=============================== TableGen Programmer's Reference =============================== .. sectnum:: .. contents:: :local: Introduction ============ The purpose of TableGen is to generate complex output files based on information from source files that are significantly easier to code than the output files would be, and also easier to maintain and modify over time. The information is coded in a declarative style involving classes and records, which are then processed by TableGen. The internalized records are passed on to various *backends*, which extract information from a subset of the records and generate one or more output files. These output files are typically ``.inc`` files for C++, but may be any type of file that the backend developer needs. This document describes the LLVM TableGen facility in detail. It is intended for the programmer who is using TableGen to produce code for a project. If you are looking for a simple overview, check out the :doc:`TableGen Overview <./index>`. The various ``*-tblgen`` commands used to invoke TableGen are described in :doc:`tblgen Family - Description to C++ Code<../CommandGuide/tblgen>`. An example of a backend is ``RegisterInfo``, which generates the register file information for a particular target machine, for use by the LLVM target-independent code generator. See :doc:`TableGen Backends <./BackEnds>` for a description of the LLVM TableGen backends, and :doc:`TableGen Backend Developer's Guide <./BackGuide>` for a guide to writing a new backend. Here are a few of the things backends can do. * Generate the register file information for a particular target machine. * Generate the instruction definitions for a target. * Generate the patterns that the code generator uses to match instructions to intermediate representation (IR) nodes. * Generate semantic attribute identifiers for Clang. * Generate abstract syntax tree (AST) declaration node definitions for Clang. * Generate AST statement node definitions for Clang. Concepts -------- TableGen source files contain two primary items: *abstract records* and *concrete records*. In this and other TableGen documents, abstract records are called *classes.* (These classes are different from C++ classes and do not map onto them.) In addition, concrete records are usually just called records, although sometimes the term *record* refers to both classes and concrete records. The distinction should be clear in context. Classes and concrete records have a unique *name*, either chosen by the programmer or generated by TableGen. Associated with that name is a list of *fields* with values and an optional list of *parent classes* (sometimes called base or super classes). The fields are the primary data that backends will process. Note that TableGen assigns no meanings to fields; the meanings are entirely up to the backends and the programs that incorporate the output of those backends. .. note:: The term "parent class" can refer to a class that is a parent of another class, and also to a class from which a concrete record inherits. This nonstandard use of the term arises because TableGen treats classes and concrete records similarly. A backend processes some subset of the concrete records built by the TableGen parser and emits the output files. These files are usually C++ ``.inc`` files that are included by the programs that require the data in those records. However, a backend can produce any type of output files. For example, it could produce a data file containing messages tagged with identifiers and substitution parameters. In a complex use case such as the LLVM code generator, there can be many concrete records and some of them can have an unexpectedly large number of fields, resulting in large output files. In order to reduce the complexity of TableGen files, classes are used to abstract out groups of record fields. For example, a few classes may abstract the concept of a machine register file, while other classes may abstract the instruction formats, and still others may abstract the individual instructions. TableGen allows an arbitrary hierarchy of classes, so that the abstract classes for two concepts can share a third superclass that abstracts common "sub-concepts" from the two original concepts. In order to make classes more useful, a concrete record (or another class) can request a class as a parent class and pass *template arguments* to it. These template arguments can be used in the fields of the parent class to initialize them in a custom manner. That is, record or class ``A`` can request parent class ``S`` with one set of template arguments, while record or class ``B`` can request ``S`` with a different set of arguments. Without template arguments, many more classes would be required, one for each combination of the template arguments. Both classes and concrete records can include fields that are uninitialized. The uninitialized "value" is represented by a question mark (``?``). Classes often have uninitialized fields that are expected to be filled in when those classes are inherited by concrete records. Even so, some fields of concrete records may remain uninitialized. TableGen provides *multiclasses* to collect a group of record definitions in one place. A multiclass is a sort of macro that can be "invoked" to define multiple concrete records all at once. A multiclass can inherit from other multiclasses, which means that the multiclass inherits all the definitions from its parent multiclasses. `Appendix C: Sample Record`_ illustrates a complex record in the Intel X86 target and the simple way in which it is defined. Source Files ============ TableGen source files are plain ASCII text files. The files can contain statements, comments, and blank lines (see `Lexical Analysis`_). The standard file extension for TableGen files is ``.td``. TableGen files can grow quite large, so there is an include mechanism that allows one file to include the content of another file (see `Include Files`_). This allows large files to be broken up into smaller ones, and also provides a simple library mechanism where multiple source files can include the same library file. TableGen supports a simple preprocessor that can be used to conditionalize portions of ``.td`` files. See `Preprocessing Facilities`_ for more information. Lexical Analysis ================ The lexical and syntax notation used here is intended to imitate `Python's`_ notation. In particular, for lexical definitions, the productions operate at the character level and there is no implied whitespace between elements. The syntax definitions operate at the token level, so there is implied whitespace between tokens. .. _`Python's`: http://docs.python.org/py3k/reference/introduction.html#notation TableGen supports BCPL-style comments (``// ...``) and nestable C-style comments (``/* ... */``). TableGen also provides simple `Preprocessing Facilities`_. Formfeed characters may be used freely in files to produce page breaks when the file is printed for review. The following are the basic punctuation tokens:: - + [ ] { } ( ) < > : ; . ... = ? # Literals -------- Numeric literals take one of the following forms: .. productionlist:: TokInteger: `DecimalInteger` | `HexInteger` | `BinInteger` DecimalInteger: ["+" | "-"] ("0"..."9")+ HexInteger: "0x" ("0"..."9" | "a"..."f" | "A"..."F")+ BinInteger: "0b" ("0" | "1")+ Observe that the :token:`DecimalInteger` token includes the optional ``+`` or ``-`` sign, unlike most languages where the sign would be treated as a unary operator. TableGen has two kinds of string literals: .. productionlist:: TokString: '"' (non-'"' characters and escapes) '"' TokCode: "[{" (shortest text not containing "}]") "}]" A :token:`TokCode` is nothing more than a multi-line string literal delimited by ``[{`` and ``}]``. It can break across lines and the line breaks are retained in the string. The current implementation accepts the following escape sequences:: \\ \' \" \t \n Identifiers ----------- TableGen has name- and identifier-like tokens, which are case-sensitive. .. productionlist:: ualpha: "a"..."z" | "A"..."Z" | "_" TokIdentifier: ("0"..."9")* `ualpha` (`ualpha` | "0"..."9")* TokVarName: "$" `ualpha` (`ualpha` | "0"..."9")* Note that, unlike most languages, TableGen allows :token:`TokIdentifier` to begin with an integer. In case of ambiguity, a token is interpreted as a numeric literal rather than an identifier. TableGen has the following reserved keywords, which cannot be used as identifiers:: assert bit bits class code dag def else false foreach defm defset defvar field if in include int let list multiclass string then true .. warning:: The ``field`` reserved word is deprecated. Bang operators -------------- TableGen provides "bang operators" that have a wide variety of uses: .. productionlist:: BangOperator: one of : !add !and !cast !con !dag : !empty !eq !filter !find !foldl : !foreach !ge !getdagop !gt !head : !if !interleave !isa !le !listconcat : !listsplat !lt !mul !ne !not : !or !setdagop !shl !size !sra : !srl !strconcat !sub !subst !substr : !tail !xor The ``!cond`` operator has a slightly different syntax compared to other bang operators, so it is defined separately: .. productionlist:: CondOperator: !cond See `Appendix A: Bang Operators`_ for a description of each bang operator. Include files ------------- TableGen has an include mechanism. The content of the included file lexically replaces the ``include`` directive and is then parsed as if it was originally in the main file. .. productionlist:: IncludeDirective: "include" `TokString` Portions of the main file and included files can be conditionalized using preprocessor directives. .. productionlist:: PreprocessorDirective: "#define" | "#ifdef" | "#ifndef" Types ===== The TableGen language is statically typed, using a simple but complete type system. Types are used to check for errors, to perform implicit conversions, and to help interface designers constrain the allowed input. Every value is required to have an associated type. TableGen supports a mixture of low-level types (e.g., ``bit``) and high-level types (e.g., ``dag``). This flexibility allows you to describe a wide range of records conveniently and compactly. .. productionlist:: Type: "bit" | "int" | "string" | "dag" :| "bits" "<" `TokInteger` ">" :| "list" "<" `Type` ">" :| `ClassID` ClassID: `TokIdentifier` ``bit`` A ``bit`` is a boolean value that can be 0 or 1. ``int`` The ``int`` type represents a simple 64-bit integer value, such as 5 or -42. ``string`` The ``string`` type represents an ordered sequence of characters of arbitrary length. ``bits<``\ *n*\ ``>`` The ``bits`` type is a fixed-sized integer of arbitrary length *n* that is treated as separate bits. These bits can be accessed individually. A field of this type is useful for representing an instruction operation code, register number, or address mode/register/displacement. The bits of the field can be set individually or as subfields. For example, in an instruction address, the addressing mode, base register number, and displacement can be set separately. ``list<``\ *type*\ ``>`` This type represents a list whose elements are of the *type* specified in angle brackets. The element type is arbitrary; it can even be another list type. List elements are indexed from 0. ``dag`` This type represents a nestable directed acyclic graph (DAG) of nodes. Each node has an *operator* and zero or more *arguments* (or *operands*). An argument can be another ``dag`` object, allowing an arbitrary tree of nodes and edges. As an example, DAGs are used to represent code patterns for use by the code generator instruction selection algorithms. See `Directed acyclic graphs (DAGs)`_ for more details; :token:`ClassID` Specifying a class name in a type context indicates that the type of the defined value must be a subclass of the specified class. This is useful in conjunction with the ``list`` type; for example, to constrain the elements of the list to a common base class (e.g., a ``list`` can only contain definitions derived from the ``Register`` class). The :token:`ClassID` must name a class that has been previously declared or defined. Values and Expressions ====================== There are many contexts in TableGen statements where a value is required. A common example is in the definition of a record, where each field is specified by a name and an optional value. TableGen allows for a reasonable number of different forms when building up value expressions. These forms allow the TableGen file to be written in a syntax that is natural for the application. Note that all of the values have rules for converting them from one type to another. For example, these rules allow you to assign a value like ``7`` to an entity of type ``bits<4>``. .. productionlist:: Value: `SimpleValue` `ValueSuffix`* :| `Value` "#" `Value` ValueSuffix: "{" `RangeList` "}" :| "[" `RangeList` "]" :| "." `TokIdentifier` RangeList: `RangePiece` ("," `RangePiece`)* RangePiece: `TokInteger` :| `TokInteger` "..." `TokInteger` :| `TokInteger` "-" `TokInteger` :| `TokInteger` `TokInteger` .. warning:: The peculiar last form of :token:`RangePiece` is due to the fact that the "``-``" is included in the :token:`TokInteger`, hence ``1-5`` gets lexed as two consecutive tokens, with values ``1`` and ``-5``, instead of "1", "-", and "5". The use of hyphen as the range punctuation is deprecated. Simple values ------------- The :token:`SimpleValue` has a number of forms. .. productionlist:: SimpleValue: `TokInteger` | `TokString`+ | `TokCode` A value can be an integer literal, a string literal, or a code literal. Multiple adjacent string literals are concatenated as in C/C++; the simple value is the concatenation of the strings. Code literals become strings and are then indistinguishable from them. .. productionlist:: SimpleValue2: "true" | "false" The ``true`` and ``false`` literals are essentially syntactic sugar for the integer values 1 and 0. They improve the readability of TableGen files when boolean values are used in field initializations, bit sequences, ``if`` statements, etc. When parsed, these literals are converted to integers. .. note:: Although ``true`` and ``false`` are literal names for 1 and 0, we recommend as a stylistic rule that you use them for boolean values only. .. productionlist:: SimpleValue3: "?" A question mark represents an uninitialized value. .. productionlist:: SimpleValue4: "{" [`ValueList`] "}" ValueList: `ValueListNE` ValueListNE: `Value` ("," `Value`)* This value represents a sequence of bits, which can be used to initialize a ``bits<``\ *n*\ ``>`` field (note the braces). When doing so, the values must represent a total of *n* bits. .. productionlist:: SimpleValue5: "[" `ValueList` "]" ["<" `Type` ">"] This value is a list initializer (note the brackets). The values in brackets are the elements of the list. The optional :token:`Type` can be used to indicate a specific element type; otherwise the element type is inferred from the given values. TableGen can usually infer the type, although sometimes not when the value is the empty list (``[]``). .. productionlist:: SimpleValue6: "(" `DagArg` [`DagArgList`] ")" DagArgList: `DagArg` ("," `DagArg`)* DagArg: `Value` [":" `TokVarName`] | `TokVarName` This represents a DAG initializer (note the parentheses). The first :token:`DagArg` is called the "operator" of the DAG and must be a record. See `Directed acyclic graphs (DAGs)`_ for more details. .. productionlist:: SimpleValue7: `TokIdentifier` The resulting value is the value of the entity named by the identifier. The possible identifiers are described here, but the descriptions will make more sense after reading the remainder of this guide. .. The code for this is exceptionally abstruse. These examples are a best-effort attempt. * A template argument of a ``class``, such as the use of ``Bar`` in:: class Foo { int Baz = Bar; } * The implicit template argument ``NAME`` in a ``class`` or ``multiclass`` definition (see `NAME`_). * A field local to a ``class``, such as the use of ``Bar`` in:: class Foo { int Bar = 5; int Baz = Bar; } * The name of a record definition, such as the use of ``Bar`` in the definition of ``Foo``:: def Bar : SomeClass { int X = 5; } def Foo { SomeClass Baz = Bar; } * A field local to a record definition, such as the use of ``Bar`` in:: def Foo { int Bar = 5; int Baz = Bar; } Fields inherited from the record's parent classes can be accessed the same way. * A template argument of a ``multiclass``, such as the use of ``Bar`` in:: multiclass Foo { def : SomeClass; } * A variable defined with the ``defvar`` or ``defset`` statements. * The iteration variable of a ``foreach``, such as the use of ``i`` in:: foreach i = 0...5 in def Foo#i; .. productionlist:: SimpleValue8: `ClassID` "<" `ValueListNE` ">" This form creates a new anonymous record definition (as would be created by an unnamed ``def`` inheriting from the given class with the given template arguments; see `def`_) and the value is that record. A field of the record can be obtained using a suffix; see `Suffixed Values`_. Invoking a class in this manner can provide a simple subroutine facility. See `Using Classes as Subroutines`_ for more information. .. productionlist:: SimpleValue9: `BangOperator` ["<" `Type` ">"] "(" `ValueListNE` ")" :| `CondOperator` "(" `CondClause` ("," `CondClause`)* ")" CondClause: `Value` ":" `Value` The bang operators provide functions that are not available with the other simple values. Except in the case of ``!cond``, a bang operator takes a list of arguments enclosed in parentheses and performs some function on those arguments, producing a value for that bang operator. The ``!cond`` operator takes a list of pairs of arguments separated by colons. See `Appendix A: Bang Operators`_ for a description of each bang operator. Suffixed values --------------- The :token:`SimpleValue` values described above can be specified with certain suffixes. The purpose of a suffix is to obtain a subvalue of the primary value. Here are the possible suffixes for some primary *value*. *value*\ ``{17}`` The final value is bit 17 of the integer *value* (note the braces). *value*\ ``{8...15}`` The final value is bits 8--15 of the integer *value*. The order of the bits can be reversed by specifying ``{15...8}``. *value*\ ``[4]`` The final value is element 4 of the list *value* (note the brackets). In other words, the brackets act as a subscripting operator on the list. This is the case only when a single element is specified. *value*\ ``[4...7,17,2...3,4]`` The final value is a new list that is a slice of the list *value*. The new list contains elements 4, 5, 6, 7, 17, 2, 3, and 4. Elements may be included multiple times and in any order. This is the result only when more than one element is specified. *value*\ ``.