Re: PATCH: Run V3 tests on installed compiler

[Daniel Jacobowitz <drow at false dot org>]

On Wed, Mar 23, 2005 at 12:35:26PM -0800, Mark Mitchell wrote:
> Mike Stump wrote:
> >Also,  I 
> >don't think it is unreasonable for the testsuite to build by  default 
> >the objects it needs during the build, and then just using  them if they 
> >like that workflow.  The difference can be put into the  site.exp file, 
> >such as:
> >
> >    set TESTING_IN_BUILD_TREE 1
> >
> >and then that can be checked; simple, quick, efficient, clear.
> 
> I think that's a mistake.
> 
> I think it's a desirable invariant to have the two testing methods (in 
> the build tree, and in the installed tree) as similar as possible.  With 
> TESTING_IN_BUILD_TREE, I think we're very likely to (a) consistently 
> break testing in the installed tree, when we update only part of some 
> conditional, or forget about it, and (b) have hard-to-explain 
> differences in results between the two approaches.  We also have to 
> maintain the build logic for the secondary objects in two places: (a) 
> Makefiles, for use in the build tree, and (b) DejaGNU, for use in 
> isntalled tree.
> 
> Obviously, there are essential differences for testing in the build 
> tree, in that compiler paths are different, -B/-L options are required, 
> etc. -- but I'd like to see us keep the set of differences as small as 
> reasonably possible.

Strongly agree.  Every year or two I need to try
./contrib/test_installed for one reason or another, and more of these
differences have popped up.

I think the only way to solve this reliably is to kill the current
build-tree testing; the previously proposed plan of a "staging"
installation directory, and letting GCC's improved relocatable install
support do the rest of the work.

-- 
Daniel Jacobowitz
CodeSourcery, LLC