[clang] [BoundsSafety] Initial documentation for -fbounds-safety (PR #70749)

[Aaron Ballman via cfe-commits]

================
@@ -0,0 +1,362 @@
+==================================================
+``-fbounds-safety``: Enforcing bounds safety for C
+==================================================
+
+.. contents::
+   :local:
+
+Overview
+========
+
+``-fbounds-safety`` is a C extension to enforce bounds safety to prevent out-of-bounds (OOB) memory accesses, which remain a major source of security vulnerabilities in C. ``-fbounds-safety`` aims to eliminate this class of bugs by turning OOB accesses into deterministic traps.
+
+The ``-fbounds-safety`` extension offers bounds annotations that programmers can use to attach bounds to pointers. For example, programmers can add the ``__counted_by(N)`` annotation to parameter ``ptr``, indicating that the pointer has ``N`` valid elements:
+
+.. code-block:: c
+
+   void foo(int *__counted_by(N) ptr, size_t N);
+
+Using this bounds information, the compiler inserts bounds checks on every pointer dereference, ensuring that the program does not access memory outside the specified bounds. The compiler requires programmers to provide enough bounds information so that the accesses can be checked at either run time or compile time — and it rejects code if it cannot.
+
+The most important contribution of ``-fbounds-safety`` is how it reduces the programmer’s annotation burden by reconciling bounds annotations at ABI boundaries with the use of implicit wide pointers (a.k.a. “fat” pointers) that carry bounds information on local variables without the need for annotations. We designed this model so that it preserves ABI compatibility with C while minimizing adoption effort.
+
+The ``-fbounds-safety`` extension has been adopted on millions of lines of production C code and proven to work in a consumer operating system setting. The extension was designed to enable incremental adoption — a key requirement in real-world settings where modifying an entire project and its dependencies all at once is often not possible. It also addresses multiple of other practical challenges that have made existing approaches to safer C dialects difficult to adopt, offering these properties that make it widely adoptable in practice:
+
+* It is designed to preserve the Application Binary Interface (ABI).
+* It interoperates well with plain C code.
+* It can be adopted partially and incrementally while still providing safety benefits.
+* It is syntactically and semantically compatible with C.
+* Consequently, source code that adopts the extension can continue to be compiled by toolchains that do not support the extension.
+* It has a relatively low adoption cost.
+* It can be implemented on top of Clang.
+
+This document discusses the key designs of ``-fbounds-safety``. The document is subject to be actively updated with a more detailed specification. The implementation plan can be found in `Implementation plans for -fbounds-safety <BoundsSafetyImplPlans.rst>`_.
+
+Programming Model
+=================
+
+Overview
+--------
+
+``-fbounds-safety`` ensures that pointers are not used to access memory beyond their bounds by performing bounds checking. If a bounds check fails, the program will deterministically trap before out-of-bounds memory is accessed.
+
+In our model, every pointer has an explicit or implicit bounds attribute that determines its bounds and ensures guaranteed bounds checking. Consider the example below where the ``__counted_by(count)`` annotation indicates that parameter ``p`` points to a buffer of integers containing ``count`` elements. An off-by-one error is present in the loop condition, leading to ``p[i]`` being out-of-bounds access during the loop’s final iteration. The compiler inserts a bounds check before ``p`` is dereferenced to ensure that the access remains within the specified bounds.
+
+.. code-block:: c
+
+   void fill_array_with_indices(int *__counted_by(count) p, unsigned count) {
+   // off-by-one error (i < count)
+      for (unsigned i = 0; i <= count; ++i) {
+         // bounds check inserted:
+         //   if (i >= count) trap();
+         p[i] = i;
+      }
+   }
+
+A bounds annotation defines an invariant for the pointer type, and the model ensures that this invariant remains true. In the example below, pointer ``p`` annotated with ``__counted_by(count)`` must always point to a memory buffer containing at least ``count`` elements of the pointee type. Increasing the value of ``count``, like in the example below, would violate this invariant and permit out-of-bounds access to the pointer. To avoid this, the compiler emits either a compile-time error or a run-time trap. Section `Maintaining correctness of bounds annotations`_ provides more details about the programming model.
