Add API for CPU cycle counter

Summary: Add an API for getting CPU cycle counter, when available. Falls back to the OS API when not.

Reviewed By: jpporto

Differential Revision: D36144469

fbshipit-source-id: b7c4168533b6b306be75fcb50603fe2152c88b4a

include/hermes/Support/OSCompat.h

@@ -190,6 +190,10 @@ std::vector<bool> sched_getaffinity();
 /// or -1 on error.
 int sched_getcpu();
 
+/// \return a monotonically increasing count of time, where the unit depends on
+/// the implementation.
+uint64_t cpu_cycle_counter();
+
 #ifdef _WINDOWS
 
 #define STDIN_FILENO 0

lib/Support/OSCompatPosix.cpp

@@ -704,6 +704,21 @@ int sched_getcpu() {
 }
 #endif
 
+uint64_t cpu_cycle_counter() {
+#if defined(__aarch64__)
+  // Clang's builtin causes SIGILL on some 64-bit ARM environments.
+  uint64_t cnt;
+  __asm __volatile("mrs %0, cntvct_el0" : "=&r"(cnt));
+  return cnt;
+#elif __has_builtin(__builtin_readcyclecounter)
+  return __builtin_readcyclecounter();
+#else
+  timespec t;
+  clock_gettime(CLOCK_MONOTONIC, &t);
+  return t.tv_sec * 1000LL * 1000LL * 1000LL + t.tv_nsec;
+#endif
+}
+
 bool set_env(const char *name, const char *value) {
   // Enforce the contract of this function that value must not be empty
   assert(*value != '\0' && "value cannot be empty string");

lib/Support/OSCompatWindows.cpp

@@ -19,6 +19,7 @@
 #define NOMINMAX // do not define min/max macros
 #include <windows.h>
 
+#include <intrin.h>
 #include <io.h>
 #include <psapi.h>
 
@@ -407,6 +408,18 @@ int sched_getcpu() {
   return -1;
 }
 
+uint64_t cpu_cycle_counter() {
+#if defined(_MSC_VER)
+  return __rdtsc();
+#elif __has_builtin(__builtin_readcyclecounter)
+  return __builtin_readcyclecounter();
+#else
+  LARGE_INTEGER cnt;
+  QueryPerformanceCounter(&cnt);
+  return cnt;
+#endif
+}
+
 bool set_env(const char *name, const char *value) {
   // Setting an env var to empty requires a lot of hacks on Windows
   assert(*value != '\0' && "value cannot be empty string");

unittests/Support/OSCompatTest.cpp

@@ -51,6 +51,12 @@ TEST(OSCompatTest, CurrentRSS) {
   EXPECT_GE(oscompat::current_rss(), beginRSS);
 }
 
+TEST(OSCompatTest, CpuCycles) {
+  uint64_t start = oscompat::cpu_cycle_counter();
+  uint64_t end = oscompat::cpu_cycle_counter();
+  EXPECT_LE(start, end);
+}
+
 #ifdef __linux__
 TEST(OSCompatTest, Scheduling) {
   // At least one CPU should be set.