Major New Features¶

• clang used to run the actual compilation in a subprocess (“clang -cc1”). Now compilations are done in-process by default. -fno-integrated-cc1 restores the former behavior. The -v and -### flags will print “(in-process)” when compilations are done in-process.
• Concepts support. Clang now supports C++2a Concepts under the -std=c++2a flag.

Non-comprehensive list of changes in this release¶

• In both C and C++ (C17 6.5.6p8, C++ [expr.add]), pointer arithmetic is only permitted within arrays. In particular, the behavior of a program is not defined if it adds a non-zero offset (or in C, any offset) to a null pointer, or if it forms a null pointer by subtracting an integer from a non-null pointer, and the LLVM optimizer now uses those guarantees for transformations. This may lead to unintended behavior in code that performs these operations. The Undefined Behavior Sanitizer -fsanitize=pointer-overflow check has been extended to detect these cases, so that code relying on them can be detected and fixed.
• The Implicit Conversion Sanitizer (-fsanitize=implicit-conversion) has learned to sanitize pre/post increment/decrement of types with bit width smaller than int.

• For X86 target, -march=skylake-avx512, -march=icelake-client, -march=icelake-server, -march=cascadelake, -march=cooperlake will default to not using 512-bit zmm registers in vectorized code unless 512-bit intrinsics are used in the source code. 512-bit operations are known to cause the CPUs to run at a lower frequency which can impact performance. This behavior can be changed by passing -mprefer-vector-width=512 on the command line.

• clang now defaults to .init_array on Linux. It used to use .ctors if the found gcc installation is older than 4.7.0. Add -fno-use-init-array to get the old behavior (.ctors).
• The behavior of the flag -flax-vector-conversions has been modified to more closely match GCC, as described below. In Clang 10 onwards, command lines specifying this flag do not permit implicit vector bitcasts between integer vectors and floating-point vectors. Such conversions are still permitted by default, however, and the default can be explicitly requested with the Clang-specific flag -flax-vector-conversions=all. In a future release of Clang, we intend to change the default to -fno-lax-vector-conversions.
• Improved support for octeon MIPS-family CPU. Added octeon+ to the list of of CPUs accepted by the driver.

New Compiler Flags¶

• The -fgnuc-version= flag now controls the value of __GNUC__ and related macros. This flag does not enable or disable any GCC extensions implemented in Clang. Setting the version to zero causes Clang to leave __GNUC__ and other GNU-namespaced macros, such as __GXX_WEAK__, undefined.
• vzeroupper insertion on X86 targets can now be disabled with -mno-vzeroupper. You can also force vzeroupper insertion to be used on CPUs that normally wouldn’t with -mvzeroupper.
• The -fno-concept-satisfaction-caching can be used to disable caching for satisfactions of Concepts. The C++2a draft standard does not currently permit this caching, but disabling it may incur significant compile-time costs. This flag is intended for experimentation purposes and may be removed at any time; please let us know if you encounter a situation where you need to specify this flag for correct program behavior.

Deprecated Compiler Flags¶

The following options are deprecated and ignored. They will be removed in future versions of Clang.

• -mmpx used to enable the __MPX__ preprocessor define for the Intel MPX instructions. There were no MPX intrinsics.
• -mno-mpx used to disable -mmpx and is the default behavior.
• -fconcepts-ts previously used to enable experimental concepts support. Use -std=c++2a instead to enable Concepts support.
• …

Modified Compiler Flags¶

• RISC-V now sets the architecture (riscv32/riscv64) based on the value provided to the -march flag, overriding the target provided by -triple.
• -flax-vector-conversions has been split into three different levels of laxness, and has been updated to match the GCC semantics:
  • -flax-vector-conversions=all: This is Clang’s current default, and permits implicit vector conversions (performed as bitcasts) between any two vector types of the same overall bit-width. Former synonym: -flax-vector-conversions (Clang <= 9).
  • -flax-vector-conversions=integer: This permits implicit vector conversions (performed as bitcasts) between any two integer vector types of the same overall bit-width. Synonym: -flax-vector-conversions (Clang >= 10).
  • -flax-vector-conversions=none: Do not perform any implicit bitcasts between vector types. Synonym: -fno-lax-vector-conversions.

