1 | // SPDX-License-Identifier: Apache-2.0 OR MIT |
2 | |
3 | #![cfg_attr (not(all(test, feature = "float" )), allow(dead_code, unused_macros))] |
4 | |
5 | #[macro_use ] |
6 | #[path = "gen/utils.rs" ] |
7 | mod gen; |
8 | |
9 | use core::sync::atomic::Ordering; |
10 | |
11 | macro_rules! static_assert { |
12 | ($cond:expr $(,)?) => {{ |
13 | let [] = [(); true as usize - $crate::utils::_assert_is_bool($cond) as usize]; |
14 | }}; |
15 | } |
16 | pub(crate) const fn _assert_is_bool(v: bool) -> bool { |
17 | v |
18 | } |
19 | |
20 | macro_rules! static_assert_layout { |
21 | ($atomic_type:ty, $value_type:ty) => { |
22 | static_assert!( |
23 | core::mem::align_of::<$atomic_type>() == core::mem::size_of::<$atomic_type>() |
24 | ); |
25 | static_assert!(core::mem::size_of::<$atomic_type>() == core::mem::size_of::<$value_type>()); |
26 | }; |
27 | } |
28 | |
29 | // #[doc = concat!(...)] requires Rust 1.54 |
30 | macro_rules! doc_comment { |
31 | ($doc:expr, $($tt:tt)*) => { |
32 | #[doc = $doc] |
33 | $($tt)* |
34 | }; |
35 | } |
36 | |
37 | // Adapted from https://github.com/BurntSushi/memchr/blob/2.4.1/src/memchr/x86/mod.rs#L9-L71. |
38 | /// # Safety |
39 | /// |
40 | /// - the caller must uphold the safety contract for the function returned by $detect_body. |
41 | /// - the memory pointed by the function pointer returned by $detect_body must be visible from any threads. |
42 | /// |
43 | /// The second requirement is always met if the function pointer is to the function definition. |
44 | /// (Currently, all uses of this macro in our code are in this case.) |
45 | #[allow (unused_macros)] |
46 | #[cfg (not(portable_atomic_no_outline_atomics))] |
47 | #[cfg (any( |
48 | target_arch = "aarch64" , |
49 | target_arch = "arm" , |
50 | target_arch = "arm64ec" , |
51 | target_arch = "powerpc64" , |
52 | target_arch = "riscv32" , |
53 | target_arch = "riscv64" , |
54 | all(target_arch = "x86_64" , not(any(target_env = "sgx" , miri))), |
55 | ))] |
56 | macro_rules! ifunc { |
57 | (unsafe fn($($arg_pat:ident: $arg_ty:ty),*) $(-> $ret_ty:ty)? { $($detect_body:tt)* }) => {{ |
58 | type FnTy = unsafe fn($($arg_ty),*) $(-> $ret_ty)?; |
59 | static FUNC: core::sync::atomic::AtomicPtr<()> |
60 | = core::sync::atomic::AtomicPtr::new(detect as *mut ()); |
61 | #[cold] |
62 | unsafe fn detect($($arg_pat: $arg_ty),*) $(-> $ret_ty)? { |
63 | let func: FnTy = { $($detect_body)* }; |
64 | FUNC.store(func as *mut (), core::sync::atomic::Ordering::Relaxed); |
65 | // SAFETY: the caller must uphold the safety contract for the function returned by $detect_body. |
66 | unsafe { func($($arg_pat),*) } |
67 | } |
68 | // SAFETY: `FnTy` is a function pointer, which is always safe to transmute with a `*mut ()`. |
69 | // (To force the caller to use unsafe block for this macro, do not use |
70 | // unsafe block here.) |
71 | let func = { |
72 | core::mem::transmute::<*mut (), FnTy>(FUNC.load(core::sync::atomic::Ordering::Relaxed)) |
73 | }; |
74 | // SAFETY: the caller must uphold the safety contract for the function returned by $detect_body. |
75 | // (To force the caller to use unsafe block for this macro, do not use |
76 | // unsafe block here.) |
77 | func($($arg_pat),*) |
78 | }}; |
