The Diagnostic*Kinds.td files¶

Diagnostics are created by adding an entry to one of the clang/Basic/Diagnostic*Kinds.td files, depending on what library will be using it. From this file, tblgen generates the unique ID of the diagnostic, the severity of the diagnostic and the English translation + format string.

There is little sanity with the naming of the unique ID’s right now. Some start with err_, warn_, ext_ to encode the severity into the name. Since the enum is referenced in the C++ code that produces the diagnostic, it is somewhat useful for it to be reasonably short.

The severity of the diagnostic comes from the set {NOTE, REMARK, WARNING, EXTENSION, EXTWARN, ERROR}. The ERROR severity is used for diagnostics indicating the program is never acceptable under any circumstances. When an error is emitted, the AST for the input code may not be fully built. The EXTENSION and EXTWARN severities are used for extensions to the language that Clang accepts. This means that Clang fully understands and can represent them in the AST, but we produce diagnostics to tell the user their code is non-portable. The difference is that the former are ignored by default, and the later warn by default. The WARNING severity is used for constructs that are valid in the currently selected source language but that are dubious in some way. The REMARK severity provides generic information about the compilation that is not necessarily related to any dubious code. The NOTE level is used to staple more information onto previous diagnostics.

These severities are mapped into a smaller set (the Diagnostic::Level enum, {Ignored, Note, Remark, Warning, Error, Fatal}) of output levels by the diagnostics subsystem based on various configuration options. Clang internally supports a fully fine grained mapping mechanism that allows you to map almost any diagnostic to the output level that you want. The only diagnostics that cannot be mapped are NOTEs, which always follow the severity of the previously emitted diagnostic and ERRORs, which can only be mapped to Fatal (it is not possible to turn an error into a warning, for example).

Diagnostic mappings are used in many ways. For example, if the user specifies -pedantic, EXTENSION maps to Warning, if they specify -pedantic-errors, it turns into Error. This is used to implement options like -Wunused_macros, -Wundef etc.

Mapping to Fatal should only be used for diagnostics that are considered so severe that error recovery won’t be able to recover sensibly from them (thus spewing a ton of bogus errors). One example of this class of error are failure to #include a file.

The Format String¶

The format string for the diagnostic is very simple, but it has some power. It takes the form of a string in English with markers that indicate where and how arguments to the diagnostic are inserted and formatted. For example, here are some simple format strings:

"binary integer literals are an extension"
"format string contains '\\0' within the string body"
"more '%%' conversions than data arguments"
"invalid operands to binary expression (%0 and %1)"
"overloaded '%0' must be a %select{unary|binary|unary or binary}2 operator"
     " (has %1 parameter%s1)"

These examples show some important points of format strings. You can use any plain ASCII character in the diagnostic string except “%” without a problem, but these are C strings, so you have to use and be aware of all the C escape sequences (as in the second example). If you want to produce a “%” in the output, use the “%%” escape sequence, like the third diagnostic. Finally, Clang uses the “%...[digit]” sequences to specify where and how arguments to the diagnostic are formatted.

Arguments to the diagnostic are numbered according to how they are specified by the C++ code that produces them, and are referenced by %0 .. %9. If you have more than 10 arguments to your diagnostic, you are doing something wrong :). Unlike printf, there is no requirement that arguments to the diagnostic end up in the output in the same order as they are specified, you could have a format string with “%1 %0” that swaps them, for example. The text in between the percent and digit are formatting instructions. If there are no instructions, the argument is just turned into a string and substituted in.

Here are some “best practices” for writing the English format string:

• Keep the string short. It should ideally fit in the 80 column limit of the DiagnosticKinds.td file. This avoids the diagnostic wrapping when printed, and forces you to think about the important point you are conveying with the diagnostic.
• Take advantage of location information. The user will be able to see the line and location of the caret, so you don’t need to tell them that the problem is with the 4th argument to the function: just point to it.
• Do not capitalize the diagnostic string, and do not end it with a period.
• If you need to quote something in the diagnostic string, use single quotes.

Diagnostics should never take random English strings as arguments: you shouldn’t use “you have a problem with %0” and pass in things like “your argument” or “your return value” as arguments. Doing this prevents translating the Clang diagnostics to other languages (because they’ll get random English words in their otherwise localized diagnostic). The exceptions to this are C/C++ language keywords (e.g., auto, const, mutable, etc) and C/C++ operators (/=). Note that things like “pointer” and “reference” are not keywords. On the other hand, you can include anything that comes from the user’s source code, including variable names, types, labels, etc. The “select” format can be used to achieve this sort of thing in a localizable way, see below.

Formatting a Diagnostic Argument¶

Arguments to diagnostics are fully typed internally, and come from a couple different classes: integers, types, names, and random strings. Depending on the class of the argument, it can be optionally formatted in different ways. This gives the DiagnosticConsumer information about what the argument means without requiring it to use a specific presentation (consider this MVC for Clang :).

Here are the different diagnostic argument formats currently supported by Clang:

“s” format

Example:

"requires %1 parameter%s1"

Class:

Integers

Description:

This is a simple formatter for integers that is useful when producing English diagnostics. When the integer is 1, it prints as nothing. When the integer is not 1, it prints as “s”. This allows some simple grammatical forms to be to be handled correctly, and eliminates the need to use gross things like "requires %1 parameter(s)".

“select” format

Example:

"must be a %select{unary|binary|unary or binary}2 operator"

Class:

Integers

Description:

This format specifier is used to merge multiple related diagnostics together into one common one, without requiring the difference to be specified as an English string argument. Instead of specifying the string, the diagnostic gets an integer argument and the format string selects the numbered option. In this case, the “%2” value must be an integer in the range [0..2]. If it is 0, it prints “unary”, if it is 1 it prints “binary” if it is 2, it prints “unary or binary”. This allows other language translations to substitute reasonable words (or entire phrases) based on the semantics of the diagnostic instead of having to do things textually. The selected string does undergo formatting.

“plural” format

Example:

"you have %1 %plural{1:mouse|:mice}1 connected to your computer"

Class:

Integers

Description:

This is a formatter for complex plural forms. It is designed to handle even the requirements of languages with very complex plural forms, as many Baltic languages have. The argument consists of a series of expression/form pairs, separated by “:”, where the first form whose expression evaluates to true is the result of the modifier.

