AArch64: Gate NEON use on runtime capability

Previously, the native backend and ABI checker assumed that NEON was
present on every AArch64 system.

Add MLK_SYS_CAP_AARCH64_NEON and gate every NEON-dependent native and
FIPS202 entry point on it. This allows integrators to provide runtime
capability detection and fall back to C.

Mark NEON requirements in the assembly ABI metadata and teach generated
ABI checks to skip unsupported kernels.

Route ABI checks through C call-stub wrappers. On AArch64, pass the
runtime capability into the assembly stub so it only saves, seeds,
captures, and restores vector registers when NEON is available. Keep the
GPR self-test unconditional and run vector corrupters only when NEON is
supported.

Extend the custom ID-register test configuration to detect AdvSIMD
through ID_AA64PFR0_EL1.

Signed-off-by: Hanno Becker <beckphan@amazon.co.uk>

Author: hanno-becker

dev/aarch64_clean/meta.h

@@ -28,6 +28,10 @@
 MLK_MUST_CHECK_RETURN_VALUE
 static MLK_INLINE int mlk_ntt_native(int16_t data[MLKEM_N])
 {
+  if (!mlk_sys_check_capability(MLK_SYS_CAP_NEON))
+  {
+    return MLK_NATIVE_FUNC_FALLBACK;
+  }
   mlk_ntt_aarch64_asm(data, mlk_aarch64_ntt_zetas_layer12345,
                       mlk_aarch64_ntt_zetas_layer67);
   return MLK_NATIVE_FUNC_SUCCESS;
@@ -36,6 +40,10 @@ static MLK_INLINE int mlk_ntt_native(int16_t data[MLKEM_N])
 MLK_MUST_CHECK_RETURN_VALUE
 static MLK_INLINE int mlk_intt_native(int16_t data[MLKEM_N])
 {
+  if (!mlk_sys_check_capability(MLK_SYS_CAP_NEON))
+  {
+    return MLK_NATIVE_FUNC_FALLBACK;
+  }
   mlk_intt_aarch64_asm(data, mlk_aarch64_invntt_zetas_layer12345,
                        mlk_aarch64_invntt_zetas_layer67);
   return MLK_NATIVE_FUNC_SUCCESS;
@@ -44,13 +52,21 @@ static MLK_INLINE int mlk_intt_native(int16_t data[MLKEM_N])
 MLK_MUST_CHECK_RETURN_VALUE
 static MLK_INLINE int mlk_poly_reduce_native(int16_t data[MLKEM_N])
 {
+  if (!mlk_sys_check_capability(MLK_SYS_CAP_NEON))
+  {
+    return MLK_NATIVE_FUNC_FALLBACK;
+  }
   mlk_poly_reduce_aarch64_asm(data);
   return MLK_NATIVE_FUNC_SUCCESS;
 }
 
