| Message ID | 1456998378-22605-1-git-send-email-james.greenhalgh@arm.com |
|---|---|
| State | New |
| Headers | show |
On Wed, Mar 09, 2016 at 12:53:02PM +0100, Rainer Orth wrote: > Richard Biener <rguenther@suse.de> writes: > > > On Thu, 3 Mar 2016, James Greenhalgh wrote: > > > >> > >> Hi, > >> > >> ARM and AArch64 will still vectorize bb-slp-34.c - we're not operating > >> with a cost model so we vectorize to a 64-bit vector of two ints, and the > >> permutes are just element swaps. > >> > >> So, don't mark this test xfail for arm/aarch64. > >> > >> Checked on x86_64-none-linux-gnu, arm-none-eabi and aarch64-none-elf with > >> no issues. > >> > >> OK? > > > > Ok. Indeed with using V2SI vectors the vectorization is valid. > > Unfortunately, the patch broke the test on sparc: > > FAIL: gcc.dg/vect/bb-slp-34.c -flto -ffat-lto-objects scan-tree-dump slp2 "basic block vectorized" > FAIL: gcc.dg/vect/bb-slp-34.c scan-tree-dump slp2 "basic block vectorized" > > While before the scan-tree-dump wasn't run due to missing vect_perm > support, it now fails. > > I believe you meant to just not xfail the test on aarch64* and arm*, Indeed. > which is what the following patch does. James, could you please test > it? Yes, that still does the right thing for me on aarch64-none-elf, and I think I now understand why. Before I touched it that said: /* { dg-final { scan-tree-dump "basic block vectorized" "slp2" { target vect_perm xfail *-*-* } } } */ Which means "only run this statement for vect_perm targets, but expect it to fail on all targets". Then I changed it to say : /* { dg-final { scan-tree-dump "basic block vectorized" "slp2" { xfail { { vect_perm } && { ! { aarch64*-*-* arm*-*-* } } } } } } */ Which means "always run this, but expect it to fail on vect_perm targets that are not arm/aarch64". Now you've changed it to say: /* { dg-final { scan-tree-dump "basic block vectorized" "slp2" { target vect_perm xfail { ! { aarch64*-*-* arm*-*-* } } } } } */ Which means "only run this for vect perm targets, and expect it to fail if the target is not arm/aarch64. Thanks for the patch, it looks right to me and is what I was originally trying to write. Sorry for the sparc (and presumably other vect_int && !vect_perm targets) break. Thanks, James
diff --git a/gcc/testsuite/gcc.dg/vect/bb-slp-34.c b/gcc/testsuite/gcc.dg/vect/bb-slp-34.c index 418f2b5..7b9511a 100644 --- a/gcc/testsuite/gcc.dg/vect/bb-slp-34.c +++ b/gcc/testsuite/gcc.dg/vect/bb-slp-34.c @@ -33,4 +33,4 @@ int main() } /* ??? XFAILed because we access "excess" elements with the permutation. */ -/* { dg-final { scan-tree-dump "basic block vectorized" "slp2" { target vect_perm xfail *-*-* } } } */ +/* { dg-final { scan-tree-dump "basic block vectorized" "slp2" { xfail { { vect_perm } && { ! { aarch64*-*-* arm*-*-* } } } } } } */