StgCmmClosure.hs 37.9 KB
Newer Older
1 2
{-# LANGUAGE CPP, RecordWildCards #-}

3 4 5
-----------------------------------------------------------------------------
--
-- Stg to C-- code generation:
6
--
7 8 9 10 11 12 13 14
-- The types   LambdaFormInfo
--             ClosureInfo
--
-- Nothing monadic in here!
--
-----------------------------------------------------------------------------

module StgCmmClosure (
Simon Marlow's avatar
Simon Marlow committed
15
        DynTag,  tagForCon, isSmallFamily,
16

17
        idPrimRep, isVoidRep, isGcPtrRep, addIdReps, addArgReps,
18
        argPrimRep,
19

20 21 22
        NonVoid(..), fromNonVoid, nonVoidIds, nonVoidStgArgs,
        assertNonVoidIds, assertNonVoidStgArgs,

23 24
        -- * LambdaFormInfo
        LambdaFormInfo,         -- Abstract
25 26 27
        StandardFormInfo,        -- ...ditto...
        mkLFThunk, mkLFReEntrant, mkConLFInfo, mkSelectorLFInfo,
        mkApLFInfo, mkLFImported, mkLFArgument, mkLFLetNoEscape,
28
        mkLFStringLit,
Simon Marlow's avatar
Simon Marlow committed
29
        lfDynTag,
Gabor Greif's avatar
Gabor Greif committed
30
        isLFThunk, isLFReEntrant, lfUpdatable,
31

32 33 34
        -- * Used by other modules
        CgLoc(..), SelfLoopInfo, CallMethod(..),
        nodeMustPointToIt, isKnownFun, funTag, tagForArity, getCallMethod,
35 36

        -- * ClosureInfo
37
        ClosureInfo,
Simon Marlow's avatar
Simon Marlow committed
38
        mkClosureInfo,
39
        mkCmmInfo,
40

41 42
        -- ** Inspection
        closureLFInfo, closureName,
Simon Marlow's avatar
Simon Marlow committed
43

44 45 46
        -- ** Labels
        -- These just need the info table label
        closureInfoLabel, staticClosureLabel,
nfrisby's avatar
nfrisby committed
47
        closureSlowEntryLabel, closureLocalEntryLabel,
48

49 50
        -- ** Predicates
        -- These are really just functions on LambdaFormInfo
Simon Marlow's avatar
Simon Marlow committed
51
        closureUpdReqd, closureSingleEntry,
52 53
        closureReEntrant, closureFunInfo,
        isToplevClosure,
54

55 56
        blackHoleOnEntry,  -- Needs LambdaFormInfo and SMRep
        isStaticClosure,   -- Needs SMPre
57

58
        -- * InfoTables
59
        mkDataConInfoTable,
60
        cafBlackHoleInfoTable,
61
        indStaticInfoTable,
62
        staticClosureNeedsLink,
63 64 65 66 67 68
    ) where

#include "../includes/MachDeps.h"

#include "HsVersions.h"

69 70
import GhcPrelude

71 72
import StgSyn
import SMRep
73
import Cmm
74
import PprCmmExpr() -- For Outputable instances
75

76
import CostCentre
77
import BlockId
78 79 80 81 82 83
import CLabel
import Id
import IdInfo
import DataCon
import Name
import Type
84
import TyCoRep
85 86
import TcType
import TyCon
87
import RepType
88 89
import BasicTypes
import Outputable
90
import DynFlags
91
import Util
92

93
import Data.Coerce (coerce)
94
import qualified Data.ByteString.Char8 as BS8
95

96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114
-----------------------------------------------------------------------------
--                Data types and synonyms
-----------------------------------------------------------------------------

-- These data types are mostly used by other modules, especially StgCmmMonad,
-- but we define them here because some functions in this module need to
-- have access to them as well

data CgLoc
  = CmmLoc CmmExpr      -- A stable CmmExpr; that is, one not mentioning
                        -- Hp, so that it remains valid across calls

  | LneLoc BlockId [LocalReg]             -- A join point
        -- A join point (= let-no-escape) should only
        -- be tail-called, and in a saturated way.
        -- To tail-call it, assign to these locals,
        -- and branch to the block id

instance Outputable CgLoc where
115 116
  ppr (CmmLoc e)    = text "cmm" <+> ppr e
  ppr (LneLoc b rs) = text "lne" <+> ppr b <+> ppr rs
117 118 119 120 121

type SelfLoopInfo = (Id, BlockId, [LocalReg])

-- used by ticky profiling
isKnownFun :: LambdaFormInfo -> Bool
122 123 124
isKnownFun LFReEntrant{} = True
isKnownFun LFLetNoEscape = True
isKnownFun _             = False
125 126


127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162
-------------------------------------
--        Non-void types
-------------------------------------
-- We frequently need the invariant that an Id or a an argument
-- is of a non-void type. This type is a witness to the invariant.

newtype NonVoid a = NonVoid a
  deriving (Eq, Show)

fromNonVoid :: NonVoid a -> a
fromNonVoid (NonVoid a) = a

instance (Outputable a) => Outputable (NonVoid a) where
  ppr (NonVoid a) = ppr a

nonVoidIds :: [Id] -> [NonVoid Id]
nonVoidIds ids = [NonVoid id | id <- ids, not (isVoidTy (idType id))]

-- | Used in places where some invariant ensures that all these Ids are
-- non-void; e.g. constructor field binders in case expressions.
-- See Note [Post-unarisation invariants] in UnariseStg.
assertNonVoidIds :: [Id] -> [NonVoid Id]
assertNonVoidIds ids = ASSERT(not (any (isVoidTy . idType) ids))
                       coerce ids

nonVoidStgArgs :: [StgArg] -> [NonVoid StgArg]
nonVoidStgArgs args = [NonVoid arg | arg <- args, not (isVoidTy (stgArgType arg))]

-- | Used in places where some invariant ensures that all these arguments are
-- non-void; e.g. constructor arguments.
-- See Note [Post-unarisation invariants] in UnariseStg.
assertNonVoidStgArgs :: [StgArg] -> [NonVoid StgArg]
assertNonVoidStgArgs args = ASSERT(not (any (isVoidTy . stgArgType) args))
                            coerce args


163
-----------------------------------------------------------------------------
164
--                Representations
165 166
-----------------------------------------------------------------------------

167 168
-- Why are these here?

