feat: start implementing arm64 support (#313)

platform/generic.h

@@ -66,10 +66,17 @@ typedef FILE fp_outfile;
 
 /* compiler-gcc.h */
 #ifdef __GNUC__
+#ifdef __aarch64__
+#define barrier() __asm__ __volatile__("" : : : "memory")
+#define mb() asm volatile("dmb ish" : : : "memory")
+#define rmb() asm volatile("dmb ishld" : : : "memory")
+#define wmb() asm volatile("dmb ishst" : : : "memory")
+#else // __aarch64__
 #define barrier() __asm__ __volatile__("" : : : "memory")
 #define mb() asm volatile("mfence" : : : "memory")
 #define rmb() asm volatile("lfence" : : : "memory")
 #define wmb() asm volatile("sfence" : : : "memory")
+#endif // ! __aarch64__
 #define read_barrier_depends()                                                                                                 \
   do {                                                                                                                         \
   } while (0)

platform/userspace-time.h

@@ -38,12 +38,6 @@ static inline __attribute__((always_inline)) u64 monotonic()
   return (u64)ts.tv_sec * (1000 * 1000 * 1000) + ts.tv_nsec;
 }
 
-#include <x86intrin.h>
-static inline __attribute__((always_inline)) u64 fp_monotonic_time_ns(void)
-{
-  return __rdtsc();
-}
-
 #ifdef __cplusplus
 }
 #endif /* __cplusplus */

util/time.h

@@ -147,7 +147,9 @@ struct monotonic_clock {
   }
 };
 
+#ifdef __x86_64__
 #include <x86intrin.h>
+#endif
 /**
  * Monotonic, std::chrono compatible clock using `__rdtsc()` as the backend.
  */
@@ -159,5 +161,13 @@ struct rdtsc_clock {
   using duration = std::chrono::duration<rep, period>;
   using time_point = std::chrono::time_point<monotonic_clock>;
 
+#ifdef __aarch64__
+  static inline __attribute__((always_inline)) time_point now() {
+	  uint64_t val;
+	  __asm__ __volatile__("mrs %0, cntvct_el0" : "=r"(val));
+	  return time_point{duration{val}};
+  }
+#else
   static inline __attribute__((always_inline)) time_point now() { return time_point{duration{__rdtsc()}}; }
+#endif
 };