import CmdLineOpts ( DynFlags, DynFlag(..), opt_OmitInterfacePragmas )
import CoreSyn
import CoreUnfold ( noUnfolding, mkTopUnfolding, okToUnfoldInHiFile )
-import CoreUtils ( exprArity, exprIsBottom )
+import CoreUtils ( exprArity )
import CoreFVs ( ruleSomeFreeVars, exprSomeFreeVars )
import CoreLint ( showPass, endPass )
import VarEnv
import VarSet
import Var ( Id, Var )
import Id ( idType, idInfo, idName, isExportedId,
- idCafInfo, mkId, isLocalId, omitIfaceSigForId,
- idFlavour, modifyIdInfo
+ idCafInfo, mkId, isLocalId, isImplicitId,
+ idFlavour, modifyIdInfo, idArity
)
import IdInfo {- loads of stuff -}
import Name ( getOccName, nameOccName, globaliseName, setNameOcc,
- localiseName, mkLocalName, isGlobalName
+ localiseName, mkLocalName, isGlobalName, isDllName
)
import OccName ( TidyOccEnv, initTidyOccEnv, tidyOccName )
import Type ( tidyTopType, tidyType, tidyTyVar )
import PrimOp ( PrimOp(..), setCCallUnique )
import HscTypes ( PersistentCompilerState( pcs_PRS ),
PersistentRenamerState( prsOrig ),
- OrigNameEnv( origNames ), OrigNameNameEnv
+ NameSupply( nsNames ), OrigNameCache
)
import UniqSupply
+import DataCon ( DataCon, dataConName )
+import Literal ( isLitLitLit )
import FiniteMap ( lookupFM, addToFM )
import Maybes ( maybeToBool, orElse )
import ErrUtils ( showPass )
+import PprCore ( pprIdCoreRule )
import SrcLoc ( noSrcLoc )
import UniqFM ( mapUFM )
import Outputable
that all Ids are unique, rather than the weaker guarantee of
no clashes which the simplifier provides.
- - Give the Id its final IdInfo; in ptic,
+ - Give each dynamic CCall occurrence a fresh unique; this is
+ rather like the cloning step above.
+
+ - Give the Id its UTTERLY FINAL IdInfo; in ptic,
* Its flavour becomes ConstantId, reflecting the fact that
from now on we regard it as a constant, not local, Id
+
* its unfolding, if it should have one
+
+ * its arity, computed from the number of visible lambdas
+
+ * its CAF info, computed from what is free in its RHS
+
Finally, substitute these new top-level binders consistently
throughout, including in unfoldings. We also tidy binders in
; let (orphans_out, _)
= initUs us1 (tidyIdRules (occ_env,subst_env) orphans_in)
- ; let prs' = prs { prsOrig = orig { origNames = orig_env' } }
+ ; let prs' = prs { prsOrig = orig { nsNames = orig_env' } }
pcs' = pcs { pcs_PRS = prs' }
; endPass dflags "Tidy Core" Opt_D_dump_simpl binds_out
-- decl. tidyTopId then does a no-op on exported binders.
prs = pcs_PRS pcs
orig = prsOrig prs
- orig_env = origNames orig
+ orig_env = nsNames orig
init_tidy_env us = (us, orig_env, initTidyOccEnv avoids, emptyVarEnv)
avoids = [getOccName bndr | bndr <- bindersOfBinds binds_in,
-> IdEnv Bool -- True <=> show unfolding
-- Step 1 from the notes above
findExternalSet binds orphan_rules
- = foldr find init_needed binds
+ = pprTrace "fes" (vcat (map pprIdCoreRule orphan_rules) $$ ppr (varSetElems orphan_rule_ids)) $
+ foldr find init_needed binds
where
orphan_rule_ids :: IdSet
orphan_rule_ids = unionVarSets [ ruleSomeFreeVars isIdAndLocal rule
\begin{code}
-type TopTidyEnv = (UniqSupply, OrigNameNameEnv, TidyOccEnv, VarEnv Var)
+type TopTidyEnv = (UniqSupply, OrigNameCache, TidyOccEnv, VarEnv Var)
-- TopTidyEnv: when tidying we need to know
-- * orig_env: Any pre-ordained Names. These may have arisen because the
tidyTopBinder mod ext_ids tidy_env rhs caf_info
env@(us, orig_env2, occ_env2, subst_env2) id
- | omitIfaceSigForId id -- Don't mess with constructors,
- = (env, id) -- record selectors, and the like
+ | isImplicitId id -- Don't mess with constructors,
+ = (env, id) -- record selectors, and the like
| otherwise
-- This function is the heart of Step 2
tidyIdInfo us tidy_env is_external unfold_info arity_info caf_info id
| opt_OmitInterfacePragmas || not is_external
-- No IdInfo if the Id isn't external, or if we don't have -O
- = mkIdInfo new_flavour
+ = mkIdInfo new_flavour caf_info
`setStrictnessInfo` strictnessInfo core_idinfo
`setArityInfo` ArityExactly arity_info
- `setCafInfo` caf_info
-- Keep strictness, arity and CAF info; it's used by the code generator
| otherwise
= let (rules', _) = initUs us (tidyRules tidy_env (specInfo core_idinfo))
in
- mkIdInfo new_flavour
+ mkIdInfo new_flavour caf_info
`setCprInfo` cprInfo core_idinfo
`setStrictnessInfo` strictnessInfo core_idinfo
`setInlinePragInfo` inlinePragInfo core_idinfo
`setUnfoldingInfo` unfold_info
- `setWorkerInfo` tidyWorker tidy_env (workerInfo core_idinfo)
+ `setWorkerInfo` tidyWorker tidy_env arity_info (workerInfo core_idinfo)
`setSpecInfo` rules'
`setArityInfo` ArityExactly arity_info
- `setCafInfo` caf_info
-- this is the final IdInfo, it must agree with the
-- code finally generated (i.e. NO more transformations
-- after this!).