79 | } |
80 | |
81 | #[allow (unused_macros)] |
82 | #[cfg (not(portable_atomic_no_outline_atomics))] |
83 | #[cfg (any( |
84 | target_arch = "aarch64" , |
85 | target_arch = "arm" , |
86 | target_arch = "arm64ec" , |
87 | target_arch = "powerpc64" , |
88 | target_arch = "riscv32" , |
89 | target_arch = "riscv64" , |
90 | all(target_arch = "x86_64" , not(any(target_env = "sgx" , miri))), |
91 | ))] |
92 | macro_rules! fn_alias { |
93 | ( |
94 | $(#[$($fn_attr:tt)*])* |
95 | $vis:vis unsafe fn($($arg_pat:ident: $arg_ty:ty),*) $(-> $ret_ty:ty)?; |
96 | $(#[$($alias_attr:tt)*])* |
97 | $new:ident = $from:ident($($last_args:tt)*); |
98 | $($rest:tt)* |
99 | ) => { |
100 | $(#[$($fn_attr)*])* |
101 | $(#[$($alias_attr)*])* |
102 | $vis unsafe fn $new($($arg_pat: $arg_ty),*) $(-> $ret_ty)? { |
103 | // SAFETY: the caller must uphold the safety contract. |
104 | unsafe { $from($($arg_pat,)* $($last_args)*) } |
105 | } |
106 | fn_alias! { |
107 | $(#[$($fn_attr)*])* |
108 | $vis unsafe fn($($arg_pat: $arg_ty),*) $(-> $ret_ty)?; |
109 | $($rest)* |
110 | } |
111 | }; |
112 | ( |
113 | $(#[$($attr:tt)*])* |
114 | $vis:vis unsafe fn($($arg_pat:ident: $arg_ty:ty),*) $(-> $ret_ty:ty)?; |
115 | ) => {} |
116 | } |
117 | |
118 | /// Make the given function const if the given condition is true. |
119 | macro_rules! const_fn { |
120 | ( |
121 | const_if: #[cfg($($cfg:tt)+)]; |
122 | $(#[$($attr:tt)*])* |
123 | $vis:vis const $($rest:tt)* |
124 | ) => { |
125 | #[cfg($($cfg)+)] |
126 | $(#[$($attr)*])* |
127 | $vis const $($rest)* |
128 | #[cfg(not($($cfg)+))] |
129 | $(#[$($attr)*])* |
130 | $vis $($rest)* |
131 | }; |
132 | } |
133 | |
134 | /// Implements `core::fmt::Debug` and `serde::{Serialize, Deserialize}` (when serde |
135 | /// feature is enabled) for atomic bool, integer, or float. |
136 | macro_rules! impl_debug_and_serde { |
137 | ($atomic_type:ident) => { |
138 | impl fmt::Debug for $atomic_type { |
139 | #[inline] // fmt is not hot path, but #[inline] on fmt seems to still be useful: https://github.com/rust-lang/rust/pull/117727 |
140 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
141 | // std atomic types use Relaxed in Debug::fmt: https://github.com/rust-lang/rust/blob/1.80.0/library/core/src/sync/atomic.rs#L2166 |
142 | fmt::Debug::fmt(&self.load(Ordering::Relaxed), f) |
143 | } |
144 | } |
145 | #[cfg(feature = "serde" )] |
146 | #[cfg_attr(docsrs, doc(cfg(feature = "serde" )))] |
147 | impl serde::ser::Serialize for $atomic_type { |
148 | #[allow(clippy::missing_inline_in_public_items)] // serde doesn't use inline on std atomic's Serialize/Deserialize impl |
149 | fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error> |
150 | where |
151 | S: serde::ser::Serializer, |
152 | { |
153 | // https://github.com/serde-rs/serde/blob/v1.0.152/serde/src/ser/impls.rs#L958-L959 |
154 | self.load(Ordering::Relaxed).serialize(serializer) |
155 | } |
156 | } |
157 | #[cfg(feature = "serde" )] |
158 | #[cfg_attr(docsrs, doc(cfg(feature = "serde" )))] |
159 | impl<'de> serde::de::Deserialize<'de> for $atomic_type { |
160 | #[allow(clippy::missing_inline_in_public_items)] // serde doesn't use inline on std atomic's Serialize/Deserialize impl |
