xref: /llvm-project/llvm/docs/ExtendingLLVM.rst (revision 0cc6b464f8adb739e6f9bcc9bda8cff9bb1f1c2a)
1============================================================
2Extending LLVM: Adding instructions, intrinsics, types, etc.
3============================================================
4
5Introduction and Warning
6========================
7
8
9During the course of using LLVM, you may wish to customize it for your research
10project or for experimentation. At this point, you may realize that you need to
11add something to LLVM, whether it be a new fundamental type, a new intrinsic
12function, or a whole new instruction.
13
14When you come to this realization, stop and think. Do you really need to extend
15LLVM? Is it a new fundamental capability that LLVM does not support at its
16current incarnation or can it be synthesized from already pre-existing LLVM
17elements? If you are not sure, ask on the `LLVM forums
18<https://discourse.llvm.org>`_. The reason is that
19extending LLVM will get involved as you need to update all the different passes
20that you intend to use with your extension, and there are ``many`` LLVM analyses
21and transformations, so it may be quite a bit of work.
22
23Adding an `intrinsic function`_ is far easier than adding an
24instruction, and is transparent to optimization passes.  If your added
25functionality can be expressed as a function call, an intrinsic function is the
26method of choice for LLVM extension.
27
28Before 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
30already-existing infrastructure, or if maybe someone else is already working on
31it. You will save yourself a lot of time and effort by doing so.
32
33.. _intrinsic function:
34
35Adding a new intrinsic function
36===============================
37
38Adding a new intrinsic function to LLVM is much easier than adding a new
39instruction.  Almost all extensions to LLVM should start as an intrinsic
40function and then be turned into an instruction if warranted.
41
42#. ``llvm/docs/LangRef.html``:
43
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.
47
48#. ``llvm/include/llvm/IR/Intrinsics*.td``:
49
50   Add an entry for your intrinsic.  Describe its memory access
51   characteristics for optimization (this controls whether it will be
52   DCE'd, CSE'd, etc). If any arguments need to be immediates, these
53   must be indicated with the ImmArg property. Note that any intrinsic
54   using one of the ``llvm_any*_ty`` types for an argument or return
55   type will be deemed by ``tblgen`` as overloaded and the
56   corresponding suffix will be required on the intrinsic's name.
57
58#. ``llvm/lib/Analysis/ConstantFolding.cpp``:
59
60   If it is possible to constant fold your intrinsic, add support to it in the
61   ``canConstantFoldCallTo`` and ``ConstantFoldCall`` functions.
62
63#. ``llvm/test/*``:
64
65   Add test cases for your test cases to the test suite
66
67Once the intrinsic has been added to the system, you must add code generator
68support for it.  Generally you must do the following steps:
69
70Add support to the .td file for the target(s) of your choice in
71``lib/Target/*/*.td``.
72
73  This is usually a matter of adding a pattern to the .td file that matches the
74  intrinsic, though it may obviously require adding the instructions you want to
75  generate as well.  There are lots of examples in the PowerPC and X86 backend
76  to follow.
77
78Adding a new SelectionDAG node
79==============================
80
81As with intrinsics, adding a new SelectionDAG node to LLVM is much easier than
82adding a new instruction.  New nodes are often added to help represent
83instructions common to many targets.  These nodes often map to an LLVM
84instruction (add, sub) or intrinsic (byteswap, population count).  In other
85cases, new nodes have been added to allow many targets to perform a common task
86(converting between floating point and integer representation) or capture more
87complicated behavior in a single node (rotate).
88
89#. ``include/llvm/CodeGen/ISDOpcodes.h``:
90
91   Add an enum value for the new SelectionDAG node.
92
93#. ``lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp``:
94
95   Add code to print the node to ``getOperationName``.  If your new node can be
96   evaluated at compile time when given constant arguments (such as an add of a
97   constant with another constant), find the ``getNode`` method that takes the
98   appropriate number of arguments, and add a case for your node to the switch
99   statement that performs constant folding for nodes that take the same number
100   of arguments as your new node.
