1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <title>Extending LLVM: Adding instructions, intrinsics, types, etc.</title>
6 <link rel="stylesheet" href="llvm.css" type="text/css">
11 <div class="doc_title">
12 Extending LLVM: Adding instructions, intrinsics, types, etc.
16 <li><a href="#introduction">Introduction and Warning</a></li>
17 <li><a href="#intrinsic">Adding a new intrinsic function</a></li>
18 <li><a href="#instruction">Adding a new instruction</a></li>
19 <li><a href="#type">Adding a new type</a>
21 <li><a href="#fund_type">Adding a new fundamental type</a></li>
22 <li><a href="#derived_type">Adding a new derived type</a></li>
26 <div class="doc_author">
27 <p>Written by <a href="http://misha.brukman.net">Misha Brukman</a>,
28 Brad Jones, and <a href="http://nondot.org/sabre">Chris Lattner</a></p>
31 <!-- *********************************************************************** -->
32 <div class="doc_section">
33 <a name="introduction">Introduction and Warning</a>
35 <!-- *********************************************************************** -->
37 <div class="doc_text">
39 <p>During the course of using LLVM, you may wish to customize it for your
40 research project or for experimentation. At this point, you may realize that
41 you need to add something to LLVM, whether it be a new fundamental type, a new
42 intrinsic function, or a whole new instruction.</p>
44 <p>When you come to this realization, stop and think. Do you really need to
45 extend LLVM? Is it a new fundamental capability that LLVM does not support at
46 its current incarnation or can it be synthesized from already pre-existing LLVM
47 elements? If you are not sure, ask on the <a
48 href="http://mail.cs.uiuc.edu/mailman/listinfo/llvmdev">LLVM-dev</a> list. The
49 reason is that extending LLVM will get involved as you need to update all the
50 different passes that you intend to use with your extension, and there are
51 <em>many</em> LLVM analyses and transformations, so it may be quite a bit of
54 <p>Adding an <a href="#intrinsic">intrinsic function</a> is easier than adding
55 an instruction, and is transparent to optimization passes which treat it as an
56 unanalyzable function. If your added functionality can be expressed as a
57 function call, an intrinsic function is the method of choice for LLVM
60 <p>Before you invest a significant amount of effort into a non-trivial
61 extension, <span class="doc_warning">ask on the list</span> if what you are
62 looking to do can be done with already-existing infrastructure, or if maybe
63 someone else is already working on it. You will save yourself a lot of time and
64 effort by doing so.</p>
68 <!-- *********************************************************************** -->
69 <div class="doc_section">
70 <a name="intrinsic">Adding a new intrinsic function</a>
72 <!-- *********************************************************************** -->
74 <div class="doc_text">
76 <p>Adding a new intrinsic function to LLVM is much easier than adding a new
77 instruction. Almost all extensions to LLVM should start as an intrinsic
78 function and then be turned into an instruction if warranted.</p>
81 <li><tt>llvm/docs/LangRef.html</tt>:
82 Document the intrinsic. Decide whether it is code generator specific and
83 what the restrictions are. Talk to other people about it so that you are
84 sure it's a good idea.</li>
86 <li><tt>llvm/include/llvm/Intrinsics.h</tt>:
87 add an enum in the <tt>llvm::Intrinsic</tt> namespace</li>
89 <li><tt>llvm/lib/VMCore/Verifier.cpp</tt>:
90 Add code to check the invariants of the intrinsic are respected.</li>
92 <li><tt>llvm/lib/VMCore/Function.cpp (<tt>Function::getIntrinsicID()</tt>)</tt>:
93 Identify the new intrinsic function, returning the enum for the intrinsic
96 <li><tt>llvm/lib/Analysis/BasicAliasAnalysis.cpp</tt>: If the new intrinsic does
97 not access memory or does not write to memory, add it to the relevant list
100 <li><tt>llvm/lib/Transforms/Utils/Local.cpp</tt>: If it is possible to constant
101 fold your intrinsic, add support to it in the <tt>canConstantFoldCallTo</tt> and
102 <tt>ConstantFoldCall</tt> functions.</li>
104 <li>Test your intrinsic</li>
106 <li><tt>llvm/test/Regression/*</tt>: add your test cases to the test suite</li>
109 <p>Once the intrinsic has been added to the system, you must add code generator
110 support for it. Generally you must do the following steps:</p>
113 <dt>Add support to the C backend in <tt>lib/Target/CBackend/</tt></dt>
115 <dd>Depending on the intrinsic, there are a few ways to implement this. First,
116 if it makes sense to lower the intrinsic to an expanded sequence of C code in
117 all cases, just emit the expansion in <tt>visitCallInst</tt>. Second, if the
118 intrinsic has some way to express it with GCC (or any other compiler)
119 extensions, it can be conditionally supported based on the compiler compiling
120 the CBE output (see llvm.prefetch for an example). Third, if the intrinsic
121 really has no way to be lowered, just have the code generator emit code that
122 prints an error message and calls abort if executed.
