- ghc-tickets - Haskell.org

[GHC] #10680: Make Backpack order-independent (again)
by GHC 05 Jan '17

05 Jan '17

#10680: Make Backpack order-independent (again) -------------------------------------+------------------------------------- Reporter: ezyang | Owner: ezyang Type: task | Status: new Priority: normal | Milestone: Component: Package | Version: 7.11 system | Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: Differential Revisions: | -------------------------------------+------------------------------------- When we moved to the new `bkp` file format, we also went back to the a format which is order-dependent: that is to say, the order in which you put the declarations matters. So if you write: {{{ unit p where module A where import B module B where ... }}} this fails to type-check, GHC complaining that `B` is not in scope. I did this, in part because it's what the Backpack paper described, and because it was "simpler" to implement. I think we should move back to an order-independent scheme, for the following reasons: 1. Haskell users are used to not needing pay particularly close attention to the ordering of their modules, and forcing people to linearize their module descriptions would be spectacularly disruptive with large amounts of modules. So un-ordered modules are "more natural. 2. Order-independence imposes some constraints on how expressive programs are (with order-dependent Backpack, you can do some pretty tricky things by ordering things certain ways); this could simplify some aspects of compiler implementation and make Backpack easier to explain. 3. A particular case of (2): it seems a lot simpler UX-wise to let a user assume that if you import a module `M` in a unit, it doesn't matter where you import it: you always get the same set of identifiers brought into scope. Thus, the incremental results of signatures should not be visible, c.f. #10679 The main idea is that only the surface-syntax is un-ordered: the internal representation of units is a DAG which we work out in an elaboration phase, not altogether unsimilar from what `GhcMake` computes. Here are the details: **The intermediate representation.** We translate into an intermediate representation which consists of a directed graph between modules, signatures and includes. Edges in the graph indicate a "depends on" relation: 1. `include p` depends on `include q` if, for some module name `H`, `p` requires `H` and `q` provides `H`. 2. A module/signature `M` depends on `include p` if `M` imports a module provided or required by `p`. 3. A module/signature `M` depends on a module/signature `S` if `M` imports `S`. 4. An `include p` depends on a module `M` if `p` requires a module named `M`. (This rule is included for completeness; we are going to disallow it shortly.) We impose one restriction: a signature cannot depend on a home module. See below for how to eliminate this restriction. Rule (2) is worth remarking upon: if a module imports a signature, it depends-on every `include` which requires that signature, as well as the relevant home signature. This could easily result in a cycle; see (2) for how to break these cycles. The consequence of this, however, is that we can factor the graph to introduce the node for the "merge of signatures", which depends on each signature and include which requires it; we can use this opportunity to build and write out the merged interface file for the unit. **Elaboration.** Take a Backpack file, construct this graph, and topsort it into a DAG of SCCs. SCCs with a single node are compileable as before. SCCs with multiple nodes will have to be managed with some mutual recursion mechanism; see refinements for more thoughts on this. **Refinements:** 1. **Can a signature depend on a (home) module?** Imports of this kind require a retypecheck loop. Consider this situation: {{{ unit p where signature H where data T module M where import H data S = S T unit q where include p module Q where import M signature H where import Q data T = T S }}} Here, signature H in q depends on Q. When we typecheck `Q`, we bring `M.S` into the type environment with a `TyThing` that describes the constructor as accepting an abstract type `T`. However, when we subsequently typecheck the local signature `H`, we must refine all `TyThing`s of `T` with the true description (e.g. constructor information). So you'll need to retypecheck `Q` (and `M`) in order to make sure the `TyThing` is correct. 2. **Can an include depend on a (home) module?** If the module has no (transitive) dependency on signatures, this is fine. However, it's easy to have a circular dependency. Consider: {{{ unit p where signature A signature B module M unit q where include p module B where import A ... }}} `B` depends on `p` for `p/A.hsig`; however, `p` depends on `B` because this module is filling a requirement. Fortunately, we can untangle this knot without any mutual recursion nonsense (and the attendant efficiency loss): `A` is just an export list, we can compute it from the abstractly type-checked version of `p` without instantiating `B`. 3. **Can we deal with include-include dependency cycles?** Yes! Just use the Backpack paper's strategy for creating a recursive unit key and compile the two packages `hs-boot` style. But I'm not planning on implementing this yet. 4. **Can we deal with signature-signature dependency cycles?** Ordered Backpack would have supported this: {{{ unit a-sig where signature A where data T unit ab-sig where include a-sig signature B where import A data S = S T signature A where import B data T = T S }}} In our model, `ab-sig` has a cycle. However, I believe any such cycle can be broken by creating sufficiently many units: {{{ unit a-sig where signature B where data T signature A where data S = S T unit b-sig where signature A where data S signature B where data T = T S unit ab-sig where include a-sig include b-sig }}} In principle, GHC could automatically break import cycles by replacing an import with an import of a reduced signature that simply has abstract type definitions. (I'm not sure this is possible for all language features.) This technique would also work for normal modules, assuming that every function is explicitly annotated with a type. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10680> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 4

