The same problem was being tracked in PR7652.
[oota-llvm.git] / docs / WritingAnLLVMBackend.html
1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2                       "http://www.w3.org/TR/html4/strict.dtd">
3 <html>
4 <head>
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="llvm.css" type="text/css">
8 </head>
9
10 <body>
11
12 <div class="doc_title">
13   Writing an LLVM Compiler Backend
14 </div>
15
16 <ol>
17   <li><a href="#intro">Introduction</a>
18   <ul>
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>
23   </ul>
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>
27   <ul>
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>
31   </ul></li>
32   <li><a href="#InstructionSet">Instruction Set</a>
33   <ul>  
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>
37   </ul></li>
38   <li><a href="#InstructionSelector">Instruction Selector</a>
39   <ul>
40     <li><a href="#LegalizePhase">The SelectionDAG Legalize Phase</a>
41     <ul>
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>       
46     </ul></li>
47     <li><a href="#callingConventions">Calling Conventions</a></li>     
48   </ul></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>
52   <ul>  
53     <li><a href="#mce">Machine Code Emitter</a></li>   
54     <li><a href="#targetJITInfo">Target JIT Info</a></li>   
55   </ul></li>
56 </ol>
57
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>
61 </div>
62
63 <!-- *********************************************************************** -->
64 <div class="doc_section">
65   <a name="intro">Introduction</a>
66 </div>
67 <!-- *********************************************************************** -->
68
69 <div class="doc_text">
70
71 <p>
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).
76 </p>
77
78 <p>
79 The backend of LLVM features a target-independent code generator that may create
80 output for several types of target CPUs &mdash; including X86, PowerPC, Alpha,
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.
83 </p>
84
85 <p>
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
91 conventions.
92 </p>
93
94 </div>
95
96 <div class="doc_subsection">
97   <a name="Audience">Audience</a>
98 </div>  
99
100 <div class="doc_text">
101
102 <p>
103 The audience for this document is anyone who needs to write an LLVM backend to
104 generate code for a specific hardware or software target.
105 </p>
106
107 </div>
108
109 <div class="doc_subsection">
110   <a name="Prerequisite">Prerequisite Reading</a>
111 </div>  
112
113 <div class="doc_text">  
114
115 <p>
116 These essential documents must be read before reading this document:
117 </p>
118
119 <ul>
120 <li><i><a href="http://www.llvm.org/docs/LangRef.html">LLVM Language Reference
121     Manual</a></i> &mdash; a reference manual for the LLVM assembly language.</li>
122
123 <li><i><a href="http://www.llvm.org/docs/CodeGenerator.html">The LLVM
124     Target-Independent Code Generator</a></i> &mdash; a guide to the components
125     (classes and code generation algorithms) for translating the LLVM internal
126     representation into machine code for a specified target.  Pay particular
127     attention to the descriptions of code generation stages: Instruction
128     Selection, Scheduling and Formation, SSA-based Optimization, Register
129     Allocation, Prolog/Epilog Code Insertion, Late Machine Code Optimizations,
130     and Code Emission.</li>
131
132 <li><i><a href="http://www.llvm.org/docs/TableGenFundamentals.html">TableGen
133     Fundamentals</a></i> &mdash;a document that describes the TableGen
134     (<tt>tblgen</tt>) application that manages domain-specific information to
135     support LLVM code generation. TableGen processes input from a target
136     description file (<tt>.td</tt> suffix) and generates C++ code that can be
137     used for code generation.</li>
138
139 <li><i><a href="http://www.llvm.org/docs/WritingAnLLVMPass.html">Writing an LLVM
140     Pass</a></i> &mdash; The assembly printer is a <tt>FunctionPass</tt>, as are
141     several SelectionDAG processing steps.</li>
142 </ul>
143
144 <p>
145 To follow the SPARC examples in this document, have a copy of
146 <i><a href="http://www.sparc.org/standards/V8.pdf">The SPARC Architecture
147 Manual, Version 8</a></i> for reference. For details about the ARM instruction
148 set, refer to the <i><a href="http://infocenter.arm.com/">ARM Architecture
149 Reference Manual</a></i>. For more about the GNU Assembler format
150 (<tt>GAS</tt>), see
151 <i><a href="http://sourceware.org/binutils/docs/as/index.html">Using As</a></i>,
152 especially for the assembly printer. <i>Using As</i> contains a list of target
153 machine dependent features.
154 </p>
155
156 </div>
157
158 <div class="doc_subsection">
159   <a name="Basic">Basic Steps</a>
160 </div>
161
162 <div class="doc_text">
163
164 <p>
165 To write a compiler backend for LLVM that converts the LLVM IR to code for a
166 specified target (machine or other language), follow these steps:
167 </p>
168
169 <ul>
170 <li>Create a subclass of the TargetMachine class that describes characteristics
171     of your target machine. Copy existing examples of specific TargetMachine
172     class and header files; for example, start with
173     <tt>SparcTargetMachine.cpp</tt> and <tt>SparcTargetMachine.h</tt>, but
174     change the file names for your target. Similarly, change code that
175     references "Sparc" to reference your target. </li>
176
177 <li>Describe the register set of the target. Use TableGen to generate code for
178     register definition, register aliases, and register classes from a
179     target-specific <tt>RegisterInfo.td</tt> input file. You should also write
180     additional code for a subclass of the TargetRegisterInfo class that
181     represents the class register file data used for register allocation and
182     also describes the interactions between registers.</li>
183
184 <li>Describe the instruction set of the target. Use TableGen to generate code
185     for target-specific instructions from target-specific versions of
186     <tt>TargetInstrFormats.td</tt> and <tt>TargetInstrInfo.td</tt>. You should
187     write additional code for a subclass of the TargetInstrInfo class to
188     represent machine instructions supported by the target machine. </li>
189
190 <li>Describe the selection and conversion of the LLVM IR from a Directed Acyclic
191     Graph (DAG) representation of instructions to native target-specific
192     instructions. Use TableGen to generate code that matches patterns and
193     selects instructions based on additional information in a target-specific
194     version of <tt>TargetInstrInfo.td</tt>. Write code
195     for <tt>XXXISelDAGToDAG.cpp</tt>, where XXX identifies the specific target,
196     to perform pattern matching and DAG-to-DAG instruction selection. Also write
197     code in <tt>XXXISelLowering.cpp</tt> to replace or remove operations and
198     data types that are not supported natively in a SelectionDAG. </li>
199
200 <li>Write code for an assembly printer that converts LLVM IR to a GAS format for
201     your target machine.  You should add assembly strings to the instructions
202     defined in your target-specific version of <tt>TargetInstrInfo.td</tt>. You
203     should also write code for a subclass of AsmPrinter that performs the
204     LLVM-to-assembly conversion and a trivial subclass of TargetAsmInfo.</li>
205
206 <li>Optionally, add support for subtargets (i.e., variants with different
207     capabilities). You should also write code for a subclass of the
208     TargetSubtarget class, which allows you to use the <tt>-mcpu=</tt>
209     and <tt>-mattr=</tt> command-line options.</li>
210
211 <li>Optionally, add JIT support and create a machine code emitter (subclass of
212     TargetJITInfo) that is used to emit binary code directly into memory. </li>
213 </ul>
214
215 <p>
216 In the <tt>.cpp</tt> and <tt>.h</tt>. files, initially stub up these methods and
217 then implement them later. Initially, you may not know which private members
218 that the class will need and which components will need to be subclassed.
219 </p>
220
221 </div>
222
223 <div class="doc_subsection">
224   <a name="Preliminaries">Preliminaries</a>
225 </div>
226
227 <div class="doc_text">
228
229 <p>
230 To actually create your compiler backend, you need to create and modify a few
231 files. The absolute minimum is discussed here. But to actually use the LLVM
232 target-independent code generator, you must perform the steps described in
233 the <a href="http://www.llvm.org/docs/CodeGenerator.html">LLVM
234 Target-Independent Code Generator</a> document.
235 </p>
236
237 <p>
238 First, you should create a subdirectory under <tt>lib/Target</tt> to hold all
239 the files related to your target. If your target is called "Dummy," create the
240 directory <tt>lib/Target/Dummy</tt>.
241 </p>
242
243 <p>
244 In this new
245 directory, create a <tt>Makefile</tt>. It is easiest to copy a
246 <tt>Makefile</tt> of another target and modify it. It should at least contain
247 the <tt>LEVEL</tt>, <tt>LIBRARYNAME</tt> and <tt>TARGET</tt> variables, and then
248 include <tt>$(LEVEL)/Makefile.common</tt>. The library can be
249 named <tt>LLVMDummy</tt> (for example, see the MIPS target). Alternatively, you
250 can split the library into <tt>LLVMDummyCodeGen</tt>
251 and <tt>LLVMDummyAsmPrinter</tt>, the latter of which should be implemented in a
252 subdirectory below <tt>lib/Target/Dummy</tt> (for example, see the PowerPC
253 target).
254 </p>
255
256 <p>
257 Note that these two naming schemes are hardcoded into <tt>llvm-config</tt>.
258 Using any other naming scheme will confuse <tt>llvm-config</tt> and produce a
259 lot of (seemingly unrelated) linker errors when linking <tt>llc</tt>.
260 </p>
261
262 <p>
263 To make your target actually do something, you need to implement a subclass of
264 <tt>TargetMachine</tt>. This implementation should typically be in the file
265 <tt>lib/Target/DummyTargetMachine.cpp</tt>, but any file in
266 the <tt>lib/Target</tt> directory will be built and should work. To use LLVM's
267 target independent code generator, you should do what all current machine
268 backends do: create a subclass of <tt>LLVMTargetMachine</tt>. (To create a
269 target from scratch, create a subclass of <tt>TargetMachine</tt>.)
270 </p>
271
272 <p>
273 To get LLVM to actually build and link your target, you need to add it to
274 the <tt>TARGETS_TO_BUILD</tt> variable. To do this, you modify the configure
275 script to know about your target when parsing the <tt>--enable-targets</tt>
276 option. Search the configure script for <tt>TARGETS_TO_BUILD</tt>, add your
277 target to the lists there (some creativity required), and then
278 reconfigure. Alternatively, you can change <tt>autotools/configure.ac</tt> and
279 regenerate configure by running <tt>./autoconf/AutoRegen.sh</tt>.
280 </p>
281
282 </div>
283
284 <!-- *********************************************************************** -->
285 <div class="doc_section">
286   <a name="TargetMachine">Target Machine</a>
287 </div>
288 <!-- *********************************************************************** -->
289
290 <div class="doc_text">
291
292 <p>
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.
300 </p>
301
302 <p>
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>.
308 </p>
309
310 <p>
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.
319 </p>
320
321 <p>
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.
325 </p>
326
327 <div class="doc_code">
328 <pre>
329 namespace llvm {
330
331 class Module;
332
333 class SparcTargetMachine : public LLVMTargetMachine {
334   const TargetData DataLayout;       // Calculates type size &amp; alignment
335   SparcSubtarget Subtarget;
336   SparcInstrInfo InstrInfo;
337   TargetFrameInfo FrameInfo;
338   
339 protected:
340   virtual const TargetAsmInfo *createTargetAsmInfo() const;
341   
342 public:
343   SparcTargetMachine(const Module &amp;M, const std::string &amp;FS);
344
345   virtual const SparcInstrInfo *getInstrInfo() const {return &amp;InstrInfo; }
346   virtual const TargetFrameInfo *getFrameInfo() const {return &amp;FrameInfo; }
347   virtual const TargetSubtarget *getSubtargetImpl() const{return &amp;Subtarget; }
348   virtual const TargetRegisterInfo *getRegisterInfo() const {
349     return &amp;InstrInfo.getRegisterInfo();
350   }
351   virtual const TargetData *getTargetData() const { return &amp;DataLayout; }
352   static unsigned getModuleMatchQuality(const Module &amp;M);
353
354   // Pass Pipeline Configuration
355   virtual bool addInstSelector(PassManagerBase &amp;PM, bool Fast);
356   virtual bool addPreEmitPass(PassManagerBase &amp;PM, bool Fast);
357 };
358
359 } // end namespace llvm
360 </pre>
361 </div>
362
363 </div>
364
365
366 <div class="doc_text">
367
368 <ul>
369 <li><tt>getInstrInfo()</tt></li>
370 <li><tt>getRegisterInfo()</tt></li>
371 <li><tt>getFrameInfo()</tt></li>
372 <li><tt>getTargetData()</tt></li>
373 <li><tt>getSubtargetImpl()</tt></li>
374 </ul>
375
376 <p>For some targets, you also need to support the following methods:</p>
377
378 <ul>
379 <li><tt>getTargetLowering()</tt></li>
380 <li><tt>getJITInfo()</tt></li>
381 </ul>
382
383 <p>
384 In addition, the <tt>XXXTargetMachine</tt> constructor should specify a
385 <tt>TargetDescription</tt> string that determines the data layout for the target
386 machine, including characteristics such as pointer size, alignment, and
387 endianness. For example, the constructor for SparcTargetMachine contains the
388 following:
389 </p>
390
391 <div class="doc_code">
392 <pre>
393 SparcTargetMachine::SparcTargetMachine(const Module &amp;M, const std::string &amp;FS)
394   : DataLayout("E-p:32:32-f128:128:128"),
395     Subtarget(M, FS), InstrInfo(Subtarget),
396     FrameInfo(TargetFrameInfo::StackGrowsDown, 8, 0) {
397 }
398 </pre>
399 </div>
400
401 </div>
402
403 <div class="doc_text">
404
405 <p>Hyphens separate portions of the <tt>TargetDescription</tt> string.</p>
406
407 <ul>
408 <li>An upper-case "<tt>E</tt>" in the string indicates a big-endian target data
409     model. a lower-case "<tt>e</tt>" indicates little-endian.</li>
410
411 <li>"<tt>p:</tt>" is followed by pointer information: size, ABI alignment, and
412     preferred alignment. If only two figures follow "<tt>p:</tt>", then the
413     first value is pointer size, and the second value is both ABI and preferred
414     alignment.</li>
415
416 <li>Then a letter for numeric type alignment: "<tt>i</tt>", "<tt>f</tt>",
417     "<tt>v</tt>", or "<tt>a</tt>" (corresponding to integer, floating point,
418     vector, or aggregate). "<tt>i</tt>", "<tt>v</tt>", or "<tt>a</tt>" are
419     followed by ABI alignment and preferred alignment. "<tt>f</tt>" is followed
420     by three values: the first indicates the size of a long double, then ABI
421     alignment, and then ABI preferred alignment.</li>
422 </ul>
423
424 </div>
425
426 <!-- *********************************************************************** -->
427 <div class="doc_section">
428   <a name="TargetRegistration">Target Registration</a>
429 </div>
430 <!-- *********************************************************************** -->
431
432 <div class="doc_text">
433
434 <p>
435 You must also register your target with the <tt>TargetRegistry</tt>, which is
436 what other LLVM tools use to be able to lookup and use your target at
437 runtime. The <tt>TargetRegistry</tt> can be used directly, but for most targets
438 there are helper templates which should take care of the work for you.</p>
439
440 <p>
441 All targets should declare a global <tt>Target</tt> object which is used to
442 represent the target during registration. Then, in the target's TargetInfo
443 library, the target should define that object and use
444 the <tt>RegisterTarget</tt> template to register the target. For example, the Sparc registration code looks like this:
445 </p>
446
447 <div class="doc_code">
448 <pre>
449 Target llvm::TheSparcTarget;
450
451 extern "C" void LLVMInitializeSparcTargetInfo() { 
452   RegisterTarget&lt;Triple::sparc, /*HasJIT=*/false&gt;
453     X(TheSparcTarget, "sparc", "Sparc");
454 }
455 </pre>
456 </div>
457
458 <p>
459 This allows the <tt>TargetRegistry</tt> to look up the target by name or by
460 target triple. In addition, most targets will also register additional features
461 which are available in separate libraries. These registration steps are
462 separate, because some clients may wish to only link in some parts of the target
463 -- the JIT code generator does not require the use of the assembler printer, for
464 example. Here is an example of registering the Sparc assembly printer:
465 </p>
466
467 <div class="doc_code">
468 <pre>
469 extern "C" void LLVMInitializeSparcAsmPrinter() { 
470   RegisterAsmPrinter&lt;SparcAsmPrinter&gt; X(TheSparcTarget);
471 }
472 </pre>
473 </div>
474
475 <p>
476 For more information, see
477 "<a href="/doxygen/TargetRegistry_8h-source.html">llvm/Target/TargetRegistry.h</a>".
