2 % (c) The AQUA Project, Glasgow University, 1993-1998
4 \section[CoreMonad]{The core pipeline monad}
7 {-# LANGUAGE UndecidableInstances #-}
10 -- * Configuration of the core-to-core passes
14 getCoreToDo, dumpSimplPhase,
17 SimplCount, doSimplTick, doFreeSimplTick, simplCountN,
18 pprSimplCount, plusSimplCount, zeroSimplCount, isZeroSimplCount, Tick(..),
23 -- ** Reading from the monad
24 getHscEnv, getRuleBase, getModule,
25 getDynFlags, getOrigNameCache,
27 -- ** Writing to the monad
30 -- ** Lifting into the monad
31 liftIO, liftIOWithCount,
32 liftIO1, liftIO2, liftIO3, liftIO4,
34 -- ** Dealing with annotations
35 getAnnotations, getFirstAnnotations,
38 showPass, endPass, endIteration, dumpIfSet,
41 putMsg, putMsgS, errorMsg, errorMsgS,
42 fatalErrorMsg, fatalErrorMsgS,
43 debugTraceMsg, debugTraceMsgS,
58 import CoreLint ( lintCoreBindings )
59 import PrelNames ( iNTERACTIVE )
61 import Module ( Module )
64 import Rules ( RuleBase )
65 import BasicTypes ( CompilerPhase(..) )
69 import IOEnv hiding ( liftIO, failM, failWithM )
70 import qualified IOEnv ( liftIO )
71 import TcEnv ( tcLookupGlobal )
72 import TcRnMonad ( TcM, initTc )
76 import qualified ErrUtils as Err
80 import UniqFM ( UniqFM, mapUFM, filterUFM )
84 import Data.List ( intersperse )
88 import qualified Data.Map as Map
92 import Prelude hiding ( read )
95 import {-# SOURCE #-} TcSplice ( lookupThName_maybe )
96 import qualified Language.Haskell.TH as TH
100 %************************************************************************
104 %************************************************************************
106 These functions are not CoreM monad stuff, but they probably ought to
107 be, and it makes a conveneint place. place for them. They print out
108 stuff before and after core passes, and do Core Lint when necessary.
111 showPass :: DynFlags -> CoreToDo -> IO ()
112 showPass dflags pass = Err.showPass dflags (showSDoc (ppr pass))
114 endPass :: DynFlags -> CoreToDo -> [CoreBind] -> [CoreRule] -> IO ()
115 endPass dflags pass = dumpAndLint dflags True pass empty (coreDumpFlag pass)
117 -- Same as endPass but doesn't dump Core even with -dverbose-core2core
118 endIteration :: DynFlags -> CoreToDo -> Int -> [CoreBind] -> [CoreRule] -> IO ()
119 endIteration dflags pass n
120 = dumpAndLint dflags False pass (ptext (sLit "iteration=") <> int n)
121 (Just Opt_D_dump_simpl_iterations)
123 dumpIfSet :: Bool -> CoreToDo -> SDoc -> SDoc -> IO ()
124 dumpIfSet dump_me pass extra_info doc
125 = Err.dumpIfSet dump_me (showSDoc (ppr pass <+> extra_info)) doc
127 dumpAndLint :: DynFlags -> Bool -> CoreToDo -> SDoc -> Maybe DynFlag
128 -> [CoreBind] -> [CoreRule] -> IO ()
129 -- The "show_all" parameter says to print dump if -dverbose-core2core is on
130 dumpAndLint dflags show_all pass extra_info mb_dump_flag binds rules
131 = do { -- Report result size if required
132 -- This has the side effect of forcing the intermediate to be evaluated
133 ; Err.debugTraceMsg dflags 2 $
134 (text " Result size =" <+> int (coreBindsSize binds))
136 -- Report verbosely, if required
137 ; let pass_name = showSDoc (ppr pass <+> extra_info)
138 dump_doc = pprCoreBindings binds
139 $$ ppUnless (null rules) pp_rules
141 ; case mb_dump_flag of
143 Just dump_flag -> Err.dumpIfSet_dyn_or dflags dump_flags pass_name dump_doc
145 dump_flags | show_all = [dump_flag, Opt_D_verbose_core2core]
146 | otherwise = [dump_flag]
149 ; when (dopt Opt_DoCoreLinting dflags) $
150 do { let (warns, errs) = lintCoreBindings binds
151 ; Err.showPass dflags ("Core Linted result of " ++ pass_name)
152 ; displayLintResults dflags pass warns errs binds } }
154 pp_rules = vcat [ blankLine
155 , ptext (sLit "------ Local rules for imported ids --------")
158 displayLintResults :: DynFlags -> CoreToDo
159 -> Bag Err.Message -> Bag Err.Message -> [CoreBind]
161 displayLintResults dflags pass warns errs binds
162 | not (isEmptyBag errs)
163 = do { printDump (vcat [ banner "errors", Err.pprMessageBag errs
164 , ptext (sLit "*** Offending Program ***")
165 , pprCoreBindings binds
166 , ptext (sLit "*** End of Offense ***") ])
167 ; Err.ghcExit dflags 1 }
169 | not (isEmptyBag warns)
170 , not (case pass of { CoreDesugar -> True; _ -> False })
171 -- Suppress warnings after desugaring pass because some
172 -- are legitimate. Notably, the desugarer generates instance
173 -- methods with INLINE pragmas that form a mutually recursive
174 -- group. Only afer a round of simplification are they unravelled.
