On Mon, Sep 13, 2021 at 01:57:52PM +0200, Christophe Lyon wrote:
> > --- gcc/testsuite/c-c++-common/gomp/atomic-29.c.jj 2021-09-10
> > 11:47:17.093164041 +0200
> > +++ gcc/testsuite/c-c++-common/gomp/atomic-29.c 2021-09-10
> > 11:52:33.428722747 +0200
> > @@ -0,0 +1,43 @@
> > +/* { dg-do compile { target c } } */
> > +/* { dg-additional-options "-O2 -fdump-tree-ompexp" } */
> > +/* { dg-final { scan-tree-dump-times "\.ATOMIC_COMPARE_EXCHANGE
> > \\\(\[^\n\r]*, 8, 5, 5\\\);" 1 "ompexp" { target sync_int_long } } } */
> > +/* { dg-final { scan-tree-dump-times "\.ATOMIC_COMPARE_EXCHANGE
> > \\\(\[^\n\r]*, 8, 4, 2\\\);" 1 "ompexp" { target sync_int_long } } } */
> > +/* { dg-final { scan-tree-dump-times "\.ATOMIC_COMPARE_EXCHANGE
> > \\\(\[^\n\r]*, 264, 5, 0\\\);" 1 "ompexp" { target sync_int_long } } } */
> > +/* { dg-final { scan-tree-dump-times "\.ATOMIC_COMPARE_EXCHANGE
> > \\\(\[^\n\r]*, 8, 0, 0\\\);" 1 "ompexp" { target sync_int_long } } } */
> > +/* { dg-final { scan-tree-dump-not "__atomic_load_8 \\\(" "ompexp" {
> > target sync_int_long } } } */
> > +
> >
>
> This test fails on arm*linux when forcing old CPU/ARCH (eg -march=armv5t).
> Not sure how easy it is to fix?
> sync_int_long returns true for arm*-*-linux-*,
> but for other arm targets, it depends on the result
> of check_effective_target_arm_acq_rel.
>
> Is it just a matter of removing arm*-*-linux-* from sync_int_long and
> always rely on arm_acq_rel?
So, atomic-28.c passes and just atomic-29.c fails?
atomic-29.c tests for double atomics aka 64-bit, so probably I should use
sync_long_long effective target instead of sync_int_long.
But, it is unclear how would that help for arm because sync_long_long is
enabled for all arm targets...
There is also sync_long_long_runtime effective target, but this is a compile
test, so it would be weird to rely on that when it ought to test the runtime
behavior of that.
Jakub
