Re: nobase.test

[Ralf Wildenhues]

Hello Peter,

* Peter O'Gorman wrote on Wed, Sep 06, 2006 at 01:08:53AM CEST:
> >>On Sun, Sep 03, 2006 at 09:33:20AM +0200, Ralf Wildenhues wrote:
> >>>I say we drop the test and require that install-sh be executable.
> >>>After all, this is so much easier to do.

> On a regular basis, I patch packages to use the autotools build  
> system that did not originally do so. Our build system unpacks the  
> original tarball and applies patches with patch. This means that  
> scripts get created without execute bits, so the build fails. The  
> idea of the test that Alexandre installed was to ensure that automake  
> would be able to continue to run without executable bits being set,  
> not so that the test could be removed when someone broke automake so  
> that it no longer works without being executable.

Ok, let me formulate this differently:

Autoconf and Automake have _never_ really supported a non-executable
`install-sh' script in the source tree.  The test exposed merely one of
several different usage cases, many of which would have always failed.
When this change was first installed, I didn't like it, but also I
did not see all the failure cases.  Now that I see them (and that is not
only the move to use AC_PROG_MKDIR_P), I think it is far too much hassle
and the risk of breaking users' package setups may be too high, IMVHO.

For example, the $(INSTALL), $(INSTALL_DATA), and $(INSTALL_PROGRAM) are
documented interfaces of Autoconf, and have never been safe for this.
The new Autoconf interface $(MKDIR_P) is documented to be used without
prefixing $(SHELL).

Maybe there is a solution to all cases.  But until then, I don't think
it's a good idea to advertise half-solutions, and employ half-tests.
And yes, that is very much IMHO.

Cheers,
Ralf