diff --git a/llvm/lib/Target/ARM/ARMInstrInfo.td b/llvm/lib/Target/ARM/ARMInstrInfo.td
--- a/llvm/lib/Target/ARM/ARMInstrInfo.td
+++ b/llvm/lib/Target/ARM/ARMInstrInfo.td
@@ -487,8 +487,12 @@
 }]>;
 def DSubReg_i32_reg : SDNodeXForm<imm, [{
   assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
-  return CurDAG->getTargetConstant(ARM::dsub_0 + N->getZExtValue()/2, SDLoc(N),
-                                   MVT::i32);
+  if (CurDAG->getDataLayout().isBigEndian())
+    return CurDAG->getTargetConstant(ARM::dsub_1 - N->getZExtValue()/2, SDLoc(N),
+                                    MVT::i32);
+  else
+    return CurDAG->getTargetConstant(ARM::dsub_0 + N->getZExtValue()/2, SDLoc(N),
+                                    MVT::i32);
 }]>;
 def DSubReg_f64_reg : SDNodeXForm<imm, [{
   assert(ARM::dsub_7 == ARM::dsub_0+7 && "Unexpected subreg numbering");
diff --git a/llvm/test/CodeGen/ARM/legalize-bitcast.ll b/llvm/test/CodeGen/ARM/legalize-bitcast.ll
--- a/llvm/test/CodeGen/ARM/legalize-bitcast.ll
+++ b/llvm/test/CodeGen/ARM/legalize-bitcast.ll
@@ -23,7 +23,7 @@
 ; CHECK-NEXT:  .LBB0_1: @ %bb.1
 ; CHECK-NEXT:    vldmia sp, {d16, d17} @ 16-byte Reload
 ; CHECK-NEXT:    vrev32.16 q8, q8
-; CHECK-NEXT:    vmov.f64 d16, d17
+; CHECK-NEXT:    @ kill: def $d16 killed $d16 killed $q8
 ; CHECK-NEXT:    vmov.32 r0, d16[0]
 ; CHECK-NEXT:    add sp, sp, #28
 ; CHECK-NEXT:    pop {r4}
diff --git a/llvm/test/CodeGen/ARM/vget_lane-be.ll b/llvm/test/CodeGen/ARM/vget_lane-be.ll
new file mode 100644
--- /dev/null
+++ b/llvm/test/CodeGen/ARM/vget_lane-be.ll
@@ -0,0 +1,51 @@
+; RUN: llc < %s -mattr=+neon | FileCheck %s
+target triple = "armebv8a-arm-none-eabihf"
+
+define i32 @ele0(<4 x i32> %a) {
+entry:
+;CHECK-LABEL: ele0:
+;CHECK-NEXT: .fnstart
+;CHECK-NEXT: @ %bb.0:
+;CHECK-NEXT: vrev64.32 q8, q0
+;CHECK-NEXT: vmov.32   r0, d16[1]
+;CHECK-NEXT: bx        lr
+
+  %vget_lane = extractelement <4 x i32> %a, i64 3
+  ret i32 %vget_lane
+}
+
+define i32 @ele1(<4 x i32> %a) {
+entry:
+;CHECK-LABEL: ele1:
+;CHECK-NEXT: .fnstart
+;CHECK-NEXT: @ %bb.0:
+;CHECK-NEXT: vrev64.32 q8, q0
+;CHECK-NEXT: vmov.32   r0, d16[0]
+;CHECK-NEXT: bx        lr
+  %vget_lane = extractelement <4 x i32> %a, i64 2
+  ret i32 %vget_lane
+}
+
+define i32 @ele2(<4 x i32> %a) {
+entry:
+;CHECK-LABEL: ele2:
+;CHECK-NEXT: .fnstart
+;CHECK-NEXT: @ %bb.0:
+;CHECK-NEXT: vrev64.32 q8, q0
+;CHECK-NEXT: vmov.32   r0, d17[1]
+;CHECK-NEXT: bx        lr
+  %vget_lane = extractelement <4 x i32> %a, i64 1
+  ret i32 %vget_lane
+}
+
+define i32 @ele3(<4 x i32> %a) {
+entry:
+;CHECK-LABEL: ele3:
+;CHECK-NEXT: .fnstart
+;CHECK-NEXT: @ %bb.0:
+;CHECK-NEXT: vrev64.32 q8, q0
+;CHECK-NEXT: vmov.32   r0, d17[0]
+;CHECK-NEXT: bx        lr
+  %vget_lane = extractelement <4 x i32> %a, i64 0
+  ret i32 %vget_lane
+}