xref: /llvm-project/llvm/docs/GlobalISel/Legalizer.rst (revision a4c7c66098c512a4f616b2072814172114e5a422)
1.. _milegalizer:
2
3Legalizer
4---------
5
6This pass transforms the generic machine instructions such that they are legal.
7
8A legal instruction is defined as:
9
10* **selectable** --- the target will later be able to select it to a
11  target-specific (non-generic) instruction. This doesn't necessarily mean that
12  :doc:`InstructionSelect` has to handle it though. It just means that
13  **something** must handle it.
14
15* operating on **vregs that can be loaded and stored** -- if necessary, the
16  target can select a ``G_LOAD``/``G_STORE`` of each gvreg operand.
17
18As opposed to SelectionDAG, there are no legalization phases.  In particular,
19'type' and 'operation' legalization are not separate.
20
21Legalization is iterative, and all state is contained in GMIR.  To maintain the
22validity of the intermediate code, instructions are introduced:
23
24* ``G_MERGE_VALUES`` --- concatenate multiple registers of the same
25  size into a single wider register.
26
27* ``G_UNMERGE_VALUES`` --- extract multiple registers of the same size
28  from a single wider register.
29
30* ``G_EXTRACT`` --- extract a simple register (as contiguous sequences of bits)
31  from a single wider register.
32
33As they are expected to be temporary byproducts of the legalization process,
34they are combined at the end of the :ref:`milegalizer` pass.
35If any remain, they are expected to always be selectable, using loads and stores
36if necessary.
37
38The legality of an instruction may only depend on the instruction itself and
39must not depend on any context in which the instruction is used. However, after
40deciding that an instruction is not legal, using the context of the instruction
41to decide how to legalize the instruction is permitted. As an example, if we
42have a ``G_FOO`` instruction of the form::
43
44  %1:_(s32) = G_CONSTANT i32 1
45  %2:_(s32) = G_FOO %0:_(s32), %1:_(s32)
46
47it's impossible to say that G_FOO is legal iff %1 is a ``G_CONSTANT`` with
48value ``1``. However, the following::
49
50  %2:_(s32) = G_FOO %0:_(s32), i32 1
51
52can say that it's legal iff operand 2 is an immediate with value ``1`` because
53that information is entirely contained within the single instruction.
54
55.. _api-legalizerinfo:
56
57API: LegalizerInfo
58^^^^^^^^^^^^^^^^^^
59
60The recommended [#legalizer-legacy-footnote]_ API looks like this::
61
62  getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL, G_AND, G_OR, G_XOR, G_SHL})
63      .legalFor({s32, s64, v2s32, v4s32, v2s64})
64      .clampScalar(0, s32, s64)
65      .widenScalarToNextPow2(0)
66      .clampNumElements(0, v2s32, v4s32)
67      .clampNumElements(0, v2s64, v2s64)
68      .moreElementsToNextPow2(0);
69
70and describes a set of rules by which we can either declare an instruction legal
71or decide which action to take to make it more legal.
72
73At the core of this ruleset is the ``LegalityQuery`` which describes the
74instruction. We use a description rather than the instruction to both allow other
75passes to determine legality without having to create an instruction and also to
76limit the information available to the predicates to that which is safe to rely
77on. Currently, the information available to the predicates that determine
78legality contains:
79
80* The opcode for the instruction
81
82* The type of each type index (see ``type0``, ``type1``, etc.)
83
84* The size in bytes and atomic ordering for each MachineMemOperand
85
86.. note::
87
88  An alternative worth investigating is to generalize the API to represent
89  actions using ``std::function`` that implements the action, instead of explicit
90  enum tokens (``Legal``, ``WidenScalar``, ...) that instruct it to call a
91  function. This would have some benefits, most notable being that Custom could
92  be removed.
93
94.. rubric:: Footnotes
95
96.. [#legalizer-legacy-footnote] An API is broadly similar to
97   SelectionDAG/TargetLowering is available but is not recommended as a more
98   powerful API is available.