175 , not opt_NoDebugOutput
176 , showLintWarnings pass
177 = printDump (banner "warnings" $$ Err.pprMessageBag warns)
179 | otherwise = return ()
181 banner string = ptext (sLit "*** Core Lint") <+> text string
182 <+> ptext (sLit ": in result of") <+> ppr pass
183 <+> ptext (sLit "***")
185 showLintWarnings :: CoreToDo -> Bool
186 -- Disable Lint warnings on the first simplifier pass, because
187 -- there may be some INLINE knots still tied, which is tiresomely noisy
188 showLintWarnings (CoreDoSimplify _ (SimplMode { sm_phase = InitialPhase })) = False
189 showLintWarnings _ = True
193 %************************************************************************
195 The CoreToDo type and related types
196 Abstraction of core-to-core passes to run.
198 %************************************************************************
201 data CoreToDo -- These are diff core-to-core passes,
202 -- which may be invoked in any order,
203 -- as many times as you like.
205 = CoreDoSimplify -- The core-to-core simplifier.
206 Int -- Max iterations
210 | CoreDoFloatOutwards FloatOutSwitches
215 | CoreDoWorkerWrapper
220 | CoreDoRuleCheck CompilerPhase String -- Check for non-application of rules
221 -- matching this string
222 | CoreDoVectorisation
223 | CoreDoNothing -- Useful when building up
224 | CoreDoPasses [CoreToDo] -- lists of these things
226 | CoreDesugar -- Not strictly a core-to-core pass, but produces
227 -- Core output, and hence useful to pass to endPass
232 coreDumpFlag :: CoreToDo -> Maybe DynFlag
233 coreDumpFlag (CoreDoSimplify {}) = Just Opt_D_dump_simpl_phases
234 coreDumpFlag CoreDoFloatInwards = Just Opt_D_verbose_core2core
235 coreDumpFlag (CoreDoFloatOutwards {}) = Just Opt_D_verbose_core2core
236 coreDumpFlag CoreLiberateCase = Just Opt_D_verbose_core2core
237 coreDumpFlag CoreDoStaticArgs = Just Opt_D_verbose_core2core
238 coreDumpFlag CoreDoStrictness = Just Opt_D_dump_stranal
239 coreDumpFlag CoreDoWorkerWrapper = Just Opt_D_dump_worker_wrapper
240 coreDumpFlag CoreDoSpecialising = Just Opt_D_dump_spec
241 coreDumpFlag CoreDoSpecConstr = Just Opt_D_dump_spec
242 coreDumpFlag CoreCSE = Just Opt_D_dump_cse
243 coreDumpFlag CoreDoVectorisation = Just Opt_D_dump_vect
244 coreDumpFlag CoreDesugar = Just Opt_D_dump_ds
245 coreDumpFlag CoreTidy = Just Opt_D_dump_simpl
246 coreDumpFlag CorePrep = Just Opt_D_dump_prep
248 coreDumpFlag CoreDoPrintCore = Nothing
249 coreDumpFlag (CoreDoRuleCheck {}) = Nothing
250 coreDumpFlag CoreDoNothing = Nothing
251 coreDumpFlag CoreDoGlomBinds = Nothing
252 coreDumpFlag (CoreDoPasses {}) = Nothing
254 instance Outputable CoreToDo where
255 ppr (CoreDoSimplify n md) = ptext (sLit "Simplifier")
257 <+> ptext (sLit "max-iterations=") <> int n
258 ppr CoreDoFloatInwards = ptext (sLit "Float inwards")
259 ppr (CoreDoFloatOutwards f) = ptext (sLit "Float out") <> parens (ppr f)
260 ppr CoreLiberateCase = ptext (sLit "Liberate case")
261 ppr CoreDoStaticArgs = ptext (sLit "Static argument")
262 ppr CoreDoStrictness = ptext (sLit "Demand analysis")
263 ppr CoreDoWorkerWrapper = ptext (sLit "Worker Wrapper binds")
264 ppr CoreDoSpecialising = ptext (sLit "Specialise")
265 ppr CoreDoSpecConstr = ptext (sLit "SpecConstr")
266 ppr CoreCSE = ptext (sLit "Common sub-expression")
267 ppr CoreDoVectorisation = ptext (sLit "Vectorisation")
268 ppr CoreDesugar = ptext (sLit "Desugar")
269 ppr CoreTidy = ptext (sLit "Tidy Core")
270 ppr CorePrep = ptext (sLit "CorePrep")
271 ppr CoreDoPrintCore = ptext (sLit "Print core")
272 ppr (CoreDoRuleCheck {}) = ptext (sLit "Rule check")
273 ppr CoreDoGlomBinds = ptext (sLit "Glom binds")
274 ppr CoreDoNothing = ptext (sLit "CoreDoNothing")
275 ppr (CoreDoPasses {}) = ptext (sLit "CoreDoPasses")
279 data SimplifierMode -- See comments in SimplMonad
281 { sm_names :: [String] -- Name(s) of the phase
282 , sm_phase :: CompilerPhase
283 , sm_rules :: Bool -- Whether RULES are enabled
284 , sm_inline :: Bool -- Whether inlining is enabled
285 , sm_case_case :: Bool -- Whether case-of-case is enabled
