[llvm] [LLVM][NVPTX] Add NVPTX codegen support for fence.proxy.tensormap (PR #100748)

[via llvm-commits]

================
@@ -251,6 +251,34 @@ Overview:
 The '``@llvm.nvvm.barrier0()``' intrinsic emits a PTX ``bar.sync 0``
 instruction, equivalent to the ``__syncthreads()`` call in CUDA.
 
+Membar/Fences
+-------------
+
+
+'``llvm.nvvm.fence.proxy.tensormap.*``'
----------------
gonzalobg wrote:

Why is `generic` dropped from PTX (`tensormap::generic`) ?

How does the user of the intrinsic know whether it is fencing "from generic to tensormap", or in the other direction "from tensormap to generic"?

How will this be disambiguated if the opposite direction needs to be added?

https://github.com/llvm/llvm-project/pull/100748