1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
6 <title>Writing an LLVM Compiler Backend</title>
7 <link rel="stylesheet" href="_static/llvm.css" type="text/css">
13 Writing an LLVM Compiler Backend
17 <li><a href="#intro">Introduction</a>
19 <li><a href="#Audience">Audience</a></li>
20 <li><a href="#Prerequisite">Prerequisite Reading</a></li>
21 <li><a href="#Basic">Basic Steps</a></li>
22 <li><a href="#Preliminaries">Preliminaries</a></li>
24 <li><a href="#TargetMachine">Target Machine</a></li>
25 <li><a href="#TargetRegistration">Target Registration</a></li>
26 <li><a href="#RegisterSet">Register Set and Register Classes</a>
28 <li><a href="#RegisterDef">Defining a Register</a></li>
29 <li><a href="#RegisterClassDef">Defining a Register Class</a></li>
30 <li><a href="#implementRegister">Implement a subclass of TargetRegisterInfo</a></li>
32 <li><a href="#InstructionSet">Instruction Set</a>
34 <li><a href="#operandMapping">Instruction Operand Mapping</a></li>
35 <li><a href="#implementInstr">Implement a subclass of TargetInstrInfo</a></li>
36 <li><a href="#branchFolding">Branch Folding and If Conversion</a></li>
38 <li><a href="#InstructionSelector">Instruction Selector</a>
40 <li><a href="#LegalizePhase">The SelectionDAG Legalize Phase</a>
42 <li><a href="#promote">Promote</a></li>
43 <li><a href="#expand">Expand</a></li>
44 <li><a href="#custom">Custom</a></li>
45 <li><a href="#legal">Legal</a></li>
47 <li><a href="#callingConventions">Calling Conventions</a></li>
49 <li><a href="#assemblyPrinter">Assembly Printer</a></li>
50 <li><a href="#subtargetSupport">Subtarget Support</a></li>
51 <li><a href="#jitSupport">JIT Support</a>
53 <li><a href="#mce">Machine Code Emitter</a></li>
54 <li><a href="#targetJITInfo">Target JIT Info</a></li>
58 <div class="doc_author">
59 <p>Written by <a href="http://www.woo.com">Mason Woo</a> and
60 <a href="http://misha.brukman.net">Misha Brukman</a></p>
63 <!-- *********************************************************************** -->
65 <a name="intro">Introduction</a>
67 <!-- *********************************************************************** -->
72 This document describes techniques for writing compiler backends that convert
73 the LLVM Intermediate Representation (IR) to code for a specified machine or
74 other languages. Code intended for a specific machine can take the form of
75 either assembly code or binary code (usable for a JIT compiler).
79 The backend of LLVM features a target-independent code generator that may create
80 output for several types of target CPUs — including X86, PowerPC, ARM,
81 and SPARC. The backend may also be used to generate code targeted at SPUs of the
82 Cell processor or GPUs to support the execution of compute kernels.
86 The document focuses on existing examples found in subdirectories
87 of <tt>llvm/lib/Target</tt> in a downloaded LLVM release. In particular, this
88 document focuses on the example of creating a static compiler (one that emits
89 text assembly) for a SPARC target, because SPARC has fairly standard
90 characteristics, such as a RISC instruction set and straightforward calling
95 <a name="Audience">Audience</a>
101 The audience for this document is anyone who needs to write an LLVM backend to
102 generate code for a specific hardware or software target.
108 <a name="Prerequisite">Prerequisite Reading</a>
114 These essential documents must be read before reading this document:
118 <li><i><a href="LangRef.html">LLVM Language Reference
119 Manual</a></i> — a reference manual for the LLVM assembly language.</li>
121 <li><i><a href="CodeGenerator.html">The LLVM
122 Target-Independent Code Generator</a></i> — a guide to the components
123 (classes and code generation algorithms) for translating the LLVM internal
124 representation into machine code for a specified target. Pay particular
125 attention to the descriptions of code generation stages: Instruction
126 Selection, Scheduling and Formation, SSA-based Optimization, Register
127 Allocation, Prolog/Epilog Code Insertion, Late Machine Code Optimizations,
128 and Code Emission.</li>
130 <li><i><a href="TableGenFundamentals.html">TableGen
131 Fundamentals</a></i> —a document that describes the TableGen
132 (<tt>tblgen</tt>) application that manages domain-specific information to
133 support LLVM code generation. TableGen processes input from a target
134 description file (<tt>.td</tt> suffix) and generates C++ code that can be
135 used for code generation.</li>
137 <li><i><a href="WritingAnLLVMPass.html">Writing an LLVM
138 Pass</a></i> — The assembly printer is a <tt>FunctionPass</tt>, as are
139 several SelectionDAG processing steps.</li>
143 To follow the SPARC examples in this document, have a copy of
144 <i><a href="http://www.sparc.org/standards/V8.pdf">The SPARC Architecture
145 Manual, Version 8</a></i> for reference. For details about the ARM instruction
146 set, refer to the <i><a href="http://infocenter.arm.com/">ARM Architecture
147 Reference Manual</a></i>. For more about the GNU Assembler format
149 <i><a href="http://sourceware.org/binutils/docs/as/index.html">Using As</a></i>,
150 especially for the assembly printer. <i>Using As</i> contains a list of target
151 machine dependent features.
157 <a name="Basic">Basic Steps</a>
163 To write a compiler backend for LLVM that converts the LLVM IR to code for a
164 specified target (machine or other language), follow these steps:
168 <li>Create a subclass of the TargetMachine class that describes characteristics
169 of your target machine. Copy existing examples of specific TargetMachine
170 class and header files; for example, start with
171 <tt>SparcTargetMachine.cpp</tt> and <tt>SparcTargetMachine.h</tt>, but
172 change the file names for your target. Similarly, change code that
173 references "Sparc" to reference your target. </li>
175 <li>Describe the register set of the target. Use TableGen to generate code for
176 register definition, register aliases, and register classes from a
177 target-specific <tt>RegisterInfo.td</tt> input file. You should also write
178 additional code for a subclass of the TargetRegisterInfo class that
179 represents the class register file data used for register allocation and
180 also describes the interactions between registers.</li>
182 <li>Describe the instruction set of the target. Use TableGen to generate code
183 for target-specific instructions from target-specific versions of
184 <tt>TargetInstrFormats.td</tt> and <tt>TargetInstrInfo.td</tt>. You should
185 write additional code for a subclass of the TargetInstrInfo class to
186 represent machine instructions supported by the target machine. </li>
188 <li>Describe the selection and conversion of the LLVM IR from a Directed Acyclic
189 Graph (DAG) representation of instructions to native target-specific
190 instructions. Use TableGen to generate code that matches patterns and
191 selects instructions based on additional information in a target-specific
192 version of <tt>TargetInstrInfo.td</tt>. Write code
193 for <tt>XXXISelDAGToDAG.cpp</tt>, where XXX identifies the specific target,
194 to perform pattern matching and DAG-to-DAG instruction selection. Also write
195 code in <tt>XXXISelLowering.cpp</tt> to replace or remove operations and
196 data types that are not supported natively in a SelectionDAG. </li>
198 <li>Write code for an assembly printer that converts LLVM IR to a GAS format for
199 your target machine. You should add assembly strings to the instructions
200 defined in your target-specific version of <tt>TargetInstrInfo.td</tt>. You
201 should also write code for a subclass of AsmPrinter that performs the
202 LLVM-to-assembly conversion and a trivial subclass of TargetAsmInfo.</li>
204 <li>Optionally, add support for subtargets (i.e., variants with different
205 capabilities). You should also write code for a subclass of the
206 TargetSubtarget class, which allows you to use the <tt>-mcpu=</tt>
207 and <tt>-mattr=</tt> command-line options.</li>
209 <li>Optionally, add JIT support and create a machine code emitter (subclass of
210 TargetJITInfo) that is used to emit binary code directly into memory. </li>
214 In the <tt>.cpp</tt> and <tt>.h</tt>. files, initially stub up these methods and
215 then implement them later. Initially, you may not know which private members
216 that the class will need and which components will need to be subclassed.
222 <a name="Preliminaries">Preliminaries</a>
228 To actually create your compiler backend, you need to create and modify a few
229 files. The absolute minimum is discussed here. But to actually use the LLVM
230 target-independent code generator, you must perform the steps described in
231 the <a href="CodeGenerator.html">LLVM
232 Target-Independent Code Generator</a> document.
236 First, you should create a subdirectory under <tt>lib/Target</tt> to hold all
237 the files related to your target. If your target is called "Dummy," create the
238 directory <tt>lib/Target/Dummy</tt>.
243 directory, create a <tt>Makefile</tt>. It is easiest to copy a
244 <tt>Makefile</tt> of another target and modify it. It should at least contain
245 the <tt>LEVEL</tt>, <tt>LIBRARYNAME</tt> and <tt>TARGET</tt> variables, and then
246 include <tt>$(LEVEL)/Makefile.common</tt>. The library can be
247 named <tt>LLVMDummy</tt> (for example, see the MIPS target). Alternatively, you
248 can split the library into <tt>LLVMDummyCodeGen</tt>
249 and <tt>LLVMDummyAsmPrinter</tt>, the latter of which should be implemented in a
250 subdirectory below <tt>lib/Target/Dummy</tt> (for example, see the PowerPC
255 Note that these two naming schemes are hardcoded into <tt>llvm-config</tt>.
256 Using any other naming scheme will confuse <tt>llvm-config</tt> and produce a
257 lot of (seemingly unrelated) linker errors when linking <tt>llc</tt>.
261 To make your target actually do something, you need to implement a subclass of
262 <tt>TargetMachine</tt>. This implementation should typically be in the file
263 <tt>lib/Target/DummyTargetMachine.cpp</tt>, but any file in
264 the <tt>lib/Target</tt> directory will be built and should work. To use LLVM's
265 target independent code generator, you should do what all current machine
266 backends do: create a subclass of <tt>LLVMTargetMachine</tt>. (To create a
267 target from scratch, create a subclass of <tt>TargetMachine</tt>.)
271 To get LLVM to actually build and link your target, you need to add it to
272 the <tt>TARGETS_TO_BUILD</tt> variable. To do this, you modify the configure
273 script to know about your target when parsing the <tt>--enable-targets</tt>
274 option. Search the configure script for <tt>TARGETS_TO_BUILD</tt>, add your
275 target to the lists there (some creativity required), and then
276 reconfigure. Alternatively, you can change <tt>autotools/configure.ac</tt> and
277 regenerate configure by running <tt>./autoconf/AutoRegen.sh</tt>.
284 <!-- *********************************************************************** -->
286 <a name="TargetMachine">Target Machine</a>
288 <!-- *********************************************************************** -->
293 <tt>LLVMTargetMachine</tt> is designed as a base class for targets implemented
294 with the LLVM target-independent code generator. The <tt>LLVMTargetMachine</tt>
295 class should be specialized by a concrete target class that implements the
296 various virtual methods. <tt>LLVMTargetMachine</tt> is defined as a subclass of
297 <tt>TargetMachine</tt> in <tt>include/llvm/Target/TargetMachine.h</tt>. The
298 <tt>TargetMachine</tt> class implementation (<tt>TargetMachine.cpp</tt>) also
299 processes numerous command-line options.
303 To create a concrete target-specific subclass of <tt>LLVMTargetMachine</tt>,
304 start by copying an existing <tt>TargetMachine</tt> class and header. You
305 should name the files that you create to reflect your specific target. For
306 instance, for the SPARC target, name the files <tt>SparcTargetMachine.h</tt> and
307 <tt>SparcTargetMachine.cpp</tt>.
