| 1 | //! Manually manage memory through raw pointers. |
| 2 | //! |
| 3 | //! *[See also the pointer primitive types](pointer).* |
| 4 | //! |
| 5 | //! # Safety |
| 6 | //! |
| 7 | //! Many functions in this module take raw pointers as arguments and read from or write to them. For |
| 8 | //! this to be safe, these pointers must be *valid* for the given access. Whether a pointer is valid |
| 9 | //! depends on the operation it is used for (read or write), and the extent of the memory that is |
| 10 | //! accessed (i.e., how many bytes are read/written) -- it makes no sense to ask "is this pointer |
| 11 | //! valid"; one has to ask "is this pointer valid for a given access". Most functions use `*mut T` |
| 12 | //! and `*const T` to access only a single value, in which case the documentation omits the size and |
| 13 | //! implicitly assumes it to be `size_of::<T>()` bytes. |
| 14 | //! |
| 15 | //! The precise rules for validity are not determined yet. The guarantees that are |
| 16 | //! provided at this point are very minimal: |
| 17 | //! |
| 18 | //! * For memory accesses of [size zero][zst], *every* pointer is valid, including the [null] |
| 19 | //! pointer. The following points are only concerned with non-zero-sized accesses. |
| 20 | //! * A [null] pointer is *never* valid. |
| 21 | //! * For a pointer to be valid, it is necessary, but not always sufficient, that the pointer be |
| 22 | //! *dereferenceable*. The [provenance] of the pointer is used to determine which [allocation] |
| 23 | //! it is derived from; a pointer is dereferenceable if the memory range of the given size |
| 24 | //! starting at the pointer is entirely contained within the bounds of that allocation. Note |
| 25 | //! that in Rust, every (stack-allocated) variable is considered a separate allocation. |
| 26 | //! * All accesses performed by functions in this module are *non-atomic* in the sense |
| 27 | //! of [atomic operations] used to synchronize between threads. This means it is |
| 28 | //! undefined behavior to perform two concurrent accesses to the same location from different |
| 29 | //! threads unless both accesses only read from memory. Notice that this explicitly |
| 30 | //! includes [`read_volatile`] and [`write_volatile`]: Volatile accesses cannot |
| 31 | //! be used for inter-thread synchronization. |
| 32 | //! * The result of casting a reference to a pointer is valid for as long as the |
| 33 | //! underlying allocation is live and no reference (just raw pointers) is used to |
| 34 | //! access the same memory. That is, reference and pointer accesses cannot be |
| 35 | //! interleaved. |
| 36 | //! |
| 37 | //! These axioms, along with careful use of [`offset`] for pointer arithmetic, |
| 38 | //! are enough to correctly implement many useful things in unsafe code. Stronger guarantees |
| 39 | //! will be provided eventually, as the [aliasing] rules are being determined. For more |
| 40 | //! information, see the [book] as well as the section in the reference devoted |
| 41 | //! to [undefined behavior][ub]. |
| 42 | //! |
| 43 | //! We say that a pointer is "dangling" if it is not valid for any non-zero-sized accesses. This |
| 44 | //! means out-of-bounds pointers, pointers to freed memory, null pointers, and pointers created with |
| 45 | //! [`NonNull::dangling`] are all dangling. |
| 46 | //! |
| 47 | //! ## Alignment |
| 48 | //! |
| 49 | //! Valid raw pointers as defined above are not necessarily properly aligned (where |
| 50 | //! "proper" alignment is defined by the pointee type, i.e., `*const T` must be |
| 51 | //! aligned to `align_of::<T>()`). However, most functions require their |
| 52 | //! arguments to be properly aligned, and will explicitly state |
| 53 | //! this requirement in their documentation. Notable exceptions to this are |
| 54 | //! [`read_unaligned`] and [`write_unaligned`]. |
| 55 | //! |
| 56 | //! When a function requires proper alignment, it does so even if the access |
| 57 | //! has size 0, i.e., even if memory is not actually touched. Consider using |
| 58 | //! [`NonNull::dangling`] in such cases. |
| 59 | //! |
| 60 | //! ## Pointer to reference conversion |
| 61 | //! |
| 62 | //! When converting a pointer to a reference (e.g. via `&*ptr` or `&mut *ptr`), |
| 63 | //! there are several rules that must be followed: |
| 64 | //! |
| 65 | //! * The pointer must be properly aligned. |
| 66 | //! |
| 67 | //! * It must be non-null. |
| 68 | //! |
| 69 | //! * It must be "dereferenceable" in the sense defined above. |
| 70 | //! |
| 71 | //! * The pointer must point to a [valid value] of type `T`. |
| 72 | //! |
| 73 | //! * You must enforce Rust's aliasing rules. The exact aliasing rules are not decided yet, so we |
| 74 | //! only give a rough overview here. The rules also depend on whether a mutable or a shared |
| 75 | //! reference is being created. |
| 76 | //! * When creating a mutable reference, then while this reference exists, the memory it points to |
| 77 | //! must not get accessed (read or written) through any other pointer or reference not derived |
| 78 | //! from this reference. |
| 79 | //! * When creating a shared reference, then while this reference exists, the memory it points to |
| 80 | //! must not get mutated (except inside `UnsafeCell`). |
| 81 | //! |
| 82 | //! If a pointer follows all of these rules, it is said to be |
| 83 | //! *convertible to a (mutable or shared) reference*. |
| 84 | // ^ we use this term instead of saying that the produced reference must |
| 85 | // be valid, as the validity of a reference is easily confused for the |
| 86 | // validity of the thing it refers to, and while the two concepts are |
| 87 | // closely related, they are not identical. |
| 88 | //! |
| 89 | //! These rules apply even if the result is unused! |
| 90 | //! (The part about being initialized is not yet fully decided, but until |
| 91 | //! it is, the only safe approach is to ensure that they are indeed initialized.) |
| 92 | //! |
| 93 | //! An example of the implications of the above rules is that an expression such |
| 94 | //! as `unsafe { &*(0 as *const u8) }` is Immediate Undefined Behavior. |
| 95 | //! |
| 96 | //! [valid value]: ../../reference/behavior-considered-undefined.html#invalid-values |
| 97 | //! |
| 98 | //! ## Allocation |
| 99 | //! |
| 100 | //! <a id="allocated-object"></a> <!-- keep old URLs working --> |
| 101 | //! |
| 102 | //! An *allocation* is a subset of program memory which is addressable |
| 103 | //! from Rust, and within which pointer arithmetic is possible. Examples of |
| 104 | //! allocations include heap allocations, stack-allocated variables, |
| 105 | //! statics, and consts. The safety preconditions of some Rust operations - |
| 106 | //! such as `offset` and field projections (`expr.field`) - are defined in |
| 107 | //! terms of the allocations on which they operate. |
| 108 | //! |
| 109 | //! An allocation has a base address, a size, and a set of memory |
| 110 | //! addresses. It is possible for an allocation to have zero size, but |
| 111 | //! such an allocation will still have a base address. The base address |
| 112 | //! of an allocation is not necessarily unique. While it is currently the |
| 113 | //! case that an allocation always has a set of memory addresses which is |
| 114 | //! fully contiguous (i.e., has no "holes"), there is no guarantee that this |
| 115 | //! will not change in the future. |
| 116 | //! |
| 117 | //! For any allocation with `base` address, `size`, and a set of |
| 118 | //! `addresses`, the following are guaranteed: |
| 119 | //! - For all addresses `a` in `addresses`, `a` is in the range `base .. (base + |
| 120 | //! size)` (note that this requires `a < base + size`, not `a <= base + size`) |
| 121 | //! - `base` is not equal to [`null()`] (i.e., the address with the numerical |
| 122 | //! value 0) |
| 123 | //! - `base + size <= usize::MAX` |
| 124 | //! - `size <= isize::MAX` |
| 125 | //! |
| 126 | //! As a consequence of these guarantees, given any address `a` within the set |
| 127 | //! of addresses of an allocation: |
| 128 | //! - It is guaranteed that `a - base` does not overflow `isize` |
| 129 | //! - It is guaranteed that `a - base` is non-negative |
| 130 | //! - It is guaranteed that, given `o = a - base` (i.e., the offset of `a` within |
| 131 | //! the allocation), `base + o` will not wrap around the address space (in |
| 132 | //! other words, will not overflow `usize`) |
| 133 | //! |
| 134 | //! [`null()`]: null |
| 135 | //! |
| 136 | //! # Provenance |
| 137 | //! |
| 138 | //! Pointers are not *simply* an "integer" or "address". For instance, it's uncontroversial |
| 139 | //! to say that a Use After Free is clearly Undefined Behavior, even if you "get lucky" |
| 140 | //! and the freed memory gets reallocated before your read/write (in fact this is the |
| 141 | //! worst-case scenario, UAFs would be much less concerning if this didn't happen!). |
| 142 | //! As another example, consider that [`wrapping_offset`] is documented to "remember" |
| 143 | //! the allocation that the original pointer points to, even if it is offset far |
| 144 | //! outside the memory range occupied by that allocation. |
| 145 | //! To rationalize claims like this, pointers need to somehow be *more* than just their addresses: |
| 146 | //! they must have **provenance**. |
| 147 | //! |
| 148 | //! A pointer value in Rust semantically contains the following information: |
| 149 | //! |
| 150 | //! * The **address** it points to, which can be represented by a `usize`. |
| 151 | //! * The **provenance** it has, defining the memory it has permission to access. Provenance can be |
| 152 | //! absent, in which case the pointer does not have permission to access any memory. |
| 153 | //! |
| 154 | //! The exact structure of provenance is not yet specified, but the permission defined by a |
| 155 | //! pointer's provenance have a *spatial* component, a *temporal* component, and a *mutability* |
| 156 | //! component: |
| 157 | //! |
| 158 | //! * Spatial: The set of memory addresses that the pointer is allowed to access. |
| 159 | //! * Temporal: The timespan during which the pointer is allowed to access those memory addresses. |
| 160 | //! * Mutability: Whether the pointer may only access the memory for reads, or also access it for |
| 161 | //! writes. Note that this can interact with the other components, e.g. a pointer might permit |
| 162 | //! mutation only for a subset of addresses, or only for a subset of its maximal timespan. |
| 163 | //! |
| 164 | //! When an [allocation] is created, it has a unique Original Pointer. For alloc |
| 165 | //! APIs this is literally the pointer the call returns, and for local variables and statics, |
| 166 | //! this is the name of the variable/static. (This is mildly overloading the term "pointer" |
| 167 | //! for the sake of brevity/exposition.) |
| 168 | //! |
| 169 | //! The Original Pointer for an allocation has provenance that constrains the *spatial* |
| 170 | //! permissions of this pointer to the memory range of the allocation, and the *temporal* |
| 171 | //! permissions to the lifetime of the allocation. Provenance is implicitly inherited by all |
| 172 | //! pointers transitively derived from the Original Pointer through operations like [`offset`], |
| 173 | //! borrowing, and pointer casts. Some operations may *shrink* the permissions of the derived |
| 174 | //! provenance, limiting how much memory it can access or how long it's valid for (i.e. borrowing a |
| 175 | //! subfield and subslicing can shrink the spatial component of provenance, and all borrowing can |
| 176 | //! shrink the temporal component of provenance). However, no operation can ever *grow* the |
| 177 | //! permissions of the derived provenance: even if you "know" there is a larger allocation, you |
| 178 | //! can't derive a pointer with a larger provenance. Similarly, you cannot "recombine" two |
| 179 | //! contiguous provenances back into one (i.e. with a `fn merge(&[T], &[T]) -> &[T]`). |
| 180 | //! |
| 181 | //! A reference to a place always has provenance over at least the memory that place occupies. |
| 182 | //! A reference to a slice always has provenance over at least the range that slice describes. |
| 183 | //! Whether and when exactly the provenance of a reference gets "shrunk" to *exactly* fit |
| 184 | //! the memory it points to is not yet determined. |
| 185 | //! |
| 186 | //! A *shared* reference only ever has provenance that permits reading from memory, |
| 187 | //! and never permits writes, except inside [`UnsafeCell`]. |
| 188 | //! |
| 189 | //! Provenance can affect whether a program has undefined behavior: |
| 190 | //! |
| 191 | //! * It is undefined behavior to access memory through a pointer that does not have provenance over |
| 192 | //! that memory. Note that a pointer "at the end" of its provenance is not actually outside its |
| 193 | //! provenance, it just has 0 bytes it can load/store. Zero-sized accesses do not require any |
| 194 | //! provenance since they access an empty range of memory. |
| 195 | //! |
| 196 | //! * It is undefined behavior to [`offset`] a pointer across a memory range that is not contained |
| 197 | //! in the allocation it is derived from, or to [`offset_from`] two pointers not derived |
| 198 | //! from the same allocation. Provenance is used to say what exactly "derived from" even |
| 199 | //! means: the lineage of a pointer is traced back to the Original Pointer it descends from, and |
| 200 | //! that identifies the relevant allocation. In particular, it's always UB to offset a |
| 201 | //! pointer derived from something that is now deallocated, except if the offset is 0. |
| 202 | //! |
| 203 | //! But it *is* still sound to: |
| 204 | //! |
| 205 | //! * Create a pointer without provenance from just an address (see [`without_provenance`]). Such a |
| 206 | //! pointer cannot be used for memory accesses (except for zero-sized accesses). This can still be |
| 207 | //! useful for sentinel values like `null` *or* to represent a tagged pointer that will never be |
| 208 | //! dereferenceable. In general, it is always sound for an integer to pretend to be a pointer "for |
| 209 | //! fun" as long as you don't use operations on it which require it to be valid (non-zero-sized |
| 210 | //! offset, read, write, etc). |
| 211 | //! |
| 212 | //! * Forge an allocation of size zero at any sufficiently aligned non-null address. |
| 213 | //! i.e. the usual "ZSTs are fake, do what you want" rules apply. |
| 214 | //! |
| 215 | //! * [`wrapping_offset`] a pointer outside its provenance. This includes pointers |
| 216 | //! which have "no" provenance. In particular, this makes it sound to do pointer tagging tricks. |
| 217 | //! |
| 218 | //! * Compare arbitrary pointers by address. Pointer comparison ignores provenance and addresses |
| 219 | //! *are* just integers, so there is always a coherent answer, even if the pointers are dangling |
| 220 | //! or from different provenances. Note that if you get "lucky" and notice that a pointer at the |
| 221 | //! end of one allocation is the "same" address as the start of another allocation, |
| 222 | //! anything you do with that fact is *probably* going to be gibberish. The scope of that |
| 223 | //! gibberish is kept under control by the fact that the two pointers *still* aren't allowed to |
| 224 | //! access the other's allocation (bytes), because they still have different provenance. |
| 225 | //! |
| 226 | //! Note that the full definition of provenance in Rust is not decided yet, as this interacts |
| 227 | //! with the as-yet undecided [aliasing] rules. |
| 228 | //! |
| 229 | //! ## Pointers Vs Integers |
| 230 | //! |
| 231 | //! From this discussion, it becomes very clear that a `usize` *cannot* accurately represent a pointer, |
| 232 | //! and converting from a pointer to a `usize` is generally an operation which *only* extracts the |
| 233 | //! address. Converting this address back into pointer requires somehow answering the question: |
| 234 | //! which provenance should the resulting pointer have? |
| 235 | //! |
| 236 | //! Rust provides two ways of dealing with this situation: *Strict Provenance* and *Exposed Provenance*. |
| 237 | //! |
| 238 | //! Note that a pointer *can* represent a `usize` (via [`without_provenance`]), so the right type to |
| 239 | //! use in situations where a value is "sometimes a pointer and sometimes a bare `usize`" is a |
| 240 | //! pointer type. |
| 241 | //! |
| 242 | //! ## Strict Provenance |
| 243 | //! |
| 244 | //! "Strict Provenance" refers to a set of APIs designed to make working with provenance more |
| 245 | //! explicit. They are intended as substitutes for casting a pointer to an integer and back. |
| 246 | //! |
| 247 | //! Entirely avoiding integer-to-pointer casts successfully side-steps the inherent ambiguity of |
| 248 | //! that operation. This benefits compiler optimizations, and it is pretty much a requirement for |
| 249 | //! using tools like [Miri] and architectures like [CHERI] that aim to detect and diagnose pointer |
| 250 | //! misuse. |
| 251 | //! |
| 252 | //! The key insight to making programming without integer-to-pointer casts *at all* viable is the |
| 253 | //! [`with_addr`] method: |
| 254 | //! |
| 255 | //! ```text |
| 256 | //! /// Creates a new pointer with the given address. |
| 257 | //! /// |
| 258 | //! /// This performs the same operation as an `addr as ptr` cast, but copies |
