1 //===- CodeGenSchedule.cpp - Scheduling MachineModels ---------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines structures to encapsulate the machine model as described in
11 // the target description.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "CodeGenSchedule.h"
16 #include "CodeGenTarget.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/Support/Debug.h"
19 #include "llvm/Support/Regex.h"
20 #include "llvm/TableGen/Error.h"
21
22 using namespace llvm;
23
24 #define DEBUG_TYPE "subtarget-emitter"
25
26 #ifndef NDEBUG
dumpIdxVec(ArrayRef<unsigned> V)27 static void dumpIdxVec(ArrayRef<unsigned> V) {
28 for (unsigned Idx : V)
29 dbgs() << Idx << ", ";
30 }
31 #endif
32
33 namespace {
34 // (instrs a, b, ...) Evaluate and union all arguments. Identical to AddOp.
35 struct InstrsOp : public SetTheory::Operator {
apply__anon50a6a0fc0111::InstrsOp36 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
37 ArrayRef<SMLoc> Loc) override {
38 ST.evaluate(Expr->arg_begin(), Expr->arg_end(), Elts, Loc);
39 }
40 };
41
42 // (instregex "OpcPat",...) Find all instructions matching an opcode pattern.
43 //
44 // TODO: Since this is a prefix match, perform a binary search over the
45 // instruction names using lower_bound. Note that the predefined instrs must be
46 // scanned linearly first. However, this is only safe if the regex pattern has
47 // no top-level bars. The DAG already has a list of patterns, so there's no
48 // reason to use top-level bars, but we need a way to verify they don't exist
49 // before implementing the optimization.
50 struct InstRegexOp : public SetTheory::Operator {
51 const CodeGenTarget &Target;
InstRegexOp__anon50a6a0fc0111::InstRegexOp52 InstRegexOp(const CodeGenTarget &t): Target(t) {}
53
apply__anon50a6a0fc0111::InstRegexOp54 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
55 ArrayRef<SMLoc> Loc) override {
56 SmallVector<Regex, 4> RegexList;
57 for (DagInit::const_arg_iterator
58 AI = Expr->arg_begin(), AE = Expr->arg_end(); AI != AE; ++AI) {
59 StringInit *SI = dyn_cast<StringInit>(*AI);
60 if (!SI)
61 PrintFatalError(Loc, "instregex requires pattern string: "
62 + Expr->getAsString());
63 std::string pat = SI->getValue();
64 // Implement a python-style prefix match.
65 if (pat[0] != '^') {
66 pat.insert(0, "^(");
67 pat.insert(pat.end(), ')');
68 }
69 RegexList.push_back(Regex(pat));
70 }
71 for (const CodeGenInstruction *Inst : Target.getInstructionsByEnumValue()) {
72 for (auto &R : RegexList) {
73 if (R.match(Inst->TheDef->getName()))
74 Elts.insert(Inst->TheDef);
75 }
76 }
77 }
78 };
79 } // end anonymous namespace
80
81 /// CodeGenModels ctor interprets machine model records and populates maps.
CodeGenSchedModels(RecordKeeper & RK,const CodeGenTarget & TGT)82 CodeGenSchedModels::CodeGenSchedModels(RecordKeeper &RK,
83 const CodeGenTarget &TGT):
84 Records(RK), Target(TGT) {
85
86 Sets.addFieldExpander("InstRW", "Instrs");
87
88 // Allow Set evaluation to recognize the dags used in InstRW records:
89 // (instrs Op1, Op1...)
90 Sets.addOperator("instrs", llvm::make_unique<InstrsOp>());
91 Sets.addOperator("instregex", llvm::make_unique<InstRegexOp>(Target));
92
93 // Instantiate a CodeGenProcModel for each SchedMachineModel with the values
94 // that are explicitly referenced in tablegen records. Resources associated
95 // with each processor will be derived later. Populate ProcModelMap with the
96 // CodeGenProcModel instances.
97 collectProcModels();
98
99 // Instantiate a CodeGenSchedRW for each SchedReadWrite record explicitly
100 // defined, and populate SchedReads and SchedWrites vectors. Implicit
101 // SchedReadWrites that represent sequences derived from expanded variant will
102 // be inferred later.
103 collectSchedRW();
104
105 // Instantiate a CodeGenSchedClass for each unique SchedRW signature directly
106 // required by an instruction definition, and populate SchedClassIdxMap. Set
107 // NumItineraryClasses to the number of explicit itinerary classes referenced
108 // by instructions. Set NumInstrSchedClasses to the number of itinerary
109 // classes plus any classes implied by instructions that derive from class
110 // Sched and provide SchedRW list. This does not infer any new classes from
111 // SchedVariant.
112 collectSchedClasses();
113
114 // Find instruction itineraries for each processor. Sort and populate
115 // CodeGenProcModel::ItinDefList. (Cycle-to-cycle itineraries). This requires
116 // all itinerary classes to be discovered.
117 collectProcItins();
118
119 // Find ItinRW records for each processor and itinerary class.
120 // (For per-operand resources mapped to itinerary classes).
121 collectProcItinRW();
122
123 // Find UnsupportedFeatures records for each processor.
124 // (For per-operand resources mapped to itinerary classes).
125 collectProcUnsupportedFeatures();
126
127 // Infer new SchedClasses from SchedVariant.
128 inferSchedClasses();
129
130 // Populate each CodeGenProcModel's WriteResDefs, ReadAdvanceDefs, and
131 // ProcResourceDefs.
132 collectProcResources();
133
134 checkCompleteness();
135 }
136
137 /// Gather all processor models.
collectProcModels()138 void CodeGenSchedModels::collectProcModels() {
139 RecVec ProcRecords = Records.getAllDerivedDefinitions("Processor");
140 std::sort(ProcRecords.begin(), ProcRecords.end(), LessRecordFieldName());
141
142 // Reserve space because we can. Reallocation would be ok.
143 ProcModels.reserve(ProcRecords.size()+1);
144
145 // Use idx=0 for NoModel/NoItineraries.
146 Record *NoModelDef = Records.getDef("NoSchedModel");
147 Record *NoItinsDef = Records.getDef("NoItineraries");
148 ProcModels.emplace_back(0, "NoSchedModel", NoModelDef, NoItinsDef);
149 ProcModelMap[NoModelDef] = 0;
150
151 // For each processor, find a unique machine model.
152 for (unsigned i = 0, N = ProcRecords.size(); i < N; ++i)
153 addProcModel(ProcRecords[i]);
154 }
155
156 /// Get a unique processor model based on the defined MachineModel and
157 /// ProcessorItineraries.
addProcModel(Record * ProcDef)158 void CodeGenSchedModels::addProcModel(Record *ProcDef) {
159 Record *ModelKey = getModelOrItinDef(ProcDef);
160 if (!ProcModelMap.insert(std::make_pair(ModelKey, ProcModels.size())).second)
161 return;
162
163 std::string Name = ModelKey->getName();
164 if (ModelKey->isSubClassOf("SchedMachineModel")) {
165 Record *ItinsDef = ModelKey->getValueAsDef("Itineraries");
166 ProcModels.emplace_back(ProcModels.size(), Name, ModelKey, ItinsDef);
167 }
168 else {
169 // An itinerary is defined without a machine model. Infer a new model.
170 if (!ModelKey->getValueAsListOfDefs("IID").empty())
171 Name = Name + "Model";
172 ProcModels.emplace_back(ProcModels.size(), Name,
173 ProcDef->getValueAsDef("SchedModel"), ModelKey);
174 }
175 DEBUG(ProcModels.back().dump());
176 }
177
178 // Recursively find all reachable SchedReadWrite records.
scanSchedRW(Record * RWDef,RecVec & RWDefs,SmallPtrSet<Record *,16> & RWSet)179 static void scanSchedRW(Record *RWDef, RecVec &RWDefs,
180 SmallPtrSet<Record*, 16> &RWSet) {
181 if (!RWSet.insert(RWDef).second)
182 return;
183 RWDefs.push_back(RWDef);
184 // Reads don't current have sequence records, but it can be added later.
185 if (RWDef->isSubClassOf("WriteSequence")) {
186 RecVec Seq = RWDef->getValueAsListOfDefs("Writes");
187 for (RecIter I = Seq.begin(), E = Seq.end(); I != E; ++I)
188 scanSchedRW(*I, RWDefs, RWSet);
189 }
190 else if (RWDef->isSubClassOf("SchedVariant")) {
191 // Visit each variant (guarded by a different predicate).
192 RecVec Vars = RWDef->getValueAsListOfDefs("Variants");
193 for (RecIter VI = Vars.begin(), VE = Vars.end(); VI != VE; ++VI) {
194 // Visit each RW in the sequence selected by the current variant.
195 RecVec Selected = (*VI)->getValueAsListOfDefs("Selected");
196 for (RecIter I = Selected.begin(), E = Selected.end(); I != E; ++I)
197 scanSchedRW(*I, RWDefs, RWSet);
198 }
199 }
200 }
201
202 // Collect and sort all SchedReadWrites reachable via tablegen records.
203 // More may be inferred later when inferring new SchedClasses from variants.
collectSchedRW()204 void CodeGenSchedModels::collectSchedRW() {
205 // Reserve idx=0 for invalid writes/reads.
206 SchedWrites.resize(1);
207 SchedReads.resize(1);
208
209 SmallPtrSet<Record*, 16> RWSet;
210
211 // Find all SchedReadWrites referenced by instruction defs.
212 RecVec SWDefs, SRDefs;
213 for (const CodeGenInstruction *Inst : Target.getInstructionsByEnumValue()) {
214 Record *SchedDef = Inst->TheDef;
215 if (SchedDef->isValueUnset("SchedRW"))
216 continue;
217 RecVec RWs = SchedDef->getValueAsListOfDefs("SchedRW");
218 for (RecIter RWI = RWs.begin(), RWE = RWs.end(); RWI != RWE; ++RWI) {
219 if ((*RWI)->isSubClassOf("SchedWrite"))
220 scanSchedRW(*RWI, SWDefs, RWSet);
221 else {
222 assert((*RWI)->isSubClassOf("SchedRead") && "Unknown SchedReadWrite");
223 scanSchedRW(*RWI, SRDefs, RWSet);
224 }
225 }
226 }
227 // Find all ReadWrites referenced by InstRW.
228 RecVec InstRWDefs = Records.getAllDerivedDefinitions("InstRW");
229 for (RecIter OI = InstRWDefs.begin(), OE = InstRWDefs.end(); OI != OE; ++OI) {
230 // For all OperandReadWrites.
231 RecVec RWDefs = (*OI)->getValueAsListOfDefs("OperandReadWrites");
232 for (RecIter RWI = RWDefs.begin(), RWE = RWDefs.end();
233 RWI != RWE; ++RWI) {
234 if ((*RWI)->isSubClassOf("SchedWrite"))
235 scanSchedRW(*RWI, SWDefs, RWSet);
236 else {
237 assert((*RWI)->isSubClassOf("SchedRead") && "Unknown SchedReadWrite");
238 scanSchedRW(*RWI, SRDefs, RWSet);
239 }
240 }
241 }
242 // Find all ReadWrites referenced by ItinRW.
243 RecVec ItinRWDefs = Records.getAllDerivedDefinitions("ItinRW");
244 for (RecIter II = ItinRWDefs.begin(), IE = ItinRWDefs.end(); II != IE; ++II) {
245 // For all OperandReadWrites.
246 RecVec RWDefs = (*II)->getValueAsListOfDefs("OperandReadWrites");
247 for (RecIter RWI = RWDefs.begin(), RWE = RWDefs.end();
248 RWI != RWE; ++RWI) {
249 if ((*RWI)->isSubClassOf("SchedWrite"))
250 scanSchedRW(*RWI, SWDefs, RWSet);
251 else {
252 assert((*RWI)->isSubClassOf("SchedRead") && "Unknown SchedReadWrite");
253 scanSchedRW(*RWI, SRDefs, RWSet);
254 }
255 }
256 }
257 // Find all ReadWrites referenced by SchedAlias. AliasDefs needs to be sorted
258 // for the loop below that initializes Alias vectors.