``\ *field* The final value is the value of the specified *field* in the specified record *value*. The paste operator ------------------ The paste operator (``#``) is the only infix operator available in TableGen expressions. It allows you to concatenate strings or lists, but has a few unusual features. The paste operator can be used when specifying the record name in a :token:`Def` or :token:`Defm` statement, in which case it must construct a string. If an operand is an undefined name (:token:`TokIdentifier`) or the name of a global :token:`Defvar` or :token:`Defset`, it is treated as a verbatim string of characters. The value of a global name is not used. The paste operator can be used in all other value expressions, in which case it can construct a string or a list. Rather oddly, but consistent with the previous case, if the *right-hand-side* operand is an undefined name or a global name, it is treated as a verbatim string of characters. The left-hand-side operand is treated normally. `Appendix B: Paste Operator Examples`_ presents examples of the behavior of the paste operator. Statements ========== The following statements may appear at the top level of TableGen source files. .. productionlist:: TableGenFile: `Statement`* Statement: `Assert` | `Class` | `Def` | `Defm` | `Defset` | `Defvar` :| `Foreach` | `If` | `Let` | `MultiClass` The following sections describe each of these top-level statements. ``class`` --- define an abstract record class --------------------------------------------- A ``class`` statement defines an abstract record class from which other classes and records can inherit. .. productionlist:: Class: "class" `ClassID` [`TemplateArgList`] `RecordBody` TemplateArgList: "<" `TemplateArgDecl` ("," `TemplateArgDecl`)* ">" TemplateArgDecl: `Type` `TokIdentifier` ["=" `Value`] A class can be parameterized by a list of "template arguments," whose values can be used in the class's record body. These template arguments are specified each time the class is inherited by another class or record. If a template argument is not assigned a default value with ``=``, it is uninitialized (has the "value" ``?``) and must be specified in the template argument list when the class is inherited (required argument). If an argument is assigned a default value, then it need not be specified in the argument list (optional argument). In the declaration, all required template arguments must precede any optional arguments. The template argument default values are evaluated from left to right. The :token:`RecordBody` is defined below. It can include a list of parent classes from which the current class inherits, along with field definitions and other statements. When a class ``C`` inherits from another class ``D``, the fields of ``D`` are effectively merged into the fields of ``C``. A given class can only be defined once. A ``class`` statement is considered to define the class if *any* of the following are true (the :token:`RecordBody` elements are described below). * The :token:`TemplateArgList` is present, or * The :token:`ParentClassList` in the :token:`RecordBody` is present, or * The :token:`Body` in the :token:`RecordBody` is present and not empty. You can declare an empty class by specifying an empty :token:`TemplateArgList` and an empty :token:`RecordBody`. This can serve as a restricted form of forward declaration. Note that records derived from a forward-declared class will inherit no fields from it, because those records are built when their declarations are parsed, and thus before the class is finally defined. .. _NAME: Every class has an implicit template argument named ``NAME`` (uppercase), which is bound to the name of the :token:`Def` or :token:`Defm` inheriting from the class. If the class is inherited by an anonymous record, the name is unspecified but globally unique. See `Examples: classes and records`_ for examples. Record Bodies ````````````` Record bodies appear in both class and record definitions. A record body can include a parent class list, which specifies the classes from which the current class or record inherits fields. Such classes are called the parent classes of the class or record. The record body also includes the main body of the definition, which contains the specification of the fields of the class or record. .. productionlist:: RecordBody: `ParentClassList` `Body` ParentClassList: [":" `ParentClassListNE`] ParentClassListNE: `ClassRef` ("," `ClassRef`)* ClassRef: (`ClassID` | `MultiClassID`) ["<" [`ValueList`] ">"] A :token:`ParentClassList` containing a :token:`MultiClassID` is valid only in the class list of a ``defm`` statement. In that case, the ID must be the name of a multiclass. .. productionlist:: Body: ";" | "{" `BodyItem`* "}" BodyItem: (`Type` | "code") `TokIdentifier` ["=" `Value`] ";" :| "let" `TokIdentifier` ["{" `RangeList` "}"] "=" `Value` ";" :| "defvar" `TokIdentifier` "=" `Value` ";" :| `Assert` A field definition in the body specifies a field to be included in the class or record. If no initial value is specified, then the field's value is uninitialized. The type must be specified; TableGen will not infer it from the value. The keyword ``code`` may be used to emphasize that the field has a string value that is code. The ``let`` form is used to reset a field to a new value. This can be done for fields defined directly in the body or fields inherited from parent classes. A :token:`RangeList` can be specified to reset certain bits in a ``bit`` field. The ``defvar`` form defines a variable whose value can be used in other value expressions within the body. The variable is not a field: it does not become a field of the class or record being defined. Variables are provided to hold temporary values while processing the body. See `Defvar in a Record Body`_ for more details. When class ``C2`` inherits from class ``C1``, it acquires all the field definitions of ``C1``. As those definitions are merged into class ``C2``, any template arguments passed to ``C1`` by ``C2`` are substituted into the definitions. In other words, the abstract record fields defined by ``C1`` are expanded with the template arguments before being merged into ``C2``. .. _def: ``def`` --- define a concrete record ------------------------------------ A ``def`` statement defines a new concrete record. .. productionlist:: Def: "def" [`NameValue`] `RecordBody` NameValue: `Value` (parsed in a special mode) The name value is optional. If specified, it is parsed in a special mode where undefined (unrecognized) identifiers are interpreted as literal strings. In particular, global identifiers are considered unrecognized. These include global variables defined by ``defvar`` and ``defset``. A record name can be the null string. If no name value is given, the record is *anonymous*. The final name of an anonymous record is unspecified but globally unique. Special handling occurs if a ``def`` appears inside a ``multiclass`` statement. See the ``multiclass`` section below for details. A record can inherit from one or more classes by specifying the :token:`ParentClassList` clause at the beginning of its record body. All of the fields in the parent classes are added to the record. If two or more parent classes provide the same field, the record ends up with the field value of the last parent class. As a special case, the name of a record can be passed as a template argument to that record's parent classes. For example: .. code-block:: text class A { dag the_dag = d; } def rec1 : A<(ops rec1)> The DAG ``(ops rec1)`` is passed as a template argument to class ``A``. Notice that the DAG includes ``rec1``, the record being defined. The steps taken to create a new record are somewhat complex. See `How records are built`_. See `Examples: classes and records`_ for examples. Examples: classes and records ----------------------------- Here is a simple TableGen file with one class and two record definitions. .. code-block:: text class C { bit V = true; } def X : C; def Y : C { let V = false; string Greeting = "Hello!"; } First, the abstract class ``C`` is defined. It has one field named ``V`` that is a bit initialized to true. Next, two records are defined, derived from class ``C``; that is, with ``C`` as their parent class. Thus they both inherit the ``V`` field. Record ``Y`` also defines another string field, ``Greeting``, which is initialized to ``"Hello!"