AArch64: Add compile-time guard for Neon

Previously, we assumed that every AArch64 system has Neon, and
that the implementation has access to it.

With this commit, use of Neon is explicitly guarded by __ARM_NEON,
similar to how all use of AVX2 is guarded by __AVX2__.

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

Author: hanno-becker

.github/workflows/ci.yml

@@ -34,6 +34,13 @@ jobs:
            arch: aarch64
            mode: native
            nix_shell: ci
+         - runner: ubuntu-24.04-arm
+           name: 'aarch64, no Neon'
+           arch: aarch64
+           mode: native
+           nix_shell: ci
+           cflags: '-march=armv8-a+nosimd'
+           extra_args: '--no-auto'
          - runner: ubuntu-24.04-arm
            name: 'x86_64, cross'
            arch: x86_64
@@ -112,6 +119,7 @@ jobs:
           nix-cache: ${{ matrix.target.mode == 'native' && 'false' || 'true' }}
           gh_token: ${{ secrets.GITHUB_TOKEN }}
           compile_mode: ${{ matrix.target.mode }}
+          cflags: ${{ matrix.target.cflags || '' }}
           extra_args: ${{ matrix.target.extra_args || '' }}
           exec_wrapper: ${{ matrix.target.exec_wrapper || '' }}
           opt: 'no_opt'
@@ -122,8 +130,9 @@ jobs:
           nix-shell: ${{ matrix.target.nix_shell }}
           gh_token: ${{ secrets.GITHUB_TOKEN }}
           compile_mode: native
-          cflags: "-DMLDSA_DEBUG -fsanitize=address -fsanitize=undefined -fno-sanitize-recover=all"
+          cflags: "${{ matrix.target.cflags || '' }} -DMLDSA_DEBUG -fsanitize=address -fsanitize=undefined -fno-sanitize-recover=all"
           ldflags: "-fsanitize=address -fsanitize=undefined -fno-sanitize-recover=all"
+          extra_args: ${{ matrix.target.extra_args || '' }}
           check_namespace: 'false'
       - name: build + test (cross, opt)
         uses: ./.github/actions/multi-functest
@@ -134,6 +143,7 @@ jobs:
           nix-cache: ${{ matrix.target.mode == 'native' && 'false' || 'true' }}
           gh_token: ${{ secrets.GITHUB_TOKEN }}
           compile_mode: ${{ matrix.target.mode }}
+          cflags: ${{ matrix.target.cflags || '' }}
           extra_args: ${{ matrix.target.extra_args || '' }}
           exec_wrapper: ${{ matrix.target.exec_wrapper || '' }}
           opt: 'opt'
@@ -148,7 +158,7 @@ jobs:
           compile_mode: ${{ matrix.target.mode }}
           extra_args: ${{ matrix.target.extra_args || '' }}
           exec_wrapper: ${{ matrix.target.exec_wrapper || '' }}
-          cflags: "-DMLDSA_DEBUG"
+          cflags: "${{ matrix.target.cflags || '' }} -DMLDSA_DEBUG"
           opt: 'opt'
   backend_tests:
     name: AArch64 FIPS202 backends (${{ matrix.backend }})

dev/fips202/aarch64/auto.h

@@ -42,6 +42,9 @@
 #if (!defined(MLD_CONFIG_NO_KEYPAIR_API) ||                                  \
      !defined(MLD_CONFIG_NO_SIGN_API) || !defined(MLD_CONFIG_REDUCE_RAM)) && \
     !defined(MLD_CONFIG_SERIAL_FIPS202_ONLY)
+/* Batched, SIMD-based Keccak-f1600 implementations. */
+#if defined(MLD_SYS_AARCH64_NEON)
+
 /*
  * Keccak-f1600x2/x4
  *
@@ -74,6 +77,8 @@
 
 #endif /* !__ARM_FEATURE_SHA3 */
 
+#endif /* MLD_SYS_AARCH64_NEON */
+
 #endif /* (!MLD_CONFIG_NO_KEYPAIR_API || !MLD_CONFIG_NO_SIGN_API || \
           !MLD_CONFIG_REDUCE_RAM) && !MLD_CONFIG_SERIAL_FIPS202_ONLY */
 

mldsa/mldsa_native.c

@@ -525,6 +525,7 @@
 #undef MLD_SYSV_ABI_SUPPORTED
 #undef MLD_SYS_AARCH64
 #undef MLD_SYS_AARCH64_EB
+#undef MLD_SYS_AARCH64_NEON
 #undef MLD_SYS_APPLE
 #undef MLD_SYS_ARMV81M_MVE
 #undef MLD_SYS_BIG_ENDIAN

