x = ...g...
+
+ {-# NOINLINE [1] g #-}
+ g y = ...
+
+ {-# RULE h g = ... #-}
+Here we must not inline g into f's RHS, even when we get to phase 0,
+because when f is later inlined into some other module we want the
+rule for h to fire.
+
+Similarly, consider
+ {-# INLINE f #-}
+ f x = ...g...
+
+ g y = ...
+and suppose that there are auto-generated specialisations and a strictness
+wrapper for g. The specialisations get activation AlwaysActive, and the
+strictness wrapper get activation (ActiveAfter 0). So the strictness
+wrepper fails the test and won't be inlined into f's InlineRule. That
+means f can inline, expose the specialised call to g, so the specialisation
+rules can fire.
+
+A note about wrappers
+~~~~~~~~~~~~~~~~~~~~~
It's also important not to inline a worker back into a wrapper.
A wrapper looks like
wraper = inline_me (\x -> ...worker... )
the wrapper (initially, the worker's only call site!). But,
if the wrapper is sure to be called, the strictness analyser will
mark it 'demanded', so when the RHS is simplified, it'll get an ArgOf
-continuation. That's why the keep_inline predicate returns True for
-ArgOf continuations. It shouldn't do any harm not to dissolve the
-inline-me note under these circumstances.
+continuation.
+
+\begin{code}
+activeUnfolding :: SimplEnv -> Id -> Bool
+activeUnfolding env
+ | not (sm_inline mode) = active_unfolding_minimal
+ | otherwise = case sm_phase mode of
+ InitialPhase -> active_unfolding_gentle
+ Phase n -> active_unfolding n
+ where
+ mode = getMode env
+
+getUnfoldingInRuleMatch :: SimplEnv -> IdUnfoldingFun
+-- When matching in RULE, we want to "look through" an unfolding
+-- (to see a constructor) if *rules* are on, even if *inlinings*
+-- are not. A notable example is DFuns, which really we want to
+-- match in rules like (op dfun) in gentle mode. Another example
+-- is 'otherwise' which we want exprIsConApp_maybe to be able to
+-- see very early on
+getUnfoldingInRuleMatch env id
+ | unf_is_active = idUnfolding id
+ | otherwise = NoUnfolding
+ where
+ mode = getMode env
+ unf_is_active
+ | not (sm_rules mode) = active_unfolding_minimal id
+ | otherwise = isActive (sm_phase mode) (idInlineActivation id)
+
+active_unfolding_minimal :: Id -> Bool
+-- Compuslory unfoldings only
+-- Ignore SimplGently, because we want to inline regardless;
+-- the Id has no top-level binding at all
+--
+-- NB: we used to have a second exception, for data con wrappers.
+-- On the grounds that we use gentle mode for rule LHSs, and
+-- they match better when data con wrappers are inlined.
+-- But that only really applies to the trivial wrappers (like (:)),
+-- and they are now constructed as Compulsory unfoldings (in MkId)
+-- so they'll happen anyway.
+active_unfolding_minimal id = isCompulsoryUnfolding (realIdUnfolding id)
+
+active_unfolding :: PhaseNum -> Id -> Bool
+active_unfolding n id = isActiveIn n (idInlineActivation id)
+
+active_unfolding_gentle :: Id -> Bool
+-- Anything that is early-active
+-- See Note [Gentle mode]
+active_unfolding_gentle id
+ = isInlinePragma prag
+ && isEarlyActive (inlinePragmaActivation prag)
+ -- NB: wrappers are not early-active
+ where
+ prag = idInlinePragma id
-Note that the result is that we do very little simplification
-inside an InlineMe.
+----------------------
+activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
+-- Nothing => No rules at all
+activeRule _dflags env
+ | not (sm_rules mode) = Nothing -- Rewriting is off
+ | otherwise = Just (isActive (sm_phase mode))
+ where
+ mode = getMode env
+\end{code}
- all xs = foldr (&&) True xs
- any p = all . map p {-# INLINE any #-}
-Problem: any won't get deforested, and so if it's exported and the
-importer doesn't use the inlining, (eg passes it as an arg) then we
-won't get deforestation at all. We havn't solved this problem yet!
+%************************************************************************
+%* *
+ preInlineUnconditionally
+%* *
+%************************************************************************
preInlineUnconditionally
~~~~~~~~~~~~~~~~~~~~~~~~
might have a BIG rhs, which will now be dup'd at every occurrenc of x.
-Evne RHSs labelled InlineMe aren't caught here, because there might be
+Even RHSs labelled InlineMe aren't caught here, because there might be
no benefit from inlining at the call site.
[Sept 01] Don't unconditionally inline a top-level thing, because that
Conclusion: inline top level things gaily until Phase 0 (the last
phase), at which point don't.
+Note [pre/postInlineUnconditionally in gentle mode]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Even in gentle mode we want to do preInlineUnconditionally. The
+reason is that too little clean-up happens if you don't inline
+use-once things. Also a bit of inlining is *good* for full laziness;
+it can expose constant sub-expressions. Example in
+spectral/mandel/Mandel.hs, where the mandelset function gets a useful
+let-float if you inline windowToViewport
+
+However, as usual for Gentle mode, do not inline things that are
+inactive in the intial stages. See Note [Gentle mode].
+
+Note [InlineRule and preInlineUnconditionally]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Surprisingly, do not pre-inline-unconditionally Ids with INLINE pragmas!
+Example
+
+ {-# INLINE f #-}
+ f :: Eq a => a -> a
+ f x = ...
+
+ fInt :: Int -> Int
+ fInt = f Int dEqInt
+
+ ...fInt...fInt...fInt...
+
+Here f occurs just once, in the RHS of f1. But if we inline it there
+we'll lose the opportunity to inline at each of fInt's call sites.
+The INLINE pragma will only inline when the application is saturated
+for exactly this reason; and we don't want PreInlineUnconditionally
+to second-guess it. A live example is Trac #3736.
+ c.f. Note [InlineRule and postInlineUnconditionally]
+
+Note [Top-level botomming Ids]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Don't inline top-level Ids that are bottoming, even if they are used just
+once, because FloatOut has gone to some trouble to extract them out.
+Inlining them won't make the program run faster!
+
+Note [Do not inline CoVars unconditionally]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Coercion variables appear inside coercions, and have a separate
+substitution, so don't inline them via the IdSubst!