| 259 | //! /// the *provenance* of `self` to the new pointer. |
| 260 | //! /// This allows us to dynamically preserve and propagate this important |
| 261 | //! /// information in a way that is otherwise impossible with a unary cast. |
| 262 | //! /// |
| 263 | //! /// This is equivalent to using `wrapping_offset` to offset `self` to the |
| 264 | //! /// given address, and therefore has all the same capabilities and restrictions. |
| 265 | //! pub fn with_addr(self, addr: usize) -> Self; |
| 266 | //! ``` |
| 267 | //! |
| 268 | //! So you're still able to drop down to the address representation and do whatever |
| 269 | //! clever bit tricks you want *as long as* you're able to keep around a pointer |
| 270 | //! into the allocation you care about that can "reconstitute" the provenance. |
| 271 | //! Usually this is very easy, because you only are taking a pointer, messing with the address, |
| 272 | //! and then immediately converting back to a pointer. To make this use case more ergonomic, |
| 273 | //! we provide the [`map_addr`] method. |
| 274 | //! |
| 275 | //! To help make it clear that code is "following" Strict Provenance semantics, we also provide an |
| 276 | //! [`addr`] method which promises that the returned address is not part of a |
| 277 | //! pointer-integer-pointer roundtrip. In the future we may provide a lint for pointer<->integer |
| 278 | //! casts to help you audit if your code conforms to strict provenance. |
| 279 | //! |
| 280 | //! ### Using Strict Provenance |
| 281 | //! |
| 282 | //! Most code needs no changes to conform to strict provenance, as the only really concerning |
| 283 | //! operation is casts from `usize` to a pointer. For code which *does* cast a `usize` to a pointer, |
| 284 | //! the scope of the change depends on exactly what you're doing. |
| 285 | //! |
| 286 | //! In general, you just need to make sure that if you want to convert a `usize` address to a |
| 287 | //! pointer and then use that pointer to read/write memory, you need to keep around a pointer |
| 288 | //! that has sufficient provenance to perform that read/write itself. In this way all of your |
| 289 | //! casts from an address to a pointer are essentially just applying offsets/indexing. |
| 290 | //! |
| 291 | //! This is generally trivial to do for simple cases like tagged pointers *as long as you |
| 292 | //! represent the tagged pointer as an actual pointer and not a `usize`*. For instance: |
| 293 | //! |
| 294 | //! ``` |
| 295 | //! unsafe { |
| 296 | //! // A flag we want to pack into our pointer |
| 297 | //! static HAS_DATA: usize = 0x1; |
| 298 | //! static FLAG_MASK: usize = !HAS_DATA; |
| 299 | //! |
| 300 | //! // Our value, which must have enough alignment to have spare least-significant-bits. |
| 301 | //! let my_precious_data: u32 = 17; |
| 302 | //! assert!(align_of::<u32>() > 1); |
| 303 | //! |
| 304 | //! // Create a tagged pointer |
| 305 | //! let ptr = &my_precious_data as *const u32; |
| 306 | //! let tagged = ptr.map_addr(|addr| addr | HAS_DATA); |
| 307 | //! |
| 308 | //! // Check the flag: |
| 309 | //! if tagged.addr() & HAS_DATA != 0 { |
| 310 | //! // Untag and read the pointer |
| 311 | //! let data = *tagged.map_addr(|addr| addr & FLAG_MASK); |
| 312 | //! assert_eq!(data, 17); |
| 313 | //! } else { |
| 314 | //! unreachable!() |
| 315 | //! } |
| 316 | //! } |
| 317 | //! ``` |
| 318 | //! |
| 319 | //! (Yes, if you've been using [`AtomicUsize`] for pointers in concurrent datastructures, you should |
| 320 | //! be using [`AtomicPtr`] instead. If that messes up the way you atomically manipulate pointers, |
| 321 | //! we would like to know why, and what needs to be done to fix it.) |
| 322 | //! |
| 323 | //! Situations where a valid pointer *must* be created from just an address, such as baremetal code |
| 324 | //! accessing a memory-mapped interface at a fixed address, cannot currently be handled with strict |
| 325 | //! provenance APIs and should use [exposed provenance](#exposed-provenance). |
| 326 | //! |
| 327 | //! ## Exposed Provenance |
| 328 | //! |
| 329 | //! As discussed above, integer-to-pointer casts are not possible with Strict Provenance APIs. |
| 330 | //! This is by design: the goal of Strict Provenance is to provide a clear specification that we are |
| 331 | //! confident can be formalized unambiguously and can be subject to precise formal reasoning. |
| 332 | //! Integer-to-pointer casts do not (currently) have such a clear specification. |
| 333 | //! |
| 334 | //! However, there exist situations where integer-to-pointer casts cannot be avoided, or |
| 335 | //! where avoiding them would require major refactoring. Legacy platform APIs also regularly assume |
| 336 | //! that `usize` can capture all the information that makes up a pointer. |
| 337 | //! Bare-metal platforms can also require the synthesis of a pointer "out of thin air" without |
| 338 | //! anywhere to obtain proper provenance from. |
| 339 | //! |
| 340 | //! Rust's model for dealing with integer-to-pointer casts is called *Exposed Provenance*. However, |
| 341 | //! the semantics of Exposed Provenance are on much less solid footing than Strict Provenance, and |
| 342 | //! at this point it is not yet clear whether a satisfying unambiguous semantics can be defined for |
| 343 | //! Exposed Provenance. (If that sounds bad, be reassured that other popular languages that provide |
| 344 | //! integer-to-pointer casts are not faring any better.) Furthermore, Exposed Provenance will not |
| 345 | //! work (well) with tools like [Miri] and [CHERI]. |
| 346 | //! |
| 347 | //! Exposed Provenance is provided by the [`expose_provenance`] and [`with_exposed_provenance`] methods, |
| 348 | //! which are equivalent to `as` casts between pointers and integers. |
| 349 | //! - [`expose_provenance`] is a lot like [`addr`], but additionally adds the provenance of the |
| 350 | //! pointer to a global list of 'exposed' provenances. (This list is purely conceptual, it exists |
| 351 | //! for the purpose of specifying Rust but is not materialized in actual executions, except in |
| 352 | //! tools like [Miri].) |
| 353 | //! Memory which is outside the control of the Rust abstract machine (MMIO registers, for example) |
| 354 | //! is always considered to be exposed, so long as this memory is disjoint from memory that will |
| 355 | //! be used by the abstract machine such as the stack, heap, and statics. |
| 356 | //! - [`with_exposed_provenance`] can be used to construct a pointer with one of these previously |
| 357 | //! 'exposed' provenances. [`with_exposed_provenance`] takes only `addr: usize` as arguments, so |
| 358 | //! unlike in [`with_addr`] there is no indication of what the correct provenance for the returned |
| 359 | //! pointer is -- and that is exactly what makes integer-to-pointer casts so tricky to rigorously |
| 360 | //! specify! The compiler will do its best to pick the right provenance for you, but currently we |
| 361 | //! cannot provide any guarantees about which provenance the resulting pointer will have. Only one |
| 362 | //! thing is clear: if there is *no* previously 'exposed' provenance that justifies the way the |
| 363 | //! returned pointer will be used, the program has undefined behavior. |
| 364 | //! |
| 365 | //! If at all possible, we encourage code to be ported to [Strict Provenance] APIs, thus avoiding |
| 366 | //! the need for Exposed Provenance. Maximizing the amount of such code is a major win for avoiding |
| 367 | //! specification complexity and to facilitate adoption of tools like [CHERI] and [Miri] that can be |
| 368 | //! a big help in increasing the confidence in (unsafe) Rust code. However, we acknowledge that this |
| 369 | //! is not always possible, and offer Exposed Provenance as a way to explicit "opt out" of the |
| 370 | //! well-defined semantics of Strict Provenance, and "opt in" to the unclear semantics of |
| 371 | //! integer-to-pointer casts. |
| 372 | //! |
| 373 | //! [aliasing]: ../../nomicon/aliasing.html |
| 374 | //! [allocation]: #allocation |
| 375 | //! [provenance]: #provenance |
| 376 | //! [book]: ../../book/ch19-01-unsafe-rust.html#dereferencing-a-raw-pointer |
| 377 | //! [ub]: ../../reference/behavior-considered-undefined.html |
| 378 | //! [zst]: ../../nomicon/exotic-sizes.html#zero-sized-types-zsts |
| 379 | //! [atomic operations]: crate::sync::atomic |
| 380 | //! [`offset`]: pointer::offset |
| 381 | //! [`offset_from`]: pointer::offset_from |
| 382 | //! [`wrapping_offset`]: pointer::wrapping_offset |
| 383 | //! [`with_addr`]: pointer::with_addr |
| 384 | //! [`map_addr`]: pointer::map_addr |
| 385 | //! [`addr`]: pointer::addr |
| 386 | //! [`AtomicUsize`]: crate::sync::atomic::AtomicUsize |
| 387 | //! [`AtomicPtr`]: crate::sync::atomic::AtomicPtr |
| 388 | //! [`expose_provenance`]: pointer::expose_provenance |
| 389 | //! [`with_exposed_provenance`]: with_exposed_provenance |
| 390 | //! [Miri]: https://github.com/rust-lang/miri |
| 391 | //! [CHERI]: https://www.cl.cam.ac.uk/research/security/ctsrd/cheri/ |
| 392 | //! [Strict Provenance]: #strict-provenance |
| 393 | //! [`UnsafeCell`]: core::cell::UnsafeCell |
| 394 | |
| 395 | #![stable (feature = "rust1" , since = "1.0.0" )] |
| 396 | // There are many unsafe functions taking pointers that don't dereference them. |
| 397 | #![allow (clippy::not_unsafe_ptr_arg_deref)] |
| 398 | |
| 399 | use crate::cmp::Ordering; |
| 400 | use crate::intrinsics::const_eval_select; |
| 401 | use crate::marker::FnPtr; |
| 402 | use crate::mem::{self, MaybeUninit, SizedTypeProperties}; |
| 403 | use crate::num::NonZero; |
| 404 | use crate::{fmt, hash, intrinsics, ub_checks}; |
| 405 | |
| 406 | mod alignment; |
| 407 | #[unstable (feature = "ptr_alignment_type" , issue = "102070" )] |
| 408 | pub use alignment::Alignment; |
| 409 | |
| 410 | mod metadata; |
| 411 | #[unstable (feature = "ptr_metadata" , issue = "81513" )] |
| 412 | pub use metadata::{DynMetadata, Pointee, Thin, from_raw_parts, from_raw_parts_mut, metadata}; |
| 413 | |
| 414 | mod non_null; |
| 415 | #[stable (feature = "nonnull" , since = "1.25.0" )] |
| 416 | pub use non_null::NonNull; |
| 417 | |
| 418 | mod unique; |
| 419 | #[unstable (feature = "ptr_internals" , issue = "none" )] |
| 420 | pub use unique::Unique; |
| 421 | |
| 422 | mod const_ptr; |
| 423 | mod mut_ptr; |
| 424 | |
| 425 | // Some functions are defined here because they accidentally got made |
| 426 | // available in this module on stable. See <https://github.com/rust-lang/rust/issues/15702>. |
| 427 | // (`transmute` also falls into this category, but it cannot be wrapped due to the |
| 428 | // check that `T` and `U` have the same size.) |
| 429 | |
| 430 | /// Copies `count * size_of::<T>()` bytes from `src` to `dst`. The source |
| 431 | /// and destination must *not* overlap. |
| 432 | /// |
| 433 | /// For regions of memory which might overlap, use [`copy`] instead. |
| 434 | /// |
| 435 | /// `copy_nonoverlapping` is semantically equivalent to C's [`memcpy`], but |
| 436 | /// with the source and destination arguments swapped, |
| 437 | /// and `count` counting the number of `T`s instead of bytes. |
| 438 | /// |
| 439 | /// The copy is "untyped" in the sense that data may be uninitialized or otherwise violate the |
| 440 | /// requirements of `T`. The initialization state is preserved exactly. |
| 441 | /// |
| 442 | /// [`memcpy`]: https://en.cppreference.com/w/c/string/byte/memcpy |
| 443 | /// |
| 444 | /// # Safety |
| 445 | /// |
| 446 | /// Behavior is undefined if any of the following conditions are violated: |
| 447 | /// |
| 448 | /// * `src` must be [valid] for reads of `count * size_of::<T>()` bytes. |
| 449 | /// |
| 450 | /// * `dst` must be [valid] for writes of `count * size_of::<T>()` bytes. |
| 451 | /// |
| 452 | /// * Both `src` and `dst` must be properly aligned. |
| 453 | /// |
| 454 | /// * The region of memory beginning at `src` with a size of `count * |
| 455 | /// size_of::<T>()` bytes must *not* overlap with the region of memory |
| 456 | /// beginning at `dst` with the same size. |
| 457 | /// |
| 458 | /// Like [`read`], `copy_nonoverlapping` creates a bitwise copy of `T`, regardless of |
| 459 | /// whether `T` is [`Copy`]. If `T` is not [`Copy`], using *both* the values |
| 460 | /// in the region beginning at `*src` and the region beginning at `*dst` can |
| 461 | /// [violate memory safety][read-ownership]. |
| 462 | /// |
| 463 | /// Note that even if the effectively copied size (`count * size_of::<T>()`) is |
| 464 | /// `0`, the pointers must be properly aligned. |
| 465 | /// |
| 466 | /// [`read`]: crate::ptr::read |
| 467 | /// [read-ownership]: crate::ptr::read#ownership-of-the-returned-value |
| 468 | /// [valid]: crate::ptr#safety |
| 469 | /// |
| 470 | /// # Examples |
| 471 | /// |
| 472 | /// Manually implement [`Vec::append`]: |
| 473 | /// |
| 474 | /// ``` |
| 475 | /// use std::ptr; |
| 476 | /// |
| 477 | /// /// Moves all the elements of `src` into `dst`, leaving `src` empty. |
| 478 | /// fn append<T>(dst: &mut Vec<T>, src: &mut Vec<T>) { |
| 479 | /// let src_len = src.len(); |
| 480 | /// let dst_len = dst.len(); |
| 481 | /// |
| 482 | /// // Ensure that `dst` has enough capacity to hold all of `src`. |
| 483 | /// dst.reserve(src_len); |
| 484 | /// |
| 485 | /// unsafe { |
| 486 | /// // The call to add is always safe because `Vec` will never |
| 487 | /// // allocate more than `isize::MAX` bytes. |
| 488 | /// let dst_ptr = dst.as_mut_ptr().add(dst_len); |
| 489 | /// let src_ptr = src.as_ptr(); |
| 490 | /// |
| 491 | /// // Truncate `src` without dropping its contents. We do this first, |
| 492 | /// // to avoid problems in case something further down panics. |
| 493 | /// src.set_len(0); |
| 494 | /// |
| 495 | /// // The two regions cannot overlap because mutable references do |
| 496 | /// // not alias, and two different vectors cannot own the same |
| 497 | /// // memory. |
| 498 | /// ptr::copy_nonoverlapping(src_ptr, dst_ptr, src_len); |
| 499 | /// |
| 500 | /// // Notify `dst` that it now holds the contents of `src`. |
| 501 | /// dst.set_len(dst_len + src_len); |
| 502 | /// } |
| 503 | /// } |
| 504 | /// |
| 505 | /// let mut a = vec!['r' ]; |
| 506 | /// let mut b = vec!['u' , 's' , 't' ]; |
| 507 | /// |
| 508 | /// append(&mut a, &mut b); |
| 509 | /// |
| 510 | /// assert_eq!(a, &['r' , 'u' , 's' , 't' ]); |
| 511 | /// assert!(b.is_empty()); |
| 512 | /// ``` |
| 513 | /// |
| 514 | /// [`Vec::append`]: ../../std/vec/struct.Vec.html#method.append |
| 515 | #[doc (alias = "memcpy" )] |
| 516 | #[stable (feature = "rust1" , since = "1.0.0" )] |
| 517 | #[rustc_const_stable (feature = "const_intrinsic_copy" , since = "1.83.0" )] |
| 518 | #[inline (always)] |
| 519 | #[cfg_attr (miri, track_caller)] // even without panics, this helps for Miri backtraces |
| 520 | #[rustc_diagnostic_item = "ptr_copy_nonoverlapping" ] |
| 521 | pub const unsafe fn copy_nonoverlapping<T>(src: *const T, dst: *mut T, count: usize) { |
| 522 | ub_checks::assert_unsafe_precondition!( |
| 523 | check_language_ub, |
| 524 | "ptr::copy_nonoverlapping requires that both pointer arguments are aligned and non-null \ |
| 525 | and the specified memory ranges do not overlap" , |
| 526 | ( |
| 527 | src: *const () = src as *const (), |
| 528 | dst: *mut () = dst as *mut (), |
| 529 | size: usize = size_of::<T>(), |
| 530 | align: usize = align_of::<T>(), |
| 531 | count: usize = count, |
| 532 | ) => { |
| 533 | let zero_size = count == 0 || size == 0; |
| 534 | ub_checks::maybe_is_aligned_and_not_null(src, align, zero_size) |
| 535 | && ub_checks::maybe_is_aligned_and_not_null(dst, align, zero_size) |
| 536 | && ub_checks::maybe_is_nonoverlapping(src, dst, size, count) |
| 537 | } |
| 538 | ); |
| 539 | |
| 540 | // SAFETY: the safety contract for `copy_nonoverlapping` must be |
| 541 | // upheld by the caller. |
| 542 | unsafe { crate::intrinsics::copy_nonoverlapping(src, dst, count) } |
| 543 | } |
| 544 | |
| 545 | /// Copies `count * size_of::<T>()` bytes from `src` to `dst`. The source |
| 546 | /// and destination may overlap. |
| 547 | /// |
| 548 | /// If the source and destination will *never* overlap, |
| 549 | /// [`copy_nonoverlapping`] can be used instead. |
| 550 | /// |
| 551 | /// `copy` is semantically equivalent to C's [`memmove`], but |
| 552 | /// with the source and destination arguments swapped, |
| 553 | /// and `count` counting the number of `T`s instead of bytes. |
| 554 | /// Copying takes place as if the bytes were copied from `src` |
| 555 | /// to a temporary array and then copied from the array to `dst`. |
| 556 | /// |
| 557 | /// The copy is "untyped" in the sense that data may be uninitialized or otherwise violate the |
| 558 | /// requirements of `T`. The initialization state is preserved exactly. |
| 559 | /// |
| 560 | /// [`memmove`]: https://en.cppreference.com/w/c/string/byte/memmove |
| 561 | /// |
| 562 | /// # Safety |
| 563 | /// |
| 564 | /// Behavior is undefined if any of the following conditions are violated: |
| 565 | /// |
| 566 | /// * `src` must be [valid] for reads of `count * size_of::<T>()` bytes. |
| 567 | /// |
| 568 | /// * `dst` must be [valid] for writes of `count * size_of::<T>()` bytes, and must remain valid even |
| 569 | /// when `src` is read for `count * size_of::<T>()` bytes. (This means if the memory ranges |
| 570 | /// overlap, the `dst` pointer must not be invalidated by `src` reads.) |
| 571 | /// |
| 572 | /// * Both `src` and `dst` must be properly aligned. |
| 573 | /// |