``. In addition, ``Y`` overrides the inherited ``V`` field, setting it to false. A class is useful for isolating the common features of multiple records in one place. A class can initialize common fields to default values, but records inheriting from that class can override the defaults. TableGen supports the definition of parameterized classes as well as nonparameterized ones. Parameterized classes specify a list of variable declarations, which may optionally have defaults, that are bound when the class is specified as a parent class of another class or record. .. code-block:: text class FPFormat val> { bits<3> Value = val; } def NotFP : FPFormat<0>; def ZeroArgFP : FPFormat<1>; def OneArgFP : FPFormat<2>; def OneArgFPRW : FPFormat<3>; def TwoArgFP : FPFormat<4>; def CompareFP : FPFormat<5>; def CondMovFP : FPFormat<6>; def SpecialFP : FPFormat<7>; The purpose of the ``FPFormat`` class is to act as a sort of enumerated type. It provides a single field, ``Value``, which holds a 3-bit number. Its template argument, ``val``, is used to set the ``Value`` field. Each of the eight records is defined with ``FPFormat`` as its parent class. The enumeration value is passed in angle brackets as the template argument. Each record will inherent the ``Value`` field with the appropriate enumeration value. Here is a more complex example of classes with template arguments. First, we define a class similar to the ``FPFormat`` class above. It takes a template argument and uses it to initialize a field named ``Value``. Then we define four records that inherit the ``Value`` field with its four different integer values. .. code-block:: text class ModRefVal val> { bits<2> Value = val; } def None : ModRefVal<0>; def Mod : ModRefVal<1>; def Ref : ModRefVal<2>; def ModRef : ModRefVal<3>; This is somewhat contrived, but let's say we would like to examine the two bits of the ``Value`` field independently. We can define a class that accepts a ``ModRefVal`` record as a template argument and splits up its value into two fields, one bit each. Then we can define records that inherit from ``ModRefBits`` and so acquire two fields from it, one for each bit in the ``ModRefVal`` record passed as the template argument. .. code-block:: text class ModRefBits { // Break the value up into its bits, which can provide a nice // interface to the ModRefVal values. bit isMod = mrv.Value{0}; bit isRef = mrv.Value{1}; } // Example uses. def foo : ModRefBits; def bar : ModRefBits; def snork : ModRefBits; This illustrates how one class can be defined to reorganize the fields in another class, thus hiding the internal representation of that other class. Running ``llvm-tblgen`` on the example prints the following definitions: .. code-block:: text def bar { // Value bit isMod = 0; bit isRef = 1; } def foo { // Value bit isMod = 1; bit isRef = 0; } def snork { // Value bit isMod = 1; bit isRef = 1; } ``let`` --- override fields in classes or records ------------------------------------------------- A ``let`` statement collects a set of field values (sometimes called *bindings*) and applies them to all the classes and records defined by statements within the scope of the ``let``. .. productionlist:: Let: "let" `LetList` "in" "{" `Statement`* "}" :| "let" `LetList` "in" `Statement` LetList: `LetItem` ("," `LetItem`)* LetItem: `TokIdentifier` ["<" `RangeList` ">"] "=" `Value` The ``let`` statement establishes a scope, which is a sequence of statements in braces or a single statement with no braces. The bindings in the :token:`LetList` apply to the statements in that scope. The field names in the :token:`LetList` must name fields in classes inherited by the classes and records defined in the statements. The field values are applied to the classes and records *after* the records inherit all the fields from their parent classes. So the ``let`` acts to override inherited field values. A ``let`` cannot override the value of a template argument. Top-level ``let`` statements are often useful when a few fields need to be overriden in several records. Here are two examples. Note that ``let`` statements can be nested. .. code-block:: text let isTerminator = true, isReturn = true, isBarrier = true, hasCtrlDep = true in def RET : I<0xC3, RawFrm, (outs), (ins), "ret", [(X86retflag 0)]>; let isCall = true in // All calls clobber the non-callee saved registers... let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6, ST0, MM0, MM1, MM2, MM3, MM4, MM5, MM6, MM7, XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7, EFLAGS] in { def CALLpcrel32 : Ii32<0xE8, RawFrm, (outs), (ins i32imm:$dst, variable_ops), "call\t${dst:call}", []>; def CALL32r : I<0xFF, MRM2r, (outs), (ins GR32:$dst, variable_ops), "call\t{*}$dst", [(X86call GR32:$dst)]>; def CALL32m : I<0xFF, MRM2m, (outs), (ins i32mem:$dst, variable_ops), "call\t{*}$dst", []>; } Note that a top-level ``let`` will not override fields defined in the classes or records themselves. ``multiclass`` --- define multiple records ------------------------------------------ While classes with template arguments are a good way to factor out commonality between multiple records, multiclasses allow a convenient method for defining many records at once. For example, consider a 3-address instruction architecture whose instructions come in two formats: ``reg = reg op reg`` and ``reg = reg op imm`` (e.g., SPARC). We would like to specify in one place that these two common formats exist, then in a separate place specify what all the operations are. The ``multiclass`` and ``defm`` statements accomplish this goal. You can think of a multiclass as a macro or template that expands into multiple records. .. productionlist:: MultiClass: "multiclass" `TokIdentifier` [`TemplateArgList`] : [":" `ParentMultiClassList`] : "{" `MultiClassStatement`+ "}" ParentMultiClassList: `MultiClassID` ("," `MultiClassID`)* MultiClassID: `TokIdentifier` MultiClassStatement: `Assert` | `Def` | `Defm` | `Defvar` | `Foreach` | `If` | `Let` As with regular classes, the multiclass has a name and can accept template arguments. A multiclass can inherit from other multiclasses, which causes the other multiclasses to be expanded and contribute to the record definitions in the inheriting multiclass. The body of the multiclass contains a series of statements that define records, using :token:`Def` and :token:`Defm`. In addition, :token:`Defvar`, :token:`Foreach`, and :token:`Let` statements can be used to factor out even more common elements. The :token:`If` and :token:`Assert` statements can also be used. Also as with regular classes, the multiclass has the implicit template argument ``NAME`` (see NAME_). When a named (non-anonymous) record is defined in a multiclass and the record's name does not include a use of the template argument ``NAME``, such a use is automatically *prepended* to the name. That is, the following are equivalent inside a multiclass:: def Foo ... def NAME # Foo ... The records defined in a multiclass are created when the multiclass is "instantiated" or "invoked" by a ``defm`` statement outside the multiclass definition. Each ``def`` statement in the multiclass produces a record. As with top-level ``def`` statements, these definitions can inherit from multiple parent classes. See `Examples: multiclasses and defms`_ for examples. ``defm`` --- invoke multiclasses to define multiple records ----------------------------------------------------------- Once multiclasses have been defined, you use the ``defm`` statement to "invoke" them and process the multiple record definitions in those multiclasses. Those record definitions are specified by ``def`` statements in the multiclasses, and indirectly by ``defm`` statements. .. productionlist:: Defm: "defm" [`NameValue`] `ParentClassList` ";" The optional :token:`NameValue` is formed in the same way as the name of a ``def``. The :token:`ParentClassList` is a colon followed by a list of at least one multiclass and any number of regular classes. The multiclasses must precede the regular classes. Note that the ``defm`` does not have a body. This statement instantiates all the records defined in all the specified multiclasses, either directly by ``def`` statements or indirectly by ``defm`` statements. These records also receive the fields defined in any regular classes included in the parent class list. This is useful for adding a common set of fields to all the records created by the ``defm``. The name is parsed in the same special mode used by ``def``. If the name is not included, an unspecified but globally unique name is provided. That is, the following examples end up with different names:: defm : SomeMultiClass<...