[GHC] #14789: GHCi fails to garbage collect declaration `l = length [1..10^8]` entered at prompt
#14789: GHCi fails to garbage collect declaration `l = length [1..10^8]` entered at prompt -------------------------------------+------------------------------------- Reporter: kabuhr | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: Component: GHCi | Version: 8.2.2 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: Runtime Unknown/Multiple | performance bug Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- This was recently reported on Stack Overflow (https://stackoverflow.com/questions/48723119/binding-len-length-xs-and- then-calculating-len-causes-ghc-to-consume-lots). With GHC 8.2.2, assigning the length of a large list as a top-level declaration at the GHCi prompt and evaluating it: {{{ Prelude> l = length [1..10^8] Prelude> l 10000000 Prelude> }}} allocates about 7 gigs of memory before displaying a result, and then memory continues to grow over the next few seconds (at the idle GHCi prompt) until 10 gigs of memory has been allocated. None of it is freed until the interpreter session ends. In contrast, directly evaluating the expression or assigning it with a let-statement does **not** exhibit this behavior: {{{ Prelude> length [1..10^8] 10000000 Prelude> let l = length [1..10^8] Prelude> l 10000000 Prelude> }}} Also, loading the top-level declaration from a file and then evaluating it at the prompt does **not** exhibit the problem. I took a look at Ticket [ticket: 14336] and applied the `-fno-it` patch referenced there. This slightly changed the pattern of memory usage: it grew to about 4 gigs before displaying the result, but then continued to grow for several seconds at the GHCi prompt before landing at about 10 gigs. I verified that the bug still exists on the current "master" branch (though I didn't check the "ghc-8.4" branch). -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14789 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14789: GHCi fails to garbage collect declaration `l = length [1..10^8]` entered at prompt -------------------------------------+------------------------------------- Reporter: kabuhr | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: Component: GHCi | Version: 8.2.2 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by bgamari): Just noticed this; quite an interesting issue. I was indeed able to reproduce this in GHC 8.4.1 and confirmed that `performMajorGC` does not cause reclaim. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14789#comment:1 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
#14789: GHCi fails to garbage collect declaration `l = length [1..10^8]` entered at prompt -------------------------------------+------------------------------------- Reporter: kabuhr | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: Component: GHCi | Version: 8.2.2 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by bgamari): Here is the code produced (from `-ddump-rn -ddump-simpl -ddump-bcos`) by the `let l` form, {{{ λ> let l = length [1..10^8] ==================== Renamer ==================== let l_a2bl = length [1 .. 10 ^ 8] ==================== Simplified expression ==================== GHC.Base.returnIO @ [()] (GHC.Types.: @ () ((Data.Foldable.length @ [] Data.Foldable.$fFoldable[] @ GHC.Integer.Type.Integer (GHC.Enum.enumFromTo @ GHC.Integer.Type.Integer GHC.Enum.$fEnumInteger 1 (GHC.Real.^ @ GHC.Integer.Type.Integer @ GHC.Integer.Type.Integer GHC.Num.$fNumInteger GHC.Real.$fIntegralInteger 10 8))) `cast` (UnsafeCo representational GHC.Types.Int () :: (GHC.Types.Int :: *) ~R# (() :: *))) (GHC.Types.