Hi Thomas, Thomas Schwinge wrote:
To allow me to progress with other work items, is the attached "OpenACC: Support GCC/C++-special 'default(_GCC_firstprivate)' clause [PR119692]" OK to push to trunk branch, with a few test cases added?
I assume this patch has been withdrawn. (Collateral from your follow-up comment in the PR). * * * I have to admit that I do not understand the idea of this patch, possibly except for using it with some testcases. Your patch email also does not really describe this - nor does the PR. However, IMHO, the real solution is -- a language hook. Namely, by using the existing 'omp_predetermined_mapping' langhook, modified as needed. It seems as if you concur, given what you point to this langhook in the PR. (After sending this patch/hack.) ->https://gcc.gnu.org/PR119692 * * *
(I might also suggest OpenACC 'default(firstprivate)' for standardization by the OpenACC Technical Committee.)
Speaking more from the OpenMP side than OpenACC side, but presumably applying to either: I find the default most useful - or default(none). The default settings are rather sensible while 'default(none)' avoids some gotchas. As a user, the most confusing bit is that scalars are firstprivatized (omp target, acc parallel), even if there are semantic and technical reasons that it often makes sense Hence, I am not sure how useful a 'default(firstprivate)' is. But I guess for some corner case, it is. In OpenMP, the defaultmap also permits to specify the preferred mapping per category (scalar, pointer, aggregate, allocatable - and 'all'), which IMHO is useful to avoid some confusions (how are scalars handled?) and permits finetuning. BTW: At least in OpenMP, some default features are not representable by mapping flags - like pointer mapping, which are privatized but the pointer target (pointee) is updated if already mapped. Tobias
