Language Compatibility

Clang strives to both conform to current language standards (up to C11 and C++11) and also to implement many widely-used extensions available in other compilers, so that most correct code will "just work" when compiled with Clang. However, Clang is more strict than other popular compilers, and may reject incorrect code that other compilers allow. This page documents common compatibility and portability issues with Clang to help you understand and fix the problem in your code when Clang emits an error message.

C compatibility

C99 inline functions

By default, Clang builds C code in GNU C17 mode, so it uses standard C99 semantics for the inline keyword. These semantics are different from those in GNU C89 mode, which is the default mode in versions of GCC prior to 5.0. For example, consider the following code:

inline int add(int i, int j) { return i + j; }

int main() {
  int i = add(4, 5);
  return i;
}

In C99, inline means that a function's definition is provided only for inlining, and that there is another definition (without inline) somewhere else in the program. That means that this program is incomplete, because if add isn't inlined (for example, when compiling without optimization), then main will have an unresolved reference to that other definition. Therefore we'll get a (correct) link-time error like this:

Undefined symbols:
  "_add", referenced from:
      _main in cc-y1jXIr.o

By contrast, GNU C89 mode (used by default in older versions of GCC) is the C89 standard plus a lot of extensions. C89 doesn't have an inline keyword, but GCC recognizes it as an extension and just treats it as a hint to the optimizer.

There are several ways to fix this problem:

All of this only applies to C code; the meaning of inline in C++ is very different from its meaning in either GNU89 or C99.

"missing" vector __builtin functions

The Intel and AMD manuals document a number "<*mmintrin.h>" header files, which define a standardized API for accessing vector operations on X86 CPUs. These functions have names like _mm_xor_ps and _mm256_addsub_pd. Compilers have leeway to implement these functions however they want. Since Clang supports an excellent set of native vector operations, the Clang headers implement these interfaces in terms of the native vector operations.

In contrast, GCC implements these functions mostly as a 1-to-1 mapping to builtin function calls, like __builtin_ia32_paddw128. These builtin functions are an internal implementation detail of GCC, and are not portable to the Intel compiler, the Microsoft compiler, or Clang. If you get build errors mentioning these, the fix is simple: switch to the *mmintrin.h functions.

The same issue occurs for NEON and Altivec for the ARM and PowerPC architectures respectively. For these, make sure to use the <arm_neon.h> and <altivec.h> headers.

For x86 architectures this script should help with the manual migration process. It will rewrite your source files in place to use the APIs instead of builtin function calls. Just call it like this:

  builtins.py *.c *.h

and it will rewrite all of the .c and .h files in the current directory to use the API calls instead of calls like __builtin_ia32_paddw128.

Lvalue casts

Old versions of GCC permit casting the left-hand side of an assignment to a different type. Clang produces an error on similar code, e.g.,

lvalue.c:2:3: error: assignment to cast is illegal, lvalue casts are not supported
  (int*)addr = val;
  ^~~~~~~~~~ ~

To fix this problem, move the cast to the right-hand side. In this example, one could use:

  addr = (float *)val;

Jumps to within __block variable scope

Clang disallows jumps into the scope of a __block variable. Variables marked with __block require special runtime initialization. A jump into the scope of a __block variable bypasses this initialization, leaving the variable's metadata in an invalid state. Consider the following code fragment:

int fetch_object_state(struct MyObject *c) {
  if (!c->active) goto error;

  __block int result;
  run_specially_somehow(^{ result = c->state; });
  return result;

 error:
  fprintf(stderr, "error while fetching object state");
  return -1;
}

GCC accepts this code, but it produces code that will usually crash when result goes out of scope if the jump is taken. (It's possible for this bug to go undetected because it often won't crash if the stack is fresh, i.e. still zeroed.) Therefore, Clang rejects this code with a hard error:

t.c:3:5: error: goto into protected scope
    goto error;
    ^
t.c:5:15: note: jump bypasses setup of __block variable
  __block int result;
              ^

The fix is to rewrite the code to not require jumping into a __block variable's scope, e.g. by limiting that scope:

  {
    __block int result;
    run_specially_somehow(^{ result = c->state; });
    return result;
  }

Non-initialization of __block variables

In the following example code, the x variable is used before it is defined:

int f0() {
  __block int x;
  return ^(){ return x; }();
}

By an accident of implementation, GCC and llvm-gcc unintentionally always zero initialized __block variables. However, any program which depends on this behavior is relying on unspecified compiler behavior. Programs must explicitly initialize all local block variables before they are used, as with other local variables.

Clang does not zero initialize local block variables, and programs which rely on such behavior will most likely break when built with Clang.

Inline assembly

In general, Clang is highly compatible with the GCC inline assembly extensions, allowing the same set of constraints, modifiers and operands as GCC inline assembly.

