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 "CodeGenInstruction.h"
17 #include "CodeGenTarget.h"
18 #include "llvm/ADT/MapVector.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/Support/Casting.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/Regex.h"
26 #include "llvm/Support/raw_ostream.h"
27 #include "llvm/TableGen/Error.h"
28 #include <algorithm>
29 #include <iterator>
30 #include <utility>
31
32 using namespace llvm;
33
34 #define DEBUG_TYPE "subtarget-emitter"
35
36 #ifndef NDEBUG
dumpIdxVec(ArrayRef<unsigned> V)37 static void dumpIdxVec(ArrayRef<unsigned> V) {
38 for (unsigned Idx : V)
39 dbgs() << Idx << ", ";
40 }
41 #endif
42
43 namespace {
44
45 // (instrs a, b, ...) Evaluate and union all arguments. Identical to AddOp.
46 struct InstrsOp : public SetTheory::Operator {
apply__anonbd1195e00111::InstrsOp47 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
48 ArrayRef<SMLoc> Loc) override {
49 ST.evaluate(Expr->arg_begin(), Expr->arg_end(), Elts, Loc);
50 }
51 };
52
53 // (instregex "OpcPat",...) Find all instructions matching an opcode pattern.
54 struct InstRegexOp : public SetTheory::Operator {
55 const CodeGenTarget &Target;
InstRegexOp__anonbd1195e00111::InstRegexOp56 InstRegexOp(const CodeGenTarget &t): Target(t) {}
57
58 /// Remove any text inside of parentheses from S.
removeParens__anonbd1195e00111::InstRegexOp59 static std::string removeParens(llvm::StringRef S) {
60 std::string Result;
61 unsigned Paren = 0;
62 // NB: We don't care about escaped parens here.
63 for (char C : S) {
64 switch (C) {
65 case '(':
66 ++Paren;
67 break;
68 case ')':
69 --Paren;
70 break;
71 default:
72 if (Paren == 0)
73 Result += C;
74 }
75 }
76 return Result;
77 }
78
apply__anonbd1195e00111::InstRegexOp79 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
80 ArrayRef<SMLoc> Loc) override {
81 ArrayRef<const CodeGenInstruction *> Instructions =
82 Target.getInstructionsByEnumValue();
83
84 unsigned NumGeneric = Target.getNumFixedInstructions();
85 unsigned NumPseudos = Target.getNumPseudoInstructions();
86 auto Generics = Instructions.slice(0, NumGeneric);
87 auto Pseudos = Instructions.slice(NumGeneric, NumPseudos);
88 auto NonPseudos = Instructions.slice(NumGeneric + NumPseudos);
89
90 for (Init *Arg : make_range(Expr->arg_begin(), Expr->arg_end())) {
91 StringInit *SI = dyn_cast<StringInit>(Arg);
92 if (!SI)
93 PrintFatalError(Loc, "instregex requires pattern string: " +
94 Expr->getAsString());
95 StringRef Original = SI->getValue();
96
97 // Extract a prefix that we can binary search on.
98 static const char RegexMetachars[] = "()^$|*+?.[]\\{}";
99 auto FirstMeta = Original.find_first_of(RegexMetachars);
100
101 // Look for top-level | or ?. We cannot optimize them to binary search.
102 if (removeParens(Original).find_first_of("|?") != std::string::npos)
103 FirstMeta = 0;
104
105 Optional<Regex> Regexpr = None;
106 StringRef Prefix = Original.substr(0, FirstMeta);
107 StringRef PatStr = Original.substr(FirstMeta);
108 if (!PatStr.empty()) {
109 // For the rest use a python-style prefix match.
110 std::string pat = PatStr;
111 if (pat[0] != '^') {
112 pat.insert(0, "^(");
113 pat.insert(pat.end(), ')');
114 }
115 Regexpr = Regex(pat);
116 }
117
118 int NumMatches = 0;
119
120 // The generic opcodes are unsorted, handle them manually.
121 for (auto *Inst : Generics) {
122 StringRef InstName = Inst->TheDef->getName();
123 if (InstName.startswith(Prefix) &&
124 (!Regexpr || Regexpr->match(InstName.substr(Prefix.size())))) {
125 Elts.insert(Inst->TheDef);
126 NumMatches++;
127 }
128 }
129
130 // Target instructions are split into two ranges: pseudo instructions
131 // first, than non-pseudos. Each range is in lexicographical order
132 // sorted by name. Find the sub-ranges that start with our prefix.
133 struct Comp {
134 bool operator()(const CodeGenInstruction *LHS, StringRef RHS) {
135 return LHS->TheDef->getName() < RHS;
136 }
137 bool operator()(StringRef LHS, const CodeGenInstruction *RHS) {
138 return LHS < RHS->TheDef->getName() &&
139 !RHS->TheDef->getName().startswith(LHS);
140 }
141 };
142 auto Range1 =
143 std::equal_range(Pseudos.begin(), Pseudos.end(), Prefix, Comp());
144 auto Range2 = std::equal_range(NonPseudos.begin(), NonPseudos.end(),
145 Prefix, Comp());
146
147 // For these ranges we know that instruction names start with the prefix.
148 // Check if there's a regex that needs to be checked.
149 const auto HandleNonGeneric = [&](const CodeGenInstruction *Inst) {
150 StringRef InstName = Inst->TheDef->getName();
151 if (!Regexpr || Regexpr->match(InstName.substr(Prefix.size()))) {
152 Elts.insert(Inst->TheDef);
153 NumMatches++;
154 }
155 };
156 std::for_each(Range1.first, Range1.second, HandleNonGeneric);
157 std::for_each(Range2.first, Range2.second, HandleNonGeneric);
158
159 if (0 == NumMatches)
160 PrintFatalError(Loc, "instregex has no matches: " + Original);
161 }
162 }
163 };
164
165 } // end anonymous namespace
166
167 /// CodeGenModels ctor interprets machine model records and populates maps.
CodeGenSchedModels(RecordKeeper & RK,const CodeGenTarget & TGT)168 CodeGenSchedModels::CodeGenSchedModels(RecordKeeper &RK,
169 const CodeGenTarget &TGT):
170 Records(RK), Target(TGT) {
171
172 Sets.addFieldExpander("InstRW", "Instrs");
173
174 // Allow Set evaluation to recognize the dags used in InstRW records:
175 // (instrs Op1, Op1...)
176 Sets.addOperator("instrs", llvm::make_unique<InstrsOp>());
177 Sets.addOperator("instregex", llvm::make_unique<InstRegexOp>(Target));
178
179 // Instantiate a CodeGenProcModel for each SchedMachineModel with the values
180 // that are explicitly referenced in tablegen records. Resources associated
181 // with each processor will be derived later. Populate ProcModelMap with the
182 // CodeGenProcModel instances.
183 collectProcModels();
184
185 // Instantiate a CodeGenSchedRW for each SchedReadWrite record explicitly
186 // defined, and populate SchedReads and SchedWrites vectors. Implicit
187 // SchedReadWrites that represent sequences derived from expanded variant will
188 // be inferred later.
189 collectSchedRW();
190
191 // Instantiate a CodeGenSchedClass for each unique SchedRW signature directly
192 // required by an instruction definition, and populate SchedClassIdxMap. Set
193 // NumItineraryClasses to the number of explicit itinerary classes referenced
194 // by instructions. Set NumInstrSchedClasses to the number of itinerary
195 // classes plus any classes implied by instructions that derive from class
196 // Sched and provide SchedRW list. This does not infer any new classes from
197 // SchedVariant.
198 collectSchedClasses();
199
200 // Find instruction itineraries for each processor. Sort and populate
201 // CodeGenProcModel::ItinDefList. (Cycle-to-cycle itineraries). This requires
202 // all itinerary classes to be discovered.
203 collectProcItins();
204
205 // Find ItinRW records for each processor and itinerary class.
206 // (For per-operand resources mapped to itinerary classes).
207 collectProcItinRW();
208
209 // Find UnsupportedFeatures records for each processor.
210 // (For per-operand resources mapped to itinerary classes).
211 collectProcUnsupportedFeatures();
212
213 // Infer new SchedClasses from SchedVariant.
214 inferSchedClasses();
215
216 // Populate each CodeGenProcModel's WriteResDefs, ReadAdvanceDefs, and
217 // ProcResourceDefs.
218 LLVM_DEBUG(
219 dbgs() << "\n+++ RESOURCE DEFINITIONS (collectProcResources) +++\n");
220 collectProcResources();
221
222 // Collect optional processor description.
223 collectOptionalProcessorInfo();
224
225 checkCompleteness();
226 }
227
collectRetireControlUnits()228 void CodeGenSchedModels::collectRetireControlUnits() {
229 RecVec Units = Records.getAllDerivedDefinitions("RetireControlUnit");
230
231 for (Record *RCU : Units) {
232 CodeGenProcModel &PM = getProcModel(RCU->getValueAsDef("SchedModel"));
233 if (PM.RetireControlUnit) {
234 PrintError(RCU->getLoc(),
235 "Expected a single RetireControlUnit definition");
236 PrintNote(PM.RetireControlUnit->getLoc(),
237 "Previous definition of RetireControlUnit was here");
238 }
239 PM.RetireControlUnit = RCU;
240 }
241 }
242
243 /// Collect optional processor information.
collectOptionalProcessorInfo()244 void CodeGenSchedModels::collectOptionalProcessorInfo() {
245 // Find register file definitions for each processor.
246 collectRegisterFiles();
247
248 // Collect processor RetireControlUnit descriptors if available.
249 collectRetireControlUnits();
250
251 // Find pfm counter definitions for each processor.
252 collectPfmCounters();
253
254 checkCompleteness();
255 }
256
257 /// Gather all processor models.
collectProcModels()258 void CodeGenSchedModels::collectProcModels() {
259 RecVec ProcRecords = Records.getAllDerivedDefinitions("Processor");
260 llvm::sort(ProcRecords.begin(), ProcRecords.end(), LessRecordFieldName());
261
262 // Reserve space because we can. Reallocation would be ok.
263 ProcModels.reserve(ProcRecords.size()+1);
264
265 // Use idx=0 for NoModel/NoItineraries.
266 Record *NoModelDef = Records.getDef("NoSchedModel");
267 Record *NoItinsDef = Records.getDef("NoItineraries");
268 ProcModels.emplace_back(0, "NoSchedModel", NoModelDef, NoItinsDef);
269 ProcModelMap[NoModelDef] = 0;
270
271 // For each processor, find a unique machine model.
272 LLVM_DEBUG(dbgs() << "+++ PROCESSOR MODELs (addProcModel) +++\n");
273 for (Record *ProcRecord : ProcRecords)
274 addProcModel(ProcRecord);
275 }
276
277 /// Get a unique processor model based on the defined MachineModel and
278 /// ProcessorItineraries.
addProcModel(Record * ProcDef)279 void CodeGenSchedModels::addProcModel(Record *ProcDef) {
280 Record *ModelKey = getModelOrItinDef(ProcDef);
281 if (!ProcModelMap.insert(std::make_pair(ModelKey, ProcModels.size())).second)
282 return;
283
284 std::string Name = ModelKey->getName();
285 if (ModelKey->isSubClassOf("SchedMachineModel")) {
286 Record *ItinsDef = ModelKey->getValueAsDef("Itineraries");
287 ProcModels.emplace_back(ProcModels.size(), Name, ModelKey, ItinsDef);
288 }
289 else {
290 // An itinerary is defined without a machine model. Infer a new model.
291 if (!ModelKey->getValueAsListOfDefs("IID").empty())
292 Name = Name + "Model";
293 ProcModels.emplace_back(ProcModels.size(), Name,
294 ProcDef->getValueAsDef("SchedModel"), ModelKey);
295 }
296 LLVM_DEBUG(ProcModels.back().dump());
297 }
298
299 // Recursively find all reachable SchedReadWrite records.
scanSchedRW(Record * RWDef,RecVec & RWDefs,SmallPtrSet<Record *,16> & RWSet)300 static void scanSchedRW(Record *RWDef, RecVec &RWDefs,
301 SmallPtrSet<Record*, 16> &RWSet) {
302 if (!RWSet.insert(RWDef).second)
303 return;
304 RWDefs.push_back(RWDef);
305 // Reads don't currently have sequence records, but it can be added later.