>; // A globally unique name. defm "" : SomeMultiClass<...>; // An empty name. The ``defm`` statement can be used in a multiclass body. When this occurs, the second variant is equivalent to:: defm NAME : SomeMultiClass<...>; More generally, when ``defm`` occurs in a multiclass and its name does not include a use of the implicit template argument ``NAME``, then ``NAME`` will be prepended automatically. That is, the following are equivalent inside a multiclass:: defm Foo : SomeMultiClass<...>; defm NAME # Foo : SomeMultiClass<...>; See `Examples: multiclasses and defms`_ for examples. Examples: multiclasses and defms -------------------------------- Here is a simple example using ``multiclass`` and ``defm``. Consider a 3-address instruction architecture whose instructions come in two formats: ``reg = reg op reg`` and ``reg = reg op imm`` (immediate). The SPARC is an example of such an architecture. .. code-block:: text def ops; def GPR; def Imm; class inst ; multiclass ri_inst { def _rr : inst; def _ri : inst; } // Define records for each instruction in the RR and RI formats. defm ADD : ri_inst<0b111, "add">; defm SUB : ri_inst<0b101, "sub">; defm MUL : ri_inst<0b100, "mul">; Each use of the ``ri_inst`` multiclass defines two records, one with the ``_rr`` suffix and one with ``_ri``. Recall that the name of the ``defm`` that uses a multiclass is prepended to the names of the records defined in that multiclass. So the resulting definitions are named:: ADD_rr, ADD_ri SUB_rr, SUB_ri MUL_rr, MUL_ri Without the ``multiclass`` feature, the instructions would have to be defined as follows. .. code-block:: text def ops; def GPR; def Imm; class inst ; class rrinst : inst; class riinst : inst; // Define records for each instruction in the RR and RI formats. def ADD_rr : rrinst<0b111, "add">; def ADD_ri : riinst<0b111, "add">; def SUB_rr : rrinst<0b101, "sub">; def SUB_ri : riinst<0b101, "sub">; def MUL_rr : rrinst<0b100, "mul">; def MUL_ri : riinst<0b100, "mul">; A ``defm`` can be used in a multiclass to "invoke" other multiclasses and create the records defined in those multiclasses in addition to the records defined in the current multiclass. In the following example, the ``basic_s`` and ``basic_p`` multiclasses contain ``defm`` statements that refer to the ``basic_r`` multiclass. The ``basic_r`` multiclass contains only ``def`` statements. .. code-block:: text class Instruction opc, string Name> { bits<4> opcode = opc; string name = Name; } multiclass basic_r opc> { def rr : Instruction; def rm : Instruction; } multiclass basic_s opc> { defm SS : basic_r; defm SD : basic_r; def X : Instruction; } multiclass basic_p opc> { defm PS : basic_r; defm PD : basic_r; def Y : Instruction; } defm ADD : basic_s<0xf>, basic_p<0xf>; The final ``defm`` creates the following records, five from the ``basic_s`` multiclass and five from the ``basic_p`` multiclass:: ADDSSrr, ADDSSrm ADDSDrr, ADDSDrm ADDX ADDPSrr, ADDPSrm ADDPDrr, ADDPDrm ADDY A ``defm`` statement, both at top level and in a multiclass, can inherit from regular classes in addition to multiclasses. The rule is that the regular classes must be listed after the multiclasses, and there must be at least one multiclass. .. code-block:: text class XD { bits<4> Prefix = 11; } class XS { bits<4> Prefix = 12; } class I op> { bits<4> opcode = op; } multiclass R { def rr : I<4>; def rm : I<2>; } multiclass Y { defm SS : R, XD; // First multiclass R, then regular class XD. defm SD : R, XS; } defm Instr : Y; This example will create four records, shown here in alphabetical order with their fields. .. code-block:: text def InstrSDrm { bits<4> opcode = { 0, 0, 1, 0 }; bits<4> Prefix = { 1, 1, 0, 0 }; } def InstrSDrr { bits<4> opcode = { 0, 1, 0, 0 }; bits<4> Prefix = { 1, 1, 0, 0 }; } def InstrSSrm { bits<4> opcode = { 0, 0, 1, 0 }; bits<4> Prefix = { 1, 0, 1, 1 }; } def InstrSSrr { bits<4> opcode = { 0, 1, 0, 0 }; bits<4> Prefix = { 1, 0, 1, 1 }; } It's also possible to use ``let`` statements inside multiclasses, providing another way to factor out commonality from the records, especially when using several levels of multiclass instantiations. .. code-block:: text multiclass basic_r opc> { let Predicates = [HasSSE2] in { def rr : Instruction; def rm : Instruction; } let Predicates = [HasSSE3] in def rx : Instruction; } multiclass basic_ss opc> { let IsDouble = false in defm SS : basic_r; let IsDouble = true in defm SD : basic_r; } defm ADD : basic_ss<0xf>; ``defset`` --- create a definition set -------------------------------------- The ``defset`` statement is used to collect a set of records into a global list of records. .. productionlist:: Defset: "defset" `Type` `TokIdentifier` "=" "{" `Statement`* "}" All records defined inside the braces via ``def`` and ``defm`` are defined as usual, and they are also collected in a global list of the given name (:token:`TokIdentifier`). The specified type must be ``list<``\ *class*\ ``>``, where *class* is some record class. The ``defset`` statement establishes a scope for its statements. It is an error to define a record in the scope of the ``defset`` that is not of type *class*. The ``defset`` statement can be nested. The inner ``defset`` adds the records to its own set, and all those records are also added to the outer set. Anonymous records created inside initialization expressions using the ``ClassID<...>`` syntax are not collected in the set. ``defvar`` --- define a variable -------------------------------- A ``defvar`` statement defines a global variable. Its value can be used throughout the statements that follow the definition. .. productionlist:: Defvar: "defvar" `TokIdentifier` "=" `Value` ";" The identifier on the left of the ``=`` is defined to be a global variable whose value is given by the value expression on the right of the ``=``. The type of the variable is automatically inferred. Once a variable has been defined, it cannot be set to another value. Variables defined in a top-level ``foreach`` go out of scope at the end of each loop iteration, so their value in one iteration is not available in the next iteration. The following ``defvar`` will not work:: defvar i = !add(i, 1) Variables can also be defined with ``defvar`` in a record body. See `Defvar in a Record Body`_ for more details. ``foreach`` --- iterate over a sequence of statements ----------------------------------------------------- The ``foreach`` statement iterates over a series of statements, varying a variable over a sequence of values. .. productionlist:: Foreach: "foreach" `ForeachIterator` "in" "{" `Statement`* "}" :| "foreach" `ForeachIterator` "in" `Statement` ForeachIterator: `TokIdentifier` "=" ("{" `RangeList` "}" | `RangePiece` | `Value`) The body of the ``foreach`` is a series of statements in braces or a single statement with no braces. The statements are re-evaluated once for each value in the range list, range piece, or single value. On each iteration, the :token:`TokIdentifier` variable is set to the value and can be used in the statements. The statement list establishes an inner scope. Variables local to a ``foreach`` go out of scope at the end of each loop iteration, so their values do not carry over from one iteration to the next. Foreach loops may be nested. The ``foreach`` statement can also be used in a record :token:`Body`. .. Note that the productions involving RangeList and RangePiece have precedence over the more generic value parsing based on the first token. .. code-block:: text foreach i = [0, 1, 2, 3] in { def R#i : Register<...>; def F#i : Register<...>; } This loop defines records named ``R0``, ``R1``, ``R2``, and ``R3``, along with ``F0``, ``F1``, ``F2``, and ``F3``. ``if`` --- select statements based on a test -------------------------------------------- The ``if`` statement allows one of two statement groups to be selected based on the value of an expression. .. productionlist:: If: "if" `Value` "then" `IfBody` :| "if" `Value` "then" `IfBody` "else" `IfBody` IfBody: "{" `Statement`* "}" | `Statement` The value expression is evaluated. If it evaluates to true (in the same sense used by the bang operators), then the statements following the ``then`` reserved word are processed. Otherwise, if there is an ``else`` reserved word, the statements following the ``else`` are processed. If the value is false and there is no ``else`` arm, no statements are processed. Because the braces around the ``then`` statements are optional, this grammar rule has the usual ambiguity with "dangling else" clauses, and it is resolved in the usual way: in a case like ``if v1 then if v2 then {...} else {...