169
idPrimRep :: Id -> PrimRep
170 171
idPrimRep id = typePrimRep1 (idType id)
    -- NB: typePrimRep1 fails on unboxed tuples,
172
    --     but by StgCmm no Ids have unboxed tuple type
173
    -- See also Note [VoidRep] in RepType
174

175 176 177
addIdReps :: [NonVoid Id] -> [NonVoid (PrimRep, Id)]
addIdReps = map (\id -> let id' = fromNonVoid id
                         in NonVoid (idPrimRep id', id'))
178

179 180 181
addArgReps :: [NonVoid StgArg] -> [NonVoid (PrimRep, StgArg)]
addArgReps = map (\arg -> let arg' = fromNonVoid arg
                           in NonVoid (argPrimRep arg', arg'))
182 183

argPrimRep :: StgArg -> PrimRep
184
argPrimRep arg = typePrimRep1 (stgArgType arg)
185 186 187


-----------------------------------------------------------------------------
188
--                LambdaFormInfo
189 190 191 192 193 194 195 196
-----------------------------------------------------------------------------

-- Information about an identifier, from the code generator's point of
-- view.  Every identifier is bound to a LambdaFormInfo in the
-- environment, which gives the code generator enough info to be able to
-- tail call or return that identifier.

data LambdaFormInfo
197 198
  = LFReEntrant         -- Reentrant closure (a function)
        TopLevelFlag    -- True if top level
199
        OneShotInfo
200 201
        !RepArity       -- Arity. Invariant: always > 0
        !Bool           -- True <=> no fvs
202 203
        ArgDescr        -- Argument descriptor (should really be in ClosureInfo)

204
  | LFThunk             -- Thunk (zero arity)
205
        TopLevelFlag
206 207
        !Bool           -- True <=> no free vars
        !Bool           -- True <=> updatable (i.e., *not* single-entry)
208
        StandardFormInfo
209
        !Bool           -- True <=> *might* be a function type
210

211 212
  | LFCon               -- A saturated constructor application
        DataCon         -- The constructor
213

214
  | LFUnknown           -- Used for function arguments and imported things.
215 216 217 218 219
                        -- We know nothing about this closure.
                        -- Treat like updatable "LFThunk"...
                        -- Imported things which we *do* know something about use
                        -- one of the other LF constructors (eg LFReEntrant for
                        -- known functions)
220
        !Bool           -- True <=> *might* be a function type
221 222 223 224
                        --      The False case is good when we want to enter it,
                        --        because then we know the entry code will do
                        --        For a function, the entry code is the fast entry point

225
  | LFUnlifted          -- A value of unboxed type;
226 227
                        -- always a value, needs evaluation

228
  | LFLetNoEscape       -- See LetNoEscape module for precise description
229

230 231

-------------------------
232
-- StandardFormInfo tells whether this thunk has one of
233 234 235 236
-- a small number of standard forms

data StandardFormInfo
  = NonStandardThunk
237
        -- The usual case: not of the standard forms
238 239

  | SelectorThunk
240 241 242 243
        -- A SelectorThunk is of form
        --      case x of
        --           con a1,..,an -> ak
        -- and the constructor is from a single-constr type.
244
       WordOff          -- 0-origin offset of ak within the "goods" of
245 246 247 248 249 250 251 252 253 254
                        -- constructor (Recall that the a1,...,an may be laid
                        -- out in the heap in a non-obvious order.)

  | ApThunk
        -- An ApThunk is of form
        --        x1 ... xn
        -- The code for the thunk just pushes x2..xn on the stack and enters x1.
        -- There are a few of these (for 1 <= n <= MAX_SPEC_AP_SIZE) pre-compiled
        -- in the RTS to save space.
        RepArity                -- Arity, n
255 256 257


------------------------------------------------------
258
--                Building LambdaFormInfo
259 260 261
------------------------------------------------------

mkLFArgument :: Id -> LambdaFormInfo
262
mkLFArgument id
263
  | isUnliftedType ty      = LFUnlifted
264
  | might_be_a_function ty = LFUnknown True
265
  | otherwise              = LFUnknown False
266 267 268 269 270 271 272 273
  where
    ty = idType id

-------------
mkLFLetNoEscape :: LambdaFormInfo
mkLFLetNoEscape = LFLetNoEscape

-------------
274 275 276
mkLFReEntrant :: TopLevelFlag    -- True of top level
              -> [Id]            -- Free vars
              -> [Id]            -- Args
277 278
              -> ArgDescr        -- Argument descriptor
              -> LambdaFormInfo
279

280 281
mkLFReEntrant _ _ [] _
  = pprPanic "mkLFReEntrant" empty
282
mkLFReEntrant top fvs args arg_descr
283 284
  = LFReEntrant top os_info (length args) (null fvs) arg_descr
  where os_info = idOneShotInfo (head args)
285 286 287 288