306 if (RWDef->isSubClassOf("WriteSequence")) {
307 RecVec Seq = RWDef->getValueAsListOfDefs("Writes");
308 for (Record *WSRec : Seq)
309 scanSchedRW(WSRec, RWDefs, RWSet);
310 }
311 else if (RWDef->isSubClassOf("SchedVariant")) {
312 // Visit each variant (guarded by a different predicate).
313 RecVec Vars = RWDef->getValueAsListOfDefs("Variants");
314 for (Record *Variant : Vars) {
315 // Visit each RW in the sequence selected by the current variant.
316 RecVec Selected = Variant->getValueAsListOfDefs("Selected");
317 for (Record *SelDef : Selected)
318 scanSchedRW(SelDef, RWDefs, RWSet);
319 }
320 }
321 }
322
323 // Collect and sort all SchedReadWrites reachable via tablegen records.
324 // More may be inferred later when inferring new SchedClasses from variants.
collectSchedRW()325 void CodeGenSchedModels::collectSchedRW() {
326 // Reserve idx=0 for invalid writes/reads.
327 SchedWrites.resize(1);
328 SchedReads.resize(1);
329
330 SmallPtrSet<Record*, 16> RWSet;
331
332 // Find all SchedReadWrites referenced by instruction defs.
333 RecVec SWDefs, SRDefs;
334 for (const CodeGenInstruction *Inst : Target.getInstructionsByEnumValue()) {
335 Record *SchedDef = Inst->TheDef;
336 if (SchedDef->isValueUnset("SchedRW"))
337 continue;
338 RecVec RWs = SchedDef->getValueAsListOfDefs("SchedRW");
339 for (Record *RW : RWs) {
340 if (RW->isSubClassOf("SchedWrite"))
341 scanSchedRW(RW, SWDefs, RWSet);
342 else {
343 assert(RW->isSubClassOf("SchedRead") && "Unknown SchedReadWrite");
344 scanSchedRW(RW, SRDefs, RWSet);
345 }
346 }
347 }
348 // Find all ReadWrites referenced by InstRW.
349 RecVec InstRWDefs = Records.getAllDerivedDefinitions("InstRW");
350 for (Record *InstRWDef : InstRWDefs) {
351 // For all OperandReadWrites.
352 RecVec RWDefs = InstRWDef->getValueAsListOfDefs("OperandReadWrites");
353 for (Record *RWDef : RWDefs) {
354 if (RWDef->isSubClassOf("SchedWrite"))
355 scanSchedRW(RWDef, SWDefs, RWSet);
356 else {
357 assert(RWDef->isSubClassOf("SchedRead") && "Unknown SchedReadWrite");
358 scanSchedRW(RWDef, SRDefs, RWSet);
359 }
360 }
361 }
362 // Find all ReadWrites referenced by ItinRW.
363 RecVec ItinRWDefs = Records.getAllDerivedDefinitions("ItinRW");
364 for (Record *ItinRWDef : ItinRWDefs) {
365 // For all OperandReadWrites.
366 RecVec RWDefs = ItinRWDef->getValueAsListOfDefs("OperandReadWrites");
367 for (Record *RWDef : RWDefs) {
368 if (RWDef->isSubClassOf("SchedWrite"))
369 scanSchedRW(RWDef, SWDefs, RWSet);
370 else {
371 assert(RWDef->isSubClassOf("SchedRead") && "Unknown SchedReadWrite");
372 scanSchedRW(RWDef, SRDefs, RWSet);
373 }
374 }
375 }
376 // Find all ReadWrites referenced by SchedAlias. AliasDefs needs to be sorted
377 // for the loop below that initializes Alias vectors.
378 RecVec AliasDefs = Records.getAllDerivedDefinitions("SchedAlias");
379 llvm::sort(AliasDefs.begin(), AliasDefs.end(), LessRecord());
380 for (Record *ADef : AliasDefs) {
381 Record *MatchDef = ADef->getValueAsDef("MatchRW");
382 Record *AliasDef = ADef->getValueAsDef("AliasRW");
383 if (MatchDef->isSubClassOf("SchedWrite")) {
384 if (!AliasDef->isSubClassOf("SchedWrite"))
385 PrintFatalError(ADef->getLoc(), "SchedWrite Alias must be SchedWrite");
386 scanSchedRW(AliasDef, SWDefs, RWSet);
387 }
388 else {
389 assert(MatchDef->isSubClassOf("SchedRead") && "Unknown SchedReadWrite");
390 if (!AliasDef->isSubClassOf("SchedRead"))
391 PrintFatalError(ADef->getLoc(), "SchedRead Alias must be SchedRead");
392 scanSchedRW(AliasDef, SRDefs, RWSet);
393 }
394 }
395 // Sort and add the SchedReadWrites directly referenced by instructions or
396 // itinerary resources. Index reads and writes in separate domains.
397 llvm::sort(SWDefs.begin(), SWDefs.end(), LessRecord());
398 for (Record *SWDef : SWDefs) {
399 assert(!getSchedRWIdx(SWDef, /*IsRead=*/false) && "duplicate SchedWrite");
400 SchedWrites.emplace_back(SchedWrites.size(), SWDef);
401 }
402 llvm::sort(SRDefs.begin(), SRDefs.end(), LessRecord());
403 for (Record *SRDef : SRDefs) {
404 assert(!getSchedRWIdx(SRDef, /*IsRead-*/true) && "duplicate SchedWrite");
405 SchedReads.emplace_back(SchedReads.size(), SRDef);
406 }
407 // Initialize WriteSequence vectors.
408 for (CodeGenSchedRW &CGRW : SchedWrites) {
409 if (!CGRW.IsSequence)
410 continue;
411 findRWs(CGRW.TheDef->getValueAsListOfDefs("Writes"), CGRW.Sequence,
412 /*IsRead=*/false);
413 }
414 // Initialize Aliases vectors.
415 for (Record *ADef : AliasDefs) {
416 Record *AliasDef = ADef->getValueAsDef("AliasRW");
417 getSchedRW(AliasDef).IsAlias = true;
418 Record *MatchDef = ADef->getValueAsDef("MatchRW");
419 CodeGenSchedRW &RW = getSchedRW(MatchDef);
420 if (RW.IsAlias)
421 PrintFatalError(ADef->getLoc(), "Cannot Alias an Alias");
422 RW.Aliases.push_back(ADef);
423 }
424 LLVM_DEBUG(
425 dbgs() << "\n+++ SCHED READS and WRITES (collectSchedRW) +++\n";
426 for (unsigned WIdx = 0, WEnd = SchedWrites.size(); WIdx != WEnd; ++WIdx) {
427 dbgs() << WIdx << ": ";
428 SchedWrites[WIdx].dump();
429 dbgs() << '\n';
430 } for (unsigned RIdx = 0, REnd = SchedReads.size(); RIdx != REnd;
431 ++RIdx) {
432 dbgs() << RIdx << ": ";
433 SchedReads[RIdx].dump();
434 dbgs() << '\n';
435 } RecVec RWDefs = Records.getAllDerivedDefinitions("SchedReadWrite");
436 for (Record *RWDef
437 : RWDefs) {
438 if (!getSchedRWIdx(RWDef, RWDef->isSubClassOf("SchedRead"))) {
439 StringRef Name = RWDef->getName();
440 if (Name != "NoWrite" && Name != "ReadDefault")
441 dbgs() << "Unused SchedReadWrite " << Name << '\n';
442 }
443 });
444 }
445
446 /// Compute a SchedWrite name from a sequence of writes.
genRWName(ArrayRef<unsigned> Seq,bool IsRead)447 std::string CodeGenSchedModels::genRWName(ArrayRef<unsigned> Seq, bool IsRead) {
448 std::string Name("(");
449 for (auto I = Seq.begin(), E = Seq.end(); I != E; ++I) {
450 if (I != Seq.begin())
451 Name += '_';
452 Name += getSchedRW(*I, IsRead).Name;
453 }
454 Name += ')';
455 return Name;
456 }
457
getSchedRWIdx(const Record * Def,bool IsRead) const458 unsigned CodeGenSchedModels::getSchedRWIdx(const Record *Def,
459 bool IsRead) const {
460 const std::vector<CodeGenSchedRW> &RWVec = IsRead ? SchedReads : SchedWrites;
461 const auto I = find_if(
462 RWVec, [Def](const CodeGenSchedRW &RW) { return RW.TheDef == Def; });
463 return I == RWVec.end() ? 0 : std::distance(RWVec.begin(), I);
464 }
465
hasReadOfWrite(Record * WriteDef) const466 bool CodeGenSchedModels::hasReadOfWrite(Record *WriteDef) const {
467 for (const CodeGenSchedRW &Read : SchedReads) {
468 Record *ReadDef = Read.TheDef;
469 if (!ReadDef || !ReadDef->isSubClassOf("ProcReadAdvance"))
470 continue;
471
472 RecVec ValidWrites = ReadDef->getValueAsListOfDefs("ValidWrites");
473 if (is_contained(ValidWrites, WriteDef)) {
474 return true;
475 }
476 }
477 return false;
478 }
479
splitSchedReadWrites(const RecVec & RWDefs,RecVec & WriteDefs,RecVec & ReadDefs)480 static void splitSchedReadWrites(const RecVec &RWDefs,
481 RecVec &WriteDefs, RecVec &ReadDefs) {
482 for (Record *RWDef : RWDefs) {
483 if (RWDef->isSubClassOf("SchedWrite"))
484 WriteDefs.push_back(RWDef);
485 else {
486 assert(RWDef->isSubClassOf("SchedRead") && "unknown SchedReadWrite");
487 ReadDefs.push_back(RWDef);
488 }
489 }
490 }
491
492 // Split the SchedReadWrites defs and call findRWs for each list.
findRWs(const RecVec & RWDefs,IdxVec & Writes,IdxVec & Reads) const493 void CodeGenSchedModels::findRWs(const RecVec &RWDefs,
494 IdxVec &Writes, IdxVec &Reads) const {
495 RecVec WriteDefs;
496 RecVec ReadDefs;
497 splitSchedReadWrites(RWDefs, WriteDefs, ReadDefs);
498 findRWs(WriteDefs, Writes, false);
499 findRWs(ReadDefs, Reads, true);
500 }
501
502 // Call getSchedRWIdx for all elements in a sequence of SchedRW defs.
findRWs(const RecVec & RWDefs,IdxVec & RWs,bool IsRead) const503 void CodeGenSchedModels::findRWs(const RecVec &RWDefs, IdxVec &RWs,
504 bool IsRead) const {
505 for (Record *RWDef : RWDefs) {
506 unsigned Idx = getSchedRWIdx(RWDef, IsRead);
507 assert(Idx && "failed to collect SchedReadWrite");
508 RWs.push_back(Idx);
509 }
510 }
511
expandRWSequence(unsigned RWIdx,IdxVec & RWSeq,bool IsRead) const512 void CodeGenSchedModels::expandRWSequence(unsigned RWIdx, IdxVec &RWSeq,
513 bool IsRead) const {
514 const CodeGenSchedRW &SchedRW = getSchedRW(RWIdx, IsRead);
515 if (!SchedRW.IsSequence) {
516 RWSeq.push_back(RWIdx);
517 return;
518 }
519 int Repeat =
520 SchedRW.TheDef ? SchedRW.TheDef->getValueAsInt("Repeat") : 1;
521 for (int i = 0; i < Repeat; ++i) {
522 for (unsigned I : SchedRW.Sequence) {
523 expandRWSequence(I, RWSeq, IsRead);
524 }
525 }
526 }
527
528 // Expand a SchedWrite as a sequence following any aliases that coincide with
529 // the given processor model.
expandRWSeqForProc(unsigned RWIdx,IdxVec & RWSeq,bool IsRead,const CodeGenProcModel & ProcModel) const530 void CodeGenSchedModels::expandRWSeqForProc(
531 unsigned RWIdx, IdxVec &RWSeq, bool IsRead,
532 const CodeGenProcModel &ProcModel) const {
533
534 const CodeGenSchedRW &SchedWrite = getSchedRW(RWIdx, IsRead);
535 Record *AliasDef = nullptr;
536 for (const Record *Rec : SchedWrite.Aliases) {
537 const CodeGenSchedRW &AliasRW = getSchedRW(Rec->getValueAsDef("AliasRW"));
538 if (Rec->getValueInit("SchedModel")->isComplete()) {
539 Record *ModelDef = Rec->getValueAsDef("SchedModel");
540 if (&getProcModel(ModelDef) != &ProcModel)
541 continue;
542 }
543 if (AliasDef)
544 PrintFatalError(AliasRW.TheDef->getLoc(), "Multiple aliases "
545 "defined for processor " + ProcModel.ModelName +
546 " Ensure only one SchedAlias exists per RW.");
547 AliasDef = AliasRW.TheDef;
548 }
549 if (AliasDef) {
550 expandRWSeqForProc(getSchedRWIdx(AliasDef, IsRead),
551 RWSeq, IsRead,ProcModel);
552 return;
553 }
554 if (!SchedWrite.IsSequence) {
555 RWSeq.push_back(RWIdx);
556 return;
557 }
558 int Repeat =
559 SchedWrite.TheDef ? SchedWrite.TheDef->getValueAsInt("Repeat") : 1;
560 for (int I = 0, E = Repeat; I < E; ++I) {
561 for (unsigned Idx : SchedWrite.Sequence) {
562 expandRWSeqForProc(Idx, RWSeq, IsRead, ProcModel);
563 }
564 }
565 }
566
567 // Find the existing SchedWrite that models this sequence of writes.
findRWForSequence(ArrayRef<unsigned> Seq,bool IsRead)568 unsigned CodeGenSchedModels::findRWForSequence(ArrayRef<unsigned> Seq,
569 bool IsRead) {
570 std::vector<CodeGenSchedRW> &RWVec = IsRead ? SchedReads : SchedWrites;
571
572 auto I = find_if(RWVec, [Seq](CodeGenSchedRW &RW) {
573 return makeArrayRef(RW.Sequence) == Seq;
574 });
575 // Index zero reserved for invalid RW.
