1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
6 <title>Extending LLVM: Adding instructions, intrinsics, types, etc.</title>
7 <link rel="stylesheet" href="llvm.css" type="text/css">
13 Extending LLVM: Adding instructions, intrinsics, types, etc.
17 <li><a href="#introduction">Introduction and Warning</a></li>
18 <li><a href="#intrinsic">Adding a new intrinsic function</a></li>
19 <li><a href="#instruction">Adding a new instruction</a></li>
20 <li><a href="#sdnode">Adding a new SelectionDAG node</a></li>
21 <li><a href="#type">Adding a new type</a>
23 <li><a href="#fund_type">Adding a new fundamental type</a></li>
24 <li><a href="#derived_type">Adding a new derived type</a></li>
28 <div class="doc_author">
29 <p>Written by <a href="http://misha.brukman.net">Misha Brukman</a>,
30 Brad Jones, Nate Begeman,
31 and <a href="http://nondot.org/sabre">Chris Lattner</a></p>
34 <!-- *********************************************************************** -->
36 <a name="introduction">Introduction and Warning</a>
38 <!-- *********************************************************************** -->
42 <p>During the course of using LLVM, you may wish to customize it for your
43 research project or for experimentation. At this point, you may realize that
44 you need to add something to LLVM, whether it be a new fundamental type, a new
45 intrinsic function, or a whole new instruction.</p>
47 <p>When you come to this realization, stop and think. Do you really need to
48 extend LLVM? Is it a new fundamental capability that LLVM does not support at
49 its current incarnation or can it be synthesized from already pre-existing LLVM
50 elements? If you are not sure, ask on the <a
51 href="http://mail.cs.uiuc.edu/mailman/listinfo/llvmdev">LLVM-dev</a> list. The
52 reason is that extending LLVM will get involved as you need to update all the
53 different passes that you intend to use with your extension, and there are
54 <em>many</em> LLVM analyses and transformations, so it may be quite a bit of
57 <p>Adding an <a href="#intrinsic">intrinsic function</a> is far easier than
58 adding an instruction, and is transparent to optimization passes. If your added
59 functionality can be expressed as a
60 function call, an intrinsic function is the method of choice for LLVM
63 <p>Before you invest a significant amount of effort into a non-trivial
64 extension, <span class="doc_warning">ask on the list</span> if what you are
65 looking to do can be done with already-existing infrastructure, or if maybe
66 someone else is already working on it. You will save yourself a lot of time and
67 effort by doing so.</p>
71 <!-- *********************************************************************** -->
73 <a name="intrinsic">Adding a new intrinsic function</a>
75 <!-- *********************************************************************** -->
79 <p>Adding a new intrinsic function to LLVM is much easier than adding a new
80 instruction. Almost all extensions to LLVM should start as an intrinsic
81 function and then be turned into an instruction if warranted.</p>
84 <li><tt>llvm/docs/LangRef.html</tt>:
85 Document the intrinsic. Decide whether it is code generator specific and
86 what the restrictions are. Talk to other people about it so that you are
87 sure it's a good idea.</li>
89 <li><tt>llvm/include/llvm/Intrinsics*.td</tt>:
90 Add an entry for your intrinsic. Describe its memory access characteristics
91 for optimization (this controls whether it will be DCE'd, CSE'd, etc). Note
92 that any intrinsic using the <tt>llvm_int_ty</tt> type for an argument will
93 be deemed by <tt>tblgen</tt> as overloaded and the corresponding suffix
94 will be required on the intrinsic's name.</li>
96 <li><tt>llvm/lib/Analysis/ConstantFolding.cpp</tt>: If it is possible to
97 constant fold your intrinsic, add support to it in the
98 <tt>canConstantFoldCallTo</tt> and <tt>ConstantFoldCall</tt> functions.</li>
100 <li><tt>llvm/test/Regression/*</tt>: Add test cases for your test cases to the
104 <p>Once the intrinsic has been added to the system, you must add code generator
105 support for it. Generally you must do the following steps:</p>
108 <dt>Add support to the C backend in <tt>lib/Target/CBackend/</tt></dt>
110 <dd>Depending on the intrinsic, there are a few ways to implement this. For
111 most intrinsics, it makes sense to add code to lower your intrinsic in
112 <tt>LowerIntrinsicCall</tt> in <tt>lib/CodeGen/IntrinsicLowering.cpp</tt>.