+
\begin{code}
preInlineUnconditionally :: SimplEnv -> TopLevelFlag -> InId -> InExpr -> Bool
preInlineUnconditionally env top_lvl bndr rhs
- | not active = False
- | opt_SimplNoPreInlining = False
+ | not active = False
+ | isStableUnfolding (idUnfolding bndr) = False -- Note [InlineRule and preInlineUnconditionally]
+ | isTopLevel top_lvl && isBottomingId bndr = False -- Note [Top-level bottoming Ids]
+ | opt_SimplNoPreInlining = False
+ | isCoVar bndr = False -- Note [Do not inline CoVars unconditionally]
| otherwise = case idOccInfo bndr of
IAmDead -> True -- Happens in ((\x.1) v)
OneOcc in_lam True int_cxt -> try_once in_lam int_cxt
- other -> False
+ _ -> False
where
- phase = getMode env
- active = case phase of
- SimplGently -> isAlwaysActive prag
- SimplPhase n -> isActive n prag
- prag = idInlinePragma bndr
-
+ mode = getMode env
+ active = isActive (sm_phase mode) act
+ -- See Note [pre/postInlineUnconditionally in gentle mode]
+ act = idInlineActivation bndr
try_once in_lam int_cxt -- There's one textual occurrence
| not in_lam = isNotTopLevel top_lvl || early_phase
| otherwise = int_cxt && canInlineInLam rhs
--- Be very careful before inlining inside a lambda, becuase (a) we must not
+-- Be very careful before inlining inside a lambda, because (a) we must not
-- invalidate occurrence information, and (b) we want to avoid pushing a
-- single allocation (here) into multiple allocations (inside lambda).
-- Inlining a *function* with a single *saturated* call would be ok, mind you.
-- canInlineInLam => free vars of rhs are (Once in_lam) or Many,
-- so substituting rhs inside a lambda doesn't change the occ info.
-- Sadly, not quite the same as exprIsHNF.
- canInlineInLam (Lit l) = True
+ canInlineInLam (Lit _) = True
canInlineInLam (Lam b e) = isRuntimeVar b || canInlineInLam e
canInlineInLam (Note _ e) = canInlineInLam e
canInlineInLam _ = False
- early_phase = case phase of
- SimplPhase 0 -> False
- other -> True
+ early_phase = case sm_phase mode of
+ Phase 0 -> False
+ _ -> True
-- If we don't have this early_phase test, consider
-- x = length [1,2,3]
-- The full laziness pass carefully floats all the cons cells to
\end{code}
+%************************************************************************
+%* *
+ postInlineUnconditionally
+%* *
+%************************************************************************
+
postInlineUnconditionally
~~~~~~~~~~~~~~~~~~~~~~~~~
@postInlineUnconditionally@ decides whether to unconditionally inline
trivial RHS. If so, we can inline and discard the binding altogether.
NB: a loop breaker has must_keep_binding = True and non-loop-breakers
-only have *forward* references Hence, it's safe to discard the binding
+only have *forward* references. Hence, it's safe to discard the binding
NOTE: This isn't our last opportunity to inline. We're at the binding
site right now, and we'll get another opportunity when we get to the
\begin{code}
postInlineUnconditionally
:: SimplEnv -> TopLevelFlag
- -> InId -- The binder (an OutId would be fine too)
+ -> OutId -- The binder (an InId would be fine too)
+ -- (*not* a CoVar)
-> OccInfo -- From the InId
-> OutExpr
-> Unfolding
-> Bool
postInlineUnconditionally env top_lvl bndr occ_info rhs unfolding
- | not active = False
- | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, dont' inline
+ | not active = False
+ | isLoopBreaker occ_info = False -- If it's a loop-breaker of any kind, don't inline
-- because it might be referred to "earlier"
- | isExportedId bndr = False
- | exprIsTrivial rhs = True
+ | isExportedId bndr = False
+ | isStableUnfolding unfolding = False -- Note [InlineRule and postInlineUnconditionally]
+ | isTopLevel top_lvl = False -- Note [Top level and postInlineUnconditionally]
+ | exprIsTrivial rhs = True
| otherwise
= case occ_info of
-- The point of examining occ_info here is that for *non-values*
-- case v of
-- True -> case x of ...
-- False -> case x of ...
- -- I'm not sure how important this is in practice
- OneOcc in_lam one_br int_cxt -- OneOcc => no code-duplication issue
+ -- This is very important in practice; e.g. wheel-seive1 doubles
+ -- in allocation if you miss this out
+ OneOcc in_lam _one_br int_cxt -- OneOcc => no code-duplication issue
-> smallEnoughToInline unfolding -- Small enough to dup
-- ToDo: consider discount on smallEnoughToInline if int_cxt is true
--
-- PRINCIPLE: when we've already simplified an expression once,
-- make sure that we only inline it if it's reasonably small.
- && ((isNotTopLevel top_lvl && not in_lam) ||
- -- But outside a lambda, we want to be reasonably aggressive
+ && (not in_lam ||
+ -- Outside a lambda, we want to be reasonably aggressive
-- about inlining into multiple branches of case
-- e.g. let x = <non-value>
-- in case y of { C1 -> ..x..; C2 -> ..x..; C3 -> ... }
-- Here x isn't mentioned in the RHS, so we don't want to
-- create the (dead) let-binding let x = (a,b) in ...
- other -> False
+ _ -> False
-- Here's an example that we don't handle well:
-- let f = if b then Left (\x.BIG) else Right (\y.BIG)
-- in \y. ....case f of {...} ....
-- Here f is used just once, and duplicating the case work is fine (exprIsCheap).
-- But
--- * We can't preInlineUnconditionally because that woud invalidate
--- the occ info for b.
--- * We can't postInlineUnconditionally because the RHS is big, and
--- that risks exponential behaviour
--- * We can't call-site inline, because the rhs is big
+-- - We can't preInlineUnconditionally because that woud invalidate
+-- the occ info for b.
+-- - We can't postInlineUnconditionally because the RHS is big, and
+-- that risks exponential behaviour
+-- - We can't call-site inline, because the rhs is big
-- Alas!
where
- active = case getMode env of
- SimplGently -> isAlwaysActive prag
- SimplPhase n -> isActive n prag
- prag = idInlinePragma bndr
-
-activeInline :: SimplEnv -> OutId -> Bool
-activeInline env id
- = case getMode env of
- SimplGently -> False
- -- No inlining at all when doing gentle stuff,
- -- except for local things that occur once
- -- The reason is that too little clean-up happens if you
- -- don't inline use-once things. Also a bit of inlining is *good* for
- -- full laziness; it can expose constant sub-expressions.
- -- Example in spectral/mandel/Mandel.hs, where the mandelset
- -- function gets a useful let-float if you inline windowToViewport
-
- -- NB: we used to have a second exception, for data con wrappers.
