- Glasgow-haskell-users

RE: Transitive inlining
by Simon Peyton-Jones 30 Dec '00

30 Dec '00

Yes, all the inlining you expect should indeed happen. If it doesn't can you show us an example? Simon | -----Original Message----- | From: Manuel M. T. Chakravarty [mailto:chak@cse.unsw.edu.au] | Sent: 11 December 2000 13:41 | To: glasgow-haskell-users(a)haskell.org | Cc: keller(a)it.uts.edu.au | Subject: Transitive inlining | | | In the context of the array library, we stumbled over | another problem. Does GHC transitive inlining across | modules? Let's say, we have the following scenario: | | module B (foo) where | | {-# INLINE foo #-} | foo .. = ...bar... | | {-# INLINE bar #-} | bar .. = ... | | | module L (baz) where | | import B | | {#- INLINE baz #-} | baz .. = ...B.foo... | | | module Main where | | import L | | main .. = ...L.baz... | | GHC in this case inlines the whole expression "...B.foo..." | in `Main', but will it inline the right hand side of `B.bar' | in `Main' (or even of `B.foo')? If the INLINE pragma is | transitive across modules boundaries, one would hope so. | | It seems that GHC is not (always?) doing this inlining, | which already for a very simple benchmark costs us a factor | of 5 runtime! (With inlining - enforced by copying | definitions manually into `Main' - the code produced by GHC | is actually a little bit faster than the code for the | corresponding C program). | | Cheers, | Manuel | | PS: I had problems building GHCi and it seems as if my | message to cvs-ghc(a)haskell.org doesn't get through | somehow... | | _______________________________________________ | Glasgow-haskell-users mailing list | Glasgow-haskell-users(a)haskell.org | http://www.haskell.org/mailman/listinfo/glasgow-haskell-users |

2 3

RULES and unfolding
by Manuel M. T. Chakravarty 30 Dec '00

30 Dec '00

Is there any heuristic in GHC that discourages putting unfoldings of functions into .hi files if the function occurs on the right handside of a RULES rule? I have code, where this seems to be the case. When I remove the rewrite rule, the function is placed into the .hi file, but with the rewrite rule, it isn't - despite an INLINE pragma for that function! Unfortunately, the example is rather big and I didn't manage to find a small piece of code showing the same behaviour yet. This currently bites me in a rule of the form {-# "" forall ... foo .. bar .. = foobar .. #-} {-# INLINE foobar #-} foobar .. = ... where I want foobar inlined. Cheers, Manuel PS: I do this only because GHC crashes when I place the rhs of the foobar definition directly into the rhs of the rule, but again I don't have a small example producing the bug *sigh*

1 0

Lambda dropping in loops
by Manuel M. T. Chakravarty 30 Dec '00

30 Dec '00

Given a function like fill 0# mpa = return () fill i# mpa = do let i'# = i# -# 1# writeMP mpa (I# i'#) e fill i'# mpa it would be nice if GHC were to lambda drop it (on the core level) into fill i# mpa = fill' i# where fill' 0# = return () fill' i# = do let i'# = i# -# 1# writeMP mpa (I# i'#) e fill' i'# as this allows it to subsequently apply the new case liberation optimisation. By performing the lambda dropping manually, I have gotten much better code in some performance sensitive definitions. Cheers, Manuel

1 0

RE: Fun with GHC's optimiser
by Simon Peyton-Jones 22 Dec '00

22 Dec '00

Manuel | I have found a way of rephrasing the definition so that it | is properly optimised by GHC. However, I think, it should | be possible to do this automatically and it is maybe not | unlike the optimisation done by simplCore/LiberateCase. I agree with your example. But I think the way to handle it is this. Suppose we have f x y = ....(f ex (C a b))... where C is a constructor, and where y is scrutinised by a case expression somewhere in f's body. Then, in this recursive call to f, we know what y will be, and we could save the case in the recursive call. So make a specialised version of f, thus: fs x a b = let y = C a b in (...original body of f...) Now add a RULE to f, saying f x (C a b) ===> fs x a b This is very much the way type specialisation works. Now every call to f that has a (C a b) argument will benefit. Including the recursive call to f in fs's RHS, so fs will become the self-recursive tight loop you want, leaving f as a sort of 'impedence matcher' for callers elsewhere. I'll try to find time to implement this. Simon

2 1

Generic Classes?
by George Russell 21 Dec '00

21 Dec '00

Are Generic classes going to be in 4.08.2, and if not when are they likely to be in a binary release? I have a potential application, though I can manage without them for the time being.