99
100Rule Processing and Declaring Rules
101"""""""""""""""""""""""""""""""""""
102
103The ``getActionDefinitionsBuilder`` function generates a ruleset for the given
104opcode(s) that rules can be added to. If multiple opcodes are given, they are
105all permanently bound to the same ruleset. The rules in a ruleset are executed
106from top to bottom and will start again from the top if an instruction is
107legalized as a result of the rules. If the ruleset is exhausted without
108satisfying any rule, then it is considered unsupported.
109
110When it doesn't declare the instruction legal, each pass over the rules may
111request that one type changes to another type. Sometimes this can cause multiple
112types to change but we avoid this as much as possible as making multiple changes
113can make it difficult to avoid infinite loops where, for example, narrowing one
114type causes another to be too small and widening that type causes the first one
115to be too big.
116
117In general, it's advisable to declare instructions legal as close to the top of
118the rule as possible and to place any expensive rules as low as possible. This
119helps with performance as testing for legality happens more often than
120legalization and legalization can require multiple passes over the rules.
121
122As a concrete example, consider the rule::
123
124  getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL, G_AND, G_OR, G_XOR, G_SHL})
125      .legalFor({s32, s64, v2s32, v4s32, v2s64})
126      .clampScalar(0, s32, s64)
127      .widenScalarToNextPow2(0);
128
129and the instruction::
130
131  %2:_(s7) = G_ADD %0:_(s7), %1:_(s7)
132
133this doesn't meet the predicate for the :ref:`.legalFor() <legalfor>` as ``s7``
134is not one of the listed types so it falls through to the
135:ref:`.clampScalar() <clampscalar>`. It does meet the predicate for this rule
136as the type is smaller than the ``s32`` and this rule instructs the legalizer
137to change type 0 to ``s32``. It then restarts from the top. This time it does
138satisfy ``.legalFor()`` and the resulting output is::
139
140  %3:_(s32) = G_ANYEXT %0:_(s7)
141  %4:_(s32) = G_ANYEXT %1:_(s7)
142  %5:_(s32) = G_ADD %3:_(s32), %4:_(s32)
143  %2:_(s7) = G_TRUNC %5:_(s32)
144
145where the ``G_ADD`` is legal and the other instructions are scheduled for
146processing by the legalizer.
147
148Rule Actions
149""""""""""""
150
151There are various rule factories that append rules to a ruleset but they have a
152few actions in common:
153
154.. _legalfor:
155
156* ``legalIf()``, ``legalFor()``, etc. declare an instruction to be legal if the
157  predicate is satisfied.
158
159* ``narrowScalarIf()``, ``narrowScalarFor()``, etc. declare an instruction to be illegal
160  if the predicate is satisfied and indicates that narrowing the scalars in one
161  of the types to a specific type would make it more legal. This action supports
162  both scalars and vectors.
163
164* ``widenScalarIf()``, ``widenScalarFor()``, etc. declare an instruction to be illegal
165  if the predicate is satisfied and indicates that widening the scalars in one
166  of the types to a specific type would make it more legal. This action supports
167  both scalars and vectors.
168
169* ``fewerElementsIf()``, ``fewerElementsFor()``, etc. declare an instruction to be
170  illegal if the predicate is satisfied and indicates reducing the number of
171  vector elements in one of the types to a specific type would make it more
172  legal. This action supports vectors.
173
174* ``moreElementsIf()``, ``moreElementsFor()``, etc. declare an instruction to be illegal
175  if the predicate is satisfied and indicates increasing the number of vector
176  elements in one of the types to a specific type would make it more legal.
177  This action supports vectors.
178
179* ``lowerIf()``, ``lowerFor()``, etc. declare an instruction to be
180  illegal if the predicate is satisfied and indicates that replacing
181  it with equivalent instruction(s) would make it more legal. Support
182  for this action differs for each opcode. These may provide an
183  optional LegalizeMutation containing a type to attempt to perform
184  the expansion in a different type.