+
+.. code-block:: c
+
+   void foo(int *__counted_by(count) p, size_t count) {
+      count++; // violates the invariant of __counted_by
+   }
+
+The requirement to annotate all pointers with explicit bounds information could present a significant adoption burden. To tackle this issue, the model incorporates the concept of a “wide pointer” (a.k.a. fat pointer) – a larger pointer that carries bounds information alongside the pointer value. Utilizing wide pointers can potentially reduce the adoption burden, as it contains bounds information internally and eliminates the need for explicit bounds annotations. However, wide pointers differ from standard C pointers in their data layout, which may result in incompatibilities with the application binary interface (ABI). Breaking the ABI complicates interoperability with external code that has not adopted the same programming model.
+
+``-fbounds-safety`` harmonizes the wide pointer and the bounds annotation approaches to reduce the adoption burden while maintaining the ABI. In this model, local variables of pointer type are implicitly treated as wide pointers, allowing them to carry bounds information without requiring explicit bounds annotations. This approach does not impact the ABI, as local variables are hidden from the ABI. Pointers associated with any other variables are treated as single object pointers (i.e., ``__single``), ensuring that they always have the tightest bounds by default and offering a strong bounds safety guarantee.
+
+By implementing default bounds annotations based on ABI visibility, a considerable portion of C code can operate without modifications within this programming model, reducing the adoption burden.
+
+The rest of the section will discuss individual bounds annotations and the programming model in more detail.
+
+Bounds annotations
+------------------
+
+Annotation for pointers to a single object
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+The C language allows pointer arithmetic on arbitrary pointers and this has been a source of many bounds safety issues. In practice, many pointers are merely pointing to a single object and incrementing or decrementing such a pointer immediately makes the pointer go out-of-bounds. To prevent this unsafety, ``-fbounds-safety`` provides the annotation ``__single`` that causes pointer arithmetic on annotated pointers to be a compile time error.
+
+* ``__single`` : indicates that the pointer is either pointing to a single object or null. Hence, pointers with ``__single`` do not permit pointer arithmetic nor being subscripted with a non-zero index. Dereferencing a ``__single`` pointer is allowed but it requires a null check. Upper and lower bounds checks are not required because the ``__single`` pointer should point to a valid object unless it’s null.
+
+We use ``__single`` as the default annotation for ABI-visible pointers. This gives strong security guarantees in that these pointers cannot be incremented or decremented unless they have an explicit, overriding bounds annotation that can be used to verify the safety of the operation. The compiler issues an error when a ``__single`` pointer is utilized for pointer arithmetic or array access, as these operations would immediately cause the pointer to exceed its bounds. Consequently, this prompts programmers to provide sufficient bounds information to pointers. In the following example, the pointer on parameter p is single-by-default, and is employed for array access. As a result, the compiler generates an error suggesting to add ``__counted_by`` to the pointer.
+
+.. code-block:: c
+
+   void fill_array_with_indices(int *p, unsigned count) {
+      for (unsigned i = 0; i < count; ++i) {
+         p[i] = i; // error
+      }
+   }
+
+
+External bounds annotations
+^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+“External” bounds annotations provide a way to express a relationship between a pointer variable and another variable (or expression) containing the bounds information of the pointer. In the following example, ``__counted_by(count)`` annotation expresses the bounds of parameter p using another parameter count. This model works naturally with many C interfaces and structs because the bounds of a pointer is often available adjacent to the pointer itself, e.g., at another parameter of the same function prototype, or at another field of the same struct declaration.
+
+.. code-block:: c
+
+   void fill_array_with_indices(int *__counted_by(count) p, size_t count) {
+      // off-by-one error
+      for (size_t i = 0; i <= count; ++i)
+         p[i] = i;
+   }
+
+External bounds annotations include ``__counted_by``, ``__sized_by``, and ``__ended_by``. These annotations do not change the pointer representation, meaning they do not have ABI implications.