311 For a target machine <tt>XXX</tt>, the implementation of
312 <tt>XXXTargetMachine</tt> must have access methods to obtain objects that
313 represent target components. These methods are named <tt>get*Info</tt>, and are
314 intended to obtain the instruction set (<tt>getInstrInfo</tt>), register set
315 (<tt>getRegisterInfo</tt>), stack frame layout (<tt>getFrameInfo</tt>), and
316 similar information. <tt>XXXTargetMachine</tt> must also implement the
317 <tt>getTargetData</tt> method to access an object with target-specific data
318 characteristics, such as data type size and alignment requirements.
322 For instance, for the SPARC target, the header file
323 <tt>SparcTargetMachine.h</tt> declares prototypes for several <tt>get*Info</tt>
324 and <tt>getTargetData</tt> methods that simply return a class member.
327 <div class="doc_code">
333 class SparcTargetMachine : public LLVMTargetMachine {
334 const TargetData DataLayout; // Calculates type size & alignment
335 SparcSubtarget Subtarget;
336 SparcInstrInfo InstrInfo;
337 TargetFrameInfo FrameInfo;
340 virtual const TargetAsmInfo *createTargetAsmInfo() const;
343 SparcTargetMachine(const Module &M, const std::string &FS);
345 virtual const SparcInstrInfo *getInstrInfo() const {return &InstrInfo; }
346 virtual const TargetFrameInfo *getFrameInfo() const {return &FrameInfo; }
347 virtual const TargetSubtarget *getSubtargetImpl() const{return &Subtarget; }
348 virtual const TargetRegisterInfo *getRegisterInfo() const {
349 return &InstrInfo.getRegisterInfo();
351 virtual const TargetData *getTargetData() const { return &DataLayout; }
352 static unsigned getModuleMatchQuality(const Module &M);
354 // Pass Pipeline Configuration
355 virtual bool addInstSelector(PassManagerBase &PM, bool Fast);
356 virtual bool addPreEmitPass(PassManagerBase &PM, bool Fast);
359 } // end namespace llvm
364 <li><tt>getInstrInfo()</tt></li>
365 <li><tt>getRegisterInfo()</tt></li>
366 <li><tt>getFrameInfo()</tt></li>
367 <li><tt>getTargetData()</tt></li>
368 <li><tt>getSubtargetImpl()</tt></li>
371 <p>For some targets, you also need to support the following methods:</p>
374 <li><tt>getTargetLowering()</tt></li>
375 <li><tt>getJITInfo()</tt></li>
379 In addition, the <tt>XXXTargetMachine</tt> constructor should specify a
380 <tt>TargetDescription</tt> string that determines the data layout for the target
381 machine, including characteristics such as pointer size, alignment, and
382 endianness. For example, the constructor for SparcTargetMachine contains the
386 <div class="doc_code">
388 SparcTargetMachine::SparcTargetMachine(const Module &M, const std::string &FS)
389 : DataLayout("E-p:32:32-f128:128:128"),
390 Subtarget(M, FS), InstrInfo(Subtarget),
391 FrameInfo(TargetFrameInfo::StackGrowsDown, 8, 0) {
396 <p>Hyphens separate portions of the <tt>TargetDescription</tt> string.</p>
399 <li>An upper-case "<tt>E</tt>" in the string indicates a big-endian target data
400 model. a lower-case "<tt>e</tt>" indicates little-endian.</li>
402 <li>"<tt>p:</tt>" is followed by pointer information: size, ABI alignment, and
403 preferred alignment. If only two figures follow "<tt>p:</tt>", then the
404 first value is pointer size, and the second value is both ABI and preferred
407 <li>Then a letter for numeric type alignment: "<tt>i</tt>", "<tt>f</tt>",
408 "<tt>v</tt>", or "<tt>a</tt>" (corresponding to integer, floating point,
409 vector, or aggregate). "<tt>i</tt>", "<tt>v</tt>", or "<tt>a</tt>" are
410 followed by ABI alignment and preferred alignment. "<tt>f</tt>" is followed
411 by three values: the first indicates the size of a long double, then ABI
412 alignment, and then ABI preferred alignment.</li>
417 <!-- *********************************************************************** -->
419 <a name="TargetRegistration">Target Registration</a>
421 <!-- *********************************************************************** -->
426 You must also register your target with the <tt>TargetRegistry</tt>, which is
427 what other LLVM tools use to be able to lookup and use your target at
428 runtime. The <tt>TargetRegistry</tt> can be used directly, but for most targets
429 there are helper templates which should take care of the work for you.</p>
432 All targets should declare a global <tt>Target</tt> object which is used to
433 represent the target during registration. Then, in the target's TargetInfo
434 library, the target should define that object and use
435 the <tt>RegisterTarget</tt> template to register the target. For example, the Sparc registration code looks like this:
438 <div class="doc_code">
440 Target llvm::TheSparcTarget;
442 extern "C" void LLVMInitializeSparcTargetInfo() {
443 RegisterTarget<Triple::sparc, /*HasJIT=*/false>
444 X(TheSparcTarget, "sparc", "Sparc");
450 This allows the <tt>TargetRegistry</tt> to look up the target by name or by
451 target triple. In addition, most targets will also register additional features
452 which are available in separate libraries. These registration steps are
453 separate, because some clients may wish to only link in some parts of the target
454 -- the JIT code generator does not require the use of the assembler printer, for
455 example. Here is an example of registering the Sparc assembly printer:
458 <div class="doc_code">
460 extern "C" void LLVMInitializeSparcAsmPrinter() {
461 RegisterAsmPrinter<SparcAsmPrinter> X(TheSparcTarget);
467 For more information, see
468 "<a href="/doxygen/TargetRegistry_8h-source.html">llvm/Target/TargetRegistry.h</a>".
473 <!-- *********************************************************************** -->
475 <a name="RegisterSet">Register Set and Register Classes</a>
477 <!-- *********************************************************************** -->
482 You should describe a concrete target-specific class that represents the
483 register file of a target machine. This class is called <tt>XXXRegisterInfo</tt>
484 (where <tt>XXX</tt> identifies the target) and represents the class register
485 file data that is used for register allocation. It also describes the
486 interactions between registers.
490 You also need to define register classes to categorize related registers. A
491 register class should be added for groups of registers that are all treated the
492 same way for some instruction. Typical examples are register classes for
493 integer, floating-point, or vector registers. A register allocator allows an
494 instruction to use any register in a specified register class to perform the
495 instruction in a similar manner. Register classes allocate virtual registers to
496 instructions from these sets, and register classes let the target-independent
497 register allocator automatically choose the actual registers.
501 Much of the code for registers, including register definition, register aliases,
502 and register classes, is generated by TableGen from <tt>XXXRegisterInfo.td</tt>
503 input files and placed in <tt>XXXGenRegisterInfo.h.inc</tt> and
504 <tt>XXXGenRegisterInfo.inc</tt> output files. Some of the code in the
505 implementation of <tt>XXXRegisterInfo</tt> requires hand-coding.
508 <!-- ======================================================================= -->
510 <a name="RegisterDef">Defining a Register</a>
516 The <tt>XXXRegisterInfo.td</tt> file typically starts with register definitions
517 for a target machine. The <tt>Register</tt> class (specified
518 in <tt>Target.td</tt>) is used to define an object for each register. The
519 specified string <tt>n</tt> becomes the <tt>Name</tt> of the register. The
520 basic <tt>Register</tt> object does not have any subregisters and does not
524 <div class="doc_code">
526 class Register<string n> {
527 string Namespace = "";
531 int SpillAlignment = 0;
532 list<Register> Aliases = [];
533 list<Register> SubRegs = [];
534 list<int> DwarfNumbers = [];
540 For example, in the <tt>X86RegisterInfo.td</tt> file, there are register
541 definitions that utilize the Register class, such as:
544 <div class="doc_code">
546 def AL : Register<"AL">, DwarfRegNum<[0, 0, 0]>;
551 This defines the register <tt>AL</tt> and assigns it values (with
552 <tt>DwarfRegNum</tt>) that are used by <tt>gcc</tt>, <tt>gdb</tt>, or a debug
553 information writer to identify a register. For register
554 <tt>AL</tt>, <tt>DwarfRegNum</tt> takes an array of 3 values representing 3
555 different modes: the first element is for X86-64, the second for exception
556 handling (EH) on X86-32, and the third is generic. -1 is a special Dwarf number
557 that indicates the gcc number is undefined, and -2 indicates the register number
558 is invalid for this mode.
562 From the previously described line in the <tt>X86RegisterInfo.td</tt> file,
563 TableGen generates this code in the <tt>X86GenRegisterInfo.inc</tt> file:
566 <div class="doc_code">
568 static const unsigned GR8[] = { X86::AL, ... };
570 const unsigned AL_AliasSet[] = { X86::AX, X86::EAX, X86::RAX, 0 };
572 const TargetRegisterDesc RegisterDescriptors[] = {
574 { "AL", "AL", AL_AliasSet, Empty_SubRegsSet, Empty_SubRegsSet, AL_SuperRegsSet }, ...
579 From the register info file, TableGen generates a <tt>TargetRegisterDesc</tt>
580 object for each register. <tt>TargetRegisterDesc</tt> is defined in
581 <tt>include/llvm/Target/TargetRegisterInfo.h</tt> with the following fields:
584 <div class="doc_code">
586 struct TargetRegisterDesc {
587 const char *AsmName; // Assembly language name for the register
588 const char *Name; // Printable name for the reg (for debugging)
589 const unsigned *AliasSet; // Register Alias Set
590 const unsigned *SubRegs; // Sub-register set
591 const unsigned *ImmSubRegs; // Immediate sub-register set
592 const unsigned *SuperRegs; // Super-register set
597 TableGen uses the entire target description file (<tt>.td</tt>) to determine
598 text names for the register (in the <tt>AsmName</tt> and <tt>Name</tt> fields of
599 <tt>TargetRegisterDesc</tt>) and the relationships of other registers to the
600 defined register (in the other <tt>TargetRegisterDesc</tt> fields). In this
601 example, other definitions establish the registers "<tt>AX</tt>",
602 "<tt>EAX</tt>", and "<tt>RAX</tt>" as aliases for one another, so TableGen
603 generates a null-terminated array (<tt>AL_AliasSet</tt>) for this register alias
608 The <tt>Register</tt> class is commonly used as a base class for more complex
609 classes. In <tt>Target.td</tt>, the <tt>Register</tt> class is the base for the
610 <tt>RegisterWithSubRegs</tt> class that is used to define registers that need to
611 specify subregisters in the <tt>SubRegs</tt> list, as shown here:
614 <div class="doc_code">
616 class RegisterWithSubRegs<string n,
617 list<Register> subregs> : Register<n> {
618 let SubRegs = subregs;
624 In <tt>SparcRegisterInfo.td</tt>, additional register classes are defined for
625 SPARC: a Register subclass, SparcReg, and further subclasses: <tt>Ri</tt>,
626 <tt>Rf</tt>, and <tt>Rd</tt>. SPARC registers are identified by 5-bit ID
627 numbers, which is a feature common to these subclasses. Note the use of
628 '<tt>let</tt>' expressions to override values that are initially defined in a
629 superclass (such as <tt>SubRegs</tt> field in the <tt>Rd</tt> class).
632 <div class="doc_code">
634 class SparcReg<string n> : Register<n> {
635 field bits<5> Num;
636 let Namespace = "SP";
638 // Ri - 32-bit integer registers
639 class Ri<bits<5> num, string n> :
643 // Rf - 32-bit floating-point registers
644 class Rf<bits<5> num, string n> :
648 // Rd - Slots in the FP register file for 64-bit
649 floating-point values.