576 return I == RWVec.end() ? 0 : std::distance(RWVec.begin(), I);
577 }
578
579 /// Add this ReadWrite if it doesn't already exist.
findOrInsertRW(ArrayRef<unsigned> Seq,bool IsRead)580 unsigned CodeGenSchedModels::findOrInsertRW(ArrayRef<unsigned> Seq,
581 bool IsRead) {
582 assert(!Seq.empty() && "cannot insert empty sequence");
583 if (Seq.size() == 1)
584 return Seq.back();
585
586 unsigned Idx = findRWForSequence(Seq, IsRead);
587 if (Idx)
588 return Idx;
589
590 std::vector<CodeGenSchedRW> &RWVec = IsRead ? SchedReads : SchedWrites;
591 unsigned RWIdx = RWVec.size();
592 CodeGenSchedRW SchedRW(RWIdx, IsRead, Seq, genRWName(Seq, IsRead));
593 RWVec.push_back(SchedRW);
594 return RWIdx;
595 }
596
597 /// Visit all the instruction definitions for this target to gather and
598 /// enumerate the itinerary classes. These are the explicitly specified
599 /// SchedClasses. More SchedClasses may be inferred.
collectSchedClasses()600 void CodeGenSchedModels::collectSchedClasses() {
601
602 // NoItinerary is always the first class at Idx=0
603 assert(SchedClasses.empty() && "Expected empty sched class");
604 SchedClasses.emplace_back(0, "NoInstrModel",
605 Records.getDef("NoItinerary"));
606 SchedClasses.back().ProcIndices.push_back(0);
607
608 // Create a SchedClass for each unique combination of itinerary class and
609 // SchedRW list.
610 for (const CodeGenInstruction *Inst : Target.getInstructionsByEnumValue()) {
611 Record *ItinDef = Inst->TheDef->getValueAsDef("Itinerary");
612 IdxVec Writes, Reads;
613 if (!Inst->TheDef->isValueUnset("SchedRW"))
614 findRWs(Inst->TheDef->getValueAsListOfDefs("SchedRW"), Writes, Reads);
615
616 // ProcIdx == 0 indicates the class applies to all processors.
617 unsigned SCIdx = addSchedClass(ItinDef, Writes, Reads, /*ProcIndices*/{0});
618 InstrClassMap[Inst->TheDef] = SCIdx;
619 }
620 // Create classes for InstRW defs.
621 RecVec InstRWDefs = Records.getAllDerivedDefinitions("InstRW");
622 llvm::sort(InstRWDefs.begin(), InstRWDefs.end(), LessRecord());
623 LLVM_DEBUG(dbgs() << "\n+++ SCHED CLASSES (createInstRWClass) +++\n");
624 for (Record *RWDef : InstRWDefs)
625 createInstRWClass(RWDef);
626
627 NumInstrSchedClasses = SchedClasses.size();
628
629 bool EnableDump = false;
630 LLVM_DEBUG(EnableDump = true);
631 if (!EnableDump)
632 return;
633
634 LLVM_DEBUG(
635 dbgs()
636 << "\n+++ ITINERARIES and/or MACHINE MODELS (collectSchedClasses) +++\n");
637 for (const CodeGenInstruction *Inst : Target.getInstructionsByEnumValue()) {
638 StringRef InstName = Inst->TheDef->getName();
639 unsigned SCIdx = getSchedClassIdx(*Inst);
640 if (!SCIdx) {
641 LLVM_DEBUG({
642 if (!Inst->hasNoSchedulingInfo)
643 dbgs() << "No machine model for " << Inst->TheDef->getName() << '\n';
644 });
645 continue;
646 }
647 CodeGenSchedClass &SC = getSchedClass(SCIdx);
648 if (SC.ProcIndices[0] != 0)
649 PrintFatalError(Inst->TheDef->getLoc(), "Instruction's sched class "
650 "must not be subtarget specific.");
651
652 IdxVec ProcIndices;
653 if (SC.ItinClassDef->getName() != "NoItinerary") {
654 ProcIndices.push_back(0);
655 dbgs() << "Itinerary for " << InstName << ": "
656 << SC.ItinClassDef->getName() << '\n';
657 }
658 if (!SC.Writes.empty()) {
659 ProcIndices.push_back(0);
660 LLVM_DEBUG({
661 dbgs() << "SchedRW machine model for " << InstName;
662 for (IdxIter WI = SC.Writes.begin(), WE = SC.Writes.end(); WI != WE;
663 ++WI)
664 dbgs() << " " << SchedWrites[*WI].Name;
665 for (IdxIter RI = SC.Reads.begin(), RE = SC.Reads.end(); RI != RE; ++RI)
666 dbgs() << " " << SchedReads[*RI].Name;
667 dbgs() << '\n';
668 });
669 }
670 const RecVec &RWDefs = SchedClasses[SCIdx].InstRWs;
671 for (Record *RWDef : RWDefs) {
672 const CodeGenProcModel &ProcModel =
673 getProcModel(RWDef->getValueAsDef("SchedModel"));
674 ProcIndices.push_back(ProcModel.Index);
675 LLVM_DEBUG(dbgs() << "InstRW on " << ProcModel.ModelName << " for "
676 << InstName);
677 IdxVec Writes;
678 IdxVec Reads;
679 findRWs(RWDef->getValueAsListOfDefs("OperandReadWrites"),
680 Writes, Reads);
681 LLVM_DEBUG({
682 for (unsigned WIdx : Writes)
683 dbgs() << " " << SchedWrites[WIdx].Name;
684 for (unsigned RIdx : Reads)
685 dbgs() << " " << SchedReads[RIdx].Name;
686 dbgs() << '\n';
687 });
688 }
689 // If ProcIndices contains zero, the class applies to all processors.
690 LLVM_DEBUG({
691 if (!std::count(ProcIndices.begin(), ProcIndices.end(), 0)) {
692 for (const CodeGenProcModel &PM : ProcModels) {
693 if (!std::count(ProcIndices.begin(), ProcIndices.end(), PM.Index))
694 dbgs() << "No machine model for " << Inst->TheDef->getName()
695 << " on processor " << PM.ModelName << '\n';
696 }
697 }
698 });
699 }
700 }
701
702 // Get the SchedClass index for an instruction.
703 unsigned
getSchedClassIdx(const CodeGenInstruction & Inst) const704 CodeGenSchedModels::getSchedClassIdx(const CodeGenInstruction &Inst) const {
705 return InstrClassMap.lookup(Inst.TheDef);
706 }
707
708 std::string
createSchedClassName(Record * ItinClassDef,ArrayRef<unsigned> OperWrites,ArrayRef<unsigned> OperReads)709 CodeGenSchedModels::createSchedClassName(Record *ItinClassDef,
710 ArrayRef<unsigned> OperWrites,
711 ArrayRef<unsigned> OperReads) {
712
713 std::string Name;
714 if (ItinClassDef && ItinClassDef->getName() != "NoItinerary")
715 Name = ItinClassDef->getName();
716 for (unsigned Idx : OperWrites) {
717 if (!Name.empty())
718 Name += '_';
719 Name += SchedWrites[Idx].Name;
720 }
721 for (unsigned Idx : OperReads) {
722 Name += '_';
723 Name += SchedReads[Idx].Name;
724 }
725 return Name;
726 }
727
createSchedClassName(const RecVec & InstDefs)728 std::string CodeGenSchedModels::createSchedClassName(const RecVec &InstDefs) {
729
730 std::string Name;
731 for (RecIter I = InstDefs.begin(), E = InstDefs.end(); I != E; ++I) {
732 if (I != InstDefs.begin())
733 Name += '_';
734 Name += (*I)->getName();
735 }
736 return Name;
737 }
738
739 /// Add an inferred sched class from an itinerary class and per-operand list of
740 /// SchedWrites and SchedReads. ProcIndices contains the set of IDs of
741 /// processors that may utilize this class.
addSchedClass(Record * ItinClassDef,ArrayRef<unsigned> OperWrites,ArrayRef<unsigned> OperReads,ArrayRef<unsigned> ProcIndices)742 unsigned CodeGenSchedModels::addSchedClass(Record *ItinClassDef,
743 ArrayRef<unsigned> OperWrites,
744 ArrayRef<unsigned> OperReads,
745 ArrayRef<unsigned> ProcIndices) {
746 assert(!ProcIndices.empty() && "expect at least one ProcIdx");
747
748 auto IsKeyEqual = [=](const CodeGenSchedClass &SC) {
749 return SC.isKeyEqual(ItinClassDef, OperWrites, OperReads);
750 };
751
752 auto I = find_if(make_range(schedClassBegin(), schedClassEnd()), IsKeyEqual);
753 unsigned Idx = I == schedClassEnd() ? 0 : std::distance(schedClassBegin(), I);
754 if (Idx || SchedClasses[0].isKeyEqual(ItinClassDef, OperWrites, OperReads)) {
755 IdxVec PI;
756 std::set_union(SchedClasses[Idx].ProcIndices.begin(),
757 SchedClasses[Idx].ProcIndices.end(),
758 ProcIndices.begin(), ProcIndices.end(),
759 std::back_inserter(PI));
760 SchedClasses[Idx].ProcIndices = std::move(PI);
761 return Idx;
762 }
763 Idx = SchedClasses.size();
764 SchedClasses.emplace_back(Idx,
765 createSchedClassName(ItinClassDef, OperWrites,
766 OperReads),
767 ItinClassDef);
768 CodeGenSchedClass &SC = SchedClasses.back();
769 SC.Writes = OperWrites;
770 SC.Reads = OperReads;
771 SC.ProcIndices = ProcIndices;
772
773 return Idx;
774 }
775
776 // Create classes for each set of opcodes that are in the same InstReadWrite
777 // definition across all processors.
createInstRWClass(Record * InstRWDef)778 void CodeGenSchedModels::createInstRWClass(Record *InstRWDef) {
779 // ClassInstrs will hold an entry for each subset of Instrs in InstRWDef that
780 // intersects with an existing class via a previous InstRWDef. Instrs that do
781 // not intersect with an existing class refer back to their former class as
782 // determined from ItinDef or SchedRW.
783 SmallMapVector<unsigned, SmallVector<Record *, 8>, 4> ClassInstrs;
784 // Sort Instrs into sets.
785 const RecVec *InstDefs = Sets.expand(InstRWDef);
786 if (InstDefs->empty())
787 PrintFatalError(InstRWDef->getLoc(), "No matching instruction opcodes");
788
789 for (Record *InstDef : *InstDefs) {
790 InstClassMapTy::const_iterator Pos = InstrClassMap.find(InstDef);
791 if (Pos == InstrClassMap.end())
792 PrintFatalError(InstDef->getLoc(), "No sched class for instruction.");
793 unsigned SCIdx = Pos->second;
794 ClassInstrs[SCIdx].push_back(InstDef);
795 }
796 // For each set of Instrs, create a new class if necessary, and map or remap
797 // the Instrs to it.
