std_detect/detect/
cache.rs

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::{AtomicUsize, Ordering};
7
8/// Sets the `bit` of `x`.
9#[inline]
10const fn set_bit(x: u128, bit: u32) -> u128 {
11    x | 1 << bit
12}
13
14/// Tests the `bit` of `x`.
15#[inline]
16const fn test_bit(x: u128, bit: u32) -> bool {
17    x & (1 << bit) != 0
18}
19
20/// Unset the `bit of `x`.
21#[inline]
22const fn unset_bit(x: u128, bit: u32) -> u128 {
23    x & !(1 << bit)
24}
25
26/// Maximum number of features that can be cached.
27const CACHE_CAPACITY: u32 = 93;
28
29/// This type is used to initialize the cache
30// The derived `Default` implementation will initialize the field to zero,
31// which is what we want.
32#[derive(Copy, Clone, Default, PartialEq, Eq)]
33pub(crate) struct Initializer(u128);
34
35// NOTE: the `debug_assert!` would catch that we do not add more Features than
36// the one fitting our cache.
37impl Initializer {
38    /// Tests the `bit` of the cache.
39    #[inline]
40    pub(crate) fn test(self, bit: u32) -> bool {
41        debug_assert!(bit < CACHE_CAPACITY, "too many features, time to increase the cache size!");
42        test_bit(self.0, bit)
43    }
44
45    /// Sets the `bit` of the cache.
46    #[inline]
47    pub(crate) fn set(&mut self, bit: u32) {
48        debug_assert!(bit < CACHE_CAPACITY, "too many features, time to increase the cache size!");
49        let v = self.0;
50        self.0 = set_bit(v, bit);
51    }
52
53    /// Unsets the `bit` of the cache.
54    #[inline]
55    pub(crate) fn unset(&mut self, bit: u32) {
56        debug_assert!(bit < CACHE_CAPACITY, "too many features, time to increase the cache size!");
57        let v = self.0;
58        self.0 = unset_bit(v, bit);
59    }
60}
61
62/// This global variable is a cache of the features supported by the CPU.
63// Note: the third slot is only used in x86
64// Another Slot can be added if needed without any change to `Initializer`
65static CACHE: [Cache; 3] = [Cache::uninitialized(), Cache::uninitialized(), Cache::uninitialized()];
66
67/// Feature cache with capacity for `size_of::<usize>() * 8 - 1` features.
68///
69/// Note: 0 is used to represent an uninitialized cache, and (at least) the most
70/// significant bit is set on any cache which has been initialized.
71///
72/// Note: we use `Relaxed` atomic operations, because we are only interested in
73/// the effects of operations on a single memory location. That is, we only need
74/// "modification order", and not the full-blown "happens before".
75struct Cache(AtomicUsize);
76
77impl Cache {
78    const CAPACITY: u32 = (core::mem::size_of::<usize>() * 8 - 1) as u32;
79    const MASK: usize = (1 << Cache::CAPACITY) - 1;
80    const INITIALIZED_BIT: usize = 1usize << Cache::CAPACITY;
81
82    /// Creates an uninitialized cache.
83    #[allow(clippy::declare_interior_mutable_const)]
84    const fn uninitialized() -> Self {
85        Cache(AtomicUsize::new(0))
86    }
87
88    /// Is the `bit` in the cache set? Returns `None` if the cache has not been initialized.
89    #[inline]
90    pub(crate) fn test(&self, bit: u32) -> Option<bool> {
91        let cached = self.0.load(Ordering::Relaxed);
92        if cached == 0 { None } else { Some(test_bit(cached as u128, bit)) }
93    }
94
95    /// Initializes the cache.
