Demand.hs 75.8 KB
Newer Older
1 2
{-# LANGUAGE CPP          #-}
{-# LANGUAGE ViewPatterns #-}
3 4 5

{-# OPTIONS_GHC -Wno-incomplete-uni-patterns #-}

Austin Seipp's avatar
Austin Seipp committed
6 7 8 9
{-
(c) The University of Glasgow 2006
(c) The GRASP/AQUA Project, Glasgow University, 1992-1998
-}
10

11 12 13 14 15
-- | A language to express the evaluation context of an expression as a
-- 'Demand' and track how an expression evaluates free variables and arguments
-- in turn as a 'DmdType'.
--
-- Lays out the abstract domain for "GHC.Core.Opt.DmdAnal".
Sylvain Henry's avatar
Sylvain Henry committed
16
module GHC.Types.Demand (
17 18 19 20 21 22 23 24 25 26 27 28
    -- * Demands
    Card(..), Demand(..), SubDemand(Prod), mkProd, viewProd,
    -- ** Algebra
    absDmd, topDmd, botDmd, seqDmd, topSubDmd,
    -- *** Least upper bound
    lubCard, lubDmd, lubSubDmd,
    -- *** Plus
    plusCard, plusDmd, plusSubDmd,
    -- *** Multiply
    multCard, multDmd, multSubDmd,
    -- ** Predicates on @Card@inalities and @Demand@s
    isAbs, isUsedOnce, isStrict,
29
    isAbsDmd, isUsedOnceDmd, isStrUsedDmd, isStrictDmd,
30 31 32 33 34 35 36
    isTopDmd, isSeqDmd, isWeakDmd,
    -- ** Special demands
    evalDmd,
    -- *** Demands used in PrimOp signatures
    lazyApply1Dmd, lazyApply2Dmd, strictOnceApply1Dmd, strictManyApply1Dmd,
    -- ** Other @Demand@ operations
    oneifyCard, oneifyDmd, strictifyDmd, strictifyDictDmd, mkWorkerDemand,
37
    peelCallDmd, peelManyCalls, mkCalledOnceDmd, mkCalledOnceDmds,
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
    addCaseBndrDmd,
    -- ** Extracting one-shot information
    argOneShots, argsOneShots, saturatedByOneShots,

    -- * Demand environments
    DmdEnv, emptyDmdEnv,
    keepAliveDmdEnv, reuseEnv,

    -- * Divergence
    Divergence(..), topDiv, botDiv, exnDiv, lubDivergence, isDeadEndDiv,

    -- * Demand types
    DmdType(..), dmdTypeDepth,
    -- ** Algebra
    nopDmdType, botDmdType,
    lubDmdType, plusDmdType, multDmdType,
    -- *** PlusDmdArg
    PlusDmdArg, mkPlusDmdArg, toPlusDmdArg,
    -- ** Other operations
    peelFV, findIdDemand, addDemand, splitDmdTy, deferAfterPreciseException,
58
    keepAliveDmdType,
59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76

    -- * Demand signatures
    StrictSig(..), mkStrictSigForArity, mkClosedStrictSig,
    splitStrictSig, strictSigDmdEnv, hasDemandEnvSig,
    nopSig, botSig, isTopSig, isDeadEndSig, appIsDeadEnd,
    -- ** Handling arity adjustments
    prependArgsStrictSig, etaConvertStrictSig,

    -- * Demand transformers from demand signatures
    DmdTransformer, dmdTransformSig, dmdTransformDataConSig, dmdTransformDictSelSig,

    -- * Trim to a type shape
    TypeShape(..), trimToType,

    -- * @seq@ing stuff
    seqDemand, seqDemandList, seqDmdType, seqStrictSig,

    -- * Zapping usage information
77
    zapUsageDemand, zapDmdEnvSig, zapUsedOnceDemand, zapUsedOnceSig
78
  ) where
sof's avatar
sof committed
79

80 81
#include "HsVersions.h"

82
import GHC.Prelude
83

84
import GHC.Types.Var ( Var, Id )
Sylvain Henry's avatar
Sylvain Henry committed
85
import GHC.Types.Var.Env
86
import GHC.Types.Var.Set
Sylvain Henry's avatar
Sylvain Henry committed
87 88
import GHC.Types.Unique.FM
import GHC.Types.Basic
89
import GHC.Data.Maybe   ( orElse )
90

Sylvain Henry's avatar
Sylvain Henry committed
91 92 93
import GHC.Core.Type    ( Type )
import GHC.Core.TyCon   ( isNewTyCon, isClassTyCon )
import GHC.Core.DataCon ( splitDataProductType_maybe )
Krzysztof Gogolewski's avatar
Krzysztof Gogolewski committed
94
import GHC.Core.Multiplicity    ( scaledThing )
95

96 97 98 99 100
import GHC.Utils.Binary
import GHC.Utils.Misc
import GHC.Utils.Outputable
import GHC.Utils.Panic

Austin Seipp's avatar
Austin Seipp committed
101 102 103
{-
************************************************************************
*                                                                      *
104
           Card: Combining Strictness and Usage
105 106 107 108
*                                                                      *
************************************************************************
-}

109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132
{- Note [Evaluation cardinalities]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The demand analyser uses an /evaluation cardinality/ of type Card,
to specify how many times a term is evaluated.  A cardinality C_lu
represents an /interval/ [l..u], meaning
    C_lu means evaluated /at least/ 'l' times and
                         /at most/  'u' times

* The lower bound corresponds to /strictness/
  Hence 'l' is either 0 (lazy)
                   or 1 (strict)

* The upper bound corresponds to /usage/
  Hence 'u' is either 0 (not used at all),
                   or 1 (used at most once)
                   or n (no information)

Intervals describe sets, so the underlying lattice is the powerset lattice.

Usually l<=u, but we also have C_10, the interval [1,0], the empty interval,
denoting the empty set.   This is the bottom element of the lattice.
-}


133
-- | Describes an interval of /evaluation cardinalities/.
134
-- See Note [Evaluation cardinalities]
135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230
data Card
  = C_00 -- ^ {0}     Absent.
  | C_01 -- ^ {0,1}   Used at most once.
  | C_0N -- ^ {0,1,n} Every possible cardinality; the top element.
  | C_11 -- ^ {1}     Strict and used once.
  | C_1N -- ^ {1,n}   Strict and used (possibly) many times.
  | C_10 -- ^ {}      The empty interval; the bottom element of the lattice.
  deriving Eq

_botCard, topCard :: Card
_botCard = C_10
topCard = C_0N

-- | True <=> lower bound is 1.
isStrict :: Card -> Bool
isStrict C_10 = True
isStrict C_11 = True
isStrict C_1N = True
isStrict _    = False

-- | True <=> upper bound is 0.
isAbs :: Card -> Bool
isAbs C_00 = True
isAbs C_10 = True -- Bottom cardinality is also absent
isAbs _    = False

-- | True <=> upper bound is 1.
isUsedOnce :: Card -> Bool
isUsedOnce C_0N = False
isUsedOnce C_1N = False
isUsedOnce _    = True

-- | Intersect with [0,1].
oneifyCard :: Card -> Card
oneifyCard C_0N = C_01
oneifyCard C_1N = C_11
oneifyCard c    = c

-- | Denotes '∪' on 'Card'.
lubCard :: Card -> Card -> Card
-- Handle C_10 (bot)
lubCard C_10 n    = n    -- bot
lubCard n    C_10 = n    -- bot
-- Handle C_0N (top)
lubCard C_0N _    = C_0N -- top
lubCard _    C_0N = C_0N -- top
-- Handle C_11
lubCard C_00 C_11 = C_01 -- {0} ∪ {1} = {0,1}
lubCard C_11 C_00 = C_01 -- {0} ∪ {1} = {0,1}
lubCard C_11 n    = n    -- {1} is a subset of all other intervals
lubCard n    C_11 = n    -- {1} is a subset of all other intervals
-- Handle C_1N
lubCard C_1N C_1N = C_1N -- reflexivity
lubCard _    C_1N = C_0N -- {0} ∪ {1,n} = top
lubCard C_1N _    = C_0N -- {0} ∪ {1,n} = top
-- Handle C_01
lubCard C_01 _    = C_01 -- {0} ∪ {0,1} = {0,1}
lubCard _    C_01 = C_01 -- {0} ∪ {0,1} = {0,1}
-- Handle C_00
lubCard C_00 C_00 = C_00 -- reflexivity

