[llvm] [AMDGPU] Fix spurious NoAlias results (PR #122309)

[Nikita Popov via llvm-commits]

================
@@ -80,10 +80,13 @@ AliasResult AMDGPUAAResult::alias(const MemoryLocation &LocA,
     } else if (const Argument *Arg = dyn_cast<Argument>(ObjA)) {
       const Function *F = Arg->getParent();
       switch (F->getCallingConv()) {
-      case CallingConv::AMDGPU_KERNEL:
+      case CallingConv::AMDGPU_KERNEL: {
         // In the kernel function, kernel arguments won't alias to (local)
         // variables in shared or private address space.
-        return AliasResult::NoAlias;
+        const auto *ObjB =
+            getUnderlyingObject(B.Ptr->stripPointerCastsForAliasAnalysis());
+        return (ObjA == ObjB) ? AliasResult::MustAlias : AliasResult::NoAlias;
----------------
nikic wrote:

> My understanding of the intent of the code is that a kernel argument can never alias with a "shared" or "private" address-space-qualified pointer that isn't based on a kernel argument (that's what (local) hints at in the comment). Perhaps this is always true for private pointers, as they can never derive from a kernel argument. But proving that a shared pointer is local to a function is going to be tricky; that's why I was trying to use underlying objects.

If that's the goal, I think the check here should be `isIdentifiedObject(ObjB) && ObjA != ObjB ? NoAlias : MayAlias`.

If it's not an identified object, you do not know whether it may be based on the kernel argument in some indirect way (e.g. via being stored and then loaded).

https://github.com/llvm/llvm-project/pull/122309