- -- On the grounds that we use gentle mode for rule LHSs, and
- -- they match better when data con wrappers are inlined.
- -- But that only really applies to the trivial wrappers (like (:)),
- -- and they are now constructed as Compulsory unfoldings (in MkId)
- -- so they'll happen anyway.
-
- SimplPhase n -> isActive n prag
- where
- prag = idInlinePragma id
-
-activeRule :: DynFlags -> SimplEnv -> Maybe (Activation -> Bool)
--- Nothing => No rules at all
-activeRule dflags env
- | not (dopt Opt_RewriteRules dflags)
- = Nothing -- Rewriting is off
- | otherwise
- = case getMode env of
- SimplGently -> Just isAlwaysActive
- -- Used to be Nothing (no rules in gentle mode)
- -- Main motivation for changing is that I wanted
- -- lift String ===> ...
- -- to work in Template Haskell when simplifying
- -- splices, so we get simpler code for literal strings
- SimplPhase n -> Just (isActive n)
+ active = isActive (sm_phase (getMode env)) (idInlineActivation bndr)
+ -- See Note [pre/postInlineUnconditionally in gentle mode]
\end{code}
+Note [Top level and postInlineUnconditionally]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+We don't do postInlineUnconditionally for top-level things (even for
+ones that are trivial):
+
+ * Doing so will inline top-level error expressions that have been
+ carefully floated out by FloatOut. More generally, it might
+ replace static allocation with dynamic.
+
+ * Even for trivial expressions there's a problem. Consider
+ {-# RULE "foo" forall (xs::[T]). reverse xs = ruggle xs #-}
+ blah xs = reverse xs
+ ruggle = sort
+ In one simplifier pass we might fire the rule, getting
+ blah xs = ruggle xs
+ but in *that* simplifier pass we must not do postInlineUnconditionally
+ on 'ruggle' because then we'll have an unbound occurrence of 'ruggle'
+
+ If the rhs is trivial it'll be inlined by callSiteInline, and then
+ the binding will be dead and discarded by the next use of OccurAnal
+
+ * There is less point, because the main goal is to get rid of local
+ bindings used in multiple case branches.
+
+
+Note [InlineRule and postInlineUnconditionally]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Do not do postInlineUnconditionally if the Id has an InlineRule, otherwise
+we lose the unfolding. Example
+
+ -- f has InlineRule with rhs (e |> co)
+ -- where 'e' is big
+ f = e |> co
+
+Then there's a danger we'll optimise to
+
+ f' = e
+ f = f' |> co
+
+and now postInlineUnconditionally, losing the InlineRule on f. Now f'
+won't inline because 'e' is too big.
+
+ c.f. Note [InlineRule and preInlineUnconditionally]
+
%************************************************************************
%* *
%************************************************************************
\begin{code}
-mkLam :: [OutBndr] -> OutExpr -> SimplM OutExpr
+mkLam :: SimplEnv -> [OutBndr] -> OutExpr -> SimplM OutExpr
-- mkLam tries three things
-- a) eta reduction, if that gives a trivial expression
-- b) eta expansion [only if there are some value lambdas]
-mkLam [] body
+mkLam _b [] body
= return body
-mkLam bndrs body
+mkLam _env bndrs body
= do { dflags <- getDOptsSmpl
; mkLam' dflags bndrs body }
where
mkLam' :: DynFlags -> [OutBndr] -> OutExpr -> SimplM OutExpr
- mkLam' dflags bndrs (Cast body@(Lam _ _) co)
+ mkLam' dflags bndrs (Cast body co)
+ | not (any bad bndrs)
-- Note [Casts and lambdas]
- = do { lam <- mkLam' dflags (bndrs ++ bndrs') body'
- ; return (mkCoerce (mkPiTypes bndrs co) lam) }
- where
- (bndrs',body') = collectBinders body
+ = do { lam <- mkLam' dflags bndrs body
+ ; return (mkCoerce (mkPiCos bndrs co) lam) }
+ where
+ co_vars = tyCoVarsOfCo co
+ bad bndr = isCoVar bndr && bndr `elemVarSet` co_vars
+
+ mkLam' dflags bndrs body@(Lam {})
+ = mkLam' dflags (bndrs ++ bndrs1) body1
+ where
+ (bndrs1, body1) = collectBinders body
mkLam' dflags bndrs body
- | dopt Opt_DoEtaReduction dflags,
- Just etad_lam <- tryEtaReduce bndrs body
+ | dopt Opt_DoEtaReduction dflags
+ , Just etad_lam <- tryEtaReduce bndrs body
= do { tick (EtaReduction (head bndrs))
; return etad_lam }
- | dopt Opt_DoLambdaEtaExpansion dflags,
- any isRuntimeVar bndrs
- = do { body' <- tryEtaExpansion dflags body
- ; return (mkLams bndrs body') }
-
- | otherwise
- = returnSmpl (mkLams bndrs body)
+ | otherwise
+ = return (mkLams bndrs body)
\end{code}
+
Note [Casts and lambdas]
~~~~~~~~~~~~~~~~~~~~~~~~
Consider
(\x. e `cast` g1) --> (\x.e) `cast` (tx -> g1)
where x:tx.
-In general, this floats casts outside lambdas, where (I hope) they might meet
-and cancel with some other cast.
-
-
--- c) floating lets out through big lambdas
--- [only if all tyvar lambdas, and only if this lambda
--- is the RHS of a let]
-
-{- Sept 01: I'm experimenting with getting the
- full laziness pass to float out past big lambdsa
- | all isTyVar bndrs, -- Only for big lambdas
- contIsRhs cont -- Only try the rhs type-lambda floating
- -- if this is indeed a right-hand side; otherwise
- -- we end up floating the thing out, only for float-in
- -- to float it right back in again!
- = tryRhsTyLam env bndrs body `thenSmpl` \ (floats, body') ->
- returnSmpl (floats, mkLams bndrs body')
--}
-
+In general, this floats casts outside lambdas, where (I hope) they
+might meet and cancel with some other cast:
+ \x. e `cast` co ===> (\x. e) `cast` (tx -> co)
+ /\a. e `cast` co ===> (/\a. e) `cast` (/\a. co)
+ /\g. e `cast` co ===> (/\g. e) `cast` (/\g. co)
+ (if not (g `in` co))
+
+Notice that it works regardless of 'e'. Originally it worked only
+if 'e' was itself a lambda, but in some cases that resulted in
+fruitless iteration in the simplifier. A good example was when
+compiling Text.ParserCombinators.ReadPrec, where we had a definition
+like (\x. Get `cast` g)
+where Get is a constructor with nonzero arity. Then mkLam eta-expanded
+the Get, and the next iteration eta-reduced it, and then eta-expanded
+it again.