-------------
mkLFThunk :: Type -> TopLevelFlag -> [Id] -> UpdateFlag -> LambdaFormInfo
mkLFThunk thunk_ty top fvs upd_flag
289
  = ASSERT( not (isUpdatable upd_flag) || not (isUnliftedType thunk_ty) )
290 291 292 293
    LFThunk top (null fvs)
            (isUpdatable upd_flag)
            NonStandardThunk
            (might_be_a_function thunk_ty)
294 295 296 297 298 299

--------------
might_be_a_function :: Type -> Bool
-- Return False only if we are *sure* it's a data type
-- Look through newtypes etc as much as poss
might_be_a_function ty
300 301
  | [LiftedRep] <- typePrimRep ty
  , Just tc <- tyConAppTyCon_maybe (unwrapType ty)
302 303 304 305
  , isDataTyCon tc
  = False
  | otherwise
  = True
306 307 308 309 310 311 312 313

-------------
mkConLFInfo :: DataCon -> LambdaFormInfo
mkConLFInfo con = LFCon con

-------------
mkSelectorLFInfo :: Id -> Int -> Bool -> LambdaFormInfo
mkSelectorLFInfo id offset updatable
314 315
  = LFThunk NotTopLevel False updatable (SelectorThunk offset)
        (might_be_a_function (idType id))
316 317 318 319 320

-------------
mkApLFInfo :: Id -> UpdateFlag -> Arity -> LambdaFormInfo
mkApLFInfo id upd_flag arity
  = LFThunk NotTopLevel (arity == 0) (isUpdatable upd_flag) (ApThunk arity)
321
        (might_be_a_function (idType id))
322 323 324 325 326 327

-------------
mkLFImported :: Id -> LambdaFormInfo
mkLFImported id
  | Just con <- isDataConWorkId_maybe id
  , isNullaryRepDataCon con
328
  = LFCon con   -- An imported nullary constructor
329 330
                -- We assume that the constructor is evaluated so that
                -- the id really does point directly to the constructor
331 332

  | arity > 0
333
  = LFReEntrant TopLevel noOneShotInfo arity True (panic "arg_descr")
334 335 336 337

  | otherwise
  = mkLFArgument id -- Not sure of exact arity
  where
338
    arity = idFunRepArity id
339

340 341 342 343
-------------
mkLFStringLit :: LambdaFormInfo
mkLFStringLit = LFUnlifted

344
-----------------------------------------------------
345
--                Dynamic pointer tagging
346 347
-----------------------------------------------------

348 349
type DynTag = Int       -- The tag on a *pointer*
                        -- (from the dynamic-tagging paper)
350

351 352 353 354 355 356 357 358 359 360 361 362
-- Note [Data constructor dynamic tags]
-- ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
--
-- The family size of a data type (the number of constructors
-- or the arity of a function) can be either:
--    * small, if the family size < 2**tag_bits
--    * big, otherwise.
--
-- Small families can have the constructor tag in the tag bits.
-- Big families only use the tag value 1 to represent evaluatedness.
-- We don't have very many tag bits: for example, we have 2 bits on
-- x86-32 and 3 bits on x86-64.
363

364 365
isSmallFamily :: DynFlags -> Int -> Bool
isSmallFamily dflags fam_size = fam_size <= mAX_PTR_TAG dflags
366

367 368
tagForCon :: DynFlags -> DataCon -> DynTag
tagForCon dflags con
369 370
  | isSmallFamily dflags fam_size = con_tag
  | otherwise                     = 1
371
  where
372
    con_tag  = dataConTag con -- NB: 1-indexed
373
    fam_size = tyConFamilySize (dataConTyCon con)
374

375 376 377 378
tagForArity :: DynFlags -> RepArity -> DynTag
tagForArity dflags arity
 | isSmallFamily dflags arity = arity
 | otherwise                  = 0
379

380
lfDynTag :: DynFlags -> LambdaFormInfo -> DynTag
381 382
-- Return the tag in the low order bits of a variable bound
-- to this LambdaForm
383 384 385
lfDynTag dflags (LFCon con)                 = tagForCon dflags con
lfDynTag dflags (LFReEntrant _ _ arity _ _) = tagForArity dflags arity
lfDynTag _      _other                      = 0
386 387 388


-----------------------------------------------------------------------------
389
--                Observing LambdaFormInfo
390 391 392 393
-----------------------------------------------------------------------------

------------
isLFThunk :: LambdaFormInfo -> Bool
394
isLFThunk (LFThunk {})  = True
395 396
isLFThunk _ = False

397 398 399
isLFReEntrant :: LambdaFormInfo -> Bool
isLFReEntrant (LFReEntrant {}) = True
isLFReEntrant _                = False
400 401

-----------------------------------------------------------------------------
402
--                Choosing SM reps
403 404
-----------------------------------------------------------------------------

405
lfClosureType :: LambdaFormInfo -> ClosureTypeInfo
406 407
lfClosureType (LFReEntrant _ _ arity _ argd) = Fun arity argd
lfClosureType (LFCon con)                    = Constr (dataConTagZ con)
408
                                                      (dataConIdentity con)
409 410
lfClosureType (LFThunk _ _ _ is_sel _)       = thunkClosureType is_sel
lfClosureType _                              = panic "lfClosureType"
411

412 413 414
thunkClosureType :: StandardFormInfo -> ClosureTypeInfo
thunkClosureType (SelectorThunk off) = ThunkSelector off
thunkClosureType _                   = Thunk
415 416 417 418 419 420 421

-- We *do* get non-updatable top-level thunks sometimes.  eg. f = g
-- gets compiled to a jump to g (if g has non-zero arity), instead of
-- messing around with update frames and PAPs.  We set the closure type
-- to FUN_STATIC in this case.

