| 1 | #[cfg (not(crossbeam_no_atomic))] |
| 2 | use core::sync::atomic::Ordering; |
| 3 | |
| 4 | /// Trait which allows reading from primitive atomic types with "consume" ordering. |
| 5 | pub trait AtomicConsume { |
| 6 | /// Type returned by `load_consume`. |
| 7 | type Val; |
| 8 | |
| 9 | /// Loads a value from the atomic using a "consume" memory ordering. |
| 10 | /// |
| 11 | /// This is similar to the "acquire" ordering, except that an ordering is |
| 12 | /// only guaranteed with operations that "depend on" the result of the load. |
| 13 | /// However consume loads are usually much faster than acquire loads on |
| 14 | /// architectures with a weak memory model since they don't require memory |
| 15 | /// fence instructions. |
| 16 | /// |
| 17 | /// The exact definition of "depend on" is a bit vague, but it works as you |
| 18 | /// would expect in practice since a lot of software, especially the Linux |
| 19 | /// kernel, rely on this behavior. |
| 20 | /// |
| 21 | /// This is currently only implemented on ARM and AArch64, where a fence |
| 22 | /// can be avoided. On other architectures this will fall back to a simple |
| 23 | /// `load(Ordering::Acquire)`. |
| 24 | fn load_consume(&self) -> Self::Val; |
| 25 | } |
| 26 | |
| 27 | #[cfg (not(crossbeam_no_atomic))] |
| 28 | // Miri and Loom don't support "consume" ordering and ThreadSanitizer doesn't treat |
| 29 | // load(Relaxed) + compiler_fence(Acquire) as "consume" load. |
| 30 | // LLVM generates machine code equivalent to fence(Acquire) in compiler_fence(Acquire) |
| 31 | // on PowerPC, MIPS, etc. (https://godbolt.org/z/hffvjvW7h), so for now the fence |
| 32 | // can be actually avoided here only on ARM and AArch64. See also |
| 33 | // https://github.com/rust-lang/rust/issues/62256. |
| 34 | #[cfg (all( |
| 35 | any(target_arch = "arm" , target_arch = "aarch64" ), |
| 36 | not(any(miri, crossbeam_loom, crossbeam_sanitize_thread)), |
| 37 | ))] |
| 38 | macro_rules! impl_consume { |
| 39 | () => { |
| 40 | #[inline] |
| 41 | fn load_consume(&self) -> Self::Val { |
| 42 | use crate::primitive::sync::atomic::compiler_fence; |
| 43 | let result = self.load(Ordering::Relaxed); |
| 44 | compiler_fence(Ordering::Acquire); |
| 45 | result |
| 46 | } |
| 47 | }; |
| 48 | } |
| 49 | |
| 50 | #[cfg (not(crossbeam_no_atomic))] |
| 51 | #[cfg (not(all( |
| 52 | any(target_arch = "arm" , target_arch = "aarch64" ), |
| 53 | not(any(miri, crossbeam_loom, crossbeam_sanitize_thread)), |
| 54 | )))] |
| 55 | macro_rules! impl_consume { |
| 56 | () => { |
| 57 | #[inline] |
| 58 | fn load_consume(&self) -> Self::Val { |
| 59 | self.load(Ordering::Acquire) |
| 60 | } |
| 61 | }; |
| 62 | } |
| 63 | |
| 64 | macro_rules! impl_atomic { |
| 65 | ($atomic:ident, $val:ty) => { |
| 66 | #[cfg(not(crossbeam_no_atomic))] |
| 67 | impl AtomicConsume for core::sync::atomic::$atomic { |
| 68 | type Val = $val; |
| 69 | impl_consume!(); |
| 70 | } |
| 71 | #[cfg(crossbeam_loom)] |
| 72 | impl AtomicConsume for loom::sync::atomic::$atomic { |
| 73 | type Val = $val; |
| 74 | impl_consume!(); |
| 75 | } |
| 76 | }; |
| 77 | } |
| 78 | |
| 79 | impl_atomic!(AtomicBool, bool); |
| 80 | impl_atomic!(AtomicUsize, usize); |
| 81 | impl_atomic!(AtomicIsize, isize); |
| 82 | impl_atomic!(AtomicU8, u8); |
| 83 | impl_atomic!(AtomicI8, i8); |
| 84 | impl_atomic!(AtomicU16, u16); |
| 85 | impl_atomic!(AtomicI16, i16); |
| 86 | #[cfg (any(target_has_atomic = "32" , not(target_pointer_width = "16" )))] |
| 87 | impl_atomic!(AtomicU32, u32); |
| 88 | #[cfg (any(target_has_atomic = "32" , not(target_pointer_width = "16" )))] |
| 89 | impl_atomic!(AtomicI32, i32); |
| 90 | #[cfg (any( |
| 91 | target_has_atomic = "64" , |
| 92 | not(any(target_pointer_width = "16" , target_pointer_width = "32" )), |
| 93 | ))] |
| 94 | impl_atomic!(AtomicU64, u64); |
| 95 | #[cfg (any( |
| 96 | target_has_atomic = "64" , |
| 97 | not(any(target_pointer_width = "16" , target_pointer_width = "32" )), |
| 98 | ))] |
| 99 | impl_atomic!(AtomicI64, i64); |
| 100 | |
| 101 | #[cfg (not(crossbeam_no_atomic))] |
| 102 | impl<T> AtomicConsume for core::sync::atomic::AtomicPtr<T> { |
| 103 | type Val = *mut T; |
| 104 | impl_consume!(); |
| 105 | } |
| 106 | |
| 107 | #[cfg (crossbeam_loom)] |
| 108 | impl<T> AtomicConsume for loom::sync::atomic::AtomicPtr<T> { |
| 109 | type Val = *mut T; |
| 110 | impl_consume!(); |
| 111 | } |
| 112 | |