TableGen Fundamentals

Written by Chris Lattner

Introduction

TableGen's purpose is to help a human develop and maintain records of domain-specific information. Because there may be a large number of these records, it is specifically designed to allow writing flexible descriptions and for common features of these records to be factored out. This reduces the amount of duplication in the description, reduces the chance of error, and makes it easier to structure domain specific information.

The core part of TableGen parses a file, instantiates the declarations, and hands the result off to a domain-specific "TableGen backend" for processing. The current major user of TableGen is the LLVM code generator.

Note that if you work on TableGen much, and use emacs or vim, that you can find an emacs "TableGen mode" and a vim language file in llvm/utils/emacs and llvm/utils/vim directory of your LLVM distribution, respectively.

Basic concepts

TableGen files consist of two key parts: 'classes' and 'definitions', both of which are considered 'records'.

TableGen records have a unique name, a list of values, and a list of superclasses. The list of values is main data that TableGen builds for each record, it is this that holds the domain specific information for the application. The interpretation of this data is left to a specific TableGen backend, but the structure and format rules are taken care of and fixed by TableGen.

TableGen definitions are the concrete form of 'records'. These generally do not have any undefined values, and are marked with the 'def' keyword.

TableGen classes are abstract records that are used to build and describe other records. These 'classes' allow the end-user to build abstractions for either the domain they are targetting (such as "Register", "RegisterClass", and "Instruction" in the LLVM code generator) or for the implementor to help factor out common properties of records (such as "FPInst", which is used to represent floating point instructions in the X86 backend). TableGen keeps track of all of the classes that are used to build up a definition, so the backend can find all definitions of a particular class, such as "Instruction".

TableGen multiclasses are groups of abstract records that are instantiated all at once. Each instantiation can result in multiple TableGen definitions.

An example record

With no other arguments, TableGen parses the specified file and prints out all of the classes, then all of the definitions. This is a good way to see what the various definitions expand to fully. Running this on the X86.td file prints this (at the time of this writing):

...
def ADDrr8 {    // Instruction X86Inst I2A8 Pattern
  string Name = "add";
  string Namespace = "X86";
  list<Register> Uses = [];
  list<Register> Defs = [];
  bit isReturn = 0;
  bit isBranch = 0;
  bit isCall = 0;
  bit isTwoAddress = 1;
  bit isTerminator = 0;
  dag Pattern = (set R8, (plus R8, R8));
  bits<8> Opcode = { 0, 0, 0, 0, 0, 0, 0, 0 };
  Format Form = MRMDestReg;
  bits<5> FormBits = { 0, 0, 0, 1, 1 };
  ArgType Type = Arg8;
  bits<3> TypeBits = { 0, 0, 1 };
  bit hasOpSizePrefix = 0;
  bit printImplicitUses = 0;
  bits<4> Prefix = { 0, 0, 0, 0 };
  FPFormat FPForm = ?;
  bits<3> FPFormBits = { 0, 0, 0 };
}
...

This definition corresponds to an 8-bit register-register add instruction in the X86. The string after the 'def' string indicates the name of the record ("ADDrr8" in this case), and the comment at the end of the line indicates the superclasses of the definition. The body of the record contains all of the data that TableGen assembled for the record, indicating that the instruction is part of the "X86" namespace, should be printed as "add" in the assembly file, it is a two-address instruction, has a particular encoding, etc. The contents and semantics of the information in the record is specific to the needs of the X86 backend, and is only shown as an example.

As you can see, a lot of information is needed for every instruction supported by the code generator, and specifying it all manually would be unmaintainble, prone to bugs, and tiring to do in the first place. Because we are using TableGen, all of the information was derived from the following definition:

def ADDrr8   : I2A8<"add", 0x00, MRMDestReg>,
               Pattern<(set R8, (plus R8, R8))>;

This definition makes use of the custom I2A8 (two address instruction with 8-bit operand) class, which is defined in the X86-specific TableGen file to factor out the common features that instructions of its class share. A key feature of TableGen is that it allows the end-user to define the abstractions they prefer to use when describing their information.

Running TableGen

TableGen runs just like any other LLVM tool. The first (optional) argument specifies the file to read. If a filename is not specified, tblgen reads from standard input.