125 <dt>Add a enum value for the SelectionDAG node in
126 <tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt></dt>
128 <dd>Also, add code to <tt>lib/CodeGen/SelectionDAG/SelectionDAG.cpp</tt> (and
129 <tt>SelectionDAGPrinter.cpp</tt>) to print the node.</dd>
131 <dt>Add code to <tt>SelectionDAG/SelectionDAGISel.cpp</tt> to recognize the
134 <dd>Presumably the intrinsic should be recognized and turned into the node you
137 <dt>Add code to <tt>SelectionDAG/LegalizeDAG.cpp</tt> to <a
138 href="CodeGenerator.html#selectiondag_legalize">legalize, promote, and
139 expand</a> the node as necessary.</dt>
141 <dd>If the intrinsic can be expanded to primitive operations, legalize can break
142 the node down into other elementary operations that are be supported.</dd>
144 <dt>Add target-specific support to specific code generators.</dt>
146 <dd>Extend the code generators you are interested in to recognize and support
147 the node, emitting the code you want.</dd>
151 Unfortunately, the process of extending the code generator to support a new node
152 is not extremely well documented. As such, it is often helpful to look at other
153 intrinsics (e.g. <tt>llvm.ctpop</tt>) to see how they are recognized and turned
154 into a node by <tt>SelectionDAGISel.cpp</tt>, legalized by
155 <tt>LegalizeDAG.cpp</tt>, then finally emitted by the various code generators.
160 <!-- *********************************************************************** -->
161 <div class="doc_section">
162 <a name="instruction">Adding a new instruction</a>
164 <!-- *********************************************************************** -->
166 <div class="doc_text">
168 <p><span class="doc_warning">WARNING: adding instructions changes the bytecode
169 format, and it will take some effort to maintain compatibility with
170 the previous version.</span> Only add an instruction if it is absolutely
175 <li><tt>llvm/include/llvm/Instruction.def</tt>:
176 add a number for your instruction and an enum name</li>
178 <li><tt>llvm/include/llvm/Instructions.h</tt>:
179 add a definition for the class that will represent your instruction</li>
181 <li><tt>llvm/include/llvm/Support/InstVisitor.h</tt>:
182 add a prototype for a visitor to your new instruction type</li>
184 <li><tt>llvm/lib/AsmParser/Lexer.l</tt>:
185 add a new token to parse your instruction from assembly text file</li>
187 <li><tt>llvm/lib/AsmParser/llvmAsmParser.y</tt>:
188 add the grammar on how your instruction can be read and what it will
189 construct as a result</li>
191 <li><tt>llvm/lib/Bytecode/Reader/Reader.cpp</tt>:
192 add a case for your instruction and how it will be parsed from bytecode</li>
194 <li><tt>llvm/lib/VMCore/Instruction.cpp</tt>:
195 add a case for how your instruction will be printed out to assembly</li>
197 <li><tt>llvm/lib/VMCore/Instructions.cpp</tt>:
198 implement the class you defined in
199 <tt>llvm/include/llvm/Instructions.h</tt></li>
201 <li>Test your instruction</li>
203 <li><tt>llvm/lib/Target/*</tt>:
204 Add support for your instruction to code generators, or add a lowering
207 <li><tt>llvm/test/Regression/*</tt>: add your test cases to the test suite.</li>
