diff --git a/builtins.cpp b/builtins.cpp
index ffb05e6d..4cd2c8c1 100644
--- a/builtins.cpp
+++ b/builtins.cpp
@@ -379,7 +379,10 @@ lSetInternalFunctions(llvm::Module *module) {
         "__ceil_uniform_float",
         "__ceil_varying_double",
         "__ceil_varying_float",
-        "__count_trailing_zeros",
+        "__count_trailing_zeros_i32",
+        "__count_trailing_zeros_i64",
+        "__count_leading_zeros_i32",
+        "__count_leading_zeros_i64",
         "__do_assert_uniform",
         "__do_assert_varying",
         "__do_print", 
diff --git a/builtins.m4 b/builtins.m4
index 2b98bd80..bcfcc840 100644
--- a/builtins.m4
+++ b/builtins.m4
@@ -1094,11 +1094,26 @@ define <$1 x i32> @__sext_varying_bool(<$1 x i32>) nounwind readnone alwaysinlin
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; count trailing zeros
 
-define i32 @__count_trailing_zeros(i32) nounwind readnone alwaysinline {
+define i32 @__count_trailing_zeros_i32(i32) nounwind readnone alwaysinline {
   %c = call i32 @llvm.cttz.i32(i32 %0)
   ret i32 %c
 }
 
+define i64 @__count_trailing_zeros_i64(i64) nounwind readnone alwaysinline {
+  %c = call i64 @llvm.cttz.i64(i64 %0)
+  ret i64 %c
+}
+
+define i32 @__count_leading_zeros_i32(i32) nounwind readnone alwaysinline {
+  %c = call i32 @llvm.ctlz.i32(i32 %0)
+  ret i32 %c
+}
+
+define i64 @__count_leading_zeros_i64(i64) nounwind readnone alwaysinline {
+  %c = call i64 @llvm.ctlz.i64(i64 %0)
+  ret i64 %c
+}
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; AOS/SOA conversion primitives
 
@@ -2500,8 +2515,11 @@ done:
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;; reduce_equal
 
-; count trailing zeros
+; count leading/trailing zeros
+declare i32 @llvm.ctlz.i32(i32)
+declare i64 @llvm.ctlz.i64(i64)
 declare i32 @llvm.cttz.i32(i32)
+declare i64 @llvm.cttz.i64(i64)
 
 define(`reduce_equal_aux', `
 define i1 @__reduce_equal_$3(<$1 x $2> %v, $2 * %samevalue,
diff --git a/ctx.cpp b/ctx.cpp
index 96c44bce..29495319 100644
--- a/ctx.cpp
+++ b/ctx.cpp
@@ -2158,13 +2158,14 @@ FunctionEmitContext::CallInst(llvm::Value *func, const FunctionType *funcType,
             // Figure out the first lane that still needs its function
             // pointer to be called.
             llvm::Value *currentMask = LoadInst(maskPtr);
-            llvm::Function *cttz = m->module->getFunction("__count_trailing_zeros");
+            llvm::Function *cttz = 
+                m->module->getFunction("__count_trailing_zeros_i32");
             assert(cttz != NULL);
             llvm::Value *firstLane = CallInst(cttz, NULL, LaneMask(currentMask),
                                               "first_lane");
 
-            // Get the pointer to the function we're going to call this time through:
-            // ftpr = func[firstLane]
+            // Get the pointer to the function we're going to call this
+            // time through: ftpr = func[firstLane]
             llvm::Value *fptr = 
                 llvm::ExtractElementInst::Create(func, firstLane, 
                                                  "extract_fptr", bblock);
diff --git a/stdlib.ispc b/stdlib.ispc
index a125d9af..52326513 100644
--- a/stdlib.ispc
+++ b/stdlib.ispc
@@ -315,6 +315,113 @@ static inline uniform int lanemask() {
     return __movmsk(__mask);
 }
 