650 class Rd<bits<5> num, string n,
651 list<Register> subregs> : SparcReg<n> {
653 let SubRegs = subregs;
659 In the <tt>SparcRegisterInfo.td</tt> file, there are register definitions that
660 utilize these subclasses of <tt>Register</tt>, such as:
663 <div class="doc_code">
665 def G0 : Ri< 0, "G0">,
666 DwarfRegNum<[0]>;
667 def G1 : Ri< 1, "G1">, DwarfRegNum<[1]>;
669 def F0 : Rf< 0, "F0">,
670 DwarfRegNum<[32]>;
671 def F1 : Rf< 1, "F1">,
672 DwarfRegNum<[33]>;
674 def D0 : Rd< 0, "F0", [F0, F1]>,
675 DwarfRegNum<[32]>;
676 def D1 : Rd< 2, "F2", [F2, F3]>,
677 DwarfRegNum<[34]>;
682 The last two registers shown above (<tt>D0</tt> and <tt>D1</tt>) are
683 double-precision floating-point registers that are aliases for pairs of
684 single-precision floating-point sub-registers. In addition to aliases, the
685 sub-register and super-register relationships of the defined register are in
686 fields of a register's TargetRegisterDesc.
691 <!-- ======================================================================= -->
693 <a name="RegisterClassDef">Defining a Register Class</a>
699 The <tt>RegisterClass</tt> class (specified in <tt>Target.td</tt>) is used to
700 define an object that represents a group of related registers and also defines
701 the default allocation order of the registers. A target description file
702 <tt>XXXRegisterInfo.td</tt> that uses <tt>Target.td</tt> can construct register
703 classes using the following class:
706 <div class="doc_code">
708 class RegisterClass<string namespace,
709 list<ValueType> regTypes, int alignment, dag regList> {
710 string Namespace = namespace;
711 list<ValueType> RegTypes = regTypes;
712 int Size = 0; // spill size, in bits; zero lets tblgen pick the size
713 int Alignment = alignment;
715 // CopyCost is the cost of copying a value between two registers
716 // default value 1 means a single instruction
717 // A negative value means copying is extremely expensive or impossible
719 dag MemberList = regList;
721 // for register classes that are subregisters of this class
722 list<RegisterClass> SubRegClassList = [];
724 code MethodProtos = [{}]; // to insert arbitrary code
725 code MethodBodies = [{}];
730 <p>To define a RegisterClass, use the following 4 arguments:</p>
733 <li>The first argument of the definition is the name of the namespace.</li>
735 <li>The second argument is a list of <tt>ValueType</tt> register type values
736 that are defined in <tt>include/llvm/CodeGen/ValueTypes.td</tt>. Defined
737 values include integer types (such as <tt>i16</tt>, <tt>i32</tt>,
738 and <tt>i1</tt> for Boolean), floating-point types
739 (<tt>f32</tt>, <tt>f64</tt>), and vector types (for example, <tt>v8i16</tt>
740 for an <tt>8 x i16</tt> vector). All registers in a <tt>RegisterClass</tt>
741 must have the same <tt>ValueType</tt>, but some registers may store vector
742 data in different configurations. For example a register that can process a
743 128-bit vector may be able to handle 16 8-bit integer elements, 8 16-bit
744 integers, 4 32-bit integers, and so on. </li>
746 <li>The third argument of the <tt>RegisterClass</tt> definition specifies the
747 alignment required of the registers when they are stored or loaded to
750 <li>The final argument, <tt>regList</tt>, specifies which registers are in this
751 class. If an alternative allocation order method is not specified, then
752 <tt>regList</tt> also defines the order of allocation used by the register
753 allocator. Besides simply listing registers with <tt>(add R0, R1, ...)</tt>,
754 more advanced set operators are available. See
755 <tt>include/llvm/Target/Target.td</tt> for more information.</li>
759 In <tt>SparcRegisterInfo.td</tt>, three RegisterClass objects are defined:
760 <tt>FPRegs</tt>, <tt>DFPRegs</tt>, and <tt>IntRegs</tt>. For all three register
761 classes, the first argument defines the namespace with the string
762 '<tt>SP</tt>'. <tt>FPRegs</tt> defines a group of 32 single-precision
763 floating-point registers (<tt>F0</tt> to <tt>F31</tt>); <tt>DFPRegs</tt> defines
764 a group of 16 double-precision registers
768 <div class="doc_code">
770 // F0, F1, F2, ..., F31
771 def FPRegs : RegisterClass<"SP", [f32], 32, (sequence "F%u", 0, 31)>;
773 def DFPRegs : RegisterClass<"SP", [f64], 64,
774 (add D0, D1, D2, D3, D4, D5, D6, D7, D8,
775 D9, D10, D11, D12, D13, D14, D15)>;
777 def IntRegs : RegisterClass<"SP", [i32], 32,
778 (add L0, L1, L2, L3, L4, L5, L6, L7,
779 I0, I1, I2, I3, I4, I5,
780 O0, O1, O2, O3, O4, O5, O7,
782 // Non-allocatable regs:
786 I7, // return address
788 G5, G6, G7 // reserved for kernel
794 Using <tt>SparcRegisterInfo.td</tt> with TableGen generates several output files
795 that are intended for inclusion in other source code that you write.
796 <tt>SparcRegisterInfo.td</tt> generates <tt>SparcGenRegisterInfo.h.inc</tt>,
797 which should be included in the header file for the implementation of the SPARC
798 register implementation that you write (<tt>SparcRegisterInfo.h</tt>). In
799 <tt>SparcGenRegisterInfo.h.inc</tt> a new structure is defined called
800 <tt>SparcGenRegisterInfo</tt> that uses <tt>TargetRegisterInfo</tt> as its
801 base. It also specifies types, based upon the defined register
802 classes: <tt>DFPRegsClass</tt>, <tt>FPRegsClass</tt>, and <tt>IntRegsClass</tt>.
806 <tt>SparcRegisterInfo.td</tt> also generates <tt>SparcGenRegisterInfo.inc</tt>,
807 which is included at the bottom of <tt>SparcRegisterInfo.cpp</tt>, the SPARC
808 register implementation. The code below shows only the generated integer
809 registers and associated register classes. The order of registers
810 in <tt>IntRegs</tt> reflects the order in the definition of <tt>IntRegs</tt> in
811 the target description file.
814 <div class="doc_code">
815 <pre> // IntRegs Register Class...
816 static const unsigned IntRegs[] = {
817 SP::L0, SP::L1, SP::L2, SP::L3, SP::L4, SP::L5,
818 SP::L6, SP::L7, SP::I0, SP::I1, SP::I2, SP::I3,
819 SP::I4, SP::I5, SP::O0, SP::O1, SP::O2, SP::O3,
820 SP::O4, SP::O5, SP::O7, SP::G1, SP::G2, SP::G3,
821 SP::G4, SP::O6, SP::I6, SP::I7, SP::G0, SP::G5,
825 // IntRegsVTs Register Class Value Types...
826 static const MVT::ValueType IntRegsVTs[] = {
830 namespace SP { // Register class instances
831 DFPRegsClass DFPRegsRegClass;
832 FPRegsClass FPRegsRegClass;
833 IntRegsClass IntRegsRegClass;
835 // IntRegs Sub-register Classess...
836 static const TargetRegisterClass* const IntRegsSubRegClasses [] = {
840 // IntRegs Super-register Classess...
841 static const TargetRegisterClass* const IntRegsSuperRegClasses [] = {
845 // IntRegs Register Class sub-classes...
846 static const TargetRegisterClass* const IntRegsSubclasses [] = {
850 // IntRegs Register Class super-classes...
851 static const TargetRegisterClass* const IntRegsSuperclasses [] = {
855 IntRegsClass::IntRegsClass() : TargetRegisterClass(IntRegsRegClassID,
856 IntRegsVTs, IntRegsSubclasses, IntRegsSuperclasses, IntRegsSubRegClasses,
857 IntRegsSuperRegClasses, 4, 4, 1, IntRegs, IntRegs + 32) {}
863 The register allocators will avoid using reserved registers, and callee saved
864 registers are not used until all the volatile registers have been used. That
865 is usually good enough, but in some cases it may be necessary to provide custom
871 <!-- ======================================================================= -->
873 <a name="implementRegister">Implement a subclass of</a>
874 <a href="CodeGenerator.html#targetregisterinfo">TargetRegisterInfo</a>
880 The final step is to hand code portions of <tt>XXXRegisterInfo</tt>, which
881 implements the interface described in <tt>TargetRegisterInfo.h</tt>. These
882 functions return <tt>0</tt>, <tt>NULL</tt>, or <tt>false</tt>, unless
883 overridden. Here is a list of functions that are overridden for the SPARC
884 implementation in <tt>SparcRegisterInfo.cpp</tt>:
888 <li><tt>getCalleeSavedRegs</tt> — Returns a list of callee-saved registers
889 in the order of the desired callee-save stack frame offset.</li>
891 <li><tt>getReservedRegs</tt> — Returns a bitset indexed by physical
892 register numbers, indicating if a particular register is unavailable.</li>
894 <li><tt>hasFP</tt> — Return a Boolean indicating if a function should have
895 a dedicated frame pointer register.</li>
897 <li><tt>eliminateCallFramePseudoInstr</tt> — If call frame setup or
898 destroy pseudo instructions are used, this can be called to eliminate
901 <li><tt>eliminateFrameIndex</tt> — Eliminate abstract frame indices from
902 instructions that may use them.</li>
904 <li><tt>emitPrologue</tt> — Insert prologue code into the function.</li>
906 <li><tt>emitEpilogue</tt> — Insert epilogue code into the function.</li>
913 <!-- *********************************************************************** -->
915 <a name="InstructionSet">Instruction Set</a>
918 <!-- *********************************************************************** -->
922 During the early stages of code generation, the LLVM IR code is converted to a
923 <tt>SelectionDAG</tt> with nodes that are instances of the <tt>SDNode</tt> class
924 containing target instructions. An <tt>SDNode</tt> has an opcode, operands, type
925 requirements, and operation properties. For example, is an operation
926 commutative, does an operation load from memory. The various operation node
927 types are described in the <tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt>
928 file (values of the <tt>NodeType</tt> enum in the <tt>ISD</tt> namespace).
932 TableGen uses the following target description (<tt>.td</tt>) input files to
933 generate much of the code for instruction definition:
937 <li><tt>Target.td</tt> — Where the <tt>Instruction</tt>, <tt>Operand</tt>,
938 <tt>InstrInfo</tt>, and other fundamental classes are defined.</li>
940 <li><tt>TargetSelectionDAG.td</tt>— Used by <tt>SelectionDAG</tt>
941 instruction selection generators, contains <tt>SDTC*</tt> classes (selection
942 DAG type constraint), definitions of <tt>SelectionDAG</tt> nodes (such as
943 <tt>imm</tt>, <tt>cond</tt>, <tt>bb</tt>, <tt>add</tt>, <tt>fadd</tt>,
944 <tt>sub</tt>), and pattern support (<tt>Pattern</tt>, <tt>Pat</tt>,
945 <tt>PatFrag</tt>, <tt>PatLeaf</tt>, <tt>ComplexPattern</tt>.</li>
947 <li><tt>XXXInstrFormats.td</tt> — Patterns for definitions of
948 target-specific instructions.</li>
950 <li><tt>XXXInstrInfo.td</tt> — Target-specific definitions of instruction
951 templates, condition codes, and instructions of an instruction set. For
952 architecture modifications, a different file name may be used. For example,
953 for Pentium with SSE instruction, this file is <tt>X86InstrSSE.td</tt>, and
954 for Pentium with MMX, this file is <tt>X86InstrMMX.td</tt>.</li>
958 There is also a target-specific <tt>XXX.td</tt> file, where <tt>XXX</tt> is the
959 name of the target. The <tt>XXX.td</tt> file includes the other <tt>.td</tt>
960 input files, but its contents are only directly important for subtargets.
