[GHC] #12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- Comparing the core output for a small program of mine I found that GHC HEAD produced binary runs 26x slower than with GHC 8.0.1. I have uploaded an example: https://gist.github.com/alexbiehl/0a1b5016223e00ae79a1399176e14eef The following is the output for the empResult function. We can see that GHC 8.0.1 nicely unboxed the accumulator in the loop. While GHC HEAD uses boxed values all over the place and doesn't even do dictionary inlining. For GHC 8.0.1 it produces: {{{#!hs -- RHS size: {terms: 175, types: 184, coercions: 4} $wempResult :: IO (Maybe (Vector ColValue)) -> State# RealWorld -> (# State# RealWorld, Either Error Int #) $wempResult = \ (ww :: IO (Maybe (Vector ColValue))) (w :: State# RealWorld) -> letrec { $sloop :: State# RealWorld -> Int# -> (# State# RealWorld, Either Error Int #) $sloop = \ (sc :: State# RealWorld) (sc1 :: Int#) -> case (ww `cast` ...) sc of _ { (# ipv, ipv1 #) -> case ipv1 of _ { Nothing -> (# ipv, Right (I# sc1) #); Just a -> case length $fVectorVectora a of _ { I# y -> case y of _ { __DEFAULT -> (# ipv, empResult2 #); 4# -> case a of _ { Vector dt dt1 dt2 -> let { $wsucc_ :: Int# -> State# RealWorld -> (# State# RealWorld, Either Error Int #) $wsucc_ = \ (ww1 :: Int#) (w1 :: State# RealWorld) -> let { $wsucc_1 :: Int# -> State# RealWorld -> (# State# RealWorld, Either Error Int #) $wsucc_1 = \ (ww2 :: Int#) (w2 :: State# RealWorld) -> case indexArray# dt2 (+# dt ww2) of _ { (# ipv2 #) -> case ipv2 of _ { __DEFAULT -> (# w2, empResult2 #); CV_Int8 dt4 -> case indexArray# dt2 (+# dt (+# ww2 1#)) of _ { (# ipv3 #) -> case ipv3 of _ { __DEFAULT -> (# w2, empResult2 #); CV_Text t -> $sloop w2 (+# sc1 1#) } }; CV_Int16 dt4 -> case indexArray# dt2 (+# dt (+# ww2 1#)) of _ { (# ipv3 #) -> case ipv3 of _ { __DEFAULT -> (# w2, empResult2 #); CV_Text t -> $sloop w2 (+# sc1 1#) } }; CV_Int32 dt4 -> case indexArray# dt2 (+# dt (+# ww2 1#)) of _ { (# ipv3 #) -> case ipv3 of _ { __DEFAULT -> (# w2, empResult2 #); CV_Text t -> $sloop w2 (+# sc1 1#) } } } } } in case indexArray# dt2 (+# dt ww1) of _ { (# ipv2 #) -> case ipv2 of _ { __DEFAULT -> (# w1, empResult2 #); CV_Int8 dt4 -> $wsucc_1 (+# ww1 1#) w1; CV_Int16 dt4 -> $wsucc_1 (+# ww1 1#) w1; CV_Int32 dt4 -> $wsucc_1 (+# ww1 1#) w1 } } } in case indexArray# dt2 dt of _ { (# ipv2 #) -> case ipv2 of _ { __DEFAULT -> (# ipv, empResult2 #); CV_Int8 dt4 -> $wsucc_ 1# ipv; CV_Int16 dt4 -> $wsucc_ 1# ipv; CV_Int32 dt4 -> $wsucc_ 1# ipv } } } } } } }; } in $sloop w 0# }}} and for GHC HEAD it produces {{{#!hs -- RHS size: {terms: 193, types: 182, coercions: 3} empResult :: Result Int empResult = case <$> $fFunctorRow $WEmp lvl19 of { Row dt fm -> case + $fNumInt (I# dt) lvl17 of dt1 { I# dt2 -> case + $fNumInt dt1 lvl17 of dt3 { I# dt4 -> case + $fNumInt dt3 lvl17 of dt5 { I# dt6 -> (\ (is :: InputStream (Vector ColValue)) -> let { lvl23 :: IO (Maybe (Vector ColValue)) lvl23 = case is of { InputStream ds1 ds2 -> ds1 } } in $! (letrec { loop :: Int -> IO (Either Error Int) loop = \ (s :: Int) -> let { lvl24 :: IO (Either Error Int) lvl24 = $! loop (+ $fNumInt s lvl17) } in let { lvl25 :: IO (Either Error Int) lvl25 = return $fMonadIO (Right s) } in >>= $fMonadIO lvl23 (\ (ma :: Maybe (Vector ColValue)) -> case ma of { Nothing -> lvl25; Just a -> case == $fEqInt dt5 (lvl13 a) of { False -> lvl21; True -> fm (\ _ (j :: Int) -> case a of { Vector dt7 dt8 dt9 -> case + $fNumInt (I# dt7) j of { I# i# -> let { $wsucc_ :: Int -> IO (Either Error Int) $wsucc_ = \ (w :: Int) -> case + $fNumInt (I# dt7) w of { I# i#1 -> case indexArray# dt9 i#1 of { (# ipv #) -> case ipv of { __DEFAULT -> lvl21; CV_Int8 dt10 -> case + $fNumInt (I# dt7) (+ $fNumInt w lvl17) of { I# i#2 -> case indexArray# dt9 i#2 of { (# ipv1 #) -> case