<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
"http://www.w3.org/TR/html4/strict.dtd">
<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
<title>Writing an LLVM Compiler Backend</title>
<link rel="stylesheet" href="llvm.css" type="text/css">
</head>
<body>
<h1>
Writing an LLVM Compiler Backend
</h1>
<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="#TargetRegistration">Target Registration</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>
<!-- *********************************************************************** -->
<h2>
<a name="intro">Introduction</a>
</h2>
<!-- *********************************************************************** -->
<div>
<p>
This document describes techniques for writing compiler backends that convert
the LLVM Intermediate Representation (IR) 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, ARM,
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>
<h3>
<a name="Audience">Audience</a>
</h3>
<div>
<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>
<h3>
<a name="Prerequisite">Prerequisite Reading</a>
</h3>
<div>
<p>
These essential documents must be read before reading this document:
</p>
<ul>
<li><i><a href="LangRef.html">LLVM Language Reference
Manual</a></i> — a reference manual for the LLVM assembly language.</li>
<li><i><a href="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 into 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="TableGenFundamentals.html">TableGen
Fundamentals</a></i> —a document that describes the TableGen
(<tt>tblgen</tt>) application that manages domain-specific information to
support LLVM code generation. TableGen processes input from a target
description file (<tt>.td</tt> suffix) and generates C++ code that can be
used for code generation.</li>
<li><i><a href="WritingAnLLVMPass.html">Writing an LLVM
Pass</a></i> — The assembly printer is a <tt>FunctionPass</tt>, as are
several SelectionDAG processing steps.</li>
</ul>
<p>
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
(<tt>GAS</tt>), 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 a list of target
machine dependent features.
</p>
</div>
<h3>
<a name="Basic">Basic Steps</a>
</h3>
<div>
<p>
To write a compiler backend for LLVM that converts the LLVM IR 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 "Sparc" 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 the 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 Directed Acyclic
Graph (DAG) 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 (i.e., 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 <tt>.cpp</tt> and <tt>.h</tt>. 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>
<h3>
<a name="Preliminaries">Preliminaries</a>
</h3>
<div>
<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="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 "Dummy," 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 <tt>LLVMDummy</tt> (for example, see the MIPS target). Alternatively, you
can split the library into <tt>LLVMDummyCodeGen</tt>
and <tt>LLVMDummyAsmPrinter</tt>, 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 a
lot 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
<tt>TargetMachine</tt>. 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 <tt>LLVMTargetMachine</tt>. (To create a
target from scratch, create a subclass of <tt>TargetMachine</tt>.)
</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>
<!-- *********************************************************************** -->
<h2>
<a name="TargetMachine">Target Machine</a>
</h2>
<!-- *********************************************************************** -->
<div>
<p>
<tt>LLVMTargetMachine</tt> is designed as a base class for targets implemented
with the LLVM target-independent code generator. The <tt>LLVMTargetMachine</tt>
class should be specialized by a concrete target class that implements the
various virtual methods. <tt>LLVMTargetMachine</tt> is defined as a subclass of
<tt>TargetMachine</tt> in <tt>include/llvm/Target/TargetMachine.h</tt>. The
<tt>TargetMachine</tt> class implementation (<tt>TargetMachine.cpp</tt>) also
processes numerous command-line options.
</p>
<p>
To create a concrete target-specific subclass of <tt>LLVMTargetMachine</tt>,
start by copying an existing <tt>TargetMachine</tt> 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 <tt>XXX</tt>, the implementation of
<tt>XXXTargetMachine</tt> 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. <tt>XXXTargetMachine</tt> 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 class="doc_code">
<pre>
namespace llvm {
class Module;
class SparcTargetMachine : public LLVMTargetMachine {
const TargetData DataLayout; // Calculates type size & alignment
SparcSubtarget Subtarget;
SparcInstrInfo InstrInfo;
TargetFrameInfo FrameInfo;
protected:
virtual const TargetAsmInfo *createTargetAsmInfo() const;
public:
SparcTargetMachine(const Module &M, const std::string &FS);
virtual const SparcInstrInfo *getInstrInfo() const {return &InstrInfo; }
virtual const TargetFrameInfo *getFrameInfo() const {return &FrameInfo; }
virtual const TargetSubtarget *getSubtargetImpl() const{return &Subtarget; }
virtual const TargetRegisterInfo *getRegisterInfo() const {
return &InstrInfo.getRegisterInfo();
}
virtual const TargetData *getTargetData() const { return &DataLayout; }
static unsigned getModuleMatchQuality(const Module &M);
// Pass Pipeline Configuration
virtual bool addInstSelector(PassManagerBase &PM, bool Fast);
virtual bool addPreEmitPass(PassManagerBase &PM, bool Fast);
};
} // end namespace llvm
</pre>
</div>
<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 <tt>XXXTargetMachine</tt> constructor should specify a
<tt>TargetDescription</tt> 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 class="doc_code">
<pre>
SparcTargetMachine::SparcTargetMachine(const Module &M, const std::string &FS)
: DataLayout("E-p:32:32-f128:128:128"),
Subtarget(M, FS), InstrInfo(Subtarget),
FrameInfo(TargetFrameInfo::StackGrowsDown, 8, 0) {
}
</pre>
</div>
<p>Hyphens separate portions of the <tt>TargetDescription</tt> string.</p>
<ul>
<li>An upper-case "<tt>E</tt>" in the string indicates a big-endian target data
model. a lower-case "<tt>e</tt>" indicates little-endian.</li>
<li>"<tt>p:</tt>" is followed by pointer information: size, ABI alignment, and
preferred alignment. If only two figures follow "<tt>p:</tt>", 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: "<tt>i</tt>", "<tt>f</tt>",
"<tt>v</tt>", or "<tt>a</tt>" (corresponding to integer, floating point,
vector, or aggregate). "<tt>i</tt>", "<tt>v</tt>", or "<tt>a</tt>" are
followed by ABI alignment and preferred alignment. "<tt>f</tt>" is followed
by three values: the first indicates the size of a long double, then ABI
alignment, and then ABI preferred alignment.</li>
</ul>
</div>
<!-- *********************************************************************** -->
<h2>
<a name="TargetRegistration">Target Registration</a>
</h2>
<!-- *********************************************************************** -->
<div>
<p>
You must also register your target with the <tt>TargetRegistry</tt>, which is
what other LLVM tools use to be able to lookup and use your target at
runtime. The <tt>TargetRegistry</tt> can be used directly, but for most targets
there are helper templates which should take care of the work for you.</p>
<p>
All targets should declare a global <tt>Target</tt> object which is used to
represent the target during registration. Then, in the target's TargetInfo
library, the target should define that object and use
the <tt>RegisterTarget</tt> template to register the target. For example, the Sparc registration code looks like this:
</p>
<div class="doc_code">
<pre>
Target llvm::TheSparcTarget;
extern "C" void LLVMInitializeSparcTargetInfo() {
RegisterTarget<Triple::sparc, /*HasJIT=*/false>
X(TheSparcTarget, "sparc", "Sparc");
}
</pre>
</div>
<p>
This allows the <tt>TargetRegistry</tt> to look up the target by name or by
target triple. In addition, most targets will also register additional features
which are available in separate libraries. These registration steps are
separate, because some clients may wish to only link in some parts of the target
-- the JIT code generator does not require the use of the assembler printer, for
example. Here is an example of registering the Sparc assembly printer:
</p>
<div class="doc_code">
<pre>
extern "C" void LLVMInitializeSparcAsmPrinter() {
RegisterAsmPrinter<SparcAsmPrinter> X(TheSparcTarget);
}
</pre>
</div>
<p>
For more information, see
"<a href="/doxygen/TargetRegistry_8h-source.html">llvm/Target/TargetRegistry.h</a>".
