[llvm] [AMDGPU] Fix negative immediate offset for unbuffered smem loads (PR #79553)

[Matt Arsenault via llvm-commits]

================
@@ -2096,21 +2099,53 @@ bool AMDGPUDAGToDAGISel::SelectSMRDBaseOffset(SDValue Addr, SDValue &SBase,
   }
   if (!N0 || !N1)
     return false;
+
+  bool Selected = false;
   if (SelectSMRDOffset(N1, SOffset, Offset, Imm32Only, IsBuffer)) {
     SBase = N0;
-    return true;
-  }
-  if (SelectSMRDOffset(N0, SOffset, Offset, Imm32Only, IsBuffer)) {
+    Selected = true;
+  } else if (SelectSMRDOffset(N0, SOffset, Offset, Imm32Only, IsBuffer)) {
     SBase = N1;
+    Selected = true;
+  }
+
+  if (Selected) {
----------------
arsenm wrote:

This is confusing, you're allowing the invalid selection and then later trying to adjust the offset. It would be clearer if you adjusted the initial selection to reject cases where the immediate cannot be used 

https://github.com/llvm/llvm-project/pull/79553