+
+Note also the side condition for the case of coercion binders.
+It does not make sense to transform
+ /\g. e `cast` g ==> (/\g.e) `cast` (/\g.g)
+because the latter is not well-kinded.
%************************************************************************
%* *
-\subsection{Eta expansion and reduction}
+ Eta expansion
%* *
%************************************************************************
-We try for eta reduction here, but *only* if we get all the
-way to an exprIsTrivial expression.
-We don't want to remove extra lambdas unless we are going
-to avoid allocating this thing altogether
-
\begin{code}
-tryEtaReduce :: [OutBndr] -> OutExpr -> Maybe OutExpr
-tryEtaReduce bndrs body
- -- We don't use CoreUtils.etaReduce, because we can be more
- -- efficient here:
- -- (a) we already have the binders
- -- (b) we can do the triviality test before computing the free vars
- = go (reverse bndrs) body
+tryEtaExpand :: SimplEnv -> OutId -> OutExpr -> SimplM (Arity, OutExpr)
+-- See Note [Eta-expanding at let bindings]
+tryEtaExpand env bndr rhs
+ = do { dflags <- getDOptsSmpl
+ ; (new_arity, new_rhs) <- try_expand dflags
+
+ ; WARN( new_arity < old_arity || new_arity < _dmd_arity,
+ (ptext (sLit "Arity decrease:") <+> (ppr bndr <+> ppr old_arity
+ <+> ppr new_arity <+> ppr _dmd_arity) $$ ppr new_rhs) )
+ -- Note [Arity decrease]
+ return (new_arity, new_rhs) }
where
- go (b : bs) (App fun arg) | ok_arg b arg = go bs fun -- Loop round
- go [] fun | ok_fun fun = Just fun -- Success!
- go _ _ = Nothing -- Failure!
-
- ok_fun fun = exprIsTrivial fun
- && not (any (`elemVarSet` (exprFreeVars fun)) bndrs)
- && (exprIsHNF fun || all ok_lam bndrs)
- ok_lam v = isTyVar v || isDictId v
- -- The exprIsHNF is because eta reduction is not
- -- valid in general: \x. bot /= bot
- -- So we need to be sure that the "fun" is a value.
+ try_expand dflags
+ | sm_eta_expand (getMode env) -- Provided eta-expansion is on
+ , not (exprIsTrivial rhs)
+ , let dicts_cheap = dopt Opt_DictsCheap dflags
+ new_arity = findArity dicts_cheap bndr rhs old_arity
+ , new_arity > rhs_arity
+ = do { tick (EtaExpansion bndr)
+ ; return (new_arity, etaExpand new_arity rhs) }
+ | otherwise
+ = return (rhs_arity, rhs)
+
+ rhs_arity = exprArity rhs
+ old_arity = idArity bndr
+ _dmd_arity = length $ fst $ splitStrictSig $ idStrictness bndr
+
+findArity :: Bool -> Id -> CoreExpr -> Arity -> Arity
+-- This implements the fixpoint loop for arity analysis
+-- See Note [Arity analysis]
+findArity dicts_cheap bndr rhs old_arity
+ = go (exprEtaExpandArity (mk_cheap_fn dicts_cheap init_cheap_app) rhs)
+ -- We always call exprEtaExpandArity once, but usually
+ -- that produces a result equal to old_arity, and then
+ -- we stop right away (since arities should not decrease)
+ -- Result: the common case is that there is just one iteration
+ where
+ go :: Arity -> Arity
+ go cur_arity
+ | cur_arity <= old_arity = cur_arity
+ | new_arity == cur_arity = cur_arity
+ | otherwise = ASSERT( new_arity < cur_arity )
+ pprTrace "Exciting arity"
+ (vcat [ ppr bndr <+> ppr cur_arity <+> ppr new_arity
+ , ppr rhs])
+ go new_arity
+ where
+ new_arity = exprEtaExpandArity (mk_cheap_fn dicts_cheap cheap_app) rhs
+
+ cheap_app :: CheapAppFun
+ cheap_app fn n_val_args
+ | fn == bndr = n_val_args < cur_arity
+ | otherwise = isCheapApp fn n_val_args
+
+ init_cheap_app :: CheapAppFun
+ init_cheap_app fn n_val_args
+ | fn == bndr = True
+ | otherwise = isCheapApp fn n_val_args
+
+mk_cheap_fn :: Bool -> CheapAppFun -> CheapFun
+mk_cheap_fn dicts_cheap cheap_app
+ | not dicts_cheap
+ = \e _ -> exprIsCheap' cheap_app e
+ | otherwise
+ = \e mb_ty -> exprIsCheap' cheap_app e
+ || case mb_ty of
+ Nothing -> False
+ Just ty -> isDictLikeTy ty
+ -- If the experimental -fdicts-cheap flag is on, we eta-expand through
+ -- dictionary bindings. This improves arities. Thereby, it also
+ -- means that full laziness is less prone to floating out the
+ -- application of a function to its dictionary arguments, which
+ -- can thereby lose opportunities for fusion. Example:
+ -- foo :: Ord a => a -> ...
+ -- foo = /\a \(d:Ord a). let d' = ...d... in \(x:a). ....
+ -- -- So foo has arity 1
+ --
+ -- f = \x. foo dInt $ bar x
+ --
+ -- The (foo DInt) is floated out, and makes ineffective a RULE
+ -- foo (bar x) = ...
--
- -- However, we always want to reduce (/\a -> f a) to f
- -- This came up in a RULE: foldr (build (/\a -> g a))
- -- did not match foldr (build (/\b -> ...something complex...))
- -- The type checker can insert these eta-expanded versions,
- -- with both type and dictionary lambdas; hence the slightly
- -- ad-hoc isDictTy
-
- ok_arg b arg = varToCoreExpr b `cheapEqExpr` arg
+ -- One could go further and make exprIsCheap reply True to any
+ -- dictionary-typed expression, but that's more work.
+ --
+ -- See Note [Dictionary-like types] in TcType.lhs for why we use
+ -- isDictLikeTy here rather than isDictTy
\end{code}
-
- Try eta expansion for RHSs
-
-We go for:
- f = \x1..xn -> N ==> f = \x1..xn y1..ym -> N y1..ym
- (n >= 0)
-
-where (in both cases)
-
- * The xi can include type variables
-
- * The yi are all value variables
-
- * N is a NORMAL FORM (i.e. no redexes anywhere)
- wanting a suitable number of extra args.
