\section{Tidying up Core}
\begin{code}
-module TidyPgm( mkBootModDetailsDs, mkBootModDetailsTc, tidyProgram ) where
+module TidyPgm( mkBootModDetailsDs, mkBootModDetailsTc,
+ tidyProgram, globaliseAndTidyId ) where
#include "HsVersions.h"
import CoreUnfold
import CoreFVs
import CoreTidy
-import PprCore
-import CoreLint
+import CoreMonad
import CoreUtils
+import Rules
+import CoreArity ( exprArity, exprBotStrictness_maybe )
+import Class ( classAllSelIds )
import VarEnv
import VarSet
import Var
import Id
-import Class
import IdInfo
import InstEnv
-import NewDemand
+import Demand
import BasicTypes
-import Name
+import Name hiding (varName)
import NameSet
import IfaceEnv
import NameEnv
-import OccName
import TcType
import DataCon
import TyCon
import Module
+import Packages( isDllName )
import HscTypes
import Maybes
-import ErrUtils
import UniqSupply
import Outputable
import FastBool hiding ( fastOr )
+import Util
+import FastString
-import Data.List ( partition )
-import Data.Maybe ( isJust )
+import Control.Monad ( when )
+import Data.List ( sortBy )
import Data.IORef ( IORef, readIORef, writeIORef )
\end{code}
-Constructing the TypeEnv, Instances, Rules from which the ModIface is
-constructed, and which goes on to subsequent modules in --make mode.
+Constructing the TypeEnv, Instances, Rules, VectInfo from which the
+ModIface is constructed, and which goes on to subsequent modules in
+--make mode.
Most of the interface file is obtained simply by serialising the
TypeEnv. One important consequence is that if the *interface file*
-> [Instance] -> [FamInstEnv.FamInst] -> IO ModDetails
mkBootModDetails hsc_env exports type_env insts fam_insts
= do { let dflags = hsc_dflags hsc_env
- ; showPass dflags "Tidy [hoot] type env"
+ ; showPass dflags CoreTidy
- ; let { insts' = tidyInstances tidyExternalId insts
+ ; let { insts' = tidyInstances globaliseAndTidyId insts
; dfun_ids = map instanceDFunId insts'
; type_env1 = tidyBootTypeEnv (availsToNameSet exports) type_env
; type_env' = extendTypeEnvWithIds type_env1 dfun_ids
-- because we don't tidy the OccNames, and if we don't remove
-- the non-exported ones we'll get many things with the
-- same name in the interface file, giving chaos.
- final_ids = [ tidyExternalId id
+ final_ids = [ globaliseAndTidyId id
| id <- typeEnvIds type_env
, isLocalId id
, keep_it id ]
keep_it id = isExportedId id || idName id `elemNameSet` exports
-tidyExternalId :: Id -> Id
+
+globaliseAndTidyId :: Id -> Id
-- Takes an LocalId with an External Name,
--- makes it into a GlobalId with VanillaIdInfo, and tidies its type
--- (NB: vanillaIdInfo makes a conservative assumption about Caf-hood.)
-tidyExternalId id
- = ASSERT2( isLocalId id && isExternalName (idName id), ppr id )
- mkVanillaGlobal (idName id) (tidyTopType (idType id))
+-- makes it into a GlobalId
+-- * unchanged Name (might be Internal or External)
+-- * unchanged details
+-- * VanillaIdInfo (makes a conservative assumption about Caf-hood)
+globaliseAndTidyId id
+ = Id.setIdType (globaliseId id) tidy_type
+ where
+ tidy_type = tidyTopType (idType id)
\end{code}
Step 1: Figure out external Ids
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Note [choosing external names]
+
+See also the section "Interface stability" in the
+RecompilationAvoidance commentary:
+ http://hackage.haskell.org/trac/ghc/wiki/Commentary/Compiler/RecompilationAvoidance
+
First we figure out which Ids are "external" Ids. An
"external" Id is one that is visible from outside the compilation
unit. These are
a) the user exported ones
b) ones mentioned in the unfoldings, workers,
or rules of externally-visible ones
-This exercise takes a sweep of the bindings bottom to top. Actually,
-in Step 2 we're also going to need to know which Ids should be
-exported with their unfoldings, so we produce not an IdSet but an
-ExtIdEnv = IdEnv Bool
+While figuring out which Ids are external, we pick a "tidy" OccName
+for each one. That is, we make its OccName distinct from the other
+external OccNames in this module, so that in interface files and
+object code we can refer to it unambiguously by its OccName. The
+OccName for each binder is prefixed by the name of the exported Id
+that references it; e.g. if "f" references "x" in its unfolding, then
+"x" is renamed to "f_x". This helps distinguish the different "x"s
+from each other, and means that if "f" is later removed, things that
+depend on the other "x"s will not need to be recompiled. Of course,
+if there are multiple "f_x"s, then we have to disambiguate somehow; we
+use "f_x0", "f_x1" etc.
+
+As far as possible we should assign names in a deterministic fashion.
+Each time this module is compiled with the same options, we should end
+up with the same set of external names with the same types. That is,
+the ABI hash in the interface should not change. This turns out to be
+quite tricky, since the order of the bindings going into the tidy
+phase is already non-deterministic, as it is based on the ordering of
+Uniques, which are assigned unpredictably.
+
+To name things in a stable way, we do a depth-first-search of the
+bindings, starting from the exports sorted by name. This way, as long
+as the bindings themselves are deterministic (they sometimes aren't!),
+the order in which they are presented to the tidying phase does not
+affect the names we assign.
Step 2: Tidy the program
~~~~~~~~~~~~~~~~~~~~~~~~
For external Ids, use the original-name cache in the NameCache
to ensure that the unique assigned is the same as the Id had
in any previous compilation run.
-
- 3. If it's an external Id, make it have a External Name, otherwise
- make it have an Internal Name.
- This is used by the code generator to decide whether
- to make the label externally visible
-
- 4. Give external Ids a "tidy" OccName. This means
- we can print them in interface files without confusing
- "x" (unique 5) with "x" (unique 10).
-
- 5. Give it its UTTERLY FINAL IdInfo; in ptic,
+
+ 3. Rename top-level Ids according to the names we chose in step 1.
