[llvm] [SystemZ] Simplify f128 atomic load/store (PR #90977)

[Ulrich Weigand via llvm-commits]

================
@@ -6829,47 +6803,87 @@ SDValue SystemZTargetLowering::combineMERGE(
   return SDValue();
 }
 
+static bool
+isI128MovedToParts(LoadSDNode *LD,
+                   SmallVector<std::pair<SDNode *, int>, 2> &Users) {
+  // Scan through all users.
+  for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
----------------
uweigand wrote:

Good point.  I've changed the logic to ensure early the loop exits early in case of more than two users.

https://github.com/llvm/llvm-project/pull/90977