Recursive functions and constant parameter closures (inlining/strictness analyzer question)

[Bulat Ziganshin]

Hello Simon,

Friday, May 30, 2008, 5:30:25 PM, you wrote:

may be i don't understand something. isn't it better to do automatic
SAT and inline results for every recursive function marked as INLINE?
it's how i want to work - just mark with INLINE speed-critical funcs.
manual checking that they are recursive and doing appropriate
transformation is too hard for me :) and, btw, how about adding
warnings about functions marked as INLINE which was not actually
inlined due to some reasons - may be very helpful for optimizing
programs without going into studying Core output

> Others have explained this nicely.  But there's a real tension
> here.  The fast version comes from a combination of (a) the static
> argument transformation, so you get the first version above, and (b)
> bodily inlining the entire function, so that at *each call site* you
> get a locally-recursive function where 'f' is known.  That's ok for
> small functions, but not so good for big ones.  Furthermore, the
> code duplication is only worthwhile if the specialisation is truly
> useful. For example, would it be better to write append like this
>   (++) xs ys = letrec app [] = ys
>                       app (x:xs) = x : app xs
>                in app xs
> and inline that at every call of (++)? Probably not.

> So that is why GHC does not automate this transformation.  If you
> know that's what you want, write a local recursion, and use an INLINE pragma.


> If someone felt like summarising this thread on the Haskell
> performance-advice wiki that would be great.
>         http://haskell.org/haskellwiki/Performance

> Meanwhile, I'll clarify in the user manual that recursive functions are not inlined.

> Simon


> _______________________________________________
> Glasgow-haskell-users mailing list
> Glasgow-haskell-users at haskell.org
> http://www.haskell.org/mailman/listinfo/glasgow-haskell-users



-- 
Best regards,
 Bulat                            mailto:Bulat.Ziganshin at gmail.com