-- | Denotes '+' on 'Card'.
plusCard :: Card -> Card -> Card
-- Handle C_00
plusCard C_00 n    = n    -- {0}+n = n
plusCard n    C_00 = n    -- {0}+n = n
-- Handle C_10
plusCard C_10 C_01 = C_11 -- These follow by applying + to lower and upper
plusCard C_10 C_0N = C_1N -- bounds individually
plusCard C_10 n    = n
plusCard C_01 C_10 = C_11
plusCard C_0N C_10 = C_1N
plusCard n    C_10 = n
-- Handle the rest (C_01, C_0N, C_11, C_1N)
plusCard C_01 C_01 = C_0N -- The upper bound is at least 1, so upper bound of
plusCard C_01 C_0N = C_0N -- the result must be 1+1 ~= N.
plusCard C_0N C_01 = C_0N -- But for the lower bound we have 4 cases where
plusCard C_0N C_0N = C_0N -- 0+0 ~= 0 (as opposed to 1), so we match on these.
plusCard _    _    = C_1N -- Otherwise we return {1,n}

-- | Denotes '*' on 'Card'.
multCard :: Card -> Card -> Card
-- Handle C_11 (neutral element)
multCard C_11 c    = c
multCard c    C_11 = c
-- Handle C_00 (annihilating element)
multCard C_00 _    = C_00
multCard _    C_00 = C_00
-- Handle C_10
multCard C_10 c    = if isStrict c then C_10 else C_00
multCard c    C_10 = if isStrict c then C_10 else C_00
-- Handle reflexive C_1N, C_01
multCard C_1N C_1N = C_1N
multCard C_01 C_01 = C_01
-- Handle C_0N and the rest (C_01, C_1N):
multCard _    _    = C_0N
231 232 233 234

{-
************************************************************************
*                                                                      *
235
           Demand: Evaluation contexts
Austin Seipp's avatar
Austin Seipp committed
236 237
*                                                                      *
************************************************************************
238
-}
239

240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290
-- | A demand describes a /scaled evaluation context/, e.g. how many times
-- and how deep the denoted thing is evaluated.
--
-- The "how many" component is represented by a 'Card'inality.
-- The "how deep" component is represented by a 'SubDemand'.
-- Examples (using Note [Demand notation]):
--
--   * 'seq' puts demand @SA@ on its argument: It evaluates the argument
--     strictly (@S@), but not any deeper (@A@).
--   * 'fst' puts demand @SP(SU,A)@ on its argument: It evaluates the argument
--     pair strictly and the first component strictly, but no nested info
--     beyond that (@U@). Its second argument is not used at all.
--   * '$' puts demand @SCS(U)@ on its first argument: It calls (@C@) the
--     argument function with one argument, exactly once (@S@). No info
--     on how the result of that call is evaluated (@U@).
--   * 'maybe' puts demand @1C1(U)@ on its second argument: It evaluates
--     the argument function lazily and calls it once when it is evaluated.
--   * @fst p + fst p@ puts demand @MP(MU,A)@ on @p@: It's @SP(SU,A)@
--     multiplied by two, so we get @M@ (used at least once, possibly multiple
--     times).
--
-- This data type is quite similar to @'Scaled' 'SubDemand'@, but it's scaled
-- by 'Card', which is an /interval/ on 'Multiplicity', the upper bound of
-- which could be used to infer uniqueness types.
data Demand
  = !Card :* !SubDemand
  deriving Eq

-- | A sub-demand describes an /evaluation context/, e.g. how deep the
-- denoted thing is evaluated. See 'Demand' for examples.
--
-- The nested 'SubDemand' @d@ of a 'Call' @Cn(d)@ is /relative/ to a single such call.
-- E.g. The expression @f 1 2 + f 3 4@ puts call demand @MCM(CS(U))@ on @f@:
-- @f@ is called exactly twice (@M@), each time exactly once (@S@) with an
-- additional argument.
--
-- The nested 'Demand's @dn@ of a 'Prod' @P(d1,d2,...)@ apply /absolutely/:
-- If @dn@ is a used once demand (cf. 'isUsedOnce'), then that means that
-- the denoted sub-expression is used once in the entire evaluation context
-- described by the surrounding 'Demand'. E.g., @UP(1U)@ means that the
-- field of the denoted expression is used at most once, although the
-- entire expression might be used many times.
--
-- See Note [Call demands are relative]
-- and Note [Demand notation].
data SubDemand
  = Poly !Card
  -- ^ Polymorphic demand, the denoted thing is evaluated arbitrarily deep,
  -- with the specified cardinality at every level.
  -- Expands to 'Call' via 'viewCall' and to 'Prod' via 'viewProd'.
  --
291 292 293 294 295
  -- @Poly n@ is semantically equivalent to @Prod [n :* Poly n, ...]@ or
  -- @Call n (Poly n)@. 'mkCall' and 'mkProd' do these rewrites.
  --
  -- In Note [Demand notation]: @U === P(U,U,...)@ and @U === CU(U)@,
  --                            @S === P(S,S,...)@ and @S === CS(S)@, and so on.
296 297 298 299 300 301 302 303
  --
  -- We only really use 'Poly' with 'C_10' (bottom), 'C_00' (absent),
  -- 'C_0N' (top) and sometimes 'C_1N', but it's simpler to treat it uniformly
  -- than to have a special constructor for each of the three cases.
  | Call !Card !SubDemand
  -- ^ @Call n sd@ describes the evaluation context of @n@ function
  -- applications, where every individual result is evaluated according to @sd@.
  -- @sd@ is /relative/ to a single call, cf. Note [Call demands are relative].
304 305
  -- Used only for values of function type. Use the smart constructor 'mkCall'
  -- whenever possible!
306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332
  | Prod ![Demand]
  -- ^ @Prod ds@ describes the evaluation context of a case scrutinisation
  -- on an expression of product type, where the product components are
  -- evaluated according to @ds@.
  deriving Eq

poly00, poly01, poly0N, poly11, poly1N, poly10 :: SubDemand
topSubDmd, botSubDmd, seqSubDmd :: SubDemand
poly00 = Poly C_00
poly01 = Poly C_01
poly0N = Poly C_0N
poly11 = Poly C_11
poly1N = Poly C_1N
poly10 = Poly C_10
topSubDmd = poly0N
botSubDmd = poly10
seqSubDmd = poly00

polyDmd :: Card -> Demand
polyDmd C_00 = C_00 :* poly00
polyDmd C_01 = C_01 :* poly01
polyDmd C_0N = C_0N :* poly0N
polyDmd C_11 = C_11 :* poly11
polyDmd C_1N = C_1N :* poly1N
polyDmd C_10 = C_10 :* poly10

-- | A smart constructor for 'Prod', applying rewrite rules along the semantic
333
-- equality @Prod [polyDmd n, ...] === polyDmd n@, simplifying to 'Poly'
334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361
-- 'SubDemand's when possible. Note that this degrades boxity information! E.g. a
-- polymorphic demand will never unbox.
mkProd :: [Demand] -> SubDemand
mkProd [] = seqSubDmd
mkProd ds@(n:*sd : _)
  | want_to_simplify n, all (== polyDmd n) ds = sd
  | otherwise                                 = Prod ds
  where
    -- We only want to simplify absent and bottom demands and unbox the others.
    -- See also Note [U should win] and Note [Don't optimise UP(U,U,...) to U].
    want_to_simplify C_00 = True
    want_to_simplify C_10 = True
    want_to_simplify _    = False

-- | @viewProd n sd@ interprets @sd@ as a 'Prod' of arity @n@, expanding 'Poly'
-- demands as necessary.
viewProd :: Arity -> SubDemand -> Maybe [Demand]
-- It's quite important that this function is optimised well;
-- it is used by lubSubDmd and plusSubDmd. Note the strict
-- application to 'polyDmd':
viewProd n (Prod ds)   | ds `lengthIs` n = Just ds
-- Note the strict application to replicate: This makes sure we don't allocate
-- a thunk for it, inlines it and lets case-of-case fire at call sites.
viewProd n (Poly card)                   = Just (replicate n $! polyDmd card)
viewProd _ _                             = Nothing
{-# INLINE viewProd #-} -- we want to fuse away the replicate and the allocation
                        -- for Arity. Otherwise, #18304 bites us.