+ If it's an external Id, make it have a External Name, otherwise
+ make it have an Internal Name. This is used by the code generator
+ to decide whether to make the label externally visible
+
+ 4. Give it its UTTERLY FINAL IdInfo; in ptic,
* its unfolding, if it should have one
* its arity, computed from the number of visible lambdas
mg_binds = binds,
mg_rules = imp_rules,
mg_vect_info = vect_info,
- mg_dir_imps = dir_imps,
- mg_anns = anns,
+ mg_anns = anns,
mg_deps = deps,
mg_foreign = foreign_stubs,
mg_hpc_info = hpc_info,
mg_modBreaks = modBreaks })
- = do { let dflags = hsc_dflags hsc_env
- ; showPass dflags "Tidy Core"
-
- ; let { omit_prags = dopt Opt_OmitInterfacePragmas dflags
- ; th = dopt Opt_TemplateHaskell dflags
- ; ext_ids = findExternalIds omit_prags binds
- ; ext_rules
- | omit_prags = []
- | otherwise = findExternalRules binds imp_rules ext_ids
- -- findExternalRules filters imp_rules to avoid binders that
- -- aren't externally visible; but the externally-visible binders
- -- are computed (by findExternalIds) assuming that all orphan
- -- rules are exported (they get their Exported flag set in the desugarer)
- -- So in fact we may export more than we need.
- -- (It's a sort of mutual recursion.)
- }
-
- ; (tidy_env, tidy_binds) <- tidyTopBinds hsc_env mod type_env ext_ids
- binds
+ = do { let { dflags = hsc_dflags hsc_env
+ ; omit_prags = dopt Opt_OmitInterfacePragmas dflags
+ ; expose_all = dopt Opt_ExposeAllUnfoldings dflags
+ ; th = xopt Opt_TemplateHaskell dflags
+ }
+ ; showPass dflags CoreTidy
+
+ ; let { implicit_binds = getImplicitBinds type_env }
+
+ ; (unfold_env, tidy_occ_env)
+ <- chooseExternalIds hsc_env mod omit_prags expose_all
+ binds implicit_binds imp_rules
+
+ ; let { ext_rules = findExternalRules omit_prags binds imp_rules unfold_env }
+ -- Glom together imp_rules and rules currently attached to binders
+ -- Then pick just the ones we need to expose
+ -- See Note [Which rules to expose]
+
+ ; let { (tidy_env, tidy_binds)
+ = tidyTopBinds hsc_env unfold_env tidy_occ_env binds }
; let { export_set = availsToNameSet exports
; final_ids = [ id | id <- bindersOfBinds tidy_binds,
-- and indeed it does, but if omit_prags is on, ext_rules is
-- empty
+ ; tidy_vect_info = tidyVectInfo tidy_env vect_info
+
-- See Note [Injecting implicit bindings]
- ; implicit_binds = getImplicitBinds type_env
; all_tidy_binds = implicit_binds ++ tidy_binds
; alg_tycons = filter isAlgTyCon (typeEnvTyCons type_env)
}
- ; endPass dflags "Tidy Core" Opt_D_dump_simpl all_tidy_binds
- ; dumpIfSet_core dflags Opt_D_dump_simpl
- "Tidy Core Rules"
- (pprRules tidy_rules)
-
- ; let dir_imp_mods = moduleEnvKeys dir_imps
-
- ; return (CgGuts { cg_module = mod,
- cg_tycons = alg_tycons,
- cg_binds = all_tidy_binds,
- cg_dir_imps = dir_imp_mods,
- cg_foreign = foreign_stubs,
+ ; endPass dflags CoreTidy all_tidy_binds tidy_rules
+
+ -- If the endPass didn't print the rules, but ddump-rules is on, print now
+ ; dumpIfSet (dopt Opt_D_dump_rules dflags
+ && (not (dopt Opt_D_dump_simpl dflags)))
+ CoreTidy
+ (ptext (sLit "rules"))
+ (pprRulesForUser tidy_rules)
+
+ -- Print one-line size info
+ ; let cs = coreBindsStats tidy_binds
+ ; when (dopt Opt_D_dump_core_stats dflags)
+ (printDump (ptext (sLit "Tidy size (terms,types,coercions)")
+ <+> ppr (moduleName mod) <> colon
+ <+> int (cs_tm cs)
+ <+> int (cs_ty cs)
+ <+> int (cs_co cs) ))
+
+ ; return (CgGuts { cg_module = mod,
+ cg_tycons = alg_tycons,
+ cg_binds = all_tidy_binds,
+ cg_foreign = foreign_stubs,
cg_dep_pkgs = dep_pkgs deps,
cg_hpc_info = hpc_info,
cg_modBreaks = modBreaks },
ModDetails { md_types = tidy_type_env,
md_rules = tidy_rules,
md_insts = tidy_insts,
- md_fam_insts = fam_insts,
+ md_vect_info = tidy_vect_info,
+ md_fam_insts = fam_insts,
md_exports = exports,
- md_anns = anns, -- are already tidy
- md_vect_info = vect_info --
+ md_anns = anns -- are already tidy
})
}
| isEnumerationTyCon tc -- For an enumeration, exposing the constructors
= True -- won't lead to the need for further exposure
-- (This includes data types with no constructors.)
- | isOpenTyCon tc -- Open type family
+ | isFamilyTyCon tc -- Open type family
= True
| otherwise -- Newtype, datatype
tidy_dfun (instanceDFunId ispec)
\end{code}
+\begin{code}
+tidyVectInfo :: TidyEnv -> VectInfo -> VectInfo
+tidyVectInfo (_, var_env) info@(VectInfo { vectInfoVar = vars
+ , vectInfoPADFun = pas
+ , vectInfoIso = isos
+ , vectInfoScalarVars = scalarVars
+ })
+ = info { vectInfoVar = tidy_vars
+ , vectInfoPADFun = tidy_pas
+ , vectInfoIso = tidy_isos
+ , vectInfoScalarVars = tidy_scalarVars
+ }
+ where
+ tidy_vars = mkVarEnv
+ $ map tidy_var_mapping
+ $ varEnvElts vars
+
+ tidy_pas = mapNameEnv tidy_snd_var pas
+ tidy_isos = mapNameEnv tidy_snd_var isos
+
+ tidy_var_mapping (from, to) = (from', (from', lookup_var to))
+ where from' = lookup_var from
+ tidy_snd_var (x, var) = (x, lookup_var var)
+
+ tidy_scalarVars = mkVarSet
+ $ map lookup_var
+ $ varSetElems scalarVars
+
+ lookup_var var = lookupWithDefaultVarEnv var_env var var
+\end{code}
+
%************************************************************************
%* *
optimisation first. (Only matters when the selector is used curried;
eg map x ys.) See Trac #2070.