mldsa/mldsa_native_asm.S

@@ -552,6 +552,7 @@
 #undef MLD_SYSV_ABI_SUPPORTED
 #undef MLD_SYS_AARCH64
 #undef MLD_SYS_AARCH64_EB
+#undef MLD_SYS_AARCH64_NEON
 #undef MLD_SYS_APPLE
 #undef MLD_SYS_ARMV81M_MVE
 #undef MLD_SYS_BIG_ENDIAN

mldsa/src/fips202/native/aarch64/auto.h

@@ -42,6 +42,9 @@
 #if (!defined(MLD_CONFIG_NO_KEYPAIR_API) ||                                  \
      !defined(MLD_CONFIG_NO_SIGN_API) || !defined(MLD_CONFIG_REDUCE_RAM)) && \
     !defined(MLD_CONFIG_SERIAL_FIPS202_ONLY)
+/* Batched, SIMD-based Keccak-f1600 implementations. */
+#if defined(MLD_SYS_AARCH64_NEON)
+
 /*
  * Keccak-f1600x2/x4
  *
@@ -74,6 +77,8 @@
 
 #endif /* !__ARM_FEATURE_SHA3 */
 
+#endif /* MLD_SYS_AARCH64_NEON */
+
 #endif /* (!MLD_CONFIG_NO_KEYPAIR_API || !MLD_CONFIG_NO_SIGN_API || \
           !MLD_CONFIG_REDUCE_RAM) && !MLD_CONFIG_SERIAL_FIPS202_ONLY */
 

mldsa/src/native/meta.h

@@ -12,7 +12,7 @@
  */
 #include "../sys.h"
 
-#ifdef MLD_SYS_AARCH64
+#ifdef MLD_SYS_AARCH64_NEON
 #include "aarch64/meta.h"
 #endif
 

mldsa/src/sys.h

@@ -39,6 +39,21 @@
 #define MLD_SYS_AARCH64
 #endif
 
+/* Check if the AArch64 compilation target supports NEON (Advanced SIMD).
+ *
+ * Some compilers also define __ARM_NEON__, but __ARM_NEON is the most reliable
+ * signal. Specifically, clang on Apple appears to keep __ARM_NEON__ set even if
+ * -march=armv8-a+nosimd is set.
+ *
+ * gcc 4.8 -- the first gcc version introducing Neon support -- sets neither
+ * __ARM_NEON nor __ARM_NEON__; in fact, there is no preprocessor signal that
+ * Neon is enabled. If you use gcc 4.8 and need Neon, you should set
+ * MLD_SYS_AARCH64_NEON manually. gcc 4.9 onwards do set __ARM_NEON.
+ */
+#if defined(MLD_SYS_AARCH64) && defined(__ARM_NEON)
+#define MLD_SYS_AARCH64_NEON
+#endif
+
 /* Check if we're running on an AArch64 big endian system. */
 #if defined(__AARCH64EB__)
 #define MLD_SYS_AARCH64_EB

scripts/autogen

@@ -4257,10 +4257,10 @@ ABI_CAPS = {
     },
     "NEON": {
         "asm_subdir": "aarch64",
-        "guard": None,
+        "guard": "MLD_SYS_AARCH64_NEON",
         "syscap": "MLD_SYS_CAP_AARCH64_NEON",
         "description": "AArch64 NEON",
-        "cflags": [],
+        "cflags": ["-march=armv8-a+simd"],
         "aux_files": [],
     },
     "SHA3": {

test/abicheck/aarch64/abicheck_aarch64.c

@@ -45,7 +45,9 @@ int check_aarch64_aapcs_compliance(struct aarch64_register_state *before,
     }
   }
 
-  /* The call stub leaves vector output untouched when NEON is unavailable. */
+  /* The call stub leaves vector output untouched when NEON was not compiled
+   * in or is unavailable at runtime. */
+#if defined(MLD_SYS_AARCH64_NEON)
   if (mld_sys_check_capability(MLD_SYS_CAP_AARCH64_NEON))
   {
     /* Check callee-saved NEON registers (d8-d15, lower 64 bits only). */
@@ -58,6 +60,7 @@ int check_aarch64_aapcs_compliance(struct aarch64_register_state *before,
       }
     }
   }
+#endif /* MLD_SYS_AARCH64_NEON */
 
   return violations;
 }

test/abicheck/aarch64/abicheck_aarch64.h

@@ -33,7 +33,11 @@ static MLD_INLINE void call_stub_aarch64(struct aarch64_register_state *input,
                                          struct aarch64_register_state *output,
                                          void (*function_ptr)(void))
 {
+#if defined(MLD_SYS_AARCH64_NEON)
   int use_neon = mld_sys_check_capability(MLD_SYS_CAP_AARCH64_NEON) != 0;
+#else
+  int use_neon = 0;
+#endif
   asm_call_stub_aarch64(input, output, function_ptr, use_neon);
 }