259 RecVec AliasDefs = Records.getAllDerivedDefinitions("SchedAlias");
260 std::sort(AliasDefs.begin(), AliasDefs.end(), LessRecord());
261 for (RecIter AI = AliasDefs.begin(), AE = AliasDefs.end(); AI != AE; ++AI) {
262 Record *MatchDef = (*AI)->getValueAsDef("MatchRW");
263 Record *AliasDef = (*AI)->getValueAsDef("AliasRW");
264 if (MatchDef->isSubClassOf("SchedWrite")) {
265 if (!AliasDef->isSubClassOf("SchedWrite"))
266 PrintFatalError((*AI)->getLoc(), "SchedWrite Alias must be SchedWrite");
267 scanSchedRW(AliasDef, SWDefs, RWSet);
268 }
269 else {
270 assert(MatchDef->isSubClassOf("SchedRead") && "Unknown SchedReadWrite");
271 if (!AliasDef->isSubClassOf("SchedRead"))
272 PrintFatalError((*AI)->getLoc(), "SchedRead Alias must be SchedRead");
273 scanSchedRW(AliasDef, SRDefs, RWSet);
274 }
275 }
276 // Sort and add the SchedReadWrites directly referenced by instructions or
277 // itinerary resources. Index reads and writes in separate domains.
278 std::sort(SWDefs.begin(), SWDefs.end(), LessRecord());
279 for (RecIter SWI = SWDefs.begin(), SWE = SWDefs.end(); SWI != SWE; ++SWI) {
280 assert(!getSchedRWIdx(*SWI, /*IsRead=*/false) && "duplicate SchedWrite");
281 SchedWrites.emplace_back(SchedWrites.size(), *SWI);
282 }
283 std::sort(SRDefs.begin(), SRDefs.end(), LessRecord());
284 for (RecIter SRI = SRDefs.begin(), SRE = SRDefs.end(); SRI != SRE; ++SRI) {
285 assert(!getSchedRWIdx(*SRI, /*IsRead-*/true) && "duplicate SchedWrite");
286 SchedReads.emplace_back(SchedReads.size(), *SRI);
287 }
288 // Initialize WriteSequence vectors.
289 for (std::vector<CodeGenSchedRW>::iterator WI = SchedWrites.begin(),
290 WE = SchedWrites.end(); WI != WE; ++WI) {
291 if (!WI->IsSequence)
292 continue;
293 findRWs(WI->TheDef->getValueAsListOfDefs("Writes"), WI->Sequence,
294 /*IsRead=*/false);
295 }
296 // Initialize Aliases vectors.
297 for (RecIter AI = AliasDefs.begin(), AE = AliasDefs.end(); AI != AE; ++AI) {
298 Record *AliasDef = (*AI)->getValueAsDef("AliasRW");
299 getSchedRW(AliasDef).IsAlias = true;
300 Record *MatchDef = (*AI)->getValueAsDef("MatchRW");
301 CodeGenSchedRW &RW = getSchedRW(MatchDef);
302 if (RW.IsAlias)
303 PrintFatalError((*AI)->getLoc(), "Cannot Alias an Alias");
304 RW.Aliases.push_back(*AI);
305 }
306 DEBUG(
307 for (unsigned WIdx = 0, WEnd = SchedWrites.size(); WIdx != WEnd; ++WIdx) {
308 dbgs() << WIdx << ": ";
309 SchedWrites[WIdx].dump();
310 dbgs() << '\n';
311 }
312 for (unsigned RIdx = 0, REnd = SchedReads.size(); RIdx != REnd; ++RIdx) {
313 dbgs() << RIdx << ": ";
314 SchedReads[RIdx].dump();
315 dbgs() << '\n';
316 }
317 RecVec RWDefs = Records.getAllDerivedDefinitions("SchedReadWrite");
318 for (RecIter RI = RWDefs.begin(), RE = RWDefs.end();
319 RI != RE; ++RI) {
320 if (!getSchedRWIdx(*RI, (*RI)->isSubClassOf("SchedRead"))) {
321 const std::string &Name = (*RI)->getName();
322 if (Name != "NoWrite" && Name != "ReadDefault")
323 dbgs() << "Unused SchedReadWrite " << (*RI)->getName() << '\n';
324 }
325 });
326 }
327
328 /// Compute a SchedWrite name from a sequence of writes.
genRWName(ArrayRef<unsigned> Seq,bool IsRead)329 std::string CodeGenSchedModels::genRWName(ArrayRef<unsigned> Seq, bool IsRead) {
330 std::string Name("(");
331 for (auto I = Seq.begin(), E = Seq.end(); I != E; ++I) {
332 if (I != Seq.begin())
333 Name += '_';
334 Name += getSchedRW(*I, IsRead).Name;
335 }
336 Name += ')';
337 return Name;
338 }
339
getSchedRWIdx(Record * Def,bool IsRead,unsigned After) const340 unsigned CodeGenSchedModels::getSchedRWIdx(Record *Def, bool IsRead,
341 unsigned After) const {
342 const std::vector<CodeGenSchedRW> &RWVec = IsRead ? SchedReads : SchedWrites;
343 assert(After < RWVec.size() && "start position out of bounds");
344 for (std::vector<CodeGenSchedRW>::const_iterator I = RWVec.begin() + After,
345 E = RWVec.end(); I != E; ++I) {
346 if (I->TheDef == Def)
347 return I - RWVec.begin();
348 }
349 return 0;
350 }
351
hasReadOfWrite(Record * WriteDef) const352 bool CodeGenSchedModels::hasReadOfWrite(Record *WriteDef) const {
353 for (unsigned i = 0, e = SchedReads.size(); i < e; ++i) {
354 Record *ReadDef = SchedReads[i].TheDef;
355 if (!ReadDef || !ReadDef->isSubClassOf("ProcReadAdvance"))
356 continue;
357
358 RecVec ValidWrites = ReadDef->getValueAsListOfDefs("ValidWrites");
359 if (std::find(ValidWrites.begin(), ValidWrites.end(), WriteDef)
360 != ValidWrites.end()) {
361 return true;
362 }
363 }
364 return false;
365 }
366
367 namespace llvm {
splitSchedReadWrites(const RecVec & RWDefs,RecVec & WriteDefs,RecVec & ReadDefs)368 void splitSchedReadWrites(const RecVec &RWDefs,
369 RecVec &WriteDefs, RecVec &ReadDefs) {
370 for (RecIter RWI = RWDefs.begin(), RWE = RWDefs.end(); RWI != RWE; ++RWI) {
371 if ((*RWI)->isSubClassOf("SchedWrite"))
372 WriteDefs.push_back(*RWI);
373 else {
374 assert((*RWI)->isSubClassOf("SchedRead") && "unknown SchedReadWrite");
375 ReadDefs.push_back(*RWI);
376 }
377 }
378 }
379 } // namespace llvm
380
381 // Split the SchedReadWrites defs and call findRWs for each list.
findRWs(const RecVec & RWDefs,IdxVec & Writes,IdxVec & Reads) const382 void CodeGenSchedModels::findRWs(const RecVec &RWDefs,
383 IdxVec &Writes, IdxVec &Reads) const {
384 RecVec WriteDefs;
385 RecVec ReadDefs;
386 splitSchedReadWrites(RWDefs, WriteDefs, ReadDefs);
387 findRWs(WriteDefs, Writes, false);
388 findRWs(ReadDefs, Reads, true);
389 }
390
391 // Call getSchedRWIdx for all elements in a sequence of SchedRW defs.
findRWs(const RecVec & RWDefs,IdxVec & RWs,bool IsRead) const392 void CodeGenSchedModels::findRWs(const RecVec &RWDefs, IdxVec &RWs,
393 bool IsRead) const {
394 for (RecIter RI = RWDefs.begin(), RE = RWDefs.end(); RI != RE; ++RI) {
395 unsigned Idx = getSchedRWIdx(*RI, IsRead);
396 assert(Idx && "failed to collect SchedReadWrite");
397 RWs.push_back(Idx);
398 }
399 }
400
expandRWSequence(unsigned RWIdx,IdxVec & RWSeq,bool IsRead) const401 void CodeGenSchedModels::expandRWSequence(unsigned RWIdx, IdxVec &RWSeq,
402 bool IsRead) const {
403 const CodeGenSchedRW &SchedRW = getSchedRW(RWIdx, IsRead);
404 if (!SchedRW.IsSequence) {
405 RWSeq.push_back(RWIdx);
406 return;
407 }
408 int Repeat =
409 SchedRW.TheDef ? SchedRW.TheDef->getValueAsInt("Repeat") : 1;
410 for (int i = 0; i < Repeat; ++i) {
411 for (IdxIter I = SchedRW.Sequence.begin(), E = SchedRW.Sequence.end();
412 I != E; ++I) {
413 expandRWSequence(*I, RWSeq, IsRead);
414 }
415 }
416 }
417
418 // Expand a SchedWrite as a sequence following any aliases that coincide with
419 // the given processor model.
expandRWSeqForProc(unsigned RWIdx,IdxVec & RWSeq,bool IsRead,const CodeGenProcModel & ProcModel) const420 void CodeGenSchedModels::expandRWSeqForProc(
421 unsigned RWIdx, IdxVec &RWSeq, bool IsRead,
422 const CodeGenProcModel &ProcModel) const {
423
424 const CodeGenSchedRW &SchedWrite = getSchedRW(RWIdx, IsRead);
425 Record *AliasDef = nullptr;
426 for (RecIter AI = SchedWrite.Aliases.begin(), AE = SchedWrite.Aliases.end();
427 AI != AE; ++AI) {
428 const CodeGenSchedRW &AliasRW = getSchedRW((*AI)->getValueAsDef("AliasRW"));
429 if ((*AI)->getValueInit("SchedModel")->isComplete()) {
430 Record *ModelDef = (*AI)->getValueAsDef("SchedModel");
431 if (&getProcModel(ModelDef) != &ProcModel)
432 continue;
433 }
434 if (AliasDef)
435 PrintFatalError(AliasRW.TheDef->getLoc(), "Multiple aliases "
436 "defined for processor " + ProcModel.ModelName +
437 " Ensure only one SchedAlias exists per RW.");
438 AliasDef = AliasRW.TheDef;
439 }
440 if (AliasDef) {
441 expandRWSeqForProc(getSchedRWIdx(AliasDef, IsRead),
442 RWSeq, IsRead,ProcModel);
443 return;
444 }
445 if (!SchedWrite.IsSequence) {
446 RWSeq.push_back(RWIdx);
447 return;
448 }
449 int Repeat =
450 SchedWrite.TheDef ? SchedWrite.TheDef->getValueAsInt("Repeat") : 1;
451 for (int i = 0; i < Repeat; ++i) {
452 for (IdxIter I = SchedWrite.Sequence.begin(), E = SchedWrite.Sequence.end();
453 I != E; ++I) {
454 expandRWSeqForProc(*I, RWSeq, IsRead, ProcModel);
455 }
456 }
457 }
458
459 // Find the existing SchedWrite that models this sequence of writes.
findRWForSequence(ArrayRef<unsigned> Seq,bool IsRead)460 unsigned CodeGenSchedModels::findRWForSequence(ArrayRef<unsigned> Seq,
461 bool IsRead) {
462 std::vector<CodeGenSchedRW> &RWVec = IsRead ? SchedReads : SchedWrites;
463
464 for (std::vector<CodeGenSchedRW>::iterator I = RWVec.begin(), E = RWVec.end();
465 I != E; ++I) {
466 if (makeArrayRef(I->Sequence) == Seq)
467 return I - RWVec.begin();
468 }
469 // Index zero reserved for invalid RW.
