Re: Loop unroll fixes

[mike stump <mrs@windriver.com>]

> Date: Tue, 18 Sep 2001 10:19:02 +0100 (BST)
> From: "Joseph S. Myers" <jsm28@cam.ac.uk>

> * Where <URL: http://gcc.gnu.org/ml/gcc/2001-05/msg01273.html > says
> (of creating synthetic testcases for bugs shown up by confidential
> code) "sometimes specific tests for fixed bugs are definitely cases
> of closing the doors after the cows have fled", this is misleading
> since such testcases may be necessary in future to debug problems
> with the fix; so creating such testcases would not be as described
> there "work that competes with everything else going on" but
> something at which a reasonable effort *must* be made before a fix
> is installed in FSF GCC.

I think that all changes to gcc should come with at least one testcase
that actually demonstrates why the change was a good idea.  If the
code merely changes performance, then the performance suite should
have a testcase that shows the performance advantage.  The last point
we currently never do, so I can be dissuaded for now that it is
unrealistic, but, I think for real bugs, it is possible to meet this
requirement.  The testsuite is invaluable to the quality of the
compiler in so many ways.  It should be a requirement, not optional.

I do realize that sometimes a testcase isn't possible, but in my
experience, those are rare.  A good example of such, would be a
testcase in the C++ framework to test throwing in/out and trough a
shared library.  The old-deja.exp framework doesn't have enough beef
to do this.  A testcase is impossible within it, without massive
modifications to old-deja.exp.

An non-example would be 112 thousand lines of customer proprietary
code which changes any time almost anything is played with.  Been
there, done that, not that hard, just time consuming to trim the
testcase and sanitize it.

If the folks that contributed the testsuite we have felt the same way
as dewar, it is clear to me that we would not have a gcc/g++
testsuite, or that it would be smaller.