New Pragmas in Clang¶

• …

Attribute Changes in Clang¶

• …

Windows Support¶

• Previous Clang versions contained a work-around to avoid an issue with the standard library headers in Visual Studio 2019 versions prior to 16.3. This work-around has now been removed, and users of Visual Studio 2019 are encouraged to upgrade to 16.3 or later, otherwise they may see link errors as below:

  error LNK2005: "bool const std::_Is_integral<int>" (??$_Is_integral@H@std@@3_NB) already defined
  

C Language Changes in Clang¶

• …

C++ Language Changes in Clang¶

• The behaviour of the gnu_inline attribute now matches GCC, for cases where used without the extern keyword. As this is a change compared to how it behaved in previous Clang versions, a warning is emitted for this combination.

Objective-C Language Changes in Clang¶

• In both Objective-C and Objective-C++, -Wcompare-distinct-pointer-types will now warn when comparing ObjC Class with an ObjC instance type pointer.

  Class clz = ...;
  MyType *instance = ...;
  bool eq = (clz == instance); // Previously undiagnosed, now warns.
  

• Objective-C++ now diagnoses conversions between Class and ObjC instance type pointers. Such conversions already emitted an on-by-default -Wincompatible-pointer-types warning in Objective-C mode, but had inadvertently been missed entirely in Objective-C++. This has been fixed, and they are now diagnosed as errors, consistent with the usual C++ treatment for conversions between unrelated pointer types.

  Class clz = ...;
  MyType *instance = ...;
  clz = instance; // Previously undiagnosed, now an error.
  instance = clz; // Previously undiagnosed, now an error.
  

  One particular issue you may run into is attempting to use a class as a key in a dictionary literal. This will now result in an error, because Class is not convertable to id<NSCopying>. (Note that this was already a warning in Objective-C mode.) While an arbitrary Class object is not guaranteed to implement NSCopying, the default metaclass implementation does. Therefore, the recommended solution is to insert an explicit cast to id, which disables the type-checking here.

Class cls = ...;

// Error: cannot convert from Class to id<NSCoding>.
NSDictionary* d = @{cls : @"Hello"};

// Fix: add an explicit cast to 'id'.
NSDictionary* d = @{(id)cls : @"Hello"};

OpenCL C Language Changes in Clang¶

…

ABI Changes in Clang¶

• gcc passes vectors of __int128 in memory on X86-64. Clang historically broke the vectors into multiple scalars using two 64-bit values for each element. Clang now matches the gcc behavior on Linux and NetBSD. You can switch back to old API behavior with flag: -fclang-abi-compat=9.0.
• RISC-V now chooses a default -march= and -mabi= to match (in almost all cases) the GCC defaults. On baremetal targets, where neither -march= nor -mabi= are specified, Clang now differs from GCC by defaulting to -march=rv32imac -mabi=ilp32 or -march=rv64imac -mabi=lp64 depending on the architecture in the target triple. These do not always match the defaults in Clang 9. We strongly suggest that you explicitly pass -march= and -mabi= when compiling for RISC-V, due to how extensible this architecture is.

OpenMP Support in Clang¶

• …

CUDA Support in Clang¶

• …

Internal API Changes¶

These are major API changes that have happened since the 9.0.0 release of Clang. If upgrading an external codebase that uses Clang as a library, this section should help get you past the largest hurdles of upgrading.

• libTooling APIs that transfer ownership of FrontendAction objects now pass them by unique_ptr, making the ownership transfer obvious in the type system. FrontendActionFactory::create() now returns a unique_ptr<FrontendAction>. runToolOnCode, runToolOnCodeWithArgs, ToolInvocation::ToolInvocation() now take a unique_ptr<FrontendAction>.

Build System Changes¶

These are major changes to the build system that have happened since the 9.0.0 release of Clang. Users of the build system should adjust accordingly.