| 574 | /// Like [`read`], `copy` creates a bitwise copy of `T`, regardless of |
| 575 | /// whether `T` is [`Copy`]. If `T` is not [`Copy`], using both the values |
| 576 | /// in the region beginning at `*src` and the region beginning at `*dst` can |
| 577 | /// [violate memory safety][read-ownership]. |
| 578 | /// |
| 579 | /// Note that even if the effectively copied size (`count * size_of::<T>()`) is |
| 580 | /// `0`, the pointers must be properly aligned. |
| 581 | /// |
| 582 | /// [`read`]: crate::ptr::read |
| 583 | /// [read-ownership]: crate::ptr::read#ownership-of-the-returned-value |
| 584 | /// [valid]: crate::ptr#safety |
| 585 | /// |
| 586 | /// # Examples |
| 587 | /// |
| 588 | /// Efficiently create a Rust vector from an unsafe buffer: |
| 589 | /// |
| 590 | /// ``` |
| 591 | /// use std::ptr; |
| 592 | /// |
| 593 | /// /// # Safety |
| 594 | /// /// |
| 595 | /// /// * `ptr` must be correctly aligned for its type and non-zero. |
| 596 | /// /// * `ptr` must be valid for reads of `elts` contiguous elements of type `T`. |
| 597 | /// /// * Those elements must not be used after calling this function unless `T: Copy`. |
| 598 | /// # #[allow (dead_code)] |
| 599 | /// unsafe fn from_buf_raw<T>(ptr: *const T, elts: usize) -> Vec<T> { |
| 600 | /// let mut dst = Vec::with_capacity(elts); |
| 601 | /// |
| 602 | /// // SAFETY: Our precondition ensures the source is aligned and valid, |
| 603 | /// // and `Vec::with_capacity` ensures that we have usable space to write them. |
| 604 | /// unsafe { ptr::copy(ptr, dst.as_mut_ptr(), elts); } |
| 605 | /// |
| 606 | /// // SAFETY: We created it with this much capacity earlier, |
| 607 | /// // and the previous `copy` has initialized these elements. |
| 608 | /// unsafe { dst.set_len(elts); } |
| 609 | /// dst |
| 610 | /// } |
| 611 | /// ``` |
| 612 | #[doc (alias = "memmove" )] |
| 613 | #[stable (feature = "rust1" , since = "1.0.0" )] |
| 614 | #[rustc_const_stable (feature = "const_intrinsic_copy" , since = "1.83.0" )] |
| 615 | #[inline (always)] |
| 616 | #[cfg_attr (miri, track_caller)] // even without panics, this helps for Miri backtraces |
| 617 | #[rustc_diagnostic_item = "ptr_copy" ] |
| 618 | pub const unsafe fn copy<T>(src: *const T, dst: *mut T, count: usize) { |
| 619 | // SAFETY: the safety contract for `copy` must be upheld by the caller. |
| 620 | unsafe { |
| 621 | ub_checks::assert_unsafe_precondition!( |
| 622 | check_language_ub, |
| 623 | "ptr::copy requires that both pointer arguments are aligned and non-null" , |
| 624 | ( |
| 625 | src: *const () = src as *const (), |
| 626 | dst: *mut () = dst as *mut (), |
| 627 | align: usize = align_of::<T>(), |
| 628 | zero_size: bool = T::IS_ZST || count == 0, |
| 629 | ) => |
| 630 | ub_checks::maybe_is_aligned_and_not_null(src, align, zero_size) |
| 631 | && ub_checks::maybe_is_aligned_and_not_null(dst, align, zero_size) |
| 632 | ); |
| 633 | crate::intrinsics::copy(src, dst, count) |
| 634 | } |
| 635 | } |
| 636 | |
| 637 | /// Sets `count * size_of::<T>()` bytes of memory starting at `dst` to |
| 638 | /// `val`. |
| 639 | /// |
| 640 | /// `write_bytes` is similar to C's [`memset`], but sets `count * |
| 641 | /// size_of::<T>()` bytes to `val`. |
| 642 | /// |
| 643 | /// [`memset`]: https://en.cppreference.com/w/c/string/byte/memset |
| 644 | /// |
| 645 | /// # Safety |
| 646 | /// |
| 647 | /// Behavior is undefined if any of the following conditions are violated: |
| 648 | /// |
| 649 | /// * `dst` must be [valid] for writes of `count * size_of::<T>()` bytes. |
| 650 | /// |
| 651 | /// * `dst` must be properly aligned. |
| 652 | /// |
| 653 | /// Note that even if the effectively copied size (`count * size_of::<T>()`) is |
| 654 | /// `0`, the pointer must be properly aligned. |
| 655 | /// |
| 656 | /// Additionally, note that changing `*dst` in this way can easily lead to undefined behavior (UB) |
| 657 | /// later if the written bytes are not a valid representation of some `T`. For instance, the |
| 658 | /// following is an **incorrect** use of this function: |
| 659 | /// |
| 660 | /// ```rust,no_run |
| 661 | /// unsafe { |
| 662 | /// let mut value: u8 = 0; |
| 663 | /// let ptr: *mut bool = &mut value as *mut u8 as *mut bool; |
| 664 | /// let _bool = ptr.read(); // This is fine, `ptr` points to a valid `bool`. |
| 665 | /// ptr.write_bytes(42u8, 1); // This function itself does not cause UB... |
| 666 | /// let _bool = ptr.read(); // ...but it makes this operation UB! ⚠️ |
| 667 | /// } |
| 668 | /// ``` |
| 669 | /// |
| 670 | /// [valid]: crate::ptr#safety |
| 671 | /// |
| 672 | /// # Examples |
| 673 | /// |
| 674 | /// Basic usage: |
| 675 | /// |
| 676 | /// ``` |
| 677 | /// use std::ptr; |
| 678 | /// |
| 679 | /// let mut vec = vec![0u32; 4]; |
| 680 | /// unsafe { |
| 681 | /// let vec_ptr = vec.as_mut_ptr(); |
| 682 | /// ptr::write_bytes(vec_ptr, 0xfe, 2); |
| 683 | /// } |
| 684 | /// assert_eq!(vec, [0xfefefefe, 0xfefefefe, 0, 0]); |
| 685 | /// ``` |
| 686 | #[doc (alias = "memset" )] |
| 687 | #[stable (feature = "rust1" , since = "1.0.0" )] |
| 688 | #[rustc_const_stable (feature = "const_ptr_write" , since = "1.83.0" )] |
| 689 | #[inline (always)] |
| 690 | #[cfg_attr (miri, track_caller)] // even without panics, this helps for Miri backtraces |
| 691 | #[rustc_diagnostic_item = "ptr_write_bytes" ] |
| 692 | pub const unsafe fn write_bytes<T>(dst: *mut T, val: u8, count: usize) { |
| 693 | // SAFETY: the safety contract for `write_bytes` must be upheld by the caller. |
| 694 | unsafe { |
| 695 | ub_checks::assert_unsafe_precondition!( |
| 696 | check_language_ub, |
| 697 | "ptr::write_bytes requires that the destination pointer is aligned and non-null" , |
| 698 | ( |
| 699 | addr: *const () = dst as *const (), |
| 700 | align: usize = align_of::<T>(), |
| 701 | zero_size: bool = T::IS_ZST || count == 0, |
| 702 | ) => ub_checks::maybe_is_aligned_and_not_null(addr, align, zero_size) |
| 703 | ); |
| 704 | crate::intrinsics::write_bytes(dst, val, count) |
| 705 | } |
| 706 | } |
| 707 | |
| 708 | /// Executes the destructor (if any) of the pointed-to value. |
| 709 | /// |
| 710 | /// This is almost the same as calling [`ptr::read`] and discarding |
| 711 | /// the result, but has the following advantages: |
| 712 | // FIXME: say something more useful than "almost the same"? |
| 713 | // There are open questions here: `read` requires the value to be fully valid, e.g. if `T` is a |
| 714 | // `bool` it must be 0 or 1, if it is a reference then it must be dereferenceable. `drop_in_place` |
| 715 | // only requires that `*to_drop` be "valid for dropping" and we have not defined what that means. In |
| 716 | // Miri it currently (May 2024) requires nothing at all for types without drop glue. |
| 717 | /// |
| 718 | /// * It is *required* to use `drop_in_place` to drop unsized types like |
| 719 | /// trait objects, because they can't be read out onto the stack and |
| 720 | /// dropped normally. |
| 721 | /// |
| 722 | /// * It is friendlier to the optimizer to do this over [`ptr::read`] when |
| 723 | /// dropping manually allocated memory (e.g., in the implementations of |
| 724 | /// `Box`/`Rc`/`Vec`), as the compiler doesn't need to prove that it's |
| 725 | /// sound to elide the copy. |
| 726 | /// |
| 727 | /// * It can be used to drop [pinned] data when `T` is not `repr(packed)` |
| 728 | /// (pinned data must not be moved before it is dropped). |
| 729 | /// |
| 730 | /// Unaligned values cannot be dropped in place, they must be copied to an aligned |
| 731 | /// location first using [`ptr::read_unaligned`]. For packed structs, this move is |
| 732 | /// done automatically by the compiler. This means the fields of packed structs |
| 733 | /// are not dropped in-place. |
| 734 | /// |
| 735 | /// [`ptr::read`]: self::read |
| 736 | /// [`ptr::read_unaligned`]: self::read_unaligned |
| 737 | /// [pinned]: crate::pin |
| 738 | /// |
| 739 | /// # Safety |
| 740 | /// |
| 741 | /// Behavior is undefined if any of the following conditions are violated: |
| 742 | /// |
| 743 | /// * `to_drop` must be [valid] for both reads and writes. |
| 744 | /// |
| 745 | /// * `to_drop` must be properly aligned, even if `T` has size 0. |
| 746 | /// |
| 747 | /// * `to_drop` must be nonnull, even if `T` has size 0. |
| 748 | /// |
| 749 | /// * The value `to_drop` points to must be valid for dropping, which may mean |
| 750 | /// it must uphold additional invariants. These invariants depend on the type |
| 751 | /// of the value being dropped. For instance, when dropping a Box, the box's |
| 752 | /// pointer to the heap must be valid. |
| 753 | /// |
| 754 | /// * While `drop_in_place` is executing, the only way to access parts of |
| 755 | /// `to_drop` is through the `&mut self` references supplied to the |
| 756 | /// `Drop::drop` methods that `drop_in_place` invokes. |
| 757 | /// |
| 758 | /// Additionally, if `T` is not [`Copy`], using the pointed-to value after |
| 759 | /// calling `drop_in_place` can cause undefined behavior. Note that `*to_drop = |
| 760 | /// foo` counts as a use because it will cause the value to be dropped |
| 761 | /// again. [`write()`] can be used to overwrite data without causing it to be |
| 762 | /// dropped. |
| 763 | /// |
| 764 | /// [valid]: self#safety |
| 765 | /// |
| 766 | /// # Examples |
| 767 | /// |
| 768 | /// Manually remove the last item from a vector: |
| 769 | /// |
| 770 | /// ``` |
| 771 | /// use std::ptr; |
| 772 | /// use std::rc::Rc; |
| 773 | /// |
| 774 | /// let last = Rc::new(1); |
| 775 | /// let weak = Rc::downgrade(&last); |
| 776 | /// |
| 777 | /// let mut v = vec![Rc::new(0), last]; |
| 778 | /// |
| 779 | /// unsafe { |
| 780 | /// // Get a raw pointer to the last element in `v`. |
| 781 | /// let ptr = &mut v[1] as *mut _; |
| 782 | /// // Shorten `v` to prevent the last item from being dropped. We do that first, |
| 783 | /// // to prevent issues if the `drop_in_place` below panics. |
| 784 | /// v.set_len(1); |
| 785 | /// // Without a call `drop_in_place`, the last item would never be dropped, |
| 786 | /// // and the memory it manages would be leaked. |
| 787 | /// ptr::drop_in_place(ptr); |
| 788 | /// } |
| 789 | /// |
| 790 | /// assert_eq!(v, &[0.into()]); |
| 791 | /// |
| 792 | /// // Ensure that the last item was dropped. |
| 793 | /// assert!(weak.upgrade().is_none()); |
| 794 | /// ``` |
| 795 | #[stable (feature = "drop_in_place" , since = "1.8.0" )] |
| 796 | #[lang = "drop_in_place" ] |
| 797 | #[allow (unconditional_recursion)] |
| 798 | #[rustc_diagnostic_item = "ptr_drop_in_place" ] |
| 799 | pub unsafe fn drop_in_place<T: ?Sized>(to_drop: *mut T) { |
| 800 | // Code here does not matter - this is replaced by the |
| 801 | // real drop glue by the compiler. |
| 802 | |
| 803 | // SAFETY: see comment above |
| 804 | unsafe { drop_in_place(to_drop) } |
| 805 | } |
| 806 | |
| 807 | /// Creates a null raw pointer. |
| 808 | /// |
| 809 | /// This function is equivalent to zero-initializing the pointer: |
| 810 | /// `MaybeUninit::<*const T>::zeroed().assume_init()`. |
| 811 | /// The resulting pointer has the address 0. |
| 812 | /// |
| 813 | /// # Examples |
| 814 | /// |
| 815 | /// ``` |
| 816 | /// use std::ptr; |
| 817 | /// |
| 818 | /// let p: *const i32 = ptr::null(); |
| 819 | /// assert!(p.is_null()); |
| 820 | /// assert_eq!(p as usize, 0); // this pointer has the address 0 |
| 821 | /// ``` |
| 822 | #[inline (always)] |
| 823 | #[must_use ] |
| 824 | #[stable (feature = "rust1" , since = "1.0.0" )] |
| 825 | #[rustc_promotable ] |
| 826 | #[rustc_const_stable (feature = "const_ptr_null" , since = "1.24.0" )] |
| 827 | #[rustc_diagnostic_item = "ptr_null" ] |
| 828 | pub const fn null<T: ?Sized + Thin>() -> *const T { |
| 829 | from_raw_parts(data_pointer:without_provenance::<()>(0), ()) |
| 830 | } |
| 831 | |
| 832 | /// Creates a null mutable raw pointer. |
| 833 | /// |
| 834 | /// This function is equivalent to zero-initializing the pointer: |
| 835 | /// `MaybeUninit::<*mut T>::zeroed().assume_init()`. |
| 836 | /// The resulting pointer has the address 0. |
| 837 | /// |
| 838 | /// # Examples |
| 839 | /// |
| 840 | /// ``` |
| 841 | /// use std::ptr; |
| 842 | /// |
| 843 | /// let p: *mut i32 = ptr::null_mut(); |
| 844 | /// assert!(p.is_null()); |
| 845 | /// assert_eq!(p as usize, 0); // this pointer has the address 0 |
| 846 | /// ``` |
| 847 | #[inline (always)] |
| 848 | #[must_use ] |
| 849 | #[stable (feature = "rust1" , since = "1.0.0" )] |
| 850 | #[rustc_promotable ] |
| 851 | #[rustc_const_stable (feature = "const_ptr_null" , since = "1.24.0" )] |
| 852 | #[rustc_diagnostic_item = "ptr_null_mut" ] |
| 853 | pub const fn null_mut<T: ?Sized + Thin>() -> *mut T { |
| 854 | from_raw_parts_mut(data_pointer:without_provenance_mut::<()>(0), ()) |
| 855 | } |
| 856 | |
| 857 | /// Creates a pointer with the given address and no [provenance][crate::ptr#provenance]. |
| 858 | /// |
| 859 | /// This is equivalent to `ptr::null().with_addr(addr)`. |
| 860 | /// |
| 861 | /// Without provenance, this pointer is not associated with any actual allocation. Such a |
| 862 | /// no-provenance pointer may be used for zero-sized memory accesses (if suitably aligned), but |
| 863 | /// non-zero-sized memory accesses with a no-provenance pointer are UB. No-provenance pointers are |
| 864 | /// little more than a `usize` address in disguise. |
| 865 | /// |
| 866 | /// This is different from `addr as *const T`, which creates a pointer that picks up a previously |
| 867 | /// exposed provenance. See [`with_exposed_provenance`] for more details on that operation. |
| 868 | /// |
| 869 | /// This is a [Strict Provenance][crate::ptr#strict-provenance] API. |
| 870 | #[inline (always)] |
| 871 | #[must_use ] |
| 872 | #[stable (feature = "strict_provenance" , since = "1.84.0" )] |
| 873 | #[rustc_const_stable (feature = "strict_provenance" , since = "1.84.0" )] |
| 874 | pub const fn without_provenance<T>(addr: usize) -> *const T { |
| 875 | without_provenance_mut(addr) |
| 876 | } |
| 877 | |
| 878 | /// Creates a new pointer that is dangling, but non-null and well-aligned. |
| 879 | /// |
| 880 | /// This is useful for initializing types which lazily allocate, like |
| 881 | /// `Vec::new` does. |
| 882 | /// |
| 883 | /// Note that the pointer value may potentially represent a valid pointer to |
| 884 | /// a `T`, which means this must not be used as a "not yet initialized" |
| 885 | /// sentinel value. Types that lazily allocate must track initialization by |
| 886 | /// some other means. |
| 887 | #[inline (always)] |
| 888 | #[must_use ] |
| 889 | #[stable (feature = "strict_provenance" , since = "1.84.0" )] |
| 890 | #[rustc_const_stable (feature = "strict_provenance" , since = "1.84.0" )] |
| 891 | pub const fn dangling<T>() -> *const T { |
| 892 | dangling_mut() |
| 893 | } |
| 894 | |
| 895 | /// Creates a pointer with the given address and no [provenance][crate::ptr#provenance]. |
| 896 | /// |
| 897 | /// This is equivalent to `ptr::null_mut().with_addr(addr)`. |
| 898 | /// |
| 899 | /// Without provenance, this pointer is not associated with any actual allocation. Such a |
| 900 | /// no-provenance pointer may be used for zero-sized memory accesses (if suitably aligned), but |
| 901 | /// non-zero-sized memory accesses with a no-provenance pointer are UB. No-provenance pointers are |
| 902 | /// little more than a `usize` address in disguise. |
| 903 | /// |
| 904 | /// This is different from `addr as *mut T`, which creates a pointer that picks up a previously |
| 905 | /// exposed provenance. See [`with_exposed_provenance_mut`] for more details on that operation. |
| 906 | /// |
| 907 | /// This is a [Strict Provenance][crate::ptr#strict-provenance] API. |
| 908 | #[inline (always)] |
| 909 | #[must_use ] |
| 910 | #[stable (feature = "strict_provenance" , since = "1.84.0" )] |
| 911 | #[rustc_const_stable (feature = "strict_provenance" , since = "1.84.0" )] |
| 912 | pub const fn without_provenance_mut<T>(addr: usize) -> *mut T { |
| 913 | // An int-to-pointer transmute currently has exactly the intended semantics: it creates a |
| 914 | // pointer without provenance. Note that this is *not* a stable guarantee about transmute |
| 915 | // semantics, it relies on sysroot crates having special status. |
| 916 | // SAFETY: every valid integer is also a valid pointer (as long as you don't dereference that |
| 917 | // pointer). |
| 918 | unsafe { mem::transmute(src:addr) } |
| 919 | } |
| 920 | |
| 921 | /// Creates a new pointer that is dangling, but non-null and well-aligned. |
| 922 | /// |
| 923 | /// This is useful for initializing types which lazily allocate, like |
| 924 | /// `Vec::new` does. |
| 925 | /// |
| 926 | /// Note that the pointer value may potentially represent a valid pointer to |
| 927 | /// a `T`, which means this must not be used as a "not yet initialized" |
| 928 | /// sentinel value. Types that lazily allocate must track initialization by |
| 929 | /// some other means. |
| 930 | #[inline (always)] |
| 931 | #[must_use ] |
| 932 | #[stable (feature = "strict_provenance" , since = "1.84.0" )] |
| 933 | #[rustc_const_stable (feature = "strict_provenance" , since = "1.84.0" )] |
| 934 | pub const fn dangling_mut<T>() -> *mut T { |
| 935 | NonNull::dangling().as_ptr() |
| 936 | } |
| 937 | |
| 938 | /// Converts an address back to a pointer, picking up some previously 'exposed' |
| 939 | /// [provenance][crate::ptr#provenance]. |
| 940 | /// |
