January 2018 - ghc-tickets

[GHC] #8487: Debugger confuses variables
by GHC 02 Jan '18

02 Jan '18

#8487: Debugger confuses variables ------------------------------------+------------------------------------- Reporter: edsko | Owner: Type: bug | 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: Related Tickets: | ------------------------------------+------------------------------------- Consider the following example: {{{ import Control.Exception f = do ma <- try $ evaluate a case ma of Right str -> return a Left err -> return $ show (err :: SomeException) where a :: String a = error "hi" }}} Load into ghci, set a breakpoint on the 5th line, then run f: {{{ :break 5 f }}} You will get {{{ Stopped at Ex.hs:(5,3)-(7,53) _result :: IO String = _ a :: Either SomeException String = Left _ }}} This doesn't make sense. `a` has type `String`; it is `ma` that has type `Either SomeException String`. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/8487> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 2

[GHC] #14610: newtype wrapping of a monadic stack kills performance
by GHC 02 Jan '18

02 Jan '18

#14610: newtype wrapping of a monadic stack kills performance -------------------------------------+------------------------------------- Reporter: mrkkrp | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.2.2 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 have a project where I in one module (A) I decided to build something like a minimal framework for the other bigger module (B) to use. One of the main points of the framework is that the monadic stacks are hidden behind `newtype`s and in those monads you can only use the functions that the module A provides. The module A can be found here: https://github.com/mrkkrp/mmark/blob/master/Text/MMark/Parser/Internal.hs There are two monadic stack wrapped with newtypes: `BParser` and `IParser`. The module B is this: https://github.com/mrkkrp/mmark/blob/master/Text/MMark/Parser.hs But it's not really relevant. Now if I change `newtype`s to `type` synonyms like so: {{{ type BParser a = ParsecT MMarkErr Text (State BlockState) a type IParser a = StateT InlineState (Parsec MMarkErr Text) a }}} and do corresponding minor corrections, I get 2x less allocations and almost 2x faster code (before): {{{ Case Allocated GCs Max with file: data/bench-yaml-block.md 119,080 0 11,088 with file: data/bench-thematic-break.md 74,368 0 10,224 with file: data/bench-heading.md 901,928 0 9,432 with file: data/bench-fenced-code-block.md 145,744 0 9,368 with file: data/bench-indented-code-block.md 124,312 0 9,368 with file: data/bench-unordered-list.md 2,010,496 1 10,784 with file: data/bench-ordered-list.md 2,025,016 1 10,728 with file: data/bench-blockquote.md 1,961,672 1 42,648 with file: data/bench-paragraph.md 2,084,104 1 42,648 with file: data/comprehensive.md 25,899,496 24 44,200 benchmarking with file: data/comprehensive.md time 3.590 ms (3.578 ms .. 3.601 ms) 1.000 R² (1.000 R² .. 1.000 R²) mean 3.555 ms (3.546 ms .. 3.565 ms) std dev 31.07 μs (24.63 μs .. 39.90 μs) }}} After: {{{ Case Allocated GCs Max with file: data/bench-yaml-block.md 116,864 0 11,088 with file: data/bench-thematic-break.md 64,776 0 10,392 with file: data/bench-heading.md 615,672 0 9,432 with file: data/bench-fenced-code-block.md 144,736 0 9,368 with file: data/bench-indented-code-block.md 123,352 0 9,368 with file: data/bench-unordered-list.md 795,072 0 41,568 with file: data/bench-ordered-list.md 808,808 0 41,512 with file: data/bench-blockquote.md 826,392 0 41,568 with file: data/bench-paragraph.md 881,432 0 41,568 with file: data/comprehensive.md 10,945,104 10 44,440 benchmarking with file: data/comprehensive.md time 2.451 ms (2.448 ms .. 2.456 ms) 1.000 R² (1.000 R² .. 1.000 R²) mean 2.432 ms (2.427 ms .. 2.437 ms) std dev 15.86 μs (13.16 μs .. 19.01 μs) }}} I'm not great at reading non-trivial core, but I gave it a shot and dumped some core. One thing I noticed that core for the module A is the same in both cases. Then the problem is an inter-module problem, probably like when you don't dump definitions into interface files with `INLINEABLE` and end up without specialization, but here it perhaps has to do with the fact that B has no information about internals of the monadic stacks from A, but I'm not sure. Here is the beginnings of core dumps (before): {{{ ==================== Tidy Core ==================== 2017-12-23 04:55:17.967944505 UTC Result size of Tidy Core = {terms: 210,169, types: 209,675, coercions: 82,492, joins: 973/3,753} }}} After: {{{ ==================== Tidy Core ==================== 2017-12-23 04:58:46.386265108 UTC Result size of Tidy Core = {terms: 301,256, types: 263,104, coercions: 28,560, joins: 1,726/5,393} }}} So it looks like `newtype` wrapping reduces GHC's ability to create join points SPJ talked about recently. I do understand that you expect a minimal reproducing example but I could not produce it even though I spent several hours in futile attempts. I started by creating a small project, defining a similar stack with a `newtype` wrapper and using it in a simple parser in another module. Then I benchmarked the parser. There is no difference between `newtype`d and code and the code that uses just type synonyms. I tried different variations, no difference. The core output is too big for me to analyze, it's like 25 Mb and 33 Mb, and I have no idea what's going on there. You're welcome to experiment with the repo, there are benchmarks for memory usage and criterion benchmarks for speed of execution: https://github.com/mrkkrp/mmark If you like I can create a separate branch with newtypes replaced with type synonyms, so you can compare the both approaches easier. I'm submitting this because my friend convinced me that it's better to let you know (even though I could not create a minimal reproducing example on my own) than to let it go completely unnoticed. -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/14610> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 23

