1 ============================================================
2 Extending LLVM: Adding instructions, intrinsics, types, etc.
3 ============================================================
5 Introduction and Warning
6 ========================
9 During the course of using LLVM, you may wish to customize it for your research
10 project or for experimentation. At this point, you may realize that you need to
11 add something to LLVM, whether it be a new fundamental type, a new intrinsic
12 function, or a whole new instruction.
14 When you come to this realization, stop and think. Do you really need to extend
15 LLVM? Is it a new fundamental capability that LLVM does not support at its
16 current incarnation or can it be synthesized from already pre-existing LLVM
17 elements? If you are not sure, ask on the `LLVM-dev
18 <http://mail.cs.uiuc.edu/mailman/listinfo/llvmdev>`_ list. The reason is that
19 extending LLVM will get involved as you need to update all the different passes
20 that you intend to use with your extension, and there are ``many`` LLVM analyses
21 and transformations, so it may be quite a bit of work.
23 Adding an `intrinsic function`_ is far easier than adding an
24 instruction, and is transparent to optimization passes. If your added
25 functionality can be expressed as a function call, an intrinsic function is the
26 method of choice for LLVM extension.
28 Before you invest a significant amount of effort into a non-trivial extension,
29 **ask on the list** if what you are looking to do can be done with
30 already-existing infrastructure, or if maybe someone else is already working on
31 it. You will save yourself a lot of time and effort by doing so.
33 .. _intrinsic function:
35 Adding a new intrinsic function
36 ===============================
38 Adding a new intrinsic function to LLVM is much easier than adding a new
39 instruction. Almost all extensions to LLVM should start as an intrinsic
40 function and then be turned into an instruction if warranted.
42 #. ``llvm/docs/LangRef.html``:
44 Document the intrinsic. Decide whether it is code generator specific and
45 what the restrictions are. Talk to other people about it so that you are
46 sure it's a good idea.
48 #. ``llvm/include/llvm/IR/Intrinsics*.td``:
50 Add an entry for your intrinsic. Describe its memory access characteristics
51 for optimization (this controls whether it will be DCE'd, CSE'd, etc). Note
52 that any intrinsic using the ``llvm_int_ty`` type for an argument will
53 be deemed by ``tblgen`` as overloaded and the corresponding suffix will
54 be required on the intrinsic's name.
56 #. ``llvm/lib/Analysis/ConstantFolding.cpp``:
58 If it is possible to constant fold your intrinsic, add support to it in the
59 ``canConstantFoldCallTo`` and ``ConstantFoldCall`` functions.
61 #. ``llvm/test/Regression/*``:
63 Add test cases for your test cases to the test suite
65 Once the intrinsic has been added to the system, you must add code generator
66 support for it. Generally you must do the following steps:
68 Add support to the .td file for the target(s) of your choice in
69 ``lib/Target/*/*.td``.
71 This is usually a matter of adding a pattern to the .td file that matches the
72 intrinsic, though it may obviously require adding the instructions you want to
73 generate as well. There are lots of examples in the PowerPC and X86 backend
76 Adding a new SelectionDAG node
77 ==============================
79 As with intrinsics, adding a new SelectionDAG node to LLVM is much easier than
80 adding a new instruction. New nodes are often added to help represent
81 instructions common to many targets. These nodes often map to an LLVM
82 instruction (add, sub) or intrinsic (byteswap, population count). In other
83 cases, new nodes have been added to allow many targets to perform a common task
84 (converting between floating point and integer representation) or capture more
85 complicated behavior in a single node (rotate).
87 #. ``include/llvm/CodeGen/ISDOpcodes.h``:
89 Add an enum value for the new SelectionDAG node.
91 #. ``lib/CodeGen/SelectionDAG/SelectionDAG.cpp``:
93 Add code to print the node to ``getOperationName``. If your new node can be
94 evaluated at compile time when given constant arguments (such as an add of a
95 constant with another constant), find the ``getNode`` method that takes the
96 appropriate number of arguments, and add a case for your node to the switch
97 statement that performs constant folding for nodes that take the same number
98 of arguments as your new node.
100 #. ``lib/CodeGen/SelectionDAG/LegalizeDAG.cpp``:
102 Add code to `legalize, promote, and expand
103 <CodeGenerator.html#selectiondag_legalize>`_ the node as necessary. At a
104 minimum, you will need to add a case statement for your node in
105 ``LegalizeOp`` which calls LegalizeOp on the node's operands, and returns a
106 new node if any of the operands changed as a result of being legalized. It
107 is likely that not all targets supported by the SelectionDAG framework will
108 natively support the new node. In this case, you must also add code in your
109 node's case statement in ``LegalizeOp`` to Expand your node into simpler,
110 legal operations. The case for ``ISD::UREM`` for expanding a remainder into
111 a divide, multiply, and a subtract is a good example.
113 #. ``lib/CodeGen/SelectionDAG/LegalizeDAG.cpp``:
115 If targets may support the new node being added only at certain sizes, you
116 will also need to add code to your node's case statement in ``LegalizeOp``
117 to Promote your node's operands to a larger size, and perform the correct
118 operation. You will also need to add code to ``PromoteOp`` to do this as
119 well. For a good example, see ``ISD::BSWAP``, which promotes its operand to
120 a wider size, performs the byteswap, and then shifts the correct bytes right
121 to emulate the narrower byteswap in the wider type.
123 #. ``lib/CodeGen/SelectionDAG/LegalizeDAG.cpp``:
125 Add a case for your node in ``ExpandOp`` to teach the legalizer how to
126 perform the action represented by the new node on a value that has been split
127 into high and low halves. This case will be used to support your node with a
128 64 bit operand on a 32 bit target.