798 for (auto &Entry : ClassInstrs) {
799 unsigned OldSCIdx = Entry.first;
800 ArrayRef<Record*> InstDefs = Entry.second;
801 // If the all instrs in the current class are accounted for, then leave
802 // them mapped to their old class.
803 if (OldSCIdx) {
804 const RecVec &RWDefs = SchedClasses[OldSCIdx].InstRWs;
805 if (!RWDefs.empty()) {
806 const RecVec *OrigInstDefs = Sets.expand(RWDefs[0]);
807 unsigned OrigNumInstrs =
808 count_if(*OrigInstDefs, [&](Record *OIDef) {
809 return InstrClassMap[OIDef] == OldSCIdx;
810 });
811 if (OrigNumInstrs == InstDefs.size()) {
812 assert(SchedClasses[OldSCIdx].ProcIndices[0] == 0 &&
813 "expected a generic SchedClass");
814 Record *RWModelDef = InstRWDef->getValueAsDef("SchedModel");
815 // Make sure we didn't already have a InstRW containing this
816 // instruction on this model.
817 for (Record *RWD : RWDefs) {
818 if (RWD->getValueAsDef("SchedModel") == RWModelDef &&
819 RWModelDef->getValueAsBit("FullInstRWOverlapCheck")) {
820 for (Record *Inst : InstDefs) {
821 PrintFatalError(InstRWDef->getLoc(), "Overlapping InstRW def " +
822 Inst->getName() + " also matches " +
823 RWD->getValue("Instrs")->getValue()->getAsString());
824 }
825 }
826 }
827 LLVM_DEBUG(dbgs() << "InstRW: Reuse SC " << OldSCIdx << ":"
828 << SchedClasses[OldSCIdx].Name << " on "
829 << RWModelDef->getName() << "\n");
830 SchedClasses[OldSCIdx].InstRWs.push_back(InstRWDef);
831 continue;
832 }
833 }
834 }
835 unsigned SCIdx = SchedClasses.size();
836 SchedClasses.emplace_back(SCIdx, createSchedClassName(InstDefs), nullptr);
837 CodeGenSchedClass &SC = SchedClasses.back();
838 LLVM_DEBUG(dbgs() << "InstRW: New SC " << SCIdx << ":" << SC.Name << " on "
839 << InstRWDef->getValueAsDef("SchedModel")->getName()
840 << "\n");
841
842 // Preserve ItinDef and Writes/Reads for processors without an InstRW entry.
843 SC.ItinClassDef = SchedClasses[OldSCIdx].ItinClassDef;
844 SC.Writes = SchedClasses[OldSCIdx].Writes;
845 SC.Reads = SchedClasses[OldSCIdx].Reads;
846 SC.ProcIndices.push_back(0);
847 // If we had an old class, copy it's InstRWs to this new class.
848 if (OldSCIdx) {
849 Record *RWModelDef = InstRWDef->getValueAsDef("SchedModel");
850 for (Record *OldRWDef : SchedClasses[OldSCIdx].InstRWs) {
851 if (OldRWDef->getValueAsDef("SchedModel") == RWModelDef) {
852 for (Record *InstDef : InstDefs) {
853 PrintFatalError(OldRWDef->getLoc(), "Overlapping InstRW def " +
854 InstDef->getName() + " also matches " +
855 OldRWDef->getValue("Instrs")->getValue()->getAsString());
856 }
857 }
858 assert(OldRWDef != InstRWDef &&
859 "SchedClass has duplicate InstRW def");
860 SC.InstRWs.push_back(OldRWDef);
861 }
862 }
863 // Map each Instr to this new class.
864 for (Record *InstDef : InstDefs)
865 InstrClassMap[InstDef] = SCIdx;
866 SC.InstRWs.push_back(InstRWDef);
867 }
868 }
869
870 // True if collectProcItins found anything.
hasItineraries() const871 bool CodeGenSchedModels::hasItineraries() const {
872 for (const CodeGenProcModel &PM : make_range(procModelBegin(),procModelEnd()))
873 if (PM.hasItineraries())
874 return true;
875 return false;
876 }
877
878 // Gather the processor itineraries.
collectProcItins()879 void CodeGenSchedModels::collectProcItins() {
880 LLVM_DEBUG(dbgs() << "\n+++ PROBLEM ITINERARIES (collectProcItins) +++\n");
881 for (CodeGenProcModel &ProcModel : ProcModels) {
882 if (!ProcModel.hasItineraries())
883 continue;
884
885 RecVec ItinRecords = ProcModel.ItinsDef->getValueAsListOfDefs("IID");
886 assert(!ItinRecords.empty() && "ProcModel.hasItineraries is incorrect");
887
888 // Populate ItinDefList with Itinerary records.
889 ProcModel.ItinDefList.resize(NumInstrSchedClasses);
890
891 // Insert each itinerary data record in the correct position within
892 // the processor model's ItinDefList.
893 for (Record *ItinData : ItinRecords) {
894 const Record *ItinDef = ItinData->getValueAsDef("TheClass");
895 bool FoundClass = false;
896
897 for (const CodeGenSchedClass &SC :
898 make_range(schedClassBegin(), schedClassEnd())) {
899 // Multiple SchedClasses may share an itinerary. Update all of them.
900 if (SC.ItinClassDef == ItinDef) {
901 ProcModel.ItinDefList[SC.Index] = ItinData;
902 FoundClass = true;
903 }
904 }
905 if (!FoundClass) {
906 LLVM_DEBUG(dbgs() << ProcModel.ItinsDef->getName()
907 << " missing class for itinerary "
908 << ItinDef->getName() << '\n');
909 }
910 }
911 // Check for missing itinerary entries.
912 assert(!ProcModel.ItinDefList[0] && "NoItinerary class can't have rec");
913 LLVM_DEBUG(
914 for (unsigned i = 1, N = ProcModel.ItinDefList.size(); i < N; ++i) {
915 if (!ProcModel.ItinDefList[i])
916 dbgs() << ProcModel.ItinsDef->getName()
917 << " missing itinerary for class " << SchedClasses[i].Name
918 << '\n';
919 });
920 }
921 }
922
923 // Gather the read/write types for each itinerary class.
collectProcItinRW()924 void CodeGenSchedModels::collectProcItinRW() {
925 RecVec ItinRWDefs = Records.getAllDerivedDefinitions("ItinRW");
926 llvm::sort(ItinRWDefs.begin(), ItinRWDefs.end(), LessRecord());
927 for (Record *RWDef : ItinRWDefs) {
928 if (!RWDef->getValueInit("SchedModel")->isComplete())
929 PrintFatalError(RWDef->getLoc(), "SchedModel is undefined");
930 Record *ModelDef = RWDef->getValueAsDef("SchedModel");
931 ProcModelMapTy::const_iterator I = ProcModelMap.find(ModelDef);
932 if (I == ProcModelMap.end()) {
933 PrintFatalError(RWDef->getLoc(), "Undefined SchedMachineModel "
934 + ModelDef->getName());
935 }
936 ProcModels[I->second].ItinRWDefs.push_back(RWDef);
937 }
938 }
939
940 // Gather the unsupported features for processor models.
collectProcUnsupportedFeatures()941 void CodeGenSchedModels::collectProcUnsupportedFeatures() {
942 for (CodeGenProcModel &ProcModel : ProcModels) {
943 for (Record *Pred : ProcModel.ModelDef->getValueAsListOfDefs("UnsupportedFeatures")) {
944 ProcModel.UnsupportedFeaturesDefs.push_back(Pred);
945 }
946 }
947 }
948
949 /// Infer new classes from existing classes. In the process, this may create new
950 /// SchedWrites from sequences of existing SchedWrites.
inferSchedClasses()951 void CodeGenSchedModels::inferSchedClasses() {
952 LLVM_DEBUG(
953 dbgs() << "\n+++ INFERRING SCHED CLASSES (inferSchedClasses) +++\n");
954 LLVM_DEBUG(dbgs() << NumInstrSchedClasses << " instr sched classes.\n");
955
956 // Visit all existing classes and newly created classes.
957 for (unsigned Idx = 0; Idx != SchedClasses.size(); ++Idx) {
958 assert(SchedClasses[Idx].Index == Idx && "bad SCIdx");
959
960 if (SchedClasses[Idx].ItinClassDef)
961 inferFromItinClass(SchedClasses[Idx].ItinClassDef, Idx);
962 if (!SchedClasses[Idx].InstRWs.empty())
963 inferFromInstRWs(Idx);
964 if (!SchedClasses[Idx].Writes.empty()) {
965 inferFromRW(SchedClasses[Idx].Writes, SchedClasses[Idx].Reads,
966 Idx, SchedClasses[Idx].ProcIndices);
967 }
968 assert(SchedClasses.size() < (NumInstrSchedClasses*6) &&
969 "too many SchedVariants");
970 }
971 }
972
973 /// Infer classes from per-processor itinerary resources.
inferFromItinClass(Record * ItinClassDef,unsigned FromClassIdx)974 void CodeGenSchedModels::inferFromItinClass(Record *ItinClassDef,
975 unsigned FromClassIdx) {
976 for (unsigned PIdx = 0, PEnd = ProcModels.size(); PIdx != PEnd; ++PIdx) {
977 const CodeGenProcModel &PM = ProcModels[PIdx];
978 // For all ItinRW entries.
979 bool HasMatch = false;
980 for (const Record *Rec : PM.ItinRWDefs) {
981 RecVec Matched = Rec->getValueAsListOfDefs("MatchedItinClasses");
982 if (!std::count(Matched.begin(), Matched.end(), ItinClassDef))
983 continue;
984 if (HasMatch)
985 PrintFatalError(Rec->getLoc(), "Duplicate itinerary class "
986 + ItinClassDef->getName()
987 + " in ItinResources for " + PM.ModelName);
988 HasMatch = true;
989 IdxVec Writes, Reads;
990 findRWs(Rec->getValueAsListOfDefs("OperandReadWrites"), Writes, Reads);
991 inferFromRW(Writes, Reads, FromClassIdx, PIdx);
992 }
993 }
994 }
995
996 /// Infer classes from per-processor InstReadWrite definitions.
inferFromInstRWs(unsigned SCIdx)997 void CodeGenSchedModels::inferFromInstRWs(unsigned SCIdx) {
998 for (unsigned I = 0, E = SchedClasses[SCIdx].InstRWs.size(); I != E; ++I) {
999 assert(SchedClasses[SCIdx].InstRWs.size() == E && "InstrRWs was mutated!");
1000 Record *Rec = SchedClasses[SCIdx].InstRWs[I];
1001 const RecVec *InstDefs = Sets.expand(Rec);
1002 RecIter II = InstDefs->begin(), IE = InstDefs->end();
1003 for (; II != IE; ++II) {
1004 if (InstrClassMap[*II] == SCIdx)
1005 break;
1006 }
1007 // If this class no longer has any instructions mapped to it, it has become
1008 // irrelevant.
1009 if (II == IE)
1010 continue;
1011 IdxVec Writes, Reads;
1012 findRWs(Rec->getValueAsListOfDefs("OperandReadWrites"), Writes, Reads);
1013 unsigned PIdx = getProcModel(Rec->getValueAsDef("SchedModel")).Index;
1014 inferFromRW(Writes, Reads, SCIdx, PIdx); // May mutate SchedClasses.
1015 }
1016 }
1017
1018 namespace {
1019
1020 // Helper for substituteVariantOperand.
1021 struct TransVariant {
1022 Record *VarOrSeqDef; // Variant or sequence.
1023 unsigned RWIdx; // Index of this variant or sequence's matched type.
1024 unsigned ProcIdx; // Processor model index or zero for any.
1025 unsigned TransVecIdx; // Index into PredTransitions::TransVec.
1026
TransVariant__anonbd1195e00711::TransVariant1027 TransVariant(Record *def, unsigned rwi, unsigned pi, unsigned ti):
1028 VarOrSeqDef(def), RWIdx(rwi), ProcIdx(pi), TransVecIdx(ti) {}
1029 };
1030
1031 // Associate a predicate with the SchedReadWrite that it guards.
