March 2009 - jhc - Haskell.org

[john@repetae.net: Re: [Haskell-cafe] I want to write a compiler]
by John Meacham 10 Mar '09

10 Mar '09

forwarding this to the jhc list as it is jhc relevant ----- Forwarded message from John Meacham <john(a)repetae.net> ----- To: haskell-cafe(a)haskell.org List-Id: The Haskell Cafe <haskell-cafe.haskell.org> Date: Mon, 9 Mar 2009 17:41:01 -0700 From: John Meacham <john(a)repetae.net> Subject: Re: [Haskell-cafe] I want to write a compiler On Mon, Mar 09, 2009 at 09:25:58AM -0500, Austin Seipp wrote: > Indeed, I stumbled upon it whilst looking at how unsafeCoerce worked > (to find out it is super-duper-special and implemented as part of E.) > I think it's actually pretty clever, and who knows, maybe it could be > useful as at least a debugging aid. :) It is actually not all that special, it is just a primitive like any other than transates to a nop, or no code. Since the optimizer can't know what is inside a primitive in general, it leaves it alone and the only affect is the type coercion implicit in its signature. When translating to Grin, the unsafeCoerces are simply dropped. Coercions used to be more complicated because once upon a time I required them to be used to translate to/from newtype declared synonyms, I found they inhibited optimizations too much so switched to the current mechanism whereby newtype expansion happens transparently in the type checker when needed. As a side effect I worked hard to try to recognize and optimize around coercions, but now that they occur fairly rarely, it isn't much of an issue any more. > This is a very good point I hadn't even thought about! Indeed, since > GRIN represents thunks in a defunctionalized way - encoded as nodes - > dealing with boxed/unboxed values becomes more of the compiler's job, > since the nature of unboxed values etc. becomes more transparent. The main reason I think it is important is because I feel not including unboxed values from the get-go was one of the biggest false starts I had when writing jhc. When I initially started writing it I considered unboxed values as a feature to be added later. When I started implementing them, I found that there were a lot of decisions I would have made differently and a lot of places they would have greatly simplified my initial design had I had them from the start. Namely, A whole host of 'compiler magic' could have gone away if I implemented all primitive types in pure haskell like I implement floats and chars now, data Float = Float_ Float32_ data Char = Char Bits32_ (where Float32_ and Bits32_ are the unboxed raw types) Since instead I had the compiler implement Int and friends directly, it suddenly had the responisibity to conjure up all sorts of operations on them, such as the Num instances, creating the big mess of compiler magic in PrimitiveOperators.hs. The more I can express in the haskell source directly the better, it raises odd questions when the compiler has to do things like implement Num instances directly, like the instances are built in, but we might not have 'imported' the types needed to declare them into the current scope, so we end up having to deal with instancse for types that don't exist yet. In any case, yeah. if you plan on unboxed types at some point, include them from the start because they can simplify a lot to offset their cost of implementation, by trying to add them on later I ended up having to pay the cost of implementing them, but also had to deal with the cruft left over from not utilizing them from the beginning. > Since you bring this up, I figure this decision also had some > influence on E in lhc/jhc, considering its type system is rich enough > to distinguish values in whnf/boxed/unboxed etc.? Yeah, there were a couple things that motivated this. A big one is that evals are bad in jhc, signifigantly moreso than traditional ghc. The points-to analysis is the main way to fight it, but I also decided I wanted to give it the best shot possible by getting rid of as many evals as I could in higher level optimizations, which necessitated a fairly rich type system since it actually made sense for jhc to keep strict track of what was in WHNF vs not even if they were both boxed. in general ghc didn't care unless a value was either CPRable or unboxable, but jhc can benefit from such knowledge no matter what the type, even if it is a partially applied function. Though, it is interesting that GHC may be moving to a jhc-style semi-tagging mechanism according to this paper[1] so more optimizations may help both equally in the future. The other main motivating factor was that I was always enamoured with the idea of jhc supporting more than one source language, in particular, I wanted to explore it having an SML front end, so wanted a core type system that was equally good at expressing lazy haskell as strict ML, and more importantly would be able to seamlessly intermingle the two without difficulty. So, when designing it, I kept that in mind, "is this equally useful as a ML compilers core?" or a "wacky ML/Haskell hybrid?". Although the ML front end is more of a thought experiment at this point, I think it helped me come up with a coherent type system that had intruiging possibilies for haskell optimizations, as well as some language implications, such as the ability to declare new 'strict' algebraic types that may lead to new ideas for language extensions. John [1] http://research.microsoft.com/en-us/um/people/simonpj/papers/ptr-tag/ -- John Meacham - ⑆repetae.net⑆john⑈ -- Haskell-Cafe mailing list Haskell-Cafe(a)haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe ----- End forwarded message ----- -- John Meacham - ⑆repetae.net⑆john⑈

