impressions on ghc-6.4

[Simon Marlow]

On 08 June 2005 12:27, Serge D. Mechveliani wrote:

> Cabal
> -----
> 
> It looks good. Thanks to developers.
> Long ago I asked for the building tool in GHC which would allow to
> avoid composing Makefile-s, with their difficult syntax and
> odd-looking system commands: `ar', `ld', and such.
> Now, Cabal does what I asked for.

Glad you like it.  See, we do listen sometimes! :-)

> Remarks:
> * maybe, `version' should not be mandatory in the package
>   specification?
> * " version: 5.02 "   (imagine "ghc-5.02")  converts to  "5.0.2" .
>   Could it allow digit sequences separated by period?

5.02 is parsed as 5.2 by the Version parser:

> readP_to_S parseVersion "5.02"
[(Version {versionBranch = [5], versionTags = []},".02"),(Version
{versionBranch = [5,2], versionTags = []},"")]

so I'm not sure what you're referring to here.  Could you give a
concrete example?

> -O, memory management
> ---------------------
> 
> In  ghc-6.2.2  the test for  dumatel-1.02
> runs 2 times faster when Dumatel is built under -O.
> And in ghc-6.4, the time does not differ.
> 
> Then, I tried the benchmark with certain mixture of list operations:
> user-written  maximum and mergeSort,  `reverse' and `zip' of GHC
> -- find enclosed the source.
> 
> -O is 1.3 faster than -Onot.
> The speed looks like in  ghc-5.02  (I think, this is all right).
> 5.02 was able to run `main' for n = 55000 in  200 Kb memory.
> 6.4 needs 2600 Kb.
> Probably, 5.02 is distinguished among other versions in memory
> management.

A difference of 2.5Mb is really very small.  If you could show that the
space usage is linear in one case and flat in another, then we have a
problem.

Cheers,
	Simon