-----------------------------------------------------------------------------
422
--                nodeMustPointToIt
423 424
-----------------------------------------------------------------------------

425
nodeMustPointToIt :: DynFlags -> LambdaFormInfo -> Bool
426
-- If nodeMustPointToIt is true, then the entry convention for
427
-- this closure has R1 (the "Node" register) pointing to the
428 429
-- closure itself --- the "self" argument

430
nodeMustPointToIt _ (LFReEntrant top _ _ no_fvs _)
431 432
  =  not no_fvs          -- Certainly if it has fvs we need to point to it
  || isNotTopLevel top   -- See Note [GC recovery]
433 434 435
        -- For lex_profiling we also access the cost centre for a
        -- non-inherited (i.e. non-top-level) function.
        -- The isNotTopLevel test above ensures this is ok.
436 437 438 439 440 441

nodeMustPointToIt dflags (LFThunk top no_fvs updatable NonStandardThunk _)
  =  not no_fvs            -- Self parameter
  || isNotTopLevel top     -- Note [GC recovery]
  || updatable             -- Need to push update frame
  || gopt Opt_SccProfilingOn dflags
442 443 444 445 446 447
          -- For the non-updatable (single-entry case):
          --
          -- True if has fvs (in which case we need access to them, and we
          --                    should black-hole it)
          -- or profiling (in which case we need to recover the cost centre
          --                 from inside it)  ToDo: do we need this even for
448 449
          --                                    top-level thunks? If not,
          --                                    isNotTopLevel subsumes this
450

451 452
nodeMustPointToIt _ (LFThunk {})        -- Node must point to a standard-form thunk
  = True
453

454
nodeMustPointToIt _ (LFCon _) = True
455

456 457 458 459 460 461 462 463 464 465
        -- Strictly speaking, the above two don't need Node to point
        -- to it if the arity = 0.  But this is a *really* unlikely
        -- situation.  If we know it's nil (say) and we are entering
        -- it. Eg: let x = [] in x then we will certainly have inlined
        -- x, since nil is a simple atom.  So we gain little by not
        -- having Node point to known zero-arity things.  On the other
        -- hand, we do lose something; Patrick's code for figuring out
        -- when something has been updated but not entered relies on
        -- having Node point to the result of an update.  SLPJ
        -- 27/11/92.
466

467
nodeMustPointToIt _ (LFUnknown _)   = True
468
nodeMustPointToIt _ LFUnlifted      = False
469
nodeMustPointToIt _ LFLetNoEscape   = False
470