211 <p>Also, you need to implement (or modify) any analyses or passes that you want
212 to understand this new instruction.</p>
217 <!-- *********************************************************************** -->
218 <div class="doc_section">
219 <a name="type">Adding a new type</a>
221 <!-- *********************************************************************** -->
223 <div class="doc_text">
225 <p><span class="doc_warning">WARNING: adding new types changes the bytecode
226 format, and will break compatibility with currently-existing LLVM
227 installations.</span> Only add new types if it is absolutely necessary.</p>
231 <!-- ======================================================================= -->
232 <div class="doc_subsection">
233 <a name="fund_type">Adding a fundamental type</a>
236 <div class="doc_text">
240 <li><tt>llvm/include/llvm/Type.h</tt>:
241 add enum for the new type; add static <tt>Type*</tt> for this type</li>
243 <li><tt>llvm/lib/VMCore/Type.cpp</tt>:
244 add mapping from <tt>TypeID</tt> => <tt>Type*</tt>;
245 initialize the static <tt>Type*</tt></li>
247 <li><tt>llvm/lib/AsmReader/Lexer.l</tt>:
248 add ability to parse in the type from text assembly</li>
250 <li><tt>llvm/lib/AsmReader/llvmAsmParser.y</tt>:
251 add a token for that type</li>
257 <!-- ======================================================================= -->
258 <div class="doc_subsection">
259 <a name="derived_type">Adding a derived type</a>
262 <div class="doc_text">
265 <li><tt>llvm/include/llvm/Type.h</tt>:
266 add enum for the new type; add a forward declaration of the type
269 <li><tt>llvm/include/llvm/DerivedTypes.h</tt>:
270 add new class to represent new class in the hierarchy; add forward
271 declaration to the TypeMap value type</li>
273 <li><tt>llvm/lib/VMCore/Type.cpp</tt>:
274 add support for derived type to:
275 <div class="doc_code">
277 std::string getTypeDescription(const Type &Ty,
278 std::vector<const Type*> &TypeStack)
279 bool TypesEqual(const Type *Ty, const Type *Ty2,
280 std::map<const Type*, const Type*> & EqTypes)
283 add necessary member functions for type, and factory methods</li>
285 <li><tt>llvm/lib/AsmReader/Lexer.l</tt>:
286 add ability to parse in the type from text assembly</li>
288 <li><tt>llvm/lib/ByteCode/Writer/Writer.cpp</tt>:
289 modify <tt>void BytecodeWriter::outputType(const Type *T)</tt> to serialize
292 <li><tt>llvm/lib/ByteCode/Reader/Reader.cpp</tt>:
293 modify <tt>const Type *BytecodeReader::ParseType()</tt> to read your data
296 <li><tt>llvm/lib/VMCore/AsmWriter.cpp</tt>:
298 <div class="doc_code">
300 void calcTypeName(const Type *Ty,
301 std::vector<const Type*> &TypeStack,
302 std::map<const Type*,std::string> &TypeNames,
303 std::string & Result)
306 to output the new derived type
314 <!-- *********************************************************************** -->
318 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
319 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
320 <a href="http://validator.w3.org/check/referer"><img
321 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!" /></a>
323 <a href="http://llvm.cs.uiuc.edu">The LLVM Compiler Infrastructure</a>
325 Last modified: $Date$