Is there a real need for -O2 in darcs.cabal?  By putting it there, -O0
can't be used without manually editing the .cabal file, which is a
major pain in the arse.

AIUI cabal uses -O2 by default anyway, cf. --enable-optimization
vs. --disable-optimization.

The issue arises because my buildbot, siva, has "only" 512MB of RAM
and I am investigating ways to avoid ghc being OOM-killed.

cabal install --help:

 -O --enable-optimization[=n]      Build with optimization (n is 0--2, default
                                   is 1)

Cabal used to use -O2 by default, but people complained about the buildtime and
Duncan reasoned that most libraries and apps wouldn't see a major difference anyway.

Whether that is true or not for Darcs is something that needs benchmarking.

OK: so when somebody has a round tuit, drop -O2 from the cabal file, build a
regular version and an -O2 version and run it through the darcs-benchmark
benchmarks?  Perhaps the unit tests as well for timing?

Eric Kow wrote:
> OK: so when somebody has a round tuit, drop -O2 from the cabal file,
> build a regular version and an -O2 version and run it through the
> darcs-benchmark benchmarks?  Perhaps the unit tests as well for
> timing?

Even if -O1 (the default) produces a binary that's less performant
than -O2, why is that important?  The current behaviour essentially
forces ALL users to use -O2, regardless of the Cabal defaults or user
preferences (i.e. ~/.cabal and --enable-optimization=N).

On Tue, Nov 24, 2009 at 03:22:38 +0000, Trent W. Buck wrote:
> Even if -O1 (the default) produces a binary that's less performant
> than -O2, why is that important?  The current behaviour essentially
> forces ALL users to use -O2, regardless of the Cabal defaults or user
> preferences (i.e. ~/.cabal and --enable-optimization=N).

Because (I claim) users generally want a fast Darcs than a fast compiled
Darcs.

These users would be forced to explicitly pass in flag (not acceptable
because they won't do it, either by oversight or otherwise) or modify
their Cabal defaults for all packages and not just Darcs (not
acceptable, ick).  What we really want is for Darcs to supply -O2 as a
default which can then be overridden.  Simpler would be if the
performance gains were negligible, in which case we should just drop the
default.

-- 
Eric Kow <http://www.nltg.brighton.ac.uk/home/Eric.Kow>
PGP Key ID: 08AC04F9