{-# LANGUAGE TupleSections #-} -- This module contains functions which implement -- the -finfo-table-map and -fdistinct-constructor-tables flags module GHC.Stg.Debug(collectDebugInformation) where import GHC.Prelude import GHC.Stg.Syntax import GHC.Types.Id import GHC.Types.Tickish import GHC.Core.DataCon import GHC.Types.IPE import GHC.Unit.Module import GHC.Types.Name ( getName, getOccName, occNameString, nameSrcSpan) import GHC.Data.FastString import GHC.Driver.Session import Control.Monad (when) import Control.Monad.Trans.Reader import Control.Monad.Trans.State import Control.Monad.Trans.Class import GHC.Types.Unique.Map import GHC.Types.SrcLoc import Control.Applicative import qualified Data.List.NonEmpty as NE import Data.List.NonEmpty (NonEmpty(..)) data SpanWithLabel = SpanWithLabel RealSrcSpan String data R = R { rDynFlags :: DynFlags, rModLocation :: ModLocation, rSpan :: Maybe SpanWithLabel } type M a = ReaderT R (State InfoTableProvMap) a withSpan :: (RealSrcSpan, String) -> M a -> M a withSpan (new_s, new_l) act = local maybe_replace act where maybe_replace r@R{ rModLocation = cur_mod, rSpan = Just (SpanWithLabel old_s _old_l) } -- prefer spans from the current module | Just (unpackFS $ srcSpanFile old_s) == ml_hs_file cur_mod , Just (unpackFS $ srcSpanFile new_s) /= ml_hs_file cur_mod = r maybe_replace r = r { rSpan = Just (SpanWithLabel new_s new_l) } collectDebugInformation :: DynFlags -> ModLocation -> [StgTopBinding] -> ([StgTopBinding], InfoTableProvMap) collectDebugInformation dflags ml bs = runState (runReaderT (mapM collectTop bs) (R dflags ml Nothing)) emptyInfoTableProvMap collectTop :: StgTopBinding -> M StgTopBinding collectTop (StgTopLifted t) = StgTopLifted <$> collectStgBind t collectTop tb = return tb collectStgBind :: StgBinding -> M StgBinding collectStgBind (StgNonRec bndr rhs) = do rhs' <- collectStgRhs bndr rhs return (StgNonRec bndr rhs') collectStgBind (StgRec pairs) = do es <- mapM (\(b, e) -> (b,) <$> collectStgRhs b e) pairs return (StgRec es) collectStgRhs :: Id -> StgRhs -> M StgRhs collectStgRhs bndr (StgRhsClosure ext cc us bs e)= do e' <- collectExpr e recordInfo bndr e' return $ StgRhsClosure ext cc us bs e' collectStgRhs _bndr (StgRhsCon cc dc _mn ticks args) = do n' <- numberDataCon dc ticks return (StgRhsCon cc dc n' ticks args) recordInfo :: Id -> StgExpr -> M () recordInfo bndr new_rhs = do modLoc <- asks rModLocation let thisFile = maybe nilFS mkFastString $ ml_hs_file modLoc -- A span from the ticks surrounding the new_rhs best_span = quickSourcePos thisFile new_rhs -- A back-up span if the bndr had a source position, many do not (think internally generated ids) bndr_span = (\s -> SpanWithLabel s (occNameString (getOccName bndr))) <$> srcSpanToRealSrcSpan (nameSrcSpan (getName bndr)) recordStgIdPosition bndr best_span bndr_span collectExpr :: StgExpr -> M StgExpr collectExpr = go where go (StgApp occ as) = return $ StgApp occ as go (StgLit lit) = return $ StgLit lit go (StgConApp dc _mn as tys) = do n' <- numberDataCon dc [] return (StgConApp dc n' as tys) go (StgOpApp op as ty) = return (StgOpApp op as ty) go (StgCase scrut bndr ty alts) = StgCase <$> collectExpr scrut <*> pure bndr <*> pure ty <*> mapM collectAlt alts go (StgLet ext bind body) = do bind' <- collectStgBind bind body' <- go body return (StgLet ext bind' body') go (StgLetNoEscape ext bind body) = do bind' <- collectStgBind bind body' <- go body return (StgLetNoEscape ext bind' body') go (StgTick tick e) = do let k = case tick of SourceNote ss fp -> withSpan (ss, fp) _ -> id e' <- k (go e) return (StgTick tick e') collectAlt :: StgAlt -> M StgAlt collectAlt (ac, bs, e) = (ac, bs, ) <$> collectExpr e -- | Try to find the best source position surrounding a 'StgExpr'. The -- heuristic strips ticks from the current expression until it finds one which -- is from the module currently being compiled. This is the same method that -- the DWARF information uses to give locations to info tables. -- -- It is usually a better alternative than using the 'RealSrcSpan' which is carefully -- propagated downwards by 'withSpan'. It's "quick" because it works only using immediate context rather -- than looking at the parent context like 'withSpan' quickSourcePos :: FastString -> StgExpr -> Maybe SpanWithLabel quickSourcePos cur_mod (StgTick (SourceNote ss m) e) | srcSpanFile ss == cur_mod = Just (SpanWithLabel ss m) | otherwise = quickSourcePos cur_mod e quickSourcePos _ _ = Nothing recordStgIdPosition :: Id -> Maybe SpanWithLabel -> Maybe SpanWithLabel -> M () recordStgIdPosition id best_span ss = do dflags <- asks rDynFlags when (gopt Opt_InfoTableMap dflags) $ do cc <- asks rSpan --Useful for debugging why a certain Id gets given a certain span --pprTraceM "recordStgIdPosition" (ppr id $$ ppr cc $$ ppr best_span $$ ppr ss) let mbspan = (\(SpanWithLabel