</p>
</div>
<!-- *********************************************************************** -->
<h2>
<a name="RegisterSet">Register Set and Register Classes</a>
</h2>
<!-- *********************************************************************** -->
<div>
<p>
You should describe a concrete target-specific class that represents the
register file of a target machine. This class is called <tt>XXXRegisterInfo</tt>
(where <tt>XXX</tt> identifies the target) and represents the class register
file data that is used for register allocation. It 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 for
integer, floating-point, or vector registers. A 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 <tt>XXXRegisterInfo</tt> requires hand-coding.
</p>
<!-- ======================================================================= -->
<h3>
<a name="RegisterDef">Defining a Register</a>
</h3>
<div>
<p>
The <tt>XXXRegisterInfo.td</tt> file typically starts with register definitions
for a target machine. The <tt>Register</tt> class (specified
in <tt>Target.td</tt>) is used to define an object for each register. The
specified string <tt>n</tt> becomes the <tt>Name</tt> of the register. The
basic <tt>Register</tt> object does not have any subregisters and does not
specify any aliases.
</p>
<div class="doc_code">
<pre>
class Register<string n> {
string Namespace = "";
string AsmName = n;
string Name = n;
int SpillSize = 0;
int SpillAlignment = 0;
list<Register> Aliases = [];
list<Register> SubRegs = [];
list<int> DwarfNumbers = [];
}
</pre>
</div>
<p>
For example, in the <tt>X86RegisterInfo.td</tt> file, there are register
definitions that utilize the Register class, such as:
</p>
<div class="doc_code">
<pre>
def AL : Register<"AL">, DwarfRegNum<[0, 0, 0]>;
</pre>
</div>
<p>
This defines the register <tt>AL</tt> and assigns it values (with
<tt>DwarfRegNum</tt>) that are used by <tt>gcc</tt>, <tt>gdb</tt>, or a debug
information writer to identify a register. For register
<tt>AL</tt>, <tt>DwarfRegNum</tt> takes an array of 3 values representing 3
different modes: the first element is for X86-64, the second for exception
handling (EH) 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 class="doc_code">
<pre>
static const unsigned GR8[] = { X86::AL, ... };
const unsigned AL_AliasSet[] = { X86::AX, X86::EAX, X86::RAX, 0 };
const TargetRegisterDesc RegisterDescriptors[] = {
...
{ "AL", "AL", AL_AliasSet, Empty_SubRegsSet, Empty_SubRegsSet, AL_SuperRegsSet }, ...
</pre>
</div>
<p>
From the register info file, TableGen generates a <tt>TargetRegisterDesc</tt>
object for each register. <tt>TargetRegisterDesc</tt> is defined in
<tt>include/llvm/Target/TargetRegisterInfo.h</tt> with the following fields:
</p>
<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>
<p>
TableGen uses the entire target description file (<tt>.td</tt>) to determine
text names for the register (in the <tt>AsmName</tt> and <tt>Name</tt> fields of
<tt>TargetRegisterDesc</tt>) and the relationships of other registers to the
defined register (in the other <tt>TargetRegisterDesc</tt> fields). In this
example, other definitions establish the registers "<tt>AX</tt>",
"<tt>EAX</tt>", and "<tt>RAX</tt>" as aliases for one another, so TableGen
generates a null-terminated array (<tt>AL_AliasSet</tt>) for this register alias
set.
</p>
<p>
The <tt>Register</tt> class is commonly used as a base class for more complex
classes. In <tt>Target.td</tt>, the <tt>Register</tt> class is the base for the
<tt>RegisterWithSubRegs</tt> class that is used to define registers that need to
specify subregisters in the <tt>SubRegs</tt> list, as shown here:
</p>
<div class="doc_code">
<pre>
class RegisterWithSubRegs<string n,
list<Register> subregs> : Register<n> {
let SubRegs = subregs;
}
</pre>
</div>
<p>
In <tt>SparcRegisterInfo.td</tt>, additional register classes are defined for
SPARC: a Register subclass, SparcReg, and further subclasses: <tt>Ri</tt>,
<tt>Rf</tt>, and <tt>Rd</tt>. SPARC registers are identified by 5-bit ID
numbers, which is a feature common to these subclasses. Note the use of
'<tt>let</tt>' expressions to override values that are initially defined in a
superclass (such as <tt>SubRegs</tt> field in the <tt>Rd</tt> class).
