<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
                      "http://www.w3.org/TR/html4/strict.dtd">
<html>
<head>
  <title>Writing an LLVM Compiler Backend</title>
  <link rel="stylesheet" href="llvm.css" type="text/css">
</head>

<body>

<div class="doc_title">
  Writing an LLVM Compiler Backend
</div>

<ol>
  <li><a href="#intro">Introduction</a>
  <ul>
    <li><a href="#Audience">Audience</a></li>
    <li><a href="#Prerequisite">Prerequisite Reading</a></li>
    <li><a href="#Basic">Basic Steps</a></li>
    <li><a href="#Preliminaries">Preliminaries</a></li>
  </ul>
  <li><a href="#TargetMachine">Target Machine</a></li>
  <li><a href="#RegisterSet">Register Set and Register Classes</a>
  <ul>
    <li><a href="#RegisterDef">Defining a Register</a></li>
    <li><a href="#RegisterClassDef">Defining a Register Class</a></li>
    <li><a href="#implementRegister">Implement a subclass of TargetRegisterInfo</a></li>
  </ul></li>
  <li><a href="#InstructionSet">Instruction Set</a>
  <ul>  
    <li><a href="#operandMapping">Instruction Operand Mapping</a></li>
    <li><a href="#implementInstr">Implement a subclass of TargetInstrInfo</a></li>
    <li><a href="#branchFolding">Branch Folding and If Conversion</a></li>
  </ul></li>
  <li><a href="#InstructionSelector">Instruction Selector</a>
  <ul>
    <li><a href="#LegalizePhase">The SelectionDAG Legalize Phase</a>
    <ul>
      <li><a href="#promote">Promote</a></li> 
      <li><a href="#expand">Expand</a></li> 
      <li><a href="#custom">Custom</a></li> 
      <li><a href="#legal">Legal</a></li>       
    </ul></li>
    <li><a href="#callingConventions">Calling Conventions</a></li>     
  </ul></li>
  <li><a href="#assemblyPrinter">Assembly Printer</a></li> 
  <li><a href="#subtargetSupport">Subtarget Support</a></li> 
  <li><a href="#jitSupport">JIT Support</a>
  <ul>  
    <li><a href="#mce">Machine Code Emitter</a></li>   
    <li><a href="#targetJITInfo">Target JIT Info</a></li>   
  </ul></li>
</ol>

<div class="doc_author">    
  <p>Written by <a href="http://www.woo.com">Mason Woo</a> and <a href="http://misha.brukman.net">Misha Brukman</a></p>
</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="intro">Introduction</a>
</div>
<!-- *********************************************************************** -->

<div class="doc_text">
<p>This document describes techniques for writing compiler backends
that convert the LLVM IR (intermediate representation) to code for a specified
machine or other languages. Code intended for a specific machine can take the
form of either assembly code or binary code (usable for a JIT compiler). </p>

<p>The backend of LLVM features a target-independent code generator
that may create output for several types of target CPUs, including X86,
PowerPC, Alpha, and SPARC. The backend may also be used to generate code
targeted at SPUs of the Cell processor or GPUs to support the execution of
compute kernels.</p>

<p>The document focuses on existing examples found in subdirectories
of <tt>llvm/lib/Target</tt> in a downloaded LLVM release. In particular, this document
focuses on the example of creating a static compiler (one that emits text
assembly) for a SPARC target, because SPARC has fairly standard
characteristics, such as a RISC instruction set and straightforward calling
conventions.</p>
</div>

<div class="doc_subsection">
  <a name="Audience">Audience</a>
</div>  

<div class="doc_text">
<p>The audience for this document is anyone who needs to write an
LLVM backend to generate code for a specific hardware or software target.</p>  
</div>

<div class="doc_subsection">
  <a name="Prerequisite">Prerequisite Reading</a>
</div>  

<div class="doc_text">  
These essential documents must be read before reading this document:  
<ul>
<li>
<i><a href="http://www.llvm.org/docs/LangRef.html">LLVM Language Reference Manual</a></i> - 
a reference manual for the LLVM assembly language
</li>
<li>
<i><a href="http://www.llvm.org/docs/CodeGenerator.html">The LLVM Target-Independent Code Generator </a></i> - 
a guide to the components (classes and code generation algorithms) for translating 
the LLVM internal representation to the machine code for a specified target. 
Pay particular attention to the descriptions of code generation stages: 
Instruction Selection, Scheduling and Formation, SSA-based Optimization, 
Register Allocation, Prolog/Epilog Code Insertion, Late Machine Code Optimizations, 
and Code Emission. 
</li>
<li>
<i><a href="http://www.llvm.org/docs/TableGenFundamentals.html">TableGen Fundamentals</a></i> - 
a document that describes the TableGen (tblgen) application that manages domain-specific 
information to support LLVM code generation. TableGen processes input from a 
target description file (.td suffix) and generates C++ code that can be used 
for code generation.
</li>
<li>
<i><a href="http://www.llvm.org/docs/WritingAnLLVMPass.html">Writing an LLVM Pass</a></i> - 
The assembly printer is a FunctionPass, as are several SelectionDAG processing steps.
</li>
</ul>
To follow the SPARC examples in this document, have a copy of 
<i><a href="http://www.sparc.org/standards/V8.pdf">The SPARC Architecture Manual, Version 8</a></i> 
for reference. For details about the ARM instruction set, refer to the 
<i><a href="http://infocenter.arm.com/">ARM Architecture Reference Manual</a></i>
For more about the GNU Assembler format (GAS), see 
<i><a href="http://sourceware.org/binutils/docs/as/index.html">Using As</a></i>
especially for the assembly printer. <i>Using As</i> contains lists of target machine dependent features. 
</div>

<div class="doc_subsection">
  <a name="Basic">Basic Steps</a>
</div>
<div class="doc_text">
<p>To write a compiler
backend for LLVM that converts the LLVM IR (intermediate representation)
to code for a specified target (machine or other language), follow these steps:</p>

<ul>
<li>
Create a subclass of the TargetMachine class that describes
characteristics of your target machine. Copy existing examples of specific
TargetMachine class and header files; for example, start with <tt>SparcTargetMachine.cpp</tt>
and <tt>SparcTargetMachine.h</tt>, but change the file names for your target. Similarly,
change code that references &quot;Sparc&quot; to reference your target. </li>

<li>Describe the register set of the target. Use TableGen to generate
code for register definition, register aliases, and register classes from a
target-specific <tt>RegisterInfo.td</tt> input file. You should also write additional
code for a subclass of TargetRegisterInfo class that represents the class
register file data used for register allocation and also describes the
interactions between registers.</li>

<li>Describe the instruction set of the target. Use TableGen to
generate code for target-specific instructions from target-specific versions of
<tt>TargetInstrFormats.td</tt> and <tt>TargetInstrInfo.td</tt>. You should write additional code
for a subclass of the TargetInstrInfo
class to represent machine
instructions supported by the target machine. </li>

<li>Describe the selection and conversion of the LLVM IR from a DAG (directed
acyclic graph) representation of instructions to native target-specific
instructions. Use TableGen to generate code that matches patterns and selects
instructions based on additional information in a target-specific version of
<tt>TargetInstrInfo.td</tt>. Write code for <tt>XXXISelDAGToDAG.cpp</tt> 
(where XXX identifies the specific target) to perform pattern
matching and DAG-to-DAG instruction selection. Also write code in <tt>XXXISelLowering.cpp</tt>
to replace or remove operations and data types that are not supported natively
in a SelectionDAG. </li>

<li>Write code for an
assembly printer that converts LLVM IR to a GAS format for your target machine.
You should add assembly strings to the instructions defined in your
target-specific version of <tt>TargetInstrInfo.td</tt>. You should also write code for a
subclass of AsmPrinter that performs the LLVM-to-assembly conversion and a
trivial subclass of TargetAsmInfo.</li>

<li>Optionally, add support for subtargets (that is, variants with
different capabilities). You should also write code for a subclass of the
TargetSubtarget class, which allows you to use the <tt>-mcpu=</tt> 
and <tt>-mattr=</tt> command-line options.</li>

<li>Optionally, add JIT support and create a machine code emitter (subclass
of TargetJITInfo) that is used to emit binary code directly into memory. </li>
</ul>

<p>In the .cpp and .h files, initially stub up these methods and
then implement them later. Initially, you may not know which private members
that the class will need and which components will need to be subclassed.</p>
</div>

<div class="doc_subsection">
  <a name="Preliminaries">Preliminaries</a>
</div>
<div class="doc_text">
<p>To actually create
your compiler backend, you need to create and modify a few files. The absolute
minimum is discussed here, but to actually use the LLVM target-independent code
generator, you must perform the steps described in the <a
href="http://www.llvm.org/docs/CodeGenerator.html">LLVM
Target-Independent Code Generator</a> document.</p>

<p>First, you should
create a subdirectory under <tt>lib/Target</tt> to hold all the files related to your
target. If your target is called &quot;Dummy&quot;, create the directory
<tt>lib/Target/Dummy</tt>.</p>

<p>In this new
directory, create a <tt>Makefile</tt>. It is easiest to copy a <tt>Makefile</tt> of another
target and modify it. It should at least contain the <tt>LEVEL</tt>, <tt>LIBRARYNAME</tt> and
<tt>TARGET</tt> variables, and then include <tt>$(LEVEL)/Makefile.common</tt>. The library can be
named LLVMDummy (for example, see the MIPS target). Alternatively, you can
split the library into LLVMDummyCodeGen and LLVMDummyAsmPrinter, the latter of
which should be implemented in a subdirectory below <tt>lib/Target/Dummy</tt> (for
example, see the PowerPC target).</p>

<p>Note that these two
naming schemes are hardcoded into <tt>llvm-config</tt>. Using any other naming scheme
will confuse <tt>llvm-config</tt> and produce lots of (seemingly unrelated) linker
errors when linking <tt>llc</tt>.</p>

<p>To make your target
actually do something, you need to implement a subclass of TargetMachine. This
implementation should typically be in the file
<tt>lib/Target/DummyTargetMachine.cpp</tt>, but any file in the <tt>lib/Target</tt> directory will
be built and should work. To use LLVM's target
independent code generator, you should do what all current machine backends do: create a subclass
of LLVMTargetMachine. (To create a target from scratch, create a subclass of
TargetMachine.)</p>

<p>To get LLVM to
actually build and link your target, you need to add it to the <tt>TARGETS_TO_BUILD</tt>
variable. To do this, you modify the configure script to know about your target
when parsing the <tt>--enable-targets</tt> option. Search the configure script for <tt>TARGETS_TO_BUILD</tt>,
add your target to the lists there (some creativity required) and then
reconfigure. Alternatively, you can change <tt>autotools/configure.ac</tt> and
regenerate configure by running <tt>./autoconf/AutoRegen.sh</tt></p>
</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="TargetMachine">Target Machine</a>
</div>
<!-- *********************************************************************** -->
<div class="doc_text">
<p>LLVMTargetMachine is designed as a base class for targets
implemented with the LLVM target-independent code generator. The
LLVMTargetMachine class should be specialized by a concrete target class that
implements the various virtual methods. LLVMTargetMachine is defined as a
subclass of TargetMachine in <tt>include/llvm/Target/TargetMachine.h</tt>. The
TargetMachine class implementation (<tt>TargetMachine.cpp</tt>) also processes numerous
command-line options.  </p>

<p>To create a concrete target-specific subclass of
LLVMTargetMachine, start by copying an existing TargetMachine class and header.
You should name the files that you create to reflect your specific target. For
instance, for the SPARC target, name the files <tt>SparcTargetMachine.h</tt> and
<tt>SparcTargetMachine.cpp</tt></p>