161 | fn deserialize<D>(deserializer: D) -> Result<Self, D::Error> |
162 | where |
163 | D: serde::de::Deserializer<'de>, |
164 | { |
165 | serde::de::Deserialize::deserialize(deserializer).map(Self::new) |
166 | } |
167 | } |
168 | }; |
169 | } |
170 | |
171 | // We do not provide `nand` because it cannot be optimized on neither x86 nor MSP430. |
172 | // https://godbolt.org/z/ahWejchbT |
173 | macro_rules! impl_default_no_fetch_ops { |
174 | ($atomic_type:ident, bool) => { |
175 | impl $atomic_type { |
176 | #[inline] |
177 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
178 | pub(crate) fn and(&self, val: bool, order: Ordering) { |
179 | self.fetch_and(val, order); |
180 | } |
181 | #[inline] |
182 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
183 | pub(crate) fn or(&self, val: bool, order: Ordering) { |
184 | self.fetch_or(val, order); |
185 | } |
186 | #[inline] |
187 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
188 | pub(crate) fn xor(&self, val: bool, order: Ordering) { |
189 | self.fetch_xor(val, order); |
190 | } |
191 | } |
192 | }; |
193 | ($atomic_type:ident, $int_type:ty) => { |
194 | impl $atomic_type { |
195 | #[inline] |
196 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
197 | pub(crate) fn add(&self, val: $int_type, order: Ordering) { |
198 | self.fetch_add(val, order); |
199 | } |
200 | #[inline] |
201 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
202 | pub(crate) fn sub(&self, val: $int_type, order: Ordering) { |
203 | self.fetch_sub(val, order); |
204 | } |
205 | #[inline] |
206 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
207 | pub(crate) fn and(&self, val: $int_type, order: Ordering) { |
208 | self.fetch_and(val, order); |
209 | } |
210 | #[inline] |
211 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
212 | pub(crate) fn or(&self, val: $int_type, order: Ordering) { |
213 | self.fetch_or(val, order); |
214 | } |
215 | #[inline] |
216 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
217 | pub(crate) fn xor(&self, val: $int_type, order: Ordering) { |
218 | self.fetch_xor(val, order); |
219 | } |
220 | } |
221 | }; |
222 | } |
223 | macro_rules! impl_default_bit_opts { |
224 | ($atomic_type:ident, $int_type:ty) => { |
225 | impl $atomic_type { |
226 | #[inline] |
227 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
228 | pub(crate) fn bit_set(&self, bit: u32, order: Ordering) -> bool { |
229 | let mask = <$int_type>::wrapping_shl(1, bit); |
230 | self.fetch_or(mask, order) & mask != 0 |
231 | } |
232 | #[inline] |
233 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
234 | pub(crate) fn bit_clear(&self, bit: u32, order: Ordering) -> bool { |
235 | let mask = <$int_type>::wrapping_shl(1, bit); |
236 | self.fetch_and(!mask, order) & mask != 0 |
237 | } |
238 | #[inline] |
239 | #[cfg_attr(miri, track_caller)] // even without panics, this helps for Miri backtraces |
240 | pub(crate) fn bit_toggle(&self, bit: u32, order: Ordering) -> bool { |
241 | let mask = <$int_type>::wrapping_shl(1, bit); |
242 | self.fetch_xor(mask, order) & mask != 0 |
243 | } |
244 | } |
245 | }; |
246 | } |
247 | |
248 | // This just outputs the input as is, but can be used like an item-level block by using it with cfg. |