</p>
<div class="doc_code">
<pre>
class SparcReg<string n> : Register<n> {
field bits<5> Num;
let Namespace = "SP";
}
// Ri - 32-bit integer registers
class Ri<bits<5> num, string n> :
SparcReg<n> {
let Num = num;
}
// Rf - 32-bit floating-point registers
class Rf<bits<5> num, string n> :
SparcReg<n> {
let Num = num;
}
// Rd - Slots in the FP register file for 64-bit
floating-point values.
class Rd<bits<5> num, string n,
list<Register> subregs> : SparcReg<n> {
let Num = num;
let SubRegs = subregs;
}
</pre>
</div>
<p>
In the <tt>SparcRegisterInfo.td</tt> file, there are register definitions that
utilize these subclasses of <tt>Register</tt>, such as:
</p>
<div class="doc_code">
<pre>
def G0 : Ri< 0, "G0">,
DwarfRegNum<[0]>;
def G1 : Ri< 1, "G1">, DwarfRegNum<[1]>;
...
def F0 : Rf< 0, "F0">,
DwarfRegNum<[32]>;
def F1 : Rf< 1, "F1">,
DwarfRegNum<[33]>;
...
def D0 : Rd< 0, "F0", [F0, F1]>,
DwarfRegNum<[32]>;
def D1 : Rd< 2, "F2", [F2, F3]>,
DwarfRegNum<[34]>;
</pre>
</div>
<p>
The last two registers shown above (<tt>D0</tt> and <tt>D1</tt>) 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's TargetRegisterDesc.
</p>
</div>
<!-- ======================================================================= -->
<h3>
<a name="RegisterClassDef">Defining a Register Class</a>
</h3>
<div>
<p>
The <tt>RegisterClass</tt> 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 class="doc_code">
<pre>
class RegisterClass<string namespace,
list<ValueType> regTypes, int alignment, dag regList> {
string Namespace = namespace;
list<ValueType> RegTypes = regTypes;
int Size = 0; // spill size, in bits; zero lets tblgen pick the size
int Alignment = alignment;
// 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;
dag MemberList = regList;
// for register classes that are subregisters of this class
list<RegisterClass> SubRegClassList = [];
code MethodProtos = [{}]; // to insert arbitrary code
code MethodBodies = [{}];
}
</pre>
</div>
<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 <tt>ValueType</tt> register type values
that are defined in <tt>include/llvm/CodeGen/ValueTypes.td</tt>. Defined
values include integer types (such as <tt>i16</tt>, <tt>i32</tt>,
and <tt>i1</tt> for Boolean), floating-point types
(<tt>f32</tt>, <tt>f64</tt>), and vector types (for example, <tt>v8i16</tt>
for an <tt>8 x i16</tt> vector). All registers in a <tt>RegisterClass</tt>
must have the same <tt>ValueType</tt>, 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 <tt>RegisterClass</tt> 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 alternative allocation order method is not specified, then
<tt>regList</tt> also defines the order of allocation used by the register
allocator. Besides simply listing registers with <tt>(add R0, R1, ...)</tt>,
more advanced set operators are available. See
<tt>include/llvm/Target/Target.td</tt> for more information.</li>
</ul>
<p>
In <tt>SparcRegisterInfo.td</tt>, three RegisterClass objects are defined:
<tt>FPRegs</tt>, <tt>DFPRegs</tt>, and <tt>IntRegs</tt>. For all three register
classes, the first argument defines the namespace with the string
'<tt>SP</tt>'. <tt>FPRegs</tt> defines a group of 32 single-precision
floating-point registers (<tt>F0</tt> to <tt>F31</tt>); <tt>DFPRegs</tt> defines
a group of 16 double-precision registers
(<tt>D0-D15</tt>).
</p>
<div class="doc_code">
<pre>
// F0, F1, F2, ..., F31
def FPRegs : RegisterClass<"SP", [f32], 32, (sequence "F%u", 0, 31)>;
def DFPRegs : RegisterClass<"SP", [f64], 64,
(add D0, D1, D2, D3, D4, D5, D6, D7, D8,
D9, D10, D11, D12, D13, D14, D15)>;
def IntRegs : RegisterClass<"SP", [i32], 32,
(add 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
)>;
</pre>
</div>
<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
<tt>SparcGenRegisterInfo</tt> that uses <tt>TargetRegisterInfo</tt> as its
base. It also specifies types, based upon the defined register
classes: <tt>DFPRegsClass</tt>, <tt>FPRegsClass</tt>, and <tt>IntRegsClass</tt>.
</p>
<p>
<tt>SparcRegisterInfo.td</tt> also generates <tt>SparcGenRegisterInfo.inc</tt>,
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 <tt>IntRegs</tt> reflects the order in the definition of <tt>IntRegs</tt> in
the target description file.
</p>
<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,
};
// IntRegsVTs Register Class Value Types...
static const MVT::ValueType IntRegsVTs[] = {
MVT::i32, MVT::Other
};
namespace SP { // Register class instances
DFPRegsClass DFPRegsRegClass;
FPRegsClass FPRegsRegClass;
IntRegsClass IntRegsRegClass;
...
// IntRegs Sub-register Classess...
static const TargetRegisterClass* const IntRegsSubRegClasses [] = {
NULL
};
...
// IntRegs Super-register Classess...
static const TargetRegisterClass* const IntRegsSuperRegClasses [] = {
NULL
};
...
// IntRegs Register Class sub-classes...
static const TargetRegisterClass* const IntRegsSubclasses [] = {
NULL
};
...
// IntRegs Register Class super-classes...
static const TargetRegisterClass* const IntRegsSuperclasses [] = {
NULL
};
IntRegsClass::IntRegsClass() : TargetRegisterClass(IntRegsRegClassID,
IntRegsVTs, IntRegsSubclasses, IntRegsSuperclasses, IntRegsSubRegClasses,
IntRegsSuperRegClasses, 4, 4, 1, IntRegs, IntRegs + 32) {}
}
</pre>
</div>
<p>
The register allocators will avoid using reserved registers, and callee saved
registers are not used until all the volatile registers have been used. That
is usually good enough, but in some cases it may be necessary to provide custom
allocation orders.
</p>
</div>
<!-- ======================================================================= -->
<h3>
<a name="implementRegister">Implement a subclass of</a>
<a href="CodeGenerator.html#targetregisterinfo">TargetRegisterInfo</a>
</h3>
<div>
<p>
The final step is to hand code portions of <tt>XXXRegisterInfo</tt>, which
implements the interface described in <tt>TargetRegisterInfo.h</tt>. These
functions return <tt>0</tt>, <tt>NULL</tt>, or <tt>false</tt>, unless
overridden. Here is 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>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>
<!-- *********************************************************************** -->
<h2>
<a name="InstructionSet">Instruction Set</a>
</h2>
<!-- *********************************************************************** -->
<div>
<p>
During the early stages of code generation, the LLVM IR code is converted to a
<tt>SelectionDAG</tt> with nodes that are instances of the <tt>SDNode</tt> class
containing target instructions. An <tt>SDNode</tt> 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 <tt>NodeType</tt> enum in the <tt>ISD</tt> namespace).