113 Second, if it makes sense to lower the intrinsic to an expanded sequence of
114 C code in all cases, just emit the expansion in <tt>visitCallInst</tt> in
115 <tt>Writer.cpp</tt>. If the intrinsic has some way to express it with GCC
116 (or any other compiler) extensions, it can be conditionally supported based
117 on the compiler compiling the CBE output (see <tt>llvm.prefetch</tt> for an
118 example). Third, if the intrinsic really has no way to be lowered, just
119 have the code generator emit code that prints an error message and calls
120 abort if executed.</dd>
122 <dt>Add support to the .td file for the target(s) of your choice in
123 <tt>lib/Target/*/*.td</tt>.</dt>
125 <dd>This is usually a matter of adding a pattern to the .td file that matches
126 the intrinsic, though it may obviously require adding the instructions you
127 want to generate as well. There are lots of examples in the PowerPC and X86
128 backend to follow.</dd>
133 <!-- *********************************************************************** -->
135 <a name="sdnode">Adding a new SelectionDAG node</a>
137 <!-- *********************************************************************** -->
141 <p>As with intrinsics, adding a new SelectionDAG node to LLVM is much easier
142 than adding a new instruction. New nodes are often added to help represent
143 instructions common to many targets. These nodes often map to an LLVM
144 instruction (add, sub) or intrinsic (byteswap, population count). In other
145 cases, new nodes have been added to allow many targets to perform a common task
146 (converting between floating point and integer representation) or capture more
147 complicated behavior in a single node (rotate).</p>
150 <li><tt>include/llvm/CodeGen/ISDOpcodes.h</tt>:
151 Add an enum value for the new SelectionDAG node.</li>
152 <li><tt>lib/CodeGen/SelectionDAG/SelectionDAG.cpp</tt>:
153 Add code to print the node to <tt>getOperationName</tt>. If your new node
154 can be evaluated at compile time when given constant arguments (such as an
155 add of a constant with another constant), find the <tt>getNode</tt> method
156 that takes the appropriate number of arguments, and add a case for your node
157 to the switch statement that performs constant folding for nodes that take
158 the same number of arguments as your new node.</li>
159 <li><tt>lib/CodeGen/SelectionDAG/LegalizeDAG.cpp</tt>:
160 Add code to <a href="CodeGenerator.html#selectiondag_legalize">legalize,
161 promote, and expand</a> the node as necessary. At a minimum, you will need
162 to add a case statement for your node in <tt>LegalizeOp</tt> which calls
163 LegalizeOp on the node's operands, and returns a new node if any of the
164 operands changed as a result of being legalized. It is likely that not all
165 targets supported by the SelectionDAG framework will natively support the
166 new node. In this case, you must also add code in your node's case
167 statement in <tt>LegalizeOp</tt> to Expand your node into simpler, legal
168 operations. The case for <tt>ISD::UREM</tt> for expanding a remainder into
169 a divide, multiply, and a subtract is a good example.</li>
170 <li><tt>lib/CodeGen/SelectionDAG/LegalizeDAG.cpp</tt>:
171 If targets may support the new node being added only at certain sizes, you
172 will also need to add code to your node's case statement in
173 <tt>LegalizeOp</tt> to Promote your node's operands to a larger size, and
174 perform the correct operation. You will also need to add code to
175 <tt>PromoteOp</tt> to do this as well. For a good example, see
177 which promotes its operand to a wider size, performs the byteswap, and then
178 shifts the correct bytes right to emulate the narrower byteswap in the
180 <li><tt>lib/CodeGen/SelectionDAG/LegalizeDAG.cpp</tt>:
181 Add a case for your node in <tt>ExpandOp</tt> to teach the legalizer how to
182 perform the action represented by the new node on a value that has been
183 split into high and low halves. This case will be used to support your
184 node with a 64 bit operand on a 32 bit target.</li>
185 <li><tt>lib/CodeGen/SelectionDAG/DAGCombiner.cpp</tt>:
186 If your node can be combined with itself, or other existing nodes in a
187 peephole-like fashion, add a visit function for it, and call that function
188 from <tt></tt>. There are several good examples for simple combines you
189 can do; <tt>visitFABS</tt> and <tt>visitSRL</tt> are good starting places.
191 <li><tt>lib/Target/PowerPC/PPCISelLowering.cpp</tt>:
192 Each target has an implementation of the <tt>TargetLowering</tt> class,
193 usually in its own file (although some targets include it in the same
194 file as the DAGToDAGISel). The default behavior for a target is to
195 assume that your new node is legal for all types that are legal for
196 that target. If this target does not natively support your node, then
197 tell the target to either Promote it (if it is supported at a larger
198 type) or Expand it. This will cause the code you wrote in
199 <tt>LegalizeOp</tt> above to decompose your new node into other legal
200 nodes for this target.</li>
201 <li><tt>lib/Target/TargetSelectionDAG.td</tt>:
202 Most current targets supported by LLVM generate code using the DAGToDAG
203 method, where SelectionDAG nodes are pattern matched to target-specific
204 nodes, which represent individual instructions. In order for the targets
205 to match an instruction to your new node, you must add a def for that node
206 to the list in this file, with the appropriate type constraints. Look at
207 <tt>add</tt>, <tt>bswap</tt>, and <tt>fadd</tt> for examples.</li>
208 <li><tt>lib/Target/PowerPC/PPCInstrInfo.td</tt>:
209 Each target has a tablegen file that describes the target's instruction
210 set. For targets that use the DAGToDAG instruction selection framework,
211 add a pattern for your new node that uses one or more target nodes.
