On 9/19/19 1:34 PM, Segher Boessenkool wrote:
Hi!
On Tue, Sep 17, 2019 at 09:45:54AM +0200, Richard Biener wrote:
The following fixes an old vectorizer issue with realignment support
(thus only powerpc is affected) and BB vectorization. The realignment
token is set up from the wrong data-ref which causes an SSA verification
failure but in other circumstances might simply generate wrong code.
Bootstrap running on x86_64-unknown-linux-gnu, I'll install this
as obvious on trunk.
PPC folks - you know best how to appropriately test a target
where we use the re-alignment optimization. IIRC on later
powerpc hardware this isn't exercised at all since we can use
unaligned accesses.
The issue is at least present on the GCC 9 branch as well but I'd
appreciate testing where it exercises the path before considering
a backport.
Is there a testcase?
Richard, can you turn the PR's reported test into a torture test case?
We post P7 big-endian results frequently to gcc-testresults, and this
bug hasn't fired on anything there, so it's not covered by existing
tests. Nothing has turned up on the testers since your patch went in,
so having the new test added should be sufficient, I'd think. P7 or
older running big-endian is what's needed to test realignment support.
Thanks,
Bill
You can use -malign-natural to get stricter alignment requirements,
that might help.
Cc:ing Bill, this is vectorizer :-)
Segher