96    #[inline]
97    fn initialize(&self, value: usize) -> usize {
98        debug_assert_eq!((value & !Cache::MASK), 0);
99        self.0.store(value | Cache::INITIALIZED_BIT, Ordering::Relaxed);
100        value
101    }
102}
103
104cfg_if::cfg_if! {
105    if #[cfg(feature = "std_detect_env_override")] {
106        #[inline]
107        fn disable_features(disable: &[u8], value: &mut Initializer) {
108            if let Ok(disable) = core::str::from_utf8(disable) {
109                for v in disable.split(" ") {
110                    let _ = super::Feature::from_str(v).map(|v| value.unset(v as u32));
111                }
112            }
113        }
114
115        #[inline]
116        fn initialize(mut value: Initializer) -> Initializer {
117            use core::ffi::CStr;
118            const RUST_STD_DETECT_UNSTABLE: &CStr = c"RUST_STD_DETECT_UNSTABLE";
119            cfg_if::cfg_if! {
120                if #[cfg(windows)] {
121                    use alloc::vec;
122                    #[link(name = "kernel32")]
123                    unsafe extern "system" {
124                        fn GetEnvironmentVariableA(name: *const u8, buffer: *mut u8, size: u32) -> u32;
125                    }
126                    let len = unsafe { GetEnvironmentVariableA(RUST_STD_DETECT_UNSTABLE.as_ptr().cast::<u8>(), core::ptr::null_mut(), 0) };
127                    if len > 0 {
128                        // +1 to include the null terminator.
129                        let mut env = vec![0; len as usize + 1];
130                        let len = unsafe { GetEnvironmentVariableA(RUST_STD_DETECT_UNSTABLE.as_ptr().cast::<u8>(), env.as_mut_ptr(), len + 1) };
131                        if len > 0 {
132                            disable_features(&env[..len as usize], &mut value);
133                        }
134                    }
135                } else {
136                    let env = unsafe {
137                        libc::getenv(RUST_STD_DETECT_UNSTABLE.as_ptr())
138                    };
139                    if !env.is_null() {
140                        let len = unsafe { libc::strlen(env) };
141                        let env = unsafe { core::slice::from_raw_parts(env as *const u8, len) };
142                        disable_features(env, &mut value);
143                    }
144                }
145            }
146            do_initialize(value);
147            value
148        }
149    } else {
150        #[inline]
151        fn initialize(value: Initializer) -> Initializer {
152            do_initialize(value);
153            value
154        }
155    }
156}
157
158#[inline]
159fn do_initialize(value: Initializer) {
160    CACHE[0].initialize((value.0) as usize & Cache::MASK);
161    CACHE[1].initialize((value.0 >> Cache::CAPACITY) as usize & Cache::MASK);
162    CACHE[2].initialize((value.0 >> (2 * Cache::CAPACITY)) as usize & Cache::MASK);
163}
164
165// We only have to detect features once, and it's fairly costly, so hint to LLVM
166// that it should assume that cache hits are more common than misses (which is
167// the point of caching). It's possibly unfortunate that this function needs to
168// reach across modules like this to call `os::detect_features`, but it produces
169// the best code out of several attempted variants.
170//
171// The `Initializer` that the cache was initialized with is returned, so that
172// the caller can call `test()` on it without having to load the value from the
173// cache again.
174#[cold]
175fn detect_and_initialize() -> Initializer {
176    initialize(super::os::detect_features())
177}
178
179/// Tests the `bit` of the storage. If the storage has not been initialized,
180/// initializes it with the result of `os::detect_features()`.
181///
182/// On its first invocation, it detects the CPU features and caches them in the
183/// `CACHE` global variable as an `AtomicU64`.
184///
185/// It uses the `Feature` variant to index into this variable as a bitset. If
186/// the bit is set, the feature is enabled, and otherwise it is disabled.
187///
188/// If the feature `std_detect_env_override` is enabled looks for the env
189/// variable `RUST_STD_DETECT_UNSTABLE` and uses its content to disable
190/// Features that would had been otherwise detected.
191#[inline]
192pub(crate) fn test(bit: u32) -> bool {
193    let (relative_bit, idx) = if bit < Cache::CAPACITY {
194        (bit, 0)
195    } else if bit < 2 * Cache::CAPACITY {
196        (bit - Cache::CAPACITY, 1)
197    } else {
198        (bit - 2 * Cache::CAPACITY, 2)
199    };
200    CACHE[idx].test(relative_bit).unwrap_or_else(|| detect_and_initialize().test(bit))
201}
std_detect/detect/cache.rs

std_detect/detect/
cache.rs
