[llvm] [AMDGPU] Correctly merge noalias scopes during lowering of LDS data. (PR #131664)

[Stanislav Mekhanoshin via llvm-commits]

================
@@ -1451,7 +1452,17 @@ class AMDGPULowerModuleLDS {
           I->setMetadata(LLVMContext::MD_alias_scope, AS);
 
           MDNode *NA = I->getMetadata(LLVMContext::MD_noalias);
-          NA = (NA ? MDNode::intersect(NA, NoAlias) : NoAlias);
+          // If domain of NoAlias (domain of LDS structure) is different
+          // than existing NA, we need to preserve exising !NoAlias
+          SmallPtrSet<const MDNode *, 16> ExistingDomains, LDSDomains;
+          ScopedNoAlias.collectScopedDomains(NA, ExistingDomains);
+          ScopedNoAlias.collectScopedDomains(NoAlias, LDSDomains);
+          auto Diff = set_difference(ExistingDomains, LDSDomains);
+          if (Diff.empty()) {
+            NA = NA ? MDNode::intersect(NA, NoAlias) : NoAlias;
+          } else {
+            NA = NA ? MDNode::concatenate(NA, NoAlias) : NoAlias;
----------------
rampitec wrote:

The code computes existing domains, but then resulting NA computation does not depend on it, it still uses the same NA and does not really use Diff or ExistingDomains. It is just replacing intersection with the concatenation, but not with the pre-existing info. What am I missing?

https://github.com/llvm/llvm-project/pull/131664