On targets that use the integrated assembler (such as most X86 targets), inline assembly is run through the integrated assembler instead of your system assembler (which is most commonly "gas", the GNU assembler). The LLVM integrated assembler is extremely compatible with GAS, but there are a couple of minor places where it is more picky, particularly due to outright GAS bugs.

One specific example is that the assembler rejects ambiguous X86 instructions that don't have suffixes. For example:

  asm("add %al, (%rax)");
  asm("addw $4, (%rax)");
  asm("add $4, (%rax)");

Both clang and GAS accept the first instruction: because the first instruction uses the 8-bit %al register as an operand, it is clear that it is an 8-bit add. The second instruction is accepted by both because the "w" suffix indicates that it is a 16-bit add. The last instruction is accepted by GAS even though there is nothing that specifies the size of the instruction (and the assembler randomly picks a 32-bit add). Because it is ambiguous, Clang rejects the instruction with this error message:

<inline asm>:3:1: error: ambiguous instructions require an explicit suffix (could be 'addb', 'addw', 'addl', or 'addq')
add $4, (%rax)
^

To fix this compatibility issue, add an explicit suffix to the instruction: this makes your code more clear and is compatible with both GCC and Clang.

Objective-C compatibility

Cast of super

GCC treats the super identifier as an expression that can, among other things, be cast to a different type. Clang treats super as a context-sensitive keyword, and will reject a type-cast of super:

super.m:11:12: error: cannot cast 'super' (it isn't an expression)
  [(Super*)super add:4];
   ~~~~~~~~^

To fix this problem, remove the type cast, e.g.

  [super add:4];

Size of interfaces

When using the "non-fragile" Objective-C ABI in use, the size of an Objective-C class may change over time as instance variables are added (or removed). For this reason, Clang rejects the application of the sizeof operator to an Objective-C class when using this ABI:

sizeof.m:4:14: error: invalid application of 'sizeof' to interface 'NSArray' in non-fragile ABI
  int size = sizeof(NSArray);
             ^     ~~~~~~~~~

Code that relies on the size of an Objective-C class is likely to be broken anyway, since that size is not actually constant. To address this problem, use the Objective-C runtime API function class_getInstanceSize():

  class_getInstanceSize([NSArray class])

Internal Objective-C types

GCC allows using pointers to internal Objective-C objects, struct objc_object*, struct objc_selector*, and struct objc_class* in place of the types id, SEL, and Class respectively. Clang treats the internal Objective-C structures as implementation detail and won't do implicit conversions:

t.mm:11:2: error: no matching function for call to 'f'
        f((struct objc_object *)p);
        ^
t.mm:5:6: note: candidate function not viable: no known conversion from 'struct objc_object *' to 'id' for 1st argument
void f(id x);
     ^

Code should use types id, SEL, and Class instead of the internal types.

C variables in @interface or @protocol

GCC allows the declaration of C variables in an @interface or @protocol declaration. Clang does not allow variable declarations to appear within these declarations unless they are marked extern.

Variables may still be declared in an @implementation.

@interface XX
int a;         // not allowed in clang
int b = 1;     // not allowed in clang
extern int c;  // allowed
@end

C++ compatibility

Variable-length arrays

GCC and C99 allow an array's size to be determined at run time. This extension is not permitted in standard C++. However, Clang supports such variable length arrays for compatibility with GNU C and C99 programs.

If you would prefer not to use this extension, you can disable it with -Werror=vla. There are several ways to fix your code:

  1. replace the variable length array with a fixed-size array if you can determine a reasonable upper bound at compile time; sometimes this is as simple as changing int size = ...; to const int size = ...; (if the initializer is a compile-time constant);
  2. use std::vector or some other suitable container type; or
  3. allocate the array on the heap instead using new Type[] - just remember to delete[] it.

Unqualified lookup in templates

Some versions of GCC accept the following invalid code:

template <typename T> T Squared(T x) {
  return Multiply(x, x);
}

int Multiply(int x, int y) {
  return x * y;
}

int main() {
  Squared(5);
}

Clang complains:

my_file.cpp:2:10: error: call to function 'Multiply' that is neither visible in the template definition nor found by argument-dependent lookup
  return Multiply(x, x);
         ^
my_file.cpp:10:3: note: in instantiation of function template specialization 'Squared<int>' requested here
  Squared(5);
  ^
my_file.cpp:5:5: note: 'Multiply' should be declared prior to the call site
int Multiply(int x, int y) {
    ^

The C++ standard says that unqualified names like Multiply are looked up in two ways.

First, the compiler does unqualified lookup in the scope where the name was written. For a template, this means the lookup is done at the point where the template is defined, not where it's instantiated. Since Multiply hasn't been declared yet at this point, unqualified lookup won't find it.

Second, if the name is called like a function, then the compiler also does argument-dependent lookup (ADL). (Sometimes unqualified lookup can suppress ADL; see [basic.lookup.argdep]p3 for more information.) In ADL, the compiler looks at the types of all the arguments to the call. When it finds a class type, it looks up the name in that class's namespace; the result is all the declarations it finds in those namespaces, plus the declarations from unqualified lookup. However, the compiler doesn't do ADL until it knows all the argument types.

In our example, Multiply is called with dependent arguments, so ADL isn't done until the template is instantiated. At that point, the arguments both have type int, which doesn't contain any class types, and so ADL doesn't look in any namespaces. Since neither form of lookup found the declaration of Multiply, the code doesn't compile.

Here's another example, this time using overloaded operators, which obey very similar rules.

#include <iostream>

template<typename T>
void Dump(const T& value) {
  std::cout << value << "\n";
}

namespace ns {
  struct Data {};
}

std::ostream& operator<<(std::ostream& out, ns::Data data) {
  return out << "Some data";
}

void Use() {
  Dump(ns::Data());
}

Again, Clang complains:

my_file2.cpp:5:13: error: call to function 'operator<<' that is neither visible in the template definition nor found by argument-dependent lookup
  std::cout << value << "\n";
            ^
my_file2.cpp:17:3: note: in instantiation of function template specialization 'Dump<ns::Data>' requested here
  Dump(ns::Data());
  ^
my_file2.cpp:12:15: note: 'operator<<' should be declared prior to the call site or in namespace 'ns'
std::ostream& operator<<(std::ostream& out, ns::Data data) {
              ^

Just like before, unqualified lookup didn't find any declarations with the name operator<<. Unlike before, the argument types both contain class types: one of them is an instance of the class template type std::basic_ostream, and the other is the type ns::Data that we declared above. Therefore, ADL will look in the namespaces std and ns for an operator<<. Since one of the argument types was still dependent during the template definition, ADL isn't done until the template is instantiated during Use, which means that the operator<< we want it to find has already been declared. Unfortunately, it was declared in the global namespace, not in either of the namespaces that ADL will look in!

There are two ways to fix this problem:

  1. Make sure the function you want to call is declared before the template that might call it. This is the only option if none of its argument types contain classes. You can do this either by moving the template definition, or by moving the function definition, or by adding a forward declaration of the function before the template.
  2. Move the function into the same namespace as one of its arguments so that ADL applies.

For more information about argument-dependent lookup, see [basic.lookup.argdep]. For more information about the ordering of lookup in templates, see [temp.dep.candidate].

Unqualified lookup into dependent bases of class templates

Some versions of GCC accept the following invalid code:

template <typename T> struct Base {
  void DoThis(T x) {}
  static void DoThat(T x) {}
};

template <typename T> struct Derived : public Base<T> {
  void Work(T x) {
    DoThis(x);  // Invalid!
    DoThat(x);  // Invalid!
  }
};