}``, the ``else`` associates with the inner ``if`` rather than the outer one. The :token:`IfBody` of the then and else arms of the ``if`` establish an inner scope. Any ``defvar`` variables defined in the bodies go out of scope when the bodies are finished (see `Defvar in a Record Body`_ for more details). The ``if`` statement can also be used in a record :token:`Body`. ``assert`` --- check that a condition is true --------------------------------------------- The ``assert`` statement checks a boolean condition to be sure that it is true and prints an error message if it is not. .. productionlist:: Assert: "assert" `condition` "," `message` ";" If the boolean condition is true, the statement does nothing. If the condition is false, it prints a nonfatal error message. The **message**, which can be an arbitrary string expression, is included in the error message as a note. The exact behavior of the ``assert`` statement depends on its placement. * At top level, the assertion is checked immediately. * In a record definition, the statement is saved and all assertions are checked after the record is completely built. * In a class definition, the assertions are saved and inherited by all the subclasses and records that inherit from the class. The assertions are then checked when the records are completely built. * In a multiclass definition, the assertions are saved with the other components of the multiclass and then checked each time the multiclass is instantiated with ``defm``. Using assertions in TableGen files can simplify record checking in TableGen backends. Here is an example of an ``assert`` in two class definitions. .. code-block:: text class PersonName { assert !le(!size(name), 32), "person name is too long: " # name; string Name = name; } class Person : PersonName { assert !and(!ge(age, 1), !le(age, 120)), "person age is invalid: " # age; int Age = age; } def Rec20 : Person<"Donald Knuth", 60> { ... } Additional Details ================== Directed acyclic graphs (DAGs) ------------------------------ A directed acyclic graph can be represented directly in TableGen using the ``dag`` datatype. A DAG node consists of an operator and zero or more arguments (or operands). Each argument can be of any desired type. By using another DAG node as an argument, an arbitrary graph of DAG nodes can be built. The syntax of a ``dag`` instance is: ``(`` *operator* *argument1*\ ``,`` *argument2*\ ``,`` ... ``)`` The operator must be present and must be a record. There can be zero or more arguments, separated by commas. The operator and arguments can have three formats. ====================== ============================================= Format Meaning ====================== ============================================= *value* argument value *value*\ ``:``\ *name* argument value and associated name *name* argument name with unset (uninitialized) value ====================== ============================================= The *value* can be any TableGen value. The *name*, if present, must be a :token:`TokVarName`, which starts with a dollar sign (``$``). The purpose of a name is to tag an operator or argument in a DAG with a particular meaning, or to associate an argument in one DAG with a like-named argument in another DAG. The following bang operators are useful for working with DAGs: ``!con``, ``!dag``, ``!empty``, ``!foreach``, ``!getdagop``, ``!setdagop``, ``!size``. Defvar in a record body ----------------------- In addition to defining global variables, the ``defvar`` statement can be used inside the :token:`Body` of a class or record definition to define local variables. The scope of the variable extends from the ``defvar`` statement to the end of the body. It cannot be set to a different value within its scope. The ``defvar`` statement can also be used in the statement list of a ``foreach``, which establishes a scope. A variable named ``V`` in an inner scope shadows (hides) any variables ``V`` in outer scopes. In particular, ``V`` in a record body shadows a global ``V``, and ``V`` in a ``foreach`` statement list shadows any ``V`` in surrounding record or global scopes. Variables defined in a ``foreach`` go out of scope at the end of each loop iteration, so their value in one iteration is not available in the next iteration. The following ``defvar`` will not work:: defvar i = !add(i, 1) How records are built --------------------- The following steps are taken by TableGen when a record is built. Classes are simply abstract records and so go through the same steps. 1. Build the record name (:token:`NameValue`) and create an empty record. 2. Parse the parent classes in the :token:`ParentClassList` from left to right, visiting each parent class's ancestor classes from top to bottom. a. Add the fields from the parent class to the record. b. Substitute the template arguments into those fields. c. Add the parent class to the record's list of inherited classes. 3. Apply any top-level ``let`` bindings to the record. Recall that top-level bindings only apply to inherited fields. 4. Parse the body of the record. * Add any fields to the record. * Modify the values of fields according to local ``let`` statements. * Define any ``defvar`` variables. 5. Make a pass over all the fields to resolve any inter-field references. 6. Add the record to the master record list. Because references between fields are resolved (step 5) after ``let`` bindings are applied (step 3), the ``let`` statement has unusual power. For example: .. code-block:: text class C { int Y = x; int Yplus1 = !add(Y, 1); int xplus1 = !add(x, 1); } let Y = 10 in { def rec1 : C<5> { } } def rec2 : C<5> { let Y = 10; } In both cases, one where a top-level ``let`` is used to bind ``Y`` and one where a local ``let`` does the same thing, the results are: .. code-block:: text def rec1 { // C int Y = 10; int Yplus1 = 11; int xplus1 = 6; } def rec2 { // C int Y = 10; int Yplus1 = 11; int xplus1 = 6; } ``Yplus1`` is 11 because the ``let Y`` is performed before the ``!add(Y, 1)`` is resolved. Use this power wisely. Using Classes as Subroutines ============================ As described in `Simple values`_, a class can be invoked in an expression and passed template arguments. This causes TableGen to create a new anonymous record inheriting from that class. As usual, the record receives all the fields defined in the class. This feature can be employed as a simple subroutine facility. The class can use the template arguments to define various variables and fields, which end up in the anonymous record. Those fields can then be retrieved in the expression invoking the class as follows. Assume that the field ``ret`` contains the final value of the subroutine. .. code-block:: text int Result = ... CalcValue.ret ...; The ``CalcValue`` class is invoked with the template argument ``arg``. It calculates a value for the ``ret`` field, which is then retrieved at the "point of call" in the initialization for the Result field. The anonymous record created in this example serves no other purpose than to carry the result value. Here is a practical example. The class ``isValidSize`` determines whether a specified number of bytes represents a valid data size. The bit ``ret`` is set appropriately. The field ``ValidSize`` obtains its initial value by invoking ``isValidSize`` with the data size and retrieving the ``ret`` field from the resulting anonymous record. .. code-block:: text class isValidSize { bit ret = !cond(!eq(size, 1): 1, !eq(size, 2): 1, !eq(size, 4): 1, !eq(size, 8): 1, !eq(size, 16): 1, true: 0); } def Data1 { int Size = ...; bit ValidSize = isValidSize.ret; } Preprocessing Facilities ======================== The preprocessor embedded in TableGen is intended only for simple conditional compilation. It supports the following directives, which are specified somewhat informally. .. productionlist:: LineBegin: beginning of line LineEnd: newline | return | EOF WhiteSpace: space | tab CComment: "/*" ... "*/" BCPLComment: "//" ... `LineEnd` WhiteSpaceOrCComment: `WhiteSpace` | `CComment` WhiteSpaceOrAnyComment: `WhiteSpace` | `CComment` | `BCPLComment` MacroName: `ualpha` (`ualpha` | "0"..."9")* PreDefine: `LineBegin` (`WhiteSpaceOrCComment`)* : "#define" (`WhiteSpace`)+ `MacroName` : (`WhiteSpaceOrAnyComment`)* `LineEnd` PreIfdef: `LineBegin` (`WhiteSpaceOrCComment`)* : ("#ifdef" | "#ifndef") (`WhiteSpace`)+ `MacroName` : (`WhiteSpaceOrAnyComment`)* `LineEnd` PreElse: `LineBegin` (`WhiteSpaceOrCComment`)* : "#else" (`WhiteSpaceOrAnyComment`)* `LineEnd` PreEndif: `LineBegin` (`WhiteSpaceOrCComment`)* : "#endif" (`WhiteSpaceOrAnyComment`)* `LineEnd` .. PreRegContentException: `PreIfdef` | `PreElse` | `PreEndif` | EOF PreRegion: .