-
-We may have to sandwich some coerces between the lambdas
-to make the types work. exprEtaExpandArity looks through coerces
-when computing arity; and etaExpand adds the coerces as necessary when
-actually computing the expansion.
-
-\begin{code}
-tryEtaExpansion :: DynFlags -> OutExpr -> SimplM OutExpr
--- There is at least one runtime binder in the binders
-tryEtaExpansion dflags body
- = getUniquesSmpl `thenSmpl` \ us ->
- returnSmpl (etaExpand fun_arity us body (exprType body))
- where
- fun_arity = exprEtaExpandArity dflags body
-\end{code}
+Note [Eta-expanding at let bindings]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+We now eta expand at let-bindings, which is where the payoff
+comes.
+
+One useful consequence is this example:
+ genMap :: C a => ...
+ {-# INLINE genMap #-}
+ genMap f xs = ...
+
+ myMap :: D a => ...
+ {-# INLINE myMap #-}
+ myMap = genMap
+
+Notice that 'genMap' should only inline if applied to two arguments.
+In the InlineRule for myMap we'll have the unfolding
+ (\d -> genMap Int (..d..))
+We do not want to eta-expand to
+ (\d f xs -> genMap Int (..d..) f xs)
+because then 'genMap' will inline, and it really shouldn't: at least
+as far as the programmer is concerned, it's not applied to two
+arguments!
+
+Note [Arity analysis]
+~~~~~~~~~~~~~~~~~~~~~
+The motivating example for arity analysis is this:
+
+ f = \x. let g = f (x+1)
+ in \y. ...g...
+
+What arity does f have? Really it should have arity 2, but a naive
+look at the RHS won't see that. You need a fixpoint analysis which
+says it has arity "infinity" the first time round.
+
+This example happens a lot; it first showed up in Andy Gill's thesis,
+fifteen years ago! It also shows up in the code for 'rnf' on lists
+in Trac #4138.
+
+The analysis is easy to achieve because exprEtaExpandArity takes an
+argument
+ type CheapFun = CoreExpr -> Maybe Type -> Bool
+used to decide if an expression is cheap enough to push inside a
+lambda. And exprIsCheap' in turn takes an argument
+ type CheapAppFun = Id -> Int -> Bool
+which tells when an application is cheap. This makes it easy to
+write the analysis loop.
+
+The analysis is cheap-and-cheerful because it doesn't deal with
+mutual recursion. But the self-recursive case is the important one.
%************************************************************************
We'd like to float this to
y1 = /\a. e1
y2 = /\a. e2
- x = /\a. C (y1 a) (y2 a)
+ x = /\a. C (y1 a) (y2 a)
for the usual reasons: we want to inline x rather vigorously.
You may think that this kind of thing is rare. But in some programs it is
abstractFloats :: [OutTyVar] -> SimplEnv -> OutExpr -> SimplM ([OutBind], OutExpr)
abstractFloats main_tvs body_env body
= ASSERT( notNull body_floats )
- do { (subst, float_binds) <- mapAccumLSmpl abstract empty_subst body_floats
- ; return (float_binds, CoreSubst.substExpr subst body) }
+ do { (subst, float_binds) <- mapAccumLM abstract empty_subst body_floats
+ ; return (float_binds, CoreSubst.substExpr (text "abstract_floats1") subst body) }
where
main_tv_set = mkVarSet main_tvs
body_floats = getFloats body_env
subst' = CoreSubst.extendIdSubst subst id poly_app
; return (subst', (NonRec poly_id poly_rhs)) }
where
- rhs' = CoreSubst.substExpr subst rhs
+ rhs' = CoreSubst.substExpr (text "abstract_floats2") subst rhs
tvs_here = varSetElems (main_tv_set `intersectVarSet` exprSomeFreeVars isTyVar rhs')
+
-- Abstract only over the type variables free in the rhs
-- wrt which the new binding is abstracted. But the naive
-- approach of abstract wrt the tyvars free in the Id's type
-- gives rise to problems. SLPJ June 98
abstract subst (Rec prs)
- = do { (poly_ids, poly_apps) <- mapAndUnzipSmpl (mk_poly tvs_here) ids
+ = do { (poly_ids, poly_apps) <- mapAndUnzipM (mk_poly tvs_here) ids
; let subst' = CoreSubst.extendSubstList subst (ids `zip` poly_apps)
- poly_rhss = [mkLams tvs_here (CoreSubst.substExpr subst' rhs) | rhs <- rhss]
+ poly_rhss = [mkLams tvs_here (CoreSubst.substExpr (text "abstract_floats3") subst' rhs)
+ | rhs <- rhss]
; return (subst', Rec (poly_ids `zip` poly_rhss)) }
where
(ids,rhss) = unzip prs
tvs_here = main_tvs
mk_poly tvs_here var
- = do { uniq <- getUniqueSmpl
+ = do { uniq <- getUniqueM
; let poly_name = setNameUnique (idName var) uniq -- Keep same name
poly_ty = mkForAllTys tvs_here (idType var) -- But new type of course
- poly_id = mkLocalId poly_name poly_ty
+ poly_id = transferPolyIdInfo var tvs_here $ -- Note [transferPolyIdInfo] in Id.lhs
+ mkLocalId poly_name poly_ty
; return (poly_id, mkTyApps (Var poly_id) (mkTyVarTys tvs_here)) }
-- In the olden days, it was crucial to copy the occInfo of the original var,
-- because we were looking at occurrence-analysed but as yet unsimplified code!
-- pinned on x.
\end{code}
+Note [Abstract over coercions]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+If a coercion variable (g :: a ~ Int) is free in the RHS, then so is the
+type variable a. Rather than sort this mess out, we simply bale out and abstract
+wrt all the type variables if any of them are coercion variables.
+
+
Historical note: if you use let-bindings instead of a substitution, beware of this:
-- Suppose we start with:
prepareAlts tries these things:
-1. If several alternatives are identical, merge them into
- a single DEFAULT alternative. I've occasionally seen this
- making a big difference:
-
- case e of =====> case e of
- C _ -> f x D v -> ....v....
- D v -> ....v.... DEFAULT -> f x
- DEFAULT -> f x
+1. Eliminate alternatives that cannot match, including the
+ DEFAULT alternative.
- The point is that we merge common RHSs, at least for the DEFAULT case.
- [One could do something more elaborate but I've never seen it needed.]
- To avoid an expensive test, we just merge branches equal to the *first*
- alternative; this picks up the common cases
- a) all branches equal
- b) some branches equal to the DEFAULT (which occurs first)
-
-2. Case merging:
- case e of b { ==> case e of b {
- p1 -> rhs1 p1 -> rhs1
- ... ...