212 Documentation for this is a bit sparse right now, but there are several
213 decent examples. See the patterns for <tt>rotl</tt> in
214 <tt>PPCInstrInfo.td</tt>.</li>
215 <li>TODO: document complex patterns.</li>
216 <li><tt>llvm/test/Regression/CodeGen/*</tt>: Add test cases for your new node
217 to the test suite. <tt>llvm/test/Regression/CodeGen/X86/bswap.ll</tt> is
223 <!-- *********************************************************************** -->
225 <a name="instruction">Adding a new instruction</a>
227 <!-- *********************************************************************** -->
231 <p><span class="doc_warning">WARNING: adding instructions changes the bitcode
232 format, and it will take some effort to maintain compatibility with
233 the previous version.</span> Only add an instruction if it is absolutely
238 <li><tt>llvm/include/llvm/Instruction.def</tt>:
239 add a number for your instruction and an enum name</li>
241 <li><tt>llvm/include/llvm/Instructions.h</tt>:
242 add a definition for the class that will represent your instruction</li>
244 <li><tt>llvm/include/llvm/Support/InstVisitor.h</tt>:
245 add a prototype for a visitor to your new instruction type</li>
247 <li><tt>llvm/lib/AsmParser/Lexer.l</tt>:
248 add a new token to parse your instruction from assembly text file</li>
250 <li><tt>llvm/lib/AsmParser/llvmAsmParser.y</tt>:
251 add the grammar on how your instruction can be read and what it will
252 construct as a result</li>
254 <li><tt>llvm/lib/Bitcode/Reader/Reader.cpp</tt>:
255 add a case for your instruction and how it will be parsed from bitcode</li>
257 <li><tt>llvm/lib/VMCore/Instruction.cpp</tt>:
258 add a case for how your instruction will be printed out to assembly</li>
260 <li><tt>llvm/lib/VMCore/Instructions.cpp</tt>:
261 implement the class you defined in
262 <tt>llvm/include/llvm/Instructions.h</tt></li>
264 <li>Test your instruction</li>
266 <li><tt>llvm/lib/Target/*</tt>:
267 Add support for your instruction to code generators, or add a lowering
270 <li><tt>llvm/test/Regression/*</tt>: add your test cases to the test suite.</li>
274 <p>Also, you need to implement (or modify) any analyses or passes that you want
275 to understand this new instruction.</p>
280 <!-- *********************************************************************** -->
282 <a name="type">Adding a new type</a>
284 <!-- *********************************************************************** -->
288 <p><span class="doc_warning">WARNING: adding new types changes the bitcode
289 format, and will break compatibility with currently-existing LLVM
290 installations.</span> Only add new types if it is absolutely necessary.</p>
292 <!-- ======================================================================= -->
294 <a name="fund_type">Adding a fundamental type</a>
301 <li><tt>llvm/include/llvm/Type.h</tt>:
302 add enum for the new type; add static <tt>Type*</tt> for this type</li>
304 <li><tt>llvm/lib/VMCore/Type.cpp</tt>:
305 add mapping from <tt>TypeID</tt> => <tt>Type*</tt>;
306 initialize the static <tt>Type*</tt></li>
308 <li><tt>llvm/lib/AsmReader/Lexer.l</tt>:
309 add ability to parse in the type from text assembly</li>
311 <li><tt>llvm/lib/AsmReader/llvmAsmParser.y</tt>:
312 add a token for that type</li>
318 <!-- ======================================================================= -->
320 <a name="derived_type">Adding a derived type</a>
326 <li><tt>llvm/include/llvm/Type.h</tt>:
327 add enum for the new type; add a forward declaration of the type
330 <li><tt>llvm/include/llvm/DerivedTypes.h</tt>:
331 add new class to represent new class in the hierarchy; add forward
332 declaration to the TypeMap value type</li>
334 <li><tt>llvm/lib/VMCore/Type.cpp</tt>:
335 add support for derived type to:
336 <div class="doc_code">
338 std::string getTypeDescription(const Type &Ty,
339 std::vector<const Type*> &TypeStack)
340 bool TypesEqual(const Type *Ty, const Type *Ty2,
341 std::map<const Type*, const Type*> & EqTypes)
344 add necessary member functions for type, and factory methods</li>
346 <li><tt>llvm/lib/AsmReader/Lexer.l</tt>:
347 add ability to parse in the type from text assembly</li>
349 <li><tt>llvm/lib/BitCode/Writer/Writer.cpp</tt>:
350 modify <tt>void BitcodeWriter::outputType(const Type *T)</tt> to serialize
353 <li><tt>llvm/lib/BitCode/Reader/Reader.cpp</tt>:
354 modify <tt>const Type *BitcodeReader::ParseType()</tt> to read your data
357 <li><tt>llvm/lib/VMCore/AsmWriter.cpp</tt>:
359 <div class="doc_code">
361 void calcTypeName(const Type *Ty,
362 std::vector<const Type*> &TypeStack,
363 std::map<const Type*,std::string> &TypeNames,
364 std::string & Result)
367 to output the new derived type
377 <!-- *********************************************************************** -->
381 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
382 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
383 <a href="http://validator.w3.org/check/referer"><img
384 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
386 <a href="http://llvm.org/">The LLVM Compiler Infrastructure</a>
388 Last modified: $Date$