Index: lib/Target/PowerPC/PPCISelLowering.cpp
===================================================================
--- lib/Target/PowerPC/PPCISelLowering.cpp
+++ lib/Target/PowerPC/PPCISelLowering.cpp
@@ -11353,9 +11353,17 @@
       if (BSwapOp.getValueType() == MVT::i16)
         BSwapOp = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, BSwapOp);
 
+      // If the type of BSWAP operand is wider than stored memory width
+      // it need to be shifted to the right side before STBRX.
+      EVT mVT = cast<StoreSDNode>(N)->getMemoryVT();
+      if (Op1VT.bitsGT(mVT)) {
+        int shift = Op1VT.getSizeInBits() - mVT.getSizeInBits();
+        BSwapOp = DAG.getNode(ISD::SRL, dl, Op1VT, BSwapOp,
+                              DAG.getConstant(shift, dl, MVT::i32));
+      }
+
       SDValue Ops[] = {
-        N->getOperand(0), BSwapOp, N->getOperand(2),
-        DAG.getValueType(N->getOperand(1).getValueType())
+        N->getOperand(0), BSwapOp, N->getOperand(2), DAG.getValueType(mVT)
       };
       return
         DAG.getMemIntrinsicNode(PPCISD::STBRX, dl, DAG.getVTList(MVT::Other),
Index: test/CodeGen/PowerPC/pr32063.ll
===================================================================
--- test/CodeGen/PowerPC/pr32063.ll
+++ test/CodeGen/PowerPC/pr32063.ll
@@ -0,0 +1,16 @@
+; RUN: llc -O2 < %s | FileCheck %s
+target triple = "powerpc64le-linux-gnu"
+
+define void @foo(i32 %v, i16* %p) {
+        %1 = and i32 %v, -65536
+        %2 = tail call i32 @llvm.bswap.i32(i32 %1)
+        %conv = trunc i32 %2 to i16
+        store i16 %conv, i16* %p
+        ret void
+
+; CHECK:     srwi
+; CHECK:     sthbrx
+; CHECK-NOT: stwbrx
+}
+
+declare i32 @llvm.bswap.i32(i32)