286 , sm_eta_expand :: Bool -- Whether eta-expansion is enabled
289 instance Outputable SimplifierMode where
290 ppr (SimplMode { sm_phase = p, sm_names = ss
291 , sm_rules = r, sm_inline = i
292 , sm_eta_expand = eta, sm_case_case = cc })
293 = ptext (sLit "SimplMode") <+> braces (
294 sep [ ptext (sLit "Phase =") <+> ppr p <+>
295 brackets (text (concat $ intersperse "," ss)) <> comma
296 , pp_flag i (sLit "inline") <> comma
297 , pp_flag r (sLit "rules") <> comma
298 , pp_flag eta (sLit "eta-expand") <> comma
299 , pp_flag cc (sLit "case-of-case") ])
301 pp_flag f s = ppUnless f (ptext (sLit "no")) <+> ptext s
306 data FloatOutSwitches = FloatOutSwitches {
307 floatOutLambdas :: Maybe Int, -- ^ Just n <=> float lambdas to top level, if
308 -- doing so will abstract over n or fewer
310 -- Nothing <=> float all lambdas to top level,
311 -- regardless of how many free variables
312 -- Just 0 is the vanilla case: float a lambda
313 -- iff it has no free vars
315 floatOutConstants :: Bool, -- ^ True <=> float constants to top level,
316 -- even if they do not escape a lambda
317 floatOutPartialApplications :: Bool -- ^ True <=> float out partial applications
318 -- based on arity information.
320 instance Outputable FloatOutSwitches where
321 ppr = pprFloatOutSwitches
323 pprFloatOutSwitches :: FloatOutSwitches -> SDoc
324 pprFloatOutSwitches sw
325 = ptext (sLit "FOS") <+> (braces $
326 sep $ punctuate comma $
327 [ ptext (sLit "Lam =") <+> ppr (floatOutLambdas sw)
328 , ptext (sLit "Consts =") <+> ppr (floatOutConstants sw)
329 , ptext (sLit "PAPs =") <+> ppr (floatOutPartialApplications sw) ])
333 %************************************************************************
335 Generating the main optimisation pipeline
337 %************************************************************************
340 getCoreToDo :: DynFlags -> [CoreToDo]
344 opt_level = optLevel dflags
345 phases = simplPhases dflags
346 max_iter = maxSimplIterations dflags
347 rule_check = ruleCheck dflags
348 strictness = dopt Opt_Strictness dflags
349 full_laziness = dopt Opt_FullLaziness dflags
350 do_specialise = dopt Opt_Specialise dflags
351 do_float_in = dopt Opt_FloatIn dflags
352 cse = dopt Opt_CSE dflags
353 spec_constr = dopt Opt_SpecConstr dflags
354 liberate_case = dopt Opt_LiberateCase dflags
355 static_args = dopt Opt_StaticArgumentTransformation dflags
356 rules_on = dopt Opt_EnableRewriteRules dflags
357 eta_expand_on = dopt Opt_DoLambdaEtaExpansion dflags
359 maybe_rule_check phase = runMaybe rule_check (CoreDoRuleCheck phase)
361 maybe_strictness_before phase
362 = runWhen (phase `elem` strictnessBefore dflags) CoreDoStrictness
364 base_mode = SimplMode { sm_phase = panic "base_mode"
366 , sm_rules = rules_on
367 , sm_eta_expand = eta_expand_on
369 , sm_case_case = True }
371 simpl_phase phase names iter
373 [ maybe_strictness_before phase
374 , CoreDoSimplify iter
375 (base_mode { sm_phase = Phase phase
376 , sm_names = names })
378 , maybe_rule_check (Phase phase)
382 = runWhen (dopt Opt_Vectorise dflags) $
383 CoreDoPasses [ simpl_gently, CoreDoVectorisation ]
385 -- By default, we have 2 phases before phase 0.
387 -- Want to run with inline phase 2 after the specialiser to give
388 -- maximum chance for fusion to work before we inline build/augment
389 -- in phase 1. This made a difference in 'ansi' where an
390 -- overloaded function wasn't inlined till too late.
392 -- Need phase 1 so that build/augment get
393 -- inlined. I found that spectral/hartel/genfft lost some useful
394 -- strictness in the function sumcode' if augment is not inlined
395 -- before strictness analysis runs
396 simpl_phases = CoreDoPasses [ simpl_phase phase ["main"] max_iter
397 | phase <- [phases, phases-1 .. 1] ]
400 -- initial simplify: mk specialiser happy: minimum effort please
401 simpl_gently = CoreDoSimplify max_iter
402 (base_mode { sm_phase = InitialPhase
403 , sm_names = ["Gentle"]
404 , sm_rules = rules_on -- Note [RULEs enabled in SimplGently]
406 , sm_case_case = False })
407 -- Don't do case-of-case transformations.