+[Oct 09: in fact, record selectors are no longer implicit Ids at all,
+because we really do want to optimise them properly. They are treated
+much like any other Id. But doing "light" optimisation on an implicit
+Id still makes sense.]
+
At one time I tried injecting the implicit bindings *early*, at the
beginning of SimplCore. But that gave rise to real difficulty,
becuase GlobalIds are supposed to have *fixed* IdInfo, but the
importing modules were expecting it to have arity 1 (Trac #2844).
It's much safer just to inject them right at the end, after tidying.
+Oh: two other reasons for injecting them late:
+
+ - If implicit Ids are already in the bindings when we start TidyPgm,
+ we'd have to be careful not to treat them as external Ids (in
+ the sense of findExternalIds); else the Ids mentioned in *their*
+ RHSs will be treated as external and you get an interface file
+ saying a18 = <blah>
+ but nothing refererring to a18 (because the implicit Id is the
+ one that does, and implicit Ids don't appear in interface files).
+
+ - More seriously, the tidied type-envt will include the implicit
+ Id replete with a18 in its unfolding; but we won't take account
+ of a18 when computing a fingerprint for the class; result chaos.
+
+There is one sort of implicit binding that is injected still later,
+namely those for data constructor workers. Reason (I think): it's
+really just a code generation trick.... binding itself makes no sense.
+See CorePrep Note [Data constructor workers].
\begin{code}
getImplicitBinds :: TypeEnv -> [CoreBind]
getImplicitBinds type_env
- = map get_defn (concatMap implicit_con_ids (typeEnvTyCons type_env)
- ++ concatMap other_implicit_ids (typeEnvElts type_env))
- -- Put the constructor wrappers first, because
- -- other implicit bindings (notably the fromT functions arising
- -- from generics) use the constructor wrappers. At least that's
- -- what External Core likes
+ = map get_defn (concatMap implicit_ids (typeEnvElts type_env))
where
- implicit_con_ids tc = mapCatMaybes dataConWrapId_maybe (tyConDataCons tc)
-
- other_implicit_ids (ATyCon tc) = filter (not . isNaughtyRecordSelector) (tyConSelIds tc)
- -- The "naughty" ones are not real functions at all
- -- They are there just so we can get decent error messages
- -- See Note [Naughty record selectors] in MkId.lhs
- other_implicit_ids (AClass cl) = classSelIds cl
- other_implicit_ids _other = []
+ implicit_ids (ATyCon tc) = mapCatMaybes dataConWrapId_maybe (tyConDataCons tc)
+ implicit_ids (AClass cls) = classAllSelIds cls
+ implicit_ids _ = []
get_defn :: Id -> CoreBind
- get_defn id = NonRec id (unfoldingTemplate (idUnfolding id))
+ get_defn id = NonRec id (unfoldingTemplate (realIdUnfolding id))
\end{code}
%* *
%************************************************************************
-\begin{code}
-type ExtIdEnv = IdEnv Bool
- -- In domain => Id is external
- -- Range = True <=> show unfolding,
- -- Always True for InlineRule
+Sete Note [choosing external names].
-findExternalIds :: Bool -> [CoreBind] -> ExtIdEnv
+\begin{code}
+type UnfoldEnv = IdEnv (Name{-new name-}, Bool {-show unfolding-})
+ -- Maps each top-level Id to its new Name (the Id is tidied in step 2)
+ -- The Unique is unchanged. If the new Name is external, it will be
+ -- visible in the interface file.
+ --
+ -- Bool => expose unfolding or not.
+
+chooseExternalIds :: HscEnv
+ -> Module
+ -> Bool -> Bool
+ -> [CoreBind]
+ -> [CoreBind]
+ -> [CoreRule]
+ -> IO (UnfoldEnv, TidyOccEnv)
-- Step 1 from the notes above
-findExternalIds omit_prags binds
- | omit_prags
- = mkVarEnv [ (id,False) | id <- bindersOfBinds binds, isExportedId id ]
- | otherwise
- = foldr find emptyVarEnv binds
- where
- find (NonRec id rhs) needed
- | need_id needed id = addExternal (id,rhs) needed
- | otherwise = needed
- find (Rec prs) needed = find_prs prs needed
-
- -- For a recursive group we have to look for a fixed point
- find_prs prs needed
- | null needed_prs = needed
- | otherwise = find_prs other_prs new_needed
- where
- (needed_prs, other_prs) = partition (need_pr needed) prs
- new_needed = foldr addExternal needed needed_prs
-
- -- The 'needed' set contains the Ids that are needed by earlier
- -- interface file emissions. If the Id isn't in this set, and isn't
- -- exported, there's no need to emit anything
- need_id needed_set id = id `elemVarEnv` needed_set || isExportedId id
- need_pr needed_set (id,_) = need_id needed_set id
-
-addExternal :: (Id,CoreExpr) -> IdEnv Bool -> IdEnv Bool
--- The Id is needed; extend the needed set
--- with it and its dependents (free vars etc)
-addExternal (id,rhs) needed
- = extendVarEnv (foldVarSet add_occ needed new_needed_ids)
- id show_unfold
- where
- add_occ id needed | id `elemVarEnv` needed = needed
- | otherwise = extendVarEnv needed id False
- -- "False" because we don't know we need the Id's unfolding
- -- Don't override existing bindings; we might have already set it to True
+chooseExternalIds hsc_env mod omit_prags expose_all binds implicit_binds imp_id_rules
+ = do { (unfold_env1,occ_env1) <- search init_work_list emptyVarEnv init_occ_env
+ ; let internal_ids = filter (not . (`elemVarEnv` unfold_env1)) binders
+ ; tidy_internal internal_ids unfold_env1 occ_env1 }
+ where
+ nc_var = hsc_NC hsc_env
+
+ -- init_ext_ids is the intial list of Ids that should be
+ -- externalised. It serves as the starting point for finding a
+ -- deterministic, tidy, renaming for all external Ids in this
+ -- module.