101
102#. ``lib/CodeGen/SelectionDAG/LegalizeDAG.cpp``:
103
104   Add code to `legalize, promote, and expand
105   <CodeGenerator.html#selectiondag_legalize>`_ the node as necessary.  At a
106   minimum, you will need to add a case statement for your node in
107   ``LegalizeOp`` which calls LegalizeOp on the node's operands, and returns a
108   new node if any of the operands changed as a result of being legalized.  It
109   is likely that not all targets supported by the SelectionDAG framework will
110   natively support the new node.  In this case, you must also add code in your
111   node's case statement in ``LegalizeOp`` to Expand your node into simpler,
112   legal operations.  The case for ``ISD::UREM`` for expanding a remainder into
113   a divide, multiply, and a subtract is a good example.
114
115#. ``lib/CodeGen/SelectionDAG/LegalizeDAG.cpp``:
116
117   If targets may support the new node being added only at certain sizes, you
118   will also need to add code to your node's case statement in ``LegalizeOp``
119   to Promote your node's operands to a larger size, and perform the correct
120   operation.  You will also need to add code to ``PromoteOp`` to do this as
121   well.  For a good example, see ``ISD::BSWAP``, which promotes its operand to
122   a wider size, performs the byteswap, and then shifts the correct bytes right
123   to emulate the narrower byteswap in the wider type.
124
125#. ``lib/CodeGen/SelectionDAG/LegalizeDAG.cpp``:
126
127   Add a case for your node in ``ExpandOp`` to teach the legalizer how to
128   perform the action represented by the new node on a value that has been split
129   into high and low halves.  This case will be used to support your node with a
130   64 bit operand on a 32 bit target.
131
132#. ``lib/CodeGen/SelectionDAG/DAGCombiner.cpp``:
133
134   If your node can be combined with itself, or other existing nodes in a
135   peephole-like fashion, add a visit function for it, and call that function
136   from. There are several good examples for simple combines you can do;
137   ``visitFABS`` and ``visitSRL`` are good starting places.
138
139#. ``lib/Target/PowerPC/PPCISelLowering.cpp``:
140
141   Each target has an implementation of the ``TargetLowering`` class, usually in
142   its own file (although some targets include it in the same file as the
143   DAGToDAGISel).  The default behavior for a target is to assume that your new
144   node is legal for all types that are legal for that target.  If this target
145   does not natively support your node, then tell the target to either Promote
146   it (if it is supported at a larger type) or Expand it.  This will cause the
147   code you wrote in ``LegalizeOp`` above to decompose your new node into other
148   legal nodes for this target.
149
150#. ``include/llvm/Target/TargetSelectionDAG.td``:
151
152   Most current targets supported by LLVM generate code using the DAGToDAG
153   method, where SelectionDAG nodes are pattern matched to target-specific
154   nodes, which represent individual instructions.  In order for the targets to
155   match an instruction to your new node, you must add a def for that node to
156   the list in this file, with the appropriate type constraints. Look at
157   ``add``, ``bswap``, and ``fadd`` for examples.
158
159#. ``lib/Target/PowerPC/PPCInstrInfo.td``:
160
161   Each target has a tablegen file that describes the target's instruction set.
162   For targets that use the DAGToDAG instruction selection framework, add a
163   pattern for your new node that uses one or more target nodes.  Documentation
164   for this is a bit sparse right now, but there are several decent examples.
165   See the patterns for ``rotl`` in ``PPCInstrInfo.td``.
166
167#. TODO: document complex patterns.
168
169#. ``llvm/test/CodeGen/*``:
170
171   Add test cases for your new node to the test suite.
172   ``llvm/test/CodeGen/X86/bswap.ll`` is a good example.
173
174Adding a new instruction
175========================
176
177.. warning::
178
179  Adding instructions changes the bitcode format, and it will take some effort
180  to maintain compatibility with the previous version. Only add an instruction
181  if it is absolutely necessary.
182
183#. ``llvm/include/llvm/IR/Instruction.def``:
184
185   add a number for your instruction and an enum name
186
187#. ``llvm/include/llvm/IR/Instructions.h``:
188
189   add a definition for the class that will represent your instruction
190
191#. ``llvm/include/llvm/IR/InstVisitor.h``:
192
193   add a prototype for a visitor to your new instruction type
194
195#. ``llvm/lib/AsmParser/LLLexer.cpp``:
196
197   add a new token to parse your instruction from assembly text file
198
199#. ``llvm/lib/AsmParser/LLParser.cpp``:
200
201   add the grammar on how your instruction can be read and what it will
202   construct as a result
203
204#. ``llvm/lib/Bitcode/Reader/BitcodeReader.cpp``:
205
206   add a case for your instruction and how it will be parsed from bitcode
207
208#. ``llvm/lib/Bitcode/Writer/BitcodeWriter.cpp``:
209
210   add a case for your instruction and how it will be parsed from bitcode
211
212#. ``llvm/lib/IR/Instruction.cpp``:
213
214   add a case for how your instruction will be printed out to assembly
215
216#. ``llvm/lib/IR/Instructions.cpp``:
217
218   implement the class you defined in ``llvm/include/llvm/Instructions.h``
219
220#. Test your instruction
221
222#. ``llvm/lib/Target/*``:
223
224   add support for your instruction to code generators, or add a lowering pass.