408 -- This makes full laziness work better
411 if opt_level == 0 then
413 simpl_phase 0 ["final"] max_iter]
414 else {- opt_level >= 1 -} [
416 -- We want to do the static argument transform before full laziness as it
417 -- may expose extra opportunities to float things outwards. However, to fix
418 -- up the output of the transformation we need at do at least one simplify
419 -- after this before anything else
420 runWhen static_args (CoreDoPasses [ simpl_gently, CoreDoStaticArgs ]),
422 -- We run vectorisation here for now, but we might also try to run
426 -- initial simplify: mk specialiser happy: minimum effort please
429 -- Specialisation is best done before full laziness
430 -- so that overloaded functions have all their dictionary lambdas manifest
431 runWhen do_specialise CoreDoSpecialising,
433 runWhen full_laziness $
434 CoreDoFloatOutwards FloatOutSwitches {
435 floatOutLambdas = Just 0,
436 floatOutConstants = True,
437 floatOutPartialApplications = False },
438 -- Was: gentleFloatOutSwitches
440 -- I have no idea why, but not floating constants to
441 -- top level is very bad in some cases.
443 -- Notably: p_ident in spectral/rewrite
444 -- Changing from "gentle" to "constantsOnly"
445 -- improved rewrite's allocation by 19%, and
446 -- made 0.0% difference to any other nofib
449 -- Not doing floatOutPartialApplications yet, we'll do
450 -- that later on when we've had a chance to get more
451 -- accurate arity information. In fact it makes no
452 -- difference at all to performance if we do it here,
453 -- but maybe we save some unnecessary to-and-fro in
456 runWhen do_float_in CoreDoFloatInwards,
460 -- Phase 0: allow all Ids to be inlined now
461 -- This gets foldr inlined before strictness analysis
463 -- At least 3 iterations because otherwise we land up with
464 -- huge dead expressions because of an infelicity in the
466 -- let k = BIG in foldr k z xs
467 -- ==> let k = BIG in letrec go = \xs -> ...(k x).... in go xs
468 -- ==> let k = BIG in letrec go = \xs -> ...(BIG x).... in go xs
470 simpl_phase 0 ["main"] (max max_iter 3),
472 runWhen strictness (CoreDoPasses [
476 simpl_phase 0 ["post-worker-wrapper"] max_iter
479 runWhen full_laziness $
480 CoreDoFloatOutwards FloatOutSwitches {
481 floatOutLambdas = floatLamArgs dflags,
482 floatOutConstants = True,
483 floatOutPartialApplications = True },
484 -- nofib/spectral/hartel/wang doubles in speed if you
485 -- do full laziness late in the day. It only happens
486 -- after fusion and other stuff, so the early pass doesn't
487 -- catch it. For the record, the redex is
488 -- f_el22 (f_el21 r_midblock)
492 -- We want CSE to follow the final full-laziness pass, because it may
493 -- succeed in commoning up things floated out by full laziness.
494 -- CSE used to rely on the no-shadowing invariant, but it doesn't any more
496 runWhen do_float_in CoreDoFloatInwards,
498 maybe_rule_check (Phase 0),
500 -- Case-liberation for -O2. This should be after
501 -- strictness analysis and the simplification which follows it.
502 runWhen liberate_case (CoreDoPasses [
504 simpl_phase 0 ["post-liberate-case"] max_iter
505 ]), -- Run the simplifier after LiberateCase to vastly
506 -- reduce the possiblility of shadowing
507 -- Reason: see Note [Shadowing] in SpecConstr.lhs
509 runWhen spec_constr CoreDoSpecConstr,
511 maybe_rule_check (Phase 0),
513 -- Final clean-up simplification:
514 simpl_phase 0 ["final"] max_iter
517 -- The core-to-core pass ordering is derived from the DynFlags:
518 runWhen :: Bool -> CoreToDo -> CoreToDo
519 runWhen True do_this = do_this
520 runWhen False _ = CoreDoNothing
522 runMaybe :: Maybe a -> (a -> CoreToDo) -> CoreToDo
523 runMaybe (Just x) f = f x
524 runMaybe Nothing _ = CoreDoNothing
526 dumpSimplPhase :: DynFlags -> SimplifierMode -> Bool
527 dumpSimplPhase dflags mode
528 | Just spec_string <- shouldDumpSimplPhase dflags
529 = match_spec spec_string
531 = dopt Opt_D_verbose_core2core dflags
534 match_spec :: String -> Bool
535 match_spec spec_string
536 = or $ map (and . map match . split ':')
537 $ split ',' spec_string
539 match :: String -> Bool
541 match s = case reads s of
542 [(n,"")] -> phase_num n
545 phase_num :: Int -> Bool
546 phase_num n = case sm_phase mode of
550 phase_name :: String -> Bool
551 phase_name s = s `elem` sm_names mode
555 Note [RULEs enabled in SimplGently]
556 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
557 RULES are enabled when doing "gentle" simplification. Two reasons:
559 * We really want the class-op cancellation to happen:
560 op (df d1 d2) --> $cop3 d1 d2
561 because this breaks the mutual recursion between 'op' and 'df'
565 to work in Template Haskell when simplifying
566 splices, so we get simpler code for literal strings
568 But watch out: list fusion can prevent floating. So use phase control
569 to switch off those rules until after floating.