Re: [GHC] #3553: parallel gc suffers badly if one thread is descheduled
by GHC 02 Jan '18

02 Jan '18

#3553: parallel gc suffers badly if one thread is descheduled -------------------------------------+------------------------------------- Reporter: simonmar | Owner: simonmar Type: bug | Status: closed Priority: normal | Milestone: 6.12.2 Component: Runtime System | Version: 6.10.4 Resolution: fixed | 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 nh2): Replying to [comment:8 wuzzeb]: > There is a recent patch to the linux kernel to implement a futux which spins in the kernel, see http://thread.gmane.org/gmane.linux.kernel/970412 This link is dead now. I suspect it was "futex: FUTEX_LOCK with optional adaptive spinning" as described by LWN: https://lwn.net/Articles/387246/ There are now also two newer articles / patch sets about futexes: * https://lwn.net/Articles/704843/ (directly related to the above one) * https://lwn.net/Articles/685769/ -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/3553#comment:16> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 0

[GHC] #14625: Casts get in the way of calculating unfolding discount
by GHC 02 Jan '18

02 Jan '18

#14625: Casts get in the way of calculating unfolding discount -------------------------------------+------------------------------------- Reporter: mpickering | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.2.2 Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: None/Unknown Unknown/Multiple | Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): Phab:D4279 | Wiki Page: -------------------------------------+------------------------------------- The dictionary argument discount was not being applied for code such as rec_sel (dict_arg 'cast' cobox). For example, adding lots of type equality constraints increases the size of `foo` so that it will no longer be inlined into `foo1` (if specialisation is turned off). {{{ {-# LANGUAGE GADTs #-} module Small where foo :: (Num a, a ~ b, a ~ b, a ~ b, a ~ b, a ~ b) => a -> b foo x = x + x foo1 :: Int -> Int foo1 x = foo x }}} -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/14625> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 0

Re: [GHC] #5927: A type-level "implies" constraint on Constraints
by GHC 02 Jan '18

02 Jan '18

#5927: A type-level "implies" constraint on Constraints -------------------------------------+------------------------------------- Reporter: illissius | Owner: (none) Type: feature request | Status: new Priority: normal | Milestone: Component: Compiler (Type | Version: 7.4.1 checker) | Keywords: Resolution: | QuantifiedContexts 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 simonpj): * keywords: => QuantifiedContexts -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/5927#comment:27> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 0

