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

[via llvm-commits]

================
@@ -790,6 +792,80 @@ void AMDGPUSwLowerLDS::translateLDSMemoryOperationsToGlobalMemory(
   }
 }
 
+void AMDGPUSwLowerLDS::lowerFlatToLocalRoundTrips(Function *Func) {
+  // After SW LDS lowering, storage that used to live in LDS now lives in global
+  // memory, but it may still be handed to a callee as a flat (generic) pointer.
+  // Such a callee can re-derive an LDS pointer with a round-trip pattern:
+  //   %p = getelementptr ..., ptr %flat_arg        ; flat, backed by global mem
+  //   %c = addrspacecast ptr %p to ptr addrspace(3)
+  //   load/store ... ptr addrspace(3) %c
+  // The addrspace(3) access would then hit dead LDS. Redo the access through
+  // the flat source pointer, which still points at the global backing buffer.
+  SmallVector<Instruction *> ToErase;
+  SmallPtrSet<AddrSpaceCastInst *, 8> Casts;
+  for (BasicBlock &BB : *Func) {
+    for (Instruction &Inst : BB) {
+      Value *Ptr = nullptr;
+      if (auto *LI = dyn_cast<LoadInst>(&Inst))
----------------
skc7 wrote:

Folded the logic into `translateLDSMemoryOperationsToGlobalMemory` path. The memory-op dispatch remains in the existing lowering path, and the new part is only pointer translation via `getFlatPtrForRoundTripLDSAccess`.

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