478 </p>
479
480 </div>
481
482 <!-- *********************************************************************** -->
483 <div class="doc_section">
484   <a name="RegisterSet">Register Set and Register Classes</a>
485 </div>
486 <!-- *********************************************************************** -->
487
488 <div class="doc_text">
489
490 <p>
491 You should describe a concrete target-specific class that represents the
492 register file of a target machine. This class is called <tt>XXXRegisterInfo</tt>
493 (where <tt>XXX</tt> identifies the target) and represents the class register
494 file data that is used for register allocation. It also describes the
495 interactions between registers.
496 </p>
497
498 <p>
499 You also need to define register classes to categorize related registers. A
500 register class should be added for groups of registers that are all treated the
501 same way for some instruction. Typical examples are register classes for
502 integer, floating-point, or vector registers. A register allocator allows an
503 instruction to use any register in a specified register class to perform the
504 instruction in a similar manner. Register classes allocate virtual registers to
505 instructions from these sets, and register classes let the target-independent
506 register allocator automatically choose the actual registers.
507 </p>
508
509 <p>
510 Much of the code for registers, including register definition, register aliases,
511 and register classes, is generated by TableGen from <tt>XXXRegisterInfo.td</tt>
512 input files and placed in <tt>XXXGenRegisterInfo.h.inc</tt> and
513 <tt>XXXGenRegisterInfo.inc</tt> output files. Some of the code in the
514 implementation of <tt>XXXRegisterInfo</tt> requires hand-coding.
515 </p>
516
517 </div>
518
519 <!-- ======================================================================= -->
520 <div class="doc_subsection">
521   <a name="RegisterDef">Defining a Register</a>
522 </div>
523
524 <div class="doc_text">
525
526 <p>
527 The <tt>XXXRegisterInfo.td</tt> file typically starts with register definitions
528 for a target machine. The <tt>Register</tt> class (specified
529 in <tt>Target.td</tt>) is used to define an object for each register. The
530 specified string <tt>n</tt> becomes the <tt>Name</tt> of the register. The
531 basic <tt>Register</tt> object does not have any subregisters and does not
532 specify any aliases.
533 </p>
534
535 <div class="doc_code">
536 <pre>
537 class Register&lt;string n&gt; {
538   string Namespace = "";
539   string AsmName = n;
540   string Name = n;
541   int SpillSize = 0;
542   int SpillAlignment = 0;
543   list&lt;Register&gt; Aliases = [];
544   list&lt;Register&gt; SubRegs = [];
545   list&lt;int&gt; DwarfNumbers = [];
546 }
547 </pre>
548 </div>
549
550 <p>
551 For example, in the <tt>X86RegisterInfo.td</tt> file, there are register
552 definitions that utilize the Register class, such as:
553 </p>
554
555 <div class="doc_code">
556 <pre>
557 def AL : Register&lt;"AL"&gt;, DwarfRegNum&lt;[0, 0, 0]&gt;;
558 </pre>
559 </div>
560
561 <p>
562 This defines the register <tt>AL</tt> and assigns it values (with
563 <tt>DwarfRegNum</tt>) that are used by <tt>gcc</tt>, <tt>gdb</tt>, or a debug
564 information writer to identify a register. For register
565 <tt>AL</tt>, <tt>DwarfRegNum</tt> takes an array of 3 values representing 3
566 different modes: the first element is for X86-64, the second for exception
567 handling (EH) on X86-32, and the third is generic. -1 is a special Dwarf number
568 that indicates the gcc number is undefined, and -2 indicates the register number
569 is invalid for this mode.
570 </p>
571
572 <p>
573 From the previously described line in the <tt>X86RegisterInfo.td</tt> file,
574 TableGen generates this code in the <tt>X86GenRegisterInfo.inc</tt> file:
575 </p>
576
577 <div class="doc_code">
578 <pre>
579 static const unsigned GR8[] = { X86::AL, ... };
580
581 const unsigned AL_AliasSet[] = { X86::AX, X86::EAX, X86::RAX, 0 };
582
583 const TargetRegisterDesc RegisterDescriptors[] = { 
584   ...
585 { "AL", "AL", AL_AliasSet, Empty_SubRegsSet, Empty_SubRegsSet, AL_SuperRegsSet }, ...
586 </pre>
587 </div>
588
589 <p>
590 From the register info file, TableGen generates a <tt>TargetRegisterDesc</tt>
591 object for each register. <tt>TargetRegisterDesc</tt> is defined in
592 <tt>include/llvm/Target/TargetRegisterInfo.h</tt> with the following fields:
593 </p>
594
595 <div class="doc_code">
596 <pre>
597 struct TargetRegisterDesc {
598   const char     *AsmName;      // Assembly language name for the register
599   const char     *Name;         // Printable name for the reg (for debugging)
600   const unsigned *AliasSet;     // Register Alias Set
601   const unsigned *SubRegs;      // Sub-register set
602   const unsigned *ImmSubRegs;   // Immediate sub-register set
603   const unsigned *SuperRegs;    // Super-register set
604 };</pre>
605 </div>
606
607 <p>
608 TableGen uses the entire target description file (<tt>.td</tt>) to determine
609 text names for the register (in the <tt>AsmName</tt> and <tt>Name</tt> fields of
610 <tt>TargetRegisterDesc</tt>) and the relationships of other registers to the
611 defined register (in the other <tt>TargetRegisterDesc</tt> fields). In this
612 example, other definitions establish the registers "<tt>AX</tt>",
613 "<tt>EAX</tt>", and "<tt>RAX</tt>" as aliases for one another, so TableGen
614 generates a null-terminated array (<tt>AL_AliasSet</tt>) for this register alias
615 set.
616 </p>
617
618 <p>
619 The <tt>Register</tt> class is commonly used as a base class for more complex
620 classes. In <tt>Target.td</tt>, the <tt>Register</tt> class is the base for the
621 <tt>RegisterWithSubRegs</tt> class that is used to define registers that need to
622 specify subregisters in the <tt>SubRegs</tt> list, as shown here:
623 </p>
624
625 <div class="doc_code">
626 <pre>
627 class RegisterWithSubRegs&lt;string n,
628 list&lt;Register&gt; subregs&gt; : Register&lt;n&gt; {
629   let SubRegs = subregs;
630 }
631 </pre>
632 </div>
633
634 <p>
635 In <tt>SparcRegisterInfo.td</tt>, additional register classes are defined for
636 SPARC: a Register subclass, SparcReg, and further subclasses: <tt>Ri</tt>,
637 <tt>Rf</tt>, and <tt>Rd</tt>. SPARC registers are identified by 5-bit ID
638 numbers, which is a feature common to these subclasses. Note the use of
639 '<tt>let</tt>' expressions to override values that are initially defined in a
640 superclass (such as <tt>SubRegs</tt> field in the <tt>Rd</tt> class).
641 </p>
642
643 <div class="doc_code">
644 <pre>
645 class SparcReg&lt;string n&gt; : Register&lt;n&gt; {
646   field bits&lt;5&gt; Num;
647   let Namespace = "SP";
648 }
649 // Ri - 32-bit integer registers
650 class Ri&lt;bits&lt;5&gt; num, string n&gt; :
651 SparcReg&lt;n&gt; {
652   let Num = num;
653 }
654 // Rf - 32-bit floating-point registers
655 class Rf&lt;bits&lt;5&gt; num, string n&gt; :
656 SparcReg&lt;n&gt; {
657   let Num = num;
658 }
659 // Rd - Slots in the FP register file for 64-bit
660 floating-point values.
661 class Rd&lt;bits&lt;5&gt; num, string n,
662 list&lt;Register&gt; subregs&gt; : SparcReg&lt;n&gt; {
663   let Num = num;
664   let SubRegs = subregs;
665 }
666 </pre>
667 </div>
668
669 <p>
670 In the <tt>SparcRegisterInfo.td</tt> file, there are register definitions that
671 utilize these subclasses of <tt>Register</tt>, such as:
672 </p>
673
674 <div class="doc_code">
675 <pre>
676 def G0 : Ri&lt; 0, "G0"&gt;,
677 DwarfRegNum&lt;[0]&gt;;
678 def G1 : Ri&lt; 1, "G1"&gt;, DwarfRegNum&lt;[1]&gt;;
679 ...
680 def F0 : Rf&lt; 0, "F0"&gt;,
681 DwarfRegNum&lt;[32]&gt;;
682 def F1 : Rf&lt; 1, "F1"&gt;,
683 DwarfRegNum&lt;[33]&gt;;
684 ...
685 def D0 : Rd&lt; 0, "F0", [F0, F1]&gt;,
686 DwarfRegNum&lt;[32]&gt;;
687 def D1 : Rd&lt; 2, "F2", [F2, F3]&gt;,
688 DwarfRegNum&lt;[34]&gt;;
689 </pre>
690 </div>
691
692 <p>
693 The last two registers shown above (<tt>D0</tt> and <tt>D1</tt>) are
694 double-precision floating-point registers that are aliases for pairs of
695 single-precision floating-point sub-registers. In addition to aliases, the
696 sub-register and super-register relationships of the defined register are in
697 fields of a register's TargetRegisterDesc.
698 </p>
699
700 </div>
701
702 <!-- ======================================================================= -->
703 <div class="doc_subsection">
704   <a name="RegisterClassDef">Defining a Register Class</a>
705 </div>
706
707 <div class="doc_text">
708
709 <p>
710 The <tt>RegisterClass</tt> class (specified in <tt>Target.td</tt>) is used to
711 define an object that represents a group of related registers and also defines
712 the default allocation order of the registers. A target description file
713 <tt>XXXRegisterInfo.td</tt> that uses <tt>Target.td</tt> can construct register
714 classes using the following class:
715 </p>
716
717 <div class="doc_code">
718 <pre>
719 class RegisterClass&lt;string namespace,
720 list&lt;ValueType&gt; regTypes, int alignment,
721                     list&lt;Register&gt; regList&gt; {
722   string Namespace = namespace;
723   list&lt;ValueType&gt; RegTypes = regTypes;
724   int Size = 0;  // spill size, in bits; zero lets tblgen pick the size
725   int Alignment = alignment;
726
727   // CopyCost is the cost of copying a value between two registers
728   // default value 1 means a single instruction
729   // A negative value means copying is extremely expensive or impossible
730   int CopyCost = 1;  
731   list&lt;Register&gt; MemberList = regList;
732   
733   // for register classes that are subregisters of this class
734   list&lt;RegisterClass&gt; SubRegClassList = [];  
735   
736   code MethodProtos = [{}];  // to insert arbitrary code
737   code MethodBodies = [{}];
738 }
739 </pre>
740 </div>
741
742 <p>To define a RegisterClass, use the following 4 arguments:</p>
743
744 <ul>
745 <li>The first argument of the definition is the name of the namespace.</li>
746
747 <li>The second argument is a list of <tt>ValueType</tt> register type values
748     that are defined in <tt>include/llvm/CodeGen/ValueTypes.td</tt>. Defined
749     values include integer types (such as <tt>i16</tt>, <tt>i32</tt>,
750     and <tt>i1</tt> for Boolean), floating-point types
751     (<tt>f32</tt>, <tt>f64</tt>), and vector types (for example, <tt>v8i16</tt>
752     for an <tt>8 x i16</tt> vector). All registers in a <tt>RegisterClass</tt>
753     must have the same <tt>ValueType</tt>, but some registers may store vector
754     data in different configurations. For example a register that can process a
755     128-bit vector may be able to handle 16 8-bit integer elements, 8 16-bit
756     integers, 4 32-bit integers, and so on. </li>
757
758 <li>The third argument of the <tt>RegisterClass</tt> definition specifies the
759     alignment required of the registers when they are stored or loaded to
760     memory.</li>
761
762 <li>The final argument, <tt>regList</tt>, specifies which registers are in this
763     class.  If an <tt>allocation_order_*</tt> method is not specified,
764     then <tt>regList</tt> also defines the order of allocation used by the
765     register allocator.</li>
766 </ul>
767
768 <p>
769 In <tt>SparcRegisterInfo.td</tt>, three RegisterClass objects are defined:
770 <tt>FPRegs</tt>, <tt>DFPRegs</tt>, and <tt>IntRegs</tt>. For all three register
771 classes, the first argument defines the namespace with the string
772 '<tt>SP</tt>'. <tt>FPRegs</tt> defines a group of 32 single-precision
773 floating-point registers (<tt>F0</tt> to <tt>F31</tt>); <tt>DFPRegs</tt> defines
774 a group of 16 double-precision registers
775 (<tt>D0-D15</tt>). For <tt>IntRegs</tt>, the <tt>MethodProtos</tt>
776 and <tt>MethodBodies</tt> methods are used by TableGen to insert the specified
777 code into generated output.