471 472 473 474 475 476
{- Note [GC recovery]
~~~~~~~~~~~~~~~~~~~~~
If we a have a local let-binding (function or thunk)
   let f = <body> in ...
AND <body> allocates, then the heap-overflow check needs to know how
to re-start the evaluation.  It uses the "self" pointer to do this.
477
So even if there are no free variables in <body>, we still make
478 479 480 481 482 483 484
nodeMustPointToIt be True for non-top-level bindings.

Why do any such bindings exist?  After all, let-floating should have
floated them out.  Well, a clever optimiser might leave one there to
avoid a space leak, deliberately recomputing a thunk.  Also (and this
really does happen occasionally) let-floating may make a function f smaller
so it can be inlined, so now (f True) may generate a local no-fv closure.
485
This actually happened during bootstrapping GHC itself, with f=mkRdrFunBind
486 487
in TcGenDeriv.) -}

488
-----------------------------------------------------------------------------
489
--                getCallMethod
490 491 492 493 494 495
-----------------------------------------------------------------------------

{- The entry conventions depend on the type of closure being entered,
whether or not it has free variables, and whether we're running
sequentially or in parallel.

496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516
Closure                           Node   Argument   Enter
Characteristics              Par   Req'd  Passing    Via
---------------------------------------------------------------------------
Unknown                     & no  & yes & stack     & node
Known fun (>1 arg), no fvs  & no  & no  & registers & fast entry (enough args)
                                                    & slow entry (otherwise)
Known fun (>1 arg), fvs     & no  & yes & registers & fast entry (enough args)
0 arg, no fvs \r,\s         & no  & no  & n/a       & direct entry
0 arg, no fvs \u            & no  & yes & n/a       & node
0 arg, fvs \r,\s,selector   & no  & yes & n/a       & node
0 arg, fvs \r,\s            & no  & yes & n/a       & direct entry
0 arg, fvs \u               & no  & yes & n/a       & node
Unknown                     & yes & yes & stack     & node
Known fun (>1 arg), no fvs  & yes & no  & registers & fast entry (enough args)
                                                    & slow entry (otherwise)
Known fun (>1 arg), fvs     & yes & yes & registers & node
0 arg, fvs \r,\s,selector   & yes & yes & n/a       & node
0 arg, no fvs \r,\s         & yes & no  & n/a       & direct entry
0 arg, no fvs \u            & yes & yes & n/a       & node
0 arg, fvs \r,\s            & yes & yes & n/a       & node
0 arg, fvs \u               & yes & yes & n/a       & node
517 518 519 520 521

When black-holing, single-entry closures could also be entered via node
(rather than directly) to catch double-entry. -}

data CallMethod
522
  = EnterIt             -- No args, not a function
523

524
  | JumpToIt BlockId [LocalReg] -- A join point or a header of a local loop
525

526
  | ReturnIt            -- It's a value (function, unboxed value,
527
                        -- or constructor), so just return it.
528

529 530
  | SlowCall                -- Unknown fun, or known fun with
                        -- too few args.
531

532
  | DirectEntry         -- Jump directly, with args in regs
533 534
        CLabel          --   The code label
        RepArity        --   Its arity
535

536 537
getCallMethod :: DynFlags
              -> Name           -- Function being applied
538 539 540 541 542
              -> Id             -- Function Id used to chech if it can refer to
                                -- CAF's and whether the function is tail-calling
                                -- itself
              -> LambdaFormInfo -- Its info
              -> RepArity       -- Number of available arguments
543
              -> RepArity       -- Number of them being void arguments
544 545 546 547 548 549
              -> CgLoc          -- Passed in from cgIdApp so that we can
                                -- handle let-no-escape bindings and self-recursive
                                -- tail calls using the same data constructor,
                                -- JumpToIt. This saves us one case branch in
                                -- cgIdApp
              -> Maybe SelfLoopInfo -- can we perform a self-recursive tail call?
550
              -> CallMethod
551

552 553 554 555
getCallMethod dflags _ id _ n_args v_args _cg_loc
              (Just (self_loop_id, block_id, args))
  | gopt Opt_Loopification dflags
  , id == self_loop_id
556
  , args `lengthIs` (n_args - v_args)
557
  -- If these patterns match then we know that:
558
  --   * loopification optimisation is turned on
559
  --   * function is performing a self-recursive call in a tail position
560 561 562
  --   * number of non-void parameters of the function matches functions arity.
  -- See Note [Self-recursive tail calls] and Note [Void arguments in
  -- self-recursive tail calls] in StgCmmExpr for more details
563 564
  = JumpToIt block_id args

565
getCallMethod dflags name id (LFReEntrant _ _ arity _ _) n_args _v_args _cg_loc
566
              _self_loop_info
567 568 569 570
  | n_args == 0 -- No args at all
  && not (gopt Opt_SccProfilingOn dflags)
     -- See Note [Evaluating functions with profiling] in rts/Apply.cmm
  = ASSERT( arity /= 0 ) ReturnIt
571
  | n_args < arity = SlowCall        -- Not enough args
572
  | otherwise      = DirectEntry (enterIdLabel dflags name (idCafInfo id)) arity
573

574
getCallMethod _ _name _ LFUnlifted n_args _v_args _cg_loc _self_loop_info
575 576
  = ASSERT( n_args == 0 ) ReturnIt

577
getCallMethod _ _name _ (LFCon _) n_args _v_args _cg_loc _self_loop_info
578
  = ASSERT( n_args == 0 ) ReturnIt
579 580
    -- n_args=0 because it'd be ill-typed to apply a saturated
    --          constructor application to anything
581

582
getCallMethod dflags name id (LFThunk _ _ updatable std_form_info is_fun)
583
              n_args _v_args _cg_loc _self_loop_info
584 585
  | is_fun      -- it *might* be a function, so we must "call" it (which is always safe)
  = SlowCall    -- We cannot just enter it [in eval/apply, the entry code
586
                -- is the fast-entry code]
587 588