1 0

Re: GHC Command Line Help
by George Russell 21 Dec '00

21 Dec '00

"Steinitz, Dominic J" wrote: > > I was experimenting with using FiniteMap. The program compiled ok but the linker gave me the following error. What do I need to include on the command line or have I not installed ghc correctly? > > [dom@lhrtba8fd85 FiniteMap]$ /usr/bin/ghc -o main test.o -L/usr/lib/ghc-4.08.1 - > lHSlang -lHSdata > /usr/lib/ghc-4.08.1/libHSdata.a(FiniteMap__1.o): In function `__init_FiniteMap': > FiniteMap__1.o(.text+0x16): undefined reference to `__init_GlaExts' > collect2: ld returned 1 exit status I suggest using /usr/bin/ghc -o main test.o -package data -package lang -fglasgow-exts or something similar (rather than trying to name the libraries directly).

1 0

Re: GHC Command Line Help
by Steinitz, Dominic J 21 Dec '00

21 Dec '00

Sorry. It's obviously -fglasgow-exts. I didn't realise it did things at the linking stage as well as the compilation stage. Dominic Steinitz 21/12/2000 16:30:15 To: glasgow-haskell-users cc: Chris.Reade djrussell B.Avery dominic molyneux bcc: Subject: GHC Command Line Help I was experimenting with using FiniteMap. The program compiled ok but the linker gave me the following error. What do I need to include on the command line or have I not installed ghc correctly? [dom@lhrtba8fd85 FiniteMap]$ /usr/bin/ghc -o main test.o -L/usr/lib/ghc-4.08.1 - lHSlang -lHSdata /usr/lib/ghc-4.08.1/libHSdata.a(FiniteMap__1.o): In function `__init_FiniteMap': FiniteMap__1.o(.text+0x16): undefined reference to `__init_GlaExts' collect2: ld returned 1 exit status ------------------------------------------------------------------------------------------------- 21st century air travel http://www.britishairways.com ------------------------------------------------------------------------------------------------- 21st century air travel http://www.britishairways.com

1 0

GHC Command Line Help
by Steinitz, Dominic J 21 Dec '00

21 Dec '00

I was experimenting with using FiniteMap. The program compiled ok but the linker gave me the following error. What do I need to include on the command line or have I not installed ghc correctly? [dom@lhrtba8fd85 FiniteMap]$ /usr/bin/ghc -o main test.o -L/usr/lib/ghc-4.08.1 - lHSlang -lHSdata /usr/lib/ghc-4.08.1/libHSdata.a(FiniteMap__1.o): In function `__init_FiniteMap': FiniteMap__1.o(.text+0x16): undefined reference to `__init_GlaExts' collect2: ld returned 1 exit status ------------------------------------------------------------------------------------------------- 21st century air travel http://www.britishairways.com

1 0

ANNOUNCE: Happy version 1.9
by Simon Marlow 21 Dec '00

21 Dec '00

ANNOUNCING Happy 1.9 - The LALR(1) Parser Generator for Haskell ----------------------------------------------------------------- I'm pleased to announce version 1.9 of Happy, the parser generator system for Haskell. Changes in this version, relative to version 1.8 (the previous full release): * A grammar may now contain several entry points, allowing several parsers to share parts of the grammar. * Some bugfixes. Happy is available in source form, which can be compiled with GHC version 4.xx (4.08.1 recommended), and we also provide binaries for some architectures. The Happy homepage with links to the various distributions lives at: http://www.haskell.org/happy/ Please send any bug reports and comments to simonmar(a)microsoft.com.

3 2

RE: The Set type
by Simon Marlow 21 Dec '00

21 Dec '00

George Russell wrote: > A minor quibble I know, but every time I use sets I want to > add/subtract single elements > to/from them. There is no function provided for doing this, > so instead you have to do > union/delete with a singleton set constructed from the > element. Done. Set now has addToSet :: Ord a => Set a -> a -> Set a delFromSet :: Ord a => Set a -> a -> Set a The names and argument order were chosen to be consistent with FiniteMap, so don't complain :-) Cheers, Simon

1 0