<p>For a target machine XXX, the implementation of XXXTargetMachine
must have access methods to obtain objects that represent target components.
These methods are named <tt>get*Info</tt> and are intended to obtain the instruction set
(<tt>getInstrInfo</tt>), register set (<tt>getRegisterInfo</tt>), stack frame layout
(<tt>getFrameInfo</tt>), and similar information. XXXTargetMachine must also implement
the <tt>getTargetData</tt> method to access an object with target-specific data
characteristics, such as data type size and alignment requirements. </p>

<p>For instance, for the SPARC target, the header file <tt>SparcTargetMachine.h</tt>
declares prototypes for several <tt>get*Info</tt> and <tt>getTargetData</tt> methods that simply
return a class member.  </p>
</div>

<div class="doc_code">
<pre>namespace llvm {

class Module;

class SparcTargetMachine : public LLVMTargetMachine {
  const TargetData DataLayout;       // Calculates type size &amp; alignment
  SparcSubtarget Subtarget;
  SparcInstrInfo InstrInfo;
  TargetFrameInfo FrameInfo;
  
protected:
  virtual const TargetAsmInfo *createTargetAsmInfo()
const;
  
public:
  SparcTargetMachine(const Module &amp;M, const std::string &amp;FS);

  virtual const SparcInstrInfo *getInstrInfo() const {return &amp;InstrInfo; }
  virtual const TargetFrameInfo *getFrameInfo() const {return &amp;FrameInfo; }
  virtual const TargetSubtarget *getSubtargetImpl() const{return &amp;Subtarget; }
  virtual const TargetRegisterInfo *getRegisterInfo() const {
    return &amp;InstrInfo.getRegisterInfo();
  }
  virtual const TargetData *getTargetData() const { return &amp;DataLayout; }
  static unsigned getModuleMatchQuality(const Module &amp;M);

  // Pass Pipeline Configuration
  virtual bool addInstSelector(PassManagerBase &amp;PM, bool Fast);
  virtual bool addPreEmitPass(PassManagerBase &amp;PM, bool Fast);
  virtual bool addAssemblyEmitter(PassManagerBase &amp;PM, bool Fast, 
                                  std::ostream &amp;Out);
};

} // end namespace llvm
</pre>
</div>

<div class="doc_text">
<ul>
<li><tt>getInstrInfo </tt></li>
<li><tt>getRegisterInfo</tt></li>
<li><tt>getFrameInfo</tt></li>
<li><tt>getTargetData</tt></li>
<li><tt>getSubtargetImpl</tt></li>
</ul>
<p>For some targets, you also need to support the following methods:
</p>

<ul>
<li><tt>getTargetLowering </tt></li>
<li><tt>getJITInfo</tt></li>
</ul>
<p>In addition, the XXXTargetMachine constructor should specify a
TargetDescription string that determines the data layout for the target machine,
including characteristics such as pointer size, alignment, and endianness. For
example, the constructor for SparcTargetMachine contains the following: </p>
</div>

<div class="doc_code">
<pre>
SparcTargetMachine::SparcTargetMachine(const Module &amp;M, const std::string &amp;FS)
  : DataLayout(&quot;E-p:32:32-f128:128:128&quot;),
    Subtarget(M, FS), InstrInfo(Subtarget),
    FrameInfo(TargetFrameInfo::StackGrowsDown, 8, 0) {
}
</pre>
</div>

<div class="doc_text">
<p>Hyphens separate portions of the TargetDescription string. </p>
<ul>
<li>The &quot;E&quot; in the string indicates a big-endian target data model; a
lower-case &quot;e&quot; would indicate little-endian. </li>
<li>&quot;p:&quot; is followed by pointer information: size, ABI alignment, and
preferred alignment. If only two figures follow &quot;p:&quot;, then the first value is
pointer size, and the second value is both ABI and preferred alignment.</li>
<li>then a letter for numeric type alignment: &quot;i&quot;, &quot;f&quot;, &quot;v&quot;, or &quot;a&quot;
(corresponding to integer, floating point, vector, or aggregate). &quot;i&quot;, &quot;v&quot;, or
&quot;a&quot; are followed by ABI alignment and preferred alignment. &quot;f&quot; is followed by
three values, the first indicates the size of a long double, then ABI alignment
and preferred alignment.</li>
</ul>
<p>You must also register your target using the RegisterTarget
template. (See the TargetMachineRegistry class.) For example, in <tt>SparcTargetMachine.cpp</tt>,
the target is registered with:</p>
</div>

<div class="doc_code">
<pre>
namespace {
  // Register the target.
  RegisterTarget&lt;SparcTargetMachine&gt;X(&quot;sparc&quot;, &quot;SPARC&quot;);
}
</pre>
</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="RegisterSet">Register Set and Register Classes</a>
</div>
<!-- *********************************************************************** -->
<div class="doc_text">
<p>You should describe
a concrete target-specific class
that represents the register file of a target machine. This class is
called XXXRegisterInfo (where XXX identifies the target) and represents the
class register file data that is used for register allocation and also
describes the interactions between registers. </p>

<p>You also need to
define register classes to categorize related registers. A register class
should be added for groups of registers that are all treated the same way for
some instruction. Typical examples are register classes that include integer,
floating-point, or vector registers. A&nbsp;register allocator allows an
instruction to use any register in a specified register class to perform the
instruction in a similar manner. Register classes allocate virtual registers to
instructions from these sets, and register classes let the target-independent
register allocator automatically choose the actual registers.</p>

<p>Much of the code for registers, including register definition,
register aliases, and register classes, is generated by TableGen from
<tt>XXXRegisterInfo.td</tt> input files and placed in <tt>XXXGenRegisterInfo.h.inc</tt> and
<tt>XXXGenRegisterInfo.inc</tt> output files. Some of the code in the implementation of
XXXRegisterInfo requires hand-coding. </p>
</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="RegisterDef">Defining a Register</a>
</div>
<div class="doc_text">
<p>The <tt>XXXRegisterInfo.td</tt> file typically starts with register definitions
for a target machine. The Register class (specified in <tt>Target.td</tt>) is used to
define an object for each register. The specified string n becomes the Name of
the register. The basic Register object does not have any subregisters and does
not specify any aliases.</p>
</div>
<div class="doc_code">
<pre>
class Register&lt;string n&gt; {
  string Namespace = &quot;&quot;;
  string AsmName = n;
  string Name = n;
  int SpillSize = 0;
  int SpillAlignment = 0;
  list&lt;Register&gt; Aliases = [];
  list&lt;Register&gt; SubRegs = [];
  list&lt;int&gt; DwarfNumbers = [];
}
</pre>
</div>

<div class="doc_text">
<p>For example, in the <tt>X86RegisterInfo.td</tt> file, there are register
definitions that utilize the Register class, such as:</p>
</div>
<div class="doc_code">
<pre>
def AL : Register&lt;&quot;AL&quot;&gt;,
DwarfRegNum&lt;[0, 0, 0]&gt;;
</pre>
</div>

<div class="doc_text">
<p>This defines the register AL and assigns it values (with
DwarfRegNum) that are used by <tt>gcc</tt>, <tt>gdb</tt>, or a debug information writer (such as
DwarfWriter in <tt>llvm/lib/CodeGen</tt>) to identify a register. For register AL,
DwarfRegNum takes an array of 3 values, representing 3 different modes: the
first element is for X86-64, the second for EH (exception handling) on X86-32,
and the third is generic. -1 is a special Dwarf number that indicates the gcc
number is undefined, and -2 indicates the register number is invalid for this
mode.</p>

<p>From the previously described line in the <tt>X86RegisterInfo.td</tt>
file, TableGen generates this code in the <tt>X86GenRegisterInfo.inc</tt> file:</p>
</div>
<div class="doc_code">
<pre>
  static const unsigned GR8[] = { X86::AL, ... };
&nbsp;
  const unsigned AL_AliasSet[] = { X86::AX, X86::EAX, X86::RAX, 0 };
&nbsp;
  const TargetRegisterDesc RegisterDescriptors[] = { 
    ...
    { &quot;AL&quot;, &quot;AL&quot;, AL_AliasSet, Empty_SubRegsSet, Empty_SubRegsSet, AL_SuperRegsSet }, ...
</pre>
</div>

<div class="doc_text">
<p>From the register info file, TableGen generates a
TargetRegisterDesc object for each register. TargetRegisterDesc is defined in
<tt>include/llvm/Target/TargetRegisterInfo.h</tt> with the following fields:</p>
</div>

<div class="doc_code">
<pre>
struct TargetRegisterDesc {
  const char     *AsmName;      // Assembly language name for the register
  const char     *Name;         // Printable name for the reg (for debugging)
  const unsigned *AliasSet;     // Register Alias Set
  const unsigned *SubRegs;      // Sub-register set
  const unsigned *ImmSubRegs;   // Immediate sub-register set
  const unsigned *SuperRegs;    // Super-register set
};</pre>
</div>

<div class="doc_text">
<p>TableGen uses the entire target description file (<tt>.td</tt>) to
determine text names for the register (in the AsmName and Name fields of
TargetRegisterDesc) and the relationships of other registers to the defined
register (in the other TargetRegisterDesc fields). In this example, other
definitions establish the registers &quot;AX&quot;, &quot;EAX&quot;, and &quot;RAX&quot; as aliases for one
another, so TableGen generates a null-terminated array (AL_AliasSet) for this
register alias set. </p>

<p>The Register class is commonly used as a base class for more
complex classes. In <tt>Target.td</tt>, the Register class is the base for the
RegisterWithSubRegs class that is used to define registers that need to specify
subregisters in the SubRegs list, as shown here:</p>
</div>
<div class="doc_code">
<pre>
class RegisterWithSubRegs&lt;string n,
list&lt;Register&gt; subregs&gt; : Register&lt;n&gt; {
  let SubRegs = subregs;
}</pre>
</div>

<div class="doc_text">
<p>In <tt>SparcRegisterInfo.td</tt>, additional register classes are defined
for SPARC: a Register subclass, SparcReg, and further subclasses: Ri, Rf, and
Rd. SPARC registers are identified by 5-bit ID numbers, which is a feature
common to these subclasses. Note the use of &lsquo;let&rsquo; expressions to override values
that are initially defined in a superclass (such as SubRegs field in the Rd
class). </p>
</div>
<div class="doc_code">
<pre>
class SparcReg&lt;string n&gt; : Register&lt;n&gt; {
  field bits&lt;5&gt; Num;
  let Namespace = &quot;SP&quot;;
}
// Ri - 32-bit integer registers
class Ri&lt;bits&lt;5&gt; num, string n&gt; :
SparcReg&lt;n&gt; {
  let Num = num;
}
// Rf - 32-bit floating-point registers
class Rf&lt;bits&lt;5&gt; num, string n&gt; :
SparcReg&lt;n&gt; {
  let Num = num;
}
// Rd - Slots in the FP register file for 64-bit
floating-point values.
class Rd&lt;bits&lt;5&gt; num, string n,
list&lt;Register&gt; subregs&gt; : SparcReg&lt;n&gt; {
  let Num = num;
  let SubRegs = subregs;
}</pre>
</div>
<div class="doc_text">
<p>In the <tt>SparcRegisterInfo.td</tt> file, there are register definitions
that utilize these subclasses of Register, such as:</p>
</div>
<div class="doc_code">
<pre>
def G0 : Ri&lt; 0, &quot;G0&quot;&gt;,
DwarfRegNum&lt;[0]&gt;;
def G1 : Ri&lt; 1, &quot;G1&quot;&gt;, DwarfRegNum&lt;[1]&gt;;
...
def F0 : Rf&lt; 0, &quot;F0&quot;&gt;,
DwarfRegNum&lt;[32]&gt;;
def F1 : Rf&lt; 1, &quot;F1&quot;&gt;,
DwarfRegNum&lt;[33]&gt;;
...
def D0 : Rd&lt; 0, &quot;F0&quot;, [F0, F1]&gt;,
DwarfRegNum&lt;[32]&gt;;
def D1 : Rd&lt; 2, &quot;F2&quot;, [F2, F3]&gt;,
DwarfRegNum&lt;[34]&gt;;
</pre>
</div>
<div class="doc_text">
<p>The last two registers shown above (D0 and D1) are double-precision
floating-point registers that are aliases for pairs of single-precision
floating-point sub-registers. In addition to aliases, the sub-register and
super-register relationships of the defined register are in fields of a
register&rsquo;s TargetRegisterDesc.</p>
</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="RegisterClassDef">Defining a Register Class</a>
</div>
<div class="doc_text">
<p>The RegisterClass class (specified in <tt>Target.td</tt>) is used to
define an object that represents a group of related registers and also defines
the default allocation order of the registers. A target description file
<tt>XXXRegisterInfo.td</tt> that uses <tt>Target.td</tt> can construct register classes using the
following class:</p>
</div>