  -- Since is_fun is False, we are *definitely* looking at a data value
ian@well-typed.com's avatar
ian@well-typed.com committed
589
  | updatable || gopt Opt_Ticky dflags -- to catch double entry
590
      {- OLD: || opt_SMP
591 592 593 594
         I decided to remove this, because in SMP mode it doesn't matter
         if we enter the same thunk multiple times, so the optimisation
         of jumping directly to the entry code is still valid.  --SDM
        -}
595
  = EnterIt
596 597 598 599 600 601

  -- even a non-updatable selector thunk can be updated by the garbage
  -- collector, so we must enter it. (#8817)
  | SelectorThunk{} <- std_form_info
  = EnterIt

602 603 604 605 606 607 608
    -- We used to have ASSERT( n_args == 0 ), but actually it is
    -- possible for the optimiser to generate
    --   let bot :: Int = error Int "urk"
    --   in (bot `cast` unsafeCoerce Int (Int -> Int)) 3
    -- This happens as a result of the case-of-error transformation
    -- So the right thing to do is just to enter the thing

609
  | otherwise        -- Jump direct to code for single-entry thunks
610
  = ASSERT( n_args == 0 )
611 612
    DirectEntry (thunkEntryLabel dflags name (idCafInfo id) std_form_info
                updatable) 0
613

614
getCallMethod _ _name _ (LFUnknown True) _n_arg _v_args _cg_locs _self_loop_info
615 616
  = SlowCall -- might be a function

617
getCallMethod _ name _ (LFUnknown False) n_args _v_args _cg_loc _self_loop_info
618
  = ASSERT2( n_args == 0, ppr name <+> ppr n_args )
619 620
    EnterIt -- Not a function

621
getCallMethod _ _name _ LFLetNoEscape _n_args _v_args (LneLoc blk_id lne_regs)
622
              _self_loop_info
623
  = JumpToIt blk_id lne_regs
624

625
getCallMethod _ _ _ _ _ _ _ _ = panic "Unknown call method"
626 627

-----------------------------------------------------------------------------
628
--              Data types for closure information
629 630 631
-----------------------------------------------------------------------------


632
{- ClosureInfo: information about a binding
633

634 635 636
   We make a ClosureInfo for each let binding (both top level and not),
   but not bindings for data constructors: for those we build a CmmInfoTable
   directly (see mkDataConInfoTable).
637

638 639
   To a first approximation:
       ClosureInfo = (LambdaFormInfo, CmmInfoTable)
Simon Marlow's avatar
Simon Marlow committed
640

641 642 643 644
   A ClosureInfo has enough information
     a) to construct the info table itself, and build other things
        related to the binding (e.g. slow entry points for a function)
     b) to allocate a closure containing that info pointer (i.e.
645
           it knows the info table label)
646 647 648 649
-}

data ClosureInfo
  = ClosureInfo {
650 651 652 653 654 655 656 657 658 659 660 661
        closureName :: !Name,           -- The thing bound to this closure
           -- we don't really need this field: it's only used in generating
           -- code for ticky and profiling, and we could pass the information
           -- around separately, but it doesn't do much harm to keep it here.

        closureLFInfo :: !LambdaFormInfo, -- NOTE: not an LFCon
          -- this tells us about what the closure contains: it's right-hand-side.

          -- the rest is just an unpacked CmmInfoTable.
        closureInfoLabel :: !CLabel,
        closureSMRep     :: !SMRep,          -- representation used by storage mgr
        closureProf      :: !ProfilingInfo
662 663
    }

Simon Marlow's avatar
Simon Marlow committed
664
-- | Convert from 'ClosureInfo' to 'CmmInfoTable'.
665 666
mkCmmInfo :: ClosureInfo -> Id -> CostCentreStack -> CmmInfoTable
mkCmmInfo ClosureInfo {..} id ccs
667 668 669
  = CmmInfoTable { cit_lbl  = closureInfoLabel
                 , cit_rep  = closureSMRep
                 , cit_prof = closureProf
670 671 672 673
                 , cit_srt  = Nothing
                 , cit_clo  = if isStaticRep closureSMRep
                                then Just (id,ccs)
                                else Nothing }
Simon Marlow's avatar
Simon Marlow committed
674

675
--------------------------------------
676
--        Building ClosureInfos
677 678
--------------------------------------

679
mkClosureInfo :: DynFlags
680 681 682 683
              -> Bool                -- Is static
              -> Id
              -> LambdaFormInfo
              -> Int -> Int        -- Total and pointer words
Simon Marlow's avatar
Simon Marlow committed
684
              -> String         -- String descriptor
685
              -> ClosureInfo
686
mkClosureInfo dflags is_static id lf_info tot_wds ptr_wds val_descr
687 688 689 690 691
  = ClosureInfo { closureName      = name
                , closureLFInfo    = lf_info
                , closureInfoLabel = info_lbl   -- These three fields are
                , closureSMRep     = sm_rep     -- (almost) an info table
                , closureProf      = prof }     -- (we don't have an SRT yet)
692
  where
Simon Marlow's avatar
Simon Marlow committed
693
    name       = idName id
694
    sm_rep     = mkHeapRep dflags is_static ptr_wds nonptr_wds (lfClosureType lf_info)
695
    prof       = mkProfilingInfo dflags id val_descr
696
    nonptr_wds = tot_wds - ptr_wds
697

698
    info_lbl = mkClosureInfoTableLabel id lf_info
699 700

--------------------------------------
701
--   Other functions over ClosureInfo
702 703
--------------------------------------

704 705 706
-- Eager blackholing is normally disabled, but can be turned on with
-- -feager-blackholing.  When it is on, we replace the info pointer of
-- the thunk with stg_EAGER_BLACKHOLE_info on entry.
707

708 709 710 711 712 713 714
-- If we wanted to do eager blackholing with slop filling,
-- we'd need to do it at the *end* of a basic block, otherwise
-- we overwrite the free variables in the thunk that we still
-- need.  We have a patch for this from Andy Cheadle, but not
-- incorporated yet. --SDM [6/2004]
--
-- Previously, eager blackholing was enabled when ticky-ticky
715 716
-- was on. But it didn't work, and it wasn't strictly necessary
-- to bring back minimal ticky-ticky, so now EAGER_BLACKHOLING
717
-- is unconditionally disabled. -- krc 1/2007
718
--
719 720
-- Static closures are never themselves black-holed.

721 722
