Decrease the precision of absl::Now in x86-64 debug builds

CycleClock::Now utilizes ABSL_INTERNAL_CYCLECLOCK_SHIFT to discourage reliance
on the raw CPU cycle counter values. As a side effect, it discourages
strictly-increasing assumption. Apply the idea to discourage over-reliance on
the precision of absl::Now/absl::GetCurrentTimeNanos.

Programs relying on a very high precision often exhibit portability issues on
machines with a lower cycle counter precision, or worse, race conditions. For
example, Apple M1 emulated x86 RDTSC only guarantees weakly-increasing RDTSC
values.

x86 clock speed is usually measured in GHz and is still precise after we drop 8
least significant bits.

PiperOrigin-RevId: 603493500
Change-Id: Ib1b00075109283f5dbcb39a43fe0ab722351a1e2

Author: MaskRay

absl/time/clock.cc

@@ -88,11 +88,25 @@ ABSL_NAMESPACE_END
 namespace absl {
 ABSL_NAMESPACE_BEGIN
 namespace time_internal {
+
+// On some processors, consecutive reads of the cycle counter may yield the
+// same value (weakly-increasing). In debug mode, clear the least significant
+// bits to discourage depending on a strictly-increasing Now() value.
+// In x86-64's debug mode, discourage depending on a strictly-increasing Now()
+// value.
+#if !defined(NDEBUG) && defined(__x86_64__)
+constexpr int64_t kCycleClockNowMask = ~int64_t{0xff};
+#else
+constexpr int64_t kCycleClockNowMask = ~int64_t{0};
+#endif
+
 // This is a friend wrapper around UnscaledCycleClock::Now()
 // (needed to access UnscaledCycleClock).
 class UnscaledCycleClockWrapperForGetCurrentTime {
  public:
-  static int64_t Now() { return base_internal::UnscaledCycleClock::Now(); }
+  static int64_t Now() {
+    return base_internal::UnscaledCycleClock::Now() & kCycleClockNowMask;
+  }
 };
 }  // namespace time_internal