572 %************************************************************************
576 %************************************************************************
579 verboseSimplStats :: Bool
580 verboseSimplStats = opt_PprStyle_Debug -- For now, anyway
582 zeroSimplCount :: DynFlags -> SimplCount
583 isZeroSimplCount :: SimplCount -> Bool
584 pprSimplCount :: SimplCount -> SDoc
585 doSimplTick, doFreeSimplTick :: Tick -> SimplCount -> SimplCount
586 plusSimplCount :: SimplCount -> SimplCount -> SimplCount
591 = VerySimplCount !Int -- Used when don't want detailed stats
594 ticks :: !Int, -- Total ticks
595 details :: !TickCounts, -- How many of each type
598 log1 :: [Tick], -- Last N events; <= opt_HistorySize,
600 log2 :: [Tick] -- Last opt_HistorySize events before that
601 -- Having log1, log2 lets us accumulate the
602 -- recent history reasonably efficiently
605 type TickCounts = Map Tick Int
607 simplCountN :: SimplCount -> Int
608 simplCountN (VerySimplCount n) = n
609 simplCountN (SimplCount { ticks = n }) = n
611 zeroSimplCount dflags
612 -- This is where we decide whether to do
613 -- the VerySimpl version or the full-stats version
614 | dopt Opt_D_dump_simpl_stats dflags
615 = SimplCount {ticks = 0, details = Map.empty,
616 n_log = 0, log1 = [], log2 = []}
620 isZeroSimplCount (VerySimplCount n) = n==0
621 isZeroSimplCount (SimplCount { ticks = n }) = n==0
623 doFreeSimplTick tick sc@SimplCount { details = dts }
624 = sc { details = dts `addTick` tick }
625 doFreeSimplTick _ sc = sc
627 doSimplTick tick sc@SimplCount { ticks = tks, details = dts, n_log = nl, log1 = l1 }
628 | nl >= opt_HistorySize = sc1 { n_log = 1, log1 = [tick], log2 = l1 }
629 | otherwise = sc1 { n_log = nl+1, log1 = tick : l1 }
631 sc1 = sc { ticks = tks+1, details = dts `addTick` tick }
633 doSimplTick _ (VerySimplCount n) = VerySimplCount (n+1)
636 -- Don't use Map.unionWith because that's lazy, and we want to
637 -- be pretty strict here!
638 addTick :: TickCounts -> Tick -> TickCounts
639 addTick fm tick = case Map.lookup tick fm of
640 Nothing -> Map.insert tick 1 fm
641 Just n -> n1 `seq` Map.insert tick n1 fm
646 plusSimplCount sc1@(SimplCount { ticks = tks1, details = dts1 })
647 sc2@(SimplCount { ticks = tks2, details = dts2 })
648 = log_base { ticks = tks1 + tks2, details = Map.unionWith (+) dts1 dts2 }
650 -- A hackish way of getting recent log info
651 log_base | null (log1 sc2) = sc1 -- Nothing at all in sc2
652 | null (log2 sc2) = sc2 { log2 = log1 sc1 }
655 plusSimplCount (VerySimplCount n) (VerySimplCount m) = VerySimplCount (n+m)
656 plusSimplCount _ _ = panic "plusSimplCount"
657 -- We use one or the other consistently
659 pprSimplCount (VerySimplCount n) = ptext (sLit "Total ticks:") <+> int n
660 pprSimplCount (SimplCount { ticks = tks, details = dts, log1 = l1, log2 = l2 })
661 = vcat [ptext (sLit "Total ticks: ") <+> int tks,
663 pprTickCounts (Map.toList dts),
664 if verboseSimplStats then
666 ptext (sLit "Log (most recent first)"),
667 nest 4 (vcat (map ppr l1) $$ vcat (map ppr l2))]
671 pprTickCounts :: [(Tick,Int)] -> SDoc
672 pprTickCounts [] = empty
673 pprTickCounts ((tick1,n1):ticks)
674 = vcat [int tot_n <+> text (tickString tick1),
675 pprTCDetails real_these,
679 tick1_tag = tickToTag tick1
680 (these, others) = span same_tick ticks
681 real_these = (tick1,n1):these
682 same_tick (tick2,_) = tickToTag tick2 == tick1_tag
683 tot_n = sum [n | (_,n) <- real_these]
685 pprTCDetails :: [(Tick, Int)] -> SDoc
687 = nest 4 (vcat [int n <+> pprTickCts tick | (tick,n) <- ticks])
693 = PreInlineUnconditionally Id
694 | PostInlineUnconditionally Id
697 | RuleFired FastString -- Rule name
700 | EtaExpansion Id -- LHS binder
701 | EtaReduction Id -- Binder on outer lambda
702 | BetaReduction Id -- Lambda binder
705 | CaseOfCase Id -- Bndr on *inner* case
706 | KnownBranch Id -- Case binder
707 | CaseMerge Id -- Binder on outer case
708 | AltMerge Id -- Case binder
709 | CaseElim Id -- Case binder
710 | CaseIdentity Id -- Case binder
