1
//===- TargetSchedule.td - Target Independent Scheduling ---*- tablegen -*-===//
3
// The LLVM Compiler Infrastructure
5
// This file is distributed under the University of Illinois Open Source
6
// License. See LICENSE.TXT for details.
8
//===----------------------------------------------------------------------===//
10
// This file defines the target-independent scheduling interfaces which should
11
// be implemented by each target which is using TableGen based scheduling.
13
// The SchedMachineModel is defined by subtargets for three categories of data:
14
// 1. Basic properties for coarse grained instruction cost model.
15
// 2. Scheduler Read/Write resources for simple per-opcode cost model.
16
// 3. Instruction itineraties for detailed reservation tables.
18
// (1) Basic properties are defined by the SchedMachineModel
19
// class. Target hooks allow subtargets to associate opcodes with
22
// (2) A per-operand machine model can be implemented in any
23
// combination of the following ways:
25
// A. Associate per-operand SchedReadWrite types with Instructions by
26
// modifying the Instruction definition to inherit from Sched. For
27
// each subtarget, define WriteRes and ReadAdvance to associate
28
// processor resources and latency with each SchedReadWrite type.
30
// B. In each instruction definition, name an ItineraryClass. For each
31
// subtarget, define ItinRW entries to map ItineraryClass to
32
// per-operand SchedReadWrite types. Unlike method A, these types may
33
// be subtarget specific and can be directly associated with resources
34
// by defining SchedWriteRes and SchedReadAdvance.
36
// C. In the subtarget, map SchedReadWrite types to specific
37
// opcodes. This overrides any SchedReadWrite types or
38
// ItineraryClasses defined by the Instruction. As in method B, the
39
// subtarget can directly associate resources with SchedReadWrite
40
// types by defining SchedWriteRes and SchedReadAdvance.
42
// D. In either the target or subtarget, define SchedWriteVariant or
43
// SchedReadVariant to map one SchedReadWrite type onto another
44
// sequence of SchedReadWrite types. This allows dynamic selection of
45
// an instruction's machine model via custom C++ code. It also allows
46
// a machine-independent SchedReadWrite type to map to a sequence of
47
// machine-dependent types.
49
// (3) A per-pipeline-stage machine model can be implemented by providing
50
// Itineraries in addition to mapping instructions to ItineraryClasses.
51
//===----------------------------------------------------------------------===//
53
// Include legacy support for instruction itineraries.
54
include "llvm/Target/TargetItinerary.td"
56
class Instruction; // Forward def
58
// DAG operator that interprets the DAG args as Instruction defs.
61
// DAG operator that interprets each DAG arg as a regex pattern for
62
// matching Instruction opcode names.
63
// The regex must match the beginning of the opcode (as in Python re.match).
64
// To avoid matching prefixes, append '$' to the pattern.
67
// Define the SchedMachineModel and provide basic properties for
68
// coarse grained instruction cost model. Default values for the
69
// properties are defined in MCSchedModel. A value of "-1" in the
70
// target description's SchedMachineModel indicates that the property
71
// is not overriden by the target.
73
// Target hooks allow subtargets to associate LoadLatency and
74
// HighLatency with groups of opcodes.
76
// See MCSchedule.h for detailed comments.
77
class SchedMachineModel {
78
int IssueWidth = -1; // Max micro-ops that may be scheduled per cycle.
79
int MinLatency = -1; // Determines which instructions are allowed in a group.
80
// (-1) inorder (0) ooo, (1): inorder +var latencies.
81
int MicroOpBufferSize = -1; // Max micro-ops that can be buffered.
82
int LoopMicroOpBufferSize = -1; // Max micro-ops that can be buffered for
83
// optimized loop dispatch/execution.
84
int LoadLatency = -1; // Cycles for loads to access the cache.
85
int HighLatency = -1; // Approximation of cycles for "high latency" ops.
86
int MispredictPenalty = -1; // Extra cycles for a mispredicted branch.
88
// Per-cycle resources tables.
89
ProcessorItineraries Itineraries = NoItineraries;
91
bit PostRAScheduler = 0; // Enable Post RegAlloc Scheduler pass.
93
// Subtargets that define a model for only a subset of instructions
94
// that have a scheduling class (itinerary class or SchedRW list)
95
// and may actually be generated for that subtarget must clear this
96
// bit. Otherwise, the scheduler considers an unmodelled opcode to
97
// be an error. This should only be set during initial bringup,
98
// or there will be no way to catch simple errors in the model
99
// resulting from changes to the instruction definitions.