| 941 | /// This is fully equivalent to `addr as *const T`. The provenance of the returned pointer is that |
| 942 | /// of *some* pointer that was previously exposed by passing it to |
| 943 | /// [`expose_provenance`][pointer::expose_provenance], or a `ptr as usize` cast. In addition, memory |
| 944 | /// which is outside the control of the Rust abstract machine (MMIO registers, for example) is |
| 945 | /// always considered to be accessible with an exposed provenance, so long as this memory is disjoint |
| 946 | /// from memory that will be used by the abstract machine such as the stack, heap, and statics. |
| 947 | /// |
| 948 | /// The exact provenance that gets picked is not specified. The compiler will do its best to pick |
| 949 | /// the "right" provenance for you (whatever that may be), but currently we cannot provide any |
| 950 | /// guarantees about which provenance the resulting pointer will have -- and therefore there |
| 951 | /// is no definite specification for which memory the resulting pointer may access. |
| 952 | /// |
| 953 | /// If there is *no* previously 'exposed' provenance that justifies the way the returned pointer |
| 954 | /// will be used, the program has undefined behavior. In particular, the aliasing rules still apply: |
| 955 | /// pointers and references that have been invalidated due to aliasing accesses cannot be used |
| 956 | /// anymore, even if they have been exposed! |
| 957 | /// |
| 958 | /// Due to its inherent ambiguity, this operation may not be supported by tools that help you to |
| 959 | /// stay conformant with the Rust memory model. It is recommended to use [Strict |
| 960 | /// Provenance][self#strict-provenance] APIs such as [`with_addr`][pointer::with_addr] wherever |
| 961 | /// possible. |
| 962 | /// |
| 963 | /// On most platforms this will produce a value with the same bytes as the address. Platforms |
| 964 | /// which need to store additional information in a pointer may not support this operation, |
| 965 | /// since it is generally not possible to actually *compute* which provenance the returned |
| 966 | /// pointer has to pick up. |
| 967 | /// |
| 968 | /// This is an [Exposed Provenance][crate::ptr#exposed-provenance] API. |
| 969 | #[must_use ] |
| 970 | #[inline (always)] |
| 971 | #[stable (feature = "exposed_provenance" , since = "1.84.0" )] |
| 972 | #[cfg_attr (miri, track_caller)] // even without panics, this helps for Miri backtraces |
| 973 | #[allow (fuzzy_provenance_casts)] // this *is* the explicit provenance API one should use instead |
| 974 | pub fn with_exposed_provenance<T>(addr: usize) -> *const T { |
| 975 | addr as *const T |
| 976 | } |
| 977 | |
| 978 | /// Converts an address back to a mutable pointer, picking up some previously 'exposed' |
| 979 | /// [provenance][crate::ptr#provenance]. |
| 980 | /// |
| 981 | /// This is fully equivalent to `addr as *mut T`. The provenance of the returned pointer is that |
| 982 | /// of *some* pointer that was previously exposed by passing it to |
| 983 | /// [`expose_provenance`][pointer::expose_provenance], or a `ptr as usize` cast. In addition, memory |
| 984 | /// which is outside the control of the Rust abstract machine (MMIO registers, for example) is |
| 985 | /// always considered to be accessible with an exposed provenance, so long as this memory is disjoint |
| 986 | /// from memory that will be used by the abstract machine such as the stack, heap, and statics. |
| 987 | /// |
| 988 | /// The exact provenance that gets picked is not specified. The compiler will do its best to pick |
| 989 | /// the "right" provenance for you (whatever that may be), but currently we cannot provide any |
| 990 | /// guarantees about which provenance the resulting pointer will have -- and therefore there |
| 991 | /// is no definite specification for which memory the resulting pointer may access. |
| 992 | /// |
| 993 | /// If there is *no* previously 'exposed' provenance that justifies the way the returned pointer |
| 994 | /// will be used, the program has undefined behavior. In particular, the aliasing rules still apply: |
| 995 | /// pointers and references that have been invalidated due to aliasing accesses cannot be used |
| 996 | /// anymore, even if they have been exposed! |
| 997 | /// |
| 998 | /// Due to its inherent ambiguity, this operation may not be supported by tools that help you to |
| 999 | /// stay conformant with the Rust memory model. It is recommended to use [Strict |
| 1000 | /// Provenance][self#strict-provenance] APIs such as [`with_addr`][pointer::with_addr] wherever |
| 1001 | /// possible. |
| 1002 | /// |
| 1003 | /// On most platforms this will produce a value with the same bytes as the address. Platforms |
| 1004 | /// which need to store additional information in a pointer may not support this operation, |
| 1005 | /// since it is generally not possible to actually *compute* which provenance the returned |
| 1006 | /// pointer has to pick up. |
| 1007 | /// |
| 1008 | /// This is an [Exposed Provenance][crate::ptr#exposed-provenance] API. |
| 1009 | #[must_use ] |
| 1010 | #[inline (always)] |
| 1011 | #[stable (feature = "exposed_provenance" , since = "1.84.0" )] |
| 1012 | #[cfg_attr (miri, track_caller)] // even without panics, this helps for Miri backtraces |
| 1013 | #[allow (fuzzy_provenance_casts)] // this *is* the explicit provenance API one should use instead |
| 1014 | pub fn with_exposed_provenance_mut<T>(addr: usize) -> *mut T { |
| 1015 | addr as *mut T |
| 1016 | } |
| 1017 | |
| 1018 | /// Converts a reference to a raw pointer. |
| 1019 | /// |
| 1020 | /// For `r: &T`, `from_ref(r)` is equivalent to `r as *const T` (except for the caveat noted below), |
| 1021 | /// but is a bit safer since it will never silently change type or mutability, in particular if the |
| 1022 | /// code is refactored. |
| 1023 | /// |
| 1024 | /// The caller must ensure that the pointee outlives the pointer this function returns, or else it |
| 1025 | /// will end up dangling. |
| 1026 | /// |
| 1027 | /// The caller must also ensure that the memory the pointer (non-transitively) points to is never |
| 1028 | /// written to (except inside an `UnsafeCell`) using this pointer or any pointer derived from it. If |
| 1029 | /// you need to mutate the pointee, use [`from_mut`]. Specifically, to turn a mutable reference `m: |
| 1030 | /// &mut T` into `*const T`, prefer `from_mut(m).cast_const()` to obtain a pointer that can later be |
| 1031 | /// used for mutation. |
| 1032 | /// |
| 1033 | /// ## Interaction with lifetime extension |
| 1034 | /// |
| 1035 | /// Note that this has subtle interactions with the rules for lifetime extension of temporaries in |
| 1036 | /// tail expressions. This code is valid, albeit in a non-obvious way: |
| 1037 | /// ```rust |
| 1038 | /// # type T = i32; |
| 1039 | /// # fn foo() -> T { 42 } |
| 1040 | /// // The temporary holding the return value of `foo` has its lifetime extended, |
| 1041 | /// // because the surrounding expression involves no function call. |
| 1042 | /// let p = &foo() as *const T; |
| 1043 | /// unsafe { p.read() }; |
| 1044 | /// ``` |
| 1045 | /// Naively replacing the cast with `from_ref` is not valid: |
| 1046 | /// ```rust,no_run |
| 1047 | /// # use std::ptr; |
| 1048 | /// # type T = i32; |
| 1049 | /// # fn foo() -> T { 42 } |
| 1050 | /// // The temporary holding the return value of `foo` does *not* have its lifetime extended, |
| 1051 | /// // because the surrounding expression involves a function call. |
| 1052 | /// let p = ptr::from_ref(&foo()); |
| 1053 | /// unsafe { p.read() }; // UB! Reading from a dangling pointer ⚠️ |
| 1054 | /// ``` |
| 1055 | /// The recommended way to write this code is to avoid relying on lifetime extension |
| 1056 | /// when raw pointers are involved: |
| 1057 | /// ```rust |
| 1058 | /// # use std::ptr; |
| 1059 | /// # type T = i32; |
| 1060 | /// # fn foo() -> T { 42 } |
| 1061 | /// let x = foo(); |
| 1062 | /// let p = ptr::from_ref(&x); |
| 1063 | /// unsafe { p.read() }; |
| 1064 | /// ``` |
| 1065 | #[inline (always)] |
| 1066 | #[must_use ] |
| 1067 | #[stable (feature = "ptr_from_ref" , since = "1.76.0" )] |
| 1068 | #[rustc_const_stable (feature = "ptr_from_ref" , since = "1.76.0" )] |
| 1069 | #[rustc_never_returns_null_ptr ] |
| 1070 | #[rustc_diagnostic_item = "ptr_from_ref" ] |
| 1071 | pub const fn from_ref<T: ?Sized>(r: &T) -> *const T { |
| 1072 | r |
| 1073 | } |
| 1074 | |
| 1075 | /// Converts a mutable reference to a raw pointer. |
| 1076 | /// |
| 1077 | /// For `r: &mut T`, `from_mut(r)` is equivalent to `r as *mut T` (except for the caveat noted |
| 1078 | /// below), but is a bit safer since it will never silently change type or mutability, in particular |
| 1079 | /// if the code is refactored. |
| 1080 | /// |
| 1081 | /// The caller must ensure that the pointee outlives the pointer this function returns, or else it |
| 1082 | /// will end up dangling. |
| 1083 | /// |
| 1084 | /// ## Interaction with lifetime extension |
| 1085 | /// |
| 1086 | /// Note that this has subtle interactions with the rules for lifetime extension of temporaries in |
| 1087 | /// tail expressions. This code is valid, albeit in a non-obvious way: |
| 1088 | /// ```rust |
| 1089 | /// # type T = i32; |
| 1090 | /// # fn foo() -> T { 42 } |
| 1091 | /// // The temporary holding the return value of `foo` has its lifetime extended, |
| 1092 | /// // because the surrounding expression involves no function call. |
| 1093 | /// let p = &mut foo() as *mut T; |
| 1094 | /// unsafe { p.write(T::default()) }; |
| 1095 | /// ``` |
| 1096 | /// Naively replacing the cast with `from_mut` is not valid: |
| 1097 | /// ```rust,no_run |
| 1098 | /// # use std::ptr; |
| 1099 | /// # type T = i32; |
| 1100 | /// # fn foo() -> T { 42 } |
| 1101 | /// // The temporary holding the return value of `foo` does *not* have its lifetime extended, |
| 1102 | /// // because the surrounding expression involves a function call. |
| 1103 | /// let p = ptr::from_mut(&mut foo()); |
| 1104 | /// unsafe { p.write(T::default()) }; // UB! Writing to a dangling pointer ⚠️ |
| 1105 | /// ``` |
| 1106 | /// The recommended way to write this code is to avoid relying on lifetime extension |
| 1107 | /// when raw pointers are involved: |
| 1108 | /// ```rust |
| 1109 | /// # use std::ptr; |
| 1110 | /// # type T = i32; |
| 1111 | /// # fn foo() -> T { 42 } |
| 1112 | /// let mut x = foo(); |
| 1113 | /// let p = ptr::from_mut(&mut x); |
| 1114 | /// unsafe { p.write(T::default()) }; |
| 1115 | /// ``` |
| 1116 | #[inline (always)] |
| 1117 | #[must_use ] |
| 1118 | #[stable (feature = "ptr_from_ref" , since = "1.76.0" )] |
| 1119 | #[rustc_const_stable (feature = "ptr_from_ref" , since = "1.76.0" )] |
| 1120 | #[rustc_never_returns_null_ptr ] |
| 1121 | pub const fn from_mut<T: ?Sized>(r: &mut T) -> *mut T { |
| 1122 | r |
| 1123 | } |
| 1124 | |
| 1125 | /// Forms a raw slice from a pointer and a length. |
| 1126 | /// |
| 1127 | /// The `len` argument is the number of **elements**, not the number of bytes. |
| 1128 | /// |
| 1129 | /// This function is safe, but actually using the return value is unsafe. |
| 1130 | /// See the documentation of [`slice::from_raw_parts`] for slice safety requirements. |
| 1131 | /// |
| 1132 | /// [`slice::from_raw_parts`]: crate::slice::from_raw_parts |
| 1133 | /// |
| 1134 | /// # Examples |
| 1135 | /// |
| 1136 | /// ```rust |
| 1137 | /// use std::ptr; |
| 1138 | /// |
| 1139 | /// // create a slice pointer when starting out with a pointer to the first element |
| 1140 | /// let x = [5, 6, 7]; |
| 1141 | /// let raw_pointer = x.as_ptr(); |
| 1142 | /// let slice = ptr::slice_from_raw_parts(raw_pointer, 3); |
| 1143 | /// assert_eq!(unsafe { &*slice }[2], 7); |
| 1144 | /// ``` |
| 1145 | /// |
| 1146 | /// You must ensure that the pointer is valid and not null before dereferencing |
| 1147 | /// the raw slice. A slice reference must never have a null pointer, even if it's empty. |
| 1148 | /// |
| 1149 | /// ```rust,should_panic |
| 1150 | /// use std::ptr; |
| 1151 | /// let danger: *const [u8] = ptr::slice_from_raw_parts(ptr::null(), 0); |
| 1152 | /// unsafe { |
| 1153 | /// danger.as_ref().expect("references must not be null" ); |
| 1154 | /// } |
| 1155 | /// ``` |
| 1156 | #[inline ] |
| 1157 | #[stable (feature = "slice_from_raw_parts" , since = "1.42.0" )] |
| 1158 | #[rustc_const_stable (feature = "const_slice_from_raw_parts" , since = "1.64.0" )] |
| 1159 | #[rustc_diagnostic_item = "ptr_slice_from_raw_parts" ] |
| 1160 | pub const fn slice_from_raw_parts<T>(data: *const T, len: usize) -> *const [T] { |
| 1161 | from_raw_parts(data, metadata:len) |
| 1162 | } |
| 1163 | |
| 1164 | /// Forms a raw mutable slice from a pointer and a length. |
| 1165 | /// |
| 1166 | /// The `len` argument is the number of **elements**, not the number of bytes. |
| 1167 | /// |
| 1168 | /// Performs the same functionality as [`slice_from_raw_parts`], except that a |
| 1169 | /// raw mutable slice is returned, as opposed to a raw immutable slice. |
| 1170 | /// |
| 1171 | /// This function is safe, but actually using the return value is unsafe. |
| 1172 | /// See the documentation of [`slice::from_raw_parts_mut`] for slice safety requirements. |
| 1173 | /// |
| 1174 | /// [`slice::from_raw_parts_mut`]: crate::slice::from_raw_parts_mut |
| 1175 | /// |
| 1176 | /// # Examples |
| 1177 | /// |
| 1178 | /// ```rust |
| 1179 | /// use std::ptr; |
| 1180 | /// |
| 1181 | /// let x = &mut [5, 6, 7]; |
| 1182 | /// let raw_pointer = x.as_mut_ptr(); |
| 1183 | /// let slice = ptr::slice_from_raw_parts_mut(raw_pointer, 3); |
| 1184 | /// |
| 1185 | /// unsafe { |
| 1186 | /// (*slice)[2] = 99; // assign a value at an index in the slice |
| 1187 | /// }; |
| 1188 | /// |
| 1189 | /// assert_eq!(unsafe { &*slice }[2], 99); |
| 1190 | /// ``` |
| 1191 | /// |
| 1192 | /// You must ensure that the pointer is valid and not null before dereferencing |
| 1193 | /// the raw slice. A slice reference must never have a null pointer, even if it's empty. |
| 1194 | /// |
| 1195 | /// ```rust,should_panic |
| 1196 | /// use std::ptr; |
| 1197 | /// let danger: *mut [u8] = ptr::slice_from_raw_parts_mut(ptr::null_mut(), 0); |
| 1198 | /// unsafe { |
| 1199 | /// danger.as_mut().expect("references must not be null" ); |
| 1200 | /// } |
| 1201 | /// ``` |
| 1202 | #[inline ] |
| 1203 | #[stable (feature = "slice_from_raw_parts" , since = "1.42.0" )] |
| 1204 | #[rustc_const_stable (feature = "const_slice_from_raw_parts_mut" , since = "1.83.0" )] |
| 1205 | #[rustc_diagnostic_item = "ptr_slice_from_raw_parts_mut" ] |
| 1206 | pub const fn slice_from_raw_parts_mut<T>(data: *mut T, len: usize) -> *mut [T] { |
| 1207 | from_raw_parts_mut(data, metadata:len) |
| 1208 | } |
| 1209 | |
| 1210 | /// Swaps the values at two mutable locations of the same type, without |
| 1211 | /// deinitializing either. |
| 1212 | /// |
| 1213 | /// But for the following exceptions, this function is semantically |
| 1214 | /// equivalent to [`mem::swap`]: |
| 1215 | /// |
| 1216 | /// * It operates on raw pointers instead of references. When references are |
| 1217 | /// available, [`mem::swap`] should be preferred. |
| 1218 | /// |
| 1219 | /// * The two pointed-to values may overlap. If the values do overlap, then the |
| 1220 | /// overlapping region of memory from `x` will be used. This is demonstrated |
| 1221 | /// in the second example below. |
| 1222 | /// |
| 1223 | /// * The operation is "untyped" in the sense that data may be uninitialized or otherwise violate |
| 1224 | /// the requirements of `T`. The initialization state is preserved exactly. |
| 1225 | /// |
| 1226 | /// # Safety |
| 1227 | /// |
| 1228 | /// Behavior is undefined if any of the following conditions are violated: |
| 1229 | /// |
| 1230 | /// * Both `x` and `y` must be [valid] for both reads and writes. They must remain valid even when the |
| 1231 | /// other pointer is written. (This means if the memory ranges overlap, the two pointers must not |
| 1232 | /// be subject to aliasing restrictions relative to each other.) |
| 1233 | /// |
| 1234 | /// * Both `x` and `y` must be properly aligned. |
| 1235 | /// |
| 1236 | /// Note that even if `T` has size `0`, the pointers must be properly aligned. |
| 1237 | /// |
| 1238 | /// [valid]: self#safety |
| 1239 | /// |
| 1240 | /// # Examples |
| 1241 | /// |
| 1242 | /// Swapping two non-overlapping regions: |
| 1243 | /// |
| 1244 | /// ``` |
| 1245 | /// use std::ptr; |
| 1246 | /// |
| 1247 | /// let mut array = [0, 1, 2, 3]; |
| 1248 | /// |
| 1249 | /// let (x, y) = array.split_at_mut(2); |
| 1250 | /// let x = x.as_mut_ptr().cast::<[u32; 2]>(); // this is `array[0..2]` |
| 1251 | /// let y = y.as_mut_ptr().cast::<[u32; 2]>(); // this is `array[2..4]` |
| 1252 | /// |
| 1253 | /// unsafe { |
| 1254 | /// ptr::swap(x, y); |
| 1255 | /// assert_eq!([2, 3, 0, 1], array); |
| 1256 | /// } |
| 1257 | /// ``` |
| 1258 | /// |
| 1259 | /// Swapping two overlapping regions: |
| 1260 | /// |
| 1261 | /// ``` |
| 1262 | /// use std::ptr; |
| 1263 | /// |
| 1264 | /// let mut array: [i32; 4] = [0, 1, 2, 3]; |
| 1265 | /// |
| 1266 | /// let array_ptr: *mut i32 = array.as_mut_ptr(); |
| 1267 | /// |
| 1268 | /// let x = array_ptr as *mut [i32; 3]; // this is `array[0..3]` |
| 1269 | /// let y = unsafe { array_ptr.add(1) } as *mut [i32; 3]; // this is `array[1..4]` |
| 1270 | /// |
| 1271 | /// unsafe { |
| 1272 | /// ptr::swap(x, y); |
| 1273 | /// // The indices `1..3` of the slice overlap between `x` and `y`. |