470 return 0;
471 }
472
473 /// Add this ReadWrite if it doesn't already exist.
findOrInsertRW(ArrayRef<unsigned> Seq,bool IsRead)474 unsigned CodeGenSchedModels::findOrInsertRW(ArrayRef<unsigned> Seq,
475 bool IsRead) {
476 assert(!Seq.empty() && "cannot insert empty sequence");
477 if (Seq.size() == 1)
478 return Seq.back();
479
480 unsigned Idx = findRWForSequence(Seq, IsRead);
481 if (Idx)
482 return Idx;
483
484 unsigned RWIdx = IsRead ? SchedReads.size() : SchedWrites.size();
485 CodeGenSchedRW SchedRW(RWIdx, IsRead, Seq, genRWName(Seq, IsRead));
486 if (IsRead)
487 SchedReads.push_back(SchedRW);
488 else
489 SchedWrites.push_back(SchedRW);
490 return RWIdx;
491 }
492
493 /// Visit all the instruction definitions for this target to gather and
494 /// enumerate the itinerary classes. These are the explicitly specified
495 /// SchedClasses. More SchedClasses may be inferred.
collectSchedClasses()496 void CodeGenSchedModels::collectSchedClasses() {
497
498 // NoItinerary is always the first class at Idx=0
499 SchedClasses.resize(1);
500 SchedClasses.back().Index = 0;
501 SchedClasses.back().Name = "NoInstrModel";
502 SchedClasses.back().ItinClassDef = Records.getDef("NoItinerary");
503 SchedClasses.back().ProcIndices.push_back(0);
504
505 // Create a SchedClass for each unique combination of itinerary class and
506 // SchedRW list.
507 for (const CodeGenInstruction *Inst : Target.getInstructionsByEnumValue()) {
508 Record *ItinDef = Inst->TheDef->getValueAsDef("Itinerary");
509 IdxVec Writes, Reads;
510 if (!Inst->TheDef->isValueUnset("SchedRW"))
511 findRWs(Inst->TheDef->getValueAsListOfDefs("SchedRW"), Writes, Reads);
512
513 // ProcIdx == 0 indicates the class applies to all processors.
514 IdxVec ProcIndices(1, 0);
515
516 unsigned SCIdx = addSchedClass(ItinDef, Writes, Reads, ProcIndices);
517 InstrClassMap[Inst->TheDef] = SCIdx;
518 }
519 // Create classes for InstRW defs.
520 RecVec InstRWDefs = Records.getAllDerivedDefinitions("InstRW");
521 std::sort(InstRWDefs.begin(), InstRWDefs.end(), LessRecord());
522 for (RecIter OI = InstRWDefs.begin(), OE = InstRWDefs.end(); OI != OE; ++OI)
523 createInstRWClass(*OI);
524
525 NumInstrSchedClasses = SchedClasses.size();
526
527 bool EnableDump = false;
528 DEBUG(EnableDump = true);
529 if (!EnableDump)
530 return;
531
532 for (const CodeGenInstruction *Inst : Target.getInstructionsByEnumValue()) {
533 std::string InstName = Inst->TheDef->getName();
534 unsigned SCIdx = InstrClassMap.lookup(Inst->TheDef);
535 if (!SCIdx) {
536 if (!Inst->hasNoSchedulingInfo)
537 dbgs() << "No machine model for " << Inst->TheDef->getName() << '\n';
538 continue;
539 }
540 CodeGenSchedClass &SC = getSchedClass(SCIdx);
541 if (SC.ProcIndices[0] != 0)
542 PrintFatalError(Inst->TheDef->getLoc(), "Instruction's sched class "
543 "must not be subtarget specific.");
544
545 IdxVec ProcIndices;
546 if (SC.ItinClassDef->getName() != "NoItinerary") {
547 ProcIndices.push_back(0);
548 dbgs() << "Itinerary for " << InstName << ": "
549 << SC.ItinClassDef->getName() << '\n';
550 }
551 if (!SC.Writes.empty()) {
552 ProcIndices.push_back(0);
553 dbgs() << "SchedRW machine model for " << InstName;
554 for (IdxIter WI = SC.Writes.begin(), WE = SC.Writes.end(); WI != WE; ++WI)
555 dbgs() << " " << SchedWrites[*WI].Name;
556 for (IdxIter RI = SC.Reads.begin(), RE = SC.Reads.end(); RI != RE; ++RI)
557 dbgs() << " " << SchedReads[*RI].Name;
558 dbgs() << '\n';
559 }
560 const RecVec &RWDefs = SchedClasses[SCIdx].InstRWs;
561 for (RecIter RWI = RWDefs.begin(), RWE = RWDefs.end();
562 RWI != RWE; ++RWI) {
563 const CodeGenProcModel &ProcModel =
564 getProcModel((*RWI)->getValueAsDef("SchedModel"));
565 ProcIndices.push_back(ProcModel.Index);
566 dbgs() << "InstRW on " << ProcModel.ModelName << " for " << InstName;
567 IdxVec Writes;
568 IdxVec Reads;
569 findRWs((*RWI)->getValueAsListOfDefs("OperandReadWrites"),
570 Writes, Reads);
571 for (IdxIter WI = Writes.begin(), WE = Writes.end(); WI != WE; ++WI)
572 dbgs() << " " << SchedWrites[*WI].Name;
573 for (IdxIter RI = Reads.begin(), RE = Reads.end(); RI != RE; ++RI)
574 dbgs() << " " << SchedReads[*RI].Name;
575 dbgs() << '\n';
576 }
577 for (std::vector<CodeGenProcModel>::iterator PI = ProcModels.begin(),
578 PE = ProcModels.end(); PI != PE; ++PI) {
579 if (!std::count(ProcIndices.begin(), ProcIndices.end(), PI->Index))
580 dbgs() << "No machine model for " << Inst->TheDef->getName()
581 << " on processor " << PI->ModelName << '\n';
582 }
583 }
584 }
585
586 /// Find an SchedClass that has been inferred from a per-operand list of
587 /// SchedWrites and SchedReads.
findSchedClassIdx(Record * ItinClassDef,ArrayRef<unsigned> Writes,ArrayRef<unsigned> Reads) const588 unsigned CodeGenSchedModels::findSchedClassIdx(Record *ItinClassDef,
589 ArrayRef<unsigned> Writes,
590 ArrayRef<unsigned> Reads) const {
591 for (SchedClassIter I = schedClassBegin(), E = schedClassEnd(); I != E; ++I) {
592 if (I->ItinClassDef == ItinClassDef && makeArrayRef(I->Writes) == Writes &&
593 makeArrayRef(I->Reads) == Reads) {
594 return I - schedClassBegin();
595 }
596 }
597 return 0;
598 }
599
600 // Get the SchedClass index for an instruction.
getSchedClassIdx(const CodeGenInstruction & Inst) const601 unsigned CodeGenSchedModels::getSchedClassIdx(
602 const CodeGenInstruction &Inst) const {
603
604 return InstrClassMap.lookup(Inst.TheDef);
605 }
606
607 std::string
createSchedClassName(Record * ItinClassDef,ArrayRef<unsigned> OperWrites,ArrayRef<unsigned> OperReads)608 CodeGenSchedModels::createSchedClassName(Record *ItinClassDef,
609 ArrayRef<unsigned> OperWrites,
610 ArrayRef<unsigned> OperReads) {
611
612 std::string Name;
613 if (ItinClassDef && ItinClassDef->getName() != "NoItinerary")
614 Name = ItinClassDef->getName();
615 for (unsigned Idx : OperWrites) {
616 if (!Name.empty())
617 Name += '_';
618 Name += SchedWrites[Idx].Name;
619 }
620 for (unsigned Idx : OperReads) {
621 Name += '_';
622 Name += SchedReads[Idx].Name;
623 }
624 return Name;
625 }
626
createSchedClassName(const RecVec & InstDefs)627 std::string CodeGenSchedModels::createSchedClassName(const RecVec &InstDefs) {
628
629 std::string Name;
630 for (RecIter I = InstDefs.begin(), E = InstDefs.end(); I != E; ++I) {
631 if (I != InstDefs.begin())
632 Name += '_';
633 Name += (*I)->getName();
634 }
635 return Name;
636 }
637
638 /// Add an inferred sched class from an itinerary class and per-operand list of
639 /// SchedWrites and SchedReads. ProcIndices contains the set of IDs of
640 /// processors that may utilize this class.
addSchedClass(Record * ItinClassDef,ArrayRef<unsigned> OperWrites,ArrayRef<unsigned> OperReads,ArrayRef<unsigned> ProcIndices)641 unsigned CodeGenSchedModels::addSchedClass(Record *ItinClassDef,
642 ArrayRef<unsigned> OperWrites,
643 ArrayRef<unsigned> OperReads,
644 ArrayRef<unsigned> ProcIndices) {
645 assert(!ProcIndices.empty() && "expect at least one ProcIdx");
646
647 unsigned Idx = findSchedClassIdx(ItinClassDef, OperWrites, OperReads);
648 if (Idx || SchedClasses[0].isKeyEqual(ItinClassDef, OperWrites, OperReads)) {
649 IdxVec PI;
650 std::set_union(SchedClasses[Idx].ProcIndices.begin(),
651 SchedClasses[Idx].ProcIndices.end(),
652 ProcIndices.begin(), ProcIndices.end(),
653 std::back_inserter(PI));
654 SchedClasses[Idx].ProcIndices.swap(PI);
655 return Idx;
656 }
657 Idx = SchedClasses.size();
658 SchedClasses.resize(Idx+1);
659 CodeGenSchedClass &SC = SchedClasses.back();
660 SC.Index = Idx;
661 SC.Name = createSchedClassName(ItinClassDef, OperWrites, OperReads);
662 SC.ItinClassDef = ItinClassDef;
663 SC.Writes = OperWrites;
664 SC.Reads = OperReads;
665 SC.ProcIndices = ProcIndices;
666
667 return Idx;
668 }
669
670 // Create classes for each set of opcodes that are in the same InstReadWrite
671 // definition across all processors.
createInstRWClass(Record * InstRWDef)672 void CodeGenSchedModels::createInstRWClass(Record *InstRWDef) {
673 // ClassInstrs will hold an entry for each subset of Instrs in InstRWDef that
674 // intersects with an existing class via a previous InstRWDef. Instrs that do
675 // not intersect with an existing class refer back to their former class as
676 // determined from ItinDef or SchedRW.
677 SmallVector<std::pair<unsigned, SmallVector<Record *, 8> >, 4> ClassInstrs;
678 // Sort Instrs into sets.
679 const RecVec *InstDefs = Sets.expand(InstRWDef);
680 if (InstDefs->empty())
681 PrintFatalError(InstRWDef->getLoc(), "No matching instruction opcodes");
682
683 for (RecIter I = InstDefs->begin(), E = InstDefs->end(); I != E; ++I) {
684 InstClassMapTy::const_iterator Pos = InstrClassMap.find(*I);
685 if (Pos == InstrClassMap.end())
686 PrintFatalError((*I)->getLoc(), "No sched class for instruction.");
687 unsigned SCIdx = Pos->second;
688 unsigned CIdx = 0, CEnd = ClassInstrs.size();
689 for (; CIdx != CEnd; ++CIdx) {
690 if (ClassInstrs[CIdx].first == SCIdx)
691 break;
692 }
693 if (CIdx == CEnd) {
694 ClassInstrs.resize(CEnd + 1);
695 ClassInstrs[CIdx].first = SCIdx;
696 }
697 ClassInstrs[CIdx].second.push_back(*I);
698 }
699 // For each set of Instrs, create a new class if necessary, and map or remap
700 // the Instrs to it.
701 unsigned CIdx = 0, CEnd = ClassInstrs.size();
702 for (; CIdx != CEnd; ++CIdx) {
703 unsigned OldSCIdx = ClassInstrs[CIdx].first;
704 ArrayRef<Record*> InstDefs = ClassInstrs[CIdx].second;
705 // If the all instrs in the current class are accounted for, then leave
706 // them mapped to their old class.