* - `PreRegContentException` :| `PreIfdef` : (`PreRegion`)* : [`PreElse`] : (`PreRegion`)* : `PreEndif` A :token:`MacroName` can be defined anywhere in a TableGen file. The name has no value; it can only be tested to see whether it is defined. A macro test region begins with an ``#ifdef`` or ``#ifndef`` directive. If the macro name is defined (``#ifdef``) or undefined (``#ifndef``), then the source code between the directive and the corresponding ``#else`` or ``#endif`` is processed. If the test fails but there is an ``#else`` clause, the source code between the ``#else`` and the ``#endif`` is processed. If the test fails and there is no ``#else`` clause, then no source code in the test region is processed. Test regions may be nested, but they must be properly nested. A region started in a file must end in that file; that is, must have its ``#endif`` in the same file. A :token:`MacroName` may be defined externally using the ``-D`` option on the ``*-tblgen`` command line:: llvm-tblgen self-reference.td -Dmacro1 -Dmacro3 Appendix A: Bang Operators ========================== Bang operators act as functions in value expressions. A bang operator takes one or more arguments, operates on them, and produces a result. If the operator produces a boolean result, the result value will be 1 for true or 0 for false. When an operator tests a boolean argument, it interprets 0 as false and non-0 as true. .. warning:: The ``!getop`` and ``!setop`` bang operators are deprecated in favor of ``!getdagop`` and ``!setdagop``. ``!add(``\ *a*\ ``,`` *b*\ ``, ...)`` This operator adds *a*, *b*, etc., and produces the sum. ``!and(``\ *a*\ ``,`` *b*\ ``, ...)`` This operator does a bitwise AND on *a*, *b*, etc., and produces the result. A logical AND can be performed if all the arguments are either 0 or 1. ``!cast<``\ *type*\ ``>(``\ *a*\ ``)`` This operator performs a cast on *a* and produces the result. If *a* is not a string, then a straightforward cast is performed, say between an ``int`` and a ``bit``, or between record types. This allows casting a record to a class. If a record is cast to ``string``, the record's name is produced. If *a* is a string, then it is treated as a record name and looked up in the list of all defined records. The resulting record is expected to be of the specified *type*. For example, if ``!cast<``\ *type*\ ``>(``\ *name*\ ``)`` appears in a multiclass definition, or in a class instantiated inside a multiclass definition, and the *name* does not reference any template arguments of the multiclass, then a record by that name must have been instantiated earlier in the source file. If *name* does reference a template argument, then the lookup is delayed until ``defm`` statements instantiating the multiclass (or later, if the defm occurs in another multiclass and template arguments of the inner multiclass that are referenced by *name* are substituted by values that themselves contain references to template arguments of the outer multiclass). If the type of *a* does not match *type*, TableGen raises an error. ``!con(``\ *a*\ ``,`` *b*\ ``, ...)`` This operator concatenates the DAG nodes *a*, *b*, etc. Their operations must equal. ``!con((op a1:$name1, a2:$name2), (op b1:$name3))`` results in the DAG node ``(op a1:$name1, a2:$name2, b1:$name3)``. ``!cond(``\ *cond1* ``:`` *val1*\ ``,`` *cond2* ``:`` *val2*\ ``, ...,`` *condn* ``:`` *valn*\ ``)`` This operator tests *cond1* and returns *val1* if the result is true. If false, the operator tests *cond2* and returns *val2* if the result is true. And so forth. An error is reported if no conditions are true. This example produces the sign word for an integer:: !cond(!lt(x, 0) : "negative", !eq(x, 0) : "zero", true : "positive") ``!dag(``\ *op*\ ``,`` *arguments*\ ``,`` *names*\ ``)`` This operator creates a DAG node with the given operator and arguments. The *arguments* and *names* arguments must be lists of equal length or uninitialized (``?``). The *names* argument must be of type ``list``. Due to limitations of the type system, *arguments* must be a list of items of a common type. In practice, this means that they should either have the same type or be records with a common parent class. Mixing ``dag`` and non-``dag`` items is not possible. However, ``?`` can be used. Example: ``!dag(op, [a1, a2, ?], ["name1", "name2", "name3"])`` results in ``(op a1-value:$name1, a2-value:$name2, ?:$name3)``. ``!empty(``\ *a*\ ``)`` This operator produces 1 if the string, list, or DAG *a* is empty; 0 otherwise. A dag is empty if it has no arguments; the operator does not count. ``!eq(`` *a*\ `,` *b*\ ``)`` This operator produces 1 if *a* is equal to *b*; 0 otherwise. The arguments must be ``bit``, ``bits``, ``int``, ``string``, or record values. Use ``!cast`` to compare other types of objects. ``!filter(``\ *var*\ ``,`` *list*\ ``,`` *predicate*\ ``)`` This operator creates a new ``list`` by filtering the elements in *list*. To perform the filtering, TableGen binds the variable *var* to each element and then evaluates the *predicate* expression, which presumably refers to *var*. The predicate must produce a boolean value (``bit``, ``bits``, or ``int``). The value is interpreted as with ``!if``: if the value is 0, the element is not included in the new list. If the value is anything else, the element is included. ``!find(``\ *string1*\ ``,`` *string2*\ [``,`` *start*]\ ``)`` This operator searches for *string2* in *string1* and produces its position. The starting position of the search may be specified by *start*, which can range between 0 and the length of *string1*; the default is 0. If the string is not found, the result is -1. ``!foldl(``\ *init*\ ``,`` *list*\ ``,`` *acc*\ ``,`` *var*\ ``,`` *expr*\ ``)`` This operator performs a left-fold over the items in *list*. The variable *acc* acts as the accumulator and is initialized to *init*. The variable *var* is bound to each element in the *list*. The expression is evaluated for each element and presumably uses *acc* and *var* to calculate the accumulated value, which ``!foldl`` stores back in *acc*. The type of *acc* is the same as *init*; the type of *var* is the same as the elements of *list*; *expr* must have the same type as *init*. The following example computes the total of the ``Number`` field in the list of records in ``RecList``:: int x = !foldl(0, RecList, total, rec, !add(total, rec.Number)); If your goal is to filter the list and produce a new list that includes only some of the elements, see ``!filter``. ``!foreach(``\ *var*\ ``,`` *sequence*\ ``,`` *expr*\ ``)`` This operator creates a new ``list``/``dag`` in which each element is a function of the corresponding element in the *sequence* ``list``/``dag``. To perform the function, TableGen binds the variable *var* to an element and then evaluates the expression. The expression presumably refers to the variable *var* and calculates the result value. If you simply want to create a list of a certain length containing the same value repeated multiple times, see ``!listsplat``. ``!ge(``\ *a*\ `,` *b*\ ``)`` This operator produces 1 if *a* is greater than or equal to *b*; 0 otherwise. The arguments must be ``bit``, ``bits``, ``int``, or ``string`` values. ``!getdagop(``\ *dag*\ ``)`` --or-- ``!getdagop<``\ *type*\ ``>(``\ *dag*\ ``)`` This operator produces the operator of the given *dag* node. Example: ``!getdagop((foo 1, 2))`` results in ``foo``. Recall that DAG operators are always records. The result of ``!getdagop`` can be used directly in a context where any record class at all is acceptable (typically placing it into another dag value). But in other contexts, it must be explicitly cast to a particular class. The ``<``\ *type*\ ``>`` syntax is provided to make this easy. For example, to assign the result to a value of type ``BaseClass``, you could write either of these:: BaseClass b = !getdagop(someDag); BaseClass b = !cast(!getdagop(someDag)); But to create a new DAG node that reuses the operator from another, no cast is necessary:: dag d = !dag(!getdagop(someDag), args, names); ``!gt(``\ *a*\ `,` *b*\ ``)`` This operator produces 1 if *a* is greater than *b*; 0 otherwise. The arguments must be ``bit``, ``bits``, ``int``, or ``string`` values. ``!head(``\ *a*\ ``)`` This operator produces the zeroth element of the list *a*. (See also ``!tail``.) ``!if(``\ *test*\ ``,`` *then*\ ``,`` *else*\ ``)`` This operator evaluates the *test*, which must produce a ``bit`` or ``int``. If the result is not 0, the *then* expression is produced; otherwise the *else* expression is produced. ``!interleave(``\ *list*\ ``,`` *delim*\ ``)`` This operator concatenates the items in the *list*, interleaving the *delim* string between each pair, and produces the resulting string. The list can be a list of string, int, bits, or bit. An empty list results in an empty string. The delimiter can be the empty string. ``!isa<``\ *type*\ ``>(``\ *a*\ ``)`` This operator produces 1 if the type of *a* is a subtype of the given *type*; 0 otherwise. ``!le(``\ *a*\ ``,`` *b*\ ``)`` This operator produces 1 if *a* is less than or equal to *b*; 0 otherwise. The arguments must be ``bit``, ``bits``, ``int``, or ``string`` values. ``!listconcat(``\ *list1*\ ``,`` *list2*\ ``, ...)