Non-comprehensive list of changes in this release¶

• Support for retpolines was added to help mitigate “branch target injection” (variant #2) of the “Spectre” speculative side channels described by Project Zero and the Spectre paper.
• Bitrig OS was merged back into OpenBSD, so Bitrig support has been removed from Clang/LLVM.
• The default value of _MSC_VER was raised from 1800 to 1911, making it compatible with the Visual Studio 2015 and 2017 C++ standard library headers. Users should generally expect this to be regularly raised to match the most recently released version of the Visual C++ compiler.
• clang now defaults to .init_array if no gcc installation can be found. If a gcc installation is found, it still prefers .ctors if the found gcc is older than 4.7.0.
• The new builtin preprocessor macros __is_target_arch, __is_target_vendor, __is_target_os, and __is_target_environment can be used to to examine the individual components of the target triple.

Improvements to Clang’s diagnostics¶

• -Wpragma-pack is a new warning that warns in the following cases:
  • When a translation unit is missing terminating #pragma pack (pop) directives.
  • When leaving an included file that changes the current alignment value, i.e. when the alignment before #include is different to the alignment after #include.
  • -Wpragma-pack-suspicious-include (disabled by default) warns on an #include when the included file contains structures or unions affected by a non-default alignment that has been specified using a #pragma pack directive prior to the #include.
• -Wobjc-messaging-id is a new, non-default warning that warns about message sends to unqualified id in Objective-C. This warning is useful for projects that would like to avoid any potential future compiler errors/warnings, as the system frameworks might add a method with the same selector which could make the message send to id ambiguous.
• -Wtautological-compare now warns when comparing an unsigned integer and 0 regardless of whether the constant is signed or unsigned.
• -Wtautological-compare now warns about comparing a signed integer and 0 when the signed integer is coerced to an unsigned type for the comparison. -Wsign-compare was adjusted not to warn in this case.
• -Wtautological-constant-compare is a new warning that warns on tautological comparisons between integer variable of the type T and the largest/smallest possible integer constant of that same type.
• For C code, -Wsign-compare, -Wsign-conversion, -Wtautological-constant-compare and -Wtautological-constant-out-of-range-compare were adjusted to use the underlying datatype of enum.
• -Wnull-pointer-arithmetic now warns about performing pointer arithmetic on a null pointer. Such pointer arithmetic has an undefined behavior if the offset is nonzero. It also now warns about arithmetic on a null pointer treated as a cast from integer to pointer (GNU extension).
• -Wzero-as-null-pointer-constant was adjusted not to warn on null pointer constants that originate from system macros, except NULL macro.
• -Wdelete-non-virtual-dtor can now fire in system headers, so that std::unique_ptr<> deleting through a non-virtual dtor is now diagnosed.
• -Wunreachable-code can now reason about __try, __except and __leave.

New Compiler Flags¶

• Clang supports the -mretpoline flag to enable retpolines. Code compiled with this flag will be hardened against variant #2 of the Spectre attack. Indirect branches from switches or gotos removed from the code, and indirect calls will be made through a “retpoline” thunk. The necessary thunks will automatically be inserted into the generated code. Clang also supports -mretpoline-external-thunk which works like -mretpoline but requires the user to provide their own thunk definitions. The external thunk names start with __x86_indirect_thunk_ and end in a register name. For 64-bit platforms, only an r11 thunk is used, but for 32-bit platforms eax, ecx, edx, and edi thunks are used.
• Clang now supports configuration files. These are collections of driver options, which can be applied by specifying the configuration file, either using command line option --config foo.cfg or encoding it into executable name foo-clang. Clang behaves as if the options from this file were inserted before the options specified in command line. This feature is primary intended to facilitate cross compilation. Details can be found in Clang Compiler User’s Manual.
• The -fdouble-square-bracket-attributes and corresponding -fno-double-square-bracket-attributes flags were added to enable or disable [[]] attributes in any language mode. Currently, only a limited number of attributes are supported outside of C++ mode. See the Clang attribute documentation for more information about which attributes are supported for each syntax.
• Added the -std=c17, -std=gnu17, and -std=iso9899:2017 language mode flags for compatibility with GCC. This enables support for the next version of the C standard, expected to be published by ISO in 2018. The only difference between the -std=c17 and -std=c11 language modes is the value of the __STDC_VERSION__ macro, as C17 is a bug fix release.
• Added the -fexperimental-isel and -fno-experimental-isel flags to enable/disable the new GlobalISel instruction selection framework. This feature is enabled by default for AArch64 at the -O0 optimization level. Support for other targets or optimization levels is currently incomplete.
• New -nostdlib++ flag to disable linking the C++ standard library. Similar to using clang instead of clang++ but doesn’t disable -lm.

Attribute Changes in Clang¶

• Clang now supports the majority of its attributes under both the GNU-style spelling (__attribute((name))) and the double square-bracket spelling in the clang vendor namespace ([[clang::name]]). Attributes whose syntax is specified by some other standard (such as CUDA and OpenCL attributes) continue to follow their respective specification.
• Added the __has_c_attribute() builtin preprocessor macro which allows users to dynamically detect whether a double square-bracket attribute is supported in C mode. This attribute syntax can be enabled with the -fdouble-square-bracket-attributes flag.
• The presence of __attribute__((availability(...))) on a declaration no longer implies default visibility for that declaration on macOS.

Windows Support¶

• Clang now has initial, preliminary support for targeting Windows on ARM64.
• clang-cl now exposes the --version flag.

C++ Language Changes in Clang¶

• Clang’s default C++ dialect is now gnu++14 instead of gnu++98. This means Clang will by default accept code using features from C++14 and conforming GNU extensions. Projects incompatible with C++14 can add -std=gnu++98 to their build settings to restore the previous behaviour.

