February 2016 - ghc-tickets

[GHC] #10397: Compiler performance regression 7.6 -> 7.8 in elimCommonBlocks
by GHC 14 Sep '17

14 Sep '17

#10397: Compiler performance regression 7.6 -> 7.8 in elimCommonBlocks -------------------------------------+------------------------------------- Reporter: | Owner: TobyGoodwin | Status: new Type: bug | Milestone: Priority: normal | Version: 7.8.4 Component: Compiler | Operating System: Unknown/Multiple Keywords: | Type of failure: None/Unknown performance | Blocked By: Architecture: | Related Tickets: Unknown/Multiple | Test Case: see ticket | Blocking: | Differential Revisions: | -------------------------------------+------------------------------------- I've been studying a compiler performance problem in ghc-7.8.4 which is not seen in ghc-7.6.3. Profiling shows the compiler spends 74.3% of its time in `elimCommonBlocks` (in `compiler/cmm/CmmCommonBlockElim.hs`). I came across the problem in real code, and rwbarton on #ghc was able to turn it into a sensible test case, which can be found at http://static.tobold.org/ghc-prof/code/ It's not really a regression, since `elimCommonBlocks` is not enabled by default in 7.6. It can be turned on with `-fnew-codegen`, and that demonstrates a similar slowdown. However 7.8 is worse than 7.6, and 7.10 slightly worse still, for reasons that I haven't explored. Compile times for `RegBig.hs` are: {{{ 7.6.3 -O2 17s 7.6.3 -O2 -fnew-codegen 177s 7.8.4 -O2 230s 7.10.1 -O2 241s }}} The problem obviously stems from the applicative expression with lots of branches `Foo <$> a <*> b <*> c <*> ...` It seems that some level of inlining occurs which means that each extra term in this expression doubles the size of the code. Incidentally, this blowup only occurs when both `RegBig.hs` ''and'' `Form.hs` are compiled with `-O2`. I haven't looked to see why the blowup occurs; although I do see that a similar blowup happens with 7.6. {{{ ... *** Simplifier: Result size of Simplifier iteration=1 = {terms: 16,525, types: 31,503, coercions: 73} Result size of Simplifier iteration=2 = {terms: 29,070, types: 45,079, coercions: 73} Result size of Simplifier iteration=3 = {terms: 50,139, types: 62,057, coercions: 73} Result size of Simplifier iteration=4 = {terms: 84,172, types: 83,805, coercions: 73} Result size of Simplifier = {terms: 84,172, types: 83,805, coercions: 73} ... }}} I've looked in some detail at what happens next, tinkering with `elimCommonBlocks` to try to understand why it runs so slowly in this case. There are two major problems. First, `elimCommonBlocks` is worse than O(n^2^), and the inlining blowup mentioned above results in a very large input: 51695 blocks in this example. (Across the project of 50 files or so where this problem originated, I see the input to `elimCommonBlocks` is normally a few hundred blocks at most.) Secondly, there is an effort made in `elimCommonBlocks` to avoid comparing whole blocks all the time, by computing a hash value of some of the block. This fails to produce sufficient unique hash values for this case. The hash must obviously omit the unique block label, and in practice all labels, and various other things, are omitted. In this case, presumably because the code is blown up from the tiny input, many many blocks are similar, and the differences are elided by the hash function. For example, 8177 blocks share the hash value 1094. These 8177 similar blocks all end up in a normal list, which is searched with Data.List.find and the full-block comparison function! Eventually this process concludes that the vast majority are in fact different: for all its hard work, `elimCommonBlocks` only finds about a dozen common blocks in total. Two blocks which both hash to 1094 are these: {{{ cCHc: _se0q::P64 = R1; _c1grM::P64 = _se0q::P64 & 7; if (_c1grM::P64 >= 2) goto c1grR; else goto c1grS; cWK7: _sig3::P64 = R1; _c1iXk::P64 = _sig3::P64 & 7; if (_c1iXk::P64 >= 2) goto c1iXp; else goto c1iXq; }}} I suggest the following steps need to be taken. == 1. Mitigation == Since this is a real problem in the 7.8 and 7.10 series, it needs a simple but effective mitigation. I suggest a test that the input to `elimCommonBlocks` is not "too big", where 1000 seems a reasonable cutoff. Another simple mitigation would be to disable common block elimination altogether (after all, `-O2` is usually understood to mean "make my code go faster even if it gets bigger"). I've tried both these possibilities with 7.8.4. The trivial patches are http://static.tobold.org/ghc-prof/0001-cbe-cutoff.patch and http://static.tobold.org/ghc-prof/0001-no-cbe.patch Compilation times for the test case `RegBig.hs` are as follows, and there is no difference in size in the `.o` file. {{{ 7.8.4 -O2 230s cbe only if input < 1000 blocks 88s no common block elimination 86s }}} == 2. Investigation == I focused on `elimCommonBlocks` thinking that the inline blowup seen earlier in the compilation is reasonable. On reflection, this is probably not so, and I think someone better acquainted with ghc internals should investigate this. == 3. Optimization == `-- TODO: Use optimization fuel` it says in `CmmCommonBlockElim.hs` Since `elimCommonBlocks` is essentially doing a `nub`, it will never perform better than O(n^2^). In fact, it's currently much worse than that as the comparison function changes as duplicate blocks are found, which in turn may create new duplicates, so we have to iterate the nub function till no more dups are found. I did experiment with improving the hash function. It's possible to include target labels, at the expense of needing to recompute block hash values on each iteration. (And currently some ''highly'' ugly code to convert labels to integers.) This considerably reduces the length of some of the hash bucket lists, and so offers a significant speedup: {{{ ghc-7.8.4 -O2 230s "hack" 170s }}} Patch for this is http://static.tobold.org/ghc-prof/0001-cbe-hacks.patch This looks like a promising approach. Including target labels doesn't eliminate all the large hash buckets, so there is scope for further improvements along these lines. I don't see any reason why the hash function should not be as discerning as the equality function, with just occasional accidental clashes. I've also pondered whether more radical rewriting of `elimCommonBlocks` might pay dividends. It's a commonplace that, if you don't need to preserve order, `map head . group . sort` is a faster `nub`. Is the order of blocks significant? I'm not sure - they all seem to end with either `goto` or `call`. I'd ''assume'' that `call` then falls through to the next block but I'm not sure. Anyway, presumably we could `zip [0..]` first, and sort again afterwards, and that should still be asymptotically faster than `nub`. Of course, this presupposes that blocks could be (sensibly? efficiently?) made an instance of `Ord`. (If they were, hash buckets could be stored in an O(1) data structure and binary search employed.) I may be able to find time to pursue some of these avenues, but as a very novice ghc hacker, I'd need some strategic direction here. Any thoughts? -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10397> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 42

