diff --git a/lib/Target/NVPTX/NVPTXISelLowering.cpp b/lib/Target/NVPTX/NVPTXISelLowering.cpp
index 828242db888..55caaad65b1 100644
--- a/lib/Target/NVPTX/NVPTXISelLowering.cpp
+++ b/lib/Target/NVPTX/NVPTXISelLowering.cpp
@@ -1595,7 +1595,7 @@ SDValue NVPTXTargetLowering::LowerFormalArguments(
             }
             Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
           }
-          InsIdx += VecSize;
+          InsIdx += NumElts;
         }
 
         if (NumElts > 0)
diff --git a/test/CodeGen/NVPTX/vec8.ll b/test/CodeGen/NVPTX/vec8.ll
new file mode 100644
index 00000000000..03f5cfc6cb0
--- /dev/null
+++ b/test/CodeGen/NVPTX/vec8.ll
@@ -0,0 +1,13 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
+
+target triple = "nvptx-unknown-cuda"
+
+; CHECK: .visible .func foo
+define void @foo(<8 x i8> %a, i8* %b) {
+  %t0 = extractelement <8 x i8> %a, i32 0
+; CHECK-DAG: ld.param.v4.u8
+; CHECK-DAG: ld.param.u32
+  store i8 %t0, i8* %b
+  ret void
+}
+