362 363 364 365 366 367 368
-- | A smart constructor for 'Call', applying rewrite rules along the semantic
-- equality @Call n (Poly n) === Poly n@, simplifying to 'Poly' 'SubDemand's
-- when possible.
mkCall :: Card -> SubDemand -> SubDemand
mkCall n cd@(Poly m) | n == m = cd
mkCall n cd                   = Call n cd

369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
-- | @viewCall sd@ interprets @sd@ as a 'Call', expanding 'Poly' demands as
-- necessary.
viewCall :: SubDemand -> Maybe (Card, SubDemand)
viewCall (Call n sd)    = Just (n, sd)
viewCall sd@(Poly card) = Just (card, sd)
viewCall _              = Nothing

topDmd, absDmd, botDmd, seqDmd :: Demand
topDmd = polyDmd C_0N
absDmd = polyDmd C_00
botDmd = polyDmd C_10
seqDmd = C_11 :* seqSubDmd

-- | Denotes '∪' on 'SubDemand'.
lubSubDmd :: SubDemand -> SubDemand -> SubDemand
-- Handle Prod
lubSubDmd (Prod ds1) (viewProd (length ds1) -> Just ds2) =
  Prod $ zipWith lubDmd ds2 ds1 -- try to fuse with ds2
-- Handle Call
lubSubDmd (Call n1 d1) (viewCall -> Just (n2, d2))
  -- See Note [Call demands are relative]
390 391 392
  | isAbs n1  = mkCall (lubCard n1 n2) (lubSubDmd botSubDmd d2)
  | isAbs n2  = mkCall (lubCard n1 n2) (lubSubDmd d1 botSubDmd)
  | otherwise = mkCall (lubCard n1 n2) (lubSubDmd d1        d2)
393 394 395 396 397 398 399 400 401
-- Handle Poly
lubSubDmd (Poly n1)  (Poly n2) = Poly (lubCard n1 n2)
-- Make use of reflexivity (so we'll match the Prod or Call cases again).
lubSubDmd sd1@Poly{} sd2       = lubSubDmd sd2 sd1
-- Otherwise (Call `lub` Prod) return Top
lubSubDmd _          _         = topSubDmd

-- | Denotes '∪' on 'Demand'.
lubDmd :: Demand -> Demand -> Demand
402
lubDmd (n1 :* sd1) (n2 :* sd2) = lubCard n1 n2 :* lubSubDmd sd1 sd2
403 404 405 406 407 408 409 410 411

-- | Denotes '+' on 'SubDemand'.
plusSubDmd :: SubDemand -> SubDemand -> SubDemand
-- Handle Prod
plusSubDmd (Prod ds1) (viewProd (length ds1) -> Just ds2) =
  Prod $ zipWith plusDmd ds2 ds1 -- try to fuse with ds2
-- Handle Call
plusSubDmd (Call n1 d1) (viewCall -> Just (n2, d2))
  -- See Note [Call demands are relative]
412 413 414
  | isAbs n1  = mkCall (plusCard n1 n2) (lubSubDmd botSubDmd d2)
  | isAbs n2  = mkCall (plusCard n1 n2) (lubSubDmd d1 botSubDmd)
  | otherwise = mkCall (plusCard n1 n2) (lubSubDmd d1        d2)
415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442
-- Handle Poly
plusSubDmd (Poly n1)  (Poly n2) = Poly (plusCard n1 n2)
-- Make use of reflexivity (so we'll match the Prod or Call cases again).
plusSubDmd sd1@Poly{} sd2       = plusSubDmd sd2 sd1
-- Otherwise (Call `lub` Prod) return Top
plusSubDmd _          _         = topSubDmd

-- | Denotes '+' on 'Demand'.
plusDmd :: Demand -> Demand -> Demand
plusDmd (n1 :* sd1) (n2 :* sd2) = plusCard n1 n2 :* plusSubDmd sd1 sd2

-- | The trivial cases of the @mult*@ functions.
-- If @multTrivial n abs a = ma@, we have the following outcomes
-- depending on @n@:
--
--   * 'C_11' => multiply by one, @ma = Just a@
--   * 'C_00', 'C_10' (e.g. @'isAbs' n@) => return the absent thing,
--      @ma = Just abs@
--   * Otherwise ('C_01', 'C_*N') it's not a trivial case, @ma = Nothing@.
multTrivial :: Card -> a -> a -> Maybe a
multTrivial C_11 _   a           = Just a
multTrivial n    abs _ | isAbs n = Just abs
multTrivial _    _   _           = Nothing

multSubDmd :: Card -> SubDemand -> SubDemand
multSubDmd n sd
  | Just sd' <- multTrivial n seqSubDmd sd = sd'
multSubDmd n (Poly n')    = Poly (multCard n n')
443
multSubDmd n (Call n' sd) = mkCall (multCard n n') sd -- See Note [Call demands are relative]
444 445 446 447 448 449 450 451 452 453
multSubDmd n (Prod ds)    = Prod (map (multDmd n) ds)

multDmd :: Card -> Demand -> Demand
multDmd n    dmd
  | Just dmd' <- multTrivial n absDmd dmd = dmd'
multDmd n (m :* dmd) = multCard n m :* multSubDmd n dmd

-- | Used to suppress pretty-printing of an uninformative demand
isTopDmd :: Demand -> Bool
isTopDmd dmd = dmd == topDmd
454

455 456
isAbsDmd :: Demand -> Bool
isAbsDmd (n :* _) = isAbs n
457

458 459 460 461
-- | Contrast with isStrictUsedDmd. See Note [Strict demands]
isStrictDmd :: Demand -> Bool
isStrictDmd (n :* _) = isStrict n

462 463 464
-- | Not absent and used strictly. See Note [Strict demands]
isStrUsedDmd :: Demand -> Bool
isStrUsedDmd (n :* _) = isStrict n && not (isAbs n)
465

466 467 468 469
isSeqDmd :: Demand -> Bool
isSeqDmd (C_11 :* sd) = sd == seqSubDmd
isSeqDmd (C_1N :* sd) = sd == seqSubDmd -- I wonder if we need this case.
isSeqDmd _            = False
470

471 472 473
-- | Is the value used at most once?
isUsedOnceDmd :: Demand -> Bool
isUsedOnceDmd (n :* _) = isUsedOnce n
474

475 476 477 478 479 480 481 482 483 484 485 486 487 488 489
-- | We try to avoid tracking weak free variable demands in strictness
-- signatures for analysis performance reasons.
-- See Note [Lazy and unleashable free variables] in "GHC.Core.Opt.DmdAnal".
isWeakDmd :: Demand -> Bool
isWeakDmd dmd@(n :* _) = not (isStrict n) && is_plus_idem_dmd dmd
  where
    -- @is_plus_idem_* thing@ checks whether @thing `plus` thing = thing@,
    -- e.g. if @thing@ is idempotent wrt. to @plus@.
    is_plus_idem_card c = plusCard c c == c
    -- is_plus_idem_dmd dmd = plusDmd dmd dmd == dmd
    is_plus_idem_dmd (n :* sd) = is_plus_idem_card n && is_plus_idem_sub_dmd sd
    -- is_plus_idem_sub_dmd sd = plusSubDmd sd sd == sd
    is_plus_idem_sub_dmd (Poly n)   = is_plus_idem_card n
    is_plus_idem_sub_dmd (Prod ds)  = all is_plus_idem_dmd ds
    is_plus_idem_sub_dmd (Call n _) = is_plus_idem_card n -- See Note [Call demands are relative]
490

