StgCmmClosure.hs 37.5 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
        ConTagZ, dataConTagZ,
17

18
        idPrimRep, isVoidRep, isGcPtrRep, addIdReps, addArgReps,
19
        argPrimRep,
20

21 22
        -- * LambdaFormInfo
        LambdaFormInfo,         -- Abstract
23 24 25
        StandardFormInfo,        -- ...ditto...
        mkLFThunk, mkLFReEntrant, mkConLFInfo, mkSelectorLFInfo,
        mkApLFInfo, mkLFImported, mkLFArgument, mkLFLetNoEscape,
Simon Marlow's avatar
Simon Marlow committed
26 27
        lfDynTag,
        maybeIsLFCon, isLFThunk, isLFReEntrant, lfUpdatable,
28

29 30 31
        -- * Used by other modules
        CgLoc(..), SelfLoopInfo, CallMethod(..),
        nodeMustPointToIt, isKnownFun, funTag, tagForArity, getCallMethod,
32 33

        -- * ClosureInfo
34
        ClosureInfo,
Simon Marlow's avatar
Simon Marlow committed
35
        mkClosureInfo,
36
        mkCmmInfo,
37

38 39
        -- ** Inspection
        closureLFInfo, closureName,
Simon Marlow's avatar
Simon Marlow committed
40

41 42 43
        -- ** Labels
        -- These just need the info table label
        closureInfoLabel, staticClosureLabel,
nfrisby's avatar
nfrisby committed
44
        closureSlowEntryLabel, closureLocalEntryLabel,
45

46 47
        -- ** Predicates
        -- These are really just functions on LambdaFormInfo
Simon Marlow's avatar
Simon Marlow committed
48
        closureUpdReqd, closureSingleEntry,
49 50
        closureReEntrant, closureFunInfo,
        isToplevClosure,
51

52 53
        blackHoleOnEntry,  -- Needs LambdaFormInfo and SMRep
        isStaticClosure,   -- Needs SMPre
54

55
        -- * InfoTables
56
        mkDataConInfoTable,
57
        cafBlackHoleInfoTable,
58
        indStaticInfoTable,
59
        staticClosureNeedsLink,
60 61 62 63 64 65 66 67 68
    ) where

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

#define FAST_STRING_NOT_NEEDED
#include "HsVersions.h"

import StgSyn
import SMRep
69
import Cmm
70
import PprCmmExpr()
71

72
import BlockId
73 74 75 76
import CLabel
import Id
import IdInfo
import DataCon
77
import FastString
78 79 80 81 82 83 84
import Name
import Type
import TypeRep
import TcType
import TyCon
import BasicTypes
import Outputable
85
import DynFlags
86
import Util
87

88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118
-----------------------------------------------------------------------------
--                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
  ppr (CmmLoc e)    = ptext (sLit "cmm") <+> ppr e
  ppr (LneLoc b rs) = ptext (sLit "lne") <+> ppr b <+> ppr rs

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

-- used by ticky profiling
isKnownFun :: LambdaFormInfo -> Bool
isKnownFun (LFReEntrant _ _ _ _) = True
isKnownFun LFLetNoEscape         = True
isKnownFun _ = False


119
-----------------------------------------------------------------------------
120
--                Representations
121 122
-----------------------------------------------------------------------------

123 124
-- Why are these here?

125 126
idPrimRep :: Id -> PrimRep
idPrimRep id = typePrimRep (idType id)
127 128
    -- NB: typePrimRep fails on unboxed tuples,
    --     but by StgCmm no Ids have unboxed tuple type
129

130 131 132 133 134 135 136 137 138 139 140
addIdReps :: [Id] -> [(PrimRep, Id)]
addIdReps ids = [(idPrimRep id, id) | id <- ids]

addArgReps :: [StgArg] -> [(PrimRep, StgArg)]
addArgReps args = [(argPrimRep arg, arg) | arg <- args]

argPrimRep :: StgArg -> PrimRep
argPrimRep arg = typePrimRep (stgArgType arg)