To be useful, one of the TableGen backends must be used. These backends are selectable on the command line (type 'tblgen --help' for a list). For example, to get a list of all of the definitions that subclass a particular type (which can be useful for building up an enum list of these records), use the --print-enums option:

$ tblgen X86.td -print-enums -class=Register
AH, AL, AX, BH, BL, BP, BX, CH, CL, CX, DH, DI, DL, DX,
EAX, EBP, EBX, ECX, EDI, EDX, ESI, ESP, FP0, FP1, FP2, FP3, FP4, FP5, FP6,
SI, SP, ST0, ST1, ST2, ST3, ST4, ST5, ST6, ST7, 

$ tblgen X86.td -print-enums -class=Instruction 
ADCrr32, ADDri16, ADDri16b, ADDri32, ADDri32b, ADDri8, ADDrr16, ADDrr32,
ADDrr8, ADJCALLSTACKDOWN, ADJCALLSTACKUP, ANDri16, ANDri16b, ANDri32, ANDri32b,
ANDri8, ANDrr16, ANDrr32, ANDrr8, BSWAPr32, CALLm32, CALLpcrel32, ...

The default backend prints out all of the records, as described above.

If you plan to use TableGen for some purpose, you will most likely have to write a backend that extracts the information specific to what you need and formats it in the appropriate way.

TableGen syntax

TableGen doesn't care about the meaning of data (that is up to the backend to define), but it does care about syntax, and it enforces a simple type system. This section describes the syntax and the constructs allowed in a TableGen file.

TableGen primitives
TableGen comments

TableGen supports BCPL style "//" comments, which run to the end of the line, and it also supports nestable "/* */" comments.

The TableGen type system

TableGen files are strongly typed, in a simple (but complete) type-system. These types are used to perform automatic conversions, check for errors, and to help interface designers constrain the input that they allow. Every value definition is required to have an associated type.

TableGen supports a mixture of very low-level types (such as bit) and very high-level types (such as dag). This flexibility is what allows it to describe a wide range of information conveniently and compactly. The TableGen types are:

To date, these types have been sufficient for describing things that TableGen has been used for, but it is straight-forward to extend this list if needed.

TableGen values and expressions

TableGen allows for a pretty reasonable number of different expression forms when building up values. These forms allow the TableGen file to be written in a natural syntax and flavor for the application. The current expression forms supported include:

Note that all of the values have rules specifying how they convert to values for different types. These rules allow you to assign a value like "7" to a "bits<4>" value, for example.

Classes and definitions

As mentioned in the intro, classes and definitions (collectively known as 'records') in TableGen are the main high-level unit of information that TableGen collects. Records are defined with a def or class keyword, the record name, and an optional list of "template arguments". If the record has superclasses, they are specified as a comma separated list that starts with a colon character (":"). If value definitions or let expressions are needed for the class, they are enclosed in curly braces ("{}"); otherwise, the record ends with a semicolon. Here is a simple TableGen file:

class C { bit V = 1; }
def X : C;
def Y : C {
  string Greeting = "hello";
}

This example defines two definitions, X and Y, both of which derive from the C class. Because of this, they both get the V bit value. The Y definition also gets the Greeting member as well.

In general, classes are useful for collecting together the commonality between a group of records and isolating it in a single place. Also, classes permit the specification of default values for their subclasses, allowing the subclasses to override them as they wish.

Value definitions

Value definitions define named entries in records. A value must be defined before it can be referred to as the operand for another value definition or before the value is reset with a let expression. A value is defined by specifying a TableGen type and a name. If an initial value is available, it may be specified after the type with an equal sign. Value definitions require terminating semicolons.

'let' expressions

A record-level let expression is used to change the value of a value definition in a record. This is primarily useful when a superclass defines a value that a derived class or definition wants to override. Let expressions consist of the 'let' keyword followed by a value name, an equal sign ("="), and a new value. For example, a new class could be added to the example above, redefining the V field for all of its subclasses:

class D : C { let V = 0; }
def Z : D;

In this case, the Z definition will have a zero value for its "V" value, despite the fact that it derives (indirectly) from the C class, because the D class overrode its value.

Class template arguments

TableGen permits the definition of parameterized classes as well as normal concrete classes. Parameterized TableGen classes specify a list of variable bindings (which may optionally have defaults) that are bound when used. Here is a simple example:

class FPFormat<bits<3> 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 SpecialFP  : FPFormat<5>;

In this case, template arguments are used as a space efficient way to specify a list of "enumeration values", each with a "Value" field set to the specified integer.