491 492
evalDmd :: Demand
evalDmd = C_1N :* topSubDmd
493

494 495 496
-- | First argument of 'GHC.Exts.maskAsyncExceptions#': @SCS(U)@.
-- Called exactly once.
strictOnceApply1Dmd :: Demand
497
strictOnceApply1Dmd = C_11 :* mkCall C_11 topSubDmd
498

499 500 501
-- | First argument of 'GHC.Exts.atomically#': @MCM(U)@.
-- Called at least once, possibly many times.
strictManyApply1Dmd :: Demand
502
strictManyApply1Dmd = C_1N :* mkCall C_1N topSubDmd
503

504 505 506
-- | First argument of catch#: @1C1(U)@.
-- Evaluates its arg lazily, but then applies it exactly once to one argument.
lazyApply1Dmd :: Demand
507
lazyApply1Dmd = C_01 :* mkCall C_01 topSubDmd
508

509 510 511
-- | Second argument of catch#: @1C1(CS(U))@.
-- Calls its arg lazily, but then applies it exactly once to an additional argument.
lazyApply2Dmd :: Demand
512
lazyApply2Dmd = C_01 :* mkCall C_01 (mkCall C_11 topSubDmd)
513

514 515 516
-- | Make a 'Demand' evaluated at-most-once.
oneifyDmd :: Demand -> Demand
oneifyDmd (n :* sd) = oneifyCard n :* sd
517

518 519 520
-- | Make a 'Demand' evaluated at-least-once (e.g. strict).
strictifyDmd :: Demand -> Demand
strictifyDmd (n :* sd) = plusCard C_10 n :* sd
521

522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551
-- | If the argument is a used non-newtype dictionary, give it strict demand.
-- Also split the product type & demand and recur in order to similarly
-- strictify the argument's contained used non-newtype superclass dictionaries.
-- We use the demand as our recursive measure to guarantee termination.
strictifyDictDmd :: Type -> Demand -> Demand
strictifyDictDmd ty (n :* Prod ds)
  | not (isAbs n)
  , Just field_tys <- as_non_newtype_dict ty
  = C_1N :* -- main idea: ensure it's strict
      if all (not . isAbsDmd) ds
        then topSubDmd -- abstract to strict w/ arbitrary component use,
                         -- since this smells like reboxing; results in CBV
                         -- boxed
                         --
                         -- TODO revisit this if we ever do boxity analysis
        else Prod (zipWith strictifyDictDmd field_tys ds)
  where
    -- | Return a TyCon and a list of field types if the given
    -- type is a non-newtype dictionary type
    as_non_newtype_dict ty
      | Just (tycon, _arg_tys, _data_con, map scaledThing -> inst_con_arg_tys)
          <- splitDataProductType_maybe ty
      , not (isNewTyCon tycon)
      , isClassTyCon tycon
      = Just inst_con_arg_tys
      | otherwise
      = Nothing
strictifyDictDmd _  dmd = dmd

-- | Wraps the 'SubDemand' with a one-shot call demand: @d@ -> @CS(d)@.
552 553
mkCalledOnceDmd :: SubDemand -> SubDemand
mkCalledOnceDmd sd = mkCall C_11 sd
554

555 556 557
-- | @mkCalledOnceDmds n d@ returns @CS(CS...(CS d))@ where there are @n@ @CS@'s.
mkCalledOnceDmds :: Arity -> SubDemand -> SubDemand
mkCalledOnceDmds arity sd = iterate mkCalledOnceDmd sd !! arity
558 559 560 561 562 563 564 565 566 567 568 569 570 571 572

-- | Peels one call level from the sub-demand, and also returns how many
-- times we entered the lambda body.
peelCallDmd :: SubDemand -> (Card, SubDemand)
peelCallDmd sd = viewCall sd `orElse` (topCard, topSubDmd)

-- Peels multiple nestings of 'Call' sub-demands and also returns
-- whether it was unsaturated in the form of a 'Card'inality, denoting
-- how many times the lambda body was entered.
-- See Note [Demands from unsaturated function calls].
peelManyCalls :: Int -> SubDemand -> Card
peelManyCalls 0 _                          = C_11
-- See Note [Call demands are relative]
peelManyCalls n (viewCall -> Just (m, sd)) = m `multCard` peelManyCalls (n-1) sd
peelManyCalls _ _                          = C_0N
573

574 575 576 577 578
-- See Note [Demand on the worker] in GHC.Core.Opt.WorkWrap
mkWorkerDemand :: Int -> Demand
mkWorkerDemand n = C_01 :* go n
  where go 0 = topSubDmd
        go n = Call C_01 $ go (n-1)
579

580 581 582 583 584 585 586 587 588
addCaseBndrDmd :: SubDemand -- On the case binder
               -> [Demand]  -- On the components of the constructor
               -> [Demand]  -- Final demands for the components of the constructor
addCaseBndrDmd (Poly n) alt_dmds
  | isAbs n   = alt_dmds
-- See Note [Demand on case-alternative binders]
addCaseBndrDmd sd       alt_dmds = zipWith plusDmd ds alt_dmds -- fuse ds!
  where
    Just ds = viewProd (length alt_dmds) sd -- Guaranteed not to be a call
589

590 591 592 593 594 595 596
argsOneShots :: StrictSig -> Arity -> [[OneShotInfo]]
-- ^ See Note [Computing one-shot info]
argsOneShots (StrictSig (DmdType _ arg_ds _)) n_val_args
  | unsaturated_call = []
  | otherwise = go arg_ds
  where
    unsaturated_call = arg_ds `lengthExceeds` n_val_args
597

598 599
    go []               = []
    go (arg_d : arg_ds) = argOneShots arg_d `cons` go arg_ds
600

601 602 603
    -- Avoid list tail like [ [], [], [] ]
    cons [] [] = []
    cons a  as = a:as
604

605 606 607 608
argOneShots :: Demand          -- ^ depending on saturation
            -> [OneShotInfo]
-- ^ See Note [Computing one-shot info]
argOneShots (_ :* sd) = go sd -- See Note [Call demands are relative]
609
  where
610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627
    go (Call n sd)
      | isUsedOnce n = OneShotLam    : go sd
      | otherwise    = NoOneShotInfo : go sd
    go _    = []

-- |
-- @saturatedByOneShots n C1(C1(...)) = True@
--   <=>
-- There are at least n nested C1(..) calls.
-- See Note [Demand on the worker] in GHC.Core.Opt.WorkWrap
saturatedByOneShots :: Int -> Demand -> Bool
saturatedByOneShots n (_ :* sd) = isUsedOnce (peelManyCalls n sd)

