On 06/28/2013 12:54 AM, Iyer, Balaji V wrote:
I agree with you and I have fixed it such that if TREE_TYPE is void then don't
bother creating a new variable.
This error happens in comma_exp.c testcase in Mac.
For example, the following expression:
array[:] = (atoi(argv[1]), (array2[0:10]+5));
In Mac, it is converting the above expression to array[:] = ((void) atoi
(argv[1]), (array2[0:10]+5))
There is a function (replace_invariant_exprs) that will go through all the
invariant expressions and replace it with a variable so that the function is
only evaluated once. In this case, there is no reason to do so and just ignore
it.
We don't need to create a variable, but if the expression has
side-effects I think we still want to move it out of the loop, right?
The Cilk+ spec doesn't seem to specify whether subexpressions of rank 0
are evaluated once per iteration, or just once total.
+ /* Sometimes, when comma_expr has a function call in it, it will
+ typecast it to void. Find_inv_trees finds those nodes and so
+ if it void type, then don't bother creating a new var to hold
+ the return value. */
+ if (VOID_TYPE_P (TREE_TYPE (t)))
+ new_var = t;
+ else
+ new_var = get_temp_regvar (TREE_TYPE (t), t);
I was suggesting
if (VOID_TYPE_P (TREE_TYPE (t)))
{
finish_expr_stmt (t);
new_var = void_zero_node;
}
Jason
