Problems dealing with GHC incompatibilities

Simon Marlow
Fri, 26 Jul 2002 16:37:17 +0100

> What version of GHC do you use at Microsoft?  Do you have a=20
> last-not-broken version
> which you use for developing GHC or what?  If there is a=20
> last-not-broken version it would
> be good to have access to it.

We generally have 'ghc' pointing to the latest stable release (i.e. now
it points to 5.04), and I usually install a snapshot whenever we make
one.  We don't develop day-to-day using snapshots, because switching
over existing builds is a little annoying.

> For quite a long time I ran an automated system which=20
> compiled GHC every night from
> CVS (using the previous version) and installed it on our=20
> system if the compile completed.  This
> actually worked quite well, though I stopped doing it because=20
> the compilation system got
> more complicated, I had problems doing it on Sparc/Solaris,=20
> and it was too much trouble.
> If the compile finished it was unusual for the compiler=20
> itself to be hopelessly broken;
> to be on the safe side you could check it was able to compile=20
> Happy and Haddock.  Something
> like this might be good.  Of course it would have to be made=20
> very clear that such snapshot
> releases were for evaluation only, and that only the official=20
> releases were to be relied on.

Our nightly builds run a 3-stage bootstrap and then the whole test
suite, so you usually have a pretty good idea by the end whether the
compiler is a good 'un or not.

You can checkout the scripts we use for the nightly builds from CVS:
just 'cvs co nightly'.  Setting up your own build using these scripts is
fairly painless; Malcolm Wallace did it recently for Sparc to bolster
our Sparc support a bit.