- pm -> rhsm pm -> rhsm
- _ -> case b of b' { pn -> let b'=b in rhsn
- pn -> rhsn ...
- ... po -> let b'=b in rhso
- po -> rhso _ -> let b'=b in rhsd
- _ -> rhsd
- }
-
- which merges two cases in one case when -- the default alternative of
- the outer case scrutises the same variable as the outer case This
- transformation is called Case Merging. It avoids that the same
- variable is scrutinised multiple times.
+2. If the DEFAULT alternative can match only one possible constructor,
+ then make that constructor explicit.
+ e.g.
+ case e of x { DEFAULT -> rhs }
+ ===>
+ case e of x { (a,b) -> rhs }
+ where the type is a single constructor type. This gives better code
+ when rhs also scrutinises x or e.
+3. Returns a list of the constructors that cannot holds in the
+ DEFAULT alternative (if there is one)
-The case where transformation (1) showed up was like this (lib/std/PrelCError.lhs):
+Here "cannot match" includes knowledge from GADTs
- x | p `is` 1 -> e1
- | p `is` 2 -> e2
- ...etc...
+It's a good idea do do this stuff before simplifying the alternatives, to
+avoid simplifying alternatives we know can't happen, and to come up with
+the list of constructors that are handled, to put into the IdInfo of the
+case binder, for use when simplifying the alternatives.
-where @is@ was something like
-
- p `is` n = p /= (-1) && p == n
+Eliminating the default alternative in (1) isn't so obvious, but it can
+happen:
-This gave rise to a horrible sequence of cases
+data Colour = Red | Green | Blue
- case p of
- (-1) -> $j p
- 1 -> e1
- DEFAULT -> $j p
+f x = case x of
+ Red -> ..
+ Green -> ..
+ DEFAULT -> h x
-and similarly in cascade for all the join points!
+h y = case y of
+ Blue -> ..
+ DEFAULT -> [ case y of ... ]
-Note [Dead binders]
-~~~~~~~~~~~~~~~~~~~~
-We do this *here*, looking at un-simplified alternatives, because we
-have to check that r doesn't mention the variables bound by the
-pattern in each alternative, so the binder-info is rather useful.
+If we inline h into f, the default case of the inlined h can't happen.
+If we don't notice this, we may end up filtering out *all* the cases
+of the inner case y, which give us nowhere to go!
\begin{code}
prepareAlts :: OutExpr -> OutId -> [InAlt] -> SimplM ([AltCon], [InAlt])
prepareAlts scrut case_bndr' alts
- = do { dflags <- getDOptsSmpl
- ; alts <- combineIdenticalAlts case_bndr' alts
-
- ; let (alts_wo_default, maybe_deflt) = findDefault alts
+ = do { let (alts_wo_default, maybe_deflt) = findDefault alts
alt_cons = [con | (con,_,_) <- alts_wo_default]
imposs_deflt_cons = nub (imposs_cons ++ alt_cons)
-- "imposs_deflt_cons" are handled
-- EITHER by the context,
-- OR by a non-DEFAULT branch in this case expression.
- ; default_alts <- prepareDefault dflags scrut case_bndr' mb_tc_app
+ ; default_alts <- prepareDefault case_bndr' mb_tc_app
imposs_deflt_cons maybe_deflt
; let trimmed_alts = filterOut impossible_alt alts_wo_default
- merged_alts = mergeAlts trimmed_alts default_alts
+ merged_alts = mergeAlts trimmed_alts default_alts
-- We need the mergeAlts in case the new default_alt
-- has turned into a constructor alternative.
-- The merge keeps the inner DEFAULT at the front, if there is one
imposs_cons = case scrut of
Var v -> otherCons (idUnfolding v)
- other -> []
+ _ -> []
impossible_alt :: CoreAlt -> Bool
impossible_alt (con, _, _) | con `elem` imposs_cons = True
impossible_alt (DataAlt con, _, _) = dataConCannotMatch inst_tys con
- impossible_alt alt = False
-
-
---------------------------------------------------
--- 1. Merge identical branches
---------------------------------------------------
-combineIdenticalAlts :: OutId -> [InAlt] -> SimplM [InAlt]
-
-combineIdenticalAlts case_bndr alts@((con1,bndrs1,rhs1) : con_alts)
- | all isDeadBinder bndrs1, -- Remember the default
- length filtered_alts < length con_alts -- alternative comes first
- -- Also Note [Dead binders]
- = do { tick (AltMerge case_bndr)
- ; return ((DEFAULT, [], rhs1) : filtered_alts) }
- where
- filtered_alts = filter keep con_alts
- keep (con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
+ impossible_alt _ = False
-combineIdenticalAlts case_bndr alts = return alts
--------------------------------------------------------------------------
--- Prepare the default alternative
--------------------------------------------------------------------------
-prepareDefault :: DynFlags
- -> OutExpr -- Scrutinee
- -> OutId -- Case binder; need just for its type. Note that as an
+prepareDefault :: OutId -- Case binder; need just for its type. Note that as an
-- OutId, it has maximum information; this is important.
-- Test simpl013 is an example
-> Maybe (TyCon, [Type]) -- Type of scrutinee, decomposed
-> Maybe InExpr -- Rhs
-> SimplM [InAlt] -- Still unsimplified
-- We use a list because it's what mergeAlts expects,
- -- And becuase case-merging can cause many to show up
-
-------- Merge nested cases ----------
-prepareDefault dflags scrut outer_bndr bndr_ty imposs_cons (Just deflt_rhs)
- | dopt Opt_CaseMerge dflags
- , Case (Var scrut_var) inner_bndr _ inner_alts <- deflt_rhs
- , scruting_same_var scrut_var
- = do { tick (CaseMerge outer_bndr)
-
- ; let munge_rhs rhs = bindCaseBndr inner_bndr (Var outer_bndr) rhs
- ; return [(con, args, munge_rhs rhs) | (con, args, rhs) <- inner_alts,
- not (con `elem` imposs_cons) ]
- -- NB: filter out any imposs_cons. Example:
- -- case x of
- -- A -> e1
- -- DEFAULT -> case x of
- -- A -> e2
- -- B -> e3
- -- When we merge, we must ensure that e1 takes
- -- precedence over e2 as the value for A!
- }
- -- Warning: don't call prepareAlts recursively!
- -- Firstly, there's no point, because inner alts have already had
- -- mkCase applied to them, so they won't have a case in their default
- -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
- -- in munge_rhs may put a case into the DEFAULT branch!
- where
- -- We are scrutinising the same variable if it's
- -- the outer case-binder, or if the outer case scrutinises a variable
- -- (and it's the same). Testing both allows us not to replace the
- -- outer scrut-var with the outer case-binder (Simplify.simplCaseBinder).