ipv1 of { __DEFAULT -> lvl21; CV_Text t -> lvl24 } } }; CV_Int16 dt10 -> case + $fNumInt (I# dt7) (+ $fNumInt w lvl17) of { I# i#2 -> case indexArray# dt9 i#2 of { (# ipv1 #) -> case ipv1 of { __DEFAULT -> lvl21; CV_Text t -> lvl24 } } }; CV_Int32 dt10 -> case + $fNumInt (I# dt7) (+ $fNumInt w lvl17) of { I# i#2 -> case indexArray# dt9 i#2 of { (# ipv1 #) -> case ipv1 of { __DEFAULT -> lvl21; CV_Text t -> lvl24 } } } } } } } in case indexArray# dt9 i# of { (# ipv #) -> case ipv of { __DEFAULT -> lvl21; CV_Int8 dt10 -> $wsucc_ (+ $fNumInt j lvl17); CV_Int16 dt10 -> $wsucc_ (+ $fNumInt j lvl17); CV_Int32 dt10 -> $wsucc_ (+ $fNumInt j lvl17) } } } }) lvl22 a $fShowColValue2 } }); } in loop) $fShowColValue2) `cast` ... } } } } }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by RyanGlScott): I'm having difficulty reproducing this with a build of GHC HEAD at commit ddc271e8ed6c5ec5e83dd50c6c5e77955a0e90ac. In fact, the Core I get looks just about identical to the one that GHC 8.0.1 produces (with `$wempResult`). FWIW, this is using `io-streams-1.3.5.0` and `vector-0.11.0.0`. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916#comment:1> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by simonpj): If GHC starts producing bad code, it's important to look into it, so thanks for reporting. Let hope it's a false alarm, though, in the light of comment:1 -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916#comment:2> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by alexbiehl): Strange, my ghc-stage2 is at the same git revision! This is how I invoked it to get to the core: {{{#!bash /Users/alexbiehl/git/ghc/inplace/bin/ghc-stage2 test.hs -O2 -threaded -rtsopts -ddump-simpl -dsupress-idinfo -dsuppress-coercions -dsuppress- type-applications -dsuppress-uniques -dsuppress-module-prefixes -fforce- recomp | less }}} (where test.hs is the decode.hs from github link) -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916#comment:3> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by RyanGlScott): Well, something must be different... {{{ ghc4/inplace/bin/ghc-stage2 Bug.hs -O2 -threaded -rtsopts -ddump-simpl -dsuppress-idinfo -dsuppress-coercions -dsuppress-type-applications -dsuppress-uniques -dsuppress-module-prefixes -fforce-recomp [1 of 1] Compiling X ( Bug.hs, Bug.o ) ... -- RHS size: {terms: 2, types: 2, coercions: 0} empResult2 :: Either Error Int empResult2 = Left Error -- RHS size: {terms: 175, types: 184, coercions: 4} $wempResult :: IO (Maybe (Vector ColValue)) -> State# RealWorld -> (# State# RealWorld, Either Error Int #) $wempResult = \ (ww :: IO (Maybe (Vector ColValue))) (w :: State# RealWorld) -> letrec { $sloop :: State# RealWorld -> Int# -> (# State# RealWorld, Either Error Int #) $sloop = \ (sc :: State# RealWorld) (sc1 :: Int#) -> case (ww `cast` ...) sc of { (# ipv, ipv1 #) -> case ipv1 of { Nothing -> (# ipv, Right (I# sc1) #); Just a -> case length $fVectorVectora a of { I# y -> case y of { __DEFAULT -> (# ipv, empResult2 #); 4# -> case a of { Vector dt dt1 dt2 -> let { $wsucc_ :: Int# -> State# RealWorld -> (# State# RealWorld, Either Error Int #) $wsucc_ = \ (ww1 :: Int#) (w1 :: State# RealWorld) -> let { $wsucc_1 :: Int# -> State# RealWorld -> (# State# RealWorld, Either Error Int #) $wsucc_1 = \ (ww2 :: Int#) (w2 :: State# RealWorld) -> case indexArray# dt2 (+# dt ww2) of { (# ipv2 #) -> case ipv2 of { __DEFAULT -> (# w2, empResult2 #); CV_Int8 dt4 -> case indexArray# dt2 (+# dt (+# ww2 1#)) of { (# ipv3 #) -> case ipv3 of { __DEFAULT -> (# w2, empResult2 #); CV_Text t -> $sloop w2 (+# sc1 1#) } }; CV_Int16 dt4 -> case indexArray# dt2 (+# dt (+# ww2 1#)) of { (# ipv3 #) -> case ipv3 of { __DEFAULT -> (# w2, empResult2 #); CV_Text t -> $sloop w2 (+# sc1 1#) } }; CV_Int32 dt4 -> case