-----------------------------------------------------------------------------
141
--                LambdaFormInfo
142 143 144 145 146 147 148 149
-----------------------------------------------------------------------------

-- 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
150 151 152 153
  = LFReEntrant         -- Reentrant closure (a function)
        TopLevelFlag    -- True if top level
        !RepArity       -- Arity. Invariant: always > 0
        !Bool           -- True <=> no fvs
154 155
        ArgDescr        -- Argument descriptor (should really be in ClosureInfo)

156
  | LFThunk             -- Thunk (zero arity)
157
        TopLevelFlag
158 159
        !Bool           -- True <=> no free vars
        !Bool           -- True <=> updatable (i.e., *not* single-entry)
160
        StandardFormInfo
161
        !Bool           -- True <=> *might* be a function type
162

163 164
  | LFCon               -- A saturated constructor application
        DataCon         -- The constructor
165

166
  | LFUnknown           -- Used for function arguments and imported things.
167 168 169 170 171 172 173 174 175 176
                        -- 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)
        !Bool                -- True <=> *might* be a function type
                        --      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

177
  | LFUnLifted          -- A value of unboxed type;
178 179
                        -- always a value, needs evaluation

180
  | LFLetNoEscape       -- See LetNoEscape module for precise description
181

182 183

-------------------------
184
-- StandardFormInfo tells whether this thunk has one of
185 186 187 188
-- a small number of standard forms

data StandardFormInfo
  = NonStandardThunk
189
        -- The usual case: not of the standard forms
190 191

  | SelectorThunk
192 193 194 195
        -- A SelectorThunk is of form
        --      case x of
        --           con a1,..,an -> ak
        -- and the constructor is from a single-constr type.
196
       WordOff          -- 0-origin offset of ak within the "goods" of
197 198 199 200 201 202 203 204 205 206
                        -- 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
207 208 209


------------------------------------------------------
210
--                Building LambdaFormInfo
211 212 213
------------------------------------------------------

mkLFArgument :: Id -> LambdaFormInfo
214 215
mkLFArgument id
  | isUnLiftedType ty             = LFUnLifted
216
  | might_be_a_function ty = LFUnknown True
217
  | otherwise                    = LFUnknown False
218 219 220 221 222 223 224 225
  where
    ty = idType id

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

-------------
226 227 228
mkLFReEntrant :: TopLevelFlag    -- True of top level
              -> [Id]            -- Free vars
              -> [Id]            -- Args
229 230
              -> ArgDescr        -- Argument descriptor
              -> LambdaFormInfo
231

232
mkLFReEntrant top fvs args arg_descr
233 234 235 236 237 238
  = LFReEntrant top (length args) (null fvs) arg_descr

-------------
mkLFThunk :: Type -> TopLevelFlag -> [Id] -> UpdateFlag -> LambdaFormInfo
mkLFThunk thunk_ty top fvs upd_flag
  = ASSERT( not (isUpdatable upd_flag) || not (isUnLiftedType thunk_ty) )
239 240 241 242
    LFThunk top (null fvs)
            (isUpdatable upd_flag)
            NonStandardThunk
            (might_be_a_function thunk_ty)
243 244 245 246 247 248

--------------
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
249 250 251 252 253 254
  | UnaryRep rep <- repType ty
  , Just tc <- tyConAppTyCon_maybe rep
  , isDataTyCon tc
  = False
  | otherwise
  = True
255 256 257 258 259 260 261 262

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

-------------
mkSelectorLFInfo :: Id -> Int -> Bool -> LambdaFormInfo
mkSelectorLFInfo id offset updatable
263 264
  = LFThunk NotTopLevel False updatable (SelectorThunk offset)
        (might_be_a_function (idType id))
265 266 267 268 269

-------------
mkApLFInfo :: Id -> UpdateFlag -> Arity -> LambdaFormInfo
mkApLFInfo id upd_flag arity
  = LFThunk NotTopLevel (arity == 0) (isUpdatable upd_flag) (ApThunk arity)
270
        (might_be_a_function (idType id))
271 272 273 274 275 276

