Re: autotest vs. in-use directories

[Ralf Wildenhues]

* Eric Blake wrote on Fri, Apr 10, 2009 at 02:29:50PM CEST:
> According to Ralf Wildenhues on 4/10/2009 12:02 AM:
> > 
> > But adding a test case that proves that all
> > three globs are necessary would be prudent.
> 
> Yep, particularly since I botched the m4 quoting for the .[!.] glob.  As
> penance, I factored this into a new m4sh macro.  I'm committing this
> (which leaves it undocumented); if we like it, then we can go one step
> further by removing the leading _ and documenting it.  Or even decide to
> fold in the subsequent AS_MKDIR_P step, to guarantee a clean dir whether
> or not it previously existed as a directory, but failing if it existed as
> a file.  [side note: Cygwin 1.5 can't touch '...' without a managed mount,
> but cygwin 1.7 removed that restriction.]

Thanks for fixing this.  BTW, I am more and more convinced that not only
should we add regression tests for all bugs that are found, but also full
coverage tests for new features we add to autotools.  IOW, ideally, each
code path added should be exercised by the test suite at least on some
system.  Of course one cannot reasonably go for all combinations of
decisions; here, judgement is called for.

Merely looking at some of the file existence tests in Automake snippets
made me find bugs in the code.

As to a public AS_CLEAN_DIR, it would probably be useful for Libtool.
However, I don't mind to postpone this and wait for some real-world
feedback on the undocumented macro for now; I'll bet some system out
there won't behave as we expect on this code.

Cheers,
Ralf