Use the 64b inner:monotonize() implementation not the 128b one for aa…

…rch64

aarch64 prior to v8.4 (FEAT_LSE2) doesn't have an instruction that guarantees
untorn 128b reads except for completing a 128b load/store exclusive pair
(ldxp/stxp) or compare-and-swap (casp) successfully. The requirement to
complete a 128b read+write atomic is actually more expensive and more unfair
than the previous implementation of monotonize() which used a Mutex on aarch64,
especially at large core counts.  For aarch64 switch to the 64b atomic
implementation which is about 13x faster for a benchmark that involves many
calls to Instant::now().

library/std/src/time/monotonic.rs

@@ -5,7 +5,7 @@ pub(super) fn monotonize(raw: time::Instant) -> time::Instant {
     inner::monotonize(raw)
 }
 
-#[cfg(all(target_has_atomic = "64", not(target_has_atomic = "128")))]
+#[cfg(any(all(target_has_atomic = "64", not(target_has_atomic = "128")), target_arch = "aarch64"))]
 pub mod inner {
     use crate::sync::atomic::AtomicU64;
     use crate::sync::atomic::Ordering::*;
@@ -70,7 +70,7 @@ pub mod inner {
     }
 }
 
-#[cfg(target_has_atomic = "128")]
+#[cfg(all(target_has_atomic = "128", not(target_arch = "aarch64")))]
 pub mod inner {
     use crate::sync::atomic::AtomicU128;
     use crate::sync::atomic::Ordering::*;