An expression can be empty, in which case it is always true. See the example at the top. Otherwise, it is a series of one or more numeric conditions, separated by “,”. If any condition matches, the expression matches. Each numeric condition can take one of three forms.

• number: A simple decimal number matches if the argument is the same as the number. Example: "%plural{1:mouse|:mice}4"
• range: A range in square brackets matches if the argument is within the range. Then range is inclusive on both ends. Example: "%plural{0:none|1:one|[2,5]:some|:many}2"
• modulo: A modulo operator is followed by a number, and equals sign and either a number or a range. The tests are the same as for plain numbers and ranges, but the argument is taken modulo the number first. Example: "%plural{%100=0:even hundred|%100=[1,50]:lower half|:everything else}1"

The parser is very unforgiving. A syntax error, even whitespace, will abort, as will a failure to match the argument against any expression.

“ordinal” format

Example:

"ambiguity in %ordinal0 argument"

Class:

Integers

Description:

This is a formatter which represents the argument number as an ordinal: the value 1 becomes 1st, 3 becomes 3rd, and so on. Values less than 1 are not supported. This formatter is currently hard-coded to use English ordinals.

“objcclass” format

Example:

"method %objcclass0 not found"

Class:

DeclarationName

Description:

This is a simple formatter that indicates the DeclarationName corresponds to an Objective-C class method selector. As such, it prints the selector with a leading “+”.

“objcinstance” format

Example:

"method %objcinstance0 not found"

Class:

DeclarationName

Description:

This is a simple formatter that indicates the DeclarationName corresponds to an Objective-C instance method selector. As such, it prints the selector with a leading “-“.

“q” format

Example:

"candidate found by name lookup is %q0"

Class:

NamedDecl *

Description:

This formatter indicates that the fully-qualified name of the declaration should be printed, e.g., “std::vector” rather than “vector”.

“diff” format

Example:

"no known conversion %diff{from $ to $|from argument type to parameter type}1,2"

Class:

QualType

Description:

This formatter takes two QualTypes and attempts to print a template difference between the two. If tree printing is off, the text inside the braces before the pipe is printed, with the formatted text replacing the $. If tree printing is on, the text after the pipe is printed and a type tree is printed after the diagnostic message.

It is really easy to add format specifiers to the Clang diagnostics system, but they should be discussed before they are added. If you are creating a lot of repetitive diagnostics and/or have an idea for a useful formatter, please bring it up on the cfe-dev mailing list.

“sub” format

Example:

Given the following record definition of type TextSubstitution:

def select_ovl_candidate : TextSubstitution<
  "%select{function|constructor}0%select{| template| %2}1">;

which can be used as

def note_ovl_candidate : Note<
  "candidate %sub{select_ovl_candidate}3,2,1 not viable">;

and will act as if it was written "candidate %select{function|constructor}3%select{| template| %1}2 not viable".

Description:

This format specifier is used to avoid repeating strings verbatim in multiple diagnostics. The argument to %sub must name a TextSubstitution tblgen record. The substitution must specify all arguments used by the substitution, and the modifier indexes in the substitution are re-numbered accordingly. The substituted text must itself be a valid format string before substitution.

Producing the Diagnostic¶

Now that you’ve created the diagnostic in the Diagnostic*Kinds.td file, you need to write the code that detects the condition in question and emits the new diagnostic. Various components of Clang (e.g., the preprocessor, Sema, etc.) provide a helper function named “Diag”. It creates a diagnostic and accepts the arguments, ranges, and other information that goes along with it.

For example, the binary expression error comes from code like this:

if (various things that are bad)
  Diag(Loc, diag::err_typecheck_invalid_operands)
    << lex->getType() << rex->getType()
    << lex->getSourceRange() << rex->getSourceRange();

This shows that use of the Diag method: it takes a location (a SourceLocation object) and a diagnostic enum value (which matches the name from Diagnostic*Kinds.td). If the diagnostic takes arguments, they are specified with the << operator: the first argument becomes %0, the second becomes %1, etc. The diagnostic interface allows you to specify arguments of many different types, including int and unsigned for integer arguments, const char* and std::string for string arguments, DeclarationName and const IdentifierInfo * for names, QualType for types, etc. SourceRanges are also specified with the << operator, but do not have a specific ordering requirement.

As you can see, adding and producing a diagnostic is pretty straightforward. The hard part is deciding exactly what you need to say to help the user, picking a suitable wording, and providing the information needed to format it correctly. The good news is that the call site that issues a diagnostic should be completely independent of how the diagnostic is formatted and in what language it is rendered.

Fix-It Hints¶

In some cases, the front end emits diagnostics when it is clear that some small change to the source code would fix the problem. For example, a missing semicolon at the end of a statement or a use of deprecated syntax that is easily rewritten into a more modern form. Clang tries very hard to emit the diagnostic and recover gracefully in these and other cases.

However, for these cases where the fix is obvious, the diagnostic can be annotated with a hint (referred to as a “fix-it hint”) that describes how to change the code referenced by the diagnostic to fix the problem. For example, it might add the missing semicolon at the end of the statement or rewrite the use of a deprecated construct into something more palatable. Here is one such example from the C++ front end, where we warn about the right-shift operator changing meaning from C++98 to C++11:

test.cpp:3:7: warning: use of right-shift operator ('>>') in template argument
                       will require parentheses in C++11
A<100 >> 2> *a;
      ^
  (       )

Here, the fix-it hint is suggesting that parentheses be added, and showing exactly where those parentheses would be inserted into the source code. The fix-it hints themselves describe what changes to make to the source code in an abstract manner, which the text diagnostic printer renders as a line of “insertions” below the caret line. Other diagnostic clients might choose to render the code differently (e.g., as markup inline) or even give the user the ability to automatically fix the problem.

Fix-it hints on errors and warnings need to obey these rules:

• Since they are automatically applied if -Xclang -fixit is passed to the driver, they should only be used when it’s very likely they match the user’s intent.
• Clang must recover from errors as if the fix-it had been applied.
• Fix-it hints on a warning must not change the meaning of the code. However, a hint may clarify the meaning as intentional, for example by adding parentheses when the precedence of operators isn’t obvious.

If a fix-it can’t obey these rules, put the fix-it on a note. Fix-its on notes are not applied automatically.

