diff --git a/kai/kai_common.h b/kai/kai_common.h
index dc815f679dbf369b7de8ae0184aa91b4acdd1a38..8e22fa3c7dee493f8924313fec34cfc7beba7fb7 100644
--- a/kai/kai_common.h
+++ b/kai/kai_common.h
@@ -104,7 +104,7 @@ inline static float kai_cast_f32_bf16(uint16_t bf16) {
 inline static uint16_t kai_cast_bf16_f32(float f32) {
     uint16_t bf16 = 0;
 #ifdef __ARM_FEATURE_BF16
-    asm("bfcvt %h[output], %s[input]" : [output] "=w"(bf16) : [input] "w"(f32));
+    __asm__ __volatile__("bfcvt %h[output], %s[input]" : [output] "=w"(bf16) : [input] "w"(f32));
 #else
     const uint32_t* i32 = (uint32_t*)(&f32);
     bf16 = (*i32 >> 16);
@@ -135,7 +135,7 @@ inline static size_t kai_roundup(size_t a, size_t b) {
 inline static uint64_t kai_get_sme_vector_length_u8(void) {
     uint64_t res = 0;
 
-    __asm __volatile(
+    __asm__ __volatile__(
         ".inst 0xd503477f  // SMSTART ZA\n"
         "cntb %0\n"
         ".inst 0xd503467f  // SMSTOP\n"
@@ -151,7 +151,7 @@ inline static uint64_t kai_get_sme_vector_length_u8(void) {
 inline static uint64_t kai_get_sme_vector_length_u16(void) {
     uint64_t res = 0;
 
-    __asm __volatile(
+    __asm__ __volatile__(
         ".inst 0xd503477f  // SMSTART ZA\n"
         "cnth %0\n"
         ".inst 0xd503467f  // SMSTOP\n"
@@ -167,7 +167,7 @@ inline static uint64_t kai_get_sme_vector_length_u16(void) {
 inline static uint64_t kai_get_sme_vector_length_u32(void) {
     uint64_t res = 0;
 
-    __asm __volatile(
+    __asm__ __volatile__(
         ".inst 0xd503477f  // SMSTART ZA\n"
         "cntw %0\n"
         ".inst 0xd503467f  // SMSTOP\n"
diff --git a/test/common/bfloat16.hpp b/test/common/bfloat16.hpp
index 94c362e395b21e7cff4a8f94dec9092b632f3345..ca0e0b371f6bad9f6b00ca55fd6a3e09c331171a 100644
--- a/test/common/bfloat16.hpp
+++ b/test/common/bfloat16.hpp
@@ -40,7 +40,7 @@ public:
     /// Creates a new object from the specified numeric value.
     BFloat16(float value) : _data(0) {
 #ifdef __ARM_FEATURE_BF16
-        asm("bfcvt %h[output], %s[input]" : [output] "=w"(_data) : [input] "w"(value));
+        __asm__ __volatile__("bfcvt %h[output], %s[input]" : [output] "=w"(_data) : [input] "w"(value));
 #else
         const uint32_t* value_i32 = reinterpret_cast<const uint32_t*>(&value);
         _data = (*value_i32 >> 16);
@@ -52,7 +52,7 @@ public:
     BFloat16& operator=(T value) {
         const auto value_f32 = static_cast<float>(value);
 #ifdef __ARM_FEATURE_BF16
-        asm("bfcvt %h[output], %s[input]" : [output] "=w"(_data) : [input] "w"(value_f32));
+        __asm__ __volatile__("bfcvt %h[output], %s[input]" : [output] "=w"(_data) : [input] "w"(value_f32));
 #else
         const uint32_t* value_i32 = reinterpret_cast<const uint32_t*>(&value_f32);
         _data = (*value_i32 >> 16);
diff --git a/test/common/round.cpp b/test/common/round.cpp
index 08f6c7edb52c41a6a7674838685891f9a6ab0d32..954d96da7d4a6aaf921920492eb06ee4f7ef54cf 100644
--- a/test/common/round.cpp
+++ b/test/common/round.cpp
@@ -13,7 +13,7 @@ namespace kai::test {
 
 int32_t round_to_nearest_even_i32(float value) {
     int32_t rounded = 0;
-    asm("fcvtns %w[output], %s[input]" : [output] "=r"(rounded) : [input] "w"(value));
+    __asm__ __volatile__("fcvtns %w[output], %s[input]" : [output] "=r"(rounded) : [input] "w"(value));
     return rounded;
 }
 
@@ -21,7 +21,7 @@ size_t round_to_nearest_even_usize(float value) {
     static_assert(sizeof(size_t) == sizeof(uint64_t));
 
     uint64_t rounded = 0;
-    asm("fcvtns %x[output], %s[input]" : [output] "=r"(rounded) : [input] "w"(value));
+    __asm__ __volatile__("fcvtns %x[output], %s[input]" : [output] "=r"(rounded) : [input] "w"(value));
     return rounded;
 }
 
diff --git a/test/common/sme.cpp b/test/common/sme.cpp
index b13e991014567bd5298c34ecf5b790d7a9df26d1..86b6bc356060431065679cf2fca7dee9e247262a 100644
--- a/test/common/sme.cpp
+++ b/test/common/sme.cpp
@@ -20,7 +20,7 @@ uint64_t get_sme_vector_length<1>() {
 
     if (res == 0) {
         if (cpu_has_sme()) {
-            __asm __volatile(
+            __asm__ __volatile__(
                 ".inst 0xd503477f  // SMSTART ZA\n"
                 "cntb %0\n"
                 ".inst 0xd503467f  // SMSTOP\n"
@@ -43,7 +43,7 @@ uint64_t get_sme_vector_length<2>() {
 
     if (res == 0) {
         if (cpu_has_sme()) {
-            __asm __volatile(
+            __asm__ __volatile__(
                 ".inst 0xd503477f  // SMSTART ZA\n"
                 "cnth %0\n"
                 ".inst 0xd503467f  // SMSTOP\n"
@@ -66,7 +66,7 @@ uint64_t get_sme_vector_length<4>() {
 
     if (res == 0) {
         if (cpu_has_sme()) {
-            __asm __volatile(
+            __asm__ __volatile__(
                 ".inst 0xd503477f  // SMSTART ZA\n"
                 "cntw %0\n"
                 ".inst 0xd503467f  // SMSTOP\n"