{- Note [Strict demands]
~~~~~~~~~~~~~~~~~~~~~~~~
'isStrUsedDmd' returns true only of demands that are
   both strict
   and  used
628 629 630

In particular, it is False for <B> (i.e. strict and not used,
cardinality C_10), which can and does arise in, say (#7319)
631 632 633 634 635 636 637 638 639 640 641
   f x = raise# <some exception>
Then 'x' is not used, so f gets strictness <B> -> .
Now the w/w generates
   fx = let x <B> = absentError "unused"
        in raise <some exception>
At this point we really don't want to convert to
   fx = case absentError "unused" of x -> raise <some exception>
Since the program is going to diverge, this swaps one error for another,
but it's really a bad idea to *ever* evaluate an absent argument.
In #7319 we get
   T7319.exe: Oops!  Entered absent arg w_s1Hd{v} [lid] [base:GHC.Base.String{tc 36u}]
642

643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681
Note [Call demands are relative]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The expression @if b then 0 else f 1 2 + f 3 4@ uses @f@ according to the demand
@UCU(CS(P(U)))@, meaning

  "f is called multiple times or not at all (CU), but each time it
   is called, it's called with *exactly one* (CS) more argument.
   Whenever it is called with two arguments, we have no info on how often
   the field of the product result is used (U)."

So the 'SubDemand' nested in a 'Call' demand is relative to exactly one call.
And that extends to the information we have how its results are used in each
call site. Consider (#18903)

  h :: Int -> Int
  h m =
    let g :: Int -> (Int,Int)
        g 1 = (m, 0)
        g n = (2 * n, 2 `div` n)
        {-# NOINLINE g #-}
    in case m of
      1 -> 0
      2 -> snd (g m)
      _ -> uncurry (+) (g m)

We want to give @g@ the demand @1C1(P(1P(U),SP(U)))@, so we see that in each call
site of @g@, we are strict in the second component of the returned pair.

This relative cardinality leads to an otherwise unexpected call to 'lubSubDmd'
in 'plusSubDmd', but if you do the math it's just the right thing.

There's one more subtlety: Since the nested demand is relative to exactly one
call, in the case where we have *at most zero calls* (e.g. CA(...)), the premise
is hurt and we can assume that the nested demand is 'botSubDmd'. That ensures
that @g@ above actually gets the @SP(U)@ demand on its second pair component,
rather than the lazy @1P(U)@ if we 'lub'bed with an absent demand.

Demand on case-alternative binders]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
682 683 684
The demand on a binder in a case alternative comes
  (a) From the demand on the binder itself
  (b) From the demand on the case binder
685
Forgetting (b) led directly to #10148.
686 687 688 689 690 691 692 693

Example. Source code:
  f x@(p,_) = if p then foo x else True

  foo (p,True) = True
  foo (p,q)    = foo (q,p)

After strictness analysis:
694
  f = \ (x_an1 [Dmd=<SP(SL),1*UP(U,1*U)>] :: (Bool, Bool)) ->
695
      case x_an1
696
      of wild_X7 [Dmd=<L,1*UP(1*U,1*U)>]
697 698 699 700 701
      { (p_an2 [Dmd=<S,1*U>], ds_dnz [Dmd=<L,A>]) ->
      case p_an2 of _ {
        False -> GHC.Types.True;
        True -> foo wild_X7 }

702
It's true that ds_dnz is *itself* absent, but the use of wild_X7 means
703
that it is very much alive and demanded.  See #10148 for how the
704 705 706 707 708
consequences play out.

This is needed even for non-product types, in case the case-binder
is used but the components of the case alternative are not.

709
Note [Don't optimise UP(U,U,...) to U]
710
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
711 712
These two SubDemands:
   UP(U,U) (@Prod [topDmd, topDmd]@)   and   U (@topSubDmd@)
713
are semantically equivalent, but we do not turn the former into
714 715 716 717 718 719 720 721 722 723
the latter, for a regrettable-subtle reason.  Consider
  f p1@(x,y) = (y,x)
  g h p2@(_,_) = h p
We want to unbox @p1@ of @f@, but not @p2@ of @g@, because @g@ only uses
@p2@ boxed and we'd have to rebox. So we give @p1@ demand UP(U,U) and @p2@
demand @U@ to inform 'GHC.Core.Opt.WorkWrap.Utils.wantToUnbox', which will
say "unbox" for @p1@ and "don't unbox" for @p2@.

So the solution is: don't aggressively collapse @Prod [topDmd, topDmd]@ to
@topSubDmd@; instead leave it as-is. In effect we are using the UseDmd to do a
724 725
little bit of boxity analysis.  Not very nice.

726 727 728 729 730 731 732 733
Note [U should win]
~~~~~~~~~~~~~~~~~~~
Both in 'lubSubDmd' and 'plusSubDmd' we want @U `plusSubDmd` UP(..)) to be @U@.
Why?  Because U carries the implication the whole thing is used, box and all,
so we don't want to w/w it, cf. Note [Don't optimise UP(U,U,...) to U].
If we use it both boxed and unboxed, then we are definitely using the box,
and so we are quite likely to pay a reboxing cost. So we make U win here.
TODO: Investigate why since 2013, we don't.
734 735 736 737 738 739 740 741 742 743 744

Example is in the Buffer argument of GHC.IO.Handle.Internals.writeCharBuffer

Baseline: (A) Not making Used win (UProd wins)
Compare with: (B) making Used win for lub and both

            Min          -0.3%     -5.6%    -10.7%    -11.0%    -33.3%
            Max          +0.3%    +45.6%    +11.5%    +11.5%     +6.9%
 Geometric Mean          -0.0%     +0.5%     +0.3%     +0.2%     -0.8%

Baseline: (B) Making Used win for both lub and both
745
Compare with: (C) making Used win for plus, but UProd win for lub
746 747 748 749 750

            Min          -0.1%     -0.3%     -7.9%     -8.0%     -6.5%
            Max          +0.1%     +1.0%    +21.0%    +21.0%     +0.5%
 Geometric Mean          +0.0%     +0.0%     -0.0%     -0.1%     -0.1%

751 752 753 754 755 756 757 758 759 760 761
Note [Computing one-shot info]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider a call
    f (\pqr. e1) (\xyz. e2) e3
where f has usage signature
    C1(C(C1(U))) C1(U) U
Then argsOneShots returns a [[OneShotInfo]] of
    [[OneShot,NoOneShotInfo,OneShot],  [OneShot]]
The occurrence analyser propagates this one-shot infor to the
binders \pqr and \xyz;
see Note [Use one-shot information] in "GHC.Core.Opt.OccurAnal".
762 763
-}

764
{- *********************************************************************
Austin Seipp's avatar
Austin Seipp committed
765
*                                                                      *
766
                 Divergence: Whether evaluation surely diverges
Austin Seipp's avatar
Austin Seipp committed
767
*                                                                      *
768
********************************************************************* -}
769

770 771 772
-- | 'Divergence' characterises whether something surely diverges.
-- Models a subset lattice of the following exhaustive set of divergence
-- results:
773 774 775
--
-- [n] nontermination (e.g. loops)
-- [i] throws imprecise exception
776 777
-- [p] throws precise exceTtion
-- [c] converges (reduces to WHNF).
778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793
--
-- The different lattice elements correspond to different subsets, indicated by
-- juxtaposition of indicators (e.g. __nc__ definitely doesn't throw an
-- exception, and may or may not reduce to WHNF).
--
-- @
--             Dunno (nipc)
--                  |
--            ExnOrDiv (nip)
--                  |
--            Diverges (ni)
-- @
--
-- As you can see, we don't distinguish __n__ and __i__.
-- See Note [Precise exceptions and strictness analysis] for why __p__ is so
-- special compared to __i__.
794
data Divergence
795 796 797 798 799
  = Diverges -- ^ Definitely throws an imprecise exception or diverges.
  | ExnOrDiv -- ^ Definitely throws a *precise* exception, an imprecise
             --   exception or diverges. Never converges, hence 'isDeadEndDiv'!
             --   See scenario 1 in Note [Precise exceptions and strictness analysis].
  | Dunno    -- ^ Might diverge, throw any kind of exception or converge.
800
  deriving Eq
801

802 803 804 805 806 807 808 809
lubDivergence :: Divergence -> Divergence -> Divergence
lubDivergence Diverges div      = div
lubDivergence div      Diverges = div
lubDivergence ExnOrDiv ExnOrDiv = ExnOrDiv
lubDivergence _        _        = Dunno
-- This needs to commute with defaultFvDmd, i.e.
-- defaultFvDmd (r1 `lubDivergence` r2) = defaultFvDmd r1 `lubDmd` defaultFvDmd r2
-- (See Note [Default demand on free variables and arguments] for why)
810

