On Tue, 22 Nov 2011, Andrew MacLeod wrote:
> int n = get_atomic_generic_size (loc, function, params);
>
> + /* Size of 0 is an error condition. */
> + if (n == 0)
> + {
> + *new_return = error_mark_node;
> + return true;
> + }
As far as I can see, get_atomic_generic_size can return 0 in various cases
where it did not print an error message:
* The "default" case in the initial switch statement. If this is not
actually reachable, use gcc_unreachable () there.
* If TYPE_SIZE_UNIT if the pointer target type is zero. This could be the
case for empty structures or zero-size arrays (both GNU extensions).
Logically there's nothing wrong with atomic operations on such a zero-size
object (they should evaluate all operands for their side effects and then
need do nothing more). But if you don't want to make them work now, there
should be a sorry () call and a PR for making them work.
In looking at that function, I also see another unrelated problem: it has
a statement "return MEMMODEL_SEQ_CST;" (i.e. return 5;) but is otherwise
meant to be returning a size not a memory model, so something doesn't make
sense there. A second unrelated problem:
void f (int n) { int a[n], b[n]; __atomic_load (&a, &b, __ATOMIC_SEQ_CST); }
ICEs in tree_low_cst. C1X doesn't require atomic operations on
variable-length types - arrays can't be _Atomic-qualified and atomic
operations are only required on atomic types - but there should be a
proper error not an ICE.
It seems a bad idea to put error_mark_node in the IR for code that has not
generated an error. So maybe you need to change get_atomic_generic_size's
interface so that errors can be unambiguously distinguished from zero-size
objects - and make the comment above that function explicitly say that on
error return it has output an error message.
Finally, the patch needs a testcase added to the testsuite for the case
reported in the PR as ICEing - and for any other problems found and fixed
in the process.
--
Joseph S. Myers
[email protected]