[GHC] #14570: Untouchable error arises from type equality, but not equivalent program with fundeps
by GHC 02 Jan '18

02 Jan '18

#14570: Untouchable error arises from type equality, but not equivalent program with fundeps -------------------------------------+------------------------------------- Reporter: lexi.lambda | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: Component: Compiler | Version: 8.0.2 (Type checker) | Keywords: | Operating System: Unknown/Multiple Architecture: | Type of failure: GHC rejects Unknown/Multiple | valid program Test Case: | Blocked By: Blocking: | Related Tickets: Differential Rev(s): | Wiki Page: -------------------------------------+------------------------------------- Given some type definitions: {{{#!hs data A data B (f :: * -> *) data X (k :: *) }}} …and this typeclass: {{{#!hs class C k a | k -> a }}} …these (highly contrived for the purposes of a minimal example) function definitions typecheck: {{{#!hs f :: forall f. (forall k. (C k (B f)) => f k) -> A f _ = undefined g :: (forall k. (C k (B X)) => X k) -> A g = f }}} However, if we use a type family instead of a class with a functional dependency: {{{#!hs type family F (k :: *) }}} …then the equivalent function definitions fail to typecheck: {{{#!hs f :: forall f. (forall k. (F k ~ B f) => f k) -> A f _ = undefined g :: (forall k. (F k ~ B X) => X k) -> A g = f }}} {{{ • Couldn't match type ‘f0’ with ‘X’ ‘f0’ is untouchable inside the constraints: F k ~ B f0 bound by a type expected by the context: F k ~ B f0 => f0 k Expected type: f0 k Actual type: X k • In the expression: f In an equation for ‘g’: g = f }}} I read Section 5.2 of the OutsideIn(X) paper, which describes touchable and untouchable type variables, and I ''sort of'' understand what’s going on here. If I add an extra argument to `f` that pushes the choice of `f` outside the inner `forall`, then the program typechecks: {{{#!hs f :: forall f a. f a -> (forall k. (F k ~ B f) => f k) -> A f _ _ = undefined g :: forall a. X a -> (forall k. (F k ~ B X) => X k) -> A g = f }}} I don’t know if this is actually a bug—it seems entirely reasonable to me that I don’t fully understand what is going on here—but I’m stumped as to ''why'' GHC rejects this program but accepts the one involving functional dependencies. I would expect it to either accept or reject both programs, given they ''seem'' equivalent. Is this just an unfortunate infelicity of the differences between how functional dependencies and type equalities are internally solved? Or is there a deeper difference between these two programs? ,,(Note: This ticket is adapted from [https://stackoverflow.com/q/47734825/465378 this Stack Overflow question].),, -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/14570> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 4

[GHC] #11113: Type family If is too strict
by GHC 01 Jan '18

01 Jan '18

#11113: Type family If is too strict -------------------------------------+------------------------------------- Reporter: olshanskydr | Owner: 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 Rev(s): | Wiki Page: -------------------------------------+------------------------------------- This code {{{#!hs type family Div (k::Nat) (n::Nat) where Div k n = If (CmpNat k n == LT) 0 (1 + Div (k-n) n) }}} is not working. When I try {{{ ghci> :t (Proxy :: Proxy (Div 100 9)) }}} it hangs on. Probably ghci is trying to calculate both '''If''' branches. If I rewrite it this way {{{#!hs type family Div (k::Nat) (n::Nat) where Div k n = Div' k n (CmpNat k n) type family Div' (k::Nat) (n::Nat) (b::Ordering) where Div' k n LT = 0 Div' k n EQ = 1 Div' k n GT = 1 + Div (k-n) n }}} it works well -- Ticket URL: <http://ghc.haskell.org/trac/ghc/ticket/11113> GHC <http://www.haskell.org/ghc/> The Glasgow Haskell Compiler

1 3