 MLK_MUST_CHECK_RETURN_VALUE
 static MLK_INLINE int mlk_poly_tomont_native(int16_t data[MLKEM_N])
 {
+  if (!mlk_sys_check_capability(MLK_SYS_CAP_NEON))
+  {
+    return MLK_NATIVE_FUNC_FALLBACK;
+  }
   mlk_poly_tomont_aarch64_asm(data);
   return MLK_NATIVE_FUNC_SUCCESS;
 }
@@ -59,6 +75,10 @@ MLK_MUST_CHECK_RETURN_VALUE
 static MLK_INLINE int mlk_poly_mulcache_compute_native(int16_t x[MLKEM_N / 2],
                                                        const int16_t y[MLKEM_N])
 {
+  if (!mlk_sys_check_capability(MLK_SYS_CAP_NEON))
+  {
+    return MLK_NATIVE_FUNC_FALLBACK;
+  }
   mlk_poly_mulcache_compute_aarch64_asm(
       x, y, mlk_aarch64_zetas_mulcache_native,
       mlk_aarch64_zetas_mulcache_twisted_native);
@@ -71,6 +91,10 @@ static MLK_INLINE int mlk_polyvec_basemul_acc_montgomery_cached_k2_native(
     int16_t r[MLKEM_N], const int16_t a[2 * MLKEM_N],
     const int16_t b[2 * MLKEM_N], const int16_t b_cache[2 * (MLKEM_N / 2)])
 {
+  if (!mlk_sys_check_capability(MLK_SYS_CAP_NEON))
+  {
+    return MLK_NATIVE_FUNC_FALLBACK;
+  }
   mlk_polyvec_basemul_acc_montgomery_cached_k2_aarch64_asm(r, a, b, b_cache);
   return MLK_NATIVE_FUNC_SUCCESS;
 }
@@ -82,6 +106,10 @@ static MLK_INLINE int mlk_polyvec_basemul_acc_montgomery_cached_k3_native(
     int16_t r[MLKEM_N], const int16_t a[3 * MLKEM_N],
     const int16_t b[3 * MLKEM_N], const int16_t b_cache[3 * (MLKEM_N / 2)])
 {
+  if (!mlk_sys_check_capability(MLK_SYS_CAP_NEON))
+  {
+    return MLK_NATIVE_FUNC_FALLBACK;
+  }
   mlk_polyvec_basemul_acc_montgomery_cached_k3_aarch64_asm(r, a, b, b_cache);
   return MLK_NATIVE_FUNC_SUCCESS;
 }
@@ -93,6 +121,10 @@ static MLK_INLINE int mlk_polyvec_basemul_acc_montgomery_cached_k4_native(
     int16_t r[MLKEM_N], const int16_t a[4 * MLKEM_N],
     const int16_t b[4 * MLKEM_N], const int16_t b_cache[4 * (MLKEM_N / 2)])
 {
+  if (!mlk_sys_check_capability(MLK_SYS_CAP_NEON))
+  {
+    return MLK_NATIVE_FUNC_FALLBACK;
+  }
   mlk_polyvec_basemul_acc_montgomery_cached_k4_aarch64_asm(r, a, b, b_cache);
   return MLK_NATIVE_FUNC_SUCCESS;
 }
@@ -102,6 +134,10 @@ MLK_MUST_CHECK_RETURN_VALUE
 static MLK_INLINE int mlk_poly_tobytes_native(uint8_t r[MLKEM_POLYBYTES],
                                               const int16_t a[MLKEM_N])
 {
+  if (!mlk_sys_check_capability(MLK_SYS_CAP_NEON))
+  {
+    return MLK_NATIVE_FUNC_FALLBACK;
+  }
   mlk_poly_tobytes_aarch64_asm(r, a);
   return MLK_NATIVE_FUNC_SUCCESS;
 }
@@ -111,8 +147,8 @@ static MLK_INLINE int mlk_rej_uniform_native(int16_t *r, unsigned len,
                                              const uint8_t *buf,
                                              unsigned buflen)
 {
-  if (len != MLKEM_N ||
-      buflen % 24 != 0) /* NEON support is mandatory for AArch64 */
+  if (!mlk_sys_check_capability(MLK_SYS_CAP_NEON) || len != MLKEM_N ||
+      buflen % 24 != 0)
   {
     return MLK_NATIVE_FUNC_FALLBACK;
   }

dev/aarch64_clean/src/intt_aarch64_asm.S

@@ -26,6 +26,7 @@
   ABI:
     Architecture: aarch64
     CallingConvention: AAPCS64
+    Features: [NEON]
     x0:
       type: buffer
       size_bytes: 512

dev/aarch64_clean/src/ntt_aarch64_asm.S

@@ -26,6 +26,7 @@
   ABI:
     Architecture: aarch64
     CallingConvention: AAPCS64
+    Features: [NEON]
     x0:
       type: buffer
       size_bytes: 512

dev/aarch64_clean/src/poly_mulcache_compute_aarch64_asm.S

@@ -10,6 +10,7 @@
   ABI:
     Architecture: aarch64
     CallingConvention: AAPCS64
+    Features: [NEON]
     x0:
       type: buffer
       size_bytes: 256

dev/aarch64_clean/src/poly_reduce_aarch64_asm.S

@@ -10,6 +10,7 @@
   ABI:
     Architecture: aarch64
     CallingConvention: AAPCS64
+    Features: [NEON]
     x0:
       type: buffer
       size_bytes: 512

dev/aarch64_clean/src/poly_tobytes_aarch64_asm.S

@@ -10,6 +10,7 @@
   ABI:
     Architecture: aarch64
     CallingConvention: AAPCS64
+    Features: [NEON]
     x0:
       type: buffer
       size_bytes: 384

dev/aarch64_clean/src/poly_tomont_aarch64_asm.S

@@ -10,6 +10,7 @@
   ABI:
     Architecture: aarch64
     CallingConvention: AAPCS64
+    Features: [NEON]
     x0:
       type: buffer
       size_bytes: 512

dev/aarch64_clean/src/polyvec_basemul_acc_montgomery_cached_k2_aarch64_asm.S

@@ -19,6 +19,7 @@
   ABI:
     Architecture: aarch64
     CallingConvention: AAPCS64
+    Features: [NEON]
     x0:
       type: buffer
       size_bytes: 512

dev/aarch64_clean/src/polyvec_basemul_acc_montgomery_cached_k3_aarch64_asm.S

@@ -19,6 +19,7 @@
   ABI:
     Architecture: aarch64
     CallingConvention: AAPCS64
+    Features: [NEON]
     x0:
       type: buffer
       size_bytes: 512

dev/aarch64_clean/src/polyvec_basemul_acc_montgomery_cached_k4_aarch64_asm.S

@@ -19,6 +19,7 @@
   ABI:
     Architecture: aarch64
     CallingConvention: AAPCS64
+    Features: [NEON]
     x0:
       type: buffer
       size_bytes: 512