249 | macro_rules! items { |
250 | ($($tt:tt)*) => { |
251 | $($tt)* |
252 | }; |
253 | } |
254 | |
255 | #[allow (dead_code)] |
256 | #[cfg (any(target_arch = "x86" , target_arch = "x86_64" ))] |
257 | // Stable version of https://doc.rust-lang.org/nightly/std/hint/fn.assert_unchecked.html. |
258 | // TODO: use real core::hint::assert_unchecked on 1.81+ https://github.com/rust-lang/rust/pull/123588 |
259 | #[inline (always)] |
260 | #[cfg_attr (all(debug_assertions, not(portable_atomic_no_track_caller)), track_caller)] |
261 | pub(crate) unsafe fn assert_unchecked(cond: bool) { |
262 | if !cond { |
263 | if cfg!(debug_assertions) { |
264 | unreachable!() |
265 | } else { |
266 | // SAFETY: the caller promised `cond` is true. |
267 | unsafe { core::hint::unreachable_unchecked() } |
268 | } |
269 | } |
270 | } |
271 | |
272 | // https://github.com/rust-lang/rust/blob/1.80.0/library/core/src/sync/atomic.rs#L3294 |
273 | #[inline ] |
274 | #[cfg_attr (all(debug_assertions, not(portable_atomic_no_track_caller)), track_caller)] |
275 | pub(crate) fn assert_load_ordering(order: Ordering) { |
276 | match order { |
277 | Ordering::Acquire | Ordering::Relaxed | Ordering::SeqCst => {} |
278 | Ordering::Release => panic!("there is no such thing as a release load" ), |
279 | Ordering::AcqRel => panic!("there is no such thing as an acquire-release load" ), |
280 | _ => unreachable!(), |
281 | } |
282 | } |
283 | |
284 | // https://github.com/rust-lang/rust/blob/1.80.0/library/core/src/sync/atomic.rs#L3279 |
285 | #[inline ] |
286 | #[cfg_attr (all(debug_assertions, not(portable_atomic_no_track_caller)), track_caller)] |
287 | pub(crate) fn assert_store_ordering(order: Ordering) { |
288 | match order { |
289 | Ordering::Release | Ordering::Relaxed | Ordering::SeqCst => {} |
290 | Ordering::Acquire => panic!("there is no such thing as an acquire store" ), |
291 | Ordering::AcqRel => panic!("there is no such thing as an acquire-release store" ), |
292 | _ => unreachable!(), |
293 | } |
294 | } |
295 | |
296 | // https://github.com/rust-lang/rust/blob/1.80.0/library/core/src/sync/atomic.rs#L3360 |
297 | #[inline ] |
298 | #[cfg_attr (all(debug_assertions, not(portable_atomic_no_track_caller)), track_caller)] |
299 | pub(crate) fn assert_compare_exchange_ordering(success: Ordering, failure: Ordering) { |
300 | match success { |
301 | Ordering::AcqRel |
302 | | Ordering::Acquire |
303 | | Ordering::Relaxed |
304 | | Ordering::Release |
305 | | Ordering::SeqCst => {} |
306 | _ => unreachable!(), |
307 | } |
308 | match failure { |
309 | Ordering::Acquire | Ordering::Relaxed | Ordering::SeqCst => {} |
310 | Ordering::Release => panic!("there is no such thing as a release failure ordering" ), |
311 | Ordering::AcqRel => panic!("there is no such thing as an acquire-release failure ordering" ), |
312 | _ => unreachable!(), |
313 | } |
314 | } |
315 | |
316 | // https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0418r2.html |
317 | // https://github.com/rust-lang/rust/pull/98383 |
318 | #[allow (dead_code)] |
319 | #[inline ] |
320 | pub(crate) fn upgrade_success_ordering(success: Ordering, failure: Ordering) -> Ordering { |
321 | match (success, failure) { |