811 812 813
-- | See Note [Asymmetry of 'plus*'], which concludes that 'plusDivergence'
-- needs to be symmetric.
-- Strictly speaking, we should have @plusDivergence Dunno Diverges = ExnOrDiv@.
814 815 816
-- But that regresses in too many places (every infinite loop, basically) to be
-- worth it and is only relevant in higher-order scenarios
-- (e.g. Divergence of @f (throwIO blah)@).
817 818 819 820 821 822 823 824 825 826 827 828
-- So 'plusDivergence' currently is 'glbDivergence', really.
plusDivergence :: Divergence -> Divergence -> Divergence
plusDivergence Dunno    Dunno    = Dunno
plusDivergence Diverges _        = Diverges
plusDivergence _        Diverges = Diverges
plusDivergence _        _        = ExnOrDiv

-- | In a non-strict scenario, we might not force the Divergence, in which case
-- we might converge, hence Dunno.
multDivergence :: Card -> Divergence -> Divergence
multDivergence n _ | not (isStrict n) = Dunno
multDivergence _ d                    = d
829

830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859
topDiv, exnDiv, botDiv :: Divergence
topDiv = Dunno
exnDiv = ExnOrDiv
botDiv = Diverges

-- | True if the 'Divergence' indicates that evaluation will not return.
-- See Note [Dead ends].
isDeadEndDiv :: Divergence -> Bool
isDeadEndDiv Diverges = True
isDeadEndDiv ExnOrDiv = True
isDeadEndDiv Dunno    = False

-- See Notes [Default demand on free variables and arguments]
-- and Scenario 1 in [Precise exceptions and strictness analysis]
defaultFvDmd :: Divergence -> Demand
defaultFvDmd Dunno    = absDmd
defaultFvDmd ExnOrDiv = absDmd -- This is the whole point of ExnOrDiv!
defaultFvDmd Diverges = botDmd -- Diverges

defaultArgDmd :: Divergence -> Demand
-- TopRes and BotRes are polymorphic, so that
--      BotRes === (Bot -> BotRes) === ...
--      TopRes === (Top -> TopRes) === ...
-- This function makes that concrete
-- Also see Note [Default demand on free variables and arguments]
defaultArgDmd Dunno    = topDmd
-- NB: not botDmd! We don't want to mask the precise exception by forcing the
-- argument. But it is still absent.
defaultArgDmd ExnOrDiv = absDmd
defaultArgDmd Diverges = botDmd
860

861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880
{- Note [Precise vs imprecise exceptions]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
An exception is considered to be /precise/ when it is thrown by the 'raiseIO#'
primop. It follows that all other primops (such as 'raise#' or
division-by-zero) throw /imprecise/ exceptions. Note that the actual type of
the exception thrown doesn't have any impact!

GHC undertakes some effort not to apply an optimisation that would mask a
/precise/ exception with some other source of nontermination, such as genuine
divergence or an imprecise exception, so that the user can reliably
intercept the precise exception with a catch handler before and after
optimisations.

See also the wiki page on precise exceptions:
https://gitlab.haskell.org/ghc/ghc/wikis/exceptions/precise-exceptions
Section 5 of "Tackling the awkward squad" talks about semantic concerns.
Imprecise exceptions are actually more interesting than precise ones (which are
fairly standard) from the perspective of semantics. See the paper "A Semantics
for Imprecise Exceptions" for more details.