All fix-it hints are described by the FixItHint class, instances of which should be attached to the diagnostic using the << operator in the same way that highlighted source ranges and arguments are passed to the diagnostic. Fix-it hints can be created with one of three constructors:

• FixItHint::CreateInsertion(Loc, Code)

  Specifies that the given Code (a string) should be inserted before the source location Loc.

• FixItHint::CreateRemoval(Range)

  Specifies that the code in the given source Range should be removed.

• FixItHint::CreateReplacement(Range, Code)

  Specifies that the code in the given source Range should be removed, and replaced with the given Code string.

The DiagnosticConsumer Interface¶

Once code generates a diagnostic with all of the arguments and the rest of the relevant information, Clang needs to know what to do with it. As previously mentioned, the diagnostic machinery goes through some filtering to map a severity onto a diagnostic level, then (assuming the diagnostic is not mapped to “Ignore”) it invokes an object that implements the DiagnosticConsumer interface with the information.

It is possible to implement this interface in many different ways. For example, the normal Clang DiagnosticConsumer (named TextDiagnosticPrinter) turns the arguments into strings (according to the various formatting rules), prints out the file/line/column information and the string, then prints out the line of code, the source ranges, and the caret. However, this behavior isn’t required.

Another implementation of the DiagnosticConsumer interface is the TextDiagnosticBuffer class, which is used when Clang is in -verify mode. Instead of formatting and printing out the diagnostics, this implementation just captures and remembers the diagnostics as they fly by. Then -verify compares the list of produced diagnostics to the list of expected ones. If they disagree, it prints out its own output. Full documentation for the -verify mode can be found in the Clang API documentation for VerifyDiagnosticConsumer.

There are many other possible implementations of this interface, and this is why we prefer diagnostics to pass down rich structured information in arguments. For example, an HTML output might want declaration names be linkified to where they come from in the source. Another example is that a GUI might let you click on typedefs to expand them. This application would want to pass significantly more information about types through to the GUI than a simple flat string. The interface allows this to happen.

Adding Translations to Clang¶

Not possible yet! Diagnostic strings should be written in UTF-8, the client can translate to the relevant code page if needed. Each translation completely replaces the format string for the diagnostic.

Immutability¶

Clang AST nodes (types, declarations, statements, expressions, and so on) are generally designed to be immutable once created. This provides a number of key benefits:

• Canonicalization of the “meaning” of nodes is possible as soon as the nodes are created, and is not invalidated by later addition of more information. For example, we canonicalize types, and use a canonicalized representation of expressions when determining whether two function template declarations involving dependent expressions declare the same entity.
• AST nodes can be reused when they have the same meaning. For example, we reuse Type nodes when representing the same type (but maintain separate TypeLocs for each instance where a type is written), and we reuse non-dependent Stmt and Expr nodes across instantiations of a template.
• Serialization and deserialization of the AST to/from AST files is simpler: we do not need to track modifications made to AST nodes imported from AST files and serialize separate “update records”.

There are unfortunately exceptions to this general approach, such as:

• The first declaration of a redeclarable entity maintains a pointer to the most recent declaration of that entity, which naturally needs to change as more declarations are parsed.
• Name lookup tables in declaration contexts change after the namespace declaration is formed.
• We attempt to maintain only a single declaration for an instantiation of a template, rather than having distinct declarations for an instantiation of the declaration versus the definition, so template instantiation often updates parts of existing declarations.
• Some parts of declarations are required to be instantiated separately (this includes default arguments and exception specifications), and such instantiations update the existing declaration.

These cases tend to be fragile; mutable AST state should be avoided where possible.

As a consequence of this design principle, we typically do not provide setters for AST state. (Some are provided for short-term modifications intended to be used immediately after an AST node is created and before it’s “published” as part of the complete AST, or where language semantics require after-the-fact updates.)

Faithfulness¶

The AST intends to provide a representation of the program that is faithful to the original source. We intend for it to be possible to write refactoring tools using only information stored in, or easily reconstructible from, the Clang AST. This means that the AST representation should either not desugar source-level constructs to simpler forms, or – where made necessary by language semantics or a clear engineering tradeoff – should desugar minimally and wrap the result in a construct representing the original source form.

For example, CXXForRangeStmt directly represents the syntactic form of a range-based for statement, but also holds a semantic representation of the range declaration and iterator declarations. It does not contain a fully-desugared ForStmt, however.

Some AST nodes (for example, ParenExpr) represent only syntax, and others (for example, ImplicitCastExpr) represent only semantics, but most nodes will represent a combination of syntax and associated semantics. Inheritance is typically used when representing different (but related) syntaxes for nodes with the same or similar semantics.

Canonical Types¶

Every instance of the Type class contains a canonical type pointer. For simple types with no typedefs involved (e.g., “int”, “int*”, “int**”), the type just points to itself. For types that have a typedef somewhere in their structure (e.g., “foo”, “foo*”, “foo**”, “bar”), the canonical type pointer points to their structurally equivalent type without any typedefs (e.g., “int”, “int*”, “int**”, and “int*” respectively).

This design provides a constant time operation (dereferencing the canonical type pointer) that gives us access to the structure of types. For example, we can trivially tell that “bar” and “foo*” are the same type by dereferencing their canonical type pointers and doing a pointer comparison (they both point to the single “int*” type).

Canonical types and typedef types bring up some complexities that must be carefully managed. Specifically, the isa/cast/dyn_cast operators generally shouldn’t be used in code that is inspecting the AST. For example, when type checking the indirection operator (unary “*” on a pointer), the type checker must verify that the operand has a pointer type. It would not be correct to check that with “isa<PointerType>(SubExpr->getType())”, because this predicate would fail if the subexpression had a typedef type.

The solution to this problem are a set of helper methods on Type, used to check their properties. In this case, it would be correct to use “SubExpr->getType()->isPointerType()” to do the check. This predicate will return true if the canonical type is a pointer, which is true any time the type is structurally a pointer type. The only hard part here is remembering not to use the isa/cast/dyn_cast operations.

The second problem we face is how to get access to the pointer type once we know it exists. To continue the example, the result type of the indirection operator is the pointee type of the subexpression. In order to determine the type, we need to get the instance of PointerType that best captures the typedef information in the program. If the type of the expression is literally a PointerType, we can return that, otherwise we have to dig through the typedefs to find the pointer type. For example, if the subexpression had type “foo*”, we could return that type as the result. If the subexpression had type “bar”, we want to return “foo*” (note that we do not want “int*”). In order to provide all of this, Type has a getAsPointerType() method that checks whether the type is structurally a PointerType and, if so, returns the best one. If not, it returns a null pointer.