322 | (Ordering::Relaxed, Ordering::Acquire) => Ordering::Acquire, |
323 | (Ordering::Release, Ordering::Acquire) => Ordering::AcqRel, |
324 | (_, Ordering::SeqCst) => Ordering::SeqCst, |
325 | _ => success, |
326 | } |
327 | } |
328 | |
329 | /// Zero-extends the given 32-bit pointer to `MaybeUninit<u64>`. |
330 | /// This is used for 64-bit architecture's 32-bit ABI (e.g., AArch64 ILP32 ABI). |
331 | /// See ptr_reg! macro in src/gen/utils.rs for details. |
332 | #[cfg (not(portable_atomic_no_asm_maybe_uninit))] |
333 | #[cfg (target_pointer_width = "32" )] |
334 | #[allow (dead_code)] |
335 | #[inline ] |
336 | pub(crate) fn zero_extend64_ptr(v: *mut ()) -> core::mem::MaybeUninit<u64> { |
337 | #[repr (C)] |
338 | struct ZeroExtended { |
339 | #[cfg (target_endian = "big" )] |
340 | pad: *mut (), |
341 | v: *mut (), |
342 | #[cfg (target_endian = "little" )] |
343 | pad: *mut (), |
344 | } |
345 | // SAFETY: we can safely transmute any 64-bit value to MaybeUninit<u64>. |
346 | unsafe { core::mem::transmute(ZeroExtended { v, pad: core::ptr::null_mut() }) } |
347 | } |
348 | |
349 | #[allow (dead_code)] |
350 | #[cfg (any( |
351 | target_arch = "aarch64" , |
352 | target_arch = "arm64ec" , |
353 | target_arch = "powerpc64" , |
354 | target_arch = "riscv64" , |
355 | target_arch = "s390x" , |
356 | target_arch = "x86_64" , |
357 | ))] |
358 | /// A 128-bit value represented as a pair of 64-bit values. |
359 | /// |
360 | /// This type is `#[repr(C)]`, both fields have the same in-memory representation |
361 | /// and are plain old data types, so access to the fields is always safe. |
362 | #[derive (Clone, Copy)] |
363 | #[repr (C)] |
364 | pub(crate) union U128 { |
365 | pub(crate) whole: u128, |
366 | pub(crate) pair: Pair<u64>, |
367 | } |
368 | #[allow (dead_code)] |
369 | #[cfg (any(target_arch = "arm" , target_arch = "riscv32" ))] |
370 | /// A 64-bit value represented as a pair of 32-bit values. |
371 | /// |
372 | /// This type is `#[repr(C)]`, both fields have the same in-memory representation |
373 | /// and are plain old data types, so access to the fields is always safe. |
374 | #[derive (Clone, Copy)] |
375 | #[repr (C)] |
376 | pub(crate) union U64 { |
377 | pub(crate) whole: u64, |
378 | pub(crate) pair: Pair<u32>, |
379 | } |
380 | #[allow (dead_code)] |
381 | #[derive (Clone, Copy)] |
382 | #[repr (C)] |
383 | pub(crate) struct Pair<T: Copy> { |
384 | // little endian order |
385 | #[cfg (any( |
386 | target_endian = "little" , |
387 | target_arch = "aarch64" , |
388 | target_arch = "arm" , |
389 | target_arch = "arm64ec" , |
390 | ))] |
391 | pub(crate) lo: T, |
392 | pub(crate) hi: T, |
393 | // big endian order |
394 | #[cfg (not(any( |
395 | target_endian = "little" , |
396 | target_arch = "aarch64" , |
397 | target_arch = "arm" , |
398 | target_arch = "arm64ec" , |
399 | )))] |
400 | pub(crate) lo: T, |
401 | } |
402 | |
403 | #[cfg (any(target_arch = "riscv32" , target_arch = "riscv64" ))] |
404 | type MinWord = u32; |
405 | #[cfg (any(target_arch = "riscv32" , target_arch = "riscv64" ))] |
406 | type RetInt = u32; |
407 | // Adapted from https://github.com/taiki-e/atomic-maybe-uninit/blob/v0.3.4/src/utils.rs#L255. |