+
+* ``__counted_by(N)`` : The pointer points to memory that contains ``N`` elements of pointee type. ``N`` is an expression of integer type which can be a simple reference to declaration, a constant including calls to constant functions, or an arithmetic expression that does not have side effect. The annotation cannot apply to pointers to incomplete types or types without size such as ``void *``.
+* ``__sized_by(N)`` : The pointer points to memory that contains ``N`` bytes. Just like the argument of ``__counted_by``, ``N`` is an expression of integer type which can be a constant, a simple reference to a declaration, or an arithmetic expression that does not have side effects. This is mainly used for pointers to incomplete types or types without size such as ``void *``.
+* ``__ended_by(P)`` : The pointer has the upper bound of value ``P``, which is one past the last element of the pointer. In other words, this annotation describes a range that starts with the pointer that has this annotation and ends with ``P`` which is the argument of the annotation. ``P`` itself may be annotated with ``__ended_by(Q)``. In this case, the end of the range extends to the pointer ``Q``.
+
+Accessing a pointer outside the specified bounds causes a run-time trap or a compile-time error. Also, the model maintains correctness of bounds annotations when the pointer and/or the related value containing the bounds information are updated or passed as arguments. This is done by compile-time restrictions or run-time checks (see Section `Maintaining correctness of bounds annotations`_ for more detail). For instance, initializing ``buf`` with ``null`` while assigning non-zero value to ``count``, as shown in the following example, would violate the ``__counted_by`` annotation because a null pointer does not point to any valid memory location. To avoid this, the compiler produces either a compile-time error or run-time trap.
+
+.. code-block:: c
+
+   void null_with_count_10(int *__counted_by(count) buf, unsigned count) {
+   buf = 0;
+   count = 10; // This is not allowed as it creates a null pointer with non-zero length
+   }
+
+However, there are use cases where a pointer is either a null pointer or is pointing to memory of the specified size. To support this idiom, ``-fbounds-safety`` provides ``*_or_null`` variants, ``__counted_by_or_null(N)``, ``__sized_by_or_null(N)``, and ``__ended_by_or_null(P)``. Accessing a pointer with any of these bounds annotations will require an extra null check to avoid a null pointer dereference.
+
+Internal bounds annotations
+^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+A wide pointer (sometimes known as a “fat” pointer) is a pointer that carries additional bounds information internally (as part of its data). The bounds require additional storage space making wide pointers larger than normal pointers, hence the name “wide pointer”. The memory layout of a wide pointer is equivalent to a struct with the pointer, upper bound, and (optionally) lower bound as its fields as shown below.
+
+.. code-block:: c
+
+   struct wide_pointer_datalayout {
+      void* pointer; // Address used for dereferences and pointer arithmetic
+      void* upper_bound; // Points one past the highest address that can be accessed
+      void* lower_bound; // (Optional) Points to lowest address that can be accessed
+   };
+
+Even with this representational change, wide pointers act syntactically as normal pointers to allow standard pointer operations, such as pointer dereference (``*p``), array subscript (``p[i]``), member access (``p->``), and pointer arithmetic, with some restrictions on bounds-unsafe uses.
+
+``-fbounds-safety`` has a set of “internal” bounds annotations to turn pointers into wide pointers. These are ``__bidi_indexable`` and ``__indexable``. When a pointer has either of these annotations, the compiler changes the pointer to the corresponding wide pointer. This means these annotations will break the ABI and will not be compatible with plain C, and thus they should generally not be used in ABI surfaces.
+
+* ``__bidi_indexable`` : A pointer with this annotation becomes a wide pointer to carry the upper bound and the lower bound, the layout of which is equivalent to ``struct { T *ptr; T *upper_bound; T *lower_bound; };``. As the name indicates, pointers with this annotation are “bidirectionally indexable”, meaning that they can be indexed with either a negative or a positive offset and the pointers can be incremented or decremented using pointer arithmetic. A ``__bidi_indexable`` pointer is allowed to hold an out-of-bounds pointer value. While creating an OOB pointer is undefined behavior in C, ``-fbounds-safety`` makes it well-defined behavior. That is, pointer arithmetic overflow with ``__bidi_indexable`` is defined as equivalent of two’s complement integer computation, and at the LLVM IR level this means ``getelementptr`` won’t get ``inbounds`` keyword. Accessing memory using the OOB pointer is prevented via a run-time bounds check.