-------------
mkLFImported :: Id -> LambdaFormInfo
mkLFImported id
  | Just con <- isDataConWorkId_maybe id
  , isNullaryRepDataCon con
277
  = LFCon con   -- An imported nullary constructor
278 279
                -- We assume that the constructor is evaluated so that
                -- the id really does point directly to the constructor
280 281 282 283 284 285 286

  | arity > 0
  = LFReEntrant TopLevel arity True (panic "arg_descr")

  | otherwise
  = mkLFArgument id -- Not sure of exact arity
  where
287
    arity = idRepArity id
288 289

-----------------------------------------------------
290
--                Dynamic pointer tagging
291 292
-----------------------------------------------------

293
type ConTagZ = Int      -- A *zero-indexed* contructor tag
294

295 296
type DynTag = Int       -- The tag on a *pointer*
                        -- (from the dynamic-tagging paper)
297

298 299 300 301 302 303 304 305 306 307 308 309
-- 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.
310

311 312
isSmallFamily :: DynFlags -> Int -> Bool
isSmallFamily dflags fam_size = fam_size <= mAX_PTR_TAG dflags
313 314 315 316 317 318

-- We keep the *zero-indexed* tag in the srt_len field of the info
-- table of a data constructor.
dataConTagZ :: DataCon -> ConTagZ
dataConTagZ con = dataConTag con - fIRST_TAG

319 320 321 322
tagForCon :: DynFlags -> DataCon -> DynTag
tagForCon dflags con
  | isSmallFamily dflags fam_size = con_tag + 1
  | otherwise                     = 1
323 324 325 326
  where
    con_tag  = dataConTagZ con
    fam_size = tyConFamilySize (dataConTyCon con)

327 328 329 330
tagForArity :: DynFlags -> RepArity -> DynTag
tagForArity dflags arity
 | isSmallFamily dflags arity = arity
 | otherwise                  = 0
331

332
lfDynTag :: DynFlags -> LambdaFormInfo -> DynTag
333 334
-- Return the tag in the low order bits of a variable bound
-- to this LambdaForm
335 336 337
lfDynTag dflags (LFCon con)               = tagForCon dflags con
lfDynTag dflags (LFReEntrant _ arity _ _) = tagForArity dflags arity
lfDynTag _      _other                    = 0
338 339 340


-----------------------------------------------------------------------------
341
--                Observing LambdaFormInfo
342 343 344 345 346 347 348 349 350
-----------------------------------------------------------------------------

-------------
maybeIsLFCon :: LambdaFormInfo -> Maybe DataCon
maybeIsLFCon (LFCon con) = Just con
maybeIsLFCon _ = Nothing

------------
isLFThunk :: LambdaFormInfo -> Bool
351
isLFThunk (LFThunk {})  = True
352 353
isLFThunk _ = False

354 355 356
isLFReEntrant :: LambdaFormInfo -> Bool
isLFReEntrant (LFReEntrant {}) = True
isLFReEntrant _                = False
357 358

-----------------------------------------------------------------------------
359
--                Choosing SM reps
360 361
-----------------------------------------------------------------------------

362 363 364 365 366 367
lfClosureType :: LambdaFormInfo -> ClosureTypeInfo
lfClosureType (LFReEntrant _ arity _ argd) = Fun arity argd
lfClosureType (LFCon con)                  = Constr (dataConTagZ con)
                                                    (dataConIdentity con)
lfClosureType (LFThunk _ _ _ is_sel _)     = thunkClosureType is_sel
lfClosureType _                            = panic "lfClosureType"
368

369 370 371
thunkClosureType :: StandardFormInfo -> ClosureTypeInfo
thunkClosureType (SelectorThunk off) = ThunkSelector off
thunkClosureType _                   = Thunk
372 373 374 375 376 377 378

-- 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.