778 </p>
779
780 <div class="doc_code">
781 <pre>
782 def FPRegs : RegisterClass&lt;"SP", [f32], 32,
783   [F0, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12, F13, F14, F15,
784    F16, F17, F18, F19, F20, F21, F22, F23, F24, F25, F26, F27, F28, F29, F30, F31]&gt;;
785
786 def DFPRegs : RegisterClass&lt;"SP", [f64], 64,
787   [D0, D1, D2, D3, D4, D5, D6, D7, D8, D9, D10, D11, D12, D13, D14, D15]&gt;;
788 &nbsp;
789 def IntRegs : RegisterClass&lt;"SP", [i32], 32,
790     [L0, L1, L2, L3, L4, L5, L6, L7,
791      I0, I1, I2, I3, I4, I5,
792      O0, O1, O2, O3, O4, O5, O7,
793      G1,
794      // Non-allocatable regs:
795      G2, G3, G4, 
796      O6,        // stack ptr
797     I6,        // frame ptr
798      I7,        // return address
799      G0,        // constant zero
800      G5, G6, G7 // reserved for kernel
801     ]&gt; {
802   let MethodProtos = [{
803     iterator allocation_order_end(const MachineFunction &amp;MF) const;
804   }];
805   let MethodBodies = [{
806     IntRegsClass::iterator
807     IntRegsClass::allocation_order_end(const MachineFunction &amp;MF) const {
808       return end() - 10  // Don't allocate special registers
809          -1;
810     }
811   }];
812 }
813 </pre>
814 </div>
815
816 <p>
817 Using <tt>SparcRegisterInfo.td</tt> with TableGen generates several output files
818 that are intended for inclusion in other source code that you write.
819 <tt>SparcRegisterInfo.td</tt> generates <tt>SparcGenRegisterInfo.h.inc</tt>,
820 which should be included in the header file for the implementation of the SPARC
821 register implementation that you write (<tt>SparcRegisterInfo.h</tt>). In
822 <tt>SparcGenRegisterInfo.h.inc</tt> a new structure is defined called
823 <tt>SparcGenRegisterInfo</tt> that uses <tt>TargetRegisterInfo</tt> as its
824 base. It also specifies types, based upon the defined register
825 classes: <tt>DFPRegsClass</tt>, <tt>FPRegsClass</tt>, and <tt>IntRegsClass</tt>.
826 </p>
827
828 <p>
829 <tt>SparcRegisterInfo.td</tt> also generates <tt>SparcGenRegisterInfo.inc</tt>,
830 which is included at the bottom of <tt>SparcRegisterInfo.cpp</tt>, the SPARC
831 register implementation. The code below shows only the generated integer
832 registers and associated register classes. The order of registers
833 in <tt>IntRegs</tt> reflects the order in the definition of <tt>IntRegs</tt> in
834 the target description file. Take special note of the use
835 of <tt>MethodBodies</tt> in <tt>SparcRegisterInfo.td</tt> to create code in
836 <tt>SparcGenRegisterInfo.inc</tt>. <tt>MethodProtos</tt> generates similar code
837 in <tt>SparcGenRegisterInfo.h.inc</tt>.
838 </p>
839
840 <div class="doc_code">
841 <pre>  // IntRegs Register Class...
842   static const unsigned IntRegs[] = {
843     SP::L0, SP::L1, SP::L2, SP::L3, SP::L4, SP::L5,
844     SP::L6, SP::L7, SP::I0, SP::I1, SP::I2, SP::I3,
845     SP::I4, SP::I5, SP::O0, SP::O1, SP::O2, SP::O3,
846     SP::O4, SP::O5, SP::O7, SP::G1, SP::G2, SP::G3,
847     SP::G4, SP::O6, SP::I6, SP::I7, SP::G0, SP::G5,
848     SP::G6, SP::G7, 
849   };
850
851   // IntRegsVTs Register Class Value Types...
852   static const MVT::ValueType IntRegsVTs[] = {
853     MVT::i32, MVT::Other
854   };
855
856 namespace SP {   // Register class instances
857   DFPRegsClass&nbsp;&nbsp;&nbsp; DFPRegsRegClass;
858   FPRegsClass&nbsp;&nbsp;&nbsp;&nbsp; FPRegsRegClass;
859   IntRegsClass&nbsp;&nbsp;&nbsp; IntRegsRegClass;
860 ...
861   // IntRegs Sub-register Classess...
862   static const TargetRegisterClass* const IntRegsSubRegClasses [] = {
863     NULL
864   };
865 ...
866   // IntRegs Super-register Classess...
867   static const TargetRegisterClass* const IntRegsSuperRegClasses [] = {
868     NULL
869   };
870 ...
871   // IntRegs Register Class sub-classes...
872   static const TargetRegisterClass* const IntRegsSubclasses [] = {
873     NULL
874   };
875 ...
876   // IntRegs Register Class super-classes...
877   static const TargetRegisterClass* const IntRegsSuperclasses [] = {
878     NULL
879   };
880 ...
881   IntRegsClass::iterator
882   IntRegsClass::allocation_order_end(const MachineFunction &amp;MF) const {
883      return end()-10  // Don't allocate special registers
884          -1;
885   }
886   
887   IntRegsClass::IntRegsClass() : TargetRegisterClass(IntRegsRegClassID, 
888     IntRegsVTs, IntRegsSubclasses, IntRegsSuperclasses, IntRegsSubRegClasses, 
889     IntRegsSuperRegClasses, 4, 4, 1, IntRegs, IntRegs + 32) {}
890 }
891 </pre>
892 </div>
893
894 </div>
895
896 <!-- ======================================================================= -->
897 <div class="doc_subsection">
898   <a name="implementRegister">Implement a subclass of</a> 
899   <a href="http://www.llvm.org/docs/CodeGenerator.html#targetregisterinfo">TargetRegisterInfo</a>
900 </div>
901
902 <div class="doc_text">
903
904 <p>
905 The final step is to hand code portions of <tt>XXXRegisterInfo</tt>, which
906 implements the interface described in <tt>TargetRegisterInfo.h</tt>. These
907 functions return <tt>0</tt>, <tt>NULL</tt>, or <tt>false</tt>, unless
908 overridden. Here is a list of functions that are overridden for the SPARC
909 implementation in <tt>SparcRegisterInfo.cpp</tt>:
910 </p>
911
912 <ul>
913 <li><tt>getCalleeSavedRegs</tt> &mdash; Returns a list of callee-saved registers
914     in the order of the desired callee-save stack frame offset.</li>
915
916 <li><tt>getReservedRegs</tt> &mdash; Returns a bitset indexed by physical
917     register numbers, indicating if a particular register is unavailable.</li>
918
919 <li><tt>hasFP</tt> &mdash; Return a Boolean indicating if a function should have
920     a dedicated frame pointer register.</li>
921
922 <li><tt>eliminateCallFramePseudoInstr</tt> &mdash; If call frame setup or
923     destroy pseudo instructions are used, this can be called to eliminate
924     them.</li>
925
926 <li><tt>eliminateFrameIndex</tt> &mdash; Eliminate abstract frame indices from
927     instructions that may use them.</li>
928
929 <li><tt>emitPrologue</tt> &mdash; Insert prologue code into the function.</li>
930
931 <li><tt>emitEpilogue</tt> &mdash; Insert epilogue code into the function.</li>
932 </ul>
933
934 </div>
935
936 <!-- *********************************************************************** -->
937 <div class="doc_section">
938   <a name="InstructionSet">Instruction Set</a>
939 </div>
940
941 <!-- *********************************************************************** -->
942 <div class="doc_text">
943
944 <p>
945 During the early stages of code generation, the LLVM IR code is converted to a
946 <tt>SelectionDAG</tt> with nodes that are instances of the <tt>SDNode</tt> class
947 containing target instructions. An <tt>SDNode</tt> has an opcode, operands, type
948 requirements, and operation properties. For example, is an operation
949 commutative, does an operation load from memory. The various operation node
950 types are described in the <tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt>
951 file (values of the <tt>NodeType</tt> enum in the <tt>ISD</tt> namespace).
952 </p>
953
954 <p>
955 TableGen uses the following target description (<tt>.td</tt>) input files to
956 generate much of the code for instruction definition:
957 </p>
958
959 <ul>
960 <li><tt>Target.td</tt> &mdash; Where the <tt>Instruction</tt>, <tt>Operand</tt>,
961     <tt>InstrInfo</tt>, and other fundamental classes are defined.</li>
962
963 <li><tt>TargetSelectionDAG.td</tt>&mdash; Used by <tt>SelectionDAG</tt>
964     instruction selection generators, contains <tt>SDTC*</tt> classes (selection
965     DAG type constraint), definitions of <tt>SelectionDAG</tt> nodes (such as
966     <tt>imm</tt>, <tt>cond</tt>, <tt>bb</tt>, <tt>add</tt>, <tt>fadd</tt>,
967     <tt>sub</tt>), and pattern support (<tt>Pattern</tt>, <tt>Pat</tt>,
968     <tt>PatFrag</tt>, <tt>PatLeaf</tt>, <tt>ComplexPattern</tt>.</li>
969
970 <li><tt>XXXInstrFormats.td</tt> &mdash; Patterns for definitions of
971     target-specific instructions.</li>
972
973 <li><tt>XXXInstrInfo.td</tt> &mdash; Target-specific definitions of instruction
974     templates, condition codes, and instructions of an instruction set. For
975     architecture modifications, a different file name may be used. For example,
976     for Pentium with SSE instruction, this file is <tt>X86InstrSSE.td</tt>, and
977     for Pentium with MMX, this file is <tt>X86InstrMMX.td</tt>.</li>
978 </ul>
979
980 <p>
981 There is also a target-specific <tt>XXX.td</tt> file, where <tt>XXX</tt> is the
982 name of the target. The <tt>XXX.td</tt> file includes the other <tt>.td</tt>
983 input files, but its contents are only directly important for subtargets.
984 </p>
985
986 <p>
987 You should describe a concrete target-specific class <tt>XXXInstrInfo</tt> that
988 represents machine instructions supported by a target machine.
989 <tt>XXXInstrInfo</tt> contains an array of <tt>XXXInstrDescriptor</tt> objects,
990 each of which describes one instruction. An instruction descriptor defines:</p>
991
992 <ul>
993 <li>Opcode mnemonic</li>
994
995 <li>Number of operands</li>
996
997 <li>List of implicit register definitions and uses</li>
998
999 <li>Target-independent properties (such as memory access, is commutable)</li>
1000
1001 <li>Target-specific flags </li>
1002 </ul>
1003
1004 <p>
1005 The Instruction class (defined in <tt>Target.td</tt>) is mostly used as a base
1006 for more complex instruction classes.
1007 </p>
1008
1009 <div class="doc_code">
1010 <pre>class Instruction {
1011   string Namespace = "";
1012   dag OutOperandList;       // An dag containing the MI def operand list.
1013   dag InOperandList;        // An dag containing the MI use operand list.
1014   string AsmString = "";    // The .s format to print the instruction with.
1015   list&lt;dag&gt; Pattern;  // Set to the DAG pattern for this instruction
1016   list&lt;Register&gt; Uses = []; 
1017   list&lt;Register&gt; Defs = [];
1018   list&lt;Predicate&gt; Predicates = [];  // predicates turned into isel match code
1019   ... remainder not shown for space ...
1020 }
1021 </pre>
1022 </div>
1023
1024 <p>
1025 A <tt>SelectionDAG</tt> node (<tt>SDNode</tt>) should contain an object
1026 representing a target-specific instruction that is defined
1027 in <tt>XXXInstrInfo.td</tt>. The instruction objects should represent
1028 instructions from the architecture manual of the target machine (such as the
1029 SPARC Architecture Manual for the SPARC target).
1030 </p>
1031
1032 <p>
1033 A single instruction from the architecture manual is often modeled as multiple
1034 target instructions, depending upon its operands. For example, a manual might
1035 describe an add instruction that takes a register or an immediate operand. An
1036 LLVM target could model this with two instructions named <tt>ADDri</tt> and
1037 <tt>ADDrr</tt>.
1038 </p>
1039
1040 <p>
1041 You should define a class for each instruction category and define each opcode
1042 as a subclass of the category with appropriate parameters such as the fixed
1043 binary encoding of opcodes and extended opcodes. You should map the register
1044 bits to the bits of the instruction in which they are encoded (for the
1045 JIT). Also you should specify how the instruction should be printed when the
1046 automatic assembly printer is used.