1032 // RWIdx is the index of the read/write variant.
1033 struct PredCheck {
1034 bool IsRead;
1035 unsigned RWIdx;
1036 Record *Predicate;
1037
PredCheck__anonbd1195e00711::PredCheck1038 PredCheck(bool r, unsigned w, Record *p): IsRead(r), RWIdx(w), Predicate(p) {}
1039 };
1040
1041 // A Predicate transition is a list of RW sequences guarded by a PredTerm.
1042 struct PredTransition {
1043 // A predicate term is a conjunction of PredChecks.
1044 SmallVector<PredCheck, 4> PredTerm;
1045 SmallVector<SmallVector<unsigned,4>, 16> WriteSequences;
1046 SmallVector<SmallVector<unsigned,4>, 16> ReadSequences;
1047 SmallVector<unsigned, 4> ProcIndices;
1048 };
1049
1050 // Encapsulate a set of partially constructed transitions.
1051 // The results are built by repeated calls to substituteVariants.
1052 class PredTransitions {
1053 CodeGenSchedModels &SchedModels;
1054
1055 public:
1056 std::vector<PredTransition> TransVec;
1057
PredTransitions(CodeGenSchedModels & sm)1058 PredTransitions(CodeGenSchedModels &sm): SchedModels(sm) {}
1059
1060 void substituteVariantOperand(const SmallVectorImpl<unsigned> &RWSeq,
1061 bool IsRead, unsigned StartIdx);
1062
1063 void substituteVariants(const PredTransition &Trans);
1064
1065 #ifndef NDEBUG
1066 void dump() const;
1067 #endif
1068
1069 private:
1070 bool mutuallyExclusive(Record *PredDef, ArrayRef<PredCheck> Term);
1071 void getIntersectingVariants(
1072 const CodeGenSchedRW &SchedRW, unsigned TransIdx,
1073 std::vector<TransVariant> &IntersectingVariants);
1074 void pushVariant(const TransVariant &VInfo, bool IsRead);
1075 };
1076
1077 } // end anonymous namespace
1078
1079 // Return true if this predicate is mutually exclusive with a PredTerm. This
1080 // degenerates into checking if the predicate is mutually exclusive with any
1081 // predicate in the Term's conjunction.
1082 //
1083 // All predicates associated with a given SchedRW are considered mutually
1084 // exclusive. This should work even if the conditions expressed by the
1085 // predicates are not exclusive because the predicates for a given SchedWrite
1086 // are always checked in the order they are defined in the .td file. Later
1087 // conditions implicitly negate any prior condition.
mutuallyExclusive(Record * PredDef,ArrayRef<PredCheck> Term)1088 bool PredTransitions::mutuallyExclusive(Record *PredDef,
1089 ArrayRef<PredCheck> Term) {
1090 for (const PredCheck &PC: Term) {
1091 if (PC.Predicate == PredDef)
1092 return false;
1093
1094 const CodeGenSchedRW &SchedRW = SchedModels.getSchedRW(PC.RWIdx, PC.IsRead);
1095 assert(SchedRW.HasVariants && "PredCheck must refer to a SchedVariant");
1096 RecVec Variants = SchedRW.TheDef->getValueAsListOfDefs("Variants");
1097 if (any_of(Variants, [PredDef](const Record *R) {
1098 return R->getValueAsDef("Predicate") == PredDef;
1099 }))
1100 return true;
1101 }
1102 return false;
1103 }
1104
hasAliasedVariants(const CodeGenSchedRW & RW,CodeGenSchedModels & SchedModels)1105 static bool hasAliasedVariants(const CodeGenSchedRW &RW,
1106 CodeGenSchedModels &SchedModels) {
1107 if (RW.HasVariants)
1108 return true;
1109
1110 for (Record *Alias : RW.Aliases) {
1111 const CodeGenSchedRW &AliasRW =
1112 SchedModels.getSchedRW(Alias->getValueAsDef("AliasRW"));
1113 if (AliasRW.HasVariants)
1114 return true;
1115 if (AliasRW.IsSequence) {
1116 IdxVec ExpandedRWs;
1117 SchedModels.expandRWSequence(AliasRW.Index, ExpandedRWs, AliasRW.IsRead);
1118 for (unsigned SI : ExpandedRWs) {
1119 if (hasAliasedVariants(SchedModels.getSchedRW(SI, AliasRW.IsRead),
1120 SchedModels))
1121 return true;
1122 }
1123 }
1124 }
1125 return false;
1126 }
1127
hasVariant(ArrayRef<PredTransition> Transitions,CodeGenSchedModels & SchedModels)1128 static bool hasVariant(ArrayRef<PredTransition> Transitions,
1129 CodeGenSchedModels &SchedModels) {
1130 for (const PredTransition &PTI : Transitions) {
1131 for (const SmallVectorImpl<unsigned> &WSI : PTI.WriteSequences)
1132 for (unsigned WI : WSI)
1133 if (hasAliasedVariants(SchedModels.getSchedWrite(WI), SchedModels))
1134 return true;
1135
1136 for (const SmallVectorImpl<unsigned> &RSI : PTI.ReadSequences)
1137 for (unsigned RI : RSI)
1138 if (hasAliasedVariants(SchedModels.getSchedRead(RI), SchedModels))
1139 return true;
1140 }
1141 return false;
1142 }
1143
1144 // Populate IntersectingVariants with any variants or aliased sequences of the
1145 // given SchedRW whose processor indices and predicates are not mutually
1146 // exclusive with the given transition.
getIntersectingVariants(const CodeGenSchedRW & SchedRW,unsigned TransIdx,std::vector<TransVariant> & IntersectingVariants)1147 void PredTransitions::getIntersectingVariants(
1148 const CodeGenSchedRW &SchedRW, unsigned TransIdx,
1149 std::vector<TransVariant> &IntersectingVariants) {
1150
1151 bool GenericRW = false;
1152
1153 std::vector<TransVariant> Variants;
1154 if (SchedRW.HasVariants) {
1155 unsigned VarProcIdx = 0;
1156 if (SchedRW.TheDef->getValueInit("SchedModel")->isComplete()) {
1157 Record *ModelDef = SchedRW.TheDef->getValueAsDef("SchedModel");
1158 VarProcIdx = SchedModels.getProcModel(ModelDef).Index;
1159 }
1160 // Push each variant. Assign TransVecIdx later.
1161 const RecVec VarDefs = SchedRW.TheDef->getValueAsListOfDefs("Variants");
1162 for (Record *VarDef : VarDefs)
1163 Variants.emplace_back(VarDef, SchedRW.Index, VarProcIdx, 0);
1164 if (VarProcIdx == 0)
1165 GenericRW = true;
1166 }
1167 for (RecIter AI = SchedRW.Aliases.begin(), AE = SchedRW.Aliases.end();
1168 AI != AE; ++AI) {
1169 // If either the SchedAlias itself or the SchedReadWrite that it aliases
1170 // to is defined within a processor model, constrain all variants to
1171 // that processor.
1172 unsigned AliasProcIdx = 0;
1173 if ((*AI)->getValueInit("SchedModel")->isComplete()) {
1174 Record *ModelDef = (*AI)->getValueAsDef("SchedModel");
1175 AliasProcIdx = SchedModels.getProcModel(ModelDef).Index;
1176 }
1177 const CodeGenSchedRW &AliasRW =
1178 SchedModels.getSchedRW((*AI)->getValueAsDef("AliasRW"));
1179
1180 if (AliasRW.HasVariants) {
1181 const RecVec VarDefs = AliasRW.TheDef->getValueAsListOfDefs("Variants");
1182 for (Record *VD : VarDefs)
1183 Variants.emplace_back(VD, AliasRW.Index, AliasProcIdx, 0);
1184 }
1185 if (AliasRW.IsSequence)
1186 Variants.emplace_back(AliasRW.TheDef, SchedRW.Index, AliasProcIdx, 0);
1187 if (AliasProcIdx == 0)
1188 GenericRW = true;
1189 }
1190 for (TransVariant &Variant : Variants) {
1191 // Don't expand variants if the processor models don't intersect.
1192 // A zero processor index means any processor.
1193 SmallVectorImpl<unsigned> &ProcIndices = TransVec[TransIdx].ProcIndices;
1194 if (ProcIndices[0] && Variant.ProcIdx) {
1195 unsigned Cnt = std::count(ProcIndices.begin(), ProcIndices.end(),
1196 Variant.ProcIdx);
1197 if (!Cnt)
1198 continue;
1199 if (Cnt > 1) {
1200 const CodeGenProcModel &PM =
1201 *(SchedModels.procModelBegin() + Variant.ProcIdx);
1202 PrintFatalError(Variant.VarOrSeqDef->getLoc(),
1203 "Multiple variants defined for processor " +
1204 PM.ModelName +
1205 " Ensure only one SchedAlias exists per RW.");
1206 }
1207 }
1208 if (Variant.VarOrSeqDef->isSubClassOf("SchedVar")) {
1209 Record *PredDef = Variant.VarOrSeqDef->getValueAsDef("Predicate");
1210 if (mutuallyExclusive(PredDef, TransVec[TransIdx].PredTerm))
1211 continue;
1212 }
1213 if (IntersectingVariants.empty()) {
1214 // The first variant builds on the existing transition.
1215 Variant.TransVecIdx = TransIdx;
1216 IntersectingVariants.push_back(Variant);
1217 }
1218 else {
1219 // Push another copy of the current transition for more variants.
1220 Variant.TransVecIdx = TransVec.size();
1221 IntersectingVariants.push_back(Variant);
1222 TransVec.push_back(TransVec[TransIdx]);
1223 }
1224 }
1225 if (GenericRW && IntersectingVariants.empty()) {
1226 PrintFatalError(SchedRW.TheDef->getLoc(), "No variant of this type has "
1227 "a matching predicate on any processor");
1228 }
1229 }
1230
1231 // Push the Reads/Writes selected by this variant onto the PredTransition
1232 // specified by VInfo.
1233 void PredTransitions::
pushVariant(const TransVariant & VInfo,bool IsRead)1234 pushVariant(const TransVariant &VInfo, bool IsRead) {
1235 PredTransition &Trans = TransVec[VInfo.TransVecIdx];
1236
1237 // If this operand transition is reached through a processor-specific alias,
1238 // then the whole transition is specific to this processor.
1239 if (VInfo.ProcIdx != 0)
1240 Trans.ProcIndices.assign(1, VInfo.ProcIdx);
1241
1242 IdxVec SelectedRWs;
1243 if (VInfo.VarOrSeqDef->isSubClassOf("SchedVar")) {
1244 Record *PredDef = VInfo.VarOrSeqDef->getValueAsDef("Predicate");
1245 Trans.PredTerm.emplace_back(IsRead, VInfo.RWIdx,PredDef);
1246 RecVec SelectedDefs = VInfo.VarOrSeqDef->getValueAsListOfDefs("Selected");
1247 SchedModels.findRWs(SelectedDefs, SelectedRWs, IsRead);
1248 }
1249 else {
1250 assert(VInfo.VarOrSeqDef->isSubClassOf("WriteSequence") &&
1251 "variant must be a SchedVariant or aliased WriteSequence");
1252 SelectedRWs.push_back(SchedModels.getSchedRWIdx(VInfo.VarOrSeqDef, IsRead));
1253 }
1254
1255 const CodeGenSchedRW &SchedRW = SchedModels.getSchedRW(VInfo.RWIdx, IsRead);
1256
1257 SmallVectorImpl<SmallVector<unsigned,4>> &RWSequences = IsRead
1258 ? Trans.ReadSequences : Trans.WriteSequences;
1259 if (SchedRW.IsVariadic) {
1260 unsigned OperIdx = RWSequences.size()-1;
1261 // Make N-1 copies of this transition's last sequence.
1262 RWSequences.insert(RWSequences.end(), SelectedRWs.size() - 1,
1263 RWSequences[OperIdx]);
1264 // Push each of the N elements of the SelectedRWs onto a copy of the last
1265 // sequence (split the current operand into N operands).
1266 // Note that write sequences should be expanded within this loop--the entire
1267 // sequence belongs to a single operand.