• In 8.0.0 and below, the install-clang-headers target would install clang’s resource directory headers. This installation is now performed by the install-clang-resource-headers target. Users of the old install-clang-headers target should switch to the new install-clang-resource-headers target. The install-clang-headers target now installs clang’s API headers (corresponding to its libraries), which is consistent with the install-llvm-headers target.
• In 9.0.0 and later Clang added a new target, clang-cpp, which generates a shared library comprised of all the clang component libraries and exporting the clang C++ APIs. Additionally the build system gained the new “CLANG_LINK_CLANG_DYLIB” option, which defaults Off, and when set to On, will force clang (and clang-based tools) to link the clang-cpp library instead of statically linking clang’s components. This option will reduce the size of binary distributions at the expense of compiler performance.
• …

AST Matchers¶

• …

clang-format¶

• The Standard style option specifies which version of C++ should be used when parsing and formatting C++ code. The set of allowed values has changed:

  • Latest will always enable new C++ language features.
  • c++03, c++11, c++14, c++17, c++20 will pin to exactly that language version.
  • Auto is the default and detects style from the code (this is unchanged).

  The previous values of Cpp03 and Cpp11 are deprecated. Note that Cpp11 is treated as Latest, as this was always clang-format’s behavior. (One motivation for this change is the new name describes the behavior better).

• Clang-format has a new option called --dry-run or -n to emit a warning for clang-format violations. This can be used together with –ferror-limit=N to limit the number of warnings per file and –Werror to make warnings into errors.

• Option IncludeIsMainSourceRegex has been added to allow for additional suffixes and file extensions to be considered as a source file for execution of logic that looks for “main include file” to put it on top.

  By default, clang-format considers source files as “main” only when they end with: .c, .cc, .cpp, .c++, .cxx, .m or .mm extensions. This config option allows to extend this set of source files considered as “main”.

  For example, if this option is configured to (Impl\.hpp)$, then a file ClassImpl.hpp is considered “main” (in addition to Class.c, Class.cc, Class.cpp and so on) and “main include file” logic will be executed (with IncludeIsMainRegex setting also being respected in later phase). Without this option set, ClassImpl.hpp would not have the main include file put on top before any other include.

• Options DeriveLineEnding and UseCRLF have been added to allow clang-format to control the newlines. DeriveLineEnding is by default true and reflects is the existing mechanism, which based is on majority rule. The new options allows this to be turned off and UseCRLF to control the decision as to which sort of line ending to use.

• Option SpaceBeforeSquareBrackets has been added to insert a space before array declarations.

  int a [5];    vs    int a[5];
  

• Clang-format now supports JavaScript null operators.

  const x = foo ?? default;
  const z = foo?.bar?.baz;
  

libclang¶

• …

Static Analyzer¶

• The Clang analyzer checker DeadStores gets a new option called WarnForDeadNestedAssignments to detect nested dead assignments (enabled by default).
• …

Undefined Behavior Sanitizer (UBSan)¶

• • The pointer-overflow check was extended added to catch the cases where a non-zero offset is applied to a null pointer, or the result of applying the offset is a null pointer.

    #include <cstdint> // for intptr_t
    
    static char *getelementpointer_inbounds(char *base, unsigned long offset) {
      // Potentially UB.
      return base + offset;
    }
    
    char *getelementpointer_unsafe(char *base, unsigned long offset) {
      // Always apply offset. UB if base is ``nullptr`` and ``offset`` is not
      // zero, or if ``base`` is non-``nullptr`` and ``offset`` is
      // ``-reinterpret_cast<intptr_t>(base)``.
      return getelementpointer_inbounds(base, offset);
    }
    
    char *getelementpointer_safe(char *base, unsigned long offset) {
      // Cast pointer to integer, perform usual arithmetic addition,
      // and cast to pointer. This is legal.
      char *computed =
          reinterpret_cast<char *>(reinterpret_cast<intptr_t>(base) + offset);
      // If either the pointer becomes non-``nullptr``, or becomes
      // ``nullptr``, we must use ``computed`` result.
      if (((base == nullptr) && (computed != nullptr)) ||
          ((base != nullptr) && (computed == nullptr)))
        return computed;
      // Else we can use ``getelementpointer_inbounds()``.
      return getelementpointer_inbounds(base, offset);
    }
    

Python Binding Changes¶

The following methods have been added:

• …