[llvm] [AMDGPU][AtomicOptimizer] Fix DT update for divergent values with Iterative strategy (PR #87605)

[Jay Foad via llvm-commits]

================
@@ -887,10 +887,18 @@ void AMDGPUAtomicOptimizerImpl::optimizeAtomic(Instruction &I,
     B.CreateBr(ComputeLoop);
 
     // Update the dominator tree for new control flow.
-    DTU.applyUpdates(
+    SmallVector<DominatorTree::UpdateType, 6> DomTreeUpdates(
         {{DominatorTree::Insert, EntryBB, ComputeLoop},
-         {DominatorTree::Insert, ComputeLoop, ComputeEnd},
-         {DominatorTree::Delete, EntryBB, SingleLaneTerminator->getParent()}});
+         {DominatorTree::Insert, ComputeLoop, ComputeEnd}});
+
+    // We're moving the terminator from EntryBB to ComputeEnd, make sure we move
+    // the DT edges as well.
+    for (auto *Succ : Terminator->successors()) {
----------------
jayfoad wrote:

Nit: could you use `ComputeEnd->successors()` here? Then you would not have to explain how `Terminator` is known to be a `BranchInst`.

https://github.com/llvm/llvm-project/pull/87605