<div class="doc_code">
<pre>
class RegisterClass&lt;string namespace,
list&lt;ValueType&gt; regTypes, int alignment,
                    list&lt;Register&gt; regList&gt; {
  string Namespace = namespace;
  list&lt;ValueType&gt; RegTypes = regTypes;
  int Size = 0;  // spill size, in bits; zero lets tblgen pick the size
  int Alignment = alignment;
&nbsp;
  // CopyCost is the cost of copying a value between two registers
  // default value 1 means a single instruction
  // A negative value means copying is extremely expensive or impossible
  int CopyCost = 1;  
  list&lt;Register&gt; MemberList = regList;
  
  // for register classes that are subregisters of this class
  list&lt;RegisterClass&gt; SubRegClassList = [];  
  
  code MethodProtos = [{}];  // to insert arbitrary code
  code MethodBodies = [{}];
}</pre>
</div>
<div class="doc_text">
<p>To define a RegisterClass, use the following 4 arguments:</p>
<ul>
<li>The first argument of the definition is the name of the
namespace. </li>

<li>The second argument is a list of ValueType register type values
that are defined in <tt>include/llvm/CodeGen/ValueTypes.td</tt>. Defined values include
integer types (such as i16, i32, and i1 for Boolean), floating-point types
(f32, f64), and vector types (for example, v8i16 for an 8 x i16 vector). All
registers in a RegisterClass must have the same ValueType, but some registers
may store vector data in different configurations. For example a register that
can process a 128-bit vector may be able to handle 16 8-bit integer elements, 8
16-bit integers, 4 32-bit integers, and so on. </li>

<li>The third argument of the RegisterClass definition specifies the
alignment required of the registers when they are stored or loaded to memory.</li>

<li>The final argument, <tt>regList</tt>, specifies which registers are in
this class.  If an <tt>allocation_order_*</tt> method is not specified, then <tt>regList</tt> also
defines the order of allocation used by the register allocator.</li>
</ul>

<p>In <tt>SparcRegisterInfo.td</tt>, three RegisterClass objects are defined:
FPRegs, DFPRegs, and IntRegs. For all three register classes, the first
argument defines the namespace with the string &ldquo;SP&rdquo;. FPRegs defines a group of 32
single-precision floating-point registers (F0 to F31); DFPRegs defines a group
of 16 double-precision registers (D0-D15). For IntRegs, the MethodProtos and
MethodBodies methods are used by TableGen to insert the specified code into generated
output.</p>
</div>
<div class="doc_code">
<pre>
def FPRegs : RegisterClass&lt;&quot;SP&quot;, [f32], 32, [F0, F1, F2, F3, F4, F5, F6, F7,   
  F8, F9, F10, F11, F12, F13, F14, F15, F16, F17, F18, F19, F20, F21, F22,
  F23, F24, F25, F26, F27, F28, F29, F30, F31]&gt;;
&nbsp;
def DFPRegs : RegisterClass&lt;&quot;SP&quot;, [f64], 64, [D0, D1, D2, D3, D4, D5, D6, D7,
  D8, D9, D10, D11, D12, D13, D14, D15]&gt;;
&nbsp;
def IntRegs : RegisterClass&lt;&quot;SP&quot;, [i32], 32, [L0, L1, L2, L3, L4, L5, L6, L7,
                                     I0, I1, I2, I3, I4, I5,
                                     O0, O1, O2, O3, O4, O5, O7,
                                     G1,
                                     // Non-allocatable regs:
                                     G2, G3, G4, 
                                     O6, // stack ptr
                                     I6, // frame ptr
                                     I7, // return address
                                     G0, // constant zero
                                     G5, G6, G7 // reserved for kernel
                                     ]&gt; {
  let MethodProtos = [{
    iterator allocation_order_end(const MachineFunction &amp;MF) const;
  }];
  let MethodBodies = [{
    IntRegsClass::iterator
    IntRegsClass::allocation_order_end(const MachineFunction &amp;MF) const {
      return end()-10  // Don't allocate special registers
         -1;  
    }
  }];
}
</pre>
</div>

<div class="doc_text">
<p>Using <tt>SparcRegisterInfo.td</tt> with TableGen generates several output
files that are intended for inclusion in other source code that you write.
<tt>SparcRegisterInfo.td</tt> generates <tt>SparcGenRegisterInfo.h.inc</tt>, which should be
included in the header file for the implementation of the SPARC register
implementation that you write (<tt>SparcRegisterInfo.h</tt>). In
<tt>SparcGenRegisterInfo.h.inc</tt> a new structure is defined called
SparcGenRegisterInfo that uses TargetRegisterInfo as its base. It also
specifies types, based upon the defined register classes: DFPRegsClass, FPRegsClass,
and IntRegsClass. </p>

<p><tt>SparcRegisterInfo.td</tt> also generates SparcGenRegisterInfo.inc,
which is included at the bottom of <tt>SparcRegisterInfo.cpp</tt>, the SPARC register
implementation. The code below shows only the generated integer registers and
associated register classes. The order of registers in IntRegs reflects the
order in the definition of IntRegs in the target description file. Take special
note of the use of MethodBodies in <tt>SparcRegisterInfo.td</tt> to create code in
<tt>SparcGenRegisterInfo.inc</tt>. MethodProtos generates similar code in
<tt>SparcGenRegisterInfo.h.inc</tt>.</p>
</div>

<div class="doc_code">
<pre>  // IntRegs Register Class...
  static const unsigned IntRegs[] = {
    SP::L0, SP::L1, SP::L2, SP::L3, SP::L4, SP::L5,
SP::L6, SP::L7, SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5, SP::O0, SP::O1,
SP::O2, SP::O3, SP::O4, SP::O5, SP::O7, SP::G1, SP::G2, SP::G3, SP::G4, SP::O6,
SP::I6, SP::I7, SP::G0, SP::G5, SP::G6, SP::G7, 
  };
&nbsp;
  // IntRegsVTs Register Class Value Types...
  static const MVT::ValueType IntRegsVTs[] = {
    MVT::i32, MVT::Other
  };
namespace SP {   // Register class instances
  DFPRegsClass&nbsp;&nbsp;&nbsp; DFPRegsRegClass;
  FPRegsClass&nbsp;&nbsp;&nbsp;&nbsp; FPRegsRegClass;
  IntRegsClass&nbsp;&nbsp;&nbsp; IntRegsRegClass;
...
&nbsp;
// IntRegs Sub-register Classess...
  static const TargetRegisterClass* const IntRegsSubRegClasses [] = {
    NULL
  };
...
// IntRegs Super-register Classess...
  static const TargetRegisterClass* const IntRegsSuperRegClasses [] = {
    NULL
  };
&nbsp;
// IntRegs Register Class sub-classes...
  static const TargetRegisterClass* const IntRegsSubclasses [] = {
    NULL
  };
...
&nbsp;
// IntRegs Register Class super-classes...
  static const TargetRegisterClass* const IntRegsSuperclasses [] = {
    NULL
  };
...
&nbsp;
  IntRegsClass::iterator
  IntRegsClass::allocation_order_end(const MachineFunction &amp;MF) const {

     return end()-10  // Don't allocate special registers
         -1; 
  }
  
IntRegsClass::IntRegsClass() : TargetRegisterClass(IntRegsRegClassID, 
   IntRegsVTs, IntRegsSubclasses, IntRegsSuperclasses, IntRegsSubRegClasses, 
   IntRegsSuperRegClasses, 4, 4, 1, IntRegs, IntRegs + 32) {}
}
</pre>
</div>
<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="implementRegister">Implement a subclass of</a> 
  <a href="http://www.llvm.org/docs/CodeGenerator.html#targetregisterinfo">TargetRegisterInfo</a>
</div>
<div class="doc_text">
<p>The final step is to hand code portions of XXXRegisterInfo, which
implements the interface described in <tt>TargetRegisterInfo.h</tt>. These functions
return 0, NULL, or false, unless overridden. Here&rsquo;s a list of functions that
are overridden for the SPARC implementation in <tt>SparcRegisterInfo.cpp</tt>:</p>
<ul>
<li><tt>getCalleeSavedRegs</tt> (returns a list of callee-saved registers in
the order of the desired callee-save stack frame offset)</li>

<li><tt>getCalleeSavedRegClasses</tt> (returns a list of preferred register
classes with which to spill each callee saved register)</li>

<li><tt>getReservedRegs</tt> (returns a bitset indexed by physical register
numbers, indicating if a particular register is unavailable)</li>

<li><tt>hasFP</tt> (return a Boolean indicating if a function should have a
dedicated frame pointer register)</li>

<li><tt>eliminateCallFramePseudoInstr</tt> (if call frame setup or destroy
pseudo instructions are used, this can be called to eliminate them)</li>

<li><tt>eliminateFrameIndex</tt> (eliminate abstract frame indices from
instructions that may use them)</li>

<li><tt>emitPrologue</tt> (insert prologue code into the function)</li>

<li><tt>emitEpilogue</tt> (insert epilogue code into the function)</li>
</ul>
</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="InstructionSet">Instruction Set</a>
</div>
<!-- *********************************************************************** -->
<div class="doc_text">
<p>During the early stages of code generation, the LLVM IR code is
converted to a SelectionDAG with nodes that are instances of the SDNode class
containing target instructions. An SDNode has an opcode, operands, type
requirements, and operation properties (for example, is an operation
commutative, does an operation load from memory). The various operation node
types are described in the <tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt> file (values
of the NodeType enum in the ISD namespace).</p>

<p>TableGen uses the following target description (.td) input files
to generate much of the code for instruction definition:</p>
<ul>
<li><tt>Target.td</tt>, where the Instruction, Operand, InstrInfo, and other
fundamental classes are defined</li>

<li><tt>TargetSelectionDAG.td</tt>, used by SelectionDAG instruction selection
generators, contains SDTC* classes (selection DAG type constraint), definitions
of SelectionDAG nodes (such as imm, cond, bb, add, fadd, sub), and pattern
support  (Pattern, Pat, PatFrag, PatLeaf, ComplexPattern)</li>

<li><tt>XXXInstrFormats.td</tt>, patterns for definitions of target-specific
instructions</li>

<li><tt>XXXInstrInfo.td</tt>, target-specific definitions of instruction
templates, condition codes, and instructions of an instruction set. (For architecture
modifications, a different file name may be used. For example, for Pentium with
SSE instruction, this file is <tt>X86InstrSSE.td</tt>, and for Pentium with MMX, this
file is <tt>X86InstrMMX.td</tt>.)</li>
</ul>
<p>There is also a target-specific <tt>XXX.td</tt> file, where XXX is the
name of the target. The <tt>XXX.td</tt> file includes the other .td input files, but its
contents are only directly important for subtargets.</p>

<p>You should describe
a concrete target-specific class
XXXInstrInfo that represents machine
instructions supported by a target machine. XXXInstrInfo contains an array of
XXXInstrDescriptor objects, each of which describes one instruction. An
instruction descriptor defines:</p>
<ul>
<li>opcode mnemonic</li>