100
bit CompleteModel = 1;
102
bit NoModel = 0; // Special tag to indicate missing machine model.
105
def NoSchedModel : SchedMachineModel {
109
// Define a kind of processor resource that may be common across
110
// similar subtargets.
111
class ProcResourceKind;
113
// Define a number of interchangeable processor resources. NumUnits
114
// determines the throughput of instructions that require the resource.
116
// An optional Super resource may be given to model these resources as
117
// a subset of the more general super resources. Using one of these
118
// resources implies using one of the super resoruces.
120
// ProcResourceUnits normally model a few buffered resources within an
121
// out-of-order engine. Buffered resources may be held for multiple
122
// clock cycles, but the scheduler does not pin them to a particular
123
// clock cycle relative to instruction dispatch. Setting BufferSize=0
124
// changes this to an in-order issue/dispatch resource. In this case,
125
// the scheduler counts down from the cycle that the instruction
126
// issues in-order, forcing a stall whenever a subsequent instruction
127
// requires the same resource until the number of ResourceCyles
128
// specified in WriteRes expire. Setting BufferSize=1 changes this to
129
// an in-order latency resource. In this case, the scheduler models
130
// producer/consumer stalls between instructions that use the
133
// Examples (all assume an out-of-order engine):
135
// Use BufferSize = -1 for "issue ports" fed by a unified reservation
136
// station. Here the size of the reservation station is modeled by
137
// MicroOpBufferSize, which should be the minimum size of either the
138
// register rename pool, unified reservation station, or reorder
141
// Use BufferSize = 0 for resources that force "dispatch/issue
142
// groups". (Different processors define dispath/issue
143
// differently. Here we refer to stage between decoding into micro-ops
144
// and moving them into a reservation station.) Normally NumMicroOps
145
// is sufficient to limit dispatch/issue groups. However, some
146
// processors can form groups of with only certain combinitions of
147
// instruction types. e.g. POWER7.
149
// Use BufferSize = 1 for in-order execution units. This is used for
150
// an in-order pipeline within an out-of-order core where scheduling
151
// dependent operations back-to-back is guaranteed to cause a
152
// bubble. e.g. Cortex-a9 floating-point.
154
// Use BufferSize > 1 for out-of-order executions units with a
155
// separate reservation station. This simply models the size of the
156
// reservation station.
158
// To model both dispatch/issue groups and in-order execution units,
159
// create two types of units, one with BufferSize=0 and one with
162
// SchedModel ties these units to a processor for any stand-alone defs
163
// of this class. Instances of subclass ProcResource will be automatically
164
// attached to a processor, so SchedModel is not needed.
165
class ProcResourceUnits<ProcResourceKind kind, int num> {
166
ProcResourceKind Kind = kind;
168
ProcResourceKind Super = ?;
170
SchedMachineModel SchedModel = ?;
173
// EponymousProcResourceKind helps implement ProcResourceUnits by
174
// allowing a ProcResourceUnits definition to reference itself. It
175
// should not be referenced anywhere else.
176
def EponymousProcResourceKind : ProcResourceKind;
178
// Subtargets typically define processor resource kind and number of
179
// units in one place.
180
class ProcResource<int num> : ProcResourceKind,
181
ProcResourceUnits<EponymousProcResourceKind, num>;
183
class ProcResGroup<list<ProcResource> resources> : ProcResourceKind {
184
list<ProcResource> Resources = resources;
185
SchedMachineModel SchedModel = ?;
189
// A target architecture may define SchedReadWrite types and associate
190
// them with instruction operands.
191
class SchedReadWrite;
193
// List the per-operand types that map to the machine model of an
194
// instruction. One SchedWrite type must be listed for each explicit
195
// def operand in order. Additional SchedWrite types may optionally be
196
// listed for implicit def operands. SchedRead types may optionally
197
// be listed for use operands in order. The order of defs relative to
198
// uses is insignificant. This way, the same SchedReadWrite list may
199
// be used for multiple forms of an operation. For example, a
200
// two-address instruction could have two tied operands or single
201
// operand that both reads and writes a reg. In both cases we have a
202
// single SchedWrite and single SchedRead in any order.
203
class Sched<list<SchedReadWrite> schedrw> {
204
list<SchedReadWrite> SchedRW = schedrw;
207
// Define a scheduler resource associated with a def operand.