[GHC] #8091: retry# lacks strictness information
by GHC 13 Sep '17

13 Sep '17

#8091: retry# lacks strictness information ------------------------------------+------------------------------------- Reporter: parcs | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.6.3 Keywords: | Operating System: Unknown/Multiple Architecture: Unknown/Multiple | Type of failure: None/Unknown Difficulty: Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | ------------------------------------+------------------------------------- The `retry#` primop should be marked as returning bottom since it never actually returns anything as far as the simplifier is concerned. This change would help the simplifier eliminate unreachable code inside STM transactions. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8091> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 13

[GHC] #8532: Hyperbolic arc cosine fails on (-1) :: Complex r.
by GHC 13 Sep '17

13 Sep '17

#8532: Hyperbolic arc cosine fails on (-1) :: Complex r. -------------------------------------+------------------------------------- Reporter: leftaroundabout | Owner: Type: bug | Status: new Priority: low | Milestone: Component: libraries/base | Version: 7.6.3 Keywords: | Operating System: Unknown/Multiple Architecture: Unknown/Multiple | Type of failure: Incorrect result Difficulty: Easy (less than 1 | at runtime hour) | Test Case: acosh(-1) :: Blocked By: | Complex Double Related Tickets: | Blocking: -------------------------------------+------------------------------------- When allowing for complex results, the hyperbolic arc cosine is continuously defined on all ℝ. In the {{{(x < (-1))}}} real ray of the complex plane, {{{acosh}}} equals {{{\z -> i * pi + acosh(abs z)}}}, which works fine for almost all arguments. Thus, {{{acosh (-1)}}} should equal {{{i * pi}}}; however due to the implementation as {{{ acosh z = log (z + (z+1) * sqrt ((z-1)/(z+1))) }}} where the denominator in the root becomes zero at {{{z = -1}}}, this comes out as {{{NaN :+ NaN}}}. Could be fixed trivially by adding a special case {{{ acosh ((-1):+0) = 0:+pi }}} to the {{{instance (RealFloat a) => Floating (Complex a)}}} in {{{Data.Complex}}}. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8532> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 13

