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="#sdnode">Adding a new SelectionDAG node</a></li>
20 <li><a href="#type">Adding a new type</a>
22 <li><a href="#fund_type">Adding a new fundamental type</a></li>
23 <li><a href="#derived_type">Adding a new derived type</a></li>
27 <div class="doc_author">
28 <p>Written by <a href="http://misha.brukman.net">Misha Brukman</a>,
29 Brad Jones, Nate Begeman,
30 and <a href="http://nondot.org/sabre">Chris Lattner</a></p>
33 <!-- *********************************************************************** -->
34 <div class="doc_section">
35 <a name="introduction">Introduction and Warning</a>
37 <!-- *********************************************************************** -->
39 <div class="doc_text">
41 <p>During the course of using LLVM, you may wish to customize it for your
42 research project or for experimentation. At this point, you may realize that
43 you need to add something to LLVM, whether it be a new fundamental type, a new
44 intrinsic function, or a whole new instruction.</p>
46 <p>When you come to this realization, stop and think. Do you really need to
47 extend LLVM? Is it a new fundamental capability that LLVM does not support at
48 its current incarnation or can it be synthesized from already pre-existing LLVM
49 elements? If you are not sure, ask on the <a
50 href="http://mail.cs.uiuc.edu/mailman/listinfo/llvmdev">LLVM-dev</a> list. The
51 reason is that extending LLVM will get involved as you need to update all the
52 different passes that you intend to use with your extension, and there are
53 <em>many</em> LLVM analyses and transformations, so it may be quite a bit of
56 <p>Adding an <a href="#intrinsic">intrinsic function</a> is far easier than
57 adding an instruction, and is transparent to optimization passes. If your added
58 functionality can be expressed as a
59 function call, an intrinsic function is the method of choice for LLVM
62 <p>Before you invest a significant amount of effort into a non-trivial
63 extension, <span class="doc_warning">ask on the list</span> if what you are
64 looking to do can be done with already-existing infrastructure, or if maybe
65 someone else is already working on it. You will save yourself a lot of time and
66 effort by doing so.</p>
70 <!-- *********************************************************************** -->
71 <div class="doc_section">
72 <a name="intrinsic">Adding a new intrinsic function</a>
74 <!-- *********************************************************************** -->
76 <div class="doc_text">
78 <p>Adding a new intrinsic function to LLVM is much easier than adding a new
79 instruction. Almost all extensions to LLVM should start as an intrinsic
80 function and then be turned into an instruction if warranted.</p>
83 <li><tt>llvm/docs/LangRef.html</tt>:
84 Document the intrinsic. Decide whether it is code generator specific and
85 what the restrictions are. Talk to other people about it so that you are
86 sure it's a good idea.</li>
88 <li><tt>llvm/include/llvm/Intrinsics*.td</tt>:
89 Add an entry for your intrinsic. Describe its memory access characteristics
90 for optimization (this controls whether it will be DCE'd, CSE'd, etc).</li>
92 <li><tt>llvm/lib/Analysis/ConstantFolding.cpp</tt>: If it is possible to
93 constant fold your intrinsic, add support to it in the
94 <tt>canConstantFoldCallTo</tt> and <tt>ConstantFoldCall</tt> functions.</li>
96 <li><tt>llvm/test/Regression/*</tt>: Add test cases for your test cases to the
100 <p>Once the intrinsic has been added to the system, you must add code generator
101 support for it. Generally you must do the following steps:</p>
104 <dt>Add support to the C backend in <tt>lib/Target/CBackend/</tt></dt>
106 <dd>Depending on the intrinsic, there are a few ways to implement this. For
107 most intrinsics, it makes sense to add code to lower your intrinsic in
108 <tt>LowerIntrinsicCall</tt> in <tt>lib/CodeGen/IntrinsicLowering.cpp</tt>.
109 Second, if it makes sense to lower the intrinsic to an expanded sequence of C
110 code in all cases, just emit the expansion in <tt>visitCallInst</tt> in
111 <tt>Writer.cpp</tt>. If the intrinsic has some way to express it with GCC
112 (or any other compiler) extensions, it can be conditionally supported based on
113 the compiler compiling the CBE output (see <tt>llvm.prefetch</tt> for an
115 Third, if the intrinsic really has no way to be lowered, just have the code
116 generator emit code that prints an error message and calls abort if executed.