707 if (OldSCIdx) {
708 const RecVec &RWDefs = SchedClasses[OldSCIdx].InstRWs;
709 if (!RWDefs.empty()) {
710 const RecVec *OrigInstDefs = Sets.expand(RWDefs[0]);
711 unsigned OrigNumInstrs = 0;
712 for (RecIter I = OrigInstDefs->begin(), E = OrigInstDefs->end();
713 I != E; ++I) {
714 if (InstrClassMap[*I] == OldSCIdx)
715 ++OrigNumInstrs;
716 }
717 if (OrigNumInstrs == InstDefs.size()) {
718 assert(SchedClasses[OldSCIdx].ProcIndices[0] == 0 &&
719 "expected a generic SchedClass");
720 DEBUG(dbgs() << "InstRW: Reuse SC " << OldSCIdx << ":"
721 << SchedClasses[OldSCIdx].Name << " on "
722 << InstRWDef->getValueAsDef("SchedModel")->getName() << "\n");
723 SchedClasses[OldSCIdx].InstRWs.push_back(InstRWDef);
724 continue;
725 }
726 }
727 }
728 unsigned SCIdx = SchedClasses.size();
729 SchedClasses.resize(SCIdx+1);
730 CodeGenSchedClass &SC = SchedClasses.back();
731 SC.Index = SCIdx;
732 SC.Name = createSchedClassName(InstDefs);
733 DEBUG(dbgs() << "InstRW: New SC " << SCIdx << ":" << SC.Name << " on "
734 << InstRWDef->getValueAsDef("SchedModel")->getName() << "\n");
735
736 // Preserve ItinDef and Writes/Reads for processors without an InstRW entry.
737 SC.ItinClassDef = SchedClasses[OldSCIdx].ItinClassDef;
738 SC.Writes = SchedClasses[OldSCIdx].Writes;
739 SC.Reads = SchedClasses[OldSCIdx].Reads;
740 SC.ProcIndices.push_back(0);
741 // Map each Instr to this new class.
742 // Note that InstDefs may be a smaller list than InstRWDef's "Instrs".
743 Record *RWModelDef = InstRWDef->getValueAsDef("SchedModel");
744 SmallSet<unsigned, 4> RemappedClassIDs;
745 for (ArrayRef<Record*>::const_iterator
746 II = InstDefs.begin(), IE = InstDefs.end(); II != IE; ++II) {
747 unsigned OldSCIdx = InstrClassMap[*II];
748 if (OldSCIdx && RemappedClassIDs.insert(OldSCIdx).second) {
749 for (RecIter RI = SchedClasses[OldSCIdx].InstRWs.begin(),
750 RE = SchedClasses[OldSCIdx].InstRWs.end(); RI != RE; ++RI) {
751 if ((*RI)->getValueAsDef("SchedModel") == RWModelDef) {
752 PrintFatalError(InstRWDef->getLoc(), "Overlapping InstRW def " +
753 (*II)->getName() + " also matches " +
754 (*RI)->getValue("Instrs")->getValue()->getAsString());
755 }
756 assert(*RI != InstRWDef && "SchedClass has duplicate InstRW def");
757 SC.InstRWs.push_back(*RI);
758 }
759 }
760 InstrClassMap[*II] = SCIdx;
761 }
762 SC.InstRWs.push_back(InstRWDef);
763 }
764 }
765
766 // True if collectProcItins found anything.
hasItineraries() const767 bool CodeGenSchedModels::hasItineraries() const {
768 for (CodeGenSchedModels::ProcIter PI = procModelBegin(), PE = procModelEnd();
769 PI != PE; ++PI) {
770 if (PI->hasItineraries())
771 return true;
772 }
773 return false;
774 }
775
776 // Gather the processor itineraries.
collectProcItins()777 void CodeGenSchedModels::collectProcItins() {
778 for (CodeGenProcModel &ProcModel : ProcModels) {
779 if (!ProcModel.hasItineraries())
780 continue;
781
782 RecVec ItinRecords = ProcModel.ItinsDef->getValueAsListOfDefs("IID");
783 assert(!ItinRecords.empty() && "ProcModel.hasItineraries is incorrect");
784
785 // Populate ItinDefList with Itinerary records.
786 ProcModel.ItinDefList.resize(NumInstrSchedClasses);
787
788 // Insert each itinerary data record in the correct position within
789 // the processor model's ItinDefList.
790 for (unsigned i = 0, N = ItinRecords.size(); i < N; i++) {
791 Record *ItinData = ItinRecords[i];
792 Record *ItinDef = ItinData->getValueAsDef("TheClass");
793 bool FoundClass = false;
794 for (SchedClassIter SCI = schedClassBegin(), SCE = schedClassEnd();
795 SCI != SCE; ++SCI) {
796 // Multiple SchedClasses may share an itinerary. Update all of them.
797 if (SCI->ItinClassDef == ItinDef) {
798 ProcModel.ItinDefList[SCI->Index] = ItinData;
799 FoundClass = true;
800 }
801 }
802 if (!FoundClass) {
803 DEBUG(dbgs() << ProcModel.ItinsDef->getName()
804 << " missing class for itinerary " << ItinDef->getName() << '\n');
805 }
806 }
807 // Check for missing itinerary entries.
808 assert(!ProcModel.ItinDefList[0] && "NoItinerary class can't have rec");
809 DEBUG(
810 for (unsigned i = 1, N = ProcModel.ItinDefList.size(); i < N; ++i) {
811 if (!ProcModel.ItinDefList[i])
812 dbgs() << ProcModel.ItinsDef->getName()
813 << " missing itinerary for class "
814 << SchedClasses[i].Name << '\n';
815 });
816 }
817 }
818
819 // Gather the read/write types for each itinerary class.
collectProcItinRW()820 void CodeGenSchedModels::collectProcItinRW() {
821 RecVec ItinRWDefs = Records.getAllDerivedDefinitions("ItinRW");
822 std::sort(ItinRWDefs.begin(), ItinRWDefs.end(), LessRecord());
823 for (RecIter II = ItinRWDefs.begin(), IE = ItinRWDefs.end(); II != IE; ++II) {
824 if (!(*II)->getValueInit("SchedModel")->isComplete())
825 PrintFatalError((*II)->getLoc(), "SchedModel is undefined");
826 Record *ModelDef = (*II)->getValueAsDef("SchedModel");
827 ProcModelMapTy::const_iterator I = ProcModelMap.find(ModelDef);
828 if (I == ProcModelMap.end()) {
829 PrintFatalError((*II)->getLoc(), "Undefined SchedMachineModel "
830 + ModelDef->getName());
831 }
832 ProcModels[I->second].ItinRWDefs.push_back(*II);
833 }
834 }
835
836 // Gather the unsupported features for processor models.
collectProcUnsupportedFeatures()837 void CodeGenSchedModels::collectProcUnsupportedFeatures() {
838 for (CodeGenProcModel &ProcModel : ProcModels) {
839 for (Record *Pred : ProcModel.ModelDef->getValueAsListOfDefs("UnsupportedFeatures")) {
840 ProcModel.UnsupportedFeaturesDefs.push_back(Pred);
841 }
842 }
843 }
844
845 /// Infer new classes from existing classes. In the process, this may create new
846 /// SchedWrites from sequences of existing SchedWrites.
inferSchedClasses()847 void CodeGenSchedModels::inferSchedClasses() {
848 DEBUG(dbgs() << NumInstrSchedClasses << " instr sched classes.\n");
849
850 // Visit all existing classes and newly created classes.
851 for (unsigned Idx = 0; Idx != SchedClasses.size(); ++Idx) {
852 assert(SchedClasses[Idx].Index == Idx && "bad SCIdx");
853
854 if (SchedClasses[Idx].ItinClassDef)
855 inferFromItinClass(SchedClasses[Idx].ItinClassDef, Idx);
856 if (!SchedClasses[Idx].InstRWs.empty())
857 inferFromInstRWs(Idx);
858 if (!SchedClasses[Idx].Writes.empty()) {
859 inferFromRW(SchedClasses[Idx].Writes, SchedClasses[Idx].Reads,
860 Idx, SchedClasses[Idx].ProcIndices);
861 }
862 assert(SchedClasses.size() < (NumInstrSchedClasses*6) &&
863 "too many SchedVariants");
864 }
865 }
866
867 /// Infer classes from per-processor itinerary resources.
inferFromItinClass(Record * ItinClassDef,unsigned FromClassIdx)868 void CodeGenSchedModels::inferFromItinClass(Record *ItinClassDef,
869 unsigned FromClassIdx) {
870 for (unsigned PIdx = 0, PEnd = ProcModels.size(); PIdx != PEnd; ++PIdx) {
871 const CodeGenProcModel &PM = ProcModels[PIdx];
872 // For all ItinRW entries.
873 bool HasMatch = false;
874 for (RecIter II = PM.ItinRWDefs.begin(), IE = PM.ItinRWDefs.end();
875 II != IE; ++II) {
876 RecVec Matched = (*II)->getValueAsListOfDefs("MatchedItinClasses");
877 if (!std::count(Matched.begin(), Matched.end(), ItinClassDef))
878 continue;
879 if (HasMatch)
880 PrintFatalError((*II)->getLoc(), "Duplicate itinerary class "
881 + ItinClassDef->getName()
882 + " in ItinResources for " + PM.ModelName);
883 HasMatch = true;
884 IdxVec Writes, Reads;
885 findRWs((*II)->getValueAsListOfDefs("OperandReadWrites"), Writes, Reads);
886 IdxVec ProcIndices(1, PIdx);
887 inferFromRW(Writes, Reads, FromClassIdx, ProcIndices);
888 }
889 }
890 }
891
892 /// Infer classes from per-processor InstReadWrite definitions.
inferFromInstRWs(unsigned SCIdx)893 void CodeGenSchedModels::inferFromInstRWs(unsigned SCIdx) {
894 for (unsigned I = 0, E = SchedClasses[SCIdx].InstRWs.size(); I != E; ++I) {
895 assert(SchedClasses[SCIdx].InstRWs.size() == E && "InstrRWs was mutated!");
896 Record *Rec = SchedClasses[SCIdx].InstRWs[I];
897 const RecVec *InstDefs = Sets.expand(Rec);
898 RecIter II = InstDefs->begin(), IE = InstDefs->end();
899 for (; II != IE; ++II) {
900 if (InstrClassMap[*II] == SCIdx)
901 break;
902 }
903 // If this class no longer has any instructions mapped to it, it has become
904 // irrelevant.
905 if (II == IE)
906 continue;
907 IdxVec Writes, Reads;
908 findRWs(Rec->getValueAsListOfDefs("OperandReadWrites"), Writes, Reads);
909 unsigned PIdx = getProcModel(Rec->getValueAsDef("SchedModel")).Index;
910 IdxVec ProcIndices(1, PIdx);
911 inferFromRW(Writes, Reads, SCIdx, ProcIndices); // May mutate SchedClasses.
912 }
913 }
914
915 namespace {
916 // Helper for substituteVariantOperand.
917 struct TransVariant {
918 Record *VarOrSeqDef; // Variant or sequence.
919 unsigned RWIdx; // Index of this variant or sequence's matched type.
920 unsigned ProcIdx; // Processor model index or zero for any.
921 unsigned TransVecIdx; // Index into PredTransitions::TransVec.
922
TransVariant__anon50a6a0fc0211::TransVariant923 TransVariant(Record *def, unsigned rwi, unsigned pi, unsigned ti):
924 VarOrSeqDef(def), RWIdx(rwi), ProcIdx(pi), TransVecIdx(ti) {}
925 };
926
927 // Associate a predicate with the SchedReadWrite that it guards.
928 // RWIdx is the index of the read/write variant.
929 struct PredCheck {
930 bool IsRead;
931 unsigned RWIdx;
932 Record *Predicate;
933
PredCheck__anon50a6a0fc0211::PredCheck934 PredCheck(bool r, unsigned w, Record *p): IsRead(r), RWIdx(w), Predicate(p) {}
935 };
936
937 // A Predicate transition is a list of RW sequences guarded by a PredTerm.
938 struct PredTransition {
939 // A predicate term is a conjunction of PredChecks.
940 SmallVector<PredCheck, 4> PredTerm;
941 SmallVector<SmallVector<unsigned,4>, 16> WriteSequences;
942 SmallVector<SmallVector<unsigned,4>, 16> ReadSequences;
943 SmallVector<unsigned, 4> ProcIndices;
944 };
945
946 // Encapsulate a set of partially constructed transitions.
947 // The results are built by repeated calls to substituteVariants.