711 | FillInCaseDefault Id -- Case binder
714 | SimplifierDone -- Ticked at each iteration of the simplifier
716 instance Outputable Tick where
717 ppr tick = text (tickString tick) <+> pprTickCts tick
719 instance Eq Tick where
720 a == b = case a `cmpTick` b of
724 instance Ord Tick where
727 tickToTag :: Tick -> Int
728 tickToTag (PreInlineUnconditionally _) = 0
729 tickToTag (PostInlineUnconditionally _) = 1
730 tickToTag (UnfoldingDone _) = 2
731 tickToTag (RuleFired _) = 3
732 tickToTag LetFloatFromLet = 4
733 tickToTag (EtaExpansion _) = 5
734 tickToTag (EtaReduction _) = 6
735 tickToTag (BetaReduction _) = 7
736 tickToTag (CaseOfCase _) = 8
737 tickToTag (KnownBranch _) = 9
738 tickToTag (CaseMerge _) = 10
739 tickToTag (CaseElim _) = 11
740 tickToTag (CaseIdentity _) = 12
741 tickToTag (FillInCaseDefault _) = 13
742 tickToTag BottomFound = 14
743 tickToTag SimplifierDone = 16
744 tickToTag (AltMerge _) = 17
746 tickString :: Tick -> String
747 tickString (PreInlineUnconditionally _) = "PreInlineUnconditionally"
748 tickString (PostInlineUnconditionally _)= "PostInlineUnconditionally"
749 tickString (UnfoldingDone _) = "UnfoldingDone"
750 tickString (RuleFired _) = "RuleFired"
751 tickString LetFloatFromLet = "LetFloatFromLet"
752 tickString (EtaExpansion _) = "EtaExpansion"
753 tickString (EtaReduction _) = "EtaReduction"
754 tickString (BetaReduction _) = "BetaReduction"
755 tickString (CaseOfCase _) = "CaseOfCase"
756 tickString (KnownBranch _) = "KnownBranch"
757 tickString (CaseMerge _) = "CaseMerge"
758 tickString (AltMerge _) = "AltMerge"
759 tickString (CaseElim _) = "CaseElim"
760 tickString (CaseIdentity _) = "CaseIdentity"
761 tickString (FillInCaseDefault _) = "FillInCaseDefault"
762 tickString BottomFound = "BottomFound"
763 tickString SimplifierDone = "SimplifierDone"
765 pprTickCts :: Tick -> SDoc
766 pprTickCts (PreInlineUnconditionally v) = ppr v
767 pprTickCts (PostInlineUnconditionally v)= ppr v
768 pprTickCts (UnfoldingDone v) = ppr v
769 pprTickCts (RuleFired v) = ppr v
770 pprTickCts LetFloatFromLet = empty
771 pprTickCts (EtaExpansion v) = ppr v
772 pprTickCts (EtaReduction v) = ppr v
773 pprTickCts (BetaReduction v) = ppr v
774 pprTickCts (CaseOfCase v) = ppr v
775 pprTickCts (KnownBranch v) = ppr v
776 pprTickCts (CaseMerge v) = ppr v
777 pprTickCts (AltMerge v) = ppr v
778 pprTickCts (CaseElim v) = ppr v
779 pprTickCts (CaseIdentity v) = ppr v
780 pprTickCts (FillInCaseDefault v) = ppr v
783 cmpTick :: Tick -> Tick -> Ordering
784 cmpTick a b = case (tickToTag a `compare` tickToTag b) of
789 cmpEqTick :: Tick -> Tick -> Ordering
790 cmpEqTick (PreInlineUnconditionally a) (PreInlineUnconditionally b) = a `compare` b
791 cmpEqTick (PostInlineUnconditionally a) (PostInlineUnconditionally b) = a `compare` b
792 cmpEqTick (UnfoldingDone a) (UnfoldingDone b) = a `compare` b
793 cmpEqTick (RuleFired a) (RuleFired b) = a `compare` b
794 cmpEqTick (EtaExpansion a) (EtaExpansion b) = a `compare` b
795 cmpEqTick (EtaReduction a) (EtaReduction b) = a `compare` b
796 cmpEqTick (BetaReduction a) (BetaReduction b) = a `compare` b
797 cmpEqTick (CaseOfCase a) (CaseOfCase b) = a `compare` b
798 cmpEqTick (KnownBranch a) (KnownBranch b) = a `compare` b
799 cmpEqTick (CaseMerge a) (CaseMerge b) = a `compare` b
800 cmpEqTick (AltMerge a) (AltMerge b) = a `compare` b
801 cmpEqTick (CaseElim a) (CaseElim b) = a `compare` b
802 cmpEqTick (CaseIdentity a) (CaseIdentity b) = a `compare` b
803 cmpEqTick (FillInCaseDefault a) (FillInCaseDefault b) = a `compare` b
808 %************************************************************************
810 Monad and carried data structure definitions
812 %************************************************************************
815 newtype CoreState = CoreState {
816 cs_uniq_supply :: UniqSupply
819 data CoreReader = CoreReader {
820 cr_hsc_env :: HscEnv,
821 cr_rule_base :: RuleBase,
825 data CoreWriter = CoreWriter {
826 cw_simpl_count :: SimplCount
829 emptyWriter :: DynFlags -> CoreWriter
