@@ -2492,26 +2492,77 @@ are listed below.
24922492 Strict Aliasing
24932493---------------
24942494
2495- Clang by default applies C/C++'s strict aliasing rules during optimizations. In
2496- cases C and C++ rules diverge, the more conservative rules are used. Clang does
2497- not make use of strict aliasing rules in all cases yet, including unions and
2498- variable-sized arrays. That may change in the future.
2499-
2500- Internally Clang encodes the strict aliasing rules in LLVM IR using type-based
2501- alias analysis (TBAA) metadata.
2502-
2503- Note that clang-cl disables strict aliasing by default, see
2504- :ref: `Strict aliasing in clang-cl. <clang_cl_strict_aliasing >`
2505-
2506- As of Clang 20, strict aliasing rules are also applied to nested pointers. The
2507- new behavior can be disabled using ``-fno-pointer-tbaa ``. Note that Clang does
2508- not apply strict aliasing rules to `void* ` pointers to avoid breaking existing
2509- code, even though this is not required by the standard.
2510-
2511- Strict aliasing violations in the source may change program behavior and
2512- ``-fno-strict-aliasing `` disables use of the strict aliasing rules. There also
2513- is an experimental :ref: `TypeSanitizer <TypeSanitizer >` to detect strict
2514- aliasing violations.
2495+ The C and C++ standards require accesses to objects in memory to use l-values of
2496+ an appropriate type for the object. This is called *strict aliasing * or
2497+ *type-based alias analysis *. Strict aliasing enhances a variety of powerful
2498+ memory optimizations, including reordering, combining, and eliminating memory
2499+ accesses. These optimizations can lead to unexpected behavior in code that
2500+ violates the strict aliasing rules. For example:
2501+
2502+ .. code-block :: c++
2503+
2504+ void advance(size_t *index, double *data) {
2505+ double value = data[*index];
2506+ / * Clang may assume that this store does not change the contents of `data `. */
2507+ *index += 1;
2508+ / * Clang may assume that this store does not change the contents of `index `. */
2509+ data[*index] = value;
2510+ / * Either of these facts may create significant optimization opportunities
2511+ if Clang is able to inline this function. */
2512+ }
2513+
2514+ Strict aliasing can be explicitly enabled with ``-fstrict-aliasing `` and
2515+ disabled with ``-fno-strict-aliasing ``. ``clang-cl `` defaults to
2516+ ``-fno-strict-aliasing ``; see :ref: `Strict aliasing in clang-cl.
2517+ <clang_cl_strict_aliasing>`. Otherwise, Clang defaults to ``-fstrict-aliasing ``.
2518+
2519+ C and C++ specify slightly different rules for strict aliasing. To improve
2520+ language interoperability, Clang allows two types to alias if either language
2521+ would permit it. This includes applying the C++ similar types rule to C,
2522+ allowing ``int ** `` to alias ``int const * const * ``. Clang also relaxes the
2523+ standard aliasing rules in the following ways:
2524+
2525+ * All integer types of the same size are permitted to alias each other,
2526+ including signed and unsigned types.
2527+ * ``void* `` is permitted to alias any pointer type, ``void** `` is permitted to
2528+ alias any pointer to pointer type, and so on.
2529+
2530+ Code which violates strict aliasing has undefined behavior. A program that
2531+ works in one version of Clang may not work in another because of changes to the
2532+ optimizer. Clang provides a `:ref:TypeSanitizer <TypeSanitizer> ` to help detect
2533+ violations of the strict aliasing rules, but it is currently still experimental.
2534+ Code that is known to violate strict aliasing should generally be built with
2535+ ``-fno-strict-aliasing `` if the violation cannot be fixed.
2536+
2537+ Clang supports several ways to fix a violation of strict aliasing:
2538+
2539+ * L-values of the character types ``char `` and ``unsigned char `` (as well as
2540+ other types, depending on the standard) are permitted to access objects of
2541+ any type.
2542+
2543+ * Library functions such as ``memcpy `` and ``memset `` are specified as treating
2544+ memory as characters and therefore are not limited by strict aliasing. If a
2545+ value of one type must be reinterpreted as another (e.g. to read the bits of a
2546+ floating-point number), use ``memcpy `` to copy the representation to an object
2547+ of the destination type. This has no overhead over a direct l-value access
2548+ because Clang should reliably optimize calls to these functions to use simple
2549+ loads and stores when they are used with small constant sizes.
2550+
2551+ * The attribute ``may_alias `` can be added to a ``typedef `` to give l-values of
2552+ that type the same aliasing power as the character types.
2553+
2554+ Clang makes a best effort to avoid obvious miscompilations from strict aliasing
2555+ by only considering type information when it cannot prove that two accesses must
2556+ refer to the same memory. However, it is not recommended that programmers
2557+ intentionally rely on this instead of using one of the solutions above because
2558+ it is too easy for the compiler's analysis to be blocked in surprising ways.
2559+
2560+ In Clang 20, Clang strengthened its implementation of strict aliasing for
2561+ accesses of pointer type. Previously, all accesses of pointer type were
2562+ permitted to alias each other, but Clang now distinguishes different pointers
2563+ by their pointee type, except as limited by the relaxations around qualifiers
2564+ and `void* ` described above. The previous behavior of treating all pointers as
2565+ aliasing can be restored using ``-fno-pointer-tbaa ``.
25152566
25162567Profile Guided Optimization
25172568---------------------------
0 commit comments