1047 </p>
1048
1049 <p>
1050 As is described in the SPARC Architecture Manual, Version 8, there are three
1051 major 32-bit formats for instructions. Format 1 is only for the <tt>CALL</tt>
1052 instruction. Format 2 is for branch on condition codes and <tt>SETHI</tt> (set
1053 high bits of a register) instructions.  Format 3 is for other instructions.
1054 </p>
1055
1056 <p>
1057 Each of these formats has corresponding classes in <tt>SparcInstrFormat.td</tt>.
1058 <tt>InstSP</tt> is a base class for other instruction classes. Additional base
1059 classes are specified for more precise formats: for example
1060 in <tt>SparcInstrFormat.td</tt>, <tt>F2_1</tt> is for <tt>SETHI</tt>,
1061 and <tt>F2_2</tt> is for branches. There are three other base
1062 classes: <tt>F3_1</tt> for register/register operations, <tt>F3_2</tt> for
1063 register/immediate operations, and <tt>F3_3</tt> for floating-point
1064 operations. <tt>SparcInstrInfo.td</tt> also adds the base class Pseudo for
1065 synthetic SPARC instructions.
1066 </p>
1067
1068 <p>
1069 <tt>SparcInstrInfo.td</tt> largely consists of operand and instruction
1070 definitions for the SPARC target. In <tt>SparcInstrInfo.td</tt>, the following
1071 target description file entry, <tt>LDrr</tt>, defines the Load Integer
1072 instruction for a Word (the <tt>LD</tt> SPARC opcode) from a memory address to a
1073 register. The first parameter, the value 3 (<tt>11<sub>2</sub></tt>), is the
1074 operation value for this category of operation. The second parameter
1075 (<tt>000000<sub>2</sub></tt>) is the specific operation value
1076 for <tt>LD</tt>/Load Word. The third parameter is the output destination, which
1077 is a register operand and defined in the <tt>Register</tt> target description
1078 file (<tt>IntRegs</tt>).
1079 </p>
1080
1081 <div class="doc_code">
1082 <pre>def LDrr : F3_1 &lt;3, 0b000000, (outs IntRegs:$dst), (ins MEMrr:$addr),
1083                  "ld [$addr], $dst",
1084                  [(set IntRegs:$dst, (load ADDRrr:$addr))]&gt;;
1085 </pre>
1086 </div>
1087
1088 <p>
1089 The fourth parameter is the input source, which uses the address
1090 operand <tt>MEMrr</tt> that is defined earlier in <tt>SparcInstrInfo.td</tt>:
1091 </p>
1092
1093 <div class="doc_code">
1094 <pre>def MEMrr : Operand&lt;i32&gt; {
1095   let PrintMethod = "printMemOperand";
1096   let MIOperandInfo = (ops IntRegs, IntRegs);
1097 }
1098 </pre>
1099 </div>
1100
1101 <p>
1102 The fifth parameter is a string that is used by the assembly printer and can be
1103 left as an empty string until the assembly printer interface is implemented. The
1104 sixth and final parameter is the pattern used to match the instruction during
1105 the SelectionDAG Select Phase described in
1106 (<a href="http://www.llvm.org/docs/CodeGenerator.html">The LLVM
1107 Target-Independent Code Generator</a>).  This parameter is detailed in the next
1108 section, <a href="#InstructionSelector">Instruction Selector</a>.
1109 </p>
1110
1111 <p>
1112 Instruction class definitions are not overloaded for different operand types, so
1113 separate versions of instructions are needed for register, memory, or immediate
1114 value operands. For example, to perform a Load Integer instruction for a Word
1115 from an immediate operand to a register, the following instruction class is
1116 defined:
1117 </p>
1118
1119 <div class="doc_code">
1120 <pre>def LDri : F3_2 &lt;3, 0b000000, (outs IntRegs:$dst), (ins MEMri:$addr),
1121                  "ld [$addr], $dst",
1122                  [(set IntRegs:$dst, (load ADDRri:$addr))]&gt;;
1123 </pre>
1124 </div>
1125
1126 <p>
1127 Writing these definitions for so many similar instructions can involve a lot of
1128 cut and paste. In td files, the <tt>multiclass</tt> directive enables the
1129 creation of templates to define several instruction classes at once (using
1130 the <tt>defm</tt> directive). For example in <tt>SparcInstrInfo.td</tt>, the
1131 <tt>multiclass</tt> pattern <tt>F3_12</tt> is defined to create 2 instruction
1132 classes each time <tt>F3_12</tt> is invoked:
1133 </p>
1134
1135 <div class="doc_code">
1136 <pre>multiclass F3_12 &lt;string OpcStr, bits&lt;6&gt; Op3Val, SDNode OpNode&gt; {
1137   def rr  : F3_1 &lt;2, Op3Val, 
1138                  (outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c),
1139                  !strconcat(OpcStr, " $b, $c, $dst"),
1140                  [(set IntRegs:$dst, (OpNode IntRegs:$b, IntRegs:$c))]&gt;;
1141   def ri  : F3_2 &lt;2, Op3Val,
1142                  (outs IntRegs:$dst), (ins IntRegs:$b, i32imm:$c),
1143                  !strconcat(OpcStr, " $b, $c, $dst"),
1144                  [(set IntRegs:$dst, (OpNode IntRegs:$b, simm13:$c))]&gt;;
1145 }
1146 </pre>
1147 </div>
1148
1149 <p>
1150 So when the <tt>defm</tt> directive is used for the <tt>XOR</tt>
1151 and <tt>ADD</tt> instructions, as seen below, it creates four instruction
1152 objects: <tt>XORrr</tt>, <tt>XORri</tt>, <tt>ADDrr</tt>, and <tt>ADDri</tt>.
1153 </p>
1154
1155 <div class="doc_code">
1156 <pre>
1157 defm XOR   : F3_12&lt;"xor", 0b000011, xor&gt;;
1158 defm ADD   : F3_12&lt;"add", 0b000000, add&gt;;
1159 </pre>
1160 </div>
1161
1162 <p>
1163 <tt>SparcInstrInfo.td</tt> also includes definitions for condition codes that
1164 are referenced by branch instructions. The following definitions
1165 in <tt>SparcInstrInfo.td</tt> indicate the bit location of the SPARC condition
1166 code. For example, the 10<sup>th</sup> bit represents the 'greater than'
1167 condition for integers, and the 22<sup>nd</sup> bit represents the 'greater
1168 than' condition for floats.
1169 </p>
1170
1171 <div class="doc_code">
1172 <pre>
1173 def ICC_NE  : ICC_VAL&lt; 9&gt;;  // Not Equal
1174 def ICC_E   : ICC_VAL&lt; 1&gt;;  // Equal
1175 def ICC_G   : ICC_VAL&lt;10&gt;;  // Greater
1176 ...
1177 def FCC_U   : FCC_VAL&lt;23&gt;;  // Unordered
1178 def FCC_G   : FCC_VAL&lt;22&gt;;  // Greater
1179 def FCC_UG  : FCC_VAL&lt;21&gt;;  // Unordered or Greater
1180 ...
1181 </pre>
1182 </div>
1183
1184 <p>
1185 (Note that <tt>Sparc.h</tt> also defines enums that correspond to the same SPARC
1186 condition codes. Care must be taken to ensure the values in <tt>Sparc.h</tt>
1187 correspond to the values in <tt>SparcInstrInfo.td</tt>. I.e.,
1188 <tt>SPCC::ICC_NE = 9</tt>, <tt>SPCC::FCC_U = 23</tt> and so on.)
1189 </p>
1190
1191 </div>
1192
1193 <!-- ======================================================================= -->
1194 <div class="doc_subsection">
1195   <a name="operandMapping">Instruction Operand Mapping</a>
1196 </div>
1197
1198 <div class="doc_text">
1199
1200 <p>
1201 The code generator backend maps instruction operands to fields in the
1202 instruction.  Operands are assigned to unbound fields in the instruction in the
1203 order they are defined. Fields are bound when they are assigned a value.  For
1204 example, the Sparc target defines the <tt>XNORrr</tt> instruction as
1205 a <tt>F3_1</tt> format instruction having three operands.
1206 </p>
1207
1208 <div class="doc_code">
1209 <pre>
1210 def XNORrr  : F3_1&lt;2, 0b000111,
1211                    (outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c),
1212                    "xnor $b, $c, $dst",
1213                    [(set IntRegs:$dst, (not (xor IntRegs:$b, IntRegs:$c)))]&gt;;
1214 </pre>
1215 </div>
1216
1217 <p>
1218 The instruction templates in <tt>SparcInstrFormats.td</tt> show the base class
1219 for <tt>F3_1</tt> is <tt>InstSP</tt>.
1220 </p>
1221
1222 <div class="doc_code">
1223 <pre>
1224 class InstSP&lt;dag outs, dag ins, string asmstr, list&lt;dag&gt; pattern&gt; : Instruction {
1225   field bits&lt;32&gt; Inst;
1226   let Namespace = "SP";
1227   bits&lt;2&gt; op;
1228   let Inst{31-30} = op;       
1229   dag OutOperandList = outs;
1230   dag InOperandList = ins;
1231   let AsmString   = asmstr;
1232   let Pattern = pattern;
1233 }
1234 </pre>
1235 </div>
1236
1237 <p><tt>InstSP</tt> leaves the <tt>op</tt> field unbound.</p>
1238
1239 <div class="doc_code">
1240 <pre>
1241 class F3&lt;dag outs, dag ins, string asmstr, list&lt;dag&gt; pattern&gt;
1242     : InstSP&lt;outs, ins, asmstr, pattern&gt; {
1243   bits&lt;5&gt; rd;
1244   bits&lt;6&gt; op3;
1245   bits&lt;5&gt; rs1;
1246   let op{1} = 1;   // Op = 2 or 3
1247   let Inst{29-25} = rd;
1248   let Inst{24-19} = op3;
1249   let Inst{18-14} = rs1;
1250 }
1251 </pre>
1252 </div>
1253
1254 <p>
1255 <tt>F3</tt> binds the <tt>op</tt> field and defines the <tt>rd</tt>,
1256 <tt>op3</tt>, and <tt>rs1</tt> fields.  <tt>F3</tt> format instructions will
1257 bind the operands <tt>rd</tt>, <tt>op3</tt>, and <tt>rs1</tt> fields.
1258 </p>
1259
1260 <div class="doc_code">
1261 <pre>
1262 class F3_1&lt;bits&lt;2&gt; opVal, bits&lt;6&gt; op3val, dag outs, dag ins,
1263            string asmstr, list&lt;dag&gt; pattern&gt; : F3&lt;outs, ins, asmstr, pattern&gt; {
1264   bits&lt;8&gt; asi = 0; // asi not currently used
1265   bits&lt;5&gt; rs2;
1266   let op         = opVal;
1267   let op3        = op3val;
1268   let Inst{13}   = 0;     // i field = 0
1269   let Inst{12-5} = asi;   // address space identifier
1270   let Inst{4-0}  = rs2;
1271 }
1272 </pre>
1273 </div>
1274
1275 <p>
1276 <tt>F3_1</tt> binds the <tt>op3</tt> field and defines the <tt>rs2</tt>
1277 fields.  <tt>F3_1</tt> format instructions will bind the operands to the <tt>rd</tt>,
1278 <tt>rs1</tt>, and <tt>rs2</tt> fields. This results in the <tt>XNORrr</tt>
1279 instruction binding <tt>$dst</tt>, <tt>$b</tt>, and <tt>$c</tt> operands to
1280 the <tt>rd</tt>, <tt>rs1</tt>, and <tt>rs2</tt> fields respectively.
1281 </p>
1282
1283 </div>
1284
1285 <!-- ======================================================================= -->
1286 <div class="doc_subsection">
1287   <a name="implementInstr">Implement a subclass of </a>
1288   <a href="http://www.llvm.org/docs/CodeGenerator.html#targetinstrinfo">TargetInstrInfo</a>
1289 </div>
1290
1291 <div class="doc_text">
1292
1293 <p>
1294 The final step is to hand code portions of <tt>XXXInstrInfo</tt>, which
1295 implements the interface described in <tt>TargetInstrInfo.h</tt>. These
1296 functions return <tt>0</tt> or a Boolean or they assert, unless
1297 overridden. Here's a list of functions that are overridden for the SPARC
1298 implementation in <tt>SparcInstrInfo.cpp</tt>:
1299 </p>
1300
1301 <ul>
1302 <li><tt>isLoadFromStackSlot</tt> &mdash; If the specified machine instruction is
1303     a direct load from a stack slot, return the register number of the
1304     destination and the <tt>FrameIndex</tt> of the stack slot.</li>
1305
1306 <li><tt>isStoreToStackSlot</tt> &mdash; If the specified machine instruction is
1307     a direct store to a stack slot, return the register number of the
1308     destination and the <tt>FrameIndex</tt> of the stack slot.</li>
1309
1310 <li><tt>copyPhysReg</tt> &mdash; Copy values between a pair of physical
1311     registers.</li>
1312
1313 <li><tt>storeRegToStackSlot</tt> &mdash; Store a register value to a stack
1314     slot.</li>
1315
1316 <li><tt>loadRegFromStackSlot</tt> &mdash; Load a register value from a stack
1317     slot.</li>
1318
1319 <li><tt>storeRegToAddr</tt> &mdash; Store a register value to memory.</li>
1320
1321 <li><tt>loadRegFromAddr</tt> &mdash; Load a register value from memory.</li>
1322
1323 <li><tt>foldMemoryOperand</tt> &mdash; Attempt to combine instructions of any
1324     load or store instruction for the specified operand(s).</li>
1325 </ul>
1326
1327 </div>
1328
1329 <!-- ======================================================================= -->
1330 <div class="doc_subsection">
1331   <a name="branchFolding">Branch Folding and If Conversion</a>
1332 </div>
1333 <div class="doc_text">
1334
1335 <p>
1336 Performance can be improved by combining instructions or by eliminating
1337 instructions that are never reached. The <tt>AnalyzeBranch</tt> method
1338 in <tt>XXXInstrInfo</tt> may be implemented to examine conditional instructions
1339 and remove unnecessary instructions. <tt>AnalyzeBranch</tt> looks at the end of
1340 a machine basic block (MBB) for opportunities for improvement, such as branch
1341 folding and if conversion. The <tt>BranchFolder</tt> and <tt>IfConverter</tt>
1342 machine function passes (see the source files <tt>BranchFolding.cpp</tt> and
1343 <tt>IfConversion.cpp</tt> in the <tt>lib/CodeGen</tt> directory) call
1344 <tt>AnalyzeBranch</tt> to improve the control flow graph that represents the
1345 instructions.