</p>
<p>
TableGen uses the following target description (<tt>.td</tt>) input files to
generate much of the code for instruction definition:
</p>
<ul>
<li><tt>Target.td</tt> — Where the <tt>Instruction</tt>, <tt>Operand</tt>,
<tt>InstrInfo</tt>, and other fundamental classes are defined.</li>
<li><tt>TargetSelectionDAG.td</tt>— Used by <tt>SelectionDAG</tt>
instruction selection generators, contains <tt>SDTC*</tt> classes (selection
DAG type constraint), definitions of <tt>SelectionDAG</tt> nodes (such as
<tt>imm</tt>, <tt>cond</tt>, <tt>bb</tt>, <tt>add</tt>, <tt>fadd</tt>,
<tt>sub</tt>), and pattern support (<tt>Pattern</tt>, <tt>Pat</tt>,
<tt>PatFrag</tt>, <tt>PatLeaf</tt>, <tt>ComplexPattern</tt>.</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 <tt>XXX</tt> is the
name of the target. The <tt>XXX.td</tt> file includes the other <tt>.td</tt>
input files, but its contents are only directly important for subtargets.
</p>
<p>
You should describe a concrete target-specific class <tt>XXXInstrInfo</tt> that
represents machine instructions supported by a target machine.
<tt>XXXInstrInfo</tt> contains an array of <tt>XXXInstrDescriptor</tt> 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 class="doc_code">
<pre>class Instruction {
string Namespace = "";
dag OutOperandList; // An dag containing the MI def operand list.
dag InOperandList; // An dag containing the MI use operand list.
string AsmString = ""; // The .s format to print the instruction with.
list<dag> Pattern; // Set to the DAG pattern for this instruction
list<Register> Uses = [];
list<Register> Defs = [];
list<Predicate> Predicates = []; // predicates turned into isel match code
... remainder not shown for space ...
}
</pre>
</div>
<p>
A <tt>SelectionDAG</tt> node (<tt>SDNode</tt>) 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. 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 <tt>ADDri</tt> and
<tt>ADDrr</tt>.
</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 <tt>CALL</tt>
instruction. Format 2 is for branch on condition codes and <tt>SETHI</tt> (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>.
<tt>InstSP</tt> is a base class for other instruction classes. Additional base
classes are specified for more precise formats: for example
in <tt>SparcInstrFormat.td</tt>, <tt>F2_1</tt> is for <tt>SETHI</tt>,
and <tt>F2_2</tt> is for branches. There are three other base
classes: <tt>F3_1</tt> for register/register operations, <tt>F3_2</tt> for
register/immediate operations, and <tt>F3_3</tt> 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, <tt>LDrr</tt>, defines the Load Integer
instruction for a Word (the <tt>LD</tt> SPARC opcode) from a memory address to a
register. The first parameter, the value 3 (<tt>11<sub>2</sub></tt>), is the
operation value for this category of operation. The second parameter
(<tt>000000<sub>2</sub></tt>) is the specific operation value
for <tt>LD</tt>/Load Word. The third parameter is the output destination, which
is a register operand and defined in the <tt>Register</tt> target description
file (<tt>IntRegs</tt>).
</p>
<div class="doc_code">
<pre>def LDrr : F3_1 <3, 0b000000, (outs IntRegs:$dst), (ins MEMrr:$addr),
"ld [$addr], $dst",
[(set IntRegs:$dst, (load ADDRrr:$addr))]>;
</pre>
</div>
<p>
The fourth parameter is the input source, which uses the address
operand <tt>MEMrr</tt> that is defined earlier in <tt>SparcInstrInfo.td</tt>:
</p>
<div class="doc_code">
<pre>def MEMrr : Operand<i32> {
let PrintMethod = "printMemOperand";
let MIOperandInfo = (ops IntRegs, IntRegs);
}
</pre>
</div>
<p>
The fifth parameter is a string that is used by the assembly printer and can be
left as an empty string until the assembly printer interface is implemented. The
sixth and final parameter is the pattern used to match the instruction during
the SelectionDAG Select Phase described in
(<a href="CodeGenerator.html">The LLVM
Target-Independent Code Generator</a>). This parameter is detailed in the next
section, <a href="#InstructionSelector">Instruction Selector</a>.
</p>
<p>
Instruction class definitions are not overloaded for different operand types, so
separate versions of instructions are needed for register, memory, or immediate
value operands. For example, to perform a Load Integer instruction for a Word
from an immediate operand to a register, the following instruction class is
defined:
</p>
<div class="doc_code">
<pre>def LDri : F3_2 <3, 0b000000, (outs IntRegs:$dst), (ins MEMri:$addr),
"ld [$addr], $dst",
[(set IntRegs:$dst, (load ADDRri:$addr))]>;
</pre>
</div>
<p>
Writing these definitions for so many similar instructions can involve a lot of
cut and paste. In td files, the <tt>multiclass</tt> directive enables the
creation of templates to define several instruction classes at once (using
the <tt>defm</tt> directive). For example in <tt>SparcInstrInfo.td</tt>, the
<tt>multiclass</tt> pattern <tt>F3_12</tt> is defined to create 2 instruction
classes each time <tt>F3_12</tt> is invoked:
</p>
<div class="doc_code">
<pre>multiclass F3_12 <string OpcStr, bits<6> Op3Val, SDNode OpNode> {
def rr : F3_1 <2, Op3Val,
(outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c),
!strconcat(OpcStr, " $b, $c, $dst"),
[(set IntRegs:$dst, (OpNode IntRegs:$b, IntRegs:$c))]>;
def ri : F3_2 <2, Op3Val,
(outs IntRegs:$dst), (ins IntRegs:$b, i32imm:$c),
!strconcat(OpcStr, " $b, $c, $dst"),
[(set IntRegs:$dst, (OpNode IntRegs:$b, simm13:$c))]>;
}
</pre>
</div>
<p>
So when the <tt>defm</tt> directive is used for the <tt>XOR</tt>
and <tt>ADD</tt> instructions, as seen below, it creates four instruction
objects: <tt>XORrr</tt>, <tt>XORri</tt>, <tt>ADDrr</tt>, and <tt>ADDri</tt>.
