March 2015 - ghc-tickets

[GHC] #9279: Optimisation bug
by GHC 18 Jan '19

18 Jan '19

#9279: Optimisation bug ------------------------------+-------------------------------------------- Reporter: simonmar | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.8.2 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: Runtime performance bug Unknown/Multiple | Test Case: Difficulty: Unknown | Blocking: Blocked By: | Related Tickets: | ------------------------------+-------------------------------------------- I have a strange problem with a local function binding that is not being inlined. I've attached the repro code, build it like this: {{{ $ ghc -O2 Haxl/Core/Monad.hs }}} There are several modules, but the one with the problem is `Haxl/Core/Monad.hs`. In particular `Haxl.Core.Monad.$fApplicativeGenHaxl2`, which contains this fragment: {{{ let { $wa4_s6YQ $wa4_s6YQ = \ w_s6YF _ ww1_s6YM _ _ w1_s6YI -> case GHC.Prim.readMutVar# ipv7_X5Ne w1_s6YI of _ { (# ipv10_X5QO, ipv11_X5QQ #) -> case ipv11_X5QQ of _ { Haxl.Core.Monad.IVarFull a2_a3kK -> case lvl6_r7qi of wild3_00 { }; Haxl.Core.Monad.IVarEmpty dt2_d4Vs -> case GHC.Prim.writeMutVar# ipv7_X5Ne (Haxl.Core.Monad.IVarFull w_s6YF) ipv10_X5QO of s2#_a5OR { __DEFAULT -> case GHC.Prim.readMutVar# dt2_d4Vs s2#_a5OR of _ { (# ipv12_X5R0, ipv13_X5R2 #) -> case GHC.Prim.readMutVar# ww1_s6YM ipv12_X5R0 of _ { (# ipv14_X5SS, ipv15_X5SU #) -> letrec { go_a5ti go_a5ti = \ ds10_a5tj -> case ds10_a5tj of _ { [] -> ipv15_X5SU; : y_a5to ys_a5tp -> GHC.Types.: (y_a5to w_s6YF) (go_a5ti ys_a5tp) }; } in case go_a5ti ipv13_X5R2 of x'_a5P6 { __DEFAULT -> case GHC.Prim.writeMutVar# ww1_s6YM x'_a5P6 ipv14_X5SS of s2#1_a5P7 { __DEFAULT -> (# s2#1_a5P7, Haxl.Core.Monad.cacheRequest6 #) } } } } } } } } in let { a2_s6lH a2_s6lH = \ w_s6YF _ w2_s6YH w3_s6YI -> case w2_s6YH of _ { Haxl.Core.Monad.SchedState ww1_s6YL ww2_s6YM ww3_s6YN ww4_s6YO -> $wa4_s6YQ w_s6YF ww1_s6YL ww2_s6YM ww3_s6YN ww4_s6YO w3_s6YI } } in }}} I want `$wa4` to be inlined at its single occurrence in `a2`. I believe the reason it is not being inlined is that `a2` is a wrapper, but the situation seems silly because the wrapper isn't going away either, so we have a redundant closure being built (this is in my inner loop). -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9279> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 26

[GHC] #8326: Place heap checks common in case alternatives before the case
by GHC 07 Jan '19