`` This operator concatenates the list arguments *list1*, *list2*, etc., and produces the resulting list. The lists must have the same element type. ``!listsplat(``\ *value*\ ``,`` *count*\ ``)`` This operator produces a list of length *count* whose elements are all equal to the *value*. For example, ``!listsplat(42, 3)`` results in ``[42, 42, 42]``. ``!lt(``\ *a*\ `,` *b*\ ``)`` This operator produces 1 if *a* is less than *b*; 0 otherwise. The arguments must be ``bit``, ``bits``, ``int``, or ``string`` values. ``!mul(``\ *a*\ ``,`` *b*\ ``, ...)`` This operator multiplies *a*, *b*, etc., and produces the product. ``!ne(``\ *a*\ `,` *b*\ ``)`` This operator produces 1 if *a* is not equal to *b*; 0 otherwise. The arguments must be ``bit``, ``bits``, ``int``, ``string``, or record values. Use ``!cast`` to compare other types of objects. ``!not(``\ *a*\ ``)`` This operator performs a logical NOT on *a*, which must be an integer. The argument 0 results in 1 (true); any other argument results in 0 (false). ``!or(``\ *a*\ ``,`` *b*\ ``, ...)`` This operator does a bitwise OR on *a*, *b*, etc., and produces the result. A logical OR can be performed if all the arguments are either 0 or 1. ``!setdagop(``\ *dag*\ ``,`` *op*\ ``)`` This operator produces a DAG node with the same arguments as *dag*, but with its operator replaced with *op*. Example: ``!setdagop((foo 1, 2), bar)`` results in ``(bar 1, 2)``. ``!shl(``\ *a*\ ``,`` *count*\ ``)`` This operator shifts *a* left logically by *count* bits and produces the resulting value. The operation is performed on a 64-bit integer; the result is undefined for shift counts outside 0...63. ``!size(``\ *a*\ ``)`` This operator produces the size of the string, list, or dag *a*. The size of a DAG is the number of arguments; the operator does not count. ``!sra(``\ *a*\ ``,`` *count*\ ``)`` This operator shifts *a* right arithmetically by *count* bits and produces the resulting value. The operation is performed on a 64-bit integer; the result is undefined for shift counts outside 0...63. ``!srl(``\ *a*\ ``,`` *count*\ ``)`` This operator shifts *a* right logically by *count* bits and produces the resulting value. The operation is performed on a 64-bit integer; the result is undefined for shift counts outside 0...63. ``!strconcat(``\ *str1*\ ``,`` *str2*\ ``, ...)`` This operator concatenates the string arguments *str1*, *str2*, etc., and produces the resulting string. ``!sub(``\ *a*\ ``,`` *b*\ ``)`` This operator subtracts *b* from *a* and produces the arithmetic difference. ``!subst(``\ *target*\ ``,`` *repl*\ ``,`` *value*\ ``)`` This operator replaces all occurrences of the *target* in the *value* with the *repl* and produces the resulting value. The *value* can be a string, in which case substring substitution is performed. The *value* can be a record name, in which case the operator produces the *repl* record if the *target* record name equals the *value* record name; otherwise it produces the *value*. ``!substr(``\ *string*\ ``,`` *start*\ [``,`` *length*]\ ``)`` This operator extracts a substring of the given *string*. The starting position of the substring is specified by *start*, which can range between 0 and the length of the string. The length of the substring is specified by *length*; if not specified, the rest of the string is extracted. The *start* and *length* arguments must be integers. ``!tail(``\ *a*\ ``)`` This operator produces a new list with all the elements of the list *a* except for the zeroth one. (See also ``!head``.) ``!xor(``\ *a*\ ``,`` *b*\ ``, ...)`` This operator does a bitwise EXCLUSIVE OR on *a*, *b*, etc., and produces the result. A logical XOR can be performed if all the arguments are either 0 or 1. Appendix B: Paste Operator Examples =================================== Here is an example illustrating the use of the paste operator in record names. .. code-block:: text defvar suffix = "_suffstring"; defvar some_ints = [0, 1, 2, 3]; def name # suffix { } foreach i = [1, 2] in { def rec # i { } } The first ``def`` does not use the value of the ``suffix`` variable. The second def does use the value of the ``i`` iterator variable, because it is not a global name. The following records are produced. .. code-block:: text def namesuffix { } def rec1 { } def rec2 { } Here is a second example illustrating the paste operator in field value expressions. .. code-block:: text def test { string strings = suffix # suffix; list integers = some_ints # [4, 5, 6]; } The ``strings`` field expression uses ``suffix`` on both sides of the paste operator. It is evaluated normally on the left hand side, but taken verbatim on the right hand side. The ``integers`` field expression uses the value of the ``some_ints`` variable and a literal list. The following record is produced. .. code-block:: text def test { string strings = "_suffstringsuffix"; list ints = [0, 1, 2, 3, 4, 5, 6]; } Appendix C: Sample Record ========================= One target machine supported by LLVM is the Intel x86. The following output from TableGen shows the record that is created to represent the 32-bit register-to-register ADD instruction. .. code-block:: text def ADD32rr { // InstructionEncoding Instruction X86Inst I ITy Sched BinOpRR BinOpRR_RF int Size = 0; string DecoderNamespace = ""; list Predicates = []; string DecoderMethod = ""; bit hasCompleteDecoder = 1; string Namespace = "X86"; dag OutOperandList = (outs GR32:$dst); dag InOperandList = (ins GR32:$src1, GR32:$src2); string AsmString = "add{l} {$src2, $src1|$src1, $src2}"; EncodingByHwMode EncodingInfos = ?; list Pattern = [(set GR32:$dst, EFLAGS, (X86add_flag GR32:$src1, GR32:$src2))]; list Uses = []; list Defs = [EFLAGS]; int CodeSize = 3; int AddedComplexity = 0; bit isPreISelOpcode = 0; bit isReturn = 0; bit isBranch = 0; bit isEHScopeReturn = 0; bit isIndirectBranch = 0; bit isCompare = 0; bit isMoveImm = 0; bit isMoveReg = 0; bit isBitcast = 0; bit isSelect = 0; bit isBarrier = 0; bit isCall = 0; bit isAdd = 0; bit isTrap = 0; bit canFoldAsLoad = 0; bit mayLoad = ?; bit mayStore = ?; bit mayRaiseFPException = 0; bit isConvertibleToThreeAddress = 1; bit isCommutable = 1; bit isTerminator = 0; bit isReMaterializable = 0; bit isPredicable = 0; bit isUnpredicable = 0; bit hasDelaySlot = 0; bit usesCustomInserter = 0; bit hasPostISelHook = 0; bit hasCtrlDep = 0; bit isNotDuplicable = 0; bit isConvergent = 0; bit isAuthenticated = 0; bit isAsCheapAsAMove = 0; bit hasExtraSrcRegAllocReq = 0; bit hasExtraDefRegAllocReq = 0; bit isRegSequence = 0; bit isPseudo = 0; bit isExtractSubreg = 0; bit isInsertSubreg = 0; bit variadicOpsAreDefs = 0; bit hasSideEffects = ?; bit isCodeGenOnly = 0; bit isAsmParserOnly = 0; bit hasNoSchedulingInfo = 0; InstrItinClass Itinerary = NoItinerary; list SchedRW = [WriteALU]; string Constraints = "$src1 = $dst"; string DisableEncoding = ""; string PostEncoderMethod = ""; bits<64> TSFlags = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0 }; string AsmMatchConverter = ""; string TwoOperandAliasConstraint = ""; string AsmVariantName = ""; bit UseNamedOperandTable = 0; bit FastISelShouldIgnore = 0; bits<8> Opcode = { 0, 0, 0, 0, 0, 0, 0, 1 }; Format Form = MRMDestReg; bits<7> FormBits = { 0, 1, 0, 1, 0, 0, 0 }; ImmType ImmT = NoImm; bit ForceDisassemble = 0; OperandSize OpSize = OpSize32; bits<2> OpSizeBits = { 1, 0 }; AddressSize AdSize = AdSizeX; bits<2> AdSizeBits = { 0, 0 }; Prefix OpPrefix = NoPrfx; bits<3> OpPrefixBits = { 0, 0, 0 }; Map OpMap = OB; bits<3> OpMapBits = { 0, 0, 0 }; bit hasREX_WPrefix = 0; FPFormat FPForm = NotFP; bit hasLockPrefix = 0; Domain ExeDomain = GenericDomain; bit hasREPPrefix = 0; Encoding OpEnc = EncNormal; bits<2> OpEncBits = { 0, 0 }; bit HasVEX_W = 0; bit IgnoresVEX_W = 0; bit EVEX_W1_VEX_W0 = 0; bit hasVEX_4V = 0; bit hasVEX_L = 0; bit ignoresVEX_L = 0; bit hasEVEX_K = 0; bit hasEVEX_Z = 0; bit hasEVEX_L2 = 0; bit hasEVEX_B = 0; bits<3> CD8_Form = { 0, 0, 0 }; int CD8_EltSize = 0; bit hasEVEX_RC = 0; bit hasNoTrackPrefix = 0; bits<7> VectSize = { 0, 0, 1, 0, 0, 0, 0 }; bits<7> CD8_Scale = { 0, 0, 0, 0, 0, 0, 0 }; string FoldGenRegForm = ?; string EVEX2VEXOverride = ?; bit isMemoryFoldable = 1; bit notEVEX2VEXConvertible = 0; } On the first line of the record, you can see that the ``ADD32rr`` record inherited from eight classes. Although the inheritance hierarchy is complex, using parent classes is much simpler than specifying the 109 individual fields for each instruction. Here is the code fragment used to define ``ADD32rr`` and multiple other ``ADD`` instructions: .. code-block:: text defm ADD : ArithBinOp_RF<0x00, 0x02, 0x04, "add", MRM0r, MRM0m, X86add_flag, add, 1, 1, 1>; The ``defm`` statement tells TableGen that ``ArithBinOp_RF`` is a multiclass, which contains multiple concrete record definitions that inherit from ``BinOpRR_RF``. That class, in turn, inherits from ``BinOpRR``, which inherits from ``ITy`` and ``Sched``, and so forth. The fields are inherited from all the parent classes; for example, ``IsIndirectBranch`` is inherited from the ``Instruction`` class.