408 | // Helper for implementing sub-word atomic operations using word-sized LL/SC loop or CAS loop. |
409 | // |
410 | // Refs: https://github.com/llvm/llvm-project/blob/llvmorg-19.1.0/llvm/lib/CodeGen/AtomicExpandPass.cpp#L737 |
411 | // (aligned_ptr, shift, mask) |
412 | #[cfg (any(target_arch = "riscv32" , target_arch = "riscv64" ))] |
413 | #[allow (dead_code)] |
414 | #[inline ] |
415 | pub(crate) fn create_sub_word_mask_values<T>(ptr: *mut T) -> (*mut MinWord, RetInt, RetInt) { |
416 | use core::mem; |
417 | // RISC-V, MIPS, SPARC, LoongArch, Xtensa: shift amount of 32-bit shift instructions is 5 bits unsigned (0-31). |
418 | // PowerPC, C-SKY: shift amount of 32-bit shift instructions is 6 bits unsigned (0-63) and shift amount 32-63 means "clear". |
419 | // Arm: shift amount of 32-bit shift instructions is 8 bits unsigned (0-255). |
420 | // Hexagon: shift amount of 32-bit shift instructions is 7 bits signed (-64-63) and negative shift amount means "reverse the direction of the shift". |
421 | // (On s390x, we don't use the mask returned from this function.) |
422 | const SHIFT_MASK: bool = !cfg!(any( |
423 | target_arch = "loongarch64" , |
424 | target_arch = "mips" , |
425 | target_arch = "mips32r6" , |
426 | target_arch = "mips64" , |
427 | target_arch = "mips64r6" , |
428 | target_arch = "riscv32" , |
429 | target_arch = "riscv64" , |
430 | target_arch = "s390x" , |
431 | target_arch = "sparc" , |
432 | target_arch = "sparc64" , |
433 | target_arch = "xtensa" , |
434 | )); |
435 | let ptr_mask = mem::size_of::<MinWord>() - 1; |
436 | let aligned_ptr = strict::with_addr(ptr, ptr as usize & !ptr_mask) as *mut MinWord; |
437 | let ptr_lsb = if SHIFT_MASK { |
438 | ptr as usize & ptr_mask |
439 | } else { |
440 | // We use 32-bit wrapping shift instructions in asm on these platforms. |
441 | ptr as usize |
442 | }; |
443 | let shift = if cfg!(any(target_endian = "little" , target_arch = "s390x" )) { |
444 | ptr_lsb.wrapping_mul(8) |
445 | } else { |
446 | (ptr_lsb ^ (mem::size_of::<MinWord>() - mem::size_of::<T>())).wrapping_mul(8) |
447 | }; |
448 | let mut mask: RetInt = (1 << (mem::size_of::<T>() * 8)) - 1; // !(0 as T) as RetInt |
449 | if SHIFT_MASK { |
450 | mask <<= shift; |
451 | } |
452 | (aligned_ptr, shift as RetInt, mask) |
453 | } |
454 | |
455 | // TODO: use stabilized core::ptr strict_provenance helpers https://github.com/rust-lang/rust/pull/130350 |
456 | #[cfg (any(miri, target_arch = "riscv32" , target_arch = "riscv64" ))] |
457 | #[allow (dead_code)] |
458 | pub(crate) mod strict { |
459 | #[inline ] |
460 | #[must_use ] |
461 | pub(crate) fn with_addr<T>(ptr: *mut T, addr: usize) -> *mut T { |
462 | // This should probably be an intrinsic to avoid doing any sort of arithmetic, but |
463 | // meanwhile, we can implement it with `wrapping_offset`, which preserves the pointer's |
464 | // provenance. |
465 | let offset = addr.wrapping_sub(ptr as usize); |
466 | (ptr as *mut u8).wrapping_add(offset) as *mut T |
467 | } |
468 | |
469 | #[cfg (miri)] |
470 | #[inline ] |
471 | #[must_use ] |
472 | pub(crate) fn map_addr<T>(ptr: *mut T, f: impl FnOnce(usize) -> usize) -> *mut T { |
473 | with_addr(ptr, f(ptr as usize)) |
474 | } |
475 | } |
476 | |