[] @ ())) ==================== Proto-BCOs ==================== ProtoBCO ExprTopLevel_E0#0 []: let sat_s3Wi = ... in ... bitmap: 0 [] ALLOC_AP 0 PUSH_BCO ProtoBCO sat_s3Wi#0 []: let sat_s3Wh = ... in ... bitmap: 0 [] ALLOC_AP 0 PUSH_BCO ProtoBCO sat_s3Wh#0 []: let sat_s3Wg = ... in ... bitmap: 0 [] ALLOC_AP 0 PUSH_BCO ProtoBCO sat_s3Wg#0 []: let sat_s3Wf = ... in ... bitmap: 0 [] PUSH_UBX (1) 8# PACK GHC.Integer.Type.S# 1 PUSH_UBX (1) 10# PACK GHC.Integer.Type.S# 1 PUSH_LL 1 0 PUSH_G GHC.Real.$fIntegralInteger PUSH_G GHC.Num.$fNumInteger PUSH_APPLY_PPPP PUSH_G GHC.Real.^ SLIDE 6 2 ENTER MKAP 0 words, 1 stkoff PUSH_UBX (1) 1# PACK GHC.Integer.Type.S# 1 PUSH_LL 1 0 PUSH_G GHC.Enum.$fEnumInteger PUSH_APPLY_PPP PUSH_G GHC.Enum.enumFromTo SLIDE 5 2 ENTER MKAP 0 words, 1 stkoff PUSH_L 0 PUSH_G Data.Foldable.$fFoldable[] PUSH_APPLY_PP PUSH_G Data.Foldable.length SLIDE 4 1 ENTER MKAP 0 words, 1 stkoff PUSH_G GHC.Types.[] PUSH_L 1 PACK : 2 PUSH_L 0 PUSH_APPLY_P PUSH_G GHC.Base.returnIO SLIDE 3 2 ENTER }}} Here is the same output for the `let`-less form, {{{ λ> l = length [1..10^8] ==================== Renamer ==================== interactive:Ghci2.l = length [1 .. 10 ^ 8] ==================== Tidy Core ==================== Result size of Tidy Core = {terms: 25, types: 11, coercions: 0, joins: 0/0} -- RHS size: {terms: 10, types: 5, coercions: 0, joins: 0/0} l :: Int [GblId] l = length @ [] Data.Foldable.$fFoldable[] @ Integer (enumFromTo @ Integer GHC.Enum.$fEnumInteger 1 (^ @ Integer @ Integer GHC.Num.$fNumInteger GHC.Real.$fIntegralInteger 10 8)) -- RHS size: {terms: 1, types: 0, coercions: 0, joins: 0/0} $trModule1_r3XB :: GHC.Prim.Addr# [GblId, Caf=NoCafRefs] $trModule1_r3XB = "interactive"# -- RHS size: {terms: 2, types: 0, coercions: 0, joins: 0/0} $trModule2_r3XM :: GHC.Types.TrName [GblId, Caf=NoCafRefs] $trModule2_r3XM = GHC.Types.TrNameS $trModule1_r3XB -- RHS size: {terms: 1, types: 0, coercions: 0, joins: 0/0} $trModule3_r3XN :: GHC.Prim.Addr# [GblId, Caf=NoCafRefs] $trModule3_r3XN = "Ghci2"# -- RHS size: {terms: 2, types: 0, coercions: 0, joins: 0/0} $trModule4_r3XO :: GHC.Types.TrName [GblId, Caf=NoCafRefs] $trModule4_r3XO = GHC.Types.TrNameS $trModule3_r3XN -- RHS size: {terms: 3, types: 0, coercions: 0, joins: 0/0} interactive:Ghci2.$trModule :: GHC.Types.Module [GblId, Caf=NoCafRefs] interactive:Ghci2.$trModule = GHC.Types.Module $trModule2_r3XM $trModule4_r3XO ==================== Proto-BCOs ==================== ProtoBCO sat_s3XU#0 []: let sat_s3XT = ... in ... bitmap: 0 [] ALLOC_AP 0 PUSH_BCO ProtoBCO sat_s3XT#0 []: let sat_s3XS = ... in ... bitmap: 0 [] PUSH_UBX (1) 8# PACK GHC.Integer.Type.S# 1 PUSH_UBX (1) 10# PACK GHC.Integer.Type.S# 1 PUSH_LL 1 0 PUSH_G GHC.Real.$fIntegralInteger PUSH_G GHC.Num.$fNumInteger PUSH_APPLY_PPPP PUSH_G GHC.Real.^ SLIDE 6 2 ENTER MKAP 0 words, 1 stkoff PUSH_UBX (1) 1# PACK GHC.Integer.Type.S# 1 PUSH_LL 1 0 PUSH_G GHC.Enum.$fEnumInteger PUSH_APPLY_PPP PUSH_G GHC.Enum.enumFromTo SLIDE 5 2 ENTER ProtoBCO Ghci2.l#0 []: Data.Foldable.length @ [] Data.Foldable.$fFoldable[] @ GHC.Integer.Type.Integer sat_s3XU bitmap: 0 [] PUSH_G sat_s3XU PUSH_G Data.Foldable.$fFoldable[] PUSH_APPLY_PP PUSH_G Data.Foldable.length ENTER ProtoBCO $trModule2_r3XM#0 []: GHC.Types.TrNameS $trModule1_r3XB bitmap: 0 [] PUSH_UBX (1) 26770624## PACK GHC.Types.TrNameS 1 ENTER ProtoBCO $trModule4_r3XO#0 []: GHC.Types.TrNameS $trModule3_r3XN bitmap: 0 [] PUSH_UBX (1) 26770592## PACK GHC.Types.TrNameS 1 ENTER ProtoBCO Ghci2.$trModule#0 []: GHC.Types.Module $trModule2_r3XM $trModule4_r3XO bitmap: 0 [] PUSH_G $trModule4_r3XO PUSH_G $trModule2_r3XM PACK GHC.Types.Module 2 ENTER }}} It looks to me like the BCO pass is somehow floating out the `enumFromTo` application, which then gets retained. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/14789#comment:2 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler
participants (1)
-
GHC