+ --
+ -- It is sorted, so that it has adeterministic order (i.e. it's the
+ -- same list every time this module is compiled), in contrast to the
+ -- bindings, which are ordered non-deterministically.
+ init_work_list = zip init_ext_ids init_ext_ids
+ init_ext_ids = sortBy (compare `on` getOccName) $
+ filter is_external binders
+
+ -- An Id should be external if either (a) it is exported or
+ -- (b) it appears in the RHS of a local rule for an imported Id.
+ -- See Note [Which rules to expose]
+ is_external id = isExportedId id || id `elemVarSet` rule_rhs_vars
+ rule_rhs_vars = foldr (unionVarSet . ruleRhsFreeVars) emptyVarSet imp_id_rules
+
+ binders = bindersOfBinds binds
+ implicit_binders = bindersOfBinds implicit_binds
+ binder_set = mkVarSet binders
+
+ avoids = [getOccName name | bndr <- binders ++ implicit_binders,
+ let name = idName bndr,
+ isExternalName name ]
+ -- In computing our "avoids" list, we must include
+ -- all implicit Ids
+ -- all things with global names (assigned once and for
+ -- all by the renamer)
+ -- since their names are "taken".
+ -- The type environment is a convenient source of such things.
+ -- In particular, the set of binders doesn't include
+ -- implicit Ids at this stage.
- new_needed_ids = (mb_unfold_ids `orElse` emptyVarSet) `unionVarSet`
- spec_ids
+ -- We also make sure to avoid any exported binders. Consider
+ -- f{-u1-} = 1 -- Local decl
+ -- ...
+ -- f{-u2-} = 2 -- Exported decl
+ --
+ -- The second exported decl must 'get' the name 'f', so we
+ -- have to put 'f' in the avoids list before we get to the first
+ -- decl. tidyTopId then does a no-op on exported binders.
+ init_occ_env = initTidyOccEnv avoids
+
+
+ search :: [(Id,Id)] -- The work-list: (external id, referrring id)
+ -- Make a tidy, external Name for the external id,
+ -- add it to the UnfoldEnv, and do the same for the
+ -- transitive closure of Ids it refers to
+ -- The referring id is used to generate a tidy
+ --- name for the external id
+ -> UnfoldEnv -- id -> (new Name, show_unfold)
+ -> TidyOccEnv -- occ env for choosing new Names
+ -> IO (UnfoldEnv, TidyOccEnv)
+
+ search [] unfold_env occ_env = return (unfold_env, occ_env)
+
+ search ((idocc,referrer) : rest) unfold_env occ_env
+ | idocc `elemVarEnv` unfold_env = search rest unfold_env occ_env
+ | otherwise = do
+ (occ_env', name') <- tidyTopName mod nc_var (Just referrer) occ_env idocc
+ let
+ (new_ids, show_unfold)
+ | omit_prags = ([], False)
+ | otherwise = addExternal expose_all refined_id
+
+ -- 'idocc' is an *occurrence*, but we need to see the
+ -- unfolding in the *definition*; so look up in binder_set
+ refined_id = case lookupVarSet binder_set idocc of
+ Just id -> id
+ Nothing -> WARN( True, ppr idocc ) idocc
+
+ unfold_env' = extendVarEnv unfold_env idocc (name',show_unfold)
+ referrer' | isExportedId refined_id = refined_id
+ | otherwise = referrer
+ --
+ search (zip new_ids (repeat referrer') ++ rest) unfold_env' occ_env'
+
+ tidy_internal :: [Id] -> UnfoldEnv -> TidyOccEnv
+ -> IO (UnfoldEnv, TidyOccEnv)
+ tidy_internal [] unfold_env occ_env = return (unfold_env,occ_env)
+ tidy_internal (id:ids) unfold_env occ_env = do
+ (occ_env', name') <- tidyTopName mod nc_var Nothing occ_env id
+ let unfold_env' = extendVarEnv unfold_env id (name',False)
+ tidy_internal ids unfold_env' occ_env'
+
+addExternal :: Bool -> Id -> ([Id],Bool)
+addExternal expose_all id = (new_needed_ids, show_unfold)
+ where
+ new_needed_ids = unfold_ids ++
+ filter (\id -> isLocalId id &&
+ not (id `elemVarSet` unfold_set))
+ (varSetElems spec_ids) -- XXX non-det ordering
idinfo = idInfo id
- dont_inline = isNeverActive (inlinePragInfo idinfo)
+ never_active = isNeverActive (inlinePragmaActivation (inlinePragInfo idinfo))
loop_breaker = isNonRuleLoopBreaker (occInfo idinfo)
- bottoming_fn = isBottomingSig (newStrictnessInfo idinfo `orElse` topSig)
+ bottoming_fn = isBottomingSig (strictnessInfo idinfo `orElse` topSig)
spec_ids = specInfoFreeVars (specInfo idinfo)
-- Stuff to do with the Id's unfolding
-- We leave the unfolding there even if there is a worker
-- In GHCI the unfolding is used by importers
show_unfold = isJust mb_unfold_ids
+ (unfold_set, unfold_ids) = mb_unfold_ids `orElse` (emptyVarSet, [])
- mb_unfold_ids :: Maybe IdSet -- Nothing => don't unfold
+ mb_unfold_ids :: Maybe (IdSet, [Id]) -- Nothing => don't unfold
mb_unfold_ids = case unfoldingInfo idinfo of
- InlineRule { uf_worker = Just wkr_id } -> Just (unitVarSet wkr_id)
- InlineRule { uf_tmpl = rhs } -> Just (exprFreeIds rhs)
- CoreUnfolding { uf_guidance = guide }
- | not bottoming_fn -- Not necessary
- , not dont_inline
- , not loop_breaker
- , not (neverUnfoldGuidance guide)
- -> Just (exprFreeIds rhs) -- The simplifier has put an up-to-date unfolding
- -- in the IdInfo, but the RHS will do just as well
-
- _ -> Nothing
+ CoreUnfolding { uf_tmpl = unf_rhs, uf_src = src, uf_guidance = guide }
+ | show_unfolding src guide
+ -> Just (unf_ext_ids src unf_rhs)
+ DFunUnfolding _ _ ops -> Just (exprsFvsInOrder (dfunArgExprs ops))
+ _ -> Nothing
+ where
+ unf_ext_ids (InlineWrapper v) _ = (unitVarSet v, [v])
+ unf_ext_ids _ unf_rhs = exprFvsInOrder unf_rhs
+ -- For a wrapper, externalise the wrapper id rather than the
+ -- fvs of the rhs. The two usually come down to the same thing
+ -- but I've seen cases where we had a wrapper id $w but a
+ -- rhs where $w had been inlined; see Trac #3922
+
+ show_unfolding unf_source unf_guidance
+ = expose_all -- 'expose_all' says to expose all
+ -- unfoldings willy-nilly
+
+ || isStableSource unf_source -- Always expose things whose
+ -- source is an inline rule
+
+ || not (bottoming_fn -- No need to inline bottom functions
+ || never_active -- Or ones that say not to
+ || loop_breaker -- Or that are loop breakers
+ || neverUnfoldGuidance unf_guidance)
+
+-- We want a deterministic free-variable list. exprFreeVars gives us
+-- a VarSet, which is in a non-deterministic order when converted to a
+-- list. Hence, here we define a free-variable finder that returns
+-- the free variables in the order that they are encountered.