</p>
<div class="doc_code">
<pre>
defm XOR : F3_12<"xor", 0b000011, xor>;
defm ADD : F3_12<"add", 0b000000, add>;
</pre>
</div>
<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 'greater than'
condition for integers, and the 22<sup>nd</sup> bit represents the 'greater
than' condition for floats.
</p>
<div class="doc_code">
<pre>
def ICC_NE : ICC_VAL< 9>; // Not Equal
def ICC_E : ICC_VAL< 1>; // Equal
def ICC_G : ICC_VAL<10>; // Greater
...
def FCC_U : FCC_VAL<23>; // Unordered
def FCC_G : FCC_VAL<22>; // Greater
def FCC_UG : FCC_VAL<21>; // Unordered or Greater
...
</pre>
</div>
<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>. I.e.,
<tt>SPCC::ICC_NE = 9</tt>, <tt>SPCC::FCC_U = 23</tt> and so on.)
</p>
<!-- ======================================================================= -->
<h3>
<a name="operandMapping">Instruction Operand Mapping</a>
</h3>
<div>
<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 <tt>XNORrr</tt> instruction as
a <tt>F3_1</tt> format instruction having three operands.
</p>
<div class="doc_code">
<pre>
def XNORrr : F3_1<2, 0b000111,
(outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c),
"xnor $b, $c, $dst",
[(set IntRegs:$dst, (not (xor IntRegs:$b, IntRegs:$c)))]>;
</pre>
</div>
<p>
The instruction templates in <tt>SparcInstrFormats.td</tt> show the base class
for <tt>F3_1</tt> is <tt>InstSP</tt>.
</p>
<div class="doc_code">
<pre>
class InstSP<dag outs, dag ins, string asmstr, list<dag> pattern> : Instruction {
field bits<32> Inst;
let Namespace = "SP";
bits<2> op;
let Inst{31-30} = op;
dag OutOperandList = outs;
dag InOperandList = ins;
let AsmString = asmstr;
let Pattern = pattern;
}
</pre>
</div>
<p><tt>InstSP</tt> leaves the <tt>op</tt> field unbound.</p>
<div class="doc_code">
<pre>
class F3<dag outs, dag ins, string asmstr, list<dag> pattern>
: InstSP<outs, ins, asmstr, pattern> {
bits<5> rd;
bits<6> op3;
bits<5> 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>
<p>
<tt>F3</tt> binds the <tt>op</tt> field and defines the <tt>rd</tt>,
<tt>op3</tt>, and <tt>rs1</tt> fields. <tt>F3</tt> format instructions will
bind the operands <tt>rd</tt>, <tt>op3</tt>, and <tt>rs1</tt> fields.
</p>
<div class="doc_code">
<pre>
class F3_1<bits<2> opVal, bits<6> op3val, dag outs, dag ins,
string asmstr, list<dag> pattern> : F3<outs, ins, asmstr, pattern> {
bits<8> asi = 0; // asi not currently used
bits<5> 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>
<p>
<tt>F3_1</tt> binds the <tt>op3</tt> field and defines the <tt>rs2</tt>
fields. <tt>F3_1</tt> format instructions will bind the operands to the <tt>rd</tt>,
<tt>rs1</tt>, and <tt>rs2</tt> fields. This results in the <tt>XNORrr</tt>
instruction binding <tt>$dst</tt>, <tt>$b</tt>, and <tt>$c</tt> operands to
the <tt>rd</tt>, <tt>rs1</tt>, and <tt>rs2</tt> fields respectively.
</p>
</div>
<!-- ======================================================================= -->
<h3>
<a name="implementInstr">Implement a subclass of </a>
<a href="CodeGenerator.html#targetinstrinfo">TargetInstrInfo</a>
</h3>
<div>
<p>
The final step is to hand code portions of <tt>XXXInstrInfo</tt>, which
implements the interface described in <tt>TargetInstrInfo.h</tt>. These
functions return <tt>0</tt> 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>isLoadFromStackSlot</tt> — If the specified machine instruction is
a direct load from a stack slot, return the register number of the
destination and the <tt>FrameIndex</tt> 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 <tt>FrameIndex</tt> of the stack slot.</li>
<li><tt>copyPhysReg</tt> — Copy values between a pair of physical
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>
<!-- ======================================================================= -->
<h3>
<a name="branchFolding">Branch Folding and If Conversion</a>
</h3>
<div>
<p>
Performance can be improved by combining instructions or by eliminating
instructions that are never reached. The <tt>AnalyzeBranch</tt> method
in <tt>XXXInstrInfo</tt> 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><tt>MachineBasicBlock &MBB</tt> — The incoming block to be
examined.</li>
<li><tt>MachineBasicBlock *&TBB</tt> — A destination block that is
returned. For a conditional branch that evaluates to true, <tt>TBB</tt> is
the destination.</li>
<li><tt>MachineBasicBlock *&FBB</tt> — For a conditional branch that
evaluates to false, <tt>FBB</tt> is returned as the destination.</li>
<li><tt>std::vector<MachineOperand> &Cond</tt> — 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 <tt>TBB</tt>
or <tt>FBB</tt>, so both parameters return <tt>NULL</tt>. 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 class="doc_code">
<pre>bool ARMInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
std::vector<MachineOperand> &Cond) const
{
MachineBasicBlock::iterator I = MBB.end();
if (I == MBB.begin() || !isUnpredicatedTerminator(--I))
return false;
</pre>
</div>
<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 <tt>TBB</tt> parameter.
</p>
<div class="doc_code">
<pre>
if (LastOpc == ARM::B || LastOpc == ARM::tB) {
TBB = LastInst->getOperand(0).getMBB();
return false;
}
</pre>
</div>
<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 <tt>TBB</tt> parameter.
