1//! Caches run-time feature detection so that it only needs to be computed
2//! once.
3
4#![allow(dead_code)] // not used on all platforms
5
6use core::sync::atomic::Ordering;
7
8use core::sync::atomic::AtomicUsize;
9
10/// Sets the `bit` of `x`.
11#[inline]
12const fn set_bit(x: u64, bit: u32) -> u64 {
13 x | 1 << bit
14}
15
16/// Tests the `bit` of `x`.
17#[inline]
18const fn test_bit(x: u64, bit: u32) -> bool {
19 x & (1 << bit) != 0
20}
21
22/// Unset the `bit of `x`.
23#[inline]
24const fn unset_bit(x: u64, bit: u32) -> u64 {
25 x & !(1 << bit)
26}
27
28/// Maximum number of features that can be cached.
29const CACHE_CAPACITY: u32 = 62;
30
31/// This type is used to initialize the cache
32// The derived `Default` implementation will initialize the field to zero,
33// which is what we want.
34#[derive(Copy, Clone, Default)]
35pub(crate) struct Initializer(u64);
36
37// NOTE: the `debug_assert!` would catch that we do not add more Features than
38// the one fitting our cache.
39impl Initializer {
40 /// Tests the `bit` of the cache.
41 #[inline]
42 pub(crate) fn test(self, bit: u32) -> bool {
43 debug_assert!(
44 bit < CACHE_CAPACITY,
45 "too many features, time to increase the cache size!"
46 );
47 test_bit(self.0, bit)
48 }
49
50 /// Sets the `bit` of the cache.
51 #[inline]
52 pub(crate) fn set(&mut self, bit: u32) {
53 debug_assert!(
54 bit < CACHE_CAPACITY,
55 "too many features, time to increase the cache size!"
56 );
57 let v = self.0;
58 self.0 = set_bit(v, bit);
59 }
60
61 /// Unsets the `bit` of the cache.
62 #[inline]
63 pub(crate) fn unset(&mut self, bit: u32) {
64 debug_assert!(
65 bit < CACHE_CAPACITY,
66 "too many features, time to increase the cache size!"
67 );
68 let v = self.0;
69 self.0 = unset_bit(v, bit);
70 }
71}
72
73/// This global variable is a cache of the features supported by the CPU.
74// Note: on x64, we only use the first slot
75static CACHE: [Cache; 2] = [Cache::uninitialized(), Cache::uninitialized()];
76
77/// Feature cache with capacity for `size_of::<usize::MAX>() * 8 - 1` features.
78///
79/// Note: 0 is used to represent an uninitialized cache, and (at least) the most
80/// significant bit is set on any cache which has been initialized.
81///
82/// Note: we use `Relaxed` atomic operations, because we are only interested in
83/// the effects of operations on a single memory location. That is, we only need
84/// "modification order", and not the full-blown "happens before".
85struct Cache(AtomicUsize);
86
87impl Cache {
88 const CAPACITY: u32 = (core::mem::size_of::<usize>() * 8 - 1) as u32;
89 const MASK: usize = (1 << Cache::CAPACITY) - 1;
90 const INITIALIZED_BIT: usize = 1usize << Cache::CAPACITY;
91
92 /// Creates an uninitialized cache.
93 #[allow(clippy::declare_interior_mutable_const)]
94 const fn uninitialized() -> Self {
95 Cache(AtomicUsize::new(0))
96 }
97
98 /// Is the `bit` in the cache set? Returns `None` if the cache has not been initialized.
99 #[inline]
100 pub(crate) fn test(&self, bit: u32) -> Option<bool> {
101 let cached = self.0.load(Ordering::Relaxed);
102 if cached == 0 {
103 None
104 } else {
105 Some(test_bit(cached as u64, bit))
106 }
107 }
108
109 /// Initializes the cache.
110 #[inline]
111 fn initialize(&self, value: usize) -> usize {
112 debug_assert_eq!((value & !Cache::MASK), 0);
113 self.0
114 .store(value | Cache::INITIALIZED_BIT, Ordering::Relaxed);
115 value
116 }
117}
118
119cfg_if::cfg_if! {
120 if #[cfg(feature = "std_detect_env_override")] {
121 #[inline]
122 fn initialize(mut value: Initializer) -> Initializer {
123 let env = unsafe {
124 libc::getenv(b"RUST_STD_DETECT_UNSTABLE\0".as_ptr() as *const libc::c_char)
125 };
126 if !env.is_null() {
127 let len = unsafe { libc::strlen(env) };
128 let env = unsafe { core::slice::from_raw_parts(env as *const u8, len) };
129 if let Ok(disable) = core::str::from_utf8(env) {
130 for v in disable.split(" ") {
131 let _ = super::Feature::from_str(v).map(|v| value.unset(v as u32));
132 }
133 }
134 }
135 do_initialize(value);
136 value
137 }
138 } else {
139 #[inline]
140 fn initialize(value: Initializer) -> Initializer {
141 do_initialize(value);
142 value
143 }
144 }
145}
146
147#[inline]
148fn do_initialize(value: Initializer) {
149 CACHE[0].initialize((value.0) as usize & Cache::MASK);
150 CACHE[1].initialize((value.0 >> Cache::CAPACITY) as usize & Cache::MASK);
151}
152
153// We only have to detect features once, and it's fairly costly, so hint to LLVM
154// that it should assume that cache hits are more common than misses (which is
155// the point of caching). It's possibly unfortunate that this function needs to
156// reach across modules like this to call `os::detect_features`, but it produces
157// the best code out of several attempted variants.
158//
159// The `Initializer` that the cache was initialized with is returned, so that
160// the caller can call `test()` on it without having to load the value from the
161// cache again.
162#[cold]
163fn detect_and_initialize() -> Initializer {
164 initialize(super::os::detect_features())
165}
166
167/// Tests the `bit` of the storage. If the storage has not been initialized,
168/// initializes it with the result of `os::detect_features()`.
169///
170/// On its first invocation, it detects the CPU features and caches them in the
171/// `CACHE` global variable as an `AtomicU64`.
172///
173/// It uses the `Feature` variant to index into this variable as a bitset. If
174/// the bit is set, the feature is enabled, and otherwise it is disabled.
175///
176/// If the feature `std_detect_env_override` is enabled looks for the env
177/// variable `RUST_STD_DETECT_UNSTABLE` and uses its content to disable
178/// Features that would had been otherwise detected.
179#[inline]
180pub(crate) fn test(bit: u32) -> bool {
181 let (relative_bit: u32, idx: i32) = if bit < Cache::CAPACITY {
182 (bit, 0)
183 } else {
184 (bit - Cache::CAPACITY, 1)
185 };
186 CACHE[idx]
187 .test(relative_bit)
188 .unwrap_or_else(|| detect_and_initialize().test(bit))
189}
190