- ghc-devs - Haskell.org

Testsuite cleaning
by Simon Peyton Jones 24 Jun '16

24 Jun '16

Thomas During debugging I often compile a single test program ghc -c T1969.hs But the new testsuite setup doesn’t remove .hi and .o files before running a test, so make TEST=T1969 says bytes allocated value is too low: … Deviation T1969(normal) bytes allocated: -95.2 % Reason? Compilation was not required! Non-perf tests fail in the same way +compilation IS NOT required *** unexpected failure for T11480b(normal) I’m sure this didn’t use to happen. It’s not fatal, because can manually remove those .o files, but it’s a bit of a nuisance. Might it be easy to restore the old behaviour? Thanks Simon

3 3

Does anyone know any easy-to-run compile-time benchmark suites?
by Ömer Sinan Ağacan 23 Jun '16

23 Jun '16

Hi all, I was wondering if anyone has or knows easy-to-run compile-time benchmarks? I'm looking for something like nofib -- ideally after a fresh build I should be able to just run `make` and get some numbers (mainly allocations) back.

3 2

isInvisible
by Simon Peyton Jones 23 Jun '16

23 Jun '16

Richard I have just spent an hour plumbing the torrid swamp of binder visibility. There is bad naming confusion. We have · Visible · Specified · Invisible The function isVisible returns True for Visible and False otherwise. But isInvisible returns True for Invisible and Specified. And partitionInvisibles, filterOutInvisibleTyVars, filterOutInvisibleTypes etc all treat Invisible and Specified the same. That is, it’s really filterOutInvisibleOrSpecifiedTypes which is terribly clumsy. Sometimes we need to pick up just the Inivisble args, and we don’t have a predicate for that (since isInvisible is taken), so you’ll see in Inst.top_instantiate we are reduced to saying (binderVisibility bndr == Invisible). It’s all very smelly. The trouble is that · In displayed types like (T a b), we display args that are Visible, but not Specified or Invisible · In visible type application like (f @t1 @t2) we can specify type args that are Visible or Specified but not Invisible. Urgh. The trouble is that “invisible” is not the negation of “visible”, which leads to all kinds of confusion. I wonder about changing to · Explicit · Specified · Implicit Then for display we display Explicit args only; for VTA we can specify Explicit or Specified args. So “not Explicit” doesn’t sound structurally like “Implicit” in the way that “not Visible” does sound like “Invisible”. Simon

2 4

Require -fexternal-interpreter support for future TH changes?
by Simon Marlow 23 Jun '16

23 Jun '16

*Background* A few months ago I added -fexternal-interpreter to GHC: - docs: http://downloads.haskell.org/~ghc/latest/docs/html/users_guide/ghci.html#gh… - wiki, rationale: https://ghc.haskell.org/trac/ghc/wiki/RemoteGHCi When -fexternal-interpreter is used, GHC runs interpreted code in a separate subprocess, and communicates with it using binary messages over a pipe. -fexternal-interpreter currently implements all of TH, quasi-quoting, annotations, and all the GHCi features except for some features of the debugger. It is also now implemented on Windows, thanks to Tamar Christina. *Proposal* I'd like to propose that going forward we commit to maintaining full support for -fexternal-interpreter, with a view to making it the default. Why? - -fexternal-interpreter will be a prerequisite for GHCJS support, so maintaining full support for TH in -fexternal-interpreter will ensure that everything that works with GHC works with GHCJS. - We will be able to make simplifications in GHC and the build system once -fexternal-interpreter is the default, because when compiling with -prof or -dynamic we won't have to compile things twice any more. - Ultimately we don't want to have two ways of doing everything, because that's harder to maintain. How? - I'll make all the TH and quasi-quoting tests run with and without -fexternal-interpreter, so it will break validate if one of these fails. *Why now?* There are some TH changes in the pipeline that will need special attention to work with -fexternal-interpreter. e.g. https://phabricator.haskell.org/D2286 and https://ghc.haskell.org/trac/ghc/wiki/TemplateHaskell/Introspective, so I'd like to raise it now so we can keep the issue in mind. Cheers Simon

4 11

How bad would it be for containers to depend on transformers?
by David Feuer 22 Jun '16

22 Jun '16

Currently, containers does not depend on transformers, so it has to duplicate its functionality or just do without. Since transformers is also a GHC boot package, I believe it should be feasible to make containers depend on it. To what extent would that reduce parallelizability of GHC builds or otherwise make people mad? David