07 Jan '19

#8326: Place heap checks common in case alternatives before the case ------------------------------------+------------------------------------- Reporter: jstolarek | Owner: Type: task | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.7 Keywords: | Operating System: Unknown/Multiple Architecture: Unknown/Multiple | Type of failure: None/Unknown Difficulty: Unknown | Test Case: Blocked By: | Blocking: 8317 Related Tickets: #1498 | ------------------------------------+------------------------------------- We would like to have functions that check whether an `Int#` is a valid tag to represent `Bool` (see Note [Optimizing isTrue#] in ghc-prim): {{{ isTrue# :: Int# -> Bool isTrue# 1# = True isTrue# _ = False isFalse# :: Int# -> Bool isFalse# 0# = True isFalse# _ = False }}} We could use them with comparison primops like this: {{{ f :: Int# -> Int f x | isTrue# (x ># 0#) = I# x | otherwise = -(I# x) }}} `isTrue#` is optimized away at the Core level: {{{ A.f = \ (x_aqM :: GHC.Prim.Int#) -> case GHC.Prim.># x_aqM 0 of _ { __DEFAULT -> GHC.Types.I# (GHC.Prim.negateInt# x_aqM); 1 -> GHC.Types.I# x_aqM } }}} but the code genrator produces very bad Cmm code, because it pushes heap checks into case alternatives: {{{ {offset cFd: // stack check if ((Sp + -16) < SpLim) goto cFr; else goto cFs; cFr: // not enough place on the stack, call GC R2 = R2; R1 = A.f_closure; call (stg_gc_fun)(R2, R1) args: 8, res: 0, upd: 8; cFs: // scrutinize (x ># 0#) _sEU::I64 = R2; _sEV::I64 = %MO_S_Gt_W64(R2, 0); if (_sEV::I64 != 1) goto cFg; else goto cFo; cFg: // False branch Hp = Hp + 16; if (Hp > HpLim) goto cFy; else goto cFx; cFy: // not enough heap, call GC HpAlloc = 16; I64[Sp - 16] = cFf; R1 = _sEV::I64; I64[Sp - 8] = _sEU::I64; Sp = Sp - 16; call stg_gc_unbx_r1(R1) returns to cFf, args: 8, res: 8, upd: 8; cFf: // re-do the False branch _sEU::I64 = I64[Sp + 8]; Sp = Sp + 16; _sEV::I64 = R1; goto cFg; cFx: // RHS of False branch I64[Hp - 8] = GHC.Types.I#_con_info; I64[Hp] = -_sEU::I64; R1 = Hp - 7; call (P64[Sp])(R1) args: 8, res: 0, upd: 8; cFo: // True branch Hp = Hp + 16; if (Hp > HpLim) goto cFv; else goto cFu; cFv: // not enough heap, call GC HpAlloc = 16; I64[Sp - 16] = cFn; R1 = _sEV::I64; I64[Sp - 8] = _sEU::I64; Sp = Sp - 16; call stg_gc_unbx_r1(R1) returns to cFn, args: 8, res: 8, upd: 8; cFn: // re-do the True branch _sEU::I64 = I64[Sp + 8]; Sp = Sp + 16; _sEV::I64 = R1; goto cFo; cFu: // RHS of True branch I64[Hp - 8] = GHC.Types.I#_con_info; I64[Hp] = _sEU::I64; R1 = Hp - 7; call (P64[Sp])(R1) args: 8, res: 0, upd: 8; } }}} This results in average 2.5% increase in binary size. By contrast, if we use `tagToEnum#` instead of `isTrue#` heap check will be placed before `case` expression and the code will be significantly shorter (this is done by a special case-on-bool optimization in the code generator - see #8317). What we would like to do here is: 1. compile case alternatives without placing heap checks inside them 2. each compiled alternative should return amount of heap it needs to allocate 3. code generator inspects amounts of heap needed by each alternative and either adds heap checks in alternatives or puts a single check before the case expression. Getting this right might be a bit tricky. 1. if all branches allocate some heap then we can just put a common heap check before the case. Note that we must allocate the higgest amount required by any of the alternatives and then alternatives that use less heap must retract the heap pointer accordingly. 2. if we have two alternatives, one of which allocates heap and the other does not, we should place the heap check only in the alternative that allocates the stack. This will solve #1498. 3. it is not clear to me what to do if we have combination of the above (more than one branch that allocates heap and at least one branch that does not). If we place heap check before the `case` expression we lose optimization of recursive functions and face the problem described in #1498. If we push heap checks into branches that allocate heap then we get code duplication, i.e. the problem that we're addressing in this ticket. I guess the only way to make correct decission here is to try different aproaches and measure their performance. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8326> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 48

[GHC] #8611: nofib’s cacheprof’s allocations nondeterminisitic
by GHC 01 Jan '19

01 Jan '19

#8611: nofib’s cacheprof’s allocations nondeterminisitic -----------------------------------------+--------------------------------- Reporter: nomeata | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: NoFib benchmark suite | Version: 7.6.3 Keywords: | Operating System: Architecture: Unknown/Multiple | Unknown/Multiple Difficulty: Unknown | Type of failure: None/Unknown Blocked By: | Test Case: Related Tickets: | Blocking: -----------------------------------------+--------------------------------- This seems to be neither expected nor desired, and hence worth investigating: {{{ ./cacheprof +RTS -t <<ghc: 127588752 bytes, 243 GCs, 2261288/6551032 avg/max bytes residency (6 samples), 15M in use, 0.00 INIT (0.00 elapsed), 0.12 MUT (0.12 elapsed), 0.05 GC (0.05 elapsed) :ghc>> ./cacheprof +RTS -t <<ghc: 127392144 bytes, 242 GCs, 2261304/6551576 avg/max bytes residency (6 samples), 15M in use, 0.00 INIT (0.00 elapsed), 0.11 MUT (0.11 elapsed), 0.04 GC (0.05 elapsed) :ghc>> }}} The allocations numbers change, when they shoudn’t. Reproduced with 7.6.3 and current HEAD. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8611> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 16