flavour -> pprTrace "tidyIdInfo" (ppr id <+> ppFlavourInfo flavour)
flavour
+
-- This is where we set names to local/global based on whether they really are
-- externally visible (see comment at the top of this module). If the name
-- was previously local, we have to give it a unique occurrence name if
local = not global
internal = not external
+------------ Worker --------------
+-- We only treat a function as having a worker if
+-- the exported arity (which is now the number of visible lambdas)
+-- is the same as the arity at the moment of the w/w split
+-- If so, we can safely omit the unfolding inside the wrapper, and
+-- instead re-generate it from the type/arity/strictness info
+-- But if the arity has changed, we just take the simple path and
+-- put the unfolding into the interface file, forgetting the fact
+-- that it's a wrapper.
+--
+-- How can this happen? Sometimes we get
+-- f = coerce t (\x y -> $wf x y)
+-- at the moment of w/w split; but the eta reducer turns it into
+-- f = coerce t $wf
+-- which is perfectly fine except that the exposed arity so far as
+-- the code generator is concerned (zero) differs from the arity
+-- when we did the split (2).
+--
+-- All this arises because we use 'arity' to mean "exactly how many
+-- top level lambdas are there" in interface files; but during the
+-- compilation of this module it means "how many things can I apply
+-- this to".
+tidyWorker tidy_env real_arity (HasWorker work_id wrap_arity)
+ | real_arity == wrap_arity
+ = HasWorker (tidyVarOcc tidy_env work_id) wrap_arity
+tidyWorker tidy_env real_arity other
+ = NoWorker
+
+------------ Rules --------------
tidyIdRules :: TidyEnv -> [IdCoreRule] -> UniqSM [IdCoreRule]
tidyIdRules env [] = returnUs []
tidyIdRules env ((fn,rule) : rules)
tidyIdRules env rules `thenUs` \ rules ->
returnUs ((tidyVarOcc env fn, rule) : rules)
-tidyWorker tidy_env (HasWorker work_id wrap_arity)
- = HasWorker (tidyVarOcc tidy_env work_id) wrap_arity
-tidyWorker tidy_env NoWorker
- = NoWorker
-
tidyRules :: TidyEnv -> CoreRules -> UniqSM CoreRules
tidyRules env (Rules rules fvs)
= mapUs (tidyRule env) rules `thenUs` \ rules ->
tidyBndr :: TidyEnv -> Var -> UniqSM (TidyEnv, Var)
tidyBndr env var
| isTyVar var = returnUs (tidyTyVar env var)
- | otherwise = tidyId env var (vanillaIdInfo `setCafInfo` NoCafRefs)
+ | otherwise = tidyId env var vanillaIdInfo
tidyBndrs :: TidyEnv -> [Var] -> UniqSM (TidyEnv, [Var])
tidyBndrs env vars = mapAccumLUs tidyBndr env vars
= tidyId env id idinfo
where
idinfo = vanillaIdInfo `setArityInfo` ArityExactly (exprArity rhs)
- `setCafInfo` NoCafRefs
-- NB: This throws away the IdInfo of the Id, which we
-- no longer need. That means we don't need to
-- run over it with env, nor renumber it.
%* *
%************************************************************************
+hasCafRefs decides whether a top-level closure can point into the dynamic heap.
+We mark such things as `MayHaveCafRefs' because this information is
+used to decide whether a particular closure needs to be referenced
+in an SRT or not.
+
+There are two reasons for setting MayHaveCafRefs:
+ a) The RHS is a CAF: a top-level updatable thunk.