-----------------------------------------------------------------------------
379
--                nodeMustPointToIt
380 381
-----------------------------------------------------------------------------

382
nodeMustPointToIt :: DynFlags -> LambdaFormInfo -> Bool
383
-- If nodeMustPointToIt is true, then the entry convention for
384
-- this closure has R1 (the "Node" register) pointing to the
385 386
-- closure itself --- the "self" argument

387
nodeMustPointToIt _ (LFReEntrant top _ no_fvs _)
388 389
  =  not no_fvs          -- Certainly if it has fvs we need to point to it
  || isNotTopLevel top   -- See Note [GC recovery]
390 391 392
        -- 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.
393 394 395 396 397 398

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
399 400 401 402 403 404
          -- 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
405 406
          --                                    top-level thunks? If not,
          --                                    isNotTopLevel subsumes this
407

408 409
nodeMustPointToIt _ (LFThunk {})        -- Node must point to a standard-form thunk
  = True
410

411
nodeMustPointToIt _ (LFCon _) = True
412

413 414 415 416 417 418 419 420 421 422
        -- 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.
423

424 425
nodeMustPointToIt _ (LFUnknown _)   = True
nodeMustPointToIt _ LFUnLifted      = False
426
nodeMustPointToIt _ LFLetNoEscape   = False
427

428 429 430 431 432 433
{- 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.
434
So even if there are no free variables in <body>, we still make
435 436 437 438 439 440 441
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.
442
This actually happened during bootsrapping GHC itself, with f=mkRdrFunBind
443 444
in TcGenDeriv.) -}

445
-----------------------------------------------------------------------------
446
--                getCallMethod
447 448 449 450 451 452
-----------------------------------------------------------------------------

{- 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.

453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473
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
474 475 476 477 478

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

data CallMethod
479
  = EnterIt             -- No args, not a function
480

481
  | JumpToIt BlockId [LocalReg] -- A join point or a header of a local loop
482

483
  | ReturnIt            -- It's a value (function, unboxed value,
484
                        -- or constructor), so just return it.
485

486 487
  | SlowCall                -- Unknown fun, or known fun with
                        -- too few args.
488

489
  | DirectEntry         -- Jump directly, with args in regs
490 491
        CLabel          --   The code label
        RepArity        --   Its arity
492

493 494
getCallMethod :: DynFlags
              -> Name           -- Function being applied
495 496 497 498 499 500 501 502 503 504 505
              -> 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
              -> 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?
506
              -> CallMethod
507

508 509
getCallMethod dflags _ id _ n_args _cg_loc (Just (self_loop_id, block_id, args))
  | gopt Opt_Loopification dflags, id == self_loop_id, n_args == length args
510
  -- If these patterns match then we know that:
511
  --   * loopification optimisation is turned on
512 513 514 515 516 517
  --   * function is performing a self-recursive call in a tail position
  --   * number of parameters of the function matches functions arity.
  -- See Note [Self-recursive tail calls] in StgCmmExpr for more details
  = JumpToIt block_id args

getCallMethod dflags _name _ lf_info _n_args _cg_loc _self_loop_info
ian@well-typed.com's avatar
ian@well-typed.com committed
518
  | nodeMustPointToIt dflags lf_info && gopt Opt_Parallel dflags
519
  =     -- If we're parallel, then we must always enter via node.
520 521
        -- The reason is that the closure may have been
        -- fetched since we allocated it.
522 523
    EnterIt

524 525
getCallMethod dflags name id (LFReEntrant _ arity _ _) n_args _cg_loc
              _self_loop_info
526
  | n_args == 0    = ASSERT( arity /= 0 )
527 528
                     ReturnIt        -- No args at all
  | n_args < arity = SlowCall        -- Not enough args
529
  | otherwise      = DirectEntry (enterIdLabel dflags name (idCafInfo id)) arity
530

531
getCallMethod _ _name _ LFUnLifted n_args _cg_loc _self_loop_info
532 533
  = ASSERT( n_args == 0 ) ReturnIt

534
getCallMethod _ _name _ (LFCon _) n_args _cg_loc _self_loop_info
535 536
  = ASSERT( n_args == 0 ) ReturnIt

