[Openmp-commits] [PATCH] D105121: [OpenMP] Avoid checking parent reference count in targetDataBegin

[Joel E. Denny via Phabricator via Openmp-commits]

jdenny added a comment.

In D105121#2867284 <https://reviews.llvm.org/D105121#2867284>, @grokos wrote:

> Thinking about it a bit more, no matter what the answer to my last question is, even the existing version of the library (without D105121 <https://reviews.llvm.org/D105121> and D104924 <https://reviews.llvm.org/D104924>) will copy the attached object to the device. So these patches result in logically equivalent behavior.

That's my understanding as well: because `IsNew=true` for `s1.p[0:10]` in your example even without D105121 <https://reviews.llvm.org/D105121> and D104924 <https://reviews.llvm.org/D104924>, the parent reference count wouldn't be checked anyway.

@RaviNarayanaswamy expressed some concern about a similar case with Fortran allocatables, but I don't know what map flags would produce the desired behavior then.  I believe the current feeling is that case can be checked after my patches land.

> @jdenny I think you can proceed with committing them as patches that simplify the code. If we need to take extra care of the corner case in the above example, then we can prepare a new patch that fixes the bug.

Thanks for the review.  I'll try to land them today.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D105121/new/

https://reviews.llvm.org/D105121