881 882 883 884 885 886 887 888 889 890 891 892 893
Note [Dead ends]
~~~~~~~~~~~~~~~~
We call an expression that either diverges or throws a precise or imprecise
exception a "dead end". We used to call such an expression just "bottoming",
but with the measures we take to preserve precise exception semantics
(see Note [Precise exceptions and strictness analysis]), that is no longer
accurate: 'exnDiv' is no longer the bottom of the Divergence lattice.

Yet externally to demand analysis, we mostly care about being able to drop dead
code etc., which is all due to the property that such an expression never
returns, hence we consider throwing a precise exception to be a dead end.
See also 'isDeadEndDiv'.

894 895
Note [Precise exceptions and strictness analysis]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
896 897 898 899 900
We have to take care to preserve precise exception semantics in strictness
analysis (#17676). There are two scenarios that need careful treatment.

The fixes were discussed at
https://gitlab.haskell.org/ghc/ghc/wikis/fixing-precise-exceptions
901

902 903 904 905 906 907
Recall that raiseIO# raises a *precise* exception, in contrast to raise# which
raises an *imprecise* exception. See Note [Precise vs imprecise exceptions].

Scenario 1: Precise exceptions in case alternatives
---------------------------------------------------
Unlike raise# (which returns botDiv), we want raiseIO# to return exnDiv.
908
Here's why. Consider this example from #13380 (similarly #17676):
909 910 911
  f x y | x>0       = raiseIO# Exc
        | y>0       = return 1
        | otherwise = return 2
912 913 914
Is 'f' strict in 'y'? One might be tempted to say yes! But that plays fast and
loose with the precise exception; after optimisation, (f 42 (error "boom"))
turns from throwing the precise Exc to throwing the imprecise user error
915 916 917
"boom". So, the defaultFvDmd of raiseIO# should be lazy (topDmd), which can be
achieved by giving it divergence exnDiv.
See Note [Default demand on free variables and arguments].
918

919 920 921
Why don't we just give it topDiv instead of introducing exnDiv?
Because then the simplifier will fail to discard raiseIO#'s continuation in
  case raiseIO# x s of { (# s', r #) -> <BIG> }
922
which we'd like to optimise to
923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938
  case raiseIO# x s of {}
Hence we came up with exnDiv. The default FV demand of exnDiv is lazy (and
its default arg dmd is absent), but otherwise (in terms of 'isDeadEndDiv') it
behaves exactly as botDiv, so that dead code elimination works as expected.
This is tracked by T13380b.

Scenario 2: Precise exceptions in case scrutinees
-------------------------------------------------
Consider (more complete examples in #148, #1592, testcase strun003)

  case foo x s of { (# s', r #) -> y }

Is this strict in 'y'? Often not! If @foo x s@ might throw a precise exception
(ultimately via raiseIO#), then we must not force 'y', which may fail to
terminate or throw an imprecise exception, until we have performed @foo x s@.

939 940 941
So we have to 'deferAfterPreciseException' (which 'lub's with 'exnDmdType' to
model the exceptional control flow) when @foo x s@ may throw a precise
exception. Motivated by T13380{d,e,f}.
942
See Note [Which scrutinees may throw precise exceptions] in "GHC.Core.Opt.DmdAnal".
943

944 945 946 947 948 949 950 951
We have to be careful not to discard dead-end Divergence from case
alternatives, though (#18086):

  m = putStrLn "foo" >> error "bar"

'm' should still have 'exnDiv', which is why it is not sufficient to lub with
'nopDmdType' (which has 'topDiv') in 'deferAfterPreciseException'.

952 953 954 955 956 957 958 959
Historical Note: This used to be called the "IO hack". But that term is rather
a bad fit because
1. It's easily confused with the "State hack", which also affects IO.
2. Neither "IO" nor "hack" is a good description of what goes on here, which
   is deferring strictness results after possibly throwing a precise exception.
   The "hack" is probably not having to defer when we can prove that the
   expression may not throw a precise exception (increasing precision of the
   analysis), but that's just a favourable guess.
960

961 962 963 964
Note [Exceptions and strictness]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
We used to smart about catching exceptions, but we aren't anymore.
See #14998 for the way it's resolved at the moment.
965

966
Here's a historic breakdown:
967

968 969 970 971 972 973 974
Apparently, exception handling prim-ops didn't use to have any special
strictness signatures, thus defaulting to nopSig, which assumes they use their
arguments lazily. Joachim was the first to realise that we could provide richer
information. Thus, in 0558911f91c (Dec 13), he added signatures to
primops.txt.pp indicating that functions like `catch#` and `catchRetry#` call
their argument, which is useful information for usage analysis. Still with a
'Lazy' strictness demand (i.e. 'lazyApply1Dmd'), though, and the world was fine.
975

976 977 978 979
In 7c0fff4 (July 15), Simon argued that giving `catch#` et al. a
'strictApply1Dmd' leads to substantial performance gains. That was at the cost
of correctness, as #10712 proved. So, back to 'lazyApply1Dmd' in
28638dfe79e (Dec 15).
980

981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028
Motivated to reproduce the gains of 7c0fff4 without the breakage of #10712,
Ben opened #11222. Simon made the demand analyser "understand catch" in
9915b656 (Jan 16) by adding a new 'catchArgDmd', which basically said to call
its argument strictly, but also swallow any thrown exceptions in
'multDivergence'. This was realized by extending the 'Str' constructor of
'ArgStr' with a 'ExnStr' field, indicating that it catches the exception, and
adding a 'ThrowsExn' constructor to the 'Divergence' lattice as an element
between 'Dunno' and 'Diverges'. Then along came #11555 and finally #13330,
so we had to revert to 'lazyApply1Dmd' again in 701256df88c (Mar 17).

This left the other variants like 'catchRetry#' having 'catchArgDmd', which is
where #14998 picked up. Item 1 was concerned with measuring the impact of also
making `catchRetry#` and `catchSTM#` have 'lazyApply1Dmd'. The result was that
there was none. We removed the last usages of 'catchArgDmd' in 00b8ecb7
(Apr 18). There was a lot of dead code resulting from that change, that we
removed in ef6b283 (Jan 19): We got rid of 'ThrowsExn' and 'ExnStr' again and
removed any code that was dealing with the peculiarities.

Where did the speed-ups vanish to? In #14998, item 3 established that
turning 'catch#' strict in its first argument didn't bring back any of the
alleged performance benefits. Item 2 of that ticket finally found out that it
was entirely due to 'catchException's new (since #11555) definition, which
was simply

    catchException !io handler = catch io handler

While 'catchException' is arguably the saner semantics for 'catch', it is an
internal helper function in "GHC.IO". Its use in
"GHC.IO.Handle.Internals.do_operation" made for the huge allocation differences:
Remove the bang and you find the regressions we originally wanted to avoid with
'catchArgDmd'. See also #exceptions_and_strictness# in "GHC.IO".

So history keeps telling us that the only possibly correct strictness annotation
for the first argument of 'catch#' is 'lazyApply1Dmd', because 'catch#' really
is not strict in its argument: Just try this in GHCi

  :set -XScopedTypeVariables
  import Control.Exception
  catch undefined (\(_ :: SomeException) -> putStrLn "you'll see this")

Any analysis that assumes otherwise will be broken in some way or another
(beyond `-fno-pendantic-bottoms`).

But then #13380 and #17676 suggest (in Mar 20) that we need to re-introduce a
subtly different variant of `ThrowsExn` (which we call `ExnOrDiv` now) that is
only used by `raiseIO#` in order to preserve precise exceptions by strictness
analysis, while not impacting the ability to eliminate dead code.
See Note [Precise exceptions and strictness analysis].
1029

1030 1031
Note [Default demand on free variables and arguments]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1032 1033 1034
Free variables not mentioned in the environment of a 'DmdType'
are demanded according to the demand type's Divergence:
  * In a Diverges (botDiv) context, that demand is botDmd
1035 1036
    (strict and absent).
  * In all other contexts, the demand is absDmd (lazy and absent).
1037 1038 1039 1040 1041 1042 1043 1044 1045 1046
This is recorded in 'defaultFvDmd'.

Similarly, we can eta-expand demand types to get demands on excess arguments
not accounted for in the type, by consulting 'defaultArgDmd':
  * In a Diverges (botDiv) context, that demand is again botDmd.
  * In a ExnOrDiv (exnDiv) context, that demand is absDmd: We surely diverge
    before evaluating the excess argument, but don't want to eagerly evaluate
    it (cf. Note [Precise exceptions and strictness analysis]).
  * In a Dunno context (topDiv), the demand is topDmd, because
    it's perfectly possible to enter the additional lambda and evaluate it
1047
    in unforeseen ways (so, not absent).
Joachim Breitner's avatar
Joachim Breitner committed
1048

1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072
Note [Bottom CPR iff Dead-Ending Divergence]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Both CPR analysis and Demand analysis handle recursive functions by doing
fixed-point iteration. To find the *least* (e.g., most informative) fixed-point,
iteration starts with the bottom element of the semantic domain. Diverging
functions generally have the bottom element as their least fixed-point.

One might think that CPR analysis and Demand analysis then agree in when a