537 538
getCallMethod dflags name id (LFThunk _ _ updatable std_form_info is_fun)
              n_args _cg_loc _self_loop_info
539 540
  | 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
541
                -- is the fast-entry code]
542 543

  -- Since is_fun is False, we are *definitely* looking at a data value
ian@well-typed.com's avatar
ian@well-typed.com committed
544
  | updatable || gopt Opt_Ticky dflags -- to catch double entry
545
      {- OLD: || opt_SMP
546 547 548 549
         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
        -}
550
  = EnterIt
551 552 553 554 555 556

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

557 558 559 560 561 562 563
    -- 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

564
  | otherwise        -- Jump direct to code for single-entry thunks
565
  = ASSERT( n_args == 0 )
566 567
    DirectEntry (thunkEntryLabel dflags name (idCafInfo id) std_form_info
                updatable) 0
568

569
getCallMethod _ _name _ (LFUnknown True) _n_arg _cg_locs _self_loop_info
570 571
  = SlowCall -- might be a function

572
getCallMethod _ name _ (LFUnknown False) n_args _cg_loc _self_loop_info
573
  = ASSERT2( n_args == 0, ppr name <+> ppr n_args )
574 575
    EnterIt -- Not a function

576 577
getCallMethod _ _name _ LFLetNoEscape _n_args (LneLoc blk_id lne_regs)
              _self_loop_info
578
  = JumpToIt blk_id lne_regs
579

580
getCallMethod _ _ _ _ _ _ _ = panic "Unknown call method"
581 582

-----------------------------------------------------------------------------
583
--                staticClosureRequired
584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605
-----------------------------------------------------------------------------

{-  staticClosureRequired is never called (hence commented out)

    SimonMar writes (Sept 07) It's an optimisation we used to apply at
    one time, I believe, but it got lost probably in the rewrite of
    the RTS/code generator.  I left that code there to remind me to
    look into whether it was worth doing sometime

{- Avoiding generating entries and info tables
   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
At present, for every function we generate all of the following,
just in case.  But they aren't always all needed, as noted below:

[NB1: all of this applies only to *functions*.  Thunks always
have closure, info table, and entry code.]

[NB2: All are needed if the function is *exported*, just to play safe.]

* Fast-entry code  ALWAYS NEEDED

* Slow-entry code
606 607 608 609 610
        Needed iff (a) we have any un-saturated calls to the function
        OR         (b) the function is passed as an arg
        OR         (c) we're in the parallel world and the function has free vars
                       [Reason: in parallel world, we always enter functions
                       with free vars via the closure.]
611 612

* The function closure
613 614 615
        Needed iff (a) we have any un-saturated calls to the function
        OR         (b) the function is passed as an arg
        OR         (c) if the function has free vars (ie not top level)
616 617 618 619 620 621 622 623 624

  Why case (a) here?  Because if the arg-satis check fails,
  UpdatePAP stuffs a pointer to the function closure in the PAP.
  [Could be changed; UpdatePAP could stuff in a code ptr instead,
   but doesn't seem worth it.]

  [NB: these conditions imply that we might need the closure
  without the slow-entry code.  Here's how.

625 626 627
        f x y = let g w = ...x..y..w...
                in
                ...(g t)...
628 629 630 631 632 633 634

  Here we need a closure for g which contains x and y,
  but since the calls are all saturated we just jump to the
  fast entry point for g, with R1 pointing to the closure for g.]


* Standard info table
635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651
        Needed iff (a) we have any un-saturated calls to the function
        OR         (b) the function is passed as an arg
        OR         (c) the function has free vars (ie not top level)

        NB.  In the sequential world, (c) is only required so that the function closure has
        an info table to point to, to keep the storage manager happy.
        If (c) alone is true we could fake up an info table by choosing
        one of a standard family of info tables, whose entry code just
        bombs out.

        [NB In the parallel world (c) is needed regardless because
        we enter functions with free vars via the closure.]