+--
+-- Note [choosing external names]
+
+exprFvsInOrder :: CoreExpr -> (VarSet, [Id])
+exprFvsInOrder e = run (dffvExpr e)
+
+exprsFvsInOrder :: [CoreExpr] -> (VarSet, [Id])
+exprsFvsInOrder es = run (mapM_ dffvExpr es)
+
+run :: DFFV () -> (VarSet, [Id])
+run (DFFV m) = case m emptyVarSet [] of
+ (set,ids,_) -> (set,ids)
+
+newtype DFFV a = DFFV (VarSet -> [Var] -> (VarSet,[Var],a))
+
+instance Monad DFFV where
+ return a = DFFV $ \set ids -> (set, ids, a)
+ (DFFV m) >>= k = DFFV $ \set ids ->
+ case m set ids of
+ (set',ids',a) -> case k a of
+ DFFV f -> f set' ids'
+
+insert :: Var -> DFFV ()
+insert v = DFFV $ \ set ids -> case () of
+ _ | v `elemVarSet` set -> (set,ids,())
+ | otherwise -> (extendVarSet set v, v:ids, ())
+
+dffvExpr :: CoreExpr -> DFFV ()
+dffvExpr e = go emptyVarSet e
+ where
+ go scope e = case e of
+ Var v | isLocalId v && not (v `elemVarSet` scope) -> insert v
+ App e1 e2 -> do go scope e1; go scope e2
+ Lam v e -> go (extendVarSet scope v) e
+ Note _ e -> go scope e
+ Cast e _ -> go scope e
+ Let (NonRec x r) e -> do go scope r; go (extendVarSet scope x) e
+ Let (Rec prs) e -> do let scope' = extendVarSetList scope (map fst prs)
+ mapM_ (go scope') (map snd prs)
+ go scope' e
+ Case e b _ as -> do go scope e
+ mapM_ (go_alt (extendVarSet scope b)) as
+ _other -> return ()
+
+ go_alt scope (_,xs,r) = go (extendVarSetList scope xs) r
\end{code}
+--------------------------------------------------------------------
+-- tidyTopName
+-- 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
+-- we intend to externalise it.
+
+\begin{code}
+tidyTopName :: Module -> IORef NameCache -> Maybe Id -> TidyOccEnv
+ -> Id -> IO (TidyOccEnv, Name)
+tidyTopName mod nc_var maybe_ref occ_env id
+ | global && internal = return (occ_env, localiseName name)
+
+ | global && external = return (occ_env, name)
+ -- Global names are assumed to have been allocated by the renamer,
+ -- so they already have the "right" unique
+ -- And it's a system-wide unique too
+
+ -- Now we get to the real reason that all this is in the IO Monad:
+ -- we have to update the name cache in a nice atomic fashion
+
+ | local && internal = do { nc <- readIORef nc_var
+ ; let (nc', new_local_name) = mk_new_local nc
+ ; writeIORef nc_var nc'
+ ; return (occ_env', new_local_name) }
+ -- Even local, internal names must get a unique occurrence, because
+ -- if we do -split-objs we externalise the name later, in the code generator
+ --
+ -- Similarly, we must make sure it has a system-wide Unique, because
+ -- the byte-code generator builds a system-wide Name->BCO symbol table
+
+ | local && external = do { nc <- readIORef nc_var
+ ; let (nc', new_external_name) = mk_new_external nc
+ ; writeIORef nc_var nc'
+ ; return (occ_env', new_external_name) }
+
+ | otherwise = panic "tidyTopName"
+ where
+ name = idName id
+ external = isJust maybe_ref
+ global = isExternalName name
+ local = not global
+ internal = not external
+ loc = nameSrcSpan name
+
+ old_occ = nameOccName name
+ new_occ
+ | Just ref <- maybe_ref, ref /= id =
+ mkOccName (occNameSpace old_occ) $
+ let
+ ref_str = occNameString (getOccName ref)
+ occ_str = occNameString old_occ
+ in
+ case occ_str of
+ '$':'w':_ -> occ_str
+ -- workers: the worker for a function already
+ -- includes the occname for its parent, so there's
+ -- no need to prepend the referrer.
+ _other | isSystemName name -> ref_str
+ | otherwise -> ref_str ++ '_' : occ_str
+ -- If this name was system-generated, then don't bother
+ -- to retain its OccName, just use the referrer. These
+ -- system-generated names will become "f1", "f2", etc. for
+ -- a referrer "f".
+ | otherwise = old_occ
+
+ (occ_env', occ') = tidyOccName occ_env new_occ
+
+ mk_new_local nc = (nc { nsUniqs = us }, mkInternalName uniq occ' loc)
+ where
+ (uniq, us) = takeUniqFromSupply (nsUniqs nc)
+
+ mk_new_external nc = allocateGlobalBinder nc mod occ' loc
+ -- If we want to externalise a currently-local name, check
+ -- whether we have already assigned a unique for it.