• Added support for some features from the C++ standard after C++17 (provisionally known as C++2a but expected to be C++20). This support can be enabled with the -std=c++2a flag. This enables:

  • Support for __VA_OPT__, to allow variadic macros to easily provide different expansions when they are invoked without variadic arguments.
  • Recognition of the <=> token (the C++2a three-way comparison operator).
  • Support for default member initializers for bit-fields.
  • Lambda capture of *this.
  • Pointer-to-member calls using const &-qualified pointers on temporary objects.

  All of these features other than __VA_OPT__ and <=> are made available with a warning in earlier C++ language modes.

• A warning has been added for a <= token followed immediately by a > character. Code containing such constructs will change meaning in C++2a due to the addition of the <=> operator.

• Clang implements the “destroying operator delete” feature described in C++ committee paper P0722R1 <http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0722r1.html>, which is targeting inclusion in C++2a but has not yet been voted into the C++ working draft. Support for this feature is enabled by the presence of the standard library type std::destroying_delete_t.

OpenCL C Language Changes in Clang¶

• Added subgroup builtins to enqueue kernel support.
• Added CL2.0 atomics as Clang builtins that now accept an additional memory scope parameter propagated to atomic IR instructions (this is to align with the corresponding change in LLVM IR) (see spec s6.13.11.4).
• Miscellaneous fixes in the CL header.
• Allow per target selection of address space during CodeGen of certain OpenCL types. Default target implementation is provided mimicking old behavior.
• Macro __IMAGE_SUPPORT__ is now automatically added (as per spec s6.10).
• Added cl_intel_subgroups and cl_intel_subgroups_short extensions.
• All function calls are marked by the convergent attribute to prevent optimizations that break SPMD program semantics. This will be removed by LLVM passes if it can be proved that the function does not use convergent operations.
• Create a kernel wrapper for enqueued blocks, which simplifies enqueue support by providing common functionality.
• Added private address space explicitly in AST and refactored address space support with several simplifications and bug fixes (PR33419 and PR33420).
• OpenCL now allows functions with empty parameters to be treated as if they had a void parameter list (inspired from C++ support). OpenCL C spec update to follow.
• General miscellaneous refactoring and cleanup of blocks support for OpenCL to remove unused parts inherited from Objective C implementation.
• Miscellaneous improvements in vector diagnostics.
• Added half float load and store builtins without enabling half as a legal type (__builtin_store_half for double, __builtin_store_halff for float, __builtin_load_half for double, __builtin_load_halff for float).

OpenMP Support in Clang¶

• Added options -f[no]-openmp-simd that support code emission only for OpenMP SIMD-based directives, like #pragma omp simd, #pragma omp parallel for simd etc. The code is emitted only for SIMD-based part of the combined directives and clauses.
• Added support for almost all target-based directives except for #pragma omp target teams distribute parallel for [simd]. Although, please note that depend clauses on target-based directives are not supported yet. Clang supports offloading to X86_64, AArch64 and PPC64[LE] devices.
• Added support for reduction-based clauses on task-based directives from upcoming OpenMP 5.0.
• The LLVM OpenMP runtime libomp now supports the OpenMP Tools Interface (OMPT) on x86, x86_64, AArch64, and PPC64 on Linux, Windows, and macOS. If you observe a measurable performance impact on one of your applications without a tool attached, please rebuild the runtime library with -DLIBOMP_OMPT_SUPPORT=OFF and file a bug at LLVM’s Bugzilla or send a message to the OpenMP development list.

AST Matchers¶

The hasDeclaration matcher now works the same for Type and QualType and only ever looks through one level of sugaring in a limited number of cases.

There are two main patterns affected by this:

• qualType(hasDeclaration(recordDecl(...))): previously, we would look through sugar like TypedefType to get at the underlying recordDecl; now, we need to explicitly remove the sugaring: qualType(hasUnqualifiedDesugaredType(hasDeclaration(recordDecl(...))))

• hasType(recordDecl(...)): hasType internally uses hasDeclaration; previously, this matcher used to match for example TypedefTypes of the RecordType, but after the change they don’t; to fix, use:

  hasType(hasUnqualifiedDesugaredType(
      recordType(hasDeclaration(recordDecl(...)))))
  

• templateSpecializationType(hasDeclaration(classTemplateDecl(...))): previously, we would directly match the underlying ClassTemplateDecl; now, we can explicitly match the ClassTemplateSpecializationDecl, but that requires to explicitly get the ClassTemplateDecl:

  templateSpecializationType(hasDeclaration(
      classTemplateSpecializationDecl(
          hasSpecializedTemplate(classTemplateDecl(...)))))
  

clang-format¶

• Option IndentPPDirectives added to indent preprocessor directives on conditionals.

  Before	After
  #if FOO #if BAR #include <foo> #endif #endif	#if FOO # if BAR # include <foo> # endif #endif

• Option -verbose added to the command line. Shows the list of processed files.

• Option IncludeBlocks added to merge and regroup multiple #include blocks during sorting.

  Before (Preserve)	Merge	Regroup
  #include "b.h" #include "a.b" #include <lib/main.h>	#include "a.h" #include "b.h" #include <lib/main.h>	#include "a.h" #include "b.h" #include <lib/main.h>

Static Analyzer¶

• The Static Analyzer can now properly detect and diagnose unary pre-/post- increment/decrement on an uninitialized value.

Undefined Behavior Sanitizer (UBSan)¶

• A minimal runtime is now available. It is suitable for use in production environments, and has a small attack surface. It only provides very basic issue logging and deduplication, and does not support -fsanitize=vptr checking.