| 1274 | /// // Reasonable results would be for to them be `[2, 3]`, so that indices `0..3` are |
| 1275 | /// // `[1, 2, 3]` (matching `y` before the `swap`); or for them to be `[0, 1]` |
| 1276 | /// // so that indices `1..4` are `[0, 1, 2]` (matching `x` before the `swap`). |
| 1277 | /// // This implementation is defined to make the latter choice. |
| 1278 | /// assert_eq!([1, 0, 1, 2], array); |
| 1279 | /// } |
| 1280 | /// ``` |
| 1281 | #[inline ] |
| 1282 | #[stable (feature = "rust1" , since = "1.0.0" )] |
| 1283 | #[rustc_const_stable (feature = "const_swap" , since = "1.85.0" )] |
| 1284 | #[rustc_diagnostic_item = "ptr_swap" ] |
| 1285 | pub const unsafe fn swap<T>(x: *mut T, y: *mut T) { |
| 1286 | // Give ourselves some scratch space to work with. |
| 1287 | // We do not have to worry about drops: `MaybeUninit` does nothing when dropped. |
| 1288 | let mut tmp: MaybeUninit = MaybeUninit::<T>::uninit(); |
| 1289 | |
| 1290 | // Perform the swap |
| 1291 | // SAFETY: the caller must guarantee that `x` and `y` are |
| 1292 | // valid for writes and properly aligned. `tmp` cannot be |
| 1293 | // overlapping either `x` or `y` because `tmp` was just allocated |
| 1294 | // on the stack as a separate allocation. |
| 1295 | unsafe { |
| 1296 | copy_nonoverlapping(src:x, dst:tmp.as_mut_ptr(), count:1); |
| 1297 | copy(src:y, dst:x, count:1); // `x` and `y` may overlap |
| 1298 | copy_nonoverlapping(src:tmp.as_ptr(), dst:y, count:1); |
| 1299 | } |
| 1300 | } |
| 1301 | |
| 1302 | /// Swaps `count * size_of::<T>()` bytes between the two regions of memory |
| 1303 | /// beginning at `x` and `y`. The two regions must *not* overlap. |
| 1304 | /// |
| 1305 | /// The operation is "untyped" in the sense that data may be uninitialized or otherwise violate the |
| 1306 | /// requirements of `T`. The initialization state is preserved exactly. |
| 1307 | /// |
| 1308 | /// # Safety |
| 1309 | /// |
| 1310 | /// Behavior is undefined if any of the following conditions are violated: |
| 1311 | /// |
| 1312 | /// * Both `x` and `y` must be [valid] for both reads and writes of `count * |
| 1313 | /// size_of::<T>()` bytes. |
| 1314 | /// |
| 1315 | /// * Both `x` and `y` must be properly aligned. |
| 1316 | /// |
| 1317 | /// * The region of memory beginning at `x` with a size of `count * |
| 1318 | /// size_of::<T>()` bytes must *not* overlap with the region of memory |
| 1319 | /// beginning at `y` with the same size. |
| 1320 | /// |
| 1321 | /// Note that even if the effectively copied size (`count * size_of::<T>()`) is `0`, |
| 1322 | /// the pointers must be properly aligned. |
| 1323 | /// |
| 1324 | /// [valid]: self#safety |
| 1325 | /// |
| 1326 | /// # Examples |
| 1327 | /// |
| 1328 | /// Basic usage: |
| 1329 | /// |
| 1330 | /// ``` |
| 1331 | /// use std::ptr; |
| 1332 | /// |
| 1333 | /// let mut x = [1, 2, 3, 4]; |
| 1334 | /// let mut y = [7, 8, 9]; |
| 1335 | /// |
| 1336 | /// unsafe { |
| 1337 | /// ptr::swap_nonoverlapping(x.as_mut_ptr(), y.as_mut_ptr(), 2); |
| 1338 | /// } |
| 1339 | /// |
| 1340 | /// assert_eq!(x, [7, 8, 3, 4]); |
| 1341 | /// assert_eq!(y, [1, 2, 9]); |
| 1342 | /// ``` |
| 1343 | /// |
| 1344 | /// # Const evaluation limitations |
| 1345 | /// |
| 1346 | /// If this function is invoked during const-evaluation, the current implementation has a small (and |
| 1347 | /// rarely relevant) limitation: if `count` is at least 2 and the data pointed to by `x` or `y` |
| 1348 | /// contains a pointer that crosses the boundary of two `T`-sized chunks of memory, the function may |
| 1349 | /// fail to evaluate (similar to a panic during const-evaluation). This behavior may change in the |
| 1350 | /// future. |
| 1351 | /// |
| 1352 | /// The limitation is illustrated by the following example: |
| 1353 | /// |
| 1354 | /// ``` |
| 1355 | /// use std::mem::size_of; |
| 1356 | /// use std::ptr; |
| 1357 | /// |
| 1358 | /// const { unsafe { |
| 1359 | /// const PTR_SIZE: usize = size_of::<*const i32>(); |
| 1360 | /// let mut data1 = [0u8; PTR_SIZE]; |
| 1361 | /// let mut data2 = [0u8; PTR_SIZE]; |
| 1362 | /// // Store a pointer in `data1`. |
| 1363 | /// data1.as_mut_ptr().cast::<*const i32>().write_unaligned(&42); |
| 1364 | /// // Swap the contents of `data1` and `data2` by swapping `PTR_SIZE` many `u8`-sized chunks. |
| 1365 | /// // This call will fail, because the pointer in `data1` crosses the boundary |
| 1366 | /// // between several of the 1-byte chunks that are being swapped here. |
| 1367 | /// //ptr::swap_nonoverlapping(data1.as_mut_ptr(), data2.as_mut_ptr(), PTR_SIZE); |
| 1368 | /// // Swap the contents of `data1` and `data2` by swapping a single chunk of size |
| 1369 | /// // `[u8; PTR_SIZE]`. That works, as there is no pointer crossing the boundary between |
| 1370 | /// // two chunks. |
| 1371 | /// ptr::swap_nonoverlapping(&mut data1, &mut data2, 1); |
| 1372 | /// // Read the pointer from `data2` and dereference it. |
| 1373 | /// let ptr = data2.as_ptr().cast::<*const i32>().read_unaligned(); |
| 1374 | /// assert!(*ptr == 42); |
| 1375 | /// } } |
| 1376 | /// ``` |
| 1377 | #[inline ] |
| 1378 | #[stable (feature = "swap_nonoverlapping" , since = "1.27.0" )] |
| 1379 | #[rustc_const_stable (feature = "const_swap_nonoverlapping" , since = "1.88.0" )] |
| 1380 | #[rustc_diagnostic_item = "ptr_swap_nonoverlapping" ] |
| 1381 | #[rustc_allow_const_fn_unstable (const_eval_select)] // both implementations behave the same |
| 1382 | #[track_caller ] |
| 1383 | pub const unsafe fn swap_nonoverlapping<T>(x: *mut T, y: *mut T, count: usize) { |
| 1384 | ub_checks::assert_unsafe_precondition!( |
| 1385 | check_library_ub, |
| 1386 | "ptr::swap_nonoverlapping requires that both pointer arguments are aligned and non-null \ |
| 1387 | and the specified memory ranges do not overlap" , |
| 1388 | ( |
| 1389 | x: *mut () = x as *mut (), |
| 1390 | y: *mut () = y as *mut (), |
| 1391 | size: usize = size_of::<T>(), |
| 1392 | align: usize = align_of::<T>(), |
| 1393 | count: usize = count, |
| 1394 | ) => { |
| 1395 | let zero_size = size == 0 || count == 0; |
| 1396 | ub_checks::maybe_is_aligned_and_not_null(x, align, zero_size) |
| 1397 | && ub_checks::maybe_is_aligned_and_not_null(y, align, zero_size) |
| 1398 | && ub_checks::maybe_is_nonoverlapping(x, y, size, count) |
| 1399 | } |
| 1400 | ); |
| 1401 | |
| 1402 | const_eval_select!( |
| 1403 | @capture[T] { x: *mut T, y: *mut T, count: usize }: |
| 1404 | if const { |
| 1405 | // At compile-time we want to always copy this in chunks of `T`, to ensure that if there |
| 1406 | // are pointers inside `T` we will copy them in one go rather than trying to copy a part |
| 1407 | // of a pointer (which would not work). |
| 1408 | // SAFETY: Same preconditions as this function |
| 1409 | unsafe { swap_nonoverlapping_const(x, y, count) } |
| 1410 | } else { |
| 1411 | // Going though a slice here helps codegen know the size fits in `isize` |
| 1412 | let slice = slice_from_raw_parts_mut(x, count); |
| 1413 | // SAFETY: This is all readable from the pointer, meaning it's one |
| 1414 | // allocation, and thus cannot be more than isize::MAX bytes. |
| 1415 | let bytes = unsafe { mem::size_of_val_raw::<[T]>(slice) }; |
| 1416 | if let Some(bytes) = NonZero::new(bytes) { |
| 1417 | // SAFETY: These are the same ranges, just expressed in a different |
| 1418 | // type, so they're still non-overlapping. |
| 1419 | unsafe { swap_nonoverlapping_bytes(x.cast(), y.cast(), bytes) }; |
| 1420 | } |
| 1421 | } |
| 1422 | ) |
| 1423 | } |
| 1424 | |
| 1425 | /// Same behavior and safety conditions as [`swap_nonoverlapping`] |
| 1426 | #[inline ] |
| 1427 | const unsafe fn swap_nonoverlapping_const<T>(x: *mut T, y: *mut T, count: usize) { |
| 1428 | let mut i: usize = 0; |
| 1429 | while i < count { |
| 1430 | // SAFETY: By precondition, `i` is in-bounds because it's below `n` |
| 1431 | let x: *mut T = unsafe { x.add(count:i) }; |
| 1432 | // SAFETY: By precondition, `i` is in-bounds because it's below `n` |
| 1433 | // and it's distinct from `x` since the ranges are non-overlapping |
| 1434 | let y: *mut T = unsafe { y.add(count:i) }; |
| 1435 | |
| 1436 | // SAFETY: we're only ever given pointers that are valid to read/write, |
| 1437 | // including being aligned, and nothing here panics so it's drop-safe. |
| 1438 | unsafe { |
| 1439 | // Note that it's critical that these use `copy_nonoverlapping`, |
| 1440 | // rather than `read`/`write`, to avoid #134713 if T has padding. |
| 1441 | let mut temp: MaybeUninit = MaybeUninit::<T>::uninit(); |
| 1442 | copy_nonoverlapping(src:x, dst:temp.as_mut_ptr(), count:1); |
| 1443 | copy_nonoverlapping(src:y, dst:x, count:1); |
| 1444 | copy_nonoverlapping(src:temp.as_ptr(), dst:y, count:1); |
| 1445 | } |
| 1446 | |
| 1447 | i += 1; |
| 1448 | } |
| 1449 | } |
| 1450 | |
| 1451 | // Don't let MIR inline this, because we really want it to keep its noalias metadata |
| 1452 | #[rustc_no_mir_inline ] |
| 1453 | #[inline ] |
| 1454 | fn swap_chunk<const N: usize>(x: &mut MaybeUninit<[u8; N]>, y: &mut MaybeUninit<[u8; N]>) { |
| 1455 | let a: MaybeUninit<[u8; N]> = *x; |
| 1456 | let b: MaybeUninit<[u8; N]> = *y; |
| 1457 | *x = b; |
| 1458 | *y = a; |
| 1459 | } |
| 1460 | |
| 1461 | #[inline ] |
| 1462 | unsafe fn swap_nonoverlapping_bytes(x: *mut u8, y: *mut u8, bytes: NonZero<usize>) { |
| 1463 | // Same as `swap_nonoverlapping::<[u8; N]>`. |
| 1464 | unsafe fn swap_nonoverlapping_chunks<const N: usize>( |
| 1465 | x: *mut MaybeUninit<[u8; N]>, |
| 1466 | y: *mut MaybeUninit<[u8; N]>, |
| 1467 | chunks: NonZero<usize>, |
| 1468 | ) { |
| 1469 | let chunks = chunks.get(); |
| 1470 | for i in 0..chunks { |
| 1471 | // SAFETY: i is in [0, chunks) so the adds and dereferences are in-bounds. |
| 1472 | unsafe { swap_chunk(&mut *x.add(i), &mut *y.add(i)) }; |
| 1473 | } |
| 1474 | } |
| 1475 | |
| 1476 | // Same as `swap_nonoverlapping_bytes`, but accepts at most 1+2+4=7 bytes |
| 1477 | #[inline ] |
| 1478 | unsafe fn swap_nonoverlapping_short(x: *mut u8, y: *mut u8, bytes: NonZero<usize>) { |
| 1479 | // Tail handling for auto-vectorized code sometimes has element-at-a-time behaviour, |
| 1480 | // see <https://github.com/rust-lang/rust/issues/134946>. |
| 1481 | // By swapping as different sizes, rather than as a loop over bytes, |
| 1482 | // we make sure not to end up with, say, seven byte-at-a-time copies. |
| 1483 | |
| 1484 | let bytes = bytes.get(); |
| 1485 | let mut i = 0; |
| 1486 | macro_rules! swap_prefix { |
| 1487 | ($($n:literal)+) => {$( |
| 1488 | if (bytes & $n) != 0 { |
| 1489 | // SAFETY: `i` can only have the same bits set as those in bytes, |
| 1490 | // so these `add`s are in-bounds of `bytes`. But the bit for |
| 1491 | // `$n` hasn't been set yet, so the `$n` bytes that `swap_chunk` |
| 1492 | // will read and write are within the usable range. |
| 1493 | unsafe { swap_chunk::<$n>(&mut*x.add(i).cast(), &mut*y.add(i).cast()) }; |
| 1494 | i |= $n; |
| 1495 | } |
| 1496 | )+}; |
| 1497 | } |
| 1498 | swap_prefix!(4 2 1); |
| 1499 | debug_assert_eq!(i, bytes); |
| 1500 | } |
| 1501 | |
| 1502 | const CHUNK_SIZE: usize = size_of::<*const ()>(); |
| 1503 | let bytes = bytes.get(); |
| 1504 | |
| 1505 | let chunks = bytes / CHUNK_SIZE; |
| 1506 | let tail = bytes % CHUNK_SIZE; |
| 1507 | if let Some(chunks) = NonZero::new(chunks) { |
| 1508 | // SAFETY: this is bytes/CHUNK_SIZE*CHUNK_SIZE bytes, which is <= bytes, |
| 1509 | // so it's within the range of our non-overlapping bytes. |
| 1510 | unsafe { swap_nonoverlapping_chunks::<CHUNK_SIZE>(x.cast(), y.cast(), chunks) }; |
| 1511 | } |
| 1512 | if let Some(tail) = NonZero::new(tail) { |
| 1513 | const { assert!(CHUNK_SIZE <= 8) }; |
| 1514 | let delta = chunks * CHUNK_SIZE; |
| 1515 | // SAFETY: the tail length is below CHUNK SIZE because of the remainder, |
| 1516 | // and CHUNK_SIZE is at most 8 by the const assert, so tail <= 7 |
| 1517 | unsafe { swap_nonoverlapping_short(x.add(delta), y.add(delta), tail) }; |
| 1518 | } |
| 1519 | } |
| 1520 | |
| 1521 | /// Moves `src` into the pointed `dst`, returning the previous `dst` value. |
| 1522 | /// |
| 1523 | /// Neither value is dropped. |
| 1524 | /// |
| 1525 | /// This function is semantically equivalent to [`mem::replace`] except that it |
| 1526 | /// operates on raw pointers instead of references. When references are |
| 1527 | /// available, [`mem::replace`] should be preferred. |
| 1528 | /// |
| 1529 | /// # Safety |
| 1530 | /// |
| 1531 | /// Behavior is undefined if any of the following conditions are violated: |
| 1532 | /// |
| 1533 | /// * `dst` must be [valid] for both reads and writes. |
| 1534 | /// |
| 1535 | /// * `dst` must be properly aligned. |
| 1536 | /// |
| 1537 | /// * `dst` must point to a properly initialized value of type `T`. |
| 1538 | /// |
| 1539 | /// Note that even if `T` has size `0`, the pointer must be properly aligned. |
| 1540 | /// |
| 1541 | /// [valid]: self#safety |
| 1542 | /// |
| 1543 | /// # Examples |
| 1544 | /// |
| 1545 | /// ``` |
| 1546 | /// use std::ptr; |
| 1547 | /// |
| 1548 | /// let mut rust = vec!['b' , 'u' , 's' , 't' ]; |
| 1549 | /// |
| 1550 | /// // `mem::replace` would have the same effect without requiring the unsafe |
| 1551 | /// // block. |
| 1552 | /// let b = unsafe { |
| 1553 | /// ptr::replace(&mut rust[0], 'r' ) |
| 1554 | /// }; |
| 1555 | /// |
| 1556 | /// assert_eq!(b, 'b' ); |
| 1557 | /// assert_eq!(rust, &['r' , 'u' , 's' , 't' ]); |
| 1558 | /// ``` |
| 1559 | #[inline ] |
| 1560 | #[stable (feature = "rust1" , since = "1.0.0" )] |
| 1561 | #[rustc_const_stable (feature = "const_replace" , since = "1.83.0" )] |
| 1562 | #[rustc_diagnostic_item = "ptr_replace" ] |
| 1563 | #[track_caller ] |
| 1564 | pub const unsafe fn replace<T>(dst: *mut T, src: T) -> T { |
| 1565 | // SAFETY: the caller must guarantee that `dst` is valid to be |
| 1566 | // cast to a mutable reference (valid for writes, aligned, initialized), |
| 1567 | // and cannot overlap `src` since `dst` must point to a distinct |
| 1568 | // allocation. |
| 1569 | unsafe { |
| 1570 | ub_checks::assert_unsafe_precondition!( |
| 1571 | check_language_ub, |
| 1572 | "ptr::replace requires that the pointer argument is aligned and non-null" , |
| 1573 | ( |
| 1574 | addr: *const () = dst as *const (), |
| 1575 | align: usize = align_of::<T>(), |
| 1576 | is_zst: bool = T::IS_ZST, |
| 1577 | ) => ub_checks::maybe_is_aligned_and_not_null(addr, align, is_zst) |
| 1578 | ); |
| 1579 | mem::replace(&mut *dst, src) |
| 1580 | } |
| 1581 | } |
| 1582 | |
| 1583 | /// Reads the value from `src` without moving it. This leaves the |
| 1584 | /// memory in `src` unchanged. |
| 1585 | /// |
| 1586 | /// # Safety |
| 1587 | /// |
| 1588 | /// Behavior is undefined if any of the following conditions are violated: |
| 1589 | /// |
| 1590 | /// * `src` must be [valid] for reads. |
| 1591 | /// |
| 1592 | /// * `src` must be properly aligned. Use [`read_unaligned`] if this is not the |
| 1593 | /// case. |
| 1594 | /// |
| 1595 | /// * `src` must point to a properly initialized value of type `T`. |
| 1596 | /// |
| 1597 | /// Note that even if `T` has size `0`, the pointer must be properly aligned. |
| 1598 | /// |
| 1599 | /// # Examples |
| 1600 | /// |
| 1601 | /// Basic usage: |
| 1602 | /// |
| 1603 | /// ``` |
| 1604 | /// let x = 12; |
| 1605 | /// let y = &x as *const i32; |
| 1606 | /// |
| 1607 | /// unsafe { |
| 1608 | /// assert_eq!(std::ptr::read(y), 12); |
| 1609 | /// } |
| 1610 | /// ``` |
| 1611 | /// |
| 1612 | /// Manually implement [`mem::swap`]: |
| 1613 | /// |
| 1614 | /// ``` |
| 1615 | /// use std::ptr; |
| 1616 | /// |
| 1617 | /// fn swap<T>(a: &mut T, b: &mut T) { |
| 1618 | /// unsafe { |
| 1619 | /// // Create a bitwise copy of the value at `a` in `tmp`. |
| 1620 | /// let tmp = ptr::read(a); |
| 1621 | /// |
| 1622 | /// // Exiting at this point (either by explicitly returning or by |
| 1623 | /// // calling a function which panics) would cause the value in `tmp` to |
| 1624 | /// // be dropped while the same value is still referenced by `a`. This |
| 1625 | /// // could trigger undefined behavior if `T` is not `Copy`. |
| 1626 | /// |
| 1627 | /// // Create a bitwise copy of the value at `b` in `a`. |
| 1628 | /// // This is safe because mutable references cannot alias. |
| 1629 | /// ptr::copy_nonoverlapping(b, a, 1); |
| 1630 | /// |
| 1631 | /// // As above, exiting here could trigger undefined behavior because |
| 1632 | /// // the same value is referenced by `a` and `b`. |
| 1633 | /// |
| 1634 | /// // Move `tmp` into `b`. |
| 1635 | /// ptr::write(b, tmp); |
| 1636 | /// |
| 1637 | /// // `tmp` has been moved (`write` takes ownership of its second argument), |
| 1638 | /// // so nothing is dropped implicitly here. |
| 1639 | /// } |
| 1640 | /// } |
| 1641 | /// |
| 1642 | /// let mut foo = "foo" .to_owned(); |
| 1643 | /// let mut bar = "bar" .to_owned(); |
| 1644 | /// |
| 1645 | /// swap(&mut foo, &mut bar); |
| 1646 | /// |
| 1647 | /// assert_eq!(foo, "bar" ); |
| 1648 | /// assert_eq!(bar, "foo" ); |
| 1649 | /// ``` |
| 1650 | /// |
| 1651 | /// ## Ownership of the Returned Value |
| 1652 | /// |
| 1653 | /// `read` creates a bitwise copy of `T`, regardless of whether `T` is [`Copy`]. |
| 1654 | /// If `T` is not [`Copy`], using both the returned value and the value at |
| 1655 | /// `*src` can violate memory safety. Note that assigning to `*src` counts as a |
| 1656 | /// use because it will attempt to drop the value at `*src`. |