</p>
<div class="doc_code">
<pre>
if ((SecondLastOpc == ARM::B || SecondLastOpc==ARM::tB) &&
(LastOpc == ARM::B || LastOpc == ARM::tB)) {
TBB = SecondLastInst->getOperand(0).getMBB();
I = LastInst;
I->eraseFromParent();
return false;
}
</pre>
</div>
<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 <tt>TBB</tt> parameter and a list of operands in
the <tt>Cond</tt> parameter to evaluate the condition.
</p>
<div class="doc_code">
<pre>
if (LastOpc == ARM::Bcc || LastOpc == ARM::tBcc) {
// Block ends with fall-through condbranch.
TBB = LastInst->getOperand(0).getMBB();
Cond.push_back(LastInst->getOperand(1));
Cond.push_back(LastInst->getOperand(2));
return false;
}
</pre>
</div>
<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
'<tt>true</tt>') in the <tt>TBB</tt> parameter and the unconditional branch
destination in the <tt>FBB</tt> (corresponding to a conditional evaluation of
'<tt>false</tt>'). A list of operands to evaluate the condition should be
returned in the <tt>Cond</tt> parameter.
</p>
<div class="doc_code">
<pre>
unsigned SecondLastOpc = SecondLastInst->getOpcode();
if ((SecondLastOpc == ARM::Bcc && LastOpc == ARM::B) ||
(SecondLastOpc == ARM::tBcc && LastOpc == ARM::tB)) {
TBB = SecondLastInst->getOperand(0).getMBB();
Cond.push_back(SecondLastInst->getOperand(1));
Cond.push_back(SecondLastInst->getOperand(2));
FBB = LastInst->getOperand(0).getMBB();
return false;
}
</pre>
</div>
<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>
<!-- *********************************************************************** -->
<h2>
<a name="InstructionSelector">Instruction Selector</a>
</h2>
<!-- *********************************************************************** -->
<div>
<p>
LLVM uses a <tt>SelectionDAG</tt> to represent LLVM IR instructions, and nodes
of the <tt>SelectionDAG</tt> 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 <tt>SelectionDAG</tt>.
</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 <tt>FunctionPass</tt> class or a subclass of <tt>FunctionPass</tt>. 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 <tt>SelectionDAG</tt> before or after specific
processing phases, use the command line options for <tt>llc</tt>, described
at <a href="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 <tt>SelectionDAG</tt> 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 class="doc_code">
<pre>
def STrr : F3_1< 3, 0b000100, (outs), (ins MEMrr:$addr, IntRegs:$src),
"st $src, [$addr]", [(store IntRegs:$src, ADDRrr:$addr)]>;
</pre>
</div>
<p>
<tt>ADDRrr</tt> is a memory mode that is also defined in
<tt>SparcInstrInfo.td</tt>:
</p>
<div class="doc_code">
<pre>
def ADDRrr : ComplexPattern<i32, 2, "SelectADDRrr", [], []>;
</pre>
</div>
<p>
The definition of <tt>ADDRrr</tt> refers to <tt>SelectADDRrr</tt>, 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 class="doc_code">
<pre>
def store : PatFrag<(ops node:$val, node:$ptr),
(st node:$val, node:$ptr), [{
if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
return !ST->isTruncatingStore() &&
ST->getAddressingMode() == ISD::UNINDEXED;
return false;
}]>;
</pre>
</div>
<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 <tt>ISD::STORE</tt> opcode.
</p>
<div class="doc_code">
<pre>
SDNode *SelectCode(SDValue N) {
...
MVT::ValueType NVT = N.getNode()->getValueType(0);
switch (N.getOpcode()) {
case ISD::STORE: {
switch (NVT) {
default:
return Select_ISD_STORE(N);
break;
}
break;
}
...
</pre>
</div>
<p>
The pattern for <tt>STrr</tt> is matched, so elsewhere in
<tt>XXXGenDAGISel.inc</tt>, code for <tt>STrr</tt> 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 class="doc_code">
<pre>
SDNode *Select_ISD_STORE(const SDValue &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;
// Pattern: (st:void IntRegs:i32:$src,
// ADDRrr:i32:$addr)<<P:Predicate_store>>
// Emits: (STrr:void ADDRrr:i32:$addr, IntRegs:i32:$src)
// Pattern complexity = 13 cost = 1 size = 0
if (SelectADDRrr(N, N2, CPTmp0, CPTmp1) &&
N1.getNode()->getValueType(0) == MVT::i32 &&
N2.getNode()->getValueType(0) == MVT::i32) {
return Emit_22(N, SP::STrr, CPTmp0, CPTmp1);
}
...
</pre>
</div>
<!-- ======================================================================= -->
<h3>
<a name="LegalizePhase">The SelectionDAG Legalize Phase</a>
</h3>
<div>
<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 <tt>XXXTargetLowering</tt> 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 class="doc_code">
<pre>
addRegisterClass(MVT::i32, SP::IntRegsRegisterClass);
addRegisterClass(MVT::f32, SP::FPRegsRegisterClass);
addRegisterClass(MVT::f64, SP::DFPRegsRegisterClass);
</pre>
</div>
<p>
You should examine the node types in the <tt>ISD</tt> 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 <tt>LegalAction</tt> 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 <tt>LegalAction</tt> values.
</p>
<!-- _______________________________________________________________________ -->
<h4>
<a name="promote">Promote</a>
</h4>
<div>
<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 class="doc_code">
<pre>
setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
</pre>
</div>
</div>
<!-- _______________________________________________________________________ -->
<h4>
<a name="expand">Expand</a>
</h4>
<div>
<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 class="doc_code">
<pre>
setOperationAction(ISD::FSIN, MVT::f32, Expand);
setOperationAction(ISD::FCOS, MVT::f32, Expand);
</pre>
</div>
</div>
<!-- _______________________________________________________________________ -->
<h4>
<a name="custom">Custom</a>
</h4>
<div>
<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 class="doc_code">
<pre>
setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
</pre>
</div>
<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 <tt>FP_TO_SINT</tt> opcode will call
the <tt>LowerFP_TO_SINT</tt> method:
</p>
<div class="doc_code">
<pre>
SDValue SparcTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
switch (Op.getOpcode()) {
case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
...
