This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: [lno] auto vectorizing on x86-64 broken
- From: Sebastian Pop <sebastian dot pop at cri dot ensmp dot fr>
- To: Andi Kleen <ak at muc dot de>
- Cc: gcc at gcc dot gnu dot org
- Date: Sat, 3 Jul 2004 16:59:07 +0200
- Subject: Re: [lno] auto vectorizing on x86-64 broken
- References: <m3acyhfzd6.fsf@averell.firstfloor.org>
On Sat, Jul 03, 2004 at 07:12:21AM +0200, Andi Kleen wrote:
>
> I changed the vectorizing testsuite in LNO to test and execute on x86-64
> (patch can be found in http://www.firstfloor.org/~andi/vect-test).
> Previously it would only execute on ppc and check log files on i386.
>
thanks. If nobody objects, I will include your patch in LNO.
> Unfortunately many of the tests fail. The compilation and
> execution works, but I get a lot of
>
> FAIL: gcc.dg/tree-ssa-vect/tree-ssa-vect-17.c scan-tree-dump-times vectorized lloops 1
>
> which means the loops weren't actually vectorized. The log file just shows
>
> vectorized 0 loops in function.
>
try to replace -fdump-tree-vect-stats with -fdump-tree-vect-details,
that will turn on much more debugging information. Then search for
"bad" that is inserted every time the vectorizer fails.
> It seems to vectorize a few loops in some test cases, but the majority
> fails. Any ideas why this doesn't work on x86-64?
>
>From what I've seen in the dumps:
get vectype for scalar type: int
op not supported by target
use not simple.
stmt not supported.
T.32_145 = T.30_142 & T.31_144;
loop_analyzer: bad operations.
The same error for char and short int, same for |, etc.
tree-ssa-vect-30.c produces an ICE. I'll look into this problem.
thanks,
Sebastian