+* ``__indexable`` : A pointer with this annotation becomes a wide pointer carrying the upper bound (but no explicit lower bound), the layout of which is equivalent to ``struct { T *ptr; T *upper_bound; };``. Since ``__indexable`` pointers do not have a separate lower bound, the pointer value itself acts as the lower bound. An ``__indexable`` pointer can only be incremented or indexed in the positive direction. Decrementing it with a known negative index triggers a compile-time error. Otherwise, the compiler inserts a run-time check to ensure pointer arithmetic doesn’t make the pointer smaller than the original ``__indexable`` pointer (Note that ``__indexable`` doesn’t have a lower bound so the pointer value is effectively the lower bound). As pointer arithmetic overflow will make the pointer smaller than the original pointer, it will cause a trap at runtime. Similar to ``__bidi_indexable``, an ``__indexable`` pointer is allowed to have a pointer value above the upper bound and creating such a pointer is well-defined behavior. Dereferencing such a pointer, however, will cause a run-time trap.
+* ``__bidi_indexable`` offers the best flexibility out of all the pointer annotations in this model, as ``__bidi_indexable`` pointers can be used for any pointer operation. However, this comes with the largest code size and memory cost out of the available pointer annotations in this model. In some cases, use of the ``__bidi_indexable`` annotation may be duplicating bounds information that exists elsewhere in the program. In such cases, using external bounds annotations may be a better choice.
+
+``__bidi_indexable`` is the default annotation for non-ABI visible pointers, such as local pointer variables — that is, if the programmer does not specify another bounds annotation, a local pointer variable is implicitly ``__bidi_indexable``. Since ``__bidi_indexable`` pointers automatically carry bounds information and have no restrictions on kinds of pointer operations that can be used with these pointers, most code inside a function works as is without modification. In the example below, ``int *buf`` doesn’t require manual annotation as it’s implicitly ``int *__bidi_indexable buf``, carrying the bounds information passed from the return value of malloc, which is necessary to insert bounds checking for ``buf[i]``.
+
+.. code-block:: c
+
+   void *__sized_by(size) malloc(size_t size);
+      int *__counted_by(n) get_array_with_0_to_n_1(size_t n) {
+      int *buf = malloc(sizeof(int) * n);
+         for (size_t i = 0; i < n; ++i)
+            buf[i] = i;
+      return buf;
+   }
+
+Annotations for sentinel-delimited arrays
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+A C string is an array of characters. The null terminator — the first null character (‘\0’) element in the array — marks the end of the string. ``-fbounds-safety`` provides ``__null_terminated`` to annotate C strings and the generalized form ``__terminated_by(T)`` to annotate pointers and arrays with an end marked by a sentinel value. The model prevents dereferencing a ``__terminated_by`` pointer beyond its end. Calculating the location of the end (i.e., the address of the sentinel value), requires reading the entire array in memory and would have some performance costs. To avoid an unintended performance hit, the model puts some restrictions on how these pointers can be used. ``__terminated_by`` pointers cannot be indexed and can only be incremented by one at a time. To allow these operations, the pointers must be explicitly converted to ``__indexable`` pointers using the intrinsic function ``__unsafe_terminated_by_to_indexable(P, T)`` (or ``__unsafe_null_terminated_to_indexable(P)``) which converts the ``__terminated_by`` pointer ``P`` to an ``__indexable`` pointer.
+
+* ``__null_terminated`` : The pointer or array is terminated by NULL or 0. Modifying the terminator or incrementing the pointer beyond it is prevented at run time.
+* ``__terminated_by(T)`` : The pointer or array is terminated by ``T`` which is a constant expression. Accessing or incrementing the pointer beyond the terminator is not allowed. This is a generalization of ``__null_terminated`` which is defined as ``__terminated_by(0)``.
+
+Annotation for interoperating with bounds-unsafe code
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+* ``__unsafe_indexable`` : A pointer with this annotation behaves the same as a plain C pointer. That is, the pointer does not have any bounds information and pointer operations are not checked.