rss d) -> (rss, d)) <$> (best_span <|> cc <|> ss) lift $ modify (\env -> env { provClosure = addToUniqMap (provClosure env) (idName id) (idType id, mbspan) }) numberDataCon :: DataCon -> [StgTickish] -> M ConstructorNumber -- Unboxed tuples and sums do not allocate so they -- have no info tables. numberDataCon dc _ | isUnboxedTupleDataCon dc = return NoNumber numberDataCon dc _ | isUnboxedSumDataCon dc = return NoNumber numberDataCon dc ts = do dflags <- asks rDynFlags if not (gopt Opt_DistinctConstructorTables dflags) then return NoNumber else do env <- lift get mcc <- asks rSpan let mbest_span = (\(SpanWithLabel rss l) -> (rss, l)) <$> (selectTick ts <|> mcc) let dcMap' = alterUniqMap (maybe (Just ((0, mbest_span) :| [] )) (\xs@((k, _):|_) -> Just ((k + 1, mbest_span) `NE.cons` xs))) (provDC env) dc lift $ put (env { provDC = dcMap' }) let r = lookupUniqMap dcMap' dc return $ case r of Nothing -> NoNumber Just res -> Numbered (fst (NE.head res)) selectTick :: [StgTickish] -> Maybe SpanWithLabel selectTick [] = Nothing selectTick (SourceNote rss d : ts ) = selectTick ts <|> Just (SpanWithLabel rss d) selectTick (_:ts) = selectTick ts {- Note [Mapping Info Tables to Source Positions] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ This note describes what the `-finfo-table-map` flag achieves. When debugging memory issues it is very useful to be able to map a specific closure to a position in the source. The prime example is being able to map a THUNK to a specific place in the source program, the mapping is usually quite precise because a fresh info table is created for each distinct THUNK. There are three parts to the implementation 1. In GHC.Stg.Debug, the SourceNote information is used in order to give a source location to some specific closures. 2. In StgToCmm, the actually used info tables are recorded in an IORef, this is important as it's hard to predict beforehand what code generation will do and which ids will end up in the generated program. 3. During code generation, a mapping from the info table to the statically determined location is emitted which can then be queried at runtime by various tools. -- Giving Source Locations to Closures At the moment thunk and constructor closures are added to the map. This information is collected in the `InfoTableProvMap` which provides a mapping from: 1. Data constructors to a list of where they are used. 2. `Name`s and where they originate from. During the CoreToStg phase, this map is populated whenever something is turned into a StgRhsClosure or an StgConApp. The current source position is recorded depending on the location indicated by the surrounding SourceNote. The functions which add information to the map are `recordStgIdPosition` and `numberDataCon`. When the -fdistinct-constructor-tables` flag is turned on then every usage of a data constructor gets its own distinct info table. This is orchestrated in `collectExpr` where an incrementing number is used to distinguish each occurrence of a data constructor. -- StgToCmm The info tables which are actually used in the generated program are recorded during the conversion from STG to Cmm. The used info tables are recorded in the `emitProc` function. All the used info tables are recorded in the `cgs_used_info` field. This step is necessary because when the information about names is collected in the previous phase it's unpredictable about which names will end up needing info tables. If you don't record which ones are actually used then you end up generating code which references info tables which don't exist. -- Code Generation The output of these two phases is combined together during code generation. A C stub is generated which creates the static map from info table pointer to the information about where that info table was created from. This is created by `ipInitCode` in the same manner as a C stub is generated for cost centres. This information can be consumed in two ways. 1. The complete mapping is emitted into the eventlog so that external tools such as eventlog2html can use the information with the heap profile by info table mode. 2. The `lookupIPE` function can be used via the `whereFrom#` primop to introspect information about a closure in a running Haskell program. Note [Distinct Info Tables for Constructors] ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ In the old times, each usage of a data constructor used the same info table. This made it impossible to distinguish which actual usuage of a data constructor was contributing primarily to the allocation in a program. Using the `-fdistinct-info-tables` flag you can cause code generation to generate a distinct info table for each usage of a constructor. Then, when inspecting the heap you can see precisely which usage of a constructor was responsible for each allocation. -}