948 class PredTransitions {
949 CodeGenSchedModels &SchedModels;
950
951 public:
952 std::vector<PredTransition> TransVec;
953
PredTransitions(CodeGenSchedModels & sm)954 PredTransitions(CodeGenSchedModels &sm): SchedModels(sm) {}
955
956 void substituteVariantOperand(const SmallVectorImpl<unsigned> &RWSeq,
957 bool IsRead, unsigned StartIdx);
958
959 void substituteVariants(const PredTransition &Trans);
960
961 #ifndef NDEBUG
962 void dump() const;
963 #endif
964
965 private:
966 bool mutuallyExclusive(Record *PredDef, ArrayRef<PredCheck> Term);
967 void getIntersectingVariants(
968 const CodeGenSchedRW &SchedRW, unsigned TransIdx,
969 std::vector<TransVariant> &IntersectingVariants);
970 void pushVariant(const TransVariant &VInfo, bool IsRead);
971 };
972 } // anonymous
973
974 // Return true if this predicate is mutually exclusive with a PredTerm. This
975 // degenerates into checking if the predicate is mutually exclusive with any
976 // predicate in the Term's conjunction.
977 //
978 // All predicates associated with a given SchedRW are considered mutually
979 // exclusive. This should work even if the conditions expressed by the
980 // predicates are not exclusive because the predicates for a given SchedWrite
981 // are always checked in the order they are defined in the .td file. Later
982 // conditions implicitly negate any prior condition.
mutuallyExclusive(Record * PredDef,ArrayRef<PredCheck> Term)983 bool PredTransitions::mutuallyExclusive(Record *PredDef,
984 ArrayRef<PredCheck> Term) {
985
986 for (ArrayRef<PredCheck>::iterator I = Term.begin(), E = Term.end();
987 I != E; ++I) {
988 if (I->Predicate == PredDef)
989 return false;
990
991 const CodeGenSchedRW &SchedRW = SchedModels.getSchedRW(I->RWIdx, I->IsRead);
992 assert(SchedRW.HasVariants && "PredCheck must refer to a SchedVariant");
993 RecVec Variants = SchedRW.TheDef->getValueAsListOfDefs("Variants");
994 for (RecIter VI = Variants.begin(), VE = Variants.end(); VI != VE; ++VI) {
995 if ((*VI)->getValueAsDef("Predicate") == PredDef)
996 return true;
997 }
998 }
999 return false;
1000 }
1001
hasAliasedVariants(const CodeGenSchedRW & RW,CodeGenSchedModels & SchedModels)1002 static bool hasAliasedVariants(const CodeGenSchedRW &RW,
1003 CodeGenSchedModels &SchedModels) {
1004 if (RW.HasVariants)
1005 return true;
1006
1007 for (RecIter I = RW.Aliases.begin(), E = RW.Aliases.end(); I != E; ++I) {
1008 const CodeGenSchedRW &AliasRW =
1009 SchedModels.getSchedRW((*I)->getValueAsDef("AliasRW"));
1010 if (AliasRW.HasVariants)
1011 return true;
1012 if (AliasRW.IsSequence) {
1013 IdxVec ExpandedRWs;
1014 SchedModels.expandRWSequence(AliasRW.Index, ExpandedRWs, AliasRW.IsRead);
1015 for (IdxIter SI = ExpandedRWs.begin(), SE = ExpandedRWs.end();
1016 SI != SE; ++SI) {
1017 if (hasAliasedVariants(SchedModels.getSchedRW(*SI, AliasRW.IsRead),
1018 SchedModels)) {
1019 return true;
1020 }
1021 }
1022 }
1023 }
1024 return false;
1025 }
1026
hasVariant(ArrayRef<PredTransition> Transitions,CodeGenSchedModels & SchedModels)1027 static bool hasVariant(ArrayRef<PredTransition> Transitions,
1028 CodeGenSchedModels &SchedModels) {
1029 for (ArrayRef<PredTransition>::iterator
1030 PTI = Transitions.begin(), PTE = Transitions.end();
1031 PTI != PTE; ++PTI) {
1032 for (SmallVectorImpl<SmallVector<unsigned,4> >::const_iterator
1033 WSI = PTI->WriteSequences.begin(), WSE = PTI->WriteSequences.end();
1034 WSI != WSE; ++WSI) {
1035 for (SmallVectorImpl<unsigned>::const_iterator
1036 WI = WSI->begin(), WE = WSI->end(); WI != WE; ++WI) {
1037 if (hasAliasedVariants(SchedModels.getSchedWrite(*WI), SchedModels))
1038 return true;
1039 }
1040 }
1041 for (SmallVectorImpl<SmallVector<unsigned,4> >::const_iterator
1042 RSI = PTI->ReadSequences.begin(), RSE = PTI->ReadSequences.end();
1043 RSI != RSE; ++RSI) {
1044 for (SmallVectorImpl<unsigned>::const_iterator
1045 RI = RSI->begin(), RE = RSI->end(); RI != RE; ++RI) {
1046 if (hasAliasedVariants(SchedModels.getSchedRead(*RI), SchedModels))
1047 return true;
1048 }
1049 }
1050 }
1051 return false;
1052 }
1053
1054 // Populate IntersectingVariants with any variants or aliased sequences of the
1055 // given SchedRW whose processor indices and predicates are not mutually
1056 // exclusive with the given transition.
getIntersectingVariants(const CodeGenSchedRW & SchedRW,unsigned TransIdx,std::vector<TransVariant> & IntersectingVariants)1057 void PredTransitions::getIntersectingVariants(
1058 const CodeGenSchedRW &SchedRW, unsigned TransIdx,
1059 std::vector<TransVariant> &IntersectingVariants) {
1060
1061 bool GenericRW = false;
1062
1063 std::vector<TransVariant> Variants;
1064 if (SchedRW.HasVariants) {
1065 unsigned VarProcIdx = 0;
1066 if (SchedRW.TheDef->getValueInit("SchedModel")->isComplete()) {
1067 Record *ModelDef = SchedRW.TheDef->getValueAsDef("SchedModel");
1068 VarProcIdx = SchedModels.getProcModel(ModelDef).Index;
1069 }
1070 // Push each variant. Assign TransVecIdx later.
1071 const RecVec VarDefs = SchedRW.TheDef->getValueAsListOfDefs("Variants");
1072 for (RecIter RI = VarDefs.begin(), RE = VarDefs.end(); RI != RE; ++RI)
1073 Variants.push_back(TransVariant(*RI, SchedRW.Index, VarProcIdx, 0));
1074 if (VarProcIdx == 0)
1075 GenericRW = true;
1076 }
1077 for (RecIter AI = SchedRW.Aliases.begin(), AE = SchedRW.Aliases.end();
1078 AI != AE; ++AI) {
1079 // If either the SchedAlias itself or the SchedReadWrite that it aliases
1080 // to is defined within a processor model, constrain all variants to
1081 // that processor.
1082 unsigned AliasProcIdx = 0;
1083 if ((*AI)->getValueInit("SchedModel")->isComplete()) {
1084 Record *ModelDef = (*AI)->getValueAsDef("SchedModel");
1085 AliasProcIdx = SchedModels.getProcModel(ModelDef).Index;
1086 }
1087 const CodeGenSchedRW &AliasRW =
1088 SchedModels.getSchedRW((*AI)->getValueAsDef("AliasRW"));
1089
1090 if (AliasRW.HasVariants) {
1091 const RecVec VarDefs = AliasRW.TheDef->getValueAsListOfDefs("Variants");
1092 for (RecIter RI = VarDefs.begin(), RE = VarDefs.end(); RI != RE; ++RI)
1093 Variants.push_back(TransVariant(*RI, AliasRW.Index, AliasProcIdx, 0));
1094 }
1095 if (AliasRW.IsSequence) {
1096 Variants.push_back(
1097 TransVariant(AliasRW.TheDef, SchedRW.Index, AliasProcIdx, 0));
1098 }
1099 if (AliasProcIdx == 0)
1100 GenericRW = true;
1101 }
1102 for (unsigned VIdx = 0, VEnd = Variants.size(); VIdx != VEnd; ++VIdx) {
1103 TransVariant &Variant = Variants[VIdx];
1104 // Don't expand variants if the processor models don't intersect.
1105 // A zero processor index means any processor.
1106 SmallVectorImpl<unsigned> &ProcIndices = TransVec[TransIdx].ProcIndices;
1107 if (ProcIndices[0] && Variants[VIdx].ProcIdx) {
1108 unsigned Cnt = std::count(ProcIndices.begin(), ProcIndices.end(),
1109 Variant.ProcIdx);
1110 if (!Cnt)
1111 continue;
1112 if (Cnt > 1) {
1113 const CodeGenProcModel &PM =
1114 *(SchedModels.procModelBegin() + Variant.ProcIdx);
1115 PrintFatalError(Variant.VarOrSeqDef->getLoc(),
1116 "Multiple variants defined for processor " +
1117 PM.ModelName +
1118 " Ensure only one SchedAlias exists per RW.");
1119 }
1120 }
1121 if (Variant.VarOrSeqDef->isSubClassOf("SchedVar")) {
1122 Record *PredDef = Variant.VarOrSeqDef->getValueAsDef("Predicate");
1123 if (mutuallyExclusive(PredDef, TransVec[TransIdx].PredTerm))
1124 continue;
1125 }
1126 if (IntersectingVariants.empty()) {
1127 // The first variant builds on the existing transition.
1128 Variant.TransVecIdx = TransIdx;
1129 IntersectingVariants.push_back(Variant);
1130 }
1131 else {
1132 // Push another copy of the current transition for more variants.
1133 Variant.TransVecIdx = TransVec.size();
1134 IntersectingVariants.push_back(Variant);
1135 TransVec.push_back(TransVec[TransIdx]);
1136 }
1137 }
1138 if (GenericRW && IntersectingVariants.empty()) {
1139 PrintFatalError(SchedRW.TheDef->getLoc(), "No variant of this type has "
1140 "a matching predicate on any processor");
1141 }
1142 }
1143
1144 // Push the Reads/Writes selected by this variant onto the PredTransition
1145 // specified by VInfo.
1146 void PredTransitions::
pushVariant(const TransVariant & VInfo,bool IsRead)1147 pushVariant(const TransVariant &VInfo, bool IsRead) {
1148
1149 PredTransition &Trans = TransVec[VInfo.TransVecIdx];
1150
1151 // If this operand transition is reached through a processor-specific alias,
1152 // then the whole transition is specific to this processor.
1153 if (VInfo.ProcIdx != 0)
1154 Trans.ProcIndices.assign(1, VInfo.ProcIdx);
1155
1156 IdxVec SelectedRWs;
1157 if (VInfo.VarOrSeqDef->isSubClassOf("SchedVar")) {
1158 Record *PredDef = VInfo.VarOrSeqDef->getValueAsDef("Predicate");
1159 Trans.PredTerm.push_back(PredCheck(IsRead, VInfo.RWIdx,PredDef));
1160 RecVec SelectedDefs = VInfo.VarOrSeqDef->getValueAsListOfDefs("Selected");
1161 SchedModels.findRWs(SelectedDefs, SelectedRWs, IsRead);
1162 }
1163 else {
1164 assert(VInfo.VarOrSeqDef->isSubClassOf("WriteSequence") &&
1165 "variant must be a SchedVariant or aliased WriteSequence");
1166 SelectedRWs.push_back(SchedModels.getSchedRWIdx(VInfo.VarOrSeqDef, IsRead));
1167 }
1168
1169 const CodeGenSchedRW &SchedRW = SchedModels.getSchedRW(VInfo.RWIdx, IsRead);
1170
1171 SmallVectorImpl<SmallVector<unsigned,4> > &RWSequences = IsRead
1172 ? Trans.ReadSequences : Trans.WriteSequences;
1173 if (SchedRW.IsVariadic) {
1174 unsigned OperIdx = RWSequences.size()-1;
1175 // Make N-1 copies of this transition's last sequence.
1176 for (unsigned i = 1, e = SelectedRWs.size(); i != e; ++i) {
1177 // Create a temporary copy the vector could reallocate.
1178 RWSequences.reserve(RWSequences.size() + 1);
1179 RWSequences.push_back(RWSequences[OperIdx]);
1180 }
1181 // Push each of the N elements of the SelectedRWs onto a copy of the last
1182 // sequence (split the current operand into N operands).
1183 // Note that write sequences should be expanded within this loop--the entire
1184 // sequence belongs to a single operand.
