https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70289
--- Comment #4 from cesar at gcc dot gnu.org --- Created attachment 38122 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38122&action=edit c test case This test cases is failing for a couple of reasons. First, the reduction variable doesn't have a data clause associated with it. Usually that isn't a problem, like it would be in c with stack-based storage, but fortran uses static allocation by default for local variables. Consequently, that static reduction variable resides in a different lto partition from the offloaded function. And that causes assertion failure in input_varpool_node. I'm not sure if we should create a static copy of that variable on the accelerator. My inclination would be to teach the gimplifier to add an present_or_copy clause for each reduction variable listed in a parallel directive. The OpenACC spec is vague on what it means by "original variable", but one could argue that it's the variable that resides on the host. The c test case reproduces the fortran problem in c using a static int reduction variable. A global reduction variable that is not associated with any data clauses will also fail in a similar fashion.