}
}
</pre>
</div>
<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 class="doc_code">
<pre>
static SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
assert(Op.getValueType() == MVT::i32);
Op = DAG.getNode(SPISD::FTOI, MVT::f32, Op.getOperand(0));
return DAG.getNode(ISD::BITCAST, MVT::i32, Op);
}
</pre>
</div>
</div>
<!-- _______________________________________________________________________ -->
<h4>
<a name="legal">Legal</a>
</h4>
<div>
<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 <tt>CTPOP</tt> (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 class="doc_code">
<pre>
setOperationAction(ISD::CTPOP, MVT::i32, Expand);
...
if (TM.getSubtarget<SparcSubtarget>().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>
</div>
<!-- ======================================================================= -->
<h3>
<a name="callingConventions">Calling Conventions</a>
</h3>
<div>
<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 <tt>CCIfType</tt> and
<tt>CCAssignToReg</tt> interfaces. If the <tt>CCIfType</tt> predicate is true
(that is, if the current argument is of type <tt>f32</tt> or <tt>f64</tt>), then
the action is performed. In this case, the <tt>CCAssignToReg</tt> action assigns
the argument value to the first available register: either <tt>R0</tt>
or <tt>R1</tt>.
</p>
<div class="doc_code">
<pre>
CCIfType<[f32,f64], CCAssignToReg<[R0, R1]>>
</pre>
</div>
<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 (<tt>CC_Sparc32</tt>). The definition of <tt>RetCC_Sparc32</tt>
(shown below) indicates which registers are used for specified scalar return
types. A single-precision float is returned to register <tt>F0</tt>, and a
double-precision float goes to register <tt>D0</tt>. A 32-bit integer is
returned in register <tt>I0</tt> or <tt>I1</tt>.
</p>
<div class="doc_code">
<pre>
def RetCC_Sparc32 : CallingConv<[
CCIfType<[i32], CCAssignToReg<[I0, I1]>>,
CCIfType<[f32], CCAssignToReg<[F0]>>,
CCIfType<[f64], CCAssignToReg<[D0]>>
]>;
</pre>
</div>
<p>
The definition of <tt>CC_Sparc32</tt> in <tt>SparcCallingConv.td</tt> introduces
<tt>CCAssignToStack</tt>, 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 class="doc_code">
<pre>
def CC_Sparc32 : CallingConv<[
// All arguments get passed in integer registers if there is space.
CCIfType<[i32, f32, f64], CCAssignToReg<[I0, I1, I2, I3, I4, I5]>>,
CCAssignToStack<4, 4>
]>;
</pre>
</div>
<p>
<tt>CCDelegateTo</tt> 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
<tt>RetCC_X86_32_C</tt> ends with <tt>CCDelegateTo</tt>. After the current value
is assigned to the register <tt>ST0</tt> or <tt>ST1</tt>,
the <tt>RetCC_X86Common</tt> is invoked.
</p>
<div class="doc_code">
<pre>
def RetCC_X86_32_C : CallingConv<[
CCIfType<[f32], CCAssignToReg<[ST0, ST1]>>,
CCIfType<[f64], CCAssignToReg<[ST0, ST1]>>,
CCDelegateTo<RetCC_X86Common>
]>;
</pre>
</div>
<p>
<tt>CCIfCC</tt> 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 <tt>Fast</tt> calling convention is in use,
then <tt>RetCC_X86_32_Fast</tt> is invoked. If the <tt>SSECall</tt> calling
convention is in use, then <tt>RetCC_X86_32_SSE</tt> is invoked.
</p>
<div class="doc_code">
<pre>
def RetCC_X86_32 : CallingConv<[
CCIfCC<"CallingConv::Fast", CCDelegateTo<RetCC_X86_32_Fast>>,
CCIfCC<"CallingConv::X86_SSECall", CCDelegateTo<RetCC_X86_32_SSE>>,
CCDelegateTo<RetCC_X86_32_C>
]>;
</pre>
</div>
<p>Other calling convention interfaces include:</p>
<ul>
<li><tt>CCIf <predicate, action></tt> — If the predicate matches,
apply the action.</li>
<li><tt>CCIfInReg <action></tt> — If the argument is marked with the
'<tt>inreg</tt>' attribute, then apply the action.</li>
<li><tt>CCIfNest <action></tt> — Inf the argument is marked with the
'<tt>nest</tt>' attribute, then apply the action.</li>
<li><tt>CCIfNotVarArg <action></tt> — If the current function does
not take a variable number of arguments, apply the action.</li>
<li><tt>CCAssignToRegWithShadow <registerList, shadowList></tt> —
similar to <tt>CCAssignToReg</tt>, but with a shadow list of registers.</li>
<li><tt>CCPassByVal <size, align></tt> — Assign value to a stack
slot with the minimum specified size and alignment.</li>
<li><tt>CCPromoteToType <type></tt> — Promote the current value to
the specified type.</li>
<li><tt>CallingConv <[actions]></tt> — Define each calling
convention that is supported.</li>
</ul>
</div>
</div>
<!-- *********************************************************************** -->
<h2>
<a name="assemblyPrinter">Assembly Printer</a>
</h2>
<!-- *********************************************************************** -->
<div>
<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 <tt>XXXTargetAsmInfo</tt> class (a subclass
of <tt>TargetAsmInfo</tt>).</li>
<li>Write <tt>XXXTargetAsmInfo.cpp</tt>, which contains target-specific values
for <tt>TargetAsmInfo</tt> properties and sometimes new implementations for
methods.</li>
<li>Write <tt>XXXAsmPrinter.cpp</tt>, which implements the <tt>AsmPrinter</tt>
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
<tt>XXXTargetAsmInfo</tt> class for use in <tt>XXXTargetAsmInfo.cpp</tt>.
Similarly, <tt>XXXTargetAsmInfo.cpp</tt> usually has a few declarations of
<tt>XXXTargetAsmInfo</tt> replacement values that override the default values
in <tt>TargetAsmInfo.cpp</tt>. For example in <tt>SparcTargetAsmInfo.cpp</tt>:
</p>
<div class="doc_code">
<pre>
SparcTargetAsmInfo::SparcTargetAsmInfo(const SparcTargetMachine &TM) {
Data16bitsDirective = "\t.half\t";
Data32bitsDirective = "\t.word\t";
Data64bitsDirective = 0; // .xword is only supported by V9.