+ -- If so, use it; if not, extend the table.
+ -- All this is done by allcoateGlobalBinder.
+ -- This is needed when *re*-compiling a module in GHCi; we must
+ -- use the same name for externally-visible things as we did before.
+\end{code}
+
\begin{code}
-findExternalRules :: [CoreBind]
- -> [CoreRule] -- Non-local rules (i.e. ones for imported fns)
- -> IdEnv a -- Ids that are exported, so we need their rules
+findExternalRules :: Bool -- Omit pragmas
+ -> [CoreBind]
+ -> [CoreRule] -- Local rules for imported fns
+ -> UnfoldEnv -- Ids that are exported, so we need their rules
-> [CoreRule]
-- The complete rules are gotten by combining
- -- a) the non-local rules
+ -- a) local rules for imported Ids
-- b) rules embedded in the top-level Ids
-findExternalRules binds non_local_rules ext_ids
- = filter (not . internal_rule) (non_local_rules ++ local_rules)
+findExternalRules omit_prags binds imp_id_rules unfold_env
+ | omit_prags = []
+ | otherwise = filterOut internal_rule (imp_id_rules ++ local_rules)
where
local_rules = [ rule
| id <- bindersOfBinds binds,
- id `elemVarEnv` ext_ids,
+ external_id id,
rule <- idCoreRules id
]
internal_rule rule
- = any internal_id (varSetElems (ruleLhsFreeIds rule))
+ = any (not . external_id) (varSetElems (ruleLhsFreeIds rule))
-- Don't export a rule whose LHS mentions a locally-defined
-- Id that is completely internal (i.e. not visible to an
-- importing module)
- internal_id id = not (id `elemVarEnv` ext_ids)
+ external_id id
+ | Just (name,_) <- lookupVarEnv unfold_env id = isExternalName name
+ | otherwise = False
\end{code}
-
+Note [Which rules to expose]
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+findExternalRules filters imp_rules to avoid binders that
+aren't externally visible; but the externally-visible binders
+are computed (by findExternalIds) assuming that all orphan
+rules are externalised (see init_ext_ids in function
+'search'). So in fact we may export more than we need.
+(It's a sort of mutual recursion.)
%************************************************************************
%* *
-- * subst_env: A Var->Var mapping that substitutes the new Var for the old
tidyTopBinds :: HscEnv
- -> Module
- -> TypeEnv
- -> ExtIdEnv
+ -> UnfoldEnv
+ -> TidyOccEnv
-> [CoreBind]
- -> IO (TidyEnv, [CoreBind])
+ -> (TidyEnv, [CoreBind])
-tidyTopBinds hsc_env mod type_env ext_ids binds
+tidyTopBinds hsc_env unfold_env init_occ_env binds
= tidy init_env binds
where
- nc_var = hsc_NC hsc_env
-
- -- We also make sure to avoid any exported binders. Consider
- -- f{-u1-} = 1 -- Local decl
- -- ...
- -- f{-u2-} = 2 -- Exported decl
- --
- -- The second exported decl must 'get' the name 'f', so we
- -- have to put 'f' in the avoids list before we get to the first
- -- decl. tidyTopId then does a no-op on exported binders.
- init_env = (initTidyOccEnv avoids, emptyVarEnv)
- avoids = [getOccName name | bndr <- typeEnvIds type_env,
- let name = idName bndr,
- isExternalName name]
- -- In computing our "avoids" list, we must include
- -- all implicit Ids
- -- all things with global names (assigned once and for
- -- all by the renamer)
- -- since their names are "taken".
- -- The type environment is a convenient source of such things.
+ init_env = (init_occ_env, emptyVarEnv)
this_pkg = thisPackage (hsc_dflags hsc_env)
- tidy env [] = return (env, [])
- tidy env (b:bs) = do { (env1, b') <- tidyTopBind this_pkg mod nc_var ext_ids env b
- ; (env2, bs') <- tidy env1 bs
- ; return (env2, b':bs') }
+ tidy env [] = (env, [])
+ tidy env (b:bs) = let (env1, b') = tidyTopBind this_pkg unfold_env env b
+ (env2, bs') = tidy env1 bs
+ in
+ (env2, b':bs')
------------------------
tidyTopBind :: PackageId
- -> Module
- -> IORef NameCache -- For allocating new unique names
- -> ExtIdEnv
- -> TidyEnv -> CoreBind
- -> IO (TidyEnv, CoreBind)
-
-tidyTopBind this_pkg mod nc_var ext_ids (occ_env1,subst1) (NonRec bndr rhs)
- = do { (occ_env2, name') <- tidyTopName mod nc_var ext_ids occ_env1 bndr
- ; let { (bndr', rhs') = tidyTopPair ext_ids tidy_env2 caf_info name' (bndr, rhs)
- ; subst2 = extendVarEnv subst1 bndr bndr'
- ; tidy_env2 = (occ_env2, subst2) }
- ; return (tidy_env2, NonRec bndr' rhs') }
+ -> UnfoldEnv
+ -> TidyEnv
+ -> CoreBind
+ -> (TidyEnv, CoreBind)
+
+tidyTopBind this_pkg unfold_env (occ_env,subst1) (NonRec bndr rhs)
+ = (tidy_env2, NonRec bndr' rhs')
where
- caf_info = hasCafRefs this_pkg subst1 (idArity bndr) rhs
-
-tidyTopBind this_pkg mod nc_var ext_ids (occ_env1,subst1) (Rec prs)
- = do { (occ_env2, names') <- tidyTopNames mod nc_var ext_ids occ_env1 bndrs
- ; let { prs' = zipWith (tidyTopPair ext_ids tidy_env2 caf_info)
- names' prs
- ; subst2 = extendVarEnvList subst1 (bndrs `zip` map fst prs')
- ; tidy_env2 = (occ_env2, subst2) }
- ; return (tidy_env2, Rec prs') }
+ Just (name',show_unfold) = lookupVarEnv unfold_env bndr
+ caf_info = hasCafRefs this_pkg subst1 (idArity bndr) rhs
+ (bndr', rhs') = tidyTopPair show_unfold tidy_env2 caf_info name' (bndr, rhs)
+ subst2 = extendVarEnv subst1 bndr bndr'
+ tidy_env2 = (occ_env, subst2)
+
+tidyTopBind this_pkg unfold_env (occ_env,subst1) (Rec prs)
+ = (tidy_env2, Rec prs')
where
+ prs' = [ tidyTopPair show_unfold tidy_env2 caf_info name' (id,rhs)
+ | (id,rhs) <- prs,
+ let (name',show_unfold) =
+ expectJust "tidyTopBind" $ lookupVarEnv unfold_env id
+ ]
+
+ subst2 = extendVarEnvList subst1 (bndrs `zip` map fst prs')
+ tidy_env2 = (occ_env, subst2)
+
bndrs = map fst prs
-- the CafInfo for a recursive group says whether *any* rhs in
| (bndr,rhs) <- prs ] = MayHaveCafRefs
| otherwise = NoCafRefs
---------------------------------------------------------------------
--- tidyTopName
--- 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
--- we intend to externalise it.