<li>number of operands</li>

<li>list of implicit register definitions and uses</li>

<li>target-independent properties (such as memory access, is
commutable)</li>

<li>target-specific flags </li>
</ul>

<p>The Instruction class (defined in <tt>Target.td</tt>) is mostly used as a
base for more complex instruction classes.</p>
</div>

<div class="doc_code">
<pre>class Instruction {
  string Namespace = &quot;&quot;;
  dag OutOperandList;       // An dag containing the MI def operand list.
  dag InOperandList;        // An dag containing the MI use operand list.
  string AsmString = &quot;&quot;;    // The .s format to print the instruction with.
  list&lt;dag&gt; Pattern;  // Set to the DAG pattern for this instruction
  list&lt;Register&gt; Uses = []; 
  list&lt;Register&gt; Defs = [];
  list&lt;Predicate&gt; Predicates = [];  // predicates turned into isel match code
  ... remainder not shown for space ...
}
</pre>
</div>
<div class="doc_text">
<p>A SelectionDAG node (SDNode) should contain an object
representing a target-specific instruction that is defined in <tt>XXXInstrInfo.td</tt>. The
instruction objects should represent instructions from the architecture manual
of the target machine (such as the
SPARC Architecture Manual for the SPARC target). </p>

<p>A single
instruction from the architecture manual is often modeled as multiple target
instructions, depending upon its operands. &nbsp;For example, a manual might
describe an add instruction that takes a register or an immediate operand. An
LLVM target could model this with two instructions named ADDri and ADDrr.</p>

<p>You should define a
class for each instruction category and define each opcode as a subclass of the
category with appropriate parameters such as the fixed binary encoding of
opcodes and extended opcodes. You should map the register bits to the bits of
the instruction in which they are encoded (for the JIT). Also you should specify
how the instruction should be printed when the automatic assembly printer is
used.</p>

<p>As is described in
the SPARC Architecture Manual, Version 8, there are three major 32-bit formats
for instructions. Format 1 is only for the CALL instruction. Format 2 is for
branch on condition codes and SETHI (set high bits of a register) instructions.
Format 3 is for other instructions. </p>

<p>Each of these
formats has corresponding classes in <tt>SparcInstrFormat.td</tt>. InstSP is a base
class for other instruction classes. Additional base classes are specified for
more precise formats: for example in <tt>SparcInstrFormat.td</tt>, F2_1 is for SETHI,
and F2_2 is for branches. There are three other base classes: F3_1 for
register/register operations, F3_2 for register/immediate operations, and F3_3 for
floating-point operations. <tt>SparcInstrInfo.td</tt> also adds the base class Pseudo for
synthetic SPARC instructions.   </p>

<p><tt>SparcInstrInfo.td</tt>
largely consists of operand and instruction definitions for the SPARC target. In
<tt>SparcInstrInfo.td</tt>, the following target description file entry, LDrr, defines
the Load Integer instruction for a Word (the LD SPARC opcode) from a memory
address to a register. The first parameter, the value 3 (11<sub>2</sub>), is
the operation value for this category of operation. The second parameter
(000000<sub>2</sub>) is the specific operation value for LD/Load Word. The
third parameter is the output destination, which is a register operand and
defined in the Register target description file (IntRegs). </p>
</div>
<div class="doc_code">
<pre>def LDrr : F3_1 &lt;3, 0b000000, (outs IntRegs:$dst), (ins MEMrr:$addr),
                 &quot;ld [$addr], $dst&quot;,
                 [(set IntRegs:$dst, (load ADDRrr:$addr))]&gt;;
</pre>
</div>

<div class="doc_text">
<p>The fourth
parameter is the input source, which uses the address operand MEMrr that is
defined earlier in <tt>SparcInstrInfo.td</tt>:</p>
</div>
<div class="doc_code">
<pre>def MEMrr : Operand&lt;i32&gt; {
  let PrintMethod = &quot;printMemOperand&quot;;
  let MIOperandInfo = (ops IntRegs, IntRegs);
}
</pre>
</div>
<div class="doc_text">
<p>The fifth parameter is a string that is used by the assembly
printer and can be left as an empty string until the assembly printer interface
is implemented. The sixth and final parameter is the pattern used to match the
instruction during the SelectionDAG Select Phase described in 
(<a href="http://www.llvm.org/docs/CodeGenerator.html">The LLVM Target-Independent Code Generator</a>).
This parameter is detailed in the next section, <a href="#InstructionSelector">Instruction Selector</a>.</p>

<p>Instruction class definitions are not overloaded for different
operand types, so separate versions of instructions are needed for register,
memory, or immediate value operands. For example, to perform a 
Load Integer instruction for a Word
from an immediate operand to a register, the following instruction class is
defined: </p>
</div>
<div class="doc_code">
<pre>def LDri : F3_2 &lt;3, 0b000000, (outs IntRegs:$dst), (ins MEMri:$addr),
                 &quot;ld [$addr], $dst&quot;,
                 [(set IntRegs:$dst, (load ADDRri:$addr))]&gt;;
</pre>
</div>
<div class="doc_text">
<p>Writing these definitions for so many similar instructions can
involve a lot of cut and paste. In td files, the <tt>multiclass</tt> directive enables
the creation of templates to define several instruction classes at once (using
the <tt>defm</tt> directive). For example in
<tt>SparcInstrInfo.td</tt>, the <tt>multiclass</tt> pattern F3_12 is defined to create 2
instruction classes each time F3_12 is invoked:  </p>
</div>
<div class="doc_code">
<pre>multiclass F3_12 &lt;string OpcStr, bits&lt;6&gt; Op3Val, SDNode OpNode&gt; {
  def rr  : F3_1 &lt;2, Op3Val, 
                 (outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c),
                 !strconcat(OpcStr, &quot; $b, $c, $dst&quot;),
                 [(set IntRegs:$dst, (OpNode IntRegs:$b, IntRegs:$c))]&gt;;
  def ri  : F3_2 &lt;2, Op3Val,
                 (outs IntRegs:$dst), (ins IntRegs:$b, i32imm:$c),
                 !strconcat(OpcStr, &quot; $b, $c, $dst&quot;),
                 [(set IntRegs:$dst, (OpNode IntRegs:$b, simm13:$c))]&gt;;
}
</pre>
</div>
<div class="doc_text">
<p>So when the <tt>defm</tt> directive is used for the XOR and ADD
instructions, as seen below, it creates four instruction objects: XORrr, XORri,
ADDrr, and ADDri.</p>
</div>
<div class="doc_code">
<pre>defm XOR   : F3_12&lt;&quot;xor&quot;, 0b000011, xor&gt;;
defm ADD   : F3_12&lt;&quot;add&quot;, 0b000000, add&gt;;
</pre>
</div>

<div class="doc_text">
<p><tt>SparcInstrInfo.td</tt>
also includes definitions for condition codes that are referenced by branch
instructions. The following definitions in <tt>SparcInstrInfo.td</tt> indicate the bit location
of the SPARC condition code; for example, the 10<sup>th</sup> bit represents
the &lsquo;greater than&rsquo; condition for integers, and the 22<sup>nd</sup> bit
represents the &lsquo;greater than&rsquo; condition for floats. </p>
</div>

<div class="doc_code">
<pre>def ICC_NE  : ICC_VAL&lt; 9&gt;;  // Not Equal
def ICC_E   : ICC_VAL&lt; 1&gt;;  // Equal
def ICC_G   : ICC_VAL&lt;10&gt;;  // Greater
...
def FCC_U   : FCC_VAL&lt;23&gt;;  // Unordered
def FCC_G   : FCC_VAL&lt;22&gt;;  // Greater
def FCC_UG  : FCC_VAL&lt;21&gt;;  // Unordered or Greater
...
</pre>
</div>

<div class="doc_text">
<p>(Note that <tt>Sparc.h</tt>
also defines enums that correspond to the same SPARC condition codes. Care must
be taken to ensure the values in <tt>Sparc.h</tt> correspond to the values in
<tt>SparcInstrInfo.td</tt>; that is, <tt>SPCC::ICC_NE = 9</tt>, <tt>SPCC::FCC_U = 23</tt> and so on.)</p>
</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="operandMapping">Instruction Operand Mapping</a>
</div>
<div class="doc_text">
<p>The code generator backend maps instruction operands to fields in
the instruction.  Operands are assigned to unbound fields in the instruction in 
the order they are defined. Fields are bound when they are assigned a value.
For example, the Sparc target defines the XNORrr instruction as a F3_1 format 
instruction having three operands.</p>
</div>

<div class="doc_code"> <pre>
def XNORrr  : F3_1&lt;2, 0b000111,
                   (outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c),
                   "xnor $b, $c, $dst",
                   [(set IntRegs:$dst, (not (xor IntRegs:$b, IntRegs:$c)))]&gt;;
</pre></div>

<div class="doc_text">
<p>The instruction templates in <tt>SparcInstrFormats.td</tt> show the base class for F3_1 is InstSP.</p>
</div>

<div class="doc_code"> <pre>
class InstSP&lt;dag outs, dag ins, string asmstr, list&lt;dag&gt; pattern&gt; : Instruction {
  field bits&lt;32&gt; Inst;
  let Namespace = "SP";
  bits&lt;2&gt; op;
  let Inst{31-30} = op;       
  dag OutOperandList = outs;
  dag InOperandList = ins;
  let AsmString   = asmstr;
  let Pattern = pattern;
}
</pre></div>
<div class="doc_text">
<p>
InstSP leaves the op field unbound.
</p>
</div>

<div class="doc_code"> <pre>
class F3&lt;dag outs, dag ins, string asmstr, list&lt;dag&gt; pattern&gt;
    : InstSP&lt;outs, ins, asmstr, pattern&gt; {
  bits&lt;5&gt; rd;
  bits&lt;6&gt; op3;
  bits&lt;5&gt; rs1;
  let op{1} = 1;   // Op = 2 or 3
  let Inst{29-25} = rd;
  let Inst{24-19} = op3;
  let Inst{18-14} = rs1;
}
</pre></div>
<div class="doc_text">
<p>
F3 binds the op field and defines the rd, op3, and rs1 fields.  F3 format instructions will
bind the operands rd, op3, and rs1 fields.
</p>
</div>

<div class="doc_code"> <pre>
class F3_1&lt;bits&lt;2&gt; opVal, bits&lt;6&gt; op3val, dag outs, dag ins,
           string asmstr, list&lt;dag&gt; pattern&gt; : F3&lt;outs, ins, asmstr, pattern&gt; {
  bits&lt;8&gt; asi = 0; // asi not currently used
  bits&lt;5&gt; rs2;
  let op         = opVal;
  let op3        = op3val;
  let Inst{13}   = 0;     // i field = 0
  let Inst{12-5} = asi;   // address space identifier
  let Inst{4-0}  = rs2;
}
</pre></div>
<div class="doc_text">
<p>
F3_1 binds the op3 field and defines the rs2 fields.  F3_1 format instructions will
bind the operands to the rd, rs1, and rs2 fields. This results in the XNORrr instruction
binding $dst, $b, and $c operands to the rd, rs1, and rs2 fields respectively.
</p>
</div>


<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="implementInstr">Implement a subclass of </a>
  <a href="http://www.llvm.org/docs/CodeGenerator.html#targetinstrinfo">TargetInstrInfo</a>
</div>

<div class="doc_text">
<p>The final step is to hand code portions of XXXInstrInfo, which
implements the interface described in <tt>TargetInstrInfo.h</tt>. These functions return
0 or a Boolean or they assert, unless overridden. Here's a list of functions
that are overridden for the SPARC implementation in <tt>SparcInstrInfo.cpp</tt>:</p>
<ul>
<li><tt>isMoveInstr</tt> (return true if the instruction is a register to
register move; false, otherwise)</li>