[GHC] #12189: sigof01m broken after D2030
by GHC 05 Jan '17

05 Jan '17

#12189: sigof01m broken after D2030 -------------------------------------+------------------------------------- Reporter: niteria | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.1 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: sigof01m | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- With [https://phabricator.haskell.org/D2030 D2030] I've changed the implementation of `Ord UnitId` to go from being `Unique` defined to being lexicographic. That somehow broke `sigof01m`. {{{ <ezyang> I think I know what the problem is (ghc --make isn't specifying a dependency from B to A) }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12189> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 2

[GHC] #10838: hsig files don't have good enough error checking for wired-in types
by GHC 05 Jan '17

05 Jan '17

#10838: hsig files don't have good enough error checking for wired-in types -------------------------------------+------------------------------------- Reporter: ezyang | Owner: ezyang Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 7.11 (Type checker) | Keywords: backpack | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: Differential Revisions: | -------------------------------------+------------------------------------- If I have the signature: {{{ {-# LANGUAGE NoImplicitPrelude #-} module MyPrelude where data Int a }}} And I compile it with `ghc MyPrelude.hsig -sig-of "MyPrelude is base:Prelude"` this won't actually give an error, despite `Int` having the wrong kind. The bug seems to be because `Int` is wired in and this case is not handled. Filing this so I don't forget to fix. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/10838> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 1

[GHC] #12997: Build broken on OpenBSD
by GHC 05 Jan '17

05 Jan '17

#12997: Build broken on OpenBSD ----------------------------------------+--------------------------------- Reporter: bgamari | Owner: Type: bug | Status: new Priority: normal | Milestone: 8.0.2 Component: Compiler | Version: 8.0.2-rc2 Keywords: | Operating System: OpenBSD Architecture: Unknown/Multiple | Type of failure: None/Unknown Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: ----------------------------------------+--------------------------------- Karel says, {{{ "/usr/local/bin/ghc" -o utils/hsc2hs/dist/build/tmp/hsc2hs -hisuf hi -osuf o -hcsuf hc -static -H32m -O -Wall -package-db libraries/bootstrapping.conf -hide-all-packages -i -iutils/hsc2hs/. -iutils/hsc2hs/dist/build -iutils/hsc2hs/dist/build/autogen -Iutils/hsc2hs/dist/build -Iutils/hsc2hs/dist/build/autogen -optP-include -optPutils/hsc2hs/dist/build/autogen/cabal_macros.h -package-id base-4.8.2.0-8dd7859ce039df9f3d3f4ff9dbf4bb93 -package-id containers-0.5.6.2-59326c33e30ec8f6afd574cbac625bbb -package-id directory-1.2.2.0-de5c44596f448d2b2988d8588c5afdef -package-id filepath-1.4.0.0-f97d1e4aebfd7a03be6980454fe31d6e -package-id process-1.2.3.0-ca3fff48cdf67a08d7e3f92817538395 -XHaskell2010 -no-user-package-db -rtsopts -odir utils/hsc2hs/dist/build -hidir utils/hsc2hs/dist/build -stubdir utils/hsc2hs/dist/build -optl-z -optlwxneeded -static -H32m -O -Wall -package-db libraries/bootstrapping.conf -hide-all-packages -i -iutils/hsc2hs/. -iutils/hsc2hs/dist/build -iutils/hsc2hs/dist/build/autogen -Iutils/hsc2hs/dist/build -Iutils/hsc2hs/dist/build/autogen -optP-include -optPutils/hsc2hs/dist/build/autogen/cabal_macros.h -package-id base-4.8.2.0-8dd7859ce039df9f3d3f4ff9dbf4bb93 -package-id containers-0.5.6.2-59326c33e30ec8f6afd574cbac625bbb -package-id directory-1.2.2.0-de5c44596f448d2b2988d8588c5afdef -package-id filepath-1.4.0.0-f97d1e4aebfd7a03be6980454fe31d6e -package-id process-1.2.3.0-ca3fff48cdf67a08d7e3f92817538395 -XHaskell2010 -no-user-package-db -rtsopts utils/hsc2hs/dist/build/Main.o utils/hsc2hs/dist/build/C.o utils/hsc2hs/dist/build/Common.o utils/hsc2hs/dist/build/CrossCodegen.o utils/hsc2hs/dist/build/DirectCodegen.o utils/hsc2hs/dist/build/Flags.o utils/hsc2hs/dist/build/HSCParser.o utils/hsc2hs/dist/build/UtilsCodegen.o utils/hsc2hs/dist/build/Paths_hsc2hs.o <no location info>: Warning: Couldn't figure out linker information! Make sure you're using GNU ld, GNU gold or the built in OS X linker, etc. gcc: wxneeded: No such file or directory compiler/ghc.mk:584: compiler/stage1/build/.depend-v.haskell: No such file or directory gmake[1]: *** [utils/hsc2hs/ghc.mk:15: utils/hsc2hs/dist/build/tmp/hsc2hs] Error 1 gmake: *** [Makefile:132: all] Error 2 }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/12997> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 7