-tidyTopNames :: Module -> IORef NameCache -> VarEnv Bool -> TidyOccEnv
- -> [Id] -> IO (TidyOccEnv, [Name])
-tidyTopNames _mod _nc_var _ext_ids occ_env [] = return (occ_env, [])
-tidyTopNames mod nc_var ext_ids occ_env (id:ids)
- = do { (occ_env1, name) <- tidyTopName mod nc_var ext_ids occ_env id
- ; (occ_env2, names) <- tidyTopNames mod nc_var ext_ids occ_env1 ids
- ; return (occ_env2, name:names) }
-
-tidyTopName :: Module -> IORef NameCache -> VarEnv Bool -> TidyOccEnv
- -> Id -> IO (TidyOccEnv, Name)
-tidyTopName mod nc_var ext_ids occ_env id
- | global && internal = return (occ_env, localiseName name)
-
- | global && external = return (occ_env, name)
- -- Global names are assumed to have been allocated by the renamer,
- -- so they already have the "right" unique
- -- And it's a system-wide unique too
-
- -- Now we get to the real reason that all this is in the IO Monad:
- -- we have to update the name cache in a nice atomic fashion
-
- | local && internal = do { nc <- readIORef nc_var
- ; let (nc', new_local_name) = mk_new_local nc
- ; writeIORef nc_var nc'
- ; return (occ_env', new_local_name) }
- -- Even local, internal names must get a unique occurrence, because
- -- if we do -split-objs we externalise the name later, in the code generator
- --
- -- Similarly, we must make sure it has a system-wide Unique, because
- -- the byte-code generator builds a system-wide Name->BCO symbol table
-
- | local && external = do { nc <- readIORef nc_var
- ; let (nc', new_external_name) = mk_new_external nc
- ; writeIORef nc_var nc'
- ; return (occ_env', new_external_name) }
-
- | otherwise = panic "tidyTopName"
- where
- name = idName id
- external = id `elemVarEnv` ext_ids
- global = isExternalName name
- local = not global
- internal = not external
- loc = nameSrcSpan name
-
- (occ_env', occ') = tidyOccName occ_env (nameOccName name)
-
- mk_new_local nc = (nc { nsUniqs = us2 }, mkInternalName uniq occ' loc)
- where
- (us1, us2) = splitUniqSupply (nsUniqs nc)
- uniq = uniqFromSupply us1
-
- mk_new_external nc = allocateGlobalBinder nc mod occ' loc
- -- If we want to externalise a currently-local name, check
- -- whether we have already assigned a unique for it.
- -- If so, use it; if not, extend the table.
- -- All this is done by allcoateGlobalBinder.
- -- This is needed when *re*-compiling a module in GHCi; we must
- -- use the same name for externally-visible things as we did before.
-
-
-----------------------------------------------------------
-tidyTopPair :: VarEnv Bool
+tidyTopPair :: Bool -- show unfolding
-> TidyEnv -- The TidyEnv is used to tidy the IdInfo
-- It is knot-tied: don't look at it!
-> CafInfo
-- group, a variable late in the group might be mentioned
-- in the IdInfo of one early in the group
-tidyTopPair ext_ids rhs_tidy_env caf_info name' (bndr, rhs)
- = (bndr', rhs')
+tidyTopPair show_unfold rhs_tidy_env caf_info name' (bndr, rhs)
+ = (bndr1, rhs1)
where
- bndr' = mkGlobalId details name' ty' idinfo'
- -- Preserve the GlobalIdDetails of existing global-ids
- details = case globalIdDetails bndr of
- NotGlobalId -> VanillaGlobal
- old_details -> old_details
- ty' = tidyTopType (idType bndr)
- rhs' = tidyExpr rhs_tidy_env rhs
- idinfo = idInfo bndr
- idinfo' = tidyTopIdInfo (isJust maybe_external)
- idinfo unfold_info
- arity caf_info
-
- -- Expose an unfolding if ext_ids tells us to
- -- Remember that ext_ids maps an Id to a Bool:
- -- True to show the unfolding, False to hide it
- maybe_external = lookupVarEnv ext_ids bndr
- show_unfold = maybe_external `orElse` False
- unfold_info | show_unfold = tidyUnfolding rhs_tidy_env rhs' (unfoldingInfo idinfo)
- | otherwise = noUnfolding
- -- NB: do *not* expose the worker if show_unfold is off,
- -- because that means this thing is a loop breaker or
- -- marked NOINLINE or something like that
- -- This is important: if you expose the worker for a loop-breaker
- -- then you can make the simplifier go into an infinite loop, because
- -- in effect the unfolding is exposed. See Trac #1709
- --
- -- You might think that if show_unfold is False, then the thing should
- -- not be w/w'd in the first place. But a legitimate reason is this:
- -- the function returns bottom
- -- In this case, show_unfold will be false (we don't expose unfoldings
- -- for bottoming functions), but we might still have a worker/wrapper
- -- split (see Note [Worker-wrapper for bottoming functions] in WorkWrap.lhs
-
- -- Usually the Id will have an accurate arity on it, because
- -- the simplifier has just run, but not always.
- -- One case I found was when the last thing the simplifier
- -- did was to let-bind a non-atomic argument and then float
- -- it to the top level. So it seems more robust just to
- -- fix it here.
