[llvm] [AMDGPU][SIMemoryLegalizer] Ensure LDS -> DMA hazards respect fences (PR #220288)

[Krzysztof Drewniak via llvm-commits]

================
@@ -8140,8 +8140,20 @@ in table :ref:`amdgpu-amdhsa-memory-model-code-sequences-gfx6-gfx9-table`.
                                                              is being released.
 
                                                          2. buffer/global/flat_atomic
-     fence        release      - singlethread *none*     *none*
+     fence        release      - singlethread *none*     1. s_waitcnt lgkmcnt(0)
                                - wavefront
+
+                                                           - Emit only on GFX9,
+                                                             before
----------------
krzysz00 wrote:

... Ok, hold on, why would that only be needed on gfx9? Can't we get this same hazard on gfx1250 - or are we requiring people who mix sync and async ops to explicitly do a workgroup fence (and then promising to not demote it)?

https://github.com/llvm/llvm-project/pull/220288