- scruting_same_var = case scrut of
- Var outer_scrut -> \ v -> v == outer_bndr || v == outer_scrut
- other -> \ v -> v == outer_bndr
--------- Fill in known constructor -----------
-prepareDefault dflags scrut case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
+prepareDefault case_bndr (Just (tycon, inst_tys)) imposs_cons (Just deflt_rhs)
| -- This branch handles the case where we are
-- scrutinisng an algebraic data type
isAlgTyCon tycon -- It's a data type, tuple, or unboxed tuples.
-- case x of { DEFAULT -> e }
-- and we don't want to fill in a default for them!
, Just all_cons <- tyConDataCons_maybe tycon
- , not (null all_cons) -- This is a tricky corner case. If the data type has no constructors,
- -- which GHC allows, then the case expression will have at most a default
- -- alternative. We don't want to eliminate that alternative, because the
- -- invariant is that there's always one alternative. It's more convenient
- -- to leave
- -- case x of { DEFAULT -> e }
- -- as it is, rather than transform it to
- -- error "case cant match"
- -- which would be quite legitmate. But it's a really obscure corner, and
- -- not worth wasting code on.
+ , not (null all_cons)
+ -- This is a tricky corner case. If the data type has no constructors,
+ -- which GHC allows, then the case expression will have at most a default
+ -- alternative. We don't want to eliminate that alternative, because the
+ -- invariant is that there's always one alternative. It's more convenient
+ -- to leave
+ -- case x of { DEFAULT -> e }
+ -- as it is, rather than transform it to
+ -- error "case cant match"
+ -- which would be quite legitmate. But it's a really obscure corner, and
+ -- not worth wasting code on.
, let imposs_data_cons = [con | DataAlt con <- imposs_cons] -- We now know it's a data type
- impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
+ impossible con = con `elem` imposs_data_cons || dataConCannotMatch inst_tys con
= case filterOut impossible all_cons of
[] -> return [] -- Eliminate the default alternative
-- altogether if it can't match
[con] -> -- It matches exactly one constructor, so fill it in
do { tick (FillInCaseDefault case_bndr)
- ; us <- getUniquesSmpl
- ; let (ex_tvs, co_tvs, arg_ids) =
- dataConRepInstPat us con inst_tys
- ; return [(DataAlt con, ex_tvs ++ co_tvs ++ arg_ids, deflt_rhs)] }
+ ; us <- getUniquesM
+ ; let (ex_tvs, arg_ids) = dataConRepInstPat us con inst_tys
+ ; return [(DataAlt con, ex_tvs ++ arg_ids, deflt_rhs)] }
- two_or_more -> return [(DEFAULT, [], deflt_rhs)]
+ _ -> return [(DEFAULT, [], deflt_rhs)]
+
+ | debugIsOn, isAlgTyCon tycon
+ , null (tyConDataCons tycon)
+ , not (isFamilyTyCon tycon || isAbstractTyCon tycon)
+ -- Check for no data constructors
+ -- This can legitimately happen for abstract types and type families,
+ -- so don't report that
+ = pprTrace "prepareDefault" (ppr case_bndr <+> ppr tycon)
+ $ return [(DEFAULT, [], deflt_rhs)]
--------- Catch-all cases -----------
-prepareDefault dflags scrut case_bndr bndr_ty imposs_cons (Just deflt_rhs)
+prepareDefault _case_bndr _bndr_ty _imposs_cons (Just deflt_rhs)
= return [(DEFAULT, [], deflt_rhs)]
-prepareDefault dflags scrut case_bndr bndr_ty imposs_cons Nothing
+prepareDefault _case_bndr _bndr_ty _imposs_cons Nothing
= return [] -- No default branch
\end{code}
-=================================================================================
+%************************************************************************
+%* *
+ mkCase
+%* *
+%************************************************************************
mkCase tries these things
-1. Eliminate the case altogether if possible
+1. Merge Nested Cases
+
+ case e of b { ==> case e of b {
+ p1 -> rhs1 p1 -> rhs1
+ ... ...
+ pm -> rhsm pm -> rhsm
+ _ -> case b of b' { pn -> let b'=b in rhsn
+ pn -> rhsn ...
+ ... po -> let b'=b in rhso
+ po -> rhso _ -> let b'=b in rhsd
+ _ -> rhsd
+ }
+
+ which merges two cases in one case when -- the default alternative of
+ the outer case scrutises the same variable as the outer case. This
+ transformation is called Case Merging. It avoids that the same
+ variable is scrutinised multiple times.
-2. Case-identity:
+2. Eliminate Identity Case
case e of ===> e
True -> True;
and similar friends.
+3. Merge identical alternatives.
+ If several alternatives are identical, merge them into
+ a single DEFAULT alternative. I've occasionally seen this
+ making a big difference:
+
+ case e of =====> case e of
+ C _ -> f x D v -> ....v....
+ D v -> ....v.... DEFAULT -> f x
+ DEFAULT -> f x
+
+ The point is that we merge common RHSs, at least for the DEFAULT case.
+ [One could do something more elaborate but I've never seen it needed.]
+ To avoid an expensive test, we just merge branches equal to the *first*
+ alternative; this picks up the common cases
+ a) all branches equal
+ b) some branches equal to the DEFAULT (which occurs first)
+
+The case where Merge Identical Alternatives transformation showed up
+was like this (base/Foreign/C/Err/Error.lhs):
+
+ x | p `is` 1 -> e1
+ | p `is` 2 -> e2
+ ...etc...
+
+where @is@ was something like
+
+ p `is` n = p /= (-1) && p == n
+
+This gave rise to a horrible sequence of cases
+
+ case p of
+ (-1) -> $j p
+ 1 -> e1
+ DEFAULT -> $j p
+
+and similarly in cascade for all the join points!