120 <dt>Add support to the .td file for the target(s) of your choice in
121 <tt>lib/Target/*/*.td</tt>.</dt>
123 <dd>This is usually a matter of adding a pattern to the .td file that matches
124 the intrinsic, though it may obviously require adding the instructions you
125 want to generate as well. There are lots of examples in the PowerPC and X86
126 backend to follow.</dd>
130 <!-- *********************************************************************** -->
131 <div class="doc_section">
132 <a name="sdnode">Adding a new SelectionDAG node</a>
134 <!-- *********************************************************************** -->
136 <div class="doc_text">
138 <p>As with intrinsics, adding a new SelectionDAG node to LLVM is much easier
139 than adding a new instruction. New nodes are often added to help represent
140 instructions common to many targets. These nodes often map to an LLVM
141 instruction (add, sub) or intrinsic (byteswap, population count). In other
142 cases, new nodes have been added to allow many targets to perform a common task
143 (converting between floating point and integer representation) or capture more
144 complicated behavior in a single node (rotate).</p>
147 <li><tt>include/llvm/CodeGen/SelectionDAGNodes.h</tt>:
148 Add an enum value for the new SelectionDAG node.</li>
149 <li><tt>lib/CodeGen/SelectionDAG/SelectionDAG.cpp</tt>:
150 Add code to print the node to <tt>getOperationName</tt>. If your new node
151 can be evaluated at compile time when given constant arguments (such as an
152 add of a constant with another constant), find the <tt>getNode</tt> method
153 that takes the appropriate number of arguments, and add a case for your node
154 to the switch statement that performs constant folding for nodes that take
155 the same number of arguments as your new node.</li>
156 <li><tt>lib/CodeGen/SelectionDAG/LegalizeDAG.cpp</tt>:
157 Add code to <a href="CodeGenerator.html#selectiondag_legalize">legalize,
158 promote, and expand</a> the node as necessary. At a minimum, you will need
159 to add a case statement for your node in <tt>LegalizeOp</tt> which calls
160 LegalizeOp on the node's operands, and returns a new node if any of the
161 operands changed as a result of being legalized. It is likely that not all
162 targets supported by the SelectionDAG framework will natively support the
163 new node. In this case, you must also add code in your node's case
164 statement in <tt>LegalizeOp</tt> to Expand your node into simpler, legal
165 operations. The case for <tt>ISD::UREM</tt> for expanding a remainder into
166 a divide, multiply, and a subtract is a good example.</li>
167 <li><tt>lib/CodeGen/SelectionDAG/LegalizeDAG.cpp</tt>:
168 If targets may support the new node being added only at certain sizes, you
169 will also need to add code to your node's case statement in
170 <tt>LegalizeOp</tt> to Promote your node's operands to a larger size, and
171 perform the correct operation. You will also need to add code to
172 <tt>PromoteOp</tt> to do this as well. For a good example, see
174 which promotes its operand to a wider size, performs the byteswap, and then
175 shifts the correct bytes right to emulate the narrower byteswap in the
177 <li><tt>lib/CodeGen/SelectionDAG/LegalizeDAG.cpp</tt>:
178 Add a case for your node in <tt>ExpandOp</tt> to teach the legalizer how to
179 perform the action represented by the new node on a value that has been
180 split into high and low halves. This case will be used to support your
181 node with a 64 bit operand on a 32 bit target.</li>
182 <li><tt>lib/CodeGen/SelectionDAG/DAGCombiner.cpp</tt>:
183 If your node can be combined with itself, or other existing nodes in a
184 peephole-like fashion, add a visit function for it, and call that function
185 from <tt></tt>. There are several good examples for simple combines you
186 can do; <tt>visitFABS</tt> and <tt>visitSRL</tt> are good starting places.
188 <li><tt>lib/Target/PowerPC/PPCISelLowering.cpp</tt>:
189 Each target has an implementation of the <tt>TargetLowering</tt> class,
190 usually in its own file (although some targets include it in the same
191 file as the DAGToDAGISel). The default behavior for a target is to
192 assume that your new node is legal for all types that are legal for
193 that target. If this target does not natively support your node, then
194 tell the target to either Promote it (if it is supported at a larger
195 type) or Expand it. This will cause the code you wrote in
196 <tt>LegalizeOp</tt> above to decompose your new node into other legal
197 nodes for this target.</li>
198 <li><tt>lib/Target/TargetSelectionDAG.td</tt>:
199 Most current targets supported by LLVM generate code using the DAGToDAG
200 method, where SelectionDAG nodes are pattern matched to target-specific
201 nodes, which represent individual instructions. In order for the targets
202 to match an instruction to your new node, you must add a def for that node
203 to the list in this file, with the appropriate type constraints. Look at
204 <tt>add</tt>, <tt>bswap</tt>, and <tt>fadd</tt> for examples.</li>
205 <li><tt>lib/Target/PowerPC/PPCInstrInfo.td</tt>:
206 Each target has a tablegen file that describes the target's instruction
207 set. For targets that use the DAGToDAG instruction selection framework,
208 add a pattern for your new node that uses one or more target nodes.
