Make memcpy_nontemporal with SVE work without crashing

getauxval is not safe to using in an ifunc, because it runs too early.
But the ABI for AArch64 passes the HWCAPS as an argument.

src/common_memcpy.cpp

@@ -109,7 +109,11 @@ void *memcpy_nontemporal_sve(void * __restrict__ dest, const void * __restrict__
 }
 #endif // SPEAD2_USE_SVE_STREAM
 
-void *(*resolve_memcpy_nontemporal())(void *, const void *, std::size_t) noexcept
+void *(*resolve_memcpy_nontemporal(
+#ifdef __aarch64__
+    std::uint64_t hwcaps  // See System V AVI for AArch64
+#endif
+))(void *, const void *, std::size_t) noexcept
 {
     /* x86 options */
 #if SPEAD2_USE_AVX512_STREAM || SPEAD2_USE_AVX_STREAM || SPEAD2_USE_SSE2_STREAM
@@ -137,7 +141,6 @@ void *(*resolve_memcpy_nontemporal())(void *, const void *, std::size_t) noexcep
 
     /* aarch64 options */
 #if SPEAD2_USE_SVE_STREAM
-    unsigned long hwcaps = getauxval(AT_HWCAP);
     if (hwcaps & HWCAP_SVE)
         return memcpy_nontemporal_sve;
 #endif
@@ -150,14 +153,20 @@ void *(*resolve_memcpy_nontemporal())(void *, const void *, std::size_t) noexcep
 
 #if SPEAD2_USE_FMV
 
-[[gnu::ifunc("_ZN6spead226resolve_memcpy_nontemporalEv")]]
+[[gnu::ifunc("_ZN6spead226resolve_memcpy_nontemporalEm")]]
 void *memcpy_nontemporal(void * __restrict__ dest, const void * __restrict__ src, std::size_t n) noexcept;
 
 #else
 
 void *memcpy_nontemporal(void * __restrict__ dest, const void * __restrict__ src, std::size_t n) noexcept
 {
-    static void *(*memcpy_nontemporal_ptr)(void * __restrict__ dest, const void * __restrict__ src, std::size_t n) noexcept = resolve_memcpy_nontemporal();
+#ifdef __aarch64__
+    static void *(*memcpy_nontemporal_ptr)(void * __restrict__ dest, const void * __restrict__ src, std::size_t n) noexcept =
+        resolve_memcpy_nontemporal(getauxval(AT_HWCAPS));
+#else
+    static void *(*memcpy_nontemporal_ptr)(void * __restrict__ dest, const void * __restrict__ src, std::size_t n) noexcept =
+        resolve_memcpy_nontemporal();
+#endif
     return memcpy_nontemporal_ptr(dest, src, n);
 }
 

src/unittest_memcpy.cpp

@@ -69,9 +69,14 @@ std::ostream &operator<<(std::ostream &o, const memcpy_function &func)
     return o << func.name;
 }
 
+static void *wrap_memcpy_nontemporal(void * __restrict__ dest, const void * __restrict__ src, std::size_t n) noexcept
+{
+    return spead2::memcpy_nontemporal(dest, src, n);
+}
+
 static const memcpy_function memcpy_functions[] =
 {
-    { "default", spead2::memcpy_nontemporal, true },
+    { "default", wrap_memcpy_nontemporal, true },
 #if SPEAD2_USE_SSE2_STREAM
     { "sse2", spead2::memcpy_nontemporal_sse2, bool(__builtin_cpu_supports("sse2")) },
 #endif