The more esoteric forms of TableGen expressions are useful in conjunction with template arguments. As an example:

class ModRefVal<bits<2> val> {
  bits<2> Value = val;
}

def None   : ModRefVal<0>;
def Mod    : ModRefVal<1>;
def Ref    : ModRefVal<2>;
def ModRef : ModRefVal<3>;

class Value<ModRefVal MR> {
  // decode some information into a more convenient format, while providing
  // a nice interface to the user of the "Value" class.
  bit isMod = MR.Value{0};
  bit isRef = MR.Value{1};

  // other stuff...
}

// Example uses
def bork : Value<Mod>;
def zork : Value<Ref>;
def hork : Value<ModRef>;

This is obviously a contrived example, but it shows how template arguments can be used to decouple the interface provided to the user of the class from the actual internal data representation expected by the class. In this case, running tblgen on the example prints the following definitions:

def bork {      // Value
  bit isMod = 1;
  bit isRef = 0;
}
def hork {      // Value
  bit isMod = 1;
  bit isRef = 1;
}
def zork {      // Value
  bit isMod = 0;
  bit isRef = 1;
}

This shows that TableGen was able to dig into the argument and extract a piece of information that was requested by the designer of the "Value" class. For more realistic examples, please see existing users of TableGen, such as the X86 backend.

Multiclass definitions and instances

While classes with template arguments are a good way to factor commonality between two instances of a definition, multiclasses allow a convenient notation for defining multiple definitions at once (instances of implicitly constructed classes). For example, consider an 3-address instruction set whose instructions come in two forms: "reg = reg op reg" and "reg = reg op imm" (e.g. SPARC). In this case, you'd like to specify in one place that this commonality exists, then in a separate place indicate what all the ops are.

Here is an example TableGen fragment that shows this idea:

def ops;
def GPR;
def Imm;
class inst<int opc, string asmstr, dag operandlist>;

multiclass ri_inst<int opc, string asmstr> {
  def _rr : inst<opc, !strconcat(asmstr, " $dst, $src1, $src2"),
                 (ops GPR:$dst, GPR:$src1, GPR:$src2)>;
  def _ri : inst<opc, !strconcat(asmstr, " $dst, $src1, $src2"),
                 (ops GPR:$dst, GPR:$src1, Imm:$src2)>;
}

// Instantiations of the ri_inst multiclass.
defm ADD : ri_inst<0b111, "add">;
defm SUB : ri_inst<0b101, "sub">;
defm MUL : ri_inst<0b100, "mul">;
...

The name of the resultant definitions has the multidef fragment names appended to them, so this defines ADD_rr, ADD_ri, SUB_rr, etc. Using a multiclass this way is exactly equivalent to instantiating the classes multiple times yourself, e.g. by writing:

def ops;
def GPR;
def Imm;
class inst<int opc, string asmstr, dag operandlist>;

class rrinst<int opc, string asmstr>
  : inst<opc, !strconcat(asmstr, " $dst, $src1, $src2"),
         (ops GPR:$dst, GPR:$src1, GPR:$src2)>;

class riinst<int opc, string asmstr>
  : inst<opc, !strconcat(asmstr, " $dst, $src1, $src2"),
         (ops GPR:$dst, GPR:$src1, Imm:$src2)>;

// Instantiations of the ri_inst multiclass.
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">;
...
File scope entities
File inclusion

TableGen supports the 'include' token, which textually substitutes the specified file in place of the include directive. The filename should be specified as a double quoted string immediately after the 'include' keyword. Example:

include "foo.td"
'let' expressions

"let" expressions at file scope are similar to "let" expressions within a record, except they can specify a value binding for multiple records at a time, and may be useful in certain other cases. File-scope let expressions are really just another way that TableGen allows the end-user to factor out commonality from the records.

File-scope "let" expressions take a comma-separated list of bindings to apply, and one of more records to bind the values in. Here are some examples:

let isTerminator = 1, isReturn = 1 in
  def RET : X86Inst<"ret", 0xC3, RawFrm, NoArg>;

let isCall = 1 in
  // All calls clobber the non-callee saved registers...
  let Defs = [EAX, ECX, EDX, FP0, FP1, FP2, FP3, FP4, FP5, FP6] in {
    def CALLpcrel32 : X86Inst<"call", 0xE8, RawFrm, NoArg>;
    def CALLr32     : X86Inst<"call", 0xFF, MRMS2r, Arg32>;
    def CALLm32     : X86Inst<"call", 0xFF, MRMS2m, Arg32>;
  }

File-scope "let" expressions are often useful when a couple of definitions need to be added to several records, and the records do not otherwise need to be opened, as in the case with the CALL* instructions above.

TableGen backends

How they work, how to write one. This section should not contain details about any particular backend, except maybe -print-enums as an example. This should highlight the APIs in TableGen/Record.h.


Valid CSS! Valid HTML 4.01! Chris Lattner
LLVM Compiler Infrastructure
Last modified: $Date: 2007/09/27 05:43:13 $