964 You should describe a concrete target-specific class <tt>XXXInstrInfo</tt> that
965 represents machine instructions supported by a target machine.
966 <tt>XXXInstrInfo</tt> contains an array of <tt>XXXInstrDescriptor</tt> objects,
967 each of which describes one instruction. An instruction descriptor defines:</p>
970 <li>Opcode mnemonic</li>
972 <li>Number of operands</li>
974 <li>List of implicit register definitions and uses</li>
976 <li>Target-independent properties (such as memory access, is commutable)</li>
978 <li>Target-specific flags </li>
982 The Instruction class (defined in <tt>Target.td</tt>) is mostly used as a base
983 for more complex instruction classes.
986 <div class="doc_code">
987 <pre>class Instruction {
988 string Namespace = "";
989 dag OutOperandList; // An dag containing the MI def operand list.
990 dag InOperandList; // An dag containing the MI use operand list.
991 string AsmString = ""; // The .s format to print the instruction with.
992 list<dag> Pattern; // Set to the DAG pattern for this instruction
993 list<Register> Uses = [];
994 list<Register> Defs = [];
995 list<Predicate> Predicates = []; // predicates turned into isel match code
996 ... remainder not shown for space ...
1002 A <tt>SelectionDAG</tt> node (<tt>SDNode</tt>) should contain an object
1003 representing a target-specific instruction that is defined
1004 in <tt>XXXInstrInfo.td</tt>. The instruction objects should represent
1005 instructions from the architecture manual of the target machine (such as the
1006 SPARC Architecture Manual for the SPARC target).
1010 A single instruction from the architecture manual is often modeled as multiple
1011 target instructions, depending upon its operands. For example, a manual might
1012 describe an add instruction that takes a register or an immediate operand. An
1013 LLVM target could model this with two instructions named <tt>ADDri</tt> and
1018 You should define a class for each instruction category and define each opcode
1019 as a subclass of the category with appropriate parameters such as the fixed
1020 binary encoding of opcodes and extended opcodes. You should map the register
1021 bits to the bits of the instruction in which they are encoded (for the
1022 JIT). Also you should specify how the instruction should be printed when the
1023 automatic assembly printer is used.
1027 As is described in the SPARC Architecture Manual, Version 8, there are three
1028 major 32-bit formats for instructions. Format 1 is only for the <tt>CALL</tt>
1029 instruction. Format 2 is for branch on condition codes and <tt>SETHI</tt> (set
1030 high bits of a register) instructions. Format 3 is for other instructions.
1034 Each of these formats has corresponding classes in <tt>SparcInstrFormat.td</tt>.
1035 <tt>InstSP</tt> is a base class for other instruction classes. Additional base
1036 classes are specified for more precise formats: for example
1037 in <tt>SparcInstrFormat.td</tt>, <tt>F2_1</tt> is for <tt>SETHI</tt>,
1038 and <tt>F2_2</tt> is for branches. There are three other base
1039 classes: <tt>F3_1</tt> for register/register operations, <tt>F3_2</tt> for
1040 register/immediate operations, and <tt>F3_3</tt> for floating-point
1041 operations. <tt>SparcInstrInfo.td</tt> also adds the base class Pseudo for
1042 synthetic SPARC instructions.
1046 <tt>SparcInstrInfo.td</tt> largely consists of operand and instruction
1047 definitions for the SPARC target. In <tt>SparcInstrInfo.td</tt>, the following
1048 target description file entry, <tt>LDrr</tt>, defines the Load Integer
1049 instruction for a Word (the <tt>LD</tt> SPARC opcode) from a memory address to a
1050 register. The first parameter, the value 3 (<tt>11<sub>2</sub></tt>), is the
1051 operation value for this category of operation. The second parameter
1052 (<tt>000000<sub>2</sub></tt>) is the specific operation value
1053 for <tt>LD</tt>/Load Word. The third parameter is the output destination, which
1054 is a register operand and defined in the <tt>Register</tt> target description
1055 file (<tt>IntRegs</tt>).
1058 <div class="doc_code">
1059 <pre>def LDrr : F3_1 <3, 0b000000, (outs IntRegs:$dst), (ins MEMrr:$addr),
1061 [(set IntRegs:$dst, (load ADDRrr:$addr))]>;
1066 The fourth parameter is the input source, which uses the address
1067 operand <tt>MEMrr</tt> that is defined earlier in <tt>SparcInstrInfo.td</tt>:
1070 <div class="doc_code">
1071 <pre>def MEMrr : Operand<i32> {
1072 let PrintMethod = "printMemOperand";
1073 let MIOperandInfo = (ops IntRegs, IntRegs);
1079 The fifth parameter is a string that is used by the assembly printer and can be
1080 left as an empty string until the assembly printer interface is implemented. The
1081 sixth and final parameter is the pattern used to match the instruction during
1082 the SelectionDAG Select Phase described in
1083 (<a href="CodeGenerator.html">The LLVM
1084 Target-Independent Code Generator</a>). This parameter is detailed in the next
1085 section, <a href="#InstructionSelector">Instruction Selector</a>.
1089 Instruction class definitions are not overloaded for different operand types, so
1090 separate versions of instructions are needed for register, memory, or immediate
1091 value operands. For example, to perform a Load Integer instruction for a Word
1092 from an immediate operand to a register, the following instruction class is
1096 <div class="doc_code">
1097 <pre>def LDri : F3_2 <3, 0b000000, (outs IntRegs:$dst), (ins MEMri:$addr),
1099 [(set IntRegs:$dst, (load ADDRri:$addr))]>;
1104 Writing these definitions for so many similar instructions can involve a lot of
1105 cut and paste. In td files, the <tt>multiclass</tt> directive enables the
1106 creation of templates to define several instruction classes at once (using
1107 the <tt>defm</tt> directive). For example in <tt>SparcInstrInfo.td</tt>, the
1108 <tt>multiclass</tt> pattern <tt>F3_12</tt> is defined to create 2 instruction
1109 classes each time <tt>F3_12</tt> is invoked:
1112 <div class="doc_code">
1113 <pre>multiclass F3_12 <string OpcStr, bits<6> Op3Val, SDNode OpNode> {
1114 def rr : F3_1 <2, Op3Val,
1115 (outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c),
1116 !strconcat(OpcStr, " $b, $c, $dst"),
1117 [(set IntRegs:$dst, (OpNode IntRegs:$b, IntRegs:$c))]>;
1118 def ri : F3_2 <2, Op3Val,
1119 (outs IntRegs:$dst), (ins IntRegs:$b, i32imm:$c),
1120 !strconcat(OpcStr, " $b, $c, $dst"),
1121 [(set IntRegs:$dst, (OpNode IntRegs:$b, simm13:$c))]>;
1127 So when the <tt>defm</tt> directive is used for the <tt>XOR</tt>
1128 and <tt>ADD</tt> instructions, as seen below, it creates four instruction
1129 objects: <tt>XORrr</tt>, <tt>XORri</tt>, <tt>ADDrr</tt>, and <tt>ADDri</tt>.
1132 <div class="doc_code">
1134 defm XOR : F3_12<"xor", 0b000011, xor>;
1135 defm ADD : F3_12<"add", 0b000000, add>;
1140 <tt>SparcInstrInfo.td</tt> also includes definitions for condition codes that
1141 are referenced by branch instructions. The following definitions
1142 in <tt>SparcInstrInfo.td</tt> indicate the bit location of the SPARC condition
1143 code. For example, the 10<sup>th</sup> bit represents the 'greater than'
1144 condition for integers, and the 22<sup>nd</sup> bit represents the 'greater
1145 than' condition for floats.
1148 <div class="doc_code">
1150 def ICC_NE : ICC_VAL< 9>; // Not Equal
1151 def ICC_E : ICC_VAL< 1>; // Equal
1152 def ICC_G : ICC_VAL<10>; // Greater
1154 def FCC_U : FCC_VAL<23>; // Unordered
1155 def FCC_G : FCC_VAL<22>; // Greater
1156 def FCC_UG : FCC_VAL<21>; // Unordered or Greater
1162 (Note that <tt>Sparc.h</tt> also defines enums that correspond to the same SPARC
1163 condition codes. Care must be taken to ensure the values in <tt>Sparc.h</tt>
1164 correspond to the values in <tt>SparcInstrInfo.td</tt>. I.e.,
1165 <tt>SPCC::ICC_NE = 9</tt>, <tt>SPCC::FCC_U = 23</tt> and so on.)
1168 <!-- ======================================================================= -->
1170 <a name="operandMapping">Instruction Operand Mapping</a>
1176 The code generator backend maps instruction operands to fields in the
1177 instruction. Operands are assigned to unbound fields in the instruction in the
1178 order they are defined. Fields are bound when they are assigned a value. For
1179 example, the Sparc target defines the <tt>XNORrr</tt> instruction as
1180 a <tt>F3_1</tt> format instruction having three operands.
1183 <div class="doc_code">
1185 def XNORrr : F3_1<2, 0b000111,
1186 (outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c),
1187 "xnor $b, $c, $dst",
1188 [(set IntRegs:$dst, (not (xor IntRegs:$b, IntRegs:$c)))]>;
1193 The instruction templates in <tt>SparcInstrFormats.td</tt> show the base class
1194 for <tt>F3_1</tt> is <tt>InstSP</tt>.
1197 <div class="doc_code">
1199 class InstSP<dag outs, dag ins, string asmstr, list<dag> pattern> : Instruction {
1200 field bits<32> Inst;
1201 let Namespace = "SP";
1203 let Inst{31-30} = op;
1204 dag OutOperandList = outs;
1205 dag InOperandList = ins;
1206 let AsmString = asmstr;
1207 let Pattern = pattern;
1212 <p><tt>InstSP</tt> leaves the <tt>op</tt> field unbound.</p>
1214 <div class="doc_code">
1216 class F3<dag outs, dag ins, string asmstr, list<dag> pattern>
1217 : InstSP<outs, ins, asmstr, pattern> {
1221 let op{1} = 1; // Op = 2 or 3
1222 let Inst{29-25} = rd;
1223 let Inst{24-19} = op3;
1224 let Inst{18-14} = rs1;
1230 <tt>F3</tt> binds the <tt>op</tt> field and defines the <tt>rd</tt>,
1231 <tt>op3</tt>, and <tt>rs1</tt> fields. <tt>F3</tt> format instructions will
1232 bind the operands <tt>rd</tt>, <tt>op3</tt>, and <tt>rs1</tt> fields.
1235 <div class="doc_code">
1237 class F3_1<bits<2> opVal, bits<6> op3val, dag outs, dag ins,
1238 string asmstr, list<dag> pattern> : F3<outs, ins, asmstr, pattern> {
1239 bits<8> asi = 0; // asi not currently used
1243 let Inst{13} = 0; // i field = 0
1244 let Inst{12-5} = asi; // address space identifier
1245 let Inst{4-0} = rs2;
1251 <tt>F3_1</tt> binds the <tt>op3</tt> field and defines the <tt>rs2</tt>
1252 fields. <tt>F3_1</tt> format instructions will bind the operands to the <tt>rd</tt>,
1253 <tt>rs1</tt>, and <tt>rs2</tt> fields. This results in the <tt>XNORrr</tt>
1254 instruction binding <tt>$dst</tt>, <tt>$b</tt>, and <tt>$c</tt> operands to
1255 the <tt>rd</tt>, <tt>rs1</tt>, and <tt>rs2</tt> fields respectively.