<li><tt>isLoadFromStackSlot</tt> (if the specified machine instruction is a
direct load from a stack slot, return the register number of the destination
and the FrameIndex of the stack slot)</li>

<li><tt>isStoreToStackSlot</tt> (if the specified machine instruction is a
direct store to a stack slot, return the register number of the destination and
the FrameIndex of the stack slot)</li>

<li><tt>copyRegToReg</tt> (copy values between a pair of registers)</li>

<li><tt>storeRegToStackSlot</tt> (store a register value to a stack slot)</li>

<li><tt>loadRegFromStackSlot</tt> (load a register value from a stack slot)</li>

<li><tt>storeRegToAddr</tt> (store a register value to memory)</li>

<li><tt>loadRegFromAddr</tt> (load a register value from memory)</li>

<li><tt>foldMemoryOperand</tt> (attempt to combine instructions of any load or
store instruction for the specified operand(s))</li>
</ul>
</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="branchFolding">Branch Folding and If Conversion</a>
</div>
<div class="doc_text">
<p>Performance can be improved by combining instructions or by eliminating
instructions that are never reached. The <tt>AnalyzeBranch</tt> method in XXXInstrInfo may
be implemented to examine conditional instructions and remove unnecessary
instructions. <tt>AnalyzeBranch</tt> looks at the end of a machine basic block (MBB) for
opportunities for improvement, such as branch folding and if conversion. The
<tt>BranchFolder</tt> and <tt>IfConverter</tt> machine function passes (see the source files
<tt>BranchFolding.cpp</tt> and <tt>IfConversion.cpp</tt> in the <tt>lib/CodeGen</tt> directory) call
<tt>AnalyzeBranch</tt> to improve the control flow graph that represents the
instructions. </p>

<p>Several implementations of <tt>AnalyzeBranch</tt> (for ARM, Alpha, and
X86) can be examined as models for your own <tt>AnalyzeBranch</tt> implementation. Since
SPARC does not implement a useful <tt>AnalyzeBranch</tt>, the ARM target implementation
is shown below.</p>

<p><tt>AnalyzeBranch</tt> returns a Boolean value and takes four parameters:</p>
<ul>
<li>MachineBasicBlock &amp;MBB &#8211; the incoming block to be
examined</li>

<li>MachineBasicBlock *&amp;TBB &#8211; a destination block that is
returned; for a conditional branch that evaluates to true, TBB is the
destination </li>

<li>MachineBasicBlock *&amp;FBB &#8211; for a conditional branch that
evaluates to false, FBB is returned as the destination</li>

<li>std::vector&lt;MachineOperand&gt; &amp;Cond &#8211; list of
operands to evaluate a condition for a conditional branch</li>
</ul>

<p>In the simplest case, if a block ends without a branch, then it
falls through to the successor block. No destination blocks are specified for
either TBB or FBB, so both parameters return NULL. The start of the <tt>AnalyzeBranch</tt>
(see code below for the ARM target) shows the function parameters and the code
for the simplest case.</p>
</div>

<div class="doc_code">
<pre>bool ARMInstrInfo::AnalyzeBranch(MachineBasicBlock &amp;MBB,
        MachineBasicBlock *&amp;TBB, MachineBasicBlock *&amp;FBB,
        std::vector&lt;MachineOperand&gt; &amp;Cond) const
{
  MachineBasicBlock::iterator I = MBB.end();
  if (I == MBB.begin() || !isUnpredicatedTerminator(--I))
    return false;
</pre>
</div>

<div class="doc_text">
<p>If a block ends with a single unconditional branch instruction,
then <tt>AnalyzeBranch</tt> (shown below) should return the destination of that branch
in the TBB parameter. </p>
</div>

<div class="doc_code">
<pre>if (LastOpc == ARM::B || LastOpc == ARM::tB) {
      TBB = LastInst-&gt;getOperand(0).getMBB();
      return false;
    }
</pre>
</div>

<div class="doc_text">
<p>If a block ends with two unconditional branches, then the second
branch is never reached. In that situation, as shown below, remove the last
branch instruction and return the penultimate branch in the TBB parameter. </p>
</div>

<div class="doc_code">
<pre>if ((SecondLastOpc == ARM::B || SecondLastOpc==ARM::tB) &amp;&amp;
      (LastOpc == ARM::B || LastOpc == ARM::tB)) {
    TBB = SecondLastInst-&gt;getOperand(0).getMBB();
    I = LastInst;
    I-&gt;eraseFromParent();
    return false;
  }
</pre>
</div>
<div class="doc_text">
<p>A block may end with a single conditional branch instruction that
falls through to successor block if the condition evaluates to false. In that
case, <tt>AnalyzeBranch</tt> (shown below) should return the destination of that
conditional branch in the TBB parameter and a list of operands in the <tt>Cond</tt>
parameter to evaluate the condition. </p>
</div>

<div class="doc_code">
<pre>if (LastOpc == ARM::Bcc || LastOpc == ARM::tBcc) {
      // Block ends with fall-through condbranch.
      TBB = LastInst-&gt;getOperand(0).getMBB();
      Cond.push_back(LastInst-&gt;getOperand(1));
      Cond.push_back(LastInst-&gt;getOperand(2));
      return false;
    }
</pre>
</div>

<div class="doc_text">
<p>If a block ends with both a conditional branch and an ensuing
unconditional branch, then <tt>AnalyzeBranch</tt> (shown below) should return the
conditional branch destination (assuming it corresponds to a conditional
evaluation of &lsquo;true&rsquo;) in the TBB parameter and the unconditional branch
destination in the FBB (corresponding to a conditional evaluation of &lsquo;false&rsquo;).
A list of operands to evaluate the condition should be returned in the <tt>Cond</tt>
parameter.</p>
</div>

<div class="doc_code">
<pre>unsigned SecondLastOpc = SecondLastInst-&gt;getOpcode();
  if ((SecondLastOpc == ARM::Bcc &amp;&amp; LastOpc == ARM::B) ||
      (SecondLastOpc == ARM::tBcc &amp;&amp; LastOpc == ARM::tB)) {
    TBB =  SecondLastInst-&gt;getOperand(0).getMBB();
    Cond.push_back(SecondLastInst-&gt;getOperand(1));
    Cond.push_back(SecondLastInst-&gt;getOperand(2));
    FBB = LastInst-&gt;getOperand(0).getMBB();
    return false;
  }
</pre>
</div>

<div class="doc_text">
<p>For the last two cases (ending with a single conditional branch or
ending with one conditional and one unconditional branch), the operands returned
in the <tt>Cond</tt> parameter can be passed to methods of other instructions to create
new branches or perform other operations. An implementation of <tt>AnalyzeBranch</tt>
requires the helper methods <tt>RemoveBranch</tt> and <tt>InsertBranch</tt> to manage subsequent
operations.</p>

<p><tt>AnalyzeBranch</tt> should return false indicating success in most circumstances.
<tt>AnalyzeBranch</tt> should only return true when the method is stumped about what to
do, for example, if a block has three terminating branches. <tt>AnalyzeBranch</tt> may
return true if it encounters a terminator it cannot handle, such as an indirect
branch.</p>
</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="InstructionSelector">Instruction Selector</a>
</div>
<!-- *********************************************************************** -->

<div class="doc_text">
<p>LLVM uses a SelectionDAG to represent LLVM IR instructions, and nodes
of the SelectionDAG ideally represent native target instructions. During code
generation, instruction selection passes are performed to convert non-native
DAG instructions into native target-specific instructions. The pass described
in <tt>XXXISelDAGToDAG.cpp</tt> is used to match patterns and perform DAG-to-DAG
instruction selection. Optionally, a pass may be defined (in
<tt>XXXBranchSelector.cpp</tt>) to perform similar DAG-to-DAG operations for branch
instructions. Later,
the code in <tt>XXXISelLowering.cpp</tt> replaces or removes operations and data types
not supported natively (legalizes) in a Selection DAG. </p>

<p>TableGen generates code for instruction selection using the
following target description input files:</p>
<ul>
<li><tt>XXXInstrInfo.td</tt> contains definitions of instructions in a
target-specific instruction set, generates <tt>XXXGenDAGISel.inc</tt>, which is included
in <tt>XXXISelDAGToDAG.cpp</tt>. </li>

<li><tt>XXXCallingConv.td</tt> contains the calling and return value conventions
for the target architecture, and it generates <tt>XXXGenCallingConv.inc</tt>, which is
included in <tt>XXXISelLowering.cpp</tt>.</li>
</ul>

<p>The implementation of an instruction selection pass must include
a header that declares the FunctionPass class or a subclass of FunctionPass. In
<tt>XXXTargetMachine.cpp</tt>, a Pass Manager (PM) should add each instruction selection
pass into the queue of passes to run.</p>

<p>The LLVM static
compiler (<tt>llc</tt>) is an excellent tool for visualizing the contents of DAGs. To display
the SelectionDAG before or after specific processing phases, use the command
line options for <tt>llc</tt>, described at <a
href="http://llvm.org/docs/CodeGenerator.html#selectiondag_process">
SelectionDAG Instruction Selection Process</a>.
</p>

<p>To describe instruction selector behavior, you should add
patterns for lowering LLVM code into a SelectionDAG as the last parameter of
the instruction definitions in <tt>XXXInstrInfo.td</tt>. For example, in
<tt>SparcInstrInfo.td</tt>, this entry defines a register store operation, and the last
parameter describes a pattern with the store DAG operator.</p>
</div>

<div class="doc_code">
<pre>def STrr  : F3_1&lt; 3, 0b000100, (outs), (ins MEMrr:$addr, IntRegs:$src),
                 &quot;st $src, [$addr]&quot;, [(store IntRegs:$src, ADDRrr:$addr)]&gt;;
</pre>
</div>

<div class="doc_text">
<p>ADDRrr is a memory mode that is also defined in <tt>SparcInstrInfo.td</tt>:</p>
</div>

<div class="doc_code">
<pre>def ADDRrr : ComplexPattern&lt;i32, 2, &quot;SelectADDRrr&quot;, [], []&gt;;
</pre>
</div>

<div class="doc_text">
<p>The definition of ADDRrr refers to SelectADDRrr, which is a function defined in an
implementation of the Instructor Selector (such as <tt>SparcISelDAGToDAG.cpp</tt>). </p>

<p>In <tt>lib/Target/TargetSelectionDAG.td</tt>, the DAG operator for store
is defined below:</p>
</div>

<div class="doc_code">
<pre>def store : PatFrag&lt;(ops node:$val, node:$ptr),
                    (st node:$val, node:$ptr), [{
  if (StoreSDNode *ST = dyn_cast&lt;StoreSDNode&gt;(N))
    return !ST-&gt;isTruncatingStore() &amp;&amp; 
           ST-&gt;getAddressingMode() == ISD::UNINDEXED;
  return false;
}]&gt;;
</pre>
</div>
<div class="doc_text">
<p><tt>XXXInstrInfo.td</tt> also generates (in <tt>XXXGenDAGISel.inc</tt>) the
<tt>SelectCode</tt> method that is used to call the appropriate processing method for an
instruction. In this example, <tt>SelectCode</tt> calls <tt>Select_ISD_STORE</tt> for the
ISD::STORE opcode.</p>
</div>

<div class="doc_code">
<pre>SDNode *SelectCode(SDValue N) {
  ... 
  MVT::ValueType NVT = N.getNode()-&gt;getValueType(0);
  switch (N.getOpcode()) {
  case ISD::STORE: {
    switch (NVT) {
    default:
      return Select_ISD_STORE(N);
      break;
    }
    break;
  }
  ...
</pre>
</div>
<div class="doc_text">
<p>The pattern for STrr is matched, so elsewhere in
<tt>XXXGenDAGISel.inc</tt>, code for STrr is created for <tt>Select_ISD_STORE</tt>. The <tt>Emit_22</tt> method
is also generated in <tt>XXXGenDAGISel.inc</tt> to complete the processing of this
instruction. </p>
</div>