[GHC] #11564: Possible overzealous unfolding
by GHC 05 Jan '17

05 Jan '17

#11564: Possible overzealous unfolding -------------------------------------+------------------------------------- Reporter: simonmar | Owner: simonpj Type: bug | Status: new Priority: high | Milestone: Component: Compiler | Version: 8.1 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: -------------------------------------+------------------------------------- I was investigating why (>>=) in the Haxl monad is being inlined more than I would expect, and I ran into something I don't fully understand, and looks dubious. Start from this standalone example: {{{ {-# LANGUAGE ScopedTypeVariables, ExistentialQuantification #-} module Haxl where import Data.IORef import Control.Exception newtype GenHaxl u a = GenHaxl { unHaxl :: Int -> IORef () -> IO (Result u a) } data Result u a = Done a | Throw SomeException | Blocked (Cont u a) data Cont u a = forall b. Cont u b :>>= (b -> GenHaxl u a) | forall b. (b -> a) :<$> (Cont u b) instance Monad (GenHaxl u) where return a = GenHaxl $ \_env _ref -> return (Done a) GenHaxl m >>= k = GenHaxl $ \env ref -> do e <- m env ref case e of Done a -> unHaxl (k a) env ref Throw e -> return (Throw e) Blocked cont -> return (Blocked (cont :>>= k)) instance Functor (GenHaxl u) instance Applicative (GenHaxl u) }}} (it could be simplified further, but I've intentionally used the exact definition of `>>=` that is used in Haxl to be sure I'm not investigating the wrong thing) Compile like this: {{{ ghc -O -c Haxl.hs }}} and look at the .hi file: {{{ ghc --show-iface Haxl.hi }}} see this: {{{ ea159c3b107c307a4e76cd310efcc8bc $fMonadGenHaxl2 :: GenHaxl u a -> (a -> GenHaxl u b) -> Int -> IORef () -> State# RealWorld -> (# State# RealWorld, Result u b #) {- Arity: 5, HasNoCafRefs, Strictness: <C(C(C(S(SS)))),1*C1(C1(C1(U(U,1*U))))><L,U><L,U><L,U><S,U>, Unfolding: InlineRule (5, True, False) (\ @ u @ a @ b (ds :: GenHaxl u a) (k :: a -> GenHaxl u b) (env :: Int) (ref :: IORef ()) (s :: State# RealWorld)[OneShot] -> case (ds `cast` (N:GenHaxl[0] <u>_P <a>_R) env ref) `cast` (N:IO[0] <Result u a>_R) s of ds1 { (#,#) ipv ipv1 -> case ipv1 of wild { Done a1 -> ((k a1) `cast` (N:GenHaxl[0] <u>_P <b>_R) env ref) `cast` (N:IO[0] <Result u b>_R) ipv Throw e -> (# ipv, Throw @ u @ b e #) Blocked cont -> (# ipv, Blocked @ u @ b (:>>= @ u @ b @ a cont k) #) } }) -} }}} That right there is the definition of `>>=`. Note that it has an `InlineRule`, which means that it will be unconditionally unfolded pretty much everywhere. I don't think this is right - there's no benefit to be had in inlining it unconditionally. I delved in a bit more, and it seems