diff --git a/llvm/lib/Target/AMDGPU/AMDGPUExportClustering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUExportClustering.cpp
--- a/llvm/lib/Target/AMDGPU/AMDGPUExportClustering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUExportClustering.cpp
@@ -81,6 +81,32 @@
   }
 }
 
+static void removeExportDependencies(ScheduleDAGInstrs *DAG, SUnit &SU) {
+  SmallVector<SDep, 2> ToAdd, ToRemove;
+
+  for (const SDep &Pred : SU.Preds) {
+    SUnit *PredSU = Pred.getSUnit();
+    if (Pred.isBarrier() && isExport(*PredSU)) {
+      ToRemove.push_back(Pred);
+      if (isExport(SU))
+        continue;
+
+      // If we remove a barrier we need to copy dependencies
+      // from the predecessor to maintain order.
+      for (const SDep &ExportPred : PredSU->Preds) {
+        SUnit *ExportPredSU = ExportPred.getSUnit();
+        if (ExportPred.isBarrier() && !isExport(*ExportPredSU))
+          ToAdd.push_back(SDep(ExportPredSU, SDep::Barrier));
+      }
+    }
+  }
+
+  for (SDep Pred : ToRemove)
+    SU.removePred(Pred);
+  for (SDep Pred : ToAdd)
+    DAG->addEdge(&SU, Pred);
+}
+
 void ExportClustering::apply(ScheduleDAGInstrs *DAG) {
   const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(DAG->TII);
 
@@ -97,15 +123,7 @@
       if (isPositionExport(TII, &SU))
         PosCount++;
     }
-
-    SmallVector<SDep, 2> ToRemove;
-    for (const SDep &Pred : SU.Preds) {
-      SUnit *PredSU = Pred.getSUnit();
-      if (Pred.isBarrier() && isExport(*PredSU))
-        ToRemove.push_back(Pred);
-    }
-    for (SDep Pred : ToRemove)
-      SU.removePred(Pred);
+    removeExportDependencies(DAG, SU);
   }
 
   // Apply clustering if there are multiple exports