Re: Summer of Code

[Martin d'Anjou]

>  So I decided to ask (specially because, in the GNU guidelines for Summer of
>  Code projects[3], they say that that list of ideas is not the only one):
>  isn't there anything that you think I could do to improve Make?

Yes!

A patch was proposed to GNU Make a while ago for GNU Make to immediately 
stop and abort after N failures:
http://lists.gnu.org/archive/html/bug-make/2009-01/msg00035.html

If you read that message in the archives, you will see that it also 
mentions one outstanding problem, but Paul D. Smith explained the 
problem and I was able to solve it using a different approach (see 
below). However what is still not solved to me is the inclusion of that 
patch in GNU Make.

So this patch has not made it into GNU Make as far as I know, and I 
would be really useful to me if it did. I will attempt to explain why.

The way GNU Make works is that it attempts to build as much of the 
desired targets as possible. When using parallel make (-j N) and when 
GNU Make encounters a prerequisite that fails to build (e.g. due to a 
syntax error), it cancels that part of the build, but instead of killing 
the other parallel jobs, it lets them come to their own completion. This 
could take hours in a large build. So you see the problem: one part of 
the build could fail after 2 minutes but you may not notice for hours! 
And I am not sitting at my terminal all the time to watch over builds: I 
use a continuous integration server. So everything must be automated.

I debated whether the CI server (jenkins) should jump in and kill the 
build, but this is not really how it works: the CI server relies on the 
exit status of the programs it runs to abort the build process: so GNU 
Make has to abort, and it has to abort after the first error. In the 
context of the CI server, incrementally building on a previously failed 
build is a no-go: it could compromise subsequent builds. So there is no 
point it letting GNU Make "build as much as it can".

I wrote up a demonstration of a simplified version of this patch, 
including the solution to the aforementioned outstanding problem. You 
can see it here:
http://ninjaverification.wordpress.com/2009/02/03/continous-integration-system-using-parallel-make/

So here is the idea: get that patch integrated with GNU Make, with all 
the quality that GNU Make developers are to expect, so it becomes part 
of the official GNU Make releases forever.

I am not a GNU Make developer, I am simply a mostly happy user which 
could be made even happier!

Thanks,
Martin d'Anjou