- arity = exprArity rhs
-
+ bndr1 = mkGlobalId details name' ty' idinfo'
+ details = idDetails bndr -- Preserve the IdDetails
+ ty' = tidyTopType (idType bndr)
+ rhs1 = tidyExpr rhs_tidy_env rhs
+ idinfo' = tidyTopIdInfo rhs_tidy_env name' rhs rhs1 (idInfo bndr)
+ show_unfold caf_info
-- tidyTopIdInfo creates the final IdInfo for top-level
-- binders. There are two delicate pieces:
-- occurrences of the binders in RHSs, and hence to occurrences in
-- unfoldings, which are inside Ids imported by GHCi. Ditto RULES.
-- CoreToStg makes use of this when constructing SRTs.
-tidyTopIdInfo :: Bool -> IdInfo -> Unfolding
- -> ArityInfo -> CafInfo
- -> IdInfo
-tidyTopIdInfo is_external idinfo unfold_info arity caf_info
+tidyTopIdInfo :: TidyEnv -> Name -> CoreExpr -> CoreExpr
+ -> IdInfo -> Bool -> CafInfo -> IdInfo
+tidyTopIdInfo rhs_tidy_env name orig_rhs tidy_rhs idinfo show_unfold caf_info
| not is_external -- For internal Ids (not externally visible)
= vanillaIdInfo -- we only need enough info for code generation
-- Arity and strictness info are enough;
-- c.f. CoreTidy.tidyLetBndr
- `setCafInfo` caf_info
- `setArityInfo` arity
- `setAllStrictnessInfo` newStrictnessInfo idinfo
+ `setCafInfo` caf_info
+ `setArityInfo` arity
+ `setStrictnessInfo` final_sig
| otherwise -- Externally-visible Ids get the whole lot
= vanillaIdInfo
`setCafInfo` caf_info
`setArityInfo` arity
- `setAllStrictnessInfo` newStrictnessInfo idinfo
- `setInlinePragInfo` inlinePragInfo idinfo
+ `setStrictnessInfo` final_sig
+ `setOccInfo` robust_occ_info
+ `setInlinePragInfo` (inlinePragInfo idinfo)
`setUnfoldingInfo` unfold_info
-- NB: we throw away the Rules
-- They have already been extracted by findExternalRules
+ where
+ is_external = isExternalName name
+
+ --------- OccInfo ------------
+ robust_occ_info = zapFragileOcc (occInfo idinfo)
+ -- It's important to keep loop-breaker information
+ -- when we are doing -fexpose-all-unfoldings
+
+ --------- Strictness ------------
+ final_sig | Just sig <- strictnessInfo idinfo
+ = WARN( _bottom_hidden sig, ppr name ) Just sig
+ | Just (_, sig) <- mb_bot_str = Just sig
+ | otherwise = Nothing
+
+ -- If the cheap-and-cheerful bottom analyser can see that
+ -- the RHS is bottom, it should jolly well be exposed
+ _bottom_hidden id_sig = case mb_bot_str of
+ Nothing -> False
+ Just (arity, _) -> not (appIsBottom id_sig arity)
+
+ mb_bot_str = exprBotStrictness_maybe orig_rhs
+
+ --------- Unfolding ------------
+ unf_info = unfoldingInfo idinfo
+ unfold_info | show_unfold = tidyUnfolding rhs_tidy_env unf_info unf_from_rhs
+ | otherwise = noUnfolding
+ unf_from_rhs = mkTopUnfolding is_bot tidy_rhs
+ is_bot = case final_sig of
+ Just sig -> isBottomingSig sig
+ Nothing -> False
+ -- NB: do *not* expose the worker if show_unfold is off,
+ -- because that means this thing is a loop breaker or
+ -- marked NOINLINE or something like that
+ -- This is important: if you expose the worker for a loop-breaker
+ -- then you can make the simplifier go into an infinite loop, because
+ -- in effect the unfolding is exposed. See Trac #1709
+ --
+ -- You might think that if show_unfold is False, then the thing should
+ -- not be w/w'd in the first place. But a legitimate reason is this:
+ -- the function returns bottom
+ -- In this case, show_unfold will be false (we don't expose unfoldings
+ -- for bottoming functions), but we might still have a worker/wrapper
+ -- split (see Note [Worker-wrapper for bottoming functions] in WorkWrap.lhs
-
-
------------- Unfolding --------------
-tidyUnfolding :: TidyEnv -> CoreExpr -> Unfolding -> Unfolding
-tidyUnfolding tidy_env _ unf@(InlineRule { uf_tmpl = rhs, uf_worker = mb_wkr })
- = unf { uf_tmpl = tidyExpr tidy_env rhs,
- uf_worker = fmap (tidyVarOcc tidy_env) mb_wkr }
-tidyUnfolding _ tidy_rhs (CoreUnfolding {})
- = mkTopUnfolding tidy_rhs
-tidyUnfolding _ _ unf = unf
+ --------- Arity ------------
+ -- Usually the Id will have an accurate arity on it, because
+ -- the simplifier has just run, but not always.
+ -- One case I found was when the last thing the simplifier
+ -- did was to let-bind a non-atomic argument and then float
+ -- it to the top level. So it seems more robust just to
+ -- fix it here.
+ arity = exprArity orig_rhs
\end{code}
%************************************************************************
\begin{code}
hasCafRefs :: PackageId -> VarEnv Var -> Arity -> CoreExpr -> CafInfo
hasCafRefs this_pkg p arity expr
- | is_caf || mentions_cafs
- = MayHaveCafRefs
+ | is_caf || mentions_cafs = MayHaveCafRefs
| otherwise = NoCafRefs
where
mentions_cafs = isFastTrue (cafRefs p expr)
- is_caf = not (arity > 0 || rhsIsStatic this_pkg expr)
+ is_dynamic_name = isDllName this_pkg
+ is_caf = not (arity > 0 || rhsIsStatic is_dynamic_name expr)
-- NB. we pass in the arity of the expression, which is expected
-- to be calculated by exprArity. This is because exprArity
cafRefs p (Note _n e) = cafRefs p e
cafRefs p (Cast e _co) = cafRefs p e
cafRefs _ (Type _) = fastBool False
+cafRefs _ (Coercion _) = fastBool False
cafRefss :: VarEnv Id -> [Expr a] -> FastBool
cafRefss _ [] = fastBool False
projects / ghc-hetmet.git / blobdiff