1268 for (IdxIter RWI = SelectedRWs.begin(), RWE = SelectedRWs.end();
1269 RWI != RWE; ++RWI, ++OperIdx) {
1270 IdxVec ExpandedRWs;
1271 if (IsRead)
1272 ExpandedRWs.push_back(*RWI);
1273 else
1274 SchedModels.expandRWSequence(*RWI, ExpandedRWs, IsRead);
1275 RWSequences[OperIdx].insert(RWSequences[OperIdx].end(),
1276 ExpandedRWs.begin(), ExpandedRWs.end());
1277 }
1278 assert(OperIdx == RWSequences.size() && "missed a sequence");
1279 }
1280 else {
1281 // Push this transition's expanded sequence onto this transition's last
1282 // sequence (add to the current operand's sequence).
1283 SmallVectorImpl<unsigned> &Seq = RWSequences.back();
1284 IdxVec ExpandedRWs;
1285 for (IdxIter RWI = SelectedRWs.begin(), RWE = SelectedRWs.end();
1286 RWI != RWE; ++RWI) {
1287 if (IsRead)
1288 ExpandedRWs.push_back(*RWI);
1289 else
1290 SchedModels.expandRWSequence(*RWI, ExpandedRWs, IsRead);
1291 }
1292 Seq.insert(Seq.end(), ExpandedRWs.begin(), ExpandedRWs.end());
1293 }
1294 }
1295
1296 // RWSeq is a sequence of all Reads or all Writes for the next read or write
1297 // operand. StartIdx is an index into TransVec where partial results
1298 // starts. RWSeq must be applied to all transitions between StartIdx and the end
1299 // of TransVec.
substituteVariantOperand(const SmallVectorImpl<unsigned> & RWSeq,bool IsRead,unsigned StartIdx)1300 void PredTransitions::substituteVariantOperand(
1301 const SmallVectorImpl<unsigned> &RWSeq, bool IsRead, unsigned StartIdx) {
1302
1303 // Visit each original RW within the current sequence.
1304 for (SmallVectorImpl<unsigned>::const_iterator
1305 RWI = RWSeq.begin(), RWE = RWSeq.end(); RWI != RWE; ++RWI) {
1306 const CodeGenSchedRW &SchedRW = SchedModels.getSchedRW(*RWI, IsRead);
1307 // Push this RW on all partial PredTransitions or distribute variants.
1308 // New PredTransitions may be pushed within this loop which should not be
1309 // revisited (TransEnd must be loop invariant).
1310 for (unsigned TransIdx = StartIdx, TransEnd = TransVec.size();
1311 TransIdx != TransEnd; ++TransIdx) {
1312 // In the common case, push RW onto the current operand's sequence.
1313 if (!hasAliasedVariants(SchedRW, SchedModels)) {
1314 if (IsRead)
1315 TransVec[TransIdx].ReadSequences.back().push_back(*RWI);
1316 else
1317 TransVec[TransIdx].WriteSequences.back().push_back(*RWI);
1318 continue;
1319 }
1320 // Distribute this partial PredTransition across intersecting variants.
1321 // This will push a copies of TransVec[TransIdx] on the back of TransVec.
1322 std::vector<TransVariant> IntersectingVariants;
1323 getIntersectingVariants(SchedRW, TransIdx, IntersectingVariants);
1324 // Now expand each variant on top of its copy of the transition.
1325 for (std::vector<TransVariant>::const_iterator
1326 IVI = IntersectingVariants.begin(),
1327 IVE = IntersectingVariants.end();
1328 IVI != IVE; ++IVI) {
1329 pushVariant(*IVI, IsRead);
1330 }
1331 }
1332 }
1333 }
1334
1335 // For each variant of a Read/Write in Trans, substitute the sequence of
1336 // Read/Writes guarded by the variant. This is exponential in the number of
1337 // variant Read/Writes, but in practice detection of mutually exclusive
1338 // predicates should result in linear growth in the total number variants.
1339 //
1340 // This is one step in a breadth-first search of nested variants.
substituteVariants(const PredTransition & Trans)1341 void PredTransitions::substituteVariants(const PredTransition &Trans) {
1342 // Build up a set of partial results starting at the back of
1343 // PredTransitions. Remember the first new transition.
1344 unsigned StartIdx = TransVec.size();
1345 TransVec.emplace_back();
1346 TransVec.back().PredTerm = Trans.PredTerm;
1347 TransVec.back().ProcIndices = Trans.ProcIndices;
1348
1349 // Visit each original write sequence.
1350 for (SmallVectorImpl<SmallVector<unsigned,4>>::const_iterator
1351 WSI = Trans.WriteSequences.begin(), WSE = Trans.WriteSequences.end();
1352 WSI != WSE; ++WSI) {
1353 // Push a new (empty) write sequence onto all partial Transitions.
1354 for (std::vector<PredTransition>::iterator I =
1355 TransVec.begin() + StartIdx, E = TransVec.end(); I != E; ++I) {
1356 I->WriteSequences.emplace_back();
1357 }
1358 substituteVariantOperand(*WSI, /*IsRead=*/false, StartIdx);
1359 }
1360 // Visit each original read sequence.
1361 for (SmallVectorImpl<SmallVector<unsigned,4>>::const_iterator
1362 RSI = Trans.ReadSequences.begin(), RSE = Trans.ReadSequences.end();
1363 RSI != RSE; ++RSI) {
1364 // Push a new (empty) read sequence onto all partial Transitions.
1365 for (std::vector<PredTransition>::iterator I =
1366 TransVec.begin() + StartIdx, E = TransVec.end(); I != E; ++I) {
1367 I->ReadSequences.emplace_back();
1368 }
1369 substituteVariantOperand(*RSI, /*IsRead=*/true, StartIdx);
1370 }
1371 }
1372
1373 // Create a new SchedClass for each variant found by inferFromRW. Pass
inferFromTransitions(ArrayRef<PredTransition> LastTransitions,unsigned FromClassIdx,CodeGenSchedModels & SchedModels)1374 static void inferFromTransitions(ArrayRef<PredTransition> LastTransitions,
1375 unsigned FromClassIdx,
1376 CodeGenSchedModels &SchedModels) {
1377 // For each PredTransition, create a new CodeGenSchedTransition, which usually
1378 // requires creating a new SchedClass.
1379 for (ArrayRef<PredTransition>::iterator
1380 I = LastTransitions.begin(), E = LastTransitions.end(); I != E; ++I) {
1381 IdxVec OperWritesVariant;
1382 transform(I->WriteSequences, std::back_inserter(OperWritesVariant),
1383 [&SchedModels](ArrayRef<unsigned> WS) {
1384 return SchedModels.findOrInsertRW(WS, /*IsRead=*/false);
1385 });
1386 IdxVec OperReadsVariant;
1387 transform(I->ReadSequences, std::back_inserter(OperReadsVariant),
1388 [&SchedModels](ArrayRef<unsigned> RS) {
1389 return SchedModels.findOrInsertRW(RS, /*IsRead=*/true);
1390 });
1391 CodeGenSchedTransition SCTrans;
1392 SCTrans.ToClassIdx =
1393 SchedModels.addSchedClass(/*ItinClassDef=*/nullptr, OperWritesVariant,
1394 OperReadsVariant, I->ProcIndices);
1395 SCTrans.ProcIndices.assign(I->ProcIndices.begin(), I->ProcIndices.end());
1396 // The final PredTerm is unique set of predicates guarding the transition.
1397 RecVec Preds;
1398 transform(I->PredTerm, std::back_inserter(Preds),
1399 [](const PredCheck &P) {
1400 return P.Predicate;
1401 });
1402 Preds.erase(std::unique(Preds.begin(), Preds.end()), Preds.end());
1403 SCTrans.PredTerm = std::move(Preds);
1404 SchedModels.getSchedClass(FromClassIdx)
1405 .Transitions.push_back(std::move(SCTrans));
1406 }
1407 }
1408
1409 // Create new SchedClasses for the given ReadWrite list. If any of the
1410 // ReadWrites refers to a SchedVariant, create a new SchedClass for each variant
1411 // of the ReadWrite list, following Aliases if necessary.
inferFromRW(ArrayRef<unsigned> OperWrites,ArrayRef<unsigned> OperReads,unsigned FromClassIdx,ArrayRef<unsigned> ProcIndices)1412 void CodeGenSchedModels::inferFromRW(ArrayRef<unsigned> OperWrites,
1413 ArrayRef<unsigned> OperReads,
1414 unsigned FromClassIdx,
1415 ArrayRef<unsigned> ProcIndices) {
1416 LLVM_DEBUG(dbgs() << "INFER RW proc("; dumpIdxVec(ProcIndices);
1417 dbgs() << ") ");
1418
1419 // Create a seed transition with an empty PredTerm and the expanded sequences
1420 // of SchedWrites for the current SchedClass.
1421 std::vector<PredTransition> LastTransitions;
1422 LastTransitions.emplace_back();
1423 LastTransitions.back().ProcIndices.append(ProcIndices.begin(),
1424 ProcIndices.end());
1425
1426 for (unsigned WriteIdx : OperWrites) {
1427 IdxVec WriteSeq;
1428 expandRWSequence(WriteIdx, WriteSeq, /*IsRead=*/false);
1429 LastTransitions[0].WriteSequences.emplace_back();
1430 SmallVectorImpl<unsigned> &Seq = LastTransitions[0].WriteSequences.back();
1431 Seq.append(WriteSeq.begin(), WriteSeq.end());
1432 LLVM_DEBUG(dbgs() << "("; dumpIdxVec(Seq); dbgs() << ") ");
1433 }
1434 LLVM_DEBUG(dbgs() << " Reads: ");
1435 for (unsigned ReadIdx : OperReads) {
1436 IdxVec ReadSeq;
1437 expandRWSequence(ReadIdx, ReadSeq, /*IsRead=*/true);
1438 LastTransitions[0].ReadSequences.emplace_back();
1439 SmallVectorImpl<unsigned> &Seq = LastTransitions[0].ReadSequences.back();
1440 Seq.append(ReadSeq.begin(), ReadSeq.end());
1441 LLVM_DEBUG(dbgs() << "("; dumpIdxVec(Seq); dbgs() << ") ");
1442 }
1443 LLVM_DEBUG(dbgs() << '\n');
1444
1445 // Collect all PredTransitions for individual operands.
1446 // Iterate until no variant writes remain.
1447 while (hasVariant(LastTransitions, *this)) {
1448 PredTransitions Transitions(*this);
1449 for (const PredTransition &Trans : LastTransitions)
1450 Transitions.substituteVariants(Trans);
1451 LLVM_DEBUG(Transitions.dump());
1452 LastTransitions.swap(Transitions.TransVec);
1453 }
1454 // If the first transition has no variants, nothing to do.
1455 if (LastTransitions[0].PredTerm.empty())
1456 return;
1457
1458 // WARNING: We are about to mutate the SchedClasses vector. Do not refer to
1459 // OperWrites, OperReads, or ProcIndices after calling inferFromTransitions.
1460 inferFromTransitions(LastTransitions, FromClassIdx, *this);
1461 }
1462
1463 // Check if any processor resource group contains all resource records in
1464 // SubUnits.
hasSuperGroup(RecVec & SubUnits,CodeGenProcModel & PM)1465 bool CodeGenSchedModels::hasSuperGroup(RecVec &SubUnits, CodeGenProcModel &PM) {
1466 for (unsigned i = 0, e = PM.ProcResourceDefs.size(); i < e; ++i) {
1467 if (!PM.ProcResourceDefs[i]->isSubClassOf("ProcResGroup"))
1468 continue;
1469 RecVec SuperUnits =
1470 PM.ProcResourceDefs[i]->getValueAsListOfDefs("Resources");
1471 RecIter RI = SubUnits.begin(), RE = SubUnits.end();
1472 for ( ; RI != RE; ++RI) {
1473 if (!is_contained(SuperUnits, *RI)) {
1474 break;
1475 }
1476 }
1477 if (RI == RE)
1478 return true;
1479 }
1480 return false;
1481 }
1482
1483 // Verify that overlapping groups have a common supergroup.
verifyProcResourceGroups(CodeGenProcModel & PM)1484 void CodeGenSchedModels::verifyProcResourceGroups(CodeGenProcModel &PM) {
1485 for (unsigned i = 0, e = PM.ProcResourceDefs.size(); i < e; ++i) {
1486 if (!PM.ProcResourceDefs[i]->isSubClassOf("ProcResGroup"))
1487 continue;
1488 RecVec CheckUnits =
1489 PM.ProcResourceDefs[i]->getValueAsListOfDefs("Resources");
1490 for (unsigned j = i+1; j < e; ++j) {
1491 if (!PM.ProcResourceDefs[j]->isSubClassOf("ProcResGroup"))
1492 continue;
1493 RecVec OtherUnits =
1494 PM.ProcResourceDefs[j]->getValueAsListOfDefs("Resources");
1495 if (std::find_first_of(CheckUnits.begin(), CheckUnits.end(),
1496 OtherUnits.begin(), OtherUnits.end())
1497 != CheckUnits.end()) {
1498 // CheckUnits and OtherUnits overlap
1499 OtherUnits.insert(OtherUnits.end(), CheckUnits.begin(),
1500 CheckUnits.end());
1501 if (!hasSuperGroup(OtherUnits, PM)) {
1502 PrintFatalError((PM.ProcResourceDefs[i])->getLoc(),
1503 "proc resource group overlaps with "
1504 + PM.ProcResourceDefs[j]->getName()
1505 + " but no supergroup contains both.");
1506 }
1507 }
1508 }
1509 }
1510 }
1511
1512 // Collect all the RegisterFile definitions available in this target.
collectRegisterFiles()1513 void CodeGenSchedModels::collectRegisterFiles() {
1514 RecVec RegisterFileDefs = Records.getAllDerivedDefinitions("RegisterFile");
1515
1516 // RegisterFiles is the vector of CodeGenRegisterFile.