this unfolding arises during worker- wrapper. Before WW we have {{{ a_sVM [LclId, Arity=5, Str=DmdType <C(C(C(S(SS)))),1*C1(C1(C1(U(U,1*U))))><L,U><L,U><L,U><S,U>, Unf=Unf{Src=<vanilla>, TopLvl=True, Value=True, ConLike=True, WorkFree=True, Expandable=True, Guidance=IF_ARGS [0 60 0 0 0] 94 60}] a_sVM = \ @ u_XQR @ a_aPN @ b_aPO ds_dQP [Dmd=<C(C(C(S(SS)))),1*C1(C1(C1(U(U,1*U))))>] k_aEC env_aED ref_aEE s_aVC [Dmd=<S,U>, OS=OneShot] -> case (((ds_dQP `cast` ...) env_aED ref_aEE) `cast` ...) s_aVC of _ [Occ=Dead, Dmd=<L,A>] { (# ipv_aVF [Dmd=<S,U>], ipv1_aVG [Dmd=<S,1*U>] #) -> case ipv1_aVG of _ [Occ=Dead, Dmd=<L,A>] { Done a_aEG -> ((((k_aEC a_aEG) `cast` ...) env_aED ref_aEE) `cast` ...) ipv_aVF; Throw e_aEH -> (# ipv_aVF, Haxl.Throw e_aEH #); Blocked cont_aEI -> (# ipv_aVF, Haxl.Blocked (Haxl.:>>= cont_aEI k_aEC) #) } } }}} and after WW we have {{{ a_sVM [LclId, Arity=5, Str=DmdType <C(C(C(S(SS)))),1*C1(C1(C1(U(U,1*U))))><L,U><L,U><L,U><S,U>, Unf=Unf{Src=InlineStable, TopLvl=True, Value=True, ConLike=True, WorkFree=True, Expandable=True, Guidance=ALWAYS_IF(arity=5,unsat_ok=True,boring_ok=False) Tmpl= \ @ u_XQR @ a_aPN @ b_aPO ds_dQP [Occ=Once] k_aEC [Occ=Once*] env_aED ref_aEE s_aVC [Occ=Once, OS=OneShot] -> case (((ds_dQP `cast` ...) env_aED ref_aEE) `cast` ...) s_aVC of _ [Occ=Dead] { (# ipv_aVF [Occ=Once*], ipv1_aVG [Occ=Once!] #) -> case ipv1_aVG of _ [Occ=Dead] { Done a_aEG [Occ=Once] -> ((((k_aEC a_aEG) `cast` ...) env_aED ref_aEE) `cast` ...) ipv_aVF; Throw e_aEH [Occ=Once] -> (# ipv_aVF, Haxl.Throw e_aEH #); Blocked cont_aEI [Occ=Once] -> (# ipv_aVF, Haxl.Blocked (Haxl.:>>= cont_aEI k_aEC) #) } }}] a_sVM = \ @ u_XQR @ a_aPN @ b_aPO ds_dQP [Dmd=<C(C(C(S(SS)))),1*C1(C1(C1(U(U,1*U))))>] k_aEC env_aED ref_aEE s_aVC [Dmd=<S,U>, OS=OneShot] -> case (((ds_dQP `cast` ...) env_aED ref_aEE) `cast` ...) s_aVC of _ [Occ=Dead, Dmd=<L,A>] { (# ipv_aVF [Dmd=<S,U>], ipv1_aVG [Dmd=<S,1*U>] #) -> case ipv1_aVG of _ [Occ=Dead, Dmd=<L,A>] { Done a_aEG -> ((((k_aEC a_aEG) `cast` ...) env_aED ref_aEE) `cast` ...) ipv_aVF; Throw e_aEH -> (# ipv_aVF, Haxl.Throw e_aEH #); Blocked cont_aEI -> (# ipv_aVF, Haxl.Blocked (Haxl.:>>= cont_aEI k_aEC) #) } } }}} For some unknown reason, this binding has acquired an always-on unfolding. There's no wrapper, we're just unfolding the whole thing. Simon, can you shed any light here? I would like to tune unfolding sizes to reduce code bloat in our codebase, but with this unfolding being unconditional it doesn't work to use `-funfolding-use-threshold`, I can only use `NOINLINE` but that's too blunt. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/11564> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 15