1346 </p>
1347
1348 <p>
1349 Several implementations of <tt>AnalyzeBranch</tt> (for ARM, Alpha, and X86) can
1350 be examined as models for your own <tt>AnalyzeBranch</tt> implementation. Since
1351 SPARC does not implement a useful <tt>AnalyzeBranch</tt>, the ARM target
1352 implementation is shown below.
1353 </p>
1354
1355 <p><tt>AnalyzeBranch</tt> returns a Boolean value and takes four parameters:</p>
1356
1357 <ul>
1358 <li><tt>MachineBasicBlock &amp;MBB</tt> &mdash; The incoming block to be
1359     examined.</li>
1360
1361 <li><tt>MachineBasicBlock *&amp;TBB</tt> &mdash; A destination block that is
1362     returned. For a conditional branch that evaluates to true, <tt>TBB</tt> is
1363     the destination.</li>
1364
1365 <li><tt>MachineBasicBlock *&amp;FBB</tt> &mdash; For a conditional branch that
1366     evaluates to false, <tt>FBB</tt> is returned as the destination.</li>
1367
1368 <li><tt>std::vector&lt;MachineOperand&gt; &amp;Cond</tt> &mdash; List of
1369     operands to evaluate a condition for a conditional branch.</li>
1370 </ul>
1371
1372 <p>
1373 In the simplest case, if a block ends without a branch, then it falls through to
1374 the successor block. No destination blocks are specified for either <tt>TBB</tt>
1375 or <tt>FBB</tt>, so both parameters return <tt>NULL</tt>. The start of
1376 the <tt>AnalyzeBranch</tt> (see code below for the ARM target) shows the
1377 function parameters and the code for the simplest case.
1378 </p>
1379
1380 <div class="doc_code">
1381 <pre>bool ARMInstrInfo::AnalyzeBranch(MachineBasicBlock &amp;MBB,
1382         MachineBasicBlock *&amp;TBB, MachineBasicBlock *&amp;FBB,
1383         std::vector&lt;MachineOperand&gt; &amp;Cond) const
1384 {
1385   MachineBasicBlock::iterator I = MBB.end();
1386   if (I == MBB.begin() || !isUnpredicatedTerminator(--I))
1387     return false;
1388 </pre>
1389 </div>
1390
1391 <p>
1392 If a block ends with a single unconditional branch instruction, then
1393 <tt>AnalyzeBranch</tt> (shown below) should return the destination of that
1394 branch in the <tt>TBB</tt> parameter.
1395 </p>
1396
1397 <div class="doc_code">
1398 <pre>
1399   if (LastOpc == ARM::B || LastOpc == ARM::tB) {
1400     TBB = LastInst-&gt;getOperand(0).getMBB();
1401     return false;
1402   }
1403 </pre>
1404 </div>
1405
1406 <p>
1407 If a block ends with two unconditional branches, then the second branch is never
1408 reached. In that situation, as shown below, remove the last branch instruction
1409 and return the penultimate branch in the <tt>TBB</tt> parameter.
1410 </p>
1411
1412 <div class="doc_code">
1413 <pre>
1414   if ((SecondLastOpc == ARM::B || SecondLastOpc==ARM::tB) &amp;&amp;
1415       (LastOpc == ARM::B || LastOpc == ARM::tB)) {
1416     TBB = SecondLastInst-&gt;getOperand(0).getMBB();
1417     I = LastInst;
1418     I-&gt;eraseFromParent();
1419     return false;
1420   }
1421 </pre>
1422 </div>
1423
1424 <p>
1425 A block may end with a single conditional branch instruction that falls through
1426 to successor block if the condition evaluates to false. In that case,
1427 <tt>AnalyzeBranch</tt> (shown below) should return the destination of that
1428 conditional branch in the <tt>TBB</tt> parameter and a list of operands in
1429 the <tt>Cond</tt> parameter to evaluate the condition.
1430 </p>
1431
1432 <div class="doc_code">
1433 <pre>
1434   if (LastOpc == ARM::Bcc || LastOpc == ARM::tBcc) {
1435     // Block ends with fall-through condbranch.
1436     TBB = LastInst-&gt;getOperand(0).getMBB();
1437     Cond.push_back(LastInst-&gt;getOperand(1));
1438     Cond.push_back(LastInst-&gt;getOperand(2));
1439     return false;
1440   }
1441 </pre>
1442 </div>
1443
1444 <p>
1445 If a block ends with both a conditional branch and an ensuing unconditional
1446 branch, then <tt>AnalyzeBranch</tt> (shown below) should return the conditional
1447 branch destination (assuming it corresponds to a conditional evaluation of
1448 '<tt>true</tt>') in the <tt>TBB</tt> parameter and the unconditional branch
1449 destination in the <tt>FBB</tt> (corresponding to a conditional evaluation of
1450 '<tt>false</tt>').  A list of operands to evaluate the condition should be
1451 returned in the <tt>Cond</tt> parameter.
1452 </p>
1453
1454 <div class="doc_code">
1455 <pre>
1456   unsigned SecondLastOpc = SecondLastInst-&gt;getOpcode();
1457
1458   if ((SecondLastOpc == ARM::Bcc &amp;&amp; LastOpc == ARM::B) ||
1459       (SecondLastOpc == ARM::tBcc &amp;&amp; LastOpc == ARM::tB)) {
1460     TBB =  SecondLastInst-&gt;getOperand(0).getMBB();
1461     Cond.push_back(SecondLastInst-&gt;getOperand(1));
1462     Cond.push_back(SecondLastInst-&gt;getOperand(2));
1463     FBB = LastInst-&gt;getOperand(0).getMBB();
1464     return false;
1465   }
1466 </pre>
1467 </div>
1468
1469 <p>
1470 For the last two cases (ending with a single conditional branch or ending with
1471 one conditional and one unconditional branch), the operands returned in
1472 the <tt>Cond</tt> parameter can be passed to methods of other instructions to
1473 create new branches or perform other operations. An implementation
1474 of <tt>AnalyzeBranch</tt> requires the helper methods <tt>RemoveBranch</tt>
1475 and <tt>InsertBranch</tt> to manage subsequent operations.
1476 </p>
1477
1478 <p>
1479 <tt>AnalyzeBranch</tt> should return false indicating success in most circumstances.
1480 <tt>AnalyzeBranch</tt> should only return true when the method is stumped about what to
1481 do, for example, if a block has three terminating branches. <tt>AnalyzeBranch</tt> may
1482 return true if it encounters a terminator it cannot handle, such as an indirect
1483 branch.
1484 </p>
1485
1486 </div>
1487
1488 <!-- *********************************************************************** -->
1489 <div class="doc_section">
1490   <a name="InstructionSelector">Instruction Selector</a>
1491 </div>
1492 <!-- *********************************************************************** -->
1493
1494 <div class="doc_text">
1495
1496 <p>
1497 LLVM uses a <tt>SelectionDAG</tt> to represent LLVM IR instructions, and nodes
1498 of the <tt>SelectionDAG</tt> ideally represent native target
1499 instructions. During code generation, instruction selection passes are performed
1500 to convert non-native DAG instructions into native target-specific
1501 instructions. The pass described in <tt>XXXISelDAGToDAG.cpp</tt> is used to
1502 match patterns and perform DAG-to-DAG instruction selection. Optionally, a pass
1503 may be defined (in <tt>XXXBranchSelector.cpp</tt>) to perform similar DAG-to-DAG
1504 operations for branch instructions. Later, the code in
1505 <tt>XXXISelLowering.cpp</tt> replaces or removes operations and data types not
1506 supported natively (legalizes) in a <tt>SelectionDAG</tt>.
1507 </p>
1508
1509 <p>
1510 TableGen generates code for instruction selection using the following target
1511 description input files:
1512 </p>
1513
1514 <ul>
1515 <li><tt>XXXInstrInfo.td</tt> &mdash; Contains definitions of instructions in a
1516     target-specific instruction set, generates <tt>XXXGenDAGISel.inc</tt>, which
1517     is included in <tt>XXXISelDAGToDAG.cpp</tt>.</li>
1518
1519 <li><tt>XXXCallingConv.td</tt> &mdash; Contains the calling and return value
1520     conventions for the target architecture, and it generates
1521     <tt>XXXGenCallingConv.inc</tt>, which is included in
1522     <tt>XXXISelLowering.cpp</tt>.</li>
1523 </ul>
1524
1525 <p>
1526 The implementation of an instruction selection pass must include a header that
1527 declares the <tt>FunctionPass</tt> class or a subclass of <tt>FunctionPass</tt>. In
1528 <tt>XXXTargetMachine.cpp</tt>, a Pass Manager (PM) should add each instruction
1529 selection pass into the queue of passes to run.
1530 </p>
1531
1532 <p>
1533 The LLVM static compiler (<tt>llc</tt>) is an excellent tool for visualizing the
1534 contents of DAGs. To display the <tt>SelectionDAG</tt> before or after specific
1535 processing phases, use the command line options for <tt>llc</tt>, described
1536 at <a href="http://llvm.org/docs/CodeGenerator.html#selectiondag_process">
1537 SelectionDAG Instruction Selection Process</a>.
1538 </p>
1539
1540 <p>
1541 To describe instruction selector behavior, you should add patterns for lowering
1542 LLVM code into a <tt>SelectionDAG</tt> as the last parameter of the instruction
1543 definitions in <tt>XXXInstrInfo.td</tt>. For example, in
1544 <tt>SparcInstrInfo.td</tt>, this entry defines a register store operation, and
1545 the last parameter describes a pattern with the store DAG operator.
1546 </p>
1547
1548 <div class="doc_code">
1549 <pre>
1550 def STrr  : F3_1&lt; 3, 0b000100, (outs), (ins MEMrr:$addr, IntRegs:$src),
1551                  "st $src, [$addr]", [(store IntRegs:$src, ADDRrr:$addr)]&gt;;
1552 </pre>
1553 </div>
1554
1555 <p>
1556 <tt>ADDRrr</tt> is a memory mode that is also defined in
1557 <tt>SparcInstrInfo.td</tt>:
1558 </p>
1559
1560 <div class="doc_code">
1561 <pre>
1562 def ADDRrr : ComplexPattern&lt;i32, 2, "SelectADDRrr", [], []&gt;;
1563 </pre>
1564 </div>
1565
1566 <p>
1567 The definition of <tt>ADDRrr</tt> refers to <tt>SelectADDRrr</tt>, which is a
1568 function defined in an implementation of the Instructor Selector (such
1569 as <tt>SparcISelDAGToDAG.cpp</tt>).
1570 </p>
1571
1572 <p>
1573 In <tt>lib/Target/TargetSelectionDAG.td</tt>, the DAG operator for store is
1574 defined below:
1575 </p>
1576
1577 <div class="doc_code">
1578 <pre>
1579 def store : PatFrag&lt;(ops node:$val, node:$ptr),
1580                     (st node:$val, node:$ptr), [{
1581   if (StoreSDNode *ST = dyn_cast&lt;StoreSDNode&gt;(N))
1582     return !ST-&gt;isTruncatingStore() &amp;&amp; 
1583            ST-&gt;getAddressingMode() == ISD::UNINDEXED;
1584   return false;
1585 }]&gt;;
1586 </pre>
1587 </div>
1588
1589 <p>
1590 <tt>XXXInstrInfo.td</tt> also generates (in <tt>XXXGenDAGISel.inc</tt>) the
1591 <tt>SelectCode</tt> method that is used to call the appropriate processing
1592 method for an instruction. In this example, <tt>SelectCode</tt>
1593 calls <tt>Select_ISD_STORE</tt> for the <tt>ISD::STORE</tt> opcode.
1594 </p>
1595
1596 <div class="doc_code">
1597 <pre>
1598 SDNode *SelectCode(SDValue N) {
1599   ... 
1600   MVT::ValueType NVT = N.getNode()-&gt;getValueType(0);
1601   switch (N.getOpcode()) {
1602   case ISD::STORE: {
1603     switch (NVT) {
1604     default:
1605       return Select_ISD_STORE(N);
1606       break;
1607     }
1608     break;
1609   }
1610   ...
1611 </pre>
1612 </div>
1613
1614 <p>
1615 The pattern for <tt>STrr</tt> is matched, so elsewhere in
1616 <tt>XXXGenDAGISel.inc</tt>, code for <tt>STrr</tt> is created for
1617 <tt>Select_ISD_STORE</tt>. The <tt>Emit_22</tt> method is also generated
1618 in <tt>XXXGenDAGISel.inc</tt> to complete the processing of this
1619 instruction.
1620 </p>
1621
1622 <div class="doc_code">
1623 <pre>
1624 SDNode *Select_ISD_STORE(const SDValue &amp;N) {
1625   SDValue Chain = N.getOperand(0);
1626   if (Predicate_store(N.getNode())) {
1627     SDValue N1 = N.getOperand(1);
1628     SDValue N2 = N.getOperand(2);
1629     SDValue CPTmp0;
1630     SDValue CPTmp1;
1631
1632     // Pattern: (st:void IntRegs:i32:$src, 
1633     //           ADDRrr:i32:$addr)&lt;&lt;P:Predicate_store&gt;&gt;
1634     // Emits: (STrr:void ADDRrr:i32:$addr, IntRegs:i32:$src)
1635     // Pattern complexity = 13  cost = 1  size = 0
1636     if (SelectADDRrr(N, N2, CPTmp0, CPTmp1) &amp;&amp;
1637         N1.getNode()-&gt;getValueType(0) == MVT::i32 &amp;&amp;
1638         N2.getNode()-&gt;getValueType(0) == MVT::i32) {
1639       return Emit_22(N, SP::STrr, CPTmp0, CPTmp1);
1640     }
1641 ...
