Svadev mailing list

[Will Dietz <willdtz AT gmail.com>]

Hi Jingyue, taking this off bugzilla since I believe this has gone
sufficiently off-topic, certainly for that particular bug report :).

On Wed, May 9, 2012 at 4:05 PM,  
<bugzilla-daemon AT llvm.org>
 wrote:
> http://llvm.org/bugs/show_bug.cgi?id=12744
>
> --- Comment #10 from Jingyue Wu 
> <wujingyue AT gmail.com>
>  2012-05-09 16:05:02 CDT ---
> I would be happy to contribute some test cases. But I am having some issues
> running the test suite under poolalloc (trunk). It says:
>
> jingyue@jingyue:~/Research/llvm/poolalloc-trunk/test$
>  make
> (cd /home/jingyue/Research/llvm/llvm-3.0/projects/test-suite/; \
>               PROJECT_DIR=/home/jingyue/Research/llvm/poolalloc-trunk make 
> -j1
> TEST=poolalloc \
>                   LARGE_PROBLEM_SIZE=1 GET_STABLE_NUMBERS=1 report
> report.html)
> /bin/sh: 1: cd: can't cd to
> /home/jingyue/Research/llvm/llvm-3.0/projects/test-suite/
> make[1]: Entering directory 
> `/home/jingyue/Research/llvm/poolalloc-trunk/test'

Well, this is the problem here :).

Looks like you either don't have the LLVM test-suite checked out, or
have it installed in a "non-standard" location.

Unfortunately atm poolalloc isn't flexible as to where it expects the
test-suite, so can you verify that you indeed have checked out the
LLVM test-suite into
"/home/jingyue/Research/llvm/llvm-3.0/projects/test-suite/"?

However, I just committed (r156516) a change that simply runs the lit
tests as the default testing target, instead of what you just ran
into: attempting to run poolalloc on all of LLVM's test-suite (which
AFAIK no one's even tried in years).

(If you're intentionally trying to run the test-suite with poolalloc,
you can still do so with a 'make fulltest')

FWIW I believe John was suggesting you write some lit tests for DS-AA,
not trying to run the test-suite (although that can be useful too).
Unfortunately due to various outstanding LLVM bugs, it's not really
practical to use "build the test-suite with -O3 and DS-AA as the alias
analysis" as a correctness test--both because it's not possible to
reliably specify that LLVM should use our AA (PR7720), and because
LLVM doesn't play nice with the AA interface it's defined (PR7615).

Finally, there are some lit tests for basicAA in
llvm/test/Analysis/BasicAA that look to be useful examples as starting
points.  In particular the ones that use '-aa-eval' more directly
reason about the AA's results, which seems a good deal less awkward
than using an optimization to indirectly test the desired AA result.
Just a thought, any tests at all would be very very welcome :).

Thanks!!

~Will