        If (c) is retained, then we'll sometimes generate an info table
        (for storage mgr purposes) without slow-entry code.  Then we need
        to use an error label in the info table to substitute for the absent
        slow entry code.
652 653 654
-}

staticClosureRequired
655 656 657 658
        :: Name
        -> StgBinderInfo
        -> LambdaFormInfo
        -> Bool
659
staticClosureRequired binder bndr_info
660
                      (LFReEntrant top_level _ _ _)        -- It's a function
661
  = ASSERT( isTopLevel top_level )
662 663
        -- Assumption: it's a top-level, no-free-var binding
        not (satCallsOnly bndr_info)
664 665 666 667 668

staticClosureRequired binder other_binder_info other_lf_info = True
-}

-----------------------------------------------------------------------------
669
--              Data types for closure information
670 671 672
-----------------------------------------------------------------------------


673
{- ClosureInfo: information about a binding
674

675 676 677
   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).
678

679 680
   To a first approximation:
       ClosureInfo = (LambdaFormInfo, CmmInfoTable)
Simon Marlow's avatar
Simon Marlow committed
681

682 683 684 685
   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.
686
           it knows the info table label)
687 688 689 690
-}

data ClosureInfo
  = ClosureInfo {
691 692 693 694 695 696 697 698 699 700 701 702
        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
703 704
    }

Simon Marlow's avatar
Simon Marlow committed
705 706
-- | Convert from 'ClosureInfo' to 'CmmInfoTable'.
mkCmmInfo :: ClosureInfo -> CmmInfoTable
707 708 709 710
mkCmmInfo ClosureInfo {..}
  = CmmInfoTable { cit_lbl  = closureInfoLabel
                 , cit_rep  = closureSMRep
                 , cit_prof = closureProf
Simon Marlow's avatar
Simon Marlow committed
711
                 , cit_srt  = NoC_SRT }
Simon Marlow's avatar
Simon Marlow committed
712

713
--------------------------------------
714
--        Building ClosureInfos
715 716
--------------------------------------

717
mkClosureInfo :: DynFlags
718 719 720 721
              -> Bool                -- Is static
              -> Id
              -> LambdaFormInfo
              -> Int -> Int        -- Total and pointer words
Simon Marlow's avatar
Simon Marlow committed
722
              -> String         -- String descriptor
723
              -> ClosureInfo
724
mkClosureInfo dflags is_static id lf_info tot_wds ptr_wds val_descr
725 726 727 728 729
  = 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)
730
  where
Simon Marlow's avatar
Simon Marlow committed
731
    name       = idName id
732
    sm_rep     = mkHeapRep dflags is_static ptr_wds nonptr_wds (lfClosureType lf_info)
733
    prof       = mkProfilingInfo dflags id val_descr
734
    nonptr_wds = tot_wds - ptr_wds
735

736
    info_lbl = mkClosureInfoTableLabel id lf_info
737 738

--------------------------------------
739
--   Other functions over ClosureInfo
740 741
--------------------------------------

742 743 744
-- 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.
745

746 747 748 749 750 751 752 753
-- 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
754 755
-- was on. But it didn't work, and it wasn't strictly necessary
-- to bring back minimal ticky-ticky, so now EAGER_BLACKHOLING
756
-- is unconditionally disabled. -- krc 1/2007
757 758 759 760 761 762 763 764 765 766
--
--
-- A single-entry (non-updatable) thunk can actually be entered
-- more than once in a parallel program, if work is duplicated
-- by two threads both entering the same updatable thunk before
-- the other has blackholed it. So, we must not eagerly
-- blackhole non-updatable thunks, or the second thread to
-- enter one will become blocked indefinitely. (They are not
-- blackholed by lazy blackholing either, since they have no
-- associated update frame.) See Trac #10414.
767 768 769

-- Static closures are never themselves black-holed.