185
186* ``libcallIf()``, ``libcallFor()``, etc. declare an instruction to be illegal if the
187  predicate is satisfied and indicates that replacing it with a libcall would
188  make it more legal. Support for this action differs for
189  each opcode.
190
191* ``customIf()``, ``customFor()``, etc. declare an instruction to be illegal if the
192  predicate is satisfied and indicates that the backend developer will supply
193  a means of making it more legal.
194
195* ``unsupportedIf()``, ``unsupportedFor()``, etc. declare an instruction to be illegal
196  if the predicate is satisfied and indicates that there is no way to make it
197  legal and the compiler should fail.
198
199* ``fallback()`` falls back on an older API and should only be used while porting
200  existing code from that API.
201
202Rule Predicates
203"""""""""""""""
204
205The rule factories also have predicates in common:
206
207* ``legal()``, ``lower()``, etc. are always satisfied.
208
209* ``legalIf()``, ``narrowScalarIf()``, etc. are satisfied if the user-supplied
210  ``LegalityPredicate`` function returns true. This predicate has access to the
211  information in the ``LegalityQuery`` to make its decision.
212  User-supplied predicates can also be combined using ``all(P0, P1, ...)``.
213
214* ``legalFor()``, ``narrowScalarFor()``, etc. are satisfied if the type matches one in
215  a given set of types. For example ``.legalFor({s16, s32})`` declares the
216  instruction legal if type 0 is either s16 or s32. Additional versions for two
217  and three type indices are generally available. For these, all the type
218  indices considered together must match all the types in one of the tuples. So
219  ``.legalFor({{s16, s32}, {s32, s64}})`` will only accept ``{s16, s32}``, or
220  ``{s32, s64}`` but will not accept ``{s16, s64}``.
221
222* ``legalForTypesWithMemSize()``, ``narrowScalarForTypesWithMemSize()``, etc. are
223  similar to ``legalFor()``, ``narrowScalarFor()``, etc. but additionally require a
224  MachineMemOperand to have a given size in each tuple.
225
226* ``legalForCartesianProduct()``, ``narrowScalarForCartesianProduct()``, etc. are
227  satisfied if each type index matches one element in each of the independent
228  sets. So ``.legalForCartesianProduct({s16, s32}, {s32, s64})`` will accept
229  ``{s16, s32}``, ``{s16, s64}``, ``{s32, s32}``, and ``{s32, s64}``.
230
231Composite Rules
232"""""""""""""""
233
234There are some composite rules for common situations built out of the above facilities:
235
236* ``widenScalarToNextPow2()`` is like ``widenScalarIf()`` but is satisfied iff the type
237  size in bits is not a power of 2 and selects a target type that is the next
238  largest power of 2.
239
240.. _clampscalar:
241
242* ``minScalar()`` is like ``widenScalarIf()`` but is satisfied iff the type
243  size in bits is smaller than the given minimum and selects the minimum as the
244  target type. Similarly, there is also a ``maxScalar()`` for the maximum and a
245  ``clampScalar()`` to do both at once.
246
247* ``minScalarSameAs()`` is like ``minScalar()`` but the minimum is taken from another
248  type index.
249
250* ``moreElementsToNextMultiple()`` is like ``moreElementsToNextPow2()`` but is based on
251  multiples of X rather than powers of 2.
252
253.. _min-legalizerinfo:
254
255Minimum Rule Set
256^^^^^^^^^^^^^^^^
257
258GlobalISel's legalizer has a great deal of flexibility in how a given target
259shapes the GMIR that the rest of the backend must handle. However, there are
260a small number of requirements that all targets must meet.
261
262Before discussing the minimum requirements, we'll need some terminology:
263
264Producer Type Set
265  The set of types which is the union of all possible types produced by at
266  least one legal instruction.