<div class="doc_code">
<pre>SDNode *Select_ISD_STORE(const SDValue &amp;N) {
  SDValue Chain = N.getOperand(0);
  if (Predicate_store(N.getNode())) {
    SDValue N1 = N.getOperand(1);
    SDValue N2 = N.getOperand(2);
    SDValue CPTmp0;
    SDValue CPTmp1;
&nbsp;
    // Pattern: (st:void IntRegs:i32:$src, 
    //           ADDRrr:i32:$addr)&lt;&lt;P:Predicate_store&gt;&gt;
    // Emits: (STrr:void ADDRrr:i32:$addr, IntRegs:i32:$src)
    // Pattern complexity = 13  cost = 1  size = 0
    if (SelectADDRrr(N, N2, CPTmp0, CPTmp1) &amp;&amp;
        N1.getNode()-&gt;getValueType(0) == MVT::i32 &amp;&amp;
        N2.getNode()-&gt;getValueType(0) == MVT::i32) {
      return Emit_22(N, SP::STrr, CPTmp0, CPTmp1);
    }
...
</pre>
</div>

<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="LegalizePhase">The SelectionDAG Legalize Phase</a>
</div>
<div class="doc_text">
<p>The Legalize phase converts a DAG to use types and operations
that are natively supported by the target. For natively unsupported types and
operations, you need to add code to the target-specific XXXTargetLowering implementation
to convert unsupported types and operations to supported ones.</p>

<p>In the constructor for the XXXTargetLowering class, first use the
<tt>addRegisterClass</tt> method to specify which types are supports and which register
classes are associated with them. The code for the register classes are generated
by TableGen from <tt>XXXRegisterInfo.td</tt> and placed in <tt>XXXGenRegisterInfo.h.inc</tt>. For
example, the implementation of the constructor for the SparcTargetLowering
class (in <tt>SparcISelLowering.cpp</tt>) starts with the following code:</p>
</div>

<div class="doc_code">
<pre>addRegisterClass(MVT::i32, SP::IntRegsRegisterClass);
addRegisterClass(MVT::f32, SP::FPRegsRegisterClass);
addRegisterClass(MVT::f64, SP::DFPRegsRegisterClass); 
</pre>
</div>

<div class="doc_text">
<p>You should examine the node types in the ISD namespace 
(<tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt>)
and determine which operations the target natively supports. For operations
that do <b>not</b> have native support, add a callback to the constructor for
the XXXTargetLowering class, so the instruction selection process knows what to
do. The TargetLowering class callback methods (declared in
<tt>llvm/Target/TargetLowering.h</tt>) are:</p>
<ul>
<li><tt>setOperationAction</tt> (general operation)</li>

<li><tt>setLoadExtAction</tt> (load with extension)</li>

<li><tt>setTruncStoreAction</tt> (truncating store)</li>

<li><tt>setIndexedLoadAction</tt> (indexed load)</li>

<li><tt>setIndexedStoreAction</tt> (indexed store)</li>

<li><tt>setConvertAction</tt> (type conversion)</li>

<li><tt>setCondCodeAction</tt> (support for a given condition code)</li>
</ul>

<p>Note: on older releases, <tt>setLoadXAction</tt> is used instead of <tt>setLoadExtAction</tt>.
Also, on older releases, <tt>setCondCodeAction</tt> may not be supported. Examine your
release to see what methods are specifically supported.</p>

<p>These callbacks are used to determine that an operation does or
does not work with a specified type (or types). And in all cases, the third
parameter is a LegalAction type enum value: <tt>Promote</tt>, <tt>Expand</tt>, 
<tt>Custom</tt>, or <tt>Legal</tt>. <tt>SparcISelLowering.cpp</tt>
contains examples of all four LegalAction values.</p>
</div>

<!-- _______________________________________________________________________ -->
<div class="doc_subsubsection">
  <a name="promote">Promote</a>
</div>

<div class="doc_text">
<p>For an operation without native support for a given type, the
specified type may be promoted to a larger type that is supported. For example,
SPARC does not support a sign-extending load for Boolean values (<tt>i1</tt> type), so
in <tt>SparcISelLowering.cpp</tt> the third 
parameter below, <tt>Promote</tt>, changes <tt>i1</tt> type
values to a large type before loading.</p>
</div>

<div class="doc_code">
<pre>setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
</pre>
</div>

<!-- _______________________________________________________________________ -->
<div class="doc_subsubsection">
  <a name="expand">Expand</a>
</div>
<div class="doc_text">
<p>For a type without native support, a value may need to be broken
down further, rather than promoted. For an operation without native support, a
combination of other operations may be used to similar effect. In SPARC, the
floating-point sine and cosine trig operations are supported by expansion to
other operations, as indicated by the third parameter, <tt>Expand</tt>, to
<tt>setOperationAction</tt>:</p>
</div>

<div class="doc_code">
<pre>setOperationAction(ISD::FSIN, MVT::f32, Expand);
setOperationAction(ISD::FCOS, MVT::f32, Expand);
</pre>
</div>

<!-- _______________________________________________________________________ -->
<div class="doc_subsubsection">
  <a name="custom">Custom</a>
</div>
<div class="doc_text">
<p>For some operations, simple type promotion or operation expansion
may be insufficient. In some cases, a special intrinsic function must be
implemented. </p>

<p>For example, a constant value may require special treatment, or
an operation may require spilling and restoring registers in the stack and
working with register allocators. </p>

<p>As seen in <tt>SparcISelLowering.cpp</tt> code below, to perform a type
conversion from a floating point value to a signed integer, first the
<tt>setOperationAction</tt> should be called with <tt>Custom</tt> as the third parameter:</p>
</div>

<div class="doc_code">
<pre>setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
</pre>
</div>    
<div class="doc_text">
<p>In the <tt>LowerOperation</tt> method, for each <tt>Custom</tt> operation, a case
statement should be added to indicate what function to call. In the following
code, an FP_TO_SINT opcode will call the <tt>LowerFP_TO_SINT</tt> method:</p>
</div>

<div class="doc_code">
<pre>SDValue SparcTargetLowering::LowerOperation(
                               SDValue Op, SelectionDAG &amp;DAG) {
  switch (Op.getOpcode()) {
  case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
  ...
  }
}
</pre>
</div>        
<div class="doc_text">
<p>Finally, the <tt>LowerFP_TO_SINT</tt> method is implemented, using an FP
register to convert the floating-point value to an integer.</p>
</div>

<div class="doc_code">
<pre>static SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &amp;DAG) {
assert(Op.getValueType() == MVT::i32);
  Op = DAG.getNode(SPISD::FTOI, MVT::f32, Op.getOperand(0));
  return DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
}
</pre>
</div>    
<!-- _______________________________________________________________________ -->
<div class="doc_subsubsection">
  <a name="legal">Legal</a>
</div>
<div class="doc_text">
<p>The <tt>Legal</tt> LegalizeAction enum value simply indicates that an
operation <b>is</b> natively supported. <tt>Legal</tt> represents the default condition,
so it is rarely used. In <tt>SparcISelLowering.cpp</tt>, the action for CTPOP (an
operation to count the bits set in an integer) is natively supported only for
SPARC v9. The following code enables the <tt>Expand</tt> conversion technique for non-v9
SPARC implementations.</p>
</div>

<div class="doc_code">
<pre>setOperationAction(ISD::CTPOP, MVT::i32, Expand);
...
if (TM.getSubtarget&lt;SparcSubtarget&gt;().isV9())
  setOperationAction(ISD::CTPOP, MVT::i32, Legal);
  case ISD::SETULT: return SPCC::ICC_CS;
  case ISD::SETULE: return SPCC::ICC_LEU;
  case ISD::SETUGT: return SPCC::ICC_GU;
  case ISD::SETUGE: return SPCC::ICC_CC;
  }
}
</pre>
</div>
<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="callingConventions">Calling Conventions</a>
</div>
<div class="doc_text">
<p>To support target-specific calling conventions, <tt>XXXGenCallingConv.td</tt>
uses interfaces (such as CCIfType and CCAssignToReg) that are defined in
<tt>lib/Target/TargetCallingConv.td</tt>. TableGen can take the target descriptor file
<tt>XXXGenCallingConv.td</tt> and generate the header file <tt>XXXGenCallingConv.inc</tt>, which
is typically included in <tt>XXXISelLowering.cpp</tt>. You can use the interfaces in
<tt>TargetCallingConv.td</tt> to specify:</p>
<ul>
<li>the order of parameter allocation</li>

<li>where parameters and return values are placed (that is, on the
stack or in registers)</li>

<li>which registers may be used</li>

<li>whether the caller or callee unwinds the stack</li>
</ul>

<p>The following example demonstrates the use of the CCIfType and
CCAssignToReg interfaces. If the CCIfType predicate is true (that is, if the
current argument is of type f32 or f64), then the action is performed. In this
case, the CCAssignToReg action assigns the argument value to the first
available register: either R0 or R1.  </p>
</div>
<div class="doc_code">
<pre>CCIfType&lt;[f32,f64], CCAssignToReg&lt;[R0, R1]&gt;&gt;
</pre>
</div>
<div class="doc_text">
<p><tt>SparcCallingConv.td</tt> contains definitions for a target-specific return-value
calling convention (RetCC_Sparc32) and a basic 32-bit C calling convention
(CC_Sparc32). The definition of RetCC_Sparc32 (shown below) indicates which
registers are used for specified scalar return types. A single-precision float
is returned to register F0, and a double-precision float goes to register D0. A
32-bit integer is returned in register I0 or I1. </p>
</div>

<div class="doc_code">
<pre>def RetCC_Sparc32 : CallingConv&lt;[
  CCIfType&lt;[i32], CCAssignToReg&lt;[I0, I1]&gt;&gt;,
  CCIfType&lt;[f32], CCAssignToReg&lt;[F0]&gt;&gt;,
  CCIfType&lt;[f64], CCAssignToReg&lt;[D0]&gt;&gt;
]&gt;;
</pre>
</div>
<div class="doc_text">
<p>The definition of CC_Sparc32 in <tt>SparcCallingConv.td</tt> introduces
CCAssignToStack, which assigns the value to a stack slot with the specified size
and alignment. In the example below, the first parameter, 4, indicates the size
of the slot, and the second parameter, also 4, indicates the stack alignment
along 4-byte units. (Special cases: if size is zero, then the ABI size is used;
if alignment is zero, then the ABI alignment is used.) </p>
</div>

<div class="doc_code">
<pre>def CC_Sparc32 : CallingConv&lt;[
  // All arguments get passed in integer registers if there is space.
  CCIfType&lt;[i32, f32, f64], CCAssignToReg&lt;[I0, I1, I2, I3, I4, I5]&gt;&gt;,
  CCAssignToStack&lt;4, 4&gt;
]&gt;;
</pre>
</div>
<div class="doc_text">
<p>CCDelegateTo is another commonly used interface, which tries to find
a specified sub-calling convention and, if a match is found, it is invoked. In
the following example (in <tt>X86CallingConv.td</tt>), the definition of RetCC_X86_32_C
ends with CCDelegateTo. After the current value is assigned to the register ST0
or ST1, the RetCC_X86Common is invoked.</p>
</div>