blackHoleOnEntry :: ClosureInfo -> Bool
blackHoleOnEntry cl_info
723
  | isStaticRep (closureSMRep cl_info)
724
  = False        -- Never black-hole a static closure
725 726

  | otherwise
727
  = case closureLFInfo cl_info of
728
      LFReEntrant {}            -> False
Simon Peyton Jones's avatar
Simon Peyton Jones committed
729 730 731 732 733 734 735
      LFLetNoEscape             -> False
      LFThunk _ _no_fvs upd _ _ -> upd   -- See Note [Black-holing non-updatable thunks]
      _other -> panic "blackHoleOnEntry"

{- Note [Black-holing non-updatable thunks]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We must not black-hole non-updatable (single-entry) thunks otherwise
736
we run into issues like #10414. Specifically:
Simon Peyton Jones's avatar
Simon Peyton Jones committed
737 738 739 740 741 742 743 744 745 746

  * There is no reason to black-hole a non-updatable thunk: it should
    not be competed for by multiple threads

  * It could, conceivably, cause a space leak if we don't black-hole
    it, if there was a live but never-followed pointer pointing to it.
    Let's hope that doesn't happen.

  * It is dangerous to black-hole a non-updatable thunk because
     - is not updated (of course)
747
     - hence, if it is black-holed and another thread tries to evaluate
Simon Peyton Jones's avatar
Simon Peyton Jones committed
748
       it, that thread will block forever
749
    This actually happened in #10414.  So we do not black-hole
Simon Peyton Jones's avatar
Simon Peyton Jones committed
750 751 752 753 754 755 756 757 758
    non-updatable thunks.

  * How could two threads evaluate the same non-updatable (single-entry)
    thunk?  See Reid Barton's example below.

  * Only eager blackholing could possibly black-hole a non-updatable
    thunk, because lazy black-holing only affects thunks with an
    update frame on the stack.

759
Here is and example due to Reid Barton (#10414):
760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799
    x = \u []  concat [[1], []]
with the following definitions,

    concat x = case x of
        []       -> []
        (:) x xs -> (++) x (concat xs)

    (++) xs ys = case xs of
        []         -> ys
        (:) x rest -> (:) x ((++) rest ys)

Where we use the syntax @\u []@ to denote an updatable thunk and @\s []@ to
denote a single-entry (i.e. non-updatable) thunk. After a thread evaluates @x@
to WHNF and calls @(++)@ the heap will contain the following thunks,

    x = 1 : y
    y = \u []  (++) [] z
    z = \s []  concat []

Now that the stage is set, consider the follow evaluations by two racing threads
A and B,

  1. Both threads enter @y@ before either is able to replace it with an
     indirection

  2. Thread A does the case analysis in @(++)@ and consequently enters @z@,
     replacing it with a black-hole

  3. At some later point thread B does the same case analysis and also attempts
     to enter @z@. However, it finds that it has been replaced with a black-hole
     so it blocks.

  4. Thread A eventually finishes evaluating @z@ (to @[]@) and updates @y@
     accordingly. It does *not* update @z@, however, as it is single-entry. This
     leaves Thread B blocked forever on a black-hole which will never be
     updated.

To avoid this sort of condition we never black-hole non-updatable thunks.
-}

800 801 802 803 804 805 806 807 808 809 810 811
isStaticClosure :: ClosureInfo -> Bool
isStaticClosure cl_info = isStaticRep (closureSMRep cl_info)

closureUpdReqd :: ClosureInfo -> Bool
closureUpdReqd ClosureInfo{ closureLFInfo = lf_info } = lfUpdatable lf_info

lfUpdatable :: LambdaFormInfo -> Bool
lfUpdatable (LFThunk _ _ upd _ _)  = upd
lfUpdatable _ = False

closureSingleEntry :: ClosureInfo -> Bool
closureSingleEntry (ClosureInfo { closureLFInfo = LFThunk _ _ upd _ _}) = not upd
812
closureSingleEntry (ClosureInfo { closureLFInfo = LFReEntrant _ OneShotLam _ _ _}) = True
813 814 815
closureSingleEntry _ = False

closureReEntrant :: ClosureInfo -> Bool
816
closureReEntrant (ClosureInfo { closureLFInfo = LFReEntrant {} }) = True
817 818
closureReEntrant _ = False

819
closureFunInfo :: ClosureInfo -> Maybe (RepArity, ArgDescr)
820 821
closureFunInfo (ClosureInfo { closureLFInfo = lf_info }) = lfFunInfo lf_info

822
lfFunInfo :: LambdaFormInfo ->  Maybe (RepArity, ArgDescr)
823 824
lfFunInfo (LFReEntrant _ _ arity _ arg_desc)  = Just (arity, arg_desc)
lfFunInfo _                                   = Nothing
825

826 827 828
funTag :: DynFlags -> ClosureInfo -> DynTag
funTag dflags (ClosureInfo { closureLFInfo = lf_info })
    = lfDynTag dflags lf_info
829 830 831 832

isToplevClosure :: ClosureInfo -> Bool
isToplevClosure (ClosureInfo { closureLFInfo = lf_info })
  = case lf_info of
833 834 835
      LFReEntrant TopLevel _ _ _ _ -> True
      LFThunk TopLevel _ _ _ _     -> True
      _other                       -> False
836 837 838 839 840

--------------------------------------
--   Label generation
--------------------------------------

841 842
staticClosureLabel :: ClosureInfo -> CLabel
staticClosureLabel = toClosureLbl .  closureInfoLabel
Simon Marlow's avatar
Simon Marlow committed
843

844 845
closureSlowEntryLabel :: ClosureInfo -> CLabel
closureSlowEntryLabel = toSlowEntryLbl . closureInfoLabel
Simon Marlow's avatar
Simon Marlow committed
846

847 848 849 850
closureLocalEntryLabel :: DynFlags -> ClosureInfo -> CLabel
closureLocalEntryLabel dflags
  | tablesNextToCode dflags = toInfoLbl  . closureInfoLabel
  | otherwise               = toEntryLbl . closureInfoLabel
batterseapower's avatar
batterseapower committed
851

852 853
mkClosureInfoTableLabel :: Id -> LambdaFormInfo -> CLabel
mkClosureInfoTableLabel id lf_info
batterseapower's avatar
batterseapower committed
854
  = case lf_info of
855
        LFThunk _ _ upd_flag (SelectorThunk offset) _
856
                      -> mkSelectorInfoLabel upd_flag offset
857

858
        LFThunk _ _ upd_flag (ApThunk arity) _
859
                      -> mkApInfoTableLabel upd_flag arity
860

861 862
        LFThunk{}     -> std_mk_lbl name cafs
        LFReEntrant{} -> std_mk_lbl name cafs
Simon Marlow's avatar
Simon Marlow committed
863
        _other        -> panic "closureInfoTableLabel"
864

865
  where
866 867
    name = idName id

Simon Marlow's avatar
Simon Marlow committed
868
