[llvm-branch-commits] [clang] [llvm] [mlir] [RFC][AMDGPU] Add EXECSYNC address space (PR #195613)

[Shilei Tian via llvm-branch-commits]

shiltian wrote:

> > Clang will generate an AS cast to flat, followed by a flat GEP, followed by a cast to AS15. This is not a bug. This is just how clang lowers things.
> 
> Clang lowering it the wrong way, introducing an unnecessary cast, is a bug

No, this is not a bug, like I said earlier. It depends on how `__named_barrier__` or whatever modifier is going to be defined. It can be an attribute about address space, or a storage attribute. If it is the latter, which is same as `__shared__`, then what clang does is right, because it has nothing to do with address space, or say, the object, from clang's perspective, is still in the default address space.

https://github.com/llvm/llvm-project/pull/195613