[GHC] #8316: GHCi debugger segfaults when trying force a certain variable
by GHC 28 Dec '18

28 Dec '18

#8316: GHCi debugger segfaults when trying force a certain variable ------------------------------------+------------------------------------- Reporter: guest | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: GHCi | Version: 7.6.3 Keywords: | Operating System: Unknown/Multiple Architecture: Unknown/Multiple | Type of failure: GHCi crash Difficulty: Unknown | Test Case: Blocked By: | Blocking: Related Tickets: | ------------------------------------+------------------------------------- The file Test.hs has following definitions: {{{ whnf :: a -> IO () whnf a = a `seq` (return ()) foo :: [Int] foo = [1..] }}} Calling ghci as: {{{ ghci Test.hs -ignore-dot-ghci }}} and bebugging foo like this: {{{ *Main> :b foo Breakpoint 0 activated at Test.hs:5:7-11 *Main> foo Stopped at Test.hs:5:7-11 _result :: [Int] = _ [Test.hs:5:7-11] *Main> :p foo foo = (_t1::[Int]) [Test.hs:5:7-11] *Main> whnf _t1 }}} results in this segault: {{{ <interactive>: internal error: TSO object entered! (GHC version 7.6.3 for x86_64_unknown_linux) Please report this as a GHC bug: http://www.haskell.org/ghc/reportabug [1] 5445 abort (core dumped) ghci Test.hs -ignore-dot-ghci }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8316> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 25

[GHC] #9627: Type error with functional dependencies
by GHC 24 Dec '18

24 Dec '18

#9627: Type error with functional dependencies -------------------------------------+------------------------------------- Reporter: augustss | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.8.3 Keywords: | Operating System: Architecture: Unknown/Multiple | Unknown/Multiple Difficulty: Unknown | Type of failure: Blocked By: | None/Unknown Related Tickets: | Test Case: | Blocking: | Differential Revisions: -------------------------------------+------------------------------------- The attached file should compile. The types T1, T2, and T3 are all equivalent types (T1 and T2 normalize to T3), but with foo::T1 and bar::T2 there is a type error. Rewriting the example using type families instead of FD makes it work. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9627> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 4

[GHC] #8949: switch -msse2 to be on by default
by GHC 21 Dec '18

21 Dec '18

#8949: switch -msse2 to be on by default ----------------------------------------------+---------------------------- Reporter: errge | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler (CodeGen) | Version: 7.9 Keywords: | Operating System: Architecture: x86 | Unknown/Multiple Difficulty: Moderate (less than a day) | Type of failure: Other Blocked By: | Test Case: Related Tickets: | Blocking: ----------------------------------------------+---------------------------- I propose msse2 to be on by default. I guess the default was chosen way back, when Pentium III support was still relevant. Nowadays we don't really win on the CPU support, because e.g. https://github.com/tibbe/hashable/blob/master/hashable.cabal is built by default with sse2 on the injected C code level. And hashable has a lot of reverse depends, therefore on an end user system (RedHat or Debian) the user is most probably unlucky with a Pentium III CPU anyways. Flipping this default would also fix the excess precision problem for most users of GHC on the i686 platform. GHC should provide a -mno-sse2 flag for the cases when this needs to be disabled. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8949> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 17

[GHC] #8655: Evaluate know-to-terminate-soon thunks
by GHC 19 Dec '18

19 Dec '18

#8655: Evaluate know-to-terminate-soon thunks ------------------------------------+------------------------------------- Reporter: nomeata | Owner: nomeata Type: task | 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: | ------------------------------------+------------------------------------- I guess I’ll better put my interior monologue in a ticket than on ghc- dev... In order to implement nested CPR (#1600), I had to enrich the CPR type to keep track of whether something, when entered, will converge for sure. My code does not solve the halting problem, but does only simple stuff, so if something is known to converge, is is very likely to be cheap. Simon suggested to measure the effect of evaluating a let-bound thunk with the known-to-terminate property, as if the body was using it strictly. This tickets tracks this suggestion, and reports progress. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8655> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 12