| 1657 | /// |
| 1658 | /// [`write()`] can be used to overwrite data without causing it to be dropped. |
| 1659 | /// |
| 1660 | /// ``` |
| 1661 | /// use std::ptr; |
| 1662 | /// |
| 1663 | /// let mut s = String::from("foo" ); |
| 1664 | /// unsafe { |
| 1665 | /// // `s2` now points to the same underlying memory as `s`. |
| 1666 | /// let mut s2: String = ptr::read(&s); |
| 1667 | /// |
| 1668 | /// assert_eq!(s2, "foo" ); |
| 1669 | /// |
| 1670 | /// // Assigning to `s2` causes its original value to be dropped. Beyond |
| 1671 | /// // this point, `s` must no longer be used, as the underlying memory has |
| 1672 | /// // been freed. |
| 1673 | /// s2 = String::default(); |
| 1674 | /// assert_eq!(s2, "" ); |
| 1675 | /// |
| 1676 | /// // Assigning to `s` would cause the old value to be dropped again, |
| 1677 | /// // resulting in undefined behavior. |
| 1678 | /// // s = String::from("bar"); // ERROR |
| 1679 | /// |
| 1680 | /// // `ptr::write` can be used to overwrite a value without dropping it. |
| 1681 | /// ptr::write(&mut s, String::from("bar" )); |
| 1682 | /// } |
| 1683 | /// |
| 1684 | /// assert_eq!(s, "bar" ); |
| 1685 | /// ``` |
| 1686 | /// |
| 1687 | /// [valid]: self#safety |
| 1688 | #[inline ] |
| 1689 | #[stable (feature = "rust1" , since = "1.0.0" )] |
| 1690 | #[rustc_const_stable (feature = "const_ptr_read" , since = "1.71.0" )] |
| 1691 | #[track_caller ] |
| 1692 | #[rustc_diagnostic_item = "ptr_read" ] |
| 1693 | pub const unsafe fn read<T>(src: *const T) -> T { |
| 1694 | // It would be semantically correct to implement this via `copy_nonoverlapping` |
| 1695 | // and `MaybeUninit`, as was done before PR #109035. Calling `assume_init` |
| 1696 | // provides enough information to know that this is a typed operation. |
| 1697 | |
| 1698 | // However, as of March 2023 the compiler was not capable of taking advantage |
| 1699 | // of that information. Thus, the implementation here switched to an intrinsic, |
| 1700 | // which lowers to `_0 = *src` in MIR, to address a few issues: |
| 1701 | // |
| 1702 | // - Using `MaybeUninit::assume_init` after a `copy_nonoverlapping` was not |
| 1703 | // turning the untyped copy into a typed load. As such, the generated |
| 1704 | // `load` in LLVM didn't get various metadata, such as `!range` (#73258), |
| 1705 | // `!nonnull`, and `!noundef`, resulting in poorer optimization. |
| 1706 | // - Going through the extra local resulted in multiple extra copies, even |
| 1707 | // in optimized MIR. (Ignoring StorageLive/Dead, the intrinsic is one |
| 1708 | // MIR statement, while the previous implementation was eight.) LLVM |
| 1709 | // could sometimes optimize them away, but because `read` is at the core |
| 1710 | // of so many things, not having them in the first place improves what we |
| 1711 | // hand off to the backend. For example, `mem::replace::<Big>` previously |
| 1712 | // emitted 4 `alloca` and 6 `memcpy`s, but is now 1 `alloc` and 3 `memcpy`s. |
| 1713 | // - In general, this approach keeps us from getting any more bugs (like |
| 1714 | // #106369) that boil down to "`read(p)` is worse than `*p`", as this |
| 1715 | // makes them look identical to the backend (or other MIR consumers). |
| 1716 | // |
| 1717 | // Future enhancements to MIR optimizations might well allow this to return |
| 1718 | // to the previous implementation, rather than using an intrinsic. |
| 1719 | |
| 1720 | // SAFETY: the caller must guarantee that `src` is valid for reads. |
| 1721 | unsafe { |
| 1722 | #[cfg (debug_assertions)] // Too expensive to always enable (for now?) |
| 1723 | ub_checks::assert_unsafe_precondition!( |
| 1724 | check_language_ub, |
| 1725 | "ptr::read requires that the pointer argument is aligned and non-null" , |
| 1726 | ( |
| 1727 | addr: *const () = src as *const (), |
| 1728 | align: usize = align_of::<T>(), |
| 1729 | is_zst: bool = T::IS_ZST, |
| 1730 | ) => ub_checks::maybe_is_aligned_and_not_null(addr, align, is_zst) |
| 1731 | ); |
| 1732 | crate::intrinsics::read_via_copy(src) |
| 1733 | } |
| 1734 | } |
| 1735 | |
| 1736 | /// Reads the value from `src` without moving it. This leaves the |
| 1737 | /// memory in `src` unchanged. |
| 1738 | /// |
| 1739 | /// Unlike [`read`], `read_unaligned` works with unaligned pointers. |
| 1740 | /// |
| 1741 | /// # Safety |
| 1742 | /// |
| 1743 | /// Behavior is undefined if any of the following conditions are violated: |
| 1744 | /// |
| 1745 | /// * `src` must be [valid] for reads. |
| 1746 | /// |
| 1747 | /// * `src` must point to a properly initialized value of type `T`. |
| 1748 | /// |
| 1749 | /// Like [`read`], `read_unaligned` creates a bitwise copy of `T`, regardless of |
| 1750 | /// whether `T` is [`Copy`]. If `T` is not [`Copy`], using both the returned |
| 1751 | /// value and the value at `*src` can [violate memory safety][read-ownership]. |
| 1752 | /// |
| 1753 | /// [read-ownership]: read#ownership-of-the-returned-value |
| 1754 | /// [valid]: self#safety |
| 1755 | /// |
| 1756 | /// ## On `packed` structs |
| 1757 | /// |
| 1758 | /// Attempting to create a raw pointer to an `unaligned` struct field with |
| 1759 | /// an expression such as `&packed.unaligned as *const FieldType` creates an |
| 1760 | /// intermediate unaligned reference before converting that to a raw pointer. |
| 1761 | /// That this reference is temporary and immediately cast is inconsequential |
| 1762 | /// as the compiler always expects references to be properly aligned. |
| 1763 | /// As a result, using `&packed.unaligned as *const FieldType` causes immediate |
| 1764 | /// *undefined behavior* in your program. |
| 1765 | /// |
| 1766 | /// Instead you must use the `&raw const` syntax to create the pointer. |
| 1767 | /// You may use that constructed pointer together with this function. |
| 1768 | /// |
| 1769 | /// An example of what not to do and how this relates to `read_unaligned` is: |
| 1770 | /// |
| 1771 | /// ``` |
| 1772 | /// #[repr(packed, C)] |
| 1773 | /// struct Packed { |
| 1774 | /// _padding: u8, |
| 1775 | /// unaligned: u32, |
| 1776 | /// } |
| 1777 | /// |
| 1778 | /// let packed = Packed { |
| 1779 | /// _padding: 0x00, |
| 1780 | /// unaligned: 0x01020304, |
| 1781 | /// }; |
| 1782 | /// |
| 1783 | /// // Take the address of a 32-bit integer which is not aligned. |
| 1784 | /// // In contrast to `&packed.unaligned as *const _`, this has no undefined behavior. |
| 1785 | /// let unaligned = &raw const packed.unaligned; |
| 1786 | /// |
| 1787 | /// let v = unsafe { std::ptr::read_unaligned(unaligned) }; |
| 1788 | /// assert_eq!(v, 0x01020304); |
| 1789 | /// ``` |
| 1790 | /// |
| 1791 | /// Accessing unaligned fields directly with e.g. `packed.unaligned` is safe however. |
| 1792 | /// |
| 1793 | /// # Examples |
| 1794 | /// |
| 1795 | /// Read a `usize` value from a byte buffer: |
| 1796 | /// |
| 1797 | /// ``` |
| 1798 | /// fn read_usize(x: &[u8]) -> usize { |
| 1799 | /// assert!(x.len() >= size_of::<usize>()); |
| 1800 | /// |
| 1801 | /// let ptr = x.as_ptr() as *const usize; |
| 1802 | /// |
| 1803 | /// unsafe { ptr.read_unaligned() } |
| 1804 | /// } |
| 1805 | /// ``` |
| 1806 | #[inline ] |
| 1807 | #[stable (feature = "ptr_unaligned" , since = "1.17.0" )] |
| 1808 | #[rustc_const_stable (feature = "const_ptr_read" , since = "1.71.0" )] |
| 1809 | #[track_caller ] |
| 1810 | #[rustc_diagnostic_item = "ptr_read_unaligned" ] |
| 1811 | pub const unsafe fn read_unaligned<T>(src: *const T) -> T { |
| 1812 | let mut tmp: MaybeUninit = MaybeUninit::<T>::uninit(); |
| 1813 | // SAFETY: the caller must guarantee that `src` is valid for reads. |
| 1814 | // `src` cannot overlap `tmp` because `tmp` was just allocated on |
| 1815 | // the stack as a separate allocation. |
| 1816 | // |
| 1817 | // Also, since we just wrote a valid value into `tmp`, it is guaranteed |
| 1818 | // to be properly initialized. |
| 1819 | unsafe { |
| 1820 | copy_nonoverlapping(src as *const u8, dst:tmp.as_mut_ptr() as *mut u8, count:size_of::<T>()); |
| 1821 | tmp.assume_init() |
| 1822 | } |
| 1823 | } |
| 1824 | |
| 1825 | /// Overwrites a memory location with the given value without reading or |
| 1826 | /// dropping the old value. |
| 1827 | /// |
| 1828 | /// `write` does not drop the contents of `dst`. This is safe, but it could leak |
| 1829 | /// allocations or resources, so care should be taken not to overwrite an object |
| 1830 | /// that should be dropped. |
| 1831 | /// |
| 1832 | /// Additionally, it does not drop `src`. Semantically, `src` is moved into the |
| 1833 | /// location pointed to by `dst`. |
| 1834 | /// |
| 1835 | /// This is appropriate for initializing uninitialized memory, or overwriting |
| 1836 | /// memory that has previously been [`read`] from. |
| 1837 | /// |
| 1838 | /// # Safety |
| 1839 | /// |
| 1840 | /// Behavior is undefined if any of the following conditions are violated: |
| 1841 | /// |
| 1842 | /// * `dst` must be [valid] for writes. |
| 1843 | /// |
| 1844 | /// * `dst` must be properly aligned. Use [`write_unaligned`] if this is not the |
| 1845 | /// case. |
| 1846 | /// |
| 1847 | /// Note that even if `T` has size `0`, the pointer must be properly aligned. |
| 1848 | /// |
| 1849 | /// [valid]: self#safety |
| 1850 | /// |
| 1851 | /// # Examples |
| 1852 | /// |
| 1853 | /// Basic usage: |
| 1854 | /// |
| 1855 | /// ``` |
| 1856 | /// let mut x = 0; |
| 1857 | /// let y = &mut x as *mut i32; |
| 1858 | /// let z = 12; |
| 1859 | /// |
| 1860 | /// unsafe { |
| 1861 | /// std::ptr::write(y, z); |
| 1862 | /// assert_eq!(std::ptr::read(y), 12); |
| 1863 | /// } |
| 1864 | /// ``` |
| 1865 | /// |
| 1866 | /// Manually implement [`mem::swap`]: |
| 1867 | /// |
| 1868 | /// ``` |
| 1869 | /// use std::ptr; |
| 1870 | /// |
| 1871 | /// fn swap<T>(a: &mut T, b: &mut T) { |
| 1872 | /// unsafe { |
| 1873 | /// // Create a bitwise copy of the value at `a` in `tmp`. |
| 1874 | /// let tmp = ptr::read(a); |
| 1875 | /// |
| 1876 | /// // Exiting at this point (either by explicitly returning or by |
| 1877 | /// // calling a function which panics) would cause the value in `tmp` to |
| 1878 | /// // be dropped while the same value is still referenced by `a`. This |
| 1879 | /// // could trigger undefined behavior if `T` is not `Copy`. |
| 1880 | /// |
| 1881 | /// // Create a bitwise copy of the value at `b` in `a`. |
| 1882 | /// // This is safe because mutable references cannot alias. |
| 1883 | /// ptr::copy_nonoverlapping(b, a, 1); |
| 1884 | /// |
| 1885 | /// // As above, exiting here could trigger undefined behavior because |
| 1886 | /// // the same value is referenced by `a` and `b`. |
| 1887 | /// |
| 1888 | /// // Move `tmp` into `b`. |
| 1889 | /// ptr::write(b, tmp); |
| 1890 | /// |
| 1891 | /// // `tmp` has been moved (`write` takes ownership of its second argument), |
| 1892 | /// // so nothing is dropped implicitly here. |
| 1893 | /// } |
| 1894 | /// } |
| 1895 | /// |
| 1896 | /// let mut foo = "foo" .to_owned(); |
| 1897 | /// let mut bar = "bar" .to_owned(); |
| 1898 | /// |
| 1899 | /// swap(&mut foo, &mut bar); |
| 1900 | /// |
| 1901 | /// assert_eq!(foo, "bar" ); |
| 1902 | /// assert_eq!(bar, "foo" ); |
| 1903 | /// ``` |
| 1904 | #[inline ] |
| 1905 | #[stable (feature = "rust1" , since = "1.0.0" )] |
| 1906 | #[rustc_const_stable (feature = "const_ptr_write" , since = "1.83.0" )] |
| 1907 | #[rustc_diagnostic_item = "ptr_write" ] |
| 1908 | #[track_caller ] |
| 1909 | pub const unsafe fn write<T>(dst: *mut T, src: T) { |
| 1910 | // Semantically, it would be fine for this to be implemented as a |
| 1911 | // `copy_nonoverlapping` and appropriate drop suppression of `src`. |
| 1912 | |
| 1913 | // However, implementing via that currently produces more MIR than is ideal. |
| 1914 | // Using an intrinsic keeps it down to just the simple `*dst = move src` in |
| 1915 | // MIR (11 statements shorter, at the time of writing), and also allows |
| 1916 | // `src` to stay an SSA value in codegen_ssa, rather than a memory one. |
| 1917 | |
| 1918 | // SAFETY: the caller must guarantee that `dst` is valid for writes. |
| 1919 | // `dst` cannot overlap `src` because the caller has mutable access |
| 1920 | // to `dst` while `src` is owned by this function. |
| 1921 | unsafe { |
| 1922 | #[cfg (debug_assertions)] // Too expensive to always enable (for now?) |
| 1923 | ub_checks::assert_unsafe_precondition!( |
| 1924 | check_language_ub, |
| 1925 | "ptr::write requires that the pointer argument is aligned and non-null" , |
| 1926 | ( |
| 1927 | addr: *mut () = dst as *mut (), |
| 1928 | align: usize = align_of::<T>(), |
| 1929 | is_zst: bool = T::IS_ZST, |
| 1930 | ) => ub_checks::maybe_is_aligned_and_not_null(addr, align, is_zst) |
| 1931 | ); |
| 1932 | intrinsics::write_via_move(dst, src) |
| 1933 | } |
| 1934 | } |
| 1935 | |
| 1936 | /// Overwrites a memory location with the given value without reading or |
| 1937 | /// dropping the old value. |
| 1938 | /// |
| 1939 | /// Unlike [`write()`], the pointer may be unaligned. |
| 1940 | /// |
| 1941 | /// `write_unaligned` does not drop the contents of `dst`. This is safe, but it |
| 1942 | /// could leak allocations or resources, so care should be taken not to overwrite |
| 1943 | /// an object that should be dropped. |
| 1944 | /// |
| 1945 | /// Additionally, it does not drop `src`. Semantically, `src` is moved into the |
| 1946 | /// location pointed to by `dst`. |
| 1947 | /// |
| 1948 | /// This is appropriate for initializing uninitialized memory, or overwriting |
| 1949 | /// memory that has previously been read with [`read_unaligned`]. |
| 1950 | /// |
| 1951 | /// # Safety |
| 1952 | /// |
| 1953 | /// Behavior is undefined if any of the following conditions are violated: |
| 1954 | /// |
| 1955 | /// * `dst` must be [valid] for writes. |
| 1956 | /// |
| 1957 | /// [valid]: self#safety |
| 1958 | /// |
| 1959 | /// ## On `packed` structs |
| 1960 | /// |
| 1961 | /// Attempting to create a raw pointer to an `unaligned` struct field with |
| 1962 | /// an expression such as `&packed.unaligned as *const FieldType` creates an |
| 1963 | /// intermediate unaligned reference before converting that to a raw pointer. |
| 1964 | /// That this reference is temporary and immediately cast is inconsequential |
| 1965 | /// as the compiler always expects references to be properly aligned. |
| 1966 | /// As a result, using `&packed.unaligned as *const FieldType` causes immediate |
| 1967 | /// *undefined behavior* in your program. |
| 1968 | /// |
| 1969 | /// Instead, you must use the `&raw mut` syntax to create the pointer. |
| 1970 | /// You may use that constructed pointer together with this function. |
| 1971 | /// |
| 1972 | /// An example of how to do it and how this relates to `write_unaligned` is: |
| 1973 | /// |
| 1974 | /// ``` |
| 1975 | /// #[repr(packed, C)] |
| 1976 | /// struct Packed { |
| 1977 | /// _padding: u8, |
| 1978 | /// unaligned: u32, |
| 1979 | /// } |
| 1980 | /// |
| 1981 | /// let mut packed: Packed = unsafe { std::mem::zeroed() }; |
| 1982 | /// |
| 1983 | /// // Take the address of a 32-bit integer which is not aligned. |
| 1984 | /// // In contrast to `&packed.unaligned as *mut _`, this has no undefined behavior. |
| 1985 | /// let unaligned = &raw mut packed.unaligned; |
| 1986 | /// |
| 1987 | /// unsafe { std::ptr::write_unaligned(unaligned, 42) }; |
| 1988 | /// |
| 1989 | /// assert_eq!({packed.unaligned}, 42); // `{...}` forces copying the field instead of creating a reference. |
| 1990 | /// ``` |
| 1991 | /// |
| 1992 | /// Accessing unaligned fields directly with e.g. `packed.unaligned` is safe however |
| 1993 | /// (as can be seen in the `assert_eq!` above). |
| 1994 | /// |
| 1995 | /// # Examples |
| 1996 | /// |
| 1997 | /// Write a `usize` value to a byte buffer: |
| 1998 | /// |
| 1999 | /// ``` |
| 2000 | /// fn write_usize(x: &mut [u8], val: usize) { |
| 2001 | /// assert!(x.len() >= size_of::<usize>()); |
| 2002 | /// |
| 2003 | /// let ptr = x.as_mut_ptr() as *mut usize; |
| 2004 | /// |
| 2005 | /// unsafe { ptr.write_unaligned(val) } |
| 2006 | /// } |
| 2007 | /// ``` |
| 2008 | #[inline ] |
| 2009 | #[stable (feature = "ptr_unaligned" , since = "1.17.0" )] |
| 2010 | #[rustc_const_stable (feature = "const_ptr_write" , since = "1.83.0" )] |
| 2011 | #[rustc_diagnostic_item = "ptr_write_unaligned" ] |
| 2012 | #[track_caller ] |
| 2013 | pub const unsafe fn write_unaligned<T>(dst: *mut T, src: T) { |
| 2014 | // SAFETY: the caller must guarantee that `dst` is valid for writes. |
| 2015 | // `dst` cannot overlap `src` because the caller has mutable access |
| 2016 | // to `dst` while `src` is owned by this function. |
| 2017 | unsafe { |
| 2018 | copy_nonoverlapping((&raw const src) as *const u8, dst as *mut u8, count:size_of::<T>()); |
| 2019 | // We are calling the intrinsic directly to avoid function calls in the generated code. |
| 2020 | intrinsics::forget(src); |
| 2021 | } |
| 2022 | } |
| 2023 | |
| 2024 | /// Performs a volatile read of the value from `src` without moving it. This |
| 2025 | /// leaves the memory in `src` unchanged. |
| 2026 | /// |
| 2027 | /// Volatile operations are intended to act on I/O memory, and are guaranteed |
| 2028 | /// to not be elided or reordered by the compiler across other volatile |
| 2029 | /// operations. |
| 2030 | /// |
| 2031 | /// # Notes |
| 2032 | /// |
| 2033 | /// Rust does not currently have a rigorously and formally defined memory model, |
| 2034 | /// so the precise semantics of what "volatile" means here is subject to change |