770 771
blackHoleOnEntry :: ClosureInfo -> Bool
blackHoleOnEntry cl_info
772
  | isStaticRep (closureSMRep cl_info)
773
  = False        -- Never black-hole a static closure
774 775

  | otherwise
776
  = case closureLFInfo cl_info of
777 778
        LFReEntrant _ _ _ _          -> False
        LFLetNoEscape                   -> False
779
        LFThunk _ _no_fvs updatable _ _ -> updatable
780
        _other -> panic "blackHoleOnEntry"      -- Should never happen
781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799

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
closureSingleEntry _ = False

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

800
closureFunInfo :: ClosureInfo -> Maybe (RepArity, ArgDescr)
801 802
closureFunInfo (ClosureInfo { closureLFInfo = lf_info }) = lfFunInfo lf_info

803
lfFunInfo :: LambdaFormInfo ->  Maybe (RepArity, ArgDescr)
804 805 806
lfFunInfo (LFReEntrant _ arity _ arg_desc)  = Just (arity, arg_desc)
lfFunInfo _                                 = Nothing

807 808 809
funTag :: DynFlags -> ClosureInfo -> DynTag
funTag dflags (ClosureInfo { closureLFInfo = lf_info })
    = lfDynTag dflags lf_info
810 811 812 813 814 815

isToplevClosure :: ClosureInfo -> Bool
isToplevClosure (ClosureInfo { closureLFInfo = lf_info })
  = case lf_info of
      LFReEntrant TopLevel _ _ _ -> True
      LFThunk TopLevel _ _ _ _   -> True
816
      _other                         -> False
817 818 819 820 821

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

822 823
staticClosureLabel :: ClosureInfo -> CLabel
staticClosureLabel = toClosureLbl .  closureInfoLabel
Simon Marlow's avatar
Simon Marlow committed
824

825 826
closureSlowEntryLabel :: ClosureInfo -> CLabel
closureSlowEntryLabel = toSlowEntryLbl . closureInfoLabel
Simon Marlow's avatar
Simon Marlow committed
827

828 829 830 831
closureLocalEntryLabel :: DynFlags -> ClosureInfo -> CLabel
closureLocalEntryLabel dflags
  | tablesNextToCode dflags = toInfoLbl  . closureInfoLabel
  | otherwise               = toEntryLbl . closureInfoLabel
batterseapower's avatar
batterseapower committed
832

833 834
mkClosureInfoTableLabel :: Id -> LambdaFormInfo -> CLabel
mkClosureInfoTableLabel id lf_info
batterseapower's avatar
batterseapower committed
835
  = case lf_info of
836
        LFThunk _ _ upd_flag (SelectorThunk offset) _
837
                      -> mkSelectorInfoLabel upd_flag offset
838

839
        LFThunk _ _ upd_flag (ApThunk arity) _
840
                      -> mkApInfoTableLabel upd_flag arity
841

842 843
        LFThunk{}     -> std_mk_lbl name cafs
        LFReEntrant{} -> std_mk_lbl name cafs
Simon Marlow's avatar
Simon Marlow committed
844
        _other        -> panic "closureInfoTableLabel"
845

846
  where
847 848
    name = idName id

Simon Marlow's avatar
Simon Marlow committed
849
    std_mk_lbl | is_local  = mkLocalInfoTableLabel
850 851
               | otherwise = mkInfoTableLabel

852 853 854 855 856 857 858
    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.

859

860
thunkEntryLabel :: DynFlags -> Name -> CafInfo -> StandardFormInfo -> Bool -> CLabel
Simon Marlow's avatar
Simon Marlow committed
861 862
-- thunkEntryLabel is a local help function, not exported.  It's used from
-- getCallMethod.
863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883
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
884 885 886 887 888 889 890 891 892 893 894 895 896 897 898


--------------------------------------
--   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@.