Clang correctly rejects it with the following errors (when Derived is eventually instantiated):
my_file.cpp:8:5: error: use of undeclared identifier 'DoThis'
    DoThis(x);
    ^
    this->
my_file.cpp:2:8: note: must qualify identifier to find this declaration in dependent base class
  void DoThis(T x) {}
       ^
my_file.cpp:9:5: error: use of undeclared identifier 'DoThat'
    DoThat(x);
    ^
    this->
my_file.cpp:3:15: note: must qualify identifier to find this declaration in dependent base class
  static void DoThat(T x) {}
Like we said above, unqualified names like DoThis and DoThat are looked up when the template Derived is defined, not when it's instantiated. When we look up a name used in a class, we usually look into the base classes. However, we can't look into the base class Base<T> because its type depends on the template argument T, so the standard says we should just ignore it. See [temp.dep]p3 for details.

The fix, as Clang tells you, is to tell the compiler that we want a class member by prefixing the calls with this->:

  void Work(T x) {
    this->DoThis(x);
    this->DoThat(x);
  }
Alternatively, you can tell the compiler exactly where to look:
  void Work(T x) {
    Base<T>::DoThis(x);
    Base<T>::DoThat(x);
  }
This works whether the methods are static or not, but be careful: if DoThis is virtual, calling it this way will bypass virtual dispatch!

Incomplete types in templates

The following code is invalid, but compilers are allowed to accept it:

  class IOOptions;
  template <class T> bool read(T &value) {
    IOOptions opts;
    return read(opts, value);
  }

  class IOOptions { bool ForceReads; };
  bool read(const IOOptions &opts, int &x);
  template bool read<>(int &);
The standard says that types which don't depend on template parameters must be complete when a template is defined if they affect the program's behavior. However, the standard also says that compilers are free to not enforce this rule. Most compilers enforce it to some extent; for example, it would be an error in GCC to write opts.ForceReads in the code above. In Clang, we feel that enforcing the rule consistently lets us provide a better experience, but unfortunately it also means we reject some code that other compilers accept.

We've explained the rule here in very imprecise terms; see [temp.res]p8 for details.

Templates with no valid instantiations

The following code contains a typo: the programmer meant init() but wrote innit() instead.

  template <class T> class Processor {
    ...
    void init();
    ...
  };
  ...
  template <class T> void process() {
    Processor<T> processor;
    processor.innit();       // <-- should be 'init()'
    ...
  }
Unfortunately, we can't flag this mistake as soon as we see it: inside a template, we're not allowed to make assumptions about "dependent types" like Processor<T>. Suppose that later on in this file the programmer adds an explicit specialization of Processor, like so:
  template <> class Processor<char*> {
    void innit();
  };
Now the program will work — as long as the programmer only ever instantiates process() with T = char*! This is why it's hard, and sometimes impossible, to diagnose mistakes in a template definition before it's instantiated.

The standard says that a template with no valid instantiations is ill-formed. Clang tries to do as much checking as possible at definition-time instead of instantiation-time: not only does this produce clearer diagnostics, but it also substantially improves compile times when using pre-compiled headers. The downside to this philosophy is that Clang sometimes fails to process files because they contain broken templates that are no longer used. The solution is simple: since the code is unused, just remove it.

Default initialization of const variable of a class type requires user-defined default constructor

If a class or struct has no user-defined default constructor, C++ doesn't allow you to default construct a const instance of it like this ([dcl.init], p9):

class Foo {
 public:
  // The compiler-supplied default constructor works fine, so we
  // don't bother with defining one.
  ...
};

void Bar() {
  const Foo foo;  // Error!
  ...
}
To fix this, you can define a default constructor for the class:
class Foo {
 public:
  Foo() {}
  ...
};

void Bar() {
  const Foo foo;  // Now the compiler is happy.
  ...
}
An upcoming change to the C++ standard is expected to weaken this rule to only apply when the compiler-supplied default constructor would leave a member uninitialized. Clang implements the more relaxed rule in version 3.8 onwards.

Parameter name lookup

Some versions of GCC allow the redeclaration of function parameter names within a function prototype in C++ code, e.g.

void f(int a, int a);

Clang diagnoses this error (where the parameter name has been redeclared). To fix this problem, rename one of the parameters.

C++11 compatibility

Deleted special member functions

In C++11, the explicit declaration of a move constructor or a move assignment operator within a class deletes the implicit declaration of the copy constructor and copy assignment operator. This change came fairly late in the C++11 standardization process, so early implementations of C++11 (including Clang before 3.0, GCC before 4.7, and Visual Studio 2010) do not implement this rule, leading them to accept this ill-formed code:

struct X {
  X(X&&); // deletes implicit copy constructor:
  // X(const X&) = delete;
};

void f(X x);
void g(X x) {
  f(x); // error: X has a deleted copy constructor
}

This affects some early C++11 code, including Boost's popular shared_ptr up to version 1.47.0. The fix for Boost's shared_ptr is available here.

Objective-C++ compatibility

Implicit downcasts

Due to a bug in its implementation, GCC allows implicit downcasts of Objective-C pointers (from a base class to a derived class) when calling functions. Such code is inherently unsafe, since the object might not actually be an instance of the derived class, and is rejected by Clang. For example, given this code:

@interface Base @end
@interface Derived : Base @end

void f(Derived *p);
void g(Base *p) {
  f(p);
}

Clang produces the following error:

downcast.mm:6:3: error: no matching function for call to 'f'
  f(p);
  ^
downcast.mm:4:6: note: candidate function not viable: cannot convert from
      superclass 'Base *' to subclass 'Derived *' for 1st argument
void f(Derived *p);
     ^

If the downcast is actually correct (e.g., because the code has already checked that the object has the appropriate type), add an explicit cast:

  f((Derived *)base);

Using class as a property name

In C and Objective-C, class is a normal identifier and can be used to name fields, ivars, methods, and so on. In C++, class is a keyword. For compatibility with existing code, Clang permits class to be used as part of a method selector in Objective-C++, but this does not extend to any other part of the language. In particular, it is impossible to use property dot syntax in Objective-C++ with the property name class, so the following code will fail to parse:

@interface I {
int cls;
}
+ (int)class;
@end

@implementation  I
- (int) Meth { return I.class; }
@end

Use explicit message-send syntax instead, i.e. [I class].