[GHC] #8598: IO hack in demand analyzer gets in the way of CPR
by GHC 19 Dec '18

19 Dec '18

#8598: IO hack in demand analyzer gets in the way of CPR ------------------------------------+------------------------------------- Reporter: nomeata | Owner: Type: task | 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: | ------------------------------------+------------------------------------- After a lot of staring at code and comparing unexpected nofib results I found the following: The IO hack in the demand analyzer (see `dmdAnalAlt` in `StrAnal.lhs` and #1592 for a good discussion) prevents CPR in any function that uses a C call. This is a small example, reduced from the `scaleFloat` method for doubles: {{{ module Float(fun) where import GHC.Float (Double(..)) import GHC.Integer (decodeDoubleInteger, encodeDoubleInteger) fun :: Double -> Double fun x | isFix = x | otherwise = case x of (D# x#) -> case decodeDoubleInteger x# of (# i, j #) -> D# (encodeDoubleInteger i j) where isFix = isDoubleFinite x == 0 foreign import ccall unsafe "isDoubleFinite" isDoubleFinite :: Double -> Int }}} Here, `fun` does current not get the CPR property, and the work gets type `GHC.Prim.Double# -> GHC.Types.Double`. Why? Because in core, there will be a {{{ case {__pkg_ccall main isDoubleFinite GHC.Prim.Double# -> GHC.Prim.State# GHC.Prim.RealWorld -> (# GHC.Prim.State# GHC.Prim.RealWorld, GHC.Prim.Int# #)}_dQD ds_dQA GHC.Prim.realWorld# of _ [Occ=Dead, Dmd=<L,A>] { (# ds_dQC [Dmd=<L,A>, OS=OneShot], ds_dQB [Dmd=<S,1*U>] #) -> ... }}} where the case body has `DmdType m {dQz-><L,1*U(U)> dQA-><L,U>}`, but `dmdAnalAlt` detects that this is evaluating a possibly exiting function and throws aways the information using `alt_ty `lubDmdType` topDmdType`. Would it be ok to `lub` only the demand on the free variables, but keep the CPR information? In nofib (if I managed to compare the right results) this does nothing for almost all benchmarks, `-9.2%` of allocations for `mandel` and `+4.9%` for reverse-complement (but these numbers are not obtained very cleanly, and all is happening on top of the better-ho-cardinality branch. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8598> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 15

[GHC] #8959: GHCi should honour UnicodeSyntax
by GHC 19 Dec '18

19 Dec '18

#8959: GHCi should honour UnicodeSyntax -------------------------------------------+------------------------------- Reporter: Fuuzetsu | Owner: Type: bug | Status: new Priority: low | Milestone: Component: Compiler | Version: 7.6.3 Keywords: | Operating System: Architecture: Unknown/Multiple | Unknown/Multiple Difficulty: Easy (less than 1 hour) | Type of failure: Blocked By: | None/Unknown Related Tickets: | Test Case: | Blocking: -------------------------------------------+------------------------------- When we load a module into GHCi with -XUnicodeSyntax, GHCi should default to printing information with the relevant unicode symbols. {{{ *Communication> let a ∷ Int; a = 7 *Communication> :t a a :: Int }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8959> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 27

[GHC] #9718: Avoid TidyPgm predicting what CorePrep will do
by GHC 10 Dec '18

10 Dec '18

#9718: Avoid TidyPgm predicting what CorePrep will do -------------------------------------+------------------------------------- Reporter: simonpj | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.8.3 Keywords: | Operating System: Architecture: Unknown/Multiple | Unknown/Multiple Difficulty: Unknown | Type of failure: Blocked By: | None/Unknown Related Tickets: | Test Case: | Blocking: | Differential Revisions: -------------------------------------+------------------------------------- At the moment the `TidyPgm` pass is forced to predict, accurately but unpleasantly, some aspects of what `CorePrep` and `Core2Stg` will do. Reason: * Each `Id` in the interface file records (among other things) the '''arity''' of the `Id`, and whether it has '''CAF references''' * We really only know these two things for sure after `CorePrep`. The conversion from Core to STG makes no structural changes. * However the result of `TidyPgm` (which preceded `CorePrep`) is used to generate the interface file. So it has to predict the arity and CAF-ref status of each `Id`. * This is not good. It restricts what `CorePrep` can do (notably, it must not change the arity of a top-level `Id`), and it leads to unsavoury code (e.g. look at the call to `CorePrep.cvtLitInteger` in `TidyPgm.cafRefsL`. It's also dangerous: an inconsistency could lead to a crash. This is a long-standing problem. My current thought for how to unravel it is this: * `TidyPgm` does not attach arity or CAF-ref info. * Instead, run `CorePrep` after `TidyPgm`, and generate accurate arity and CAF-ref info * Then use that auxiliary mapping during the conversion from tidied program to `ModIface`. I don't think this would be hard. It would mean that the tidied program and the core-prep'd program would have to exist in memory at the same time. An alternative would be to generate the `ModIface` from the tidied program ''sans'' arity and CAF-ref info, and then, after `CorePrep` run over it to add arity and CAF-ref info. (You'd have to do this before generating the fingerprints.) The advantage of this is that the `ModIface` can be a lot smaller than the code for the entire module. A long-standing wart which needs some careful attention. Simon -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/9718> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 42