1 0

darcs patch: update building documentation (and 11 more)
by John Meacham 07 Mar '09

07 Mar '09

Wed Mar 4 19:37:39 PST 2009 John Meacham <john(a)repetae.net> * update building documentation Thu Mar 5 02:06:33 PST 2009 John Meacham <john(a)repetae.net> * fix foreign import of c_strcmp to have pointers as arguments Thu Mar 5 02:16:54 PST 2009 John Meacham <john(a)repetae.net> * fix Grin.Devolve to catch all lifted functions in a single pass and not have to recurse Thu Mar 5 03:04:59 PST 2009 John Meacham <john(a)repetae.net> * remove superfluous 'Atom' argument to the tyvar [binary format change] Thu Mar 5 03:06:30 PST 2009 John Meacham <john(a)repetae.net> * update version number to 0.6.0 Thu Mar 5 17:40:27 PST 2009 John Meacham <john(a)repetae.net> * code cleanups and tweak strictness and UNPACK pragmas Thu Mar 5 20:43:27 PST 2009 John Meacham <john(a)repetae.net> * don't include '()' in parameters to primitives. Thu Mar 5 20:44:08 PST 2009 John Meacham <john(a)repetae.net> * add various ways to test compile time properties in Jhc.Options Thu Mar 5 21:30:12 PST 2009 John Meacham <john(a)repetae.net> * clean up handling of Jhc.Options options Fri Mar 6 00:58:52 PST 2009 John Meacham <john(a)repetae.net> * clean up core and grin dumping some Fri Mar 6 00:59:22 PST 2009 John Meacham <john(a)repetae.net> * add back in an optimization step after lambda lifting to fix up some core that translates to pessismal grin Fri Mar 6 02:36:07 PST 2009 John Meacham <john(a)repetae.net> * minor improvements

1 0

new packages for jhc, applicative and containers
by John Meacham 05 Mar '09

05 Mar '09

I am happy to announce that as of the 0.5.20090304 release, the default jhc install comes with a few more packages that make it a lot easier to compile random code out there, applicative: Control.Applicative, Control.Arrow, Control.Category, Data.Foldable, Data.Traversable containers: Data.Graph, Data.IntMap, Data.IntSet, Data.Map, Data.Sequence, Data.Set, Data.Tree They are included in the tarball and rpms and the default build. you have to pull containers from the darcs repository as mentioned on the building page, but other than that it builds directly from the shared containers source with no hackery required! Information on installing jhc is here: http://repetae.net/computer/jhc/building.shtml and the main jhc page is here: http://repetae.net/computer/jhc John -- John Meacham - ⑆repetae.net⑆john⑈

1 0

darcs patch: clean up renaming code, collect free var... (and 6 more)
by John Meacham 05 Mar '09

05 Mar '09

Wed Mar 4 17:12:29 PST 2009 John Meacham <john(a)repetae.net> * clean up renaming code, collect free variables from SPECIALIZATION pragmas before renaming Wed Mar 4 17:12:34 PST 2009 John Meacham <john(a)repetae.net> * clean out unneeeded hsQualTypeHsContext function Wed Mar 4 17:13:15 PST 2009 John Meacham <john(a)repetae.net> * perform specialization of SPECIALIZE pragmas that have free variables properly Wed Mar 4 17:54:37 PST 2009 John Meacham <john(a)repetae.net> * disable bang patterns for now Wed Mar 4 18:07:13 PST 2009 John Meacham <john(a)repetae.net> * add applicative and containers to distributed libraries Wed Mar 4 18:07:39 PST 2009 John Meacham <john(a)repetae.net> tagged neopdoxhedvo Wed Mar 4 18:08:02 PST 2009 John Meacham <john(a)repetae.net> * update datestamp

1 0

darcs patch: clean up renaming code, collect free var... (and 6 more)
by John Meacham 05 Mar '09