Re: [GHC] #6166: Performance regression in mwc-random since 7.0.x
by GHC 05 Jan '17

05 Jan '17

#6166: Performance regression in mwc-random since 7.0.x -------------------------------------+------------------------------------- Reporter: bos | Owner: Type: bug | Status: closed Priority: high | Milestone: 8.2.1 Component: Compiler | Version: 7.4.2 Resolution: fixed | Keywords: Operating System: Unknown/Multiple | Architecture: x86_64 Type of failure: Runtime | (amd64) performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by simonpj): There's a little-known file `nofib/Simon-nofib-notes` that contains per- benchmark commentary. Could you add the info above to it? The notes concentrate on aspects of nofib that have proved fragile, and are supposed to stop you reinventing the wheel when you are investigating a nofib perf blip. Thanks Simon -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/6166#comment:22> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 0

Re: [GHC] #6166: Performance regression in mwc-random since 7.0.x
by GHC 05 Jan '17

05 Jan '17

#6166: Performance regression in mwc-random since 7.0.x -------------------------------------+------------------------------------- Reporter: bos | Owner: Type: bug | Status: closed Priority: high | Milestone: 8.2.1 Component: Compiler | Version: 7.4.2 Resolution: fixed | Keywords: Operating System: Unknown/Multiple | Architecture: x86_64 Type of failure: Runtime | (amd64) performance bug | Test Case: Blocked By: | Blocking: Related Tickets: | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Changes (by rwbarton): * status: new => closed * resolution: => fixed Comment: The original version of mwc-random and the minimized test cases in the comments are a bit fragile, in that they rely crucially on `blocks` being floated out of the IO action which contains it, ideally to top level. I'm not sure why, but GHC 7.8.4 isn't doing this floating out in the versions marked SLOW. For better or worse, GHC 7.10.1 does float out `blocks` to top level in the SLOW versions, and I also checked that it produces efficient code for the original version mwc-random-0.13.1.0 that prompted this report. Without a way to reproduce this in a recent version, I'm going to assert that the underlying issue was probably fixed between 7.8 and 7.10. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/6166#comment:21> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 0

[GHC] #13055: Adding INLINABLE pragma causes GHC panic
by GHC 04 Jan '17

04 Jan '17

#13055: Adding INLINABLE pragma causes GHC panic -------------------------------------+------------------------------------- Reporter: mpickering | Owner: Type: bug | Status: new Priority: high | 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: -------------------------------------+------------------------------------- If you add an `INLINABLE` pragma to `<**>` in `GHC.Base` then `stage2` fails to build with an `idInfo` panic. I don't really know where to begin debugging this but it seems quite bad! Here is a build where it fails: https://phabricator.haskell.org/harbormaster/build/17389/ Here is a branch with one commit with the change: https://github.com/ghc/ghc/tree/wip/inlinable-bug -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/13055> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 7

[GHC] #13053: Inferred type for hole is not general enough
by GHC 04 Jan '17

04 Jan '17

#13053: Inferred type for hole is not general enough -------------------------------------+------------------------------------- Reporter: osa1 | Owner: Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.1 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: -------------------------------------+------------------------------------- As noted by @deech in #13050, inferred types for holes are sometimes not general enough. Example: {{{ Prelude> :set -XNoMonomorphismRestriction -- added this just to make sure Prelude> :t _ 1 2 <interactive>:1:1: error: • Found hole: _ :: Integer -> Integer -> t Where: ‘t’ is a rigid type variable bound by the inferred type of it :: t at <interactive>:1:1 • In the expression: _ In the expression: _ 1 2 }}} A more general type would be `(Num a, Num b) => a -> b -> t`. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/13053> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 1

Re: [GHC] #605: Optimisation: strict enumerations
by GHC 04 Jan '17

04 Jan '17

#605: Optimisation: strict enumerations -------------------------------------+------------------------------------- Reporter: simonmar | Owner: Type: task | Status: new Priority: normal | Milestone: 8.2.1 Component: Compiler | Version: None Resolution: None | Keywords: Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: N/A Blocked By: | Blocking: Related Tickets: #6135 | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Changes (by dfeuer): * cc: dfeuer (added) * milestone: ⊥ => 8.2.1 Comment: Setting a milestone because osa1 is working actively in this area. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/605#comment:22> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 0