| 2035 | /// over time. That being said, the semantics will almost always end up pretty |
| 2036 | /// similar to [C11's definition of volatile][c11]. |
| 2037 | /// |
| 2038 | /// The compiler shouldn't change the relative order or number of volatile |
| 2039 | /// memory operations. However, volatile memory operations on zero-sized types |
| 2040 | /// (e.g., if a zero-sized type is passed to `read_volatile`) are noops |
| 2041 | /// and may be ignored. |
| 2042 | /// |
| 2043 | /// [c11]: http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1570.pdf |
| 2044 | /// |
| 2045 | /// # Safety |
| 2046 | /// |
| 2047 | /// Behavior is undefined if any of the following conditions are violated: |
| 2048 | /// |
| 2049 | /// * `src` must be [valid] for reads. |
| 2050 | /// |
| 2051 | /// * `src` must be properly aligned. |
| 2052 | /// |
| 2053 | /// * `src` must point to a properly initialized value of type `T`. |
| 2054 | /// |
| 2055 | /// Like [`read`], `read_volatile` creates a bitwise copy of `T`, regardless of |
| 2056 | /// whether `T` is [`Copy`]. If `T` is not [`Copy`], using both the returned |
| 2057 | /// value and the value at `*src` can [violate memory safety][read-ownership]. |
| 2058 | /// However, storing non-[`Copy`] types in volatile memory is almost certainly |
| 2059 | /// incorrect. |
| 2060 | /// |
| 2061 | /// Note that even if `T` has size `0`, the pointer must be properly aligned. |
| 2062 | /// |
| 2063 | /// [valid]: self#safety |
| 2064 | /// [read-ownership]: read#ownership-of-the-returned-value |
| 2065 | /// |
| 2066 | /// Just like in C, whether an operation is volatile has no bearing whatsoever |
| 2067 | /// on questions involving concurrent access from multiple threads. Volatile |
| 2068 | /// accesses behave exactly like non-atomic accesses in that regard. In particular, |
| 2069 | /// a race between a `read_volatile` and any write operation to the same location |
| 2070 | /// is undefined behavior. |
| 2071 | /// |
| 2072 | /// # Examples |
| 2073 | /// |
| 2074 | /// Basic usage: |
| 2075 | /// |
| 2076 | /// ``` |
| 2077 | /// let x = 12; |
| 2078 | /// let y = &x as *const i32; |
| 2079 | /// |
| 2080 | /// unsafe { |
| 2081 | /// assert_eq!(std::ptr::read_volatile(y), 12); |
| 2082 | /// } |
| 2083 | /// ``` |
| 2084 | #[inline ] |
| 2085 | #[stable (feature = "volatile" , since = "1.9.0" )] |
| 2086 | #[track_caller ] |
| 2087 | #[rustc_diagnostic_item = "ptr_read_volatile" ] |
| 2088 | pub unsafe fn read_volatile<T>(src: *const T) -> T { |
| 2089 | // SAFETY: the caller must uphold the safety contract for `volatile_load`. |
| 2090 | unsafe { |
| 2091 | ub_checks::assert_unsafe_precondition!( |
| 2092 | check_language_ub, |
| 2093 | "ptr::read_volatile requires that the pointer argument is aligned and non-null" , |
| 2094 | ( |
| 2095 | addr: *const () = src as *const (), |
| 2096 | align: usize = align_of::<T>(), |
| 2097 | is_zst: bool = T::IS_ZST, |
| 2098 | ) => ub_checks::maybe_is_aligned_and_not_null(addr, align, is_zst) |
| 2099 | ); |
| 2100 | intrinsics::volatile_load(src) |
| 2101 | } |
| 2102 | } |
| 2103 | |
| 2104 | /// Performs a volatile write of a memory location with the given value without |
| 2105 | /// reading or dropping the old value. |
| 2106 | /// |
| 2107 | /// Volatile operations are intended to act on I/O memory, and are guaranteed |
| 2108 | /// to not be elided or reordered by the compiler across other volatile |
| 2109 | /// operations. |
| 2110 | /// |
| 2111 | /// `write_volatile` does not drop the contents of `dst`. This is safe, but it |
| 2112 | /// could leak allocations or resources, so care should be taken not to overwrite |
| 2113 | /// an object that should be dropped. |
| 2114 | /// |
| 2115 | /// Additionally, it does not drop `src`. Semantically, `src` is moved into the |
| 2116 | /// location pointed to by `dst`. |
| 2117 | /// |
| 2118 | /// # Notes |
| 2119 | /// |
| 2120 | /// Rust does not currently have a rigorously and formally defined memory model, |
| 2121 | /// so the precise semantics of what "volatile" means here is subject to change |
| 2122 | /// over time. That being said, the semantics will almost always end up pretty |
| 2123 | /// similar to [C11's definition of volatile][c11]. |
| 2124 | /// |
| 2125 | /// The compiler shouldn't change the relative order or number of volatile |
| 2126 | /// memory operations. However, volatile memory operations on zero-sized types |
| 2127 | /// (e.g., if a zero-sized type is passed to `write_volatile`) are noops |
| 2128 | /// and may be ignored. |
| 2129 | /// |
| 2130 | /// [c11]: http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1570.pdf |
| 2131 | /// |
| 2132 | /// # Safety |
| 2133 | /// |
| 2134 | /// Behavior is undefined if any of the following conditions are violated: |
| 2135 | /// |
| 2136 | /// * `dst` must be [valid] for writes. |
| 2137 | /// |
| 2138 | /// * `dst` must be properly aligned. |
| 2139 | /// |
| 2140 | /// Note that even if `T` has size `0`, the pointer must be properly aligned. |
| 2141 | /// |
| 2142 | /// [valid]: self#safety |
| 2143 | /// |
| 2144 | /// Just like in C, whether an operation is volatile has no bearing whatsoever |
| 2145 | /// on questions involving concurrent access from multiple threads. Volatile |
| 2146 | /// accesses behave exactly like non-atomic accesses in that regard. In particular, |
| 2147 | /// a race between a `write_volatile` and any other operation (reading or writing) |
| 2148 | /// on the same location is undefined behavior. |
| 2149 | /// |
| 2150 | /// # Examples |
| 2151 | /// |
| 2152 | /// Basic usage: |
| 2153 | /// |
| 2154 | /// ``` |
| 2155 | /// let mut x = 0; |
| 2156 | /// let y = &mut x as *mut i32; |
| 2157 | /// let z = 12; |
| 2158 | /// |
| 2159 | /// unsafe { |
| 2160 | /// std::ptr::write_volatile(y, z); |
| 2161 | /// assert_eq!(std::ptr::read_volatile(y), 12); |
| 2162 | /// } |
| 2163 | /// ``` |
| 2164 | #[inline ] |
| 2165 | #[stable (feature = "volatile" , since = "1.9.0" )] |
| 2166 | #[rustc_diagnostic_item = "ptr_write_volatile" ] |
| 2167 | #[track_caller ] |
| 2168 | pub unsafe fn write_volatile<T>(dst: *mut T, src: T) { |
| 2169 | // SAFETY: the caller must uphold the safety contract for `volatile_store`. |
| 2170 | unsafe { |
| 2171 | ub_checks::assert_unsafe_precondition!( |
| 2172 | check_language_ub, |
| 2173 | "ptr::write_volatile requires that the pointer argument is aligned and non-null" , |
| 2174 | ( |
| 2175 | addr: *mut () = dst as *mut (), |
| 2176 | align: usize = align_of::<T>(), |
| 2177 | is_zst: bool = T::IS_ZST, |
| 2178 | ) => ub_checks::maybe_is_aligned_and_not_null(addr, align, is_zst) |
| 2179 | ); |
| 2180 | intrinsics::volatile_store(dst, val:src); |
| 2181 | } |
| 2182 | } |
| 2183 | |
| 2184 | /// Align pointer `p`. |
| 2185 | /// |
| 2186 | /// Calculate offset (in terms of elements of `size_of::<T>()` stride) that has to be applied |
| 2187 | /// to pointer `p` so that pointer `p` would get aligned to `a`. |
| 2188 | /// |
| 2189 | /// # Safety |
| 2190 | /// `a` must be a power of two. |
| 2191 | /// |
| 2192 | /// # Notes |
| 2193 | /// This implementation has been carefully tailored to not panic. It is UB for this to panic. |
| 2194 | /// The only real change that can be made here is change of `INV_TABLE_MOD_16` and associated |
| 2195 | /// constants. |
| 2196 | /// |
| 2197 | /// If we ever decide to make it possible to call the intrinsic with `a` that is not a |
| 2198 | /// power-of-two, it will probably be more prudent to just change to a naive implementation rather |
| 2199 | /// than trying to adapt this to accommodate that change. |
| 2200 | /// |
| 2201 | /// Any questions go to @nagisa. |
| 2202 | #[allow (ptr_to_integer_transmute_in_consts)] |
| 2203 | pub(crate) unsafe fn align_offset<T: Sized>(p: *const T, a: usize) -> usize { |
| 2204 | // FIXME(#75598): Direct use of these intrinsics improves codegen significantly at opt-level <= |
| 2205 | // 1, where the method versions of these operations are not inlined. |
| 2206 | use intrinsics::{ |
| 2207 | assume, cttz_nonzero, exact_div, mul_with_overflow, unchecked_rem, unchecked_shl, |
| 2208 | unchecked_shr, unchecked_sub, wrapping_add, wrapping_mul, wrapping_sub, |
| 2209 | }; |
| 2210 | |
| 2211 | /// Calculate multiplicative modular inverse of `x` modulo `m`. |
| 2212 | /// |
| 2213 | /// This implementation is tailored for `align_offset` and has following preconditions: |
| 2214 | /// |
| 2215 | /// * `m` is a power-of-two; |
| 2216 | /// * `x < m`; (if `x ≥ m`, pass in `x % m` instead) |
| 2217 | /// |
| 2218 | /// Implementation of this function shall not panic. Ever. |
| 2219 | #[inline ] |
| 2220 | const unsafe fn mod_inv(x: usize, m: usize) -> usize { |
| 2221 | /// Multiplicative modular inverse table modulo 2⁴ = 16. |
| 2222 | /// |
| 2223 | /// Note, that this table does not contain values where inverse does not exist (i.e., for |
| 2224 | /// `0⁻¹ mod 16`, `2⁻¹ mod 16`, etc.) |
| 2225 | const INV_TABLE_MOD_16: [u8; 8] = [1, 11, 13, 7, 9, 3, 5, 15]; |
| 2226 | /// Modulo for which the `INV_TABLE_MOD_16` is intended. |
| 2227 | const INV_TABLE_MOD: usize = 16; |
| 2228 | |
| 2229 | // SAFETY: `m` is required to be a power-of-two, hence non-zero. |
| 2230 | let m_minus_one = unsafe { unchecked_sub(m, 1) }; |
| 2231 | let mut inverse = INV_TABLE_MOD_16[(x & (INV_TABLE_MOD - 1)) >> 1] as usize; |
| 2232 | let mut mod_gate = INV_TABLE_MOD; |
| 2233 | // We iterate "up" using the following formula: |
| 2234 | // |
| 2235 | // $$ xy ≡ 1 (mod 2ⁿ) → xy (2 - xy) ≡ 1 (mod 2²ⁿ) $$ |
| 2236 | // |
| 2237 | // This application needs to be applied at least until `2²ⁿ ≥ m`, at which point we can |
| 2238 | // finally reduce the computation to our desired `m` by taking `inverse mod m`. |
| 2239 | // |
| 2240 | // This computation is `O(log log m)`, which is to say, that on 64-bit machines this loop |
| 2241 | // will always finish in at most 4 iterations. |
| 2242 | loop { |
| 2243 | // y = y * (2 - xy) mod n |
| 2244 | // |
| 2245 | // Note, that we use wrapping operations here intentionally – the original formula |
| 2246 | // uses e.g., subtraction `mod n`. It is entirely fine to do them `mod |
| 2247 | // usize::MAX` instead, because we take the result `mod n` at the end |
| 2248 | // anyway. |
| 2249 | if mod_gate >= m { |
| 2250 | break; |
| 2251 | } |
| 2252 | inverse = wrapping_mul(inverse, wrapping_sub(2usize, wrapping_mul(x, inverse))); |
| 2253 | let (new_gate, overflow) = mul_with_overflow(mod_gate, mod_gate); |
| 2254 | if overflow { |
| 2255 | break; |
| 2256 | } |
| 2257 | mod_gate = new_gate; |
| 2258 | } |
| 2259 | inverse & m_minus_one |
| 2260 | } |
| 2261 | |
| 2262 | let stride = size_of::<T>(); |
| 2263 | |
| 2264 | let addr: usize = p.addr(); |
| 2265 | |
| 2266 | // SAFETY: `a` is a power-of-two, therefore non-zero. |
| 2267 | let a_minus_one = unsafe { unchecked_sub(a, 1) }; |
| 2268 | |
| 2269 | if stride == 0 { |
| 2270 | // SPECIAL_CASE: handle 0-sized types. No matter how many times we step, the address will |
| 2271 | // stay the same, so no offset will be able to align the pointer unless it is already |
| 2272 | // aligned. This branch _will_ be optimized out as `stride` is known at compile-time. |
| 2273 | let p_mod_a = addr & a_minus_one; |
| 2274 | return if p_mod_a == 0 { 0 } else { usize::MAX }; |
| 2275 | } |
| 2276 | |
| 2277 | // SAFETY: `stride == 0` case has been handled by the special case above. |
| 2278 | let a_mod_stride = unsafe { unchecked_rem(a, stride) }; |
| 2279 | if a_mod_stride == 0 { |
| 2280 | // SPECIAL_CASE: In cases where the `a` is divisible by `stride`, byte offset to align a |
| 2281 | // pointer can be computed more simply through `-p (mod a)`. In the off-chance the byte |
| 2282 | // offset is not a multiple of `stride`, the input pointer was misaligned and no pointer |
| 2283 | // offset will be able to produce a `p` aligned to the specified `a`. |
| 2284 | // |
| 2285 | // The naive `-p (mod a)` equation inhibits LLVM's ability to select instructions |
| 2286 | // like `lea`. We compute `(round_up_to_next_alignment(p, a) - p)` instead. This |
| 2287 | // redistributes operations around the load-bearing, but pessimizing `and` instruction |
| 2288 | // sufficiently for LLVM to be able to utilize the various optimizations it knows about. |
| 2289 | // |
| 2290 | // LLVM handles the branch here particularly nicely. If this branch needs to be evaluated |
| 2291 | // at runtime, it will produce a mask `if addr_mod_stride == 0 { 0 } else { usize::MAX }` |
| 2292 | // in a branch-free way and then bitwise-OR it with whatever result the `-p mod a` |
| 2293 | // computation produces. |
| 2294 | |
| 2295 | let aligned_address = wrapping_add(addr, a_minus_one) & wrapping_sub(0, a); |
| 2296 | let byte_offset = wrapping_sub(aligned_address, addr); |
| 2297 | // FIXME: Remove the assume after <https://github.com/llvm/llvm-project/issues/62502> |
| 2298 | // SAFETY: Masking by `-a` can only affect the low bits, and thus cannot have reduced |
| 2299 | // the value by more than `a-1`, so even though the intermediate values might have |
| 2300 | // wrapped, the byte_offset is always in `[0, a)`. |
| 2301 | unsafe { assume(byte_offset < a) }; |
| 2302 | |
| 2303 | // SAFETY: `stride == 0` case has been handled by the special case above. |
| 2304 | let addr_mod_stride = unsafe { unchecked_rem(addr, stride) }; |
| 2305 | |
| 2306 | return if addr_mod_stride == 0 { |
| 2307 | // SAFETY: `stride` is non-zero. This is guaranteed to divide exactly as well, because |
| 2308 | // addr has been verified to be aligned to the original type’s alignment requirements. |
| 2309 | unsafe { exact_div(byte_offset, stride) } |
| 2310 | } else { |
| 2311 | usize::MAX |
| 2312 | }; |
| 2313 | } |
| 2314 | |
| 2315 | // GENERAL_CASE: From here on we’re handling the very general case where `addr` may be |
| 2316 | // misaligned, there isn’t an obvious relationship between `stride` and `a` that we can take an |
| 2317 | // advantage of, etc. This case produces machine code that isn’t particularly high quality, |
| 2318 | // compared to the special cases above. The code produced here is still within the realm of |
| 2319 | // miracles, given the situations this case has to deal with. |
| 2320 | |
| 2321 | // SAFETY: a is power-of-two hence non-zero. stride == 0 case is handled above. |
| 2322 | // FIXME(const-hack) replace with min |
| 2323 | let gcdpow = unsafe { |
| 2324 | let x = cttz_nonzero(stride); |
| 2325 | let y = cttz_nonzero(a); |
| 2326 | if x < y { x } else { y } |
| 2327 | }; |
| 2328 | // SAFETY: gcdpow has an upper-bound that’s at most the number of bits in a `usize`. |
| 2329 | let gcd = unsafe { unchecked_shl(1usize, gcdpow) }; |
| 2330 | // SAFETY: gcd is always greater or equal to 1. |
| 2331 | if addr & unsafe { unchecked_sub(gcd, 1) } == 0 { |
| 2332 | // This branch solves for the following linear congruence equation: |
| 2333 | // |
| 2334 | // ` p + so = 0 mod a ` |
| 2335 | // |
| 2336 | // `p` here is the pointer value, `s` - stride of `T`, `o` offset in `T`s, and `a` - the |
| 2337 | // requested alignment. |
| 2338 | // |
| 2339 | // With `g = gcd(a, s)`, and the above condition asserting that `p` is also divisible by |
| 2340 | // `g`, we can denote `a' = a/g`, `s' = s/g`, `p' = p/g`, then this becomes equivalent to: |
| 2341 | // |
| 2342 | // ` p' + s'o = 0 mod a' ` |
| 2343 | // ` o = (a' - (p' mod a')) * (s'^-1 mod a') ` |
| 2344 | // |
| 2345 | // The first term is "the relative alignment of `p` to `a`" (divided by the `g`), the |
| 2346 | // second term is "how does incrementing `p` by `s` bytes change the relative alignment of |
| 2347 | // `p`" (again divided by `g`). Division by `g` is necessary to make the inverse well |
| 2348 | // formed if `a` and `s` are not co-prime. |
| 2349 | // |
| 2350 | // Furthermore, the result produced by this solution is not "minimal", so it is necessary |
| 2351 | // to take the result `o mod lcm(s, a)`. This `lcm(s, a)` is the same as `a'`. |
| 2352 | |
| 2353 | // SAFETY: `gcdpow` has an upper-bound not greater than the number of trailing 0-bits in |
| 2354 | // `a`. |
| 2355 | let a2 = unsafe { unchecked_shr(a, gcdpow) }; |
| 2356 | // SAFETY: `a2` is non-zero. Shifting `a` by `gcdpow` cannot shift out any of the set bits |
| 2357 | // in `a` (of which it has exactly one). |
| 2358 | let a2minus1 = unsafe { unchecked_sub(a2, 1) }; |
| 2359 | // SAFETY: `gcdpow` has an upper-bound not greater than the number of trailing 0-bits in |
| 2360 | // `a`. |
| 2361 | let s2 = unsafe { unchecked_shr(stride & a_minus_one, gcdpow) }; |
| 2362 | // SAFETY: `gcdpow` has an upper-bound not greater than the number of trailing 0-bits in |
| 2363 | // `a`. Furthermore, the subtraction cannot overflow, because `a2 = a >> gcdpow` will |
| 2364 | // always be strictly greater than `(p % a) >> gcdpow`. |
| 2365 | let minusp2 = unsafe { unchecked_sub(a2, unchecked_shr(addr & a_minus_one, gcdpow)) }; |
| 2366 | // SAFETY: `a2` is a power-of-two, as proven above. `s2` is strictly less than `a2` |
| 2367 | // because `(s % a) >> gcdpow` is strictly less than `a >> gcdpow`. |
| 2368 | return wrapping_mul(minusp2, unsafe { mod_inv(s2, a2) }) & a2minus1; |
| 2369 | } |
| 2370 | |
| 2371 | // Cannot be aligned at all. |
| 2372 | usize::MAX |
| 2373 | } |
| 2374 | |
| 2375 | /// Compares raw pointers for equality. |
| 2376 | /// |
| 2377 | /// This is the same as using the `==` operator, but less generic: |
| 2378 | /// the arguments have to be `*const T` raw pointers, |