1642 </pre>
1643 </div>
1644
1645 </div>
1646
1647 <!-- ======================================================================= -->
1648 <div class="doc_subsection">
1649   <a name="LegalizePhase">The SelectionDAG Legalize Phase</a>
1650 </div>
1651
1652 <div class="doc_text">
1653
1654 <p>
1655 The Legalize phase converts a DAG to use types and operations that are natively
1656 supported by the target. For natively unsupported types and operations, you need
1657 to add code to the target-specific XXXTargetLowering implementation to convert
1658 unsupported types and operations to supported ones.
1659 </p>
1660
1661 <p>
1662 In the constructor for the <tt>XXXTargetLowering</tt> class, first use the
1663 <tt>addRegisterClass</tt> method to specify which types are supports and which
1664 register classes are associated with them. The code for the register classes are
1665 generated by TableGen from <tt>XXXRegisterInfo.td</tt> and placed
1666 in <tt>XXXGenRegisterInfo.h.inc</tt>. For example, the implementation of the
1667 constructor for the SparcTargetLowering class (in
1668 <tt>SparcISelLowering.cpp</tt>) starts with the following code:
1669 </p>
1670
1671 <div class="doc_code">
1672 <pre>
1673 addRegisterClass(MVT::i32, SP::IntRegsRegisterClass);
1674 addRegisterClass(MVT::f32, SP::FPRegsRegisterClass);
1675 addRegisterClass(MVT::f64, SP::DFPRegsRegisterClass); 
1676 </pre>
1677 </div>
1678
1679 <p>
1680 You should examine the node types in the <tt>ISD</tt> namespace
1681 (<tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt>) and determine which
1682 operations the target natively supports. For operations that do <b>not</b> have
1683 native support, add a callback to the constructor for the XXXTargetLowering
1684 class, so the instruction selection process knows what to do. The TargetLowering
1685 class callback methods (declared in <tt>llvm/Target/TargetLowering.h</tt>) are:
1686 </p>
1687
1688 <ul>
1689 <li><tt>setOperationAction</tt> &mdash; General operation.</li>
1690
1691 <li><tt>setLoadExtAction</tt> &mdash; Load with extension.</li>
1692
1693 <li><tt>setTruncStoreAction</tt> &mdash; Truncating store.</li>
1694
1695 <li><tt>setIndexedLoadAction</tt> &mdash; Indexed load.</li>
1696
1697 <li><tt>setIndexedStoreAction</tt> &mdash; Indexed store.</li>
1698
1699 <li><tt>setConvertAction</tt> &mdash; Type conversion.</li>
1700
1701 <li><tt>setCondCodeAction</tt> &mdash; Support for a given condition code.</li>
1702 </ul>
1703
1704 <p>
1705 Note: on older releases, <tt>setLoadXAction</tt> is used instead
1706 of <tt>setLoadExtAction</tt>.  Also, on older releases,
1707 <tt>setCondCodeAction</tt> may not be supported. Examine your release
1708 to see what methods are specifically supported.
1709 </p>
1710
1711 <p>
1712 These callbacks are used to determine that an operation does or does not work
1713 with a specified type (or types). And in all cases, the third parameter is
1714 a <tt>LegalAction</tt> type enum value: <tt>Promote</tt>, <tt>Expand</tt>,
1715 <tt>Custom</tt>, or <tt>Legal</tt>. <tt>SparcISelLowering.cpp</tt>
1716 contains examples of all four <tt>LegalAction</tt> values.
1717 </p>
1718
1719 </div>
1720
1721 <!-- _______________________________________________________________________ -->
1722 <div class="doc_subsubsection">
1723   <a name="promote">Promote</a>
1724 </div>
1725
1726 <div class="doc_text">
1727
1728 <p>
1729 For an operation without native support for a given type, the specified type may
1730 be promoted to a larger type that is supported. For example, SPARC does not
1731 support a sign-extending load for Boolean values (<tt>i1</tt> type), so
1732 in <tt>SparcISelLowering.cpp</tt> the third parameter below, <tt>Promote</tt>,
1733 changes <tt>i1</tt> type values to a large type before loading.
1734 </p>
1735
1736 <div class="doc_code">
1737 <pre>
1738 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
1739 </pre>
1740 </div>
1741
1742 </div>
1743
1744 <!-- _______________________________________________________________________ -->
1745 <div class="doc_subsubsection">
1746   <a name="expand">Expand</a>
1747 </div>
1748
1749 <div class="doc_text">
1750
1751 <p>
1752 For a type without native support, a value may need to be broken down further,
1753 rather than promoted. For an operation without native support, a combination of
1754 other operations may be used to similar effect. In SPARC, the floating-point
1755 sine and cosine trig operations are supported by expansion to other operations,
1756 as indicated by the third parameter, <tt>Expand</tt>, to
1757 <tt>setOperationAction</tt>:
1758 </p>
1759
1760 <div class="doc_code">
1761 <pre>
1762 setOperationAction(ISD::FSIN, MVT::f32, Expand);
1763 setOperationAction(ISD::FCOS, MVT::f32, Expand);
1764 </pre>
1765 </div>
1766
1767 </div>
1768
1769 <!-- _______________________________________________________________________ -->
1770 <div class="doc_subsubsection">
1771   <a name="custom">Custom</a>
1772 </div>
1773
1774 <div class="doc_text">
1775
1776 <p>
1777 For some operations, simple type promotion or operation expansion may be
1778 insufficient. In some cases, a special intrinsic function must be implemented.
1779 </p>
1780
1781 <p>
1782 For example, a constant value may require special treatment, or an operation may
1783 require spilling and restoring registers in the stack and working with register
1784 allocators.
1785 </p>
1786
1787 <p>
1788 As seen in <tt>SparcISelLowering.cpp</tt> code below, to perform a type
1789 conversion from a floating point value to a signed integer, first the
1790 <tt>setOperationAction</tt> should be called with <tt>Custom</tt> as the third
1791 parameter:
1792 </p>
1793
1794 <div class="doc_code">
1795 <pre>
1796 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
1797 </pre>
1798 </div>    
1799
1800 <p>
1801 In the <tt>LowerOperation</tt> method, for each <tt>Custom</tt> operation, a
1802 case statement should be added to indicate what function to call. In the
1803 following code, an <tt>FP_TO_SINT</tt> opcode will call
1804 the <tt>LowerFP_TO_SINT</tt> method:
1805 </p>
1806
1807 <div class="doc_code">
1808 <pre>
1809 SDValue SparcTargetLowering::LowerOperation(SDValue Op, SelectionDAG &amp;DAG) {
1810   switch (Op.getOpcode()) {
1811   case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
1812   ...
1813   }
1814 }
1815 </pre>
1816 </div>
1817
1818 <p>
1819 Finally, the <tt>LowerFP_TO_SINT</tt> method is implemented, using an FP
1820 register to convert the floating-point value to an integer.
1821 </p>
1822
1823 <div class="doc_code">
1824 <pre>
1825 static SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &amp;DAG) {
1826   assert(Op.getValueType() == MVT::i32);
1827   Op = DAG.getNode(SPISD::FTOI, MVT::f32, Op.getOperand(0));
1828   return DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
1829 }
1830 </pre>
1831 </div>    
1832
1833 </div>
1834
1835 <!-- _______________________________________________________________________ -->
1836 <div class="doc_subsubsection">
1837   <a name="legal">Legal</a>
1838 </div>
1839
1840 <div class="doc_text">
1841
1842 <p>
1843 The <tt>Legal</tt> LegalizeAction enum value simply indicates that an
1844 operation <b>is</b> natively supported. <tt>Legal</tt> represents the default
1845 condition, so it is rarely used. In <tt>SparcISelLowering.cpp</tt>, the action
1846 for <tt>CTPOP</tt> (an operation to count the bits set in an integer) is
1847 natively supported only for SPARC v9. The following code enables
1848 the <tt>Expand</tt> conversion technique for non-v9 SPARC implementations.
1849 </p>
1850
1851 <div class="doc_code">
1852 <pre>
1853 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
1854 ...
1855 if (TM.getSubtarget&lt;SparcSubtarget&gt;().isV9())
1856   setOperationAction(ISD::CTPOP, MVT::i32, Legal);
1857   case ISD::SETULT: return SPCC::ICC_CS;
1858   case ISD::SETULE: return SPCC::ICC_LEU;
1859   case ISD::SETUGT: return SPCC::ICC_GU;
1860   case ISD::SETUGE: return SPCC::ICC_CC;
1861   }
1862 }
1863 </pre>
1864 </div>
1865
1866 </div>
1867
1868 <!-- ======================================================================= -->
1869 <div class="doc_subsection">
1870   <a name="callingConventions">Calling Conventions</a>
1871 </div>
1872
1873 <div class="doc_text">
1874
1875 <p>
1876 To support target-specific calling conventions, <tt>XXXGenCallingConv.td</tt>
1877 uses interfaces (such as CCIfType and CCAssignToReg) that are defined in
1878 <tt>lib/Target/TargetCallingConv.td</tt>. TableGen can take the target
1879 descriptor file <tt>XXXGenCallingConv.td</tt> and generate the header
1880 file <tt>XXXGenCallingConv.inc</tt>, which is typically included
1881 in <tt>XXXISelLowering.cpp</tt>. You can use the interfaces in
1882 <tt>TargetCallingConv.td</tt> to specify:
1883 </p>
1884
1885 <ul>
1886 <li>The order of parameter allocation.</li>
1887
1888 <li>Where parameters and return values are placed (that is, on the stack or in
1889     registers).</li>
1890
1891 <li>Which registers may be used.</li>
1892
1893 <li>Whether the caller or callee unwinds the stack.</li>
1894 </ul>
1895
1896 <p>
1897 The following example demonstrates the use of the <tt>CCIfType</tt> and
1898 <tt>CCAssignToReg</tt> interfaces. If the <tt>CCIfType</tt> predicate is true
1899 (that is, if the current argument is of type <tt>f32</tt> or <tt>f64</tt>), then
1900 the action is performed. In this case, the <tt>CCAssignToReg</tt> action assigns
1901 the argument value to the first available register: either <tt>R0</tt>
1902 or <tt>R1</tt>.
1903 </p>
1904
1905 <div class="doc_code">
1906 <pre>
1907 CCIfType&lt;[f32,f64], CCAssignToReg&lt;[R0, R1]&gt;&gt;
1908 </pre>
1909 </div>
1910
1911 <p>
1912 <tt>SparcCallingConv.td</tt> contains definitions for a target-specific
1913 return-value calling convention (RetCC_Sparc32) and a basic 32-bit C calling
1914 convention (<tt>CC_Sparc32</tt>). The definition of <tt>RetCC_Sparc32</tt>
1915 (shown below) indicates which registers are used for specified scalar return
1916 types. A single-precision float is returned to register <tt>F0</tt>, and a
1917 double-precision float goes to register <tt>D0</tt>. A 32-bit integer is
1918 returned in register <tt>I0</tt> or <tt>I1</tt>.
1919 </p>
1920
1921 <div class="doc_code">
1922 <pre>
1923 def RetCC_Sparc32 : CallingConv&lt;[
1924   CCIfType&lt;[i32], CCAssignToReg&lt;[I0, I1]&gt;&gt;,
1925   CCIfType&lt;[f32], CCAssignToReg&lt;[F0]&gt;&gt;,
1926   CCIfType&lt;[f64], CCAssignToReg&lt;[D0]&gt;&gt;
1927 ]&gt;;
1928 </pre>
1929 </div>
1930
1931 <p>
1932 The definition of <tt>CC_Sparc32</tt> in <tt>SparcCallingConv.td</tt> introduces
1933 <tt>CCAssignToStack</tt>, which assigns the value to a stack slot with the
1934 specified size and alignment. In the example below, the first parameter, 4,
1935 indicates the size of the slot, and the second parameter, also 4, indicates the
1936 stack alignment along 4-byte units. (Special cases: if size is zero, then the
1937 ABI size is used; if alignment is zero, then the ABI alignment is used.)
1938 </p>
1939
1940 <div class="doc_code">
1941 <pre>
1942 def CC_Sparc32 : CallingConv&lt;[
1943   // All arguments get passed in integer registers if there is space.
1944   CCIfType&lt;[i32, f32, f64], CCAssignToReg&lt;[I0, I1, I2, I3, I4, I5]&gt;&gt;,
1945   CCAssignToStack&lt;4, 4&gt;
1946 ]&gt;;
1947 </pre>
1948 </div>
1949
1950 <p>
1951 <tt>CCDelegateTo</tt> is another commonly used interface, which tries to find a
1952 specified sub-calling convention, and, if a match is found, it is invoked. In
1953 the following example (in <tt>X86CallingConv.td</tt>), the definition of
1954 <tt>RetCC_X86_32_C</tt> ends with <tt>CCDelegateTo</tt>. After the current value
1955 is assigned to the register <tt>ST0</tt> or <tt>ST1</tt>,
1956 the <tt>RetCC_X86Common</tt> is invoked.