05 Mar '09

Wed Mar 4 17:12:29 PST 2009 John Meacham <john(a)repetae.net> * clean up renaming code, collect free variables from SPECIALIZATION pragmas before renaming Wed Mar 4 17:12:34 PST 2009 John Meacham <john(a)repetae.net> * clean out unneeeded hsQualTypeHsContext function Wed Mar 4 17:13:15 PST 2009 John Meacham <john(a)repetae.net> * perform specialization of SPECIALIZE pragmas that have free variables properly Wed Mar 4 17:54:37 PST 2009 John Meacham <john(a)repetae.net> * disable bang patterns for now Wed Mar 4 18:07:13 PST 2009 John Meacham <john(a)repetae.net> * add applicative and containers to distributed libraries Wed Mar 4 18:07:39 PST 2009 John Meacham <john(a)repetae.net> tagged neopdoxhedvo Wed Mar 4 18:08:02 PST 2009 John Meacham <john(a)repetae.net> * update datestamp

1 0

darcs patch: parse SPECIALIZE instance pragmas (and 2 more)
by John Meacham 04 Mar '09

04 Mar '09

Wed Mar 4 03:18:05 PST 2009 John Meacham <john(a)repetae.net> * parse SPECIALIZE instance pragmas Wed Mar 4 03:18:31 PST 2009 John Meacham <john(a)repetae.net> * add 'containers' package to distributed packages Wed Mar 4 03:22:34 PST 2009 John Meacham <john(a)repetae.net> * fix name of Read class

1 0

darcs patch: major clean ups to Desugar module, remov... (and 2 more)
by John Meacham 01 Mar '09

01 Mar '09

Fri Feb 27 20:52:44 PST 2009 John Meacham <john(a)repetae.net> * major clean ups to Desugar module, remove almost 200 lines of code inherited from hatchet Fri Feb 27 21:07:36 PST 2009 John Meacham <john(a)repetae.net> * add Eq and Ord istances for ForeignPtr Fri Feb 27 21:28:50 PST 2009 John Meacham <john(a)repetae.net> * add hIsEOF and hWaitForInput to standard library

1 0

darcs patch: thread options into the renamer and simp... (and 4 more)
by John Meacham 01 Mar '09

01 Mar '09

Fri Feb 27 18:14:01 PST 2009 John Meacham <john(a)repetae.net> * thread options into the renamer and simplify renamer some by leaving wrapping of constructors to the type checker Fri Feb 27 18:37:28 PST 2009 John Meacham <john(a)repetae.net> * added HsLet, HsDo, and HsLocated to syntax traversal code Fri Feb 27 19:45:20 PST 2009 John Meacham <john(a)repetae.net> * don't desugar list comprehensions until after renaming has occured, always bind to the prelude functions in list comprehension expansion Fri Feb 27 20:15:35 PST 2009 John Meacham <john(a)repetae.net> * clean up some old cruft dealing with names Fri Feb 27 20:23:30 PST 2009 John Meacham <john(a)repetae.net> * don't desugar do bindings until after renaming

1 0

darcs patch: various library fixes, added minusPtr, o... (and 2 more)
by John Meacham 01 Mar '09

01 Mar '09

Fri Feb 27 06:25:45 PST 2009 John Meacham <john(a)repetae.net> * various library fixes, added minusPtr, openBinaryFile, hGetLine, fixed type for C finalizers Fri Feb 27 06:26:07 PST 2009 John Meacham <john(a)repetae.net> * allow foreign import specification parameters to appear in arbitrary order Fri Feb 27 06:30:40 PST 2009 John Meacham <john(a)repetae.net> * clean up old renaming code for HsExp

1 0

darcs patch: allow foreign exporting functions with u... (and 5 more)
by John Meacham 01 Mar '09

01 Mar '09

Fri Feb 27 00:31:38 PST 2009 John Meacham <john(a)repetae.net> * allow foreign exporting functions with unboxed arguments and return values Fri Feb 27 01:03:27 PST 2009 John Meacham <john(a)repetae.net> * clean up code some, add Addr_ and FunAddr_ unboxed newtypes for when you need to use a specific calling convention with foreign code Fri Feb 27 02:42:45 PST 2009 John Meacham <john(a)repetae.net> * add stub typeable library Fri Feb 27 02:50:42 PST 2009 John Meacham <john(a)repetae.net> * fix creation of complicated conjured type names, don't print built in data table entries unless requested Fri Feb 27 04:39:45 PST 2009 John Meacham <john(a)repetae.net> * add ffi export and import regression test Fri Feb 27 04:42:54 PST 2009 John Meacham <john(a)repetae.net> * add System.Environment

1 0