[GHC] #8583: Associated pattern synonyms
by GHC 11 Sep '17

11 Sep '17

#8583: Associated pattern synonyms ------------------------------------+------------------------------------- Reporter: cactus | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: Keywords: | Operating System: Unknown/Multiple Architecture: Unknown/Multiple | Type of failure: None/Unknown Difficulty: Unknown | Test Case: Blocked By: 5144 | Blocking: Related Tickets: | ------------------------------------+------------------------------------- The PatternSynonyms wiki page has a section on (typeclass-)associated pattern synonyms: {{{ class ListLike l where pattern Nil :: l a pattern Cons :: a -> l a -> a isNil :: l a -> Bool isNil Nil = True isNil (Cons _ _) = False append :: l a -> l a -> l a instance ListLike [] where pattern Nil = [] pattern Cons x xs = x:xs append = (++) headOf :: (ListLike l) => l a -> Maybe a headOf Nil = Nothing headOf (Cons x _) = Just x }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8583> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 18

[GHC] #10818: GHC 7.10.2 takes much longer to compile some packages
by GHC 11 Sep '17

11 Sep '17

#10818: GHC 7.10.2 takes much longer to compile some packages -------------------------------------+------------------------------------- Reporter: kazu- | Owner: yamamoto | Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.10.2 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: Differential Revisions: | -------------------------------------+------------------------------------- If we compile the iproute package by GHC, GHC 7.10.2 take much longer than GHC 7.8.4. See the report from Travis CI: https://travis-ci.org/kazu-yamamoto/iproute/builds/77427248 I can reproduce this on my local MacBookAir: - GHC 7.8.4: cabal build 8.81s user 0.94s system 103% cpu 9.430 total - GHC 7.10.2: cabal build 37.86s user 3.49s system 98% cpu 42.066 total -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10818> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 25

[GHC] #9661: Branchless ==# is compiled to branchy code
by GHC 08 Sep '17

08 Sep '17

#9661: Branchless ==# is compiled to branchy code -------------------------------------+------------------------------------- Reporter: dfeuer | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.9 Keywords: | Operating System: Architecture: Unknown/Multiple | Unknown/Multiple Difficulty: Unknown | Type of failure: Runtime Blocked By: | performance bug Related Tickets: | Test Case: | Blocking: | Differential Revisions: -------------------------------------+------------------------------------- This started as a comment on #6135, but Jan Stolarek requested that I file a separate report. The branchless tests `<#`, `>#`, `<=#`, and `>=#`, and their wordy cousins, seem to work properly, but `==#` and `eqWord#` don't. If I write {{{#!hs foo x = tagToEnum# ((x <# 3#) `orI#` (x ># 100#) `orI#` (x ==# 12#) `orI#` (x ==# 15#)) }}} I get (in 7.8.3 and in 7.9) {{{#!hs IsDigit.foo :: GHC.Prim.Int# -> GHC.Types.Bool [GblId, Arity=1, Caf=NoCafRefs, Str=DmdType <S,1*U>, Unf=Unf{Src=InlineStable, TopLvl=True, Value=True, ConLike=True, WorkFree=True, Expandable=True, Guidance=ALWAYS_IF(arity=1,unsat_ok=True,boring_ok=False) Tmpl= \ (x3_a2if [Occ=Once!] :: GHC.Prim.Int#) -> case x3_a2if of wild_Xe { __DEFAULT -> GHC.Prim.tagToEnum# @ GHC.Types.Bool (GHC.Prim.orI# (GHC.Prim.<# wild_Xe 3) (GHC.Prim.># wild_Xe 100)); 12 -> GHC.Types.True; 15 -> GHC.Types.True }}] IsDigit.foo = \ (x3_a2if :: GHC.Prim.Int#) -> case x3_a2if of wild_Xe { __DEFAULT -> GHC.Prim.tagToEnum# @ GHC.Types.Bool (GHC.Prim.orI# (GHC.Prim.<# wild_Xe 3) (GHC.Prim.># wild_Xe 100)); 12 -> GHC.Types.True; 15 -> GHC.Types.True }}} and branching assembly to match. Jan Stolarek indicates this is probably a problem with constant folding. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9661> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 30