830 emptyWriter dflags = CoreWriter {
831 cw_simpl_count = zeroSimplCount dflags
834 plusWriter :: CoreWriter -> CoreWriter -> CoreWriter
835 plusWriter w1 w2 = CoreWriter {
836 cw_simpl_count = (cw_simpl_count w1) `plusSimplCount` (cw_simpl_count w2)
839 type CoreIOEnv = IOEnv CoreReader
841 -- | The monad used by Core-to-Core passes to access common state, register simplification
842 -- statistics and so on
843 newtype CoreM a = CoreM { unCoreM :: CoreState -> CoreIOEnv (a, CoreState, CoreWriter) }
845 instance Functor CoreM where
850 instance Monad CoreM where
851 return x = CoreM (\s -> nop s x)
852 mx >>= f = CoreM $ \s -> do
853 (x, s', w1) <- unCoreM mx s
854 (y, s'', w2) <- unCoreM (f x) s'
855 return (y, s'', w1 `plusWriter` w2)
857 instance Applicative CoreM where
861 -- For use if the user has imported Control.Monad.Error from MTL
862 -- Requires UndecidableInstances
863 instance MonadPlus IO => MonadPlus CoreM where
864 mzero = CoreM (const mzero)
865 m `mplus` n = CoreM (\rs -> unCoreM m rs `mplus` unCoreM n rs)
867 instance MonadUnique CoreM where
868 getUniqueSupplyM = do
869 us <- getS cs_uniq_supply
870 let (us1, us2) = splitUniqSupply us
871 modifyS (\s -> s { cs_uniq_supply = us2 })
879 -> IO (a, SimplCount)
880 runCoreM hsc_env rule_base us mod m =
881 liftM extract $ runIOEnv reader $ unCoreM m state
883 reader = CoreReader {
884 cr_hsc_env = hsc_env,
885 cr_rule_base = rule_base,
892 extract :: (a, CoreState, CoreWriter) -> (a, SimplCount)
893 extract (value, _, writer) = (value, cw_simpl_count writer)
898 %************************************************************************
900 Core combinators, not exported
902 %************************************************************************
906 nop :: CoreState -> a -> CoreIOEnv (a, CoreState, CoreWriter)
909 return (x, s, emptyWriter $ (hsc_dflags . cr_hsc_env) r)
911 read :: (CoreReader -> a) -> CoreM a
912 read f = CoreM (\s -> getEnv >>= (\r -> nop s (f r)))
914 getS :: (CoreState -> a) -> CoreM a
915 getS f = CoreM (\s -> nop s (f s))
917 modifyS :: (CoreState -> CoreState) -> CoreM ()
918 modifyS f = CoreM (\s -> nop (f s) ())
920 write :: CoreWriter -> CoreM ()
921 write w = CoreM (\s -> return ((), s, w))
925 \subsection{Lifting IO into the monad}
929 -- | Lift an 'IOEnv' operation into 'CoreM'
930 liftIOEnv :: CoreIOEnv a -> CoreM a
931 liftIOEnv mx = CoreM (\s -> mx >>= (\x -> nop s x))
933 instance MonadIO CoreM where
934 liftIO = liftIOEnv . IOEnv.liftIO
936 -- | Lift an 'IO' operation into 'CoreM' while consuming its 'SimplCount'
937 liftIOWithCount :: IO (SimplCount, a) -> CoreM a
938 liftIOWithCount what = liftIO what >>= (\(count, x) -> addSimplCount count >> return x)
943 %************************************************************************
945 Reader, writer and state accessors
947 %************************************************************************
951 getHscEnv :: CoreM HscEnv
952 getHscEnv = read cr_hsc_env
954 getRuleBase :: CoreM RuleBase
955 getRuleBase = read cr_rule_base
957 getModule :: CoreM Module
958 getModule = read cr_module
960 addSimplCount :: SimplCount -> CoreM ()
961 addSimplCount count = write (CoreWriter { cw_simpl_count = count })
963 -- Convenience accessors for useful fields of HscEnv
965 getDynFlags :: CoreM DynFlags
966 getDynFlags = fmap hsc_dflags getHscEnv
968 -- | The original name cache is the current mapping from 'Module' and
969 -- 'OccName' to a compiler-wide unique 'Name'
970 getOrigNameCache :: CoreM OrigNameCache
971 getOrigNameCache = do
972 nameCacheRef <- fmap hsc_NC getHscEnv
973 liftIO $ fmap nsNames $ readIORef nameCacheRef
978 %************************************************************************
980 Dealing with annotations
982 %************************************************************************
985 -- | Get all annotations of a given type. This happens lazily, that is
986 -- no deserialization will take place until the [a] is actually demanded and
987 -- the [a] can also be empty (the UniqFM is not filtered).