130 #. ``lib/CodeGen/SelectionDAG/DAGCombiner.cpp``:
132 If your node can be combined with itself, or other existing nodes in a
133 peephole-like fashion, add a visit function for it, and call that function
134 from. There are several good examples for simple combines you can do;
135 ``visitFABS`` and ``visitSRL`` are good starting places.
137 #. ``lib/Target/PowerPC/PPCISelLowering.cpp``:
139 Each target has an implementation of the ``TargetLowering`` class, usually in
140 its own file (although some targets include it in the same file as the
141 DAGToDAGISel). The default behavior for a target is to assume that your new
142 node is legal for all types that are legal for that target. If this target
143 does not natively support your node, then tell the target to either Promote
144 it (if it is supported at a larger type) or Expand it. This will cause the
145 code you wrote in ``LegalizeOp`` above to decompose your new node into other
146 legal nodes for this target.
148 #. ``lib/Target/TargetSelectionDAG.td``:
150 Most current targets supported by LLVM generate code using the DAGToDAG
151 method, where SelectionDAG nodes are pattern matched to target-specific
152 nodes, which represent individual instructions. In order for the targets to
153 match an instruction to your new node, you must add a def for that node to
154 the list in this file, with the appropriate type constraints. Look at
155 ``add``, ``bswap``, and ``fadd`` for examples.
157 #. ``lib/Target/PowerPC/PPCInstrInfo.td``:
159 Each target has a tablegen file that describes the target's instruction set.
160 For targets that use the DAGToDAG instruction selection framework, add a
161 pattern for your new node that uses one or more target nodes. Documentation
162 for this is a bit sparse right now, but there are several decent examples.
163 See the patterns for ``rotl`` in ``PPCInstrInfo.td``.
165 #. TODO: document complex patterns.
167 #. ``llvm/test/Regression/CodeGen/*``:
169 Add test cases for your new node to the test suite.
170 ``llvm/test/Regression/CodeGen/X86/bswap.ll`` is a good example.
172 Adding a new instruction
173 ========================
177 Adding instructions changes the bitcode format, and it will take some effort
178 to maintain compatibility with the previous version. Only add an instruction
179 if it is absolutely necessary.
181 #. ``llvm/include/llvm/Instruction.def``:
183 add a number for your instruction and an enum name
185 #. ``llvm/include/llvm/Instructions.h``:
187 add a definition for the class that will represent your instruction
189 #. ``llvm/include/llvm/Support/InstVisitor.h``:
191 add a prototype for a visitor to your new instruction type
193 #. ``llvm/lib/AsmParser/Lexer.l``:
195 add a new token to parse your instruction from assembly text file
197 #. ``llvm/lib/AsmParser/llvmAsmParser.y``:
199 add the grammar on how your instruction can be read and what it will
200 construct as a result
202 #. ``llvm/lib/Bitcode/Reader/Reader.cpp``:
204 add a case for your instruction and how it will be parsed from bitcode
206 #. ``llvm/lib/VMCore/Instruction.cpp``:
208 add a case for how your instruction will be printed out to assembly
210 #. ``llvm/lib/VMCore/Instructions.cpp``:
212 implement the class you defined in ``llvm/include/llvm/Instructions.h``
214 #. Test your instruction
216 #. ``llvm/lib/Target/*``:
218 add support for your instruction to code generators, or add a lowering pass.
220 #. ``llvm/test/Regression/*``:
222 add your test cases to the test suite.
224 Also, you need to implement (or modify) any analyses or passes that you want to
225 understand this new instruction.
232 Adding new types changes the bitcode format, and will break compatibility with
233 currently-existing LLVM installations. Only add new types if it is absolutely
236 Adding a fundamental type
237 -------------------------
239 #. ``llvm/include/llvm/Type.h``:
241 add enum for the new type; add static ``Type*`` for this type
243 #. ``llvm/lib/VMCore/Type.cpp``:
245 add mapping from ``TypeID`` => ``Type*``; initialize the static ``Type*``
247 #. ``llvm/lib/AsmReader/Lexer.l``:
249 add ability to parse in the type from text assembly
251 #. ``llvm/lib/AsmReader/llvmAsmParser.y``:
253 add a token for that type
255 Adding a derived type
256 ---------------------
258 #. ``llvm/include/llvm/Type.h``:
260 add enum for the new type; add a forward declaration of the type also
262 #. ``llvm/include/llvm/DerivedTypes.h``:
264 add new class to represent new class in the hierarchy; add forward
265 declaration to the TypeMap value type
267 #. ``llvm/lib/VMCore/Type.cpp``:
269 add support for derived type to:
273 std::string getTypeDescription(const Type &Ty,
274 std::vector<const Type*> &TypeStack)
275 bool TypesEqual(const Type *Ty, const Type *Ty2,
276 std::map<const Type*, const Type*> &EqTypes)
278 add necessary member functions for type, and factory methods
280 #. ``llvm/lib/AsmReader/Lexer.l``:
282 add ability to parse in the type from text assembly
284 #. ``llvm/lib/Bitcode/Writer/Writer.cpp``:
286 modify ``void BitcodeWriter::outputType(const Type *T)`` to serialize your
289 #. ``llvm/lib/Bitcode/Reader/Reader.cpp``:
291 modify ``const Type *BitcodeReader::ParseType()`` to read your data type
293 #. ``llvm/lib/VMCore/AsmWriter.cpp``:
299 void calcTypeName(const Type *Ty,
300 std::vector<const Type*> &TypeStack,
301 std::map<const Type*,std::string> &TypeNames,
304 to output the new derived type