1260 <!-- ======================================================================= -->
1262 <a name="implementInstr">Implement a subclass of </a>
1263 <a href="CodeGenerator.html#targetinstrinfo">TargetInstrInfo</a>
1269 The final step is to hand code portions of <tt>XXXInstrInfo</tt>, which
1270 implements the interface described in <tt>TargetInstrInfo.h</tt>. These
1271 functions return <tt>0</tt> or a Boolean or they assert, unless
1272 overridden. Here's a list of functions that are overridden for the SPARC
1273 implementation in <tt>SparcInstrInfo.cpp</tt>:
1277 <li><tt>isLoadFromStackSlot</tt> — If the specified machine instruction is
1278 a direct load from a stack slot, return the register number of the
1279 destination and the <tt>FrameIndex</tt> of the stack slot.</li>
1281 <li><tt>isStoreToStackSlot</tt> — If the specified machine instruction is
1282 a direct store to a stack slot, return the register number of the
1283 destination and the <tt>FrameIndex</tt> of the stack slot.</li>
1285 <li><tt>copyPhysReg</tt> — Copy values between a pair of physical
1288 <li><tt>storeRegToStackSlot</tt> — Store a register value to a stack
1291 <li><tt>loadRegFromStackSlot</tt> — Load a register value from a stack
1294 <li><tt>storeRegToAddr</tt> — Store a register value to memory.</li>
1296 <li><tt>loadRegFromAddr</tt> — Load a register value from memory.</li>
1298 <li><tt>foldMemoryOperand</tt> — Attempt to combine instructions of any
1299 load or store instruction for the specified operand(s).</li>
1304 <!-- ======================================================================= -->
1306 <a name="branchFolding">Branch Folding and If Conversion</a>
1311 Performance can be improved by combining instructions or by eliminating
1312 instructions that are never reached. The <tt>AnalyzeBranch</tt> method
1313 in <tt>XXXInstrInfo</tt> may be implemented to examine conditional instructions
1314 and remove unnecessary instructions. <tt>AnalyzeBranch</tt> looks at the end of
1315 a machine basic block (MBB) for opportunities for improvement, such as branch
1316 folding and if conversion. The <tt>BranchFolder</tt> and <tt>IfConverter</tt>
1317 machine function passes (see the source files <tt>BranchFolding.cpp</tt> and
1318 <tt>IfConversion.cpp</tt> in the <tt>lib/CodeGen</tt> directory) call
1319 <tt>AnalyzeBranch</tt> to improve the control flow graph that represents the
1324 Several implementations of <tt>AnalyzeBranch</tt> (for ARM, Alpha, and X86) can
1325 be examined as models for your own <tt>AnalyzeBranch</tt> implementation. Since
1326 SPARC does not implement a useful <tt>AnalyzeBranch</tt>, the ARM target
1327 implementation is shown below.
1330 <p><tt>AnalyzeBranch</tt> returns a Boolean value and takes four parameters:</p>
1333 <li><tt>MachineBasicBlock &MBB</tt> — The incoming block to be
1336 <li><tt>MachineBasicBlock *&TBB</tt> — A destination block that is
1337 returned. For a conditional branch that evaluates to true, <tt>TBB</tt> is
1338 the destination.</li>
1340 <li><tt>MachineBasicBlock *&FBB</tt> — For a conditional branch that
1341 evaluates to false, <tt>FBB</tt> is returned as the destination.</li>
1343 <li><tt>std::vector<MachineOperand> &Cond</tt> — List of
1344 operands to evaluate a condition for a conditional branch.</li>
1348 In the simplest case, if a block ends without a branch, then it falls through to
1349 the successor block. No destination blocks are specified for either <tt>TBB</tt>
1350 or <tt>FBB</tt>, so both parameters return <tt>NULL</tt>. The start of
1351 the <tt>AnalyzeBranch</tt> (see code below for the ARM target) shows the
1352 function parameters and the code for the simplest case.
1355 <div class="doc_code">
1356 <pre>bool ARMInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
1357 MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
1358 std::vector<MachineOperand> &Cond) const
1360 MachineBasicBlock::iterator I = MBB.end();
1361 if (I == MBB.begin() || !isUnpredicatedTerminator(--I))
1367 If a block ends with a single unconditional branch instruction, then
1368 <tt>AnalyzeBranch</tt> (shown below) should return the destination of that
1369 branch in the <tt>TBB</tt> parameter.
1372 <div class="doc_code">
1374 if (LastOpc == ARM::B || LastOpc == ARM::tB) {
1375 TBB = LastInst->getOperand(0).getMBB();
1382 If a block ends with two unconditional branches, then the second branch is never
1383 reached. In that situation, as shown below, remove the last branch instruction
1384 and return the penultimate branch in the <tt>TBB</tt> parameter.
1387 <div class="doc_code">
1389 if ((SecondLastOpc == ARM::B || SecondLastOpc==ARM::tB) &&
1390 (LastOpc == ARM::B || LastOpc == ARM::tB)) {
1391 TBB = SecondLastInst->getOperand(0).getMBB();
1393 I->eraseFromParent();
1400 A block may end with a single conditional branch instruction that falls through
1401 to successor block if the condition evaluates to false. In that case,
1402 <tt>AnalyzeBranch</tt> (shown below) should return the destination of that
1403 conditional branch in the <tt>TBB</tt> parameter and a list of operands in
1404 the <tt>Cond</tt> parameter to evaluate the condition.
1407 <div class="doc_code">
1409 if (LastOpc == ARM::Bcc || LastOpc == ARM::tBcc) {
1410 // Block ends with fall-through condbranch.
1411 TBB = LastInst->getOperand(0).getMBB();
1412 Cond.push_back(LastInst->getOperand(1));
1413 Cond.push_back(LastInst->getOperand(2));
1420 If a block ends with both a conditional branch and an ensuing unconditional
1421 branch, then <tt>AnalyzeBranch</tt> (shown below) should return the conditional
1422 branch destination (assuming it corresponds to a conditional evaluation of
1423 '<tt>true</tt>') in the <tt>TBB</tt> parameter and the unconditional branch
1424 destination in the <tt>FBB</tt> (corresponding to a conditional evaluation of
1425 '<tt>false</tt>'). A list of operands to evaluate the condition should be
1426 returned in the <tt>Cond</tt> parameter.
1429 <div class="doc_code">
1431 unsigned SecondLastOpc = SecondLastInst->getOpcode();
1433 if ((SecondLastOpc == ARM::Bcc && LastOpc == ARM::B) ||
1434 (SecondLastOpc == ARM::tBcc && LastOpc == ARM::tB)) {
1435 TBB = SecondLastInst->getOperand(0).getMBB();
1436 Cond.push_back(SecondLastInst->getOperand(1));
1437 Cond.push_back(SecondLastInst->getOperand(2));
1438 FBB = LastInst->getOperand(0).getMBB();
1445 For the last two cases (ending with a single conditional branch or ending with
1446 one conditional and one unconditional branch), the operands returned in
1447 the <tt>Cond</tt> parameter can be passed to methods of other instructions to
1448 create new branches or perform other operations. An implementation
1449 of <tt>AnalyzeBranch</tt> requires the helper methods <tt>RemoveBranch</tt>
1450 and <tt>InsertBranch</tt> to manage subsequent operations.
1454 <tt>AnalyzeBranch</tt> should return false indicating success in most circumstances.
1455 <tt>AnalyzeBranch</tt> should only return true when the method is stumped about what to
1456 do, for example, if a block has three terminating branches. <tt>AnalyzeBranch</tt> may
1457 return true if it encounters a terminator it cannot handle, such as an indirect
1465 <!-- *********************************************************************** -->
1467 <a name="InstructionSelector">Instruction Selector</a>
1469 <!-- *********************************************************************** -->
1474 LLVM uses a <tt>SelectionDAG</tt> to represent LLVM IR instructions, and nodes
1475 of the <tt>SelectionDAG</tt> ideally represent native target
1476 instructions. During code generation, instruction selection passes are performed
1477 to convert non-native DAG instructions into native target-specific
1478 instructions. The pass described in <tt>XXXISelDAGToDAG.cpp</tt> is used to
1479 match patterns and perform DAG-to-DAG instruction selection. Optionally, a pass
1480 may be defined (in <tt>XXXBranchSelector.cpp</tt>) to perform similar DAG-to-DAG
1481 operations for branch instructions. Later, the code in
1482 <tt>XXXISelLowering.cpp</tt> replaces or removes operations and data types not
1483 supported natively (legalizes) in a <tt>SelectionDAG</tt>.
1487 TableGen generates code for instruction selection using the following target
1488 description input files:
1492 <li><tt>XXXInstrInfo.td</tt> — Contains definitions of instructions in a
1493 target-specific instruction set, generates <tt>XXXGenDAGISel.inc</tt>, which
1494 is included in <tt>XXXISelDAGToDAG.cpp</tt>.</li>
1496 <li><tt>XXXCallingConv.td</tt> — Contains the calling and return value
1497 conventions for the target architecture, and it generates
1498 <tt>XXXGenCallingConv.inc</tt>, which is included in
1499 <tt>XXXISelLowering.cpp</tt>.</li>
1503 The implementation of an instruction selection pass must include a header that
1504 declares the <tt>FunctionPass</tt> class or a subclass of <tt>FunctionPass</tt>. In
1505 <tt>XXXTargetMachine.cpp</tt>, a Pass Manager (PM) should add each instruction
1506 selection pass into the queue of passes to run.
1510 The LLVM static compiler (<tt>llc</tt>) is an excellent tool for visualizing the
1511 contents of DAGs. To display the <tt>SelectionDAG</tt> before or after specific
1512 processing phases, use the command line options for <tt>llc</tt>, described
1513 at <a href="CodeGenerator.html#selectiondag_process">
1514 SelectionDAG Instruction Selection Process</a>.
1518 To describe instruction selector behavior, you should add patterns for lowering
1519 LLVM code into a <tt>SelectionDAG</tt> as the last parameter of the instruction
1520 definitions in <tt>XXXInstrInfo.td</tt>. For example, in
1521 <tt>SparcInstrInfo.td</tt>, this entry defines a register store operation, and
1522 the last parameter describes a pattern with the store DAG operator.
1525 <div class="doc_code">
1527 def STrr : F3_1< 3, 0b000100, (outs), (ins MEMrr:$addr, IntRegs:$src),
1528 "st $src, [$addr]", [(store IntRegs:$src, ADDRrr:$addr)]>;
1533 <tt>ADDRrr</tt> is a memory mode that is also defined in
1534 <tt>SparcInstrInfo.td</tt>:
1537 <div class="doc_code">
1539 def ADDRrr : ComplexPattern<i32, 2, "SelectADDRrr", [], []>;
1544 The definition of <tt>ADDRrr</tt> refers to <tt>SelectADDRrr</tt>, which is a
1545 function defined in an implementation of the Instructor Selector (such
1546 as <tt>SparcISelDAGToDAG.cpp</tt>).
1550 In <tt>lib/Target/TargetSelectionDAG.td</tt>, the DAG operator for store is
1554 <div class="doc_code">
1556 def store : PatFrag<(ops node:$val, node:$ptr),
1557 (st node:$val, node:$ptr), [{
1558 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
1559 return !ST->isTruncatingStore() &&
1560 ST->getAddressingMode() == ISD::UNINDEXED;
1567 <tt>XXXInstrInfo.td</tt> also generates (in <tt>XXXGenDAGISel.inc</tt>) the
1568 <tt>SelectCode</tt> method that is used to call the appropriate processing
1569 method for an instruction. In this example, <tt>SelectCode</tt>
1570 calls <tt>Select_ISD_STORE</tt> for the <tt>ISD::STORE</tt> opcode.
