diff --git a/builtins/target-nvptx.ll b/builtins/target-nvptx.ll
index 5334f882..4dab86c4 100644
--- a/builtins/target-nvptx.ll
+++ b/builtins/target-nvptx.ll
@@ -1044,6 +1044,24 @@ define  i1 @__reduce_equal_double(<1 x double> %vv, double * %samevalue,
 
 }
 
+;;;;;;;;;;; shuffle
+define(`shuffle1', `
+define <1 x $1> @__shuffle_$1(<1 x $1>, <1 x i32>) nounwind readnone alwaysinline 
+{
+  %val  = extractelement <1 x $1> %0, i32 0
+  %lane = extractelement <1 x i32> %1, i32 0
+  %rets = tail call $1 @__shfl_$1_nvptx($1 %val, i32 %lane)
+  %retv = insertelement <1 x $1> undef, $1 %rets, i32 0
+  ret <1 x $1> %retv
+}
+')
+shuffle1(i8)
+shuffle1(i16)
+shuffle1(i32)
+shuffle1(i64)
+shuffle1(float)
+shuffle1(double)
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; unaligned loads/loads+broadcasts
 
diff --git a/builtins/util-nvptx.m4 b/builtins/util-nvptx.m4
index c948cfc3..76fc7f2b 100644
--- a/builtins/util-nvptx.m4
+++ b/builtins/util-nvptx.m4
@@ -815,18 +815,6 @@ define <WIDTH x $1> @__shift_$1(<WIDTH x $1>, i32) nounwind readnone alwaysinlin
 }
 
 
-define <WIDTH x $1> @__shuffle_$1(<WIDTH x $1>, <WIDTH x i32>) nounwind readnone alwaysinline {
-forloop(i, 0, eval(WIDTH-1), `  
-  %index_`'i = extractelement <WIDTH x i32> %1, i32 i')
-forloop(i, 0, eval(WIDTH-1), `  
-  %v_`'i = extractelement <WIDTH x $1> %0, i32 %index_`'i')
-
-  %ret_0 = insertelement <WIDTH x $1> undef, $1 %v_0, i32 0
-forloop(i, 1, eval(WIDTH-1), `  %ret_`'i = insertelement <WIDTH x $1> %ret_`'eval(i-1), $1 %v_`'i, i32 i
-')
-  ret <WIDTH x $1> %ret_`'eval(WIDTH-1)
-}
-
 define <WIDTH x $1> @__shuffle2_$1(<WIDTH x $1>, <WIDTH x $1>, <WIDTH x i32>) nounwind readnone alwaysinline {
   %v2 = shufflevector <WIDTH x $1> %0, <WIDTH x $1> %1, <eval(2*WIDTH) x i32> <
       forloop(i, 0, eval(2*WIDTH-2), `i32 i, ') i32 eval(2*WIDTH-1)