+* ``__unsafe_indexable`` can be used to mark pointers from system headers or pointers from code that has not adopted -fbounds safety. This enables interoperation between code using ``-fbounds-safety`` and code that does not.
+
+Default pointer types
+---------------------
+
+ABI visibility and default annotations
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Requiring ``-fbounds-safety`` adopters to add bounds annotations to all pointers in the codebase would be a significant adoption burden. To avoid this and to secure all pointers by default, ``-fbounds-safety`` applies default bounds annotations to pointer types.
+
+Default annotations apply to pointer types of declarations
+
+``-fbounds-safety`` applies default bounds annotations to pointer types used in declarations. The default annotations are determined by the ABI visibility of the pointer. A pointer type is ABI-visible if changing its size or representation affects the ABI. For instance, changing the size of a type used in a function parameter will affect the ABI and thus pointers used in function parameters are ABI-visible pointers. On the other hand, changing the types of local variables won’t have such ABI implications. Hence, ``-fbounds-safety`` considers the outermost pointer types of local variables as non-ABI visible. The rest of the pointers such as nested pointer types, pointer types of global variables, struct fields, and function prototypes are considered ABI-visible.
+
+All ABI-visible pointers are treated as ``__single`` by default unless annotated otherwise. This default both preserves ABI and makes these pointers safe by default. This behavior can be controlled with pragma to set the default annotation for ABI-visible pointers to be either ``__single``, ``__bidi_indexable``, ``__indexable``, or ``__unsafe_indexable``. For instance, ``__ptrcheck_abi_assume_unsafe_indexable()`` will make all ABI-visible pointers be ``__unsafe_indexable``.
+Non-ABI visible pointers — the outermost pointer types of local variables — are ``__bidi_indexable`` by default, so that these pointers have the bounds information necessary to perform bounds checks without the need for a manual annotation.
+All ``const char`` pointers are ``__null_terminated`` by default.
+In system headers, the default pointer attribute for ABI-visible pointers is set to ``__unsafe_indexable`` by default.
+
+ABI implications of default bounds annotations
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Although modifying types of a local variable doesn’t impact the ABI, taking the address of such a modified type could create a pointer type that has an ABI mismatch. Looking at the following example, ``int *local`` is implicitly ``int *__bidi_indexable`` and thus the type of ``&local`` is a pointer to ``int *__bidi_indexable``. On the other hand, in ``void foo(int **)``, the parameter type is a pointer to ``int *__single`` (i.e., ``void foo(int *__single *__single)``) (or a pointer to ``int *__unsafe_indexable`` if it’s from a system header). The compiler reports an error for casts between pointers whose elements have incompatible pointer attributes. This way, ``-fbounds-safety`` prevents pointers that are implicitly ``__bidi_indexable`` from silently escaping thereby breaking the ABI.
+
+.. code-block:: c
+
+   void foo(int **);
+
+   void bar(void) {
+   int *local = 0;
+   foo(&local); // error: passing 'int *__bidi_indexable*__bidi_indexable' to parameter of incompatible nested pointer type 'int *__single*__single'
+   }
+
+Default pointer types in ``sizeof()``
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+A pointer type in ``sizeof()`` does not have an implicit bounds annotation. When a bounds attribute is not specified, the evaluated pointer type is treated identically to a plain C pointer type. Therefore, ``sizeof(int*)`` remains the same with or without ``-fbounds-safety``. That said, programmers can explicitly add attribute to the types, e.g., ``sizeof(int *__bidi_indexable)``, in which case the sizeof evaluates to the size of type ``int *__bidi_indexable`` (the value equivalent to ``3 * sizeof(int*)``).
----------------
AaronBallman wrote:

So this means we can *never* change the internal representation of a wide pointer... is that acceptable? (e.g., what if we decide we can improve memory pressure by storing (ptr, small offset) for `__indexable`` instead of (ptr, ptr) due to local analysis determining we can save the space?

Also, we should explain how this impacts point alignment and `alignof` behavior.

https://github.com/llvm/llvm-project/pull/70749