moved tsc and asm utilities to rx

rx/include/rx/SharedAtomic.hpp

@@ -6,19 +6,10 @@
 #include <functional>
 #include <limits>
 #include <system_error>
-#include <thread>
 #include <type_traits>
+#include "asm.hpp"
 
 namespace rx {
-inline void yield() { std::this_thread::yield(); }
-inline void relax() {
-#if defined(__GNUC__) && (defined __i386__ || defined __x86_64__)
-  __builtin_ia32_pause();
-#else
-  yield();
-#endif
-}
-
 static constexpr auto kRelaxSpinCount = 12;
 static constexpr auto kSpinCount = 16;
 
@@ -31,7 +22,7 @@ bool try_spin_wait(auto &&pred) {
     }
 
     if (i < kRelaxSpinCount) {
-      relax();
+      pause();
     } else {
       yield();
     }

rx/include/rx/align.hpp

@@ -1,14 +1,21 @@
 #pragma once
 
-#include <cstdint>
+#include <type_traits>
 
 namespace rx {
-inline constexpr std::uint64_t alignUp(std::uint64_t value,
-                                       std::uint64_t alignment) {
-  return (value + (alignment - 1)) & ~(alignment - 1);
+template <typename T, typename U>
+  requires std::is_unsigned_v<T>
+inline constexpr std::make_unsigned_t<std::common_type_t<T, U>>
+alignUp(T value, U alignment) {
+  return static_cast<std::make_unsigned_t<std::common_type_t<T, U>>>(
+      (value + (alignment - 1)) & ~(alignment - 1));
 }
-inline constexpr std::uint64_t alignDown(std::uint64_t value,
-                                         std::uint64_t alignment) {
-  return value & ~(alignment - 1);
+
+template <typename T, typename U>
+  requires std::is_unsigned_v<T>
+inline constexpr std::make_unsigned_t<std::common_type_t<T, U>>
+alignDown(T value, U alignment) {
+  return static_cast<std::make_unsigned_t<std::common_type_t<T, U>>>(
+      value & ~(alignment - 1));
 }
 } // namespace rx

rx/include/rx/asm.hpp

@@ -1,6 +1,8 @@
 #pragma once
 
+#include "rx/tsc.hpp"
 #include "types.hpp"
+#include <thread>
 #include <atomic>
 
 extern bool g_use_rtm;
@@ -275,6 +277,16 @@ inline void pause() {
 #endif
 }
 
+inline void yield() { std::this_thread::yield(); }
+
+// Synchronization helper (cache-friendly busy waiting)
+inline void busy_wait(usz cycles = 3000) {
+  const u64 stop = get_tsc() + cycles;
+  do
+    pause();
+  while (get_tsc() < stop);
+}
+
 // Align to power of 2
 template <typename T, typename U>
   requires std::is_unsigned_v<T>
@@ -312,12 +324,6 @@ constexpr T rational_mul(T value, std::type_identity_t<T> numerator,
     return static_cast<T>(value * u64{numerator} / u64{denominator});
   }
 
-#if is_u128_emulated
-  if constexpr (sizeof(T) <= sizeof(u128) / 2) {
-    return static_cast<T>(u128_from_mul(value, numerator) / u64{denominator});
-  }
-#endif
-
   return static_cast<T>(value / denominator * numerator +
                         (value % denominator) * numerator / denominator);
 }

rx/include/rx/tsc.hpp

@@ -0,0 +1,28 @@
+#pragma once
+
+#include "types.hpp"
+#include <cstdint>
+
+#ifdef _M_X64
+#ifdef _MSC_VER
+extern "C" std::uint64_t __rdtsc();
+#else
+#include <immintrin.h>
+#endif
+#endif
+
+namespace rx {
+inline std::uint64_t get_tsc() {
+#if defined(ARCH_ARM64)
+  std::uint64_t r = 0;
+  __asm__ volatile("mrs %0, cntvct_el0" : "=r"(r));
+  return r;
+#elif defined(_M_X64)
+  return __rdtsc();
+#elif defined(ARCH_X64)
+  return __builtin_ia32_rdtsc();
+#else
+#error "Missing rx::get_tsc() implementation"
+#endif
+}
+} // namespace rx

rx/include/rx/types.hpp

@@ -106,20 +106,6 @@ template <typename F> fn_helper(F &&f) -> fn_helper<F>;
                       [[maybe_unused]] auto &&z,                               \
                       [[maybe_unused]] auto &&w) { return (__VA_ARGS__); })
 
-#if __cpp_lib_bit_cast < 201806L
-namespace std {
-template <typename To, typename From>
-[[nodiscard]] constexpr To bit_cast(const From &from) noexcept {
-  return __builtin_bit_cast(To, from);
-}
-} // namespace std
-#endif
-
-#if defined(__INTELLISENSE__) || (defined(__clang__) && (__clang_major__ <= 16))
-#define consteval constexpr
-#define constinit
-#endif
-
 // FIXME: move to ps3 kernel implementation
 using schar = signed char;
 using uchar = unsigned char;
@@ -206,9 +192,9 @@ public:
 };
 
 #if defined(ARCH_X64) && !defined(_MSC_VER)
-using __m128i = long long __attribute__((vector_size(16)));
-using __m128d = double __attribute__((vector_size(16)));
-using __m128 = float __attribute__((vector_size(16)));
+using __m128i = long long __attribute__((vector_size(16), aligned(16)));
+using __m128d = double __attribute__((vector_size(16), aligned(16)));
+using __m128 = float __attribute__((vector_size(16), aligned(16)));
 #endif
 
 #ifndef _MSC_VER

rx/src/SharedMutex.cpp

@@ -1,14 +1,7 @@
 #include "SharedMutex.hpp"
+#include "asm.hpp"
 #include <syscall.h>
 #include <unistd.h>
-#include <xmmintrin.h>
-
-static void busy_wait(unsigned long long cycles = 3000) {
-  const auto stop = __builtin_ia32_rdtsc() + cycles;
-  do
-    _mm_pause();
-  while (__builtin_ia32_rdtsc() < stop);
-}
 
 namespace rx {
 void shared_mutex::impl_lock_shared(unsigned val) {