899 900
mkProfilingInfo :: DynFlags -> Id -> String -> ProfilingInfo
mkProfilingInfo dflags id val_descr
ian@well-typed.com's avatar
ian@well-typed.com committed
901
  | not (gopt Opt_SccProfilingOn dflags) = NoProfilingInfo
902 903 904 905
  | otherwise = ProfilingInfo ty_descr_w8 val_descr_w8
  where
    ty_descr_w8  = stringToWord8s (getTyDescription (idType id))
    val_descr_w8 = stringToWord8s val_descr
906 907 908 909 910

getTyDescription :: Type -> String
getTyDescription ty
  = case (tcSplitSigmaTy ty) of { (_, _, tau_ty) ->
    case tau_ty of
911 912 913 914
      TyVarTy _                            -> "*"
      AppTy fun _                   -> getTyDescription fun
      FunTy _ res                   -> '-' : '>' : fun_result res
      TyConApp tycon _              -> getOccString tycon
915
      ForAllTy _ ty          -> getTyDescription ty
916
      LitTy n                -> getTyLitDescription n
917 918 919
    }
  where
    fun_result (FunTy _ res) = '>' : fun_result res
920
    fun_result other             = getTyDescription other
921

922 923 924
getTyLitDescription :: TyLit -> String
getTyLitDescription l =
  case l of
925 926
    NumTyLit n -> show n
    StrTyLit n -> show n
927

928
--------------------------------------
Simon Marlow's avatar
Simon Marlow committed
929
--   CmmInfoTable-related things
930 931
--------------------------------------

932 933
mkDataConInfoTable :: DynFlags -> DataCon -> Bool -> Int -> Int -> CmmInfoTable
mkDataConInfoTable dflags data_con is_static ptr_wds nonptr_wds
934 935 936 937 938 939 940 941 942 943
 = CmmInfoTable { cit_lbl  = info_lbl
                , cit_rep  = sm_rep
                , cit_prof = prof
                , cit_srt  = NoC_SRT }
 where
   name = dataConName data_con

   info_lbl | is_static = mkStaticInfoTableLabel name NoCafRefs
            | otherwise = mkConInfoTableLabel    name NoCafRefs

944
   sm_rep = mkHeapRep dflags is_static ptr_wds nonptr_wds cl_type
945

946
   cl_type = Constr (dataConTagZ data_con) (dataConIdentity data_con)
947

ian@well-typed.com's avatar
ian@well-typed.com committed
948
   prof | not (gopt Opt_SccProfilingOn dflags) = NoProfilingInfo
949
        | otherwise                            = ProfilingInfo ty_descr val_descr
950 951 952

   ty_descr  = stringToWord8s $ occNameString $ getOccName $ dataConTyCon data_con
   val_descr = stringToWord8s $ occNameString $ getOccName data_con
Simon Marlow's avatar
Simon Marlow committed
953

Simon Marlow's avatar
Simon Marlow committed
954 955 956 957 958 959 960 961 962
-- 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
                 , cit_srt  = NoC_SRT }
Simon Marlow's avatar
Simon Marlow committed
963

964 965 966 967 968 969 970
indStaticInfoTable :: CmmInfoTable
indStaticInfoTable
  = CmmInfoTable { cit_lbl  = mkIndStaticInfoLabel
                 , cit_rep  = indStaticRep
                 , cit_prof = NoProfilingInfo
                 , cit_srt  = NoC_SRT }

Simon Marlow's avatar
Simon Marlow committed
971
staticClosureNeedsLink :: Bool -> CmmInfoTable -> Bool
Simon Marlow's avatar
Simon Marlow committed
972 973
-- 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
974 975
--         a) it has an SRT
--        b) it's a constructor with one or more pointer fields
Simon Marlow's avatar
Simon Marlow committed
976 977
-- In case (b), the constructor's fields themselves play the role
-- of the SRT.
Simon Marlow's avatar
Simon Marlow committed
978 979 980 981 982 983 984
--
-- At this point, the cit_srt field has not been calculated (that
-- happens right at the end of the Cmm pipeline), but we do have the
-- VarSet of CAFs that CoreToStg attached, and if that is empty there
-- will definitely not be an SRT.
--
staticClosureNeedsLink has_srt CmmInfoTable{ cit_rep = smrep }
Simon Marlow's avatar
Simon Marlow committed
985
  | isConRep smrep         = not (isStaticNoCafCon smrep)
Simon Marlow's avatar
Simon Marlow committed
986
  | otherwise              = has_srt -- needsSRT (cit_srt info_tbl)