1573 <div class="doc_code">
1575 SDNode *SelectCode(SDValue N) {
1577 MVT::ValueType NVT = N.getNode()->getValueType(0);
1578 switch (N.getOpcode()) {
1582 return Select_ISD_STORE(N);
1592 The pattern for <tt>STrr</tt> is matched, so elsewhere in
1593 <tt>XXXGenDAGISel.inc</tt>, code for <tt>STrr</tt> is created for
1594 <tt>Select_ISD_STORE</tt>. The <tt>Emit_22</tt> method is also generated
1595 in <tt>XXXGenDAGISel.inc</tt> to complete the processing of this
1599 <div class="doc_code">
1601 SDNode *Select_ISD_STORE(const SDValue &N) {
1602 SDValue Chain = N.getOperand(0);
1603 if (Predicate_store(N.getNode())) {
1604 SDValue N1 = N.getOperand(1);
1605 SDValue N2 = N.getOperand(2);
1609 // Pattern: (st:void IntRegs:i32:$src,
1610 // ADDRrr:i32:$addr)<<P:Predicate_store>>
1611 // Emits: (STrr:void ADDRrr:i32:$addr, IntRegs:i32:$src)
1612 // Pattern complexity = 13 cost = 1 size = 0
1613 if (SelectADDRrr(N, N2, CPTmp0, CPTmp1) &&
1614 N1.getNode()->getValueType(0) == MVT::i32 &&
1615 N2.getNode()->getValueType(0) == MVT::i32) {
1616 return Emit_22(N, SP::STrr, CPTmp0, CPTmp1);
1622 <!-- ======================================================================= -->
1624 <a name="LegalizePhase">The SelectionDAG Legalize Phase</a>
1630 The Legalize phase converts a DAG to use types and operations that are natively
1631 supported by the target. For natively unsupported types and operations, you need
1632 to add code to the target-specific XXXTargetLowering implementation to convert
1633 unsupported types and operations to supported ones.
1637 In the constructor for the <tt>XXXTargetLowering</tt> class, first use the
1638 <tt>addRegisterClass</tt> method to specify which types are supports and which
1639 register classes are associated with them. The code for the register classes are
1640 generated by TableGen from <tt>XXXRegisterInfo.td</tt> and placed
1641 in <tt>XXXGenRegisterInfo.h.inc</tt>. For example, the implementation of the
1642 constructor for the SparcTargetLowering class (in
1643 <tt>SparcISelLowering.cpp</tt>) starts with the following code:
1646 <div class="doc_code">
1648 addRegisterClass(MVT::i32, SP::IntRegsRegisterClass);
1649 addRegisterClass(MVT::f32, SP::FPRegsRegisterClass);
1650 addRegisterClass(MVT::f64, SP::DFPRegsRegisterClass);
1655 You should examine the node types in the <tt>ISD</tt> namespace
1656 (<tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt>) and determine which
1657 operations the target natively supports. For operations that do <b>not</b> have
1658 native support, add a callback to the constructor for the XXXTargetLowering
1659 class, so the instruction selection process knows what to do. The TargetLowering
1660 class callback methods (declared in <tt>llvm/Target/TargetLowering.h</tt>) are:
1664 <li><tt>setOperationAction</tt> — General operation.</li>
1666 <li><tt>setLoadExtAction</tt> — Load with extension.</li>
1668 <li><tt>setTruncStoreAction</tt> — Truncating store.</li>
1670 <li><tt>setIndexedLoadAction</tt> — Indexed load.</li>
1672 <li><tt>setIndexedStoreAction</tt> — Indexed store.</li>
1674 <li><tt>setConvertAction</tt> — Type conversion.</li>
1676 <li><tt>setCondCodeAction</tt> — Support for a given condition code.</li>
1680 Note: on older releases, <tt>setLoadXAction</tt> is used instead
1681 of <tt>setLoadExtAction</tt>. Also, on older releases,
1682 <tt>setCondCodeAction</tt> may not be supported. Examine your release
1683 to see what methods are specifically supported.
1687 These callbacks are used to determine that an operation does or does not work
1688 with a specified type (or types). And in all cases, the third parameter is
1689 a <tt>LegalAction</tt> type enum value: <tt>Promote</tt>, <tt>Expand</tt>,
1690 <tt>Custom</tt>, or <tt>Legal</tt>. <tt>SparcISelLowering.cpp</tt>
1691 contains examples of all four <tt>LegalAction</tt> values.
1694 <!-- _______________________________________________________________________ -->
1696 <a name="promote">Promote</a>
1702 For an operation without native support for a given type, the specified type may
1703 be promoted to a larger type that is supported. For example, SPARC does not
1704 support a sign-extending load for Boolean values (<tt>i1</tt> type), so
1705 in <tt>SparcISelLowering.cpp</tt> the third parameter below, <tt>Promote</tt>,
1706 changes <tt>i1</tt> type values to a large type before loading.
1709 <div class="doc_code">
1711 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
1717 <!-- _______________________________________________________________________ -->
1719 <a name="expand">Expand</a>
1725 For a type without native support, a value may need to be broken down further,
1726 rather than promoted. For an operation without native support, a combination of
1727 other operations may be used to similar effect. In SPARC, the floating-point
1728 sine and cosine trig operations are supported by expansion to other operations,
1729 as indicated by the third parameter, <tt>Expand</tt>, to
1730 <tt>setOperationAction</tt>:
1733 <div class="doc_code">
1735 setOperationAction(ISD::FSIN, MVT::f32, Expand);
1736 setOperationAction(ISD::FCOS, MVT::f32, Expand);
1742 <!-- _______________________________________________________________________ -->
1744 <a name="custom">Custom</a>
1750 For some operations, simple type promotion or operation expansion may be
1751 insufficient. In some cases, a special intrinsic function must be implemented.
1755 For example, a constant value may require special treatment, or an operation may
1756 require spilling and restoring registers in the stack and working with register
1761 As seen in <tt>SparcISelLowering.cpp</tt> code below, to perform a type
1762 conversion from a floating point value to a signed integer, first the
1763 <tt>setOperationAction</tt> should be called with <tt>Custom</tt> as the third
1767 <div class="doc_code">
1769 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
1774 In the <tt>LowerOperation</tt> method, for each <tt>Custom</tt> operation, a
1775 case statement should be added to indicate what function to call. In the
1776 following code, an <tt>FP_TO_SINT</tt> opcode will call
1777 the <tt>LowerFP_TO_SINT</tt> method:
1780 <div class="doc_code">
1782 SDValue SparcTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
1783 switch (Op.getOpcode()) {
1784 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
1792 Finally, the <tt>LowerFP_TO_SINT</tt> method is implemented, using an FP
1793 register to convert the floating-point value to an integer.
1796 <div class="doc_code">
1798 static SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
1799 assert(Op.getValueType() == MVT::i32);
1800 Op = DAG.getNode(SPISD::FTOI, MVT::f32, Op.getOperand(0));
1801 return DAG.getNode(ISD::BITCAST, MVT::i32, Op);
1808 <!-- _______________________________________________________________________ -->
1810 <a name="legal">Legal</a>
1816 The <tt>Legal</tt> LegalizeAction enum value simply indicates that an
1817 operation <b>is</b> natively supported. <tt>Legal</tt> represents the default
1818 condition, so it is rarely used. In <tt>SparcISelLowering.cpp</tt>, the action
1819 for <tt>CTPOP</tt> (an operation to count the bits set in an integer) is
1820 natively supported only for SPARC v9. The following code enables
1821 the <tt>Expand</tt> conversion technique for non-v9 SPARC implementations.
1824 <div class="doc_code">
1826 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
1828 if (TM.getSubtarget<SparcSubtarget>().isV9())
1829 setOperationAction(ISD::CTPOP, MVT::i32, Legal);
1830 case ISD::SETULT: return SPCC::ICC_CS;
1831 case ISD::SETULE: return SPCC::ICC_LEU;
1832 case ISD::SETUGT: return SPCC::ICC_GU;
1833 case ISD::SETUGE: return SPCC::ICC_CC;
1843 <!-- ======================================================================= -->
1845 <a name="callingConventions">Calling Conventions</a>
1851 To support target-specific calling conventions, <tt>XXXGenCallingConv.td</tt>
1852 uses interfaces (such as CCIfType and CCAssignToReg) that are defined in
1853 <tt>lib/Target/TargetCallingConv.td</tt>. TableGen can take the target
1854 descriptor file <tt>XXXGenCallingConv.td</tt> and generate the header
1855 file <tt>XXXGenCallingConv.inc</tt>, which is typically included
1856 in <tt>XXXISelLowering.cpp</tt>. You can use the interfaces in
1857 <tt>TargetCallingConv.td</tt> to specify:
1861 <li>The order of parameter allocation.</li>
1863 <li>Where parameters and return values are placed (that is, on the stack or in
1866 <li>Which registers may be used.</li>
1868 <li>Whether the caller or callee unwinds the stack.</li>
1872 The following example demonstrates the use of the <tt>CCIfType</tt> and
1873 <tt>CCAssignToReg</tt> interfaces. If the <tt>CCIfType</tt> predicate is true
1874 (that is, if the current argument is of type <tt>f32</tt> or <tt>f64</tt>), then
1875 the action is performed. In this case, the <tt>CCAssignToReg</tt> action assigns
1876 the argument value to the first available register: either <tt>R0</tt>
1880 <div class="doc_code">
1882 CCIfType<[f32,f64], CCAssignToReg<[R0, R1]>>
1887 <tt>SparcCallingConv.td</tt> contains definitions for a target-specific
1888 return-value calling convention (RetCC_Sparc32) and a basic 32-bit C calling
1889 convention (<tt>CC_Sparc32</tt>). The definition of <tt>RetCC_Sparc32</tt>
1890 (shown below) indicates which registers are used for specified scalar return
1891 types. A single-precision float is returned to register <tt>F0</tt>, and a
1892 double-precision float goes to register <tt>D0</tt>. A 32-bit integer is
1893 returned in register <tt>I0</tt> or <tt>I1</tt>.
1896 <div class="doc_code">
1898 def RetCC_Sparc32 : CallingConv<[
1899 CCIfType<[i32], CCAssignToReg<[I0, I1]>>,
1900 CCIfType<[f32], CCAssignToReg<[F0]>>,
1901 CCIfType<[f64], CCAssignToReg<[D0]>>
1907 The definition of <tt>CC_Sparc32</tt> in <tt>SparcCallingConv.td</tt> introduces
1908 <tt>CCAssignToStack</tt>, which assigns the value to a stack slot with the
1909 specified size and alignment. In the example below, the first parameter, 4,
1910 indicates the size of the slot, and the second parameter, also 4, indicates the
1911 stack alignment along 4-byte units. (Special cases: if size is zero, then the
1912 ABI size is used; if alignment is zero, then the ABI alignment is used.)
1915 <div class="doc_code">
1917 def CC_Sparc32 : CallingConv<[
1918 // All arguments get passed in integer registers if there is space.
1919 CCIfType<[i32, f32, f64], CCAssignToReg<[I0, I1, I2, I3, I4, I5]>>,
1920 CCAssignToStack<4, 4>
1926 <tt>CCDelegateTo</tt> is another commonly used interface, which tries to find a
1927 specified sub-calling convention, and, if a match is found, it is invoked. In
1928 the following example (in <tt>X86CallingConv.td</tt>), the definition of
1929 <tt>RetCC_X86_32_C</tt> ends with <tt>CCDelegateTo</tt>. After the current value
1930 is assigned to the register <tt>ST0</tt> or <tt>ST1</tt>,
1931 the <tt>RetCC_X86Common</tt> is invoked.