225
226#. ``llvm/test/*``:
227
228   add your test cases to the test suite.
229
230Also, you need to implement (or modify) any analyses or passes that you want to
231understand this new instruction.
232
233Adding a new type
234=================
235
236.. warning::
237
238  Adding new types changes the bitcode format, and will break compatibility with
239  currently-existing LLVM installations. Only add new types if it is absolutely
240  necessary.
241
242Adding a fundamental type
243-------------------------
244
245#. ``llvm/include/llvm/IR/Type.h``:
246
247   add enum for the new type; add static ``Type*`` for this type
248
249#. ``llvm/lib/IR/Type.cpp`` and ``llvm/lib/CodeGen/ValueTypes.cpp``:
250
251   add mapping from ``TypeID`` => ``Type*``; initialize the static ``Type*``
252
253#. ``llvm/include/llvm-c/Core.h`` and ``llvm/lib/IR/Core.cpp``:
254
255   add enum ``LLVMTypeKind`` and modify
256   ``LLVMTypeKind LLVMGetTypeKind(LLVMTypeRef Ty)`` for the new type
257
258#. ``llvm/lib/AsmParser/LLLexer.cpp``:
259
260   add ability to parse in the type from text assembly
261
262#. ``llvm/lib/AsmParser/LLParser.cpp``:
263
264   add a token for that type
265
266#. ``llvm/lib/Bitcode/Writer/BitcodeWriter.cpp``:
267
268   modify ``void ModuleBitcodeWriter::writeTypeTable()`` to serialize your type
269
270#. ``llvm/lib/Bitcode/Reader/BitcodeReader.cpp``:
271
272   modify ``Error BitcodeReader::parseTypeTableBody()`` to read your data type
273
274#. ``include/llvm/Bitcode/LLVMBitCodes.h``:
275
276   add enum ``TypeCodes`` for the new type
277
278Adding a derived type
279---------------------
280
281#. ``llvm/include/llvm/IR/Type.h``:
282
283   add enum for the new type; add a forward declaration of the type also
284
285#. ``llvm/include/llvm/IR/DerivedTypes.h``:
286
287   add new class to represent new class in the hierarchy; add forward
288   declaration to the TypeMap value type
289
290#. ``llvm/lib/IR/Type.cpp`` and ``llvm/lib/CodeGen/ValueTypes.cpp``:
291
292   add support for derived type, notably `enum TypeID` and `is`, `get` methods.
293
294#. ``llvm/include/llvm-c/Core.h`` and ``llvm/lib/IR/Core.cpp``:
295
296   add enum ``LLVMTypeKind`` and modify
297   `LLVMTypeKind LLVMGetTypeKind(LLVMTypeRef Ty)` for the new type
298
299#. ``llvm/lib/AsmParser/LLLexer.cpp``:
300
301   modify ``lltok::Kind LLLexer::LexIdentifier()`` to add ability to
302   parse in the type from text assembly
303
304#. ``llvm/lib/Bitcode/Writer/BitcodeWriter.cpp``:
305
306   modify ``void ModuleBitcodeWriter::writeTypeTable()`` to serialize your type
307
308#. ``llvm/lib/Bitcode/Reader/BitcodeReader.cpp``:
309
310   modify ``Error BitcodeReader::parseTypeTableBody()`` to read your data type
311
312#. ``include/llvm/Bitcode/LLVMBitCodes.h``:
313
314   add enum ``TypeCodes`` for the new type
315
316#. ``llvm/lib/IR/AsmWriter.cpp``:
317
318   modify ``void TypePrinting::print(Type *Ty, raw_ostream &OS)``
319   to output the new derived type
320