+
\begin{code}
-mkCase :: OutExpr -> OutId -> OutType
- -> [OutAlt] -- Increasing order
- -> SimplM OutExpr
+mkCase, mkCase1, mkCase2
+ :: DynFlags
+ -> OutExpr -> OutId
+ -> [OutAlt] -- Alternatives in standard (increasing) order
+ -> SimplM OutExpr
--------------------------------------------------
--- 1. Check for empty alternatives
+-- 1. Merge Nested Cases
--------------------------------------------------
--- This isn't strictly an error. It's possible that the simplifer might "see"
--- that an inner case has no accessible alternatives before it "sees" that the
--- entire branch of an outer case is inaccessible. So we simply
--- put an error case here insteadd
-mkCase scrut case_bndr ty []
- = pprTrace "mkCase: null alts" (ppr case_bndr <+> ppr scrut) $
- return (mkApps (Var rUNTIME_ERROR_ID)
- [Type ty, Lit (mkStringLit "Impossible alternative")])
+mkCase dflags scrut outer_bndr ((DEFAULT, _, deflt_rhs) : outer_alts)
+ | dopt Opt_CaseMerge dflags
+ , Case (Var inner_scrut_var) inner_bndr _ inner_alts <- deflt_rhs
+ , inner_scrut_var == outer_bndr
+ = do { tick (CaseMerge outer_bndr)
+
+ ; let wrap_alt (con, args, rhs) = ASSERT( outer_bndr `notElem` args )
+ (con, args, wrap_rhs rhs)
+ -- Simplifier's no-shadowing invariant should ensure
+ -- that outer_bndr is not shadowed by the inner patterns
+ wrap_rhs rhs = Let (NonRec inner_bndr (Var outer_bndr)) rhs
+ -- The let is OK even for unboxed binders,
+
+ wrapped_alts | isDeadBinder inner_bndr = inner_alts
+ | otherwise = map wrap_alt inner_alts
+
+ merged_alts = mergeAlts outer_alts wrapped_alts
+ -- NB: mergeAlts gives priority to the left
+ -- case x of
+ -- A -> e1
+ -- DEFAULT -> case x of
+ -- A -> e2
+ -- B -> e3
+ -- When we merge, we must ensure that e1 takes
+ -- precedence over e2 as the value for A!
+
+ ; mkCase1 dflags scrut outer_bndr merged_alts
+ }
+ -- Warning: don't call mkCase recursively!
+ -- Firstly, there's no point, because inner alts have already had
+ -- mkCase applied to them, so they won't have a case in their default
+ -- Secondly, if you do, you get an infinite loop, because the bindCaseBndr
+ -- in munge_rhs may put a case into the DEFAULT branch!
+mkCase dflags scrut bndr alts = mkCase1 dflags scrut bndr alts
--------------------------------------------------
--- 2. Identity case
+-- 2. Eliminate Identity Case
--------------------------------------------------
-mkCase scrut case_bndr ty alts -- Identity case
+mkCase1 _dflags scrut case_bndr alts -- Identity case
| all identity_alt alts
- = tick (CaseIdentity case_bndr) `thenSmpl_`
- returnSmpl (re_cast scrut)
+ = do { tick (CaseIdentity case_bndr)
+ ; return (re_cast scrut) }
where
identity_alt (con, args, rhs) = check_eq con args (de_cast rhs)
check_eq (LitAlt lit') _ (Lit lit) = lit == lit'
check_eq (DataAlt con) args rhs = rhs `cheapEqExpr` mkConApp con (arg_tys ++ varsToCoreExprs args)
|| rhs `cheapEqExpr` Var case_bndr
- check_eq con args rhs = False
+ check_eq _ _ _ = False
arg_tys = map Type (tyConAppArgs (idType case_bndr))
re_cast scrut = case head alts of
(_,_,Cast _ co) -> Cast scrut co
- other -> scrut
+ _ -> scrut
+--------------------------------------------------
+-- 3. Merge Identical Alternatives
+--------------------------------------------------
+mkCase1 dflags scrut case_bndr ((_con1,bndrs1,rhs1) : con_alts)
+ | all isDeadBinder bndrs1 -- Remember the default
+ , length filtered_alts < length con_alts -- alternative comes first
+ -- Also Note [Dead binders]
+ = do { tick (AltMerge case_bndr)
+ ; mkCase2 dflags scrut case_bndr alts' }
+ where
+ alts' = (DEFAULT, [], rhs1) : filtered_alts
+ filtered_alts = filter keep con_alts
+ keep (_con,bndrs,rhs) = not (all isDeadBinder bndrs && rhs `cheapEqExpr` rhs1)
+mkCase1 dflags scrut bndr alts = mkCase2 dflags scrut bndr alts
--------------------------------------------------
-- Catch-all
--------------------------------------------------
-mkCase scrut bndr ty alts = returnSmpl (Case scrut bndr ty alts)
+mkCase2 _dflags scrut bndr alts
+ = return (Case scrut bndr (coreAltsType alts) alts)
\end{code}
-
-When adding auxiliary bindings for the case binder, it's worth checking if
-its dead, because it often is, and occasionally these mkCase transformations
-cascade rather nicely.
-
-\begin{code}
-bindCaseBndr bndr rhs body
- | isDeadBinder bndr = body
- | otherwise = bindNonRec bndr rhs body
-\end{code}
+Note [Dead binders]
+~~~~~~~~~~~~~~~~~~~~
+Note that dead-ness is maintained by the simplifier, so that it is
+accurate after simplification as well as before.
+
+
+Note [Cascading case merge]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Case merging should cascade in one sweep, because it
+happens bottom-up
+
+ case e of a {
+ DEFAULT -> case a of b
+ DEFAULT -> case b of c {
+ DEFAULT -> e
+ A -> ea
+ B -> eb
+ C -> ec
+==>
+ case e of a {
+ DEFAULT -> case a of b
+ DEFAULT -> let c = b in e
+ A -> let c = b in ea
+ B -> eb
+ C -> ec
+==>
+ case e of a {
+ DEFAULT -> let b = a in let c = b in e
+ A -> let b = a in let c = b in ea
+ B -> let b = a in eb
+ C -> ec
+
+
+However here's a tricky case that we still don't catch, and I don't
+see how to catch it in one pass:
+
+ case x of c1 { I# a1 ->
+ case a1 of c2 ->
+ 0 -> ...
+ DEFAULT -> case x of c3 { I# a2 ->
+ case a2 of ...
+
+After occurrence analysis (and its binder-swap) we get this
+
+ case x of c1 { I# a1 ->
+ let x = c1 in -- Binder-swap addition
+ case a1 of c2 ->
+ 0 -> ...
+ DEFAULT -> case x of c3 { I# a2 ->
+ case a2 of ...
+
+When we simplify the inner case x, we'll see that
+x=c1=I# a1. So we'll bind a2 to a1, and get
+
+ case x of c1 { I# a1 ->
+ case a1 of c2 ->
+ 0 -> ...
+ DEFAULT -> case a1 of ...
+
+This is corect, but we can't do a case merge in this sweep
+because c2 /= a1. Reason: the binding c1=I# a1 went inwards
+without getting changed to c1=I# c2.
+
+I don't think this is worth fixing, even if I knew how. It'll
+all come out in the next pass anyway.
+
+
\ No newline at end of file
projects / ghc-hetmet.git / blobdiff