diff --git a/builtins/target-nvptx.ll b/builtins/target-nvptx.ll
index 9151adf8..1e8d0ae5 100644
--- a/builtins/target-nvptx.ll
+++ b/builtins/target-nvptx.ll
@@ -722,9 +722,16 @@ svml_stubs(double,d,WIDTH)
 
 define  i64 @__movmsk(<1 x i1>) nounwind readnone alwaysinline {
   %v = extractelement <1 x i1> %0, i32 0
-  %call = call i32 @__ballot_nvptx(i1 zeroext %v)
-  %v64 = zext i32 %call to i64
-  ret i64 %v64
+;; if 0
+  ;; this one fails with ./tests/popcnt-4.ispc and others ...
+;; %v0  = call i32 @__ballot_nvptx(i1 %v)
+;; %v64 = zext i32 %v0 to i64
+
+;; else 
+   ;; this one just copies mask  
+    %v64 = zext i1 %v to i64
+;; endif 
+    ret i64 %v64
 }
 
 define  i1 @__any(<1 x i1>) nounwind readnone alwaysinline {
diff --git a/builtins/util-nvptx.m4 b/builtins/util-nvptx.m4
index 04c63758..93870f15 100644
--- a/builtins/util-nvptx.m4
+++ b/builtins/util-nvptx.m4
@@ -2693,44 +2693,10 @@ declare i8* @malloc(i32)
 declare i32 @posix_memalign(i8**, i32, i32)
 declare void @free(i8 *)
 
-define noalias i8 * @__new_uniform_32rt(i64 %size) {
-  %ptr = alloca i8*
-  %conv = trunc i64 %size to i32
-  %alignment = load i32* @memory_alignment
-  %call1 = call i32 @posix_memalign(i8** %ptr, i32 %alignment, i32 %conv)
-  %ptr_val = load i8** %ptr
-  ret i8* %ptr_val
-}
-
-define <WIDTH x i64> @__new_varying32_32rt(<WIDTH x i32> %size, <WIDTH x MASK> %mask) {
-  %ret = alloca <WIDTH x i64>
-  store <WIDTH x i64> zeroinitializer, <WIDTH x i64> * %ret
-  %ret64 = bitcast <WIDTH x i64> * %ret to i64 *
-  %alignment = load i32* @memory_alignment
-
-  per_lane(WIDTH, <WIDTH x MASK> %mask, `
-    %sz_LANE_ID = extractelement <WIDTH x i32> %size, i32 LANE
-    %store_LANE_ID = getelementptr i64 * %ret64, i32 LANE
-    %ptr_LANE_ID = bitcast i64* %store_LANE_ID to i8**
-    %call_LANE_ID = call i32 @posix_memalign(i8** %ptr_LANE_ID, i32 %alignment, i32 %sz_LANE_ID)')
-
-  %r = load <WIDTH x i64> * %ret
-  ret <WIDTH x i64> %r
-}
-
-define void @__delete_uniform_32rt(i8 * %ptr) {
-  call void @free(i8 * %ptr)
-  ret void
-}
-
-define void @__delete_varying_32rt(<WIDTH x i64> %ptr, <WIDTH x MASK> %mask) {
-  per_lane(WIDTH, <WIDTH x MASK> %mask, `
-      %iptr_LANE_ID = extractelement <WIDTH x i64> %ptr, i32 LANE
-      %ptr_LANE_ID = inttoptr i64 %iptr_LANE_ID to i8 *
-      call void @free(i8 * %ptr_LANE_ID)
-  ')
-  ret void
-}
+declare noalias i8 * @__new_uniform_32rt(i64 %size);
+declare <WIDTH x i64> @__new_varying32_32rt(<WIDTH x i32> %size, <WIDTH x MASK> %mask);
+declare void @__delete_uniform_32rt(i8 * %ptr);
+declare void @__delete_varying_32rt(<WIDTH x i64> %ptr, <WIDTH x MASK> %mask);
 
 ',
 RUNTIME, `64',