1517 for (Record *RF : RegisterFileDefs) {
1518 // For each register file definition, construct a CodeGenRegisterFile object
1519 // and add it to the appropriate scheduling model.
1520 CodeGenProcModel &PM = getProcModel(RF->getValueAsDef("SchedModel"));
1521 PM.RegisterFiles.emplace_back(CodeGenRegisterFile(RF->getName(),RF));
1522 CodeGenRegisterFile &CGRF = PM.RegisterFiles.back();
1523
1524 // Now set the number of physical registers as well as the cost of registers
1525 // in each register class.
1526 CGRF.NumPhysRegs = RF->getValueAsInt("NumPhysRegs");
1527 RecVec RegisterClasses = RF->getValueAsListOfDefs("RegClasses");
1528 std::vector<int64_t> RegisterCosts = RF->getValueAsListOfInts("RegCosts");
1529 for (unsigned I = 0, E = RegisterClasses.size(); I < E; ++I) {
1530 int Cost = RegisterCosts.size() > I ? RegisterCosts[I] : 1;
1531 CGRF.Costs.emplace_back(RegisterClasses[I], Cost);
1532 }
1533 }
1534 }
1535
1536 // Collect all the RegisterFile definitions available in this target.
collectPfmCounters()1537 void CodeGenSchedModels::collectPfmCounters() {
1538 for (Record *Def : Records.getAllDerivedDefinitions("PfmIssueCounter")) {
1539 CodeGenProcModel &PM = getProcModel(Def->getValueAsDef("SchedModel"));
1540 PM.PfmIssueCounterDefs.emplace_back(Def);
1541 }
1542 for (Record *Def : Records.getAllDerivedDefinitions("PfmCycleCounter")) {
1543 CodeGenProcModel &PM = getProcModel(Def->getValueAsDef("SchedModel"));
1544 if (PM.PfmCycleCounterDef) {
1545 PrintFatalError(Def->getLoc(),
1546 "multiple cycle counters for " +
1547 Def->getValueAsDef("SchedModel")->getName());
1548 }
1549 PM.PfmCycleCounterDef = Def;
1550 }
1551 }
1552
1553 // Collect and sort WriteRes, ReadAdvance, and ProcResources.
collectProcResources()1554 void CodeGenSchedModels::collectProcResources() {
1555 ProcResourceDefs = Records.getAllDerivedDefinitions("ProcResourceUnits");
1556 ProcResGroups = Records.getAllDerivedDefinitions("ProcResGroup");
1557
1558 // Add any subtarget-specific SchedReadWrites that are directly associated
1559 // with processor resources. Refer to the parent SchedClass's ProcIndices to
1560 // determine which processors they apply to.
1561 for (const CodeGenSchedClass &SC :
1562 make_range(schedClassBegin(), schedClassEnd())) {
1563 if (SC.ItinClassDef) {
1564 collectItinProcResources(SC.ItinClassDef);
1565 continue;
1566 }
1567
1568 // This class may have a default ReadWrite list which can be overriden by
1569 // InstRW definitions.
1570 for (Record *RW : SC.InstRWs) {
1571 Record *RWModelDef = RW->getValueAsDef("SchedModel");
1572 unsigned PIdx = getProcModel(RWModelDef).Index;
1573 IdxVec Writes, Reads;
1574 findRWs(RW->getValueAsListOfDefs("OperandReadWrites"), Writes, Reads);
1575 collectRWResources(Writes, Reads, PIdx);
1576 }
1577
1578 collectRWResources(SC.Writes, SC.Reads, SC.ProcIndices);
1579 }
1580 // Add resources separately defined by each subtarget.
1581 RecVec WRDefs = Records.getAllDerivedDefinitions("WriteRes");
1582 for (Record *WR : WRDefs) {
1583 Record *ModelDef = WR->getValueAsDef("SchedModel");
1584 addWriteRes(WR, getProcModel(ModelDef).Index);
1585 }
1586 RecVec SWRDefs = Records.getAllDerivedDefinitions("SchedWriteRes");
1587 for (Record *SWR : SWRDefs) {
1588 Record *ModelDef = SWR->getValueAsDef("SchedModel");
1589 addWriteRes(SWR, getProcModel(ModelDef).Index);
1590 }
1591 RecVec RADefs = Records.getAllDerivedDefinitions("ReadAdvance");
1592 for (Record *RA : RADefs) {
1593 Record *ModelDef = RA->getValueAsDef("SchedModel");
1594 addReadAdvance(RA, getProcModel(ModelDef).Index);
1595 }
1596 RecVec SRADefs = Records.getAllDerivedDefinitions("SchedReadAdvance");
1597 for (Record *SRA : SRADefs) {
1598 if (SRA->getValueInit("SchedModel")->isComplete()) {
1599 Record *ModelDef = SRA->getValueAsDef("SchedModel");
1600 addReadAdvance(SRA, getProcModel(ModelDef).Index);
1601 }
1602 }
1603 // Add ProcResGroups that are defined within this processor model, which may
1604 // not be directly referenced but may directly specify a buffer size.
1605 RecVec ProcResGroups = Records.getAllDerivedDefinitions("ProcResGroup");
1606 for (Record *PRG : ProcResGroups) {
1607 if (!PRG->getValueInit("SchedModel")->isComplete())
1608 continue;
1609 CodeGenProcModel &PM = getProcModel(PRG->getValueAsDef("SchedModel"));
1610 if (!is_contained(PM.ProcResourceDefs, PRG))
1611 PM.ProcResourceDefs.push_back(PRG);
1612 }
1613 // Add ProcResourceUnits unconditionally.
1614 for (Record *PRU : Records.getAllDerivedDefinitions("ProcResourceUnits")) {
1615 if (!PRU->getValueInit("SchedModel")->isComplete())
1616 continue;
1617 CodeGenProcModel &PM = getProcModel(PRU->getValueAsDef("SchedModel"));
1618 if (!is_contained(PM.ProcResourceDefs, PRU))
1619 PM.ProcResourceDefs.push_back(PRU);
1620 }
1621 // Finalize each ProcModel by sorting the record arrays.
1622 for (CodeGenProcModel &PM : ProcModels) {
1623 llvm::sort(PM.WriteResDefs.begin(), PM.WriteResDefs.end(),
1624 LessRecord());
1625 llvm::sort(PM.ReadAdvanceDefs.begin(), PM.ReadAdvanceDefs.end(),
1626 LessRecord());
1627 llvm::sort(PM.ProcResourceDefs.begin(), PM.ProcResourceDefs.end(),
1628 LessRecord());
1629 LLVM_DEBUG(
1630 PM.dump();
1631 dbgs() << "WriteResDefs: "; for (RecIter RI = PM.WriteResDefs.begin(),
1632 RE = PM.WriteResDefs.end();
1633 RI != RE; ++RI) {
1634 if ((*RI)->isSubClassOf("WriteRes"))
1635 dbgs() << (*RI)->getValueAsDef("WriteType")->getName() << " ";
1636 else
1637 dbgs() << (*RI)->getName() << " ";
1638 } dbgs() << "\nReadAdvanceDefs: ";
1639 for (RecIter RI = PM.ReadAdvanceDefs.begin(),
1640 RE = PM.ReadAdvanceDefs.end();
1641 RI != RE; ++RI) {
1642 if ((*RI)->isSubClassOf("ReadAdvance"))
1643 dbgs() << (*RI)->getValueAsDef("ReadType")->getName() << " ";
1644 else
1645 dbgs() << (*RI)->getName() << " ";
1646 } dbgs()
1647 << "\nProcResourceDefs: ";
1648 for (RecIter RI = PM.ProcResourceDefs.begin(),
1649 RE = PM.ProcResourceDefs.end();
1650 RI != RE; ++RI) { dbgs() << (*RI)->getName() << " "; } dbgs()
1651 << '\n');
1652 verifyProcResourceGroups(PM);
1653 }
1654
1655 ProcResourceDefs.clear();
1656 ProcResGroups.clear();
1657 }
1658
checkCompleteness()1659 void CodeGenSchedModels::checkCompleteness() {
1660 bool Complete = true;
1661 bool HadCompleteModel = false;
1662 for (const CodeGenProcModel &ProcModel : procModels()) {
1663 const bool HasItineraries = ProcModel.hasItineraries();
1664 if (!ProcModel.ModelDef->getValueAsBit("CompleteModel"))
1665 continue;
1666 for (const CodeGenInstruction *Inst : Target.getInstructionsByEnumValue()) {
1667 if (Inst->hasNoSchedulingInfo)
1668 continue;
1669 if (ProcModel.isUnsupported(*Inst))
1670 continue;
1671 unsigned SCIdx = getSchedClassIdx(*Inst);
1672 if (!SCIdx) {
1673 if (Inst->TheDef->isValueUnset("SchedRW") && !HadCompleteModel) {
1674 PrintError("No schedule information for instruction '"
1675 + Inst->TheDef->getName() + "'");
1676 Complete = false;
1677 }
1678 continue;
1679 }
1680
1681 const CodeGenSchedClass &SC = getSchedClass(SCIdx);
1682 if (!SC.Writes.empty())
1683 continue;
1684 if (HasItineraries && SC.ItinClassDef != nullptr &&
1685 SC.ItinClassDef->getName() != "NoItinerary")
1686 continue;
1687
1688 const RecVec &InstRWs = SC.InstRWs;
1689 auto I = find_if(InstRWs, [&ProcModel](const Record *R) {
1690 return R->getValueAsDef("SchedModel") == ProcModel.ModelDef;
1691 });
1692 if (I == InstRWs.end()) {
1693 PrintError("'" + ProcModel.ModelName + "' lacks information for '" +
1694 Inst->TheDef->getName() + "'");
1695 Complete = false;
1696 }
1697 }
1698 HadCompleteModel = true;
1699 }
1700 if (!Complete) {
1701 errs() << "\n\nIncomplete schedule models found.\n"
1702 << "- Consider setting 'CompleteModel = 0' while developing new models.\n"
1703 << "- Pseudo instructions can be marked with 'hasNoSchedulingInfo = 1'.\n"
1704 << "- Instructions should usually have Sched<[...]> as a superclass, "
1705 "you may temporarily use an empty list.\n"
1706 << "- Instructions related to unsupported features can be excluded with "
1707 "list<Predicate> UnsupportedFeatures = [HasA,..,HasY]; in the "
1708 "processor model.\n\n";
1709 PrintFatalError("Incomplete schedule model");
1710 }
1711 }
1712
1713 // Collect itinerary class resources for each processor.
collectItinProcResources(Record * ItinClassDef)1714 void CodeGenSchedModels::collectItinProcResources(Record *ItinClassDef) {
1715 for (unsigned PIdx = 0, PEnd = ProcModels.size(); PIdx != PEnd; ++PIdx) {
1716 const CodeGenProcModel &PM = ProcModels[PIdx];
1717 // For all ItinRW entries.
