On 10/28/2015 06:39 AM, Richard Biener wrote:
On Wed, Oct 28, 2015 at 1:34 PM, Richard Biener
<richard.guent...@gmail.com> wrote:
On Tue, Oct 27, 2015 at 8:49 PM, David Malcolm
<dmalc...@redhat.com> wrote:
gcc/testsuite/ChangeLog: * unittests/test-gimple.c: New file.
I fear that keeping tests like this up-to-date and working is more
of a headache than the value we get from it.
Perhaps. I'm not sure either way on that specific example. It might be
more interesting to do something like that for RTL which doesn't have as
much static structure. And it'd provide a basis for testing code to
bring more static structure to RTL.
So - with our testsuite and passes we have the issues that we'd like
to have regression tests that nail down the IL input to a specific
pass. That rots over time (obviously) so many tests are no longer
testing what they are supposed to test.
Yes, absolutely. But I think that class of problems it outside the
scope of what David is trying to tackle.
Essentially this kind of testing would be a good replacement for most of
the dump-scanning tests we have.
Some years ago I proposed -Otest to make the input into passes a
little bit more stable (basically run only a single pass). It
requires the desired pass input to be created from the lowering
stages only of course (you could enable more than one pass but then
things get less stable again).
I must have missed that. But it's an interesting idea.
Ideally we'd get progress with the GIMPLE FE up to the point that
textual input to passes works (we need pass manager support for this
as well, of course).
We haven't really started planning for next year yet within Red Hat, but
pushing the gimple front-end has already been mentioned at least once as
a potential project.
Jeff