208
class SchedWrite : SchedReadWrite;
209
def NoWrite : SchedWrite;
211
// Define a scheduler resource associated with a use operand.
212
class SchedRead : SchedReadWrite;
214
// Define a SchedWrite that is modeled as a sequence of other
215
// SchedWrites with additive latency. This allows a single operand to
216
// be mapped the resources composed from a set of previously defined
219
// If the final write in this sequence is a SchedWriteVariant marked
220
// Variadic, then the list of prior writes are distributed across all
221
// operands after resolving the predicate for the final write.
223
// SchedModel silences warnings but is ignored.
224
class WriteSequence<list<SchedWrite> writes, int rep = 1> : SchedWrite {
225
list<SchedWrite> Writes = writes;
227
SchedMachineModel SchedModel = ?;
230
// Define values common to WriteRes and SchedWriteRes.
232
// SchedModel ties these resources to a processor.
233
class ProcWriteResources<list<ProcResourceKind> resources> {
234
list<ProcResourceKind> ProcResources = resources;
235
list<int> ResourceCycles = [];
240
// Allow a processor to mark some scheduling classes as unsupported
241
// for stronger verification.
243
SchedMachineModel SchedModel = ?;
246
// Define the resources and latency of a SchedWrite. This will be used
247
// directly by targets that have no itinerary classes. In this case,
248
// SchedWrite is defined by the target, while WriteResources is
249
// defined by the subtarget, and maps the SchedWrite to processor
252
// If a target already has itinerary classes, SchedWriteResources can
253
// be used instead to define subtarget specific SchedWrites and map
254
// them to processor resources in one place. Then ItinRW can map
255
// itinerary classes to the subtarget's SchedWrites.
257
// ProcResources indicates the set of resources consumed by the write.
258
// Optionally, ResourceCycles indicates the number of cycles the
259
// resource is consumed. Each ResourceCycles item is paired with the
260
// ProcResource item at the same position in its list. Since
261
// ResourceCycles are rarely specialized, the list may be
262
// incomplete. By default, resources are consumed for a single cycle,
263
// regardless of latency, which models a fully pipelined processing
264
// unit. A value of 0 for ResourceCycles means that the resource must
265
// be available but is not consumed, which is only relevant for
266
// unbuffered resources.
268
// By default, each SchedWrite takes one micro-op, which is counted
269
// against the processor's IssueWidth limit. If an instruction can
270
// write multiple registers with a single micro-op, the subtarget
271
// should define one of the writes to be zero micro-ops. If a
272
// subtarget requires multiple micro-ops to write a single result, it
273
// should either override the write's NumMicroOps to be greater than 1
274
// or require additional writes. Extra writes can be required either
275
// by defining a WriteSequence, or simply listing extra writes in the
276
// instruction's list of writers beyond the number of "def"
277
// operands. The scheduler assumes that all micro-ops must be
278
// dispatched in the same cycle. These micro-ops may be required to
279
// begin or end the current dispatch group.
280
class WriteRes<SchedWrite write, list<ProcResourceKind> resources>
281
: ProcWriteResources<resources> {
282
SchedWrite WriteType = write;
285
// Directly name a set of WriteResources defining a new SchedWrite
286
// type at the same time. This class is unaware of its SchedModel so
287
// must be referenced by InstRW or ItinRW.
288
class SchedWriteRes<list<ProcResourceKind> resources> : SchedWrite,
289
ProcWriteResources<resources>;
291
// Define values common to ReadAdvance and SchedReadAdvance.
293
// SchedModel ties these resources to a processor.
294
class ProcReadAdvance<int cycles, list<SchedWrite> writes = []> {
296
list<SchedWrite> ValidWrites = writes;
297
// Allow a processor to mark some scheduling classes as unsupported
298
// for stronger verification.
300
SchedMachineModel SchedModel = ?;
303
// A processor may define a ReadAdvance associated with a SchedRead
304
// to reduce latency of a prior write by N cycles. A negative advance
305
// effectively increases latency, which may be used for cross-domain
308
// A ReadAdvance may be associated with a list of SchedWrites
309
// to implement pipeline bypass. The Writes list may be empty to
310
// indicate operands that are always read this number of Cycles later
311
// than a normal register read, allowing the read's parent instruction
312
// to issue earlier relative to the writer.