<div class="doc_code">
<pre>def RetCC_X86_32_C : CallingConv&lt;[
  CCIfType&lt;[f32], CCAssignToReg&lt;[ST0, ST1]&gt;&gt;,
  CCIfType&lt;[f64], CCAssignToReg&lt;[ST0, ST1]&gt;&gt;,
  CCDelegateTo&lt;RetCC_X86Common&gt;
]&gt;;
</pre>
</div>
<div class="doc_text">
<p>CCIfCC is an interface that attempts to match the given name to
the current calling convention. If the name identifies the current calling
convention, then a specified action is invoked. In the following example (in
<tt>X86CallingConv.td</tt>), if the Fast calling convention is in use, then RetCC_X86_32_Fast
is invoked. If the SSECall calling convention is in use, then RetCC_X86_32_SSE
is invoked. </p>
</div>

<div class="doc_code">
<pre>def RetCC_X86_32 : CallingConv&lt;[
  CCIfCC&lt;&quot;CallingConv::Fast&quot;, CCDelegateTo&lt;RetCC_X86_32_Fast&gt;&gt;,
  CCIfCC&lt;&quot;CallingConv::X86_SSECall&quot;, CCDelegateTo&lt;RetCC_X86_32_SSE&gt;&gt;,
  CCDelegateTo&lt;RetCC_X86_32_C&gt;
]&gt;;
</pre>
</div>
<div class="doc_text">
<p>Other calling convention interfaces include:</p>
<ul>
<li>CCIf &lt;predicate, action&gt; - if the predicate matches, apply
the action</li>

<li>CCIfInReg &lt;action&gt; - if the argument is marked with the
&lsquo;inreg&rsquo; attribute, then apply the action </li>

<li>CCIfNest &lt;action&gt; - if the argument is marked with the
&lsquo;nest&rsquo; attribute, then apply the action</li>

<li>CCIfNotVarArg &lt;action&gt; - if the current function does not
take a variable number of arguments, apply the action</li>

<li>CCAssignToRegWithShadow &lt;registerList, shadowList&gt; -
similar to CCAssignToReg, but with a shadow list of registers</li>

<li>CCPassByVal &lt;size, align&gt; - assign value to a stack slot
with the minimum specified size and alignment </li>

<li>CCPromoteToType &lt;type&gt; - promote the current value to the specified
type</li>

<li>CallingConv &lt;[actions]&gt; - define each calling convention
that is supported</li>
</ul>
</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="assemblyPrinter">Assembly Printer</a>
</div>
<!-- *********************************************************************** -->

<div class="doc_text">
<p>During the code
emission stage, the code generator may utilize an LLVM pass to produce assembly
output. To do this, you want to implement the code for a printer that converts
LLVM IR to a GAS-format assembly language for your target machine, using the
following steps:</p>
<ul>
<li>Define all the assembly strings for your target, adding them to
the instructions defined in the <tt>XXXInstrInfo.td</tt> file. 
(See <a href="#InstructionSet">Instruction Set</a>.)
TableGen will produce an output file (<tt>XXXGenAsmWriter.inc</tt>) with an
implementation of the <tt>printInstruction</tt> method for the XXXAsmPrinter class.</li>

<li>Write <tt>XXXTargetAsmInfo.h</tt>, which contains the bare-bones
declaration of the XXXTargetAsmInfo class (a subclass of TargetAsmInfo). </li>

<li>Write <tt>XXXTargetAsmInfo.cpp</tt>, which contains target-specific values
for TargetAsmInfo properties and sometimes new implementations for methods</li>

<li>Write <tt>XXXAsmPrinter.cpp</tt>, which implements the AsmPrinter class
that performs the LLVM-to-assembly conversion. </li>
</ul>

<p>The code in <tt>XXXTargetAsmInfo.h</tt> is usually a trivial declaration
of the XXXTargetAsmInfo class for use in <tt>XXXTargetAsmInfo.cpp</tt>. Similarly,
<tt>XXXTargetAsmInfo.cpp</tt> usually has a few declarations of XXXTargetAsmInfo replacement
values that override the default values in <tt>TargetAsmInfo.cpp</tt>. For example in
<tt>SparcTargetAsmInfo.cpp</tt>, </p>
</div>

<div class="doc_code">
<pre>SparcTargetAsmInfo::SparcTargetAsmInfo(const SparcTargetMachine &amp;TM) {
  Data16bitsDirective = &quot;\t.half\t&quot;;
  Data32bitsDirective = &quot;\t.word\t&quot;;
  Data64bitsDirective = 0;  // .xword is only supported by V9.
  ZeroDirective = &quot;\t.skip\t&quot;;
  CommentString = &quot;!&quot;;
  ConstantPoolSection = &quot;\t.section \&quot;.rodata\&quot;,#alloc\n&quot;;
}
</pre>
</div>
<div class="doc_text">
<p>The X86 assembly printer implementation (X86TargetAsmInfo) is an
example where the target specific TargetAsmInfo class uses overridden methods:
<tt>ExpandInlineAsm</tt> and <tt>PreferredEHDataFormat</tt>. </p>

<p>A target-specific implementation of AsmPrinter is written in
<tt>XXXAsmPrinter.cpp</tt>, which implements the AsmPrinter class that converts the LLVM
to printable assembly. The implementation must include the following headers
that have declarations for the AsmPrinter and MachineFunctionPass classes. The
MachineFunctionPass is a subclass of FunctionPass. </p>
</div>

<div class="doc_code">
<pre>#include &quot;llvm/CodeGen/AsmPrinter.h&quot;
#include &quot;llvm/CodeGen/MachineFunctionPass.h&quot; 
</pre>
</div>

<div class="doc_text">
<p>As a FunctionPass, AsmPrinter first calls <tt>doInitialization</tt> to set
up the AsmPrinter. In SparcAsmPrinter, a Mangler object is instantiated to
process variable names.</p>

<p>In <tt>XXXAsmPrinter.cpp</tt>, the <tt>runOnMachineFunction</tt> method (declared
in MachineFunctionPass) must be implemented for XXXAsmPrinter. In
MachineFunctionPass, the <tt>runOnFunction</tt> method invokes <tt>runOnMachineFunction</tt>.
Target-specific implementations of <tt>runOnMachineFunction</tt> differ, but generally
do the following to process each machine function:</p>
<ul>
<li>call <tt>SetupMachineFunction</tt> to perform initialization</li>

<li>call <tt>EmitConstantPool</tt> to print out (to the output stream)
constants which have been spilled to memory </li>

<li>call <tt>EmitJumpTableInfo</tt> to print out jump tables used by the
current function </li>

<li>print out the label for the current function</li>

<li>print out the code for the function, including basic block labels
and the assembly for the instruction (using <tt>printInstruction</tt>)</li>
</ul>
<p>The XXXAsmPrinter implementation must also include the code
generated by TableGen that is output in the <tt>XXXGenAsmWriter.inc</tt> file. The code
in <tt>XXXGenAsmWriter.inc</tt> contains an implementation of the <tt>printInstruction</tt>
method that may call these methods:</p>
<ul>
<li><tt>printOperand</tt></li>

<li><tt>printMemOperand</tt></li>

<li><tt>printCCOperand (for conditional statements)</tt></li>

<li><tt>printDataDirective</tt></li>

<li><tt>printDeclare</tt></li>

<li><tt>printImplicitDef</tt></li>

<li><tt>printInlineAsm</tt></li>

<li><tt>printLabel</tt></li>

<li><tt>printPICJumpTableEntry</tt></li>

<li><tt>printPICJumpTableSetLabel</tt></li>
</ul>

<p>The implementations of <tt>printDeclare</tt>, <tt>printImplicitDef</tt>,
<tt>printInlineAsm</tt>, and <tt>printLabel</tt> in <tt>AsmPrinter.cpp</tt> are generally adequate for
printing assembly and do not need to be overridden. (<tt>printBasicBlockLabel</tt> is
another method that is implemented in <tt>AsmPrinter.cpp</tt> that may be directly used
in an implementation of XXXAsmPrinter.)</p>

<p>The <tt>printOperand</tt> method is implemented with a long switch/case
statement for the type of operand: register, immediate, basic block, external
symbol, global address, constant pool index, or jump table index. For an
instruction with a memory address operand, the <tt>printMemOperand</tt> method should be
implemented to generate the proper output. Similarly, <tt>printCCOperand</tt> should be
used to print a conditional operand. </p>

<p><tt>doFinalization</tt> should be overridden in XXXAsmPrinter, and
it should be called to shut down the assembly printer. During <tt>doFinalization</tt>,
global variables and constants are printed to output.</p>
</div>
<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="subtargetSupport">Subtarget Support</a>
</div>
<!-- *********************************************************************** -->

<div class="doc_text">
<p>Subtarget support is used to inform the code generation process
of instruction set variations for a given chip set.  For example, the LLVM
SPARC implementation provided covers three major versions of the SPARC
microprocessor architecture: Version 8 (V8, which is a 32-bit architecture),
Version 9 (V9, a 64-bit architecture), and the UltraSPARC architecture. V8 has
16 double-precision floating-point registers that are also usable as either 32
single-precision or 8 quad-precision registers.  V8 is also purely big-endian. V9
has 32 double-precision floating-point registers that are also usable as 16
quad-precision registers, but cannot be used as single-precision registers. The
UltraSPARC architecture combines V9 with UltraSPARC Visual Instruction Set
extensions.</p>

<p>If subtarget support is needed, you should implement a
target-specific XXXSubtarget class for your architecture. This class should
process the command-line options <tt>&#8211;mcpu=</tt> and <tt>&#8211;mattr=</tt></p>

<p>TableGen uses definitions in the <tt>Target.td</tt> and <tt>Sparc.td</tt> files to
generate code in <tt>SparcGenSubtarget.inc</tt>. In <tt>Target.td</tt>, shown below, the
SubtargetFeature interface is defined. The first 4 string parameters of the
SubtargetFeature interface are a feature name, an attribute set by the feature,
the value of the attribute, and a description of the feature. (The fifth
parameter is a list of features whose presence is implied, and its default
value is an empty array.)</p>
</div>

<div class="doc_code">
<pre>class SubtargetFeature&lt;string n, string a,  string v, string d,
                       list&lt;SubtargetFeature&gt; i = []&gt; {
  string Name = n;
  string Attribute = a;
  string Value = v;
  string Desc = d;
  list&lt;SubtargetFeature&gt; Implies = i;
}
</pre>
</div>
<div class="doc_text">
<p>In the <tt>Sparc.td</tt> file, the SubtargetFeature is used to define the
following features.  </p>
</div>

<div class="doc_code">
<pre>def FeatureV9 : SubtargetFeature&lt;&quot;v9&quot;, &quot;IsV9&quot;, &quot;true&quot;,
                     &quot;Enable SPARC-V9 instructions&quot;&gt;;
def FeatureV8Deprecated : SubtargetFeature&lt;&quot;deprecated-v8&quot;, 
                     &quot;V8DeprecatedInsts&quot;, &quot;true&quot;,
                     &quot;Enable deprecated V8 instructions in V9 mode&quot;&gt;;
def FeatureVIS : SubtargetFeature&lt;&quot;vis&quot;, &quot;IsVIS&quot;, &quot;true&quot;,
                     &quot;Enable UltraSPARC Visual Instruction Set extensions&quot;&gt;;
</pre>
</div>

<div class="doc_text">
<p>Elsewhere in <tt>Sparc.td</tt>, the Proc class is defined and then is used
to define particular SPARC processor subtypes that may have the previously
described features. </p>
</div>