indexArray# dt2 (+# dt (+# ww2 1#)) of { (# ipv3 #) -> case ipv3 of { __DEFAULT -> (# w2, empResult2 #); CV_Text t -> $sloop w2 (+# sc1 1#) } } } } } in case indexArray# dt2 (+# dt ww1) of { (# ipv2 #) -> case ipv2 of { __DEFAULT -> (# w1, empResult2 #); CV_Int8 dt4 -> $wsucc_1 (+# ww1 1#) w1; CV_Int16 dt4 -> $wsucc_1 (+# ww1 1#) w1; CV_Int32 dt4 -> $wsucc_1 (+# ww1 1#) w1 } } } in case indexArray# dt2 dt of { (# ipv2 #) -> case ipv2 of { __DEFAULT -> (# ipv, empResult2 #); CV_Int8 dt4 -> $wsucc_ 1# ipv; CV_Int16 dt4 -> $wsucc_ 1# ipv; CV_Int32 dt4 -> $wsucc_ 1# ipv } } } } } } }; } in $sloop w 0# -- RHS size: {terms: 8, types: 16, coercions: 0} empResult1 :: InputStream (Vector ColValue) -> State# RealWorld -> (# State# RealWorld, Either Error Int #) empResult1 = \ (w :: InputStream (Vector ColValue)) (w1 :: State# RealWorld) -> case w of { InputStream ww1 ww2 -> $wempResult ww1 w1 } -- RHS size: {terms: 1, types: 0, coercions: 13} empResult :: Result Int empResult = empResult1 `cast` ... }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916#comment:4> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by rwbarton): I haven't looked at the Core at all, but maybe the original reporter is comparing a release (perf) build of GHC 8.0.1 to a devel build of HEAD? The difference could be due to different strictness/unfolding info in a library. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916#comment:5> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by alexbiehl): That could be the case indeed: I invoked `hadrian -j --flavour=quickest` to build the compiler. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916#comment:6> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by RyanGlScott): Replying to [comment:6 alexbiehl]:
That could be the case indeed: I invoked `hadrian -j --flavour=quickest` to build the compiler.
I think `quickest` might very well be the culprit. From `mk/build.mk`: {{{ # Even faster build. NOT RECOMMENDED: the libraries will be # completely unoptimised, so any code built with this compiler # (including stage2) will run very slowly: #BuildFlavour = quickest }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916#comment:7> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by rwbarton): Right, `quickest` is basically designed to see whether the stage2 compiler works at all, as fast as possible. So it doesn't bother optimizing the libraries (and you will get a slow stage2 compiler that generates slow code). -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916#comment:8> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: closed Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Resolution: invalid | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Changes (by mpickering): * status: new => closed * resolution: => invalid -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916#comment:9> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
#12916: GHC 8.0.1 vs GHC HEAD (8.1.20161202) doesn't do inlining and unboxing very well -------------------------------------+------------------------------------- Reporter: alexbiehl | Owner: Type: bug | Status: closed Priority: normal | Milestone: Component: Compiler | Version: 8.0.1 Resolution: invalid | Keywords: Operating System: Unknown/Multiple | Architecture: | Unknown/Multiple Type of failure: None/Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by alexbiehl): Ok, thanks for the explanation! But one last question: Why is GHC 8.0.1 and HEAD even in perf flavour not able to inline the length function on the vector? {{{#!hs ... Just a -> case length $fVectorVectora a of { I# y -> case y of { ... }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12916#comment:10> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler
participants (1)
-
GHC