209 Documentation for this is a bit sparse right now, but there are several
210 decent examples. See the patterns for <tt>rotl</tt> in
211 <tt>PPCInstrInfo.td</tt>.</li>
212 <li>TODO: document complex patterns.</li>
213 <li><tt>llvm/test/Regression/CodeGen/*</tt>: Add test cases for your new node
214 to the test suite. <tt>llvm/test/Regression/CodeGen/X86/bswap.ll</tt> is
220 <!-- *********************************************************************** -->
221 <div class="doc_section">
222 <a name="instruction">Adding a new instruction</a>
224 <!-- *********************************************************************** -->
226 <div class="doc_text">
228 <p><span class="doc_warning">WARNING: adding instructions changes the bytecode
229 format, and it will take some effort to maintain compatibility with
230 the previous version.</span> Only add an instruction if it is absolutely
235 <li><tt>llvm/include/llvm/Instruction.def</tt>:
236 add a number for your instruction and an enum name</li>
238 <li><tt>llvm/include/llvm/Instructions.h</tt>:
239 add a definition for the class that will represent your instruction</li>
241 <li><tt>llvm/include/llvm/Support/InstVisitor.h</tt>:
242 add a prototype for a visitor to your new instruction type</li>
244 <li><tt>llvm/lib/AsmParser/Lexer.l</tt>:
245 add a new token to parse your instruction from assembly text file</li>
247 <li><tt>llvm/lib/AsmParser/llvmAsmParser.y</tt>:
248 add the grammar on how your instruction can be read and what it will
249 construct as a result</li>
251 <li><tt>llvm/lib/Bytecode/Reader/Reader.cpp</tt>:
252 add a case for your instruction and how it will be parsed from bytecode</li>
254 <li><tt>llvm/lib/VMCore/Instruction.cpp</tt>:
255 add a case for how your instruction will be printed out to assembly</li>
257 <li><tt>llvm/lib/VMCore/Instructions.cpp</tt>:
258 implement the class you defined in
259 <tt>llvm/include/llvm/Instructions.h</tt></li>
261 <li>Test your instruction</li>
263 <li><tt>llvm/lib/Target/*</tt>:
264 Add support for your instruction to code generators, or add a lowering
267 <li><tt>llvm/test/Regression/*</tt>: add your test cases to the test suite.</li>
271 <p>Also, you need to implement (or modify) any analyses or passes that you want
272 to understand this new instruction.</p>
277 <!-- *********************************************************************** -->
278 <div class="doc_section">
279 <a name="type">Adding a new type</a>
281 <!-- *********************************************************************** -->
283 <div class="doc_text">
285 <p><span class="doc_warning">WARNING: adding new types changes the bytecode
286 format, and will break compatibility with currently-existing LLVM
287 installations.</span> Only add new types if it is absolutely necessary.</p>
291 <!-- ======================================================================= -->
292 <div class="doc_subsection">
293 <a name="fund_type">Adding a fundamental type</a>
296 <div class="doc_text">
300 <li><tt>llvm/include/llvm/Type.h</tt>:
301 add enum for the new type; add static <tt>Type*</tt> for this type</li>
303 <li><tt>llvm/lib/VMCore/Type.cpp</tt>:
304 add mapping from <tt>TypeID</tt> => <tt>Type*</tt>;
305 initialize the static <tt>Type*</tt></li>
307 <li><tt>llvm/lib/AsmReader/Lexer.l</tt>:
308 add ability to parse in the type from text assembly</li>
310 <li><tt>llvm/lib/AsmReader/llvmAsmParser.y</tt>:
311 add a token for that type</li>
317 <!-- ======================================================================= -->
318 <div class="doc_subsection">
319 <a name="derived_type">Adding a derived type</a>
322 <div class="doc_text">
325 <li><tt>llvm/include/llvm/Type.h</tt>:
326 add enum for the new type; add a forward declaration of the type
329 <li><tt>llvm/include/llvm/DerivedTypes.h</tt>:
330 add new class to represent new class in the hierarchy; add forward
331 declaration to the TypeMap value type</li>
333 <li><tt>llvm/lib/VMCore/Type.cpp</tt>:
334 add support for derived type to:
335 <div class="doc_code">
337 std::string getTypeDescription(const Type &Ty,
338 std::vector<const Type*> &TypeStack)
339 bool TypesEqual(const Type *Ty, const Type *Ty2,
340 std::map<const Type*, const Type*> & EqTypes)
343 add necessary member functions for type, and factory methods</li>
345 <li><tt>llvm/lib/AsmReader/Lexer.l</tt>:
346 add ability to parse in the type from text assembly</li>
348 <li><tt>llvm/lib/ByteCode/Writer/Writer.cpp</tt>:
349 modify <tt>void BytecodeWriter::outputType(const Type *T)</tt> to serialize
352 <li><tt>llvm/lib/ByteCode/Reader/Reader.cpp</tt>:
353 modify <tt>const Type *BytecodeReader::ParseType()</tt> to read your data
356 <li><tt>llvm/lib/VMCore/AsmWriter.cpp</tt>:
358 <div class="doc_code">
360 void calcTypeName(const Type *Ty,
361 std::vector<const Type*> &TypeStack,
362 std::map<const Type*,std::string> &TypeNames,
363 std::string & Result)
366 to output the new derived type
374 <!-- *********************************************************************** -->
378 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
379 src="http://jigsaw.w3.org/css-validator/images/vcss" alt="Valid CSS!"></a>
380 <a href="http://validator.w3.org/check/referer"><img
381 src="http://www.w3.org/Icons/valid-html401" alt="Valid HTML 4.01!" /></a>
383 <a href="http://llvm.org">The LLVM Compiler Infrastructure</a>
385 Last modified: $Date$