1957 </p>
1958
1959 <div class="doc_code">
1960 <pre>
1961 def RetCC_X86_32_C : CallingConv&lt;[
1962   CCIfType&lt;[f32], CCAssignToReg&lt;[ST0, ST1]&gt;&gt;,
1963   CCIfType&lt;[f64], CCAssignToReg&lt;[ST0, ST1]&gt;&gt;,
1964   CCDelegateTo&lt;RetCC_X86Common&gt;
1965 ]&gt;;
1966 </pre>
1967 </div>
1968
1969 <p>
1970 <tt>CCIfCC</tt> is an interface that attempts to match the given name to the
1971 current calling convention. If the name identifies the current calling
1972 convention, then a specified action is invoked. In the following example (in
1973 <tt>X86CallingConv.td</tt>), if the <tt>Fast</tt> calling convention is in use,
1974 then <tt>RetCC_X86_32_Fast</tt> is invoked. If the <tt>SSECall</tt> calling
1975 convention is in use, then <tt>RetCC_X86_32_SSE</tt> is invoked.
1976 </p>
1977
1978 <div class="doc_code">
1979 <pre>
1980 def RetCC_X86_32 : CallingConv&lt;[
1981   CCIfCC&lt;"CallingConv::Fast", CCDelegateTo&lt;RetCC_X86_32_Fast&gt;&gt;,
1982   CCIfCC&lt;"CallingConv::X86_SSECall", CCDelegateTo&lt;RetCC_X86_32_SSE&gt;&gt;,
1983   CCDelegateTo&lt;RetCC_X86_32_C&gt;
1984 ]&gt;;
1985 </pre>
1986 </div>
1987
1988 <p>Other calling convention interfaces include:</p>
1989
1990 <ul>
1991 <li><tt>CCIf &lt;predicate, action&gt;</tt> &mdash; If the predicate matches,
1992     apply the action.</li>
1993
1994 <li><tt>CCIfInReg &lt;action&gt;</tt> &mdash; If the argument is marked with the
1995     '<tt>inreg</tt>' attribute, then apply the action.</li>
1996
1997 <li><tt>CCIfNest &lt;action&gt;</tt> &mdash; Inf the argument is marked with the
1998     '<tt>nest</tt>' attribute, then apply the action.</li>
1999
2000 <li><tt>CCIfNotVarArg &lt;action&gt;</tt> &mdash; If the current function does
2001     not take a variable number of arguments, apply the action.</li>
2002
2003 <li><tt>CCAssignToRegWithShadow &lt;registerList, shadowList&gt;</tt> &mdash;
2004     similar to <tt>CCAssignToReg</tt>, but with a shadow list of registers.</li>
2005
2006 <li><tt>CCPassByVal &lt;size, align&gt;</tt> &mdash; Assign value to a stack
2007     slot with the minimum specified size and alignment.</li>
2008
2009 <li><tt>CCPromoteToType &lt;type&gt;</tt> &mdash; Promote the current value to
2010     the specified type.</li>
2011
2012 <li><tt>CallingConv &lt;[actions]&gt;</tt> &mdash; Define each calling
2013     convention that is supported.</li>
2014 </ul>
2015
2016 </div>
2017
2018 <!-- *********************************************************************** -->
2019 <div class="doc_section">
2020   <a name="assemblyPrinter">Assembly Printer</a>
2021 </div>
2022 <!-- *********************************************************************** -->
2023
2024 <div class="doc_text">
2025
2026 <p>
2027 During the code emission stage, the code generator may utilize an LLVM pass to
2028 produce assembly output. To do this, you want to implement the code for a
2029 printer that converts LLVM IR to a GAS-format assembly language for your target
2030 machine, using the following steps:
2031 </p>
2032
2033 <ul>
2034 <li>Define all the assembly strings for your target, adding them to the
2035     instructions defined in the <tt>XXXInstrInfo.td</tt> file.
2036     (See <a href="#InstructionSet">Instruction Set</a>.)  TableGen will produce
2037     an output file (<tt>XXXGenAsmWriter.inc</tt>) with an implementation of
2038     the <tt>printInstruction</tt> method for the XXXAsmPrinter class.</li>
2039
2040 <li>Write <tt>XXXTargetAsmInfo.h</tt>, which contains the bare-bones declaration
2041     of the <tt>XXXTargetAsmInfo</tt> class (a subclass
2042     of <tt>TargetAsmInfo</tt>).</li>
2043
2044 <li>Write <tt>XXXTargetAsmInfo.cpp</tt>, which contains target-specific values
2045     for <tt>TargetAsmInfo</tt> properties and sometimes new implementations for
2046     methods.</li>
2047
2048 <li>Write <tt>XXXAsmPrinter.cpp</tt>, which implements the <tt>AsmPrinter</tt>
2049     class that performs the LLVM-to-assembly conversion.</li>
2050 </ul>
2051
2052 <p>
2053 The code in <tt>XXXTargetAsmInfo.h</tt> is usually a trivial declaration of the
2054 <tt>XXXTargetAsmInfo</tt> class for use in <tt>XXXTargetAsmInfo.cpp</tt>.
2055 Similarly, <tt>XXXTargetAsmInfo.cpp</tt> usually has a few declarations of
2056 <tt>XXXTargetAsmInfo</tt> replacement values that override the default values
2057 in <tt>TargetAsmInfo.cpp</tt>. For example in <tt>SparcTargetAsmInfo.cpp</tt>:
2058 </p>
2059
2060 <div class="doc_code">
2061 <pre>
2062 SparcTargetAsmInfo::SparcTargetAsmInfo(const SparcTargetMachine &amp;TM) {
2063   Data16bitsDirective = "\t.half\t";
2064   Data32bitsDirective = "\t.word\t";
2065   Data64bitsDirective = 0;  // .xword is only supported by V9.
2066   ZeroDirective = "\t.skip\t";
2067   CommentString = "!";
2068   ConstantPoolSection = "\t.section \".rodata\",#alloc\n";
2069 }
2070 </pre>
2071 </div>
2072
2073 <p>
2074 The X86 assembly printer implementation (<tt>X86TargetAsmInfo</tt>) is an
2075 example where the target specific <tt>TargetAsmInfo</tt> class uses an 
2076 overridden methods: <tt>ExpandInlineAsm</tt>.
2077 </p>
2078
2079 <p>
2080 A target-specific implementation of AsmPrinter is written in
2081 <tt>XXXAsmPrinter.cpp</tt>, which implements the <tt>AsmPrinter</tt> class that
2082 converts the LLVM to printable assembly. The implementation must include the
2083 following headers that have declarations for the <tt>AsmPrinter</tt> and
2084 <tt>MachineFunctionPass</tt> classes. The <tt>MachineFunctionPass</tt> is a
2085 subclass of <tt>FunctionPass</tt>.
2086 </p>
2087
2088 <div class="doc_code">
2089 <pre>
2090 #include "llvm/CodeGen/AsmPrinter.h"
2091 #include "llvm/CodeGen/MachineFunctionPass.h" 
2092 </pre>
2093 </div>
2094
2095 <p>
2096 As a <tt>FunctionPass</tt>, <tt>AsmPrinter</tt> first
2097 calls <tt>doInitialization</tt> to set up the <tt>AsmPrinter</tt>. In
2098 <tt>SparcAsmPrinter</tt>, a <tt>Mangler</tt> object is instantiated to process
2099 variable names.
2100 </p>
2101
2102 <p>
2103 In <tt>XXXAsmPrinter.cpp</tt>, the <tt>runOnMachineFunction</tt> method
2104 (declared in <tt>MachineFunctionPass</tt>) must be implemented
2105 for <tt>XXXAsmPrinter</tt>. In <tt>MachineFunctionPass</tt>,
2106 the <tt>runOnFunction</tt> method invokes <tt>runOnMachineFunction</tt>.
2107 Target-specific implementations of <tt>runOnMachineFunction</tt> differ, but
2108 generally do the following to process each machine function:
2109 </p>
2110
2111 <ul>
2112 <li>Call <tt>SetupMachineFunction</tt> to perform initialization.</li>
2113
2114 <li>Call <tt>EmitConstantPool</tt> to print out (to the output stream) constants
2115     which have been spilled to memory.</li>
2116
2117 <li>Call <tt>EmitJumpTableInfo</tt> to print out jump tables used by the current
2118     function.</li>
2119
2120 <li>Print out the label for the current function.</li>
2121
2122 <li>Print out the code for the function, including basic block labels and the
2123     assembly for the instruction (using <tt>printInstruction</tt>)</li>
2124 </ul>
2125
2126 <p>
2127 The <tt>XXXAsmPrinter</tt> implementation must also include the code generated
2128 by TableGen that is output in the <tt>XXXGenAsmWriter.inc</tt> file. The code
2129 in <tt>XXXGenAsmWriter.inc</tt> contains an implementation of the
2130 <tt>printInstruction</tt> method that may call these methods:
2131 </p>
2132
2133 <ul>
2134 <li><tt>printOperand</tt></li>
2135
2136 <li><tt>printMemOperand</tt></li>
2137
2138 <li><tt>printCCOperand (for conditional statements)</tt></li>
2139
2140 <li><tt>printDataDirective</tt></li>
2141
2142 <li><tt>printDeclare</tt></li>
2143
2144 <li><tt>printImplicitDef</tt></li>
2145
2146 <li><tt>printInlineAsm</tt></li>
2147 </ul>
2148
2149 <p>
2150 The implementations of <tt>printDeclare</tt>, <tt>printImplicitDef</tt>,
2151 <tt>printInlineAsm</tt>, and <tt>printLabel</tt> in <tt>AsmPrinter.cpp</tt> are
2152 generally adequate for printing assembly and do not need to be
2153 overridden.
2154 </p>
2155
2156 <p>
2157 The <tt>printOperand</tt> method is implemented with a long switch/case
2158 statement for the type of operand: register, immediate, basic block, external
2159 symbol, global address, constant pool index, or jump table index. For an
2160 instruction with a memory address operand, the <tt>printMemOperand</tt> method
2161 should be implemented to generate the proper output. Similarly,
2162 <tt>printCCOperand</tt> should be used to print a conditional operand.
2163 </p>
2164
2165 <p><tt>doFinalization</tt> should be overridden in <tt>XXXAsmPrinter</tt>, and
2166 it should be called to shut down the assembly printer. During
2167 <tt>doFinalization</tt>, global variables and constants are printed to
2168 output.
2169 </p>
2170
2171 </div>
2172
2173 <!-- *********************************************************************** -->
2174 <div class="doc_section">
2175   <a name="subtargetSupport">Subtarget Support</a>
2176 </div>
2177 <!-- *********************************************************************** -->
2178
2179 <div class="doc_text">
2180
2181 <p>
2182 Subtarget support is used to inform the code generation process of instruction
2183 set variations for a given chip set.  For example, the LLVM SPARC implementation
2184 provided covers three major versions of the SPARC microprocessor architecture:
2185 Version 8 (V8, which is a 32-bit architecture), Version 9 (V9, a 64-bit
2186 architecture), and the UltraSPARC architecture. V8 has 16 double-precision
2187 floating-point registers that are also usable as either 32 single-precision or 8
2188 quad-precision registers.  V8 is also purely big-endian. V9 has 32
2189 double-precision floating-point registers that are also usable as 16
2190 quad-precision registers, but cannot be used as single-precision registers. The
2191 UltraSPARC architecture combines V9 with UltraSPARC Visual Instruction Set
2192 extensions.
2193 </p>
2194
2195 <p>
2196 If subtarget support is needed, you should implement a target-specific
2197 XXXSubtarget class for your architecture. This class should process the
2198 command-line options <tt>-mcpu=</tt> and <tt>-mattr=</tt>.
2199 </p>
2200
2201 <p>
2202 TableGen uses definitions in the <tt>Target.td</tt> and <tt>Sparc.td</tt> files
2203 to generate code in <tt>SparcGenSubtarget.inc</tt>. In <tt>Target.td</tt>, shown
2204 below, the <tt>SubtargetFeature</tt> interface is defined. The first 4 string
2205 parameters of the <tt>SubtargetFeature</tt> interface are a feature name, an
2206 attribute set by the feature, the value of the attribute, and a description of
2207 the feature. (The fifth parameter is a list of features whose presence is
2208 implied, and its default value is an empty array.)
2209 </p>
2210
2211 <div class="doc_code">
2212 <pre>
2213 class SubtargetFeature&lt;string n, string a,  string v, string d,
2214                        list&lt;SubtargetFeature&gt; i = []&gt; {
2215   string Name = n;
2216   string Attribute = a;
2217   string Value = v;
2218   string Desc = d;
2219   list&lt;SubtargetFeature&gt; Implies = i;
2220 }
2221 </pre>
2222 </div>
2223
2224 <p>
2225 In the <tt>Sparc.td</tt> file, the SubtargetFeature is used to define the
2226 following features.
2227 </p>
2228
2229 <div class="doc_code">
2230 <pre>
2231 def FeatureV9 : SubtargetFeature&lt;"v9", "IsV9", "true",
2232                      "Enable SPARC-V9 instructions"&gt;;
2233 def FeatureV8Deprecated : SubtargetFeature&lt;"deprecated-v8", 
2234                      "V8DeprecatedInsts", "true",
2235                      "Enable deprecated V8 instructions in V9 mode"&gt;;
2236 def FeatureVIS : SubtargetFeature&lt;"vis", "IsVIS", "true",
2237                      "Enable UltraSPARC Visual Instruction Set extensions"&gt;;
2238 </pre>
2239 </div>
2240
2241 <p>
2242 Elsewhere in <tt>Sparc.td</tt>, the Proc class is defined and then is used to
2243 define particular SPARC processor subtypes that may have the previously
2244 described features.