267
268Consumer Type Set
269  The set of types which is the union of all possible types consumed by at
270  least one legal instruction.
271
272Both sets are often identical but there's no guarantee of that. For example,
273it's not uncommon to be unable to consume s64 but still be able to produce it
274for a few specific instructions.
275
276Minimum Rules For Scalars
277"""""""""""""""""""""""""
278
279* G_ANYEXT must be legal for all inputs from the producer type set and all larger
280  outputs from the consumer type set.
281* G_TRUNC must be legal for all inputs from the producer type set and all
282  smaller outputs from the consumer type set.
283
284G_ANYEXT, and G_TRUNC have mandatory legality since the GMIR requires a means to
285connect operations with different type sizes. They are usually trivial to support
286since G_ANYEXT doesn't define the value of the additional bits and G_TRUNC is
287discarding bits. The other conversions can be lowered into G_ANYEXT/G_TRUNC
288with some additional operations that are subject to further legalization. For
289example, G_SEXT can lower to::
290
291  %1 = G_ANYEXT %0
292  %2 = G_CONSTANT ...
293  %3 = G_SHL %1, %2
294  %4 = G_ASHR %3, %2
295
296and the G_CONSTANT/G_SHL/G_ASHR can further lower to other operations or target
297instructions. Similarly, G_FPEXT has no legality requirement since it can lower
298to a G_ANYEXT followed by a target instruction.
299
300G_MERGE_VALUES and G_UNMERGE_VALUES do not have legality requirements since the
301former can lower to G_ANYEXT and some other legalizable instructions, while the
302latter can lower to some legalizable instructions followed by G_TRUNC.
303
304Minimum Legality For Vectors
305""""""""""""""""""""""""""""
306
307Within the vector types, there aren't any defined conversions in LLVM IR as
308vectors are often converted by reinterpreting the bits or by decomposing the
309vector and reconstituting it as a different type. As such, G_BITCAST is the
310only operation to account for. We generally don't require that it's legal
311because it can usually be lowered to COPY (or to nothing using
312replaceAllUses()). However, there are situations where G_BITCAST is non-trivial
313(e.g. little-endian vectors of big-endian data such as on big-endian MIPS MSA and
314big-endian ARM NEON, see `_i_bitcast`). To account for this G_BITCAST must be
315legal for all type combinations that change the bit pattern in the value.
316
317There are no legality requirements for G_BUILD_VECTOR, or G_BUILD_VECTOR_TRUNC
318since these can be handled by:
319* Declaring them legal.
320* Scalarizing them.
321* Lowering them to G_TRUNC+G_ANYEXT and some legalizable instructions.
322* Lowering them to target instructions which are legal by definition.
323
324The same reasoning also allows G_UNMERGE_VALUES to lack legality requirements
325for vector inputs.
326
327Minimum Legality for Pointers
328"""""""""""""""""""""""""""""
329
330There are no minimum rules for pointers since G_INTTOPTR and G_PTRTOINT can
331be selected to a COPY from register class to another by the legalizer.
332
333Minimum Legality For Operations
334"""""""""""""""""""""""""""""""
335
336The rules for G_ANYEXT, G_MERGE_VALUES, G_BITCAST, G_BUILD_VECTOR,
337G_BUILD_VECTOR_TRUNC, G_CONCAT_VECTORS, G_UNMERGE_VALUES, G_PTRTOINT, and
338G_INTTOPTR have already been noted above. In addition to those, the following
339operations have requirements:
340
341* For every type that can be produced by any instruction, G_IMPLICIT_DEF must be
342  legal.
343* G_PHI must be legal for all types in the producer and consumer typesets. This
344  is usually trivial as it requires no code to be selected.
345* At least one G_FRAME_INDEX must be legal
346* At least one G_BLOCK_ADDR must be legal
347
348There are many other operations you'd expect to have legality requirements but
349they can be lowered to target instructions which are legal by definition.
350