1934 <div class="doc_code">
1936 def RetCC_X86_32_C : CallingConv<[
1937 CCIfType<[f32], CCAssignToReg<[ST0, ST1]>>,
1938 CCIfType<[f64], CCAssignToReg<[ST0, ST1]>>,
1939 CCDelegateTo<RetCC_X86Common>
1945 <tt>CCIfCC</tt> is an interface that attempts to match the given name to the
1946 current calling convention. If the name identifies the current calling
1947 convention, then a specified action is invoked. In the following example (in
1948 <tt>X86CallingConv.td</tt>), if the <tt>Fast</tt> calling convention is in use,
1949 then <tt>RetCC_X86_32_Fast</tt> is invoked. If the <tt>SSECall</tt> calling
1950 convention is in use, then <tt>RetCC_X86_32_SSE</tt> is invoked.
1953 <div class="doc_code">
1955 def RetCC_X86_32 : CallingConv<[
1956 CCIfCC<"CallingConv::Fast", CCDelegateTo<RetCC_X86_32_Fast>>,
1957 CCIfCC<"CallingConv::X86_SSECall", CCDelegateTo<RetCC_X86_32_SSE>>,
1958 CCDelegateTo<RetCC_X86_32_C>
1963 <p>Other calling convention interfaces include:</p>
1966 <li><tt>CCIf <predicate, action></tt> — If the predicate matches,
1967 apply the action.</li>
1969 <li><tt>CCIfInReg <action></tt> — If the argument is marked with the
1970 '<tt>inreg</tt>' attribute, then apply the action.</li>
1972 <li><tt>CCIfNest <action></tt> — Inf the argument is marked with the
1973 '<tt>nest</tt>' attribute, then apply the action.</li>
1975 <li><tt>CCIfNotVarArg <action></tt> — If the current function does
1976 not take a variable number of arguments, apply the action.</li>
1978 <li><tt>CCAssignToRegWithShadow <registerList, shadowList></tt> —
1979 similar to <tt>CCAssignToReg</tt>, but with a shadow list of registers.</li>
1981 <li><tt>CCPassByVal <size, align></tt> — Assign value to a stack
1982 slot with the minimum specified size and alignment.</li>
1984 <li><tt>CCPromoteToType <type></tt> — Promote the current value to
1985 the specified type.</li>
1987 <li><tt>CallingConv <[actions]></tt> — Define each calling
1988 convention that is supported.</li>
1995 <!-- *********************************************************************** -->
1997 <a name="assemblyPrinter">Assembly Printer</a>
1999 <!-- *********************************************************************** -->
2004 During the code emission stage, the code generator may utilize an LLVM pass to
2005 produce assembly output. To do this, you want to implement the code for a
2006 printer that converts LLVM IR to a GAS-format assembly language for your target
2007 machine, using the following steps:
2011 <li>Define all the assembly strings for your target, adding them to the
2012 instructions defined in the <tt>XXXInstrInfo.td</tt> file.
2013 (See <a href="#InstructionSet">Instruction Set</a>.) TableGen will produce
2014 an output file (<tt>XXXGenAsmWriter.inc</tt>) with an implementation of
2015 the <tt>printInstruction</tt> method for the XXXAsmPrinter class.</li>
2017 <li>Write <tt>XXXTargetAsmInfo.h</tt>, which contains the bare-bones declaration
2018 of the <tt>XXXTargetAsmInfo</tt> class (a subclass
2019 of <tt>TargetAsmInfo</tt>).</li>
2021 <li>Write <tt>XXXTargetAsmInfo.cpp</tt>, which contains target-specific values
2022 for <tt>TargetAsmInfo</tt> properties and sometimes new implementations for
2025 <li>Write <tt>XXXAsmPrinter.cpp</tt>, which implements the <tt>AsmPrinter</tt>
2026 class that performs the LLVM-to-assembly conversion.</li>
2030 The code in <tt>XXXTargetAsmInfo.h</tt> is usually a trivial declaration of the
2031 <tt>XXXTargetAsmInfo</tt> class for use in <tt>XXXTargetAsmInfo.cpp</tt>.
2032 Similarly, <tt>XXXTargetAsmInfo.cpp</tt> usually has a few declarations of
2033 <tt>XXXTargetAsmInfo</tt> replacement values that override the default values
2034 in <tt>TargetAsmInfo.cpp</tt>. For example in <tt>SparcTargetAsmInfo.cpp</tt>:
2037 <div class="doc_code">
2039 SparcTargetAsmInfo::SparcTargetAsmInfo(const SparcTargetMachine &TM) {
2040 Data16bitsDirective = "\t.half\t";
2041 Data32bitsDirective = "\t.word\t";
2042 Data64bitsDirective = 0; // .xword is only supported by V9.
2043 ZeroDirective = "\t.skip\t";
2044 CommentString = "!";
2045 ConstantPoolSection = "\t.section \".rodata\",#alloc\n";
2051 The X86 assembly printer implementation (<tt>X86TargetAsmInfo</tt>) is an
2052 example where the target specific <tt>TargetAsmInfo</tt> class uses an
2053 overridden methods: <tt>ExpandInlineAsm</tt>.
2057 A target-specific implementation of AsmPrinter is written in
2058 <tt>XXXAsmPrinter.cpp</tt>, which implements the <tt>AsmPrinter</tt> class that
2059 converts the LLVM to printable assembly. The implementation must include the
2060 following headers that have declarations for the <tt>AsmPrinter</tt> and
2061 <tt>MachineFunctionPass</tt> classes. The <tt>MachineFunctionPass</tt> is a
2062 subclass of <tt>FunctionPass</tt>.
2065 <div class="doc_code">
2067 #include "llvm/CodeGen/AsmPrinter.h"
2068 #include "llvm/CodeGen/MachineFunctionPass.h"
2073 As a <tt>FunctionPass</tt>, <tt>AsmPrinter</tt> first
2074 calls <tt>doInitialization</tt> to set up the <tt>AsmPrinter</tt>. In
2075 <tt>SparcAsmPrinter</tt>, a <tt>Mangler</tt> object is instantiated to process
2080 In <tt>XXXAsmPrinter.cpp</tt>, the <tt>runOnMachineFunction</tt> method
2081 (declared in <tt>MachineFunctionPass</tt>) must be implemented
2082 for <tt>XXXAsmPrinter</tt>. In <tt>MachineFunctionPass</tt>,
2083 the <tt>runOnFunction</tt> method invokes <tt>runOnMachineFunction</tt>.
2084 Target-specific implementations of <tt>runOnMachineFunction</tt> differ, but
2085 generally do the following to process each machine function:
2089 <li>Call <tt>SetupMachineFunction</tt> to perform initialization.</li>
2091 <li>Call <tt>EmitConstantPool</tt> to print out (to the output stream) constants
2092 which have been spilled to memory.</li>
2094 <li>Call <tt>EmitJumpTableInfo</tt> to print out jump tables used by the current
2097 <li>Print out the label for the current function.</li>
2099 <li>Print out the code for the function, including basic block labels and the
2100 assembly for the instruction (using <tt>printInstruction</tt>)</li>
2104 The <tt>XXXAsmPrinter</tt> implementation must also include the code generated
2105 by TableGen that is output in the <tt>XXXGenAsmWriter.inc</tt> file. The code
2106 in <tt>XXXGenAsmWriter.inc</tt> contains an implementation of the
2107 <tt>printInstruction</tt> method that may call these methods:
2111 <li><tt>printOperand</tt></li>
2113 <li><tt>printMemOperand</tt></li>
2115 <li><tt>printCCOperand (for conditional statements)</tt></li>
2117 <li><tt>printDataDirective</tt></li>
2119 <li><tt>printDeclare</tt></li>
2121 <li><tt>printImplicitDef</tt></li>
2123 <li><tt>printInlineAsm</tt></li>
2127 The implementations of <tt>printDeclare</tt>, <tt>printImplicitDef</tt>,
2128 <tt>printInlineAsm</tt>, and <tt>printLabel</tt> in <tt>AsmPrinter.cpp</tt> are
2129 generally adequate for printing assembly and do not need to be
2134 The <tt>printOperand</tt> method is implemented with a long switch/case
2135 statement for the type of operand: register, immediate, basic block, external
2136 symbol, global address, constant pool index, or jump table index. For an
2137 instruction with a memory address operand, the <tt>printMemOperand</tt> method
2138 should be implemented to generate the proper output. Similarly,
2139 <tt>printCCOperand</tt> should be used to print a conditional operand.
2142 <p><tt>doFinalization</tt> should be overridden in <tt>XXXAsmPrinter</tt>, and
2143 it should be called to shut down the assembly printer. During
2144 <tt>doFinalization</tt>, global variables and constants are printed to
2150 <!-- *********************************************************************** -->
2152 <a name="subtargetSupport">Subtarget Support</a>
2154 <!-- *********************************************************************** -->
2159 Subtarget support is used to inform the code generation process of instruction
2160 set variations for a given chip set. For example, the LLVM SPARC implementation
2161 provided covers three major versions of the SPARC microprocessor architecture:
2162 Version 8 (V8, which is a 32-bit architecture), Version 9 (V9, a 64-bit
2163 architecture), and the UltraSPARC architecture. V8 has 16 double-precision
2164 floating-point registers that are also usable as either 32 single-precision or 8
2165 quad-precision registers. V8 is also purely big-endian. V9 has 32
2166 double-precision floating-point registers that are also usable as 16
2167 quad-precision registers, but cannot be used as single-precision registers. The
2168 UltraSPARC architecture combines V9 with UltraSPARC Visual Instruction Set
2173 If subtarget support is needed, you should implement a target-specific
2174 XXXSubtarget class for your architecture. This class should process the
2175 command-line options <tt>-mcpu=</tt> and <tt>-mattr=</tt>.
2179 TableGen uses definitions in the <tt>Target.td</tt> and <tt>Sparc.td</tt> files
2180 to generate code in <tt>SparcGenSubtarget.inc</tt>. In <tt>Target.td</tt>, shown
2181 below, the <tt>SubtargetFeature</tt> interface is defined. The first 4 string
2182 parameters of the <tt>SubtargetFeature</tt> interface are a feature name, an
2183 attribute set by the feature, the value of the attribute, and a description of
2184 the feature. (The fifth parameter is a list of features whose presence is
2185 implied, and its default value is an empty array.)