This structure is somewhat mystical, but after meditating on it, it will make sense to you :).

Redeclarations and Overloads¶

Within a translation unit, it is common for an entity to be declared several times. For example, we might declare a function “f” and then later re-declare it as part of an inlined definition:

void f(int x, int y, int z = 1);

inline void f(int x, int y, int z) { /* ...  */ }

The representation of “f” differs in the source-centric and semantics-centric views of a declaration context. In the source-centric view, all redeclarations will be present, in the order they occurred in the source code, making this view suitable for clients that wish to see the structure of the source code. In the semantics-centric view, only the most recent “f” will be found by the lookup, since it effectively replaces the first declaration of “f”.

(Note that because f can be redeclared at block scope, or in a friend declaration, etc. it is possible that the declaration of f found by name lookup will not be the most recent one.)

In the semantics-centric view, overloading of functions is represented explicitly. For example, given two declarations of a function “g” that are overloaded, e.g.,

void g();
void g(int);

the DeclContext::lookup operation will return a DeclContext::lookup_result that contains a range of iterators over declarations of “g”. Clients that perform semantic analysis on a program that is not concerned with the actual source code will primarily use this semantics-centric view.

Lexical and Semantic Contexts¶

Each declaration has two potentially different declaration contexts: a lexical context, which corresponds to the source-centric view of the declaration context, and a semantic context, which corresponds to the semantics-centric view. The lexical context is accessible via Decl::getLexicalDeclContext while the semantic context is accessible via Decl::getDeclContext, both of which return DeclContext pointers. For most declarations, the two contexts are identical. For example:

class X {
public:
  void f(int x);
};

Here, the semantic and lexical contexts of X::f are the DeclContext associated with the class X (itself stored as a RecordDecl AST node). However, we can now define X::f out-of-line:

void X::f(int x = 17) { /* ...  */ }

This definition of “f” has different lexical and semantic contexts. The lexical context corresponds to the declaration context in which the actual declaration occurred in the source code, e.g., the translation unit containing X. Thus, this declaration of X::f can be found by traversing the declarations provided by [decls_begin(), decls_end()) in the translation unit.

The semantic context of X::f corresponds to the class X, since this member function is (semantically) a member of X. Lookup of the name f into the DeclContext associated with X will then return the definition of X::f (including information about the default argument).

Transparent Declaration Contexts¶

In C and C++, there are several contexts in which names that are logically declared inside another declaration will actually “leak” out into the enclosing scope from the perspective of name lookup. The most obvious instance of this behavior is in enumeration types, e.g.,

enum Color {
  Red,
  Green,
  Blue
};

Here, Color is an enumeration, which is a declaration context that contains the enumerators Red, Green, and Blue. Thus, traversing the list of declarations contained in the enumeration Color will yield Red, Green, and Blue. However, outside of the scope of Color one can name the enumerator Red without qualifying the name, e.g.,

Color c = Red;

There are other entities in C++ that provide similar behavior. For example, linkage specifications that use curly braces:

extern "C" {
  void f(int);
  void g(int);
}
// f and g are visible here

For source-level accuracy, we treat the linkage specification and enumeration type as a declaration context in which its enclosed declarations (“Red”, “Green”, and “Blue”; “f” and “g”) are declared. However, these declarations are visible outside of the scope of the declaration context.

These language features (and several others, described below) have roughly the same set of requirements: declarations are declared within a particular lexical context, but the declarations are also found via name lookup in scopes enclosing the declaration itself. This feature is implemented via transparent declaration contexts (see DeclContext::isTransparentContext()), whose declarations are visible in the nearest enclosing non-transparent declaration context. This means that the lexical context of the declaration (e.g., an enumerator) will be the transparent DeclContext itself, as will the semantic context, but the declaration will be visible in every outer context up to and including the first non-transparent declaration context (since transparent declaration contexts can be nested).

The transparent DeclContexts are:

• Enumerations (but not C++11 “scoped enumerations”):

  enum Color {
    Red,
    Green,
    Blue
  };
  // Red, Green, and Blue are in scope
  

• C++ linkage specifications:

  extern "C" {
    void f(int);
    void g(int);
  }
  // f and g are in scope
  

• Anonymous unions and structs:

  struct LookupTable {
    bool IsVector;
    union {
      std::vector<Item> *Vector;
      std::set<Item> *Set;
    };
  };
  
  LookupTable LT;
  LT.Vector = 0; // Okay: finds Vector inside the unnamed union
  

• C++11 inline namespaces:

  namespace mylib {
    inline namespace debug {
      class X;
    }
  }
  mylib::X *xp; // okay: mylib::X refers to mylib::debug::X
  

Multiply-Defined Declaration Contexts¶

C++ namespaces have the interesting property that the namespace can be defined multiple times, and the declarations provided by each namespace definition are effectively merged (from the semantic point of view). For example, the following two code snippets are semantically indistinguishable:

// Snippet #1:
namespace N {
  void f();
}
namespace N {
  void f(int);
}

// Snippet #2:
namespace N {
  void f();
  void f(int);
}

In Clang’s representation, the source-centric view of declaration contexts will actually have two separate NamespaceDecl nodes in Snippet #1, each of which is a declaration context that contains a single declaration of “f”. However, the semantics-centric view provided by name lookup into the namespace N for “f” will return a DeclContext::lookup_result that contains a range of iterators over declarations of “f”.

DeclContext manages multiply-defined declaration contexts internally. The function DeclContext::getPrimaryContext retrieves the “primary” context for a given DeclContext instance, which is the DeclContext responsible for maintaining the lookup table used for the semantics-centric view. Given a DeclContext, one can obtain the set of declaration contexts that are semantically connected to this declaration context, in source order, including this context (which will be the only result, for non-namespace contexts) via DeclContext::collectAllContexts. Note that these functions are used internally within the lookup and insertion methods of the DeclContext, so the vast majority of clients can ignore them.

Because the same entity can be defined multiple times in different modules, it is also possible for there to be multiple definitions of (for instance) a CXXRecordDecl, all of which describe a definition of the same class. In such a case, only one of those “definitions” is considered by Clang to be the definiition of the class, and the others are treated as non-defining declarations that happen to also contain member declarations. Corresponding members in each definition of such multiply-defined classes are identified either by redeclaration chains (if the members are Redeclarable) or by simply a pointer to the canonical declaration (if the declarations are not Redeclarable – in that case, a Mergeable base class is used instead).