1185 for (IdxIter RWI = SelectedRWs.begin(), RWE = SelectedRWs.end();
1186 RWI != RWE; ++RWI, ++OperIdx) {
1187 IdxVec ExpandedRWs;
1188 if (IsRead)
1189 ExpandedRWs.push_back(*RWI);
1190 else
1191 SchedModels.expandRWSequence(*RWI, ExpandedRWs, IsRead);
1192 RWSequences[OperIdx].insert(RWSequences[OperIdx].end(),
1193 ExpandedRWs.begin(), ExpandedRWs.end());
1194 }
1195 assert(OperIdx == RWSequences.size() && "missed a sequence");
1196 }
1197 else {
1198 // Push this transition's expanded sequence onto this transition's last
1199 // sequence (add to the current operand's sequence).
1200 SmallVectorImpl<unsigned> &Seq = RWSequences.back();
1201 IdxVec ExpandedRWs;
1202 for (IdxIter RWI = SelectedRWs.begin(), RWE = SelectedRWs.end();
1203 RWI != RWE; ++RWI) {
1204 if (IsRead)
1205 ExpandedRWs.push_back(*RWI);
1206 else
1207 SchedModels.expandRWSequence(*RWI, ExpandedRWs, IsRead);
1208 }
1209 Seq.insert(Seq.end(), ExpandedRWs.begin(), ExpandedRWs.end());
1210 }
1211 }
1212
1213 // RWSeq is a sequence of all Reads or all Writes for the next read or write
1214 // operand. StartIdx is an index into TransVec where partial results
1215 // starts. RWSeq must be applied to all transitions between StartIdx and the end
1216 // of TransVec.
substituteVariantOperand(const SmallVectorImpl<unsigned> & RWSeq,bool IsRead,unsigned StartIdx)1217 void PredTransitions::substituteVariantOperand(
1218 const SmallVectorImpl<unsigned> &RWSeq, bool IsRead, unsigned StartIdx) {
1219
1220 // Visit each original RW within the current sequence.
1221 for (SmallVectorImpl<unsigned>::const_iterator
1222 RWI = RWSeq.begin(), RWE = RWSeq.end(); RWI != RWE; ++RWI) {
1223 const CodeGenSchedRW &SchedRW = SchedModels.getSchedRW(*RWI, IsRead);
1224 // Push this RW on all partial PredTransitions or distribute variants.
1225 // New PredTransitions may be pushed within this loop which should not be
1226 // revisited (TransEnd must be loop invariant).
1227 for (unsigned TransIdx = StartIdx, TransEnd = TransVec.size();
1228 TransIdx != TransEnd; ++TransIdx) {
1229 // In the common case, push RW onto the current operand's sequence.
1230 if (!hasAliasedVariants(SchedRW, SchedModels)) {
1231 if (IsRead)
1232 TransVec[TransIdx].ReadSequences.back().push_back(*RWI);
1233 else
1234 TransVec[TransIdx].WriteSequences.back().push_back(*RWI);
1235 continue;
1236 }
1237 // Distribute this partial PredTransition across intersecting variants.
1238 // This will push a copies of TransVec[TransIdx] on the back of TransVec.
1239 std::vector<TransVariant> IntersectingVariants;
1240 getIntersectingVariants(SchedRW, TransIdx, IntersectingVariants);
1241 // Now expand each variant on top of its copy of the transition.
1242 for (std::vector<TransVariant>::const_iterator
1243 IVI = IntersectingVariants.begin(),
1244 IVE = IntersectingVariants.end();
1245 IVI != IVE; ++IVI) {
1246 pushVariant(*IVI, IsRead);
1247 }
1248 }
1249 }
1250 }
1251
1252 // For each variant of a Read/Write in Trans, substitute the sequence of
1253 // Read/Writes guarded by the variant. This is exponential in the number of
1254 // variant Read/Writes, but in practice detection of mutually exclusive
1255 // predicates should result in linear growth in the total number variants.
1256 //
1257 // This is one step in a breadth-first search of nested variants.
substituteVariants(const PredTransition & Trans)1258 void PredTransitions::substituteVariants(const PredTransition &Trans) {
1259 // Build up a set of partial results starting at the back of
1260 // PredTransitions. Remember the first new transition.
1261 unsigned StartIdx = TransVec.size();
1262 TransVec.resize(TransVec.size() + 1);
1263 TransVec.back().PredTerm = Trans.PredTerm;
1264 TransVec.back().ProcIndices = Trans.ProcIndices;
1265
1266 // Visit each original write sequence.
1267 for (SmallVectorImpl<SmallVector<unsigned,4> >::const_iterator
1268 WSI = Trans.WriteSequences.begin(), WSE = Trans.WriteSequences.end();
1269 WSI != WSE; ++WSI) {
1270 // Push a new (empty) write sequence onto all partial Transitions.
1271 for (std::vector<PredTransition>::iterator I =
1272 TransVec.begin() + StartIdx, E = TransVec.end(); I != E; ++I) {
1273 I->WriteSequences.resize(I->WriteSequences.size() + 1);
1274 }
1275 substituteVariantOperand(*WSI, /*IsRead=*/false, StartIdx);
1276 }
1277 // Visit each original read sequence.
1278 for (SmallVectorImpl<SmallVector<unsigned,4> >::const_iterator
1279 RSI = Trans.ReadSequences.begin(), RSE = Trans.ReadSequences.end();
1280 RSI != RSE; ++RSI) {
1281 // Push a new (empty) read sequence onto all partial Transitions.
1282 for (std::vector<PredTransition>::iterator I =
1283 TransVec.begin() + StartIdx, E = TransVec.end(); I != E; ++I) {
1284 I->ReadSequences.resize(I->ReadSequences.size() + 1);
1285 }
1286 substituteVariantOperand(*RSI, /*IsRead=*/true, StartIdx);
1287 }
1288 }
1289
1290 // Create a new SchedClass for each variant found by inferFromRW. Pass
inferFromTransitions(ArrayRef<PredTransition> LastTransitions,unsigned FromClassIdx,CodeGenSchedModels & SchedModels)1291 static void inferFromTransitions(ArrayRef<PredTransition> LastTransitions,
1292 unsigned FromClassIdx,
1293 CodeGenSchedModels &SchedModels) {
1294 // For each PredTransition, create a new CodeGenSchedTransition, which usually
1295 // requires creating a new SchedClass.
1296 for (ArrayRef<PredTransition>::iterator
1297 I = LastTransitions.begin(), E = LastTransitions.end(); I != E; ++I) {
1298 IdxVec OperWritesVariant;
1299 for (SmallVectorImpl<SmallVector<unsigned,4> >::const_iterator
1300 WSI = I->WriteSequences.begin(), WSE = I->WriteSequences.end();
1301 WSI != WSE; ++WSI) {
1302 // Create a new write representing the expanded sequence.
1303 OperWritesVariant.push_back(
1304 SchedModels.findOrInsertRW(*WSI, /*IsRead=*/false));
1305 }
1306 IdxVec OperReadsVariant;
1307 for (SmallVectorImpl<SmallVector<unsigned,4> >::const_iterator
1308 RSI = I->ReadSequences.begin(), RSE = I->ReadSequences.end();
1309 RSI != RSE; ++RSI) {
1310 // Create a new read representing the expanded sequence.
1311 OperReadsVariant.push_back(
1312 SchedModels.findOrInsertRW(*RSI, /*IsRead=*/true));
1313 }
1314 IdxVec ProcIndices(I->ProcIndices.begin(), I->ProcIndices.end());
1315 CodeGenSchedTransition SCTrans;
1316 SCTrans.ToClassIdx =
1317 SchedModels.addSchedClass(/*ItinClassDef=*/nullptr, OperWritesVariant,
1318 OperReadsVariant, ProcIndices);
1319 SCTrans.ProcIndices = ProcIndices;
1320 // The final PredTerm is unique set of predicates guarding the transition.
1321 RecVec Preds;
1322 for (SmallVectorImpl<PredCheck>::const_iterator
1323 PI = I->PredTerm.begin(), PE = I->PredTerm.end(); PI != PE; ++PI) {
1324 Preds.push_back(PI->Predicate);
1325 }
1326 RecIter PredsEnd = std::unique(Preds.begin(), Preds.end());
1327 Preds.resize(PredsEnd - Preds.begin());
1328 SCTrans.PredTerm = Preds;
1329 SchedModels.getSchedClass(FromClassIdx).Transitions.push_back(SCTrans);
1330 }
1331 }
1332
1333 // Create new SchedClasses for the given ReadWrite list. If any of the
1334 // ReadWrites refers to a SchedVariant, create a new SchedClass for each variant
1335 // of the ReadWrite list, following Aliases if necessary.
inferFromRW(ArrayRef<unsigned> OperWrites,ArrayRef<unsigned> OperReads,unsigned FromClassIdx,ArrayRef<unsigned> ProcIndices)1336 void CodeGenSchedModels::inferFromRW(ArrayRef<unsigned> OperWrites,
1337 ArrayRef<unsigned> OperReads,
1338 unsigned FromClassIdx,
1339 ArrayRef<unsigned> ProcIndices) {
1340 DEBUG(dbgs() << "INFER RW proc("; dumpIdxVec(ProcIndices); dbgs() << ") ");
1341
1342 // Create a seed transition with an empty PredTerm and the expanded sequences
1343 // of SchedWrites for the current SchedClass.
1344 std::vector<PredTransition> LastTransitions;
1345 LastTransitions.resize(1);
1346 LastTransitions.back().ProcIndices.append(ProcIndices.begin(),
1347 ProcIndices.end());
1348
1349 for (unsigned WriteIdx : OperWrites) {
1350 IdxVec WriteSeq;
1351 expandRWSequence(WriteIdx, WriteSeq, /*IsRead=*/false);
1352 unsigned Idx = LastTransitions[0].WriteSequences.size();
1353 LastTransitions[0].WriteSequences.resize(Idx + 1);
1354 SmallVectorImpl<unsigned> &Seq = LastTransitions[0].WriteSequences[Idx];
1355 for (IdxIter WI = WriteSeq.begin(), WE = WriteSeq.end(); WI != WE; ++WI)
1356 Seq.push_back(*WI);
1357 DEBUG(dbgs() << "("; dumpIdxVec(Seq); dbgs() << ") ");
1358 }
1359 DEBUG(dbgs() << " Reads: ");
1360 for (unsigned ReadIdx : OperReads) {
1361 IdxVec ReadSeq;
1362 expandRWSequence(ReadIdx, ReadSeq, /*IsRead=*/true);
1363 unsigned Idx = LastTransitions[0].ReadSequences.size();
1364 LastTransitions[0].ReadSequences.resize(Idx + 1);
1365 SmallVectorImpl<unsigned> &Seq = LastTransitions[0].ReadSequences[Idx];
1366 for (IdxIter RI = ReadSeq.begin(), RE = ReadSeq.end(); RI != RE; ++RI)
1367 Seq.push_back(*RI);
1368 DEBUG(dbgs() << "("; dumpIdxVec(Seq); dbgs() << ") ");
1369 }
1370 DEBUG(dbgs() << '\n');
1371
1372 // Collect all PredTransitions for individual operands.
1373 // Iterate until no variant writes remain.
1374 while (hasVariant(LastTransitions, *this)) {
1375 PredTransitions Transitions(*this);
1376 for (std::vector<PredTransition>::const_iterator
1377 I = LastTransitions.begin(), E = LastTransitions.end();
1378 I != E; ++I) {
1379 Transitions.substituteVariants(*I);
1380 }
1381 DEBUG(Transitions.dump());
1382 LastTransitions.swap(Transitions.TransVec);
1383 }
1384 // If the first transition has no variants, nothing to do.
1385 if (LastTransitions[0].PredTerm.empty())
1386 return;
1387
1388 // WARNING: We are about to mutate the SchedClasses vector. Do not refer to
1389 // OperWrites, OperReads, or ProcIndices after calling inferFromTransitions.