2188 <div class="doc_code">
2190 class SubtargetFeature<string n, string a, string v, string d,
2191 list<SubtargetFeature> i = []> {
2193 string Attribute = a;
2196 list<SubtargetFeature> Implies = i;
2202 In the <tt>Sparc.td</tt> file, the SubtargetFeature is used to define the
2206 <div class="doc_code">
2208 def FeatureV9 : SubtargetFeature<"v9", "IsV9", "true",
2209 "Enable SPARC-V9 instructions">;
2210 def FeatureV8Deprecated : SubtargetFeature<"deprecated-v8",
2211 "V8DeprecatedInsts", "true",
2212 "Enable deprecated V8 instructions in V9 mode">;
2213 def FeatureVIS : SubtargetFeature<"vis", "IsVIS", "true",
2214 "Enable UltraSPARC Visual Instruction Set extensions">;
2219 Elsewhere in <tt>Sparc.td</tt>, the Proc class is defined and then is used to
2220 define particular SPARC processor subtypes that may have the previously
2224 <div class="doc_code">
2226 class Proc<string Name, list<SubtargetFeature> Features>
2227 : Processor<Name, NoItineraries, Features>;
2229 def : Proc<"generic", []>;
2230 def : Proc<"v8", []>;
2231 def : Proc<"supersparc", []>;
2232 def : Proc<"sparclite", []>;
2233 def : Proc<"f934", []>;
2234 def : Proc<"hypersparc", []>;
2235 def : Proc<"sparclite86x", []>;
2236 def : Proc<"sparclet", []>;
2237 def : Proc<"tsc701", []>;
2238 def : Proc<"v9", [FeatureV9]>;
2239 def : Proc<"ultrasparc", [FeatureV9, FeatureV8Deprecated]>;
2240 def : Proc<"ultrasparc3", [FeatureV9, FeatureV8Deprecated]>;
2241 def : Proc<"ultrasparc3-vis", [FeatureV9, FeatureV8Deprecated, FeatureVIS]>;
2246 From <tt>Target.td</tt> and <tt>Sparc.td</tt> files, the resulting
2247 SparcGenSubtarget.inc specifies enum values to identify the features, arrays of
2248 constants to represent the CPU features and CPU subtypes, and the
2249 ParseSubtargetFeatures method that parses the features string that sets
2250 specified subtarget options. The generated <tt>SparcGenSubtarget.inc</tt> file
2251 should be included in the <tt>SparcSubtarget.cpp</tt>. The target-specific
2252 implementation of the XXXSubtarget method should follow this pseudocode:
2255 <div class="doc_code">
2257 XXXSubtarget::XXXSubtarget(const Module &M, const std::string &FS) {
2258 // Set the default features
2259 // Determine default and user specified characteristics of the CPU
2260 // Call ParseSubtargetFeatures(FS, CPU) to parse the features string
2261 // Perform any additional operations
2268 <!-- *********************************************************************** -->
2270 <a name="jitSupport">JIT Support</a>
2272 <!-- *********************************************************************** -->
2277 The implementation of a target machine optionally includes a Just-In-Time (JIT)
2278 code generator that emits machine code and auxiliary structures as binary output
2279 that can be written directly to memory. To do this, implement JIT code
2280 generation by performing the following steps:
2284 <li>Write an <tt>XXXCodeEmitter.cpp</tt> file that contains a machine function
2285 pass that transforms target-machine instructions into relocatable machine
2288 <li>Write an <tt>XXXJITInfo.cpp</tt> file that implements the JIT interfaces for
2289 target-specific code-generation activities, such as emitting machine code
2292 <li>Modify <tt>XXXTargetMachine</tt> so that it provides a
2293 <tt>TargetJITInfo</tt> object through its <tt>getJITInfo</tt> method.</li>
2297 There are several different approaches to writing the JIT support code. For
2298 instance, TableGen and target descriptor files may be used for creating a JIT
2299 code generator, but are not mandatory. For the Alpha and PowerPC target
2300 machines, TableGen is used to generate <tt>XXXGenCodeEmitter.inc</tt>, which
2301 contains the binary coding of machine instructions and the
2302 <tt>getBinaryCodeForInstr</tt> method to access those codes. Other JIT
2303 implementations do not.
2307 Both <tt>XXXJITInfo.cpp</tt> and <tt>XXXCodeEmitter.cpp</tt> must include the
2308 <tt>llvm/CodeGen/MachineCodeEmitter.h</tt> header file that defines the
2309 <tt>MachineCodeEmitter</tt> class containing code for several callback functions
2310 that write data (in bytes, words, strings, etc.) to the output stream.
2313 <!-- ======================================================================= -->
2315 <a name="mce">Machine Code Emitter</a>
2321 In <tt>XXXCodeEmitter.cpp</tt>, a target-specific of the <tt>Emitter</tt> class
2322 is implemented as a function pass (subclass
2323 of <tt>MachineFunctionPass</tt>). The target-specific implementation
2324 of <tt>runOnMachineFunction</tt> (invoked by
2325 <tt>runOnFunction</tt> in <tt>MachineFunctionPass</tt>) iterates through the
2326 <tt>MachineBasicBlock</tt> calls <tt>emitInstruction</tt> to process each
2327 instruction and emit binary code. <tt>emitInstruction</tt> is largely
2328 implemented with case statements on the instruction types defined in
2329 <tt>XXXInstrInfo.h</tt>. For example, in <tt>X86CodeEmitter.cpp</tt>,
2330 the <tt>emitInstruction</tt> method is built around the following switch/case
2334 <div class="doc_code">
2336 switch (Desc->TSFlags & X86::FormMask) {
2337 case X86II::Pseudo: // for not yet implemented instructions
2338 ... // or pseudo-instructions
2340 case X86II::RawFrm: // for instructions with a fixed opcode value
2343 case X86II::AddRegFrm: // for instructions that have one register operand
2344 ... // added to their opcode
2346 case X86II::MRMDestReg:// for instructions that use the Mod/RM byte
2347 ... // to specify a destination (register)
2349 case X86II::MRMDestMem:// for instructions that use the Mod/RM byte
2350 ... // to specify a destination (memory)
2352 case X86II::MRMSrcReg: // for instructions that use the Mod/RM byte
2353 ... // to specify a source (register)
2355 case X86II::MRMSrcMem: // for instructions that use the Mod/RM byte
2356 ... // to specify a source (memory)
2358 case X86II::MRM0r: case X86II::MRM1r: // for instructions that operate on
2359 case X86II::MRM2r: case X86II::MRM3r: // a REGISTER r/m operand and
2360 case X86II::MRM4r: case X86II::MRM5r: // use the Mod/RM byte and a field
2361 case X86II::MRM6r: case X86II::MRM7r: // to hold extended opcode data
2364 case X86II::MRM0m: case X86II::MRM1m: // for instructions that operate on
2365 case X86II::MRM2m: case X86II::MRM3m: // a MEMORY r/m operand and
2366 case X86II::MRM4m: case X86II::MRM5m: // use the Mod/RM byte and a field
2367 case X86II::MRM6m: case X86II::MRM7m: // to hold extended opcode data
2370 case X86II::MRMInitReg: // for instructions whose source and
2371 ... // destination are the same register
2378 The implementations of these case statements often first emit the opcode and
2379 then get the operand(s). Then depending upon the operand, helper methods may be
2380 called to process the operand(s). For example, in <tt>X86CodeEmitter.cpp</tt>,
2381 for the <tt>X86II::AddRegFrm</tt> case, the first data emitted
2382 (by <tt>emitByte</tt>) is the opcode added to the register operand. Then an
2383 object representing the machine operand, <tt>MO1</tt>, is extracted. The helper
2384 methods such as <tt>isImmediate</tt>,
2385 <tt>isGlobalAddress</tt>, <tt>isExternalSymbol</tt>, <tt>isConstantPoolIndex</tt>, and
2386 <tt>isJumpTableIndex</tt> determine the operand
2387 type. (<tt>X86CodeEmitter.cpp</tt> also has private methods such
2388 as <tt>emitConstant</tt>, <tt>emitGlobalAddress</tt>,
2389 <tt>emitExternalSymbolAddress</tt>, <tt>emitConstPoolAddress</tt>,
2390 and <tt>emitJumpTableAddress</tt> that emit the data into the output stream.)
2393 <div class="doc_code">
2395 case X86II::AddRegFrm:
2396 MCE.emitByte(BaseOpcode + getX86RegNum(MI.getOperand(CurOp++).getReg()));
2398 if (CurOp != NumOps) {
2399 const MachineOperand &MO1 = MI.getOperand(CurOp++);
2400 unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2401 if (MO1.isImmediate())
2402 emitConstant(MO1.getImm(), Size);
2404 unsigned rt = Is64BitMode ? X86::reloc_pcrel_word
2405 : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word);
2406 if (Opcode == X86::MOV64ri)
2407 rt = X86::reloc_absolute_dword; // FIXME: add X86II flag?
2408 if (MO1.isGlobalAddress()) {
2409 bool NeedStub = isa<Function>(MO1.getGlobal());
2410 bool isLazy = gvNeedsLazyPtr(MO1.getGlobal());
2411 emitGlobalAddress(MO1.getGlobal(), rt, MO1.getOffset(), 0,
2413 } else if (MO1.isExternalSymbol())
2414 emitExternalSymbolAddress(MO1.getSymbolName(), rt);
2415 else if (MO1.isConstantPoolIndex())
2416 emitConstPoolAddress(MO1.getIndex(), rt);
2417 else if (MO1.isJumpTableIndex())
2418 emitJumpTableAddress(MO1.getIndex(), rt);
2426 In the previous example, <tt>XXXCodeEmitter.cpp</tt> uses the
2427 variable <tt>rt</tt>, which is a RelocationType enum that may be used to
2428 relocate addresses (for example, a global address with a PIC base offset). The
2429 <tt>RelocationType</tt> enum for that target is defined in the short
2430 target-specific <tt>XXXRelocations.h</tt> file. The <tt>RelocationType</tt> is used by
2431 the <tt>relocate</tt> method defined in <tt>XXXJITInfo.cpp</tt> to rewrite
2432 addresses for referenced global symbols.
2436 For example, <tt>X86Relocations.h</tt> specifies the following relocation types
2437 for the X86 addresses. In all four cases, the relocated value is added to the
2438 value already in memory. For <tt>reloc_pcrel_word</tt>
2439 and <tt>reloc_picrel_word</tt>, there is an additional initial adjustment.
2442 <div class="doc_code">
2444 enum RelocationType {
2445 reloc_pcrel_word = 0, // add reloc value after adjusting for the PC loc
2446 reloc_picrel_word = 1, // add reloc value after adjusting for the PIC base
2447 reloc_absolute_word = 2, // absolute relocation; no additional adjustment
2448 reloc_absolute_dword = 3 // absolute relocation; no additional adjustment
2455 <!-- ======================================================================= -->
2457 <a name="targetJITInfo">Target JIT Info</a>
2463 <tt>XXXJITInfo.cpp</tt> implements the JIT interfaces for target-specific
2464 code-generation activities, such as emitting machine code and stubs. At minimum,
2465 a target-specific version of <tt>XXXJITInfo</tt> implements the following:
2469 <li><tt>getLazyResolverFunction</tt> — Initializes the JIT, gives the
2470 target a function that is used for compilation.</li>
2472 <li><tt>emitFunctionStub</tt> — Returns a native function with a specified
2473 address for a callback function.</li>
2475 <li><tt>relocate</tt> — Changes the addresses of referenced globals, based
2476 on relocation types.</li>
2478 <li>Callback function that are wrappers to a function stub that is used when the
2479 real target is not initially known.</li>
2483 <tt>getLazyResolverFunction</tt> is generally trivial to implement. It makes the
2484 incoming parameter as the global <tt>JITCompilerFunction</tt> and returns the
2485 callback function that will be used a function wrapper. For the Alpha target
2486 (in <tt>AlphaJITInfo.cpp</tt>), the <tt>getLazyResolverFunction</tt>
2487 implementation is simply:
2490 <div class="doc_code">
2492 TargetJITInfo::LazyResolverFn AlphaJITInfo::getLazyResolverFunction(
2494 JITCompilerFunction = F;
2495 return AlphaCompilationCallback;
2501 For the X86 target, the <tt>getLazyResolverFunction</tt> implementation is a
2502 little more complication, because it returns a different callback function for
2503 processors with SSE instructions and XMM registers.
2507 The callback function initially saves and later restores the callee register
2508 values, incoming arguments, and frame and return address. The callback function
2509 needs low-level access to the registers or stack, so it is typically implemented
2517 <!-- *********************************************************************** -->
2521 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
2522 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
2523 <a href="http://validator.w3.org/check/referer"><img
2524 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
2526 <a href="http://www.woo.com">Mason Woo</a> and <a href="http://misha.brukman.net">Misha Brukman</a><br>
2527 <a href="http://llvm.org/">The LLVM Compiler Infrastructure</a>
2529 Last modified: $Date$