989 -- This should be done once at the start of a Core-to-Core pass that uses
992 -- See Note [Annotations]
993 getAnnotations :: Typeable a => ([Word8] -> a) -> ModGuts -> CoreM (UniqFM [a])
994 getAnnotations deserialize guts = do
996 ann_env <- liftIO $ prepareAnnotations hsc_env (Just guts)
997 return (deserializeAnns deserialize ann_env)
999 -- | Get at most one annotation of a given type per Unique.
1000 getFirstAnnotations :: Typeable a => ([Word8] -> a) -> ModGuts -> CoreM (UniqFM a)
1001 getFirstAnnotations deserialize guts
1002 = liftM (mapUFM head . filterUFM (not . null))
1003 $ getAnnotations deserialize guts
1009 A Core-to-Core pass that wants to make use of annotations calls
1010 getAnnotations or getFirstAnnotations at the beginning to obtain a UniqFM with
1011 annotations of a specific type. This produces all annotations from interface
1012 files read so far. However, annotations from interface files read during the
1013 pass will not be visible until getAnnotations is called again. This is similar
1014 to how rules work and probably isn't too bad.
1016 The current implementation could be optimised a bit: when looking up
1017 annotations for a thing from the HomePackageTable, we could search directly in
1018 the module where the thing is defined rather than building one UniqFM which
1019 contains all annotations we know of. This would work because annotations can
1020 only be given to things defined in the same module. However, since we would
1021 only want to deserialise every annotation once, we would have to build a cache
1022 for every module in the HTP. In the end, it's probably not worth it as long as
1023 we aren't using annotations heavily.
1025 %************************************************************************
1027 Direct screen output
1029 %************************************************************************
1033 msg :: (DynFlags -> SDoc -> IO ()) -> SDoc -> CoreM ()
1035 dflags <- getDynFlags
1036 liftIO $ how dflags doc
1038 -- | Output a String message to the screen
1039 putMsgS :: String -> CoreM ()
1040 putMsgS = putMsg . text
1042 -- | Output a message to the screen
1043 putMsg :: SDoc -> CoreM ()
1044 putMsg = msg Err.putMsg
1046 -- | Output a string error to the screen
1047 errorMsgS :: String -> CoreM ()
1048 errorMsgS = errorMsg . text
1050 -- | Output an error to the screen
1051 errorMsg :: SDoc -> CoreM ()
1052 errorMsg = msg Err.errorMsg
1054 -- | Output a fatal string error to the screen. Note this does not by itself cause the compiler to die
1055 fatalErrorMsgS :: String -> CoreM ()
1056 fatalErrorMsgS = fatalErrorMsg . text
1058 -- | Output a fatal error to the screen. Note this does not by itself cause the compiler to die
1059 fatalErrorMsg :: SDoc -> CoreM ()
1060 fatalErrorMsg = msg Err.fatalErrorMsg
1062 -- | Output a string debugging message at verbosity level of @-v@ or higher
1063 debugTraceMsgS :: String -> CoreM ()
1064 debugTraceMsgS = debugTraceMsg . text
1066 -- | Outputs a debugging message at verbosity level of @-v@ or higher
1067 debugTraceMsg :: SDoc -> CoreM ()
1068 debugTraceMsg = msg (flip Err.debugTraceMsg 3)
1070 -- | Show some labelled 'SDoc' if a particular flag is set or at a verbosity level of @-v -ddump-most@ or higher
1071 dumpIfSet_dyn :: DynFlag -> String -> SDoc -> CoreM ()
1072 dumpIfSet_dyn flag str = msg (\dflags -> Err.dumpIfSet_dyn dflags flag str)
1077 initTcForLookup :: HscEnv -> TcM a -> IO a
1078 initTcForLookup hsc_env = liftM (expectJust "initTcInteractive" . snd) . initTc hsc_env HsSrcFile False iNTERACTIVE
1083 %************************************************************************
1087 %************************************************************************
1090 instance MonadThings CoreM where
1091 lookupThing name = do
1092 hsc_env <- getHscEnv
1093 liftIO $ initTcForLookup hsc_env (tcLookupGlobal name)
1096 %************************************************************************
1098 Template Haskell interoperability
1100 %************************************************************************
1104 -- | Attempt to convert a Template Haskell name to one that GHC can
1105 -- understand. Original TH names such as those you get when you use
1106 -- the @'foo@ syntax will be translated to their equivalent GHC name
1107 -- exactly. Qualified or unqualifed TH names will be dynamically bound
1108 -- to names in the module being compiled, if possible. Exact TH names
1109 -- will be bound to the name they represent, exactly.
1110 thNameToGhcName :: TH.Name -> CoreM (Maybe Name)
1111 thNameToGhcName th_name = do
1112 hsc_env <- getHscEnv
1113 liftIO $ initTcForLookup hsc_env (lookupThName_maybe th_name)