ZeroDirective = "\t.skip\t";
CommentString = "!";
ConstantPoolSection = "\t.section \".rodata\",#alloc\n";
}
</pre>
</div>
<p>
The X86 assembly printer implementation (<tt>X86TargetAsmInfo</tt>) is an
example where the target specific <tt>TargetAsmInfo</tt> class uses an
overridden methods: <tt>ExpandInlineAsm</tt>.
</p>
<p>
A target-specific implementation of AsmPrinter is written in
<tt>XXXAsmPrinter.cpp</tt>, which implements the <tt>AsmPrinter</tt> class that
converts the LLVM to printable assembly. The implementation must include the
following headers that have declarations for the <tt>AsmPrinter</tt> and
<tt>MachineFunctionPass</tt> classes. The <tt>MachineFunctionPass</tt> is a
subclass of <tt>FunctionPass</tt>.
</p>
<div class="doc_code">
<pre>
#include "llvm/CodeGen/AsmPrinter.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
</pre>
</div>
<p>
As a <tt>FunctionPass</tt>, <tt>AsmPrinter</tt> first
calls <tt>doInitialization</tt> to set up the <tt>AsmPrinter</tt>. In
<tt>SparcAsmPrinter</tt>, a <tt>Mangler</tt> object is instantiated to process
variable names.
</p>
<p>
In <tt>XXXAsmPrinter.cpp</tt>, the <tt>runOnMachineFunction</tt> method
(declared in <tt>MachineFunctionPass</tt>) must be implemented
for <tt>XXXAsmPrinter</tt>. In <tt>MachineFunctionPass</tt>,
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 <tt>XXXAsmPrinter</tt> 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>
</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.
</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 <tt>XXXAsmPrinter</tt>, 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>
<!-- *********************************************************************** -->
<h2>
<a name="subtargetSupport">Subtarget Support</a>
</h2>
<!-- *********************************************************************** -->
<div>
<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>-mcpu=</tt> and <tt>-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 <tt>SubtargetFeature</tt> interface is defined. The first 4 string
parameters of the <tt>SubtargetFeature</tt> 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 class="doc_code">
<pre>
class SubtargetFeature<string n, string a, string v, string d,
list<SubtargetFeature> i = []> {
string Name = n;
string Attribute = a;
string Value = v;
string Desc = d;
list<SubtargetFeature> Implies = i;
}
</pre>
</div>
<p>
In the <tt>Sparc.td</tt> file, the SubtargetFeature is used to define the
following features.
</p>
<div class="doc_code">
<pre>
def FeatureV9 : SubtargetFeature<"v9", "IsV9", "true",
"Enable SPARC-V9 instructions">;
def FeatureV8Deprecated : SubtargetFeature<"deprecated-v8",
"V8DeprecatedInsts", "true",
"Enable deprecated V8 instructions in V9 mode">;
def FeatureVIS : SubtargetFeature<"vis", "IsVIS", "true",
"Enable UltraSPARC Visual Instruction Set extensions">;
</pre>
</div>
<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 class="doc_code">
<pre>
class Proc<string Name, list<SubtargetFeature> Features>
: Processor<Name, NoItineraries, Features>;
def : Proc<"generic", []>;
def : Proc<"v8", []>;
def : Proc<"supersparc", []>;
def : Proc<"sparclite", []>;
def : Proc<"f934", []>;
def : Proc<"hypersparc", []>;
def : Proc<"sparclite86x", []>;
def : Proc<"sparclet", []>;
def : Proc<"tsc701", []>;
def : Proc<"v9", [FeatureV9]>;
def : Proc<"ultrasparc", [FeatureV9, FeatureV8Deprecated]>;
def : Proc<"ultrasparc3", [FeatureV9, FeatureV8Deprecated]>;
def : Proc<"ultrasparc3-vis", [FeatureV9, FeatureV8Deprecated, FeatureVIS]>;
</pre>
</div>
<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 class="doc_code">
<pre>
XXXSubtarget::XXXSubtarget(const Module &M, const std::string &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>
<!-- *********************************************************************** -->
<h2>
<a name="jitSupport">JIT Support</a>
</h2>
<!-- *********************************************************************** -->
<div>
<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 <tt>XXXTargetMachine</tt> so that it provides a
<tt>TargetJITInfo</tt> 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
<tt>MachineCodeEmitter</tt> class containing code for several callback functions
that write data (in bytes, words, strings, etc.) to the output stream.
</p>
<!-- ======================================================================= -->
<h3>
<a name="mce">Machine Code Emitter</a>
</h3>
<div>
<p>
In <tt>XXXCodeEmitter.cpp</tt>, a target-specific of the <tt>Emitter</tt> class
is implemented as a function pass (subclass
of <tt>MachineFunctionPass</tt>). The target-specific implementation
of <tt>runOnMachineFunction</tt> (invoked by
<tt>runOnFunction</tt> in <tt>MachineFunctionPass</tt>) iterates through the
<tt>MachineBasicBlock</tt> 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 class="doc_code">
<pre>
switch (Desc->TSFlags & 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>
<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, <tt>MO1</tt>, 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 class="doc_code">
<pre>
case X86II::AddRegFrm:
MCE.emitByte(BaseOpcode + getX86RegNum(MI.getOperand(CurOp++).getReg()));
if (CurOp != NumOps) {
const MachineOperand &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<Function>(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>
<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
<tt>RelocationType</tt> enum for that target is defined in the short
target-specific <tt>XXXRelocations.h</tt> file. The <tt>RelocationType</tt> 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 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>
<!-- ======================================================================= -->
<h3>
<a name="targetJITInfo">Target JIT Info</a>
</h3>
<div>
<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 <tt>XXXJITInfo</tt> implements the following:
</p>
<ul>
<li><tt>getLazyResolverFunction</tt> — Initializes the JIT, gives the
target a function that is used for compilation.</li>
<li><tt>emitFunctionStub</tt> — Returns a native function with a specified
address for a callback function.</li>
<li><tt>relocate</tt> — 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 <tt>JITCompilerFunction</tt> 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 class="doc_code">
<pre>
TargetJITInfo::LazyResolverFn AlphaJITInfo::getLazyResolverFunction(
JITCompilerFn F) {
JITCompilerFunction = F;
return AlphaCompilationCallback;
}
</pre>
</div>
<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>
</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: 2012-03-01 07:14:19 -0800 (Thu, 01 Mar 2012) $
</address>
</body>
</html>