[Nix-dev] Re: Accepting -jX on builds

[Ludovic Courtès]

Hello!

Eelco Dolstra <e.dolstra at tudelft.nl>
writes:

> I used to be against this because the purity risk, but if it's turned *on* on a
> per-package basis, it's probably an acceptable risk.  (We also have a stack of
> 8-core machines here that we should put to good use ;-)

I used to be in favor of something like that but...

It's very hard to check whether a package is parallel-build-safe.  You'd
have to, e.g., build it tens of times and if the output is the same for
all builds, you'd get *some* confidence.

> PS: I have this wild idea of making a front end for Nix that reads a Makefile
> and basically converts it to a Nix expression - that is, every Make rule becomes
> a derivation.  Then you get all the parallelism you want and it would still be
> pure.  Or conversely, if your Makefile isn't pure (e.g. if build actions
> interfere with each other), it would deterministically fail.  As an added bonus,
> you would never have to write another "make clean" rule :-)

Incidentally, there was once work about non-timestamp-based dependencies
in GNU Make, which would allow the use of crytographic hashes to
determine whether a target is up-to-date:

  http://lists.gnu.org/archive/html/make-alpha/2007-10/msg00007.html

  http://lists.gnu.org/archive/html/make-alpha/2007-04/msg00008.html

That was apparently never integrated, though.

Thanks,
Ludo'.