1390 inferFromTransitions(LastTransitions, FromClassIdx, *this);
1391 }
1392
1393 // Check if any processor resource group contains all resource records in
1394 // SubUnits.
hasSuperGroup(RecVec & SubUnits,CodeGenProcModel & PM)1395 bool CodeGenSchedModels::hasSuperGroup(RecVec &SubUnits, CodeGenProcModel &PM) {
1396 for (unsigned i = 0, e = PM.ProcResourceDefs.size(); i < e; ++i) {
1397 if (!PM.ProcResourceDefs[i]->isSubClassOf("ProcResGroup"))
1398 continue;
1399 RecVec SuperUnits =
1400 PM.ProcResourceDefs[i]->getValueAsListOfDefs("Resources");
1401 RecIter RI = SubUnits.begin(), RE = SubUnits.end();
1402 for ( ; RI != RE; ++RI) {
1403 if (std::find(SuperUnits.begin(), SuperUnits.end(), *RI)
1404 == SuperUnits.end()) {
1405 break;
1406 }
1407 }
1408 if (RI == RE)
1409 return true;
1410 }
1411 return false;
1412 }
1413
1414 // Verify that overlapping groups have a common supergroup.
verifyProcResourceGroups(CodeGenProcModel & PM)1415 void CodeGenSchedModels::verifyProcResourceGroups(CodeGenProcModel &PM) {
1416 for (unsigned i = 0, e = PM.ProcResourceDefs.size(); i < e; ++i) {
1417 if (!PM.ProcResourceDefs[i]->isSubClassOf("ProcResGroup"))
1418 continue;
1419 RecVec CheckUnits =
1420 PM.ProcResourceDefs[i]->getValueAsListOfDefs("Resources");
1421 for (unsigned j = i+1; j < e; ++j) {
1422 if (!PM.ProcResourceDefs[j]->isSubClassOf("ProcResGroup"))
1423 continue;
1424 RecVec OtherUnits =
1425 PM.ProcResourceDefs[j]->getValueAsListOfDefs("Resources");
1426 if (std::find_first_of(CheckUnits.begin(), CheckUnits.end(),
1427 OtherUnits.begin(), OtherUnits.end())
1428 != CheckUnits.end()) {
1429 // CheckUnits and OtherUnits overlap
1430 OtherUnits.insert(OtherUnits.end(), CheckUnits.begin(),
1431 CheckUnits.end());
1432 if (!hasSuperGroup(OtherUnits, PM)) {
1433 PrintFatalError((PM.ProcResourceDefs[i])->getLoc(),
1434 "proc resource group overlaps with "
1435 + PM.ProcResourceDefs[j]->getName()
1436 + " but no supergroup contains both.");
1437 }
1438 }
1439 }
1440 }
1441 }
1442
1443 // Collect and sort WriteRes, ReadAdvance, and ProcResources.
collectProcResources()1444 void CodeGenSchedModels::collectProcResources() {
1445 ProcResourceDefs = Records.getAllDerivedDefinitions("ProcResourceUnits");
1446 ProcResGroups = Records.getAllDerivedDefinitions("ProcResGroup");
1447
1448 // Add any subtarget-specific SchedReadWrites that are directly associated
1449 // with processor resources. Refer to the parent SchedClass's ProcIndices to
1450 // determine which processors they apply to.
1451 for (SchedClassIter SCI = schedClassBegin(), SCE = schedClassEnd();
1452 SCI != SCE; ++SCI) {
1453 if (SCI->ItinClassDef)
1454 collectItinProcResources(SCI->ItinClassDef);
1455 else {
1456 // This class may have a default ReadWrite list which can be overriden by
1457 // InstRW definitions.
1458 if (!SCI->InstRWs.empty()) {
1459 for (RecIter RWI = SCI->InstRWs.begin(), RWE = SCI->InstRWs.end();
1460 RWI != RWE; ++RWI) {
1461 Record *RWModelDef = (*RWI)->getValueAsDef("SchedModel");
1462 IdxVec ProcIndices(1, getProcModel(RWModelDef).Index);
1463 IdxVec Writes, Reads;
1464 findRWs((*RWI)->getValueAsListOfDefs("OperandReadWrites"),
1465 Writes, Reads);
1466 collectRWResources(Writes, Reads, ProcIndices);
1467 }
1468 }
1469 collectRWResources(SCI->Writes, SCI->Reads, SCI->ProcIndices);
1470 }
1471 }
1472 // Add resources separately defined by each subtarget.
1473 RecVec WRDefs = Records.getAllDerivedDefinitions("WriteRes");
1474 for (RecIter WRI = WRDefs.begin(), WRE = WRDefs.end(); WRI != WRE; ++WRI) {
1475 Record *ModelDef = (*WRI)->getValueAsDef("SchedModel");
1476 addWriteRes(*WRI, getProcModel(ModelDef).Index);
1477 }
1478 RecVec SWRDefs = Records.getAllDerivedDefinitions("SchedWriteRes");
1479 for (RecIter WRI = SWRDefs.begin(), WRE = SWRDefs.end(); WRI != WRE; ++WRI) {
1480 Record *ModelDef = (*WRI)->getValueAsDef("SchedModel");
1481 addWriteRes(*WRI, getProcModel(ModelDef).Index);
1482 }
1483 RecVec RADefs = Records.getAllDerivedDefinitions("ReadAdvance");
1484 for (RecIter RAI = RADefs.begin(), RAE = RADefs.end(); RAI != RAE; ++RAI) {
1485 Record *ModelDef = (*RAI)->getValueAsDef("SchedModel");
1486 addReadAdvance(*RAI, getProcModel(ModelDef).Index);
1487 }
1488 RecVec SRADefs = Records.getAllDerivedDefinitions("SchedReadAdvance");
1489 for (RecIter RAI = SRADefs.begin(), RAE = SRADefs.end(); RAI != RAE; ++RAI) {
1490 if ((*RAI)->getValueInit("SchedModel")->isComplete()) {
1491 Record *ModelDef = (*RAI)->getValueAsDef("SchedModel");
1492 addReadAdvance(*RAI, getProcModel(ModelDef).Index);
1493 }
1494 }
1495 // Add ProcResGroups that are defined within this processor model, which may
1496 // not be directly referenced but may directly specify a buffer size.
1497 RecVec ProcResGroups = Records.getAllDerivedDefinitions("ProcResGroup");
1498 for (RecIter RI = ProcResGroups.begin(), RE = ProcResGroups.end();
1499 RI != RE; ++RI) {
1500 if (!(*RI)->getValueInit("SchedModel")->isComplete())
1501 continue;
1502 CodeGenProcModel &PM = getProcModel((*RI)->getValueAsDef("SchedModel"));
1503 RecIter I = std::find(PM.ProcResourceDefs.begin(),
1504 PM.ProcResourceDefs.end(), *RI);
1505 if (I == PM.ProcResourceDefs.end())
1506 PM.ProcResourceDefs.push_back(*RI);
1507 }
1508 // Finalize each ProcModel by sorting the record arrays.
1509 for (CodeGenProcModel &PM : ProcModels) {
1510 std::sort(PM.WriteResDefs.begin(), PM.WriteResDefs.end(),
1511 LessRecord());
1512 std::sort(PM.ReadAdvanceDefs.begin(), PM.ReadAdvanceDefs.end(),
1513 LessRecord());
1514 std::sort(PM.ProcResourceDefs.begin(), PM.ProcResourceDefs.end(),
1515 LessRecord());
1516 DEBUG(
1517 PM.dump();
1518 dbgs() << "WriteResDefs: ";
1519 for (RecIter RI = PM.WriteResDefs.begin(),
1520 RE = PM.WriteResDefs.end(); RI != RE; ++RI) {
1521 if ((*RI)->isSubClassOf("WriteRes"))
1522 dbgs() << (*RI)->getValueAsDef("WriteType")->getName() << " ";
1523 else
1524 dbgs() << (*RI)->getName() << " ";
1525 }
1526 dbgs() << "\nReadAdvanceDefs: ";
1527 for (RecIter RI = PM.ReadAdvanceDefs.begin(),
1528 RE = PM.ReadAdvanceDefs.end(); RI != RE; ++RI) {
1529 if ((*RI)->isSubClassOf("ReadAdvance"))
1530 dbgs() << (*RI)->getValueAsDef("ReadType")->getName() << " ";
1531 else
1532 dbgs() << (*RI)->getName() << " ";
1533 }
1534 dbgs() << "\nProcResourceDefs: ";
1535 for (RecIter RI = PM.ProcResourceDefs.begin(),
1536 RE = PM.ProcResourceDefs.end(); RI != RE; ++RI) {
1537 dbgs() << (*RI)->getName() << " ";
1538 }
1539 dbgs() << '\n');
1540 verifyProcResourceGroups(PM);
1541 }
1542
1543 ProcResourceDefs.clear();
1544 ProcResGroups.clear();
1545 }
1546
checkCompleteness()1547 void CodeGenSchedModels::checkCompleteness() {
1548 bool Complete = true;
1549 bool HadCompleteModel = false;
1550 for (const CodeGenProcModel &ProcModel : procModels()) {
1551 if (!ProcModel.ModelDef->getValueAsBit("CompleteModel"))
1552 continue;
1553 for (const CodeGenInstruction *Inst : Target.getInstructionsByEnumValue()) {
1554 if (Inst->hasNoSchedulingInfo)
1555 continue;
1556 if (ProcModel.isUnsupported(*Inst))
1557 continue;
1558 unsigned SCIdx = getSchedClassIdx(*Inst);
1559 if (!SCIdx) {
1560 if (Inst->TheDef->isValueUnset("SchedRW") && !HadCompleteModel) {
1561 PrintError("No schedule information for instruction '"
1562 + Inst->TheDef->getName() + "'");
1563 Complete = false;
1564 }
1565 continue;
1566 }
1567
1568 const CodeGenSchedClass &SC = getSchedClass(SCIdx);
1569 if (!SC.Writes.empty())
1570 continue;
1571 if (SC.ItinClassDef != nullptr)
1572 continue;
1573
1574 const RecVec &InstRWs = SC.InstRWs;
1575 auto I = std::find_if(InstRWs.begin(), InstRWs.end(),
1576 [&ProcModel] (const Record *R) {
1577 return R->getValueAsDef("SchedModel") ==
1578 ProcModel.ModelDef;
1579 });
1580 if (I == InstRWs.end()) {
1581 PrintError("'" + ProcModel.ModelName + "' lacks information for '" +
1582 Inst->TheDef->getName() + "'");
1583 Complete = false;
1584 }
1585 }
1586 HadCompleteModel = true;
1587 }
1588 if (!Complete) {
1589 errs() << "\n\nIncomplete schedule models found.\n"
1590 << "- Consider setting 'CompleteModel = 0' while developing new models.\n"
1591 << "- Pseudo instructions can be marked with 'hasNoSchedulingInfo = 1'.\n"
1592 << "- Instructions should usually have Sched<[...]> as a superclass, "
1593 "you may temporarily use an empty list.\n"
1594 << "- Instructions related to unsupported features can be excluded with "
1595 "list<Predicate> UnsupportedFeatures = [HasA,..,HasY]; in the "
1596 "processor model.\n\n";
1597 PrintFatalError("Incomplete schedule model");
1598 }
1599 }
1600
1601 // Collect itinerary class resources for each processor.
collectItinProcResources(Record * ItinClassDef)1602 void CodeGenSchedModels::collectItinProcResources(Record *ItinClassDef) {
1603 for (unsigned PIdx = 0, PEnd = ProcModels.size(); PIdx != PEnd; ++PIdx) {
1604 const CodeGenProcModel &PM = ProcModels[PIdx];
1605 // For all ItinRW entries.