2 1

Inferring instance constraints with DeriveAnyClass
by Ryan Scott 21 Jun '16

21 Jun '16

Andres, I'm trying to think of a proper solution to Trac #12144 [1]. This bug triggers when you try to use DeriveAnyClass in a somewhat exotic fashion: {-# LANGUAGE DeriveAnyClass, KindSignatures #-} class C (a :: * -> *) data T a = MkT (a -> Int) deriving C This currently gives a GHC panic: ghc: panic! (the 'impossible' happened) (GHC version 8.0.1 for x86_64-unknown-linux): contravariant This baffled me until I realized why it's happening: for typeclasses of kind * -> *, DeriveAnyClass simply re-uses the same algorithm that DeriveFunctor uses for coming up with an instance context. For instance, if you have: data T f a = MkT a (f a) deriving (Functor, C) Then it would generate two instances of the form: instance Functor f => Functor (T f) where ... instance C f => C (T f) where ... But #12144 reveals the fatal downside of doing this: DeriveFunctor has special knowledge about type parameters in contravariant positions, but this doesn't even make sense to think about with a class like C! (The only reason GHC won't panic if a Functor instance is derived for T is because there are Functor-specific checks that cause an error message to pop up before the panic can be reached.) My question is then: why does DeriveAnyClass take the bizarre approach of co-opting the DeriveFunctor algorithm? Andres, you originally proposed this in #7346 [2], but I don't quite understand why you wanted to do it this way. Couldn't we infer the context simply from the contexts of the default method type signatures? This is a question that Reid Barton has also asked, to which José Pedro Magalhães answered in the negatory [3]. But Pedro's reasoning has never quite made sense to me, because we've been able to typecheck constraints arising from default method type signatures for a long time, so why would it be impractical to do so in this case? I'd appreciate hearing a more detailed explanation on this issue, because at the moment, I am completely stuck on figuring out how one might fix #12144. Regards, Ryan S. ----- [1] https://ghc.haskell.org/trac/ghc/ticket/12144 [2] https://ghc.haskell.org/trac/ghc/ticket/7346 [3] https://ghc.haskell.org/trac/ghc/ticket/5462#comment:30

4 6

Simon's email classified as spam
by Simon Peyton Jones 19 Jun '16

19 Jun '16

Dear GHC devs/users This is another test to see if email from me, relayed via Haskell.org, ends up in your spam folder. Gershom thinks he’s fixed it (below). Can I trespass on your patience once more? Just let me know if this email ends up in your inbox or spam. Can you cc John and Gershom (but perhaps not everyone else)? Thanks Simon | From: Gershom B [mailto:gershomb@gmail.com] | Sent: 18 June 2016 18:53 | To: Simon Peyton Jones <simonpj(a)microsoft.com>; John Wiegley | <johnw(a)newartisans.com> | Cc: Michael Burge <michaelburge(a)pobox.com> | Subject: Re: FW: CMM-to-SAM: Register allocation weirdness | | Simon — I just found two possible sources of the problem (first: the top | level config didn’t take hold due to other errors when updating — fixed that, | and second, it might be possible the top level config isn’t retroactively | applied to all lists — so i added the config to the relevant lists directly). | | I think if you try one more time it might work (fingers crossed).

