[llvm] [AMDGPU] Fix LDS access via flat pointer argument in amdgpu-sw-lower-lds (PR #209842)

[via llvm-commits]

================
@@ -1320,6 +1419,20 @@ bool AMDGPUSwLowerLDS::run() {
     Changed = true;
   }
 
+  // Lower any remaining non-kernel that reaches lowered LDS through a flat
+  // round trip (addrspacecast flat->local). Functions handled above already had
+  // these accesses translated there, so skip them. No detection is needed:
+  // collapsing the round trip is valid regardless of the pointer's origin.
+  for (Function &F : M) {
+    if (F.isDeclaration() || AMDGPU::isKernel(F))
+      continue;
+    if (FuncLDSAccessInfo.NonKernelToLDSAccessMap.contains(&F) ||
+        FuncLDSAccessInfo.NonKernelsWithLDSArgument.contains(&F))
+      continue;
+    if (lowerNonKernelLDSFlatArgAccesses(&F))
+      Changed = true;
+  }
+
----------------
skc7 wrote:

Yes, this PR only fixes the bug, a helper function gets LDS as a flat pointer, casts it back to addrspace(3), and ends up reading dead LDS after relocation. This fix is what unblocks the rocPRIM (ROCM-24185) failure now. I'm happy to do the redesign, if necessary, as a follow-up.

https://github.com/llvm/llvm-project/pull/209842