<div class="doc_code">
<pre>class Proc&lt;string Name, list&lt;SubtargetFeature&gt; Features&gt;
 : Processor&lt;Name, NoItineraries, Features&gt;;
&nbsp;
def : Proc&lt;&quot;generic&quot;,         []&gt;;
def : Proc&lt;&quot;v8&quot;,              []&gt;;
def : Proc&lt;&quot;supersparc&quot;,      []&gt;;
def : Proc&lt;&quot;sparclite&quot;,       []&gt;;
def : Proc&lt;&quot;f934&quot;,            []&gt;;
def : Proc&lt;&quot;hypersparc&quot;,      []&gt;;
def : Proc&lt;&quot;sparclite86x&quot;,    []&gt;;
def : Proc&lt;&quot;sparclet&quot;,        []&gt;;
def : Proc&lt;&quot;tsc701&quot;,          []&gt;;
def : Proc&lt;&quot;v9&quot;,              [FeatureV9]&gt;;
def : Proc&lt;&quot;ultrasparc&quot;,      [FeatureV9, FeatureV8Deprecated]&gt;;
def : Proc&lt;&quot;ultrasparc3&quot;,     [FeatureV9, FeatureV8Deprecated]&gt;;
def : Proc&lt;&quot;ultrasparc3-vis&quot;, [FeatureV9, FeatureV8Deprecated, FeatureVIS]&gt;;
</pre>
</div>

<div class="doc_text">
<p>From <tt>Target.td</tt> and <tt>Sparc.td</tt> files, the resulting
SparcGenSubtarget.inc specifies enum values to identify the features, arrays of
constants to represent the CPU features and CPU subtypes, and the
ParseSubtargetFeatures method that parses the features string that sets
specified subtarget options. The generated <tt>SparcGenSubtarget.inc</tt> file should be
included in the <tt>SparcSubtarget.cpp</tt>. The target-specific implementation of the XXXSubtarget
method should follow this pseudocode:</p>
</div>

<div class="doc_code">
<pre>XXXSubtarget::XXXSubtarget(const Module &amp;M, const std::string &amp;FS) {
  // Set the default features
  // Determine default and user specified characteristics of the CPU
  // Call ParseSubtargetFeatures(FS, CPU) to parse the features string
  // Perform any additional operations
}
</pre>
</div>

<!-- *********************************************************************** -->
<div class="doc_section">
  <a name="jitSupport">JIT Support</a>
</div>
<!-- *********************************************************************** -->

<div class="doc_text">
<p>The implementation of a target machine optionally includes a Just-In-Time
(JIT) code generator that emits machine code and auxiliary structures as binary
output that can be written directly to memory. 
To do this, implement JIT code generation by performing the following
steps:</p>
<ul>
<li>Write an <tt>XXXCodeEmitter.cpp</tt> file that contains a machine function
pass that transforms target-machine instructions into relocatable machine code.</li>

<li>Write an <tt>XXXJITInfo.cpp</tt> file that implements the JIT interfaces
for target-specific code-generation
activities, such as emitting machine code and stubs. </li>

<li>Modify XXXTargetMachine so that it provides a TargetJITInfo
object through its <tt>getJITInfo</tt> method. </li>
</ul>

<p>There are several different approaches to writing the JIT support
code. For instance, TableGen and target descriptor files may be used for
creating a JIT code generator, but are not mandatory. For the Alpha and PowerPC
target machines, TableGen is used to generate <tt>XXXGenCodeEmitter.inc</tt>, which
contains the binary coding of machine instructions and the
<tt>getBinaryCodeForInstr</tt> method to access those codes. Other JIT implementations
do not.</p>

<p>Both <tt>XXXJITInfo.cpp</tt> and <tt>XXXCodeEmitter.cpp</tt> must include the
<tt>llvm/CodeGen/MachineCodeEmitter.h</tt> header file that defines the MachineCodeEmitter
class containing code for several callback functions that write data (in bytes,
words, strings, etc.) to the output stream.</p>
</div>
<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="mce">Machine Code Emitter</a>
</div>

<div class="doc_text">
<p>In <tt>XXXCodeEmitter.cpp</tt>, a target-specific of the Emitter class is
implemented as a function pass (subclass of MachineFunctionPass). The
target-specific implementation of <tt>runOnMachineFunction</tt> (invoked by
<tt>runOnFunction</tt> in MachineFunctionPass) iterates through the MachineBasicBlock
calls <tt>emitInstruction</tt> to process each instruction and emit binary code. <tt>emitInstruction</tt>
is largely implemented with case statements on the instruction types defined in
<tt>XXXInstrInfo.h</tt>. For example, in <tt>X86CodeEmitter.cpp</tt>, the <tt>emitInstruction</tt> method
is built around the following switch/case statements:</p>
</div>

<div class="doc_code">
<pre>switch (Desc-&gt;TSFlags &amp; X86::FormMask) {
case X86II::Pseudo:  // for not yet implemented instructions 
   ...               // or pseudo-instructions
   break;
case X86II::RawFrm:  // for instructions with a fixed opcode value
   ...
   break;
case X86II::AddRegFrm: // for instructions that have one register operand 
   ...                 // added to their opcode
   break;
case X86II::MRMDestReg:// for instructions that use the Mod/RM byte
   ...                 // to specify a destination (register)
   break;
case X86II::MRMDestMem:// for instructions that use the Mod/RM byte
   ...                 // to specify a destination (memory)
   break;
case X86II::MRMSrcReg: // for instructions that use the Mod/RM byte
   ...                 // to specify a source (register)
   break;
case X86II::MRMSrcMem: // for instructions that use the Mod/RM byte
   ...                 // to specify a source (memory)
   break;
case X86II::MRM0r: case X86II::MRM1r:  // for instructions that operate on 
case X86II::MRM2r: case X86II::MRM3r:  // a REGISTER r/m operand and
case X86II::MRM4r: case X86II::MRM5r:  // use the Mod/RM byte and a field
case X86II::MRM6r: case X86II::MRM7r:  // to hold extended opcode data
   ...  
   break;
case X86II::MRM0m: case X86II::MRM1m:  // for instructions that operate on
case X86II::MRM2m: case X86II::MRM3m:  // a MEMORY r/m operand and
case X86II::MRM4m: case X86II::MRM5m:  // use the Mod/RM byte and a field
case X86II::MRM6m: case X86II::MRM7m:  // to hold extended opcode data
   ...  
   break;
case X86II::MRMInitReg: // for instructions whose source and
   ...                  // destination are the same register
   break;
}
</pre>
</div>
<div class="doc_text">
<p>The implementations of these case statements often first emit the
opcode and then get the operand(s). Then depending upon the operand, helper
methods may be called to process the operand(s). For example, in <tt>X86CodeEmitter.cpp</tt>,
for the <tt>X86II::AddRegFrm</tt> case, the first data emitted (by <tt>emitByte</tt>) is the
opcode added to the register operand. Then an object representing the machine
operand, MO1, is extracted. The helper methods such as <tt>isImmediate</tt>,
<tt>isGlobalAddress</tt>, <tt>isExternalSymbol</tt>, <tt>isConstantPoolIndex</tt>, and 
<tt>isJumpTableIndex</tt>
determine the operand type. (<tt>X86CodeEmitter.cpp</tt> also has private methods such
as <tt>emitConstant</tt>, <tt>emitGlobalAddress</tt>, 
<tt>emitExternalSymbolAddress</tt>, <tt>emitConstPoolAddress</tt>,
and <tt>emitJumpTableAddress</tt> that emit the data into the output stream.) </p>
</div>

<div class="doc_code">
<pre>case X86II::AddRegFrm:
  MCE.emitByte(BaseOpcode + getX86RegNum(MI.getOperand(CurOp++).getReg()));
  
  if (CurOp != NumOps) {
    const MachineOperand &amp;MO1 = MI.getOperand(CurOp++);
    unsigned Size = X86InstrInfo::sizeOfImm(Desc);
    if (MO1.isImmediate())
      emitConstant(MO1.getImm(), Size);
    else {
      unsigned rt = Is64BitMode ? X86::reloc_pcrel_word
        : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word);
      if (Opcode == X86::MOV64ri) 
        rt = X86::reloc_absolute_dword;  // FIXME: add X86II flag?
      if (MO1.isGlobalAddress()) {
        bool NeedStub = isa&lt;Function&gt;(MO1.getGlobal());
        bool isLazy = gvNeedsLazyPtr(MO1.getGlobal());
        emitGlobalAddress(MO1.getGlobal(), rt, MO1.getOffset(), 0,
                          NeedStub, isLazy);
      } else if (MO1.isExternalSymbol())
        emitExternalSymbolAddress(MO1.getSymbolName(), rt);
      else if (MO1.isConstantPoolIndex())
        emitConstPoolAddress(MO1.getIndex(), rt);
      else if (MO1.isJumpTableIndex())
        emitJumpTableAddress(MO1.getIndex(), rt);
    }
  }
  break;
</pre>
</div>
<div class="doc_text">
<p>In the previous example, <tt>XXXCodeEmitter.cpp</tt> uses the variable <tt>rt</tt>,
which is a RelocationType enum that may be used to relocate addresses (for
example, a global address with a PIC base offset). The RelocationType enum for
that target is defined in the short target-specific <tt>XXXRelocations.h</tt> file. The
RelocationType is used by the <tt>relocate</tt> method defined in <tt>XXXJITInfo.cpp</tt> to
rewrite addresses for referenced global symbols.</p>

<p>For example, <tt>X86Relocations.h</tt> specifies the following relocation
types for the X86 addresses. In all four cases, the relocated value is added to
the value already in memory. For <tt>reloc_pcrel_word</tt> and <tt>reloc_picrel_word</tt>, 
there is an additional initial adjustment.</p>
</div>

<div class="doc_code">
<pre>enum RelocationType {
  reloc_pcrel_word = 0,  // add reloc value after adjusting for the PC loc
  reloc_picrel_word = 1, // add reloc value after adjusting for the PIC base
  reloc_absolute_word = 2, // absolute relocation; no additional adjustment 
  reloc_absolute_dword = 3 // absolute relocation; no additional adjustment
};
</pre>
</div>
<!-- ======================================================================= -->
<div class="doc_subsection">
  <a name="targetJITInfo">Target JIT Info</a>
</div>
<div class="doc_text">
<p><tt>XXXJITInfo.cpp</tt> implements the JIT interfaces for target-specific code-generation
activities, such as emitting machine code and stubs. At minimum, 
a target-specific version of XXXJITInfo implements the following:</p>
<ul>
<li><tt>getLazyResolverFunction</tt> &#8211; initializes the JIT, gives the
target a function that is used for compilation </li>

<li><tt>emitFunctionStub</tt> &#8211; returns a native function with a
specified address for a callback function</li>

<li><tt>relocate</tt> &#8211; changes the addresses of referenced globals,
based on relocation types</li>

<li>callback function that are wrappers to a function stub that is
used when the real target is not initially known </li>
</ul>

<p><tt>getLazyResolverFunction</tt> is generally trivial to implement. It
makes the incoming parameter as the global JITCompilerFunction and returns the
callback function that will be used a function wrapper. For the Alpha target
(in <tt>AlphaJITInfo.cpp</tt>), the <tt>getLazyResolverFunction</tt> implementation is simply:</p>
</div>

<div class="doc_code">
<pre>TargetJITInfo::LazyResolverFn AlphaJITInfo::getLazyResolverFunction(  
                                            JITCompilerFn F) 
{
  JITCompilerFunction = F;
  return AlphaCompilationCallback;
}
</pre>
</div>
<div class="doc_text">
<p>For the X86 target, the <tt>getLazyResolverFunction</tt> implementation is
a little more complication, because it returns a different callback function
for processors with SSE instructions and XMM registers. </p>

<p>The callback function initially saves and later restores the
callee register values, incoming arguments, and frame and return address. The
callback function needs low-level access to the registers or stack, so it is typically
implemented with assembler. </p>
</div>

<!-- *********************************************************************** -->

<hr>
<address>
  <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
  src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
  <a href="http://validator.w3.org/check/referer"><img
  src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>

  <a href="http://www.woo.com">Mason Woo</a> and <a href="http://misha.brukman.net">Misha Brukman</a><br>
  <a href="http://llvm.org">The LLVM Compiler Infrastructure</a>
  <br>
  Last modified: $Date$
</address>

</body>
</html>