[GHC] #10365: Make Semigroup as a superclass of Monoid
by GHC 07 Sep '17

07 Sep '17

#10365: Make Semigroup as a superclass of Monoid -------------------------------------+------------------------------------- Reporter: gidyn | Owner: Type: feature | Status: new request | Milestone: Priority: normal | Version: 7.10.1 Component: | Operating System: Unknown/Multiple libraries/base | Type of failure: None/Unknown Keywords: | Blocked By: Architecture: | Related Tickets: Unknown/Multiple | Test Case: | Blocking: | Differential Revisions: | -------------------------------------+------------------------------------- See http://haskell.1045720.n5.nabble.com/Proposal-Make-Semigroup-as-a -superclass-of-Monoid-td5767835.html -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10365> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 45

[GHC] #9797: Investigate rewriting `>>=` to `*>` or `>>` for appropriate types
by GHC 05 Sep '17

05 Sep '17

#9797: Investigate rewriting `>>=` to `*>` or `>>` for appropriate types -------------------------------------+------------------------------------- Reporter: dfeuer | Owner: dfeuer Type: task | Status: new Priority: normal | Milestone: 7.10.1 Component: Core Libraries | Version: 7.9 Keywords: | Operating System: Architecture: Unknown/Multiple | Unknown/Multiple Difficulty: Unknown | Type of failure: Runtime Blocked By: | performance bug Related Tickets: | Test Case: | Blocking: | Differential Revisions: -------------------------------------+------------------------------------- When we see `m >>= (\_ -> n)` with a type that admits an optimized `>>` or `<$`, we should try to take advantage of that. I don't currently know if this applies to any types under "GHC HQ" control. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9797> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 4

[GHC] #10328: Control.Monad exports lead to weird Haddocks
by GHC 05 Sep '17

05 Sep '17

#10328: Control.Monad exports lead to weird Haddocks -------------------------------------+------------------------------------- Reporter: dfeuer | Owner: ekmett Type: feature | Status: new request | Milestone: Priority: normal | Version: 7.10.1 Component: Core | Operating System: Unknown/Multiple Libraries | Type of failure: Documentation Keywords: | bug Architecture: | Blocked By: Unknown/Multiple | Related Tickets: Test Case: | Blocking: | Differential Revisions: | -------------------------------------+------------------------------------- `Control.Monad` exports, and therefore documents, the `Functor` class (but only the `fmap` member) and the `Monad` class. It does not export, and therefore does not document, the `Applicative` class. This is all pretty weird. Since `Functor` is in the prelude (and probably most/all alternate preludes), is there actually a reason to export it from `Control.Monad`? -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10328> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 4

[GHC] #9756: Warn about unnecessary unsafeCoerce
by GHC 05 Sep '17

05 Sep '17

#9756: Warn about unnecessary unsafeCoerce -------------------------------------+------------------------------------- Reporter: dfeuer | Owner: Type: feature request | Status: new Priority: normal | Milestone: Component: Compiler (Type | Version: 7.9 checker) | Operating System: Keywords: | Unknown/Multiple Architecture: Unknown/Multiple | Type of failure: Difficulty: Unknown | None/Unknown Blocked By: | Test Case: Related Tickets: | Blocking: | Differential Revisions: -------------------------------------+------------------------------------- `unsafeCoerce` has been around a lot longer than `coerce`, so older code tends to use it even if `coerce` could do the job. It would be nice if the type checker could produce a warning when it detected such a situation. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9756> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 4