Abstract Syntax Graph¶

Despite the name, the Clang AST is not a tree. It is a directed graph with cycles. One example of a cycle is the connection between a ClassTemplateDecl and its “templated” CXXRecordDecl. The templated CXXRecordDecl represents all the fields and methods inside the class template, while the ClassTemplateDecl holds the information which is related to being a template, i.e. template arguments, etc. We can get the templated class (the CXXRecordDecl) of a ClassTemplateDecl with ClassTemplateDecl::getTemplatedDecl(). And we can get back a pointer of the “described” class template from the templated class: CXXRecordDecl::getDescribedTemplate(). So, this is a cycle between two nodes: between the templated and the described node. There may be various other kinds of cycles in the AST especially in case of declarations.

Structural Equivalency¶

Importing one AST node copies that node into the destination ASTContext. To copy one node means that we create a new node in the “to” context then we set its properties to be equal to the properties of the source node. Before the copy, we make sure that the source node is not structurally equivalent to any existing node in the destination context. If it happens to be equivalent then we skip the copy.

The informal definition of structural equivalency is the following: Two nodes are structurally equivalent if they are

• builtin types and refer to the same type, e.g. int and int are structurally equivalent,
• function types and all their parameters have structurally equivalent types,
• record types and all their fields in order of their definition have the same identifier names and structurally equivalent types,
• variable or function declarations and they have the same identifier name and their types are structurally equivalent.

In C, two types are structurally equivalent if they are compatible types. For a formal definition of compatible types, please refer to 6.2.7/1 in the C11 standard. However, there is no definition for compatible types in the C++ standard. Still, we extend the definition of structural equivalency to templates and their instantiations similarly: besides checking the previously mentioned properties, we have to check for equivalent template parameters/arguments, etc.

The structural equivalent check can be and is used independently from the ASTImporter, e.g. the clang::Sema class uses it also.

The equivalence of nodes may depend on the equivalency of other pairs of nodes. Thus, the check is implemented as a parallel graph traversal. We traverse through the nodes of both graphs at the same time. The actual implementation is similar to breadth-first-search. Let’s say we start the traverse with the <A,B> pair of nodes. Whenever the traversal reaches a pair <X,Y> then the following statements are true:

• A and X are nodes from the same ASTContext.
• B and Y are nodes from the same ASTContext.
• A and B may or may not be from the same ASTContext.
• if A == X and B == Y (pointer equivalency) then (there is a cycle during the traverse)
  • A and B are structurally equivalent if and only if
    • All dependent nodes on the path from <A,B> to <X,Y> are structurally equivalent.

When we compare two classes or enums and one of them is incomplete or has unloaded external lexical declarations then we cannot descend to compare their contained declarations. So in these cases they are considered equal if they have the same names. This is the way how we compare forward declarations with definitions.

Redeclaration Chains¶

The early version of the ASTImporter’s merge mechanism squashed the declarations, i.e. it aimed to have only one declaration instead of maintaining a whole redeclaration chain. This early approach simply skipped importing a function prototype, but it imported a definition. To demonstrate the problem with this approach let’s consider an empty “to” context and the following virtual function declarations of f in the “from” context:

struct B { virtual void f(); };
void B::f() {} // <-- let's import this definition

If we imported the definition with the “squashing” approach then we would end-up having one declaration which is indeed a definition, but isVirtual() returns false for it. The reason is that the definition is indeed not virtual, it is the property of the prototype!

Consequently, we must either set the virtual flag for the definition (but then we create a malformed AST which the parser would never create), or we import the whole redeclaration chain of the function. The most recent version of the ASTImporter uses the latter mechanism. We do import all function declarations - regardless if they are definitions or prototypes - in the order as they appear in the “from” context.

If we have an existing definition in the “to” context, then we cannot import another definition, we will use the existing definition. However, we can import prototype(s): we chain the newly imported prototype(s) to the existing definition. Whenever we import a new prototype from a third context, that will be added to the end of the redeclaration chain. This may result in long redeclaration chains in certain cases, e.g. if we import from several translation units which include the same header with the prototype.

To mitigate the problem of long redeclaration chains of free functions, we could compare prototypes to see if they have the same properties and if yes then we could merge these prototypes. The implementation of squashing of prototypes for free functions is future work.

Chaining functions this way ensures that we do copy all information from the source AST. Nonetheless, there is a problem with member functions: While we can have many prototypes for free functions, we must have only one prototype for a member function.

void f(); // OK
void f(); // OK

struct X {
  void f(); // OK
  void f(); // ERROR
};
void X::f() {} // OK

Thus, prototypes of member functions must be squashed, we cannot just simply attach a new prototype to the existing in-class prototype. Consider the following contexts:

// "to" context
struct X {
  void f(); // D0
};

// "from" context
struct X {
  void f(); // D1
};
void X::f() {} // D2

When we import the prototype and the definition of f from the “from” context, then the resulting redecl chain will look like this D0 -> D2', where D2' is the copy of D2 in the “to” context.

Generally speaking, when we import declarations (like enums and classes) we do attach the newly imported declaration to the existing redeclaration chain (if there is structural equivalency). We do not import, however, the whole redeclaration chain as we do in case of functions. Up till now, we haven’t found any essential property of forward declarations which is similar to the case of the virtual flag in a member function prototype. In the future, this may change, though.

Traversal during the Import¶

The node specific import mechanisms are implemented in ASTNodeImporter::VisitNode() functions, e.g. VisitFunctionDecl(). When we import a declaration then first we import everything which is needed to call the constructor of that declaration node. Everything which can be set later is set after the node is created. For example, in case of a FunctionDecl we first import the declaration context in which the function is declared, then we create the FunctionDecl and only then we import the body of the function. This means there are implicit dependencies between AST nodes. These dependencies determine the order in which we visit nodes in the “from” context. As with the regular graph traversal algorithms like DFS, we keep track which nodes we have already visited in ASTImporter::ImportedDecls. Whenever we create a node then we immediately add that to the ImportedDecls. We must not start the import of any other declarations before we keep track of the newly created one. This is essential, otherwise, we would not be able to handle circular dependencies. To enforce this, we wrap all constructor calls of all AST nodes in GetImportedOrCreateDecl(). This wrapper ensures that all newly created declarations are immediately marked as imported; also, if a declaration is already marked as imported then we just return its counterpart in the “to” context. Consequently, calling a declaration’s ::Create() function directly would lead to errors, please don’t do that!

