diff --git a/examples/intrinsics/generic-16.h b/examples/intrinsics/generic-16.h
index e2426f7f..3ac77a41 100644
--- a/examples/intrinsics/generic-16.h
+++ b/examples/intrinsics/generic-16.h
@@ -938,7 +938,7 @@ REDUCE_MINMAX(uint64_t, __vec16_i64, __reduce_max_uint64, >)
 ///////////////////////////////////////////////////////////////////////////
 // masked load/store
 
-static FORCEINLINE __vec16_i8 __masked_load_8(unsigned char *p,
+static FORCEINLINE __vec16_i8 __masked_load_8(void *p,
                                               __vec16_i1 mask) {
     __vec16_i8 ret;
     int8_t *ptr = (int8_t *)p;
@@ -948,7 +948,7 @@ static FORCEINLINE __vec16_i8 __masked_load_8(unsigned char *p,
     return ret;
 }
 
-static FORCEINLINE __vec16_i16 __masked_load_16(unsigned char *p,
+static FORCEINLINE __vec16_i16 __masked_load_16(void *p,
                                                 __vec16_i1 mask) {
     __vec16_i16 ret;
     int16_t *ptr = (int16_t *)p;
@@ -958,7 +958,7 @@ static FORCEINLINE __vec16_i16 __masked_load_16(unsigned char *p,
     return ret;
 }
 
-static FORCEINLINE __vec16_i32 __masked_load_32(unsigned char *p,
+static FORCEINLINE __vec16_i32 __masked_load_32(void *p,
                                                 __vec16_i1 mask) {
     __vec16_i32 ret;
     int32_t *ptr = (int32_t *)p;
@@ -968,7 +968,7 @@ static FORCEINLINE __vec16_i32 __masked_load_32(unsigned char *p,
     return ret;
 }
 
-static FORCEINLINE __vec16_i64 __masked_load_64(unsigned char *p,
+static FORCEINLINE __vec16_i64 __masked_load_64(void *p,
                                                 __vec16_i1 mask) {
     __vec16_i64 ret;
     int64_t *ptr = (int64_t *)p;
@@ -978,7 +978,7 @@ static FORCEINLINE __vec16_i64 __masked_load_64(unsigned char *p,
     return ret;
 }
 
-static FORCEINLINE void __masked_store_8(unsigned char *p, __vec16_i8 val,
+static FORCEINLINE void __masked_store_8(void *p, __vec16_i8 val,
                                          __vec16_i1 mask) {
     int8_t *ptr = (int8_t *)p;
     for (int i = 0; i < 16; ++i)
@@ -986,7 +986,7 @@ static FORCEINLINE void __masked_store_8(unsigned char *p, __vec16_i8 val,
             ptr[i] = val.v[i];
 }
 
-static FORCEINLINE void __masked_store_16(unsigned char *p, __vec16_i16 val,
+static FORCEINLINE void __masked_store_16(void *p, __vec16_i16 val,
                                           __vec16_i1 mask) {
     int16_t *ptr = (int16_t *)p;
     for (int i = 0; i < 16; ++i)
@@ -994,7 +994,7 @@ static FORCEINLINE void __masked_store_16(unsigned char *p, __vec16_i16 val,
             ptr[i] = val.v[i];
 }
 
-static FORCEINLINE void __masked_store_32(unsigned char *p, __vec16_i32 val,
+static FORCEINLINE void __masked_store_32(void *p, __vec16_i32 val,
                                           __vec16_i1 mask) {
     int32_t *ptr = (int32_t *)p;
     for (int i = 0; i < 16; ++i)
@@ -1002,7 +1002,7 @@ static FORCEINLINE void __masked_store_32(unsigned char *p, __vec16_i32 val,
             ptr[i] = val.v[i];
 }
 
-static FORCEINLINE void __masked_store_64(unsigned char *p, __vec16_i64 val,
+static FORCEINLINE void __masked_store_64(void *p, __vec16_i64 val,
                                           __vec16_i1 mask) {
     int64_t *ptr = (int64_t *)p;
     for (int i = 0; i < 16; ++i)
diff --git a/examples/intrinsics/sse4.h b/examples/intrinsics/sse4.h
index 0662a4c0..5e5131f3 100644
--- a/examples/intrinsics/sse4.h
+++ b/examples/intrinsics/sse4.h
@@ -2328,7 +2328,7 @@ static FORCEINLINE uint64_t __reduce_max_uint64(__vec4_i64 v) {
 ///////////////////////////////////////////////////////////////////////////
 // masked load/store
 
