arm crc32 fix

diff --git a/galerautils/src/gu_crc32c_arm64.c b/galerautils/src/gu_crc32c_arm64.c
index 62c0a0fa..f8647abf 100644
--- a/galerautils/src/gu_crc32c_arm64.c
+++ b/galerautils/src/gu_crc32c_arm64.c
@@ -20,7 +20,6 @@
 #include <stdbool.h>
 
 #include <arm_acle.h>
-#include <arm_neon.h>
 
 static inline gu_crc32c_t
 crc32c_arm64_tail7(gu_crc32c_t state, const uint8_t* ptr, size_t len)
@@ -69,25 +68,36 @@ gu_crc32c_arm64(gu_crc32c_t state, const void* data, size_t len)
     return crc32c_arm64_tail7(state, ptr, len);
 }
 
+#include <asm/hwcap.h>
 #include <sys/auxv.h>
 
-#ifndef HWCAP_CRC32
-#define HWCAP_CRC32 (1 << 7)
+#if defined(HWCAP_CRC32)
+#    define GU_AT_HWCAP    AT_HWCAP
+#    define GU_HWCAP_CRC32 HWCAP_CRC32
+#elif defined(HWCAP2_CRC32)
+#    define GU_AT_HWCAP    AT_HWCAP2
+#    define GU_HWCAP_CRC32 HWCAP2_CRC32
 #endif /* HWCAP_CRC32 */
 
 gu_crc32c_func_t
 gu_crc32c_hardware()
 {
-    unsigned long int const hwcaps = getauxval(AT_HWCAP);
-    if (hwcaps & HWCAP_CRC32)
+#if defined(GU_AT_HWCAP)
+    unsigned long int const hwcaps = getauxval(GU_AT_HWCAP);
+    if (hwcaps & GU_HWCAP_CRC32)
     {
         gu_info ("CRC-32C: using hardware acceleration.");
         return gu_crc32c_arm64;
     }
     else
     {
+        gu_info ("CRC-32C: hardware does not have CRC-32C capabilities.");
         return NULL;
     }
+#else
+    gu_info ("CRC-32C: compiled without hardware acceleration support.");
+    return NULL;
+#endif /* GU_AT_HWCAP */
 }
 
 #endif /* GU_CRC32C_ARM64 */
diff --git a/galerautils/tests/crc32c_bench.cpp b/galerautils/tests/crc32c_bench.cpp
index 74dd93fb..a713d338 100644
--- a/galerautils/tests/crc32c_bench.cpp
+++ b/galerautils/tests/crc32c_bench.cpp
@@ -93,6 +93,8 @@ run_bench_with_impl(gu_crc32c_func_t impl,
               << std::fixed << duration << '\t' << result << '\n';
 }
 
+static gu_crc32c_func_t configured_impl;
+
 static void
 one_length(size_t const len, size_t const reps)
 {
@@ -109,7 +111,8 @@ one_length(size_t const len, size_t const reps)
 #endif /* GU_CRC32C_X86 */
 
 #if defined(GU_CRC32C_ARM64)
-    run_bench_with_impl(gu_crc32c_arm64,        len, reps, "GU arm64   ");
+    if (gu_crc32c_arm64 == configured_impl)
+        run_bench_with_impl(gu_crc32c_arm64,    len, reps, "GU arm64   ");
 #endif /* GU_CRC32C_X86 */
 }
 
@@ -117,9 +120,11 @@ int main()
 {
     gu_crc32c_configure(); // compute SW lookup tables
 
+    configured_impl = gu_crc32c_func;
+
     one_length(11,  1<<22 /* 4M   */);
     one_length(31,  1<<21 /* 2M   */);
     one_length(64,  1<<20 /* 1M   */);
     one_length(512, 1<<17 /* 128K */);
-    one_length(1<<20 /* 1M */,    64);
+    one_length(1<<20,  64 /* 1M   */);
 }
diff --git a/galerautils/tests/gu_crc32c_test.c b/galerautils/tests/gu_crc32c_test.c
index 96efb0bd..c3795107 100644
--- a/galerautils/tests/gu_crc32c_test.c
+++ b/galerautils/tests/gu_crc32c_test.c
@@ -144,8 +144,13 @@ END_TEST
 #if defined(GU_CRC32C_ARM64)
 START_TEST(test_gu_crc32c_arm64)
 {
-    gu_crc32c_func = gu_crc32c_arm64;
-    test_function();
+    gu_crc32c_func = gu_crc32c_hardware();
+
+    if (NULL != gu_crc32c_func)
+    {
+        ck_assert(gu_crc32c_arm64 == gu_crc32c_func);
+        test_function();
+    }
 }
 END_TEST
 #endif /* GU_CRC32C_ARM64 */