Even with the use of GetImportedOrCreateDecl() there is still a probability of having an infinite import recursion if things are imported from each other in wrong way. Imagine that during the import of A, the import of B is requested before we could create the node for A (the constructor needs a reference to B). And the same could be true for the import of B (A is requested to be imported before we could create the node for B). In case of the templated-described swing we take extra attention to break the cyclical dependency: we import and set the described template only after the CXXRecordDecl is created. As a best practice, before creating the node in the “to” context, avoid importing of other nodes which are not needed for the constructor of node A.

Error Handling¶

Every import function returns with either an llvm::Error or an llvm::Expected<T> object. This enforces to check the return value of the import functions. If there was an error during one import then we return with that error. (Exception: when we import the members of a class, we collect the individual errors with each member and we concatenate them in one Error object.) We cache these errors in cases of declarations. During the next import call if there is an existing error we just return with that. So, clients of the library receive an Error object, which they must check.

During import of a specific declaration, it may happen that some AST nodes had already been created before we recognize an error. In this case, we signal back the error to the caller, but the “to” context remains polluted with those nodes which had been created. Ideally, those nodes should not had been created, but that time we did not know about the error, the error happened later. Since the AST is immutable (most of the cases we can’t remove existing nodes) we choose to mark these nodes as erroneous.

We cache the errors associated with declarations in the “from” context in ASTImporter::ImportDeclErrors and the ones which are associated with the “to” context in ASTImporterSharedState::ImportErrors. Note that, there may be several ASTImporter objects which import into the same “to” context but from different “from” contexts; in this case, they have to share the associated errors of the “to” context.

When an error happens, that propagates through the call stack, through all the dependant nodes. However, in case of dependency cycles, this is not enough, because we strive to mark the erroneous nodes so clients can act upon. In those cases, we have to keep track of the errors for those nodes which are intermediate nodes of a cycle.

An import path is the list of the AST nodes which we visit during an Import call. If node A depends on node B then the path contains an A->B edge. From the call stack of the import functions, we can read the very same path.

Now imagine the following AST, where the -> represents dependency in terms of the import (all nodes are declarations).

A->B->C->D
   `->E

We would like to import A. The import behaves like a DFS, so we will visit the nodes in this order: ABCDE. During the visitation we will have the following import paths:

A
AB
ABC
ABCD
ABC
AB
ABE
AB
A

If during the visit of E there is an error then we set an error for E, then as the call stack shrinks for B, then for A:

A
AB
ABC
ABCD
ABC
AB
ABE // Error! Set an error to E
AB  // Set an error to B
A   // Set an error to A

However, during the import we could import C and D without any error and they are independent of A,B and E. We must not set up an error for C and D. So, at the end of the import we have an entry in ImportDeclErrors for A,B,E but not for C,D.

Now, what happens if there is a cycle in the import path? Let’s consider this AST:

A->B->C->A
   `->E

During the visitation, we will have the below import paths and if during the visit of E there is an error then we will set up an error for E,B,A. But what’s up with C?

A
AB
ABC
ABCA
ABC
AB
ABE // Error! Set an error to E
AB  // Set an error to B
A   // Set an error to A

This time we know that both B and C are dependent on A. This means we must set up an error for C too. As the call stack reverses back we get to A and we must set up an error to all nodes which depend on A (this includes C). But C is no longer on the import path, it just had been previously. Such a situation can happen only if during the visitation we had a cycle. If we didn’t have any cycle, then the normal way of passing an Error object through the call stack could handle the situation. This is why we must track cycles during the import process for each visited declaration.

Lookup Problems¶

When we import a declaration from the source context then we check whether we already have a structurally equivalent node with the same name in the “to” context. If the “from” node is a definition and the found one is also a definition, then we do not create a new node, instead, we mark the found node as the imported node. If the found definition and the one we want to import have the same name but they are structurally in-equivalent, then we have an ODR violation in case of C++. If the “from” node is not a definition then we add that to the redeclaration chain of the found node. This behaviour is essential when we merge ASTs from different translation units which include the same header file(s). For example, we want to have only one definition for the class template std::vector, even if we included <vector> in several translation units.

To find a structurally equivalent node we can use the regular C/C++ lookup functions: DeclContext::noload_lookup() and DeclContext::localUncachedLookup(). These functions do respect the C/C++ name hiding rules, thus you cannot find certain declarations in a given declaration context. For instance, unnamed declarations (anonymous structs), non-first friend declarations and template specializations are hidden. This is a problem, because if we use the regular C/C++ lookup then we create redundant AST nodes during the merge! Also, having two instances of the same node could result in false structural in-equivalencies of other nodes which depend on the duplicated node. Because of these reasons, we created a lookup class which has the sole purpose to register all declarations, so later they can be looked up by subsequent import requests. This is the ASTImporterLookupTable class. This lookup table should be shared amongst the different ASTImporter instances if they happen to import to the very same “to” context. This is why we can use the importer specific lookup only via the ASTImporterSharedState class.

Class Template Instantiations¶

Different translation units may have class template instantiations with the same template arguments, but with a different set of instantiated MethodDecls and FieldDecls. Consider the following files:

// x.h
template <typename T>
struct X {
    int a{0}; // FieldDecl with InitListExpr
    X(char) : a(3) {}     // (1)
    X(int) {}             // (2)
};

// foo.cpp
void foo() {
    // ClassTemplateSpec with ctor (1): FieldDecl without InitlistExpr
    X<char> xc('c');
}

// bar.cpp
void bar() {
    // ClassTemplateSpec with ctor (2): FieldDecl WITH InitlistExpr
    X<char> xc(1);
}

In foo.cpp we use the constructor with number (1), which explicitly initializes the member a to 3, thus the InitListExpr {0} is not used here and the AST node is not instantiated. However, in the case of bar.cpp we use the constructor with number (2), which does not explicitly initialize the a member, so the default InitListExpr is needed and thus instantiated. When we merge the AST of foo.cpp and bar.cpp we must create an AST node for the class template instantiation of X<char> which has all the required nodes. Therefore, when we find an existing ClassTemplateSpecializationDecl then we merge the fields of the ClassTemplateSpecializationDecl in the “from” context in a way that the InitListExpr is copied if not existent yet. The same merge mechanism should be done in the cases of instantiated default arguments and exception specifications of functions.