2 1

CMM-to-ASM: Register allocation wierdness
by Carter Schonwald 19 Jun '16

19 Jun '16

Agreed. There's also some other mismatches between ghc and llvm in a few fun / interesting ways! There's a lot of room for improvement in both code gens, but there's also a lot of room to improve the ease of experimenting with improvements. Eg we don't have a peephole pass per target, so those get hacked into the pretty printing code last time I checked On Thursday, June 16, 2016, Ben Gamari <ben(a)smart-cactus.org <javascript:_e(%7B%7D,'cvml','ben(a)smart-cactus.org');>> wrote: > > Ccing David Spitzenberg, who has thought about proc-point splitting, which > is relevant for reasons that we will see below. > > > Harendra Kumar <harendra.kumar(a)gmail.com> writes: > > > On 16 June 2016 at 13:59, Ben Gamari <ben(a)smart-cactus.org> wrote: > >> > >> It actually came to my attention while researching this that the > >> -fregs-graph flag is currently silently ignored [2]. Unfortunately this > >> means you'll need to build a new compiler if you want to try using it. > > > > Yes I did try -fregs-graph and -fregs-iterative both. To debug why > nothing > > changed I had to compare the executables produced with and without the > > flags and found them identical. A note in the manual could have saved me > > some time since that's the first place to go for help. I was wondering > if I > > am making a mistake in the build and if it is not being rebuilt > > properly. Your note confirms my observation, it indeed does not change > > anything. > > > Indeed; I've opened D2335 [1] to reenable -fregs-graph and add an > appropriate note to the users guide. > > >> All-in-all, the graph coloring allocator is in great need of some love; > >> Harendra, perhaps you'd like to have a try at dusting it off and perhaps > >> look into why it regresses in compiler performance? It would be great if > >> we could use it by default. > > > > Yes, I can try that. In fact I was going in that direction and then > stopped > > to look at what llvm does. llvm gave me impressive results in some cases > > though not so great in others. I compared the code generated by llvm and > it > > perhaps did a better job in theory (used fewer instructions) but due to > > more spilling the end result was pretty similar. > > > For the record, I have also struggled with register spilling issues in > the past. See, for instance, #10012, which describes a behavior which > arises from the C-- sinking pass's unwillingness to duplicate code > across branches. While in general it's good to avoid the code bloat that > this duplication implies, in the case shown in that ticket duplicating > the computation would be significantly less code than the bloat from > spilling the needed results. > > > But I found a few interesting optimizations that llvm did. For example, > > there was a heap adjustment and check in the looping path which was > > redundant and was readjusted in the loop itself without use. LLVM either > > removed the redundant _adjustments_ in the loop or moved them out of the > > loop. But it did not remove the corresponding heap _checks_. That makes > me > > wonder if the redundant heap checks can also be moved or removed. If we > can > > do some sort of loop analysis at the CMM level itself and avoid or remove > > the redundant heap adjustments as well as checks or at least float them > out > > of the cycle wherever possible. That sort of optimization can make a > > significant difference to my case at least. Since we are explicitly aware > > of the heap at the CMM level there may be an opportunity to do better > than > > llvm if we optimize the generated CMM or the generation of CMM itself. > > > Very interesting, thanks for writing this down! Indeed if these checks > really are redundant then we should try to avoid them. Do you have any > code you could share that demosntrates this? > > It would be great to open Trac tickets to track some of the optimization > opportunities that you noted we may be missing. Trac tickets are far > easier to track over longer durations than mailing list conversations, > which tend to get lost in the noise after a few weeks pass. > > > A thought that came to my mind was whether we should focus on getting > > better code out of the llvm backend or the native code generator. LLVM > > seems pretty good at the specialized task of code generation and low > level > > optimization, it is well funded, widely used and has a big community > > support. That allows us to leverage that huge effort and take advantage > of > > the new developments. Does it make sense to outsource the code generation > > and low level optimization tasks to llvm and ghc focussing on higher > level > > optimizations which are harder to do at the llvm level? What are the > > downsides of using llvm exclusively in future? > > > > There is indeed a question of where we wish to focus our optimization > efforts. However, I think using LLVM exclusively would be a mistake. > LLVM is a rather large dependency that has in the past been rather > difficult to track (this is why we now only target one LLVM release in a > given GHC release). Moreover, it's significantly slower than our > existing native code generator. There are a number of reasons for this, > some of which are fixable. For instance, we currently make no effort to > tell > LLVM which passes are worth running and which we've handled; this is > something which should be fixed but will require a rather significant > investment by someone to determine how GHC's and LLVM's passes overlap, > how they interact, and generally which are helpful (see GHC #11295). > > Furthermore, there are a few annoying impedance mismatches between Cmm > and LLVM's representation. This can be seen in our treatment of proc > points: when we need to take the address of a block within a function > LLVM requires that we break the block into a separate procedure, hiding > many potential optimizations from the optimizer. This was discussed > further on this list earlier this year [2]. It would be great to > eliminate proc-point splitting but doing so will almost certainly > require cooperation from LLVM. > > Cheers, > > - Ben > > > [1] https://phabricator.haskell.org/D2335 > [2] https://mail.haskell.org/pipermail/ghc-devs/2015-November/010535.html >

1 0

Re: Optimizing "counting" GADTs
by David Feuer 19 Jun '16

19 Jun '16

Partially. Unfortunately, bidirectional pattern synonyms tie the types of the pattern synonyms to the types of the smart constructors for no good reason, making them (currently) inappropriate. But fixing that problem would offer one way to this optimization, I think. On May 25, 2016 8:37 PM, "Carter Schonwald" <carter.schonwald(a)gmail.com> wrote: could this be simulated/modeled with pattern synonyms? On Wed, May 25, 2016 at 7:51 PM, David Feuer <david.feuer(a)gmail.com> wrote: > I've started a wiki page, > https://ghc.haskell.org/trac/ghc/wiki/OptimizeCountingGADTs , to consider > optimizing GADTs that look like natural numbers but that possibly have > "heavy zeros". Please take a look. > > _______________________________________________ > ghc-devs mailing list > ghc-devs(a)haskell.org > http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs > >

4 7

Force GC calls out of the straight line execution path
by Harendra Kumar 17 Jun '16

17 Jun '16

Hi, I noticed in the generated code (llvm as well as native) that in some cases the GC calls are in the straight path and the regular code is out of the straight line path. Like this: => 0x408fc0: lea 0x30(%r12),%rax => 0x408fc5: cmp 0x358(%r13),%rax => 0x408fcc: jbe 0x408fe9 # notice jbe instead of ja i.e. branch taken in normal case I tried to count in how many cases its happening in my executable and found that its only a small percentage (4-6%) of cases but those cases include the code which runs 99% of the time in my benchmark. Though it does not make a whole lot of difference but the difference is perceptible and especially when it is in a tight loop. Is it possible to somehow force all the GC calls out of the line during code generation? Has it been thought/discussed before? -harendra

3 4