+///////////////////////////////////////////////////////////////////////////
+// count leading/trailing zeros
+
+static inline uniform unsigned int32
+count_leading_zeros(uniform unsigned int32 v) {
+    return __count_leading_zeros_i32(v);
+}
+
+static inline uniform unsigned int64
+count_leading_zeros(uniform unsigned int64 v) {
+    return __count_leading_zeros_i64(v);
+}
+
+static inline uniform unsigned int32
+count_trailing_zeros(uniform unsigned int32 v) {
+    return __count_trailing_zeros_i32(v);
+}
+
+static inline uniform unsigned int64
+count_trailing_zeros(uniform unsigned int64 v) {
+    return __count_trailing_zeros_i64(v);
+}
+
+static inline uniform int32
+count_leading_zeros(uniform int32 v) {
+    return __count_leading_zeros_i32(v);
+}
+
+static inline uniform int64
+count_leading_zeros(uniform int64 v) {
+    return __count_leading_zeros_i64(v);
+}
+
+static inline uniform int32
+count_trailing_zeros(uniform int32 v) {
+    return __count_trailing_zeros_i32(v);
+}
+
+static inline uniform int64
+count_trailing_zeros(uniform int64 v) {
+    return __count_trailing_zeros_i64(v);
+}
+
+static inline unsigned int32
+count_leading_zeros(unsigned int32 v) {
+    unsigned int32 r;
+    for (uniform int i = 0; i < programCount; ++i)
+        r = insert(r, i, __count_leading_zeros_i32(extract(v, i)));
+    return r;
+}
+
+static inline unsigned int64
+count_leading_zeros(unsigned int64 v) {
+    unsigned int64 r;
+    for (uniform int i = 0; i < programCount; ++i)
+        r = insert(r, i, __count_leading_zeros_i64(extract(v, i)));
+    return r;
+}
+
+static inline unsigned int32
+count_trailing_zeros(unsigned int32 v) {
+    unsigned int32 r;
+    for (uniform int i = 0; i < programCount; ++i)
+        r = insert(r, i, __count_trailing_zeros_i32(extract(v, i)));
+    return r;
+}
+
+static inline unsigned int64
+count_trailing_zeros(unsigned int64 v) {
+    unsigned int64 r;
+    for (uniform int i = 0; i < programCount; ++i)
+        r = insert(r, i, __count_trailing_zeros_i64(extract(v, i)));
+    return r;
+}
+
+static inline int32
+count_leading_zeros(int32 v) {
+    int32 r;
+    for (uniform int i = 0; i < programCount; ++i)
+        r = insert(r, i, __count_leading_zeros_i32(extract(v, i)));
+    return r;
+}
+
+static inline int64
+count_leading_zeros(int64 v) {
+    int64 r;
+    for (uniform int i = 0; i < programCount; ++i)
+        r = insert(r, i, __count_leading_zeros_i64(extract(v, i)));
+    return r;
+}
+
+static inline int32
+count_trailing_zeros(int32 v) {
+    int32 r;
+    for (uniform int i = 0; i < programCount; ++i)
+        r = insert(r, i, __count_trailing_zeros_i32(extract(v, i)));
+    return r;
+}
+
+static inline int64
+count_trailing_zeros(int64 v) {
+    int64 r;
+    for (uniform int i = 0; i < programCount; ++i)
+        r = insert(r, i, __count_trailing_zeros_i64(extract(v, i)));
+    return r;
+}
+
 ///////////////////////////////////////////////////////////////////////////
 // AOS/SOA conversion
 
diff --git a/tests/count-leading-trailing-zeros-1.ispc b/tests/count-leading-trailing-zeros-1.ispc
new file mode 100644
index 00000000..221d066d
--- /dev/null
+++ b/tests/count-leading-trailing-zeros-1.ispc
@@ -0,0 +1,11 @@
+
+export uniform int width() { return programCount; }
+
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    RET[programIndex] = count_trailing_zeros(0xf0);
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 4;
+}
diff --git a/tests/count-leading-trailing-zeros-2.ispc b/tests/count-leading-trailing-zeros-2.ispc
new file mode 100644
index 00000000..b94f5508
--- /dev/null
+++ b/tests/count-leading-trailing-zeros-2.ispc
@@ -0,0 +1,11 @@
+
+export uniform int width() { return programCount; }
+
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    RET[programIndex] = count_leading_zeros((int32)0xf0);
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 24;
+}
diff --git a/tests/count-leading-trailing-zeros-3.ispc b/tests/count-leading-trailing-zeros-3.ispc
new file mode 100644
index 00000000..4812efa6
--- /dev/null
+++ b/tests/count-leading-trailing-zeros-3.ispc
@@ -0,0 +1,11 @@
+
+export uniform int width() { return programCount; }
+
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    RET[programIndex] = count_leading_zeros((unsigned int64)0xf0);
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 56;
+}
diff --git a/tests/count-leading-trailing-zeros-4.ispc b/tests/count-leading-trailing-zeros-4.ispc
new file mode 100644
index 00000000..5cef2b7a
--- /dev/null
+++ b/tests/count-leading-trailing-zeros-4.ispc
@@ -0,0 +1,12 @@
+
+export uniform int width() { return programCount; }
+
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    int32 i = (1 << programIndex);
+    RET[programIndex] = count_leading_zeros(i);
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 31-programIndex;
+}
diff --git a/tests/count-leading-trailing-zeros-5.ispc b/tests/count-leading-trailing-zeros-5.ispc
new file mode 100644
index 00000000..f872d099
--- /dev/null
+++ b/tests/count-leading-trailing-zeros-5.ispc
@@ -0,0 +1,12 @@
+
+export uniform int width() { return programCount; }
+
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    unsigned int64 i = ((unsigned int64)1 << (50+programIndex));
+    RET[programIndex] = count_trailing_zeros(i);
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = 50+programIndex;
+}