Visibility of Declarations¶

During import of a global variable with external visibility, the lookup will find variables (with the same name) but with static visibility (linkage). Clearly, we cannot put them into the same redeclaration chain. The same is true the in case of functions. Also, we have to take care of other kinds of declarations like enums, classes, etc. if they are in anonymous namespaces. Therefore, we filter the lookup results and consider only those which have the same visibility as the declaration we currently import.

We consider two declarations in two anonymous namespaces to have the same visibility only if they are imported from the same AST context.

Strategies to Handle Conflicting Names¶

During the import we lookup existing declarations with the same name. We filter the lookup results based on their visibility. If any of the found declarations are not structurally equivalent then we bumped to a name conflict error (ODR violation in C++). In this case, we return with an Error and we set up the Error object for the declaration. However, some clients of the ASTImporter may require a different, perhaps less conservative and more liberal error handling strategy.

E.g. static analysis clients may benefit if the node is created even if there is a name conflict. During the CTU analysis of certain projects, we recognized that there are global declarations which collide with declarations from other translation units, but they are not referenced outside from their translation unit. These declarations should be in an unnamed namespace ideally. If we treat these collisions liberally then CTU analysis can find more results. Note, the feature be able to choose between name conflict handling strategies is still an ongoing work.

Basic Blocks¶

Concretely, an instance of CFG is a collection of basic blocks. Each basic block is an instance of CFGBlock, which simply contains an ordered sequence of Stmt* (each referring to statements in the AST). The ordering of statements within a block indicates unconditional flow of control from one statement to the next. Conditional control-flow is represented using edges between basic blocks. The statements within a given CFGBlock can be traversed using the CFGBlock::*iterator interface.

A CFG object owns the instances of CFGBlock within the control-flow graph it represents. Each CFGBlock within a CFG is also uniquely numbered (accessible via CFGBlock::getBlockID()). Currently the number is based on the ordering the blocks were created, but no assumptions should be made on how CFGBlocks are numbered other than their numbers are unique and that they are numbered from 0..N-1 (where N is the number of basic blocks in the CFG).

Entry and Exit Blocks¶

Each instance of CFG contains two special blocks: an entry block (accessible via CFG::getEntry()), which has no incoming edges, and an exit block (accessible via CFG::getExit()), which has no outgoing edges. Neither block contains any statements, and they serve the role of providing a clear entrance and exit for a body of code such as a function body. The presence of these empty blocks greatly simplifies the implementation of many analyses built on top of CFGs.

Conditional Control-Flow¶

Conditional control-flow (such as those induced by if-statements and loops) is represented as edges between CFGBlocks. Because different C language constructs can induce control-flow, each CFGBlock also records an extra Stmt* that represents the terminator of the block. A terminator is simply the statement that caused the control-flow, and is used to identify the nature of the conditional control-flow between blocks. For example, in the case of an if-statement, the terminator refers to the IfStmt object in the AST that represented the given branch.

To illustrate, consider the following code example:

int foo(int x) {
  x = x + 1;
  if (x > 2)
    x++;
  else {
    x += 2;
    x *= 2;
  }

  return x;
}

After invoking the parser+semantic analyzer on this code fragment, the AST of the body of foo is referenced by a single Stmt*. We can then construct an instance of CFG representing the control-flow graph of this function body by single call to a static class method:

Stmt *FooBody = ...
std::unique_ptr<CFG> FooCFG = CFG::buildCFG(FooBody);

Along with providing an interface to iterate over its CFGBlocks, the CFG class also provides methods that are useful for debugging and visualizing CFGs. For example, the method CFG::dump() dumps a pretty-printed version of the CFG to standard error. This is especially useful when one is using a debugger such as gdb. For example, here is the output of FooCFG->dump():

[ B5 (ENTRY) ]
   Predecessors (0):
   Successors (1): B4

[ B4 ]
   1: x = x + 1
   2: (x > 2)
   T: if [B4.2]
   Predecessors (1): B5
   Successors (2): B3 B2

[ B3 ]
   1: x++
   Predecessors (1): B4
   Successors (1): B1

[ B2 ]
   1: x += 2
   2: x *= 2
   Predecessors (1): B4
   Successors (1): B1

[ B1 ]
   1: return x;
   Predecessors (2): B2 B3
   Successors (1): B0

[ B0 (EXIT) ]
   Predecessors (1): B1
   Successors (0):

For each block, the pretty-printed output displays for each block the number of predecessor blocks (blocks that have outgoing control-flow to the given block) and successor blocks (blocks that have control-flow that have incoming control-flow from the given block). We can also clearly see the special entry and exit blocks at the beginning and end of the pretty-printed output. For the entry block (block B5), the number of predecessor blocks is 0, while for the exit block (block B0) the number of successor blocks is 0.

The most interesting block here is B4, whose outgoing control-flow represents the branching caused by the sole if-statement in foo. Of particular interest is the second statement in the block, (x > 2), and the terminator, printed as if [B4.2]. The second statement represents the evaluation of the condition of the if-statement, which occurs before the actual branching of control-flow. Within the CFGBlock for B4, the Stmt* for the second statement refers to the actual expression in the AST for (x > 2). Thus pointers to subclasses of Expr can appear in the list of statements in a block, and not just subclasses of Stmt that refer to proper C statements.

The terminator of block B4 is a pointer to the IfStmt object in the AST. The pretty-printer outputs if [B4.2] because the condition expression of the if-statement has an actual place in the basic block, and thus the terminator is essentially referring to the expression that is the second statement of block B4 (i.e., B4.2). In this manner, conditions for control-flow (which also includes conditions for loops and switch statements) are hoisted into the actual basic block.

