diff --git a/llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp b/llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
index 415164fc9e2cb..710662dc107c5 100644
--- a/llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
@@ -2337,6 +2337,11 @@ class BitPermutationSelector {
                         ANDIVal, ANDISVal), 0);
     }
 
+    // Caller assumes ResNo == 0, but we might have ResNo != 0 after
+    // optimizing away a permutation.  Kludge with an extra node.
+    if (Res.getResNo() != 0)
+      return CurDAG->getMachineNode(PPC::OR, dl, MVT::i32, Res, Res);
+
     return Res.getNode();
   }
 
diff --git a/llvm/test/CodeGen/PowerPC/lwzu-i48.ll b/llvm/test/CodeGen/PowerPC/lwzu-i48.ll
new file mode 100644
index 0000000000000..8205f5976e876
--- /dev/null
+++ b/llvm/test/CodeGen/PowerPC/lwzu-i48.ll
@@ -0,0 +1,19 @@
+; RUN: llc -mtriple=powerpc-unknown-openbsd < %s | FileCheck %s
+
+; BitPermutationSelector in PPCISelDAGToDAG.cpp was taking the wrong
+; result of a load <pre-inc> after optimizing away a permutation.
+; Here, the big end of i48 %3 was %1 but should be %0.
+
+define i32 @hop(ptr %out, ptr %in) {
+entry:
+  %0 = getelementptr i8, ptr %in, i32 28
+  %1 = load i32, ptr %0, align 4
+  %2 = ptrtoint ptr %0 to i48
+  %3 = shl i48 %2, 16
+  store i48 %3, ptr %out, align 4
+  ret i32 %1
+}
+; The stw should store POINTER, not VALUE.
+; CHECK:        lwzu [[VALUE:[0-9]+]], 28([[POINTER:[0-9]+]])
+; CHECK:        mr [[MOVED:[0-9]+]], [[POINTER]]
+; CHECK:        stw [[MOVED]], 0({{[0-9]+}})