1718 bool HasMatch = false;
1719 for (RecIter II = PM.ItinRWDefs.begin(), IE = PM.ItinRWDefs.end();
1720 II != IE; ++II) {
1721 RecVec Matched = (*II)->getValueAsListOfDefs("MatchedItinClasses");
1722 if (!std::count(Matched.begin(), Matched.end(), ItinClassDef))
1723 continue;
1724 if (HasMatch)
1725 PrintFatalError((*II)->getLoc(), "Duplicate itinerary class "
1726 + ItinClassDef->getName()
1727 + " in ItinResources for " + PM.ModelName);
1728 HasMatch = true;
1729 IdxVec Writes, Reads;
1730 findRWs((*II)->getValueAsListOfDefs("OperandReadWrites"), Writes, Reads);
1731 collectRWResources(Writes, Reads, PIdx);
1732 }
1733 }
1734 }
1735
collectRWResources(unsigned RWIdx,bool IsRead,ArrayRef<unsigned> ProcIndices)1736 void CodeGenSchedModels::collectRWResources(unsigned RWIdx, bool IsRead,
1737 ArrayRef<unsigned> ProcIndices) {
1738 const CodeGenSchedRW &SchedRW = getSchedRW(RWIdx, IsRead);
1739 if (SchedRW.TheDef) {
1740 if (!IsRead && SchedRW.TheDef->isSubClassOf("SchedWriteRes")) {
1741 for (unsigned Idx : ProcIndices)
1742 addWriteRes(SchedRW.TheDef, Idx);
1743 }
1744 else if (IsRead && SchedRW.TheDef->isSubClassOf("SchedReadAdvance")) {
1745 for (unsigned Idx : ProcIndices)
1746 addReadAdvance(SchedRW.TheDef, Idx);
1747 }
1748 }
1749 for (RecIter AI = SchedRW.Aliases.begin(), AE = SchedRW.Aliases.end();
1750 AI != AE; ++AI) {
1751 IdxVec AliasProcIndices;
1752 if ((*AI)->getValueInit("SchedModel")->isComplete()) {
1753 AliasProcIndices.push_back(
1754 getProcModel((*AI)->getValueAsDef("SchedModel")).Index);
1755 }
1756 else
1757 AliasProcIndices = ProcIndices;
1758 const CodeGenSchedRW &AliasRW = getSchedRW((*AI)->getValueAsDef("AliasRW"));
1759 assert(AliasRW.IsRead == IsRead && "cannot alias reads to writes");
1760
1761 IdxVec ExpandedRWs;
1762 expandRWSequence(AliasRW.Index, ExpandedRWs, IsRead);
1763 for (IdxIter SI = ExpandedRWs.begin(), SE = ExpandedRWs.end();
1764 SI != SE; ++SI) {
1765 collectRWResources(*SI, IsRead, AliasProcIndices);
1766 }
1767 }
1768 }
1769
1770 // Collect resources for a set of read/write types and processor indices.
collectRWResources(ArrayRef<unsigned> Writes,ArrayRef<unsigned> Reads,ArrayRef<unsigned> ProcIndices)1771 void CodeGenSchedModels::collectRWResources(ArrayRef<unsigned> Writes,
1772 ArrayRef<unsigned> Reads,
1773 ArrayRef<unsigned> ProcIndices) {
1774 for (unsigned Idx : Writes)
1775 collectRWResources(Idx, /*IsRead=*/false, ProcIndices);
1776
1777 for (unsigned Idx : Reads)
1778 collectRWResources(Idx, /*IsRead=*/true, ProcIndices);
1779 }
1780
1781 // Find the processor's resource units for this kind of resource.
findProcResUnits(Record * ProcResKind,const CodeGenProcModel & PM,ArrayRef<SMLoc> Loc) const1782 Record *CodeGenSchedModels::findProcResUnits(Record *ProcResKind,
1783 const CodeGenProcModel &PM,
1784 ArrayRef<SMLoc> Loc) const {
1785 if (ProcResKind->isSubClassOf("ProcResourceUnits"))
1786 return ProcResKind;
1787
1788 Record *ProcUnitDef = nullptr;
1789 assert(!ProcResourceDefs.empty());
1790 assert(!ProcResGroups.empty());
1791
1792 for (Record *ProcResDef : ProcResourceDefs) {
1793 if (ProcResDef->getValueAsDef("Kind") == ProcResKind
1794 && ProcResDef->getValueAsDef("SchedModel") == PM.ModelDef) {
1795 if (ProcUnitDef) {
1796 PrintFatalError(Loc,
1797 "Multiple ProcessorResourceUnits associated with "
1798 + ProcResKind->getName());
1799 }
1800 ProcUnitDef = ProcResDef;
1801 }
1802 }
1803 for (Record *ProcResGroup : ProcResGroups) {
1804 if (ProcResGroup == ProcResKind
1805 && ProcResGroup->getValueAsDef("SchedModel") == PM.ModelDef) {
1806 if (ProcUnitDef) {
1807 PrintFatalError(Loc,
1808 "Multiple ProcessorResourceUnits associated with "
1809 + ProcResKind->getName());
1810 }
1811 ProcUnitDef = ProcResGroup;
1812 }
1813 }
1814 if (!ProcUnitDef) {
1815 PrintFatalError(Loc,
1816 "No ProcessorResources associated with "
1817 + ProcResKind->getName());
1818 }
1819 return ProcUnitDef;
1820 }
1821
1822 // Iteratively add a resource and its super resources.
addProcResource(Record * ProcResKind,CodeGenProcModel & PM,ArrayRef<SMLoc> Loc)1823 void CodeGenSchedModels::addProcResource(Record *ProcResKind,
1824 CodeGenProcModel &PM,
1825 ArrayRef<SMLoc> Loc) {
1826 while (true) {
1827 Record *ProcResUnits = findProcResUnits(ProcResKind, PM, Loc);
1828
1829 // See if this ProcResource is already associated with this processor.
1830 if (is_contained(PM.ProcResourceDefs, ProcResUnits))
1831 return;
1832
1833 PM.ProcResourceDefs.push_back(ProcResUnits);
1834 if (ProcResUnits->isSubClassOf("ProcResGroup"))
1835 return;
1836
1837 if (!ProcResUnits->getValueInit("Super")->isComplete())
1838 return;
1839
1840 ProcResKind = ProcResUnits->getValueAsDef("Super");
1841 }
1842 }
1843
1844 // Add resources for a SchedWrite to this processor if they don't exist.
addWriteRes(Record * ProcWriteResDef,unsigned PIdx)1845 void CodeGenSchedModels::addWriteRes(Record *ProcWriteResDef, unsigned PIdx) {
1846 assert(PIdx && "don't add resources to an invalid Processor model");
1847
1848 RecVec &WRDefs = ProcModels[PIdx].WriteResDefs;
1849 if (is_contained(WRDefs, ProcWriteResDef))
1850 return;
1851 WRDefs.push_back(ProcWriteResDef);
1852
1853 // Visit ProcResourceKinds referenced by the newly discovered WriteRes.
1854 RecVec ProcResDefs = ProcWriteResDef->getValueAsListOfDefs("ProcResources");
1855 for (RecIter WritePRI = ProcResDefs.begin(), WritePRE = ProcResDefs.end();
1856 WritePRI != WritePRE; ++WritePRI) {
1857 addProcResource(*WritePRI, ProcModels[PIdx], ProcWriteResDef->getLoc());
1858 }
1859 }
1860
1861 // Add resources for a ReadAdvance to this processor if they don't exist.
addReadAdvance(Record * ProcReadAdvanceDef,unsigned PIdx)1862 void CodeGenSchedModels::addReadAdvance(Record *ProcReadAdvanceDef,
1863 unsigned PIdx) {
1864 RecVec &RADefs = ProcModels[PIdx].ReadAdvanceDefs;
1865 if (is_contained(RADefs, ProcReadAdvanceDef))
1866 return;
1867 RADefs.push_back(ProcReadAdvanceDef);
1868 }
1869
getProcResourceIdx(Record * PRDef) const1870 unsigned CodeGenProcModel::getProcResourceIdx(Record *PRDef) const {
1871 RecIter PRPos = find(ProcResourceDefs, PRDef);
1872 if (PRPos == ProcResourceDefs.end())
1873 PrintFatalError(PRDef->getLoc(), "ProcResource def is not included in "
1874 "the ProcResources list for " + ModelName);
1875 // Idx=0 is reserved for invalid.
1876 return 1 + (PRPos - ProcResourceDefs.begin());
1877 }
1878
isUnsupported(const CodeGenInstruction & Inst) const1879 bool CodeGenProcModel::isUnsupported(const CodeGenInstruction &Inst) const {
1880 for (const Record *TheDef : UnsupportedFeaturesDefs) {
1881 for (const Record *PredDef : Inst.TheDef->getValueAsListOfDefs("Predicates")) {
1882 if (TheDef->getName() == PredDef->getName())
1883 return true;
1884 }
1885 }
1886 return false;
1887 }
1888
1889 #ifndef NDEBUG
dump() const1890 void CodeGenProcModel::dump() const {
1891 dbgs() << Index << ": " << ModelName << " "
1892 << (ModelDef ? ModelDef->getName() : "inferred") << " "
1893 << (ItinsDef ? ItinsDef->getName() : "no itinerary") << '\n';
1894 }
1895
dump() const1896 void CodeGenSchedRW::dump() const {
1897 dbgs() << Name << (IsVariadic ? " (V) " : " ");
1898 if (IsSequence) {
1899 dbgs() << "(";
1900 dumpIdxVec(Sequence);
1901 dbgs() << ")";
1902 }
1903 }
1904
dump(const CodeGenSchedModels * SchedModels) const1905 void CodeGenSchedClass::dump(const CodeGenSchedModels* SchedModels) const {
1906 dbgs() << "SCHEDCLASS " << Index << ":" << Name << '\n'
1907 << " Writes: ";
1908 for (unsigned i = 0, N = Writes.size(); i < N; ++i) {
1909 SchedModels->getSchedWrite(Writes[i]).dump();
1910 if (i < N-1) {
1911 dbgs() << '\n';
1912 dbgs().indent(10);
1913 }
1914 }
1915 dbgs() << "\n Reads: ";
1916 for (unsigned i = 0, N = Reads.size(); i < N; ++i) {
1917 SchedModels->getSchedRead(Reads[i]).dump();
1918 if (i < N-1) {
1919 dbgs() << '\n';
1920 dbgs().indent(10);
1921 }
1922 }
1923 dbgs() << "\n ProcIdx: "; dumpIdxVec(ProcIndices); dbgs() << '\n';
1924 if (!Transitions.empty()) {
1925 dbgs() << "\n Transitions for Proc ";
1926 for (const CodeGenSchedTransition &Transition : Transitions) {
1927 dumpIdxVec(Transition.ProcIndices);
1928 }
1929 }
1930 }
1931
dump() const1932 void PredTransitions::dump() const {
1933 dbgs() << "Expanded Variants:\n";
1934 for (std::vector<PredTransition>::const_iterator
1935 TI = TransVec.begin(), TE = TransVec.end(); TI != TE; ++TI) {
1936 dbgs() << "{";
1937 for (SmallVectorImpl<PredCheck>::const_iterator
1938 PCI = TI->PredTerm.begin(), PCE = TI->PredTerm.end();
1939 PCI != PCE; ++PCI) {
1940 if (PCI != TI->PredTerm.begin())
1941 dbgs() << ", ";
1942 dbgs() << SchedModels.getSchedRW(PCI->RWIdx, PCI->IsRead).Name
1943 << ":" << PCI->Predicate->getName();
1944 }
1945 dbgs() << "},\n => {";
1946 for (SmallVectorImpl<SmallVector<unsigned,4>>::const_iterator
1947 WSI = TI->WriteSequences.begin(), WSE = TI->WriteSequences.end();
1948 WSI != WSE; ++WSI) {
1949 dbgs() << "(";
1950 for (SmallVectorImpl<unsigned>::const_iterator
1951 WI = WSI->begin(), WE = WSI->end(); WI != WE; ++WI) {
1952 if (WI != WSI->begin())
1953 dbgs() << ", ";
1954 dbgs() << SchedModels.getSchedWrite(*WI).Name;
1955 }
1956 dbgs() << "),";
1957 }
1958 dbgs() << "}\n";
1959 }
1960 }
1961 #endif // NDEBUG
1962