    std_mk_lbl | is_local  = mkLocalInfoTableLabel
869 870
               | otherwise = mkInfoTableLabel

871 872 873 874 875 876 877
    cafs     = idCafInfo id
    is_local = isDataConWorkId id
       -- Make the _info pointer for the implicit datacon worker
       -- binding local. The reason we can do this is that importing
       -- code always either uses the _closure or _con_info. By the
       -- invariants in CorePrep anything else gets eta expanded.

878

879
thunkEntryLabel :: DynFlags -> Name -> CafInfo -> StandardFormInfo -> Bool -> CLabel
Simon Marlow's avatar
Simon Marlow committed
880 881
-- thunkEntryLabel is a local help function, not exported.  It's used from
-- getCallMethod.
882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902
thunkEntryLabel dflags _thunk_id _ (ApThunk arity) upd_flag
  = enterApLabel dflags upd_flag arity
thunkEntryLabel dflags _thunk_id _ (SelectorThunk offset) upd_flag
  = enterSelectorLabel dflags upd_flag offset
thunkEntryLabel dflags thunk_id c _ _
  = enterIdLabel dflags thunk_id c

enterApLabel :: DynFlags -> Bool -> Arity -> CLabel
enterApLabel dflags is_updatable arity
  | tablesNextToCode dflags = mkApInfoTableLabel is_updatable arity
  | otherwise               = mkApEntryLabel is_updatable arity

enterSelectorLabel :: DynFlags -> Bool -> WordOff -> CLabel
enterSelectorLabel dflags upd_flag offset
  | tablesNextToCode dflags = mkSelectorInfoLabel upd_flag offset
  | otherwise               = mkSelectorEntryLabel upd_flag offset

enterIdLabel :: DynFlags -> Name -> CafInfo -> CLabel
enterIdLabel dflags id c
  | tablesNextToCode dflags = mkInfoTableLabel id c
  | otherwise               = mkEntryLabel id c
903 904 905 906 907 908 909 910 911 912 913 914 915 916 917


--------------------------------------
--   Profiling
--------------------------------------

-- Profiling requires two pieces of information to be determined for
-- each closure's info table --- description and type.

-- The description is stored directly in the @CClosureInfoTable@ when the
-- info table is built.

-- The type is determined from the type information stored with the @Id@
-- in the closure info using @closureTypeDescr@.

918 919
mkProfilingInfo :: DynFlags -> Id -> String -> ProfilingInfo
mkProfilingInfo dflags id val_descr
ian@well-typed.com's avatar
ian@well-typed.com committed
920
  | not (gopt Opt_SccProfilingOn dflags) = NoProfilingInfo
921
  | otherwise = ProfilingInfo ty_descr_w8 (BS8.pack val_descr)
922
  where
923
    ty_descr_w8  = BS8.pack (getTyDescription (idType id))
924 925 926 927 928

getTyDescription :: Type -> String
getTyDescription ty
  = case (tcSplitSigmaTy ty) of { (_, _, tau_ty) ->
    case tau_ty of
929 930 931
      TyVarTy _              -> "*"
      AppTy fun _            -> getTyDescription fun
      TyConApp tycon _       -> getOccString tycon
932
      FunTy {}              -> '-' : fun_result tau_ty
933
      ForAllTy _  ty         -> getTyDescription ty
934
      LitTy n                -> getTyLitDescription n
935 936
      CastTy ty _            -> getTyDescription ty
      CoercionTy co          -> pprPanic "getTyDescription" (ppr co)
937 938
    }
  where
939 940
    fun_result (FunTy { ft_res = res }) = '>' : fun_result res
    fun_result other                    = getTyDescription other
941

942 943 944
getTyLitDescription :: TyLit -> String
getTyLitDescription l =
  case l of
945 946
    NumTyLit n -> show n
    StrTyLit n -> show n
947

948
--------------------------------------
949
--   CmmInfoTable-related things
950 951
--------------------------------------

952 953
mkDataConInfoTable :: DynFlags -> DataCon -> Bool -> Int -> Int -> CmmInfoTable
mkDataConInfoTable dflags data_con is_static ptr_wds nonptr_wds
954 955 956
 = CmmInfoTable { cit_lbl  = info_lbl
                , cit_rep  = sm_rep
                , cit_prof = prof
957
                , cit_srt  = Nothing
Simon Marlow's avatar
Simon Marlow committed
958
                , cit_clo  = Nothing }
959 960
 where
   name = dataConName data_con
Simon Marlow's avatar
Simon Marlow committed
961
   info_lbl = mkConInfoTableLabel name NoCafRefs
962
   sm_rep = mkHeapRep dflags is_static ptr_wds nonptr_wds cl_type
963
   cl_type = Constr (dataConTagZ data_con) (dataConIdentity data_con)
964 965
                  -- We keep the *zero-indexed* tag in the srt_len field
                  -- of the info table of a data constructor.
966

ian@well-typed.com's avatar
ian@well-typed.com committed
967
   prof | not (gopt Opt_SccProfilingOn dflags) = NoProfilingInfo
968
        | otherwise                            = ProfilingInfo ty_descr val_descr
969

970 971
   ty_descr  = BS8.pack $ occNameString $ getOccName $ dataConTyCon data_con
   val_descr = BS8.pack $ occNameString $ getOccName data_con
972

Simon Marlow's avatar
Simon Marlow committed
973 974 975 976 977 978 979 980
-- We need a black-hole closure info to pass to @allocDynClosure@ when we
-- want to allocate the black hole on entry to a CAF.

cafBlackHoleInfoTable :: CmmInfoTable
cafBlackHoleInfoTable
  = CmmInfoTable { cit_lbl  = mkCAFBlackHoleInfoTableLabel
                 , cit_rep  = blackHoleRep
                 , cit_prof = NoProfilingInfo
981 982
                 , cit_srt  = Nothing
                 , cit_clo  = Nothing }
983

984 985 986 987 988
indStaticInfoTable :: CmmInfoTable
indStaticInfoTable
  = CmmInfoTable { cit_lbl  = mkIndStaticInfoLabel
                 , cit_rep  = indStaticRep
                 , cit_prof = NoProfilingInfo
989 990
                 , cit_srt  = Nothing
                 , cit_clo  = Nothing }
991

Simon Marlow's avatar
Simon Marlow committed
992
staticClosureNeedsLink :: Bool -> CmmInfoTable -> Bool
993 994
-- A static closure needs a link field to aid the GC when traversing
-- the static closure graph.  But it only needs such a field if either
Simon Marlow's avatar
Simon Marlow committed
995
--        a) it has an SRT
996
--        b) it's a constructor with one or more pointer fields
997 998
-- In case (b), the constructor's fields themselves play the role
-- of the SRT.
Simon Marlow's avatar
Simon Marlow committed
999
staticClosureNeedsLink has_srt CmmInfoTable{ cit_rep = smrep }
1000
  | isConRep smrep         = not (isStaticNoCafCon smrep)
1001
  | otherwise              = has_srt