+ b) The RHS refers to something that MayHaveCafRefs
+
+Possible improvement: In an effort to keep the number of CAFs (and
+hence the size of the SRTs) down, we could also look at the expression and
+decide whether it requires a small bounded amount of heap, so we can ignore
+it as a CAF. In these cases however, we would need to use an additional
+CAF list to keep track of non-collectable CAFs.
+
\begin{code}
hasCafRefs :: (Id -> Bool) -> CoreExpr -> CafInfo
+-- Only called for the RHS of top-level lets
hasCafRefss :: (Id -> Bool) -> [CoreExpr] -> CafInfo
-- predicate returns True for a given Id if we look at this Id when
-- calculating the result. Used to *avoid* looking at the CafInfo
then MayHaveCafRefs
else NoCafRefs
+ -- used for recursive groups. The whole group is set to
+ -- "MayHaveCafRefs" if at least one of the group is a CAF or
+ -- refers to any CAFs.
hasCafRefss p exprs = if any isCAF exprs || isFastTrue (cafRefss p exprs)
then MayHaveCafRefs
else NoCafRefs
cafRefss p [] = fastBool False
cafRefss p (e:es) = cafRefs p e `fastOr` cafRefss p es
--- Decide whether a closure looks like a CAF or not. In an effort to
--- keep the number of CAFs (and hence the size of the SRTs) down, we
--- would also like to look at the expression and decide whether it
--- requires a small bounded amount of heap, so we can ignore it as a CAF.
--- In these cases, we need to use an additional CAF list to keep track of
--- non-collectable CAFs.
---
--- We mark real CAFs as `MayHaveCafRefs' because this information is used
--- to decide whether a particular closure needs to be referenced in an
--- SRT or not.
isCAF :: CoreExpr -> Bool
- -- special case for expressions which are always bottom,
- -- such as 'error "..."'. We don't need to record it as
- -- a CAF, since it can only be entered once.
-isCAF e
- | not_function && is_bottom = False
- | not_function && updatable = True
- | otherwise = False
- where
- not_function = exprArity e == 0
- is_bottom = exprIsBottom e
- updatable = True {- ToDo: check type for onceness? -}
+-- Only called for the RHS of top-level lets
+isCAF e = not (rhsIsNonUpd e)
+ {- ToDo: check type for onceness, i.e. non-updatable thunks? -}
+
+rhsIsNonUpd :: CoreExpr -> Bool
+ -- True => Value-lambda, constructor, PAP
+ -- This is a bit like CoreUtils.exprIsValue, with the following differences:
+ -- a) scc "foo" (\x -> ...) is updatable (so we catch the right SCC)
+ --
+ -- b) (C x xs), where C is a contructors is updatable if the application is
+ -- dynamic: see isDynConApp
+ --
+ -- c) don't look through unfolding of f in (f x). I'm suspicious of this one
+
+rhsIsNonUpd (Lam b e) = isId b || rhsIsNonUpd e
+rhsIsNonUpd (Note (SCC _) e) = False
+rhsIsNonUpd (Note _ e) = rhsIsNonUpd e
+rhsIsNonUpd other_expr
+ = go other_expr 0 []
+ where
+ go (Var f) n_args args = idAppIsNonUpd f n_args args
+
+ go (App f a) n_args args
+ | isTypeArg a = go f n_args args
+ | otherwise = go f (n_args + 1) (a:args)
+
+ go (Note (SCC _) f) n_args args = False
+ go (Note _ f) n_args args = go f n_args args
+
+ go other n_args args = False
+
+idAppIsNonUpd :: Id -> Int -> [CoreExpr] -> Bool
+idAppIsNonUpd id n_val_args args
+ = case idFlavour id of
+ DataConId con | not (isDynConApp con args) -> True
+ other -> n_val_args < idArity id
+
+isDynConApp :: DataCon -> [CoreExpr] -> Bool
+isDynConApp con args = isDllName (dataConName con) || any isDynArg args
+-- Top-level constructor applications can usually be allocated
+-- statically, but they can't if
+-- a) the constructor, or any of the arguments, come from another DLL
+-- b) any of the arguments are LitLits
+-- (because we can't refer to static labels in other DLLs).
+-- If this happens we simply make the RHS into an updatable thunk,
+-- and 'exectute' it rather than allocating it statically.
+-- All this should match the decision in (see CoreToStg.coreToStgRhs)
+
+
+isDynArg :: CoreExpr -> Bool
+isDynArg (Var v) = isDllName (idName v)
+isDynArg (Note _ e) = isDynArg e
+isDynArg (Lit lit) = isLitLitLit lit
+isDynArg (App e _) = isDynArg e -- must be a type app
+isDynArg (Lam _ e) = isDynArg e -- must be a type lam
\end{code}