Re: [Haskell-cafe] Profiling CAFs (re-post)

10 Oct 2006

      On Tue, Oct 10, 2006 at 01:59:23PM +0100, Ian Lynagh wrote:
...
To: Matthias Fischmann 
Cc: haskell-cafe@haskell.org
From: Ian Lynagh 
Date: Tue, 10 Oct 2006 13:59:23 +0100
Subject: Re: [Haskell-cafe] Profiling CAFs (re-post)
On Tue, Oct 10, 2006 at 01:31:58PM +0200, Matthias Fischmann wrote:
...
What qualifies as constant applicable form, and why is it not
  labelled in a more informative way?
CAFs are, AIUI, things that are just values (i.e. things that don't take
an argument) that have been floated up to the top level.
Ok, this is consistent with the documentation.  However, it doesn't
explain why we would need to treat them differently.  In particular, I
don't understand why I wouldn't want to know their (module-global)
names.
...
Compiling with -caf-all might give you more useful information.
Oops.  I thought i had that in my Makefile, but appearently i was
wrong...  If I add it, this is what happens:

======================================================================
module Main where

x = f [1..5] (f [2..] [3..])

f xs ys = l
    where
    l = [ if s then x else y | (x, y) <- zip xs ys ]
    s = g xs ys
    g [] _ = True
    g _ [] = False
    g (x:xs) (y:ys) = g xs ys

main = print (show x)
======================================================================

$ ghc -prof -caf-all Main.hs -o Main  # (ghc 6.4)
/tmp/ghc22775.hc:1475: error: redefinition of `Mainmain_CAF_cc_ccs'
/tmp/ghc22775.hc:1470: error: `Mainmain_CAF_cc_ccs' previously defined here
/tmp/ghc22775.hc:1490: error: redefinition of `Mainsat_CAF_cc_ccs'
/tmp/ghc22775.hc:1480: error: `Mainsat_CAF_cc_ccs' previously defined here
/tmp/ghc22775.hc:1495: error: redefinition of `Mainsat_CAF_cc_ccs'
/tmp/ghc22775.hc:1490: error: `Mainsat_CAF_cc_ccs' previously defined here

$ ghc -prof -auto-all -caf-all Main.hs -o Main
/tmp/ghc22771.hc:1517: error: redefinition of `Mainmain_CAF_cc_ccs'
/tmp/ghc22771.hc:1512: error: `Mainmain_CAF_cc_ccs' previously defined here
...
If that doesn't help then you might find it helpful to look at heap
profiles rather than just the normal profiler output.
Section 5.4, prof.hp, yes.  Probably should have thought of that
myself.  Will do.
...
...
Why are there functions that inherit all of their (considerable)
  time and space consumption from elsewhere, but nothing in the
  list would allow for such a rich inheritage?
I didn't understand that. If it's possible to give a small example then
that might help?
Small is harder in this case than with type errors.  But I'll try as
soon as I know what's wrong with the above.

thanks,
matthias