2245 </p>
2246
2247 <div class="doc_code">
2248 <pre>
2249 class Proc&lt;string Name, list&lt;SubtargetFeature&gt; Features&gt;
2250   : Processor&lt;Name, NoItineraries, Features&gt;;
2251 &nbsp;
2252 def : Proc&lt;"generic",         []&gt;;
2253 def : Proc&lt;"v8",              []&gt;;
2254 def : Proc&lt;"supersparc",      []&gt;;
2255 def : Proc&lt;"sparclite",       []&gt;;
2256 def : Proc&lt;"f934",            []&gt;;
2257 def : Proc&lt;"hypersparc",      []&gt;;
2258 def : Proc&lt;"sparclite86x",    []&gt;;
2259 def : Proc&lt;"sparclet",        []&gt;;
2260 def : Proc&lt;"tsc701",          []&gt;;
2261 def : Proc&lt;"v9",              [FeatureV9]&gt;;
2262 def : Proc&lt;"ultrasparc",      [FeatureV9, FeatureV8Deprecated]&gt;;
2263 def : Proc&lt;"ultrasparc3",     [FeatureV9, FeatureV8Deprecated]&gt;;
2264 def : Proc&lt;"ultrasparc3-vis", [FeatureV9, FeatureV8Deprecated, FeatureVIS]&gt;;
2265 </pre>
2266 </div>
2267
2268 <p>
2269 From <tt>Target.td</tt> and <tt>Sparc.td</tt> files, the resulting
2270 SparcGenSubtarget.inc specifies enum values to identify the features, arrays of
2271 constants to represent the CPU features and CPU subtypes, and the
2272 ParseSubtargetFeatures method that parses the features string that sets
2273 specified subtarget options. The generated <tt>SparcGenSubtarget.inc</tt> file
2274 should be included in the <tt>SparcSubtarget.cpp</tt>. The target-specific
2275 implementation of the XXXSubtarget method should follow this pseudocode:
2276 </p>
2277
2278 <div class="doc_code">
2279 <pre>
2280 XXXSubtarget::XXXSubtarget(const Module &amp;M, const std::string &amp;FS) {
2281   // Set the default features
2282   // Determine default and user specified characteristics of the CPU
2283   // Call ParseSubtargetFeatures(FS, CPU) to parse the features string
2284   // Perform any additional operations
2285 }
2286 </pre>
2287 </div>
2288
2289 </div>
2290
2291 <!-- *********************************************************************** -->
2292 <div class="doc_section">
2293   <a name="jitSupport">JIT Support</a>
2294 </div>
2295 <!-- *********************************************************************** -->
2296
2297 <div class="doc_text">
2298
2299 <p>
2300 The implementation of a target machine optionally includes a Just-In-Time (JIT)
2301 code generator that emits machine code and auxiliary structures as binary output
2302 that can be written directly to memory.  To do this, implement JIT code
2303 generation by performing the following steps:
2304 </p>
2305
2306 <ul>
2307 <li>Write an <tt>XXXCodeEmitter.cpp</tt> file that contains a machine function
2308     pass that transforms target-machine instructions into relocatable machine
2309     code.</li>
2310
2311 <li>Write an <tt>XXXJITInfo.cpp</tt> file that implements the JIT interfaces for
2312     target-specific code-generation activities, such as emitting machine code
2313     and stubs.</li>
2314
2315 <li>Modify <tt>XXXTargetMachine</tt> so that it provides a
2316     <tt>TargetJITInfo</tt> object through its <tt>getJITInfo</tt> method.</li>
2317 </ul>
2318
2319 <p>
2320 There are several different approaches to writing the JIT support code. For
2321 instance, TableGen and target descriptor files may be used for creating a JIT
2322 code generator, but are not mandatory. For the Alpha and PowerPC target
2323 machines, TableGen is used to generate <tt>XXXGenCodeEmitter.inc</tt>, which
2324 contains the binary coding of machine instructions and the
2325 <tt>getBinaryCodeForInstr</tt> method to access those codes. Other JIT
2326 implementations do not.
2327 </p>
2328
2329 <p>
2330 Both <tt>XXXJITInfo.cpp</tt> and <tt>XXXCodeEmitter.cpp</tt> must include the
2331 <tt>llvm/CodeGen/MachineCodeEmitter.h</tt> header file that defines the
2332 <tt>MachineCodeEmitter</tt> class containing code for several callback functions
2333 that write data (in bytes, words, strings, etc.) to the output stream.
2334 </p>
2335
2336 </div>
2337
2338 <!-- ======================================================================= -->
2339 <div class="doc_subsection">
2340   <a name="mce">Machine Code Emitter</a>
2341 </div>
2342
2343 <div class="doc_text">
2344
2345 <p>
2346 In <tt>XXXCodeEmitter.cpp</tt>, a target-specific of the <tt>Emitter</tt> class
2347 is implemented as a function pass (subclass
2348 of <tt>MachineFunctionPass</tt>). The target-specific implementation
2349 of <tt>runOnMachineFunction</tt> (invoked by
2350 <tt>runOnFunction</tt> in <tt>MachineFunctionPass</tt>) iterates through the
2351 <tt>MachineBasicBlock</tt> calls <tt>emitInstruction</tt> to process each
2352 instruction and emit binary code. <tt>emitInstruction</tt> is largely
2353 implemented with case statements on the instruction types defined in
2354 <tt>XXXInstrInfo.h</tt>. For example, in <tt>X86CodeEmitter.cpp</tt>,
2355 the <tt>emitInstruction</tt> method is built around the following switch/case
2356 statements:
2357 </p>
2358
2359 <div class="doc_code">
2360 <pre>
2361 switch (Desc-&gt;TSFlags &amp; X86::FormMask) {
2362 case X86II::Pseudo:  // for not yet implemented instructions 
2363    ...               // or pseudo-instructions
2364    break;
2365 case X86II::RawFrm:  // for instructions with a fixed opcode value
2366    ...
2367    break;
2368 case X86II::AddRegFrm: // for instructions that have one register operand 
2369    ...                 // added to their opcode
2370    break;
2371 case X86II::MRMDestReg:// for instructions that use the Mod/RM byte
2372    ...                 // to specify a destination (register)
2373    break;
2374 case X86II::MRMDestMem:// for instructions that use the Mod/RM byte
2375    ...                 // to specify a destination (memory)
2376    break;
2377 case X86II::MRMSrcReg: // for instructions that use the Mod/RM byte
2378    ...                 // to specify a source (register)
2379    break;
2380 case X86II::MRMSrcMem: // for instructions that use the Mod/RM byte
2381    ...                 // to specify a source (memory)
2382    break;
2383 case X86II::MRM0r: case X86II::MRM1r:  // for instructions that operate on 
2384 case X86II::MRM2r: case X86II::MRM3r:  // a REGISTER r/m operand and
2385 case X86II::MRM4r: case X86II::MRM5r:  // use the Mod/RM byte and a field
2386 case X86II::MRM6r: case X86II::MRM7r:  // to hold extended opcode data
2387    ...  
2388    break;
2389 case X86II::MRM0m: case X86II::MRM1m:  // for instructions that operate on
2390 case X86II::MRM2m: case X86II::MRM3m:  // a MEMORY r/m operand and
2391 case X86II::MRM4m: case X86II::MRM5m:  // use the Mod/RM byte and a field
2392 case X86II::MRM6m: case X86II::MRM7m:  // to hold extended opcode data
2393    ...  
2394    break;
2395 case X86II::MRMInitReg: // for instructions whose source and
2396    ...                  // destination are the same register
2397    break;
2398 }
2399 </pre>
2400 </div>
2401
2402 <p>
2403 The implementations of these case statements often first emit the opcode and
2404 then get the operand(s). Then depending upon the operand, helper methods may be
2405 called to process the operand(s). For example, in <tt>X86CodeEmitter.cpp</tt>,
2406 for the <tt>X86II::AddRegFrm</tt> case, the first data emitted
2407 (by <tt>emitByte</tt>) is the opcode added to the register operand. Then an
2408 object representing the machine operand, <tt>MO1</tt>, is extracted. The helper
2409 methods such as <tt>isImmediate</tt>,
2410 <tt>isGlobalAddress</tt>, <tt>isExternalSymbol</tt>, <tt>isConstantPoolIndex</tt>, and 
2411 <tt>isJumpTableIndex</tt> determine the operand
2412 type. (<tt>X86CodeEmitter.cpp</tt> also has private methods such
2413 as <tt>emitConstant</tt>, <tt>emitGlobalAddress</tt>,
2414 <tt>emitExternalSymbolAddress</tt>, <tt>emitConstPoolAddress</tt>,
2415 and <tt>emitJumpTableAddress</tt> that emit the data into the output stream.)
2416 </p>
2417
2418 <div class="doc_code">
2419 <pre>
2420 case X86II::AddRegFrm:
2421   MCE.emitByte(BaseOpcode + getX86RegNum(MI.getOperand(CurOp++).getReg()));
2422   
2423   if (CurOp != NumOps) {
2424     const MachineOperand &amp;MO1 = MI.getOperand(CurOp++);
2425     unsigned Size = X86InstrInfo::sizeOfImm(Desc);
2426     if (MO1.isImmediate())
2427       emitConstant(MO1.getImm(), Size);
2428     else {
2429       unsigned rt = Is64BitMode ? X86::reloc_pcrel_word
2430         : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word);
2431       if (Opcode == X86::MOV64ri) 
2432         rt = X86::reloc_absolute_dword;  // FIXME: add X86II flag?
2433       if (MO1.isGlobalAddress()) {
2434         bool NeedStub = isa&lt;Function&gt;(MO1.getGlobal());
2435         bool isLazy = gvNeedsLazyPtr(MO1.getGlobal());
2436         emitGlobalAddress(MO1.getGlobal(), rt, MO1.getOffset(), 0,
2437                           NeedStub, isLazy);
2438       } else if (MO1.isExternalSymbol())
2439         emitExternalSymbolAddress(MO1.getSymbolName(), rt);
2440       else if (MO1.isConstantPoolIndex())
2441         emitConstPoolAddress(MO1.getIndex(), rt);
2442       else if (MO1.isJumpTableIndex())
2443         emitJumpTableAddress(MO1.getIndex(), rt);
2444     }
2445   }
2446   break;
2447 </pre>
2448 </div>
2449
2450 <p>
2451 In the previous example, <tt>XXXCodeEmitter.cpp</tt> uses the
2452 variable <tt>rt</tt>, which is a RelocationType enum that may be used to
2453 relocate addresses (for example, a global address with a PIC base offset). The
2454 <tt>RelocationType</tt> enum for that target is defined in the short
2455 target-specific <tt>XXXRelocations.h</tt> file. The <tt>RelocationType</tt> is used by
2456 the <tt>relocate</tt> method defined in <tt>XXXJITInfo.cpp</tt> to rewrite
2457 addresses for referenced global symbols.
2458 </p>
2459
2460 <p>
2461 For example, <tt>X86Relocations.h</tt> specifies the following relocation types
2462 for the X86 addresses. In all four cases, the relocated value is added to the
2463 value already in memory. For <tt>reloc_pcrel_word</tt>
2464 and <tt>reloc_picrel_word</tt>, there is an additional initial adjustment.
2465 </p>
2466
2467 <div class="doc_code">
2468 <pre>
2469 enum RelocationType {
2470   reloc_pcrel_word = 0,    // add reloc value after adjusting for the PC loc
2471   reloc_picrel_word = 1,   // add reloc value after adjusting for the PIC base
2472   reloc_absolute_word = 2, // absolute relocation; no additional adjustment 
2473   reloc_absolute_dword = 3 // absolute relocation; no additional adjustment
2474 };
2475 </pre>
2476 </div>
2477
2478 </div>
2479
2480 <!-- ======================================================================= -->
2481 <div class="doc_subsection">
2482   <a name="targetJITInfo">Target JIT Info</a>
2483 </div>
2484
2485 <div class="doc_text">
2486
2487 <p>
2488 <tt>XXXJITInfo.cpp</tt> implements the JIT interfaces for target-specific
2489 code-generation activities, such as emitting machine code and stubs. At minimum,
2490 a target-specific version of <tt>XXXJITInfo</tt> implements the following:
2491 </p>
2492
2493 <ul>
2494 <li><tt>getLazyResolverFunction</tt> &mdash; Initializes the JIT, gives the
2495     target a function that is used for compilation.</li>
2496
2497 <li><tt>emitFunctionStub</tt> &mdash; Returns a native function with a specified
2498     address for a callback function.</li>
2499
2500 <li><tt>relocate</tt> &mdash; Changes the addresses of referenced globals, based
2501     on relocation types.</li>
2502
2503 <li>Callback function that are wrappers to a function stub that is used when the
2504     real target is not initially known.</li>
2505 </ul>
2506
2507 <p>
2508 <tt>getLazyResolverFunction</tt> is generally trivial to implement. It makes the
2509 incoming parameter as the global <tt>JITCompilerFunction</tt> and returns the
2510 callback function that will be used a function wrapper. For the Alpha target
2511 (in <tt>AlphaJITInfo.cpp</tt>), the <tt>getLazyResolverFunction</tt>
2512 implementation is simply:
2513 </p>
2514
2515 <div class="doc_code">
2516 <pre>
2517 TargetJITInfo::LazyResolverFn AlphaJITInfo::getLazyResolverFunction(  
2518                                             JITCompilerFn F) {
2519   JITCompilerFunction = F;
2520   return AlphaCompilationCallback;
2521 }
2522 </pre>
2523 </div>
2524
2525 <p>
2526 For the X86 target, the <tt>getLazyResolverFunction</tt> implementation is a
2527 little more complication, because it returns a different callback function for
2528 processors with SSE instructions and XMM registers.
2529 </p>
2530
2531 <p>
2532 The callback function initially saves and later restores the callee register
2533 values, incoming arguments, and frame and return address. The callback function
2534 needs low-level access to the registers or stack, so it is typically implemented
2535 with assembler.
2536 </p>
2537
2538 </div>
2539
2540 <!-- *********************************************************************** -->
2541
2542 <hr>
2543 <address>
2544   <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
2545   src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
2546   <a href="http://validator.w3.org/check/referer"><img
2547   src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
2548
2549   <a href="http://www.woo.com">Mason Woo</a> and <a href="http://misha.brukman.net">Misha Brukman</a><br>
2550   <a href="http://llvm.org">The LLVM Compiler Infrastructure</a>
2551   <br>
2552   Last modified: $Date$
2553 </address>
2554
2555 </body>
2556 </html>