Implementation Approach¶

After trying several different approaches, we’ve finally converged on a design (Note, at the time of this writing, not all of this has been implemented, consider this a design goal!). Our basic approach is to define a single recursive evaluation method (Expr::Evaluate), which is implemented in AST/ExprConstant.cpp. Given an expression with “scalar” type (integer, fp, complex, or pointer) this method returns the following information:

• Whether the expression is an integer constant expression, a general constant that was folded but has no side effects, a general constant that was folded but that does have side effects, or an uncomputable/unfoldable value.
• If the expression was computable in any way, this method returns the APValue for the result of the expression.
• If the expression is not evaluatable at all, this method returns information on one of the problems with the expression. This includes a SourceLocation for where the problem is, and a diagnostic ID that explains the problem. The diagnostic should have ERROR type.
• If the expression is not an integer constant expression, this method returns information on one of the problems with the expression. This includes a SourceLocation for where the problem is, and a diagnostic ID that explains the problem. The diagnostic should have EXTENSION type.

This information gives various clients the flexibility that they want, and we will eventually have some helper methods for various extensions. For example, Sema should have a Sema::VerifyIntegerConstantExpression method, which calls Evaluate on the expression. If the expression is not foldable, the error is emitted, and it would return true. If the expression is not an i-c-e, the EXTENSION diagnostic is emitted. Finally it would return false to indicate that the AST is OK.

Other clients can use the information in other ways, for example, codegen can just use expressions that are foldable in any way.

Extensions¶

This section describes how some of the various extensions Clang supports interacts with constant evaluation:

• __extension__: The expression form of this extension causes any evaluatable subexpression to be accepted as an integer constant expression.
• __builtin_constant_p: This returns true (as an integer constant expression) if the operand evaluates to either a numeric value (that is, not a pointer cast to integral type) of integral, enumeration, floating or complex type, or if it evaluates to the address of the first character of a string literal (possibly cast to some other type). As a special case, if __builtin_constant_p is the (potentially parenthesized) condition of a conditional operator expression (“?:”), only the true side of the conditional operator is considered, and it is evaluated with full constant folding.
• __builtin_choose_expr: The condition is required to be an integer constant expression, but we accept any constant as an “extension of an extension”. This only evaluates one operand depending on which way the condition evaluates.
• __builtin_classify_type: This always returns an integer constant expression.
• __builtin_inf, nan, ...: These are treated just like a floating-point literal.
• __builtin_abs, copysign, ...: These are constant folded as general constant expressions.
• __builtin_strlen and strlen: These are constant folded as integer constant expressions if the argument is a string literal.

Attribute Basics¶

Attributes in Clang are handled in three stages: parsing into a parsed attribute representation, conversion from a parsed attribute into a semantic attribute, and then the semantic handling of the attribute.

Parsing of the attribute is determined by the various syntactic forms attributes can take, such as GNU, C++11, and Microsoft style attributes, as well as other information provided by the table definition of the attribute. Ultimately, the parsed representation of an attribute object is an ParsedAttr object. These parsed attributes chain together as a list of parsed attributes attached to a declarator or declaration specifier. The parsing of attributes is handled automatically by Clang, except for attributes spelled as keywords. When implementing a keyword attribute, the parsing of the keyword and creation of the ParsedAttr object must be done manually.

Eventually, Sema::ProcessDeclAttributeList() is called with a Decl and an ParsedAttr, at which point the parsed attribute can be transformed into a semantic attribute. The process by which a parsed attribute is converted into a semantic attribute depends on the attribute definition and semantic requirements of the attribute. The end result, however, is that the semantic attribute object is attached to the Decl object, and can be obtained by a call to Decl::getAttr<T>().

The structure of the semantic attribute is also governed by the attribute definition given in Attr.td. This definition is used to automatically generate functionality used for the implementation of the attribute, such as a class derived from clang::Attr, information for the parser to use, automated semantic checking for some attributes, etc.

include/clang/Basic/Attr.td¶

The first step to adding a new attribute to Clang is to add its definition to include/clang/Basic/Attr.td. This tablegen definition must derive from the Attr (tablegen, not semantic) type, or one of its derivatives. Most attributes will derive from the InheritableAttr type, which specifies that the attribute can be inherited by later redeclarations of the Decl it is associated with. InheritableParamAttr is similar to InheritableAttr, except that the attribute is written on a parameter instead of a declaration. If the attribute is intended to apply to a type instead of a declaration, such an attribute should derive from TypeAttr, and will generally not be given an AST representation. (Note that this document does not cover the creation of type attributes.) An attribute that inherits from IgnoredAttr is parsed, but will generate an ignored attribute diagnostic when used, which may be useful when an attribute is supported by another vendor but not supported by clang.

The definition will specify several key pieces of information, such as the semantic name of the attribute, the spellings the attribute supports, the arguments the attribute expects, and more. Most members of the Attr tablegen type do not require definitions in the derived definition as the default suffice. However, every attribute must specify at least a spelling list, a subject list, and a documentation list.

clang-tblgen -gen-attr-docs -I /path/to/clang/include /path/to/clang/include/clang/Basic/Attr.td -o /path/to/clang/docs/AttributeReference.rst

Boilerplate¶

All semantic processing of declaration attributes happens in lib/Sema/SemaDeclAttr.cpp, and generally starts in the ProcessDeclAttribute() function. If the attribute is a “simple” attribute – meaning that it requires no custom semantic processing aside from what is automatically provided, add a call to handleSimpleAttribute<YourAttr>(S, D, Attr); to the switch statement. Otherwise, write a new handleYourAttr() function, and add that to the switch statement. Please do not implement handling logic directly in the case for the attribute.

Unless otherwise specified by the attribute definition, common semantic checking of the parsed attribute is handled automatically. This includes diagnosing parsed attributes that do not appertain to the given Decl, ensuring the correct minimum number of arguments are passed, etc.

If the attribute adds additional warnings, define a DiagGroup in include/clang/Basic/DiagnosticGroups.td named after the attribute’s Spelling with “_”s replaced by “-“s. If there is only a single diagnostic, it is permissible to use InGroup<DiagGroup<"your-attribute">> directly in DiagnosticSemaKinds.td

All semantic diagnostics generated for your attribute, including automatically- generated ones (such as subjects and argument counts), should have a corresponding test case.

Semantic handling¶

Most attributes are implemented to have some effect on the compiler. For instance, to modify the way code is generated, or to add extra semantic checks for an analysis pass, etc. Having added the attribute definition and conversion to the semantic representation for the attribute, what remains is to implement the custom logic requiring use of the attribute.

The clang::Decl object can be queried for the presence or absence of an attribute using hasAttr<T>(). To obtain a pointer to the semantic representation of the attribute, getAttr<T> may be used.