diff --git a/builtins/target-nvptx.ll b/builtins/target-nvptx.ll
index e93c65fe..a2fdacd6 100644
--- a/builtins/target-nvptx.ll
+++ b/builtins/target-nvptx.ll
@@ -1125,6 +1125,22 @@ rotate(i64)
 rotate(float)
 rotate(double)
 
+define(`broadcast', `
+define <1 x $1> @__broadcast_$1(<1 x $1>, i32) nounwind readnone alwaysinline 
+{
+  %val  = extractelement <1 x $1> %0, i32 0
+  %rets = tail call $1 @__shfl_$1_nvptx($1 %val, i32 %1)
+  %retv = insertelement <1 x $1> undef, $1 %rets, i32 0
+  ret <1 x $1> %retv
+}
+')
+broadcast(i8)
+broadcast(i16)
+broadcast(i32)
+broadcast(i64)
+broadcast(float)
+broadcast(double)
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; unaligned loads/loads+broadcasts
 
diff --git a/builtins/util-nvptx.m4 b/builtins/util-nvptx.m4
index aa3d4e82..ef7ce3f8 100644
--- a/builtins/util-nvptx.m4
+++ b/builtins/util-nvptx.m4
@@ -757,13 +757,6 @@ divert`'dnl
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 define(`shuffles', `
-define <WIDTH x $1> @__broadcast_$1(<WIDTH x $1>, i32) nounwind readnone alwaysinline {
-  %v = extractelement <WIDTH x $1> %0, i32 %1
-  %broadcast_init = insertelement <WIDTH x $1> undef, $1 %v, i32 0
-  %broadcast = shufflevector <WIDTH x $1> %broadcast_init, <WIDTH x $1> undef, <WIDTH x i32> zeroinitializer 
-  ret <WIDTH x $1> %broadcast
-}
-
 ')
 
 define(`define_shuffles',`