313
class ReadAdvance<SchedRead read, int cycles, list<SchedWrite> writes = []>
314
: ProcReadAdvance<cycles, writes> {
315
SchedRead ReadType = read;
318
// Directly associate a new SchedRead type with a delay and optional
319
// pipeline bypess. For use with InstRW or ItinRW.
320
class SchedReadAdvance<int cycles, list<SchedWrite> writes = []> : SchedRead,
321
ProcReadAdvance<cycles, writes>;
323
// Define SchedRead defaults. Reads seldom need special treatment.
324
def ReadDefault : SchedRead;
325
def NoReadAdvance : SchedReadAdvance<0>;
327
// Define shared code that will be in the same scope as all
328
// SchedPredicates. Available variables are:
329
// (const MachineInstr *MI, const TargetSchedModel *SchedModel)
330
class PredicateProlog<code c> {
334
// Define a predicate to determine which SchedVariant applies to a
335
// particular MachineInstr. The code snippet is used as an
336
// if-statement's expression. Available variables are MI, SchedModel,
337
// and anything defined in a PredicateProlog.
339
// SchedModel silences warnings but is ignored.
340
class SchedPredicate<code pred> {
341
SchedMachineModel SchedModel = ?;
342
code Predicate = pred;
344
def NoSchedPred : SchedPredicate<[{true}]>;
346
// Associate a predicate with a list of SchedReadWrites. By default,
347
// the selected SchedReadWrites are still associated with a single
348
// operand and assumed to execute sequentially with additive
349
// latency. However, if the parent SchedWriteVariant or
350
// SchedReadVariant is marked "Variadic", then each Selected
351
// SchedReadWrite is mapped in place to the instruction's variadic
352
// operands. In this case, latency is not additive. If the current Variant
353
// is already part of a Sequence, then that entire chain leading up to
354
// the Variant is distributed over the variadic operands.
355
class SchedVar<SchedPredicate pred, list<SchedReadWrite> selected> {
356
SchedPredicate Predicate = pred;
357
list<SchedReadWrite> Selected = selected;
360
// SchedModel silences warnings but is ignored.
361
class SchedVariant<list<SchedVar> variants> {
362
list<SchedVar> Variants = variants;
364
SchedMachineModel SchedModel = ?;
367
// A SchedWriteVariant is a single SchedWrite type that maps to a list
368
// of SchedWrite types under the conditions defined by its predicates.
370
// A Variadic write is expanded to cover multiple "def" operands. The
371
// SchedVariant's Expansion list is then interpreted as one write
372
// per-operand instead of the usual sequential writes feeding a single
374
class SchedWriteVariant<list<SchedVar> variants> : SchedWrite,
375
SchedVariant<variants> {
378
// A SchedReadVariant is a single SchedRead type that maps to a list
379
// of SchedRead types under the conditions defined by its predicates.
381
// A Variadic write is expanded to cover multiple "readsReg" operands as
383
class SchedReadVariant<list<SchedVar> variants> : SchedRead,
384
SchedVariant<variants> {
387
// Map a set of opcodes to a list of SchedReadWrite types. This allows
388
// the subtarget to easily override specific operations.
390
// SchedModel ties this opcode mapping to a processor.
391
class InstRW<list<SchedReadWrite> rw, dag instrlist> {
392
list<SchedReadWrite> OperandReadWrites = rw;
393
dag Instrs = instrlist;
394
SchedMachineModel SchedModel = ?;
397
// Map a set of itinerary classes to SchedReadWrite resources. This is
398
// used to bootstrap a target (e.g. ARM) when itineraries already
399
// exist and changing InstrInfo is undesirable.
401
// SchedModel ties this ItineraryClass mapping to a processor.
402
class ItinRW<list<SchedReadWrite> rw, list<InstrItinClass> iic> {
403
list<InstrItinClass> MatchedItinClasses = iic;
404
list<SchedReadWrite> OperandReadWrites = rw;
405
SchedMachineModel SchedModel = ?;
408
// Alias a target-defined SchedReadWrite to a processor specific
409
// SchedReadWrite. This allows a subtarget to easily map a
410
// SchedReadWrite type onto a WriteSequence, SchedWriteVariant, or
413
// SchedModel will usually be provided by surrounding let statement
414
// and ties this SchedAlias mapping to a processor.
415
class SchedAlias<SchedReadWrite match, SchedReadWrite alias> {
416
SchedReadWrite MatchRW = match;
417
SchedReadWrite AliasRW = alias;
418
SchedMachineModel SchedModel = ?;