function gets a bottom denotation. E.g., whenever it has 'botCpr', it should
also have 'botDiv'. But that is not the case, because strictness analysis has to
be careful around precise exceptions, see Note [Precise vs imprecise exceptions].

So Demand analysis gives some diverging functions 'exnDiv' (which is *not* the
bottom element) when the CPR signature says 'botCpr', and that's OK. Here's an
example (from #18086) where that is the case:

ioTest :: IO ()
ioTest = do
  putStrLn "hi"
  undefined

However, one can loosely say that we give a function 'botCpr' whenever its
'Divergence' is 'exnDiv' or 'botDiv', i.e., dead-ending. But that's just
a consequence of fixed-point iteration, it's not important that they agree.
1073

Austin Seipp's avatar
Austin Seipp committed
1074 1075
************************************************************************
*                                                                      *
1076
           Demand environments and types
Austin Seipp's avatar
Austin Seipp committed
1077 1078 1079
*                                                                      *
************************************************************************
-}
1080

1081 1082
-- Subject to Note [Default demand on free variables and arguments]
type DmdEnv = VarEnv Demand
1083

1084 1085
emptyDmdEnv :: VarEnv Demand
emptyDmdEnv = emptyVarEnv
1086

1087 1088 1089 1090
multDmdEnv :: Card -> DmdEnv -> DmdEnv
multDmdEnv n env
  | Just env' <- multTrivial n emptyDmdEnv env = env'
  | otherwise                                  = mapVarEnv (multDmd n) env
1091

1092 1093
reuseEnv :: DmdEnv -> DmdEnv
reuseEnv = multDmdEnv C_1N
1094

1095 1096 1097 1098 1099 1100 1101 1102 1103 1104
-- | @keepAliveDmdType dt vs@ makes sure that the Ids in @vs@ have
-- /some/ usage in the returned demand types -- they are not Absent.
-- See Note [Absence analysis for stable unfoldings and RULES]
--     in "GHC.Core.Opt.DmdAnal".
keepAliveDmdEnv :: DmdEnv -> IdSet -> DmdEnv
keepAliveDmdEnv env vs
  = nonDetStrictFoldVarSet add env vs
  where
    add :: Id -> DmdEnv -> DmdEnv
    add v env = extendVarEnv_C add_dmd env v topDmd
1105

1106 1107 1108 1109 1110
    add_dmd :: Demand -> Demand -> Demand
    -- If the existing usage is Absent, make it used
    -- Otherwise leave it alone
    add_dmd dmd _ | isAbsDmd dmd = topDmd
                  | otherwise    = dmd
1111

1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122
-- | Characterises how an expression
--    * Evaluates its free variables ('dt_env')
--    * Evaluates its arguments ('dt_args')
--    * Diverges on every code path or not ('dt_div')
data DmdType
  = DmdType
  { dt_env  :: DmdEnv     -- ^ Demand on explicitly-mentioned free variables
  , dt_args :: [Demand]   -- ^ Demand on arguments
  , dt_div  :: Divergence -- ^ Whether evaluation diverges.
                          -- See Note [Demand type Divergence]
  }
1123

1124
instance Eq DmdType where
1125 1126
  (==) (DmdType fv1 ds1 div1)
       (DmdType fv2 ds2 div2) = nonDetUFMToList fv1 == nonDetUFMToList fv2
niteria's avatar
niteria committed
1127 1128 1129
         -- It's OK to use nonDetUFMToList here because we're testing for
         -- equality and even though the lists will be in some arbitrary
         -- Unique order, it is the same order for both
1130
                              && ds1 == ds2 && div1 == div2
1131

1132 1133
-- | Compute the least upper bound of two 'DmdType's elicited /by the same
-- incoming demand/!
1134
lubDmdType :: DmdType -> DmdType -> DmdType
1135
lubDmdType d1 d2
1136
  = DmdType lub_fv lub_ds lub_div
1137
  where
1138
    n = max (dmdTypeDepth d1) (dmdTypeDepth d2)
1139 1140
    (DmdType fv1 ds1 r1) = etaExpandDmdType n d1
    (DmdType fv2 ds2 r2) = etaExpandDmdType n d2
1141

1142
    lub_fv  = plusVarEnv_CD lubDmd fv1 (defaultFvDmd r1) fv2 (defaultFvDmd r2)
1143
    lub_ds  = zipWithEqual "lubDmdType" lubDmd ds1 ds2
1144
    lub_div = lubDivergence r1 r2
1145

1146
type PlusDmdArg = (DmdEnv, Divergence)
1147

1148 1149
mkPlusDmdArg :: DmdEnv -> PlusDmdArg
mkPlusDmdArg env = (env, topDiv)
1150

1151 1152
toPlusDmdArg :: DmdType -> PlusDmdArg
toPlusDmdArg (DmdType fv _ r) = (fv, r)
1153

1154 1155 1156 1157
plusDmdType :: DmdType -> PlusDmdArg -> DmdType
plusDmdType (DmdType fv1 ds1 r1) (fv2, t2)
    -- See Note [Asymmetry of 'plus*']
    -- 'plus' takes the argument/result info from its *first* arg,
1158
    -- using its second arg just for its free-var info.
1159
  = DmdType (plusVarEnv_CD plusDmd fv1 (defaultFvDmd r1) fv2 (defaultFvDmd t2))
1160
            ds1
1161
            (r1 `plusDivergence` t2)
1162

1163
botDmdType :: DmdType
1164
botDmdType = DmdType emptyDmdEnv [] botDiv
1165

1166 1167 1168 1169 1170
-- | The demand type of doing nothing (lazy, absent, no Divergence
-- information). Note that it is ''not'' the top of the lattice (which would be
-- "may use everything"), so it is (no longer) called topDmdType.
nopDmdType :: DmdType
nopDmdType = DmdType emptyDmdEnv [] topDiv
1171

1172 1173 1174
isTopDmdType :: DmdType -> Bool
isTopDmdType (DmdType env args div)
  = div == topDiv && null args && isEmptyVarEnv env
1175

1176 1177 1178 1179 1180
-- | The demand type of an unspecified expression that is guaranteed to
-- throw a (precise or imprecise) exception or diverge.
exnDmdType :: DmdType
exnDmdType = DmdType emptyDmdEnv [] exnDiv

1181
dmdTypeDepth :: DmdType -> Arity
1182
dmdTypeDepth = length . dt_args
1183

1184 1185 1186 1187
-- | This makes sure we can use the demand type with n arguments after eta
-- expansion, where n must not be lower than the demand types depth.
-- It appends the argument list with the correct 'defaultArgDmd'.
etaExpandDmdType :: Arity -> DmdType -> DmdType
1188
etaExpandDmdType n d@DmdType{dt_args = ds, dt_div = div}
1189
  | n == depth = d
1190
  | n >  depth = d{dt_args = inc_ds}
1191
  | otherwise  = pprPanic "etaExpandDmdType: arity decrease" (ppr n $$ ppr d)
1192
  where depth = length ds
1193 1194 1195 1196 1197 1198 1199
        -- Arity increase:
        --  * Demands on FVs are still valid
        --  * Demands on args also valid, plus we can extend with defaultArgDmd
        --    as appropriate for the given Divergence
        --  * Divergence is still valid:
        --    - A dead end after 2 arguments stays a dead end after 3 arguments
        --    - The remaining case is Dunno, which is already topDiv
1200
        inc_ds = take n (ds ++ repeat (defaultArgDmd div))
1201 1202 1203 1204 1205

-- | A conservative approximation for a given 'DmdType' in case of an arity
-- decrease. Currently, it's just nopDmdType.
decreaseArityDmdType :: DmdType -> DmdType
decreaseArityDmdType _ = nopDmdType
1206

1207 1208 1209 1210
splitDmdTy :: DmdType -> (Demand, DmdType)
-- Split off one function argument
-- We already have a suitable demand on all
-- free vars, so no need to add more!
1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234
splitDmdTy ty@DmdType{dt_args=dmd:args} = (dmd, ty{dt_args=args})
splitDmdTy ty@DmdType{dt_div=div}       = (defaultArgDmd div, ty)

multDmdType :: Card -> DmdType -> DmdType
multDmdType n (DmdType fv args res_ty)
  = -- pprTrace "multDmdType" (ppr n $$ ppr fv $$ ppr (multDmdEnv n fv)) $
    DmdType (multDmdEnv n fv)
            (map (multDmd n) args)
            (multDivergence n res_ty)

peelFV :: DmdType -> Var -> (DmdType, Demand)
peelFV (DmdType fv ds res) id = -- pprTrace "rfv" (ppr id <+> ppr dmd $$ ppr fv)
                               (DmdType fv' ds res, dmd)
  where
  fv' = fv `delVarEnv` id
  -- See Note [Default demand on free variables and arguments]
  dmd  = lookupVarEnv fv id `orElse` defaultFvDmd res

addDemand :: Demand -> DmdType -> DmdType
addDemand dmd (DmdType fv ds res) = DmdType fv (dmd:ds) res

findIdDemand :: DmdType -> Var -> Demand
findIdDemand (DmdType fv _ res) id
  = lookupVarEnv fv id `orElse` defaultFvDmd res
1235

1236
-- | When e is evaluated after executing an IO action that may throw a precise
1237 1238 1239 1240 1241 1242 1243 1244
-- exception, we act as if there is an additional control flow path that is
-- taken if e throws a precise exception. The demand type of this control flow
-- path
--   * is lazy and absent ('topDmd') in all free variables and arguments
--   * has 'exnDiv' 'Divergence' result
-- So we can simply take a variant of 'nopDmdType', 'exnDmdType'.
-- Why not 'nopDmdType'? Because then the result of 'e' can never be 'exnDiv'!
-- That means failure to drop dead-ends, see #18086.
1245 1246
-- See Note [Precise exceptions and strictness analysis]
deferAfterPreciseException :: DmdType -> DmdType
1247
deferAfterPreciseException = lubDmdType exnDmdType
1248

1249 1250 1251 1252 1253
-- | See 'keepAliveDmdEnv'.
keepAliveDmdType :: DmdType -> VarSet -> DmdType
keepAliveDmdType (DmdType fvs ds res) vars =
  DmdType (fvs `keepAliveDmdEnv` vars) ds res

1254 1255 1256 1257 1258 1259 1260
{-
Note [Demand type Divergence]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In contrast to StrictSigs, DmdTypes are elicited under a specific incoming demand.
This is described in detail in Note [Understanding DmdType and StrictSig].
Here, we'll focus on what that means for a DmdType's Divergence in a higher-order
scenario.
1261

1262 1263 1264 1265 1266 1267 1268 1269 1270
Consider
  err x y = x `seq` y `seq` error (show x)
this has a strictness signature of
  <SU><SU>b
meaning that we don't know what happens when we call err in weaker contexts than
CS(CS(U)), like @err `seq` ()@ (SU) and @err 1 `seq` ()@ (CS(U)). We
may not unleash the botDiv, hence assume topDiv. Of course, in
@err 1 2 `seq` ()@ the incoming demand CS(CS(S)) is strong enough and we see
that the expression diverges.
Joachim Breitner's avatar
Joachim Breitner committed
1271

1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297