-static FORCEINLINE __vec4_i8 __masked_load_8(unsigned char *p, 
+static FORCEINLINE __vec4_i8 __masked_load_8(void *p, 
                                              __vec4_i1 mask) {
     int8_t r[4];
     int8_t *ptr = (int8_t *)p;
@@ -2348,7 +2348,7 @@ static FORCEINLINE __vec4_i8 __masked_load_8(unsigned char *p,
     return __vec4_i8(r[0], r[1], r[2], r[3]);
 }
 
-static FORCEINLINE __vec4_i16 __masked_load_16(unsigned char *p, 
+static FORCEINLINE __vec4_i16 __masked_load_16(void *p, 
                                                __vec4_i1 mask) {
     int16_t r[4];
     int16_t *ptr = (int16_t *)p;
@@ -2372,7 +2372,7 @@ static FORCEINLINE __vec4_i16 __masked_load_16(unsigned char *p,
     return __vec4_i16(r[0], r[1], r[2], r[3]);
 }
 
-static FORCEINLINE __vec4_i32 __masked_load_32(unsigned char *p, 
+static FORCEINLINE __vec4_i32 __masked_load_32(void *p, 
                                                __vec4_i1 mask) {
     __m128i r = _mm_set_epi32(0, 0, 0, 0);
     int32_t *ptr = (int32_t *)p;
@@ -2395,7 +2395,7 @@ static FORCEINLINE __vec4_i32 __masked_load_32(unsigned char *p,
     return r;
 }
 
-static FORCEINLINE __vec4_i64 __masked_load_64(unsigned char *p, 
+static FORCEINLINE __vec4_i64 __masked_load_64(void *p, 
                                                __vec4_i1 mask) {
     uint64_t r[4];
     uint64_t *ptr = (uint64_t *)p;
@@ -2418,7 +2418,7 @@ static FORCEINLINE __vec4_i64 __masked_load_64(unsigned char *p,
     return __vec4_i64(r[0], r[1], r[2], r[3]);
 }
 
-static FORCEINLINE void __masked_store_8(unsigned char *p, __vec4_i8 val, 
+static FORCEINLINE void __masked_store_8(void *p, __vec4_i8 val, 
                                          __vec4_i1 mask) {
     int8_t *ptr = (int8_t *)p;
 
@@ -2439,7 +2439,7 @@ static FORCEINLINE void __masked_store_8(unsigned char *p, __vec4_i8 val,
         ptr[3] = _mm_extract_epi8(val.v, 3);
 }
 
-static FORCEINLINE void __masked_store_16(unsigned char *p, __vec4_i16 val, __vec4_i1 mask) {
+static FORCEINLINE void __masked_store_16(void *p, __vec4_i16 val, __vec4_i1 mask) {
     int16_t *ptr = (int16_t *)p;
 
     uint32_t m = _mm_extract_ps(mask.v, 0);
@@ -2459,7 +2459,7 @@ static FORCEINLINE void __masked_store_16(unsigned char *p, __vec4_i16 val, __ve
         ptr[3] = _mm_extract_epi16(val.v, 3);
 }
 
-static FORCEINLINE void __masked_store_32(unsigned char *p, __vec4_i32 val, 
+static FORCEINLINE void __masked_store_32(void *p, __vec4_i32 val, 
                                           __vec4_i1 mask) {
     int32_t *ptr = (int32_t *)p;
     uint32_t m = _mm_extract_ps(mask.v, 0);
@@ -2479,7 +2479,7 @@ static FORCEINLINE void __masked_store_32(unsigned char *p, __vec4_i32 val,
         ptr[3] = _mm_extract_epi32(val.v, 3);
 }
 
-static FORCEINLINE void __masked_store_64(unsigned char *p, __vec4_i64 val, 
+static FORCEINLINE void __masked_store_64(void *p, __vec4_i64 val, 
                                           __vec4_i1 mask) {
     int64_t *ptr = (int64_t *)p;
     uint32_t m = _mm_extract_ps(mask.v, 0);