1606 bool HasMatch = false;
1607 for (RecIter II = PM.ItinRWDefs.begin(), IE = PM.ItinRWDefs.end();
1608 II != IE; ++II) {
1609 RecVec Matched = (*II)->getValueAsListOfDefs("MatchedItinClasses");
1610 if (!std::count(Matched.begin(), Matched.end(), ItinClassDef))
1611 continue;
1612 if (HasMatch)
1613 PrintFatalError((*II)->getLoc(), "Duplicate itinerary class "
1614 + ItinClassDef->getName()
1615 + " in ItinResources for " + PM.ModelName);
1616 HasMatch = true;
1617 IdxVec Writes, Reads;
1618 findRWs((*II)->getValueAsListOfDefs("OperandReadWrites"), Writes, Reads);
1619 IdxVec ProcIndices(1, PIdx);
1620 collectRWResources(Writes, Reads, ProcIndices);
1621 }
1622 }
1623 }
1624
collectRWResources(unsigned RWIdx,bool IsRead,ArrayRef<unsigned> ProcIndices)1625 void CodeGenSchedModels::collectRWResources(unsigned RWIdx, bool IsRead,
1626 ArrayRef<unsigned> ProcIndices) {
1627 const CodeGenSchedRW &SchedRW = getSchedRW(RWIdx, IsRead);
1628 if (SchedRW.TheDef) {
1629 if (!IsRead && SchedRW.TheDef->isSubClassOf("SchedWriteRes")) {
1630 for (unsigned Idx : ProcIndices)
1631 addWriteRes(SchedRW.TheDef, Idx);
1632 }
1633 else if (IsRead && SchedRW.TheDef->isSubClassOf("SchedReadAdvance")) {
1634 for (unsigned Idx : ProcIndices)
1635 addReadAdvance(SchedRW.TheDef, Idx);
1636 }
1637 }
1638 for (RecIter AI = SchedRW.Aliases.begin(), AE = SchedRW.Aliases.end();
1639 AI != AE; ++AI) {
1640 IdxVec AliasProcIndices;
1641 if ((*AI)->getValueInit("SchedModel")->isComplete()) {
1642 AliasProcIndices.push_back(
1643 getProcModel((*AI)->getValueAsDef("SchedModel")).Index);
1644 }
1645 else
1646 AliasProcIndices = ProcIndices;
1647 const CodeGenSchedRW &AliasRW = getSchedRW((*AI)->getValueAsDef("AliasRW"));
1648 assert(AliasRW.IsRead == IsRead && "cannot alias reads to writes");
1649
1650 IdxVec ExpandedRWs;
1651 expandRWSequence(AliasRW.Index, ExpandedRWs, IsRead);
1652 for (IdxIter SI = ExpandedRWs.begin(), SE = ExpandedRWs.end();
1653 SI != SE; ++SI) {
1654 collectRWResources(*SI, IsRead, AliasProcIndices);
1655 }
1656 }
1657 }
1658
1659 // Collect resources for a set of read/write types and processor indices.
collectRWResources(ArrayRef<unsigned> Writes,ArrayRef<unsigned> Reads,ArrayRef<unsigned> ProcIndices)1660 void CodeGenSchedModels::collectRWResources(ArrayRef<unsigned> Writes,
1661 ArrayRef<unsigned> Reads,
1662 ArrayRef<unsigned> ProcIndices) {
1663
1664 for (unsigned Idx : Writes)
1665 collectRWResources(Idx, /*IsRead=*/false, ProcIndices);
1666
1667 for (unsigned Idx : Reads)
1668 collectRWResources(Idx, /*IsRead=*/true, ProcIndices);
1669 }
1670
1671
1672 // Find the processor's resource units for this kind of resource.
findProcResUnits(Record * ProcResKind,const CodeGenProcModel & PM) const1673 Record *CodeGenSchedModels::findProcResUnits(Record *ProcResKind,
1674 const CodeGenProcModel &PM) const {
1675 if (ProcResKind->isSubClassOf("ProcResourceUnits"))
1676 return ProcResKind;
1677
1678 Record *ProcUnitDef = nullptr;
1679 assert(!ProcResourceDefs.empty());
1680 assert(!ProcResGroups.empty());
1681
1682 for (RecIter RI = ProcResourceDefs.begin(), RE = ProcResourceDefs.end();
1683 RI != RE; ++RI) {
1684
1685 if ((*RI)->getValueAsDef("Kind") == ProcResKind
1686 && (*RI)->getValueAsDef("SchedModel") == PM.ModelDef) {
1687 if (ProcUnitDef) {
1688 PrintFatalError((*RI)->getLoc(),
1689 "Multiple ProcessorResourceUnits associated with "
1690 + ProcResKind->getName());
1691 }
1692 ProcUnitDef = *RI;
1693 }
1694 }
1695 for (RecIter RI = ProcResGroups.begin(), RE = ProcResGroups.end();
1696 RI != RE; ++RI) {
1697
1698 if (*RI == ProcResKind
1699 && (*RI)->getValueAsDef("SchedModel") == PM.ModelDef) {
1700 if (ProcUnitDef) {
1701 PrintFatalError((*RI)->getLoc(),
1702 "Multiple ProcessorResourceUnits associated with "
1703 + ProcResKind->getName());
1704 }
1705 ProcUnitDef = *RI;
1706 }
1707 }
1708 if (!ProcUnitDef) {
1709 PrintFatalError(ProcResKind->getLoc(),
1710 "No ProcessorResources associated with "
1711 + ProcResKind->getName());
1712 }
1713 return ProcUnitDef;
1714 }
1715
1716 // Iteratively add a resource and its super resources.
addProcResource(Record * ProcResKind,CodeGenProcModel & PM)1717 void CodeGenSchedModels::addProcResource(Record *ProcResKind,
1718 CodeGenProcModel &PM) {
1719 for (;;) {
1720 Record *ProcResUnits = findProcResUnits(ProcResKind, PM);
1721
1722 // See if this ProcResource is already associated with this processor.
1723 RecIter I = std::find(PM.ProcResourceDefs.begin(),
1724 PM.ProcResourceDefs.end(), ProcResUnits);
1725 if (I != PM.ProcResourceDefs.end())
1726 return;
1727
1728 PM.ProcResourceDefs.push_back(ProcResUnits);
1729 if (ProcResUnits->isSubClassOf("ProcResGroup"))
1730 return;
1731
1732 if (!ProcResUnits->getValueInit("Super")->isComplete())
1733 return;
1734
1735 ProcResKind = ProcResUnits->getValueAsDef("Super");
1736 }
1737 }
1738
1739 // Add resources for a SchedWrite to this processor if they don't exist.
addWriteRes(Record * ProcWriteResDef,unsigned PIdx)1740 void CodeGenSchedModels::addWriteRes(Record *ProcWriteResDef, unsigned PIdx) {
1741 assert(PIdx && "don't add resources to an invalid Processor model");
1742
1743 RecVec &WRDefs = ProcModels[PIdx].WriteResDefs;
1744 RecIter WRI = std::find(WRDefs.begin(), WRDefs.end(), ProcWriteResDef);
1745 if (WRI != WRDefs.end())
1746 return;
1747 WRDefs.push_back(ProcWriteResDef);
1748
1749 // Visit ProcResourceKinds referenced by the newly discovered WriteRes.
1750 RecVec ProcResDefs = ProcWriteResDef->getValueAsListOfDefs("ProcResources");
1751 for (RecIter WritePRI = ProcResDefs.begin(), WritePRE = ProcResDefs.end();
1752 WritePRI != WritePRE; ++WritePRI) {
1753 addProcResource(*WritePRI, ProcModels[PIdx]);
1754 }
1755 }
1756
1757 // Add resources for a ReadAdvance to this processor if they don't exist.
addReadAdvance(Record * ProcReadAdvanceDef,unsigned PIdx)1758 void CodeGenSchedModels::addReadAdvance(Record *ProcReadAdvanceDef,
1759 unsigned PIdx) {
1760 RecVec &RADefs = ProcModels[PIdx].ReadAdvanceDefs;
1761 RecIter I = std::find(RADefs.begin(), RADefs.end(), ProcReadAdvanceDef);
1762 if (I != RADefs.end())
1763 return;
1764 RADefs.push_back(ProcReadAdvanceDef);
1765 }
1766
getProcResourceIdx(Record * PRDef) const1767 unsigned CodeGenProcModel::getProcResourceIdx(Record *PRDef) const {
1768 RecIter PRPos = std::find(ProcResourceDefs.begin(), ProcResourceDefs.end(),
1769 PRDef);
1770 if (PRPos == ProcResourceDefs.end())
1771 PrintFatalError(PRDef->getLoc(), "ProcResource def is not included in "
1772 "the ProcResources list for " + ModelName);
1773 // Idx=0 is reserved for invalid.
1774 return 1 + (PRPos - ProcResourceDefs.begin());
1775 }
1776
isUnsupported(const CodeGenInstruction & Inst) const1777 bool CodeGenProcModel::isUnsupported(const CodeGenInstruction &Inst) const {
1778 for (const Record *TheDef : UnsupportedFeaturesDefs) {
1779 for (const Record *PredDef : Inst.TheDef->getValueAsListOfDefs("Predicates")) {
1780 if (TheDef->getName() == PredDef->getName())
1781 return true;
1782 }
1783 }
1784 return false;
1785 }
1786
1787 #ifndef NDEBUG
dump() const1788 void CodeGenProcModel::dump() const {
1789 dbgs() << Index << ": " << ModelName << " "
1790 << (ModelDef ? ModelDef->getName() : "inferred") << " "
1791 << (ItinsDef ? ItinsDef->getName() : "no itinerary") << '\n';
1792 }
1793
dump() const1794 void CodeGenSchedRW::dump() const {
1795 dbgs() << Name << (IsVariadic ? " (V) " : " ");
1796 if (IsSequence) {
1797 dbgs() << "(";
1798 dumpIdxVec(Sequence);
1799 dbgs() << ")";
1800 }
1801 }
1802
dump(const CodeGenSchedModels * SchedModels) const1803 void CodeGenSchedClass::dump(const CodeGenSchedModels* SchedModels) const {
1804 dbgs() << "SCHEDCLASS " << Index << ":" << Name << '\n'
1805 << " Writes: ";
1806 for (unsigned i = 0, N = Writes.size(); i < N; ++i) {
1807 SchedModels->getSchedWrite(Writes[i]).dump();
1808 if (i < N-1) {
1809 dbgs() << '\n';
1810 dbgs().indent(10);
1811 }
1812 }
1813 dbgs() << "\n Reads: ";
1814 for (unsigned i = 0, N = Reads.size(); i < N; ++i) {
1815 SchedModels->getSchedRead(Reads[i]).dump();
1816 if (i < N-1) {
1817 dbgs() << '\n';
1818 dbgs().indent(10);
1819 }
1820 }
1821 dbgs() << "\n ProcIdx: "; dumpIdxVec(ProcIndices); dbgs() << '\n';
1822 if (!Transitions.empty()) {
1823 dbgs() << "\n Transitions for Proc ";
1824 for (std::vector<CodeGenSchedTransition>::const_iterator
1825 TI = Transitions.begin(), TE = Transitions.end(); TI != TE; ++TI) {
1826 dumpIdxVec(TI->ProcIndices);
1827 }
1828 }
1829 }
1830
dump() const1831 void PredTransitions::dump() const {
1832 dbgs() << "Expanded Variants:\n";
1833 for (std::vector<PredTransition>::const_iterator
1834 TI = TransVec.begin(), TE = TransVec.end(); TI != TE; ++TI) {
1835 dbgs() << "{";
1836 for (SmallVectorImpl<PredCheck>::const_iterator
1837 PCI = TI->PredTerm.begin(), PCE = TI->PredTerm.end();
1838 PCI != PCE; ++PCI) {
1839 if (PCI != TI->PredTerm.begin())
1840 dbgs() << ", ";
1841 dbgs() << SchedModels.getSchedRW(PCI->RWIdx, PCI->IsRead).Name
1842 << ":" << PCI->Predicate->getName();
1843 }
1844 dbgs() << "},\n => {";
1845 for (SmallVectorImpl<SmallVector<unsigned,4> >::const_iterator
1846 WSI = TI->WriteSequences.begin(), WSE = TI->WriteSequences.end();
1847 WSI != WSE; ++WSI) {
1848 dbgs() << "(";
1849 for (SmallVectorImpl<unsigned>::const_iterator
1850 WI = WSI->begin(), WE = WSI->end(); WI != WE; ++WI) {
1851 if (WI != WSI->begin())
1852 dbgs() << ", ";
1853 dbgs() << SchedModels.getSchedWrite(*WI).Name;
1854 }
1855 dbgs() << "),";
1856 }
1857 dbgs() << "}\n";
1858 }
1859 }
1860 #endif // NDEBUG
1861