ghc --make, recompilation problems? (was: GHC 6.2)

[Simon Marlow]

> >We're removing a lot of old/deprecated stuff in 6.2, so I 
> want to be sure we haven't broken 
> >anything before doing the release.  The best estimate I can 
> give is 4-6 weeks, but it could be 
> >sooner if things settle down quickly.
> 
> talking about broken things: I've been reluctant to mention it so far,
> because I can't even give a reproducable example, and the problem 
> only occurs on very large projects, so without a clue as to what 
> might cause it, I certainly couldn't narrow down such an example, 
> should I manage to pin down one. 
> 
> there seems to be a problem in ghc --make's recompilation dependency
> checks, which means that recompilation might produce inconsistent
> executables. the usual effect is a problem-free rebuild, followed by
> inexplicable crashes of the executable. removing the .hi/.o files and
> calling ghc --make again reliably solves the problem, which is why
> I suspect recompilation-only dependency checks (type/class changes
> not propagated to dependent modules?).
> 
> sorry about this rather useless piece of information, but perhaps 
> other people on this list have other parts of the puzzle, and it would
> be nice to get rid of this (haskell programs don't crash like this!-).

We have two known bugs in the recompilation checker.  One is related to
switching the optimisation level while developing a program with --make:
the workaround is not to do this.

The other bug is unfortunately rather more serious  and doesn't have a
workaround.  Essentially it means that --make might not recompile as
much of the program as it should, and you might end up with an
inconsistent executable (which is likely to crash).  Simon P.J. noticed
the bug while revamping the front end recently, and we've verified that
it can indeed lead to crashing programs.  There isn't a simple fix, nor
a workaround, although avoiding -O in conjunction with --make might help
matters.

6.2 will have the bug (because the fix is in Simon P.J.'s rewrite of the
front-end, which is scheduled for 6.4).  That is unless we can think of
a way to fix it that doesn't involve writing hundreds of lines of new
code.

Note that removing all the .o & .hi files and building from scratch will
always result in a working program.  The bug only affects recompilation
after making changes to one or more modules in the program.

This bug has basically been there ever since we had --make, and it has
only been seen in the wild perhaps a handful of times, so it's arguably
not a showstopper.  However, it is rather disturbing.

Cheers,
	Simon