| 2379 | /// not anything that implements `PartialEq`. |
| 2380 | /// |
| 2381 | /// This can be used to compare `&T` references (which coerce to `*const T` implicitly) |
| 2382 | /// by their address rather than comparing the values they point to |
| 2383 | /// (which is what the `PartialEq for &T` implementation does). |
| 2384 | /// |
| 2385 | /// When comparing wide pointers, both the address and the metadata are tested for equality. |
| 2386 | /// However, note that comparing trait object pointers (`*const dyn Trait`) is unreliable: pointers |
| 2387 | /// to values of the same underlying type can compare inequal (because vtables are duplicated in |
| 2388 | /// multiple codegen units), and pointers to values of *different* underlying type can compare equal |
| 2389 | /// (since identical vtables can be deduplicated within a codegen unit). |
| 2390 | /// |
| 2391 | /// # Examples |
| 2392 | /// |
| 2393 | /// ``` |
| 2394 | /// use std::ptr; |
| 2395 | /// |
| 2396 | /// let five = 5; |
| 2397 | /// let other_five = 5; |
| 2398 | /// let five_ref = &five; |
| 2399 | /// let same_five_ref = &five; |
| 2400 | /// let other_five_ref = &other_five; |
| 2401 | /// |
| 2402 | /// assert!(five_ref == same_five_ref); |
| 2403 | /// assert!(ptr::eq(five_ref, same_five_ref)); |
| 2404 | /// |
| 2405 | /// assert!(five_ref == other_five_ref); |
| 2406 | /// assert!(!ptr::eq(five_ref, other_five_ref)); |
| 2407 | /// ``` |
| 2408 | /// |
| 2409 | /// Slices are also compared by their length (fat pointers): |
| 2410 | /// |
| 2411 | /// ``` |
| 2412 | /// let a = [1, 2, 3]; |
| 2413 | /// assert!(std::ptr::eq(&a[..3], &a[..3])); |
| 2414 | /// assert!(!std::ptr::eq(&a[..2], &a[..3])); |
| 2415 | /// assert!(!std::ptr::eq(&a[0..2], &a[1..3])); |
| 2416 | /// ``` |
| 2417 | #[stable (feature = "ptr_eq" , since = "1.17.0" )] |
| 2418 | #[inline (always)] |
| 2419 | #[must_use = "pointer comparison produces a value" ] |
| 2420 | #[rustc_diagnostic_item = "ptr_eq" ] |
| 2421 | #[allow (ambiguous_wide_pointer_comparisons)] // it's actually clear here |
| 2422 | pub fn eq<T: ?Sized>(a: *const T, b: *const T) -> bool { |
| 2423 | a == b |
| 2424 | } |
| 2425 | |
| 2426 | /// Compares the *addresses* of the two pointers for equality, |
| 2427 | /// ignoring any metadata in fat pointers. |
| 2428 | /// |
| 2429 | /// If the arguments are thin pointers of the same type, |
| 2430 | /// then this is the same as [`eq`]. |
| 2431 | /// |
| 2432 | /// # Examples |
| 2433 | /// |
| 2434 | /// ``` |
| 2435 | /// use std::ptr; |
| 2436 | /// |
| 2437 | /// let whole: &[i32; 3] = &[1, 2, 3]; |
| 2438 | /// let first: &i32 = &whole[0]; |
| 2439 | /// |
| 2440 | /// assert!(ptr::addr_eq(whole, first)); |
| 2441 | /// assert!(!ptr::eq::<dyn std::fmt::Debug>(whole, first)); |
| 2442 | /// ``` |
| 2443 | #[stable (feature = "ptr_addr_eq" , since = "1.76.0" )] |
| 2444 | #[inline (always)] |
| 2445 | #[must_use = "pointer comparison produces a value" ] |
| 2446 | pub fn addr_eq<T: ?Sized, U: ?Sized>(p: *const T, q: *const U) -> bool { |
| 2447 | (p as *const ()) == (q as *const ()) |
| 2448 | } |
| 2449 | |
| 2450 | /// Compares the *addresses* of the two function pointers for equality. |
| 2451 | /// |
| 2452 | /// This is the same as `f == g`, but using this function makes clear that the potentially |
| 2453 | /// surprising semantics of function pointer comparison are involved. |
| 2454 | /// |
| 2455 | /// There are **very few guarantees** about how functions are compiled and they have no intrinsic |
| 2456 | /// “identity”; in particular, this comparison: |
| 2457 | /// |
| 2458 | /// * May return `true` unexpectedly, in cases where functions are equivalent. |
| 2459 | /// |
| 2460 | /// For example, the following program is likely (but not guaranteed) to print `(true, true)` |
| 2461 | /// when compiled with optimization: |
| 2462 | /// |
| 2463 | /// ``` |
| 2464 | /// let f: fn(i32) -> i32 = |x| x; |
| 2465 | /// let g: fn(i32) -> i32 = |x| x + 0; // different closure, different body |
| 2466 | /// let h: fn(u32) -> u32 = |x| x + 0; // different signature too |
| 2467 | /// dbg!(std::ptr::fn_addr_eq(f, g), std::ptr::fn_addr_eq(f, h)); // not guaranteed to be equal |
| 2468 | /// ``` |
| 2469 | /// |
| 2470 | /// * May return `false` in any case. |
| 2471 | /// |
| 2472 | /// This is particularly likely with generic functions but may happen with any function. |
| 2473 | /// (From an implementation perspective, this is possible because functions may sometimes be |
| 2474 | /// processed more than once by the compiler, resulting in duplicate machine code.) |
| 2475 | /// |
| 2476 | /// Despite these false positives and false negatives, this comparison can still be useful. |
| 2477 | /// Specifically, if |
| 2478 | /// |
| 2479 | /// * `T` is the same type as `U`, `T` is a [subtype] of `U`, or `U` is a [subtype] of `T`, and |
| 2480 | /// * `ptr::fn_addr_eq(f, g)` returns true, |
| 2481 | /// |
| 2482 | /// then calling `f` and calling `g` will be equivalent. |
| 2483 | /// |
| 2484 | /// |
| 2485 | /// # Examples |
| 2486 | /// |
| 2487 | /// ``` |
| 2488 | /// use std::ptr; |
| 2489 | /// |
| 2490 | /// fn a() { println!("a" ); } |
| 2491 | /// fn b() { println!("b" ); } |
| 2492 | /// assert!(!ptr::fn_addr_eq(a as fn(), b as fn())); |
| 2493 | /// ``` |
| 2494 | /// |
| 2495 | /// [subtype]: https://doc.rust-lang.org/reference/subtyping.html |
| 2496 | #[stable (feature = "ptr_fn_addr_eq" , since = "1.85.0" )] |
| 2497 | #[inline (always)] |
| 2498 | #[must_use = "function pointer comparison produces a value" ] |
| 2499 | pub fn fn_addr_eq<T: FnPtr, U: FnPtr>(f: T, g: U) -> bool { |
| 2500 | f.addr() == g.addr() |
| 2501 | } |
| 2502 | |
| 2503 | /// Hash a raw pointer. |
| 2504 | /// |
| 2505 | /// This can be used to hash a `&T` reference (which coerces to `*const T` implicitly) |
| 2506 | /// by its address rather than the value it points to |
| 2507 | /// (which is what the `Hash for &T` implementation does). |
| 2508 | /// |
| 2509 | /// # Examples |
| 2510 | /// |
| 2511 | /// ``` |
| 2512 | /// use std::hash::{DefaultHasher, Hash, Hasher}; |
| 2513 | /// use std::ptr; |
| 2514 | /// |
| 2515 | /// let five = 5; |
| 2516 | /// let five_ref = &five; |
| 2517 | /// |
| 2518 | /// let mut hasher = DefaultHasher::new(); |
| 2519 | /// ptr::hash(five_ref, &mut hasher); |
| 2520 | /// let actual = hasher.finish(); |
| 2521 | /// |
| 2522 | /// let mut hasher = DefaultHasher::new(); |
| 2523 | /// (five_ref as *const i32).hash(&mut hasher); |
| 2524 | /// let expected = hasher.finish(); |
| 2525 | /// |
| 2526 | /// assert_eq!(actual, expected); |
| 2527 | /// ``` |
| 2528 | #[stable (feature = "ptr_hash" , since = "1.35.0" )] |
| 2529 | pub fn hash<T: ?Sized, S: hash::Hasher>(hashee: *const T, into: &mut S) { |
| 2530 | use crate::hash::Hash; |
| 2531 | hashee.hash(state:into); |
| 2532 | } |
| 2533 | |
| 2534 | #[stable (feature = "fnptr_impls" , since = "1.4.0" )] |
| 2535 | impl<F: FnPtr> PartialEq for F { |
| 2536 | #[inline ] |
| 2537 | fn eq(&self, other: &Self) -> bool { |
| 2538 | self.addr() == other.addr() |
| 2539 | } |
| 2540 | } |
| 2541 | #[stable (feature = "fnptr_impls" , since = "1.4.0" )] |
| 2542 | impl<F: FnPtr> Eq for F {} |
| 2543 | |
| 2544 | #[stable (feature = "fnptr_impls" , since = "1.4.0" )] |
| 2545 | impl<F: FnPtr> PartialOrd for F { |
| 2546 | #[inline ] |
| 2547 | fn partial_cmp(&self, other: &Self) -> Option<Ordering> { |
| 2548 | self.addr().partial_cmp(&other.addr()) |
| 2549 | } |
| 2550 | } |
| 2551 | #[stable (feature = "fnptr_impls" , since = "1.4.0" )] |
| 2552 | impl<F: FnPtr> Ord for F { |
| 2553 | #[inline ] |
| 2554 | fn cmp(&self, other: &Self) -> Ordering { |
| 2555 | self.addr().cmp(&other.addr()) |
| 2556 | } |
| 2557 | } |
| 2558 | |
| 2559 | #[stable (feature = "fnptr_impls" , since = "1.4.0" )] |
| 2560 | impl<F: FnPtr> hash::Hash for F { |
| 2561 | fn hash<HH: hash::Hasher>(&self, state: &mut HH) { |
| 2562 | state.write_usize(self.addr() as _) |
| 2563 | } |
| 2564 | } |
| 2565 | |
| 2566 | #[stable (feature = "fnptr_impls" , since = "1.4.0" )] |
| 2567 | impl<F: FnPtr> fmt::Pointer for F { |
| 2568 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
| 2569 | fmt::pointer_fmt_inner(self.addr() as _, f) |
| 2570 | } |
| 2571 | } |
| 2572 | |
| 2573 | #[stable (feature = "fnptr_impls" , since = "1.4.0" )] |
| 2574 | impl<F: FnPtr> fmt::Debug for F { |
| 2575 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
| 2576 | fmt::pointer_fmt_inner(self.addr() as _, f) |
| 2577 | } |
| 2578 | } |
| 2579 | |
| 2580 | /// Creates a `const` raw pointer to a place, without creating an intermediate reference. |
| 2581 | /// |
| 2582 | /// `addr_of!(expr)` is equivalent to `&raw const expr`. The macro is *soft-deprecated*; |
| 2583 | /// use `&raw const` instead. |
| 2584 | /// |
| 2585 | /// It is still an open question under which conditions writing through an `addr_of!`-created |
| 2586 | /// pointer is permitted. If the place `expr` evaluates to is based on a raw pointer, then the |
| 2587 | /// result of `addr_of!` inherits all permissions from that raw pointer. However, if the place is |
| 2588 | /// based on a reference, local variable, or `static`, then until all details are decided, the same |
| 2589 | /// rules as for shared references apply: it is UB to write through a pointer created with this |
| 2590 | /// operation, except for bytes located inside an `UnsafeCell`. Use `&raw mut` (or [`addr_of_mut`]) |
| 2591 | /// to create a raw pointer that definitely permits mutation. |
| 2592 | /// |
| 2593 | /// Creating a reference with `&`/`&mut` is only allowed if the pointer is properly aligned |
| 2594 | /// and points to initialized data. For cases where those requirements do not hold, |
| 2595 | /// raw pointers should be used instead. However, `&expr as *const _` creates a reference |
| 2596 | /// before casting it to a raw pointer, and that reference is subject to the same rules |
| 2597 | /// as all other references. This macro can create a raw pointer *without* creating |
| 2598 | /// a reference first. |
| 2599 | /// |
| 2600 | /// See [`addr_of_mut`] for how to create a pointer to uninitialized data. |
| 2601 | /// Doing that with `addr_of` would not make much sense since one could only |
| 2602 | /// read the data, and that would be Undefined Behavior. |
| 2603 | /// |
| 2604 | /// # Safety |
| 2605 | /// |
| 2606 | /// The `expr` in `addr_of!(expr)` is evaluated as a place expression, but never loads from the |
| 2607 | /// place or requires the place to be dereferenceable. This means that `addr_of!((*ptr).field)` |
| 2608 | /// still requires the projection to `field` to be in-bounds, using the same rules as [`offset`]. |
| 2609 | /// However, `addr_of!(*ptr)` is defined behavior even if `ptr` is null, dangling, or misaligned. |
| 2610 | /// |
| 2611 | /// Note that `Deref`/`Index` coercions (and their mutable counterparts) are applied inside |
| 2612 | /// `addr_of!` like everywhere else, in which case a reference is created to call `Deref::deref` or |
| 2613 | /// `Index::index`, respectively. The statements above only apply when no such coercions are |
| 2614 | /// applied. |
| 2615 | /// |
| 2616 | /// [`offset`]: pointer::offset |
| 2617 | /// |
| 2618 | /// # Example |
| 2619 | /// |
| 2620 | /// **Correct usage: Creating a pointer to unaligned data** |
| 2621 | /// |
| 2622 | /// ``` |
| 2623 | /// use std::ptr; |
| 2624 | /// |
| 2625 | /// #[repr(packed)] |
| 2626 | /// struct Packed { |
| 2627 | /// f1: u8, |
| 2628 | /// f2: u16, |
| 2629 | /// } |
| 2630 | /// |
| 2631 | /// let packed = Packed { f1: 1, f2: 2 }; |
| 2632 | /// // `&packed.f2` would create an unaligned reference, and thus be Undefined Behavior! |
| 2633 | /// let raw_f2 = ptr::addr_of!(packed.f2); |
| 2634 | /// assert_eq!(unsafe { raw_f2.read_unaligned() }, 2); |
| 2635 | /// ``` |
| 2636 | /// |
| 2637 | /// **Incorrect usage: Out-of-bounds fields projection** |
| 2638 | /// |
| 2639 | /// ```rust,no_run |
| 2640 | /// use std::ptr; |
| 2641 | /// |
| 2642 | /// #[repr(C)] |
| 2643 | /// struct MyStruct { |
| 2644 | /// field1: i32, |
| 2645 | /// field2: i32, |
| 2646 | /// } |
| 2647 | /// |
| 2648 | /// let ptr: *const MyStruct = ptr::null(); |
| 2649 | /// let fieldptr = unsafe { ptr::addr_of!((*ptr).field2) }; // Undefined Behavior ⚠️ |
| 2650 | /// ``` |
| 2651 | /// |
| 2652 | /// The field projection `.field2` would offset the pointer by 4 bytes, |
| 2653 | /// but the pointer is not in-bounds of an allocation for 4 bytes, |
| 2654 | /// so this offset is Undefined Behavior. |
| 2655 | /// See the [`offset`] docs for a full list of requirements for inbounds pointer arithmetic; the |
| 2656 | /// same requirements apply to field projections, even inside `addr_of!`. (In particular, it makes |
| 2657 | /// no difference whether the pointer is null or dangling.) |
| 2658 | #[stable (feature = "raw_ref_macros" , since = "1.51.0" )] |
| 2659 | #[rustc_macro_transparency = "semitransparent" ] |
| 2660 | pub macro addr_of($place:expr) { |
| 2661 | &raw const $place |
| 2662 | } |
| 2663 | |
| 2664 | /// Creates a `mut` raw pointer to a place, without creating an intermediate reference. |
| 2665 | /// |
| 2666 | /// `addr_of_mut!(expr)` is equivalent to `&raw mut expr`. The macro is *soft-deprecated*; |
| 2667 | /// use `&raw mut` instead. |
| 2668 | /// |
| 2669 | /// Creating a reference with `&`/`&mut` is only allowed if the pointer is properly aligned |
| 2670 | /// and points to initialized data. For cases where those requirements do not hold, |
| 2671 | /// raw pointers should be used instead. However, `&mut expr as *mut _` creates a reference |
| 2672 | /// before casting it to a raw pointer, and that reference is subject to the same rules |
| 2673 | /// as all other references. This macro can create a raw pointer *without* creating |
| 2674 | /// a reference first. |
| 2675 | /// |
| 2676 | /// # Safety |
| 2677 | /// |
| 2678 | /// The `expr` in `addr_of_mut!(expr)` is evaluated as a place expression, but never loads from the |
| 2679 | /// place or requires the place to be dereferenceable. This means that `addr_of_mut!((*ptr).field)` |
| 2680 | /// still requires the projection to `field` to be in-bounds, using the same rules as [`offset`]. |
| 2681 | /// However, `addr_of_mut!(*ptr)` is defined behavior even if `ptr` is null, dangling, or misaligned. |
| 2682 | /// |
| 2683 | /// Note that `Deref`/`Index` coercions (and their mutable counterparts) are applied inside |
| 2684 | /// `addr_of_mut!` like everywhere else, in which case a reference is created to call `Deref::deref` |
| 2685 | /// or `Index::index`, respectively. The statements above only apply when no such coercions are |
| 2686 | /// applied. |
| 2687 | /// |
| 2688 | /// [`offset`]: pointer::offset |
| 2689 | /// |
| 2690 | /// # Examples |
| 2691 | /// |
| 2692 | /// **Correct usage: Creating a pointer to unaligned data** |
| 2693 | /// |
| 2694 | /// ``` |
| 2695 | /// use std::ptr; |
| 2696 | /// |
| 2697 | /// #[repr(packed)] |
| 2698 | /// struct Packed { |
| 2699 | /// f1: u8, |
| 2700 | /// f2: u16, |
| 2701 | /// } |
| 2702 | /// |
| 2703 | /// let mut packed = Packed { f1: 1, f2: 2 }; |
| 2704 | /// // `&mut packed.f2` would create an unaligned reference, and thus be Undefined Behavior! |
| 2705 | /// let raw_f2 = ptr::addr_of_mut!(packed.f2); |
| 2706 | /// unsafe { raw_f2.write_unaligned(42); } |
| 2707 | /// assert_eq!({packed.f2}, 42); // `{...}` forces copying the field instead of creating a reference. |
| 2708 | /// ``` |
| 2709 | /// |
| 2710 | /// **Correct usage: Creating a pointer to uninitialized data** |
| 2711 | /// |
| 2712 | /// ```rust |
| 2713 | /// use std::{ptr, mem::MaybeUninit}; |
| 2714 | /// |
| 2715 | /// struct Demo { |
| 2716 | /// field: bool, |
| 2717 | /// } |
| 2718 | /// |
| 2719 | /// let mut uninit = MaybeUninit::<Demo>::uninit(); |
| 2720 | /// // `&uninit.as_mut().field` would create a reference to an uninitialized `bool`, |
| 2721 | /// // and thus be Undefined Behavior! |
| 2722 | /// let f1_ptr = unsafe { ptr::addr_of_mut!((*uninit.as_mut_ptr()).field) }; |
| 2723 | /// unsafe { f1_ptr.write(true); } |
| 2724 | /// let init = unsafe { uninit.assume_init() }; |
| 2725 | /// ``` |
| 2726 | /// |
| 2727 | /// **Incorrect usage: Out-of-bounds fields projection** |
| 2728 | /// |
| 2729 | /// ```rust,no_run |
| 2730 | /// use std::ptr; |
| 2731 | /// |
| 2732 | /// #[repr(C)] |
| 2733 | /// struct MyStruct { |
| 2734 | /// field1: i32, |
| 2735 | /// field2: i32, |
| 2736 | /// } |
| 2737 | /// |
| 2738 | /// let ptr: *mut MyStruct = ptr::null_mut(); |
| 2739 | /// let fieldptr = unsafe { ptr::addr_of_mut!((*ptr).field2) }; // Undefined Behavior ⚠️ |
| 2740 | /// ``` |
| 2741 | /// |
| 2742 | /// The field projection `.field2` would offset the pointer by 4 bytes, |
| 2743 | /// but the pointer is not in-bounds of an allocation for 4 bytes, |
| 2744 | /// so this offset is Undefined Behavior. |
| 2745 | /// See the [`offset`] docs for a full list of requirements for inbounds pointer arithmetic; the |
| 2746 | /// same requirements apply to field projections, even inside `addr_of_mut!`. (In particular, it |
| 2747 | /// makes no difference whether the pointer is null or dangling.) |
| 2748 | #[stable (feature = "raw_ref_macros" , since = "1.51.0" )] |
| 2749 | #[rustc_macro_transparency = "semitransparent" ] |
| 2750 | pub macro addr_of_mut($place:expr) { |
| 2751 | &raw mut $place |
| 2752 | } |
| 2753 | |